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
41static 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 */
47struct 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 */
67struct 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
86static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
87static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
88static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
89
90static 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
110static 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
127static 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
149int 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
162void 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
181void 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
201void 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
215struct 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
246void 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
254void 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
263static 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 = &__get_cpu_var(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
281static 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
293static 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
357out_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
367static 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
445out:
446	kfree(dma_pages);
447
448	return ret;
449}
450
451void 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
468static 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
499static 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
512static 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 */
526static 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 */
546static 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 */
570static 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
679out:
680	mutex_unlock(&pool->flush_lock);
681	if (waitqueue_active(&pool->flush_wait))
682		wake_up(&pool->flush_wait);
683out_nolock:
684	return ret;
685}
686
687static 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
694void 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
729void 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
743void *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		return ibmr;
764
765	ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
766	if (ret == 0)
767		*key_ret = ibmr->fmr->rkey;
768	else
769		printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
770
771	ibmr->device = rds_ibdev;
772	rds_ibdev = NULL;
773
774 out:
775	if (ret) {
776		if (ibmr)
777			rds_ib_free_mr(ibmr, 0);
778		ibmr = ERR_PTR(ret);
779	}
780	if (rds_ibdev)
781		rds_ib_dev_put(rds_ibdev);
782	return ibmr;
783}
784
785