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>
42 * This is stored as mr->r_trans_private.
45 struct rds_ib_device *device;
46 struct rds_ib_mr_pool *pool;
48 struct list_head list;
49 unsigned int remap_count;
51 struct scatterlist *sg;
58 * Our own little FMR pool
60 struct rds_ib_mr_pool {
61 struct mutex flush_lock; /* serialize fmr invalidate */
62 struct work_struct flush_worker; /* flush worker */
64 spinlock_t list_lock; /* protect variables below */
65 atomic_t item_count; /* total # of MRs */
66 atomic_t dirty_count; /* # dirty of MRs */
67 struct list_head drop_list; /* MRs that have reached their max_maps limit */
68 struct list_head free_list; /* unused MRs */
69 struct list_head clean_list; /* unused & unamapped MRs */
70 atomic_t free_pinned; /* memory pinned by free MRs */
71 unsigned long max_items;
72 unsigned long max_items_soft;
73 unsigned long max_free_pinned;
74 struct ib_fmr_attr fmr_attr;
77 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all);
78 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
79 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
81 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
83 struct rds_ib_device *rds_ibdev;
84 struct rds_ib_ipaddr *i_ipaddr;
86 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
88 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
89 if (i_ipaddr->ipaddr == ipaddr) {
100 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
102 struct rds_ib_ipaddr *i_ipaddr;
104 i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
108 i_ipaddr->ipaddr = ipaddr;
110 spin_lock_irq(&rds_ibdev->spinlock);
111 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
112 spin_unlock_irq(&rds_ibdev->spinlock);
117 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
119 struct rds_ib_ipaddr *i_ipaddr;
120 struct rds_ib_ipaddr *to_free = NULL;
123 spin_lock_irq(&rds_ibdev->spinlock);
124 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
125 if (i_ipaddr->ipaddr == ipaddr) {
126 list_del_rcu(&i_ipaddr->list);
131 spin_unlock_irq(&rds_ibdev->spinlock);
139 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
141 struct rds_ib_device *rds_ibdev_old;
143 rds_ibdev_old = rds_ib_get_device(ipaddr);
145 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
147 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
150 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
152 struct rds_ib_connection *ic = conn->c_transport_data;
154 /* conn was previously on the nodev_conns_list */
155 spin_lock_irq(&ib_nodev_conns_lock);
156 BUG_ON(list_empty(&ib_nodev_conns));
157 BUG_ON(list_empty(&ic->ib_node));
158 list_del(&ic->ib_node);
160 spin_lock_irq(&rds_ibdev->spinlock);
161 list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
162 spin_unlock_irq(&rds_ibdev->spinlock);
163 spin_unlock_irq(&ib_nodev_conns_lock);
165 ic->rds_ibdev = rds_ibdev;
168 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
170 struct rds_ib_connection *ic = conn->c_transport_data;
172 /* place conn on nodev_conns_list */
173 spin_lock(&ib_nodev_conns_lock);
175 spin_lock_irq(&rds_ibdev->spinlock);
176 BUG_ON(list_empty(&ic->ib_node));
177 list_del(&ic->ib_node);
178 spin_unlock_irq(&rds_ibdev->spinlock);
180 list_add_tail(&ic->ib_node, &ib_nodev_conns);
182 spin_unlock(&ib_nodev_conns_lock);
184 ic->rds_ibdev = NULL;
187 void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock)
189 struct rds_ib_connection *ic, *_ic;
192 /* avoid calling conn_destroy with irqs off */
193 spin_lock_irq(list_lock);
194 list_splice(list, &tmp_list);
195 INIT_LIST_HEAD(list);
196 spin_unlock_irq(list_lock);
198 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
199 rds_conn_destroy(ic->conn);
202 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
204 struct rds_ib_mr_pool *pool;
206 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
208 return ERR_PTR(-ENOMEM);
210 INIT_LIST_HEAD(&pool->free_list);
211 INIT_LIST_HEAD(&pool->drop_list);
212 INIT_LIST_HEAD(&pool->clean_list);
213 mutex_init(&pool->flush_lock);
214 spin_lock_init(&pool->list_lock);
215 INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
217 pool->fmr_attr.max_pages = fmr_message_size;
218 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
219 pool->fmr_attr.page_shift = PAGE_SHIFT;
220 pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
222 /* We never allow more than max_items MRs to be allocated.
223 * When we exceed more than max_items_soft, we start freeing
224 * items more aggressively.
