xref: /openbmc/linux/drivers/iommu/iova.c (revision 3381df09)
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 bool has_iova_flush_queue(struct iova_domain *iovad)
58 {
59 	return !!iovad->fq;
60 }
61 
62 static void free_iova_flush_queue(struct iova_domain *iovad)
63 {
64 	if (!has_iova_flush_queue(iovad))
65 		return;
66 
67 	if (timer_pending(&iovad->fq_timer))
68 		del_timer(&iovad->fq_timer);
69 
70 	fq_destroy_all_entries(iovad);
71 
72 	free_percpu(iovad->fq);
73 
74 	iovad->fq         = NULL;
75 	iovad->flush_cb   = NULL;
76 	iovad->entry_dtor = NULL;
77 }
78 
79 int init_iova_flush_queue(struct iova_domain *iovad,
80 			  iova_flush_cb flush_cb, iova_entry_dtor entry_dtor)
81 {
82 	struct iova_fq __percpu *queue;
83 	int cpu;
84 
85 	atomic64_set(&iovad->fq_flush_start_cnt,  0);
86 	atomic64_set(&iovad->fq_flush_finish_cnt, 0);
87 
88 	queue = alloc_percpu(struct iova_fq);
89 	if (!queue)
90 		return -ENOMEM;
91 
92 	iovad->flush_cb   = flush_cb;
93 	iovad->entry_dtor = entry_dtor;
94 
95 	for_each_possible_cpu(cpu) {
96 		struct iova_fq *fq;
97 
98 		fq = per_cpu_ptr(queue, cpu);
99 		fq->head = 0;
100 		fq->tail = 0;
101 
102 		spin_lock_init(&fq->lock);
103 	}
104 
105 	smp_wmb();
106 
107 	iovad->fq = queue;
108 
109 	timer_setup(&iovad->fq_timer, fq_flush_timeout, 0);
110 	atomic_set(&iovad->fq_timer_on, 0);
111 
112 	return 0;
113 }
114 EXPORT_SYMBOL_GPL(init_iova_flush_queue);
115 
116 static struct rb_node *
117 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
118 {
119 	if (limit_pfn <= iovad->dma_32bit_pfn)
120 		return iovad->cached32_node;
121 
122 	return iovad->cached_node;
123 }
124 
125 static void
126 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
127 {
128 	if (new->pfn_hi < iovad->dma_32bit_pfn)
129 		iovad->cached32_node = &new->node;
130 	else
131 		iovad->cached_node = &new->node;
132 }
133 
134 static void
135 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
136 {
137 	struct iova *cached_iova;
138 
139 	cached_iova = rb_entry(iovad->cached32_node, struct iova, node);
140 	if (free == cached_iova ||
141 	    (free->pfn_hi < iovad->dma_32bit_pfn &&
142 	     free->pfn_lo >= cached_iova->pfn_lo)) {
143 		iovad->cached32_node = rb_next(&free->node);
144 		iovad->max32_alloc_size = iovad->dma_32bit_pfn;
145 	}
146 
147 	cached_iova = rb_entry(iovad->cached_node, struct iova, node);
148 	if (free->pfn_lo >= cached_iova->pfn_lo)
149 		iovad->cached_node = rb_next(&free->node);
150 }
151 
152 /* Insert the iova into domain rbtree by holding writer lock */
153 static void
154 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
155 		   struct rb_node *start)
156 {
157 	struct rb_node **new, *parent = NULL;
158 
159 	new = (start) ? &start : &(root->rb_node);
160 	/* Figure out where to put new node */
161 	while (*new) {
162 		struct iova *this = rb_entry(*new, struct iova, node);
163 
164 		parent = *new;
165 
166 		if (iova->pfn_lo < this->pfn_lo)
167 			new = &((*new)->rb_left);
168 		else if (iova->pfn_lo > this->pfn_lo)
169 			new = &((*new)->rb_right);
170 		else {
171 			WARN_ON(1); /* this should not happen */
172 			return;
173 		}
174 	}
175 	/* Add new node and rebalance tree. */
176 	rb_link_node(&iova->node, parent, new);
177 	rb_insert_color(&iova->node, root);
178 }
179 
180 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
