1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright © 2006-2009, Intel Corporation.
4 *
5 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
6 */
7
8#include <linux/iova.h>
9#include <linux/kmemleak.h>
10#include <linux/module.h>
11#include <linux/slab.h>
12#include <linux/smp.h>
13#include <linux/bitops.h>
14#include <linux/cpu.h>
15#include <linux/workqueue.h>
16
17/* The anchor node sits above the top of the usable address space */
18#define IOVA_ANCHOR ~0UL
19
20#define IOVA_RANGE_CACHE_MAX_SIZE 6 /* log of max cached IOVA range size (in pages) */
21
22static bool iova_rcache_insert(struct iova_domain *iovad,
23 unsigned long pfn,
24 unsigned long size);
25static unsigned long iova_rcache_get(struct iova_domain *iovad,
26 unsigned long size,
27 unsigned long limit_pfn);
28static void free_iova_rcaches(struct iova_domain *iovad);
29static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad);
30static void free_global_cached_iovas(struct iova_domain *iovad);
31
32static struct iova *to_iova(struct rb_node *node)
33{
34 return rb_entry(node, struct iova, node);
35}
36
37void
38init_iova_domain(struct iova_domain *iovad, unsigned long granule,
39 unsigned long start_pfn)
40{
41 /*
42 * IOVA granularity will normally be equal to the smallest
43 * supported IOMMU page size; both *must* be capable of
44 * representing individual CPU pages exactly.
45 */
46 BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
47
48 spin_lock_init(&iovad->iova_rbtree_lock);
49 iovad->rbroot = RB_ROOT;
50 iovad->cached_node = &iovad->anchor.node;
51 iovad->cached32_node = &iovad->anchor.node;
52 iovad->granule = granule;
53 iovad->start_pfn = start_pfn;
54 iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
55 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
56 iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
57 rb_link_node(node: &iovad->anchor.node, NULL, rb_link: &iovad->rbroot.rb_node);
58 rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
59}
60EXPORT_SYMBOL_GPL(init_iova_domain);
61
62static struct rb_node *
63__get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
64{
65 if (limit_pfn <= iovad->dma_32bit_pfn)
66 return iovad->cached32_node;
67
68 return iovad->cached_node;
69}
70
71static void
72__cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
73{
74 if (new->pfn_hi < iovad->dma_32bit_pfn)
75 iovad->cached32_node = &new->node;
76 else
77 iovad->cached_node = &new->node;
78}
79
80static void
81__cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
82{
83 struct iova *cached_iova;
84
85 cached_iova = to_iova(node: iovad->cached32_node);
86 if (free == cached_iova ||
87 (free->pfn_hi < iovad->dma_32bit_pfn &&
88 free->pfn_lo >= cached_iova->pfn_lo))
89 iovad->cached32_node = rb_next(&free->node);
90
91 if (free->pfn_lo < iovad->dma_32bit_pfn)
92 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
93
94 cached_iova = to_iova(node: iovad->cached_node);
95 if (free->pfn_lo >= cached_iova->pfn_lo)
96 iovad->cached_node = rb_next(&free->node);
97}
98
99static struct rb_node *iova_find_limit(struct iova_domain *iovad, unsigned long limit_pfn)
100{
101 struct rb_node *node, *next;
102 /*
103 * Ideally what we'd like to judge here is whether limit_pfn is close
104 * enough to the highest-allocated IOVA that starting the allocation
105 * walk from the anchor node will be quicker than this initial work to
106 * find an exact starting point (especially if that ends up being the
107 * anchor node anyway). This is an incredibly crude approximation which
108 * only really helps the most likely case, but is at least trivially easy.
