2 * Copyright (c) 2006 Oracle. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
35 #include <linux/rculist.h>
36 #include <linux/llist.h>
41 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
42 #define CLEAN_LIST_BUSY_BIT 0
45 * This is stored as mr->r_trans_private.
48 struct rds_ib_device *device;
49 struct rds_ib_mr_pool *pool;
52 struct llist_node llnode;
54 /* unmap_list is for freeing */
55 struct list_head unmap_list;
56 unsigned int remap_count;
58 struct scatterlist *sg;
65 * Our own little FMR pool
67 struct rds_ib_mr_pool {
68 struct mutex flush_lock; /* serialize fmr invalidate */
69 struct delayed_work flush_worker; /* flush worker */
71 atomic_t item_count; /* total # of MRs */
72 atomic_t dirty_count; /* # dirty of MRs */
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;
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;
86 struct workqueue_struct *rds_ib_fmr_wq;
88 int rds_ib_fmr_init(void)
90 rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
96 /* By the time this is called all the IB devices should have been torn down and
97 * had their pools freed. As each pool is freed its work struct is waited on,
98 * so the pool flushing work queue should be idle by the time we get here.
100 void rds_ib_fmr_exit(void)
102 destroy_workqueue(rds_ib_fmr_wq);
105 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
106 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
107 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
109 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
111 struct rds_ib_device *rds_ibdev;
112 struct rds_ib_ipaddr *i_ipaddr;
115 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
116 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
117 if (i_ipaddr->ipaddr == ipaddr) {
118 atomic_inc(&rds_ibdev->refcount);
129 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
131 struct rds_ib_ipaddr *i_ipaddr;
133 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
137 i_ipaddr->ipaddr = ipaddr;
139 spin_lock_irq(&rds_ibdev->spinlock);
140 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
141 spin_unlock_irq(&rds_ibdev->spinlock);
146 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
148 struct rds_ib_ipaddr *i_ipaddr;
149 struct rds_ib_ipaddr *to_free = NULL;
152 spin_lock_irq(&rds_ibdev->spinlock);
153 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
154 if (i_ipaddr->ipaddr == ipaddr) {
155 list_del_rcu(&i_ipaddr->list);
160 spin_unlock_irq(&rds_ibdev->spinlock);
163 kfree_rcu(to_free, rcu);
166 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
168 struct rds_ib_device *rds_ibdev_old;
170 rds_ibdev_old = rds_ib_get_device(ipaddr);
172 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
174 if (rds_ibdev_old != rds_ibdev) {
175 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
176 rds_ib_dev_put(rds_ibdev_old);
177 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
179 rds_ib_dev_put(rds_ibdev_old);
184 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
186 struct rds_ib_connection *ic = conn->c_transport_data;
188 /* conn was previously on the nodev_conns_list */
189 spin_lock_irq(&ib_nodev_conns_lock);
190 BUG_ON(list_empty(&ib_nodev_conns));
191 BUG_ON(list_empty(&ic->ib_node));
192 list_del(&ic->ib_node);
194 spin_lock(&rds_ibdev->spinlock);
195 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
196 spin_unlock(&rds_ibdev->spinlock);
197 spin_unlock_irq(&ib_nodev_conns_lock);
199 ic->rds_ibdev = rds_ibdev;
200 atomic_inc(&rds_ibdev->refcount);
203 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
205 struct rds_ib_connection *ic = conn->c_transport_data;
207 /* place conn on nodev_conns_list */
208 spin_lock(&ib_nodev_conns_lock);
210 spin_lock_irq(&rds_ibdev->spinlock);
211 BUG_ON(list_empty(&ic->ib_node));
212 list_del(&ic->ib_node);
213 spin_unlock_irq(&rds_ibdev->spinlock);
215 list_add_tail(&ic->ib_node, &ib_nodev_conns);
217 spin_unlock(&ib_nodev_conns_lock);
219 ic->rds_ibdev = NULL;
220 rds_ib_dev_put(rds_ibdev);
223 void rds_ib_destroy_nodev_conns(void)
225 struct rds_ib_connection *ic, *_ic;
228 /* avoid calling conn_destroy with irqs off */
229 spin_lock_irq(&ib_nodev_conns_lock);
230 list_splice(&ib_nodev_conns, &tmp_list);
231 spin_unlock_irq(&ib_nodev_conns_lock);
233 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
234 rds_conn_destroy(ic->conn);
237 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
239 struct rds_ib_mr_pool *pool;
241 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
243 return ERR_PTR(-ENOMEM);
245 init_llist_head(&pool->free_list);
246 init_llist_head(&pool->drop_list);
247 init_llist_head(&pool->clean_list);
248 mutex_init(&pool->flush_lock);
249 init_waitqueue_head(&pool->flush_wait);
250 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
252 pool->fmr_attr.max_pages = fmr_message_size;
253 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
254 pool->fmr_attr.page_shift = PAGE_SHIFT;
255 pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
257 /* We never allow more than max_items MRs to be allocated.
