Loading...
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
38#include <linux/uaccess.h>
39#include <asm/tlbflush.h>
40#include <asm/shmparam.h>
41
42#include "internal.h"
43
44struct vfree_deferred {
45 struct llist_head list;
46 struct work_struct wq;
47};
48static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
49
50static void __vunmap(const void *, int);
51
52static void free_work(struct work_struct *w)
53{
54 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
55 struct llist_node *t, *llnode;
56
57 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
58 __vunmap((void *)llnode, 1);
59}
60
61/*** Page table manipulation functions ***/
62
63static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
64{
65 pte_t *pte;
66
67 pte = pte_offset_kernel(pmd, addr);
68 do {
69 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
70 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
71 } while (pte++, addr += PAGE_SIZE, addr != end);
72}
73
74static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
75{
76 pmd_t *pmd;
77 unsigned long next;
78
79 pmd = pmd_offset(pud, addr);
80 do {
81 next = pmd_addr_end(addr, end);
82 if (pmd_clear_huge(pmd))
83 continue;
84 if (pmd_none_or_clear_bad(pmd))
85 continue;
86 vunmap_pte_range(pmd, addr, next);
87 } while (pmd++, addr = next, addr != end);
88}
89
90static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
91{
92 pud_t *pud;
93 unsigned long next;
94
95 pud = pud_offset(p4d, addr);
96 do {
97 next = pud_addr_end(addr, end);
98 if (pud_clear_huge(pud))
99 continue;
100 if (pud_none_or_clear_bad(pud))
101 continue;
102 vunmap_pmd_range(pud, addr, next);
103 } while (pud++, addr = next, addr != end);
104}
105
106static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
107{
108 p4d_t *p4d;
109 unsigned long next;
110
111 p4d = p4d_offset(pgd, addr);
112 do {
113 next = p4d_addr_end(addr, end);
114 if (p4d_clear_huge(p4d))
115 continue;
116 if (p4d_none_or_clear_bad(p4d))
117 continue;
118 vunmap_pud_range(p4d, addr, next);
119 } while (p4d++, addr = next, addr != end);
120}
121
122static void vunmap_page_range(unsigned long addr, unsigned long end)
123{
124 pgd_t *pgd;
125 unsigned long next;
126
127 BUG_ON(addr >= end);
128 pgd = pgd_offset_k(addr);
129 do {
130 next = pgd_addr_end(addr, end);
131 if (pgd_none_or_clear_bad(pgd))
132 continue;
133 vunmap_p4d_range(pgd, addr, next);
134 } while (pgd++, addr = next, addr != end);
135}
136
137static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
138 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
139{
140 pte_t *pte;
141
142 /*
143 * nr is a running index into the array which helps higher level
144 * callers keep track of where we're up to.
145 */
146
147 pte = pte_alloc_kernel(pmd, addr);
148 if (!pte)
149 return -ENOMEM;
150 do {
151 struct page *page = pages[*nr];
152
153 if (WARN_ON(!pte_none(*pte)))
154 return -EBUSY;
155 if (WARN_ON(!page))
156 return -ENOMEM;
157 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
158 (*nr)++;
159 } while (pte++, addr += PAGE_SIZE, addr != end);
160 return 0;
161}
162
163static int vmap_pmd_range(pud_t *pud, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
165{
166 pmd_t *pmd;
167 unsigned long next;
168
169 pmd = pmd_alloc(&init_mm, pud, addr);
170 if (!pmd)
171 return -ENOMEM;
172 do {
173 next = pmd_addr_end(addr, end);
174 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
175 return -ENOMEM;
176 } while (pmd++, addr = next, addr != end);
177 return 0;
178}
179
180static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
181 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
182{
183 pud_t *pud;
184 unsigned long next;
185
186 pud = pud_alloc(&init_mm, p4d, addr);
187 if (!pud)
188 return -ENOMEM;
189 do {
190 next = pud_addr_end(addr, end);
191 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
192 return -ENOMEM;
193 } while (pud++, addr = next, addr != end);
194 return 0;
195}
196
197static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
198 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
199{
200 p4d_t *p4d;
201 unsigned long next;
202
203 p4d = p4d_alloc(&init_mm, pgd, addr);
204 if (!p4d)
205 return -ENOMEM;
206 do {
207 next = p4d_addr_end(addr, end);
208 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
209 return -ENOMEM;
210 } while (p4d++, addr = next, addr != end);
211 return 0;
212}
213
214/*
215 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
216 * will have pfns corresponding to the "pages" array.
217 *
218 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
219 */
220static int vmap_page_range_noflush(unsigned long start, unsigned long end,
221 pgprot_t prot, struct page **pages)
222{
223 pgd_t *pgd;
224 unsigned long next;
225 unsigned long addr = start;
226 int err = 0;
227 int nr = 0;
228
229 BUG_ON(addr >= end);
230 pgd = pgd_offset_k(addr);
231 do {
232 next = pgd_addr_end(addr, end);
233 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
234 if (err)
235 return err;
236 } while (pgd++, addr = next, addr != end);
237
238 return nr;
239}
240
241static int vmap_page_range(unsigned long start, unsigned long end,
242 pgprot_t prot, struct page **pages)
243{
244 int ret;
245
246 ret = vmap_page_range_noflush(start, end, prot, pages);
247 flush_cache_vmap(start, end);
248 return ret;
249}
250
251int is_vmalloc_or_module_addr(const void *x)
252{
253 /*
254 * ARM, x86-64 and sparc64 put modules in a special place,
255 * and fall back on vmalloc() if that fails. Others
256 * just put it in the vmalloc space.
257 */
258#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
259 unsigned long addr = (unsigned long)x;
260 if (addr >= MODULES_VADDR && addr < MODULES_END)
261 return 1;
262#endif
263 return is_vmalloc_addr(x);
264}
265
266/*
267 * Walk a vmap address to the struct page it maps.
268 */
269struct page *vmalloc_to_page(const void *vmalloc_addr)
270{
271 unsigned long addr = (unsigned long) vmalloc_addr;
272 struct page *page = NULL;
273 pgd_t *pgd = pgd_offset_k(addr);
274 p4d_t *p4d;
275 pud_t *pud;
276 pmd_t *pmd;
277 pte_t *ptep, pte;
278
279 /*
280 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
281 * architectures that do not vmalloc module space
282 */
283 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
284
285 if (pgd_none(*pgd))
286 return NULL;
287 p4d = p4d_offset(pgd, addr);
288 if (p4d_none(*p4d))
289 return NULL;
290 pud = pud_offset(p4d, addr);
291
292 /*
293 * Don't dereference bad PUD or PMD (below) entries. This will also
294 * identify huge mappings, which we may encounter on architectures
295 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
296 * identified as vmalloc addresses by is_vmalloc_addr(), but are
297 * not [unambiguously] associated with a struct page, so there is
298 * no correct value to return for them.
299 */
300 WARN_ON_ONCE(pud_bad(*pud));
301 if (pud_none(*pud) || pud_bad(*pud))
302 return NULL;
303 pmd = pmd_offset(pud, addr);
304 WARN_ON_ONCE(pmd_bad(*pmd));
305 if (pmd_none(*pmd) || pmd_bad(*pmd))
306 return NULL;
307
308 ptep = pte_offset_map(pmd, addr);
309 pte = *ptep;
310 if (pte_present(pte))
311 page = pte_page(pte);
312 pte_unmap(ptep);
313 return page;
314}
315EXPORT_SYMBOL(vmalloc_to_page);
316
317/*
318 * Map a vmalloc()-space virtual address to the physical page frame number.
319 */
320unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
321{
322 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
323}
324EXPORT_SYMBOL(vmalloc_to_pfn);
325
326
327/*** Global kva allocator ***/
328
329#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
330#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
331
332
333static DEFINE_SPINLOCK(vmap_area_lock);
334/* Export for kexec only */
335LIST_HEAD(vmap_area_list);
336static LLIST_HEAD(vmap_purge_list);
337static struct rb_root vmap_area_root = RB_ROOT;
338static bool vmap_initialized __read_mostly;
339
340/*
341 * This kmem_cache is used for vmap_area objects. Instead of
342 * allocating from slab we reuse an object from this cache to
343 * make things faster. Especially in "no edge" splitting of
344 * free block.
345 */
346static struct kmem_cache *vmap_area_cachep;
347
348/*
349 * This linked list is used in pair with free_vmap_area_root.
350 * It gives O(1) access to prev/next to perform fast coalescing.
351 */
352static LIST_HEAD(free_vmap_area_list);
353
354/*
355 * This augment red-black tree represents the free vmap space.
356 * All vmap_area objects in this tree are sorted by va->va_start
357 * address. It is used for allocation and merging when a vmap
358 * object is released.
359 *
360 * Each vmap_area node contains a maximum available free block
361 * of its sub-tree, right or left. Therefore it is possible to
362 * find a lowest match of free area.
363 */
364static struct rb_root free_vmap_area_root = RB_ROOT;
365
366/*
367 * Preload a CPU with one object for "no edge" split case. The
368 * aim is to get rid of allocations from the atomic context, thus
369 * to use more permissive allocation masks.
370 */
371static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
372
373static __always_inline unsigned long
374va_size(struct vmap_area *va)
375{
376 return (va->va_end - va->va_start);
377}
378
379static __always_inline unsigned long
380get_subtree_max_size(struct rb_node *node)
381{
382 struct vmap_area *va;
383
384 va = rb_entry_safe(node, struct vmap_area, rb_node);
385 return va ? va->subtree_max_size : 0;
386}
387
388/*
389 * Gets called when remove the node and rotate.
390 */
391static __always_inline unsigned long
392compute_subtree_max_size(struct vmap_area *va)
393{
394 return max3(va_size(va),
395 get_subtree_max_size(va->rb_node.rb_left),
396 get_subtree_max_size(va->rb_node.rb_right));
397}
398
399RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
400 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
401
402static void purge_vmap_area_lazy(void);
403static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
404static unsigned long lazy_max_pages(void);
405
406static atomic_long_t nr_vmalloc_pages;
407
408unsigned long vmalloc_nr_pages(void)
409{
410 return atomic_long_read(&nr_vmalloc_pages);
411}
412
413static struct vmap_area *__find_vmap_area(unsigned long addr)
414{
415 struct rb_node *n = vmap_area_root.rb_node;
416
417 while (n) {
418 struct vmap_area *va;
419
420 va = rb_entry(n, struct vmap_area, rb_node);
421 if (addr < va->va_start)
422 n = n->rb_left;
423 else if (addr >= va->va_end)
424 n = n->rb_right;
425 else
426 return va;
427 }
428
429 return NULL;
430}
431
432/*
433 * This function returns back addresses of parent node
434 * and its left or right link for further processing.
435 */
436static __always_inline struct rb_node **
437find_va_links(struct vmap_area *va,
438 struct rb_root *root, struct rb_node *from,
439 struct rb_node **parent)
440{
441 struct vmap_area *tmp_va;
442 struct rb_node **link;
443
444 if (root) {
445 link = &root->rb_node;
446 if (unlikely(!*link)) {
447 *parent = NULL;
448 return link;
449 }
450 } else {
451 link = &from;
452 }
453
454 /*
455 * Go to the bottom of the tree. When we hit the last point
456 * we end up with parent rb_node and correct direction, i name
457 * it link, where the new va->rb_node will be attached to.
458 */
459 do {
460 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
461
462 /*
463 * During the traversal we also do some sanity check.
464 * Trigger the BUG() if there are sides(left/right)
465 * or full overlaps.
466 */
467 if (va->va_start < tmp_va->va_end &&
468 va->va_end <= tmp_va->va_start)
469 link = &(*link)->rb_left;
470 else if (va->va_end > tmp_va->va_start &&
471 va->va_start >= tmp_va->va_end)
472 link = &(*link)->rb_right;
473 else
474 BUG();
475 } while (*link);
476
477 *parent = &tmp_va->rb_node;
478 return link;
479}
480
481static __always_inline struct list_head *
482get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
483{
484 struct list_head *list;
485
486 if (unlikely(!parent))
487 /*
488 * The red-black tree where we try to find VA neighbors
489 * before merging or inserting is empty, i.e. it means
490 * there is no free vmap space. Normally it does not
491 * happen but we handle this case anyway.
492 */
493 return NULL;
494
495 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
496 return (&parent->rb_right == link ? list->next : list);
497}
498
499static __always_inline void
500link_va(struct vmap_area *va, struct rb_root *root,
501 struct rb_node *parent, struct rb_node **link, struct list_head *head)
502{
503 /*
504 * VA is still not in the list, but we can
505 * identify its future previous list_head node.
506 */
507 if (likely(parent)) {
508 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
509 if (&parent->rb_right != link)
510 head = head->prev;
511 }
512
513 /* Insert to the rb-tree */
514 rb_link_node(&va->rb_node, parent, link);
515 if (root == &free_vmap_area_root) {
516 /*
517 * Some explanation here. Just perform simple insertion
518 * to the tree. We do not set va->subtree_max_size to
519 * its current size before calling rb_insert_augmented().
520 * It is because of we populate the tree from the bottom
521 * to parent levels when the node _is_ in the tree.
522 *
523 * Therefore we set subtree_max_size to zero after insertion,
524 * to let __augment_tree_propagate_from() puts everything to
525 * the correct order later on.
526 */
527 rb_insert_augmented(&va->rb_node,
528 root, &free_vmap_area_rb_augment_cb);
529 va->subtree_max_size = 0;
530 } else {
531 rb_insert_color(&va->rb_node, root);
532 }
533
534 /* Address-sort this list */
535 list_add(&va->list, head);
536}
537
538static __always_inline void
539unlink_va(struct vmap_area *va, struct rb_root *root)
540{
541 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
542 return;
543
544 if (root == &free_vmap_area_root)
545 rb_erase_augmented(&va->rb_node,
546 root, &free_vmap_area_rb_augment_cb);
547 else
548 rb_erase(&va->rb_node, root);
549
550 list_del(&va->list);
551 RB_CLEAR_NODE(&va->rb_node);
552}
553
554#if DEBUG_AUGMENT_PROPAGATE_CHECK
555static void
556augment_tree_propagate_check(struct rb_node *n)
557{
558 struct vmap_area *va;
559 struct rb_node *node;
560 unsigned long size;
561 bool found = false;
562
563 if (n == NULL)
564 return;
565
566 va = rb_entry(n, struct vmap_area, rb_node);
567 size = va->subtree_max_size;
568 node = n;
569
570 while (node) {
571 va = rb_entry(node, struct vmap_area, rb_node);
572
573 if (get_subtree_max_size(node->rb_left) == size) {
574 node = node->rb_left;
575 } else {
576 if (va_size(va) == size) {
577 found = true;
578 break;
579 }
580
581 node = node->rb_right;
582 }
583 }
584
585 if (!found) {
586 va = rb_entry(n, struct vmap_area, rb_node);
587 pr_emerg("tree is corrupted: %lu, %lu\n",
588 va_size(va), va->subtree_max_size);
589 }
590
591 augment_tree_propagate_check(n->rb_left);
592 augment_tree_propagate_check(n->rb_right);
593}
594#endif
595
596/*
597 * This function populates subtree_max_size from bottom to upper
598 * levels starting from VA point. The propagation must be done
599 * when VA size is modified by changing its va_start/va_end. Or
600 * in case of newly inserting of VA to the tree.
601 *
602 * It means that __augment_tree_propagate_from() must be called:
603 * - After VA has been inserted to the tree(free path);
604 * - After VA has been shrunk(allocation path);
605 * - After VA has been increased(merging path).
606 *
607 * Please note that, it does not mean that upper parent nodes
608 * and their subtree_max_size are recalculated all the time up
609 * to the root node.
610 *
611 * 4--8
612 * /\
613 * / \
614 * / \
615 * 2--2 8--8
616 *
617 * For example if we modify the node 4, shrinking it to 2, then
618 * no any modification is required. If we shrink the node 2 to 1
619 * its subtree_max_size is updated only, and set to 1. If we shrink
620 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
621 * node becomes 4--6.
622 */
623static __always_inline void
624augment_tree_propagate_from(struct vmap_area *va)
625{
626 struct rb_node *node = &va->rb_node;
627 unsigned long new_va_sub_max_size;
628
629 while (node) {
630 va = rb_entry(node, struct vmap_area, rb_node);
631 new_va_sub_max_size = compute_subtree_max_size(va);
632
633 /*
634 * If the newly calculated maximum available size of the
635 * subtree is equal to the current one, then it means that
636 * the tree is propagated correctly. So we have to stop at
637 * this point to save cycles.
638 */
639 if (va->subtree_max_size == new_va_sub_max_size)
640 break;
641
642 va->subtree_max_size = new_va_sub_max_size;
643 node = rb_parent(&va->rb_node);
644 }
645
646#if DEBUG_AUGMENT_PROPAGATE_CHECK
647 augment_tree_propagate_check(free_vmap_area_root.rb_node);
648#endif
649}
650
651static void
652insert_vmap_area(struct vmap_area *va,
653 struct rb_root *root, struct list_head *head)
654{
655 struct rb_node **link;
656 struct rb_node *parent;
657
658 link = find_va_links(va, root, NULL, &parent);
659 link_va(va, root, parent, link, head);
660}
661
662static void
663insert_vmap_area_augment(struct vmap_area *va,
664 struct rb_node *from, struct rb_root *root,
665 struct list_head *head)
666{
667 struct rb_node **link;
668 struct rb_node *parent;
669
670 if (from)
671 link = find_va_links(va, NULL, from, &parent);
672 else
673 link = find_va_links(va, root, NULL, &parent);
674
675 link_va(va, root, parent, link, head);
676 augment_tree_propagate_from(va);
677}
678
679/*
680 * Merge de-allocated chunk of VA memory with previous
681 * and next free blocks. If coalesce is not done a new
682 * free area is inserted. If VA has been merged, it is
683 * freed.
684 */
685static __always_inline void
686merge_or_add_vmap_area(struct vmap_area *va,
687 struct rb_root *root, struct list_head *head)
688{
689 struct vmap_area *sibling;
690 struct list_head *next;
691 struct rb_node **link;
692 struct rb_node *parent;
693 bool merged = false;
694
695 /*
696 * Find a place in the tree where VA potentially will be
697 * inserted, unless it is merged with its sibling/siblings.
698 */
699 link = find_va_links(va, root, NULL, &parent);
700
701 /*
702 * Get next node of VA to check if merging can be done.
703 */
704 next = get_va_next_sibling(parent, link);
705 if (unlikely(next == NULL))
706 goto insert;
707
708 /*
709 * start end
710 * | |
711 * |<------VA------>|<-----Next----->|
712 * | |
713 * start end
714 */
715 if (next != head) {
716 sibling = list_entry(next, struct vmap_area, list);
717 if (sibling->va_start == va->va_end) {
718 sibling->va_start = va->va_start;
719
720 /* Check and update the tree if needed. */
721 augment_tree_propagate_from(sibling);
722
723 /* Free vmap_area object. */
724 kmem_cache_free(vmap_area_cachep, va);
725
726 /* Point to the new merged area. */
727 va = sibling;
728 merged = true;
729 }
730 }
731
732 /*
733 * start end
734 * | |
735 * |<-----Prev----->|<------VA------>|
736 * | |
737 * start end
738 */
739 if (next->prev != head) {
740 sibling = list_entry(next->prev, struct vmap_area, list);
741 if (sibling->va_end == va->va_start) {
742 sibling->va_end = va->va_end;
743
744 /* Check and update the tree if needed. */
745 augment_tree_propagate_from(sibling);
746
747 if (merged)
748 unlink_va(va, root);
749
750 /* Free vmap_area object. */
751 kmem_cache_free(vmap_area_cachep, va);
752 return;
753 }
754 }
755
756insert:
757 if (!merged) {
758 link_va(va, root, parent, link, head);
759 augment_tree_propagate_from(va);
760 }
761}
762
763static __always_inline bool
764is_within_this_va(struct vmap_area *va, unsigned long size,
765 unsigned long align, unsigned long vstart)
766{
767 unsigned long nva_start_addr;
768
769 if (va->va_start > vstart)
770 nva_start_addr = ALIGN(va->va_start, align);
771 else
772 nva_start_addr = ALIGN(vstart, align);
773
774 /* Can be overflowed due to big size or alignment. */
775 if (nva_start_addr + size < nva_start_addr ||
776 nva_start_addr < vstart)
777 return false;
778
779 return (nva_start_addr + size <= va->va_end);
780}
781
782/*
783 * Find the first free block(lowest start address) in the tree,
784 * that will accomplish the request corresponding to passing
785 * parameters.
786 */
787static __always_inline struct vmap_area *
788find_vmap_lowest_match(unsigned long size,
789 unsigned long align, unsigned long vstart)
790{
791 struct vmap_area *va;
792 struct rb_node *node;
793 unsigned long length;
794
795 /* Start from the root. */
796 node = free_vmap_area_root.rb_node;
797
798 /* Adjust the search size for alignment overhead. */
799 length = size + align - 1;
800
801 while (node) {
802 va = rb_entry(node, struct vmap_area, rb_node);
803
804 if (get_subtree_max_size(node->rb_left) >= length &&
805 vstart < va->va_start) {
806 node = node->rb_left;
807 } else {
808 if (is_within_this_va(va, size, align, vstart))
809 return va;
810
811 /*
812 * Does not make sense to go deeper towards the right
813 * sub-tree if it does not have a free block that is
814 * equal or bigger to the requested search length.
815 */
816 if (get_subtree_max_size(node->rb_right) >= length) {
817 node = node->rb_right;
818 continue;
819 }
820
821 /*
822 * OK. We roll back and find the first right sub-tree,
823 * that will satisfy the search criteria. It can happen
824 * only once due to "vstart" restriction.