109 */
110 if (limit_pfn > iovad->dma_32bit_pfn)
111 return &iovad->anchor.node;
112
113 node = iovad->rbroot.rb_node;
114 while (to_iova(node)->pfn_hi < limit_pfn)
115 node = node->rb_right;
116
117search_left:
118 while (node->rb_left && to_iova(node: node->rb_left)->pfn_lo >= limit_pfn)
119 node = node->rb_left;
120
121 if (!node->rb_left)
122 return node;
123
124 next = node->rb_left;
125 while (next->rb_right) {
126 next = next->rb_right;
127 if (to_iova(node: next)->pfn_lo >= limit_pfn) {
128 node = next;
129 goto search_left;
130 }
131 }
132
133 return node;
134}
135
136/* Insert the iova into domain rbtree by holding writer lock */
137static void
138iova_insert_rbtree(struct rb_root *root, struct iova *iova,
139 struct rb_node *start)
140{
141 struct rb_node **new, *parent = NULL;
142
143 new = (start) ? &start : &(root->rb_node);
144 /* Figure out where to put new node */
145 while (*new) {
146 struct iova *this = to_iova(node: *new);
147
148 parent = *new;
149
150 if (iova->pfn_lo < this->pfn_lo)
151 new = &((*new)->rb_left);
152 else if (iova->pfn_lo > this->pfn_lo)
153 new = &((*new)->rb_right);
154 else {
155 WARN_ON(1); /* this should not happen */
156 return;
157 }
158 }
159 /* Add new node and rebalance tree. */
160 rb_link_node(node: &iova->node, parent, rb_link: new);
161 rb_insert_color(&iova->node, root);
162}
163
164static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
165 unsigned long size, unsigned long limit_pfn,
166 struct iova *new, bool size_aligned)
167{
168 struct rb_node *curr, *prev;
169 struct iova *curr_iova;
170 unsigned long flags;
171 unsigned long new_pfn, retry_pfn;
172 unsigned long align_mask = ~0UL;
173 unsigned long high_pfn = limit_pfn, low_pfn = iovad->start_pfn;
174
175 if (size_aligned)
176 align_mask <<= fls_long(l: size - 1);
177
178 /* Walk the tree backwards */
179 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
180 if (limit_pfn <= iovad->dma_32bit_pfn &&
181 size >= iovad->max32_alloc_size)
182 goto iova32_full;
183
184 curr = __get_cached_rbnode(iovad, limit_pfn);
185 curr_iova = to_iova(node: curr);
186 retry_pfn = curr_iova->pfn_hi;
187
188retry:
189 do {
190 high_pfn = min(high_pfn, curr_iova->pfn_lo);
191 new_pfn = (high_pfn - size) & align_mask;
192 prev = curr;
193 curr = rb_prev(curr);
194 curr_iova = to_iova(node: curr);
195 } while (curr && new_pfn <= curr_iova->pfn_hi && new_pfn >= low_pfn);
196
197 if (high_pfn < size || new_pfn < low_pfn) {
198 if (low_pfn == iovad->start_pfn && retry_pfn < limit_pfn) {
199 high_pfn = limit_pfn;
200 low_pfn = retry_pfn + 1;
201 curr = iova_find_limit(iovad, limit_pfn);
202 curr_iova = to_iova(node: curr);
203 goto retry;
204 }
205 iovad->max32_alloc_size = size;
206 goto iova32_full;
207 }
208
209 /* pfn_lo will point to size aligned address if size_aligned is set */
210 new->pfn_lo = new_pfn;
211 new->pfn_hi = new->pfn_lo + size - 1;
212
213 /* If we have 'prev', it's a valid place to start the insertion. */
214 iova_insert_rbtree(root: &iovad->rbroot, iova: new, start: prev);
215 __cached_rbnode_insert_update(iovad, new);
216
217 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
218 return 0;
219
220iova32_full:
221 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
222 return -ENOMEM;
223}
224
225static struct kmem_cache *iova_cache;
226static unsigned int iova_cache_users;
227static DEFINE_MUTEX(iova_cache_mutex);
228
229static struct iova *alloc_iova_mem(void)
230{
231 return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
232}
233
234static void free_iova_mem(struct iova *iova)
235{
236 if (iova->pfn_lo != IOVA_ANCHOR)
237 kmem_cache_free(s: iova_cache, objp: iova);
238}
239
240/**
241 * alloc_iova - allocates an iova
242 * @iovad: - iova domain in question
243 * @size: - size of page frames to allocate
244 * @limit_pfn: - max limit address
245 * @size_aligned: - set if size_aligned address range is required
246 * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
247 * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
248 * flag is set then the allocated address iova->pfn_lo will be naturally
249 * aligned on roundup_power_of_two(size).
