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 static void rds_ib_frag_drop_page(struct rds_page_frag
*frag
)
48 rdsdebug("frag %p page %p\n", frag
, frag
->f_page
);
49 __free_page(frag
->f_page
);
53 static void rds_ib_frag_free(struct rds_page_frag
*frag
)
55 rdsdebug("frag %p page %p\n", frag
, frag
->f_page
);
57 kmem_cache_free(rds_ib_frag_slab
, frag
);
61 * We map a page at a time. Its fragments are posted in order. This
62 * is called in fragment order as the fragments get send completion events.
63 * Only the last frag in the page performs the unmapping.
65 * It's OK for ring cleanup to call this in whatever order it likes because
66 * DMA is not in flight and so we can unmap while other ring entries still
67 * hold page references in their frags.
69 static void rds_ib_recv_unmap_page(struct rds_ib_connection
*ic
,
70 struct rds_ib_recv_work
*recv
)
72 struct rds_page_frag
*frag
= recv
->r_frag
;
74 rdsdebug("recv %p frag %p page %p\n", recv
, frag
, frag
->f_page
);
76 ib_dma_unmap_page(ic
->i_cm_id
->device
,
78 RDS_FRAG_SIZE
, DMA_FROM_DEVICE
);
82 void rds_ib_recv_init_ring(struct rds_ib_connection
*ic
)
84 struct rds_ib_recv_work
*recv
;
87 for (i
= 0, recv
= ic
->i_recvs
; i
< ic
->i_recv_ring
.w_nr
; i
++, recv
++) {
93 recv
->r_wr
.next
= NULL
;
95 recv
->r_wr
.sg_list
= recv
->r_sge
;
96 recv
->r_wr
.num_sge
= RDS_IB_RECV_SGE
;
98 sge
= &recv
->r_sge
[0];
99 sge
->addr
= ic
->i_recv_hdrs_dma
+ (i
* sizeof(struct rds_header
));
100 sge
->length
= sizeof(struct rds_header
);
101 sge
->lkey
= ic
->i_mr
->lkey
;
103 sge
= &recv
->r_sge
[1];
105 sge
->length
= RDS_FRAG_SIZE
;
106 sge
->lkey
= ic
->i_mr
->lkey
;
110 static void rds_ib_recv_clear_one(struct rds_ib_connection
*ic
,
111 struct rds_ib_recv_work
*recv
)
114 rds_inc_put(&recv
->r_ibinc
->ii_inc
);
115 recv
->r_ibinc
= NULL
;
118 rds_ib_recv_unmap_page(ic
, recv
);
119 if (recv
->r_frag
->f_page
)
120 rds_ib_frag_drop_page(recv
->r_frag
);
121 rds_ib_frag_free(recv
->r_frag
);
126 void rds_ib_recv_clear_ring(struct rds_ib_connection
*ic
)
130 for (i
= 0; i
< ic
->i_recv_ring
.w_nr
; i
++)
131 rds_ib_recv_clear_one(ic
, &ic
->i_recvs
[i
]);
133 if (ic
->i_frag
.f_page
)
134 rds_ib_frag_drop_page(&ic
->i_frag
);
137 static int rds_ib_recv_refill_one(struct rds_connection
*conn
,
138 struct rds_ib_recv_work
*recv
)
140 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
145 if (!recv
->r_ibinc
) {
146 if (!atomic_add_unless(&rds_ib_allocation
, 1, rds_ib_sysctl_max_recv_allocation
)) {
147 rds_ib_stats_inc(s_ib_rx_alloc_limit
);
150 recv
->r_ibinc
= kmem_cache_alloc(rds_ib_incoming_slab
, GFP_NOWAIT
);
151 if (!recv
->r_ibinc
) {
152 atomic_dec(&rds_ib_allocation
);
155 INIT_LIST_HEAD(&recv
->r_ibinc
->ii_frags
);
156 rds_inc_init(&recv
->r_ibinc
->ii_inc
, conn
, conn
->c_faddr
);
160 recv
->r_frag
= kmem_cache_alloc(rds_ib_frag_slab
, GFP_NOWAIT
);
163 INIT_LIST_HEAD(&recv
->r_frag
->f_item
);
164 recv
->r_frag
->f_page
= NULL
;
167 if (!ic
->i_frag
.f_page
) {
168 ic
->i_frag
.f_page
= alloc_page(GFP_NOWAIT
);
169 if (!ic
->i_frag
.f_page
)
171 ic
->i_frag
.f_offset
= 0;
174 dma_addr
= ib_dma_map_page(ic
->i_cm_id
->device
,
179 if (ib_dma_mapping_error(ic
->i_cm_id
->device
, dma_addr
))
183 * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap()
184 * must be called on this recv. This happens as completions hit
185 * in order or on connection shutdown.
