Commit | Line | Data |
---|---|---|
08b48a1e AG |
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> | |
5a0e3ad6 | 34 | #include <linux/slab.h> |
764f2dd9 | 35 | #include <linux/rculist.h> |
1bc144b6 | 36 | #include <linux/llist.h> |
08b48a1e AG |
37 | |
38 | #include "rds.h" | |
08b48a1e AG |
39 | #include "ib.h" |
40 | ||
6fa70da6 CM |
41 | static DEFINE_PER_CPU(unsigned long, clean_list_grace); |
42 | #define CLEAN_LIST_BUSY_BIT 0 | |
08b48a1e AG |
43 | |
44 | /* | |
45 | * This is stored as mr->r_trans_private. | |
46 | */ | |
47 | struct rds_ib_mr { | |
48 | struct rds_ib_device *device; | |
49 | struct rds_ib_mr_pool *pool; | |
50 | struct ib_fmr *fmr; | |
6fa70da6 | 51 | |
1bc144b6 | 52 | struct llist_node llnode; |
6fa70da6 CM |
53 | |
54 | /* unmap_list is for freeing */ | |
55 | struct list_head unmap_list; | |
08b48a1e AG |
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 | */ | |
67 | struct rds_ib_mr_pool { | |
68 | struct mutex flush_lock; /* serialize fmr invalidate */ | |
7a0ff5db | 69 | struct delayed_work flush_worker; /* flush worker */ |
08b48a1e | 70 | |
08b48a1e AG |
71 | atomic_t item_count; /* total # of MRs */ |
72 | atomic_t dirty_count; /* # dirty of MRs */ | |
6fa70da6 | 73 | |
1bc144b6 HY |
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 */ | |
6fa70da6 CM |
77 | wait_queue_head_t flush_wait; |
78 | ||
08b48a1e AG |
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 | ||
6fa70da6 | 86 | static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **); |
08b48a1e AG |
87 | static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr); |
88 | static void rds_ib_mr_pool_flush_worker(struct work_struct *work); | |
89 | ||
90 | static 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 | ||
ea819867 ZB |
95 | rcu_read_lock(); |
96 | list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) { | |
764f2dd9 | 97 | list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { |
08b48a1e | 98 | if (i_ipaddr->ipaddr == ipaddr) { |
3e0249f9 | 99 | atomic_inc(&rds_ibdev->refcount); |
764f2dd9 | 100 | rcu_read_unlock(); |
08b48a1e AG |
101 | return rds_ibdev; |
102 | } | |
103 | } | |
08b48a1e | 104 | } |
ea819867 | 105 | rcu_read_unlock(); |
08b48a1e AG |
106 | |
107 | return NULL; | |
108 | } | |
109 | ||
110 | static 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); | |
764f2dd9 | 121 | list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list); |
08b48a1e AG |
122 | spin_unlock_irq(&rds_ibdev->spinlock); |
123 | ||
124 | return 0; | |
125 | } | |
126 | ||
127 | static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) | |
128 | { | |
4a81802b | 129 | struct rds_ib_ipaddr *i_ipaddr; |
764f2dd9 CM |
130 | struct rds_ib_ipaddr *to_free = NULL; |
131 | ||
08b48a1e AG |
132 | |
133 | spin_lock_irq(&rds_ibdev->spinlock); | |
764f2dd9 | 134 | list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { |
08b48a1e | 135 | if (i_ipaddr->ipaddr == ipaddr) { |
764f2dd9 CM |
136 | list_del_rcu(&i_ipaddr->list); |
137 | to_free = i_ipaddr; | |
08b48a1e AG |
138 | break; |
139 | } | |
140 | } | |
141 | spin_unlock_irq(&rds_ibdev->spinlock); | |
764f2dd9 CM |
142 | |
143 | if (to_free) { | |
144 | synchronize_rcu(); | |
145 | kfree(to_free); | |
146 | } | |
08b48a1e AG |
147 | } |
148 | ||
