af_unix: limit unix_tot_inflight
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / net / rds / ib_send.c
blob71f373c421bc4d8b6a97f2be7201a2dc8c2290e8
1 /*
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
12 * conditions are met:
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
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
30 * SOFTWARE.
33 #include <linux/kernel.h>
34 #include <linux/in.h>
35 #include <linux/device.h>
36 #include <linux/dmapool.h>
38 #include "rds.h"
39 #include "ib.h"
41 static char *rds_ib_wc_status_strings[] = {
42 #define RDS_IB_WC_STATUS_STR(foo) \
43 [IB_WC_##foo] = __stringify(IB_WC_##foo)
44 RDS_IB_WC_STATUS_STR(SUCCESS),
45 RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
46 RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
47 RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
48 RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
49 RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
50 RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
51 RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
52 RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
53 RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
54 RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
55 RDS_IB_WC_STATUS_STR(REM_OP_ERR),
56 RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
57 RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
58 RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
59 RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
60 RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
61 RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
62 RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
63 RDS_IB_WC_STATUS_STR(FATAL_ERR),
64 RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
65 RDS_IB_WC_STATUS_STR(GENERAL_ERR),
66 #undef RDS_IB_WC_STATUS_STR
69 char *rds_ib_wc_status_str(enum ib_wc_status status)
71 return rds_str_array(rds_ib_wc_status_strings,
72 ARRAY_SIZE(rds_ib_wc_status_strings), status);
76 * Convert IB-specific error message to RDS error message and call core
77 * completion handler.
79 static void rds_ib_send_complete(struct rds_message *rm,
80 int wc_status,
81 void (*complete)(struct rds_message *rm, int status))
83 int notify_status;
85 switch (wc_status) {
86 case IB_WC_WR_FLUSH_ERR:
87 return;
89 case IB_WC_SUCCESS:
90 notify_status = RDS_RDMA_SUCCESS;
91 break;
93 case IB_WC_REM_ACCESS_ERR:
94 notify_status = RDS_RDMA_REMOTE_ERROR;
95 break;
97 default:
98 notify_status = RDS_RDMA_OTHER_ERROR;
99 break;
101 complete(rm, notify_status);
104 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
105 struct rm_data_op *op,
106 int wc_status)
108 if (op->op_nents)
109 ib_dma_unmap_sg(ic->i_cm_id->device,
110 op->op_sg, op->op_nents,
111 DMA_TO_DEVICE);
114 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
115 struct rm_rdma_op *op,
116 int wc_status)
118 if (op->op_mapped) {
119 ib_dma_unmap_sg(ic->i_cm_id->device,
120 op->op_sg, op->op_nents,
121 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
122 op->op_mapped = 0;
125 /* If the user asked for a completion notification on this
126 * message, we can implement three different semantics:
127 * 1. Notify when we received the ACK on the RDS message
128 * that was queued with the RDMA. This provides reliable
129 * notification of RDMA status at the expense of a one-way
130 * packet delay.
131 * 2. Notify when the IB stack gives us the completion event for
132 * the RDMA operation.
133 * 3. Notify when the IB stack gives us the completion event for
134 * the accompanying RDS messages.
135 * Here, we implement approach #3. To implement approach #2,
136 * we would need to take an event for the rdma WR. To implement #1,
137 * don't call rds_rdma_send_complete at all, and fall back to the notify
138 * handling in the ACK processing code.
140 * Note: There's no need to explicitly sync any RDMA buffers using
141 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
142 * operation itself unmapped the RDMA buffers, which takes care
143 * of synching.
145 rds_ib_send_complete(container_of(op, struct rds_message, rdma),
146 wc_status, rds_rdma_send_complete);
148 if (op->op_write)
149 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
150 else
151 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
154 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
155 struct rm_atomic_op *op,
156 int wc_status)
158 /* unmap atomic recvbuf */
159 if (op->op_mapped) {
160 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
161 DMA_FROM_DEVICE);
162 op->op_mapped = 0;
165 rds_ib_send_complete(container_of(op, struct rds_message, atomic),
166 wc_status, rds_atomic_send_complete);
168 if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
169 rds_ib_stats_inc(s_ib_atomic_cswp);
170 else
171 rds_ib_stats_inc(s_ib_atomic_fadd);
175 * Unmap the resources associated with a struct send_work.
177 * Returns the rm for no good reason other than it is unobtainable
178 * other than by switching on wr.opcode, currently, and the caller,
179 * the event handler, needs it.
