xref: /openbmc/linux/drivers/iommu/iova.c (revision b830f94f)
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 static bool iova_rcache_insert(struct iova_domain *iovad,
19 			       unsigned long pfn,
20 			       unsigned long size);
21 static unsigned long iova_rcache_get(struct iova_domain *iovad,
22 				     unsigned long size,
23 				     unsigned long limit_pfn);
24 static void init_iova_rcaches(struct iova_domain *iovad);
25 static void free_iova_rcaches(struct iova_domain *iovad);
26 static void fq_destroy_all_entries(struct iova_domain *iovad);
27 static void fq_flush_timeout(struct timer_list *t);
28 
29 void
30 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
31 	unsigned long start_pfn)
32 {
33 	/*
34 	 * IOVA granularity will normally be equal to the smallest
35 	 * supported IOMMU page size; both *must* be capable of
36 	 * representing individual CPU pages exactly.
37 	 */
38 	BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
39 
40 	spin_lock_init(&iovad->iova_rbtree_lock);
41 	iovad->rbroot = RB_ROOT;
42 	iovad->cached_node = &iovad->anchor.node;
43 	iovad->cached32_node = &iovad->anchor.node;
44 	iovad->granule = granule;
45 	iovad->start_pfn = start_pfn;
46 	iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
47 	iovad->max32_alloc_size = iovad->dma_32bit_pfn;
48 	iovad->flush_cb = NULL;
49 	iovad->fq = NULL;
50 	iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
51 	rb_link_node(&iovad->anchor.node, NULL, &iovad->rbroot.rb_node);
52 	rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
53 	init_iova_rcaches(iovad);
54 }
55 EXPORT_SYMBOL_GPL(init_iova_domain);
56 
57 static void free_iova_flush_queue(struct iova_domain *iovad)
58 {
59 	if (!iovad->fq)
60 		return;
61 
62 	if (timer_pending(&iovad->fq_timer))
63 		del_timer(&iovad->fq_timer);
64 
65 	fq_destroy_all_entries(iovad);
66 
67 	free_percpu(iovad->fq);
68 
69 	iovad->fq         = NULL;
70 	iovad->flush_cb   = NULL;
71 	iovad->entry_dtor = NULL;
72 }
73 
74 int init_iova_flush_queue(struct iova_domain *iovad,
75 			  iova_flush_cb flush_cb, iova_entry_dtor entry_dtor)
76 {
77 	int cpu;
78 
79 	atomic64_set(&iovad->fq_flush_start_cnt,  0);
80 	atomic64_set(&iovad->fq_flush_finish_cnt, 0);
81 
82 	iovad->fq = alloc_percpu(struct iova_fq);
83 	if (!iovad->fq)
84 		return -ENOMEM;
85 
86 	iovad->flush_cb   = flush_cb;
87 	iovad->entry_dtor = entry_dtor;
88 
89 	for_each_possible_cpu(cpu) {
90 		struct iova_fq *fq;
91 
92 		fq = per_cpu_ptr(iovad->fq, cpu);
93 		fq->head = 0;
94 		fq->tail = 0;
95 
96 		spin_lock_init(&fq->lock);
97 	}
98 
99 	timer_setup(&iovad->fq_timer, fq_flush_timeout, 0);
100 	atomic_set(&iovad->fq_timer_on, 0);
101 
102 	return 0;
103 }
104 EXPORT_SYMBOL_GPL(init_iova_flush_queue);
105 
106 static struct rb_node *
107 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
108 {
109 	if (limit_pfn <= iovad->dma_32bit_pfn)
110 		return iovad->cached32_node;
111 
112 	return iovad->cached_node;
113 }
114 
115 static void
116 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
117 {
118 	if (new->pfn_hi < iovad->dma_32bit_pfn)
119 		iovad->cached32_node = &new->node;
120 	else
121 		iovad->cached_node = &new->node;
122 }
123 
124 static void
125 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
126 {
127 	struct iova *cached_iova;
128 
129 	cached_iova = rb_entry(iovad->cached32_node, struct iova, node);
130 	if (free->pfn_hi < iovad->dma_32bit_pfn &&
131 	    free->pfn_lo >= cached_iova->pfn_lo) {
132 		iovad->cached32_node = rb_next(&free->node);
