projects/linux·blob master · refs

1045 lines · 30.9 KiB

1/*
2 * Copyright (c) 2006, 2019 Oracle and/or its affiliates. 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#include "ib_mr.h"
43
44/*
45 * Convert IB-specific error message to RDS error message and call core
46 * completion handler.
47 */
48static void rds_ib_send_complete(struct rds_message *rm,
49 int wc_status,
50 void (*complete)(struct rds_message *rm, int status))
51{
52 int notify_status;
53
54 switch (wc_status) {
55 case IB_WC_WR_FLUSH_ERR:
56 return;
57
58 case IB_WC_SUCCESS:
59 notify_status = RDS_RDMA_SUCCESS;
60 break;
61
62 case IB_WC_REM_ACCESS_ERR:
63 notify_status = RDS_RDMA_REMOTE_ERROR;
64 break;
65
66 default:
67 notify_status = RDS_RDMA_OTHER_ERROR;
68 break;
69 }
70 complete(rm, notify_status);
71}
72
73static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
74 struct rm_data_op *op,
75 int wc_status)
76{
77 if (op->op_nents)
78 ib_dma_unmap_sg(ic->i_cm_id->device,
79 op->op_sg, op->op_nents,
80 DMA_TO_DEVICE);
81}
82
83static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
84 struct rm_rdma_op *op,
85 int wc_status)
86{
87 if (op->op_mapped) {
88 ib_dma_unmap_sg(ic->i_cm_id->device,
89 op->op_sg, op->op_nents,
90 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
91 op->op_mapped = 0;
92 }
93
94 /* If the user asked for a completion notification on this
95 * message, we can implement three different semantics:
96 * 1. Notify when we received the ACK on the RDS message
97 * that was queued with the RDMA. This provides reliable
98 * notification of RDMA status at the expense of a one-way
99 * packet delay.
100 * 2. Notify when the IB stack gives us the completion event for
101 * the RDMA operation.
102 * 3. Notify when the IB stack gives us the completion event for
103 * the accompanying RDS messages.
104 * Here, we implement approach #3. To implement approach #2,
105 * we would need to take an event for the rdma WR. To implement #1,
106 * don't call rds_rdma_send_complete at all, and fall back to the notify
107 * handling in the ACK processing code.
108 *
109 * Note: There's no need to explicitly sync any RDMA buffers using
110 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
111 * operation itself unmapped the RDMA buffers, which takes care
112 * of synching.
113 */
114 rds_ib_send_complete(container_of(op, struct rds_message, rdma),
115 wc_status, rds_rdma_send_complete);
116
117 if (op->op_write)
118 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
119 else
120 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
121}
122
123static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
124 struct rm_atomic_op *op,
125 int wc_status)
126{
127 /* unmap atomic recvbuf */
128 if (op->op_mapped) {
129 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
130 DMA_FROM_DEVICE);
131 op->op_mapped = 0;
132 }
133
134 rds_ib_send_complete(container_of(op, struct rds_message, atomic),
135 wc_status, rds_atomic_send_complete);
136
137 if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
138 rds_ib_stats_inc(s_ib_atomic_cswp);
139 else
140 rds_ib_stats_inc(s_ib_atomic_fadd);
141}
142
143/*
144 * Unmap the resources associated with a struct send_work.
145 *
146 * Returns the rm for no good reason other than it is unobtainable
147 * other than by switching on wr.opcode, currently, and the caller,
148 * the event handler, needs it.
