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, gfp_t gfp)
103 struct rds_ib_cache_head *head;
106 cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp);
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, gfp_t gfp)
125 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp);
127 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp);
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);
197 atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
198 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
201 /* Recycle inc after freeing attached frags */
202 void rds_ib_inc_free(struct rds_incoming *inc)
204 struct rds_ib_incoming *ibinc;
205 struct rds_page_frag *frag;
206 struct rds_page_frag *pos;
207 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
209 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
211 /* Free attached frags */
212 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
213 list_del_init(&frag->f_item);
214 rds_ib_frag_free(ic, frag);
216 BUG_ON(!list_empty(&ibinc->ii_frags));
218 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
219 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
222 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
223 struct rds_ib_recv_work *recv)
226 rds_inc_put(&recv->r_ibinc->ii_inc);
227 recv->r_ibinc = NULL;
230 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
231 rds_ib_frag_free(ic, recv->r_frag);
236 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
240 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
241 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
244 static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
247 struct rds_ib_incoming *ibinc;
248 struct list_head *cache_item;
251 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
253 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
255 avail_allocs = atomic_add_unless(&rds_ib_allocation,
256 1, rds_ib_sysctl_max_recv_allocation);
258 rds_ib_stats_inc(s_ib_rx_alloc_limit);
261 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
263 atomic_dec(&rds_ib_allocation);
266 rds_ib_stats_inc(s_ib_rx_total_incs);
268 INIT_LIST_HEAD(&ibinc->ii_frags);
269 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
274 static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
275 gfp_t slab_mask, gfp_t page_mask)
277 struct rds_page_frag *frag;
278 struct list_head *cache_item;
281 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
283 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
284 atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs);
285 rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE);
287 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
291 sg_init_table(&frag->f_sg, 1);
292 ret = rds_page_remainder_alloc(&frag->f_sg,
293 RDS_FRAG_SIZE, page_mask);
295 kmem_cache_free(rds_ib_frag_slab, frag);
298 rds_ib_stats_inc(s_ib_rx_total_frags);
301 INIT_LIST_HEAD(&frag->f_item);
306 static int rds_ib_recv_refill_one(struct rds_connection *conn,
307 struct rds_ib_recv_work *recv, gfp_t gfp)
309 struct rds_ib_connection *ic = conn->c_transport_data;
312 gfp_t slab_mask = GFP_NOWAIT;
313 gfp_t page_mask = GFP_NOWAIT;
315 if (gfp & __GFP_DIRECT_RECLAIM) {
316 slab_mask = GFP_KERNEL;
317 page_mask = GFP_HIGHUSER;
320 if (!ic->i_cache_incs.ready)
321 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
322 if (!ic->i_cache_frags.ready)
323 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
326 * ibinc was taken from recv if recv contained the start of a message.
327 * recvs that were continuations will still have this allocated.
329 if (!recv->r_ibinc) {
330 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
335 WARN_ON(recv->r_frag); /* leak! */
336 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
340 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
344 sge = &recv->r_sge[0];
345 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
346 sge->length = sizeof(struct rds_header);
348 sge = &recv->r_sge[1];
349 sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg);
350 sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg);
357 static int acquire_refill(struct rds_connection *conn)
359 return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0;
362 static void release_refill(struct rds_connection *conn)
364 clear_bit(RDS_RECV_REFILL, &conn->c_flags);
365 smp_mb__after_atomic();
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.
372 if (waitqueue_active(&conn->c_waitq))
373 wake_up_all(&conn->c_waitq);
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
381 * -1 is returned if posting fails due to temporary resource exhaustion.
383 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
385 struct rds_ib_connection *ic = conn->c_transport_data;
386 struct rds_ib_recv_work *recv;
387 struct ib_recv_wr *failed_wr;
388 unsigned int posted = 0;
390 bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM);
393 /* the goal here is to just make sure that someone, somewhere
394 * is posting buffers. If we can't get the refill lock,
395 * let them do their thing
397 if (!acquire_refill(conn))
400 while ((prefill || rds_conn_up(conn)) &&
401 rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
402 if (pos >= ic->i_recv_ring.w_nr) {
403 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
408 recv = &ic->i_recvs[pos];
409 ret = rds_ib_recv_refill_one(conn, recv, gfp);
414 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv,
415 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
416 (long) ib_sg_dma_address(
418 &recv->r_frag->f_sg));
420 /* XXX when can this fail? */
421 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
423 rds_ib_conn_error(conn, "recv post on "
424 "%pI4 returned %d, disconnecting and "
425 "reconnecting\n", &conn->c_faddr,
433 /* We're doing flow control - update the window. */
434 if (ic->i_flowctl && posted)
435 rds_ib_advertise_credits(conn, posted);
438 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
440 release_refill(conn);
442 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
443 * in this case the ring being low is going to lead to more interrupts
444 * and we can safely let the softirq code take care of it unless the
445 * ring is completely empty.
