Loading...
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/slab.h>
35#include <linux/rculist.h>
36#include <linux/llist.h>
37
38#include "rds_single_path.h"
39#include "ib_mr.h"
40#include "rds.h"
41
42struct workqueue_struct *rds_ib_mr_wq;
43struct rds_ib_dereg_odp_mr {
44 struct work_struct work;
45 struct ib_mr *mr;
46};
47
48static void rds_ib_odp_mr_worker(struct work_struct *work);
49
50static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
51{
52 struct rds_ib_device *rds_ibdev;
53 struct rds_ib_ipaddr *i_ipaddr;
54
55 rcu_read_lock();
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);
60 rcu_read_unlock();
61 return rds_ibdev;
62 }
63 }
64 }
65 rcu_read_unlock();
66
67 return NULL;
68}
69
70static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
71{
72 struct rds_ib_ipaddr *i_ipaddr;
73
74 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
75 if (!i_ipaddr)
76 return -ENOMEM;
77
78 i_ipaddr->ipaddr = ipaddr;
79
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);
83
84 return 0;
85}
86
87static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
88{
89 struct rds_ib_ipaddr *i_ipaddr;
90 struct rds_ib_ipaddr *to_free = NULL;
91
92
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);
97 to_free = i_ipaddr;
98 break;
99 }
100 }
101 spin_unlock_irq(&rds_ibdev->spinlock);
102
103 if (to_free)
104 kfree_rcu(to_free, rcu);
105}
106
107int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
108 struct in6_addr *ipaddr)
109{
110 struct rds_ib_device *rds_ibdev_old;
111
112 rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
113 if (!rds_ibdev_old)
114 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
115
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]);
120 }
121 rds_ib_dev_put(rds_ibdev_old);
122
123 return 0;
124}
125
126void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
127{
128 struct rds_ib_connection *ic = conn->c_transport_data;
129
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);
135
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);
140
141 ic->rds_ibdev = rds_ibdev;
142 refcount_inc(&rds_ibdev->refcount);
143}
144
145void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
146{
147 struct rds_ib_connection *ic = conn->c_transport_data;
148
149 /* place conn on nodev_conns_list */
150 spin_lock(&ib_nodev_conns_lock);
151
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);
156
157 list_add_tail(&ic->ib_node, &ib_nodev_conns);
158
159 spin_unlock(&ib_nodev_conns_lock);
160
161 ic->rds_ibdev = NULL;
162 rds_ib_dev_put(rds_ibdev);
163}
164
165void rds_ib_destroy_nodev_conns(void)
166{
167 struct rds_ib_connection *ic, *_ic;
168 LIST_HEAD(tmp_list);
169
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);
174
175 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
176 rds_conn_destroy(ic->conn);
177}
178
179void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
180{
181 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
182
183 iinfo->rdma_mr_max = pool_1m->max_items;
184 iinfo->rdma_mr_size = pool_1m->max_pages;
185}
186
187#if IS_ENABLED(CONFIG_IPV6)
188void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
189 struct rds6_info_rdma_connection *iinfo6)
190{
191 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
192
193 iinfo6->rdma_mr_max = pool_1m->max_items;
194 iinfo6->rdma_mr_size = pool_1m->max_pages;
195}
196#endif
197
198struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
199{
200 struct rds_ib_mr *ibmr = NULL;
201 struct llist_node *ret;
202 unsigned long flags;
203
204 spin_lock_irqsave(&pool->clean_lock, flags);
205 ret = llist_del_first(&pool->clean_list);
206 spin_unlock_irqrestore(&pool->clean_lock, flags);
207 if (ret) {
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);
211 else
212 rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
213 }
214
215 return ibmr;
216}
217
218void rds_ib_sync_mr(void *trans_private, int direction)
219{
