Loading...
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/device.h>
36#include <linux/dmapool.h>
37#include <linux/ratelimit.h>
38
39#include "rds.h"
40#include "ib.h"
41
42static char *rds_ib_wc_status_strings[] = {
43#define RDS_IB_WC_STATUS_STR(foo) \
44 [IB_WC_##foo] = __stringify(IB_WC_##foo)
45 RDS_IB_WC_STATUS_STR(SUCCESS),
46 RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
47 RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
48 RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
49 RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
50 RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
51 RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
52 RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
53 RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
54 RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
55 RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
56 RDS_IB_WC_STATUS_STR(REM_OP_ERR),
57 RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
58 RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
59 RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
60 RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
61 RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
62 RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
63 RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
64 RDS_IB_WC_STATUS_STR(FATAL_ERR),
65 RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
66 RDS_IB_WC_STATUS_STR(GENERAL_ERR),
67#undef RDS_IB_WC_STATUS_STR
68};
69
70char *rds_ib_wc_status_str(enum ib_wc_status status)
71{
72 return rds_str_array(rds_ib_wc_status_strings,
73 ARRAY_SIZE(rds_ib_wc_status_strings), status);
74}
75
76/*
77 * Convert IB-specific error message to RDS error message and call core
78 * completion handler.
79 */
80static void rds_ib_send_complete(struct rds_message *rm,
81 int wc_status,
82 void (*complete)(struct rds_message *rm, int status))
83{
84 int notify_status;
85
86 switch (wc_status) {
87 case IB_WC_WR_FLUSH_ERR:
88 return;
89
90 case IB_WC_SUCCESS:
91 notify_status = RDS_RDMA_SUCCESS;
92 break;
93
94 case IB_WC_REM_ACCESS_ERR:
95 notify_status = RDS_RDMA_REMOTE_ERROR;
96 break;
97
98 default:
99 notify_status = RDS_RDMA_OTHER_ERROR;
100 break;
101 }
102 complete(rm, notify_status);
103}
104
105static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
106 struct rm_data_op *op,
107 int wc_status)
108{
109 if (op->op_nents)
110 ib_dma_unmap_sg(ic->i_cm_id->device,
111 op->op_sg, op->op_nents,
112 DMA_TO_DEVICE);
113}
114
115static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
116 struct rm_rdma_op *op,
117 int wc_status)
118{
119 if (op->op_mapped) {
120 ib_dma_unmap_sg(ic->i_cm_id->device,
121 op->op_sg, op->op_nents,
122 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
123 op->op_mapped = 0;
124 }
125
126 /* If the user asked for a completion notification on this
127 * message, we can implement three different semantics:
128 * 1. Notify when we received the ACK on the RDS message
129 * that was queued with the RDMA. This provides reliable
130 * notification of RDMA status at the expense of a one-way
131 * packet delay.
132 * 2. Notify when the IB stack gives us the completion event for
133 * the RDMA operation.
134 * 3. Notify when the IB stack gives us the completion event for
135 * the accompanying RDS messages.
136 * Here, we implement approach #3. To implement approach #2,
137 * we would need to take an event for the rdma WR. To implement #1,
138 * don't call rds_rdma_send_complete at all, and fall back to the notify
139 * handling in the ACK processing code.
140 *
141 * Note: There's no need to explicitly sync any RDMA buffers using
142 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
143 * operation itself unmapped the RDMA buffers, which takes care
144 * of synching.
145 */
146 rds_ib_send_complete(container_of(op, struct rds_message, rdma),
147 wc_status, rds_rdma_send_complete);
148
149 if (op->op_write)
150 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
151 else
152 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
153}
154
155static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
156 struct rm_atomic_op *op,
157 int wc_status)
158{
159 /* unmap atomic recvbuf */
160 if (op->op_mapped) {
161 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
162 DMA_FROM_DEVICE);
163 op->op_mapped = 0;
164 }
165
166 rds_ib_send_complete(container_of(op, struct rds_message, atomic),
167 wc_status, rds_atomic_send_complete);
168
169 if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
170 rds_ib_stats_inc(s_ib_atomic_cswp);
171 else
172 rds_ib_stats_inc(s_ib_atomic_fadd);
173}
174
175/*
176 * Unmap the resources associated with a struct send_work.
177 *
178 * Returns the rm for no good reason other than it is unobtainable
179 * other than by switching on wr.opcode, currently, and the caller,
180 * the event handler, needs it.
181 */
182static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
183 struct rds_ib_send_work *send,
184 int wc_status)
185{
186 struct rds_message *rm = NULL;
187
188 /* In the error case, wc.opcode sometimes contains garbage */
189 switch (send->s_wr.opcode) {
190 case IB_WR_SEND:
191 if (send->s_op) {
192 rm = container_of(send->s_op, struct rds_message, data);
193 rds_ib_send_unmap_data(ic, send->s_op, wc_status);
194 }
195 break;
196 case IB_WR_RDMA_WRITE:
197 case IB_WR_RDMA_READ:
198 if (send->s_op) {
199 rm = container_of(send->s_op, struct rds_message, rdma);
200 rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
201 }
202 break;
203 case IB_WR_ATOMIC_FETCH_AND_ADD:
204 case IB_WR_ATOMIC_CMP_AND_SWP:
205 if (send->s_op) {
206 rm = container_of(send->s_op, struct rds_message, atomic);
207 rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
208 }
209 break;
210 default:
211 printk_ratelimited(KERN_NOTICE
212 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
213 __func__, send->s_wr.opcode);
214 break;
215 }
216
217 send->s_wr.opcode = 0xdead;
218
219 return rm;
220}
221
222void rds_ib_send_init_ring(struct rds_ib_connection *ic)
223{
224 struct rds_ib_send_work *send;
225 u32 i;
226
227 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
228 struct ib_sge *sge;
229
230 send->s_op = NULL;
231
232 send->s_wr.wr_id = i;
233 send->s_wr.sg_list = send->s_sge;
234 send->s_wr.ex.imm_data = 0;
235
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;
240
241 send->s_sge[1].lkey = ic->i_mr->lkey;
242 }
243}
244
245void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
246{
247 struct rds_ib_send_work *send;
248 u32 i;
249
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);
253 }
254}
255
256/*
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.
