2 * Copyright (c) 2006, 2017 Oracle and/or its affiliates. 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
);
171 atomic_dec(&rds_ib_allocation
);
174 rds_ib_cache_xfer_to_ready(&ic
->i_cache_frags
);
175 rds_ib_cache_splice_all_lists(&ic
->i_cache_frags
, &list
);
176 free_percpu(ic
->i_cache_frags
.percpu
);
178 list_for_each_entry_safe(frag
, frag_tmp
, &list
, f_cache_entry
) {
179 list_del(&frag
->f_cache_entry
);
180 WARN_ON(!list_empty(&frag
->f_item
));
181 kmem_cache_free(rds_ib_frag_slab
, frag
);
186 static void rds_ib_recv_cache_put(struct list_head
*new_item
,
187 struct rds_ib_refill_cache
*cache
);
188 static struct list_head
*rds_ib_recv_cache_get(struct rds_ib_refill_cache
*cache
);
191 /* Recycle frag and attached recv buffer f_sg */
192 static void rds_ib_frag_free(struct rds_ib_connection
*ic
,
193 struct rds_page_frag
*frag
)
195 rdsdebug("frag %p page %p\n", frag
, sg_page(&frag
->f_sg
));
197 rds_ib_recv_cache_put(&frag
->f_cache_entry
, &ic
->i_cache_frags
);
198 atomic_add(RDS_FRAG_SIZE
/ SZ_1K
, &ic
->i_cache_allocs
);
199 rds_ib_stats_add(s_ib_recv_added_to_cache
, RDS_FRAG_SIZE
);
202 /* Recycle inc after freeing attached frags */
203 void rds_ib_inc_free(struct rds_incoming
*inc
)
205 struct rds_ib_incoming
*ibinc
;
206 struct rds_page_frag
*frag
;
207 struct rds_page_frag
*pos
;
208 struct rds_ib_connection
*ic
= inc
->i_conn
->c_transport_data
;
210 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
212 /* Free attached frags */
213 list_for_each_entry_safe(frag
, pos
, &ibinc
->ii_frags
, f_item
) {
214 list_del_init(&frag
->f_item
);
215 rds_ib_frag_free(ic
, frag
);
217 BUG_ON(!list_empty(&ibinc
->ii_frags
));
219 rdsdebug("freeing ibinc %p inc %p\n", ibinc
, inc
);
220 rds_ib_recv_cache_put(&ibinc
->ii_cache_entry
, &ic
->i_cache_incs
);
223 static void rds_ib_recv_clear_one(struct rds_ib_connection
*ic
,
224 struct rds_ib_recv_work
*recv
)
227 rds_inc_put(&recv
->r_ibinc
->ii_inc
);
228 recv
->r_ibinc
= NULL
;
231 ib_dma_unmap_sg(ic
->i_cm_id
->device
, &recv
->r_frag
->f_sg
, 1, DMA_FROM_DEVICE
);
232 rds_ib_frag_free(ic
, recv
->r_frag
);
237 void rds_ib_recv_clear_ring(struct rds_ib_connection
*ic
)
241 for (i
= 0; i
< ic
->i_recv_ring
.w_nr
; i
++)
242 rds_ib_recv_clear_one(ic
, &ic
->i_recvs
[i
]);
245 static struct rds_ib_incoming
*rds_ib_refill_one_inc(struct rds_ib_connection
*ic
,
248 struct rds_ib_incoming
*ibinc
;
249 struct list_head
*cache_item
;
252 cache_item
= rds_ib_recv_cache_get(&ic
->i_cache_incs
);
254 ibinc
= container_of(cache_item
, struct rds_ib_incoming
, ii_cache_entry
);
256 avail_allocs
= atomic_add_unless(&rds_ib_allocation
,
257 1, rds_ib_sysctl_max_recv_allocation
);
259 rds_ib_stats_inc(s_ib_rx_alloc_limit
);
262 ibinc
= kmem_cache_alloc(rds_ib_incoming_slab
, slab_mask
);
264 atomic_dec(&rds_ib_allocation
);
267 rds_ib_stats_inc(s_ib_rx_total_incs
);
269 INIT_LIST_HEAD(&ibinc
->ii_frags
);
