2 * Copyright (c) 2006 Oracle. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
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>
42 static struct kmem_cache *rds_ib_incoming_slab;
43 static struct kmem_cache *rds_ib_frag_slab;
44 static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
46 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
48 struct rds_ib_recv_work *recv;
51 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
57 recv->r_wr.next = NULL;
59 recv->r_wr.sg_list = recv->r_sge;
60 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
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;
67 sge = &recv->r_sge[1];
69 sge->length = RDS_FRAG_SIZE;
70 sge->lkey = ic->i_mr->lkey;
75 * The entire 'from' list, including the from element itself, is put on
76 * to the tail of the 'to' list.
78 static void list_splice_entire_tail(struct list_head *from,
81 struct list_head *from_last = from->prev;
83 list_splice_tail(from_last, to);
84 list_add_tail(from_last, to);
87 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
89 struct list_head *tmp;
91 tmp = xchg(&cache->xfer, NULL);
94 list_splice_entire_tail(tmp, cache->ready);
100 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
102 struct rds_ib_cache_head *head;
105 cache->percpu = alloc_percpu(struct rds_ib_cache_head);
109 for_each_possible_cpu(cpu) {
110 head = per_cpu_ptr(cache->percpu, cpu);
120 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
124 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
128 free_percpu(ic->i_cache_incs.percpu);
134 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
135 struct list_head *caller_list)
137 struct rds_ib_cache_head *head;
140 for_each_possible_cpu(cpu) {
141 head = per_cpu_ptr(cache->percpu, cpu);
143 list_splice_entire_tail(head->first, caller_list);
149 list_splice_entire_tail(cache->ready, caller_list);
154 void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
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;
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);
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);
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);
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);
184 static void rds_ib_recv_cache_put(struct list_head *new_item,
185 struct rds_ib_refill_cache *cache);
186 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
189 /* Recycle frag and attached recv buffer f_sg */
190 static void rds_ib_frag_free(struct rds_ib_connection *ic,
191 struct rds_page_frag *frag)
193 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
195 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
198 /* Recycle inc after freeing attached frags */
199 void rds_ib_inc_free(struct rds_incoming *inc)
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;
206 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
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);
213 BUG_ON(!list_empty(&ibinc->ii_frags));
215 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
216 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
219 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
220 struct rds_ib_recv_work *recv)
223 rds_inc_put(&recv->r_ibinc->ii_inc);
224 recv->r_ibinc = NULL;
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);
233 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
237 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
238 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
241 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
244 struct rds_ib_incoming *ibinc;
245 struct list_head *cache_item;
248 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
250 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
252 avail_allocs = atomic_add_unless(&rds_ib_allocation,
253 1, rds_ib_sysctl_max_recv_allocation);
255 rds_ib_stats_inc(s_ib_rx_alloc_limit);
258 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
260 atomic_dec(&rds_ib_allocation);
264 INIT_LIST_HEAD(&ibinc->ii_frags);
265 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
270 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
271 gfp_t slab_mask, gfp_t page_mask)
273 struct rds_page_frag *frag;
274 struct list_head *cache_item;
277 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
279 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
281 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
285 sg_init_table(&frag->f_sg, 1);
286 ret = rds_page_remainder_alloc(&frag->f_sg,
287 RDS_FRAG_SIZE, page_mask);
289 kmem_cache_free(rds_ib_frag_slab, frag);
294 INIT_LIST_HEAD(&frag->f_item);
299 static int rds_ib_recv_refill_one(struct rds_connection *conn,
300 struct rds_ib_recv_work *recv, int prefill)
302 struct rds_ib_connection *ic = conn->c_transport_data;
305 gfp_t slab_mask = GFP_NOWAIT;
306 gfp_t page_mask = GFP_NOWAIT;
309 slab_mask = GFP_KERNEL;
310 page_mask = GFP_HIGHUSER;
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);
319 * ibinc was taken from recv if recv contained the start of a message.
320 * recvs that were continuations will still have this allocated.
322 if (!recv->r_ibinc) {
323 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
328 WARN_ON(recv->r_frag); /* leak! */
329 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
333 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
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);
341 sge = &recv->r_sge[1];
342 sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg);
343 sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg);
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
355 * -1 is returned if posting fails due to temporary resource exhaustion.
