1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2004 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */ 27 /* All Rights Reserved */ 28 /* 29 * Portions of this source code were derived from Berkeley 30 * 4.3 BSD under license from the Regents of the University of 31 * California. 32 */ 33 34 #pragma ident "%Z%%M% %I% %E% SMI" 35 36 /* 37 * Server side of RPC over RDMA in the kernel. 38 */ 39 40 #include <sys/param.h> 41 #include <sys/types.h> 42 #include <sys/user.h> 43 #include <sys/sysmacros.h> 44 #include <sys/proc.h> 45 #include <sys/file.h> 46 #include <sys/errno.h> 47 #include <sys/kmem.h> 48 #include <sys/debug.h> 49 #include <sys/systm.h> 50 #include <sys/cmn_err.h> 51 #include <sys/kstat.h> 52 #include <sys/vtrace.h> 53 #include <sys/debug.h> 54 55 #include <rpc/types.h> 56 #include <rpc/xdr.h> 57 #include <rpc/auth.h> 58 #include <rpc/clnt.h> 59 #include <rpc/rpc_msg.h> 60 #include <rpc/svc.h> 61 #include <rpc/rpc_rdma.h> 62 #include <sys/ddi.h> 63 #include <sys/sunddi.h> 64 65 #include <inet/common.h> 66 #include <inet/ip.h> 67 #include <inet/ip6.h> 68 69 /* 70 * RDMA transport specific data associated with SVCMASTERXPRT 71 */ 72 struct rdma_data { 73 SVCMASTERXPRT *rd_xprt; /* back ptr to SVCMASTERXPRT */ 74 struct rdma_svc_data rd_data; /* rdma data */ 75 rdma_mod_t *r_mod; /* RDMA module containing ops ptr */ 76 }; 77 78 /* 79 * Plugin connection specific data stashed away in clone SVCXPRT 80 */ 81 struct clone_rdma_data { 82 CONN *conn; /* RDMA connection */ 83 rdma_buf_t rpcbuf; /* RPC req/resp buffer */ 84 }; 85 86 #ifdef DEBUG 87 int rdma_svc_debug = 0; 88 #endif 89 90 #define MAXADDRLEN 128 /* max length for address mask */ 91 92 /* 93 * Routines exported through ops vector. 94 */ 95 static bool_t svc_rdma_krecv(SVCXPRT *, mblk_t *, struct rpc_msg *); 96 static bool_t svc_rdma_ksend(SVCXPRT *, struct rpc_msg *); 97 static bool_t svc_rdma_kgetargs(SVCXPRT *, xdrproc_t, caddr_t); 98 static bool_t svc_rdma_kfreeargs(SVCXPRT *, xdrproc_t, caddr_t); 99 void svc_rdma_kdestroy(SVCMASTERXPRT *); 100 static int svc_rdma_kdup(struct svc_req *, caddr_t, int, 101 struct dupreq **, bool_t *); 102 static void svc_rdma_kdupdone(struct dupreq *, caddr_t, 103 void (*)(), int, int); 104 static int32_t *svc_rdma_kgetres(SVCXPRT *, int); 105 static void svc_rdma_kfreeres(SVCXPRT *); 106 static void svc_rdma_kclone_destroy(SVCXPRT *); 107 static void svc_rdma_kstart(SVCMASTERXPRT *); 108 void svc_rdma_kstop(SVCMASTERXPRT *); 109 110 /* 111 * Server transport operations vector. 112 */ 113 struct svc_ops rdma_svc_ops = { 114 svc_rdma_krecv, /* Get requests */ 115 svc_rdma_kgetargs, /* Deserialize arguments */ 116 svc_rdma_ksend, /* Send reply */ 117 svc_rdma_kfreeargs, /* Free argument data space */ 118 svc_rdma_kdestroy, /* Destroy transport handle */ 119 svc_rdma_kdup, /* Check entry in dup req cache */ 120 svc_rdma_kdupdone, /* Mark entry in dup req cache as done */ 121 svc_rdma_kgetres, /* Get pointer to response buffer */ 122 svc_rdma_kfreeres, /* Destroy pre-serialized response header */ 123 svc_rdma_kclone_destroy, /* Destroy a clone xprt */ 124 svc_rdma_kstart /* Tell `ready-to-receive' to rpcmod */ 125 }; 126 127 /* 128 * Server statistics 129 * NOTE: This structure type is duplicated in the NFS fast path. 130 */ 131 struct { 132 kstat_named_t rscalls; 133 kstat_named_t rsbadcalls; 134 kstat_named_t rsnullrecv; 135 kstat_named_t rsbadlen; 136 kstat_named_t rsxdrcall; 137 kstat_named_t rsdupchecks; 138 kstat_named_t rsdupreqs; 139 kstat_named_t rslongrpcs; 140 } rdmarsstat = { 141 { "calls", KSTAT_DATA_UINT64 }, 142 { "badcalls", KSTAT_DATA_UINT64 }, 143 { "nullrecv", KSTAT_DATA_UINT64 }, 144 { "badlen", KSTAT_DATA_UINT64 }, 145 { "xdrcall", KSTAT_DATA_UINT64 }, 146 { "dupchecks", KSTAT_DATA_UINT64 }, 147 { "dupreqs", KSTAT_DATA_UINT64 }, 148 { "longrpcs", KSTAT_DATA_UINT64 } 149 }; 150 151 kstat_named_t *rdmarsstat_ptr = (kstat_named_t *)&rdmarsstat; 152 uint_t rdmarsstat_ndata = sizeof (rdmarsstat) / sizeof (kstat_named_t); 153 154 #define RSSTAT_INCR(x) rdmarsstat.x.value.ui64++ 155 156 /* 157 * Create a transport record. 158 * The transport record, output buffer, and private data structure 159 * are allocated. The output buffer is serialized into using xdrmem. 