1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright © 2006-2009, Intel Corporation.
5 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
8 #include <linux/iova.h>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/smp.h>
12 #include <linux/bitops.h>
13 #include <linux/cpu.h>
15 /* The anchor node sits above the top of the usable address space */
16 #define IOVA_ANCHOR ~0UL
18 static bool iova_rcache_insert(struct iova_domain *iovad,
21 static unsigned long iova_rcache_get(struct iova_domain *iovad,
23 unsigned long limit_pfn);
24 static void init_iova_rcaches(struct iova_domain *iovad);
25 static void free_iova_rcaches(struct iova_domain *iovad);
26 static void fq_destroy_all_entries(struct iova_domain *iovad);
27 static void fq_flush_timeout(struct timer_list *t);
30 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
31 unsigned long start_pfn)
34 * IOVA granularity will normally be equal to the smallest
35 * supported IOMMU page size; both *must* be capable of
36 * representing individual CPU pages exactly.
38 BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
40 spin_lock_init(&iovad->iova_rbtree_lock);
41 iovad->rbroot = RB_ROOT;
42 iovad->cached_node = &iovad->anchor.node;
43 iovad->cached32_node = &iovad->anchor.node;
44 iovad->granule = granule;
45 iovad->start_pfn = start_pfn;
46 iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
47 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
48 iovad->flush_cb = NULL;
50 iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
51 rb_link_node(&iovad->anchor.node, NULL, &iovad->rbroot.rb_node);
52 rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
53 init_iova_rcaches(iovad);
55 EXPORT_SYMBOL_GPL(init_iova_domain);
57 bool has_iova_flush_queue(struct iova_domain *iovad)
62 static void free_iova_flush_queue(struct iova_domain *iovad)
64 if (!has_iova_flush_queue(iovad))
67 del_timer_sync(&iovad->fq_timer);
69 fq_destroy_all_entries(iovad);
71 free_percpu(iovad->fq);
74 iovad->flush_cb = NULL;
75 iovad->entry_dtor = NULL;
78 int init_iova_flush_queue(struct iova_domain *iovad,
79 iova_flush_cb flush_cb, iova_entry_dtor entry_dtor)
81 struct iova_fq __percpu *queue;
84 atomic64_set(&iovad->fq_flush_start_cnt, 0);
85 atomic64_set(&iovad->fq_flush_finish_cnt, 0);
87 queue = alloc_percpu(struct iova_fq);
91 iovad->flush_cb = flush_cb;
92 iovad->entry_dtor = entry_dtor;
94 for_each_possible_cpu(cpu) {
97 fq = per_cpu_ptr(queue, cpu);
101 spin_lock_init(&fq->lock);
108 timer_setup(&iovad->fq_timer, fq_flush_timeout, 0);
109 atomic_set(&iovad->fq_timer_on, 0);
113 EXPORT_SYMBOL_GPL(init_iova_flush_queue);
115 static struct rb_node *
116 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
118 if (limit_pfn <= iovad->dma_32bit_pfn)
119 return iovad->cached32_node;
121 return iovad->cached_node;
125 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
127 if (new->pfn_hi < iovad->dma_32bit_pfn)
128 iovad->cached32_node = &new->node;
130 iovad->cached_node = &new->node;
134 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
136 struct iova *cached_iova;
138 cached_iova = rb_entry(iovad->cached32_node, struct iova, node);
139 if (free == cached_iova ||
140 (free->pfn_hi < iovad->dma_32bit_pfn &&
141 free->pfn_lo >= cached_iova->pfn_lo))
142 iovad->cached32_node = rb_next(&free->node);
144 if (free->pfn_lo < iovad->dma_32bit_pfn)
145 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
147 cached_iova = rb_entry(iovad->cached_node, struct iova, node);
148 if (free->pfn_lo >= cached_iova->pfn_lo)
