2 * Copyright (c) 2006 Oracle. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
35 #include <linux/device.h>
36 #include <linux/dmapool.h>
37 #include <linux/ratelimit.h>
42 static void rds_iw_send_rdma_complete(struct rds_message *rm,
48 case IB_WC_WR_FLUSH_ERR:
52 notify_status = RDS_RDMA_SUCCESS;
55 case IB_WC_REM_ACCESS_ERR:
56 notify_status = RDS_RDMA_REMOTE_ERROR;
60 notify_status = RDS_RDMA_OTHER_ERROR;
63 rds_rdma_send_complete(rm, notify_status);
66 static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic,
67 struct rm_rdma_op *op)
70 ib_dma_unmap_sg(ic->i_cm_id->device,
71 op->op_sg, op->op_nents,
72 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
77 static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
78 struct rds_iw_send_work *send,
81 struct rds_message *rm = send->s_rm;
83 rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
85 ib_dma_unmap_sg(ic->i_cm_id->device,
86 rm->data.op_sg, rm->data.op_nents,
89 if (rm->rdma.op_active) {
90 rds_iw_send_unmap_rdma(ic, &rm->rdma);
92 /* If the user asked for a completion notification on this
93 * message, we can implement three different semantics:
94 * 1. Notify when we received the ACK on the RDS message
95 * that was queued with the RDMA. This provides reliable
96 * notification of RDMA status at the expense of a one-way
98 * 2. Notify when the IB stack gives us the completion event for
100 * 3. Notify when the IB stack gives us the completion event for
101 * the accompanying RDS messages.
102 * Here, we implement approach #3. To implement approach #2,
103 * call rds_rdma_send_complete from the cq_handler. To implement #1,
104 * don't call rds_rdma_send_complete at all, and fall back to the notify
105 * handling in the ACK processing code.
107 * Note: There's no need to explicitly sync any RDMA buffers using
108 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
109 * operation itself unmapped the RDMA buffers, which takes care
112 rds_iw_send_rdma_complete(rm, wc_status);
114 if (rm->rdma.op_write)
115 rds_stats_add(s_send_rdma_bytes, rm->rdma.op_bytes);
117 rds_stats_add(s_recv_rdma_bytes, rm->rdma.op_bytes);
120 /* If anyone waited for this message to get flushed out, wake
122 rds_message_unmapped(rm);
128 void rds_iw_send_init_ring(struct rds_iw_connection *ic)
130 struct rds_iw_send_work *send;
133 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
138 send->s_mapping = NULL;
140 send->s_wr.next = NULL;
141 send->s_wr.wr_id = i;
142 send->s_wr.sg_list = send->s_sge;
143 send->s_wr.num_sge = 1;
144 send->s_wr.opcode = IB_WR_SEND;
145 send->s_wr.send_flags = 0;
146 send->s_wr.ex.imm_data = 0;
148 sge = rds_iw_data_sge(ic, send->s_sge);
151 sge = rds_iw_header_sge(ic, send->s_sge);
152 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
153 sge->length = sizeof(struct rds_header);
156 send->s_mr = ib_alloc_fast_reg_mr(ic->i_pd, fastreg_message_size);
157 if (IS_ERR(send->s_mr)) {
158 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed\n");
162 send->s_page_list = ib_alloc_fast_reg_page_list(
163 ic->i_cm_id->device, fastreg_message_size);
164 if (IS_ERR(send->s_page_list)) {
165 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
171 void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
173 struct rds_iw_send_work *send;
176 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
178 ib_dereg_mr(send->s_mr);
179 BUG_ON(!send->s_page_list);
180 ib_free_fast_reg_page_list(send->s_page_list);
181 if (send->s_wr.opcode == 0xdead)
184 rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
186 rds_iw_send_unmap_rdma(ic, send->s_op);
191 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
192 * operations performed in the send path. As the sender allocs and potentially
193 * unallocs the next free entry in the ring it doesn't alter which is
194 * the next to be freed, which is what this is concerned with.
