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/slab.h>
35#include <linux/pci.h>
36#include <linux/dma-mapping.h>
37#include <rdma/rdma_cm.h>
38
39#include "rds.h"
40#include "ib.h"
41
42static struct kmem_cache *rds_ib_incoming_slab;
43static struct kmem_cache *rds_ib_frag_slab;
44static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
45
46void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
47{
48 struct rds_ib_recv_work *recv;
49 u32 i;
50
51 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
52 struct ib_sge *sge;
53
54 recv->r_ibinc = NULL;
55 recv->r_frag = NULL;
56
57 recv->r_wr.next = NULL;
58 recv->r_wr.wr_id = i;
59 recv->r_wr.sg_list = recv->r_sge;
60 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
61
62 sge = &recv->r_sge[0];
63 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
64 sge->length = sizeof(struct rds_header);
65 sge->lkey = ic->i_mr->lkey;
66
67 sge = &recv->r_sge[1];
68 sge->addr = 0;
69 sge->length = RDS_FRAG_SIZE;
70 sge->lkey = ic->i_mr->lkey;
71 }
72}
73
74/*
75 * The entire 'from' list, including the from element itself, is put on
76 * to the tail of the 'to' list.
77 */
78static void list_splice_entire_tail(struct list_head *from,
79 struct list_head *to)
80{
81 struct list_head *from_last = from->prev;
82
83 list_splice_tail(from_last, to);
84 list_add_tail(from_last, to);
85}
86
87static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
88{
89 struct list_head *tmp;
90
91 tmp = xchg(&cache->xfer, NULL);
92 if (tmp) {
93 if (cache->ready)
94 list_splice_entire_tail(tmp, cache->ready);
95 else
96 cache->ready = tmp;
97 }
98}
99
100static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
101{
102 struct rds_ib_cache_head *head;
103 int cpu;
104
105 cache->percpu = alloc_percpu(struct rds_ib_cache_head);
106 if (!cache->percpu)
107 return -ENOMEM;
108
109 for_each_possible_cpu(cpu) {
110 head = per_cpu_ptr(cache->percpu, cpu);
111 head->first = NULL;
112 head->count = 0;
113 }
114 cache->xfer = NULL;
115 cache->ready = NULL;
116
117 return 0;
118}
119
120int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
121{
122 int ret;
123
124 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
125 if (!ret) {
126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
127 if (ret)
128 free_percpu(ic->i_cache_incs.percpu);
129 }
130
131 return ret;
132}
133
134static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
135 struct list_head *caller_list)
136{
137 struct rds_ib_cache_head *head;
138 int cpu;
139
140 for_each_possible_cpu(cpu) {
141 head = per_cpu_ptr(cache->percpu, cpu);
142 if (head->first) {
143 list_splice_entire_tail(head->first, caller_list);
144 head->first = NULL;
145 }
146 }
147
148 if (cache->ready) {
149 list_splice_entire_tail(cache->ready, caller_list);
150 cache->ready = NULL;
151 }
152}
153
154void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
155{
156 struct rds_ib_incoming *inc;
157 struct rds_ib_incoming *inc_tmp;
158 struct rds_page_frag *frag;
159 struct rds_page_frag *frag_tmp;
160 LIST_HEAD(list);
161
162 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
163 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
164 free_percpu(ic->i_cache_incs.percpu);
165
166 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
167 list_del(&inc->ii_cache_entry);
168 WARN_ON(!list_empty(&inc->ii_frags));
169 kmem_cache_free(rds_ib_incoming_slab, inc);
170 }
171
172 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
173 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
174 free_percpu(ic->i_cache_frags.percpu);
175
176 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
177 list_del(&frag->f_cache_entry);
178 WARN_ON(!list_empty(&frag->f_item));
179 kmem_cache_free(rds_ib_frag_slab, frag);
180 }
181}
182
183/* fwd decl */
184static void rds_ib_recv_cache_put(struct list_head *new_item,
185 struct rds_ib_refill_cache *cache);
186static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
187
188
189/* Recycle frag and attached recv buffer f_sg */
190static void rds_ib_frag_free(struct rds_ib_connection *ic,
191 struct rds_page_frag *frag)
192{
193 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
194
195 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
196}
197
198/* Recycle inc after freeing attached frags */
199void rds_ib_inc_free(struct rds_incoming *inc)
200{
201 struct rds_ib_incoming *ibinc;
202 struct rds_page_frag *frag;
203 struct rds_page_frag *pos;
204 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
205
206 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
207
208 /* Free attached frags */
209 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
210 list_del_init(&frag->f_item);
211 rds_ib_frag_free(ic, frag);
212 }
213 BUG_ON(!list_empty(&ibinc->ii_frags));
214
215 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
216 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
217}
218
219static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
220 struct rds_ib_recv_work *recv)
221{
222 if (recv->r_ibinc) {
223 rds_inc_put(&recv->r_ibinc->ii_inc);
224 recv->r_ibinc = NULL;
225 }
226 if (recv->r_frag) {
227 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
228 rds_ib_frag_free(ic, recv->r_frag);
229 recv->r_frag = NULL;
230 }
231}
232
233void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
234{
235 u32 i;
236
237 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
238 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
239}
240
241static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
242 gfp_t slab_mask)
243{
244 struct rds_ib_incoming *ibinc;
245 struct list_head *cache_item;
246 int avail_allocs;
247
248 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
249 if (cache_item) {
250 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
251 } else {
252 avail_allocs = atomic_add_unless(&rds_ib_allocation,
253 1, rds_ib_sysctl_max_recv_allocation);
254 if (!avail_allocs) {
255 rds_ib_stats_inc(s_ib_rx_alloc_limit);
256 return NULL;
257 }
258 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
259 if (!ibinc) {
260 atomic_dec(&rds_ib_allocation);
261 return NULL;
262 }
263 }
264 INIT_LIST_HEAD(&ibinc->ii_frags);
265 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
266
267 return ibinc;
268}
269
270static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
271 gfp_t slab_mask, gfp_t page_mask)
272{
273 struct rds_page_frag *frag;
274 struct list_head *cache_item;
275 int ret;
276
277 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
278 if (cache_item) {
279 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
280 } else {
281 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
282 if (!frag)
283 return NULL;
284
285 sg_init_table(&frag->f_sg, 1);
286 ret = rds_page_remainder_alloc(&frag->f_sg,
287 RDS_FRAG_SIZE, page_mask);
288 if (ret) {
289 kmem_cache_free(rds_ib_frag_slab, frag);
290 return NULL;
291 }
292 }
293
294 INIT_LIST_HEAD(&frag->f_item);
295
296 return frag;
297}
298
299static int rds_ib_recv_refill_one(struct rds_connection *conn,
300 struct rds_ib_recv_work *recv, int prefill)
301{
302 struct rds_ib_connection *ic = conn->c_transport_data;
303 struct ib_sge *sge;
304 int ret = -ENOMEM;
305 gfp_t slab_mask = GFP_NOWAIT;
306 gfp_t page_mask = GFP_NOWAIT;
307
308 if (prefill) {
309 slab_mask = GFP_KERNEL;
310 page_mask = GFP_HIGHUSER;
311 }
312
313 if (!ic->i_cache_incs.ready)
314 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
315 if (!ic->i_cache_frags.ready)
316 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
317
318 /*
319 * ibinc was taken from recv if recv contained the start of a message.
320 * recvs that were continuations will still have this allocated.