149 iovad->cached_node = rb_next(&free->node);
152 /* Insert the iova into domain rbtree by holding writer lock */
154 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
155 struct rb_node *start)
157 struct rb_node **new, *parent = NULL;
159 new = (start) ? &start : &(root->rb_node);
160 /* Figure out where to put new node */
162 struct iova *this = rb_entry(*new, struct iova, node);
166 if (iova->pfn_lo < this->pfn_lo)
167 new = &((*new)->rb_left);
168 else if (iova->pfn_lo > this->pfn_lo)
169 new = &((*new)->rb_right);
171 WARN_ON(1); /* this should not happen */
175 /* Add new node and rebalance tree. */
176 rb_link_node(&iova->node, parent, new);
177 rb_insert_color(&iova->node, root);
180 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
181 unsigned long size, unsigned long limit_pfn,
182 struct iova *new, bool size_aligned)
184 struct rb_node *curr, *prev;
185 struct iova *curr_iova;
187 unsigned long new_pfn;
188 unsigned long align_mask = ~0UL;
191 align_mask <<= fls_long(size - 1);
193 /* Walk the tree backwards */
194 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
195 if (limit_pfn <= iovad->dma_32bit_pfn &&
196 size >= iovad->max32_alloc_size)
199 curr = __get_cached_rbnode(iovad, limit_pfn);
200 curr_iova = rb_entry(curr, struct iova, node);
202 limit_pfn = min(limit_pfn, curr_iova->pfn_lo);
203 new_pfn = (limit_pfn - size) & align_mask;
205 curr = rb_prev(curr);
206 curr_iova = rb_entry(curr, struct iova, node);
207 } while (curr && new_pfn <= curr_iova->pfn_hi);
209 if (limit_pfn < size || new_pfn < iovad->start_pfn) {
210 iovad->max32_alloc_size = size;
214 /* pfn_lo will point to size aligned address if size_aligned is set */
215 new->pfn_lo = new_pfn;
216 new->pfn_hi = new->pfn_lo + size - 1;
218 /* If we have 'prev', it's a valid place to start the insertion. */
219 iova_insert_rbtree(&iovad->rbroot, new, prev);
220 __cached_rbnode_insert_update(iovad, new);
222 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
226 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
230 static struct kmem_cache *iova_cache;
231 static unsigned int iova_cache_users;
232 static DEFINE_MUTEX(iova_cache_mutex);
234 struct iova *alloc_iova_mem(void)
236 return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
238 EXPORT_SYMBOL(alloc_iova_mem);
240 void free_iova_mem(struct iova *iova)
242 if (iova->pfn_lo != IOVA_ANCHOR)
243 kmem_cache_free(iova_cache, iova);
245 EXPORT_SYMBOL(free_iova_mem);
247 int iova_cache_get(void)
249 mutex_lock(&iova_cache_mutex);
250 if (!iova_cache_users) {
251 iova_cache = kmem_cache_create(
252 "iommu_iova", sizeof(struct iova), 0,
253 SLAB_HWCACHE_ALIGN, NULL);
255 mutex_unlock(&iova_cache_mutex);
256 pr_err("Couldn't create iova cache\n");
262 mutex_unlock(&iova_cache_mutex);
266 EXPORT_SYMBOL_GPL(iova_cache_get);
268 void iova_cache_put(void)
270 mutex_lock(&iova_cache_mutex);
271 if (WARN_ON(!iova_cache_users)) {
272 mutex_unlock(&iova_cache_mutex);
276 if (!iova_cache_users)
277 kmem_cache_destroy(iova_cache);
278 mutex_unlock(&iova_cache_mutex);
280 EXPORT_SYMBOL_GPL(iova_cache_put);
283 * alloc_iova - allocates an iova
284 * @iovad: - iova domain in question
285 * @size: - size of page frames to allocate
286 * @limit_pfn: - max limit address
287 * @size_aligned: - set if size_aligned address range is required
288 * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
289 * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
290 * flag is set then the allocated address iova->pfn_lo will be naturally
291 * aligned on roundup_power_of_two(size).