825 */
826 while ((node = rb_parent(node))) {
827 va = rb_entry(node, struct vmap_area, rb_node);
828 if (is_within_this_va(va, size, align, vstart))
829 return va;
830
831 if (get_subtree_max_size(node->rb_right) >= length &&
832 vstart <= va->va_start) {
833 node = node->rb_right;
834 break;
835 }
836 }
837 }
838 }
839
840 return NULL;
841}
842
843#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
844#include <linux/random.h>
845
846static struct vmap_area *
847find_vmap_lowest_linear_match(unsigned long size,
848 unsigned long align, unsigned long vstart)
849{
850 struct vmap_area *va;
851
852 list_for_each_entry(va, &free_vmap_area_list, list) {
853 if (!is_within_this_va(va, size, align, vstart))
854 continue;
855
856 return va;
857 }
858
859 return NULL;
860}
861
862static void
863find_vmap_lowest_match_check(unsigned long size)
864{
865 struct vmap_area *va_1, *va_2;
866 unsigned long vstart;
867 unsigned int rnd;
868
869 get_random_bytes(&rnd, sizeof(rnd));
870 vstart = VMALLOC_START + rnd;
871
872 va_1 = find_vmap_lowest_match(size, 1, vstart);
873 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
874
875 if (va_1 != va_2)
876 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
877 va_1, va_2, vstart);
878}
879#endif
880
881enum fit_type {
882 NOTHING_FIT = 0,
883 FL_FIT_TYPE = 1, /* full fit */
884 LE_FIT_TYPE = 2, /* left edge fit */
885 RE_FIT_TYPE = 3, /* right edge fit */
886 NE_FIT_TYPE = 4 /* no edge fit */
887};
888
889static __always_inline enum fit_type
890classify_va_fit_type(struct vmap_area *va,
891 unsigned long nva_start_addr, unsigned long size)
892{
893 enum fit_type type;
894
895 /* Check if it is within VA. */
896 if (nva_start_addr < va->va_start ||
897 nva_start_addr + size > va->va_end)
898 return NOTHING_FIT;
899
900 /* Now classify. */
901 if (va->va_start == nva_start_addr) {
902 if (va->va_end == nva_start_addr + size)
903 type = FL_FIT_TYPE;
904 else
905 type = LE_FIT_TYPE;
906 } else if (va->va_end == nva_start_addr + size) {
907 type = RE_FIT_TYPE;
908 } else {
909 type = NE_FIT_TYPE;
910 }
911
912 return type;
913}
914
915static __always_inline int
916adjust_va_to_fit_type(struct vmap_area *va,
917 unsigned long nva_start_addr, unsigned long size,
918 enum fit_type type)
919{
920 struct vmap_area *lva = NULL;
921
922 if (type == FL_FIT_TYPE) {
923 /*
924 * No need to split VA, it fully fits.
925 *
926 * | |
927 * V NVA V
928 * |---------------|
929 */
930 unlink_va(va, &free_vmap_area_root);
931 kmem_cache_free(vmap_area_cachep, va);
932 } else if (type == LE_FIT_TYPE) {
933 /*
934 * Split left edge of fit VA.
935 *
936 * | |
937 * V NVA V R
938 * |-------|-------|
939 */
940 va->va_start += size;
941 } else if (type == RE_FIT_TYPE) {
942 /*
943 * Split right edge of fit VA.
944 *
945 * | |
946 * L V NVA V
947 * |-------|-------|
948 */
949 va->va_end = nva_start_addr;
950 } else if (type == NE_FIT_TYPE) {
951 /*
952 * Split no edge of fit VA.
953 *
954 * | |
955 * L V NVA V R
956 * |---|-------|---|
957 */
958 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
959 if (unlikely(!lva)) {
960 /*
961 * For percpu allocator we do not do any pre-allocation
962 * and leave it as it is. The reason is it most likely
963 * never ends up with NE_FIT_TYPE splitting. In case of
964 * percpu allocations offsets and sizes are aligned to
965 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
966 * are its main fitting cases.
967 *
968 * There are a few exceptions though, as an example it is
969 * a first allocation (early boot up) when we have "one"
970 * big free space that has to be split.
971 */
972 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
973 if (!lva)
974 return -1;
975 }
976
977 /*
978 * Build the remainder.
979 */
980 lva->va_start = va->va_start;
981 lva->va_end = nva_start_addr;
982
983 /*
984 * Shrink this VA to remaining size.
985 */
986 va->va_start = nva_start_addr + size;
987 } else {
988 return -1;
989 }
990
991 if (type != FL_FIT_TYPE) {
992 augment_tree_propagate_from(va);
993
994 if (lva) /* type == NE_FIT_TYPE */
995 insert_vmap_area_augment(lva, &va->rb_node,
996 &free_vmap_area_root, &free_vmap_area_list);
997 }
998
999 return 0;
1000}
1001
1002/*
1003 * Returns a start address of the newly allocated area, if success.
1004 * Otherwise a vend is returned that indicates failure.
1005 */
1006static __always_inline unsigned long
1007__alloc_vmap_area(unsigned long size, unsigned long align,
1008 unsigned long vstart, unsigned long vend)
1009{
1010 unsigned long nva_start_addr;
1011 struct vmap_area *va;
1012 enum fit_type type;
1013 int ret;
1014
1015 va = find_vmap_lowest_match(size, align, vstart);
1016 if (unlikely(!va))
1017 return vend;
1018
1019 if (va->va_start > vstart)
1020 nva_start_addr = ALIGN(va->va_start, align);
1021 else
1022 nva_start_addr = ALIGN(vstart, align);
1023
1024 /* Check the "vend" restriction. */
1025 if (nva_start_addr + size > vend)
1026 return vend;
1027
1028 /* Classify what we have found. */
1029 type = classify_va_fit_type(va, nva_start_addr, size);
1030 if (WARN_ON_ONCE(type == NOTHING_FIT))
1031 return vend;
1032
1033 /* Update the free vmap_area. */
1034 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1035 if (ret)
1036 return vend;
1037
1038#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1039 find_vmap_lowest_match_check(size);
1040#endif
1041
1042 return nva_start_addr;
1043}
1044
1045/*
1046 * Allocate a region of KVA of the specified size and alignment, within the
1047 * vstart and vend.
1048 */
1049static struct vmap_area *alloc_vmap_area(unsigned long size,
1050 unsigned long align,
1051 unsigned long vstart, unsigned long vend,
1052 int node, gfp_t gfp_mask)
1053{
1054 struct vmap_area *va, *pva;
1055 unsigned long addr;
1056 int purged = 0;
1057
1058 BUG_ON(!size);
1059 BUG_ON(offset_in_page(size));
1060 BUG_ON(!is_power_of_2(align));
1061
1062 if (unlikely(!vmap_initialized))
1063 return ERR_PTR(-EBUSY);
1064
1065 might_sleep();
1066
1067 va = kmem_cache_alloc_node(vmap_area_cachep,
1068 gfp_mask & GFP_RECLAIM_MASK, node);
1069 if (unlikely(!va))
1070 return ERR_PTR(-ENOMEM);
1071
1072 /*
1073 * Only scan the relevant parts containing pointers to other objects
1074 * to avoid false negatives.
1075 */
1076 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
1077
1078retry:
1079 /*
1080 * Preload this CPU with one extra vmap_area object to ensure
1081 * that we have it available when fit type of free area is
1082 * NE_FIT_TYPE.
1083 *
1084 * The preload is done in non-atomic context, thus it allows us
1085 * to use more permissive allocation masks to be more stable under
1086 * low memory condition and high memory pressure.
1087 *
1088 * Even if it fails we do not really care about that. Just proceed
1089 * as it is. "overflow" path will refill the cache we allocate from.
1090 */
1091 preempt_disable();
1092 if (!__this_cpu_read(ne_fit_preload_node)) {
1093 preempt_enable();
1094 pva = kmem_cache_alloc_node(vmap_area_cachep, GFP_KERNEL, node);
1095 preempt_disable();
1096
1097 if (__this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva)) {
1098 if (pva)
1099 kmem_cache_free(vmap_area_cachep, pva);
1100 }
1101 }
1102
1103 spin_lock(&vmap_area_lock);
1104 preempt_enable();
1105
1106 /*
1107 * If an allocation fails, the "vend" address is
1108 * returned. Therefore trigger the overflow path.
1109 */
1110 addr = __alloc_vmap_area(size, align, vstart, vend);
1111 if (unlikely(addr == vend))
1112 goto overflow;
1113
1114 va->va_start = addr;
1115 va->va_end = addr + size;
1116 va->vm = NULL;
1117 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1118
1119 spin_unlock(&vmap_area_lock);
1120
1121 BUG_ON(!IS_ALIGNED(va->va_start, align));
1122 BUG_ON(va->va_start < vstart);
1123 BUG_ON(va->va_end > vend);
1124
1125 return va;
1126
1127overflow:
1128 spin_unlock(&vmap_area_lock);
1129 if (!purged) {
1130 purge_vmap_area_lazy();
1131 purged = 1;
1132 goto retry;
1133 }
1134
1135 if (gfpflags_allow_blocking(gfp_mask)) {
1136 unsigned long freed = 0;
1137 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1138 if (freed > 0) {
1139 purged = 0;
1140 goto retry;
1141 }
1142 }
1143
1144 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1145 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1146 size);
1147
1148 kmem_cache_free(vmap_area_cachep, va);
1149 return ERR_PTR(-EBUSY);
1150}
1151
1152int register_vmap_purge_notifier(struct notifier_block *nb)
1153{
1154 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1155}
1156EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1157
1158int unregister_vmap_purge_notifier(struct notifier_block *nb)
1159{
1160 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1161}
1162EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1163
1164static void __free_vmap_area(struct vmap_area *va)
1165{
1166 /*
1167 * Remove from the busy tree/list.
1168 */
1169 unlink_va(va, &vmap_area_root);
1170
1171 /*
1172 * Merge VA with its neighbors, otherwise just add it.
1173 */
1174 merge_or_add_vmap_area(va,
1175 &free_vmap_area_root, &free_vmap_area_list);
1176}
1177
1178/*
1179 * Free a region of KVA allocated by alloc_vmap_area
1180 */
1181static void free_vmap_area(struct vmap_area *va)
1182{
1183 spin_lock(&vmap_area_lock);
1184 __free_vmap_area(va);
1185 spin_unlock(&vmap_area_lock);
1186}
1187
1188/*
1189 * Clear the pagetable entries of a given vmap_area
1190 */
1191static void unmap_vmap_area(struct vmap_area *va)
1192{
1193 vunmap_page_range(va->va_start, va->va_end);
1194}
1195
1196/*
1197 * lazy_max_pages is the maximum amount of virtual address space we gather up
1198 * before attempting to purge with a TLB flush.
1199 *
1200 * There is a tradeoff here: a larger number will cover more kernel page tables
1201 * and take slightly longer to purge, but it will linearly reduce the number of
1202 * global TLB flushes that must be performed. It would seem natural to scale
1203 * this number up linearly with the number of CPUs (because vmapping activity
1204 * could also scale linearly with the number of CPUs), however it is likely
1205 * that in practice, workloads might be constrained in other ways that mean
1206 * vmap activity will not scale linearly with CPUs. Also, I want to be
1207 * conservative and not introduce a big latency on huge systems, so go with
1208 * a less aggressive log scale. It will still be an improvement over the old
1209 * code, and it will be simple to change the scale factor if we find that it
1210 * becomes a problem on bigger systems.
1211 */
1212static unsigned long lazy_max_pages(void)
1213{
1214 unsigned int log;
1215
1216 log = fls(num_online_cpus());
1217
1218 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1219}
1220
1221static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1222
1223/*
1224 * Serialize vmap purging. There is no actual criticial section protected
1225 * by this look, but we want to avoid concurrent calls for performance
1226 * reasons and to make the pcpu_get_vm_areas more deterministic.
1227 */
1228static DEFINE_MUTEX(vmap_purge_lock);
1229
1230/* for per-CPU blocks */
1231static void purge_fragmented_blocks_allcpus(void);
1232
1233/*
1234 * called before a call to iounmap() if the caller wants vm_area_struct's
1235 * immediately freed.
1236 */
1237void set_iounmap_nonlazy(void)
1238{
1239 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1240}
1241
1242/*
1243 * Purges all lazily-freed vmap areas.
1244 */
1245static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1246{
1247 unsigned long resched_threshold;
1248 struct llist_node *valist;
1249 struct vmap_area *va;
1250 struct vmap_area *n_va;
1251
1252 lockdep_assert_held(&vmap_purge_lock);
1253
1254 valist = llist_del_all(&vmap_purge_list);
1255 if (unlikely(valist == NULL))
1256 return false;
1257
1258 /*
1259 * First make sure the mappings are removed from all page-tables
1260 * before they are freed.
1261 */
1262 vmalloc_sync_all();
1263
1264 /*
1265 * TODO: to calculate a flush range without looping.
1266 * The list can be up to lazy_max_pages() elements.
1267 */
1268 llist_for_each_entry(va, valist, purge_list) {
1269 if (va->va_start < start)
1270 start = va->va_start;
1271 if (va->va_end > end)
1272 end = va->va_end;
1273 }
1274
1275 flush_tlb_kernel_range(start, end);
1276 resched_threshold = lazy_max_pages() << 1;
1277
1278 spin_lock(&vmap_area_lock);
1279 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1280 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1281
1282 /*
1283 * Finally insert or merge lazily-freed area. It is
1284 * detached and there is no need to "unlink" it from
1285 * anything.
1286 */
1287 merge_or_add_vmap_area(va,
1288 &free_vmap_area_root, &free_vmap_area_list);
1289
1290 atomic_long_sub(nr, &vmap_lazy_nr);
1291
1292 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1293 cond_resched_lock(&vmap_area_lock);
1294 }
1295 spin_unlock(&vmap_area_lock);
1296 return true;
1297}
1298
1299/*
1300 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1301 * is already purging.
1302 */
1303static void try_purge_vmap_area_lazy(void)
1304{
1305 if (mutex_trylock(&vmap_purge_lock)) {
1306 __purge_vmap_area_lazy(ULONG_MAX, 0);
1307 mutex_unlock(&vmap_purge_lock);
1308 }
1309}
1310
1311/*
1312 * Kick off a purge of the outstanding lazy areas.
1313 */
1314static void purge_vmap_area_lazy(void)
1315{
1316 mutex_lock(&vmap_purge_lock);
1317 purge_fragmented_blocks_allcpus();
1318 __purge_vmap_area_lazy(ULONG_MAX, 0);
1319 mutex_unlock(&vmap_purge_lock);
1320}
1321
1322/*
1323 * Free a vmap area, caller ensuring that the area has been unmapped
1324 * and flush_cache_vunmap had been called for the correct range
1325 * previously.
1326 */
1327static void free_vmap_area_noflush(struct vmap_area *va)
1328{
1329 unsigned long nr_lazy;
1330
1331 spin_lock(&vmap_area_lock);
1332 unlink_va(va, &vmap_area_root);
1333 spin_unlock(&vmap_area_lock);
1334
1335 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1336 PAGE_SHIFT, &vmap_lazy_nr);
1337
1338 /* After this point, we may free va at any time */
1339 llist_add(&va->purge_list, &vmap_purge_list);
1340
1341 if (unlikely(nr_lazy > lazy_max_pages()))
1342 try_purge_vmap_area_lazy();
1343}
1344
1345/*
1346 * Free and unmap a vmap area
1347 */
1348static void free_unmap_vmap_area(struct vmap_area *va)
1349{
1350 flush_cache_vunmap(va->va_start, va->va_end);
1351 unmap_vmap_area(va);
1352 if (debug_pagealloc_enabled())
1353 flush_tlb_kernel_range(va->va_start, va->va_end);
1354
1355 free_vmap_area_noflush(va);
1356}
1357
1358static struct vmap_area *find_vmap_area(unsigned long addr)
1359{
1360 struct vmap_area *va;
1361
1362 spin_lock(&vmap_area_lock);
1363 va = __find_vmap_area(addr);
1364 spin_unlock(&vmap_area_lock);
1365
1366 return va;
1367}
1368
1369/*** Per cpu kva allocator ***/
1370
1371/*
1372 * vmap space is limited especially on 32 bit architectures. Ensure there is
1373 * room for at least 16 percpu vmap blocks per CPU.
1374 */
1375/*
1376 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1377 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1378 * instead (we just need a rough idea)
1379 */
1380#if BITS_PER_LONG == 32
1381#define VMALLOC_SPACE (128UL*1024*1024)
1382#else
1383#define VMALLOC_SPACE (128UL*1024*1024*1024)
1384#endif
1385
1386#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1387#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1388#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1389#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1390#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1391#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1392#define VMAP_BBMAP_BITS \
1393 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1394 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1395 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1396
1397#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1398
1399struct vmap_block_queue {
1400 spinlock_t lock;
1401 struct list_head free;
1402};
1403
1404struct vmap_block {
1405 spinlock_t lock;
1406 struct vmap_area *va;
1407 unsigned long free, dirty;
1408 unsigned long dirty_min, dirty_max; /*< dirty range */
1409 struct list_head free_list;
1410 struct rcu_head rcu_head;
1411 struct list_head purge;
1412};
1413
1414/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1415static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1416
1417/*
1418 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1419 * in the free path. Could get rid of this if we change the API to return a
1420 * "cookie" from alloc, to be passed to free. But no big deal yet.
1421 */
1422static DEFINE_SPINLOCK(vmap_block_tree_lock);
1423static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
1424
1425/*
1426 * We should probably have a fallback mechanism to allocate virtual memory
1427 * out of partially filled vmap blocks. However vmap block sizing should be
1428 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1429 * big problem.
1430 */
1431
1432static unsigned long addr_to_vb_idx(unsigned long addr)
1433{
1434 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1435 addr /= VMAP_BLOCK_SIZE;
1436 return addr;
1437}
1438
1439static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1440{
1441 unsigned long addr;
1442
1443 addr = va_start + (pages_off << PAGE_SHIFT);
1444 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1445 return (void *)addr;
1446}
1447
1448/**
1449 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1450 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1451 * @order: how many 2^order pages should be occupied in newly allocated block
1452 * @gfp_mask: flags for the page level allocator
1453 *
1454 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1455 */
1456static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1457{
1458 struct vmap_block_queue *vbq;
1459 struct vmap_block *vb;
1460 struct vmap_area *va;
1461 unsigned long vb_idx;
1462 int node, err;
1463 void *vaddr;
1464
1465 node = numa_node_id();
1466
1467 vb = kmalloc_node(sizeof(struct vmap_block),
1468 gfp_mask & GFP_RECLAIM_MASK, node);
1469 if (unlikely(!vb))
1470 return ERR_PTR(-ENOMEM);
1471
1472 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1473 VMALLOC_START, VMALLOC_END,
1474 node, gfp_mask);
1475 if (IS_ERR(va)) {
1476 kfree(vb);
1477 return ERR_CAST(va);
1478 }
1479
1480 err = radix_tree_preload(gfp_mask);
1481 if (unlikely(err)) {
1482 kfree(vb);
1483 free_vmap_area(va);
1484 return ERR_PTR(err);
1485 }
1486
1487 vaddr = vmap_block_vaddr(va->va_start, 0);
1488 spin_lock_init(&vb->lock);
1489 vb->va = va;
1490 /* At least something should be left free */
1491 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1492 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1493 vb->dirty = 0;
1494 vb->dirty_min = VMAP_BBMAP_BITS;
1495 vb->dirty_max = 0;
1496 INIT_LIST_HEAD(&vb->free_list);
1497
1498 vb_idx = addr_to_vb_idx(va->va_start);
1499 spin_lock(&vmap_block_tree_lock);
1500 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
1501 spin_unlock(&vmap_block_tree_lock);
1502 BUG_ON(err);
1503 radix_tree_preload_end();
1504
1505 vbq = &get_cpu_var(vmap_block_queue);
1506 spin_lock(&vbq->lock);
1507 list_add_tail_rcu(&vb->free_list, &vbq->free);
1508 spin_unlock(&vbq->lock);
1509 put_cpu_var(vmap_block_queue);
1510
1511 return vaddr;
1512}
1513
1514static void free_vmap_block(struct vmap_block *vb)
1515{
1516 struct vmap_block *tmp;
1517 unsigned long vb_idx;
1518
1519 vb_idx = addr_to_vb_idx(vb->va->va_start);
1520 spin_lock(&vmap_block_tree_lock);
1521 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
1522 spin_unlock(&vmap_block_tree_lock);
1523 BUG_ON(tmp != vb);
1524
1525 free_vmap_area_noflush(vb->va);
1526 kfree_rcu(vb, rcu_head);
1527}
1528
1529static void purge_fragmented_blocks(int cpu)
1530{
1531 LIST_HEAD(purge);
1532 struct vmap_block *vb;
1533 struct vmap_block *n_vb;
1534 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1535
1536 rcu_read_lock();
1537 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1538
1539 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1540 continue;
1541
1542 spin_lock(&vb->lock);
1543 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1544 vb->free = 0; /* prevent further allocs after releasing lock */
1545 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1546 vb->dirty_min = 0;
1547 vb->dirty_max = VMAP_BBMAP_BITS;
1548 spin_lock(&vbq->lock);
1549 list_del_rcu(&vb->free_list);
1550 spin_unlock(&vbq->lock);
1551 spin_unlock(&vb->lock);
1552 list_add_tail(&vb->purge, &purge);
1553 } else
1554 spin_unlock(&vb->lock);
1555 }
1556 rcu_read_unlock();
1557
1558 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1559 list_del(&vb->purge);
1560 free_vmap_block(vb);
1561 }
1562}
1563
1564static void purge_fragmented_blocks_allcpus(void)
1565{
1566 int cpu;
1567
1568 for_each_possible_cpu(cpu)
1569 purge_fragmented_blocks(cpu);
1570}
1571
1572static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1573{
1574 struct vmap_block_queue *vbq;
1575 struct vmap_block *vb;
1576 void *vaddr = NULL;
1577 unsigned int order;
1578
1579 BUG_ON(offset_in_page(size));
1580 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1581 if (WARN_ON(size == 0)) {
1582 /*
1583 * Allocating 0 bytes isn't what caller wants since
1584 * get_order(0) returns funny result. Just warn and terminate
1585 * early.
