Merge tag 'reset-for-v5.3' of git://git.pengutronix.de/git/pza/linux into arm/drivers
[sfrench/cifs-2.6.git] / net / rds / ib_recv.c
1 /*
2  * Copyright (c) 2006, 2017 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
43 static struct kmem_cache *rds_ib_incoming_slab;
44 static struct kmem_cache *rds_ib_frag_slab;
45 static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
46
47 void 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 * sizeof(struct rds_header));
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  */
79 static 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
88 static 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
101 static 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
121 int 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
135 static 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
155 void 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 */
186 static void rds_ib_recv_cache_put(struct list_head *new_item,
187                                   struct rds_ib_refill_cache *cache);
188 static 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 */
192 static 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 */
203 void 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
223 static 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
237 void 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
245 static 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
275 static 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
307 static 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_NOWAIT;
314         gfp_t page_mask = GFP_NOWAIT;
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) * sizeof(struct rds_header);
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;
354 out:
355         return ret;
356 }
357
358 static int acquire_refill(struct rds_connection *conn)
359 {
360         return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
361 }
362
363 static 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  */
381 void 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         u32 pos;
389
390         /* the goal here is to just make sure that someone, somewhere
391          * is posting buffers.  If we can't get the refill lock,
392          * let them do their thing
393          */
394         if (!acquire_refill(conn))
395                 return;
396
397         while ((prefill || rds_conn_up(conn)) &&
398                rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
399                 if (pos >= ic->i_recv_ring.w_nr) {
400                         printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
401                                         pos);
402                         break;
403                 }
404
405                 recv = &ic->i_recvs[pos];
406                 ret = rds_ib_recv_refill_one(conn, recv, gfp);
407                 if (ret) {
408                         break;
409                 }
410
411                 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
412                          recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
413                          (long)sg_dma_address(&recv->r_frag->f_sg));
414
415                 /* XXX when can this fail? */
416                 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL);
417                 if (ret) {
418                         rds_ib_conn_error(conn, "recv post on "
419                                "%pI6c returned %d, disconnecting and "
420                                "reconnecting\n", &conn->c_faddr,
421                                ret);
422                         break;
423                 }
424
425                 posted++;
426         }
427
428         /* We're doing flow control - update the window. */
429         if (ic->i_flowctl && posted)
430                 rds_ib_advertise_credits(conn, posted);
431
432         if (ret)
433                 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
434
435         release_refill(conn);
436
437         /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
438          * in this case the ring being low is going to lead to more interrupts
439          * and we can safely let the softirq code take care of it unless the
440          * ring is completely empty.
441          *
442          * if we're called from krdsd, we'll be GFP_KERNEL.  In this case
443          * we might have raced with the softirq code while we had the refill
444          * lock held.  Use rds_ib_ring_low() instead of ring_empty to decide
445          * if we should requeue.
446          */
447         if (rds_conn_up(conn) &&
448             ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
449             rds_ib_ring_empty(&ic->i_recv_ring))) {
450                 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
451         }
452 }
453
454 /*
455  * We want to recycle several types of recv allocations, like incs and frags.
456  * To use this, the *_free() function passes in the ptr to a list_head within
457  * the recyclee, as well as the cache to put it on.
458  *
459  * First, we put the memory on a percpu list. When this reaches a certain size,
460  * We move it to an intermediate non-percpu list in a lockless manner, with some
461  * xchg/compxchg wizardry.
462  *
463  * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
464  * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
465  * list_empty() will return true with one element is actually present.