259 */
260static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
261{
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);
266}
267
268/*
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.
273 */
274void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
275{
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;
286
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);
292
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);
299
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;
305 }
306
307 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
308
309 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
310
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++;
315
316 rm = rds_ib_send_unmap_op(ic, send, wc.status);
317
318 if (send->s_queued + HZ/2 < jiffies)
319 rds_ib_stats_inc(s_ib_tx_stalled);
320
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);
326 }
327 rds_message_put(rm);
328 send->s_op = NULL;
329 }
330
331 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
332 }
333
334 rds_ib_ring_free(&ic->i_send_ring, completed);
335 rds_ib_sub_signaled(ic, nr_sig);
336 nr_sig = 0;
337
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);
341
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));
348 }
349 }
350}
351
352/*
353 * This is the main function for allocating credits when sending
354 * messages.
355 *
356 * Conceptually, we have two counters:
357 * - send credits: this tells us how many WRs we're allowed
358 * to submit without overruning the receiver's queue. For
359 * each SEND WR we post, we decrement this by one.
360 *
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.
366 *
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).
374 *
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.
379 *
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.
385 *
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.
390 *
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.
395 */
396int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
397 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
398{
399 unsigned int avail, posted, got = 0, advertise;
400 long oldval, newval;
401
402 *adv_credits = 0;
403 if (!ic->i_flowctl)
404 return wanted;
405
406try_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);
411
412 rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
413 wanted, avail, posted);
414
415 /* The last credit must be used to send a credit update. */
416 if (avail && !posted)
417 avail--;
418
419 if (avail < wanted) {
420 struct rds_connection *conn = ic->i_cm_id->context;
421
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;
428 }
429 newval -= IB_SET_SEND_CREDITS(got);
430
431 /*
432 * If need_posted is non-zero, then the caller wants
433 * the posted regardless of whether any send credits are
434 * available.
435 */
436 if (posted && (got || need_posted)) {
437 advertise = min_t(unsigned int, posted, max_posted);
438 newval -= IB_SET_POST_CREDITS(advertise);
439 }
440
441 /* Finally bill everything */
442 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
443 goto try_again;
444
445 *adv_credits = advertise;
446 return got;
447}
448
449void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
450{
451 struct rds_ib_connection *ic = conn->c_transport_data;
452
453 if (credits == 0)
454 return;
455
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" : "");
460
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);
464
465 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
466
467 rds_ib_stats_inc(s_ib_rx_credit_updates);
468}
469
470void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
471{
472 struct rds_ib_connection *ic = conn->c_transport_data;
473
474 if (posted == 0)
475 return;
476
477 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
478
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)
484 *
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.
490 */
491 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
492 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
493}
494
495static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
496 struct rds_ib_send_work *send,
497 bool notify)
498{
499 /*
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.
503 */
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;
508 }
509 return 0;
510}
511
512/*
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.
520 *
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.
524 */
525int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
526 unsigned int hdr_off, unsigned int sg, unsigned int off)
527{
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;
546
547 BUG_ON(off % RDS_FRAG_SIZE);
548 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
549
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 scat = &rm->data.op_sg[sg];
555 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
556 return sizeof(struct rds_header) + ret;
557 }
558
559 /* FIXME we may overallocate here */
560 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
561 i = 1;
562 else
563 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
564
565 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
566 if (work_alloc == 0) {
567 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
568 rds_ib_stats_inc(s_ib_tx_ring_full);
569 ret = -ENOMEM;
570 goto out;
571 }
572
573 if (ic->i_flowctl) {
574 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
575 adv_credits += posted;
576 if (credit_alloc < work_alloc) {
577 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
578 work_alloc = credit_alloc;
579 flow_controlled = 1;
580 }
581 if (work_alloc == 0) {
582 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
583 rds_ib_stats_inc(s_ib_tx_throttle);
584 ret = -ENOMEM;
585 goto out;
586 }
587 }
588
589 /* map the message the first time we see it */
590 if (!ic->i_data_op) {
591 if (rm->data.op_nents) {
592 rm->data.op_count = ib_dma_map_sg(dev,
593 rm->data.op_sg,
594 rm->data.op_nents,
595 DMA_TO_DEVICE);
596 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
597 if (rm->data.op_count == 0) {
598 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
599 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
600 ret = -ENOMEM; /* XXX ? */
601 goto out;
602 }
603 } else {
604 rm->data.op_count = 0;
605 }
606
607 rds_message_addref(rm);
608 ic->i_data_op = &rm->data;
609
610 /* Finalize the header */
611 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
612 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
613 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
614 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
615
616 /* If it has a RDMA op, tell the peer we did it. This is
617 * used by the peer to release use-once RDMA MRs. */
618 if (rm->rdma.op_active) {
619 struct rds_ext_header_rdma ext_hdr;
620
621 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
622 rds_message_add_extension(&rm->m_inc.i_hdr,
623 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
624 }
625 if (rm->m_rdma_cookie) {
626 rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
627 rds_rdma_cookie_key(rm->m_rdma_cookie),
628 rds_rdma_cookie_offset(rm->m_rdma_cookie));
629 }
630
631 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
632 * we should not do this unless we have a chance of at least
633 * sticking the header into the send ring. Which is why we
634 * should call rds_ib_ring_alloc first. */
635 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
636 rds_message_make_checksum(&rm->m_inc.i_hdr);
637
638 /*
639 * Update adv_credits since we reset the ACK_REQUIRED bit.
640 */
641 if (ic->i_flowctl) {
642 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
643 adv_credits += posted;
644 BUG_ON(adv_credits > 255);
645 }
646 }
647
648 /* Sometimes you want to put a fence between an RDMA
649 * READ and the following SEND.
650 * We could either do this all the time
651 * or when requested by the user. Right now, we let
652 * the application choose.