187 recv
->r_frag
->f_page
= ic
->i_frag
.f_page
;
188 recv
->r_frag
->f_offset
= ic
->i_frag
.f_offset
;
189 recv
->r_frag
->f_mapped
= dma_addr
;
191 sge
= &recv
->r_sge
[0];
192 sge
->addr
= ic
->i_recv_hdrs_dma
+ (recv
- ic
->i_recvs
) * sizeof(struct rds_header
);
193 sge
->length
= sizeof(struct rds_header
);
195 sge
= &recv
->r_sge
[1];
196 sge
->addr
= dma_addr
;
197 sge
->length
= RDS_FRAG_SIZE
;
199 get_page(recv
->r_frag
->f_page
);
201 if (ic
->i_frag
.f_offset
< RDS_PAGE_LAST_OFF
) {
202 ic
->i_frag
.f_offset
+= RDS_FRAG_SIZE
;
204 put_page(ic
->i_frag
.f_page
);
205 ic
->i_frag
.f_page
= NULL
;
206 ic
->i_frag
.f_offset
= 0;
215 * This tries to allocate and post unused work requests after making sure that
216 * they have all the allocations they need to queue received fragments into
217 * sockets. The i_recv_mutex is held here so that ring_alloc and _unalloc
218 * pairs don't go unmatched.
220 * -1 is returned if posting fails due to temporary resource exhaustion.
222 int rds_ib_recv_refill(struct rds_connection
*conn
, int prefill
)
224 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
225 struct rds_ib_recv_work
*recv
;
226 struct ib_recv_wr
*failed_wr
;
227 unsigned int posted
= 0;
231 while ((prefill
|| rds_conn_up(conn
)) &&
232 rds_ib_ring_alloc(&ic
->i_recv_ring
, 1, &pos
)) {
233 if (pos
>= ic
->i_recv_ring
.w_nr
) {
234 printk(KERN_NOTICE
"Argh - ring alloc returned pos=%u\n",
240 recv
= &ic
->i_recvs
[pos
];
241 ret
= rds_ib_recv_refill_one(conn
, recv
);
247 /* XXX when can this fail? */
248 ret
= ib_post_recv(ic
->i_cm_id
->qp
, &recv
->r_wr
, &failed_wr
);
249 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv
,
250 recv
->r_ibinc
, recv
->r_frag
->f_page
,
251 (long) recv
->r_frag
->f_mapped
, ret
);
253 rds_ib_conn_error(conn
, "recv post on "
254 "%pI4 returned %d, disconnecting and "
255 "reconnecting\n", &conn
->c_faddr
,
264 /* We're doing flow control - update the window. */
265 if (ic
->i_flowctl
&& posted
)
266 rds_ib_advertise_credits(conn
, posted
);
269 rds_ib_ring_unalloc(&ic
->i_recv_ring
, 1);
273 static void rds_ib_inc_purge(struct rds_incoming
*inc
)
275 struct rds_ib_incoming
*ibinc
;
276 struct rds_page_frag
*frag
;
277 struct rds_page_frag
*pos
;
279 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
280 rdsdebug("purging ibinc %p inc %p\n", ibinc
, inc
);
282 list_for_each_entry_safe(frag
, pos
, &ibinc
->ii_frags
, f_item
) {
283 list_del_init(&frag
->f_item
);
284 rds_ib_frag_drop_page(frag
);
285 rds_ib_frag_free(frag
);
289 void rds_ib_inc_free(struct rds_incoming
*inc
)
291 struct rds_ib_incoming
*ibinc
;
293 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
295 rds_ib_inc_purge(inc
);
296 rdsdebug("freeing ibinc %p inc %p\n", ibinc
, inc
);
297 BUG_ON(!list_empty(&ibinc
->ii_frags
));
298 kmem_cache_free(rds_ib_incoming_slab
, ibinc
);
299 atomic_dec(&rds_ib_allocation
);
