]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0-only | |
2 | /* | |
3 | * linux/mm/vmalloc.c | |
4 | * | |
5 | * Copyright (C) 1993 Linus Torvalds | |
6 | * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 | |
7 | * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000 | |
8 | * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002 | |
9 | * Numa awareness, Christoph Lameter, SGI, June 2005 | |
10 | */ | |
11 | ||
12 | #include <linux/vmalloc.h> | |
13 | #include <linux/mm.h> | |
14 | #include <linux/module.h> | |
15 | #include <linux/highmem.h> | |
16 | #include <linux/sched/signal.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/spinlock.h> | |
19 | #include <linux/interrupt.h> | |
20 | #include <linux/proc_fs.h> | |
21 | #include <linux/seq_file.h> | |
22 | #include <linux/set_memory.h> | |
23 | #include <linux/debugobjects.h> | |
24 | #include <linux/kallsyms.h> | |
25 | #include <linux/list.h> | |
26 | #include <linux/notifier.h> | |
27 | #include <linux/rbtree.h> | |
28 | #include <linux/radix-tree.h> | |
29 | #include <linux/rcupdate.h> | |
30 | #include <linux/pfn.h> | |
31 | #include <linux/kmemleak.h> | |
32 | #include <linux/atomic.h> | |
33 | #include <linux/compiler.h> | |
34 | #include <linux/llist.h> | |
35 | #include <linux/bitops.h> | |
36 | #include <linux/rbtree_augmented.h> | |
37 | #include <linux/overflow.h> | |
38 | ||
39 | #include <linux/uaccess.h> | |
40 | #include <asm/tlbflush.h> | |
41 | #include <asm/shmparam.h> | |
42 | ||
43 | #include "internal.h" | |
44 | #include "pgalloc-track.h" | |
45 | ||
46 | bool is_vmalloc_addr(const void *x) | |
47 | { | |
48 | unsigned long addr = (unsigned long)x; | |
49 | ||
50 | return addr >= VMALLOC_START && addr < VMALLOC_END; | |
51 | } | |
52 | EXPORT_SYMBOL(is_vmalloc_addr); | |
53 | ||
54 | struct vfree_deferred { | |
55 | struct llist_head list; | |
56 | struct work_struct wq; | |
57 | }; | |
58 | static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred); | |
59 | ||
60 | static void __vunmap(const void *, int); | |
61 | ||
62 | static void free_work(struct work_struct *w) | |
63 | { | |
64 | struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq); | |
65 | struct llist_node *t, *llnode; | |
66 | ||
67 | llist_for_each_safe(llnode, t, llist_del_all(&p->list)) | |
68 | __vunmap((void *)llnode, 1); | |
69 | } | |
70 | ||
71 | /*** Page table manipulation functions ***/ | |
72 | ||
73 | static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end, | |
74 | pgtbl_mod_mask *mask) | |
75 | { | |
76 | pte_t *pte; | |
77 | ||
78 | pte = pte_offset_kernel(pmd, addr); | |
79 | do { | |
80 | pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte); | |
81 | WARN_ON(!pte_none(ptent) && !pte_present(ptent)); | |
82 | } while (pte++, addr += PAGE_SIZE, addr != end); | |
83 | *mask |= PGTBL_PTE_MODIFIED; | |
84 | } | |
85 | ||
86 | static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end, | |
87 | pgtbl_mod_mask *mask) | |
88 | { | |
89 | pmd_t *pmd; | |
90 | unsigned long next; | |
91 | int cleared; | |
92 | ||
93 | pmd = pmd_offset(pud, addr); | |
94 | do { | |
95 | next = pmd_addr_end(addr, end); | |
96 | ||
97 | cleared = pmd_clear_huge(pmd); | |
98 | if (cleared || pmd_bad(*pmd)) | |
99 | *mask |= PGTBL_PMD_MODIFIED; | |
100 | ||
101 | if (cleared) | |
102 | continue; | |
103 | if (pmd_none_or_clear_bad(pmd)) | |
104 | continue; | |
105 | vunmap_pte_range(pmd, addr, next, mask); | |
106 | } while (pmd++, addr = next, addr != end); | |
107 | } | |
108 | ||
109 | static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end, | |
110 | pgtbl_mod_mask *mask) | |
111 | { | |
112 | pud_t *pud; | |
113 | unsigned long next; | |
114 | int cleared; | |
115 | ||
116 | pud = pud_offset(p4d, addr); | |
117 | do { | |
118 | next = pud_addr_end(addr, end); | |
119 | ||
120 | cleared = pud_clear_huge(pud); | |
121 | if (cleared || pud_bad(*pud)) | |
122 | *mask |= PGTBL_PUD_MODIFIED; | |
123 | ||
124 | if (cleared) | |
125 | continue; | |
126 | if (pud_none_or_clear_bad(pud)) | |
127 | continue; | |
128 | vunmap_pmd_range(pud, addr, next, mask); | |
129 | } while (pud++, addr = next, addr != end); | |
130 | } | |
131 | ||
132 | static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end, | |
133 | pgtbl_mod_mask *mask) | |
134 | { | |
135 | p4d_t *p4d; | |
136 | unsigned long next; | |
137 | int cleared; | |
138 | ||
139 | p4d = p4d_offset(pgd, addr); | |
140 | do { | |
141 | next = p4d_addr_end(addr, end); | |
142 | ||
143 | cleared = p4d_clear_huge(p4d); | |
144 | if (cleared || p4d_bad(*p4d)) | |
145 | *mask |= PGTBL_P4D_MODIFIED; | |
146 | ||
147 | if (cleared) | |
148 | continue; | |
149 | if (p4d_none_or_clear_bad(p4d)) | |
150 | continue; | |
151 | vunmap_pud_range(p4d, addr, next, mask); | |
152 | } while (p4d++, addr = next, addr != end); | |
153 | } | |
154 | ||
155 | /** | |
156 | * unmap_kernel_range_noflush - unmap kernel VM area | |
157 | * @start: start of the VM area to unmap | |
158 | * @size: size of the VM area to unmap | |
159 | * | |
160 | * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify | |
161 | * should have been allocated using get_vm_area() and its friends. | |
162 | * | |
163 | * NOTE: | |
164 | * This function does NOT do any cache flushing. The caller is responsible | |
165 | * for calling flush_cache_vunmap() on to-be-mapped areas before calling this | |
166 | * function and flush_tlb_kernel_range() after. | |
167 | */ | |
168 | void unmap_kernel_range_noflush(unsigned long start, unsigned long size) | |
169 | { | |
170 | unsigned long end = start + size; | |
171 | unsigned long next; | |
172 | pgd_t *pgd; | |
173 | unsigned long addr = start; | |
174 | pgtbl_mod_mask mask = 0; | |
175 | ||
176 | BUG_ON(addr >= end); | |
177 | start = addr; | |
178 | pgd = pgd_offset_k(addr); | |
179 | do { | |
180 | next = pgd_addr_end(addr, end); | |
181 | if (pgd_bad(*pgd)) | |
182 | mask |= PGTBL_PGD_MODIFIED; | |
183 | if (pgd_none_or_clear_bad(pgd)) | |
184 | continue; | |
185 | vunmap_p4d_range(pgd, addr, next, &mask); | |
186 | } while (pgd++, addr = next, addr != end); | |
187 | ||
188 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) | |
189 | arch_sync_kernel_mappings(start, end); | |
190 | } | |
191 | ||
192 | static int vmap_pte_range(pmd_t *pmd, unsigned long addr, | |
193 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, | |
194 | pgtbl_mod_mask *mask) | |
195 | { | |
196 | pte_t *pte; | |
197 | ||
198 | /* | |
199 | * nr is a running index into the array which helps higher level | |
200 | * callers keep track of where we're up to. | |
201 | */ | |
202 | ||
203 | pte = pte_alloc_kernel_track(pmd, addr, mask); | |
204 | if (!pte) | |
205 | return -ENOMEM; | |
206 | do { | |
207 | struct page *page = pages[*nr]; | |
208 | ||
209 | if (WARN_ON(!pte_none(*pte))) | |
210 | return -EBUSY; | |
211 | if (WARN_ON(!page)) | |
212 | return -ENOMEM; | |
213 | set_pte_at(&init_mm, addr, pte, mk_pte(page, prot)); | |
214 | (*nr)++; | |
215 | } while (pte++, addr += PAGE_SIZE, addr != end); | |
216 | *mask |= PGTBL_PTE_MODIFIED; | |
217 | return 0; | |
218 | } | |
219 | ||
220 | static int vmap_pmd_range(pud_t *pud, unsigned long addr, | |
221 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, | |
222 | pgtbl_mod_mask *mask) | |
223 | { | |
224 | pmd_t *pmd; | |
225 | unsigned long next; | |
226 | ||
227 | pmd = pmd_alloc_track(&init_mm, pud, addr, mask); | |
228 | if (!pmd) | |
229 | return -ENOMEM; | |
230 | do { | |
231 | next = pmd_addr_end(addr, end); | |
232 | if (vmap_pte_range(pmd, addr, next, prot, pages, nr, mask)) | |
233 | return -ENOMEM; | |
234 | } while (pmd++, addr = next, addr != end); | |
235 | return 0; | |
236 | } | |
237 | ||
238 | static int vmap_pud_range(p4d_t *p4d, unsigned long addr, | |
239 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, | |
240 | pgtbl_mod_mask *mask) | |
241 | { | |
242 | pud_t *pud; | |
243 | unsigned long next; | |
244 | ||
245 | pud = pud_alloc_track(&init_mm, p4d, addr, mask); | |
246 | if (!pud) | |
247 | return -ENOMEM; | |
248 | do { | |
249 | next = pud_addr_end(addr, end); | |
250 | if (vmap_pmd_range(pud, addr, next, prot, pages, nr, mask)) | |
251 | return -ENOMEM; | |
252 | } while (pud++, addr = next, addr != end); | |
253 | return 0; | |
254 | } | |
255 | ||
256 | static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, | |
257 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, | |
258 | pgtbl_mod_mask *mask) | |
259 | { | |
260 | p4d_t *p4d; | |
261 | unsigned long next; | |
262 | ||
263 | p4d = p4d_alloc_track(&init_mm, pgd, addr, mask); | |
264 | if (!p4d) | |
265 | return -ENOMEM; | |
266 | do { | |
267 | next = p4d_addr_end(addr, end); | |
268 | if (vmap_pud_range(p4d, addr, next, prot, pages, nr, mask)) | |
269 | return -ENOMEM; | |
270 | } while (p4d++, addr = next, addr != end); | |
271 | return 0; | |
272 | } | |
273 | ||
274 | /** | |
275 | * map_kernel_range_noflush - map kernel VM area with the specified pages | |
276 | * @addr: start of the VM area to map | |
277 | * @size: size of the VM area to map | |
278 | * @prot: page protection flags to use | |
279 | * @pages: pages to map | |
280 | * | |
281 | * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should | |
282 | * have been allocated using get_vm_area() and its friends. | |
283 | * | |
284 | * NOTE: | |
285 | * This function does NOT do any cache flushing. The caller is responsible for | |
286 | * calling flush_cache_vmap() on to-be-mapped areas before calling this | |
287 | * function. | |
288 | * | |
289 | * RETURNS: | |
290 | * 0 on success, -errno on failure. | |
291 | */ | |
292 | int map_kernel_range_noflush(unsigned long addr, unsigned long size, | |
293 | pgprot_t prot, struct page **pages) | |
294 | { | |
295 | unsigned long start = addr; | |
296 | unsigned long end = addr + size; | |
297 | unsigned long next; | |
298 | pgd_t *pgd; | |
299 | int err = 0; | |
300 | int nr = 0; | |
301 | pgtbl_mod_mask mask = 0; | |
302 | ||
303 | BUG_ON(addr >= end); | |
304 | pgd = pgd_offset_k(addr); | |
305 | do { | |
306 | next = pgd_addr_end(addr, end); | |
307 | if (pgd_bad(*pgd)) | |
308 | mask |= PGTBL_PGD_MODIFIED; | |
309 | err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr, &mask); | |
310 | if (err) | |
311 | return err; | |
312 | } while (pgd++, addr = next, addr != end); | |
313 | ||
314 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) | |
315 | arch_sync_kernel_mappings(start, end); | |
316 | ||
317 | return 0; | |
318 | } | |
319 | ||
320 | int map_kernel_range(unsigned long start, unsigned long size, pgprot_t prot, | |
321 | struct page **pages) | |
322 | { | |
323 | int ret; | |
324 | ||
325 | ret = map_kernel_range_noflush(start, size, prot, pages); | |
326 | flush_cache_vmap(start, start + size); | |
327 | return ret; | |
328 | } | |
329 | ||
330 | int is_vmalloc_or_module_addr(const void *x) | |
331 | { | |
332 | /* | |
333 | * ARM, x86-64 and sparc64 put modules in a special place, | |
334 | * and fall back on vmalloc() if that fails. Others | |
335 | * just put it in the vmalloc space. | |
336 | */ | |
337 | #if defined(CONFIG_MODULES) && defined(MODULES_VADDR) | |
338 | unsigned long addr = (unsigned long)x; | |
339 | if (addr >= MODULES_VADDR && addr < MODULES_END) | |
340 | return 1; | |
341 | #endif | |
342 | return is_vmalloc_addr(x); | |
343 | } | |
344 | ||
345 | /* | |
346 | * Walk a vmap address to the struct page it maps. | |
347 | */ | |
348 | struct page *vmalloc_to_page(const void *vmalloc_addr) | |
349 | { | |
350 | unsigned long addr = (unsigned long) vmalloc_addr; | |
351 | struct page *page = NULL; | |
352 | pgd_t *pgd = pgd_offset_k(addr); | |
353 | p4d_t *p4d; | |
354 | pud_t *pud; | |
355 | pmd_t *pmd; | |
356 | pte_t *ptep, pte; | |
357 | ||
358 | /* | |
359 | * XXX we might need to change this if we add VIRTUAL_BUG_ON for | |
360 | * architectures that do not vmalloc module space | |
361 | */ | |
362 | VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr)); | |
363 | ||
364 | if (pgd_none(*pgd)) | |
365 | return NULL; | |
366 | p4d = p4d_offset(pgd, addr); | |
367 | if (p4d_none(*p4d)) | |
368 | return NULL; | |
369 | pud = pud_offset(p4d, addr); | |
370 | ||
371 | /* | |
372 | * Don't dereference bad PUD or PMD (below) entries. This will also | |
373 | * identify huge mappings, which we may encounter on architectures | |
374 | * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be | |
375 | * identified as vmalloc addresses by is_vmalloc_addr(), but are | |
376 | * not [unambiguously] associated with a struct page, so there is | |
377 | * no correct value to return for them. | |
378 | */ | |
379 | WARN_ON_ONCE(pud_bad(*pud)); | |
380 | if (pud_none(*pud) || pud_bad(*pud)) | |
381 | return NULL; | |
382 | pmd = pmd_offset(pud, addr); | |
383 | WARN_ON_ONCE(pmd_bad(*pmd)); | |
384 | if (pmd_none(*pmd) || pmd_bad(*pmd)) | |
385 | return NULL; | |
386 | ||
387 | ptep = pte_offset_map(pmd, addr); | |
388 | pte = *ptep; | |
389 | if (pte_present(pte)) | |
390 | page = pte_page(pte); | |
391 | pte_unmap(ptep); | |
392 | return page; | |
393 | } | |
394 | EXPORT_SYMBOL(vmalloc_to_page); | |
395 | ||
396 | /* | |
397 | * Map a vmalloc()-space virtual address to the physical page frame number. | |
398 | */ | |
399 | unsigned long vmalloc_to_pfn(const void *vmalloc_addr) | |
400 | { | |
401 | return page_to_pfn(vmalloc_to_page(vmalloc_addr)); | |
402 | } | |
403 | EXPORT_SYMBOL(vmalloc_to_pfn); | |
404 | ||
405 | ||
406 | /*** Global kva allocator ***/ | |
407 | ||
408 | #define DEBUG_AUGMENT_PROPAGATE_CHECK 0 | |
409 | #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0 | |
410 | ||
411 | ||
412 | static DEFINE_SPINLOCK(vmap_area_lock); | |
413 | static DEFINE_SPINLOCK(free_vmap_area_lock); | |
414 | /* Export for kexec only */ | |
415 | LIST_HEAD(vmap_area_list); | |
416 | static LLIST_HEAD(vmap_purge_list); | |
417 | static struct rb_root vmap_area_root = RB_ROOT; | |
418 | static bool vmap_initialized __read_mostly; | |
419 | ||
420 | /* | |
421 | * This kmem_cache is used for vmap_area objects. Instead of | |
422 | * allocating from slab we reuse an object from this cache to | |
423 | * make things faster. Especially in "no edge" splitting of | |
424 | * free block. | |
425 | */ | |
426 | static struct kmem_cache *vmap_area_cachep; | |
427 | ||
428 | /* | |
429 | * This linked list is used in pair with free_vmap_area_root. | |
430 | * It gives O(1) access to prev/next to perform fast coalescing. | |
431 | */ | |
432 | static LIST_HEAD(free_vmap_area_list); | |
433 | ||
434 | /* | |
435 | * This augment red-black tree represents the free vmap space. | |
436 | * All vmap_area objects in this tree are sorted by va->va_start | |
437 | * address. It is used for allocation and merging when a vmap | |
438 | * object is released. | |
439 | * | |
440 | * Each vmap_area node contains a maximum available free block | |
441 | * of its sub-tree, right or left. Therefore it is possible to | |
442 | * find a lowest match of free area. | |
443 | */ | |
444 | static struct rb_root free_vmap_area_root = RB_ROOT; | |
445 | ||
446 | /* | |
447 | * Preload a CPU with one object for "no edge" split case. The | |
448 | * aim is to get rid of allocations from the atomic context, thus | |
449 | * to use more permissive allocation masks. | |
450 | */ | |
451 | static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node); | |
452 | ||
453 | static __always_inline unsigned long | |
454 | va_size(struct vmap_area *va) | |
455 | { | |
456 | return (va->va_end - va->va_start); | |
457 | } | |
458 | ||
459 | static __always_inline unsigned long | |
460 | get_subtree_max_size(struct rb_node *node) | |