447 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
448 * we might have raced with the softirq code while we had the refill
449 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
450 * if we should requeue.
452 if (rds_conn_up(conn) &&
453 ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) ||
454 rds_ib_ring_empty(&ic->i_recv_ring))) {
455 queue_delayed_work(rds_wq, &conn->c_recv_w, 1);
460 * We want to recycle several types of recv allocations, like incs and frags.
461 * To use this, the *_free() function passes in the ptr to a list_head within
462 * the recyclee, as well as the cache to put it on.
464 * First, we put the memory on a percpu list. When this reaches a certain size,
465 * We move it to an intermediate non-percpu list in a lockless manner, with some
466 * xchg/compxchg wizardry.
468 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
469 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
470 * list_empty() will return true with one element is actually present.
472 static void rds_ib_recv_cache_put(struct list_head *new_item,
473 struct rds_ib_refill_cache *cache)
476 struct list_head *old, *chpfirst;
478 local_irq_save(flags);
480 chpfirst = __this_cpu_read(cache->percpu->first);
482 INIT_LIST_HEAD(new_item);
483 else /* put on front */
484 list_add_tail(new_item, chpfirst);
486 __this_cpu_write(cache->percpu->first, new_item);
487 __this_cpu_inc(cache->percpu->count);
489 if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT)
493 * Return our per-cpu first list to the cache's xfer by atomically
494 * grabbing the current xfer list, appending it to our per-cpu list,
495 * and then atomically returning that entire list back to the
496 * cache's xfer list as long as it's still empty.
499 old = xchg(&cache->xfer, NULL);
501 list_splice_entire_tail(old, chpfirst);
502 old = cmpxchg(&cache->xfer, NULL, chpfirst);
506 __this_cpu_write(cache->percpu->first, NULL);
507 __this_cpu_write(cache->percpu->count, 0);
509 local_irq_restore(flags);
512 static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
514 struct list_head *head = cache->ready;
517 if (!list_empty(head)) {
518 cache->ready = head->next;
527 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
529 struct rds_ib_incoming *ibinc;
530 struct rds_page_frag *frag;
531 unsigned long to_copy;
532 unsigned long frag_off = 0;
537 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
538 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
539 len = be32_to_cpu(inc->i_hdr.h_len);
541 while (iov_iter_count(to) && copied < len) {
542 if (frag_off == RDS_FRAG_SIZE) {
543 frag = list_entry(frag->f_item.next,
544 struct rds_page_frag, f_item);
547 to_copy = min_t(unsigned long, iov_iter_count(to),
548 RDS_FRAG_SIZE - frag_off);
549 to_copy = min_t(unsigned long, to_copy, len - copied);
551 /* XXX needs + offset for multiple recvs per page */
552 rds_stats_add(s_copy_to_user, to_copy);
553 ret = copy_page_to_iter(sg_page(&frag->f_sg),
554 frag->f_sg.offset + frag_off,
567 /* ic starts out kzalloc()ed */
568 void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
570 struct ib_send_wr *wr = &ic->i_ack_wr;
571 struct ib_sge *sge = &ic->i_ack_sge;
573 sge->addr = ic->i_ack_dma;
574 sge->length = sizeof(struct rds_header);
575 sge->lkey = ic->i_pd->local_dma_lkey;
579 wr->opcode = IB_WR_SEND;
580 wr->wr_id = RDS_IB_ACK_WR_ID;
581 wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
585 * You'd think that with reliable IB connections you wouldn't need to ack
586 * messages that have been received. The problem is that IB hardware generates
587 * an ack message before it has DMAed the message into memory. This creates a
588 * potential message loss if the HCA is disabled for any reason between when it
589 * sends the ack and before the message is DMAed and processed. This is only a
590 * potential issue if another HCA is available for fail-over.
