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 /* for per-CPU blocks */
513 static void purge_fragmented_blocks_allcpus(void);
514
515 /*
516 * called before a call to iounmap() if the caller wants vm_area_struct's
517 * immediately freed.
518 */
519 void set_iounmap_nonlazy(void)
520 {
521 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
522 }
523
524 /*
525 * Purges all lazily-freed vmap areas.
526 *
527 * If sync is 0 then don't purge if there is already a purge in progress.
528 * If force_flush is 1, then flush kernel TLBs between *start and *end even
529 * if we found no lazy vmap areas to unmap (callers can use this to optimise
530 * their own TLB flushing).
531 * Returns with *start = min(*start, lowest purged address)
532 * *end = max(*end, highest purged address)
533 */
534 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
535 int sync, int force_flush)
536 {
537 static DEFINE_SPINLOCK(purge_lock);
538 LIST_HEAD(valist);
539 struct vmap_area *va;
540 struct vmap_area *n_va;
541 int nr = 0;
542
543 /*
544 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
545 * should not expect such behaviour. This just simplifies locking for
546 * the case that isn't actually used at the moment anyway.
547 */
548 if (!sync && !force_flush) {
549 if (!spin_trylock(&purge_lock))
550 return;
551 } else
552 spin_lock(&purge_lock);
553
554 if (sync)
555 purge_fragmented_blocks_allcpus();
556
557 rcu_read_lock();
558 list_for_each_entry_rcu(va, &vmap_area_list, list) {
559 if (va->flags & VM_LAZY_FREE) {
560 if (va->va_start < *start)
561 *start = va->va_start;
562 if (va->va_end > *end)
563 *end = va->va_end;
564 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
565 unmap_vmap_area(va);
566 list_add_tail(&va->purge_list, &valist);
567 va->flags |= VM_LAZY_FREEING;
568 va->flags &= ~VM_LAZY_FREE;
569 }
570 }
571 rcu_read_unlock();
572
573 if (nr)
574 atomic_sub(nr, &vmap_lazy_nr);
575
576 if (nr || force_flush)
577 flush_tlb_kernel_range(*start, *end);
578
579 if (nr) {
580 spin_lock(&vmap_area_lock);
581 list_for_each_entry_safe(va, n_va, &valist, purge_list)
582 __free_vmap_area(va);
583 spin_unlock(&vmap_area_lock);
584 }
585 spin_unlock(&purge_lock);
586 }
587
588 /*
589 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
590 * is already purging.
591 */
592 static void try_purge_vmap_area_lazy(void)
593 {
594 unsigned long start = ULONG_MAX, end = 0;
595
596 __purge_vmap_area_lazy(&start, &end, 0, 0);
597 }
598
599 /*
600 * Kick off a purge of the outstanding lazy areas.
601 */
602 static void purge_vmap_area_lazy(void)
603 {
604 unsigned long start = ULONG_MAX, end = 0;
605
606 __purge_vmap_area_lazy(&start, &end, 1, 0);
607 }
608
609 /*
610 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
611 * called for the correct range previously.
612 */
613 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
614 {
615 va->flags |= VM_LAZY_FREE;
616 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
617 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
618 try_purge_vmap_area_lazy();
619 }
620
621 /*
622 * Free and unmap a vmap area
623 */
624 static void free_unmap_vmap_area(struct vmap_area *va)
625 {
626 flush_cache_vunmap(va->va_start, va->va_end);
627 free_unmap_vmap_area_noflush(va);
628 }
629
630 static struct vmap_area *find_vmap_area(unsigned long addr)
631 {
632 struct vmap_area *va;
633
634 spin_lock(&vmap_area_lock);
635 va = __find_vmap_area(addr);
636 spin_unlock(&vmap_area_lock);
637
638 return va;
639 }
640
641 static void free_unmap_vmap_area_addr(unsigned long addr)
642 {
643 struct vmap_area *va;
644
645 va = find_vmap_area(addr);
646 BUG_ON(!va);
647 free_unmap_vmap_area(va);
648 }
649
650
651 /*** Per cpu kva allocator ***/
652
653 /*
654 * vmap space is limited especially on 32 bit architectures. Ensure there is
655 * room for at least 16 percpu vmap blocks per CPU.
656 */
657 /*
658 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
659 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
660 * instead (we just need a rough idea)
661 */
662 #if BITS_PER_LONG == 32
663 #define VMALLOC_SPACE (128UL*1024*1024)
664 #else
665 #define VMALLOC_SPACE (128UL*1024*1024*1024)
666 #endif
667
668 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
669 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
670 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
671 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
672 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
673 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
674 #define VMAP_BBMAP_BITS \
675 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
676 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
677 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
678
679 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
680
681 static bool vmap_initialized __read_mostly = false;
682
683 struct vmap_block_queue {
684 spinlock_t lock;
685 struct list_head free;
686 };
687
688 struct vmap_block {
689 spinlock_t lock;
690 struct vmap_area *va;
691 struct vmap_block_queue *vbq;
692 unsigned long free, dirty;
693 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
694 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
695 struct list_head free_list;
696 struct rcu_head rcu_head;
697 struct list_head purge;
698 };
699
700 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
701 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
702
703 /*
704 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
705 * in the free path. Could get rid of this if we change the API to return a
706 * "cookie" from alloc, to be passed to free. But no big deal yet.
707 */
708 static DEFINE_SPINLOCK(vmap_block_tree_lock);
709 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
710
711 /*
712 * We should probably have a fallback mechanism to allocate virtual memory
713 * out of partially filled vmap blocks. However vmap block sizing should be
714 * fairly reasonable according to the vmalloc size, so it shouldn't be a
715 * big problem.