133 		iovad->max32_alloc_size = iovad->dma_32bit_pfn;
134 	}
135 
136 	cached_iova = rb_entry(iovad->cached_node, struct iova, node);
137 	if (free->pfn_lo >= cached_iova->pfn_lo)
138 		iovad->cached_node = rb_next(&free->node);
139 }
140 
141 /* Insert the iova into domain rbtree by holding writer lock */
142 static void
143 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
144 		   struct rb_node *start)
145 {
146 	struct rb_node **new, *parent = NULL;
147 
148 	new = (start) ? &start : &(root->rb_node);
149 	/* Figure out where to put new node */
150 	while (*new) {
151 		struct iova *this = rb_entry(*new, struct iova, node);
152 
153 		parent = *new;
154 
155 		if (iova->pfn_lo < this->pfn_lo)
156 			new = &((*new)->rb_left);
157 		else if (iova->pfn_lo > this->pfn_lo)
158 			new = &((*new)->rb_right);
159 		else {
160 			WARN_ON(1); /* this should not happen */
161 			return;
162 		}
163 	}
164 	/* Add new node and rebalance tree. */
165 	rb_link_node(&iova->node, parent, new);
166 	rb_insert_color(&iova->node, root);
167 }
168 
169 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
170 		unsigned long size, unsigned long limit_pfn,
171 			struct iova *new, bool size_aligned)
172 {
173 	struct rb_node *curr, *prev;
174 	struct iova *curr_iova;
175 	unsigned long flags;
176 	unsigned long new_pfn;
177 	unsigned long align_mask = ~0UL;
178 
179 	if (size_aligned)
180 		align_mask <<= fls_long(size - 1);
181 
182 	/* Walk the tree backwards */
183 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
184 	if (limit_pfn <= iovad->dma_32bit_pfn &&
185 			size >= iovad->max32_alloc_size)
186 		goto iova32_full;
187 
188 	curr = __get_cached_rbnode(iovad, limit_pfn);
189 	curr_iova = rb_entry(curr, struct iova, node);
190 	do {
191 		limit_pfn = min(limit_pfn, curr_iova->pfn_lo);
192 		new_pfn = (limit_pfn - size) & align_mask;
193 		prev = curr;
194 		curr = rb_prev(curr);
195 		curr_iova = rb_entry(curr, struct iova, node);
196 	} while (curr && new_pfn <= curr_iova->pfn_hi);
197 
198 	if (limit_pfn < size || new_pfn < iovad->start_pfn) {
199 		iovad->max32_alloc_size = size;
200 		goto iova32_full;
201 	}
202 
203 	/* pfn_lo will point to size aligned address if size_aligned is set */
204 	new->pfn_lo = new_pfn;
205 	new->pfn_hi = new->pfn_lo + size - 1;
206 
207 	/* If we have 'prev', it's a valid place to start the insertion. */
208 	iova_insert_rbtree(&iovad->rbroot, new, prev);
209 	__cached_rbnode_insert_update(iovad, new);
210 
211 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
212 	return 0;
213 
214 iova32_full:
215 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
216 	return -ENOMEM;
217 }
218 
219 static struct kmem_cache *iova_cache;
220 static unsigned int iova_cache_users;
221 static DEFINE_MUTEX(iova_cache_mutex);
222 
223 struct iova *alloc_iova_mem(void)
224 {
225 	return kmem_cache_alloc(iova_cache, GFP_ATOMIC);
226 }
227 EXPORT_SYMBOL(alloc_iova_mem);
228 
229 void free_iova_mem(struct iova *iova)
230 {
231 	if (iova->pfn_lo != IOVA_ANCHOR)
232 		kmem_cache_free(iova_cache, iova);
233 }
234 EXPORT_SYMBOL(free_iova_mem);
235 
236 int iova_cache_get(void)
237 {
238 	mutex_lock(&iova_cache_mutex);
239 	if (!iova_cache_users) {
240 		iova_cache = kmem_cache_create(
241 			"iommu_iova", sizeof(struct iova), 0,
242 			SLAB_HWCACHE_ALIGN, NULL);
243 		if (!iova_cache) {
244 			mutex_unlock(&iova_cache_mutex);
245 			printk(KERN_ERR "Couldn't create iova cache\n");
246 			return -ENOMEM;
247 		}
248 	}
249 
250 	iova_cache_users++;