270 rds_inc_init(&ibinc
->ii_inc
, ic
->conn
, &ic
->conn
->c_faddr
);
275 static struct rds_page_frag
*rds_ib_refill_one_frag(struct rds_ib_connection
*ic
,
276 gfp_t slab_mask
, gfp_t page_mask
)
278 struct rds_page_frag
*frag
;
279 struct list_head
*cache_item
;
282 cache_item
= rds_ib_recv_cache_get(&ic
->i_cache_frags
);
284 frag
= container_of(cache_item
, struct rds_page_frag
, f_cache_entry
);
285 atomic_sub(RDS_FRAG_SIZE
/ SZ_1K
, &ic
->i_cache_allocs
);
286 rds_ib_stats_add(s_ib_recv_added_to_cache
, RDS_FRAG_SIZE
);
288 frag
= kmem_cache_alloc(rds_ib_frag_slab
, slab_mask
);
292 sg_init_table(&frag
->f_sg
, 1);
293 ret
= rds_page_remainder_alloc(&frag
->f_sg
,
294 RDS_FRAG_SIZE
, page_mask
);
296 kmem_cache_free(rds_ib_frag_slab
, frag
);
299 rds_ib_stats_inc(s_ib_rx_total_frags
);
302 INIT_LIST_HEAD(&frag
->f_item
);
307 static int rds_ib_recv_refill_one(struct rds_connection
*conn
,
308 struct rds_ib_recv_work
*recv
, gfp_t gfp
)
310 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
313 gfp_t slab_mask
= GFP_NOWAIT
;
314 gfp_t page_mask
= GFP_NOWAIT
;
316 if (gfp
& __GFP_DIRECT_RECLAIM
) {
317 slab_mask
= GFP_KERNEL
;
318 page_mask
= GFP_HIGHUSER
;
321 if (!ic
->i_cache_incs
.ready
)
322 rds_ib_cache_xfer_to_ready(&ic
->i_cache_incs
);
323 if (!ic
->i_cache_frags
.ready
)
324 rds_ib_cache_xfer_to_ready(&ic
->i_cache_frags
);
327 * ibinc was taken from recv if recv contained the start of a message.
328 * recvs that were continuations will still have this allocated.
330 if (!recv
->r_ibinc
) {
331 recv
->r_ibinc
= rds_ib_refill_one_inc(ic
, slab_mask
);
336 WARN_ON(recv
->r_frag
); /* leak! */
337 recv
->r_frag
= rds_ib_refill_one_frag(ic
, slab_mask
, page_mask
);
341 ret
= ib_dma_map_sg(ic
->i_cm_id
->device
, &recv
->r_frag
->f_sg
,
345 sge
= &recv
->r_sge
[0];
346 sge
->addr
= ic
->i_recv_hdrs_dma
+ (recv
- ic
->i_recvs
) * sizeof(struct rds_header
);
347 sge
->length
= sizeof(struct rds_header
);
349 sge
= &recv
->r_sge
[1];
350 sge
->addr
= sg_dma_address(&recv
->r_frag
->f_sg
);
351 sge
->length
= sg_dma_len(&recv
->r_frag
->f_sg
);
358 static int acquire_refill(struct rds_connection
*conn
)
360 return test_and_set_bit(RDS_RECV_REFILL
, &conn
->c_flags
) == 0;
363 static void release_refill(struct rds_connection
*conn
)
365 clear_bit(RDS_RECV_REFILL
, &conn
->c_flags
);
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 void rds_ib_recv_refill(struct rds_connection
*conn
, int prefill
, gfp_t gfp
)
383 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
384 struct rds_ib_recv_work
*recv
;
385 unsigned int posted
= 0;
387 bool can_wait
= !!(gfp
& __GFP_DIRECT_RECLAIM
);
388 bool must_wake
= false;
391 /* the goal here is to just make sure that someone, somewhere
392 * is posting buffers. If we can't get the refill lock,
393 * let them do their thing
395 if (!acquire_refill(conn
))
398 while ((prefill
|| rds_conn_up(conn
)) &&
399 rds_ib_ring_alloc(&ic
->i_recv_ring