461 | { | |
462 | struct vmap_area *va; | |
463 | ||
464 | va = rb_entry_safe(node, struct vmap_area, rb_node); | |
465 | return va ? va->subtree_max_size : 0; | |
466 | } | |
467 | ||
468 | /* | |
469 | * Gets called when remove the node and rotate. | |
470 | */ | |
471 | static __always_inline unsigned long | |
472 | compute_subtree_max_size(struct vmap_area *va) | |
473 | { | |
474 | return max3(va_size(va), | |
475 | get_subtree_max_size(va->rb_node.rb_left), | |
476 | get_subtree_max_size(va->rb_node.rb_right)); | |
477 | } | |
478 | ||
479 | RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb, | |
480 | struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size) | |
481 | ||
482 | static void purge_vmap_area_lazy(void); | |
483 | static BLOCKING_NOTIFIER_HEAD(vmap_notify_list); | |
484 | static unsigned long lazy_max_pages(void); | |
485 | ||
486 | static atomic_long_t nr_vmalloc_pages; | |
487 | ||
488 | unsigned long vmalloc_nr_pages(void) | |
489 | { | |
490 | return atomic_long_read(&nr_vmalloc_pages); | |
491 | } | |
492 | ||
493 | static struct vmap_area *__find_vmap_area(unsigned long addr) | |
494 | { | |
495 | struct rb_node *n = vmap_area_root.rb_node; | |
496 | ||
497 | while (n) { | |
498 | struct vmap_area *va; | |
499 | ||
500 | va = rb_entry(n, struct vmap_area, rb_node); | |
501 | if (addr < va->va_start) | |
502 | n = n->rb_left; | |
503 | else if (addr >= va->va_end) | |
504 | n = n->rb_right; | |
505 | else | |
506 | return va; | |
507 | } | |
508 | ||
509 | return NULL; | |
510 | } | |
511 | ||
512 | /* | |
513 | * This function returns back addresses of parent node | |
514 | * and its left or right link for further processing. | |
515 | */ | |
516 | static __always_inline struct rb_node ** | |
517 | find_va_links(struct vmap_area *va, | |
518 | struct rb_root *root, struct rb_node *from, | |
519 | struct rb_node **parent) | |
520 | { | |
521 | struct vmap_area *tmp_va; | |
522 | struct rb_node **link; | |
523 | ||
524 | if (root) { | |
525 | link = &root->rb_node; | |
526 | if (unlikely(!*link)) { | |
527 | *parent = NULL; | |
528 | return link; | |
529 | } | |
530 | } else { | |
531 | link = &from; | |
532 | } | |
533 | ||
534 | /* | |
535 | * Go to the bottom of the tree. When we hit the last point | |
536 | * we end up with parent rb_node and correct direction, i name | |
537 | * it link, where the new va->rb_node will be attached to. | |
538 | */ | |
539 | do { | |
540 | tmp_va = rb_entry(*link, struct vmap_area, rb_node); | |
541 | ||
542 | /* | |
543 | * During the traversal we also do some sanity check. | |
544 | * Trigger the BUG() if there are sides(left/right) | |
545 | * or full overlaps. | |
546 | */ | |
547 | if (va->va_start < tmp_va->va_end && | |
548 | va->va_end <= tmp_va->va_start) | |
549 | link = &(*link)->rb_left; | |
550 | else if (va->va_end > tmp_va->va_start && | |
551 | va->va_start >= tmp_va->va_end) | |
552 | link = &(*link)->rb_right; | |
553 | else | |
554 | BUG(); | |
555 | } while (*link); | |
556 | ||
557 | *parent = &tmp_va->rb_node; | |
558 | return link; | |
559 | } | |
560 | ||
561 | static __always_inline struct list_head * | |
562 | get_va_next_sibling(struct rb_node *parent, struct rb_node **link) | |
563 | { | |
564 | struct list_head *list; | |
565 | ||
566 | if (unlikely(!parent)) | |
567 | /* | |
568 | * The red-black tree where we try to find VA neighbors | |
569 | * before merging or inserting is empty, i.e. it means | |
570 | * there is no free vmap space. Normally it does not | |
571 | * happen but we handle this case anyway. | |
572 | */ | |
573 | return NULL; | |
574 | ||
575 | list = &rb_entry(parent, struct vmap_area, rb_node)->list; | |
576 | return (&parent->rb_right == link ? list->next : list); | |
577 | } | |
578 | ||
579 | static __always_inline void | |
580 | link_va(struct vmap_area *va, struct rb_root *root, | |
581 | struct rb_node *parent, struct rb_node **link, struct list_head *head) | |
582 | { | |
583 | /* | |
584 | * VA is still not in the list, but we can | |
585 | * identify its future previous list_head node. | |
586 | */ | |
587 | if (likely(parent)) { | |
588 | head = &rb_entry(parent, struct vmap_area, rb_node)->list; | |
589 | if (&parent->rb_right != link) | |
590 | head = head->prev; | |
591 | } | |
592 | ||
593 | /* Insert to the rb-tree */ | |
594 | rb_link_node(&va->rb_node, parent, link); | |
595 | if (root == &free_vmap_area_root) { | |
596 | /* | |
597 | * Some explanation here. Just perform simple insertion | |
598 | * to the tree. We do not set va->subtree_max_size to | |
599 | * its current size before calling rb_insert_augmented(). | |
600 | * It is because of we populate the tree from the bottom | |
601 | * to parent levels when the node _is_ in the tree. | |
602 | * | |
603 | * Therefore we set subtree_max_size to zero after insertion, | |
604 | * to let __augment_tree_propagate_from() puts everything to | |
605 | * the correct order later on. | |
606 | */ | |
607 | rb_insert_augmented(&va->rb_node, | |
608 | root, &free_vmap_area_rb_augment_cb); | |
609 | va->subtree_max_size = 0; | |
610 | } else { | |
611 | rb_insert_color(&va->rb_node, root); | |
612 | } | |
613 | ||
614 | /* Address-sort this list */ | |
615 | list_add(&va->list, head); | |
616 | } | |
617 | ||
618 | static __always_inline void | |
619 | unlink_va(struct vmap_area *va, struct rb_root *root) | |
620 | { | |
621 | if (WARN_ON(RB_EMPTY_NODE(&va->rb_node))) | |
622 | return; | |
623 | ||
624 | if (root == &free_vmap_area_root) | |
625 | rb_erase_augmented(&va->rb_node, | |
626 | root, &free_vmap_area_rb_augment_cb); | |
627 | else | |
628 | rb_erase(&va->rb_node, root); | |
629 | ||
630 | list_del(&va->list); | |
631 | RB_CLEAR_NODE(&va->rb_node); | |
632 | } | |
633 | ||
634 | #if DEBUG_AUGMENT_PROPAGATE_CHECK | |
635 | static void | |
636 | augment_tree_propagate_check(struct rb_node *n) | |
637 | { | |
638 | struct vmap_area *va; | |
639 | struct rb_node *node; | |
640 | unsigned long size; | |
641 | bool found = false; | |
642 | ||
643 | if (n == NULL) | |
644 | return; | |
645 | ||
646 | va = rb_entry(n, struct vmap_area, rb_node); | |
647 | size = va->subtree_max_size; | |
648 | node = n; | |
649 | ||
650 | while (node) { | |
651 | va = rb_entry(node, struct vmap_area, rb_node); | |
652 | ||
653 | if (get_subtree_max_size(node->rb_left) == size) { | |
654 | node = node->rb_left; | |
655 | } else { | |
656 | if (va_size(va) == size) { | |
657 | found = true; | |
658 | break; | |
659 | } | |
660 | ||
661 | node = node->rb_right; | |
662 | } | |
663 | } | |
664 | ||
665 | if (!found) { | |
666 | va = rb_entry(n, struct vmap_area, rb_node); | |
667 | pr_emerg("tree is corrupted: %lu, %lu\n", | |
668 | va_size(va), va->subtree_max_size); | |
669 | } | |
670 | ||
671 | augment_tree_propagate_check(n->rb_left); | |
672 | augment_tree_propagate_check(n->rb_right); | |
673 | } | |
674 | #endif | |
675 | ||
676 | /* | |
677 | * This function populates subtree_max_size from bottom to upper | |
678 | * levels starting from VA point. The propagation must be done | |
679 | * when VA size is modified by changing its va_start/va_end. Or | |
680 | * in case of newly inserting of VA to the tree. | |
681 | * | |
682 | * It means that __augment_tree_propagate_from() must be called: | |
683 | * - After VA has been inserted to the tree(free path); | |
684 | * - After VA has been shrunk(allocation path); | |
685 | * - After VA has been increased(merging path). | |
686 | * | |
687 | * Please note that, it does not mean that upper parent nodes | |
688 | * and their subtree_max_size are recalculated all the time up | |
689 | * to the root node. | |
690 | * | |
691 | * 4--8 | |
692 | * /\ | |
693 | * / \ | |
694 | * / \ | |
695 | * 2--2 8--8 | |
696 | * | |
697 | * For example if we modify the node 4, shrinking it to 2, then | |
698 | * no any modification is required. If we shrink the node 2 to 1 | |
699 | * its subtree_max_size is updated only, and set to 1. If we shrink | |
700 | * the node 8 to 6, then its subtree_max_size is set to 6 and parent | |
701 | * node becomes 4--6. | |
702 | */ | |
703 | static __always_inline void | |
704 | augment_tree_propagate_from(struct vmap_area *va) | |
705 | { | |
706 | struct rb_node *node = &va->rb_node; | |
707 | unsigned long new_va_sub_max_size; | |
708 | ||
709 | while (node) { | |
710 | va = rb_entry(node, struct vmap_area, rb_node); | |
711 | new_va_sub_max_size = compute_subtree_max_size(va); | |
712 | ||
713 | /* | |
714 | * If the newly calculated maximum available size of the | |
715 | * subtree is equal to the current one, then it means that | |
716 | * the tree is propagated correctly. So we have to stop at | |
717 | * this point to save cycles. | |
718 | */ | |
719 | if (va->subtree_max_size == new_va_sub_max_size) | |
720 | break; | |
721 | ||
722 | va->subtree_max_size = new_va_sub_max_size; | |
723 | node = rb_parent(&va->rb_node); | |
724 | } | |
725 | ||
726 | #if DEBUG_AUGMENT_PROPAGATE_CHECK | |
727 | augment_tree_propagate_check(free_vmap_area_root.rb_node); | |
728 | #endif | |
729 | } | |
730 | ||
731 | static void | |
732 | insert_vmap_area(struct vmap_area *va, | |
733 | struct rb_root *root, struct list_head *head) | |
734 | { | |
735 | struct rb_node **link; | |
736 | struct rb_node *parent; | |
737 | ||
738 | link = find_va_links(va, root, NULL, &parent); | |
739 | link_va(va, root, parent, link, head); | |
740 | } | |
741 | ||
742 | static void | |
743 | insert_vmap_area_augment(struct vmap_area *va, | |
744 | struct rb_node *from, struct rb_root *root, | |
745 | struct list_head *head) | |
746 | { | |
747 | struct rb_node **link; | |
748 | struct rb_node *parent; | |
749 | ||
750 | if (from) | |
751 | link = find_va_links(va, NULL, from, &parent); | |
752 | else | |
753 | link = find_va_links(va, root, NULL, &parent); | |
754 | ||
755 | link_va(va, root, parent, link, head); | |
756 | augment_tree_propagate_from(va); | |
757 | } | |
758 | ||
759 | /* | |
760 | * Merge de-allocated chunk of VA memory with previous | |
761 | * and next free blocks. If coalesce is not done a new | |
762 | * free area is inserted. If VA has been merged, it is | |
763 | * freed. | |
764 | */ | |
765 | static __always_inline struct vmap_area * | |
766 | merge_or_add_vmap_area(struct vmap_area *va, | |
767 | struct rb_root *root, struct list_head *head) | |
768 | { | |
769 | struct vmap_area *sibling; | |
770 | struct list_head *next; | |
771 | struct rb_node **link; | |
772 | struct rb_node *parent; | |
773 | bool merged = false; | |
774 | ||
775 | /* | |
776 | * Find a place in the tree where VA potentially will be | |
777 | * inserted, unless it is merged with its sibling/siblings. | |
778 | */ | |
779 | link = find_va_links(va, root, NULL, &parent); | |
780 | ||
781 | /* | |
782 | * Get next node of VA to check if merging can be done. | |
783 | */ | |
784 | next = get_va_next_sibling(parent, link); | |
785 | if (unlikely(next == NULL)) | |
786 | goto insert; | |
787 | ||
788 | /* | |
789 | * start end | |
790 | * | | | |
791 | * |<------VA------>|<-----Next----->| | |
792 | * | | | |
793 | * start end | |
794 | */ | |
795 | if (next != head) { | |
796 | sibling = list_entry(next, struct vmap_area, list); | |
797 | if (sibling->va_start == va->va_end) { | |
798 | sibling->va_start = va->va_start; | |
799 | ||
800 | /* Check and update the tree if needed. */ | |
801 | augment_tree_propagate_from(sibling); | |
802 | ||
803 | /* Free vmap_area object. */ | |
804 | kmem_cache_free(vmap_area_cachep, va); | |
805 | ||
806 | /* Point to the new merged area. */ | |
807 | va = sibling; | |
808 | merged = true; | |
809 | } | |
810 | } | |
811 | ||
812 | /* | |
813 | * start end | |
814 | * | | | |
815 | * |<-----Prev----->|<------VA------>| | |
816 | * | | | |
817 | * start end | |
818 | */ | |
819 | if (next->prev != head) { | |
820 | sibling = list_entry(next->prev, struct vmap_area, list); | |
821 | if (sibling->va_end == va->va_start) { | |
822 | sibling->va_end = va->va_end; | |
823 | ||
824 | /* Check and update the tree if needed. */ | |
825 | augment_tree_propagate_from(sibling); | |
826 | ||
827 | if (merged) | |
828 | unlink_va(va, root); | |
829 | ||
830 | /* Free vmap_area object. */ | |
831 | kmem_cache_free(vmap_area_cachep, va); | |
832 | ||
833 | /* Point to the new merged area. */ | |
834 | va = sibling; | |
835 | merged = true; | |
836 | } | |
837 | } | |
838 | ||
839 | insert: | |
840 | if (!merged) { | |
841 | link_va(va, root, parent, link, head); | |
842 | augment_tree_propagate_from(va); | |
843 | } | |
844 | ||
845 | return va; | |
846 | } | |
847 | ||
848 | static __always_inline bool | |
849 | is_within_this_va(struct vmap_area *va, unsigned long size, | |
850 | unsigned long align, unsigned long vstart) | |
851 | { | |
852 | unsigned long nva_start_addr; | |
853 | ||
854 | if (va->va_start > vstart) | |
855 | nva_start_addr = ALIGN(va->va_start, align); | |
856 | else | |
857 | nva_start_addr = ALIGN(vstart, align); | |
858 | ||
859 | /* Can be overflowed due to big size or alignment. */ | |
860 | if (nva_start_addr + size < nva_start_addr || | |
861 | nva_start_addr < vstart) | |
862 | return false; | |
863 | ||
864 | return (nva_start_addr + size <= va->va_end); | |
865 | } | |
866 | ||
867 | /* | |
868 | * Find the first free block(lowest start address) in the tree, | |
869 | * that will accomplish the request corresponding to passing | |
870 | * parameters. | |
871 | */ | |
872 | static __always_inline struct vmap_area * | |
873 | find_vmap_lowest_match(unsigned long size, | |
874 | unsigned long align, unsigned long vstart) | |
875 | { | |
876 | struct vmap_area *va; | |
877 | struct rb_node *node; | |
878 | unsigned long length; | |
879 | ||
880 | /* Start from the root. */ | |
881 | node = free_vmap_area_root.rb_node; | |
882 | ||
883 | /* Adjust the search size for alignment overhead. */ | |
884 | length = size + align - 1; | |
885 | ||
886 | while (node) { | |
887 | va = rb_entry(node, struct vmap_area, rb_node); | |
888 | ||
889 | if (get_subtree_max_size(node->rb_left) >= length && | |
890 | vstart < va->va_start) { | |
891 | node = node->rb_left; | |
892 | } else { | |
893 | if (is_within_this_va(va, size, align, vstart)) | |
894 | return va; | |
895 | ||
896 | /* | |
897 | * Does not make sense to go deeper towards the right | |
898 | * sub-tree if it does not have a free block that is | |
899 | * equal or bigger to the requested search length. | |
900 | */ | |
901 | if (get_subtree_max_size(node->rb_right) >= length) { | |
902 | node = node->rb_right; | |
903 | continue; | |
904 | } | |
905 | ||
906 | /* | |
907 | * OK. We roll back and find the first right sub-tree, | |
908 | * that will satisfy the search criteria. It can happen | |
909 | * only once due to "vstart" restriction. | |
910 | */ | |
911 | while ((node = rb_parent(node))) { | |
912 | va = rb_entry(node, struct vmap_area, rb_node); | |
913 | if (is_within_this_va(va, size, align, vstart)) | |
914 | return va; | |
915 | ||
916 | if (get_subtree_max_size(node->rb_right) >= length && | |
917 | vstart <= va->va_start) { | |
918 | node = node->rb_right; | |
919 | break; | |
920 | } | |
921 | } | |
922 | } | |
923 | } | |
924 | ||
925 | return NULL; | |
926 | } | |
927 | ||
928 | #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK | |
929 | #include <linux/random.h> | |
930 | ||
931 | static struct vmap_area * | |
932 | find_vmap_lowest_linear_match(unsigned long size, | |
933 | unsigned long align, unsigned long vstart) | |
934 | { | |
935 | struct vmap_area *va; | |
936 | ||
937 | list_for_each_entry(va, &free_vmap_area_list, list) { | |
