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>
14 #include <linux/workqueue.h>
16 /* The anchor node sits above the top of the usable address space */
17 #define IOVA_ANCHOR ~0UL
19 #define IOVA_RANGE_CACHE_MAX_SIZE 6 /* log of max cached IOVA range size (in pages) */
21 static bool iova_rcache_insert(struct iova_domain *iovad,
24 static unsigned long iova_rcache_get(struct iova_domain *iovad,
26 unsigned long limit_pfn);
27 static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad);
28 static void free_iova_rcaches(struct iova_domain *iovad);
30 unsigned long iova_rcache_range(void)
32 return PAGE_SIZE << (IOVA_RANGE_CACHE_MAX_SIZE - 1);
35 static int iova_cpuhp_dead(unsigned int cpu, struct hlist_node *node)
37 struct iova_domain *iovad;
39 iovad = hlist_entry_safe(node, struct iova_domain, cpuhp_dead);
41 free_cpu_cached_iovas(cpu, iovad);
45 static void free_global_cached_iovas(struct iova_domain *iovad);
47 static struct iova *to_iova(struct rb_node *node)
49 return rb_entry(node, struct iova, node);
53 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
54 unsigned long start_pfn)
57 * IOVA granularity will normally be equal to the smallest
58 * supported IOMMU page size; both *must* be capable of
59 * representing individual CPU pages exactly.
61 BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
63 spin_lock_init(&iovad->iova_rbtree_lock);
64 iovad->rbroot = RB_ROOT;
65 iovad->cached_node = &iovad->anchor.node;
66 iovad->cached32_node = &iovad->anchor.node;
67 iovad->granule = granule;
68 iovad->start_pfn = start_pfn;
69 iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
70 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
71 iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
72 rb_link_node(&iovad->anchor.node, NULL, &iovad->rbroot.rb_node);
73 rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
75 EXPORT_SYMBOL_GPL(init_iova_domain);
77 static struct rb_node *
78 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
80 if (limit_pfn <= iovad->dma_32bit_pfn)
81 return iovad->cached32_node;
83 return iovad->cached_node;
87 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
89 if (new->pfn_hi < iovad->dma_32bit_pfn)
90 iovad->cached32_node = &new->node;
92 iovad->cached_node = &new->node;
96 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
98 struct iova *cached_iova;
100 cached_iova = to_iova(iovad->cached32_node);
101 if (free == cached_iova ||
102 (free->pfn_hi < iovad->dma_32bit_pfn &&
103 free->pfn_lo >= cached_iova->pfn_lo))
104 iovad->cached32_node = rb_next(&free->node);
106 if (free->pfn_lo < iovad->dma_32bit_pfn)
107 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
109 cached_iova = to_iova(iovad->cached_node);
110 if (free->pfn_lo >= cached_iova->pfn_lo)
111 iovad->cached_node = rb_next(&free->node);
114 static struct rb_node *iova_find_limit(struct iova_domain *iovad, unsigned long limit_pfn)
116 struct rb_node *node, *next;
118 * Ideally what we'd like to judge here is whether limit_pfn is close
119 * enough to the highest-allocated IOVA that starting the allocation
120 * walk from the anchor node will be quicker than this initial work to
121 * find an exact starting point (especially if that ends up being the
122 * anchor node anyway). This is an incredibly crude approximation which
123 * only really helps the most likely case, but is at least trivially easy.
