blob: 0f551a4a44cddc7a042d47bbcd85c7126569ee69 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Christoph Lameter930fc452005-10-29 18:15:41 -07008 * Numa awareness, Christoph Lameter, SGI, June 2005
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 */
10
Nick Piggindb64fe02008-10-18 20:27:03 -070011#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
Alexey Dobriyand43c36d2009-10-07 17:09:06 +040015#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
Alexey Dobriyan5f6a6a92008-10-06 03:50:47 +040019#include <linux/proc_fs.h>
Christoph Lametera10aa572008-04-28 02:12:40 -070020#include <linux/seq_file.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070021#include <linux/debugobjects.h>
Christoph Lameter23016962008-04-28 02:12:42 -070022#include <linux/kallsyms.h>
Nick Piggindb64fe02008-10-18 20:27:03 -070023#include <linux/list.h>
24#include <linux/rbtree.h>
25#include <linux/radix-tree.h>
26#include <linux/rcupdate.h>
Tejun Heof0aa6612009-02-20 16:29:08 +090027#include <linux/pfn.h>
Catalin Marinas89219d32009-06-11 13:23:19 +010028#include <linux/kmemleak.h>
Nick Piggindb64fe02008-10-18 20:27:03 -070029#include <asm/atomic.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <asm/uaccess.h>
31#include <asm/tlbflush.h>
David Miller2dca6992009-09-21 12:22:34 -070032#include <asm/shmparam.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
34
Nick Piggindb64fe02008-10-18 20:27:03 -070035/*** Page table manipulation functions ***/
Adrian Bunkb2213852006-09-25 23:31:02 -070036
Linus Torvalds1da177e2005-04-16 15:20:36 -070037static 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
Nick Piggindb64fe02008-10-18 20:27:03 -070048static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -070049{
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
Nick Piggindb64fe02008-10-18 20:27:03 -070062static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -070063{
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
Nick Piggindb64fe02008-10-18 20:27:03 -070076static void vunmap_page_range(unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -070077{
78 pgd_t *pgd;
79 unsigned long next;
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
81 BUG_ON(addr >= end);
82 pgd = pgd_offset_k(addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -070083 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);
Linus Torvalds1da177e2005-04-16 15:20:36 -070089}
90
91static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
Nick Piggindb64fe02008-10-18 20:27:03 -070092 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070093{
94 pte_t *pte;
95
Nick Piggindb64fe02008-10-18 20:27:03 -070096 /*
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
Hugh Dickins872fec12005-10-29 18:16:21 -0700101 pte = pte_alloc_kernel(pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102 if (!pte)
103 return -ENOMEM;
104 do {
Nick Piggindb64fe02008-10-18 20:27:03 -0700105 struct page *page = pages[*nr];
106
107 if (WARN_ON(!pte_none(*pte)))
108 return -EBUSY;
109 if (WARN_ON(!page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 return -ENOMEM;
111 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
Nick Piggindb64fe02008-10-18 20:27:03 -0700112 (*nr)++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 } while (pte++, addr += PAGE_SIZE, addr != end);
114 return 0;
115}
116
Nick Piggindb64fe02008-10-18 20:27:03 -0700117static int vmap_pmd_range(pud_t *pud, unsigned long addr,
118 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119{
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);
Nick Piggindb64fe02008-10-18 20:27:03 -0700128 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 return -ENOMEM;
130 } while (pmd++, addr = next, addr != end);
131 return 0;
132}
133
Nick Piggindb64fe02008-10-18 20:27:03 -0700134static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
135 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136{
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);
Nick Piggindb64fe02008-10-18 20:27:03 -0700145 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 return -ENOMEM;
147 } while (pud++, addr = next, addr != end);
148 return 0;
149}
150
Nick Piggindb64fe02008-10-18 20:27:03 -0700151/*
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 */
Tejun Heo8fc48982009-02-20 16:29:08 +0900157static int vmap_page_range_noflush(unsigned long start, unsigned long end,
158 pgprot_t prot, struct page **pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
160 pgd_t *pgd;
161 unsigned long next;
Adam Lackorzynski2e4e27c2009-01-04 12:00:46 -0800162 unsigned long addr = start;
Nick Piggindb64fe02008-10-18 20:27:03 -0700163 int err = 0;
164 int nr = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
166 BUG_ON(addr >= end);
167 pgd = pgd_offset_k(addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 do {
169 next = pgd_addr_end(addr, end);
Nick Piggindb64fe02008-10-18 20:27:03 -0700170 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 if (err)
Figo.zhangbf88c8c2009-09-21 17:01:47 -0700172 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 } while (pgd++, addr = next, addr != end);
Nick Piggindb64fe02008-10-18 20:27:03 -0700174
Nick Piggindb64fe02008-10-18 20:27:03 -0700175 return nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176}
177
Tejun Heo8fc48982009-02-20 16:29:08 +0900178static 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
KAMEZAWA Hiroyuki81ac3ad2009-09-22 16:45:49 -0700188int is_vmalloc_or_module_addr(const void *x)
Linus Torvalds73bdf0a2008-10-15 08:35:12 -0700189{
190 /*
Russell Kingab4f2ee2008-11-06 17:11:07 +0000191 * ARM, x86-64 and sparc64 put modules in a special place,
Linus Torvalds73bdf0a2008-10-15 08:35:12 -0700192 * 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
Christoph Lameter48667e72008-02-04 22:28:31 -0800203/*
Nick Piggindb64fe02008-10-18 20:27:03 -0700204 * Walk a vmap address to the struct page it maps.
Christoph Lameter48667e72008-02-04 22:28:31 -0800205 */
Christoph Lameterb3bdda02008-02-04 22:28:32 -0800206struct page *vmalloc_to_page(const void *vmalloc_addr)
Christoph Lameter48667e72008-02-04 22:28:31 -0800207{
208 unsigned long addr = (unsigned long) vmalloc_addr;
209 struct page *page = NULL;
210 pgd_t *pgd = pgd_offset_k(addr);
Christoph Lameter48667e72008-02-04 22:28:31 -0800211
Ingo Molnar7aa413d2008-06-19 13:28:11 +0200212 /*
213 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
214 * architectures that do not vmalloc module space
215 */
Linus Torvalds73bdf0a2008-10-15 08:35:12 -0700216 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
Jiri Slaby59ea7462008-06-12 13:56:40 +0200217
Christoph Lameter48667e72008-02-04 22:28:31 -0800218 if (!pgd_none(*pgd)) {
Nick Piggindb64fe02008-10-18 20:27:03 -0700219 pud_t *pud = pud_offset(pgd, addr);
Christoph Lameter48667e72008-02-04 22:28:31 -0800220 if (!pud_none(*pud)) {
Nick Piggindb64fe02008-10-18 20:27:03 -0700221 pmd_t *pmd = pmd_offset(pud, addr);
Christoph Lameter48667e72008-02-04 22:28:31 -0800222 if (!pmd_none(*pmd)) {
Nick Piggindb64fe02008-10-18 20:27:03 -0700223 pte_t *ptep, pte;
224
Christoph Lameter48667e72008-02-04 22:28:31 -0800225 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}
235EXPORT_SYMBOL(vmalloc_to_page);
236
237/*
238 * Map a vmalloc()-space virtual address to the physical page frame number.