225 * Make sure that max_items > max_items_soft > max_items / 2
227 pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
228 pool->max_items = rds_ibdev->max_fmrs;
233 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
235 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
237 iinfo->rdma_mr_max = pool->max_items;
238 iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
241 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
243 flush_workqueue(rds_wq);
244 rds_ib_flush_mr_pool(pool, 1);
245 WARN_ON(atomic_read(&pool->item_count));
246 WARN_ON(atomic_read(&pool->free_pinned));
250 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
252 struct rds_ib_mr *ibmr = NULL;
255 spin_lock_irqsave(&pool->list_lock, flags);
256 if (!list_empty(&pool->clean_list)) {
257 ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list);
258 list_del_init(&ibmr->list);
260 spin_unlock_irqrestore(&pool->list_lock, flags);
265 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
267 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
268 struct rds_ib_mr *ibmr = NULL;
269 int err = 0, iter = 0;
272 ibmr = rds_ib_reuse_fmr(pool);
276 /* No clean MRs - now we have the choice of either
277 * allocating a fresh MR up to the limit imposed by the
278 * driver, or flush any dirty unused MRs.
279 * We try to avoid stalling in the send path if possible,
280 * so we allocate as long as we're allowed to.
282 * We're fussy with enforcing the FMR limit, though. If the driver
283 * tells us we can't use more than N fmrs, we shouldn't start
285 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
288 atomic_dec(&pool->item_count);
291 rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
292 return ERR_PTR(-EAGAIN);
295 /* We do have some empty MRs. Flush them out. */
296 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
297 rds_ib_flush_mr_pool(pool, 0);
300 ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
306 memset(ibmr, 0, sizeof(*ibmr));
308 ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
309 (IB_ACCESS_LOCAL_WRITE |
310 IB_ACCESS_REMOTE_READ |
311 IB_ACCESS_REMOTE_WRITE|
312 IB_ACCESS_REMOTE_ATOMIC),
315 if (IS_ERR(ibmr->fmr)) {
316 err = PTR_ERR(ibmr->fmr);
318 printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
322 rds_ib_stats_inc(s_ib_rdma_mr_alloc);
328 ib_dealloc_fmr(ibmr->fmr);
331 atomic_dec(&pool->item_count);
335 static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
336 struct scatterlist *sg, unsigned int nents)
338 struct ib_device *dev = rds_ibdev->dev;
339 struct scatterlist *scat = sg;
343 int page_cnt, sg_dma_len;
347 sg_dma_len = ib_dma_map_sg(dev, sg, nents,
349 if (unlikely(!sg_dma_len)) {
350 printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
357 for (i = 0; i < sg_dma_len; ++i) {
358 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
359 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
361 if (dma_addr & ~PAGE_MASK) {
367 if ((dma_addr + dma_len) & ~PAGE_MASK) {
368 if (i < sg_dma_len - 1)
377 page_cnt += len >> PAGE_SHIFT;
378 if (page_cnt > fmr_message_size)
381 dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
382 rdsibdev_to_node(rds_ibdev));
387 for (i = 0; i < sg_dma_len; ++i) {
388 unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
389 u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
391 for (j = 0; j < dma_len; j += PAGE_SIZE)
392 dma_pages[page_cnt++] =
393 (dma_addr & PAGE_MASK) + j;
396 ret = ib_map_phys_fmr(ibmr->fmr,
397 dma_pages, page_cnt, io_addr);
401 /* Success - we successfully remapped the MR, so we can
402 * safely tear down the old mapping. */
403 rds_ib_teardown_mr(ibmr);
406 ibmr->sg_len = nents;
407 ibmr->sg_dma_len = sg_dma_len;
410 rds_ib_stats_inc(s_ib_rdma_mr_used);
419 void rds_ib_sync_mr(void *trans_private, int direction)
421 struct rds_ib_mr *ibmr = trans_private;
422 struct rds_ib_device *rds_ibdev = ibmr->device;
425 case DMA_FROM_DEVICE:
426 ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
427 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
430 ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
431 ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
436 static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
438 struct rds_ib_device *rds_ibdev = ibmr->device;
440 if (ibmr->sg_dma_len) {
441 ib_dma_unmap_sg(rds_ibdev->dev,
442 ibmr->sg, ibmr->sg_len,
444 ibmr->sg_dma_len = 0;
447 /* Release the s/g list */
451 for (i = 0; i < ibmr->sg_len; ++i) {
452 struct page *page = sg_page(&ibmr->sg[i]);
454 /* FIXME we need a way to tell a r/w MR
456 BUG_ON(irqs_disabled());
457 set_page_dirty(page);
467 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
469 unsigned int pinned = ibmr->sg_len;
471 __rds_ib_teardown_mr(ibmr);
473 struct rds_ib_device *rds_ibdev = ibmr->device;
474 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
476 atomic_sub(pinned, &pool->free_pinned);
480 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
482 unsigned int item_count;
484 item_count = atomic_read(&pool->item_count);
492 * Flush our pool of MRs.
