xref: /openbmc/linux/net/rds/ib_rdma.c (revision e2f1cf25)
1 /*
2  * Copyright (c) 2006 Oracle.  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.h"
39 #include "ib.h"
40 
41 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
42 #define CLEAN_LIST_BUSY_BIT 0
43 
44 /*
45  * This is stored as mr->r_trans_private.
46  */
47 struct rds_ib_mr {
48 	struct rds_ib_device	*device;
49 	struct rds_ib_mr_pool	*pool;
50 	struct ib_fmr		*fmr;
51 
52 	struct llist_node	llnode;
53 
54 	/* unmap_list is for freeing */
55 	struct list_head	unmap_list;
56 	unsigned int		remap_count;
57 
58 	struct scatterlist	*sg;
59 	unsigned int		sg_len;
60 	u64			*dma;
61 	int			sg_dma_len;
62 };
63 
64 /*
65  * Our own little FMR pool
66  */
67 struct rds_ib_mr_pool {
68 	struct mutex		flush_lock;		/* serialize fmr invalidate */
69 	struct delayed_work	flush_worker;		/* flush worker */
70 
71 	atomic_t		item_count;		/* total # of MRs */
72 	atomic_t		dirty_count;		/* # dirty of MRs */
73 
74 	struct llist_head	drop_list;		/* MRs that have reached their max_maps limit */
75 	struct llist_head	free_list;		/* unused MRs */
76 	struct llist_head	clean_list;		/* global unused & unamapped MRs */
77 	wait_queue_head_t	flush_wait;
78 
79 	atomic_t		free_pinned;		/* memory pinned by free MRs */
80 	unsigned long		max_items;
81 	unsigned long		max_items_soft;
82 	unsigned long		max_free_pinned;
83 	struct ib_fmr_attr	fmr_attr;
84 };
85 
86 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
87 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
88 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
89 
90 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
91 {
92 	struct rds_ib_device *rds_ibdev;
93 	struct rds_ib_ipaddr *i_ipaddr;
94 
95 	rcu_read_lock();
96 	list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
97 		list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
98 			if (i_ipaddr->ipaddr == ipaddr) {
99 				atomic_inc(&rds_ibdev->refcount);
100 				rcu_read_unlock();
101 				return rds_ibdev;
102 			}
103 		}
104 	}
105 	rcu_read_unlock();
106 
107 	return NULL;
108 }
109 
110 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
111 {
112 	struct rds_ib_ipaddr *i_ipaddr;
113 
114 	i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
115 	if (!i_ipaddr)
116 		return -ENOMEM;
117 
118 	i_ipaddr->ipaddr = ipaddr;
119 
120 	spin_lock_irq(&rds_ibdev->spinlock);
121 	list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
122 	spin_unlock_irq(&rds_ibdev->spinlock);
123 
124 	return 0;
125 }
126 
127 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
128 {
129 	struct rds_ib_ipaddr *i_ipaddr;
130 	struct rds_ib_ipaddr *to_free = NULL;
131 
132 
133 	spin_lock_irq(&rds_ibdev->spinlock);
134 	list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
135 		if (i_ipaddr->ipaddr == ipaddr) {
136 			list_del_rcu(&i_ipaddr->list);
137 			to_free = i_ipaddr;
138 			break;
139 		}
140 	}
141 	spin_unlock_irq(&rds_ibdev->spinlock);
142 
143 	if (to_free) {
144 		synchronize_rcu();
145 		kfree(to_free);
146 	}
147 }
148 
149 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
150 {
151 	struct rds_ib_device *rds_ibdev_old;
152 
153 	rds_ibdev_old = rds_ib_get_device(ipaddr);
154 	if (rds_ibdev_old) {
155 		rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
156 		rds_ib_dev_put(rds_ibdev_old);
157 	}
158 
159 	return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
160 }
161 
162 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
163 {
164 	struct rds_ib_connection *ic = conn->c_transport_data;
165 
166 	/* conn was previously on the nodev_conns_list */
167 	spin_lock_irq(&ib_nodev_conns_lock);
168 	BUG_ON(list_empty(&ib_nodev_conns));
169 	BUG_ON(list_empty(&ic->ib_node));
170 	list_del(&ic->ib_node);
171 
172 	spin_lock(&rds_ibdev->spinlock);
173 	list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
174 	spin_unlock(&rds_ibdev->spinlock);
175 	spin_unlock_irq(&ib_nodev_conns_lock);
176 
