projects/linux·blob master · refs

1096 lines · 31.1 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/sched/clock.h>
35#include <linux/slab.h>
36#include <linux/pci.h>
37#include <linux/dma-mapping.h>
38#include <rdma/rdma_cm.h>
39
40#include "rds_single_path.h"
41#include "rds.h"
42#include "ib.h"
43
44static struct kmem_cache *rds_ib_incoming_slab;
45static struct kmem_cache *rds_ib_frag_slab;
46static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
47
48void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
49{
50 struct rds_ib_recv_work *recv;
51 u32 i;
52
53 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
54 struct ib_sge *sge;
55
56 recv->r_ibinc = NULL;
57 recv->r_frag = NULL;
58
59 recv->r_wr.next = NULL;
60 recv->r_wr.wr_id = i;
61 recv->r_wr.sg_list = recv->r_sge;
62 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
63
64 sge = &recv->r_sge[0];
65 sge->addr = ic->i_recv_hdrs_dma[i];
66 sge->length = sizeof(struct rds_header);
67 sge->lkey = ic->i_pd->local_dma_lkey;
68
69 sge = &recv->r_sge[1];
70 sge->addr = 0;
71 sge->length = RDS_FRAG_SIZE;
72 sge->lkey = ic->i_pd->local_dma_lkey;
73 }
74}
75
76/*
77 * The entire 'from' list, including the from element itself, is put on
78 * to the tail of the 'to' list.
79 */
80static void list_splice_entire_tail(struct list_head *from,
81 struct list_head *to)
82{
83 struct list_head *from_last = from->prev;
84
85 list_splice_tail(from_last, to);
86 list_add_tail(from_last, to);
87}
88
89static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
90{
91 struct list_head *tmp;
92
93 tmp = xchg(&cache->xfer, NULL);
94 if (tmp) {
95 if (cache->ready)
96 list_splice_entire_tail(tmp, cache->ready);
97 else
98 cache->ready = tmp;
99 }
100}
101
102static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp)
103{
104 struct rds_ib_cache_head *head;
105 int cpu;
106
107 cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
108 if (!cache->percpu)
109 return -ENOMEM;
110
111 for_each_possible_cpu(cpu) {
112 head = per_cpu_ptr(cache->percpu, cpu);
113 head->first = NULL;
114 head->count = 0;
115 }
116 cache->xfer = NULL;
117 cache->ready = NULL;
118
119 return 0;
120}
121
122int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp)
123{
124 int ret;
125
126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
127 if (!ret) {
128 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
129 if (ret)
130 free_percpu(ic->i_cache_incs.percpu);
131 }
132
133 return ret;
134}
135
136static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
137 struct list_head *caller_list)
138{
139 struct rds_ib_cache_head *head;
140 int cpu;
141
142 for_each_possible_cpu(cpu) {
143 head = per_cpu_ptr(cache->percpu, cpu);
144 if (head->first) {
145 list_splice_entire_tail(head->first, caller_list);
146 head->first = NULL;
147 }
148 }
149
150 if (cache->ready) {
151 list_splice_entire_tail(cache->ready, caller_list);
152 cache->ready = NULL;
153 }
154}
155
156void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
157{
158 struct rds_ib_incoming *inc;
159 struct rds_ib_incoming *inc_tmp;
160 struct rds_page_frag *frag;
161 struct rds_page_frag *frag_tmp;
162 LIST_HEAD(list);
163
164 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
165 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
166 free_percpu(ic->i_cache_incs.percpu);
167
168 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
169 list_del(&inc->ii_cache_entry);
170 WARN_ON(!list_empty(&inc->ii_frags));
171 kmem_cache_free(rds_ib_incoming_slab, inc);
172 atomic_dec(&rds_ib_allocation);
173 }
174
175 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
176 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
177 free_percpu(ic->i_cache_frags.percpu);
178
179 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
180 list_del(&frag->f_cache_entry);
181 WARN_ON(!list_empty(&frag->f_item));
182 kmem_cache_free(rds_ib_frag_slab, frag);
183 }
184}
185
186/* fwd decl */
187static void rds_ib_recv_cache_put(struct list_head *new_item,
188 struct rds_ib_refill_cache *cache);
189static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