, 1, &pos
)) {
400 if (pos
>= ic
->i_recv_ring
.w_nr
) {
401 printk(KERN_NOTICE
"Argh - ring alloc returned pos=%u\n",
406 recv
= &ic
->i_recvs
[pos
];
407 ret
= rds_ib_recv_refill_one(conn
, recv
, gfp
);
413 rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv
,
414 recv
->r_ibinc
, sg_page(&recv
->r_frag
->f_sg
),
415 (long)sg_dma_address(&recv
->r_frag
->f_sg
));
417 /* XXX when can this fail? */
418 ret
= ib_post_recv(ic
->i_cm_id
->qp
, &recv
->r_wr
, NULL
);
420 rds_ib_conn_error(conn
, "recv post on "
421 "%pI6c returned %d, disconnecting and "
422 "reconnecting\n", &conn
->c_faddr
,
429 if ((posted
> 128 && need_resched()) || posted
> 8192) {
435 /* We're doing flow control - update the window. */
436 if (ic
->i_flowctl
&& posted
)
437 rds_ib_advertise_credits(conn
, posted
);
440 rds_ib_ring_unalloc(&ic
->i_recv_ring
, 1);
442 release_refill(conn
);
444 /* if we're called from the softirq handler, we'll be GFP_NOWAIT.
445 * in this case the ring being low is going to lead to more interrupts
446 * and we can safely let the softirq code take care of it unless the
447 * ring is completely empty.
449 * if we're called from krdsd, we'll be GFP_KERNEL. In this case
450 * we might have raced with the softirq code while we had the refill
451 * lock held. Use rds_ib_ring_low() instead of ring_empty to decide
452 * if we should requeue.
454 if (rds_conn_up(conn
) &&
456 (can_wait
&& rds_ib_ring_low(&ic
->i_recv_ring
)) ||
457 rds_ib_ring_empty(&ic
->i_recv_ring
))) {
458 queue_delayed_work(rds_wq
, &conn
->c_recv_w
, 1);
465 * We want to recycle several types of recv allocations, like incs and frags.
466 * To use this, the *_free() function passes in the ptr to a list_head within
467 * the recyclee, as well as the cache to put it on.
469 * First, we put the memory on a percpu list. When this reaches a certain size,
470 * We move it to an intermediate non-percpu list in a lockless manner, with some
471 * xchg/compxchg wizardry.
473 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
474 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
475 * list_empty() will return true with one element is actually present.
477 static void rds_ib_recv_cache_put(struct list_head
*new_item
,
478 struct rds_ib_refill_cache
*cache
)
481 struct list_head
*old
, *chpfirst
;
483 local_irq_save(flags
);
485 chpfirst
= __this_cpu_read(cache
->percpu
->first
);
487 INIT_LIST_HEAD(new_item
);
488 else /* put on front */
489 list_add_tail(new_item
, chpfirst
);
491 __this_cpu_write(cache
->percpu
->first
, new_item
);
492 __this_cpu_inc(cache
->percpu
->count
);
494 if (__this_cpu_read(cache
->percpu
->count
) < RDS_IB_RECYCLE_BATCH_COUNT
)
498 * Return our per-cpu first list to the cache's xfer by atomically
499 * grabbing the current xfer list, appending it to our per-cpu list,
500 * and then atomically returning that entire list back to the
501 * cache's xfer list as long as it's still empty.
