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
47 enum {
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
83 struct 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 kthread_worker *worker;
100 struct kthread_work work;
101
102 atomic_t req_ser;
103 atomic_t flush_ser;
104
105 wait_queue_head_t force_wait;
106 };
107
ib_fmr_hash(u64 first_page)108 static inline u32 ib_fmr_hash(u64 first_page)
109 {
110 return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
111 (IB_FMR_HASH_SIZE - 1);
112 }
113
114 /* Caller must hold pool_lock */
ib_fmr_cache_lookup(struct ib_fmr_pool * pool,u64 * page_list,int page_list_len,u64 io_virtual_address)115 static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
116 u64 *page_list,
117 int page_list_len,
118 u64 io_virtual_address)
119 {
120 struct hlist_head *bucket;
121 struct ib_pool_fmr *fmr;
122
123 if (!pool->cache_bucket)
124 return NULL;
125
126 bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
127
128 hlist_for_each_entry(fmr, bucket, cache_node)
129 if (io_virtual_address == fmr->io_virtual_address &&
130 page_list_len == fmr->page_list_len &&
131 !memcmp(page_list, fmr->page_list,
132 page_list_len * sizeof *page_list))
133 return fmr;
134
135 return NULL;
136 }
137
ib_fmr_batch_release(struct ib_fmr_pool * pool)138 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
139 {
140 int ret;
141 struct ib_pool_fmr *fmr;
142 LIST_HEAD(unmap_list);
143 LIST_HEAD(fmr_list);
144
145 spin_lock_irq(&pool->pool_lock);
146
147 list_for_each_entry(fmr, &pool->dirty_list, list) {
148 hlist_del_init(&fmr->cache_node);
149 fmr->remap_count = 0;
150 list_add_tail(&fmr->fmr->list, &fmr_list);
151
152 #ifdef DEBUG
153 if (fmr->ref_count !=0) {
154 pr_warn(PFX "Unmapping FMR 0x%08x with ref count %d\n",
155 fmr, fmr->ref_count);
156 }
157 #endif
158 }
159
160 list_splice_init(&pool->dirty_list, &unmap_list);
161 pool->dirty_len = 0;
162
163 spin_unlock_irq(&pool->pool_lock);
164
165 if (list_empty(&unmap_list)) {
166 return;
167 }
168
169 ret = ib_unmap_fmr(&fmr_list);
170 if (ret)
171 pr_warn(PFX "ib_unmap_fmr returned %d\n", ret);
172
173 spin_lock_irq(&pool->pool_lock);
174 list_splice(&unmap_list, &pool->free_list);
175 spin_unlock_irq(&pool->pool_lock);
176 }
177
ib_fmr_cleanup_func(struct kthread_work * work)178 static void ib_fmr_cleanup_func(struct kthread_work *work)
179 {
180 struct ib_fmr_pool *pool = container_of(work, struct ib_fmr_pool, work);
181
182 ib_fmr_batch_release(pool);
183 atomic_inc(&pool->flush_ser);
184 wake_up_interruptible(&pool->force_wait);
185
186 if (pool->flush_function)
187 pool->flush_function(pool, pool->flush_arg);
188
189 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0)
190 kthread_queue_work(pool->worker, &pool->work);
191 }
192
193 /**
194 * ib_create_fmr_pool - Create an FMR pool
195 * @pd:Protection domain for FMRs
196 * @params:FMR pool parameters
197 *
198 * Create a pool of FMRs. Return value is pointer to new pool or
199 * error code if creation failed.