160 * There is one transport record per user process which implements a 161 * set of services. 162 */ 163 /* ARGSUSED */ 164 int 165 svc_rdma_kcreate(char *netid, SVC_CALLOUT_TABLE *sct, int id, 166 rdma_xprt_group_t *started_xprts) 167 { 168 int error; 169 SVCMASTERXPRT *xprt; 170 struct rdma_data *rd; 171 rdma_registry_t *rmod; 172 rdma_xprt_record_t *xprt_rec; 173 queue_t *q; 174 175 /* 176 * modload the RDMA plugins is not already done. 177 */ 178 if (!rdma_modloaded) { 179 mutex_enter(&rdma_modload_lock); 180 if (!rdma_modloaded) { 181 error = rdma_modload(); 182 } 183 mutex_exit(&rdma_modload_lock); 184 185 if (error) 186 return (error); 187 } 188 189 /* 190 * master_xprt_count is the count of master transport handles 191 * that were successfully created and are ready to recieve for 192 * RDMA based access. 193 */ 194 error = 0; 195 xprt_rec = NULL; 196 rw_enter(&rdma_lock, RW_READER); 197 if (rdma_mod_head == NULL) { 198 started_xprts->rtg_count = 0; 199 rw_exit(&rdma_lock); 200 if (rdma_dev_available) 201 return (EPROTONOSUPPORT); 202 else 203 return (ENODEV); 204 } 205 206 /* 207 * If we have reached here, then atleast one RDMA plugin has loaded. 208 * Create a master_xprt, make it start listenining on the device, 209 * if an error is generated, record it, we might need to shut 210 * the master_xprt. 211 * SVC_START() calls svc_rdma_kstart which calls plugin binding 212 * routines. 213 */ 214 for (rmod = rdma_mod_head; rmod != NULL; rmod = rmod->r_next) { 215 216 /* 217 * One SVCMASTERXPRT per RDMA plugin. 218 */ 219 xprt = kmem_zalloc(sizeof (*xprt), KM_SLEEP); 220 xprt->xp_ops = &rdma_svc_ops; 221 xprt->xp_sct = sct; 222 xprt->xp_type = T_RDMA; 223 mutex_init(&xprt->xp_req_lock, NULL, MUTEX_DEFAULT, NULL); 224 mutex_init(&xprt->xp_thread_lock, NULL, MUTEX_DEFAULT, NULL); 225 xprt->xp_req_head = (mblk_t *)0; 226 xprt->xp_req_tail = (mblk_t *)0; 227 xprt->xp_threads = 0; 228 xprt->xp_detached_threads = 0; 229 230 rd = kmem_zalloc(sizeof (*rd), KM_SLEEP); 231 xprt->xp_p2 = (caddr_t)rd; 232 rd->rd_xprt = xprt; 233 rd->r_mod = rmod->r_mod; 234 235 q = &rd->rd_data.q; 236 xprt->xp_wq = q; 237 q->q_ptr = &rd->rd_xprt; 238 xprt->xp_netid = NULL; 239 240 if (netid != NULL) { 241 xprt->xp_netid = kmem_alloc(strlen(netid) + 1, 242 KM_SLEEP); 243 (void) strcpy(xprt->xp_netid, netid); 244 } 245 246 xprt->xp_addrmask.maxlen = 247 xprt->xp_addrmask.len = sizeof (struct sockaddr_in); 248 xprt->xp_addrmask.buf = 249 kmem_zalloc(xprt->xp_addrmask.len, KM_SLEEP); 250 ((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_addr.s_addr = 251 (uint32_t)~0; 252 ((struct sockaddr_in *)xprt->xp_addrmask.buf)->sin_family = 253 (ushort_t)~0; 254 255 /* 256 * Each of the plugins will have their own Service ID 257 * to listener specific mapping, like port number for VI 258 * and service name for IB. 259 */ 260 rd->rd_data.svcid = id; 261 error = svc_xprt_register(xprt, id); 262 if (error) { 263 cmn_err(CE_WARN, "svc_rdma_kcreate: svc_xprt_register" 264 "failed"); 265 goto cleanup; 266 } 267 268 SVC_START(xprt); 269 if (!rd->rd_data.active) { 270 svc_xprt_unregister(xprt); 271 error = rd->rd_data.err_code; 272 goto cleanup; 273 } 274 275 /* 276 * This is set only when there is atleast one or more 277 * transports successfully created. We insert the pointer 278 * to the created RDMA master xprt into a separately maintained 279 * list. This way we can easily reference it later to cleanup, 280 * when NFS kRPC service pool is going away/unregistered. 281 */ 282 started_xprts->rtg_count ++; 283 xprt_rec = kmem_alloc(sizeof (*xprt_rec), KM_SLEEP); 284 xprt_rec->rtr_xprt_ptr = xprt; 285 xprt_rec->rtr_next = started_xprts->rtg_listhead; 286 started_xprts->rtg_listhead = xprt_rec; 287 continue; 288 cleanup: 289 SVC_DESTROY(xprt); 290 if (error == RDMA_FAILED) 291 error = EPROTONOSUPPORT; 292 } 293 294 rw_exit(&rdma_lock); 295 296 /* 297 * Don't return any error even if a single plugin was started 298 * successfully. 