Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[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 <rdma/ib_fmr_pool.h>
43
44 #include "core_priv.h"
45
46 #define PFX "fmr_pool: "
47
48 enum {
49         IB_FMR_MAX_REMAPS = 32,
50
51         IB_FMR_HASH_BITS  = 8,
52         IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
53         IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
54 };
55
56 /*
57  * If an FMR is not in use, then the list member will point to either
58  * its pool's free_list (if the FMR can be mapped again; that is,
59  * remap_count < pool->max_remaps) or its pool's dirty_list (if the
60  * FMR needs to be unmapped before being remapped).  In either of
61  * these cases it is a bug if the ref_count is not 0.  In other words,
62  * if ref_count is > 0, then the list member must not be linked into
63  * either free_list or dirty_list.
64  *
65  * The cache_node member is used to link the FMR into a cache bucket
66  * (if caching is enabled).  This is independent of the reference
67  * count of the FMR.  When a valid FMR is released, its ref_count is
68  * decremented, and if ref_count reaches 0, the FMR is placed in
69  * either free_list or dirty_list as appropriate.  However, it is not
70  * removed from the cache and may be "revived" if a call to
71  * ib_fmr_register_physical() occurs before the FMR is remapped.  In
72  * this case we just increment the ref_count and remove the FMR from
73  * free_list/dirty_list.
74  *
75  * Before we remap an FMR from free_list, we remove it from the cache
76  * (to prevent another user from obtaining a stale FMR).  When an FMR
77  * is released, we add it to the tail of the free list, so that our
78  * cache eviction policy is "least recently used."
79  *
80  * All manipulation of ref_count, list and cache_node is protected by
81  * pool_lock to maintain consistency.
82  */
83
84 struct ib_fmr_pool {
85         spinlock_t                pool_lock;
86
87         int                       pool_size;
88         int                       max_pages;
89         int                       max_remaps;
90         int                       dirty_watermark;
91         int                       dirty_len;
92         struct list_head          free_list;
93         struct list_head          dirty_list;
94         struct hlist_head        *cache_bucket;
95
96         void                     (*flush_function)(struct ib_fmr_pool *pool,
97                                                    void *              arg);
98         void                     *flush_arg;
99
100         struct task_struct       *thread;
101
102         atomic_t                  req_ser;
103         atomic_t                  flush_ser;
104
105         wait_queue_head_t         force_wait;
106 };
107
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 */
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         struct hlist_node *pos;
123
124         if (!pool->cache_bucket)
125                 return NULL;
126
127         bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
128
129         hlist_for_each_entry(fmr, pos, bucket, cache_node)
130                 if (io_virtual_address == fmr->io_virtual_address &&
131                     page_list_len      == fmr->page_list_len      &&
132                     !memcmp(page_list, fmr->page_list,
133                             page_list_len * sizeof *page_list))
134                         return fmr;
135
136         return NULL;
137 }
138
139 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
140 {
141         int                 ret;
142         struct ib_pool_fmr *fmr;
143         LIST_HEAD(unmap_list);
144         LIST_HEAD(fmr_list);
145
146         spin_lock_irq(&pool->pool_lock);
147
148         list_for_each_entry(fmr, &pool->dirty_list, list) {
149                 hlist_del_init(&fmr->cache_node);
150                 fmr->remap_count = 0;
151                 list_add_tail(&fmr->fmr->list, &fmr_list);
152
153 #ifdef DEBUG
154                 if (fmr->ref_count !=0) {
155                         printk(KERN_WARNING PFX "Unmapping FMR 0x%08x with ref count %d\n",
156                                fmr, fmr->ref_count);
157                 }
158 #endif
159         }
160
161         list_splice(&pool->dirty_list, &unmap_list);
162         INIT_LIST_HEAD(&pool->dirty_list);
163         pool->dirty_len = 0;
164
165         spin_unlock_irq(&pool->pool_lock);
