xref: /openbmc/linux/mm/vmalloc.c (revision 9d56dd3b083a3bec56e9da35ce07baca81030b03)
1 /*
2  *  linux/mm/vmalloc.c
3  *
4  *  Copyright (C) 1993  Linus Torvalds
5  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6  *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7  *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8  *  Numa awareness, Christoph Lameter, SGI, June 2005
9  */
10 
11 #include <linux/vmalloc.h>
12 #include <linux/mm.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/debugobjects.h>
22 #include <linux/kallsyms.h>
23 #include <linux/list.h>
24 #include <linux/rbtree.h>
25 #include <linux/radix-tree.h>
26 #include <linux/rcupdate.h>
27 #include <linux/pfn.h>
28 #include <linux/kmemleak.h>
29 #include <asm/atomic.h>
30 #include <asm/uaccess.h>
31 #include <asm/tlbflush.h>
32 #include <asm/shmparam.h>
33 
34 
35 /*** Page table manipulation functions ***/
36 
37 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
38 {
39 	pte_t *pte;
40 
41 	pte = pte_offset_kernel(pmd, addr);
42 	do {
43 		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
44 		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
45 	} while (pte++, addr += PAGE_SIZE, addr != end);
46 }
47 
48 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
49 {
50 	pmd_t *pmd;
51 	unsigned long next;
52 
53 	pmd = pmd_offset(pud, addr);
54 	do {
55 		next = pmd_addr_end(addr, end);
56 		if (pmd_none_or_clear_bad(pmd))
57 			continue;
58 		vunmap_pte_range(pmd, addr, next);
59 	} while (pmd++, addr = next, addr != end);
60 }
61 
62 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
63 {
64 	pud_t *pud;
65 	unsigned long next;
66 
67 	pud = pud_offset(pgd, addr);
68 	do {
69 		next = pud_addr_end(addr, end);
70 		if (pud_none_or_clear_bad(pud))
71 			continue;
72 		vunmap_pmd_range(pud, addr, next);
73 	} while (pud++, addr = next, addr != end);
74 }
75 
76 static void vunmap_page_range(unsigned long addr, unsigned long end)
77 {
78 	pgd_t *pgd;
79 	unsigned long next;
80 
81 	BUG_ON(addr >= end);
82 	pgd = pgd_offset_k(addr);
83 	do {
84 		next = pgd_addr_end(addr, end);
85 		if (pgd_none_or_clear_bad(pgd))
86 			continue;
87 		vunmap_pud_range(pgd, addr, next);
88 	} while (pgd++, addr = next, addr != end);
89 }
90 
91 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
92 		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
93 {
94 	pte_t *pte;
95 
96 	/*
97 	 * nr is a running index into the array which helps higher level
98 	 * callers keep track of where we're up to.
99 	 */
100 
101 	pte = pte_alloc_kernel(pmd, addr);
102 	if (!pte)
103 		return -ENOMEM;
104 	do {
105 		struct page *page = pages[*nr];
106 
107 		if (WARN_ON(!pte_none(*pte)))
108 			return -EBUSY;
109 		if (WARN_ON(!page))
110 			return -ENOMEM;
111 		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
112 		(*nr)++;
113 	} while (pte++, addr += PAGE_SIZE, addr != end);
114 	return 0;
115 }
116 
117 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
118 		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
119 {
120 	pmd_t *pmd;
121 	unsigned long next;
122 
123 	pmd = pmd_alloc(&init_mm, pud, addr);
124 	if (!pmd)
125 		return -ENOMEM;
126 	do {
127 		next = pmd_addr_end(addr, end);
128 		if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
129 			return -ENOMEM;
130 	} while (pmd++, addr = next, addr != end);
131 	return 0;
132 }
133 
134 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
135 		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
136 {
137 	pud_t *pud;
138 	unsigned long next;
139 
140 	pud = pud_alloc(&init_mm, pgd, addr);
141 	if (!pud)
142 		return -ENOMEM;
143 	do {
144 		next = pud_addr_end(addr, end);
145 		if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
146 			return -ENOMEM;
147 	} while (pud++, addr = next, addr != end);
148 	return 0;
149 }
150 
151 /*
152  * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
153  * will have pfns corresponding to the "pages" array.
154  *
155  * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
156  */
157 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
158 				   pgprot_t prot, struct page **pages)
159 {
160 	pgd_t *pgd;
161 	unsigned long next;
162 	unsigned long addr = start;
163 	int err = 0;
164 	int nr = 0;
165 
166 	BUG_ON(addr >= end);
167 	pgd = pgd_offset_k(addr);
168 	do {
169 		next = pgd_addr_end(addr, end);
170 		err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
171 		if (err)
172 			return err;
173 	} while (pgd++, addr = next, addr != end);
174 
175 	return nr;
176 }
177 
178 static int vmap_page_range(unsigned long start, unsigned long end,
179 			   pgprot_t prot, struct page **pages)
180 {
181 	int ret;
182 
183 	ret = vmap_page_range_noflush(start, end, prot, pages);
184 	flush_cache_vmap(start, end);
185 	return ret;
186 }
187 
188 int is_vmalloc_or_module_addr(const void *x)
189 {
190 	/*
191 	 * ARM, x86-64 and sparc64 put modules in a special place,
192 	 * and fall back on vmalloc() if that fails. Others
193 	 * just put it in the vmalloc space.
194 	 */
195 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
196 	unsigned long addr = (unsigned long)x;
197 	if (addr >= MODULES_VADDR && addr < MODULES_END)
198 		return 1;
199 #endif
200 	return is_vmalloc_addr(x);
201 }
202 
203 /*
204  * Walk a vmap address to the struct page it maps.
205  */
206 struct page *vmalloc_to_page(const void *vmalloc_addr)
207 {
208 	unsigned long addr = (unsigned long) vmalloc_addr;
209 	struct page *page = NULL;
210 	pgd_t *pgd = pgd_offset_k(addr);
211 
212 	/*
213 	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
214 	 * architectures that do not vmalloc module space
215 	 */
216 	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
217 
218 	if (!pgd_none(*pgd)) {
219 		pud_t *pud = pud_offset(pgd, addr);
220 		if (!pud_none(*pud)) {
221 			pmd_t *pmd = pmd_offset(pud, addr);
222 			if (!pmd_none(*pmd)) {
223 				pte_t *ptep, pte;
224 
225 				ptep = pte_offset_map(pmd, addr);
226 				pte = *ptep;
227 				if (pte_present(pte))
228 					page = pte_page(pte);
229 				pte_unmap(ptep);
230 			}
231 		}
232 	}
233 	return page;
234 }
235 EXPORT_SYMBOL(vmalloc_to_page);
236 
237 /*
238  * Map a vmalloc()-space virtual address to the physical page frame number.
239  */
240 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
241 {
242 	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
243 }
244 EXPORT_SYMBOL(vmalloc_to_pfn);
245 
246 
247 /*** Global kva allocator ***/
248 
249 #define VM_LAZY_FREE	0x01
250 #define VM_LAZY_FREEING	0x02
251 #define VM_VM_AREA	0x04
252 
253 struct vmap_area {
254 	unsigned long va_start;
255 	unsigned long va_end;
256 	unsigned long flags;
257 	struct rb_node rb_node;		/* address sorted rbtree */
258 	struct list_head list;		/* address sorted list */
259 	struct list_head purge_list;	/* "lazy purge" list */
260 	void *private;
261 	struct rcu_head rcu_head;
262 };
263 
264 static DEFINE_SPINLOCK(vmap_area_lock);
265 static struct rb_root vmap_area_root = RB_ROOT;
266 static LIST_HEAD(vmap_area_list);
267 static unsigned long vmap_area_pcpu_hole;
268 
269 static struct vmap_area *__find_vmap_area(unsigned long addr)
270 {
271 	struct rb_node *n = vmap_area_root.rb_node;
272 
273 	while (n) {
274 		struct vmap_area *va;
275 
276 		va = rb_entry(n, struct vmap_area, rb_node);
277 		if (addr < va->va_start)
278 			n = n->rb_left;
279 		else if (addr > va->va_start)
280 			n = n->rb_right;
281 		else
282 			return va;
283 	}
284 
285 	return NULL;
286 }
287 
288 static void __insert_vmap_area(struct vmap_area *va)
289 {
290 	struct rb_node **p = &vmap_area_root.rb_node;
291 	struct rb_node *parent = NULL;
292 	struct rb_node *tmp;
293 
294 	while (*p) {
295 		struct vmap_area *tmp;
296 
297 		parent = *p;
298 		tmp = rb_entry(parent, struct vmap_area, rb_node);
299 		if (va->va_start < tmp->va_end)
300 			p = &(*p)->rb_left;
301 		else if (va->va_end > tmp->va_start)
302 			p = &(*p)->rb_right;
303 		else
304 			BUG();
305 	}
306 
307 	rb_link_node(&va->rb_node, parent, p);
308 	rb_insert_color(&va->rb_node, &vmap_area_root);
309 
310 	/* address-sort this list so it is usable like the vmlist */
311 	tmp = rb_prev(&va->rb_node);
312 	if (tmp) {
313 		struct vmap_area *prev;
314 		prev = rb_entry(tmp, struct vmap_area, rb_node);
315 		list_add_rcu(&va->list, &prev->list);
316 	} else
317 		list_add_rcu(&va->list, &vmap_area_list);
318 }
319 
320 static void purge_vmap_area_lazy(void);
321 
322 /*
323  * Allocate a region of KVA of the specified size and alignment, within the
324  * vstart and vend.