504 old
= xchg(&cache
->xfer
, NULL
);
506 list_splice_entire_tail(old
, chpfirst
);
507 old
= cmpxchg(&cache
->xfer
, NULL
, chpfirst
);
511 __this_cpu_write(cache
->percpu
->first
, NULL
);
512 __this_cpu_write(cache
->percpu
->count
, 0);
514 local_irq_restore(flags
);
517 static struct list_head
*rds_ib_recv_cache_get(struct rds_ib_refill_cache
*cache
)
519 struct list_head
*head
= cache
->ready
;
522 if (!list_empty(head
)) {
523 cache
->ready
= head
->next
;
532 int rds_ib_inc_copy_to_user(struct rds_incoming
*inc
, struct iov_iter
*to
)
534 struct rds_ib_incoming
*ibinc
;
535 struct rds_page_frag
*frag
;
536 unsigned long to_copy
;
537 unsigned long frag_off
= 0;
542 ibinc
= container_of(inc
, struct rds_ib_incoming
, ii_inc
);
543 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
544 len
= be32_to_cpu(inc
->i_hdr
.h_len
);
546 while (iov_iter_count(to
) && copied
< len
) {
547 if (frag_off
== RDS_FRAG_SIZE
) {
548 frag
= list_entry(frag
->f_item
.next
,
549 struct rds_page_frag
, f_item
);
552 to_copy
= min_t(unsigned long, iov_iter_count(to
),
553 RDS_FRAG_SIZE
- frag_off
);
554 to_copy
= min_t(unsigned long, to_copy
, len
- copied
);
556 /* XXX needs + offset for multiple recvs per page */
557 rds_stats_add(s_copy_to_user
, to_copy
);
558 ret
= copy_page_to_iter(sg_page(&frag
->f_sg
),
559 frag
->f_sg
.offset
+ frag_off
,
572 /* ic starts out kzalloc()ed */
573 void rds_ib_recv_init_ack(struct rds_ib_connection
*ic
)
575 struct ib_send_wr
*wr
= &ic
->i_ack_wr
;
576 struct ib_sge
*sge
= &ic
->i_ack_sge
;
578 sge
->addr
= ic
->i_ack_dma
;
579 sge
->length
= sizeof(struct rds_header
);
580 sge
->lkey
= ic
->i_pd
->local_dma_lkey
;
584 wr
->opcode
= IB_WR_SEND
;
585 wr
->wr_id
= RDS_IB_ACK_WR_ID
;
586 wr
->send_flags
= IB_SEND_SIGNALED
| IB_SEND_SOLICITED
;
590 * You'd think that with reliable IB connections you wouldn't need to ack
591 * messages that have been received. The problem is that IB hardware generates
592 * an ack message before it has DMAed the message into memory. This creates a
593 * potential message loss if the HCA is disabled for any reason between when it
594 * sends the ack and before the message is DMAed and processed. This is only a
595 * potential issue if another HCA is available for fail-over.
597 * When the remote host receives our ack they'll free the sent message from
598 * their send queue. To decrease the latency of this we always send an ack
599 * immediately after we've received messages.
601 * For simplicity, we only have one ack in flight at a time. This puts
602 * pressure on senders to have deep enough send queues to absorb the latency of
603 * a single ack frame being in flight. This might not be good enough.
605 * This is implemented by have a long-lived send_wr and sge which point to a
606 * statically allocated ack frame. This ack wr does not fall under the ring
607 * accounting that the tx and rx wrs do. The QP attribute specifically makes
608 * room for it beyond the ring size. Send completion notices its special
609 * wr_id and avoids working with the ring in that case.