177 	ic->rds_ibdev = rds_ibdev;
178 	atomic_inc(&rds_ibdev->refcount);
179 }
180 
181 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
182 {
183 	struct rds_ib_connection *ic = conn->c_transport_data;
184 
185 	/* place conn on nodev_conns_list */
186 	spin_lock(&ib_nodev_conns_lock);
187 
188 	spin_lock_irq(&rds_ibdev->spinlock);
189 	BUG_ON(list_empty(&ic->ib_node));
190 	list_del(&ic->ib_node);
191 	spin_unlock_irq(&rds_ibdev->spinlock);
192 
193 	list_add_tail(&ic->ib_node, &ib_nodev_conns);
194 
195 	spin_unlock(&ib_nodev_conns_lock);
196 
197 	ic->rds_ibdev = NULL;
198 	rds_ib_dev_put(rds_ibdev);
199 }
200 
201 void rds_ib_destroy_nodev_conns(void)
202 {
203 	struct rds_ib_connection *ic, *_ic;
204 	LIST_HEAD(tmp_list);
205 
206 	/* avoid calling conn_destroy with irqs off */
207 	spin_lock_irq(&ib_nodev_conns_lock);
208 	list_splice(&ib_nodev_conns, &tmp_list);
209 	spin_unlock_irq(&ib_nodev_conns_lock);
210 
211 	list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
212 		rds_conn_destroy(ic->conn);
213 }
214 
215 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
216 {
217 	struct rds_ib_mr_pool *pool;
218 
219 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
220 	if (!pool)
221 		return ERR_PTR(-ENOMEM);
222 
223 	init_llist_head(&pool->free_list);
224 	init_llist_head(&pool->drop_list);
225 	init_llist_head(&pool->clean_list);
226 	mutex_init(&pool->flush_lock);
227 	init_waitqueue_head(&pool->flush_wait);
228 	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
229 
230 	pool->fmr_attr.max_pages = fmr_message_size;
231 	pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
232 	pool->fmr_attr.page_shift = PAGE_SHIFT;
233 	pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
234 
235 	/* We never allow more than max_items MRs to be allocated.
236 	 * When we exceed more than max_items_soft, we start freeing
237 	 * items more aggressively.
238 	 * Make sure that max_items > max_items_soft > max_items / 2
239 	 */
240 	pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
241 	pool->max_items = rds_ibdev->max_fmrs;
242 
243 	return pool;
244 }
245 
246 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
247 {
248 	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
249 
250 	iinfo->rdma_mr_max = pool->max_items;
251 	iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
252 }
253 
254 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
255 {
256 	cancel_delayed_work_sync(&pool->flush_worker);
257 	rds_ib_flush_mr_pool(pool, 1, NULL);
258 	WARN_ON(atomic_read(&pool->item_count));
259 	WARN_ON(atomic_read(&pool->free_pinned));
260 	kfree(pool);
261 }
262 
263 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
264 {
265 	struct rds_ib_mr *ibmr = NULL;
266 	struct llist_node *ret;
267 	unsigned long *flag;
268 
269 	preempt_disable();
270 	flag = this_cpu_ptr(&clean_list_grace);
271 	set_bit(CLEAN_LIST_BUSY_BIT, flag);
272 	ret = llist_del_first(&pool->clean_list);
273 	if (ret)
274 		ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
275 
276 	clear_bit(CLEAN_LIST_BUSY_BIT, flag);
277 	preempt_enable();
278 	return ibmr;
279 }
280 
281 static inline void wait_clean_list_grace(void)
282 {
283 	int cpu;
284 	unsigned long *flag;
285 
286 	for_each_online_cpu(cpu) {
287 		flag = &per_cpu(clean_list_grace, cpu);
288 		while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
289 			cpu_relax();
290 	}
291 }
292 
293 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
294 {
295 	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
296 	struct rds_ib_mr *ibmr = NULL;
297 	int err = 0, iter = 0;
298 
299 	if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
300 		schedule_delayed_work(&pool->flush_worker, 10);
301 
302 	while (1) {
303 		ibmr = rds_ib_reuse_fmr(pool);
304 		if (ibmr)
305 			return ibmr;
306 
307 		/* No clean MRs - now we have the choice of either
308 		 * allocating a fresh MR up to the limit imposed by the
309 		 * driver, or flush any dirty unused MRs.
