[PATCH] IB: Eliminate MAD cache leak associated with local completions
[linux-2.6] / drivers / infiniband / core / fmr_pool.c
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  * $Id: fmr_pool.c 2730 2005-06-28 16:43:03Z sean.hefty $
34  */
35
36 #include <linux/errno.h>
37 #include <linux/spinlock.h>
38 #include <linux/slab.h>
39 #include <linux/jhash.h>
40 #include <linux/kthread.h>
41
42 #include <ib_fmr_pool.h>
43
44 #include "core_priv.h"
45
46 enum {
47         IB_FMR_MAX_REMAPS = 32,
48
49         IB_FMR_HASH_BITS  = 8,
50         IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
51         IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
52 };
53
54 /*
55  * If an FMR is not in use, then the list member will point to either
56  * its pool's free_list (if the FMR can be mapped again; that is,
57  * remap_count < IB_FMR_MAX_REMAPS) or its pool's dirty_list (if the
58  * FMR needs to be unmapped before being remapped).  In either of
59  * these cases it is a bug if the ref_count is not 0.  In other words,
60  * if ref_count is > 0, then the list member must not be linked into
61  * either free_list or dirty_list.
62  *
63  * The cache_node member is used to link the FMR into a cache bucket
64  * (if caching is enabled).  This is independent of the reference
65  * count of the FMR.  When a valid FMR is released, its ref_count is
66  * decremented, and if ref_count reaches 0, the FMR is placed in
67  * either free_list or dirty_list as appropriate.  However, it is not
68  * removed from the cache and may be "revived" if a call to
69  * ib_fmr_register_physical() occurs before the FMR is remapped.  In
70  * this case we just increment the ref_count and remove the FMR from
71  * free_list/dirty_list.
72  *
73  * Before we remap an FMR from free_list, we remove it from the cache
74  * (to prevent another user from obtaining a stale FMR).  When an FMR
75  * is released, we add it to the tail of the free list, so that our
76  * cache eviction policy is "least recently used."
77  *
78  * All manipulation of ref_count, list and cache_node is protected by
79  * pool_lock to maintain consistency.
80  */
81
82 struct ib_fmr_pool {
83         spinlock_t                pool_lock;
84
85         int                       pool_size;
86         int                       max_pages;
87         int                       dirty_watermark;
88         int                       dirty_len;
89         struct list_head          free_list;
90         struct list_head          dirty_list;
91         struct hlist_head        *cache_bucket;
92
93         void                     (*flush_function)(struct ib_fmr_pool *pool,
94                                                    void *              arg);
95         void                     *flush_arg;
96
97         struct task_struct       *thread;
98
99         atomic_t                  req_ser;
100         atomic_t                  flush_ser;
101
102         wait_queue_head_t         force_wait;
103 };
104
105 static inline u32 ib_fmr_hash(u64 first_page)
106 {
107         return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
108                 (IB_FMR_HASH_SIZE - 1);
109 }
110
111 /* Caller must hold pool_lock */
112 static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
113                                                       u64 *page_list,
114                                                       int  page_list_len,
115                                                       u64  io_virtual_address)
116 {
117         struct hlist_head *bucket;
118         struct ib_pool_fmr *fmr;
119         struct hlist_node *pos;
120
121         if (!pool->cache_bucket)
122                 return NULL;
123
124         bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
125
126         hlist_for_each_entry(fmr, pos, bucket, cache_node)
127                 if (io_virtual_address == fmr->io_virtual_address &&
128                     page_list_len      == fmr->page_list_len      &&
129                     !memcmp(page_list, fmr->page_list,
130                             page_list_len * sizeof *page_list))
131                         return fmr;
132
133         return NULL;
134 }
135
136 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
137 {
138         int                 ret;
139         struct ib_pool_fmr *fmr;
140         LIST_HEAD(unmap_list);
141         LIST_HEAD(fmr_list);
142
143         spin_lock_irq(&pool->pool_lock);
144
145         list_for_each_entry(fmr, &pool->dirty_list, list) {
146                 hlist_del_init(&fmr->cache_node);
147                 fmr->remap_count = 0;
148                 list_add_tail(&fmr->fmr->list, &fmr_list);
149
150 #ifdef DEBUG
151                 if (fmr->ref_count !=0) {
152                         printk(KERN_WARNING "Unmapping FMR 0x%08x with ref count %d",
153                                fmr, fmr->ref_count);
154                 }
155 #endif
156         }
157
158         list_splice(&pool->dirty_list, &unmap_list);
159         INIT_LIST_HEAD(&pool->dirty_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                 printk(KERN_WARNING "ib_unmap_fmr returned %d", 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
177 static int ib_fmr_cleanup_thread(void *pool_ptr)
178 {
179         struct ib_fmr_pool *pool = pool_ptr;
180
181         do {
182                 if (pool->dirty_len >= pool->dirty_watermark ||
183                     atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
184                         ib_fmr_batch_release(pool);
185
186                         atomic_inc(&pool->flush_ser);
187                         wake_up_interruptible(&pool->force_wait);
188
189                         if (pool->flush_function)
190                                 pool->flush_function(pool, pool->flush_arg);
191                 }
192
193                 set_current_state(TASK_INTERRUPTIBLE);
194                 if (pool->dirty_len < pool->dirty_watermark &&
195                     atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
196                     !kthread_should_stop())
197                         schedule();
198                 __set_current_state(TASK_RUNNING);
199         } while (!kthread_should_stop());
200
201         return 0;
202 }
203
204 /**
205  * ib_create_fmr_pool - Create an FMR pool
206  * @pd:Protection domain for FMRs
207  * @params:FMR pool parameters
208  *
209  * Create a pool of FMRs.  Return value is pointer to new pool or
210  * error code if creation failed.
