/* * Copyright (c) 2009 Mellanox Technologies Ltd. All rights reserved. * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenIB.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "sdp.h" SDP_MODPARAM_INT(rcvbuf_initial_size, 32 * 1024, "Receive buffer initial size in bytes."); SDP_MODPARAM_SINT(rcvbuf_scale, 0x8, "Receive buffer size scale factor."); /* Like tcp_fin - called when SDP_MID_DISCONNECT is received */ static void sdp_handle_disconn(struct sdp_sock *ssk) { sdp_dbg(ssk->socket, "%s\n", __func__); SDP_WLOCK_ASSERT(ssk); if (TCPS_HAVERCVDFIN(ssk->state) == 0) socantrcvmore(ssk->socket); switch (ssk->state) { case TCPS_SYN_RECEIVED: case TCPS_ESTABLISHED: ssk->state = TCPS_CLOSE_WAIT; break; case TCPS_FIN_WAIT_1: /* Received a reply FIN - start Infiniband tear down */ sdp_dbg(ssk->socket, "%s: Starting Infiniband tear down sending DREQ\n", __func__); sdp_cancel_dreq_wait_timeout(ssk); ssk->qp_active = 0; if (ssk->id) { struct rdma_cm_id *id; id = ssk->id; SDP_WUNLOCK(ssk); rdma_disconnect(id); SDP_WLOCK(ssk); } else { sdp_warn(ssk->socket, "%s: ssk->id is NULL\n", __func__); return; } break; case TCPS_TIME_WAIT: /* This is a mutual close situation and we've got the DREQ from the peer before the SDP_MID_DISCONNECT */ break; case TCPS_CLOSED: /* FIN arrived after IB teardown started - do nothing */ sdp_dbg(ssk->socket, "%s: fin in state %s\n", __func__, sdp_state_str(ssk->state)); return; default: sdp_warn(ssk->socket, "%s: FIN in unexpected state. state=%d\n", __func__, ssk->state); break; } } static int sdp_post_recv(struct sdp_sock *ssk) { struct sdp_buf *rx_req; int i, rc; u64 addr; struct ib_device *dev; struct ib_recv_wr rx_wr = { NULL }; struct ib_sge ibsge[SDP_MAX_RECV_SGES]; struct ib_sge *sge = ibsge; struct ib_recv_wr *bad_wr; struct mbuf *mb, *m; struct sdp_bsdh *h; int id = ring_head(ssk->rx_ring); /* Now, allocate and repost recv */ sdp_prf(ssk->socket, mb, "Posting mb"); mb = m_getm2(NULL, ssk->recv_bytes, M_NOWAIT, MT_DATA, M_PKTHDR); if (mb == NULL) { /* Retry so we can't stall out with no memory. */ if (!rx_ring_posted(ssk)) queue_work(rx_comp_wq, &ssk->rx_comp_work); return -1; } for (m = mb; m != NULL; m = m->m_next) { m->m_len = M_SIZE(m); mb->m_pkthdr.len += m->m_len; } h = mtod(mb, struct sdp_bsdh *); rx_req = ssk->rx_ring.buffer + (id & (SDP_RX_SIZE - 1)); rx_req->mb = mb; dev = ssk->ib_device; for (i = 0; mb != NULL; i++, mb = mb->m_next, sge++) { addr = ib_dma_map_single(dev, mb->m_data, mb->m_len, DMA_TO_DEVICE); /* TODO: proper error handling */ BUG_ON(ib_dma_mapping_error(dev, addr)); BUG_ON(i >= SDP_MAX_RECV_SGES); rx_req->mapping[i] = addr; sge->addr = addr; sge->length = mb->m_len; sge->lkey = ssk->sdp_dev->mr->lkey; } rx_wr.next = NULL; rx_wr.wr_id = id | SDP_OP_RECV; rx_wr.sg_list = ibsge; rx_wr.num_sge = i; rc = ib_post_recv(ssk->qp, &rx_wr, &bad_wr); if (unlikely(rc)) { sdp_warn(ssk->socket, "ib_post_recv failed. status %d\n", rc); sdp_cleanup_sdp_buf(ssk, rx_req, DMA_FROM_DEVICE); m_freem(mb); sdp_notify(ssk, ECONNRESET); return -1; } atomic_inc(&ssk->rx_ring.head); SDPSTATS_COUNTER_INC(post_recv); return 0; } static inline int sdp_post_recvs_needed(struct sdp_sock *ssk) { unsigned long bytes_in_process; unsigned long max_bytes; int buffer_size; int posted; if (!ssk->qp_active || !ssk->socket) return 0; posted = rx_ring_posted(ssk); if (posted >= SDP_RX_SIZE) return 0; if (posted < SDP_MIN_TX_CREDITS) return 1; buffer_size = ssk->recv_bytes; max_bytes = max(ssk->socket->so_snd.sb_hiwat, (1 + SDP_MIN_TX_CREDITS) * buffer_size); max_bytes *= rcvbuf_scale; /* * Compute bytes in the receive queue and socket buffer. */ bytes_in_process = (posted - SDP_MIN_TX_CREDITS) * buffer_size; bytes_in_process += sbused(&ssk->socket->so_rcv); return bytes_in_process < max_bytes; } static inline void sdp_post_recvs(struct sdp_sock *ssk) { while (sdp_post_recvs_needed(ssk)) if (sdp_post_recv(ssk)) return; } static inline struct mbuf * sdp_sock_queue_rcv_mb(struct socket *sk, struct mbuf *mb) { struct sdp_sock *ssk = sdp_sk(sk); struct sdp_bsdh *h; h = mtod(mb, struct sdp_bsdh *); #ifdef SDP_ZCOPY SDP_SKB_CB(mb)->seq = rcv_nxt(ssk); if (h->mid == SDP_MID_SRCAVAIL) { struct sdp_srcah *srcah = (struct sdp_srcah *)(h+1); struct rx_srcavail_state *rx_sa; ssk->srcavail_cancel_mseq = 0; ssk->rx_sa = rx_sa = RX_SRCAVAIL_STATE(mb) = kzalloc( sizeof(struct rx_srcavail_state), M_NOWAIT); rx_sa->mseq = ntohl(h->mseq); rx_sa->used = 0; rx_sa->len = mb_len = ntohl(srcah->len); rx_sa->rkey = ntohl(srcah->rkey); rx_sa->vaddr = be64_to_cpu(srcah->vaddr); rx_sa->flags = 0; if (ssk->tx_sa) { sdp_dbg_data(ssk->socket, "got RX SrcAvail while waiting " "for TX SrcAvail. waking up TX SrcAvail" "to be aborted\n"); wake_up(sk->sk_sleep); } atomic_add(mb->len, &ssk->rcv_nxt); sdp_dbg_data(sk, "queueing SrcAvail. mb_len = %d vaddr = %lld\n", mb_len, rx_sa->vaddr); } else #endif { atomic_add(mb->m_pkthdr.len, &ssk->rcv_nxt); } m_adj(mb, SDP_HEAD_SIZE); SOCKBUF_LOCK(&sk->so_rcv); if (unlikely(h->flags & SDP_OOB_PRES)) sdp_urg(ssk, mb); sbappend_locked(&sk->so_rcv, mb); sorwakeup_locked(sk); return mb; } static int sdp_get_recv_bytes(struct sdp_sock *ssk, u32 new_size) { return MIN(new_size, SDP_MAX_PACKET); } int sdp_init_buffers(struct sdp_sock *ssk, u32 new_size) { ssk->recv_bytes = sdp_get_recv_bytes(ssk, new_size); sdp_post_recvs(ssk); return 0; } int sdp_resize_buffers(struct sdp_sock *ssk, u32 new_size) { u32 curr_size = ssk->recv_bytes; u32 