321 */
322 if (!recv->r_ibinc) {
323 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
324 if (!recv->r_ibinc)
325 goto out;
326 }
327
328 WARN_ON(recv->r_frag); /* leak! */
329 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
330 if (!recv->r_frag)
331 goto out;
332
333 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
334 1, DMA_FROM_DEVICE);
335 WARN_ON(ret != 1);
336
337 sge = &recv->r_sge[0];
338 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
339 sge->length = sizeof(struct rds_header);
340
341 sge = &recv->r_sge[1];
342 sge->addr = sg_dma_address(&recv->r_frag->f_sg);
343 sge->length = sg_dma_len(&recv->r_frag->f_sg);
344
345 ret = 0;
346out:
347 return ret;
348}
349
350/*
351 * This tries to allocate and post unused work requests after making sure that
352 * they have all the allocations they need to queue received fragments into
353 * sockets.
354 *
355 * -1 is returned if posting fails due to temporary resource exhaustion.
356 */
357void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
358{
359 struct rds_ib_connection *ic = conn->c_transport_data;
360 struct rds_ib_recv_work *recv;
361 struct ib_recv_wr *failed_wr;
362 unsigned int posted = 0;
363 int ret = 0;
364 u32 pos;
365
366 while ((prefill || rds_conn_up(conn)) &&
367 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
368 if (pos >= ic->i_recv_ring.w_nr) {
369 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
370 pos);
371 break;
372 }
373
374 recv = &ic->i_recvs[pos];
375 ret = rds_ib_recv_refill_one(conn, recv, prefill);
376 if (ret) {
377 break;
378 }
379
380 /* XXX when can this fail? */
381 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
382 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
383 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
384 (long) sg_dma_address(&recv->r_frag->f_sg), ret);
385 if (ret) {
386 rds_ib_conn_error(conn, "recv post on "
387 "%pI4 returned %d, disconnecting and "
388 "reconnecting\n", &conn->c_faddr,
389 ret);
390 break;
391 }
392
393 posted++;
394 }
395
396 /* We're doing flow control - update the window. */
397 if (ic->i_flowctl && posted)
398 rds_ib_advertise_credits(conn, posted);
399
400 if (ret)
401 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
402}
403
404/*
405 * We want to recycle several types of recv allocations, like incs and frags.
406 * To use this, the *_free() function passes in the ptr to a list_head within
407 * the recyclee, as well as the cache to put it on.
408 *
409 * First, we put the memory on a percpu list. When this reaches a certain size,
410 * We move it to an intermediate non-percpu list in a lockless manner, with some
411 * xchg/compxchg wizardry.
412 *
413 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
414 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
415 * list_empty() will return true with one element is actually present.
416 */
417static void rds_ib_recv_cache_put(struct list_head *new_item,
418 struct rds_ib_refill_cache *cache)
419{
420 unsigned long flags;
421 struct rds_ib_cache_head *chp;
422 struct list_head *old;
423
424 local_irq_save(flags);
425
426 chp = per_cpu_ptr(cache->percpu, smp_processor_id());
427 if (!chp->first)
428 INIT_LIST_HEAD(new_item);
429 else /* put on front */
430 list_add_tail(new_item, chp->first);
431 chp->first = new_item;
432 chp->count++;
433
434 if (chp->count < RDS_IB_RECYCLE_BATCH_COUNT)
435 goto end;
436
437 /*
438 * Return our per-cpu first list to the cache's xfer by atomically
439 * grabbing the current xfer list, appending it to our per-cpu list,
440 * and then atomically returning that entire list back to the
441 * cache's xfer list as long as it's still empty.
442 */
443 do {
444 old = xchg(&cache->xfer, NULL);
445 if (old)
446 list_splice_entire_tail(old, chp->first);
447 old = cmpxchg(&cache->xfer, NULL, chp->first);
448 } while (old);
449
450 chp->first = NULL;
451 chp->count = 0;
452end:
453 local_irq_restore(flags);
454}
455
456static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
457{
458 struct list_head *head = cache->ready;
459
460 if (head) {
461 if (!list_empty(head)) {
462 cache->ready = head->next;
463 list_del_init(head);
464 } else
465 cache->ready = NULL;
466 }
467
468 return head;
469}
470
471int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
472 size_t size)
473{
474 struct rds_ib_incoming *ibinc;
475 struct rds_page_frag *frag;
476 struct iovec *iov = first_iov;
477 unsigned long to_copy;
478 unsigned long frag_off = 0;
479 unsigned long iov_off = 0;
480 int copied = 0;
481 int ret;
482 u32 len;
483
484 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
485 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
486 len = be32_to_cpu(inc->i_hdr.h_len);
487
488 while (copied < size && copied < len) {
489 if (frag_off == RDS_FRAG_SIZE) {
490 frag = list_entry(frag->f_item.next,
491 struct rds_page_frag, f_item);
492 frag_off = 0;
493 }
494 while (iov_off == iov->iov_len) {
495 iov_off = 0;
496 iov++;
497 }
498
499 to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
500 to_copy = min_t(size_t, to_copy, size - copied);
501 to_copy = min_t(unsigned long, to_copy, len - copied);
502
503 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
504 "[%p, %u] + %lu\n",
505 to_copy, iov->iov_base, iov->iov_len, iov_off,
506 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);
507
508 /* XXX needs + offset for multiple recvs per page */
509 ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
510 frag->f_sg.offset + frag_off,
511 iov->iov_base + iov_off,
512 to_copy);
513 if (ret) {
514 copied = ret;
515 break;
516 }
517
518 iov_off += to_copy;
519 frag_off += to_copy;
520 copied += to_copy;
521 }
522
523 return copied;
524}
525
526/* ic starts out kzalloc()ed */
527void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
528{
529 struct ib_send_wr *wr = &ic->i_ack_wr;
530 struct ib_sge *sge = &ic->i_ack_sge;
531
532 sge->addr = ic->i_ack_dma;
533 sge->length = sizeof(struct rds_header);
534 sge->lkey = ic->i_mr->lkey;
535
536 wr->sg_list = sge;
537 wr->num_sge = 1;
538 wr->opcode = IB_WR_SEND;
539 wr->wr_id = RDS_IB_ACK_WR_ID;
540 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
541}
542
543/*
544 * You'd think that with reliable IB connections you wouldn't need to ack
545 * messages that have been received. The problem is that IB hardware generates
546 * an ack message before it has DMAed the message into memory. This creates a
547 * potential message loss if the HCA is disabled for any reason between when it
548 * sends the ack and before the message is DMAed and processed. This is only a
549 * potential issue if another HCA is available for fail-over.
550 *
551 * When the remote host receives our ack they'll free the sent message from
552 * their send queue. To decrease the latency of this we always send an ack
553 * immediately after we've received messages.
554 *
555 * For simplicity, we only have one ack in flight at a time. This puts
556 * pressure on senders to have deep enough send queues to absorb the latency of
557 * a single ack frame being in flight. This might not be good enough.
558 *
559 * This is implemented by have a long-lived send_wr and sge which point to a
560 * statically allocated ack frame. This ack wr does not fall under the ring
561 * accounting that the tx and rx wrs do. The QP attribute specifically makes
562 * room for it beyond the ring size. Send completion notices its special
563 * wr_id and avoids working with the ring in that case.
