2 * Copyright (c) 2006, 2018 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/rculist.h>
36 #include <linux/llist.h>
38 #include "rds_single_path.h"
42 struct workqueue_struct
*rds_ib_mr_wq
;
43 struct rds_ib_dereg_odp_mr
{
44 struct work_struct work
;
48 static void rds_ib_odp_mr_worker(struct work_struct
*work
);
50 static struct rds_ib_device
*rds_ib_get_device(__be32 ipaddr
)
52 struct rds_ib_device
*rds_ibdev
;
53 struct rds_ib_ipaddr
*i_ipaddr
;
56 list_for_each_entry_rcu(rds_ibdev
, &rds_ib_devices
, list
) {
57 list_for_each_entry_rcu(i_ipaddr
, &rds_ibdev
->ipaddr_list
, list
) {
58 if (i_ipaddr
->ipaddr
== ipaddr
) {
59 refcount_inc(&rds_ibdev
->refcount
);
70 static int rds_ib_add_ipaddr(struct rds_ib_device
*rds_ibdev
, __be32 ipaddr
)
72 struct rds_ib_ipaddr
*i_ipaddr
;
74 i_ipaddr
= kmalloc(sizeof *i_ipaddr
, GFP_KERNEL
);
78 i_ipaddr
->ipaddr
= ipaddr
;
80 spin_lock_irq(&rds_ibdev
->spinlock
);
81 list_add_tail_rcu(&i_ipaddr
->list
, &rds_ibdev
->ipaddr_list
);
82 spin_unlock_irq(&rds_ibdev
->spinlock
);
87 static void rds_ib_remove_ipaddr(struct rds_ib_device
*rds_ibdev
, __be32 ipaddr
)
89 struct rds_ib_ipaddr
*i_ipaddr
;
90 struct rds_ib_ipaddr
*to_free
= NULL
;
93 spin_lock_irq(&rds_ibdev
->spinlock
);
94 list_for_each_entry_rcu(i_ipaddr
, &rds_ibdev
->ipaddr_list
, list
) {
95 if (i_ipaddr
->ipaddr
== ipaddr
) {
96 list_del_rcu(&i_ipaddr
->list
);
101 spin_unlock_irq(&rds_ibdev
->spinlock
);
104 kfree_rcu(to_free
, rcu
);
107 int rds_ib_update_ipaddr(struct rds_ib_device
*rds_ibdev
,
108 struct in6_addr
*ipaddr
)
110 struct rds_ib_device
*rds_ibdev_old
;
112 rds_ibdev_old
= rds_ib_get_device(ipaddr
->s6_addr32
[3]);
114 return rds_ib_add_ipaddr(rds_ibdev
, ipaddr
->s6_addr32
[3]);
116 if (rds_ibdev_old
!= rds_ibdev
) {
117 rds_ib_remove_ipaddr(rds_ibdev_old
, ipaddr
->s6_addr32
[3]);
118 rds_ib_dev_put(rds_ibdev_old
);
119 return rds_ib_add_ipaddr(rds_ibdev
, ipaddr
->s6_addr32
[3]);
121 rds_ib_dev_put(rds_ibdev_old
);
126 void rds_ib_add_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
)
128 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
130 /* conn was previously on the nodev_conns_list */
131 spin_lock_irq(&ib_nodev_conns_lock
);
132 BUG_ON(list_empty(&ib_nodev_conns
));
133 BUG_ON(list_empty(&ic
->ib_node
));
134 list_del(&ic
->ib_node
);
136 spin_lock(&rds_ibdev
->spinlock
);
137 list_add_tail(&ic
->ib_node
, &rds_ibdev
->conn_list
);
138 spin_unlock(&rds_ibdev
->spinlock
);
139 spin_unlock_irq(&ib_nodev_conns_lock
);
141 ic
->rds_ibdev
= rds_ibdev
;
142 refcount_inc(&rds_ibdev
->refcount
);
145 void rds_ib_remove_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
)
147 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
149 /* place conn on nodev_conns_list */
150 spin_lock(&ib_nodev_conns_lock
);
152 spin_lock_irq(&rds_ibdev
->spinlock
);
153 BUG_ON(list_empty(&ic
->ib_node
));
154 list_del(&ic
->ib_node
);
155 spin_unlock_irq(&rds_ibdev
->spinlock
);
157 list_add_tail(&ic
->ib_node
, &ib_nodev_conns
);
159 spin_unlock(&ib_nodev_conns_lock
);
161 ic
->rds_ibdev
= NULL
;
162 rds_ib_dev_put(rds_ibdev
);
165 void rds_ib_destroy_nodev_conns(void)
167 struct rds_ib_connection
*ic
, *_ic
;