251 	mutex_unlock(&iova_cache_mutex);
252 
253 	return 0;
254 }
255 EXPORT_SYMBOL_GPL(iova_cache_get);
256 
257 void iova_cache_put(void)
258 {
259 	mutex_lock(&iova_cache_mutex);
260 	if (WARN_ON(!iova_cache_users)) {
261 		mutex_unlock(&iova_cache_mutex);
262 		return;
263 	}
264 	iova_cache_users--;
265 	if (!iova_cache_users)
266 		kmem_cache_destroy(iova_cache);
267 	mutex_unlock(&iova_cache_mutex);
268 }
269 EXPORT_SYMBOL_GPL(iova_cache_put);
270 
271 /**
272  * alloc_iova - allocates an iova
273  * @iovad: - iova domain in question
274  * @size: - size of page frames to allocate
275  * @limit_pfn: - max limit address
276  * @size_aligned: - set if size_aligned address range is required
277  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
278  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
279  * flag is set then the allocated address iova->pfn_lo will be naturally
280  * aligned on roundup_power_of_two(size).
281  */
282 struct iova *
283 alloc_iova(struct iova_domain *iovad, unsigned long size,
284 	unsigned long limit_pfn,
285 	bool size_aligned)
286 {
287 	struct iova *new_iova;
288 	int ret;
289 
290 	new_iova = alloc_iova_mem();
291 	if (!new_iova)
292 		return NULL;
293 
294 	ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
295 			new_iova, size_aligned);
296 
297 	if (ret) {
298 		free_iova_mem(new_iova);
299 		return NULL;
300 	}
301 
302 	return new_iova;
303 }
304 EXPORT_SYMBOL_GPL(alloc_iova);
305 
306 static struct iova *
307 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
308 {
309 	struct rb_node *node = iovad->rbroot.rb_node;
310 
311 	assert_spin_locked(&iovad->iova_rbtree_lock);
312 
313 	while (node) {
314 		struct iova *iova = rb_entry(node, struct iova, node);
315 
316 		if (pfn < iova->pfn_lo)
317 			node = node->rb_left;
318 		else if (pfn > iova->pfn_hi)
319 			node = node->rb_right;
320 		else
321 			return iova;	/* pfn falls within iova's range */
322 	}
323 
324 	return NULL;
325 }
326 
327 static void private_free_iova(struct iova_domain *iovad, struct iova *iova)
328 {
329 	assert_spin_locked(&iovad->iova_rbtree_lock);
330 	__cached_rbnode_delete_update(iovad, iova);
331 	rb_erase(&iova->node, &iovad->rbroot);
332 	free_iova_mem(iova);
333 }
334 
335 /**
336  * find_iova - finds an iova for a given pfn
337  * @iovad: - iova domain in question.
338  * @pfn: - page frame number
339  * This function finds and returns an iova belonging to the
340  * given doamin which matches the given pfn.
341  */
342 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
343 {
344 	unsigned long flags;
345 	struct iova *iova;
346 
347 	/* Take the lock so that no other thread is manipulating the rbtree */
348 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
349 	iova = private_find_iova(iovad, pfn);
350 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
351 	return iova;
352 }
353 EXPORT_SYMBOL_GPL(find_iova);
354 
355 /**
356  * __free_iova - frees the given iova
357  * @iovad: iova domain in question.
358  * @iova: iova in question.
359  * Frees the given iova belonging to the giving domain
360  */
361 void
362 __free_iova(struct iova_domain *iovad, struct iova *iova)
363 {
364 	unsigned long flags;
365 
366 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
367 	private_free_iova(iovad, iova);
368 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
369 }
370 EXPORT_SYMBOL_GPL(__free_iova);
371 
372 /**
373  * free_iova - finds and frees the iova for a given pfn
374  * @iovad: - iova domain in question.
