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_mr(ic->i_pd, IB_MR_TYPE_MEM_REG,
157 fastreg_message_size);
158 if (IS_ERR(send->s_mr)) {
159 printk(KERN_WARNING "RDS/IW: ib_alloc_mr failed\n");
163 send->s_page_list = ib_alloc_fast_reg_page_list(
164 ic->i_cm_id->device, fastreg_message_size);
165 if (IS_ERR(send->s_page_list)) {
166 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
172 void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
174 struct rds_iw_send_work *send;
177 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
179 ib_dereg_mr(send->s_mr);
180 BUG_ON(!send->s_page_list);
181 ib_free_fast_reg_page_list(send->s_page_list);
182 if (send->s_wr.opcode == 0xdead)
185 rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
187 rds_iw_send_unmap_rdma(ic, send->s_op);
192 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
193 * operations performed in the send path. As the sender allocs and potentially
194 * unallocs the next free entry in the ring it doesn't alter which is
195 * the next to be freed, which is what this is concerned with.
197 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
199 struct rds_connection *conn = context;
200 struct rds_iw_connection *ic = conn->c_transport_data;
202 struct rds_iw_send_work *send;
208 rdsdebug("cq %p conn %p\n", cq, conn);
209 rds_iw_stats_inc(s_iw_tx_cq_call);
210 ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
212 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
214 while (ib_poll_cq(cq, 1, &wc) > 0) {
215 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
216 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
217 be32_to_cpu(wc.ex.imm_data));
218 rds_iw_stats_inc(s_iw_tx_cq_event);
220 if (wc.status != IB_WC_SUCCESS) {
221 printk(KERN_ERR "WC Error: status = %d opcode = %d\n", wc.status, wc.opcode);
225 if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
226 ic->i_fastreg_posted = 0;
230 if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
231 ic->i_fastreg_posted = 1;
235 if (wc.wr_id == RDS_IW_ACK_WR_ID) {
236 if (time_after(jiffies, ic->i_ack_queued + HZ/2))
237 rds_iw_stats_inc(s_iw_tx_stalled);
238 rds_iw_ack_send_complete(ic);
242 oldest = rds_iw_ring_oldest(&ic->i_send_ring);
244 completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
246 for (i = 0; i < completed; i++) {
247 send = &ic->i_sends[oldest];
249 /* In the error case, wc.opcode sometimes contains garbage */
250 switch (send->s_wr.opcode) {
253 rds_iw_send_unmap_rm(ic, send, wc.status);
255 case IB_WR_FAST_REG_MR:
256 case IB_WR_RDMA_WRITE:
257 case IB_WR_RDMA_READ:
258 case IB_WR_RDMA_READ_WITH_INV:
259 /* Nothing to be done - the SG list will be unmapped
260 * when the SEND completes. */
263 printk_ratelimited(KERN_NOTICE
264 "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
265 __func__, send->s_wr.opcode);
269 send->s_wr.opcode = 0xdead;
270 send->s_wr.num_sge = 1;
271 if (time_after(jiffies, send->s_queued + HZ/2))
272 rds_iw_stats_inc(s_iw_tx_stalled);
274 /* If a RDMA operation produced an error, signal this right
275 * away. If we don't, the subsequent SEND that goes with this
276 * RDMA will be canceled with ERR_WFLUSH, and the application
277 * never learn that the RDMA failed. */
278 if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
279 struct rds_message *rm;
281 rm = rds_send_get_message(conn, send->s_op);
283 rds_iw_send_rdma_complete(rm, wc.status);
286 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
289 rds_iw_ring_free(&ic->i_send_ring, completed);
291 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
292 test_bit(0, &conn->c_map_queued))
293 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
295 /* We expect errors as the qp is drained during shutdown */
296 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
297 rds_iw_conn_error(conn,
298 "send completion on %pI4 "
299 "had status %u, disconnecting and reconnecting\n",
300 &conn->c_faddr, wc.status);
306 * This is the main function for allocating credits when sending
309 * Conceptually, we have two counters:
310 * - send credits: this tells us how many WRs we're allowed
311 * to submit without overruning the receiver's queue. For
312 * each SEND WR we post, we decrement this by one.