325  */
326 static struct vmap_area *alloc_vmap_area(unsigned long size,
327 				unsigned long align,
328 				unsigned long vstart, unsigned long vend,
329 				int node, gfp_t gfp_mask)
330 {
331 	struct vmap_area *va;
332 	struct rb_node *n;
333 	unsigned long addr;
334 	int purged = 0;
335 
336 	BUG_ON(!size);
337 	BUG_ON(size & ~PAGE_MASK);
338 
339 	va = kmalloc_node(sizeof(struct vmap_area),
340 			gfp_mask & GFP_RECLAIM_MASK, node);
341 	if (unlikely(!va))
342 		return ERR_PTR(-ENOMEM);
343 
344 retry:
345 	addr = ALIGN(vstart, align);
346 
347 	spin_lock(&vmap_area_lock);
348 	if (addr + size - 1 < addr)
349 		goto overflow;
350 
351 	/* XXX: could have a last_hole cache */
352 	n = vmap_area_root.rb_node;
353 	if (n) {
354 		struct vmap_area *first = NULL;
355 
356 		do {
357 			struct vmap_area *tmp;
358 			tmp = rb_entry(n, struct vmap_area, rb_node);
359 			if (tmp->va_end >= addr) {
360 				if (!first && tmp->va_start < addr + size)
361 					first = tmp;
362 				n = n->rb_left;
363 			} else {
364 				first = tmp;
365 				n = n->rb_right;
366 			}
367 		} while (n);
368 
369 		if (!first)
370 			goto found;
371 
372 		if (first->va_end < addr) {
373 			n = rb_next(&first->rb_node);
374 			if (n)
375 				first = rb_entry(n, struct vmap_area, rb_node);
376 			else
377 				goto found;
378 		}
379 
380 		while (addr + size > first->va_start && addr + size <= vend) {
381 			addr = ALIGN(first->va_end + PAGE_SIZE, align);
382 			if (addr + size - 1 < addr)
383 				goto overflow;
384 
385 			n = rb_next(&first->rb_node);
386 			if (n)
387 				first = rb_entry(n, struct vmap_area, rb_node);
388 			else
389 				goto found;
390 		}
391 	}
392 found:
393 	if (addr + size > vend) {
394 overflow:
395 		spin_unlock(&vmap_area_lock);
396 		if (!purged) {
397 			purge_vmap_area_lazy();
398 			purged = 1;
399 			goto retry;
400 		}
401 		if (printk_ratelimit())
402 			printk(KERN_WARNING
403 				"vmap allocation for size %lu failed: "
404 				"use vmalloc=<size> to increase size.\n", size);
405 		kfree(va);
406 		return ERR_PTR(-EBUSY);
407 	}
408 
409 	BUG_ON(addr & (align-1));
410 
411 	va->va_start = addr;
412 	va->va_end = addr + size;
413 	va->flags = 0;
414 	__insert_vmap_area(va);
415 	spin_unlock(&vmap_area_lock);
416 
417 	return va;
418 }
419 
420 static void rcu_free_va(struct rcu_head *head)
421 {
422 	struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
423 
424 	kfree(va);
425 }
426 
427 static void __free_vmap_area(struct vmap_area *va)
428 {
429 	BUG_ON(RB_EMPTY_NODE(&va->rb_node));
430 	rb_erase(&va->rb_node, &vmap_area_root);
431 	RB_CLEAR_NODE(&va->rb_node);
432 	list_del_rcu(&va->list);
433 
434 	/*
435 	 * Track the highest possible candidate for pcpu area
436 	 * allocation.  Areas outside of vmalloc area can be returned
437 	 * here too, consider only end addresses which fall inside
438 	 * vmalloc area proper.
439 	 */
440 	if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
441 		vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
442 
443 	call_rcu(&va->rcu_head, rcu_free_va);
444 }
445 
446 /*
447  * Free a region of KVA allocated by alloc_vmap_area
448  */
449 static void free_vmap_area(struct vmap_area *va)
450 {
451 	spin_lock(&vmap_area_lock);
452 	__free_vmap_area(va);
453 	spin_unlock(&vmap_area_lock);
454 }
455 
456 /*
457  * Clear the pagetable entries of a given vmap_area
458  */
459 static void unmap_vmap_area(struct vmap_area *va)
460 {
461 	vunmap_page_range(va->va_start, va->va_end);
462 }
463 
464 static void vmap_debug_free_range(unsigned long start, unsigned long end)
465 {
466 	/*
467 	 * Unmap page tables and force a TLB flush immediately if
468 	 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
469 	 * bugs similarly to those in linear kernel virtual address
470 	 * space after a page has been freed.
471 	 *
472 	 * All the lazy freeing logic is still retained, in order to
473 	 * minimise intrusiveness of this debugging feature.
474 	 *
475 	 * This is going to be *slow* (linear kernel virtual address
476 	 * debugging doesn't do a broadcast TLB flush so it is a lot
477 	 * faster).
478 	 */
479 #ifdef CONFIG_DEBUG_PAGEALLOC
480 	vunmap_page_range(start, end);
481 	flush_tlb_kernel_range(start, end);
482 #endif
483 }
484 
485 /*
486  * lazy_max_pages is the maximum amount of virtual address space we gather up
487  * before attempting to purge with a TLB flush.
488  *
489  * There is a tradeoff here: a larger number will cover more kernel page tables
490  * and take slightly longer to purge, but it will linearly reduce the number of
491  * global TLB flushes that must be performed. It would seem natural to scale
492  * this number up linearly with the number of CPUs (because vmapping activity
493  * could also scale linearly with the number of CPUs), however it is likely
494  * that in practice, workloads might be constrained in other ways that mean
495  * vmap activity will not scale linearly with CPUs. Also, I want to be
496  * conservative and not introduce a big latency on huge systems, so go with
497  * a less aggressive log scale. It will still be an improvement over the old
498  * code, and it will be simple to change the scale factor if we find that it
499  * becomes a problem on bigger systems.
500  */
501 static unsigned long lazy_max_pages(void)
502 {
503 	unsigned int log;
504 
505 	log = fls(num_online_cpus());
506 
507 	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
508 }
509 
510 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
511 
512 /*
513  * Purges all lazily-freed vmap areas.
514  *
515  * If sync is 0 then don't purge if there is already a purge in progress.
516  * If force_flush is 1, then flush kernel TLBs between *start and *end even
517  * if we found no lazy vmap areas to unmap (callers can use this to optimise
518  * their own TLB flushing).
519  * Returns with *start = min(*start, lowest purged address)
520  *              *end = max(*end, highest purged address)
521  */
522 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
523 					int sync, int force_flush)
524 {
525 	static DEFINE_SPINLOCK(purge_lock);
526 	LIST_HEAD(valist);
527 	struct vmap_area *va;
528 	struct vmap_area *n_va;
529 	int nr = 0;
530 
531 	/*
532 	 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
533 	 * should not expect such behaviour. This just simplifies locking for
534 	 * the case that isn't actually used at the moment anyway.
535 	 */
536 	if (!sync && !force_flush) {
537 		if (!spin_trylock(&purge_lock))
538 			return;
539 	} else
540 		spin_lock(&purge_lock);
541 
542 	rcu_read_lock();
543 	list_for_each_entry_rcu(va, &vmap_area_list, list) {
544 		if (va->flags & VM_LAZY_FREE) {
545 			if (va->va_start < *start)
546 				*start = va->va_start;
547 			if (va->va_end > *end)
548 				*end = va->va_end;
549 			nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
550 			unmap_vmap_area(va);
551 			list_add_tail(&va->purge_list, &valist);
552 			va->flags |= VM_LAZY_FREEING;
553 			va->flags &= ~VM_LAZY_FREE;
554 		}
555 	}
556 	rcu_read_unlock();
557 
558 	if (nr) {
559 		BUG_ON(nr > atomic_read(&vmap_lazy_nr));
560 		atomic_sub(nr, &vmap_lazy_nr);
561 	}
562 
563 	if (nr || force_flush)
564 		flush_tlb_kernel_range(*start, *end);
565 
566 	if (nr) {
567 		spin_lock(&vmap_area_lock);
568 		list_for_each_entry_safe(va, n_va, &valist, purge_list)
569 			__free_vmap_area(va);
570 		spin_unlock(&vmap_area_lock);
571 	}
572 	spin_unlock(&purge_lock);
573 }
574 
575 /*
576  * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
577  * is already purging.
578  */
579 static void try_purge_vmap_area_lazy(void)
580 {
581 	unsigned long start = ULONG_MAX, end = 0;
582 
583 	__purge_vmap_area_lazy(&start, &end, 0, 0);
584 }
585 
586 /*
587  * Kick off a purge of the outstanding lazy areas.
588  */
589 static void purge_vmap_area_lazy(void)
590 {
591 	unsigned long start = ULONG_MAX, end = 0;
592 
593 	__purge_vmap_area_lazy(&start, &end, 1, 0);
594 }
595 
596 /*
597  * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
598  * called for the correct range previously.