592 * When the remote host receives our ack they'll free the sent message from
593 * their send queue. To decrease the latency of this we always send an ack
594 * immediately after we've received messages.
596 * For simplicity, we only have one ack in flight at a time. This puts
597 * pressure on senders to have deep enough send queues to absorb the latency of
598 * a single ack frame being in flight. This might not be good enough.
600 * This is implemented by have a long-lived send_wr and sge which point to a
601 * statically allocated ack frame. This ack wr does not fall under the ring
602 * accounting that the tx and rx wrs do. The QP attribute specifically makes
603 * room for it beyond the ring size. Send completion notices its special
604 * wr_id and avoids working with the ring in that case.
606 #ifndef KERNEL_HAS_ATOMIC64
607 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
611 spin_lock_irqsave(&ic->i_ack_lock, flags);
612 ic->i_ack_next = seq;
614 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
615 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
618 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
623 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
625 spin_lock_irqsave(&ic->i_ack_lock, flags);
626 seq = ic->i_ack_next;
627 spin_unlock_irqrestore(&ic->i_ack_lock, flags);
632 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required)
634 atomic64_set(&ic->i_ack_next, seq);
636 smp_mb__before_atomic();
637 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
641 static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
643 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
644 smp_mb__after_atomic();
646 return atomic64_read(&ic->i_ack_next);
651 static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
653 struct rds_header *hdr = ic->i_ack;
654 struct ib_send_wr *failed_wr;
658 seq = rds_ib_get_ack(ic);
660 rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
661 rds_message_populate_header(hdr, 0, 0, 0);
662 hdr->h_ack = cpu_to_be64(seq);
663 hdr->h_credit = adv_credits;
664 rds_message_make_checksum(hdr);
665 ic->i_ack_queued = jiffies;
667 ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
669 /* Failed to send. Release the WR, and
672 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
673 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
675 rds_ib_stats_inc(s_ib_ack_send_failure);
677 rds_ib_conn_error(ic->conn, "sending ack failed\n");
679 rds_ib_stats_inc(s_ib_ack_sent);
683 * There are 3 ways of getting acknowledgements to the peer:
684 * 1. We call rds_ib_attempt_ack from the recv completion handler
685 * to send an ACK-only frame.
686 * However, there can be only one such frame in the send queue
687 * at any time, so we may have to postpone it.
688 * 2. When another (data) packet is transmitted while there's
689 * an ACK in the queue, we piggyback the ACK sequence number
690 * on the data packet.
691 * 3. If the ACK WR is done sending, we get called from the
692 * send queue completion handler, and check whether there's
693 * another ACK pending (postponed because the WR was on the
694 * queue). If so, we transmit it.
696 * We maintain 2 variables:
697 * - i_ack_flags, which keeps track of whether the ACK WR
698 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
699 * - i_ack_next, which is the last sequence number we received
701 * Potentially, send queue and receive queue handlers can run concurrently.
702 * It would be nice to not have to use a spinlock to synchronize things,
703 * but the one problem that rules this out is that 64bit updates are
704 * not atomic on all platforms. Things would be a lot simpler if
705 * we had atomic64 or maybe cmpxchg64 everywhere.
707 * Reconnecting complicates this picture just slightly. When we
708 * reconnect, we may be seeing duplicate packets. The peer
709 * is retransmitting them, because it hasn't seen an ACK for
710 * them. It is important that we ACK these.
712 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
713 * this flag set *MUST* be acknowledged immediately.
717 * When we get here, we're called from the recv queue handler.
718 * Check whether we ought to transmit an ACK.
720 void rds_ib_attempt_ack(struct rds_ib_connection *ic)
722 unsigned int adv_credits;
724 if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
727 if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
728 rds_ib_stats_inc(s_ib_ack_send_delayed);
732 /* Can we get a send credit? */
733 if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
734 rds_ib_stats_inc(s_ib_tx_throttle);
735 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
739 clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
740 rds_ib_send_ack(ic, adv_credits);
744 * We get here from the send completion handler, when the
745 * adapter tells us the ACK frame was sent.
747 void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
749 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
750 rds_ib_attempt_ack(ic);
754 * This is called by the regular xmit code when it wants to piggyback
755 * an ACK on an outgoing frame.