314 * - posted credits: this tells us how many WRs we recently
315 * posted to the receive queue. This value is transferred
316 * to the peer as a "credit update" in a RDS header field.
317 * Every time we transmit credits to the peer, we subtract
318 * the amount of transferred credits from this counter.
320 * It is essential that we avoid situations where both sides have
321 * exhausted their send credits, and are unable to send new credits
322 * to the peer. We achieve this by requiring that we send at least
323 * one credit update to the peer before exhausting our credits.
324 * When new credits arrive, we subtract one credit that is withheld
325 * until we've posted new buffers and are ready to transmit these
326 * credits (see rds_iw_send_add_credits below).
328 * The RDS send code is essentially single-threaded; rds_send_xmit
329 * grabs c_send_lock to ensure exclusive access to the send ring.
330 * However, the ACK sending code is independent and can race with
333 * In the send path, we need to update the counters for send credits
334 * and the counter of posted buffers atomically - when we use the
335 * last available credit, we cannot allow another thread to race us
336 * and grab the posted credits counter. Hence, we have to use a
337 * spinlock to protect the credit counter, or use atomics.
339 * Spinlocks shared between the send and the receive path are bad,
340 * because they create unnecessary delays. An early implementation
341 * using a spinlock showed a 5% degradation in throughput at some
344 * This implementation avoids spinlocks completely, putting both
345 * counters into a single atomic, and updating that atomic using
346 * atomic_add (in the receive path, when receiving fresh credits),
347 * and using atomic_cmpxchg when updating the two counters.
349 int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
350 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
352 unsigned int avail, posted, got = 0, advertise;
361 oldval = newval = atomic_read(&ic->i_credits);
362 posted = IB_GET_POST_CREDITS(oldval);
363 avail = IB_GET_SEND_CREDITS(oldval);
365 rdsdebug("wanted=%u credits=%u posted=%u\n",
366 wanted, avail, posted);
368 /* The last credit must be used to send a credit update. */
369 if (avail && !posted)
372 if (avail < wanted) {
373 struct rds_connection *conn = ic->i_cm_id->context;
375 /* Oops, there aren't that many credits left! */
376 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
379 /* Sometimes you get what you want, lalala. */
382 newval -= IB_SET_SEND_CREDITS(got);
385 * If need_posted is non-zero, then the caller wants
386 * the posted regardless of whether any send credits are
389 if (posted && (got || need_posted)) {
390 advertise = min_t(unsigned int, posted, max_posted);
391 newval -= IB_SET_POST_CREDITS(advertise);
394 /* Finally bill everything */
395 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
398 *adv_credits = advertise;
402 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
404 struct rds_iw_connection *ic = conn->c_transport_data;
409 rdsdebug("credits=%u current=%u%s\n",
411 IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
412 test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
414 atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
415 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
416 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
418 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
420 rds_iw_stats_inc(s_iw_rx_credit_updates);
423 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
425 struct rds_iw_connection *ic = conn->c_transport_data;
430 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
432 /* Decide whether to send an update to the peer now.
433 * If we would send a credit update for every single buffer we
434 * post, we would end up with an ACK storm (ACK arrives,
435 * consumes buffer, we refill the ring, send ACK to remote
436 * advertising the newly posted buffer... ad inf)
438 * Performance pretty much depends on how often we send
439 * credit updates - too frequent updates mean lots of ACKs.
440 * Too infrequent updates, and the peer will run out of
441 * credits and has to throttle.
442 * For the time being, 16 seems to be a good compromise.