196 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
198 struct rds_connection *conn = context;
199 struct rds_iw_connection *ic = conn->c_transport_data;
201 struct rds_iw_send_work *send;
207 rdsdebug("cq %p conn %p\n", cq, conn);
208 rds_iw_stats_inc(s_iw_tx_cq_call);
209 ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
211 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
213 while (ib_poll_cq(cq, 1, &wc) > 0) {
214 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
215 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
216 be32_to_cpu(wc.ex.imm_data));
217 rds_iw_stats_inc(s_iw_tx_cq_event);
219 if (wc.status != IB_WC_SUCCESS) {
220 printk(KERN_ERR "WC Error: status = %d opcode = %d\n", wc.status, wc.opcode);
224 if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
225 ic->i_fastreg_posted = 0;
229 if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
230 ic->i_fastreg_posted = 1;
234 if (wc.wr_id == RDS_IW_ACK_WR_ID) {
235 if (ic->i_ack_queued + HZ/2 < jiffies)
236 rds_iw_stats_inc(s_iw_tx_stalled);
237 rds_iw_ack_send_complete(ic);
241 oldest = rds_iw_ring_oldest(&ic->i_send_ring);
243 completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
245 for (i = 0; i < completed; i++) {
246 send = &ic->i_sends[oldest];
248 /* In the error case, wc.opcode sometimes contains garbage */
249 switch (send->s_wr.opcode) {
252 rds_iw_send_unmap_rm(ic, send, wc.status);
254 case IB_WR_FAST_REG_MR:
255 case IB_WR_RDMA_WRITE:
256 case IB_WR_RDMA_READ:
257 case IB_WR_RDMA_READ_WITH_INV:
258 /* Nothing to be done - the SG list will be unmapped
259 * when the SEND completes. */
262 printk_ratelimited(KERN_NOTICE
263 "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
264 __func__, send->s_wr.opcode);
268 send->s_wr.opcode = 0xdead;
269 send->s_wr.num_sge = 1;
270 if (send->s_queued + HZ/2 < jiffies)
271 rds_iw_stats_inc(s_iw_tx_stalled);
273 /* If a RDMA operation produced an error, signal this right
274 * away. If we don't, the subsequent SEND that goes with this
275 * RDMA will be canceled with ERR_WFLUSH, and the application
276 * never learn that the RDMA failed. */
277 if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
278 struct rds_message *rm;
280 rm = rds_send_get_message(conn, send->s_op);
282 rds_iw_send_rdma_complete(rm, wc.status);
285 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
288 rds_iw_ring_free(&ic->i_send_ring, completed);
290 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
291 test_bit(0, &conn->c_map_queued))
292 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
294 /* We expect errors as the qp is drained during shutdown */
295 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
296 rds_iw_conn_error(conn,
297 "send completion on %pI4 "
298 "had status %u, disconnecting and reconnecting\n",
299 &conn->c_faddr, wc.status);
305 * This is the main function for allocating credits when sending
308 * Conceptually, we have two counters:
309 * - send credits: this tells us how many WRs we're allowed
310 * to submit without overruning the receiver's queue. For
311 * each SEND WR we post, we decrement this by one.
313 * - posted credits: this tells us how many WRs we recently
314 * posted to the receive queue. This value is transferred
315 * to the peer as a "credit update" in a RDS header field.
316 * Every time we transmit credits to the peer, we subtract
317 * the amount of transferred credits from this counter.
319 * It is essential that we avoid situations where both sides have
320 * exhausted their send credits, and are unable to send new credits
321 * to the peer. We achieve this by requiring that we send at least
322 * one credit update to the peer before exhausting our credits.
323 * When new credits arrive, we subtract one credit that is withheld
324 * until we've posted new buffers and are ready to transmit these
325 * credits (see rds_iw_send_add_credits below).
327 * The RDS send code is essentially single-threaded; rds_send_xmit
328 * grabs c_send_lock to ensure exclusive access to the send ring.
329 * However, the ACK sending code is independent and can race with
332 * In the send path, we need to update the counters for send credits
333 * and the counter of posted buffers atomically - when we use the
334 * last available credit, we cannot allow another thread to race us
335 * and grab the posted credits counter. Hence, we have to use a
336 * spinlock to protect the credit counter, or use atomics.
338 * Spinlocks shared between the send and the receive path are bad,
339 * because they create unnecessary delays. An early implementation
340 * using a spinlock showed a 5% degradation in throughput at some
343 * This implementation avoids spinlocks completely, putting both
344 * counters into a single atomic, and updating that atomic using
345 * atomic_add (in the receive path, when receiving fresh credits),
346 * and using atomic_cmpxchg when updating the two counters.
