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>
39 #include "rds_single_path.h"
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);
47 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
49 struct rds_ib_recv_work *recv;
52 for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
58 recv->r_wr.next = NULL;
60 recv->r_wr.sg_list = recv->r_sge;
61 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
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;
68 sge = &recv->r_sge[1];
70 sge->length = RDS_FRAG_SIZE;
71 sge->lkey = ic->i_pd->local_dma_lkey;
76 * The entire 'from' list, including the from element itself, is put on
77 * to the tail of the 'to' list.
79 static void list_splice_entire_tail(struct list_head *from,
82 struct list_head *from_last = from->prev;
84 list_splice_tail(from_last, to);
85 list_add_tail(from_last, to);
88 static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
90 struct list_head *tmp;
92 tmp = xchg(&cache->xfer, NULL);
95 list_splice_entire_tail(tmp, cache->ready);
101 static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
103 struct rds_ib_cache_head *head;
106 cache->percpu = alloc_percpu(struct rds_ib_cache_head);
110 for_each_possible_cpu(cpu) {
111 head = per_cpu_ptr(cache->percpu, cpu);
121 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
125 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
127 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
129 free_percpu(ic->i_cache_incs.percpu);
135 static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
136 struct list_head *caller_list)
138 struct rds_ib_cache_head *head;
141 for_each_possible_cpu(cpu) {
142 head = per_cpu_ptr(cache->percpu, cpu);
144 list_splice_entire_tail(head->first, caller_list);
150 list_splice_entire_tail(cache->ready, caller_list);
155 void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
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;
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);
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);
173 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
174 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
175 free_percpu(ic->i_cache_frags.percpu);
177 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
178 list_del(&frag->f_cache_entry);
179 WARN_ON(!list_empty(&frag->f_item));
180 kmem_cache_free(rds_ib_frag_slab, frag);
185 static void rds_ib_recv_cache_put(struct list_head *new_item,
186 struct rds_ib_refill_cache *cache);
187 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
190 /* Recycle frag and attached recv buffer f_sg */
191 static void rds_ib_frag_free(struct rds_ib_connection *ic,
192 struct rds_page_frag *frag)
194 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
196 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
199 /* Recycle inc after freeing attached frags */
200 void rds_ib_inc_free(struct rds_incoming *inc)
202 struct rds_ib_incoming *ibinc;
203 struct rds_page_frag *frag;
204 struct rds_page_frag *pos;
205 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
207 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
209 /* Free attached frags */
210 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
211 list_del_init(&frag->f_item);
212 rds_ib_frag_free(ic, frag);
214 BUG_ON(!list_empty(&ibinc->ii_frags));
216 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
217 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
220 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
221 struct rds_ib_recv_work *recv)
224 rds_inc_put(&recv->r_ibinc->ii_inc);
225 recv->r_ibinc = NULL;
228 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
229 rds_ib_frag_free(ic, recv->r_frag);
234 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
238 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
239 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
242 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
245 struct rds_ib_incoming *ibinc;
246 struct list_head *cache_item;
249 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
251 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
253 avail_allocs = atomic_add_unless(&rds_ib_allocation,
254 1, rds_ib_sysctl_max_recv_allocation);
256 rds_ib_stats_inc(s_ib_rx_alloc_limit);
259 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
261 atomic_dec(&rds_ib_allocation);
265 INIT_LIST_HEAD(&ibinc->ii_frags);
266 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
271 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
272 gfp_t slab_mask, gfp_t page_mask)
274 struct rds_page_frag *frag;
275 struct list_head *cache_item;
278 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
280 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
282 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
286 sg_init_table(&frag->f_sg, 1);
287 ret = rds_page_remainder_alloc(&frag->f_sg,
288 RDS_FRAG_SIZE, page_mask);
290 kmem_cache_free(rds_ib_frag_slab, frag);
295 INIT_LIST_HEAD(&frag->f_item);
300 static int rds_ib_recv_refill_one(struct rds_connection *conn,
301 struct rds_ib_recv_work *recv, gfp_t gfp)
303 struct rds_ib_connection *ic = conn->c_transport_data;
306 gfp_t slab_mask = GFP_NOWAIT;
307 gfp_t page_mask = GFP_NOWAIT;
309 if (gfp & __GFP_DIRECT_RECLAIM) {
310 slab_mask = GFP_KERNEL;
311 page_mask = GFP_HIGHUSER;
314 if (!ic->i_cache_incs.ready)
315 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
316 if (!ic->i_cache_frags.ready)
317 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
320 * ibinc was taken from recv if recv contained the start of a message.