149 */
150static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
151 struct rds_ib_send_work *send,
152 int wc_status)
153{
154 struct rds_message *rm = NULL;
155
156 /* In the error case, wc.opcode sometimes contains garbage */
157 switch (send->s_wr.opcode) {
158 case IB_WR_SEND:
159 if (send->s_op) {
160 rm = container_of(send->s_op, struct rds_message, data);
161 rds_ib_send_unmap_data(ic, send->s_op, wc_status);
162 }
163 break;
164 case IB_WR_RDMA_WRITE:
165 case IB_WR_RDMA_READ:
166 if (send->s_op) {
167 rm = container_of(send->s_op, struct rds_message, rdma);
168 rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
169 }
170 break;
171 case IB_WR_ATOMIC_FETCH_AND_ADD:
172 case IB_WR_ATOMIC_CMP_AND_SWP:
173 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
174 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
175 if (send->s_op) {
176 rm = container_of(send->s_op, struct rds_message, atomic);
177 rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
178 }
179 break;
180 default:
181 printk_ratelimited(KERN_NOTICE
182 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
183 __func__, send->s_wr.opcode);
184 break;
185 }
186
187 send->s_wr.opcode = 0xdead;
188
189 return rm;
190}
191
192void rds_ib_send_init_ring(struct rds_ib_connection *ic)
193{
194 struct rds_ib_send_work *send;
195 u32 i;
196
197 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
198 struct ib_sge *sge;
199
200 send->s_op = NULL;
201
202 send->s_wr.wr_id = i;
203 send->s_wr.sg_list = send->s_sge;
204 send->s_wr.ex.imm_data = 0;
205
206 sge = &send->s_sge[0];
207 sge->addr = ic->i_send_hdrs_dma[i];
208
209 sge->length = sizeof(struct rds_header);
210 sge->lkey = ic->i_pd->local_dma_lkey;
211
212 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
213 }
214}
215
216void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
217{
218 struct rds_ib_send_work *send;
219 u32 i;
220
221 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
222 if (send->s_op && send->s_wr.opcode != 0xdead)
223 rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
224 }
225}
226
227/*
228 * The only fast path caller always has a non-zero nr, so we don't
229 * bother testing nr before performing the atomic sub.
230 */
231static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
232{
233 if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
234 waitqueue_active(&rds_ib_ring_empty_wait))
235 wake_up(&rds_ib_ring_empty_wait);
236 BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
237}
238
239/*
240 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
241 * operations performed in the send path. As the sender allocs and potentially
242 * unallocs the next free entry in the ring it doesn't alter which is
243 * the next to be freed, which is what this is concerned with.
244 */
245void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
246{
247 struct rds_message *rm = NULL;
248 struct rds_connection *conn = ic->conn;
249 struct rds_ib_send_work *send;
250 u32 completed;
251 u32 oldest;
252 u32 i = 0;
253 int nr_sig = 0;
254
255
256 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
257 (unsigned long long)wc->wr_id, wc->status,
258 ib_wc_status_msg(wc->status), wc->byte_len,
259 be32_to_cpu(wc->ex.imm_data));
260 rds_ib_stats_inc(s_ib_tx_cq_event);
261
262 if (wc->wr_id == RDS_IB_ACK_WR_ID) {
263 if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
264 rds_ib_stats_inc(s_ib_tx_stalled);
265 rds_ib_ack_send_complete(ic);
266 return;
267 }
268
269 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
270
271 completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
272
273 for (i = 0; i < completed; i++) {
274 send = &ic->i_sends[oldest];
275 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
276 nr_sig++;
277
278 rm = rds_ib_send_unmap_op(ic, send, wc->status);
279
280 if (time_after(jiffies, send->s_queued + HZ / 2))
281 rds_ib_stats_inc(s_ib_tx_stalled);
282
283 if (send->s_op) {
284 if (send->s_op == rm->m_final_op) {
285 /* If anyone waited for this message to get
286 * flushed out, wake them up now
287 */
288 rds_message_unmapped(rm);
289 }
290 rds_message_put(rm);
291 send->s_op = NULL;
292 }
293
294 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
295 }
296
297 rds_ib_ring_free(&ic->i_send_ring, completed);
298 rds_ib_sub_signaled(ic, nr_sig);
299
300 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
301 test_bit(0, &conn->c_map_queued))
302 queue_delayed_work(conn->c_path->cp_wq, &conn->c_send_w, 0);
303
304 /* We expect errors as the qp is drained during shutdown */
305 if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
306 rds_ib_conn_error(conn, "send completion on <%pI6c,%pI6c,%d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
307 &conn->c_laddr, &conn->c_faddr,
308 conn->c_tos, wc->status,
309 ib_wc_status_msg(wc->status), wc->vendor_err);
310 }
311}
312
313/*
314 * This is the main function for allocating credits when sending
315 * messages.