599  */
600 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
601 {
602 	va->flags |= VM_LAZY_FREE;
603 	atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
604 	if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
605 		try_purge_vmap_area_lazy();
606 }
607 
608 /*
609  * Free and unmap a vmap area
610  */
611 static void free_unmap_vmap_area(struct vmap_area *va)
612 {
613 	flush_cache_vunmap(va->va_start, va->va_end);
614 	free_unmap_vmap_area_noflush(va);
615 }
616 
617 static struct vmap_area *find_vmap_area(unsigned long addr)
618 {
619 	struct vmap_area *va;
620 
621 	spin_lock(&vmap_area_lock);
622 	va = __find_vmap_area(addr);
623 	spin_unlock(&vmap_area_lock);
624 
625 	return va;
626 }
627 
628 static void free_unmap_vmap_area_addr(unsigned long addr)
629 {
630 	struct vmap_area *va;
631 
632 	va = find_vmap_area(addr);
633 	BUG_ON(!va);
634 	free_unmap_vmap_area(va);
635 }
636 
637 
638 /*** Per cpu kva allocator ***/
639 
640 /*
641  * vmap space is limited especially on 32 bit architectures. Ensure there is
642  * room for at least 16 percpu vmap blocks per CPU.
643  */
644 /*
645  * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
646  * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
647  * instead (we just need a rough idea)
648  */
649 #if BITS_PER_LONG == 32
650 #define VMALLOC_SPACE		(128UL*1024*1024)
651 #else
652 #define VMALLOC_SPACE		(128UL*1024*1024*1024)
653 #endif
654 
655 #define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
656 #define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
657 #define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
658 #define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
659 #define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
660 #define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
661 #define VMAP_BBMAP_BITS		VMAP_MIN(VMAP_BBMAP_BITS_MAX,		\
662 					VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
663 						VMALLOC_PAGES / NR_CPUS / 16))
664 
665 #define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
666 
667 static bool vmap_initialized __read_mostly = false;
668 
669 struct vmap_block_queue {
670 	spinlock_t lock;
671 	struct list_head free;
672 	struct list_head dirty;
673 	unsigned int nr_dirty;
674 };
675 
676 struct vmap_block {
677 	spinlock_t lock;
678 	struct vmap_area *va;
679 	struct vmap_block_queue *vbq;
680 	unsigned long free, dirty;
681 	DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
682 	DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
683 	union {
684 		struct list_head free_list;
685 		struct rcu_head rcu_head;
686 	};
687 };
688 
689 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
690 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
691 
692 /*
693  * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
694  * in the free path. Could get rid of this if we change the API to return a
695  * "cookie" from alloc, to be passed to free. But no big deal yet.
696  */
697 static DEFINE_SPINLOCK(vmap_block_tree_lock);
698 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
699 
700 /*
701  * We should probably have a fallback mechanism to allocate virtual memory
702  * out of partially filled vmap blocks. However vmap block sizing should be
703  * fairly reasonable according to the vmalloc size, so it shouldn't be a
704  * big problem.
705  */
706 
707 static unsigned long addr_to_vb_idx(unsigned long addr)
708 {
709 	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
710 	addr /= VMAP_BLOCK_SIZE;
711 	return addr;
712 }
713 
714 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
715 {
716 	struct vmap_block_queue *vbq;
717 	struct vmap_block *vb;
718 	struct vmap_area *va;
719 	unsigned long vb_idx;
720 	int node, err;
721 
722 	node = numa_node_id();
723 
724 	vb = kmalloc_node(sizeof(struct vmap_block),
725 			gfp_mask & GFP_RECLAIM_MASK, node);
726 	if (unlikely(!vb))
727 		return ERR_PTR(-ENOMEM);
728 
729 	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
730 					VMALLOC_START, VMALLOC_END,
731 					node, gfp_mask);
732 	if (unlikely(IS_ERR(va))) {
733 		kfree(vb);
734 		return ERR_PTR(PTR_ERR(va));
735 	}
736 
737 	err = radix_tree_preload(gfp_mask);
738 	if (unlikely(err)) {
739 		kfree(vb);
740 		free_vmap_area(va);
741 		return ERR_PTR(err);
742 	}
743 
744 	spin_lock_init(&vb->lock);
745 	vb->va = va;
746 	vb->free = VMAP_BBMAP_BITS;
747 	vb->dirty = 0;
748 	bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
749 	bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
750 	INIT_LIST_HEAD(&vb->free_list);
751 
752 	vb_idx = addr_to_vb_idx(va->va_start);
753 	spin_lock(&vmap_block_tree_lock);
754 	err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
755 	spin_unlock(&vmap_block_tree_lock);
756 	BUG_ON(err);
757 	radix_tree_preload_end();
758 
759 	vbq = &get_cpu_var(vmap_block_queue);
760 	vb->vbq = vbq;
761 	spin_lock(&vbq->lock);
762 	list_add(&vb->free_list, &vbq->free);
763 	spin_unlock(&vbq->lock);
764 	put_cpu_var(vmap_block_queue);
765 
766 	return vb;
767 }
768 
769 static void rcu_free_vb(struct rcu_head *head)
770 {
771 	struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
772 
773 	kfree(vb);
774 }
775 
776 static void free_vmap_block(struct vmap_block *vb)
777 {
778 	struct vmap_block *tmp;
779 	unsigned long vb_idx;
780 
781 	BUG_ON(!list_empty(&vb->free_list));
782 
783 	vb_idx = addr_to_vb_idx(vb->va->va_start);
784 	spin_lock(&vmap_block_tree_lock);
785 	tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
786 	spin_unlock(&vmap_block_tree_lock);
787 	BUG_ON(tmp != vb);
788 
789 	free_unmap_vmap_area_noflush(vb->va);
790 	call_rcu(&vb->rcu_head, rcu_free_vb);
791 }
792 
793 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
794 {
795 	struct vmap_block_queue *vbq;
796 	struct vmap_block *vb;
797 	unsigned long addr = 0;
798 	unsigned int order;
799 
800 	BUG_ON(size & ~PAGE_MASK);
801 	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
802 	order = get_order(size);
803 
804 again:
805 	rcu_read_lock();
806 	vbq = &get_cpu_var(vmap_block_queue);
807 	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
808 		int i;
809 
810 		spin_lock(&vb->lock);
811 		i = bitmap_find_free_region(vb->alloc_map,
812 						VMAP_BBMAP_BITS, order);
813 
814 		if (i >= 0) {
815 			addr = vb->va->va_start + (i << PAGE_SHIFT);
816 			BUG_ON(addr_to_vb_idx(addr) !=
817 					addr_to_vb_idx(vb->va->va_start));
818 			vb->free -= 1UL << order;
819 			if (vb->free == 0) {
820 				spin_lock(&vbq->lock);
821 				list_del_init(&vb->free_list);
822 				spin_unlock(&vbq->lock);
823 			}
824 			spin_unlock(&vb->lock);
825 			break;
826 		}
827 		spin_unlock(&vb->lock);
828 	}
829 	put_cpu_var(vmap_block_queue);
830 	rcu_read_unlock();
831 
832 	if (!addr) {
833 		vb = new_vmap_block(gfp_mask);
834 		if (IS_ERR(vb))
835 			return vb;
836 		goto again;
837 	}
838 
839 	return (void *)addr;
840 }
841 
842 static void vb_free(const void *addr, unsigned long size)
843 {
844 	unsigned long offset;
845 	unsigned long vb_idx;
846 	unsigned int order;
847 	struct vmap_block *vb;
848 
849 	BUG_ON(size & ~PAGE_MASK);
850 	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
851 
852 	flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
853 
854 	order = get_order(size);
855 
856 	offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
857 
858 	vb_idx = addr_to_vb_idx((unsigned long)addr);
859 	rcu_read_lock();
860 	vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
861 	rcu_read_unlock();
862 	BUG_ON(!vb);
863 
864 	spin_lock(&vb->lock);
865 	bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
866 
867 	vb->dirty += 1UL << order;
868 	if (vb->dirty == VMAP_BBMAP_BITS) {
869 		BUG_ON(vb->free || !list_empty(&vb->free_list));
870 		spin_unlock(&vb->lock);
871 		free_vmap_block(vb);
872 	} else
873 		spin_unlock(&vb->lock);
874 }
875 
876 /**
877  * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
878  *
879  * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
880  * to amortize TLB flushing overheads. What this means is that any page you
881  * have now, may, in a former life, have been mapped into kernel virtual
882  * address by the vmap layer and so there might be some CPUs with TLB entries
883  * still referencing that page (additional to the regular 1:1 kernel mapping).
884  *
885  * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
886  * be sure that none of the pages we have control over will have any aliases
887  * from the vmap layer.