444 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
445 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
449 rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
450 struct rds_iw_send_work *send, unsigned int pos,
451 unsigned long buffer, unsigned int length,
456 WARN_ON(pos != send - ic->i_sends);
458 send->s_wr.send_flags = send_flags;
459 send->s_wr.opcode = IB_WR_SEND;
460 send->s_wr.num_sge = 2;
461 send->s_wr.next = NULL;
462 send->s_queued = jiffies;
466 sge = rds_iw_data_sge(ic, send->s_sge);
468 sge->length = length;
469 sge->lkey = rds_iw_local_dma_lkey(ic);
471 sge = rds_iw_header_sge(ic, send->s_sge);
473 /* We're sending a packet with no payload. There is only
475 send->s_wr.num_sge = 1;
476 sge = &send->s_sge[0];
479 sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
480 sge->length = sizeof(struct rds_header);
481 sge->lkey = rds_iw_local_dma_lkey(ic);
485 * This can be called multiple times for a given message. The first time
486 * we see a message we map its scatterlist into the IB device so that
487 * we can provide that mapped address to the IB scatter gather entries
488 * in the IB work requests. We translate the scatterlist into a series
489 * of work requests that fragment the message. These work requests complete
490 * in order so we pass ownership of the message to the completion handler
491 * once we send the final fragment.
493 * The RDS core uses the c_send_lock to only enter this function once
494 * per connection. This makes sure that the tx ring alloc/unalloc pairs
495 * don't get out of sync and confuse the ring.
497 int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
498 unsigned int hdr_off, unsigned int sg, unsigned int off)
500 struct rds_iw_connection *ic = conn->c_transport_data;
501 struct ib_device *dev = ic->i_cm_id->device;
502 struct rds_iw_send_work *send = NULL;
503 struct rds_iw_send_work *first;
504 struct rds_iw_send_work *prev;
505 struct ib_send_wr *failed_wr;
506 struct scatterlist *scat;
516 int flow_controlled = 0;
518 BUG_ON(off % RDS_FRAG_SIZE);
519 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
521 /* Fastreg support */
522 if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
527 /* FIXME we may overallocate here */
528 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
531 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
533 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
534 if (work_alloc == 0) {
535 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
536 rds_iw_stats_inc(s_iw_tx_ring_full);
541 credit_alloc = work_alloc;
543 credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
544 adv_credits += posted;
545 if (credit_alloc < work_alloc) {
546 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
547 work_alloc = credit_alloc;
550 if (work_alloc == 0) {
551 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
552 rds_iw_stats_inc(s_iw_tx_throttle);
558 /* map the message the first time we see it */
561 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
562 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
563 rm->m_inc.i_hdr.h_flags,
564 be32_to_cpu(rm->m_inc.i_hdr.h_len));
566 if (rm->data.op_nents) {
567 rm->data.op_count = ib_dma_map_sg(dev,
571 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
572 if (rm->data.op_count == 0) {
573 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
574 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
575 ret = -ENOMEM; /* XXX ? */
579 rm->data.op_count = 0;
582 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
583 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
584 rds_message_addref(rm);
585 rm->data.op_dmasg = 0;
586 rm->data.op_dmaoff = 0;
589 /* Finalize the header */
590 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
591 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
592 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
593 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
595 /* If it has a RDMA op, tell the peer we did it. This is
596 * used by the peer to release use-once RDMA MRs. */
597 if (rm->rdma.op_active) {
598 struct rds_ext_header_rdma ext_hdr;
600 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
601 rds_message_add_extension(&rm->m_inc.i_hdr,
602 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
604 if (rm->m_rdma_cookie) {
605 rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
606 rds_rdma_cookie_key(rm->m_rdma_cookie),
607 rds_rdma_cookie_offset(rm->m_rdma_cookie));
610 /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
611 * we should not do this unless we have a chance of at least
612 * sticking the header into the send ring. Which is why we
613 * should call rds_iw_ring_alloc first. */
614 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
615 rds_message_make_checksum(&rm->m_inc.i_hdr);
618 * Update adv_credits since we reset the ACK_REQUIRED bit.
