2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
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>
41 #include <rdma/ib_fmr_pool.h>
43 #include "core_priv.h"
45 #define PFX "fmr_pool: "
48 IB_FMR_MAX_REMAPS = 32,
51 IB_FMR_HASH_SIZE = 1 << IB_FMR_HASH_BITS,
52 IB_FMR_HASH_MASK = IB_FMR_HASH_SIZE - 1
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.
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.
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."
79 * All manipulation of ref_count, list and cache_node is protected by
80 * pool_lock to maintain consistency.
91 struct list_head free_list;
92 struct list_head dirty_list;
93 struct hlist_head *cache_bucket;
95 void (*flush_function)(struct ib_fmr_pool *pool,
99 struct kthread_worker *worker;
100 struct kthread_work work;
105 wait_queue_head_t force_wait;
108 static inline u32 ib_fmr_hash(u64 first_page)
110 return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
111 (IB_FMR_HASH_SIZE - 1);
114 /* Caller must hold pool_lock */
115 static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
118 u64 io_virtual_address)
120 struct hlist_head *bucket;
121 struct ib_pool_fmr *fmr;
123 if (!pool->cache_bucket)
126 bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
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))
138 static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
141 struct ib_pool_fmr *fmr;
142 LIST_HEAD(unmap_list);
145 spin_lock_irq(&pool->pool_lock);
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);
153 if (fmr->ref_count !=0) {
154 pr_warn(PFX "Unmapping FMR 0x%08x with ref count %d\n",
155 fmr, fmr->ref_count);
160 list_splice_init(&pool->dirty_list, &unmap_list);
163 spin_unlock_irq(&pool->pool_lock);
165 if (list_empty(&unmap_list)) {
169 ret = ib_unmap_fmr(&fmr_list);
171 pr_warn(PFX "ib_unmap_fmr returned %d\n", ret);
173 spin_lock_irq(&pool->pool_lock);
174 list_splice(&unmap_list, &pool->free_list);
175 spin_unlock_irq(&pool->pool_lock);
178 static void ib_fmr_cleanup_func(struct kthread_work *work)
180 struct ib_fmr_pool *pool = container_of(work, struct ib_fmr_pool, work);
182 ib_fmr_batch_release(pool);
183 atomic_inc(&pool->flush_ser);
184 wake_up_interruptible(&pool->force_wait);
186 if (pool->flush_function)
187 pool->flush_function(pool, pool->flush_arg);
189 if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0)
190 kthread_queue_work(pool->worker, &pool->work);
194 * ib_create_fmr_pool - Create an FMR pool
195 * @pd:Protection domain for FMRs
196 * @params:FMR pool parameters
198 * Create a pool of FMRs. Return value is pointer to new pool or
199 * error code if creation failed.
201 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd *pd,
202 struct ib_fmr_pool_param *params)
204 struct ib_device *device;
205 struct ib_fmr_pool *pool;
211 return ERR_PTR(-EINVAL);
214 if (!device->alloc_fmr || !device->dealloc_fmr ||
215 !device->map_phys_fmr || !device->unmap_fmr) {
216 dev_info(&device->dev, "Device does not support FMRs\n");
217 return ERR_PTR(-ENOSYS);
220 if (!device->attrs.max_map_per_fmr)
221 max_remaps = IB_FMR_MAX_REMAPS;
223 max_remaps = device->attrs.max_map_per_fmr;
225 pool = kmalloc(sizeof *pool, GFP_KERNEL);
227 return ERR_PTR(-ENOMEM);
229 pool->cache_bucket = NULL;
230 pool->flush_function = params->flush_function;
231 pool->flush_arg = params->flush_arg;
233 INIT_LIST_HEAD(&pool->free_list);
234 INIT_LIST_HEAD(&pool->dirty_list);
238 kmalloc_array(IB_FMR_HASH_SIZE,
239 sizeof(*pool->cache_bucket),
241 if (!pool->cache_bucket) {
246 for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
247 INIT_HLIST_HEAD(pool->cache_bucket + i);
251 pool->max_pages = params->max_pages_per_fmr;
252 pool->max_remaps = max_remaps;
253 pool->dirty_watermark = params->dirty_watermark;
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);
261 kthread_create_worker(0, "ib_fmr(%s)", dev_name(&device->dev));
262 if (IS_ERR(pool->worker)) {
263 pr_warn(PFX "couldn't start cleanup kthread worker\n");
264 ret = PTR_ERR(pool->worker);
267 kthread_init_work(&pool->work, ib_fmr_cleanup_func);
270 struct ib_pool_fmr *fmr;
271 struct ib_fmr_attr fmr_attr = {
272 .max_pages = params->max_pages_per_fmr,
273 .max_maps = pool->max_remaps,
274 .page_shift = params->page_shift
276 int bytes_per_fmr = sizeof *fmr;
278 if (pool->cache_bucket)
279 bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
281 for (i = 0; i < params->pool_size; ++i) {
282 fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
287 fmr->remap_count = 0;
289 INIT_HLIST_NODE(&fmr->cache_node);
291 fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
292 if (IS_ERR(fmr->fmr)) {
293 pr_warn(PFX "fmr_create failed for FMR %d\n",
299 list_add_tail(&fmr->list, &pool->free_list);
307 kfree(pool->cache_bucket);
313 ib_destroy_fmr_pool(pool);
315 return ERR_PTR(-ENOMEM);
317 EXPORT_SYMBOL(ib_create_fmr_pool);
320 * ib_destroy_fmr_pool - Free FMR pool
321 * @pool:FMR pool to free
323 * Destroy an FMR pool and free all associated resources.