181 		unsigned long size, unsigned long limit_pfn,
182 			struct iova *new, bool size_aligned)
183 {
184 	struct rb_node *curr, *prev;
185 	struct iova *curr_iova;
186 	unsigned long flags;
187 	unsigned long new_pfn;
188 	unsigned long align_mask = ~0UL;
189 
190 	if (size_aligned)
191 		align_mask <<= fls_long(size - 1);
192 
193 	/* Walk the tree backwards */
194 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
195 	if (limit_pfn <= iovad->dma_32bit_pfn &&
196 			size >= iovad->max32_alloc_size)
197 		goto iova32_full;
198 
199 	curr = __get_cached_rbnode(iovad, limit_pfn);
200 	curr_iova = rb_entry(curr, struct iova, node);
201 	do {
202 		limit_pfn = min(limit_pfn, curr_iova->pfn_lo);
203 		new_pfn = (limit_pfn - size) & align_mask;
204 		prev = curr;
205 		curr = rb_prev(curr);
206 		curr_iova = rb_entry(curr, struct iova, node);
207 	} while (curr && new_pfn <= curr_iova->pfn_hi);
208 
209 	if (limit_pfn < size || new_pfn < iovad->start_pfn) {
210 		iovad->max32_alloc_size = size;
211 		goto iova32_full;
212 	}
213 
214 	/* pfn_lo will point to size aligned address if size_aligned is set */
215 	new->pfn_lo = new_pfn;
216 	new->pfn_hi = new->pfn_lo + size - 1;
217 
218 	/* If we have 'prev', it's a valid place to start the insertion. */
219 	iova_insert_rbtree(&iovad->rbroot, new, prev);
220 	__cached_rbnode_insert_update(iovad, new);
221 
222 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
223 	return 0;
224 
225 iova32_full:
226 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
227 	return -ENOMEM;
228 }
229 
230 static struct kmem_cache *iova_cache;
231 static unsigned int iova_cache_users;
232 static DEFINE_MUTEX(iova_cache_mutex);
233 
234 struct iova *alloc_iova_mem(void)
235 {
236 	return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
237 }
238 EXPORT_SYMBOL(alloc_iova_mem);
239 
240 void free_iova_mem(struct iova *iova)
241 {
242 	if (iova->pfn_lo != IOVA_ANCHOR)
243 		kmem_cache_free(iova_cache, iova);
244 }
245 EXPORT_SYMBOL(free_iova_mem);
246 
247 int iova_cache_get(void)
248 {
249 	mutex_lock(&iova_cache_mutex);
250 	if (!iova_cache_users) {
251 		iova_cache = kmem_cache_create(
252 			"iommu_iova", sizeof(struct iova), 0,
253 			SLAB_HWCACHE_ALIGN, NULL);
254 		if (!iova_cache) {
255 			mutex_unlock(&iova_cache_mutex);
256 			printk(KERN_ERR "Couldn't create iova cache\n");
257 			return -ENOMEM;
258 		}
259 	}
260 
261 	iova_cache_users++;
262 	mutex_unlock(&iova_cache_mutex);
263 
264 	return 0;
265 }
266 EXPORT_SYMBOL_GPL(iova_cache_get);
267 
268 void iova_cache_put(void)
269 {
270 	mutex_lock(&iova_cache_mutex);
271 	if (WARN_ON(!iova_cache_users)) {
272 		mutex_unlock(&iova_cache_mutex);
273 		return;
274 	}
275 	iova_cache_users--;
276 	if (!iova_cache_users)
277 		kmem_cache_destroy(iova_cache);
278 	mutex_unlock(&iova_cache_mutex);
279 }
280 EXPORT_SYMBOL_GPL(iova_cache_put);
281 
282 /**
283  * alloc_iova - allocates an iova
284  * @iovad: - iova domain in question
285  * @size: - size of page frames to allocate
286  * @limit_pfn: - max limit address
287  * @size_aligned: - set if size_aligned address range is required
288  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
289  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
290  * flag is set then the allocated address iova->pfn_lo will be naturally
291  * aligned on roundup_power_of_two(size).