653 */
654 if (rm->rdma.op_active && rm->rdma.op_fence)
655 send_flags = IB_SEND_FENCE;
656
657 /* Each frag gets a header. Msgs may be 0 bytes */
658 send = &ic->i_sends[pos];
659 first = send;
660 prev = NULL;
661 scat = &ic->i_data_op->op_sg[sg];
662 i = 0;
663 do {
664 unsigned int len = 0;
665
666 /* Set up the header */
667 send->s_wr.send_flags = send_flags;
668 send->s_wr.opcode = IB_WR_SEND;
669 send->s_wr.num_sge = 1;
670 send->s_wr.next = NULL;
671 send->s_queued = jiffies;
672 send->s_op = NULL;
673
674 send->s_sge[0].addr = ic->i_send_hdrs_dma
675 + (pos * sizeof(struct rds_header));
676 send->s_sge[0].length = sizeof(struct rds_header);
677
678 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
679
680 /* Set up the data, if present */
681 if (i < work_alloc
682 && scat != &rm->data.op_sg[rm->data.op_count]) {
683 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
684 send->s_wr.num_sge = 2;
685
686 send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
687 send->s_sge[1].length = len;
688
689 bytes_sent += len;
690 off += len;
691 if (off == ib_sg_dma_len(dev, scat)) {
692 scat++;
693 off = 0;
694 }
695 }
696
697 rds_ib_set_wr_signal_state(ic, send, 0);
698
699 /*
700 * Always signal the last one if we're stopping due to flow control.
701 */
702 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
703 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
704
705 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
706 nr_sig++;
707
708 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
709 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
710
711 if (ic->i_flowctl && adv_credits) {
712 struct rds_header *hdr = &ic->i_send_hdrs[pos];
713
714 /* add credit and redo the header checksum */
715 hdr->h_credit = adv_credits;
716 rds_message_make_checksum(hdr);
717 adv_credits = 0;
718 rds_ib_stats_inc(s_ib_tx_credit_updates);
719 }
720
721 if (prev)
722 prev->s_wr.next = &send->s_wr;
723 prev = send;
724
725 pos = (pos + 1) % ic->i_send_ring.w_nr;
726 send = &ic->i_sends[pos];
727 i++;
728
729 } while (i < work_alloc
730 && scat != &rm->data.op_sg[rm->data.op_count]);
731
732 /* Account the RDS header in the number of bytes we sent, but just once.
733 * The caller has no concept of fragmentation. */
734 if (hdr_off == 0)
735 bytes_sent += sizeof(struct rds_header);
736
737 /* if we finished the message then send completion owns it */
738 if (scat == &rm->data.op_sg[rm->data.op_count]) {
739 prev->s_op = ic->i_data_op;
740 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
741 ic->i_data_op = NULL;
742 }
743
744 /* Put back wrs & credits we didn't use */
745 if (i < work_alloc) {
746 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
747 work_alloc = i;
748 }
749 if (ic->i_flowctl && i < credit_alloc)
750 rds_ib_send_add_credits(conn, credit_alloc - i);
751
752 if (nr_sig)
753 atomic_add(nr_sig, &ic->i_signaled_sends);
754
755 /* XXX need to worry about failed_wr and partial sends. */
756 failed_wr = &first->s_wr;
757 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
758 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
759 first, &first->s_wr, ret, failed_wr);
760 BUG_ON(failed_wr != &first->s_wr);
761 if (ret) {
762 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
763 "returned %d\n", &conn->c_faddr, ret);
764 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
765 rds_ib_sub_signaled(ic, nr_sig);
766 if (prev->s_op) {
767 ic->i_data_op = prev->s_op;
768 prev->s_op = NULL;
769 }
770
771 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
772 goto out;
773 }
774
775 ret = bytes_sent;
776out:
777 BUG_ON(adv_credits);
778 return ret;
779}
780
781/*
782 * Issue atomic operation.
783 * A simplified version of the rdma case, we always map 1 SG, and
784 * only 8 bytes, for the return value from the atomic operation.
785 */
786int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
787{
788 struct rds_ib_connection *ic = conn->c_transport_data;
789 struct rds_ib_send_work *send = NULL;
790 struct ib_send_wr *failed_wr;
791 struct rds_ib_device *rds_ibdev;
792 u32 pos;
793 u32 work_alloc;
794 int ret;
795 int nr_sig = 0;
796
797 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
798
799 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
800 if (work_alloc != 1) {
801 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
802 rds_ib_stats_inc(s_ib_tx_ring_full);
803 ret = -ENOMEM;
804 goto out;
805 }
806
807 /* address of send request in ring */
808 send = &ic->i_sends[pos];
809 send->s_queued = jiffies;
810
811 if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
812 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
813 send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
814 send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
815 send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
816 send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
817 } else { /* FADD */
818 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
819 send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
820 send->s_wr.wr.atomic.swap = 0;
821 send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
822 send->s_wr.wr.atomic.swap_mask = 0;
823 }
824 nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
825 send->s_wr.num_sge = 1;
826 send->s_wr.next = NULL;
827 send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
828 send->s_wr.wr.atomic.rkey = op->op_rkey;
829 send->s_op = op;
830 rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
831
832 /* map 8 byte retval buffer to the device */
833 ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
834 rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
835 if (ret != 1) {
836 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
837 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
838 ret = -ENOMEM; /* XXX ? */
839 goto out;
840 }
841
842 /* Convert our struct scatterlist to struct ib_sge */
843 send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
844 send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
845 send->s_sge[0].lkey = ic->i_mr->lkey;
846
847 rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
848 send->s_sge[0].addr, send->s_sge[0].length);
849
850 if (nr_sig)
851 atomic_add(nr_sig, &ic->i_signaled_sends);
852
853 failed_wr = &send->s_wr;
854 ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
855 rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
856 send, &send->s_wr, ret, failed_wr);
857 BUG_ON(failed_wr != &send->s_wr);
858 if (ret) {
859 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
860 "returned %d\n", &conn->c_faddr, ret);
861 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
862 rds_ib_sub_signaled(ic, nr_sig);
863 goto out;
864 }
865
866 if (unlikely(failed_wr != &send->s_wr)) {
867 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
868 BUG_ON(failed_wr != &send->s_wr);
869 }
870
871out:
872 return ret;
873}
874
875int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
876{
877 struct rds_ib_connection *ic = conn->c_transport_data;
878 struct rds_ib_send_work *send = NULL;
879 struct rds_ib_send_work *first;
880 struct rds_ib_send_work *prev;
881 struct ib_send_wr *failed_wr;
882 struct scatterlist *scat;
883 unsigned long len;
884 u64 remote_addr = op->op_remote_addr;
885 u32 max_sge = ic->rds_ibdev->max_sge;
886 u32 pos;
887 u32 work_alloc;
888 u32 i;
889 u32 j;
890 int sent;
891 int ret;
892 int num_sge;
893 int nr_sig = 0;
894
895 /* map the op the first time we see it */
896 if (!op->op_mapped) {
897 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
898 op->op_sg, op->op_nents, (op->op_write) ?