466  */
467 static void rds_ib_recv_cache_put(struct list_head *new_item,
468                                  struct rds_ib_refill_cache *cache)
469 {
470         unsigned long flags;
471         struct list_head *old, *chpfirst;
472
473         local_irq_save(flags);
474
475         chpfirst = __this_cpu_read(cache->percpu->first);
476         if (!chpfirst)
477                 INIT_LIST_HEAD(new_item);
478         else /* put on front */
479                 list_add_tail(new_item, chpfirst);
480
481         __this_cpu_write(cache->percpu->first, new_item);
482         __this_cpu_inc(cache->percpu->count);
483
484         if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
485                 goto end;
486
487         /*
488          * Return our per-cpu first list to the cache's xfer by atomically
489          * grabbing the current xfer list, appending it to our per-cpu list,
490          * and then atomically returning that entire list back to the
491          * cache's xfer list as long as it's still empty.
492          */
493         do {
494                 old = xchg(&cache->xfer, NULL);
495                 if (old)
496                         list_splice_entire_tail(old, chpfirst);
497                 old = cmpxchg(&cache->xfer, NULL, chpfirst);
498         } while (old);
499
500
501         __this_cpu_write(cache->percpu->first, NULL);
502         __this_cpu_write(cache->percpu->count, 0);
503 end:
504         local_irq_restore(flags);
505 }
506
507 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
508 {
509         struct list_head *head = cache->ready;
510
511         if (head) {
512                 if (!list_empty(head)) {
513                         cache->ready = head->next;
514                         list_del_init(head);
515                 } else
516                         cache->ready = NULL;
517         }
518
519         return head;
520 }
521
522 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
523 {
524         struct rds_ib_incoming *ibinc;
525         struct rds_page_frag *frag;
526         unsigned long to_copy;
527         unsigned long frag_off = 0;
528         int copied = 0;
529         int ret;
530         u32 len;
531
532         ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
533         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
534         len = be32_to_cpu(inc->i_hdr.h_len);
535
536         while (iov_iter_count(to) && copied < len) {
537                 if (frag_off == RDS_FRAG_SIZE) {
538                         frag = list_entry(frag->f_item.next,
539                                           struct rds_page_frag, f_item);
540                         frag_off = 0;
541                 }
542                 to_copy = min_t(unsigned long, iov_iter_count(to),
543                                 RDS_FRAG_SIZE - frag_off);
544                 to_copy = min_t(unsigned long, to_copy, len - copied);
545
546                 /* XXX needs + offset for multiple recvs per page */
547                 rds_stats_add(s_copy_to_user, to_copy);
548                 ret = copy_page_to_iter(sg_page(&frag->f_sg),
549                                         frag->f_sg.offset + frag_off,
550                                         to_copy,
551                                         to);
552                 if (ret != to_copy)
553                         return -EFAULT;
554
555                 frag_off += to_copy;
556                 copied += to_copy;
557         }
558
559         return copied;
560 }
561
562 /* ic starts out kzalloc()ed */
563 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
564 {
565         struct ib_send_wr *wr = &ic->i_ack_wr;
566         struct ib_sge *sge = &ic->i_ack_sge;
567
568         sge->addr = ic->i_ack_dma;
569         sge->length = sizeof(struct rds_header);
570         sge->lkey = ic->i_pd->local_dma_lkey;
571
572         wr->sg_list = sge;
573         wr->num_sge = 1;
574         wr->opcode = IB_WR_SEND;
575         wr->wr_id = RDS_IB_ACK_WR_ID;
576         wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
577 }
578
579 /*
580  * You'd think that with reliable IB connections you wouldn't need to ack
581  * messages that have been received.  The problem is that IB hardware generates
582  * an ack message before it has DMAed the message into memory.  This creates a
583  * potential message loss if the HCA is disabled for any reason between when it
584  * sends the ack and before the message is DMAed and processed.  This is only a
585  * potential issue if another HCA is available for fail-over.
586  *
587  * When the remote host receives our ack they'll free the sent message from
588  * their send queue.  To decrease the latency of this we always send an ack
589  * immediately after we've received messages.
590  *
591  * For simplicity, we only have one ack in flight at a time.  This puts
592  * pressure on senders to have deep enough send queues to absorb the latency of
593  * a single ack frame being in flight.  This might not be good enough.
