Loading...
Note: File does not exist in v6.8.
1/*
2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
13 * conditions are met:
14 *
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
18 *
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <linux/errno.h>
35#include <linux/spinlock.h>
36#include <linux/export.h>
37#include <linux/slab.h>
38#include <linux/jhash.h>
39#include <linux/kthread.h>
40
41#include <rdma/ib_fmr_pool.h>
42
43#include "core_priv.h"
44
45#define PFX "fmr_pool: "
46
47enum {
48 IB_FMR_MAX_REMAPS = 32,
49
50 IB_FMR_HASH_BITS = 8,
51 IB_FMR_HASH_SIZE = 1 << IB_FMR_HASH_BITS,
52 IB_FMR_HASH_MASK = IB_FMR_HASH_SIZE - 1
53};
54
55/*
56 * If an FMR is not in use, then the list member will point to either
57 * its pool's free_list (if the FMR can be mapped again; that is,
58 * remap_count < pool->max_remaps) or its pool's dirty_list (if the
59 * FMR needs to be unmapped before being remapped). In either of
60 * these cases it is a bug if the ref_count is not 0. In other words,
61 * if ref_count is > 0, then the list member must not be linked into
62 * either free_list or dirty_list.
63 *
64 * The cache_node member is used to link the FMR into a cache bucket
65 * (if caching is enabled). This is independent of the reference
66 * count of the FMR. When a valid FMR is released, its ref_count is
67 * decremented, and if ref_count reaches 0, the FMR is placed in
68 * either free_list or dirty_list as appropriate. However, it is not
69 * removed from the cache and may be "revived" if a call to
70 * ib_fmr_register_physical() occurs before the FMR is remapped. In
71 * this case we just increment the ref_count and remove the FMR from
72 * free_list/dirty_list.
73 *
74 * Before we remap an FMR from free_list, we remove it from the cache
75 * (to prevent another user from obtaining a stale FMR). When an FMR
76 * is released, we add it to the tail of the free list, so that our
77 * cache eviction policy is "least recently used."
78 *
79 * All manipulation of ref_count, list and cache_node is protected by
80 * pool_lock to maintain consistency.
81 */
82
83struct ib_fmr_pool {
84 spinlock_t pool_lock;
85
86 int pool_size;
87 int max_pages;
88 int max_remaps;
89 int dirty_watermark;
90 int dirty_len;
91 struct list_head free_list;
92 struct list_head dirty_list;
93 struct hlist_head *cache_bucket;
94
95 void (*flush_function)(struct ib_fmr_pool *pool,
96 void * arg);
97 void *flush_arg;
98
99 struct task_struct *thread;
100
101 atomic_t req_ser;
102 atomic_t flush_ser;
103
104 wait_queue_head_t force_wait;
105};
106
107static inline u32 ib_fmr_hash(u64 first_page)
108{
109 return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
110 (IB_FMR_HASH_SIZE - 1);
111}
112
113/* Caller must hold pool_lock */
114static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
115 u64 *page_list,
116 int page_list_len,
117 u64 io_virtual_address)
118{
119 struct hlist_head *bucket;
120 struct ib_pool_fmr *fmr;
121
122 if (!pool->cache_bucket)
123 return NULL;
124
125 bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
126
127 hlist_for_each_entry(fmr, bucket, cache_node)
128 if (io_virtual_address == fmr->io_virtual_address &&
129 page_list_len == fmr->page_list_len &&
130 !memcmp(page_list, fmr->page_list,
131 page_list_len * sizeof *page_list))
132 return fmr;
133
134 return NULL;
135}
136
137static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
138{
139 int ret;