899 DMA_TO_DEVICE : DMA_FROM_DEVICE);
900 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
901 if (op->op_count == 0) {
902 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
903 ret = -ENOMEM; /* XXX ? */
904 goto out;
905 }
906
907 op->op_mapped = 1;
908 }
909
910 /*
911 * Instead of knowing how to return a partial rdma read/write we insist that there
912 * be enough work requests to send the entire message.
913 */
914 i = ceil(op->op_count, max_sge);
915
916 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
917 if (work_alloc != i) {
918 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
919 rds_ib_stats_inc(s_ib_tx_ring_full);
920 ret = -ENOMEM;
921 goto out;
922 }
923
924 send = &ic->i_sends[pos];
925 first = send;
926 prev = NULL;
927 scat = &op->op_sg[0];
928 sent = 0;
929 num_sge = op->op_count;
930
931 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
932 send->s_wr.send_flags = 0;
933 send->s_queued = jiffies;
934 send->s_op = NULL;
935
936 nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
937
938 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
939 send->s_wr.wr.rdma.remote_addr = remote_addr;
940 send->s_wr.wr.rdma.rkey = op->op_rkey;
941
942 if (num_sge > max_sge) {
943 send->s_wr.num_sge = max_sge;
944 num_sge -= max_sge;
945 } else {
946 send->s_wr.num_sge = num_sge;
947 }
948
949 send->s_wr.next = NULL;
950
951 if (prev)
952 prev->s_wr.next = &send->s_wr;
953
954 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
955 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
956 send->s_sge[j].addr =
957 ib_sg_dma_address(ic->i_cm_id->device, scat);
958 send->s_sge[j].length = len;
959 send->s_sge[j].lkey = ic->i_mr->lkey;
960
961 sent += len;
962 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
963
964 remote_addr += len;
965 scat++;
966 }
967
968 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
969 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
970
971 prev = send;
972 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
973 send = ic->i_sends;
974 }
975
976 /* give a reference to the last op */
977 if (scat == &op->op_sg[op->op_count]) {
978 prev->s_op = op;
979 rds_message_addref(container_of(op, struct rds_message, rdma));
980 }
981
982 if (i < work_alloc) {
983 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
984 work_alloc = i;
985 }
986
987 if (nr_sig)
988 atomic_add(nr_sig, &ic->i_signaled_sends);
989
990 failed_wr = &first->s_wr;
991 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
992 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
993 first, &first->s_wr, ret, failed_wr);
994 BUG_ON(failed_wr != &first->s_wr);
995 if (ret) {
996 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
997 "returned %d\n", &conn->c_faddr, ret);
998 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
999 rds_ib_sub_signaled(ic, nr_sig);
1000 goto out;
1001 }
1002
1003 if (unlikely(failed_wr != &first->s_wr)) {
1004 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1005 BUG_ON(failed_wr != &first->s_wr);
1006 }
1007
1008
1009out:
1010 return ret;
1011}
1012
1013void rds_ib_xmit_complete(struct rds_connection *conn)
1014{
1015 struct rds_ib_connection *ic = conn->c_transport_data;
1016
1017 /* We may have a pending ACK or window update we were unable
1018 * to send previously (due to flow control). Try again. */
1019 rds_ib_attempt_ack(ic);
1020}
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/device.h>
36#include <linux/dmapool.h>
37#include <linux/ratelimit.h>
38
39#include "rds_single_path.h"
40#include "rds.h"
41#include "ib.h"
42
43/*
44 * Convert IB-specific error message to RDS error message and call core
45 * completion handler.
46 */
47static void rds_ib_send_complete(struct rds_message *rm,
48 int wc_status,
49 void (*complete)(struct rds_message *rm, int status))
50{
51 int notify_status;
52
53 switch (wc_status) {
54 case IB_WC_WR_FLUSH_ERR:
55 return;
56
57 case IB_WC_SUCCESS:
58 notify_status = RDS_RDMA_SUCCESS;
59 break;
60
61 case IB_WC_REM_ACCESS_ERR:
62 notify_status = RDS_RDMA_REMOTE_ERROR;
63 break;
64
65 default:
66 notify_status = RDS_RDMA_OTHER_ERROR;
67 break;
68 }
69 complete(rm, notify_status);
70}
71
72static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
73 struct rm_data_op *op,
74 int wc_status)
75{
76 if (op->op_nents)
77 ib_dma_unmap_sg(ic->i_cm_id->device,
78 op->op_sg, op->op_nents,
79 DMA_TO_DEVICE);
80}
81
82static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
83 struct rm_rdma_op *op,
84 int wc_status)
85{
86 if (op->op_mapped) {
87 ib_dma_unmap_sg(ic->i_cm_id->device,
88 op->op_sg, op->op_nents,
89 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
90 op->op_mapped = 0;
91 }
92
93 /* If the user asked for a completion notification on this
94 * message, we can implement three different semantics:
95 * 1. Notify when we received the ACK on the RDS message
96 * that was queued with the RDMA. This provides reliable
97 * notification of RDMA status at the expense of a one-way
98 * packet delay.
99 * 2. Notify when the IB stack gives us the completion event for
100 * the RDMA operation.
101 * 3. Notify when the IB stack gives us the completion event for
102 * the accompanying RDS messages.
103 * Here, we implement approach #3. To implement approach #2,
104 * we would need to take an event for the rdma WR. To implement #1,
105 * don't call rds_rdma_send_complete at all, and fall back to the notify
106 * handling in the ACK processing code.
107 *
108 * Note: There's no need to explicitly sync any RDMA buffers using
109 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
110 * operation itself unmapped the RDMA buffers, which takes care
111 * of synching.
