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>
41 * This is stored as mr->r_trans_private.
44 struct rds_ib_device
*device
;
45 struct rds_ib_mr_pool
*pool
;
47 struct list_head list
;
48 unsigned int remap_count
;
50 struct scatterlist
*sg
;
57 * Our own little FMR pool
59 struct rds_ib_mr_pool
{
60 struct mutex flush_lock
; /* serialize fmr invalidate */
61 struct work_struct flush_worker
; /* flush worker */
63 spinlock_t list_lock
; /* protect variables below */
64 atomic_t item_count
; /* total # of MRs */
65 atomic_t dirty_count
; /* # dirty of MRs */
66 struct list_head drop_list
; /* MRs that have reached their max_maps limit */
67 struct list_head free_list
; /* unused MRs */
68 struct list_head clean_list
; /* unused & unamapped MRs */
69 atomic_t free_pinned
; /* memory pinned by free MRs */
70 unsigned long max_items
;
71 unsigned long max_items_soft
;
72 unsigned long max_free_pinned
;
73 struct ib_fmr_attr fmr_attr
;
76 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool
*pool
, int free_all
);
77 static void rds_ib_teardown_mr(struct rds_ib_mr
*ibmr
);
78 static void rds_ib_mr_pool_flush_worker(struct work_struct
*work
);
80 static struct rds_ib_device
*rds_ib_get_device(__be32 ipaddr
)
82 struct rds_ib_device
*rds_ibdev
;
83 struct rds_ib_ipaddr
*i_ipaddr
;
85 list_for_each_entry(rds_ibdev
, &rds_ib_devices
, list
) {
86 spin_lock_irq(&rds_ibdev
->spinlock
);
87 list_for_each_entry(i_ipaddr
, &rds_ibdev
->ipaddr_list
, list
) {
88 if (i_ipaddr
->ipaddr
== ipaddr
) {
89 spin_unlock_irq(&rds_ibdev
->spinlock
);
93 spin_unlock_irq(&rds_ibdev
->spinlock
);
99 static int rds_ib_add_ipaddr(struct rds_ib_device
*rds_ibdev
, __be32 ipaddr
)
101 struct rds_ib_ipaddr
*i_ipaddr
;
103 i_ipaddr
= kmalloc(sizeof *i_ipaddr
, GFP_KERNEL
);
107 i_ipaddr
->ipaddr
= ipaddr
;
109 spin_lock_irq(&rds_ibdev
->spinlock
);
110 list_add_tail(&i_ipaddr
->list
, &rds_ibdev
->ipaddr_list
);
111 spin_unlock_irq(&rds_ibdev
->spinlock
);
116 static void rds_ib_remove_ipaddr(struct rds_ib_device
*rds_ibdev
, __be32 ipaddr
)
118 struct rds_ib_ipaddr
*i_ipaddr
, *next
;
120 spin_lock_irq(&rds_ibdev
->spinlock
);
121 list_for_each_entry_safe(i_ipaddr
, next
, &rds_ibdev
->ipaddr_list
, list
) {
122 if (i_ipaddr
->ipaddr
== ipaddr
) {
123 list_del(&i_ipaddr
->list
);
128 spin_unlock_irq(&rds_ibdev
->spinlock
);
131 int rds_ib_update_ipaddr(struct rds_ib_device
*rds_ibdev
, __be32 ipaddr
)
133 struct rds_ib_device
*rds_ibdev_old
;
135 rds_ibdev_old
= rds_ib_get_device(ipaddr
);
137 rds_ib_remove_ipaddr(rds_ibdev_old
, ipaddr
);
139 return rds_ib_add_ipaddr(rds_ibdev
, ipaddr
);
142 void rds_ib_add_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
)
144 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
146 /* conn was previously on the nodev_conns_list */
147 spin_lock_irq(&ib_nodev_conns_lock
);
148 BUG_ON(list_empty(&ib_nodev_conns
));
149 BUG_ON(list_empty(&ic
->ib_node
));
150 list_del(&ic
->ib_node
);
152 spin_lock_irq(&rds_ibdev
->spinlock
);
153 list_add_tail(&ic
->ib_node
, &rds_ibdev
->conn_list
);
154 spin_unlock_irq(&rds_ibdev
->spinlock
);
155 spin_unlock_irq(&ib_nodev_conns_lock
);
157 ic
->rds_ibdev
= rds_ibdev
;