564 */
565#ifndef KERNEL_HAS_ATOMIC64
566static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
567 int ack_required)
568{
569 unsigned long flags;
570
571 spin_lock_irqsave(&ic->i_ack_lock, flags);
572 ic->i_ack_next = seq;
573 if (ack_required)
574 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
575 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
576}
577
578static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
579{
580 unsigned long flags;
581 u64 seq;
582
583 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
584
585 spin_lock_irqsave(&ic->i_ack_lock, flags);
586 seq = ic->i_ack_next;
587 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
588
589 return seq;
590}
591#else
592static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
593 int ack_required)
594{
595 atomic64_set(&ic->i_ack_next, seq);
596 if (ack_required) {
597 smp_mb__before_clear_bit();
598 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
599 }
600}
601
602static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
603{
604 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
605 smp_mb__after_clear_bit();
606
607 return atomic64_read(&ic->i_ack_next);
608}
609#endif
610
611
612static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
613{
614 struct rds_header *hdr = ic->i_ack;
615 struct ib_send_wr *failed_wr;
616 u64 seq;
617 int ret;
618
619 seq = rds_ib_get_ack(ic);
620
621 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
622 rds_message_populate_header(hdr, 0, 0, 0);
623 hdr->h_ack = cpu_to_be64(seq);
624 hdr->h_credit = adv_credits;
625 rds_message_make_checksum(hdr);
626 ic->i_ack_queued = jiffies;
627
628 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
629 if (unlikely(ret)) {
630 /* Failed to send. Release the WR, and
631 * force another ACK.
632 */
633 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
634 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
635
636 rds_ib_stats_inc(s_ib_ack_send_failure);
637
638 rds_ib_conn_error(ic->conn, "sending ack failed\n");
639 } else
640 rds_ib_stats_inc(s_ib_ack_sent);
641}
642
643/*
644 * There are 3 ways of getting acknowledgements to the peer:
645 * 1. We call rds_ib_attempt_ack from the recv completion handler
646 * to send an ACK-only frame.
647 * However, there can be only one such frame in the send queue
648 * at any time, so we may have to postpone it.
649 * 2. When another (data) packet is transmitted while there's
650 * an ACK in the queue, we piggyback the ACK sequence number
651 * on the data packet.
652 * 3. If the ACK WR is done sending, we get called from the
653 * send queue completion handler, and check whether there's
654 * another ACK pending (postponed because the WR was on the
655 * queue). If so, we transmit it.
656 *
657 * We maintain 2 variables:
658 * - i_ack_flags, which keeps track of whether the ACK WR
659 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
660 * - i_ack_next, which is the last sequence number we received
661 *
662 * Potentially, send queue and receive queue handlers can run concurrently.
663 * It would be nice to not have to use a spinlock to synchronize things,
664 * but the one problem that rules this out is that 64bit updates are
665 * not atomic on all platforms. Things would be a lot simpler if
666 * we had atomic64 or maybe cmpxchg64 everywhere.
667 *
668 * Reconnecting complicates this picture just slightly. When we
669 * reconnect, we may be seeing duplicate packets. The peer
670 * is retransmitting them, because it hasn't seen an ACK for
671 * them. It is important that we ACK these.
672 *
673 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
674 * this flag set *MUST* be acknowledged immediately.
675 */
676
677/*
678 * When we get here, we're called from the recv queue handler.
679 * Check whether we ought to transmit an ACK.
680 */
681void rds_ib_attempt_ack(struct rds_ib_connection *ic)
682{
683 unsigned int adv_credits;
684
685 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
686 return;
687
688 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
689 rds_ib_stats_inc(s_ib_ack_send_delayed);
690 return;
691 }
692
693 /* Can we get a send credit? */
694 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
695 rds_ib_stats_inc(s_ib_tx_throttle);
696 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
697 return;
698 }
699
700 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
701 rds_ib_send_ack(ic, adv_credits);
702}
703
704/*
705 * We get here from the send completion handler, when the
706 * adapter tells us the ACK frame was sent.
707 */
708void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
709{
710 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
711 rds_ib_attempt_ack(ic);
712}
713
714/*
715 * This is called by the regular xmit code when it wants to piggyback
716 * an ACK on an outgoing frame.
717 */
718u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
719{
720 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
721 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
722 return rds_ib_get_ack(ic);
723}
724
725/*
726 * It's kind of lame that we're copying from the posted receive pages into
727 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
728 * them. But receiving new congestion bitmaps should be a *rare* event, so
729 * hopefully we won't need to invest that complexity in making it more
730 * efficient. By copying we can share a simpler core with TCP which has to
731 * copy.
732 */
733static void rds_ib_cong_recv(struct rds_connection *conn,
734 struct rds_ib_incoming *ibinc)
735{
736 struct rds_cong_map *map;
737 unsigned int map_off;
738 unsigned int map_page;
739 struct rds_page_frag *frag;
740 unsigned long frag_off;
741 unsigned long to_copy;
742 unsigned long copied;
743 uint64_t uncongested = 0;
744 void *addr;
745
746 /* catch completely corrupt packets */
747 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
748 return;
749
750 map = conn->c_fcong;
751 map_page = 0;
752 map_off = 0;
753
754 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
755 frag_off = 0;
756
757 copied = 0;
758
759 while (copied < RDS_CONG_MAP_BYTES) {
760 uint64_t *src, *dst;
761 unsigned int k;
762
763 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
764 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
765
766 addr = kmap_atomic(sg_page(&frag->f_sg));
767
768 src = addr + frag_off;
769 dst = (void *)map->m_page_addrs[map_page] + map_off;
770 for (k = 0; k < to_copy; k += 8) {
771 /* Record ports that became uncongested, ie
772 * bits that changed from 0 to 1. */
773 uncongested |= ~(*src) & *dst;
774 *dst++ = *src++;
775 }
776 kunmap_atomic(addr);
777
778 copied += to_copy;
779
780 map_off += to_copy;
781 if (map_off == PAGE_SIZE) {
782 map_off = 0;
783 map_page++;
784 }
785
786 frag_off += to_copy;
787 if (frag_off == RDS_FRAG_SIZE) {
788 frag = list_entry(frag->f_item.next,
789 struct rds_page_frag, f_item);
790 frag_off = 0;
791 }
792 }
793
794 /* the congestion map is in little endian order */
795 uncongested = le64_to_cpu(uncongested);
796
797 rds_cong_map_updated(map, uncongested);
798}
799
800/*
801 * Rings are posted with all the allocations they'll need to queue the
802 * incoming message to the receiving socket so this can't fail.
803 * All fragments start with a header, so we can make sure we're not receiving
804 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
805 */
806struct rds_ib_ack_state {
807 u64 ack_next;
808 u64 ack_recv;
809 unsigned int ack_required:1;
810 unsigned int ack_next_valid:1;
811 unsigned int ack_recv_valid:1;
812};
813
814static void rds_ib_process_recv(struct rds_connection *conn,
815 struct rds_ib_recv_work *recv, u32 data_len,
816 struct rds_ib_ack_state *state)
817{
818 struct rds_ib_connection *ic = conn->c_transport_data;
819 struct rds_ib_incoming *ibinc = ic->i_ibinc;
820 struct rds_header *ihdr, *hdr;
821
822 /* XXX shut down the connection if port 0,0 are seen? */
823
824 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
825 data_len);
826
827 if (data_len < sizeof(struct rds_header)) {
828 rds_ib_conn_error(conn, "incoming message "
829 "from %pI4 didn't include a "
830 "header, disconnecting and "
831 "reconnecting\n",
832 &conn->c_faddr);
833 return;
834 }
835 data_len -= sizeof(struct rds_header);
836
837 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
838
839 /* Validate the checksum. */
840 if (!rds_message_verify_checksum(ihdr)) {
841 rds_ib_conn_error(conn, "incoming message "
842 "from %pI4 has corrupted header - "
843 "forcing a reconnect\n",
844 &conn->c_faddr);
845 rds_stats_inc(s_recv_drop_bad_checksum);
846 return;
847 }
848
849 /* Process the ACK sequence which comes with every packet */
850 state->ack_recv = be64_to_cpu(ihdr->h_ack);
851 state->ack_recv_valid = 1;
852
853 /* Process the credits update if there was one */
854 if (ihdr->h_credit)
855 rds_ib_send_add_credits(conn, ihdr->h_credit);
856
857 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
858 /* This is an ACK-only packet. The fact that it gets
859 * special treatment here is that historically, ACKs
860 * were rather special beasts.