250 */
251struct iova *
252alloc_iova(struct iova_domain *iovad, unsigned long size,
253 unsigned long limit_pfn,
254 bool size_aligned)
255{
256 struct iova *new_iova;
257 int ret;
258
259 new_iova = alloc_iova_mem();
260 if (!new_iova)
261 return NULL;
262
263 ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn: limit_pfn + 1,
264 new: new_iova, size_aligned);
265
266 if (ret) {
267 free_iova_mem(iova: new_iova);
268 return NULL;
269 }
270
271 return new_iova;
272}
273EXPORT_SYMBOL_GPL(alloc_iova);
274
275static struct iova *
276private_find_iova(struct iova_domain *iovad, unsigned long pfn)
277{
278 struct rb_node *node = iovad->rbroot.rb_node;
279
280 assert_spin_locked(&iovad->iova_rbtree_lock);
281
282 while (node) {
283 struct iova *iova = to_iova(node);
284
285 if (pfn < iova->pfn_lo)
286 node = node->rb_left;
287 else if (pfn > iova->pfn_hi)
288 node = node->rb_right;
289 else
290 return iova; /* pfn falls within iova's range */
291 }
292
293 return NULL;
294}
295
296static void remove_iova(struct iova_domain *iovad, struct iova *iova)
297{
298 assert_spin_locked(&iovad->iova_rbtree_lock);
299 __cached_rbnode_delete_update(iovad, free: iova);
300 rb_erase(&iova->node, &iovad->rbroot);
301}
302
303/**
304 * find_iova - finds an iova for a given pfn
305 * @iovad: - iova domain in question.
306 * @pfn: - page frame number
307 * This function finds and returns an iova belonging to the
308 * given domain which matches the given pfn.
309 */
310struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
311{
312 unsigned long flags;
313 struct iova *iova;
314
315 /* Take the lock so that no other thread is manipulating the rbtree */
316 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
317 iova = private_find_iova(iovad, pfn);
318 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
319 return iova;
320}
321EXPORT_SYMBOL_GPL(find_iova);
322
323/**
324 * __free_iova - frees the given iova
325 * @iovad: iova domain in question.
326 * @iova: iova in question.
327 * Frees the given iova belonging to the giving domain
328 */
329void
330__free_iova(struct iova_domain *iovad, struct iova *iova)
331{
332 unsigned long flags;
333
334 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
335 remove_iova(iovad, iova);
336 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
337 free_iova_mem(iova);
338}
339EXPORT_SYMBOL_GPL(__free_iova);
340
341/**
342 * free_iova - finds and frees the iova for a given pfn
343 * @iovad: - iova domain in question.
344 * @pfn: - pfn that is allocated previously
345 * This functions finds an iova for a given pfn and then
346 * frees the iova from that domain.
347 */
348void
349free_iova(struct iova_domain *iovad, unsigned long pfn)
350{
351 unsigned long flags;
352 struct iova *iova;
353
354 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
355 iova = private_find_iova(iovad, pfn);
356 if (!iova) {
357 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
358 return;
359 }
360 remove_iova(iovad, iova);
361 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
362 free_iova_mem(iova);
363}
364EXPORT_SYMBOL_GPL(free_iova);
365
366/**
367 * alloc_iova_fast - allocates an iova from rcache
368 * @iovad: - iova domain in question
369 * @size: - size of page frames to allocate
370 * @limit_pfn: - max limit address
371 * @flush_rcache: - set to flush rcache on regular allocation failure
372 * This function tries to satisfy an iova allocation from the rcache,
373 * and falls back to regular allocation on failure. If regular allocation
374 * fails too and the flush_rcache flag is set then the rcache will be flushed.