1586 */
1587 return NULL;
1588 }
1589 order = get_order(size);
1590
1591 rcu_read_lock();
1592 vbq = &get_cpu_var(vmap_block_queue);
1593 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1594 unsigned long pages_off;
1595
1596 spin_lock(&vb->lock);
1597 if (vb->free < (1UL << order)) {
1598 spin_unlock(&vb->lock);
1599 continue;
1600 }
1601
1602 pages_off = VMAP_BBMAP_BITS - vb->free;
1603 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1604 vb->free -= 1UL << order;
1605 if (vb->free == 0) {
1606 spin_lock(&vbq->lock);
1607 list_del_rcu(&vb->free_list);
1608 spin_unlock(&vbq->lock);
1609 }
1610
1611 spin_unlock(&vb->lock);
1612 break;
1613 }
1614
1615 put_cpu_var(vmap_block_queue);
1616 rcu_read_unlock();
1617
1618 /* Allocate new block if nothing was found */
1619 if (!vaddr)
1620 vaddr = new_vmap_block(order, gfp_mask);
1621
1622 return vaddr;
1623}
1624
1625static void vb_free(const void *addr, unsigned long size)
1626{
1627 unsigned long offset;
1628 unsigned long vb_idx;
1629 unsigned int order;
1630 struct vmap_block *vb;
1631
1632 BUG_ON(offset_in_page(size));
1633 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1634
1635 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1636
1637 order = get_order(size);
1638
1639 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1640 offset >>= PAGE_SHIFT;
1641
1642 vb_idx = addr_to_vb_idx((unsigned long)addr);
1643 rcu_read_lock();
1644 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1645 rcu_read_unlock();
1646 BUG_ON(!vb);
1647
1648 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1649
1650 if (debug_pagealloc_enabled())
1651 flush_tlb_kernel_range((unsigned long)addr,
1652 (unsigned long)addr + size);
1653
1654 spin_lock(&vb->lock);
1655
1656 /* Expand dirty range */
1657 vb->dirty_min = min(vb->dirty_min, offset);
1658 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1659
1660 vb->dirty += 1UL << order;
1661 if (vb->dirty == VMAP_BBMAP_BITS) {
1662 BUG_ON(vb->free);
1663 spin_unlock(&vb->lock);
1664 free_vmap_block(vb);
1665 } else
1666 spin_unlock(&vb->lock);
1667}
1668
1669static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1670{
1671 int cpu;
1672
1673 if (unlikely(!vmap_initialized))
1674 return;
1675
1676 might_sleep();
1677
1678 for_each_possible_cpu(cpu) {
1679 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1680 struct vmap_block *vb;
1681
1682 rcu_read_lock();
1683 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1684 spin_lock(&vb->lock);
1685 if (vb->dirty) {
1686 unsigned long va_start = vb->va->va_start;
1687 unsigned long s, e;
1688
1689 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1690 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1691
1692 start = min(s, start);
1693 end = max(e, end);
1694
1695 flush = 1;
1696 }
1697 spin_unlock(&vb->lock);
1698 }
1699 rcu_read_unlock();
1700 }
1701
1702 mutex_lock(&vmap_purge_lock);
1703 purge_fragmented_blocks_allcpus();
1704 if (!__purge_vmap_area_lazy(start, end) && flush)
1705 flush_tlb_kernel_range(start, end);
1706 mutex_unlock(&vmap_purge_lock);
1707}
1708
1709/**
1710 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1711 *
1712 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1713 * to amortize TLB flushing overheads. What this means is that any page you
1714 * have now, may, in a former life, have been mapped into kernel virtual
1715 * address by the vmap layer and so there might be some CPUs with TLB entries
1716 * still referencing that page (additional to the regular 1:1 kernel mapping).
1717 *
1718 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1719 * be sure that none of the pages we have control over will have any aliases
1720 * from the vmap layer.
1721 */
1722void vm_unmap_aliases(void)
1723{
1724 unsigned long start = ULONG_MAX, end = 0;
1725 int flush = 0;
1726
1727 _vm_unmap_aliases(start, end, flush);
1728}
1729EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1730
1731/**
1732 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1733 * @mem: the pointer returned by vm_map_ram
1734 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1735 */
1736void vm_unmap_ram(const void *mem, unsigned int count)
1737{
1738 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1739 unsigned long addr = (unsigned long)mem;
1740 struct vmap_area *va;
1741
1742 might_sleep();
1743 BUG_ON(!addr);
1744 BUG_ON(addr < VMALLOC_START);
1745 BUG_ON(addr > VMALLOC_END);
1746 BUG_ON(!PAGE_ALIGNED(addr));
1747
1748 if (likely(count <= VMAP_MAX_ALLOC)) {
1749 debug_check_no_locks_freed(mem, size);
1750 vb_free(mem, size);
1751 return;
1752 }
1753
1754 va = find_vmap_area(addr);
1755 BUG_ON(!va);
1756 debug_check_no_locks_freed((void *)va->va_start,
1757 (va->va_end - va->va_start));
1758 free_unmap_vmap_area(va);
1759}
1760EXPORT_SYMBOL(vm_unmap_ram);
1761
1762/**
1763 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1764 * @pages: an array of pointers to the pages to be mapped
1765 * @count: number of pages
1766 * @node: prefer to allocate data structures on this node
1767 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1768 *
1769 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1770 * faster than vmap so it's good. But if you mix long-life and short-life
1771 * objects with vm_map_ram(), it could consume lots of address space through
1772 * fragmentation (especially on a 32bit machine). You could see failures in
1773 * the end. Please use this function for short-lived objects.
1774 *
1775 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1776 */
1777void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1778{
1779 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1780 unsigned long addr;
1781 void *mem;
1782
1783 if (likely(count <= VMAP_MAX_ALLOC)) {
1784 mem = vb_alloc(size, GFP_KERNEL);
1785 if (IS_ERR(mem))
1786 return NULL;
1787 addr = (unsigned long)mem;
1788 } else {
1789 struct vmap_area *va;
1790 va = alloc_vmap_area(size, PAGE_SIZE,
1791 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1792 if (IS_ERR(va))
1793 return NULL;
1794
1795 addr = va->va_start;
1796 mem = (void *)addr;
1797 }
1798 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1799 vm_unmap_ram(mem, count);
1800 return NULL;
1801 }
1802 return mem;
1803}
1804EXPORT_SYMBOL(vm_map_ram);
1805
1806static struct vm_struct *vmlist __initdata;
1807
1808/**
1809 * vm_area_add_early - add vmap area early during boot
1810 * @vm: vm_struct to add
1811 *
1812 * This function is used to add fixed kernel vm area to vmlist before
1813 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1814 * should contain proper values and the other fields should be zero.
1815 *
1816 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1817 */
1818void __init vm_area_add_early(struct vm_struct *vm)
1819{
1820 struct vm_struct *tmp, **p;
1821
1822 BUG_ON(vmap_initialized);
1823 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1824 if (tmp->addr >= vm->addr) {
1825 BUG_ON(tmp->addr < vm->addr + vm->size);
1826 break;
1827 } else
1828 BUG_ON(tmp->addr + tmp->size > vm->addr);
1829 }
1830 vm->next = *p;
1831 *p = vm;
1832}
1833
1834/**
1835 * vm_area_register_early - register vmap area early during boot
1836 * @vm: vm_struct to register
1837 * @align: requested alignment
1838 *
1839 * This function is used to register kernel vm area before
1840 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1841 * proper values on entry and other fields should be zero. On return,
1842 * vm->addr contains the allocated address.
1843 *
1844 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1845 */
1846void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1847{
1848 static size_t vm_init_off __initdata;
1849 unsigned long addr;
1850
1851 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1852 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1853
1854 vm->addr = (void *)addr;
1855
1856 vm_area_add_early(vm);
1857}
1858
1859static void vmap_init_free_space(void)
1860{
1861 unsigned long vmap_start = 1;
1862 const unsigned long vmap_end = ULONG_MAX;
1863 struct vmap_area *busy, *free;
1864
1865 /*
1866 * B F B B B F
1867 * -|-----|.....|-----|-----|-----|.....|-
1868 * | The KVA space |
1869 * |<--------------------------------->|
1870 */
1871 list_for_each_entry(busy, &vmap_area_list, list) {
1872 if (busy->va_start - vmap_start > 0) {
1873 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1874 if (!WARN_ON_ONCE(!free)) {
1875 free->va_start = vmap_start;
1876 free->va_end = busy->va_start;
1877
1878 insert_vmap_area_augment(free, NULL,
1879 &free_vmap_area_root,
1880 &free_vmap_area_list);
1881 }
1882 }
1883
1884 vmap_start = busy->va_end;
1885 }
1886
1887 if (vmap_end - vmap_start > 0) {
1888 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1889 if (!WARN_ON_ONCE(!free)) {
1890 free->va_start = vmap_start;
1891 free->va_end = vmap_end;
1892
1893 insert_vmap_area_augment(free, NULL,
1894 &free_vmap_area_root,
1895 &free_vmap_area_list);
1896 }
1897 }
1898}
1899
1900void __init vmalloc_init(void)
1901{
1902 struct vmap_area *va;
1903 struct vm_struct *tmp;
1904 int i;
1905
1906 /*
1907 * Create the cache for vmap_area objects.
1908 */
1909 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1910
1911 for_each_possible_cpu(i) {
1912 struct vmap_block_queue *vbq;
1913 struct vfree_deferred *p;
1914
1915 vbq = &per_cpu(vmap_block_queue, i);
1916 spin_lock_init(&vbq->lock);
1917 INIT_LIST_HEAD(&vbq->free);
1918 p = &per_cpu(vfree_deferred, i);
1919 init_llist_head(&p->list);
1920 INIT_WORK(&p->wq, free_work);
1921 }
1922
1923 /* Import existing vmlist entries. */
1924 for (tmp = vmlist; tmp; tmp = tmp->next) {
1925 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1926 if (WARN_ON_ONCE(!va))
1927 continue;
1928
1929 va->va_start = (unsigned long)tmp->addr;
1930 va->va_end = va->va_start + tmp->size;
1931 va->vm = tmp;
1932 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1933 }
1934
1935 /*
1936 * Now we can initialize a free vmap space.
1937 */
1938 vmap_init_free_space();
1939 vmap_initialized = true;
1940}
1941
1942/**
1943 * map_kernel_range_noflush - map kernel VM area with the specified pages
1944 * @addr: start of the VM area to map
1945 * @size: size of the VM area to map
1946 * @prot: page protection flags to use
1947 * @pages: pages to map
1948 *
1949 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1950 * specify should have been allocated using get_vm_area() and its
1951 * friends.
1952 *
1953 * NOTE:
1954 * This function does NOT do any cache flushing. The caller is
1955 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1956 * before calling this function.
1957 *
1958 * RETURNS:
1959 * The number of pages mapped on success, -errno on failure.
1960 */
1961int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1962 pgprot_t prot, struct page **pages)
1963{
1964 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1965}
1966
1967/**
1968 * unmap_kernel_range_noflush - unmap kernel VM area
1969 * @addr: start of the VM area to unmap
1970 * @size: size of the VM area to unmap
1971 *
1972 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1973 * specify should have been allocated using get_vm_area() and its
1974 * friends.
1975 *
1976 * NOTE:
1977 * This function does NOT do any cache flushing. The caller is
1978 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1979 * before calling this function and flush_tlb_kernel_range() after.
1980 */
1981void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1982{
1983 vunmap_page_range(addr, addr + size);
1984}
1985EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1986
1987/**
1988 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1989 * @addr: start of the VM area to unmap
1990 * @size: size of the VM area to unmap
1991 *
1992 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1993 * the unmapping and tlb after.
1994 */
1995void unmap_kernel_range(unsigned long addr, unsigned long size)
1996{
1997 unsigned long end = addr + size;
1998
1999 flush_cache_vunmap(addr, end);
2000 vunmap_page_range(addr, end);
2001 flush_tlb_kernel_range(addr, end);
2002}
2003EXPORT_SYMBOL_GPL(unmap_kernel_range);
2004
2005int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
2006{
2007 unsigned long addr = (unsigned long)area->addr;
2008 unsigned long end = addr + get_vm_area_size(area);
2009 int err;
2010
2011 err = vmap_page_range(addr, end, prot, pages);
2012
2013 return err > 0 ? 0 : err;
2014}
2015EXPORT_SYMBOL_GPL(map_vm_area);
2016
2017static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2018 unsigned long flags, const void *caller)
2019{
2020 spin_lock(&vmap_area_lock);
2021 vm->flags = flags;
2022 vm->addr = (void *)va->va_start;
2023 vm->size = va->va_end - va->va_start;
2024 vm->caller = caller;
2025 va->vm = vm;
2026 spin_unlock(&vmap_area_lock);
2027}
2028
2029static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2030{
2031 /*
2032 * Before removing VM_UNINITIALIZED,
2033 * we should make sure that vm has proper values.
2034 * Pair with smp_rmb() in show_numa_info().
2035 */
2036 smp_wmb();
2037 vm->flags &= ~VM_UNINITIALIZED;
2038}
2039
2040static struct vm_struct *__get_vm_area_node(unsigned long size,
2041 unsigned long align, unsigned long flags, unsigned long start,
2042 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2043{
2044 struct vmap_area *va;
2045 struct vm_struct *area;
2046
2047 BUG_ON(in_interrupt());
2048 size = PAGE_ALIGN(size);
2049 if (unlikely(!size))
2050 return NULL;
2051
2052 if (flags & VM_IOREMAP)
2053 align = 1ul << clamp_t(int, get_count_order_long(size),
2054 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2055
2056 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2057 if (unlikely(!area))
2058 return NULL;
2059
2060 if (!(flags & VM_NO_GUARD))
2061 size += PAGE_SIZE;
2062
2063 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2064 if (IS_ERR(va)) {
2065 kfree(area);
2066 return NULL;
2067 }
2068
2069 setup_vmalloc_vm(area, va, flags, caller);
2070
2071 return area;
2072}
2073
2074struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
2075 unsigned long start, unsigned long end)
2076{
2077 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2078 GFP_KERNEL, __builtin_return_address(0));
2079}
2080EXPORT_SYMBOL_GPL(__get_vm_area);
2081
2082struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2083 unsigned long start, unsigned long end,
2084 const void *caller)
2085{
2086 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2087 GFP_KERNEL, caller);
2088}
2089
2090/**
2091 * get_vm_area - reserve a contiguous kernel virtual area
2092 * @size: size of the area
2093 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2094 *
2095 * Search an area of @size in the kernel virtual mapping area,
2096 * and reserved it for out purposes. Returns the area descriptor
2097 * on success or %NULL on failure.
2098 *
2099 * Return: the area descriptor on success or %NULL on failure.
2100 */
2101struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2102{
2103 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2104 NUMA_NO_NODE, GFP_KERNEL,
2105 __builtin_return_address(0));
2106}
2107
2108struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2109 const void *caller)
2110{
2111 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2112 NUMA_NO_NODE, GFP_KERNEL, caller);
2113}
2114
2115/**
2116 * find_vm_area - find a continuous kernel virtual area
2117 * @addr: base address
2118 *
2119 * Search for the kernel VM area starting at @addr, and return it.
2120 * It is up to the caller to do all required locking to keep the returned
2121 * pointer valid.
2122 *
2123 * Return: pointer to the found area or %NULL on faulure
2124 */
2125struct vm_struct *find_vm_area(const void *addr)
2126{
2127 struct vmap_area *va;
2128
2129 va = find_vmap_area((unsigned long)addr);
2130 if (!va)
2131 return NULL;
2132
2133 return va->vm;
2134}
2135
2136/**
2137 * remove_vm_area - find and remove a continuous kernel virtual area
2138 * @addr: base address
2139 *
2140 * Search for the kernel VM area starting at @addr, and remove it.
2141 * This function returns the found VM area, but using it is NOT safe
2142 * on SMP machines, except for its size or flags.
2143 *
2144 * Return: pointer to the found area or %NULL on faulure
2145 */
2146struct vm_struct *remove_vm_area(const void *addr)
2147{
2148 struct vmap_area *va;
2149
2150 might_sleep();
2151
2152 spin_lock(&vmap_area_lock);
2153 va = __find_vmap_area((unsigned long)addr);
2154 if (va && va->vm) {
2155 struct vm_struct *vm = va->vm;
2156
2157 va->vm = NULL;
2158 spin_unlock(&vmap_area_lock);
2159
2160 kasan_free_shadow(vm);
2161 free_unmap_vmap_area(va);
2162
2163 return vm;
2164 }
2165
2166 spin_unlock(&vmap_area_lock);
2167 return NULL;
2168}
2169
2170static inline void set_area_direct_map(const struct vm_struct *area,
2171 int (*set_direct_map)(struct page *page))
2172{
2173 int i;
2174
2175 for (i = 0; i < area->nr_pages; i++)
2176 if (page_address(area->pages[i]))
2177 set_direct_map(area->pages[i]);
2178}
2179
2180/* Handle removing and resetting vm mappings related to the vm_struct. */
2181static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2182{
2183 unsigned long start = ULONG_MAX, end = 0;
2184 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2185 int flush_dmap = 0;
2186 int i;
2187
2188 remove_vm_area(area->addr);
2189
2190 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2191 if (!flush_reset)
2192 return;
2193
2194 /*
2195 * If not deallocating pages, just do the flush of the VM area and
2196 * return.
2197 */
2198 if (!deallocate_pages) {
2199 vm_unmap_aliases();
2200 return;
2201 }
2202
2203 /*
2204 * If execution gets here, flush the vm mapping and reset the direct
2205 * map. Find the start and end range of the direct mappings to make sure
2206 * the vm_unmap_aliases() flush includes the direct map.
2207 */
2208 for (i = 0; i < area->nr_pages; i++) {
2209 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2210 if (addr) {
2211 start = min(addr, start);
2212 end = max(addr + PAGE_SIZE, end);
2213 flush_dmap = 1;
2214 }
2215 }
2216
2217 /*
2218 * Set direct map to something invalid so that it won't be cached if
2219 * there are any accesses after the TLB flush, then flush the TLB and
2220 * reset the direct map permissions to the default.
2221 */
2222 set_area_direct_map(area, set_direct_map_invalid_noflush);
2223 _vm_unmap_aliases(start, end, flush_dmap);
2224 set_area_direct_map(area, set_direct_map_default_noflush);
2225}
2226
2227static void __vunmap(const void *addr, int deallocate_pages)
2228{
2229 struct vm_struct *area;
2230
2231 if (!addr)
2232 return;
2233
2234 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2235 addr))
2236 return;
2237
2238 area = find_vm_area(addr);
2239 if (unlikely(!area)) {
2240 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2241 addr);
2242 return;
2243 }
2244
2245 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2246 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2247
2248 vm_remove_mappings(area, deallocate_pages);
2249
2250 if (deallocate_pages) {
2251 int i;
2252
2253 for (i = 0; i < area->nr_pages; i++) {
2254 struct page *page = area->pages[i];
2255
2256 BUG_ON(!page);
2257 __free_pages(page, 0);
2258 }
2259 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2260
2261 kvfree(area->pages);
2262 }
2263
2264 kfree(area);
2265 return;
2266}
2267
2268static inline void __vfree_deferred(const void *addr)
2269{
2270 /*
2271 * Use raw_cpu_ptr() because this can be called from preemptible
2272 * context. Preemption is absolutely fine here, because the llist_add()
2273 * implementation is lockless, so it works even if we are adding to
2274 * nother cpu's list. schedule_work() should be fine with this too.
2275 */
2276 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2277
2278 if (llist_add((struct llist_node *)addr, &p->list))
2279 schedule_work(&p->wq);
2280}
2281
2282/**
2283 * vfree_atomic - release memory allocated by vmalloc()
2284 * @addr: memory base address
2285 *
2286 * This one is just like vfree() but can be called in any atomic context
2287 * except NMIs.
2288 */
2289void vfree_atomic(const void *addr)
2290{
2291 BUG_ON(in_nmi());
2292
2293 kmemleak_free(addr);
2294
2295 if (!addr)
2296 return;
2297 __vfree_deferred(addr);
2298}
2299
2300static void __vfree(const void *addr)
2301{
2302 if (unlikely(in_interrupt()))
2303 __vfree_deferred(addr);
2304 else
2305 __vunmap(addr, 1);
2306}
2307
2308/**
2309 * vfree - release memory allocated by vmalloc()
2310 * @addr: memory base address
2311 *
2312 * Free the virtually continuous memory area starting at @addr, as
2313 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2314 * NULL, no operation is performed.
2315 *
2316 * Must not be called in NMI context (strictly speaking, only if we don't
2317 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2318 * conventions for vfree() arch-depenedent would be a really bad idea)
2319 *
2320 * May sleep if called *not* from interrupt context.
2321 *
2322 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2323 */
2324void vfree(const void *addr)
2325{
2326 BUG_ON(in_nmi());
2327
2328 kmemleak_free(addr);
2329
2330 might_sleep_if(!in_interrupt());
2331
2332 if (!addr)
2333 return;
2334
2335 __vfree(addr);
2336}
2337EXPORT_SYMBOL(vfree);
2338
2339/**
2340 * vunmap - release virtual mapping obtained by vmap()
2341 * @addr: memory base address
2342 *
2343 * Free the virtually contiguous memory area starting at @addr,
2344 * which was created from the page array passed to vmap().
2345 *
2346 * Must not be called in interrupt context.
2347 */
2348void vunmap(const void *addr)
2349{
2350 BUG_ON(in_interrupt());
2351 might_sleep();
2352 if (addr)
2353 __vunmap(addr, 0);
2354}
2355EXPORT_SYMBOL(vunmap);
2356
2357/**
2358 * vmap - map an array of pages into virtually contiguous space
2359 * @pages: array of page pointers
2360 * @count: number of pages to map
2361 * @flags: vm_area->flags
2362 * @prot: page protection for the mapping
2363 *
2364 * Maps @count pages from @pages into contiguous kernel virtual
2365 * space.
