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_pd
->local_dma_lkey
;
67 sge
= &recv
->r_sge
[1];
69 sge
->length
= RDS_FRAG_SIZE
;
70 sge
->lkey
= ic
->i_pd
->local_dma_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
, gfp_t gfp
)
302 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
305 gfp_t slab_mask
= GFP_NOWAIT
;
306 gfp_t page_mask
= GFP_NOWAIT
;
308 if (gfp
& __GFP_DIRECT_RECLAIM
) {
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
);
350 static int acquire_refill(struct rds_connection
*conn
)
352 return test_and_set_bit(RDS_RECV_REFILL
, &conn
->c_flags
) == 0;
355 static void release_refill(struct rds_connection
*conn
)
357 clear_bit(RDS_RECV_REFILL
, &conn
->c_flags
);
359 /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
360 * hot path and finding waiters is very rare. We don't want to walk
361 * the system-wide hashed waitqueue buckets in the fast path only to
362 * almost never find waiters.
364 if (waitqueue_active(&conn
->c_waitq
))
365 wake_up_all(&conn
->c_waitq
);
369 * This tries to allocate and post unused work requests after making sure that
370 * they have all the allocations they need to queue received fragments into
373 * -1 is returned if posting fails due to temporary resource exhaustion.
375 void rds_ib_recv_refill(struct rds_connection
*conn
, int prefill
, gfp_t gfp
)
377 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
378 struct rds_ib_recv_work
*recv
;
379 struct ib_recv_wr
*failed_wr
;
380 unsigned int posted
= 0;
382 bool can_wait
= !!(gfp
& __GFP_DIRECT_RECLAIM
);
385 /* the goal here is to just make sure that someone, somewhere
386 * is posting buffers. If we can't get the refill lock,
387 * let them do their thing
389 if (!acquire_refill(conn
))
392 while ((prefill
|| rds_conn_up(conn
)) &&
393 rds_ib_ring_alloc(&ic
->i_recv_ring
, 1, &pos
)) {
394 if (pos
>= ic
->i_recv_ring
.w_nr
) {
395 printk(KERN_NOTICE
"Argh - ring alloc returned pos=%u\n",
400 recv
= &ic
->i_recvs
[pos
];
401 ret
= rds_ib_recv_refill_one(conn
, recv
, gfp
);
406 /* XXX when can this fail? */
407 ret
= ib_post_recv(ic
->i_cm_id
->qp
, &recv
->r_wr
, &failed_wr
);
408 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv
,
409 recv
->r_ibinc
, sg_page(&recv
->r_frag
->f_sg
),
410 (long) ib_sg_dma_address(
412 &recv
->r_frag
->f_sg
),
415 rds_ib_conn_error(conn
, "recv post on "
416 "%pI4 returned %d, disconnecting and "
417 "reconnecting\n", &conn
->c_faddr
,
425 /* We're doing flow control - update the window. */
426 if (ic
->i_flowctl
&& posted
)
427 rds_ib_advertise_credits(conn
, posted
);
430 rds_ib_ring_unalloc(&ic
->i_recv_ring
, 1);
432 release_refill(conn
);
434 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
435 * in this case the ring being low is going to lead to more interrupts
436 * and we can safely let the softirq code take care of it unless the
437 * ring is completely empty.
439 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
440 * we might have raced with the softirq code while we had the refill
441 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
442 * if we should requeue.
444 if (rds_conn_up(conn
) &&
445 ((can_wait
&& rds_ib_ring_low(&ic
->i_recv_ring
)) ||
446 rds_ib_ring_empty(&ic
->i_recv_ring
))) {
447 queue_delayed_work(rds_wq
, &conn
->c_recv_w
, 1);
452 * We want to recycle several types of recv allocations, like incs and frags.
453 * To use this, the *_free() function passes in the ptr to a list_head within
454 * the recyclee, as well as the cache to put it on.
456 * First, we put the memory on a percpu list. When this reaches a certain size,
457 * We move it to an intermediate non-percpu list in a lockless manner, with some
458 * xchg/compxchg wizardry.
460 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
461 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
462 * list_empty() will return true with one element is actually present.