861 */
862 rds_ib_stats_inc(s_ib_ack_received);
863
864 /*
865 * Usually the frags make their way on to incs and are then freed as
866 * the inc is freed. We don't go that route, so we have to drop the
867 * page ref ourselves. We can't just leave the page on the recv
868 * because that confuses the dma mapping of pages and each recv's use
869 * of a partial page.
870 *
871 * FIXME: Fold this into the code path below.
872 */
873 rds_ib_frag_free(ic, recv->r_frag);
874 recv->r_frag = NULL;
875 return;
876 }
877
878 /*
879 * If we don't already have an inc on the connection then this
880 * fragment has a header and starts a message.. copy its header
881 * into the inc and save the inc so we can hang upcoming fragments
882 * off its list.
883 */
884 if (!ibinc) {
885 ibinc = recv->r_ibinc;
886 recv->r_ibinc = NULL;
887 ic->i_ibinc = ibinc;
888
889 hdr = &ibinc->ii_inc.i_hdr;
890 memcpy(hdr, ihdr, sizeof(*hdr));
891 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
892
893 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
894 ic->i_recv_data_rem, hdr->h_flags);
895 } else {
896 hdr = &ibinc->ii_inc.i_hdr;
897 /* We can't just use memcmp here; fragments of a
898 * single message may carry different ACKs */
899 if (hdr->h_sequence != ihdr->h_sequence ||
900 hdr->h_len != ihdr->h_len ||
901 hdr->h_sport != ihdr->h_sport ||
902 hdr->h_dport != ihdr->h_dport) {
903 rds_ib_conn_error(conn,
904 "fragment header mismatch; forcing reconnect\n");
905 return;
906 }
907 }
908
909 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
910 recv->r_frag = NULL;
911
912 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
913 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
914 else {
915 ic->i_recv_data_rem = 0;
916 ic->i_ibinc = NULL;
917
918 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
919 rds_ib_cong_recv(conn, ibinc);
920 else {
921 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
922 &ibinc->ii_inc, GFP_ATOMIC);
923 state->ack_next = be64_to_cpu(hdr->h_sequence);
924 state->ack_next_valid = 1;
925 }
926
927 /* Evaluate the ACK_REQUIRED flag *after* we received
928 * the complete frame, and after bumping the next_rx
929 * sequence. */
930 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
931 rds_stats_inc(s_recv_ack_required);
932 state->ack_required = 1;
933 }
934
935 rds_inc_put(&ibinc->ii_inc);
936 }
937}
938
939/*
940 * Plucking the oldest entry from the ring can be done concurrently with
941 * the thread refilling the ring. Each ring operation is protected by
942 * spinlocks and the transient state of refilling doesn't change the
943 * recording of which entry is oldest.
944 *
945 * This relies on IB only calling one cq comp_handler for each cq so that
946 * there will only be one caller of rds_recv_incoming() per RDS connection.
947 */
948void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
949{
950 struct rds_connection *conn = context;
951 struct rds_ib_connection *ic = conn->c_transport_data;
952
953 rdsdebug("conn %p cq %p\n", conn, cq);
954
955 rds_ib_stats_inc(s_ib_rx_cq_call);
956
957 tasklet_schedule(&ic->i_recv_tasklet);
958}
959
960static inline void rds_poll_cq(struct rds_ib_connection *ic,
961 struct rds_ib_ack_state *state)
962{
963 struct rds_connection *conn = ic->conn;
964 struct ib_wc wc;
965 struct rds_ib_recv_work *recv;
966
967 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
968 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
969 (unsigned long long)wc.wr_id, wc.status,
970 rds_ib_wc_status_str(wc.status), wc.byte_len,
971 be32_to_cpu(wc.ex.imm_data));
972 rds_ib_stats_inc(s_ib_rx_cq_event);
973
974 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
975
976 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
977
978 /*
979 * Also process recvs in connecting state because it is possible
980 * to get a recv completion _before_ the rdmacm ESTABLISHED
981 * event is processed.
982 */
983 if (wc.status == IB_WC_SUCCESS) {
984 rds_ib_process_recv(conn, recv, wc.byte_len, state);
985 } else {
986 /* We expect errors as the qp is drained during shutdown */
987 if (rds_conn_up(conn) || rds_conn_connecting(conn))
988 rds_ib_conn_error(conn, "recv completion on %pI4 had "
989 "status %u (%s), disconnecting and "
990 "reconnecting\n", &conn->c_faddr,
991 wc.status,
992 rds_ib_wc_status_str(wc.status));
993 }
994
995 /*
996 * It's very important that we only free this ring entry if we've truly
997 * freed the resources allocated to the entry. The refilling path can
998 * leak if we don't.
999 */
1000 rds_ib_ring_free(&ic->i_recv_ring, 1);
1001 }
1002}
1003
1004void rds_ib_recv_tasklet_fn(unsigned long data)
1005{
1006 struct rds_ib_connection *ic = (struct rds_ib_connection *) data;
1007 struct rds_connection *conn = ic->conn;
1008 struct rds_ib_ack_state state = { 0, };
1009
1010 rds_poll_cq(ic, &state);
1011 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
1012 rds_poll_cq(ic, &state);
1013
1014 if (state.ack_next_valid)
1015 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
1016 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
1017 rds_send_drop_acked(conn, state.ack_recv, NULL);
1018 ic->i_ack_recv = state.ack_recv;
1019 }
1020 if (rds_conn_up(conn))
1021 rds_ib_attempt_ack(ic);
1022
1023 /* If we ever end up with a really empty receive ring, we're
1024 * in deep trouble, as the sender will definitely see RNR
1025 * timeouts. */
1026 if (rds_ib_ring_empty(&ic->i_recv_ring))
1027 rds_ib_stats_inc(s_ib_rx_ring_empty);
1028
1029 if (rds_ib_ring_low(&ic->i_recv_ring))
1030 rds_ib_recv_refill(conn, 0);
1031}
1032
1033int rds_ib_recv(struct rds_connection *conn)
1034{
1035 struct rds_ib_connection *ic = conn->c_transport_data;
1036 int ret = 0;
1037
1038 rdsdebug("conn %p\n", conn);
1039 if (rds_conn_up(conn))
1040 rds_ib_attempt_ack(ic);
1041
1042 return ret;
1043}
1044
1045int rds_ib_recv_init(void)
1046{
1047 struct sysinfo si;
1048 int ret = -ENOMEM;
1049
1050 /* Default to 30% of all available RAM for recv memory */
1051 si_meminfo(&si);
1052 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1053
1054 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1055 sizeof(struct rds_ib_incoming),
1056 0, SLAB_HWCACHE_ALIGN, NULL);
1057 if (!rds_ib_incoming_slab)
1058 goto out;
1059
1060 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1061 sizeof(struct rds_page_frag),
1062 0, SLAB_HWCACHE_ALIGN, NULL);
1063 if (!rds_ib_frag_slab)
1064 kmem_cache_destroy(rds_ib_incoming_slab);
1065 else
1066 ret = 0;
1067out:
1068 return ret;
1069}
1070
1071void rds_ib_recv_exit(void)
1072{
1073 kmem_cache_destroy(rds_ib_incoming_slab);
1074 kmem_cache_destroy(rds_ib_frag_slab);
1075}
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/slab.h>
35#include <linux/pci.h>
36#include <linux/dma-mapping.h>
37#include <rdma/rdma_cm.h>
38
39#include "rds_single_path.h"
40#include "rds.h"
41#include "ib.h"
42
43static struct kmem_cache *rds_ib_incoming_slab;
44static struct kmem_cache *rds_ib_frag_slab;
45static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
46
47void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
48{
49 struct rds_ib_recv_work *recv;
50 u32 i;
51
52 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
53 struct ib_sge *sge;
54
55 recv->r_ibinc = NULL;
56 recv->r_frag = NULL;
57
58 recv->r_wr.next = NULL;
59 recv->r_wr.wr_id = i;
60 recv->r_wr.sg_list = recv->r_sge;
61 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
62
63 sge = &recv->r_sge[0];
64 sge->addr = ic->i_recv_hdrs_dma[i];
65 sge->length = sizeof(struct rds_header);
66 sge->lkey = ic->i_pd->local_dma_lkey;
67
68 sge = &recv->r_sge[1];
69 sge->addr = 0;
70 sge->length = RDS_FRAG_SIZE;
71 sge->lkey = ic->i_pd->local_dma_lkey;
72 }
73}
74
75/*
76 * The entire 'from' list, including the from element itself, is put on
77 * to the tail of the 'to' list.