611 #ifndef KERNEL_HAS_ATOMIC64
612 void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
, int ack_required
)
616 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
617 ic
->i_ack_next
= seq
;
619 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
620 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
623 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
628 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
630 spin_lock_irqsave(&ic
->i_ack_lock
, flags
);
631 seq
= ic
->i_ack_next
;
632 spin_unlock_irqrestore(&ic
->i_ack_lock
, flags
);
637 void rds_ib_set_ack(struct rds_ib_connection
*ic
, u64 seq
, int ack_required
)
639 atomic64_set(&ic
->i_ack_next
, seq
);
641 smp_mb__before_atomic();
642 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
646 static u64
rds_ib_get_ack(struct rds_ib_connection
*ic
)
648 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
649 smp_mb__after_atomic();
651 return atomic64_read(&ic
->i_ack_next
);
656 static void rds_ib_send_ack(struct rds_ib_connection
*ic
, unsigned int adv_credits
)
658 struct rds_header
*hdr
= ic
->i_ack
;
662 seq
= rds_ib_get_ack(ic
);
664 rdsdebug("send_ack: ic %p ack %llu\n", ic
, (unsigned long long) seq
);
665 rds_message_populate_header(hdr
, 0, 0, 0);
666 hdr
->h_ack
= cpu_to_be64(seq
);
667 hdr
->h_credit
= adv_credits
;
668 rds_message_make_checksum(hdr
);
669 ic
->i_ack_queued
= jiffies
;
671 ret
= ib_post_send(ic
->i_cm_id
->qp
, &ic
->i_ack_wr
, NULL
);
673 /* Failed to send. Release the WR, and
676 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
677 set_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
679 rds_ib_stats_inc(s_ib_ack_send_failure
);
681 rds_ib_conn_error(ic
->conn
, "sending ack failed\n");
683 rds_ib_stats_inc(s_ib_ack_sent
);
687 * There are 3 ways of getting acknowledgements to the peer:
688 * 1. We call rds_ib_attempt_ack from the recv completion handler
689 * to send an ACK-only frame.
690 * However, there can be only one such frame in the send queue
691 * at any time, so we may have to postpone it.
692 * 2. When another (data) packet is transmitted while there's
693 * an ACK in the queue, we piggyback the ACK sequence number
694 * on the data packet.
695 * 3. If the ACK WR is done sending, we get called from the
696 * send queue completion handler, and check whether there's
697 * another ACK pending (postponed because the WR was on the
698 * queue). If so, we transmit it.
700 * We maintain 2 variables:
701 * - i_ack_flags, which keeps track of whether the ACK WR
702 * is currently in the send queue or not (IB_ACK_IN_FLIGHT)
703 * - i_ack_next, which is the last sequence number we received
705 * Potentially, send queue and receive queue handlers can run concurrently.
706 * It would be nice to not have to use a spinlock to synchronize things,
707 * but the one problem that rules this out is that 64bit updates are
708 * not atomic on all platforms. Things would be a lot simpler if
709 * we had atomic64 or maybe cmpxchg64 everywhere.
711 * Reconnecting complicates this picture just slightly. When we
712 * reconnect, we may be seeing duplicate packets. The peer
713 * is retransmitting them, because it hasn't seen an ACK for
714 * them. It is important that we ACK these.
716 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
717 * this flag set *MUST* be acknowledged immediately.
721 * When we get here, we're called from the recv queue handler.
722 * Check whether we ought to transmit an ACK.
724 void rds_ib_attempt_ack(struct rds_ib_connection
*ic
)
726 unsigned int adv_credits
;
728 if (!test_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
731 if (test_and_set_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
)) {
732 rds_ib_stats_inc(s_ib_ack_send_delayed
);
736 /* Can we get a send credit? */
737 if (!rds_ib_send_grab_credits(ic
, 1, &adv_credits
, 0, RDS_MAX_ADV_CREDIT
)) {
738 rds_ib_stats_inc(s_ib_tx_throttle
);
739 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
743 clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
);
744 rds_ib_send_ack(ic
, adv_credits
);
748 * We get here from the send completion handler, when the
749 * adapter tells us the ACK frame was sent.
751 void rds_ib_ack_send_complete(struct rds_ib_connection
*ic
)
753 clear_bit(IB_ACK_IN_FLIGHT
, &ic
->i_ack_flags
);
754 rds_ib_attempt_ack(ic
);
758 * This is called by the regular xmit code when it wants to piggyback
759 * an ACK on an outgoing frame.
