Merge tag 'io_uring-6.2-2023-01-13' of git://git.kernel.dk/linux
[linux-2.6-block.git] / drivers / iommu / iova.c
CommitLineData
3b20eb23 1// SPDX-License-Identifier: GPL-2.0-only
f8de50eb 2/*
a15a519e 3 * Copyright © 2006-2009, Intel Corporation.
f8de50eb 4 *
98bcef56 5 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
f8de50eb
KA
6 */
7
38717946 8#include <linux/iova.h>
15bbdec3 9#include <linux/module.h>
85b45456 10#include <linux/slab.h>
9257b4a2
OP
11#include <linux/smp.h>
12#include <linux/bitops.h>
aaffaa8a 13#include <linux/cpu.h>
9257b4a2 14
bb68b2fb
RM
15/* The anchor node sits above the top of the usable address space */
16#define IOVA_ANCHOR ~0UL
17
32e92d9f
JG
18#define IOVA_RANGE_CACHE_MAX_SIZE 6 /* log of max cached IOVA range size (in pages) */
19
9257b4a2
OP
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);
149448b3 26static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad);
9257b4a2 27static void free_iova_rcaches(struct iova_domain *iovad);
f598a497 28
6d9870b7
JG
29unsigned long iova_rcache_range(void)
30{
31 return PAGE_SIZE << (IOVA_RANGE_CACHE_MAX_SIZE - 1);
32}
33
f598a497
JG
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
6fa3525b 44static void free_global_cached_iovas(struct iova_domain *iovad);
85b45456 45
7ae31cec
RM
46static struct iova *to_iova(struct rb_node *node)
47{
48 return rb_entry(node, struct iova, node);
49}
50
f8de50eb 51void
0fb5fe87 52init_iova_domain(struct iova_domain *iovad, unsigned long granule,
aa3ac946 53 unsigned long start_pfn)
f8de50eb 54{
0fb5fe87
RM
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
f8de50eb
KA
62 spin_lock_init(&iovad->iova_rbtree_lock);
63 iovad->rbroot = RB_ROOT;
973f5fbe
RM
64 iovad->cached_node = &iovad->anchor.node;
65 iovad->cached32_node = &iovad->anchor.node;
0fb5fe87 66 iovad->granule = granule;
1b722500 67 iovad->start_pfn = start_pfn;
aa3ac946 68 iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
bee60e94 69 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
bb68b2fb
RM
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);
f8de50eb 73}
9b41760b 74EXPORT_SYMBOL_GPL(init_iova_domain);
f8de50eb
KA
75
76static struct rb_node *
973f5fbe 77__get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
f8de50eb 78{
973f5fbe
RM
79 if (limit_pfn <= iovad->dma_32bit_pfn)
80 return iovad->cached32_node;
e60aa7b5 81
973f5fbe 82 return iovad->cached_node;
f8de50eb
KA
83}
84
85static void
e60aa7b5 86__cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
f8de50eb 87{
e60aa7b5
RM
88 if (new->pfn_hi < iovad->dma_32bit_pfn)
89 iovad->cached32_node = &new->node;
90 else
91 iovad->cached_node = &new->node;
f8de50eb
KA
92}
93
94static void
95__cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
96{
97 struct iova *cached_iova;
f8de50eb 98
7ae31cec 99 cached_iova = to_iova(iovad->cached32_node);
9eed17d3
CW
100 if (free == cached_iova ||
101 (free->pfn_hi < iovad->dma_32bit_pfn &&
5b61343b 102 free->pfn_lo >= cached_iova->pfn_lo))
e60aa7b5 103 iovad->cached32_node = rb_next(&free->node);
5b61343b
RM
104
105 if (free->pfn_lo < iovad->dma_32bit_pfn)
bee60e94 106 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
e60aa7b5 107
7ae31cec 108 cached_iova = to_iova(iovad->cached_node);
973f5fbe 109 if (free->pfn_lo >= cached_iova->pfn_lo)
e60aa7b5 110 iovad->cached_node = rb_next(&free->node);
f8de50eb
KA
111}
112
371d7955
RM
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
d751751a
MS
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) {
7ae31cec 160 struct iova *this = to_iova(*new);
d751751a
MS
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
ddf02886 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)
f8de50eb 181{
