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>
37 #include <linux/delay.h>
42 static DEFINE_PER_CPU(unsigned long, clean_list_grace);
43 #define CLEAN_LIST_BUSY_BIT 0
46 * This is stored as mr->r_trans_private.
49 struct rds_ib_device *device;
50 struct rds_ib_mr_pool *pool;
53 struct llist_node llnode;
55 /* unmap_list is for freeing */
56 struct list_head unmap_list;
57 unsigned int remap_count;
59 struct scatterlist *sg;
66 * Our own little FMR pool
68 struct rds_ib_mr_pool {
69 struct mutex flush_lock; /* serialize fmr invalidate */
70 struct delayed_work flush_worker; /* flush worker */
72 atomic_t item_count; /* total # of MRs */
73 atomic_t dirty_count; /* # dirty of MRs */
75 struct llist_head drop_list; /* MRs that have reached their max_maps limit */
76 struct llist_head free_list; /* unused MRs */
77 struct llist_head clean_list; /* global unused & unamapped MRs */
78 wait_queue_head_t flush_wait;
80 atomic_t free_pinned; /* memory pinned by free MRs */
81 unsigned long max_items;
82 unsigned long max_items_soft;
83 unsigned long max_free_pinned;
84 struct ib_fmr_attr fmr_attr;
87 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
88 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
89 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
91 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
93 struct rds_ib_device *rds_ibdev;
94 struct rds_ib_ipaddr *i_ipaddr;
97 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
98 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
99 if (i_ipaddr->ipaddr == ipaddr) {
100 atomic_inc(&rds_ibdev->refcount);
111 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
113 struct rds_ib_ipaddr *i_ipaddr;
115 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
119 i_ipaddr->ipaddr = ipaddr;
121 spin_lock_irq(&rds_ibdev->spinlock);
122 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
123 spin_unlock_irq(&rds_ibdev->spinlock);
128 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
130 struct rds_ib_ipaddr *i_ipaddr;
131 struct rds_ib_ipaddr *to_free = NULL;
134 spin_lock_irq(&rds_ibdev->spinlock);
135 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
136 if (i_ipaddr->ipaddr == ipaddr) {
137 list_del_rcu(&i_ipaddr->list);
142 spin_unlock_irq(&rds_ibdev->spinlock);
150 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
152 struct rds_ib_device *rds_ibdev_old;
154 rds_ibdev_old = rds_ib_get_device(ipaddr);
156 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
157 rds_ib_dev_put(rds_ibdev_old);
160 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
163 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
165 struct rds_ib_connection *ic = conn->c_transport_data;
167 /* conn was previously on the nodev_conns_list */
168 spin_lock_irq(&ib_nodev_conns_lock);
169 BUG_ON(list_empty(&ib_nodev_conns));
170 BUG_ON(list_empty(&ic->ib_node));
171 list_del(&ic->ib_node);
173 spin_lock(&rds_ibdev->spinlock);
174 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
175 spin_unlock(&rds_ibdev->spinlock);
176 spin_unlock_irq(&ib_nodev_conns_lock);
178 ic->rds_ibdev = rds_ibdev;
179 atomic_inc(&rds_ibdev->refcount);
182 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
184 struct rds_ib_connection *ic = conn->c_transport_data;
186 /* place conn on nodev_conns_list */
187 spin_lock(&ib_nodev_conns_lock);
189 spin_lock_irq(&rds_ibdev->spinlock);
190 BUG_ON(list_empty(&ic->ib_node));
191 list_del(&ic->ib_node);
192 spin_unlock_irq(&rds_ibdev->spinlock);
194 list_add_tail(&ic->ib_node, &ib_nodev_conns);
196 spin_unlock(&ib_nodev_conns_lock);
198 ic->rds_ibdev = NULL;
199 rds_ib_dev_put(rds_ibdev);
202 void rds_ib_destroy_nodev_conns(void)
204 struct rds_ib_connection *ic, *_ic;
207 /* avoid calling conn_destroy with irqs off */
208 spin_lock_irq(&ib_nodev_conns_lock);
209 list_splice(&ib_nodev_conns, &tmp_list);
210 spin_unlock_irq(&ib_nodev_conns_lock);
212 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
213 rds_conn_destroy(ic->conn);
216 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
218 struct rds_ib_mr_pool *pool;
220 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
222 return ERR_PTR(-ENOMEM);
224 init_llist_head(&pool->free_list);
225 init_llist_head(&pool->drop_list);
226 init_llist_head(&pool->clean_list);
227 mutex_init(&pool->flush_lock);
228 init_waitqueue_head(&pool->flush_wait);
229 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
231 pool->fmr_attr.max_pages = fmr_message_size;
232 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
233 pool->fmr_attr.page_shift = PAGE_SHIFT;
234 pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
236 /* We never allow more than max_items MRs to be allocated.