170 /* avoid calling conn_destroy with irqs off */
171 spin_lock_irq(&ib_nodev_conns_lock
);
172 list_splice(&ib_nodev_conns
, &tmp_list
);
173 spin_unlock_irq(&ib_nodev_conns_lock
);
175 list_for_each_entry_safe(ic
, _ic
, &tmp_list
, ib_node
)
176 rds_conn_destroy(ic
->conn
);
179 void rds_ib_get_mr_info(struct rds_ib_device
*rds_ibdev
, struct rds_info_rdma_connection
*iinfo
)
181 struct rds_ib_mr_pool
*pool_1m
= rds_ibdev
->mr_1m_pool
;
183 iinfo
->rdma_mr_max
= pool_1m
->max_items
;
184 iinfo
->rdma_mr_size
= pool_1m
->max_pages
;
187 #if IS_ENABLED(CONFIG_IPV6)
188 void rds6_ib_get_mr_info(struct rds_ib_device
*rds_ibdev
,
189 struct rds6_info_rdma_connection
*iinfo6
)
191 struct rds_ib_mr_pool
*pool_1m
= rds_ibdev
->mr_1m_pool
;
193 iinfo6
->rdma_mr_max
= pool_1m
->max_items
;
194 iinfo6
->rdma_mr_size
= pool_1m
->max_pages
;
198 struct rds_ib_mr
*rds_ib_reuse_mr(struct rds_ib_mr_pool
*pool
)
200 struct rds_ib_mr
*ibmr
= NULL
;
201 struct llist_node
*ret
;
204 spin_lock_irqsave(&pool
->clean_lock
, flags
);
205 ret
= llist_del_first(&pool
->clean_list
);
206 spin_unlock_irqrestore(&pool
->clean_lock
, flags
);
208 ibmr
= llist_entry(ret
, struct rds_ib_mr
, llnode
);
209 if (pool
->pool_type
== RDS_IB_MR_8K_POOL
)
210 rds_ib_stats_inc(s_ib_rdma_mr_8k_reused
);
212 rds_ib_stats_inc(s_ib_rdma_mr_1m_reused
);
218 void rds_ib_sync_mr(void *trans_private
, int direction
)
220 struct rds_ib_mr
*ibmr
= trans_private
;
221 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
227 case DMA_FROM_DEVICE
:
228 ib_dma_sync_sg_for_cpu(rds_ibdev
->dev
, ibmr
->sg
,
229 ibmr
->sg_dma_len
, DMA_BIDIRECTIONAL
);
232 ib_dma_sync_sg_for_device(rds_ibdev
->dev
, ibmr
->sg
,
233 ibmr
->sg_dma_len
, DMA_BIDIRECTIONAL
);
238 void __rds_ib_teardown_mr(struct rds_ib_mr
*ibmr
)
240 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
242 if (ibmr
->sg_dma_len
) {
243 ib_dma_unmap_sg(rds_ibdev
->dev
,
244 ibmr
->sg
, ibmr
->sg_len
,
246 ibmr
->sg_dma_len
= 0;
249 /* Release the s/g list */
253 for (i
= 0; i
< ibmr
->sg_len
; ++i
) {
254 struct page
*page
= sg_page(&ibmr
->sg
[i
]);
256 /* FIXME we need a way to tell a r/w MR
258 WARN_ON(!page
->mapping
&& irqs_disabled());
259 set_page_dirty(page
);
269 void rds_ib_teardown_mr(struct rds_ib_mr
*ibmr
)
271 unsigned int pinned
= ibmr
->sg_len
;
273 __rds_ib_teardown_mr(ibmr
);
275 struct rds_ib_mr_pool
*pool
= ibmr
->pool
;
277 atomic_sub(pinned
, &pool
->free_pinned
);
281 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool
*pool
, int free_all
)
283 unsigned int item_count
;
285 item_count
= atomic_read(&pool
->item_count
);
293 * given an llist of mrs, put them all into the list_head for more processing
295 static unsigned int llist_append_to_list(struct llist_head
*llist
,
296 struct list_head
*list
)
298 struct rds_ib_mr
*ibmr
;
299 struct llist_node
*node
;
300 struct llist_node
*next
;
301 unsigned int count
= 0;
303 node
= llist_del_all(llist
);
306 ibmr
= llist_entry(node
, struct rds_ib_mr
, llnode
);
307 list_add_tail(&ibmr
->unmap_list
, list
);
315 * this takes a list head of mrs and turns it into linked llist nodes
316 * of clusters. Each cluster has linked llist nodes of
317 * MR_CLUSTER_SIZE mrs that are ready for reuse.