78 */
79static void list_splice_entire_tail(struct list_head *from,
80 struct list_head *to)
81{
82 struct list_head *from_last = from->prev;
83
84 list_splice_tail(from_last, to);
85 list_add_tail(from_last, to);
86}
87
88static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
89{
90 struct list_head *tmp;
91
92 tmp = xchg(&cache->xfer, NULL);
93 if (tmp) {
94 if (cache->ready)
95 list_splice_entire_tail(tmp, cache->ready);
96 else
97 cache->ready = tmp;
98 }
99}
100
101static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
102{
103 struct rds_ib_cache_head *head;
104 int cpu;
105
106 cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
107 if (!cache->percpu)
108 return -ENOMEM;
109
110 for_each_possible_cpu(cpu) {
111 head = per_cpu_ptr(cache->percpu, cpu);
112 head->first = NULL;
113 head->count = 0;
114 }
115 cache->xfer = NULL;
116 cache->ready = NULL;
117
118 return 0;
119}
120
121int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
122{
123 int ret;
124
125 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
126 if (!ret) {
127 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
128 if (ret)
129 free_percpu(ic->i_cache_incs.percpu);
130 }
131
132 return ret;
133}
134
135static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
136 struct list_head *caller_list)
137{
138 struct rds_ib_cache_head *head;
139 int cpu;
140
141 for_each_possible_cpu(cpu) {
142 head = per_cpu_ptr(cache->percpu, cpu);
143 if (head->first) {
144 list_splice_entire_tail(head->first, caller_list);
145 head->first = NULL;
146 }
147 }
148
149 if (cache->ready) {
150 list_splice_entire_tail(cache->ready, caller_list);
151 cache->ready = NULL;
152 }
153}
154
155void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
156{
157 struct rds_ib_incoming *inc;
158 struct rds_ib_incoming *inc_tmp;
159 struct rds_page_frag *frag;
160 struct rds_page_frag *frag_tmp;
161 LIST_HEAD(list);
162
163 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
164 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
165 free_percpu(ic->i_cache_incs.percpu);
166
167 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
168 list_del(&inc->ii_cache_entry);
169 WARN_ON(!list_empty(&inc->ii_frags));
170 kmem_cache_free(rds_ib_incoming_slab, inc);
171 atomic_dec(&rds_ib_allocation);
172 }
173
174 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
175 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
176 free_percpu(ic->i_cache_frags.percpu);
177
178 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
179 list_del(&frag->f_cache_entry);
180 WARN_ON(!list_empty(&frag->f_item));
181 kmem_cache_free(rds_ib_frag_slab, frag);
182 }
183}
184
185/* fwd decl */
186static void rds_ib_recv_cache_put(struct list_head *new_item,
187 struct rds_ib_refill_cache *cache);
188static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
189
190
191/* Recycle frag and attached recv buffer f_sg */
192static void rds_ib_frag_free(struct rds_ib_connection *ic,
193 struct rds_page_frag *frag)
194{
195 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
196
197 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
198 atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
199 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
200}
201
202/* Recycle inc after freeing attached frags */
203void rds_ib_inc_free(struct rds_incoming *inc)
204{
205 struct rds_ib_incoming *ibinc;
206 struct rds_page_frag *frag;
207 struct rds_page_frag *pos;
208 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
209
210 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
211
212 /* Free attached frags */
213 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
214 list_del_init(&frag->f_item);
215 rds_ib_frag_free(ic, frag);
216 }
217 BUG_ON(!list_empty(&ibinc->ii_frags));
218
219 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
220 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
221}
222
223static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
224 struct rds_ib_recv_work *recv)
225{
226 if (recv->r_ibinc) {
227 rds_inc_put(&recv->r_ibinc->ii_inc);
228 recv->r_ibinc = NULL;
229 }
230 if (recv->r_frag) {
231 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
232 rds_ib_frag_free(ic, recv->r_frag);
233 recv->r_frag = NULL;
234 }
235}
236
237void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
238{
239 u32 i;
240
241 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
242 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
243}
244
245static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
246 gfp_t slab_mask)
247{
248 struct rds_ib_incoming *ibinc;
249 struct list_head *cache_item;
250 int avail_allocs;
251
252 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
253 if (cache_item) {
254 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
255 } else {
256 avail_allocs = atomic_add_unless(&rds_ib_allocation,
257 1, rds_ib_sysctl_max_recv_allocation);
258 if (!avail_allocs) {
259 rds_ib_stats_inc(s_ib_rx_alloc_limit);
260 return NULL;
261 }
262 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
263 if (!ibinc) {
264 atomic_dec(&rds_ib_allocation);
265 return NULL;
266 }
267 rds_ib_stats_inc(s_ib_rx_total_incs);
268 }
269 INIT_LIST_HEAD(&ibinc->ii_frags);
270 rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
271
272 return ibinc;
273}
274
275static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
276 gfp_t slab_mask, gfp_t page_mask)
277{
278 struct rds_page_frag *frag;
279 struct list_head *cache_item;
280 int ret;
281
282 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
283 if (cache_item) {
284 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
285 atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
286 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
287 } else {
288 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
289 if (!frag)
290 return NULL;
291
292 sg_init_table(&frag->f_sg, 1);
293 ret = rds_page_remainder_alloc(&frag->f_sg,
294 RDS_FRAG_SIZE, page_mask);
295 if (ret) {
296 kmem_cache_free(rds_ib_frag_slab, frag);
297 return NULL;
298 }
299 rds_ib_stats_inc(s_ib_rx_total_frags);
300 }
301
302 INIT_LIST_HEAD(&frag->f_item);
303
304 return frag;
305}
306
307static int rds_ib_recv_refill_one(struct rds_connection *conn,
308 struct rds_ib_recv_work *recv, gfp_t gfp)
309{
310 struct rds_ib_connection *ic = conn->c_transport_data;
311 struct ib_sge *sge;
312 int ret = -ENOMEM;
313 gfp_t slab_mask = gfp;
314 gfp_t page_mask = gfp;
315
316 if (gfp & __GFP_DIRECT_RECLAIM) {
317 slab_mask = GFP_KERNEL;
318 page_mask = GFP_HIGHUSER;
319 }
320
321 if (!ic->i_cache_incs.ready)
322 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
323 if (!ic->i_cache_frags.ready)
324 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
325
326 /*
327 * ibinc was taken from recv if recv contained the start of a message.
328 * recvs that were continuations will still have this allocated.