294 alloc_iova(struct iova_domain *iovad, unsigned long size,
295 unsigned long limit_pfn,
298 struct iova *new_iova;
301 new_iova = alloc_iova_mem();
305 ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
306 new_iova, size_aligned);
309 free_iova_mem(new_iova);
315 EXPORT_SYMBOL_GPL(alloc_iova);
318 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
320 struct rb_node *node = iovad->rbroot.rb_node;
322 assert_spin_locked(&iovad->iova_rbtree_lock);
325 struct iova *iova = rb_entry(node, struct iova, node);
327 if (pfn < iova->pfn_lo)
328 node = node->rb_left;
329 else if (pfn > iova->pfn_hi)
330 node = node->rb_right;
332 return iova; /* pfn falls within iova's range */
338 static void private_free_iova(struct iova_domain *iovad, struct iova *iova)
340 assert_spin_locked(&iovad->iova_rbtree_lock);
341 __cached_rbnode_delete_update(iovad, iova);
342 rb_erase(&iova->node, &iovad->rbroot);
347 * find_iova - finds an iova for a given pfn
348 * @iovad: - iova domain in question.
349 * @pfn: - page frame number
350 * This function finds and returns an iova belonging to the
351 * given doamin which matches the given pfn.
353 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
358 /* Take the lock so that no other thread is manipulating the rbtree */
359 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
360 iova = private_find_iova(iovad, pfn);
361 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
364 EXPORT_SYMBOL_GPL(find_iova);
367 * __free_iova - frees the given iova
368 * @iovad: iova domain in question.
369 * @iova: iova in question.
370 * Frees the given iova belonging to the giving domain
373 __free_iova(struct iova_domain *iovad, struct iova *iova)
377 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
378 private_free_iova(iovad, iova);
379 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
381 EXPORT_SYMBOL_GPL(__free_iova);
384 * free_iova - finds and frees the iova for a given pfn
385 * @iovad: - iova domain in question.
386 * @pfn: - pfn that is allocated previously
387 * This functions finds an iova for a given pfn and then
388 * frees the iova from that domain.
391 free_iova(struct iova_domain *iovad, unsigned long pfn)
393 struct iova *iova = find_iova(iovad, pfn);
396 __free_iova(iovad, iova);
399 EXPORT_SYMBOL_GPL(free_iova);
402 * alloc_iova_fast - allocates an iova from rcache
403 * @iovad: - iova domain in question
404 * @size: - size of page frames to allocate
405 * @limit_pfn: - max limit address
406 * @flush_rcache: - set to flush rcache on regular allocation failure
407 * This function tries to satisfy an iova allocation from the rcache,
408 * and falls back to regular allocation on failure. If regular allocation
409 * fails too and the flush_rcache flag is set then the rcache will be flushed.
412 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
413 unsigned long limit_pfn, bool flush_rcache)
415 unsigned long iova_pfn;
416 struct iova *new_iova;
418 iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
423 new_iova = alloc_iova(iovad, size, limit_pfn, true);
430 /* Try replenishing IOVAs by flushing rcache. */
431 flush_rcache = false;
432 for_each_online_cpu(cpu)
433 free_cpu_cached_iovas(cpu, iovad);
437 return new_iova->pfn_lo;
439 EXPORT_SYMBOL_GPL(alloc_iova_fast);
442 * free_iova_fast - free iova pfn range into rcache
443 * @iovad: - iova domain in question.
444 * @pfn: - pfn that is allocated previously
445 * @size: - # of pages in range
446 * This functions frees an iova range by trying to put it into the rcache,
447 * falling back to regular iova deallocation via free_iova() if this fails.