299 */ 300 if (started_xprts->rtg_count == 0) 301 return (error); 302 return (0); 303 } 304 305 /* 306 * Cleanup routine for freeing up memory allocated by 307 * svc_rdma_kcreate() 308 */ 309 void 310 svc_rdma_kdestroy(SVCMASTERXPRT *xprt) 311 { 312 struct rdma_data *rd = (struct rdma_data *)xprt->xp_p2; 313 314 315 mutex_destroy(&xprt->xp_req_lock); 316 mutex_destroy(&xprt->xp_thread_lock); 317 kmem_free(xprt->xp_netid, strlen(xprt->xp_netid) + 1); 318 kmem_free(rd, sizeof (*rd)); 319 kmem_free(xprt->xp_addrmask.buf, xprt->xp_addrmask.maxlen); 320 kmem_free(xprt, sizeof (*xprt)); 321 } 322 323 324 static void 325 svc_rdma_kstart(SVCMASTERXPRT *xprt) 326 { 327 struct rdma_svc_data *svcdata; 328 rdma_mod_t *rmod; 329 330 svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data; 331 rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod; 332 333 /* 334 * Create a listener for module at this port 335 */ 336 337 (*rmod->rdma_ops->rdma_svc_listen)(svcdata); 338 } 339 340 void 341 svc_rdma_kstop(SVCMASTERXPRT *xprt) 342 { 343 struct rdma_svc_data *svcdata; 344 rdma_mod_t *rmod; 345 346 svcdata = &((struct rdma_data *)xprt->xp_p2)->rd_data; 347 rmod = ((struct rdma_data *)xprt->xp_p2)->r_mod; 348 349 /* 350 * Call the stop listener routine for each plugin. 351 */ 352 (*rmod->rdma_ops->rdma_svc_stop)(svcdata); 353 if (svcdata->active) 354 cmn_err(CE_WARN, "rdma_stop: Failed to shutdown RDMA based kRPC" 355 " listener"); 356 } 357 358 /* ARGSUSED */ 359 static void 360 svc_rdma_kclone_destroy(SVCXPRT *clone_xprt) 361 { 362 } 363 364 static bool_t 365 svc_rdma_krecv(SVCXPRT *clone_xprt, mblk_t *mp, struct rpc_msg *msg) 366 { 367 XDR *xdrs; 368 rdma_stat status; 369 struct recv_data *rdp = (struct recv_data *)mp->b_rptr; 370 CONN *conn; 371 struct clone_rdma_data *vd; 372 struct clist *cl; 373 uint_t vers, op, pos; 374 uint32_t xid; 375 376 vd = (struct clone_rdma_data *)clone_xprt->xp_p2buf; 377 RSSTAT_INCR(rscalls); 378 conn = rdp->conn; 379 380 /* 381 * Post a receive descriptor on this 382 * endpoint to ensure all packets are received. 383 */ 384 status = rdma_svc_postrecv(conn); 385 if (status != RDMA_SUCCESS) { 386 cmn_err(CE_NOTE, 387 "svc_rdma_krecv: rdma_svc_postrecv failed %d", status); 388 } 389 390 if (rdp->status != 0) { 391 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 392 RDMA_REL_CONN(conn); 393 RSSTAT_INCR(rsbadcalls); 394 freeb(mp); 395 return (FALSE); 396 } 397 398 /* 399 * Decode rpc message 400 */ 401 xdrs = &clone_xprt->xp_xdrin; 402 xdrmem_create(xdrs, rdp->rpcmsg.addr, rdp->rpcmsg.len, XDR_DECODE); 403 404 /* 405 * Get the XID 406 */ 407 /* 408 * Treat xid as opaque (xid is the first entity 409 * in the rpc rdma message). 410 */ 411 xid = *(uint32_t *)rdp->rpcmsg.addr; 412 /* Skip xid and set the xdr position accordingly. */ 413 XDR_SETPOS(xdrs, sizeof (uint32_t)); 414 if (! xdr_u_int(xdrs, &vers) || 415 ! xdr_u_int(xdrs, &op)) { 416 cmn_err(CE_WARN, "svc_rdma_krecv: xdr_u_int failed"); 417 XDR_DESTROY(xdrs); 418 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 419 RDMA_REL_CONN(conn); 420 freeb(mp); 421 RSSTAT_INCR(rsbadcalls); 422 return (FALSE); 423 } 424 if (op == RDMA_DONE) { 425 /* 426 * Should not get RDMA_DONE 427 */ 428 freeb(mp); 429 XDR_DESTROY(xdrs); 430 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 431 RDMA_REL_CONN(conn); 432 RSSTAT_INCR(rsbadcalls); 433 return (FALSE); /* no response */ 434 } 435 436 #ifdef DEBUG 437 if (rdma_svc_debug) 438 printf("svc_rdma_krecv: recv'd call xid %u\n", xid); 439 #endif 440 /* 441 * Now decode the chunk list 442 */ 443 cl = NULL; 444 if (! xdr_do_clist(xdrs, &cl)) { 445 cmn_err(CE_WARN, "svc_rdma_krecv: xdr_do_clist failed"); 446 } 447 448 /* 449 * A chunk at 0 offset indicates that the RPC call message 450 * is in a chunk. Get the RPC call message chunk. 