620 rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
621 adv_credits += posted;
622 BUG_ON(adv_credits > 255);
625 send = &ic->i_sends[pos];
628 scat = &rm->data.op_sg[rm->data.op_dmasg];
632 /* Sometimes you want to put a fence between an RDMA
633 * READ and the following SEND.
634 * We could either do this all the time
635 * or when requested by the user. Right now, we let
636 * the application choose.
638 if (rm->rdma.op_active && rm->rdma.op_fence)
639 send_flags = IB_SEND_FENCE;
642 * We could be copying the header into the unused tail of the page.
643 * That would need to be changed in the future when those pages might
644 * be mapped userspace pages or page cache pages. So instead we always
645 * use a second sge and our long-lived ring of mapped headers. We send
646 * the header after the data so that the data payload can be aligned on
650 /* handle a 0-len message */
651 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
652 rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
656 /* if there's data reference it with a chain of work reqs */
657 for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
660 send = &ic->i_sends[pos];
662 len = min(RDS_FRAG_SIZE,
663 ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
664 rds_iw_xmit_populate_wr(ic, send, pos,
665 ib_sg_dma_address(dev, scat) + rm->data.op_dmaoff, len,
669 * We want to delay signaling completions just enough to get
670 * the batching benefits but not so much that we create dead time
673 if (ic->i_unsignaled_wrs-- == 0) {
674 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
675 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
678 ic->i_unsignaled_bytes -= len;
679 if (ic->i_unsignaled_bytes <= 0) {
680 ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
681 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
685 * Always signal the last one if we're stopping due to flow control.
687 if (flow_controlled && i == (work_alloc-1))
688 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
690 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
691 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
694 rm->data.op_dmaoff += len;
695 if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
697 rm->data.op_dmaoff = 0;
702 /* Tack on the header after the data. The header SGE should already
703 * have been set up to point to the right header buffer. */
704 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
707 struct rds_header *hdr = &ic->i_send_hdrs[pos];
709 printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
710 be16_to_cpu(hdr->h_dport),
712 be32_to_cpu(hdr->h_len));
715 struct rds_header *hdr = &ic->i_send_hdrs[pos];
717 /* add credit and redo the header checksum */
718 hdr->h_credit = adv_credits;
719 rds_message_make_checksum(hdr);
721 rds_iw_stats_inc(s_iw_tx_credit_updates);
725 prev->s_wr.next = &send->s_wr;
728 pos = (pos + 1) % ic->i_send_ring.w_nr;
731 /* Account the RDS header in the number of bytes we sent, but just once.
732 * The caller has no concept of fragmentation. */
734 sent += sizeof(struct rds_header);
736 /* if we finished the message then send completion owns it */
737 if (scat == &rm->data.op_sg[rm->data.op_count]) {
738 prev->s_rm = ic->i_rm;
739 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
743 if (i < work_alloc) {
744 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
747 if (ic->i_flowctl && i < credit_alloc)
748 rds_iw_send_add_credits(conn, credit_alloc - i);
750 /* XXX need to worry about failed_wr and partial sends. */
751 failed_wr = &first->s_wr;
752 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
753 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
754 first, &first->s_wr, ret, failed_wr);
755 BUG_ON(failed_wr != &first->s_wr);
757 printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
758 "returned %d\n", &conn->c_faddr, ret);
759 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
761 ic->i_rm = prev->s_rm;
773 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)
775 BUG_ON(nent > send->s_page_list->max_page_list_len);
777 * Perform a WR for the fast_reg_mr. Each individual page
778 * in the sg list is added to the fast reg page list and placed
779 * inside the fast_reg_mr WR.