973f5fbe
RM
182 struct rb_node *curr, *prev;
183 struct iova *curr_iova;
f8de50eb 184 unsigned long flags;
4e89dce7 185 unsigned long new_pfn, retry_pfn;
086c83ac 186 unsigned long align_mask = ~0UL;
4e89dce7 187 unsigned long high_pfn = limit_pfn, low_pfn = iovad->start_pfn;
086c83ac
ZL
188
189 if (size_aligned)
190 align_mask <<= fls_long(size - 1);
f8de50eb
KA
191
192 /* Walk the tree backwards */
193 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
bee60e94
GK
194 if (limit_pfn <= iovad->dma_32bit_pfn &&
195 size >= iovad->max32_alloc_size)
196 goto iova32_full;
197
973f5fbe 198 curr = __get_cached_rbnode(iovad, limit_pfn);
7ae31cec 199 curr_iova = to_iova(curr);
4e89dce7
VJ
200 retry_pfn = curr_iova->pfn_hi + 1;
201
202retry:
973f5fbe 203 do {
4e89dce7
VJ
204 high_pfn = min(high_pfn, curr_iova->pfn_lo);
205 new_pfn = (high_pfn - size) & align_mask;
ddf02886 206 prev = curr;
f8de50eb 207 curr = rb_prev(curr);
7ae31cec 208 curr_iova = to_iova(curr);
4e89dce7
VJ
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;
371d7955 215 curr = iova_find_limit(iovad, limit_pfn);
7ae31cec 216 curr_iova = to_iova(curr);
4e89dce7
VJ
217 goto retry;
218 }
80ef4464 219 iovad->max32_alloc_size = size;
bee60e94 220 goto iova32_full;
80ef4464 221 }
f76aec76
KA
222
223 /* pfn_lo will point to size aligned address if size_aligned is set */
086c83ac 224 new->pfn_lo = new_pfn;
f76aec76 225 new->pfn_hi = new->pfn_lo + size - 1;
f8de50eb 226
d751751a
MS
227 /* If we have 'prev', it's a valid place to start the insertion. */
228 iova_insert_rbtree(&iovad->rbroot, new, prev);
e60aa7b5 229 __cached_rbnode_insert_update(iovad, new);
ddf02886 230
f8de50eb
KA
231 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
232 return 0;
bee60e94
GK
233
234iova32_full:
bee60e94
GK
235 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
236 return -ENOMEM;
f8de50eb
KA
237}
238
ae1ff3d6
SA
239static struct kmem_cache *iova_cache;
240static unsigned int iova_cache_users;
241static DEFINE_MUTEX(iova_cache_mutex);
242
51b70b81 243static struct iova *alloc_iova_mem(void)
ae1ff3d6 244{
944c9175 245 return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
ae1ff3d6 246}
ae1ff3d6 247
176cfc18 248static void free_iova_mem(struct iova *iova)
ae1ff3d6 249{
bb68b2fb
RM
250 if (iova->pfn_lo != IOVA_ANCHOR)
251 kmem_cache_free(iova_cache, iova);
ae1ff3d6 252}
ae1ff3d6
SA
253
254int iova_cache_get(void)
255{
256 mutex_lock(&iova_cache_mutex);
257 if (!iova_cache_users) {
f598a497
JG
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
ae1ff3d6
SA
268 iova_cache = kmem_cache_create(
269 "iommu_iova", sizeof(struct iova), 0,
270 SLAB_HWCACHE_ALIGN, NULL);
271 if (!iova_cache) {
f598a497 272 cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
ae1ff3d6 273 mutex_unlock(&iova_cache_mutex);
3a0ce12e 274 pr_err("Couldn't create iova cache\n");
ae1ff3d6
SA
275 return -ENOMEM;
276 }
277 }
278
279 iova_cache_users++;
280 mutex_unlock(&iova_cache_mutex);
281
282 return 0;
283}
9b41760b 284EXPORT_SYMBOL_GPL(iova_cache_get);
ae1ff3d6
SA
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--;
f598a497
JG
294 if (!iova_cache_users) {
295 cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
ae1ff3d6 296 kmem_cache_destroy(iova_cache);
f598a497 297 }
ae1ff3d6
SA
298 mutex_unlock(&iova_cache_mutex);
299}
9b41760b 300EXPORT_SYMBOL_GPL(iova_cache_put);
ae1ff3d6 301
f8de50eb
KA
302/**
303 * alloc_iova - allocates an iova
07db0409
MI
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
1b722500
RM
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
f76aec76
KA
310 * flag is set then the allocated address iova->pfn_lo will be naturally
311 * aligned on roundup_power_of_two(size).