292  */
293 struct iova *
294 alloc_iova(struct iova_domain *iovad, unsigned long size,
295 	unsigned long limit_pfn,
296 	bool size_aligned)
297 {
298 	struct iova *new_iova;
299 	int ret;
300 
301 	new_iova = alloc_iova_mem();
302 	if (!new_iova)
303 		return NULL;
304 
305 	ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
306 			new_iova, size_aligned);
307 
308 	if (ret) {
309 		free_iova_mem(new_iova);
310 		return NULL;
311 	}
312 
313 	return new_iova;
314 }
315 EXPORT_SYMBOL_GPL(alloc_iova);
316 
317 static struct iova *
318 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
319 {
320 	struct rb_node *node = iovad->rbroot.rb_node;
321 
322 	assert_spin_locked(&iovad->iova_rbtree_lock);
323 
324 	while (node) {
325 		struct iova *iova = rb_entry(node, struct iova, node);
326 
327 		if (pfn < iova->pfn_lo)
328 			node = node->rb_left;
329 		else if (pfn > iova->pfn_hi)
330 			node = node->rb_right;
331 		else
332 			return iova;	/* pfn falls within iova's range */
333 	}
334 
335 	return NULL;
336 }
337 
338 static void private_free_iova(struct iova_domain *iovad, struct iova *iova)
339 {
340 	assert_spin_locked(&iovad->iova_rbtree_lock);
341 	__cached_rbnode_delete_update(iovad, iova);
342 	rb_erase(&iova->node, &iovad->rbroot);
343 	free_iova_mem(iova);
344 }
345 
346 /**
347  * find_iova - finds an iova for a given pfn
348  * @iovad: - iova domain in question.
349  * @pfn: - page frame number
350  * This function finds and returns an iova belonging to the
351  * given doamin which matches the given pfn.
352  */
353 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
354 {
355 	unsigned long flags;
356 	struct iova *iova;
357 
358 	/* Take the lock so that no other thread is manipulating the rbtree */
359 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
360 	iova = private_find_iova(iovad, pfn);
361 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
362 	return iova;
363 }
364 EXPORT_SYMBOL_GPL(find_iova);
365 
366 /**
367  * __free_iova - frees the given iova
368  * @iovad: iova domain in question.
369  * @iova: iova in question.
370  * Frees the given iova belonging to the giving domain
371  */
372 void
373 __free_iova(struct iova_domain *iovad, struct iova *iova)
374 {
375 	unsigned long flags;
376 
377 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
378 	private_free_iova(iovad, iova);
379 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
380 }
381 EXPORT_SYMBOL_GPL(__free_iova);
382 
383 /**
384  * free_iova - finds and frees the iova for a given pfn
385  * @iovad: - iova domain in question.
386  * @pfn: - pfn that is allocated previously
387  * This functions finds an iova for a given pfn and then
388  * frees the iova from that domain.
389  */
390 void
391 free_iova(struct iova_domain *iovad, unsigned long pfn)
392 {
393 	struct iova *iova = find_iova(iovad, pfn);
394 
395 	if (iova)
396 		__free_iova(iovad, iova);
397 
398 }
399 EXPORT_SYMBOL_GPL(free_iova);
400 
401 /**
402  * alloc_iova_fast - allocates an iova from rcache
403  * @iovad: - iova domain in question
404  * @size: - size of page frames to allocate
405  * @limit_pfn: - max limit address
406  * @flush_rcache: - set to flush rcache on regular allocation failure
407  * This function tries to satisfy an iova allocation from the rcache,
408  * and falls back to regular allocation on failure. If regular allocation
409  * fails too and the flush_rcache flag is set then the rcache will be flushed.