316 *
317 * Conceptually, we have two counters:
318 * - send credits: this tells us how many WRs we're allowed
319 * to submit without overruning the receiver's queue. For
320 * each SEND WR we post, we decrement this by one.
321 *
322 * - posted credits: this tells us how many WRs we recently
323 * posted to the receive queue. This value is transferred
324 * to the peer as a "credit update" in a RDS header field.
325 * Every time we transmit credits to the peer, we subtract
326 * the amount of transferred credits from this counter.
327 *
328 * It is essential that we avoid situations where both sides have
329 * exhausted their send credits, and are unable to send new credits
330 * to the peer. We achieve this by requiring that we send at least
331 * one credit update to the peer before exhausting our credits.
332 * When new credits arrive, we subtract one credit that is withheld
333 * until we've posted new buffers and are ready to transmit these
334 * credits (see rds_ib_send_add_credits below).
335 *
336 * The RDS send code is essentially single-threaded; rds_send_xmit
337 * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
338 * However, the ACK sending code is independent and can race with
339 * message SENDs.
340 *
341 * In the send path, we need to update the counters for send credits
342 * and the counter of posted buffers atomically - when we use the
343 * last available credit, we cannot allow another thread to race us
344 * and grab the posted credits counter. Hence, we have to use a
345 * spinlock to protect the credit counter, or use atomics.
346 *
347 * Spinlocks shared between the send and the receive path are bad,
348 * because they create unnecessary delays. An early implementation
349 * using a spinlock showed a 5% degradation in throughput at some
350 * loads.
351 *
352 * This implementation avoids spinlocks completely, putting both
353 * counters into a single atomic, and updating that atomic using
354 * atomic_add (in the receive path, when receiving fresh credits),
355 * and using atomic_cmpxchg when updating the two counters.
356 */
357int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
358 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
359{
360 unsigned int avail, posted, got = 0, advertise;
361 long oldval, newval;
362
363 *adv_credits = 0;
364 if (!ic->i_flowctl)
365 return wanted;
366
367try_again:
368 advertise = 0;
369 oldval = newval = atomic_read(&ic->i_credits);
370 posted = IB_GET_POST_CREDITS(oldval);
371 avail = IB_GET_SEND_CREDITS(oldval);
372
373 rdsdebug("wanted=%u credits=%u posted=%u\n",
374 wanted, avail, posted);
375
376 /* The last credit must be used to send a credit update. */
377 if (avail && !posted)
378 avail--;
379
380 if (avail < wanted) {
381 struct rds_connection *conn = ic->i_cm_id->context;
382
383 /* Oops, there aren't that many credits left! */
384 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
385 got = avail;
386 } else {
387 /* Sometimes you get what you want, lalala. */
388 got = wanted;
389 }
390 newval -= IB_SET_SEND_CREDITS(got);
391
392 /*
393 * If need_posted is non-zero, then the caller wants
394 * the posted regardless of whether any send credits are
395 * available.
396 */
397 if (posted && (got || need_posted)) {
398 advertise = min_t(unsigned int, posted, max_posted);
399 newval -= IB_SET_POST_CREDITS(advertise);
400 }
401
402 /* Finally bill everything */
403 if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
404 goto try_again;
405
406 *adv_credits = advertise;
407 return got;
408}
409
410void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
411{
412 struct rds_ib_connection *ic = conn->c_transport_data;
413
414 if (credits == 0)
415 return;
416
417 rdsdebug("credits=%u current=%u%s\n",
418 credits,
419 IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
420 test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
421
422 atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
423 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
424 queue_delayed_work(conn->c_path->cp_wq, &conn->c_send_w, 0);
425
426 WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
427
428 rds_ib_stats_inc(s_ib_rx_credit_updates);
429}
430
431void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
432{
433 struct rds_ib_connection *ic = conn->c_transport_data;
434
435 if (posted == 0)
436 return;
437
438 atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
439
440 /* Decide whether to send an update to the peer now.