594  *
595  * This is implemented by have a long-lived send_wr and sge which point to a
596  * statically allocated ack frame.  This ack wr does not fall under the ring
597  * accounting that the tx and rx wrs do.  The QP attribute specifically makes
598  * room for it beyond the ring size.  Send completion notices its special
599  * wr_id and avoids working with the ring in that case.
600  */
601 #ifndef KERNEL_HAS_ATOMIC64
602 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
603 {
604         unsigned long flags;
605
606         spin_lock_irqsave(&ic->i_ack_lock, flags);
607         ic->i_ack_next = seq;
608         if (ack_required)
609                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
610         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
611 }
612
613 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
614 {
615         unsigned long flags;
616         u64 seq;
617
618         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
619
620         spin_lock_irqsave(&ic->i_ack_lock, flags);
621         seq = ic->i_ack_next;
622         spin_unlock_irqrestore(&ic->i_ack_lock, flags);
623
624         return seq;
625 }
626 #else
627 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
628 {
629         atomic64_set(&ic->i_ack_next, seq);
630         if (ack_required) {
631                 smp_mb__before_atomic();
632                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
633         }
634 }
635
636 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
637 {
638         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
639         smp_mb__after_atomic();
640
641         return atomic64_read(&ic->i_ack_next);
642 }
643 #endif
644
645
646 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
647 {
648         struct rds_header *hdr = ic->i_ack;
649         u64 seq;
650         int ret;
651
652         seq = rds_ib_get_ack(ic);
653
654         rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
655         rds_message_populate_header(hdr, 0, 0, 0);
656         hdr->h_ack = cpu_to_be64(seq);
657         hdr->h_credit = adv_credits;
658         rds_message_make_checksum(hdr);
659         ic->i_ack_queued = jiffies;
660
661         ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL);
662         if (unlikely(ret)) {
663                 /* Failed to send. Release the WR, and
664                  * force another ACK.
665                  */
666                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
667                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
668
669                 rds_ib_stats_inc(s_ib_ack_send_failure);
670
671                 rds_ib_conn_error(ic->conn, "sending ack failed\n");
672         } else
673                 rds_ib_stats_inc(s_ib_ack_sent);
674 }
675
676 /*
677  * There are 3 ways of getting acknowledgements to the peer:
678  *  1.  We call rds_ib_attempt_ack from the recv completion handler
679  *      to send an ACK-only frame.
680  *      However, there can be only one such frame in the send queue
681  *      at any time, so we may have to postpone it.
682  *  2.  When another (data) packet is transmitted while there's
683  *      an ACK in the queue, we piggyback the ACK sequence number
684  *      on the data packet.
685  *  3.  If the ACK WR is done sending, we get called from the
686  *      send queue completion handler, and check whether there's
687  *      another ACK pending (postponed because the WR was on the
688  *      queue). If so, we transmit it.
689  *
690  * We maintain 2 variables:
691  *  -   i_ack_flags, which keeps track of whether the ACK WR
692  *      is currently in the send queue or not (IB_ACK_IN_FLIGHT)
693  *  -   i_ack_next, which is the last sequence number we received
694  *
695  * Potentially, send queue and receive queue handlers can run concurrently.
696  * It would be nice to not have to use a spinlock to synchronize things,
697  * but the one problem that rules this out is that 64bit updates are
698  * not atomic on all platforms. Things would be a lot simpler if
699  * we had atomic64 or maybe cmpxchg64 everywhere.
700  *
701  * Reconnecting complicates this picture just slightly. When we
702  * reconnect, we may be seeing duplicate packets. The peer
703  * is retransmitting them, because it hasn't seen an ACK for
704  * them. It is important that we ACK these.
705  *
706  * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
707  * this flag set *MUST* be acknowledged immediately.
708  */
709
710 /*
711  * When we get here, we're called from the recv queue handler.