f8de50eb
KA
312 */
313struct iova *
314alloc_iova(struct iova_domain *iovad, unsigned long size,
f76aec76
KA
315 unsigned long limit_pfn,
316 bool size_aligned)
f8de50eb 317{
f8de50eb
KA
318 struct iova *new_iova;
319 int ret;
320
321 new_iova = alloc_iova_mem();
322 if (!new_iova)
323 return NULL;
324
757c370f 325 ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
ddf02886 326 new_iova, size_aligned);
f8de50eb
KA
327
328 if (ret) {
f8de50eb
KA
329 free_iova_mem(new_iova);
330 return NULL;
331 }
332
f8de50eb
KA
333 return new_iova;
334}
9b41760b 335EXPORT_SYMBOL_GPL(alloc_iova);
f8de50eb 336
9257b4a2
OP
337static struct iova *
338private_find_iova(struct iova_domain *iovad, unsigned long pfn)
f8de50eb 339{
9257b4a2
OP
340 struct rb_node *node = iovad->rbroot.rb_node;
341
342 assert_spin_locked(&iovad->iova_rbtree_lock);
f8de50eb 343
f8de50eb 344 while (node) {
7ae31cec 345 struct iova *iova = to_iova(node);
f8de50eb 346
f8de50eb
KA
347 if (pfn < iova->pfn_lo)
348 node = node->rb_left;
2070f940 349 else if (pfn > iova->pfn_hi)
f8de50eb 350 node = node->rb_right;
2070f940
ZL
351 else
352 return iova; /* pfn falls within iova's range */
f8de50eb
KA
353 }
354
f8de50eb
KA
355 return NULL;
356}
9257b4a2 357
7978724f 358static void remove_iova(struct iova_domain *iovad, struct iova *iova)
9257b4a2
OP
359{
360 assert_spin_locked(&iovad->iova_rbtree_lock);
361 __cached_rbnode_delete_update(iovad, iova);
362 rb_erase(&iova->node, &iovad->rbroot);
9257b4a2
OP
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
6775ae90 370 * given domain which matches the given pfn.
9257b4a2
OP
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}
9b41760b 383EXPORT_SYMBOL_GPL(find_iova);
f8de50eb
KA
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);
7978724f 397 remove_iova(iovad, iova);
f8de50eb 398 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
7978724f 399 free_iova_mem(iova);
f8de50eb 400}
9b41760b 401EXPORT_SYMBOL_GPL(__free_iova);
f8de50eb
KA
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{
3a651b3a
CW
413 unsigned long flags;
414 struct iova *iova;
733cac2a 415
3a651b3a
CW
416 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
417 iova = private_find_iova(iovad, pfn);
7978724f
XC
418 if (!iova) {
419 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
420 return;
421 }
422 remove_iova(iovad, iova);
3a651b3a 423 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
7978724f 424 free_iova_mem(iova);
f8de50eb 425}
9b41760b 426EXPORT_SYMBOL_GPL(free_iova);
f8de50eb 427
9257b4a2
OP
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
538d5b33 433 * @flush_rcache: - set to flush rcache on regular allocation failure
9257b4a2 434 * This function tries to satisfy an iova allocation from the rcache,
538d5b33
TN
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.
9257b4a2
OP
437*/
438unsigned long
439alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
538d5b33 440 unsigned long limit_pfn, bool flush_rcache)
9257b4a2 441{
9257b4a2
OP
442 unsigned long iova_pfn;
443 struct iova *new_iova;
444
972bf252
JG
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
b826ee9a 454 iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
9257b4a2
OP
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
538d5b33 463 if (!flush_rcache)
9257b4a2
OP
464 return 0;
465
466 /* Try replenishing IOVAs by flushing rcache. */
538d5b33 467 flush_rcache = false;
9257b4a2
OP
468 for_each_online_cpu(cpu)
469 free_cpu_cached_iovas(cpu, iovad);
6fa3525b 470 free_global_cached_iovas(iovad);
9257b4a2
OP
471 goto retry;
472 }
473
474 return new_iova->pfn_lo;
475}
a93a9626 476EXPORT_SYMBOL_GPL(alloc_iova_fast);
9257b4a2
OP
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}
a93a9626 494EXPORT_SYMBOL_GPL(free_iova_fast);
9257b4a2 495
32e92d9f
JG
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
f8de50eb 503/**
6775ae90 504 * put_iova_domain - destroys the iova domain
f8de50eb
KA
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{
7595dc58 510 struct iova *iova, *tmp;
f8de50eb 511
32e92d9f
JG
512 if (iovad->rcaches)
513 iova_domain_free_rcaches(iovad);
514
7595dc58 515 rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
f8de50eb 516 free_iova_mem(iova);
f8de50eb 517}
9b41760b 518EXPORT_SYMBOL_GPL(put_iova_domain);
f8de50eb
KA
519
520static int
521__is_range_overlap(struct rb_node *node,
522 unsigned long pfn_lo, unsigned long pfn_hi)
523{
7ae31cec 524 struct iova *iova = to_iova(node);
f8de50eb
KA
525
526 if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
527 return 1;
528 return 0;
529}
530
75f05569
JL
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
f8de50eb
KA
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
75f05569
JL
551 iova = alloc_and_init_iova(pfn_lo, pfn_hi);