441 * If we would send a credit update for every single buffer we
442 * post, we would end up with an ACK storm (ACK arrives,
443 * consumes buffer, we refill the ring, send ACK to remote
444 * advertising the newly posted buffer... ad inf)
445 *
446 * Performance pretty much depends on how often we send
447 * credit updates - too frequent updates mean lots of ACKs.
448 * Too infrequent updates, and the peer will run out of
449 * credits and has to throttle.
450 * For the time being, 16 seems to be a good compromise.
451 */
452 if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
453 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
454}
455
456static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
457 struct rds_ib_send_work *send,
458 bool notify)
459{
460 /*
461 * We want to delay signaling completions just enough to get
462 * the batching benefits but not so much that we create dead time
463 * on the wire.
464 */
465 if (ic->i_unsignaled_wrs-- == 0 || notify) {
466 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
467 send->s_wr.send_flags |= IB_SEND_SIGNALED;
468 return 1;
469 }
470 return 0;
471}
472
473/*
474 * This can be called multiple times for a given message. The first time
475 * we see a message we map its scatterlist into the IB device so that
476 * we can provide that mapped address to the IB scatter gather entries
477 * in the IB work requests. We translate the scatterlist into a series
478 * of work requests that fragment the message. These work requests complete
479 * in order so we pass ownership of the message to the completion handler
480 * once we send the final fragment.
481 *
482 * The RDS core uses the c_send_lock to only enter this function once
483 * per connection. This makes sure that the tx ring alloc/unalloc pairs
484 * don't get out of sync and confuse the ring.
485 */
486int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
487 unsigned int hdr_off, unsigned int sg, unsigned int off)
488{
489 struct rds_ib_connection *ic = conn->c_transport_data;
490 struct ib_device *dev = ic->i_cm_id->device;
491 struct rds_ib_send_work *send = NULL;
492 struct rds_ib_send_work *first;
493 struct rds_ib_send_work *prev;
494 const struct ib_send_wr *failed_wr;
495 struct scatterlist *scat;
496 u32 pos;
497 u32 i;
498 u32 work_alloc;
499 u32 credit_alloc = 0;
500 u32 posted;
501 u32 adv_credits = 0;
502 int send_flags = 0;
503 int bytes_sent = 0;
504 int ret;
505 int flow_controlled = 0;
506 int nr_sig = 0;
507
508 BUG_ON(off % RDS_FRAG_SIZE);
509 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
510
511 /* Do not send cong updates to IB loopback */
512 if (conn->c_loopback
513 && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
514 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
515 scat = &rm->data.op_sg[sg];
516 ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
517 return sizeof(struct rds_header) + ret;
518 }
519
520 /* FIXME we may overallocate here */
521 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
522 i = 1;
523 else
524 i = DIV_ROUND_UP(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
525
526 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
527 if (work_alloc == 0) {
528 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
529 rds_ib_stats_inc(s_ib_tx_ring_full);
530 ret = -ENOMEM;
531 goto out;
532 }
533
534 if (ic->i_flowctl) {