888  */
889 void vm_unmap_aliases(void)
890 {
891 	unsigned long start = ULONG_MAX, end = 0;
892 	int cpu;
893 	int flush = 0;
894 
895 	if (unlikely(!vmap_initialized))
896 		return;
897 
898 	for_each_possible_cpu(cpu) {
899 		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
900 		struct vmap_block *vb;
901 
902 		rcu_read_lock();
903 		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
904 			int i;
905 
906 			spin_lock(&vb->lock);
907 			i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
908 			while (i < VMAP_BBMAP_BITS) {
909 				unsigned long s, e;
910 				int j;
911 				j = find_next_zero_bit(vb->dirty_map,
912 					VMAP_BBMAP_BITS, i);
913 
914 				s = vb->va->va_start + (i << PAGE_SHIFT);
915 				e = vb->va->va_start + (j << PAGE_SHIFT);
916 				vunmap_page_range(s, e);
917 				flush = 1;
918 
919 				if (s < start)
920 					start = s;
921 				if (e > end)
922 					end = e;
923 
924 				i = j;
925 				i = find_next_bit(vb->dirty_map,
926 							VMAP_BBMAP_BITS, i);
927 			}
928 			spin_unlock(&vb->lock);
929 		}
930 		rcu_read_unlock();
931 	}
932 
933 	__purge_vmap_area_lazy(&start, &end, 1, flush);
934 }
935 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
936 
937 /**
938  * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
939  * @mem: the pointer returned by vm_map_ram
940  * @count: the count passed to that vm_map_ram call (cannot unmap partial)
941  */
942 void vm_unmap_ram(const void *mem, unsigned int count)
943 {
944 	unsigned long size = count << PAGE_SHIFT;
945 	unsigned long addr = (unsigned long)mem;
946 
947 	BUG_ON(!addr);
948 	BUG_ON(addr < VMALLOC_START);
949 	BUG_ON(addr > VMALLOC_END);
950 	BUG_ON(addr & (PAGE_SIZE-1));
951 
952 	debug_check_no_locks_freed(mem, size);
953 	vmap_debug_free_range(addr, addr+size);
954 
955 	if (likely(count <= VMAP_MAX_ALLOC))
956 		vb_free(mem, size);
957 	else
958 		free_unmap_vmap_area_addr(addr);
959 }
960 EXPORT_SYMBOL(vm_unmap_ram);
961 
962 /**
963  * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
964  * @pages: an array of pointers to the pages to be mapped
965  * @count: number of pages
966  * @node: prefer to allocate data structures on this node
967  * @prot: memory protection to use. PAGE_KERNEL for regular RAM
968  *
969  * Returns: a pointer to the address that has been mapped, or %NULL on failure
970  */
971 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
972 {
973 	unsigned long size = count << PAGE_SHIFT;
974 	unsigned long addr;
975 	void *mem;
976 
977 	if (likely(count <= VMAP_MAX_ALLOC)) {
978 		mem = vb_alloc(size, GFP_KERNEL);
979 		if (IS_ERR(mem))
980 			return NULL;
981 		addr = (unsigned long)mem;
982 	} else {
983 		struct vmap_area *va;
984 		va = alloc_vmap_area(size, PAGE_SIZE,
985 				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
986 		if (IS_ERR(va))
987 			return NULL;
988 
989 		addr = va->va_start;
990 		mem = (void *)addr;
991 	}
992 	if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
993 		vm_unmap_ram(mem, count);
994 		return NULL;
995 	}
996 	return mem;
997 }
998 EXPORT_SYMBOL(vm_map_ram);
999 
1000 /**
1001  * vm_area_register_early - register vmap area early during boot
1002  * @vm: vm_struct to register
1003  * @align: requested alignment
1004  *
1005  * This function is used to register kernel vm area before
1006  * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1007  * proper values on entry and other fields should be zero.  On return,
1008  * vm->addr contains the allocated address.
1009  *
1010  * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1011  */
1012 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1013 {
1014 	static size_t vm_init_off __initdata;
1015 	unsigned long addr;
1016 
1017 	addr = ALIGN(VMALLOC_START + vm_init_off, align);
1018 	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1019 
1020 	vm->addr = (void *)addr;
1021 
1022 	vm->next = vmlist;
1023 	vmlist = vm;
1024 }
1025 
1026 void __init vmalloc_init(void)
1027 {
1028 	struct vmap_area *va;
1029 	struct vm_struct *tmp;
1030 	int i;
1031 
1032 	for_each_possible_cpu(i) {
1033 		struct vmap_block_queue *vbq;
1034 
1035 		vbq = &per_cpu(vmap_block_queue, i);
1036 		spin_lock_init(&vbq->lock);
1037 		INIT_LIST_HEAD(&vbq->free);
1038 		INIT_LIST_HEAD(&vbq->dirty);
1039 		vbq->nr_dirty = 0;
1040 	}
1041 
1042 	/* Import existing vmlist entries. */
1043 	for (tmp = vmlist; tmp; tmp = tmp->next) {
1044 		va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1045 		va->flags = tmp->flags | VM_VM_AREA;
1046 		va->va_start = (unsigned long)tmp->addr;
1047 		va->va_end = va->va_start + tmp->size;
1048 		__insert_vmap_area(va);
1049 	}
1050 
1051 	vmap_area_pcpu_hole = VMALLOC_END;
1052 
1053 	vmap_initialized = true;
1054 }
1055 
1056 /**
1057  * map_kernel_range_noflush - map kernel VM area with the specified pages
1058  * @addr: start of the VM area to map
1059  * @size: size of the VM area to map
1060  * @prot: page protection flags to use
1061  * @pages: pages to map
1062  *
1063  * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1064  * specify should have been allocated using get_vm_area() and its
1065  * friends.
1066  *
1067  * NOTE:
1068  * This function does NOT do any cache flushing.  The caller is
1069  * responsible for calling flush_cache_vmap() on to-be-mapped areas
1070  * before calling this function.
1071  *
1072  * RETURNS:
1073  * The number of pages mapped on success, -errno on failure.
1074  */
1075 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1076 			     pgprot_t prot, struct page **pages)
1077 {
1078 	return vmap_page_range_noflush(addr, addr + size, prot, pages);
1079 }
1080 
1081 /**
1082  * unmap_kernel_range_noflush - unmap kernel VM area
1083  * @addr: start of the VM area to unmap
1084  * @size: size of the VM area to unmap
1085  *
1086  * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1087  * specify should have been allocated using get_vm_area() and its
1088  * friends.
1089  *
1090  * NOTE:
1091  * This function does NOT do any cache flushing.  The caller is
1092  * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1093  * before calling this function and flush_tlb_kernel_range() after.
1094  */
1095 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1096 {
1097 	vunmap_page_range(addr, addr + size);
1098 }
1099 
1100 /**
1101  * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1102  * @addr: start of the VM area to unmap
1103  * @size: size of the VM area to unmap
1104  *
1105  * Similar to unmap_kernel_range_noflush() but flushes vcache before
1106  * the unmapping and tlb after.
1107  */
1108 void unmap_kernel_range(unsigned long addr, unsigned long size)
1109 {
1110 	unsigned long end = addr + size;
1111 
1112 	flush_cache_vunmap(addr, end);
1113 	vunmap_page_range(addr, end);
1114 	flush_tlb_kernel_range(addr, end);
1115 }
1116 
1117 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1118 {
1119 	unsigned long addr = (unsigned long)area->addr;
1120 	unsigned long end = addr + area->size - PAGE_SIZE;
1121 	int err;
1122 
1123 	err = vmap_page_range(addr, end, prot, *pages);
1124 	if (err > 0) {
1125 		*pages += err;
1126 		err = 0;
1127 	}
1128 
1129 	return err;
1130 }
1131 EXPORT_SYMBOL_GPL(map_vm_area);
1132 
1133 /*** Old vmalloc interfaces ***/
1134 DEFINE_RWLOCK(vmlist_lock);
1135 struct vm_struct *vmlist;
1136 
1137 static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1138 			      unsigned long flags, void *caller)
1139 {
1140 	struct vm_struct *tmp, **p;
1141 
1142 	vm->flags = flags;
1143 	vm->addr = (void *)va->va_start;
1144 	vm->size = va->va_end - va->va_start;
1145 	vm->caller = caller;
1146 	va->private = vm;
1147 	va->flags |= VM_VM_AREA;
1148 
1149 	write_lock(&vmlist_lock);
1150 	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1151 		if (tmp->addr >= vm->addr)
1152 			break;
1153 	}
1154 	vm->next = *p;
1155 	*p = vm;
1156 	write_unlock(&vmlist_lock);
1157 }
1158 
1159 static struct vm_struct *__get_vm_area_node(unsigned long size,
1160 		unsigned long align, unsigned long flags, unsigned long start,
1161 		unsigned long end, int node, gfp_t gfp_mask, void *caller)
1162 {
1163 	static struct vmap_area *va;
1164 	struct vm_struct *area;
1165 
1166 	BUG_ON(in_interrupt());
1167 	if (flags & VM_IOREMAP) {
1168 		int bit = fls(size);
1169 
1170 		if (bit > IOREMAP_MAX_ORDER)
1171 			bit = IOREMAP_MAX_ORDER;
1172 		else if (bit < PAGE_SHIFT)
1173 			bit = PAGE_SHIFT;
1174 
1175 		align = 1ul << bit;
1176 	}
1177 
1178 	size = PAGE_ALIGN(size);
1179 	if (unlikely(!size))
1180 		return NULL;
1181 
1182 	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1183 	if (unlikely(!area))
1184 		return NULL;
1185 
1186 	/*
1187 	 * We always allocate a guard page.
1188 	 */
1189 	size += PAGE_SIZE;
1190 
1191 	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1192 	if (IS_ERR(va)) {
1193 		kfree(area);
1194 		return NULL;
1195 	}
1196 
1197 	insert_vmalloc_vm(area, va, flags, caller);
1198 	return area;
1199 }
1200 
1201 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1202 				unsigned long start, unsigned long end)
1203 {
1204 	return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1205 						__builtin_return_address(0));
1206 }
1207 EXPORT_SYMBOL_GPL(__get_vm_area);
1208 
1209 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1210 				       unsigned long start, unsigned long end,
1211 				       void *caller)
1212 {
1213 	return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1214 				  caller);
1215 }
1216 
1217 /**
1218  *	get_vm_area  -  reserve a contiguous kernel virtual area
1219  *	@size:		size of the area
1220  *	@flags:		%VM_IOREMAP for I/O mappings or VM_ALLOC
1221  *
1222  *	Search an area of @size in the kernel virtual mapping area,
1223  *	and reserved it for out purposes.  Returns the area descriptor
1224  *	on success or %NULL on failure.