464 static void rds_ib_recv_cache_put(struct list_head
*new_item
,
465 struct rds_ib_refill_cache
*cache
)
468 struct list_head
*old
, *chpfirst
;
470 local_irq_save(flags
);
472 chpfirst
= __this_cpu_read(cache
->percpu
->first
);
474 INIT_LIST_HEAD(new_item
);
475 else /* put on front */
476 list_add_tail(new_item
, chpfirst
);
478 __this_cpu_write(cache
->percpu
->first
, new_item
);
479 __this_cpu_inc(cache
->percpu
->count
);
481 if (__this_cpu_read(cache
->percpu
->count
) < RDS_IB_RECYCLE_BATCH_COUNT
)
485 * Return our per-cpu first list to the cache's xfer by atomically
486 * grabbing the current xfer list, appending it to our per-cpu list,
487 * and then atomically returning that entire list back to the
488 * cache's xfer list as long as it's still empty.
491 old
= xchg(&cache
->xfer
, NULL
);
493 list_splice_entire_tail(old
, chpfirst
);
494 old
= cmpxchg(&cache
->xfer
, NULL
, chpfirst
);
498 __this_cpu_write(cache
->percpu
->first
, NULL
);
499 __this_cpu_write(cache
->percpu
->count
, 0);
501 local_irq_restore(flags
);
504 static struct list_head
*rds_ib_recv_cache_get(struct rds_ib_refill_cache
*cache
)
506 struct list_head
*head
= cache
->ready
;
509 if (!list_empty(head
)) {
510 cache
->ready
= head
->next
;
519 int rds_ib_inc_copy_to_user(struct rds_incoming
*inc
, struct iov_iter
*to
)
521 struct rds_ib_incoming
*ibinc
;
522 struct rds_page_frag
*frag
;
523 unsigned long to_copy
;
524 unsigned long frag_off
= 0;
529 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
530 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
531 len
= be32_to_cpu(inc
->i_hdr
.h_len
);
533 while (iov_iter_count(to
) && copied
< len
) {
534 if (frag_off
== RDS_FRAG_SIZE
) {
535 frag
= list_entry(frag
->f_item
.next
,
536 struct rds_page_frag
, f_item
);
539 to_copy
= min_t(unsigned long, iov_iter_count(to
),
540 RDS_FRAG_SIZE
- frag_off
);
541 to_copy
= min_t(unsigned long, to_copy
, len
- copied
);
543 /* XXX needs + offset for multiple recvs per page */
544 rds_stats_add(s_copy_to_user
, to_copy
);
545 ret
= copy_page_to_iter(sg_page(&frag
->f_sg
),
546 frag
->f_sg
.offset
+ frag_off
,
559 /* ic starts out kzalloc()ed */
560 void rds_ib_recv_init_ack(struct rds_ib_connection
*ic
)
562 struct ib_send_wr
*wr
= &ic
->i_ack_wr
;
563 struct ib_sge
*sge
= &ic
->i_ack_sge
;
565 sge
->addr
= ic
->i_ack_dma
;
566 sge
->length
= sizeof(struct rds_header
);
567 sge
->lkey
= ic
->i_pd
->local_dma_lkey
;
571 wr
->opcode
= IB_WR_SEND
;
572 wr
->wr_id
= RDS_IB_ACK_WR_ID
;
573 wr
->send_flags
= IB_SEND_SIGNALED
| IB_SEND_SOLICITED
;
577 * You'd think that with reliable IB connections you wouldn't need to ack
578 * messages that have been received. The problem is that IB hardware generates
579 * an ack message before it has DMAed the message into memory. This creates a
580 * potential message loss if the HCA is disabled for any reason between when it
581 * sends the ack and before the message is DMAed and processed. This is only a
582 * potential issue if another HCA is available for fail-over.
584 * When the remote host receives our ack they'll free the sent message from
585 * their send queue. To decrease the latency of this we always send an ack
586 * immediately after we've received messages.
588 * For simplicity, we only have one ack in flight at a time. This puts
589 * pressure on senders to have deep enough send queues to absorb the latency of
590 * a single ack frame being in flight. This might not be good enough.
592 * This is implemented by have a long-lived send_wr and sge which point to a
593 * statically allocated ack frame. This ack wr does not fall under the ring
594 * accounting that the tx and rx wrs do. The QP attribute specifically makes
595 * room for it beyond the ring size. Send completion notices its special
596 * wr_id and avoids working with the ring in that case.