200 */
ib_create_fmr_pool(struct ib_pd * pd,struct ib_fmr_pool_param * params)201 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd *pd,
202 struct ib_fmr_pool_param *params)
203 {
204 struct ib_device *device;
205 struct ib_fmr_pool *pool;
206 int i;
207 int ret;
208 int max_remaps;
209
210 if (!params)
211 return ERR_PTR(-EINVAL);
212
213 device = pd->device;
214 if (!device->alloc_fmr || !device->dealloc_fmr ||
215 !device->map_phys_fmr || !device->unmap_fmr) {
216 pr_info(PFX "Device %s does not support FMRs\n", device->name);
217 return ERR_PTR(-ENOSYS);
218 }
219
220 if (!device->attrs.max_map_per_fmr)
221 max_remaps = IB_FMR_MAX_REMAPS;
222 else
223 max_remaps = device->attrs.max_map_per_fmr;
224
225 pool = kmalloc(sizeof *pool, GFP_KERNEL);
226 if (!pool)
227 return ERR_PTR(-ENOMEM);
228
229 pool->cache_bucket = NULL;
230 pool->flush_function = params->flush_function;
231 pool->flush_arg = params->flush_arg;
232
233 INIT_LIST_HEAD(&pool->free_list);
234 INIT_LIST_HEAD(&pool->dirty_list);
235
236 if (params->cache) {
237 pool->cache_bucket =
238 kmalloc_array(IB_FMR_HASH_SIZE,
239 sizeof(*pool->cache_bucket),
240 GFP_KERNEL);
241 if (!pool->cache_bucket) {
242 ret = -ENOMEM;
243 goto out_free_pool;
244 }
245
246 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
247 INIT_HLIST_HEAD(pool->cache_bucket + i);
248 }
249
250 pool->pool_size = 0;
251 pool->max_pages = params->max_pages_per_fmr;
252 pool->max_remaps = max_remaps;
253 pool->dirty_watermark = params->dirty_watermark;
254 pool->dirty_len = 0;
255 spin_lock_init(&pool->pool_lock);
256 atomic_set(&pool->req_ser, 0);
257 atomic_set(&pool->flush_ser, 0);
258 init_waitqueue_head(&pool->force_wait);
259
260 pool->worker = kthread_create_worker(0, "ib_fmr(%s)", device->name);
261 if (IS_ERR(pool->worker)) {
262 pr_warn(PFX "couldn't start cleanup kthread worker\n");
263 ret = PTR_ERR(pool->worker);
264 goto out_free_pool;
265 }
266 kthread_init_work(&pool->work, ib_fmr_cleanup_func);
267
268 {
269 struct ib_pool_fmr *fmr;
270 struct ib_fmr_attr fmr_attr = {
271 .max_pages = params->max_pages_per_fmr,
272 .max_maps = pool->max_remaps,
273 .page_shift = params->page_shift
274 };
275 int bytes_per_fmr = sizeof *fmr;
276
277 if (pool->cache_bucket)
278 bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
279
280 for (i = 0; i < params->pool_size; ++i) {
281 fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
282 if (!fmr)
283 goto out_fail;
284
285 fmr->pool = pool;
286 fmr->remap_count = 0;
287 fmr->ref_count = 0;
288 INIT_HLIST_NODE(&fmr->cache_node);
289
290 fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
291 if (IS_ERR(fmr->fmr)) {
292 pr_warn(PFX "fmr_create failed for FMR %d\n",
293 i);
294 kfree(fmr);
295 goto out_fail;
296 }
297
298 list_add_tail(&fmr->list, &pool->free_list);
299 ++pool->pool_size;
300 }
301 }
302
303 return pool;
304
305 out_free_pool:
306 kfree(pool->cache_bucket);
307 kfree(pool);
308
309 return ERR_PTR(ret);
310
311 out_fail:
312 ib_destroy_fmr_pool(pool);
313
314 return ERR_PTR(-ENOMEM);
315 }
316 EXPORT_SYMBOL(ib_create_fmr_pool);
317
318 /**
319 * ib_destroy_fmr_pool - Free FMR pool
320 * @pool:FMR pool to free
321 *
322 * Destroy an FMR pool and free all associated resources.
323 */
ib_destroy_fmr_pool(struct ib_fmr_pool * pool)324 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
325 {
326 struct ib_pool_fmr *fmr;
327 struct ib_pool_fmr *tmp;
328 LIST_HEAD(fmr_list);
329 int i;
330
331 kthread_destroy_worker(pool->worker);
332 ib_fmr_batch_release(pool);
333
334 i = 0;
335 list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
336 if (fmr->remap_count) {
337 INIT_LIST_HEAD(&fmr_list);
338 list_add_tail(&fmr->fmr->list, &fmr_list);
339 ib_unmap_fmr(&fmr_list);
340 }
341 ib_dealloc_fmr(fmr->fmr);
342 list_del(&fmr->list);
343 kfree(fmr);
344 ++i;
345 }
346
347 if (i < pool->pool_size)
348 pr_warn(PFX "pool still has %d regions registered\n",
349 pool->pool_size - i);
350
351 kfree(pool->cache_bucket);
352 kfree(pool);
353 }
354 EXPORT_SYMBOL(ib_destroy_fmr_pool);
355
356 /**
357 * ib_flush_fmr_pool - Invalidate all unmapped FMRs
358 * @pool:FMR pool to flush
359 *
360 * Ensure that all unmapped FMRs are fully invalidated.
361 */
ib_flush_fmr_pool(struct ib_fmr_pool * pool)362 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
363 {
364 int serial;
365 struct ib_pool_fmr *fmr, *next;
366
367 /*
368 * The free_list holds FMRs that may have been used
369 * but have not been remapped enough times to be dirty.
370 * Put them on the dirty list now so that the cleanup
371 * thread will reap them too.