757 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
759 if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
760 rds_ib_stats_inc(s_ib_ack_send_piggybacked);
761 return rds_ib_get_ack(ic);
765 * It's kind of lame that we're copying from the posted receive pages into
766 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
767 * them. But receiving new congestion bitmaps should be a *rare* event, so
768 * hopefully we won't need to invest that complexity in making it more
769 * efficient. By copying we can share a simpler core with TCP which has to
772 static void rds_ib_cong_recv(struct rds_connection *conn,
773 struct rds_ib_incoming *ibinc)
775 struct rds_cong_map *map;
776 unsigned int map_off;
777 unsigned int map_page;
778 struct rds_page_frag *frag;
779 unsigned long frag_off;
780 unsigned long to_copy;
781 unsigned long copied;
782 uint64_t uncongested = 0;
785 /* catch completely corrupt packets */
786 if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
793 frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
798 while (copied < RDS_CONG_MAP_BYTES) {
802 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
803 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
805 addr = kmap_atomic(sg_page(&frag->f_sg));
807 src = addr + frag->f_sg.offset + frag_off;
808 dst = (void *)map->m_page_addrs[map_page] + map_off;
809 for (k = 0; k < to_copy; k += 8) {
810 /* Record ports that became uncongested, ie
811 * bits that changed from 0 to 1. */
812 uncongested |= ~(*src) & *dst;
820 if (map_off == PAGE_SIZE) {
826 if (frag_off == RDS_FRAG_SIZE) {
827 frag = list_entry(frag->f_item.next,
828 struct rds_page_frag, f_item);
833 /* the congestion map is in little endian order */
834 uncongested = le64_to_cpu(uncongested);
836 rds_cong_map_updated(map, uncongested);
839 static void rds_ib_process_recv(struct rds_connection *conn,
840 struct rds_ib_recv_work *recv, u32 data_len,
841 struct rds_ib_ack_state *state)
843 struct rds_ib_connection *ic = conn->c_transport_data;
844 struct rds_ib_incoming *ibinc = ic->i_ibinc;
845 struct rds_header *ihdr, *hdr;
847 /* XXX shut down the connection if port 0,0 are seen? */
849 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
852 if (data_len < sizeof(struct rds_header)) {
853 rds_ib_conn_error(conn, "incoming message "
854 "from %pI4 didn't include a "
855 "header, disconnecting and "
860 data_len -= sizeof(struct rds_header);
862 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
864 /* Validate the checksum. */
865 if (!rds_message_verify_checksum(ihdr)) {
866 rds_ib_conn_error(conn, "incoming message "
867 "from %pI4 has corrupted header - "
868 "forcing a reconnect\n",
870 rds_stats_inc(s_recv_drop_bad_checksum);
874 /* Process the ACK sequence which comes with every packet */
875 state->ack_recv = be64_to_cpu(ihdr->h_ack);
876 state->ack_recv_valid = 1;
878 /* Process the credits update if there was one */
880 rds_ib_send_add_credits(conn, ihdr->h_credit);
882 if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
883 /* This is an ACK-only packet. The fact that it gets
884 * special treatment here is that historically, ACKs
885 * were rather special beasts.
887 rds_ib_stats_inc(s_ib_ack_received);
890 * Usually the frags make their way on to incs and are then freed as
891 * the inc is freed. We don't go that route, so we have to drop the
892 * page ref ourselves. We can't just leave the page on the recv
893 * because that confuses the dma mapping of pages and each recv's use
896 * FIXME: Fold this into the code path below.
