xref: /openbmc/linux/net/rds/ib_rdma.c (revision dc6a81c3)
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 struct rds_ib_dereg_odp_mr {
44 	struct work_struct work;
45 	struct ib_mr *mr;
46 };
47 
48 static void rds_ib_odp_mr_worker(struct work_struct *work);
49 
50 static 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 
70 static 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 
87 static 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 
107 int 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 
126 void 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 
145 void 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 
165 void 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 
179 void 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->fmr_attr.max_pages;
185 }
186 
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)
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->fmr_attr.max_pages;
195 }
196 #endif
197 
198 struct 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 
218 void 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 
238 void __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 
269 void 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 
281 static 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  */
295 static 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  */
319 static 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  */
342 int 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 	if (pool->use_fastreg)
410 		rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
411 	else
412 		rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
413 
414 	if (!list_empty(&unmap_list)) {
415 		unsigned long flags;
416 
417 		list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
418 		if (ibmr_ret) {
419 			*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
420 			clean_nodes = clean_nodes->next;
421 		}
422 		/* more than one entry in llist nodes */
423 		if (clean_nodes) {
424 			spin_lock_irqsave(&pool->clean_lock, flags);
425 			llist_add_batch(clean_nodes, clean_tail,
426 					&pool->clean_list);
427 			spin_unlock_irqrestore(&pool->clean_lock, flags);
428 		}
429 	}
430 
431 	atomic_sub(unpinned, &pool->free_pinned);
432 	atomic_sub(dirty_to_clean, &pool->dirty_count);
433 	atomic_sub(nfreed, &pool->item_count);
434 
435 out:
436 	mutex_unlock(&pool->flush_lock);
437 	if (waitqueue_active(&pool->flush_wait))
438 		wake_up(&pool->flush_wait);
439 out_nolock:
440 	return 0;
441 }
442 
443 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
444 {
445 	struct rds_ib_mr *ibmr = NULL;
446 	int iter = 0;
447 
448 	while (1) {
449 		ibmr = rds_ib_reuse_mr(pool);
450 		if (ibmr)
451 			return ibmr;
452 
453 		if (atomic_inc_return(&pool->item_count) <= pool->max_items)
454 			break;
455 
456 		atomic_dec(&pool->item_count);
457 
458 		if (++iter > 2) {
459 			if (pool->pool_type == RDS_IB_MR_8K_POOL)
460 				rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
461 			else
462 				rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
463 			break;
464 		}
465 
466 		/* We do have some empty MRs. Flush them out. */
467 		if (pool->pool_type == RDS_IB_MR_8K_POOL)
468 			rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
469 		else
470 			rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
471 
472 		rds_ib_flush_mr_pool(pool, 0, &ibmr);
473 		if (ibmr)
474 			return ibmr;
475 	}
476 
477 	return NULL;
478 }
479 
480 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
481 {
482 	struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
483 
484 	rds_ib_flush_mr_pool(pool, 0, NULL);
485 }
486 
487 void rds_ib_free_mr(void *trans_private, int invalidate)
488 {
489 	struct rds_ib_mr *ibmr = trans_private;
490 	struct rds_ib_mr_pool *pool = ibmr->pool;
491 	struct rds_ib_device *rds_ibdev = ibmr->device;
492 
493 	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
494 
495 	if (ibmr->odp) {
496 		/* A MR created and marked as use_once. We use delayed work,
497 		 * because there is a change that we are in interrupt and can't
498 		 * call to ib_dereg_mr() directly.
499 		 */
500 		INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
501 		queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
502 		return;
503 	}
504 
505 	/* Return it to the pool's free list */
506 	if (rds_ibdev->use_fastreg)
507 		rds_ib_free_frmr_list(ibmr);
508 	else
509 		rds_ib_free_fmr_list(ibmr);
510 
511 	atomic_add(ibmr->sg_len, &pool->free_pinned);
512 	atomic_inc(&pool->dirty_count);
513 
514 	/* If we've pinned too many pages, request a flush */
515 	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
516 	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
517 		queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
518 
519 	if (invalidate) {
520 		if (likely(!in_interrupt())) {
521 			rds_ib_flush_mr_pool(pool, 0, NULL);
522 		} else {
523 			/* We get here if the user created a MR marked
524 			 * as use_once and invalidate at the same time.