357 void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
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;
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",
374 recv = &ic->i_recvs[pos];
375 ret = rds_ib_recv_refill_one(conn, recv, prefill);
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) ib_sg_dma_address(
386 &recv->r_frag->f_sg),
389 rds_ib_conn_error(conn, "recv post on "
390 "%pI4 returned %d, disconnecting and "
391 "reconnecting\n", &conn->c_faddr,
399 /* We're doing flow control - update the window. */
400 if (ic->i_flowctl && posted)
401 rds_ib_advertise_credits(conn, posted);
404 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
408 * We want to recycle several types of recv allocations, like incs and frags.
409 * To use this, the *_free() function passes in the ptr to a list_head within
410 * the recyclee, as well as the cache to put it on.
412 * First, we put the memory on a percpu list. When this reaches a certain size,
413 * We move it to an intermediate non-percpu list in a lockless manner, with some
414 * xchg/compxchg wizardry.
416 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
417 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
418 * list_empty() will return true with one element is actually present.
420 static void rds_ib_recv_cache_put(struct list_head *new_item,
421 struct rds_ib_refill_cache *cache)
424 struct list_head *old;
425 struct list_head __percpu *chpfirst;
427 local_irq_save(flags);
429 chpfirst = __this_cpu_read(cache->percpu->first);
431 INIT_LIST_HEAD(new_item);
432 else /* put on front */
433 list_add_tail(new_item, chpfirst);
435 __this_cpu_write(chpfirst, new_item);
436 __this_cpu_inc(cache->percpu->count);
438 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
442 * Return our per-cpu first list to the cache's xfer by atomically
443 * grabbing the current xfer list, appending it to our per-cpu list,
444 * and then atomically returning that entire list back to the
445 * cache's xfer list as long as it's still empty.
448 old = xchg(&cache->xfer, NULL);
450 list_splice_entire_tail(old, chpfirst);
451 old = cmpxchg(&cache->xfer, NULL, chpfirst);
455 __this_cpu_write(chpfirst, NULL);
456 __this_cpu_write(cache->percpu->count, 0);
458 local_irq_restore(flags);
461 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
463 struct list_head *head = cache->ready;
466 if (!list_empty(head)) {
467 cache->ready = head->next;
476 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
479 struct rds_ib_incoming *ibinc;
480 struct rds_page_frag *frag;
481 struct iovec *iov = first_iov;
482 unsigned long to_copy;
483 unsigned long frag_off = 0;
484 unsigned long iov_off = 0;
489 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
490 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
491 len = be32_to_cpu(inc->i_hdr.h_len);
493 while (copied < size && copied < len) {
494 if (frag_off == RDS_FRAG_SIZE) {
495 frag = list_entry(frag->f_item.next,
496 struct rds_page_frag, f_item);
499 while (iov_off == iov->iov_len) {
504 to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
505 to_copy = min_t(size_t, to_copy, size - copied);
506 to_copy = min_t(unsigned long, to_copy, len - copied);
508 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
510 to_copy, iov->iov_base, iov->iov_len, iov_off,
511 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);
513 /* XXX needs + offset for multiple recvs per page */
514 ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
515 frag->f_sg.offset + frag_off,
516 iov->iov_base + iov_off,
531 /* ic starts out kzalloc()ed */
532 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
534 struct ib_send_wr *wr = &ic->i_ack_wr;
535 struct ib_sge *sge = &ic->i_ack_sge;
537 sge->addr = ic->i_ack_dma;
538 sge->length = sizeof(struct rds_header);
539 sge->lkey = ic->i_mr->lkey;
543 wr->opcode = IB_WR_SEND;
544 wr->wr_id = RDS_IB_ACK_WR_ID;
545 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
549 * You'd think that with reliable IB connections you wouldn't need to ack
550 * messages that have been received. The problem is that IB hardware generates
551 * an ack message before it has DMAed the message into memory. This creates a
552 * potential message loss if the HCA is disabled for any reason between when it
553 * sends the ack and before the message is DMAed and processed. This is only a
554 * potential issue if another HCA is available for fail-over.
556 * When the remote host receives our ack they'll free the sent message from
557 * their send queue. To decrease the latency of this we always send an ack
558 * immediately after we've received messages.
560 * For simplicity, we only have one ack in flight at a time. This puts
561 * pressure on senders to have deep enough send queues to absorb the latency of
562 * a single ack frame being in flight. This might not be good enough.