211  */
212 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
213                                        struct ib_fmr_pool_param *params)
214 {
215         struct ib_device   *device;
216         struct ib_fmr_pool *pool;
217         int i;
218         int ret;
219
220         if (!params)
221                 return ERR_PTR(-EINVAL);
222
223         device = pd->device;
224         if (!device->alloc_fmr    || !device->dealloc_fmr  ||
225             !device->map_phys_fmr || !device->unmap_fmr) {
226                 printk(KERN_WARNING "Device %s does not support fast memory regions",
227                        device->name);
228                 return ERR_PTR(-ENOSYS);
229         }
230
231         pool = kmalloc(sizeof *pool, GFP_KERNEL);
232         if (!pool) {
233                 printk(KERN_WARNING "couldn't allocate pool struct");
234                 return ERR_PTR(-ENOMEM);
235         }
236
237         pool->cache_bucket   = NULL;
238
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                         printk(KERN_WARNING "Failed to allocate cache in pool");
251                         ret = -ENOMEM;
252                         goto out_free_pool;
253                 }
254
255                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
256                         INIT_HLIST_HEAD(pool->cache_bucket + i);
257         }
258
259         pool->pool_size       = 0;
260         pool->max_pages       = params->max_pages_per_fmr;
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_create(ib_fmr_cleanup_thread,
269                                       pool,
270                                       "ib_fmr(%s)",
271                                       device->name);
272         if (IS_ERR(pool->thread)) {
273                 printk(KERN_WARNING "couldn't start cleanup thread");
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 attr = {
281                         .max_pages = params->max_pages_per_fmr,
282                         .max_maps  = IB_FMR_MAX_REMAPS,
283                         .page_size = PAGE_SHIFT
284                 };
285
286                 for (i = 0; i < params->pool_size; ++i) {
287                         fmr = kmalloc(sizeof *fmr + params->max_pages_per_fmr * sizeof (u64),
288                                       GFP_KERNEL);
289                         if (!fmr) {
290                                 printk(KERN_WARNING "failed to allocate fmr struct "
291                                        "for FMR %d", i);
292                                 goto out_fail;
293                         }
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, &attr);
301                         if (IS_ERR(fmr->fmr)) {
302                                 printk(KERN_WARNING "fmr_create failed for FMR %d", i);
303                                 kfree(fmr);
304                                 goto out_fail;
305                         }
306
307                         list_add_tail(&fmr->list, &pool->free_list);
308                         ++pool->pool_size;
309                 }
310         }
311
312         return pool;
313
314  out_free_pool:
315         kfree(pool->cache_bucket);
316         kfree(pool);
317
318         return ERR_PTR(ret);
319
320  out_fail:
321         ib_destroy_fmr_pool(pool);
322
323         return ERR_PTR(-ENOMEM);
324 }
325 EXPORT_SYMBOL(ib_create_fmr_pool);
326
327 /**
328  * ib_destroy_fmr_pool - Free FMR pool
329  * @pool:FMR pool to free
330  *
331  * Destroy an FMR pool and free all associated resources.
332  */
333 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
334 {
335         struct ib_pool_fmr *fmr;
336         struct ib_pool_fmr *tmp;
337         int                 i;
338
339         kthread_stop(pool->thread);
340         ib_fmr_batch_release(pool);
341
342         i = 0;
343         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
344                 ib_dealloc_fmr(fmr->fmr);
345                 list_del(&fmr->list);
346                 kfree(fmr);
347                 ++i;
348         }
349
350         if (i < pool->pool_size)
351                 printk(KERN_WARNING "pool still has %d regions registered",
352                        pool->pool_size - i);
353
354         kfree(pool->cache_bucket);
355         kfree(pool);
356 }
357 EXPORT_SYMBOL(ib_destroy_fmr_pool);
358
359 /**
360  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
361  * @pool:FMR pool to flush
362  *
363  * Ensure that all unmapped FMRs are fully invalidated.