329 */
330 if (!recv->r_ibinc) {
331 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
332 if (!recv->r_ibinc)
333 goto out;
334 }
335
336 WARN_ON(recv->r_frag); /* leak! */
337 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
338 if (!recv->r_frag)
339 goto out;
340
341 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
342 1, DMA_FROM_DEVICE);
343 WARN_ON(ret != 1);
344
345 sge = &recv->r_sge[0];
346 sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
347 sge->length = sizeof(struct rds_header);
348
349 sge = &recv->r_sge[1];
350 sge->addr = sg_dma_address(&recv->r_frag->f_sg);
351 sge->length = sg_dma_len(&recv->r_frag->f_sg);
352
353 ret = 0;
354out:
355 return ret;
356}
357
358static int acquire_refill(struct rds_connection *conn)
359{
360 return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
361}
362
363static void release_refill(struct rds_connection *conn)
364{
365 clear_bit(RDS_RECV_REFILL, &conn->c_flags);
366
367 /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
368 * hot path and finding waiters is very rare. We don't want to walk
369 * the system-wide hashed waitqueue buckets in the fast path only to
370 * almost never find waiters.
371 */
372 if (waitqueue_active(&conn->c_waitq))
373 wake_up_all(&conn->c_waitq);
374}
375
376/*
377 * This tries to allocate and post unused work requests after making sure that
378 * they have all the allocations they need to queue received fragments into
379 * sockets.
380 */
381void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
382{
383 struct rds_ib_connection *ic = conn->c_transport_data;
384 struct rds_ib_recv_work *recv;
385 unsigned int posted = 0;
386 int ret = 0;
387 bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
388 bool must_wake = false;
389 u32 pos;
390
391 /* the goal here is to just make sure that someone, somewhere
392 * is posting buffers. If we can't get the refill lock,
393 * let them do their thing
394 */
395 if (!acquire_refill(conn))
396 return;
397
398 while ((prefill || rds_conn_up(conn)) &&
399 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
400 if (pos >= ic->i_recv_ring.w_nr) {
401 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
402 pos);
403 break;
404 }
405
406 recv = &ic->i_recvs[pos];
407 ret = rds_ib_recv_refill_one(conn, recv, gfp);
408 if (ret) {
409 must_wake = true;
410 break;
411 }
412
413 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
414 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
415 (long)sg_dma_address(&recv->r_frag->f_sg));
416
417 /* XXX when can this fail? */
418 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
419 if (ret) {
420 rds_ib_conn_error(conn, "recv post on "
421 "%pI6c returned %d, disconnecting and "
422 "reconnecting\n", &conn->c_faddr,
423 ret);
424 break;
425 }
426
427 posted++;
428
429 if ((posted > 128 && need_resched()) || posted > 8192) {
430 must_wake = true;
431 break;
432 }
433 }
434
435 /* We're doing flow control - update the window. */
436 if (ic->i_flowctl && posted)
437 rds_ib_advertise_credits(conn, posted);
438
439 if (ret)
440 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
441
442 release_refill(conn);
443
444 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
445 * in this case the ring being low is going to lead to more interrupts
446 * and we can safely let the softirq code take care of it unless the
447 * ring is completely empty.
448 *
449 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
450 * we might have raced with the softirq code while we had the refill
451 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
452 * if we should requeue.
453 */
454 if (rds_conn_up(conn) &&
455 (must_wake ||
456 (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
457 rds_ib_ring_empty(&ic->i_recv_ring))) {
458 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
459 }
460 if (can_wait)
461 cond_resched();
462}
463
464/*
465 * We want to recycle several types of recv allocations, like incs and frags.
466 * To use this, the *_free() function passes in the ptr to a list_head within
467 * the recyclee, as well as the cache to put it on.
468 *
469 * First, we put the memory on a percpu list. When this reaches a certain size,
470 * We move it to an intermediate non-percpu list in a lockless manner, with some
471 * xchg/compxchg wizardry.
472 *
473 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
474 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
475 * list_empty() will return true with one element is actually present.
476 */
477static void rds_ib_recv_cache_put(struct list_head *new_item,
478 struct rds_ib_refill_cache *cache)
479{
480 unsigned long flags;
481 struct list_head *old, *chpfirst;
482
483 local_irq_save(flags);
484
485 chpfirst = __this_cpu_read(cache->percpu->first);
486 if (!chpfirst)
487 INIT_LIST_HEAD(new_item);
488 else /* put on front */
489 list_add_tail(new_item, chpfirst);
490
491 __this_cpu_write(cache->percpu->first, new_item);
492 __this_cpu_inc(cache->percpu->count);
493
494 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
495 goto end;
496
497 /*
498 * Return our per-cpu first list to the cache's xfer by atomically
499 * grabbing the current xfer list, appending it to our per-cpu list,
500 * and then atomically returning that entire list back to the
501 * cache's xfer list as long as it's still empty.
502 */
503 do {
504 old = xchg(&cache->xfer, NULL);
505 if (old)
506 list_splice_entire_tail(old, chpfirst);
507 old = cmpxchg(&cache->xfer, NULL, chpfirst);
508 } while (old);
509
510
511 __this_cpu_write(cache->percpu->first, NULL);
512 __this_cpu_write(cache->percpu->count, 0);
513end:
514 local_irq_restore(flags);
515}
516
517static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
518{
519 struct list_head *head = cache->ready;
520
521 if (head) {
522 if (!list_empty(head)) {
523 cache->ready = head->next;
524 list_del_init(head);
525 } else
526 cache->ready = NULL;
527 }
528
529 return head;
530}
531
532int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
533{
534 struct rds_ib_incoming *ibinc;
535 struct rds_page_frag *frag;
536 unsigned long to_copy;
537 unsigned long frag_off = 0;
538 int copied = 0;
539 int ret;
540 u32 len;
541
542 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
543 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
544 len = be32_to_cpu(inc->i_hdr.h_len);
545
546 while (iov_iter_count(to) && copied < len) {
547 if (frag_off == RDS_FRAG_SIZE) {
548 frag = list_entry(frag->f_item.next,
549 struct rds_page_frag, f_item);
550 frag_off = 0;
551 }
552 to_copy = min_t(unsigned long, iov_iter_count(to),
553 RDS_FRAG_SIZE - frag_off);
554 to_copy = min_t(unsigned long, to_copy, len - copied);
555
556 /* XXX needs + offset for multiple recvs per page */
557 rds_stats_add(s_copy_to_user, to_copy);
558 ret = copy_page_to_iter(sg_page(&frag->f_sg),
559 frag->f_sg.offset + frag_off,
560 to_copy,
561 to);
562 if (ret != to_copy)
563 return -EFAULT;
564
565 frag_off += to_copy;
566 copied += to_copy;
567 }
568
569 return copied;
570}
571
572/* ic starts out kzalloc()ed */
573void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
574{
575 struct ib_send_wr *wr = &ic->i_ack_wr;
576 struct ib_sge *sge = &ic->i_ack_sge;
577
578 sge->addr = ic->i_ack_dma;
579 sge->length = sizeof(struct rds_header);
580 sge->lkey = ic->i_pd->local_dma_lkey;
581
582 wr->sg_list = sge;
583 wr->num_sge = 1;
584 wr->opcode = IB_WR_SEND;
585 wr->wr_id = RDS_IB_ACK_WR_ID;
586 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
587}
588
589/*
590 * You'd think that with reliable IB connections you wouldn't need to ack
591 * messages that have been received. The problem is that IB hardware generates
592 * an ack message before it has DMAed the message into memory. This creates a
593 * potential message loss if the HCA is disabled for any reason between when it
594 * sends the ack and before the message is DMAed and processed. This is only a
595 * potential issue if another HCA is available for fail-over.
596 *
597 * When the remote host receives our ack they'll free the sent message from
598 * their send queue. To decrease the latency of this we always send an ack
599 * immediately after we've received messages.