938 | if (!is_within_this_va(va, size, align, vstart)) | |
939 | continue; | |
940 | ||
941 | return va; | |
942 | } | |
943 | ||
944 | return NULL; | |
945 | } | |
946 | ||
947 | static void | |
948 | find_vmap_lowest_match_check(unsigned long size) | |
949 | { | |
950 | struct vmap_area *va_1, *va_2; | |
951 | unsigned long vstart; | |
952 | unsigned int rnd; | |
953 | ||
954 | get_random_bytes(&rnd, sizeof(rnd)); | |
955 | vstart = VMALLOC_START + rnd; | |
956 | ||
957 | va_1 = find_vmap_lowest_match(size, 1, vstart); | |
958 | va_2 = find_vmap_lowest_linear_match(size, 1, vstart); | |
959 | ||
960 | if (va_1 != va_2) | |
961 | pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n", | |
962 | va_1, va_2, vstart); | |
963 | } | |
964 | #endif | |
965 | ||
966 | enum fit_type { | |
967 | NOTHING_FIT = 0, | |
968 | FL_FIT_TYPE = 1, /* full fit */ | |
969 | LE_FIT_TYPE = 2, /* left edge fit */ | |
970 | RE_FIT_TYPE = 3, /* right edge fit */ | |
971 | NE_FIT_TYPE = 4 /* no edge fit */ | |
972 | }; | |
973 | ||
974 | static __always_inline enum fit_type | |
975 | classify_va_fit_type(struct vmap_area *va, | |
976 | unsigned long nva_start_addr, unsigned long size) | |
977 | { | |
978 | enum fit_type type; | |
979 | ||
980 | /* Check if it is within VA. */ | |
981 | if (nva_start_addr < va->va_start || | |
982 | nva_start_addr + size > va->va_end) | |
983 | return NOTHING_FIT; | |
984 | ||
985 | /* Now classify. */ | |
986 | if (va->va_start == nva_start_addr) { | |
987 | if (va->va_end == nva_start_addr + size) | |
988 | type = FL_FIT_TYPE; | |
989 | else | |
990 | type = LE_FIT_TYPE; | |
991 | } else if (va->va_end == nva_start_addr + size) { | |
992 | type = RE_FIT_TYPE; | |
993 | } else { | |
994 | type = NE_FIT_TYPE; | |
995 | } | |
996 | ||
997 | return type; | |
998 | } | |
999 | ||
1000 | static __always_inline int | |
1001 | adjust_va_to_fit_type(struct vmap_area *va, | |
1002 | unsigned long nva_start_addr, unsigned long size, | |
1003 | enum fit_type type) | |
1004 | { | |
1005 | struct vmap_area *lva = NULL; | |
1006 | ||
1007 | if (type == FL_FIT_TYPE) { | |
1008 | /* | |
1009 | * No need to split VA, it fully fits. | |
1010 | * | |
1011 | * | | | |
1012 | * V NVA V | |
1013 | * |---------------| | |
1014 | */ | |
1015 | unlink_va(va, &free_vmap_area_root); | |
1016 | kmem_cache_free(vmap_area_cachep, va); | |
1017 | } else if (type == LE_FIT_TYPE) { | |
1018 | /* | |
1019 | * Split left edge of fit VA. | |
1020 | * | |
1021 | * | | | |
1022 | * V NVA V R | |
1023 | * |-------|-------| | |
1024 | */ | |
1025 | va->va_start += size; | |
1026 | } else if (type == RE_FIT_TYPE) { | |
1027 | /* | |
1028 | * Split right edge of fit VA. | |
1029 | * | |
1030 | * | | | |
1031 | * L V NVA V | |
1032 | * |-------|-------| | |
1033 | */ | |
1034 | va->va_end = nva_start_addr; | |
1035 | } else if (type == NE_FIT_TYPE) { | |
1036 | /* | |
1037 | * Split no edge of fit VA. | |
1038 | * | |
1039 | * | | | |
1040 | * L V NVA V R | |
1041 | * |---|-------|---| | |
1042 | */ | |
1043 | lva = __this_cpu_xchg(ne_fit_preload_node, NULL); | |
1044 | if (unlikely(!lva)) { | |
1045 | /* | |
1046 | * For percpu allocator we do not do any pre-allocation | |
1047 | * and leave it as it is. The reason is it most likely | |
1048 | * never ends up with NE_FIT_TYPE splitting. In case of | |
1049 | * percpu allocations offsets and sizes are aligned to | |
1050 | * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE | |
1051 | * are its main fitting cases. | |
1052 | * | |
1053 | * There are a few exceptions though, as an example it is | |
1054 | * a first allocation (early boot up) when we have "one" | |
1055 | * big free space that has to be split. | |
1056 | * | |
1057 | * Also we can hit this path in case of regular "vmap" | |
1058 | * allocations, if "this" current CPU was not preloaded. | |
1059 | * See the comment in alloc_vmap_area() why. If so, then | |
1060 | * GFP_NOWAIT is used instead to get an extra object for | |
1061 | * split purpose. That is rare and most time does not | |
1062 | * occur. | |
1063 | * | |
1064 | * What happens if an allocation gets failed. Basically, | |
1065 | * an "overflow" path is triggered to purge lazily freed | |
1066 | * areas to free some memory, then, the "retry" path is | |
1067 | * triggered to repeat one more time. See more details | |
1068 | * in alloc_vmap_area() function. | |
1069 | */ | |
1070 | lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT); | |
1071 | if (!lva) | |
1072 | return -1; | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * Build the remainder. | |
1077 | */ | |
1078 | lva->va_start = va->va_start; | |
1079 | lva->va_end = nva_start_addr; | |
1080 | ||
1081 | /* | |
1082 | * Shrink this VA to remaining size. | |
1083 | */ | |
1084 | va->va_start = nva_start_addr + size; | |
1085 | } else { | |
1086 | return -1; | |
1087 | } | |
1088 | ||
1089 | if (type != FL_FIT_TYPE) { | |
1090 | augment_tree_propagate_from(va); | |
1091 | ||
1092 | if (lva) /* type == NE_FIT_TYPE */ | |
1093 | insert_vmap_area_augment(lva, &va->rb_node, | |
1094 | &free_vmap_area_root, &free_vmap_area_list); | |
1095 | } | |
1096 | ||
1097 | return 0; | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * Returns a start address of the newly allocated area, if success. | |
1102 | * Otherwise a vend is returned that indicates failure. | |
1103 | */ | |
1104 | static __always_inline unsigned long | |
1105 | __alloc_vmap_area(unsigned long size, unsigned long align, | |
1106 | unsigned long vstart, unsigned long vend) | |
1107 | { | |
1108 | unsigned long nva_start_addr; | |
1109 | struct vmap_area *va; | |
1110 | enum fit_type type; | |
1111 | int ret; | |
1112 | ||
1113 | va = find_vmap_lowest_match(size, align, vstart); | |
1114 | if (unlikely(!va)) | |
1115 | return vend; | |
1116 | ||
1117 | if (va->va_start > vstart) | |
1118 | nva_start_addr = ALIGN(va->va_start, align); | |
1119 | else | |
1120 | nva_start_addr = ALIGN(vstart, align); | |
1121 | ||
1122 | /* Check the "vend" restriction. */ | |
1123 | if (nva_start_addr + size > vend) | |
1124 | return vend; | |
1125 | ||
1126 | /* Classify what we have found. */ | |
1127 | type = classify_va_fit_type(va, nva_start_addr, size); | |
1128 | if (WARN_ON_ONCE(type == NOTHING_FIT)) | |
1129 | return vend; | |
1130 | ||
1131 | /* Update the free vmap_area. */ | |
1132 | ret = adjust_va_to_fit_type(va, nva_start_addr, size, type); | |
1133 | if (ret) | |
1134 | return vend; | |
1135 | ||
1136 | #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK | |
1137 | find_vmap_lowest_match_check(size); | |
1138 | #endif | |
1139 | ||
1140 | return nva_start_addr; | |
1141 | } | |
1142 | ||
1143 | /* | |
1144 | * Free a region of KVA allocated by alloc_vmap_area | |
1145 | */ | |
1146 | static void free_vmap_area(struct vmap_area *va) | |
1147 | { | |
1148 | /* | |
1149 | * Remove from the busy tree/list. | |
1150 | */ | |
1151 | spin_lock(&vmap_area_lock); | |
1152 | unlink_va(va, &vmap_area_root); | |
1153 | spin_unlock(&vmap_area_lock); | |
1154 | ||
1155 | /* | |
1156 | * Insert/Merge it back to the free tree/list. | |
1157 | */ | |
1158 | spin_lock(&free_vmap_area_lock); | |
1159 | merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list); | |
1160 | spin_unlock(&free_vmap_area_lock); | |
1161 | } | |
1162 | ||
1163 | /* | |
1164 | * Allocate a region of KVA of the specified size and alignment, within the | |
1165 | * vstart and vend. | |
1166 | */ | |
1167 | static struct vmap_area *alloc_vmap_area(unsigned long size, | |
1168 | unsigned long align, | |
1169 | unsigned long vstart, unsigned long vend, | |
1170 | int node, gfp_t gfp_mask) | |
1171 | { | |
1172 | struct vmap_area *va, *pva; | |
1173 | unsigned long addr; | |
1174 | int purged = 0; | |
1175 | int ret; | |
1176 | ||
1177 | BUG_ON(!size); | |
1178 | BUG_ON(offset_in_page(size)); | |
1179 | BUG_ON(!is_power_of_2(align)); | |
1180 | ||
1181 | if (unlikely(!vmap_initialized)) | |
1182 | return ERR_PTR(-EBUSY); | |
1183 | ||
1184 | might_sleep(); | |
1185 | gfp_mask = gfp_mask & GFP_RECLAIM_MASK; | |
1186 | ||
1187 | va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node); | |
1188 | if (unlikely(!va)) | |
1189 | return ERR_PTR(-ENOMEM); | |
1190 | ||
1191 | /* | |
1192 | * Only scan the relevant parts containing pointers to other objects | |
1193 | * to avoid false negatives. | |
1194 | */ | |
1195 | kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask); | |
1196 | ||
1197 | retry: | |
1198 | /* | |
1199 | * Preload this CPU with one extra vmap_area object. It is used | |
1200 | * when fit type of free area is NE_FIT_TYPE. Please note, it | |
1201 | * does not guarantee that an allocation occurs on a CPU that | |
1202 | * is preloaded, instead we minimize the case when it is not. | |
1203 | * It can happen because of cpu migration, because there is a | |
1204 | * race until the below spinlock is taken. | |
1205 | * | |
1206 | * The preload is done in non-atomic context, thus it allows us | |
1207 | * to use more permissive allocation masks to be more stable under | |
1208 | * low memory condition and high memory pressure. In rare case, | |
1209 | * if not preloaded, GFP_NOWAIT is used. | |
1210 | * | |
1211 | * Set "pva" to NULL here, because of "retry" path. | |
1212 | */ | |
1213 | pva = NULL; | |
1214 | ||
1215 | if (!this_cpu_read(ne_fit_preload_node)) | |
1216 | /* | |
1217 | * Even if it fails we do not really care about that. | |
1218 | * Just proceed as it is. If needed "overflow" path | |
1219 | * will refill the cache we allocate from. | |
1220 | */ | |
1221 | pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node); | |
1222 | ||
1223 | spin_lock(&free_vmap_area_lock); | |
1224 | ||
1225 | if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva)) | |
1226 | kmem_cache_free(vmap_area_cachep, pva); | |
1227 | ||
1228 | /* | |
1229 | * If an allocation fails, the "vend" address is | |
1230 | * returned. Therefore trigger the overflow path. | |
1231 | */ | |
1232 | addr = __alloc_vmap_area(size, align, vstart, vend); | |
1233 | spin_unlock(&free_vmap_area_lock); | |
1234 | ||
1235 | if (unlikely(addr == vend)) | |
1236 | goto overflow; | |
1237 | ||
1238 | va->va_start = addr; | |
1239 | va->va_end = addr + size; | |
1240 | va->vm = NULL; | |
1241 | ||
1242 | ||
1243 | spin_lock(&vmap_area_lock); | |
1244 | insert_vmap_area(va, &vmap_area_root, &vmap_area_list); | |
1245 | spin_unlock(&vmap_area_lock); | |
1246 | ||
1247 | BUG_ON(!IS_ALIGNED(va->va_start, align)); | |
1248 | BUG_ON(va->va_start < vstart); | |
1249 | BUG_ON(va->va_end > vend); | |
1250 | ||
1251 | ret = kasan_populate_vmalloc(addr, size); | |
1252 | if (ret) { | |
1253 | free_vmap_area(va); | |
1254 | return ERR_PTR(ret); | |
1255 | } | |
1256 | ||
1257 | return va; | |
1258 | ||
1259 | overflow: | |
1260 | if (!purged) { | |
1261 | purge_vmap_area_lazy(); | |
1262 | purged = 1; | |
1263 | goto retry; | |
1264 | } | |
1265 | ||
1266 | if (gfpflags_allow_blocking(gfp_mask)) { | |
1267 | unsigned long freed = 0; | |
1268 | blocking_notifier_call_chain(&vmap_notify_list, 0, &freed); | |
1269 | if (freed > 0) { | |
1270 | purged = 0; | |
1271 | goto retry; | |
1272 | } | |
1273 | } | |
1274 | ||
1275 | if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) | |
1276 | pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n", | |
1277 | size); | |
1278 | ||
1279 | kmem_cache_free(vmap_area_cachep, va); | |
1280 | return ERR_PTR(-EBUSY); | |
1281 | } | |
1282 | ||
1283 | int register_vmap_purge_notifier(struct notifier_block *nb) | |
1284 | { | |
1285 | return blocking_notifier_chain_register(&vmap_notify_list, nb); | |
1286 | } | |
1287 | EXPORT_SYMBOL_GPL(register_vmap_purge_notifier); | |
1288 | ||
1289 | int unregister_vmap_purge_notifier(struct notifier_block *nb) | |
1290 | { | |
1291 | return blocking_notifier_chain_unregister(&vmap_notify_list, nb); | |
1292 | } | |
1293 | EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier); | |
1294 | ||
1295 | /* | |
1296 | * lazy_max_pages is the maximum amount of virtual address space we gather up | |
1297 | * before attempting to purge with a TLB flush. | |
1298 | * | |
1299 | * There is a tradeoff here: a larger number will cover more kernel page tables | |
1300 | * and take slightly longer to purge, but it will linearly reduce the number of | |
1301 | * global TLB flushes that must be performed. It would seem natural to scale | |
1302 | * this number up linearly with the number of CPUs (because vmapping activity | |
1303 | * could also scale linearly with the number of CPUs), however it is likely | |
1304 | * that in practice, workloads might be constrained in other ways that mean | |
1305 | * vmap activity will not scale linearly with CPUs. Also, I want to be | |
1306 | * conservative and not introduce a big latency on huge systems, so go with | |
1307 | * a less aggressive log scale. It will still be an improvement over the old | |
1308 | * code, and it will be simple to change the scale factor if we find that it | |
1309 | * becomes a problem on bigger systems. | |
1310 | */ | |
1311 | static unsigned long lazy_max_pages(void) | |
1312 | { | |
1313 | unsigned int log; | |
1314 | ||
1315 | log = fls(num_online_cpus()); | |
1316 | ||
1317 | return log * (32UL * 1024 * 1024 / PAGE_SIZE); | |
1318 | } | |
1319 | ||
1320 | static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0); | |
1321 | ||
1322 | /* | |
1323 | * Serialize vmap purging. There is no actual criticial section protected | |
1324 | * by this look, but we want to avoid concurrent calls for performance | |
1325 | * reasons and to make the pcpu_get_vm_areas more deterministic. | |
1326 | */ | |
1327 | static DEFINE_MUTEX(vmap_purge_lock); | |
1328 | ||
1329 | /* for per-CPU blocks */ | |
1330 | static void purge_fragmented_blocks_allcpus(void); | |
1331 | ||
1332 | /* | |
1333 | * called before a call to iounmap() if the caller wants vm_area_struct's | |
1334 | * immediately freed. | |
1335 | */ | |
1336 | void set_iounmap_nonlazy(void) | |
1337 | { | |
1338 | atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1); | |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * Purges all lazily-freed vmap areas. | |
1343 | */ | |
1344 | static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end) | |
1345 | { | |
1346 | unsigned long resched_threshold; | |
1347 | struct llist_node *valist; | |
1348 | struct vmap_area *va; | |
1349 | struct vmap_area *n_va; | |
1350 | ||
1351 | lockdep_assert_held(&vmap_purge_lock); | |
1352 | ||
1353 | valist = llist_del_all(&vmap_purge_list); | |
1354 | if (unlikely(valist == NULL)) | |
1355 | return false; | |
1356 | ||
1357 | /* | |
1358 | * TODO: to calculate a flush range without looping. | |
1359 | * The list can be up to lazy_max_pages() elements. | |
1360 | */ | |
1361 | llist_for_each_entry(va, valist, purge_list) { | |
1362 | if (va->va_start < start) | |
1363 | start = va->va_start; | |
1364 | if (va->va_end > end) | |
1365 | end = va->va_end; | |
1366 | } | |
1367 | ||
1368 | flush_tlb_kernel_range(start, end); | |
1369 | resched_threshold = lazy_max_pages() << 1; | |
1370 | ||
1371 | spin_lock(&free_vmap_area_lock); | |
1372 | llist_for_each_entry_safe(va, n_va, valist, purge_list) { | |
1373 | unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT; | |
1374 | unsigned long orig_start = va->va_start; | |
1375 | unsigned long orig_end = va->va_end; | |
1376 | ||
1377 | /* | |
1378 | * Finally insert or merge lazily-freed area. It is | |
1379 | * detached and there is no need to "unlink" it from | |
1380 | * anything. | |
1381 | */ | |
1382 | va = merge_or_add_vmap_area(va, &free_vmap_area_root, | |
1383 | &free_vmap_area_list); | |