410 */
411 unsigned long
412 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
413 		unsigned long limit_pfn, bool flush_rcache)
414 {
415 	unsigned long iova_pfn;
416 	struct iova *new_iova;
417 
418 	iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
419 	if (iova_pfn)
420 		return iova_pfn;
421 
422 retry:
423 	new_iova = alloc_iova(iovad, size, limit_pfn, true);
424 	if (!new_iova) {
425 		unsigned int cpu;
426 
427 		if (!flush_rcache)
428 			return 0;
429 
430 		/* Try replenishing IOVAs by flushing rcache. */
431 		flush_rcache = false;
432 		for_each_online_cpu(cpu)
433 			free_cpu_cached_iovas(cpu, iovad);
434 		goto retry;
435 	}
436 
437 	return new_iova->pfn_lo;
438 }
439 EXPORT_SYMBOL_GPL(alloc_iova_fast);
440 
441 /**
442  * free_iova_fast - free iova pfn range into rcache
443  * @iovad: - iova domain in question.
444  * @pfn: - pfn that is allocated previously
445  * @size: - # of pages in range
446  * This functions frees an iova range by trying to put it into the rcache,
447  * falling back to regular iova deallocation via free_iova() if this fails.
448  */
449 void
450 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
451 {
452 	if (iova_rcache_insert(iovad, pfn, size))
453 		return;
454 
455 	free_iova(iovad, pfn);
456 }
457 EXPORT_SYMBOL_GPL(free_iova_fast);
458 
459 #define fq_ring_for_each(i, fq) \
460 	for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
461 
462 static inline bool fq_full(struct iova_fq *fq)
463 {
464 	assert_spin_locked(&fq->lock);
465 	return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
466 }
467 
468 static inline unsigned fq_ring_add(struct iova_fq *fq)
469 {
470 	unsigned idx = fq->tail;
471 
472 	assert_spin_locked(&fq->lock);
473 
474 	fq->tail = (idx + 1) % IOVA_FQ_SIZE;
475 
476 	return idx;
477 }
478 
479 static void fq_ring_free(struct iova_domain *iovad, struct iova_fq *fq)
480 {
481 	u64 counter = atomic64_read(&iovad->fq_flush_finish_cnt);
482 	unsigned idx;
483 
484 	assert_spin_locked(&fq->lock);
485 
486 	fq_ring_for_each(idx, fq) {
487 
488 		if (fq->entries[idx].counter >= counter)
489 			break;
490 
491 		if (iovad->entry_dtor)
492 			iovad->entry_dtor(fq->entries[idx].data);
493 
494 		free_iova_fast(iovad,
495 			       fq->entries[idx].iova_pfn,
496 			       fq->entries[idx].pages);
497 
498 		fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
499 	}
500 }
501 
502 static void iova_domain_flush(struct iova_domain *iovad)
503 {
504 	atomic64_inc(&iovad->fq_flush_start_cnt);
505 	iovad->flush_cb(iovad);
506 	atomic64_inc(&iovad->fq_flush_finish_cnt);
507 }
508 
509 static void fq_destroy_all_entries(struct iova_domain *iovad)
510 {
511 	int cpu;
512 
513 	/*
514 	 * This code runs when the iova_domain is being detroyed, so don't
515 	 * bother to free iovas, just call the entry_dtor on all remaining
516 	 * entries.