716 */
717
718 static unsigned long addr_to_vb_idx(unsigned long addr)
719 {
720 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
721 addr /= VMAP_BLOCK_SIZE;
722 return addr;
723 }
724
725 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
726 {
727 struct vmap_block_queue *vbq;
728 struct vmap_block *vb;
729 struct vmap_area *va;
730 unsigned long vb_idx;
731 int node, err;
732
733 node = numa_node_id();
734
735 vb = kmalloc_node(sizeof(struct vmap_block),
736 gfp_mask & GFP_RECLAIM_MASK, node);
737 if (unlikely(!vb))
738 return ERR_PTR(-ENOMEM);
739
740 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
741 VMALLOC_START, VMALLOC_END,
742 node, gfp_mask);
743 if (unlikely(IS_ERR(va))) {
744 kfree(vb);
745 return ERR_PTR(PTR_ERR(va));
746 }
747
748 err = radix_tree_preload(gfp_mask);
749 if (unlikely(err)) {
750 kfree(vb);
751 free_vmap_area(va);
752 return ERR_PTR(err);
753 }
754
755 spin_lock_init(&vb->lock);
756 vb->va = va;
757 vb->free = VMAP_BBMAP_BITS;
758 vb->dirty = 0;
759 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
760 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
761 INIT_LIST_HEAD(&vb->free_list);
762
763 vb_idx = addr_to_vb_idx(va->va_start);
764 spin_lock(&vmap_block_tree_lock);
765 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
766 spin_unlock(&vmap_block_tree_lock);
767 BUG_ON(err);
768 radix_tree_preload_end();
769
770 vbq = &get_cpu_var(vmap_block_queue);
771 vb->vbq = vbq;
772 spin_lock(&vbq->lock);
773 list_add_rcu(&vb->free_list, &vbq->free);
774 spin_unlock(&vbq->lock);
775 put_cpu_var(vmap_cpu_blocks);
776
777 return vb;
778 }
779
780 static void rcu_free_vb(struct rcu_head *head)
781 {
782 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
783
784 kfree(vb);
785 }
786
787 static void free_vmap_block(struct vmap_block *vb)
788 {
789 struct vmap_block *tmp;
790 unsigned long vb_idx;
791
792 vb_idx = addr_to_vb_idx(vb->va->va_start);
793 spin_lock(&vmap_block_tree_lock);
794 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
795 spin_unlock(&vmap_block_tree_lock);
796 BUG_ON(tmp != vb);
797
798 free_unmap_vmap_area_noflush(vb->va);
799 call_rcu(&vb->rcu_head, rcu_free_vb);
800 }
801
802 static void purge_fragmented_blocks(int cpu)
803 {
804 LIST_HEAD(purge);
805 struct vmap_block *vb;
806 struct vmap_block *n_vb;
807 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
808
809 rcu_read_lock();
810 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
811
812 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
813 continue;
814
815 spin_lock(&vb->lock);
816 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
817 vb->free = 0; /* prevent further allocs after releasing lock */
818 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
819 bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
820 bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
821 spin_lock(&vbq->lock);
822 list_del_rcu(&vb->free_list);
823 spin_unlock(&vbq->lock);
824 spin_unlock(&vb->lock);
825 list_add_tail(&vb->purge, &purge);
826 } else
827 spin_unlock(&vb->lock);
828 }
829 rcu_read_unlock();
830
831 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
832 list_del(&vb->purge);
833 free_vmap_block(vb);
834 }
835 }
836
837 static void purge_fragmented_blocks_thiscpu(void)
838 {
839 purge_fragmented_blocks(smp_processor_id());
840 }
841
842 static void purge_fragmented_blocks_allcpus(void)
843 {
844 int cpu;
845
846 for_each_possible_cpu(cpu)
847 purge_fragmented_blocks(cpu);
848 }
849
850 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
851 {
852 struct vmap_block_queue *vbq;
853 struct vmap_block *vb;
854 unsigned long addr = 0;
855 unsigned int order;
856 int purge = 0;
857
858 BUG_ON(size & ~PAGE_MASK);
859 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
860 order = get_order(size);
861
862 again:
863 rcu_read_lock();
864 vbq = &get_cpu_var(vmap_block_queue);
865 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
866 int i;
867
868 spin_lock(&vb->lock);
869 if (vb->free < 1UL << order)
870 goto next;
871 i = bitmap_find_free_region(vb->alloc_map,
872 VMAP_BBMAP_BITS, order);
873
874 if (i < 0) {
875 if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
876 /* fragmented and no outstanding allocations */
877 BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
878 purge = 1;
879 }
880 goto next;
881 }
882 addr = vb->va->va_start + (i << PAGE_SHIFT);
883 BUG_ON(addr_to_vb_idx(addr) !=
884 addr_to_vb_idx(vb->va->va_start));
885 vb->free -= 1UL << order;
886 if (vb->free == 0) {
887 spin_lock(&vbq->lock);
888 list_del_rcu(&vb->free_list);
889 spin_unlock(&vbq->lock);
890 }
891 spin_unlock(&vb->lock);
892 break;
893 next:
894 spin_unlock(&vb->lock);
895 }
896
897 if (purge)
898 purge_fragmented_blocks_thiscpu();
899
900 put_cpu_var(vmap_cpu_blocks);
901 rcu_read_unlock();
902
903 if (!addr) {
904 vb = new_vmap_block(gfp_mask);
905 if (IS_ERR(vb))
906 return vb;
907 goto again;
908 }
909
910 return (void *)addr;
911 }
912
913 static void vb_free(const void *addr, unsigned long size)
914 {
915 unsigned long offset;
916 unsigned long vb_idx;
917 unsigned int order;
918 struct vmap_block *vb;
919
920 BUG_ON(size & ~PAGE_MASK);
921 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
922
923 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
924
925 order = get_order(size);
926
927 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
928
929 vb_idx = addr_to_vb_idx((unsigned long)addr);
930 rcu_read_lock();
931 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
932 rcu_read_unlock();
933 BUG_ON(!vb);
934
935 spin_lock(&vb->lock);
936 BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
937
938 vb->dirty += 1UL << order;
939 if (vb->dirty == VMAP_BBMAP_BITS) {
940 BUG_ON(vb->free);
941 spin_unlock(&vb->lock);
942 free_vmap_block(vb);
943 } else
944 spin_unlock(&vb->lock);
945 }
946
947 /**
948 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
949 *
950 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
951 * to amortize TLB flushing overheads. What this means is that any page you
952 * have now, may, in a former life, have been mapped into kernel virtual
953 * address by the vmap layer and so there might be some CPUs with TLB entries
954 * still referencing that page (additional to the regular 1:1 kernel mapping).
955 *
956 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
957 * be sure that none of the pages we have control over will have any aliases
958 * from the vmap layer.
959 */
960 void vm_unmap_aliases(void)
961 {
962 unsigned long start = ULONG_MAX, end = 0;
963 int cpu;
964 int flush = 0;
965
966 if (unlikely(!vmap_initialized))
967 return;
968
969 for_each_possible_cpu(cpu) {
970 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
971 struct vmap_block *vb;
972
973 rcu_read_lock();
974 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
975 int i;
976
977 spin_lock(&vb->lock);
978 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
979 while (i < VMAP_BBMAP_BITS) {
980 unsigned long s, e;
981 int j;
982 j = find_next_zero_bit(vb->dirty_map,
983 VMAP_BBMAP_BITS, i);
984
985 s = vb->va->va_start + (i << PAGE_SHIFT);
986 e = vb->va->va_start + (j << PAGE_SHIFT);
987 vunmap_page_range(s, e);
988 flush = 1;
989
990 if (s < start)
991 start = s;
992 if (e > end)
993 end = e;
994
995 i = j;
996 i = find_next_bit(vb->dirty_map,
997 VMAP_BBMAP_BITS, i);
998 }
999 spin_unlock(&vb->lock);
1000 }
1001 rcu_read_unlock();
1002 }
1003
1004 __purge_vmap_area_lazy(&start, &end, 1, flush);
1005 }
1006 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1007
1008 /**
1009 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1010 * @mem: the pointer returned by vm_map_ram
1011 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1012 */
1013 void vm_unmap_ram(const void *mem, unsigned int count)
1014 {
1015 unsigned long size = count << PAGE_SHIFT;
1016 unsigned long addr = (unsigned long)mem;
1017
1018 BUG_ON(!addr);
1019 BUG_ON(addr < VMALLOC_START);
1020 BUG_ON(addr > VMALLOC_END);
1021 BUG_ON(addr & (PAGE_SIZE-1));
1022
1023 debug_check_no_locks_freed(mem, size);
1024 vmap_debug_free_range(addr, addr+size);
1025
1026 if (likely(count <= VMAP_MAX_ALLOC))
1027 vb_free(mem, size);
1028 else
1029 free_unmap_vmap_area_addr(addr);
1030 }
1031 EXPORT_SYMBOL(vm_unmap_ram);
1032
1033 /**
1034 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1035 * @pages: an array of pointers to the pages to be mapped
1036 * @count: number of pages
1037 * @node: prefer to allocate data structures on this node
1038 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1039 *
1040 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1041 */
1042 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1043 {
1044 unsigned long size = count << PAGE_SHIFT;
1045 unsigned long addr;
1046 void *mem;
1047
1048 if (likely(count <= VMAP_MAX_ALLOC)) {
1049 mem = vb_alloc(size, GFP_KERNEL);
1050 if (IS_ERR(mem))
1051 return NULL;
1052 addr = (unsigned long)mem;
1053 } else {
1054 struct vmap_area *va;
1055 va = alloc_vmap_area(size, PAGE_SIZE,
1056 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1057 if (IS_ERR(va))
1058 return NULL;
1059
1060 addr = va->va_start;
1061 mem = (void *)addr;
1062 }
1063 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1064 vm_unmap_ram(mem, count);
1065 return NULL;
1066 }
1067 return mem;
1068 }
1069 EXPORT_SYMBOL(vm_map_ram);
1070
1071 /**
1072 * vm_area_register_early - register vmap area early during boot
1073 * @vm: vm_struct to register
1074 * @align: requested alignment
1075 *
1076 * This function is used to register kernel vm area before
1077 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1078 * proper values on entry and other fields should be zero. On return,
1079 * vm->addr contains the allocated address.