112 */
113 rds_ib_send_complete(container_of(op, struct rds_message, rdma),
114 wc_status, rds_rdma_send_complete);
115
116 if (op->op_write)
117 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
118 else
119 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
120}
121
122static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
123 struct rm_atomic_op *op,
124 int wc_status)
125{
126 /* unmap atomic recvbuf */
127 if (op->op_mapped) {
128 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
129 DMA_FROM_DEVICE);
130 op->op_mapped = 0;
131 }
132
133 rds_ib_send_complete(container_of(op, struct rds_message, atomic),
134 wc_status, rds_atomic_send_complete);
135
136 if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
137 rds_ib_stats_inc(s_ib_atomic_cswp);
138 else
139 rds_ib_stats_inc(s_ib_atomic_fadd);
140}
141
142/*
143 * Unmap the resources associated with a struct send_work.
144 *
145 * Returns the rm for no good reason other than it is unobtainable
146 * other than by switching on wr.opcode, currently, and the caller,
147 * the event handler, needs it.
148 */
149static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
150 struct rds_ib_send_work *send,
151 int wc_status)
152{
153 struct rds_message *rm = NULL;
154
155 /* In the error case, wc.opcode sometimes contains garbage */
156 switch (send->s_wr.opcode) {
157 case IB_WR_SEND:
158 if (send->s_op) {
159 rm = container_of(send->s_op, struct rds_message, data);
160 rds_ib_send_unmap_data(ic, send->s_op, wc_status);
161 }
162 break;
163 case IB_WR_RDMA_WRITE:
164 case IB_WR_RDMA_READ:
165 if (send->s_op) {
166 rm = container_of(send->s_op, struct rds_message, rdma);
167 rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
168 }
169 break;
170 case IB_WR_ATOMIC_FETCH_AND_ADD:
171 case IB_WR_ATOMIC_CMP_AND_SWP:
172 if (send->s_op) {
173 rm = container_of(send->s_op, struct rds_message, atomic);
174 rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
175 }
176 break;
177 default:
178 printk_ratelimited(KERN_NOTICE
179 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
180 __func__, send->s_wr.opcode);
181 break;
182 }
183
184 send->s_wr.opcode = 0xdead;
185
186 return rm;
187}
188
189void rds_ib_send_init_ring(struct rds_ib_connection *ic)
190{
191 struct rds_ib_send_work *send;
192 u32 i;
193
194 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
195 struct ib_sge *sge;
196
197 send->s_op = NULL;
198
199 send->s_wr.wr_id = i;
200 send->s_wr.sg_list = send->s_sge;
201 send->s_wr.ex.imm_data = 0;
202
203 sge = &send->s_sge[0];
204 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
205 sge->length = sizeof(struct rds_header);
206 sge->lkey = ic->i_pd->local_dma_lkey;
207
208 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
209 }
210}
211
212void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
213{
214 struct rds_ib_send_work *send;
215 u32 i;
216
217 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
218 if (send->s_op && send->s_wr.opcode != 0xdead)
219 rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
220 }
221}
222
223/*
224 * The only fast path caller always has a non-zero nr, so we don't
225 * bother testing nr before performing the atomic sub.
226 */
227static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
228{
229 if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
230 waitqueue_active(&rds_ib_ring_empty_wait))
231 wake_up(&rds_ib_ring_empty_wait);
232 BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
233}
234
235/*
236 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
237 * operations performed in the send path. As the sender allocs and potentially
238 * unallocs the next free entry in the ring it doesn't alter which is
239 * the next to be freed, which is what this is concerned with.
240 */
241void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
242{
243 struct rds_message *rm = NULL;
244 struct rds_connection *conn = ic->conn;
245 struct rds_ib_send_work *send;
246 u32 completed;
247 u32 oldest;
248 u32 i = 0;
249 int nr_sig = 0;
250
251
252 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
253 (unsigned long long)wc->wr_id, wc->status,
254 ib_wc_status_msg(wc->status), wc->byte_len,
255 be32_to_cpu(wc->ex.imm_data));
256 rds_ib_stats_inc(s_ib_tx_cq_event);
257
258 if (wc->wr_id == RDS_IB_ACK_WR_ID) {
259 if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
260 rds_ib_stats_inc(s_ib_tx_stalled);
261 rds_ib_ack_send_complete(ic);
262 return;
263 }
264
265 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
266
267 completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
268
269 for (i = 0; i < completed; i++) {
270 send = &ic->i_sends[oldest];
271 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
272 nr_sig++;
273
274 rm = rds_ib_send_unmap_op(ic, send, wc->status);
275
276 if (time_after(jiffies, send->s_queued + HZ / 2))
277 rds_ib_stats_inc(s_ib_tx_stalled);
278
279 if (send->s_op) {
280 if (send->s_op == rm->m_final_op) {
281 /* If anyone waited for this message to get
282 * flushed out, wake them up now
283 */
284 rds_message_unmapped(rm);
285 }
286 rds_message_put(rm);
287 send->s_op = NULL;
288 }
289
290 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
291 }
292
293 rds_ib_ring_free(&ic->i_send_ring, completed);
294 rds_ib_sub_signaled(ic, nr_sig);
295 nr_sig = 0;
296
297 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
298 test_bit(0, &conn->c_map_queued))
299 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
300
301 /* We expect errors as the qp is drained during shutdown */
302 if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
303 rds_ib_conn_error(conn, "send completion on %pI4 had status %u (%s), disconnecting and reconnecting\n",
304 &conn->c_faddr, wc->status,
305 ib_wc_status_msg(wc->status));
306 }
307}
308
309/*
310 * This is the main function for allocating credits when sending
311 * messages.
312 *
313 * Conceptually, we have two counters:
314 * - send credits: this tells us how many WRs we're allowed
315 * to submit without overruning the receiver's queue. For
316 * each SEND WR we post, we decrement this by one.
317 *
318 * - posted credits: this tells us how many WRs we recently
319 * posted to the receive queue. This value is transferred
320 * to the peer as a "credit update" in a RDS header field.
321 * Every time we transmit credits to the peer, we subtract
322 * the amount of transferred credits from this counter.
323 *
324 * It is essential that we avoid situations where both sides have
325 * exhausted their send credits, and are unable to send new credits
326 * to the peer. We achieve this by requiring that we send at least
327 * one credit update to the peer before exhausting our credits.
328 * When new credits arrive, we subtract one credit that is withheld
329 * until we've posted new buffers and are ready to transmit these
330 * credits (see rds_ib_send_add_credits below).