517 	 */
518 	if (!iovad->entry_dtor)
519 		return;
520 
521 	for_each_possible_cpu(cpu) {
522 		struct iova_fq *fq = per_cpu_ptr(iovad->fq, cpu);
523 		int idx;
524 
525 		fq_ring_for_each(idx, fq)
526 			iovad->entry_dtor(fq->entries[idx].data);
527 	}
528 }
529 
530 static void fq_flush_timeout(struct timer_list *t)
531 {
532 	struct iova_domain *iovad = from_timer(iovad, t, fq_timer);
533 	int cpu;
534 
535 	atomic_set(&iovad->fq_timer_on, 0);
536 	iova_domain_flush(iovad);
537 
538 	for_each_possible_cpu(cpu) {
539 		unsigned long flags;
540 		struct iova_fq *fq;
541 
542 		fq = per_cpu_ptr(iovad->fq, cpu);
543 		spin_lock_irqsave(&fq->lock, flags);
544 		fq_ring_free(iovad, fq);
545 		spin_unlock_irqrestore(&fq->lock, flags);
546 	}
547 }
548 
549 void queue_iova(struct iova_domain *iovad,
550 		unsigned long pfn, unsigned long pages,
551 		unsigned long data)
552 {
553 	struct iova_fq *fq = raw_cpu_ptr(iovad->fq);
554 	unsigned long flags;
555 	unsigned idx;
556 
557 	spin_lock_irqsave(&fq->lock, flags);
558 
559 	/*
560 	 * First remove all entries from the flush queue that have already been
561 	 * flushed out on another CPU. This makes the fq_full() check below less
562 	 * likely to be true.
563 	 */
564 	fq_ring_free(iovad, fq);
565 
566 	if (fq_full(fq)) {
567 		iova_domain_flush(iovad);
568 		fq_ring_free(iovad, fq);
569 	}
570 
571 	idx = fq_ring_add(fq);
572 
573 	fq->entries[idx].iova_pfn = pfn;
574 	fq->entries[idx].pages    = pages;
575 	fq->entries[idx].data     = data;
576 	fq->entries[idx].counter  = atomic64_read(&iovad->fq_flush_start_cnt);
577 
578 	spin_unlock_irqrestore(&fq->lock, flags);
579 
580 	/* Avoid false sharing as much as possible. */
581 	if (!atomic_read(&iovad->fq_timer_on) &&
582 	    !atomic_cmpxchg(&iovad->fq_timer_on, 0, 1))
583 		mod_timer(&iovad->fq_timer,
584 			  jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
585 }
586 EXPORT_SYMBOL_GPL(queue_iova);
587 
588 /**
589  * put_iova_domain - destroys the iova doamin
590  * @iovad: - iova domain in question.
591  * All the iova's in that domain are destroyed.
592  */
593 void put_iova_domain(struct iova_domain *iovad)
594 {
595 	struct iova *iova, *tmp;
596 
597 	free_iova_flush_queue(iovad);
598 	free_iova_rcaches(iovad);
599 	rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
600 		free_iova_mem(iova);
601 }
602 EXPORT_SYMBOL_GPL(put_iova_domain);
603 
604 static int
605 __is_range_overlap(struct rb_node *node,
606 	unsigned long pfn_lo, unsigned long pfn_hi)
607 {
608 	struct iova *iova = rb_entry(node, struct iova, node);
609 
610 	if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
611 		return 1;
612 	return 0;
613 }
614 
615 static inline struct iova *
616 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
617 {
618 	struct iova *iova;
619 
620 	iova = alloc_iova_mem();
621 	if (iova) {
622 		iova->pfn_lo = pfn_lo;
623 		iova->pfn_hi = pfn_hi;
624 	}
625 
626 	return iova;
627 }
628 
629 static struct iova *
630 __insert_new_range(struct iova_domain *iovad,
631 	unsigned long pfn_lo, unsigned long pfn_hi)
632 {
633 	struct iova *iova;
634 
635 	iova = alloc_and_init_iova(pfn_lo, pfn_hi);
636 	if (iova)
637 		iova_insert_rbtree(&iovad->rbroot, iova, NULL);
638 
639 	return iova;
640 }
641 
642 static void
643 __adjust_overlap_range(struct iova *iova,
644 	unsigned long *pfn_lo, unsigned long *pfn_hi)
645 {
646 	if (*pfn_lo < iova->pfn_lo)
647 		iova->pfn_lo = *pfn_lo;
648 	if (*pfn_hi > iova->pfn_hi)
649 		*pfn_lo = iova->pfn_hi + 1;
650 }
651 
652 /**
653  * reserve_iova - reserves an iova in the given range
654  * @iovad: - iova domain pointer
655  * @pfn_lo: - lower page frame address
656  * @pfn_hi:- higher pfn adderss
657  * This function allocates reserves the address range from pfn_lo to pfn_hi so
658  * that this address is not dished out as part of alloc_iova.
