/* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License, Version 1.0 only * (the "License"). You may not use this file except in compliance * with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright 2005 Sun Microsystems, Inc. All rights reserved. * Use is subject to license terms. */ #pragma ident "%Z%%M% %I% %E% SMI" /* * Inter-Domain Network * * IDN Protocol functions to support domain link/unlink/reconfig. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* kas decl */ #include #include #define IDNBUG_CPUPERBOARD extern pri_t maxclsyspri; extern u_longlong_t gettick(); clock_t idn_xmit_monitor_freq = 50; static int idn_connect(int domid); static int idn_disconnect(int domid, idn_fin_t fintype, idn_finarg_t finarg, idn_finsync_t finsync); static void idn_deconfig(int domid); static void idn_unlink_domainset(domainset_t domset, idn_fin_t fintype, idn_finarg_t finarg, idn_finopt_t finopt, boardset_t idnset); static void idn_retry_execute(void *arg); static void idn_retry_submit(void (*func)(uint_t token, void *arg), void *arg, uint_t token, clock_t ticks); static void idn_shutdown_datapath(domainset_t domset, int force); static mblk_t *idn_fill_buffer(caddr_t bufp, int size, mblk_t *mp, uchar_t **data_rptrp); static ushort_t idn_cksum(register ushort_t *hdrp, register int count); static int idn_mark_awol(int domid, clock_t *atime); static void idn_recv_proto(idn_protomsg_t *hp); static void idn_send_config(int domid, int phase); static void idn_recv_config(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static int idn_send_master_config(int domid, int phase); static int idn_send_slave_config(int domid, int phase); static uint_t idn_check_master_config(int domid, uint_t *exp, uint_t *act); static uint_t idn_check_slave_config(int domid, uint_t *exp, uint_t *act); static int idn_recv_config_done(int domid); static void idn_nego_cleanup_check(int domid, int new_masterid, int new_cpuid); static void idn_recv_cmd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static int idn_recv_data(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static int idn_send_data_loopback(idn_netaddr_t dst_netaddr, queue_t *wq, mblk_t *mp); static void idn_send_dataresp(int domid, idn_nack_t nacktype); static int idn_send_mboxdata(int domid, struct idn *sip, int channel, caddr_t bufp); static int idn_recv_mboxdata(int channel, caddr_t bufp); static int idn_program_hardware(int domid); static int idn_deprogram_hardware(int domid); static void idn_send_cmd_nackresp(int domid, idn_msgtype_t *mtp, idn_cmd_t cmdtype, idn_nack_t nacktype); static void idn_local_cmd(idn_cmd_t cmdtype, uint_t arg1, uint_t arg2, uint_t arg3); static void idn_terminate_cmd(int domid, int serrno); static void idn_mboxarea_init(idn_mboxtbl_t *mtp, register int ntbls); static void idn_mainmbox_activate(int domid); static void idn_mainmbox_deactivate(ushort_t domset); static void idn_mainmbox_chan_register(int domid, idn_mainmbox_t *send_mmp, idn_mainmbox_t *recv_mmp, int channel); static int idn_mainmbox_chan_unregister(ushort_t domset, int channel); static int idn_mainmbox_flush(int domid, idn_mainmbox_t *mmp); static void idn_mainmbox_reset(int domid, idn_mainmbox_t *cmp); static int idn_activate_channel(idn_chanset_t chanset, idn_chanop_t chanop); static void idn_deactivate_channel(idn_chanset_t chanset, idn_chanop_t chanop); static int idn_deactivate_channel_services(int channel, idn_chanop_t chanop); static int idn_activate_channel_services(int channel); static void idn_chan_server(idn_chansvr_t **cspp); #if 0 static void idn_chan_flush(idn_chansvr_t *csp); #endif /* 0 */ static void idn_chan_action(int channel, idn_chanaction_t chanaction, int wait); static void idn_chan_addmbox(int channel, ushort_t domset); static void idn_chan_delmbox(int channel, ushort_t domset); static void idn_submit_chanactivate_job(int channel); static void idn_exec_chanactivate(void *chn); static void idn_link_established(void *arg); static void idn_prealloc_slab(int nslabs); static void idn_recv_slaballoc_req(int domid, idn_msgtype_t *mtp, uint_t slab_size); static void idn_send_slaballoc_resp(int domid, idn_msgtype_t *mtp, uint_t slab_offset, uint_t slab_size, int serrno); static void idn_recv_slaballoc_resp(int domid, smr_offset_t slab_offset, uint_t slab_size, int serrno); static void idn_recv_slabreap_req(int domid, idn_msgtype_t *mtp, int nslabs); static void idn_recv_slabreap_resp(int domid, int nslabs, int serrno); static void idn_send_slabreap_resp(int domid, idn_msgtype_t *mtp, int nslabs, int serrno); static void idn_recv_slabfree_req(int domid, idn_msgtype_t *mtp, smr_offset_t slab_offset, uint_t slab_size); static void idn_recv_slabfree_resp(int domid, uint_t slab_offset, uint_t slab_size, int serrno); static void idn_send_slabfree_resp(int domid, idn_msgtype_t *mtp, uint_t slab_offset, uint_t slab_size, int serrno); static void idn_retry_nodename_req(void *arg); static void idn_send_nodename_req(int domid); static void idn_send_nodename_resp(int domid, idn_msgtype_t *mtp, uint_t bufoffset, int serrno); static void idn_recv_nodename_req(int domid, idn_msgtype_t *mtp, uint_t bufoffset); static void idn_recv_nodename_resp(int domid, uint_t bufoffset, int serrno); static void idn_protocol_server(int *id); static void idn_protocol_server_killall(); static void idn_protojob_free(idn_protojob_t *jp); static int idn_xstate_transfunc(int domid, void *transarg); static int idn_xphase_transition(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_sync_enter(int domid, idn_synccmd_t cmd, domainset_t xset, domainset_t rset, int (*transfunc)(), void *transarg); static domainset_t idn_sync_register(int domid, idn_synccmd_t cmd, domainset_t ready_set, idn_syncreg_t regtype); static void idn_sync_register_awol(int domid); static int idn_verify_config_mbox(int domid); static int idn_select_master(int domid, int rmasterid, int rcpuid); static int valid_mtu(uint_t mtu); static int valid_bufsize(uint_t bufsize); static int valid_slabsize(int slabsize); static int valid_nwrsize(int nwrsize); static int idn_master_init(); static void idn_master_deinit(); static void idn_send_acknack(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static int idn_send_nego(int domid, idn_msgtype_t *mtp, domainset_t conset); static void idn_retry_nego(uint_t token, void *arg); static int idn_check_nego(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_nego_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_error_nego(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_nego_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_nego_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_final_nego(int domid); static void idn_exit_nego(int domid, uint_t msgtype); static int idn_send_con(int domid, idn_msgtype_t *mtp, idn_con_t contype, domainset_t conset); static void idn_retry_con(uint_t token, void *arg); static int idn_check_con(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_con_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_error_con(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_con_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_con_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_final_con(int domid); static void idn_exit_con(int domid, uint_t msgtype); static int idn_send_fin(int domid, idn_msgtype_t *mtp, idn_fin_t fintype, idn_finarg_t finarg, idn_finopt_t finopt, domainset_t finset, uint_t finmaster); static void idn_retry_fin(uint_t token, void *arg); static int idn_check_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_error_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static int idn_check_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_error_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_action_fin_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs); static void idn_final_fin(int domid); static void idn_exit_fin(int domid, uint_t msgtype); /* * We keep a small cache of protojob structures just * in case allocation within idn_handler comes back * with nothing from the land of kmem. */ idn_protojob_t idn_protojob_cache[IDN_DMV_PENDING_MAX]; idn_protojob_t *idn_protojob_cache_list; kmutex_t idn_protojob_cache_lock; /* * - receive message. * - call check-function for current state. * - if (check-function == ok) then * call action-function for current state. * else * call error-function for current state. * - transition state based on check results. * - if (next state == final state) then * call final-function. */ static idn_xphase_t xphase_nego = { IDNP_NEGO, { { IDNDS_NEGO_PEND, idn_check_nego, idn_action_nego_pend, idn_error_nego}, { IDNDS_NEGO_SENT, idn_check_nego, idn_action_nego_sent, idn_error_nego}, { IDNDS_NEGO_RCVD, NULL, idn_action_nego_rcvd, NULL }, { IDNDS_CONFIG, NULL, NULL, NULL }, }, idn_final_nego, idn_exit_nego }; static idn_xphase_t xphase_con = { IDNP_CON, { { IDNDS_CON_PEND, idn_check_con, idn_action_con_pend, idn_error_con}, { IDNDS_CON_SENT, idn_check_con, idn_action_con_sent, idn_error_con}, { IDNDS_CON_RCVD, NULL, idn_action_con_rcvd, NULL }, { IDNDS_CON_READY, NULL, NULL, NULL }, }, idn_final_con, idn_exit_con }; static idn_xphase_t xphase_fin = { IDNP_FIN, { { IDNDS_FIN_PEND, idn_check_fin_pend, idn_action_fin_pend, idn_error_fin_pend }, { IDNDS_FIN_SENT, idn_check_fin_sent, idn_action_fin_sent, idn_error_fin_sent }, { IDNDS_FIN_RCVD, NULL, idn_action_fin_rcvd, NULL }, { IDNDS_DMAP, NULL, NULL, NULL }, }, idn_final_fin, idn_exit_fin }; static int idnxs_state_table[4][5][2] = { { /* IDNXS_PEND */ { IDNXS_SENT, IDNXS_PEND }, /* 0 */ { IDNXS_RCVD, IDNXS_PEND }, /* msg */ { IDNXS_NIL, IDNXS_PEND }, /* msg+ack */ { IDNXS_PEND, IDNXS_NIL }, /* ack */ { IDNXS_PEND, IDNXS_NIL }, /* nack */ }, { /* IDNXS_SENT */ { IDNXS_NIL, IDNXS_NIL }, /* 0 */ { IDNXS_RCVD, IDNXS_PEND }, /* msg */ { IDNXS_FINAL, IDNXS_PEND }, /* msg+ack */ { IDNXS_NIL, IDNXS_NIL }, /* ack */ { IDNXS_PEND, IDNXS_NIL }, /* nack */ }, { /* IDNXS_RCVD */ { IDNXS_NIL, IDNXS_NIL }, /* 0 */ { IDNXS_NIL, IDNXS_NIL }, /* msg */ { IDNXS_FINAL, IDNXS_NIL }, /* msg+ack */ { IDNXS_FINAL, IDNXS_NIL }, /* ack */ { IDNXS_PEND, IDNXS_NIL }, /* nack */ }, { /* IDNXS_FINAL */ { IDNXS_NIL, IDNXS_NIL }, /* 0 */ { IDNXS_NIL, IDNXS_NIL }, /* msg */ { IDNXS_NIL, IDNXS_NIL }, /* msg+ack */ { IDNXS_NIL, IDNXS_NIL }, /* ack */ { IDNXS_NIL, IDNXS_NIL }, /* nack */ } }; /* * NONE Respective domain does not have a master. * OTHER Respective domain has a master different * than either local or remote. * LOCAL Respective domain has chosen local as master. * REMOTE Respective domain has chosen remote as master. * * Actions: * VOTE Compare votes and select one. * VOTE_RCFG Compare votes and Reconfigure * if necessary, i.e. remote won. * CONNECT Connect to remote's OTHER if different * than our local master. * LOCAL Local domain is winner. * REMOTE Remote domain is winner. * WAIT Wait for remote to connect to our * master if his is different. * ERROR An impossible condition. * * Index: * 0 = Local * 1 = Remote */ static idn_master_select_t master_select_table[4][4] = { { /* local remote */ MASTER_SELECT_VOTE, /* NONE NONE */ MASTER_SELECT_CONNECT, /* NONE OTHER */ MASTER_SELECT_LOCAL, /* NONE LOCAL */ MASTER_SELECT_REMOTE /* NONE REMOTE */ }, { MASTER_SELECT_WAIT, /* OTHER NONE */ MASTER_SELECT_CONNECT, /* OTHER OTHER */ MASTER_SELECT_WAIT, /* OTHER LOCAL */ MASTER_SELECT_WAIT /* OTHER REMOTE */ }, { MASTER_SELECT_LOCAL, /* LOCAL NONE */ MASTER_SELECT_CONNECT, /* LOCAL OTHER */ MASTER_SELECT_LOCAL, /* LOCAL LOCAL */ MASTER_SELECT_VOTE_RCFG /* LOCAL REMOTE */ }, { MASTER_SELECT_REMOTE, /* REMOTE NONE */ MASTER_SELECT_CONNECT, /* REMOTE OTHER */ MASTER_SELECT_ERROR, /* REMOTE LOCAL */ MASTER_SELECT_REMOTE /* REMOTE REMOTE */ } }; void idn_assign_cookie(int domid) { static ushort_t num = 0; ushort_t cookie; procname_t proc = "idn_assign_cookie"; if ((cookie = idn_domain[domid].dcookie_recv) != 0) return; cookie = (ushort_t)(((uint64_t)&idn_domain[domid] >> 8) & 0xff); while ((cookie ^= num++ & 0xff) == 0) ; PR_PROTO("%s:%d: assigned RECV cookie 0x%x\n", proc, domid, cookie); idn_domain[domid].dcookie_recv = cookie; } void idn_update_priority(int domid, int pri) { idn_domain_t *dp; procname_t proc = "idn_update_priority"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; if (pri >= IDNVOTE_MINPRI) { dp->dvote.v.priority = pri & IDNVOTE_PRI_MASK; PR_PROTO("%s:%d: SETTING PRIORITY to req(%d) " "(localpri = 0x%x)\n", proc, domid, pri, IDNVOTE_PRIVALUE(dp->dvote)); } else { PR_PROTO("%s:%d: PRIORITIES UNCHANGED (pri = 0x%x)\n", proc, domid, IDNVOTE_PRIVALUE(dp->dvote)); } } /* * Initiate a link between the local domain and the remote domain * containing the given cpuid. */ int idn_link(int domid, int cpuid, int pri, int waittime, idnsb_error_t *sep) { int rv; idn_domain_t *dp; void *opcookie; procname_t proc = "idn_link"; if ((cpuid < 0) || (cpuid >= NCPU)) { cmn_err(CE_WARN, "IDN: 201: (LINK) invalid CPU ID (%d)", cpuid); return (EINVAL); } if (waittime < 0) { cmn_err(CE_WARN, "IDN: 202: (LINK) invalid time-out value (%d)", waittime); return (EINVAL); } if (!VALID_DOMAINID(domid)) { cmn_err(CE_WARN, "IDN: 203: (LINK) invalid domain ID (%d)", domid); return (EINVAL); } if (domid == idn.localid) return (0); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); dp = &idn_domain[domid]; switch (dp->dstate) { case IDNDS_CLOSED: break; case IDNDS_CONNECTED: #ifdef DEBUG cmn_err(CE_NOTE, "!IDN: domain %d (CPU ID %d) already connected", domid, cpuid); #endif /* DEBUG */ IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return (0); default: cmn_err(CE_WARN, "IDN: 204: domain %d state (%s) inappropriate", domid, idnds_str[dp->dstate]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return (EINVAL); } rv = idn_open_domain(domid, cpuid, 0); if (rv != 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to open-domain(%d,%d)", proc, domid, cpuid); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return (EIO); } IDN_DLOCK_EXCL(idn.localid); idn_update_priority(idn.localid, pri); IDN_DUNLOCK(idn.localid); if (waittime > 0) opcookie = idn_init_op(IDNOP_CONNECTED, DOMAINSET(domid), sep); idn_connect(domid); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); PR_PROTO("%s:%d: ALLOCATED idn_link(%d)\n", proc, domid, cpuid); if (waittime > 0) { boardset_t domset = 0; /* * Well we've successfully allocated a domain id, * but the link may not be fully established yet. * Need to wait since it happens asynchronously. */ PR_PROTO("%s:%d: WAITING for op(%s) for (domset 0%x)...\n", proc, domid, idnop_str[IDNOP_CONNECTED], DOMAINSET(domid)); rv = idn_wait_op(opcookie, &domset, waittime); } #ifdef DEBUG if (rv == 0) { if (waittime > 0) { PR_PROTO("%s:%d: connect SUCCEEDED (cpu %d)\n", proc, domid, cpuid); } else { PR_PROTO("%s:%d: connect KICKED OFF (cpu %d)\n", proc, domid, cpuid); } } else { PR_PROTO("%s:%d: connect FAILED (cpu %d)\n", proc, domid, cpuid); } #endif /* DEBUG */ return (rv); } /* * Unlink the given domain from any domain cluster of * which it might be a member. Force indicates that domain * should not go AWOL and if it's currently AWOL to close * and remove it. * IMPORTANT: If the (hard) force flag is set, the caller is * assumed to GUARANTEE that the given domain will * not attempt to communicate with the local domain * in any manner. */ int idn_unlink(int domid, boardset_t idnset, idn_fin_t fintype, idn_finopt_t finopt, int waittime, idnsb_error_t *sep) { int rv = 0; domainset_t domset; void *opcookie; procname_t proc = "idn_unlink"; if (waittime < 0) { cmn_err(CE_WARN, "IDN: 202: (UNLINK) invalid time-out value (%d)", waittime); SET_IDNKERR_IDNERR(sep, IDNKERR_INVALID_WTIME); SET_IDNKERR_PARAM0(sep, waittime); return (EINVAL); } if (!VALID_DOMAINID(domid)) { cmn_err(CE_WARN, "IDN: 203: (UNLINK) invalid domain ID (%d)", domid); SET_IDNKERR_IDNERR(sep, IDNKERR_INVALID_DOMAIN); SET_IDNKERR_PARAM0(sep, domid); SET_IDNKERR_PARAM1(sep, -1); return (EINVAL); } if (idn.localid == IDN_NIL_DOMID) { #ifdef DEBUG cmn_err(CE_NOTE, "!IDN: %s: local domain not connected to an IDNnet", proc); #endif /* DEBUG */ return (0); } /* * Lock ordering protocols requires that we grab the * global lock _before_ the local domain's lock. * However, non-local domains must have their lock * grabbed _before_ the global lock. */ IDN_SYNC_LOCK(); IDN_GLOCK_EXCL(); domset = idn.domset.ds_trans_on | idn.domset.ds_trans_off; if ((idn.state == IDNGS_OFFLINE) && !domset) { #ifdef DEBUG cmn_err(CE_WARN, "!IDN: %s: local domain not connected to an IDNnet", proc); #endif /* DEBUG */ IDN_GUNLOCK(); IDN_SYNC_UNLOCK(); return (0); } if ((domid == IDN_NIL_DOMID) || (domid == idn.localid)) { domid = idn.localid; IDN_GSTATE_TRANSITION(IDNGS_DISCONNECT); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); domset = DOMAINSET_ALL; DOMAINSET_DEL(domset, idn.localid); } else { domset = DOMAINSET(domid); } IDN_GUNLOCK(); if (waittime > 0) opcookie = idn_init_op(IDNOP_DISCONNECTED, domset, sep); idn_unlink_domainset(domset, fintype, IDNFIN_ARG_NONE, finopt, idnset); IDN_SYNC_UNLOCK(); if (waittime > 0) { /* * Well the unlink has successfully kicked off. * Since process is asynchronous we need to wait * for it to complete. */ PR_PROTO("%s:%d: WAITING for op(%s) for (domset 0%x)...\n", proc, domid, idnop_str[IDNOP_DISCONNECTED], domset); rv = idn_wait_op(opcookie, &domset, waittime); } if (rv == 0) { if (waittime > 0) { PR_PROTO("%s:%d: disconnect SUCCEEDED\n", proc, domid); } else { PR_PROTO("%s:%d: disconnect KICKED OFF\n", proc, domid); } } else { PR_PROTO("%s:%d: disconnect FAILED\n", proc, domid); } return (rv); } static void idn_unlink_domainset(domainset_t domset, idn_fin_t fintype, idn_finarg_t finarg, idn_finopt_t finopt, boardset_t idnset) { int d; domainset_t offset; procname_t proc = "idn_unlink_domainset"; ASSERT(IDN_SYNC_IS_LOCKED()); /* * Determine subset for which we have * no active connections. */ offset = domset & ~(idn.domset.ds_trans_on | idn.domset.ds_connected | idn.domset.ds_trans_off | idn.domset.ds_relink); /* * Determine subset that are really candidates. * Note that we include those already down the path * since it's possible a request came in to upgrade * their fintype (e.g. NORMAL->FORCE_SOFT). */ domset &= ~offset; if (offset) idn_update_op(IDNOP_DISCONNECTED, offset, NULL); IDN_GLOCK_EXCL(); if ((finopt == IDNFIN_OPT_RELINK) && (idn.state != IDNGS_DISCONNECT)) { /* * Don't add domains already transitioning off. * If they caught on an earlier Reconfig wave then * they'll already be in ds_relink anyway. Otherwise, * once a domain is transition off we can't upgrade * him to a RELINK. */ #ifdef DEBUG if (idn.domset.ds_hitlist & domset) { PR_HITLIST("%s: domset=%x, hitlist=%x, trans_off=%x " "-> relink = %x -> %x\n", proc, domset, idn.domset.ds_hitlist, idn.domset.ds_relink, idn.domset.ds_trans_off, idn.domset.ds_relink | (domset & ~idn.domset.ds_trans_off)); } #endif /* DEBUG */ domset &= ~idn.domset.ds_trans_off; idn.domset.ds_relink |= domset; } else { idn.domset.ds_relink &= ~domset; } /* * Update the ds_trans_on/off so we don't waste * time talking to these folks. */ idn.domset.ds_trans_on &= ~domset; idn.domset.ds_trans_off |= domset; if (domset == 0) { if ((idn.domset.ds_trans_on | idn.domset.ds_connected | idn.domset.ds_trans_off | idn.domset.ds_relink) == 0) { PR_HITLIST("%s:%x: HITLIST %x -> 0\n", proc, domset, idn.domset.ds_hitlist); idn.domset.ds_hitlist = 0; IDN_GSTATE_TRANSITION(IDNGS_OFFLINE); } IDN_GUNLOCK(); return; } IDN_GUNLOCK(); for (d = 0; d < MAX_DOMAINS; d++) { idn_domain_t *dp; idn_fin_t ftype; if (!DOMAIN_IN_SET(domset, d)) continue; dp = &idn_domain[d]; IDN_DLOCK_EXCL(d); IDN_HISTORY_LOG(IDNH_RELINK, d, dp->dstate, idn.domset.ds_relink); ftype = fintype; if ((dp->dcpu != IDN_NIL_DCPU) && dp->dhw.dh_boardset) { /* * If domain is not in the IDNSET passed * down then we need to upgrade this to * hard-force in order to prevent possible * system failures (arbstop). This is simply * extra protection beyond that checked by * the SSP. IDNSET contains the set of boards * that have a "link" to the local domain, * including the SMD regs. */ if ((idnset & dp->dhw.dh_boardset) == 0) { PR_PROTO("%s:%d: boardset 0x%x " "NOT in IDNSET 0x%x\n", proc, d, dp->dhw.dh_boardset, idnset); if (ftype != IDNFIN_FORCE_HARD) cmn_err(CE_NOTE, "!IDN: 222: no IDN linkage " "found (b=0x%x, i=0x%x) " "upgrading unlink %s to %s", dp->dhw.dh_boardset, idnset, idnfin_str[ftype], idnfin_str[IDNFIN_FORCE_HARD]); ftype = IDNFIN_FORCE_HARD; } else { PR_PROTO("%s:%d: boardset 0x%x " "FOUND in IDNSET 0x%x\n", proc, d, dp->dhw.dh_boardset, idnset); } } idn_disconnect(d, ftype, finarg, IDNDS_SYNC_TYPE(dp)); IDN_DUNLOCK(d); } } /* * Return w/locks held. */ static int idn_connect(int domid) { idn_xdcargs_t xargs; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_connect"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(dp->dcpu != IDN_NIL_DCPU); if (dp->dstate != IDNDS_CLOSED) { if (DOMAIN_IN_SET(idn.domset.ds_trans_on | idn.domset.ds_connected, domid)) { PR_PROTO("%s:%d: already connected or " "in-progress\n", proc, domid); } else { PR_PROTO("%s:%d: current state (%s) != " "CLOSED\n", proc, domid, idnds_str[dp->dstate]); } return (-1); } ASSERT(!DOMAIN_IN_SET(idn.domset.ds_connected, domid)); ASSERT(!DOMAIN_IN_SET(idn.domset.ds_trans_off, domid)); dp->dxp = &xphase_nego; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); idn_xphase_transition(domid, NULL, xargs); return (0); } /* * Return w/locks held. */ static int idn_disconnect(int domid, idn_fin_t fintype, idn_finarg_t finarg, idn_finsync_t finsync) { int new_masterid, new_cpuid = IDN_NIL_DCPU; uint_t token; uint_t finmaster; idn_xdcargs_t xargs; idn_finopt_t finopt; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_disconnect"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (dp->dstate == IDNDS_CLOSED) { PR_PROTO("%s:%d: already CLOSED\n", proc, domid); idn_update_op(IDNOP_DISCONNECTED, DOMAINSET(domid), NULL); return (-1); } /* * Terminate any outstanding commands that were * targeted towards this domain. */ idn_terminate_cmd(domid, ECANCELED); /* * Terminate any and all retries that may have * outstanding for this domain. */ token = IDN_RETRY_TOKEN(domid, IDN_RETRY_TYPEALL); (void) idn_retry_terminate(token); /* * Stop all outstanding message timers for * this guy. */ IDN_MSGTIMER_STOP(domid, 0, 0); dp->dxp = &xphase_fin; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); if ((int)dp->dfin < (int)fintype) { /* * You can only upgrade a fin type. * We don't allow it to be downgraded * as it's too dangerous since some * state may have been blown away while * we were fin'ing at a higher level. */ IDN_FSTATE_TRANSITION(dp, fintype); } dp->dfin_sync = finsync; PR_PROTO("%s:%d: disconnect synchronously = %s\n", proc, domid, (finsync == IDNFIN_SYNC_OFF) ? "OFF" : (finsync == IDNFIN_SYNC_NO) ? "NO" : "YES"); IDN_GLOCK_SHARED(); if (DOMAIN_IN_SET(idn.domset.ds_relink, domid) && (idn.state != IDNGS_DISCONNECT)) { finopt = IDNFIN_OPT_RELINK; } else { finopt = IDNFIN_OPT_UNLINK; PR_HITLIST("%s:%d: HITLIST %x -> %x\n", proc, domid, idn.domset.ds_hitlist, idn.domset.ds_hitlist | DOMAINSET(domid)); DOMAINSET_ADD(idn.domset.ds_hitlist, domid); } CLR_XARGS(xargs); SET_XARGS_FIN_TYPE(xargs, dp->dfin); SET_XARGS_FIN_ARG(xargs, finarg); SET_XARGS_FIN_OPT(xargs, finopt); SET_XARGS_FIN_DOMSET(xargs, 0); /* unused when msg = 0 */ new_masterid = IDN_GET_NEW_MASTERID(); IDN_GUNLOCK(); if (new_masterid != IDN_NIL_DOMID) new_cpuid = idn_domain[new_masterid].dcpu; finmaster = MAKE_FIN_MASTER(new_masterid, new_cpuid); SET_XARGS_FIN_MASTER(xargs, finmaster); idn_xphase_transition(domid, NULL, xargs); return (0); } static int idn_next_xstate(idn_xstate_t o_xstate, int err, uint_t msg) { int index; procname_t proc = "idn_next_xstate"; ASSERT(((int)o_xstate >= 0) && ((int)o_xstate <= 4)); if (!msg) index = 0; else if ((msg & IDNP_MSGTYPE_MASK) == 0) index = (msg & IDNP_ACK) ? 3 : (msg & IDNP_NACK) ? 4 : -1; else index = (msg & IDNP_ACK) ? 2 : !(msg & IDNP_ACKNACK_MASK) ? 1 : -1; if (index == -1) { STRING(str); INUM2STR(msg, str); PR_PROTO("%s: (msg = 0x%x(%s))\n", proc, msg, str); return (IDNXS_NIL); } if (err == -1) { int n_xstate; /* * Caller is just interested in querying is this * is a valid message to receive in the current * xstate. A return value of IDNXS_NIL indicates * that it's not. A return value of non-IDNXS_NIL * indicates it's cool. An invalid message is * determined by both err & !err states being IDNXS_NIL. */ n_xstate = idnxs_state_table[(int)o_xstate][index][0]; if (n_xstate != IDNXS_NIL) return (n_xstate); else return (idnxs_state_table[(int)o_xstate][index][1]); } else { return (idnxs_state_table[(int)o_xstate][index][err ? 1 : 0]); } } static int idn_select_candidate(domainset_t master_set) { int d, best_id = IDN_NIL_DOMID; uint_t best_vote = 0; idn_domain_t *dp; procname_t proc = "idn_select_candidate"; ASSERT(IDN_SYNC_IS_LOCKED()); if (master_set == 0) { PR_PROTO("%s: %x -> %d\n", proc, master_set, IDN_NIL_DOMID); return (IDN_NIL_DOMID); } for (d = 0; d < MAX_DOMAINS; d++) { uint_t vote; idn_vote_t v; if (!DOMAIN_IN_SET(master_set, d)) continue; dp = &idn_domain[d]; if ((dp->domid == IDN_NIL_DOMID) || (dp->dcpu == IDN_NIL_DCPU) || ((v.ticket = dp->dvote.ticket) == 0)) continue; vote = IDNVOTE_ELECT(v); if (vote > best_vote) { best_vote = vote; best_id = d; } } PR_PROTO("%s: %x -> %d\n", proc, master_set, best_id); return (best_id); } /* * If a non-zero value is returned then GLOCK will have been dropped. * Otherwise, routine returns with all incoming locks still held. */ static int idn_select_master(int domid, int rmasterid, int rcpuid) { char *sel; int lmasterid, masterid; int do_reconfig = 0; int lindex, rindex; idn_domain_t *ldp, *rdp; uint_t rvote, lvote; idn_master_select_t select; procname_t proc = "idn_select_master"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_GLOCK_IS_EXCL()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); PR_PROTO("%s:%d: lmasterid = %d, rmasterid = %d, rcpuid = %d\n", proc, domid, IDN_GET_MASTERID(), rmasterid, rcpuid); IDN_DLOCK_EXCL(idn.localid); ldp = &idn_domain[idn.localid]; rdp = &idn_domain[domid]; /* * Clear master bits since mastership is derived from * other information (local/remote idn.masterid/idn.new_masterid) * and we don't want the vote master bit to confuse matters. */ lvote = IDNVOTE_ELECT(ldp->dvote); rvote = IDNVOTE_ELECT(rdp->dvote); lmasterid = IDN_GET_MASTERID(); lindex = (lmasterid == IDN_NIL_DOMID) ? MASTER_IS_NONE : (lmasterid == idn.localid) ? MASTER_IS_LOCAL : (lmasterid == domid) ? MASTER_IS_REMOTE : MASTER_IS_OTHER; rindex = (rmasterid == IDN_NIL_DOMID) ? MASTER_IS_NONE : (rmasterid == domid) ? MASTER_IS_REMOTE : (rmasterid == idn.localid) ? MASTER_IS_LOCAL : MASTER_IS_OTHER; select = master_select_table[lindex][rindex]; masterid = IDN_NIL_DOMID; /* * Each case is responsible for dropping DLOCK(localid) * and GLOCK if it doesn't select a master, unless a * reconfig is necessary. */ switch (select) { case MASTER_SELECT_VOTE_RCFG: sel = "VOTE_RECONFIG"; if (lvote > rvote) { /* * If the local domain is the winner then remote * domain will have to Reconfig. We'll continue * through the connection process anyway. The * remote domains will tell us to back-off while * Reconfigs, but that's okay as we'll keep retrying. */ masterid = idn.localid; } else if (lvote < rvote) { do_reconfig = 1; /* * GLOCK will get dropped once reconfig * is kicked off. */ } else { cmn_err(CE_WARN, "IDN: 206: cannot link domains " "with equal votes (L(%d),R(%d),0x%x)", idn.localid, domid, rvote); IDN_GUNLOCK(); } IDN_DUNLOCK(idn.localid); break; case MASTER_SELECT_VOTE: sel = "VOTE"; if (lvote > rvote) { masterid = idn.localid; ldp->dvote.v.master = 1; rdp->dvote.v.master = 0; } else if (lvote < rvote) { masterid = domid; ldp->dvote.v.master = 0; rdp->dvote.v.master = 1; } else { cmn_err(CE_WARN, "IDN: 206: cannot link domains " "with equal votes (L(%d),R(%d),0x%x)", idn.localid, domid, rvote); } ASSERT(IDN_GET_MASTERID() == IDN_NIL_DOMID); if (masterid != IDN_NIL_DOMID) { IDN_SET_MASTERID(masterid); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); } else { IDN_GUNLOCK(); } IDN_DUNLOCK(idn.localid); break; case MASTER_SELECT_REMOTE: sel = "REMOTE"; masterid = domid; if (IDN_GET_MASTERID() == IDN_NIL_DOMID) { IDN_SET_MASTERID(masterid); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); ldp->dvote.v.master = 0; rdp->dvote.v.master = 1; } ASSERT(IDN_GET_MASTERID() == domid); IDN_DUNLOCK(idn.localid); break; case MASTER_SELECT_LOCAL: sel = "LOCAL"; masterid = idn.localid; if (IDN_GET_MASTERID() == IDN_NIL_DOMID) { IDN_SET_MASTERID(masterid); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); ldp->dvote.v.master = 1; rdp->dvote.v.master = 0; } ASSERT(IDN_GET_MASTERID() == idn.localid); IDN_DUNLOCK(idn.localid); break; case MASTER_SELECT_CONNECT: sel = "CONNECT"; if (rmasterid == lmasterid) { /* * Local and remote have same master, * let him come onboard. */ masterid = lmasterid; IDN_DUNLOCK(idn.localid); } else { int rv; IDN_DUNLOCK(idn.localid); IDN_GUNLOCK(); IDN_DLOCK_EXCL(rmasterid); PR_PROTO("%s:%d: attempting connect w/remote " "master %d\n", proc, domid, rmasterid); rv = idn_open_domain(rmasterid, rcpuid, 0); if (rv == 0) { idn_connect(rmasterid); } else if (rv < 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to " "open-domain(%d,%d)", proc, rmasterid, rcpuid); } else { /* * Must already have a connection going. */ PR_PROTO("%s:%d: failed " "idn_open_domain(%d,%d,0) " "(rv = %d)\n", proc, domid, rmasterid, rcpuid, rv); } IDN_DUNLOCK(rmasterid); } break; case MASTER_SELECT_WAIT: sel = "WAIT"; /* * If the remote domain has the same master as the local * domain then there's no need to wait. */ if (rmasterid == lmasterid) { masterid = lmasterid; } else { IDN_GUNLOCK(); } IDN_DUNLOCK(idn.localid); break; case MASTER_SELECT_ERROR: sel = "ERROR"; /* * Hit impossible condition. */ cmn_err(CE_WARN, "IDN: 207: local/remote master-id conflict " "(%d.lmasterid = %d, %d.rmasterid = %d)", idn.localid, lmasterid, domid, rmasterid); IDN_GUNLOCK(); IDN_DUNLOCK(idn.localid); break; default: cmn_err(CE_WARN, "IDN: 208: %s: unknown case (%d)", proc, (int)select); IDN_GUNLOCK(); IDN_DUNLOCK(idn.localid); ASSERT(0); break; } if (masterid == IDN_NIL_DOMID) { PR_PROTO("%s:%d: NO MASTER SELECTED (rmstr=%d) sel=%s\n", proc, domid, rmasterid, sel); } else { PR_PROTO("%s:%d: MASTER SELECTED = %d (%s)\n", proc, domid, masterid, (masterid == idn.localid) ? "LOCAL" : (masterid == domid) ? "REMOTE" : "OTHER"); } if (do_reconfig) { domainset_t dis_set; /* * Local domain already has a master. * Need to dismantle all connections * and reestablish one with new master. */ IDN_GKSTAT_GLOBAL_EVENT(gk_reconfigs, gk_reconfig_last); PR_PROTO("%s:%d: RECONFIG new masterid = %d\n", proc, domid, domid); IDN_GSTATE_TRANSITION(IDNGS_RECONFIG); IDN_SET_NEW_MASTERID(domid); IDN_GUNLOCK(); dis_set = idn.domset.ds_trans_on | idn.domset.ds_connected; DOMAINSET_DEL(dis_set, domid); idn_unlink_domainset(dis_set, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNFIN_OPT_RELINK, BOARDSET_ALL); } return ((masterid == IDN_NIL_DOMID) ? -1 : 0); } /*ARGSUSED1*/ static void idn_retry_query(uint_t token, void *arg) { idn_retry_t rtype = IDN_RETRY_TOKEN2TYPE(token); int d, domid = IDN_RETRY_TOKEN2DOMID(token); idn_domain_t *dp = &idn_domain[domid]; idn_synccmd_t sync_cmd; domainset_t query_set, my_ready_set; procname_t proc = "idn_retry_query"; IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); switch (rtype) { case IDNRETRY_CONQ: sync_cmd = IDNSYNC_CONNECT; my_ready_set = idn.domset.ds_ready_on | idn.domset.ds_connected; my_ready_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(my_ready_set, idn.localid); break; case IDNRETRY_FINQ: sync_cmd = IDNSYNC_DISCONNECT; my_ready_set = idn.domset.ds_ready_off | ~idn.domset.ds_connected; break; default: IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } if (dp->dsync.s_cmd == sync_cmd) my_ready_set |= dp->dsync.s_set_rdy; query_set = idn_sync_register(domid, sync_cmd, 0, IDNSYNC_REG_QUERY); PR_PROTO("%s:%d: query_set = 0x%x\n", proc, domid, query_set); if (query_set == 0) { IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(query_set, d)) continue; dp = &idn_domain[d]; if (d != domid) IDN_DLOCK_EXCL(d); if ((dp->dsync.s_cmd == sync_cmd) || (!dp->dcookie_send && (rtype == IDNRETRY_CONQ))) { if (d != domid) IDN_DUNLOCK(d); continue; } IDN_SYNC_QUERY_UPDATE(domid, d); if (rtype == IDNRETRY_CONQ) idn_send_con(d, NULL, IDNCON_QUERY, my_ready_set); else idn_send_fin(d, NULL, IDNFIN_QUERY, IDNFIN_ARG_NONE, IDNFIN_OPT_NONE, my_ready_set, NIL_FIN_MASTER); if (d != domid) IDN_DUNLOCK(d); } IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); } static int idn_send_nego(int domid, idn_msgtype_t *mtp, domainset_t conset) { idn_domain_t *ldp, *dp; int d, masterid; uint_t dmask; uint_t acknack; uint_t ticket; idnneg_dset_t dset; idn_msgtype_t mt; procname_t proc = "idn_send_nego"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (mtp) { acknack = mtp->mt_mtype & IDNP_ACKNACK_MASK; mt.mt_mtype = mtp->mt_mtype; mt.mt_atype = mtp->mt_atype; mt.mt_cookie = mtp->mt_cookie; } else { acknack = 0; mt.mt_mtype = IDNP_NEGO; mt.mt_atype = 0; mt.mt_cookie = IDN_TIMER_PUBLIC_COOKIE; } IDN_GLOCK_SHARED(); dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; if ((idn.state == IDNGS_RECONFIG) || ((masterid = IDN_GET_MASTERID()) == IDN_NIL_DOMID)) { masterid = IDN_GET_NEW_MASTERID(); if ((masterid == idn.localid) || (masterid == domid)) { /* * We only send the new-master "hint" to * "other" domains. If the new-master is * ourself or we're talking to the new-master * then we need to be accurate about our * real master so that the correct master * is selected. */ masterid = IDN_NIL_DOMID; } } DOMAINSET_DEL(conset, idn.localid); DOMAINSET_DEL(conset, domid); /* * Exclude domains from conset that are on * remote domain's hitlist. He's not interested * in hearing about them. SSP is probably requesting * such domains be unlinked - will eventually get to * local domain. */ conset &= ~idn.domset.ds_hitlist; if ((masterid != IDN_NIL_DOMID) && DOMAIN_IN_SET(idn.domset.ds_hitlist, masterid)) { PR_PROTO("%s:%d: masterid(%d) on hitlist(0x%x) -> -1\n", proc, domid, masterid, idn.domset.ds_hitlist); /* * Yikes, our chosen master is on the hitlist! */ masterid = IDN_NIL_DOMID; } dmask = IDNNEG_DSET_MYMASK(); IDNNEG_DSET_INIT(dset, dmask); for (d = 0; d < MAX_DOMAINS; d++) { int cpuid; if (!DOMAIN_IN_SET(conset, d)) continue; if ((cpuid = idn_domain[d].dcpu) == IDN_NIL_DCPU) { ASSERT(d != masterid); continue; } IDNNEG_DSET_SET(dset, d, cpuid, dmask); } IDNNEG_DSET_SET_MASTER(dset, domid, masterid); ASSERT((masterid != IDN_NIL_DOMID) ? (idn_domain[masterid].dcpu != IDN_NIL_DCPU) : 1); IDN_GUNLOCK(); IDN_DLOCK_SHARED(idn.localid); ticket = IDNVOTE_BASICS(ldp->dvote); /* * We just want to send basic vote components without an * indication of mastership (master bit) since that's primarily * for local domain's usage. There is more correct master * indications in the DSET. Recall that if we were in a * Reconfig we would have transmitted the "new_masterid" * which might conflict with the local domain's vote.v.master * bit if he was originally the master prior to the Reconfig. */ PR_PROTO("%s:%d: sending nego%sto (cpu %d) " "[v=0x%x, cs=0x%x, mstr=%d]\n", proc, domid, (acknack & IDNP_ACK) ? "+ack " : (acknack & IDNP_NACK) ? "+nack " : " ", dp->dcpu, ticket, conset, masterid); IDN_MSGTIMER_START(domid, IDNP_NEGO, 0, idn_msg_waittime[IDNP_NEGO], &mt.mt_cookie); IDNXDC(domid, &mt, ticket, dset[0], dset[1], dset[2]); IDN_DUNLOCK(idn.localid); return (0); } static int idn_recv_nego(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs, ushort_t dcookie) { uint_t msg = mtp->mt_mtype; idn_msgtype_t mt; idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t nargs; procname_t proc = "idn_recv_nego"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp->mt_cookie; #ifdef DEBUG if (DOMAIN_IN_SET(idn.domset.ds_hitlist, domid)) { PR_HITLIST("%s:%d: dcpu=%d, dstate=%s, msg=%x, " "hitlist=%x\n", proc, domid, dp->dcpu, idnds_str[dp->dstate], msg, idn.domset.ds_hitlist); } #endif /* DEBUG */ if (dp->dcpu == IDN_NIL_DCPU) { int cpuid; uint_t ticket; /* * Brandnew link. Need to open a new domain entry. */ ticket = GET_XARGS_NEGO_TICKET(xargs); cpuid = dp->dcpu_last; ASSERT(VALID_CPUID(cpuid)); if (idn_open_domain(domid, cpuid, ticket) != 0) { PR_PROTO("%s:%d: FAILED to open doamin " "(ticket = 0x%x)\n", proc, domid, ticket); return (-1); } } if ((msg & IDNP_MSGTYPE_MASK) == IDNP_NEGO) { PR_PROTO("%s:%d: assigned SEND cookie 0x%x\n", proc, domid, dcookie); dp->dcookie_send = dcookie; } if ((dp->dxp == NULL) && IDNDS_IS_CLOSED(dp)) { dp->dxp = &xphase_nego; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); } else if (dp->dxp != &xphase_nego) { if (msg & IDNP_MSGTYPE_MASK) { /* * If we already have a connection to somebody * trying to initiate a connection to us, then * possibly we've awaken from a coma or he did. * In any case, dismantle current connection * and attempt to establish a new one. */ if (dp->dstate == IDNDS_CONNECTED) { DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNFIN_SYNC_YES); } else { mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; CLR_XARGS(nargs); if (DOMAIN_IN_SET(idn.domset.ds_hitlist, domid)) { SET_XARGS_NACK_TYPE(nargs, IDNNACK_EXIT); } else { int new_masterid; int new_cpuid = IDN_NIL_DCPU; SET_XARGS_NACK_TYPE(nargs, IDNNACK_RETRY); IDN_GLOCK_SHARED(); new_masterid = IDN_GET_NEW_MASTERID(); if (new_masterid == IDN_NIL_DOMID) new_masterid = IDN_GET_MASTERID(); if (new_masterid != IDN_NIL_DOMID) { idn_domain_t *mdp; mdp = &idn_domain[new_masterid]; new_cpuid = mdp->dcpu; } SET_XARGS_NACK_ARG1(nargs, new_masterid); SET_XARGS_NACK_ARG2(nargs, new_cpuid); IDN_GUNLOCK(); } idn_send_acknack(domid, &mt, nargs); } } return (0); } idn_xphase_transition(domid, mtp, xargs); return (0); } /*ARGSUSED1*/ static void idn_retry_nego(uint_t token, void *arg) { int domid = IDN_RETRY_TOKEN2DOMID(token); int new_masterid; idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t xargs; procname_t proc = "idn_retry_nego"; ASSERT(IDN_RETRY_TOKEN2TYPE(token) == IDNRETRY_NEGO); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); if (dp->dxp != &xphase_nego) { STRING(str); #ifdef DEBUG if (dp->dxp) { INUM2STR(dp->dxp->xt_msgtype, str); } #endif /* DEBUG */ PR_PROTO("%s:%d: dxp(%s) != NEGO...bailing...