451 */ 452 if (cl != NULL && op == RDMA_NOMSG) { 453 struct clist *cllong; /* Long RPC chunk */ 454 455 /* Remove RPC call message chunk from chunklist */ 456 cllong = cl; 457 cl = cl->c_next; 458 cllong->c_next = NULL; 459 460 /* Allocate and register memory for the RPC call msg chunk */ 461 cllong->c_daddr = (uint64)(uintptr_t) 462 kmem_alloc(cllong->c_len, KM_SLEEP); 463 if (cllong->c_daddr == NULL) { 464 cmn_err(CE_WARN, 465 "svc_rdma_krecv: no memory for rpc call"); 466 XDR_DESTROY(xdrs); 467 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 468 RDMA_REL_CONN(conn); 469 freeb(mp); 470 RSSTAT_INCR(rsbadcalls); 471 clist_free(cl); 472 clist_free(cllong); 473 return (FALSE); 474 } 475 status = clist_register(conn, cllong, 0); 476 if (status) { 477 cmn_err(CE_WARN, 478 "svc_rdma_krecv: clist_register failed"); 479 kmem_free((void *)(uintptr_t)cllong->c_daddr, 480 cllong->c_len); 481 XDR_DESTROY(xdrs); 482 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 483 RDMA_REL_CONN(conn); 484 freeb(mp); 485 RSSTAT_INCR(rsbadcalls); 486 clist_free(cl); 487 clist_free(cllong); 488 return (FALSE); 489 } 490 491 /* 492 * Now read the RPC call message in 493 */ 494 status = RDMA_READ(conn, cllong, WAIT); 495 if (status) { 496 cmn_err(CE_WARN, 497 "svc_rdma_krecv: rdma_read failed %d", status); 498 (void) clist_deregister(conn, cllong, 0); 499 kmem_free((void *)(uintptr_t)cllong->c_daddr, 500 cllong->c_len); 501 XDR_DESTROY(xdrs); 502 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 503 RDMA_REL_CONN(conn); 504 freeb(mp); 505 RSSTAT_INCR(rsbadcalls); 506 clist_free(cl); 507 clist_free(cllong); 508 return (FALSE); 509 } 510 /* 511 * Sync memory for CPU after DMA 512 */ 513 status = clist_syncmem(conn, cllong, 0); 514 515 /* 516 * Deregister the chunk 517 */ 518 (void) clist_deregister(conn, cllong, 0); 519 520 /* 521 * Setup the XDR for the RPC call message 522 */ 523 xdrrdma_create(xdrs, (caddr_t)(uintptr_t)cllong->c_daddr, 524 cllong->c_len, 0, cl, XDR_DECODE, conn); 525 vd->rpcbuf.type = CHUNK_BUFFER; 526 vd->rpcbuf.addr = (caddr_t)(uintptr_t)cllong->c_daddr; 527 vd->rpcbuf.len = cllong->c_len; 528 vd->rpcbuf.handle.mrc_rmr = 0; 529 530 /* 531 * Free the chunk element with the Long RPC details and 532 * the message received. 533 */ 534 clist_free(cllong); 535 RDMA_BUF_FREE(conn, &rdp->rpcmsg); 536 } else { 537 pos = XDR_GETPOS(xdrs); 538 539 /* 540 * Now the RPC call message header 541 */ 542 xdrrdma_create(xdrs, rdp->rpcmsg.addr + pos, 543 rdp->rpcmsg.len - pos, 0, cl, XDR_DECODE, conn); 544 vd->rpcbuf = rdp->rpcmsg; 545 } 546 if (! xdr_callmsg(xdrs, msg)) { 547 cmn_err(CE_WARN, "svc_rdma_krecv: xdr_callmsg failed"); 548 if (cl != NULL) 549 clist_free(cl); 550 XDR_DESTROY(xdrs); 551 rdma_buf_free(conn, &vd->rpcbuf); 552 RDMA_REL_CONN(conn); 553 freeb(mp); 554 RSSTAT_INCR(rsxdrcall); 555 RSSTAT_INCR(rsbadcalls); 556 return (FALSE); 557 } 558 559 /* 560 * Point the remote transport address in the service_transport 561 * handle at the address in the request. 562 */ 563 clone_xprt->xp_rtaddr.buf = conn->c_raddr.buf; 564 clone_xprt->xp_rtaddr.len = conn->c_raddr.len; 565 clone_xprt->xp_rtaddr.maxlen = conn->c_raddr.len; 566 567 #ifdef DEBUG 568 if (rdma_svc_debug) { 569 struct sockaddr_in *sin4; 570 char print_addr[INET_ADDRSTRLEN]; 571 572 sin4 = (struct sockaddr_in *)clone_xprt->xp_rtaddr.buf; 573 bzero(print_addr, INET_ADDRSTRLEN); 574 (void) inet_ntop(AF_INET, 575 &sin4->sin_addr, print_addr, INET_ADDRSTRLEN); 576 cmn_err(CE_NOTE, 577 "svc_rdma_krecv: remote clnt_addr: %s", print_addr); 578 } 579 #endif 580 581 clone_xprt->xp_xid = xid; 582 vd->conn = conn; 583 freeb(mp); 584 return (TRUE); 585 } 586 587 /* 588 * Send rpc reply. 589 */ 590 static bool_t 591 svc_rdma_ksend(SVCXPRT *clone_xprt, struct rpc_msg *msg) 592 { 593 struct clone_rdma_data *vd; 594 XDR *xdrs = &(clone_xprt->xp_xdrout), rxdrs; 595 int retval = FALSE; 596 xdrproc_t xdr_results; 597 caddr_t xdr_location; 598 bool_t has_args, reg = FALSE; 599 uint_t len, op; 600 uint_t vers; 601 struct clist *cl = NULL, *cle = NULL; 602 struct clist *sendlist = NULL; 603 int status; 604 int msglen; 605 rdma_buf_t clmsg, longreply, rpcreply; 606 607 vd = (struct clone_rdma_data *)clone_xprt->xp_p2buf; 608 609 /* 610 * If there is a result procedure specified in the reply message, 611 * it will be processed in the xdr_replymsg and SVCAUTH_WRAP. 612 * We need to make sure it won't be processed twice, so we null 613 * it for xdr_replymsg here. 614 */ 615 has_args = FALSE; 616 if (msg->rm_reply.rp_stat == MSG_ACCEPTED && 617 msg->rm_reply.rp_acpt.ar_stat == SUCCESS) { 618 if ((xdr_results = msg->acpted_rply.ar_results.proc) != NULL) { 619 has_args = TRUE; 620 xdr_location = msg->acpted_rply.ar_results.where; 621 msg->acpted_rply.ar_results.proc = xdr_void; 622 msg->acpted_rply.ar_results.where = NULL; 623 } 624 } 625 626 /* 627 * Get the size of the rpc reply message. Need this 628 * to determine if the rpc reply message will fit in 629 * the pre-allocated RDMA buffers. If the rpc reply 630 * message length is greater that the pre-allocated 631 * buffers then, a one time use buffer is allocated 632 * and registered for this rpc reply. 