659  */
660 struct iova *
661 reserve_iova(struct iova_domain *iovad,
662 	unsigned long pfn_lo, unsigned long pfn_hi)
663 {
664 	struct rb_node *node;
665 	unsigned long flags;
666 	struct iova *iova;
667 	unsigned int overlap = 0;
668 
669 	/* Don't allow nonsensical pfns */
670 	if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
671 		return NULL;
672 
673 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
674 	for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
675 		if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
676 			iova = rb_entry(node, struct iova, node);
677 			__adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
678 			if ((pfn_lo >= iova->pfn_lo) &&
679 				(pfn_hi <= iova->pfn_hi))
680 				goto finish;
681 			overlap = 1;
682 
683 		} else if (overlap)
684 				break;
685 	}
686 
687 	/* We are here either because this is the first reserver node
688 	 * or need to insert remaining non overlap addr range
689 	 */
690 	iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
691 finish:
692 
693 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
694 	return iova;
695 }
696 EXPORT_SYMBOL_GPL(reserve_iova);
697 
698 /**
699  * copy_reserved_iova - copies the reserved between domains
700  * @from: - source doamin from where to copy
701  * @to: - destination domin where to copy
702  * This function copies reserved iova's from one doamin to
703  * other.
704  */
705 void
706 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
707 {
708 	unsigned long flags;
709 	struct rb_node *node;
710 
711 	spin_lock_irqsave(&from->iova_rbtree_lock, flags);
712 	for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
713 		struct iova *iova = rb_entry(node, struct iova, node);
714 		struct iova *new_iova;
715 
716 		if (iova->pfn_lo == IOVA_ANCHOR)
717 			continue;
718 
719 		new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
720 		if (!new_iova)
721 			printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
722 				iova->pfn_lo, iova->pfn_lo);
723 	}
724 	spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
725 }
726 EXPORT_SYMBOL_GPL(copy_reserved_iova);
727 
728 struct iova *
729 split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
730 		      unsigned long pfn_lo, unsigned long pfn_hi)
731 {
732 	unsigned long flags;
733 	struct iova *prev = NULL, *next = NULL;
734 
735 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
736 	if (iova->pfn_lo < pfn_lo) {
737 		prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
738 		if (prev == NULL)
739 			goto error;
740 	}
741 	if (iova->pfn_hi > pfn_hi) {
742 		next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
743 		if (next == NULL)
744 			goto error;
745 	}
746 
747 	__cached_rbnode_delete_update(iovad, iova);
748 	rb_erase(&iova->node, &iovad->rbroot);
749 
750 	if (prev) {
751 		iova_insert_rbtree(&iovad->rbroot, prev, NULL);
752 		iova->pfn_lo = pfn_lo;
753 	}
754 	if (next) {
755 		iova_insert_rbtree(&iovad->rbroot, next, NULL);
756 		iova->pfn_hi = pfn_hi;
757 	}
758 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
759 
760 	return iova;
761 
762 error:
763 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
764 	if (prev)
765 		free_iova_mem(prev);
766 	return NULL;
767 }
768 
769 /*
770  * Magazine caches for IOVA ranges.  For an introduction to magazines,
771  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
772  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
773  * For simplicity, we use a static magazine size and don't implement the
774  * dynamic size tuning described in the paper.