125 if (limit_pfn > iovad->dma_32bit_pfn)
126 return &iovad->anchor.node;
128 node = iovad->rbroot.rb_node;
129 while (to_iova(node)->pfn_hi < limit_pfn)
130 node = node->rb_right;
133 while (node->rb_left && to_iova(node->rb_left)->pfn_lo >= limit_pfn)
134 node = node->rb_left;
139 next = node->rb_left;
140 while (next->rb_right) {
141 next = next->rb_right;
142 if (to_iova(next)->pfn_lo >= limit_pfn) {
151 /* Insert the iova into domain rbtree by holding writer lock */
153 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
154 struct rb_node *start)
156 struct rb_node **new, *parent = NULL;
158 new = (start) ? &start : &(root->rb_node);
159 /* Figure out where to put new node */
161 struct iova *this = to_iova(*new);
165 if (iova->pfn_lo < this->pfn_lo)
166 new = &((*new)->rb_left);
167 else if (iova->pfn_lo > this->pfn_lo)
168 new = &((*new)->rb_right);
170 WARN_ON(1); /* this should not happen */
174 /* Add new node and rebalance tree. */
175 rb_link_node(&iova->node, parent, new);
176 rb_insert_color(&iova->node, root);
179 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
180 unsigned long size, unsigned long limit_pfn,
181 struct iova *new, bool size_aligned)
183 struct rb_node *curr, *prev;
184 struct iova *curr_iova;
186 unsigned long new_pfn, retry_pfn;
187 unsigned long align_mask = ~0UL;
188 unsigned long high_pfn = limit_pfn, low_pfn = iovad->start_pfn;
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 = to_iova(curr);
201 retry_pfn = curr_iova->pfn_hi;
205 high_pfn = min(high_pfn, curr_iova->pfn_lo);
206 new_pfn = (high_pfn - size) & align_mask;
208 curr = rb_prev(curr);
209 curr_iova = to_iova(curr);
210 } while (curr && new_pfn <= curr_iova->pfn_hi && new_pfn >= low_pfn);
212 if (high_pfn < size || new_pfn < low_pfn) {
213 if (low_pfn == iovad->start_pfn && retry_pfn < limit_pfn) {
214 high_pfn = limit_pfn;
215 low_pfn = retry_pfn + 1;
216 curr = iova_find_limit(iovad, limit_pfn);
217 curr_iova = to_iova(curr);
220 iovad->max32_alloc_size = size;
224 /* pfn_lo will point to size aligned address if size_aligned is set */
225 new->pfn_lo = new_pfn;
226 new->pfn_hi = new->pfn_lo + size - 1;
228 /* If we have 'prev', it's a valid place to start the insertion. */
229 iova_insert_rbtree(&iovad->rbroot, new, prev);
230 __cached_rbnode_insert_update(iovad, new);
232 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
236 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
240 static struct kmem_cache *iova_cache;
241 static unsigned int iova_cache_users;
242 static DEFINE_MUTEX(iova_cache_mutex);
244 static struct iova *alloc_iova_mem(void)
246 return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
249 static void free_iova_mem(struct iova *iova)
251 if (iova->pfn_lo != IOVA_ANCHOR)
252 kmem_cache_free(iova_cache, iova);
255 int iova_cache_get(void)
257 mutex_lock(&iova_cache_mutex);
258 if (!iova_cache_users) {
261 ret = cpuhp_setup_state_multi(CPUHP_IOMMU_IOVA_DEAD, "iommu/iova:dead", NULL,
264 mutex_unlock(&iova_cache_mutex);
265 pr_err("Couldn't register cpuhp handler\n");
269 iova_cache = kmem_cache_create(
270 "iommu_iova", sizeof(struct iova), 0,
271 SLAB_HWCACHE_ALIGN, NULL);
273 cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
274 mutex_unlock(&iova_cache_mutex);
275 pr_err("Couldn't create iova cache\n");
281 mutex_unlock(&iova_cache_mutex);
285 EXPORT_SYMBOL_GPL(iova_cache_get);
287 void iova_cache_put(void)
289 mutex_lock(&iova_cache_mutex);
290 if (WARN_ON(!iova_cache_users)) {
291 mutex_unlock(&iova_cache_mutex);
295 if (!iova_cache_users) {
296 cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
297 kmem_cache_destroy(iova_cache);
299 mutex_unlock(&iova_cache_mutex);
301 EXPORT_SYMBOL_GPL(iova_cache_put);
304 * alloc_iova - allocates an iova
305 * @iovad: - iova domain in question
306 * @size: - size of page frames to allocate
307 * @limit_pfn: - max limit address
308 * @size_aligned: - set if size_aligned address range is required
309 * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
310 * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
311 * flag is set then the allocated address iova->pfn_lo will be naturally
312 * aligned on roundup_power_of_two(size).