1080 *
1081 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1082 */
1083 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1084 {
1085 static size_t vm_init_off __initdata;
1086 unsigned long addr;
1087
1088 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1089 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1090
1091 vm->addr = (void *)addr;
1092
1093 vm->next = vmlist;
1094 vmlist = vm;
1095 }
1096
1097 void __init vmalloc_init(void)
1098 {
1099 struct vmap_area *va;
1100 struct vm_struct *tmp;
1101 int i;
1102
1103 for_each_possible_cpu(i) {
1104 struct vmap_block_queue *vbq;
1105
1106 vbq = &per_cpu(vmap_block_queue, i);
1107 spin_lock_init(&vbq->lock);
1108 INIT_LIST_HEAD(&vbq->free);
1109 }
1110
1111 /* Import existing vmlist entries. */
1112 for (tmp = vmlist; tmp; tmp = tmp->next) {
1113 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1114 va->flags = tmp->flags | VM_VM_AREA;
1115 va->va_start = (unsigned long)tmp->addr;
1116 va->va_end = va->va_start + tmp->size;
1117 __insert_vmap_area(va);
1118 }
1119
1120 vmap_area_pcpu_hole = VMALLOC_END;
1121
1122 vmap_initialized = true;
1123 }
1124
1125 /**
1126 * map_kernel_range_noflush - map kernel VM area with the specified pages
1127 * @addr: start of the VM area to map
1128 * @size: size of the VM area to map
1129 * @prot: page protection flags to use
1130 * @pages: pages to map
1131 *
1132 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1133 * specify should have been allocated using get_vm_area() and its
1134 * friends.
1135 *
1136 * NOTE:
1137 * This function does NOT do any cache flushing. The caller is
1138 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1139 * before calling this function.
1140 *
1141 * RETURNS:
1142 * The number of pages mapped on success, -errno on failure.
1143 */
1144 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1145 pgprot_t prot, struct page **pages)
1146 {
1147 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1148 }
1149
1150 /**
1151 * unmap_kernel_range_noflush - unmap kernel VM area
1152 * @addr: start of the VM area to unmap
1153 * @size: size of the VM area to unmap
1154 *
1155 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1156 * specify should have been allocated using get_vm_area() and its
1157 * friends.
1158 *
1159 * NOTE:
1160 * This function does NOT do any cache flushing. The caller is
1161 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1162 * before calling this function and flush_tlb_kernel_range() after.
1163 */
1164 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1165 {
1166 vunmap_page_range(addr, addr + size);
1167 }
1168
1169 /**
1170 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1171 * @addr: start of the VM area to unmap
1172 * @size: size of the VM area to unmap
1173 *
1174 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1175 * the unmapping and tlb after.
1176 */
1177 void unmap_kernel_range(unsigned long addr, unsigned long size)
1178 {
1179 unsigned long end = addr + size;
1180
1181 flush_cache_vunmap(addr, end);
1182 vunmap_page_range(addr, end);
1183 flush_tlb_kernel_range(addr, end);
1184 }
1185
1186 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1187 {
1188 unsigned long addr = (unsigned long)area->addr;
1189 unsigned long end = addr + area->size - PAGE_SIZE;
1190 int err;
1191
1192 err = vmap_page_range(addr, end, prot, *pages);
1193 if (err > 0) {
1194 *pages += err;
1195 err = 0;
1196 }
1197
1198 return err;
1199 }
1200 EXPORT_SYMBOL_GPL(map_vm_area);
1201
1202 /*** Old vmalloc interfaces ***/
1203 DEFINE_RWLOCK(vmlist_lock);
1204 struct vm_struct *vmlist;
1205
1206 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1207 unsigned long flags, void *caller)
1208 {
1209 vm->flags = flags;
1210 vm->addr = (void *)va->va_start;
1211 vm->size = va->va_end - va->va_start;
1212 vm->caller = caller;
1213 va->private = vm;
1214 va->flags |= VM_VM_AREA;
1215 }
1216
1217 static void insert_vmalloc_vmlist(struct vm_struct *vm)
1218 {
1219 struct vm_struct *tmp, **p;
1220
1221 vm->flags &= ~VM_UNLIST;
1222
1223 write_lock(&vmlist_lock);
1224 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1225 if (tmp->addr >= vm->addr)
1226 break;
1227 }
1228 vm->next = *p;
1229 *p = vm;
1230 write_unlock(&vmlist_lock);
1231 }
1232
1233 static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1234 unsigned long flags, void *caller)
1235 {
1236 setup_vmalloc_vm(vm, va, flags, caller);
1237 insert_vmalloc_vmlist(vm);
1238 }
1239
1240 static struct vm_struct *__get_vm_area_node(unsigned long size,
1241 unsigned long align, unsigned long flags, unsigned long start,
1242 unsigned long end, int node, gfp_t gfp_mask, void *caller)
1243 {
1244 static struct vmap_area *va;
1245 struct vm_struct *area;
1246
1247 BUG_ON(in_interrupt());
1248 if (flags & VM_IOREMAP) {
1249 int bit = fls(size);
1250
1251 if (bit > IOREMAP_MAX_ORDER)
1252 bit = IOREMAP_MAX_ORDER;
1253 else if (bit < PAGE_SHIFT)
1254 bit = PAGE_SHIFT;
1255
1256 align = 1ul << bit;
1257 }
1258
1259 size = PAGE_ALIGN(size);
1260 if (unlikely(!size))
1261 return NULL;
1262
1263 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1264 if (unlikely(!area))
1265 return NULL;
1266
1267 /*
1268 * We always allocate a guard page.
1269 */
1270 size += PAGE_SIZE;
1271
1272 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1273 if (IS_ERR(va)) {
1274 kfree(area);
1275 return NULL;
1276 }
1277
1278 /*
1279 * When this function is called from __vmalloc_node,
1280 * we do not add vm_struct to vmlist here to avoid
1281 * accessing uninitialized members of vm_struct such as
1282 * pages and nr_pages fields. They will be set later.