149 | int 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); | |
e1f475a7 | 154 | if (!rds_ibdev_old) |
155 | return rds_ib_add_ipaddr(rds_ibdev, ipaddr); | |
156 | ||
157 | if (rds_ibdev_old != rds_ibdev) { | |
08b48a1e | 158 | rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr); |
3e0249f9 | 159 | rds_ib_dev_put(rds_ibdev_old); |
e1f475a7 | 160 | return rds_ib_add_ipaddr(rds_ibdev, ipaddr); |
3e0249f9 | 161 | } |
e1f475a7 | 162 | rds_ib_dev_put(rds_ibdev_old); |
08b48a1e | 163 | |
e1f475a7 | 164 | return 0; |
08b48a1e AG |
165 | } |
166 | ||
745cbcca | 167 | void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) |
08b48a1e AG |
168 | { |
169 | struct rds_ib_connection *ic = conn->c_transport_data; | |
170 | ||
171 | /* conn was previously on the nodev_conns_list */ | |
172 | spin_lock_irq(&ib_nodev_conns_lock); | |
173 | BUG_ON(list_empty(&ib_nodev_conns)); | |
174 | BUG_ON(list_empty(&ic->ib_node)); | |
175 | list_del(&ic->ib_node); | |
08b48a1e | 176 | |
aef3ea33 | 177 | spin_lock(&rds_ibdev->spinlock); |
08b48a1e | 178 | list_add_tail(&ic->ib_node, &rds_ibdev->conn_list); |
aef3ea33 | 179 | spin_unlock(&rds_ibdev->spinlock); |
745cbcca | 180 | spin_unlock_irq(&ib_nodev_conns_lock); |
08b48a1e AG |
181 | |
182 | ic->rds_ibdev = rds_ibdev; | |
3e0249f9 | 183 | atomic_inc(&rds_ibdev->refcount); |
08b48a1e AG |
184 | } |
185 | ||
745cbcca | 186 | void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) |
08b48a1e | 187 | { |
745cbcca | 188 | struct rds_ib_connection *ic = conn->c_transport_data; |
08b48a1e | 189 | |
745cbcca AG |
190 | /* place conn on nodev_conns_list */ |
191 | spin_lock(&ib_nodev_conns_lock); | |
08b48a1e | 192 | |
745cbcca AG |
193 | spin_lock_irq(&rds_ibdev->spinlock); |
194 | BUG_ON(list_empty(&ic->ib_node)); | |
195 | list_del(&ic->ib_node); | |
196 | spin_unlock_irq(&rds_ibdev->spinlock); | |
197 | ||
198 | list_add_tail(&ic->ib_node, &ib_nodev_conns); | |
199 | ||
200 | spin_unlock(&ib_nodev_conns_lock); | |
201 | ||
202 | ic->rds_ibdev = NULL; | |
3e0249f9 | 203 | rds_ib_dev_put(rds_ibdev); |
08b48a1e AG |
204 | } |
205 | ||
8aeb1ba6 | 206 | void rds_ib_destroy_nodev_conns(void) |
08b48a1e AG |
207 | { |
208 | struct rds_ib_connection *ic, *_ic; | |
209 | LIST_HEAD(tmp_list); | |
210 | ||
211 | /* avoid calling conn_destroy with irqs off */ | |
8aeb1ba6 ZB |
212 | spin_lock_irq(&ib_nodev_conns_lock); |
213 | list_splice(&ib_nodev_conns, &tmp_list); | |
214 | spin_unlock_irq(&ib_nodev_conns_lock); | |
08b48a1e | 215 | |
433d308d | 216 | list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) |
08b48a1e | 217 | rds_conn_destroy(ic->conn); |
08b48a1e AG |
218 | } |
219 | ||
220 | struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev) | |
221 | { | |
222 | struct rds_ib_mr_pool *pool; | |
223 | ||
224 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); | |
225 | if (!pool) | |
226 | return ERR_PTR(-ENOMEM); | |
227 | ||
1bc144b6 HY |
228 | init_llist_head(&pool->free_list); |
229 | init_llist_head(&pool->drop_list); | |
230 | init_llist_head(&pool->clean_list); | |
08b48a1e | 231 | mutex_init(&pool->flush_lock); |
6fa70da6 | 232 | init_waitqueue_head(&pool->flush_wait); |
7a0ff5db | 233 | INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker); |
08b48a1e AG |
234 | |
235 | pool->fmr_attr.max_pages = fmr_message_size; | |