258 * When we exceed more than max_items_soft, we start freeing
259 * items more aggressively.
260 * Make sure that max_items > max_items_soft > max_items / 2
262 pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
263 pool->max_items = rds_ibdev->max_fmrs;
268 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
270 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
272 iinfo->rdma_mr_max = pool->max_items;
273 iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
276 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
278 cancel_delayed_work_sync(&pool->flush_worker);
279 rds_ib_flush_mr_pool(pool, 1, NULL);
280 WARN_ON(atomic_read(&pool->item_count));
281 WARN_ON(atomic_read(&pool->free_pinned));
285 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
287 struct rds_ib_mr *ibmr = NULL;
288 struct llist_node *ret;
292 flag = this_cpu_ptr(&clean_list_grace);
293 set_bit(CLEAN_LIST_BUSY_BIT, flag);
294 ret = llist_del_first(&pool->clean_list);
296 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
298 clear_bit(CLEAN_LIST_BUSY_BIT, flag);
303 static inline void wait_clean_list_grace(void)
308 for_each_online_cpu(cpu) {
309 flag = &per_cpu(clean_list_grace, cpu);
310 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
315 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
317 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
318 struct rds_ib_mr *ibmr = NULL;
319 int err = 0, iter = 0;
321 if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
322 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
325 ibmr = rds_ib_reuse_fmr(pool);
329 /* No clean MRs - now we have the choice of either
330 * allocating a fresh MR up to the limit imposed by the
331 * driver, or flush any dirty unused MRs.
332 * We try to avoid stalling in the send path if possible,
333 * so we allocate as long as we're allowed to.
335 * We're fussy with enforcing the FMR limit, though. If the driver
336 * tells us we can't use more than N fmrs, we shouldn't start
338 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
341 atomic_dec(&pool->item_count);
344 rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
345 return ERR_PTR(-EAGAIN);
348 /* We do have some empty MRs. Flush them out. */
349 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
350 rds_ib_flush_mr_pool(pool, 0, &ibmr);
355 ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
361 ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
362 (IB_ACCESS_LOCAL_WRITE |
363 IB_ACCESS_REMOTE_READ |
364 IB_ACCESS_REMOTE_WRITE|
365 IB_ACCESS_REMOTE_ATOMIC),
367 if (IS_ERR(ibmr->fmr)) {
368 err = PTR_ERR(ibmr->fmr);
370 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
374 rds_ib_stats_inc(s_ib_rdma_mr_alloc);
380 ib_dealloc_fmr(ibmr->fmr);
383 atomic_dec(&pool->item_count);
387 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
388 struct scatterlist *sg, unsigned int nents)
390 struct ib_device *dev = rds_ibdev->dev;
391 struct scatterlist *scat = sg;
395 int page_cnt, sg_dma_len;
399 sg_dma_len = ib_dma_map_sg(dev, sg, nents,
401 if (unlikely(!sg_dma_len)) {
402 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
409 for (i = 0; i < sg_dma_len; ++i) {
410 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
411 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
413 if (dma_addr & ~PAGE_MASK) {
419 if ((dma_addr + dma_len) & ~PAGE_MASK) {
420 if (i < sg_dma_len - 1)
429 page_cnt += len >> PAGE_SHIFT;
430 if (page_cnt > fmr_message_size)
433 dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
434 rdsibdev_to_node(rds_ibdev));
439 for (i = 0; i < sg_dma_len; ++i) {
440 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
441 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
