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