236 | pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps; | |
a870d627 | 237 | pool->fmr_attr.page_shift = PAGE_SHIFT; |
08b48a1e AG |
238 | pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4; |
239 | ||
240 | /* We never allow more than max_items MRs to be allocated. | |
241 | * When we exceed more than max_items_soft, we start freeing | |
242 | * items more aggressively. | |
243 | * Make sure that max_items > max_items_soft > max_items / 2 | |
244 | */ | |
245 | pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4; | |
246 | pool->max_items = rds_ibdev->max_fmrs; | |
247 | ||
248 | return pool; | |
249 | } | |
250 | ||
251 | void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo) | |
252 | { | |
253 | struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; | |
254 | ||
255 | iinfo->rdma_mr_max = pool->max_items; | |
256 | iinfo->rdma_mr_size = pool->fmr_attr.max_pages; | |
257 | } | |
258 | ||
259 | void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool) | |
260 | { | |
7a0ff5db | 261 | cancel_delayed_work_sync(&pool->flush_worker); |
6fa70da6 | 262 | rds_ib_flush_mr_pool(pool, 1, NULL); |
571c02fa AG |
263 | WARN_ON(atomic_read(&pool->item_count)); |
264 | WARN_ON(atomic_read(&pool->free_pinned)); | |
08b48a1e AG |
265 | kfree(pool); |
266 | } | |
267 | ||
268 | static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool) | |
269 | { | |
270 | struct rds_ib_mr *ibmr = NULL; | |
1bc144b6 | 271 | struct llist_node *ret; |
6fa70da6 | 272 | unsigned long *flag; |
08b48a1e | 273 | |
6fa70da6 | 274 | preempt_disable(); |
903ceff7 | 275 | flag = this_cpu_ptr(&clean_list_grace); |
6fa70da6 | 276 | set_bit(CLEAN_LIST_BUSY_BIT, flag); |
1bc144b6 | 277 | ret = llist_del_first(&pool->clean_list); |
6fa70da6 | 278 | if (ret) |
1bc144b6 | 279 | ibmr = llist_entry(ret, struct rds_ib_mr, llnode); |
08b48a1e | 280 | |
6fa70da6 CM |
281 | clear_bit(CLEAN_LIST_BUSY_BIT, flag); |
282 | preempt_enable(); | |
08b48a1e AG |
283 | return ibmr; |
284 | } | |
285 | ||
6fa70da6 CM |
286 | static inline void wait_clean_list_grace(void) |
287 | { | |
288 | int cpu; | |
289 | unsigned long *flag; | |
290 | ||
291 | for_each_online_cpu(cpu) { | |
292 | flag = &per_cpu(clean_list_grace, cpu); | |
293 | while (test_bit(CLEAN_LIST_BUSY_BIT, flag)) | |
294 | cpu_relax(); | |
295 | } | |
296 | } | |
297 | ||
08b48a1e AG |
298 | static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev) |
299 | { | |
300 | struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; | |
301 | struct rds_ib_mr *ibmr = NULL; | |
302 | int err = 0, iter = 0; | |
303 | ||
8576f374 | 304 | if (atomic_read(&pool->dirty_count) >= pool->max_items / 10) |
c534a107 | 305 | schedule_delayed_work(&pool->flush_worker, 10); |
8576f374 | 306 | |
08b48a1e AG |
307 | while (1) { |
308 | ibmr = rds_ib_reuse_fmr(pool); | |
309 | if (ibmr) | |
310 | return ibmr; | |
311 | ||
312 | /* No clean MRs - now we have the choice of either | |
313 | * allocating a fresh MR up to the limit imposed by the | |
314 | * driver, or flush any dirty unused MRs. | |
315 | * We try to avoid stalling in the send path if possible, | |
316 | * so we allocate as long as we're allowed to. | |
317 | * | |
318 | * We're fussy with enforcing the FMR limit, though. If the driver | |
