projects/linux·blob v2.6.21 · refs

749 lines · 17.3 KiB

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 <[email protected]>, 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/mm.h>
12#include <linux/module.h>
13#include <linux/highmem.h>
14#include <linux/slab.h>
15#include <linux/spinlock.h>
16#include <linux/interrupt.h>
17
18#include <linux/vmalloc.h>
19
20#include <asm/uaccess.h>
21#include <asm/tlbflush.h>
22
23
24DEFINE_RWLOCK(vmlist_lock);
25struct vm_struct *vmlist;
26
27static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
28 int node);
29
30static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
31{
32 pte_t *pte;
33
34 pte = pte_offset_kernel(pmd, addr);
35 do {
36 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
37 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
38 } while (pte++, addr += PAGE_SIZE, addr != end);
39}
40
41static inline void vunmap_pmd_range(pud_t *pud, unsigned long addr,
42 unsigned long end)
43{
44 pmd_t *pmd;
45 unsigned long next;
46
47 pmd = pmd_offset(pud, addr);
48 do {
49 next = pmd_addr_end(addr, end);
50 if (pmd_none_or_clear_bad(pmd))
51 continue;
52 vunmap_pte_range(pmd, addr, next);
53 } while (pmd++, addr = next, addr != end);
54}
55
56static inline void vunmap_pud_range(pgd_t *pgd, unsigned long addr,
57 unsigned long end)
58{
59 pud_t *pud;
60 unsigned long next;
61
62 pud = pud_offset(pgd, addr);
63 do {
64 next = pud_addr_end(addr, end);
65 if (pud_none_or_clear_bad(pud))
66 continue;
67 vunmap_pmd_range(pud, addr, next);
68 } while (pud++, addr = next, addr != end);
69}
70
71void unmap_vm_area(struct vm_struct *area)
72{
73 pgd_t *pgd;
74 unsigned long next;
75 unsigned long addr = (unsigned long) area->addr;
76 unsigned long end = addr + area->size;
77
78 BUG_ON(addr >= end);
79 pgd = pgd_offset_k(addr);
80 flush_cache_vunmap(addr, end);
81 do {
82 next = pgd_addr_end(addr, end);
83 if (pgd_none_or_clear_bad(pgd))
84 continue;
85 vunmap_pud_range(pgd, addr, next);
86 } while (pgd++, addr = next, addr != end);
87 flush_tlb_kernel_range((unsigned long) area->addr, end);
88}
89
90static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
91 unsigned long end, pgprot_t prot, struct page ***pages)
92{
93 pte_t *pte;
94
95 pte = pte_alloc_kernel(pmd, addr);
96 if (!pte)
97 return -ENOMEM;
98 do {
99 struct page *page = **pages;
100 WARN_ON(!pte_none(*pte));
101 if (!page)
102 return -ENOMEM;
103 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
104 (*pages)++;
105 } while (pte++, addr += PAGE_SIZE, addr != end);
106 return 0;
107}
108
109static inline int vmap_pmd_range(pud_t *pud, unsigned long addr,
110 unsigned long end, pgprot_t prot, struct page ***pages)
111{
112 pmd_t *pmd;
113 unsigned long next;
114
115 pmd = pmd_alloc(&init_mm, pud, addr);
116 if (!pmd)
117 return -ENOMEM;
118 do {
119 next = pmd_addr_end(addr, end);
120 if (vmap_pte_range(pmd, addr, next, prot, pages))
121 return -ENOMEM;
122 } while (pmd++, addr = next, addr != end);
123 return 0;
124}
125
126static inline int vmap_pud_range(pgd_t *pgd, unsigned long addr,
127 unsigned long end, pgprot_t prot, struct page ***pages)
128{
129 pud_t *pud;
130 unsigned long next;
131
132 pud = pud_alloc(&init_mm, pgd, addr);
133 if (!pud)
134 return -ENOMEM;
135 do {
136 next = pud_addr_end(addr, end);
137 if (vmap_pmd_range(pud, addr, next, prot, pages))
138 return -ENOMEM;
139 } while (pud++, addr = next, addr != end);
140 return 0;
141}
142
143int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
144{
145 pgd_t *pgd;
146 unsigned long next;
147 unsigned long addr = (unsigned long) area->addr;
148 unsigned long end = addr + area->size - PAGE_SIZE;
149 int err;
150
151 BUG_ON(addr >= end);
152 pgd = pgd_offset_k(addr);