220 struct rds_ib_mr *ibmr = trans_private;
221 struct rds_ib_device *rds_ibdev = ibmr->device;
222
223 if (ibmr->odp)
224 return;
225
226 switch (direction) {
227 case DMA_FROM_DEVICE:
228 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
229 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
230 break;
231 case DMA_TO_DEVICE:
232 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
233 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
234 break;
235 }
236}
237
238void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
239{
240 struct rds_ib_device *rds_ibdev = ibmr->device;
241
242 if (ibmr->sg_dma_len) {
243 ib_dma_unmap_sg(rds_ibdev->dev,
244 ibmr->sg, ibmr->sg_len,
245 DMA_BIDIRECTIONAL);
246 ibmr->sg_dma_len = 0;
247 }
248
249 /* Release the s/g list */
250 if (ibmr->sg_len) {
251 unsigned int i;
252
253 for (i = 0; i < ibmr->sg_len; ++i) {
254 struct page *page = sg_page(&ibmr->sg[i]);
255
256 /* FIXME we need a way to tell a r/w MR
257 * from a r/o MR */
258 WARN_ON(!page->mapping && irqs_disabled());
259 set_page_dirty(page);
260 put_page(page);
261 }
262 kfree(ibmr->sg);
263
264 ibmr->sg = NULL;
265 ibmr->sg_len = 0;
266 }
267}
268
269void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
270{
271 unsigned int pinned = ibmr->sg_len;
272
273 __rds_ib_teardown_mr(ibmr);
274 if (pinned) {
275 struct rds_ib_mr_pool *pool = ibmr->pool;
276
277 atomic_sub(pinned, &pool->free_pinned);
278 }
279}
280
281static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
282{
283 unsigned int item_count;
284
285 item_count = atomic_read(&pool->item_count);
286 if (free_all)
287 return item_count;
288
289 return 0;
290}
291
292/*
293 * given an llist of mrs, put them all into the list_head for more processing
294 */
295static unsigned int llist_append_to_list(struct llist_head *llist,
296 struct list_head *list)
297{
298 struct rds_ib_mr *ibmr;
299 struct llist_node *node;
300 struct llist_node *next;
301 unsigned int count = 0;
302
303 node = llist_del_all(llist);
304 while (node) {
305 next = node->next;
306 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
307 list_add_tail(&ibmr->unmap_list, list);
308 node = next;
309 count++;
310 }
311 return count;
312}
313
314/*
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.
318 */
319static void list_to_llist_nodes(struct list_head *list,
320 struct llist_node **nodes_head,
321 struct llist_node **nodes_tail)
322{
323 struct rds_ib_mr *ibmr;
324 struct llist_node *cur = NULL;
325 struct llist_node **next = nodes_head;
326
327 list_for_each_entry(ibmr, list, unmap_list) {
328 cur = &ibmr->llnode;
329 *next = cur;
330 next = &cur->next;
331 }
332 *next = NULL;
333 *nodes_tail = cur;
334}
335
336/*
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.
341 */
342int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
343 int free_all, struct rds_ib_mr **ibmr_ret)
344{
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;
351
352 if (pool->pool_type == RDS_IB_MR_8K_POOL)
353 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
354 else
355 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
356
357 if (ibmr_ret) {
358 DEFINE_WAIT(wait);
359 while (!mutex_trylock(&pool->flush_lock)) {
360 ibmr = rds_ib_reuse_mr(pool);
361 if (ibmr) {
362 *ibmr_ret = ibmr;
363 finish_wait(&pool->flush_wait, &wait);
364 goto out_nolock;
365 }
366
367 prepare_to_wait(&pool->flush_wait, &wait,
368 TASK_UNINTERRUPTIBLE);
369 if (llist_empty(&pool->clean_list))
370 schedule();
371
372 ibmr = rds_ib_reuse_mr(pool);
373 if (ibmr) {
374 *ibmr_ret = ibmr;
375 finish_wait(&pool->flush_wait, &wait);
376 goto out_nolock;
377 }
378 }
379 finish_wait(&pool->flush_wait, &wait);
380 } else
381 mutex_lock(&pool->flush_lock);
382
383 if (ibmr_ret) {
384 ibmr = rds_ib_reuse_mr(pool);
385 if (ibmr) {
386 *ibmr_ret = ibmr;
387 goto out;
388 }
389 }
390
391 /* Get the list of all MRs to be dropped. Ordering matters -
392 * we want to put drop_list ahead of free_list.