319 | * tells us we can't use more than N fmrs, we shouldn't start | |
320 | * arguing with it */ | |
321 | if (atomic_inc_return(&pool->item_count) <= pool->max_items) | |
322 | break; | |
323 | ||
324 | atomic_dec(&pool->item_count); | |
325 | ||
326 | if (++iter > 2) { | |
327 | rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted); | |
328 | return ERR_PTR(-EAGAIN); | |
329 | } | |
330 | ||
331 | /* We do have some empty MRs. Flush them out. */ | |
332 | rds_ib_stats_inc(s_ib_rdma_mr_pool_wait); | |
6fa70da6 CM |
333 | rds_ib_flush_mr_pool(pool, 0, &ibmr); |
334 | if (ibmr) | |
335 | return ibmr; | |
08b48a1e AG |
336 | } |
337 | ||
e4c52c98 | 338 | ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev)); |
08b48a1e AG |
339 | if (!ibmr) { |
340 | err = -ENOMEM; | |
341 | goto out_no_cigar; | |
342 | } | |
343 | ||
38a4e5e6 CM |
344 | memset(ibmr, 0, sizeof(*ibmr)); |
345 | ||
08b48a1e AG |
346 | ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd, |
347 | (IB_ACCESS_LOCAL_WRITE | | |
348 | IB_ACCESS_REMOTE_READ | | |
15133f6e AG |
349 | IB_ACCESS_REMOTE_WRITE| |
350 | IB_ACCESS_REMOTE_ATOMIC), | |
08b48a1e AG |
351 | &pool->fmr_attr); |
352 | if (IS_ERR(ibmr->fmr)) { | |
353 | err = PTR_ERR(ibmr->fmr); | |
354 | ibmr->fmr = NULL; | |
355 | printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err); | |
356 | goto out_no_cigar; | |
357 | } | |
358 | ||
359 | rds_ib_stats_inc(s_ib_rdma_mr_alloc); | |
360 | return ibmr; | |
361 | ||
362 | out_no_cigar: | |
363 | if (ibmr) { | |
364 | if (ibmr->fmr) | |
365 | ib_dealloc_fmr(ibmr->fmr); | |
366 | kfree(ibmr); | |
367 | } | |
368 | atomic_dec(&pool->item_count); | |
369 | return ERR_PTR(err); | |
370 | } | |
371 | ||
372 | static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr, | |
373 | struct scatterlist *sg, unsigned int nents) | |
374 | { | |
375 | struct ib_device *dev = rds_ibdev->dev; | |
376 | struct scatterlist *scat = sg; | |
377 | u64 io_addr = 0; | |
378 | u64 *dma_pages; | |
379 | u32 len; | |
380 | int page_cnt, sg_dma_len; | |
381 | int i, j; | |
382 | int ret; | |
383 | ||
384 | sg_dma_len = ib_dma_map_sg(dev, sg, nents, | |
385 | DMA_BIDIRECTIONAL); | |
386 | if (unlikely(!sg_dma_len)) { | |
387 | printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n"); | |
388 | return -EBUSY; | |
389 | } | |
390 | ||
391 | len = 0; | |
392 | page_cnt = 0; | |
393 | ||
394 | for (i = 0; i < sg_dma_len; ++i) { | |
395 | unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]); | |
396 | u64 dma_addr = ib_sg_dma_address(dev, &scat[i]); | |
397 | ||
a870d627 | 398 | if (dma_addr & ~PAGE_MASK) { |
08b48a1e AG |
399 | if (i > 0) |
400 | return -EINVAL; | |
401 | else | |
402 | ++page_cnt; | |
403 | } | |
a870d627 | 404 | if ((dma_addr + dma_len) & ~PAGE_MASK) { |
08b48a1e AG |
405 | if (i < sg_dma_len - 1) |
406 | return -EINVAL; | |
407 | else | |
408 | ++page_cnt; | |
409 | } | |
410 | ||
411 | len += dma_len; | |
412 | } | |
413 | ||
a870d627 | 414 | page_cnt += len >> PAGE_SHIFT; |
08b48a1e AG |
415 | if (page_cnt > fmr_message_size) |
416 | return -EINVAL; | |
417 | ||
e4c52c98 AG |
418 | dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC, |
419 | rdsibdev_to_node(rds_ibdev)); | |
08b48a1e AG |
420 | if (!dma_pages) |
421 | return -ENOMEM; | |
422 | ||