max_size = SDP_MAX_PACKET; if (new_size > curr_size && new_size <= max_size) { ssk->recv_bytes = sdp_get_recv_bytes(ssk, new_size); return 0; } return -1; } static void sdp_handle_resize_request(struct sdp_sock *ssk, struct sdp_chrecvbuf *buf) { if (sdp_resize_buffers(ssk, ntohl(buf->size)) == 0) ssk->recv_request_head = ring_head(ssk->rx_ring) + 1; else ssk->recv_request_head = ring_tail(ssk->rx_ring); ssk->recv_request = 1; } static void sdp_handle_resize_ack(struct sdp_sock *ssk, struct sdp_chrecvbuf *buf) { u32 new_size = ntohl(buf->size); if (new_size > ssk->xmit_size_goal) ssk->xmit_size_goal = new_size; } static struct mbuf * sdp_recv_completion(struct sdp_sock *ssk, int id) { struct sdp_buf *rx_req; struct ib_device *dev; struct mbuf *mb; if (unlikely(id != ring_tail(ssk->rx_ring))) { printk(KERN_WARNING "Bogus recv completion id %d tail %d\n", id, ring_tail(ssk->rx_ring)); return NULL; } dev = ssk->ib_device; rx_req = &ssk->rx_ring.buffer[id & (SDP_RX_SIZE - 1)]; mb = rx_req->mb; sdp_cleanup_sdp_buf(ssk, rx_req, DMA_FROM_DEVICE); atomic_inc(&ssk->rx_ring.tail); atomic_dec(&ssk->remote_credits); return mb; } /* socket lock should be taken before calling this */ static int sdp_process_rx_ctl_mb(struct sdp_sock *ssk, struct mbuf *mb) { struct sdp_bsdh *h; struct socket *sk; SDP_WLOCK_ASSERT(ssk); sk = ssk->socket; h = mtod(mb, struct sdp_bsdh *); switch (h->mid) { case SDP_MID_DATA: case SDP_MID_SRCAVAIL: sdp_dbg(sk, "DATA after socket rcv was shutdown\n"); /* got data in RCV_SHUTDOWN */ if (ssk->state == TCPS_FIN_WAIT_1) { sdp_dbg(sk, "RX data when state = FIN_WAIT1\n"); sdp_notify(ssk, ECONNRESET); } m_freem(mb); break; #ifdef SDP_ZCOPY case SDP_MID_RDMARDCOMPL: m_freem(mb); break; case SDP_MID_SENDSM: sdp_handle_sendsm(ssk, ntohl(h->mseq_ack)); m_freem(mb); break; case SDP_MID_SRCAVAIL_CANCEL: sdp_dbg_data(sk, "Handling SrcAvailCancel\n"); sdp_prf(sk, NULL, "Handling SrcAvailCancel"); if (ssk->rx_sa) { ssk->srcavail_cancel_mseq = ntohl(h->mseq); ssk->rx_sa->flags |= RX_SA_ABORTED; ssk->rx_sa = NULL; /* TODO: change it into SDP_MID_DATA and get the dirty logic from recvmsg */ } else { sdp_dbg(sk, "Got SrcAvailCancel - " "but no SrcAvail in process\n"); } m_freem(mb); break; case SDP_MID_SINKAVAIL: sdp_dbg_data(sk, "Got SinkAvail - not supported: ignored\n"); sdp_prf(sk, NULL, "Got SinkAvail - not supported: ignored"); /* FALLTHROUGH */ #endif case SDP_MID_ABORT: sdp_dbg_data(sk, "Handling ABORT\n"); sdp_prf(sk, NULL, "Handling ABORT"); sdp_notify(ssk, ECONNRESET); m_freem(mb); break; case SDP_MID_DISCONN: sdp_dbg_data(sk, "Handling DISCONN\n"); sdp_prf(sk, NULL, "Handling DISCONN"); sdp_handle_disconn(ssk); break; case SDP_MID_CHRCVBUF: sdp_dbg_data(sk, "Handling