375*/
376unsigned long
377alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
378 unsigned long limit_pfn, bool flush_rcache)
379{
380 unsigned long iova_pfn;
381 struct iova *new_iova;
382
383 /*
384 * Freeing non-power-of-two-sized allocations back into the IOVA caches
385 * will come back to bite us badly, so we have to waste a bit of space
386 * rounding up anything cacheable to make sure that can't happen. The
387 * order of the unadjusted size will still match upon freeing.
388 */
389 if (size < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1)))
390 size = roundup_pow_of_two(size);
391
392 iova_pfn = iova_rcache_get(iovad, size, limit_pfn: limit_pfn + 1);
393 if (iova_pfn)
394 return iova_pfn;
395
396retry:
397 new_iova = alloc_iova(iovad, size, limit_pfn, true);
398 if (!new_iova) {
399 unsigned int cpu;
400
401 if (!flush_rcache)
402 return 0;
403
404 /* Try replenishing IOVAs by flushing rcache. */
405 flush_rcache = false;
406 for_each_online_cpu(cpu)
407 free_cpu_cached_iovas(cpu, iovad);
408 free_global_cached_iovas(iovad);
409 goto retry;
410 }
411
412 return new_iova->pfn_lo;
413}
414EXPORT_SYMBOL_GPL(alloc_iova_fast);
415
416/**
417 * free_iova_fast - free iova pfn range into rcache
418 * @iovad: - iova domain in question.
419 * @pfn: - pfn that is allocated previously
420 * @size: - # of pages in range
421 * This functions frees an iova range by trying to put it into the rcache,
422 * falling back to regular iova deallocation via free_iova() if this fails.
423 */
424void
425free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
426{
427 if (iova_rcache_insert(iovad, pfn, size))
428 return;
429
430 free_iova(iovad, pfn);
431}
432EXPORT_SYMBOL_GPL(free_iova_fast);
433
434static void iova_domain_free_rcaches(struct iova_domain *iovad)
435{
436 cpuhp_state_remove_instance_nocalls(state: CPUHP_IOMMU_IOVA_DEAD,
437 node: &iovad->cpuhp_dead);
438 free_iova_rcaches(iovad);
439}
440
441/**
442 * put_iova_domain - destroys the iova domain
443 * @iovad: - iova domain in question.
444 * All the iova's in that domain are destroyed.
445 */
446void put_iova_domain(struct iova_domain *iovad)
447{
448 struct iova *iova, *tmp;
449
450 if (iovad->rcaches)
451 iova_domain_free_rcaches(iovad);
452
453 rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
454 free_iova_mem(iova);
455}
456EXPORT_SYMBOL_GPL(put_iova_domain);
457
458static int
459__is_range_overlap(struct rb_node *node,
460 unsigned long pfn_lo, unsigned long pfn_hi)
461{
462 struct iova *iova = to_iova(node);
463
464 if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
465 return 1;
466 return 0;
467}
468
469static inline struct iova *
470alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
471{
472 struct iova *iova;
473
474 iova = alloc_iova_mem();
475 if (iova) {
476 iova->pfn_lo = pfn_lo;
477 iova->pfn_hi = pfn_hi;
478 }
479
480 return iova;
481}
482
483static struct iova *
484__insert_new_range(struct iova_domain *iovad,
485 unsigned long pfn_lo, unsigned long pfn_hi)
486{
487 struct iova *iova;
488
489 iova = alloc_and_init_iova(pfn_lo, pfn_hi);
490 if (iova)
491 iova_insert_rbtree(root: &iovad->rbroot, iova, NULL);
492
493 return iova;
494}
495
496static void
497__adjust_overlap_range(struct iova *iova,
498 unsigned long *pfn_lo, unsigned long *pfn_hi)
499{
500 if (*pfn_lo < iova->pfn_lo)
501 iova->pfn_lo = *pfn_lo;
502 if (*pfn_hi > iova->pfn_hi)
503 *pfn_lo = iova->pfn_hi + 1;
504}
505
506/**
507 * reserve_iova - reserves an iova in the given range
508 * @iovad: - iova domain pointer
509 * @pfn_lo: - lower page frame address
510 * @pfn_hi:- higher pfn address
511 * This function allocates reserves the address range from pfn_lo to pfn_hi so
512 * that this address is not dished out as part of alloc_iova.