166
167         if (list_empty(&unmap_list)) {
168                 return;
169         }
170
171         ret = ib_unmap_fmr(&fmr_list);
172         if (ret)
173                 printk(KERN_WARNING PFX "ib_unmap_fmr returned %d\n", ret);
174
175         spin_lock_irq(&pool->pool_lock);
176         list_splice(&unmap_list, &pool->free_list);
177         spin_unlock_irq(&pool->pool_lock);
178 }
179
180 static int ib_fmr_cleanup_thread(void *pool_ptr)
181 {
182         struct ib_fmr_pool *pool = pool_ptr;
183
184         do {
185                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
186                         ib_fmr_batch_release(pool);
187
188                         atomic_inc(&pool->flush_ser);
189                         wake_up_interruptible(&pool->force_wait);
190
191                         if (pool->flush_function)
192                                 pool->flush_function(pool, pool->flush_arg);
193                 }
194
195                 set_current_state(TASK_INTERRUPTIBLE);
196                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
197                     !kthread_should_stop())
198                         schedule();
199                 __set_current_state(TASK_RUNNING);
200         } while (!kthread_should_stop());
201
202         return 0;
203 }
204
205 /**
206  * ib_create_fmr_pool - Create an FMR pool
207  * @pd:Protection domain for FMRs
208  * @params:FMR pool parameters
209  *
210  * Create a pool of FMRs.  Return value is pointer to new pool or
211  * error code if creation failed.
212  */
213 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
214                                        struct ib_fmr_pool_param *params)
215 {
216         struct ib_device   *device;
217         struct ib_fmr_pool *pool;
218         struct ib_device_attr *attr;
219         int i;
220         int ret;
221         int max_remaps;
222
223         if (!params)
224                 return ERR_PTR(-EINVAL);
225
226         device = pd->device;
227         if (!device->alloc_fmr    || !device->dealloc_fmr  ||
228             !device->map_phys_fmr || !device->unmap_fmr) {
229                 printk(KERN_INFO PFX "Device %s does not support FMRs\n",
230                        device->name);
231                 return ERR_PTR(-ENOSYS);
232         }
233
234         attr = kmalloc(sizeof *attr, GFP_KERNEL);
235         if (!attr) {
236                 printk(KERN_WARNING PFX "couldn't allocate device attr struct\n");
237                 return ERR_PTR(-ENOMEM);
238         }
239
240         ret = ib_query_device(device, attr);
241         if (ret) {
242                 printk(KERN_WARNING PFX "couldn't query device: %d\n", ret);
243                 kfree(attr);
244                 return ERR_PTR(ret);
245         }
246
247         if (!attr->max_map_per_fmr)
248                 max_remaps = IB_FMR_MAX_REMAPS;
249         else
250                 max_remaps = attr->max_map_per_fmr;
251
252         kfree(attr);
253
254         pool = kmalloc(sizeof *pool, GFP_KERNEL);
255         if (!pool) {
256                 printk(KERN_WARNING PFX "couldn't allocate pool struct\n");
257                 return ERR_PTR(-ENOMEM);
258         }
259
260         pool->cache_bucket   = NULL;
261
262         pool->flush_function = params->flush_function;
263         pool->flush_arg      = params->flush_arg;
264
265         INIT_LIST_HEAD(&pool->free_list);
266         INIT_LIST_HEAD(&pool->dirty_list);
267
268         if (params->cache) {
269                 pool->cache_bucket =
270                         kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
271                                 GFP_KERNEL);
272                 if (!pool->cache_bucket) {
273                         printk(KERN_WARNING PFX "Failed to allocate cache in pool\n");
274                         ret = -ENOMEM;
275                         goto out_free_pool;
276                 }
277
278                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
279                         INIT_HLIST_HEAD(pool->cache_bucket + i);
280         }
281
282         pool->pool_size       = 0;
283         pool->max_pages       = params->max_pages_per_fmr;
284         pool->max_remaps      = max_remaps;
285         pool->dirty_watermark = params->dirty_watermark;
286         pool->dirty_len       = 0;
287         spin_lock_init(&pool->pool_lock);