RX CHRCVBUF\n"); sdp_handle_resize_request(ssk, (struct sdp_chrecvbuf *)(h+1)); m_freem(mb); break; case SDP_MID_CHRCVBUF_ACK: sdp_dbg_data(sk, "Handling RX CHRCVBUF_ACK\n"); sdp_handle_resize_ack(ssk, (struct sdp_chrecvbuf *)(h+1)); m_freem(mb); break; default: /* TODO: Handle other messages */ sdp_warn(sk, "SDP: FIXME MID %d\n", h->mid); m_freem(mb); } return 0; } static int sdp_process_rx_mb(struct sdp_sock *ssk, struct mbuf *mb) { struct socket *sk; struct sdp_bsdh *h; unsigned long mseq_ack; int credits_before; h = mtod(mb, struct sdp_bsdh *); sk = ssk->socket; /* * If another thread is in so_pcbfree this may be partially torn * down but no further synchronization is required as the destroying * thread will wait for receive to shutdown before discarding the * socket. */ if (sk == NULL) { m_freem(mb); return 0; } SDPSTATS_HIST_LINEAR(credits_before_update, tx_credits(ssk)); mseq_ack = ntohl(h->mseq_ack); credits_before = tx_credits(ssk); atomic_set(&ssk->tx_ring.credits, mseq_ack - ring_head(ssk->tx_ring) + 1 + ntohs(h->bufs)); if (mseq_ack >= ssk->nagle_last_unacked) ssk->nagle_last_unacked = 0; sdp_prf1(ssk->socket, mb, "RX %s +%d c:%d->%d mseq:%d ack:%d\n", mid2str(h->mid), ntohs(h->bufs), credits_before, tx_credits(ssk), ntohl(h->mseq), ntohl(h->mseq_ack)); if (unlikely(h->mid == SDP_MID_DATA && mb->m_pkthdr.len == SDP_HEAD_SIZE)) { /* Credit update is valid even after RCV_SHUTDOWN */ m_freem(mb); return 0; } if ((h->mid != SDP_MID_DATA && h->mid != SDP_MID_SRCAVAIL) || TCPS_HAVERCVDFIN(ssk->state)) { sdp_prf(sk, NULL, "Control mb - queing to control queue"); #ifdef SDP_ZCOPY if (h->mid == SDP_MID_SRCAVAIL_CANCEL) { sdp_dbg_data(sk, "Got SrcAvailCancel. " "seq: 0x%d seq_ack: 0x%d\n", ntohl(h->mseq), ntohl(h->mseq_ack)); ssk->srcavail_cancel_mseq = ntohl(h->mseq); } if (h->mid == SDP_MID_RDMARDCOMPL) { struct sdp_rrch *rrch = (struct sdp_rrch *)(h+1); sdp_dbg_data(sk, "RdmaRdCompl message arrived\n"); sdp_handle_rdma_read_compl(ssk, ntohl(h->mseq_ack), ntohl(rrch->len)); } #endif mb->m_nextpkt = NULL; if (ssk->rx_ctl_tail) ssk->rx_ctl_tail->m_nextpkt = mb; else ssk->rx_ctl_q = mb; ssk->rx_ctl_tail = mb; return 0; } sdp_prf1(sk, NULL, "queueing %s mb\n", mid2str(h->mid)); mb = sdp_sock_queue_rcv_mb(sk, mb); return 0; } /* called only from irq */ static struct mbuf * sdp_process_rx_wc(struct sdp_sock *ssk, struct ib_wc *wc) { struct mbuf *mb; struct sdp_bsdh *h; struct socket *sk = ssk->socket; int mseq; mb = sdp_recv_completion(ssk, wc->wr_id); if (unlikely(!mb)) return NULL; if (unlikely(wc->status)) { if (ssk->qp_active && sk) { sdp_dbg(sk, "Recv completion with error. " "Status %d, vendor: %d\n", wc->status, wc->vendor_err); sdp_abort(sk); ssk->qp_active = 0; } m_freem(mb); return NULL; } sdp_dbg_data(sk, "Recv completion. ID %d Length %d\n", (int)wc->wr_id, wc->byte_len); if (unlikely(wc->byte_len < sizeof(struct sdp_bsdh))) { sdp_warn(sk, "SDP BUG! byte_len %d < %zd\n", wc->byte_len, sizeof(struct sdp_bsdh)); m_freem(mb); return NULL; } /* Use m_adj to trim the tail of data we didn't use. */ m_adj(mb, -(mb->m_pkthdr.len - wc->byte_len)); h = mtod(mb, struct sdp_bsdh *); SDP_DUMP_PACKET(ssk->socket, "RX", mb, h); ssk->rx_packets++; ssk->rx_bytes += mb->m_pkthdr.len; mseq = ntohl(h->mseq); atomic_set(&ssk->mseq_ack, mseq); if (mseq != (int)wc->wr_id) sdp_warn(sk, "SDP BUG! mseq %d != wrid %d\n", mseq, (int)wc->wr_id); return mb; } /* Wakeup writers if we now have credits. */ static void sdp_bzcopy_write_space(struct sdp_sock *ssk) { struct socket *sk = ssk->socket; if (tx_credits(ssk) >= ssk->min_bufs && sk) sowwakeup(sk); } /* only from interrupt. */ static int sdp_poll_rx_cq(struct sdp_sock *ssk) { struct ib_cq *cq = ssk->rx_ring.cq; struct ib_wc ibwc[SDP_NUM_WC]; int n, i; int wc_processed = 0; struct mbuf *mb; do { n = ib_poll_cq(cq, SDP_NUM_WC, ibwc); for (i = 0; i < n; ++i) { struct ib_wc *wc = &ibwc[i]; BUG_ON(!(wc->wr_id & SDP_OP_RECV)); mb = sdp_process_rx_wc(ssk, wc); if (!mb) continue; sdp_process_rx_mb(ssk, mb); wc_processed++; } } while (n == SDP_NUM_WC); if (wc_processed) sdp_bzcopy_write_space(ssk); return wc_processed; } static void sdp_rx_comp_work(struct work_struct *work) { struct sdp_sock *ssk = container_of(work, struct sdp_sock, rx_comp_work); sdp_prf(ssk->socket, NULL, "%s", __func__); SDP_WLOCK(ssk); if (unlikely(!ssk->qp)) { sdp_prf(ssk->socket, NULL, "qp was destroyed"); goto out; } if (unlikely(!ssk->rx_ring.cq)) { sdp_prf(ssk->socket, NULL, "rx_ring.cq is NULL"); goto out; } if (unlikely(!ssk->poll_cq)) { struct rdma_cm_id *id = ssk->id; if (id && id->qp) rdma_notify(id, IB_EVENT_COMM_EST); goto out; } sdp_do_posts(ssk); out: SDP_WUNLOCK(ssk); } void sdp_do_posts(struct sdp_sock *ssk) { struct socket *sk = ssk->socket; int xmit_poll_force; struct mbuf *mb; SDP_WLOCK_ASSERT(ssk); if (!ssk->qp_active) { sdp_dbg(sk, "QP is deactivated\n"); return; } while ((mb = ssk->rx_ctl_q)) { ssk->rx_ctl_q = mb->m_nextpkt; mb->m_nextpkt = NULL; sdp_process_rx_ctl_mb(ssk, mb); } if (ssk->state == TCPS_TIME_WAIT) return; if (!ssk->rx_ring.cq || !ssk->tx_ring.cq) return; sdp_post_recvs(ssk); if (tx_ring_posted(ssk)) sdp_xmit_poll(ssk, 1); sdp_post_sends(ssk, M_NOWAIT); xmit_poll_force = tx_credits(ssk) < SDP_MIN_TX_CREDITS; if (credit_update_needed(ssk) || xmit_poll_force) { /* if has pending tx because run out of tx_credits - xmit it */ sdp_prf(sk, NULL, "Processing to free pending