513 */
514struct iova *
515reserve_iova(struct iova_domain *iovad,
516 unsigned long pfn_lo, unsigned long pfn_hi)
517{
518 struct rb_node *node;
519 unsigned long flags;
520 struct iova *iova;
521 unsigned int overlap = 0;
522
523 /* Don't allow nonsensical pfns */
524 if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
525 return NULL;
526
527 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
528 for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
529 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
530 iova = to_iova(node);
531 __adjust_overlap_range(iova, pfn_lo: &pfn_lo, pfn_hi: &pfn_hi);
532 if ((pfn_lo >= iova->pfn_lo) &&
533 (pfn_hi <= iova->pfn_hi))
534 goto finish;
535 overlap = 1;
536
537 } else if (overlap)
538 break;
539 }
540
541 /* We are here either because this is the first reserver node
542 * or need to insert remaining non overlap addr range
543 */
544 iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
545finish:
546
547 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
548 return iova;
549}
550EXPORT_SYMBOL_GPL(reserve_iova);
551
552/*
553 * Magazine caches for IOVA ranges. For an introduction to magazines,
554 * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
555 * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
556 * For simplicity, we use a static magazine size and don't implement the
557 * dynamic size tuning described in the paper.
558 */
559
560/*
561 * As kmalloc's buffer size is fixed to power of 2, 127 is chosen to
562 * assure size of 'iova_magazine' to be 1024 bytes, so that no memory
563 * will be wasted. Since only full magazines are inserted into the depot,
564 * we don't need to waste PFN capacity on a separate list head either.
565 */
566#define IOVA_MAG_SIZE 127
567
568#define IOVA_DEPOT_DELAY msecs_to_jiffies(100)
569
570struct iova_magazine {
571 union {
572 unsigned long size;
573 struct iova_magazine *next;
574 };
575 unsigned long pfns[IOVA_MAG_SIZE];
576};
577static_assert(!(sizeof(struct iova_magazine) & (sizeof(struct iova_magazine) - 1)));
578
579struct iova_cpu_rcache {
580 spinlock_t lock;
581 struct iova_magazine *loaded;
582 struct iova_magazine *prev;
583};
584
585struct iova_rcache {
586 spinlock_t lock;
587 unsigned int depot_size;
588 struct iova_magazine *depot;
589 struct iova_cpu_rcache __percpu *cpu_rcaches;
590 struct iova_domain *iovad;
591 struct delayed_work work;
592};
593
594static struct kmem_cache *iova_magazine_cache;
595
596unsigned long iova_rcache_range(void)
597{
598 return PAGE_SIZE << (IOVA_RANGE_CACHE_MAX_SIZE - 1);
599}
600
601static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
602{
603 struct iova_magazine *mag;
604
605 mag = kmem_cache_alloc(iova_magazine_cache, flags);
606 if (mag)
607 mag->size = 0;
608
609 return mag;
610}
611
612static void iova_magazine_free(struct iova_magazine *mag)
613{
614 kmem_cache_free(s: iova_magazine_cache, objp: mag);
615}
616
617static void
618iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
619{
620 unsigned long flags;
621 int i;
622
623 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
624
625 for (i = 0 ; i < mag->size; ++i) {
626 struct iova *iova = private_find_iova(iovad, pfn: mag->pfns[i]);
627
628 if (WARN_ON(!iova))
629 continue;
630
631 remove_iova(iovad, iova);
632 free_iova_mem(iova);
633 }
634
635 spin_unlock_irqrestore(lock: &iovad->iova_rbtree_lock, flags);
636
637 mag->size = 0;
638}
639
640static bool iova_magazine_full(struct iova_magazine *mag)
641{
642 return mag->size == IOVA_MAG_SIZE;
643}
644
645static bool iova_magazine_empty(struct iova_magazine *mag)
646{
647 return mag->size == 0;
648}
649
650static unsigned long iova_magazine_pop(struct iova_magazine *mag,
651 unsigned long limit_pfn)
652{
653 int i;
654 unsigned long pfn;
655
656 /* Only fall back to the rbtree if we have no suitable pfns at all */
657 for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
658 if (i == 0)
659 return 0;
660
661 /* Swap it to pop it */
662 pfn = mag->pfns[i];
663 mag->pfns[i] = mag->pfns[--mag->size];
664
665 return pfn;
666}
667
668static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
669{
670 mag->pfns[mag->size++] = pfn;
671}
672
673static struct iova_magazine *iova_depot_pop(struct iova_rcache *rcache)
674{
675 struct iova_magazine *mag = rcache->depot;
676
677 /*
678 * As the mag->next pointer is moved to rcache->depot and reset via
679 * the mag->size assignment, mark it as a transient false positive.