393 */
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);
396 if (free_all) {
397 unsigned long flags;
398
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);
402 }
403
404 free_goal = rds_ib_flush_goal(pool, free_all);
405
406 if (list_empty(&unmap_list))
407 goto out;
408
409 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
410
411 if (!list_empty(&unmap_list)) {
412 unsigned long flags;
413
414 list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
415 if (ibmr_ret) {
416 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
417 clean_nodes = clean_nodes->next;
418 }
419 /* more than one entry in llist nodes */
420 if (clean_nodes) {
421 spin_lock_irqsave(&pool->clean_lock, flags);
422 llist_add_batch(clean_nodes, clean_tail,
423 &pool->clean_list);
424 spin_unlock_irqrestore(&pool->clean_lock, flags);
425 }
426 }
427
428 atomic_sub(unpinned, &pool->free_pinned);
429 atomic_sub(dirty_to_clean, &pool->dirty_count);
430 atomic_sub(nfreed, &pool->item_count);
431
432out:
433 mutex_unlock(&pool->flush_lock);
434 if (waitqueue_active(&pool->flush_wait))
435 wake_up(&pool->flush_wait);
436out_nolock:
437 return 0;
438}
439
440struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
441{
442 struct rds_ib_mr *ibmr = NULL;
443 int iter = 0;
444
445 while (1) {
446 ibmr = rds_ib_reuse_mr(pool);
447 if (ibmr)
448 return ibmr;
449
450 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
451 break;
452
453 atomic_dec(&pool->item_count);
454
455 if (++iter > 2) {
456 if (pool->pool_type == RDS_IB_MR_8K_POOL)
457 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
458 else
459 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
460 break;
461 }
462
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);
466 else
467 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
468
469 rds_ib_flush_mr_pool(pool, 0, &ibmr);
470 if (ibmr)
471 return ibmr;
472 }
473
474 return NULL;
475}
476
477static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
478{
479 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
480
481 rds_ib_flush_mr_pool(pool, 0, NULL);
482}
483
484void rds_ib_free_mr(void *trans_private, int invalidate)
485{
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;
489
490 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
491
492 if (ibmr->odp) {
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.
496 */
497 INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
498 queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
499 return;
500 }
501
502 /* Return it to the pool's free list */
503 rds_ib_free_frmr_list(ibmr);
504
505 atomic_add(ibmr->sg_len, &pool->free_pinned);
506 atomic_inc(&pool->dirty_count);
507
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);
512
513 if (invalidate) {
514 if (likely(!in_interrupt())) {
515 rds_ib_flush_mr_pool(pool, 0, NULL);
516 } else {
517 /* We get here if the user created a MR marked
518 * as use_once and invalidate at the same time.