315 alloc_iova(struct iova_domain *iovad, unsigned long size,
316 unsigned long limit_pfn,
319 struct iova *new_iova;
322 new_iova = alloc_iova_mem();
326 ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
327 new_iova, size_aligned);
330 free_iova_mem(new_iova);
336 EXPORT_SYMBOL_GPL(alloc_iova);
339 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
341 struct rb_node *node = iovad->rbroot.rb_node;
343 assert_spin_locked(&iovad->iova_rbtree_lock);
346 struct iova *iova = to_iova(node);
348 if (pfn < iova->pfn_lo)
349 node = node->rb_left;
350 else if (pfn > iova->pfn_hi)
351 node = node->rb_right;
353 return iova; /* pfn falls within iova's range */
359 static void remove_iova(struct iova_domain *iovad, struct iova *iova)
361 assert_spin_locked(&iovad->iova_rbtree_lock);
362 __cached_rbnode_delete_update(iovad, iova);
363 rb_erase(&iova->node, &iovad->rbroot);
367 * find_iova - finds an iova for a given pfn
368 * @iovad: - iova domain in question.
369 * @pfn: - page frame number
370 * This function finds and returns an iova belonging to the
371 * given domain which matches the given pfn.
373 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
378 /* Take the lock so that no other thread is manipulating the rbtree */
379 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
380 iova = private_find_iova(iovad, pfn);
381 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
384 EXPORT_SYMBOL_GPL(find_iova);
387 * __free_iova - frees the given iova
388 * @iovad: iova domain in question.
389 * @iova: iova in question.
390 * Frees the given iova belonging to the giving domain
393 __free_iova(struct iova_domain *iovad, struct iova *iova)
397 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
398 remove_iova(iovad, iova);
399 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
402 EXPORT_SYMBOL_GPL(__free_iova);
405 * free_iova - finds and frees the iova for a given pfn
406 * @iovad: - iova domain in question.
407 * @pfn: - pfn that is allocated previously
408 * This functions finds an iova for a given pfn and then
409 * frees the iova from that domain.
412 free_iova(struct iova_domain *iovad, unsigned long pfn)
417 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
418 iova = private_find_iova(iovad, pfn);
420 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
423 remove_iova(iovad, iova);
424 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
427 EXPORT_SYMBOL_GPL(free_iova);
430 * alloc_iova_fast - allocates an iova from rcache
431 * @iovad: - iova domain in question
432 * @size: - size of page frames to allocate
433 * @limit_pfn: - max limit address
434 * @flush_rcache: - set to flush rcache on regular allocation failure
435 * This function tries to satisfy an iova allocation from the rcache,
436 * and falls back to regular allocation on failure. If regular allocation
437 * fails too and the flush_rcache flag is set then the rcache will be flushed.
440 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
441 unsigned long limit_pfn, bool flush_rcache)
443 unsigned long iova_pfn;
444 struct iova *new_iova;
447 * Freeing non-power-of-two-sized allocations back into the IOVA caches
448 * will come back to bite us badly, so we have to waste a bit of space
449 * rounding up anything cacheable to make sure that can't happen. The
450 * order of the unadjusted size will still match upon freeing.
452 if (size < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1)))
453 size = roundup_pow_of_two(size);
455 iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
460 new_iova = alloc_iova(iovad, size, limit_pfn, true);
467 /* Try replenishing IOVAs by flushing rcache. */
468 flush_rcache = false;
469 for_each_online_cpu(cpu)
470 free_cpu_cached_iovas(cpu, iovad);
471 free_global_cached_iovas(iovad);
475 return new_iova->pfn_lo;
477 EXPORT_SYMBOL_GPL(alloc_iova_fast);
480 * free_iova_fast - free iova pfn range into rcache
481 * @iovad: - iova domain in question.