331 *
332 * The RDS send code is essentially single-threaded; rds_send_xmit
333 * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
334 * However, the ACK sending code is independent and can race with
335 * message SENDs.
336 *
337 * In the send path, we need to update the counters for send credits
338 * and the counter of posted buffers atomically - when we use the
339 * last available credit, we cannot allow another thread to race us
340 * and grab the posted credits counter. Hence, we have to use a
341 * spinlock to protect the credit counter, or use atomics.
342 *
343 * Spinlocks shared between the send and the receive path are bad,
344 * because they create unnecessary delays. An early implementation
345 * using a spinlock showed a 5% degradation in throughput at some
346 * loads.
347 *
348 * This implementation avoids spinlocks completely, putting both
349 * counters into a single atomic, and updating that atomic using
350 * atomic_add (in the receive path, when receiving fresh credits),
351 * and using atomic_cmpxchg when updating the two counters.
352 */
353int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
354 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
355{
356 unsigned int avail, posted, got = 0, advertise;
357 long oldval, newval;
358
359 *adv_credits = 0;
360 if (!ic->i_flowctl)
361 return wanted;
362
363try_again:
364 advertise = 0;
365 oldval = newval = atomic_read(&ic->i_credits);
366 posted = IB_GET_POST_CREDITS(oldval);
367 avail = IB_GET_SEND_CREDITS(oldval);
368
369 rdsdebug("wanted=%u credits=%u posted=%u\n",
370 wanted, avail, posted);
371
372 /* The last credit must be used to send a credit update. */
373 if (avail && !posted)
374 avail--;
375
376 if (avail < wanted) {
377 struct rds_connection *conn = ic->i_cm_id->context;
378
379 /* Oops, there aren't that many credits left! */
380 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
381 got = avail;
382 } else {
383 /* Sometimes you get what you want, lalala. */
384 got = wanted;
385 }
386 newval -= IB_SET_SEND_CREDITS(got);
387
388 /*
389 * If need_posted is non-zero, then the caller wants
390 * the posted regardless of whether any send credits are
391 * available.
392 */
393 if (posted && (got || need_posted)) {
394 advertise = min_t(unsigned int, posted, max_posted);
395 newval -= IB_SET_POST_CREDITS(advertise);
396 }
397
398 /* Finally bill everything */
399 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
400 goto try_again;
401
402 *adv_credits = advertise;
403 return got;
404}
405
406void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
407{
408 struct rds_ib_connection *ic = conn->c_transport_data;
409
410 if (credits == 0)
411 return;
412
413 rdsdebug("credits=%u current=%u%s\n",
414 credits,
415 IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
416 test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
417
418 atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
419 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
420 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
421
422 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
423
424 rds_ib_stats_inc(s_ib_rx_credit_updates);
425}
426
427void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
428{
429 struct rds_ib_connection *ic = conn->c_transport_data;
430
431 if (posted == 0)
432 return;
433
434 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
435
436 /* Decide whether to send an update to the peer now.
437 * If we would send a credit update for every single buffer we
438 * post, we would end up with an ACK storm (ACK arrives,
439 * consumes buffer, we refill the ring, send ACK to remote
440 * advertising the newly posted buffer... ad inf)
441 *
442 * Performance pretty much depends on how often we send
443 * credit updates - too frequent updates mean lots of ACKs.
444 * Too infrequent updates, and the peer will run out of
445 * credits and has to throttle.
446 * For the time being, 16 seems to be a good compromise.
447 */
448 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
449 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
450}
451
452static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
453 struct rds_ib_send_work *send,
454 bool notify)
455{
456 /*
457 * We want to delay signaling completions just enough to get
458 * the batching benefits but not so much that we create dead time
459 * on the wire.
460 */
461 if (ic->i_unsignaled_wrs-- == 0 || notify) {
462 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
463 send->s_wr.send_flags |= IB_SEND_SIGNALED;
464 return 1;
465 }
466 return 0;
467}
468
469/*
470 * This can be called multiple times for a given message. The first time
471 * we see a message we map its scatterlist into the IB device so that
472 * we can provide that mapped address to the IB scatter gather entries
473 * in the IB work requests. We translate the scatterlist into a series
474 * of work requests that fragment the message. These work requests complete
475 * in order so we pass ownership of the message to the completion handler
476 * once we send the final fragment.
477 *
478 * The RDS core uses the c_send_lock to only enter this function once
479 * per connection. This makes sure that the tx ring alloc/unalloc pairs
480 * don't get out of sync and confuse the ring.
481 */
482int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
483 unsigned int hdr_off, unsigned int sg, unsigned int off)
484{
485 struct rds_ib_connection *ic = conn->c_transport_data;
486 struct ib_device *dev = ic->i_cm_id->device;
487 struct rds_ib_send_work *send = NULL;
488 struct rds_ib_send_work *first;
489 struct rds_ib_send_work *prev;
490 struct ib_send_wr *failed_wr;
491 struct scatterlist *scat;
492 u32 pos;
493 u32 i;
494 u32 work_alloc;
495 u32 credit_alloc = 0;
496 u32 posted;
497 u32 adv_credits = 0;
498 int send_flags = 0;
499 int bytes_sent = 0;
500 int ret;
501 int flow_controlled = 0;
502 int nr_sig = 0;
503
504 BUG_ON(off % RDS_FRAG_SIZE);
505 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
506
507 /* Do not send cong updates to IB loopback */
508 if (conn->c_loopback
509 && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
510 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
511 scat = &rm->data.op_sg[sg];
512 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
513 return sizeof(struct rds_header) + ret;
514 }
515
516 /* FIXME we may overallocate here */
517 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
518 i = 1;
519 else
520 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
521
522 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
523 if (work_alloc == 0) {
524 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
525 rds_ib_stats_inc(s_ib_tx_ring_full);
526 ret = -ENOMEM;
527 goto out;
528 }
529
530 if (ic->i_flowctl) {
531 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
532 adv_credits += posted;
533 if (credit_alloc < work_alloc) {
534 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
535 work_alloc = credit_alloc;
536 flow_controlled = 1;
537 }
538 if (work_alloc == 0) {
539 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
540 rds_ib_stats_inc(s_ib_tx_throttle);
541 ret = -ENOMEM;
542 goto out;
543 }
544 }
545
546 /* map the message the first time we see it */
547 if (!ic->i_data_op) {
548 if (rm->data.op_nents) {
549 rm->data.op_count = ib_dma_map_sg(dev,
550 rm->data.op_sg,
551 rm->data.op_nents,
552 DMA_TO_DEVICE);
553 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
554 if (rm->data.op_count == 0) {
555 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
556 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
557 ret = -ENOMEM; /* XXX ? */
558 goto out;
559 }
560 } else {
561 rm->data.op_count = 0;
562 }
563
564 rds_message_addref(rm);
565 rm->data.op_dmasg = 0;
566 rm->data.op_dmaoff = 0;
567 ic->i_data_op = &rm->data;
568
569 /* Finalize the header */
570 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
571 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
572 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
573 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
574
575 /* If it has a RDMA op, tell the peer we did it. This is
576 * used by the peer to release use-once RDMA MRs. */
577 if (rm->rdma.op_active) {
578 struct rds_ext_header_rdma ext_hdr;
579
580 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
581 rds_message_add_extension(&rm->m_inc.i_hdr,
582 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
583 }
584 if (rm->m_rdma_cookie) {
585 rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
586 rds_rdma_cookie_key(rm->m_rdma_cookie),
587 rds_rdma_cookie_offset(rm->m_rdma_cookie));
588 }
589
590 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
591 * we should not do this unless we have a chance of at least
592 * sticking the header into the send ring. Which is why we
593 * should call rds_ib_ring_alloc first. */
594 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
595 rds_message_make_checksum(&rm->m_inc.i_hdr);
596
597 /*
598 * Update adv_credits since we reset the ACK_REQUIRED bit.