300 BUG_ON(atomic_read(&rds_ib_allocation
) < 0);
303 int rds_ib_inc_copy_to_user(struct rds_incoming
*inc
, struct iovec
*first_iov
,
306 struct rds_ib_incoming
*ibinc
;
307 struct rds_page_frag
*frag
;
308 struct iovec
*iov
= first_iov
;
309 unsigned long to_copy
;
310 unsigned long frag_off
= 0;
311 unsigned long iov_off
= 0;
316 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
317 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
318 len
= be32_to_cpu(inc
->i_hdr
.h_len
);
320 while (copied
< size
&& copied
< len
) {
321 if (frag_off
== RDS_FRAG_SIZE
) {
322 frag
= list_entry(frag
->f_item
.next
,
323 struct rds_page_frag
, f_item
);
326 while (iov_off
== iov
->iov_len
) {
331 to_copy
= min(iov
->iov_len
- iov_off
, RDS_FRAG_SIZE
- frag_off
);
332 to_copy
= min_t(size_t, to_copy
, size
- copied
);
333 to_copy
= min_t(unsigned long, to_copy
, len
- copied
);
335 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
337 to_copy
, iov
->iov_base
, iov
->iov_len
, iov_off
,
338 frag
->f_page
, frag
->f_offset
, frag_off
);
340 /* XXX needs + offset for multiple recvs per page */
341 ret
= rds_page_copy_to_user(frag
->f_page
,
342 frag
->f_offset
+ frag_off
,
343 iov
->iov_base
+ iov_off
,
358 /* ic starts out kzalloc()ed */
359 void rds_ib_recv_init_ack(struct rds_ib_connection
*ic
)
361 struct ib_send_wr
*wr
= &ic
->i_ack_wr
;
362 struct ib_sge
*sge
= &ic
->i_ack_sge
;
364 sge
->addr
= ic
->i_ack_dma
;
365 sge
->length
= sizeof(struct rds_header
);
366 sge
->lkey
= ic
->i_mr
->lkey
;
370 wr
->opcode
= IB_WR_SEND
;
371 wr
->wr_id
= RDS_IB_ACK_WR_ID
;
372 wr
->send_flags
= IB_SEND_SIGNALED
| IB_SEND_SOLICITED
;
376 * You'd think that with reliable IB connections you wouldn't need to ack
377 * messages that have been received. The problem is that IB hardware generates
378 * an ack message before it has DMAed the message into memory. This creates a
379 * potential message loss if the HCA is disabled for any reason between when it
380 * sends the ack and before the message is DMAed and processed. This is only a
381 * potential issue if another HCA is available for fail-over.
383 * When the remote host receives our ack they'll free the sent message from
384 * their send queue. To decrease the latency of this we always send an ack
385 * immediately after we've received messages.
387 * For simplicity, we only have one ack in flight at a time. This puts
388 * pressure on senders to have deep enough send queues to absorb the latency of
389 * a single ack frame being in flight. This might not be good enough.
391 * This is implemented by have a long-lived send_wr and sge which point to a
392 * statically allocated ack frame. This ack wr does not fall under the ring
393 * accounting that the tx and rx wrs do. The QP attribute specifically makes
394 * room for it beyond the ring size. Send completion notices its special
395 * wr_id and avoids working with the ring in that case.