140 struct ib_pool_fmr *fmr;
141 LIST_HEAD(unmap_list);
142 LIST_HEAD(fmr_list);
143
144 spin_lock_irq(&pool->pool_lock);
145
146 list_for_each_entry(fmr, &pool->dirty_list, list) {
147 hlist_del_init(&fmr->cache_node);
148 fmr->remap_count = 0;
149 list_add_tail(&fmr->fmr->list, &fmr_list);
150
151#ifdef DEBUG
152 if (fmr->ref_count !=0) {
153 pr_warn(PFX "Unmapping FMR 0x%08x with ref count %d\n",
154 fmr, fmr->ref_count);
155 }
156#endif
157 }
158
159 list_splice_init(&pool->dirty_list, &unmap_list);
160 pool->dirty_len = 0;
161
162 spin_unlock_irq(&pool->pool_lock);
163
164 if (list_empty(&unmap_list)) {
165 return;
166 }
167
168 ret = ib_unmap_fmr(&fmr_list);
169 if (ret)
170 pr_warn(PFX "ib_unmap_fmr returned %d\n", ret);
171
172 spin_lock_irq(&pool->pool_lock);
173 list_splice(&unmap_list, &pool->free_list);
174 spin_unlock_irq(&pool->pool_lock);
175}
176
177static int ib_fmr_cleanup_thread(void *pool_ptr)
178{
179 struct ib_fmr_pool *pool = pool_ptr;
180
181 do {
182 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
183 ib_fmr_batch_release(pool);
184
185 atomic_inc(&pool->flush_ser);
186 wake_up_interruptible(&pool->force_wait);
187
188 if (pool->flush_function)
189 pool->flush_function(pool, pool->flush_arg);
190 }
191
192 set_current_state(TASK_INTERRUPTIBLE);
193 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
194 !kthread_should_stop())
195 schedule();
196 __set_current_state(TASK_RUNNING);
197 } while (!kthread_should_stop());
198
199 return 0;
200}
201
202/**
203 * ib_create_fmr_pool - Create an FMR pool
204 * @pd:Protection domain for FMRs
205 * @params:FMR pool parameters
206 *
207 * Create a pool of FMRs. Return value is pointer to new pool or
208 * error code if creation failed.
209 */
210struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd *pd,
211 struct ib_fmr_pool_param *params)
212{
213 struct ib_device *device;
214 struct ib_fmr_pool *pool;
215 int i;
216 int ret;
217 int max_remaps;
218
219 if (!params)
220 return ERR_PTR(-EINVAL);
221
222 device = pd->device;
223 if (!device->alloc_fmr || !device->dealloc_fmr ||
224 !device->map_phys_fmr || !device->unmap_fmr) {
225 pr_info(PFX "Device %s does not support FMRs\n", device->name);
226 return ERR_PTR(-ENOSYS);
227 }
228
229 if (!device->attrs.max_map_per_fmr)
230 max_remaps = IB_FMR_MAX_REMAPS;
231 else
232 max_remaps = device->attrs.max_map_per_fmr;
233
234 pool = kmalloc(sizeof *pool, GFP_KERNEL);
235 if (!pool)
236 return ERR_PTR(-ENOMEM);
237
238 pool->cache_bucket = NULL;
239 pool->flush_function = params->flush_function;
240 pool->flush_arg = params->flush_arg;
241
242 INIT_LIST_HEAD(&pool->free_list);
243 INIT_LIST_HEAD(&pool->dirty_list);
244
245 if (params->cache) {
246 pool->cache_bucket =
247 kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
248 GFP_KERNEL);
249 if (!pool->cache_bucket) {
250 ret = -ENOMEM;
251 goto out_free_pool;
252 }
253
254 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
255 INIT_HLIST_HEAD(pool->cache_bucket + i);
256 }
257
258 pool->pool_size = 0;
259 pool->max_pages = params->max_pages_per_fmr;
260 pool->max_remaps = max_remaps;
261 pool->dirty_watermark = params->dirty_watermark;
262 pool->dirty_len = 0;
263 spin_lock_init(&pool->pool_lock);
264 atomic_set(&pool->req_ser, 0);
265 atomic_set(&pool->flush_ser, 0);
266 init_waitqueue_head(&pool->force_wait);
267
268 pool->thread = kthread_run(ib_fmr_cleanup_thread,
269 pool,
270 "ib_fmr(%s)",