319 static void list_to_llist_nodes(struct list_head
*list
,
320 struct llist_node
**nodes_head
,
321 struct llist_node
**nodes_tail
)
323 struct rds_ib_mr
*ibmr
;
324 struct llist_node
*cur
= NULL
;
325 struct llist_node
**next
= nodes_head
;
327 list_for_each_entry(ibmr
, list
, unmap_list
) {
337 * Flush our pool of MRs.
338 * At a minimum, all currently unused MRs are unmapped.
339 * If the number of MRs allocated exceeds the limit, we also try
340 * to free as many MRs as needed to get back to this limit.
342 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool
*pool
,
343 int free_all
, struct rds_ib_mr
**ibmr_ret
)
345 struct rds_ib_mr
*ibmr
;
346 struct llist_node
*clean_nodes
;
347 struct llist_node
*clean_tail
;
348 LIST_HEAD(unmap_list
);
349 unsigned long unpinned
= 0;
350 unsigned int nfreed
= 0, dirty_to_clean
= 0, free_goal
;
352 if (pool
->pool_type
== RDS_IB_MR_8K_POOL
)
353 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush
);
355 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush
);
359 while (!mutex_trylock(&pool
->flush_lock
)) {
360 ibmr
= rds_ib_reuse_mr(pool
);
363 finish_wait(&pool
->flush_wait
, &wait
);
367 prepare_to_wait(&pool
->flush_wait
, &wait
,
368 TASK_UNINTERRUPTIBLE
);
369 if (llist_empty(&pool
->clean_list
))
372 ibmr
= rds_ib_reuse_mr(pool
);
375 finish_wait(&pool
->flush_wait
, &wait
);
379 finish_wait(&pool
->flush_wait
, &wait
);
381 mutex_lock(&pool
->flush_lock
);
384 ibmr
= rds_ib_reuse_mr(pool
);
391 /* Get the list of all MRs to be dropped. Ordering matters -
392 * we want to put drop_list ahead of free_list.
394 dirty_to_clean
= llist_append_to_list(&pool
->drop_list
, &unmap_list
);
395 dirty_to_clean
+= llist_append_to_list(&pool
->free_list
, &unmap_list
);
399 spin_lock_irqsave(&pool
->clean_lock
, flags
);
400 llist_append_to_list(&pool
->clean_list
, &unmap_list
);
401 spin_unlock_irqrestore(&pool
->clean_lock
, flags
);
404 free_goal
= rds_ib_flush_goal(pool
, free_all
);
406 if (list_empty(&unmap_list
))
409 rds_ib_unreg_frmr(&unmap_list
, &nfreed
, &unpinned
, free_goal
);
411 if (!list_empty(&unmap_list
)) {
414 list_to_llist_nodes(&unmap_list
, &clean_nodes
, &clean_tail
);
416 *ibmr_ret
= llist_entry(clean_nodes
, struct rds_ib_mr
, llnode
);
417 clean_nodes
= clean_nodes
->next
;
419 /* more than one entry in llist nodes */
421 spin_lock_irqsave(&pool
->clean_lock
, flags
);
422 llist_add_batch(clean_nodes
, clean_tail
,
424 spin_unlock_irqrestore(&pool
->clean_lock
, flags
);
428 atomic_sub(unpinned
, &pool
->free_pinned
);
429 atomic_sub(dirty_to_clean
, &pool
->dirty_count
);
430 atomic_sub(nfreed
, &pool
->item_count
);
433 mutex_unlock(&pool
->flush_lock
);
434 if (waitqueue_active(&pool
->flush_wait
))
435 wake_up(&pool
->flush_wait
);
440 struct rds_ib_mr
*rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool
*pool
)
442 struct rds_ib_mr
*ibmr
= NULL
;
446 ibmr
= rds_ib_reuse_mr(pool