375  * @pfn: - pfn that is allocated previously
376  * This functions finds an iova for a given pfn and then
377  * frees the iova from that domain.
378  */
379 void
380 free_iova(struct iova_domain *iovad, unsigned long pfn)
381 {
382 	struct iova *iova = find_iova(iovad, pfn);
383 
384 	if (iova)
385 		__free_iova(iovad, iova);
386 
387 }
388 EXPORT_SYMBOL_GPL(free_iova);
389 
390 /**
391  * alloc_iova_fast - allocates an iova from rcache
392  * @iovad: - iova domain in question
393  * @size: - size of page frames to allocate
394  * @limit_pfn: - max limit address
395  * @flush_rcache: - set to flush rcache on regular allocation failure
396  * This function tries to satisfy an iova allocation from the rcache,
397  * and falls back to regular allocation on failure. If regular allocation
398  * fails too and the flush_rcache flag is set then the rcache will be flushed.
399 */
400 unsigned long
401 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
402 		unsigned long limit_pfn, bool flush_rcache)
403 {
404 	unsigned long iova_pfn;
405 	struct iova *new_iova;
406 
407 	iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
408 	if (iova_pfn)
409 		return iova_pfn;
410 
411 retry:
412 	new_iova = alloc_iova(iovad, size, limit_pfn, true);
413 	if (!new_iova) {
414 		unsigned int cpu;
415 
416 		if (!flush_rcache)
417 			return 0;
418 
419 		/* Try replenishing IOVAs by flushing rcache. */
420 		flush_rcache = false;
421 		for_each_online_cpu(cpu)
422 			free_cpu_cached_iovas(cpu, iovad);
423 		goto retry;
424 	}
425 
426 	return new_iova->pfn_lo;
427 }
428 EXPORT_SYMBOL_GPL(alloc_iova_fast);
429 
430 /**
431  * free_iova_fast - free iova pfn range into rcache
432  * @iovad: - iova domain in question.
433  * @pfn: - pfn that is allocated previously
434  * @size: - # of pages in range
435  * This functions frees an iova range by trying to put it into the rcache,
436  * falling back to regular iova deallocation via free_iova() if this fails.
437  */
438 void
439 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
440 {
441 	if (iova_rcache_insert(iovad, pfn, size))
442 		return;
443 
444 	free_iova(iovad, pfn);
445 }
446 EXPORT_SYMBOL_GPL(free_iova_fast);
447 
448 #define fq_ring_for_each(i, fq) \
449 	for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
450 
451 static inline bool fq_full(struct iova_fq *fq)
452 {
453 	assert_spin_locked(&fq->lock);
454 	return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
455 }
456 
457 static inline unsigned fq_ring_add(struct iova_fq *fq)
458 {
459 	unsigned idx = fq->tail;
460 
461 	assert_spin_locked(&fq->lock);
462 
463 	fq->tail = (idx + 1) % IOVA_FQ_SIZE;
464 
465 	return idx;
466 }
467 
468 static void fq_ring_free(struct iova_domain *iovad, struct iova_fq *fq)
469 {
470 	u64 counter = atomic64_read(&iovad->fq_flush_finish_cnt);
471 	unsigned idx;
472 
473 	assert_spin_locked(&fq->lock);
474 
475 	fq_ring_for_each(idx, fq) {
476 
477 		if (fq->entries[idx].counter >= counter)
478 			break;
479 
480 		if (iovad->entry_dtor)
481 			iovad->entry_dtor(fq->entries[idx].data);
482 
483 		free_iova_fast(iovad,
484 			       fq->entries[idx].iova_pfn,
485 			       fq->entries[idx].pages);
486 
487 		fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
488 	}
489 }
490 
491 static void iova_domain_flush(struct iova_domain *iovad)
492 {
493 	atomic64_inc(&iovad->fq_flush_start_cnt);
494 	iovad->flush_cb(iovad);
495 	atomic64_inc(&iovad->fq_flush_finish_cnt);
496 }
497 
498 static void fq_destroy_all_entries(struct iova_domain *iovad)
499 {
500 	int cpu;
501 
502 	/*
503 	 * This code runs when the iova_domain is being detroyed, so don't
504 	 * bother to free iovas, just call the entry_dtor on all remaining
505 	 * entries.