\n", proc, domid, dp->dxp ? str : "NULL"); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } if (dp->dxstate != IDNXS_PEND) { PR_PROTO("%s:%d: xstate(%s) != %s...bailing\n", proc, domid, idnxs_str[dp->dxstate], idnxs_str[IDNXS_PEND]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } IDN_GLOCK_SHARED(); if (idn.state == IDNGS_RECONFIG) { /* * Have to try again later after * reconfig has completed. */ PR_PROTO("%s:%d: reconfig in-progress...try later\n", proc, domid); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[IDNP_NEGO]); IDN_GUNLOCK(); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } new_masterid = IDN_GET_NEW_MASTERID(); if ((idn.state == IDNGS_CONNECT) && (new_masterid != IDN_NIL_DOMID) && (domid != new_masterid) && (idn.localid != new_masterid)) { /* * We have a new master pending and this * guy isn't it. Wait until the local domain * has a chance to connect with the new * master before going forward with this * guy. */ PR_PROTO("%s:%d: waiting for connect to new master %d\n", proc, domid, IDN_GET_NEW_MASTERID()); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[IDNP_NEGO]); IDN_GUNLOCK(); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } IDN_GUNLOCK(); idn_xphase_transition(domid, NULL, xargs); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); } static int idn_check_nego(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { int d, new_masterid, masterid; int cpuid, m_cpuid = -1; uint_t dmask; uint_t msg = mtp ? mtp->mt_mtype : 0; idn_domain_t *dp, *ldp; domainset_t con_set, pending_set; idnneg_dset_t dset; procname_t proc = "idn_check_nego"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; if (msg & IDNP_NACK) { if (GET_XARGS_NACK_TYPE(xargs) == IDNNACK_EXIT) { PR_HITLIST("%s:%d(%s): (msg=%x) EXIT received, " "adding to hitlist %x -> %x\n", proc, domid, idnds_str[dp->dstate], msg, idn.domset.ds_hitlist, idn.domset.ds_hitlist | DOMAINSET(domid)); DOMAINSET_ADD(idn.domset.ds_hitlist, domid); return (-1); } else { return (0); } } if (DOMAIN_IN_SET(idn.domset.ds_hitlist, domid)) { PR_HITLIST("%s:%d(%s): (msg=%x) domain in hitlist (%x) - " "exiting phase\n", proc, domid, idnds_str[dp->dstate], msg, idn.domset.ds_hitlist); return (-1); } if ((dp->dstate == IDNDS_NEGO_PEND) && (msg & IDNP_MSGTYPE_MASK) && (msg & IDNP_ACK)) /* nego+ack */ return (1); dmask = (uint_t)-1; IDN_GLOCK_EXCL(); if (idn.state == IDNGS_DISCONNECT) { PR_PROTO("%s:%d: DISCONNECT in-progress >>> EXIT\n", proc, domid); IDN_GUNLOCK(); return (-1); } else if (idn.state == IDNGS_OFFLINE) { IDN_GSTATE_TRANSITION(IDNGS_CONNECT); IDN_PREP_HWINIT(); IDN_DLOCK_EXCL(idn.localid); ldp->dvote.v.connected = 0; IDN_DUNLOCK(idn.localid); } if (!DOMAIN_IN_SET(idn.domset.ds_trans_on, domid)) { DOMAINSET_ADD(idn.domset.ds_trans_on, domid); IDN_HISTORY_LOG(IDNH_NEGO, domid, idn.domset.ds_trans_on, idn.domset.ds_connected); } switch (idn.state) { case IDNGS_RECONFIG: PR_PROTO("%s:%d: RECONFIG in-progress >>> RETRY\n", proc, domid); IDN_GUNLOCK(); return (1); case IDNGS_CONNECT: new_masterid = IDN_GET_NEW_MASTERID(); if ((new_masterid != IDN_NIL_DOMID) && (domid != new_masterid) && (idn.localid != new_masterid)) { PR_PROTO("%s:%d: waiting for connect to " "new master %d\n", proc, domid, IDN_GET_NEW_MASTERID()); IDN_GUNLOCK(); return (1); } break; default: break; } ASSERT((idn.state == IDNGS_CONNECT) || (idn.state == IDNGS_ONLINE)); con_set = 0; if (msg) { idn_domain_t *mdp; idn_vote_t vote; vote.ticket = GET_XARGS_NEGO_TICKET(xargs); /* * Sender should note have set master bit, * but just in case clear it so local domain * doesn't get confused. */ vote.v.master = 0; dp->dvote.ticket = vote.ticket; GET_XARGS_NEGO_DSET(xargs, dset); /*LINTED*/ IDNNEG_DSET_GET_MASK(dset, domid, dmask); IDNNEG_DSET_GET_MASTER(dset, new_masterid); if (new_masterid == IDNNEG_NO_MASTER) { new_masterid = IDN_NIL_DOMID; } else { /* * Remote domain has a master. Find * his cpuid in the dset. We may need * it to initiate a connection. */ if (new_masterid == domid) { m_cpuid = dp->dcpu; } else { IDNNEG_DSET_GET(dset, new_masterid, m_cpuid, dmask); if (m_cpuid == -1) { /* * Something is bogus if remote domain * is reporting a valid masterid, but * doesn't have the cpuid for it. */ cmn_err(CE_WARN, "IDN: 209: remote domain (ID " "%d, CPU %d) reporting master " "(ID %d) without CPU ID", domid, dp->dcpu, new_masterid); DOMAINSET_ADD(idn.domset.ds_hitlist, domid); IDN_GUNLOCK(); return (-1); } } } for (d = 0; d < MAX_DOMAINS; d++) { if ((d == idn.localid) || (d == domid)) continue; IDNNEG_DSET_GET(dset, d, cpuid, dmask); if (cpuid != -1) { DOMAINSET_ADD(con_set, d); } } #ifdef DEBUG if (idn.domset.ds_hitlist) { PR_HITLIST("%s:%d: con_set %x -> %x (hitlist = %x)\n", proc, domid, con_set, con_set & ~idn.domset.ds_hitlist, idn.domset.ds_hitlist); } #endif /* DEBUG */ con_set &= ~idn.domset.ds_hitlist; ASSERT(!DOMAIN_IN_SET(con_set, idn.localid)); ASSERT(!DOMAIN_IN_SET(con_set, domid)); if ((new_masterid != IDN_NIL_DOMID) && DOMAIN_IN_SET(idn.domset.ds_hitlist, new_masterid)) { PR_HITLIST("%s:%d: new_mstr %d -> -1 (hitlist = %x)\n", proc, domid, new_masterid, idn.domset.ds_hitlist); IDN_GUNLOCK(); return (1); } if (idn_select_master(domid, new_masterid, m_cpuid) < 0) { /* * Returns w/GLOCK dropped if error. */ return (1); } masterid = IDN_GET_MASTERID(); ASSERT(masterid != IDN_NIL_DOMID); if (idn.state == IDNGS_CONNECT) { /* * This is the initial connection for * the local domain. */ IDN_DLOCK_EXCL(idn.localid); if (masterid == idn.localid) { if (idn_master_init() < 0) { cmn_err(CE_WARN, "IDN: 210: failed to init " "MASTER context"); ldp->dvote.v.master = 0; IDN_DUNLOCK(idn.localid); IDN_GSTATE_TRANSITION(IDNGS_DISCONNECT); IDN_SET_MASTERID(IDN_NIL_DOMID); IDN_GUNLOCK(); return (-1); } DSLAB_LOCK_EXCL(idn.localid); ldp->dslab_state = DSLAB_STATE_LOCAL; DSLAB_UNLOCK(idn.localid); ldp->dvote.v.connected = 1; } else { /* * Either the remote domain is the * master or its a new slave trying * to connect to us. We can't allow * further progress until we've * sync'd up with the master. */ if (masterid != domid) { IDN_DUNLOCK(idn.localid); IDN_GUNLOCK(); return (1); } DSLAB_LOCK_EXCL(idn.localid); ldp->dslab_state = DSLAB_STATE_REMOTE; DSLAB_UNLOCK(idn.localid); } IDN_DUNLOCK(idn.localid); /* * We've sync'd up with the new master. */ IDN_GSTATE_TRANSITION(IDNGS_ONLINE); } mdp = &idn_domain[masterid]; if ((masterid != domid) && !IDNDS_CONFIG_DONE(mdp)) { /* * We can't progress any further with * other domains until we've exchanged all * the necessary CFG info with the master, * i.e. until we have a mailbox area from * which we can allocate mailboxes to * other domains. */ PR_PROTO("%s:%d: still exchanging CFG " "w/master(%d)\n", proc, domid, masterid); IDN_GUNLOCK(); return (1); } DSLAB_LOCK_EXCL(domid); dp->dslab_state = ldp->dslab_state; DSLAB_UNLOCK(domid); if (idn.state != IDNGS_ONLINE) { IDN_GSTATE_TRANSITION(IDNGS_ONLINE); } } IDN_GUNLOCK(); pending_set = con_set; pending_set &= ~(idn.domset.ds_trans_on | idn.domset.ds_connected); idn.domset.ds_trans_on |= pending_set; con_set |= idn.domset.ds_trans_on | idn.domset.ds_connected; con_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(con_set, idn.localid); if (dp->dsync.s_cmd != IDNSYNC_CONNECT) { idn_sync_exit(domid, IDNSYNC_DISCONNECT); idn_sync_enter(domid, IDNSYNC_CONNECT, con_set, DOMAINSET(idn.localid), idn_xstate_transfunc, (void *)IDNP_CON); } /* * Get this domain registered as an expected domain on * the remaining domains in the CONNECT synchronization. */ (void) idn_sync_register(domid, IDNSYNC_CONNECT, 0, IDNSYNC_REG_NEW); /* * Note that if (msg == 0), i.e. then there will be * no dset and also pending_set will be 0. * So, the following loop will never attempt to * look at the dset unless (msg != 0), implying * that we've been through the initial code above * and have initialized dmask. */ ASSERT(pending_set ? (dmask != (uint_t)-1) : 1); for (d = 0; d < MAX_DOMAINS; d++) { int rv; if (!DOMAIN_IN_SET(pending_set, d)) continue; ASSERT((d != idn.localid) && (d != domid)); dp = &idn_domain[d]; IDNNEG_DSET_GET(dset, d, cpuid, dmask); if (cpuid == -1) { PR_PROTO("%s:%d: failed to get cpuid from dset " "for domain %d (pset = 0x%x)\n", proc, domid, d, pending_set); DOMAINSET_DEL(idn.domset.ds_trans_on, d); continue; } IDN_DLOCK_EXCL(d); if ((rv = idn_open_domain(d, cpuid, 0)) != 0) { PR_PROTO("%s:%d: failed " "idn_open_domain(%d,%d,0) (rv = %d)\n", proc, domid, d, cpuid, rv); if (rv < 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to " "open-domain(%d,%d)", proc, d, cpuid); DOMAINSET_DEL(idn.domset.ds_trans_on, d); } else if (DOMAIN_IN_SET(idn.domset.ds_trans_off, d)) { /* * We've requested to connect to a domain * from which we're disconnecting. We * better mark this guy for relinking. */ DOMAINSET_ADD(idn.domset.ds_relink, d); IDN_HISTORY_LOG(IDNH_RELINK, d, dp->dstate, idn.domset.ds_relink); } IDN_DUNLOCK(d); continue; } idn_connect(d); IDN_DUNLOCK(d); } return (0); } /*ARGSUSED*/ static void idn_action_nego_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; idn_msgtype_t mt; domainset_t con_set; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); con_set = idn.domset.ds_trans_on | idn.domset.ds_connected; con_set &= ~idn.domset.ds_trans_off; if (!msg) { idn_send_nego(domid, NULL, con_set); } else { mt.mt_mtype = IDNP_NEGO | IDNP_ACK; mt.mt_atype = 0; mt.mt_cookie = mtp->mt_cookie; idn_send_nego(domid, &mt, con_set); } } /*ARGSUSED*/ static void idn_error_nego(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { int new_masterid, new_cpuid; int retry = 1; uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t token; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) { idn_nack_t nack; nack = GET_XARGS_NACK_TYPE(xargs); switch (nack) { case IDNNACK_RETRY: new_masterid = (int)GET_XARGS_NACK_ARG1(xargs); new_cpuid = (int)GET_XARGS_NACK_ARG2(xargs); break; case IDNNACK_EXIT: retry = 0; /*FALLTHROUGH*/ default: new_masterid = IDN_NIL_DOMID; new_cpuid = IDN_NIL_DCPU; break; } idn_nego_cleanup_check(domid, new_masterid, new_cpuid); } if (msg & IDNP_MSGTYPE_MASK) { idn_msgtype_t mt; idn_xdcargs_t nargs; mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; mt.mt_cookie = mtp->mt_cookie; CLR_XARGS(nargs); SET_XARGS_NACK_TYPE(nargs, IDNNACK_RETRY); IDN_GLOCK_SHARED(); new_masterid = IDN_GET_NEW_MASTERID(); if (new_masterid == IDN_NIL_DOMID) new_masterid = IDN_GET_MASTERID(); if (new_masterid != IDN_NIL_DOMID) new_cpuid = idn_domain[new_masterid].dcpu; else new_cpuid = IDN_NIL_DCPU; SET_XARGS_NACK_ARG1(nargs, new_masterid); SET_XARGS_NACK_ARG2(nargs, new_cpuid); IDN_GUNLOCK(); idn_send_acknack(domid, &mt, nargs); } if (retry) { token = IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[(int)IDNRETRY_NEGO]); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_RESET_COOKIES(domid); idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNDS_SYNC_TYPE(&idn_domain[domid])); } } /*ARGSUSED*/ static void idn_action_nego_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; domainset_t conset; idn_msgtype_t mt; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp ? mtp->mt_cookie : 0; conset = idn.domset.ds_trans_on | idn.domset.ds_connected; conset &= ~idn.domset.ds_trans_off; if ((msg & IDNP_ACKNACK_MASK) == 0) { /* * nego */ mt.mt_mtype = IDNP_NEGO | IDNP_ACK; mt.mt_atype = 0; idn_send_nego(domid, &mt, conset); } else if (msg & IDNP_MSGTYPE_MASK) { int d; idn_xdcargs_t nargs; idnneg_dset_t dset; uint_t dmask; idn_vote_t vote; mt.mt_mtype = IDNP_ACK; mt.mt_atype = msg; DOMAINSET_DEL(conset, idn.localid); DOMAINSET_DEL(conset, domid); dmask = IDNNEG_DSET_MYMASK(); IDNNEG_DSET_INIT(dset, dmask); for (d = 0; d < MAX_DOMAINS; d++) { int cpuid; if (!DOMAIN_IN_SET(conset, d)) continue; if ((cpuid = idn_domain[d].dcpu) == IDN_NIL_DCPU) continue; IDNNEG_DSET_SET(dset, d, cpuid, dmask); } IDNNEG_DSET_SET_MASTER(dset, domid, IDN_GET_MASTERID()); ASSERT((IDN_GET_MASTERID() != IDN_NIL_DOMID) ? (idn_domain[IDN_GET_MASTERID()].dcpu != IDN_NIL_DCPU) : 1); vote.ticket = idn_domain[idn.localid].dvote.ticket; vote.v.master = 0; CLR_XARGS(nargs); SET_XARGS_NEGO_TICKET(nargs, vote.ticket); SET_XARGS_NEGO_DSET(nargs, dset); /* * nego+ack */ idn_send_acknack(domid, &mt, nargs); } else { uint_t token; int new_masterid, new_cpuid; int retry = 1; idn_nack_t nack; /* * nack - retry * * It's possible if we've made it this far that * we may have already chosen a master and this * dude might be it! If it is we need to clean up. */ nack = GET_XARGS_NACK_TYPE(xargs); switch (nack) { case IDNNACK_RETRY: new_masterid = (int)GET_XARGS_NACK_ARG1(xargs); new_cpuid = (int)GET_XARGS_NACK_ARG2(xargs); break; case IDNNACK_EXIT: retry = 0; /*FALLTHROUGH*/ default: new_masterid = IDN_NIL_DOMID; new_cpuid = IDN_NIL_DCPU; break; } idn_nego_cleanup_check(domid, new_masterid, new_cpuid); if (retry) { token = IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[(int)IDNRETRY_NEGO]); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_RESET_COOKIES(domid); idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNDS_SYNC_TYPE(&idn_domain[domid])); } } } /*ARGSUSED*/ static void idn_action_nego_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) { uint_t token; int new_masterid, new_cpuid; int retry = 1; idn_nack_t nack; /* * nack - retry. * * At this stage of receiving a nack we need to * check whether we need to start over again with * selecting a new master. */ nack = GET_XARGS_NACK_TYPE(xargs); switch (nack) { case IDNNACK_RETRY: new_masterid = (int)GET_XARGS_NACK_ARG1(xargs); new_cpuid = (int)GET_XARGS_NACK_ARG2(xargs); break; case IDNNACK_EXIT: retry = 0; /*FALLTHROUGH*/ default: new_masterid = IDN_NIL_DOMID; new_cpuid = IDN_NIL_DCPU; break; } idn_nego_cleanup_check(domid, new_masterid, new_cpuid); if (retry) { token = IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[(int)IDNRETRY_NEGO]); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_RESET_COOKIES(domid); idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNDS_SYNC_TYPE(&idn_domain[domid])); } } } static void idn_final_nego(int domid) { idn_domain_t *dp = &idn_domain[domid]; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); (void) idn_retry_terminate(IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO)); ASSERT(dp->dstate == IDNDS_CONFIG); dp->dxp = NULL; IDN_XSTATE_TRANSITION(dp, IDNXS_NIL); idn_send_config(domid, 1); } /* */ /*ARGSUSED1*/ static void idn_exit_nego(int domid, uint_t msgtype) { idn_domain_t *dp; idn_fin_t fintype; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; fintype = msgtype ? IDNFIN_NORMAL : IDNFIN_FORCE_HARD; (void) idn_retry_terminate(IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO)); ASSERT(!DOMAIN_IN_SET(idn.domset.ds_connected, domid)); ASSERT(!DOMAIN_IN_SET(idn.domset.ds_ready_on, domid)); ASSERT(dp->dxp == &xphase_nego); idn_nego_cleanup_check(domid, IDN_NIL_DOMID, IDN_NIL_DCPU); IDN_GLOCK_SHARED(); if ((idn.state != IDNGS_DISCONNECT) && !DOMAIN_IN_SET(idn.domset.ds_hitlist, domid)) { DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); } else { idn_update_op(IDNOP_ERROR, DOMAINSET(domid), NULL); DOMAINSET_DEL(idn.domset.ds_relink, domid); } IDN_GUNLOCK(); /* * Reset send cookie to 0 so that receiver does not validate * cookie. This is necessary since at this early stage it's * possible we may not have exchanged appropriate cookies. */ IDN_RESET_COOKIES(domid); idn_disconnect(domid, fintype, IDNFIN_ARG_NONE, IDNDS_SYNC_TYPE(dp)); } static void idn_nego_cleanup_check(int domid, int new_masterid, int new_cpuid) { idn_domain_t *ldp, *dp; procname_t proc = "idn_nego_cleanup_check"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; IDN_GLOCK_EXCL(); if (((idn.state == IDNGS_ONLINE) && !idn.domset.ds_connected) || (idn.state == IDNGS_CONNECT)) { domainset_t trans_on; int masterid; int retry_domid = IDN_NIL_DOMID; int rv; IDN_DLOCK_EXCL(idn.localid); masterid = (idn.state == IDNGS_ONLINE) ? IDN_GET_MASTERID() : IDN_GET_NEW_MASTERID(); trans_on = idn.domset.ds_trans_on; DOMAINSET_DEL(trans_on, domid); if (trans_on == 0) { int d; domainset_t relink = idn.domset.ds_relink; /* * This was the only guy we were trying * to connect with. */ ASSERT((idn.state == IDNGS_ONLINE) ? ((idn.localid == masterid) || (domid == masterid)) : 1); if (idn.localid == masterid) idn_master_deinit(); ldp->dvote.v.connected = 0; ldp->dvote.v.master = 0; dp->dvote.v.master = 0; IDN_SET_MASTERID(IDN_NIL_DOMID); IDN_SET_NEW_MASTERID(new_masterid); IDN_GSTATE_TRANSITION(IDNGS_CONNECT); IDN_PREP_HWINIT(); IDN_DUNLOCK(idn.localid); IDN_GUNLOCK(); /* * If there's a new master available then * just try and relink with him unless * it's ourself. */ if ((new_masterid != IDN_NIL_DOMID) && (new_masterid != idn.localid) && (new_masterid != domid)) { IDN_DLOCK_EXCL(new_masterid); rv = idn_open_domain(new_masterid, new_cpuid, 0); if (rv < 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to " "open-domain(%d,%d)", proc, new_masterid, new_cpuid); IDN_GLOCK_EXCL(); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); IDN_GUNLOCK(); } else { relink = DOMAINSET(new_masterid); } IDN_DUNLOCK(new_masterid); } DOMAINSET_DEL(relink, domid); if (relink) for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(relink, d)) continue; retry_domid = d; break; } } else if (domid == masterid) { /* * There are other domains we were trying * to connect to. As long as the chosen * master was somebody other then this * domain that nack'd us, life is cool, but * if it was this remote domain we'll need * to start over. */ IDN_DUNLOCK(idn.localid); dp->dvote.v.master = 0; IDN_SET_MASTERID(IDN_NIL_DOMID); IDN_SET_NEW_MASTERID(new_masterid); if (idn.state == IDNGS_ONLINE) { IDN_GKSTAT_GLOBAL_EVENT(gk_reconfigs, gk_reconfig_last); IDN_GSTATE_TRANSITION(IDNGS_RECONFIG); IDN_GUNLOCK(); idn_unlink_domainset(trans_on, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNFIN_OPT_RELINK, BOARDSET_ALL); } else if ((new_masterid != IDN_NIL_DOMID) && (new_masterid != idn.localid) && (new_masterid != domid) && !DOMAIN_IN_SET(trans_on, new_masterid)) { IDN_GUNLOCK(); IDN_DLOCK_EXCL(new_masterid); rv = idn_open_domain(new_masterid, new_cpuid, 0); IDN_GLOCK_EXCL(); IDN_DUNLOCK(new_masterid); if (rv < 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to " "open-domain(%d,%d)", proc, new_masterid, new_cpuid); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); new_masterid = IDN_NIL_DOMID; } else { retry_domid = new_masterid; } IDN_GUNLOCK(); } else { IDN_GUNLOCK(); } } else { IDN_DUNLOCK(idn.localid); IDN_GUNLOCK(); } if (retry_domid != IDN_NIL_DOMID) { uint_t token; idn_domain_t *rdp = &idn_domain[retry_domid]; IDN_DLOCK_EXCL(retry_domid); rdp->dxp = &xphase_nego; IDN_XSTATE_TRANSITION(rdp, IDNXS_PEND); IDN_DUNLOCK(retry_domid); token = IDN_RETRY_TOKEN(retry_domid, IDNRETRY_NEGO); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[(int)IDNRETRY_NEGO]); } } else { IDN_GUNLOCK(); } } static int idn_send_con(int domid, idn_msgtype_t *mtp, idn_con_t contype, domainset_t conset) { idn_msgtype_t mt; uint_t acknack; procname_t proc = "idn_send_con"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (mtp) { acknack = mtp->mt_mtype & IDNP_ACKNACK_MASK; mt.mt_mtype = mtp->mt_mtype; mt.mt_atype = mtp->mt_atype; mt.mt_cookie = mtp->mt_cookie; } else { acknack = 0; mt.mt_mtype = IDNP_CON; mt.mt_atype = 0; /* * For simple CON queries we want a unique * timer assigned. For others, they * effectively share one. */ if (contype == IDNCON_QUERY) mt.mt_cookie = 0; else mt.mt_cookie = IDN_TIMER_PUBLIC_COOKIE; } ASSERT((contype == IDNCON_QUERY) ? idn_domain[domid].dcookie_send : 1); PR_PROTO("%s:%d: sending con%sto (cpu %d) [ct=%s, cs=0x%x]\n", proc, domid, (acknack & IDNP_ACK) ? "+ack " : (acknack & IDNP_NACK) ? "+nack " : " ", idn_domain[domid].dcpu, idncon_str[contype], conset); IDN_MSGTIMER_START(domid, IDNP_CON, (ushort_t)contype, idn_msg_waittime[IDNP_CON], &mt.mt_cookie); IDNXDC(domid, &mt, (uint_t)contype, (uint_t)conset, 0, 0); return (0); } /* * Must leave w/DLOCK dropped and SYNC_LOCK held. */ static int idn_recv_con(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t msgarg = mtp ? mtp->mt_atype : 0; idn_con_t contype; domainset_t my_ready_set, ready_set; idn_msgtype_t mt; idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t aargs; procname_t proc = "idn_recv_con"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp ? mtp->mt_cookie : 0; contype = GET_XARGS_CON_TYPE(xargs); ready_set = GET_XARGS_CON_DOMSET(xargs); CLR_XARGS(aargs); if (!(msg & IDNP_NACK) && (contype == IDNCON_QUERY)) { domainset_t query_set; query_set = idn_sync_register(domid, IDNSYNC_CONNECT, ready_set, IDNSYNC_REG_REG); my_ready_set = idn.domset.ds_connected | idn.domset.ds_ready_on; my_ready_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(my_ready_set, idn.localid); if (msg & IDNP_MSGTYPE_MASK) { mt.mt_mtype = IDNP_ACK; mt.mt_atype = IDNP_CON; SET_XARGS_CON_TYPE(aargs, contype); SET_XARGS_CON_DOMSET(aargs, my_ready_set); idn_send_acknack(domid, &mt, aargs); } if (query_set) { uint_t token; token = IDN_RETRY_TOKEN(domid, IDNRETRY_CONQ); idn_retry_submit(idn_retry_query, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CONQ]); } return (0); } if (dp->dxp == NULL) { STRING(mstr); STRING(lstr); /* * Must have received an inappropriate error * message as we should already be registered * by the time we reach here. */ INUM2STR(msg, mstr); INUM2STR(msgarg, lstr); PR_PROTO("%s:%d: ERROR: NOT YET REGISTERED (%s/%s)\n", proc, domid, mstr, lstr); if (msg & IDNP_MSGTYPE_MASK) { mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; SET_XARGS_NACK_TYPE(aargs, IDNNACK_RETRY); idn_send_acknack(domid, &mt, aargs); } return (-1); } idn_xphase_transition(domid, mtp, xargs); return (0); } /*ARGSUSED1*/ static void idn_retry_con(uint_t token, void *arg) { int domid = IDN_RETRY_TOKEN2DOMID(token); idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t xargs; procname_t proc = "idn_retry_con"; ASSERT(IDN_RETRY_TOKEN2TYPE(token) == IDNRETRY_CON); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); if (dp->dxp != &xphase_con) { STRING(str); #ifdef DEBUG if (dp->dxp) { INUM2STR(dp->dxp->xt_msgtype, str); } #endif /* DEBUG */ PR_PROTO("%s:%d: dxp(%s) != CON...bailing...\n", proc, domid, dp->dxp ? str : "NULL"); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } if ((dp->dsync.s_cmd != IDNSYNC_CONNECT) || (dp->dxstate != IDNXS_PEND)) { PR_PROTO("%s:%d: cmd (%s) and/or xstate (%s) not " "expected (%s/%s)\n", proc, domid, idnsync_str[dp->dsync.s_cmd], idnxs_str[dp->dxstate], idnsync_str[IDNSYNC_CONNECT], idnxs_str[IDNXS_PEND]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } idn_xphase_transition(domid, NULL, xargs); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); } static int idn_check_con(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { int ready; uint_t msg = mtp ? mtp->mt_mtype : 0; idn_domain_t *dp = &idn_domain[domid]; domainset_t ready_set, my_ready_set, query_set; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) return (0); if ((dp->dstate == IDNDS_CON_PEND) && (msg & IDNP_MSGTYPE_MASK) && (msg & IDNP_ACK)) /* con+ack */ return (1); if (msg == 0) { ready_set = idn.domset.ds_connected & ~idn.domset.ds_trans_off; } else { ready_set = GET_XARGS_CON_DOMSET(xargs); DOMAINSET_ADD(idn.domset.ds_ready_on, domid); } DOMAINSET_ADD(ready_set, idn.localid); query_set = idn_sync_register(domid, IDNSYNC_CONNECT, ready_set, IDNSYNC_REG_REG); /* * No need to query this domain as he's already * in the CON sequence. */ DOMAINSET_DEL(query_set, domid); ready = (dp->dsync.s_set_exp == dp->dsync.s_set_rdy) ? 1 : 0; if (ready) { DOMAINSET_DEL(idn.domset.ds_ready_on, domid); DOMAINSET_ADD(idn.domset.ds_connected, domid); } if (query_set) { int d; my_ready_set = idn.domset.ds_ready_on | idn.domset.ds_connected; my_ready_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(my_ready_set, idn.localid); for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(query_set, d)) continue; dp = &idn_domain[d]; IDN_DLOCK_EXCL(d); if ((dp->dsync.s_cmd == IDNSYNC_CONNECT) || !dp->dcookie_send) { IDN_DUNLOCK(d); continue; } IDN_SYNC_QUERY_UPDATE(domid, d); idn_send_con(d, NULL, IDNCON_QUERY, my_ready_set); IDN_DUNLOCK(d); } } return (!msg ? 0 : (ready ? 0 : 1)); } /*ARGSUSED2*/ static void idn_error_con(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t token; uint_t msg = mtp ? mtp->mt_mtype : 0; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_MSGTYPE_MASK) { idn_msgtype_t mt; idn_xdcargs_t nargs; mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; mt.mt_cookie = mtp->mt_cookie; CLR_XARGS(nargs); SET_XARGS_NACK_TYPE(nargs, IDNNACK_RETRY); idn_send_acknack(domid, &mt, nargs); } token = IDN_RETRY_TOKEN(domid, IDNRETRY_CON); idn_retry_submit(idn_retry_con, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CON]); } /*ARGSUSED*/ static void idn_action_con_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; idn_domain_t *dp = &idn_domain[domid]; idn_msgtype_t mt; domainset_t my_ready_set; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); my_ready_set = dp->dsync.s_set_rdy | idn.domset.ds_ready_on | idn.domset.ds_connected; my_ready_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(my_ready_set, idn.localid); if (!msg) { idn_send_con(domid, NULL, IDNCON_NORMAL, my_ready_set); } else { mt.mt_mtype = IDNP_CON | IDNP_ACK; mt.mt_atype = 0; mt.mt_cookie = mtp->mt_cookie; idn_send_con(domid, &mt, IDNCON_NORMAL, my_ready_set); } } static void idn_action_con_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; idn_domain_t *dp = &idn_domain[domid]; idn_con_t contype; domainset_t my_ready_set; idn_msgtype_t mt; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp ? mtp->mt_cookie : 0; my_ready_set = dp->dsync.s_set_rdy | idn.domset.ds_ready_on | idn.domset.ds_connected; my_ready_set &= ~idn.domset.ds_trans_off; DOMAINSET_ADD(my_ready_set, idn.localid); contype = GET_XARGS_CON_TYPE(xargs); if ((msg & IDNP_ACKNACK_MASK) == 0) { /* * con */ mt.mt_mtype = IDNP_CON | IDNP_ACK; mt.mt_atype = 0; idn_send_con(domid, &mt, contype, my_ready_set); } else if (msg & IDNP_MSGTYPE_MASK) { idn_xdcargs_t cargs; mt.mt_mtype = IDNP_ACK; mt.mt_atype = msg; CLR_XARGS(cargs); SET_XARGS_CON_TYPE(cargs, contype); SET_XARGS_CON_DOMSET(cargs, my_ready_set); /* * con+ack */ idn_send_acknack(domid, &mt, cargs); } else { uint_t token; /* * nack - retry */ token = IDN_RETRY_TOKEN(domid, IDNRETRY_CON); idn_retry_submit(idn_retry_con, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CON]); } } /*ARGSUSED*/ static void idn_action_con_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) { uint_t token; /* * nack - retry */ token = IDN_RETRY_TOKEN(domid, IDNRETRY_CON); idn_retry_submit(idn_retry_con, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CON]); } } static void idn_final_con(int domid) { uint_t targ; uint_t token = IDN_RETRY_TOKEN(domid, IDNRETRY_CON); idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_final_con"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); (void) idn_retry_terminate(token); dp->dxp = NULL; IDN_XSTATE_TRANSITION(dp, IDNXS_NIL); idn_sync_exit(domid, IDNSYNC_CONNECT); CHECKPOINT_OPENED(IDNSB_CHKPT_LINK, dp->dhw.dh_boardset, 1); DOMAINSET_DEL(idn.domset.ds_trans_on, domid); DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_FSTATE_TRANSITION(dp, IDNFIN_OFF); PR_PROTO("%s:%d: CONNECTED\n", proc, domid); if (idn.domset.ds_trans_on == 0) { if ((idn.domset.ds_trans_off | idn.domset.ds_relink) == 0) { PR_HITLIST("%s:%d: HITLIST %x -> 0\n", proc, domid, idn.domset.ds_hitlist); idn.domset.ds_hitlist = 0; } PR_PROTO("%s:%d: ALL CONNECTED ************ " "(0x%x + 0x%x) = 0x%x\n", proc, domid, DOMAINSET(idn.localid), idn.domset.ds_connected, DOMAINSET(idn.localid) | idn.domset.ds_connected); } else { PR_PROTO("%s:%d: >>> ds_trans_on = 0x%x, ds_ready_on = 0x%x\n", proc, domid, idn.domset.ds_trans_on, idn.domset.ds_ready_on); } if (idn_verify_config_mbox(domid)) { idnsb_error_t idnerr; /* * Mailbox is not cool. Need to disconnect. */ INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EPROTO); SET_IDNKERR_IDNERR(&idnerr, IDNKERR_SMR_CORRUPTED); SET_IDNKERR_PARAM0(&idnerr, domid); idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); /* * We cannot disconnect from an individual domain * unless all domains are attempting to disconnect * from him also, especially now since we touched * the SMR and now we have a potential cache conflicts * with the other domains with respect to this * domain. Disconnect attempt will effectively * shutdown connection with respective domain * which is the effect we really want anyway. */ idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_SMRBAD, IDNFIN_SYNC_YES); return; } if (lock_try(&idn.first_swlink)) { /* * This is our first connection. Need to * kick some stuff into gear. */ idndl_dlpi_init(); (void) idn_activate_channel(CHANSET_ALL, IDNCHAN_ONLINE); targ = 0xf0; } else { targ = 0; } idn_mainmbox_activate(domid); idn_update_op(IDNOP_CONNECTED, DOMAINSET(domid), NULL); IDN_GKSTAT_GLOBAL_EVENT(gk_links, gk_link_last); membar_stst_ldst(); IDN_DSTATE_TRANSITION(dp, IDNDS_CONNECTED); /* * Need to kick off initial commands in background. * We do not want to do them within the context of * a protocol server because they may sleep and thus * cause the protocol server to incur a soft-deadlock, * i.e. he's sleeping waiting in the slab-waiting area * for a response that will arrive on his protojob * queue, but which he obviously can't process since * he's not waiting on his protojob queue. */ targ |= domid & 0x0f; (void) timeout(idn_link_established, (void *)(uintptr_t)targ, 50); cmn_err(CE_NOTE, "!IDN: 200: link (domain %d, CPU %d) connected", dp->domid, dp->dcpu); } static void idn_exit_con(int domid, uint_t msgtype) { idn_domain_t *dp = &idn_domain[domid]; idn_fin_t fintype; procname_t proc = "idn_exit_con"; STRING(str); ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); INUM2STR(msgtype, str); PR_PROTO("%s:%d: msgtype = 0x%x(%s)\n", proc, domid, msgtype, str); fintype = msgtype ? IDNFIN_NORMAL : IDNFIN_FORCE_HARD; IDN_GLOCK_SHARED(); if (idn.state != IDNGS_DISCONNECT) { DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); } IDN_GUNLOCK(); idn_disconnect(domid, fintype, IDNFIN_ARG_NONE, IDNDS_SYNC_TYPE(dp)); } static int idn_send_fin(int domid, idn_msgtype_t *mtp, idn_fin_t fintype, idn_finarg_t finarg, idn_finopt_t finopt, domainset_t finset, uint_t finmaster) { int need_timer = 1; uint_t acknack; uint_t fintypearg = 0; idn_msgtype_t mt; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_send_fin"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_HELD(domid)); ASSERT((fintype != IDNFIN_QUERY) ? (finopt != IDNFIN_OPT_NONE) : 1); if (mtp) { acknack = mtp->mt_mtype & IDNP_ACKNACK_MASK; mt.mt_mtype = mtp->mt_mtype; mt.mt_atype = mtp->mt_atype; mt.mt_cookie = mtp->mt_cookie; } else { acknack = 0; mt.mt_mtype = IDNP_FIN; mt.mt_atype = 0; /* * For simple FIN queries we want a unique * timer assigned. For others, they * effectively share one. */ if (fintype == IDNFIN_QUERY) mt.mt_cookie = 0; else mt.mt_cookie = IDN_TIMER_PUBLIC_COOKIE; } PR_PROTO("%s:%d: sending fin%sto (cpu %d) " "[ft=%s, fa=%s, fs=0x%x, fo=%s, fm=(%d,%d)]\n", proc, domid, (acknack & IDNP_ACK) ? "+ack " : (acknack & IDNP_NACK) ? "+nack " : " ", dp->dcpu, idnfin_str[fintype], idnfinarg_str[finarg], (int)finset, idnfinopt_str[finopt], FIN_MASTER_DOMID(finmaster), FIN_MASTER_CPUID(finmaster)); if (need_timer) { IDN_MSGTIMER_START(domid, IDNP_FIN, (ushort_t)fintype, idn_msg_waittime[IDNP_FIN], &mt.mt_cookie); } SET_FIN_TYPE(fintypearg, fintype); SET_FIN_ARG(fintypearg, finarg); IDNXDC(domid, &mt, fintypearg, (uint_t)finset, (uint_t)finopt, finmaster); return (0); } /* * Must leave w/DLOCK dropped and SYNC_LOCK held. */ static int idn_recv_fin(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; idn_fin_t fintype; idn_finarg_t finarg; idn_finopt_t finopt; domainset_t my_ready_set, ready_set; idn_msgtype_t mt; idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t aargs; procname_t proc = "idn_recv_fin"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp ? mtp->mt_cookie : 0; fintype = GET_XARGS_FIN_TYPE(xargs); finarg = GET_XARGS_FIN_ARG(xargs); ready_set = GET_XARGS_FIN_DOMSET(xargs); finopt = GET_XARGS_FIN_OPT(xargs); CLR_XARGS(aargs); if (msg & IDNP_NACK) { PR_PROTO("%s:%d: received NACK (type = %s)\n", proc, domid, idnnack_str[xargs[0]]); } else { PR_PROTO("%s:%d: fintype = %s, finopt = %s, " "finarg = %s, ready_set = 0x%x\n", proc, domid, idnfin_str[fintype], idnfinopt_str[finopt], idnfinarg_str[finarg], ready_set); } if (!