239 */
Christoph Lameterb3bdda02008-02-04 22:28:32 -0800240unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
Christoph Lameter48667e72008-02-04 22:28:31 -0800241{
242 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
243}
244EXPORT_SYMBOL(vmalloc_to_pfn);
245
Nick Piggindb64fe02008-10-18 20:27:03 -0700246
247/*** Global kva allocator ***/
248
249#define VM_LAZY_FREE 0x01
250#define VM_LAZY_FREEING 0x02
251#define VM_VM_AREA 0x04
252
253struct 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
264static DEFINE_SPINLOCK(vmap_area_lock);
265static struct rb_root vmap_area_root = RB_ROOT;
266static LIST_HEAD(vmap_area_list);
Tejun Heoca23e402009-08-14 15:00:52 +0900267static unsigned long vmap_area_pcpu_hole;
Nick Piggindb64fe02008-10-18 20:27:03 -0700268
269static struct vmap_area *__find_vmap_area(unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270{
Nick Piggindb64fe02008-10-18 20:27:03 -0700271 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
288static 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
320static 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 */
326static 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;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 unsigned long addr;
Nick Piggindb64fe02008-10-18 20:27:03 -0700334 int purged = 0;
335
Nick Piggin77669702009-02-27 14:03:03 -0800336 BUG_ON(!size);
Nick Piggindb64fe02008-10-18 20:27:03 -0700337 BUG_ON(size & ~PAGE_MASK);
338
Nick Piggindb64fe02008-10-18 20:27:03 -0700339 va = kmalloc_node(sizeof(struct vmap_area),
340 gfp_mask & GFP_RECLAIM_MASK, node);
341 if (unlikely(!va))
342 return ERR_PTR(-ENOMEM);
343
344retry:
Glauber Costa0ae15132008-11-19 15:36:33 -0800345 addr = ALIGN(vstart, align);
346
Nick Piggindb64fe02008-10-18 20:27:03 -0700347 spin_lock(&vmap_area_lock);
Nick Piggin77669702009-02-27 14:03:03 -0800348 if (addr + size - 1 < addr)
349 goto overflow;
350
Nick Piggindb64fe02008-10-18 20:27:03 -0700351 /* 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
Nick Pigginf011c2d2008-11-19 15:36:32 -0800380 while (addr + size > first->va_start && addr + size <= vend) {
Nick Piggindb64fe02008-10-18 20:27:03 -0700381 addr = ALIGN(first->va_end + PAGE_SIZE, align);
Nick Piggin77669702009-02-27 14:03:03 -0800382 if (addr + size - 1 < addr)
383 goto overflow;
Nick Piggindb64fe02008-10-18 20:27:03 -0700384
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 }
392found:
393 if (addr + size > vend) {
Nick Piggin77669702009-02-27 14:03:03 -0800394overflow:
Nick Piggindb64fe02008-10-18 20:27:03 -0700395 spin_unlock(&vmap_area_lock);
396 if (!purged) {
397 purge_vmap_area_lazy();
398 purged = 1;
399 goto retry;
400 }
401 if (printk_ratelimit())
Glauber Costac1279c42009-01-06 14:39:18 -0800402 printk(KERN_WARNING
403 "vmap allocation for size %lu failed: "
404 "use vmalloc=<size> to increase size.\n", size);
Ralph Wuerthner2498ce42009-05-06 16:02:59 -0700405 kfree(va);
Nick Piggindb64fe02008-10-18 20:27:03 -0700406 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
420static 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
427static 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
Tejun Heoca23e402009-08-14 15:00:52 +0900434 /*
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
Nick Piggindb64fe02008-10-18 20:27:03 -0700443 call_rcu(&va->rcu_head, rcu_free_va);
444}
445
446/*
447 * Free a region of KVA allocated by alloc_vmap_area
448 */
449static 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 */
459static void unmap_vmap_area(struct vmap_area *va)
460{
461 vunmap_page_range(va->va_start, va->va_end);
462}
463
Nick Piggincd528582009-01-06 14:39:20 -0800464static 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
Nick Piggindb64fe02008-10-18 20:27:03 -0700485/*
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 */
501static 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
510static 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 */
522static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
523 int sync, int force_flush)
524{
Andrew Morton46666d82009-01-15 13:51:15 -0800525 static DEFINE_SPINLOCK(purge_lock);
Nick Piggindb64fe02008-10-18 20:27:03 -0700526 LIST_HEAD(valist);
527 struct vmap_area *va;
Vegard Nossumcbb76672009-02-27 14:03:04 -0800528 struct vmap_area *n_va;
Nick Piggindb64fe02008-10-18 20:27:03 -0700529 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) {
Andrew Morton46666d82009-01-15 13:51:15 -0800537 if (!spin_trylock(&purge_lock))
Nick Piggindb64fe02008-10-18 20:27:03 -0700538 return;
539 } else
Andrew Morton46666d82009-01-15 13:51:15 -0800540 spin_lock(&purge_lock);
Nick Piggindb64fe02008-10-18 20:27:03 -0700541
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);
Vegard Nossumcbb76672009-02-27 14:03:04 -0800568 list_for_each_entry_safe(va, n_va, &valist, purge_list)
Nick Piggindb64fe02008-10-18 20:27:03 -0700569 __free_vmap_area(va);
570 spin_unlock(&vmap_area_lock);
571 }
Andrew Morton46666d82009-01-15 13:51:15 -0800572 spin_unlock(&purge_lock);
Nick Piggindb64fe02008-10-18 20:27:03 -0700573}
574
575/*
Nick Piggin496850e2008-11-19 15:36:33 -0800576 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
577 * is already purging.
578 */
579static 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/*
Nick Piggindb64fe02008-10-18 20:27:03 -0700587 * Kick off a purge of the outstanding lazy areas.
588 */
589static void purge_vmap_area_lazy(void)
590{
591 unsigned long start = ULONG_MAX, end = 0;
592
Nick Piggin496850e2008-11-19 15:36:33 -0800593 __purge_vmap_area_lazy(&start, &end, 1, 0);
Nick Piggindb64fe02008-10-18 20:27:03 -0700594}
595
596/*
Nick Pigginb29acbd2008-12-01 13:13:47 -0800597 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
598 * called for the correct range previously.
Nick Piggindb64fe02008-10-18 20:27:03 -0700599 */
Nick Pigginb29acbd2008-12-01 13:13:47 -0800600static void free_unmap_vmap_area_noflush(struct vmap_area *va)
Nick Piggindb64fe02008-10-18 20:27:03 -0700601{
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()))
Nick Piggin496850e2008-11-19 15:36:33 -0800605 try_purge_vmap_area_lazy();
Nick Piggindb64fe02008-10-18 20:27:03 -0700606}
607
Nick Pigginb29acbd2008-12-01 13:13:47 -0800608/*
609 * Free and unmap a vmap area
610 */
611static 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
Nick Piggindb64fe02008-10-18 20:27:03 -0700617static 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
628static 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
Jeremy Fitzhardinge9b463332008-10-28 19:22:34 +1100667static bool vmap_initialized __read_mostly = false;
668
Nick Piggindb64fe02008-10-18 20:27:03 -0700669struct vmap_block_queue {
670 spinlock_t lock;
671 struct list_head free;
672 struct list_head dirty;
673 unsigned int nr_dirty;
674};
675
676struct 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 {
MinChan Kimd0868172009-03-31 15:19:26 -0700684 struct list_head free_list;
Nick Piggindb64fe02008-10-18 20:27:03 -0700685 struct rcu_head rcu_head;
686 };
687};
688
689/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
690static 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 */
697static DEFINE_SPINLOCK(vmap_block_tree_lock);
698static 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
707static 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
714static 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);
Nick Piggindb64fe02008-10-18 20:27:03 -0700751
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_cpu_blocks);
765
766 return vb;
767}
768
769static 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
776static void free_vmap_block(struct vmap_block *vb)
777{
778 struct vmap_block *tmp;
779 unsigned long vb_idx;
780
MinChan Kimd0868172009-03-31 15:19:26 -0700781 BUG_ON(!list_empty(&vb->free_list));
Nick Piggindb64fe02008-10-18 20:27:03 -0700782
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
Nick Pigginb29acbd2008-12-01 13:13:47 -0800789 free_unmap_vmap_area_noflush(vb->va);
Nick Piggindb64fe02008-10-18 20:27:03 -0700790 call_rcu(&vb->rcu_head, rcu_free_vb);
791}
792
793static 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
804again:
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_cpu_blocks);
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
842static 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);
Nick Pigginb29acbd2008-12-01 13:13:47 -0800851
852 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
853
Nick Piggindb64fe02008-10-18 20:27:03 -0700854 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);
MinChan Kimd0868172009-03-31 15:19:26 -0700866
Nick Piggindb64fe02008-10-18 20:27:03 -0700867 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 */
889void vm_unmap_aliases(void)
890{
891 unsigned long start = ULONG_MAX, end = 0;
892 int cpu;
893 int flush = 0;
894
