1 // SPDX-License-Identifier: GPL-2.0-only
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
12 #include <linux/vmalloc.h>
14 #include <linux/module.h>
15 #include <linux/highmem.h>
16 #include <linux/sched/signal.h>
17 #include <linux/slab.h>
18 #include <linux/spinlock.h>
19 #include <linux/interrupt.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/set_memory.h>
23 #include <linux/debugobjects.h>
24 #include <linux/kallsyms.h>
25 #include <linux/list.h>
26 #include <linux/notifier.h>
27 #include <linux/rbtree.h>
28 #include <linux/radix-tree.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/llist.h>
35 #include <linux/bitops.h>
36 #include <linux/rbtree_augmented.h>
37 #include <linux/overflow.h>
39 #include <linux/uaccess.h>
40 #include <asm/tlbflush.h>
41 #include <asm/shmparam.h>
45 bool is_vmalloc_addr(const void *x
)
47 unsigned long addr
= (unsigned long)x
;
49 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
51 EXPORT_SYMBOL(is_vmalloc_addr
);
53 struct vfree_deferred
{
54 struct llist_head list
;
55 struct work_struct wq
;
57 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
59 static void __vunmap(const void *, int);
61 static void free_work(struct work_struct
*w
)
63 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
64 struct llist_node
*t
, *llnode
;
66 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
67 __vunmap((void *)llnode
, 1);
70 /*** Page table manipulation functions ***/
72 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
)
76 pte
= pte_offset_kernel(pmd
, addr
);
78 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
79 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
80 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
83 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
)
88 pmd
= pmd_offset(pud
, addr
);
90 next
= pmd_addr_end(addr
, end
);
91 if (pmd_clear_huge(pmd
))
93 if (pmd_none_or_clear_bad(pmd
))
95 vunmap_pte_range(pmd
, addr
, next
);
96 } while (pmd
++, addr
= next
, addr
!= end
);
99 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
)
104 pud
= pud_offset(p4d
, addr
);
106 next
= pud_addr_end(addr
, end
);
107 if (pud_clear_huge(pud
))
109 if (pud_none_or_clear_bad(pud
))
111 vunmap_pmd_range(pud
, addr
, next
);
112 } while (pud
++, addr
= next
, addr
!= end
);
115 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
)
120 p4d
= p4d_offset(pgd
, addr
);
122 next
= p4d_addr_end(addr
, end
);
123 if (p4d_clear_huge(p4d
))
125 if (p4d_none_or_clear_bad(p4d
))
127 vunmap_pud_range(p4d
, addr
, next
);
128 } while (p4d
++, addr
= next
, addr
!= end
);
131 static void vunmap_page_range(unsigned long addr
, unsigned long end
)
137 pgd
= pgd_offset_k(addr
);
139 next
= pgd_addr_end(addr
, end
);
140 if (pgd_none_or_clear_bad(pgd
))
142 vunmap_p4d_range(pgd
, addr
, next
);
143 } while (pgd
++, addr
= next
, addr
!= end
);
146 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
147 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
152 * nr is a running index into the array which helps higher level
153 * callers keep track of where we're up to.
156 pte
= pte_alloc_kernel(pmd
, addr
);
160 struct page
*page
= pages
[*nr
];
162 if (WARN_ON(!pte_none(*pte
)))
166 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
168 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
172 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
173 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
178 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
182 next
= pmd_addr_end(addr
, end
);
183 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
185 } while (pmd
++, addr
= next
, addr
!= end
);
189 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
190 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
195 pud
= pud_alloc(&init_mm
, p4d
, addr
);
199 next
= pud_addr_end(addr
, end
);
200 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
202 } while (pud
++, addr
= next
, addr
!= end
);
206 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
207 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
212 p4d
= p4d_alloc(&init_mm
, pgd
, addr
);
216 next
= p4d_addr_end(addr
, end
);
217 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
))
219 } while (p4d
++, addr
= next
, addr
!= end
);
224 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
225 * will have pfns corresponding to the "pages" array.