160 void rds_ib_remove_conn(struct rds_ib_device
*rds_ibdev
, struct rds_connection
*conn
)
162 struct rds_ib_connection
*ic
= conn
->c_transport_data
;
164 /* place conn on nodev_conns_list */
165 spin_lock(&ib_nodev_conns_lock
);
167 spin_lock_irq(&rds_ibdev
->spinlock
);
168 BUG_ON(list_empty(&ic
->ib_node
));
169 list_del(&ic
->ib_node
);
170 spin_unlock_irq(&rds_ibdev
->spinlock
);
172 list_add_tail(&ic
->ib_node
, &ib_nodev_conns
);
174 spin_unlock(&ib_nodev_conns_lock
);
176 ic
->rds_ibdev
= NULL
;
179 void __rds_ib_destroy_conns(struct list_head
*list
, spinlock_t
*list_lock
)
181 struct rds_ib_connection
*ic
, *_ic
;
184 /* avoid calling conn_destroy with irqs off */
185 spin_lock_irq(list_lock
);
186 list_splice(list
, &tmp_list
);
187 INIT_LIST_HEAD(list
);
188 spin_unlock_irq(list_lock
);
190 list_for_each_entry_safe(ic
, _ic
, &tmp_list
, ib_node
)
191 rds_conn_destroy(ic
->conn
);
194 struct rds_ib_mr_pool
*rds_ib_create_mr_pool(struct rds_ib_device
*rds_ibdev
)
196 struct rds_ib_mr_pool
*pool
;
198 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
200 return ERR_PTR(-ENOMEM
);
202 INIT_LIST_HEAD(&pool
->free_list
);
203 INIT_LIST_HEAD(&pool
->drop_list
);
204 INIT_LIST_HEAD(&pool
->clean_list
);
205 mutex_init(&pool
->flush_lock
);
206 spin_lock_init(&pool
->list_lock
);
207 INIT_WORK(&pool
->flush_worker
, rds_ib_mr_pool_flush_worker
);
209 pool
->fmr_attr
.max_pages
= fmr_message_size
;
210 pool
->fmr_attr
.max_maps
= rds_ibdev
->fmr_max_remaps
;
211 pool
->fmr_attr
.page_shift
= PAGE_SHIFT
;
212 pool
->max_free_pinned
= rds_ibdev
->max_fmrs
* fmr_message_size
/ 4;
214 /* We never allow more than max_items MRs to be allocated.
215 * When we exceed more than max_items_soft, we start freeing
216 * items more aggressively.
217 * Make sure that max_items > max_items_soft > max_items / 2
219 pool
->max_items_soft
= rds_ibdev
->max_fmrs
* 3 / 4;
220 pool
->max_items
= rds_ibdev
->max_fmrs
;
225 void rds_ib_get_mr_info(struct rds_ib_device
*rds_ibdev
, struct rds_info_rdma_connection
*iinfo
)
227 struct rds_ib_mr_pool
*pool
= rds_ibdev
->mr_pool
;
229 iinfo
->rdma_mr_max
= pool
->max_items
;
230 iinfo
->rdma_mr_size
= pool
->fmr_attr
.max_pages
;
233 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool
*pool
)
235 flush_workqueue(rds_wq
);
236 rds_ib_flush_mr_pool(pool
, 1);
237 BUG_ON(atomic_read(&pool
->item_count
));
238 BUG_ON(atomic_read(&pool
->free_pinned
));
242 static inline struct rds_ib_mr
*rds_ib_reuse_fmr(struct rds_ib_mr_pool
*pool
)
244 struct rds_ib_mr
*ibmr
= NULL
;
247 spin_lock_irqsave(&pool
->list_lock
, flags
);
248 if (!list_empty(&pool
->clean_list
)) {
249 ibmr
= list_entry(pool
->clean_list
.next
, struct rds_ib_mr
, list
);
250 list_del_init(&ibmr
->list
);
252 spin_unlock_irqrestore(&pool
->list_lock
, flags
);
257 static struct rds_ib_mr
*rds_ib_alloc_fmr(struct rds_ib_device
*rds_ibdev
)
259 struct rds_ib_mr_pool
*pool
= rds_ibdev
->mr_pool
;
260 struct rds_ib_mr
*ibmr
= NULL
;
261 int err
= 0, iter
= 0;
264 ibmr
= rds_ib_reuse_fmr(pool
);
268 /* No clean MRs - now we have the choice of either
269 * allocating a fresh MR up to the limit imposed by the
270 * driver, or flush any dirty unused MRs.