Jeremy Fitzhardinge9b463332008-10-28 19:22:34 +1100895 if (unlikely(!vmap_initialized))
896 return;
897
Nick Piggindb64fe02008-10-18 20:27:03 -0700898 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}
935EXPORT_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 */
942void 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);
Nick Piggincd528582009-01-06 14:39:20 -0800953 vmap_debug_free_range(addr, addr+size);
Nick Piggindb64fe02008-10-18 20:27:03 -0700954
955 if (likely(count <= VMAP_MAX_ALLOC))
956 vb_free(mem, size);
957 else
958 free_unmap_vmap_area_addr(addr);
959}
960EXPORT_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
Randy Dunlape99c97a2008-10-29 14:01:09 -0700968 *
969 * Returns: a pointer to the address that has been mapped, or %NULL on failure
Nick Piggindb64fe02008-10-18 20:27:03 -0700970 */
971void *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}
998EXPORT_SYMBOL(vm_map_ram);
999
Tejun Heof0aa6612009-02-20 16:29:08 +09001000/**
1001 * vm_area_register_early - register vmap area early during boot
1002 * @vm: vm_struct to register
Tejun Heoc0c0a292009-02-24 11:57:21 +09001003 * @align: requested alignment
Tejun Heof0aa6612009-02-20 16:29:08 +09001004 *
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 */
Tejun Heoc0c0a292009-02-24 11:57:21 +09001012void __init vm_area_register_early(struct vm_struct *vm, size_t align)
Tejun Heof0aa6612009-02-20 16:29:08 +09001013{
1014 static size_t vm_init_off __initdata;
Tejun Heoc0c0a292009-02-24 11:57:21 +09001015 unsigned long addr;
Tejun Heof0aa6612009-02-20 16:29:08 +09001016
Tejun Heoc0c0a292009-02-24 11:57:21 +09001017 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;
Tejun Heof0aa6612009-02-20 16:29:08 +09001021
1022 vm->next = vmlist;
1023 vmlist = vm;
1024}
1025
Nick Piggindb64fe02008-10-18 20:27:03 -07001026void __init vmalloc_init(void)
1027{
Ivan Kokshaysky822c18f2009-01-15 13:50:48 -08001028 struct vmap_area *va;
1029 struct vm_struct *tmp;
Nick Piggindb64fe02008-10-18 20:27:03 -07001030 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 }
Jeremy Fitzhardinge9b463332008-10-28 19:22:34 +11001041
Ivan Kokshaysky822c18f2009-01-15 13:50:48 -08001042 /* Import existing vmlist entries. */
1043 for (tmp = vmlist; tmp; tmp = tmp->next) {
Pekka Enberg43ebdac2009-05-25 15:01:35 +03001044 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
Ivan Kokshaysky822c18f2009-01-15 13:50:48 -08001045 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 }
Tejun Heoca23e402009-08-14 15:00:52 +09001050
1051 vmap_area_pcpu_hole = VMALLOC_END;
1052
Jeremy Fitzhardinge9b463332008-10-28 19:22:34 +11001053 vmap_initialized = true;
Nick Piggindb64fe02008-10-18 20:27:03 -07001054}
1055
Tejun Heo8fc48982009-02-20 16:29:08 +09001056/**
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 */
1075int 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 */
1095void 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 */
Nick Piggindb64fe02008-10-18 20:27:03 -07001108void unmap_kernel_range(unsigned long addr, unsigned long size)
1109{
1110 unsigned long end = addr + size;
Tejun Heof6fcba72009-02-20 15:38:48 -08001111
1112 flush_cache_vunmap(addr, end);
Nick Piggindb64fe02008-10-18 20:27:03 -07001113 vunmap_page_range(addr, end);
1114 flush_tlb_kernel_range(addr, end);
1115}
1116
1117int 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}
1131EXPORT_SYMBOL_GPL(map_vm_area);
1132
1133/*** Old vmalloc interfaces ***/
1134DEFINE_RWLOCK(vmlist_lock);
1135struct vm_struct *vmlist;
1136
Tejun Heocf88c792009-08-14 15:00:52 +09001137static 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
Nick Piggindb64fe02008-10-18 20:27:03 -07001159static struct vm_struct *__get_vm_area_node(unsigned long size,
David Miller2dca6992009-09-21 12:22:34 -07001160 unsigned long align, unsigned long flags, unsigned long start,
1161 unsigned long end, int node, gfp_t gfp_mask, void *caller)
Nick Piggindb64fe02008-10-18 20:27:03 -07001162{
1163 static struct vmap_area *va;
1164 struct vm_struct *area;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
Giridhar Pemmasani52fd24c2006-10-28 10:38:34 -07001166 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 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 }
Nick Piggindb64fe02008-10-18 20:27:03 -07001177
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 size = PAGE_ALIGN(size);
OGAWA Hirofumi31be8302006-11-16 01:19:29 -08001179 if (unlikely(!size))
1180 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Tejun Heocf88c792009-08-14 15:00:52 +09001182 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 if (unlikely(!area))
1184 return NULL;
1185
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 /*
1187 * We always allocate a guard page.
1188 */
1189 size += PAGE_SIZE;
1190
Nick Piggindb64fe02008-10-18 20:27:03 -07001191 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1192 if (IS_ERR(va)) {
1193 kfree(area);
1194 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196
Tejun Heocf88c792009-08-14 15:00:52 +09001197 insert_vmalloc_vm(area, va, flags, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 return area;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199}
1200
Christoph Lameter930fc452005-10-29 18:15:41 -07001201struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1202 unsigned long start, unsigned long end)
1203{
David Miller2dca6992009-09-21 12:22:34 -07001204 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
Christoph Lameter23016962008-04-28 02:12:42 -07001205 __builtin_return_address(0));
Christoph Lameter930fc452005-10-29 18:15:41 -07001206}
Rusty Russell5992b6d2007-07-19 01:49:21 -07001207EXPORT_SYMBOL_GPL(__get_vm_area);
Christoph Lameter930fc452005-10-29 18:15:41 -07001208
Benjamin Herrenschmidtc2968612009-02-18 14:48:12 -08001209struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1210 unsigned long start, unsigned long end,
1211 void *caller)
1212{
David Miller2dca6992009-09-21 12:22:34 -07001213 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
Benjamin Herrenschmidtc2968612009-02-18 14:48:12 -08001214 caller);
1215}
1216
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217/**
Simon Arlott183ff222007-10-20 01:27:18 +02001218 * get_vm_area - reserve a contiguous kernel virtual area
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 * @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 */
1226struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1227{
David Miller2dca6992009-09-21 12:22:34 -07001228 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
Christoph Lameter23016962008-04-28 02:12:42 -07001229 -1, GFP_KERNEL, __builtin_return_address(0));
1230}
1231
1232struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1233 void *caller)
1234{
David Miller2dca6992009-09-21 12:22:34 -07001235 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
Christoph Lameter23016962008-04-28 02:12:42 -07001236 -1, GFP_KERNEL, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237}
1238
Giridhar Pemmasani52fd24c2006-10-28 10:38:34 -07001239struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1240 int node, gfp_t gfp_mask)
Christoph Lameter930fc452005-10-29 18:15:41 -07001241{
David Miller2dca6992009-09-21 12:22:34 -07001242 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1243 node, gfp_mask, __builtin_return_address(0));
Christoph Lameter930fc452005-10-29 18:15:41 -07001244}
1245
Nick Piggindb64fe02008-10-18 20:27:03 -07001246static struct vm_struct *find_vm_area(const void *addr)
Nick Piggin83342312006-06-23 02:03:20 -07001247{
Nick Piggindb64fe02008-10-18 20:27:03 -07001248 struct vmap_area *va;
Nick Piggin83342312006-06-23 02:03:20 -07001249
Nick Piggindb64fe02008-10-18 20:27:03 -07001250 va = find_vmap_area((unsigned long)addr);
1251 if (va && va->flags & VM_VM_AREA)
1252 return va->private;
Nick Piggin83342312006-06-23 02:03:20 -07001253
Andi Kleen7856dfe2005-05-20 14:27:57 -07001254 return NULL;
Andi Kleen7856dfe2005-05-20 14:27:57 -07001255}
1256
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257/**
Simon Arlott183ff222007-10-20 01:27:18 +02001258 * remove_vm_area - find and remove a continuous kernel virtual area
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 * @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
Andi Kleen7856dfe2005-05-20 14:27:57 -07001263 * on SMP machines, except for its size or flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 */
Christoph Lameterb3bdda02008-02-04 22:28:32 -08001265struct vm_struct *remove_vm_area(const void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266{
Nick Piggindb64fe02008-10-18 20:27:03 -07001267 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;
KAMEZAWA Hiroyukidd32c272009-09-21 17:02:32 -07001273 /*
1274 * remove from list and disallow access to this vm_struct
1275 * before unmap. (address range confliction is maintained by
1276 * vmap.)