598 #ifndef KERNEL_HAS_ATOMIC64
599 void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
, int ack_required
)
603 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
604 ic
->i_ack_next
= seq
;
606 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
607 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
610 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
615 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
617 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
618 seq
= ic
->i_ack_next
;
619 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
624 void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
, int ack_required
)
626 atomic64_set(&ic
->i_ack_next
, seq
);
628 smp_mb__before_atomic();
629 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
633 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
635 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
636 smp_mb__after_atomic();
638 return atomic64_read(&ic
->i_ack_next
);
643 static void rds_ib_send_ack(struct rds_ib_connection
*ic
, unsigned int adv_credits
)
645 struct rds_header
*hdr
= ic
->i_ack
;
646 struct ib_send_wr
*failed_wr
;
650 seq
= rds_ib_get_ack(ic
);
652 rdsdebug("send_ack: ic %p ack %llu\n", ic
, (unsigned long long) seq
);
653 rds_message_populate_header(hdr
, 0, 0, 0);
654 hdr
->h_ack
= cpu_to_be64(seq
);
655 hdr
->h_credit
= adv_credits
;
656 rds_message_make_checksum(hdr
);
657 ic
->i_ack_queued
= jiffies
;
659 ret
= ib_post_send(ic
->i_cm_id
->qp
, &ic
->i_ack_wr
, &failed_wr
);
661 /* Failed to send. Release the WR, and
664 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
665 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
667 rds_ib_stats_inc(s_ib_ack_send_failure
);
669 rds_ib_conn_error(ic
->conn
, "sending ack failed\n");
671 rds_ib_stats_inc(s_ib_ack_sent
);
675 * There are 3 ways of getting acknowledgements to the peer:
676 * 1. We call rds_ib_attempt_ack from the recv completion handler
677 * to send an ACK-only frame.
678 * However, there can be only one such frame in the send queue
679 * at any time, so we may have to postpone it.
680 * 2. When another (data) packet is transmitted while there's
681 * an ACK in the queue, we piggyback the ACK sequence number
682 * on the data packet.
683 * 3. If the ACK WR is done sending, we get called from the
684 * send queue completion handler, and check whether there's
685 * another ACK pending (postponed because the WR was on the
686 * queue). If so, we transmit it.
688 * We maintain 2 variables:
689 * - i_ack_flags, which keeps track of whether the ACK WR
690 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
691 * - i_ack_next, which is the last sequence number we received
693 * Potentially, send queue and receive queue handlers can run concurrently.
694 * It would be nice to not have to use a spinlock to synchronize things,
695 * but the one problem that rules this out is that 64bit updates are
696 * not atomic on all platforms. Things would be a lot simpler if
697 * we had atomic64 or maybe cmpxchg64 everywhere.
699 * Reconnecting complicates this picture just slightly. When we
700 * reconnect, we may be seeing duplicate packets. The peer
701 * is retransmitting them, because it hasn't seen an ACK for
702 * them. It is important that we ACK these.
704 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
705 * this flag set *MUST* be acknowledged immediately.
709 * When we get here, we're called from the recv queue handler.
710 * Check whether we ought to transmit an ACK.
712 void rds_ib_attempt_ack(struct rds_ib_connection
*ic
)
714 unsigned int adv_credits
;
716 if (!test_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
719 if (test_and_set_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
)) {
720 rds_ib_stats_inc(s_ib_ack_send_delayed
);
724 /* Can we get a send credit? */
725 if (!rds_ib_send_grab_credits(ic
, 1, &adv_credits
, 0, RDS_MAX_ADV_CREDIT
)) {
726 rds_ib_stats_inc(s_ib_tx_throttle
);
727 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
731 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
732 rds_ib_send_ack(ic
, adv_credits
);
736 * We get here from the send completion handler, when the
737 * adapter tells us the ACK frame was sent.
739 void rds_ib_ack_send_complete(struct rds_ib_connection
*ic
)
741 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
742 rds_ib_attempt_ack(ic
);
746 * This is called by the regular xmit code when it wants to piggyback
747 * an ACK on an outgoing frame.