680 */
681 kmemleak_transient_leak(ptr: mag->next);
682 rcache->depot = mag->next;
683 mag->size = IOVA_MAG_SIZE;
684 rcache->depot_size--;
685 return mag;
686}
687
688static void iova_depot_push(struct iova_rcache *rcache, struct iova_magazine *mag)
689{
690 mag->next = rcache->depot;
691 rcache->depot = mag;
692 rcache->depot_size++;
693}
694
695static void iova_depot_work_func(struct work_struct *work)
696{
697 struct iova_rcache *rcache = container_of(work, typeof(*rcache), work.work);
698 struct iova_magazine *mag = NULL;
699 unsigned long flags;
700
701 spin_lock_irqsave(&rcache->lock, flags);
702 if (rcache->depot_size > num_online_cpus())
703 mag = iova_depot_pop(rcache);
704 spin_unlock_irqrestore(lock: &rcache->lock, flags);
705
706 if (mag) {
707 iova_magazine_free_pfns(mag, iovad: rcache->iovad);
708 iova_magazine_free(mag);
709 schedule_delayed_work(dwork: &rcache->work, IOVA_DEPOT_DELAY);
710 }
711}
712
713int iova_domain_init_rcaches(struct iova_domain *iovad)
714{
715 unsigned int cpu;
716 int i, ret;
717
718 iovad->rcaches = kcalloc(IOVA_RANGE_CACHE_MAX_SIZE,
719 sizeof(struct iova_rcache),
720 GFP_KERNEL);
721 if (!iovad->rcaches)
722 return -ENOMEM;
723
724 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
725 struct iova_cpu_rcache *cpu_rcache;
726 struct iova_rcache *rcache;
727
728 rcache = &iovad->rcaches[i];
729 spin_lock_init(&rcache->lock);
730 rcache->iovad = iovad;
731 INIT_DELAYED_WORK(&rcache->work, iova_depot_work_func);
732 rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache),
733 cache_line_size());
734 if (!rcache->cpu_rcaches) {
735 ret = -ENOMEM;
736 goto out_err;
737 }
738 for_each_possible_cpu(cpu) {
739 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
740
741 spin_lock_init(&cpu_rcache->lock);
742 cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
743 cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
744 if (!cpu_rcache->loaded || !cpu_rcache->prev) {
745 ret = -ENOMEM;
746 goto out_err;
747 }
748 }
749 }
750
751 ret = cpuhp_state_add_instance_nocalls(state: CPUHP_IOMMU_IOVA_DEAD,
752 node: &iovad->cpuhp_dead);
753 if (ret)
754 goto out_err;
755 return 0;
756
757out_err:
758 free_iova_rcaches(iovad);
759 return ret;
760}
761EXPORT_SYMBOL_GPL(iova_domain_init_rcaches);
762
763/*
764 * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
765 * return true on success. Can fail if rcache is full and we can't free
766 * space, and free_iova() (our only caller) will then return the IOVA
767 * range to the rbtree instead.