1277 */
Nick Piggindb64fe02008-10-18 20:27:03 -07001278 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
KAMEZAWA Hiroyukidd32c272009-09-21 17:02:32 -07001284 vmap_debug_free_range(va->va_start, va->va_end);
1285 free_unmap_vmap_area(va);
1286 vm->size -= PAGE_SIZE;
1287
Nick Piggindb64fe02008-10-18 20:27:03 -07001288 return vm;
1289 }
1290 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291}
1292
Christoph Lameterb3bdda02008-02-04 22:28:32 -08001293static void __vunmap(const void *addr, int deallocate_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294{
1295 struct vm_struct *area;
1296
1297 if (!addr)
1298 return;
1299
1300 if ((PAGE_SIZE-1) & (unsigned long)addr) {
Arjan van de Ven4c8573e2008-07-25 19:45:37 -07001301 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 return;
1303 }
1304
1305 area = remove_vm_area(addr);
1306 if (unlikely(!area)) {
Arjan van de Ven4c8573e2008-07-25 19:45:37 -07001307 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 return;
1310 }
1311
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07001312 debug_check_no_locks_freed(addr, area->size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001313 debug_check_no_obj_freed(addr, area->size);
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07001314
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 if (deallocate_pages) {
1316 int i;
1317
1318 for (i = 0; i < area->nr_pages; i++) {
Christoph Lameterbf53d6f2008-02-04 22:28:34 -08001319 struct page *page = area->pages[i];
1320
1321 BUG_ON(!page);
1322 __free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 }
1324
Jan Kiszka8757d5f2006-07-14 00:23:56 -07001325 if (area->flags & VM_VPAGES)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 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()
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 * @addr: memory base address
1338 *
Simon Arlott183ff222007-10-20 01:27:18 +02001339 * Free the virtually continuous memory area starting at @addr, as
Pekka Enberg80e93ef2005-09-09 13:10:16 -07001340 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1341 * NULL, no operation is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07001343 * Must not be called in interrupt context.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 */
Christoph Lameterb3bdda02008-02-04 22:28:32 -08001345void vfree(const void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346{
1347 BUG_ON(in_interrupt());
Catalin Marinas89219d32009-06-11 13:23:19 +01001348
1349 kmemleak_free(addr);
1350
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 __vunmap(addr, 1);
1352}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353EXPORT_SYMBOL(vfree);
1354
1355/**
1356 * vunmap - release virtual mapping obtained by vmap()
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 * @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 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07001362 * Must not be called in interrupt context.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 */
Christoph Lameterb3bdda02008-02-04 22:28:32 -08001364void vunmap(const void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365{
1366 BUG_ON(in_interrupt());
Peter Zijlstra34754b62009-02-25 16:04:03 +01001367 might_sleep();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 __vunmap(addr, 0);
1369}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370EXPORT_SYMBOL(vunmap);
1371
1372/**
1373 * vmap - map an array of pages into virtually contiguous space
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 * @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 */
1382void *vmap(struct page **pages, unsigned int count,
1383 unsigned long flags, pgprot_t prot)
1384{
1385 struct vm_struct *area;
1386
Peter Zijlstra34754b62009-02-25 16:04:03 +01001387 might_sleep();
1388
Jan Beulich44813742009-09-21 17:03:05 -07001389 if (count > totalram_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 return NULL;
1391
Christoph Lameter23016962008-04-28 02:12:42 -07001392 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1393 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 if (!area)
1395 return NULL;
Christoph Lameter23016962008-04-28 02:12:42 -07001396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 if (map_vm_area(area, prot, &pages)) {
1398 vunmap(area->addr);
1399 return NULL;
1400 }
1401
1402 return area->addr;
1403}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404EXPORT_SYMBOL(vmap);
1405
David Miller2dca6992009-09-21 12:22:34 -07001406static void *__vmalloc_node(unsigned long size, unsigned long align,
1407 gfp_t gfp_mask, pgprot_t prot,
Nick Piggindb64fe02008-10-18 20:27:03 -07001408 int node, void *caller);
Adrian Bunke31d9eb2008-02-04 22:29:09 -08001409static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
Christoph Lameter23016962008-04-28 02:12:42 -07001410 pgprot_t prot, int node, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411{
1412 struct page **pages;
1413 unsigned int nr_pages, array_size, i;
1414
1415 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1416 array_size = (nr_pages * sizeof(struct page *));
1417
1418 area->nr_pages = nr_pages;
1419 /* Please note that the recursion is strictly bounded. */
Jan Kiszka8757d5f2006-07-14 00:23:56 -07001420 if (array_size > PAGE_SIZE) {
David Miller2dca6992009-09-21 12:22:34 -07001421 pages = __vmalloc_node(array_size, 1, gfp_mask | __GFP_ZERO,
Christoph Lameter23016962008-04-28 02:12:42 -07001422 PAGE_KERNEL, node, caller);
Jan Kiszka8757d5f2006-07-14 00:23:56 -07001423 area->flags |= VM_VPAGES;
Andrew Morton286e1ea2006-10-17 00:09:57 -07001424 } else {
1425 pages = kmalloc_node(array_size,
Christoph Lameter6cb06222007-10-16 01:25:41 -07001426 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
Andrew Morton286e1ea2006-10-17 00:09:57 -07001427 node);
1428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429 area->pages = pages;
Christoph Lameter23016962008-04-28 02:12:42 -07001430 area->caller = caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 if (!area->pages) {
1432 remove_vm_area(area->addr);
1433 kfree(area);
1434 return NULL;
1435 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436
1437 for (i = 0; i < area->nr_pages; i++) {
Christoph Lameterbf53d6f2008-02-04 22:28:34 -08001438 struct page *page;
1439
Christoph Lameter930fc452005-10-29 18:15:41 -07001440 if (node < 0)
Christoph Lameterbf53d6f2008-02-04 22:28:34 -08001441 page = alloc_page(gfp_mask);
Christoph Lameter930fc452005-10-29 18:15:41 -07001442 else
Christoph Lameterbf53d6f2008-02-04 22:28:34 -08001443 page = alloc_pages_node(node, gfp_mask, 0);
1444
1445 if (unlikely(!page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 /* Successfully allocated i pages, free them in __vunmap() */
1447 area->nr_pages = i;
1448 goto fail;
1449 }
Christoph Lameterbf53d6f2008-02-04 22:28:34 -08001450 area->pages[i] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 }
1452
1453 if (map_vm_area(area, prot, &pages))
1454 goto fail;
1455 return area->addr;
1456
1457fail:
1458 vfree(area->addr);
1459 return NULL;
1460}
1461
Christoph Lameter930fc452005-10-29 18:15:41 -07001462void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1463{
Catalin Marinas89219d32009-06-11 13:23:19 +01001464 void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1465 __builtin_return_address(0));
1466
1467 /*
1468 * A ref_count = 3 is needed because the vm_struct and vmap_area
1469 * structures allocated in the __get_vm_area_node() function contain
1470 * references to the virtual address of the vmalloc'ed block.
1471 */
1472 kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1473
1474 return addr;
Christoph Lameter930fc452005-10-29 18:15:41 -07001475}
1476
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477/**
Christoph Lameter930fc452005-10-29 18:15:41 -07001478 * __vmalloc_node - allocate virtually contiguous memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 * @size: allocation size
David Miller2dca6992009-09-21 12:22:34 -07001480 * @align: desired alignment
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 * @gfp_mask: flags for the page level allocator
1482 * @prot: protection mask for the allocated pages
Randy Dunlapd44e0782005-11-07 01:01:10 -08001483 * @node: node to use for allocation or -1
Randy Dunlapc85d1942008-05-01 04:34:48 -07001484 * @caller: caller's return address
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 *
1486 * Allocate enough pages to cover @size from the page level
1487 * allocator with @gfp_mask flags. Map them into contiguous
1488 * kernel virtual space, using a pagetable protection of @prot.