1384 | ||
1385 | if (is_vmalloc_or_module_addr((void *)orig_start)) | |
1386 | kasan_release_vmalloc(orig_start, orig_end, | |
1387 | va->va_start, va->va_end); | |
1388 | ||
1389 | atomic_long_sub(nr, &vmap_lazy_nr); | |
1390 | ||
1391 | if (atomic_long_read(&vmap_lazy_nr) < resched_threshold) | |
1392 | cond_resched_lock(&free_vmap_area_lock); | |
1393 | } | |
1394 | spin_unlock(&free_vmap_area_lock); | |
1395 | return true; | |
1396 | } | |
1397 | ||
1398 | /* | |
1399 | * Kick off a purge of the outstanding lazy areas. Don't bother if somebody | |
1400 | * is already purging. | |
1401 | */ | |
1402 | static void try_purge_vmap_area_lazy(void) | |
1403 | { | |
1404 | if (mutex_trylock(&vmap_purge_lock)) { | |
1405 | __purge_vmap_area_lazy(ULONG_MAX, 0); | |
1406 | mutex_unlock(&vmap_purge_lock); | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | /* | |
1411 | * Kick off a purge of the outstanding lazy areas. | |
1412 | */ | |
1413 | static void purge_vmap_area_lazy(void) | |
1414 | { | |
1415 | mutex_lock(&vmap_purge_lock); | |
1416 | purge_fragmented_blocks_allcpus(); | |
1417 | __purge_vmap_area_lazy(ULONG_MAX, 0); | |
1418 | mutex_unlock(&vmap_purge_lock); | |
1419 | } | |
1420 | ||
1421 | /* | |
1422 | * Free a vmap area, caller ensuring that the area has been unmapped | |
1423 | * and flush_cache_vunmap had been called for the correct range | |
1424 | * previously. | |
1425 | */ | |
1426 | static void free_vmap_area_noflush(struct vmap_area *va) | |
1427 | { | |
1428 | unsigned long nr_lazy; | |
1429 | ||
1430 | spin_lock(&vmap_area_lock); | |
1431 | unlink_va(va, &vmap_area_root); | |
1432 | spin_unlock(&vmap_area_lock); | |
1433 | ||
1434 | nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> | |
1435 | PAGE_SHIFT, &vmap_lazy_nr); | |
1436 | ||
1437 | /* After this point, we may free va at any time */ | |
1438 | llist_add(&va->purge_list, &vmap_purge_list); | |
1439 | ||
1440 | if (unlikely(nr_lazy > lazy_max_pages())) | |
1441 | try_purge_vmap_area_lazy(); | |
1442 | } | |
1443 | ||
1444 | /* | |
1445 | * Free and unmap a vmap area | |
1446 | */ | |
1447 | static void free_unmap_vmap_area(struct vmap_area *va) | |
1448 | { | |
1449 | flush_cache_vunmap(va->va_start, va->va_end); | |
1450 | unmap_kernel_range_noflush(va->va_start, va->va_end - va->va_start); | |
1451 | if (debug_pagealloc_enabled_static()) | |
1452 | flush_tlb_kernel_range(va->va_start, va->va_end); | |
1453 | ||
1454 | free_vmap_area_noflush(va); | |
1455 | } | |
1456 | ||
1457 | static struct vmap_area *find_vmap_area(unsigned long addr) | |
1458 | { | |
1459 | struct vmap_area *va; | |
1460 | ||
1461 | spin_lock(&vmap_area_lock); | |
1462 | va = __find_vmap_area(addr); | |
1463 | spin_unlock(&vmap_area_lock); | |
1464 | ||
1465 | return va; | |
1466 | } | |
1467 | ||
1468 | /*** Per cpu kva allocator ***/ | |
1469 | ||
1470 | /* | |
1471 | * vmap space is limited especially on 32 bit architectures. Ensure there is | |
1472 | * room for at least 16 percpu vmap blocks per CPU. | |
1473 | */ | |
1474 | /* | |
1475 | * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able | |
1476 | * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess | |
1477 | * instead (we just need a rough idea) | |
1478 | */ | |
1479 | #if BITS_PER_LONG == 32 | |
1480 | #define VMALLOC_SPACE (128UL*1024*1024) | |
1481 | #else | |
1482 | #define VMALLOC_SPACE (128UL*1024*1024*1024) | |
1483 | #endif | |
1484 | ||
1485 | #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE) | |
1486 | #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */ | |
1487 | #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */ | |
1488 | #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2) | |
1489 | #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */ | |
1490 | #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */ | |
1491 | #define VMAP_BBMAP_BITS \ | |
1492 | VMAP_MIN(VMAP_BBMAP_BITS_MAX, \ | |
1493 | VMAP_MAX(VMAP_BBMAP_BITS_MIN, \ | |
1494 | VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16)) | |
1495 | ||
1496 | #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE) | |
1497 | ||
1498 | struct vmap_block_queue { | |
1499 | spinlock_t lock; | |
1500 | struct list_head free; | |
1501 | }; | |
1502 | ||
1503 | struct vmap_block { | |
1504 | spinlock_t lock; | |
1505 | struct vmap_area *va; | |
1506 | unsigned long free, dirty; | |
1507 | unsigned long dirty_min, dirty_max; /*< dirty range */ | |
1508 | struct list_head free_list; | |
1509 | struct rcu_head rcu_head; | |
1510 | struct list_head purge; | |
1511 | }; | |
1512 | ||
1513 | /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */ | |
1514 | static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue); | |
1515 | ||
1516 | /* | |
1517 | * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block | |
1518 | * in the free path. Could get rid of this if we change the API to return a | |
1519 | * "cookie" from alloc, to be passed to free. But no big deal yet. | |
1520 | */ | |
1521 | static DEFINE_SPINLOCK(vmap_block_tree_lock); | |
1522 | static RADIX_TREE(vmap_block_tree, GFP_ATOMIC); | |
1523 | ||
1524 | /* | |
1525 | * We should probably have a fallback mechanism to allocate virtual memory | |
1526 | * out of partially filled vmap blocks. However vmap block sizing should be | |
1527 | * fairly reasonable according to the vmalloc size, so it shouldn't be a | |
1528 | * big problem. | |
1529 | */ | |
1530 | ||
1531 | static unsigned long addr_to_vb_idx(unsigned long addr) | |
1532 | { | |
1533 | addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1); | |
1534 | addr /= VMAP_BLOCK_SIZE; | |
1535 | return addr; | |
1536 | } | |
1537 | ||
1538 | static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off) | |
1539 | { | |
1540 | unsigned long addr; | |
1541 | ||
1542 | addr = va_start + (pages_off << PAGE_SHIFT); | |
1543 | BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start)); | |
1544 | return (void *)addr; | |
1545 | } | |
1546 | ||
1547 | /** | |
1548 | * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this | |
1549 | * block. Of course pages number can't exceed VMAP_BBMAP_BITS | |
1550 | * @order: how many 2^order pages should be occupied in newly allocated block | |
1551 | * @gfp_mask: flags for the page level allocator | |
1552 | * | |
1553 | * Return: virtual address in a newly allocated block or ERR_PTR(-errno) | |
1554 | */ | |
1555 | static void *new_vmap_block(unsigned int order, gfp_t gfp_mask) | |
1556 | { | |
1557 | struct vmap_block_queue *vbq; | |
1558 | struct vmap_block *vb; | |
1559 | struct vmap_area *va; | |
1560 | unsigned long vb_idx; | |
1561 | int node, err; | |
1562 | void *vaddr; | |
1563 | ||
1564 | node = numa_node_id(); | |
1565 | ||
1566 | vb = kmalloc_node(sizeof(struct vmap_block), | |
1567 | gfp_mask & GFP_RECLAIM_MASK, node); | |
1568 | if (unlikely(!vb)) | |
1569 | return ERR_PTR(-ENOMEM); | |
1570 | ||
1571 | va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE, | |
1572 | VMALLOC_START, VMALLOC_END, | |
1573 | node, gfp_mask); | |
1574 | if (IS_ERR(va)) { | |
1575 | kfree(vb); | |
1576 | return ERR_CAST(va); | |
1577 | } | |
1578 | ||
1579 | err = radix_tree_preload(gfp_mask); | |
1580 | if (unlikely(err)) { | |
1581 | kfree(vb); | |
1582 | free_vmap_area(va); | |
1583 | return ERR_PTR(err); | |
1584 | } | |
1585 | ||
1586 | vaddr = vmap_block_vaddr(va->va_start, 0); | |
1587 | spin_lock_init(&vb->lock); | |
1588 | vb->va = va; | |
1589 | /* At least something should be left free */ | |
1590 | BUG_ON(VMAP_BBMAP_BITS <= (1UL << order)); | |
1591 | vb->free = VMAP_BBMAP_BITS - (1UL << order); | |
1592 | vb->dirty = 0; | |
1593 | vb->dirty_min = VMAP_BBMAP_BITS; | |
1594 | vb->dirty_max = 0; | |
1595 | INIT_LIST_HEAD(&vb->free_list); | |
1596 | ||
1597 | vb_idx = addr_to_vb_idx(va->va_start); | |
1598 | spin_lock(&vmap_block_tree_lock); | |
1599 | err = radix_tree_insert(&vmap_block_tree, vb_idx, vb); | |
1600 | spin_unlock(&vmap_block_tree_lock); | |
1601 | BUG_ON(err); | |
1602 | radix_tree_preload_end(); | |
1603 | ||
1604 | vbq = &get_cpu_var(vmap_block_queue); | |
1605 | spin_lock(&vbq->lock); | |
1606 | list_add_tail_rcu(&vb->free_list, &vbq->free); | |
1607 | spin_unlock(&vbq->lock); | |
1608 | put_cpu_var(vmap_block_queue); | |
1609 | ||
1610 | return vaddr; | |
1611 | } | |
1612 | ||
1613 | static void free_vmap_block(struct vmap_block *vb) | |
1614 | { | |
1615 | struct vmap_block *tmp; | |
1616 | unsigned long vb_idx; | |
1617 | ||
1618 | vb_idx = addr_to_vb_idx(vb->va->va_start); | |
1619 | spin_lock(&vmap_block_tree_lock); | |
1620 | tmp = radix_tree_delete(&vmap_block_tree, vb_idx); | |
1621 | spin_unlock(&vmap_block_tree_lock); | |
1622 | BUG_ON(tmp != vb); | |
1623 | ||
1624 | free_vmap_area_noflush(vb->va); | |
1625 | kfree_rcu(vb, rcu_head); | |
1626 | } | |
1627 | ||
1628 | static void purge_fragmented_blocks(int cpu) | |
1629 | { | |
1630 | LIST_HEAD(purge); | |
1631 | struct vmap_block *vb; | |
1632 | struct vmap_block *n_vb; | |
1633 | struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu); | |
1634 | ||
1635 | rcu_read_lock(); | |
1636 | list_for_each_entry_rcu(vb, &vbq->free, free_list) { | |
1637 | ||
1638 | if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS)) | |
1639 | continue; | |
1640 | ||
1641 | spin_lock(&vb->lock); | |
1642 | if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) { | |
1643 | vb->free = 0; /* prevent further allocs after releasing lock */ | |
1644 | vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */ | |
1645 | vb->dirty_min = 0; | |
1646 | vb->dirty_max = VMAP_BBMAP_BITS; | |
1647 | spin_lock(&vbq->lock); | |
1648 | list_del_rcu(&vb->free_list); | |
1649 | spin_unlock(&vbq->lock); | |
1650 | spin_unlock(&vb->lock); | |
1651 | list_add_tail(&vb->purge, &purge); | |
1652 | } else | |
1653 | spin_unlock(&vb->lock); | |
1654 | } | |
1655 | rcu_read_unlock(); | |
1656 | ||
1657 | list_for_each_entry_safe(vb, n_vb, &purge, purge) { | |
1658 | list_del(&vb->purge); | |
1659 | free_vmap_block(vb); | |
1660 | } | |
1661 | } | |
1662 | ||
1663 | static void purge_fragmented_blocks_allcpus(void) | |
1664 | { | |
1665 | int cpu; | |
1666 | ||
1667 | for_each_possible_cpu(cpu) | |
1668 | purge_fragmented_blocks(cpu); | |
1669 | } | |
1670 | ||
1671 | static void *vb_alloc(unsigned long size, gfp_t gfp_mask) | |
1672 | { | |
1673 | struct vmap_block_queue *vbq; | |
1674 | struct vmap_block *vb; | |
1675 | void *vaddr = NULL; | |
1676 | unsigned int order; | |
1677 | ||
1678 | BUG_ON(offset_in_page(size)); | |
1679 | BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC); | |
1680 | if (WARN_ON(size == 0)) { | |
1681 | /* | |
1682 | * Allocating 0 bytes isn't what caller wants since | |
1683 | * get_order(0) returns funny result. Just warn and terminate | |
1684 | * early. | |
1685 | */ | |
1686 | return NULL; | |
1687 | } | |
1688 | order = get_order(size); | |
1689 | ||
1690 | rcu_read_lock(); | |
1691 | vbq = &get_cpu_var(vmap_block_queue); | |
1692 | list_for_each_entry_rcu(vb, &vbq->free, free_list) { | |
1693 | unsigned long pages_off; | |
1694 | ||
1695 | spin_lock(&vb->lock); | |
1696 | if (vb->free < (1UL << order)) { | |
1697 | spin_unlock(&vb->lock); | |
1698 | continue; | |
1699 | } | |
1700 | ||
1701 | pages_off = VMAP_BBMAP_BITS - vb->free; | |
1702 | vaddr = vmap_block_vaddr(vb->va->va_start, pages_off); | |
1703 | vb->free -= 1UL << order; | |
1704 | if (vb->free == 0) { | |
1705 | spin_lock(&vbq->lock); | |
1706 | list_del_rcu(&vb->free_list); | |
1707 | spin_unlock(&vbq->lock); | |
1708 | } | |
1709 | ||
1710 | spin_unlock(&vb->lock); | |
1711 | break; | |
1712 | } | |
1713 | ||
1714 | put_cpu_var(vmap_block_queue); | |
1715 | rcu_read_unlock(); | |
1716 | ||
1717 | /* Allocate new block if nothing was found */ | |
1718 | if (!vaddr) | |
1719 | vaddr = new_vmap_block(order, gfp_mask); | |
1720 | ||
1721 | return vaddr; | |
1722 | } | |
1723 | ||
1724 | static void vb_free(unsigned long addr, unsigned long size) | |
1725 | { | |
1726 | unsigned long offset; | |
1727 | unsigned long vb_idx; | |
1728 | unsigned int order; | |
1729 | struct vmap_block *vb; | |
1730 | ||
1731 | BUG_ON(offset_in_page(size)); | |
1732 | BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC); | |
1733 | ||
1734 | flush_cache_vunmap(addr, addr + size); | |
1735 | ||
1736 | order = get_order(size); | |
1737 | ||
1738 | offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT; | |
1739 | ||
1740 | vb_idx = addr_to_vb_idx(addr); | |
1741 | rcu_read_lock(); | |
1742 | vb = radix_tree_lookup(&vmap_block_tree, vb_idx); | |
1743 | rcu_read_unlock(); | |
1744 | BUG_ON(!vb); | |
1745 | ||
1746 | unmap_kernel_range_noflush(addr, size); | |
1747 | ||
1748 | if (debug_pagealloc_enabled_static()) | |
1749 | flush_tlb_kernel_range(addr, addr + size); | |
1750 | ||
1751 | spin_lock(&vb->lock); | |
1752 | ||
1753 | /* Expand dirty range */ | |
1754 | vb->dirty_min = min(vb->dirty_min, offset); | |
1755 | vb->dirty_max = max(vb->dirty_max, offset + (1UL << order)); | |
1756 | ||
1757 | vb->dirty += 1UL << order; | |
1758 | if (vb->dirty == VMAP_BBMAP_BITS) { | |
1759 | BUG_ON(vb->free); | |
1760 | spin_unlock(&vb->lock); | |
1761 | free_vmap_block(vb); | |
1762 | } else | |
1763 | spin_unlock(&vb->lock); | |
1764 | } | |
1765 | ||
1766 | static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush) | |
1767 | { | |
1768 | int cpu; | |
1769 | ||
1770 | if (unlikely(!vmap_initialized)) | |
1771 | return; | |
1772 | ||
1773 | might_sleep(); | |
1774 | ||
1775 | for_each_possible_cpu(cpu) { | |
1776 | struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu); | |
1777 | struct vmap_block *vb; | |
1778 | ||
1779 | rcu_read_lock(); | |
1780 | list_for_each_entry_rcu(vb, &vbq->free, free_list) { | |
1781 | spin_lock(&vb->lock); | |
1782 | if (vb->dirty) { | |
1783 | unsigned long va_start = vb->va->va_start; | |
1784 | unsigned long s, e; | |
1785 | ||
1786 | s = va_start + (vb->dirty_min << PAGE_SHIFT); | |
1787 | e = va_start + (vb->dirty_max << PAGE_SHIFT); | |
1788 | ||
1789 | start = min(s, start); | |
1790 | end = max(e, end); | |
1791 | ||
1792 | flush = 1; | |
1793 | } | |
1794 | spin_unlock(&vb->lock); | |
1795 | } | |
1796 | rcu_read_unlock(); | |
1797 | } | |
1798 | ||
1799 | mutex_lock(&vmap_purge_lock); | |
1800 | purge_fragmented_blocks_allcpus(); | |
1801 | if (!__purge_vmap_area_lazy(start, end) && flush) | |
1802 | flush_tlb_kernel_range(start, end); | |
1803 | mutex_unlock(&vmap_purge_lock); | |
1804 | } | |
1805 | ||
1806 | /** | |
1807 | * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer | |
1808 | * | |
1809 | * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily | |
1810 | * to amortize TLB flushing overheads. What this means is that any page you | |
1811 | * have now, may, in a former life, have been mapped into kernel virtual | |
1812 | * address by the vmap layer and so there might be some CPUs with TLB entries | |
1813 | * still referencing that page (additional to the regular 1:1 kernel mapping). | |
1814 | * | |
1815 | * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can | |
1816 | * be sure that none of the pages we have control over will have any aliases | |
1817 | * from the vmap layer. | |
1818 | */ | |
1819 | void vm_unmap_aliases(void) | |
1820 | { | |
1821 | unsigned long start = ULONG_MAX, end = 0; | |
1822 | int flush = 0; | |
1823 | ||
1824 | _vm_unmap_aliases(start, end, flush); | |
1825 | } | |
1826 | EXPORT_SYMBOL_GPL(vm_unmap_aliases); | |
1827 | ||
1828 | /** | |
1829 | * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram | |
1830 | * @mem: the pointer returned by vm_map_ram | |