712  * Check whether we ought to transmit an ACK.
713  */
714 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
715 {
716         unsigned int adv_credits;
717
718         if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
719                 return;
720
721         if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
722                 rds_ib_stats_inc(s_ib_ack_send_delayed);
723                 return;
724         }
725
726         /* Can we get a send credit? */
727         if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
728                 rds_ib_stats_inc(s_ib_tx_throttle);
729                 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
730                 return;
731         }
732
733         clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
734         rds_ib_send_ack(ic, adv_credits);
735 }
736
737 /*
738  * We get here from the send completion handler, when the
739  * adapter tells us the ACK frame was sent.
740  */
741 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
742 {
743         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
744         rds_ib_attempt_ack(ic);
745 }
746
747 /*
748  * This is called by the regular xmit code when it wants to piggyback
749  * an ACK on an outgoing frame.
750  */
751 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
752 {
753         if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
754                 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
755         return rds_ib_get_ack(ic);
756 }
757
758 /*
759  * It's kind of lame that we're copying from the posted receive pages into
760  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
761  * them.  But receiving new congestion bitmaps should be a *rare* event, so
762  * hopefully we won't need to invest that complexity in making it more
763  * efficient.  By copying we can share a simpler core with TCP which has to
764  * copy.
765  */
766 static void rds_ib_cong_recv(struct rds_connection *conn,
767                               struct rds_ib_incoming *ibinc)
768 {
769         struct rds_cong_map *map;
770         unsigned int map_off;
771         unsigned int map_page;
772         struct rds_page_frag *frag;
773         unsigned long frag_off;
774         unsigned long to_copy;
775         unsigned long copied;
776         __le64 uncongested = 0;
777         void *addr;
778
779         /* catch completely corrupt packets */
780         if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
781                 return;
782
783         map = conn->c_fcong;
784         map_page = 0;
785         map_off = 0;
786
787         frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
788         frag_off = 0;
789
790         copied = 0;
791
792         while (copied < RDS_CONG_MAP_BYTES) {
793                 __le64 *src, *dst;
794                 unsigned int k;
795
796                 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
797                 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
798
799                 addr = kmap_atomic(sg_page(&frag->f_sg));
800
801                 src = addr + frag->f_sg.offset + frag_off;
802                 dst = (void *)map->m_page_addrs[map_page] + map_off;
803                 for (k = 0; k < to_copy; k += 8) {
804                         /* Record ports that became uncongested, ie
805                          * bits that changed from 0 to 1. */
806                         uncongested |= ~(*src) & *dst;
807                         *dst++ = *src++;
808                 }
809                 kunmap_atomic(addr);
810
811                 copied += to_copy;
812
813                 map_off += to_copy;
814                 if (map_off == PAGE_SIZE) {
815                         map_off = 0;
816                         map_page++;
817                 }
818
819                 frag_off += to_copy;
820                 if (frag_off == RDS_FRAG_SIZE) {
821                         frag = list_entry(frag->f_item.next,
822                                           struct rds_page_frag, f_item);
823                         frag_off = 0;
824                 }
825         }
826
827         /* the congestion map is in little endian order */
828         rds_cong_map_updated(map, le64_to_cpu(uncongested));
829 }
830
831 static void rds_ib_process_recv(struct rds_connection *conn,
832                                 struct rds_ib_recv_work *recv, u32 data_len,
833                                 struct rds_ib_ack_state *state)
834 {
835         struct rds_ib_connection *ic = conn->c_transport_data;
836         struct rds_ib_incoming *ibinc = ic->i_ibinc;
837         struct rds_header *ihdr, *hdr;
838
839         /* XXX shut down the connection if port 0,0 are seen? */
840
841         rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
842                  data_len);
843
844         if (data_len < sizeof(struct rds_header)) {
845                 rds_ib_conn_error(conn, "incoming message "
846                        "from %pI6c didn't include a "
847                        "header, disconnecting and "
848                        "reconnecting\n",
849                        &conn->c_faddr);
850                 return;
851         }
852         data_len -= sizeof(struct rds_header);
853
854         ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
855
856         /* Validate the checksum. */
857         if (!rds_message_verify_checksum(ihdr)) {
858                 rds_ib_conn_error(conn, "incoming message "
859                        "from %pI6c has corrupted header - "
860                        "forcing a reconnect\n",
861                        &conn->c_faddr);
862                 rds_stats_inc(s_recv_drop_bad_checksum);
863                 return;
864         }
865
866         /* Process the ACK sequence which comes with every packet */
867         state->ack_recv = be64_to_cpu(ihdr->h_ack);
868         state->ack_recv_valid = 1;
869
870         /* Process the credits update if there was one */
871         if (ihdr->h_credit)
872                 rds_ib_send_add_credits(conn, ihdr->h_credit);
873
874         if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
875                 /* This is an ACK-only packet. The fact that it gets
876                  * special treatment here is that historically, ACKs
877                  * were rather special beasts.