181 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
182 struct rds_ib_send_work *send,
183 int wc_status)
185 struct rds_message *rm = NULL;
187 /* In the error case, wc.opcode sometimes contains garbage */
188 switch (send->s_wr.opcode) {
189 case IB_WR_SEND:
190 if (send->s_op) {
191 rm = container_of(send->s_op, struct rds_message, data);
192 rds_ib_send_unmap_data(ic, send->s_op, wc_status);
194 break;
195 case IB_WR_RDMA_WRITE:
196 case IB_WR_RDMA_READ:
197 if (send->s_op) {
198 rm = container_of(send->s_op, struct rds_message, rdma);
199 rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
201 break;
202 case IB_WR_ATOMIC_FETCH_AND_ADD:
203 case IB_WR_ATOMIC_CMP_AND_SWP:
204 if (send->s_op) {
205 rm = container_of(send->s_op, struct rds_message, atomic);
206 rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
208 break;
209 default:
210 if (printk_ratelimit())
211 printk(KERN_NOTICE
212 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
213 __func__, send->s_wr.opcode);
214 break;
217 send->s_wr.opcode = 0xdead;
219 return rm;
222 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
224 struct rds_ib_send_work *send;
225 u32 i;
227 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
228 struct ib_sge *sge;
230 send->s_op = NULL;
232 send->s_wr.wr_id = i;
233 send->s_wr.sg_list = send->s_sge;
234 send->s_wr.ex.imm_data = 0;
236 sge = &send->s_sge[0];
237 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
238 sge->length = sizeof(struct rds_header);
239 sge->lkey = ic->i_mr->lkey;
241 send->s_sge[1].lkey = ic->i_mr->lkey;
245 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
247 struct rds_ib_send_work *send;
248 u32 i;
250 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
251 if (send->s_op && send->s_wr.opcode != 0xdead)
252 rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
257 * The only fast path caller always has a non-zero nr, so we don't
258 * bother testing nr before performing the atomic sub.
260 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
262 if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
263 waitqueue_active(&rds_ib_ring_empty_wait))
264 wake_up(&rds_ib_ring_empty_wait);
265 BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
269 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
270 * operations performed in the send path. As the sender allocs and potentially
271 * unallocs the next free entry in the ring it doesn't alter which is
272 * the next to be freed, which is what this is concerned with.
274 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
276 struct rds_connection *conn = context;
277 struct rds_ib_connection *ic = conn->c_transport_data;
278 struct rds_message *rm = NULL;
279 struct ib_wc wc;
280 struct rds_ib_send_work *send;
281 u32 completed;
282 u32 oldest;
283 u32 i = 0;
284 int ret;
285 int nr_sig = 0;
287 rdsdebug("cq %p conn %p\n", cq, conn);
288 rds_ib_stats_inc(s_ib_tx_cq_call);
289 ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
290 if (ret)
291 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
293 while (ib_poll_cq(cq, 1, &wc) > 0) {
294 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
295 (unsigned long long)wc.wr_id, wc.status,
296 rds_ib_wc_status_str(wc.status), wc.byte_len,
297 be32_to_cpu(wc.ex.imm_data));
298 rds_ib_stats_inc(s_ib_tx_cq_event);
300 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
301 if (ic->i_ack_queued + HZ/2 < jiffies)
302 rds_ib_stats_inc(s_ib_tx_stalled);
303 rds_ib_ack_send_complete(ic);
304 continue;
307 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
309 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
311 for (i = 0; i < completed; i++) {
312 send = &ic->i_sends[oldest];
313 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
314 nr_sig++;
316 rm = rds_ib_send_unmap_op(ic, send, wc.status);
318 if (send->s_queued + HZ/2 < jiffies)
319 rds_ib_stats_inc(s_ib_tx_stalled);
321 if (send->s_op) {
322 if (send->s_op == rm->m_final_op) {
323 /* If anyone waited for this message to get flushed out, wake
324 * them up now */
325 rds_message_unmapped(rm);
327 rds_message_put(rm);
328 send->s_op = NULL;
331 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
334 rds_ib_ring_free(&ic->i_send_ring, completed);
335 rds_ib_sub_signaled(ic, nr_sig);
336 nr_sig = 0;
338 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
339 test_bit(0, &conn->c_map_queued))
340 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
342 /* We expect errors as the qp is drained during shutdown */
343 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
344 rds_ib_conn_error(conn, "send completion on %pI4 had status "
345 "%u (%s), disconnecting and reconnecting\n",
346 &conn->c_faddr, wc.status,
347 rds_ib_wc_status_str(wc.status));
353 * This is the main function for allocating credits when sending
354 * messages.