1489 */
David Miller2dca6992009-09-21 12:22:34 -07001490static void *__vmalloc_node(unsigned long size, unsigned long align,
1491 gfp_t gfp_mask, pgprot_t prot,
1492 int node, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493{
1494 struct vm_struct *area;
Catalin Marinas89219d32009-06-11 13:23:19 +01001495 void *addr;
1496 unsigned long real_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
1498 size = PAGE_ALIGN(size);
Jan Beulich44813742009-09-21 17:03:05 -07001499 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 return NULL;
1501
David Miller2dca6992009-09-21 12:22:34 -07001502 area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
1503 VMALLOC_END, node, gfp_mask, caller);
Christoph Lameter23016962008-04-28 02:12:42 -07001504
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 if (!area)
1506 return NULL;
1507
Catalin Marinas89219d32009-06-11 13:23:19 +01001508 addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1509
1510 /*
1511 * A ref_count = 3 is needed because the vm_struct and vmap_area
1512 * structures allocated in the __get_vm_area_node() function contain
1513 * references to the virtual address of the vmalloc'ed block.
1514 */
1515 kmemleak_alloc(addr, real_size, 3, gfp_mask);
1516
1517 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518}
1519
Christoph Lameter930fc452005-10-29 18:15:41 -07001520void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1521{
David Miller2dca6992009-09-21 12:22:34 -07001522 return __vmalloc_node(size, 1, gfp_mask, prot, -1,
Christoph Lameter23016962008-04-28 02:12:42 -07001523 __builtin_return_address(0));
Christoph Lameter930fc452005-10-29 18:15:41 -07001524}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525EXPORT_SYMBOL(__vmalloc);
1526
1527/**
1528 * vmalloc - allocate virtually contiguous memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 * @size: allocation size
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 * Allocate enough pages to cover @size from the page level
1531 * allocator and map them into contiguous kernel virtual space.
1532 *
Michael Opdenackerc1c88972006-10-03 23:21:02 +02001533 * For tight control over page level allocator and protection flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 * use __vmalloc() instead.
1535 */
1536void *vmalloc(unsigned long size)
1537{
David Miller2dca6992009-09-21 12:22:34 -07001538 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
Christoph Lameter23016962008-04-28 02:12:42 -07001539 -1, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541EXPORT_SYMBOL(vmalloc);
1542
Christoph Lameter930fc452005-10-29 18:15:41 -07001543/**
Rolf Eike Beeread04082006-09-27 01:50:13 -07001544 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1545 * @size: allocation size
Nick Piggin83342312006-06-23 02:03:20 -07001546 *
Rolf Eike Beeread04082006-09-27 01:50:13 -07001547 * The resulting memory area is zeroed so it can be mapped to userspace
1548 * without leaking data.
Nick Piggin83342312006-06-23 02:03:20 -07001549 */
1550void *vmalloc_user(unsigned long size)
1551{
1552 struct vm_struct *area;
1553 void *ret;
1554
David Miller2dca6992009-09-21 12:22:34 -07001555 ret = __vmalloc_node(size, SHMLBA,
1556 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
Glauber Costa84877842009-01-06 14:39:19 -08001557 PAGE_KERNEL, -1, __builtin_return_address(0));
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001558 if (ret) {
Nick Piggindb64fe02008-10-18 20:27:03 -07001559 area = find_vm_area(ret);
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001560 area->flags |= VM_USERMAP;
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001561 }
Nick Piggin83342312006-06-23 02:03:20 -07001562 return ret;
1563}
1564EXPORT_SYMBOL(vmalloc_user);
1565
1566/**
Christoph Lameter930fc452005-10-29 18:15:41 -07001567 * vmalloc_node - allocate memory on a specific node
Christoph Lameter930fc452005-10-29 18:15:41 -07001568 * @size: allocation size
Randy Dunlapd44e0782005-11-07 01:01:10 -08001569 * @node: numa node
Christoph Lameter930fc452005-10-29 18:15:41 -07001570 *
1571 * Allocate enough pages to cover @size from the page level
1572 * allocator and map them into contiguous kernel virtual space.
1573 *
Michael Opdenackerc1c88972006-10-03 23:21:02 +02001574 * For tight control over page level allocator and protection flags
Christoph Lameter930fc452005-10-29 18:15:41 -07001575 * use __vmalloc() instead.
1576 */
1577void *vmalloc_node(unsigned long size, int node)
1578{
David Miller2dca6992009-09-21 12:22:34 -07001579 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
Christoph Lameter23016962008-04-28 02:12:42 -07001580 node, __builtin_return_address(0));
Christoph Lameter930fc452005-10-29 18:15:41 -07001581}
1582EXPORT_SYMBOL(vmalloc_node);
1583
Pavel Pisa4dc3b162005-05-01 08:59:25 -07001584#ifndef PAGE_KERNEL_EXEC
1585# define PAGE_KERNEL_EXEC PAGE_KERNEL
1586#endif
1587
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588/**
1589 * vmalloc_exec - allocate virtually contiguous, executable memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 * @size: allocation size
1591 *
1592 * Kernel-internal function to allocate enough pages to cover @size
1593 * the page level allocator and map them into contiguous and
1594 * executable kernel virtual space.
1595 *
Michael Opdenackerc1c88972006-10-03 23:21:02 +02001596 * For tight control over page level allocator and protection flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 * use __vmalloc() instead.
1598 */
1599
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600void *vmalloc_exec(unsigned long size)
1601{
David Miller2dca6992009-09-21 12:22:34 -07001602 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
Glauber Costa84877842009-01-06 14:39:19 -08001603 -1, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604}
1605
Andi Kleen0d08e0d2007-05-02 19:27:12 +02001606#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
Benjamin Herrenschmidt7ac674f2007-07-19 01:49:10 -07001607#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
Andi Kleen0d08e0d2007-05-02 19:27:12 +02001608#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
Benjamin Herrenschmidt7ac674f2007-07-19 01:49:10 -07001609#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
Andi Kleen0d08e0d2007-05-02 19:27:12 +02001610#else
1611#define GFP_VMALLOC32 GFP_KERNEL
1612#endif
1613
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614/**
1615 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 * @size: allocation size
1617 *
1618 * Allocate enough 32bit PA addressable pages to cover @size from the
1619 * page level allocator and map them into contiguous kernel virtual space.
1620 */
1621void *vmalloc_32(unsigned long size)
1622{
David Miller2dca6992009-09-21 12:22:34 -07001623 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
Glauber Costa84877842009-01-06 14:39:19 -08001624 -1, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626EXPORT_SYMBOL(vmalloc_32);
1627
Nick Piggin83342312006-06-23 02:03:20 -07001628/**
Rolf Eike Beeread04082006-09-27 01:50:13 -07001629 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
Nick Piggin83342312006-06-23 02:03:20 -07001630 * @size: allocation size
Rolf Eike Beeread04082006-09-27 01:50:13 -07001631 *
1632 * The resulting memory area is 32bit addressable and zeroed so it can be
1633 * mapped to userspace without leaking data.
Nick Piggin83342312006-06-23 02:03:20 -07001634 */
1635void *vmalloc_32_user(unsigned long size)
1636{
1637 struct vm_struct *area;
1638 void *ret;
1639
David Miller2dca6992009-09-21 12:22:34 -07001640 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
Glauber Costa84877842009-01-06 14:39:19 -08001641 -1, __builtin_return_address(0));
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001642 if (ret) {
Nick Piggindb64fe02008-10-18 20:27:03 -07001643 area = find_vm_area(ret);
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001644 area->flags |= VM_USERMAP;
Eric Dumazet2b4ac442006-11-10 12:27:48 -08001645 }
Nick Piggin83342312006-06-23 02:03:20 -07001646 return ret;
1647}
1648EXPORT_SYMBOL(vmalloc_32_user);
1649
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001650/*
1651 * small helper routine , copy contents to buf from addr.
1652 * If the page is not present, fill zero.
1653 */
1654
1655static int aligned_vread(char *buf, char *addr, unsigned long count)
1656{
1657 struct page *p;
1658 int copied = 0;
1659
1660 while (count) {
1661 unsigned long offset, length;
1662
1663 offset = (unsigned long)addr & ~PAGE_MASK;
1664 length = PAGE_SIZE - offset;
1665 if (length > count)
1666 length = count;
1667 p = vmalloc_to_page(addr);
1668 /*
1669 * To do safe access to this _mapped_ area, we need
1670 * lock. But adding lock here means that we need to add
1671 * overhead of vmalloc()/vfree() calles for this _debug_
1672 * interface, rarely used. Instead of that, we'll use
1673 * kmap() and get small overhead in this access function.