898 rds_ib_frag_free(ic, recv->r_frag);
904 * If we don't already have an inc on the connection then this
905 * fragment has a header and starts a message.. copy its header
906 * into the inc and save the inc so we can hang upcoming fragments
910 ibinc = recv->r_ibinc;
911 recv->r_ibinc = NULL;
914 hdr = &ibinc->ii_inc.i_hdr;
915 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] =
917 memcpy(hdr, ihdr, sizeof(*hdr));
918 ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);
919 ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] =
922 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
923 ic->i_recv_data_rem, hdr->h_flags);
925 hdr = &ibinc->ii_inc.i_hdr;
926 /* We can't just use memcmp here; fragments of a
927 * single message may carry different ACKs */
928 if (hdr->h_sequence != ihdr->h_sequence ||
929 hdr->h_len != ihdr->h_len ||
930 hdr->h_sport != ihdr->h_sport ||
931 hdr->h_dport != ihdr->h_dport) {
932 rds_ib_conn_error(conn,
933 "fragment header mismatch; forcing reconnect\n");
938 list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
941 if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
942 ic->i_recv_data_rem -= RDS_FRAG_SIZE;
944 ic->i_recv_data_rem = 0;
947 if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
948 rds_ib_cong_recv(conn, ibinc);
950 rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
951 &ibinc->ii_inc, GFP_ATOMIC);
952 state->ack_next = be64_to_cpu(hdr->h_sequence);
953 state->ack_next_valid = 1;
956 /* Evaluate the ACK_REQUIRED flag *after* we received
957 * the complete frame, and after bumping the next_rx
959 if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
960 rds_stats_inc(s_recv_ack_required);
961 state->ack_required = 1;
964 rds_inc_put(&ibinc->ii_inc);
968 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
970 struct rds_ib_ack_state *state)
972 struct rds_connection *conn = ic->conn;
973 struct rds_ib_recv_work *recv;
975 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
976 (unsigned long long)wc->wr_id, wc->status,
977 ib_wc_status_msg(wc->status), wc->byte_len,
978 be32_to_cpu(wc->ex.imm_data));
980 rds_ib_stats_inc(s_ib_rx_cq_event);
981 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
982 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1,
985 /* Also process recvs in connecting state because it is possible
986 * to get a recv completion _before_ the rdmacm ESTABLISHED
987 * event is processed.
989 if (wc->status == IB_WC_SUCCESS) {
990 rds_ib_process_recv(conn, recv, wc->byte_len, state);
992 /* We expect errors as the qp is drained during shutdown */
993 if (rds_conn_up(conn) || rds_conn_connecting(conn))
994 rds_ib_conn_error(conn, "recv completion on <%pI4,%pI4> had status %u (%s), disconnecting and reconnecting\n",
995 &conn->c_laddr, &conn->c_faddr,
997 ib_wc_status_msg(wc->status));
1000 /* rds_ib_process_recv() doesn't always consume the frag, and
1001 * we might not have called it at all if the wc didn't indicate
1002 * success. We already unmapped the frag's pages, though, and
1003 * the following rds_ib_ring_free() call tells the refill path
1004 * that it will not find an allocated frag here. Make sure we
1005 * keep that promise by freeing a frag that's still on the ring.
1008 rds_ib_frag_free(ic, recv->r_frag);
1009 recv->r_frag = NULL;
1011 rds_ib_ring_free(&ic->i_recv_ring, 1);
1013 /* If we ever end up with a really empty receive ring, we're
1014 * in deep trouble, as the sender will definitely see RNR
1016 if (rds_ib_ring_empty(&ic->i_recv_ring))
1017 rds_ib_stats_inc(s_ib_rx_ring_empty);
1019 if (rds_ib_ring_low(&ic->i_recv_ring)) {
1020 rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
1021 rds_ib_stats_inc(s_ib_rx_refill_from_cq);
1025 int rds_ib_recv_path(struct rds_conn_path *cp)
1027 struct rds_connection *conn = cp->cp_conn;
1028 struct rds_ib_connection *ic = conn->c_transport_data;
1031 rdsdebug("conn %p\n", conn);
1032 if (rds_conn_up(conn)) {
1033 rds_ib_attempt_ack(ic);
1034 rds_ib_recv_refill(conn, 0, GFP_KERNEL);
1035 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
1041 int rds_ib_recv_init(void)
1046 /* Default to 30% of all available RAM for recv memory */
1048 rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;
1050 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
1051 sizeof(struct rds_ib_incoming),
1052 0, SLAB_HWCACHE_ALIGN, NULL);
1053 if (!rds_ib_incoming_slab)
1056 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
1057 sizeof(struct rds_page_frag),
1058 0, SLAB_HWCACHE_ALIGN, NULL);
1059 if (!rds_ib_frag_slab) {
1060 kmem_cache_destroy(rds_ib_incoming_slab);
1061 rds_ib_incoming_slab = NULL;
1068 void rds_ib_recv_exit(void)
1070 kmem_cache_destroy(rds_ib_incoming_slab);
1071 kmem_cache_destroy(rds_ib_frag_slab);