356 * Conceptually, we have two counters:
357 * - send credits: this tells us how many WRs we're allowed
358 * to submit without overruning the reciever's queue. For
359 * each SEND WR we post, we decrement this by one.
361 * - posted credits: this tells us how many WRs we recently
362 * posted to the receive queue. This value is transferred
363 * to the peer as a "credit update" in a RDS header field.
364 * Every time we transmit credits to the peer, we subtract
365 * the amount of transferred credits from this counter.
367 * It is essential that we avoid situations where both sides have
368 * exhausted their send credits, and are unable to send new credits
369 * to the peer. We achieve this by requiring that we send at least
370 * one credit update to the peer before exhausting our credits.
371 * When new credits arrive, we subtract one credit that is withheld
372 * until we've posted new buffers and are ready to transmit these
373 * credits (see rds_ib_send_add_credits below).
375 * The RDS send code is essentially single-threaded; rds_send_xmit
376 * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
377 * However, the ACK sending code is independent and can race with
378 * message SENDs.
380 * In the send path, we need to update the counters for send credits
381 * and the counter of posted buffers atomically - when we use the
382 * last available credit, we cannot allow another thread to race us
383 * and grab the posted credits counter. Hence, we have to use a
384 * spinlock to protect the credit counter, or use atomics.
386 * Spinlocks shared between the send and the receive path are bad,
387 * because they create unnecessary delays. An early implementation
388 * using a spinlock showed a 5% degradation in throughput at some
389 * loads.
391 * This implementation avoids spinlocks completely, putting both
392 * counters into a single atomic, and updating that atomic using
393 * atomic_add (in the receive path, when receiving fresh credits),
394 * and using atomic_cmpxchg when updating the two counters.
396 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
397 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
399 unsigned int avail, posted, got = 0, advertise;
400 long oldval, newval;
402 *adv_credits = 0;
403 if (!ic->i_flowctl)
404 return wanted;
406 try_again:
407 advertise = 0;
408 oldval = newval = atomic_read(&ic->i_credits);
409 posted = IB_GET_POST_CREDITS(oldval);
410 avail = IB_GET_SEND_CREDITS(oldval);
412 rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
413 wanted, avail, posted);
415 /* The last credit must be used to send a credit update. */
416 if (avail && !posted)
417 avail--;
419 if (avail < wanted) {
420 struct rds_connection *conn = ic->i_cm_id->context;
422 /* Oops, there aren't that many credits left! */
423 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
424 got = avail;
425 } else {
426 /* Sometimes you get what you want, lalala. */
427 got = wanted;
429 newval -= IB_SET_SEND_CREDITS(got);
432 * If need_posted is non-zero, then the caller wants
433 * the posted regardless of whether any send credits are
434 * available.
436 if (posted && (got || need_posted)) {
437 advertise = min_t(unsigned int, posted, max_posted);
438 newval -= IB_SET_POST_CREDITS(advertise);
441 /* Finally bill everything */
442 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
443 goto try_again;
445 *adv_credits = advertise;
446 return got;
449 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
451 struct rds_ib_connection *ic = conn->c_transport_data;
453 if (credits == 0)
454 return;
456 rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
457 credits,
458 IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
459 test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
461 atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
462 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
463 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
465 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
467 rds_ib_stats_inc(s_ib_rx_credit_updates);
470 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
472 struct rds_ib_connection *ic = conn->c_transport_data;
474 if (posted == 0)
475 return;
477 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
479 /* Decide whether to send an update to the peer now.
480 * If we would send a credit update for every single buffer we
481 * post, we would end up with an ACK storm (ACK arrives,
482 * consumes buffer, we refill the ring, send ACK to remote
483 * advertising the newly posted buffer... ad inf)
485 * Performance pretty much depends on how often we send
486 * credit updates - too frequent updates mean lots of ACKs.
487 * Too infrequent updates, and the peer will run out of
488 * credits and has to throttle.
489 * For the time being, 16 seems to be a good compromise.
491 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
492 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
495 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
496 struct rds_ib_send_work *send,
497 bool notify)
500 * We want to delay signaling completions just enough to get
501 * the batching benefits but not so much that we create dead time
502 * on the wire.