1831 | * @count: the count passed to that vm_map_ram call (cannot unmap partial) | |
1832 | */ | |
1833 | void vm_unmap_ram(const void *mem, unsigned int count) | |
1834 | { | |
1835 | unsigned long size = (unsigned long)count << PAGE_SHIFT; | |
1836 | unsigned long addr = (unsigned long)mem; | |
1837 | struct vmap_area *va; | |
1838 | ||
1839 | might_sleep(); | |
1840 | BUG_ON(!addr); | |
1841 | BUG_ON(addr < VMALLOC_START); | |
1842 | BUG_ON(addr > VMALLOC_END); | |
1843 | BUG_ON(!PAGE_ALIGNED(addr)); | |
1844 | ||
1845 | kasan_poison_vmalloc(mem, size); | |
1846 | ||
1847 | if (likely(count <= VMAP_MAX_ALLOC)) { | |
1848 | debug_check_no_locks_freed(mem, size); | |
1849 | vb_free(addr, size); | |
1850 | return; | |
1851 | } | |
1852 | ||
1853 | va = find_vmap_area(addr); | |
1854 | BUG_ON(!va); | |
1855 | debug_check_no_locks_freed((void *)va->va_start, | |
1856 | (va->va_end - va->va_start)); | |
1857 | free_unmap_vmap_area(va); | |
1858 | } | |
1859 | EXPORT_SYMBOL(vm_unmap_ram); | |
1860 | ||
1861 | /** | |
1862 | * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space) | |
1863 | * @pages: an array of pointers to the pages to be mapped | |
1864 | * @count: number of pages | |
1865 | * @node: prefer to allocate data structures on this node | |
1866 | * | |
1867 | * If you use this function for less than VMAP_MAX_ALLOC pages, it could be | |
1868 | * faster than vmap so it's good. But if you mix long-life and short-life | |
1869 | * objects with vm_map_ram(), it could consume lots of address space through | |
1870 | * fragmentation (especially on a 32bit machine). You could see failures in | |
1871 | * the end. Please use this function for short-lived objects. | |
1872 | * | |
1873 | * Returns: a pointer to the address that has been mapped, or %NULL on failure | |
1874 | */ | |
1875 | void *vm_map_ram(struct page **pages, unsigned int count, int node) | |
1876 | { | |
1877 | unsigned long size = (unsigned long)count << PAGE_SHIFT; | |
1878 | unsigned long addr; | |
1879 | void *mem; | |
1880 | ||
1881 | if (likely(count <= VMAP_MAX_ALLOC)) { | |
1882 | mem = vb_alloc(size, GFP_KERNEL); | |
1883 | if (IS_ERR(mem)) | |
1884 | return NULL; | |
1885 | addr = (unsigned long)mem; | |
1886 | } else { | |
1887 | struct vmap_area *va; | |
1888 | va = alloc_vmap_area(size, PAGE_SIZE, | |
1889 | VMALLOC_START, VMALLOC_END, node, GFP_KERNEL); | |
1890 | if (IS_ERR(va)) | |
1891 | return NULL; | |
1892 | ||
1893 | addr = va->va_start; | |
1894 | mem = (void *)addr; | |
1895 | } | |
1896 | ||
1897 | kasan_unpoison_vmalloc(mem, size); | |
1898 | ||
1899 | if (map_kernel_range(addr, size, PAGE_KERNEL, pages) < 0) { | |
1900 | vm_unmap_ram(mem, count); | |
1901 | return NULL; | |
1902 | } | |
1903 | return mem; | |
1904 | } | |
1905 | EXPORT_SYMBOL(vm_map_ram); | |
1906 | ||
1907 | static struct vm_struct *vmlist __initdata; | |
1908 | ||
1909 | /** | |
1910 | * vm_area_add_early - add vmap area early during boot | |
1911 | * @vm: vm_struct to add | |
1912 | * | |
1913 | * This function is used to add fixed kernel vm area to vmlist before | |
1914 | * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags | |
1915 | * should contain proper values and the other fields should be zero. | |
1916 | * | |
1917 | * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING. | |
1918 | */ | |
1919 | void __init vm_area_add_early(struct vm_struct *vm) | |
1920 | { | |
1921 | struct vm_struct *tmp, **p; | |
1922 | ||
1923 | BUG_ON(vmap_initialized); | |
1924 | for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) { | |
1925 | if (tmp->addr >= vm->addr) { | |
1926 | BUG_ON(tmp->addr < vm->addr + vm->size); | |
1927 | break; | |
1928 | } else | |
1929 | BUG_ON(tmp->addr + tmp->size > vm->addr); | |
1930 | } | |
1931 | vm->next = *p; | |
1932 | *p = vm; | |
1933 | } | |
1934 | ||
1935 | /** | |
1936 | * vm_area_register_early - register vmap area early during boot | |
1937 | * @vm: vm_struct to register | |
1938 | * @align: requested alignment | |
1939 | * | |
1940 | * This function is used to register kernel vm area before | |
1941 | * vmalloc_init() is called. @vm->size and @vm->flags should contain | |
1942 | * proper values on entry and other fields should be zero. On return, | |
1943 | * vm->addr contains the allocated address. | |
1944 | * | |
1945 | * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING. | |
1946 | */ | |
1947 | void __init vm_area_register_early(struct vm_struct *vm, size_t align) | |
1948 | { | |
1949 | static size_t vm_init_off __initdata; | |
1950 | unsigned long addr; | |
1951 | ||
1952 | addr = ALIGN(VMALLOC_START + vm_init_off, align); | |
1953 | vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START; | |
1954 | ||
1955 | vm->addr = (void *)addr; | |
1956 | ||
1957 | vm_area_add_early(vm); | |
1958 | } | |
1959 | ||
1960 | static void vmap_init_free_space(void) | |
1961 | { | |
1962 | unsigned long vmap_start = 1; | |
1963 | const unsigned long vmap_end = ULONG_MAX; | |
1964 | struct vmap_area *busy, *free; | |
1965 | ||
1966 | /* | |
1967 | * B F B B B F | |
1968 | * -|-----|.....|-----|-----|-----|.....|- | |
1969 | * | The KVA space | | |
1970 | * |<--------------------------------->| | |
1971 | */ | |
1972 | list_for_each_entry(busy, &vmap_area_list, list) { | |
1973 | if (busy->va_start - vmap_start > 0) { | |
1974 | free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); | |
1975 | if (!WARN_ON_ONCE(!free)) { | |
1976 | free->va_start = vmap_start; | |
1977 | free->va_end = busy->va_start; | |
1978 | ||
1979 | insert_vmap_area_augment(free, NULL, | |
1980 | &free_vmap_area_root, | |
1981 | &free_vmap_area_list); | |
1982 | } | |
1983 | } | |
1984 | ||
1985 | vmap_start = busy->va_end; | |
1986 | } | |
1987 | ||
1988 | if (vmap_end - vmap_start > 0) { | |
1989 | free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); | |
1990 | if (!WARN_ON_ONCE(!free)) { | |
1991 | free->va_start = vmap_start; | |
1992 | free->va_end = vmap_end; | |
1993 | ||
1994 | insert_vmap_area_augment(free, NULL, | |
1995 | &free_vmap_area_root, | |
1996 | &free_vmap_area_list); | |
1997 | } | |
1998 | } | |
1999 | } | |
2000 | ||
2001 | void __init vmalloc_init(void) | |
2002 | { | |
2003 | struct vmap_area *va; | |
2004 | struct vm_struct *tmp; | |
2005 | int i; | |
2006 | ||
2007 | /* | |
2008 | * Create the cache for vmap_area objects. | |
2009 | */ | |
2010 | vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC); | |
2011 | ||
2012 | for_each_possible_cpu(i) { | |
2013 | struct vmap_block_queue *vbq; | |
2014 | struct vfree_deferred *p; | |
2015 | ||
2016 | vbq = &per_cpu(vmap_block_queue, i); | |
2017 | spin_lock_init(&vbq->lock); | |
2018 | INIT_LIST_HEAD(&vbq->free); | |
2019 | p = &per_cpu(vfree_deferred, i); | |
2020 | init_llist_head(&p->list); | |
2021 | INIT_WORK(&p->wq, free_work); | |
2022 | } | |
2023 | ||
2024 | /* Import existing vmlist entries. */ | |
2025 | for (tmp = vmlist; tmp; tmp = tmp->next) { | |
2026 | va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); | |
2027 | if (WARN_ON_ONCE(!va)) | |
2028 | continue; | |
2029 | ||
2030 | va->va_start = (unsigned long)tmp->addr; | |
2031 | va->va_end = va->va_start + tmp->size; | |
2032 | va->vm = tmp; | |
2033 | insert_vmap_area(va, &vmap_area_root, &vmap_area_list); | |
2034 | } | |
2035 | ||
2036 | /* | |
2037 | * Now we can initialize a free vmap space. | |
2038 | */ | |
2039 | vmap_init_free_space(); | |
2040 | vmap_initialized = true; | |
2041 | } | |
2042 | ||
2043 | /** | |
2044 | * unmap_kernel_range - unmap kernel VM area and flush cache and TLB | |
2045 | * @addr: start of the VM area to unmap | |
2046 | * @size: size of the VM area to unmap | |
2047 | * | |
2048 | * Similar to unmap_kernel_range_noflush() but flushes vcache before | |
2049 | * the unmapping and tlb after. | |
2050 | */ | |
2051 | void unmap_kernel_range(unsigned long addr, unsigned long size) | |
2052 | { | |
2053 | unsigned long end = addr + size; | |
2054 | ||
2055 | flush_cache_vunmap(addr, end); | |
2056 | unmap_kernel_range_noflush(addr, size); | |
2057 | flush_tlb_kernel_range(addr, end); | |
2058 | } | |
2059 | ||
2060 | static inline void setup_vmalloc_vm_locked(struct vm_struct *vm, | |
2061 | struct vmap_area *va, unsigned long flags, const void *caller) | |
2062 | { | |
2063 | vm->flags = flags; | |
2064 | vm->addr = (void *)va->va_start; | |
2065 | vm->size = va->va_end - va->va_start; | |
2066 | vm->caller = caller; | |
2067 | va->vm = vm; | |
2068 | } | |
2069 | ||
2070 | static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va, | |
2071 | unsigned long flags, const void *caller) | |
2072 | { | |
2073 | spin_lock(&vmap_area_lock); | |
2074 | setup_vmalloc_vm_locked(vm, va, flags, caller); | |
2075 | spin_unlock(&vmap_area_lock); | |
2076 | } | |
2077 | ||
2078 | static void clear_vm_uninitialized_flag(struct vm_struct *vm) | |
2079 | { | |
2080 | /* | |
2081 | * Before removing VM_UNINITIALIZED, | |
2082 | * we should make sure that vm has proper values. | |
2083 | * Pair with smp_rmb() in show_numa_info(). | |
2084 | */ | |
2085 | smp_wmb(); | |
2086 | vm->flags &= ~VM_UNINITIALIZED; | |
2087 | } | |
2088 | ||
2089 | static struct vm_struct *__get_vm_area_node(unsigned long size, | |
2090 | unsigned long align, unsigned long flags, unsigned long start, | |
2091 | unsigned long end, int node, gfp_t gfp_mask, const void *caller) | |
2092 | { | |
2093 | struct vmap_area *va; | |
2094 | struct vm_struct *area; | |
2095 | unsigned long requested_size = size; | |
2096 | ||
2097 | BUG_ON(in_interrupt()); | |
2098 | size = PAGE_ALIGN(size); | |
2099 | if (unlikely(!size)) | |
2100 | return NULL; | |
2101 | ||
2102 | if (flags & VM_IOREMAP) | |
2103 | align = 1ul << clamp_t(int, get_count_order_long(size), | |
2104 | PAGE_SHIFT, IOREMAP_MAX_ORDER); | |
2105 | ||
2106 | area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node); | |
2107 | if (unlikely(!area)) | |
2108 | return NULL; | |
2109 | ||
2110 | if (!(flags & VM_NO_GUARD)) | |
2111 | size += PAGE_SIZE; | |
2112 | ||
2113 | va = alloc_vmap_area(size, align, start, end, node, gfp_mask); | |
2114 | if (IS_ERR(va)) { | |
2115 | kfree(area); | |
2116 | return NULL; | |
2117 | } | |
2118 | ||
2119 | kasan_unpoison_vmalloc((void *)va->va_start, requested_size); | |
2120 | ||
2121 | setup_vmalloc_vm(area, va, flags, caller); | |
2122 | ||
2123 | return area; | |
2124 | } | |
2125 | ||
2126 | struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags, | |
2127 | unsigned long start, unsigned long end, | |
2128 | const void *caller) | |
2129 | { | |
2130 | return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE, | |
2131 | GFP_KERNEL, caller); | |
2132 | } | |
2133 | ||
2134 | /** | |
2135 | * get_vm_area - reserve a contiguous kernel virtual area | |
2136 | * @size: size of the area | |
2137 | * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC | |
2138 | * | |
2139 | * Search an area of @size in the kernel virtual mapping area, | |
2140 | * and reserved it for out purposes. Returns the area descriptor | |
2141 | * on success or %NULL on failure. | |
2142 | * | |
2143 | * Return: the area descriptor on success or %NULL on failure. | |
2144 | */ | |
2145 | struct vm_struct *get_vm_area(unsigned long size, unsigned long flags) | |
2146 | { | |
2147 | return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END, | |
2148 | NUMA_NO_NODE, GFP_KERNEL, | |
2149 | __builtin_return_address(0)); | |
2150 | } | |
2151 | ||
2152 | struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags, | |
2153 | const void *caller) | |
2154 | { | |
2155 | return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END, | |
2156 | NUMA_NO_NODE, GFP_KERNEL, caller); | |
2157 | } | |
2158 | ||
2159 | /** | |
2160 | * find_vm_area - find a continuous kernel virtual area | |
2161 | * @addr: base address | |
2162 | * | |
2163 | * Search for the kernel VM area starting at @addr, and return it. | |
2164 | * It is up to the caller to do all required locking to keep the returned | |
2165 | * pointer valid. | |
2166 | * | |
2167 | * Return: pointer to the found area or %NULL on faulure | |
2168 | */ | |
2169 | struct vm_struct *find_vm_area(const void *addr) | |
2170 | { | |
2171 | struct vmap_area *va; | |
2172 | ||
2173 | va = find_vmap_area((unsigned long)addr); | |
2174 | if (!va) | |
2175 | return NULL; | |
2176 | ||
2177 | return va->vm; | |
2178 | } | |
2179 | ||
2180 | /** | |
2181 | * remove_vm_area - find and remove a continuous kernel virtual area | |
2182 | * @addr: base address | |
2183 | * | |
2184 | * Search for the kernel VM area starting at @addr, and remove it. | |
2185 | * This function returns the found VM area, but using it is NOT safe | |
2186 | * on SMP machines, except for its size or flags. | |
2187 | * | |
2188 | * Return: pointer to the found area or %NULL on faulure | |
2189 | */ | |
2190 | struct vm_struct *remove_vm_area(const void *addr) | |
2191 | { | |
2192 | struct vmap_area *va; | |
2193 | ||
2194 | might_sleep(); | |
2195 | ||
2196 | spin_lock(&vmap_area_lock); | |
2197 | va = __find_vmap_area((unsigned long)addr); | |
2198 | if (va && va->vm) { | |
2199 | struct vm_struct *vm = va->vm; | |
2200 | ||
2201 | va->vm = NULL; | |
2202 | spin_unlock(&vmap_area_lock); | |
2203 | ||
2204 | kasan_free_shadow(vm); | |
2205 | free_unmap_vmap_area(va); | |
2206 | ||
2207 | return vm; | |
2208 | } | |
2209 | ||
2210 | spin_unlock(&vmap_area_lock); | |
2211 | return NULL; | |
2212 | } | |
2213 | ||
2214 | static inline void set_area_direct_map(const struct vm_struct *area, | |
2215 | int (*set_direct_map)(struct page *page)) | |
2216 | { | |
2217 | int i; | |
2218 | ||
2219 | for (i = 0; i < area->nr_pages; i++) | |
2220 | if (page_address(area->pages[i])) | |
2221 | set_direct_map(area->pages[i]); | |
2222 | } | |
2223 | ||
2224 | /* Handle removing and resetting vm mappings related to the vm_struct. */ | |
2225 | static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) | |
2226 | { | |
2227 | unsigned long start = ULONG_MAX, end = 0; | |
2228 | int flush_reset = area->flags & VM_FLUSH_RESET_PERMS; | |
2229 | int flush_dmap = 0; | |
2230 | int i; | |
2231 | ||
2232 | remove_vm_area(area->addr); | |
2233 | ||
2234 | /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */ | |
2235 | if (!flush_reset) | |
2236 | return; | |
2237 | ||
2238 | /* | |
2239 | * If not deallocating pages, just do the flush of the VM area and | |
2240 | * return. | |
2241 | */ | |
2242 | if (!deallocate_pages) { | |
2243 | vm_unmap_aliases(); | |
2244 | return; | |
2245 | } | |
2246 | ||
2247 | /* | |
2248 | * If execution gets here, flush the vm mapping and reset the direct | |
2249 | * map. Find the start and end range of the direct mappings to make sure | |
2250 | * the vm_unmap_aliases() flush includes the direct map. | |
2251 | */ | |
2252 | for (i = 0; i < area->nr_pages; i++) { | |
2253 | unsigned long addr = (unsigned long)page_address(area->pages[i]); | |
2254 | if (addr) { | |
2255 | start = min(addr, start); | |
2256 | end = max(addr + PAGE_SIZE, end); | |
2257 | flush_dmap = 1; | |
2258 | } | |
2259 | } | |
2260 | ||
2261 | /* | |
2262 | * Set direct map to something invalid so that it won't be cached if | |
2263 | * there are any accesses after the TLB flush, then flush the TLB and | |
2264 | * reset the direct map permissions to the default. | |
2265 | */ | |
2266 | set_area_direct_map(area, set_direct_map_invalid_noflush); | |
2267 | _vm_unmap_aliases(start, end, flush_dmap); | |
2268 | set_area_direct_map(area, set_direct_map_default_noflush); | |
2269 | } | |
2270 | ||
2271 | static void __vunmap(const void *addr, int deallocate_pages) | |
2272 | { | |
2273 | struct vm_struct *area; | |