372 */
373 spin_lock_irq(&pool->pool_lock);
374 list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
375 if (fmr->remap_count > 0)
376 list_move(&fmr->list, &pool->dirty_list);
377 }
378 spin_unlock_irq(&pool->pool_lock);
379
380 serial = atomic_inc_return(&pool->req_ser);
381 kthread_queue_work(pool->worker, &pool->work);
382
383 if (wait_event_interruptible(pool->force_wait,
384 atomic_read(&pool->flush_ser) - serial >= 0))
385 return -EINTR;
386
387 return 0;
388 }
389 EXPORT_SYMBOL(ib_flush_fmr_pool);
390
391 /**
392 * ib_fmr_pool_map_phys - Map an FMR from an FMR pool.
393 * @pool_handle: FMR pool to allocate FMR from
394 * @page_list: List of pages to map
395 * @list_len: Number of pages in @page_list
396 * @io_virtual_address: I/O virtual address for new FMR
397 */
ib_fmr_pool_map_phys(struct ib_fmr_pool * pool_handle,u64 * page_list,int list_len,u64 io_virtual_address)398 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
399 u64 *page_list,
400 int list_len,
401 u64 io_virtual_address)
402 {
403 struct ib_fmr_pool *pool = pool_handle;
404 struct ib_pool_fmr *fmr;
405 unsigned long flags;
406 int result;
407
408 if (list_len < 1 || list_len > pool->max_pages)
409 return ERR_PTR(-EINVAL);
410
411 spin_lock_irqsave(&pool->pool_lock, flags);
412 fmr = ib_fmr_cache_lookup(pool,
413 page_list,
414 list_len,
415 io_virtual_address);
416 if (fmr) {
417 /* found in cache */
418 ++fmr->ref_count;
419 if (fmr->ref_count == 1) {
420 list_del(&fmr->list);
421 }
422
423 spin_unlock_irqrestore(&pool->pool_lock, flags);
424
425 return fmr;
426 }
427
428 if (list_empty(&pool->free_list)) {
429 spin_unlock_irqrestore(&pool->pool_lock, flags);
430 return ERR_PTR(-EAGAIN);
431 }
432
433 fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
434 list_del(&fmr->list);
435 hlist_del_init(&fmr->cache_node);
436 spin_unlock_irqrestore(&pool->pool_lock, flags);
437
438 result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
439 io_virtual_address);
440
441 if (result) {
442 spin_lock_irqsave(&pool->pool_lock, flags);
443 list_add(&fmr->list, &pool->free_list);
444 spin_unlock_irqrestore(&pool->pool_lock, flags);
445
446 pr_warn(PFX "fmr_map returns %d\n", result);
447
448 return ERR_PTR(result);
449 }
450
451 ++fmr->remap_count;
452 fmr->ref_count = 1;
453
454 if (pool->cache_bucket) {
455 fmr->io_virtual_address = io_virtual_address;
456 fmr->page_list_len = list_len;
457 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
458
459 spin_lock_irqsave(&pool->pool_lock, flags);
460 hlist_add_head(&fmr->cache_node,
461 pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
462 spin_unlock_irqrestore(&pool->pool_lock, flags);
463 }
464
465 return fmr;
466 }
467 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
468
469 /**
470 * ib_fmr_pool_unmap - Unmap FMR
471 * @fmr:FMR to unmap
472 *
473 * Unmap an FMR. The FMR mapping may remain valid until the FMR is
474 * reused (or until ib_flush_fmr_pool() is called).
475 */
ib_fmr_pool_unmap(struct ib_pool_fmr * fmr)476 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
477 {
478 struct ib_fmr_pool *pool;
479 unsigned long flags;
480
481 pool = fmr->pool;
482
483 spin_lock_irqsave(&pool->pool_lock, flags);
484
485 --fmr->ref_count;
486 if (!fmr->ref_count) {
487 if (fmr->remap_count < pool->max_remaps) {
488 list_add_tail(&fmr->list, &pool->free_list);
489 } else {
490 list_add_tail(&fmr->list, &pool->dirty_list);
491 if (++pool->dirty_len >= pool->dirty_watermark) {
492 atomic_inc(&pool->req_ser);
493 kthread_queue_work(pool->worker, &pool->work);
494 }
495 }
496 }
497
498 #ifdef DEBUG
499 if (fmr->ref_count < 0)
500 pr_warn(PFX "FMR %p has ref count %d < 0\n",
501 fmr, fmr->ref_count);
502 #endif
503
504 spin_unlock_irqrestore(&pool->pool_lock, flags);
505
506 return 0;
507 }
508 EXPORT_SYMBOL(ib_fmr_pool_unmap);
509