519 */
520 queue_delayed_work(rds_ib_mr_wq,
521 &pool->flush_worker, 10);
522 }
523 }
524
525 rds_ib_dev_put(rds_ibdev);
526}
527
528void rds_ib_flush_mrs(void)
529{
530 struct rds_ib_device *rds_ibdev;
531
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);
536
537 if (rds_ibdev->mr_1m_pool)
538 rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
539 }
540 up_read(&rds_ib_devices_lock);
541}
542
543u32 rds_ib_get_lkey(void *trans_private)
544{
545 struct rds_ib_mr *ibmr = trans_private;
546
547 return ibmr->u.mr->lkey;
548}
549
550void *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)
554{
555 struct rds_ib_device *rds_ibdev;
556 struct rds_ib_mr *ibmr = NULL;
557 struct rds_ib_connection *ic = NULL;
558 int ret;
559
560 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
561 if (!rds_ibdev) {
562 ret = -ENODEV;
563 goto out;
564 }
565
566 if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
567 u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
568 int access_flags =
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 = {};
573 struct ib_mr *ib_mr;
574
575 if (!rds_ibdev->odp_capable) {
576 ret = -EOPNOTSUPP;
577 goto out;
578 }
579
580 ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
581 access_flags);
582
583 if (IS_ERR(ib_mr)) {
584 rdsdebug("rds_ib_get_user_mr returned %d\n",
585 IS_ERR(ib_mr));
586 ret = PTR_ERR(ib_mr);
587 goto out;
588 }
589 if (key_ret)
590 *key_ret = ib_mr->rkey;
591
592 ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
593 if (!ibmr) {
594 ib_dereg_mr(ib_mr);
595 ret = -ENOMEM;
596 goto out;
597 }
598 ibmr->u.mr = ib_mr;
599 ibmr->odp = 1;
600
601 sge.addr = virt_addr;
602 sge.length = length;
603 sge.lkey = ib_mr->lkey;
604
605 ib_advise_mr(rds_ibdev->pd,
606 IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
607 IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
608 return ibmr;
609 }
610
611 if (conn)
612 ic = conn->c_transport_data;
613
614 if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
615 ret = -ENODEV;
616 goto out;
617 }
618
619 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
620 if (IS_ERR(ibmr)) {
621 ret = PTR_ERR(ibmr);
622 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
623 } else {
624 return ibmr;
625 }
626
627 out:
628 if (rds_ibdev)
629 rds_ib_dev_put(rds_ibdev);
630
631 return ERR_PTR(ret);
632}
633
634void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
635{
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));
640 kfree(pool);
641}
642
643struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
644 int pool_type)
645{
646 struct rds_ib_mr_pool *pool;
647
648 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
649 if (!pool)
650 return ERR_PTR(-ENOMEM);
651
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);
660
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;
665 } else {
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;
669 }
670
671 pool->max_free_pinned = pool->max_items * pool->max_pages / 4;
672 pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
673
674 return pool;
675}
676
677int rds_ib_mr_init(void)
678{
679 rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
680 if (!rds_ib_mr_wq)
681 return -ENOMEM;
682 return 0;
683}
684
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.
688 */
689void rds_ib_mr_exit(void)
690{
691 destroy_workqueue(rds_ib_mr_wq);
692}
693
694static void rds_ib_odp_mr_worker(struct work_struct *work)
695{
696 struct rds_ib_mr *ibmr;
697
698 ibmr = container_of(work, struct rds_ib_mr, work.work);
699 ib_dereg_mr(ibmr->u.mr);
700 kfree(ibmr);
701}
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/slab.h>
35#include <linux/rculist.h>
36#include <linux/llist.h>
37
38#include "rds_single_path.h"
39#include "ib_mr.h"
40
41struct workqueue_struct *rds_ib_mr_wq;
42