237 * When we exceed more than max_items_soft, we start freeing
238 * items more aggressively.
239 * Make sure that max_items > max_items_soft > max_items / 2
241 pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
242 pool->max_items = rds_ibdev->max_fmrs;
247 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
249 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
251 iinfo->rdma_mr_max = pool->max_items;
252 iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
255 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
257 cancel_delayed_work_sync(&pool->flush_worker);
258 rds_ib_flush_mr_pool(pool, 1, NULL);
259 WARN_ON(atomic_read(&pool->item_count));
260 WARN_ON(atomic_read(&pool->free_pinned));
264 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
266 struct rds_ib_mr *ibmr = NULL;
267 struct llist_node *ret;
271 flag = this_cpu_ptr(&clean_list_grace);
272 set_bit(CLEAN_LIST_BUSY_BIT, flag);
273 ret = llist_del_first(&pool->clean_list);
275 ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
277 clear_bit(CLEAN_LIST_BUSY_BIT, flag);
282 static inline void wait_clean_list_grace(void)
287 for_each_online_cpu(cpu) {
288 flag = &per_cpu(clean_list_grace, cpu);
289 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
294 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
296 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
297 struct rds_ib_mr *ibmr = NULL;
298 int err = 0, iter = 0;
300 if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
301 schedule_delayed_work(&pool->flush_worker, 10);
304 ibmr = rds_ib_reuse_fmr(pool);
308 /* No clean MRs - now we have the choice of either
309 * allocating a fresh MR up to the limit imposed by the
310 * driver, or flush any dirty unused MRs.
311 * We try to avoid stalling in the send path if possible,
312 * so we allocate as long as we're allowed to.
314 * We're fussy with enforcing the FMR limit, though. If the driver
315 * tells us we can't use more than N fmrs, we shouldn't start
317 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
320 atomic_dec(&pool->item_count);
323 rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
324 return ERR_PTR(-EAGAIN);
327 /* We do have some empty MRs. Flush them out. */
328 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
329 rds_ib_flush_mr_pool(pool, 0, &ibmr);
334 ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
340 memset(ibmr, 0, sizeof(*ibmr));
342 ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
343 (IB_ACCESS_LOCAL_WRITE |
344 IB_ACCESS_REMOTE_READ |
345 IB_ACCESS_REMOTE_WRITE|
346 IB_ACCESS_REMOTE_ATOMIC),
348 if (IS_ERR(ibmr->fmr)) {
349 err = PTR_ERR(ibmr->fmr);
351 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
355 rds_ib_stats_inc(s_ib_rdma_mr_alloc);
361 ib_dealloc_fmr(ibmr->fmr);
364 atomic_dec(&pool->item_count);
368 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
369 struct scatterlist *sg, unsigned int nents)
371 struct ib_device *dev = rds_ibdev->dev;
372 struct scatterlist *scat = sg;
376 int page_cnt, sg_dma_len;
380 sg_dma_len = ib_dma_map_sg(dev, sg, nents,
382 if (unlikely(!sg_dma_len)) {
383 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
390 for (i = 0; i < sg_dma_len; ++i) {
391 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
392 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
394 if (dma_addr & ~PAGE_MASK) {
400 if ((dma_addr + dma_len) & ~PAGE_MASK) {
401 if (i < sg_dma_len - 1)
410 page_cnt += len >> PAGE_SHIFT;
411 if (page_cnt > fmr_message_size)
414 dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
415 rdsibdev_to_node(rds_ibdev));
420 for (i = 0; i < sg_dma_len; ++i) {
421 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
422 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