2366 *
2367 * Return: the address of the area or %NULL on failure
2368 */
2369void *vmap(struct page **pages, unsigned int count,
2370 unsigned long flags, pgprot_t prot)
2371{
2372 struct vm_struct *area;
2373 unsigned long size; /* In bytes */
2374
2375 might_sleep();
2376
2377 if (count > totalram_pages())
2378 return NULL;
2379
2380 size = (unsigned long)count << PAGE_SHIFT;
2381 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2382 if (!area)
2383 return NULL;
2384
2385 if (map_vm_area(area, prot, pages)) {
2386 vunmap(area->addr);
2387 return NULL;
2388 }
2389
2390 return area->addr;
2391}
2392EXPORT_SYMBOL(vmap);
2393
2394static void *__vmalloc_node(unsigned long size, unsigned long align,
2395 gfp_t gfp_mask, pgprot_t prot,
2396 int node, const void *caller);
2397static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2398 pgprot_t prot, int node)
2399{
2400 struct page **pages;
2401 unsigned int nr_pages, array_size, i;
2402 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2403 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2404 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2405 0 :
2406 __GFP_HIGHMEM;
2407
2408 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2409 array_size = (nr_pages * sizeof(struct page *));
2410
2411 /* Please note that the recursion is strictly bounded. */
2412 if (array_size > PAGE_SIZE) {
2413 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2414 PAGE_KERNEL, node, area->caller);
2415 } else {
2416 pages = kmalloc_node(array_size, nested_gfp, node);
2417 }
2418
2419 if (!pages) {
2420 remove_vm_area(area->addr);
2421 kfree(area);
2422 return NULL;
2423 }
2424
2425 area->pages = pages;
2426 area->nr_pages = nr_pages;
2427
2428 for (i = 0; i < area->nr_pages; i++) {
2429 struct page *page;
2430
2431 if (node == NUMA_NO_NODE)
2432 page = alloc_page(alloc_mask|highmem_mask);
2433 else
2434 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2435
2436 if (unlikely(!page)) {
2437 /* Successfully allocated i pages, free them in __vunmap() */
2438 area->nr_pages = i;
2439 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2440 goto fail;
2441 }
2442 area->pages[i] = page;
2443 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
2444 cond_resched();
2445 }
2446 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2447
2448 if (map_vm_area(area, prot, pages))
2449 goto fail;
2450 return area->addr;
2451
2452fail:
2453 warn_alloc(gfp_mask, NULL,
2454 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2455 (area->nr_pages*PAGE_SIZE), area->size);
2456 __vfree(area->addr);
2457 return NULL;
2458}
2459
2460/**
2461 * __vmalloc_node_range - allocate virtually contiguous memory
2462 * @size: allocation size
2463 * @align: desired alignment
2464 * @start: vm area range start
2465 * @end: vm area range end
2466 * @gfp_mask: flags for the page level allocator
2467 * @prot: protection mask for the allocated pages
2468 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2469 * @node: node to use for allocation or NUMA_NO_NODE
2470 * @caller: caller's return address
2471 *
2472 * Allocate enough pages to cover @size from the page level
2473 * allocator with @gfp_mask flags. Map them into contiguous
2474 * kernel virtual space, using a pagetable protection of @prot.
2475 *
2476 * Return: the address of the area or %NULL on failure
2477 */
2478void *__vmalloc_node_range(unsigned long size, unsigned long align,
2479 unsigned long start, unsigned long end, gfp_t gfp_mask,
2480 pgprot_t prot, unsigned long vm_flags, int node,
2481 const void *caller)
2482{
2483 struct vm_struct *area;
2484 void *addr;
2485 unsigned long real_size = size;
2486
2487 size = PAGE_ALIGN(size);
2488 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2489 goto fail;
2490
2491 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
2492 vm_flags, start, end, node, gfp_mask, caller);
2493 if (!area)
2494 goto fail;
2495
2496 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2497 if (!addr)
2498 return NULL;
2499
2500 /*
2501 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2502 * flag. It means that vm_struct is not fully initialized.
2503 * Now, it is fully initialized, so remove this flag here.
2504 */
2505 clear_vm_uninitialized_flag(area);
2506
2507 kmemleak_vmalloc(area, size, gfp_mask);
2508
2509 return addr;
2510
2511fail:
2512 warn_alloc(gfp_mask, NULL,
2513 "vmalloc: allocation failure: %lu bytes", real_size);
2514 return NULL;
2515}
2516
2517/*
2518 * This is only for performance analysis of vmalloc and stress purpose.
2519 * It is required by vmalloc test module, therefore do not use it other
2520 * than that.
2521 */
2522#ifdef CONFIG_TEST_VMALLOC_MODULE
2523EXPORT_SYMBOL_GPL(__vmalloc_node_range);
2524#endif
2525
2526/**
2527 * __vmalloc_node - allocate virtually contiguous memory
2528 * @size: allocation size
2529 * @align: desired alignment
2530 * @gfp_mask: flags for the page level allocator
2531 * @prot: protection mask for the allocated pages
2532 * @node: node to use for allocation or NUMA_NO_NODE
2533 * @caller: caller's return address
2534 *
2535 * Allocate enough pages to cover @size from the page level
2536 * allocator with @gfp_mask flags. Map them into contiguous
2537 * kernel virtual space, using a pagetable protection of @prot.
2538 *
2539 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2540 * and __GFP_NOFAIL are not supported
2541 *
2542 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2543 * with mm people.
2544 *
2545 * Return: pointer to the allocated memory or %NULL on error
2546 */
2547static void *__vmalloc_node(unsigned long size, unsigned long align,
2548 gfp_t gfp_mask, pgprot_t prot,
2549 int node, const void *caller)
2550{
2551 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2552 gfp_mask, prot, 0, node, caller);
2553}
2554
2555void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
2556{
2557 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
2558 __builtin_return_address(0));
2559}
2560EXPORT_SYMBOL(__vmalloc);
2561
2562static inline void *__vmalloc_node_flags(unsigned long size,
2563 int node, gfp_t flags)
2564{
2565 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
2566 node, __builtin_return_address(0));
2567}
2568
2569
2570void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
2571 void *caller)
2572{
2573 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
2574}
2575
2576/**
2577 * vmalloc - allocate virtually contiguous memory
2578 * @size: allocation size
2579 *
2580 * Allocate enough pages to cover @size from the page level
2581 * allocator and map them into contiguous kernel virtual space.
2582 *
2583 * For tight control over page level allocator and protection flags
2584 * use __vmalloc() instead.
2585 *
2586 * Return: pointer to the allocated memory or %NULL on error
2587 */
2588void *vmalloc(unsigned long size)
2589{
2590 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2591 GFP_KERNEL);
2592}
2593EXPORT_SYMBOL(vmalloc);
2594
2595/**
2596 * vzalloc - allocate virtually contiguous memory with zero fill
2597 * @size: allocation size
2598 *
2599 * Allocate enough pages to cover @size from the page level
2600 * allocator and map them into contiguous kernel virtual space.
2601 * The memory allocated is set to zero.
2602 *
2603 * For tight control over page level allocator and protection flags
2604 * use __vmalloc() instead.
2605 *
2606 * Return: pointer to the allocated memory or %NULL on error
2607 */
2608void *vzalloc(unsigned long size)
2609{
2610 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2611 GFP_KERNEL | __GFP_ZERO);
2612}
2613EXPORT_SYMBOL(vzalloc);
2614
2615/**
2616 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2617 * @size: allocation size
2618 *
2619 * The resulting memory area is zeroed so it can be mapped to userspace
2620 * without leaking data.
2621 *
2622 * Return: pointer to the allocated memory or %NULL on error
2623 */
2624void *vmalloc_user(unsigned long size)
2625{
2626 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2627 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2628 VM_USERMAP, NUMA_NO_NODE,
2629 __builtin_return_address(0));
2630}
2631EXPORT_SYMBOL(vmalloc_user);
2632
2633/**
2634 * vmalloc_node - allocate memory on a specific node
2635 * @size: allocation size
2636 * @node: numa node
2637 *
2638 * Allocate enough pages to cover @size from the page level
2639 * allocator and map them into contiguous kernel virtual space.
2640 *
2641 * For tight control over page level allocator and protection flags
2642 * use __vmalloc() instead.
2643 *
2644 * Return: pointer to the allocated memory or %NULL on error
2645 */
2646void *vmalloc_node(unsigned long size, int node)
2647{
2648 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
2649 node, __builtin_return_address(0));
2650}
2651EXPORT_SYMBOL(vmalloc_node);
2652
2653/**
2654 * vzalloc_node - allocate memory on a specific node with zero fill
2655 * @size: allocation size
2656 * @node: numa node
2657 *
2658 * Allocate enough pages to cover @size from the page level
2659 * allocator and map them into contiguous kernel virtual space.
2660 * The memory allocated is set to zero.
2661 *
2662 * For tight control over page level allocator and protection flags
2663 * use __vmalloc_node() instead.
2664 *
2665 * Return: pointer to the allocated memory or %NULL on error
2666 */
2667void *vzalloc_node(unsigned long size, int node)
2668{
2669 return __vmalloc_node_flags(size, node,
2670 GFP_KERNEL | __GFP_ZERO);
2671}
2672EXPORT_SYMBOL(vzalloc_node);
2673
2674/**
2675 * vmalloc_exec - allocate virtually contiguous, executable memory
2676 * @size: allocation size
2677 *
2678 * Kernel-internal function to allocate enough pages to cover @size
2679 * the page level allocator and map them into contiguous and
2680 * executable kernel virtual space.
2681 *
2682 * For tight control over page level allocator and protection flags
2683 * use __vmalloc() instead.
2684 *
2685 * Return: pointer to the allocated memory or %NULL on error
2686 */
2687void *vmalloc_exec(unsigned long size)
2688{
2689 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
2690 GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
2691 NUMA_NO_NODE, __builtin_return_address(0));
2692}
2693
2694#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2695#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2696#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2697#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2698#else
2699/*
2700 * 64b systems should always have either DMA or DMA32 zones. For others
2701 * GFP_DMA32 should do the right thing and use the normal zone.
2702 */
2703#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2704#endif
2705
2706/**
2707 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2708 * @size: allocation size
2709 *
2710 * Allocate enough 32bit PA addressable pages to cover @size from the
2711 * page level allocator and map them into contiguous kernel virtual space.
2712 *
2713 * Return: pointer to the allocated memory or %NULL on error
2714 */
2715void *vmalloc_32(unsigned long size)
2716{
2717 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
2718 NUMA_NO_NODE, __builtin_return_address(0));
2719}
2720EXPORT_SYMBOL(vmalloc_32);
2721
2722/**
2723 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2724 * @size: allocation size
2725 *
2726 * The resulting memory area is 32bit addressable and zeroed so it can be
2727 * mapped to userspace without leaking data.
2728 *
2729 * Return: pointer to the allocated memory or %NULL on error
2730 */
2731void *vmalloc_32_user(unsigned long size)
2732{
2733 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2734 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2735 VM_USERMAP, NUMA_NO_NODE,
2736 __builtin_return_address(0));
2737}
2738EXPORT_SYMBOL(vmalloc_32_user);
2739
2740/*
2741 * small helper routine , copy contents to buf from addr.
2742 * If the page is not present, fill zero.
2743 */
2744
2745static int aligned_vread(char *buf, char *addr, unsigned long count)
2746{
2747 struct page *p;
2748 int copied = 0;
2749
2750 while (count) {
2751 unsigned long offset, length;
2752
2753 offset = offset_in_page(addr);
2754 length = PAGE_SIZE - offset;
2755 if (length > count)
2756 length = count;
2757 p = vmalloc_to_page(addr);
2758 /*
2759 * To do safe access to this _mapped_ area, we need
2760 * lock. But adding lock here means that we need to add
2761 * overhead of vmalloc()/vfree() calles for this _debug_
2762 * interface, rarely used. Instead of that, we'll use
2763 * kmap() and get small overhead in this access function.
2764 */
2765 if (p) {
2766 /*
2767 * we can expect USER0 is not used (see vread/vwrite's
2768 * function description)
2769 */
2770 void *map = kmap_atomic(p);
2771 memcpy(buf, map + offset, length);
2772 kunmap_atomic(map);
2773 } else
2774 memset(buf, 0, length);
2775
2776 addr += length;
2777 buf += length;
2778 copied += length;
2779 count -= length;
2780 }
2781 return copied;
2782}
2783
2784static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2785{
2786 struct page *p;
2787 int copied = 0;
2788
2789 while (count) {
2790 unsigned long offset, length;
2791
2792 offset = offset_in_page(addr);
2793 length = PAGE_SIZE - offset;
2794 if (length > count)
2795 length = count;
2796 p = vmalloc_to_page(addr);
2797 /*
2798 * To do safe access to this _mapped_ area, we need
2799 * lock. But adding lock here means that we need to add
2800 * overhead of vmalloc()/vfree() calles for this _debug_
2801 * interface, rarely used. Instead of that, we'll use
2802 * kmap() and get small overhead in this access function.
2803 */
2804 if (p) {
2805 /*
2806 * we can expect USER0 is not used (see vread/vwrite's
2807 * function description)
2808 */
2809 void *map = kmap_atomic(p);
2810 memcpy(map + offset, buf, length);
2811 kunmap_atomic(map);
2812 }
2813 addr += length;
2814 buf += length;
2815 copied += length;
2816 count -= length;
2817 }
2818 return copied;
2819}
2820
2821/**
2822 * vread() - read vmalloc area in a safe way.
2823 * @buf: buffer for reading data
2824 * @addr: vm address.
2825 * @count: number of bytes to be read.
2826 *
2827 * This function checks that addr is a valid vmalloc'ed area, and
2828 * copy data from that area to a given buffer. If the given memory range
2829 * of [addr...addr+count) includes some valid address, data is copied to
2830 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2831 * IOREMAP area is treated as memory hole and no copy is done.
2832 *
2833 * If [addr...addr+count) doesn't includes any intersects with alive
2834 * vm_struct area, returns 0. @buf should be kernel's buffer.
2835 *
2836 * Note: In usual ops, vread() is never necessary because the caller
2837 * should know vmalloc() area is valid and can use memcpy().
2838 * This is for routines which have to access vmalloc area without
2839 * any information, as /dev/kmem.
2840 *
2841 * Return: number of bytes for which addr and buf should be increased
2842 * (same number as @count) or %0 if [addr...addr+count) doesn't
2843 * include any intersection with valid vmalloc area
2844 */
2845long vread(char *buf, char *addr, unsigned long count)
2846{
2847 struct vmap_area *va;
2848 struct vm_struct *vm;
2849 char *vaddr, *buf_start = buf;
2850 unsigned long buflen = count;
2851 unsigned long n;
2852
2853 /* Don't allow overflow */
2854 if ((unsigned long) addr + count < count)
2855 count = -(unsigned long) addr;
2856
2857 spin_lock(&vmap_area_lock);
2858 list_for_each_entry(va, &vmap_area_list, list) {
2859 if (!count)
2860 break;
2861
2862 if (!va->vm)
2863 continue;
2864
2865 vm = va->vm;
2866 vaddr = (char *) vm->addr;
2867 if (addr >= vaddr + get_vm_area_size(vm))
2868 continue;
2869 while (addr < vaddr) {
2870 if (count == 0)
2871 goto finished;
2872 *buf = '\0';
2873 buf++;
2874 addr++;
2875 count--;
2876 }
2877 n = vaddr + get_vm_area_size(vm) - addr;
2878 if (n > count)
2879 n = count;
2880 if (!(vm->flags & VM_IOREMAP))
2881 aligned_vread(buf, addr, n);
2882 else /* IOREMAP area is treated as memory hole */
2883 memset(buf, 0, n);
2884 buf += n;
2885 addr += n;
2886 count -= n;
2887 }
2888finished:
2889 spin_unlock(&vmap_area_lock);
2890
2891 if (buf == buf_start)
2892 return 0;
2893 /* zero-fill memory holes */
2894 if (buf != buf_start + buflen)
2895 memset(buf, 0, buflen - (buf - buf_start));
2896
2897 return buflen;
2898}
2899
2900/**
2901 * vwrite() - write vmalloc area in a safe way.
2902 * @buf: buffer for source data
2903 * @addr: vm address.
2904 * @count: number of bytes to be read.
2905 *
2906 * This function checks that addr is a valid vmalloc'ed area, and
2907 * copy data from a buffer to the given addr. If specified range of
2908 * [addr...addr+count) includes some valid address, data is copied from
2909 * proper area of @buf. If there are memory holes, no copy to hole.
2910 * IOREMAP area is treated as memory hole and no copy is done.
2911 *
2912 * If [addr...addr+count) doesn't includes any intersects with alive
2913 * vm_struct area, returns 0. @buf should be kernel's buffer.
2914 *
2915 * Note: In usual ops, vwrite() is never necessary because the caller
2916 * should know vmalloc() area is valid and can use memcpy().
2917 * This is for routines which have to access vmalloc area without
2918 * any information, as /dev/kmem.
2919 *
2920 * Return: number of bytes for which addr and buf should be
2921 * increased (same number as @count) or %0 if [addr...addr+count)
2922 * doesn't include any intersection with valid vmalloc area
2923 */
2924long vwrite(char *buf, char *addr, unsigned long count)
2925{
2926 struct vmap_area *va;
2927 struct vm_struct *vm;
2928 char *vaddr;
2929 unsigned long n, buflen;
2930 int copied = 0;
2931
2932 /* Don't allow overflow */
2933 if ((unsigned long) addr + count < count)
2934 count = -(unsigned long) addr;
2935 buflen = count;
2936
2937 spin_lock(&vmap_area_lock);
2938 list_for_each_entry(va, &vmap_area_list, list) {
2939 if (!count)
2940 break;
2941
2942 if (!va->vm)
2943 continue;
2944
2945 vm = va->vm;
2946 vaddr = (char *) vm->addr;
2947 if (addr >= vaddr + get_vm_area_size(vm))
2948 continue;
2949 while (addr < vaddr) {
2950 if (count == 0)
2951 goto finished;
2952 buf++;
2953 addr++;
2954 count--;
2955 }
2956 n = vaddr + get_vm_area_size(vm) - addr;
2957 if (n > count)
2958 n = count;
2959 if (!(vm->flags & VM_IOREMAP)) {
2960 aligned_vwrite(buf, addr, n);
2961 copied++;
2962 }
2963 buf += n;
2964 addr += n;
2965 count -= n;
2966 }
2967finished:
2968 spin_unlock(&vmap_area_lock);
2969 if (!copied)
2970 return 0;
2971 return buflen;
2972}
2973
2974/**
2975 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2976 * @vma: vma to cover
2977 * @uaddr: target user address to start at
2978 * @kaddr: virtual address of vmalloc kernel memory
2979 * @size: size of map area
2980 *
2981 * Returns: 0 for success, -Exxx on failure
2982 *
2983 * This function checks that @kaddr is a valid vmalloc'ed area,
2984 * and that it is big enough to cover the range starting at
2985 * @uaddr in @vma. Will return failure if that criteria isn't
2986 * met.
2987 *
2988 * Similar to remap_pfn_range() (see mm/memory.c)
2989 */
2990int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2991 void *kaddr, unsigned long size)
2992{
2993 struct vm_struct *area;
2994
2995 size = PAGE_ALIGN(size);
2996
2997 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2998 return -EINVAL;
2999
3000 area = find_vm_area(kaddr);
3001 if (!area)
3002 return -EINVAL;
3003
3004 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3005 return -EINVAL;
3006
3007 if (kaddr + size > area->addr + get_vm_area_size(area))
3008 return -EINVAL;
3009
3010 do {
3011 struct page *page = vmalloc_to_page(kaddr);
3012 int ret;
3013
3014 ret = vm_insert_page(vma, uaddr, page);
3015 if (ret)
3016 return ret;
3017
3018 uaddr += PAGE_SIZE;
3019 kaddr += PAGE_SIZE;
3020 size -= PAGE_SIZE;
3021 } while (size > 0);
3022
3023 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3024
3025 return 0;
3026}
3027EXPORT_SYMBOL(remap_vmalloc_range_partial);
3028
3029/**
3030 * remap_vmalloc_range - map vmalloc pages to userspace
3031 * @vma: vma to cover (map full range of vma)
3032 * @addr: vmalloc memory
3033 * @pgoff: number of pages into addr before first page to map
3034 *
3035 * Returns: 0 for success, -Exxx on failure
3036 *
3037 * This function checks that addr is a valid vmalloc'ed area, and
3038 * that it is big enough to cover the vma. Will return failure if
3039 * that criteria isn't met.
3040 *
3041 * Similar to remap_pfn_range() (see mm/memory.c)
3042 */
3043int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3044 unsigned long pgoff)
3045{
3046 return remap_vmalloc_range_partial(vma, vma->vm_start,
3047 addr + (pgoff << PAGE_SHIFT),
3048 vma->vm_end - vma->vm_start);
3049}
3050EXPORT_SYMBOL(remap_vmalloc_range);
3051
3052/*
3053 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3054 * have one.
3055 *
3056 * The purpose of this function is to make sure the vmalloc area
3057 * mappings are identical in all page-tables in the system.
3058 */
3059void __weak vmalloc_sync_all(void)
3060{
3061}
3062
3063
3064static int f(pte_t *pte, unsigned long addr, void *data)
3065{
3066 pte_t ***p = data;
3067
3068 if (p) {
3069 *(*p) = pte;
3070 (*p)++;
3071 }
3072 return 0;
3073}
3074
3075/**
3076 * alloc_vm_area - allocate a range of kernel address space
3077 * @size: size of the area
3078 * @ptes: returns the PTEs for the address space
3079 *
3080 * Returns: NULL on failure, vm_struct on success
3081 *
3082 * This function reserves a range of kernel address space, and
3083 * allocates pagetables to map that range. No actual mappings
3084 * are created.
3085 *
3086 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3087 * allocated for the VM area are returned.
3088 */
3089struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3090{
3091 struct vm_struct *area;
3092
3093 area = get_vm_area_caller(size, VM_IOREMAP,
3094 __builtin_return_address(0));
3095 if (area == NULL)
3096 return NULL;
3097
3098 /*
3099 * This ensures that page tables are constructed for this region
3100 * of kernel virtual address space and mapped into init_mm.
3101 */
3102 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3103 size, f, ptes ? &ptes : NULL)) {
3104 free_vm_area(area);
3105 return NULL;
3106 }
3107
3108 return area;
3109}
3110EXPORT_SYMBOL_GPL(alloc_vm_area);
3111
3112void free_vm_area(struct vm_struct *area)
3113{
3114 struct vm_struct *ret;
3115 ret = remove_vm_area(area->addr);
3116 BUG_ON(ret != area);
3117 kfree(area);
3118}
3119EXPORT_SYMBOL_GPL(free_vm_area);
3120
3121#ifdef CONFIG_SMP
3122static struct vmap_area *node_to_va(struct rb_node *n)
3123{
3124 return rb_entry_safe(n, struct vmap_area, rb_node);
3125}
3126
3127/**
3128 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3129 * @addr: target address
3130 *
3131 * Returns: vmap_area if it is found. If there is no such area
3132 * the first highest(reverse order) vmap_area is returned
3133 * i.e. va->va_start < addr && va->va_end < addr or NULL
3134 * if there are no any areas before @addr.