506 	 */
507 	if (!iovad->entry_dtor)
508 		return;
509 
510 	for_each_possible_cpu(cpu) {
511 		struct iova_fq *fq = per_cpu_ptr(iovad->fq, cpu);
512 		int idx;
513 
514 		fq_ring_for_each(idx, fq)
515 			iovad->entry_dtor(fq->entries[idx].data);
516 	}
517 }
518 
519 static void fq_flush_timeout(struct timer_list *t)
520 {
521 	struct iova_domain *iovad = from_timer(iovad, t, fq_timer);
522 	int cpu;
523 
524 	atomic_set(&iovad->fq_timer_on, 0);
525 	iova_domain_flush(iovad);
526 
527 	for_each_possible_cpu(cpu) {
528 		unsigned long flags;
529 		struct iova_fq *fq;
530 
531 		fq = per_cpu_ptr(iovad->fq, cpu);
532 		spin_lock_irqsave(&fq->lock, flags);
533 		fq_ring_free(iovad, fq);
534 		spin_unlock_irqrestore(&fq->lock, flags);
535 	}
536 }
537 
538 void queue_iova(struct iova_domain *iovad,
539 		unsigned long pfn, unsigned long pages,
540 		unsigned long data)
541 {
542 	struct iova_fq *fq = raw_cpu_ptr(iovad->fq);
543 	unsigned long flags;
544 	unsigned idx;
545 
546 	spin_lock_irqsave(&fq->lock, flags);
547 
548 	/*
549 	 * First remove all entries from the flush queue that have already been
550 	 * flushed out on another CPU. This makes the fq_full() check below less
551 	 * likely to be true.
552 	 */
553 	fq_ring_free(iovad, fq);
554 
555 	if (fq_full(fq)) {
556 		iova_domain_flush(iovad);
557 		fq_ring_free(iovad, fq);
558 	}
559 
560 	idx = fq_ring_add(fq);
561 
562 	fq->entries[idx].iova_pfn = pfn;
563 	fq->entries[idx].pages    = pages;
564 	fq->entries[idx].data     = data;
565 	fq->entries[idx].counter  = atomic64_read(&iovad->fq_flush_start_cnt);
566 
567 	spin_unlock_irqrestore(&fq->lock, flags);
568 
569 	if (atomic_cmpxchg(&iovad->fq_timer_on, 0, 1) == 0)
570 		mod_timer(&iovad->fq_timer,
571 			  jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
572 }
573 EXPORT_SYMBOL_GPL(queue_iova);
574 
575 /**
576  * put_iova_domain - destroys the iova doamin
577  * @iovad: - iova domain in question.
578  * All the iova's in that domain are destroyed.
579  */
580 void put_iova_domain(struct iova_domain *iovad)
581 {
582 	struct iova *iova, *tmp;
583 
584 	free_iova_flush_queue(iovad);
585 	free_iova_rcaches(iovad);
586 	rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
587 		free_iova_mem(iova);
588 }
589 EXPORT_SYMBOL_GPL(put_iova_domain);
590 
591 static int
592 __is_range_overlap(struct rb_node *node,
593 	unsigned long pfn_lo, unsigned long pfn_hi)
594 {
595 	struct iova *iova = rb_entry(node, struct iova, node);
596 
597 	if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
598 		return 1;
599 	return 0;
600 }
601 
602 static inline struct iova *
603 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
604 {
605 	struct iova *iova;
606 
607 	iova = alloc_iova_mem();
608 	if (iova) {
609 		iova->pfn_lo = pfn_lo;
610 		iova->pfn_hi = pfn_hi;
611 	}
612 
613 	return iova;
614 }
615 
616 static struct iova *
617 __insert_new_range(struct iova_domain *iovad,
618 	unsigned long pfn_lo, unsigned long pfn_hi)
619 {
620 	struct iova *iova;
621 
622 	iova = alloc_and_init_iova(pfn_lo, pfn_hi);
623 	if (iova)
624 		iova_insert_rbtree(&iovad->rbroot, iova, NULL);
625 
626 	return iova;
627 }
628 
629 static void
630 __adjust_overlap_range(struct iova *iova,
631 	unsigned long *pfn_lo, unsigned long *pfn_hi)
632 {
633 	if (*pfn_lo < iova->pfn_lo)
634 		iova->pfn_lo = *pfn_lo;
635 	if (*pfn_hi > iova->pfn_hi)
636 		*pfn_lo = iova->pfn_hi + 1;
637 }
638 
639 /**
640  * reserve_iova - reserves an iova in the given range
641  * @iovad: - iova domain pointer
642  * @pfn_lo: - lower page frame address
643  * @pfn_hi:- higher pfn adderss
644  * This function allocates reserves the address range from pfn_lo to pfn_hi so
645  * that this address is not dished out as part of alloc_iova.