633 */ 634 msglen = xdr_sizeof(xdr_replymsg, msg); 635 if (has_args && msg->rm_reply.rp_acpt.ar_verf.oa_flavor != RPCSEC_GSS) { 636 msglen += xdrrdma_sizeof(xdr_results, xdr_location, 637 rdma_minchunk); 638 if (msglen > RPC_MSG_SZ) { 639 640 /* 641 * Allocate chunk buffer for rpc reply 642 */ 643 rpcreply.type = CHUNK_BUFFER; 644 rpcreply.addr = kmem_zalloc(msglen, KM_SLEEP); 645 cle = kmem_zalloc(sizeof (*cle), KM_SLEEP); 646 cle->c_xdroff = 0; 647 cle->c_len = rpcreply.len = msglen; 648 cle->c_saddr = (uint64)(uintptr_t)rpcreply.addr; 649 cle->c_next = NULL; 650 xdrrdma_create(xdrs, rpcreply.addr, msglen, 651 rdma_minchunk, cle, XDR_ENCODE, NULL); 652 op = RDMA_NOMSG; 653 } else { 654 /* 655 * Get a pre-allocated buffer for rpc reply 656 */ 657 rpcreply.type = SEND_BUFFER; 658 if (RDMA_BUF_ALLOC(vd->conn, &rpcreply)) { 659 cmn_err(CE_WARN, 660 "svc_rdma_ksend: no free buffers!"); 661 return (retval); 662 } 663 xdrrdma_create(xdrs, rpcreply.addr, rpcreply.len, 664 rdma_minchunk, NULL, XDR_ENCODE, NULL); 665 op = RDMA_MSG; 666 } 667 668 /* 669 * Initialize the XDR encode stream. 670 */ 671 msg->rm_xid = clone_xprt->xp_xid; 672 673 if (!(xdr_replymsg(xdrs, msg) && 674 (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs, 675 xdr_results, xdr_location)))) { 676 rdma_buf_free(vd->conn, &rpcreply); 677 if (cle) 678 clist_free(cle); 679 cmn_err(CE_WARN, 680 "svc_rdma_ksend: xdr_replymsg/SVCAUTH_WRAP " 681 "failed"); 682 goto out; 683 } 684 len = XDR_GETPOS(xdrs); 685 } 686 if (has_args && msg->rm_reply.rp_acpt.ar_verf.oa_flavor == RPCSEC_GSS) { 687 688 /* 689 * For RPCSEC_GSS since we cannot accurately presize the 690 * buffer required for encoding, we assume that its going 691 * to be a Long RPC to start with. We also create the 692 * the XDR stream with min_chunk set to 0 which instructs 693 * the XDR layer to not chunk the incoming byte stream. 694 */ 695 msglen += 2 * MAX_AUTH_BYTES + 2 * sizeof (struct opaque_auth); 696 msglen += xdr_sizeof(xdr_results, xdr_location); 697 698 /* 699 * Long RPC. Allocate one time use custom buffer. 700 */ 701 longreply.type = CHUNK_BUFFER; 702 longreply.addr = kmem_zalloc(msglen, KM_SLEEP); 703 cle = kmem_zalloc(sizeof (*cle), KM_SLEEP); 704 cle->c_xdroff = 0; 705 cle->c_len = longreply.len = msglen; 706 cle->c_saddr = (uint64)(uintptr_t)longreply.addr; 707 cle->c_next = NULL; 708 xdrrdma_create(xdrs, longreply.addr, msglen, 0, cle, 709 XDR_ENCODE, NULL); 710 op = RDMA_NOMSG; 711 /* 712 * Initialize the XDR encode stream. 713 */ 714 msg->rm_xid = clone_xprt->xp_xid; 715 716 if (!(xdr_replymsg(xdrs, msg) && 717 (!has_args || SVCAUTH_WRAP(&clone_xprt->xp_auth, xdrs, 718 xdr_results, xdr_location)))) { 719 if (longreply.addr != xdrs->x_base) { 720 longreply.addr = xdrs->x_base; 721 longreply.len = xdr_getbufsize(xdrs); 722 } 723 rdma_buf_free(vd->conn, &longreply); 724 if (cle) 725 clist_free(cle); 726 cmn_err(CE_WARN, 727 "svc_rdma_ksend: xdr_replymsg/SVCAUTH_WRAP " 728 "failed"); 729 goto out; 730 } 731 732 /* 733 * If we had to allocate a new buffer while encoding 734 * then update the addr and len. 735 */ 736 if (longreply.addr != xdrs->x_base) { 737 longreply.addr = xdrs->x_base; 738 longreply.len = xdr_getbufsize(xdrs); 739 } 740 741 len = XDR_GETPOS(xdrs); 742 743 /* 744 * If it so happens that the encoded message is after all 745 * not long enough to be a Long RPC then allocate a 746 * SEND_BUFFER and copy the encoded message into it. 747 */ 748 if (len > RPC_MSG_SZ) { 749 rpcreply.type = CHUNK_BUFFER; 750 rpcreply.addr = longreply.addr; 751 rpcreply.len = longreply.len; 752 } else { 753 clist_free(cle); 754 XDR_DESTROY(xdrs); 755 /* 756 * Get a pre-allocated buffer for rpc reply 757 */ 758 rpcreply.type = SEND_BUFFER; 759 if (RDMA_BUF_ALLOC(vd->conn, &rpcreply)) { 760 cmn_err(CE_WARN, 761 "svc_rdma_ksend: no free