IB/fmr_pool: Allocate page list for pool FMRs only when caching enabled
[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, *next;
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         /*
162          * The free_list may hold FMRs that have been put there
163          * because they haven't reached the max_remap count.
164          * Invalidate their mapping as well.
165          */
166         list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
167                 if (fmr->remap_count == 0)
168                         continue;
169                 hlist_del_init(&fmr->cache_node);
170                 fmr->remap_count = 0;
171                 list_add_tail(&fmr->fmr->list, &fmr_list);
172                 list_move(&fmr->list, &unmap_list);
173         }
174
175         list_splice(&pool->dirty_list, &unmap_list);
176         INIT_LIST_HEAD(&pool->dirty_list);
177         pool->dirty_len = 0;
178
179         spin_unlock_irq(&pool->pool_lock);
180
181         if (list_empty(&unmap_list)) {
182                 return;
183         }
184
185         ret = ib_unmap_fmr(&fmr_list);
186         if (ret)
187                 printk(KERN_WARNING PFX "ib_unmap_fmr returned %d\n", ret);
188
189         spin_lock_irq(&pool->pool_lock);
190         list_splice(&unmap_list, &pool->free_list);
191         spin_unlock_irq(&pool->pool_lock);
192 }
193
194 static int ib_fmr_cleanup_thread(void *pool_ptr)
195 {
196         struct ib_fmr_pool *pool = pool_ptr;
197
198         do {
199                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0) {
200                         ib_fmr_batch_release(pool);
201
202                         atomic_inc(&pool->flush_ser);
203                         wake_up_interruptible(&pool->force_wait);
204
205                         if (pool->flush_function)
206                                 pool->flush_function(pool, pool->flush_arg);
207                 }
208
209                 set_current_state(TASK_INTERRUPTIBLE);
210                 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) >= 0 &&
211                     !kthread_should_stop())
212                         schedule();
213                 __set_current_state(TASK_RUNNING);
214         } while (!kthread_should_stop());
215
216         return 0;
217 }
218
219 /**
220  * ib_create_fmr_pool - Create an FMR pool
221  * @pd:Protection domain for FMRs
222  * @params:FMR pool parameters
223  *
224  * Create a pool of FMRs.  Return value is pointer to new pool or
225  * error code if creation failed.
226  */
227 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
228                                        struct ib_fmr_pool_param *params)
229 {
230         struct ib_device   *device;
231         struct ib_fmr_pool *pool;
232         struct ib_device_attr *attr;
233         int i;
234         int ret;
235         int max_remaps;
236
237         if (!params)
238                 return ERR_PTR(-EINVAL);
239
240         device = pd->device;
241         if (!device->alloc_fmr    || !device->dealloc_fmr  ||
242             !device->map_phys_fmr || !device->unmap_fmr) {
243                 printk(KERN_INFO PFX "Device %s does not support FMRs\n",
244                        device->name);
245                 return ERR_PTR(-ENOSYS);
246         }
247
248         attr = kmalloc(sizeof *attr, GFP_KERNEL);
249         if (!attr) {
250                 printk(KERN_WARNING PFX "couldn't allocate device attr struct\n");
251                 return ERR_PTR(-ENOMEM);
252         }
253
254         ret = ib_query_device(device, attr);
255         if (ret) {
256                 printk(KERN_WARNING PFX "couldn't query device: %d\n", ret);
257                 kfree(attr);
258                 return ERR_PTR(ret);
259         }
260
261         if (!attr->max_map_per_fmr)
262                 max_remaps = IB_FMR_MAX_REMAPS;
263         else
264                 max_remaps = attr->max_map_per_fmr;
265
266         kfree(attr);
267
268         pool = kmalloc(sizeof *pool, GFP_KERNEL);
269         if (!pool) {
270                 printk(KERN_WARNING PFX "couldn't allocate pool struct\n");
271                 return ERR_PTR(-ENOMEM);
272         }
273
274         pool->cache_bucket   = NULL;
275
276         pool->flush_function = params->flush_function;
277         pool->flush_arg      = params->flush_arg;
278
279         INIT_LIST_HEAD(&pool->free_list);
280         INIT_LIST_HEAD(&pool->dirty_list);
281
282         if (params->cache) {
283                 pool->cache_bucket =
284                         kmalloc(IB_FMR_HASH_SIZE * sizeof *pool->cache_bucket,