504 if (ic->i_unsignaled_wrs-- == 0 || notify) {
505 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
506 send->s_wr.send_flags |= IB_SEND_SIGNALED;
507 return 1;
509 return 0;
513 * This can be called multiple times for a given message. The first time
514 * we see a message we map its scatterlist into the IB device so that
515 * we can provide that mapped address to the IB scatter gather entries
516 * in the IB work requests. We translate the scatterlist into a series
517 * of work requests that fragment the message. These work requests complete
518 * in order so we pass ownership of the message to the completion handler
519 * once we send the final fragment.
521 * The RDS core uses the c_send_lock to only enter this function once
522 * per connection. This makes sure that the tx ring alloc/unalloc pairs
523 * don't get out of sync and confuse the ring.
525 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
526 unsigned int hdr_off, unsigned int sg, unsigned int off)
528 struct rds_ib_connection *ic = conn->c_transport_data;
529 struct ib_device *dev = ic->i_cm_id->device;
530 struct rds_ib_send_work *send = NULL;
531 struct rds_ib_send_work *first;
532 struct rds_ib_send_work *prev;
533 struct ib_send_wr *failed_wr;
534 struct scatterlist *scat;
535 u32 pos;
536 u32 i;
537 u32 work_alloc;
538 u32 credit_alloc = 0;
539 u32 posted;
540 u32 adv_credits = 0;
541 int send_flags = 0;
542 int bytes_sent = 0;
543 int ret;
544 int flow_controlled = 0;
545 int nr_sig = 0;
547 BUG_ON(off % RDS_FRAG_SIZE);
548 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
550 /* Do not send cong updates to IB loopback */
551 if (conn->c_loopback
552 && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
553 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
554 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
557 /* FIXME we may overallocate here */
558 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
559 i = 1;
560 else
561 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
563 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
564 if (work_alloc == 0) {
565 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
566 rds_ib_stats_inc(s_ib_tx_ring_full);
567 ret = -ENOMEM;
568 goto out;
571 if (ic->i_flowctl) {
572 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
573 adv_credits += posted;
574 if (credit_alloc < work_alloc) {
575 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
576 work_alloc = credit_alloc;
577 flow_controlled = 1;
579 if (work_alloc == 0) {
580 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
581 rds_ib_stats_inc(s_ib_tx_throttle);
582 ret = -ENOMEM;
583 goto out;
587 /* map the message the first time we see it */
588 if (!ic->i_data_op) {
589 if (rm->data.op_nents) {
590 rm->data.op_count = ib_dma_map_sg(dev,
591 rm->data.op_sg,
592 rm->data.op_nents,
593 DMA_TO_DEVICE);
594 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
595 if (rm->data.op_count == 0) {
596 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
597 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
598 ret = -ENOMEM; /* XXX ? */
599 goto out;
601 } else {
602 rm->data.op_count = 0;
605 rds_message_addref(rm);
606 ic->i_data_op = &rm->data;
608 /* Finalize the header */
609 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
610 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
611 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
612 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
614 /* If it has a RDMA op, tell the peer we did it. This is
615 * used by the peer to release use-once RDMA MRs. */
616 if (rm->rdma.op_active) {
617 struct rds_ext_header_rdma ext_hdr;
619 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
620 rds_message_add_extension(&rm->m_inc.i_hdr,
621 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
623 if (rm->m_rdma_cookie) {
624 rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
625 rds_rdma_cookie_key(rm->m_rdma_cookie),
626 rds_rdma_cookie_offset(rm->m_rdma_cookie));
629 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
630 * we should not do this unless we have a chance of at least
631 * sticking the header into the send ring. Which is why we
632 * should call rds_ib_ring_alloc first. */
633 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
634 rds_message_make_checksum(&rm->m_inc.i_hdr);
637 * Update adv_credits since we reset the ACK_REQUIRED bit.
639 if (ic->i_flowctl) {
640 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
641 adv_credits += posted;
642 BUG_ON(adv_credits > 255);
646 /* Sometimes you want to put a fence between an RDMA
647 * READ and the following SEND.
648 * We could either do this all the time
649 * or when requested by the user. Right now, we let
650 * the application choose.