482 * @pfn: - pfn that is allocated previously
483 * @size: - # of pages in range
484 * This functions frees an iova range by trying to put it into the rcache,
485 * falling back to regular iova deallocation via free_iova() if this fails.
488 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
490 if (iova_rcache_insert(iovad, pfn, size))
493 free_iova(iovad, pfn);
495 EXPORT_SYMBOL_GPL(free_iova_fast);
497 static void iova_domain_free_rcaches(struct iova_domain *iovad)
499 cpuhp_state_remove_instance_nocalls(CPUHP_IOMMU_IOVA_DEAD,
501 free_iova_rcaches(iovad);
505 * put_iova_domain - destroys the iova domain
506 * @iovad: - iova domain in question.
507 * All the iova's in that domain are destroyed.
509 void put_iova_domain(struct iova_domain *iovad)
511 struct iova *iova, *tmp;
514 iova_domain_free_rcaches(iovad);
516 rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
519 EXPORT_SYMBOL_GPL(put_iova_domain);
522 __is_range_overlap(struct rb_node *node,
523 unsigned long pfn_lo, unsigned long pfn_hi)
525 struct iova *iova = to_iova(node);
527 if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
532 static inline struct iova *
533 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
537 iova = alloc_iova_mem();
539 iova->pfn_lo = pfn_lo;
540 iova->pfn_hi = pfn_hi;
547 __insert_new_range(struct iova_domain *iovad,
548 unsigned long pfn_lo, unsigned long pfn_hi)
552 iova = alloc_and_init_iova(pfn_lo, pfn_hi);
554 iova_insert_rbtree(&iovad->rbroot, iova, NULL);
560 __adjust_overlap_range(struct iova *iova,
561 unsigned long *pfn_lo, unsigned long *pfn_hi)
563 if (*pfn_lo < iova->pfn_lo)
564 iova->pfn_lo = *pfn_lo;
565 if (*pfn_hi > iova->pfn_hi)
566 *pfn_lo = iova->pfn_hi + 1;
570 * reserve_iova - reserves an iova in the given range
571 * @iovad: - iova domain pointer
572 * @pfn_lo: - lower page frame address
573 * @pfn_hi:- higher pfn adderss
574 * This function allocates reserves the address range from pfn_lo to pfn_hi so
575 * that this address is not dished out as part of alloc_iova.
578 reserve_iova(struct iova_domain *iovad,
579 unsigned long pfn_lo, unsigned long pfn_hi)
581 struct rb_node *node;
584 unsigned int overlap = 0;
586 /* Don't allow nonsensical pfns */
587 if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
590 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
591 for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
592 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
593 iova = to_iova(node);
594 __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
595 if ((pfn_lo >= iova->pfn_lo) &&
596 (pfn_hi <= iova->pfn_hi))
604 /* We are here either because this is the first reserver node
605 * or need to insert remaining non overlap addr range
607 iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
610 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
613 EXPORT_SYMBOL_GPL(reserve_iova);
616 * Magazine caches for IOVA ranges. For an introduction to magazines,
617 * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
618 * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
619 * For simplicity, we use a static magazine size and don't implement the
620 * dynamic size tuning described in the paper.
624 * As kmalloc's buffer size is fixed to power of 2, 127 is chosen to
625 * assure size of 'iova_magazine' to be 1024 bytes, so that no memory
626 * will be wasted. Since only full magazines are inserted into the depot,
627 * we don't need to waste PFN capacity on a separate list head either.