364  */
365 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
366 {
367         int serial;
368
369         atomic_inc(&pool->req_ser);
370         /*
371          * It's OK if someone else bumps req_ser again here -- we'll
372          * just wait a little longer.
373          */
374         serial = atomic_read(&pool->req_ser);
375
376         wake_up_process(pool->thread);
377
378         if (wait_event_interruptible(pool->force_wait,
379                                      atomic_read(&pool->flush_ser) -
380                                      atomic_read(&pool->req_ser) >= 0))
381                 return -EINTR;
382
383         return 0;
384 }
385 EXPORT_SYMBOL(ib_flush_fmr_pool);
386
387 /**
388  * ib_fmr_pool_map_phys -
389  * @pool:FMR pool to allocate FMR from
390  * @page_list:List of pages to map
391  * @list_len:Number of pages in @page_list
392  * @io_virtual_address:I/O virtual address for new FMR
393  *
394  * Map an FMR from an FMR pool.
395  */
396 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
397                                          u64                *page_list,
398                                          int                 list_len,
399                                          u64                *io_virtual_address)
400 {
401         struct ib_fmr_pool *pool = pool_handle;
402         struct ib_pool_fmr *fmr;
403         unsigned long       flags;
404         int                 result;
405
406         if (list_len < 1 || list_len > pool->max_pages)
407                 return ERR_PTR(-EINVAL);
408
409         spin_lock_irqsave(&pool->pool_lock, flags);
410         fmr = ib_fmr_cache_lookup(pool,
411                                   page_list,
412                                   list_len,
413                                   *io_virtual_address);
414         if (fmr) {
415                 /* found in cache */
416                 ++fmr->ref_count;
417                 if (fmr->ref_count == 1) {
418                         list_del(&fmr->list);
419                 }
420
421                 spin_unlock_irqrestore(&pool->pool_lock, flags);
422
423                 return fmr;
424         }
425
426         if (list_empty(&pool->free_list)) {
427                 spin_unlock_irqrestore(&pool->pool_lock, flags);
428                 return ERR_PTR(-EAGAIN);
429         }
430
431         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
432         list_del(&fmr->list);
433         hlist_del_init(&fmr->cache_node);
434         spin_unlock_irqrestore(&pool->pool_lock, flags);
435
436         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
437                                  *io_virtual_address);
438
439         if (result) {
440                 spin_lock_irqsave(&pool->pool_lock, flags);
441                 list_add(&fmr->list, &pool->free_list);
442                 spin_unlock_irqrestore(&pool->pool_lock, flags);
443
444                 printk(KERN_WARNING "fmr_map returns %d\n",
445                        result);
446
447                 return ERR_PTR(result);
448         }
449
450         ++fmr->remap_count;
451         fmr->ref_count = 1;
452
453         if (pool->cache_bucket) {
454                 fmr->io_virtual_address = *io_virtual_address;
455                 fmr->page_list_len      = list_len;
456                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
457
458                 spin_lock_irqsave(&pool->pool_lock, flags);
459                 hlist_add_head(&fmr->cache_node,
460                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
461                 spin_unlock_irqrestore(&pool->pool_lock, flags);
462         }
463
464         return fmr;
465 }
466 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
467
468 /**
469  * ib_fmr_pool_unmap - Unmap FMR
470  * @fmr:FMR to unmap
471  *
472  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
473  * reused (or until ib_flush_fmr_pool() is called).
474  */
475 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
476 {
477         struct ib_fmr_pool *pool;
478         unsigned long flags;
479
480         pool = fmr->pool;
481
482         spin_lock_irqsave(&pool->pool_lock, flags);
483
484         --fmr->ref_count;
485         if (!fmr->ref_count) {
486                 if (fmr->remap_count < IB_FMR_MAX_REMAPS) {
487                         list_add_tail(&fmr->list, &pool->free_list);
488                 } else {
489                         list_add_tail(&fmr->list, &pool->dirty_list);
490                         ++pool->dirty_len;
491                         wake_up_process(pool->thread);
492                 }
493         }
494
495 #ifdef DEBUG
496         if (fmr->ref_count < 0)
497                 printk(KERN_WARNING "FMR %p has ref count %d < 0",
498                        fmr, fmr->ref_count);
499 #endif
500
501         spin_unlock_irqrestore(&pool->pool_lock, flags);
502
503         return 0;
504 }
505 EXPORT_SYMBOL(ib_fmr_pool_unmap);