878                  */
879                 rds_ib_stats_inc(s_ib_ack_received);
880
881                 /*
882                  * Usually the frags make their way on to incs and are then freed as
883                  * the inc is freed.  We don't go that route, so we have to drop the
884                  * page ref ourselves.  We can't just leave the page on the recv
885                  * because that confuses the dma mapping of pages and each recv's use
886                  * of a partial page.
887                  *
888                  * FIXME: Fold this into the code path below.
889                  */
890                 rds_ib_frag_free(ic, recv->r_frag);
891                 recv->r_frag = NULL;
892                 return;
893         }
894
895         /*
896          * If we don't already have an inc on the connection then this
897          * fragment has a header and starts a message.. copy its header
898          * into the inc and save the inc so we can hang upcoming fragments
899          * off its list.
900          */
901         if (!ibinc) {
902                 ibinc = recv->r_ibinc;
903                 recv->r_ibinc = NULL;
904                 ic->i_ibinc = ibinc;
905
906                 hdr = &ibinc->ii_inc.i_hdr;
907                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
908                                 local_clock();
909                 memcpy(hdr, ihdr, sizeof(*hdr));
910                 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
911                 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
912                                 local_clock();
913
914                 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
915                          ic->i_recv_data_rem, hdr->h_flags);
916         } else {
917                 hdr = &ibinc->ii_inc.i_hdr;
918                 /* We can't just use memcmp here; fragments of a
919                  * single message may carry different ACKs */
920                 if (hdr->h_sequence != ihdr->h_sequence ||
921                     hdr->h_len != ihdr->h_len ||
922                     hdr->h_sport != ihdr->h_sport ||
923                     hdr->h_dport != ihdr->h_dport) {
924                         rds_ib_conn_error(conn,
925                                 "fragment header mismatch; forcing reconnect\n");
926                         return;
927                 }
928         }
929
930         list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
931         recv->r_frag = NULL;
932
933         if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
934                 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
935         else {
936                 ic->i_recv_data_rem = 0;
937                 ic->i_ibinc = NULL;
938
939                 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) {
940                         rds_ib_cong_recv(conn, ibinc);
941                 } else {
942                         rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr,
943                                           &ibinc->ii_inc, GFP_ATOMIC);
944                         state->ack_next = be64_to_cpu(hdr->h_sequence);
945                         state->ack_next_valid = 1;
946                 }
947
948                 /* Evaluate the ACK_REQUIRED flag *after* we received
949                  * the complete frame, and after bumping the next_rx
950                  * sequence. */
951                 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
952                         rds_stats_inc(s_recv_ack_required);
953                         state->ack_required = 1;
954                 }
955
956                 rds_inc_put(&ibinc->ii_inc);
957         }
958 }
959
960 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
961                              struct ib_wc *wc,
962                              struct rds_ib_ack_state *state)
963 {
964         struct rds_connection *conn = ic->conn;
965         struct rds_ib_recv_work *recv;
966
967         rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
968                  (unsigned long long)wc->wr_id, wc->status,
969                  ib_wc_status_msg(wc->status), wc->byte_len,
970                  be32_to_cpu(wc->ex.imm_data));
971
972         rds_ib_stats_inc(s_ib_rx_cq_event);
973         recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
974         ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
975                         DMA_FROM_DEVICE);
976
977         /* Also process recvs in connecting state because it is possible
978          * to get a recv completion _before_ the rdmacm ESTABLISHED
979          * event is processed.