397 #ifndef KERNEL_HAS_ATOMIC64
398 static void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
,
403 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
404 ic
->i_ack_next
= seq
;
406 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
407 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
410 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
415 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
417 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
418 seq
= ic
->i_ack_next
;
419 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
424 static void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
,
427 atomic64_set(&ic
->i_ack_next
, seq
);
429 smp_mb__before_clear_bit();
430 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
434 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
436 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
437 smp_mb__after_clear_bit();
439 return atomic64_read(&ic
->i_ack_next
);
444 static void rds_ib_send_ack(struct rds_ib_connection
*ic
, unsigned int adv_credits
)
446 struct rds_header
*hdr
= ic
->i_ack
;
447 struct ib_send_wr
*failed_wr
;
451 seq
= rds_ib_get_ack(ic
);
453 rdsdebug("send_ack: ic %p ack %llu\n", ic
, (unsigned long long) seq
);
454 rds_message_populate_header(hdr
, 0, 0, 0);
455 hdr
->h_ack
= cpu_to_be64(seq
);
456 hdr
->h_credit
= adv_credits
;
457 rds_message_make_checksum(hdr
);
458 ic
->i_ack_queued
= jiffies
;
460 ret
= ib_post_send(ic
->i_cm_id
->qp
, &ic
->i_ack_wr
, &failed_wr
);
462 /* Failed to send. Release the WR, and
465 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
466 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
468 rds_ib_stats_inc(s_ib_ack_send_failure
);
470 rds_ib_conn_error(ic
->conn
, "sending ack failed\n");
472 rds_ib_stats_inc(s_ib_ack_sent
);
476 * There are 3 ways of getting acknowledgements to the peer:
477 * 1. We call rds_ib_attempt_ack from the recv completion handler
478 * to send an ACK-only frame.
479 * However, there can be only one such frame in the send queue
480 * at any time, so we may have to postpone it.
481 * 2. When another (data) packet is transmitted while there's
482 * an ACK in the queue, we piggyback the ACK sequence number
483 * on the data packet.
484 * 3. If the ACK WR is done sending, we get called from the
485 * send queue completion handler, and check whether there's
486 * another ACK pending (postponed because the WR was on the
487 * queue). If so, we transmit it.
489 * We maintain 2 variables:
490 * - i_ack_flags, which keeps track of whether the ACK WR
491 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
492 * - i_ack_next, which is the last sequence number we received
494 * Potentially, send queue and receive queue handlers can run concurrently.
495 * It would be nice to not have to use a spinlock to synchronize things,
496 * but the one problem that rules this out is that 64bit updates are
497 * not atomic on all platforms. Things would be a lot simpler if
498 * we had atomic64 or maybe cmpxchg64 everywhere.
500 * Reconnecting complicates this picture just slightly. When we
501 * reconnect, we may be seeing duplicate packets. The peer
502 * is retransmitting them, because it hasn't seen an ACK for
503 * them. It is important that we ACK these.
505 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
506 * this flag set *MUST* be acknowledged immediately.
510 * When we get here, we're called from the recv queue handler.
511 * Check whether we ought to transmit an ACK.
513 void rds_ib_attempt_ack(struct rds_ib_connection
*ic
)
515 unsigned int adv_credits
;
517 if (!test_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
520 if (test_and_set_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
)) {
521 rds_ib_stats_inc(s_ib_ack_send_delayed
);
525 /* Can we get a send credit? */
526 if (!rds_ib_send_grab_credits(ic
, 1, &adv_credits
, 0, RDS_MAX_ADV_CREDIT
)) {
527 rds_ib_stats_inc(s_ib_tx_throttle
);
528 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
532 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
533 rds_ib_send_ack(ic
, adv_credits
);
537 * We get here from the send completion handler, when the
538 * adapter tells us the ACK frame was sent.
540 void rds_ib_ack_send_complete(struct rds_ib_connection
*ic
)
542 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
543 rds_ib_attempt_ack(ic
);
547 * This is called by the regular xmit code when it wants to piggyback
548 * an ACK on an outgoing frame.