2274 | ||
2275 | if (!addr) | |
2276 | return; | |
2277 | ||
2278 | if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n", | |
2279 | addr)) | |
2280 | return; | |
2281 | ||
2282 | area = find_vm_area(addr); | |
2283 | if (unlikely(!area)) { | |
2284 | WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", | |
2285 | addr); | |
2286 | return; | |
2287 | } | |
2288 | ||
2289 | debug_check_no_locks_freed(area->addr, get_vm_area_size(area)); | |
2290 | debug_check_no_obj_freed(area->addr, get_vm_area_size(area)); | |
2291 | ||
2292 | kasan_poison_vmalloc(area->addr, area->size); | |
2293 | ||
2294 | vm_remove_mappings(area, deallocate_pages); | |
2295 | ||
2296 | if (deallocate_pages) { | |
2297 | int i; | |
2298 | ||
2299 | for (i = 0; i < area->nr_pages; i++) { | |
2300 | struct page *page = area->pages[i]; | |
2301 | ||
2302 | BUG_ON(!page); | |
2303 | __free_pages(page, 0); | |
2304 | } | |
2305 | atomic_long_sub(area->nr_pages, &nr_vmalloc_pages); | |
2306 | ||
2307 | kvfree(area->pages); | |
2308 | } | |
2309 | ||
2310 | kfree(area); | |
2311 | return; | |
2312 | } | |
2313 | ||
2314 | static inline void __vfree_deferred(const void *addr) | |
2315 | { | |
2316 | /* | |
2317 | * Use raw_cpu_ptr() because this can be called from preemptible | |
2318 | * context. Preemption is absolutely fine here, because the llist_add() | |
2319 | * implementation is lockless, so it works even if we are adding to | |
2320 | * another cpu's list. schedule_work() should be fine with this too. | |
2321 | */ | |
2322 | struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred); | |
2323 | ||
2324 | if (llist_add((struct llist_node *)addr, &p->list)) | |
2325 | schedule_work(&p->wq); | |
2326 | } | |
2327 | ||
2328 | /** | |
2329 | * vfree_atomic - release memory allocated by vmalloc() | |
2330 | * @addr: memory base address | |
2331 | * | |
2332 | * This one is just like vfree() but can be called in any atomic context | |
2333 | * except NMIs. | |
2334 | */ | |
2335 | void vfree_atomic(const void *addr) | |
2336 | { | |
2337 | BUG_ON(in_nmi()); | |
2338 | ||
2339 | kmemleak_free(addr); | |
2340 | ||
2341 | if (!addr) | |
2342 | return; | |
2343 | __vfree_deferred(addr); | |
2344 | } | |
2345 | ||
2346 | static void __vfree(const void *addr) | |
2347 | { | |
2348 | if (unlikely(in_interrupt())) | |
2349 | __vfree_deferred(addr); | |
2350 | else | |
2351 | __vunmap(addr, 1); | |
2352 | } | |
2353 | ||
2354 | /** | |
2355 | * vfree - release memory allocated by vmalloc() | |
2356 | * @addr: memory base address | |
2357 | * | |
2358 | * Free the virtually continuous memory area starting at @addr, as | |
2359 | * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is | |
2360 | * NULL, no operation is performed. | |
2361 | * | |
2362 | * Must not be called in NMI context (strictly speaking, only if we don't | |
2363 | * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling | |
2364 | * conventions for vfree() arch-depenedent would be a really bad idea) | |
2365 | * | |
2366 | * May sleep if called *not* from interrupt context. | |
2367 | * | |
2368 | * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node) | |
2369 | */ | |
2370 | void vfree(const void *addr) | |
2371 | { | |
2372 | BUG_ON(in_nmi()); | |
2373 | ||
2374 | kmemleak_free(addr); | |
2375 | ||
2376 | might_sleep_if(!in_interrupt()); | |
2377 | ||
2378 | if (!addr) | |
2379 | return; | |
2380 | ||
2381 | __vfree(addr); | |
2382 | } | |
2383 | EXPORT_SYMBOL(vfree); | |
2384 | ||
2385 | /** | |
2386 | * vunmap - release virtual mapping obtained by vmap() | |
2387 | * @addr: memory base address | |
2388 | * | |
2389 | * Free the virtually contiguous memory area starting at @addr, | |
2390 | * which was created from the page array passed to vmap(). | |
2391 | * | |
2392 | * Must not be called in interrupt context. | |
2393 | */ | |
2394 | void vunmap(const void *addr) | |
2395 | { | |
2396 | BUG_ON(in_interrupt()); | |
2397 | might_sleep(); | |
2398 | if (addr) | |
2399 | __vunmap(addr, 0); | |
2400 | } | |
2401 | EXPORT_SYMBOL(vunmap); | |
2402 | ||
2403 | /** | |
2404 | * vmap - map an array of pages into virtually contiguous space | |
2405 | * @pages: array of page pointers | |
2406 | * @count: number of pages to map | |
2407 | * @flags: vm_area->flags | |
2408 | * @prot: page protection for the mapping | |
2409 | * | |
2410 | * Maps @count pages from @pages into contiguous kernel virtual | |
2411 | * space. | |
2412 | * | |
2413 | * Return: the address of the area or %NULL on failure | |
2414 | */ | |
2415 | void *vmap(struct page **pages, unsigned int count, | |
2416 | unsigned long flags, pgprot_t prot) | |
2417 | { | |
2418 | struct vm_struct *area; | |
2419 | unsigned long size; /* In bytes */ | |
2420 | ||
2421 | might_sleep(); | |
2422 | ||
2423 | if (count > totalram_pages()) | |
2424 | return NULL; | |
2425 | ||
2426 | size = (unsigned long)count << PAGE_SHIFT; | |
2427 | area = get_vm_area_caller(size, flags, __builtin_return_address(0)); | |
2428 | if (!area) | |
2429 | return NULL; | |
2430 | ||
2431 | if (map_kernel_range((unsigned long)area->addr, size, pgprot_nx(prot), | |
2432 | pages) < 0) { | |
2433 | vunmap(area->addr); | |
2434 | return NULL; | |
2435 | } | |
2436 | ||
2437 | return area->addr; | |
2438 | } | |
2439 | EXPORT_SYMBOL(vmap); | |
2440 | ||
2441 | static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask, | |
2442 | pgprot_t prot, int node) | |
2443 | { | |
2444 | struct page **pages; | |
2445 | unsigned int nr_pages, array_size, i; | |
2446 | const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO; | |
2447 | const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN; | |
2448 | const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ? | |
2449 | 0 : | |
2450 | __GFP_HIGHMEM; | |
2451 | ||
2452 | nr_pages = get_vm_area_size(area) >> PAGE_SHIFT; | |
2453 | array_size = (nr_pages * sizeof(struct page *)); | |
2454 | ||
2455 | /* Please note that the recursion is strictly bounded. */ | |
2456 | if (array_size > PAGE_SIZE) { | |
2457 | pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask, | |
2458 | node, area->caller); | |
2459 | } else { | |
2460 | pages = kmalloc_node(array_size, nested_gfp, node); | |
2461 | } | |
2462 | ||
2463 | if (!pages) { | |
2464 | remove_vm_area(area->addr); | |
2465 | kfree(area); | |
2466 | return NULL; | |
2467 | } | |
2468 | ||
2469 | area->pages = pages; | |
2470 | area->nr_pages = nr_pages; | |
2471 | ||
2472 | for (i = 0; i < area->nr_pages; i++) { | |
2473 | struct page *page; | |
2474 | ||
2475 | if (node == NUMA_NO_NODE) | |
2476 | page = alloc_page(alloc_mask|highmem_mask); | |
2477 | else | |
2478 | page = alloc_pages_node(node, alloc_mask|highmem_mask, 0); | |
2479 | ||
2480 | if (unlikely(!page)) { | |
2481 | /* Successfully allocated i pages, free them in __vunmap() */ | |
2482 | area->nr_pages = i; | |
2483 | atomic_long_add(area->nr_pages, &nr_vmalloc_pages); | |
2484 | goto fail; | |
2485 | } | |
2486 | area->pages[i] = page; | |
2487 | if (gfpflags_allow_blocking(gfp_mask)) | |
2488 | cond_resched(); | |
2489 | } | |
2490 | atomic_long_add(area->nr_pages, &nr_vmalloc_pages); | |
2491 | ||
2492 | if (map_kernel_range((unsigned long)area->addr, get_vm_area_size(area), | |
2493 | prot, pages) < 0) | |
2494 | goto fail; | |
2495 | ||
2496 | return area->addr; | |
2497 | ||
2498 | fail: | |
2499 | warn_alloc(gfp_mask, NULL, | |
2500 | "vmalloc: allocation failure, allocated %ld of %ld bytes", | |
2501 | (area->nr_pages*PAGE_SIZE), area->size); | |
2502 | __vfree(area->addr); | |
2503 | return NULL; | |
2504 | } | |
2505 | ||
2506 | /** | |
2507 | * __vmalloc_node_range - allocate virtually contiguous memory | |
2508 | * @size: allocation size | |
2509 | * @align: desired alignment | |
2510 | * @start: vm area range start | |
2511 | * @end: vm area range end | |
2512 | * @gfp_mask: flags for the page level allocator | |
2513 | * @prot: protection mask for the allocated pages | |
2514 | * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD) | |
2515 | * @node: node to use for allocation or NUMA_NO_NODE | |
2516 | * @caller: caller's return address | |
2517 | * | |
2518 | * Allocate enough pages to cover @size from the page level | |
2519 | * allocator with @gfp_mask flags. Map them into contiguous | |
2520 | * kernel virtual space, using a pagetable protection of @prot. | |
2521 | * | |
2522 | * Return: the address of the area or %NULL on failure | |
2523 | */ | |
2524 | void *__vmalloc_node_range(unsigned long size, unsigned long align, | |
2525 | unsigned long start, unsigned long end, gfp_t gfp_mask, | |
2526 | pgprot_t prot, unsigned long vm_flags, int node, | |
2527 | const void *caller) | |
2528 | { | |
2529 | struct vm_struct *area; | |
2530 | void *addr; | |
2531 | unsigned long real_size = size; | |
2532 | ||
2533 | size = PAGE_ALIGN(size); | |
2534 | if (!size || (size >> PAGE_SHIFT) > totalram_pages()) | |
2535 | goto fail; | |
2536 | ||
2537 | area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED | | |
2538 | vm_flags, start, end, node, gfp_mask, caller); | |
2539 | if (!area) | |
2540 | goto fail; | |
2541 | ||
2542 | addr = __vmalloc_area_node(area, gfp_mask, prot, node); | |
2543 | if (!addr) | |
2544 | return NULL; | |
2545 | ||
2546 | /* | |
2547 | * In this function, newly allocated vm_struct has VM_UNINITIALIZED | |
2548 | * flag. It means that vm_struct is not fully initialized. | |
2549 | * Now, it is fully initialized, so remove this flag here. | |
2550 | */ | |
2551 | clear_vm_uninitialized_flag(area); | |
2552 | ||
2553 | kmemleak_vmalloc(area, size, gfp_mask); | |
2554 | ||
2555 | return addr; | |
2556 | ||
2557 | fail: | |
2558 | warn_alloc(gfp_mask, NULL, | |
2559 | "vmalloc: allocation failure: %lu bytes", real_size); | |
2560 | return NULL; | |
2561 | } | |
2562 | ||
2563 | /** | |
2564 | * __vmalloc_node - allocate virtually contiguous memory | |
2565 | * @size: allocation size | |
2566 | * @align: desired alignment | |
2567 | * @gfp_mask: flags for the page level allocator | |
2568 | * @node: node to use for allocation or NUMA_NO_NODE | |
2569 | * @caller: caller's return address | |
2570 | * | |
2571 | * Allocate enough pages to cover @size from the page level allocator with | |
2572 | * @gfp_mask flags. Map them into contiguous kernel virtual space. | |
2573 | * | |
2574 | * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL | |
2575 | * and __GFP_NOFAIL are not supported | |
2576 | * | |
2577 | * Any use of gfp flags outside of GFP_KERNEL should be consulted | |
2578 | * with mm people. | |
2579 | * | |
2580 | * Return: pointer to the allocated memory or %NULL on error | |
2581 | */ | |
2582 | void *__vmalloc_node(unsigned long size, unsigned long align, | |
2583 | gfp_t gfp_mask, int node, const void *caller) | |
2584 | { | |
2585 | return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END, | |
2586 | gfp_mask, PAGE_KERNEL, 0, node, caller); | |
2587 | } | |
2588 | /* | |
2589 | * This is only for performance analysis of vmalloc and stress purpose. | |
2590 | * It is required by vmalloc test module, therefore do not use it other | |
2591 | * than that. | |
2592 | */ | |
2593 | #ifdef CONFIG_TEST_VMALLOC_MODULE | |
2594 | EXPORT_SYMBOL_GPL(__vmalloc_node); | |
2595 | #endif | |
2596 | ||
2597 | void *__vmalloc(unsigned long size, gfp_t gfp_mask) | |
2598 | { | |
2599 | return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE, | |
2600 | __builtin_return_address(0)); | |
2601 | } | |
2602 | EXPORT_SYMBOL(__vmalloc); | |
2603 | ||
2604 | /** | |
2605 | * vmalloc - allocate virtually contiguous memory | |
2606 | * @size: allocation size | |
2607 | * | |
2608 | * Allocate enough pages to cover @size from the page level | |
2609 | * allocator and map them into contiguous kernel virtual space. | |
2610 | * | |
2611 | * For tight control over page level allocator and protection flags | |
2612 | * use __vmalloc() instead. | |
2613 | * | |
2614 | * Return: pointer to the allocated memory or %NULL on error | |
2615 | */ | |
2616 | void *vmalloc(unsigned long size) | |
2617 | { | |
2618 | return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE, | |
2619 | __builtin_return_address(0)); | |
2620 | } | |
2621 | EXPORT_SYMBOL(vmalloc); | |
2622 | ||
2623 | /** | |
2624 | * vzalloc - allocate virtually contiguous memory with zero fill | |
2625 | * @size: allocation size | |
2626 | * | |
2627 | * Allocate enough pages to cover @size from the page level | |
2628 | * allocator and map them into contiguous kernel virtual space. | |
2629 | * The memory allocated is set to zero. | |
2630 | * | |
2631 | * For tight control over page level allocator and protection flags | |
2632 | * use __vmalloc() instead. | |
2633 | * | |
2634 | * Return: pointer to the allocated memory or %NULL on error | |
2635 | */ | |
2636 | void *vzalloc(unsigned long size) | |
2637 | { | |
2638 | return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE, | |
2639 | __builtin_return_address(0)); | |
2640 | } | |
2641 | EXPORT_SYMBOL(vzalloc); | |
2642 | ||
2643 | /** | |
2644 | * vmalloc_user - allocate zeroed virtually contiguous memory for userspace | |
2645 | * @size: allocation size | |
2646 | * | |
2647 | * The resulting memory area is zeroed so it can be mapped to userspace | |
2648 | * without leaking data. | |
2649 | * | |
2650 | * Return: pointer to the allocated memory or %NULL on error | |
2651 | */ | |
2652 | void *vmalloc_user(unsigned long size) | |
2653 | { | |
2654 | return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END, | |
2655 | GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL, | |
2656 | VM_USERMAP, NUMA_NO_NODE, | |
2657 | __builtin_return_address(0)); | |
2658 | } | |
2659 | EXPORT_SYMBOL(vmalloc_user); | |
2660 | ||
2661 | /** | |
2662 | * vmalloc_node - allocate memory on a specific node | |
2663 | * @size: allocation size | |
2664 | * @node: numa node | |
2665 | * | |
2666 | * Allocate enough pages to cover @size from the page level | |
2667 | * allocator and map them into contiguous kernel virtual space. | |
2668 | * | |
2669 | * For tight control over page level allocator and protection flags | |
2670 | * use __vmalloc() instead. | |
2671 | * | |
2672 | * Return: pointer to the allocated memory or %NULL on error | |
2673 | */ | |
2674 | void *vmalloc_node(unsigned long size, int node) | |
2675 | { | |
2676 | return __vmalloc_node(size, 1, GFP_KERNEL, node, | |
2677 | __builtin_return_address(0)); | |
2678 | } | |
2679 | EXPORT_SYMBOL(vmalloc_node); | |
2680 | ||
2681 | /** | |
2682 | * vzalloc_node - allocate memory on a specific node with zero fill | |
2683 | * @size: allocation size | |
2684 | * @node: numa node | |
2685 | * | |
2686 | * Allocate enough pages to cover @size from the page level | |
2687 | * allocator and map them into contiguous kernel virtual space. | |
2688 | * The memory allocated is set to zero. | |
2689 | * | |
2690 | * Return: pointer to the allocated memory or %NULL on error | |
2691 | */ | |
2692 | void *vzalloc_node(unsigned long size, int node) | |
2693 | { | |
2694 | return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node, | |
2695 | __builtin_return_address(0)); | |
2696 | } | |
2697 | EXPORT_SYMBOL(vzalloc_node); | |
2698 | ||
2699 | #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32) | |
2700 | #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL) | |
2701 | #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA) | |
2702 | #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL) | |
2703 | #else | |
2704 | /* | |
2705 | * 64b systems should always have either DMA or DMA32 zones. For others | |
2706 | * GFP_DMA32 should do the right thing and use the normal zone. | |