310 		 * We try to avoid stalling in the send path if possible,
311 		 * so we allocate as long as we're allowed to.
312 		 *
313 		 * We're fussy with enforcing the FMR limit, though. If the driver
314 		 * tells us we can't use more than N fmrs, we shouldn't start
315 		 * arguing with it */
316 		if (atomic_inc_return(&pool->item_count) <= pool->max_items)
317 			break;
318 
319 		atomic_dec(&pool->item_count);
320 
321 		if (++iter > 2) {
322 			rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
323 			return ERR_PTR(-EAGAIN);
324 		}
325 
326 		/* We do have some empty MRs. Flush them out. */
327 		rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
328 		rds_ib_flush_mr_pool(pool, 0, &ibmr);
329 		if (ibmr)
330 			return ibmr;
331 	}
332 
333 	ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
334 	if (!ibmr) {
335 		err = -ENOMEM;
336 		goto out_no_cigar;
337 	}
338 
339 	memset(ibmr, 0, sizeof(*ibmr));
340 
341 	ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
342 			(IB_ACCESS_LOCAL_WRITE |
343 			 IB_ACCESS_REMOTE_READ |
344 			 IB_ACCESS_REMOTE_WRITE|
345 			 IB_ACCESS_REMOTE_ATOMIC),
346 			&pool->fmr_attr);
347 	if (IS_ERR(ibmr->fmr)) {
348 		err = PTR_ERR(ibmr->fmr);
349 		ibmr->fmr = NULL;
350 		printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
351 		goto out_no_cigar;
352 	}
353 
354 	rds_ib_stats_inc(s_ib_rdma_mr_alloc);
355 	return ibmr;
356 
357 out_no_cigar:
358 	if (ibmr) {
359 		if (ibmr->fmr)
360 			ib_dealloc_fmr(ibmr->fmr);
361 		kfree(ibmr);
362 	}
363 	atomic_dec(&pool->item_count);
364 	return ERR_PTR(err);
365 }
366 
367 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
368 	       struct scatterlist *sg, unsigned int nents)
369 {
370 	struct ib_device *dev = rds_ibdev->dev;
371 	struct scatterlist *scat = sg;
372 	u64 io_addr = 0;
373 	u64 *dma_pages;
374 	u32 len;
375 	int page_cnt, sg_dma_len;
376 	int i, j;
377 	int ret;
378 
379 	sg_dma_len = ib_dma_map_sg(dev, sg, nents,
380 				 DMA_BIDIRECTIONAL);
381 	if (unlikely(!sg_dma_len)) {
382 		printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
383 		return -EBUSY;
384 	}
385 
386 	len = 0;
387 	page_cnt = 0;
388 
389 	for (i = 0; i < sg_dma_len; ++i) {
390 		unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
391 		u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
392 
393 		if (dma_addr & ~PAGE_MASK) {
394 			if (i > 0)
395 				return -EINVAL;
396 			else
397 				++page_cnt;
398 		}
399 		if ((dma_addr + dma_len) & ~PAGE_MASK) {
400 			if (i < sg_dma_len - 1)
401 				return -EINVAL;
402 			else
403 				++page_cnt;
404 		}
405 
406 		len += dma_len;
407 	}
408 
409 	page_cnt += len >> PAGE_SHIFT;
410 	if (page_cnt > fmr_message_size)
411 		return -EINVAL;
412 
413 	dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
414 				 rdsibdev_to_node(rds_ibdev));
415 	if (!dma_pages)
416 		return -ENOMEM;
417 
418 	page_cnt = 0;
419 	for (i = 0; i < sg_dma_len; ++i) {
420 		unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
421 		u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
422 
423 		for (j = 0; j < dma_len; j += PAGE_SIZE)