564 * This is implemented by have a long-lived send_wr and sge which point to a
565 * statically allocated ack frame. This ack wr does not fall under the ring
566 * accounting that the tx and rx wrs do. The QP attribute specifically makes
567 * room for it beyond the ring size. Send completion notices its special
568 * wr_id and avoids working with the ring in that case.
570 #ifndef KERNEL_HAS_ATOMIC64
571 static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
576 spin_lock_irqsave(&ic->i_ack_lock, flags);
577 ic->i_ack_next = seq;
579 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
580 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
583 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
588 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
590 spin_lock_irqsave(&ic->i_ack_lock, flags);
591 seq = ic->i_ack_next;
592 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
597 static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
600 atomic64_set(&ic->i_ack_next, seq);
602 smp_mb__before_clear_bit();
603 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
607 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
609 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
610 smp_mb__after_clear_bit();
612 return atomic64_read(&ic->i_ack_next);
617 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
619 struct rds_header *hdr = ic->i_ack;
620 struct ib_send_wr *failed_wr;
624 seq = rds_ib_get_ack(ic);
626 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
627 rds_message_populate_header(hdr, 0, 0, 0);
628 hdr->h_ack = cpu_to_be64(seq);
629 hdr->h_credit = adv_credits;
630 rds_message_make_checksum(hdr);
631 ic->i_ack_queued = jiffies;
633 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
635 /* Failed to send. Release the WR, and
638 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
639 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
641 rds_ib_stats_inc(s_ib_ack_send_failure);
643 rds_ib_conn_error(ic->conn, "sending ack failed\n");
645 rds_ib_stats_inc(s_ib_ack_sent);
649 * There are 3 ways of getting acknowledgements to the peer:
650 * 1. We call rds_ib_attempt_ack from the recv completion handler
651 * to send an ACK-only frame.
652 * However, there can be only one such frame in the send queue
653 * at any time, so we may have to postpone it.
654 * 2. When another (data) packet is transmitted while there's
655 * an ACK in the queue, we piggyback the ACK sequence number
656 * on the data packet.
657 * 3. If the ACK WR is done sending, we get called from the
658 * send queue completion handler, and check whether there's
659 * another ACK pending (postponed because the WR was on the
660 * queue). If so, we transmit it.
662 * We maintain 2 variables:
663 * - i_ack_flags, which keeps track of whether the ACK WR
664 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
665 * - i_ack_next, which is the last sequence number we received
667 * Potentially, send queue and receive queue handlers can run concurrently.
668 * It would be nice to not have to use a spinlock to synchronize things,
669 * but the one problem that rules this out is that 64bit updates are
670 * not atomic on all platforms. Things would be a lot simpler if
671 * we had atomic64 or maybe cmpxchg64 everywhere.
673 * Reconnecting complicates this picture just slightly. When we
674 * reconnect, we may be seeing duplicate packets. The peer
675 * is retransmitting them, because it hasn't seen an ACK for
676 * them. It is important that we ACK these.
678 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
679 * this flag set *MUST* be acknowledged immediately.
683 * When we get here, we're called from the recv queue handler.
684 * Check whether we ought to transmit an ACK.
686 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
688 unsigned int adv_credits;
690 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
693 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
694 rds_ib_stats_inc(s_ib_ack_send_delayed);
698 /* Can we get a send credit? */
699 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
700 rds_ib_stats_inc(s_ib_tx_throttle);
701 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
705 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
706 rds_ib_send_ack(ic, adv_credits);
710 * We get here from the send completion handler, when the
711 * adapter tells us the ACK frame was sent.
713 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
715 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
716 rds_ib_attempt_ack(ic);
720 * This is called by the regular xmit code when it wants to piggyback
721 * an ACK on an outgoing frame.