190
191
192/* Recycle frag and attached recv buffer f_sg */
193static void rds_ib_frag_free(struct rds_ib_connection *ic,
194 struct rds_page_frag *frag)
195{
196 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
197
198 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
199 atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
200 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
201}
202
203/* Recycle inc after freeing attached frags */
204void rds_ib_inc_free(struct rds_incoming *inc)
205{
206 struct rds_ib_incoming *ibinc;
207 struct rds_page_frag *frag;
208 struct rds_page_frag *pos;
209 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
210
211 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
212
213 /* Free attached frags */
214 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
215 list_del_init(&frag->f_item);
216 rds_ib_frag_free(ic, frag);
217 }
218 BUG_ON(!list_empty(&ibinc->ii_frags));
219
220 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
221 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
222}
223
224static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
225 struct rds_ib_recv_work *recv)
226{
227 if (recv->r_ibinc) {
228 rds_inc_put(&recv->r_ibinc->ii_inc);
229 recv->r_ibinc = NULL;
230 }
231 if (recv->r_frag) {
232 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
233 rds_ib_frag_free(ic, recv->r_frag);
234 recv->r_frag = NULL;
235 }
236}
237
238void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
239{
240 u32 i;
241
242 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
243 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
244}
245
246static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
247 gfp_t slab_mask)
248{
249 struct rds_ib_incoming *ibinc;
250 struct list_head *cache_item;
251 int avail_allocs;
252
253 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
254 if (cache_item) {
255 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
256 } else {
257 avail_allocs = atomic_add_unless(&rds_ib_allocation,
258 1, rds_ib_sysctl_max_recv_allocation);
259 if (!avail_allocs) {
260 rds_ib_stats_inc(s_ib_rx_alloc_limit);
261 return NULL;
262 }
263 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
264 if (!ibinc) {
265 atomic_dec(&rds_ib_allocation);
266 return NULL;
267 }
268 rds_ib_stats_inc(s_ib_rx_total_incs);
269 }
270 INIT_LIST_HEAD(&ibinc->ii_frags);
271 rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr);
272
273 return ibinc;
274}
275
276static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
277 gfp_t slab_mask, gfp_t page_mask)
278{
279 struct rds_page_frag *frag;
280 struct list_head *cache_item;
281 int ret;
282
283 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
284 if (cache_item) {
285 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
286 atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
287 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
288 } else {
289 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
290 if (!frag)
291 return NULL;
292
293 sg_init_table(&frag->f_sg, 1);
294 ret = rds_page_remainder_alloc(&frag->f_sg,
295 RDS_FRAG_SIZE, page_mask);
296 if (ret) {
297 kmem_cache_free(rds_ib_frag_slab, frag);
298 return NULL;
299 }
300 rds_ib_stats_inc(s_ib_rx_total_frags);
301 }
302
303 INIT_LIST_HEAD(&frag->f_item);
304
305 return frag;
306}
307
308static int rds_ib_recv_refill_one(struct rds_connection *conn,
309 struct rds_ib_recv_work *recv, gfp_t gfp)
310{
311 struct rds_ib_connection *ic = conn->c_transport_data;
312 struct ib_sge *sge;
313 int ret = -ENOMEM;
314 gfp_t slab_mask = gfp;
315 gfp_t page_mask = gfp;
316
317 if (gfp & __GFP_DIRECT_RECLAIM) {
318 slab_mask = GFP_KERNEL;
319 page_mask = GFP_HIGHUSER;
320 }
321
322 if (!ic->i_cache_incs.ready)
323 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
324 if (!ic->i_cache_frags.ready)
325 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
326
327 /*
328 * ibinc was taken from recv if recv contained the start of a message.
329 * recvs that were continuations will still have this allocated.