(msg & IDNP_NACK) && (fintype == IDNFIN_QUERY)) { domainset_t query_set; query_set = idn_sync_register(domid, IDNSYNC_DISCONNECT, ready_set, IDNSYNC_REG_REG); my_ready_set = ~idn.domset.ds_connected | idn.domset.ds_ready_off; if (msg & IDNP_MSGTYPE_MASK) { mt.mt_mtype = IDNP_ACK; mt.mt_atype = IDNP_FIN; SET_XARGS_FIN_TYPE(aargs, fintype); SET_XARGS_FIN_ARG(aargs, finarg); SET_XARGS_FIN_DOMSET(aargs, my_ready_set); SET_XARGS_FIN_OPT(aargs, IDNFIN_OPT_NONE); SET_XARGS_FIN_MASTER(aargs, NIL_FIN_MASTER); idn_send_acknack(domid, &mt, aargs); } if (query_set) { uint_t token; token = IDN_RETRY_TOKEN(domid, IDNRETRY_FINQ); idn_retry_submit(idn_retry_query, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FINQ]); } return (0); } if (dp->dxp != &xphase_fin) { uint_t token; if (IDNDS_IS_CLOSED(dp)) { PR_PROTO("%s:%d: domain already closed (%s)\n", proc, domid, idnds_str[dp->dstate]); if (msg & IDNP_MSGTYPE_MASK) { /* * fin or fin+ack. */ mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; SET_XARGS_NACK_TYPE(aargs, IDNNACK_NOCONN); idn_send_acknack(domid, &mt, aargs); } return (0); } dp->dfin_sync = IDNDS_SYNC_TYPE(dp); /* * Need to do some clean-up ala idn_disconnect(). * * Terminate any outstanding commands that were * targeted towards this domain. */ idn_terminate_cmd(domid, ECANCELED); /* * Terminate any and all retries that may have * outstanding for this domain. */ token = IDN_RETRY_TOKEN(domid, IDN_RETRY_TYPEALL); (void) idn_retry_terminate(token); /* * Stop all outstanding message timers for * this guy. */ IDN_MSGTIMER_STOP(domid, 0, 0); dp->dxp = &xphase_fin; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); } if (msg & IDNP_NACK) { idn_nack_t nack; nack = GET_XARGS_NACK_TYPE(xargs); if (nack == IDNNACK_NOCONN) { /* * We're trying to FIN with somebody we're * already disconnected from. Need to * speed this guy through. */ DOMAINSET_ADD(idn.domset.ds_ready_off, domid); (void) idn_sync_register(domid, IDNSYNC_DISCONNECT, DOMAINSET_ALL, IDNSYNC_REG_REG); ready_set = (uint_t)DOMAINSET_ALL; /* * Need to transform message to allow us to * pass this guy right through and not waste time * talking to him. */ IDN_FSTATE_TRANSITION(dp, IDNFIN_FORCE_HARD); switch (dp->dstate) { case IDNDS_FIN_PEND: mtp->mt_mtype = 0; mtp->mt_atype = 0; break; case IDNDS_FIN_SENT: mtp->mt_mtype = IDNP_FIN | IDNP_ACK; mtp->mt_atype = 0; break; case IDNDS_FIN_RCVD: mtp->mt_mtype = IDNP_ACK; mtp->mt_atype = IDNP_FIN | IDNP_ACK; break; default: #ifdef DEBUG cmn_err(CE_PANIC, "%s:%d: UNEXPECTED state = %s", proc, domid, idnds_str[dp->dstate]); #endif /* DEBUG */ break; } } fintype = (uint_t)dp->dfin; finopt = DOMAIN_IN_SET(idn.domset.ds_relink, domid) ? IDNFIN_OPT_RELINK : IDNFIN_OPT_UNLINK; CLR_XARGS(xargs); SET_XARGS_FIN_TYPE(xargs, fintype); SET_XARGS_FIN_ARG(xargs, finarg); SET_XARGS_FIN_DOMSET(xargs, ready_set); SET_XARGS_FIN_OPT(xargs, finopt); SET_XARGS_FIN_MASTER(xargs, NIL_FIN_MASTER); } idn_xphase_transition(domid, mtp, xargs); return (0); } /*ARGSUSED1*/ static void idn_retry_fin(uint_t token, void *arg) { int domid = IDN_RETRY_TOKEN2DOMID(token); int new_masterid, new_cpuid = IDN_NIL_DCPU; uint_t finmaster; idn_domain_t *dp = &idn_domain[domid]; idn_xdcargs_t xargs; idn_finopt_t finopt; procname_t proc = "idn_retry_fin"; ASSERT(IDN_RETRY_TOKEN2TYPE(token) == IDNRETRY_FIN); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); if (dp->dxp != &xphase_fin) { PR_PROTO("%s:%d: dxp(0x%p) != xstate_fin(0x%p)...bailing\n", proc, domid, dp->dxp, &xphase_fin); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } if (dp->dxstate != IDNXS_PEND) { PR_PROTO("%s:%d: xstate(%s) != %s...bailing\n", proc, domid, idnxs_str[dp->dxstate], idnxs_str[IDNXS_PEND]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } finopt = DOMAIN_IN_SET(idn.domset.ds_relink, domid) ? IDNFIN_OPT_RELINK : IDNFIN_OPT_UNLINK; CLR_XARGS(xargs); SET_XARGS_FIN_TYPE(xargs, dp->dfin); /*LINTED*/ SET_XARGS_FIN_ARG(xargs, IDNFIN_ARG_NONE); SET_XARGS_FIN_OPT(xargs, finopt); SET_XARGS_FIN_DOMSET(xargs, 0); /* unused when msg == 0 */ IDN_GLOCK_SHARED(); new_masterid = IDN_GET_NEW_MASTERID(); IDN_GUNLOCK(); if (new_masterid != IDN_NIL_DOMID) new_cpuid = idn_domain[new_masterid].dcpu; finmaster = MAKE_FIN_MASTER(new_masterid, new_cpuid); SET_XARGS_FIN_MASTER(xargs, finmaster); idn_xphase_transition(domid, NULL, xargs); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); } static int idn_check_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { idn_domain_t *dp = &idn_domain[domid]; idn_fin_t fintype; idn_finopt_t finopt; idn_finarg_t finarg; int ready; int finmasterid; int fincpuid; uint_t finmaster; uint_t msg = mtp ? mtp->mt_mtype : 0; domainset_t query_set, ready_set, conn_set; domainset_t my_ready_set, shutdown_set; procname_t proc = "idn_check_fin_pend"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) return (0); if ((dp->dstate == IDNDS_FIN_PEND) && (msg & IDNP_MSGTYPE_MASK) && (msg & IDNP_ACK)) /* fin+ack */ return (1); query_set = 0; if (!DOMAIN_IN_SET(idn.domset.ds_trans_off, domid)) { /* * Can't remove domain from ds_connected yet, * since he's still officially connected until * we get an ACK from him. */ DOMAINSET_DEL(idn.domset.ds_trans_on, domid); DOMAINSET_ADD(idn.domset.ds_trans_off, domid); } IDN_GLOCK_SHARED(); conn_set = (idn.domset.ds_connected | idn.domset.ds_trans_on) & ~idn.domset.ds_trans_off; if ((idn.state == IDNGS_DISCONNECT) || (idn.state == IDNGS_RECONFIG) || (domid == IDN_GET_MASTERID()) || !conn_set) { /* * If we're disconnecting, reconfiguring, * unlinking from the master, or unlinking * the last of our connections, then we need * to shutdown all the channels. */ shutdown_set = DOMAINSET_ALL; } else { shutdown_set = DOMAINSET(domid); } IDN_GUNLOCK(); idn_shutdown_datapath(shutdown_set, (dp->dfin == IDNFIN_FORCE_HARD)); IDN_GLOCK_EXCL(); /* * Remap the SMR back to our local space if the remote * domain going down is the master. We do this now before * flushing caches. This will help guarantee that any * accidental accesses to the SMR after the cache flush * will only go to local memory. */ if ((domid == IDN_GET_MASTERID()) && (idn.smr.rempfn != PFN_INVALID)) { PR_PROTO("%s:%d: deconfiging CURRENT MASTER - SMR remap\n", proc, domid); IDN_DLOCK_EXCL(idn.localid); /* * We're going to remap the SMR, * so gotta blow away our local * pointer to the mbox table. */ idn_domain[idn.localid].dmbox.m_tbl = NULL; IDN_DUNLOCK(idn.localid); idn.smr.rempfn = PFN_INVALID; idn.smr.rempfnlim = PFN_INVALID; smr_remap(&kas, idn.smr.vaddr, idn.smr.locpfn, IDN_SMR_SIZE); } IDN_GUNLOCK(); if (DOMAIN_IN_SET(idn.domset.ds_flush, domid)) { idnxf_flushall_ecache(); CHECKPOINT_CLOSED(IDNSB_CHKPT_CACHE, dp->dhw.dh_boardset, 2); DOMAINSET_DEL(idn.domset.ds_flush, domid); } fintype = GET_XARGS_FIN_TYPE(xargs); finarg = GET_XARGS_FIN_ARG(xargs); ready_set = GET_XARGS_FIN_DOMSET(xargs); finopt = GET_XARGS_FIN_OPT(xargs); ASSERT(fintype != IDNFIN_QUERY); if (!VALID_FIN(fintype)) { /* * If for some reason remote domain * sent us an invalid FIN type, * override it to a NORMAL fin. */ PR_PROTO("%s:%d: WARNING invalid fintype (%d) -> %s(%d)\n", proc, domid, (int)fintype, idnfin_str[IDNFIN_NORMAL], (int)IDNFIN_NORMAL); fintype = IDNFIN_NORMAL; } if (!VALID_FINOPT(finopt)) { PR_PROTO("%s:%d: WARNING invalid finopt (%d) -> %s(%d)\n", proc, domid, (int)finopt, idnfinopt_str[IDNFIN_OPT_UNLINK], (int)IDNFIN_OPT_UNLINK); finopt = IDNFIN_OPT_UNLINK; } finmaster = GET_XARGS_FIN_MASTER(xargs); finmasterid = FIN_MASTER_DOMID(finmaster); fincpuid = FIN_MASTER_CPUID(finmaster); if ((finarg != IDNFIN_ARG_NONE) && !DOMAIN_IN_SET(idn.domset.ds_hitlist, domid)) { idnsb_error_t idnerr; INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EPROTO); SET_IDNKERR_IDNERR(&idnerr, FINARG2IDNKERR(finarg)); SET_IDNKERR_PARAM0(&idnerr, domid); if (IDNFIN_ARG_IS_FATAL(finarg)) { finopt = IDNFIN_OPT_UNLINK; DOMAINSET_DEL(idn.domset.ds_relink, domid); DOMAINSET_ADD(idn.domset.ds_hitlist, domid); if (idn.domset.ds_connected == 0) { domainset_t domset; IDN_GLOCK_EXCL(); domset = ~idn.domset.ds_relink; if (idn.domset.ds_relink == 0) { IDN_GSTATE_TRANSITION(IDNGS_DISCONNECT); } domset &= ~idn.domset.ds_hitlist; /* * The primary domain we were trying to * connect to fin'd us with a fatal argument. * Something isn't cool in our IDN environment, * e.g. corrupted SMR or non-compatible CONFIG * parameters. In any case we need to dismantle * ourselves completely. */ IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); IDN_GUNLOCK(); IDN_DUNLOCK(domid); DOMAINSET_DEL(domset, idn.localid); DOMAINSET_DEL(domset, domid); idn_update_op(IDNOP_ERROR, DOMAINSET_ALL, &idnerr); PR_HITLIST("%s:%d: unlink_domainset(%x) " "due to CFG error (relink=%x, " "hitlist=%x)\n", proc, domid, domset, idn.domset.ds_relink, idn.domset.ds_hitlist); idn_unlink_domainset(domset, IDNFIN_NORMAL, finarg, IDNFIN_OPT_UNLINK, BOARDSET_ALL); IDN_DLOCK_EXCL(domid); } PR_HITLIST("%s:%d: CFG error, (conn=%x, relink=%x, " "hitlist=%x)\n", proc, domid, idn.domset.ds_connected, idn.domset.ds_relink, idn.domset.ds_hitlist); } idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); } if ((finmasterid != IDN_NIL_DOMID) && (!VALID_DOMAINID(finmasterid) || DOMAIN_IN_SET(idn.domset.ds_hitlist, domid))) { PR_HITLIST("%s:%d: finmasterid = %d -> -1, relink=%x, " "hitlist=%x\n", proc, domid, finmasterid, idn.domset.ds_relink, idn.domset.ds_hitlist); PR_PROTO("%s:%d: WARNING invalid finmasterid (%d) -> -1\n", proc, domid, finmasterid); finmasterid = IDN_NIL_DOMID; } IDN_GLOCK_EXCL(); if ((finopt == IDNFIN_OPT_RELINK) && (idn.state != IDNGS_DISCONNECT)) { DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); DOMAINSET_ADD(idn.domset.ds_hitlist, domid); } if ((domid == IDN_GET_NEW_MASTERID()) && !DOMAIN_IN_SET(idn.domset.ds_relink, domid)) { IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); } if ((idn.state != IDNGS_DISCONNECT) && (idn.state != IDNGS_RECONFIG) && (domid == IDN_GET_MASTERID())) { domainset_t dis_set, master_candidates; IDN_GKSTAT_GLOBAL_EVENT(gk_reconfigs, gk_reconfig_last); IDN_GSTATE_TRANSITION(IDNGS_RECONFIG); IDN_GUNLOCK(); if ((finmasterid != IDN_NIL_DOMID) && (finmasterid != idn.localid)) { if (finmasterid != domid) IDN_DLOCK_EXCL(finmasterid); if (idn_open_domain(finmasterid, fincpuid, 0) < 0) { cmn_err(CE_WARN, "IDN: 205: (%s) failed to " "open-domain(%d,%d)", proc, finmasterid, fincpuid); if (finmasterid != domid) IDN_DUNLOCK(finmasterid); finmasterid = IDN_NIL_DOMID; } if (finmasterid != domid) IDN_DUNLOCK(finmasterid); } IDN_GLOCK_EXCL(); if (finmasterid == IDN_NIL_DOMID) { int m; master_candidates = idn.domset.ds_trans_on | idn.domset.ds_connected | idn.domset.ds_relink; master_candidates &= ~(idn.domset.ds_trans_off & ~idn.domset.ds_relink); DOMAINSET_DEL(master_candidates, domid); /* * Local domain gets to participate also. */ DOMAINSET_ADD(master_candidates, idn.localid); m = idn_select_candidate(master_candidates); IDN_SET_NEW_MASTERID(m); } else { IDN_SET_NEW_MASTERID(finmasterid); } IDN_GUNLOCK(); dis_set = idn.domset.ds_trans_on | idn.domset.ds_connected; DOMAINSET_DEL(dis_set, domid); idn_unlink_domainset(dis_set, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNFIN_OPT_RELINK, BOARDSET_ALL); } else { IDN_GUNLOCK(); } /* * My local ready-set are those domains from which I * have confirmed no datapaths exist. */ my_ready_set = ~idn.domset.ds_connected; switch (dp->dfin) { case IDNFIN_NORMAL: case IDNFIN_FORCE_SOFT: case IDNFIN_FORCE_HARD: if (fintype < dp->dfin) { /* * Remote domain has requested a * FIN of lower priority than what * we're currently running. Just * leave the priority where it is. */ break; } /*FALLTHROUGH*/ default: IDN_FSTATE_TRANSITION(dp, fintype); break; } ASSERT(dp->dfin_sync != IDNFIN_SYNC_OFF); if (msg == 0) { /* * Local domain is initiating a FIN sequence * to remote domid. Note that remote domain * remains in ds_connected even though he's * in thet ready-set from the local domain's * perspective. We can't remove him from * ds_connected until we get a confirmed message * from him indicating he has ceased communication. */ ready_set = my_ready_set; } else { /* * Remote domain initiated a FIN sequence * to local domain. This implies that he * has shutdown his datapath to us. Since * we shutdown our datapath to him, we're * effectively now in his ready-set. */ DOMAINSET_ADD(ready_set, idn.localid); /* * Since we know both sides of the connection * have ceased, this remote domain is effectively * considered disconnected. */ DOMAINSET_ADD(idn.domset.ds_ready_off, domid); } if (dp->dfin == IDNFIN_FORCE_HARD) { /* * If we're doing a hard disconnect * of this domain then we want to * blow straight through and not * waste time trying to talk to the * remote domain nor to domains we * believe are AWOL. Although we will * try and do it cleanly with * everybody else. */ DOMAINSET_ADD(my_ready_set, domid); my_ready_set |= idn.domset.ds_awol; ready_set = DOMAINSET_ALL; } else if (dp->dfin_sync == IDNFIN_SYNC_NO) { /* * If we're not fin'ing this domain * synchronously then the only * expected domain set is himself. */ ready_set |= ~DOMAINSET(domid); my_ready_set |= ~DOMAINSET(domid); } if (dp->dsync.s_cmd != IDNSYNC_DISCONNECT) { idn_sync_exit(domid, IDNSYNC_CONNECT); idn_sync_enter(domid, IDNSYNC_DISCONNECT, DOMAINSET_ALL, my_ready_set, idn_xstate_transfunc, (void *)IDNP_FIN); } query_set = idn_sync_register(domid, IDNSYNC_DISCONNECT, ready_set, IDNSYNC_REG_REG); /* * No need to query this domain as he's already * in the FIN sequence. */ DOMAINSET_DEL(query_set, domid); ready = (dp->dsync.s_set_exp == dp->dsync.s_set_rdy) ? 1 : 0; if (ready) { DOMAINSET_DEL(idn.domset.ds_ready_off, domid); DOMAINSET_DEL(idn.domset.ds_connected, domid); } if (query_set) { int d; my_ready_set = idn.domset.ds_ready_off | ~idn.domset.ds_connected; for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(query_set, d)) continue; dp = &idn_domain[d]; IDN_DLOCK_EXCL(d); if (dp->dsync.s_cmd == IDNSYNC_DISCONNECT) { IDN_DUNLOCK(d); continue; } IDN_SYNC_QUERY_UPDATE(domid, d); idn_send_fin(d, NULL, IDNFIN_QUERY, IDNFIN_ARG_NONE, IDNFIN_OPT_NONE, my_ready_set, NIL_FIN_MASTER); IDN_DUNLOCK(d); } } return (!msg ? 0 : (ready ? 0 : 1)); } /*ARGSUSED*/ static void idn_error_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t token; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_HELD(domid)); /* * Don't communicate with domains that * we're forcing a hard disconnect. */ if ((idn_domain[domid].dfin != IDNFIN_FORCE_HARD) && (msg & IDNP_MSGTYPE_MASK)) { idn_msgtype_t mt; idn_xdcargs_t nargs; mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; mt.mt_cookie = mtp->mt_cookie; CLR_XARGS(nargs); SET_XARGS_NACK_TYPE(nargs, IDNNACK_RETRY); idn_send_acknack(domid, &mt, nargs); } token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } static void idn_action_fin_pend(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { idn_domain_t *dp = &idn_domain[domid]; domainset_t my_ready_set; idn_finopt_t finopt; idn_finarg_t finarg; uint_t finmaster; int new_masterid, new_cpuid = IDN_NIL_DCPU; uint_t msg = mtp ? mtp->mt_mtype : 0; idn_msgtype_t mt; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_HELD(domid)); my_ready_set = dp->dsync.s_set_rdy | idn.domset.ds_ready_off | ~idn.domset.ds_connected; ASSERT(xargs[0] != (uint_t)IDNFIN_QUERY); finarg = GET_XARGS_FIN_ARG(xargs); finopt = DOMAIN_IN_SET(idn.domset.ds_relink, domid) ? IDNFIN_OPT_RELINK : IDNFIN_OPT_UNLINK; mt.mt_cookie = mtp ? mtp->mt_cookie : 0; IDN_GLOCK_SHARED(); new_masterid = IDN_GET_NEW_MASTERID(); IDN_GUNLOCK(); if (new_masterid != IDN_NIL_DOMID) new_cpuid = idn_domain[new_masterid].dcpu; finmaster = MAKE_FIN_MASTER(new_masterid, new_cpuid); if (dp->dfin == IDNFIN_FORCE_HARD) { ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (!msg) { mt.mt_mtype = IDNP_FIN | IDNP_ACK; mt.mt_atype = 0; } else { mt.mt_mtype = IDNP_ACK; mt.mt_atype = IDNP_FIN | IDNP_ACK; } idn_xphase_transition(domid, &mt, xargs); } else if (!msg) { idn_send_fin(domid, NULL, dp->dfin, finarg, finopt, my_ready_set, finmaster); } else if ((msg & IDNP_ACKNACK_MASK) == 0) { /* * fin */ mt.mt_mtype = IDNP_FIN | IDNP_ACK; mt.mt_atype = 0; idn_send_fin(domid, &mt, dp->dfin, finarg, finopt, my_ready_set, finmaster); } else { uint_t token; /* * nack - retry */ token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } } static int idn_check_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { int ready; uint_t msg = mtp ? mtp->mt_mtype : 0; idn_fin_t fintype; idn_finopt_t finopt; idn_domain_t *dp = &idn_domain[domid]; domainset_t query_set, ready_set; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) return (0); fintype = GET_XARGS_FIN_TYPE(xargs); ready_set = GET_XARGS_FIN_DOMSET(xargs); finopt = GET_XARGS_FIN_OPT(xargs); ASSERT(fintype != IDNFIN_QUERY); if (!VALID_FIN(fintype)) { /* * If for some reason remote domain * sent us an invalid FIN type, * override it to a NORMAL fin. */ fintype = IDNFIN_NORMAL; } if (!VALID_FINOPT(finopt)) { finopt = IDNFIN_OPT_UNLINK; } IDN_GLOCK_SHARED(); if ((finopt == IDNFIN_OPT_RELINK) && (idn.state != IDNGS_DISCONNECT)) { DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); } else { DOMAINSET_DEL(idn.domset.ds_relink, domid); } IDN_GUNLOCK(); switch (dp->dfin) { case IDNFIN_NORMAL: case IDNFIN_FORCE_SOFT: case IDNFIN_FORCE_HARD: if (fintype < dp->dfin) { /* * Remote domain has requested a * FIN of lower priority than what * we're current running. Just * leave the priority where it is. */ break; } /*FALLTHROUGH*/ default: IDN_FSTATE_TRANSITION(dp, fintype); break; } if (dp->dfin == IDNFIN_FORCE_HARD) { /* * If we're doing a hard disconnect * of this domain then we want to * blow straight through and not * waste time trying to talk to the * remote domain. By registering him * as ready with respect to all * possible domains he'll transition * immediately. Note that we'll still * try and do it coherently with * other domains to which we're connected. */ ready_set = DOMAINSET_ALL; } else { DOMAINSET_ADD(ready_set, idn.localid); } DOMAINSET_ADD(idn.domset.ds_ready_off, domid); query_set = idn_sync_register(domid, IDNSYNC_DISCONNECT, ready_set, IDNSYNC_REG_REG); /* * No need to query this domain as he's already * in the FIN sequence. */ DOMAINSET_DEL(query_set, domid); ready = (dp->dsync.s_set_exp == dp->dsync.s_set_rdy) ? 1 : 0; if (ready) { DOMAINSET_DEL(idn.domset.ds_ready_off, domid); DOMAINSET_DEL(idn.domset.ds_connected, domid); } if (query_set) { int d; domainset_t my_ready_set; my_ready_set = idn.domset.ds_ready_off | ~idn.domset.ds_connected; for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(query_set, d)) continue; dp = &idn_domain[d]; IDN_DLOCK_EXCL(d); if (dp->dsync.s_cmd == IDNSYNC_DISCONNECT) { IDN_DUNLOCK(d); continue; } IDN_SYNC_QUERY_UPDATE(domid, d); idn_send_fin(d, NULL, IDNFIN_QUERY, IDNFIN_ARG_NONE, IDNFIN_OPT_NONE, my_ready_set, NIL_FIN_MASTER); IDN_DUNLOCK(d); } } return ((ready > 0) ? 0 : 1); } /*ARGSUSED*/ static void idn_error_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t token; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); /* * Don't communicate with domains that * we're forcing a hard disconnect. */ if ((idn_domain[domid].dfin != IDNFIN_FORCE_HARD) && (msg & IDNP_MSGTYPE_MASK)) { idn_msgtype_t mt; idn_xdcargs_t nargs; mt.mt_mtype = IDNP_NACK; mt.mt_atype = msg; mt.mt_cookie = mtp->mt_cookie; CLR_XARGS(nargs); SET_XARGS_NACK_TYPE(nargs, IDNNACK_RETRY); idn_send_acknack(domid, &mt, nargs); } token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } static void idn_action_fin_sent(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; int new_masterid, new_cpuid = IDN_NIL_DCPU; uint_t finmaster; idn_msgtype_t mt; idn_finopt_t finopt; idn_finarg_t finarg; domainset_t my_ready_set; idn_domain_t *dp = &idn_domain[domid]; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); mt.mt_cookie = mtp ? mtp->mt_cookie : 0; finopt = DOMAIN_IN_SET(idn.domset.ds_relink, domid) ? IDNFIN_OPT_RELINK : IDNFIN_OPT_UNLINK; finarg = GET_XARGS_FIN_ARG(xargs); my_ready_set = dp->dsync.s_set_rdy | idn.domset.ds_ready_off | ~idn.domset.ds_connected; IDN_GLOCK_SHARED(); new_masterid = IDN_GET_NEW_MASTERID(); IDN_GUNLOCK(); if (new_masterid != IDN_NIL_DOMID) new_cpuid = idn_domain[new_masterid].dcpu; finmaster = MAKE_FIN_MASTER(new_masterid, new_cpuid); if ((msg & IDNP_ACKNACK_MASK) == 0) { /* * fin */ if (dp->dfin == IDNFIN_FORCE_HARD) { mt.mt_mtype = IDNP_ACK; mt.mt_atype = IDNP_FIN | IDNP_ACK; idn_xphase_transition(domid, &mt, xargs); } else { mt.mt_mtype = IDNP_FIN | IDNP_ACK; mt.mt_atype = 0; idn_send_fin(domid, &mt, dp->dfin, finarg, finopt, my_ready_set, finmaster); } } else if (msg & IDNP_MSGTYPE_MASK) { /* * fin+ack */ if (dp->dfin != IDNFIN_FORCE_HARD) { idn_xdcargs_t fargs; mt.mt_mtype = IDNP_ACK; mt.mt_atype = msg; CLR_XARGS(fargs); SET_XARGS_FIN_TYPE(fargs, dp->dfin); SET_XARGS_FIN_ARG(fargs, finarg); SET_XARGS_FIN_DOMSET(fargs, my_ready_set); SET_XARGS_FIN_OPT(fargs, finopt); SET_XARGS_FIN_MASTER(fargs, finmaster); idn_send_acknack(domid, &mt, fargs); } } else { uint_t token; /* * nack - retry */ token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } } /*ARGSUSED*/ static void idn_action_fin_rcvd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (msg & IDNP_NACK) { uint_t token; /* * nack - retry. */ token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } } static void idn_final_fin(int domid) { int do_relink; int rv, d, new_masterid = IDN_NIL_DOMID; idn_gstate_t next_gstate; domainset_t relinkset; uint_t token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_domain_t *ldp, *dp = &idn_domain[domid]; procname_t proc = "idn_final_fin"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(dp->dstate == IDNDS_DMAP); (void) idn_retry_terminate(token); dp->dxp = NULL; IDN_XSTATE_TRANSITION(dp, IDNXS_NIL); idn_sync_exit(domid, IDNSYNC_DISCONNECT); DOMAINSET_DEL(idn.domset.ds_trans_off, domid); do_relink = DOMAIN_IN_SET(idn.domset.ds_relink, domid) ? 1 : 0; /* * idn_deconfig will idn_close_domain. */ idn_deconfig(domid); PR_PROTO("%s:%d: DISCONNECTED\n", proc, domid); IDN_GLOCK_EXCL(); /* * It's important that this update-op occur within * the context of holding the glock(EXCL). There is * still some additional state stuff to cleanup which * will be completed once the glock is dropped in * this flow. Which means anybody that's doing a * SSI_INFO and waiting on glock will not actually * run until the clean-up is completed, which is what * we want. Recall that a separate thread processes * the SSI_LINK/UNLINK calls and when they complete * (i.e. are awakened) they will immediately SSI_INFO * and we don't want them to prematurely pick up stale * information. */ idn_update_op(IDNOP_DISCONNECTED, DOMAINSET(domid), NULL); ASSERT(idn.state != IDNGS_OFFLINE); ASSERT(!DOMAIN_IN_SET(idn.domset.ds_trans_on, domid)); if (domid == IDN_GET_MASTERID()) { IDN_SET_MASTERID(IDN_NIL_DOMID); dp->dvote.v.master = 0; } if ((domid == IDN_GET_NEW_MASTERID()) && !do_relink) { IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); } if (idn.state == IDNGS_RECONFIG) new_masterid = IDN_GET_NEW_MASTERID(); if ((idn.domset.ds_trans_on | idn.domset.ds_trans_off | idn.domset.ds_relink) == 0) { PR_HITLIST("%s:%d: HITLIST %x -> 0\n", proc, domid, idn.domset.ds_hitlist); idn.domset.ds_hitlist = 0; } if (idn.domset.ds_connected || idn.domset.ds_trans_off) { PR_PROTO("%s:%d: ds_connected = 0x%x, ds_trans_off = 0x%x\n", proc, domid, idn.domset.ds_connected, idn.domset.ds_trans_off); IDN_GUNLOCK(); goto fin_done; } IDN_DLOCK_EXCL(idn.localid); ldp = &idn_domain[idn.localid]; if (idn.domset.ds_trans_on != 0) { ASSERT((idn.state != IDNGS_DISCONNECT) && (idn.state != IDNGS_OFFLINE)); switch (idn.state) { case IDNGS_CONNECT: if (idn.localid == IDN_GET_MASTERID()) { idn_master_deinit(); IDN_SET_MASTERID(IDN_NIL_DOMID); ldp->dvote.v.master = 0; } /*FALLTHROUGH*/ case IDNGS_ONLINE: next_gstate = idn.state; break; case IDNGS_RECONFIG: if (idn.localid == IDN_GET_MASTERID()) { idn_master_deinit(); IDN_SET_MASTERID(IDN_NIL_DOMID); ldp->dvote.v.master = 0; } ASSERT(IDN_GET_MASTERID() == IDN_NIL_DOMID); next_gstate = IDNGS_CONNECT; ldp->dvote.v.connected = 0; /* * Need to do HWINIT since we won't * be transitioning through OFFLINE * which would normally be caught in * idn_check_nego() when we * initially go to CONNECT. */ IDN_PREP_HWINIT(); break; case IDNGS_DISCONNECT: case IDNGS_OFFLINE: cmn_err(CE_WARN, "IDN: 211: disconnect domain %d, " "unexpected Gstate (%s)", domid, idngs_str[idn.state]); IDN_DUNLOCK(idn.localid); IDN_GUNLOCK(); goto fin_done; default: /* * XXX * Go into FATAL state? */ cmn_err(CE_PANIC, "IDN: 212: disconnect domain %d, " "bad Gstate (%d)", domid, idn.state); /* not reached */ break; } } else { if (idn.localid == IDN_GET_MASTERID()) { idn_master_deinit(); IDN_SET_MASTERID(IDN_NIL_DOMID); ldp->dvote.v.master = 0; } next_gstate = IDNGS_OFFLINE; if (idn.domset.ds_relink == 0) { IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); } } IDN_DUNLOCK(idn.localid); /* * If we reach here we've effectively disconnected all * existing links, however new ones may be pending. */ PR_PROTO("%s:%d: ALL DISCONNECTED *****************\n", proc, domid); IDN_GSTATE_TRANSITION(next_gstate); ASSERT((idn.state == IDNGS_OFFLINE) ? (IDN_GET_MASTERID() == IDN_NIL_DOMID) : 1); IDN_GUNLOCK(); /* * If we have no new masterid and yet there are relinkers * out there, then force us to attempt to link with one * of them. */ if ((new_masterid == IDN_NIL_DOMID) && idn.domset.ds_relink) new_masterid = idn.localid; if (new_masterid != IDN_NIL_DOMID) { /* * If the local domain is the selected * master then we'll want to initiate * a link with one of the other candidates. * If not, then we want to initiate a link * with the master only. */ relinkset = (new_masterid == idn.localid) ? idn.domset.ds_relink : DOMAINSET(new_masterid); DOMAINSET_DEL(relinkset, idn.localid); for (d = 0; d < MAX_DOMAINS; d++) { int lock_held; if (!DOMAIN_IN_SET(relinkset, d)) continue; if (d == domid) { do_relink = 0; lock_held = 0; } else { IDN_DLOCK_EXCL(d); lock_held = 1; } rv = idn_open_domain(d, -1, 0); if (rv == 0) { rv = idn_connect(d); if (lock_held) IDN_DUNLOCK(d); /* * If we're able to kick off at * least one connect then that's * good enough for now. The others * will fall into place normally. */ if (rv == 0) break; } else if (rv < 0) { if (lock_held) IDN_DUNLOCK(d); cmn_err(CE_WARN, "IDN: 205: (%s.1) failed to " "open-domain(%d,%d)", proc, domid, -1); DOMAINSET_DEL(idn.domset.ds_relink, d); } else { if (lock_held) IDN_DUNLOCK(d); PR_PROTO("%s:%d: failed to " "re-open domain %d " "(cpu %d) [rv = %d]\n", proc, domid, d, idn_domain[d].dcpu, rv); } } } fin_done: if (do_relink) { ASSERT(IDN_DLOCK_IS_EXCL(domid)); rv = idn_open_domain(domid, -1, 0); if (rv == 0) { (void) idn_connect(domid); } else if (rv < 0) { cmn_err(CE_WARN, "IDN: 205: (%s.2) failed to " "open-domain(%d,%d)", proc, domid, -1); DOMAINSET_DEL(idn.domset.ds_relink, domid); } } } static void idn_exit_fin(int domid, uint_t msgtype) { idn_domain_t *dp = &idn_domain[domid]; uint_t token; procname_t proc = "idn_exit_fin"; STRING(str); ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); INUM2STR(msgtype, str); PR_PROTO("%s:%d: msgtype = 0x%x(%s)\n", proc, domid, msgtype, str); token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); (void) idn_retry_terminate(token); DOMAINSET_DEL(idn.domset.ds_ready_off, domid); dp->dxp = &xphase_fin; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); } /* * Must return w/locks held. */ static int idn_xphase_transition(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t msgarg = mtp ? mtp->mt_atype : 0; idn_xphase_t *xp; idn_domain_t *dp; int (*cfunc)(int, idn_msgtype_t *, idn_xdcargs_t); void (*ffunc)(int); void (*afunc)(int, idn_msgtype_t *, idn_xdcargs_t); void (*efunc)(int, idn_msgtype_t *, idn_xdcargs_t); void (*xfunc)(int, uint_t); int err = 0; uint_t msgtype; idn_xstate_t o_xstate, n_xstate; procname_t proc = "idn_xphase_transition"; STRING(mstr); STRING(astr); ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); INUM2STR(msg, mstr); INUM2STR(msgarg, astr); dp = &idn_domain[domid]; if ((xp = dp->dxp) == NULL) { PR_PROTO("%s:%d: WARNING: domain xsp is NULL (msg = %s, " "msgarg = %s) <<<<<<<<<<<<\n", proc, domid, mstr, astr); return (-1); } o_xstate = dp->dxstate; xfunc = xp->xt_exit; if ((msgtype = (msg & IDNP_MSGTYPE_MASK)) == 0) msgtype = msgarg & IDNP_MSGTYPE_MASK; if ((o_xstate == IDNXS_PEND) && msg && ((msg & IDNP_ACKNACK_MASK) == msg)) { PR_PROTO("%s:%d: unwanted acknack received (o_xstate = %s, " "msg = %s/%s - dropping message\n", proc, domid, idnxs_str[(int)o_xstate], mstr, astr); return (0); } /* * Validate that message received is following * the expected protocol for the current state. */ if (idn_next_xstate(o_xstate, -1, msg) == IDNXS_NIL) { PR_PROTO("%s:%d: WARNING: o_xstate = %s, msg = %s -> NIL " "<<<<<<<<<\n", proc, domid, idnxs_str[(int)o_xstate], mstr); if (xfunc) (*xfunc)(domid, msgtype); return (-1); } if (msg || msgarg) { /* * Verify that message type is correct for * the given xstate. */ if (msgtype != xp->xt_msgtype) { STRING(xstr); STRING(tstr); INUM2STR(xp->xt_msgtype, xstr); INUM2STR(msgtype, tstr); PR_PROTO("%s:%d: WARNING: msg expected %s(0x%x), " "actual %s(0x%x) [msg=%s(0x%x), " "msgarg=%s(0x%x)]\n", proc, domid, xstr, xp->xt_msgtype, tstr, msgtype, mstr, msg, astr, msgarg); if (xfunc) (*xfunc)(domid, msgtype); return (-1); } } cfunc = xp->xt_trans[(int)o_xstate].t_check; if (cfunc && ((err = (*cfunc)(domid, mtp, xargs)) < 0)) { if (o_xstate != IDNXS_PEND) { IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); } if (xfunc) (*xfunc)(domid, msgtype); return (-1); } n_xstate = idn_next_xstate(o_xstate, err, msg); if (n_xstate == IDNXS_NIL) { PR_PROTO("%s:%d: WARNING: n_xstate = %s, msg = %s -> NIL " "<<<<<<<<<\n", proc, domid, idnxs_str[(int)n_xstate], mstr); if (xfunc) (*xfunc)(domid, msgtype); return (-1); } if (n_xstate != o_xstate) { IDN_XSTATE_TRANSITION(dp, n_xstate); } if (err) { if ((efunc = xp->xt_trans[(int)o_xstate].t_error) != NULL) (*efunc)(domid, mtp, xargs); } else if ((afunc = xp->xt_trans[(int)o_xstate].t_action) != NULL) { (*afunc)(domid, mtp, xargs); } if ((n_xstate == IDNXS_FINAL) && ((ffunc = xp->xt_final) != NULL)) (*ffunc)(domid); return (0); } /* * Entered and returns w/DLOCK & SYNC_LOCK held. */ static int idn_xstate_transfunc(int domid, void *transarg) { uint_t msg = (uint_t)(uintptr_t)transarg; uint_t token; procname_t proc = "idn_xstate_transfunc"; ASSERT(IDN_SYNC_IS_LOCKED()); switch (msg) { case IDNP_CON: DOMAINSET_ADD(idn.domset.ds_connected, domid); break; case IDNP_FIN: DOMAINSET_DEL(idn.domset.ds_connected, domid); break; default: PR_PROTO("%s:%d: ERROR: unknown msg (0x%x) <<<<<<<<\n", proc, domid, msg); return (0); } token = IDN_RETRY_TOKEN(domid, (msg == IDNP_CON) ? IDNRETRY_CON : IDNRETRY_FIN); if (msg == IDNP_CON) idn_retry_submit(idn_retry_con, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CON]); else idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); return (1); } /* * Entered and returns w/DLOCK & SYNC_LOCK held. */ static void idn_sync_enter(int domid, idn_synccmd_t cmd, domainset_t xset, domainset_t rset, int (*transfunc)(), void *transarg) { int z; idn_syncop_t *sp; idn_synczone_t *zp; procname_t proc = "idn_sync_enter"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); z = IDN_SYNC_GETZONE(cmd); ASSERT(z >= 0); zp = &idn.sync.sz_zone[z]; PR_SYNC("%s:%d: cmd=%s(%d), z=%d, xs=0x%x, rx=0x%x, cnt=%d\n", proc, domid, idnsync_str[cmd], cmd, z, xset, rset, zp->sc_cnt); sp = &idn_domain[domid].dsync; sp->s_domid = domid; sp->s_cmd = cmd; sp->s_msg = 0; sp->s_set_exp = xset; sp->s_set_rdy = rset; sp->s_transfunc = transfunc; sp->s_transarg = transarg; IDN_SYNC_QUERY_INIT(domid); sp->s_next = zp->sc_op; zp->sc_op = sp; zp->sc_cnt++; } /* * Entered and returns w/DLOCK & SYNC_LOCK held. */ void idn_sync_exit(int domid, idn_synccmd_t cmd) { int d, z, zone, tot_queries, tot_domains; idn_syncop_t *sp; idn_synczone_t *zp = NULL; procname_t proc = "idn_sync_exit"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); sp = &idn_domain[domid].dsync; z = IDN_SYNC_GETZONE(sp->s_cmd); zone = IDN_SYNC_GETZONE(cmd); PR_SYNC("%s:%d: cmd=%s(%d) (z=%d, zone=%d)\n", proc, domid, idnsync_str[cmd], cmd, z, zone); #ifdef DEBUG if (z != -1) { tot_queries = tot_domains = 0; for (d = 0; d < MAX_DOMAINS; d++) { int qv; if ((qv = sp->s_query[d]) > 0) { tot_queries += qv; tot_domains++; PR_SYNC("%s:%d: query_count = %d\n", proc, domid, qv); } } PR_SYNC("%s:%d: tot_queries = %d, tot_domaines = %d\n", proc, domid, tot_queries, tot_domains); } #endif /* DEBUG */ zp = (z != -1) ? &idn.sync.sz_zone[z] : NULL; if (zp) { idn_syncop_t **spp; for (spp = &zp->sc_op; *spp; spp = &((*spp)->s_next)) { if (*spp == sp) { *spp = sp->s_next; sp->s_next = NULL; zp->sc_cnt--; break; } } } sp->s_cmd = IDNSYNC_NIL; for (z = 0; z < IDN_SYNC_NUMZONE; z++) { idn_syncop_t **spp, **nspp; if ((zone != -1) && (z != zone)) continue; zp = &idn.sync.sz_zone[z]; for (spp = &zp->sc_op; *spp; spp = nspp) { sp = *spp; nspp = &sp->s_next; if (!DOMAIN_IN_SET(sp->s_set_exp, domid)) continue; DOMAINSET_DEL(sp->s_set_exp, domid); DOMAINSET_DEL(sp->s_set_rdy, domid); if ((sp->s_set_exp == sp->s_set_rdy) && sp->s_transfunc) { int delok; ASSERT(sp->s_domid != domid); PR_SYNC("%s:%d invoking transfunc " "for domain %d\n", proc, domid, sp->s_domid); delok = (*sp->s_transfunc)(sp->s_domid, sp->s_transarg); if (delok) { *spp = sp->s_next; sp->s_next = NULL; zp->sc_cnt--; nspp = spp; } } } } } /* * Entered and returns w/DLOCK & SYNC_LOCK held. */ static domainset_t idn_sync_register(int domid, idn_synccmd_t cmd, domainset_t ready_set, idn_syncreg_t regtype) { int z; idn_synczone_t *zp; idn_syncop_t *sp, **spp, **nspp; domainset_t query_set = 0, trans_set; procname_t proc = "idn_sync_register"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); if ((z = IDN_SYNC_GETZONE(cmd)) == -1) { PR_SYNC("%s:%d: ERROR: unexpected sync cmd(%d)\n", proc, domid, cmd); return (0); } /* * Find out what domains are in transition with respect * to given command. There will be no need to query * these folks. */ trans_set = IDN_SYNC_GETTRANS(cmd); zp = &idn.sync.sz_zone[z]; PR_SYNC("%s:%d: cmd=%s(%d), z=%d, rset=0x%x, " "regtype=%s(%d), sc_op=%s\n", proc, domid, idnsync_str[cmd], cmd, z, ready_set, idnreg_str[regtype], regtype, zp->sc_op ? idnsync_str[zp->sc_op->s_cmd] : "NULL"); for (spp = &zp->sc_op; *spp; spp = nspp) { sp = *spp; nspp = &sp->s_next; if (regtype == IDNSYNC_REG_NEW) { DOMAINSET_ADD(sp->s_set_exp, domid); PR_SYNC("%s:%d: adding new to %d (exp=0x%x)\n", proc, domid, sp->s_domid, sp->s_set_exp); } else if (regtype == IDNSYNC_REG_QUERY) { query_set |= ~sp->s_set_rdy & sp->s_set_exp; continue; } if (!DOMAIN_IN_SET(sp->s_set_exp, domid)) continue; if (!DOMAIN_IN_SET(ready_set, sp->s_domid)) { /* * Given domid doesn't have a desired * domain in his ready-set. We'll need * to query him again. */ DOMAINSET_ADD(query_set, domid); continue; } /* * If we reach here, then an expected domain * has marked its respective datapath to * sp->s_domid as down (i.e. in his ready_set). */ DOMAINSET_ADD(sp->s_set_rdy, domid); PR_SYNC("%s:%d: mark READY for domain %d " "(r=0x%x, x=0x%x)\n", proc, domid, sp->s_domid, sp->s_set_rdy, sp->s_set_exp); query_set |= ~sp->s_set_rdy & sp->s_set_exp; if (sp->s_set_exp == sp->s_set_rdy) { #ifdef DEBUG if (sp->s_msg == 0) { sp->s_msg = 1; PR_SYNC("%s:%d: >>>>>>>>>>> DOMAIN %d " "ALL CHECKED IN (0x%x)\n", proc, domid, sp->s_domid, sp->s_set_exp); } #endif /* DEBUG */ if ((sp->s_domid != domid) && sp->s_transfunc) { int delok; PR_SYNC("%s:%d invoking transfunc " "for domain %d\n", proc, domid, sp->s_domid); delok = (*sp->s_transfunc)(sp->s_domid, sp->s_transarg); if (delok) { *spp = sp->s_next; sp->s_next = NULL; zp->sc_cnt--; nspp = spp; } } } } PR_SYNC("%s:%d: trans_set = 0x%x, query_set = 0x%x -> 0x%x\n", proc, domid, trans_set, query_set, query_set & ~trans_set); query_set &= ~trans_set; return (query_set); } static void idn_sync_register_awol(int domid) { int z; idn_synccmd_t cmd = IDNSYNC_DISCONNECT; idn_synczone_t *zp; idn_syncop_t *sp; procname_t proc = "idn_sync_register_awol"; ASSERT(IDN_SYNC_IS_LOCKED()); if ((z = IDN_SYNC_GETZONE(cmd)) == -1) { PR_SYNC("%s:%d: ERROR: unexpected sync cmd(%d)\n", proc, domid, cmd); return; } zp = &idn.sync.sz_zone[z]; PR_SYNC("%s:%d: cmd=%s(%d), z=%d (domain %d = AWOL)\n", proc, domid, idnsync_str[cmd], cmd, z, domid); for (sp = zp->sc_op; sp; sp = sp->s_next) { idn_domain_t *dp; dp = &idn_domain[sp->s_domid]; if (dp->dfin == IDNFIN_FORCE_HARD) { DOMAINSET_ADD(sp->s_set_rdy, domid); PR_SYNC("%s:%d: adding new to %d (rdy=0x%x)\n", proc, domid, sp->s_domid, sp->s_set_rdy); } } } static void idn_link_established(void *arg) { int first_link; int domid, masterid; uint_t info = (uint_t)(uintptr_t)arg; first_link = (int)(info & 0xf0); domid = (int)(info & 0x0f); IDN_GLOCK_SHARED(); masterid = IDN_GET_MASTERID(); if ((masterid == IDN_NIL_DOMID) || (idn_domain[masterid].dstate != IDNDS_CONNECTED)) { /* * No point in doing this unless we're connected * to the master. */ if ((masterid != IDN_NIL_DOMID) && (idn.state == IDNGS_ONLINE)) { /* * As long as we're still online keep * trying. */ (void) timeout(idn_link_established, arg, 50); } IDN_GUNLOCK(); return; } IDN_GUNLOCK(); if (first_link && IDN_SLAB_PREALLOC) idn_prealloc_slab(IDN_SLAB_PREALLOC); /* * No guarantee, but it might save a little * time. */ if (idn_domain[domid].dstate == IDNDS_CONNECTED) { /* * Get the remote domain's dname. */ idn_send_nodename_req(domid); } /* * May have had some streams backed up waiting for * this connection. Prod them. */ rw_enter(&idn.struprwlock, RW_READER); mutex_enter(&idn.sipwenlock); idndl_wenable(NULL); mutex_exit(&idn.sipwenlock); rw_exit(&idn.struprwlock); } /* * Send the following chunk of data received from above onto * the IDN wire. This is raw data as far as the IDN driver * is concerned. * Returns: * IDNXMIT_LOOP - Msg handled in loopback and thus * still active (i.e. don't free). * IDNXMIT_OKAY - Data handled (freemsg). * IDNXMIT_DROP - Packet should be dropped. * IDNXMIT_RETRY - Packet should be requeued and retried. * IDNXMIT_REQUEUE - Packet should be requeued, but not * immediatetly retried. */ int idn_send_data(int dst_domid, idn_netaddr_t dst_netaddr, queue_t *wq, mblk_t *mp) { int pktcnt = 0; int msglen; int rv = IDNXMIT_OKAY; int xfersize = 0; caddr_t iobufp, iodatap; uchar_t *data_rptr; int cpuindex; int serrno; int channel; int retry_reclaim; idn_chansvr_t *csp = NULL; uint_t netports = 0; struct idnstr *stp; struct idn *sip; idn_domain_t *dp; struct ether_header *ehp; smr_pkthdr_t *hdrp; idn_msgtype_t mt; procname_t proc = "idn_send_data"; #ifdef DEBUG size_t orig_msglen = msgsize(mp); #endif /* DEBUG */ ASSERT(DB_TYPE(mp) == M_DATA); mt.mt_mtype = IDNP_DATA; mt.mt_atype = 0; mt.mt_cookie = 0; channel = (int)dst_netaddr.net.chan; msglen = msgdsize(mp); PR_DATA("%s:%d: (netaddr 0x%x) msgsize=%ld, msgdsize=%d\n", proc, dst_domid, dst_netaddr.netaddr, msgsize(mp), msglen); ASSERT(wq->q_ptr); stp = (struct idnstr *)wq->q_ptr; sip = stp->ss_sip; ASSERT(sip); if (msglen < 0) { /* * No data to send. That was easy! */ PR_DATA("%s:%d: BAD msg length (%d) (netaddr 0x%x)\n", proc, dst_domid, msglen, dst_netaddr.netaddr); return (IDNXMIT_DROP); } ASSERT(RW_READ_HELD(&stp->ss_rwlock)); if (dst_domid == IDN_NIL_DOMID) { cmn_err(CE_WARN, "IDN: 213: no destination specified " "(d=%d, c=%d, n=0x%x)", dst_domid, dst_netaddr.net.chan, dst_netaddr.net.netid); IDN_KSTAT_INC(sip, si_nolink); IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; goto nocando; } ehp = (struct ether_header *)mp->b_rptr; PR_DATA("%s:%d: destination channel = %d\n", proc, dst_domid, channel); #ifdef DEBUG { uchar_t echn; echn = (uchar_t) ehp->ether_shost.ether_addr_octet[IDNETHER_CHANNEL]; ASSERT((uchar_t)channel == echn); } #endif /* DEBUG */ ASSERT(msglen <= IDN_DATA_SIZE); dp = &idn_domain[dst_domid]; /* * Get reader lock. We hold for the duration * of the transfer so that our state doesn't * change during this activity. Note that since * we grab the reader lock, we can still permit * simultaneous tranfers from different threads * to the same domain. * Before we waste a bunch of time gathering locks, etc. * do a an unprotected check to make sure things are * semi-copesetic. If these values are in flux, * that's okay. */ if ((dp->dstate != IDNDS_CONNECTED) || (idn.state != IDNGS_ONLINE)) { IDN_KSTAT_INC(sip, si_linkdown); if (idn.state != IDNGS_ONLINE) { rv = IDNXMIT_REQUEUE; } else { IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; } goto nocando; } if (idn.chan_servers[channel].ch_send.c_checkin) { /* * Gotta bail, somethin' s'up. */ rv = IDNXMIT_REQUEUE; goto nocando; } csp = &idn.chan_servers[channel]; IDN_CHAN_LOCK_SEND(csp); if (dst_netaddr.net.netid == IDN_BROADCAST_ALLNETID) { /* * We're doing a broadcast. Need to set * up IDN netaddr's one at a time. * We set the ethernet destination to the same * instance as the sending address. The instance * numbers effectively represent subnets. */ dst_netaddr.net.netid = dp->dnetid; (void) idndl_domain_etheraddr(dst_domid, channel, &ehp->ether_dhost); if (dst_domid == idn.localid) { mblk_t *nmp; /* * If this is a broadcast and going to * the local domain, then we need to make * a private copy of the message since * the current one will be reused when * transmitting to other domains. */ PR_DATA("%s:%d: dup broadcast msg for local domain\n", proc, dst_domid); if ((nmp = copymsg(mp)) == NULL) { /* * Couldn't get a duplicate copy. */ IDN_CHAN_UNLOCK_SEND(csp); csp = NULL; IDN_KSTAT_INC(sip, si_allocbfail); IDN_KSTAT_INC(sip, si_noxmtbuf); rv = IDNXMIT_DROP; goto nocando; } mp = nmp; } } if (dp->dnetid != dst_netaddr.net.netid) { PR_DATA("%s:%d: dest netid (0x%x) != expected (0x%x)\n", proc, dst_domid, (uint_t)dst_netaddr.net.netid, (uint_t)dp->dnetid); IDN_CHAN_UNLOCK_SEND(csp); csp = NULL; IDN_KSTAT_INC(sip, si_nolink); IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; goto nocando; } if (dst_domid == idn.localid) { int lbrv; /* * Sending to our local domain! Loopback. * Note that idn_send_data_loop returning 0 * does not mean the message can now be freed. * We need to return (-1) so that caller doesn't * try to free mblk. */ IDN_CHAN_UNLOCK_SEND(csp); rw_exit(&stp->ss_rwlock); lbrv = idn_send_data_loopback(dst_netaddr, wq, mp); rw_enter(&stp->ss_rwlock, RW_READER); if (lbrv == 0) { return (IDNXMIT_LOOP); } else { IDN_KSTAT_INC(sip, si_macxmt_errors); return (IDNXMIT_DROP); } } if (dp->dstate != IDNDS_CONNECTED) { /* * Can't send data unless a link has already been * established with the target domain. Normally, * a user cannot set the remote netaddr unless a * link has already been established, however it * is possible the connection may have become * disconnected since that time. */ IDN_CHAN_UNLOCK_SEND(csp); csp = NULL; IDN_KSTAT_INC(sip, si_linkdown); IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; goto nocando; } /* * Need to make sure the channel is active and that the * domain to which we're sending is allowed to receive stuff. */ if (!IDN_CHANNEL_IS_SEND_ACTIVE(csp)) { int not_active; /* * See if we can activate channel. */ IDN_CHAN_UNLOCK_SEND(csp); not_active = idn_activate_channel(CHANSET(channel), IDNCHAN_OPEN); if (!not_active) { /* * Only grab the lock for a recheck if we were * able to activate the channel. */ IDN_CHAN_LOCK_SEND(csp); } /* * Verify channel still active now that we have the lock. */ if (not_active || !IDN_CHANNEL_IS_SEND_ACTIVE(csp)) { if (!not_active) { /* * Only need to drop the lock if it was * acquired while we thought we had * activated the channel. */ IDN_CHAN_UNLOCK_SEND(csp); } ASSERT(!IDN_CHAN_SEND_IS_LOCKED(csp)); /* * Damn! Must have went inactive during the window * before we regrabbed the send lock. Oh well, can't * spend all day doing this, bail out. Set csp to * NULL to prevent inprogress update at bottom. */ csp = NULL; /* * Channel is not active, should not be used. */ PR_DATA("%s:%d: dest channel %d NOT ACTIVE\n", proc, dst_domid, channel); IDN_KSTAT_INC(sip, si_linkdown); rv = IDNXMIT_REQUEUE; goto nocando; } ASSERT(IDN_CHAN_SEND_IS_LOCKED(csp)); } /* * If we made it here then the channel is active * Make sure the target domain is registered to receive stuff, * i.e. we're still linked. */ if (!IDN_CHAN_DOMAIN_IS_REGISTERED(csp, dst_domid)) { /* * If domain is not even registered with this channel * then we have no business being here. Doesn't matter * whether it's active or not. */ PR_DATA("%s:%d: domain not registered with channel %d\n", proc, dst_domid, channel); /* * Set csp to NULL to prevent in-progress update below. */ IDN_CHAN_UNLOCK_SEND(csp); csp = NULL; IDN_KSTAT_INC(sip, si_linkdown); IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; goto nocando; } IDN_CHAN_SEND_INPROGRESS(csp); IDN_CHAN_UNLOCK_SEND(csp); /* * Find a target cpu to send interrupt to if * it becomes necessary (i.e. remote channel * server is idle). */ cpuindex = dp->dcpuindex; /* * dcpuindex is atomically incremented, but other than * that is not well protected and that's okay. The * intention is to simply spread around the interrupts * at the destination domain, however we don't have to * anal about it. If we hit the same cpu multiple times * in a row that's okay, it will only be for a very short * period anyway before the cpuindex is incremented * to the next cpu. */ if (cpuindex < NCPU) { ATOMIC_INC(dp->dcpuindex); } if (dp->dcpuindex >= NCPU) dp->dcpuindex = 0; IDN_ASSIGN_DCPU(dp, cpuindex); #ifdef XXX_DLPI_UNFRIENDLY { ushort_t dstport = (ushort_t)dp->dcpu; /* * XXX * This is not DLPI friendly, but we need some way * of distributing our XDC interrupts to the cpus * on the remote domain in a relatively random fashion * while trying to remain constant for an individual * network connection. Don't want the target network * appl pinging around cpus thrashing the caches. * So, we'll pick target cpus based on the destination * TCP/IP port (socket). The (simple) alternative to * this is to simply send all messages destined for * particular domain to the same cpu (dcpu), but * will lower our bandwidth and introduce a lot of * contention on that target cpu. */ if (ehp->ether_type == ETHERTYPE_IP) { ipha_t *ipha; uchar_t *dstporta; int hdr_length; mblk_t *nmp = mp; uchar_t *rptr = mp->b_rptr + sizeof (struct ether_header); if (nmp->b_wptr <= rptr) { /* * Only the ethernet header was contained * in the first block. Check for the * next packet. */ if ((nmp = mp->b_cont) != NULL) rptr = nmp->b_rptr; } /* * If we still haven't found the IP header packet * then don't bother. Can't search forever. */ if (nmp && ((nmp->b_wptr - rptr) >= IP_SIMPLE_HDR_LENGTH)) { ipha = (ipha_t *)ALIGN32(rptr); ASSERT(DB_TYPE(mp) == M_DATA); hdr_length = IPH_HDR_LENGTH(ipha); switch (ipha->ipha_protocol) { case IPPROTO_UDP: case IPPROTO_TCP: /* * TCP/UDP Protocol Header (1st word) * 0 15,16 31 * ----------------------- * | src port | dst port | * ----------------------- */ dstporta = (uchar_t *)ipha + hdr_length; netports = *(uint_t *)dstporta; dstporta += 2; dstport = *(ushort_t *)dstporta; break; default: break; } } } IDN_ASSIGN_DCPU(dp, dstport); PR_DATA("%s:%d: (dstport %d) assigned %d\n", proc, dst_domid, (int)dstport, dp->dcpu); } #endif /* XXX_DLPI_UNFRIENDLY */ data_rptr = mp->b_rptr; ASSERT(dp->dcpu != IDN_NIL_DCPU); ASSERT(idn_domain[dst_domid].dmbox.m_send); retry_reclaim = 1; retry: if ((dp->dio >= IDN_RECLAIM_MIN) || dp->diowanted) { int reclaim_req; /* * Reclaim however many outstanding buffers * there are up to IDN_RECLAIM_MAX if it's set. */ reclaim_req = dp->diowanted ? -1 : IDN_RECLAIM_MAX ? MIN(dp->dio, IDN_RECLAIM_MAX) : dp->dio; (void) idn_reclaim_mboxdata(dst_domid, channel, reclaim_req); } if (dp->dio >= IDN_WINDOW_EMAX) { if (lock_try(&dp->diocheck)) { IDN_MSGTIMER_START(dst_domid, IDNP_DATA, 0, idn_msg_waittime[IDNP_DATA], &mt.mt_cookie); /* * We have exceeded the minimum window for * outstanding I/O buffers to this domain. * Need to start the MSG timer to check for * possible response from remote domain. * The remote domain may be hung. Send a * wakeup! Specify all channels for given * domain since we don't know precisely which * is backed up (dio is global). */ IDNXDC(dst_domid, &mt, (uint_t)dst_netaddr.net.chan, 0, 0, 0); } /* * Yikes! We have exceeded the maximum window * which means no more packets going to remote * domain until he frees some up. */ IDN_KSTAT_INC(sip, si_txmax); IDN_KSTAT_INC(sip, si_macxmt_errors); rv = IDNXMIT_DROP; goto nocando; } /* * Allocate a SMR I/O buffer and send it. */ if (msglen == 0) { /* * A zero length messages is effectively a signal * to just send an interrupt to the remote domain. */ IDN_MSGTIMER_START(dst_domid, IDNP_DATA, 0, idn_msg_waittime[IDNP_DATA], &mt.mt_cookie); IDNXDC(dst_domid, &mt, (uint_t)dst_netaddr.net.chan, 0, 0, 0); } for (; (msglen > 0) && mp; msglen -= xfersize) { int xrv; smr_offset_t bufoffset; #ifdef DEBUG int n_xfersize; #endif /* DEBUG */ ASSERT(msglen <= IDN_DATA_SIZE); xfersize = msglen; serrno = smr_buf_alloc(dst_domid, xfersize, &iobufp); if (serrno) { PR_DATA("%s:%d: failed to alloc SMR I/O buffer " "(serrno = %d)\n", proc, dst_domid, serrno); /* * Failure is either due to a timeout waiting * for the master to give us a slab, OR the * local domain exhausted its slab quota! * In either case we'll have to bail from * here and let higher layers decide what * to do. * We also could have had locking problems. * A negative serrno indicates we lost the lock * on dst_domid, so no need in dropping lock. */ if (lock_try(&dp->diowanted) && retry_reclaim) { /* * We were the first to acquire the * lock indicating that it wasn't * set on entry to idn_send_data. * So, let's go back and see if we * can't reclaim some buffers and * try again. * It's very likely diowanted will be * enough to prevent us from looping * on retrying here, however to protect * against the small window where a * race condition might exist, we use * the retry_reclaim flag so that we * don't retry more than once. */ retry_reclaim = 0; goto retry; } rv = (serrno > 0) ? serrno : -serrno; IDN_KSTAT_INC(sip, si_notbufs); IDN_KSTAT_INC(sip, si_noxmtbuf); /* MIB II */ switch (rv) { case ENOMEM: case EBUSY: case ENOLCK: case ETIMEDOUT: case EDQUOT: /* * These are all transient conditions * which should be recoverable over * time. */ rv = IDNXMIT_REQUEUE; break; default: rv = IDNXMIT_DROP; break; } goto nocando; } lock_clear(&dp->diowanted); hdrp = IDN_BUF2HDR(iobufp); bufoffset = (smr_offset_t)IDN_ALIGNPTR(sizeof (smr_pkthdr_t), data_rptr); /* * If the alignment of bufoffset took us pass the * length of a smr_pkthdr_t then we need to possibly * lower xfersize since it was calulated based on * a perfect alignment. However, if we're in DLPI * mode then shouldn't be necessary since the length * of the incoming packet (mblk) should have already * taken into consideration this possible adjustment. */ #ifdef DEBUG if (bufoffset != sizeof (smr_pkthdr_t)) PR_DATA("%s:%d: offset ALIGNMENT (%lu -> %u) " "(data_rptr = %p)\n", proc, dst_domid, sizeof (smr_pkthdr_t), bufoffset, data_rptr); n_xfersize = MIN(xfersize, (IDN_SMR_BUFSIZE - bufoffset)); if (xfersize != n_xfersize) { PR_DATA("%s:%d: xfersize ADJUST (%d -> %d)\n", proc, dst_domid, xfersize, n_xfersize); cmn_err(CE_WARN, "%s: ERROR (xfersize = %d, > " "bufsize(%d)-bufoffset(%d) = %d)", proc, xfersize, IDN_SMR_BUFSIZE, bufoffset, IDN_SMR_BUFSIZE - bufoffset); } #endif /* DEBUG */ xfersize = MIN(xfersize, (int)(IDN_SMR_BUFSIZE - bufoffset)); iodatap = IDN_BUF2DATA(iobufp, bufoffset); mp = idn_fill_buffer(iodatap, xfersize, mp, &data_rptr); hdrp->b_netaddr = dst_netaddr.netaddr; hdrp->b_netports = netports; hdrp->b_offset = bufoffset; hdrp->b_length = xfersize; hdrp->b_next = IDN_NIL_SMROFFSET; hdrp->b_rawio = 0; hdrp->b_cksum = IDN_CKSUM_PKT(hdrp); xrv = idn_send_mboxdata(dst_domid, sip, channel, iobufp); if (xrv) { /* * Reclaim packet. * Return error on this packet so it can be retried * (putbq). Note that it should be safe to assume * that this for-loop is only executed once when in * DLPI mode and so no need to worry about fractured * mblk packet. */ PR_DATA("%s:%d: DATA XFER to chan %d FAILED " "(ret=%d)\n", proc, dst_domid, channel, xrv); smr_buf_free(dst_domid, iobufp, xfersize); PR_DATA("%s:%d: (line %d) dec(dio) -> %d\n", proc, dst_domid, __LINE__, dp->dio); rv = IDNXMIT_DROP; IDN_KSTAT_INC(sip, si_macxmt_errors); goto nocando; } else { pktcnt++; /* * Packet will get freed on a subsequent send * when we reclaim buffers that the receivers * has finished consuming. */ } } #ifdef DEBUG if (pktcnt > 1) cmn_err(CE_WARN, "%s: ERROR: sent multi-pkts (%d), len = %ld", proc, pktcnt, orig_msglen); #endif /* DEBUG */ PR_DATA("%s:%d: SENT %d packets (%d @ 0x%x)\n", proc, dst_domid, pktcnt, dst_netaddr.net.chan, dst_netaddr.net.netid); IDN_CHAN_LOCK_SEND(csp); IDN_CHAN_SEND_DONE(csp); IDN_CHAN_UNLOCK_SEND(csp); return (IDNXMIT_OKAY); nocando: if (csp) { IDN_CHAN_LOCK_SEND(csp); IDN_CHAN_SEND_DONE(csp); IDN_CHAN_UNLOCK_SEND(csp); } if (rv == IDNXMIT_REQUEUE) { /* * Better kick off monitor to check when * it's ready to reenable the queues for * this channel. */ idn_xmit_monitor_kickoff(channel); } return (rv); } /* * Function to support local loopback testing of IDN driver. * Primarily geared towards measuring stream-head and IDN driver * overhead with respect to data messages. Setting idn_strhead_only * allows routine to focus on stream-head overhead by simply putting * the message straight to the 'next' queue of the destination * read-queue. Current implementation puts the message directly to * the read-queue thus sending the message right back to the IDN driver * as though the data came in off the wire. No need to worry about * any IDN layers attempting to ack data as that's normally handled * by idnh_recv_data. * * dst_netaddr = destination port-n-addr on local domain. * wq = write queue from whence message came. * mp = the (data-only) message. * * Returns 0 Indicates data handled. * errno EAGAIN indicates data can be retried. * Other errno's indicate failure to handle. */ static int idn_send_data_loopback(idn_netaddr_t dst_netaddr, queue_t *wq, mblk_t *mp) { register struct idnstr *stp; struct idn *sip; int rv = 0; procname_t proc = "idn_send_data_loopback"; if (dst_netaddr.net.netid != idn_domain[idn.localid].dnetid) { PR_DATA("%s: dst_netaddr.net.netid 0x%x != local 0x%x\n", proc, dst_netaddr.net.netid, idn_domain[idn.localid].dnetid); rv = EADDRNOTAVAIL; goto done; } stp = (struct idnstr *)wq->q_ptr; if (!stp || !stp->ss_rq) { rv = EDESTADDRREQ; goto done; } sip = stp->ss_sip; idndl_read(sip, mp); rv = 0; done: return (rv); } /* * Fill bufp with as much data as possible from the message pointed * to by mp up to size bytes. * Save our current read pointer in the variable parameter (data_rptrp) * so we know where to start on the next go around. Don't want to * bump the actual b_rptr in the mblk because the mblk may need to * be reused, e.g. broadcast. * Return the mblk pointer to the position we had to stop. */ static mblk_t * idn_fill_buffer(caddr_t bufp, int size, mblk_t *mp, uchar_t **data_rptrp) { int copysize; ASSERT(bufp && size); if (mp == NULL) return (NULL); while ((size > 0) && mp) { copysize = MIN(mp->b_wptr - (*data_rptrp), size); if (copysize > 0) { /* * If there's data to copy, do it. */ bcopy((*data_rptrp), bufp, copysize); (*data_rptrp) += copysize; bufp += copysize; size -= copysize; } if (mp->b_wptr <= (*data_rptrp)) { /* * If we emptied the mblk, then * move on to the next one. */ for (mp = mp->b_cont; mp && (mp->b_datap->db_type != M_DATA); mp = mp->b_cont) ; if (mp) *data_rptrp = mp->b_rptr; } } return (mp); } /* * Messages received here do NOT arrive on a stream, but are * instead handled via the idn_protocol_servers. This routine * is effectively the job processor for the protocol servers. */ static void idn_recv_proto(idn_protomsg_t *hp) { int domid, cpuid; int sync_lock = 0; idn_domain_t *dp; register uint_t mtype; register uint_t msgtype, acktype; idn_msgtype_t mt; ushort_t dcookie, tcookie; procname_t proc = "idn_recv_proto"; if (idn.state == IDNGS_IGNORE) { /* * Fault injection to simulate non-responsive domain. */ return; } domid = hp->m_domid; cpuid = hp->m_cpuid; msgtype = hp->m_msgtype; acktype = hp->m_acktype; dcookie = IDN_DCOOKIE(hp->m_cookie); tcookie = IDN_TCOOKIE(hp->m_cookie); /* * msgtype = Is the type of message we received, * e.g. nego, ack, nego+ack, etc. * * acktype = If we received a pure ack or nack * then this variable is set to the * type of message that was ack/nack'd. */ if ((mtype = msgtype & IDNP_MSGTYPE_MASK) == 0) { /* * Received a pure ack/nack. */ mtype = acktype & IDNP_MSGTYPE_MASK; } if (!VALID_MSGTYPE(mtype)) { PR_PROTO("%s:%d: ERROR: invalid message type (0x%x)\n", proc, domid, mtype); return; } if (!VALID_CPUID(cpuid)) { PR_PROTO("%s:%d: ERROR: invalid cpuid (%d)\n", proc, domid, cpuid); return; } /* * No pure data packets should reach this level. * Data+ack messages will reach here, but only * for the purpose of stopping the timer which * happens by default when this routine is called. */ ASSERT(msgtype != IDNP_DATA); /* * We should never receive a request from ourself, * except for commands in the case of broadcasts! */ if ((domid == idn.localid) && (mtype != IDNP_CMD)) { char str[15]; inum2str(hp->m_msgtype, str); cmn_err(CE_WARN, "IDN: 214: received message (%s[0x%x]) from self " "(domid %d)", str, hp->m_msgtype, domid); return; } IDN_SYNC_LOCK(); /* * Set a flag indicating whether we really need * SYNC-LOCK. We'll drop it in a little bit if * we really don't need it. */ switch (mtype) { case IDNP_CON: case IDNP_FIN: case IDNP_NEGO: sync_lock = 1; break; default: break; } dp = &idn_domain[domid]; IDN_DLOCK_EXCL(domid); /* * The only messages we do _not_ check the cookie are: * nego * nego+ack * fin - if received cookie is 0. * fin+ack - if received cookie is 0. * ack/fin - if received cookie is 0. * nack/fin - if received cookie is 0. */ if (((msgtype & IDNP_MSGTYPE_MASK) != IDNP_NEGO) && ((mtype != IDNP_FIN) || (dcookie && dp->dcookie_recv))) { if (dp->dcookie_recv != dcookie) { dp->dcookie_errcnt++; if (dp->dcookie_err == 0) { /* * Set cookie error to prevent a * possible flood of bogus cookies * and thus error messages. */ dp->dcookie_err = 1; cmn_err(CE_WARN, "IDN: 215: invalid cookie (0x%x) " "for message (0x%x) from domain %d", dcookie, hp->m_msgtype, domid); PR_PROTO("%s:%d: received cookie (0x%x), " "expected (0x%x) [errcnt = %d]\n", proc, domid, dcookie, dp->dcookie_recv, dp->dcookie_errcnt); } IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); return; } } dp->dcookie_err = 0; IDN_GLOCK_EXCL(); idn_clear_awol(domid); IDN_GUNLOCK(); if (!sync_lock) /* really don't need SYNC-LOCK past here */ IDN_SYNC_UNLOCK(); /* * Stop any timers that may have been outstanding for * this domain, for this particular message type. * Note that CFG timers are directly managed by * config recv/send code. */ if ((mtype != IDNP_CFG) && (msgtype & IDNP_ACKNACK_MASK) && tcookie) { IDN_MSGTIMER_STOP(domid, mtype, tcookie); } /* * Keep track of the last cpu to send us a message. * If the domain has not yet been assigned, we'll need * this cpuid in order to send back a respond. */ dp->dcpu_last = cpuid; mt.mt_mtype = (ushort_t)msgtype; mt.mt_atype = (ushort_t)acktype; mt.mt_cookie = tcookie; switch (mtype) { case IDNP_NEGO: idn_recv_nego(domid, &mt, hp->m_xargs, dcookie); break; case IDNP_CFG: idn_recv_config(domid, &mt, hp->m_xargs); break; case IDNP_CON: idn_recv_con(domid, &mt, hp->m_xargs); break; case IDNP_FIN: idn_recv_fin(domid, &mt, hp->m_xargs); break; case IDNP_CMD: idn_recv_cmd(domid, &mt, hp->m_xargs); break; case IDNP_DATA: ASSERT(msgtype & IDNP_ACKNACK_MASK); /* * When doing the fast track we simply process * possible nack error conditions. The actual * processing of the SMR data buffer is taken * care of in idnh_recv_dataack. When NOT doing * the fast track, we do all the processing here * in the protocol server. */ idn_recv_data(domid, &mt, hp->m_xargs); break; default: /* * Should be receiving 0 inum and 0 acknack. */ #ifdef DEBUG cmn_err(CE_PANIC, #else /* DEBUG */ cmn_err(CE_WARN, #endif /* DEBUG */ "IDN: 216: (0x%x)msgtype/(0x%x)acktype rcvd from " "domain %d", msgtype, acktype, domid); break; } IDN_DUNLOCK(domid); /* * All receiving routines are responsible for dropping drwlock. */ if (sync_lock) IDN_SYNC_UNLOCK(); } /* * Once the CONFIG state is hit we immediately blast out all * of our config info. This guarantees that the CONFIG state * effectively signifies that the sender has sent _all_ of * their config info. */ static void idn_send_config(int domid, int phase) { idn_domain_t *dp; int rv; clock_t cfg_waittime = idn_msg_waittime[IDNP_CFG]; procname_t proc = "idn_send_config"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; ASSERT(dp->dstate == IDNDS_CONFIG); if (phase == 1) { /* * Reset stuff in dtmp to 0: * dcfgphase * dcksum * dncfgitems * dmaxnets * dmboxpernet */ dp->dtmp = 0; } if (dp->dcfgsnddone) { if (!dp->dcfgrcvdone) { IDN_MSGTIMER_START(domid, IDNP_CFG, 0, cfg_waittime, NULL); } return; } IDN_DLOCK_SHARED(idn.localid); PR_PROTO("%s:%d: sending %s config (phase %d)\n", proc, domid, idn_domain[idn.localid].dvote.v.master ? "MASTER" : "SLAVE", phase); if (idn_domain[idn.localid].dvote.v.master) rv = idn_send_master_config(domid, phase); else rv = idn_send_slave_config(domid, phase); IDN_DUNLOCK(idn.localid); if (rv >= 0) { if (rv == 1) { dp->dcfgsnddone = 1; PR_PROTO("%s:%d: SEND config DONE\n", proc, domid); if (!dp->dcfgrcvdone) { IDN_MSGTIMER_START(domid, IDNP_CFG, 0, cfg_waittime, NULL); } } else { IDN_MSGTIMER_START(domid, IDNP_CFG, 0, cfg_waittime, NULL); } } } /* * Clear out the mailbox table. * NOTE: This routine touches the SMR. */ static void idn_reset_mboxtbl(idn_mboxtbl_t *mtp) { int qi; idn_mboxmsg_t *mp = &mtp->mt_queue[0]; qi = 0; do { mp[qi].ms_bframe = 0; mp[qi].ms_owner = 0; mp[qi].ms_flag = 0; IDN_MMBOXINDEX_INC(qi); } while (qi); } static int idn_get_mbox_config(int domid, int *mindex, smr_offset_t *mtable, smr_offset_t *mdomain) { idn_domain_t *dp, *ldp; dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(IDN_DLOCK_IS_SHARED(idn.localid)); ASSERT(IDN_GET_MASTERID() != IDN_NIL_DOMID); /* * Get SMR offset of receive mailbox assigned * to respective domain. If I'm a slave then * my dmbox.m_tbl will not have been assigned yet. * Instead of sending the actual offset I send * the master his assigned index. Since the * master knows what offset it will assign to * me he can determine his assigned (recv) mailbox * based on the offset and given index. The local * domain can also use this information once the * dmbox.m_tbl is received to properly assign the * correct mbox offset to the master. */ if (ldp->dmbox.m_tbl == NULL) { /* * Local domain has not yet been assigned a * (recv) mailbox table. This must be the * initial connection of this domain. */ ASSERT(dp->dvote.v.master && !ldp->dvote.v.master); ASSERT(mindex); *mindex = domid; } else { idn_mboxtbl_t *mtp; mtp = IDN_MBOXTBL_PTR(ldp->dmbox.m_tbl, domid); ASSERT(mdomain); *mdomain = IDN_ADDR2OFFSET(mtp); if (ldp->dvote.v.master) { /* * Need to calculate mailbox table to * assign to the given domain. Since * I'm the master his mailbox is in * the (all-domains) mailbox table. */ mtp = IDN_MBOXAREA_BASE(idn.mboxarea, domid); ASSERT(mtable); *mtable = IDN_ADDR2OFFSET(mtp); dp->dmbox.m_tbl = mtp; } } return (0); } /* * RETURNS: * 1 Unexpected/unnecessary phase. * 0 Successfully handled, timer needed. */ static int idn_send_master_config(int domid, int phase) { idn_cfgsubtype_t cfg_subtype; int rv = 0; idn_domain_t *dp, *ldp; idn_msgtype_t mt; int nmcadr; uint_t barpfn, larpfn; uint_t cpus_u32, cpus_l32; uint_t mcadr[3]; smr_offset_t mbox_table, mbox_domain; register int b, p, m; procname_t proc = "idn_send_master_config"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(IDN_DLOCK_IS_SHARED(idn.localid)); dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; ASSERT(dp->dstate == IDNDS_CONFIG); ASSERT(dp->dvote.v.master == 0); ASSERT(ldp->dvote.v.master == 1); mt.mt_mtype = IDNP_CFG; mt.mt_atype = 0; mt.mt_cookie = 0; m = 0; mcadr[0] = mcadr[1] = mcadr[2] = 0; cfg_subtype.val = 0; switch (phase) { case 1: mbox_table = mbox_domain = IDN_NIL_SMROFFSET; idn_get_mbox_config(domid, NULL, &mbox_table, &mbox_domain); /* * ---------------------------------------------------- * Send: SLABSIZE, DATAMBOX.DOMAIN, DATAMBOX.TABLE * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_SLAB); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_DATAMBOX, IDNCFGARG_DATAMBOX_DOMAIN); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_DATAMBOX, IDNCFGARG_DATAMBOX_TABLE); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; ASSERT(mbox_domain != IDN_NIL_SMROFFSET); ASSERT(mbox_table != IDN_NIL_SMROFFSET); PR_PROTO("%s:%d:%d: sending SLABSIZE (%d), " "DATAMBOX.DOMAIN (0x%x), DATAMBOX.TABLE (0x%x)\n", proc, domid, phase, IDN_SLAB_BUFCOUNT, mbox_domain, mbox_table); IDNXDC(domid, &mt, cfg_subtype.val, IDN_SLAB_BUFCOUNT, mbox_domain, mbox_table); break; case 2: barpfn = idn.smr.locpfn; larpfn = barpfn + (uint_t)btop(MB2B(IDN_SMR_SIZE)); /* * ---------------------------------------------------- * Send: NETID, BARLAR * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_NETID, 0); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_BARLAR, IDNCFGARG_BARLAR_BAR); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_BARLAR, IDNCFGARG_BARLAR_LAR); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending NETID (%d), " "BARPFN/LARPFN (0x%x/0x%x)\n", proc, domid, phase, ldp->dnetid, barpfn, larpfn); IDNXDC(domid, &mt, cfg_subtype.val, (uint_t)ldp->dnetid, barpfn, larpfn); break; case 3: nmcadr = ldp->dhw.dh_nmcadr; cpus_u32 = UPPER32_CPUMASK(ldp->dcpuset); cpus_l32 = LOWER32_CPUMASK(ldp->dcpuset); /* * ---------------------------------------------------- * Send: CPUSET, NMCADR * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_CPUSET, IDNCFGARG_CPUSET_UPPER); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_CPUSET, IDNCFGARG_CPUSET_LOWER); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_NMCADR, 0); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending CPUSET (0x%x.%x), NMCADR (%d)\n", proc, domid, phase, cpus_u32, cpus_l32, nmcadr); IDNXDC(domid, &mt, cfg_subtype.val, cpus_u32, cpus_l32, nmcadr); break; case 4: /* * ---------------------------------------------------- * Send: BOARDSET, MTU, BUFSIZE * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_BOARDSET, 0); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_MTU); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_BUF); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending BOARDSET (0x%x), MTU (0x%lx), " "BUFSIZE (0x%x)\n", proc, domid, phase, ldp->dhw.dh_boardset, IDN_MTU, IDN_SMR_BUFSIZE); IDNXDC(domid, &mt, cfg_subtype.val, ldp->dhw.dh_boardset, IDN_MTU, IDN_SMR_BUFSIZE); break; case 5: /* * ---------------------------------------------------- * Send: MAXNETS, MBOXPERNET, CKSUM * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_DATASVR, IDNCFGARG_DATASVR_MAXNETS); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_DATASVR, IDNCFGARG_DATASVR_MBXPERNET); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_OPTIONS, IDNCFGARG_CHECKSUM); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending MAXNETS (%d), " "MBOXPERNET (%d), CKSUM (%d)\n", proc, domid, phase, IDN_MAX_NETS, IDN_MBOX_PER_NET, IDN_CHECKSUM); IDNXDC(domid, &mt, cfg_subtype.val, IDN_MAX_NETS, IDN_MBOX_PER_NET, IDN_CHECKSUM); break; case 6: /* * ---------------------------------------------------- * Send: NWRSIZE (piggyback on MCADRs) * ---------------------------------------------------- */ cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_NWR); mcadr[0] = IDN_NWR_SIZE; m = 1; /*FALLTHROUGH*/ default: /* case 7 and above */ /* * ---------------------------------------------------- * Send: MCADR's * ---------------------------------------------------- * First need to figure how many we've already sent * based on what phase of CONFIG we're in. * ---------------------------------------------------- */ if (phase > 6) { p = ((phase - 7) * 3) + 2; for (b = 0; (b < MAX_BOARDS) && (p > 0); b++) if (ldp->dhw.dh_mcadr[b]) p--; } else { b = 0; } for (; (b < MAX_BOARDS) && (m < 3); b++) { if (ldp->dhw.dh_mcadr[b] == 0) continue; mcadr[m] = ldp->dhw.dh_mcadr[b]; cfg_subtype.param.p[m] = IDN_CFGPARAM(IDNCFG_MCADR, b); m++; } if (m > 0) { if (phase == 6) { PR_PROTO("%s:%d:%d: sending NWRSIZE (%d), " "MCADRs (0x%x, 0x%x)\n", proc, domid, phase, mcadr[0], mcadr[1], mcadr[2]); } else { PR_PROTO("%s:%d:%d: sending MCADRs " "(0x%x, 0x%x, 0x%x)\n", proc, domid, phase, mcadr[0], mcadr[1], mcadr[2]); } cfg_subtype.info.num = m; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; IDNXDC(domid, &mt, cfg_subtype.val, mcadr[0], mcadr[1], mcadr[2]); } else { rv = 1; } break; } return (rv); } /* * RETURNS: * 1 Unexpected/unnecessary phase. * 0 Successfully handled. */ static int idn_send_slave_config(int domid, int phase) { idn_cfgsubtype_t cfg_subtype; int rv = 0; idn_domain_t *dp, *ldp; smr_offset_t mbox_domain; idn_msgtype_t mt; int mbox_index; uint_t cpus_u32, cpus_l32; procname_t proc = "idn_send_slave_config"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(IDN_DLOCK_IS_SHARED(idn.localid)); mt.mt_mtype = IDNP_CFG; mt.mt_atype = 0; dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; ASSERT(dp->dstate == IDNDS_CONFIG); ASSERT(ldp->dvote.v.master == 0); switch (phase) { case 1: mbox_index = IDN_NIL_DOMID; mbox_domain = IDN_NIL_SMROFFSET; idn_get_mbox_config(domid, &mbox_index, NULL, &mbox_domain); /* * ---------------------------------------------------- * Send: DATAMBOX.DOMAIN or DATAMBOX.INDEX, * DATASVR.MAXNETS, DATASVR.MBXPERNET * ---------------------------------------------------- */ cfg_subtype.val = 0; if (mbox_index == IDN_NIL_DOMID) { ASSERT(mbox_domain != IDN_NIL_SMROFFSET); cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_DATAMBOX, IDNCFGARG_DATAMBOX_DOMAIN); } else { /* * Should only be sending Index to * the master and not another slave. */ ASSERT(dp->dvote.v.master); ASSERT(mbox_domain == IDN_NIL_SMROFFSET); cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_DATAMBOX, IDNCFGARG_DATAMBOX_INDEX); } cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_DATASVR, IDNCFGARG_DATASVR_MAXNETS); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_DATASVR, IDNCFGARG_DATASVR_MBXPERNET); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending DATAMBOX.%s (0x%x), " "MAXNETS (%d), MBXPERNET (%d)\n", proc, domid, phase, (IDN_CFGPARAM_ARG(cfg_subtype.param.p[0]) == IDNCFGARG_DATAMBOX_INDEX) ? "INDEX" : "DOMAIN", (mbox_index == IDN_NIL_DOMID) ? mbox_domain : mbox_index, IDN_MAX_NETS, IDN_MBOX_PER_NET); IDNXDC(domid, &mt, cfg_subtype.val, ((mbox_index == IDN_NIL_DOMID) ? mbox_domain : mbox_index), IDN_MAX_NETS, IDN_MBOX_PER_NET); break; case 2: cpus_u32 = UPPER32_CPUMASK(ldp->dcpuset); cpus_l32 = LOWER32_CPUMASK(ldp->dcpuset); /* * ---------------------------------------------------- * Send: NETID, CPUSET * ---------------------------------------------------- */ cfg_subtype.val = 0; cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_NETID, 0); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_CPUSET, IDNCFGARG_CPUSET_UPPER); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_CPUSET, IDNCFGARG_CPUSET_LOWER); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending NETID (%d), " "CPUSET (0x%x.