288         atomic_set(&pool->req_ser,   0);
289         atomic_set(&pool->flush_ser, 0);
290         init_waitqueue_head(&pool->force_wait);
291
292         pool->thread = kthread_run(ib_fmr_cleanup_thread,
293                                    pool,
294                                    "ib_fmr(%s)",
295                                    device->name);
296         if (IS_ERR(pool->thread)) {
297                 printk(KERN_WARNING PFX "couldn't start cleanup thread\n");
298                 ret = PTR_ERR(pool->thread);
299                 goto out_free_pool;
300         }
301
302         {
303                 struct ib_pool_fmr *fmr;
304                 struct ib_fmr_attr fmr_attr = {
305                         .max_pages  = params->max_pages_per_fmr,
306                         .max_maps   = pool->max_remaps,
307                         .page_shift = params->page_shift
308                 };
309                 int bytes_per_fmr = sizeof *fmr;
310
311                 if (pool->cache_bucket)
312                         bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
313
314                 for (i = 0; i < params->pool_size; ++i) {
315                         fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
316                         if (!fmr) {
317                                 printk(KERN_WARNING PFX "failed to allocate fmr "
318                                        "struct for FMR %d\n", i);
319                                 goto out_fail;
320                         }
321
322                         fmr->pool             = pool;
323                         fmr->remap_count      = 0;
324                         fmr->ref_count        = 0;
325                         INIT_HLIST_NODE(&fmr->cache_node);
326
327                         fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
328                         if (IS_ERR(fmr->fmr)) {
329                                 printk(KERN_WARNING PFX "fmr_create failed "
330                                        "for FMR %d\n", i);
331                                 kfree(fmr);
332                                 goto out_fail;
333                         }
334
335                         list_add_tail(&fmr->list, &pool->free_list);
336                         ++pool->pool_size;
337                 }
338         }
339
340         return pool;
341
342  out_free_pool:
343         kfree(pool->cache_bucket);
344         kfree(pool);
345
346         return ERR_PTR(ret);
347
348  out_fail:
349         ib_destroy_fmr_pool(pool);
350
351         return ERR_PTR(-ENOMEM);
352 }
353 EXPORT_SYMBOL(ib_create_fmr_pool);
354
355 /**
356  * ib_destroy_fmr_pool - Free FMR pool
357  * @pool:FMR pool to free
358  *
359  * Destroy an FMR pool and free all associated resources.
360  */
361 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
362 {
363         struct ib_pool_fmr *fmr;
364         struct ib_pool_fmr *tmp;
365         LIST_HEAD(fmr_list);
366         int                 i;
367
368         kthread_stop(pool->thread);
369         ib_fmr_batch_release(pool);
370
371         i = 0;
372         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
373                 if (fmr->remap_count) {
374                         INIT_LIST_HEAD(&fmr_list);
375                         list_add_tail(&fmr->fmr->list, &fmr_list);
376                         ib_unmap_fmr(&fmr_list);
377                 }
378                 ib_dealloc_fmr(fmr->fmr);
379                 list_del(&fmr->list);
380                 kfree(fmr);
381                 ++i;
382         }
383
384         if (i < pool->pool_size)
385                 printk(KERN_WARNING PFX "pool still has %d regions registered\n",
386                        pool->pool_size - i);
387
388         kfree(pool->cache_bucket);
389         kfree(pool);
390 }
391 EXPORT_SYMBOL(ib_destroy_fmr_pool);
392
393 /**
394  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
395  * @pool:FMR pool to flush
396  *
397  * Ensure that all unmapped FMRs are fully invalidated.
398  */
399 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
400 {
401         int serial;
402         struct ib_pool_fmr *fmr, *next;
403
404         /*
405          * The free_list holds FMRs that may have been used
406          * but have not been remapped enough times to be dirty.
407          * Put them on the dirty list now so that the cleanup
408          * thread will reap them too.