153 do {
154 next = pgd_addr_end(addr, end);
155 err = vmap_pud_range(pgd, addr, next, prot, pages);
156 if (err)
157 break;
158 } while (pgd++, addr = next, addr != end);
159 flush_cache_vmap((unsigned long) area->addr, end);
160 return err;
161}
162
163static struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long flags,
164 unsigned long start, unsigned long end,
165 int node, gfp_t gfp_mask)
166{
167 struct vm_struct **p, *tmp, *area;
168 unsigned long align = 1;
169 unsigned long addr;
170
171 BUG_ON(in_interrupt());
172 if (flags & VM_IOREMAP) {
173 int bit = fls(size);
174
175 if (bit > IOREMAP_MAX_ORDER)
176 bit = IOREMAP_MAX_ORDER;
177 else if (bit < PAGE_SHIFT)
178 bit = PAGE_SHIFT;
179
180 align = 1ul << bit;
181 }
182 addr = ALIGN(start, align);
183 size = PAGE_ALIGN(size);
184 if (unlikely(!size))
185 return NULL;
186
187 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_LEVEL_MASK, node);
188 if (unlikely(!area))
189 return NULL;
190
191 /*
192 * We always allocate a guard page.
193 */
194 size += PAGE_SIZE;
195
196 write_lock(&vmlist_lock);
197 for (p = &vmlist; (tmp = *p) != NULL ;p = &tmp->next) {
198 if ((unsigned long)tmp->addr < addr) {
199 if((unsigned long)tmp->addr + tmp->size >= addr)
200 addr = ALIGN(tmp->size +
201 (unsigned long)tmp->addr, align);
202 continue;
203 }
204 if ((size + addr) < addr)
205 goto out;
206 if (size + addr <= (unsigned long)tmp->addr)
207 goto found;
208 addr = ALIGN(tmp->size + (unsigned long)tmp->addr, align);
209 if (addr > end - size)
210 goto out;
211 }
212
213found:
214 area->next = *p;
215 *p = area;
216
217 area->flags = flags;
218 area->addr = (void *)addr;
219 area->size = size;
220 area->pages = NULL;
221 area->nr_pages = 0;
222 area->phys_addr = 0;
223 write_unlock(&vmlist_lock);
224
225 return area;
226
227out:
228 write_unlock(&vmlist_lock);
229 kfree(area);
230 if (printk_ratelimit())
231 printk(KERN_WARNING "allocation failed: out of vmalloc space - use vmalloc=<size> to increase size.\n");
232 return NULL;
233}
234
235struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
236 unsigned long start, unsigned long end)
237{
238 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL);
239}
240
241/**
242 * get_vm_area - reserve a contingous kernel virtual area
243 * @size: size of the area
244 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
245 *
246 * Search an area of @size in the kernel virtual mapping area,
247 * and reserved it for out purposes. Returns the area descriptor
248 * on success or %NULL on failure.
249 */
250struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
251{
252 return __get_vm_area(size, flags, VMALLOC_START, VMALLOC_END);
253}
254
255struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
256 int node, gfp_t gfp_mask)
257{
258 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
259 gfp_mask);
260}
261
262/* Caller must hold vmlist_lock */
263static struct vm_struct *__find_vm_area(void *addr)
264{
265 struct vm_struct *tmp;
266
267 for (tmp = vmlist; tmp != NULL; tmp = tmp->next) {
268 if (tmp->addr == addr)
269 break;
270 }
271
272 return tmp;
273}
274
275/* Caller must hold vmlist_lock */
276static struct vm_struct *__remove_vm_area(void *addr)
277{
278 struct vm_struct **p, *tmp;
279
280 for (p = &vmlist ; (tmp = *p) != NULL ;p = &tmp->next) {
281 if (tmp->addr == addr)
282 goto found;
283 }
284 return NULL;
285
286found:
287 unmap_vm_area(tmp);
288 *p = tmp->next;
289
290 /*
291 * Remove the guard page.
292 */
293 tmp->size -= PAGE_SIZE;
294 return tmp;
295}
296
297/**
298 * remove_vm_area - find and remove a contingous kernel virtual area
299 * @addr: base address
300 *
301 * Search for the kernel VM area starting at @addr, and remove it.