227 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
229 static int vmap_page_range_noflush(unsigned long start
, unsigned long end
,
230 pgprot_t prot
, struct page
**pages
)
234 unsigned long addr
= start
;
239 pgd
= pgd_offset_k(addr
);
241 next
= pgd_addr_end(addr
, end
);
242 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
);
245 } while (pgd
++, addr
= next
, addr
!= end
);
250 static int vmap_page_range(unsigned long start
, unsigned long end
,
251 pgprot_t prot
, struct page
**pages
)
255 ret
= vmap_page_range_noflush(start
, end
, prot
, pages
);
256 flush_cache_vmap(start
, end
);
260 int is_vmalloc_or_module_addr(const void *x
)
263 * ARM, x86-64 and sparc64 put modules in a special place,
264 * and fall back on vmalloc() if that fails. Others
265 * just put it in the vmalloc space.
267 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
268 unsigned long addr
= (unsigned long)x
;
269 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
272 return is_vmalloc_addr(x
);
276 * Walk a vmap address to the struct page it maps.
278 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
280 unsigned long addr
= (unsigned long) vmalloc_addr
;
281 struct page
*page
= NULL
;
282 pgd_t
*pgd
= pgd_offset_k(addr
);
289 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
290 * architectures that do not vmalloc module space
292 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
296 p4d
= p4d_offset(pgd
, addr
);
299 pud
= pud_offset(p4d
, addr
);
302 * Don't dereference bad PUD or PMD (below) entries. This will also
303 * identify huge mappings, which we may encounter on architectures
304 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
305 * identified as vmalloc addresses by is_vmalloc_addr(), but are
306 * not [unambiguously] associated with a struct page, so there is
307 * no correct value to return for them.
309 WARN_ON_ONCE(pud_bad(*pud
));
310 if (pud_none(*pud
) || pud_bad(*pud
))
312 pmd
= pmd_offset(pud
, addr
);
313 WARN_ON_ONCE(pmd_bad(*pmd
));
314 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
317 ptep
= pte_offset_map(pmd
, addr
);
319 if (pte_present(pte
))
320 page
= pte_page(pte
);
324 EXPORT_SYMBOL(vmalloc_to_page
);
327 * Map a vmalloc()-space virtual address to the physical page frame number.
329 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
331 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
333 EXPORT_SYMBOL(vmalloc_to_pfn
);
336 /*** Global kva allocator ***/
338 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
339 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
342 static DEFINE_SPINLOCK(vmap_area_lock
);
343 static DEFINE_SPINLOCK(free_vmap_area_lock
);
344 /* Export for kexec only */
345 LIST_HEAD(vmap_area_list
);
346 static LLIST_HEAD(vmap_purge_list
);
347 static struct rb_root vmap_area_root
= RB_ROOT
;
348 static bool vmap_initialized __read_mostly
;
351 * This kmem_cache is used for vmap_area objects. Instead of
352 * allocating from slab we reuse an object from this cache to
353 * make things faster. Especially in "no edge" splitting of
356 static struct kmem_cache
*vmap_area_cachep
;
359 * This linked list is used in pair with free_vmap_area_root.
360 * It gives O(1) access to prev/next to perform fast coalescing.
362 static LIST_HEAD(free_vmap_area_list
);
365 * This augment red-black tree represents the free vmap space.
366 * All vmap_area objects in this tree are sorted by va->va_start
367 * address. It is used for allocation and merging when a vmap
368 * object is released.
370 * Each vmap_area node contains a maximum available free block
371 * of its sub-tree, right or left. Therefore it is possible to
372 * find a lowest match of free area.
374 static struct rb_root free_vmap_area_root
= RB_ROOT
;
377 * Preload a CPU with one object for "no edge" split case. The
378 * aim is to get rid of allocations from the atomic context, thus
379 * to use more permissive allocation masks.
381 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
383 static __always_inline
unsigned long
384 va_size(struct vmap_area
*va
)
386 return (va
->va_end
- va
->va_start
);
389 static __always_inline
unsigned long
390 get_subtree_max_size(struct rb_node
*node
)
392 struct vmap_area
*va
;
394 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
395 return va
? va
->subtree_max_size
: 0;
399 * Gets called when remove the node and rotate.