330 */
331 if (!recv->r_ibinc) {
332 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
333 if (!recv->r_ibinc)
334 goto out;
335 }
336
337 WARN_ON(recv->r_frag); /* leak! */
338 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
339 if (!recv->r_frag)
340 goto out;
341
342 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
343 1, DMA_FROM_DEVICE);
344 WARN_ON(ret != 1);
345
346 sge = &recv->r_sge[0];
347 sge->addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
348 sge->length = sizeof(struct rds_header);
349
350 sge = &recv->r_sge[1];
351 sge->addr = sg_dma_address(&recv->r_frag->f_sg);
352 sge->length = sg_dma_len(&recv->r_frag->f_sg);
353
354 ret = 0;
355out:
356 return ret;
357}
358
359static int acquire_refill(struct rds_connection *conn)
360{
361 return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
362}
363
364static void release_refill(struct rds_connection *conn)
365{
366 clear_bit_unlock(RDS_RECV_REFILL, &conn->c_flags);
367
368 /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
369 * hot path and finding waiters is very rare. We don't want to walk
370 * the system-wide hashed waitqueue buckets in the fast path only to
371 * almost never find waiters.
372 */
373 if (wq_has_sleeper(&conn->c_waitq))
374 wake_up_all(&conn->c_waitq);
375}
376
377/*
378 * This tries to allocate and post unused work requests after making sure that
379 * they have all the allocations they need to queue received fragments into
380 * sockets.
381 */
382void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
383{
384 struct rds_ib_connection *ic = conn->c_transport_data;
385 struct rds_ib_recv_work *recv;
386 unsigned int posted = 0;
387 int ret = 0;
388 bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
389 bool must_wake = false;
390 u32 pos;
391
392 /* the goal here is to just make sure that someone, somewhere
393 * is posting buffers. If we can't get the refill lock,
394 * let them do their thing. The holder may also be
395 * rds_conn_shutdown() tearing the path down, in which case
396 * there is nothing to post.
397 */
398 if (!acquire_refill(conn))
399 return;
400
401 while ((prefill || rds_conn_up(conn)) &&
402 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
403 if (pos >= ic->i_recv_ring.w_nr) {
404 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
405 pos);
406 break;
407 }
408
409 recv = &ic->i_recvs[pos];
410 ret = rds_ib_recv_refill_one(conn, recv, gfp);
411 if (ret) {
412 must_wake = true;
413 break;
414 }
415
416 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
417 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
418 (long)sg_dma_address(&recv->r_frag->f_sg));
419
420 /* XXX when can this fail? */
421 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
422 if (ret) {
423 rds_ib_conn_error(conn, "recv post on "
424 "%pI6c returned %d, disconnecting and "
425 "reconnecting\n", &conn->c_faddr,
426 ret);
427 break;
428 }
429
430 posted++;
431
432 if ((posted > 128 && need_resched()) || posted > 8192) {
433 must_wake = true;
434 break;
435 }
436 }
437
438 /* We're doing flow control - update the window. */
439 if (ic->i_flowctl && posted)
440 rds_ib_advertise_credits(conn, posted);
441
442 if (ret)
443 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
444
445 release_refill(conn);
446
447 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
448 * in this case the ring being low is going to lead to more interrupts
449 * and we can safely let the softirq code take care of it unless the
450 * ring is completely empty.
451 *
452 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
453 * we might have raced with the softirq code while we had the refill
454 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
455 * if we should requeue.
456 */
457 if (rds_conn_up(conn) &&
458 (must_wake ||
459 (can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
460 rds_ib_ring_empty(&ic->i_recv_ring))) {
461 queue_delayed_work(conn->c_path->cp_wq, &conn->c_recv_w, 1);
462 }
463 if (can_wait)
464 cond_resched();
465}
466
467/*
468 * We want to recycle several types of recv allocations, like incs and frags.
469 * To use this, the *_free() function passes in the ptr to a list_head within
470 * the recyclee, as well as the cache to put it on.
471 *
472 * First, we put the memory on a percpu list. When this reaches a certain size,
473 * We move it to an intermediate non-percpu list in a lockless manner, with some
474 * xchg/compxchg wizardry.