424 for (j = 0; j < dma_len; j += PAGE_SIZE)
425 dma_pages[page_cnt++] =
426 (dma_addr & PAGE_MASK) + j;
429 ret = ib_map_phys_fmr(ibmr->fmr,
430 dma_pages, page_cnt, io_addr);
434 /* Success - we successfully remapped the MR, so we can
435 * safely tear down the old mapping. */
436 rds_ib_teardown_mr(ibmr);
439 ibmr->sg_len = nents;
440 ibmr->sg_dma_len = sg_dma_len;
443 rds_ib_stats_inc(s_ib_rdma_mr_used);
452 void rds_ib_sync_mr(void *trans_private, int direction)
454 struct rds_ib_mr *ibmr = trans_private;
455 struct rds_ib_device *rds_ibdev = ibmr->device;
458 case DMA_FROM_DEVICE:
459 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
460 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
463 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
464 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
469 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
471 struct rds_ib_device *rds_ibdev = ibmr->device;
473 if (ibmr->sg_dma_len) {
474 ib_dma_unmap_sg(rds_ibdev->dev,
475 ibmr->sg, ibmr->sg_len,
477 ibmr->sg_dma_len = 0;
480 /* Release the s/g list */
484 for (i = 0; i < ibmr->sg_len; ++i) {
485 struct page *page = sg_page(&ibmr->sg[i]);
487 /* FIXME we need a way to tell a r/w MR
489 BUG_ON(irqs_disabled());
490 set_page_dirty(page);
500 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
502 unsigned int pinned = ibmr->sg_len;
504 __rds_ib_teardown_mr(ibmr);
506 struct rds_ib_device *rds_ibdev = ibmr->device;
507 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
509 atomic_sub(pinned, &pool->free_pinned);
513 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
515 unsigned int item_count;
517 item_count = atomic_read(&pool->item_count);
525 * given an llist of mrs, put them all into the list_head for more processing
527 static void llist_append_to_list(struct llist_head *llist, struct list_head *list)
529 struct rds_ib_mr *ibmr;
530 struct llist_node *node;
531 struct llist_node *next;
533 node = llist_del_all(llist);
536 ibmr = llist_entry(node, struct rds_ib_mr, llnode);
537 list_add_tail(&ibmr->unmap_list, list);
543 * this takes a list head of mrs and turns it into linked llist nodes
544 * of clusters. Each cluster has linked llist nodes of
545 * MR_CLUSTER_SIZE mrs that are ready for reuse.
547 static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
548 struct list_head *list,
549 struct llist_node **nodes_head,
550 struct llist_node **nodes_tail)
552 struct rds_ib_mr *ibmr;
553 struct llist_node *cur = NULL;
554 struct llist_node **next = nodes_head;
556 list_for_each_entry(ibmr, list, unmap_list) {
566 * Flush our pool of MRs.
567 * At a minimum, all currently unused MRs are unmapped.
568 * If the number of MRs allocated exceeds the limit, we also try
569 * to free as many MRs as needed to get back to this limit.
571 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
572 int free_all, struct rds_ib_mr **ibmr_ret)
574 struct rds_ib_mr *ibmr, *next;
575 struct llist_node *clean_nodes;
576 struct llist_node *clean_tail;
577 LIST_HEAD(unmap_list);
579 unsigned long unpinned = 0;
580 unsigned int nfreed = 0, ncleaned = 0, free_goal;
583 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
587 while(!mutex_trylock(&pool->flush_lock)) {
588 ibmr = rds_ib_reuse_fmr(pool);
591 finish_wait(&pool->flush_wait, &wait);
595 prepare_to_wait(&pool->flush_wait, &wait,
596 TASK_UNINTERRUPTIBLE);
597 if (llist_empty(&pool->clean_list))
600 ibmr = rds_ib_reuse_fmr(pool);
603 finish_wait(&pool->flush_wait, &wait);
607 finish_wait(&pool->flush_wait, &wait);
609 mutex_lock(&pool->flush_lock);
612 ibmr = rds_ib_reuse_fmr(pool);
619 /* Get the list of all MRs to be dropped. Ordering matters -
620 * we want to put drop_list ahead of free_list.