749 u64
rds_ib_piggyb_ack(struct rds_ib_connection
*ic
)
751 if (test_and_clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
752 rds_ib_stats_inc(s_ib_ack_send_piggybacked
);
753 return rds_ib_get_ack(ic
);
757 * It's kind of lame that we're copying from the posted receive pages into
758 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
759 * them. But receiving new congestion bitmaps should be a *rare* event, so
760 * hopefully we won't need to invest that complexity in making it more
761 * efficient. By copying we can share a simpler core with TCP which has to
764 static void rds_ib_cong_recv(struct rds_connection
*conn
,
765 struct rds_ib_incoming
*ibinc
)
767 struct rds_cong_map
*map
;
768 unsigned int map_off
;
769 unsigned int map_page
;
770 struct rds_page_frag
*frag
;
771 unsigned long frag_off
;
772 unsigned long to_copy
;
773 unsigned long copied
;
774 uint64_t uncongested
= 0;
777 /* catch completely corrupt packets */
778 if (be32_to_cpu(ibinc
->ii_inc
.i_hdr
.h_len
) != RDS_CONG_MAP_BYTES
)
785 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
790 while (copied
< RDS_CONG_MAP_BYTES
) {
794 to_copy
= min(RDS_FRAG_SIZE
- frag_off
, PAGE_SIZE
- map_off
);
795 BUG_ON(to_copy
& 7); /* Must be 64bit aligned. */
797 addr
= kmap_atomic(sg_page(&frag
->f_sg
));
799 src
= addr
+ frag
->f_sg
.offset
+ frag_off
;
800 dst
= (void *)map
->m_page_addrs
[map_page
] + map_off
;
801 for (k
= 0; k
< to_copy
; k
+= 8) {
802 /* Record ports that became uncongested, ie
803 * bits that changed from 0 to 1. */
804 uncongested
|= ~(*src
) & *dst
;
812 if (map_off
== PAGE_SIZE
) {
818 if (frag_off
== RDS_FRAG_SIZE
) {
819 frag
= list_entry(frag
->f_item
.next
,
820 struct rds_page_frag
, f_item
);
825 /* the congestion map is in little endian order */
826 uncongested
= le64_to_cpu(uncongested
);
828 rds_cong_map_updated(map
, uncongested
);
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
)
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
;
839 /* XXX shut down the connection if port 0,0 are seen? */
841 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic
, ibinc
, recv
,
844 if (data_len
< sizeof(struct rds_header
)) {
845 rds_ib_conn_error(conn
, "incoming message "
846 "from %pI4 didn't include a "
847 "header, disconnecting and "
852 data_len
-= sizeof(struct rds_header
);
854 ihdr
= &ic
->i_recv_hdrs
[recv
- ic
->i_recvs
];
856 /* Validate the checksum. */
857 if (!rds_message_verify_checksum(ihdr
)) {
858 rds_ib_conn_error(conn
, "incoming message "
859 "from %pI4 has corrupted header - "
860 "forcing a reconnect\n",
862 rds_stats_inc(s_recv_drop_bad_checksum
);
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;
870 /* Process the credits update if there was one */
872 rds_ib_send_add_credits(conn
, ihdr
->h_credit
);
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.
879 rds_ib_stats_inc(s_ib_ack_received
);
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
888 * FIXME: Fold this into the code path below.
890 rds_ib_frag_free(ic
, recv
->r_frag
);
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
902 ibinc
= recv
->r_ibinc
;
903 recv
->r_ibinc
= NULL
;
906 hdr
= &ibinc
->ii_inc
.i_hdr
;
907 memcpy(hdr
, ihdr
, sizeof(*hdr
));
908 ic
->i_recv_data_rem
= be32_to_cpu(hdr
->h_len
);
910 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic
, ibinc
,
911 ic
->i_recv_data_rem
, hdr
->h_flags
);
913 hdr
= &ibinc
->ii_inc
.i_hdr
;
914 /* We can't just use memcmp here; fragments of a
915 * single message may carry different ACKs */
916 if (hdr
->h_sequence
!= ihdr
->h_sequence
||
917 hdr
->h_len
!= ihdr
->h_len
||
918 hdr
->h_sport
!= ihdr
->h_sport
||
919 hdr
->h_dport
!= ihdr
->h_dport
) {