723 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
725 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
726 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
727 return rds_ib_get_ack(ic);
731 * It's kind of lame that we're copying from the posted receive pages into
732 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
733 * them. But receiving new congestion bitmaps should be a *rare* event, so
734 * hopefully we won't need to invest that complexity in making it more
735 * efficient. By copying we can share a simpler core with TCP which has to
738 static void rds_ib_cong_recv(struct rds_connection *conn,
739 struct rds_ib_incoming *ibinc)
741 struct rds_cong_map *map;
742 unsigned int map_off;
743 unsigned int map_page;
744 struct rds_page_frag *frag;
745 unsigned long frag_off;
746 unsigned long to_copy;
747 unsigned long copied;
748 uint64_t uncongested = 0;
751 /* catch completely corrupt packets */
752 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
759 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
764 while (copied < RDS_CONG_MAP_BYTES) {
768 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
769 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
771 addr = kmap_atomic(sg_page(&frag->f_sg));
773 src = addr + frag_off;
774 dst = (void *)map->m_page_addrs[map_page] + map_off;
775 for (k = 0; k < to_copy; k += 8) {
776 /* Record ports that became uncongested, ie
777 * bits that changed from 0 to 1. */
778 uncongested |= ~(*src) & *dst;
786 if (map_off == PAGE_SIZE) {
792 if (frag_off == RDS_FRAG_SIZE) {
793 frag = list_entry(frag->f_item.next,
794 struct rds_page_frag, f_item);
799 /* the congestion map is in little endian order */
800 uncongested = le64_to_cpu(uncongested);
802 rds_cong_map_updated(map, uncongested);
806 * Rings are posted with all the allocations they'll need to queue the
807 * incoming message to the receiving socket so this can't fail.
808 * All fragments start with a header, so we can make sure we're not receiving
809 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
811 struct rds_ib_ack_state {
814 unsigned int ack_required:1;
815 unsigned int ack_next_valid:1;
816 unsigned int ack_recv_valid:1;
819 static void rds_ib_process_recv(struct rds_connection *conn,
820 struct rds_ib_recv_work *recv, u32 data_len,
821 struct rds_ib_ack_state *state)
823 struct rds_ib_connection *ic = conn->c_transport_data;
824 struct rds_ib_incoming *ibinc = ic->i_ibinc;
825 struct rds_header *ihdr, *hdr;
827 /* XXX shut down the connection if port 0,0 are seen? */
829 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
832 if (data_len < sizeof(struct rds_header)) {
833 rds_ib_conn_error(conn, "incoming message "
834 "from %pI4 didn't include a "
835 "header, disconnecting and "
840 data_len -= sizeof(struct rds_header);
842 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
844 /* Validate the checksum. */
845 if (!rds_message_verify_checksum(ihdr)) {
846 rds_ib_conn_error(conn, "incoming message "
847 "from %pI4 has corrupted header - "
848 "forcing a reconnect\n",
850 rds_stats_inc(s_recv_drop_bad_checksum);
854 /* Process the ACK sequence which comes with every packet */
855 state->ack_recv = be64_to_cpu(ihdr->h_ack);
856 state->ack_recv_valid = 1;
858 /* Process the credits update if there was one */
860 rds_ib_send_add_credits(conn, ihdr->h_credit);
862 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
863 /* This is an ACK-only packet. The fact that it gets
864 * special treatment here is that historically, ACKs
865 * were rather special beasts.
867 rds_ib_stats_inc(s_ib_ack_received);
870 * Usually the frags make their way on to incs and are then freed as
871 * the inc is freed. We don't go that route, so we have to drop the
872 * page ref ourselves. We can't just leave the page on the recv
873 * because that confuses the dma mapping of pages and each recv's use
876 * FIXME: Fold this into the code path below.
878 rds_ib_frag_free(ic, recv->r_frag);
884 * If we don't already have an inc on the connection then this
885 * fragment has a header and starts a message.. copy its header
886 * into the inc and save the inc so we can hang upcoming fragments
890 ibinc = recv->r_ibinc;
891 recv->r_ibinc = NULL;
894 hdr = &ibinc->ii_inc.i_hdr;
895 memcpy(hdr, ihdr, sizeof(*hdr));
896 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
898 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
899 ic->i_recv_data_rem, hdr->h_flags);
901 hdr = &ibinc->ii_inc.i_hdr;
902 /* We can't just use memcmp here; fragments of a
903 * single message may carry different ACKs */
904 if (hdr->h_sequence != ihdr->h_sequence ||
905 hdr->h_len != ihdr->h_len ||
906 hdr->h_sport != ihdr->h_sport ||
907 hdr->h_dport != ihdr->h_dport) {
908 rds_ib_conn_error(conn,
909 "fragment header mismatch; forcing reconnect\n");
914 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
917 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
918 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
920 ic->i_recv_data_rem = 0;
923 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
924 rds_ib_cong_recv(conn, ibinc);
926 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
927 &ibinc->ii_inc, GFP_ATOMIC);
928 state->ack_next = be64_to_cpu(hdr->h_sequence);
929 state->ack_next_valid = 1;
932 /* Evaluate the ACK_REQUIRED flag *after* we received
933 * the complete frame, and after bumping the next_rx
935 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
936 rds_stats_inc(s_recv_ack_required);
937 state->ack_required = 1;
940 rds_inc_put(&ibinc->ii_inc);
945 * Plucking the oldest entry from the ring can be done concurrently with
946 * the thread refilling the ring. Each ring operation is protected by
947 * spinlocks and the transient state of refilling doesn't change the
948 * recording of which entry is oldest.