424 			dma_pages[page_cnt++] =
425 				(dma_addr & PAGE_MASK) + j;
426 	}
427 
428 	ret = ib_map_phys_fmr(ibmr->fmr,
429 				   dma_pages, page_cnt, io_addr);
430 	if (ret)
431 		goto out;
432 
433 	/* Success - we successfully remapped the MR, so we can
434 	 * safely tear down the old mapping. */
435 	rds_ib_teardown_mr(ibmr);
436 
437 	ibmr->sg = scat;
438 	ibmr->sg_len = nents;
439 	ibmr->sg_dma_len = sg_dma_len;
440 	ibmr->remap_count++;
441 
442 	rds_ib_stats_inc(s_ib_rdma_mr_used);
443 	ret = 0;
444 
445 out:
446 	kfree(dma_pages);
447 
448 	return ret;
449 }
450 
451 void rds_ib_sync_mr(void *trans_private, int direction)
452 {
453 	struct rds_ib_mr *ibmr = trans_private;
454 	struct rds_ib_device *rds_ibdev = ibmr->device;
455 
456 	switch (direction) {
457 	case DMA_FROM_DEVICE:
458 		ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
459 			ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
460 		break;
461 	case DMA_TO_DEVICE:
462 		ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
463 			ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
464 		break;
465 	}
466 }
467 
468 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
469 {
470 	struct rds_ib_device *rds_ibdev = ibmr->device;
471 
472 	if (ibmr->sg_dma_len) {
473 		ib_dma_unmap_sg(rds_ibdev->dev,
474 				ibmr->sg, ibmr->sg_len,
475 				DMA_BIDIRECTIONAL);
476 		ibmr->sg_dma_len = 0;
477 	}
478 
479 	/* Release the s/g list */
480 	if (ibmr->sg_len) {
481 		unsigned int i;
482 
483 		for (i = 0; i < ibmr->sg_len; ++i) {
484 			struct page *page = sg_page(&ibmr->sg[i]);
485 
486 			/* FIXME we need a way to tell a r/w MR
487 			 * from a r/o MR */
488 			BUG_ON(irqs_disabled());
489 			set_page_dirty(page);
490 			put_page(page);
491 		}
492 		kfree(ibmr->sg);
493 
494 		ibmr->sg = NULL;
495 		ibmr->sg_len = 0;
496 	}
497 }
498 
499 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
500 {
501 	unsigned int pinned = ibmr->sg_len;
502 
503 	__rds_ib_teardown_mr(ibmr);
504 	if (pinned) {
505 		struct rds_ib_device *rds_ibdev = ibmr->device;
506 		struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
507 
508 		atomic_sub(pinned, &pool->free_pinned);
509 	}
510 }
511 
512 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
513 {
514 	unsigned int item_count;
515 
516 	item_count = atomic_read(&pool->item_count);
517 	if (free_all)
518 		return item_count;
519 
520 	return 0;
521 }
522 
523 /*
524  * given an llist of mrs, put them all into the list_head for more processing
525  */
526 static void llist_append_to_list(struct llist_head *llist, struct list_head *list)
527 {
528 	struct rds_ib_mr *ibmr;
529 	struct llist_node *node;
530 	struct llist_node *next;
531 
532 	node = llist_del_all(llist);
533 	while (node) {
534 		next = node->next;
535 		ibmr = llist_entry(node, struct rds_ib_mr, llnode);
536 		list_add_tail(&ibmr->unmap_list, list);
537 		node = next;
538 	}
539 }
540 
541 /*
542  * this takes a list head of mrs and turns it into linked llist nodes
543  * of clusters.  Each cluster has linked llist nodes of
544  * MR_CLUSTER_SIZE mrs that are ready for reuse.