475 *
476 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
477 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
478 * list_empty() will return true with one element is actually present.
479 */
480static void rds_ib_recv_cache_put(struct list_head *new_item,
481 struct rds_ib_refill_cache *cache)
482{
483 unsigned long flags;
484 struct list_head *old, *chpfirst;
485
486 local_irq_save(flags);
487
488 chpfirst = __this_cpu_read(cache->percpu->first);
489 if (!chpfirst)
490 INIT_LIST_HEAD(new_item);
491 else /* put on front */
492 list_add_tail(new_item, chpfirst);
493
494 __this_cpu_write(cache->percpu->first, new_item);
495 __this_cpu_inc(cache->percpu->count);
496
497 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
498 goto end;
499
500 /*
501 * Return our per-cpu first list to the cache's xfer by atomically
502 * grabbing the current xfer list, appending it to our per-cpu list,
503 * and then atomically returning that entire list back to the
504 * cache's xfer list as long as it's still empty.
505 */
506 do {
507 old = xchg(&cache->xfer, NULL);
508 if (old)
509 list_splice_entire_tail(old, chpfirst);
510 old = cmpxchg(&cache->xfer, NULL, chpfirst);
511 } while (old);
512
513
514 __this_cpu_write(cache->percpu->first, NULL);
515 __this_cpu_write(cache->percpu->count, 0);
516end:
517 local_irq_restore(flags);
518}
519
520static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
521{
522 struct list_head *head = cache->ready;
523
524 if (head) {
525 if (!list_empty(head)) {
526 cache->ready = head->next;
527 list_del_init(head);
528 } else
529 cache->ready = NULL;
530 }
531
532 return head;
533}
534
535int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
536{
537 struct rds_ib_incoming *ibinc;
538 struct rds_page_frag *frag;
539 unsigned long to_copy;
540 unsigned long frag_off = 0;
541 int copied = 0;
542 int ret;
543 u32 len;
544
545 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
546 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
547 len = be32_to_cpu(inc->i_hdr.h_len);
548
549 while (iov_iter_count(to) && copied < len) {
550 if (frag_off == RDS_FRAG_SIZE) {
551 frag = list_entry(frag->f_item.next,
552 struct rds_page_frag, f_item);
553 frag_off = 0;
554 }
555 to_copy = min_t(unsigned long, iov_iter_count(to),
556 RDS_FRAG_SIZE - frag_off);
557 to_copy = min_t(unsigned long, to_copy, len - copied);
558
559 /* XXX needs + offset for multiple recvs per page */
560 rds_stats_add(s_copy_to_user, to_copy);
561 ret = copy_page_to_iter(sg_page(&frag->f_sg),
562 frag->f_sg.offset + frag_off,
563 to_copy,
564 to);
565 if (ret != to_copy)
566 return -EFAULT;
567
568 frag_off += to_copy;
569 copied += to_copy;
570 }
571
572 return copied;
573}
574
575/* ic starts out kzalloc()ed */
576void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
577{
578 struct ib_send_wr *wr = &ic->i_ack_wr;
579 struct ib_sge *sge = &ic->i_ack_sge;
580
581 sge->addr = ic->i_ack_dma;
582 sge->length = sizeof(struct rds_header);
583 sge->lkey = ic->i_pd->local_dma_lkey;
584
585 wr->sg_list = sge;
586 wr->num_sge = 1;
587 wr->opcode = IB_WR_SEND;
588 wr->wr_id = RDS_IB_ACK_WR_ID;
589 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
590}
591
592/*
593 * You'd think that with reliable IB connections you wouldn't need to ack
594 * messages that have been received. The problem is that IB hardware generates
595 * an ack message before it has DMAed the message into memory. This creates a
596 * potential message loss if the HCA is disabled for any reason between when it
597 * sends the ack and before the message is DMAed and processed. This is only a
598 * potential issue if another HCA is available for fail-over.
599 *
600 * When the remote host receives our ack they'll free the sent message from
601 * their send queue. To decrease the latency of this we always send an ack
602 * immediately after we've received messages.