423 | page_cnt = 0; | |
424 | for (i = 0; i < sg_dma_len; ++i) { | |
425 | unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]); | |
426 | u64 dma_addr = ib_sg_dma_address(dev, &scat[i]); | |
427 | ||
a870d627 | 428 | for (j = 0; j < dma_len; j += PAGE_SIZE) |
08b48a1e | 429 | dma_pages[page_cnt++] = |
a870d627 | 430 | (dma_addr & PAGE_MASK) + j; |
08b48a1e AG |
431 | } |
432 | ||
433 | ret = ib_map_phys_fmr(ibmr->fmr, | |
434 | dma_pages, page_cnt, io_addr); | |
435 | if (ret) | |
436 | goto out; | |
437 | ||
438 | /* Success - we successfully remapped the MR, so we can | |
439 | * safely tear down the old mapping. */ | |
440 | rds_ib_teardown_mr(ibmr); | |
441 | ||
442 | ibmr->sg = scat; | |
443 | ibmr->sg_len = nents; | |
444 | ibmr->sg_dma_len = sg_dma_len; | |
445 | ibmr->remap_count++; | |
446 | ||
447 | rds_ib_stats_inc(s_ib_rdma_mr_used); | |
448 | ret = 0; | |
449 | ||
450 | out: | |
451 | kfree(dma_pages); | |
452 | ||
453 | return ret; | |
454 | } | |
455 | ||
456 | void rds_ib_sync_mr(void *trans_private, int direction) | |
457 | { | |
458 | struct rds_ib_mr *ibmr = trans_private; | |
459 | struct rds_ib_device *rds_ibdev = ibmr->device; | |
460 | ||
461 | switch (direction) { | |
462 | case DMA_FROM_DEVICE: | |
463 | ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg, | |
464 | ibmr->sg_dma_len, DMA_BIDIRECTIONAL); | |
465 | break; | |
466 | case DMA_TO_DEVICE: | |
467 | ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg, | |
468 | ibmr->sg_dma_len, DMA_BIDIRECTIONAL); | |
469 | break; | |
470 | } | |
471 | } | |
472 | ||
473 | static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr) | |
474 | { | |
475 | struct rds_ib_device *rds_ibdev = ibmr->device; | |
476 | ||
477 | if (ibmr->sg_dma_len) { | |
478 | ib_dma_unmap_sg(rds_ibdev->dev, | |
479 | ibmr->sg, ibmr->sg_len, | |
480 | DMA_BIDIRECTIONAL); | |
481 | ibmr->sg_dma_len = 0; | |
482 | } | |
483 | ||
484 | /* Release the s/g list */ | |
485 | if (ibmr->sg_len) { | |
486 | unsigned int i; | |
487 | ||
488 | for (i = 0; i < ibmr->sg_len; ++i) { | |
489 | struct page *page = sg_page(&ibmr->sg[i]); | |
490 | ||
491 | /* FIXME we need a way to tell a r/w MR | |
492 | * from a r/o MR */ | |
9e2effba | 493 | BUG_ON(irqs_disabled()); |
08b48a1e AG |
494 | set_page_dirty(page); |
495 | put_page(page); | |
496 | } | |
497 | kfree(ibmr->sg); | |
498 | ||
499 | ibmr->sg = NULL; | |
500 | ibmr->sg_len = 0; | |
501 | } | |
502 | } | |
503 | ||
504 | static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr) | |
505 | { | |
506 | unsigned int pinned = ibmr->sg_len; | |
507 | ||
508 | __rds_ib_teardown_mr(ibmr); | |
509 | if (pinned) { | |
510 | struct rds_ib_device *rds_ibdev = ibmr->device; | |
511 | struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; | |
512 | ||
513 | atomic_sub(pinned, &pool->free_pinned); | |
514 | } | |
515 | } | |
516 | ||
517 | static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all) | |
518 | { | |
519 | unsigned int item_count; | |
520 | ||
521 | item_count = atomic_read(&pool->item_count); | |
522 | if (free_all) | |
523 | return item_count; | |
524 | ||
525 | return 0; | |
526 | } | |
527 | ||
6fa70da6 | 528 | /* |
1bc144b6 | 529 | * given an llist of mrs, put them all into the list_head for more processing |
6fa70da6 | 530 | */ |