348 int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
349 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
351 unsigned int avail, posted, got = 0, advertise;
360 oldval = newval = atomic_read(&ic->i_credits);
361 posted = IB_GET_POST_CREDITS(oldval);
362 avail = IB_GET_SEND_CREDITS(oldval);
364 rdsdebug("rds_iw_send_grab_credits(%u): credits=%u posted=%u\n",
365 wanted, avail, posted);
367 /* The last credit must be used to send a credit update. */
368 if (avail && !posted)
371 if (avail < wanted) {
372 struct rds_connection *conn = ic->i_cm_id->context;
374 /* Oops, there aren't that many credits left! */
375 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
378 /* Sometimes you get what you want, lalala. */
381 newval -= IB_SET_SEND_CREDITS(got);
384 * If need_posted is non-zero, then the caller wants
385 * the posted regardless of whether any send credits are
388 if (posted && (got || need_posted)) {
389 advertise = min_t(unsigned int, posted, max_posted);
390 newval -= IB_SET_POST_CREDITS(advertise);
393 /* Finally bill everything */
394 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
397 *adv_credits = advertise;
401 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
403 struct rds_iw_connection *ic = conn->c_transport_data;
408 rdsdebug("rds_iw_send_add_credits(%u): current=%u%s\n",
410 IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
411 test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
413 atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
414 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
415 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
417 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
419 rds_iw_stats_inc(s_iw_rx_credit_updates);
422 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
424 struct rds_iw_connection *ic = conn->c_transport_data;
429 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
431 /* Decide whether to send an update to the peer now.
432 * If we would send a credit update for every single buffer we
433 * post, we would end up with an ACK storm (ACK arrives,
434 * consumes buffer, we refill the ring, send ACK to remote
435 * advertising the newly posted buffer... ad inf)
437 * Performance pretty much depends on how often we send
438 * credit updates - too frequent updates mean lots of ACKs.
439 * Too infrequent updates, and the peer will run out of
440 * credits and has to throttle.
441 * For the time being, 16 seems to be a good compromise.
443 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
444 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
448 rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
449 struct rds_iw_send_work *send, unsigned int pos,
450 unsigned long buffer, unsigned int length,
455 WARN_ON(pos != send - ic->i_sends);
457 send->s_wr.send_flags = send_flags;
458 send->s_wr.opcode = IB_WR_SEND;
459 send->s_wr.num_sge = 2;
460 send->s_wr.next = NULL;
461 send->s_queued = jiffies;
465 sge = rds_iw_data_sge(ic, send->s_sge);
467 sge->length = length;
468 sge->lkey = rds_iw_local_dma_lkey(ic);
470 sge = rds_iw_header_sge(ic, send->s_sge);
472 /* We're sending a packet with no payload. There is only
474 send->s_wr.num_sge = 1;
475 sge = &send->s_sge[0];
478 sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
479 sge->length = sizeof(struct rds_header);
480 sge->lkey = rds_iw_local_dma_lkey(ic);
484 * This can be called multiple times for a given message. The first time
485 * we see a message we map its scatterlist into the IB device so that
486 * we can provide that mapped address to the IB scatter gather entries
487 * in the IB work requests. We translate the scatterlist into a series
488 * of work requests that fragment the message. These work requests complete
489 * in order so we pass ownership of the message to the completion handler
490 * once we send the final fragment.
492 * The RDS core uses the c_send_lock to only enter this function once
493 * per connection. This makes sure that the tx ring alloc/unalloc pairs
494 * don't get out of sync and confuse the ring.
496 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
497 unsigned int hdr_off, unsigned int sg, unsigned int off)
499 struct rds_iw_connection *ic = conn->c_transport_data;
500 struct ib_device *dev = ic->i_cm_id->device;
501 struct rds_iw_send_work *send = NULL;
502 struct rds_iw_send_work *first;
503 struct rds_iw_send_work *prev;
504 struct ib_send_wr *failed_wr;
505 struct scatterlist *scat;
515 int flow_controlled = 0;
517 BUG_ON(off % RDS_FRAG_SIZE);
518 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
520 /* Fastreg support */
521 if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
526 /* FIXME we may overallocate here */
527 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
530 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
532 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
533 if (work_alloc == 0) {
534 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
535 rds_iw_stats_inc(s_iw_tx_ring_full);
540 credit_alloc = work_alloc;
542 credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
543 adv_credits += posted;
544 if (credit_alloc < work_alloc) {
545 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
546 work_alloc = credit_alloc;
549 if (work_alloc == 0) {
550 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
551 rds_iw_stats_inc(s_iw_tx_throttle);
557 /* map the message the first time we see it */
560 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
561 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
562 rm->m_inc.i_hdr.h_flags,
563 be32_to_cpu(rm->m_inc.i_hdr.h_len));
565 if (rm->data.op_nents) {
566 rm->data.op_count = ib_dma_map_sg(dev,
570 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
571 if (rm->data.op_count == 0) {
572 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
573 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
574 ret = -ENOMEM; /* XXX ? */
578 rm->data.op_count = 0;
581 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
582 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
583 rds_message_addref(rm);
586 /* Finalize the header */
587 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
588 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
589 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
590 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
592 /* If it has a RDMA op, tell the peer we did it. This is
593 * used by the peer to release use-once RDMA MRs. */
594 if (rm->rdma.op_active) {
595 struct rds_ext_header_rdma ext_hdr;
597 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
598 rds_message_add_extension(&rm->m_inc.i_hdr,
599 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
601 if (rm->m_rdma_cookie) {
602 rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
603 rds_rdma_cookie_key(rm->m_rdma_cookie),
604 rds_rdma_cookie_offset(rm->m_rdma_cookie));
607 /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
608 * we should not do this unless we have a chance of at least
609 * sticking the header into the send ring. Which is why we
610 * should call rds_iw_ring_alloc first. */
611 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
612 rds_message_make_checksum(&rm->m_inc.i_hdr);
615 * Update adv_credits since we reset the ACK_REQUIRED bit.