321 * recvs that were continuations will still have this allocated.
323 if (!recv->r_ibinc) {
324 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
329 WARN_ON(recv->r_frag); /* leak! */
330 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
334 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
338 sge = &recv->r_sge[0];
339 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
340 sge->length = sizeof(struct rds_header);
342 sge = &recv->r_sge[1];
343 sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg);
344 sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg);
351 static int acquire_refill(struct rds_connection *conn)
353 return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
356 static void release_refill(struct rds_connection *conn)
358 clear_bit(RDS_RECV_REFILL, &conn->c_flags);
360 /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
361 * hot path and finding waiters is very rare. We don't want to walk
362 * the system-wide hashed waitqueue buckets in the fast path only to
363 * almost never find waiters.
365 if (waitqueue_active(&conn->c_waitq))
366 wake_up_all(&conn->c_waitq);
370 * This tries to allocate and post unused work requests after making sure that
371 * they have all the allocations they need to queue received fragments into
374 * -1 is returned if posting fails due to temporary resource exhaustion.
376 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
378 struct rds_ib_connection *ic = conn->c_transport_data;
379 struct rds_ib_recv_work *recv;
380 struct ib_recv_wr *failed_wr;
381 unsigned int posted = 0;
383 bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
386 /* the goal here is to just make sure that someone, somewhere
387 * is posting buffers. If we can't get the refill lock,
388 * let them do their thing
390 if (!acquire_refill(conn))
393 while ((prefill || rds_conn_up(conn)) &&
394 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
395 if (pos >= ic->i_recv_ring.w_nr) {
396 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
401 recv = &ic->i_recvs[pos];
402 ret = rds_ib_recv_refill_one(conn, recv, gfp);
407 /* XXX when can this fail? */
408 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
409 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
410 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
411 (long) ib_sg_dma_address(
413 &recv->r_frag->f_sg),
416 rds_ib_conn_error(conn, "recv post on "
417 "%pI4 returned %d, disconnecting and "
418 "reconnecting\n", &conn->c_faddr,
426 /* We're doing flow control - update the window. */
427 if (ic->i_flowctl && posted)
428 rds_ib_advertise_credits(conn, posted);
431 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
433 release_refill(conn);
435 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
436 * in this case the ring being low is going to lead to more interrupts
437 * and we can safely let the softirq code take care of it unless the
438 * ring is completely empty.
440 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
441 * we might have raced with the softirq code while we had the refill
442 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
443 * if we should requeue.
445 if (rds_conn_up(conn) &&
446 ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
447 rds_ib_ring_empty(&ic->i_recv_ring))) {
448 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
453 * We want to recycle several types of recv allocations, like incs and frags.
454 * To use this, the *_free() function passes in the ptr to a list_head within
455 * the recyclee, as well as the cache to put it on.
457 * First, we put the memory on a percpu list. When this reaches a certain size,
458 * We move it to an intermediate non-percpu list in a lockless manner, with some
459 * xchg/compxchg wizardry.
461 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
462 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
463 * list_empty() will return true with one element is actually present.
465 static void rds_ib_recv_cache_put(struct list_head *new_item,
466 struct rds_ib_refill_cache *cache)
469 struct list_head *old, *chpfirst;
471 local_irq_save(flags);
473 chpfirst = __this_cpu_read(cache->percpu->first);
475 INIT_LIST_HEAD(new_item);
476 else /* put on front */
477 list_add_tail(new_item, chpfirst);
479 __this_cpu_write(cache->percpu->first, new_item);
480 __this_cpu_inc(cache->percpu->count);
482 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
486 * Return our per-cpu first list to the cache's xfer by atomically
487 * grabbing the current xfer list, appending it to our per-cpu list,
488 * and then atomically returning that entire list back to the
489 * cache's xfer list as long as it's still empty.