271 device->name);
272 if (IS_ERR(pool->thread)) {
273 pr_warn(PFX "couldn't start cleanup thread\n");
274 ret = PTR_ERR(pool->thread);
275 goto out_free_pool;
276 }
277
278 {
279 struct ib_pool_fmr *fmr;
280 struct ib_fmr_attr fmr_attr = {
281 .max_pages = params->max_pages_per_fmr,
282 .max_maps = pool->max_remaps,
283 .page_shift = params->page_shift
284 };
285 int bytes_per_fmr = sizeof *fmr;
286
287 if (pool->cache_bucket)
288 bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
289
290 for (i = 0; i < params->pool_size; ++i) {
291 fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
292 if (!fmr)
293 goto out_fail;
294
295 fmr->pool = pool;
296 fmr->remap_count = 0;
297 fmr->ref_count = 0;
298 INIT_HLIST_NODE(&fmr->cache_node);
299
300 fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
301 if (IS_ERR(fmr->fmr)) {
302 pr_warn(PFX "fmr_create failed for FMR %d\n",
303 i);
304 kfree(fmr);
305 goto out_fail;
306 }
307
308 list_add_tail(&fmr->list, &pool->free_list);
309 ++pool->pool_size;
310 }
311 }
312
313 return pool;
314
315 out_free_pool:
316 kfree(pool->cache_bucket);
317 kfree(pool);
318
319 return ERR_PTR(ret);
320
321 out_fail:
322 ib_destroy_fmr_pool(pool);
323
324 return ERR_PTR(-ENOMEM);
325}
326EXPORT_SYMBOL(ib_create_fmr_pool);
327
328/**
329 * ib_destroy_fmr_pool - Free FMR pool
330 * @pool:FMR pool to free
331 *
332 * Destroy an FMR pool and free all associated resources.
333 */
334void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
335{
336 struct ib_pool_fmr *fmr;
337 struct ib_pool_fmr *tmp;
338 LIST_HEAD(fmr_list);
339 int i;
340
341 kthread_stop(pool->thread);
342 ib_fmr_batch_release(pool);
343
344 i = 0;
345 list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
346 if (fmr->remap_count) {
347 INIT_LIST_HEAD(&fmr_list);
348 list_add_tail(&fmr->fmr->list, &fmr_list);
349 ib_unmap_fmr(&fmr_list);
350 }
351 ib_dealloc_fmr(fmr->fmr);
352 list_del(&fmr->list);
353 kfree(fmr);
354 ++i;
355 }
356
357 if (i < pool->pool_size)
358 pr_warn(PFX "pool still has %d regions registered\n",
359 pool->pool_size - i);
360
361 kfree(pool->cache_bucket);
362 kfree(pool);
363}
364EXPORT_SYMBOL(ib_destroy_fmr_pool);
365
366/**
367 * ib_flush_fmr_pool - Invalidate all unmapped FMRs
368 * @pool:FMR pool to flush
369 *
370 * Ensure that all unmapped FMRs are fully invalidated.
371 */
372int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
373{
374 int serial;
375 struct ib_pool_fmr *fmr, *next;
376
377 /*
378 * The free_list holds FMRs that may have been used
379 * but have not been remapped enough times to be dirty.
380 * Put them on the dirty list now so that the cleanup
381 * thread will reap them too.
382 */
383 spin_lock_irq(&pool->pool_lock);
384 list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
385 if (fmr->remap_count > 0)
386 list_move(&fmr->list, &pool->dirty_list);
387 }
388 spin_unlock_irq(&pool->pool_lock);
389
390 serial = atomic_inc_return(&pool->req_ser);
391 wake_up_process(pool->thread);
392
393 if (wait_event_interruptible(pool->force_wait,
394 atomic_read(&pool->flush_ser) - serial >= 0))
395 return -EINTR;
396
397 return 0;
398}
399EXPORT_SYMBOL(ib_flush_fmr_pool);
400
401/**
402 * ib_fmr_pool_map_phys -
403 * @pool:FMR pool to allocate FMR from
404 * @page_list:List of pages to map
405 * @list_len:Number of pages in @page_list
406 * @io_virtual_address:I/O virtual address for new FMR
407 *
408 * Map an FMR from an FMR pool.