buffers!"); 762 rdma_buf_free(vd->conn, &longreply); 763 return (retval); 764 } 765 bcopy(longreply.addr, rpcreply.addr, len); 766 xdrrdma_create(xdrs, rpcreply.addr, len, 0, NULL, 767 XDR_ENCODE, NULL); 768 rdma_buf_free(vd->conn, &longreply); 769 op = RDMA_MSG; 770 } 771 } 772 773 if (has_args == FALSE) { 774 775 if (msglen > RPC_MSG_SZ) { 776 777 /* 778 * Allocate chunk buffer for rpc reply 779 */ 780 rpcreply.type = CHUNK_BUFFER; 781 rpcreply.addr = kmem_zalloc(msglen, KM_SLEEP); 782 cle = kmem_zalloc(sizeof (*cle), KM_SLEEP); 783 cle->c_xdroff = 0; 784 cle->c_len = rpcreply.len = msglen; 785 cle->c_saddr = (uint64)(uintptr_t)rpcreply.addr; 786 cle->c_next = NULL; 787 xdrrdma_create(xdrs, rpcreply.addr, msglen, 788 rdma_minchunk, cle, XDR_ENCODE, NULL); 789 op = RDMA_NOMSG; 790 } else { 791 /* 792 * Get a pre-allocated buffer for rpc reply 793 */ 794 rpcreply.type = SEND_BUFFER; 795 if (RDMA_BUF_ALLOC(vd->conn, &rpcreply)) { 796 cmn_err(CE_WARN, 797 "svc_rdma_ksend: no free buffers!"); 798 return (retval); 799 } 800 xdrrdma_create(xdrs, rpcreply.addr, rpcreply.len, 801 rdma_minchunk, NULL, XDR_ENCODE, NULL); 802 op = RDMA_MSG; 803 } 804 805 /* 806 * Initialize the XDR encode stream. 807 */ 808 msg->rm_xid = clone_xprt->xp_xid; 809 810 if (!xdr_replymsg(xdrs, msg)) { 811 rdma_buf_free(vd->conn, &rpcreply); 812 if (cle) 813 clist_free(cle); 814 cmn_err(CE_WARN, 815 "svc_rdma_ksend: xdr_replymsg/SVCAUTH_WRAP " 816 "failed"); 817 goto out; 818 } 819 len = XDR_GETPOS(xdrs); 820 } 821 822 /* 823 * Get clist and a buffer for sending it across 824 */ 825 cl = xdrrdma_clist(xdrs); 826 clmsg.type = SEND_BUFFER; 827 if (RDMA_BUF_ALLOC(vd->conn, &clmsg)) { 828 rdma_buf_free(vd->conn, &rpcreply); 829 cmn_err(CE_WARN, "svc_rdma_ksend: no free buffers!!"); 830 goto out; 831 } 832 833 /* 834 * Now register the chunks in the list 835 */ 836 if (cl != NULL) { 837 status = clist_register(vd->conn, cl, 1); 838 if (status != RDMA_SUCCESS) { 839 rdma_buf_free(vd->conn, &clmsg); 840 cmn_err(CE_WARN, 841 "svc_rdma_ksend: clist register failed"); 842 goto out; 843 } 844 reg = TRUE; 845 } 846 847 /* 848 * XDR the XID, vers, and op 849 */ 850 /* 851 * Treat xid as opaque (xid is the first entity 852 * in the rpc rdma message). 853 */ 854 vers = RPCRDMA_VERS; 855 xdrs = &rxdrs; 856 xdrmem_create(xdrs, clmsg.addr, clmsg.len, XDR_ENCODE); 857 (*(uint32_t *)clmsg.addr) = msg->rm_xid; 858 /* Skip xid and set the xdr position accordingly. */ 859 XDR_SETPOS(xdrs, sizeof (uint32_t)); 860 if (! xdr_u_int(xdrs, &vers) || 861 ! xdr_u_int(xdrs, &op)) { 862 rdma_buf_free(vd->conn, &rpcreply); 863 rdma_buf_free(vd->conn, &clmsg); 864 cmn_err(CE_WARN, "svc_rdma_ksend: xdr_u_int failed"); 865 goto out; 866 } 867 868 /* 869 * Now XDR the chunk list 870 */ 871 (void) xdr_do_clist(xdrs, &cl); 872 873 clist_add(&sendlist, 0, XDR_GETPOS(xdrs), &clmsg.handle, clmsg.addr, 874 NULL, NULL); 875 876 if (op == RDMA_MSG) { 877 clist_add(&sendlist, 0, len, &rpcreply.handle, rpcreply.addr, 878 NULL, NULL); 879 } else { 880 cl->c_len = len; 881 RSSTAT_INCR(rslongrpcs); 882 } 883 884 /* 885 * Send the reply message to the client 886 */ 887 if (cl != NULL) { 888 status = clist_syncmem(vd->conn, cl, 1); 889 if (status != RDMA_SUCCESS) { 890 rdma_buf_free(vd->conn, &rpcreply); 891 rdma_buf_free(vd->conn, &clmsg); 892 goto out; 893 } 894 #ifdef DEBUG 895 if (rdma_svc_debug) 896 printf("svc_rdma_ksend: chunk response len %d xid %u\n", 897 cl->c_len, msg->rm_xid); 898 #endif 899 /* 900 * Post a receive buffer because we expect a RDMA_DONE 901 * message. 902 */ 903 status = rdma_svc_postrecv(vd->conn); 904 905 /* 906 * Send the RPC reply message and wait for RDMA_DONE 907 */ 908 status = RDMA_SEND_RESP(vd->conn, sendlist, msg->rm_xid); 909 if (status != RDMA_SUCCESS) { 910 #ifdef DEBUG 911 if (rdma_svc_debug) 912 cmn_err(CE_NOTE, "svc_rdma_ksend: " 913 "rdma_send_resp failed %d", status); 914 #endif 915 goto out; 916 } 917 #ifdef DEBUG 918 if (rdma_svc_debug) 919 printf("svc_rdma_ksend: got RDMA_DONE xid %u\n", msg->rm_xid); 920 #endif 921 } else { 922 #ifdef DEBUG 923 if (rdma_svc_debug) 924 printf("svc_rdma_ksend: msg response xid %u\n", msg->rm_xid); 925 #endif 926 status = RDMA_SEND(vd->conn, sendlist, msg->rm_xid); 927 if (status != RDMA_SUCCESS) { 928 #ifdef DEBUG 929 if (rdma_svc_debug) 930 cmn_err(CE_NOTE, "svc_rdma_ksend: " 931 "rdma_send failed %d", status); 932 #endif 933 goto out; 934 } 935 } 936 937 retval = TRUE; 938 out: 939 /* 940 * Deregister the chunks 941 */ 942 if (cl != NULL) { 943 if (reg) 944 (void) clist_deregister(vd->conn, cl, 1); 945 if (op == RDMA_NOMSG) { 946 /* 947 * Long RPC reply in chunk. Free it up. 948 */ 949 rdma_buf_free(vd->conn, &rpcreply); 950 } 951 clist_free(cl); 952 } 953 954 /* 955 * Free up sendlist chunks 956 */ 957 if (sendlist != NULL) 958 clist_free(sendlist); 959 960 /* 961 * Destroy private data for xdr rdma 962 */ 963 XDR_DESTROY(&(clone_xprt->xp_xdrout)); 964 965 /* 966 * This is completely disgusting. If public is set it is 967 * a pointer to a structure whose first field is the address 968 * of the function to free that structure and any related 969 * stuff. (see rrokfree in nfs_xdr.c). 970 */ 971 if (xdrs->x_public) { 972 /* LINTED pointer alignment */ 973 (**((int (**)())xdrs->x_public))(xdrs->x_public); 974 } 975 976 return (retval); 977 } 978 979 /* 980 * Deserialize arguments. 981 */ 982 static bool_t 983 svc_rdma_kgetargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, caddr_t args_ptr) 984 { 985 if ((SVCAUTH_UNWRAP(&clone_xprt->xp_auth, &clone_xprt->xp_xdrin, 986 xdr_args, args_ptr)) != TRUE) 987 return (FALSE); 988 return (TRUE); 989 } 990 991 static bool_t 992 svc_rdma_kfreeargs(SVCXPRT *clone_xprt, xdrproc_t xdr_args, 993 caddr_t args_ptr) 994 { 995 struct clone_rdma_data *vd; 996 bool_t retval; 997 998 vd = (struct clone_rdma_data *)clone_xprt->xp_p2buf; 999 if (args_ptr) { 1000 XDR *xdrs = &clone_xprt->xp_xdrin; 1001 struct clist *cl; 1002 1003 cl = xdrrdma_clist(xdrs); 1004 if (cl != NULL) 1005 clist_free(cl); 1006 1007 xdrs->x_op = XDR_FREE; 1008 retval = (*xdr_args)(xdrs, args_ptr); 1009 } 1010 XDR_DESTROY(&(clone_xprt->xp_xdrin)); 1011 rdma_buf_free(vd->conn, &vd->rpcbuf); 1012 RDMA_REL_CONN(vd->conn); 1013 return (retval); 1014 } 1015 1016 /* ARGSUSED */ 1017 static int32_t * 1018 svc_rdma_kgetres(SVCXPRT *clone_xprt, int size) 1019 { 1020 return (NULL); 1021 } 1022 1023 /* ARGSUSED */ 1024 static void 1025 svc_rdma_kfreeres(SVCXPRT *clone_xprt) 1026 { 1027 } 1028 1029 /* 1030 * the dup cacheing routines below provide a cache of non-failure 1031 * transaction id's. rpc service routines can use this to detect 1032 * retransmissions and re-send a non-failure response. 1033 */ 1034 1035 /* 1036 * MAXDUPREQS is the number of cached items. It should be adjusted 1037 * to the service load so that there is likely to be a response entry 1038 * when the first retransmission comes in. 1039 */ 1040 #define MAXDUPREQS 1024 1041 1042 /* 1043 * This should be appropriately scaled to MAXDUPREQS. 1044 */ 1045 #define DRHASHSZ 257 1046 1047 #if ((DRHASHSZ & (DRHASHSZ - 1)) == 0) 1048 #define XIDHASH(xid) ((xid) & (DRHASHSZ - 1)) 1049 #else 1050 #define XIDHASH(xid) ((xid) % DRHASHSZ) 1051 #endif 1052 #define DRHASH(dr) XIDHASH((dr)->dr_xid) 1053 #define REQTOXID(req) ((req)->rq_xprt->xp_xid) 1054 1055 static int rdmandupreqs = 0; 1056 static int rdmamaxdupreqs = MAXDUPREQS; 1057 static kmutex_t rdmadupreq_lock; 1058 static struct dupreq *rdmadrhashtbl[DRHASHSZ]; 1059 static int rdmadrhashstat[DRHASHSZ]; 1060 1061 static void unhash(struct dupreq *); 1062 1063 /* 1064 * rdmadrmru points to the head of a circular linked list in lru order. 1065 * rdmadrmru->dr_next == drlru 1066 */ 1067 struct dupreq *rdmadrmru; 1068 1069 /* 1070 * svc_rdma_kdup searches the request cache and returns 0 if the 1071 * request is not found in the cache. If it is found, then it 1072 * returns the state of the request (in progress or done) and 1073 * the status or attributes that were part of the original reply. 1074 */ 1075 static int 1076 svc_rdma_kdup(struct svc_req *req, caddr_t res, int size, struct dupreq **drpp, 1077 bool_t *dupcachedp) 1078 { 1079 struct dupreq *dr; 1080 uint32_t xid; 1081 uint32_t drhash; 1082 int status; 1083 1084 xid = REQTOXID(req); 1085 mutex_enter(&rdmadupreq_lock); 1086 RSSTAT_INCR(rsdupchecks); 1087 /* 1088 * Check to see whether an entry already exists in the cache. 