603 *
604 * For simplicity, we only have one ack in flight at a time. This puts
605 * pressure on senders to have deep enough send queues to absorb the latency of
606 * a single ack frame being in flight. This might not be good enough.
607 *
608 * This is implemented by have a long-lived send_wr and sge which point to a
609 * statically allocated ack frame. This ack wr does not fall under the ring
610 * accounting that the tx and rx wrs do. The QP attribute specifically makes
611 * room for it beyond the ring size. Send completion notices its special
612 * wr_id and avoids working with the ring in that case.
613 */
614#ifndef KERNEL_HAS_ATOMIC64
615void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
616{
617 unsigned long flags;
618
619 spin_lock_irqsave(&ic->i_ack_lock, flags);
620 ic->i_ack_next = seq;
621 if (ack_required)
622 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
623 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
624}
625
626static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
627{
628 unsigned long flags;
629 u64 seq;
630
631 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
632
633 spin_lock_irqsave(&ic->i_ack_lock, flags);
634 seq = ic->i_ack_next;
635 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
636
637 return seq;
638}
639#else
640void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
641{
642 atomic64_set(&ic->i_ack_next, seq);
643 if (ack_required) {
644 smp_mb__before_atomic();
645 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
646 }
647}
648
649static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
650{
651 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
652 smp_mb__after_atomic();
653
654 return atomic64_read(&ic->i_ack_next);
655}
656#endif
657
658
659static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
660{
661 struct rds_header *hdr = ic->i_ack;
662 u64 seq;
663 int ret;
664
665 seq = rds_ib_get_ack(ic);
666
667 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
668
669 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, ic->i_ack_dma,
670 sizeof(*hdr), DMA_TO_DEVICE);
671 rds_message_populate_header(hdr, 0, 0, 0);
672 hdr->h_ack = cpu_to_be64(seq);
673 hdr->h_credit = adv_credits;
674 rds_message_make_checksum(hdr);
675 ib_dma_sync_single_for_device(ic->rds_ibdev->dev, ic->i_ack_dma,
676 sizeof(*hdr), DMA_TO_DEVICE);
677
678 ic->i_ack_queued = jiffies;
679
680 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
681 if (unlikely(ret)) {
682 /* Failed to send. Release the WR, and
683 * force another ACK.
684 */
685 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
686 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
687
688 rds_ib_stats_inc(s_ib_ack_send_failure);
689
690 rds_ib_conn_error(ic->conn, "sending ack failed\n");
691 } else
692 rds_ib_stats_inc(s_ib_ack_sent);
693}
694
695/*
696 * There are 3 ways of getting acknowledgements to the peer:
697 * 1. We call rds_ib_attempt_ack from the recv completion handler
698 * to send an ACK-only frame.
699 * However, there can be only one such frame in the send queue
700 * at any time, so we may have to postpone it.
701 * 2. When another (data) packet is transmitted while there's
702 * an ACK in the queue, we piggyback the ACK sequence number
703 * on the data packet.
704 * 3. If the ACK WR is done sending, we get called from the
705 * send queue completion handler, and check whether there's
706 * another ACK pending (postponed because the WR was on the
707 * queue). If so, we transmit it.
708 *
709 * We maintain 2 variables:
710 * - i_ack_flags, which keeps track of whether the ACK WR
711 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
712 * - i_ack_next, which is the last sequence number we received
713 *
714 * Potentially, send queue and receive queue handlers can run concurrently.
715 * It would be nice to not have to use a spinlock to synchronize things,
716 * but the one problem that rules this out is that 64bit updates are
717 * not atomic on all platforms. Things would be a lot simpler if
718 * we had atomic64 or maybe cmpxchg64 everywhere.
719 *
720 * Reconnecting complicates this picture just slightly. When we
721 * reconnect, we may be seeing duplicate packets. The peer
722 * is retransmitting them, because it hasn't seen an ACK for
723 * them. It is important that we ACK these.
724 *
725 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
726 * this flag set *MUST* be acknowledged immediately.
727 */
728
729/*
730 * When we get here, we're called from the recv queue handler.