761 u64
rds_ib_piggyb_ack(struct rds_ib_connection
*ic
)
763 if (test_and_clear_bit(IB_ACK_REQUESTED
, &ic
->i_ack_flags
))
764 rds_ib_stats_inc(s_ib_ack_send_piggybacked
);
765 return rds_ib_get_ack(ic
);
769 * It's kind of lame that we're copying from the posted receive pages into
770 * long-lived bitmaps. We could have posted the bitmaps and rdma written into
771 * them. But receiving new congestion bitmaps should be a *rare* event, so
772 * hopefully we won't need to invest that complexity in making it more
773 * efficient. By copying we can share a simpler core with TCP which has to
776 static void rds_ib_cong_recv(struct rds_connection
*conn
,
777 struct rds_ib_incoming
*ibinc
)
779 struct rds_cong_map
*map
;
780 unsigned int map_off
;
781 unsigned int map_page
;
782 struct rds_page_frag
*frag
;
783 unsigned long frag_off
;
784 unsigned long to_copy
;
785 unsigned long copied
;
786 __le64 uncongested
= 0;
789 /* catch completely corrupt packets */
790 if (be32_to_cpu(ibinc
->ii_inc
.i_hdr
.h_len
) != RDS_CONG_MAP_BYTES
)
797 frag
= list_entry(ibinc
->ii_frags
.next
, struct rds_page_frag
, f_item
);
802 while (copied
< RDS_CONG_MAP_BYTES
) {
806 to_copy
= min(RDS_FRAG_SIZE
- frag_off
, PAGE_SIZE
- map_off
);
807 BUG_ON(to_copy
& 7); /* Must be 64bit aligned. */
809 addr
= kmap_atomic(sg_page(&frag
->f_sg
));
811 src
= addr
+ frag
->f_sg
.offset
+ frag_off
;
812 dst
= (void *)map
->m_page_addrs
[map_page
] + map_off
;
813 for (k
= 0; k
< to_copy
; k
+= 8) {
814 /* Record ports that became uncongested, ie
815 * bits that changed from 0 to 1. */
816 uncongested
|= ~(*src
) & *dst
;
824 if (map_off
== PAGE_SIZE
) {
830 if (frag_off
== RDS_FRAG_SIZE
) {
831 frag
= list_entry(frag
->f_item
.next
,
832 struct rds_page_frag
, f_item
);
837 /* the congestion map is in little endian order */
838 rds_cong_map_updated(map
, le64_to_cpu(uncongested
));
841 static void rds_ib_process_recv(struct rds_connection
*conn
,
842 struct rds_ib_recv_work
*recv
, u32 data_len
,
843 struct rds_ib_ack_state
*state
)
845 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
846 struct rds_ib_incoming
*ibinc
= ic
->i_ibinc
;
847 struct rds_header
*ihdr
, *hdr
;
849 /* XXX shut down the connection if port 0,0 are seen? */
851 rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic
, ibinc
, recv
,
854 if (data_len
< sizeof(struct rds_header
)) {
855 rds_ib_conn_error(conn
, "incoming message "
856 "from %pI6c didn't include a "
857 "header, disconnecting and "
862 data_len
-= sizeof(struct rds_header
);
864 ihdr
= &ic
->i_recv_hdrs
[recv
- ic
->i_recvs
];
866 /* Validate the checksum. */
867 if (!rds_message_verify_checksum(ihdr
)) {
868 rds_ib_conn_error(conn
, "incoming message "
869 "from %pI6c has corrupted header - "
870 "forcing a reconnect\n",
872 rds_stats_inc(s_recv_drop_bad_checksum
);
876 /* Process the ACK sequence which comes with every packet */
877 state
->ack_recv
= be64_to_cpu(ihdr
->h_ack
);
878 state
->ack_recv_valid
= 1;
880 /* Process the credits update if there was one */
882 rds_ib_send_add_credits(conn
, ihdr
->h_credit
);
884 if (ihdr
->h_sport
== 0 && ihdr
->h_dport
== 0 && data_len
== 0) {
885 /* This is an ACK-only packet. The fact that it gets
886 * special treatment here is that historically, ACKs
887 * were rather special beasts.