%x)\n", proc, domid, phase, ldp->dnetid, cpus_u32, cpus_l32); IDNXDC(domid, &mt, cfg_subtype.val, (uint_t)ldp->dnetid, cpus_u32, cpus_l32); break; case 3: /* * ---------------------------------------------------- * Send: BOARDSET, MTU, BUFSIZE * ---------------------------------------------------- */ cfg_subtype.val = 0; cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_BOARDSET, 0); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_MTU); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_BUF); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending BOARDSET (0x%x), MTU (0x%lx), " "BUFSIZE (0x%x)\n", proc, domid, phase, ldp->dhw.dh_boardset, IDN_MTU, IDN_SMR_BUFSIZE); IDNXDC(domid, &mt, cfg_subtype.val, ldp->dhw.dh_boardset, IDN_MTU, IDN_SMR_BUFSIZE); break; case 4: /* * ---------------------------------------------------- * Send: SLABSIZE, OPTIONS.CHECKSUM, NWR_SIZE * ---------------------------------------------------- */ cfg_subtype.val = 0; cfg_subtype.param.p[0] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_SLAB); cfg_subtype.param.p[1] = IDN_CFGPARAM(IDNCFG_OPTIONS, IDNCFGARG_CHECKSUM); cfg_subtype.param.p[2] = IDN_CFGPARAM(IDNCFG_SIZE, IDNCFGARG_SIZE_NWR); cfg_subtype.info.num = 3; cfg_subtype.info.phase = phase; dp->dcfgphase = phase; PR_PROTO("%s:%d:%d: sending SLABSIZE (%d), CKSUM (%d), " "NWRSIZE (%d)\n", proc, domid, phase, IDN_SLAB_BUFCOUNT, IDN_CHECKSUM, IDN_NWR_SIZE); IDNXDC(domid, &mt, cfg_subtype.val, IDN_SLAB_BUFCOUNT, IDN_CHECKSUM, IDN_NWR_SIZE); break; default: rv = 1; break; } return (rv); } #define CFG_FATAL ((uint_t)-1) /* reset link */ #define CFG_CONTINUE 0x0000 /* looking for more */ #define CFG_DONE 0x0001 /* got everything expected */ #define CFG_ERR_MTU 0x0002 #define CFG_ERR_BUF 0x0004 #define CFG_ERR_SLAB 0x0008 #define CFG_ERR_NWR 0x0010 #define CFG_ERR_NETS 0x0020 #define CFG_ERR_MBOX 0x0040 #define CFG_ERR_NMCADR 0x0080 #define CFG_ERR_MCADR 0x0100 #define CFG_ERR_CKSUM 0x0200 #define CFG_ERR_SMR 0x0400 #define CFG_MAX_ERRORS 16 #define CFGERR2IDNKERR(ce) \ (((ce) & CFG_ERR_MTU) ? IDNKERR_CONFIG_MTU : \ ((ce) & CFG_ERR_BUF) ? IDNKERR_CONFIG_BUF : \ ((ce) & CFG_ERR_SLAB) ? IDNKERR_CONFIG_SLAB : \ ((ce) & CFG_ERR_NWR) ? IDNKERR_CONFIG_NWR : \ ((ce) & CFG_ERR_NETS) ? IDNKERR_CONFIG_NETS : \ ((ce) & CFG_ERR_MBOX) ? IDNKERR_CONFIG_MBOX : \ ((ce) & CFG_ERR_NMCADR) ? IDNKERR_CONFIG_NMCADR : \ ((ce) & CFG_ERR_MCADR) ? IDNKERR_CONFIG_MCADR : \ ((ce) & CFG_ERR_CKSUM) ? IDNKERR_CONFIG_CKSUM : \ ((ce) & CFG_ERR_SMR) ? IDNKERR_CONFIG_SMR : 0) #define CFGERR2FINARG(ce) \ (((ce) & CFG_ERR_MTU) ? IDNFIN_ARG_CFGERR_MTU : \ ((ce) & CFG_ERR_BUF) ? IDNFIN_ARG_CFGERR_BUF : \ ((ce) & CFG_ERR_SLAB) ? IDNFIN_ARG_CFGERR_SLAB : \ ((ce) & CFG_ERR_NWR) ? IDNFIN_ARG_CFGERR_NWR : \ ((ce) & CFG_ERR_NETS) ? IDNFIN_ARG_CFGERR_NETS : \ ((ce) & CFG_ERR_MBOX) ? IDNFIN_ARG_CFGERR_MBOX : \ ((ce) & CFG_ERR_NMCADR) ? IDNFIN_ARG_CFGERR_NMCADR : \ ((ce) & CFG_ERR_MCADR) ? IDNFIN_ARG_CFGERR_MCADR : \ ((ce) & CFG_ERR_CKSUM) ? IDNFIN_ARG_CFGERR_CKSUM : \ ((ce) & CFG_ERR_SMR) ? IDNFIN_ARG_CFGERR_SMR : IDNFIN_ARG_NONE) /* * Called when some CFG messages arrive. We use dncfgitems to count the * total number of items received so far since we'll receive multiple CFG * messages during the CONFIG phase. Note that dncfgitems is initialized * in idn_send_config. */ static void idn_recv_config(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp->mt_mtype; uint_t rv, rv_expected, rv_actual; int pnum; int phase; register int p; register int c; idn_mainmbox_t *mmp; register uint_t subtype, subtype_arg; idn_domain_t *dp; int index; idn_domain_t *ldp = &idn_domain[idn.localid]; idn_mboxtbl_t *mbtp; idn_cfgsubtype_t cfg_subtype; idn_xdcargs_t cfg_arg; idn_msgtype_t mt; idnsb_error_t idnerr; procname_t proc = "idn_recv_config"; ASSERT(domid != idn.localid); GET_XARGS(xargs, &cfg_subtype.val, &cfg_arg[0], &cfg_arg[1], &cfg_arg[2]); cfg_arg[3] = 0; dp = &idn_domain[domid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (dp->dstate != IDNDS_CONFIG) { /* * Not ready to receive config info. * Drop whatever he sent us. Let the * timer continue and timeout if needed. */ PR_PROTO("%s:%d: WARNING state(%s) != CONFIG\n", proc, domid, idnds_str[dp->dstate]); return; } if ((msg & IDNP_ACKNACK_MASK) || dp->dcfgsnddone) { IDN_MSGTIMER_STOP(domid, IDNP_CFG, 0); } if (msg & IDNP_ACKNACK_MASK) { /* * ack/cfg */ phase = GET_XARGS_CFG_PHASE(xargs); PR_PROTO("%s:%d: received ACK for CFG phase %d\n", proc, domid, phase); if (phase != (int)dp->dcfgphase) { /* * Phase is not what we were * expecting. Something got lost * in the shuffle. Restart the * timer and let it timeout if necessary * and reestablish the connection. */ IDN_MSGTIMER_START(domid, IDNP_CFG, dp->dcfgphase, idn_msg_waittime[IDNP_CFG], NULL); } else { idn_send_config(domid, phase + 1); if (dp->dcfgsnddone && dp->dcfgrcvdone) { IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); if (dp->dstate == IDNDS_CONFIG) { dp->dxp = &xphase_con; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); bzero(xargs, sizeof (xargs)); idn_xphase_transition(domid, NULL, xargs); } IDN_SYNC_UNLOCK(); } } return; } pnum = (int)cfg_subtype.info.num; phase = (int)cfg_subtype.info.phase; for (p = 0; p < pnum; p++) { int board; #ifdef DEBUG uint_t val; char *str; val = 0; str = NULL; #define RCVCFG(s, v) { str = (s); val = (v); } #else #define RCVCFG(s, v) {} #endif /* DEBUG */ subtype = IDN_CFGPARAM_TYPE(cfg_subtype.param.p[p]); subtype_arg = IDN_CFGPARAM_ARG(cfg_subtype.param.p[p]); switch (subtype) { case IDNCFG_BARLAR: IDN_GLOCK_EXCL(); switch (subtype_arg) { case IDNCFGARG_BARLAR_BAR: if (idn.smr.rempfn == PFN_INVALID) { idn.smr.rempfn = (pfn_t)cfg_arg[p]; dp->dncfgitems++; RCVCFG("BARLAR_BAR", cfg_arg[p]); } break; case IDNCFGARG_BARLAR_LAR: if (idn.smr.rempfnlim == PFN_INVALID) { idn.smr.rempfnlim = (pfn_t)cfg_arg[p]; dp->dncfgitems++; RCVCFG("BARLAR_LAR", cfg_arg[p]); } break; default: cmn_err(CE_WARN, "IDN 217: unknown CFGARG type (%d) " "from domain %d", subtype_arg, domid); break; } IDN_GUNLOCK(); break; case IDNCFG_MCADR: board = subtype_arg; if ((board >= 0) && (board < MAX_BOARDS) && (dp->dhw.dh_mcadr[board] == 0)) { dp->dhw.dh_mcadr[board] = cfg_arg[p]; dp->dncfgitems++; RCVCFG("MCADR", cfg_arg[p]); } break; case IDNCFG_NMCADR: if (dp->dhw.dh_nmcadr == 0) { dp->dhw.dh_nmcadr = cfg_arg[p]; dp->dncfgitems++; RCVCFG("NMCADR", cfg_arg[p]); } break; case IDNCFG_CPUSET: switch (subtype_arg) { case IDNCFGARG_CPUSET_UPPER: { cpuset_t tmpset; MAKE64_CPUMASK(tmpset, cfg_arg[p], 0); CPUSET_OR(dp->dcpuset, tmpset); dp->dncfgitems++; RCVCFG("CPUSET_UPPER", cfg_arg[p]); break; } case IDNCFGARG_CPUSET_LOWER: { cpuset_t tmpset; MAKE64_CPUMASK(tmpset, 0, cfg_arg[p]); CPUSET_OR(dp->dcpuset, tmpset); dp->dncfgitems++; RCVCFG("CPUSET_LOWER", cfg_arg[p]); break; } default: ASSERT(0); break; } break; case IDNCFG_NETID: if (dp->dnetid == (ushort_t)-1) { dp->dnetid = (ushort_t)cfg_arg[p]; dp->dncfgitems++; RCVCFG("NETID", cfg_arg[p]); } break; case IDNCFG_BOARDSET: if ((dp->dhw.dh_boardset & cfg_arg[p]) == dp->dhw.dh_boardset) { /* * Boardset better include what we * already know about. */ dp->dhw.dh_boardset = cfg_arg[p]; dp->dncfgitems++; RCVCFG("BOARDSET", cfg_arg[p]); } break; case IDNCFG_SIZE: switch (subtype_arg) { case IDNCFGARG_SIZE_MTU: if (dp->dmtu == 0) { dp->dmtu = cfg_arg[p]; dp->dncfgitems++; RCVCFG("MTU", cfg_arg[p]); } break; case IDNCFGARG_SIZE_BUF: if (dp->dbufsize == 0) { dp->dbufsize = cfg_arg[p]; dp->dncfgitems++; RCVCFG("BUFSIZE", cfg_arg[p]); } break; case IDNCFGARG_SIZE_SLAB: if (dp->dslabsize == 0) { dp->dslabsize = (short)cfg_arg[p]; dp->dncfgitems++; RCVCFG("SLABSIZE", cfg_arg[p]); } break; case IDNCFGARG_SIZE_NWR: if (dp->dnwrsize == 0) { dp->dnwrsize = (short)cfg_arg[p]; dp->dncfgitems++; RCVCFG("NWRSIZE", cfg_arg[p]); } break; default: ASSERT(0); break; } break; case IDNCFG_DATAMBOX: switch (subtype_arg) { case IDNCFGARG_DATAMBOX_TABLE: if (ldp->dmbox.m_tbl || !dp->dvote.v.master || !VALID_NWROFFSET(cfg_arg[p], 4)) { /* * Only a master should be * sending us a datambox table. */ break; } IDN_DLOCK_EXCL(idn.localid); ldp->dmbox.m_tbl = (idn_mboxtbl_t *) IDN_OFFSET2ADDR(cfg_arg[p]); IDN_DUNLOCK(idn.localid); dp->dncfgitems++; RCVCFG("DATAMBOX.TABLE", cfg_arg[p]); break; case IDNCFGARG_DATAMBOX_DOMAIN: if (dp->dmbox.m_send->mm_smr_mboxp || !VALID_NWROFFSET(cfg_arg[p], 4)) break; mbtp = (idn_mboxtbl_t *) IDN_OFFSET2ADDR(cfg_arg[p]); mmp = dp->dmbox.m_send; for (c = 0; c < IDN_MAX_NETS; c++) { mutex_enter(&mmp[c].mm_mutex); mmp[c].mm_smr_mboxp = mbtp; mutex_exit(&mmp[c].mm_mutex); IDN_MBOXTBL_PTR_INC(mbtp); } if (c <= 0) break; dp->dncfgitems++; RCVCFG("DATAMBOX.DOMAIN", cfg_arg[p]); break; case IDNCFGARG_DATAMBOX_INDEX: if (!ldp->dvote.v.master || dp->dmbox.m_send->mm_smr_mboxp) { /* * If I'm not the master then * I can't handle processing a * mailbox index. * OR, if I already have the send * mailbox, I'm done with this * config item. */ break; } ASSERT(dp->dmbox.m_tbl); index = (int)cfg_arg[p]; /* * The given index is the local domain's * index into the remote domain's mailbox * table that contains the mailbox that * remote domain wants the local domain to * use as the send mailbox for messages * destined for the remote domain. * I.e. from the remote domain's * perspective, this is his receive * mailbox. */ mbtp = IDN_MBOXTBL_PTR(dp->dmbox.m_tbl, index); mmp = dp->dmbox.m_send; for (c = 0; c < IDN_MAX_NETS; c++) { mutex_enter(&mmp[c].mm_mutex); mmp[c].mm_smr_mboxp = mbtp; mutex_exit(&mmp[c].mm_mutex); IDN_MBOXTBL_PTR_INC(mbtp); } if (c <= 0) break; dp->dncfgitems++; RCVCFG("DATAMBOX.INDEX", cfg_arg[p]); break; default: ASSERT(0); break; } break; case IDNCFG_DATASVR: switch (subtype_arg) { case IDNCFGARG_DATASVR_MAXNETS: if (dp->dmaxnets) break; dp->dmaxnets = (uint_t)(cfg_arg[p] & 0x3f); dp->dncfgitems++; RCVCFG("DATASVR.MAXNETS", cfg_arg[p]); break; case IDNCFGARG_DATASVR_MBXPERNET: if (dp->dmboxpernet) break; dp->dmboxpernet = (uint_t)(cfg_arg[p] & 0x1ff); dp->dncfgitems++; RCVCFG("DATASVR.MBXPERNET", cfg_arg[p]); break; default: ASSERT(0); break; } break; case IDNCFG_OPTIONS: switch (subtype_arg) { case IDNCFGARG_CHECKSUM: if (dp->dcksum) break; if ((cfg_arg[p] & 0xff) == 0) dp->dcksum = 1; /* off */ else dp->dcksum = 2; /* on */ dp->dncfgitems++; RCVCFG("OPTIONS.CHECKSUM", cfg_arg[p]); break; default: ASSERT(0); break; } default: break; } #ifdef DEBUG PR_PROTO("%s:%d: received %s (0x%x)\n", proc, domid, str ? str : "", val); #endif /* DEBUG */ } mt.mt_mtype = IDNP_ACK; mt.mt_atype = IDNP_CFG; mt.mt_cookie = mtp->mt_cookie; CLR_XARGS(cfg_arg); SET_XARGS_CFG_PHASE(cfg_arg, phase); idn_send_acknack(domid, &mt, cfg_arg); rv_expected = rv_actual = 0; if (dp->dvote.v.master == 0) { /* * Remote domain is a slave, check if we've received * all that we were expecting, and if so transition to * the next state. */ rv = idn_check_slave_config(domid, &rv_expected, &rv_actual); } else { /* * Remote domain is a master, check if this slave has * received all that it was expecting, and if so * transition to the next state. */ rv = idn_check_master_config(domid, &rv_expected, &rv_actual); } switch (rv) { case CFG_DONE: /* * All config info received that was expected, wrap up. */ if (!idn_recv_config_done(domid) && dp->dvote.v.master) { IDN_DLOCK_EXCL(idn.localid); ldp->dvote.v.connected = 1; IDN_DUNLOCK(idn.localid); } break; case CFG_CONTINUE: /* * If we're not done sending our own config, then * there's no need to set a timer since one will * automatically be set when we send a config * message waiting for an acknowledgement. */ if (dp->dcfgsnddone) { /* * We haven't yet received all the config * information we were expecting. Need to * restart CFG timer if we've sent everything.. */ IDN_MSGTIMER_START(domid, IDNP_CFG, 0, idn_msg_waittime[IDNP_CFG], NULL); } break; case CFG_FATAL: /* * Fatal error occurred during config exchange. * We need to shutdown connection in this * case, so initiate a (non-relink) FIN. * so let's get the show on the road. */ IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); /* * If the state has changed from CONFIG * then somebody else has taken over * control of this domain so we can just * bail out. */ if (dp->dstate == IDNDS_CONFIG) { INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EPROTO); SET_IDNKERR_IDNERR(&idnerr, IDNKERR_CONFIG_FATAL); SET_IDNKERR_PARAM0(&idnerr, domid); idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); /* * Keep this guy around so we can try again. */ DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); idn_disconnect(domid, IDNFIN_NORMAL, IDNFIN_ARG_CFGERR_FATAL, IDNFIN_SYNC_NO); } IDN_SYNC_UNLOCK(); break; default: /* parameter conflict */ IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); if (dp->dstate != IDNDS_CONFIG) { /* * Hmmm...changed in the short period * we had dropped the lock, oh well. */ IDN_SYNC_UNLOCK(); break; } c = 0; for (p = 0; p < CFG_MAX_ERRORS; p++) if (rv & (1 << p)) c++; INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EINVAL); SET_IDNKERR_PARAM0(&idnerr, domid); if (c > 1) { SET_IDNKERR_IDNERR(&idnerr, IDNKERR_CONFIG_MULTIPLE); SET_IDNKERR_PARAM1(&idnerr, c); } else { SET_IDNKERR_IDNERR(&idnerr, CFGERR2IDNKERR(rv)); SET_IDNKERR_PARAM1(&idnerr, rv_expected); SET_IDNKERR_PARAM2(&idnerr, rv_actual); } /* * Any parameter conflicts are grounds for dismissal. */ if (idn.domset.ds_connected == 0) { domainset_t domset; /* * We have no other connections yet. * We must blow out of here completely * unless we have relinkers left from * a RECONFIG. */ IDN_GLOCK_EXCL(); domset = ~idn.domset.ds_relink; if (idn.domset.ds_relink == 0) { IDN_GSTATE_TRANSITION(IDNGS_DISCONNECT); } domset &= ~idn.domset.ds_hitlist; IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); IDN_GUNLOCK(); IDN_DUNLOCK(domid); DOMAINSET_DEL(domset, idn.localid); idn_update_op(IDNOP_ERROR, DOMAINSET_ALL, &idnerr); PR_HITLIST("%s:%d: unlink_domainset(%x) due to " "CFG error (relink=%x, hitlist=%x)\n", proc, domid, domset, idn.domset.ds_relink, idn.domset.ds_hitlist); idn_unlink_domainset(domset, IDNFIN_NORMAL, CFGERR2FINARG(rv), IDNFIN_OPT_UNLINK, BOARDSET_ALL); IDN_SYNC_UNLOCK(); IDN_DLOCK_EXCL(domid); } else { PR_HITLIST("%s:%d: idn_disconnect(%d) due to CFG " "error (conn=%x, relink=%x, hitlist=%x)\n", proc, domid, domid, idn.domset.ds_connected, idn.domset.ds_relink, idn.domset.ds_hitlist); /* * If we have other connections then * we're only going to blow away this * single connection. */ idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); DOMAINSET_DEL(idn.domset.ds_relink, domid); idn_disconnect(domid, IDNFIN_NORMAL, CFGERR2FINARG(rv), IDNFIN_SYNC_NO); IDN_SYNC_UNLOCK(); } break; } } /* * Called by master or slave which expects exactly the following * with respect to config info received from a SLAVE: * IDNCFG_CPUSET * IDNCFG_NETID * IDNCFG_BOARDSET * IDNCFG_SIZE (MTU, BUF, SLAB, NWR) * IDNCFG_DATAMBOX (DOMAIN or INDEX if caller is master) * IDNCFG_DATASVR (MAXNETS, MBXPERNET) * IDNCFG_OPTIONS (CHECKSUM) */ static uint_t idn_check_slave_config(int domid, uint_t *exp, uint_t *act) { uint_t rv = 0; idn_domain_t *ldp, *dp; procname_t proc = "idn_check_slave_config"; dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; ASSERT(domid != idn.localid); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(dp->dstate == IDNDS_CONFIG); PR_PROTO("%s:%d: number received %d, number expected %d\n", proc, domid, (int)dp->dncfgitems, IDN_SLAVE_NCFGITEMS); if ((int)dp->dncfgitems < IDN_SLAVE_NCFGITEMS) return (CFG_CONTINUE); if ((dp->dnetid == (ushort_t)-1) || CPUSET_ISNULL(dp->dcpuset) || (dp->dhw.dh_boardset == 0) || (dp->dmbox.m_send->mm_smr_mboxp == NULL) || (dp->dmaxnets == 0) || (dp->dmboxpernet == 0) || (dp->dcksum == 0) || (dp->dmtu == 0) || (dp->dbufsize == 0) || (dp->dslabsize == 0) || (dp->dnwrsize == 0)) { /* * We received our IDN_SLAVE_NCFGITEMS config items, * but not all what we were expecting! Gotta nack and * close connection. */ cmn_err(CE_WARN, "IDN: 218: missing some required config items from " "domain %d", domid); rv = CFG_FATAL; goto done; } if (!valid_mtu(dp->dmtu)) { cmn_err(CE_WARN, "IDN: 219: remote domain %d MTU (%d) invalid " "(local.mtu = %d)", dp->domid, dp->dmtu, ldp->dmtu); *exp = (uint_t)ldp->dmtu; *act = (uint_t)dp->dmtu; rv |= CFG_ERR_MTU; } if (!valid_bufsize(dp->dbufsize)) { cmn_err(CE_WARN, "IDN: 220: remote domain %d BUFSIZE (%d) invalid " "(local.bufsize = %d)", dp->domid, dp->dbufsize, ldp->dbufsize); *exp = (uint_t)ldp->dbufsize; *act = (uint_t)dp->dbufsize; rv |= CFG_ERR_BUF; } if (!valid_slabsize((int)dp->dslabsize)) { cmn_err(CE_WARN, "IDN: 221: remote domain %d SLABSIZE (%d) invalid " "(local.slabsize = %d)", dp->domid, dp->dslabsize, ldp->dslabsize); *exp = (uint_t)ldp->dslabsize; *act = (uint_t)dp->dslabsize; rv |= CFG_ERR_SLAB; } if (!valid_nwrsize((int)dp->dnwrsize)) { cmn_err(CE_WARN, "IDN: 223: remote domain %d NWRSIZE (%d) invalid " "(local.nwrsize = %d)", dp->domid, dp->dnwrsize, ldp->dnwrsize); *exp = (uint_t)ldp->dnwrsize; *act = (uint_t)dp->dnwrsize; rv |= CFG_ERR_NWR; } if ((int)dp->dmaxnets != IDN_MAX_NETS) { cmn_err(CE_WARN, "IDN: 224: remote domain %d MAX_NETS (%d) invalid " "(local.maxnets = %d)", dp->domid, (int)dp->dmaxnets, IDN_MAX_NETS); *exp = (uint_t)IDN_MAX_NETS; *act = (uint_t)dp->dmaxnets; rv |= CFG_ERR_NETS; } if ((int)dp->dmboxpernet != IDN_MBOX_PER_NET) { cmn_err(CE_WARN, "IDN: 225: remote domain %d MBOX_PER_NET (%d) " "invalid (local.mboxpernet = %d)", dp->domid, (int)dp->dmboxpernet, IDN_MBOX_PER_NET); *exp = (uint_t)IDN_MBOX_PER_NET; *act = (uint_t)dp->dmboxpernet; rv |= CFG_ERR_MBOX; } if ((dp->dcksum - 1) != (uchar_t)IDN_CHECKSUM) { cmn_err(CE_WARN, "IDN: 226: remote domain %d CHECKSUM flag (%d) " "mismatches local domain's (%d)", dp->domid, (int)dp->dcksum - 1, IDN_CHECKSUM); *exp = (uint_t)IDN_CHECKSUM; *act = (uint_t)(dp->dcksum - 1); rv |= CFG_ERR_CKSUM; } done: return (rv ? rv : CFG_DONE); } /* * Called by slave ONLY which expects exactly the following * config info from the MASTER: * IDNCFG_BARLAR * IDNCFG_MCADR * IDNCFG_NMCADR * IDNCFG_CPUSET * IDNCFG_NETID * IDNCFG_BOARDSET * IDNCFG_SIZE (MTU, BUF, SLAB, NWR) * IDNCFG_DATAMBOX (TABLE, DOMAIN) * IDNCFG_DATASVR (MAXNETS, MBXPERNET) * IDNCFG_OPTIONS (CHECKSUM) */ static uint_t idn_check_master_config(int domid, uint_t *exp, uint_t *act) { uint_t rv = 0; int nmcadr; int total_expitems; int p, m, err; idn_domain_t *dp; idn_domain_t *ldp = &idn_domain[idn.localid]; procname_t proc = "idn_check_master_config"; dp = &idn_domain[domid]; ASSERT(IDN_GET_MASTERID() != idn.localid); ASSERT(domid != idn.localid); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(dp->dstate == IDNDS_CONFIG); PR_PROTO("%s:%d: number received %d, minimum number expected %d\n", proc, domid, (int)dp->dncfgitems, IDN_MASTER_NCFGITEMS); if ((int)dp->dncfgitems < IDN_MASTER_NCFGITEMS) return (CFG_CONTINUE); /* * We have at least IDN_MASTER_NCFGITEMS items which * means we have at least one MCADR. Need to make sure * we have all that we're expecting, NMCADR. */ total_expitems = IDN_MASTER_NCFGITEMS + dp->dhw.dh_nmcadr - 1; if ((dp->dhw.dh_nmcadr == 0) || ((int)dp->dncfgitems < total_expitems)) { /* * We have not yet received all the MCADRs * we're expecting. */ PR_PROTO("%s:%d: haven't received all MCADRs yet.\n", proc, domid); return (CFG_CONTINUE); } nmcadr = 0; for (p = 0; p < MAX_BOARDS; p++) if (dp->dhw.dh_mcadr[p] != 0) nmcadr++; IDN_GLOCK_SHARED(); if ((idn.smr.rempfn == PFN_INVALID) || (idn.smr.rempfnlim == PFN_INVALID) || (dp->dnetid == (ushort_t)-1) || CPUSET_ISNULL(dp->dcpuset) || (dp->dhw.dh_boardset == 0) || (nmcadr != dp->dhw.dh_nmcadr) || (dp->dmbox.m_send->mm_smr_mboxp == NULL) || (ldp->dmbox.m_tbl == NULL) || (dp->dmaxnets == 0) || (dp->dmboxpernet == 0) || (dp->dcksum == 0) || (dp->dmtu == 0) || (dp->dbufsize == 0) || (dp->dnwrsize == 0)) { IDN_GUNLOCK(); /* * We received all of our config items, but not * all what we were expecting! Gotta reset and * close connection. */ cmn_err(CE_WARN, "IDN: 227: missing some required config items from " "domain %d", domid); rv = CFG_FATAL; goto done; } if ((idn.smr.rempfnlim - idn.smr.rempfn) > btop(MB2B(IDN_SMR_SIZE))) { /* * The master's SMR region is larger than * mine! This means that this domain may * receive I/O buffers which are out of the * range of this local domain's SMR virtual * address space. The master SMR has to be * no larger than the local SMR in order to * guarantee enough local virtual addresses * to see all of the SMR space. * XXX - Possibly add negotiating SMR size. * Try to create a new virtual mapping. * Could let domains negotiate SMR size. * Winning size would have to be smallest * in DC. If so, how to handle incoming * domains with even smaller SMRs? * - Could either disallow connection * - Could reconfigure to use smaller SMR. */ cmn_err(CE_WARN, "IDN: 228: master's SMR (%ld) larger than " "local's SMR (%ld)", idn.smr.rempfnlim - idn.smr.rempfn, btop(MB2B(IDN_SMR_SIZE))); *exp = (uint_t)IDN_SMR_SIZE; *act = (uint_t)B2MB(ptob(idn.smr.rempfnlim - idn.smr.rempfn)); rv |= CFG_ERR_SMR; } IDN_GUNLOCK(); if (!valid_mtu(dp->dmtu)) { cmn_err(CE_WARN, "IDN: 219: remote domain %d MTU (%d) invalid " "(local.mtu = %d)", dp->domid, dp->dmtu, ldp->dmtu); *exp = (uint_t)ldp->dmtu; *act = (uint_t)dp->dmtu; rv |= CFG_ERR_MTU; } if (!valid_bufsize(dp->dbufsize)) { cmn_err(CE_WARN, "IDN: 220: remote domain %d BUFSIZE (%d) invalid " "(local.bufsize = %d)", dp->domid, dp->dbufsize, ldp->dbufsize); *exp = (uint_t)ldp->dbufsize; *act = (uint_t)dp->dbufsize; rv |= CFG_ERR_BUF; } if (!valid_nwrsize((int)dp->dnwrsize)) { cmn_err(CE_WARN, "IDN: 223: remote domain %d NWRSIZE (%d) invalid " "(local.nwrsize = %d)", dp->domid, dp->dnwrsize, ldp->dnwrsize); *exp = (uint_t)ldp->dnwrsize; *act = (uint_t)dp->dnwrsize; rv |= CFG_ERR_NWR; } if ((int)dp->dmaxnets != IDN_MAX_NETS) { cmn_err(CE_WARN, "IDN: 224: remote domain %d MAX_NETS (%d) invalid " "(local.maxnets = %d)", dp->domid, (int)dp->dmaxnets, IDN_MAX_NETS); *exp = (uint_t)IDN_MAX_NETS; *act = (uint_t)dp->dmaxnets; rv |= CFG_ERR_NETS; } if ((int)dp->dmboxpernet != IDN_MBOX_PER_NET) { cmn_err(CE_WARN, "IDN: 225: remote domain %d MBOX_PER_NET (%d) " "invalid (local.mboxpernet = %d)", dp->domid, (int)dp->dmboxpernet, IDN_MBOX_PER_NET); *exp = (uint_t)IDN_MBOX_PER_NET; *act = (uint_t)dp->dmboxpernet; rv |= CFG_ERR_MBOX; } if ((dp->dcksum - 1) != (uchar_t)IDN_CHECKSUM) { cmn_err(CE_WARN, "IDN: 226: remote domain %d CHECKSUM flag (%d) " "mismatches local domain's (%d)", dp->domid, (int)dp->dcksum - 1, IDN_CHECKSUM); *exp = (uint_t)IDN_CHECKSUM; *act = (uint_t)(dp->dcksum - 1); rv |= CFG_ERR_CKSUM; } nmcadr = 0; err = 0; for (m = 0; m < MAX_BOARDS; m++) { if (!BOARD_IN_SET(dp->dhw.dh_boardset, m) && dp->dhw.dh_mcadr[m]) { cmn_err(CE_WARN, "IDN: 229: remote domain %d boardset (0x%x) " "conflicts with MCADR(board %d) [0x%x]", dp->domid, (uint_t)dp->dhw.dh_boardset, m, dp->dhw.dh_mcadr[m]); err++; } if (dp->dhw.dh_mcadr[m]) nmcadr++; } if (err) { *exp = 0; *act = err; rv |= CFG_ERR_MCADR; } else if (nmcadr != dp->dhw.dh_nmcadr) { cmn_err(CE_WARN, "IDN: 230: remote domain %d reported number of " "MCADRs (%d) mismatches received (%d)", dp->domid, dp->dhw.dh_nmcadr, nmcadr); *exp = (uint_t)dp->dhw.dh_nmcadr; *act = (uint_t)nmcadr; rv |= CFG_ERR_NMCADR; } done: return (rv ? rv : CFG_DONE); } static int idn_recv_config_done(int domid) { boardset_t b_conflicts; cpuset_t p_conflicts; register int p, i; register idn_domain_t *dp; idnsb_error_t idnerr; procname_t proc = "idn_recv_config_done"; ASSERT(domid != IDN_NIL_DOMID); dp = &idn_domain[domid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); /* * Well, we received all that we were expecting * so stop any CFG timers we had going. */ IDN_MSGTIMER_STOP(domid, IDNP_CFG, 0); dp->dncpus = 0; for (p = 0; p < NCPU; p++) if (CPU_IN_SET(dp->dcpuset, p)) dp->dncpus++; dp->dhw.dh_nboards = 0; for (p = 0; p < MAX_BOARDS; p++) if (BOARD_IN_SET(dp->dhw.dh_boardset, p)) dp->dhw.dh_nboards++; IDN_GLOCK_EXCL(); /* * Verify dcpuset and dhw.dh_boardset don't * conflict with any existing DC member. */ b_conflicts = idn.dc_boardset & dp->dhw.dh_boardset; CPUSET_ZERO(p_conflicts); CPUSET_OR(p_conflicts, idn.dc_cpuset); CPUSET_AND(p_conflicts, dp->dcpuset); if (b_conflicts || !CPUSET_ISNULL(p_conflicts)) { if (b_conflicts) { cmn_err(CE_WARN, "IDN: 231: domain %d boardset " "(0x%x) conflicts with existing " "IDN boardset (0x%x)", domid, dp->dhw.dh_boardset, b_conflicts); } if (!CPUSET_ISNULL(p_conflicts)) { cmn_err(CE_WARN, "IDN: 232: domain %d cpuset " "(0x%x.%0x) conflicts with existing " "IDN cpuset (0x%x.%0x)", domid, UPPER32_CPUMASK(dp->dcpuset), LOWER32_CPUMASK(dp->dcpuset), UPPER32_CPUMASK(p_conflicts), LOWER32_CPUMASK(p_conflicts)); } IDN_GUNLOCK(); /* * Need to disconnect and not retry with this guy. */ IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_DLOCK_EXCL(domid); INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EPROTO); SET_IDNKERR_IDNERR(&idnerr, IDNKERR_CONFIG_FATAL); SET_IDNKERR_PARAM0(&idnerr, domid); idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); idn_disconnect(domid, IDNFIN_FORCE_HARD, IDNFIN_ARG_CFGERR_FATAL, IDNFIN_SYNC_NO); IDN_SYNC_UNLOCK(); return (-1); } idn_mainmbox_reset(domid, dp->dmbox.m_send); idn_mainmbox_reset(domid, dp->dmbox.m_recv); #ifdef IDNBUG_CPUPERBOARD /* * We only allow connections to domains whose (mem) boards * all have at least one cpu. This is necessary so that * we can program the CICs of that respective board. This * is primarily only a requirement if the remote domain * is the master _and_ has the SMR in that particular board. * To simplify the checking we simply restrict connections to * domains that have at least one cpu on all boards that * contain memory. */ if (!idn_cpu_per_board((void *)NULL, dp->dcpuset, &dp->dhw)) { cmn_err(CE_WARN, "IDN: 233: domain %d missing CPU per " "memory boardset (0x%x), CPU boardset (0x%x)", domid, dp->dhw.dh_boardset, cpuset2boardset(dp->dcpuset)); IDN_GUNLOCK(); /* * Need to disconnect and not retry with this guy. */ IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); DOMAINSET_DEL(idn.domset.ds_relink, domid); IDN_DLOCK_EXCL(domid); INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EINVAL); SET_IDNKERR_IDNERR(&idnerr, IDNKERR_CPU_CONFIG); SET_IDNKERR_PARAM0(&idnerr, domid); idn_update_op(IDNOP_ERROR, DOMAINSET(domid), &idnerr); idn_disconnect(domid, IDNFIN_FORCE_HARD, IDNFIN_ARG_CPUCFG, IDNFIN_SYNC_NO); IDN_SYNC_UNLOCK(); return (-1); } #endif /* IDNBUG_CPUPERBOARD */ CPUSET_OR(idn.dc_cpuset, dp->dcpuset); idn.dc_boardset |= dp->dhw.dh_boardset; IDN_GUNLOCK(); /* * Set up the portmap for this domain. */ i = -1; for (p = 0; p < NCPU; p++) { BUMP_INDEX(dp->dcpuset, i); dp->dcpumap[p] = (uchar_t)i; } /* * Got everything we need from the remote * domain, now we can program hardware as needed. */ if (idn_program_hardware(domid) != 0) { domainset_t domset; /* * Yikes! Failed to program hardware. * Gotta bail. */ cmn_err(CE_WARN, "IDN: 234: failed to program hardware for domain %d " "(boardset = 0x%x)", domid, dp->dhw.dh_boardset); IDN_DUNLOCK(domid); /* * If we're having problems programming our * hardware we better unlink completely from * the IDN before things get really bad. */ IDN_SYNC_LOCK(); IDN_GLOCK_EXCL(); IDN_GSTATE_TRANSITION(IDNGS_DISCONNECT); domset = DOMAINSET_ALL; DOMAINSET_DEL(domset, idn.localid); IDN_SET_NEW_MASTERID(IDN_NIL_DOMID); IDN_GUNLOCK(); INIT_IDNKERR(&idnerr); SET_IDNKERR_ERRNO(&idnerr, EINVAL); SET_IDNKERR_IDNERR(&idnerr, IDNKERR_HW_ERROR); SET_IDNKERR_PARAM0(&idnerr, domid); idn_update_op(IDNOP_ERROR, DOMAINSET_ALL, &idnerr); idn_unlink_domainset(domset, IDNFIN_NORMAL, IDNFIN_ARG_HWERR, IDNFIN_OPT_UNLINK, BOARDSET_ALL); IDN_SYNC_UNLOCK(); IDN_DLOCK_EXCL(domid); return (-1); } /* * Now that hardware has been programmed we can * remap the SMR into our local space, if necessary. */ IDN_GLOCK_EXCL(); if (domid == IDN_GET_MASTERID()) { /* * No need to worry about disabling the data * server since at this stage there is only * one and he doesn't go active until his * mailbox (dmbox.m_recv->mm_smr_mboxp) is set up. */ smr_remap(&kas, idn.smr.vaddr, idn.smr.rempfn, IDN_SMR_SIZE); } IDN_GUNLOCK(); /* * There is no need to ACK the CFG messages since remote * domain would not progress to the next state (CON_SENT) * unless he has received everything. */ dp->dcfgrcvdone = 1; PR_PROTO("%s:%d: RECV config DONE\n", proc, domid); if (dp->dcfgsnddone) { idn_xdcargs_t xargs; /* * Well, we've received all that we were expecting, * but we don't know if the remote domain has * received all that it was expecting from us, * although we know we transferred everything * so let's get the show on the road. */ IDN_DUNLOCK(domid); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(domid); /* * If the state has changed from CONFIG * then somebody else has taken over * control of this domain so we can just * bail out. */ if (dp->dstate == IDNDS_CONFIG) { dp->dxp = &xphase_con; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); bzero(xargs, sizeof (xargs)); idn_xphase_transition(domid, NULL, xargs); } IDN_SYNC_UNLOCK(); } return (0); } static int idn_verify_config_mbox(int domid) { idn_domain_t *ldp, *dp; idn_mainmbox_t *mmp; idn_mboxtbl_t *mtp; int c, rv = 0; uint_t activeptr, readyptr; ushort_t mbox_csum; ASSERT(IDN_DLOCK_IS_EXCL(domid)); dp = &idn_domain[domid]; ldp = &idn_domain[idn.localid]; /* * The master will have assigned us the dmbox.m_tbl * from which we assign our receive mailboxes. * The first (0) entry contains the cookie used * for verification. */ IDN_DLOCK_SHARED(idn.localid); /* * Now that we have an assigned mboxtbl from the * master, we can determine which receive mailbox * we indirectly assigned to him at the time we * sent him his MBOX_INDEX. Prep it, however note * that the master will have not been able to * validate it because of the chicken 'n egg * problem between a master and slave. Thus we * need to reset the cookie after the prep. */ mmp = dp->dmbox.m_recv; mtp = IDN_MBOXTBL_PTR(ldp->dmbox.m_tbl, domid); for (c = 0; c < IDN_MAX_NETS; c++) { mutex_enter(&mmp[c].mm_mutex); ASSERT(!mmp[c].mm_smr_mboxp); mmp[c].mm_smr_mboxp = mtp; mbox_csum = IDN_CKSUM_MBOX(&mtp->mt_header); if (!VALID_MBOXHDR(&mtp->mt_header, c, mbox_csum)) { cmn_err(CE_WARN, "IDN: 235: [recv] mailbox (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, c); cmn_err(CE_CONT, "IDN: 235: [recv] expected (cookie 0x%x, " "cksum 0x%x) actual (cookie 0x%x, " "cksum 0x%x)\n", IDN_GET_MBOXHDR_COOKIE(&mtp->mt_header), (int)mtp->mt_header.mh_cksum, IDN_MAKE_MBOXHDR_COOKIE(0, 0, c), (int)mbox_csum); mutex_exit(&mmp[c].mm_mutex); rv = -1; break; } activeptr = mtp->mt_header.mh_svr_active_ptr; readyptr = mtp->mt_header.mh_svr_ready_ptr; /* * Verify pointers are valid. */ if (!activeptr || !VALID_NWROFFSET(activeptr, 2) || !readyptr || !VALID_NWROFFSET(readyptr, 2)) { cmn_err(CE_WARN, "IDN: 235: [recv] mailbox (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, c); cmn_err(CE_CONT, "IDN: 235: [recv] activeptr (0x%x), " "readyptr (0x%x)\n", activeptr, readyptr); mutex_exit(&mmp[c].mm_mutex); rv = -1; break; } mmp[c].mm_smr_activep = (ushort_t *)IDN_OFFSET2ADDR(activeptr); mmp[c].mm_smr_readyp = (ushort_t *)IDN_OFFSET2ADDR(readyptr); mutex_exit(&mmp[c].mm_mutex); IDN_MBOXTBL_PTR_INC(mtp); } IDN_DUNLOCK(idn.localid); if (rv) return (rv); /* * Now we need to translate SMR offsets for send mailboxes * to actual virtual addresses. */ mmp = dp->dmbox.m_send; for (c = 0; c < IDN_MAX_NETS; mmp++, c++) { mutex_enter(&mmp->mm_mutex); if ((mtp = mmp->mm_smr_mboxp) == NULL) { mutex_exit(&mmp->mm_mutex); rv = -1; break; } mbox_csum = IDN_CKSUM_MBOX(&mtp->mt_header); if (!VALID_MBOXHDR(&mtp->mt_header, c, mbox_csum)) { cmn_err(CE_WARN, "IDN: 235: [send] mailbox (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, c); cmn_err(CE_CONT, "IDN: 235: [send] expected (cookie 0x%x, " "cksum 0x%x) actual (cookie 0x%x, " "cksum 0x%x)\n", IDN_GET_MBOXHDR_COOKIE(&mtp->mt_header), (int)mtp->mt_header.mh_cksum, IDN_MAKE_MBOXHDR_COOKIE(0, 0, c), (int)mbox_csum); mutex_exit(&mmp->mm_mutex); rv = -1; break; } activeptr = mtp->mt_header.mh_svr_active_ptr; readyptr = mtp->mt_header.mh_svr_ready_ptr; /* * Paranoid check. */ if (!activeptr || !VALID_NWROFFSET(activeptr, 2) || !readyptr || !VALID_NWROFFSET(readyptr, 2)) { cmn_err(CE_WARN, "IDN: 235: [send] mailbox (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, c); cmn_err(CE_CONT, "IDN: 235: [send] activeptr (0x%x), " "readyptr (0x%x)\n", activeptr, readyptr); mutex_exit(&mmp->mm_mutex); rv = -1; break; } mmp->mm_smr_activep = (ushort_t *)IDN_OFFSET2ADDR(activeptr); mmp->mm_smr_readyp = (ushort_t *)IDN_OFFSET2ADDR(readyptr); idn_reset_mboxtbl(mtp); mutex_exit(&mmp->mm_mutex); IDN_MBOXTBL_PTR_INC(mtp); } return (rv); } /* * The BUFSIZEs between domains have to be equal so that slave buffers * and the master's slabpool are consistent. * The MTUs between domains have to be equal so they can transfer * packets consistently without possible data truncation. * * ZZZ - Perhaps these could be negotiated? */ static int valid_mtu(uint_t mtu) { return ((mtu == idn_domain[idn.localid].dmtu) && mtu); } static int valid_bufsize(uint_t bufsize) { return ((bufsize == idn_domain[idn.localid].dbufsize) && bufsize); } static int valid_slabsize(int slabsize) { return ((slabsize == idn_domain[idn.localid].dslabsize) && slabsize); } static int valid_nwrsize(int nwrsize) { return ((nwrsize == idn_domain[idn.localid].dnwrsize) && nwrsize); } static int idn_program_hardware(int domid) { int rv, is_master; idn_domain_t *dp; uint_t *mcadrp; pfn_t rem_pfn, rem_pfnlimit; procname_t proc = "idn_program_hardware"; PR_PROTO("%s:%d: program hw in domain %d w.r.t remote domain %d\n", proc, domid, idn.localid, domid); dp = &idn_domain[domid]; ASSERT(domid != idn.localid); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(dp->dstate == IDNDS_CONFIG); IDN_GLOCK_EXCL(); if (DOMAIN_IN_SET(idn.domset.ds_hwlinked, domid)) { IDN_GUNLOCK(); return (0); } DOMAINSET_ADD(idn.domset.ds_flush, domid); CHECKPOINT_OPENED(IDNSB_CHKPT_CACHE, dp->dhw.dh_boardset, 1); if (domid != IDN_GET_MASTERID()) { /* * If the remote domain is a slave, then * all we have to program is the CIC sm_mask. */ is_master = 0; if ((idn.localid == IDN_GET_MASTERID()) && lock_try(&idn.first_hwlink)) { /* * This is our first HW link and I'm the * master, which means we need to program * our local bar/lar. */ ASSERT(idn.first_hwmasterid == (short)IDN_NIL_DOMID); idn.first_hwmasterid = (short)idn.localid; rem_pfn = idn.smr.locpfn; rem_pfnlimit = idn.smr.locpfn + btop(MB2B(IDN_SMR_SIZE)); } else { /* * Otherwise, just a slave linking to * another slave. No bar/lar updating * necessary. */ rem_pfn = rem_pfnlimit = PFN_INVALID; } mcadrp = NULL; } else { /* * If the remote domain is a master, then * we need to program the CIC sm_mask/sm_bar/sm_lar, * and PC's. */ is_master = 1; rem_pfn = idn.smr.rempfn; rem_pfnlimit = idn.smr.rempfnlim; mcadrp = dp->dhw.dh_mcadr; ASSERT(idn.first_hwmasterid == (short)IDN_NIL_DOMID); idn.first_hwmasterid = (short)domid; } PR_PROTO("%s:%d: ADD bset (0x%x)\n", proc, domid, dp->dhw.dh_boardset); rv = idnxf_shmem_add(is_master, dp->dhw.dh_boardset, rem_pfn, rem_pfnlimit, mcadrp); if (rv == 0) { DOMAINSET_ADD(idn.domset.ds_hwlinked, domid); } else { if (rem_pfn == idn.smr.locpfn) lock_clear(&idn.first_hwlink); if (idn.first_hwmasterid == (short)domid) idn.first_hwmasterid = (short)IDN_NIL_DOMID; (void) idnxf_shmem_sub(is_master, dp->dhw.dh_boardset); } IDN_GUNLOCK(); return (rv); } static int idn_deprogram_hardware(int domid) { int rv, is_master; idn_domain_t *dp; procname_t proc = "idn_deprogram_hardware"; dp = &idn_domain[domid]; ASSERT(domid != idn.localid); ASSERT(IDN_DLOCK_IS_EXCL(domid)); /* * Need to take into consideration what boards remote * domain was connected to. If we don't have a connection to * them ourself, then we better remove them now , otherwise * they'll never be removed (unless we link to them at some point). */ #if 0 DEBUG_USECDELAY(500000); #endif /* 0 */ IDN_GLOCK_EXCL(); if (!DOMAIN_IN_SET(idn.domset.ds_hwlinked, domid)) { IDN_GUNLOCK(); return (0); } PR_PROTO("%s:%d: DEprogram hw in domain %d w.r.t remote domain %d\n", proc, domid, idn.localid, domid); /* * It's possible to come through this flow for domains that * have not been programmed, i.e. not in idn.hwlinked_domset, * so don't bother asserting that they might be in there. * This can occur if we lose a domain during the config/syn * sequence. If this occurs we won't know whether the remote * domain has programmed its hardware or not. If it has then * it will have to go through the DMAP sequence and thus we * have to go through it also. So, if we reach at least the * CONFIG state, we need to go through the DMAP handshake. */ PR_PROTO("%s:%d: SUB bset (0x%x)\n", proc, domid, dp->dhw.dh_boardset); if (idn.first_hwmasterid == (short)domid) { is_master = 1; idn.first_hwmasterid = (short)IDN_NIL_DOMID; } else { is_master = 0; } rv = idnxf_shmem_sub(is_master, dp->dhw.dh_boardset); if (rv == 0) DOMAINSET_DEL(idn.domset.ds_hwlinked, domid); IDN_GUNLOCK(); return (rv); } /* * Remember can't send slabs back to master at this point. * Entered with write-drwlock held. * Returns with drwlock dropped. */ static void idn_deconfig(int domid) { idn_domain_t *dp, *ldp; smr_slab_t *sp; int c, masterid; procname_t proc = "idn_deconfig"; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_DLOCK_IS_EXCL(domid)); ASSERT(domid != idn.localid); ldp = &idn_domain[idn.localid]; dp = &idn_domain[domid]; ASSERT(dp->dstate == IDNDS_DMAP); PR_PROTO("%s:%d: (dio=%d, dioerr=%d, dnslabs=%d)\n", proc, domid, dp->dio, dp->dioerr, dp->dnslabs); IDN_GLOCK_EXCL(); masterid = IDN_GET_MASTERID(); idn.dc_boardset &= ~dp->dhw.dh_boardset; for (c = 0; c < NCPU; c++) { if (CPU_IN_SET(dp->dcpuset, c)) { CPUSET_DEL(idn.dc_cpuset, c); } } IDN_GUNLOCK(); (void) smr_buf_free_all(domid); if (idn.localid == masterid) { /* * Since I'm the master there may * have been slabs in this domain's * idn_domain[] entry. */ DSLAB_LOCK_EXCL(domid); if ((sp = dp->dslab) != NULL) { PR_PROTO("%s:%d: freeing up %d dead slabs\n", proc, domid, dp->dnslabs); smr_slab_free(domid, sp); dp->dslab = NULL; dp->dnslabs = 0; dp->dslab_state = DSLAB_STATE_UNKNOWN; } DSLAB_UNLOCK(domid); } else if (domid == masterid) { /* * We're shutting down the master! * We need to blow away our local slab * data structures. * Since I'm not the master, there should * be no slab structures in the given * domain's idn_domain[] entry. They should * only exist in the local domain's entry. */ DSLAB_LOCK_EXCL(idn.localid); ASSERT(dp->dslab == NULL); #ifdef DEBUG { int nbusy = 0; uint_t dommask = 0; for (sp = ldp->dslab; sp; sp = sp->sl_next) { smr_slabbuf_t *bp; if (!smr_slab_busy(sp)) continue; nbusy++; for (bp = sp->sl_inuse; bp; bp = bp->sb_next) if (bp->sb_domid != IDN_NIL_DOMID) DOMAINSET_ADD(dommask, bp->sb_domid); } if (nbusy) PR_PROTO("%s:%d: found %d busy slabs " "(dommask = 0x%x)\n", proc, domid, nbusy, dommask); } #endif /* DEBUG */ if ((sp = ldp->dslab) != NULL) { PR_PROTO("%s:%d: freeing up %d local slab " "structs\n", proc, domid, ldp->dnslabs); smr_slab_garbage_collection(sp); ldp->dslab = NULL; ldp->dnslabs = 0; ldp->dslab_state = DSLAB_STATE_UNKNOWN; } DSLAB_UNLOCK(idn.localid); } if (dp->dio) { PR_PROTO("%s:%d: reset dio (%d) to 0\n", proc, domid, dp->dio); dp->dio = 0; } dp->dioerr = 0; PR_PROTO("%s:%d: reset diocheck (%x) to 0\n", proc, domid, dp->diocheck); lock_clear(&dp->diocheck); CHECKPOINT_CLOSED(IDNSB_CHKPT_LINK, dp->dhw.dh_boardset, 2); /* * Should have already flush our memory before * reaching this stage. The issue is that by the * time we reach here the remote domains may have * already reprogrammed their hardware and so flushing * out caches now could result in a arbstop/hang * if we have data that needs to go back to one * of the remote domains that has already reprogrammed * its hardware. */ ASSERT(!DOMAIN_IN_SET(idn.domset.ds_flush, domid)); (void) idn_deprogram_hardware(domid); /* * XXX - what to do if we * fail to program hardware * probably should panic since * demise of system may be near? * Sufficient to just shutdown network? */ IDN_DSTATE_TRANSITION(dp, IDNDS_CLOSED); idn_close_domain(domid); } /* * If we're sending a Reset we better make sure we don't have any * references or traffic headed in the direction of this guy, since * when he receives the reset, he'll start shutting down which means * we effectively have to shutdown _before_ sending the reset. * DO NOT HOLD ANY DOMAIN RWLOCKS ON ENTRY. Could result in deadlock * due to channel server looping back through STREAMs and attempting * to acquire domain lock, i.e. channel server will never "stop". */ static void idn_shutdown_datapath(domainset_t domset, int force) { int do_allchan; idn_domain_t *dp; register int d; procname_t proc = "idn_shutdown_datapath"; PR_CHAN("%s: domset = 0x%x\n", proc, (uint_t)domset); do_allchan = (domset == DOMAINSET_ALL) ? 