1674 */
1675 if (p) {
1676 /*
1677 * we can expect USER0 is not used (see vread/vwrite's
1678 * function description)
1679 */
1680 void *map = kmap_atomic(p, KM_USER0);
1681 memcpy(buf, map + offset, length);
1682 kunmap_atomic(map, KM_USER0);
1683 } else
1684 memset(buf, 0, length);
1685
1686 addr += length;
1687 buf += length;
1688 copied += length;
1689 count -= length;
1690 }
1691 return copied;
1692}
1693
1694static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1695{
1696 struct page *p;
1697 int copied = 0;
1698
1699 while (count) {
1700 unsigned long offset, length;
1701
1702 offset = (unsigned long)addr & ~PAGE_MASK;
1703 length = PAGE_SIZE - offset;
1704 if (length > count)
1705 length = count;
1706 p = vmalloc_to_page(addr);
1707 /*
1708 * To do safe access to this _mapped_ area, we need
1709 * lock. But adding lock here means that we need to add
1710 * overhead of vmalloc()/vfree() calles for this _debug_
1711 * interface, rarely used. Instead of that, we'll use
1712 * kmap() and get small overhead in this access function.
1713 */
1714 if (p) {
1715 /*
1716 * we can expect USER0 is not used (see vread/vwrite's
1717 * function description)
1718 */
1719 void *map = kmap_atomic(p, KM_USER0);
1720 memcpy(map + offset, buf, length);
1721 kunmap_atomic(map, KM_USER0);
1722 }
1723 addr += length;
1724 buf += length;
1725 copied += length;
1726 count -= length;
1727 }
1728 return copied;
1729}
1730
1731/**
1732 * vread() - read vmalloc area in a safe way.
1733 * @buf: buffer for reading data
1734 * @addr: vm address.
1735 * @count: number of bytes to be read.
1736 *
1737 * Returns # of bytes which addr and buf should be increased.
1738 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1739 * includes any intersect with alive vmalloc area.
1740 *
1741 * This function checks that addr is a valid vmalloc'ed area, and
1742 * copy data from that area to a given buffer. If the given memory range
1743 * of [addr...addr+count) includes some valid address, data is copied to
1744 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1745 * IOREMAP area is treated as memory hole and no copy is done.
1746 *
1747 * If [addr...addr+count) doesn't includes any intersects with alive
1748 * vm_struct area, returns 0.
1749 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1750 * the caller should guarantee KM_USER0 is not used.
1751 *
1752 * Note: In usual ops, vread() is never necessary because the caller
1753 * should know vmalloc() area is valid and can use memcpy().
1754 * This is for routines which have to access vmalloc area without
1755 * any informaion, as /dev/kmem.
1756 *
1757 */
1758
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759long vread(char *buf, char *addr, unsigned long count)
1760{
1761 struct vm_struct *tmp;
1762 char *vaddr, *buf_start = buf;
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001763 unsigned long buflen = count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 unsigned long n;
1765
1766 /* Don't allow overflow */
1767 if ((unsigned long) addr + count < count)
1768 count = -(unsigned long) addr;
1769
1770 read_lock(&vmlist_lock);
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001771 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 vaddr = (char *) tmp->addr;
1773 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1774 continue;
1775 while (addr < vaddr) {
1776 if (count == 0)
1777 goto finished;
1778 *buf = '\0';
1779 buf++;
1780 addr++;
1781 count--;
1782 }
1783 n = vaddr + tmp->size - PAGE_SIZE - addr;
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001784 if (n > count)
1785 n = count;
1786 if (!(tmp->flags & VM_IOREMAP))
1787 aligned_vread(buf, addr, n);
1788 else /* IOREMAP area is treated as memory hole */
1789 memset(buf, 0, n);
1790 buf += n;
1791 addr += n;
1792 count -= n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 }
1794finished:
1795 read_unlock(&vmlist_lock);
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001796
1797 if (buf == buf_start)
1798 return 0;
1799 /* zero-fill memory holes */
1800 if (buf != buf_start + buflen)
1801 memset(buf, 0, buflen - (buf - buf_start));
1802
1803 return buflen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804}
1805
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001806/**
1807 * vwrite() - write vmalloc area in a safe way.
1808 * @buf: buffer for source data
1809 * @addr: vm address.
1810 * @count: number of bytes to be read.
1811 *
1812 * Returns # of bytes which addr and buf should be incresed.
1813 * (same number to @count).
1814 * If [addr...addr+count) doesn't includes any intersect with valid
1815 * vmalloc area, returns 0.
1816 *
1817 * This function checks that addr is a valid vmalloc'ed area, and
1818 * copy data from a buffer to the given addr. If specified range of
1819 * [addr...addr+count) includes some valid address, data is copied from
1820 * proper area of @buf. If there are memory holes, no copy to hole.
1821 * IOREMAP area is treated as memory hole and no copy is done.
1822 *
1823 * If [addr...addr+count) doesn't includes any intersects with alive
1824 * vm_struct area, returns 0.
1825 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1826 * the caller should guarantee KM_USER0 is not used.
1827 *
1828 * Note: In usual ops, vwrite() is never necessary because the caller
1829 * should know vmalloc() area is valid and can use memcpy().
1830 * This is for routines which have to access vmalloc area without
1831 * any informaion, as /dev/kmem.
1832 *
1833 * The caller should guarantee KM_USER1 is not used.
1834 */
1835
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836long vwrite(char *buf, char *addr, unsigned long count)
1837{
1838 struct vm_struct *tmp;
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001839 char *vaddr;
1840 unsigned long n, buflen;
1841 int copied = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842
1843 /* Don't allow overflow */
1844 if ((unsigned long) addr + count < count)
1845 count = -(unsigned long) addr;
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001846 buflen = count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
1848 read_lock(&vmlist_lock);
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001849 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 vaddr = (char *) tmp->addr;
1851 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1852 continue;
1853 while (addr < vaddr) {
1854 if (count == 0)
1855 goto finished;
1856 buf++;
1857 addr++;
1858 count--;
1859 }
1860 n = vaddr + tmp->size - PAGE_SIZE - addr;
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001861 if (n > count)
1862 n = count;
1863 if (!(tmp->flags & VM_IOREMAP)) {
1864 aligned_vwrite(buf, addr, n);
1865 copied++;
1866 }
1867 buf += n;
1868 addr += n;
1869 count -= n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 }
1871finished:
1872 read_unlock(&vmlist_lock);
KAMEZAWA Hiroyukid0107eb2009-09-21 17:02:34 -07001873 if (!copied)
1874 return 0;
1875 return buflen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876}
Nick Piggin83342312006-06-23 02:03:20 -07001877
1878/**
1879 * remap_vmalloc_range - map vmalloc pages to userspace
Nick Piggin83342312006-06-23 02:03:20 -07001880 * @vma: vma to cover (map full range of vma)
1881 * @addr: vmalloc memory
1882 * @pgoff: number of pages into addr before first page to map
Randy Dunlap76824862008-03-19 17:00:40 -07001883 *
1884 * Returns: 0 for success, -Exxx on failure
Nick Piggin83342312006-06-23 02:03:20 -07001885 *
1886 * This function checks that addr is a valid vmalloc'ed area, and
1887 * that it is big enough to cover the vma. Will return failure if
1888 * that criteria isn't met.