889 rds_ib_stats_inc(s_ib_ack_received
);
892 * Usually the frags make their way on to incs and are then freed as
893 * the inc is freed. We don't go that route, so we have to drop the
894 * page ref ourselves. We can't just leave the page on the recv
895 * because that confuses the dma mapping of pages and each recv's use
898 * FIXME: Fold this into the code path below.
900 rds_ib_frag_free(ic
, recv
->r_frag
);
906 * If we don't already have an inc on the connection then this
907 * fragment has a header and starts a message.. copy its header
908 * into the inc and save the inc so we can hang upcoming fragments
912 ibinc
= recv
->r_ibinc
;
913 recv
->r_ibinc
= NULL
;
916 hdr
= &ibinc
->ii_inc
.i_hdr
;
917 ibinc
->ii_inc
.i_rx_lat_trace
[RDS_MSG_RX_HDR
] =
919 memcpy(hdr
, ihdr
, sizeof(*hdr
));
920 ic
->i_recv_data_rem
= be32_to_cpu(hdr
->h_len
);
921 ibinc
->ii_inc
.i_rx_lat_trace
[RDS_MSG_RX_START
] =
924 rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic
, ibinc
,
925 ic
->i_recv_data_rem
, hdr
->h_flags
);
927 hdr
= &ibinc
->ii_inc
.i_hdr
;
928 /* We can't just use memcmp here; fragments of a
929 * single message may carry different ACKs */
930 if (hdr
->h_sequence
!= ihdr
->h_sequence
||
931 hdr
->h_len
!= ihdr
->h_len
||
932 hdr
->h_sport
!= ihdr
->h_sport
||
933 hdr
->h_dport
!= ihdr
->h_dport
) {
934 rds_ib_conn_error(conn
,
935 "fragment header mismatch; forcing reconnect\n");
940 list_add_tail(&recv
->r_frag
->f_item
, &ibinc
->ii_frags
);
943 if (ic
->i_recv_data_rem
> RDS_FRAG_SIZE
)
944 ic
->i_recv_data_rem
-= RDS_FRAG_SIZE
;
946 ic
->i_recv_data_rem
= 0;
949 if (ibinc
->ii_inc
.i_hdr
.h_flags
== RDS_FLAG_CONG_BITMAP
) {
950 rds_ib_cong_recv(conn
, ibinc
);
952 rds_recv_incoming(conn
, &conn
->c_faddr
, &conn
->c_laddr
,
953 &ibinc
->ii_inc
, GFP_ATOMIC
);
954 state
->ack_next
= be64_to_cpu(hdr
->h_sequence
);
955 state
->ack_next_valid
= 1;
958 /* Evaluate the ACK_REQUIRED flag *after* we received
959 * the complete frame, and after bumping the next_rx
961 if (hdr
->h_flags
& RDS_FLAG_ACK_REQUIRED
) {
962 rds_stats_inc(s_recv_ack_required
);
963 state
->ack_required
= 1;
966 rds_inc_put(&ibinc
->ii_inc
);
970 void rds_ib_recv_cqe_handler(struct rds_ib_connection
*ic
,
972 struct rds_ib_ack_state
*state
)
974 struct rds_connection
*conn
= ic
->conn
;
975 struct rds_ib_recv_work
*recv
;
977 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
978 (unsigned long long)wc
->wr_id
, wc
->status
,
979 ib_wc_status_msg(wc
->status
), wc
->byte_len
,
980 be32_to_cpu(wc
->ex
.imm_data
));
982 rds_ib_stats_inc(s_ib_rx_cq_event
);
983 recv
= &ic
->i_recvs
[rds_ib_ring_oldest(&ic
->i_recv_ring
)];
984 ib_dma_unmap_sg(ic
->i_cm_id
->device
, &recv
->r_frag
->f_sg
, 1,
987 /* Also process recvs in connecting state because it is possible
988 * to get a recv completion _before_ the rdmacm ESTABLISHED
989 * event is processed.