646  */
647 struct iova *
648 reserve_iova(struct iova_domain *iovad,
649 	unsigned long pfn_lo, unsigned long pfn_hi)
650 {
651 	struct rb_node *node;
652 	unsigned long flags;
653 	struct iova *iova;
654 	unsigned int overlap = 0;
655 
656 	/* Don't allow nonsensical pfns */
657 	if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
658 		return NULL;
659 
660 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
661 	for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
662 		if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
663 			iova = rb_entry(node, struct iova, node);
664 			__adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
665 			if ((pfn_lo >= iova->pfn_lo) &&
666 				(pfn_hi <= iova->pfn_hi))
667 				goto finish;
668 			overlap = 1;
669 
670 		} else if (overlap)
671 				break;
672 	}
673 
674 	/* We are here either because this is the first reserver node
675 	 * or need to insert remaining non overlap addr range
676 	 */
677 	iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
678 finish:
679 
680 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
681 	return iova;
682 }
683 EXPORT_SYMBOL_GPL(reserve_iova);
684 
685 /**
686  * copy_reserved_iova - copies the reserved between domains
687  * @from: - source doamin from where to copy
688  * @to: - destination domin where to copy
689  * This function copies reserved iova's from one doamin to
690  * other.
691  */
692 void
693 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
694 {
695 	unsigned long flags;
696 	struct rb_node *node;
697 
698 	spin_lock_irqsave(&from->iova_rbtree_lock, flags);
699 	for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
700 		struct iova *iova = rb_entry(node, struct iova, node);
701 		struct iova *new_iova;
702 
703 		if (iova->pfn_lo == IOVA_ANCHOR)
704 			continue;
705 
706 		new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
707 		if (!new_iova)
708 			printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
709 				iova->pfn_lo, iova->pfn_lo);
710 	}
711 	spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
712 }
713 EXPORT_SYMBOL_GPL(copy_reserved_iova);
714 
715 struct iova *
716 split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
717 		      unsigned long pfn_lo, unsigned long pfn_hi)
718 {
719 	unsigned long flags;
720 	struct iova *prev = NULL, *next = NULL;
721 
722 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
723 	if (iova->pfn_lo < pfn_lo) {
724 		prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
725 		if (prev == NULL)
726 			goto error;
727 	}
728 	if (iova->pfn_hi > pfn_hi) {
729 		next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
730 		if (next == NULL)
731 			goto error;
732 	}
733 
734 	__cached_rbnode_delete_update(iovad, iova);
735 	rb_erase(&iova->node, &iovad->rbroot);
736 
737 	if (prev) {
738 		iova_insert_rbtree(&iovad->rbroot, prev, NULL);
739 		iova->pfn_lo = pfn_lo;
740 	}
741 	if (next) {
742 		iova_insert_rbtree(&iovad->rbroot, next, NULL);
743 		iova->pfn_hi = pfn_hi;
744 	}
745 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
746 
747 	return iova;
748 
749 error:
750 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
751 	if (prev)
752 		free_iova_mem(prev);
753 	return NULL;
754 }
755 
756 /*
757  * Magazine caches for IOVA ranges.  For an introduction to magazines,
758  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
759  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
760  * For simplicity, we use a static magazine size and don't implement the
761  * dynamic size tuning described in the paper.