271 * We try to avoid stalling in the send path if possible,
272 * so we allocate as long as we're allowed to.
274 * We're fussy with enforcing the FMR limit, though. If the driver
275 * tells us we can't use more than N fmrs, we shouldn't start
277 if (atomic_inc_return(&pool
->item_count
) <= pool
->max_items
)
280 atomic_dec(&pool
->item_count
);
283 rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted
);
284 return ERR_PTR(-EAGAIN
);
287 /* We do have some empty MRs. Flush them out. */
288 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait
);
289 rds_ib_flush_mr_pool(pool
, 0);
292 ibmr
= kzalloc(sizeof(*ibmr
), GFP_KERNEL
);
298 ibmr
->fmr
= ib_alloc_fmr(rds_ibdev
->pd
,
299 (IB_ACCESS_LOCAL_WRITE
|
300 IB_ACCESS_REMOTE_READ
|
301 IB_ACCESS_REMOTE_WRITE
),
303 if (IS_ERR(ibmr
->fmr
)) {
304 err
= PTR_ERR(ibmr
->fmr
);
306 printk(KERN_WARNING
"RDS/IB: ib_alloc_fmr failed (err=%d)\n", err
);
310 rds_ib_stats_inc(s_ib_rdma_mr_alloc
);
316 ib_dealloc_fmr(ibmr
->fmr
);
319 atomic_dec(&pool
->item_count
);
323 static int rds_ib_map_fmr(struct rds_ib_device
*rds_ibdev
, struct rds_ib_mr
*ibmr
,
324 struct scatterlist
*sg
, unsigned int nents
)
326 struct ib_device
*dev
= rds_ibdev
->dev
;
327 struct scatterlist
*scat
= sg
;
331 int page_cnt
, sg_dma_len
;
335 sg_dma_len
= ib_dma_map_sg(dev
, sg
, nents
,
337 if (unlikely(!sg_dma_len
)) {
338 printk(KERN_WARNING
"RDS/IB: dma_map_sg failed!\n");
345 for (i
= 0; i
< sg_dma_len
; ++i
) {
346 unsigned int dma_len
= ib_sg_dma_len(dev
, &scat
[i
]);
347 u64 dma_addr
= ib_sg_dma_address(dev
, &scat
[i
]);
349 if (dma_addr
& ~PAGE_MASK
) {
355 if ((dma_addr
+ dma_len
) & ~PAGE_MASK
) {
356 if (i
< sg_dma_len
- 1)
365 page_cnt
+= len
>> PAGE_SHIFT
;
366 if (page_cnt
> fmr_message_size
)
369 dma_pages
= kmalloc(sizeof(u64
) * page_cnt
, GFP_ATOMIC
);
374 for (i
= 0; i
< sg_dma_len
; ++i
) {
375 unsigned int dma_len
= ib_sg_dma_len(dev
, &scat
[i
]);
376 u64 dma_addr
= ib_sg_dma_address(dev
, &scat
[i
]);
378 for (j
= 0; j
< dma_len
; j
+= PAGE_SIZE
)
379 dma_pages
[page_cnt
++] =
380 (dma_addr
& PAGE_MASK
) + j
;
383 ret
= ib_map_phys_fmr(ibmr
->fmr
,
384 dma_pages
, page_cnt
, io_addr
);
388 /* Success - we successfully remapped the MR, so we can
389 * safely tear down the old mapping. */
390 rds_ib_teardown_mr(ibmr
);
393 ibmr
->sg_len
= nents
;
394 ibmr
->sg_dma_len
= sg_dma_len
;
397 rds_ib_stats_inc(s_ib_rdma_mr_used
);
406 void rds_ib_sync_mr(void *trans_private
, int direction
)
408 struct rds_ib_mr
*ibmr
= trans_private
;
409 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
412 case DMA_FROM_DEVICE
:
