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
42 struct workqueue_struct *rds_ib_mr_wq;
43
44 static void rds_ib_odp_mr_worker(struct work_struct *work);
45
rds_ib_get_device(__be32 ipaddr)46 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
47 {
48 struct rds_ib_device *rds_ibdev;
49 struct rds_ib_ipaddr *i_ipaddr;
50
51 rcu_read_lock();
52 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
53 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
54 if (i_ipaddr->ipaddr == ipaddr) {
55 refcount_inc(&rds_ibdev->refcount);
56 rcu_read_unlock();
57 return rds_ibdev;
58 }
59 }
60 }
61 rcu_read_unlock();
62
63 return NULL;
64 }
65
rds_ib_add_ipaddr(struct rds_ib_device * rds_ibdev,__be32 ipaddr)66 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
67 {
68 struct rds_ib_ipaddr *i_ipaddr;
69
70 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
71 if (!i_ipaddr)
72 return -ENOMEM;
73
74 i_ipaddr->ipaddr = ipaddr;
75
76 spin_lock_irq(&rds_ibdev->spinlock);
77 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
78 spin_unlock_irq(&rds_ibdev->spinlock);
79
80 return 0;
81 }
82
rds_ib_remove_ipaddr(struct rds_ib_device * rds_ibdev,__be32 ipaddr)83 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
84 {
85 struct rds_ib_ipaddr *i_ipaddr;
86 struct rds_ib_ipaddr *to_free = NULL;
87
88
89 spin_lock_irq(&rds_ibdev->spinlock);
90 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
91 if (i_ipaddr->ipaddr == ipaddr) {
92 list_del_rcu(&i_ipaddr->list);
93 to_free = i_ipaddr;
94 break;
95 }
96 }
97 spin_unlock_irq(&rds_ibdev->spinlock);
98
99 if (to_free)
100 kfree_rcu(to_free, rcu);
101 }
102
rds_ib_update_ipaddr(struct rds_ib_device * rds_ibdev,struct in6_addr * ipaddr)103 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
104 struct in6_addr *ipaddr)
105 {
106 struct rds_ib_device *rds_ibdev_old;
107
108 rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
109 if (!rds_ibdev_old)
110 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
111
112 if (rds_ibdev_old != rds_ibdev) {
113 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]);
114 rds_ib_dev_put(rds_ibdev_old);
115 return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
116 }
117 rds_ib_dev_put(rds_ibdev_old);
118
119 return 0;
120 }
121
rds_ib_add_conn(struct rds_ib_device * rds_ibdev,struct rds_connection * conn)122 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
123 {
124 struct rds_ib_connection *ic = conn->c_transport_data;
125
126 /* conn was previously on the nodev_conns_list */
127 spin_lock_irq(&ib_nodev_conns_lock);
128 BUG_ON(list_empty(&ib_nodev_conns));
129 BUG_ON(list_empty(&ic->ib_node));
130 list_del(&ic->ib_node);
131
132 spin_lock(&rds_ibdev->spinlock);
133 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
134 spin_unlock(&rds_ibdev->spinlock);
135 spin_unlock_irq(&ib_nodev_conns_lock);
136
137 ic->rds_ibdev = rds_ibdev;
138 refcount_inc(&rds_ibdev->refcount);
139 }
140
rds_ib_remove_conn(struct rds_ib_device * rds_ibdev,struct rds_connection * conn)141 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
142 {
143 struct rds_ib_connection *ic = conn->c_transport_data;
144
145 /* place conn on nodev_conns_list */
146 spin_lock(&ib_nodev_conns_lock);
147
148 spin_lock_irq(&rds_ibdev->spinlock);
149 BUG_ON(list_empty(&ic->ib_node));
150 list_del(&ic->ib_node);
151 spin_unlock_irq(&rds_ibdev->spinlock);
152
153 list_add_tail(&ic->ib_node, &ib_nodev_conns);
154
155 spin_unlock(&ib_nodev_conns_lock);
156
157 ic->rds_ibdev = NULL;
158 rds_ib_dev_put(rds_ibdev);
159 }
160
rds_ib_destroy_nodev_conns(void)161 void rds_ib_destroy_nodev_conns(void)
162 {
163 struct rds_ib_connection *ic, *_ic;
164 LIST_HEAD(tmp_list);
165
166 /* avoid calling conn_destroy with irqs off */
167 spin_lock_irq(&ib_nodev_conns_lock);
168 list_splice(&ib_nodev_conns, &tmp_list);
169 spin_unlock_irq(&ib_nodev_conns_lock);
170
171 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
172 rds_conn_destroy(ic->conn);
173 }
174
rds_ib_get_mr_info(struct rds_ib_device * rds_ibdev,struct rds_info_rdma_connection * iinfo)175 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