1283 * To distinguish it from others, we use a VM_UNLIST flag.
1284 */
1285 if (flags & VM_UNLIST)
1286 setup_vmalloc_vm(area, va, flags, caller);
1287 else
1288 insert_vmalloc_vm(area, va, flags, caller);
1289
1290 return area;
1291 }
1292
1293 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1294 unsigned long start, unsigned long end)
1295 {
1296 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1297 __builtin_return_address(0));
1298 }
1299 EXPORT_SYMBOL_GPL(__get_vm_area);
1300
1301 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1302 unsigned long start, unsigned long end,
1303 void *caller)
1304 {
1305 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1306 caller);
1307 }
1308
1309 /**
1310 * get_vm_area - reserve a contiguous kernel virtual area
1311 * @size: size of the area
1312 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1313 *
1314 * Search an area of @size in the kernel virtual mapping area,
1315 * and reserved it for out purposes. Returns the area descriptor
1316 * on success or %NULL on failure.
1317 */
1318 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1319 {
1320 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1321 -1, GFP_KERNEL, __builtin_return_address(0));
1322 }
1323
1324 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1325 void *caller)
1326 {
1327 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1328 -1, GFP_KERNEL, caller);
1329 }
1330
1331 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1332 int node, gfp_t gfp_mask)
1333 {
1334 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1335 node, gfp_mask, __builtin_return_address(0));
1336 }
1337
1338 static struct vm_struct *find_vm_area(const void *addr)
1339 {
1340 struct vmap_area *va;
1341
1342 va = find_vmap_area((unsigned long)addr);
1343 if (va && va->flags & VM_VM_AREA)
1344 return va->private;
1345
1346 return NULL;
1347 }
1348
1349 /**
1350 * remove_vm_area - find and remove a continuous kernel virtual area
1351 * @addr: base address
1352 *
1353 * Search for the kernel VM area starting at @addr, and remove it.
1354 * This function returns the found VM area, but using it is NOT safe
1355 * on SMP machines, except for its size or flags.
1356 */
1357 struct vm_struct *remove_vm_area(const void *addr)
1358 {
1359 struct vmap_area *va;
1360
1361 va = find_vmap_area((unsigned long)addr);
1362 if (va && va->flags & VM_VM_AREA) {
1363 struct vm_struct *vm = va->private;
1364
1365 if (!(vm->flags & VM_UNLIST)) {
1366 struct vm_struct *tmp, **p;
1367 /*
1368 * remove from list and disallow access to
1369 * this vm_struct before unmap. (address range
1370 * confliction is maintained by vmap.)
1371 */
1372 write_lock(&vmlist_lock);
1373 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1374 ;
1375 *p = tmp->next;
1376 write_unlock(&vmlist_lock);
1377 }
1378
1379 vmap_debug_free_range(va->va_start, va->va_end);
1380 free_unmap_vmap_area(va);
1381 vm->size -= PAGE_SIZE;
1382
1383 return vm;
1384 }
1385 return NULL;
1386 }
1387
1388 static void __vunmap(const void *addr, int deallocate_pages)
1389 {
1390 struct vm_struct *area;
1391
1392 if (!addr)
1393 return;
1394
1395 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1396 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1397 return;
1398 }
1399
1400 area = remove_vm_area(addr);
1401 if (unlikely(!area)) {
1402 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1403 addr);
1404 return;
1405 }
1406
1407 debug_check_no_locks_freed(addr, area->size);
1408 debug_check_no_obj_freed(addr, area->size);
1409
1410 if (deallocate_pages) {
1411 int i;
1412
1413 for (i = 0; i < area->nr_pages; i++) {
1414 struct page *page = area->pages[i];
1415
1416 BUG_ON(!page);
1417 __free_page(page);
1418 }
1419
1420 if (area->flags & VM_VPAGES)
1421 vfree(area->pages);
1422 else
1423 kfree(area->pages);
1424 }
1425
1426 kfree(area);
1427 return;
1428 }
1429
1430 /**
1431 * vfree - release memory allocated by vmalloc()
1432 * @addr: memory base address
1433 *
1434 * Free the virtually continuous memory area starting at @addr, as
1435 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1436 * NULL, no operation is performed.
1437 *
1438 * Must not be called in interrupt context.
1439 */
1440 void vfree(const void *addr)
1441 {
1442 BUG_ON(in_interrupt());
1443
1444 kmemleak_free(addr);
1445
1446 __vunmap(addr, 1);
1447 }
1448 EXPORT_SYMBOL(vfree);
1449
1450 /**
1451 * vunmap - release virtual mapping obtained by vmap()
1452 * @addr: memory base address
1453 *
1454 * Free the virtually contiguous memory area starting at @addr,
1455 * which was created from the page array passed to vmap().
1456 *
1457 * Must not be called in interrupt context.
1458 */
1459 void vunmap(const void *addr)
1460 {
1461 BUG_ON(in_interrupt());
1462 might_sleep();
1463 __vunmap(addr, 0);
1464 }
1465 EXPORT_SYMBOL(vunmap);
1466
1467 /**
1468 * vmap - map an array of pages into virtually contiguous space
1469 * @pages: array of page pointers
1470 * @count: number of pages to map
1471 * @flags: vm_area->flags
1472 * @prot: page protection for the mapping
1473 *
1474 * Maps @count pages from @pages into contiguous kernel virtual
1475 * space.
1476 */
1477 void *vmap(struct page **pages, unsigned int count,
1478 unsigned long flags, pgprot_t prot)
1479 {
1480 struct vm_struct *area;
1481
1482 might_sleep();
1483
1484 if (count > totalram_pages)
1485 return NULL;
1486
1487 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1488 __builtin_return_address(0));
1489 if (!area)
1490 return NULL;
1491
1492 if (map_vm_area(area, prot, &pages)) {
1493 vunmap(area->addr);
1494 return NULL;
1495 }
1496
1497 return area->addr;
1498 }
1499 EXPORT_SYMBOL(vmap);
1500
1501 static void *__vmalloc_node(unsigned long size, unsigned long align,
1502 gfp_t gfp_mask, pgprot_t prot,
1503 int node, void *caller);
1504 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1505 pgprot_t prot, int node, void *caller)
1506 {
1507 struct page **pages;
1508 unsigned int nr_pages, array_size, i;
1509
1510 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1511 array_size = (nr_pages * sizeof(struct page *));
1512
1513 area->nr_pages = nr_pages;
1514 /* Please note that the recursion is strictly bounded. */
1515 if (array_size > PAGE_SIZE) {
1516 pages = __vmalloc_node(array_size, 1, gfp_mask | __GFP_ZERO,
1517 PAGE_KERNEL, node, caller);
1518 area->flags |= VM_VPAGES;
1519 } else {
1520 pages = kmalloc_node(array_size,
1521 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1522 node);
1523 }
1524 area->pages = pages;
1525 area->caller = caller;
1526 if (!area->pages) {
1527 remove_vm_area(area->addr);
1528 kfree(area);
1529 return NULL;
1530 }
1531
1532 for (i = 0; i < area->nr_pages; i++) {
1533 struct page *page;
1534
1535 if (node < 0)
1536 page = alloc_page(gfp_mask);
1537 else
1538 page = alloc_pages_node(node, gfp_mask, 0);
1539
1540 if (unlikely(!page)) {
1541 /* Successfully allocated i pages, free them in __vunmap() */
1542 area->nr_pages = i;
1543 goto fail;
1544 }
1545 area->pages[i] = page;
1546 }
1547
1548 if (map_vm_area(area, prot, &pages))
1549 goto fail;
1550 return area->addr;
1551
1552 fail:
1553 vfree(area->addr);
1554 return NULL;
1555 }
1556
1557 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1558 {
1559 void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1560 __builtin_return_address(0));
1561
1562 /*
1563 * A ref_count = 3 is needed because the vm_struct and vmap_area
1564 * structures allocated in the __get_vm_area_node() function contain
1565 * references to the virtual address of the vmalloc'ed block.