535 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
536 adv_credits += posted;
537 if (credit_alloc < work_alloc) {
538 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
539 work_alloc = credit_alloc;
540 flow_controlled = 1;
541 }
542 if (work_alloc == 0) {
543 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
544 rds_ib_stats_inc(s_ib_tx_throttle);
545 ret = -ENOMEM;
546 goto out;
547 }
548 }
549
550 /* map the message the first time we see it */
551 if (!ic->i_data_op) {
552 if (rm->data.op_nents) {
553 rm->data.op_count = ib_dma_map_sg(dev,
554 rm->data.op_sg,
555 rm->data.op_nents,
556 DMA_TO_DEVICE);
557 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
558 if (rm->data.op_count == 0) {
559 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
560 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
561 ret = -ENOMEM; /* XXX ? */
562 goto out;
563 }
564 } else {
565 rm->data.op_count = 0;
566 }
567
568 rds_message_addref(rm);
569 rm->data.op_dmasg = 0;
570 rm->data.op_dmaoff = 0;
571 ic->i_data_op = &rm->data;
572
573 /* Finalize the header */
574 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
575 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
576 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
577 rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
578
579 /* If it has a RDMA op, tell the peer we did it. This is
580 * used by the peer to release use-once RDMA MRs. */
581 if (rm->rdma.op_active) {
582 struct rds_ext_header_rdma ext_hdr = {};
583 struct rds_ext_header_rdma_bytes
584 rdma_bytes_ext_hdr = {};
585
586 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
587 if (rds_message_add_extension(&rm->m_inc.i_hdr,
588 RDS_EXTHDR_RDMA,
589 &ext_hdr)) {
590 /* prepare the rdma bytes ext header */
591 rdma_bytes_ext_hdr.h_rflags =
592 rm->rdma.op_write ?
593 RDS_FLAG_RDMA_WR_BYTES :
594 RDS_FLAG_RDMA_RD_BYTES;
595 rdma_bytes_ext_hdr.h_rdma_bytes =
596 cpu_to_be32(rm->rdma.op_bytes);
597 } else {
598 rdsdebug("RDS_EXTHDR_RDMA dropped");
599 }
600
601 if (rds_message_add_extension(&rm->m_inc.i_hdr,
602 RDS_EXTHDR_RDMA_BYTES,
603 &rdma_bytes_ext_hdr)) {
604 /* rdma bytes ext header was added successfully,
605 * notify the remote side via flag in header
606 */
607 rm->m_inc.i_hdr.h_flags |=
608 RDS_FLAG_EXTHDR_EXTENSION;
609 } else {
610 rdsdebug("RDS_EXTHDR_RDMA_BYTES dropped");
611 }
612 }
613 if (rm->m_rdma_cookie &&
614 !rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
615 rds_rdma_cookie_key(rm->m_rdma_cookie),
616 rds_rdma_cookie_offset(rm->m_rdma_cookie))) {
617 rdsdebug("RDS_EXTHDR_RDMA_DEST dropped\n");
618 }
619
620 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
621 * we should not do this unless we have a chance of at least
622 * sticking the header into the send ring. Which is why we
623 * should call rds_ib_ring_alloc first. */
624 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
625 rds_message_make_checksum(&rm->m_inc.i_hdr);
626
627 /*
628 * Update adv_credits since we reset the ACK_REQUIRED bit.
629 */
630 if (ic->i_flowctl) {
631 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
632 adv_credits += posted;
633 BUG_ON(adv_credits > 255);
634 }
635 }
636
637 /* Sometimes you want to put a fence between an RDMA
638 * READ and the following SEND.
639 * We could either do this all the time
640 * or when requested by the user. Right now, we let
641 * the application choose.