43static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
44{
45 struct rds_ib_device *rds_ibdev;
46 struct rds_ib_ipaddr *i_ipaddr;
47
48 rcu_read_lock();
49 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
50 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
51 if (i_ipaddr->ipaddr == ipaddr) {
52 refcount_inc(&rds_ibdev->refcount);
53 rcu_read_unlock();
54 return rds_ibdev;
55 }
56 }
57 }
58 rcu_read_unlock();
59
60 return NULL;
61}
62
63static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
64{
65 struct rds_ib_ipaddr *i_ipaddr;
66
67 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
68 if (!i_ipaddr)
69 return -ENOMEM;
70
71 i_ipaddr->ipaddr = ipaddr;
72
73 spin_lock_irq(&rds_ibdev->spinlock);
74 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
75 spin_unlock_irq(&rds_ibdev->spinlock);
76
77 return 0;
78}
79
80static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
81{
82 struct rds_ib_ipaddr *i_ipaddr;
83 struct rds_ib_ipaddr *to_free = NULL;
84
85
86 spin_lock_irq(&rds_ibdev->spinlock);
87 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
88 if (i_ipaddr->ipaddr == ipaddr) {
89 list_del_rcu(&i_ipaddr->list);
90 to_free = i_ipaddr;
91 break;
92 }
93 }
94 spin_unlock_irq(&rds_ibdev->spinlock);
95
96 if (to_free)
97 kfree_rcu(to_free, rcu);
98}
99
100int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
101 struct in6_addr *ipaddr)
102{
103 struct rds_ib_device *rds_ibdev_old;
104
105 rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
106 if (!rds_ibdev_old)
107 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
108
109 if (rds_ibdev_old != rds_ibdev) {
110 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]);
111 rds_ib_dev_put(rds_ibdev_old);
112 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
113 }
114 rds_ib_dev_put(rds_ibdev_old);
115
116 return 0;
117}
118
119void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
120{
121 struct rds_ib_connection *ic = conn->c_transport_data;
122
123 /* conn was previously on the nodev_conns_list */
124 spin_lock_irq(&ib_nodev_conns_lock);
125 BUG_ON(list_empty(&ib_nodev_conns));
126 BUG_ON(list_empty(&ic->ib_node));
127 list_del(&ic->ib_node);
128
129 spin_lock(&rds_ibdev->spinlock);
130 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
131 spin_unlock(&rds_ibdev->spinlock);
132 spin_unlock_irq(&ib_nodev_conns_lock);
133
134 ic->rds_ibdev = rds_ibdev;
135 refcount_inc(&rds_ibdev->refcount);
136}
137
138void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
139{
140 struct rds_ib_connection *ic = conn->c_transport_data;
141
142 /* place conn on nodev_conns_list */
143 spin_lock(&ib_nodev_conns_lock);
144
145 spin_lock_irq(&rds_ibdev->spinlock);
146 BUG_ON(list_empty(&ic->ib_node));
147 list_del(&ic->ib_node);
148 spin_unlock_irq(&rds_ibdev->spinlock);
149
150 list_add_tail(&ic->ib_node, &ib_nodev_conns);
151
152 spin_unlock(&ib_nodev_conns_lock);
153
154 ic->rds_ibdev = NULL;
155 rds_ib_dev_put(rds_ibdev);
156}
157
158void rds_ib_destroy_nodev_conns(void)
159{
160 struct rds_ib_connection *ic, *_ic;
161 LIST_HEAD(tmp_list);
162
163 /* avoid calling conn_destroy with irqs off */
164 spin_lock_irq(&ib_nodev_conns_lock);
165 list_splice(&ib_nodev_conns, &tmp_list);
166 spin_unlock_irq(&ib_nodev_conns_lock);
167
168 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
169 rds_conn_destroy(ic->conn);
170}
171
172void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
173{
174 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
175
176 iinfo->rdma_mr_max = pool_1m->max_items;
177 iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
178}
179
180#if IS_ENABLED(CONFIG_IPV6)
181void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
182 struct rds6_info_rdma_connection *iinfo6)
183{
184 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