781 send->s_wr.opcode = IB_WR_FAST_REG_MR;
782 send->s_wr.wr.fast_reg.length = len;
783 send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
784 send->s_wr.wr.fast_reg.page_list = send->s_page_list;
785 send->s_wr.wr.fast_reg.page_list_len = nent;
786 send->s_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
787 send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
788 send->s_wr.wr.fast_reg.iova_start = sg_addr;
790 ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
793 int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
795 struct rds_iw_connection *ic = conn->c_transport_data;
796 struct rds_iw_send_work *send = NULL;
797 struct rds_iw_send_work *first;
798 struct rds_iw_send_work *prev;
799 struct ib_send_wr *failed_wr;
800 struct rds_iw_device *rds_iwdev;
801 struct scatterlist *scat;
803 u64 remote_addr = op->op_remote_addr;
812 rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
814 /* map the message the first time we see it */
815 if (!op->op_mapped) {
816 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
817 op->op_sg, op->op_nents, (op->op_write) ?
818 DMA_TO_DEVICE : DMA_FROM_DEVICE);
819 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
820 if (op->op_count == 0) {
821 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
822 ret = -ENOMEM; /* XXX ? */
830 /* Alloc space on the send queue for the fastreg */
831 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
832 if (work_alloc != 1) {
833 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
834 rds_iw_stats_inc(s_iw_tx_ring_full);
841 * Instead of knowing how to return a partial rdma read/write we insist that there
842 * be enough work requests to send the entire message.
844 i = ceil(op->op_count, rds_iwdev->max_sge);
846 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
847 if (work_alloc != i) {
848 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
849 rds_iw_stats_inc(s_iw_tx_ring_full);
854 send = &ic->i_sends[pos];
856 first = prev = &ic->i_sends[fr_pos];
861 scat = &op->op_sg[0];
863 num_sge = op->op_count;
865 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
866 send->s_wr.send_flags = 0;
867 send->s_queued = jiffies;
870 * We want to delay signaling completions just enough to get
871 * the batching benefits but not so much that we create dead time on the wire.
873 if (ic->i_unsignaled_wrs-- == 0) {
874 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
875 send->s_wr.send_flags = IB_SEND_SIGNALED;
878 /* To avoid the need to have the plumbing to invalidate the fastreg_mr used
879 * for local access after RDS is finished with it, using
880 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
883 send->s_wr.opcode = IB_WR_RDMA_WRITE;
885 send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
887 send->s_wr.wr.rdma.remote_addr = remote_addr;
888 send->s_wr.wr.rdma.rkey = op->op_rkey;
891 if (num_sge > rds_iwdev->max_sge) {
892 send->s_wr.num_sge = rds_iwdev->max_sge;
893 num_sge -= rds_iwdev->max_sge;
895 send->s_wr.num_sge = num_sge;
897 send->s_wr.next = NULL;
900 prev->s_wr.next = &send->s_wr;
902 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
903 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
905 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
906 send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
908 send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
909 send->s_sge[j].length = len;
910 send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
914 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
920 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
921 send->s_wr.num_sge = 1;
922 send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
923 send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
924 send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
927 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
928 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
931 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
935 /* if we finished the message then send completion owns it */
936 if (scat == &op->op_sg[op->op_count])
937 first->s_wr.send_flags = IB_SEND_SIGNALED;
939 if (i < work_alloc) {
940 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
944 /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
945 * recommended. Putting the lkey on the wire is a security hole, as it can
946 * allow for memory access to all of memory on the remote system. Some
947 * adapters do not allow using the lkey for this at all. To bypass this use a
948 * fastreg_mr (or possibly a dma_mr)
951 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
952 op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
956 failed_wr = &first->s_wr;
957 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
958 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
959 first, &first->s_wr, ret, failed_wr);
960 BUG_ON(failed_wr != &first->s_wr);
962 printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
963 "returned %d\n", &conn->c_faddr, ret);
964 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
972 void rds_iw_xmit_complete(struct rds_connection *conn)
974 struct rds_iw_connection *ic = conn->c_transport_data;
976 /* We may have a pending ACK or window update we were unable
977 * to send previously (due to flow control). Try again. */
978 rds_iw_attempt_ack(ic);