325 void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
327 struct ib_pool_fmr *fmr;
328 struct ib_pool_fmr *tmp;
332 kthread_destroy_worker(pool->worker);
333 ib_fmr_batch_release(pool);
336 list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
337 if (fmr->remap_count) {
338 INIT_LIST_HEAD(&fmr_list);
339 list_add_tail(&fmr->fmr->list, &fmr_list);
340 ib_unmap_fmr(&fmr_list);
342 ib_dealloc_fmr(fmr->fmr);
343 list_del(&fmr->list);
348 if (i < pool->pool_size)
349 pr_warn(PFX "pool still has %d regions registered\n",
350 pool->pool_size - i);
352 kfree(pool->cache_bucket);
355 EXPORT_SYMBOL(ib_destroy_fmr_pool);
358 * ib_flush_fmr_pool - Invalidate all unmapped FMRs
359 * @pool:FMR pool to flush
361 * Ensure that all unmapped FMRs are fully invalidated.
363 int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
366 struct ib_pool_fmr *fmr, *next;
369 * The free_list holds FMRs that may have been used
370 * but have not been remapped enough times to be dirty.
371 * Put them on the dirty list now so that the cleanup
372 * thread will reap them too.
374 spin_lock_irq(&pool->pool_lock);
375 list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
376 if (fmr->remap_count > 0)
377 list_move(&fmr->list, &pool->dirty_list);
379 spin_unlock_irq(&pool->pool_lock);
381 serial = atomic_inc_return(&pool->req_ser);
382 kthread_queue_work(pool->worker, &pool->work);
384 if (wait_event_interruptible(pool->force_wait,
385 atomic_read(&pool->flush_ser) - serial >= 0))
390 EXPORT_SYMBOL(ib_flush_fmr_pool);
393 * ib_fmr_pool_map_phys - Map an FMR from an FMR pool.
394 * @pool_handle: FMR pool to allocate FMR from
395 * @page_list: List of pages to map
396 * @list_len: Number of pages in @page_list
397 * @io_virtual_address: I/O virtual address for new FMR
399 struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
402 u64 io_virtual_address)
404 struct ib_fmr_pool *pool = pool_handle;
405 struct ib_pool_fmr *fmr;
409 if (list_len < 1 || list_len > pool->max_pages)
410 return ERR_PTR(-EINVAL);
412 spin_lock_irqsave(&pool->pool_lock, flags);
413 fmr = ib_fmr_cache_lookup(pool,
420 if (fmr->ref_count == 1) {
421 list_del(&fmr->list);
424 spin_unlock_irqrestore(&pool->pool_lock, flags);
429 if (list_empty(&pool->free_list)) {
430 spin_unlock_irqrestore(&pool->pool_lock, flags);
431 return ERR_PTR(-EAGAIN);
434 fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
435 list_del(&fmr->list);
436 hlist_del_init(&fmr->cache_node);
437 spin_unlock_irqrestore(&pool->pool_lock, flags);
439 result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
443 spin_lock_irqsave(&pool->pool_lock, flags);
444 list_add(&fmr->list, &pool->free_list);
445 spin_unlock_irqrestore(&pool->pool_lock, flags);
447 pr_warn(PFX "fmr_map returns %d\n", result);
449 return ERR_PTR(result);
455 if (pool->cache_bucket) {
456 fmr->io_virtual_address = io_virtual_address;
457 fmr->page_list_len = list_len;
458 memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
460 spin_lock_irqsave(&pool->pool_lock, flags);
461 hlist_add_head(&fmr->cache_node,
462 pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
463 spin_unlock_irqrestore(&pool->pool_lock, flags);
468 EXPORT_SYMBOL(ib_fmr_pool_map_phys);
471 * ib_fmr_pool_unmap - Unmap FMR
474 * Unmap an FMR. The FMR mapping may remain valid until the FMR is
475 * reused (or until ib_flush_fmr_pool() is called).
477 int ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
479 struct ib_fmr_pool *pool;
484 spin_lock_irqsave(&pool->pool_lock, flags);
487 if (!fmr->ref_count) {
488 if (fmr->remap_count < pool->max_remaps) {
489 list_add_tail(&fmr->list, &pool->free_list);
491 list_add_tail(&fmr->list, &pool->dirty_list);
492 if (++pool->dirty_len >= pool->dirty_watermark) {
493 atomic_inc(&pool->req_ser);
494 kthread_queue_work(pool->worker, &pool->work);
500 if (fmr->ref_count < 0)
501 pr_warn(PFX "FMR %p has ref count %d < 0\n",
502 fmr, fmr->ref_count);
505 spin_unlock_irqrestore(&pool->pool_lock, flags);
509 EXPORT_SYMBOL(ib_fmr_pool_unmap);