409          */
410         spin_lock_irq(&pool->pool_lock);
411         list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
412                 if (fmr->remap_count > 0)
413                         list_move(&fmr->list, &pool->dirty_list);
414         }
415         spin_unlock_irq(&pool->pool_lock);
416
417         serial = atomic_inc_return(&pool->req_ser);
418         wake_up_process(pool->thread);
419
420         if (wait_event_interruptible(pool->force_wait,
421                                      atomic_read(&pool->flush_ser) - serial >= 0))
422                 return -EINTR;
423
424         return 0;
425 }
426 EXPORT_SYMBOL(ib_flush_fmr_pool);
427
428 /**
429  * ib_fmr_pool_map_phys -
430  * @pool:FMR pool to allocate FMR from
431  * @page_list:List of pages to map
432  * @list_len:Number of pages in @page_list
433  * @io_virtual_address:I/O virtual address for new FMR
434  *
435  * Map an FMR from an FMR pool.
436  */
437 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
438                                          u64                *page_list,
439                                          int                 list_len,
440                                          u64                 io_virtual_address)
441 {
442         struct ib_fmr_pool *pool = pool_handle;
443         struct ib_pool_fmr *fmr;
444         unsigned long       flags;
445         int                 result;
446
447         if (list_len < 1 || list_len > pool->max_pages)
448                 return ERR_PTR(-EINVAL);
449
450         spin_lock_irqsave(&pool->pool_lock, flags);
451         fmr = ib_fmr_cache_lookup(pool,
452                                   page_list,
453                                   list_len,
454                                   io_virtual_address);
455         if (fmr) {
456                 /* found in cache */
457                 ++fmr->ref_count;
458                 if (fmr->ref_count == 1) {
459                         list_del(&fmr->list);
460                 }
461
462                 spin_unlock_irqrestore(&pool->pool_lock, flags);
463
464                 return fmr;
465         }
466
467         if (list_empty(&pool->free_list)) {
468                 spin_unlock_irqrestore(&pool->pool_lock, flags);
469                 return ERR_PTR(-EAGAIN);
470         }
471
472         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
473         list_del(&fmr->list);
474         hlist_del_init(&fmr->cache_node);
475         spin_unlock_irqrestore(&pool->pool_lock, flags);
476
477         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
478                                  io_virtual_address);
479
480         if (result) {
481                 spin_lock_irqsave(&pool->pool_lock, flags);
482                 list_add(&fmr->list, &pool->free_list);
483                 spin_unlock_irqrestore(&pool->pool_lock, flags);
484
485                 printk(KERN_WARNING PFX "fmr_map returns %d\n", result);
486
487                 return ERR_PTR(result);
488         }
489
490         ++fmr->remap_count;
491         fmr->ref_count = 1;
492
493         if (pool->cache_bucket) {
494                 fmr->io_virtual_address = io_virtual_address;
495                 fmr->page_list_len      = list_len;
496                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
497
498                 spin_lock_irqsave(&pool->pool_lock, flags);
499                 hlist_add_head(&fmr->cache_node,
500                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
501                 spin_unlock_irqrestore(&pool->pool_lock, flags);
502         }
503
504         return fmr;
505 }
506 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
507
508 /**
509  * ib_fmr_pool_unmap - Unmap FMR
510  * @fmr:FMR to unmap
511  *
512  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
513  * reused (or until ib_flush_fmr_pool() is called).
514  */
515 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
516 {
517         struct ib_fmr_pool *pool;
518         unsigned long flags;
519
520         pool = fmr->pool;
521
522         spin_lock_irqsave(&pool->pool_lock, flags);
523
524         --fmr->ref_count;
525         if (!fmr->ref_count) {
526                 if (fmr->remap_count < pool->max_remaps) {
527                         list_add_tail(&fmr->list, &pool->free_list);
528                 } else {
529                         list_add_tail(&fmr->list, &pool->dirty_list);
530                         if (++pool->dirty_len >= pool->dirty_watermark) {
531                                 atomic_inc(&pool->req_ser);
532                                 wake_up_process(pool->thread);
533                         }
534                 }
535         }
536
537 #ifdef DEBUG
538         if (fmr->ref_count < 0)
539                 printk(KERN_WARNING PFX "FMR %p has ref count %d < 0\n",
540                        fmr, fmr->ref_count);
541 #endif
542
543         spin_unlock_irqrestore(&pool->pool_lock, flags);
544
545         return 0;
546 }
547 EXPORT_SYMBOL(ib_fmr_pool_unmap);