599 */
600 if (ic->i_flowctl) {
601 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
602 adv_credits += posted;
603 BUG_ON(adv_credits > 255);
604 }
605 }
606
607 /* Sometimes you want to put a fence between an RDMA
608 * READ and the following SEND.
609 * We could either do this all the time
610 * or when requested by the user. Right now, we let
611 * the application choose.
612 */
613 if (rm->rdma.op_active && rm->rdma.op_fence)
614 send_flags = IB_SEND_FENCE;
615
616 /* Each frag gets a header. Msgs may be 0 bytes */
617 send = &ic->i_sends[pos];
618 first = send;
619 prev = NULL;
620 scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
621 i = 0;
622 do {
623 unsigned int len = 0;
624
625 /* Set up the header */
626 send->s_wr.send_flags = send_flags;
627 send->s_wr.opcode = IB_WR_SEND;
628 send->s_wr.num_sge = 1;
629 send->s_wr.next = NULL;
630 send->s_queued = jiffies;
631 send->s_op = NULL;
632
633 send->s_sge[0].addr = ic->i_send_hdrs_dma
634 + (pos * sizeof(struct rds_header));
635 send->s_sge[0].length = sizeof(struct rds_header);
636
637 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
638
639 /* Set up the data, if present */
640 if (i < work_alloc
641 && scat != &rm->data.op_sg[rm->data.op_count]) {
642 len = min(RDS_FRAG_SIZE,
643 ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
644 send->s_wr.num_sge = 2;
645
646 send->s_sge[1].addr = ib_sg_dma_address(dev, scat);
647 send->s_sge[1].addr += rm->data.op_dmaoff;
648 send->s_sge[1].length = len;
649
650 bytes_sent += len;
651 rm->data.op_dmaoff += len;
652 if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
653 scat++;
654 rm->data.op_dmasg++;
655 rm->data.op_dmaoff = 0;
656 }
657 }
658
659 rds_ib_set_wr_signal_state(ic, send, 0);
660
661 /*
662 * Always signal the last one if we're stopping due to flow control.
663 */
664 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
665 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
666
667 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
668 nr_sig++;
669
670 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
671 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
672
673 if (ic->i_flowctl && adv_credits) {
674 struct rds_header *hdr = &ic->i_send_hdrs[pos];
675
676 /* add credit and redo the header checksum */
677 hdr->h_credit = adv_credits;
678 rds_message_make_checksum(hdr);
679 adv_credits = 0;
680 rds_ib_stats_inc(s_ib_tx_credit_updates);
681 }
682
683 if (prev)
684 prev->s_wr.next = &send->s_wr;
685 prev = send;
686
687 pos = (pos + 1) % ic->i_send_ring.w_nr;
688 send = &ic->i_sends[pos];
689 i++;
690
691 } while (i < work_alloc
692 && scat != &rm->data.op_sg[rm->data.op_count]);
693
694 /* Account the RDS header in the number of bytes we sent, but just once.
695 * The caller has no concept of fragmentation. */
696 if (hdr_off == 0)
697 bytes_sent += sizeof(struct rds_header);
698
699 /* if we finished the message then send completion owns it */
700 if (scat == &rm->data.op_sg[rm->data.op_count]) {
701 prev->s_op = ic->i_data_op;
702 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
703 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) {
704 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
705 prev->s_wr.send_flags |= IB_SEND_SIGNALED;
706 nr_sig++;
707 }
708 ic->i_data_op = NULL;
709 }
710
711 /* Put back wrs & credits we didn't use */
712 if (i < work_alloc) {
713 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
714 work_alloc = i;
715 }
716 if (ic->i_flowctl && i < credit_alloc)
717 rds_ib_send_add_credits(conn, credit_alloc - i);
718
719 if (nr_sig)
720 atomic_add(nr_sig, &ic->i_signaled_sends);
721
722 /* XXX need to worry about failed_wr and partial sends. */
723 failed_wr = &first->s_wr;
724 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
725 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
726 first, &first->s_wr, ret, failed_wr);
727 BUG_ON(failed_wr != &first->s_wr);
728 if (ret) {
729 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
730 "returned %d\n", &conn->c_faddr, ret);
731 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
732 rds_ib_sub_signaled(ic, nr_sig);
733 if (prev->s_op) {
734 ic->i_data_op = prev->s_op;
735 prev->s_op = NULL;
736 }
737
738 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
739 goto out;
740 }
741
742 ret = bytes_sent;
743out:
744 BUG_ON(adv_credits);
745 return ret;
746}
747
748/*
749 * Issue atomic operation.
750 * A simplified version of the rdma case, we always map 1 SG, and
751 * only 8 bytes, for the return value from the atomic operation.