185
186 iinfo6->rdma_mr_max = pool_1m->max_items;
187 iinfo6->rdma_mr_size = pool_1m->fmr_attr.max_pages;
188}
189#endif
190
191struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
192{
193 struct rds_ib_mr *ibmr = NULL;
194 struct llist_node *ret;
195 unsigned long flags;
196
197 spin_lock_irqsave(&pool->clean_lock, flags);
198 ret = llist_del_first(&pool->clean_list);
199 spin_unlock_irqrestore(&pool->clean_lock, flags);
200 if (ret) {
201 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
202 if (pool->pool_type == RDS_IB_MR_8K_POOL)
203 rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
204 else
205 rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
206 }
207
208 return ibmr;
209}
210
211void rds_ib_sync_mr(void *trans_private, int direction)
212{
213 struct rds_ib_mr *ibmr = trans_private;
214 struct rds_ib_device *rds_ibdev = ibmr->device;
215
216 switch (direction) {
217 case DMA_FROM_DEVICE:
218 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
219 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
220 break;
221 case DMA_TO_DEVICE:
222 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
223 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
224 break;
225 }
226}
227
228void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
229{
230 struct rds_ib_device *rds_ibdev = ibmr->device;
231
232 if (ibmr->sg_dma_len) {
233 ib_dma_unmap_sg(rds_ibdev->dev,
234 ibmr->sg, ibmr->sg_len,
235 DMA_BIDIRECTIONAL);
236 ibmr->sg_dma_len = 0;
237 }
238
239 /* Release the s/g list */
240 if (ibmr->sg_len) {
241 unsigned int i;
242
243 for (i = 0; i < ibmr->sg_len; ++i) {
244 struct page *page = sg_page(&ibmr->sg[i]);
245
246 /* FIXME we need a way to tell a r/w MR
247 * from a r/o MR */
248 WARN_ON(!page->mapping && irqs_disabled());
249 set_page_dirty(page);
250 put_page(page);
251 }
252 kfree(ibmr->sg);
253
254 ibmr->sg = NULL;
255 ibmr->sg_len = 0;
256 }
257}
258
259void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
260{
261 unsigned int pinned = ibmr->sg_len;
262
263 __rds_ib_teardown_mr(ibmr);
264 if (pinned) {
265 struct rds_ib_mr_pool *pool = ibmr->pool;
266
267 atomic_sub(pinned, &pool->free_pinned);
268 }
269}
270
271static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
272{
273 unsigned int item_count;
274
275 item_count = atomic_read(&pool->item_count);
276 if (free_all)
277 return item_count;
278
279 return 0;
280}
281
282/*
283 * given an llist of mrs, put them all into the list_head for more processing
284 */
285static unsigned int llist_append_to_list(struct llist_head *llist,
286 struct list_head *list)
287{
288 struct rds_ib_mr *ibmr;
289 struct llist_node *node;
290 struct llist_node *next;
291 unsigned int count = 0;
292
293 node = llist_del_all(llist);
294 while (node) {
295 next = node->next;
296 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
297 list_add_tail(&ibmr->unmap_list, list);
298 node = next;
299 count++;
300 }
301 return count;
302}
303
304/*
305 * this takes a list head of mrs and turns it into linked llist nodes
306 * of clusters. Each cluster has linked llist nodes of
307 * MR_CLUSTER_SIZE mrs that are ready for reuse.
308 */
309static void list_to_llist_nodes(struct list_head *list,
310 struct llist_node **nodes_head,
311 struct llist_node **nodes_tail)
312{
313 struct rds_ib_mr *ibmr;
314 struct llist_node *cur = NULL;
315 struct llist_node **next = nodes_head;
316
317 list_for_each_entry(ibmr, list, unmap_list) {
318 cur = &ibmr->llnode;
319 *next = cur;
320 next = &cur->next;
321 }
322 *next = NULL;
323 *nodes_tail = cur;
324}
325
326/*
327 * Flush our pool of MRs.
328 * At a minimum, all currently unused MRs are unmapped.
329 * If the number of MRs allocated exceeds the limit, we also try
330 * to free as many MRs as needed to get back to this limit.