409 */
410struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
411 u64 *page_list,
412 int list_len,
413 u64 io_virtual_address)
414{
415 struct ib_fmr_pool *pool = pool_handle;
416 struct ib_pool_fmr *fmr;
417 unsigned long flags;
418 int result;
419
420 if (list_len < 1 || list_len > pool->max_pages)
421 return ERR_PTR(-EINVAL);
422
423 spin_lock_irqsave(&pool->pool_lock, flags);
424 fmr = ib_fmr_cache_lookup(pool,
425 page_list,
426 list_len,
427 io_virtual_address);
428 if (fmr) {
429 /* found in cache */
430 ++fmr->ref_count;
431 if (fmr->ref_count == 1) {
432 list_del(&fmr->list);
433 }
434
435 spin_unlock_irqrestore(&pool->pool_lock, flags);
436
437 return fmr;
438 }
439
440 if (list_empty(&pool->free_list)) {
441 spin_unlock_irqrestore(&pool->pool_lock, flags);
442 return ERR_PTR(-EAGAIN);
443 }
444
445 fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
446 list_del(&fmr->list);
447 hlist_del_init(&fmr->cache_node);
448 spin_unlock_irqrestore(&pool->pool_lock, flags);
449
450 result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
451 io_virtual_address);
452
453 if (result) {
454 spin_lock_irqsave(&pool->pool_lock, flags);
455 list_add(&fmr->list, &pool->free_list);
456 spin_unlock_irqrestore(&pool->pool_lock, flags);
457
458 pr_warn(PFX "fmr_map returns %d\n", result);
459
460 return ERR_PTR(result);
461 }
462
463 ++fmr->remap_count;
464 fmr->ref_count = 1;
465
466 if (pool->cache_bucket) {
467 fmr->io_virtual_address = io_virtual_address;
468 fmr->page_list_len = list_len;
469 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
470
471 spin_lock_irqsave(&pool->pool_lock, flags);
472 hlist_add_head(&fmr->cache_node,
473 pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
474 spin_unlock_irqrestore(&pool->pool_lock, flags);
475 }
476
477 return fmr;
478}
479EXPORT_SYMBOL(ib_fmr_pool_map_phys);
480
481/**
482 * ib_fmr_pool_unmap - Unmap FMR
483 * @fmr:FMR to unmap
484 *
485 * Unmap an FMR. The FMR mapping may remain valid until the FMR is
486 * reused (or until ib_flush_fmr_pool() is called).
487 */
488int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
489{
490 struct ib_fmr_pool *pool;
491 unsigned long flags;
492
493 pool = fmr->pool;
494
495 spin_lock_irqsave(&pool->pool_lock, flags);
496
497 --fmr->ref_count;
498 if (!fmr->ref_count) {
499 if (fmr->remap_count < pool->max_remaps) {
500 list_add_tail(&fmr->list, &pool->free_list);
501 } else {
502 list_add_tail(&fmr->list, &pool->dirty_list);
503 if (++pool->dirty_len >= pool->dirty_watermark) {
504 atomic_inc(&pool->req_ser);
505 wake_up_process(pool->thread);
506 }
507 }
508 }
509
510#ifdef DEBUG
511 if (fmr->ref_count < 0)
512 pr_warn(PFX "FMR %p has ref count %d < 0\n",
513 fmr, fmr->ref_count);
514#endif
515
516 spin_unlock_irqrestore(&pool->pool_lock, flags);
517
518 return 0;
519}
520EXPORT_SYMBOL(ib_fmr_pool_unmap);