629 #define IOVA_MAG_SIZE 127
631 #define IOVA_DEPOT_DELAY msecs_to_jiffies(100)
633 struct iova_magazine {
636 struct iova_magazine *next;
638 unsigned long pfns[IOVA_MAG_SIZE];
640 static_assert(!(sizeof(struct iova_magazine) & (sizeof(struct iova_magazine) - 1)));
642 struct iova_cpu_rcache {
644 struct iova_magazine *loaded;
645 struct iova_magazine *prev;
650 unsigned int depot_size;
651 struct iova_magazine *depot;
652 struct iova_cpu_rcache __percpu *cpu_rcaches;
653 struct iova_domain *iovad;
654 struct delayed_work work;
657 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
659 struct iova_magazine *mag;
661 mag = kmalloc(sizeof(*mag), flags);
668 static void iova_magazine_free(struct iova_magazine *mag)
674 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
679 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
681 for (i = 0 ; i < mag->size; ++i) {
682 struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
687 remove_iova(iovad, iova);
691 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
696 static bool iova_magazine_full(struct iova_magazine *mag)
698 return mag->size == IOVA_MAG_SIZE;
701 static bool iova_magazine_empty(struct iova_magazine *mag)
703 return mag->size == 0;
706 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
707 unsigned long limit_pfn)
712 /* Only fall back to the rbtree if we have no suitable pfns at all */
713 for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
717 /* Swap it to pop it */
719 mag->pfns[i] = mag->pfns[--mag->size];
724 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
726 mag->pfns[mag->size++] = pfn;
729 static struct iova_magazine *iova_depot_pop(struct iova_rcache *rcache)
731 struct iova_magazine *mag = rcache->depot;
733 rcache->depot = mag->next;
734 mag->size = IOVA_MAG_SIZE;
735 rcache->depot_size--;
739 static void iova_depot_push(struct iova_rcache *rcache, struct iova_magazine *mag)
741 mag->next = rcache->depot;
743 rcache->depot_size++;
746 static void iova_depot_work_func(struct work_struct *work)
748 struct iova_rcache *rcache = container_of(work, typeof(*rcache), work.work);
749 struct iova_magazine *mag = NULL;
752 spin_lock_irqsave(&rcache->lock, flags);
753 if (rcache->depot_size > num_online_cpus())
754 mag = iova_depot_pop(rcache);
755 spin_unlock_irqrestore(&rcache->lock, flags);
758 iova_magazine_free_pfns(mag, rcache->iovad);
759 iova_magazine_free(mag);
760 schedule_delayed_work(&rcache->work, IOVA_DEPOT_DELAY);
764 int iova_domain_init_rcaches(struct iova_domain *iovad)
769 iovad->rcaches = kcalloc(IOVA_RANGE_CACHE_MAX_SIZE,
770 sizeof(struct iova_rcache),
775 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
776 struct iova_cpu_rcache *cpu_rcache;
777 struct iova_rcache *rcache;
779 rcache = &iovad->rcaches[i];
780 spin_lock_init(&rcache->lock);
781 rcache->iovad = iovad;
782 INIT_DELAYED_WORK(&rcache->work, iova_depot_work_func);
783 rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache),
785 if (!rcache->cpu_rcaches) {
789 for_each_possible_cpu(cpu) {
790 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
792 spin_lock_init(&cpu_rcache->lock);
793 cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
794 cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
795 if (!cpu_rcache->loaded || !cpu_rcache->prev) {
802 ret = cpuhp_state_add_instance_nocalls(CPUHP_IOMMU_IOVA_DEAD,
809 free_iova_rcaches(iovad);
812 EXPORT_SYMBOL_GPL(iova_domain_init_rcaches);
815 * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
816 * return true on success. Can fail if rcache is full and we can't free
817 * space, and free_iova() (our only caller) will then return the IOVA
818 * range to the rbtree instead.