401 static __always_inline
unsigned long
402 compute_subtree_max_size(struct vmap_area
*va
)
404 return max3(va_size(va
),
405 get_subtree_max_size(va
->rb_node
.rb_left
),
406 get_subtree_max_size(va
->rb_node
.rb_right
));
409 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
410 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
412 static void purge_vmap_area_lazy(void);
413 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
414 static unsigned long lazy_max_pages(void);
416 static atomic_long_t nr_vmalloc_pages
;
418 unsigned long vmalloc_nr_pages(void)
420 return atomic_long_read(&nr_vmalloc_pages
);
423 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
425 struct rb_node
*n
= vmap_area_root
.rb_node
;
428 struct vmap_area
*va
;
430 va
= rb_entry(n
, struct vmap_area
, rb_node
);
431 if (addr
< va
->va_start
)
433 else if (addr
>= va
->va_end
)
443 * This function returns back addresses of parent node
444 * and its left or right link for further processing.
446 static __always_inline
struct rb_node
**
447 find_va_links(struct vmap_area
*va
,
448 struct rb_root
*root
, struct rb_node
*from
,
449 struct rb_node
**parent
)
451 struct vmap_area
*tmp_va
;
452 struct rb_node
**link
;
455 link
= &root
->rb_node
;
456 if (unlikely(!*link
)) {
465 * Go to the bottom of the tree. When we hit the last point
466 * we end up with parent rb_node and correct direction, i name
467 * it link, where the new va->rb_node will be attached to.
470 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
473 * During the traversal we also do some sanity check.
474 * Trigger the BUG() if there are sides(left/right)
477 if (va
->va_start
< tmp_va
->va_end
&&
478 va
->va_end
<= tmp_va
->va_start
)
479 link
= &(*link
)->rb_left
;
480 else if (va
->va_end
> tmp_va
->va_start
&&
481 va
->va_start
>= tmp_va
->va_end
)
482 link
= &(*link
)->rb_right
;
487 *parent
= &tmp_va
->rb_node
;
491 static __always_inline
struct list_head
*
492 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
494 struct list_head
*list
;
496 if (unlikely(!parent
))
498 * The red-black tree where we try to find VA neighbors
499 * before merging or inserting is empty, i.e. it means
500 * there is no free vmap space. Normally it does not
501 * happen but we handle this case anyway.
505 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
506 return (&parent
->rb_right
== link
? list
->next
: list
);
509 static __always_inline
void
510 link_va(struct vmap_area
*va
, struct rb_root
*root
,
511 struct rb_node
*parent
, struct rb_node
**link
, struct list_head
*head
)
514 * VA is still not in the list, but we can
515 * identify its future previous list_head node.
517 if (likely(parent
)) {
518 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
519 if (&parent
->rb_right
!= link
)
523 /* Insert to the rb-tree */
524 rb_link_node(&va
->rb_node
, parent
, link
);
525 if (root
== &free_vmap_area_root
) {
527 * Some explanation here. Just perform simple insertion
528 * to the tree. We do not set va->subtree_max_size to
529 * its current size before calling rb_insert_augmented().
530 * It is because of we populate the tree from the bottom
531 * to parent levels when the node _is_ in the tree.
533 * Therefore we set subtree_max_size to zero after insertion,
534 * to let __augment_tree_propagate_from() puts everything to
535 * the correct order later on.