1225  */
1226 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1227 {
1228 	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1229 				-1, GFP_KERNEL, __builtin_return_address(0));
1230 }
1231 
1232 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1233 				void *caller)
1234 {
1235 	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1236 						-1, GFP_KERNEL, caller);
1237 }
1238 
1239 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1240 				   int node, gfp_t gfp_mask)
1241 {
1242 	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1243 				  node, gfp_mask, __builtin_return_address(0));
1244 }
1245 
1246 static struct vm_struct *find_vm_area(const void *addr)
1247 {
1248 	struct vmap_area *va;
1249 
1250 	va = find_vmap_area((unsigned long)addr);
1251 	if (va && va->flags & VM_VM_AREA)
1252 		return va->private;
1253 
1254 	return NULL;
1255 }
1256 
1257 /**
1258  *	remove_vm_area  -  find and remove a continuous kernel virtual area
1259  *	@addr:		base address
1260  *
1261  *	Search for the kernel VM area starting at @addr, and remove it.
1262  *	This function returns the found VM area, but using it is NOT safe
1263  *	on SMP machines, except for its size or flags.
1264  */
1265 struct vm_struct *remove_vm_area(const void *addr)
1266 {
1267 	struct vmap_area *va;
1268 
1269 	va = find_vmap_area((unsigned long)addr);
1270 	if (va && va->flags & VM_VM_AREA) {
1271 		struct vm_struct *vm = va->private;
1272 		struct vm_struct *tmp, **p;
1273 		/*
1274 		 * remove from list and disallow access to this vm_struct
1275 		 * before unmap. (address range confliction is maintained by
1276 		 * vmap.)
1277 		 */
1278 		write_lock(&vmlist_lock);
1279 		for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1280 			;
1281 		*p = tmp->next;
1282 		write_unlock(&vmlist_lock);
1283 
1284 		vmap_debug_free_range(va->va_start, va->va_end);
1285 		free_unmap_vmap_area(va);
1286 		vm->size -= PAGE_SIZE;
1287 
1288 		return vm;
1289 	}
1290 	return NULL;
1291 }
1292 
1293 static void __vunmap(const void *addr, int deallocate_pages)
1294 {
1295 	struct vm_struct *area;
1296 
1297 	if (!addr)
1298 		return;
1299 
1300 	if ((PAGE_SIZE-1) & (unsigned long)addr) {
1301 		WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1302 		return;
1303 	}
1304 
1305 	area = remove_vm_area(addr);
1306 	if (unlikely(!area)) {
1307 		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1308 				addr);
1309 		return;
1310 	}
1311 
1312 	debug_check_no_locks_freed(addr, area->size);
1313 	debug_check_no_obj_freed(addr, area->size);
1314 
1315 	if (deallocate_pages) {
1316 		int i;
1317 
1318 		for (i = 0; i < area->nr_pages; i++) {
1319 			struct page *page = area->pages[i];
1320 
1321 			BUG_ON(!page);
1322 			__free_page(page);
1323 		}
1324 
1325 		if (area->flags & VM_VPAGES)
1326 			vfree(area->pages);
1327 		else
1328 			kfree(area->pages);
1329 	}
1330 
1331 	kfree(area);
1332 	return;
1333 }
1334 
1335 /**
1336  *	vfree  -  release memory allocated by vmalloc()
1337  *	@addr:		memory base address
1338  *
1339  *	Free the virtually continuous memory area starting at @addr, as
1340  *	obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1341  *	NULL, no operation is performed.
1342  *
1343  *	Must not be called in interrupt context.
1344  */
1345 void vfree(const void *addr)
1346 {
1347 	BUG_ON(in_interrupt());
1348 
1349 	kmemleak_free(addr);
1350 
1351 	__vunmap(addr, 1);
1352 }
1353 EXPORT_SYMBOL(vfree);
1354 
1355 /**
1356  *	vunmap  -  release virtual mapping obtained by vmap()
1357  *	@addr:		memory base address
1358  *
1359  *	Free the virtually contiguous memory area starting at @addr,
1360  *	which was created from the page array passed to vmap().
1361  *
1362  *	Must not be called in interrupt context.
1363  */
1364 void vunmap(const void *addr)
1365 {
1366 	BUG_ON(in_interrupt());
1367 	might_sleep();
1368 	__vunmap(addr, 0);
1369 }
1370 EXPORT_SYMBOL(vunmap);
1371 
1372 /**
1373  *	vmap  -  map an array of pages into virtually contiguous space
1374  *	@pages:		array of page pointers
1375  *	@count:		number of pages to map
1376  *	@flags:		vm_area->flags
1377  *	@prot:		page protection for the mapping
1378  *
1379  *	Maps @count pages from @pages into contiguous kernel virtual
1380  *	space.
1381  */
1382 void *vmap(struct page **pages, unsigned int count,
1383 		unsigned long flags, pgprot_t prot)
1384 {
1385 	struct vm_struct *area;
1386 
1387 	might_sleep();
1388 
1389 	if (count > totalram_pages)
1390 		return NULL;
1391 
1392 	area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1393 					__builtin_return_address(0));
1394 	if (!area)
1395 		return NULL;
1396 
1397 	if (map_vm_area(area, prot, &pages)) {
1398 		vunmap(area->addr);
1399 		return NULL;
1400 	}
1401 
1402 	return area->addr;
1403 }
1404 EXPORT_SYMBOL(vmap);
1405 
1406 static void *__vmalloc_node(unsigned long size, unsigned long align,
1407 			    gfp_t gfp_mask, pgprot_t prot,
1408 			    int node, void *caller);
1409 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1410 				 pgprot_t prot, int node, void *caller)
1411 {
1412 	struct page **pages;
1413 	unsigned int nr_pages, array_size, i;
1414 	gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1415 
1416 	nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1417 	array_size = (nr_pages * sizeof(struct page *));
1418 
1419 	area->nr_pages = nr_pages;
1420 	/* Please note that the recursion is strictly bounded. */
1421 	if (array_size > PAGE_SIZE) {
1422 		pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1423 				PAGE_KERNEL, node, caller);
1424 		area->flags |= VM_VPAGES;
1425 	} else {
1426 		pages = kmalloc_node(array_size, nested_gfp, node);
1427 	}
1428 	area->pages = pages;
1429 	area->caller = caller;
1430 	if (!area->pages) {
1431 		remove_vm_area(area->addr);
1432 		kfree(area);
1433 		return NULL;
1434 	}
1435 
1436 	for (i = 0; i < area->nr_pages; i++) {
1437 		struct page *page;
1438 
1439 		if (node < 0)
1440 			page = alloc_page(gfp_mask);
1441 		else
1442 			page = alloc_pages_node(node, gfp_mask, 0);
1443 
1444 		if (unlikely(!page)) {
1445 			/* Successfully allocated i pages, free them in __vunmap() */
1446 			area->nr_pages = i;
1447 			goto fail;
1448 		}
1449 		area->pages[i] = page;
1450 	}
1451 
1452 	if (map_vm_area(area, prot, &pages))
1453 		goto fail;
1454 	return area->addr;
1455 
1456 fail:
1457 	vfree(area->addr);
1458 	return NULL;
1459 }
1460 
1461 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1462 {
1463 	void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1464 					 __builtin_return_address(0));
1465 
1466 	/*
1467 	 * A ref_count = 3 is needed because the vm_struct and vmap_area
1468 	 * structures allocated in the __get_vm_area_node() function contain
1469 	 * references to the virtual address of the vmalloc'ed block.
1470 	 */
1471 	kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1472 
1473 	return addr;
1474 }
1475 
1476 /**
1477  *	__vmalloc_node  -  allocate virtually contiguous memory
1478  *	@size:		allocation size
1479  *	@align:		desired alignment
1480  *	@gfp_mask:	flags for the page level allocator
1481  *	@prot:		protection mask for the allocated pages
1482  *	@node:		node to use for allocation or -1
1483  *	@caller:	caller's return address
1484  *
1485  *	Allocate enough pages to cover @size from the page level
1486  *	allocator with @gfp_mask flags.  Map them into contiguous
1487  *	kernel virtual space, using a pagetable protection of @prot.
1488  */
1489 static void *__vmalloc_node(unsigned long size, unsigned long align,
1490 			    gfp_t gfp_mask, pgprot_t prot,
1491 			    int node, void *caller)
1492 {
1493 	struct vm_struct *area;
1494 	void *addr;
1495 	unsigned long real_size = size;
1496 
1497 	size = PAGE_ALIGN(size);
1498 	if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1499 		return NULL;
1500 
1501 	area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
1502 				  VMALLOC_END, node, gfp_mask, caller);
1503 
1504 	if (!area)
1505 		return NULL;
1506 
1507 	addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1508 
1509 	/*
1510 	 * A ref_count = 3 is needed because the vm_struct and vmap_area
1511 	 * structures allocated in the __get_vm_area_node() function contain
1512 	 * references to the virtual address of the vmalloc'ed block.
1513 	 */
1514 	kmemleak_alloc(addr, real_size, 3, gfp_mask);
1515 
1516 	return addr;
1517 }
1518 
1519 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1520 {
1521 	return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1522 				__builtin_return_address(0));
1523 }
1524 EXPORT_SYMBOL(__vmalloc);
1525 
1526 /**
1527  *	vmalloc  -  allocate virtually contiguous memory
1528  *	@size:		allocation size
1529  *	Allocate enough pages to cover @size from the page level
1530  *	allocator and map them into contiguous kernel virtual space.
1531  *
1532  *	For tight control over page level allocator and protection flags
1533  *	use __vmalloc() instead.