642 */
643 if (rm->rdma.op_active && rm->rdma.op_fence)
644 send_flags = IB_SEND_FENCE;
645
646 /* Each frag gets a header. Msgs may be 0 bytes */
647 send = &ic->i_sends[pos];
648 first = send;
649 prev = NULL;
650 scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
651 i = 0;
652 do {
653 unsigned int len = 0;
654
655 /* Set up the header */
656 send->s_wr.send_flags = send_flags;
657 send->s_wr.opcode = IB_WR_SEND;
658 send->s_wr.num_sge = 1;
659 send->s_wr.next = NULL;
660 send->s_queued = jiffies;
661 send->s_op = NULL;
662
663 send->s_sge[0].addr = ic->i_send_hdrs_dma[pos];
664
665 send->s_sge[0].length = sizeof(struct rds_header);
666 send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
667
668 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev,
669 ic->i_send_hdrs_dma[pos],
670 sizeof(struct rds_header),
671 DMA_TO_DEVICE);
672 memcpy(ic->i_send_hdrs[pos], &rm->m_inc.i_hdr,
673 sizeof(struct rds_header));
674
675
676 /* Set up the data, if present */
677 if (i < work_alloc
678 && scat != &rm->data.op_sg[rm->data.op_count]) {
679 len = min(RDS_FRAG_SIZE,
680 sg_dma_len(scat) - rm->data.op_dmaoff);
681 send->s_wr.num_sge = 2;
682
683 send->s_sge[1].addr = sg_dma_address(scat);
684 send->s_sge[1].addr += rm->data.op_dmaoff;
685 send->s_sge[1].length = len;
686 send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
687
688 bytes_sent += len;
689 rm->data.op_dmaoff += len;
690 if (rm->data.op_dmaoff == sg_dma_len(scat)) {
691 scat++;
692 rm->data.op_dmasg++;
693 rm->data.op_dmaoff = 0;
694 }
695 }
696
697 rds_ib_set_wr_signal_state(ic, send, false);
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 rds_ib_set_wr_signal_state(ic, send, true);
704 send->s_wr.send_flags |= IB_SEND_SOLICITED;
705 }
706
707 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
708 nr_sig++;
709
710 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
711 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
712
713 if (ic->i_flowctl && adv_credits) {
714 struct rds_header *hdr = ic->i_send_hdrs[pos];
715
716 /* add credit and redo the header checksum */
717 hdr->h_credit = adv_credits;
718 rds_message_make_checksum(hdr);
719 adv_credits = 0;
720 rds_ib_stats_inc(s_ib_tx_credit_updates);
721 }
722 ib_dma_sync_single_for_device(ic->rds_ibdev->dev,
723 ic->i_send_hdrs_dma[pos],
724 sizeof(struct rds_header),
725 DMA_TO_DEVICE);
726
727 if (prev)
728 prev->s_wr.next = &send->s_wr;
729 prev = send;
730
731 pos = (pos + 1) % ic->i_send_ring.w_nr;
732 send = &ic->i_sends[pos];
733 i++;
734
735 } while (i < work_alloc
736 && scat != &rm->data.op_sg[rm->data.op_count]);
737
738 /* Account the RDS header in the number of bytes we sent, but just once.
739 * The caller has no concept of fragmentation. */
740 if (hdr_off == 0)
741 bytes_sent += sizeof(struct rds_header);
742
743 /* if we finished the message then send completion owns it */
744 if (scat == &rm->data.op_sg[rm->data.op_count]) {
745 prev->s_op = ic->i_data_op;
746 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
747 if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED))
748 nr_sig += rds_ib_set_wr_signal_state(ic, prev, true);
749 ic->i_data_op = NULL;
750 }
751
752 /* Put back wrs & credits we didn't use */
753 if (i < work_alloc) {
754 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
755 work_alloc = i;
756 }
757 if (ic->i_flowctl && i < credit_alloc)
758 rds_ib_send_add_credits(conn, credit_alloc - i);
759
760 if (nr_sig)
761 atomic_add(nr_sig, &ic->i_signaled_sends);
762
763 /* XXX need to worry about failed_wr and partial sends. */
764 failed_wr = &first->s_wr;
765 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
766 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
767 first, &first->s_wr, ret, failed_wr);
768 BUG_ON(failed_wr != &first->s_wr);
769 if (ret) {
770 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI6c "
771 "returned %d\n", &conn->c_faddr, ret);
772 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
773 rds_ib_sub_signaled(ic, nr_sig);
774 if (prev->s_op) {
775 ic->i_data_op = prev->s_op;
776 prev->s_op = NULL;
777 }
778
779 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
780 goto out;
781 }
782
783 ret = bytes_sent;
784out:
785 BUG_ON(adv_credits);
786 return ret;
787}
788
789/*
790 * Issue atomic operation.