552 if (iova)
d751751a 553 iova_insert_rbtree(&iovad->rbroot, iova, NULL);
f8de50eb 554
f8de50eb
KA
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
bb68b2fb
RM
585 /* Don't allow nonsensical pfns */
586 if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
587 return NULL;
588
3d39cecc 589 spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
f8de50eb
KA
590 for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
591 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
7ae31cec 592 iova = to_iova(node);
f8de50eb
KA
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
25985edc 603 /* We are here either because this is the first reserver node
f8de50eb
KA
604 * or need to insert remaining non overlap addr range
605 */
606 iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
607finish:
608
3d39cecc 609 spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
f8de50eb
KA
610 return iova;
611}
9b41760b 612EXPORT_SYMBOL_GPL(reserve_iova);
f8de50eb 613
9257b4a2
OP
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
b4c9bf17
FT
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
32e92d9f 628#define MAX_GLOBAL_MAGS 32 /* magazines per bin */
9257b4a2
OP
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
32e92d9f
JG
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
9257b4a2
OP
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
9257b4a2
OP
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
d3e3d2be
RM
669 if (WARN_ON(!iova))
670 continue;
671
7978724f
XC
672 remove_iova(iovad, iova);
673 free_iova_mem(iova);
9257b4a2
OP
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{
a390bde7 683 return mag->size == IOVA_MAG_SIZE;
9257b4a2
OP
684}
685
686static bool iova_magazine_empty(struct iova_magazine *mag)
687{
a390bde7 688 return mag->size == 0;
9257b4a2
OP
689}
690
691static unsigned long iova_magazine_pop(struct iova_magazine *mag,
692 unsigned long limit_pfn)
693{
e8b19840
RM
694 int i;
695 unsigned long pfn;
696
e8b19840
RM
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];
9257b4a2 705
e8b19840 706 return pfn;
9257b4a2
OP
707}
708
709static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
710{
9257b4a2
OP
711 mag->pfns[mag->size++] = pfn;
712}
713
32e92d9f 714int iova_domain_init_rcaches(struct iova_domain *iovad)
9257b4a2 715{
9257b4a2 716 unsigned int cpu;
32e92d9f
JG
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;
9257b4a2
OP
724
725 for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
32e92d9f
JG
726 struct iova_cpu_rcache *cpu_rcache;
727 struct iova_rcache *rcache;
728
9257b4a2
OP
729 rcache = &iovad->rcaches[i];
730 spin_lock_init(&rcache->lock);
731 rcache->depot_size = 0;
32e92d9f
JG
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 }
9257b4a2
OP
738 for_each_possible_cpu(cpu) {
739 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
32e92d9f 740
9257b4a2
OP
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);
32e92d9f
JG
744 if (!cpu_rcache->loaded || !cpu_rcache->prev) {
745 ret = -ENOMEM;
746 goto out_err;
747 }
9257b4a2
OP
748 }
749 }
32e92d9f
JG
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;
9257b4a2 760}
32e92d9f 761EXPORT_SYMBOL_GPL(iova_domain_init_rcaches);
9257b4a2
OP
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
aaffaa8a 778 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
9257b4a2
OP
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
aaffaa8a 841 cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
9257b4a2
OP
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
189cb8fe 878 if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
9257b4a2
OP
879 return 0;
880
b826ee9a 881 return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
9257b4a2
OP
882}
883
9257b4a2
OP
884/*
885 * free rcache data structures.
886 */
887static void free_iova_rcaches(struct iova_domain *iovad)
888{
889 struct iova_rcache *rcache;
7595dc58 890 struct iova_cpu_rcache *cpu_rcache;
9257b4a2
OP
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];
32e92d9f
JG
896 if (!rcache->cpu_rcaches)
897 break;
7595dc58
RM
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 }
9257b4a2 903 free_percpu(rcache->cpu_rcaches);
7595dc58 904 for (j = 0; j < rcache->depot_size; ++j)
9257b4a2 905 iova_magazine_free(rcache->depot[j]);
9257b4a2 906 }
32e92d9f
JG
907
908 kfree(iovad->rcaches);
909 iovad->rcaches = NULL;
9257b4a2
OP
910}
911
912/*
913 * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
914 */
149448b3 915static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
9257b4a2
OP
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
6fa3525b
VJ
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}
15bbdec3
SA
952MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
953MODULE_LICENSE("GPL");