1534  */
1535 void *vmalloc(unsigned long size)
1536 {
1537 	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1538 					-1, __builtin_return_address(0));
1539 }
1540 EXPORT_SYMBOL(vmalloc);
1541 
1542 /**
1543  * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1544  * @size: allocation size
1545  *
1546  * The resulting memory area is zeroed so it can be mapped to userspace
1547  * without leaking data.
1548  */
1549 void *vmalloc_user(unsigned long size)
1550 {
1551 	struct vm_struct *area;
1552 	void *ret;
1553 
1554 	ret = __vmalloc_node(size, SHMLBA,
1555 			     GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1556 			     PAGE_KERNEL, -1, __builtin_return_address(0));
1557 	if (ret) {
1558 		area = find_vm_area(ret);
1559 		area->flags |= VM_USERMAP;
1560 	}
1561 	return ret;
1562 }
1563 EXPORT_SYMBOL(vmalloc_user);
1564 
1565 /**
1566  *	vmalloc_node  -  allocate memory on a specific node
1567  *	@size:		allocation size
1568  *	@node:		numa node
1569  *
1570  *	Allocate enough pages to cover @size from the page level
1571  *	allocator and map them into contiguous kernel virtual space.
1572  *
1573  *	For tight control over page level allocator and protection flags
1574  *	use __vmalloc() instead.
1575  */
1576 void *vmalloc_node(unsigned long size, int node)
1577 {
1578 	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1579 					node, __builtin_return_address(0));
1580 }
1581 EXPORT_SYMBOL(vmalloc_node);
1582 
1583 #ifndef PAGE_KERNEL_EXEC
1584 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1585 #endif
1586 
1587 /**
1588  *	vmalloc_exec  -  allocate virtually contiguous, executable memory
1589  *	@size:		allocation size
1590  *
1591  *	Kernel-internal function to allocate enough pages to cover @size
1592  *	the page level allocator and map them into contiguous and
1593  *	executable kernel virtual space.
1594  *
1595  *	For tight control over page level allocator and protection flags
1596  *	use __vmalloc() instead.
1597  */
1598 
1599 void *vmalloc_exec(unsigned long size)
1600 {
1601 	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1602 			      -1, __builtin_return_address(0));
1603 }
1604 
1605 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1606 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1607 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1608 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1609 #else
1610 #define GFP_VMALLOC32 GFP_KERNEL
1611 #endif
1612 
1613 /**
1614  *	vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1615  *	@size:		allocation size
1616  *
1617  *	Allocate enough 32bit PA addressable pages to cover @size from the
1618  *	page level allocator and map them into contiguous kernel virtual space.
1619  */
1620 void *vmalloc_32(unsigned long size)
1621 {
1622 	return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1623 			      -1, __builtin_return_address(0));
1624 }
1625 EXPORT_SYMBOL(vmalloc_32);
1626 
1627 /**
1628  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1629  *	@size:		allocation size
1630  *
1631  * The resulting memory area is 32bit addressable and zeroed so it can be
1632  * mapped to userspace without leaking data.
1633  */
1634 void *vmalloc_32_user(unsigned long size)
1635 {
1636 	struct vm_struct *area;
1637 	void *ret;
1638 
1639 	ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1640 			     -1, __builtin_return_address(0));
1641 	if (ret) {
1642 		area = find_vm_area(ret);
1643 		area->flags |= VM_USERMAP;
1644 	}
1645 	return ret;
1646 }
1647 EXPORT_SYMBOL(vmalloc_32_user);
1648 
1649 /*
1650  * small helper routine , copy contents to buf from addr.
1651  * If the page is not present, fill zero.
1652  */
1653 
1654 static int aligned_vread(char *buf, char *addr, unsigned long count)
1655 {
1656 	struct page *p;
1657 	int copied = 0;
1658 
1659 	while (count) {
1660 		unsigned long offset, length;
1661 
1662 		offset = (unsigned long)addr & ~PAGE_MASK;
1663 		length = PAGE_SIZE - offset;
1664 		if (length > count)
1665 			length = count;
1666 		p = vmalloc_to_page(addr);
1667 		/*
1668 		 * To do safe access to this _mapped_ area, we need
1669 		 * lock. But adding lock here means that we need to add
1670 		 * overhead of vmalloc()/vfree() calles for this _debug_
1671 		 * interface, rarely used. Instead of that, we'll use
1672 		 * kmap() and get small overhead in this access function.
1673 		 */
1674 		if (p) {
1675 			/*
1676 			 * we can expect USER0 is not used (see vread/vwrite's
1677 			 * function description)
1678 			 */
1679 			void *map = kmap_atomic(p, KM_USER0);
1680 			memcpy(buf, map + offset, length);
1681 			kunmap_atomic(map, KM_USER0);
1682 		} else
1683 			memset(buf, 0, length);
1684 
1685 		addr += length;
1686 		buf += length;
1687 		copied += length;
1688 		count -= length;
1689 	}
1690 	return copied;
1691 }
1692 
1693 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1694 {
1695 	struct page *p;
1696 	int copied = 0;
1697 
1698 	while (count) {
1699 		unsigned long offset, length;
1700 
1701 		offset = (unsigned long)addr & ~PAGE_MASK;
1702 		length = PAGE_SIZE - offset;
1703 		if (length > count)
1704 			length = count;
1705 		p = vmalloc_to_page(addr);
1706 		/*
1707 		 * To do safe access to this _mapped_ area, we need
1708 		 * lock. But adding lock here means that we need to add
1709 		 * overhead of vmalloc()/vfree() calles for this _debug_
1710 		 * interface, rarely used. Instead of that, we'll use
1711 		 * kmap() and get small overhead in this access function.
1712 		 */
1713 		if (p) {
1714 			/*
1715 			 * we can expect USER0 is not used (see vread/vwrite's
1716 			 * function description)
1717 			 */
1718 			void *map = kmap_atomic(p, KM_USER0);
1719 			memcpy(map + offset, buf, length);
1720 			kunmap_atomic(map, KM_USER0);
1721 		}
1722 		addr += length;
1723 		buf += length;
1724 		copied += length;
1725 		count -= length;
1726 	}
1727 	return copied;
1728 }
1729 
1730 /**
1731  *	vread() -  read vmalloc area in a safe way.
1732  *	@buf:		buffer for reading data
1733  *	@addr:		vm address.
1734  *	@count:		number of bytes to be read.
1735  *
1736  *	Returns # of bytes which addr and buf should be increased.
1737  *	(same number to @count). Returns 0 if [addr...addr+count) doesn't
1738  *	includes any intersect with alive vmalloc area.
1739  *
1740  *	This function checks that addr is a valid vmalloc'ed area, and
1741  *	copy data from that area to a given buffer. If the given memory range
1742  *	of [addr...addr+count) includes some valid address, data is copied to
1743  *	proper area of @buf. If there are memory holes, they'll be zero-filled.
1744  *	IOREMAP area is treated as memory hole and no copy is done.
1745  *
1746  *	If [addr...addr+count) doesn't includes any intersects with alive
1747  *	vm_struct area, returns 0.
1748  *	@buf should be kernel's buffer. Because	this function uses KM_USER0,
1749  *	the caller should guarantee KM_USER0 is not used.
1750  *
1751  *	Note: In usual ops, vread() is never necessary because the caller
1752  *	should know vmalloc() area is valid and can use memcpy().
1753  *	This is for routines which have to access vmalloc area without
1754  *	any informaion, as /dev/kmem.
1755  *
1756  */
1757 
1758 long vread(char *buf, char *addr, unsigned long count)
1759 {
1760 	struct vm_struct *tmp;
1761 	char *vaddr, *buf_start = buf;
1762 	unsigned long buflen = count;
1763 	unsigned long n;
1764 
1765 	/* Don't allow overflow */
1766 	if ((unsigned long) addr + count < count)
1767 		count = -(unsigned long) addr;
1768 
1769 	read_lock(&vmlist_lock);
1770 	for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1771 		vaddr = (char *) tmp->addr;
1772 		if (addr >= vaddr + tmp->size - PAGE_SIZE)
1773 			continue;
1774 		while (addr < vaddr) {
1775 			if (count == 0)
1776 				goto finished;
1777 			*buf = '\0';
1778 			buf++;
1779 			addr++;
1780 			count--;
1781 		}
1782 		n = vaddr + tmp->size - PAGE_SIZE - addr;
1783 		if (n > count)
1784 			n = count;
1785 		if (!(tmp->flags & VM_IOREMAP))
1786 			aligned_vread(buf, addr, n);
1787 		else /* IOREMAP area is treated as memory hole */
1788 			memset(buf, 0, n);
1789 		buf += n;
1790 		addr += n;
1791 		count -= n;
1792 	}
1793 finished:
1794 	read_unlock(&vmlist_lock);
1795 
1796 	if (buf == buf_start)
1797 		return 0;
1798 	/* zero-fill memory holes */
1799 	if (buf != buf_start + buflen)
1800 		memset(buf, 0, buflen - (buf - buf_start));
1801 
1802 	return buflen;
1803 }
1804 
1805 /**
1806  *	vwrite() -  write vmalloc area in a safe way.
1807  *	@buf:		buffer for source data
1808  *	@addr:		vm address.
1809  *	@count:		number of bytes to be read.
1810  *
1811  *	Returns # of bytes which addr and buf should be incresed.
1812  *	(same number to @count).
1813  *	If [addr...addr+count) doesn't includes any intersect with valid
1814  *	vmalloc area, returns 0.