sends"); sdp_xmit_poll(ssk, xmit_poll_force); sdp_prf(sk, NULL, "Sending credit update"); sdp_post_sends(ssk, M_NOWAIT); } } int sdp_process_rx(struct sdp_sock *ssk) { int wc_processed = 0; int credits_before; if (!rx_ring_trylock(&ssk->rx_ring)) { sdp_dbg(ssk->socket, "ring destroyed. not polling it\n"); return 0; } credits_before = tx_credits(ssk); wc_processed = sdp_poll_rx_cq(ssk); sdp_prf(ssk->socket, NULL, "processed %d", wc_processed); if (wc_processed) { sdp_prf(ssk->socket, NULL, "credits: %d -> %d", credits_before, tx_credits(ssk)); queue_work(rx_comp_wq, &ssk->rx_comp_work); } sdp_arm_rx_cq(ssk); rx_ring_unlock(&ssk->rx_ring); return (wc_processed); } static void sdp_rx_irq(struct ib_cq *cq, void *cq_context) { struct socket *sk = cq_context; struct sdp_sock *ssk = sdp_sk(sk); if (cq != ssk->rx_ring.cq) { sdp_dbg(sk, "cq = %p, ssk->cq = %p\n", cq, ssk->rx_ring.cq); return; } SDPSTATS_COUNTER_INC(rx_int_count); sdp_prf(sk, NULL, "rx irq"); sdp_process_rx(ssk); } static void sdp_rx_ring_purge(struct sdp_sock *ssk) { while (rx_ring_posted(ssk) > 0) { struct mbuf *mb; mb = sdp_recv_completion(ssk, ring_tail(ssk->rx_ring)); if (!mb) break; m_freem(mb); } } void sdp_rx_ring_init(struct sdp_sock *ssk) { ssk->rx_ring.buffer = NULL; ssk->rx_ring.destroyed = 0; rw_init(&ssk->rx_ring.destroyed_lock, "sdp rx lock"); } static void sdp_rx_cq_event_handler(struct ib_event *event, void *data) { } int sdp_rx_ring_create(struct sdp_sock *ssk, struct ib_device *device) { struct ib_cq *rx_cq; int rc = 0; sdp_dbg(ssk->socket, "rx ring created"); INIT_WORK(&ssk->rx_comp_work, sdp_rx_comp_work); atomic_set(&ssk->rx_ring.head, 1); atomic_set(&ssk->rx_ring.tail, 1); ssk->rx_ring.buffer = kmalloc( sizeof *ssk->rx_ring.buffer * SDP_RX_SIZE, GFP_KERNEL); if (!ssk->rx_ring.buffer) { sdp_warn(ssk->socket, "Unable to allocate RX Ring size %zd.\n", sizeof(*ssk->rx_ring.buffer) * SDP_RX_SIZE); return -ENOMEM; } rx_cq = ib_create_cq(device, sdp_rx_irq, sdp_rx_cq_event_handler, ssk->socket, SDP_RX_SIZE, 0); if (IS_ERR(rx_cq)) { rc = PTR_ERR(rx_cq); sdp_warn(ssk->socket, "Unable to allocate RX CQ: %d.\n", rc); goto err_cq; } sdp_sk(ssk->socket)->rx_ring.cq = rx_cq; sdp_arm_rx_cq(ssk); return 0; err_cq: kfree(ssk->rx_ring.buffer); ssk->rx_ring.buffer = NULL; return rc; } void sdp_rx_ring_destroy(struct sdp_sock *ssk) { cancel_work_sync(&ssk->rx_comp_work); rx_ring_destroy_lock(&ssk->rx_ring); if (ssk->rx_ring.buffer) { sdp_rx_ring_purge(ssk); kfree(ssk->rx_ring.buffer); ssk->rx_ring.buffer = NULL; } if (ssk->rx_ring.cq) { if (ib_destroy_cq(ssk->rx_ring.cq)) { sdp_warn(ssk->socket, "destroy cq(%p) failed\n", ssk->rx_ring.cq); } else { ssk->rx_ring.cq = NULL; } } WARN_ON(ring_head(ssk->rx_ring) != ring_tail(ssk->rx_ring)); }