3135 */
3136static struct vmap_area *
3137pvm_find_va_enclose_addr(unsigned long addr)
3138{
3139 struct vmap_area *va, *tmp;
3140 struct rb_node *n;
3141
3142 n = free_vmap_area_root.rb_node;
3143 va = NULL;
3144
3145 while (n) {
3146 tmp = rb_entry(n, struct vmap_area, rb_node);
3147 if (tmp->va_start <= addr) {
3148 va = tmp;
3149 if (tmp->va_end >= addr)
3150 break;
3151
3152 n = n->rb_right;
3153 } else {
3154 n = n->rb_left;
3155 }
3156 }
3157
3158 return va;
3159}
3160
3161/**
3162 * pvm_determine_end_from_reverse - find the highest aligned address
3163 * of free block below VMALLOC_END
3164 * @va:
3165 * in - the VA we start the search(reverse order);
3166 * out - the VA with the highest aligned end address.
3167 *
3168 * Returns: determined end address within vmap_area
3169 */
3170static unsigned long
3171pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3172{
3173 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3174 unsigned long addr;
3175
3176 if (likely(*va)) {
3177 list_for_each_entry_from_reverse((*va),
3178 &free_vmap_area_list, list) {
3179 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3180 if ((*va)->va_start < addr)
3181 return addr;
3182 }
3183 }
3184
3185 return 0;
3186}
3187
3188/**
3189 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3190 * @offsets: array containing offset of each area
3191 * @sizes: array containing size of each area
3192 * @nr_vms: the number of areas to allocate
3193 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3194 *
3195 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3196 * vm_structs on success, %NULL on failure
3197 *
3198 * Percpu allocator wants to use congruent vm areas so that it can
3199 * maintain the offsets among percpu areas. This function allocates
3200 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3201 * be scattered pretty far, distance between two areas easily going up
3202 * to gigabytes. To avoid interacting with regular vmallocs, these
3203 * areas are allocated from top.
3204 *
3205 * Despite its complicated look, this allocator is rather simple. It
3206 * does everything top-down and scans free blocks from the end looking
3207 * for matching base. While scanning, if any of the areas do not fit the
3208 * base address is pulled down to fit the area. Scanning is repeated till
3209 * all the areas fit and then all necessary data structures are inserted
3210 * and the result is returned.
3211 */
3212struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3213 const size_t *sizes, int nr_vms,
3214 size_t align)
3215{
3216 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3217 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3218 struct vmap_area **vas, *va;
3219 struct vm_struct **vms;
3220 int area, area2, last_area, term_area;
3221 unsigned long base, start, size, end, last_end;
3222 bool purged = false;
3223 enum fit_type type;
3224
3225 /* verify parameters and allocate data structures */
3226 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3227 for (last_area = 0, area = 0; area < nr_vms; area++) {
3228 start = offsets[area];
3229 end = start + sizes[area];
3230
3231 /* is everything aligned properly? */
3232 BUG_ON(!IS_ALIGNED(offsets[area], align));
3233 BUG_ON(!IS_ALIGNED(sizes[area], align));
3234
3235 /* detect the area with the highest address */
3236 if (start > offsets[last_area])
3237 last_area = area;
3238
3239 for (area2 = area + 1; area2 < nr_vms; area2++) {
3240 unsigned long start2 = offsets[area2];
3241 unsigned long end2 = start2 + sizes[area2];
3242
3243 BUG_ON(start2 < end && start < end2);
3244 }
3245 }
3246 last_end = offsets[last_area] + sizes[last_area];
3247
3248 if (vmalloc_end - vmalloc_start < last_end) {
3249 WARN_ON(true);
3250 return NULL;
3251 }
3252
3253 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3254 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3255 if (!vas || !vms)
3256 goto err_free2;
3257
3258 for (area = 0; area < nr_vms; area++) {
3259 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3260 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3261 if (!vas[area] || !vms[area])
3262 goto err_free;
3263 }
3264retry:
3265 spin_lock(&vmap_area_lock);
3266
3267 /* start scanning - we scan from the top, begin with the last area */
3268 area = term_area = last_area;
3269 start = offsets[area];
3270 end = start + sizes[area];
3271
3272 va = pvm_find_va_enclose_addr(vmalloc_end);
3273 base = pvm_determine_end_from_reverse(&va, align) - end;
3274
3275 while (true) {
3276 /*
3277 * base might have underflowed, add last_end before
3278 * comparing.
3279 */
3280 if (base + last_end < vmalloc_start + last_end)
3281 goto overflow;
3282
3283 /*
3284 * Fitting base has not been found.
3285 */
3286 if (va == NULL)
3287 goto overflow;
3288
3289 /*
3290 * If required width exeeds current VA block, move
3291 * base downwards and then recheck.
3292 */
3293 if (base + end > va->va_end) {
3294 base = pvm_determine_end_from_reverse(&va, align) - end;
3295 term_area = area;
3296 continue;
3297 }
3298
3299 /*
3300 * If this VA does not fit, move base downwards and recheck.
3301 */
3302 if (base + start < va->va_start) {
3303 va = node_to_va(rb_prev(&va->rb_node));
3304 base = pvm_determine_end_from_reverse(&va, align) - end;
3305 term_area = area;
3306 continue;
3307 }
3308
3309 /*
3310 * This area fits, move on to the previous one. If
3311 * the previous one is the terminal one, we're done.
3312 */
3313 area = (area + nr_vms - 1) % nr_vms;
3314 if (area == term_area)
3315 break;
3316
3317 start = offsets[area];
3318 end = start + sizes[area];
3319 va = pvm_find_va_enclose_addr(base + end);
3320 }
3321
3322 /* we've found a fitting base, insert all va's */
3323 for (area = 0; area < nr_vms; area++) {
3324 int ret;
3325
3326 start = base + offsets[area];
3327 size = sizes[area];
3328
3329 va = pvm_find_va_enclose_addr(start);
3330 if (WARN_ON_ONCE(va == NULL))
3331 /* It is a BUG(), but trigger recovery instead. */
3332 goto recovery;
3333
3334 type = classify_va_fit_type(va, start, size);
3335 if (WARN_ON_ONCE(type == NOTHING_FIT))
3336 /* It is a BUG(), but trigger recovery instead. */
3337 goto recovery;
3338
3339 ret = adjust_va_to_fit_type(va, start, size, type);
3340 if (unlikely(ret))
3341 goto recovery;
3342
3343 /* Allocated area. */
3344 va = vas[area];
3345 va->va_start = start;
3346 va->va_end = start + size;
3347
3348 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
3349 }
3350
3351 spin_unlock(&vmap_area_lock);
3352
3353 /* insert all vm's */
3354 for (area = 0; area < nr_vms; area++)
3355 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
3356 pcpu_get_vm_areas);
3357
3358 kfree(vas);
3359 return vms;
3360
3361recovery:
3362 /* Remove previously inserted areas. */
3363 while (area--) {
3364 __free_vmap_area(vas[area]);
3365 vas[area] = NULL;
3366 }
3367
3368overflow:
3369 spin_unlock(&vmap_area_lock);
3370 if (!purged) {
3371 purge_vmap_area_lazy();
3372 purged = true;
3373
3374 /* Before "retry", check if we recover. */
3375 for (area = 0; area < nr_vms; area++) {
3376 if (vas[area])
3377 continue;
3378
3379 vas[area] = kmem_cache_zalloc(
3380 vmap_area_cachep, GFP_KERNEL);
3381 if (!vas[area])
3382 goto err_free;
3383 }
3384
3385 goto retry;
3386 }
3387
3388err_free:
3389 for (area = 0; area < nr_vms; area++) {
3390 if (vas[area])
3391 kmem_cache_free(vmap_area_cachep, vas[area]);
3392
3393 kfree(vms[area]);
3394 }
3395err_free2:
3396 kfree(vas);
3397 kfree(vms);
3398 return NULL;
3399}
3400
3401/**
3402 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3403 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3404 * @nr_vms: the number of allocated areas
3405 *
3406 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3407 */
3408void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3409{
3410 int i;
3411
3412 for (i = 0; i < nr_vms; i++)
3413 free_vm_area(vms[i]);
3414 kfree(vms);
3415}
3416#endif /* CONFIG_SMP */
3417
3418#ifdef CONFIG_PROC_FS
3419static void *s_start(struct seq_file *m, loff_t *pos)
3420 __acquires(&vmap_area_lock)
3421{
3422 spin_lock(&vmap_area_lock);
3423 return seq_list_start(&vmap_area_list, *pos);
3424}
3425
3426static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3427{
3428 return seq_list_next(p, &vmap_area_list, pos);
3429}
3430
3431static void s_stop(struct seq_file *m, void *p)
3432 __releases(&vmap_area_lock)
3433{
3434 spin_unlock(&vmap_area_lock);
3435}
3436
3437static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3438{
3439 if (IS_ENABLED(CONFIG_NUMA)) {
3440 unsigned int nr, *counters = m->private;
3441
3442 if (!counters)
3443 return;
3444
3445 if (v->flags & VM_UNINITIALIZED)
3446 return;
3447 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3448 smp_rmb();
3449
3450 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3451
3452 for (nr = 0; nr < v->nr_pages; nr++)
3453 counters[page_to_nid(v->pages[nr])]++;
3454
3455 for_each_node_state(nr, N_HIGH_MEMORY)
3456 if (counters[nr])
3457 seq_printf(m, " N%u=%u", nr, counters[nr]);
3458 }
3459}
3460
3461static void show_purge_info(struct seq_file *m)
3462{
3463 struct llist_node *head;
3464 struct vmap_area *va;
3465
3466 head = READ_ONCE(vmap_purge_list.first);
3467 if (head == NULL)
3468 return;
3469
3470 llist_for_each_entry(va, head, purge_list) {
3471 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3472 (void *)va->va_start, (void *)va->va_end,
3473 va->va_end - va->va_start);
3474 }
3475}
3476
3477static int s_show(struct seq_file *m, void *p)
3478{
3479 struct vmap_area *va;
3480 struct vm_struct *v;
3481
3482 va = list_entry(p, struct vmap_area, list);
3483
3484 /*
3485 * s_show can encounter race with remove_vm_area, !vm on behalf
3486 * of vmap area is being tear down or vm_map_ram allocation.
3487 */
3488 if (!va->vm) {
3489 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3490 (void *)va->va_start, (void *)va->va_end,
3491 va->va_end - va->va_start);
3492
3493 return 0;
3494 }
3495
3496 v = va->vm;
3497
3498 seq_printf(m, "0x%pK-0x%pK %7ld",
3499 v->addr, v->addr + v->size, v->size);
3500
3501 if (v->caller)
3502 seq_printf(m, " %pS", v->caller);
3503
3504 if (v->nr_pages)
3505 seq_printf(m, " pages=%d", v->nr_pages);
3506
3507 if (v->phys_addr)
3508 seq_printf(m, " phys=%pa", &v->phys_addr);
3509
3510 if (v->flags & VM_IOREMAP)
3511 seq_puts(m, " ioremap");
3512
3513 if (v->flags & VM_ALLOC)
3514 seq_puts(m, " vmalloc");
3515
3516 if (v->flags & VM_MAP)
3517 seq_puts(m, " vmap");
3518
3519 if (v->flags & VM_USERMAP)
3520 seq_puts(m, " user");
3521
3522 if (v->flags & VM_DMA_COHERENT)
3523 seq_puts(m, " dma-coherent");
3524
3525 if (is_vmalloc_addr(v->pages))
3526 seq_puts(m, " vpages");
3527
3528 show_numa_info(m, v);
3529 seq_putc(m, '\n');
3530
3531 /*
3532 * As a final step, dump "unpurged" areas. Note,
3533 * that entire "/proc/vmallocinfo" output will not
3534 * be address sorted, because the purge list is not
3535 * sorted.
3536 */
3537 if (list_is_last(&va->list, &vmap_area_list))
3538 show_purge_info(m);
3539
3540 return 0;
3541}
3542
3543static const struct seq_operations vmalloc_op = {
3544 .start = s_start,
3545 .next = s_next,
3546 .stop = s_stop,
3547 .show = s_show,
3548};
3549
3550static int __init proc_vmalloc_init(void)
3551{
3552 if (IS_ENABLED(CONFIG_NUMA))
3553 proc_create_seq_private("vmallocinfo", 0400, NULL,
3554 &vmalloc_op,
3555 nr_node_ids * sizeof(unsigned int), NULL);
3556 else
3557 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3558 return 0;
3559}
3560module_init(proc_vmalloc_init);
3561
3562#endif
1/*
2 * linux/mm/vmalloc.c
3 *
4 * Copyright (C) 1993 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8 * Numa awareness, Christoph Lameter, SGI, June 2005
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/debugobjects.h>
22#include <linux/kallsyms.h>
23#include <linux/list.h>
24#include <linux/rbtree.h>
25#include <linux/radix-tree.h>
26#include <linux/rcupdate.h>
27#include <linux/pfn.h>
28#include <linux/kmemleak.h>
29#include <linux/atomic.h>
30#include <linux/compiler.h>
31#include <linux/llist.h>
32#include <linux/bitops.h>
33
34#include <asm/uaccess.h>
35#include <asm/tlbflush.h>
36#include <asm/shmparam.h>
37
38#include "internal.h"
39
40struct vfree_deferred {
41 struct llist_head list;
42 struct work_struct wq;
43};
44static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
45
46static void __vunmap(const void *, int);
47
48static void free_work(struct work_struct *w)
49{
50 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
51 struct llist_node *llnode = llist_del_all(&p->list);
52 while (llnode) {
53 void *p = llnode;
54 llnode = llist_next(llnode);
55 __vunmap(p, 1);
56 }
57}
58
59/*** Page table manipulation functions ***/
60
61static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
62{
63 pte_t *pte;
64
65 pte = pte_offset_kernel(pmd, addr);
66 do {
67 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
68 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
69 } while (pte++, addr += PAGE_SIZE, addr != end);
70}
71
72static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
73{
74 pmd_t *pmd;
75 unsigned long next;
76
77 pmd = pmd_offset(pud, addr);
78 do {
79 next = pmd_addr_end(addr, end);
80 if (pmd_clear_huge(pmd))
81 continue;
82 if (pmd_none_or_clear_bad(pmd))
83 continue;
84 vunmap_pte_range(pmd, addr, next);
85 } while (pmd++, addr = next, addr != end);
86}
87
88static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
89{
90 pud_t *pud;
91 unsigned long next;
92
93 pud = pud_offset(pgd, addr);
94 do {
95 next = pud_addr_end(addr, end);
96 if (pud_clear_huge(pud))
97 continue;
98 if (pud_none_or_clear_bad(pud))
99 continue;
100 vunmap_pmd_range(pud, addr, next);
101 } while (pud++, addr = next, addr != end);
102}
103
104static void vunmap_page_range(unsigned long addr, unsigned long end)
105{
106 pgd_t *pgd;
107 unsigned long next;
108
109 BUG_ON(addr >= end);
110 pgd = pgd_offset_k(addr);
111 do {
112 next = pgd_addr_end(addr, end);
113 if (pgd_none_or_clear_bad(pgd))
114 continue;
115 vunmap_pud_range(pgd, addr, next);
116 } while (pgd++, addr = next, addr != end);
117}
118
119static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
120 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
121{
122 pte_t *pte;
123
124 /*
125 * nr is a running index into the array which helps higher level
126 * callers keep track of where we're up to.
127 */
128
129 pte = pte_alloc_kernel(pmd, addr);
130 if (!pte)
131 return -ENOMEM;
132 do {
133 struct page *page = pages[*nr];
134
135 if (WARN_ON(!pte_none(*pte)))
136 return -EBUSY;
137 if (WARN_ON(!page))
138 return -ENOMEM;
139 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
140 (*nr)++;
141 } while (pte++, addr += PAGE_SIZE, addr != end);
142 return 0;
143}
144
145static int vmap_pmd_range(pud_t *pud, unsigned long addr,
146 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
147{
148 pmd_t *pmd;
149 unsigned long next;
150
151 pmd = pmd_alloc(&init_mm, pud, addr);
152 if (!pmd)
153 return -ENOMEM;
154 do {
155 next = pmd_addr_end(addr, end);
156 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
157 return -ENOMEM;
158 } while (pmd++, addr = next, addr != end);
159 return 0;
160}
161
162static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
163 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
164{
165 pud_t *pud;
166 unsigned long next;
167
168 pud = pud_alloc(&init_mm, pgd, addr);
169 if (!pud)
170 return -ENOMEM;
171 do {
172 next = pud_addr_end(addr, end);
173 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
174 return -ENOMEM;
175 } while (pud++, addr = next, addr != end);
176 return 0;
177}
178
179/*
180 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
181 * will have pfns corresponding to the "pages" array.
182 *
183 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
184 */
185static int vmap_page_range_noflush(unsigned long start, unsigned long end,
186 pgprot_t prot, struct page **pages)
187{
188 pgd_t *pgd;
189 unsigned long next;
190 unsigned long addr = start;
191 int err = 0;
192 int nr = 0;
193
194 BUG_ON(addr >= end);
195 pgd = pgd_offset_k(addr);
196 do {
197 next = pgd_addr_end(addr, end);
198 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
199 if (err)
200 return err;
201 } while (pgd++, addr = next, addr != end);
202
203 return nr;
204}
205
206static int vmap_page_range(unsigned long start, unsigned long end,
207 pgprot_t prot, struct page **pages)
208{
209 int ret;
210
211 ret = vmap_page_range_noflush(start, end, prot, pages);
212 flush_cache_vmap(start, end);
213 return ret;
214}
215
216int is_vmalloc_or_module_addr(const void *x)
217{
218 /*
219 * ARM, x86-64 and sparc64 put modules in a special place,
220 * and fall back on vmalloc() if that fails. Others
221 * just put it in the vmalloc space.
222 */
223#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
224 unsigned long addr = (unsigned long)x;
225 if (addr >= MODULES_VADDR && addr < MODULES_END)
226 return 1;
227#endif
228 return is_vmalloc_addr(x);
229}
230
231/*
232 * Walk a vmap address to the struct page it maps.
233 */
234struct page *vmalloc_to_page(const void *vmalloc_addr)
235{
236 unsigned long addr = (unsigned long) vmalloc_addr;
237 struct page *page = NULL;
238 pgd_t *pgd = pgd_offset_k(addr);
239
240 /*
241 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
242 * architectures that do not vmalloc module space
243 */
244 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
245
246 if (!pgd_none(*pgd)) {
247 pud_t *pud = pud_offset(pgd, addr);
248 if (!pud_none(*pud)) {
249 pmd_t *pmd = pmd_offset(pud, addr);
250 if (!pmd_none(*pmd)) {
251 pte_t *ptep, pte;
252
253 ptep = pte_offset_map(pmd, addr);
254 pte = *ptep;
255 if (pte_present(pte))
256 page = pte_page(pte);
257 pte_unmap(ptep);
258 }
259 }
260 }
261 return page;
262}
263EXPORT_SYMBOL(vmalloc_to_page);
264
265/*
266 * Map a vmalloc()-space virtual address to the physical page frame number.
267 */
268unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
269{
270 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
271}
272EXPORT_SYMBOL(vmalloc_to_pfn);
273
274
275/*** Global kva allocator ***/
276
277#define VM_LAZY_FREE 0x01
278#define VM_LAZY_FREEING 0x02
279#define VM_VM_AREA 0x04
280
281static DEFINE_SPINLOCK(vmap_area_lock);
282/* Export for kexec only */
283LIST_HEAD(vmap_area_list);
284static struct rb_root vmap_area_root = RB_ROOT;
285
286/* The vmap cache globals are protected by vmap_area_lock */
287static struct rb_node *free_vmap_cache;
288static unsigned long cached_hole_size;
289static unsigned long cached_vstart;
290static unsigned long cached_align;
291
292static unsigned long vmap_area_pcpu_hole;
293
294static struct vmap_area *__find_vmap_area(unsigned long addr)
295{
296 struct rb_node *n = vmap_area_root.rb_node;
297
298 while (n) {
299 struct vmap_area *va;
300
301 va = rb_entry(n, struct vmap_area, rb_node);
302 if (addr < va->va_start)
303 n = n->rb_left;
304 else if (addr >= va->va_end)
305 n = n->rb_right;
306 else
307 return va;
308 }
309
310 return NULL;
311}
312
313static void __insert_vmap_area(struct vmap_area *va)
314{
315 struct rb_node **p = &vmap_area_root.rb_node;
316 struct rb_node *parent = NULL;
317 struct rb_node *tmp;
318
319 while (*p) {
320 struct vmap_area *tmp_va;
321
322 parent = *p;
323 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
324 if (va->va_start < tmp_va->va_end)
325 p = &(*p)->rb_left;
326 else if (va->va_end > tmp_va->va_start)
327 p = &(*p)->rb_right;
328 else
329 BUG();
330 }
331
332 rb_link_node(&va->rb_node, parent, p);
333 rb_insert_color(&va->rb_node, &vmap_area_root);
334
335 /* address-sort this list */
336 tmp = rb_prev(&va->rb_node);
337 if (tmp) {
338 struct vmap_area *prev;
339 prev = rb_entry(tmp, struct vmap_area, rb_node);
340 list_add_rcu(&va->list, &prev->list);
341 } else
342 list_add_rcu(&va->list, &vmap_area_list);
343}
344
345static void purge_vmap_area_lazy(void);
346
347/*
348 * Allocate a region of KVA of the specified size and alignment, within the
349 * vstart and vend.
350 */
351static struct vmap_area *alloc_vmap_area(unsigned long size,
352 unsigned long align,
353 unsigned long vstart, unsigned long vend,
354 int node, gfp_t gfp_mask)
355{
356 struct vmap_area *va;
357 struct rb_node *n;
358 unsigned long addr;
359 int purged = 0;
360 struct vmap_area *first;
361
362 BUG_ON(!size);
363 BUG_ON(offset_in_page(size));
364 BUG_ON(!is_power_of_2(align));
365
366 va = kmalloc_node(sizeof(struct vmap_area),
367 gfp_mask & GFP_RECLAIM_MASK, node);
368 if (unlikely(!va))
369 return ERR_PTR(-ENOMEM);
370
371 /*
372 * Only scan the relevant parts containing pointers to other objects
373 * to avoid false negatives.
374 */
375 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
376
377retry:
378 spin_lock(&vmap_area_lock);
379 /*
380 * Invalidate cache if we have more permissive parameters.