791 * A simplified version of the rdma case, we always map 1 SG, and
792 * only 8 bytes, for the return value from the atomic operation.
793 */
794int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
795{
796 struct rds_ib_connection *ic = conn->c_transport_data;
797 struct rds_ib_send_work *send = NULL;
798 const struct ib_send_wr *failed_wr;
799 u32 pos;
800 u32 work_alloc;
801 int ret;
802 int nr_sig = 0;
803
804 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
805 if (work_alloc != 1) {
806 rds_ib_stats_inc(s_ib_tx_ring_full);
807 ret = -ENOMEM;
808 goto out;
809 }
810
811 /* address of send request in ring */
812 send = &ic->i_sends[pos];
813 send->s_queued = jiffies;
814
815 if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
816 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
817 send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
818 send->s_atomic_wr.swap = op->op_m_cswp.swap;
819 send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
820 send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
821 } else { /* FADD */
822 send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
823 send->s_atomic_wr.compare_add = op->op_m_fadd.add;
824 send->s_atomic_wr.swap = 0;
825 send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
826 send->s_atomic_wr.swap_mask = 0;
827 }
828 send->s_wr.send_flags = 0;
829 nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
830 send->s_atomic_wr.wr.num_sge = 1;
831 send->s_atomic_wr.wr.next = NULL;
832 send->s_atomic_wr.remote_addr = op->op_remote_addr;
833 send->s_atomic_wr.rkey = op->op_rkey;
834 send->s_op = op;
835 rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
836
837 /* map 8 byte retval buffer to the device */
838 ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
839 rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
840 if (ret != 1) {
841 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
842 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
843 ret = -ENOMEM; /* XXX ? */
844 goto out;
845 }
846
847 /* Convert our struct scatterlist to struct ib_sge */
848 send->s_sge[0].addr = sg_dma_address(op->op_sg);
849 send->s_sge[0].length = sg_dma_len(op->op_sg);
850 send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
851
852 rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
853 send->s_sge[0].addr, send->s_sge[0].length);
854
855 if (nr_sig)
856 atomic_add(nr_sig, &ic->i_signaled_sends);
857
858 failed_wr = &send->s_atomic_wr.wr;
859 ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
860 rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
861 send, &send->s_atomic_wr, ret, failed_wr);
862 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
863 if (ret) {
864 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI6c "
865 "returned %d\n", &conn->c_faddr, ret);
866 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
867 rds_ib_sub_signaled(ic, nr_sig);
868 goto out;
869 }
870
871 if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
872 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
873 BUG_ON(failed_wr != &send->s_atomic_wr.wr);
874 }
875
876out:
877 return ret;
878}
879
880int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
881{
882 struct rds_ib_connection *ic = conn->c_transport_data;
883 struct rds_ib_send_work *send = NULL;
884 struct rds_ib_send_work *first;
885 struct rds_ib_send_work *prev;
886 const struct ib_send_wr *failed_wr;
887 struct scatterlist *scat;
888 unsigned long len;
889 u64 remote_addr = op->op_remote_addr;
890 u32 max_sge = ic->rds_ibdev->max_sge;
891 u32 pos;
892 u32 work_alloc;
893 u32 i;
894 u32 j;
895 int sent;
896 int ret;
897 int num_sge;
898 int nr_sig = 0;
899 u64 odp_addr = op->op_odp_addr;
900 u32 odp_lkey = 0;
901
902 /* map the op the first time we see it */
903 if (!op->op_odp_mr) {
904 if (!op->op_mapped) {
905 op->op_count =
906 ib_dma_map_sg(ic->i_cm_id->device, op->op_sg,
907 op->op_nents,
908 (op->op_write) ? DMA_TO_DEVICE :
909 DMA_FROM_DEVICE);
910 rdsdebug("ic %p mapping op %p: %d\n", ic, op,
911 op->op_count);
912 if (op->op_count == 0) {
913 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
914 ret = -ENOMEM; /* XXX ? */
915 goto out;
916 }
917 op->op_mapped = 1;
918 }
919 } else {
920 op->op_count = op->op_nents;
921 odp_lkey = rds_ib_get_lkey(op->op_odp_mr->r_trans_private);
922 }
923
924 /*
925 * Instead of knowing how to return a partial rdma read/write we insist that there
926 * be enough work requests to send the entire message.