1089 */ 1090 dr = rdmadrhashtbl[XIDHASH(xid)]; 1091 while (dr != NULL) { 1092 if (dr->dr_xid == xid && 1093 dr->dr_proc == req->rq_proc && 1094 dr->dr_prog == req->rq_prog && 1095 dr->dr_vers == req->rq_vers && 1096 dr->dr_addr.len == req->rq_xprt->xp_rtaddr.len && 1097 bcmp((caddr_t)dr->dr_addr.buf, 1098 (caddr_t)req->rq_xprt->xp_rtaddr.buf, 1099 dr->dr_addr.len) == 0) { 1100 status = dr->dr_status; 1101 if (status == DUP_DONE) { 1102 bcopy(dr->dr_resp.buf, res, size); 1103 if (dupcachedp != NULL) 1104 *dupcachedp = (dr->dr_resfree != NULL); 1105 } else { 1106 dr->dr_status = DUP_INPROGRESS; 1107 *drpp = dr; 1108 } 1109 RSSTAT_INCR(rsdupreqs); 1110 mutex_exit(&rdmadupreq_lock); 1111 return (status); 1112 } 1113 dr = dr->dr_chain; 1114 } 1115 1116 /* 1117 * There wasn't an entry, either allocate a new one or recycle 1118 * an old one. 1119 */ 1120 if (rdmandupreqs < rdmamaxdupreqs) { 1121 dr = kmem_alloc(sizeof (*dr), KM_NOSLEEP); 1122 if (dr == NULL) { 1123 mutex_exit(&rdmadupreq_lock); 1124 return (DUP_ERROR); 1125 } 1126 dr->dr_resp.buf = NULL; 1127 dr->dr_resp.maxlen = 0; 1128 dr->dr_addr.buf = NULL; 1129 dr->dr_addr.maxlen = 0; 1130 if (rdmadrmru) { 1131 dr->dr_next = rdmadrmru->dr_next; 1132 rdmadrmru->dr_next = dr; 1133 } else { 1134 dr->dr_next = dr; 1135 } 1136 rdmandupreqs++; 1137 } else { 1138 dr = rdmadrmru->dr_next; 1139 while (dr->dr_status == DUP_INPROGRESS) { 1140 dr = dr->dr_next; 1141 if (dr == rdmadrmru->dr_next) { 1142 cmn_err(CE_WARN, "svc_rdma_kdup no slots free"); 1143 mutex_exit(&rdmadupreq_lock); 1144 return (DUP_ERROR); 1145 } 1146 } 1147 unhash(dr); 1148 if (dr->dr_resfree) { 1149 (*dr->dr_resfree)(dr->dr_resp.buf); 1150 } 1151 } 1152 dr->dr_resfree = NULL; 1153 rdmadrmru = dr; 1154 1155 dr->dr_xid = REQTOXID(req); 1156 dr->dr_prog = req->rq_prog; 1157 dr->dr_vers = req->rq_vers; 1158 dr->dr_proc = req->rq_proc; 1159 if (dr->dr_addr.maxlen < req->rq_xprt->xp_rtaddr.len) { 1160 if (dr->dr_addr.buf != NULL) 1161 kmem_free(dr->dr_addr.buf, dr->dr_addr.maxlen); 1162 dr->dr_addr.maxlen = req->rq_xprt->xp_rtaddr.len; 1163 dr->dr_addr.buf = kmem_alloc(dr->dr_addr.maxlen, KM_NOSLEEP); 1164 if (dr->dr_addr.buf == NULL) { 1165 dr->dr_addr.maxlen = 0; 1166 dr->dr_status = DUP_DROP; 1167 mutex_exit(&rdmadupreq_lock); 1168 return (DUP_ERROR); 1169 } 1170 } 1171 dr->dr_addr.len = req->rq_xprt->xp_rtaddr.len; 1172 bcopy(req->rq_xprt->xp_rtaddr.buf, dr->dr_addr.buf, dr->dr_addr.len); 1173 if (dr->dr_resp.maxlen < size) { 1174 if (dr->dr_resp.buf != NULL) 1175 kmem_free(dr->dr_resp.buf, dr->dr_resp.maxlen); 1176 dr->dr_resp.maxlen = (unsigned int)size; 1177 dr->dr_resp.buf = kmem_alloc(size, KM_NOSLEEP); 1178 if (dr->dr_resp.buf == NULL) { 1179 dr->dr_resp.maxlen = 0; 1180 dr->dr_status = DUP_DROP; 1181 mutex_exit(&rdmadupreq_lock); 1182 return (DUP_ERROR); 1183 } 1184 } 1185 dr->dr_status = DUP_INPROGRESS; 1186 1187 drhash = (uint32_t)DRHASH(dr); 1188 dr->dr_chain = rdmadrhashtbl[drhash]; 1189 rdmadrhashtbl[drhash] = dr; 1190 rdmadrhashstat[drhash]++; 1191 mutex_exit(&rdmadupreq_lock); 1192 *drpp = dr; 1193 return (DUP_NEW); 1194 } 1195 1196 /* 1197 * svc_rdma_kdupdone marks the request done (DUP_DONE or DUP_DROP) 1198 * and stores the response. 1199 */ 1200 static void 1201 svc_rdma_kdupdone(struct dupreq *dr, caddr_t res, void (*dis_resfree)(), 1202 int size, int status) 1203 { 1204 ASSERT(dr->dr_resfree == NULL); 1205 if (status == DUP_DONE) { 1206 bcopy(res, dr->dr_resp.buf, size); 1207 dr->dr_resfree = dis_resfree; 1208 } 1209 dr->dr_status = status; 1210 } 1211 1212 /* 1213 * This routine expects that the mutex, rdmadupreq_lock, is already held. 1214 */ 1215 static void 1216 unhash(struct dupreq *dr) 1217 { 1218 struct dupreq *drt; 1219 struct dupreq *drtprev = NULL; 1220 uint32_t drhash; 1221 1222 ASSERT(MUTEX_HELD(&rdmadupreq_lock)); 1223 1224 drhash = (uint32_t)DRHASH(dr); 1225 drt = rdmadrhashtbl[drhash]; 1226 while (drt != NULL) { 1227 if (drt == dr) { 1228 rdmadrhashstat[drhash]--; 1229 if (drtprev == NULL) { 1230 rdmadrhashtbl[drhash] = drt->dr_chain; 1231 } else { 1232 drtprev->dr_chain = drt->dr_chain; 1233 } 1234 return; 1235 } 1236 drtprev = drt; 1237 drt = drt->dr_chain; 1238 } 1239 } 1240