176 {
177 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
178
179 iinfo->rdma_mr_max = pool_1m->max_items;
180 iinfo->rdma_mr_size = pool_1m->max_pages;
181 }
182
183 #if IS_ENABLED(CONFIG_IPV6)
rds6_ib_get_mr_info(struct rds_ib_device * rds_ibdev,struct rds6_info_rdma_connection * iinfo6)184 void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
185 struct rds6_info_rdma_connection *iinfo6)
186 {
187 struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
188
189 iinfo6->rdma_mr_max = pool_1m->max_items;
190 iinfo6->rdma_mr_size = pool_1m->max_pages;
191 }
192 #endif
193
rds_ib_reuse_mr(struct rds_ib_mr_pool * pool)194 struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
195 {
196 struct rds_ib_mr *ibmr = NULL;
197 struct llist_node *ret;
198 unsigned long flags;
199
200 spin_lock_irqsave(&pool->clean_lock, flags);
201 ret = llist_del_first(&pool->clean_list);
202 spin_unlock_irqrestore(&pool->clean_lock, flags);
203 if (ret) {
204 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
205 if (pool->pool_type == RDS_IB_MR_8K_POOL)
206 rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
207 else
208 rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
209 }
210
211 return ibmr;
212 }
213
rds_ib_sync_mr(void * trans_private,int direction)214 void rds_ib_sync_mr(void *trans_private, int direction)
215 {
216 struct rds_ib_mr *ibmr = trans_private;
217 struct rds_ib_device *rds_ibdev = ibmr->device;
218
219 if (ibmr->odp)
220 return;
221
222 switch (direction) {
223 case DMA_FROM_DEVICE:
224 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
225 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
226 break;
227 case DMA_TO_DEVICE:
228 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
229 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
230 break;
231 }
232 }
233
__rds_ib_teardown_mr(struct rds_ib_mr * ibmr)234 void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
235 {
236 struct rds_ib_device *rds_ibdev = ibmr->device;
237
238 if (ibmr->sg_dma_len) {
239 ib_dma_unmap_sg(rds_ibdev->dev,
240 ibmr->sg, ibmr->sg_len,
241 DMA_BIDIRECTIONAL);
242 ibmr->sg_dma_len = 0;
243 }
244
245 /* Release the s/g list */
246 if (ibmr->sg_len) {
247 unsigned int i;
248
249 for (i = 0; i < ibmr->sg_len; ++i) {
250 struct page *page = sg_page(&ibmr->sg[i]);
251
252 /* FIXME we need a way to tell a r/w MR
253 * from a r/o MR */
254 WARN_ON(!page->mapping && irqs_disabled());
255 set_page_dirty(page);
256 put_page(page);
257 }
258 kfree(ibmr->sg);
259
260 ibmr->sg = NULL;
261 ibmr->sg_len = 0;
262 }
263 }
264
rds_ib_teardown_mr(struct rds_ib_mr * ibmr)265 void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
266 {
267 unsigned int pinned = ibmr->sg_len;
268
269 __rds_ib_teardown_mr(ibmr);
270 if (pinned) {
271 struct rds_ib_mr_pool *pool = ibmr->pool;
272
273 atomic_sub(pinned, &pool->free_pinned);
274 }
275 }
276
rds_ib_flush_goal(struct rds_ib_mr_pool * pool,int free_all)277 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
278 {
279 unsigned int item_count;
280
281 item_count = atomic_read(&pool->item_count);
282 if (free_all)
283 return item_count;
284
285 return 0;
286 }
287
288 /*
289 * given an llist of mrs, put them all into the list_head for more processing
290 */
llist_append_to_list(struct llist_head * llist,struct list_head * list)291 static unsigned int llist_append_to_list(struct llist_head *llist,
292 struct list_head *list)
293 {
294 struct rds_ib_mr *ibmr;
295 struct llist_node *node;
296 struct llist_node *next;
297 unsigned int count = 0;
298
299 node = llist_del_all(llist);
300 while (node) {
301 next = node->next;
302 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
303 list_add_tail(&ibmr->unmap_list, list);
304 node = next;
305 count++;
306 }
307 return count;
308 }
309
310 /*
311 * this takes a list head of mrs and turns it into linked llist nodes
312 * of clusters. Each cluster has linked llist nodes of
313 * MR_CLUSTER_SIZE mrs that are ready for reuse.