622 llist_append_to_list(&pool->drop_list, &unmap_list);
623 llist_append_to_list(&pool->free_list, &unmap_list);
625 llist_append_to_list(&pool->clean_list, &unmap_list);
627 free_goal = rds_ib_flush_goal(pool, free_all);
629 if (list_empty(&unmap_list))
632 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
633 list_for_each_entry(ibmr, &unmap_list, unmap_list)
634 list_add(&ibmr->fmr->list, &fmr_list);
636 ret = ib_unmap_fmr(&fmr_list);
638 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
640 /* Now we can destroy the DMA mapping and unpin any pages */
641 list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
642 unpinned += ibmr->sg_len;
643 __rds_ib_teardown_mr(ibmr);
644 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
645 rds_ib_stats_inc(s_ib_rdma_mr_free);
646 list_del(&ibmr->unmap_list);
647 ib_dealloc_fmr(ibmr->fmr);
654 if (!list_empty(&unmap_list)) {
655 /* we have to make sure that none of the things we're about
656 * to put on the clean list would race with other cpus trying
657 * to pull items off. The llist would explode if we managed to
658 * remove something from the clean list and then add it back again
659 * while another CPU was spinning on that same item in llist_del_first.
661 * This is pretty unlikely, but just in case wait for an llist grace period
662 * here before adding anything back into the clean list.
664 wait_clean_list_grace();
666 list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
668 *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
670 /* more than one entry in llist nodes */
671 if (clean_nodes->next)
672 llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
676 atomic_sub(unpinned, &pool->free_pinned);
677 atomic_sub(ncleaned, &pool->dirty_count);
678 atomic_sub(nfreed, &pool->item_count);
681 mutex_unlock(&pool->flush_lock);
682 if (waitqueue_active(&pool->flush_wait))
683 wake_up(&pool->flush_wait);
688 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
690 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
692 rds_ib_flush_mr_pool(pool, 0, NULL);
695 void rds_ib_free_mr(void *trans_private, int invalidate)
697 struct rds_ib_mr *ibmr = trans_private;
698 struct rds_ib_device *rds_ibdev = ibmr->device;
699 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
701 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
703 /* Return it to the pool's free list */
704 if (ibmr->remap_count >= pool->fmr_attr.max_maps)
705 llist_add(&ibmr->llnode, &pool->drop_list);
707 llist_add(&ibmr->llnode, &pool->free_list);
709 atomic_add(ibmr->sg_len, &pool->free_pinned);
710 atomic_inc(&pool->dirty_count);
712 /* If we've pinned too many pages, request a flush */
713 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
714 atomic_read(&pool->dirty_count) >= pool->max_items / 10)
715 schedule_delayed_work(&pool->flush_worker, 10);
718 if (likely(!in_interrupt())) {
719 rds_ib_flush_mr_pool(pool, 0, NULL);
721 /* We get here if the user created a MR marked
722 * as use_once and invalidate at the same time. */
723 schedule_delayed_work(&pool->flush_worker, 10);
727 rds_ib_dev_put(rds_ibdev);
730 void rds_ib_flush_mrs(void)
732 struct rds_ib_device *rds_ibdev;
734 down_read(&rds_ib_devices_lock);
735 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
736 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
739 rds_ib_flush_mr_pool(pool, 0, NULL);
741 up_read(&rds_ib_devices_lock);
744 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
745 struct rds_sock *rs, u32 *key_ret)
747 struct rds_ib_device *rds_ibdev;
748 struct rds_ib_mr *ibmr = NULL;
751 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
757 if (!rds_ibdev->mr_pool) {
762 ibmr = rds_ib_alloc_fmr(rds_ibdev);
764 rds_ib_dev_put(rds_ibdev);
768 ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
770 *key_ret = ibmr->fmr->rkey;
772 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
774 ibmr->device = rds_ibdev;
780 rds_ib_free_mr(ibmr, 0);
784 rds_ib_dev_put(rds_ibdev);