617 rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
618 adv_credits += posted;
619 BUG_ON(adv_credits > 255);
622 send = &ic->i_sends[pos];
625 scat = &rm->data.op_sg[sg];
629 /* Sometimes you want to put a fence between an RDMA
630 * READ and the following SEND.
631 * We could either do this all the time
632 * or when requested by the user. Right now, we let
633 * the application choose.
635 if (rm->rdma.op_active && rm->rdma.op_fence)
636 send_flags = IB_SEND_FENCE;
639 * We could be copying the header into the unused tail of the page.
640 * That would need to be changed in the future when those pages might
641 * be mapped userspace pages or page cache pages. So instead we always
642 * use a second sge and our long-lived ring of mapped headers. We send
643 * the header after the data so that the data payload can be aligned on
647 /* handle a 0-len message */
648 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
649 rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
653 /* if there's data reference it with a chain of work reqs */
654 for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
657 send = &ic->i_sends[pos];
659 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
660 rds_iw_xmit_populate_wr(ic, send, pos,
661 ib_sg_dma_address(dev, scat) + off, len,
665 * We want to delay signaling completions just enough to get
666 * the batching benefits but not so much that we create dead time
669 if (ic->i_unsignaled_wrs-- == 0) {
670 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
671 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
674 ic->i_unsignaled_bytes -= len;
675 if (ic->i_unsignaled_bytes <= 0) {
676 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
677 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
681 * Always signal the last one if we're stopping due to flow control.
683 if (flow_controlled && i == (work_alloc-1))
684 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
686 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
687 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
691 if (off == ib_sg_dma_len(dev, scat)) {
697 /* Tack on the header after the data. The header SGE should already
698 * have been set up to point to the right header buffer. */
699 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
702 struct rds_header *hdr = &ic->i_send_hdrs[pos];
704 printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
705 be16_to_cpu(hdr->h_dport),
707 be32_to_cpu(hdr->h_len));
710 struct rds_header *hdr = &ic->i_send_hdrs[pos];
712 /* add credit and redo the header checksum */
713 hdr->h_credit = adv_credits;
714 rds_message_make_checksum(hdr);
716 rds_iw_stats_inc(s_iw_tx_credit_updates);
720 prev->s_wr.next = &send->s_wr;
723 pos = (pos + 1) % ic->i_send_ring.w_nr;
726 /* Account the RDS header in the number of bytes we sent, but just once.
727 * The caller has no concept of fragmentation. */
729 sent += sizeof(struct rds_header);
731 /* if we finished the message then send completion owns it */
732 if (scat == &rm->data.op_sg[rm->data.op_count]) {
733 prev->s_rm = ic->i_rm;
734 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
738 if (i < work_alloc) {
739 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
742 if (ic->i_flowctl && i < credit_alloc)
743 rds_iw_send_add_credits(conn, credit_alloc - i);
745 /* XXX need to worry about failed_wr and partial sends. */
746 failed_wr = &first->s_wr;
747 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
748 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
749 first, &first->s_wr, ret, failed_wr);
750 BUG_ON(failed_wr != &first->s_wr);
752 printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
753 "returned %d\n", &conn->c_faddr, ret);
754 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
756 ic->i_rm = prev->s_rm;
768 static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rds_iw_connection *ic, struct rds_iw_send_work *send, int nent, int len, u64 sg_addr)
770 BUG_ON(nent > send->s_page_list->max_page_list_len);
772 * Perform a WR for the fast_reg_mr. Each individual page
773 * in the sg list is added to the fast reg page list and placed
774 * inside the fast_reg_mr WR.