331 */
332int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
333 int free_all, struct rds_ib_mr **ibmr_ret)
334{
335 struct rds_ib_mr *ibmr;
336 struct llist_node *clean_nodes;
337 struct llist_node *clean_tail;
338 LIST_HEAD(unmap_list);
339 unsigned long unpinned = 0;
340 unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
341
342 if (pool->pool_type == RDS_IB_MR_8K_POOL)
343 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
344 else
345 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
346
347 if (ibmr_ret) {
348 DEFINE_WAIT(wait);
349 while (!mutex_trylock(&pool->flush_lock)) {
350 ibmr = rds_ib_reuse_mr(pool);
351 if (ibmr) {
352 *ibmr_ret = ibmr;
353 finish_wait(&pool->flush_wait, &wait);
354 goto out_nolock;
355 }
356
357 prepare_to_wait(&pool->flush_wait, &wait,
358 TASK_UNINTERRUPTIBLE);
359 if (llist_empty(&pool->clean_list))
360 schedule();
361
362 ibmr = rds_ib_reuse_mr(pool);
363 if (ibmr) {
364 *ibmr_ret = ibmr;
365 finish_wait(&pool->flush_wait, &wait);
366 goto out_nolock;
367 }
368 }
369 finish_wait(&pool->flush_wait, &wait);
370 } else
371 mutex_lock(&pool->flush_lock);
372
373 if (ibmr_ret) {
374 ibmr = rds_ib_reuse_mr(pool);
375 if (ibmr) {
376 *ibmr_ret = ibmr;
377 goto out;
378 }
379 }
380
381 /* Get the list of all MRs to be dropped. Ordering matters -
382 * we want to put drop_list ahead of free_list.
383 */
384 dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
385 dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
386 if (free_all) {
387 unsigned long flags;
388
389 spin_lock_irqsave(&pool->clean_lock, flags);
390 llist_append_to_list(&pool->clean_list, &unmap_list);
391 spin_unlock_irqrestore(&pool->clean_lock, flags);
392 }
393
394 free_goal = rds_ib_flush_goal(pool, free_all);
395
396 if (list_empty(&unmap_list))
397 goto out;
398
399 if (pool->use_fastreg)
400 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
401 else
402 rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
403
404 if (!list_empty(&unmap_list)) {
405 unsigned long flags;
406
407 list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
408 if (ibmr_ret) {
409 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
410 clean_nodes = clean_nodes->next;
411 }
412 /* more than one entry in llist nodes */
413 if (clean_nodes) {
414 spin_lock_irqsave(&pool->clean_lock, flags);
415 llist_add_batch(clean_nodes, clean_tail,
416 &pool->clean_list);
417 spin_unlock_irqrestore(&pool->clean_lock, flags);
418 }
419 }
420
421 atomic_sub(unpinned, &pool->free_pinned);
422 atomic_sub(dirty_to_clean, &pool->dirty_count);
423 atomic_sub(nfreed, &pool->item_count);
424
425out:
426 mutex_unlock(&pool->flush_lock);
427 if (waitqueue_active(&pool->flush_wait))
428 wake_up(&pool->flush_wait);
429out_nolock:
430 return 0;
431}
432
433struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
434{
435 struct rds_ib_mr *ibmr = NULL;
436 int iter = 0;
437
438 while (1) {
439 ibmr = rds_ib_reuse_mr(pool);
440 if (ibmr)
441 return ibmr;
442
443 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
444 break;
445
446 atomic_dec(&pool->item_count);
447
448 if (++iter > 2) {
449 if (pool->pool_type == RDS_IB_MR_8K_POOL)
450 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
451 else
452 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
453 break;
454 }
455
456 /* We do have some empty MRs. Flush them out. */
457 if (pool->pool_type == RDS_IB_MR_8K_POOL)
458 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
459 else
460 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
461
462 rds_ib_flush_mr_pool(pool, 0, &ibmr);
463 if (ibmr)
464 return ibmr;
465 }
466
467 return NULL;
468}
469
470static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
471{
472 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
473
474 rds_ib_flush_mr_pool(pool, 0, NULL);
475}
476
477void rds_ib_free_mr(void *trans_private, int invalidate)
478{
479 struct rds_ib_mr *ibmr = trans_private;
480 struct rds_ib_mr_pool *pool = ibmr->pool;
481 struct rds_ib_device *rds_ibdev = ibmr->device;
482
483 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
484
485 /* Return it to the pool's free list */
486 if (rds_ibdev->use_fastreg)
487 rds_ib_free_frmr_list(ibmr);
488 else
489 rds_ib_free_fmr_list(ibmr);
490
491 atomic_add(ibmr->sg_len, &pool->free_pinned);
492 atomic_inc(&pool->dirty_count);
493
494 /* If we've pinned too many pages, request a flush */
495 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
496 atomic_read(&pool->dirty_count) >= pool->max_items / 5)
497 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
498
499 if (invalidate) {
500 if (likely(!in_interrupt())) {
501 rds_ib_flush_mr_pool(pool, 0, NULL);
502 } else {
503 /* We get here if the user created a MR marked
504 * as use_once and invalidate at the same time.