537 rb_insert_augmented(&va
->rb_node
,
538 root
, &free_vmap_area_rb_augment_cb
);
539 va
->subtree_max_size
= 0;
541 rb_insert_color(&va
->rb_node
, root
);
544 /* Address-sort this list */
545 list_add(&va
->list
, head
);
548 static __always_inline
void
549 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
551 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
554 if (root
== &free_vmap_area_root
)
555 rb_erase_augmented(&va
->rb_node
,
556 root
, &free_vmap_area_rb_augment_cb
);
558 rb_erase(&va
->rb_node
, root
);
561 RB_CLEAR_NODE(&va
->rb_node
);
564 #if DEBUG_AUGMENT_PROPAGATE_CHECK
566 augment_tree_propagate_check(struct rb_node
*n
)
568 struct vmap_area
*va
;
569 struct rb_node
*node
;
576 va
= rb_entry(n
, struct vmap_area
, rb_node
);
577 size
= va
->subtree_max_size
;
581 va
= rb_entry(node
, struct vmap_area
, rb_node
);
583 if (get_subtree_max_size(node
->rb_left
) == size
) {
584 node
= node
->rb_left
;
586 if (va_size(va
) == size
) {
591 node
= node
->rb_right
;
596 va
= rb_entry(n
, struct vmap_area
, rb_node
);
597 pr_emerg("tree is corrupted: %lu, %lu\n",
598 va_size(va
), va
->subtree_max_size
);
601 augment_tree_propagate_check(n
->rb_left
);
602 augment_tree_propagate_check(n
->rb_right
);
607 * This function populates subtree_max_size from bottom to upper
608 * levels starting from VA point. The propagation must be done
609 * when VA size is modified by changing its va_start/va_end. Or
610 * in case of newly inserting of VA to the tree.
612 * It means that __augment_tree_propagate_from() must be called:
613 * - After VA has been inserted to the tree(free path);
614 * - After VA has been shrunk(allocation path);
615 * - After VA has been increased(merging path).
617 * Please note that, it does not mean that upper parent nodes
618 * and their subtree_max_size are recalculated all the time up
627 * For example if we modify the node 4, shrinking it to 2, then
628 * no any modification is required. If we shrink the node 2 to 1
629 * its subtree_max_size is updated only, and set to 1. If we shrink
630 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
633 static __always_inline
void
634 augment_tree_propagate_from(struct vmap_area
*va
)
636 struct rb_node
*node
= &va
->rb_node
;
637 unsigned long new_va_sub_max_size
;
640 va
= rb_entry(node
, struct vmap_area
, rb_node
);
641 new_va_sub_max_size
= compute_subtree_max_size(va
);
644 * If the newly calculated maximum available size of the
645 * subtree is equal to the current one, then it means that
646 * the tree is propagated correctly. So we have to stop at
647 * this point to save cycles.
649 if (va
->subtree_max_size
== new_va_sub_max_size
)
652 va
->subtree_max_size
= new_va_sub_max_size
;
653 node
= rb_parent(&va
->rb_node
);
656 #if DEBUG_AUGMENT_PROPAGATE_CHECK
657 augment_tree_propagate_check(free_vmap_area_root
.rb_node
);
662 insert_vmap_area(struct vmap_area
*va
,
663 struct rb_root
*root
, struct list_head
*head
)
665 struct rb_node
**link
;
666 struct rb_node
*parent
;
668 link
= find_va_links(va
, root
, NULL
, &parent
);
669 link_va(va
, root
, parent
, link
, head
);
673 insert_vmap_area_augment(struct vmap_area
*va
,
674 struct rb_node
*from
, struct rb_root
*root
,
675 struct list_head
*head
)
677 struct rb_node
**link
;
678 struct rb_node
*parent
;
681 link
= find_va_links(va
, NULL
, from
, &parent
);
683 link
= find_va_links(va
, root
, NULL
, &parent
);
685 link_va(va
, root
, parent
, link
, head
);
686 augment_tree_propagate_from(va
);
690 * Merge de-allocated chunk of VA memory with previous
691 * and next free blocks. If coalesce is not done a new
692 * free area is inserted. If VA has been merged, it is
695 static __always_inline
struct vmap_area
*
696 merge_or_add_vmap_area(struct vmap_area
*va
,
697 struct rb_root
*root
, struct list_head
*head
)
699 struct vmap_area
*sibling
;
700 struct list_head
*next
;
701 struct rb_node
**link
;
702 struct rb_node
*parent
;
706 * Find a place in the tree where VA potentially will be
707 * inserted, unless it is merged with its sibling/siblings.