450 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
452 if (iova_rcache_insert(iovad, pfn, size))
455 free_iova(iovad, pfn);
457 EXPORT_SYMBOL_GPL(free_iova_fast);
459 #define fq_ring_for_each(i, fq) \
460 for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
462 static inline bool fq_full(struct iova_fq *fq)
464 assert_spin_locked(&fq->lock);
465 return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
468 static inline unsigned fq_ring_add(struct iova_fq *fq)
470 unsigned idx = fq->tail;
472 assert_spin_locked(&fq->lock);
474 fq->tail = (idx + 1) % IOVA_FQ_SIZE;
479 static void fq_ring_free(struct iova_domain *iovad, struct iova_fq *fq)
481 u64 counter = atomic64_read(&iovad->fq_flush_finish_cnt);
484 assert_spin_locked(&fq->lock);
486 fq_ring_for_each(idx, fq) {
488 if (fq->entries[idx].counter >= counter)
491 if (iovad->entry_dtor)
492 iovad->entry_dtor(fq->entries[idx].data);
494 free_iova_fast(iovad,
495 fq->entries[idx].iova_pfn,
496 fq->entries[idx].pages);
498 fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
502 static void iova_domain_flush(struct iova_domain *iovad)
504 atomic64_inc(&iovad->fq_flush_start_cnt);
505 iovad->flush_cb(iovad);
506 atomic64_inc(&iovad->fq_flush_finish_cnt);
509 static void fq_destroy_all_entries(struct iova_domain *iovad)
514 * This code runs when the iova_domain is being detroyed, so don't
515 * bother to free iovas, just call the entry_dtor on all remaining
518 if (!iovad->entry_dtor)
521 for_each_possible_cpu(cpu) {
522 struct iova_fq *fq = per_cpu_ptr(iovad->fq, cpu);
525 fq_ring_for_each(idx, fq)
526 iovad->entry_dtor(fq->entries[idx].data);
530 static void fq_flush_timeout(struct timer_list *t)
532 struct iova_domain *iovad = from_timer(iovad, t, fq_timer);
535 atomic_set(&iovad->fq_timer_on, 0);
536 iova_domain_flush(iovad);
538 for_each_possible_cpu(cpu) {
542 fq = per_cpu_ptr(iovad->fq, cpu);
543 spin_lock_irqsave(&fq->lock, flags);
544 fq_ring_free(iovad, fq);
545 spin_unlock_irqrestore(&fq->lock, flags);
549 void queue_iova(struct iova_domain *iovad,
550 unsigned long pfn, unsigned long pages,
553 struct iova_fq *fq = raw_cpu_ptr(iovad->fq);
557 spin_lock_irqsave(&fq->lock, flags);
560 * First remove all entries from the flush queue that have already been
561 * flushed out on another CPU. This makes the fq_full() check below less
564 fq_ring_free(iovad, fq);
567 iova_domain_flush(iovad);
568 fq_ring_free(iovad, fq);
571 idx = fq_ring_add(fq);
573 fq->entries[idx].iova_pfn = pfn;
574 fq->entries[idx].pages = pages;
575 fq->entries[idx].data = data;
576 fq->entries[idx].counter = atomic64_read(&iovad->fq_flush_start_cnt);
578 spin_unlock_irqrestore(&fq->lock, flags);
580 /* Avoid false sharing as much as possible. */
581 if (!atomic_read(&iovad->fq_timer_on) &&
582 !atomic_xchg(&iovad->fq_timer_on, 1))
583 mod_timer(&iovad->fq_timer,
584 jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
586 EXPORT_SYMBOL_GPL(queue_iova);
589 * put_iova_domain - destroys the iova doamin
590 * @iovad: - iova domain in question.
591 * All the iova's in that domain are destroyed.
593 void put_iova_domain(struct iova_domain *iovad)
595 struct iova *iova, *tmp;
597 free_iova_flush_queue(iovad);
598 free_iova_rcaches(iovad);
599 rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
602 EXPORT_SYMBOL_GPL(put_iova_domain);
605 __is_range_overlap(struct rb_node *node,
606 unsigned long pfn_lo, unsigned long pfn_hi)
608 struct iova *iova = rb_entry(node, struct iova, node);
610 if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
615 static inline struct iova *
616 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
620 iova = alloc_iova_mem();
622 iova->pfn_lo = pfn_lo;
623 iova->pfn_hi = pfn_hi;
630 __insert_new_range(struct iova_domain *iovad,
631 unsigned long pfn_lo, unsigned long pfn_hi)
635 iova = alloc_and_init_iova(pfn_lo, pfn_hi);
637 iova_insert_rbtree(&iovad->rbroot, iova, NULL);
643 __adjust_overlap_range(struct iova *iova,
644 unsigned long *pfn_lo, unsigned long *pfn_hi)
646 if (*pfn_lo < iova->pfn_lo)
647 iova->pfn_lo = *pfn_lo;
648 if (*pfn_hi > iova->pfn_hi)
649 *pfn_lo = iova->pfn_hi + 1;
653 * reserve_iova - reserves an iova in the given range
654 * @iovad: - iova domain pointer
655 * @pfn_lo: - lower page frame address
656 * @pfn_hi:- higher pfn adderss
657 * This function allocates reserves the address range from pfn_lo to pfn_hi so
658 * that this address is not dished out as part of alloc_iova.