550 u64
rds_ib_piggyb_ack(struct rds_ib_connection
*ic
)
552 if (test_and_clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
553 rds_ib_stats_inc(s_ib_ack_send_piggybacked
);
554 return rds_ib_get_ack(ic
);
558 * It's kind of lame that we're copying from the posted receive pages into
559 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
560 * them. But receiving new congestion bitmaps should be a *rare* event, so
561 * hopefully we won't need to invest that complexity in making it more
562 * efficient. By copying we can share a simpler core with TCP which has to
565 static void rds_ib_cong_recv(struct rds_connection
*conn
,
566 struct rds_ib_incoming
*ibinc
)
568 struct rds_cong_map
*map
;
569 unsigned int map_off
;
570 unsigned int map_page
;
571 struct rds_page_frag
*frag
;
572 unsigned long frag_off
;
573 unsigned long to_copy
;
574 unsigned long copied
;
575 uint64_t uncongested
= 0;
578 /* catch completely corrupt packets */
579 if (be32_to_cpu(ibinc
->ii_inc
.i_hdr
.h_len
) != RDS_CONG_MAP_BYTES
)
586 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
591 while (copied
< RDS_CONG_MAP_BYTES
) {
595 to_copy
= min(RDS_FRAG_SIZE
- frag_off
, PAGE_SIZE
- map_off
);
596 BUG_ON(to_copy
& 7); /* Must be 64bit aligned. */
598 addr
= kmap_atomic(frag
->f_page
, KM_SOFTIRQ0
);
600 src
= addr
+ frag_off
;
601 dst
= (void *)map
->m_page_addrs
[map_page
] + map_off
;
602 for (k
= 0; k
< to_copy
; k
+= 8) {
603 /* Record ports that became uncongested, ie
604 * bits that changed from 0 to 1. */
605 uncongested
|= ~(*src
) & *dst
;
608 kunmap_atomic(addr
, KM_SOFTIRQ0
);
613 if (map_off
== PAGE_SIZE
) {
619 if (frag_off
== RDS_FRAG_SIZE
) {
620 frag
= list_entry(frag
->f_item
.next
,
621 struct rds_page_frag
, f_item
);
626 /* the congestion map is in little endian order */
627 uncongested
= le64_to_cpu(uncongested
);
629 rds_cong_map_updated(map
, uncongested
);
633 * Rings are posted with all the allocations they'll need to queue the
634 * incoming message to the receiving socket so this can't fail.
635 * All fragments start with a header, so we can make sure we're not receiving
636 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
638 struct rds_ib_ack_state
{
641 unsigned int ack_required
:1;
642 unsigned int ack_next_valid
:1;
643 unsigned int ack_recv_valid
:1;
646 static void rds_ib_process_recv(struct rds_connection
*conn
,
647 struct rds_ib_recv_work
*recv
, u32 data_len
,
648 struct rds_ib_ack_state
*state
)
650 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
651 struct rds_ib_incoming
*ibinc
= ic
->i_ibinc
;
652 struct rds_header
*ihdr
, *hdr
;
654 /* XXX shut down the connection if port 0,0 are seen? */
656 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic
, ibinc
, recv
,
659 if (data_len
< sizeof(struct rds_header
)) {
660 rds_ib_conn_error(conn
, "incoming message "
661 "from %pI4 didn't inclue a "
662 "header, disconnecting and "
667 data_len
-= sizeof(struct rds_header
);
669 ihdr
= &ic
->i_recv_hdrs
[recv
- ic
->i_recvs
];
671 /* Validate the checksum. */
672 if (!rds_message_verify_checksum(ihdr
)) {
673 rds_ib_conn_error(conn
, "incoming message "
674 "from %pI4 has corrupted header - "
675 "forcing a reconnect\n",
677 rds_stats_inc(s_recv_drop_bad_checksum
);
681 /* Process the ACK sequence which comes with every packet */
682 state
->ack_recv
= be64_to_cpu(ihdr
->h_ack
);
683 state
->ack_recv_valid
= 1;
685 /* Process the credits update if there was one */
687 rds_ib_send_add_credits(conn
, ihdr
->h_credit
);
689 if (ihdr
->h_sport
== 0 && ihdr
->h_dport
== 0 && data_len
== 0) {
690 /* This is an ACK-only packet. The fact that it gets
691 * special treatment here is that historically, ACKs
692 * were rather special beasts.