1bc144b6 | 531 | static void llist_append_to_list(struct llist_head *llist, struct list_head *list) |
6fa70da6 CM |
532 | { |
533 | struct rds_ib_mr *ibmr; | |
1bc144b6 HY |
534 | struct llist_node *node; |
535 | struct llist_node *next; | |
536 | ||
537 | node = llist_del_all(llist); | |
538 | while (node) { | |
539 | next = node->next; | |
540 | ibmr = llist_entry(node, struct rds_ib_mr, llnode); | |
6fa70da6 | 541 | list_add_tail(&ibmr->unmap_list, list); |
1bc144b6 | 542 | node = next; |
6fa70da6 CM |
543 | } |
544 | } | |
545 | ||
546 | /* | |
1bc144b6 HY |
547 | * this takes a list head of mrs and turns it into linked llist nodes |
548 | * of clusters. Each cluster has linked llist nodes of | |
549 | * MR_CLUSTER_SIZE mrs that are ready for reuse. | |
6fa70da6 | 550 | */ |
1bc144b6 HY |
551 | static void list_to_llist_nodes(struct rds_ib_mr_pool *pool, |
552 | struct list_head *list, | |
553 | struct llist_node **nodes_head, | |
554 | struct llist_node **nodes_tail) | |
6fa70da6 CM |
555 | { |
556 | struct rds_ib_mr *ibmr; | |
1bc144b6 HY |
557 | struct llist_node *cur = NULL; |
558 | struct llist_node **next = nodes_head; | |
6fa70da6 CM |
559 | |
560 | list_for_each_entry(ibmr, list, unmap_list) { | |
1bc144b6 HY |
561 | cur = &ibmr->llnode; |
562 | *next = cur; | |
563 | next = &cur->next; | |
6fa70da6 | 564 | } |
1bc144b6 HY |
565 | *next = NULL; |
566 | *nodes_tail = cur; | |
6fa70da6 CM |
567 | } |
568 | ||
08b48a1e AG |
569 | /* |
570 | * Flush our pool of MRs. | |
571 | * At a minimum, all currently unused MRs are unmapped. | |
572 | * If the number of MRs allocated exceeds the limit, we also try | |
573 | * to free as many MRs as needed to get back to this limit. | |
574 | */ | |
6fa70da6 CM |
575 | static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, |
576 | int free_all, struct rds_ib_mr **ibmr_ret) | |
08b48a1e AG |
577 | { |
578 | struct rds_ib_mr *ibmr, *next; | |
1bc144b6 HY |
579 | struct llist_node *clean_nodes; |
580 | struct llist_node *clean_tail; | |
08b48a1e AG |
581 | LIST_HEAD(unmap_list); |
582 | LIST_HEAD(fmr_list); | |
583 | unsigned long unpinned = 0; | |
08b48a1e AG |
584 | unsigned int nfreed = 0, ncleaned = 0, free_goal; |
585 | int ret = 0; | |
586 | ||
587 | rds_ib_stats_inc(s_ib_rdma_mr_pool_flush); | |
588 | ||
6fa70da6 CM |
589 | if (ibmr_ret) { |
590 | DEFINE_WAIT(wait); | |
591 | while(!mutex_trylock(&pool->flush_lock)) { | |
592 | ibmr = rds_ib_reuse_fmr(pool); | |
593 | if (ibmr) { | |
594 | *ibmr_ret = ibmr; | |
595 | finish_wait(&pool->flush_wait, &wait); | |
596 | goto out_nolock; | |
597 | } | |
598 | ||
599 | prepare_to_wait(&pool->flush_wait, &wait, | |
600 | TASK_UNINTERRUPTIBLE); | |
1bc144b6 | 601 | if (llist_empty(&pool->clean_list)) |
6fa70da6 CM |
602 | schedule(); |
603 | ||
604 | ibmr = rds_ib_reuse_fmr(pool); | |
605 | if (ibmr) { | |
606 | *ibmr_ret = ibmr; | |
607 | finish_wait(&pool->flush_wait, &wait); | |
608 | goto out_nolock; | |
609 | } | |
610 | } | |
611 | finish_wait(&pool->flush_wait, &wait); | |
612 | } else | |
613 | mutex_lock(&pool->flush_lock); | |
614 | ||
615 | if (ibmr_ret) { | |
616 | ibmr = rds_ib_reuse_fmr(pool); | |
617 | if (ibmr) { | |
618 | *ibmr_ret = ibmr; | |
619 | goto out; | |
620 | } | |
621 | } | |
08b48a1e | 622 | |