709 link
= find_va_links(va
, root
, NULL
, &parent
);
712 * Get next node of VA to check if merging can be done.
714 next
= get_va_next_sibling(parent
, link
);
715 if (unlikely(next
== NULL
))
721 * |<------VA------>|<-----Next----->|
726 sibling
= list_entry(next
, struct vmap_area
, list
);
727 if (sibling
->va_start
== va
->va_end
) {
728 sibling
->va_start
= va
->va_start
;
730 /* Check and update the tree if needed. */
731 augment_tree_propagate_from(sibling
);
733 /* Free vmap_area object. */
734 kmem_cache_free(vmap_area_cachep
, va
);
736 /* Point to the new merged area. */
745 * |<-----Prev----->|<------VA------>|
749 if (next
->prev
!= head
) {
750 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
751 if (sibling
->va_end
== va
->va_start
) {
752 sibling
->va_end
= va
->va_end
;
754 /* Check and update the tree if needed. */
755 augment_tree_propagate_from(sibling
);
760 /* Free vmap_area object. */
761 kmem_cache_free(vmap_area_cachep
, va
);
763 /* Point to the new merged area. */
771 link_va(va
, root
, parent
, link
, head
);
772 augment_tree_propagate_from(va
);
778 static __always_inline
bool
779 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
780 unsigned long align
, unsigned long vstart
)
782 unsigned long nva_start_addr
;
784 if (va
->va_start
> vstart
)
785 nva_start_addr
= ALIGN(va
->va_start
, align
);
787 nva_start_addr
= ALIGN(vstart
, align
);
789 /* Can be overflowed due to big size or alignment. */
790 if (nva_start_addr
+ size
< nva_start_addr
||
791 nva_start_addr
< vstart
)
794 return (nva_start_addr
+ size
<= va
->va_end
);
798 * Find the first free block(lowest start address) in the tree,
799 * that will accomplish the request corresponding to passing
802 static __always_inline
struct vmap_area
*
803 find_vmap_lowest_match(unsigned long size
,
804 unsigned long align
, unsigned long vstart
)
806 struct vmap_area
*va
;
807 struct rb_node
*node
;
808 unsigned long length
;
810 /* Start from the root. */
811 node
= free_vmap_area_root
.rb_node
;
813 /* Adjust the search size for alignment overhead. */
814 length
= size
+ align
- 1;
817 va
= rb_entry(node
, struct vmap_area
, rb_node
);
819 if (get_subtree_max_size(node
->rb_left
) >= length
&&
820 vstart
< va
->va_start
) {
821 node
= node
->rb_left
;
823 if (is_within_this_va(va
, size
, align
, vstart
))
827 * Does not make sense to go deeper towards the right
828 * sub-tree if it does not have a free block that is
829 * equal or bigger to the requested search length.
831 if (get_subtree_max_size(node
->rb_right
) >= length
) {
832 node
= node
->rb_right
;
837 * OK. We roll back and find the first right sub-tree,
838 * that will satisfy the search criteria. It can happen
839 * only once due to "vstart" restriction.