492 old = xchg(&cache->xfer, NULL);
494 list_splice_entire_tail(old, chpfirst);
495 old = cmpxchg(&cache->xfer, NULL, chpfirst);
499 __this_cpu_write(cache->percpu->first, NULL);
500 __this_cpu_write(cache->percpu->count, 0);
502 local_irq_restore(flags);
505 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
507 struct list_head *head = cache->ready;
510 if (!list_empty(head)) {
511 cache->ready = head->next;
520 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
522 struct rds_ib_incoming *ibinc;
523 struct rds_page_frag *frag;
524 unsigned long to_copy;
525 unsigned long frag_off = 0;
530 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
531 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
532 len = be32_to_cpu(inc->i_hdr.h_len);
534 while (iov_iter_count(to) && copied < len) {
535 if (frag_off == RDS_FRAG_SIZE) {
536 frag = list_entry(frag->f_item.next,
537 struct rds_page_frag, f_item);
540 to_copy = min_t(unsigned long, iov_iter_count(to),
541 RDS_FRAG_SIZE - frag_off);
542 to_copy = min_t(unsigned long, to_copy, len - copied);
544 /* XXX needs + offset for multiple recvs per page */
545 rds_stats_add(s_copy_to_user, to_copy);
546 ret = copy_page_to_iter(sg_page(&frag->f_sg),
547 frag->f_sg.offset + frag_off,
560 /* ic starts out kzalloc()ed */
561 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
563 struct ib_send_wr *wr = &ic->i_ack_wr;
564 struct ib_sge *sge = &ic->i_ack_sge;
566 sge->addr = ic->i_ack_dma;
567 sge->length = sizeof(struct rds_header);
568 sge->lkey = ic->i_pd->local_dma_lkey;
572 wr->opcode = IB_WR_SEND;
573 wr->wr_id = RDS_IB_ACK_WR_ID;
574 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
578 * You'd think that with reliable IB connections you wouldn't need to ack
579 * messages that have been received. The problem is that IB hardware generates
580 * an ack message before it has DMAed the message into memory. This creates a
581 * potential message loss if the HCA is disabled for any reason between when it
582 * sends the ack and before the message is DMAed and processed. This is only a
583 * potential issue if another HCA is available for fail-over.
585 * When the remote host receives our ack they'll free the sent message from
586 * their send queue. To decrease the latency of this we always send an ack
587 * immediately after we've received messages.
589 * For simplicity, we only have one ack in flight at a time. This puts
590 * pressure on senders to have deep enough send queues to absorb the latency of
591 * a single ack frame being in flight. This might not be good enough.
593 * This is implemented by have a long-lived send_wr and sge which point to a
594 * statically allocated ack frame. This ack wr does not fall under the ring
595 * accounting that the tx and rx wrs do. The QP attribute specifically makes
596 * room for it beyond the ring size. Send completion notices its special
597 * wr_id and avoids working with the ring in that case.
599 #ifndef KERNEL_HAS_ATOMIC64
600 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
604 spin_lock_irqsave(&ic->i_ack_lock, flags);
605 ic->i_ack_next = seq;
607 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
608 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
611 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
616 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
618 spin_lock_irqsave(&ic->i_ack_lock, flags);
619 seq = ic->i_ack_next;
620 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
625 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
627 atomic64_set(&ic->i_ack_next, seq);
629 smp_mb__before_atomic();
630 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
634 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
636 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
637 smp_mb__after_atomic();
639 return atomic64_read(&ic->i_ack_next);
644 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
646 struct rds_header *hdr = ic->i_ack;
647 struct ib_send_wr *failed_wr;
651 seq = rds_ib_get_ack(ic);
653 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
654 rds_message_populate_header(hdr, 0, 0, 0);
655 hdr->h_ack = cpu_to_be64(seq);
656 hdr->h_credit = adv_credits;
657 rds_message_make_checksum(hdr);
658 ic->i_ack_queued = jiffies;
660 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
662 /* Failed to send. Release the WR, and
665 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
666 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
668 rds_ib_stats_inc(s_ib_ack_send_failure);
670 rds_ib_conn_error(ic->conn, "sending ack failed\n");
672 rds_ib_stats_inc(s_ib_ack_sent);
676 * There are 3 ways of getting acknowledgements to the peer:
677 * 1. We call rds_ib_attempt_ack from the recv completion handler
678 * to send an ACK-only frame.
679 * However, there can be only one such frame in the send queue
680 * at any time, so we may have to postpone it.
681 * 2. When another (data) packet is transmitted while there's
682 * an ACK in the queue, we piggyback the ACK sequence number
683 * on the data packet.
684 * 3. If the ACK WR is done sending, we get called from the
685 * send queue completion handler, and check whether there's
686 * another ACK pending (postponed because the WR was on the
687 * queue). If so, we transmit it.
689 * We maintain 2 variables:
690 * - i_ack_flags, which keeps track of whether the ACK WR
691 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
692 * - i_ack_next, which is the last sequence number we received
694 * Potentially, send queue and receive queue handlers can run concurrently.
695 * It would be nice to not have to use a spinlock to synchronize things,
696 * but the one problem that rules this out is that 64bit updates are
697 * not atomic on all platforms. Things would be a lot simpler if
698 * we had atomic64 or maybe cmpxchg64 everywhere.