285                                 GFP_KERNEL);
286                 if (!pool->cache_bucket) {
287                         printk(KERN_WARNING PFX "Failed to allocate cache in pool\n");
288                         ret = -ENOMEM;
289                         goto out_free_pool;
290                 }
291
292                 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
293                         INIT_HLIST_HEAD(pool->cache_bucket + i);
294         }
295
296         pool->pool_size       = 0;
297         pool->max_pages       = params->max_pages_per_fmr;
298         pool->max_remaps      = max_remaps;
299         pool->dirty_watermark = params->dirty_watermark;
300         pool->dirty_len       = 0;
301         spin_lock_init(&pool->pool_lock);
302         atomic_set(&pool->req_ser,   0);
303         atomic_set(&pool->flush_ser, 0);
304         init_waitqueue_head(&pool->force_wait);
305
306         pool->thread = kthread_run(ib_fmr_cleanup_thread,
307                                    pool,
308                                    "ib_fmr(%s)",
309                                    device->name);
310         if (IS_ERR(pool->thread)) {
311                 printk(KERN_WARNING PFX "couldn't start cleanup thread\n");
312                 ret = PTR_ERR(pool->thread);
313                 goto out_free_pool;
314         }
315
316         {
317                 struct ib_pool_fmr *fmr;
318                 struct ib_fmr_attr fmr_attr = {
319                         .max_pages  = params->max_pages_per_fmr,
320                         .max_maps   = pool->max_remaps,
321                         .page_shift = params->page_shift
322                 };
323                 int bytes_per_fmr = sizeof *fmr;
324
325                 if (pool->cache_bucket)
326                         bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
327
328                 for (i = 0; i < params->pool_size; ++i) {
329                         fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
330                         if (!fmr) {
331                                 printk(KERN_WARNING PFX "failed to allocate fmr "
332                                        "struct for FMR %d\n", i);
333                                 goto out_fail;
334                         }
335
336                         fmr->pool             = pool;
337                         fmr->remap_count      = 0;
338                         fmr->ref_count        = 0;
339                         INIT_HLIST_NODE(&fmr->cache_node);
340
341                         fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
342                         if (IS_ERR(fmr->fmr)) {
343                                 printk(KERN_WARNING PFX "fmr_create failed "
344                                        "for FMR %d\n", i);
345                                 kfree(fmr);
346                                 goto out_fail;
347                         }
348
349                         list_add_tail(&fmr->list, &pool->free_list);
350                         ++pool->pool_size;
351                 }
352         }
353
354         return pool;
355
356  out_free_pool:
357         kfree(pool->cache_bucket);
358         kfree(pool);
359
360         return ERR_PTR(ret);
361
362  out_fail:
363         ib_destroy_fmr_pool(pool);
364
365         return ERR_PTR(-ENOMEM);
366 }
367 EXPORT_SYMBOL(ib_create_fmr_pool);
368
369 /**
370  * ib_destroy_fmr_pool - Free FMR pool
371  * @pool:FMR pool to free
372  *
373  * Destroy an FMR pool and free all associated resources.
374  */
375 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
376 {
377         struct ib_pool_fmr *fmr;
378         struct ib_pool_fmr *tmp;
379         LIST_HEAD(fmr_list);
380         int                 i;
381
382         kthread_stop(pool->thread);
383         ib_fmr_batch_release(pool);
384
385         i = 0;
386         list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
387                 ib_dealloc_fmr(fmr->fmr);
388                 list_del(&fmr->list);
389                 kfree(fmr);
390                 ++i;
391         }
392
393         if (i < pool->pool_size)
394                 printk(KERN_WARNING PFX "pool still has %d regions registered\n",
395                        pool->pool_size - i);
396
397         kfree(pool->cache_bucket);
398         kfree(pool);
399 }
400 EXPORT_SYMBOL(ib_destroy_fmr_pool);
401
402 /**
403  * ib_flush_fmr_pool - Invalidate all unmapped FMRs
404  * @pool:FMR pool to flush
405  *
406  * Ensure that all unmapped FMRs are fully invalidated.