775  */
776 
777 #define IOVA_MAG_SIZE 128
778 
779 struct iova_magazine {
780 	unsigned long size;
781 	unsigned long pfns[IOVA_MAG_SIZE];
782 };
783 
784 struct iova_cpu_rcache {
785 	spinlock_t lock;
786 	struct iova_magazine *loaded;
787 	struct iova_magazine *prev;
788 };
789 
790 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
791 {
792 	return kzalloc(sizeof(struct iova_magazine), flags);
793 }
794 
795 static void iova_magazine_free(struct iova_magazine *mag)
796 {
797 	kfree(mag);
798 }
799 
800 static void
801 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
802 {
803 	unsigned long flags;
804 	int i;
805 
806 	if (!mag)
807 		return;
808 
809 	spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
810 
811 	for (i = 0 ; i < mag->size; ++i) {
812 		struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
813 
814 		BUG_ON(!iova);
815 		private_free_iova(iovad, iova);
816 	}
817 
818 	spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
819 
820 	mag->size = 0;
821 }
822 
823 static bool iova_magazine_full(struct iova_magazine *mag)
824 {
825 	return (mag && mag->size == IOVA_MAG_SIZE);
826 }
827 
828 static bool iova_magazine_empty(struct iova_magazine *mag)
829 {
830 	return (!mag || mag->size == 0);
831 }
832 
833 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
834 				       unsigned long limit_pfn)
835 {
836 	int i;
837 	unsigned long pfn;
838 
839 	BUG_ON(iova_magazine_empty(mag));
840 
841 	/* Only fall back to the rbtree if we have no suitable pfns at all */
842 	for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
843 		if (i == 0)
844 			return 0;
845 
846 	/* Swap it to pop it */
847 	pfn = mag->pfns[i];
848 	mag->pfns[i] = mag->pfns[--mag->size];
849 
850 	return pfn;
851 }
852 
853 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
854 {
855 	BUG_ON(iova_magazine_full(mag));
856 
857 	mag->pfns[mag->size++] = pfn;
858 }
859 
860 static void init_iova_rcaches(struct iova_domain *iovad)
861 {
862 	struct iova_cpu_rcache *cpu_rcache;
863 	struct iova_rcache *rcache;
864 	unsigned int cpu;
865 	int i;
866 
867 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
868 		rcache = &iovad->rcaches[i];
869 		spin_lock_init(&rcache->lock);
870 		rcache->depot_size = 0;
871 		rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
872 		if (WARN_ON(!rcache->cpu_rcaches))
873 			continue;
874 		for_each_possible_cpu(cpu) {
875 			cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
876 			spin_lock_init(&cpu_rcache->lock);
877 			cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
878 			cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
879 		}
880 	}
881 }
882 
883 /*
884  * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
885  * return true on success.  Can fail if rcache is full and we can't free
886  * space, and free_iova() (our only caller) will then return the IOVA
887  * range to the rbtree instead.
888  */
889 static bool __iova_rcache_insert(struct iova_domain *iovad,
890 				 struct iova_rcache *rcache,
891 				 unsigned long iova_pfn)
892 {
893 	struct iova_magazine *mag_to_free = NULL;
894 	struct iova_cpu_rcache *cpu_rcache;
895 	bool can_insert = false;
896 	unsigned long flags;
897 
898 	cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
899 	spin_lock_irqsave(&cpu_rcache->lock, flags);
900 
901 	if (!iova_magazine_full(cpu_rcache->loaded)) {
902 		can_insert = true;
903 	} else if (!iova_magazine_full(cpu_rcache->prev)) {
904 		swap(cpu_rcache->prev, cpu_rcache->loaded);
905 		can_insert = true;
906 	} else {
907 		struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
908 
909 		if (new_mag) {
910 			spin_lock(&rcache->lock);
911 			if (rcache->depot_size < MAX_GLOBAL_MAGS) {
912 				rcache->depot[rcache->depot_size++] =
913 						cpu_rcache->loaded;
914 			} else {
915 				mag_to_free = cpu_rcache->loaded;
916 			}
917 			spin_unlock(&rcache->lock);
918 
919 			cpu_rcache->loaded = new_mag;
920 			can_insert = true;
921 		}
922 	}
923 
924 	if (can_insert)
925 		iova_magazine_push(cpu_rcache->loaded, iova_pfn);
926 
927 	spin_unlock_irqrestore(&cpu_rcache->lock, flags);
928 
929 	if (mag_to_free) {
930 		iova_magazine_free_pfns(mag_to_free, iovad);
931 		iova_magazine_free(mag_to_free);
932 	}
933 
934 	return can_insert;
935 }
936 
937 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
938 			       unsigned long size)
939 {
940 	unsigned int log_size = order_base_2(size);
941 
942 	if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
943 		return false;
944 
945 	return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
946 }
947 
948 /*
949  * Caller wants to allocate a new IOVA range from 'rcache'.  If we can
950  * satisfy the request, return a matching non-NULL range and remove
951  * it from the 'rcache'.