1889 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08001890 * Similar to remap_pfn_range() (see mm/memory.c)
Nick Piggin83342312006-06-23 02:03:20 -07001891 */
1892int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1893 unsigned long pgoff)
1894{
1895 struct vm_struct *area;
1896 unsigned long uaddr = vma->vm_start;
1897 unsigned long usize = vma->vm_end - vma->vm_start;
Nick Piggin83342312006-06-23 02:03:20 -07001898
1899 if ((PAGE_SIZE-1) & (unsigned long)addr)
1900 return -EINVAL;
1901
Nick Piggindb64fe02008-10-18 20:27:03 -07001902 area = find_vm_area(addr);
Nick Piggin83342312006-06-23 02:03:20 -07001903 if (!area)
Nick Piggindb64fe02008-10-18 20:27:03 -07001904 return -EINVAL;
Nick Piggin83342312006-06-23 02:03:20 -07001905
1906 if (!(area->flags & VM_USERMAP))
Nick Piggindb64fe02008-10-18 20:27:03 -07001907 return -EINVAL;
Nick Piggin83342312006-06-23 02:03:20 -07001908
1909 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
Nick Piggindb64fe02008-10-18 20:27:03 -07001910 return -EINVAL;
Nick Piggin83342312006-06-23 02:03:20 -07001911
1912 addr += pgoff << PAGE_SHIFT;
1913 do {
1914 struct page *page = vmalloc_to_page(addr);
Nick Piggindb64fe02008-10-18 20:27:03 -07001915 int ret;
1916
Nick Piggin83342312006-06-23 02:03:20 -07001917 ret = vm_insert_page(vma, uaddr, page);
1918 if (ret)
1919 return ret;
1920
1921 uaddr += PAGE_SIZE;
1922 addr += PAGE_SIZE;
1923 usize -= PAGE_SIZE;
1924 } while (usize > 0);
1925
1926 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1927 vma->vm_flags |= VM_RESERVED;
1928
Nick Piggindb64fe02008-10-18 20:27:03 -07001929 return 0;
Nick Piggin83342312006-06-23 02:03:20 -07001930}
1931EXPORT_SYMBOL(remap_vmalloc_range);
1932
Christoph Hellwig1eeb66a2007-05-08 00:27:03 -07001933/*
1934 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1935 * have one.
1936 */
1937void __attribute__((weak)) vmalloc_sync_all(void)
1938{
1939}
Jeremy Fitzhardinge5f4352f2007-07-17 18:37:04 -07001940
1941
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001942static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
Jeremy Fitzhardinge5f4352f2007-07-17 18:37:04 -07001943{
1944 /* apply_to_page_range() does all the hard work. */
1945 return 0;
1946}
1947
1948/**
1949 * alloc_vm_area - allocate a range of kernel address space
1950 * @size: size of the area
Randy Dunlap76824862008-03-19 17:00:40 -07001951 *
1952 * Returns: NULL on failure, vm_struct on success
Jeremy Fitzhardinge5f4352f2007-07-17 18:37:04 -07001953 *
1954 * This function reserves a range of kernel address space, and
1955 * allocates pagetables to map that range. No actual mappings
1956 * are created. If the kernel address space is not shared
1957 * between processes, it syncs the pagetable across all
1958 * processes.
1959 */
1960struct vm_struct *alloc_vm_area(size_t size)
1961{
1962 struct vm_struct *area;
1963
Christoph Lameter23016962008-04-28 02:12:42 -07001964 area = get_vm_area_caller(size, VM_IOREMAP,
1965 __builtin_return_address(0));
Jeremy Fitzhardinge5f4352f2007-07-17 18:37:04 -07001966 if (area == NULL)
1967 return NULL;
1968
1969 /*
1970 * This ensures that page tables are constructed for this region
1971 * of kernel virtual address space and mapped into init_mm.
1972 */
1973 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1974 area->size, f, NULL)) {
1975 free_vm_area(area);
1976 return NULL;
1977 }
1978
1979 /* Make sure the pagetables are constructed in process kernel
1980 mappings */
1981 vmalloc_sync_all();
1982
1983 return area;
1984}
1985EXPORT_SYMBOL_GPL(alloc_vm_area);
1986
1987void free_vm_area(struct vm_struct *area)
1988{
1989 struct vm_struct *ret;
1990 ret = remove_vm_area(area->addr);
1991 BUG_ON(ret != area);
1992 kfree(area);
1993}
1994EXPORT_SYMBOL_GPL(free_vm_area);
Christoph Lametera10aa572008-04-28 02:12:40 -07001995
Tejun Heoca23e402009-08-14 15:00:52 +09001996static struct vmap_area *node_to_va(struct rb_node *n)
1997{
1998 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
1999}
2000
2001/**
2002 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2003 * @end: target address
2004 * @pnext: out arg for the next vmap_area
2005 * @pprev: out arg for the previous vmap_area
2006 *
2007 * Returns: %true if either or both of next and prev are found,
2008 * %false if no vmap_area exists
2009 *
2010 * Find vmap_areas end addresses of which enclose @end. ie. if not
2011 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2012 */
2013static bool pvm_find_next_prev(unsigned long end,
2014 struct vmap_area **pnext,
2015 struct vmap_area **pprev)
2016{
2017 struct rb_node *n = vmap_area_root.rb_node;
2018 struct vmap_area *va = NULL;
2019
2020 while (n) {
2021 va = rb_entry(n, struct vmap_area, rb_node);
2022 if (end < va->va_end)
2023 n = n->rb_left;
2024 else if (end > va->va_end)
2025 n = n->rb_right;
2026 else
2027 break;
2028 }
2029
2030 if (!va)
2031 return false;
2032
2033 if (va->va_end > end) {
2034 *pnext = va;
2035 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2036 } else {
2037 *pprev = va;
2038 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2039 }
2040 return true;
2041}
2042
2043/**
2044 * pvm_determine_end - find the highest aligned address between two vmap_areas
2045 * @pnext: in/out arg for the next vmap_area
2046 * @pprev: in/out arg for the previous vmap_area
2047 * @align: alignment
2048 *
2049 * Returns: determined end address
2050 *
2051 * Find the highest aligned address between *@pnext and *@pprev below
2052 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2053 * down address is between the end addresses of the two vmap_areas.
2054 *
2055 * Please note that the address returned by this function may fall
2056 * inside *@pnext vmap_area. The caller is responsible for checking
2057 * that.
2058 */
2059static unsigned long pvm_determine_end(struct vmap_area **pnext,
2060 struct vmap_area **pprev,
2061 unsigned long align)
2062{
2063 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2064 unsigned long addr;
2065
2066 if (*pnext)
2067 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2068 else
2069 addr = vmalloc_end;
2070
2071 while (*pprev && (*pprev)->va_end > addr) {
2072 *pnext = *pprev;
2073 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2074 }
2075
2076 return addr;
2077}
2078
2079/**
2080 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2081 * @offsets: array containing offset of each area
2082 * @sizes: array containing size of each area
2083 * @nr_vms: the number of areas to allocate
2084 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2085 * @gfp_mask: allocation mask
2086 *
2087 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2088 * vm_structs on success, %NULL on failure
2089 *
2090 * Percpu allocator wants to use congruent vm areas so that it can
2091 * maintain the offsets among percpu areas. This function allocates
2092 * congruent vmalloc areas for it. These areas tend to be scattered
2093 * pretty far, distance between two areas easily going up to
2094 * gigabytes. To avoid interacting with regular vmallocs, these areas
2095 * are allocated from top.
2096 *
2097 * Despite its complicated look, this allocator is rather simple. It
2098 * does everything top-down and scans areas from the end looking for
2099 * matching slot. While scanning, if any of the areas overlaps with
2100 * existing vmap_area, the base address is pulled down to fit the
2101 * area. Scanning is repeated till all the areas fit and then all
2102 * necessary data structres are inserted and the result is returned.
2103 */
2104struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2105 const size_t *sizes, int nr_vms,
2106 size_t align, gfp_t gfp_mask)
2107{
2108 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2109 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2110 struct vmap_area **vas, *prev, *next;
2111 struct vm_struct **vms;
2112 int area, area2, last_area, term_area;
2113 unsigned long base, start, end, last_end;
2114 bool purged = false;
2115
2116 gfp_mask &= GFP_RECLAIM_MASK;
2117
2118 /* verify parameters and allocate data structures */
2119 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2120 for (last_area = 0, area = 0; area < nr_vms; area++) {
2121 start = offsets[area];
2122 end = start + sizes[area];
2123
2124 /* is everything aligned properly? */
2125 BUG_ON(!IS_ALIGNED(offsets[area], align));
2126 BUG_ON(!IS_ALIGNED(sizes[area], align));
2127
2128 /* detect the area with the highest address */
2129 if (start > offsets[last_area])
2130 last_area = area;
2131
2132 for (area2 = 0; area2 < nr_vms; area2++) {
2133 unsigned long start2 = offsets[area2];
2134 unsigned long end2 = start2 + sizes[area2];
2135
2136 if (area2 == area)
2137 continue;
2138
2139 BUG_ON(start2 >= start && start2 < end);
2140 BUG_ON(end2 <= end && end2 > start);
2141 }
2142 }
2143 last_end = offsets[last_area] + sizes[last_area];
2144
2145 if (vmalloc_end - vmalloc_start < last_end) {
2146 WARN_ON(true);
2147 return NULL;
2148 }
2149
2150 vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2151 vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2152 if (!vas || !vms)
2153 goto err_free;
2154
2155 for (area = 0; area < nr_vms; area++) {
2156 vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2157 vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2158 if (!vas[area] || !vms[area])
2159 goto err_free;
2160 }
2161retry:
2162 spin_lock(&vmap_area_lock);
2163
2164 /* start scanning - we scan from the top, begin with the last area */
2165 area = term_area = last_area;
2166 start = offsets[area];
2167 end = start + sizes[area];
2168
2169 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2170 base = vmalloc_end - last_end;
2171 goto found;
2172 }
2173 base = pvm_determine_end(&next, &prev, align) - end;
2174
2175 while (true) {
2176 BUG_ON(next && next->va_end <= base + end);
2177 BUG_ON(prev && prev->va_end > base + end);
2178
2179 /*
2180 * base might have underflowed, add last_end before
2181 * comparing.