752 */
753int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
754{
755 struct rds_ib_connection *ic = conn->c_transport_data;
756 struct rds_ib_send_work *send = NULL;
757 struct ib_send_wr *failed_wr;
758 struct rds_ib_device *rds_ibdev;
759 u32 pos;
760 u32 work_alloc;
761 int ret;
762 int nr_sig = 0;
763
764 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
765
766 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
767 if (work_alloc != 1) {
768 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
769 rds_ib_stats_inc(s_ib_tx_ring_full);
770 ret = -ENOMEM;
771 goto out;
772 }
773
774 /* address of send request in ring */
775 send = &ic->i_sends[pos];
776 send->s_queued = jiffies;
777
778 if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
779 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
780 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
781 send->s_atomic_wr.swap = op->op_m_cswp.swap;
782 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
783 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
784 } else { /* FADD */
785 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
786 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
787 send->s_atomic_wr.swap = 0;
788 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
789 send->s_atomic_wr.swap_mask = 0;
790 }
791 nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
792 send->s_atomic_wr.wr.num_sge = 1;
793 send->s_atomic_wr.wr.next = NULL;
794 send->s_atomic_wr.remote_addr = op->op_remote_addr;
795 send->s_atomic_wr.rkey = op->op_rkey;
796 send->s_op = op;
797 rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
798
799 /* map 8 byte retval buffer to the device */
800 ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
801 rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
802 if (ret != 1) {
803 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
804 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
805 ret = -ENOMEM; /* XXX ? */
806 goto out;
807 }
808
809 /* Convert our struct scatterlist to struct ib_sge */
810 send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
811 send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
812 send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
813
814 rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
815 send->s_sge[0].addr, send->s_sge[0].length);
816
817 if (nr_sig)
818 atomic_add(nr_sig, &ic->i_signaled_sends);
819
820 failed_wr = &send->s_atomic_wr.wr;
821 ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
822 rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
823 send, &send->s_atomic_wr, ret, failed_wr);
824 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
825 if (ret) {
826 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
827 "returned %d\n", &conn->c_faddr, ret);
828 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
829 rds_ib_sub_signaled(ic, nr_sig);
830 goto out;
831 }
832
833 if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
834 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
835 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
836 }
837
838out:
839 return ret;
840}
841
842int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
843{
844 struct rds_ib_connection *ic = conn->c_transport_data;
845 struct rds_ib_send_work *send = NULL;
846 struct rds_ib_send_work *first;
847 struct rds_ib_send_work *prev;
848 struct ib_send_wr *failed_wr;
849 struct scatterlist *scat;
850 unsigned long len;
851 u64 remote_addr = op->op_remote_addr;
852 u32 max_sge = ic->rds_ibdev->max_sge;
853 u32 pos;
854 u32 work_alloc;
855 u32 i;
856 u32 j;
857 int sent;
858 int ret;
859 int num_sge;
860 int nr_sig = 0;
861
862 /* map the op the first time we see it */
863 if (!op->op_mapped) {
864 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
865 op->op_sg, op->op_nents, (op->op_write) ?
866 DMA_TO_DEVICE : DMA_FROM_DEVICE);
867 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
868 if (op->op_count == 0) {
869 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
870 ret = -ENOMEM; /* XXX ? */
871 goto out;
872 }
873
874 op->op_mapped = 1;
875 }
876
877 /*
878 * Instead of knowing how to return a partial rdma read/write we insist that there
879 * be enough work requests to send the entire message.
880 */
881 i = ceil(op->op_count, max_sge);
882
883 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
884 if (work_alloc != i) {
885 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
886 rds_ib_stats_inc(s_ib_tx_ring_full);
887 ret = -ENOMEM;
888 goto out;
889 }
890
891 send = &ic->i_sends[pos];
892 first = send;
893 prev = NULL;
894 scat = &op->op_sg[0];
895 sent = 0;
896 num_sge = op->op_count;
897
898 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
899 send->s_wr.send_flags = 0;
900 send->s_queued = jiffies;
901 send->s_op = NULL;
902
903 nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
904
905 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
906 send->s_rdma_wr.remote_addr = remote_addr;
907 send->s_rdma_wr.rkey = op->op_rkey;
908
909 if (num_sge > max_sge) {
910 send->s_rdma_wr.wr.num_sge = max_sge;
911 num_sge -= max_sge;
912 } else {
913 send->s_rdma_wr.wr.num_sge = num_sge;
914 }
915
916 send->s_rdma_wr.wr.next = NULL;
917
918 if (prev)
919 prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
920
921 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
922 scat != &op->op_sg[op->op_count]; j++) {
923 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
924 send->s_sge[j].addr =
925 ib_sg_dma_address(ic->i_cm_id->device, scat);
926 send->s_sge[j].length = len;
927 send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
928
929 sent += len;
930 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
931
932 remote_addr += len;
933 scat++;
934 }
935
936 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
937 &send->s_rdma_wr.wr,
938 send->s_rdma_wr.wr.num_sge,
939 send->s_rdma_wr.wr.next);
940
941 prev = send;
942 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
943 send = ic->i_sends;
944 }
945
946 /* give a reference to the last op */
947 if (scat == &op->op_sg[op->op_count]) {
948 prev->s_op = op;
949 rds_message_addref(container_of(op, struct rds_message, rdma));
950 }
951
952 if (i < work_alloc) {
953 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
954 work_alloc = i;
955 }
956
957 if (nr_sig)
958 atomic_add(nr_sig, &ic->i_signaled_sends);
959
960 failed_wr = &first->s_rdma_wr.wr;
961 ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
962 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
963 first, &first->s_rdma_wr.wr, ret, failed_wr);
964 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
965 if (ret) {
966 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
967 "returned %d\n", &conn->c_faddr, ret);
968 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
969 rds_ib_sub_signaled(ic, nr_sig);
970 goto out;
971 }
972
973 if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
974 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
975 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
976 }
977
978
979out:
980 return ret;
981}
982
983void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
984{
985 struct rds_connection *conn = cp->cp_conn;
986 struct rds_ib_connection *ic = conn->c_transport_data;
987
988 /* We may have a pending ACK or window update we were unable
989 * to send previously (due to flow control). Try again. */
990 rds_ib_attempt_ack(ic);
991}