820 static bool __iova_rcache_insert(struct iova_domain *iovad,
821 struct iova_rcache *rcache,
822 unsigned long iova_pfn)
824 struct iova_cpu_rcache *cpu_rcache;
825 bool can_insert = false;
828 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
829 spin_lock_irqsave(&cpu_rcache->lock, flags);
831 if (!iova_magazine_full(cpu_rcache->loaded)) {
833 } else if (!iova_magazine_full(cpu_rcache->prev)) {
834 swap(cpu_rcache->prev, cpu_rcache->loaded);
837 struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
840 spin_lock(&rcache->lock);
841 iova_depot_push(rcache, cpu_rcache->loaded);
842 spin_unlock(&rcache->lock);
843 schedule_delayed_work(&rcache->work, IOVA_DEPOT_DELAY);
845 cpu_rcache->loaded = new_mag;
851 iova_magazine_push(cpu_rcache->loaded, iova_pfn);
853 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
858 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
861 unsigned int log_size = order_base_2(size);
863 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
866 return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
870 * Caller wants to allocate a new IOVA range from 'rcache'. If we can
871 * satisfy the request, return a matching non-NULL range and remove
872 * it from the 'rcache'.
874 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
875 unsigned long limit_pfn)
877 struct iova_cpu_rcache *cpu_rcache;
878 unsigned long iova_pfn = 0;
879 bool has_pfn = false;
882 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
883 spin_lock_irqsave(&cpu_rcache->lock, flags);
885 if (!iova_magazine_empty(cpu_rcache->loaded)) {
887 } else if (!iova_magazine_empty(cpu_rcache->prev)) {
888 swap(cpu_rcache->prev, cpu_rcache->loaded);
891 spin_lock(&rcache->lock);
893 iova_magazine_free(cpu_rcache->loaded);
894 cpu_rcache->loaded = iova_depot_pop(rcache);
897 spin_unlock(&rcache->lock);
901 iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
903 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
909 * Try to satisfy IOVA allocation range from rcache. Fail if requested
910 * size is too big or the DMA limit we are given isn't satisfied by the
911 * top element in the magazine.
913 static unsigned long iova_rcache_get(struct iova_domain *iovad,
915 unsigned long limit_pfn)
917 unsigned int log_size = order_base_2(size);
919 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
922 return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
926 * free rcache data structures.
928 static void free_iova_rcaches(struct iova_domain *iovad)
930 struct iova_rcache *rcache;
931 struct iova_cpu_rcache *cpu_rcache;
934 for (int i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
935 rcache = &iovad->rcaches[i];
936 if (!rcache->cpu_rcaches)
938 for_each_possible_cpu(cpu) {
939 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
940 iova_magazine_free(cpu_rcache->loaded);
941 iova_magazine_free(cpu_rcache->prev);
943 free_percpu(rcache->cpu_rcaches);
944 cancel_delayed_work_sync(&rcache->work);
945 while (rcache->depot)
946 iova_magazine_free(iova_depot_pop(rcache));
949 kfree(iovad->rcaches);
950 iovad->rcaches = NULL;
954 * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
956 static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
958 struct iova_cpu_rcache *cpu_rcache;
959 struct iova_rcache *rcache;
963 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
964 rcache = &iovad->rcaches[i];
965 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
966 spin_lock_irqsave(&cpu_rcache->lock, flags);
967 iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
968 iova_magazine_free_pfns(cpu_rcache->prev, iovad);
969 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
974 * free all the IOVA ranges of global cache
976 static void free_global_cached_iovas(struct iova_domain *iovad)
978 struct iova_rcache *rcache;
981 for (int i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
982 rcache = &iovad->rcaches[i];
983 spin_lock_irqsave(&rcache->lock, flags);
984 while (rcache->depot) {
985 struct iova_magazine *mag = iova_depot_pop(rcache);
987 iova_magazine_free_pfns(mag, iovad);
988 iova_magazine_free(mag);
990 spin_unlock_irqrestore(&rcache->lock, flags);
993 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
994 MODULE_LICENSE("GPL");