731 * Check whether we ought to transmit an ACK.
732 */
733void rds_ib_attempt_ack(struct rds_ib_connection *ic)
734{
735 unsigned int adv_credits;
736
737 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
738 return;
739
740 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
741 rds_ib_stats_inc(s_ib_ack_send_delayed);
742 return;
743 }
744
745 /* Can we get a send credit? */
746 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
747 rds_ib_stats_inc(s_ib_tx_throttle);
748 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
749 return;
750 }
751
752 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
753 rds_ib_send_ack(ic, adv_credits);
754}
755
756/*
757 * We get here from the send completion handler, when the
758 * adapter tells us the ACK frame was sent.
759 */
760void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
761{
762 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
763 rds_ib_attempt_ack(ic);
764}
765
766/*
767 * This is called by the regular xmit code when it wants to piggyback
768 * an ACK on an outgoing frame.
769 */
770u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
771{
772 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
773 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
774 return rds_ib_get_ack(ic);
775}
776
777/*
778 * It's kind of lame that we're copying from the posted receive pages into
779 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
780 * them. But receiving new congestion bitmaps should be a *rare* event, so
781 * hopefully we won't need to invest that complexity in making it more
782 * efficient. By copying we can share a simpler core with TCP which has to
783 * copy.
784 */
785static void rds_ib_cong_recv(struct rds_connection *conn,
786 struct rds_ib_incoming *ibinc)
787{
788 struct rds_cong_map *map;
789 unsigned int map_off;
790 unsigned int map_page;
791 struct rds_page_frag *frag;
792 unsigned long frag_off;
793 unsigned long to_copy;
794 unsigned long copied;
795 __le64 uncongested = 0;
796 void *addr;
797
798 /* catch completely corrupt packets */
799 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
800 return;
801
802 map = conn->c_fcong;
803 map_page = 0;
804 map_off = 0;
805
806 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
807 frag_off = 0;
808
809 copied = 0;
810
811 while (copied < RDS_CONG_MAP_BYTES) {
812 __le64 *src, *dst;
813 unsigned int k;
814
815 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
816 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
817
818 addr = kmap_atomic(sg_page(&frag->f_sg));
819
820 src = addr + frag->f_sg.offset + frag_off;
821 dst = (void *)map->m_page_addrs[map_page] + map_off;
822 for (k = 0; k < to_copy; k += 8) {
823 /* Record ports that became uncongested, ie
824 * bits that changed from 0 to 1. */
825 uncongested |= ~(*src) & *dst;
826 *dst++ = *src++;
827 }
828 kunmap_atomic(addr);
829
830 copied += to_copy;
831
832 map_off += to_copy;
833 if (map_off == PAGE_SIZE) {
834 map_off = 0;
835 map_page++;
836 }
837
838 frag_off += to_copy;
839 if (frag_off == RDS_FRAG_SIZE) {
840 frag = list_entry(frag->f_item.next,
841 struct rds_page_frag, f_item);
842 frag_off = 0;
843 }
844 }
845
846 /* the congestion map is in little endian order */
847 rds_cong_map_updated(map, le64_to_cpu(uncongested));
848}
849
850static void rds_ib_process_recv(struct rds_connection *conn,
851 struct rds_ib_recv_work *recv, u32 data_len,
852 struct rds_ib_ack_state *state)
853{
854 struct rds_ib_connection *ic = conn->c_transport_data;
855 struct rds_ib_incoming *ibinc = ic->i_ibinc;
856 struct rds_header *ihdr, *hdr;
857 dma_addr_t dma_addr = ic->i_recv_hdrs_dma[recv - ic->i_recvs];
858
859 /* XXX shut down the connection if port 0,0 are seen? */
860
861 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
862 data_len);
863
864 if (data_len < sizeof(struct rds_header)) {
865 rds_ib_conn_error(conn, "incoming message "
866 "from %pI6c didn't include a "
867 "header, disconnecting and "
868 "reconnecting\n",
869 &conn->c_faddr);
870 return;
871 }
872 data_len -= sizeof(struct rds_header);
873
874 ihdr = ic->i_recv_hdrs[recv - ic->i_recvs];
875
876 ib_dma_sync_single_for_cpu(ic->rds_ibdev->dev, dma_addr,
877 sizeof(*ihdr), DMA_FROM_DEVICE);
878 /* Validate the checksum. */
879 if (!rds_message_verify_checksum(ihdr)) {
880 rds_ib_conn_error(conn, "incoming message "
881 "from %pI6c has corrupted header - "
882 "forcing a reconnect\n",
883 &conn->c_faddr);
884 rds_stats_inc(s_recv_drop_bad_checksum);
885 goto done;
886 }
887
888 /* Process the ACK sequence which comes with every packet */
889 state->ack_recv = be64_to_cpu(ihdr->h_ack);
890 state->ack_recv_valid = 1;
891
892 /* Process the credits update if there was one */
893 if (ihdr->h_credit)
894 rds_ib_send_add_credits(conn, ihdr->h_credit);
895
896 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
897 /* This is an ACK-only packet. The fact that it gets
898 * special treatment here is that historically, ACKs
899 * were rather special beasts.