545  */
546 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
547 				struct list_head *list,
548 				struct llist_node **nodes_head,
549 				struct llist_node **nodes_tail)
550 {
551 	struct rds_ib_mr *ibmr;
552 	struct llist_node *cur = NULL;
553 	struct llist_node **next = nodes_head;
554 
555 	list_for_each_entry(ibmr, list, unmap_list) {
556 		cur = &ibmr->llnode;
557 		*next = cur;
558 		next = &cur->next;
559 	}
560 	*next = NULL;
561 	*nodes_tail = cur;
562 }
563 
564 /*
565  * Flush our pool of MRs.
566  * At a minimum, all currently unused MRs are unmapped.
567  * If the number of MRs allocated exceeds the limit, we also try
568  * to free as many MRs as needed to get back to this limit.
569  */
570 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
571 			        int free_all, struct rds_ib_mr **ibmr_ret)
572 {
573 	struct rds_ib_mr *ibmr, *next;
574 	struct llist_node *clean_nodes;
575 	struct llist_node *clean_tail;
576 	LIST_HEAD(unmap_list);
577 	LIST_HEAD(fmr_list);
578 	unsigned long unpinned = 0;
579 	unsigned int nfreed = 0, ncleaned = 0, free_goal;
580 	int ret = 0;
581 
582 	rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
583 
584 	if (ibmr_ret) {
585 		DEFINE_WAIT(wait);
586 		while(!mutex_trylock(&pool->flush_lock)) {
587 			ibmr = rds_ib_reuse_fmr(pool);
588 			if (ibmr) {
589 				*ibmr_ret = ibmr;
590 				finish_wait(&pool->flush_wait, &wait);
591 				goto out_nolock;
592 			}
593 
594 			prepare_to_wait(&pool->flush_wait, &wait,
595 					TASK_UNINTERRUPTIBLE);
596 			if (llist_empty(&pool->clean_list))
597 				schedule();
598 
599 			ibmr = rds_ib_reuse_fmr(pool);
600 			if (ibmr) {
601 				*ibmr_ret = ibmr;
602 				finish_wait(&pool->flush_wait, &wait);
603 				goto out_nolock;
604 			}
605 		}
606 		finish_wait(&pool->flush_wait, &wait);
607 	} else
608 		mutex_lock(&pool->flush_lock);
609 
610 	if (ibmr_ret) {
611 		ibmr = rds_ib_reuse_fmr(pool);
612 		if (ibmr) {
613 			*ibmr_ret = ibmr;
614 			goto out;
615 		}
616 	}
617 
618 	/* Get the list of all MRs to be dropped. Ordering matters -
619 	 * we want to put drop_list ahead of free_list.
620 	 */
621 	llist_append_to_list(&pool->drop_list, &unmap_list);
622 	llist_append_to_list(&pool->free_list, &unmap_list);
623 	if (free_all)
624 		llist_append_to_list(&pool->clean_list, &unmap_list);
625 
626 	free_goal = rds_ib_flush_goal(pool, free_all);
627 
628 	if (list_empty(&unmap_list))
629 		goto out;
630 
631 	/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
632 	list_for_each_entry(ibmr, &unmap_list, unmap_list)
633 		list_add(&ibmr->fmr->list, &fmr_list);
634 
635 	ret = ib_unmap_fmr(&fmr_list);
636 	if (ret)
637 		printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
638 
639 	/* Now we can destroy the DMA mapping and unpin any pages */
640 	list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
641 		unpinned += ibmr->sg_len;
642 		__rds_ib_teardown_mr(ibmr);
643 		if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
644 			rds_ib_stats_inc(s_ib_rdma_mr_free);
645 			list_del(&ibmr->unmap_list);
646 			ib_dealloc_fmr(ibmr->fmr);
647 			kfree(ibmr);
648 			nfreed++;
649 		}
650 		ncleaned++;
651 	}
652 
653 	if (!list_empty(&unmap_list)) {
654 		/* we have to make sure that none of the things we're about
655 		 * to put on the clean list would race with other cpus trying
656 		 * to pull items off.  The llist would explode if we managed to
657 		 * remove something from the clean list and then add it back again
658 		 * while another CPU was spinning on that same item in llist_del_first.