1566 */
1567 kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1568
1569 return addr;
1570 }
1571
1572 /**
1573 * __vmalloc_node - allocate virtually contiguous memory
1574 * @size: allocation size
1575 * @align: desired alignment
1576 * @gfp_mask: flags for the page level allocator
1577 * @prot: protection mask for the allocated pages
1578 * @node: node to use for allocation or -1
1579 * @caller: caller's return address
1580 *
1581 * Allocate enough pages to cover @size from the page level
1582 * allocator with @gfp_mask flags. Map them into contiguous
1583 * kernel virtual space, using a pagetable protection of @prot.
1584 */
1585 static void *__vmalloc_node(unsigned long size, unsigned long align,
1586 gfp_t gfp_mask, pgprot_t prot,
1587 int node, void *caller)
1588 {
1589 struct vm_struct *area;
1590 void *addr;
1591 unsigned long real_size = size;
1592
1593 size = PAGE_ALIGN(size);
1594 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1595 return NULL;
1596
1597 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1598 VMALLOC_START, VMALLOC_END, node,
1599 gfp_mask, caller);
1600
1601 if (!area)
1602 return NULL;
1603
1604 addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1605
1606 /*
1607 * In this function, newly allocated vm_struct is not added
1608 * to vmlist at __get_vm_area_node(). so, it is added here.
1609 */
1610 insert_vmalloc_vmlist(area);
1611
1612 /*
1613 * A ref_count = 3 is needed because the vm_struct and vmap_area
1614 * structures allocated in the __get_vm_area_node() function contain
1615 * references to the virtual address of the vmalloc'ed block.
1616 */
1617 kmemleak_alloc(addr, real_size, 3, gfp_mask);
1618
1619 return addr;
1620 }
1621
1622 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1623 {
1624 return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1625 __builtin_return_address(0));
1626 }
1627 EXPORT_SYMBOL(__vmalloc);
1628
1629 /**
1630 * vmalloc - allocate virtually contiguous memory
1631 * @size: allocation size
1632 * Allocate enough pages to cover @size from the page level
1633 * allocator and map them into contiguous kernel virtual space.
1634 *
1635 * For tight control over page level allocator and protection flags
1636 * use __vmalloc() instead.
1637 */
1638 void *vmalloc(unsigned long size)
1639 {
1640 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1641 -1, __builtin_return_address(0));
1642 }
1643 EXPORT_SYMBOL(vmalloc);
1644
1645 /**
1646 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1647 * @size: allocation size
1648 *
1649 * The resulting memory area is zeroed so it can be mapped to userspace
1650 * without leaking data.
1651 */
1652 void *vmalloc_user(unsigned long size)
1653 {
1654 struct vm_struct *area;
1655 void *ret;
1656
1657 ret = __vmalloc_node(size, SHMLBA,
1658 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1659 PAGE_KERNEL, -1, __builtin_return_address(0));
1660 if (ret) {
1661 area = find_vm_area(ret);
1662 area->flags |= VM_USERMAP;
1663 }
1664 return ret;
1665 }
1666 EXPORT_SYMBOL(vmalloc_user);
1667
1668 /**
1669 * vmalloc_node - allocate memory on a specific node
1670 * @size: allocation size
1671 * @node: numa node
1672 *
1673 * Allocate enough pages to cover @size from the page level
1674 * allocator and map them into contiguous kernel virtual space.
1675 *
1676 * For tight control over page level allocator and protection flags
1677 * use __vmalloc() instead.
1678 */
1679 void *vmalloc_node(unsigned long size, int node)
1680 {
1681 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1682 node, __builtin_return_address(0));
1683 }
1684 EXPORT_SYMBOL(vmalloc_node);
1685
1686 #ifndef PAGE_KERNEL_EXEC
1687 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1688 #endif
1689
1690 /**
1691 * vmalloc_exec - allocate virtually contiguous, executable memory
1692 * @size: allocation size
1693 *
1694 * Kernel-internal function to allocate enough pages to cover @size
1695 * the page level allocator and map them into contiguous and
1696 * executable kernel virtual space.
1697 *
1698 * For tight control over page level allocator and protection flags
1699 * use __vmalloc() instead.
1700 */
1701
1702 void *vmalloc_exec(unsigned long size)
1703 {
1704 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1705 -1, __builtin_return_address(0));
1706 }
1707
1708 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1709 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1710 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1711 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1712 #else
1713 #define GFP_VMALLOC32 GFP_KERNEL
1714 #endif
1715
1716 /**
1717 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1718 * @size: allocation size
1719 *
1720 * Allocate enough 32bit PA addressable pages to cover @size from the
1721 * page level allocator and map them into contiguous kernel virtual space.
1722 */
1723 void *vmalloc_32(unsigned long size)
1724 {
1725 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1726 -1, __builtin_return_address(0));
1727 }
1728 EXPORT_SYMBOL(vmalloc_32);
1729
1730 /**
1731 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1732 * @size: allocation size
1733 *
1734 * The resulting memory area is 32bit addressable and zeroed so it can be
1735 * mapped to userspace without leaking data.
1736 */
1737 void *vmalloc_32_user(unsigned long size)
1738 {
1739 struct vm_struct *area;
1740 void *ret;
1741
1742 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1743 -1, __builtin_return_address(0));
1744 if (ret) {
1745 area = find_vm_area(ret);
1746 area->flags |= VM_USERMAP;
1747 }
1748 return ret;
1749 }
1750 EXPORT_SYMBOL(vmalloc_32_user);
1751
1752 /*
1753 * small helper routine , copy contents to buf from addr.
1754 * If the page is not present, fill zero.
1755 */
1756
1757 static int aligned_vread(char *buf, char *addr, unsigned long count)
1758 {
1759 struct page *p;
1760 int copied = 0;
1761
1762 while (count) {
1763 unsigned long offset, length;
1764
1765 offset = (unsigned long)addr & ~PAGE_MASK;
1766 length = PAGE_SIZE - offset;
1767 if (length > count)
1768 length = count;
1769 p = vmalloc_to_page(addr);
1770 /*
1771 * To do safe access to this _mapped_ area, we need
1772 * lock. But adding lock here means that we need to add
1773 * overhead of vmalloc()/vfree() calles for this _debug_
1774 * interface, rarely used. Instead of that, we'll use
1775 * kmap() and get small overhead in this access function.