443 for (j = 0; j < dma_len; j += PAGE_SIZE)
444 dma_pages[page_cnt++] =
445 (dma_addr & PAGE_MASK) + j;
448 ret = ib_map_phys_fmr(ibmr->fmr,
449 dma_pages, page_cnt, io_addr);
453 /* Success - we successfully remapped the MR, so we can
454 * safely tear down the old mapping. */
455 rds_ib_teardown_mr(ibmr);
458 ibmr->sg_len = nents;
459 ibmr->sg_dma_len = sg_dma_len;
462 rds_ib_stats_inc(s_ib_rdma_mr_used);
471 void rds_ib_sync_mr(void *trans_private, int direction)
473 struct rds_ib_mr *ibmr = trans_private;
474 struct rds_ib_device *rds_ibdev = ibmr->device;
477 case DMA_FROM_DEVICE:
478 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
479 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
482 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
483 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
488 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
490 struct rds_ib_device *rds_ibdev = ibmr->device;
492 if (ibmr->sg_dma_len) {
493 ib_dma_unmap_sg(rds_ibdev->dev,
494 ibmr->sg, ibmr->sg_len,
496 ibmr->sg_dma_len = 0;
499 /* Release the s/g list */
503 for (i = 0; i < ibmr->sg_len; ++i) {
504 struct page *page = sg_page(&ibmr->sg[i]);
506 /* FIXME we need a way to tell a r/w MR
508 WARN_ON(!page->mapping && irqs_disabled());
509 set_page_dirty(page);
519 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
521 unsigned int pinned = ibmr->sg_len;
523 __rds_ib_teardown_mr(ibmr);
525 struct rds_ib_mr_pool *pool = ibmr->pool;
527 atomic_sub(pinned, &pool->free_pinned);
531 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
533 unsigned int item_count;
535 item_count = atomic_read(&pool->item_count);
543 * given an llist of mrs, put them all into the list_head for more processing
545 static unsigned int llist_append_to_list(struct llist_head *llist,
546 struct list_head *list)
548 struct rds_ib_mr *ibmr;
549 struct llist_node *node;
550 struct llist_node *next;
551 unsigned int count = 0;
553 node = llist_del_all(llist);
556 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
557 list_add_tail(&ibmr->unmap_list, list);
565 * this takes a list head of mrs and turns it into linked llist nodes
566 * of clusters. Each cluster has linked llist nodes of
567 * MR_CLUSTER_SIZE mrs that are ready for reuse.
569 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
570 struct list_head *list,
571 struct llist_node **nodes_head,
572 struct llist_node **nodes_tail)
574 struct rds_ib_mr *ibmr;
575 struct llist_node *cur = NULL;
576 struct llist_node **next = nodes_head;
578 list_for_each_entry(ibmr, list, unmap_list) {
588 * Flush our pool of MRs.
589 * At a minimum, all currently unused MRs are unmapped.
590 * If the number of MRs allocated exceeds the limit, we also try
591 * to free as many MRs as needed to get back to this limit.
593 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
594 int free_all, struct rds_ib_mr **ibmr_ret)
596 struct rds_ib_mr *ibmr, *next;
597 struct llist_node *clean_nodes;
598 struct llist_node *clean_tail;
599 LIST_HEAD(unmap_list);
601 unsigned long unpinned = 0;
602 unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
605 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
609 while(!mutex_trylock(&pool->flush_lock)) {
610 ibmr = rds_ib_reuse_fmr(pool);
613 finish_wait(&pool->flush_wait, &wait);
617 prepare_to_wait(&pool->flush_wait, &wait,
618 TASK_UNINTERRUPTIBLE);
619 if (llist_empty(&pool->clean_list))
622 ibmr = rds_ib_reuse_fmr(pool);
625 finish_wait(&pool->flush_wait, &wait);
629 finish_wait(&pool->flush_wait, &wait);
631 mutex_lock(&pool->flush_lock);
634 ibmr = rds_ib_reuse_fmr(pool);
641 /* Get the list of all MRs to be dropped. Ordering matters -
642 * we want to put drop_list ahead of free_list.