505 */
506 queue_delayed_work(rds_ib_mr_wq,
507 &pool->flush_worker, 10);
508 }
509 }
510
511 rds_ib_dev_put(rds_ibdev);
512}
513
514void rds_ib_flush_mrs(void)
515{
516 struct rds_ib_device *rds_ibdev;
517
518 down_read(&rds_ib_devices_lock);
519 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
520 if (rds_ibdev->mr_8k_pool)
521 rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
522
523 if (rds_ibdev->mr_1m_pool)
524 rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
525 }
526 up_read(&rds_ib_devices_lock);
527}
528
529void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
530 struct rds_sock *rs, u32 *key_ret,
531 struct rds_connection *conn)
532{
533 struct rds_ib_device *rds_ibdev;
534 struct rds_ib_mr *ibmr = NULL;
535 struct rds_ib_connection *ic = NULL;
536 int ret;
537
538 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
539 if (!rds_ibdev) {
540 ret = -ENODEV;
541 goto out;
542 }
543
544 if (conn)
545 ic = conn->c_transport_data;
546
547 if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
548 ret = -ENODEV;
549 goto out;
550 }
551
552 if (rds_ibdev->use_fastreg)
553 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
554 else
555 ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
556 if (IS_ERR(ibmr)) {
557 ret = PTR_ERR(ibmr);
558 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
559 } else {
560 return ibmr;
561 }
562
563 out:
564 if (rds_ibdev)
565 rds_ib_dev_put(rds_ibdev);
566
567 return ERR_PTR(ret);
568}
569
570void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
571{
572 cancel_delayed_work_sync(&pool->flush_worker);
573 rds_ib_flush_mr_pool(pool, 1, NULL);
574 WARN_ON(atomic_read(&pool->item_count));
575 WARN_ON(atomic_read(&pool->free_pinned));
576 kfree(pool);
577}
578
579struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
580 int pool_type)
581{
582 struct rds_ib_mr_pool *pool;
583
584 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
585 if (!pool)
586 return ERR_PTR(-ENOMEM);
587
588 pool->pool_type = pool_type;
589 init_llist_head(&pool->free_list);
590 init_llist_head(&pool->drop_list);
591 init_llist_head(&pool->clean_list);
592 spin_lock_init(&pool->clean_lock);
593 mutex_init(&pool->flush_lock);
594 init_waitqueue_head(&pool->flush_wait);
595 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
596
597 if (pool_type == RDS_IB_MR_1M_POOL) {
598 /* +1 allows for unaligned MRs */
599 pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1;
600 pool->max_items = rds_ibdev->max_1m_mrs;
601 } else {
602 /* pool_type == RDS_IB_MR_8K_POOL */
603 pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1;
604 pool->max_items = rds_ibdev->max_8k_mrs;
605 }
606
607 pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
608 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
609 pool->fmr_attr.page_shift = PAGE_SHIFT;
610 pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
611 pool->use_fastreg = rds_ibdev->use_fastreg;
612
613 return pool;
614}
615
616int rds_ib_mr_init(void)
617{
618 rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
619 if (!rds_ib_mr_wq)
620 return -ENOMEM;
621 return 0;
622}
623
624/* By the time this is called all the IB devices should have been torn down and
625 * had their pools freed. As each pool is freed its work struct is waited on,
626 * so the pool flushing work queue should be idle by the time we get here.
627 */
628void rds_ib_mr_exit(void)
629{
630 destroy_workqueue(rds_ib_mr_wq);
631}