762  */
763 
764 #define IOVA_MAG_SIZE 128
765 
766 struct iova_magazine {
767 	unsigned long size;
768 	unsigned long pfns[IOVA_MAG_SIZE];
769 };
770 
771 struct iova_cpu_rcache {
772 	spinlock_t lock;
773 	struct iova_magazine *loaded;
774 	struct iova_magazine *prev;
775 };
776 
777 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
778 {
779 	return kzalloc(sizeof(struct iova_magazine), flags);
780 }
781 
782 static void iova_magazine_free(struct iova_magazine *mag)
783 {
784 	kfree(mag);
785 }
786 
787 static void
788 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
789 {
790 	unsigned long flags;
791 	int i;
792 
793 	if (!mag)
794 		return;
795 
796 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
797 
798 	for (i = 0 ; i < mag->size; ++i) {
799 		struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
800 
801 		BUG_ON(!iova);
802 		private_free_iova(iovad, iova);
803 	}
804 
805 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
806 
807 	mag->size = 0;
808 }
809 
810 static bool iova_magazine_full(struct iova_magazine *mag)
811 {
812 	return (mag && mag->size == IOVA_MAG_SIZE);
813 }
814 
815 static bool iova_magazine_empty(struct iova_magazine *mag)
816 {
817 	return (!mag || mag->size == 0);
818 }
819 
820 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
821 				       unsigned long limit_pfn)
822 {
823 	int i;
824 	unsigned long pfn;
825 
826 	BUG_ON(iova_magazine_empty(mag));
827 
828 	/* Only fall back to the rbtree if we have no suitable pfns at all */
829 	for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
830 		if (i == 0)
831 			return 0;
832 
833 	/* Swap it to pop it */
834 	pfn = mag->pfns[i];
835 	mag->pfns[i] = mag->pfns[--mag->size];
836 
837 	return pfn;
838 }
839 
840 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
841 {
842 	BUG_ON(iova_magazine_full(mag));
843 
844 	mag->pfns[mag->size++] = pfn;
845 }
846 
847 static void init_iova_rcaches(struct iova_domain *iovad)
848 {
849 	struct iova_cpu_rcache *cpu_rcache;
850 	struct iova_rcache *rcache;
851 	unsigned int cpu;
852 	int i;
853 
854 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
855 		rcache = &iovad->rcaches[i];
856 		spin_lock_init(&rcache->lock);
857 		rcache->depot_size = 0;
858 		rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
859 		if (WARN_ON(!rcache->cpu_rcaches))
860 			continue;
861 		for_each_possible_cpu(cpu) {
862 			cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
863 			spin_lock_init(&cpu_rcache->lock);
864 			cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
865 			cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
866 		}
867 	}
868 }
869 
870 /*
871  * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
872  * return true on success.  Can fail if rcache is full and we can't free
873  * space, and free_iova() (our only caller) will then return the IOVA
874  * range to the rbtree instead.
875  */
876 static bool __iova_rcache_insert(struct iova_domain *iovad,
877 				 struct iova_rcache *rcache,
878 				 unsigned long iova_pfn)
879 {
880 	struct iova_magazine *mag_to_free = NULL;
881 	struct iova_cpu_rcache *cpu_rcache;
882 	bool can_insert = false;
883 	unsigned long flags;
884 
885 	cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
886 	spin_lock_irqsave(&cpu_rcache->lock, flags);
887 
888 	if (!iova_magazine_full(cpu_rcache->loaded)) {
889 		can_insert = true;
890 	} else if (!iova_magazine_full(cpu_rcache->prev)) {
891 		swap(cpu_rcache->prev, cpu_rcache->loaded);
892 		can_insert = true;
893 	} else {
894 		struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
895 
896 		if (new_mag) {
897 			spin_lock(&rcache->lock);
898 			if (rcache->depot_size < MAX_GLOBAL_MAGS) {
899 				rcache->depot[rcache->depot_size++] =
900 						cpu_rcache->loaded;
901 			} else {
902 				mag_to_free = cpu_rcache->loaded;
903 			}
904 			spin_unlock(&rcache->lock);
905 
906 			cpu_rcache->loaded = new_mag;
907 			can_insert = true;
908 		}
909 	}
910 
911 	if (can_insert)
912 		iova_magazine_push(cpu_rcache->loaded, iova_pfn);
913 
914 	spin_unlock_irqrestore(&cpu_rcache->lock, flags);
915 
916 	if (mag_to_free) {
917 		iova_magazine_free_pfns(mag_to_free, iovad);
918 		iova_magazine_free(mag_to_free);
919 	}
920 
921 	return can_insert;
922 }
923 
924 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
925 			       unsigned long size)
926 {
927 	unsigned int log_size = order_base_2(size);
928 
929 	if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
930 		return false;
931 
932 	return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
933 }
934 
935 /*
936  * Caller wants to allocate a new IOVA range from 'rcache'.  If we can
937  * satisfy the request, return a matching non-NULL range and remove
938  * it from the 'rcache'.