302 * This function returns the found VM area, but using it is NOT safe
303 * on SMP machines, except for its size or flags.
304 */
305struct vm_struct *remove_vm_area(void *addr)
306{
307 struct vm_struct *v;
308 write_lock(&vmlist_lock);
309 v = __remove_vm_area(addr);
310 write_unlock(&vmlist_lock);
311 return v;
312}
313
314void __vunmap(void *addr, int deallocate_pages)
315{
316 struct vm_struct *area;
317
318 if (!addr)
319 return;
320
321 if ((PAGE_SIZE-1) & (unsigned long)addr) {
322 printk(KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
323 WARN_ON(1);
324 return;
325 }
326
327 area = remove_vm_area(addr);
328 if (unlikely(!area)) {
329 printk(KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
330 addr);
331 WARN_ON(1);
332 return;
333 }
334
335 debug_check_no_locks_freed(addr, area->size);
336
337 if (deallocate_pages) {
338 int i;
339
340 for (i = 0; i < area->nr_pages; i++) {
341 BUG_ON(!area->pages[i]);
342 __free_page(area->pages[i]);
343 }
344
345 if (area->flags & VM_VPAGES)
346 vfree(area->pages);
347 else
348 kfree(area->pages);
349 }
350
351 kfree(area);
352 return;
353}
354
355/**
356 * vfree - release memory allocated by vmalloc()
357 * @addr: memory base address
358 *
359 * Free the virtually contiguous memory area starting at @addr, as
360 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
361 * NULL, no operation is performed.
362 *
363 * Must not be called in interrupt context.
364 */
365void vfree(void *addr)
366{
367 BUG_ON(in_interrupt());
368 __vunmap(addr, 1);
369}
370EXPORT_SYMBOL(vfree);
371
372/**
373 * vunmap - release virtual mapping obtained by vmap()
374 * @addr: memory base address
375 *
376 * Free the virtually contiguous memory area starting at @addr,
377 * which was created from the page array passed to vmap().
378 *
379 * Must not be called in interrupt context.
380 */
381void vunmap(void *addr)
382{
383 BUG_ON(in_interrupt());
384 __vunmap(addr, 0);
385}
386EXPORT_SYMBOL(vunmap);
387
388/**
389 * vmap - map an array of pages into virtually contiguous space
390 * @pages: array of page pointers
391 * @count: number of pages to map
392 * @flags: vm_area->flags
393 * @prot: page protection for the mapping
394 *
395 * Maps @count pages from @pages into contiguous kernel virtual
396 * space.
397 */
398void *vmap(struct page **pages, unsigned int count,
399 unsigned long flags, pgprot_t prot)
400{
401 struct vm_struct *area;
402
403 if (count > num_physpages)
404 return NULL;
405
406 area = get_vm_area((count << PAGE_SHIFT), flags);
407 if (!area)
408 return NULL;
409 if (map_vm_area(area, prot, &pages)) {
410 vunmap(area->addr);
411 return NULL;
412 }
413
414 return area->addr;
415}
416EXPORT_SYMBOL(vmap);
417
418void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
419 pgprot_t prot, int node)
420{
421 struct page **pages;
422 unsigned int nr_pages, array_size, i;
423
424 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
425 array_size = (nr_pages * sizeof(struct page *));
426
427 area->nr_pages = nr_pages;
428 /* Please note that the recursion is strictly bounded. */
429 if (array_size > PAGE_SIZE) {
430 pages = __vmalloc_node(array_size, gfp_mask, PAGE_KERNEL, node);
431 area->flags |= VM_VPAGES;
432 } else {
433 pages = kmalloc_node(array_size,
434 (gfp_mask & ~(__GFP_HIGHMEM | __GFP_ZERO)),
435 node);
436 }
437 area->pages = pages;
438 if (!area->pages) {
439 remove_vm_area(area->addr);
440 kfree(area);
441 return NULL;
442 }
443 memset(area->pages, 0, array_size);
444
445 for (i = 0; i < area->nr_pages; i++) {
446 if (node < 0)
447 area->pages[i] = alloc_page(gfp_mask);
448 else
449 area->pages[i] = alloc_pages_node(node, gfp_mask, 0);
450 if (unlikely(!area->pages[i])) {
451 /* Successfully allocated i pages, free them in __vunmap() */
452 area->nr_pages = i;
453 goto fail;
454 }
455 }
456
457 if (map_vm_area(area, prot, &pages))
458 goto fail;
459 return area->addr;
460
461fail:
462 vfree(area->addr);
463 return NULL;
464}
465
466void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
467{
468 return __vmalloc_area_node(area, gfp_mask, prot, -1);
469}
470
471/**
472 * __vmalloc_node - allocate virtually contiguous memory
473 * @size: allocation size
474 * @gfp_mask: flags for the page level allocator
475 * @prot: protection mask for the allocated pages
476 * @node: node to use for allocation or -1
477 *
478 * Allocate enough pages to cover @size from the page level
479 * allocator with @gfp_mask flags. Map them into contiguous
480 * kernel virtual space, using a pagetable protection of @prot.