2707 | */ | |
2708 | #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL | |
2709 | #endif | |
2710 | ||
2711 | /** | |
2712 | * vmalloc_32 - allocate virtually contiguous memory (32bit addressable) | |
2713 | * @size: allocation size | |
2714 | * | |
2715 | * Allocate enough 32bit PA addressable pages to cover @size from the | |
2716 | * page level allocator and map them into contiguous kernel virtual space. | |
2717 | * | |
2718 | * Return: pointer to the allocated memory or %NULL on error | |
2719 | */ | |
2720 | void *vmalloc_32(unsigned long size) | |
2721 | { | |
2722 | return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE, | |
2723 | __builtin_return_address(0)); | |
2724 | } | |
2725 | EXPORT_SYMBOL(vmalloc_32); | |
2726 | ||
2727 | /** | |
2728 | * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory | |
2729 | * @size: allocation size | |
2730 | * | |
2731 | * The resulting memory area is 32bit addressable and zeroed so it can be | |
2732 | * mapped to userspace without leaking data. | |
2733 | * | |
2734 | * Return: pointer to the allocated memory or %NULL on error | |
2735 | */ | |
2736 | void *vmalloc_32_user(unsigned long size) | |
2737 | { | |
2738 | return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END, | |
2739 | GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL, | |
2740 | VM_USERMAP, NUMA_NO_NODE, | |
2741 | __builtin_return_address(0)); | |
2742 | } | |
2743 | EXPORT_SYMBOL(vmalloc_32_user); | |
2744 | ||
2745 | /* | |
2746 | * small helper routine , copy contents to buf from addr. | |
2747 | * If the page is not present, fill zero. | |
2748 | */ | |
2749 | ||
2750 | static int aligned_vread(char *buf, char *addr, unsigned long count) | |
2751 | { | |
2752 | struct page *p; | |
2753 | int copied = 0; | |
2754 | ||
2755 | while (count) { | |
2756 | unsigned long offset, length; | |
2757 | ||
2758 | offset = offset_in_page(addr); | |
2759 | length = PAGE_SIZE - offset; | |
2760 | if (length > count) | |
2761 | length = count; | |
2762 | p = vmalloc_to_page(addr); | |
2763 | /* | |
2764 | * To do safe access to this _mapped_ area, we need | |
2765 | * lock. But adding lock here means that we need to add | |
2766 | * overhead of vmalloc()/vfree() calles for this _debug_ | |
2767 | * interface, rarely used. Instead of that, we'll use | |
2768 | * kmap() and get small overhead in this access function. | |
2769 | */ | |
2770 | if (p) { | |
2771 | /* | |
2772 | * we can expect USER0 is not used (see vread/vwrite's | |
2773 | * function description) | |
2774 | */ | |
2775 | void *map = kmap_atomic(p); | |
2776 | memcpy(buf, map + offset, length); | |
2777 | kunmap_atomic(map); | |
2778 | } else | |
2779 | memset(buf, 0, length); | |
2780 | ||
2781 | addr += length; | |
2782 | buf += length; | |
2783 | copied += length; | |
2784 | count -= length; | |
2785 | } | |
2786 | return copied; | |
2787 | } | |
2788 | ||
2789 | static int aligned_vwrite(char *buf, char *addr, unsigned long count) | |
2790 | { | |
2791 | struct page *p; | |
2792 | int copied = 0; | |
2793 | ||
2794 | while (count) { | |
2795 | unsigned long offset, length; | |
2796 | ||
2797 | offset = offset_in_page(addr); | |
2798 | length = PAGE_SIZE - offset; | |
2799 | if (length > count) | |
2800 | length = count; | |
2801 | p = vmalloc_to_page(addr); | |
2802 | /* | |
2803 | * To do safe access to this _mapped_ area, we need | |
2804 | * lock. But adding lock here means that we need to add | |
2805 | * overhead of vmalloc()/vfree() calles for this _debug_ | |
2806 | * interface, rarely used. Instead of that, we'll use | |
2807 | * kmap() and get small overhead in this access function. | |
2808 | */ | |
2809 | if (p) { | |
2810 | /* | |
2811 | * we can expect USER0 is not used (see vread/vwrite's | |
2812 | * function description) | |
2813 | */ | |
2814 | void *map = kmap_atomic(p); | |
2815 | memcpy(map + offset, buf, length); | |
2816 | kunmap_atomic(map); | |
2817 | } | |
2818 | addr += length; | |
2819 | buf += length; | |
2820 | copied += length; | |
2821 | count -= length; | |
2822 | } | |
2823 | return copied; | |
2824 | } | |
2825 | ||
2826 | /** | |
2827 | * vread() - read vmalloc area in a safe way. | |
2828 | * @buf: buffer for reading data | |
2829 | * @addr: vm address. | |
2830 | * @count: number of bytes to be read. | |
2831 | * | |
2832 | * This function checks that addr is a valid vmalloc'ed area, and | |
2833 | * copy data from that area to a given buffer. If the given memory range | |
2834 | * of [addr...addr+count) includes some valid address, data is copied to | |
2835 | * proper area of @buf. If there are memory holes, they'll be zero-filled. | |
2836 | * IOREMAP area is treated as memory hole and no copy is done. | |
2837 | * | |
2838 | * If [addr...addr+count) doesn't includes any intersects with alive | |
2839 | * vm_struct area, returns 0. @buf should be kernel's buffer. | |
2840 | * | |
2841 | * Note: In usual ops, vread() is never necessary because the caller | |
2842 | * should know vmalloc() area is valid and can use memcpy(). | |
2843 | * This is for routines which have to access vmalloc area without | |
2844 | * any information, as /dev/kmem. | |
2845 | * | |
2846 | * Return: number of bytes for which addr and buf should be increased | |
2847 | * (same number as @count) or %0 if [addr...addr+count) doesn't | |
2848 | * include any intersection with valid vmalloc area | |
2849 | */ | |
2850 | long vread(char *buf, char *addr, unsigned long count) | |
2851 | { | |
2852 | struct vmap_area *va; | |
2853 | struct vm_struct *vm; | |
2854 | char *vaddr, *buf_start = buf; | |
2855 | unsigned long buflen = count; | |
2856 | unsigned long n; | |
2857 | ||
2858 | /* Don't allow overflow */ | |
2859 | if ((unsigned long) addr + count < count) | |
2860 | count = -(unsigned long) addr; | |
2861 | ||
2862 | spin_lock(&vmap_area_lock); | |
2863 | list_for_each_entry(va, &vmap_area_list, list) { | |
2864 | if (!count) | |
2865 | break; | |
2866 | ||
2867 | if (!va->vm) | |
2868 | continue; | |
2869 | ||
2870 | vm = va->vm; | |
2871 | vaddr = (char *) vm->addr; | |
2872 | if (addr >= vaddr + get_vm_area_size(vm)) | |
2873 | continue; | |
2874 | while (addr < vaddr) { | |
2875 | if (count == 0) | |
2876 | goto finished; | |
2877 | *buf = '\0'; | |
2878 | buf++; | |
2879 | addr++; | |
2880 | count--; | |
2881 | } | |
2882 | n = vaddr + get_vm_area_size(vm) - addr; | |
2883 | if (n > count) | |
2884 | n = count; | |
2885 | if (!(vm->flags & VM_IOREMAP)) | |
2886 | aligned_vread(buf, addr, n); | |
2887 | else /* IOREMAP area is treated as memory hole */ | |
2888 | memset(buf, 0, n); | |
2889 | buf += n; | |
2890 | addr += n; | |
2891 | count -= n; | |
2892 | } | |
2893 | finished: | |
2894 | spin_unlock(&vmap_area_lock); | |
2895 | ||
2896 | if (buf == buf_start) | |
2897 | return 0; | |
2898 | /* zero-fill memory holes */ | |
2899 | if (buf != buf_start + buflen) | |
2900 | memset(buf, 0, buflen - (buf - buf_start)); | |
2901 | ||
2902 | return buflen; | |
2903 | } | |
2904 | ||
2905 | /** | |
2906 | * vwrite() - write vmalloc area in a safe way. | |
2907 | * @buf: buffer for source data | |
2908 | * @addr: vm address. | |
2909 | * @count: number of bytes to be read. | |
2910 | * | |
2911 | * This function checks that addr is a valid vmalloc'ed area, and | |
2912 | * copy data from a buffer to the given addr. If specified range of | |
2913 | * [addr...addr+count) includes some valid address, data is copied from | |
2914 | * proper area of @buf. If there are memory holes, no copy to hole. | |
2915 | * IOREMAP area is treated as memory hole and no copy is done. | |
2916 | * | |
2917 | * If [addr...addr+count) doesn't includes any intersects with alive | |
2918 | * vm_struct area, returns 0. @buf should be kernel's buffer. | |
2919 | * | |
2920 | * Note: In usual ops, vwrite() is never necessary because the caller | |
2921 | * should know vmalloc() area is valid and can use memcpy(). | |
2922 | * This is for routines which have to access vmalloc area without | |
2923 | * any information, as /dev/kmem. | |
2924 | * | |
2925 | * Return: number of bytes for which addr and buf should be | |
2926 | * increased (same number as @count) or %0 if [addr...addr+count) | |
2927 | * doesn't include any intersection with valid vmalloc area | |
2928 | */ | |
2929 | long vwrite(char *buf, char *addr, unsigned long count) | |
2930 | { | |
2931 | struct vmap_area *va; | |
2932 | struct vm_struct *vm; | |
2933 | char *vaddr; | |
2934 | unsigned long n, buflen; | |
2935 | int copied = 0; | |
2936 | ||
2937 | /* Don't allow overflow */ | |
2938 | if ((unsigned long) addr + count < count) | |
2939 | count = -(unsigned long) addr; | |
2940 | buflen = count; | |
2941 | ||
2942 | spin_lock(&vmap_area_lock); | |
2943 | list_for_each_entry(va, &vmap_area_list, list) { | |
2944 | if (!count) | |
2945 | break; | |
2946 | ||
2947 | if (!va->vm) | |
2948 | continue; | |
2949 | ||
2950 | vm = va->vm; | |
2951 | vaddr = (char *) vm->addr; | |
2952 | if (addr >= vaddr + get_vm_area_size(vm)) | |
2953 | continue; | |
2954 | while (addr < vaddr) { | |
2955 | if (count == 0) | |
2956 | goto finished; | |
2957 | buf++; | |
2958 | addr++; | |
2959 | count--; | |
2960 | } | |
2961 | n = vaddr + get_vm_area_size(vm) - addr; | |
2962 | if (n > count) | |
2963 | n = count; | |
2964 | if (!(vm->flags & VM_IOREMAP)) { | |
2965 | aligned_vwrite(buf, addr, n); | |
2966 | copied++; | |
2967 | } | |
2968 | buf += n; | |
2969 | addr += n; | |
2970 | count -= n; | |
2971 | } | |
2972 | finished: | |
2973 | spin_unlock(&vmap_area_lock); | |
2974 | if (!copied) | |
2975 | return 0; | |
2976 | return buflen; | |
2977 | } | |
2978 | ||
2979 | /** | |
2980 | * remap_vmalloc_range_partial - map vmalloc pages to userspace | |
2981 | * @vma: vma to cover | |
2982 | * @uaddr: target user address to start at | |
2983 | * @kaddr: virtual address of vmalloc kernel memory | |
2984 | * @pgoff: offset from @kaddr to start at | |
2985 | * @size: size of map area | |
2986 | * | |
2987 | * Returns: 0 for success, -Exxx on failure | |
2988 | * | |
2989 | * This function checks that @kaddr is a valid vmalloc'ed area, | |
2990 | * and that it is big enough to cover the range starting at | |
2991 | * @uaddr in @vma. Will return failure if that criteria isn't | |
2992 | * met. | |
2993 | * | |
2994 | * Similar to remap_pfn_range() (see mm/memory.c) | |
2995 | */ | |
2996 | int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr, | |
2997 | void *kaddr, unsigned long pgoff, | |
2998 | unsigned long size) | |
2999 | { | |
3000 | struct vm_struct *area; | |
3001 | unsigned long off; | |
3002 | unsigned long end_index; | |
3003 | ||
3004 | if (check_shl_overflow(pgoff, PAGE_SHIFT, &off)) | |
3005 | return -EINVAL; | |
3006 | ||
3007 | size = PAGE_ALIGN(size); | |
3008 | ||
3009 | if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr)) | |
3010 | return -EINVAL; | |
3011 | ||
3012 | area = find_vm_area(kaddr); | |
3013 | if (!area) | |
3014 | return -EINVAL; | |
3015 | ||
3016 | if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT))) | |
3017 | return -EINVAL; | |
3018 | ||
3019 | if (check_add_overflow(size, off, &end_index) || | |
3020 | end_index > get_vm_area_size(area)) | |
3021 | return -EINVAL; | |
3022 | kaddr += off; | |
3023 | ||
3024 | do { | |
3025 | struct page *page = vmalloc_to_page(kaddr); | |
3026 | int ret; | |
3027 | ||
3028 | ret = vm_insert_page(vma, uaddr, page); | |
3029 | if (ret) | |
3030 | return ret; | |
3031 | ||
3032 | uaddr += PAGE_SIZE; | |
3033 | kaddr += PAGE_SIZE; | |
3034 | size -= PAGE_SIZE; | |
3035 | } while (size > 0); | |
3036 | ||
3037 | vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP; | |
3038 | ||
3039 | return 0; | |
3040 | } | |
3041 | EXPORT_SYMBOL(remap_vmalloc_range_partial); | |
3042 | ||
3043 | /** | |
3044 | * remap_vmalloc_range - map vmalloc pages to userspace | |
3045 | * @vma: vma to cover (map full range of vma) | |
3046 | * @addr: vmalloc memory | |
3047 | * @pgoff: number of pages into addr before first page to map | |
3048 | * | |
3049 | * Returns: 0 for success, -Exxx on failure | |
3050 | * | |
3051 | * This function checks that addr is a valid vmalloc'ed area, and | |
3052 | * that it is big enough to cover the vma. Will return failure if | |
3053 | * that criteria isn't met. | |
3054 | * | |
3055 | * Similar to remap_pfn_range() (see mm/memory.c) | |
3056 | */ | |
3057 | int remap_vmalloc_range(struct vm_area_struct *vma, void *addr, | |
3058 | unsigned long pgoff) | |
3059 | { | |
3060 | return remap_vmalloc_range_partial(vma, vma->vm_start, | |
3061 | addr, pgoff, | |
3062 | vma->vm_end - vma->vm_start); | |
3063 | } | |
3064 | EXPORT_SYMBOL(remap_vmalloc_range); | |
3065 | ||
3066 | static int f(pte_t *pte, unsigned long addr, void *data) | |
3067 | { | |
3068 | pte_t ***p = data; | |
3069 | ||
3070 | if (p) { | |
3071 | *(*p) = pte; | |
3072 | (*p)++; | |
3073 | } | |
3074 | return 0; | |
3075 | } | |
3076 | ||
3077 | /** | |
3078 | * alloc_vm_area - allocate a range of kernel address space | |
3079 | * @size: size of the area | |
3080 | * @ptes: returns the PTEs for the address space | |
3081 | * | |
3082 | * Returns: NULL on failure, vm_struct on success | |
3083 | * | |
3084 | * This function reserves a range of kernel address space, and | |
3085 | * allocates pagetables to map that range. No actual mappings | |
3086 | * are created. | |
3087 | * | |
3088 | * If @ptes is non-NULL, pointers to the PTEs (in init_mm) | |
3089 | * allocated for the VM area are returned. | |
3090 | */ | |
3091 | struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes) | |
3092 | { | |
3093 | struct vm_struct *area; | |
3094 | ||
3095 | area = get_vm_area_caller(size, VM_IOREMAP, | |
3096 | __builtin_return_address(0)); | |
3097 | if (area == NULL) | |
3098 | return NULL; | |
3099 | ||
3100 | /* | |
3101 | * This ensures that page tables are constructed for this region | |
3102 | * of kernel virtual address space and mapped into init_mm. | |
3103 | */ | |
3104 | if (apply_to_page_range(&init_mm, (unsigned long)area->addr, | |
3105 | size, f, ptes ? &ptes : NULL)) { | |
3106 | free_vm_area(area); | |
3107 | return NULL; | |
3108 | } | |
3109 | ||
3110 | return area; | |
3111 | } | |
3112 | EXPORT_SYMBOL_GPL(alloc_vm_area); | |
3113 | ||
3114 | void free_vm_area(struct vm_struct *area) | |
3115 | { | |
3116 | struct vm_struct *ret; | |
3117 | ret = remove_vm_area(area->addr); | |
3118 | BUG_ON(ret != area); | |
3119 | kfree(area); | |
3120 | } | |
3121 | EXPORT_SYMBOL_GPL(free_vm_area); | |
3122 | ||
3123 | #ifdef CONFIG_SMP | |
3124 | static struct vmap_area *node_to_va(struct rb_node *n) | |
3125 | { | |
3126 | return rb_entry_safe(n, struct vmap_area, rb_node); | |
3127 | } | |
3128 | ||
3129 | /** | |
3130 | * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to | |
3131 | * @addr: target address | |
3132 | * | |
3133 | * Returns: vmap_area if it is found. If there is no such area | |
3134 | * the first highest(reverse order) vmap_area is returned | |
3135 | * i.e. va->va_start < addr && va->va_end < addr or NULL | |
3136 | * if there are no any areas before @addr. | |
3137 | */ | |
3138 | static struct vmap_area * | |
3139 | pvm_find_va_enclose_addr(unsigned long addr) | |
3140 | { | |
3141 | struct vmap_area *va, *tmp; | |
3142 | struct rb_node *n; | |
3143 | ||
3144 | n = free_vmap_area_root.rb_node; | |
3145 | va = NULL; | |
3146 | ||
3147 | while (n) { | |
3148 | tmp = rb_entry(n, struct vmap_area, rb_node); | |
3149 | if (tmp->va_start <= addr) { | |
3150 | va = tmp; | |
3151 | if (tmp->va_end >= addr) | |
3152 | break; | |
3153 | ||
3154 | n = n->rb_right; | |