659 		 *
660 		 * This is pretty unlikely, but just in case  wait for an llist grace period
661 		 * here before adding anything back into the clean list.
662 		 */
663 		wait_clean_list_grace();
664 
665 		list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
666 		if (ibmr_ret)
667 			*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
668 
669 		/* more than one entry in llist nodes */
670 		if (clean_nodes->next)
671 			llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
672 
673 	}
674 
675 	atomic_sub(unpinned, &pool->free_pinned);
676 	atomic_sub(ncleaned, &pool->dirty_count);
677 	atomic_sub(nfreed, &pool->item_count);
678 
679 out:
680 	mutex_unlock(&pool->flush_lock);
681 	if (waitqueue_active(&pool->flush_wait))
682 		wake_up(&pool->flush_wait);
683 out_nolock:
684 	return ret;
685 }
686 
687 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
688 {
689 	struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
690 
691 	rds_ib_flush_mr_pool(pool, 0, NULL);
692 }
693 
694 void rds_ib_free_mr(void *trans_private, int invalidate)
695 {
696 	struct rds_ib_mr *ibmr = trans_private;
697 	struct rds_ib_device *rds_ibdev = ibmr->device;
698 	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
699 
700 	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
701 
702 	/* Return it to the pool's free list */
703 	if (ibmr->remap_count >= pool->fmr_attr.max_maps)
704 		llist_add(&ibmr->llnode, &pool->drop_list);
705 	else
706 		llist_add(&ibmr->llnode, &pool->free_list);
707 
708 	atomic_add(ibmr->sg_len, &pool->free_pinned);
709 	atomic_inc(&pool->dirty_count);
710 
711 	/* If we've pinned too many pages, request a flush */
712 	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
713 	    atomic_read(&pool->dirty_count) >= pool->max_items / 10)
714 		schedule_delayed_work(&pool->flush_worker, 10);
715 
716 	if (invalidate) {
717 		if (likely(!in_interrupt())) {
718 			rds_ib_flush_mr_pool(pool, 0, NULL);
719 		} else {
720 			/* We get here if the user created a MR marked
721 			 * as use_once and invalidate at the same time. */
722 			schedule_delayed_work(&pool->flush_worker, 10);
723 		}
724 	}
725 
726 	rds_ib_dev_put(rds_ibdev);
727 }
728 
729 void rds_ib_flush_mrs(void)
730 {
731 	struct rds_ib_device *rds_ibdev;
732 
733 	down_read(&rds_ib_devices_lock);
734 	list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
735 		struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
736 
737 		if (pool)
738 			rds_ib_flush_mr_pool(pool, 0, NULL);
739 	}
740 	up_read(&rds_ib_devices_lock);
741 }
742 
743 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
744 		    struct rds_sock *rs, u32 *key_ret)
745 {
746 	struct rds_ib_device *rds_ibdev;
747 	struct rds_ib_mr *ibmr = NULL;
748 	int ret;
749 
750 	rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
751 	if (!rds_ibdev) {
752 		ret = -ENODEV;
753 		goto out;
754 	}
755 
756 	if (!rds_ibdev->mr_pool) {
757 		ret = -ENODEV;
758 		goto out;
759 	}
760 
761 	ibmr = rds_ib_alloc_fmr(rds_ibdev);
762 	if (IS_ERR(ibmr)) {
763 		rds_ib_dev_put(rds_ibdev);
764 		return ibmr;
765 	}
766 
767 	ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
768 	if (ret == 0)
769 		*key_ret = ibmr->fmr->rkey;
770 	else
771 		printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
772 
773 	ibmr->device = rds_ibdev;
774 	rds_ibdev = NULL;
775 
776  out:
777 	if (ret) {
778 		if (ibmr)
779 			rds_ib_free_mr(ibmr, 0);
780 		ibmr = ERR_PTR(ret);
781 	}
782 	if (rds_ibdev)
783 		rds_ib_dev_put(rds_ibdev);
784 	return ibmr;
785 }
786 
787