700 * Reconnecting complicates this picture just slightly. When we
701 * reconnect, we may be seeing duplicate packets. The peer
702 * is retransmitting them, because it hasn't seen an ACK for
703 * them. It is important that we ACK these.
705 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
706 * this flag set *MUST* be acknowledged immediately.
710 * When we get here, we're called from the recv queue handler.
711 * Check whether we ought to transmit an ACK.
713 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
715 unsigned int adv_credits;
717 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
720 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
721 rds_ib_stats_inc(s_ib_ack_send_delayed);
725 /* Can we get a send credit? */
726 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
727 rds_ib_stats_inc(s_ib_tx_throttle);
728 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
732 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
733 rds_ib_send_ack(ic, adv_credits);
737 * We get here from the send completion handler, when the
738 * adapter tells us the ACK frame was sent.
740 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
742 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
743 rds_ib_attempt_ack(ic);
747 * This is called by the regular xmit code when it wants to piggyback
748 * an ACK on an outgoing frame.
750 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
752 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
753 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
754 return rds_ib_get_ack(ic);
758 * It's kind of lame that we're copying from the posted receive pages into
759 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
760 * them. But receiving new congestion bitmaps should be a *rare* event, so
761 * hopefully we won't need to invest that complexity in making it more
762 * efficient. By copying we can share a simpler core with TCP which has to
765 static void rds_ib_cong_recv(struct rds_connection *conn,
766 struct rds_ib_incoming *ibinc)
768 struct rds_cong_map *map;
769 unsigned int map_off;
770 unsigned int map_page;
771 struct rds_page_frag *frag;
772 unsigned long frag_off;
773 unsigned long to_copy;
774 unsigned long copied;
775 uint64_t uncongested = 0;
778 /* catch completely corrupt packets */
779 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
786 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
791 while (copied < RDS_CONG_MAP_BYTES) {
795 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
796 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
798 addr = kmap_atomic(sg_page(&frag->f_sg));
800 src = addr + frag->f_sg.offset + frag_off;
801 dst = (void *)map->m_page_addrs[map_page] + map_off;
802 for (k = 0; k < to_copy; k += 8) {
803 /* Record ports that became uncongested, ie
804 * bits that changed from 0 to 1. */
805 uncongested |= ~(*src) & *dst;
813 if (map_off == PAGE_SIZE) {
819 if (frag_off == RDS_FRAG_SIZE) {
820 frag = list_entry(frag->f_item.next,
821 struct rds_page_frag, f_item);
826 /* the congestion map is in little endian order */
827 uncongested = le64_to_cpu(uncongested);
829 rds_cong_map_updated(map, uncongested);
832 static void rds_ib_process_recv(struct rds_connection *conn,
833 struct rds_ib_recv_work *recv, u32 data_len,
834 struct rds_ib_ack_state *state)
836 struct rds_ib_connection *ic = conn->c_transport_data;
837 struct rds_ib_incoming *ibinc = ic->i_ibinc;
838 struct rds_header *ihdr, *hdr;
840 /* XXX shut down the connection if port 0,0 are seen? */
842 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
845 if (data_len < sizeof(struct rds_header)) {
846 rds_ib_conn_error(conn, "incoming message "
847 "from %pI4 didn't include a "
848 "header, disconnecting and "
853 data_len -= sizeof(struct rds_header);
855 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
857 /* Validate the checksum. */
858 if (!rds_message_verify_checksum(ihdr)) {
859 rds_ib_conn_error(conn, "incoming message "
860 "from %pI4 has corrupted header - "
861 "forcing a reconnect\n",
863 rds_stats_inc(s_recv_drop_bad_checksum);
867 /* Process the ACK sequence which comes with every packet */
868 state->ack_recv = be64_to_cpu(ihdr->h_ack);
869 state->ack_recv_valid = 1;
871 /* Process the credits update if there was one */
873 rds_ib_send_add_credits(conn, ihdr->h_credit);
875 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
876 /* This is an ACK-only packet. The fact that it gets
877 * special treatment here is that historically, ACKs
878 * were rather special beasts.
880 rds_ib_stats_inc(s_ib_ack_received);
883 * Usually the frags make their way on to incs and are then freed as
884 * the inc is freed. We don't go that route, so we have to drop the
885 * page ref ourselves. We can't just leave the page on the recv
886 * because that confuses the dma mapping of pages and each recv's use
889 * FIXME: Fold this into the code path below.