661 reserve_iova(struct iova_domain *iovad,
662 unsigned long pfn_lo, unsigned long pfn_hi)
664 struct rb_node *node;
667 unsigned int overlap = 0;
669 /* Don't allow nonsensical pfns */
670 if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
673 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
674 for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
675 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
676 iova = rb_entry(node, struct iova, node);
677 __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
678 if ((pfn_lo >= iova->pfn_lo) &&
679 (pfn_hi <= iova->pfn_hi))
687 /* We are here either because this is the first reserver node
688 * or need to insert remaining non overlap addr range
690 iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
693 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
696 EXPORT_SYMBOL_GPL(reserve_iova);
699 * copy_reserved_iova - copies the reserved between domains
700 * @from: - source doamin from where to copy
701 * @to: - destination domin where to copy
702 * This function copies reserved iova's from one doamin to
706 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
709 struct rb_node *node;
711 spin_lock_irqsave(&from->iova_rbtree_lock, flags);
712 for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
713 struct iova *iova = rb_entry(node, struct iova, node);
714 struct iova *new_iova;
716 if (iova->pfn_lo == IOVA_ANCHOR)
719 new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
721 pr_err("Reserve iova range %lx@%lx failed\n",
722 iova->pfn_lo, iova->pfn_lo);
724 spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
726 EXPORT_SYMBOL_GPL(copy_reserved_iova);
729 split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
730 unsigned long pfn_lo, unsigned long pfn_hi)
733 struct iova *prev = NULL, *next = NULL;
735 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
736 if (iova->pfn_lo < pfn_lo) {
737 prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
741 if (iova->pfn_hi > pfn_hi) {
742 next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
747 __cached_rbnode_delete_update(iovad, iova);
748 rb_erase(&iova->node, &iovad->rbroot);
751 iova_insert_rbtree(&iovad->rbroot, prev, NULL);
752 iova->pfn_lo = pfn_lo;
755 iova_insert_rbtree(&iovad->rbroot, next, NULL);
756 iova->pfn_hi = pfn_hi;
758 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
763 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
770 * Magazine caches for IOVA ranges. For an introduction to magazines,
771 * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
772 * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
773 * For simplicity, we use a static magazine size and don't implement the
774 * dynamic size tuning described in the paper.
777 #define IOVA_MAG_SIZE 128
779 struct iova_magazine {
781 unsigned long pfns[IOVA_MAG_SIZE];
784 struct iova_cpu_rcache {
786 struct iova_magazine *loaded;
787 struct iova_magazine *prev;
790 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
792 return kzalloc(sizeof(struct iova_magazine), flags);
795 static void iova_magazine_free(struct iova_magazine *mag)
801 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
809 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
811 for (i = 0 ; i < mag->size; ++i) {
812 struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
817 private_free_iova(iovad, iova);
820 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
825 static bool iova_magazine_full(struct iova_magazine *mag)
827 return (mag && mag->size == IOVA_MAG_SIZE);
830 static bool iova_magazine_empty(struct iova_magazine *mag)
832 return (!mag || mag->size == 0);
835 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
836 unsigned long limit_pfn)
841 BUG_ON(iova_magazine_empty(mag));
843 /* Only fall back to the rbtree if we have no suitable pfns at all */
844 for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
848 /* Swap it to pop it */
850 mag->pfns[i] = mag->pfns[--mag->size];
855 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
857 BUG_ON(iova_magazine_full(mag));
859 mag->pfns[mag->size++] = pfn;
862 static void init_iova_rcaches(struct iova_domain *iovad)
864 struct iova_cpu_rcache *cpu_rcache;
865 struct iova_rcache *rcache;
869 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
870 rcache = &iovad->rcaches[i];
871 spin_lock_init(&rcache->lock);
872 rcache->depot_size = 0;
873 rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
874 if (WARN_ON(!rcache->cpu_rcaches))
876 for_each_possible_cpu(cpu) {
877 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
878 spin_lock_init(&cpu_rcache->lock);
879 cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
880 cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
886 * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
887 * return true on success. Can fail if rcache is full and we can't free
888 * space, and free_iova() (our only caller) will then return the IOVA
889 * range to the rbtree instead.