314 */
list_to_llist_nodes(struct list_head * list,struct llist_node ** nodes_head,struct llist_node ** nodes_tail)315 static void list_to_llist_nodes(struct list_head *list,
316 struct llist_node **nodes_head,
317 struct llist_node **nodes_tail)
318 {
319 struct rds_ib_mr *ibmr;
320 struct llist_node *cur = NULL;
321 struct llist_node **next = nodes_head;
322
323 list_for_each_entry(ibmr, list, unmap_list) {
324 cur = &ibmr->llnode;
325 *next = cur;
326 next = &cur->next;
327 }
328 *next = NULL;
329 *nodes_tail = cur;
330 }
331
332 /*
333 * Flush our pool of MRs.
334 * At a minimum, all currently unused MRs are unmapped.
335 * If the number of MRs allocated exceeds the limit, we also try
336 * to free as many MRs as needed to get back to this limit.
337 */
rds_ib_flush_mr_pool(struct rds_ib_mr_pool * pool,int free_all,struct rds_ib_mr ** ibmr_ret)338 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
339 int free_all, struct rds_ib_mr **ibmr_ret)
340 {
341 struct rds_ib_mr *ibmr;
342 struct llist_node *clean_nodes;
343 struct llist_node *clean_tail;
344 LIST_HEAD(unmap_list);
345 unsigned long unpinned = 0;
346 unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
347
348 if (pool->pool_type == RDS_IB_MR_8K_POOL)
349 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
350 else
351 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
352
353 if (ibmr_ret) {
354 DEFINE_WAIT(wait);
355 while (!mutex_trylock(&pool->flush_lock)) {
356 ibmr = rds_ib_reuse_mr(pool);
357 if (ibmr) {
358 *ibmr_ret = ibmr;
359 finish_wait(&pool->flush_wait, &wait);
360 goto out_nolock;
361 }
362
363 prepare_to_wait(&pool->flush_wait, &wait,
364 TASK_UNINTERRUPTIBLE);
365 if (llist_empty(&pool->clean_list))
366 schedule();
367
368 ibmr = rds_ib_reuse_mr(pool);
369 if (ibmr) {
370 *ibmr_ret = ibmr;
371 finish_wait(&pool->flush_wait, &wait);
372 goto out_nolock;
373 }
374 }
375 finish_wait(&pool->flush_wait, &wait);
376 } else
377 mutex_lock(&pool->flush_lock);
378
379 if (ibmr_ret) {
380 ibmr = rds_ib_reuse_mr(pool);
381 if (ibmr) {
382 *ibmr_ret = ibmr;
383 goto out;
384 }
385 }
386
387 /* Get the list of all MRs to be dropped. Ordering matters -
388 * we want to put drop_list ahead of free_list.