891 rds_ib_frag_free(ic, recv->r_frag);
897 * If we don't already have an inc on the connection then this
898 * fragment has a header and starts a message.. copy its header
899 * into the inc and save the inc so we can hang upcoming fragments
903 ibinc = recv->r_ibinc;
904 recv->r_ibinc = NULL;
907 hdr = &ibinc->ii_inc.i_hdr;
908 memcpy(hdr, ihdr, sizeof(*hdr));
909 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
911 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
912 ic->i_recv_data_rem, hdr->h_flags);
914 hdr = &ibinc->ii_inc.i_hdr;
915 /* We can't just use memcmp here; fragments of a
916 * single message may carry different ACKs */
917 if (hdr->h_sequence != ihdr->h_sequence ||
918 hdr->h_len != ihdr->h_len ||
919 hdr->h_sport != ihdr->h_sport ||
920 hdr->h_dport != ihdr->h_dport) {
921 rds_ib_conn_error(conn,
922 "fragment header mismatch; forcing reconnect\n");
927 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
930 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
931 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
933 ic->i_recv_data_rem = 0;
936 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
937 rds_ib_cong_recv(conn, ibinc);
939 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
940 &ibinc->ii_inc, GFP_ATOMIC);
941 state->ack_next = be64_to_cpu(hdr->h_sequence);
942 state->ack_next_valid = 1;
945 /* Evaluate the ACK_REQUIRED flag *after* we received
946 * the complete frame, and after bumping the next_rx
948 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
949 rds_stats_inc(s_recv_ack_required);
950 state->ack_required = 1;
953 rds_inc_put(&ibinc->ii_inc);
957 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
959 struct rds_ib_ack_state *state)
961 struct rds_connection *conn = ic->conn;
962 struct rds_ib_recv_work *recv;
964 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
965 (unsigned long long)wc->wr_id, wc->status,
966 ib_wc_status_msg(wc->status), wc->byte_len,
967 be32_to_cpu(wc->ex.imm_data));
969 rds_ib_stats_inc(s_ib_rx_cq_event);
970 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
971 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
974 /* Also process recvs in connecting state because it is possible
975 * to get a recv completion _before_ the rdmacm ESTABLISHED
976 * event is processed.
978 if (wc->status == IB_WC_SUCCESS) {
979 rds_ib_process_recv(conn, recv, wc->byte_len, state);
981 /* We expect errors as the qp is drained during shutdown */
982 if (rds_conn_up(conn) || rds_conn_connecting(conn))
983 rds_ib_conn_error(conn, "recv completion on %pI4 had status %u (%s), disconnecting and reconnecting\n",
986 ib_wc_status_msg(wc->status));
989 /* rds_ib_process_recv() doesn't always consume the frag, and
990 * we might not have called it at all if the wc didn't indicate
991 * success. We already unmapped the frag's pages, though, and
992 * the following rds_ib_ring_free() call tells the refill path
993 * that it will not find an allocated frag here. Make sure we
994 * keep that promise by freeing a frag that's still on the ring.
997 rds_ib_frag_free(ic, recv->r_frag);
1000 rds_ib_ring_free(&ic->i_recv_ring, 1);
1002 /* If we ever end up with a really empty receive ring, we're
1003 * in deep trouble, as the sender will definitely see RNR
1005 if (rds_ib_ring_empty(&ic->i_recv_ring))
1006 rds_ib_stats_inc(s_ib_rx_ring_empty);
1008 if (rds_ib_ring_low(&ic->i_recv_ring))
1009 rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1012 int rds_ib_recv_path(struct rds_conn_path *cp)
1014 struct rds_connection *conn = cp->cp_conn;
1015 struct rds_ib_connection *ic = conn->c_transport_data;
1018 rdsdebug("conn %p\n", conn);
1019 if (rds_conn_up(conn)) {
1020 rds_ib_attempt_ack(ic);
1021 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1027 int rds_ib_recv_init(void)
1032 /* Default to 30% of all available RAM for recv memory */
1034 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1036 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1037 sizeof(struct rds_ib_incoming),
1038 0, SLAB_HWCACHE_ALIGN, NULL);
1039 if (!rds_ib_incoming_slab)
1042 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1043 sizeof(struct rds_page_frag),
1044 0, SLAB_HWCACHE_ALIGN, NULL);
1045 if (!rds_ib_frag_slab) {
1046 kmem_cache_destroy(rds_ib_incoming_slab);
1047 rds_ib_incoming_slab = NULL;
1054 void rds_ib_recv_exit(void)
1056 kmem_cache_destroy(rds_ib_incoming_slab);
1057 kmem_cache_destroy(rds_ib_frag_slab);