652 if (rm->rdma.op_active && rm->rdma.op_fence)
653 send_flags = IB_SEND_FENCE;
655 /* Each frag gets a header. Msgs may be 0 bytes */
656 send = &ic->i_sends[pos];
657 first = send;
658 prev = NULL;
659 scat = &ic->i_data_op->op_sg[sg];
660 i = 0;
661 do {
662 unsigned int len = 0;
664 /* Set up the header */
665 send->s_wr.send_flags = send_flags;
666 send->s_wr.opcode = IB_WR_SEND;
667 send->s_wr.num_sge = 1;
668 send->s_wr.next = NULL;
669 send->s_queued = jiffies;
670 send->s_op = NULL;
672 send->s_sge[0].addr = ic->i_send_hdrs_dma
673 + (pos * sizeof(struct rds_header));
674 send->s_sge[0].length = sizeof(struct rds_header);
676 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
678 /* Set up the data, if present */
679 if (i < work_alloc
680 && scat != &rm->data.op_sg[rm->data.op_count]) {
681 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
682 send->s_wr.num_sge = 2;
684 send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
685 send->s_sge[1].length = len;
687 bytes_sent += len;
688 off += len;
689 if (off == ib_sg_dma_len(dev, scat)) {
690 scat++;
691 off = 0;
695 rds_ib_set_wr_signal_state(ic, send, 0);
698 * Always signal the last one if we're stopping due to flow control.
700 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
701 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
703 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
704 nr_sig++;
706 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
707 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
709 if (ic->i_flowctl && adv_credits) {
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);
715 adv_credits = 0;
716 rds_ib_stats_inc(s_ib_tx_credit_updates);
719 if (prev)
720 prev->s_wr.next = &send->s_wr;
721 prev = send;
723 pos = (pos + 1) % ic->i_send_ring.w_nr;
724 send = &ic->i_sends[pos];
725 i++;
727 } while (i < work_alloc
728 && scat != &rm->data.op_sg[rm->data.op_count]);
730 /* Account the RDS header in the number of bytes we sent, but just once.
731 * The caller has no concept of fragmentation. */
732 if (hdr_off == 0)
733 bytes_sent += sizeof(struct rds_header);
735 /* if we finished the message then send completion owns it */
736 if (scat == &rm->data.op_sg[rm->data.op_count]) {
737 prev->s_op = ic->i_data_op;
738 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
739 ic->i_data_op = NULL;
742 /* Put back wrs & credits we didn't use */
743 if (i < work_alloc) {
744 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
745 work_alloc = i;
747 if (ic->i_flowctl && i < credit_alloc)
748 rds_ib_send_add_credits(conn, credit_alloc - i);
750 if (nr_sig)
751 atomic_add(nr_sig, &ic->i_signaled_sends);
753 /* XXX need to worry about failed_wr and partial sends. */
754 failed_wr = &first->s_wr;
755 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
756 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
757 first, &first->s_wr, ret, failed_wr);
758 BUG_ON(failed_wr != &first->s_wr);
759 if (ret) {
760 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
761 "returned %d\n", &conn->c_faddr, ret);
762 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
763 rds_ib_sub_signaled(ic, nr_sig);
764 if (prev->s_op) {
765 ic->i_data_op = prev->s_op;
766 prev->s_op = NULL;
769 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
770 goto out;
773 ret = bytes_sent;
774 out:
775 BUG_ON(adv_credits);
776 return ret;
780 * Issue atomic operation.
781 * A simplified version of the rdma case, we always map 1 SG, and
782 * only 8 bytes, for the return value from the atomic operation.
784 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
786 struct rds_ib_connection *ic = conn->c_transport_data;
787 struct rds_ib_send_work *send = NULL;
788 struct ib_send_wr *failed_wr;
789 struct rds_ib_device *rds_ibdev;
790 u32 pos;
791 u32 work_alloc;
792 int ret;
793 int nr_sig = 0;
795 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
797 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
798 if (work_alloc != 1) {
799 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
800 rds_ib_stats_inc(s_ib_tx_ring_full);
801 ret = -ENOMEM;
802 goto out;
805 /* address of send request in ring */
806 send = &ic->i_sends[pos];
807 send->s_queued = jiffies;
809 if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
810 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
811 send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
812 send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
813 send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
814 send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
815 } else { /* FADD */
816 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
817 send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
818 send->s_wr.wr.atomic.swap = 0;
819 send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
820 send->s_wr.wr.atomic.swap_mask = 0;
822 nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
823 send->s_wr.num_sge = 1;
824 send->s_wr.next = NULL;
825 send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
826 send->s_wr.wr.atomic.rkey = op->op_rkey;
827 send->s_op = op;
828 rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
830 /* map 8 byte retval buffer to the device */