08b48a1e | 623 | /* Get the list of all MRs to be dropped. Ordering matters - |
6fa70da6 CM |
624 | * we want to put drop_list ahead of free_list. |
625 | */ | |
1bc144b6 HY |
626 | llist_append_to_list(&pool->drop_list, &unmap_list); |
627 | llist_append_to_list(&pool->free_list, &unmap_list); | |
08b48a1e | 628 | if (free_all) |
1bc144b6 | 629 | llist_append_to_list(&pool->clean_list, &unmap_list); |
08b48a1e AG |
630 | |
631 | free_goal = rds_ib_flush_goal(pool, free_all); | |
632 | ||
633 | if (list_empty(&unmap_list)) | |
634 | goto out; | |
635 | ||
636 | /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */ | |
6fa70da6 | 637 | list_for_each_entry(ibmr, &unmap_list, unmap_list) |
08b48a1e | 638 | list_add(&ibmr->fmr->list, &fmr_list); |
6fa70da6 | 639 | |
08b48a1e AG |
640 | ret = ib_unmap_fmr(&fmr_list); |
641 | if (ret) | |
642 | printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret); | |
643 | ||
644 | /* Now we can destroy the DMA mapping and unpin any pages */ | |
6fa70da6 | 645 | list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) { |
08b48a1e AG |
646 | unpinned += ibmr->sg_len; |
647 | __rds_ib_teardown_mr(ibmr); | |
648 | if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) { | |
649 | rds_ib_stats_inc(s_ib_rdma_mr_free); | |
6fa70da6 | 650 | list_del(&ibmr->unmap_list); |
08b48a1e AG |
651 | ib_dealloc_fmr(ibmr->fmr); |
652 | kfree(ibmr); | |
653 | nfreed++; | |
654 | } | |
655 | ncleaned++; | |
656 | } | |
657 | ||
6fa70da6 CM |
658 | if (!list_empty(&unmap_list)) { |
659 | /* we have to make sure that none of the things we're about | |
660 | * to put on the clean list would race with other cpus trying | |
1bc144b6 | 661 | * to pull items off. The llist would explode if we managed to |
6fa70da6 | 662 | * remove something from the clean list and then add it back again |
1bc144b6 | 663 | * while another CPU was spinning on that same item in llist_del_first. |
6fa70da6 | 664 | * |
1bc144b6 | 665 | * This is pretty unlikely, but just in case wait for an llist grace period |
6fa70da6 CM |
666 | * here before adding anything back into the clean list. |
667 | */ | |
668 | wait_clean_list_grace(); | |
669 | ||
1bc144b6 | 670 | list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail); |
6fa70da6 | 671 | if (ibmr_ret) |
1bc144b6 | 672 | *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode); |
6fa70da6 | 673 | |
1bc144b6 HY |
674 | /* more than one entry in llist nodes */ |
675 | if (clean_nodes->next) | |
676 | llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list); | |
6fa70da6 CM |
677 | |
678 | } | |
08b48a1e AG |
679 | |
680 | atomic_sub(unpinned, &pool->free_pinned); | |
681 | atomic_sub(ncleaned, &pool->dirty_count); | |
682 | atomic_sub(nfreed, &pool->item_count); | |
683 | ||
684 | out: | |
685 | mutex_unlock(&pool->flush_lock); | |
6fa70da6 CM |
686 | if (waitqueue_active(&pool->flush_wait)) |
687 | wake_up(&pool->flush_wait); | |
688 | out_nolock: | |
08b48a1e AG |
689 | return ret; |
690 | } | |
691 | ||
692 | static void rds_ib_mr_pool_flush_worker(struct work_struct *work) | |
693 | { | |
7a0ff5db | 694 | struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work); |
08b48a1e | 695 | |
6fa70da6 | 696 | rds_ib_flush_mr_pool(pool, 0, NULL); |
08b48a1e AG |
697 | } |
698 | ||