381 * cached_hole_size notes the largest hole noticed _below_
382 * the vmap_area cached in free_vmap_cache: if size fits
383 * into that hole, we want to scan from vstart to reuse
384 * the hole instead of allocating above free_vmap_cache.
385 * Note that __free_vmap_area may update free_vmap_cache
386 * without updating cached_hole_size or cached_align.
387 */
388 if (!free_vmap_cache ||
389 size < cached_hole_size ||
390 vstart < cached_vstart ||
391 align < cached_align) {
392nocache:
393 cached_hole_size = 0;
394 free_vmap_cache = NULL;
395 }
396 /* record if we encounter less permissive parameters */
397 cached_vstart = vstart;
398 cached_align = align;
399
400 /* find starting point for our search */
401 if (free_vmap_cache) {
402 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
403 addr = ALIGN(first->va_end, align);
404 if (addr < vstart)
405 goto nocache;
406 if (addr + size < addr)
407 goto overflow;
408
409 } else {
410 addr = ALIGN(vstart, align);
411 if (addr + size < addr)
412 goto overflow;
413
414 n = vmap_area_root.rb_node;
415 first = NULL;
416
417 while (n) {
418 struct vmap_area *tmp;
419 tmp = rb_entry(n, struct vmap_area, rb_node);
420 if (tmp->va_end >= addr) {
421 first = tmp;
422 if (tmp->va_start <= addr)
423 break;
424 n = n->rb_left;
425 } else
426 n = n->rb_right;
427 }
428
429 if (!first)
430 goto found;
431 }
432
433 /* from the starting point, walk areas until a suitable hole is found */
434 while (addr + size > first->va_start && addr + size <= vend) {
435 if (addr + cached_hole_size < first->va_start)
436 cached_hole_size = first->va_start - addr;
437 addr = ALIGN(first->va_end, align);
438 if (addr + size < addr)
439 goto overflow;
440
441 if (list_is_last(&first->list, &vmap_area_list))
442 goto found;
443
444 first = list_next_entry(first, list);
445 }
446
447found:
448 if (addr + size > vend)
449 goto overflow;
450
451 va->va_start = addr;
452 va->va_end = addr + size;
453 va->flags = 0;
454 __insert_vmap_area(va);
455 free_vmap_cache = &va->rb_node;
456 spin_unlock(&vmap_area_lock);
457
458 BUG_ON(!IS_ALIGNED(va->va_start, align));
459 BUG_ON(va->va_start < vstart);
460 BUG_ON(va->va_end > vend);
461
462 return va;
463
464overflow:
465 spin_unlock(&vmap_area_lock);
466 if (!purged) {
467 purge_vmap_area_lazy();
468 purged = 1;
469 goto retry;
470 }
471 if (printk_ratelimit())
472 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
473 size);
474 kfree(va);
475 return ERR_PTR(-EBUSY);
476}
477
478static void __free_vmap_area(struct vmap_area *va)
479{
480 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
481
482 if (free_vmap_cache) {
483 if (va->va_end < cached_vstart) {
484 free_vmap_cache = NULL;
485 } else {
486 struct vmap_area *cache;
487 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
488 if (va->va_start <= cache->va_start) {
489 free_vmap_cache = rb_prev(&va->rb_node);
490 /*
491 * We don't try to update cached_hole_size or
492 * cached_align, but it won't go very wrong.
493 */
494 }
495 }
496 }
497 rb_erase(&va->rb_node, &vmap_area_root);
498 RB_CLEAR_NODE(&va->rb_node);
499 list_del_rcu(&va->list);
500
501 /*
502 * Track the highest possible candidate for pcpu area
503 * allocation. Areas outside of vmalloc area can be returned
504 * here too, consider only end addresses which fall inside
505 * vmalloc area proper.
506 */
507 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
508 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
509
510 kfree_rcu(va, rcu_head);
511}
512
513/*
514 * Free a region of KVA allocated by alloc_vmap_area
515 */
516static void free_vmap_area(struct vmap_area *va)
517{
518 spin_lock(&vmap_area_lock);
519 __free_vmap_area(va);
520 spin_unlock(&vmap_area_lock);
521}
522
523/*
524 * Clear the pagetable entries of a given vmap_area
525 */
526static void unmap_vmap_area(struct vmap_area *va)
527{
528 vunmap_page_range(va->va_start, va->va_end);
529}
530
531static void vmap_debug_free_range(unsigned long start, unsigned long end)
532{
533 /*
534 * Unmap page tables and force a TLB flush immediately if pagealloc
535 * debugging is enabled. This catches use after free bugs similarly to
536 * those in linear kernel virtual address space after a page has been
537 * freed.
538 *
539 * All the lazy freeing logic is still retained, in order to minimise
540 * intrusiveness of this debugging feature.
541 *
542 * This is going to be *slow* (linear kernel virtual address debugging
543 * doesn't do a broadcast TLB flush so it is a lot faster).
544 */
545 if (debug_pagealloc_enabled()) {
546 vunmap_page_range(start, end);
547 flush_tlb_kernel_range(start, end);
548 }
549}
550
551/*
552 * lazy_max_pages is the maximum amount of virtual address space we gather up
553 * before attempting to purge with a TLB flush.
554 *
555 * There is a tradeoff here: a larger number will cover more kernel page tables
556 * and take slightly longer to purge, but it will linearly reduce the number of
557 * global TLB flushes that must be performed. It would seem natural to scale
558 * this number up linearly with the number of CPUs (because vmapping activity
559 * could also scale linearly with the number of CPUs), however it is likely
560 * that in practice, workloads might be constrained in other ways that mean
561 * vmap activity will not scale linearly with CPUs. Also, I want to be
562 * conservative and not introduce a big latency on huge systems, so go with
563 * a less aggressive log scale. It will still be an improvement over the old
564 * code, and it will be simple to change the scale factor if we find that it
565 * becomes a problem on bigger systems.
566 */
567static unsigned long lazy_max_pages(void)
568{
569 unsigned int log;
570
571 log = fls(num_online_cpus());
572
573 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
574}
575
576static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
577
578/* for per-CPU blocks */
579static void purge_fragmented_blocks_allcpus(void);
580
581/*
582 * called before a call to iounmap() if the caller wants vm_area_struct's
583 * immediately freed.
584 */
585void set_iounmap_nonlazy(void)
586{
587 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
588}
589
590/*
591 * Purges all lazily-freed vmap areas.
592 *
593 * If sync is 0 then don't purge if there is already a purge in progress.
594 * If force_flush is 1, then flush kernel TLBs between *start and *end even
595 * if we found no lazy vmap areas to unmap (callers can use this to optimise
596 * their own TLB flushing).
597 * Returns with *start = min(*start, lowest purged address)
598 * *end = max(*end, highest purged address)
599 */
600static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
601 int sync, int force_flush)
602{
603 static DEFINE_SPINLOCK(purge_lock);
604 LIST_HEAD(valist);
605 struct vmap_area *va;
606 struct vmap_area *n_va;
607 int nr = 0;
608
609 /*
610 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
611 * should not expect such behaviour. This just simplifies locking for
612 * the case that isn't actually used at the moment anyway.
613 */
614 if (!sync && !force_flush) {
615 if (!spin_trylock(&purge_lock))
616 return;
617 } else
618 spin_lock(&purge_lock);
619
620 if (sync)
621 purge_fragmented_blocks_allcpus();
622
623 rcu_read_lock();
624 list_for_each_entry_rcu(va, &vmap_area_list, list) {
625 if (va->flags & VM_LAZY_FREE) {
626 if (va->va_start < *start)
627 *start = va->va_start;
628 if (va->va_end > *end)
629 *end = va->va_end;
630 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
631 list_add_tail(&va->purge_list, &valist);
632 va->flags |= VM_LAZY_FREEING;
633 va->flags &= ~VM_LAZY_FREE;
634 }
635 }
636 rcu_read_unlock();
637
638 if (nr)
639 atomic_sub(nr, &vmap_lazy_nr);
640
641 if (nr || force_flush)
642 flush_tlb_kernel_range(*start, *end);
643
644 if (nr) {
645 spin_lock(&vmap_area_lock);
646 list_for_each_entry_safe(va, n_va, &valist, purge_list)
647 __free_vmap_area(va);
648 spin_unlock(&vmap_area_lock);
649 }
650 spin_unlock(&purge_lock);
651}
652
653/*
654 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
655 * is already purging.
656 */
657static void try_purge_vmap_area_lazy(void)
658{
659 unsigned long start = ULONG_MAX, end = 0;
660
661 __purge_vmap_area_lazy(&start, &end, 0, 0);
662}
663
664/*
665 * Kick off a purge of the outstanding lazy areas.
666 */
667static void purge_vmap_area_lazy(void)
668{
669 unsigned long start = ULONG_MAX, end = 0;
670
671 __purge_vmap_area_lazy(&start, &end, 1, 0);
672}
673
674/*
675 * Free a vmap area, caller ensuring that the area has been unmapped
676 * and flush_cache_vunmap had been called for the correct range
677 * previously.
678 */
679static void free_vmap_area_noflush(struct vmap_area *va)
680{
681 va->flags |= VM_LAZY_FREE;
682 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
683 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
684 try_purge_vmap_area_lazy();
685}
686
687/*
688 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
689 * called for the correct range previously.
690 */
691static void free_unmap_vmap_area_noflush(struct vmap_area *va)
692{
693 unmap_vmap_area(va);
694 free_vmap_area_noflush(va);
695}
696
697/*
698 * Free and unmap a vmap area
699 */
700static void free_unmap_vmap_area(struct vmap_area *va)
701{
702 flush_cache_vunmap(va->va_start, va->va_end);
703 free_unmap_vmap_area_noflush(va);
704}
705
706static struct vmap_area *find_vmap_area(unsigned long addr)
707{
708 struct vmap_area *va;
709
710 spin_lock(&vmap_area_lock);
711 va = __find_vmap_area(addr);
712 spin_unlock(&vmap_area_lock);
713
714 return va;
715}
716
717static void free_unmap_vmap_area_addr(unsigned long addr)
718{
719 struct vmap_area *va;
720
721 va = find_vmap_area(addr);
722 BUG_ON(!va);
723 free_unmap_vmap_area(va);
724}
725
726
727/*** Per cpu kva allocator ***/
728
729/*
730 * vmap space is limited especially on 32 bit architectures. Ensure there is
731 * room for at least 16 percpu vmap blocks per CPU.
732 */
733/*
734 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
735 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
736 * instead (we just need a rough idea)
737 */
738#if BITS_PER_LONG == 32
739#define VMALLOC_SPACE (128UL*1024*1024)
740#else
741#define VMALLOC_SPACE (128UL*1024*1024*1024)
742#endif
743
744#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
745#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
746#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
747#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
748#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
749#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
750#define VMAP_BBMAP_BITS \
751 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
752 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
753 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
754
755#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
756
757static bool vmap_initialized __read_mostly = false;
758
759struct vmap_block_queue {
760 spinlock_t lock;
761 struct list_head free;
762};
763
764struct vmap_block {
765 spinlock_t lock;
766 struct vmap_area *va;
767 unsigned long free, dirty;
768 unsigned long dirty_min, dirty_max; /*< dirty range */
769 struct list_head free_list;
770 struct rcu_head rcu_head;
771 struct list_head purge;
772};
773
774/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
775static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
776
777/*
778 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
779 * in the free path. Could get rid of this if we change the API to return a
780 * "cookie" from alloc, to be passed to free. But no big deal yet.
781 */
782static DEFINE_SPINLOCK(vmap_block_tree_lock);
783static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
784
785/*
786 * We should probably have a fallback mechanism to allocate virtual memory
787 * out of partially filled vmap blocks. However vmap block sizing should be
788 * fairly reasonable according to the vmalloc size, so it shouldn't be a
789 * big problem.
790 */
791
792static unsigned long addr_to_vb_idx(unsigned long addr)
793{
794 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
795 addr /= VMAP_BLOCK_SIZE;
796 return addr;
797}
798
799static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
800{
801 unsigned long addr;
802
803 addr = va_start + (pages_off << PAGE_SHIFT);
804 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
805 return (void *)addr;
806}
807
808/**
809 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
810 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
811 * @order: how many 2^order pages should be occupied in newly allocated block
812 * @gfp_mask: flags for the page level allocator
813 *
814 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
815 */
816static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
817{
818 struct vmap_block_queue *vbq;
819 struct vmap_block *vb;
820 struct vmap_area *va;
821 unsigned long vb_idx;
822 int node, err;
823 void *vaddr;
824
825 node = numa_node_id();
826
827 vb = kmalloc_node(sizeof(struct vmap_block),
828 gfp_mask & GFP_RECLAIM_MASK, node);
829 if (unlikely(!vb))
830 return ERR_PTR(-ENOMEM);
831
832 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
833 VMALLOC_START, VMALLOC_END,
834 node, gfp_mask);
835 if (IS_ERR(va)) {
836 kfree(vb);
837 return ERR_CAST(va);
838 }
839
840 err = radix_tree_preload(gfp_mask);
841 if (unlikely(err)) {
842 kfree(vb);
843 free_vmap_area(va);
844 return ERR_PTR(err);
845 }
846
847 vaddr = vmap_block_vaddr(va->va_start, 0);
848 spin_lock_init(&vb->lock);
849 vb->va = va;
850 /* At least something should be left free */
851 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
852 vb->free = VMAP_BBMAP_BITS - (1UL << order);
853 vb->dirty = 0;
854 vb->dirty_min = VMAP_BBMAP_BITS;
855 vb->dirty_max = 0;
856 INIT_LIST_HEAD(&vb->free_list);
857
858 vb_idx = addr_to_vb_idx(va->va_start);
859 spin_lock(&vmap_block_tree_lock);
860 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
861 spin_unlock(&vmap_block_tree_lock);
862 BUG_ON(err);
863 radix_tree_preload_end();
864
865 vbq = &get_cpu_var(vmap_block_queue);
866 spin_lock(&vbq->lock);
867 list_add_tail_rcu(&vb->free_list, &vbq->free);
868 spin_unlock(&vbq->lock);
869 put_cpu_var(vmap_block_queue);
870
871 return vaddr;
872}
873
874static void free_vmap_block(struct vmap_block *vb)
875{
876 struct vmap_block *tmp;
877 unsigned long vb_idx;
878
879 vb_idx = addr_to_vb_idx(vb->va->va_start);
880 spin_lock(&vmap_block_tree_lock);
881 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
882 spin_unlock(&vmap_block_tree_lock);
883 BUG_ON(tmp != vb);
884
885 free_vmap_area_noflush(vb->va);
886 kfree_rcu(vb, rcu_head);
887}
888
889static void purge_fragmented_blocks(int cpu)
890{
891 LIST_HEAD(purge);
892 struct vmap_block *vb;
893 struct vmap_block *n_vb;
894 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
895
896 rcu_read_lock();
897 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
898
899 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
900 continue;
901
902 spin_lock(&vb->lock);
903 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
904 vb->free = 0; /* prevent further allocs after releasing lock */
905 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
906 vb->dirty_min = 0;
907 vb->dirty_max = VMAP_BBMAP_BITS;
908 spin_lock(&vbq->lock);
909 list_del_rcu(&vb->free_list);
910 spin_unlock(&vbq->lock);
911 spin_unlock(&vb->lock);
912 list_add_tail(&vb->purge, &purge);
913 } else
914 spin_unlock(&vb->lock);
915 }
916 rcu_read_unlock();
917
918 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
919 list_del(&vb->purge);
920 free_vmap_block(vb);
921 }
922}
923
924static void purge_fragmented_blocks_allcpus(void)
925{
926 int cpu;
927
928 for_each_possible_cpu(cpu)
929 purge_fragmented_blocks(cpu);
930}
931
932static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
933{
934 struct vmap_block_queue *vbq;
935 struct vmap_block *vb;
936 void *vaddr = NULL;
937 unsigned int order;
938
939 BUG_ON(offset_in_page(size));
940 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
941 if (WARN_ON(size == 0)) {
942 /*
943 * Allocating 0 bytes isn't what caller wants since
944 * get_order(0) returns funny result. Just warn and terminate
945 * early.
946 */
947 return NULL;
948 }
949 order = get_order(size);
950
951 rcu_read_lock();
952 vbq = &get_cpu_var(vmap_block_queue);
953 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
954 unsigned long pages_off;
955
956 spin_lock(&vb->lock);
957 if (vb->free < (1UL << order)) {
958 spin_unlock(&vb->lock);
959 continue;
960 }
961
962 pages_off = VMAP_BBMAP_BITS - vb->free;
963 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
964 vb->free -= 1UL << order;
965 if (vb->free == 0) {
966 spin_lock(&vbq->lock);
967 list_del_rcu(&vb->free_list);
968 spin_unlock(&vbq->lock);
969 }
970
971 spin_unlock(&vb->lock);
972 break;
973 }
974
975 put_cpu_var(vmap_block_queue);
976 rcu_read_unlock();
977
978 /* Allocate new block if nothing was found */
979 if (!vaddr)
980 vaddr = new_vmap_block(order, gfp_mask);
981
982 return vaddr;
983}
984
985static void vb_free(const void *addr, unsigned long size)
986{
987 unsigned long offset;
988 unsigned long vb_idx;
989 unsigned int order;
990 struct vmap_block *vb;
991
992 BUG_ON(offset_in_page(size));
993 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
994
995 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
996
997 order = get_order(size);
998
999 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1000 offset >>= PAGE_SHIFT;
1001
1002 vb_idx = addr_to_vb_idx((unsigned long)addr);
1003 rcu_read_lock();
1004 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1005 rcu_read_unlock();
1006 BUG_ON(!vb);
1007
1008 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1009
1010 spin_lock(&vb->lock);
1011
1012 /* Expand dirty range */
1013 vb->dirty_min = min(vb->dirty_min, offset);
1014 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1015
1016 vb->dirty += 1UL << order;
1017 if (vb->dirty == VMAP_BBMAP_BITS) {
1018 BUG_ON(vb->free);
1019 spin_unlock(&vb->lock);
1020 free_vmap_block(vb);
1021 } else
1022 spin_unlock(&vb->lock);
1023}
1024
1025/**
1026 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1027 *
1028 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1029 * to amortize TLB flushing overheads. What this means is that any page you
1030 * have now, may, in a former life, have been mapped into kernel virtual
1031 * address by the vmap layer and so there might be some CPUs with TLB entries
1032 * still referencing that page (additional to the regular 1:1 kernel mapping).
1033 *
1034 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1035 * be sure that none of the pages we have control over will have any aliases
1036 * from the vmap layer.
1037 */
1038void vm_unmap_aliases(void)
1039{
1040 unsigned long start = ULONG_MAX, end = 0;
1041 int cpu;
1042 int flush = 0;
1043
1044 if (unlikely(!vmap_initialized))
1045 return;
1046
1047 for_each_possible_cpu(cpu) {
1048 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1049 struct vmap_block *vb;
1050
1051 rcu_read_lock();
1052 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1053 spin_lock(&vb->lock);
1054 if (vb->dirty) {
1055 unsigned long va_start = vb->va->va_start;
1056 unsigned long s, e;
1057
1058 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1059 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1060
1061 start = min(s, start);
1062 end = max(e, end);
1063
1064 flush = 1;
1065 }
1066 spin_unlock(&vb->lock);
1067 }
1068 rcu_read_unlock();
1069 }
1070
1071 __purge_vmap_area_lazy(&start, &end, 1, flush);
1072}
1073EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1074
1075/**
1076 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1077 * @mem: the pointer returned by vm_map_ram
1078 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1079 */
1080void vm_unmap_ram(const void *mem, unsigned int count)
1081{
1082 unsigned long size = count << PAGE_SHIFT;
1083 unsigned long addr = (unsigned long)mem;
1084
1085 BUG_ON(!addr);
1086 BUG_ON(addr < VMALLOC_START);
1087 BUG_ON(addr > VMALLOC_END);
1088 BUG_ON(!PAGE_ALIGNED(addr));
1089
1090 debug_check_no_locks_freed(mem, size);
1091 vmap_debug_free_range(addr, addr+size);
1092
1093 if (likely(count <= VMAP_MAX_ALLOC))
1094 vb_free(mem, size);
1095 else
1096 free_unmap_vmap_area_addr(addr);
1097}
1098EXPORT_SYMBOL(vm_unmap_ram);
1099
1100/**
1101 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1102 * @pages: an array of pointers to the pages to be mapped
1103 * @count: number of pages
1104 * @node: prefer to allocate data structures on this node
1105 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1106 *
1107 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1108 * faster than vmap so it's good. But if you mix long-life and short-life
1109 * objects with vm_map_ram(), it could consume lots of address space through
1110 * fragmentation (especially on a 32bit machine). You could see failures in
1111 * the end. Please use this function for short-lived objects.
1112 *
1113 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1114 */
1115void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1116{
1117 unsigned long size = count << PAGE_SHIFT;
1118 unsigned long addr;
1119 void *mem;
1120
1121 if (likely(count <= VMAP_MAX_ALLOC)) {
1122 mem = vb_alloc(size, GFP_KERNEL);
1123 if (IS_ERR(mem))
1124 return NULL;
1125 addr = (unsigned long)mem;
1126 } else {
1127 struct vmap_area *va;
1128 va = alloc_vmap_area(size, PAGE_SIZE,
1129 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1130 if (IS_ERR(va))
1131 return NULL;
1132
1133 addr = va->va_start;
1134 mem = (void *)addr;
1135 }
1136 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1137 vm_unmap_ram(mem, count);
1138 return NULL;
1139 }
1140 return mem;
1141}
1142EXPORT_SYMBOL(vm_map_ram);
1143
1144static struct vm_struct *vmlist __initdata;
1145/**
1146 * vm_area_add_early - add vmap area early during boot
1147 * @vm: vm_struct to add
1148 *
1149 * This function is used to add fixed kernel vm area to vmlist before
1150 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1151 * should contain proper values and the other fields should be zero.
1152 *
1153 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1154 */
1155void __init vm_area_add_early(struct vm_struct *vm)
1156{
1157 struct vm_struct *tmp, **p;
1158
1159 BUG_ON(vmap_initialized);
1160 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1161 if (tmp->addr >= vm->addr) {
1162 BUG_ON(tmp->addr < vm->addr + vm->size);
1163 break;
1164 } else
1165 BUG_ON(tmp->addr + tmp->size > vm->addr);
1166 }
1167 vm->next = *p;
1168 *p = vm;
1169}
1170
1171/**
1172 * vm_area_register_early - register vmap area early during boot
1173 * @vm: vm_struct to register
1174 * @align: requested alignment
1175 *
1176 * This function is used to register kernel vm area before
1177 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1178 * proper values on entry and other fields should be zero. On return,
1179 * vm->addr contains the allocated address.