939  */
940 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
941 				       unsigned long limit_pfn)
942 {
943 	struct iova_cpu_rcache *cpu_rcache;
944 	unsigned long iova_pfn = 0;
945 	bool has_pfn = false;
946 	unsigned long flags;
947 
948 	cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
949 	spin_lock_irqsave(&cpu_rcache->lock, flags);
950 
951 	if (!iova_magazine_empty(cpu_rcache->loaded)) {
952 		has_pfn = true;
953 	} else if (!iova_magazine_empty(cpu_rcache->prev)) {
954 		swap(cpu_rcache->prev, cpu_rcache->loaded);
955 		has_pfn = true;
956 	} else {
957 		spin_lock(&rcache->lock);
958 		if (rcache->depot_size > 0) {
959 			iova_magazine_free(cpu_rcache->loaded);
960 			cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
961 			has_pfn = true;
962 		}
963 		spin_unlock(&rcache->lock);
964 	}
965 
966 	if (has_pfn)
967 		iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
968 
969 	spin_unlock_irqrestore(&cpu_rcache->lock, flags);
970 
971 	return iova_pfn;
972 }
973 
974 /*
975  * Try to satisfy IOVA allocation range from rcache.  Fail if requested
976  * size is too big or the DMA limit we are given isn't satisfied by the
977  * top element in the magazine.
978  */
979 static unsigned long iova_rcache_get(struct iova_domain *iovad,
980 				     unsigned long size,
981 				     unsigned long limit_pfn)
982 {
983 	unsigned int log_size = order_base_2(size);
984 
985 	if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
986 		return 0;
987 
988 	return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
989 }
990 
991 /*
992  * free rcache data structures.
993  */
994 static void free_iova_rcaches(struct iova_domain *iovad)
995 {
996 	struct iova_rcache *rcache;
997 	struct iova_cpu_rcache *cpu_rcache;
998 	unsigned int cpu;
999 	int i, j;
1000 
1001 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1002 		rcache = &iovad->rcaches[i];
1003 		for_each_possible_cpu(cpu) {
1004 			cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1005 			iova_magazine_free(cpu_rcache->loaded);
1006 			iova_magazine_free(cpu_rcache->prev);
1007 		}
1008 		free_percpu(rcache->cpu_rcaches);
1009 		for (j = 0; j < rcache->depot_size; ++j)
1010 			iova_magazine_free(rcache->depot[j]);
1011 	}
1012 }
1013 
1014 /*
1015  * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
1016  */
1017 void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
1018 {
1019 	struct iova_cpu_rcache *cpu_rcache;
1020 	struct iova_rcache *rcache;
1021 	unsigned long flags;
1022 	int i;
1023 
1024 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1025 		rcache = &iovad->rcaches[i];
1026 		cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1027 		spin_lock_irqsave(&cpu_rcache->lock, flags);
1028 		iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
1029 		iova_magazine_free_pfns(cpu_rcache->prev, iovad);
1030 		spin_unlock_irqrestore(&cpu_rcache->lock, flags);
1031 	}
1032 }
1033 
1034 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
1035 MODULE_LICENSE("GPL");
1036