1815  *
1816  *	This function checks that addr is a valid vmalloc'ed area, and
1817  *	copy data from a buffer to the given addr. If specified range of
1818  *	[addr...addr+count) includes some valid address, data is copied from
1819  *	proper area of @buf. If there are memory holes, no copy to hole.
1820  *	IOREMAP area is treated as memory hole and no copy is done.
1821  *
1822  *	If [addr...addr+count) doesn't includes any intersects with alive
1823  *	vm_struct area, returns 0.
1824  *	@buf should be kernel's buffer. Because	this function uses KM_USER0,
1825  *	the caller should guarantee KM_USER0 is not used.
1826  *
1827  *	Note: In usual ops, vwrite() is never necessary because the caller
1828  *	should know vmalloc() area is valid and can use memcpy().
1829  *	This is for routines which have to access vmalloc area without
1830  *	any informaion, as /dev/kmem.
1831  *
1832  *	The caller should guarantee KM_USER1 is not used.
1833  */
1834 
1835 long vwrite(char *buf, char *addr, unsigned long count)
1836 {
1837 	struct vm_struct *tmp;
1838 	char *vaddr;
1839 	unsigned long n, buflen;
1840 	int copied = 0;
1841 
1842 	/* Don't allow overflow */
1843 	if ((unsigned long) addr + count < count)
1844 		count = -(unsigned long) addr;
1845 	buflen = count;
1846 
1847 	read_lock(&vmlist_lock);
1848 	for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1849 		vaddr = (char *) tmp->addr;
1850 		if (addr >= vaddr + tmp->size - PAGE_SIZE)
1851 			continue;
1852 		while (addr < vaddr) {
1853 			if (count == 0)
1854 				goto finished;
1855 			buf++;
1856 			addr++;
1857 			count--;
1858 		}
1859 		n = vaddr + tmp->size - PAGE_SIZE - addr;
1860 		if (n > count)
1861 			n = count;
1862 		if (!(tmp->flags & VM_IOREMAP)) {
1863 			aligned_vwrite(buf, addr, n);
1864 			copied++;
1865 		}
1866 		buf += n;
1867 		addr += n;
1868 		count -= n;
1869 	}
1870 finished:
1871 	read_unlock(&vmlist_lock);
1872 	if (!copied)
1873 		return 0;
1874 	return buflen;
1875 }
1876 
1877 /**
1878  *	remap_vmalloc_range  -  map vmalloc pages to userspace
1879  *	@vma:		vma to cover (map full range of vma)
1880  *	@addr:		vmalloc memory
1881  *	@pgoff:		number of pages into addr before first page to map
1882  *
1883  *	Returns:	0 for success, -Exxx on failure
1884  *
1885  *	This function checks that addr is a valid vmalloc'ed area, and
1886  *	that it is big enough to cover the vma. Will return failure if
1887  *	that criteria isn't met.
1888  *
1889  *	Similar to remap_pfn_range() (see mm/memory.c)
1890  */
1891 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1892 						unsigned long pgoff)
1893 {
1894 	struct vm_struct *area;
1895 	unsigned long uaddr = vma->vm_start;
1896 	unsigned long usize = vma->vm_end - vma->vm_start;
1897 
1898 	if ((PAGE_SIZE-1) & (unsigned long)addr)
1899 		return -EINVAL;
1900 
1901 	area = find_vm_area(addr);
1902 	if (!area)
1903 		return -EINVAL;
1904 
1905 	if (!(area->flags & VM_USERMAP))
1906 		return -EINVAL;
1907 
1908 	if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1909 		return -EINVAL;
1910 
1911 	addr += pgoff << PAGE_SHIFT;
1912 	do {
1913 		struct page *page = vmalloc_to_page(addr);
1914 		int ret;
1915 
1916 		ret = vm_insert_page(vma, uaddr, page);
1917 		if (ret)
1918 			return ret;
1919 
1920 		uaddr += PAGE_SIZE;
1921 		addr += PAGE_SIZE;
1922 		usize -= PAGE_SIZE;
1923 	} while (usize > 0);
1924 
1925 	/* Prevent "things" like memory migration? VM_flags need a cleanup... */
1926 	vma->vm_flags |= VM_RESERVED;
1927 
1928 	return 0;
1929 }
1930 EXPORT_SYMBOL(remap_vmalloc_range);
1931 
1932 /*
1933  * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1934  * have one.
1935  */
1936 void  __attribute__((weak)) vmalloc_sync_all(void)
1937 {
1938 }
1939 
1940 
1941 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1942 {
1943 	/* apply_to_page_range() does all the hard work. */
1944 	return 0;
1945 }
1946 
1947 /**
1948  *	alloc_vm_area - allocate a range of kernel address space
1949  *	@size:		size of the area
1950  *
1951  *	Returns:	NULL on failure, vm_struct on success
1952  *
1953  *	This function reserves a range of kernel address space, and
1954  *	allocates pagetables to map that range.  No actual mappings
1955  *	are created.  If the kernel address space is not shared
1956  *	between processes, it syncs the pagetable across all
1957  *	processes.
1958  */
1959 struct vm_struct *alloc_vm_area(size_t size)
1960 {
1961 	struct vm_struct *area;
1962 
1963 	area = get_vm_area_caller(size, VM_IOREMAP,
1964 				__builtin_return_address(0));
1965 	if (area == NULL)
1966 		return NULL;
1967 
1968 	/*
1969 	 * This ensures that page tables are constructed for this region
1970 	 * of kernel virtual address space and mapped into init_mm.
1971 	 */
1972 	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1973 				area->size, f, NULL)) {
1974 		free_vm_area(area);
1975 		return NULL;
1976 	}
1977 
1978 	/* Make sure the pagetables are constructed in process kernel
1979 	   mappings */
1980 	vmalloc_sync_all();
1981 
1982 	return area;
1983 }
1984 EXPORT_SYMBOL_GPL(alloc_vm_area);
1985 
1986 void free_vm_area(struct vm_struct *area)
1987 {
1988 	struct vm_struct *ret;
1989 	ret = remove_vm_area(area->addr);
1990 	BUG_ON(ret != area);
1991 	kfree(area);
1992 }
1993 EXPORT_SYMBOL_GPL(free_vm_area);
1994 
1995 static struct vmap_area *node_to_va(struct rb_node *n)
1996 {
1997 	return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
1998 }
1999 
2000 /**
2001  * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2002  * @end: target address
2003  * @pnext: out arg for the next vmap_area
2004  * @pprev: out arg for the previous vmap_area
2005  *
2006  * Returns: %true if either or both of next and prev are found,
2007  *	    %false if no vmap_area exists
2008  *
2009  * Find vmap_areas end addresses of which enclose @end.  ie. if not
2010  * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2011  */
2012 static bool pvm_find_next_prev(unsigned long end,
2013 			       struct vmap_area **pnext,
2014 			       struct vmap_area **pprev)
2015 {
2016 	struct rb_node *n = vmap_area_root.rb_node;
2017 	struct vmap_area *va = NULL;
2018 
2019 	while (n) {
2020 		va = rb_entry(n, struct vmap_area, rb_node);
2021 		if (end < va->va_end)
2022 			n = n->rb_left;
2023 		else if (end > va->va_end)
2024 			n = n->rb_right;
2025 		else
2026 			break;
2027 	}
2028 
2029 	if (!va)
2030 		return false;
2031 
2032 	if (va->va_end > end) {
2033 		*pnext = va;
2034 		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2035 	} else {
2036 		*pprev = va;
2037 		*pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2038 	}
2039 	return true;
2040 }
2041 
2042 /**
2043  * pvm_determine_end - find the highest aligned address between two vmap_areas
2044  * @pnext: in/out arg for the next vmap_area
2045  * @pprev: in/out arg for the previous vmap_area
2046  * @align: alignment
2047  *
2048  * Returns: determined end address
2049  *
2050  * Find the highest aligned address between *@pnext and *@pprev below
2051  * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2052  * down address is between the end addresses of the two vmap_areas.
2053  *
2054  * Please note that the address returned by this function may fall
2055  * inside *@pnext vmap_area.  The caller is responsible for checking
2056  * that.
2057  */
2058 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2059 				       struct vmap_area **pprev,
2060 				       unsigned long align)
2061 {
2062 	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2063 	unsigned long addr;
2064 
2065 	if (*pnext)
2066 		addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2067 	else
2068 		addr = vmalloc_end;
2069 
2070 	while (*pprev && (*pprev)->va_end > addr) {
2071 		*pnext = *pprev;
2072 		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2073 	}
2074 
2075 	return addr;
2076 }
2077 
2078 /**
2079  * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2080  * @offsets: array containing offset of each area
2081  * @sizes: array containing size of each area
2082  * @nr_vms: the number of areas to allocate
2083  * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2084  * @gfp_mask: allocation mask
2085  *
2086  * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2087  *	    vm_structs on success, %NULL on failure
2088  *
2089  * Percpu allocator wants to use congruent vm areas so that it can
2090  * maintain the offsets among percpu areas.  This function allocates
2091  * congruent vmalloc areas for it.  These areas tend to be scattered
2092  * pretty far, distance between two areas easily going up to
2093  * gigabytes.  To avoid interacting with regular vmallocs, these areas
2094  * are allocated from top.
2095  *
2096  * Despite its complicated look, this allocator is rather simple.  It
2097  * does everything top-down and scans areas from the end looking for
2098  * matching slot.  While scanning, if any of the areas overlaps with
2099  * existing vmap_area, the base address is pulled down to fit the
2100  * area.  Scanning is repeated till all the areas fit and then all
2101  * necessary data structres are inserted and the result is returned.