776 send->s_wr.opcode = IB_WR_FAST_REG_MR;
777 send->s_wr.wr.fast_reg.length = len;
778 send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
779 send->s_wr.wr.fast_reg.page_list = send->s_page_list;
780 send->s_wr.wr.fast_reg.page_list_len = nent;
781 send->s_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
782 send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
783 send->s_wr.wr.fast_reg.iova_start = sg_addr;
785 ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
788 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
790 struct rds_iw_connection *ic = conn->c_transport_data;
791 struct rds_iw_send_work *send = NULL;
792 struct rds_iw_send_work *first;
793 struct rds_iw_send_work *prev;
794 struct ib_send_wr *failed_wr;
795 struct rds_iw_device *rds_iwdev;
796 struct scatterlist *scat;
798 u64 remote_addr = op->op_remote_addr;
807 rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
809 /* map the message the first time we see it */
810 if (!op->op_mapped) {
811 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
812 op->op_sg, op->op_nents, (op->op_write) ?
813 DMA_TO_DEVICE : DMA_FROM_DEVICE);
814 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
815 if (op->op_count == 0) {
816 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
817 ret = -ENOMEM; /* XXX ? */
825 /* Alloc space on the send queue for the fastreg */
826 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
827 if (work_alloc != 1) {
828 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
829 rds_iw_stats_inc(s_iw_tx_ring_full);
836 * Instead of knowing how to return a partial rdma read/write we insist that there
837 * be enough work requests to send the entire message.
839 i = ceil(op->op_count, rds_iwdev->max_sge);
841 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
842 if (work_alloc != i) {
843 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
844 rds_iw_stats_inc(s_iw_tx_ring_full);
849 send = &ic->i_sends[pos];
851 first = prev = &ic->i_sends[fr_pos];
856 scat = &op->op_sg[0];
858 num_sge = op->op_count;
860 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
861 send->s_wr.send_flags = 0;
862 send->s_queued = jiffies;
865 * We want to delay signaling completions just enough to get
866 * the batching benefits but not so much that we create dead time on the wire.
868 if (ic->i_unsignaled_wrs-- == 0) {
869 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
870 send->s_wr.send_flags = IB_SEND_SIGNALED;
873 /* To avoid the need to have the plumbing to invalidate the fastreg_mr used
874 * for local access after RDS is finished with it, using
875 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
878 send->s_wr.opcode = IB_WR_RDMA_WRITE;
880 send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
882 send->s_wr.wr.rdma.remote_addr = remote_addr;
883 send->s_wr.wr.rdma.rkey = op->op_rkey;
886 if (num_sge > rds_iwdev->max_sge) {
887 send->s_wr.num_sge = rds_iwdev->max_sge;
888 num_sge -= rds_iwdev->max_sge;
890 send->s_wr.num_sge = num_sge;
892 send->s_wr.next = NULL;
895 prev->s_wr.next = &send->s_wr;
897 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
898 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
900 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
901 send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
903 send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
904 send->s_sge[j].length = len;
905 send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
909 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
915 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
916 send->s_wr.num_sge = 1;
917 send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
918 send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
919 send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
922 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
923 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
926 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
930 /* if we finished the message then send completion owns it */
931 if (scat == &op->op_sg[op->op_count])
932 first->s_wr.send_flags = IB_SEND_SIGNALED;
934 if (i < work_alloc) {
935 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
939 /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
940 * recommended. Putting the lkey on the wire is a security hole, as it can
941 * allow for memory access to all of memory on the remote system. Some
942 * adapters do not allow using the lkey for this at all. To bypass this use a
943 * fastreg_mr (or possibly a dma_mr)
946 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
947 op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
951 failed_wr = &first->s_wr;
952 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
953 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
954 first, &first->s_wr, ret, failed_wr);
955 BUG_ON(failed_wr != &first->s_wr);
957 printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
958 "returned %d\n", &conn->c_faddr, ret);
959 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
967 void rds_iw_xmit_complete(struct rds_connection *conn)
969 struct rds_iw_connection *ic = conn->c_transport_data;
971 /* We may have a pending ACK or window update we were unable
972 * to send previously (due to flow control). Try again. */
973 rds_iw_attempt_ack(ic);