841 while ((node
= rb_parent(node
))) {
842 va
= rb_entry(node
, struct vmap_area
, rb_node
);
843 if (is_within_this_va(va
, size
, align
, vstart
))
846 if (get_subtree_max_size(node
->rb_right
) >= length
&&
847 vstart
<= va
->va_start
) {
848 node
= node
->rb_right
;
858 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
859 #include <linux/random.h>
861 static struct vmap_area
*
862 find_vmap_lowest_linear_match(unsigned long size
,
863 unsigned long align
, unsigned long vstart
)
865 struct vmap_area
*va
;
867 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
868 if (!is_within_this_va(va
, size
, align
, vstart
))
878 find_vmap_lowest_match_check(unsigned long size
)
880 struct vmap_area
*va_1
, *va_2
;
881 unsigned long vstart
;
884 get_random_bytes(&rnd
, sizeof(rnd
));
885 vstart
= VMALLOC_START
+ rnd
;
887 va_1
= find_vmap_lowest_match(size
, 1, vstart
);
888 va_2
= find_vmap_lowest_linear_match(size
, 1, vstart
);
891 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
898 FL_FIT_TYPE
= 1, /* full fit */
899 LE_FIT_TYPE
= 2, /* left edge fit */
900 RE_FIT_TYPE
= 3, /* right edge fit */
901 NE_FIT_TYPE
= 4 /* no edge fit */
904 static __always_inline
enum fit_type
905 classify_va_fit_type(struct vmap_area
*va
,
906 unsigned long nva_start_addr
, unsigned long size
)
910 /* Check if it is within VA. */
911 if (nva_start_addr
< va
->va_start
||
912 nva_start_addr
+ size
> va
->va_end
)
916 if (va
->va_start
== nva_start_addr
) {
917 if (va
->va_end
== nva_start_addr
+ size
)
921 } else if (va
->va_end
== nva_start_addr
+ size
) {
930 static __always_inline
int
931 adjust_va_to_fit_type(struct vmap_area
*va
,
932 unsigned long nva_start_addr
, unsigned long size
,
935 struct vmap_area
*lva
= NULL
;
937 if (type
== FL_FIT_TYPE
) {
939 * No need to split VA, it fully fits.
945 unlink_va(va
, &free_vmap_area_root
);
946 kmem_cache_free(vmap_area_cachep
, va
);
947 } else if (type
== LE_FIT_TYPE
) {
949 * Split left edge of fit VA.
955 va
->va_start
+= size
;
956 } else if (type
== RE_FIT_TYPE
) {
958 * Split right edge of fit VA.
964 va
->va_end
= nva_start_addr
;
965 } else if (type
== NE_FIT_TYPE
) {
967 * Split no edge of fit VA.
973 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
974 if (unlikely(!lva
)) {
976 * For percpu allocator we do not do any pre-allocation
977 * and leave it as it is. The reason is it most likely
978 * never ends up with NE_FIT_TYPE splitting. In case of
979 * percpu allocations offsets and sizes are aligned to
980 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
981 * are its main fitting cases.
983 * There are a few exceptions though, as an example it is
984 * a first allocation (early boot up) when we have "one"
985 * big free space that has to be split.
987 * Also we can hit this path in case of regular "vmap"
988 * allocations, if "this" current CPU was not preloaded.
989 * See the comment in alloc_vmap_area() why. If so, then
990 * GFP_NOWAIT is used instead to get an extra object for
991 * split purpose. That is rare and most time does not
994 * What happens if an allocation gets failed. Basically,
995 * an "overflow" path is triggered to purge lazily freed
996 * areas to free some memory, then, the "retry" path is
997 * triggered to repeat one more time. See more details
998 * in alloc_vmap_area() function.
1000 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1006 * Build the remainder.
1008 lva
->va_start
= va
->va_start
;
1009 lva
->va_end
= nva_start_addr
;
1012 * Shrink this VA to remaining size.
1014 va
->va_start
= nva_start_addr
+ size
;
1019 if (type
!= FL_FIT_TYPE
) {
1020 augment_tree_propagate_from(va
);
1022 if (lva
) /* type == NE_FIT_TYPE */
1023 insert_vmap_area_augment(lva
, &va
->rb_node
,
1024 &free_vmap_area_root
, &free_vmap_area_list
);
1031 * Returns a start address of the newly allocated area, if success.
1032 * Otherwise a vend is returned that indicates failure.