644 dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
645 dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
647 llist_append_to_list(&pool->clean_list, &unmap_list);
649 free_goal = rds_ib_flush_goal(pool, free_all);
651 if (list_empty(&unmap_list))
654 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
655 list_for_each_entry(ibmr, &unmap_list, unmap_list)
656 list_add(&ibmr->fmr->list, &fmr_list);
658 ret = ib_unmap_fmr(&fmr_list);
660 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
662 /* Now we can destroy the DMA mapping and unpin any pages */
663 list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
664 unpinned += ibmr->sg_len;
665 __rds_ib_teardown_mr(ibmr);
666 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
667 rds_ib_stats_inc(s_ib_rdma_mr_free);
668 list_del(&ibmr->unmap_list);
669 ib_dealloc_fmr(ibmr->fmr);
675 if (!list_empty(&unmap_list)) {
676 /* we have to make sure that none of the things we're about
677 * to put on the clean list would race with other cpus trying
678 * to pull items off. The llist would explode if we managed to
679 * remove something from the clean list and then add it back again
680 * while another CPU was spinning on that same item in llist_del_first.
682 * This is pretty unlikely, but just in case wait for an llist grace period
683 * here before adding anything back into the clean list.
685 wait_clean_list_grace();
687 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
689 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
691 /* more than one entry in llist nodes */
692 if (clean_nodes->next)
693 llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
697 atomic_sub(unpinned, &pool->free_pinned);
698 atomic_sub(dirty_to_clean, &pool->dirty_count);
699 atomic_sub(nfreed, &pool->item_count);
702 mutex_unlock(&pool->flush_lock);
703 if (waitqueue_active(&pool->flush_wait))
704 wake_up(&pool->flush_wait);
709 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
711 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
713 rds_ib_flush_mr_pool(pool, 0, NULL);
716 void rds_ib_free_mr(void *trans_private, int invalidate)
718 struct rds_ib_mr *ibmr = trans_private;
719 struct rds_ib_mr_pool *pool = ibmr->pool;
720 struct rds_ib_device *rds_ibdev = ibmr->device;
722 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
724 /* Return it to the pool's free list */
725 if (ibmr->remap_count >= pool->fmr_attr.max_maps)
726 llist_add(&ibmr->llnode, &pool->drop_list);
728 llist_add(&ibmr->llnode, &pool->free_list);
730 atomic_add(ibmr->sg_len, &pool->free_pinned);
731 atomic_inc(&pool->dirty_count);
733 /* If we've pinned too many pages, request a flush */
734 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
735 atomic_read(&pool->dirty_count) >= pool->max_items / 5)
736 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
739 if (likely(!in_interrupt())) {
740 rds_ib_flush_mr_pool(pool, 0, NULL);
742 /* We get here if the user created a MR marked
743 * as use_once and invalidate at the same time.
745 queue_delayed_work(rds_ib_fmr_wq,
746 &pool->flush_worker, 10);
750 rds_ib_dev_put(rds_ibdev);
753 void rds_ib_flush_mrs(void)
755 struct rds_ib_device *rds_ibdev;
757 down_read(&rds_ib_devices_lock);
758 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
759 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
762 rds_ib_flush_mr_pool(pool, 0, NULL);
764 up_read(&rds_ib_devices_lock);
767 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
768 struct rds_sock *rs, u32 *key_ret)
770 struct rds_ib_device *rds_ibdev;
771 struct rds_ib_mr *ibmr = NULL;
774 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
780 if (!rds_ibdev->mr_pool) {
785 ibmr = rds_ib_alloc_fmr(rds_ibdev);
787 rds_ib_dev_put(rds_ibdev);
791 ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
793 *key_ret = ibmr->fmr->rkey;
795 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
797 ibmr->device = rds_ibdev;
803 rds_ib_free_mr(ibmr, 0);
807 rds_ib_dev_put(rds_ibdev);