1776 */
1777 if (p) {
1778 /*
1779 * we can expect USER0 is not used (see vread/vwrite's
1780 * function description)
1781 */
1782 void *map = kmap_atomic(p, KM_USER0);
1783 memcpy(buf, map + offset, length);
1784 kunmap_atomic(map, KM_USER0);
1785 } else
1786 memset(buf, 0, length);
1787
1788 addr += length;
1789 buf += length;
1790 copied += length;
1791 count -= length;
1792 }
1793 return copied;
1794 }
1795
1796 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1797 {
1798 struct page *p;
1799 int copied = 0;
1800
1801 while (count) {
1802 unsigned long offset, length;
1803
1804 offset = (unsigned long)addr & ~PAGE_MASK;
1805 length = PAGE_SIZE - offset;
1806 if (length > count)
1807 length = count;
1808 p = vmalloc_to_page(addr);
1809 /*
1810 * To do safe access to this _mapped_ area, we need
1811 * lock. But adding lock here means that we need to add
1812 * overhead of vmalloc()/vfree() calles for this _debug_
1813 * interface, rarely used. Instead of that, we'll use
1814 * kmap() and get small overhead in this access function.
1815 */
1816 if (p) {
1817 /*
1818 * we can expect USER0 is not used (see vread/vwrite's
1819 * function description)
1820 */
1821 void *map = kmap_atomic(p, KM_USER0);
1822 memcpy(map + offset, buf, length);
1823 kunmap_atomic(map, KM_USER0);
1824 }
1825 addr += length;
1826 buf += length;
1827 copied += length;
1828 count -= length;
1829 }
1830 return copied;
1831 }
1832
1833 /**
1834 * vread() - read vmalloc area in a safe way.
1835 * @buf: buffer for reading data
1836 * @addr: vm address.
1837 * @count: number of bytes to be read.
1838 *
1839 * Returns # of bytes which addr and buf should be increased.
1840 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1841 * includes any intersect with alive vmalloc area.
1842 *
1843 * This function checks that addr is a valid vmalloc'ed area, and
1844 * copy data from that area to a given buffer. If the given memory range
1845 * of [addr...addr+count) includes some valid address, data is copied to
1846 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1847 * IOREMAP area is treated as memory hole and no copy is done.
1848 *
1849 * If [addr...addr+count) doesn't includes any intersects with alive
1850 * vm_struct area, returns 0.
1851 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1852 * the caller should guarantee KM_USER0 is not used.
1853 *
1854 * Note: In usual ops, vread() is never necessary because the caller
1855 * should know vmalloc() area is valid and can use memcpy().
1856 * This is for routines which have to access vmalloc area without
1857 * any informaion, as /dev/kmem.
1858 *
1859 */
1860
1861 long vread(char *buf, char *addr, unsigned long count)
1862 {
1863 struct vm_struct *tmp;
1864 char *vaddr, *buf_start = buf;
1865 unsigned long buflen = count;
1866 unsigned long n;
1867
1868 /* Don't allow overflow */
1869 if ((unsigned long) addr + count < count)
1870 count = -(unsigned long) addr;
1871
1872 read_lock(&vmlist_lock);
1873 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1874 vaddr = (char *) tmp->addr;
1875 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1876 continue;
1877 while (addr < vaddr) {
1878 if (count == 0)
1879 goto finished;
1880 *buf = '\0';
1881 buf++;
1882 addr++;
1883 count--;
1884 }
1885 n = vaddr + tmp->size - PAGE_SIZE - addr;
1886 if (n > count)
1887 n = count;
1888 if (!(tmp->flags & VM_IOREMAP))
1889 aligned_vread(buf, addr, n);
1890 else /* IOREMAP area is treated as memory hole */
1891 memset(buf, 0, n);
1892 buf += n;
1893 addr += n;
1894 count -= n;
1895 }
1896 finished:
1897 read_unlock(&vmlist_lock);
1898
1899 if (buf == buf_start)
1900 return 0;
1901 /* zero-fill memory holes */
1902 if (buf != buf_start + buflen)
1903 memset(buf, 0, buflen - (buf - buf_start));
1904
1905 return buflen;
1906 }
1907
1908 /**
1909 * vwrite() - write vmalloc area in a safe way.
1910 * @buf: buffer for source data
1911 * @addr: vm address.
1912 * @count: number of bytes to be read.
1913 *
1914 * Returns # of bytes which addr and buf should be incresed.
1915 * (same number to @count).
1916 * If [addr...addr+count) doesn't includes any intersect with valid
1917 * vmalloc area, returns 0.
1918 *
1919 * This function checks that addr is a valid vmalloc'ed area, and
1920 * copy data from a buffer to the given addr. If specified range of
1921 * [addr...addr+count) includes some valid address, data is copied from
1922 * proper area of @buf. If there are memory holes, no copy to hole.
1923 * IOREMAP area is treated as memory hole and no copy is done.
1924 *
1925 * If [addr...addr+count) doesn't includes any intersects with alive
1926 * vm_struct area, returns 0.
1927 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1928 * the caller should guarantee KM_USER0 is not used.
1929 *
1930 * Note: In usual ops, vwrite() is never necessary because the caller
1931 * should know vmalloc() area is valid and can use memcpy().
1932 * This is for routines which have to access vmalloc area without
1933 * any informaion, as /dev/kmem.
1934 *
1935 * The caller should guarantee KM_USER1 is not used.
1936 */
1937
1938 long vwrite(char *buf, char *addr, unsigned long count)
1939 {
1940 struct vm_struct *tmp;
1941 char *vaddr;
1942 unsigned long n, buflen;
1943 int copied = 0;
1944
1945 /* Don't allow overflow */
1946 if ((unsigned long) addr + count < count)
1947 count = -(unsigned long) addr;
1948 buflen = count;
1949
1950 read_lock(&vmlist_lock);
1951 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1952 vaddr = (char *) tmp->addr;
1953 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1954 continue;
1955 while (addr < vaddr) {
1956 if (count == 0)
1957 goto finished;
1958 buf++;
1959 addr++;
1960 count--;
1961 }
1962 n = vaddr + tmp->size - PAGE_SIZE - addr;
1963 if (n > count)
1964 n = count;
1965 if (!(tmp->flags & VM_IOREMAP)) {
1966 aligned_vwrite(buf, addr, n);
1967 copied++;
1968 }
1969 buf += n;
1970 addr += n;
1971 count -= n;
1972 }
1973 finished:
1974 read_unlock(&vmlist_lock);
1975 if (!copied)
1976 return 0;
1977 return buflen;
1978 }
1979
1980 /**
1981 * remap_vmalloc_range - map vmalloc pages to userspace
1982 * @vma: vma to cover (map full range of vma)
1983 * @addr: vmalloc memory
1984 * @pgoff: number of pages into addr before first page to map
1985 *
1986 * Returns: 0 for success, -Exxx on failure
1987 *
1988 * This function checks that addr is a valid vmalloc'ed area, and
1989 * that it is big enough to cover the vma. Will return failure if
1990 * that criteria isn't met.