952  */
953 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
954 				       unsigned long limit_pfn)
955 {
956 	struct iova_cpu_rcache *cpu_rcache;
957 	unsigned long iova_pfn = 0;
958 	bool has_pfn = false;
959 	unsigned long flags;
960 
961 	cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
962 	spin_lock_irqsave(&cpu_rcache->lock, flags);
963 
964 	if (!iova_magazine_empty(cpu_rcache->loaded)) {
965 		has_pfn = true;
966 	} else if (!iova_magazine_empty(cpu_rcache->prev)) {
967 		swap(cpu_rcache->prev, cpu_rcache->loaded);
968 		has_pfn = true;
969 	} else {
970 		spin_lock(&rcache->lock);
971 		if (rcache->depot_size > 0) {
972 			iova_magazine_free(cpu_rcache->loaded);
973 			cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
974 			has_pfn = true;
975 		}
976 		spin_unlock(&rcache->lock);
977 	}
978 
979 	if (has_pfn)
980 		iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
981 
982 	spin_unlock_irqrestore(&cpu_rcache->lock, flags);
983 
984 	return iova_pfn;
985 }
986 
987 /*
988  * Try to satisfy IOVA allocation range from rcache.  Fail if requested
989  * size is too big or the DMA limit we are given isn't satisfied by the
990  * top element in the magazine.
991  */
992 static unsigned long iova_rcache_get(struct iova_domain *iovad,
993 				     unsigned long size,
994 				     unsigned long limit_pfn)
995 {
996 	unsigned int log_size = order_base_2(size);
997 
998 	if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
999 		return 0;
1000 
1001 	return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
1002 }
1003 
1004 /*
1005  * free rcache data structures.
1006  */
1007 static void free_iova_rcaches(struct iova_domain *iovad)
1008 {
1009 	struct iova_rcache *rcache;
1010 	struct iova_cpu_rcache *cpu_rcache;
1011 	unsigned int cpu;
1012 	int i, j;
1013 
1014 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1015 		rcache = &iovad->rcaches[i];
1016 		for_each_possible_cpu(cpu) {
1017 			cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1018 			iova_magazine_free(cpu_rcache->loaded);
1019 			iova_magazine_free(cpu_rcache->prev);
1020 		}
1021 		free_percpu(rcache->cpu_rcaches);
1022 		for (j = 0; j < rcache->depot_size; ++j)
1023 			iova_magazine_free(rcache->depot[j]);
1024 	}
1025 }
1026 
1027 /*
1028  * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
1029  */
1030 void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
1031 {
1032 	struct iova_cpu_rcache *cpu_rcache;
1033 	struct iova_rcache *rcache;
1034 	unsigned long flags;
1035 	int i;
1036 
1037 	for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1038 		rcache = &iovad->rcaches[i];
1039 		cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1040 		spin_lock_irqsave(&cpu_rcache->lock, flags);
1041 		iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
1042 		iova_magazine_free_pfns(cpu_rcache->prev, iovad);
1043 		spin_unlock_irqrestore(&cpu_rcache->lock, flags);
1044 	}
1045 }
1046 
1047 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
1048 MODULE_LICENSE("GPL");
1049