950 * This relies on IB only calling one cq comp_handler for each cq so that
951 * there will only be one caller of rds_recv_incoming() per RDS connection.
953 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
955 struct rds_connection *conn = context;
956 struct rds_ib_connection *ic = conn->c_transport_data;
958 rdsdebug("conn %p cq %p\n", conn, cq);
960 rds_ib_stats_inc(s_ib_rx_cq_call);
962 tasklet_schedule(&ic->i_recv_tasklet);
965 static inline void rds_poll_cq(struct rds_ib_connection *ic,
966 struct rds_ib_ack_state *state)
968 struct rds_connection *conn = ic->conn;
970 struct rds_ib_recv_work *recv;
972 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
973 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
974 (unsigned long long)wc.wr_id, wc.status,
975 rds_ib_wc_status_str(wc.status), wc.byte_len,
976 be32_to_cpu(wc.ex.imm_data));
977 rds_ib_stats_inc(s_ib_rx_cq_event);
979 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
981 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
984 * Also process recvs in connecting state because it is possible
985 * to get a recv completion _before_ the rdmacm ESTABLISHED
986 * event is processed.
988 if (wc.status == IB_WC_SUCCESS) {
989 rds_ib_process_recv(conn, recv, wc.byte_len, state);
991 /* We expect errors as the qp is drained during shutdown */
992 if (rds_conn_up(conn) || rds_conn_connecting(conn))
993 rds_ib_conn_error(conn, "recv completion on %pI4 had "
994 "status %u (%s), disconnecting and "
995 "reconnecting\n", &conn->c_faddr,
997 rds_ib_wc_status_str(wc.status));
1001 * It's very important that we only free this ring entry if we've truly
1002 * freed the resources allocated to the entry. The refilling path can
1005 rds_ib_ring_free(&ic->i_recv_ring, 1);
1009 void rds_ib_recv_tasklet_fn(unsigned long data)
1011 struct rds_ib_connection *ic = (struct rds_ib_connection *) data;
1012 struct rds_connection *conn = ic->conn;
1013 struct rds_ib_ack_state state = { 0, };
1015 rds_poll_cq(ic, &state);
1016 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
1017 rds_poll_cq(ic, &state);
1019 if (state.ack_next_valid)
1020 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
1021 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
1022 rds_send_drop_acked(conn, state.ack_recv, NULL);
1023 ic->i_ack_recv = state.ack_recv;
1025 if (rds_conn_up(conn))
1026 rds_ib_attempt_ack(ic);
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
1031 if (rds_ib_ring_empty(&ic->i_recv_ring))
1032 rds_ib_stats_inc(s_ib_rx_ring_empty);
1034 if (rds_ib_ring_low(&ic->i_recv_ring))
1035 rds_ib_recv_refill(conn, 0);
1038 int rds_ib_recv(struct rds_connection *conn)
1040 struct rds_ib_connection *ic = conn->c_transport_data;
1043 rdsdebug("conn %p\n", conn);
1044 if (rds_conn_up(conn))
1045 rds_ib_attempt_ack(ic);
1050 int rds_ib_recv_init(void)
1055 /* Default to 30% of all available RAM for recv memory */
1057 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1059 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1060 sizeof(struct rds_ib_incoming),
1061 0, SLAB_HWCACHE_ALIGN, NULL);
1062 if (!rds_ib_incoming_slab)
1065 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1066 sizeof(struct rds_page_frag),
1067 0, SLAB_HWCACHE_ALIGN, NULL);
1068 if (!rds_ib_frag_slab)
1069 kmem_cache_destroy(rds_ib_incoming_slab);
1076 void rds_ib_recv_exit(void)
1078 kmem_cache_destroy(rds_ib_incoming_slab);
1079 kmem_cache_destroy(rds_ib_frag_slab);