1180 *
1181 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1182 */
1183void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1184{
1185 static size_t vm_init_off __initdata;
1186 unsigned long addr;
1187
1188 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1189 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1190
1191 vm->addr = (void *)addr;
1192
1193 vm_area_add_early(vm);
1194}
1195
1196void __init vmalloc_init(void)
1197{
1198 struct vmap_area *va;
1199 struct vm_struct *tmp;
1200 int i;
1201
1202 for_each_possible_cpu(i) {
1203 struct vmap_block_queue *vbq;
1204 struct vfree_deferred *p;
1205
1206 vbq = &per_cpu(vmap_block_queue, i);
1207 spin_lock_init(&vbq->lock);
1208 INIT_LIST_HEAD(&vbq->free);
1209 p = &per_cpu(vfree_deferred, i);
1210 init_llist_head(&p->list);
1211 INIT_WORK(&p->wq, free_work);
1212 }
1213
1214 /* Import existing vmlist entries. */
1215 for (tmp = vmlist; tmp; tmp = tmp->next) {
1216 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1217 va->flags = VM_VM_AREA;
1218 va->va_start = (unsigned long)tmp->addr;
1219 va->va_end = va->va_start + tmp->size;
1220 va->vm = tmp;
1221 __insert_vmap_area(va);
1222 }
1223
1224 vmap_area_pcpu_hole = VMALLOC_END;
1225
1226 vmap_initialized = true;
1227}
1228
1229/**
1230 * map_kernel_range_noflush - map kernel VM area with the specified pages
1231 * @addr: start of the VM area to map
1232 * @size: size of the VM area to map
1233 * @prot: page protection flags to use
1234 * @pages: pages to map
1235 *
1236 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1237 * specify should have been allocated using get_vm_area() and its
1238 * friends.
1239 *
1240 * NOTE:
1241 * This function does NOT do any cache flushing. The caller is
1242 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1243 * before calling this function.
1244 *
1245 * RETURNS:
1246 * The number of pages mapped on success, -errno on failure.
1247 */
1248int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1249 pgprot_t prot, struct page **pages)
1250{
1251 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1252}
1253
1254/**
1255 * unmap_kernel_range_noflush - unmap kernel VM area
1256 * @addr: start of the VM area to unmap
1257 * @size: size of the VM area to unmap
1258 *
1259 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1260 * specify should have been allocated using get_vm_area() and its
1261 * friends.
1262 *
1263 * NOTE:
1264 * This function does NOT do any cache flushing. The caller is
1265 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1266 * before calling this function and flush_tlb_kernel_range() after.
1267 */
1268void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1269{
1270 vunmap_page_range(addr, addr + size);
1271}
1272EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1273
1274/**
1275 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1276 * @addr: start of the VM area to unmap
1277 * @size: size of the VM area to unmap
1278 *
1279 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1280 * the unmapping and tlb after.
1281 */
1282void unmap_kernel_range(unsigned long addr, unsigned long size)
1283{
1284 unsigned long end = addr + size;
1285
1286 flush_cache_vunmap(addr, end);
1287 vunmap_page_range(addr, end);
1288 flush_tlb_kernel_range(addr, end);
1289}
1290EXPORT_SYMBOL_GPL(unmap_kernel_range);
1291
1292int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1293{
1294 unsigned long addr = (unsigned long)area->addr;
1295 unsigned long end = addr + get_vm_area_size(area);
1296 int err;
1297
1298 err = vmap_page_range(addr, end, prot, pages);
1299
1300 return err > 0 ? 0 : err;
1301}
1302EXPORT_SYMBOL_GPL(map_vm_area);
1303
1304static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1305 unsigned long flags, const void *caller)
1306{
1307 spin_lock(&vmap_area_lock);
1308 vm->flags = flags;
1309 vm->addr = (void *)va->va_start;
1310 vm->size = va->va_end - va->va_start;
1311 vm->caller = caller;
1312 va->vm = vm;
1313 va->flags |= VM_VM_AREA;
1314 spin_unlock(&vmap_area_lock);
1315}
1316
1317static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1318{
1319 /*
1320 * Before removing VM_UNINITIALIZED,
1321 * we should make sure that vm has proper values.
1322 * Pair with smp_rmb() in show_numa_info().
1323 */
1324 smp_wmb();
1325 vm->flags &= ~VM_UNINITIALIZED;
1326}
1327
1328static struct vm_struct *__get_vm_area_node(unsigned long size,
1329 unsigned long align, unsigned long flags, unsigned long start,
1330 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1331{
1332 struct vmap_area *va;
1333 struct vm_struct *area;
1334
1335 BUG_ON(in_interrupt());
1336 if (flags & VM_IOREMAP)
1337 align = 1ul << clamp_t(int, fls_long(size),
1338 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1339
1340 size = PAGE_ALIGN(size);
1341 if (unlikely(!size))
1342 return NULL;
1343
1344 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1345 if (unlikely(!area))
1346 return NULL;
1347
1348 if (!(flags & VM_NO_GUARD))
1349 size += PAGE_SIZE;
1350
1351 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1352 if (IS_ERR(va)) {
1353 kfree(area);
1354 return NULL;
1355 }
1356
1357 setup_vmalloc_vm(area, va, flags, caller);
1358
1359 return area;
1360}
1361
1362struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1363 unsigned long start, unsigned long end)
1364{
1365 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1366 GFP_KERNEL, __builtin_return_address(0));
1367}
1368EXPORT_SYMBOL_GPL(__get_vm_area);
1369
1370struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1371 unsigned long start, unsigned long end,
1372 const void *caller)
1373{
1374 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1375 GFP_KERNEL, caller);
1376}
1377
1378/**
1379 * get_vm_area - reserve a contiguous kernel virtual area
1380 * @size: size of the area
1381 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1382 *
1383 * Search an area of @size in the kernel virtual mapping area,
1384 * and reserved it for out purposes. Returns the area descriptor
1385 * on success or %NULL on failure.
1386 */
1387struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1388{
1389 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1390 NUMA_NO_NODE, GFP_KERNEL,
1391 __builtin_return_address(0));
1392}
1393
1394struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1395 const void *caller)
1396{
1397 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1398 NUMA_NO_NODE, GFP_KERNEL, caller);
1399}
1400
1401/**
1402 * find_vm_area - find a continuous kernel virtual area
1403 * @addr: base address
1404 *
1405 * Search for the kernel VM area starting at @addr, and return it.
1406 * It is up to the caller to do all required locking to keep the returned
1407 * pointer valid.
1408 */
1409struct vm_struct *find_vm_area(const void *addr)
1410{
1411 struct vmap_area *va;
1412
1413 va = find_vmap_area((unsigned long)addr);
1414 if (va && va->flags & VM_VM_AREA)
1415 return va->vm;
1416
1417 return NULL;
1418}
1419
1420/**
1421 * remove_vm_area - find and remove a continuous kernel virtual area
1422 * @addr: base address
1423 *
1424 * Search for the kernel VM area starting at @addr, and remove it.
1425 * This function returns the found VM area, but using it is NOT safe
1426 * on SMP machines, except for its size or flags.
1427 */
1428struct vm_struct *remove_vm_area(const void *addr)
1429{
1430 struct vmap_area *va;
1431
1432 va = find_vmap_area((unsigned long)addr);
1433 if (va && va->flags & VM_VM_AREA) {
1434 struct vm_struct *vm = va->vm;
1435
1436 spin_lock(&vmap_area_lock);
1437 va->vm = NULL;
1438 va->flags &= ~VM_VM_AREA;
1439 spin_unlock(&vmap_area_lock);
1440
1441 vmap_debug_free_range(va->va_start, va->va_end);
1442 kasan_free_shadow(vm);
1443 free_unmap_vmap_area(va);
1444
1445 return vm;
1446 }
1447 return NULL;
1448}
1449
1450static void __vunmap(const void *addr, int deallocate_pages)
1451{
1452 struct vm_struct *area;
1453
1454 if (!addr)
1455 return;
1456
1457 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1458 addr))
1459 return;
1460
1461 area = remove_vm_area(addr);
1462 if (unlikely(!area)) {
1463 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1464 addr);
1465 return;
1466 }
1467
1468 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1469 debug_check_no_obj_freed(addr, get_vm_area_size(area));
1470
1471 if (deallocate_pages) {
1472 int i;
1473
1474 for (i = 0; i < area->nr_pages; i++) {
1475 struct page *page = area->pages[i];
1476
1477 BUG_ON(!page);
1478 __free_kmem_pages(page, 0);
1479 }
1480
1481 kvfree(area->pages);
1482 }
1483
1484 kfree(area);
1485 return;
1486}
1487
1488/**
1489 * vfree - release memory allocated by vmalloc()
1490 * @addr: memory base address
1491 *
1492 * Free the virtually continuous memory area starting at @addr, as
1493 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1494 * NULL, no operation is performed.
1495 *
1496 * Must not be called in NMI context (strictly speaking, only if we don't
1497 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1498 * conventions for vfree() arch-depenedent would be a really bad idea)
1499 *
1500 * NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1501 */
1502void vfree(const void *addr)
1503{
1504 BUG_ON(in_nmi());
1505
1506 kmemleak_free(addr);
1507
1508 if (!addr)
1509 return;
1510 if (unlikely(in_interrupt())) {
1511 struct vfree_deferred *p = this_cpu_ptr(&vfree_deferred);
1512 if (llist_add((struct llist_node *)addr, &p->list))
1513 schedule_work(&p->wq);
1514 } else
1515 __vunmap(addr, 1);
1516}
1517EXPORT_SYMBOL(vfree);
1518
1519/**
1520 * vunmap - release virtual mapping obtained by vmap()
1521 * @addr: memory base address
1522 *
1523 * Free the virtually contiguous memory area starting at @addr,
1524 * which was created from the page array passed to vmap().
1525 *
1526 * Must not be called in interrupt context.
1527 */
1528void vunmap(const void *addr)
1529{
1530 BUG_ON(in_interrupt());
1531 might_sleep();
1532 if (addr)
1533 __vunmap(addr, 0);
1534}
1535EXPORT_SYMBOL(vunmap);
1536
1537/**
1538 * vmap - map an array of pages into virtually contiguous space
1539 * @pages: array of page pointers
1540 * @count: number of pages to map
1541 * @flags: vm_area->flags
1542 * @prot: page protection for the mapping
1543 *
1544 * Maps @count pages from @pages into contiguous kernel virtual
1545 * space.
1546 */
1547void *vmap(struct page **pages, unsigned int count,
1548 unsigned long flags, pgprot_t prot)
1549{
1550 struct vm_struct *area;
1551
1552 might_sleep();
1553
1554 if (count > totalram_pages)
1555 return NULL;
1556
1557 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1558 __builtin_return_address(0));
1559 if (!area)
1560 return NULL;
1561
1562 if (map_vm_area(area, prot, pages)) {
1563 vunmap(area->addr);
1564 return NULL;
1565 }
1566
1567 return area->addr;
1568}
1569EXPORT_SYMBOL(vmap);
1570
1571static void *__vmalloc_node(unsigned long size, unsigned long align,
1572 gfp_t gfp_mask, pgprot_t prot,
1573 int node, const void *caller);
1574static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1575 pgprot_t prot, int node)
1576{
1577 const int order = 0;
1578 struct page **pages;
1579 unsigned int nr_pages, array_size, i;
1580 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1581 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1582
1583 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1584 array_size = (nr_pages * sizeof(struct page *));
1585
1586 area->nr_pages = nr_pages;
1587 /* Please note that the recursion is strictly bounded. */
1588 if (array_size > PAGE_SIZE) {
1589 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1590 PAGE_KERNEL, node, area->caller);
1591 } else {
1592 pages = kmalloc_node(array_size, nested_gfp, node);
1593 }
1594 area->pages = pages;
1595 if (!area->pages) {
1596 remove_vm_area(area->addr);
1597 kfree(area);
1598 return NULL;
1599 }
1600
1601 for (i = 0; i < area->nr_pages; i++) {
1602 struct page *page;
1603
1604 if (node == NUMA_NO_NODE)
1605 page = alloc_kmem_pages(alloc_mask, order);
1606 else
1607 page = alloc_kmem_pages_node(node, alloc_mask, order);
1608
1609 if (unlikely(!page)) {
1610 /* Successfully allocated i pages, free them in __vunmap() */
1611 area->nr_pages = i;
1612 goto fail;
1613 }
1614 area->pages[i] = page;
1615 if (gfpflags_allow_blocking(gfp_mask))
1616 cond_resched();
1617 }
1618
1619 if (map_vm_area(area, prot, pages))
1620 goto fail;
1621 return area->addr;
1622
1623fail:
1624 warn_alloc_failed(gfp_mask, order,
1625 "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1626 (area->nr_pages*PAGE_SIZE), area->size);
1627 vfree(area->addr);
1628 return NULL;
1629}
1630
1631/**
1632 * __vmalloc_node_range - allocate virtually contiguous memory
1633 * @size: allocation size
1634 * @align: desired alignment
1635 * @start: vm area range start
1636 * @end: vm area range end
1637 * @gfp_mask: flags for the page level allocator
1638 * @prot: protection mask for the allocated pages
1639 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1640 * @node: node to use for allocation or NUMA_NO_NODE
1641 * @caller: caller's return address
1642 *
1643 * Allocate enough pages to cover @size from the page level
1644 * allocator with @gfp_mask flags. Map them into contiguous
1645 * kernel virtual space, using a pagetable protection of @prot.
1646 */
1647void *__vmalloc_node_range(unsigned long size, unsigned long align,
1648 unsigned long start, unsigned long end, gfp_t gfp_mask,
1649 pgprot_t prot, unsigned long vm_flags, int node,
1650 const void *caller)
1651{
1652 struct vm_struct *area;
1653 void *addr;
1654 unsigned long real_size = size;
1655
1656 size = PAGE_ALIGN(size);
1657 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1658 goto fail;
1659
1660 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1661 vm_flags, start, end, node, gfp_mask, caller);
1662 if (!area)
1663 goto fail;
1664
1665 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1666 if (!addr)
1667 return NULL;
1668
1669 /*
1670 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1671 * flag. It means that vm_struct is not fully initialized.
1672 * Now, it is fully initialized, so remove this flag here.
1673 */
1674 clear_vm_uninitialized_flag(area);
1675
1676 /*
1677 * A ref_count = 2 is needed because vm_struct allocated in
1678 * __get_vm_area_node() contains a reference to the virtual address of
1679 * the vmalloc'ed block.
1680 */
1681 kmemleak_alloc(addr, real_size, 2, gfp_mask);
1682
1683 return addr;
1684
1685fail:
1686 warn_alloc_failed(gfp_mask, 0,
1687 "vmalloc: allocation failure: %lu bytes\n",
1688 real_size);
1689 return NULL;
1690}
1691
1692/**
1693 * __vmalloc_node - allocate virtually contiguous memory
1694 * @size: allocation size
1695 * @align: desired alignment
1696 * @gfp_mask: flags for the page level allocator
1697 * @prot: protection mask for the allocated pages
1698 * @node: node to use for allocation or NUMA_NO_NODE
1699 * @caller: caller's return address
1700 *
1701 * Allocate enough pages to cover @size from the page level
1702 * allocator with @gfp_mask flags. Map them into contiguous
1703 * kernel virtual space, using a pagetable protection of @prot.
1704 */
1705static void *__vmalloc_node(unsigned long size, unsigned long align,
1706 gfp_t gfp_mask, pgprot_t prot,
1707 int node, const void *caller)
1708{
1709 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1710 gfp_mask, prot, 0, node, caller);
1711}
1712
1713void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1714{
1715 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1716 __builtin_return_address(0));
1717}
1718EXPORT_SYMBOL(__vmalloc);
1719
1720static inline void *__vmalloc_node_flags(unsigned long size,
1721 int node, gfp_t flags)
1722{
1723 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1724 node, __builtin_return_address(0));
1725}
1726
1727/**
1728 * vmalloc - allocate virtually contiguous memory
1729 * @size: allocation size
1730 * Allocate enough pages to cover @size from the page level
1731 * allocator and map them into contiguous kernel virtual space.
1732 *
1733 * For tight control over page level allocator and protection flags
1734 * use __vmalloc() instead.
1735 */
1736void *vmalloc(unsigned long size)
1737{
1738 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1739 GFP_KERNEL | __GFP_HIGHMEM);
1740}
1741EXPORT_SYMBOL(vmalloc);
1742
1743/**
1744 * vzalloc - allocate virtually contiguous memory with zero fill
1745 * @size: allocation size
1746 * Allocate enough pages to cover @size from the page level
1747 * allocator and map them into contiguous kernel virtual space.
1748 * The memory allocated is set to zero.
1749 *
1750 * For tight control over page level allocator and protection flags
1751 * use __vmalloc() instead.
1752 */
1753void *vzalloc(unsigned long size)
1754{
1755 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1756 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1757}
1758EXPORT_SYMBOL(vzalloc);
1759
1760/**
1761 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1762 * @size: allocation size
1763 *
1764 * The resulting memory area is zeroed so it can be mapped to userspace
1765 * without leaking data.
1766 */
1767void *vmalloc_user(unsigned long size)
1768{
1769 struct vm_struct *area;
1770 void *ret;
1771
1772 ret = __vmalloc_node(size, SHMLBA,
1773 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1774 PAGE_KERNEL, NUMA_NO_NODE,
1775 __builtin_return_address(0));
1776 if (ret) {
1777 area = find_vm_area(ret);
1778 area->flags |= VM_USERMAP;
1779 }
1780 return ret;
1781}
1782EXPORT_SYMBOL(vmalloc_user);
1783
1784/**
1785 * vmalloc_node - allocate memory on a specific node
1786 * @size: allocation size
1787 * @node: numa node
1788 *
1789 * Allocate enough pages to cover @size from the page level
1790 * allocator and map them into contiguous kernel virtual space.
1791 *
1792 * For tight control over page level allocator and protection flags
1793 * use __vmalloc() instead.
1794 */
1795void *vmalloc_node(unsigned long size, int node)
1796{
1797 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1798 node, __builtin_return_address(0));
1799}
1800EXPORT_SYMBOL(vmalloc_node);
1801
1802/**
1803 * vzalloc_node - allocate memory on a specific node with zero fill
1804 * @size: allocation size
1805 * @node: numa node
1806 *
1807 * Allocate enough pages to cover @size from the page level
1808 * allocator and map them into contiguous kernel virtual space.
1809 * The memory allocated is set to zero.
1810 *
1811 * For tight control over page level allocator and protection flags
1812 * use __vmalloc_node() instead.
1813 */
1814void *vzalloc_node(unsigned long size, int node)
1815{
1816 return __vmalloc_node_flags(size, node,
1817 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1818}
1819EXPORT_SYMBOL(vzalloc_node);
1820
1821#ifndef PAGE_KERNEL_EXEC
1822# define PAGE_KERNEL_EXEC PAGE_KERNEL
1823#endif
1824
1825/**
1826 * vmalloc_exec - allocate virtually contiguous, executable memory
1827 * @size: allocation size
1828 *
1829 * Kernel-internal function to allocate enough pages to cover @size
1830 * the page level allocator and map them into contiguous and
1831 * executable kernel virtual space.
1832 *
1833 * For tight control over page level allocator and protection flags
1834 * use __vmalloc() instead.
1835 */
1836
1837void *vmalloc_exec(unsigned long size)
1838{
1839 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1840 NUMA_NO_NODE, __builtin_return_address(0));
1841}
1842
1843#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1844#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1845#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1846#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1847#else
1848#define GFP_VMALLOC32 GFP_KERNEL
1849#endif
1850
1851/**
1852 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1853 * @size: allocation size
1854 *
1855 * Allocate enough 32bit PA addressable pages to cover @size from the
1856 * page level allocator and map them into contiguous kernel virtual space.
1857 */
1858void *vmalloc_32(unsigned long size)
1859{
1860 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1861 NUMA_NO_NODE, __builtin_return_address(0));
1862}
1863EXPORT_SYMBOL(vmalloc_32);
1864
1865/**
1866 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1867 * @size: allocation size
1868 *
1869 * The resulting memory area is 32bit addressable and zeroed so it can be
1870 * mapped to userspace without leaking data.
1871 */
1872void *vmalloc_32_user(unsigned long size)
1873{
1874 struct vm_struct *area;
1875 void *ret;
1876
1877 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1878 NUMA_NO_NODE, __builtin_return_address(0));
1879 if (ret) {
1880 area = find_vm_area(ret);
1881 area->flags |= VM_USERMAP;
1882 }
1883 return ret;
1884}
1885EXPORT_SYMBOL(vmalloc_32_user);
1886
1887/*
1888 * small helper routine , copy contents to buf from addr.
1889 * If the page is not present, fill zero.
1890 */
1891
1892static int aligned_vread(char *buf, char *addr, unsigned long count)
1893{
1894 struct page *p;
1895 int copied = 0;
1896
1897 while (count) {
1898 unsigned long offset, length;
1899
1900 offset = offset_in_page(addr);
1901 length = PAGE_SIZE - offset;
1902 if (length > count)
1903 length = count;
1904 p = vmalloc_to_page(addr);
1905 /*
1906 * To do safe access to this _mapped_ area, we need
1907 * lock. But adding lock here means that we need to add
1908 * overhead of vmalloc()/vfree() calles for this _debug_
1909 * interface, rarely used. Instead of that, we'll use
1910 * kmap() and get small overhead in this access function.
1911 */
1912 if (p) {
1913 /*
1914 * we can expect USER0 is not used (see vread/vwrite's
1915 * function description)
1916 */
1917 void *map = kmap_atomic(p);
1918 memcpy(buf, map + offset, length);
1919 kunmap_atomic(map);
1920 } else
1921 memset(buf, 0, length);
1922
1923 addr += length;
1924 buf += length;
1925 copied += length;
1926 count -= length;
1927 }
1928 return copied;
1929}
1930
1931static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1932{
1933 struct page *p;
1934 int copied = 0;
1935
1936 while (count) {
1937 unsigned long offset, length;
1938
1939 offset = offset_in_page(addr);
1940 length = PAGE_SIZE - offset;
1941 if (length > count)
1942 length = count;
1943 p = vmalloc_to_page(addr);
1944 /*
1945 * To do safe access to this _mapped_ area, we need
1946 * lock. But adding lock here means that we need to add
1947 * overhead of vmalloc()/vfree() calles for this _debug_
1948 * interface, rarely used. Instead of that, we'll use
1949 * kmap() and get small overhead in this access function.