900 */
901 rds_ib_stats_inc(s_ib_ack_received);
902
903 /*
904 * Usually the frags make their way on to incs and are then freed as
905 * the inc is freed. We don't go that route, so we have to drop the
906 * page ref ourselves. We can't just leave the page on the recv
907 * because that confuses the dma mapping of pages and each recv's use
908 * of a partial page.
909 *
910 * FIXME: Fold this into the code path below.
911 */
912 rds_ib_frag_free(ic, recv->r_frag);
913 recv->r_frag = NULL;
914 goto done;
915 }
916
917 /*
918 * If we don't already have an inc on the connection then this
919 * fragment has a header and starts a message.. copy its header
920 * into the inc and save the inc so we can hang upcoming fragments
921 * off its list.
922 */
923 if (!ibinc) {
924 ibinc = recv->r_ibinc;
925 recv->r_ibinc = NULL;
926 ic->i_ibinc = ibinc;
927
928 hdr = &ibinc->ii_inc.i_hdr;
929 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
930 local_clock();
931 memcpy(hdr, ihdr, sizeof(*hdr));
932 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
933 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
934 local_clock();
935
936 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
937 ic->i_recv_data_rem, hdr->h_flags);
938 } else {
939 hdr = &ibinc->ii_inc.i_hdr;
940 /* We can't just use memcmp here; fragments of a
941 * single message may carry different ACKs */
942 if (hdr->h_sequence != ihdr->h_sequence ||
943 hdr->h_len != ihdr->h_len ||
944 hdr->h_sport != ihdr->h_sport ||
945 hdr->h_dport != ihdr->h_dport) {
946 rds_ib_conn_error(conn,
947 "fragment header mismatch; forcing reconnect\n");
948 goto done;
949 }
950 }
951
952 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
953 recv->r_frag = NULL;
954
955 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
956 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
957 else {
958 ic->i_recv_data_rem = 0;
959 ic->i_ibinc = NULL;
960
961 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
962 rds_ib_cong_recv(conn, ibinc);
963 } else {
964 rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
965 &ibinc->ii_inc, GFP_ATOMIC);
966 state->ack_next = be64_to_cpu(hdr->h_sequence);
967 state->ack_next_valid = 1;
968 }
969
970 /* Evaluate the ACK_REQUIRED flag *after* we received
971 * the complete frame, and after bumping the next_rx
972 * sequence. */
973 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
974 rds_stats_inc(s_recv_ack_required);
975 state->ack_required = 1;
976 }
977
978 rds_inc_put(&ibinc->ii_inc);
979 }
980done:
981 ib_dma_sync_single_for_device(ic->rds_ibdev->dev, dma_addr,
982 sizeof(*ihdr), DMA_FROM_DEVICE);
983}
984
985void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
986 struct ib_wc *wc,
987 struct rds_ib_ack_state *state)
988{
989 struct rds_connection *conn = ic->conn;
990 struct rds_ib_recv_work *recv;
991
992 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
993 (unsigned long long)wc->wr_id, wc->status,
994 ib_wc_status_msg(wc->status), wc->byte_len,
995 be32_to_cpu(wc->ex.imm_data));
996
997 rds_ib_stats_inc(s_ib_rx_cq_event);
998 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
999 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
1000 DMA_FROM_DEVICE);
1001
1002 /* Also process recvs in connecting state because it is possible
1003 * to get a recv completion _before_ the rdmacm ESTABLISHED
1004 * event is processed.