927 */
928 i = DIV_ROUND_UP(op->op_count, max_sge);
929
930 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
931 if (work_alloc != i) {
932 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
933 rds_ib_stats_inc(s_ib_tx_ring_full);
934 ret = -ENOMEM;
935 goto out;
936 }
937
938 send = &ic->i_sends[pos];
939 first = send;
940 prev = NULL;
941 scat = &op->op_sg[0];
942 sent = 0;
943 num_sge = op->op_count;
944
945 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
946 send->s_wr.send_flags = 0;
947 send->s_queued = jiffies;
948 send->s_op = NULL;
949
950 if (!op->op_notify)
951 nr_sig += rds_ib_set_wr_signal_state(ic, send,
952 op->op_notify);
953
954 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
955 send->s_rdma_wr.remote_addr = remote_addr;
956 send->s_rdma_wr.rkey = op->op_rkey;
957
958 if (num_sge > max_sge) {
959 send->s_rdma_wr.wr.num_sge = max_sge;
960 num_sge -= max_sge;
961 } else {
962 send->s_rdma_wr.wr.num_sge = num_sge;
963 }
964
965 send->s_rdma_wr.wr.next = NULL;
966
967 if (prev)
968 prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
969
970 for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
971 scat != &op->op_sg[op->op_count]; j++) {
972 len = sg_dma_len(scat);
973 if (!op->op_odp_mr) {
974 send->s_sge[j].addr = sg_dma_address(scat);
975 send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
976 } else {
977 send->s_sge[j].addr = odp_addr;
978 send->s_sge[j].lkey = odp_lkey;
979 }
980 send->s_sge[j].length = len;
981
982 sent += len;
983 rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
984
985 remote_addr += len;
986 odp_addr += len;
987 scat++;
988 }
989
990 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
991 &send->s_rdma_wr.wr,
992 send->s_rdma_wr.wr.num_sge,
993 send->s_rdma_wr.wr.next);
994
995 prev = send;
996 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
997 send = ic->i_sends;
998 }
999
1000 /* give a reference to the last op */
1001 if (scat == &op->op_sg[op->op_count]) {
1002 prev->s_op = op;
1003 rds_message_addref(container_of(op, struct rds_message, rdma));
1004 }
1005
1006 if (i < work_alloc) {
1007 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
1008 work_alloc = i;
1009 }
1010
1011 if (nr_sig)
1012 atomic_add(nr_sig, &ic->i_signaled_sends);
1013
1014 failed_wr = &first->s_rdma_wr.wr;
1015 ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
1016 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
1017 first, &first->s_rdma_wr.wr, ret, failed_wr);
1018 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
1019 if (ret) {
1020 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI6c "
1021 "returned %d\n", &conn->c_faddr, ret);
1022 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
1023 rds_ib_sub_signaled(ic, nr_sig);
1024 goto out;
1025 }
1026
1027 if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
1028 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1029 BUG_ON(failed_wr != &first->s_rdma_wr.wr);
1030 }
1031
1032
1033out:
1034 return ret;
1035}
1036
1037void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
1038{
1039 struct rds_connection *conn = cp->cp_conn;
1040 struct rds_ib_connection *ic = conn->c_transport_data;
1041
1042 /* We may have a pending ACK or window update we were unable
1043 * to send previously (due to flow control). Try again. */
1044 rds_ib_attempt_ack(ic);
1045}