1950 */
1951 if (p) {
1952 /*
1953 * we can expect USER0 is not used (see vread/vwrite's
1954 * function description)
1955 */
1956 void *map = kmap_atomic(p);
1957 memcpy(map + offset, buf, length);
1958 kunmap_atomic(map);
1959 }
1960 addr += length;
1961 buf += length;
1962 copied += length;
1963 count -= length;
1964 }
1965 return copied;
1966}
1967
1968/**
1969 * vread() - read vmalloc area in a safe way.
1970 * @buf: buffer for reading data
1971 * @addr: vm address.
1972 * @count: number of bytes to be read.
1973 *
1974 * Returns # of bytes which addr and buf should be increased.
1975 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1976 * includes any intersect with alive vmalloc area.
1977 *
1978 * This function checks that addr is a valid vmalloc'ed area, and
1979 * copy data from that area to a given buffer. If the given memory range
1980 * of [addr...addr+count) includes some valid address, data is copied to
1981 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1982 * IOREMAP area is treated as memory hole and no copy is done.
1983 *
1984 * If [addr...addr+count) doesn't includes any intersects with alive
1985 * vm_struct area, returns 0. @buf should be kernel's buffer.
1986 *
1987 * Note: In usual ops, vread() is never necessary because the caller
1988 * should know vmalloc() area is valid and can use memcpy().
1989 * This is for routines which have to access vmalloc area without
1990 * any informaion, as /dev/kmem.
1991 *
1992 */
1993
1994long vread(char *buf, char *addr, unsigned long count)
1995{
1996 struct vmap_area *va;
1997 struct vm_struct *vm;
1998 char *vaddr, *buf_start = buf;
1999 unsigned long buflen = count;
2000 unsigned long n;
2001
2002 /* Don't allow overflow */
2003 if ((unsigned long) addr + count < count)
2004 count = -(unsigned long) addr;
2005
2006 spin_lock(&vmap_area_lock);
2007 list_for_each_entry(va, &vmap_area_list, list) {
2008 if (!count)
2009 break;
2010
2011 if (!(va->flags & VM_VM_AREA))
2012 continue;
2013
2014 vm = va->vm;
2015 vaddr = (char *) vm->addr;
2016 if (addr >= vaddr + get_vm_area_size(vm))
2017 continue;
2018 while (addr < vaddr) {
2019 if (count == 0)
2020 goto finished;
2021 *buf = '\0';
2022 buf++;
2023 addr++;
2024 count--;
2025 }
2026 n = vaddr + get_vm_area_size(vm) - addr;
2027 if (n > count)
2028 n = count;
2029 if (!(vm->flags & VM_IOREMAP))
2030 aligned_vread(buf, addr, n);
2031 else /* IOREMAP area is treated as memory hole */
2032 memset(buf, 0, n);
2033 buf += n;
2034 addr += n;
2035 count -= n;
2036 }
2037finished:
2038 spin_unlock(&vmap_area_lock);
2039
2040 if (buf == buf_start)
2041 return 0;
2042 /* zero-fill memory holes */
2043 if (buf != buf_start + buflen)
2044 memset(buf, 0, buflen - (buf - buf_start));
2045
2046 return buflen;
2047}
2048
2049/**
2050 * vwrite() - write vmalloc area in a safe way.
2051 * @buf: buffer for source data
2052 * @addr: vm address.
2053 * @count: number of bytes to be read.
2054 *
2055 * Returns # of bytes which addr and buf should be incresed.
2056 * (same number to @count).
2057 * If [addr...addr+count) doesn't includes any intersect with valid
2058 * vmalloc area, returns 0.
2059 *
2060 * This function checks that addr is a valid vmalloc'ed area, and
2061 * copy data from a buffer to the given addr. If specified range of
2062 * [addr...addr+count) includes some valid address, data is copied from
2063 * proper area of @buf. If there are memory holes, no copy to hole.
2064 * IOREMAP area is treated as memory hole and no copy is done.
2065 *
2066 * If [addr...addr+count) doesn't includes any intersects with alive
2067 * vm_struct area, returns 0. @buf should be kernel's buffer.
2068 *
2069 * Note: In usual ops, vwrite() is never necessary because the caller
2070 * should know vmalloc() area is valid and can use memcpy().
2071 * This is for routines which have to access vmalloc area without
2072 * any informaion, as /dev/kmem.
2073 */
2074
2075long vwrite(char *buf, char *addr, unsigned long count)
2076{
2077 struct vmap_area *va;
2078 struct vm_struct *vm;
2079 char *vaddr;
2080 unsigned long n, buflen;
2081 int copied = 0;
2082
2083 /* Don't allow overflow */
2084 if ((unsigned long) addr + count < count)
2085 count = -(unsigned long) addr;
2086 buflen = count;
2087
2088 spin_lock(&vmap_area_lock);
2089 list_for_each_entry(va, &vmap_area_list, list) {
2090 if (!count)
2091 break;
2092
2093 if (!(va->flags & VM_VM_AREA))
2094 continue;
2095
2096 vm = va->vm;
2097 vaddr = (char *) vm->addr;
2098 if (addr >= vaddr + get_vm_area_size(vm))
2099 continue;
2100 while (addr < vaddr) {
2101 if (count == 0)
2102 goto finished;
2103 buf++;
2104 addr++;
2105 count--;
2106 }
2107 n = vaddr + get_vm_area_size(vm) - addr;
2108 if (n > count)
2109 n = count;
2110 if (!(vm->flags & VM_IOREMAP)) {
2111 aligned_vwrite(buf, addr, n);
2112 copied++;
2113 }
2114 buf += n;
2115 addr += n;
2116 count -= n;
2117 }
2118finished:
2119 spin_unlock(&vmap_area_lock);
2120 if (!copied)
2121 return 0;
2122 return buflen;
2123}
2124
2125/**
2126 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2127 * @vma: vma to cover
2128 * @uaddr: target user address to start at
2129 * @kaddr: virtual address of vmalloc kernel memory
2130 * @size: size of map area
2131 *
2132 * Returns: 0 for success, -Exxx on failure
2133 *
2134 * This function checks that @kaddr is a valid vmalloc'ed area,
2135 * and that it is big enough to cover the range starting at
2136 * @uaddr in @vma. Will return failure if that criteria isn't
2137 * met.
2138 *
2139 * Similar to remap_pfn_range() (see mm/memory.c)
2140 */
2141int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2142 void *kaddr, unsigned long size)
2143{
2144 struct vm_struct *area;
2145
2146 size = PAGE_ALIGN(size);
2147
2148 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2149 return -EINVAL;
2150
2151 area = find_vm_area(kaddr);
2152 if (!area)
2153 return -EINVAL;
2154
2155 if (!(area->flags & VM_USERMAP))
2156 return -EINVAL;
2157
2158 if (kaddr + size > area->addr + area->size)
2159 return -EINVAL;
2160
2161 do {
2162 struct page *page = vmalloc_to_page(kaddr);
2163 int ret;
2164
2165 ret = vm_insert_page(vma, uaddr, page);
2166 if (ret)
2167 return ret;
2168
2169 uaddr += PAGE_SIZE;
2170 kaddr += PAGE_SIZE;
2171 size -= PAGE_SIZE;
2172 } while (size > 0);
2173
2174 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2175
2176 return 0;
2177}
2178EXPORT_SYMBOL(remap_vmalloc_range_partial);
2179
2180/**
2181 * remap_vmalloc_range - map vmalloc pages to userspace
2182 * @vma: vma to cover (map full range of vma)
2183 * @addr: vmalloc memory
2184 * @pgoff: number of pages into addr before first page to map
2185 *
2186 * Returns: 0 for success, -Exxx on failure
2187 *
2188 * This function checks that addr is a valid vmalloc'ed area, and
2189 * that it is big enough to cover the vma. Will return failure if
2190 * that criteria isn't met.
2191 *
2192 * Similar to remap_pfn_range() (see mm/memory.c)
2193 */
2194int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2195 unsigned long pgoff)
2196{
2197 return remap_vmalloc_range_partial(vma, vma->vm_start,
2198 addr + (pgoff << PAGE_SHIFT),
2199 vma->vm_end - vma->vm_start);
2200}
2201EXPORT_SYMBOL(remap_vmalloc_range);
2202
2203/*
2204 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2205 * have one.
2206 */
2207void __weak vmalloc_sync_all(void)
2208{
2209}
2210
2211
2212static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2213{
2214 pte_t ***p = data;
2215
2216 if (p) {
2217 *(*p) = pte;
2218 (*p)++;
2219 }
2220 return 0;
2221}
2222
2223/**
2224 * alloc_vm_area - allocate a range of kernel address space
2225 * @size: size of the area
2226 * @ptes: returns the PTEs for the address space
2227 *
2228 * Returns: NULL on failure, vm_struct on success
2229 *
2230 * This function reserves a range of kernel address space, and
2231 * allocates pagetables to map that range. No actual mappings
2232 * are created.
2233 *
2234 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2235 * allocated for the VM area are returned.
2236 */
2237struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2238{
2239 struct vm_struct *area;
2240
2241 area = get_vm_area_caller(size, VM_IOREMAP,
2242 __builtin_return_address(0));
2243 if (area == NULL)
2244 return NULL;
2245
2246 /*
2247 * This ensures that page tables are constructed for this region
2248 * of kernel virtual address space and mapped into init_mm.
2249 */
2250 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2251 size, f, ptes ? &ptes : NULL)) {
2252 free_vm_area(area);
2253 return NULL;
2254 }
2255
2256 return area;
2257}
2258EXPORT_SYMBOL_GPL(alloc_vm_area);
2259
2260void free_vm_area(struct vm_struct *area)
2261{
2262 struct vm_struct *ret;
2263 ret = remove_vm_area(area->addr);
2264 BUG_ON(ret != area);
2265 kfree(area);
2266}
2267EXPORT_SYMBOL_GPL(free_vm_area);
2268
2269#ifdef CONFIG_SMP
2270static struct vmap_area *node_to_va(struct rb_node *n)
2271{
2272 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2273}
2274
2275/**
2276 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2277 * @end: target address
2278 * @pnext: out arg for the next vmap_area
2279 * @pprev: out arg for the previous vmap_area
2280 *
2281 * Returns: %true if either or both of next and prev are found,
2282 * %false if no vmap_area exists
2283 *
2284 * Find vmap_areas end addresses of which enclose @end. ie. if not
2285 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2286 */
2287static bool pvm_find_next_prev(unsigned long end,
2288 struct vmap_area **pnext,
2289 struct vmap_area **pprev)
2290{
2291 struct rb_node *n = vmap_area_root.rb_node;
2292 struct vmap_area *va = NULL;
2293
2294 while (n) {
2295 va = rb_entry(n, struct vmap_area, rb_node);
2296 if (end < va->va_end)
2297 n = n->rb_left;
2298 else if (end > va->va_end)
2299 n = n->rb_right;
2300 else
2301 break;
2302 }
2303
2304 if (!va)
2305 return false;
2306
2307 if (va->va_end > end) {
2308 *pnext = va;
2309 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2310 } else {
2311 *pprev = va;
2312 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2313 }
2314 return true;
2315}
2316
2317/**
2318 * pvm_determine_end - find the highest aligned address between two vmap_areas
2319 * @pnext: in/out arg for the next vmap_area
2320 * @pprev: in/out arg for the previous vmap_area
2321 * @align: alignment
2322 *
2323 * Returns: determined end address
2324 *
2325 * Find the highest aligned address between *@pnext and *@pprev below
2326 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2327 * down address is between the end addresses of the two vmap_areas.
2328 *
2329 * Please note that the address returned by this function may fall
2330 * inside *@pnext vmap_area. The caller is responsible for checking
2331 * that.
2332 */
2333static unsigned long pvm_determine_end(struct vmap_area **pnext,
2334 struct vmap_area **pprev,
2335 unsigned long align)
2336{
2337 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2338 unsigned long addr;
2339
2340 if (*pnext)
2341 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2342 else
2343 addr = vmalloc_end;
2344
2345 while (*pprev && (*pprev)->va_end > addr) {
2346 *pnext = *pprev;
2347 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2348 }
2349
2350 return addr;
2351}
2352
2353/**
2354 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2355 * @offsets: array containing offset of each area
2356 * @sizes: array containing size of each area
2357 * @nr_vms: the number of areas to allocate
2358 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2359 *
2360 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2361 * vm_structs on success, %NULL on failure
2362 *
2363 * Percpu allocator wants to use congruent vm areas so that it can
2364 * maintain the offsets among percpu areas. This function allocates
2365 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2366 * be scattered pretty far, distance between two areas easily going up
2367 * to gigabytes. To avoid interacting with regular vmallocs, these
2368 * areas are allocated from top.
2369 *
2370 * Despite its complicated look, this allocator is rather simple. It
2371 * does everything top-down and scans areas from the end looking for
2372 * matching slot. While scanning, if any of the areas overlaps with
2373 * existing vmap_area, the base address is pulled down to fit the
2374 * area. Scanning is repeated till all the areas fit and then all
2375 * necessary data structres are inserted and the result is returned.
2376 */
2377struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2378 const size_t *sizes, int nr_vms,
2379 size_t align)
2380{
2381 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2382 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2383 struct vmap_area **vas, *prev, *next;
2384 struct vm_struct **vms;
2385 int area, area2, last_area, term_area;
2386 unsigned long base, start, end, last_end;
2387 bool purged = false;
2388
2389 /* verify parameters and allocate data structures */
2390 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2391 for (last_area = 0, area = 0; area < nr_vms; area++) {
2392 start = offsets[area];
2393 end = start + sizes[area];
2394
2395 /* is everything aligned properly? */
2396 BUG_ON(!IS_ALIGNED(offsets[area], align));
2397 BUG_ON(!IS_ALIGNED(sizes[area], align));
2398
2399 /* detect the area with the highest address */
2400 if (start > offsets[last_area])
2401 last_area = area;
2402
2403 for (area2 = 0; area2 < nr_vms; area2++) {
2404 unsigned long start2 = offsets[area2];
2405 unsigned long end2 = start2 + sizes[area2];
2406
2407 if (area2 == area)
2408 continue;
2409
2410 BUG_ON(start2 >= start && start2 < end);
2411 BUG_ON(end2 <= end && end2 > start);
2412 }
2413 }
2414 last_end = offsets[last_area] + sizes[last_area];
2415
2416 if (vmalloc_end - vmalloc_start < last_end) {
2417 WARN_ON(true);
2418 return NULL;
2419 }
2420
2421 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2422 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2423 if (!vas || !vms)
2424 goto err_free2;
2425
2426 for (area = 0; area < nr_vms; area++) {
2427 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2428 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2429 if (!vas[area] || !vms[area])
2430 goto err_free;
2431 }
2432retry:
2433 spin_lock(&vmap_area_lock);
2434
2435 /* start scanning - we scan from the top, begin with the last area */
2436 area = term_area = last_area;
2437 start = offsets[area];
2438 end = start + sizes[area];
2439
2440 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2441 base = vmalloc_end - last_end;
2442 goto found;
2443 }
2444 base = pvm_determine_end(&next, &prev, align) - end;
2445
2446 while (true) {
2447 BUG_ON(next && next->va_end <= base + end);
2448 BUG_ON(prev && prev->va_end > base + end);
2449
2450 /*
2451 * base might have underflowed, add last_end before
2452 * comparing.
2453 */
2454 if (base + last_end < vmalloc_start + last_end) {
2455 spin_unlock(&vmap_area_lock);
2456 if (!purged) {
2457 purge_vmap_area_lazy();
2458 purged = true;
2459 goto retry;
2460 }
2461 goto err_free;
2462 }
2463
2464 /*
2465 * If next overlaps, move base downwards so that it's
2466 * right below next and then recheck.
2467 */
2468 if (next && next->va_start < base + end) {
2469 base = pvm_determine_end(&next, &prev, align) - end;
2470 term_area = area;
2471 continue;
2472 }
2473
2474 /*
2475 * If prev overlaps, shift down next and prev and move
2476 * base so that it's right below new next and then
2477 * recheck.
2478 */
2479 if (prev && prev->va_end > base + start) {
2480 next = prev;
2481 prev = node_to_va(rb_prev(&next->rb_node));
2482 base = pvm_determine_end(&next, &prev, align) - end;
2483 term_area = area;
2484 continue;
2485 }
2486
2487 /*
2488 * This area fits, move on to the previous one. If
2489 * the previous one is the terminal one, we're done.
2490 */
2491 area = (area + nr_vms - 1) % nr_vms;
2492 if (area == term_area)
2493 break;
2494 start = offsets[area];
2495 end = start + sizes[area];
2496 pvm_find_next_prev(base + end, &next, &prev);
2497 }
2498found:
2499 /* we've found a fitting base, insert all va's */
2500 for (area = 0; area < nr_vms; area++) {
2501 struct vmap_area *va = vas[area];
2502
2503 va->va_start = base + offsets[area];
2504 va->va_end = va->va_start + sizes[area];
2505 __insert_vmap_area(va);
2506 }
2507
2508 vmap_area_pcpu_hole = base + offsets[last_area];
2509
2510 spin_unlock(&vmap_area_lock);
2511
2512 /* insert all vm's */
2513 for (area = 0; area < nr_vms; area++)
2514 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2515 pcpu_get_vm_areas);
2516
2517 kfree(vas);
2518 return vms;
2519
2520err_free:
2521 for (area = 0; area < nr_vms; area++) {
2522 kfree(vas[area]);
2523 kfree(vms[area]);
2524 }
2525err_free2:
2526 kfree(vas);
2527 kfree(vms);
2528 return NULL;
2529}
2530
2531/**
2532 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2533 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2534 * @nr_vms: the number of allocated areas
2535 *
2536 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2537 */
2538void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2539{
2540 int i;
2541
2542 for (i = 0; i < nr_vms; i++)
2543 free_vm_area(vms[i]);
2544 kfree(vms);
2545}
2546#endif /* CONFIG_SMP */
2547
2548#ifdef CONFIG_PROC_FS
2549static void *s_start(struct seq_file *m, loff_t *pos)
2550 __acquires(&vmap_area_lock)
2551{
2552 loff_t n = *pos;
2553 struct vmap_area *va;
2554
2555 spin_lock(&vmap_area_lock);
2556 va = list_first_entry(&vmap_area_list, typeof(*va), list);
2557 while (n > 0 && &va->list != &vmap_area_list) {
2558 n--;
2559 va = list_next_entry(va, list);
2560 }
2561 if (!n && &va->list != &vmap_area_list)
2562 return va;
2563
2564 return NULL;
2565
2566}
2567
2568static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2569{
2570 struct vmap_area *va = p, *next;
2571
2572 ++*pos;
2573 next = list_next_entry(va, list);
2574 if (&next->list != &vmap_area_list)
2575 return next;
2576
2577 return NULL;
2578}
2579
2580static void s_stop(struct seq_file *m, void *p)
2581 __releases(&vmap_area_lock)
2582{
2583 spin_unlock(&vmap_area_lock);
2584}
2585
2586static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2587{
2588 if (IS_ENABLED(CONFIG_NUMA)) {
2589 unsigned int nr, *counters = m->private;
2590
2591 if (!counters)
2592 return;
2593
2594 if (v->flags & VM_UNINITIALIZED)
2595 return;
2596 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2597 smp_rmb();
2598
2599 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2600
2601 for (nr = 0; nr < v->nr_pages; nr++)
2602 counters[page_to_nid(v->pages[nr])]++;
2603
2604 for_each_node_state(nr, N_HIGH_MEMORY)
2605 if (counters[nr])
2606 seq_printf(m, " N%u=%u", nr, counters[nr]);
2607 }
2608}
2609
2610static int s_show(struct seq_file *m, void *p)
2611{
2612 struct vmap_area *va = p;
2613 struct vm_struct *v;
2614
2615 /*
2616 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2617 * behalf of vmap area is being tear down or vm_map_ram allocation.
2618 */
2619 if (!(va->flags & VM_VM_AREA))
2620 return 0;
2621
2622 v = va->vm;
2623
2624 seq_printf(m, "0x%pK-0x%pK %7ld",
2625 v->addr, v->addr + v->size, v->size);
2626
2627 if (v->caller)
2628 seq_printf(m, " %pS", v->caller);
2629
2630 if (v->nr_pages)
2631 seq_printf(m, " pages=%d", v->nr_pages);
2632
2633 if (v->phys_addr)
2634 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2635
2636 if (v->flags & VM_IOREMAP)
2637 seq_puts(m, " ioremap");
2638
2639 if (v->flags & VM_ALLOC)
2640 seq_puts(m, " vmalloc");
2641
2642 if (v->flags & VM_MAP)
2643 seq_puts(m, " vmap");
2644
2645 if (v->flags & VM_USERMAP)
2646 seq_puts(m, " user");
2647
2648 if (is_vmalloc_addr(v->pages))
2649 seq_puts(m, " vpages");
2650
2651 show_numa_info(m, v);
2652 seq_putc(m, '\n');
2653 return 0;
2654}
2655
2656static const struct seq_operations vmalloc_op = {
2657 .start = s_start,
2658 .next = s_next,
2659 .stop = s_stop,
2660 .show = s_show,
2661};
2662
2663static int vmalloc_open(struct inode *inode, struct file *file)
2664{
2665 if (IS_ENABLED(CONFIG_NUMA))
2666 return seq_open_private(file, &vmalloc_op,
2667 nr_node_ids * sizeof(unsigned int));
2668 else
2669 return seq_open(file, &vmalloc_op);
2670}
2671
2672static const struct file_operations proc_vmalloc_operations = {
2673 .open = vmalloc_open,
2674 .read = seq_read,
2675 .llseek = seq_lseek,
2676 .release = seq_release_private,
2677};
2678
2679static int __init proc_vmalloc_init(void)
2680{
2681 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2682 return 0;
2683}
2684module_init(proc_vmalloc_init);
2685
2686#endif
2687