600 *
601 * For simplicity, we only have one ack in flight at a time. This puts
602 * pressure on senders to have deep enough send queues to absorb the latency of
603 * a single ack frame being in flight. This might not be good enough.
604 *
605 * This is implemented by have a long-lived send_wr and sge which point to a
606 * statically allocated ack frame. This ack wr does not fall under the ring
607 * accounting that the tx and rx wrs do. The QP attribute specifically makes
608 * room for it beyond the ring size. Send completion notices its special
609 * wr_id and avoids working with the ring in that case.
610 */
611#ifndef KERNEL_HAS_ATOMIC64
612void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
613{
614 unsigned long flags;
615
616 spin_lock_irqsave(&ic->i_ack_lock, flags);
617 ic->i_ack_next = seq;
618 if (ack_required)
619 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
620 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
621}
622
623static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
624{
625 unsigned long flags;
626 u64 seq;
627
628 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
629
630 spin_lock_irqsave(&ic->i_ack_lock, flags);
631 seq = ic->i_ack_next;
632 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
633
634 return seq;
635}
636#else
637void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
638{
639 atomic64_set(&ic->i_ack_next, seq);
640 if (ack_required) {
641 smp_mb__before_atomic();
642 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
643 }
644}
645
646static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
647{
648 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
649 smp_mb__after_atomic();
650
651 return atomic64_read(&ic->i_ack_next);
652}
653#endif
654
655
656static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
657{
658 struct rds_header *hdr = ic->i_ack;
659 u64 seq;
660 int ret;
661
662 seq = rds_ib_get_ack(ic);
663
664 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
665
666 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
667 sizeof(*hdr), DMA_TO_DEVICE);
668 rds_message_populate_header(hdr, 0, 0, 0);
669 hdr->h_ack = cpu_to_be64(seq);
670 hdr->h_credit = adv_credits;
671 rds_message_make_checksum(hdr);
672 ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
673 sizeof(*hdr), DMA_TO_DEVICE);
674
675 ic->i_ack_queued = jiffies;
676
677 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
678 if (unlikely(ret)) {
679 /* Failed to send. Release the WR, and
680 * force another ACK.
681 */
682 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
683 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
684
685 rds_ib_stats_inc(s_ib_ack_send_failure);
686
687 rds_ib_conn_error(ic->conn, "sending ack failed\n");
688 } else
689 rds_ib_stats_inc(s_ib_ack_sent);
690}
691
692/*
693 * There are 3 ways of getting acknowledgements to the peer:
694 * 1. We call rds_ib_attempt_ack from the recv completion handler
695 * to send an ACK-only frame.
696 * However, there can be only one such frame in the send queue
697 * at any time, so we may have to postpone it.
698 * 2. When another (data) packet is transmitted while there's
699 * an ACK in the queue, we piggyback the ACK sequence number
700 * on the data packet.
701 * 3. If the ACK WR is done sending, we get called from the
702 * send queue completion handler, and check whether there's
703 * another ACK pending (postponed because the WR was on the
704 * queue). If so, we transmit it.
705 *
706 * We maintain 2 variables:
707 * - i_ack_flags, which keeps track of whether the ACK WR
708 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
709 * - i_ack_next, which is the last sequence number we received
710 *
711 * Potentially, send queue and receive queue handlers can run concurrently.
712 * It would be nice to not have to use a spinlock to synchronize things,
713 * but the one problem that rules this out is that 64bit updates are
714 * not atomic on all platforms. Things would be a lot simpler if
715 * we had atomic64 or maybe cmpxchg64 everywhere.
716 *
717 * Reconnecting complicates this picture just slightly. When we
718 * reconnect, we may be seeing duplicate packets. The peer
719 * is retransmitting them, because it hasn't seen an ACK for
720 * them. It is important that we ACK these.
721 *
722 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
723 * this flag set *MUST* be acknowledged immediately.
724 */
725
726/*
727 * When we get here, we're called from the recv queue handler.
728 * Check whether we ought to transmit an ACK.
729 */
730void rds_ib_attempt_ack(struct rds_ib_connection *ic)
731{
732 unsigned int adv_credits;
733
734 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
735 return;
736
737 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
738 rds_ib_stats_inc(s_ib_ack_send_delayed);
739 return;
740 }
741
742 /* Can we get a send credit? */
743 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
744 rds_ib_stats_inc(s_ib_tx_throttle);
745 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
746 return;
747 }
748
749 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
750 rds_ib_send_ack(ic, adv_credits);
751}
752
753/*
754 * We get here from the send completion handler, when the
755 * adapter tells us the ACK frame was sent.
756 */
757void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
758{
759 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
760 rds_ib_attempt_ack(ic);
761}
762
763/*
764 * This is called by the regular xmit code when it wants to piggyback
765 * an ACK on an outgoing frame.
766 */
767u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
768{
769 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
770 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
771 return rds_ib_get_ack(ic);
772}
773
774/*
775 * It's kind of lame that we're copying from the posted receive pages into
776 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
777 * them. But receiving new congestion bitmaps should be a *rare* event, so
778 * hopefully we won't need to invest that complexity in making it more
779 * efficient. By copying we can share a simpler core with TCP which has to
780 * copy.
781 */
782static void rds_ib_cong_recv(struct rds_connection *conn,
783 struct rds_ib_incoming *ibinc)
784{
785 struct rds_cong_map *map;
786 unsigned int map_off;
787 unsigned int map_page;
788 struct rds_page_frag *frag;
789 unsigned long frag_off;
790 unsigned long to_copy;
791 unsigned long copied;
792 __le64 uncongested = 0;
793 void *addr;
794
795 /* catch completely corrupt packets */
796 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
797 return;
798
799 map = conn->c_fcong;
800 map_page = 0;
801 map_off = 0;
802
803 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
804 frag_off = 0;
805
806 copied = 0;
807
808 while (copied < RDS_CONG_MAP_BYTES) {
809 __le64 *src, *dst;
810 unsigned int k;
811
812 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
813 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
814
815 addr = kmap_atomic(sg_page(&frag->f_sg));
816
817 src = addr + frag->f_sg.offset + frag_off;
818 dst = (void *)map->m_page_addrs[map_page] + map_off;
819 for (k = 0; k < to_copy; k += 8) {
820 /* Record ports that became uncongested, ie
821 * bits that changed from 0 to 1. */
822 uncongested |= ~(*src) & *dst;
823 *dst++ = *src++;
824 }
825 kunmap_atomic(addr);
826
827 copied += to_copy;
828
829 map_off += to_copy;
830 if (map_off == PAGE_SIZE) {
831 map_off = 0;
832 map_page++;
833 }
834
835 frag_off += to_copy;
836 if (frag_off == RDS_FRAG_SIZE) {
837 frag = list_entry(frag->f_item.next,
838 struct rds_page_frag, f_item);
839 frag_off = 0;
840 }
841 }
842
843 /* the congestion map is in little endian order */
844 rds_cong_map_updated(map, le64_to_cpu(uncongested));
845}
846
847static void rds_ib_process_recv(struct rds_connection *conn,
848 struct rds_ib_recv_work *recv, u32 data_len,
849 struct rds_ib_ack_state *state)
850{
851 struct rds_ib_connection *ic = conn->c_transport_data;
852 struct rds_ib_incoming *ibinc = ic->i_ibinc;
853 struct rds_header *ihdr, *hdr;
854 dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
855
856 /* XXX shut down the connection if port 0,0 are seen? */
857
858 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
859 data_len);
860
861 if (data_len < sizeof(struct rds_header)) {
862 rds_ib_conn_error(conn, "incoming message "
863 "from %pI6c didn't include a "
864 "header, disconnecting and "
865 "reconnecting\n",
866 &conn->c_faddr);
867 return;
868 }
869 data_len -= sizeof(struct rds_header);
870
871 ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
872
873 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
874 sizeof(*ihdr), DMA_FROM_DEVICE);
875 /* Validate the checksum. */
876 if (!rds_message_verify_checksum(ihdr)) {
877 rds_ib_conn_error(conn, "incoming message "
878 "from %pI6c has corrupted header - "
879 "forcing a reconnect\n",
880 &conn->c_faddr);
881 rds_stats_inc(s_recv_drop_bad_checksum);
882 goto done;
883 }
884
885 /* Process the ACK sequence which comes with every packet */
886 state->ack_recv = be64_to_cpu(ihdr->h_ack);
887 state->ack_recv_valid = 1;
888
889 /* Process the credits update if there was one */
890 if (ihdr->h_credit)
891 rds_ib_send_add_credits(conn, ihdr->h_credit);
892
893 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
894 /* This is an ACK-only packet. The fact that it gets
895 * special treatment here is that historically, ACKs
896 * were rather special beasts.