694 rds_ib_stats_inc(s_ib_ack_received
);
697 * Usually the frags make their way on to incs and are then freed as
698 * the inc is freed. We don't go that route, so we have to drop the
699 * page ref ourselves. We can't just leave the page on the recv
700 * because that confuses the dma mapping of pages and each recv's use
701 * of a partial page. We can leave the frag, though, it will be
704 * FIXME: Fold this into the code path below.
706 rds_ib_frag_drop_page(recv
->r_frag
);
711 * If we don't already have an inc on the connection then this
712 * fragment has a header and starts a message.. copy its header
713 * into the inc and save the inc so we can hang upcoming fragments
717 ibinc
= recv
->r_ibinc
;
718 recv
->r_ibinc
= NULL
;
721 hdr
= &ibinc
->ii_inc
.i_hdr
;
722 memcpy(hdr
, ihdr
, sizeof(*hdr
));
723 ic
->i_recv_data_rem
= be32_to_cpu(hdr
->h_len
);
725 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic
, ibinc
,
726 ic
->i_recv_data_rem
, hdr
->h_flags
);
728 hdr
= &ibinc
->ii_inc
.i_hdr
;
729 /* We can't just use memcmp here; fragments of a
730 * single message may carry different ACKs */
731 if (hdr
->h_sequence
!= ihdr
->h_sequence
||
732 hdr
->h_len
!= ihdr
->h_len
||
733 hdr
->h_sport
!= ihdr
->h_sport
||
734 hdr
->h_dport
!= ihdr
->h_dport
) {
735 rds_ib_conn_error(conn
,
736 "fragment header mismatch; forcing reconnect\n");
741 list_add_tail(&recv
->r_frag
->f_item
, &ibinc
->ii_frags
);
744 if (ic
->i_recv_data_rem
> RDS_FRAG_SIZE
)
745 ic
->i_recv_data_rem
-= RDS_FRAG_SIZE
;
747 ic
->i_recv_data_rem
= 0;
750 if (ibinc
->ii_inc
.i_hdr
.h_flags
== RDS_FLAG_CONG_BITMAP
)
751 rds_ib_cong_recv(conn
, ibinc
);
753 rds_recv_incoming(conn
, conn
->c_faddr
, conn
->c_laddr
,
754 &ibinc
->ii_inc
, GFP_ATOMIC
,
756 state
->ack_next
= be64_to_cpu(hdr
->h_sequence
);
757 state
->ack_next_valid
= 1;
760 /* Evaluate the ACK_REQUIRED flag *after* we received
761 * the complete frame, and after bumping the next_rx
763 if (hdr
->h_flags
& RDS_FLAG_ACK_REQUIRED
) {
764 rds_stats_inc(s_recv_ack_required
);
765 state
->ack_required
= 1;
768 rds_inc_put(&ibinc
->ii_inc
);
773 * Plucking the oldest entry from the ring can be done concurrently with
774 * the thread refilling the ring. Each ring operation is protected by
775 * spinlocks and the transient state of refilling doesn't change the
776 * recording of which entry is oldest.
778 * This relies on IB only calling one cq comp_handler for each cq so that
779 * there will only be one caller of rds_recv_incoming() per RDS connection.