768 */
769static bool __iova_rcache_insert(struct iova_domain *iovad,
770 struct iova_rcache *rcache,
771 unsigned long iova_pfn)
772{
773 struct iova_cpu_rcache *cpu_rcache;
774 bool can_insert = false;
775 unsigned long flags;
776
777 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
778 spin_lock_irqsave(&cpu_rcache->lock, flags);
779
780 if (!iova_magazine_full(mag: cpu_rcache->loaded)) {
781 can_insert = true;
782 } else if (!iova_magazine_full(mag: cpu_rcache->prev)) {
783 swap(cpu_rcache->prev, cpu_rcache->loaded);
784 can_insert = true;
785 } else {
786 struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
787
788 if (new_mag) {
789 spin_lock(lock: &rcache->lock);
790 iova_depot_push(rcache, mag: cpu_rcache->loaded);
791 spin_unlock(lock: &rcache->lock);
792 schedule_delayed_work(dwork: &rcache->work, IOVA_DEPOT_DELAY);
793
794 cpu_rcache->loaded = new_mag;
795 can_insert = true;
796 }
797 }
798
799 if (can_insert)
800 iova_magazine_push(mag: cpu_rcache->loaded, pfn: iova_pfn);
801
802 spin_unlock_irqrestore(lock: &cpu_rcache->lock, flags);
803
804 return can_insert;
805}
806
807static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
808 unsigned long size)
809{
810 unsigned int log_size = order_base_2(size);
811
812 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
813 return false;
814
815 return __iova_rcache_insert(iovad, rcache: &iovad->rcaches[log_size], iova_pfn: pfn);
816}
817
818/*
819 * Caller wants to allocate a new IOVA range from 'rcache'. If we can
820 * satisfy the request, return a matching non-NULL range and remove
821 * it from the 'rcache'.
822 */
823static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
824 unsigned long limit_pfn)
825{
826 struct iova_cpu_rcache *cpu_rcache;
827 unsigned long iova_pfn = 0;
828 bool has_pfn = false;
829 unsigned long flags;
830
831 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
832 spin_lock_irqsave(&cpu_rcache->lock, flags);
833
834 if (!iova_magazine_empty(mag: cpu_rcache->loaded)) {
835 has_pfn = true;
836 } else if (!iova_magazine_empty(mag: cpu_rcache->prev)) {
837 swap(cpu_rcache->prev, cpu_rcache->loaded);
838 has_pfn = true;
839 } else {
840 spin_lock(lock: &rcache->lock);
841 if (rcache->depot) {
842 iova_magazine_free(mag: cpu_rcache->loaded);
843 cpu_rcache->loaded = iova_depot_pop(rcache);
844 has_pfn = true;
845 }
846 spin_unlock(lock: &rcache->lock);
847 }
848
849 if (has_pfn)
850 iova_pfn = iova_magazine_pop(mag: cpu_rcache->loaded, limit_pfn);
851
852 spin_unlock_irqrestore(lock: &cpu_rcache->lock, flags);
853
854 return iova_pfn;
855}
856
857/*
858 * Try to satisfy IOVA allocation range from rcache. Fail if requested
859 * size is too big or the DMA limit we are given isn't satisfied by the
860 * top element in the magazine.
861 */
862static unsigned long iova_rcache_get(struct iova_domain *iovad,
863 unsigned long size,
864 unsigned long limit_pfn)
865{
866 unsigned int log_size = order_base_2(size);
867
868 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
869 return 0;
870
871 return __iova_rcache_get(rcache: &iovad->rcaches[log_size], limit_pfn: limit_pfn - size);
872}
873
874/*
875 * free rcache data structures.