407  */
408 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
409 {
410         int serial = atomic_inc_return(&pool->req_ser);
411
412         wake_up_process(pool->thread);
413
414         if (wait_event_interruptible(pool->force_wait,
415                                      atomic_read(&pool->flush_ser) - serial >= 0))
416                 return -EINTR;
417
418         return 0;
419 }
420 EXPORT_SYMBOL(ib_flush_fmr_pool);
421
422 /**
423  * ib_fmr_pool_map_phys -
424  * @pool:FMR pool to allocate FMR from
425  * @page_list:List of pages to map
426  * @list_len:Number of pages in @page_list
427  * @io_virtual_address:I/O virtual address for new FMR
428  *
429  * Map an FMR from an FMR pool.
430  */
431 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
432                                          u64                *page_list,
433                                          int                 list_len,
434                                          u64                 io_virtual_address)
435 {
436         struct ib_fmr_pool *pool = pool_handle;
437         struct ib_pool_fmr *fmr;
438         unsigned long       flags;
439         int                 result;
440
441         if (list_len < 1 || list_len > pool->max_pages)
442                 return ERR_PTR(-EINVAL);
443
444         spin_lock_irqsave(&pool->pool_lock, flags);
445         fmr = ib_fmr_cache_lookup(pool,
446                                   page_list,
447                                   list_len,
448                                   io_virtual_address);
449         if (fmr) {
450                 /* found in cache */
451                 ++fmr->ref_count;
452                 if (fmr->ref_count == 1) {
453                         list_del(&fmr->list);
454                 }
455
456                 spin_unlock_irqrestore(&pool->pool_lock, flags);
457
458                 return fmr;
459         }
460
461         if (list_empty(&pool->free_list)) {
462                 spin_unlock_irqrestore(&pool->pool_lock, flags);
463                 return ERR_PTR(-EAGAIN);
464         }
465
466         fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
467         list_del(&fmr->list);
468         hlist_del_init(&fmr->cache_node);
469         spin_unlock_irqrestore(&pool->pool_lock, flags);
470
471         result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
472                                  io_virtual_address);
473
474         if (result) {
475                 spin_lock_irqsave(&pool->pool_lock, flags);
476                 list_add(&fmr->list, &pool->free_list);
477                 spin_unlock_irqrestore(&pool->pool_lock, flags);
478
479                 printk(KERN_WARNING PFX "fmr_map returns %d\n", result);
480
481                 return ERR_PTR(result);
482         }
483
484         ++fmr->remap_count;
485         fmr->ref_count = 1;
486
487         if (pool->cache_bucket) {
488                 fmr->io_virtual_address = io_virtual_address;
489                 fmr->page_list_len      = list_len;
490                 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
491
492                 spin_lock_irqsave(&pool->pool_lock, flags);
493                 hlist_add_head(&fmr->cache_node,
494                                pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
495                 spin_unlock_irqrestore(&pool->pool_lock, flags);
496         }
497
498         return fmr;
499 }
500 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
501
502 /**
503  * ib_fmr_pool_unmap - Unmap FMR
504  * @fmr:FMR to unmap
505  *
506  * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
507  * reused (or until ib_flush_fmr_pool() is called).
508  */
509 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
510 {
511         struct ib_fmr_pool *pool;
512         unsigned long flags;
513
514         pool = fmr->pool;
515
516         spin_lock_irqsave(&pool->pool_lock, flags);
517
518         --fmr->ref_count;
519         if (!fmr->ref_count) {
520                 if (fmr->remap_count < pool->max_remaps) {
521                         list_add_tail(&fmr->list, &pool->free_list);
522                 } else {
523                         list_add_tail(&fmr->list, &pool->dirty_list);
524                         if (++pool->dirty_len >= pool->dirty_watermark) {
525                                 atomic_inc(&pool->req_ser);
526                                 wake_up_process(pool->thread);
527                         }
528                 }
529         }
530
531 #ifdef DEBUG
532         if (fmr->ref_count < 0)
533                 printk(KERN_WARNING PFX "FMR %p has ref count %d < 0\n",
534                        fmr, fmr->ref_count);
535 #endif
536
537         spin_unlock_irqrestore(&pool->pool_lock, flags);
538
539         return 0;
540 }
541 EXPORT_SYMBOL(ib_fmr_pool_unmap);