781 void rds_ib_recv_cq_comp_handler(struct ib_cq
*cq
, void *context
)
783 struct rds_connection
*conn
= context
;
784 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
786 rdsdebug("conn %p cq %p\n", conn
, cq
);
788 rds_ib_stats_inc(s_ib_rx_cq_call
);
790 tasklet_schedule(&ic
->i_recv_tasklet
);
793 static inline void rds_poll_cq(struct rds_ib_connection
*ic
,
794 struct rds_ib_ack_state
*state
)
796 struct rds_connection
*conn
= ic
->conn
;
798 struct rds_ib_recv_work
*recv
;
800 while (ib_poll_cq(ic
->i_recv_cq
, 1, &wc
) > 0) {
801 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
802 (unsigned long long)wc
.wr_id
, wc
.status
, wc
.byte_len
,
803 be32_to_cpu(wc
.ex
.imm_data
));
804 rds_ib_stats_inc(s_ib_rx_cq_event
);
806 recv
= &ic
->i_recvs
[rds_ib_ring_oldest(&ic
->i_recv_ring
)];
808 rds_ib_recv_unmap_page(ic
, recv
);
811 * Also process recvs in connecting state because it is possible
812 * to get a recv completion _before_ the rdmacm ESTABLISHED
813 * event is processed.
815 if (rds_conn_up(conn
) || rds_conn_connecting(conn
)) {
816 /* We expect errors as the qp is drained during shutdown */
817 if (wc
.status
== IB_WC_SUCCESS
) {
818 rds_ib_process_recv(conn
, recv
, wc
.byte_len
, state
);
820 rds_ib_conn_error(conn
, "recv completion on "
821 "%pI4 had status %u, disconnecting and "
822 "reconnecting\n", &conn
->c_faddr
,
827 rds_ib_ring_free(&ic
->i_recv_ring
, 1);
831 void rds_ib_recv_tasklet_fn(unsigned long data
)
833 struct rds_ib_connection
*ic
= (struct rds_ib_connection
*) data
;
834 struct rds_connection
*conn
= ic
->conn
;
835 struct rds_ib_ack_state state
= { 0, };
837 rds_poll_cq(ic
, &state
);
838 ib_req_notify_cq(ic
->i_recv_cq
, IB_CQ_SOLICITED
);
839 rds_poll_cq(ic
, &state
);
841 if (state
.ack_next_valid
)
842 rds_ib_set_ack(ic
, state
.ack_next
, state
.ack_required
);
843 if (state
.ack_recv_valid
&& state
.ack_recv
> ic
->i_ack_recv
) {
844 rds_send_drop_acked(conn
, state
.ack_recv
, NULL
);
845 ic
->i_ack_recv
= state
.ack_recv
;
847 if (rds_conn_up(conn
))
848 rds_ib_attempt_ack(ic
);
850 /* If we ever end up with a really empty receive ring, we're
851 * in deep trouble, as the sender will definitely see RNR
853 if (rds_ib_ring_empty(&ic
->i_recv_ring
))
854 rds_ib_stats_inc(s_ib_rx_ring_empty
);
856 if (rds_ib_ring_low(&ic
->i_recv_ring
))
857 rds_ib_recv_refill(conn
, 0);
860 int rds_ib_recv(struct rds_connection
*conn
)
862 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
865 rdsdebug("conn %p\n", conn
);
866 if (rds_conn_up(conn
))
867 rds_ib_attempt_ack(ic
);
872 int __init
rds_ib_recv_init(void)
877 /* Default to 30% of all available RAM for recv memory */
879 rds_ib_sysctl_max_recv_allocation
= si
.totalram
/ 3 * PAGE_SIZE
/ RDS_FRAG_SIZE
;
881 rds_ib_incoming_slab
= kmem_cache_create("rds_ib_incoming",
882 sizeof(struct rds_ib_incoming
),
884 if (!rds_ib_incoming_slab
)
887 rds_ib_frag_slab
= kmem_cache_create("rds_ib_frag",
888 sizeof(struct rds_page_frag
),
890 if (!rds_ib_frag_slab
)
891 kmem_cache_destroy(rds_ib_incoming_slab
);
898 void rds_ib_recv_exit(void)
900 kmem_cache_destroy(rds_ib_incoming_slab
);
901 kmem_cache_destroy(rds_ib_frag_slab
);