980          */
981         if (wc->status == IB_WC_SUCCESS) {
982                 rds_ib_process_recv(conn, recv, wc->byte_len, state);
983         } else {
984                 /* We expect errors as the qp is drained during shutdown */
985                 if (rds_conn_up(conn) || rds_conn_connecting(conn))
986                         rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), disconnecting and reconnecting\n",
987                                           &conn->c_laddr, &conn->c_faddr,
988                                           conn->c_tos, wc->status,
989                                           ib_wc_status_msg(wc->status));
990         }
991
992         /* rds_ib_process_recv() doesn't always consume the frag, and
993          * we might not have called it at all if the wc didn't indicate
994          * success. We already unmapped the frag's pages, though, and
995          * the following rds_ib_ring_free() call tells the refill path
996          * that it will not find an allocated frag here. Make sure we
997          * keep that promise by freeing a frag that's still on the ring.
998          */
999         if (recv->r_frag) {
1000                 rds_ib_frag_free(ic, recv->r_frag);
1001                 recv->r_frag = NULL;
1002         }
1003         rds_ib_ring_free(&ic->i_recv_ring, 1);
1004
1005         /* If we ever end up with a really empty receive ring, we're
1006          * in deep trouble, as the sender will definitely see RNR
1007          * timeouts. */
1008         if (rds_ib_ring_empty(&ic->i_recv_ring))
1009                 rds_ib_stats_inc(s_ib_rx_ring_empty);
1010
1011         if (rds_ib_ring_low(&ic->i_recv_ring)) {
1012                 rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1013                 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1014         }
1015 }
1016
1017 int rds_ib_recv_path(struct rds_conn_path *cp)
1018 {
1019         struct rds_connection *conn = cp->cp_conn;
1020         struct rds_ib_connection *ic = conn->c_transport_data;
1021
1022         rdsdebug("conn %p\n", conn);
1023         if (rds_conn_up(conn)) {
1024                 rds_ib_attempt_ack(ic);
1025                 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1026                 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1027         }
1028
1029         return 0;
1030 }
1031
1032 int rds_ib_recv_init(void)
1033 {
1034         struct sysinfo si;
1035         int ret = -ENOMEM;
1036
1037         /* Default to 30% of all available RAM for recv memory */
1038         si_meminfo(&si);
1039         rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1040
1041         rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1042                                         sizeof(struct rds_ib_incoming),
1043                                         0, SLAB_HWCACHE_ALIGN, NULL);
1044         if (!rds_ib_incoming_slab)
1045                 goto out;
1046
1047         rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1048                                         sizeof(struct rds_page_frag),
1049                                         0, SLAB_HWCACHE_ALIGN, NULL);
1050         if (!rds_ib_frag_slab) {
1051                 kmem_cache_destroy(rds_ib_incoming_slab);
1052                 rds_ib_incoming_slab = NULL;
1053         } else
1054                 ret = 0;
1055 out:
1056         return ret;
1057 }
1058
1059 void rds_ib_recv_exit(void)
1060 {
1061         WARN_ON(atomic_read(&rds_ib_allocation));
1062
1063         kmem_cache_destroy(rds_ib_incoming_slab);
1064         kmem_cache_destroy(rds_ib_frag_slab);
1065 }