413 ib_dma_sync_sg_for_cpu(rds_ibdev
->dev
, ibmr
->sg
,
414 ibmr
->sg_dma_len
, DMA_BIDIRECTIONAL
);
417 ib_dma_sync_sg_for_device(rds_ibdev
->dev
, ibmr
->sg
,
418 ibmr
->sg_dma_len
, DMA_BIDIRECTIONAL
);
423 static void __rds_ib_teardown_mr(struct rds_ib_mr
*ibmr
)
425 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
427 if (ibmr
->sg_dma_len
) {
428 ib_dma_unmap_sg(rds_ibdev
->dev
,
429 ibmr
->sg
, ibmr
->sg_len
,
431 ibmr
->sg_dma_len
= 0;
434 /* Release the s/g list */
438 for (i
= 0; i
< ibmr
->sg_len
; ++i
) {
439 struct page
*page
= sg_page(&ibmr
->sg
[i
]);
441 /* FIXME we need a way to tell a r/w MR
443 set_page_dirty(page
);
453 static void rds_ib_teardown_mr(struct rds_ib_mr
*ibmr
)
455 unsigned int pinned
= ibmr
->sg_len
;
457 __rds_ib_teardown_mr(ibmr
);
459 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
460 struct rds_ib_mr_pool
*pool
= rds_ibdev
->mr_pool
;
462 atomic_sub(pinned
, &pool
->free_pinned
);
466 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool
*pool
, int free_all
)
468 unsigned int item_count
;
470 item_count
= atomic_read(&pool
->item_count
);
478 * Flush our pool of MRs.
479 * At a minimum, all currently unused MRs are unmapped.
480 * If the number of MRs allocated exceeds the limit, we also try
481 * to free as many MRs as needed to get back to this limit.
483 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool
*pool
, int free_all
)
485 struct rds_ib_mr
*ibmr
, *next
;
486 LIST_HEAD(unmap_list
);
488 unsigned long unpinned
= 0;
490 unsigned int nfreed
= 0, ncleaned
= 0, free_goal
;
493 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush
);
495 mutex_lock(&pool
->flush_lock
);
497 spin_lock_irqsave(&pool
->list_lock
, flags
);
498 /* Get the list of all MRs to be dropped. Ordering matters -
499 * we want to put drop_list ahead of free_list. */
500 list_splice_init(&pool
->free_list
, &unmap_list
);
501 list_splice_init(&pool
->drop_list
, &unmap_list
);
503 list_splice_init(&pool
->clean_list
, &unmap_list
);
504 spin_unlock_irqrestore(&pool
->list_lock
, flags
);
506 free_goal
= rds_ib_flush_goal(pool
, free_all
);
508 if (list_empty(&unmap_list
))
511 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
512 list_for_each_entry(ibmr
, &unmap_list
, list
)
513 list_add(&ibmr
->fmr
->list
, &fmr_list
);
514 ret
= ib_unmap_fmr(&fmr_list
);
516 printk(KERN_WARNING
"RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret
);
518 /* Now we can destroy the DMA mapping and unpin any pages */
519 list_for_each_entry_safe(ibmr
, next
, &unmap_list
, list
) {
520 unpinned
+= ibmr
->sg_len
;
521 __rds_ib_teardown_mr(ibmr
);
522 if (nfreed