1005 */
1006 if (wc->status == IB_WC_SUCCESS) {
1007 rds_ib_process_recv(conn, recv, wc->byte_len, state);
1008 } else {
1009 /* We expect errors as the qp is drained during shutdown */
1010 if (rds_conn_up(conn) || rds_conn_connecting(conn))
1011 rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), vendor err 0x%x, disconnecting and reconnecting\n",
1012 &conn->c_laddr, &conn->c_faddr,
1013 conn->c_tos, wc->status,
1014 ib_wc_status_msg(wc->status),
1015 wc->vendor_err);
1016 }
1017
1018 /* rds_ib_process_recv() doesn't always consume the frag, and
1019 * we might not have called it at all if the wc didn't indicate
1020 * success. We already unmapped the frag's pages, though, and
1021 * the following rds_ib_ring_free() call tells the refill path
1022 * that it will not find an allocated frag here. Make sure we
1023 * keep that promise by freeing a frag that's still on the ring.
1024 */
1025 if (recv->r_frag) {
1026 rds_ib_frag_free(ic, recv->r_frag);
1027 recv->r_frag = NULL;
1028 }
1029 rds_ib_ring_free(&ic->i_recv_ring, 1);
1030
1031 /* If we ever end up with a really empty receive ring, we're
1032 * in deep trouble, as the sender will definitely see RNR
1033 * timeouts. */
1034 if (rds_ib_ring_empty(&ic->i_recv_ring))
1035 rds_ib_stats_inc(s_ib_rx_ring_empty);
1036
1037 if (rds_ib_ring_low(&ic->i_recv_ring)) {
1038 rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1039 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1040 }
1041}
1042
1043int rds_ib_recv_path(struct rds_conn_path *cp)
1044{
1045 struct rds_connection *conn = cp->cp_conn;
1046 struct rds_ib_connection *ic = conn->c_transport_data;
1047
1048 rdsdebug("conn %p\n", conn);
1049 if (rds_conn_up(conn)) {
1050 rds_ib_attempt_ack(ic);
1051 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1052 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1053 }
1054
1055 return 0;
1056}
1057
1058int rds_ib_recv_init(void)
1059{
1060 struct sysinfo si;
1061 int ret = -ENOMEM;
1062
1063 /* Default to 30% of all available RAM for recv memory */
1064 si_meminfo(&si);
1065 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1066
1067 rds_ib_incoming_slab =
1068 kmem_cache_create_usercopy("rds_ib_incoming",
1069 sizeof(struct rds_ib_incoming),
1070 0, SLAB_HWCACHE_ALIGN,
1071 offsetof(struct rds_ib_incoming,
1072 ii_inc.i_usercopy),
1073 sizeof(struct rds_inc_usercopy),
1074 NULL);
1075 if (!rds_ib_incoming_slab)
1076 goto out;
1077
1078 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1079 sizeof(struct rds_page_frag),
1080 0, SLAB_HWCACHE_ALIGN, NULL);
1081 if (!rds_ib_frag_slab) {
1082 kmem_cache_destroy(rds_ib_incoming_slab);
1083 rds_ib_incoming_slab = NULL;
1084 } else
1085 ret = 0;
1086out:
1087 return ret;
1088}
1089
1090void rds_ib_recv_exit(void)
1091{
1092 WARN_ON(atomic_read(&rds_ib_allocation));
1093
1094 kmem_cache_destroy(rds_ib_incoming_slab);
1095 kmem_cache_destroy(rds_ib_frag_slab);
1096}