3155 | } else { | |
3156 | n = n->rb_left; | |
3157 | } | |
3158 | } | |
3159 | ||
3160 | return va; | |
3161 | } | |
3162 | ||
3163 | /** | |
3164 | * pvm_determine_end_from_reverse - find the highest aligned address | |
3165 | * of free block below VMALLOC_END | |
3166 | * @va: | |
3167 | * in - the VA we start the search(reverse order); | |
3168 | * out - the VA with the highest aligned end address. | |
3169 | * | |
3170 | * Returns: determined end address within vmap_area | |
3171 | */ | |
3172 | static unsigned long | |
3173 | pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align) | |
3174 | { | |
3175 | unsigned long vmalloc_end = VMALLOC_END & ~(align - 1); | |
3176 | unsigned long addr; | |
3177 | ||
3178 | if (likely(*va)) { | |
3179 | list_for_each_entry_from_reverse((*va), | |
3180 | &free_vmap_area_list, list) { | |
3181 | addr = min((*va)->va_end & ~(align - 1), vmalloc_end); | |
3182 | if ((*va)->va_start < addr) | |
3183 | return addr; | |
3184 | } | |
3185 | } | |
3186 | ||
3187 | return 0; | |
3188 | } | |
3189 | ||
3190 | /** | |
3191 | * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator | |
3192 | * @offsets: array containing offset of each area | |
3193 | * @sizes: array containing size of each area | |
3194 | * @nr_vms: the number of areas to allocate | |
3195 | * @align: alignment, all entries in @offsets and @sizes must be aligned to this | |
3196 | * | |
3197 | * Returns: kmalloc'd vm_struct pointer array pointing to allocated | |
3198 | * vm_structs on success, %NULL on failure | |
3199 | * | |
3200 | * Percpu allocator wants to use congruent vm areas so that it can | |
3201 | * maintain the offsets among percpu areas. This function allocates | |
3202 | * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to | |
3203 | * be scattered pretty far, distance between two areas easily going up | |
3204 | * to gigabytes. To avoid interacting with regular vmallocs, these | |
3205 | * areas are allocated from top. | |
3206 | * | |
3207 | * Despite its complicated look, this allocator is rather simple. It | |
3208 | * does everything top-down and scans free blocks from the end looking | |
3209 | * for matching base. While scanning, if any of the areas do not fit the | |
3210 | * base address is pulled down to fit the area. Scanning is repeated till | |
3211 | * all the areas fit and then all necessary data structures are inserted | |
3212 | * and the result is returned. | |
3213 | */ | |
3214 | struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets, | |
3215 | const size_t *sizes, int nr_vms, | |
3216 | size_t align) | |
3217 | { | |
3218 | const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align); | |
3219 | const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1); | |
3220 | struct vmap_area **vas, *va; | |
3221 | struct vm_struct **vms; | |
3222 | int area, area2, last_area, term_area; | |
3223 | unsigned long base, start, size, end, last_end, orig_start, orig_end; | |
3224 | bool purged = false; | |
3225 | enum fit_type type; | |
3226 | ||
3227 | /* verify parameters and allocate data structures */ | |
3228 | BUG_ON(offset_in_page(align) || !is_power_of_2(align)); | |
3229 | for (last_area = 0, area = 0; area < nr_vms; area++) { | |
3230 | start = offsets[area]; | |
3231 | end = start + sizes[area]; | |
3232 | ||
3233 | /* is everything aligned properly? */ | |
3234 | BUG_ON(!IS_ALIGNED(offsets[area], align)); | |
3235 | BUG_ON(!IS_ALIGNED(sizes[area], align)); | |
3236 | ||
3237 | /* detect the area with the highest address */ | |
3238 | if (start > offsets[last_area]) | |
3239 | last_area = area; | |
3240 | ||
3241 | for (area2 = area + 1; area2 < nr_vms; area2++) { | |
3242 | unsigned long start2 = offsets[area2]; | |
3243 | unsigned long end2 = start2 + sizes[area2]; | |
3244 | ||
3245 | BUG_ON(start2 < end && start < end2); | |
3246 | } | |
3247 | } | |
3248 | last_end = offsets[last_area] + sizes[last_area]; | |
3249 | ||
3250 | if (vmalloc_end - vmalloc_start < last_end) { | |
3251 | WARN_ON(true); | |
3252 | return NULL; | |
3253 | } | |
3254 | ||
3255 | vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL); | |
3256 | vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL); | |
3257 | if (!vas || !vms) | |
3258 | goto err_free2; | |
3259 | ||
3260 | for (area = 0; area < nr_vms; area++) { | |
3261 | vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL); | |
3262 | vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL); | |
3263 | if (!vas[area] || !vms[area]) | |
3264 | goto err_free; | |
3265 | } | |
3266 | retry: | |
3267 | spin_lock(&free_vmap_area_lock); | |
3268 | ||
3269 | /* start scanning - we scan from the top, begin with the last area */ | |
3270 | area = term_area = last_area; | |
3271 | start = offsets[area]; | |
3272 | end = start + sizes[area]; | |
3273 | ||
3274 | va = pvm_find_va_enclose_addr(vmalloc_end); | |
3275 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
3276 | ||
3277 | while (true) { | |
3278 | /* | |
3279 | * base might have underflowed, add last_end before | |
3280 | * comparing. | |
3281 | */ | |
3282 | if (base + last_end < vmalloc_start + last_end) | |
3283 | goto overflow; | |
3284 | ||
3285 | /* | |
3286 | * Fitting base has not been found. | |
3287 | */ | |
3288 | if (va == NULL) | |
3289 | goto overflow; | |
3290 | ||
3291 | /* | |
3292 | * If required width exceeds current VA block, move | |
3293 | * base downwards and then recheck. | |
3294 | */ | |
3295 | if (base + end > va->va_end) { | |
3296 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
3297 | term_area = area; | |
3298 | continue; | |
3299 | } | |
3300 | ||
3301 | /* | |
3302 | * If this VA does not fit, move base downwards and recheck. | |
3303 | */ | |
3304 | if (base + start < va->va_start) { | |
3305 | va = node_to_va(rb_prev(&va->rb_node)); | |
3306 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
3307 | term_area = area; | |
3308 | continue; | |
3309 | } | |
3310 | ||
3311 | /* | |
3312 | * This area fits, move on to the previous one. If | |
3313 | * the previous one is the terminal one, we're done. | |
3314 | */ | |
3315 | area = (area + nr_vms - 1) % nr_vms; | |
3316 | if (area == term_area) | |
3317 | break; | |
3318 | ||
3319 | start = offsets[area]; | |
3320 | end = start + sizes[area]; | |
3321 | va = pvm_find_va_enclose_addr(base + end); | |
3322 | } | |
3323 | ||
3324 | /* we've found a fitting base, insert all va's */ | |
3325 | for (area = 0; area < nr_vms; area++) { | |
3326 | int ret; | |
3327 | ||
3328 | start = base + offsets[area]; | |
3329 | size = sizes[area]; | |
3330 | ||
3331 | va = pvm_find_va_enclose_addr(start); | |
3332 | if (WARN_ON_ONCE(va == NULL)) | |
3333 | /* It is a BUG(), but trigger recovery instead. */ | |
3334 | goto recovery; | |
3335 | ||
3336 | type = classify_va_fit_type(va, start, size); | |
3337 | if (WARN_ON_ONCE(type == NOTHING_FIT)) | |
3338 | /* It is a BUG(), but trigger recovery instead. */ | |
3339 | goto recovery; | |
3340 | ||
3341 | ret = adjust_va_to_fit_type(va, start, size, type); | |
3342 | if (unlikely(ret)) | |
3343 | goto recovery; | |
3344 | ||
3345 | /* Allocated area. */ | |
3346 | va = vas[area]; | |
3347 | va->va_start = start; | |
3348 | va->va_end = start + size; | |
3349 | } | |
3350 | ||
3351 | spin_unlock(&free_vmap_area_lock); | |
3352 | ||
3353 | /* populate the kasan shadow space */ | |
3354 | for (area = 0; area < nr_vms; area++) { | |
3355 | if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area])) | |
3356 | goto err_free_shadow; | |
3357 | ||
3358 | kasan_unpoison_vmalloc((void *)vas[area]->va_start, | |
3359 | sizes[area]); | |
3360 | } | |
3361 | ||
3362 | /* insert all vm's */ | |
3363 | spin_lock(&vmap_area_lock); | |
3364 | for (area = 0; area < nr_vms; area++) { | |
3365 | insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list); | |
3366 | ||
3367 | setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC, | |
3368 | pcpu_get_vm_areas); | |
3369 | } | |
3370 | spin_unlock(&vmap_area_lock); | |
3371 | ||
3372 | kfree(vas); | |
3373 | return vms; | |
3374 | ||
3375 | recovery: | |
3376 | /* | |
3377 | * Remove previously allocated areas. There is no | |
3378 | * need in removing these areas from the busy tree, | |
3379 | * because they are inserted only on the final step | |
3380 | * and when pcpu_get_vm_areas() is success. | |
3381 | */ | |
3382 | while (area--) { | |
3383 | orig_start = vas[area]->va_start; | |
3384 | orig_end = vas[area]->va_end; | |
3385 | va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root, | |
3386 | &free_vmap_area_list); | |
3387 | kasan_release_vmalloc(orig_start, orig_end, | |
3388 | va->va_start, va->va_end); | |
3389 | vas[area] = NULL; | |
3390 | } | |
3391 | ||
3392 | overflow: | |
3393 | spin_unlock(&free_vmap_area_lock); | |
3394 | if (!purged) { | |
3395 | purge_vmap_area_lazy(); | |
3396 | purged = true; | |
3397 | ||
3398 | /* Before "retry", check if we recover. */ | |
3399 | for (area = 0; area < nr_vms; area++) { | |
3400 | if (vas[area]) | |
3401 | continue; | |
3402 | ||
3403 | vas[area] = kmem_cache_zalloc( | |
3404 | vmap_area_cachep, GFP_KERNEL); | |
3405 | if (!vas[area]) | |
3406 | goto err_free; | |
3407 | } | |
3408 | ||
3409 | goto retry; | |
3410 | } | |
3411 | ||
3412 | err_free: | |
3413 | for (area = 0; area < nr_vms; area++) { | |
3414 | if (vas[area]) | |
3415 | kmem_cache_free(vmap_area_cachep, vas[area]); | |
3416 | ||
3417 | kfree(vms[area]); | |
3418 | } | |
3419 | err_free2: | |
3420 | kfree(vas); | |
3421 | kfree(vms); | |
3422 | return NULL; | |
3423 | ||
3424 | err_free_shadow: | |
3425 | spin_lock(&free_vmap_area_lock); | |
3426 | /* | |
3427 | * We release all the vmalloc shadows, even the ones for regions that | |
3428 | * hadn't been successfully added. This relies on kasan_release_vmalloc | |
3429 | * being able to tolerate this case. | |
3430 | */ | |
3431 | for (area = 0; area < nr_vms; area++) { | |
3432 | orig_start = vas[area]->va_start; | |
3433 | orig_end = vas[area]->va_end; | |
3434 | va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root, | |
3435 | &free_vmap_area_list); | |
3436 | kasan_release_vmalloc(orig_start, orig_end, | |
3437 | va->va_start, va->va_end); | |
3438 | vas[area] = NULL; | |
3439 | kfree(vms[area]); | |
3440 | } | |
3441 | spin_unlock(&free_vmap_area_lock); | |
3442 | kfree(vas); | |
3443 | kfree(vms); | |
3444 | return NULL; | |
3445 | } | |
3446 | ||
3447 | /** | |
3448 | * pcpu_free_vm_areas - free vmalloc areas for percpu allocator | |
3449 | * @vms: vm_struct pointer array returned by pcpu_get_vm_areas() | |
3450 | * @nr_vms: the number of allocated areas | |
3451 | * | |
3452 | * Free vm_structs and the array allocated by pcpu_get_vm_areas(). | |
3453 | */ | |
3454 | void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms) | |
3455 | { | |
3456 | int i; | |
3457 | ||
3458 | for (i = 0; i < nr_vms; i++) | |
3459 | free_vm_area(vms[i]); | |
3460 | kfree(vms); | |
3461 | } | |
3462 | #endif /* CONFIG_SMP */ | |
3463 | ||
3464 | #ifdef CONFIG_PROC_FS | |
3465 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3466 | __acquires(&vmap_purge_lock) | |
3467 | __acquires(&vmap_area_lock) | |
3468 | { | |
3469 | mutex_lock(&vmap_purge_lock); | |
3470 | spin_lock(&vmap_area_lock); | |
3471 | ||
3472 | return seq_list_start(&vmap_area_list, *pos); | |
3473 | } | |
3474 | ||
3475 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3476 | { | |
3477 | return seq_list_next(p, &vmap_area_list, pos); | |
3478 | } | |
3479 | ||
3480 | static void s_stop(struct seq_file *m, void *p) | |
3481 | __releases(&vmap_purge_lock) | |
3482 | __releases(&vmap_area_lock) | |
3483 | { | |
3484 | mutex_unlock(&vmap_purge_lock); | |
3485 | spin_unlock(&vmap_area_lock); | |
3486 | } | |
3487 | ||
3488 | static void show_numa_info(struct seq_file *m, struct vm_struct *v) | |
3489 | { | |
3490 | if (IS_ENABLED(CONFIG_NUMA)) { | |
3491 | unsigned int nr, *counters = m->private; | |
3492 | ||
3493 | if (!counters) | |
3494 | return; | |
3495 | ||
3496 | if (v->flags & VM_UNINITIALIZED) | |
3497 | return; | |
3498 | /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */ | |
3499 | smp_rmb(); | |
3500 | ||
3501 | memset(counters, 0, nr_node_ids * sizeof(unsigned int)); | |
3502 | ||
3503 | for (nr = 0; nr < v->nr_pages; nr++) | |
3504 | counters[page_to_nid(v->pages[nr])]++; | |
3505 | ||
3506 | for_each_node_state(nr, N_HIGH_MEMORY) | |
3507 | if (counters[nr]) | |
3508 | seq_printf(m, " N%u=%u", nr, counters[nr]); | |
3509 | } | |
3510 | } | |
3511 | ||
3512 | static void show_purge_info(struct seq_file *m) | |
3513 | { | |
3514 | struct llist_node *head; | |
3515 | struct vmap_area *va; | |
3516 | ||
3517 | head = READ_ONCE(vmap_purge_list.first); | |
3518 | if (head == NULL) | |
3519 | return; | |
3520 | ||
3521 | llist_for_each_entry(va, head, purge_list) { | |
3522 | seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n", | |
3523 | (void *)va->va_start, (void *)va->va_end, | |
3524 | va->va_end - va->va_start); | |
3525 | } | |
3526 | } | |
3527 | ||
3528 | static int s_show(struct seq_file *m, void *p) | |
3529 | { | |
3530 | struct vmap_area *va; | |
3531 | struct vm_struct *v; | |
3532 | ||
3533 | va = list_entry(p, struct vmap_area, list); | |
3534 | ||
3535 | /* | |
3536 | * s_show can encounter race with remove_vm_area, !vm on behalf | |
3537 | * of vmap area is being tear down or vm_map_ram allocation. | |
3538 | */ | |
3539 | if (!va->vm) { | |
3540 | seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n", | |
3541 | (void *)va->va_start, (void *)va->va_end, | |
3542 | va->va_end - va->va_start); | |
3543 | ||
3544 | return 0; | |
3545 | } | |
3546 | ||
3547 | v = va->vm; | |
3548 | ||
3549 | seq_printf(m, "0x%pK-0x%pK %7ld", | |
3550 | v->addr, v->addr + v->size, v->size); | |
3551 | ||
3552 | if (v->caller) | |
3553 | seq_printf(m, " %pS", v->caller); | |
3554 | ||
3555 | if (v->nr_pages) | |
3556 | seq_printf(m, " pages=%d", v->nr_pages); | |
3557 | ||
3558 | if (v->phys_addr) | |
3559 | seq_printf(m, " phys=%pa", &v->phys_addr); | |
3560 | ||
3561 | if (v->flags & VM_IOREMAP) | |
3562 | seq_puts(m, " ioremap"); | |
3563 | ||
3564 | if (v->flags & VM_ALLOC) | |
3565 | seq_puts(m, " vmalloc"); | |
3566 | ||
3567 | if (v->flags & VM_MAP) | |
3568 | seq_puts(m, " vmap"); | |
3569 | ||
3570 | if (v->flags & VM_USERMAP) | |
3571 | seq_puts(m, " user"); | |
3572 | ||
3573 | if (v->flags & VM_DMA_COHERENT) | |
3574 | seq_puts(m, " dma-coherent"); | |
3575 | ||
3576 | if (is_vmalloc_addr(v->pages)) | |
3577 | seq_puts(m, " vpages"); | |
3578 | ||
3579 | show_numa_info(m, v); | |
3580 | seq_putc(m, '\n'); | |
3581 | ||
3582 | /* | |
3583 | * As a final step, dump "unpurged" areas. Note, | |
3584 | * that entire "/proc/vmallocinfo" output will not | |
3585 | * be address sorted, because the purge list is not | |
3586 | * sorted. | |
3587 | */ | |
3588 | if (list_is_last(&va->list, &vmap_area_list)) | |
3589 | show_purge_info(m); | |
3590 | ||
3591 | return 0; | |
3592 | } | |
3593 | ||
3594 | static const struct seq_operations vmalloc_op = { | |
3595 | .start = s_start, | |
3596 | .next = s_next, | |
3597 | .stop = s_stop, | |
3598 | .show = s_show, | |
3599 | }; | |
3600 | ||
3601 | static int __init proc_vmalloc_init(void) | |
3602 | { | |
3603 | if (IS_ENABLED(CONFIG_NUMA)) | |
3604 | proc_create_seq_private("vmallocinfo", 0400, NULL, | |
3605 | &vmalloc_op, | |
3606 | nr_node_ids * sizeof(unsigned int), NULL); | |
3607 | else | |
3608 | proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op); | |
3609 | return 0; | |
3610 | } | |
3611 | module_init(proc_vmalloc_init); | |
3612 | ||
3613 | #endif |