891 static bool __iova_rcache_insert(struct iova_domain *iovad,
892 struct iova_rcache *rcache,
893 unsigned long iova_pfn)
895 struct iova_magazine *mag_to_free = NULL;
896 struct iova_cpu_rcache *cpu_rcache;
897 bool can_insert = false;
900 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
901 spin_lock_irqsave(&cpu_rcache->lock, flags);
903 if (!iova_magazine_full(cpu_rcache->loaded)) {
905 } else if (!iova_magazine_full(cpu_rcache->prev)) {
906 swap(cpu_rcache->prev, cpu_rcache->loaded);
909 struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
912 spin_lock(&rcache->lock);
913 if (rcache->depot_size < MAX_GLOBAL_MAGS) {
914 rcache->depot[rcache->depot_size++] =
917 mag_to_free = cpu_rcache->loaded;
919 spin_unlock(&rcache->lock);
921 cpu_rcache->loaded = new_mag;
927 iova_magazine_push(cpu_rcache->loaded, iova_pfn);
929 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
932 iova_magazine_free_pfns(mag_to_free, iovad);
933 iova_magazine_free(mag_to_free);
939 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
942 unsigned int log_size = order_base_2(size);
944 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
947 return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
951 * Caller wants to allocate a new IOVA range from 'rcache'. If we can
952 * satisfy the request, return a matching non-NULL range and remove
953 * it from the 'rcache'.
955 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
956 unsigned long limit_pfn)
958 struct iova_cpu_rcache *cpu_rcache;
959 unsigned long iova_pfn = 0;
960 bool has_pfn = false;
963 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
964 spin_lock_irqsave(&cpu_rcache->lock, flags);
966 if (!iova_magazine_empty(cpu_rcache->loaded)) {
968 } else if (!iova_magazine_empty(cpu_rcache->prev)) {
969 swap(cpu_rcache->prev, cpu_rcache->loaded);
972 spin_lock(&rcache->lock);
973 if (rcache->depot_size > 0) {
974 iova_magazine_free(cpu_rcache->loaded);
975 cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
978 spin_unlock(&rcache->lock);
982 iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
984 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
990 * Try to satisfy IOVA allocation range from rcache. Fail if requested
991 * size is too big or the DMA limit we are given isn't satisfied by the
992 * top element in the magazine.
994 static unsigned long iova_rcache_get(struct iova_domain *iovad,
996 unsigned long limit_pfn)
998 unsigned int log_size = order_base_2(size);
1000 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
1003 return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
1007 * free rcache data structures.
1009 static void free_iova_rcaches(struct iova_domain *iovad)
1011 struct iova_rcache *rcache;
1012 struct iova_cpu_rcache *cpu_rcache;
1016 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1017 rcache = &iovad->rcaches[i];
1018 for_each_possible_cpu(cpu) {
1019 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1020 iova_magazine_free(cpu_rcache->loaded);
1021 iova_magazine_free(cpu_rcache->prev);
1023 free_percpu(rcache->cpu_rcaches);
1024 for (j = 0; j < rcache->depot_size; ++j)
1025 iova_magazine_free(rcache->depot[j]);
1030 * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
1032 void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
1034 struct iova_cpu_rcache *cpu_rcache;
1035 struct iova_rcache *rcache;
1036 unsigned long flags;
1039 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1040 rcache = &iovad->rcaches[i];
1041 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1042 spin_lock_irqsave(&cpu_rcache->lock, flags);
1043 iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
1044 iova_magazine_free_pfns(cpu_rcache->prev, iovad);
1045 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
1049 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
1050 MODULE_LICENSE("GPL");