);
450 if (atomic_inc_return(&pool
->item_count
) <= pool
->max_items
)
453 atomic_dec(&pool
->item_count
);
456 if (pool
->pool_type
== RDS_IB_MR_8K_POOL
)
457 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted
);
459 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted
);
463 /* We do have some empty MRs. Flush them out. */
464 if (pool
->pool_type
== RDS_IB_MR_8K_POOL
)
465 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait
);
467 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait
);
469 rds_ib_flush_mr_pool(pool
, 0, &ibmr
);
477 static void rds_ib_mr_pool_flush_worker(struct work_struct
*work
)
479 struct rds_ib_mr_pool
*pool
= container_of(work
, struct rds_ib_mr_pool
, flush_worker
.work
);
481 rds_ib_flush_mr_pool(pool
, 0, NULL
);
484 void rds_ib_free_mr(void *trans_private
, int invalidate
)
486 struct rds_ib_mr
*ibmr
= trans_private
;
487 struct rds_ib_mr_pool
*pool
= ibmr
->pool
;
488 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
490 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr
->sg_len
);
493 /* A MR created and marked as use_once. We use delayed work,
494 * because there is a change that we are in interrupt and can't
495 * call to ib_dereg_mr() directly.
497 INIT_DELAYED_WORK(&ibmr
->work
, rds_ib_odp_mr_worker
);
498 queue_delayed_work(rds_ib_mr_wq
, &ibmr
->work
, 0);
502 /* Return it to the pool's free list */
503 rds_ib_free_frmr_list(ibmr
);
505 atomic_add(ibmr
->sg_len
, &pool
->free_pinned
);
506 atomic_inc(&pool
->dirty_count
);
508 /* If we've pinned too many pages, request a flush */
509 if (atomic_read(&pool
->free_pinned
) >= pool
->max_free_pinned
||
510 atomic_read(&pool
->dirty_count
) >= pool
->max_items
/ 5)
511 queue_delayed_work(rds_ib_mr_wq
, &pool
->flush_worker
, 10);
514 if (likely(!in_interrupt())) {
515 rds_ib_flush_mr_pool(pool
, 0, NULL
);
517 /* We get here if the user created a MR marked
518 * as use_once and invalidate at the same time.
520 queue_delayed_work(rds_ib_mr_wq
,
521 &pool
->flush_worker
, 10);
525 rds_ib_dev_put(rds_ibdev
);
528 void rds_ib_flush_mrs(void)
530 struct rds_ib_device
*rds_ibdev
;
532 down_read(&rds_ib_devices_lock
);
533 list_for_each_entry(rds_ibdev
, &rds_ib_devices
, list
) {
534 if (rds_ibdev
->mr_8k_pool
)
535 rds_ib_flush_mr_pool(rds_ibdev
->mr_8k_pool
, 0, NULL
);
537 if (rds_ibdev
->mr_1m_pool
)
538 rds_ib_flush_mr_pool(rds_ibdev
->mr_1m_pool
, 0, NULL
);
540 up_read(&rds_ib_devices_lock
);
543 u32
rds_ib_get_lkey(void *trans_private
)
545 struct rds_ib_mr
*ibmr
= trans_private
;
547 return ibmr
->u
.mr
->lkey
;
550 void *rds_ib_get_mr(struct scatterlist
*sg
, unsigned long nents
,
551 struct rds_sock
*rs
, u32
*key_ret
,
552 struct rds_connection
*conn
,
553 u64 start
, u64 length
, int need_odp
)
555 struct rds_ib_device
*rds_ibdev
;
556 struct rds_ib_mr
*ibmr
= NULL
;
557 struct rds_ib_connection
*ic
= NULL
;
560 rds_ibdev
= rds_ib_get_device(rs