493 * At a minimum, all currently unused MRs are unmapped.
494 * If the number of MRs allocated exceeds the limit, we also try
495 * to free as many MRs as needed to get back to this limit.
497 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
499 struct rds_ib_mr *ibmr, *next;
500 LIST_HEAD(unmap_list);
502 unsigned long unpinned = 0;
504 unsigned int nfreed = 0, ncleaned = 0, free_goal;
507 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
509 mutex_lock(&pool->flush_lock);
511 spin_lock_irqsave(&pool->list_lock, flags);
512 /* Get the list of all MRs to be dropped. Ordering matters -
513 * we want to put drop_list ahead of free_list. */
514 list_splice_init(&pool->free_list, &unmap_list);
515 list_splice_init(&pool->drop_list, &unmap_list);
517 list_splice_init(&pool->clean_list, &unmap_list);
518 spin_unlock_irqrestore(&pool->list_lock, flags);
520 free_goal = rds_ib_flush_goal(pool, free_all);
522 if (list_empty(&unmap_list))
525 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
526 list_for_each_entry(ibmr, &unmap_list, list)
527 list_add(&ibmr->fmr->list, &fmr_list);
528 ret = ib_unmap_fmr(&fmr_list);
530 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
532 /* Now we can destroy the DMA mapping and unpin any pages */
533 list_for_each_entry_safe(ibmr, next, &unmap_list, list) {
534 unpinned += ibmr->sg_len;
535 __rds_ib_teardown_mr(ibmr);
536 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
537 rds_ib_stats_inc(s_ib_rdma_mr_free);
538 list_del(&ibmr->list);
539 ib_dealloc_fmr(ibmr->fmr);
546 spin_lock_irqsave(&pool->list_lock, flags);
547 list_splice(&unmap_list, &pool->clean_list);
548 spin_unlock_irqrestore(&pool->list_lock, flags);
550 atomic_sub(unpinned, &pool->free_pinned);
551 atomic_sub(ncleaned, &pool->dirty_count);
552 atomic_sub(nfreed, &pool->item_count);
555 mutex_unlock(&pool->flush_lock);
559 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
561 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker);
563 rds_ib_flush_mr_pool(pool, 0);
566 void rds_ib_free_mr(void *trans_private, int invalidate)
568 struct rds_ib_mr *ibmr = trans_private;
569 struct rds_ib_device *rds_ibdev = ibmr->device;
570 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
573 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
575 /* Return it to the pool's free list */
576 spin_lock_irqsave(&pool->list_lock, flags);
577 if (ibmr->remap_count >= pool->fmr_attr.max_maps)
578 list_add(&ibmr->list, &pool->drop_list);
580 list_add(&ibmr->list, &pool->free_list);
582 atomic_add(ibmr->sg_len, &pool->free_pinned);
583 atomic_inc(&pool->dirty_count);
584 spin_unlock_irqrestore(&pool->list_lock, flags);
586 /* If we've pinned too many pages, request a flush */
587 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
588 atomic_read(&pool->dirty_count) >= pool->max_items / 10)
589 queue_work(rds_wq, &pool->flush_worker);
592 if (likely(!in_interrupt())) {
593 rds_ib_flush_mr_pool(pool, 0);
595 /* We get here if the user created a MR marked
596 * as use_once and invalidate at the same time. */
597 queue_work(rds_wq, &pool->flush_worker);
602 void rds_ib_flush_mrs(void)
604 struct rds_ib_device *rds_ibdev;
606 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
607 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
610 rds_ib_flush_mr_pool(pool, 0);
614 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
615 struct rds_sock *rs, u32 *key_ret)
617 struct rds_ib_device *rds_ibdev;
618 struct rds_ib_mr *ibmr = NULL;
621 rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
627 if (!rds_ibdev->mr_pool) {
632 ibmr = rds_ib_alloc_fmr(rds_ibdev);
636 ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
638 *key_ret = ibmr->fmr->rkey;
640 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
642 ibmr->device = rds_ibdev;
647 rds_ib_free_mr(ibmr, 0);