389 */
390 dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
391 dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
392 if (free_all) {
393 unsigned long flags;
394
395 spin_lock_irqsave(&pool->clean_lock, flags);
396 llist_append_to_list(&pool->clean_list, &unmap_list);
397 spin_unlock_irqrestore(&pool->clean_lock, flags);
398 }
399
400 free_goal = rds_ib_flush_goal(pool, free_all);
401
402 if (list_empty(&unmap_list))
403 goto out;
404
405 rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
406
407 if (!list_empty(&unmap_list)) {
408 unsigned long flags;
409
410 list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
411 if (ibmr_ret) {
412 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
413 clean_nodes = clean_nodes->next;
414 }
415 /* more than one entry in llist nodes */
416 if (clean_nodes) {
417 spin_lock_irqsave(&pool->clean_lock, flags);
418 llist_add_batch(clean_nodes, clean_tail,
419 &pool->clean_list);
420 spin_unlock_irqrestore(&pool->clean_lock, flags);
421 }
422 }
423
424 atomic_sub(unpinned, &pool->free_pinned);
425 atomic_sub(dirty_to_clean, &pool->dirty_count);
426 atomic_sub(nfreed, &pool->item_count);
427
428 out:
429 mutex_unlock(&pool->flush_lock);
430 if (waitqueue_active(&pool->flush_wait))
431 wake_up(&pool->flush_wait);
432 out_nolock:
433 return 0;
434 }
435
rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool * pool)436 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
437 {
438 struct rds_ib_mr *ibmr = NULL;
439 int iter = 0;
440
441 while (1) {
442 ibmr = rds_ib_reuse_mr(pool);
443 if (ibmr)
444 return ibmr;
445
446 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
447 break;
448
449 atomic_dec(&pool->item_count);
450
451 if (++iter > 2) {
452 if (pool->pool_type == RDS_IB_MR_8K_POOL)
453 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
454 else
455 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
456 break;
457 }
458
459 /* We do have some empty MRs. Flush them out. */
460 if (pool->pool_type == RDS_IB_MR_8K_POOL)
461 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
462 else
463 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
464
465 rds_ib_flush_mr_pool(pool, 0, &ibmr);
466 if (ibmr)
467 return ibmr;
468 }
469
470 return NULL;
471 }
472
rds_ib_mr_pool_flush_worker(struct work_struct * work)473 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
474 {
475 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
476
477 rds_ib_flush_mr_pool(pool, 0, NULL);
478 }
479
rds_ib_free_mr(void * trans_private,int invalidate)480 void rds_ib_free_mr(void *trans_private, int invalidate)
481 {
482 struct rds_ib_mr *ibmr = trans_private;
483 struct rds_ib_mr_pool *pool = ibmr->pool;
484 struct rds_ib_device *rds_ibdev = ibmr->device;
485
486 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
487
488 if (ibmr->odp) {
489 /* A MR created and marked as use_once. We use delayed work,
490 * because there is a change that we are in interrupt and can't
491 * call to ib_dereg_mr() directly.
492 */
493 INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
494 queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
495 return;
496 }
497
498 /* Return it to the pool's free list */
499 rds_ib_free_frmr_list(ibmr);
500
501 atomic_add(ibmr->sg_len, &pool->free_pinned);
502 atomic_inc(&pool->dirty_count);
503
504 /* If we've pinned too many pages, request a flush */
505 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
506 atomic_read(&pool->dirty_count) >= pool->max_items / 5)
507 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
508
509 if (invalidate) {
510 if (likely(!in_interrupt())) {
511 rds_ib_flush_mr_pool(pool, 0, NULL);
512 } else {
513 /* We get here if the user created a MR marked
514 * as use_once and invalidate at the same time.
515 */
516 queue_delayed_work(rds_ib_mr_wq,
517 &pool->flush_worker, 10);
518 }
519 }
520
521 rds_ib_dev_put(rds_ibdev);
522 }
523
rds_ib_flush_mrs(void)524 void rds_ib_flush_mrs(void)
525 {
526 struct rds_ib_device *rds_ibdev;
527
528 down_read(&rds_ib_devices_lock);
529 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
530 if (rds_ibdev->mr_8k_pool)
531 rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
532
533 if (rds_ibdev->mr_1m_pool)
534 rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
535 }
536 up_read(&rds_ib_devices_lock);
537 }
538
rds_ib_get_lkey(void * trans_private)539 u32 rds_ib_get_lkey(void *trans_private)
540 {
541 struct rds_ib_mr *ibmr = trans_private;