1991 *
1992 * Similar to remap_pfn_range() (see mm/memory.c)
1993 */
1994 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1995 unsigned long pgoff)
1996 {
1997 struct vm_struct *area;
1998 unsigned long uaddr = vma->vm_start;
1999 unsigned long usize = vma->vm_end - vma->vm_start;
2000
2001 if ((PAGE_SIZE-1) & (unsigned long)addr)
2002 return -EINVAL;
2003
2004 area = find_vm_area(addr);
2005 if (!area)
2006 return -EINVAL;
2007
2008 if (!(area->flags & VM_USERMAP))
2009 return -EINVAL;
2010
2011 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2012 return -EINVAL;
2013
2014 addr += pgoff << PAGE_SHIFT;
2015 do {
2016 struct page *page = vmalloc_to_page(addr);
2017 int ret;
2018
2019 ret = vm_insert_page(vma, uaddr, page);
2020 if (ret)
2021 return ret;
2022
2023 uaddr += PAGE_SIZE;
2024 addr += PAGE_SIZE;
2025 usize -= PAGE_SIZE;
2026 } while (usize > 0);
2027
2028 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2029 vma->vm_flags |= VM_RESERVED;
2030
2031 return 0;
2032 }
2033 EXPORT_SYMBOL(remap_vmalloc_range);
2034
2035 /*
2036 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2037 * have one.
2038 */
2039 void __attribute__((weak)) vmalloc_sync_all(void)
2040 {
2041 }
2042
2043
2044 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2045 {
2046 /* apply_to_page_range() does all the hard work. */
2047 return 0;
2048 }
2049
2050 /**
2051 * alloc_vm_area - allocate a range of kernel address space
2052 * @size: size of the area
2053 *
2054 * Returns: NULL on failure, vm_struct on success
2055 *
2056 * This function reserves a range of kernel address space, and
2057 * allocates pagetables to map that range. No actual mappings
2058 * are created. If the kernel address space is not shared
2059 * between processes, it syncs the pagetable across all
2060 * processes.
2061 */
2062 struct vm_struct *alloc_vm_area(size_t size)
2063 {
2064 struct vm_struct *area;
2065
2066 area = get_vm_area_caller(size, VM_IOREMAP,
2067 __builtin_return_address(0));
2068 if (area == NULL)
2069 return NULL;
2070
2071 /*
2072 * This ensures that page tables are constructed for this region
2073 * of kernel virtual address space and mapped into init_mm.
2074 */
2075 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2076 area->size, f, NULL)) {
2077 free_vm_area(area);
2078 return NULL;
2079 }
2080
2081 /* Make sure the pagetables are constructed in process kernel
2082 mappings */
2083 vmalloc_sync_all();
2084
2085 return area;
2086 }
2087 EXPORT_SYMBOL_GPL(alloc_vm_area);
2088
2089 void free_vm_area(struct vm_struct *area)
2090 {
2091 struct vm_struct *ret;
2092 ret = remove_vm_area(area->addr);
2093 BUG_ON(ret != area);
2094 kfree(area);
2095 }
2096 EXPORT_SYMBOL_GPL(free_vm_area);
2097
2098 #ifndef CONFIG_HAVE_LEGACY_PER_CPU_AREA
2099 static struct vmap_area *node_to_va(struct rb_node *n)
2100 {
2101 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2102 }
2103
2104 /**
2105 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2106 * @end: target address
2107 * @pnext: out arg for the next vmap_area
2108 * @pprev: out arg for the previous vmap_area
2109 *
2110 * Returns: %true if either or both of next and prev are found,
2111 * %false if no vmap_area exists
2112 *
2113 * Find vmap_areas end addresses of which enclose @end. ie. if not
2114 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2115 */
2116 static bool pvm_find_next_prev(unsigned long end,
2117 struct vmap_area **pnext,
2118 struct vmap_area **pprev)
2119 {
2120 struct rb_node *n = vmap_area_root.rb_node;
2121 struct vmap_area *va = NULL;
2122
2123 while (n) {
2124 va = rb_entry(n, struct vmap_area, rb_node);
2125 if (end < va->va_end)
2126 n = n->rb_left;
2127 else if (end > va->va_end)
2128 n = n->rb_right;
2129 else
2130 break;
2131 }
2132
2133 if (!va)
2134 return false;
2135
2136 if (va->va_end > end) {
2137 *pnext = va;
2138 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2139 } else {
2140 *pprev = va;
2141 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2142 }
2143 return true;
2144 }
2145
2146 /**
2147 * pvm_determine_end - find the highest aligned address between two vmap_areas
2148 * @pnext: in/out arg for the next vmap_area
2149 * @pprev: in/out arg for the previous vmap_area
2150 * @align: alignment
2151 *
2152 * Returns: determined end address
2153 *
2154 * Find the highest aligned address between *@pnext and *@pprev below
2155 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2156 * down address is between the end addresses of the two vmap_areas.
2157 *
2158 * Please note that the address returned by this function may fall
2159 * inside *@pnext vmap_area. The caller is responsible for checking
2160 * that.
2161 */
2162 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2163 struct vmap_area **pprev,
2164 unsigned long align)
2165 {
2166 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2167 unsigned long addr;
2168
2169 if (*pnext)
2170 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2171 else
2172 addr = vmalloc_end;
2173
2174 while (*pprev && (*pprev)->va_end > addr) {
2175 *pnext = *pprev;
2176 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2177 }
2178
2179 return addr;
2180 }
2181
2182 /**
2183 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2184 * @offsets: array containing offset of each area
2185 * @sizes: array containing size of each area
2186 * @nr_vms: the number of areas to allocate
2187 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2188 * @gfp_mask: allocation mask
2189 *
2190 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2191 * vm_structs on success, %NULL on failure
2192 *
2193 * Percpu allocator wants to use congruent vm areas so that it can
2194 * maintain the offsets among percpu areas. This function allocates
2195 * congruent vmalloc areas for it. These areas tend to be scattered
2196 * pretty far, distance between two areas easily going up to
2197 * gigabytes. To avoid interacting with regular vmallocs, these areas
2198 * are allocated from top.
2199 *
2200 * Despite its complicated look, this allocator is rather simple. It
2201 * does everything top-down and scans areas from the end looking for
2202 * matching slot. While scanning, if any of the areas overlaps with
2203 * existing vmap_area, the base address is pulled down to fit the
2204 * area. Scanning is repeated till all the areas fit and then all
2205 * necessary data structres are inserted and the result is returned.
2206 */
2207 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2208 const size_t *sizes, int nr_vms,
2209 size_t align, gfp_t gfp_mask)
2210 {
2211 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2212 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2213 struct vmap_area **vas, *prev, *next;
2214 struct vm_struct **vms;
2215 int area, area2, last_area, term_area;
2216 unsigned long base, start, end, last_end;
2217 bool purged = false;
2218
2219 gfp_mask &= GFP_RECLAIM_MASK;
2220
2221 /* verify parameters and allocate data structures */
2222 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2223 for (last_area = 0, area = 0; area < nr_vms; area++) {
2224 start = offsets[area];
2225 end = start + sizes[area];
2226
2227 /* is everything aligned properly? */
2228 BUG_ON(!IS_ALIGNED(offsets[area], align));
2229 BUG_ON(!IS_ALIGNED(sizes[area], align));
2230
2231 /* detect the area with the highest address */
2232 if (start > offsets[last_area])
2233 last_area = area;
2234
2235 for (area2 = 0; area2 < nr_vms; area2++) {
2236 unsigned long start2 = offsets[area2];
2237 unsigned long end2 = start2 + sizes[area2];
2238
2239 if (area2 == area)
2240 continue;
2241
2242 BUG_ON(start2 >= start && start2 < end);
2243 BUG_ON(end2 <= end && end2 > start);
2244 }
2245 }
2246 last_end = offsets[last_area] + sizes[last_area];
2247
2248 if (vmalloc_end - vmalloc_start < last_end) {
2249 WARN_ON(true);
2250 return NULL;
2251 }
2252
2253 vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2254 vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2255 if (!vas || !vms)
2256 goto err_free;
2257
2258 for (area = 0; area < nr_vms; area++) {
2259 vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2260 vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2261 if (!vas[area] || !vms[area])
2262 goto err_free;
2263 }
2264 retry:
2265 spin_lock(&vmap_area_lock);
2266
2267 /* start scanning - we scan from the top, begin with the last area */
2268 area = term_area = last_area;
2269 start = offsets[area];
2270 end = start + sizes[area];
2271
2272 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2273 base = vmalloc_end - last_end;
2274 goto found;
2275 }
2276 base = pvm_determine_end(&next, &prev, align) - end;
2277
2278 while (true) {
2279 BUG_ON(next && next->va_end <= base + end);
2280 BUG_ON(prev && prev->va_end > base + end);
2281
2282 /*
2283 * base might have underflowed, add last_end before
2284 * comparing.