->rs_bound_addr
.s6_addr32
[3]);
566 if (need_odp
== ODP_ZEROBASED
|| need_odp
== ODP_VIRTUAL
) {
567 u64 virt_addr
= need_odp
== ODP_ZEROBASED
? 0 : start
;
569 (IB_ACCESS_LOCAL_WRITE
| IB_ACCESS_REMOTE_READ
|
570 IB_ACCESS_REMOTE_WRITE
| IB_ACCESS_REMOTE_ATOMIC
|
571 IB_ACCESS_ON_DEMAND
);
572 struct ib_sge sge
= {};
575 if (!rds_ibdev
->odp_capable
) {
580 ib_mr
= ib_reg_user_mr(rds_ibdev
->pd
, start
, length
, virt_addr
,
584 rdsdebug("rds_ib_get_user_mr returned %d\n",
586 ret
= PTR_ERR(ib_mr
);
590 *key_ret
= ib_mr
->rkey
;
592 ibmr
= kzalloc(sizeof(*ibmr
), GFP_KERNEL
);
601 sge
.addr
= virt_addr
;
603 sge
.lkey
= ib_mr
->lkey
;
605 ib_advise_mr(rds_ibdev
->pd
,
606 IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE
,
607 IB_UVERBS_ADVISE_MR_FLAG_FLUSH
, &sge
, 1);
612 ic
= conn
->c_transport_data
;
614 if (!rds_ibdev
->mr_8k_pool
|| !rds_ibdev
->mr_1m_pool
) {
619 ibmr
= rds_ib_reg_frmr(rds_ibdev
, ic
, sg
, nents
, key_ret
);
622 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret
);
629 rds_ib_dev_put(rds_ibdev
);
634 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool
*pool
)
636 cancel_delayed_work_sync(&pool
->flush_worker
);
637 rds_ib_flush_mr_pool(pool
, 1, NULL
);
638 WARN_ON(atomic_read(&pool
->item_count
));
639 WARN_ON(atomic_read(&pool
->free_pinned
));
643 struct rds_ib_mr_pool
*rds_ib_create_mr_pool(struct rds_ib_device
*rds_ibdev
,
646 struct rds_ib_mr_pool
*pool
;
648 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
650 return ERR_PTR(-ENOMEM
);
652 pool
->pool_type
= pool_type
;
653 init_llist_head(&pool
->free_list
);
654 init_llist_head(&pool
->drop_list
);
655 init_llist_head(&pool
->clean_list
);
656 spin_lock_init(&pool
->clean_lock
);
657 mutex_init(&pool
->flush_lock
);
658 init_waitqueue_head(&pool
->flush_wait
);
659 INIT_DELAYED_WORK(&pool
->flush_worker
, rds_ib_mr_pool_flush_worker
);
661 if (pool_type
== RDS_IB_MR_1M_POOL
) {
662 /* +1 allows for unaligned MRs */
663 pool
->max_pages
= RDS_MR_1M_MSG_SIZE
+ 1;
664 pool
->max_items
= rds_ibdev
->max_1m_mrs
;
666 /* pool_type == RDS_IB_MR_8K_POOL */
667 pool
->max_pages
= RDS_MR_8K_MSG_SIZE
+ 1;
668 pool
->max_items
= rds_ibdev
->max_8k_mrs
;
671 pool
->max_free_pinned
= pool
->max_items
* pool
->max_pages
/ 4;
672 pool
->max_items_soft
= rds_ibdev
->max_mrs
* 3 / 4;
677 int rds_ib_mr_init(void)
679 rds_ib_mr_wq
= alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM
, 0);
685 /* By the time this is called all the IB devices should have been torn down and
686 * had their pools freed. As each pool is freed its work struct is waited on,
687 * so the pool flushing work queue should be idle by the time we get here.
689 void rds_ib_mr_exit(void)
691 destroy_workqueue(rds_ib_mr_wq
);
694 static void rds_ib_odp_mr_worker(struct work_struct
*work
)
696 struct rds_ib_mr
*ibmr
;
698 ibmr
= container_of(work
, struct rds_ib_mr
, work
.work
);
699 ib_dereg_mr(ibmr
->u
.mr
);