481 */
482static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
483 int node)
484{
485 struct vm_struct *area;
486
487 size = PAGE_ALIGN(size);
488 if (!size || (size >> PAGE_SHIFT) > num_physpages)
489 return NULL;
490
491 area = get_vm_area_node(size, VM_ALLOC, node, gfp_mask);
492 if (!area)
493 return NULL;
494
495 return __vmalloc_area_node(area, gfp_mask, prot, node);
496}
497
498void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
499{
500 return __vmalloc_node(size, gfp_mask, prot, -1);
501}
502EXPORT_SYMBOL(__vmalloc);
503
504/**
505 * vmalloc - allocate virtually contiguous memory
506 * @size: allocation size
507 * Allocate enough pages to cover @size from the page level
508 * allocator and map them into contiguous kernel virtual space.
509 *
510 * For tight control over page level allocator and protection flags
511 * use __vmalloc() instead.
512 */
513void *vmalloc(unsigned long size)
514{
515 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
516}
517EXPORT_SYMBOL(vmalloc);
518
519/**
520 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
521 * @size: allocation size
522 *
523 * The resulting memory area is zeroed so it can be mapped to userspace
524 * without leaking data.
525 */
526void *vmalloc_user(unsigned long size)
527{
528 struct vm_struct *area;
529 void *ret;
530
531 ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
532 if (ret) {
533 write_lock(&vmlist_lock);
534 area = __find_vm_area(ret);
535 area->flags |= VM_USERMAP;
536 write_unlock(&vmlist_lock);
537 }
538 return ret;
539}
540EXPORT_SYMBOL(vmalloc_user);
541
542/**
543 * vmalloc_node - allocate memory on a specific node
544 * @size: allocation size
545 * @node: numa node
546 *
547 * Allocate enough pages to cover @size from the page level
548 * allocator and map them into contiguous kernel virtual space.
549 *
550 * For tight control over page level allocator and protection flags
551 * use __vmalloc() instead.
552 */
553void *vmalloc_node(unsigned long size, int node)
554{
555 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL, node);
556}
557EXPORT_SYMBOL(vmalloc_node);
558
559#ifndef PAGE_KERNEL_EXEC
560# define PAGE_KERNEL_EXEC PAGE_KERNEL
561#endif
562
563/**
564 * vmalloc_exec - allocate virtually contiguous, executable memory
565 * @size: allocation size
566 *
567 * Kernel-internal function to allocate enough pages to cover @size
568 * the page level allocator and map them into contiguous and
569 * executable kernel virtual space.
570 *
571 * For tight control over page level allocator and protection flags
572 * use __vmalloc() instead.
573 */
574
575void *vmalloc_exec(unsigned long size)
576{
577 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
578}
579
580/**
581 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
582 * @size: allocation size
583 *
584 * Allocate enough 32bit PA addressable pages to cover @size from the
585 * page level allocator and map them into contiguous kernel virtual space.
586 */
587void *vmalloc_32(unsigned long size)
588{
589 return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
590}
591EXPORT_SYMBOL(vmalloc_32);
592
593/**
594 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
595 * @size: allocation size
596 *
597 * The resulting memory area is 32bit addressable and zeroed so it can be
598 * mapped to userspace without leaking data.