831 ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
832 rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
833 if (ret != 1) {
834 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
835 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
836 ret = -ENOMEM; /* XXX ? */
837 goto out;
840 /* Convert our struct scatterlist to struct ib_sge */
841 send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
842 send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
843 send->s_sge[0].lkey = ic->i_mr->lkey;
845 rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
846 send->s_sge[0].addr, send->s_sge[0].length);
848 if (nr_sig)
849 atomic_add(nr_sig, &ic->i_signaled_sends);
851 failed_wr = &send->s_wr;
852 ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
853 rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
854 send, &send->s_wr, ret, failed_wr);
855 BUG_ON(failed_wr != &send->s_wr);
856 if (ret) {
857 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
858 "returned %d\n", &conn->c_faddr, ret);
859 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
860 rds_ib_sub_signaled(ic, nr_sig);
861 goto out;
864 if (unlikely(failed_wr != &send->s_wr)) {
865 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
866 BUG_ON(failed_wr != &send->s_wr);
869 out:
870 return ret;
873 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
875 struct rds_ib_connection *ic = conn->c_transport_data;
876 struct rds_ib_send_work *send = NULL;
877 struct rds_ib_send_work *first;
878 struct rds_ib_send_work *prev;
879 struct ib_send_wr *failed_wr;
880 struct scatterlist *scat;
881 unsigned long len;
882 u64 remote_addr = op->op_remote_addr;
883 u32 max_sge = ic->rds_ibdev->max_sge;
884 u32 pos;
885 u32 work_alloc;
886 u32 i;
887 u32 j;
888 int sent;
889 int ret;
890 int num_sge;
891 int nr_sig = 0;
893 /* map the op the first time we see it */
894 if (!op->op_mapped) {
895 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
896 op->op_sg, op->op_nents, (op->op_write) ?
897 DMA_TO_DEVICE : DMA_FROM_DEVICE);
898 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
899 if (op->op_count == 0) {
900 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
901 ret = -ENOMEM; /* XXX ? */
902 goto out;
905 op->op_mapped = 1;
909 * Instead of knowing how to return a partial rdma read/write we insist that there
910 * be enough work requests to send the entire message.
912 i = ceil(op->op_count, max_sge);
914 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
915 if (work_alloc != i) {
916 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
917 rds_ib_stats_inc(s_ib_tx_ring_full);
918 ret = -ENOMEM;
919 goto out;
922 send = &ic->i_sends[pos];
923 first = send;
924 prev = NULL;
925 scat = &op->op_sg[0];
926 sent = 0;
927 num_sge = op->op_count;
929 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
930 send->s_wr.send_flags = 0;
931 send->s_queued = jiffies;
932 send->s_op = NULL;
934 nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
936 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
937 send->s_wr.wr.rdma.remote_addr = remote_addr;
938 send->s_wr.wr.rdma.rkey = op->op_rkey;
940 if (num_sge > max_sge) {
941 send->s_wr.num_sge = max_sge;
942 num_sge -= max_sge;
943 } else {
944 send->s_wr.num_sge = num_sge;
947 send->s_wr.next = NULL;
949 if (prev)
950 prev->s_wr.next = &send->s_wr;
952 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
953 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
954 send->s_sge[j].addr =
955 ib_sg_dma_address(ic->i_cm_id->device, scat);
956 send->s_sge[j].length = len;
957 send->s_sge[j].lkey = ic->i_mr->lkey;
959 sent += len;
960 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
962 remote_addr += len;
963 scat++;
966 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
967 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
969 prev = send;
970 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
971 send = ic->i_sends;
974 /* give a reference to the last op */
975 if (scat == &op->op_sg[op->op_count]) {
976 prev->s_op = op;
977 rds_message_addref(container_of(op, struct rds_message, rdma));
980 if (i < work_alloc) {
981 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
982 work_alloc = i;
985 if (nr_sig)
986 atomic_add(nr_sig, &ic->i_signaled_sends);
988 failed_wr = &first->s_wr;
989 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
990 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
991 first, &first->s_wr, ret, failed_wr);
992 BUG_ON(failed_wr != &first->s_wr);
993 if (ret) {
994 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
995 "returned %d\n", &conn->c_faddr, ret);
996 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
997 rds_ib_sub_signaled(ic, nr_sig);
998 goto out;
1001 if (unlikely(failed_wr != &first->s_wr)) {
1002 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1003 BUG_ON(failed_wr != &first->s_wr);
1007 out:
1008 return ret;
1011 void rds_ib_xmit_complete(struct rds_connection *conn)
1013 struct rds_ib_connection *ic = conn->c_transport_data;
1015 /* We may have a pending ACK or window update we were unable
1016 * to send previously (due to flow control). Try again. */
1017 rds_ib_attempt_ack(ic);