699 | void rds_ib_free_mr(void *trans_private, int invalidate) | |
700 | { | |
701 | struct rds_ib_mr *ibmr = trans_private; | |
702 | struct rds_ib_device *rds_ibdev = ibmr->device; | |
703 | struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; | |
08b48a1e AG |
704 | |
705 | rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len); | |
706 | ||
707 | /* Return it to the pool's free list */ | |
08b48a1e | 708 | if (ibmr->remap_count >= pool->fmr_attr.max_maps) |
1bc144b6 | 709 | llist_add(&ibmr->llnode, &pool->drop_list); |
08b48a1e | 710 | else |
1bc144b6 | 711 | llist_add(&ibmr->llnode, &pool->free_list); |
08b48a1e AG |
712 | |
713 | atomic_add(ibmr->sg_len, &pool->free_pinned); | |
714 | atomic_inc(&pool->dirty_count); | |
08b48a1e AG |
715 | |
716 | /* If we've pinned too many pages, request a flush */ | |
f64f9e71 JP |
717 | if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || |
718 | atomic_read(&pool->dirty_count) >= pool->max_items / 10) | |
c534a107 | 719 | schedule_delayed_work(&pool->flush_worker, 10); |
08b48a1e AG |
720 | |
721 | if (invalidate) { | |
722 | if (likely(!in_interrupt())) { | |
6fa70da6 | 723 | rds_ib_flush_mr_pool(pool, 0, NULL); |
08b48a1e AG |
724 | } else { |
725 | /* We get here if the user created a MR marked | |
726 | * as use_once and invalidate at the same time. */ | |
c534a107 | 727 | schedule_delayed_work(&pool->flush_worker, 10); |
08b48a1e AG |
728 | } |
729 | } | |
3e0249f9 ZB |
730 | |
731 | rds_ib_dev_put(rds_ibdev); | |
08b48a1e AG |
732 | } |
733 | ||
734 | void rds_ib_flush_mrs(void) | |
735 | { | |
736 | struct rds_ib_device *rds_ibdev; | |
737 | ||
ea819867 | 738 | down_read(&rds_ib_devices_lock); |
08b48a1e AG |
739 | list_for_each_entry(rds_ibdev, &rds_ib_devices, list) { |
740 | struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; | |
741 | ||
742 | if (pool) | |
6fa70da6 | 743 | rds_ib_flush_mr_pool(pool, 0, NULL); |
08b48a1e | 744 | } |
ea819867 | 745 | up_read(&rds_ib_devices_lock); |
08b48a1e AG |
746 | } |
747 | ||
748 | void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, | |
749 | struct rds_sock *rs, u32 *key_ret) | |
750 | { | |
751 | struct rds_ib_device *rds_ibdev; | |
752 | struct rds_ib_mr *ibmr = NULL; | |
753 | int ret; | |
754 | ||
755 | rds_ibdev = rds_ib_get_device(rs->rs_bound_addr); | |
756 | if (!rds_ibdev) { | |
757 | ret = -ENODEV; | |
758 | goto out; | |
759 | } | |
760 | ||
761 | if (!rds_ibdev->mr_pool) { | |
762 | ret = -ENODEV; | |
763 | goto out; | |
764 | } | |
765 | ||
766 | ibmr = rds_ib_alloc_fmr(rds_ibdev); | |
4fabb594 WW |
767 | if (IS_ERR(ibmr)) { |
768 | rds_ib_dev_put(rds_ibdev); | |
08b48a1e | 769 | return ibmr; |
4fabb594 | 770 | } |
08b48a1e AG |
771 | |
772 | ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents); | |
773 | if (ret == 0) | |
774 | *key_ret = ibmr->fmr->rkey; | |
775 | else | |
776 | printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret); | |
777 | ||
778 | ibmr->device = rds_ibdev; | |
3e0249f9 | 779 | rds_ibdev = NULL; |
08b48a1e AG |
780 | |
781 | out: | |
782 | if (ret) { | |
783 | if (ibmr) | |
784 | rds_ib_free_mr(ibmr, 0); | |
785 | ibmr = ERR_PTR(ret); | |
786 | } | |
3e0249f9 ZB |
787 | if (rds_ibdev) |
788 | rds_ib_dev_put(rds_ibdev); | |
08b48a1e AG |
789 | return ibmr; |
790 | } | |
6fa70da6 | 791 |