2285 */
2286 if (base + last_end < vmalloc_start + last_end) {
2287 spin_unlock(&vmap_area_lock);
2288 if (!purged) {
2289 purge_vmap_area_lazy();
2290 purged = true;
2291 goto retry;
2292 }
2293 goto err_free;
2294 }
2295
2296 /*
2297 * If next overlaps, move base downwards so that it's
2298 * right below next and then recheck.
2299 */
2300 if (next && next->va_start < base + end) {
2301 base = pvm_determine_end(&next, &prev, align) - end;
2302 term_area = area;
2303 continue;
2304 }
2305
2306 /*
2307 * If prev overlaps, shift down next and prev and move
2308 * base so that it's right below new next and then
2309 * recheck.
2310 */
2311 if (prev && prev->va_end > base + start) {
2312 next = prev;
2313 prev = node_to_va(rb_prev(&next->rb_node));
2314 base = pvm_determine_end(&next, &prev, align) - end;
2315 term_area = area;
2316 continue;
2317 }
2318
2319 /*
2320 * This area fits, move on to the previous one. If
2321 * the previous one is the terminal one, we're done.
2322 */
2323 area = (area + nr_vms - 1) % nr_vms;
2324 if (area == term_area)
2325 break;
2326 start = offsets[area];
2327 end = start + sizes[area];
2328 pvm_find_next_prev(base + end, &next, &prev);
2329 }
2330 found:
2331 /* we've found a fitting base, insert all va's */
2332 for (area = 0; area < nr_vms; area++) {
2333 struct vmap_area *va = vas[area];
2334
2335 va->va_start = base + offsets[area];
2336 va->va_end = va->va_start + sizes[area];
2337 __insert_vmap_area(va);
2338 }
2339
2340 vmap_area_pcpu_hole = base + offsets[last_area];
2341
2342 spin_unlock(&vmap_area_lock);
2343
2344 /* insert all vm's */
2345 for (area = 0; area < nr_vms; area++)
2346 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2347 pcpu_get_vm_areas);
2348
2349 kfree(vas);
2350 return vms;
2351
2352 err_free:
2353 for (area = 0; area < nr_vms; area++) {
2354 if (vas)
2355 kfree(vas[area]);
2356 if (vms)
2357 kfree(vms[area]);
2358 }
2359 kfree(vas);
2360 kfree(vms);
2361 return NULL;
2362 }
2363 #endif
2364
2365 /**
2366 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2367 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2368 * @nr_vms: the number of allocated areas
2369 *
2370 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2371 */
2372 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2373 {
2374 int i;
2375
2376 for (i = 0; i < nr_vms; i++)
2377 free_vm_area(vms[i]);
2378 kfree(vms);
2379 }
2380
2381 #ifdef CONFIG_PROC_FS
2382 static void *s_start(struct seq_file *m, loff_t *pos)
2383 {
2384 loff_t n = *pos;
2385 struct vm_struct *v;
2386
2387 read_lock(&vmlist_lock);
2388 v = vmlist;
2389 while (n > 0 && v) {
2390 n--;
2391 v = v->next;
2392 }
2393 if (!n)
2394 return v;
2395
2396 return NULL;
2397
2398 }
2399
2400 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2401 {
2402 struct vm_struct *v = p;
2403
2404 ++*pos;
2405 return v->next;
2406 }
2407
2408 static void s_stop(struct seq_file *m, void *p)
2409 {
2410 read_unlock(&vmlist_lock);
2411 }
2412
2413 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2414 {
2415 if (NUMA_BUILD) {
2416 unsigned int nr, *counters = m->private;
2417
2418 if (!counters)
2419 return;
2420
2421 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2422
2423 for (nr = 0; nr < v->nr_pages; nr++)
2424 counters[page_to_nid(v->pages[nr])]++;
2425
2426 for_each_node_state(nr, N_HIGH_MEMORY)
2427 if (counters[nr])
2428 seq_printf(m, " N%u=%u", nr, counters[nr]);
2429 }
2430 }
2431
2432 static int s_show(struct seq_file *m, void *p)
2433 {
2434 struct vm_struct *v = p;
2435
2436 seq_printf(m, "0x%p-0x%p %7ld",
2437 v->addr, v->addr + v->size, v->size);
2438
2439 if (v->caller) {
2440 char buff[KSYM_SYMBOL_LEN];
2441
2442 seq_putc(m, ' ');
2443 sprint_symbol(buff, (unsigned long)v->caller);
2444 seq_puts(m, buff);
2445 }
2446
2447 if (v->nr_pages)
2448 seq_printf(m, " pages=%d", v->nr_pages);
2449
2450 if (v->phys_addr)
2451 seq_printf(m, " phys=%lx", v->phys_addr);
2452
2453 if (v->flags & VM_IOREMAP)
2454 seq_printf(m, " ioremap");
2455
2456 if (v->flags & VM_ALLOC)
2457 seq_printf(m, " vmalloc");
2458
2459 if (v->flags & VM_MAP)
2460 seq_printf(m, " vmap");
2461
2462 if (v->flags & VM_USERMAP)
2463 seq_printf(m, " user");
2464
2465 if (v->flags & VM_VPAGES)
2466 seq_printf(m, " vpages");
2467
2468 show_numa_info(m, v);
2469 seq_putc(m, '\n');
2470 return 0;
2471 }
2472
2473 static const struct seq_operations vmalloc_op = {
2474 .start = s_start,
2475 .next = s_next,
2476 .stop = s_stop,
2477 .show = s_show,
2478 };
2479
2480 static int vmalloc_open(struct inode *inode, struct file *file)
2481 {
2482 unsigned int *ptr = NULL;
2483 int ret;
2484
2485 if (NUMA_BUILD)
2486 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2487 ret = seq_open(file, &vmalloc_op);
2488 if (!ret) {
2489 struct seq_file *m = file->private_data;
2490 m->private = ptr;
2491 } else
2492 kfree(ptr);
2493 return ret;
2494 }
2495
2496 static const struct file_operations proc_vmalloc_operations = {
2497 .open = vmalloc_open,
2498 .read = seq_read,
2499 .llseek = seq_lseek,
2500 .release = seq_release_private,
2501 };
2502
2503 static int __init proc_vmalloc_init(void)
2504 {
2505 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2506 return 0;
2507 }
2508 module_init(proc_vmalloc_init);
2509 #endif
2510
2511
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.