897 */
898 rds_ib_stats_inc(s_ib_ack_received);
899
900 /*
901 * Usually the frags make their way on to incs and are then freed as
902 * the inc is freed. We don't go that route, so we have to drop the
903 * page ref ourselves. We can't just leave the page on the recv
904 * because that confuses the dma mapping of pages and each recv's use
905 * of a partial page.
906 *
907 * FIXME: Fold this into the code path below.
908 */
909 rds_ib_frag_free(ic, recv->r_frag);
910 recv->r_frag = NULL;
911 goto done;
912 }
913
914 /*
915 * If we don't already have an inc on the connection then this
916 * fragment has a header and starts a message.. copy its header
917 * into the inc and save the inc so we can hang upcoming fragments
918 * off its list.
919 */
920 if (!ibinc) {
921 ibinc = recv->r_ibinc;
922 recv->r_ibinc = NULL;
923 ic->i_ibinc = ibinc;
924
925 hdr = &ibinc->ii_inc.i_hdr;
926 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
927 local_clock();
928 memcpy(hdr, ihdr, sizeof(*hdr));
929 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
930 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
931 local_clock();
932
933 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
934 ic->i_recv_data_rem, hdr->h_flags);
935 } else {
936 hdr = &ibinc->ii_inc.i_hdr;
937 /* We can't just use memcmp here; fragments of a
938 * single message may carry different ACKs */
939 if (hdr->h_sequence != ihdr->h_sequence ||
940 hdr->h_len != ihdr->h_len ||
941 hdr->h_sport != ihdr->h_sport ||
942 hdr->h_dport != ihdr->h_dport) {
943 rds_ib_conn_error(conn,
944 "fragment header mismatch; forcing reconnect\n");
945 goto done;
946 }
947 }
948
949 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
950 recv->r_frag = NULL;
951
952 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
953 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
954 else {
955 ic->i_recv_data_rem = 0;
956 ic->i_ibinc = NULL;
957
958 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
959 rds_ib_cong_recv(conn, ibinc);
960 } else {
961 rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
962 &ibinc->ii_inc, GFP_ATOMIC);
963 state->ack_next = be64_to_cpu(hdr->h_sequence);
964 state->ack_next_valid = 1;
965 }
966
967 /* Evaluate the ACK_REQUIRED flag *after* we received
968 * the complete frame, and after bumping the next_rx
969 * sequence. */
970 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
971 rds_stats_inc(s_recv_ack_required);
972 state->ack_required = 1;
973 }
974
975 rds_inc_put(&ibinc->ii_inc);
976 }
977done:
978 ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
979 sizeof(*ihdr), DMA_FROM_DEVICE);
980}
981
982void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
983 struct ib_wc *wc,
984 struct rds_ib_ack_state *state)
985{
986 struct rds_connection *conn = ic->conn;
987 struct rds_ib_recv_work *recv;
988
989 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
990 (unsigned long long)wc->wr_id, wc->status,
991 ib_wc_status_msg(wc->status), wc->byte_len,
992 be32_to_cpu(wc->ex.imm_data));
993
994 rds_ib_stats_inc(s_ib_rx_cq_event);
995 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
996 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
997 DMA_FROM_DEVICE);
998
999 /* Also process recvs in connecting state because it is possible
1000 * to get a recv completion _before_ the rdmacm ESTABLISHED
1001 * event is processed.
1002 */
1003 if (wc->status == IB_WC_SUCCESS) {
1004 rds_ib_process_recv(conn, recv, wc->byte_len, state);
1005 } else {
1006 /* We expect errors as the qp is drained during shutdown */
1007 if (rds_conn_up(conn) || rds_conn_connecting(conn))
1008 rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
1009 &conn->c_laddr, &conn->c_faddr,
1010 conn->c_tos, wc->status,
1011 ib_wc_status_msg(wc->status),
1012 wc->vendor_err);
1013 }
1014
1015 /* rds_ib_process_recv() doesn't always consume the frag, and
1016 * we might not have called it at all if the wc didn't indicate
1017 * success. We already unmapped the frag's pages, though, and
1018 * the following rds_ib_ring_free() call tells the refill path
1019 * that it will not find an allocated frag here. Make sure we
1020 * keep that promise by freeing a frag that's still on the ring.
1021 */
1022 if (recv->r_frag) {
1023 rds_ib_frag_free(ic, recv->r_frag);
1024 recv->r_frag = NULL;
1025 }
1026 rds_ib_ring_free(&ic->i_recv_ring, 1);
1027
1028 /* If we ever end up with a really empty receive ring, we're
1029 * in deep trouble, as the sender will definitely see RNR
1030 * timeouts. */
1031 if (rds_ib_ring_empty(&ic->i_recv_ring))
1032 rds_ib_stats_inc(s_ib_rx_ring_empty);
1033
1034 if (rds_ib_ring_low(&ic->i_recv_ring)) {
1035 rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
1036 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1037 }
1038}
1039
1040int rds_ib_recv_path(struct rds_conn_path *cp)
1041{
1042 struct rds_connection *conn = cp->cp_conn;
1043 struct rds_ib_connection *ic = conn->c_transport_data;
1044
1045 rdsdebug("conn %p\n", conn);
1046 if (rds_conn_up(conn)) {
1047 rds_ib_attempt_ack(ic);
1048 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1049 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1050 }
1051
1052 return 0;
1053}
1054
1055int rds_ib_recv_init(void)
1056{
1057 struct sysinfo si;
1058 int ret = -ENOMEM;
1059
1060 /* Default to 30% of all available RAM for recv memory */
1061 si_meminfo(&si);
1062 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1063
1064 rds_ib_incoming_slab =
1065 kmem_cache_create_usercopy("rds_ib_incoming",
1066 sizeof(struct rds_ib_incoming),
1067 0, SLAB_HWCACHE_ALIGN,
1068 offsetof(struct rds_ib_incoming,
1069 ii_inc.i_usercopy),
1070 sizeof(struct rds_inc_usercopy),
1071 NULL);
1072 if (!rds_ib_incoming_slab)
1073 goto out;
1074
1075 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1076 sizeof(struct rds_page_frag),
1077 0, SLAB_HWCACHE_ALIGN, NULL);
1078 if (!rds_ib_frag_slab) {
1079 kmem_cache_destroy(rds_ib_incoming_slab);
1080 rds_ib_incoming_slab = NULL;
1081 } else
1082 ret = 0;
1083out:
1084 return ret;
1085}
1086
1087void rds_ib_recv_exit(void)
1088{
1089 WARN_ON(atomic_read(&rds_ib_allocation));
1090
1091 kmem_cache_destroy(rds_ib_incoming_slab);
1092 kmem_cache_destroy(rds_ib_frag_slab);
1093}