1 : 0; DOMAINSET_DEL(domset, idn.localid); if (do_allchan) { /* * Need to stop all outgoing and * incoming SMR references. */ idn_deactivate_channel(CHANSET_ALL, IDNCHAN_OFFLINE); } /* * If force is set then we don't want to reference * the SMR at all, so deactivate the domains from * channels first. This will result in the mainmbox-flush * routines to just clean up without referencing the * SMR space. */ if (force) idn_mainmbox_deactivate(domset); /* * Flush out mailboxes (clear smr reference). */ for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(domset, d)) continue; dp = &idn_domain[d]; if ((dp->dmbox.m_send == NULL) && (dp->dmbox.m_recv == NULL)) continue; IDN_MBOX_LOCK(d); if (dp->dmbox.m_send) (void) idn_mainmbox_flush(d, dp->dmbox.m_send); if (dp->dmbox.m_recv) (void) idn_mainmbox_flush(d, dp->dmbox.m_recv); IDN_MBOX_UNLOCK(d); } /* * Deactivate all domain references also. * Only necessary if it wasn't already done above. */ if (!force) idn_mainmbox_deactivate(domset); } void idn_send_cmd(int domid, idn_cmd_t cmdtype, uint_t arg1, uint_t arg2, uint_t arg3) { idn_msgtype_t mt; procname_t proc = "idn_send_cmd"; mt.mt_mtype = IDNP_CMD; mt.mt_atype = 0; mt.mt_cookie = 0; ASSERT(IDN_DLOCK_IS_HELD(domid)); PR_PROTO("%s:%d: sending command %s\n", proc, domid, VALID_IDNCMD(cmdtype) ? idncmd_str[cmdtype] : "unknown"); IDN_MSGTIMER_START(domid, IDNP_CMD, (ushort_t)cmdtype, idn_msg_waittime[IDNP_CMD], &mt.mt_cookie); IDNXDC(domid, &mt, (uint_t)cmdtype, arg1, arg2, arg3); } void idn_send_cmdresp(int domid, idn_msgtype_t *mtp, idn_cmd_t cmdtype, uint_t arg1, uint_t arg2, uint_t cerrno) { idn_msgtype_t mt; ASSERT(IDN_DLOCK_IS_HELD(domid)); if (domid == idn.localid) { /* * It's possible local domain received a command * from itself. However, we cannot send a normal * "ack" response (XDC) to ourself. */ return; } mt.mt_mtype = IDNP_CMD | IDNP_ACK; mt.mt_atype = 0; mt.mt_cookie = mtp->mt_cookie; IDNXDC(domid, &mt, (uint_t)cmdtype, arg1, arg2, cerrno); } static void idn_send_cmd_nackresp(int domid, idn_msgtype_t *mtp, idn_cmd_t cmdtype, idn_nack_t nacktype) { idn_msgtype_t mt; if (domid == idn.localid) return; mt.mt_mtype = IDNP_CMD | IDNP_NACK; mt.mt_atype = 0; mt.mt_cookie = mtp->mt_cookie; (void) IDNXDC(domid, &mt, (uint_t)cmdtype, (uint_t)nacktype, 0, 0); } void idn_broadcast_cmd(idn_cmd_t cmdtype, uint_t arg1, uint_t arg2, uint_t arg3) { idn_msgtype_t mt; domainset_t domset; procname_t proc = "idn_broadcast_cmd"; IDN_GLOCK_SHARED(); domset = idn.domset.ds_connected; DOMAINSET_DEL(domset, idn.localid); PR_PROTO("%s: broadcasting command (%s) to domainset 0x%x\n", proc, VALID_IDNCMD(cmdtype) ? idncmd_str[cmdtype] : "unknown", domset); mt.mt_mtype = IDNP_CMD; mt.mt_atype = 0; mt.mt_cookie = 0; IDNXDC_BROADCAST(domset, &mt, (uint_t)cmdtype, arg1, arg2, arg3); IDN_GUNLOCK(); /* * This is a broadcast which means local domain needs * to process it also. Since we can't XDC to ourselves * we simply call a local function. */ idn_local_cmd(cmdtype, arg1, arg2, arg3); } /* * Since xargs[0] contains the cmdtype, only xargs[1], xargs[2], xargs[3] * are valid possible response arguments. */ static void idn_recv_cmd(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { uint_t msg = mtp->mt_mtype; register idn_domain_t *dp; idn_cmd_t cmdtype; uint_t acknack; uint_t cmdarg1, cmdarg2, cmdarg3; int islocal; int unsup_cmd_sent, unsup_cmd_recvd; procname_t proc = "idn_recv_cmd"; acknack = msg & IDNP_ACKNACK_MASK; GET_XARGS(xargs, &cmdtype, &cmdarg1, &cmdarg2, &cmdarg3); dp = &idn_domain[domid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); IDN_GLOCK_SHARED(); islocal = (domid == idn.localid); ASSERT(!acknack || (acknack & IDNP_ACKNACK_MASK)); PR_PROTO("%s:%d: (local=%d) acknack=0x%x, cmdtype=%s(%d), " "a1=0x%x, a2=0x%x, a3=0x%x\n", proc, domid, islocal, acknack, VALID_IDNCMD(cmdtype) ? idncmd_str[cmdtype] : "unknown", cmdtype, cmdarg1, cmdarg2, cmdarg3); unsup_cmd_sent = unsup_cmd_recvd = 0; if ((IDN_GET_MASTERID() == IDN_NIL_DOMID) || (dp->dstate != IDNDS_CONNECTED)) { /* * Commands cannot be handled without a valid * master. If this is a request then nack him. */ PR_PROTO("%s:%d: cannot process CMD w/o master (%d, %s)\n", proc, domid, IDN_GET_MASTERID(), idnds_str[dp->dstate]); if (!islocal && !(acknack & IDNP_ACKNACK_MASK)) idn_send_cmd_nackresp(domid, mtp, cmdtype, IDNNACK_NOCONN); IDN_GUNLOCK(); return; } IDN_GUNLOCK(); if (acknack & IDNP_ACKNACK_MASK) { idn_nack_t nack; /* * Receiving a cmd+ack or cmd+nack in response to some * earlier command we must have issued. * If the response is a nack, there are two possibilites: * * 1. Remote domain failed to allocate due * to limited resources. * * 2. Remote domain does not support this * particular command. * * In the case of #2, the argument immediately after * the cmdtype (xargs[1]) will be (-1). */ nack = (idn_nack_t)cmdarg1; if ((acknack & IDNP_NACK) && (nack == IDNNACK_BADCMD)) unsup_cmd_sent++; if (islocal) { /* * Shouldn't be receiving local commands w/acks. */ cmdtype = (idn_cmd_t)0; } switch (cmdtype) { case IDNCMD_SLABALLOC: idn_recv_slaballoc_resp(domid, cmdarg1, cmdarg2, cmdarg3); break; case IDNCMD_SLABFREE: idn_recv_slabfree_resp(domid, cmdarg1, cmdarg2, cmdarg3); break; case IDNCMD_SLABREAP: /* * We only care if successful. */ if (acknack & IDNP_ACK) idn_recv_slabreap_resp(domid, cmdarg1, cmdarg3); break; case IDNCMD_NODENAME: if ((acknack & IDNP_NACK) == 0) { idn_recv_nodename_resp(domid, cmdarg1, cmdarg3); break; } switch (nack) { case IDNNACK_NOCONN: case IDNNACK_RETRY: /* * Remote domain was not quite * ready, try again. */ PR_PROTO("%s:%d: remote not ready " "for %s - retrying " "[dstate=%s]\n", proc, domid, idncmd_str[IDNCMD_NODENAME], idnds_str[dp->dstate]); if (dp->dstate == IDNDS_CONNECTED) (void) timeout(idn_retry_nodename_req, (void *)(uintptr_t)domid, hz); default: break; } break; default: /* * Unsupported command. */ unsup_cmd_recvd++; break; } if (unsup_cmd_sent) { PR_PROTO("%s:%d: unsupported command " "requested (0x%x)\n", proc, domid, cmdtype); } if (unsup_cmd_recvd) { PR_PROTO("%s:%d: unsupported command " "response (0x%x)\n", proc, domid, cmdtype); } } else { /* * Receiving a regular cmd from a remote domain. */ switch (cmdtype) { case IDNCMD_SLABALLOC: idn_recv_slaballoc_req(domid, mtp, cmdarg1); break; case IDNCMD_SLABFREE: idn_recv_slabfree_req(domid, mtp, cmdarg1, cmdarg2); break; case IDNCMD_SLABREAP: idn_recv_slabreap_req(domid, mtp, cmdarg1); break; case IDNCMD_NODENAME: idn_recv_nodename_req(domid, mtp, cmdarg1); break; default: /* * Unsupported command. */ unsup_cmd_recvd++; break; } if (!islocal && unsup_cmd_recvd) { /* * Received an unsupported IDN command. */ idn_send_cmd_nackresp(domid, mtp, cmdtype, IDNNACK_BADCMD); } } } /* * This is a supporting routine for idn_broadcast_cmd() to * handle processing of the requested command for the local * domain. Currently the only support broadcast command * supported is reaping. */ /*ARGSUSED2*/ static void idn_local_cmd(idn_cmd_t cmdtype, uint_t arg1, uint_t arg2, uint_t arg3) { idn_protojob_t *jp; idn_domain_t *ldp = &idn_domain[idn.localid]; procname_t proc = "idn_local_cmd"; PR_PROTO("%s: submitting local command %s on domain %d\n", proc, VALID_IDNCMD(cmdtype) ? idncmd_str[cmdtype] : "unknown", idn.localid); jp = idn_protojob_alloc(KM_SLEEP); jp->j_msg.m_domid = ldp->domid; jp->j_msg.m_msgtype = IDNP_CMD; jp->j_msg.m_cookie = ldp->dcookie_recv; SET_XARGS(jp->j_msg.m_xargs, cmdtype, arg1, arg2, arg3); idn_protojob_submit(ldp->domid, jp); } /* * Terminate any outstanding commands that may have * been targeted for the given domain. A command is * designated as outstanding if it has an active timer. * * serrno = ECANCELED. */ static void idn_terminate_cmd(int domid, int serrno) { idn_domain_t *dp; idn_timer_t *tplist = NULL, *tp; procname_t proc = "idn_terminate_cmd"; dp = &idn_domain[domid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); IDN_MSGTIMER_GET(dp, IDNP_CMD, tplist, 0); /* * At this point the timers are effectively terminated * since when they're t_onq indication is set false. */ if (tplist == NULL) { PR_PROTO("%s:%d: no outstanding cmds found\n", proc, domid); /* * There is a window where we may have caught a * request just prior to issuing the actual * command (SLABALLOC). We're guaranteed if there * was, then he will have at least registered. * So, if we abort the command now, he'll catch * it before going to sleep. * Drop through. */ } ASSERT(tplist ? (tplist->t_back->t_forw == NULL) : 1); for (tp = tplist; tp; tp = tp->t_forw) { ASSERT(tp->t_type == IDNP_CMD); PR_PROTO("%s:%d: found outstanding cmd: %s\n", proc, domid, idncmd_str[tp->t_subtype]); switch (tp->t_subtype) { case IDNCMD_SLABALLOC: /* * Outstanding slaballoc request may have * slab waiters hanging around. Need to * tell them to bail out. The given domain * must be the master if we have an outstanding * command to him. This also presumes that * if there are any waiters they're only in * the local domain's waiting area (i.e. we're * a slave). */ #ifdef DEBUG IDN_GLOCK_SHARED(); ASSERT(domid == IDN_GET_MASTERID()); ASSERT(idn.localid != IDN_GET_MASTERID()); IDN_GUNLOCK(); #endif /* DEBUG */ (void) smr_slabwaiter_abort(idn.localid, serrno); break; case IDNCMD_SLABFREE: case IDNCMD_SLABREAP: case IDNCMD_NODENAME: /* * Nothing really waiting for these operations * so no biggy if we just drop. * Note that NODENAME may have an outstanding * buffer, however that will be reclaimed * when we actually unlink from domain. */ break; default: ASSERT(0); break; } } /* * As mentioned before the timers are effectively no-op'd * once they're dequeued, however let's cleanup house and * get rid of the useless entries in the timeout queue. */ if (tplist) { IDN_TIMER_STOPALL(tplist); } if (idn_domain[idn.localid].dvote.v.master) { /* * I'm the master so it's possible I had * outstanding commands (SLABALLOC) waiting * to be satisfied for the given domain. * Since we're forcing an error it's okay * to continue holding onto the drwlock. */ PR_PROTO("%s:%d: abort slaballoc waiters\n", proc, domid); (void) smr_slabwaiter_abort(domid, serrno); } else if (dp->dvote.v.master) { PR_PROTO("%s:%d: abort (local domain) slaballoc waiters\n", proc, domid); (void) smr_slabwaiter_abort(idn.localid, serrno); } } static void idn_send_acknack(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_send_acknack"; ASSERT(mtp ? (mtp->mt_mtype & IDNP_ACKNACK_MASK) : 1); ASSERT(domid != IDN_NIL_DOMID); #ifdef DEBUG { STRING(mstr); STRING(astr); INUM2STR(mtp->mt_mtype, mstr); INUM2STR(mtp->mt_atype, astr); if (mtp->mt_mtype & IDNP_ACK) { PR_PROTO("%s:%d: dstate=%s, msg=(%s/%s), " "a1=0x%x, a2=0x%x, a3=0x%x, a4 = 0x%x\n", proc, domid, idnds_str[dp->dstate], astr, mstr, xargs[0], xargs[1], xargs[2], xargs[3]); } else { idn_nack_t nack; nack = GET_XARGS_NACK_TYPE(xargs); PR_PROTO("%s:%d: dstate=%s, msg=(%s/%s), " "nack=%s(0x%x)\n", proc, domid, idnds_str[dp->dstate], astr, mstr, idnnack_str[nack], (uint_t)nack); } } #endif /* DEBUG */ (void) IDNXDC(domid, mtp, xargs[0], xargs[1], xargs[2], xargs[3]); } /*ARGSUSED0*/ static void idn_prealloc_slab(int nslabs) { register int s, serrno; smr_slab_t *sp; idn_domain_t *ldp = &idn_domain[idn.localid]; procname_t proc = "idn_prealloc_slab"; IDN_GLOCK_SHARED(); DSLAB_LOCK_SHARED(idn.localid); if ((idn.state != IDNGS_ONLINE) || (ldp->dnslabs > 0)) { /* * Not in the proper state or slab already allocated. */ DSLAB_UNLOCK(idn.localid); IDN_GUNLOCK(); return; } IDN_GUNLOCK(); ASSERT(!ldp->dslab); serrno = 0; for (s = 0; (s < nslabs) && ((int)ldp->dnslabs < nslabs); s++) { /* * Returns with ldp->drwlock dropped. */ serrno = smr_slab_alloc(idn.localid, &sp); if (serrno != 0) { PR_PROTO("%s: FAILED to pre-alloc'd " "slab (serrno = %d)\n", proc, serrno); break; } /* * State may have changed since smr_slab_alloc * temporarily drops drwlock. Make sure we're * still connected. */ PR_PROTO("%s: SUCCESSFULLY pre-alloc'd slab\n", proc); if (idn.state != IDNGS_ONLINE) { PR_PROTO("%s: Lost connection..leaving\n", proc); break; } } DSLAB_UNLOCK(idn.localid); } /* * Received a request from a remote domain to * allocate a slab from the master SMR for him. * Allocate slab and return the response. */ static void idn_recv_slaballoc_req(int domid, idn_msgtype_t *mtp, uint_t slab_size) { register idn_domain_t *dp; procname_t proc = "idn_recv_slaballoc_req"; PR_PROTO("%s: slaballoc req from domain %d (size=0x%x)\n", proc, domid, slab_size); dp = &idn_domain[domid]; ASSERT(IDN_DLOCK_IS_EXCL(domid)); IDN_GLOCK_SHARED(); if (idn.localid != IDN_GET_MASTERID()) { IDN_GUNLOCK(); /* * It's a fatal error if the remote domain thinks * we're the master. */ idn_send_slaballoc_resp(domid, mtp, 0, 0, EACCES); } else if (dp->dstate != IDNDS_CONNECTED) { IDN_GUNLOCK(); /* * It's a fatal error if we don't yet have a * connection established with the requestor. */ idn_send_slaballoc_resp(domid, mtp, 0, 0, ENOLINK); } else { int serrno; smr_slab_t *sp; smr_offset_t slab_offset; IDN_GUNLOCK(); DSLAB_LOCK_SHARED(domid); IDN_DUNLOCK(domid); /* * We're connected and we're the master. * smr_slab_alloc() returns with dp->drwlock dropped. */ if ((serrno = smr_slab_alloc(domid, &sp)) == 0) { /* * Successfully allocated slab for remote slave. */ slab_offset = IDN_ADDR2OFFSET(sp->sl_start); slab_size = sp->sl_end - sp->sl_start; ASSERT((slab_offset != 0) && (slab_size != 0)); } else { slab_offset = slab_size = 0; } DSLAB_UNLOCK(domid); /* * The drwlock is dropped during smr_slab_alloc. * During that time our connection with the given * domain may have changed. Better check again. */ IDN_DLOCK_SHARED(domid); if ((dp->dstate != IDNDS_CONNECTED) && !serrno) { /* * Connection broke. Keep the slab here. */ DSLAB_LOCK_EXCL(domid); IDN_DUNLOCK(domid); smr_slab_free(domid, sp); DSLAB_UNLOCK(domid); slab_offset = slab_size = 0; serrno = ECANCELED; IDN_DLOCK_SHARED(domid); } /* * Send response. * Note that smr_slab_alloc automatically installs * slab into domains respective idn_domain entry * to be associated with that domain. */ idn_send_slaballoc_resp(domid, mtp, slab_offset, slab_size, serrno); } } static void idn_send_slaballoc_resp(int domid, idn_msgtype_t *mtp, smr_offset_t slab_offset, uint_t slab_size, int serrno) { procname_t proc = "idn_send_slaballoc_resp"; PR_PROTO("%s: slaballoc resp to domain %d (off=0x%x, size=0x%x) " "[serrno = %d]\n", proc, domid, slab_offset, slab_size, serrno); idn_send_cmdresp(domid, mtp, IDNCMD_SLABALLOC, slab_offset, slab_size, serrno); } /* * Received the ack or nack to a previous allocation request * made by the local domain to the master for a slab. Need * to "put" the response into the waiting area for any * waiters. */ static void idn_recv_slaballoc_resp(int domid, smr_offset_t slab_offset, uint_t slab_size, int serrno) { smr_slab_t *sp = NULL; int rv; procname_t proc = "idn_recv_slaballoc_resp"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); PR_PROTO("%s: slaballoc resp from domain %d (off=0x%x, size=0x%x) " "[serrno = %d]\n", proc, domid, slab_offset, slab_size, serrno); if (!serrno) { IDN_GLOCK_SHARED(); if (domid != IDN_GET_MASTERID()) { /* * We should only be receiving responses from * our master. This is either a bogus message * or an old response. In either case dump it. */ PR_PROTO("%s: BOGUS slaballoc resp from domid %d " "(master = %d)\n", proc, domid, IDN_GET_MASTERID()); serrno = EPROTO; } IDN_GUNLOCK(); if (!serrno && !VALID_NWROFFSET(slab_offset, IDN_SMR_BUFSIZE)) { PR_PROTO("%s: slab offset (0x%x) out of range " "(0-0x%lx)\n", proc, slab_offset, MB2B(IDN_NWR_SIZE)); serrno = EPROTO; } else if (!serrno) { sp = GETSTRUCT(smr_slab_t, 1); sp->sl_start = IDN_OFFSET2ADDR(slab_offset); sp->sl_end = sp->sl_start + slab_size; smr_alloc_buflist(sp); } } /* * Always "put" slabs back to yourself since you're a slave. * Note that we set the forceflag so that even if there are * no waiters we still install the slab for the domain. */ if (!serrno) { DSLAB_LOCK_EXCL(idn.localid); } rv = smr_slaballoc_put(idn.localid, sp, 1, serrno); if (!serrno) { DSLAB_UNLOCK(idn.localid); } if (rv < 0) { /* * Some kind of error trying to install response. * If there was a valid slab sent to us, we'll * just have to send it back. */ PR_PROTO("%s: failed to install response in waiting area\n", proc); if (slab_size != 0) { PR_PROTO("%s: sending slab back to domain %d " "(master = %d)\n", proc, domid, IDN_GET_MASTERID()); idn_send_cmd(domid, IDNCMD_SLABFREE, slab_offset, slab_size, 0); } if (sp) { smr_free_buflist(sp); FREESTRUCT(sp, smr_slab_t, 1); } } } /* * Note that slab reaping is effectively performed asynchronously * since the request will be received a protocol server. */ static void idn_recv_slabreap_req(int domid, idn_msgtype_t *mtp, int nslabs) { procname_t proc = "idn_recv_slabreap_req"; PR_PROTO("%s: slab reap request (nslabs = %d)\n", proc, nslabs); ASSERT(IDN_DLOCK_IS_EXCL(domid)); IDN_GLOCK_SHARED(); if (domid != IDN_GET_MASTERID()) { /* * Only the master can request that slabs be reaped. */ IDN_GUNLOCK(); PR_PROTO("%s: only master can request slab reaping\n", proc); idn_send_cmdresp(domid, mtp, IDNCMD_SLABREAP, 0, 0, EACCES); return; } IDN_GUNLOCK(); if (nslabs != 0) { IDN_DUNLOCK(domid); smr_slab_reap(idn.localid, &nslabs); IDN_DLOCK_SHARED(domid); } PR_PROTO("%s: slab reap result (nslabs = %d)\n", proc, nslabs); /* * Go ahead and send the reap response back before we start * free'ing off the individual slabs. */ idn_send_slabreap_resp(domid, mtp, nslabs, 0); } static void idn_recv_slabreap_resp(int domid, int nslabs, int serrno) { procname_t proc = "idn_recv_slabreap_resp"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); if ((idn.localid != IDN_GET_MASTERID()) || (idn.localid == domid)) { PR_PROTO("%s: unexpected slabreap resp received " "(domid = %d)\n", proc, domid); ASSERT(0); return; } PR_PROTO("%s: recvd reap response from domain %d for %d slabs " "[serrno = %d]\n", proc, domid, nslabs, serrno); } /* * Not really necessary to send slabreap response. * XXX - perhaps useful to master for accounting or * throttling of further reaping? */ static void idn_send_slabreap_resp(int domid, idn_msgtype_t *mtp, int nslabs, int serrno) { idn_send_cmdresp(domid, mtp, IDNCMD_SLABREAP, nslabs, 0, serrno); } /* * Slave -> Master ONLY * Master never sends slabfree request to itself. */ static void idn_recv_slabfree_req(int domid, idn_msgtype_t *mtp, smr_offset_t slab_offset, uint_t slab_size) { smr_slab_t *sp; int serrno; caddr_t s_start, s_end; procname_t proc = "idn_recv_slabfree_req"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (domid == IDN_GET_MASTERID()) { PR_PROTO("%s: unexpected slabfree req received (domid = %d)\n", proc, domid); idn_send_slabfree_resp(domid, mtp, slab_offset, slab_size, EACCES); return; } if (slab_size > IDN_SLAB_SIZE) { PR_PROTO("%s: unexpected slab size. exp %d, recvd %d\n", proc, IDN_SLAB_SIZE, slab_size); idn_send_slabfree_resp(domid, mtp, slab_offset, slab_size, EINVAL); return; } s_start = IDN_OFFSET2ADDR(slab_offset); s_end = s_start + slab_size; /* * Master has received a SLABFREE request (effectively a response * to some earlier SLABREAP request. * Find the slab associated with this slab and free it up. */ DSLAB_LOCK_EXCL(domid); if ((sp = smr_slaballoc_get(domid, s_start, s_end)) != NULL) { smr_slab_free(domid, sp); serrno = 0; } else { serrno = EINVAL; } DSLAB_UNLOCK(domid); idn_send_slabfree_resp(domid, mtp, slab_offset, slab_size, serrno); } /* * Master -> Slave ONLY */ static void idn_recv_slabfree_resp(int domid, uint_t slab_offset, uint_t slab_size, int serrno) { procname_t proc = "idn_recv_slabfree_resp"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (domid != IDN_GET_MASTERID()) { PR_PROTO("%s: unexpected slabfree resp received (domid = %d)\n", proc, domid); ASSERT(0); return; } if (slab_size > IDN_SLAB_SIZE) { PR_PROTO("%s: unexpected slab size. exp %d, recvd %d\n", proc, IDN_SLAB_SIZE, slab_size); ASSERT(0); return; } PR_PROTO("%s: recvd free resp from dom %d " "- slab (off/size) 0x%x/0x%x [serrno = %d]\n", proc, domid, slab_offset, slab_size, serrno); } static void idn_send_slabfree_resp(int domid, idn_msgtype_t *mtp, uint_t slab_offset, uint_t slab_size, int serrno) { idn_send_cmdresp(domid, mtp, IDNCMD_SLABFREE, slab_offset, slab_size, serrno); } static void idn_retry_nodename_req(void *arg) { int domid = (int)(uintptr_t)arg; idn_send_nodename_req(domid); } static void idn_send_nodename_req(int domid) { caddr_t b_bufp; smr_offset_t bufoffset; int serrno; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_send_nodename_req"; /* * Need to drop domain lock across * SMR allocation. */ serrno = smr_buf_alloc(domid, MAXDNAME+1, &b_bufp); IDN_DLOCK_SHARED(domid); if (dp->dstate != IDNDS_CONNECTED) { /* * Lost connection. */ PR_PROTO("%s:%d: connection lost [dstate = %s]\n", proc, domid, idnds_str[dp->dstate]); IDN_DUNLOCK(domid); if (!serrno) (void) smr_buf_free(domid, b_bufp, MAXDNAME+1); return; } if (serrno) { /* * Failed to allocate buffer, but still have * connection so keep trying. We may have queried * the master a little too earlier. */ PR_PROTO("%s:%d: buffer alloc failed [dstate = %s]\n", proc, domid, idnds_str[dp->dstate]); (void) timeout(idn_retry_nodename_req, (void *)(uintptr_t)domid, hz); IDN_DUNLOCK(domid); return; } *b_bufp = (char)MAXDNAME; bufoffset = IDN_ADDR2OFFSET(b_bufp); idn_send_cmd(domid, IDNCMD_NODENAME, bufoffset, 0, 0); IDN_DUNLOCK(domid); } static void idn_send_nodename_resp(int domid, idn_msgtype_t *mtp, smr_offset_t bufoffset, int serrno) { idn_send_cmdresp(domid, mtp, IDNCMD_NODENAME, (uint_t)bufoffset, 0, serrno); } static void idn_recv_nodename_req(int domid, idn_msgtype_t *mtp, smr_offset_t bufoffset) { caddr_t b_bufp; int length; idn_domain_t *ldp = &idn_domain[idn.localid]; procname_t proc = "idn_recv_nodename_req"; IDN_DLOCK_EXCL(idn.localid); if (!strlen(ldp->dname)) { if (!strlen(utsname.nodename)) { /* * Local domain's nodename hasn't been * set yet. */ IDN_DUNLOCK(idn.localid); idn_send_cmd_nackresp(domid, mtp, IDNCMD_NODENAME, IDNNACK_RETRY); return; } strncpy(ldp->dname, utsname.nodename, MAXDNAME - 1); } IDN_DLOCK_DOWNGRADE(idn.localid); if (!VALID_NWROFFSET(bufoffset, IDN_SMR_BUFSIZE)) { PR_PROTO("%s:%d: invalid SMR offset received (0x%x)\n", proc, domid, bufoffset); IDN_DUNLOCK(idn.localid); idn_send_nodename_resp(domid, mtp, bufoffset, EINVAL); return; } b_bufp = IDN_OFFSET2ADDR(bufoffset); length = (int)(*b_bufp++ & 0xff); if (length < strlen(ldp->dname)) { PR_PROTO("%s:%d: buffer not big enough (req %lu, got %d)\n", proc, domid, strlen(ldp->dname), length); IDN_DUNLOCK(idn.localid); idn_send_nodename_resp(domid, mtp, bufoffset, EINVAL); return; } strncpy(b_bufp, ldp->dname, MAXDNAME); b_bufp[MAXDNAME-1] = 0; IDN_DUNLOCK(idn.localid); idn_send_nodename_resp(domid, mtp, bufoffset, 0); } static void idn_recv_nodename_resp(int domid, smr_offset_t bufoffset, int serrno) { caddr_t b_bufp; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_recv_nodename_resp"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); if (!VALID_NWROFFSET(bufoffset, IDN_SMR_BUFSIZE)) { PR_PROTO("%s:%d: invalid SMR offset received (0x%x)\n", proc, domid, bufoffset); return; } if (serrno == 0) { b_bufp = IDN_OFFSET2ADDR(bufoffset) + 1; b_bufp[MAXDNAME-1] = 0; if (strlen(b_bufp) > 0) { strncpy(dp->dname, b_bufp, MAXDNAME); PR_PROTO("%s:%d: received nodename(%s)\n", proc, domid, dp->dname); } } (void) smr_buf_free(domid, b_bufp - 1, MAXDNAME + 1); } /* * The master allocations the SMR management structures. */ static int idn_master_init() { idn_domain_t *ldp = &idn_domain[idn.localid]; size_t reserved_size = 0; caddr_t reserved_area = NULL; procname_t proc = "idn_master_init"; ASSERT(IDN_GLOCK_IS_EXCL()); ASSERT(IDN_DLOCK_IS_EXCL(idn.localid)); if (idn.mboxarea != NULL) { PR_PROTO("%s: master data already initialized\n", proc); return (0); } PR_PROTO("%s: initializing master data (domid = %d)\n", proc, idn.localid); /* * Reserve an area of the SMR for mailbox usage. * This area is allocated to other domains via * the master. Round it up to IDN_SMR_BUFSIZE multiple. */ reserved_size = IDNROUNDUP(IDN_MBOXAREA_SIZE, IDN_SMR_BUFSIZE); PR_PROTO("%s: reserving %lu bytes for mailbox area\n", proc, reserved_size); #ifdef DEBUG if (reserved_size > (size_t)IDN_SLAB_SIZE) { PR_PROTO("%s: WARNING mbox area (%ld) > slab size (%d)\n", proc, reserved_size, IDN_SLAB_SIZE); } #endif /* DEBUG */ /* * Initialize the pool of slabs and SMR I/O buffers. */ if (smr_slabpool_init(reserved_size, &reserved_area) != 0) { idn_master_deinit(); return (-1); } ASSERT(idn.mboxarea == NULL); ASSERT(reserved_area); bzero(reserved_area, reserved_size); idn.mboxarea = (idn_mboxtbl_t *)reserved_area; ldp->dmbox.m_tbl = IDN_MBOXAREA_BASE(idn.mboxarea, idn.localid); /* * Initialize the SMR pointers in the entire * mailbox table. */ idn_mboxarea_init(idn.mboxarea, IDN_MBOXAREA_SIZE / IDN_MBOXTBL_SIZE); return (0); } static void idn_master_deinit() { idn_domain_t *ldp; smr_slab_t *sp; procname_t proc = "idn_master_deinit"; ASSERT(IDN_GLOCK_IS_EXCL()); ASSERT(IDN_DLOCK_IS_EXCL(idn.localid)); if (idn.mboxarea == NULL) { PR_PROTO("%s: master data already deinitialized\n", proc); return; } ldp = &idn_domain[idn.localid]; PR_PROTO("%s: deinitializing master data (domid = %d)\n", proc, idn.localid); ldp->dmbox.m_tbl = NULL; idn.mboxarea = NULL; /* * Master may still be holding onto slabs of his own. */ DSLAB_LOCK_EXCL(idn.localid); sp = ldp->dslab; ldp->dslab = NULL; ldp->dnslabs = 0; if (sp) smr_slab_free(idn.localid, sp); ldp->dslab_state = DSLAB_STATE_UNKNOWN; DSLAB_UNLOCK(idn.localid); smr_slabpool_deinit(); } static int idn_mark_awol(int domid, clock_t *atime) { clock_t awol; idn_domain_t *dp = &idn_domain[domid]; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_GLOCK_IS_EXCL()); if (!DOMAIN_IN_SET(idn.domset.ds_awol, domid)) { DOMAINSET_ADD(idn.domset.ds_awol, domid); idn.nawols++; } awol = lbolt; if (dp->dawol.a_count++ == 0) dp->dawol.a_time = awol; dp->dawol.a_last = awol; if ((awol - dp->dawol.a_msg) >= (clock_t)(idn_awolmsg_interval * hz)) dp->dawol.a_msg = awol; else awol = 0; *atime = awol; idn_awol_event_set(dp->dhw.dh_boardset); return (dp->dawol.a_count); } void idn_clear_awol(int domid) { idn_domain_t *dp = &idn_domain[domid]; ASSERT(IDN_SYNC_IS_LOCKED()); ASSERT(IDN_GLOCK_IS_EXCL()); if (DOMAIN_IN_SET(idn.domset.ds_awol, domid)) { DOMAINSET_DEL(idn.domset.ds_awol, domid); idn.nawols--; } if (dp->dawol.a_count > 0) { dp->dawol.a_count = 0; dp->dawol.a_last = dp->dawol.a_time; dp->dawol.a_time = 0; dp->dawol.a_msg = 0; idn_awol_event_clear(dp->dhw.dh_boardset); } } /* * A timer expired. */ void idn_timer_expired(void *arg) { idn_domain_t *dp; char *op = "UNKNOWN"; clock_t awol = 0; int awolcount, dcpu, domid; idn_timer_t *tp = (idn_timer_t *)arg; idn_timerq_t *tq = NULL; uint_t token; char dname[MAXDNAME]; procname_t proc = "idn_timer_expired"; STRING(str); tq = tp->t_q; ASSERT(tp->t_domid != IDN_NIL_DOMID); IDN_TIMERQ_LOCK(tq); INUM2STR(tp->t_type, str); if (tp->t_onq == 0) { PR_TIMER("%s: timer CAUGHT TERMINATION (type = %s)\n", proc, str); /* * Timer was dequeued. Somebody is trying * to shut it down. */ IDN_TIMERQ_UNLOCK(tq); return; } IDN_TIMER_DEQUEUE(tq, tp); IDN_TIMERQ_UNLOCK(tq); IDN_SYNC_LOCK(); IDN_DLOCK_EXCL(tp->t_domid); domid = tp->t_domid; dp = &idn_domain[domid]; strcpy(dname, dp->dname); dcpu = dp->dcpu; IDN_TIMER_EXEC(tp); #ifdef DEBUG PR_TIMER("%s:%d: [%s] timer EXPIRED (C=0x%x, P=0x%llx, X=0x%llx)\n", proc, tp->t_domid, str, tp->t_cookie, tp->t_posttime, tp->t_exectime); #endif /* DEBUG */ /* * IMPORTANT: * Each case is responsible for dropping SYNC_LOCK & DLOCK. */ switch (tp->t_type) { case IDNP_DATA: IDN_SYNC_UNLOCK(); /* * Timed out waiting for a data packet response. * We can't close domain since he may just be * temporarily AWOL. * Note that dio and diocheck do not get cleared. * This is taken care of when the domain restarts * or is fatally closed. * We only need a reader lock for this. */ IDN_DLOCK_DOWNGRADE(domid); if (dp->diocheck && dp->dmbox.m_send) { (void) idn_reclaim_mboxdata(domid, 0, -1); if (dp->dio >= IDN_WINDOW_EMAX) { idn_msgtype_t mt; /* * Restart timer for another * go around. */ IDN_MSGTIMER_START(domid, IDNP_DATA, 0, idn_msg_waittime[IDNP_DATA], &mt.mt_cookie); } else { lock_clear(&dp->diocheck); } } IDN_DUNLOCK(domid); break; case IDNP_NEGO: /* * If we're not in a NEGO transition, then * just ignore this timeout. */ if (dp->dxp == &xphase_nego) { uint_t token; IDN_GLOCK_EXCL(); op = "CONNECT"; awolcount = idn_mark_awol(domid, &awol); IDN_GUNLOCK(); idn_nego_cleanup_check(domid, IDN_NIL_DOMID, IDN_NIL_DCPU); IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); token = IDN_RETRY_TOKEN(domid, IDNRETRY_NEGO); idn_retry_submit(idn_retry_nego, NULL, token, idn_msg_retrytime[(int)IDNRETRY_NEGO]); } IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; case IDNP_CMD: /* * Timeouts on commands typically mean that the * the master is not responding. Furthermore, we * can't FORCE a FIN disconnect since at this stage * we are CONNECTED and thus other domains may * have cache entries that we're sharing with them. * Only choice is to completely disconnect from * IDN and try to reestablish connection. * * However, timeouts attempting to get nodename * are not fatal. Although we don't want to retry * either since each timeout is a lost buffer to * the remote domain. */ if (tp->t_subtype == (ushort_t)IDNCMD_NODENAME) { PR_PROTO("%s:%d: timedout waiting for nodename\n", proc, domid); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; } IDN_GLOCK_EXCL(); if (idn.state == IDNGS_ONLINE) { domainset_t domset; int masterid = IDN_GET_MASTERID(); IDN_GKSTAT_GLOBAL_EVENT(gk_reconfigs, gk_reconfig_last); PR_PROTO("%s:%d: RECONFIG trying old masterid = %d\n", proc, domid, masterid); IDN_GSTATE_TRANSITION(IDNGS_RECONFIG); IDN_SET_NEW_MASTERID(masterid); IDN_GUNLOCK(); IDN_DUNLOCK(domid); domset = idn.domset.ds_trans_on | idn.domset.ds_connected; idn_unlink_domainset(domset, IDNFIN_NORMAL, IDNFIN_ARG_NONE, IDNFIN_OPT_RELINK, BOARDSET_ALL); } else { IDN_GUNLOCK(); IDN_DUNLOCK(domid); } IDN_SYNC_UNLOCK(); break; case IDNP_CON: if (tp->t_subtype == (ushort_t)IDNCON_QUERY) { /* * Timed out sending a CON-query. This is * non-fatal. We simply need to retry. */ IDN_GLOCK_EXCL(); op = "CONNECT"; awolcount = idn_mark_awol(domid, &awol); IDN_GUNLOCK(); token = IDN_RETRY_TOKEN(domid, IDNRETRY_CONQ); idn_retry_submit(idn_retry_query, NULL, token, idn_msg_retrytime[(int)IDNRETRY_CONQ]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; } /*FALLTHROUGH*/ case IDNP_CFG: /* * Any timeouts here we simply try to disconnect * and reestablish the link. Since we haven't * reached the connected state w.r.t. this domain * we put his fin state to FORCE-HARD in order * to shoot right through without involving other * domains. Recall that other domains may have * established connections with the given domain * which means any FIN queries to them will always * return connected to the given domain. Since * neither the given domain nor the local domain * plan on disconnecting from the IDN the connection * to the other domains will remain thereby preventing * the local FIN from ever completing. Recall that * a FIN depends on all member domains FIN'ing also. */ IDN_GLOCK_EXCL(); op = "CONNECT"; awolcount = idn_mark_awol(domid, &awol); IDN_GUNLOCK(); DOMAINSET_ADD(idn.domset.ds_relink, domid); IDN_HISTORY_LOG(IDNH_RELINK, domid, dp->dstate, idn.domset.ds_relink); idn_disconnect(domid, IDNFIN_FORCE_SOFT, IDNFIN_ARG_NONE, IDNFIN_SYNC_NO); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; case IDNP_FIN: /* * Timeouts here simply try to retry. */ IDN_GLOCK_EXCL(); op = "DISCONNECT"; awolcount = idn_mark_awol(domid, &awol); IDN_GUNLOCK(); if (tp->t_subtype == (ushort_t)IDNFIN_QUERY) { int d; domainset_t rdyset; /* * Timed out sending a FIN-query. This is * non-fatal. We simply need to retry. * If we were doing a forced unlink of any * domains, we don't want this awol guy * to hold us up. Looks for any forced * unlinks and make them "ready" with * respect to this awol domain. */ rdyset = 0; for (d = 0; d < MAX_DOMAINS; d++) { if (FIN_IS_FORCE(idn_domain[d].dfin)) { DOMAINSET_ADD(rdyset, d); } } if (rdyset) (void) idn_sync_register(domid, IDNSYNC_DISCONNECT, rdyset, IDNSYNC_REG_REG); token = IDN_RETRY_TOKEN(domid, IDNRETRY_FINQ); idn_retry_submit(idn_retry_query, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FINQ]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; } if (dp->dfin == IDNFIN_FORCE_SOFT) { IDN_FSTATE_TRANSITION(dp, IDNFIN_FORCE_HARD); } /* * Anybody that was waiting on this domain and * had a hard-force in action gets this guy for * free in their base ready-set. */ idn_sync_register_awol(domid); dp->dxp = &xphase_fin; IDN_XSTATE_TRANSITION(dp, IDNXS_PEND); token = IDN_RETRY_TOKEN(domid, IDNRETRY_FIN); idn_retry_submit(idn_retry_fin, NULL, token, idn_msg_retrytime[(int)IDNRETRY_FIN]); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; default: ASSERT(0); IDN_DUNLOCK(domid); IDN_SYNC_UNLOCK(); break; } IDN_TIMER_FREE(tp); if (awol) { if (strlen(dname) > 0) { cmn_err(CE_WARN, "IDN: 236: domain (%s) [ID %d] not " "responding to %s [#%d]", dname, domid, op, awolcount); } else { cmn_err(CE_WARN, "IDN: 236: domain [ID %d, CPU %d] not " "responding to %s [#%d]", domid, dcpu, op, awolcount); } } } #if 0 static int idn_retry_check(uint_t token) { int i, count = 0; int domid = IDN_RETRY_TOKEN2DOMID(token); int key = IDN_RETRY_TOKEN2TYPE(token); idn_retry_job_t *rp; idn_retry_queue_t *qp; qp = &idn.retryqueue; mutex_enter(&qp->rq_mutex); for (i = 0, rp = qp->rq_jobs; i < qp->rq_count; i++, rp = rp->rj_next) if ((domid == IDN_RETRY_TOKEN2DOMID(rp->rj_token)) && ((key == IDN_RETRY_TYPEALL) || (rp->rj_token == token))) count++; mutex_exit(&qp->rq_mutex); return (count); } #endif /* 0 */ static void idn_retry_execute(void *arg) { idn_retry_job_t *rp = (idn_retry_job_t *)arg; idn_retry_queue_t *qp; qp = &idn.retryqueue; mutex_enter(&qp->rq_mutex); if (rp->rj_onq == 0) { /* * Job has already been claimed by * retry termination routine. * Bail out. */ mutex_exit(&qp->rq_mutex); return; } rp->rj_next->rj_prev = rp->rj_prev; rp->rj_prev->rj_next = rp->rj_next; if (--(qp->rq_count) == 0) qp->rq_jobs = NULL; else if (qp->rq_jobs == rp) qp->rq_jobs = rp->rj_next; mutex_exit(&qp->rq_mutex); (*rp->rj_func)(rp->rj_token, rp->rj_arg); IDNRETRY_FREEJOB(rp); } /* * */ static void idn_retry_submit(void (*func)(uint_t token, void *arg), void *arg, uint_t token, clock_t ticks) { idn_retry_job_t *rp, *cp; idn_retry_queue_t *qp; int c; procname_t proc = "idn_retry_submit"; if (ticks < 0) { PR_PROTO("%s: (token = 0x%x) WARNING ticks = %ld\n", proc, token, ticks); return; } if (ticks == 0) /* At least one tick to get into background */ ticks++; PR_PROTO("%s: token = 0x%x\n", proc, token); qp = &idn.retryqueue; mutex_enter(&qp->rq_mutex); for (c = 0, cp = qp->rq_jobs; c < qp->rq_count; cp = cp->rj_next, c++) { if (cp->rj_token == token) { PR_PROTO("%s: token = (%d,0x%x) already present\n", proc, IDN_RETRY_TOKEN2DOMID(token), IDN_RETRY_TOKEN2TYPE(token)); break; } } if (c < qp->rq_count) { mutex_exit(&qp->rq_mutex); return; } rp = IDNRETRY_ALLOCJOB(); rp->rj_func = func; rp->rj_arg = arg; rp->rj_token = token; rp->rj_prev = rp->rj_next = rp; if (qp->rq_jobs == NULL) { qp->rq_jobs = rp; } else { rp->rj_next = qp->rq_jobs; rp->rj_prev = qp->rq_jobs->rj_prev; rp->rj_next->rj_prev = rp; rp->rj_prev->rj_next = rp; } rp->rj_onq = 1; qp->rq_count++; rp->rj_id = timeout(idn_retry_execute, (caddr_t)rp, ticks); mutex_exit(&qp->rq_mutex); } int idn_retry_terminate(uint_t token) { int i, domid; uint_t key, count; idn_retry_job_t *rp, *nrp, *fp; idn_retry_queue_t *qp; procname_t proc = "idn_retry_terminate"; key = IDN_RETRY_TOKEN2TYPE(token); domid = IDN_RETRY_TOKEN2DOMID(token); fp = NULL; qp = &idn.retryqueue; mutex_enter(&qp->rq_mutex); for (i = count = 0, rp = qp->rq_jobs; i < qp->rq_count; i++) { nrp = rp->rj_next; if ((domid == IDN_RETRY_TOKEN2DOMID(rp->rj_token)) && ((key == IDN_RETRY_TYPEALL) || (rp->rj_token == token))) { /* * Turn off onq field as a signal to * the execution routine that this * retry has been terminated. This * is necessary since we can't untimeout * while holding the rq_mutex otherwise * we'll deadlock with the execution * routine. We'll untimeout these guys * _after_ we drop rq_mutex. */ rp->rj_onq = 0; rp->rj_next->rj_prev = rp->rj_prev; rp->rj_prev->rj_next = rp->rj_next; if (qp->rq_jobs == rp) qp->rq_jobs = rp->rj_next; rp->rj_next = fp; fp = rp; count++; } rp = nrp; } if ((qp->rq_count -= count) == 0) qp->rq_jobs = NULL; mutex_exit(&qp->rq_mutex); PR_PROTO("%s: token = (%d,0x%x), dequeued = %d\n", proc, domid, key, count); for (; fp; fp = nrp) { (void) untimeout(fp->rj_id); nrp = fp->rj_next; IDNRETRY_FREEJOB(fp); } return (count); } /* * ----------------------------------------------------------------------- * The sole purpose of the idn_protocol_server is to manage the IDN * protocols between the various domains. These messages do _not_ go * through the regular streams queues since they are not dependent on * any user process or module necessarily having the IDN driver open. * There may be multiple instances of these servers to enhance performance * of domain management. Each server is assigned a idn_protoqueue_t * from which to obtain the work they need to do. * ----------------------------------------------------------------------- */ int idn_protocol_init(int nservers) { int i; idn_protojob_t *jp; register idn_protoqueue_t *protoq; if (nservers <= 0) { cmn_err(CE_WARN, "IDN: 237: invalid number (%d) of protocol servers", nservers); return (-1); } idn.protocol.p_jobpool = kmem_cache_create("idn_protocol_jobcache", sizeof (idn_protojob_t), 0, NULL, NULL, NULL, NULL, NULL, 0); if (idn.protocol.p_jobpool == NULL) { cmn_err(CE_WARN, "IDN: 238: kmem_cache_create(jobcache) failed"); return (-1); } /* * Initialize static cache for protojob. */ mutex_init(&idn_protojob_cache_lock, NULL, MUTEX_DRIVER, NULL); jp = &idn_protojob_cache[0]; for (i = 1; i < IDN_DMV_PENDING_MAX; jp = jp->j_next, i++) { jp->j_cache = 1; jp->j_next = &idn_protojob_cache[i]; } jp->j_cache = 1; jp->j_next = NULL; idn_protojob_cache_list = &idn_protojob_cache[0]; /* * Init morgue semaphore. */ sema_init(&idn.protocol.p_morgue, 0, NULL, SEMA_DEFAULT, NULL); /* * Alloc server queues. */ idn.protocol.p_serverq = GETSTRUCT(idn_protoqueue_t, nservers); /* * Init server queues. */ protoq = idn.protocol.p_serverq; for (i = 0; i < nservers; protoq++, i++) { mutex_init(&protoq->q_mutex, NULL, MUTEX_DRIVER, NULL); cv_init(&protoq->q_cv, NULL, CV_DEFAULT, NULL); protoq->q_id = i; protoq->q_joblist = NULL; protoq->q_joblist_tail = NULL; protoq->q_die = 0; protoq->q_morgue = &idn.protocol.p_morgue; /* * Create protocol server thread. */ protoq->q_threadp = thread_create(NULL, 0, idn_protocol_server, (caddr_t)&i, sizeof (i), &p0, TS_RUN, maxclsyspri); } /* * The servers are kept in the p_server[] array, however * we'll build a linked list of them to facilitate debugging. */ protoq = idn.protocol.p_serverq; for (i = 0; i < (nservers - 1); protoq++, i++) protoq->q_next = (protoq + 1); protoq->q_next = NULL; idn.nservers = nservers; return (idn.nservers); } void idn_protocol_deinit() { register int i; int nservers; register idn_protoqueue_t *protoq; nservers = idn.nservers; if (nservers <= 0) return; /* * Make sure the servers are dead. */ idn_protocol_server_killall(); ASSERT(idn.nservers == 0); /* * Destroy the mutexes. */ protoq = idn.protocol.p_serverq; for (i = 0; i < nservers; protoq++, i++) { mutex_destroy(&protoq->q_mutex); cv_destroy(&protoq->q_cv); } /* * Free up the protoqueue memory. */ FREESTRUCT(idn.protocol.p_serverq, idn_protoqueue_t, nservers); idn.protocol.p_serverq = NULL; /* * Destroy the morgue semaphore. */ sema_destroy(&idn.protocol.p_morgue); if (idn.protocol.p_jobpool) { kmem_cache_destroy(idn.protocol.p_jobpool); idn.protocol.p_jobpool = NULL; } } static void idn_protocol_server(int *id) { idn_protoqueue_t *pq; idn_protojob_t *jl; register idn_protojob_t *jp; procname_t proc = "idn_protocol_server"; if (id == NULL) { PR_PROTO("%s: id == NULL, thread exiting\n", proc); return; } ASSERT((*id >= 0) && (*id < idn_protocol_nservers)); pq = &idn.protocol.p_serverq[*id]; ASSERT(pq->q_id == *id); PR_PROTO("%s: id %d starting up (pq = 0x%p)\n", proc, pq->q_id, pq); /*CONSTCOND*/ while (1) { mutex_enter(&pq->q_mutex); while (((jl = pq->q_joblist) == NULL) && !pq->q_die) cv_wait(&pq->q_cv, &pq->q_mutex); pq->q_joblist = pq->q_joblist_tail = NULL; if (pq->q_die) { /* * We've been killed. Need to check-in * at the morgue. */ pq->q_threadp = NULL; mutex_exit(&pq->q_mutex); PR_PROTO("%s: thread (%d) killed...bye bye\n", proc, pq->q_id); for (jp = jl; jp; jp = jl) { jl = jp->j_next; idn_protojob_free(jp); } sema_v(pq->q_morgue); thread_exit(); /*NOTREACHED*/ } mutex_exit(&pq->q_mutex); /* * We can process the jobs asynchronously while more are * put on. */ for (jp = jl; jp; jp = jl) { jl = jp->j_next; idn_recv_proto(&(jp->j_msg)); idn_protojob_free(jp); } } } /* * Kill off all the protocol servers. */ static void idn_protocol_server_killall() { register idn_protoqueue_t *pq; int i; procname_t proc = "idn_protocol_server_killall"; PR_PROTO("%s: killing off %d protocol servers\n", proc, idn.nservers); pq = idn.protocol.p_serverq; for (i = 0; i < idn.nservers; pq++, i++) { mutex_enter(&pq->q_mutex); pq->q_die = 1; cv_signal(&pq->q_cv); mutex_exit(&pq->q_mutex); } while (idn.nservers > 0) { sema_p(&idn.protocol.p_morgue); idn.nservers--; } } idn_protojob_t * idn_protojob_alloc(int kmflag) { idn_protojob_t *jp; jp = kmem_cache_alloc(idn.protocol.p_jobpool, kmflag); if (jp == NULL) { mutex_enter(&idn_protojob_cache_lock); if ((jp = idn_protojob_cache_list) != NULL) idn_protojob_cache_list = jp->j_next; mutex_exit(&idn_protojob_cache_lock); } else { jp->j_cache = 0; } return (jp); } static void idn_protojob_free(idn_protojob_t *jp) { ASSERT(jp); if (jp->j_cache) { mutex_enter(&idn_protojob_cache_lock); jp->j_next = idn_protojob_cache_list; idn_protojob_cache_list = jp; mutex_exit(&idn_protojob_cache_lock); } else { kmem_cache_free(idn.protocol.p_jobpool, (void *)jp); } } void idn_protojob_submit(int cookie, idn_protojob_t *jp) { idn_protoqueue_t *pq; int serverid; procname_t proc = "idn_protojob_submit"; STRING(str); if (jp == NULL) return; serverid = IDN_PROTOCOL_SERVER_HASH(cookie); pq = &idn.protocol.p_serverq[serverid]; INUM2STR(jp->j_msg.m_msgtype, str); PR_PROTO("%s: job (d=%d, m=0x%x, %s) submitted to " "protocol server %d\n", proc, jp->j_msg.m_domid, jp->j_msg.m_msgtype, str, serverid); mutex_enter(&pq->q_mutex); /* * Can't submit jobs to dying servers. */ if (!pq->q_die) { if (pq->q_joblist_tail) { pq->q_joblist_tail->j_next = jp; pq->q_joblist_tail = jp; } else { pq->q_joblist = pq->q_joblist_tail = jp; } jp->j_next = NULL; cv_signal(&pq->q_cv); } else { PR_PROTO("%s: protocol server dead. freeing protojob\n", proc); idn_protojob_free(jp); } mutex_exit(&pq->q_mutex); } static void idn_mboxarea_init(idn_mboxtbl_t *mtp, register int ntbls) { register int d; caddr_t state_ptr = NULL, mtbasep = (caddr_t)mtp; idn_mboxtbl_t *amtp; procname_t proc = "idn_mboxarea_init"; ASSERT(mtp && (ntbls > 0)); PR_PROTO("%s: init mboxtbl (0x%p) ntbls = %d\n", proc, mtp, ntbls); for (d = 0; d < ntbls; d++) { register int pd, sd; register int ch; mtp->mt_header.mh_svr_active = 0; mtp->mt_header.mh_svr_ready = 0; /* * Initialize the header of each mbox table * with a cookie for identity. */ /* * Format: 0xc0c0DSCC * D = primary domain * S = sub-domain of primary * CC = channel of sub-domain. */ pd = (d / MAX_DOMAINS) / IDN_MAX_NETS; sd = (d / IDN_MAX_NETS) % MAX_DOMAINS; ch = d % IDN_MAX_NETS; /* * We point all sub-domains in the same channel * to the same active sync flag since a single server * services all domains in the same channel. */ amtp = IDN_MBOXTBL_ABS_PTR(mtbasep, pd, 0, ch); state_ptr = (caddr_t)&amtp->mt_header.mh_svr_active; mtp->mt_header.mh_svr_active_ptr = IDN_ADDR2OFFSET(state_ptr); state_ptr = (caddr_t)&amtp->mt_header.mh_svr_ready; mtp->mt_header.mh_svr_ready_ptr = IDN_ADDR2OFFSET(state_ptr); mtp->mt_header.mh_cookie = IDN_MAKE_MBOXHDR_COOKIE(pd, sd, ch); mtp->mt_header.mh_cksum = IDN_CKSUM_MBOX(&mtp->mt_header); IDN_MBOXTBL_PTR_INC(mtp); } /* * Now that the master has initialized the entire mailbox * region the referenced memory may not necessarily be up-to-date * with respect to the actual SMR memory due to caching. * In order to make sure future connecting domains get a * consistent picture of the mailbox region, it's necessary * for the master to flush its caches. */ PR_PROTO("%s: flushing ecache's of local (master) domain\n", proc); idnxf_flushall_ecache(); } idn_mainmbox_t * idn_mainmbox_init(int domid, int mbx) { idn_mainmbox_t *mmp; int c; idn_mainmbox_t *cmp; procname_t proc = "idn_mainmbox_init"; ASSERT(idn_domain[domid].dcpu != IDN_NIL_DCPU); ASSERT(IDN_DLOCK_IS_HELD(domid)); PR_PROTO("%s: initializing main %s mailbox for domain %d\n", proc, IDNMBOX_IS_RECV(mbx) ? "RECV" : "SEND", domid); cmp = GETSTRUCT(idn_mainmbox_t, IDN_MAX_NETS); for (c = 0; c < IDN_MAX_NETS; c++) { mmp = &cmp[c]; mmp->mm_channel = (short)c; mutex_init(&mmp->mm_mutex, NULL, MUTEX_DRIVER, NULL); mmp->mm_domid = (short)domid; mmp->mm_type = mbx; } mmp = cmp; /* * The actual SMR mailbox (mmp->mm_smr_mboxp) gets setup * when the SMR is setup. */ return (mmp); } static void idn_mainmbox_reset(int domid, idn_mainmbox_t *cmp) { idn_mainmbox_t *mmp; int c; procname_t proc = "idn_mainmbox_reset"; ASSERT(IDN_DLOCK_IS_EXCL(domid)); PR_PROTO("%s: reseting main %s mailbox for domain %d\n", proc, IDNMBOX_IS_RECV(cmp->mm_type) ? "RECV" : "SEND", domid); for (c = 0; c < IDN_MAX_NETS; c++) { mmp = &cmp[c]; mmp->mm_channel = (short)c; mmp->mm_domid = (short)domid; mmp->mm_count = 0; mmp->mm_flags = 0; mmp->mm_qiget = mmp->mm_qiput = 0; mmp->mm_csp = NULL; ASSERT(mmp->mm_type == cmp->mm_type); } } void idn_mainmbox_deinit(int domid, idn_mainmbox_t *mmp) { procname_t proc = "idn_mainmbox_deinit"; ASSERT(IDN_DLOCK_IS_HELD(domid)); PR_PROTO("%s: deinitializing main %s mailbox for domain %d\n", proc, IDNMBOX_IS_RECV(mmp->mm_type) ? "RECV" : "SEND", domid); ASSERT(idn_domain_is_registered(domid, -1, NULL) == 0); FREESTRUCT(mmp, idn_mainmbox_t, IDN_MAX_NETS); } static void idn_mainmbox_activate(int domid) { register int c; idn_domain_t *dp = &idn_domain[domid]; procname_t proc = "idn_mainmbox_activate"; ASSERT(IDN_DLOCK_IS_HELD(domid)); PR_PROTO("%s:%d: activating main mailbox\n", proc, domid); for (c = 0; c < IDN_MAX_NETS; c++) idn_mainmbox_chan_register(domid, &dp->dmbox.m_send[c], &dp->dmbox.m_recv[c], c); } /* * Called upon disabling the SMR to deactivate all the mailboxes * so that they no longer reference the SMR that's going away. * * stopall - Indicates to stop all channel services, across the board. */ static void idn_mainmbox_deactivate(ushort_t domset) { int svr_count; procname_t proc = "idn_mainmbox_deactivate"; if (domset == 0) return; PR_PROTO("%s: %s deactivating main mailboxes for domset 0x%x\n", proc, (domset == (ushort_t)-1) ? "STOP-ALL" : "NORMAL", domset); svr_count = idn_mainmbox_chan_unregister(domset, -1); PR_PROTO("%s: deactivated %d chansvrs (domset 0x%x)\n", proc, svr_count, domset); } static void idn_mainmbox_chan_register(int domid, idn_mainmbox_t *send_mmp, idn_mainmbox_t *recv_mmp, int channel) { ASSERT(IDN_DLOCK_IS_HELD(domid)); /* * Obtain receive mailbox lock first. */ mutex_enter(&recv_mmp->mm_mutex); mutex_enter(&send_mmp->mm_mutex); ASSERT(recv_mmp->mm_channel == (short)channel); ASSERT(send_mmp->mm_channel == (short)channel); recv_mmp->mm_csp = &idn.chan_servers[channel]; recv_mmp->mm_count = 0; recv_mmp->mm_dropped = 0; recv_mmp->mm_flags = 0; send_mmp->mm_csp = &idn.chan_servers[channel]; send_mmp->mm_count = 0; send_mmp->mm_dropped = 0; send_mmp->mm_flags = 0; mutex_exit(&send_mmp->mm_mutex); mutex_exit(&recv_mmp->mm_mutex); /* * We have to add ourselves to the respective * channel server's service table. * Note that the channel may not necessarily be * active at this time. */ ASSERT(idn.chan_servers); /* * Have to get the channel server under * control so we can add ourselves. * Returns w/c_mutex. */ IDN_CHAN_LOCK_GLOBAL(&idn.chan_servers[channel]); /* * Add the following domain (mailbox) for monitoring * by the respective channel server. */ idn_chan_addmbox(channel, DOMAINSET(domid)); IDN_CHAN_UNLOCK_GLOBAL(&idn.chan_servers[channel]); } /* * Unregister the given domain from the specified channel(s) for monitoring. */ static int idn_mainmbox_chan_unregister(ushort_t domset, int channel) { int c, dd_count; int min_chan, max_chan; procname_t proc = "idn_mainmbox_chan_unregister"; PR_CHAN("%s: deactivating main mailboxes (channel %d) " "for domset 0x%x\n", proc, channel, domset); if (channel == -1) { min_chan = 0; max_chan = IDN_MAX_NETS - 1; } else { min_chan = max_chan = channel; } /* * Point all the data dispatchers to the same morgue * so we can kill them all at once. */ dd_count = 0; for (c = min_chan; c <= max_chan; c++) { /* * Have to get the channel server under * control so we can remove ourselves. * Returns w/c_mutex held. */ IDN_CHAN_LOCK_GLOBAL(&idn.chan_servers[c]); /* * Delete the following domain (mailbox) from * monitoring by the respective channel server. */ idn_chan_delmbox(c, (ushort_t)domset); IDN_CHAN_UNLOCK_GLOBAL(&idn.chan_servers[c]); dd_count++; } PR_CHAN("%s: deactivated %d channel mboxes for domset 0x%x, chan %d\n", proc, dd_count, domset, channel); return (dd_count); } /* * Check if the given domain is registered with the given channel(s). */ int idn_domain_is_registered(int domid, int channel, idn_chanset_t *chansetp) { int regcount; int c, min_chan, max_chan; idn_chanset_t chanset; procname_t proc = "idn_domain_is_registered"; CHANSET_ZERO(chanset); if (idn.chan_servers == NULL) { PR_CHAN("%s: idn.chan_servers == NULL!!\n", proc); return (0); } if (channel == -1) { min_chan = 0; max_chan = IDN_MAX_NETS - 1; } else { min_chan = max_chan = channel; } regcount = 0; for (c = min_chan; c <= max_chan; c++) { idn_chansvr_t *csp; csp = &idn.chan_servers[c]; IDN_CHAN_LOCK_SEND(csp); /* * Don't really need recv side lock since registeration * can't change while we're holding send side. * No need to wait for send side to actually suspend * since all we want to do is prevent the registered * information from changing. */ if (IDN_CHAN_DOMAIN_IS_REGISTERED(csp, domid)) { regcount++; CHANSET_ADD(chanset, c); } IDN_CHAN_UNLOCK_SEND(csp); } PR_CHAN("%s: domid %d mbox reg'd with %d channels [0x%x] (req=%d)\n", proc, domid, regcount, chanset, channel); if (chansetp) *chansetp = chanset; return (regcount); } static int idn_mainmbox_flush(int domid, idn_mainmbox_t *mmp) { register int qi; register idn_mboxmsg_t *mqp; int total_count = 0; int c, count; int mbox_type; char *mbox_str; int lost_io, total_lost_io = 0; idn_chanset_t chanset; procname_t proc = "idn_mainmbox_flush"; if (mmp == NULL) return (0); CHANSET_ZERO(chanset); mbox_type = mmp->mm_type; ASSERT((mbox_type == IDNMMBOX_TYPE_SEND) || (mbox_type == IDNMMBOX_TYPE_RECV)); mbox_str = (mbox_type == IDNMMBOX_TYPE_SEND) ? "SEND" : "RECV"; /* * Determine which channels this domain is registered * with. If he's not registered with any, then we * can't touch the SMR. */ (void) idn_domain_is_registered(domid, -1, &chanset); for (c = 0; c < IDN_MAX_NETS; c++) { ushort_t mbox_csum; if (mmp[c].mm_smr_mboxp == NULL) continue; mutex_enter(&mmp[c].mm_mutex); ASSERT(mmp[c].mm_type == mbox_type); if (CHAN_IN_SET(chanset, c) == 0) { /* * Domain is no longer registered. * DON'T TOUCH THE SMR - IT'S POISON! */ if (mmp[c].mm_smr_mboxp) { PR_CHAN("%s:%d:%s: domain unregistered " "w/chan %d - DUMPING SMR reference\n", proc, domid, mbox_str, c); lost_io = IDN_MMBOXINDEX_DIFF(mmp[c].mm_qiput, mmp[c].mm_qiget); #ifdef DEBUG if (mbox_type == IDNMMBOX_TYPE_RECV) { PR_CHAN("%s:%d:%s: blowing away %d " "incoming pkts\n", proc, domid, mbox_str, lost_io); } else { PR_CHAN("%s:%d:%s: blowing away %d/%d " "outstanding pkts\n", proc, domid, mbox_str, lost_io, idn_domain[domid].dio); } #endif /* DEBUG */ } mmp[c].mm_qiput = mmp[c].mm_qiget = 0; mmp[c].mm_smr_mboxp = NULL; total_lost_io += lost_io; } if (mmp[c].mm_smr_mboxp) { mbox_csum = IDN_CKSUM_MBOX(&mmp[c].mm_smr_mboxp->mt_header); if (!VALID_NWRADDR(mmp[c].mm_smr_mboxp, 4) || !VALID_MBOXHDR(&mmp[c].mm_smr_mboxp->mt_header, c, mbox_csum)) { lost_io = IDN_MMBOXINDEX_DIFF(mmp[c].mm_qiput, mmp[c].mm_qiget); #ifdef DEBUG if (mbox_type == IDNMMBOX_TYPE_RECV) { PR_CHAN("%s:%d:%s: bad mbox. blowing " "away %d incoming pkts\n", proc, domid, mbox_str, lost_io); } else { PR_CHAN("%s:%d:%s: bad mbox. blowing " "away %d/%d outstanding pkts\n", proc, domid, mbox_str, lost_io, idn_domain[domid].dio); } #endif /* DEBUG */ mmp[c].mm_smr_mboxp = NULL; mmp[c].mm_qiput = mmp[c].mm_qiget = 0; total_lost_io += lost_io; } } if (mmp[c].mm_smr_mboxp == NULL) { mutex_exit(&mmp[c].mm_mutex); continue; } mqp = &mmp[c].mm_smr_mboxp->mt_queue[0]; qi = 0; count = 0; /* * It's quite possible the remote domain may be accessing * these mailbox entries at the exact same time we're * clearing the owner bit. That's okay. All we're trying * to do at this point is to minimize the number of packets * the remote domain might try to process unnecessarily. */ do { if (mqp[qi].ms_owner) count++; mqp[qi].ms_owner = 0; IDN_MMBOXINDEX_INC(qi); } while (qi); lost_io = IDN_MMBOXINDEX_DIFF(mmp[c].mm_qiput, mmp[c].mm_qiget); total_lost_io += lost_io; mmp[c].mm_qiput = mmp[c].mm_qiget = 0; mmp[c].mm_smr_mboxp = NULL; mutex_exit(&mmp[c].mm_mutex); total_count += count; PR_CHAN("%s:%d:%s: flushed out %d mbox entries for chan %d\n", proc, domid, mbox_str, count, c); } if (total_lost_io && (mbox_type == IDNMMBOX_TYPE_SEND)) { int lost_bufs; /* * If we lost all our outstanding I/O. We could * possible could have slabs now with mistakenly * outstanding I/O buffers. Need to clean them up. * Clean up of leftovers our self. */ lost_bufs = smr_buf_free_all(domid); PR_CHAN("%s:%d:%s: flushed %d/%d buffers from slabs\n", proc, domid, mbox_str, lost_bufs, total_lost_io); } PR_CHAN("%s:%d:%s: flushed total of %d mailbox entries (lost %d)\n", proc, domid, mbox_str, total_count, total_lost_io); return (total_count); } void idn_chanserver_bind(int net, int cpuid) { int ocpuid; cpu_t *cp; idn_chansvr_t *csp; kthread_id_t tp; procname_t proc = "idn_chanserver_bind"; csp = &idn.chan_servers[net]; IDN_CHAN_LOCK_GLOBAL(csp); mutex_enter(&cpu_lock); /* protect checking cpu_ready_set */ ocpuid = csp->ch_bound_cpuid; cp = cpu_get(cpuid); if ((cpuid != -1) && ((cp == NULL) || !cpu_is_online(cp))) { mutex_exit(&cpu_lock); cmn_err(CE_WARN, "IDN: 239: invalid CPU ID (%d) specified for " "IDN net %d", cpuid, net); IDN_CHAN_UNLOCK_GLOBAL(csp); return; } if ((tp = csp->ch_recv_threadp) == NULL) { /* * Thread is not yet active. Set ch_bound_cpuid * so when thread activates it will automatically * bind itself. */ csp->ch_bound_cpuid = -1; csp->ch_bound_cpuid_pending = cpuid; } else { if (ocpuid != -1) { thread_affinity_clear(tp); csp->ch_bound_cpuid = -1; } if (cpuid >= 0) { thread_affinity_set(tp, cpuid); csp->ch_bound_cpuid = cpuid; } csp->ch_bound_cpuid_pending = -1; } mutex_exit(&cpu_lock); PR_CHAN("%s: bound net/channel (%d) from cpuid %d to%scpuid %d\n", proc, net, ocpuid, tp ? " " : " (pending) ", cpuid); IDN_CHAN_UNLOCK_GLOBAL(csp); } #ifdef DEBUG static idn_mboxhdr_t *prev_mhp[IDN_MAXMAX_NETS]; #endif /* DEBUG */ /* * Get access to the respective channel server's synchronization * header which resides in SMR space. */ static idn_mboxhdr_t * idn_chan_server_syncheader(int channel) { idn_domain_t *ldp = &idn_domain[idn.localid]; idn_mboxtbl_t *mtp; idn_mboxhdr_t *mhp; ushort_t mbox_csum; procname_t proc = "idn_chan_server_syncheader"; ASSERT(IDN_CHAN_RECV_IS_LOCKED(&idn.chan_servers[channel])); IDN_DLOCK_SHARED(idn.localid); if (ldp->dmbox.m_tbl == NULL) { PR_CHAN("%s: local dmbox.m_tbl == NULL\n", proc); IDN_DUNLOCK(idn.localid); return (NULL); } mtp = IDN_MBOXTBL_PTR_CHAN(ldp->dmbox.m_tbl, channel); mhp = &mtp->mt_header; mbox_csum = IDN_CKSUM_MBOX(&mtp->mt_header); #ifdef DEBUG if (mhp != prev_mhp[channel]) { prev_mhp[channel] = mhp; PR_CHAN("%s: chan_server (%d) cookie = 0x%x (exp 0x%x)\n", proc, channel, IDN_GET_MBOXHDR_COOKIE(mhp), IDN_MAKE_MBOXHDR_COOKIE(0, 0, channel)); PR_CHAN("%s: chan_server (%d) actv_ptr = 0x%x (exp 0x%x)\n", proc, channel, mhp->mh_svr_active_ptr, IDN_ADDR2OFFSET(&mhp->mh_svr_active)); PR_CHAN("%s: chan_server (%d) ready_ptr = 0x%x (exp 0x%x)\n", proc, channel, mhp->mh_svr_ready_ptr, IDN_ADDR2OFFSET(&mhp->mh_svr_ready)); PR_CHAN("%s: chan_server (%d) mbox_cksum = 0x%x (exp 0x%x)\n", proc, channel, (int)mhp->mh_cksum, (int)mbox_csum); } #endif /* DEBUG */ if ((IDN_ADDR2OFFSET(&mhp->mh_svr_active) != mhp->mh_svr_active_ptr) || (IDN_ADDR2OFFSET(&mhp->mh_svr_ready) != mhp->mh_svr_ready_ptr) || !VALID_MBOXHDR(mhp, channel, mbox_csum)) { idn_chansvr_t *csp; csp = &idn.chan_servers[channel]; if (IDN_CHANNEL_IS_RECV_CORRUPTED(csp) == 0) { IDN_CHANSVC_MARK_RECV_CORRUPTED(csp); cmn_err(CE_WARN, "IDN: 240: (channel %d) SMR CORRUPTED " "- RELINK", channel); cmn_err(CE_CONT, "IDN: 240: (channel %d) cookie " "(expected 0x%x, actual 0x%x)\n", channel, IDN_MAKE_MBOXHDR_COOKIE(0, 0, channel), mhp->mh_cookie); cmn_err(CE_CONT, "IDN: 240: (channel %d) actv_flg " "(expected 0x%x, actual 0x%x)\n", channel, mhp->mh_svr_active_ptr, IDN_ADDR2OFFSET(&mhp->mh_svr_active)); cmn_err(CE_CONT, "IDN: 240: (channel %d) ready_flg " "(expected 0x%x, actual 0x%x)\n", channel, mhp->mh_svr_ready_ptr, IDN_ADDR2OFFSET(&mhp->mh_svr_ready)); } mhp = NULL; } IDN_DUNLOCK(idn.localid); PR_CHAN("%s: channel(%d) mainhp = 0x%p\n", proc, channel, mhp); return (mhp); } #define CHANSVR_SYNC_CACHE(csp, mmp, chan) \ { \ ASSERT(IDN_CHAN_RECV_IS_LOCKED(csp)); \ if ((csp)->ch_recv_changed) { \ register int _d; \ (csp)->ch_recv_scanset = (csp)->ch_recv_scanset_pending; \ (csp)->ch_recv_domset = (csp)->ch_recv_domset_pending; \ for (_d = 0; _d < MAX_DOMAINS; _d++) { \ if (DOMAIN_IN_SET((csp)->ch_recv_domset, _d)) { \ (mmp)[_d] = \ &idn_domain[_d].dmbox.m_recv[chan]; \ } else { \ (mmp)[_d] = NULL; \ } \ } \ (csp)->ch_recv_changed = 0; \ } \ } #define CHANSVR_NEXT_DOMID(csp, i, d) \ { \ (i) = ((i) + 1) & (MAX_DOMAINS - 1); \ (d) = (int)(((csp)->ch_recv_scanset >> ((i) << 2)) & 0xf); \ } #define CHANSVR_RESET_INDEX(i) ((i) = -1) #ifdef DEBUG static idn_mainmbox_t *Mmp[IDN_MAXMAX_NETS][MAX_DOMAINS]; #endif /* DEBUG */ static void idn_chan_server(idn_chansvr_t **cspp) { idn_mboxhdr_t *mainhp; register idn_chansvr_t *csp; register idn_mboxmsg_t *mqp; #ifdef DEBUG idn_mainmbox_t **mmp; #else idn_mainmbox_t *mmp[MAX_DOMAINS]; #endif /* DEBUG */ register int qi; struct idn *sip; int channel; int cpuid; int empty; int tot_pktcount, tot_dropcount; register int index; register int domid; register int idleloops; procname_t proc = "idn_chan_server"; #ifdef DEBUG mmp = &Mmp[(*cspp)->ch_id][0]; bzero(mmp, MAX_DOMAINS * sizeof (idn_mainmbox_t *)); #else /* DEBUG */ bzero(mmp, sizeof (mmp)); #endif /* DEBUG */ tot_pktcount = tot_dropcount = 0; ASSERT(cspp && *cspp); csp = *cspp; channel = csp->ch_id; sip = IDN_INST2SIP(channel); ASSERT(sip); PR_CHAN("%s: CHANNEL SERVER (channel %d) GOING ACTIVE...\n", proc, channel); IDN_CHAN_LOCK_RECV(csp); IDN_CHAN_RECV_INPROGRESS(csp); ASSERT(csp->ch_recv_threadp == curthread); mutex_enter(&cpu_lock); if ((cpuid = csp->ch_bound_cpuid_pending) != -1) { cpu_t *cp = cpu_get(cpuid); /* * We've been requested to bind to * a particular cpu. */ if ((cp == NULL) || !cpu_is_online(cp)) { /* * Cpu seems to have gone away or gone offline * since originally requested. */ mutex_exit(&cpu_lock); cmn_err(CE_WARN, "IDN: 239: invalid CPU ID (%d) specified for " "IDN net %d", cpuid, channel); } else { csp->ch_bound_cpuid = cpuid; affinity_set(csp->ch_bound_cpuid); mutex_exit(&cpu_lock); } csp->ch_bound_cpuid_pending = -1; } else { mutex_exit(&cpu_lock); } if (csp->ch_bound_cpuid != -1) { PR_CHAN("%s: thread bound to cpuid %d\n", proc, csp->ch_bound_cpuid); } /* * Only the first (main) mbox header is used for * synchronization with data delivery since there is * only data server for all mailboxes for this * given channel. */ CHANSVR_SYNC_CACHE(csp, mmp, channel); mainhp = ((csp->ch_recv_domcount > 0) && IDN_CHANNEL_IS_RECV_ACTIVE(csp)) ? idn_chan_server_syncheader(channel) : NULL; if (mainhp && IDN_CHANNEL_IS_RECV_ACTIVE(csp)) mainhp->mh_svr_active = 1; ASSERT(csp->ch_recv_domcount ? (csp->ch_recv_scanset && csp->ch_recv_domset) : 1); IDN_CHAN_UNLOCK_RECV(csp); empty = 0; idleloops = 0; CHANSVR_RESET_INDEX(index); /* * --------------------------------------------- */ /*CONSTCOND*/ while (1) { register int pktcount; register int dropcount; ushort_t mbox_csum; idn_mboxtbl_t *smr_mboxp; /* points to SMR space */ register smr_offset_t bufoffset; #ifdef DEBUG register smr_pkthdr_t *hdrp; idn_netaddr_t netaddr; #endif /* DEBUG */ /* * Speed through and find the next available domid. */ CHANSVR_NEXT_DOMID(csp, index, domid); if (!index) { /* * We only check state changes when * we wrap around. Done for performance. */ if (!IDN_CHANNEL_IS_RECV_ACTIVE(csp) || csp->ch_recv.c_checkin || (idn.state != IDNGS_ONLINE)) { PR_DATA("%s: (channel %d) %s\n", proc, channel, IDN_CHANNEL_IS_DETACHED(csp) ? "DEAD" : IDN_CHANNEL_IS_PENDING(csp) ? "IDLED" : IDN_CHANNEL_IS_ACTIVE(csp) ? "ACTIVE" : "DISABLED"); goto cc_sleep; } } if (csp->ch_recv.c_checkin) goto cc_sleep; if (empty == csp->ch_recv_domcount) { empty = 0; goto cc_slowdown; } ASSERT(mmp[domid] != NULL); mutex_enter(&mmp[domid]->mm_mutex); if ((smr_mboxp = mmp[domid]->mm_smr_mboxp) == NULL) { /* * Somebody is trying to shut things down. */ empty++; mutex_exit(&mmp[domid]->mm_mutex); continue; } ASSERT(mmp[domid]->mm_channel == (short)channel); /* * We don't care if the mm_smr_mboxp is nullified * after this point. The thread attempting to shut * us down has to formally pause this channel before * anything is official anyway. So, we can continue * with our local SMR reference until the thread * shutting us down really stops us. * * Need to get the qiget index _before_ we drop the * lock since it might get flushed (idn_mainmbox_flush) * once we drop the mm_mutex. * * We prefer not to hold the mm_mutex across the * idn_recv_mboxdata() call since that may be time- * consuming. */ qi = mmp[domid]->mm_qiget; /* * Check the mailbox header if checksum is turned on. */ mbox_csum = IDN_CKSUM_MBOX(&smr_mboxp->mt_header); if (!VALID_MBOXHDR(&smr_mboxp->mt_header, channel, mbox_csum)) { IDN_KSTAT_INC(sip, si_mboxcrc); IDN_KSTAT_INC(sip, si_ierrors); if (!(mmp[domid]->mm_flags & IDNMMBOX_FLAG_CORRUPTED)) { cmn_err(CE_WARN, "IDN: 241: [recv] (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, channel); mmp[domid]->mm_flags |= IDNMMBOX_FLAG_CORRUPTED; } empty = 0; mutex_exit(&mmp[domid]->mm_mutex); goto cc_sleep; } mutex_exit(&mmp[domid]->mm_mutex); mqp = &smr_mboxp->mt_queue[0]; pktcount = dropcount = 0; if (mqp[qi].ms_owner == 0) goto cc_next; bufoffset = IDN_BFRAME2OFFSET(mqp[qi].ms_bframe); if (!VALID_NWROFFSET(bufoffset, IDN_SMR_BUFSIZE)) { /* ASSERT(0); */ mqp[qi].ms_flag |= IDN_MBOXMSG_FLAG_ERR_BADOFFSET; mqp[qi].ms_owner = 0; IDN_MMBOXINDEX_INC(qi); dropcount++; IDN_KSTAT_INC(sip, si_smraddr); IDN_KSTAT_INC(sip, si_ierrors); } else { PR_DATA("%s: (channel %d) pkt (off 0x%x, " "qiget %d) from domain %d\n", proc, channel, bufoffset, qi, domid); #ifdef DEBUG hdrp = IDN_BUF2HDR(IDN_OFFSET2ADDR(bufoffset)); netaddr.netaddr = hdrp->b_netaddr; ASSERT(netaddr.net.chan == (ushort_t)channel); #endif /* DEBUG */ if (idn_recv_mboxdata(channel, IDN_OFFSET2ADDR(bufoffset)) < 0) { mutex_enter(&mmp[domid]->mm_mutex); if (!(mmp[domid]->mm_flags & IDNMMBOX_FLAG_CORRUPTED)) { cmn_err(CE_WARN, "IDN: 241: [recv] (domain " "%d, channel %d) SMR " "CORRUPTED - RELINK", domid, channel); mmp[domid]->mm_flags |= IDNMMBOX_FLAG_CORRUPTED; } mutex_exit(&mmp[domid]->mm_mutex); } mqp[qi].ms_owner = 0; IDN_MMBOXINDEX_INC(qi); pktcount++; } cc_next: mutex_enter(&mmp[domid]->mm_mutex); if (mmp[domid]->mm_smr_mboxp) { if (dropcount) mmp[domid]->mm_dropped += dropcount; mmp[domid]->mm_qiget = qi; mmp[domid]->mm_count += pktcount; } mutex_exit(&mmp[domid]->mm_mutex); if (pktcount == 0) { empty++; } else { csp->ch_recv_waittime = IDN_NETSVR_WAIT_MIN; empty = 0; idleloops = 0; PR_DATA("%s: (channel %d) dom=%d, pktcnt=%d\n", proc, channel, domid, pktcount); } continue; cc_slowdown: #ifdef DEBUG if (idleloops == 0) { PR_DATA("%s: (channel %d) going SOFT IDLE...\n", proc, channel); } #endif /* DEBUG */ if (idleloops++ < IDN_NETSVR_SPIN_COUNT) { /* * At this level we only busy-wait. * Get back into action. */ continue; } idleloops = 0; cc_sleep: if (mainhp) mainhp->mh_svr_active = 0; IDN_CHAN_LOCK_RECV(csp); cc_die: ASSERT(IDN_CHAN_RECV_IS_LOCKED(csp)); if (!IDN_CHANNEL_IS_RECV_ACTIVE(csp) && IDN_CHANNEL_IS_DETACHED(csp)) { /* * Time to die... */ PR_CHAN("%s: (channel %d) serviced %d " "packets, drop = %d\n", proc, channel, tot_pktcount, tot_dropcount); PR_CHAN("%s: (channel %d) TERMINATING\n", proc, channel); PR_CHAN("%s: (channel %d) ch_morguep = %p\n", proc, channel, csp->ch_recv_morguep); csp->ch_recv_threadp = NULL; #ifdef DEBUG for (index = 0; index < csp->ch_recv_domcount; index++) { if ((int)((csp->ch_recv_scanset >> (index*4)) & 0xf) == domid) { PR_DATA("%s: WARNING (channel %d) " "DROPPING domid %d...\n", proc, channel, domid); } } #endif /* DEBUG */ IDN_CHAN_RECV_DONE(csp); sema_v(csp->ch_recv_morguep); IDN_CHAN_UNLOCK_RECV(csp); thread_exit(); /* not reached */ } do { if (IDN_CHANNEL_IS_DETACHED(csp)) { PR_CHAN("%s: (channel %d) going to DIE...\n", proc, channel); goto cc_die; } #ifdef DEBUG if (IDN_CHANNEL_IS_RECV_ACTIVE(csp) && (csp->ch_recv_waittime <= IDN_NETSVR_WAIT_MAX)) { PR_CHAN("%s: (channel %d) going SOFT IDLE " "(waittime = %d ticks)...\n", proc, channel, csp->ch_recv_waittime); } else { PR_CHAN("%s: (channel %d) going " "HARD IDLE...\n", proc, channel); } #endif /* DEBUG */ IDN_CHAN_RECV_DONE(csp); /* * If we're being asked to check-in then * go into a hard sleep. Want to give the * thread requesting us to checkin a chance. */ while (csp->ch_recv.c_checkin) cv_wait(&csp->ch_recv_cv, &csp->ch_recv.c_mutex); if (csp->ch_recv_waittime > IDN_NETSVR_WAIT_MAX) cv_wait(&csp->ch_recv_cv, &csp->ch_recv.c_mutex); else (void) cv_timedwait(&csp->ch_recv_cv, &csp->ch_recv.c_mutex, lbolt + csp->ch_recv_waittime); IDN_CHAN_RECV_INPROGRESS(csp); IDN_KSTAT_INC(sip, si_sigsvr); if (csp->ch_recv_waittime <= IDN_NETSVR_WAIT_MAX) csp->ch_recv_waittime <<= IDN_NETSVR_WAIT_SHIFT; } while (!IDN_CHANNEL_IS_RECV_ACTIVE(csp)); /* * Before we see the world (and touch SMR space), * see if we've been told to die. */ mainhp = NULL; /* * The world may have changed since we were * asleep. Need to resync cache and check for a * new syncheader. * * Reset chansvr cache against any changes in * mbox fields we need (mm_qiget). */ CHANSVR_SYNC_CACHE(csp, mmp, channel); if (csp->ch_recv_domcount <= 0) { /* * Everybody disappeared on us. * Go back to sleep. */ goto cc_die; } ASSERT(csp->ch_recv_scanset && csp->ch_recv_domset); mainhp = idn_chan_server_syncheader(channel); if (mainhp == NULL) { /* * Bummer...we're idling... */ goto cc_die; } mainhp->mh_svr_active = 1; IDN_CHAN_UNLOCK_RECV(csp); /* * Reset the domid index after sleeping. */ CHANSVR_RESET_INDEX(index); empty = 0; idleloops = 0; } } #if 0 /* * We maintain a separate function for flushing the STREAMs * queue of a channel because it must be done outside the * context of the idn_chan_action routine. The streams flush * cannot occur inline with the idn_chan_action because * the act of flushing may cause IDN send functions to be called * directly and thus locks to be obtained which could result * in deadlocks. */ static void idn_chan_flush(idn_chansvr_t *csp) { queue_t *rq; struct idn *sip; int flush_type = 0; idn_chaninfo_t *csend, *crecv; procname_t proc = "idn_chan_flush"; csend = &csp->ch_send; crecv = &csp->ch_recv; mutex_enter(&crecv->c_mutex); mutex_enter(&csend->c_mutex); if (crecv->c_state & IDN_CHANSVC_STATE_FLUSH) flush_type |= FLUSHR; if (csend->c_state & IDN_CHANSVC_STATE_FLUSH) flush_type |= FLUSHW; if (flush_type) { rq = NULL; rw_enter(&idn.struprwlock, RW_READER); if ((sip = IDN_INST2SIP(csp->ch_id)) != NULL) rq = sip->si_ipq; rw_exit(&idn.struprwlock); if (rq) { /* * Flush the STREAM if possible * to get the channel server coherent * enough to respond to us. */ PR_CHAN("%s: sending FLUSH (%x) to channel %d\n", proc, flush_type, csp->ch_id); (void) putnextctl1(rq, M_FLUSH, flush_type); } crecv->c_state &= ~IDN_CHANSVC_STATE_FLUSH; csend->c_state &= ~IDN_CHANSVC_STATE_FLUSH; if (crecv->c_waiters) cv_broadcast(&crecv->c_cv); } mutex_exit(&csend->c_mutex); mutex_exit(&crecv->c_mutex); } #endif /* 0 */ /* * Locks are with respect to SEND/RECV locks (c_mutex). * * STOP/SUSPEND/DETACH * - Entered with locks dropped, leave with locks held. * DETACH - lock dropped manually. * RESTART/RESUME * - Entered with locks held, leave with locks dropped. * ATTACH * - both enter and leave with locks dropped. */ static void idn_chan_action(int channel, idn_chanaction_t chanaction, int wait) { uchar_t clr_state, set_state; uint_t is_running; domainset_t closed_slabwaiters = 0; struct idn *sip; idn_chansvr_t *csp; idn_chaninfo_t *csend, *crecv; procname_t proc = "idn_chan_action"; ASSERT((channel >= 0) && (channel < IDN_MAX_NETS)); ASSERT(idn.chan_servers); csp = &idn.chan_servers[channel]; PR_CHAN("%s: requesting %s for channel %d\n", proc, chanaction_str[(int)chanaction], channel); csend = &csp->ch_send; crecv = &csp->ch_recv; ASSERT(IDN_CHAN_GLOBAL_IS_LOCKED(csp)); clr_state = set_state = 0; switch (chanaction) { case IDNCHAN_ACTION_DETACH: clr_state = IDN_CHANSVC_STATE_MASK; /*FALLTHROUGH*/ case IDNCHAN_ACTION_STOP: clr_state |= IDN_CHANSVC_STATE_ENABLED; /*FALLTHROUGH*/ case IDNCHAN_ACTION_SUSPEND: clr_state |= IDN_CHANSVC_STATE_ACTIVE; /* * Must maintain this locking order. * Set asynchronous check-in flags. */ crecv->c_checkin = 1; csend->c_checkin = 1; is_running = 0; if ((csend->c_inprogress || crecv->c_inprogress) && wait && (csp->ch_recv_threadp != curthread)) { rw_enter(&idn.struprwlock, RW_READER); if ((sip = IDN_INST2SIP(channel)) != NULL) { /* * Temporarily turn off the STREAM * to give a chance to breath. */ is_running = sip->si_flags & IDNRUNNING; if (is_running) sip->si_flags &= ~IDNRUNNING; } rw_exit(&idn.struprwlock); } mutex_enter(&crecv->c_mutex); crecv->c_state &= ~clr_state; mutex_enter(&csend->c_mutex); csend->c_state &= ~clr_state; /* * It's possible the channel server could come * through this flow itself due to putting data upstream * that ultimately turned around and came back down for * sending. If this is the case we certainly don't * want to cv_wait, otherwise we'll obviously deadlock * waiting for ourself. So, only block if somebody * other than the channel server we're attempting to * suspend/stop. */ if (wait && (csp->ch_recv_threadp != curthread)) { int do_flush = 0; if (csend->c_inprogress || crecv->c_inprogress) do_flush++; if (do_flush) { rw_enter(&idn.struprwlock, RW_READER); if ((sip = IDN_INST2SIP(channel)) != NULL) { /* * Temporarily turn off the STREAM * to give a chance to breath. */ if (sip->si_flags & IDNRUNNING) { is_running = 1; sip->si_flags &= ~IDNRUNNING; } } rw_exit(&idn.struprwlock); } /* * If we have any senders in-progress * it's possible they're stuck waiting * down in smr_buf_alloc which may never * arrive if we're in an unlink process. * Rather than wait for it to timeout * let's be proactive so we can disconnect * asap. */ closed_slabwaiters = csp->ch_reg_domset; DOMAINSET_ADD(closed_slabwaiters, idn.localid); if (closed_slabwaiters) smr_slabwaiter_close(closed_slabwaiters); do { /* * It's possible due to a STREAMs * loopback from read queue to write queue * that receiver and sender may be same * thread, i.e. receiver's inprogress * flag will never clear until sender's * inprogress flag clears. So, we wait * for sender's inprogress first. */ while (csend->c_inprogress) { mutex_exit(&crecv->c_mutex); while (csend->c_inprogress) { csend->c_waiters++; cv_wait(&csend->c_cv, &csend->c_mutex); csend->c_waiters--; } /* * Maintain lock ordering. * Eventually we will catch * him due to the flag settings. */ mutex_exit(&csend->c_mutex); mutex_enter(&crecv->c_mutex); mutex_enter(&csend->c_mutex); } if (crecv->c_inprogress) { mutex_exit(&csend->c_mutex); while (crecv->c_inprogress) { crecv->c_waiters++; cv_wait(&crecv->c_cv, &crecv->c_mutex); crecv->c_waiters--; } mutex_enter(&csend->c_mutex); } } while (csend->c_inprogress); } if (is_running) { /* * Restore the IDNRUNNING bit in * the flags to let them know the * channel is still alive. */ rw_enter(&idn.struprwlock, RW_READER); if ((sip = IDN_INST2SIP(channel)) != NULL) sip->si_flags |= IDNRUNNING; rw_exit(&idn.struprwlock); } if (closed_slabwaiters) { /* * We can reopen now since at this point no new * slabwaiters will attempt to come in and wait. */ smr_slabwaiter_open(csp->ch_reg_domset); } crecv->c_checkin = 0; csend->c_checkin = 0; /* * ALL leave with locks held. */ PR_CHAN("%s: action (%s) for channel %d - COMPLETED\n", proc, chanaction_str[(int)chanaction], channel); break; case IDNCHAN_ACTION_ATTACH: mutex_enter(&crecv->c_mutex); mutex_enter(&csend->c_mutex); set_state |= csp->ch_state & IDN_CHANSVC_STATE_ATTACHED; /*FALLTHROUGH*/ case IDNCHAN_ACTION_RESTART: set_state |= csp->ch_state & IDN_CHANSVC_STATE_ENABLED; /*FALLTHROUGH*/ case IDNCHAN_ACTION_RESUME: ASSERT(IDN_CHAN_LOCAL_IS_LOCKED(csp)); set_state |= csp->ch_state & IDN_CHANSVC_STATE_ACTIVE; crecv->c_state |= set_state; csend->c_state |= set_state; /* * The channel server itself could come through this * flow, so obviously no point in attempting to wake * ourself up!. */ if (csp->ch_recv_threadp && (csp->ch_recv_threadp != curthread)) cv_signal(&csp->ch_recv_cv); PR_CHAN("%s: action (%s) for channel %d - COMPLETED\n", proc, chanaction_str[(int)chanaction], channel); /* * Leaves with lock released. */ mutex_exit(&csend->c_mutex); mutex_exit(&crecv->c_mutex); break; default: ASSERT(0); break; } } static void idn_chan_addmbox(int channel, ushort_t domset) { idn_chansvr_t *csp; register int d; procname_t proc = "idn_chan_addmbox"; PR_CHAN("%s: adding domset 0x%x main mailboxes to channel %d\n", proc, domset, channel); ASSERT(idn.chan_servers); csp = &idn.chan_servers[channel]; /* * Adding domains to a channel can be * asynchonous, so we don't bother waiting. */ IDN_CHANNEL_SUSPEND(channel, 0); /* * Now we have the sending and receiving sides blocked * for this channel. */ for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(domset, d)) continue; if (IDN_CHAN_DOMAIN_IS_REGISTERED(csp, d)) { DOMAINSET_DEL(domset, d); continue; } IDN_CHANSVR_SCANSET_ADD_PENDING(csp, d); DOMAINSET_ADD(csp->ch_recv_domset_pending, d); IDN_CHAN_DOMAIN_REGISTER(csp, d); PR_CHAN("%s: domain %d (channel %d) RECV (pending) " "scanset = 0x%lx\n", proc, d, channel, csp->ch_recv_scanset_pending); PR_CHAN("%s: domain %d (channel %d) domset = 0x%x\n", proc, d, channel, (uint_t)csp->ch_reg_domset); CHECKPOINT_OPENED(IDNSB_CHKPT_CHAN, idn_domain[d].dhw.dh_boardset, 1); } if (domset) csp->ch_recv_changed = 1; IDN_CHANNEL_RESUME(channel); } static void idn_chan_delmbox(int channel, ushort_t domset) { idn_chansvr_t *csp; register int d; procname_t proc = "idn_chan_delmbox"; PR_CHAN("%s: deleting domset 0x%x main mailboxes from channel %d\n", proc, domset, channel); ASSERT(idn.chan_servers); csp = &idn.chan_servers[channel]; /* * Here we have to wait for the channel server * as it's vital that we don't return without guaranteeing * that the given domset is no longer registered. */ IDN_CHANNEL_SUSPEND(channel, 1); /* * Now we have the sending and receiving sides blocked * for this channel. */ for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(domset, d)) continue; if (!IDN_CHAN_DOMAIN_IS_REGISTERED(csp, d)) { DOMAINSET_DEL(domset, d); continue; } /* * This domain has a mailbox hanging on this channel. * Get him out. * * First remove him from the receive side. */ ASSERT(csp->ch_recv_domcount > 0); IDN_CHANSVR_SCANSET_DEL_PENDING(csp, d); DOMAINSET_DEL(csp->ch_recv_domset_pending, d); IDN_CHAN_DOMAIN_UNREGISTER(csp, d); PR_CHAN("%s: domain %d (channel %d) RECV (pending) " "scanset = 0x%lx\n", proc, d, channel, csp->ch_recv_scanset_pending); PR_CHAN("%s: domain %d (channel %d) domset = 0x%x\n", proc, d, channel, (uint_t)csp->ch_reg_domset); CHECKPOINT_CLOSED(IDNSB_CHKPT_CHAN, idn_domain[d].dhw.dh_boardset, 2); } if (domset) csp->ch_recv_changed = 1; IDN_CHANNEL_RESUME(channel); } static int idn_valid_etherheader(struct ether_header *ehp) { uchar_t *eap; eap = &ehp->ether_dhost.ether_addr_octet[0]; if ((eap[IDNETHER_ZERO] != 0) && (eap[IDNETHER_ZERO] != 0xff)) return (0); if ((eap[IDNETHER_COOKIE1] != IDNETHER_COOKIE1_VAL) && (eap[IDNETHER_COOKIE1] != 0xff)) return (0); if ((eap[IDNETHER_COOKIE2] != IDNETHER_COOKIE2_VAL) && (eap[IDNETHER_COOKIE2] != 0xff)) return (0); if ((eap[IDNETHER_RESERVED] != IDNETHER_RESERVED_VAL) && (eap[IDNETHER_RESERVED] != 0xff)) return (0); if (!VALID_UCHANNEL(eap[IDNETHER_CHANNEL]) && (eap[IDNETHER_CHANNEL] != 0xff)) return (0); if (!VALID_UDOMAINID(IDN_NETID2DOMID(eap[IDNETHER_NETID])) && (eap[IDNETHER_NETID] != 0xff)) return (0); return (1); } /* * Packet header has already been filled in. * RETURNS: 0 * ENOLINK * EPROTO * ENOSPC */ /*ARGSUSED*/ static int idn_send_mboxdata(int domid, struct idn *sip, int channel, caddr_t bufp) { idn_mainmbox_t *mmp; idn_mboxmsg_t *mqp; smr_pkthdr_t *hdrp; smr_offset_t bufoffset; idn_netaddr_t dst; ushort_t mbox_csum; int rv = 0; int pktlen, qi; procname_t proc = "idn_send_mboxdata"; mmp = idn_domain[domid].dmbox.m_send; if (mmp == NULL) { PR_DATA("%s: dmbox.m_send == NULL\n", proc); IDN_KSTAT_INC(sip, si_linkdown); return (ENOLINK); } mmp += channel; mutex_enter(&mmp->mm_mutex); if (mmp->mm_smr_mboxp == NULL) { PR_DATA("%s: (d %d, chn %d) mm_smr_mboxp == NULL\n", proc, domid, channel); IDN_KSTAT_INC(sip, si_linkdown); rv = ENOLINK; goto send_err; } mbox_csum = IDN_CKSUM_MBOX(&mmp->mm_smr_mboxp->mt_header); if (mbox_csum != mmp->mm_smr_mboxp->mt_header.mh_cksum) { PR_DATA("%s: (d %d, chn %d) mbox hdr cksum (%d) " "!= actual (%d)\n", proc, domid, channel, mbox_csum, mmp->mm_smr_mboxp->mt_header.mh_cksum); if ((mmp->mm_flags & IDNMMBOX_FLAG_CORRUPTED) == 0) { cmn_err(CE_WARN, "IDN: 241: [send] (domain %d, " "channel %d) SMR CORRUPTED - RELINK", domid, channel); mmp->mm_flags |= IDNMMBOX_FLAG_CORRUPTED; } IDN_KSTAT_INC(sip, si_mboxcrc); IDN_KSTAT_INC(sip, si_oerrors); rv = EPROTO; goto send_err; } bufoffset = IDN_ADDR2OFFSET(bufp); hdrp = IDN_BUF2HDR(bufp); pktlen = hdrp->b_length; dst.netaddr = hdrp->b_netaddr; ASSERT(dst.net.chan == (ushort_t)channel); mqp = &mmp->mm_smr_mboxp->mt_queue[0]; qi = mmp->mm_qiput; if (mqp[qi].ms_owner) { PR_DATA("%s: mailbox FULL (qiput=%d, qiget=%d)\n", proc, mmp->mm_qiput, mmp->mm_qiget); IDN_KSTAT_INC(sip, si_txfull); rv = ENOSPC; goto send_err; } if (mqp[qi].ms_flag & IDN_MBOXMSG_FLAG_RECLAIM) { smr_offset_t recl_bufoffset; /* * Remote domain finished with mailbox entry, * however it has not been reclaimed yet. A reclaim * was done before coming into this routine, however * timing may have been such that the entry became * free just after the reclamation, but before * entry into here. Go ahead and reclaim this entry. */ recl_bufoffset = IDN_BFRAME2OFFSET(mqp[qi].ms_bframe); PR_DATA("%s: attempting reclaim (domain %d) " "(qiput=%d, b_off=0x%x)\n", proc, domid, qi, recl_bufoffset); if (VALID_NWROFFSET(recl_bufoffset, IDN_SMR_BUFSIZE)) { int recl; caddr_t b_bufp; smr_pkthdr_t *b_hdrp; b_bufp = IDN_OFFSET2ADDR(recl_bufoffset); b_hdrp = IDN_BUF2HDR(b_bufp); if (IDN_CKSUM_PKT(b_hdrp) != b_hdrp->b_cksum) { IDN_KSTAT_INC(sip, si_crc); IDN_KSTAT_INC(sip, si_fcs_errors); IDN_KSTAT_INC(sip, si_reclaim); IDN_KSTAT_INC(sip, si_oerrors); } recl = smr_buf_free(domid, b_bufp, b_hdrp->b_length); #ifdef DEBUG if (recl == 0) { PR_DATA("%s: SUCCESSFULLY reclaimed buf " "(domain %d)\n", proc, domid); } else { PR_DATA("%s: WARNING: reclaim failed (FREE) " "(domain %d)\n", proc, domid); } #endif /* DEBUG */ } else { IDN_KSTAT_INC(sip, si_smraddr); IDN_KSTAT_INC(sip, si_reclaim); PR_DATA("%s: WARNING: reclaim failed (BAD OFFSET) " "(domain %d)\n", proc, domid); } } if (*mmp->mm_smr_readyp == 0) { mmp->mm_qiput = qi; IDN_KSTAT_INC(sip, si_linkdown); rv = ENOLINK; goto send_err; } mqp[qi].ms_flag = IDN_MBOXMSG_FLAG_RECLAIM; mqp[qi].ms_bframe = IDN_OFFSET2BFRAME(bufoffset); /* membar_stst(); */ mqp[qi].ms_owner = 1; IDN_MMBOXINDEX_INC(qi); mmp->mm_qiput = qi; mmp->mm_count++; if ((*mmp->mm_smr_readyp) && !(*mmp->mm_smr_activep)) { idn_msgtype_t mt; mt.mt_mtype = IDNP_DATA; mt.mt_atype = 0; IDN_KSTAT_INC(sip, si_xdcall); (void) IDNXDC(domid, &mt, (uint_t)dst.net.chan, 0, 0, 0); } mutex_exit(&mmp->mm_mutex); IDN_KSTAT_INC(sip, si_opackets); IDN_KSTAT_INC(sip, si_opackets64); IDN_KSTAT_ADD(sip, si_xmtbytes, pktlen); IDN_KSTAT_ADD(sip, si_obytes64, (uint64_t)pktlen); return (0); send_err: mmp->mm_dropped++; mutex_exit(&mmp->mm_mutex); return (rv); } static int idn_recv_mboxdata(int channel, caddr_t bufp) { smr_pkthdr_t *hdrp; struct idn *sip; mblk_t *mp = nilp(mblk_t); int pktlen; int apktlen; int rv = 0; smr_offset_t bufoffset; ushort_t csum; idn_netaddr_t dst, daddr; procname_t proc = "idn_recv_mboxdata"; hdrp = IDN_BUF2HDR(bufp); csum = IDN_CKSUM_PKT(hdrp); sip = IDN_INST2SIP(channel); if (sip == NULL) { /*LINTED*/ sip = IDN_INST2SIP(0); } ASSERT(sip); if (csum != hdrp->b_cksum) { PR_DATA("%s: bad checksum(%x) != expected(%x)\n", proc, (uint_t)csum, (uint_t)hdrp->b_cksum); IDN_KSTAT_INC(sip, si_crc); IDN_KSTAT_INC(sip, si_fcs_errors); rv = -1; goto recv_err; } daddr.net.chan = (ushort_t)channel; daddr.net.netid = (ushort_t)idn.localid; dst.netaddr = hdrp->b_netaddr; bufoffset = hdrp->b_offset; if (dst.netaddr != daddr.netaddr) { PR_DATA("%s: wrong dest netaddr (0x%x), expected (0x%x)\n", proc, dst.netaddr, daddr.netaddr); IDN_KSTAT_INC(sip, si_nolink); IDN_KSTAT_INC(sip, si_macrcv_errors); goto recv_err; } pktlen = hdrp->b_length; apktlen = pktlen; if ((pktlen <= 0) || (pktlen > IDN_DATA_SIZE)) { PR_DATA("%s: invalid packet length (%d) <= 0 || > %lu\n", proc, pktlen, IDN_DATA_SIZE); IDN_KSTAT_INC(sip, si_buff); IDN_KSTAT_INC(sip, si_toolong_errors); goto recv_err; } mp = allocb(apktlen + IDN_ALIGNSIZE, BPRI_LO); if (mp == nilp(mblk_t)) { PR_DATA("%s: allocb(pkt) failed\n", proc); IDN_KSTAT_INC(sip, si_allocbfail); IDN_KSTAT_INC(sip, si_norcvbuf); /* MIB II */ goto recv_err; } ASSERT(DB_TYPE(mp) == M_DATA); /* * Copy data packet into its streams buffer. * Align pointers for maximum bcopy performance. */ mp->b_rptr = (uchar_t *)IDN_ALIGNPTR(mp->b_rptr, bufoffset); bcopy(IDN_BUF2DATA(bufp, bufoffset), mp->b_rptr, apktlen); mp->b_wptr = mp->b_rptr + pktlen; if (IDN_CHECKSUM && !idn_valid_etherheader((struct ether_header *)mp->b_rptr)) { freeb(mp); mp = nilp(mblk_t); PR_DATA("%s: etherheader CORRUPTED\n", proc); IDN_KSTAT_INC(sip, si_crc); IDN_KSTAT_INC(sip, si_fcs_errors); rv = -1; goto recv_err; } idndl_read(NULL, mp); recv_err: if (mp == nilp(mblk_t)) { IDN_KSTAT_INC(sip, si_ierrors); } return (rv); } /* * When on shutdown path (idn_active_resources) must call * idn_mainmbox_flush() _BEFORE_ calling idn_reclaim_mboxdata() * for any final data. This is necessary incase the mailboxes * have been unregistered. If they have then idn_mainmbox_flush() * will set mm_smr_mboxp to NULL which prevents us from touching * poison SMR space. */ int idn_reclaim_mboxdata(int domid, int channel, int nbufs) { idn_mainmbox_t *mmp; idn_mboxmsg_t *mqp; smr_pkthdr_t *hdrp; idn_domain_t *dp; int qi; int mi; int reclaim_cnt = 0; int free_cnt; ushort_t csum; struct idn *sip; smr_offset_t reclaim_list, curr, prev; procname_t proc = "idn_reclaim_mboxdata"; sip = IDN_INST2SIP(channel); if (sip == NULL) { /*LINTED*/ sip = IDN_INST2SIP(0); } ASSERT(sip); dp = &idn_domain[domid]; PR_DATA("%s: requested %d buffers from domain %d\n", proc, nbufs, domid); if (lock_try(&dp->dreclaim_inprogress) == 0) { /* * Reclaim is already in progress, don't * bother. */ PR_DATA("%s: reclaim already in progress\n", proc); return (0); } if (dp->dmbox.m_send == NULL) return (0); reclaim_list = curr = prev = IDN_NIL_SMROFFSET; mi = (int)dp->dreclaim_index; do { ushort_t mbox_csum; mmp = &dp->dmbox.m_send[mi]; /* do-while continues down */ ASSERT(mmp); if (mutex_tryenter(&mmp->mm_mutex) == 0) { /* * This channel is busy, move on. */ IDN_MBOXCHAN_INC(mi); continue; } if (mmp->mm_smr_mboxp == NULL) { PR_DATA("%s: no smr pointer for domid %d, chan %d\n", proc, domid, (int)mmp->mm_channel); ASSERT(mmp->mm_qiget == mmp->mm_qiput); mutex_exit(&mmp->mm_mutex); IDN_MBOXCHAN_INC(mi); continue; } mbox_csum = IDN_CKSUM_MBOX(&mmp->mm_smr_mboxp->mt_header); if (mbox_csum != mmp->mm_smr_mboxp->mt_header.mh_cksum) { PR_DATA("%s: (d %d, chn %d) mbox hdr " "cksum (%d) != actual (%d)\n", proc, domid, (int)mmp->mm_channel, mbox_csum, mmp->mm_smr_mboxp->mt_header.mh_cksum); IDN_KSTAT_INC(sip, si_mboxcrc); IDN_KSTAT_INC(sip, si_oerrors); mutex_exit(&mmp->mm_mutex); IDN_MBOXCHAN_INC(mi); continue; } mqp = &mmp->mm_smr_mboxp->mt_queue[0]; qi = mmp->mm_qiget; while (!mqp[qi].ms_owner && (mqp[qi].ms_flag & IDN_MBOXMSG_FLAG_RECLAIM) && nbufs) { idn_mboxmsg_t *msp; int badbuf; badbuf = 0; msp = &mqp[qi]; if (msp->ms_flag & IDN_MBOXMSG_FLAG_ERRMASK) { PR_DATA("%s: msg.flag ERROR(0x%x) (off=0x%x, " "domid=%d, qiget=%d)\n", proc, (uint_t)(msp->ms_flag & IDN_MBOXMSG_FLAG_ERRMASK), IDN_BFRAME2OFFSET(msp->ms_bframe), domid, qi); } prev = curr; curr = IDN_BFRAME2OFFSET(mqp[qi].ms_bframe); if (!VALID_NWROFFSET(curr, IDN_SMR_BUFSIZE)) { badbuf = 1; IDN_KSTAT_INC(sip, si_reclaim); } else { /* * Put the buffers onto a list that will be * formally reclaimed down below. This allows * us to free up mboxq entries as fast as * possible. */ hdrp = IDN_BUF2HDR(IDN_OFFSET2ADDR(curr)); csum = IDN_CKSUM_PKT(hdrp); if (csum != hdrp->b_cksum) { badbuf = 1; IDN_KSTAT_INC(sip, si_crc); IDN_KSTAT_INC(sip, si_fcs_errors); IDN_KSTAT_INC(sip, si_reclaim); if (!(mmp->mm_flags & IDNMMBOX_FLAG_CORRUPTED)) { cmn_err(CE_WARN, "IDN: 241: [send] " "(domain %d, channel " "%d) SMR CORRUPTED - " "RELINK", domid, channel); mmp->mm_flags |= IDNMMBOX_FLAG_CORRUPTED; } } else if (reclaim_list == IDN_NIL_SMROFFSET) { reclaim_list = curr; } else { caddr_t bufp; bufp = IDN_OFFSET2ADDR(prev); hdrp = IDN_BUF2HDR(bufp); hdrp->b_next = curr; } } mqp[qi].ms_flag = 0; IDN_MMBOXINDEX_INC(qi); if (!badbuf) { nbufs--; reclaim_cnt++; } if (qi == mmp->mm_qiget) break; } mmp->mm_qiget = qi; mutex_exit(&mmp->mm_mutex); IDN_MBOXCHAN_INC(mi); } while ((mi != (int)dp->dreclaim_index) && nbufs); dp->dreclaim_index = (uchar_t)mi; if (reclaim_list != IDN_NIL_SMROFFSET) { hdrp = IDN_BUF2HDR(IDN_OFFSET2ADDR(curr)); hdrp->b_next = IDN_NIL_SMROFFSET; } PR_DATA("%s: reclaimed %d buffers from domain %d\n", proc, reclaim_cnt, domid); if (reclaim_cnt == 0) { lock_clear(&dp->dreclaim_inprogress); return (0); } /* * Now actually go and reclaim (free) the buffers. */ free_cnt = 0; for (curr = reclaim_list; curr != IDN_NIL_SMROFFSET; ) { caddr_t bufp; bufp = IDN_OFFSET2ADDR(curr); hdrp = IDN_BUF2HDR(bufp); csum = IDN_CKSUM_PKT(hdrp); if (csum != hdrp->b_cksum) { /* * Once corruption is detected we * can't trust our list any further. * These buffers are effectively lost. */ cmn_err(CE_WARN, "IDN: 241: [send] (domain %d, channel %d) SMR " "CORRUPTED - RELINK", domid, channel); break; } curr = hdrp->b_next; if (!smr_buf_free(domid, bufp, hdrp->b_length)) free_cnt++; } if ((dp->dio < IDN_WINDOW_EMAX) && dp->diocheck) { lock_clear(&dp->diocheck); IDN_MSGTIMER_STOP(domid, IDNP_DATA, 0); } #ifdef DEBUG if (free_cnt != reclaim_cnt) { PR_DATA("%s: *** WARNING *** freecnt(%d) != reclaim_cnt (%d)\n", proc, free_cnt, reclaim_cnt); } #endif /* DEBUG */ lock_clear(&dp->dreclaim_inprogress); return (reclaim_cnt); } void idn_signal_data_server(int domid, ushort_t channel) { idn_nack_t nacktype = 0; idn_domain_t *dp; idn_chansvr_t *csp; int c, min_chan, max_chan; idn_mainmbox_t *mmp; procname_t proc = "idn_signal_data_server"; if (domid == IDN_NIL_DOMID) return; dp = &idn_domain[domid]; if (dp->dawol.a_count > 0) { /* * Domain was previously AWOL, but no longer. */ IDN_SYNC_LOCK(); IDN_GLOCK_EXCL(); idn_clear_awol(domid); IDN_GUNLOCK(); IDN_SYNC_UNLOCK(); } /* * Do a precheck before wasting time trying to acquire the lock. */ if ((dp->dstate != IDNDS_CONNECTED) || !IDN_DLOCK_TRY_SHARED(domid)) { /* * Either we're not connected or somebody is busy working * on the domain. Bail on the signal for now, we'll catch * it on the next go around. */ return; } /* * We didn't have the drwlock on the first check of dstate, * but now that we do, make sure the world hasn't changed! */ if (dp->dstate != IDNDS_CONNECTED) { /* * If we reach here, then no connection. * Send no response if this is the case. */ nacktype = IDNNACK_NOCONN; goto send_dresp; } /* * No need to worry about locking mainmbox * because we're already holding reader * lock on domain, plus we're just reading * fields in the mainmbox which only change * (or go away) when the writer lock is * held on the domain. */ if ((mmp = dp->dmbox.m_recv) == NULL) { /* * No local mailbox. */ nacktype = IDNNACK_BADCFG; goto send_dresp; } if ((channel != IDN_BROADCAST_ALLCHAN) && (channel >= IDN_MAX_NETS)) { nacktype = IDNNACK_BADCHAN; goto send_dresp; } if (channel == IDN_BROADCAST_ALLCHAN) { PR_DATA("%s: requested signal to ALL channels on domain %d\n", proc, domid); min_chan = 0; max_chan = IDN_MAX_NETS - 1; } else { PR_DATA("%s: requested signal to channel %d on domain %d\n", proc, channel, domid); min_chan = max_chan = (int)channel; } mmp += min_chan; for (c = min_chan; c <= max_chan; mmp++, c++) { /* * We do a quick check for a pending channel. * If pending it will need activation and we rather * do that through a separate (proto) thread. */ csp = &idn.chan_servers[c]; if (csp->ch_recv.c_checkin) { PR_DATA("%s: chansvr (%d) for domid %d CHECK-IN\n", proc, c, domid); continue; } if (IDN_CHAN_TRYLOCK_RECV(csp) == 0) { /* * Failed to grab lock, server must be active. */ PR_DATA("%s: chansvr (%d) for domid %d already actv\n", proc, c, domid); continue; } if (IDN_CHANNEL_IS_PENDING(csp)) { /* * Lock is pending. Submit asynchronous * job to activate and move-on. */ IDN_CHAN_UNLOCK_RECV(csp); idn_submit_chanactivate_job(c); continue; } /* * If he ain't active, we ain't talkin'. */ if (IDN_CHANNEL_IS_RECV_ACTIVE(csp) == 0) { IDN_CHAN_UNLOCK_RECV(csp); PR_DATA("%s: chansvr (%d) for domid %d inactive\n", proc, c, domid); continue; } if (mutex_tryenter(&mmp->mm_mutex) == 0) { IDN_CHAN_UNLOCK_RECV(csp); continue; } if (mmp->mm_csp != csp) { /* * Not registered. */ mutex_exit(&mmp->mm_mutex); IDN_CHAN_UNLOCK_RECV(csp); continue; } if (mmp->mm_smr_mboxp == NULL) { /* * No SMR mailbox. */ mutex_exit(&mmp->mm_mutex); IDN_CHAN_UNLOCK_RECV(csp); continue; } mutex_exit(&mmp->mm_mutex); if (csp->ch_recv.c_inprogress) { /* * Data server is already active. */ IDN_CHAN_UNLOCK_RECV(csp); PR_DATA("%s: chansvr (%d) for domid %d already actv\n", proc, c, domid); continue; } ASSERT(csp == &idn.chan_servers[c]); PR_DATA("%s: signaling data dispatcher for chan %d dom %d\n", proc, c, domid); ASSERT(csp); cv_signal(&csp->ch_recv_cv); IDN_CHAN_UNLOCK_RECV(csp); } if (!nacktype || (channel == IDN_BROADCAST_ALLCHAN)) { /* * If there were no real errors or we were * handling multiple channels, then just * return. */ IDN_DUNLOCK(domid); return; } send_dresp: PR_DATA("%s: sending NACK (%s) back to domain %d (cpu %d)\n", proc, idnnack_str[nacktype], domid, idn_domain[domid].dcpu); idn_send_dataresp(domid, nacktype); IDN_DUNLOCK(domid); } /*ARGSUSED*/ static int idn_recv_data(int domid, idn_msgtype_t *mtp, idn_xdcargs_t xargs) { #ifdef DEBUG uint_t msg = mtp ? mtp->mt_mtype : 0; uint_t msgarg = mtp ? mtp->mt_atype : 0; procname_t proc = "idn_recv_data"; PR_PROTO("%s:%d: DATA message received (msg = 0x%x, msgarg = 0x%x)\n", proc, domid, msg, msgarg); PR_PROTO("%s:%d: xargs = (0x%x, 0x%x, 0x%x, 0x%x)\n", proc, domid, xargs[0], xargs[1], xargs[2], xargs[3]); #endif /* DEBUG */ return (0); } /* * Only used when sending a negative response. */ static void idn_send_dataresp(int domid, idn_nack_t nacktype) { idn_msgtype_t mt; ASSERT(IDN_DLOCK_IS_HELD(domid)); if (idn_domain[domid].dcpu == IDN_NIL_DCPU) return; mt.mt_mtype = IDNP_NACK; mt.mt_atype = IDNP_DATA; (void) IDNXDC(domid, &mt, (uint_t)nacktype, 0, 0, 0); } /* * Checksum routine used in checksum smr_pkthdr_t and idn_mboxhdr_t. */ static ushort_t idn_cksum(register ushort_t *hdrp, register int count) { register int i; register ushort_t sum = 0; for (i = 0; i < count; i++) sum += hdrp[i]; sum = (sum >> 16) + (sum & 0xffff); sum += (sum >> 16); return (~sum); } /* * ------------------------------------------------ */ int idn_open_channel(int channel) { int masterid; idn_chansvr_t *csp; struct idn *sip; procname_t proc = "idn_open_channel"; if (channel >= IDN_MAX_NETS) { cmn_err(CE_WARN, "IDN: 242: maximum channels (%d) already open", IDN_MAX_NETS); return (-1); } IDN_GLOCK_EXCL(); ASSERT(idn.chan_servers != NULL); csp = &idn.chan_servers[channel]; IDN_CHAN_LOCK_GLOBAL(csp); if (IDN_CHANNEL_IS_ATTACHED(csp)) { PR_CHAN("%s: channel %d already open\n", proc, channel); IDN_CHAN_UNLOCK_GLOBAL(csp); IDN_GUNLOCK(); return (0); } /* * Need to zero out the kstats now that we're activating * this channel. */ for (sip = idn.sip; sip; sip = sip->si_nextp) { if (sip->si_dip && (ddi_get_instance(sip->si_dip) == channel)) { bzero(&sip->si_kstat, sizeof (sip->si_kstat)); break; } } IDN_CHANSVC_MARK_ATTACHED(csp); idn.nchannels++; CHANSET_ADD(idn.chanset, channel); IDN_CHANNEL_ATTACH(channel); IDN_CHAN_UNLOCK_GLOBAL(csp); /* * We increase our window threshold each time a channel * is opened. */ ASSERT(idn.nchannels > 0); IDN_WINDOW_EMAX = IDN_WINDOW_MAX + ((idn.nchannels - 1) * IDN_WINDOW_INCR); PR_CHAN("%s: channel %d is OPEN (nchannels = %d)\n", proc, channel, idn.nchannels); masterid = IDN_GET_MASTERID(); IDN_GUNLOCK(); /* * Check if there is an active master to which * we're connected. If so, then activate channel. */ if (masterid != IDN_NIL_DOMID) { idn_domain_t *dp; dp = &idn_domain[masterid]; IDN_DLOCK_SHARED(masterid); if (dp->dvote.v.master && (dp->dstate == IDNDS_CONNECTED)) (void) idn_activate_channel(CHANSET(channel), IDNCHAN_ONLINE); IDN_DUNLOCK(masterid); } return (0); } void idn_close_channel(int channel, idn_chanop_t chanop) { idn_chansvr_t *csp; procname_t proc = "idn_close_channel"; ASSERT(idn.chan_servers != NULL); csp = &idn.chan_servers[channel]; IDN_GLOCK_EXCL(); IDN_CHAN_LOCK_GLOBAL(csp); if (IDN_CHANNEL_IS_DETACHED(csp)) { PR_CHAN("%s: channel %d already closed\n", proc, channel); IDN_CHAN_UNLOCK_GLOBAL(csp); IDN_GUNLOCK(); return; } IDN_CHAN_UNLOCK_GLOBAL(csp); idn_deactivate_channel(CHANSET(channel), chanop); IDN_CHAN_LOCK_GLOBAL(csp); if (chanop == IDNCHAN_HARD_CLOSE) { idn.nchannels--; CHANSET_DEL(idn.chanset, channel); /* * We increase our window threshold each time a channel * is opened. */ if (idn.nchannels <= 0) IDN_WINDOW_EMAX = 0; else IDN_WINDOW_EMAX = IDN_WINDOW_MAX + ((idn.nchannels - 1) * IDN_WINDOW_INCR); } PR_CHAN("%s: channel %d is (%s) CLOSED (nchannels = %d)\n", proc, channel, (chanop == IDNCHAN_SOFT_CLOSE) ? "SOFT" : (chanop == IDNCHAN_HARD_CLOSE) ? "HARD" : "OFFLINE", idn.nchannels); IDN_CHAN_UNLOCK_GLOBAL(csp); IDN_GUNLOCK(); } static int idn_activate_channel(idn_chanset_t chanset, idn_chanop_t chanop) { int c, rv = 0; procname_t proc = "idn_activate_channel"; PR_CHAN("%s: chanset = 0x%x, chanop = %s\n", proc, chanset, chanop_str[chanop]); if (idn.state != IDNGS_ONLINE) { /* * Can't activate any channels unless local * domain is connected and thus has a master. */ PR_CHAN("%s: local domain not connected. no data servers\n", proc); return (-1); } for (c = 0; c < IDN_MAX_NETS; c++) { idn_chansvr_t *csp; idn_mboxhdr_t *mainhp; struct idn *sip; if (!CHAN_IN_SET(chanset, c)) continue; csp = &idn.chan_servers[c]; if (chanop == IDNCHAN_ONLINE) { IDN_CHAN_LOCK_GLOBAL(csp); } else { /* * We don't wait to grab the global lock * if IDNCHAN_OPEN since these occur along * critical data paths and will be retried * anyway if needed. */ if (IDN_CHAN_TRYLOCK_GLOBAL(csp) == 0) { PR_CHAN("%s: failed to acquire global " "lock for channel %d\n", proc, c); continue; } } if (!IDN_CHANNEL_IS_ATTACHED(csp)) { PR_CHAN("%s: channel %d NOT open\n", proc, c); IDN_CHAN_UNLOCK_GLOBAL(csp); continue; } if (IDN_CHANNEL_IS_ACTIVE(csp)) { PR_CHAN("%s: channel %d already active\n", proc, c); rv++; IDN_CHAN_UNLOCK_GLOBAL(csp); continue; } /* * Channel activation can happen asynchronously. */ IDN_CHANNEL_SUSPEND(c, 0); if (IDN_CHANNEL_IS_PENDING(csp) && (chanop == IDNCHAN_OPEN)) { PR_CHAN("%s: ACTIVATING channel %d\n", proc, c); if (idn_activate_channel_services(c) >= 0) { PR_CHAN("%s: Setting channel %d ACTIVE\n", proc, c); IDN_CHANSVC_MARK_ACTIVE(csp); rv++; } } else if (!IDN_CHANNEL_IS_PENDING(csp) && (chanop == IDNCHAN_ONLINE)) { PR_CHAN("%s: Setting channel %d PENDING\n", proc, c); IDN_CHANSVC_MARK_PENDING(csp); } /* * Don't syncheader (i.e. touch SMR) unless * channel is at least ENABLED. For a DISABLED * channel, the SMR may be invalid so do NOT * touch it. */ if (IDN_CHANNEL_IS_ENABLED(csp) && ((mainhp = idn_chan_server_syncheader(c)) != NULL)) { PR_CHAN("%s: marking chansvr (mhp=0x%p) %d READY\n", proc, mainhp, c); mainhp->mh_svr_ready = 1; } IDN_CHANNEL_RESUME(c); sip = IDN_INST2SIP(c); ASSERT(sip); if (sip->si_wantw) { mutex_enter(&idn.sipwenlock); idndl_wenable(sip); mutex_exit(&idn.sipwenlock); } IDN_CHAN_UNLOCK_GLOBAL(csp); } /* * Returns "not active", i.e. value of 0 indicates * no channels are activated. */ return (rv == 0); } static void idn_deactivate_channel(idn_chanset_t chanset, idn_chanop_t chanop) { int c; procname_t proc = "idn_deactivate_channel"; PR_CHAN("%s: chanset = 0x%x, chanop = %s\n", proc, chanset, chanop_str[chanop]); for (c = 0; c < IDN_MAX_NETS; c++) { idn_chansvr_t *csp; idn_mboxhdr_t *mainhp; if (!CHAN_IN_SET(chanset, c)) continue; csp = &idn.chan_servers[c]; IDN_CHAN_LOCK_GLOBAL(csp); if (((chanop == IDNCHAN_SOFT_CLOSE) && !IDN_CHANNEL_IS_ACTIVE(csp)) || ((chanop == IDNCHAN_HARD_CLOSE) && IDN_CHANNEL_IS_DETACHED(csp)) || ((chanop == IDNCHAN_OFFLINE) && !IDN_CHANNEL_IS_ENABLED(csp))) { ASSERT(!IDN_CHANNEL_IS_RECV_ACTIVE(csp)); ASSERT(!IDN_CHANNEL_IS_SEND_ACTIVE(csp)); PR_CHAN("%s: channel %d already deactivated\n", proc, c); IDN_CHAN_UNLOCK_GLOBAL(csp); continue; } switch (chanop) { case IDNCHAN_OFFLINE: IDN_CHANSVC_MARK_IDLE(csp); IDN_CHANSVC_MARK_DISABLED(csp); IDN_CHANNEL_STOP(c, 1); mainhp = idn_chan_server_syncheader(c); if (mainhp != NULL) mainhp->mh_svr_ready = 0; break; case IDNCHAN_HARD_CLOSE: IDN_CHANSVC_MARK_DETACHED(csp); IDN_CHANNEL_DETACH(c, 1); mainhp = idn_chan_server_syncheader(c); if (mainhp != NULL) mainhp->mh_svr_ready = 0; break; default: IDN_CHANSVC_MARK_IDLE(csp); IDN_CHANNEL_SUSPEND(c, 1); ASSERT(IDN_CHANNEL_IS_ATTACHED(csp)); break; } lock_clear(&csp->ch_actvlck); lock_clear(&csp->ch_initlck); PR_CHAN("%s: DEACTIVATING channel %d (%s)\n", proc, c, chanop_str[chanop]); PR_CHAN("%s: removing chanset 0x%x data svrs for " "each domain link\n", proc, chanset); (void) idn_deactivate_channel_services(c, chanop); } /* * Returns with channels unlocked. */ } /* * The priority of the channel server must be less than that * of the protocol server since the protocol server tasks * are (can be) of more importance. * * Possible range: 60-99. */ static pri_t idn_chansvr_pri = (7 * MAXCLSYSPRI) / 8; static int idn_activate_channel_services(int channel) { idn_chansvr_t *csp; procname_t proc = "idn_activate_channel_services"; ASSERT((channel >= 0) && (channel < IDN_MAX_NETS)); csp = &idn.chan_servers[channel]; ASSERT(IDN_CHAN_GLOBAL_IS_LOCKED(csp)); ASSERT(IDN_CHAN_LOCAL_IS_LOCKED(csp)); if (csp->ch_recv_threadp) { /* * There's an existing dispatcher! * Must have been idle'd during an earlier * stint. */ ASSERT(csp->ch_id == (uchar_t)channel); PR_CHAN("%s: existing chansvr FOUND for (c=%d)\n", proc, channel); if (IDN_CHANNEL_IS_PENDING(csp) == 0) return (-1); PR_CHAN("%s: chansvr (c=%d) Rstate = 0x%x, Sstate = 0x%x\n", proc, channel, csp->ch_recv.c_state, csp->ch_send.c_state); cv_signal(&csp->ch_recv_cv); return (0); } if (IDN_CHANNEL_IS_PENDING(csp) == 0) return (-1); csp->ch_id = (uchar_t)channel; PR_CHAN("%s: init channel %d server\n", proc, channel); csp->ch_recv_morguep = GETSTRUCT(ksema_t, 1); sema_init(csp->ch_recv_morguep, 0, NULL, SEMA_DRIVER, NULL); csp->ch_recv.c_inprogress = 0; csp->ch_recv.c_waiters = 0; csp->ch_recv.c_checkin = 0; csp->ch_recv_changed = 1; csp->ch_recv_domset = csp->ch_reg_domset; csp->ch_recv_waittime = IDN_NETSVR_WAIT_MIN; csp->ch_recv_threadp = thread_create(NULL, 0, idn_chan_server, &csp, sizeof (csp), &p0, TS_RUN, idn_chansvr_pri); csp->ch_send.c_inprogress = 0; csp->ch_send.c_waiters = 0; csp->ch_send.c_checkin = 0; return (0); } /* * This routine can handle terminating a set of channel * servers all at once, however currently only used * for serial killing, i.e. one-at-a-time. * * Entered with RECV locks held on chanset. * Acquires SEND locks if needed. * Leaves with all RECV and SEND locks dropped. */ static int idn_deactivate_channel_services(int channel, idn_chanop_t chanop) { idn_chansvr_t *csp; int cs_count; int c; idn_chanset_t chanset; ksema_t *central_morguep = NULL; procname_t proc = "idn_deactivate_channel_services"; ASSERT(idn.chan_servers); PR_CHAN("%s: deactivating channel %d services\n", proc, channel); /* * XXX * Old code allowed us to deactivate multiple channel * servers at once. Keep for now just in case. */ chanset = CHANSET(channel); /* * Point all the data dispatchers to the same morgue * so we can kill them all at once. */ cs_count = 0; for (c = 0; c < IDN_MAX_NETS; c++) { if (!CHAN_IN_SET(chanset, c)) continue; csp = &idn.chan_servers[c]; ASSERT(IDN_CHAN_GLOBAL_IS_LOCKED(csp)); ASSERT(IDN_CHAN_LOCAL_IS_LOCKED(csp)); if (csp->ch_recv_threadp == NULL) { /* * No channel server home. * But we're still holding the c_mutex. * At mark him idle incase we start him up. */ PR_CHAN("%s: no channel server found for chan %d\n", proc, c); IDN_CHAN_UNLOCK_LOCAL(csp); IDN_CHAN_UNLOCK_GLOBAL(csp); continue; } ASSERT(csp->ch_id == (uchar_t)c); /* * Okay, now we've blocked the send and receive sides. */ if ((chanop == IDNCHAN_SOFT_CLOSE) || (chanop == IDNCHAN_OFFLINE)) { /* * We set turned off the ACTIVE flag, but there's * no guarantee he stopped because of it. He may * have already been sleeping. We need to be * sure he recognizes the IDLE, so we need to * signal him and give him a chance to see it. */ cv_signal(&csp->ch_recv_cv); IDN_CHAN_UNLOCK_LOCAL(csp); IDN_CHAN_UNLOCK_GLOBAL(csp); cs_count++; continue; } PR_CHAN("%s: pointing chansvr %d to morgue (0x%p)\n", proc, c, central_morguep ? central_morguep : csp->ch_recv_morguep); if (central_morguep == NULL) { central_morguep = csp->ch_recv_morguep; } else { sema_destroy(csp->ch_recv_morguep); FREESTRUCT(csp->ch_recv_morguep, ksema_t, 1); csp->ch_recv_morguep = central_morguep; } cv_signal(&csp->ch_recv_cv); if (csp->ch_recv.c_waiters > 0) cv_broadcast(&csp->ch_recv.c_cv); /* * Save any existing binding for next reincarnation. * Note that we're holding the local and global * locks so we're protected against others touchers * of the ch_bound_cpuid fields. */ csp->ch_bound_cpuid_pending = csp->ch_bound_cpuid; csp->ch_bound_cpuid = -1; IDN_CHAN_UNLOCK_LOCAL(csp); IDN_CHAN_UNLOCK_GLOBAL(csp); cs_count++; } PR_CHAN("%s: signaled %d chansvrs for chanset 0x%x\n", proc, cs_count, chanset); if ((chanop == IDNCHAN_SOFT_CLOSE) || (chanop == IDNCHAN_OFFLINE)) return (cs_count); PR_CHAN("%s: waiting for %d (chnset=0x%x) chan svrs to term\n", proc, cs_count, chanset); PR_CHAN("%s: morguep = 0x%p\n", proc, central_morguep); ASSERT((cs_count > 0) ? (central_morguep != NULL) : 1); while (cs_count-- > 0) sema_p(central_morguep); if (central_morguep) { sema_destroy(central_morguep); FREESTRUCT(central_morguep, ksema_t, 1); } return (cs_count); } int idn_chanservers_init() { int c; idn_chansvr_t *csp; if (idn.chan_servers) return (0); idn.chan_servers = GETSTRUCT(idn_chansvr_t, IDN_MAXMAX_NETS); for (c = 0; c < IDN_MAXMAX_NETS; c++) { csp = &idn.chan_servers[c]; mutex_init(&csp->ch_send.c_mutex, NULL, MUTEX_DEFAULT, NULL); mutex_init(&csp->ch_recv.c_mutex, NULL, MUTEX_DEFAULT, NULL); cv_init(&csp->ch_send.c_cv, NULL, CV_DRIVER, NULL); cv_init(&csp->ch_recv.c_cv, NULL, CV_DRIVER, NULL); cv_init(&csp->ch_recv_cv, NULL, CV_DRIVER, NULL); csp->ch_bound_cpuid = -1; csp->ch_bound_cpuid_pending = -1; } return (c); } void idn_chanservers_deinit() { int c; idn_chansvr_t *csp; if (idn.chan_servers == NULL) return; for (c = 0; c < IDN_MAXMAX_NETS; c++) { csp = &idn.chan_servers[c]; mutex_destroy(&csp->ch_send.c_mutex); mutex_destroy(&csp->ch_recv.c_mutex); cv_destroy(&csp->ch_send.c_cv); cv_destroy(&csp->ch_recv.c_cv); cv_destroy(&csp->ch_recv_cv); } FREESTRUCT(idn.chan_servers, idn_chansvr_t, IDN_MAXMAX_NETS); idn.chan_servers = NULL; } static void idn_exec_chanactivate(void *chn) { int not_active, channel; idn_chansvr_t *csp; channel = (int)(uintptr_t)chn; IDN_GLOCK_SHARED(); if (idn.chan_servers == NULL) { IDN_GUNLOCK(); return; } csp = &idn.chan_servers[channel]; if (IDN_CHAN_TRYLOCK_GLOBAL(csp) == 0) { /* * If we can't grab the global lock, then * something is up, skip out. */ IDN_GUNLOCK(); return; } IDN_GUNLOCK(); if (IDN_CHANNEL_IS_PENDING(csp) && lock_try(&csp->ch_actvlck)) { IDN_CHAN_UNLOCK_GLOBAL(csp); not_active = idn_activate_channel(CHANSET(channel), IDNCHAN_OPEN); if (not_active) lock_clear(&csp->ch_actvlck); } else { IDN_CHAN_UNLOCK_GLOBAL(csp); } } /* * Delayed activation of channel. We don't want to do this within * idn_signal_data_server() since that's called within the context * of an XDC handler so we submit it as a timeout() call to be short * as soon as possible. * The ch_initlck & ch_actvlck are used to synchronize activation * of the channel so that we don't have multiple idn_activate_channel's * attempting to activate the same channel. */ static void idn_submit_chanactivate_job(int channel) { idn_chansvr_t *csp; if (idn.chan_servers == NULL) return; csp = &idn.chan_servers[channel]; if (lock_try(&csp->ch_initlck) == 0) return; (void) timeout(idn_exec_chanactivate, (caddr_t)(uintptr_t)channel, 1); } /*ARGSUSED0*/ static void idn_xmit_monitor(void *unused) { int c, d; idn_chansvr_t *csp; idn_chanset_t wake_set; domainset_t conset; smr_slab_t *sp; procname_t proc = "idn_xmit_monitor"; CHANSET_ZERO(wake_set); mutex_enter(&idn.xmit_lock); if ((idn.xmit_tid == NULL) || !idn.xmit_chanset_wanted) { idn.xmit_tid = NULL; mutex_exit(&idn.xmit_lock); PR_XMON("%s: bailing out\n", proc); return; } /* * No point in transmitting unless state * is ONLINE. */ if (idn.state != IDNGS_ONLINE) goto retry; conset = idn.domset.ds_connected; /* * Try and reclaim some buffers if possible. */ for (d = 0; d < MAX_DOMAINS; d++) { if (!DOMAIN_IN_SET(conset, d)) continue; if (!IDN_DLOCK_TRY_SHARED(d)) continue; if (idn_domain[d].dcpu != IDN_NIL_DCPU) (void) idn_reclaim_mboxdata(d, 0, -1); IDN_DUNLOCK(d); } /* * Now check if we were successful in getting * any buffers. */ DSLAB_LOCK_SHARED(idn.localid); sp = idn_domain[idn.localid].dslab; for (; sp; sp = sp->sl_next) if (sp->sl_free) break; DSLAB_UNLOCK(idn.localid); /* * If there are no buffers available, * no point in reenabling the queues. */ if (sp == NULL) goto retry; CHANSET_ZERO(wake_set); for (c = 0; c < IDN_MAX_NETS; c++) { int pending_bits; struct idn *sip; if (!CHAN_IN_SET(idn.xmit_chanset_wanted, c)) continue; csp = &idn.chan_servers[c]; if (!IDN_CHAN_TRYLOCK_GLOBAL(csp)) continue; pending_bits = csp->ch_state & IDN_CHANSVC_PENDING_BITS; sip = IDN_INST2SIP(c); if (!csp->ch_send.c_checkin && (pending_bits == IDN_CHANSVC_PENDING_BITS) && sip && (sip->si_flags & IDNRUNNING)) { IDN_CHAN_UNLOCK_GLOBAL(csp); CHANSET_ADD(wake_set, c); PR_XMON("%s: QENABLE for channel %d\n", proc, c); rw_enter(&idn.struprwlock, RW_READER); mutex_enter(&idn.sipwenlock); idndl_wenable(sip); mutex_exit(&idn.sipwenlock); rw_exit(&idn.struprwlock); } else { IDN_CHAN_UNLOCK_GLOBAL(csp); } } /* * Clear the channels we enabled. */ idn.xmit_chanset_wanted &= ~wake_set; retry: if (idn.xmit_chanset_wanted == 0) idn.xmit_tid = NULL; else idn.xmit_tid = timeout(idn_xmit_monitor, NULL, idn_xmit_monitor_freq); mutex_exit(&idn.xmit_lock); } void idn_xmit_monitor_kickoff(int chan_wanted) { procname_t proc = "idn_xmit_monitor_kickoff"; mutex_enter(&idn.xmit_lock); if (chan_wanted < 0) { /* * Wants all channels. */ idn.xmit_chanset_wanted = CHANSET_ALL; } else { CHANSET_ADD(idn.xmit_chanset_wanted, chan_wanted); } if (idn.xmit_tid != (timeout_id_t)NULL) { /* * A monitor is already running, so * he will catch the new "wants" when * he comes around. */ mutex_exit(&idn.xmit_lock); return; } PR_XMON("%s: xmit_mon kicked OFF (chanset = 0x%x)\n", proc, idn.xmit_chanset_wanted); idn.xmit_tid = timeout(idn_xmit_monitor, NULL, idn_xmit_monitor_freq); mutex_exit(&idn.xmit_lock); }