599 */
600void *vmalloc_32_user(unsigned long size)
601{
602 struct vm_struct *area;
603 void *ret;
604
605 ret = __vmalloc(size, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
606 if (ret) {
607 write_lock(&vmlist_lock);
608 area = __find_vm_area(ret);
609 area->flags |= VM_USERMAP;
610 write_unlock(&vmlist_lock);
611 }
612 return ret;
613}
614EXPORT_SYMBOL(vmalloc_32_user);
615
616long vread(char *buf, char *addr, unsigned long count)
617{
618 struct vm_struct *tmp;
619 char *vaddr, *buf_start = buf;
620 unsigned long n;
621
622 /* Don't allow overflow */
623 if ((unsigned long) addr + count < count)
624 count = -(unsigned long) addr;
625
626 read_lock(&vmlist_lock);
627 for (tmp = vmlist; tmp; tmp = tmp->next) {
628 vaddr = (char *) tmp->addr;
629 if (addr >= vaddr + tmp->size - PAGE_SIZE)
630 continue;
631 while (addr < vaddr) {
632 if (count == 0)
633 goto finished;
634 *buf = '\0';
635 buf++;
636 addr++;
637 count--;
638 }
639 n = vaddr + tmp->size - PAGE_SIZE - addr;
640 do {
641 if (count == 0)
642 goto finished;
643 *buf = *addr;
644 buf++;
645 addr++;
646 count--;
647 } while (--n > 0);
648 }
649finished:
650 read_unlock(&vmlist_lock);
651 return buf - buf_start;
652}
653
654long vwrite(char *buf, char *addr, unsigned long count)
655{
656 struct vm_struct *tmp;
657 char *vaddr, *buf_start = buf;
658 unsigned long n;
659
660 /* Don't allow overflow */
661 if ((unsigned long) addr + count < count)
662 count = -(unsigned long) addr;
663
664 read_lock(&vmlist_lock);
665 for (tmp = vmlist; tmp; tmp = tmp->next) {
666 vaddr = (char *) tmp->addr;
667 if (addr >= vaddr + tmp->size - PAGE_SIZE)
668 continue;
669 while (addr < vaddr) {
670 if (count == 0)
671 goto finished;
672 buf++;
673 addr++;
674 count--;
675 }
676 n = vaddr + tmp->size - PAGE_SIZE - addr;
677 do {
678 if (count == 0)
679 goto finished;
680 *addr = *buf;
681 buf++;
682 addr++;
683 count--;
684 } while (--n > 0);
685 }
686finished:
687 read_unlock(&vmlist_lock);
688 return buf - buf_start;
689}
690
691/**
692 * remap_vmalloc_range - map vmalloc pages to userspace
693 * @vma: vma to cover (map full range of vma)
694 * @addr: vmalloc memory
695 * @pgoff: number of pages into addr before first page to map
696 * @returns: 0 for success, -Exxx on failure
697 *
698 * This function checks that addr is a valid vmalloc'ed area, and
699 * that it is big enough to cover the vma. Will return failure if
700 * that criteria isn't met.
701 *
702 * Similar to remap_pfn_range() (see mm/memory.c)
703 */
704int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
705 unsigned long pgoff)
706{
707 struct vm_struct *area;
708 unsigned long uaddr = vma->vm_start;
709 unsigned long usize = vma->vm_end - vma->vm_start;
710 int ret;
711
712 if ((PAGE_SIZE-1) & (unsigned long)addr)
713 return -EINVAL;
714
715 read_lock(&vmlist_lock);
716 area = __find_vm_area(addr);
717 if (!area)
718 goto out_einval_locked;
719
720 if (!(area->flags & VM_USERMAP))
721 goto out_einval_locked;
722
723 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
724 goto out_einval_locked;
725 read_unlock(&vmlist_lock);
726
727 addr += pgoff << PAGE_SHIFT;
728 do {
729 struct page *page = vmalloc_to_page(addr);
730 ret = vm_insert_page(vma, uaddr, page);
731 if (ret)
732 return ret;
733
734 uaddr += PAGE_SIZE;
735 addr += PAGE_SIZE;
736 usize -= PAGE_SIZE;
737 } while (usize > 0);
738
739 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
740 vma->vm_flags |= VM_RESERVED;
741
742 return ret;
743
744out_einval_locked:
745 read_unlock(&vmlist_lock);
746 return -EINVAL;
747}
748EXPORT_SYMBOL(remap_vmalloc_range);
749