2102  */
2103 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2104 				     const size_t *sizes, int nr_vms,
2105 				     size_t align, gfp_t gfp_mask)
2106 {
2107 	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2108 	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2109 	struct vmap_area **vas, *prev, *next;
2110 	struct vm_struct **vms;
2111 	int area, area2, last_area, term_area;
2112 	unsigned long base, start, end, last_end;
2113 	bool purged = false;
2114 
2115 	gfp_mask &= GFP_RECLAIM_MASK;
2116 
2117 	/* verify parameters and allocate data structures */
2118 	BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2119 	for (last_area = 0, area = 0; area < nr_vms; area++) {
2120 		start = offsets[area];
2121 		end = start + sizes[area];
2122 
2123 		/* is everything aligned properly? */
2124 		BUG_ON(!IS_ALIGNED(offsets[area], align));
2125 		BUG_ON(!IS_ALIGNED(sizes[area], align));
2126 
2127 		/* detect the area with the highest address */
2128 		if (start > offsets[last_area])
2129 			last_area = area;
2130 
2131 		for (area2 = 0; area2 < nr_vms; area2++) {
2132 			unsigned long start2 = offsets[area2];
2133 			unsigned long end2 = start2 + sizes[area2];
2134 
2135 			if (area2 == area)
2136 				continue;
2137 
2138 			BUG_ON(start2 >= start && start2 < end);
2139 			BUG_ON(end2 <= end && end2 > start);
2140 		}
2141 	}
2142 	last_end = offsets[last_area] + sizes[last_area];
2143 
2144 	if (vmalloc_end - vmalloc_start < last_end) {
2145 		WARN_ON(true);
2146 		return NULL;
2147 	}
2148 
2149 	vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2150 	vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2151 	if (!vas || !vms)
2152 		goto err_free;
2153 
2154 	for (area = 0; area < nr_vms; area++) {
2155 		vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2156 		vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2157 		if (!vas[area] || !vms[area])
2158 			goto err_free;
2159 	}
2160 retry:
2161 	spin_lock(&vmap_area_lock);
2162 
2163 	/* start scanning - we scan from the top, begin with the last area */
2164 	area = term_area = last_area;
2165 	start = offsets[area];
2166 	end = start + sizes[area];
2167 
2168 	if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2169 		base = vmalloc_end - last_end;
2170 		goto found;
2171 	}
2172 	base = pvm_determine_end(&next, &prev, align) - end;
2173 
2174 	while (true) {
2175 		BUG_ON(next && next->va_end <= base + end);
2176 		BUG_ON(prev && prev->va_end > base + end);
2177 
2178 		/*
2179 		 * base might have underflowed, add last_end before
2180 		 * comparing.
2181 		 */
2182 		if (base + last_end < vmalloc_start + last_end) {
2183 			spin_unlock(&vmap_area_lock);
2184 			if (!purged) {
2185 				purge_vmap_area_lazy();
2186 				purged = true;
2187 				goto retry;
2188 			}
2189 			goto err_free;
2190 		}
2191 
2192 		/*
2193 		 * If next overlaps, move base downwards so that it's
2194 		 * right below next and then recheck.
2195 		 */
2196 		if (next && next->va_start < base + end) {
2197 			base = pvm_determine_end(&next, &prev, align) - end;
2198 			term_area = area;
2199 			continue;
2200 		}
2201 
2202 		/*
2203 		 * If prev overlaps, shift down next and prev and move
2204 		 * base so that it's right below new next and then
2205 		 * recheck.
2206 		 */
2207 		if (prev && prev->va_end > base + start)  {
2208 			next = prev;
2209 			prev = node_to_va(rb_prev(&next->rb_node));
2210 			base = pvm_determine_end(&next, &prev, align) - end;
2211 			term_area = area;
2212 			continue;
2213 		}
2214 
2215 		/*
2216 		 * This area fits, move on to the previous one.  If
2217 		 * the previous one is the terminal one, we're done.
2218 		 */
2219 		area = (area + nr_vms - 1) % nr_vms;
2220 		if (area == term_area)
2221 			break;
2222 		start = offsets[area];
2223 		end = start + sizes[area];
2224 		pvm_find_next_prev(base + end, &next, &prev);
2225 	}
2226 found:
2227 	/* we've found a fitting base, insert all va's */
2228 	for (area = 0; area < nr_vms; area++) {
2229 		struct vmap_area *va = vas[area];
2230 
2231 		va->va_start = base + offsets[area];
2232 		va->va_end = va->va_start + sizes[area];
2233 		__insert_vmap_area(va);
2234 	}
2235 
2236 	vmap_area_pcpu_hole = base + offsets[last_area];
2237 
2238 	spin_unlock(&vmap_area_lock);
2239 
2240 	/* insert all vm's */
2241 	for (area = 0; area < nr_vms; area++)
2242 		insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2243 				  pcpu_get_vm_areas);
2244 
2245 	kfree(vas);
2246 	return vms;
2247 
2248 err_free:
2249 	for (area = 0; area < nr_vms; area++) {
2250 		if (vas)
2251 			kfree(vas[area]);
2252 		if (vms)
2253 			kfree(vms[area]);
2254 	}
2255 	kfree(vas);
2256 	kfree(vms);
2257 	return NULL;
2258 }
2259 
2260 /**
2261  * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2262  * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2263  * @nr_vms: the number of allocated areas
2264  *
2265  * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2266  */
2267 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2268 {
2269 	int i;
2270 
2271 	for (i = 0; i < nr_vms; i++)
2272 		free_vm_area(vms[i]);
2273 	kfree(vms);
2274 }
2275 
2276 #ifdef CONFIG_PROC_FS
2277 static void *s_start(struct seq_file *m, loff_t *pos)
2278 {
2279 	loff_t n = *pos;
2280 	struct vm_struct *v;
2281 
2282 	read_lock(&vmlist_lock);
2283 	v = vmlist;
2284 	while (n > 0 && v) {
2285 		n--;
2286 		v = v->next;
2287 	}
2288 	if (!n)
2289 		return v;
2290 
2291 	return NULL;
2292 
2293 }
2294 
2295 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2296 {
2297 	struct vm_struct *v = p;
2298 
2299 	++*pos;
2300 	return v->next;
2301 }
2302 
2303 static void s_stop(struct seq_file *m, void *p)
2304 {
2305 	read_unlock(&vmlist_lock);
2306 }
2307 
2308 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2309 {
2310 	if (NUMA_BUILD) {
2311 		unsigned int nr, *counters = m->private;
2312 
2313 		if (!counters)
2314 			return;
2315 
2316 		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2317 
2318 		for (nr = 0; nr < v->nr_pages; nr++)
2319 			counters[page_to_nid(v->pages[nr])]++;
2320 
2321 		for_each_node_state(nr, N_HIGH_MEMORY)
2322 			if (counters[nr])
2323 				seq_printf(m, " N%u=%u", nr, counters[nr]);
2324 	}
2325 }
2326 
2327 static int s_show(struct seq_file *m, void *p)
2328 {
2329 	struct vm_struct *v = p;
2330 
2331 	seq_printf(m, "0x%p-0x%p %7ld",
2332 		v->addr, v->addr + v->size, v->size);
2333 
2334 	if (v->caller) {
2335 		char buff[KSYM_SYMBOL_LEN];
2336 
2337 		seq_putc(m, ' ');
2338 		sprint_symbol(buff, (unsigned long)v->caller);
2339 		seq_puts(m, buff);
2340 	}
2341 
2342 	if (v->nr_pages)
2343 		seq_printf(m, " pages=%d", v->nr_pages);
2344 
2345 	if (v->phys_addr)
2346 		seq_printf(m, " phys=%lx", v->phys_addr);
2347 
2348 	if (v->flags & VM_IOREMAP)
2349 		seq_printf(m, " ioremap");
2350 
2351 	if (v->flags & VM_ALLOC)
2352 		seq_printf(m, " vmalloc");
2353 
2354 	if (v->flags & VM_MAP)
2355 		seq_printf(m, " vmap");
2356 
2357 	if (v->flags & VM_USERMAP)
2358 		seq_printf(m, " user");
2359 
2360 	if (v->flags & VM_VPAGES)
2361 		seq_printf(m, " vpages");
2362 
2363 	show_numa_info(m, v);
2364 	seq_putc(m, '\n');
2365 	return 0;
2366 }
2367 
2368 static const struct seq_operations vmalloc_op = {
2369 	.start = s_start,
2370 	.next = s_next,
2371 	.stop = s_stop,
2372 	.show = s_show,
2373 };
2374 
2375 static int vmalloc_open(struct inode *inode, struct file *file)
2376 {
2377 	unsigned int *ptr = NULL;
2378 	int ret;
2379 
2380 	if (NUMA_BUILD)
2381 		ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2382 	ret = seq_open(file, &vmalloc_op);
2383 	if (!ret) {
2384 		struct seq_file *m = file->private_data;
2385 		m->private = ptr;
2386 	} else
2387 		kfree(ptr);
2388 	return ret;
2389 }
2390 
2391 static const struct file_operations proc_vmalloc_operations = {
2392 	.open		= vmalloc_open,
2393 	.read		= seq_read,
2394 	.llseek		= seq_lseek,
2395 	.release	= seq_release_private,
2396 };
2397 
2398 static int __init proc_vmalloc_init(void)
2399 {
2400 	proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2401 	return 0;
2402 }
2403 module_init(proc_vmalloc_init);
2404 #endif
2405 
2406