991 if (wc
->status
== IB_WC_SUCCESS
) {
992 rds_ib_process_recv(conn
, recv
, wc
->byte_len
, state
);
994 /* We expect errors as the qp is drained during shutdown */
995 if (rds_conn_up(conn
) || rds_conn_connecting(conn
))
996 rds_ib_conn_error(conn
, "recv completion on <%pI6c,%pI6c, %d> had status %u (%s), disconnecting and reconnecting\n",
997 &conn
->c_laddr
, &conn
->c_faddr
,
998 conn
->c_tos
, wc
->status
,
999 ib_wc_status_msg(wc
->status
));
1002 /* rds_ib_process_recv() doesn't always consume the frag, and
1003 * we might not have called it at all if the wc didn't indicate
1004 * success. We already unmapped the frag's pages, though, and
1005 * the following rds_ib_ring_free() call tells the refill path
1006 * that it will not find an allocated frag here. Make sure we
1007 * keep that promise by freeing a frag that's still on the ring.
1010 rds_ib_frag_free(ic
, recv
->r_frag
);
1011 recv
->r_frag
= NULL
;
1013 rds_ib_ring_free(&ic
->i_recv_ring
, 1);
1015 /* If we ever end up with a really empty receive ring, we're
1016 * in deep trouble, as the sender will definitely see RNR
1018 if (rds_ib_ring_empty(&ic
->i_recv_ring
))
1019 rds_ib_stats_inc(s_ib_rx_ring_empty
);
1021 if (rds_ib_ring_low(&ic
->i_recv_ring
)) {
1022 rds_ib_recv_refill(conn
, 0, GFP_NOWAIT
);
1023 rds_ib_stats_inc(s_ib_rx_refill_from_cq
);
1027 int rds_ib_recv_path(struct rds_conn_path
*cp
)
1029 struct rds_connection
*conn
= cp
->cp_conn
;
1030 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
1032 rdsdebug("conn %p\n", conn
);
1033 if (rds_conn_up(conn
)) {
1034 rds_ib_attempt_ack(ic
);
1035 rds_ib_recv_refill(conn
, 0, GFP_KERNEL
);
1036 rds_ib_stats_inc(s_ib_rx_refill_from_thread
);
1042 int rds_ib_recv_init(void)
1047 /* Default to 30% of all available RAM for recv memory */
1049 rds_ib_sysctl_max_recv_allocation
= si
.totalram
/ 3 * PAGE_SIZE
/ RDS_FRAG_SIZE
;
1051 rds_ib_incoming_slab
=
1052 kmem_cache_create_usercopy("rds_ib_incoming",
1053 sizeof(struct rds_ib_incoming
),
1054 0, SLAB_HWCACHE_ALIGN
,
1055 offsetof(struct rds_ib_incoming
,
1057 sizeof(struct rds_inc_usercopy
),
1059 if (!rds_ib_incoming_slab
)
1062 rds_ib_frag_slab
= kmem_cache_create("rds_ib_frag",
1063 sizeof(struct rds_page_frag
),
1064 0, SLAB_HWCACHE_ALIGN
, NULL
);
1065 if (!rds_ib_frag_slab
) {
1066 kmem_cache_destroy(rds_ib_incoming_slab
);
1067 rds_ib_incoming_slab
= NULL
;
1074 void rds_ib_recv_exit(void)
1076 WARN_ON(atomic_read(&rds_ib_allocation
));
1078 kmem_cache_destroy(rds_ib_incoming_slab
);
1079 kmem_cache_destroy(rds_ib_frag_slab
);