2182 */
2183 if (base + last_end < vmalloc_start + last_end) {
2184 spin_unlock(&vmap_area_lock);
2185 if (!purged) {
2186 purge_vmap_area_lazy();
2187 purged = true;
2188 goto retry;
2189 }
2190 goto err_free;
2191 }
2192
2193 /*
2194 * If next overlaps, move base downwards so that it's
2195 * right below next and then recheck.
2196 */
2197 if (next && next->va_start < base + end) {
2198 base = pvm_determine_end(&next, &prev, align) - end;
2199 term_area = area;
2200 continue;
2201 }
2202
2203 /*
2204 * If prev overlaps, shift down next and prev and move
2205 * base so that it's right below new next and then
2206 * recheck.
2207 */
2208 if (prev && prev->va_end > base + start) {
2209 next = prev;
2210 prev = node_to_va(rb_prev(&next->rb_node));
2211 base = pvm_determine_end(&next, &prev, align) - end;
2212 term_area = area;
2213 continue;
2214 }
2215
2216 /*
2217 * This area fits, move on to the previous one. If
2218 * the previous one is the terminal one, we're done.
2219 */
2220 area = (area + nr_vms - 1) % nr_vms;
2221 if (area == term_area)
2222 break;
2223 start = offsets[area];
2224 end = start + sizes[area];
2225 pvm_find_next_prev(base + end, &next, &prev);
2226 }
2227found:
2228 /* we've found a fitting base, insert all va's */
2229 for (area = 0; area < nr_vms; area++) {
2230 struct vmap_area *va = vas[area];
2231
2232 va->va_start = base + offsets[area];
2233 va->va_end = va->va_start + sizes[area];
2234 __insert_vmap_area(va);
2235 }
2236
2237 vmap_area_pcpu_hole = base + offsets[last_area];
2238
2239 spin_unlock(&vmap_area_lock);
2240
2241 /* insert all vm's */
2242 for (area = 0; area < nr_vms; area++)
2243 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2244 pcpu_get_vm_areas);
2245
2246 kfree(vas);
2247 return vms;
2248
2249err_free:
2250 for (area = 0; area < nr_vms; area++) {
2251 if (vas)
2252 kfree(vas[area]);
2253 if (vms)
2254 kfree(vms[area]);
2255 }
2256 kfree(vas);
2257 kfree(vms);
2258 return NULL;
2259}
2260
2261/**
2262 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2263 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2264 * @nr_vms: the number of allocated areas
2265 *
2266 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2267 */
2268void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2269{
2270 int i;
2271
2272 for (i = 0; i < nr_vms; i++)
2273 free_vm_area(vms[i]);
2274 kfree(vms);
2275}
Christoph Lametera10aa572008-04-28 02:12:40 -07002276
2277#ifdef CONFIG_PROC_FS
2278static void *s_start(struct seq_file *m, loff_t *pos)
2279{
2280 loff_t n = *pos;
2281 struct vm_struct *v;
2282
2283 read_lock(&vmlist_lock);
2284 v = vmlist;
2285 while (n > 0 && v) {
2286 n--;
2287 v = v->next;
2288 }
2289 if (!n)
2290 return v;
2291
2292 return NULL;
2293
2294}
2295
2296static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2297{
2298 struct vm_struct *v = p;
2299
2300 ++*pos;
2301 return v->next;
2302}
2303
2304static void s_stop(struct seq_file *m, void *p)
2305{
2306 read_unlock(&vmlist_lock);
2307}
2308
Eric Dumazeta47a1262008-07-23 21:27:38 -07002309static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2310{
2311 if (NUMA_BUILD) {
2312 unsigned int nr, *counters = m->private;
2313
2314 if (!counters)
2315 return;
2316
2317 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2318
2319 for (nr = 0; nr < v->nr_pages; nr++)
2320 counters[page_to_nid(v->pages[nr])]++;
2321
2322 for_each_node_state(nr, N_HIGH_MEMORY)
2323 if (counters[nr])
2324 seq_printf(m, " N%u=%u", nr, counters[nr]);
2325 }
2326}
2327
Christoph Lametera10aa572008-04-28 02:12:40 -07002328static int s_show(struct seq_file *m, void *p)
2329{
2330 struct vm_struct *v = p;
2331
2332 seq_printf(m, "0x%p-0x%p %7ld",
2333 v->addr, v->addr + v->size, v->size);
2334
Christoph Lameter23016962008-04-28 02:12:42 -07002335 if (v->caller) {
Hugh Dickins9c246242008-12-09 13:14:27 -08002336 char buff[KSYM_SYMBOL_LEN];
Christoph Lameter23016962008-04-28 02:12:42 -07002337
2338 seq_putc(m, ' ');
2339 sprint_symbol(buff, (unsigned long)v->caller);
2340 seq_puts(m, buff);
2341 }
2342
Christoph Lametera10aa572008-04-28 02:12:40 -07002343 if (v->nr_pages)
2344 seq_printf(m, " pages=%d", v->nr_pages);
2345
2346 if (v->phys_addr)
2347 seq_printf(m, " phys=%lx", v->phys_addr);
2348
2349 if (v->flags & VM_IOREMAP)
2350 seq_printf(m, " ioremap");
2351
2352 if (v->flags & VM_ALLOC)
2353 seq_printf(m, " vmalloc");
2354
2355 if (v->flags & VM_MAP)
2356 seq_printf(m, " vmap");
2357
2358 if (v->flags & VM_USERMAP)
2359 seq_printf(m, " user");
2360
2361 if (v->flags & VM_VPAGES)
2362 seq_printf(m, " vpages");
2363
Eric Dumazeta47a1262008-07-23 21:27:38 -07002364 show_numa_info(m, v);
Christoph Lametera10aa572008-04-28 02:12:40 -07002365 seq_putc(m, '\n');
2366 return 0;
2367}
2368
Alexey Dobriyan5f6a6a92008-10-06 03:50:47 +04002369static const struct seq_operations vmalloc_op = {
Christoph Lametera10aa572008-04-28 02:12:40 -07002370 .start = s_start,
2371 .next = s_next,
2372 .stop = s_stop,
2373 .show = s_show,
2374};
Alexey Dobriyan5f6a6a92008-10-06 03:50:47 +04002375
2376static int vmalloc_open(struct inode *inode, struct file *file)
2377{
2378 unsigned int *ptr = NULL;
2379 int ret;
2380
2381 if (NUMA_BUILD)
2382 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2383 ret = seq_open(file, &vmalloc_op);
2384 if (!ret) {
2385 struct seq_file *m = file->private_data;
2386 m->private = ptr;
2387 } else
2388 kfree(ptr);
2389 return ret;
2390}
2391
2392static const struct file_operations proc_vmalloc_operations = {
2393 .open = vmalloc_open,
2394 .read = seq_read,
2395 .llseek = seq_lseek,
2396 .release = seq_release_private,
2397};
2398
2399static int __init proc_vmalloc_init(void)
2400{
2401 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2402 return 0;
2403}
2404module_init(proc_vmalloc_init);
Christoph Lametera10aa572008-04-28 02:12:40 -07002405#endif
2406