< free_goal
|| ibmr
->remap_count
>= pool
->fmr_attr
.max_maps
) {
523 rds_ib_stats_inc(s_ib_rdma_mr_free
);
524 list_del(&ibmr
->list
);
525 ib_dealloc_fmr(ibmr
->fmr
);
532 spin_lock_irqsave(&pool
->list_lock
, flags
);
533 list_splice(&unmap_list
, &pool
->clean_list
);
534 spin_unlock_irqrestore(&pool
->list_lock
, flags
);
536 atomic_sub(unpinned
, &pool
->free_pinned
);
537 atomic_sub(ncleaned
, &pool
->dirty_count
);
538 atomic_sub(nfreed
, &pool
->item_count
);
541 mutex_unlock(&pool
->flush_lock
);
545 static void rds_ib_mr_pool_flush_worker(struct work_struct
*work
)
547 struct rds_ib_mr_pool
*pool
= container_of(work
, struct rds_ib_mr_pool
, flush_worker
);
549 rds_ib_flush_mr_pool(pool
, 0);
552 void rds_ib_free_mr(void *trans_private
, int invalidate
)
554 struct rds_ib_mr
*ibmr
= trans_private
;
555 struct rds_ib_device
*rds_ibdev
= ibmr
->device
;
556 struct rds_ib_mr_pool
*pool
= rds_ibdev
->mr_pool
;
559 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr
->sg_len
);
561 /* Return it to the pool's free list */
562 spin_lock_irqsave(&pool
->list_lock
, flags
);
563 if (ibmr
->remap_count
>= pool
->fmr_attr
.max_maps
)
564 list_add(&ibmr
->list
, &pool
->drop_list
);
566 list_add(&ibmr
->list
, &pool
->free_list
);
568 atomic_add(ibmr
->sg_len
, &pool
->free_pinned
);
569 atomic_inc(&pool
->dirty_count
);
570 spin_unlock_irqrestore(&pool
->list_lock
, flags
);
572 /* If we've pinned too many pages, request a flush */
573 if (atomic_read(&pool
->free_pinned
) >= pool
->max_free_pinned
||
574 atomic_read(&pool
->dirty_count
) >= pool
->max_items
/ 10)
575 queue_work(rds_wq
, &pool
->flush_worker
);
578 if (likely(!in_interrupt())) {
579 rds_ib_flush_mr_pool(pool
, 0);
581 /* We get here if the user created a MR marked
582 * as use_once and invalidate at the same time. */
583 queue_work(rds_wq
, &pool
->flush_worker
);
588 void rds_ib_flush_mrs(void)
590 struct rds_ib_device
*rds_ibdev
;
592 list_for_each_entry(rds_ibdev
, &rds_ib_devices
, list
) {
593 struct rds_ib_mr_pool
*pool
= rds_ibdev
->mr_pool
;
596 rds_ib_flush_mr_pool(pool
, 0);
600 void *rds_ib_get_mr(struct scatterlist
*sg
, unsigned long nents
,
601 struct rds_sock
*rs
, u32
*key_ret
)
603 struct rds_ib_device
*rds_ibdev
;
604 struct rds_ib_mr
*ibmr
= NULL
;
607 rds_ibdev
= rds_ib_get_device(rs
->rs_bound_addr
);
613 if (!rds_ibdev
->mr_pool
) {
618 ibmr
= rds_ib_alloc_fmr(rds_ibdev
);
622 ret
= rds_ib_map_fmr(rds_ibdev
, ibmr
, sg
, nents
);
624 *key_ret
= ibmr
->fmr
->rkey
;
626 printk(KERN_WARNING
"RDS/IB: map_fmr failed (errno=%d)\n", ret
);
628 ibmr
->device
= rds_ibdev
;
633 rds_ib_free_mr(ibmr
, 0);