876 */
877static void free_iova_rcaches(struct iova_domain *iovad)
878{
879 struct iova_rcache *rcache;
880 struct iova_cpu_rcache *cpu_rcache;
881 unsigned int cpu;
882
883 for (int i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
884 rcache = &iovad->rcaches[i];
885 if (!rcache->cpu_rcaches)
886 break;
887 for_each_possible_cpu(cpu) {
888 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
889 iova_magazine_free(mag: cpu_rcache->loaded);
890 iova_magazine_free(mag: cpu_rcache->prev);
891 }
892 free_percpu(pdata: rcache->cpu_rcaches);
893 cancel_delayed_work_sync(dwork: &rcache->work);
894 while (rcache->depot)
895 iova_magazine_free(mag: iova_depot_pop(rcache));
896 }
897
898 kfree(objp: iovad->rcaches);
899 iovad->rcaches = NULL;
900}
901
902/*
903 * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
904 */
905static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
906{
907 struct iova_cpu_rcache *cpu_rcache;
908 struct iova_rcache *rcache;
909 unsigned long flags;
910 int i;
911
912 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
913 rcache = &iovad->rcaches[i];
914 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
915 spin_lock_irqsave(&cpu_rcache->lock, flags);
916 iova_magazine_free_pfns(mag: cpu_rcache->loaded, iovad);
917 iova_magazine_free_pfns(mag: cpu_rcache->prev, iovad);
918 spin_unlock_irqrestore(lock: &cpu_rcache->lock, flags);
919 }
920}
921
922/*
923 * free all the IOVA ranges of global cache
924 */
925static void free_global_cached_iovas(struct iova_domain *iovad)
926{
927 struct iova_rcache *rcache;
928 unsigned long flags;
929
930 for (int i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
931 rcache = &iovad->rcaches[i];
932 spin_lock_irqsave(&rcache->lock, flags);
933 while (rcache->depot) {
934 struct iova_magazine *mag = iova_depot_pop(rcache);
935
936 iova_magazine_free_pfns(mag, iovad);
937 iova_magazine_free(mag);
938 }
939 spin_unlock_irqrestore(lock: &rcache->lock, flags);
940 }
941}
942
943static int iova_cpuhp_dead(unsigned int cpu, struct hlist_node *node)
944{
945 struct iova_domain *iovad;
946
947 iovad = hlist_entry_safe(node, struct iova_domain, cpuhp_dead);
948
949 free_cpu_cached_iovas(cpu, iovad);
950 return 0;
951}
952
953int iova_cache_get(void)
954{
955 int err = -ENOMEM;
956
957 mutex_lock(lock: &iova_cache_mutex);
958 if (!iova_cache_users) {
959 iova_cache = kmem_cache_create("iommu_iova", sizeof(struct iova), 0,
960 SLAB_HWCACHE_ALIGN, NULL);
961 if (!iova_cache)
962 goto out_err;
963
964 iova_magazine_cache = kmem_cache_create("iommu_iova_magazine",
965 sizeof(struct iova_magazine),
966 0, SLAB_HWCACHE_ALIGN, NULL);
967 if (!iova_magazine_cache)
968 goto out_err;
969
970 err = cpuhp_setup_state_multi(state: CPUHP_IOMMU_IOVA_DEAD, name: "iommu/iova:dead",
971 NULL, teardown: iova_cpuhp_dead);
972 if (err) {
973 pr_err("IOVA: Couldn't register cpuhp handler: %pe\n", ERR_PTR(err));
974 goto out_err;
975 }
976 }
977
978 iova_cache_users++;
979 mutex_unlock(lock: &iova_cache_mutex);
980
981 return 0;
982
983out_err:
984 kmem_cache_destroy(s: iova_cache);
985 kmem_cache_destroy(s: iova_magazine_cache);
986 mutex_unlock(lock: &iova_cache_mutex);
987 return err;
988}
989EXPORT_SYMBOL_GPL(iova_cache_get);
990
991void iova_cache_put(void)
992{
993 mutex_lock(lock: &iova_cache_mutex);
994 if (WARN_ON(!iova_cache_users)) {
995 mutex_unlock(lock: &iova_cache_mutex);
996 return;
997 }
998 iova_cache_users--;
999 if (!iova_cache_users) {
1000 cpuhp_remove_multi_state(state: CPUHP_IOMMU_IOVA_DEAD);
1001 kmem_cache_destroy(s: iova_cache);
1002 kmem_cache_destroy(s: iova_magazine_cache);
1003 }
1004 mutex_unlock(lock: &iova_cache_mutex);
1005}
1006EXPORT_SYMBOL_GPL(iova_cache_put);
1007
1008MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
1009MODULE_DESCRIPTION("IOMMU I/O Virtual Address management");
1010MODULE_LICENSE("GPL");
1011