920 rds_ib_conn_error(conn
,
921 "fragment header mismatch; forcing reconnect\n");
926 list_add_tail(&recv
->r_frag
->f_item
, &ibinc
->ii_frags
);
929 if (ic
->i_recv_data_rem
> RDS_FRAG_SIZE
)
930 ic
->i_recv_data_rem
-= RDS_FRAG_SIZE
;
932 ic
->i_recv_data_rem
= 0;
935 if (ibinc
->ii_inc
.i_hdr
.h_flags
== RDS_FLAG_CONG_BITMAP
)
936 rds_ib_cong_recv(conn
, ibinc
);
938 rds_recv_incoming(conn
, conn
->c_faddr
, conn
->c_laddr
,
939 &ibinc
->ii_inc
, GFP_ATOMIC
);
940 state
->ack_next
= be64_to_cpu(hdr
->h_sequence
);
941 state
->ack_next_valid
= 1;
944 /* Evaluate the ACK_REQUIRED flag *after* we received
945 * the complete frame, and after bumping the next_rx
947 if (hdr
->h_flags
& RDS_FLAG_ACK_REQUIRED
) {
948 rds_stats_inc(s_recv_ack_required
);
949 state
->ack_required
= 1;
952 rds_inc_put(&ibinc
->ii_inc
);
956 void rds_ib_recv_cqe_handler(struct rds_ib_connection
*ic
,
958 struct rds_ib_ack_state
*state
)
960 struct rds_connection
*conn
= ic
->conn
;
961 struct rds_ib_recv_work
*recv
;
963 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
964 (unsigned long long)wc
->wr_id
, wc
->status
,
965 ib_wc_status_msg(wc
->status
), wc
->byte_len
,
966 be32_to_cpu(wc
->ex
.imm_data
));
968 rds_ib_stats_inc(s_ib_rx_cq_event
);
969 recv
= &ic
->i_recvs
[rds_ib_ring_oldest(&ic
->i_recv_ring
)];
970 ib_dma_unmap_sg(ic
->i_cm_id
->device
, &recv
->r_frag
->f_sg
, 1,
973 /* Also process recvs in connecting state because it is possible
974 * to get a recv completion _before_ the rdmacm ESTABLISHED
975 * event is processed.
977 if (wc
->status
== IB_WC_SUCCESS
) {
978 rds_ib_process_recv(conn
, recv
, wc
->byte_len
, state
);
980 /* We expect errors as the qp is drained during shutdown */
981 if (rds_conn_up(conn
) || rds_conn_connecting(conn
))
982 rds_ib_conn_error(conn
, "recv completion on %pI4 had status %u (%s), disconnecting and reconnecting\n",
985 ib_wc_status_msg(wc
->status
));
988 /* rds_ib_process_recv() doesn't always consume the frag, and
989 * we might not have called it at all if the wc didn't indicate
990 * success. We already unmapped the frag's pages, though, and
991 * the following rds_ib_ring_free() call tells the refill path
992 * that it will not find an allocated frag here. Make sure we
993 * keep that promise by freeing a frag that's still on the ring.
996 rds_ib_frag_free(ic
, recv
->r_frag
);
999 rds_ib_ring_free(&ic
->i_recv_ring
, 1);
1001 /* If we ever end up with a really empty receive ring, we're
1002 * in deep trouble, as the sender will definitely see RNR
1004 if (rds_ib_ring_empty(&ic
->i_recv_ring
))
1005 rds_ib_stats_inc(s_ib_rx_ring_empty
);
1007 if (rds_ib_ring_low(&ic
->i_recv_ring
))
1008 rds_ib_recv_refill(conn
, 0, GFP_NOWAIT
);
1011 int rds_ib_recv(struct rds_connection
*conn
)
1013 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
1016 rdsdebug("conn %p\n", conn
);
1017 if (rds_conn_up(conn
)) {
1018 rds_ib_attempt_ack(ic
);
1019 rds_ib_recv_refill(conn
, 0, GFP_KERNEL
);
1025 int rds_ib_recv_init(void)
1030 /* Default to 30% of all available RAM for recv memory */
1032 rds_ib_sysctl_max_recv_allocation
= si
.totalram
/ 3 * PAGE_SIZE
/ RDS_FRAG_SIZE
;
1034 rds_ib_incoming_slab
= kmem_cache_create("rds_ib_incoming",
1035 sizeof(struct rds_ib_incoming
),
1036 0, SLAB_HWCACHE_ALIGN
, NULL
);
1037 if (!rds_ib_incoming_slab
)
1040 rds_ib_frag_slab
= kmem_cache_create("rds_ib_frag",
1041 sizeof(struct rds_page_frag
),
1042 0, SLAB_HWCACHE_ALIGN
, NULL
);
1043 if (!rds_ib_frag_slab
) {
1044 kmem_cache_destroy(rds_ib_incoming_slab
);
1045 rds_ib_incoming_slab
= NULL
;
1052 void rds_ib_recv_exit(void)
1054 kmem_cache_destroy(rds_ib_incoming_slab
);
1055 kmem_cache_destroy(rds_ib_frag_slab
);