542
543 return ibmr->u.mr->lkey;
544 }
545
rds_ib_get_mr(struct scatterlist * sg,unsigned long nents,struct rds_sock * rs,u32 * key_ret,struct rds_connection * conn,u64 start,u64 length,int need_odp)546 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
547 struct rds_sock *rs, u32 *key_ret,
548 struct rds_connection *conn,
549 u64 start, u64 length, int need_odp)
550 {
551 struct rds_ib_device *rds_ibdev;
552 struct rds_ib_mr *ibmr = NULL;
553 struct rds_ib_connection *ic = NULL;
554 int ret;
555
556 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
557 if (!rds_ibdev) {
558 ret = -ENODEV;
559 goto out;
560 }
561
562 if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
563 u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
564 int access_flags =
565 (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
566 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
567 IB_ACCESS_ON_DEMAND);
568 struct ib_sge sge = {};
569 struct ib_mr *ib_mr;
570
571 if (!rds_ibdev->odp_capable) {
572 ret = -EOPNOTSUPP;
573 goto out;
574 }
575
576 ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
577 access_flags);
578
579 if (IS_ERR(ib_mr)) {
580 rdsdebug("rds_ib_get_user_mr returned %d\n",
581 IS_ERR(ib_mr));
582 ret = PTR_ERR(ib_mr);
583 goto out;
584 }
585 if (key_ret)
586 *key_ret = ib_mr->rkey;
587
588 ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
589 if (!ibmr) {
590 ib_dereg_mr(ib_mr);
591 ret = -ENOMEM;
592 goto out;
593 }
594 ibmr->u.mr = ib_mr;
595 ibmr->odp = 1;
596
597 sge.addr = virt_addr;
598 sge.length = length;
599 sge.lkey = ib_mr->lkey;
600
601 ib_advise_mr(rds_ibdev->pd,
602 IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
603 IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
604 return ibmr;
605 }
606
607 if (conn)
608 ic = conn->c_transport_data;
609
610 if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
611 ret = -ENODEV;
612 goto out;
613 }
614
615 ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
616 if (IS_ERR(ibmr)) {
617 ret = PTR_ERR(ibmr);
618 pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
619 } else {
620 return ibmr;
621 }
622
623 out:
624 if (rds_ibdev)
625 rds_ib_dev_put(rds_ibdev);
626
627 return ERR_PTR(ret);
628 }
629
rds_ib_destroy_mr_pool(struct rds_ib_mr_pool * pool)630 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
631 {
632 cancel_delayed_work_sync(&pool->flush_worker);
633 rds_ib_flush_mr_pool(pool, 1, NULL);
634 WARN_ON(atomic_read(&pool->item_count));
635 WARN_ON(atomic_read(&pool->free_pinned));
636 kfree(pool);
637 }
638
rds_ib_create_mr_pool(struct rds_ib_device * rds_ibdev,int pool_type)639 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
640 int pool_type)
641 {
642 struct rds_ib_mr_pool *pool;
643
644 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
645 if (!pool)
646 return ERR_PTR(-ENOMEM);
647
648 pool->pool_type = pool_type;
649 init_llist_head(&pool->free_list);
650 init_llist_head(&pool->drop_list);
651 init_llist_head(&pool->clean_list);
652 spin_lock_init(&pool->clean_lock);
653 mutex_init(&pool->flush_lock);
654 init_waitqueue_head(&pool->flush_wait);
655 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
656
657 if (pool_type == RDS_IB_MR_1M_POOL) {
658 /* +1 allows for unaligned MRs */
659 pool->max_pages = RDS_MR_1M_MSG_SIZE + 1;
660 pool->max_items = rds_ibdev->max_1m_mrs;
661 } else {
662 /* pool_type == RDS_IB_MR_8K_POOL */
663 pool->max_pages = RDS_MR_8K_MSG_SIZE + 1;
664 pool->max_items = rds_ibdev->max_8k_mrs;
665 }
666
667 pool->max_free_pinned = pool->max_items * pool->max_pages / 4;
668 pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
669
670 return pool;
671 }
672
rds_ib_mr_init(void)673 int rds_ib_mr_init(void)
674 {
675 rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
676 if (!rds_ib_mr_wq)
677 return -ENOMEM;
678 return 0;
679 }
680
681 /* By the time this is called all the IB devices should have been torn down and
682 * had their pools freed. As each pool is freed its work struct is waited on,
683 * so the pool flushing work queue should be idle by the time we get here.
684 */
rds_ib_mr_exit(void)685 void rds_ib_mr_exit(void)
686 {
687 destroy_workqueue(rds_ib_mr_wq);
688 }
689
rds_ib_odp_mr_worker(struct work_struct * work)690 static void rds_ib_odp_mr_worker(struct work_struct *work)
691 {
692 struct rds_ib_mr *ibmr;
693
694 ibmr = container_of(work, struct rds_ib_mr, work.work);
695 ib_dereg_mr(ibmr->u.mr);
696 kfree(ibmr);
697 }
698