525 			 */
526 			queue_delayed_work(rds_ib_mr_wq,
527 					   &pool->flush_worker, 10);
528 		}
529 	}
530 
531 	rds_ib_dev_put(rds_ibdev);
532 }
533 
534 void rds_ib_flush_mrs(void)
535 {
536 	struct rds_ib_device *rds_ibdev;
537 
538 	down_read(&rds_ib_devices_lock);
539 	list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
540 		if (rds_ibdev->mr_8k_pool)
541 			rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
542 
543 		if (rds_ibdev->mr_1m_pool)
544 			rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
545 	}
546 	up_read(&rds_ib_devices_lock);
547 }
548 
549 u32 rds_ib_get_lkey(void *trans_private)
550 {
551 	struct rds_ib_mr *ibmr = trans_private;
552 
553 	return ibmr->u.mr->lkey;
554 }
555 
556 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
557 		    struct rds_sock *rs, u32 *key_ret,
558 		    struct rds_connection *conn,
559 		    u64 start, u64 length, int need_odp)
560 {
561 	struct rds_ib_device *rds_ibdev;
562 	struct rds_ib_mr *ibmr = NULL;
563 	struct rds_ib_connection *ic = NULL;
564 	int ret;
565 
566 	rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
567 	if (!rds_ibdev) {
568 		ret = -ENODEV;
569 		goto out;
570 	}
571 
572 	if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
573 		u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
574 		int access_flags =
575 			(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
576 			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
577 			 IB_ACCESS_ON_DEMAND);
578 		struct ib_sge sge = {};
579 		struct ib_mr *ib_mr;
580 
581 		if (!rds_ibdev->odp_capable) {
582 			ret = -EOPNOTSUPP;
583 			goto out;
584 		}
585 
586 		ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
587 				       access_flags);
588 
589 		if (IS_ERR(ib_mr)) {
590 			rdsdebug("rds_ib_get_user_mr returned %d\n",
591 				 IS_ERR(ib_mr));
592 			ret = PTR_ERR(ib_mr);
593 			goto out;
594 		}
595 		if (key_ret)
596 			*key_ret = ib_mr->rkey;
597 
598 		ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
599 		if (!ibmr) {
600 			ib_dereg_mr(ib_mr);
601 			ret = -ENOMEM;
602 			goto out;
603 		}
604 		ibmr->u.mr = ib_mr;
605 		ibmr->odp = 1;
606 
607 		sge.addr = virt_addr;
608 		sge.length = length;
609 		sge.lkey = ib_mr->lkey;
610 
611 		ib_advise_mr(rds_ibdev->pd,
612 			     IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
613 			     IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
614 		return ibmr;
615 	}
616 
617 	if (conn)
618 		ic = conn->c_transport_data;
619 
620 	if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
621 		ret = -ENODEV;
622 		goto out;
623 	}
624 
625 	if (rds_ibdev->use_fastreg)
626 		ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
627 	else
628 		ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
629 	if (IS_ERR(ibmr)) {
630 		ret = PTR_ERR(ibmr);
631 		pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
632 	} else {
633 		return ibmr;
634 	}
635 
636  out:
637 	if (rds_ibdev)
638 		rds_ib_dev_put(rds_ibdev);
639 
640 	return ERR_PTR(ret);
641 }
642 
643 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
644 {
645 	cancel_delayed_work_sync(&pool->flush_worker);
646 	rds_ib_flush_mr_pool(pool, 1, NULL);
647 	WARN_ON(atomic_read(&pool->item_count));
648 	WARN_ON(atomic_read(&pool->free_pinned));
649 	kfree(pool);
650 }
651 
652 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
653 					     int pool_type)
654 {
655 	struct rds_ib_mr_pool *pool;
656 
657 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
658 	if (!pool)
659 		return ERR_PTR(-ENOMEM);
660 
661 	pool->pool_type = pool_type;
662 	init_llist_head(&pool->free_list);
663 	init_llist_head(&pool->drop_list);
664 	init_llist_head(&pool->clean_list);
665 	spin_lock_init(&pool->clean_lock);
666 	mutex_init(&pool->flush_lock);
667 	init_waitqueue_head(&pool->flush_wait);
668 	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
669 
670 	if (pool_type == RDS_IB_MR_1M_POOL) {
671 		/* +1 allows for unaligned MRs */
672 		pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1;
673 		pool->max_items = rds_ibdev->max_1m_mrs;
674 	} else {
675 		/* pool_type == RDS_IB_MR_8K_POOL */
676 		pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1;
677 		pool->max_items = rds_ibdev->max_8k_mrs;
678 	}
679 
680 	pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
681 	pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
682 	pool->fmr_attr.page_shift = PAGE_SHIFT;
683 	pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
684 	pool->use_fastreg = rds_ibdev->use_fastreg;
685 
686 	return pool;
687 }
688 
689 int rds_ib_mr_init(void)
690 {
691 	rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
692 	if (!rds_ib_mr_wq)
693 		return -ENOMEM;
694 	return 0;
695 }
696 
697 /* By the time this is called all the IB devices should have been torn down and
698  * had their pools freed.  As each pool is freed its work struct is waited on,
699  * so the pool flushing work queue should be idle by the time we get here.
700  */
701 void rds_ib_mr_exit(void)
702 {
703 	destroy_workqueue(rds_ib_mr_wq);
704 }
705 
706 static void rds_ib_odp_mr_worker(struct work_struct  *work)
707 {
708 	struct rds_ib_mr *ibmr;
709 
710 	ibmr = container_of(work, struct rds_ib_mr, work.work);
711 	ib_dereg_mr(ibmr->u.mr);
712 	kfree(ibmr);
713 }
714