1034 static __always_inline
unsigned long
1035 __alloc_vmap_area(unsigned long size
, unsigned long align
,
1036 unsigned long vstart
, unsigned long vend
)
1038 unsigned long nva_start_addr
;
1039 struct vmap_area
*va
;
1043 va
= find_vmap_lowest_match(size
, align
, vstart
);
1047 if (va
->va_start
> vstart
)
1048 nva_start_addr
= ALIGN(va
->va_start
, align
);
1050 nva_start_addr
= ALIGN(vstart
, align
);
1052 /* Check the "vend" restriction. */
1053 if (nva_start_addr
+ size
> vend
)
1056 /* Classify what we have found. */
1057 type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1058 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
1061 /* Update the free vmap_area. */
1062 ret
= adjust_va_to_fit_type(va
, nva_start_addr
, size
, type
);
1066 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1067 find_vmap_lowest_match_check(size
);
1070 return nva_start_addr
;
1074 * Free a region of KVA allocated by alloc_vmap_area
1076 static void free_vmap_area(struct vmap_area
*va
)
1079 * Remove from the busy tree/list.
1081 spin_lock(&vmap_area_lock
);
1082 unlink_va(va
, &vmap_area_root
);
1083 spin_unlock(&vmap_area_lock
);
1086 * Insert/Merge it back to the free tree/list.
1088 spin_lock(&free_vmap_area_lock
);
1089 merge_or_add_vmap_area(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1090 spin_unlock(&free_vmap_area_lock
);
1094 * Allocate a region of KVA of the specified size and alignment, within the
1097 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1098 unsigned long align
,
1099 unsigned long vstart
, unsigned long vend
,
1100 int node
, gfp_t gfp_mask
)
1102 struct vmap_area
*va
, *pva
;
1108 BUG_ON(offset_in_page(size
));
1109 BUG_ON(!is_power_of_2(align
));
1111 if (unlikely(!vmap_initialized
))
1112 return ERR_PTR(-EBUSY
);
1115 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1117 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1119 return ERR_PTR(-ENOMEM
);
1122 * Only scan the relevant parts containing pointers to other objects
1123 * to avoid false negatives.
1125 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1129 * Preload this CPU with one extra vmap_area object. It is used
1130 * when fit type of free area is NE_FIT_TYPE. Please note, it
1131 * does not guarantee that an allocation occurs on a CPU that
1132 * is preloaded, instead we minimize the case when it is not.
1133 * It can happen because of cpu migration, because there is a
1134 * race until the below spinlock is taken.
1136 * The preload is done in non-atomic context, thus it allows us
1137 * to use more permissive allocation masks to be more stable under
1138 * low memory condition and high memory pressure. In rare case,
1139 * if not preloaded, GFP_NOWAIT is used.
1141 * Set "pva" to NULL here, because of "retry" path.
1145 if (!this_cpu_read(ne_fit_preload_node
))
1147 * Even if it fails we do not really care about that.
1148 * Just proceed as it is. If needed "overflow" path
1149 * will refill the cache we allocate from.
1151 pva
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1153 spin_lock(&free_vmap_area_lock
);
1155 if (pva
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, pva
))
1156 kmem_cache_free(vmap_area_cachep
, pva
);
1159 * If an allocation fails, the "vend" address is
1160 * returned. Therefore trigger the overflow path.
1162 addr
= __alloc_vmap_area(size
, align
, vstart
, vend
);
1163 spin_unlock(&free_vmap_area_lock
);
1165 if (unlikely(addr
== vend
))
1168 va
->va_start
= addr
;
1169 va
->va_end
= addr
+ size
;
1173 spin_lock(&vmap_area_lock
);
1174 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1175 spin_unlock(&vmap_area_lock
);
1177 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1178 BUG_ON(va
->va_start
< vstart
);
1179 BUG_ON(va
->va_end
> vend
);
1181 ret
= kasan_populate_vmalloc(addr
, size
);
1184 return ERR_PTR(ret
);
1191 purge_vmap_area_lazy();
1196 if (gfpflags_allow_blocking(gfp_mask
)) {
1197 unsigned long freed
= 0;
1198 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1205 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1206 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1209 kmem_cache_free(vmap_area_cachep
, va
);
1210 return ERR_PTR(-EBUSY
);
1213 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1215 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1217 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1219 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1221 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1223 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1226 * Clear the pagetable entries of a given vmap_area
1228 static void unmap_vmap_area(struct vmap_area
*va
)
1230 vunmap_page_range(va
->va_start
, va
->va_end
);
1234 * lazy_max_pages is the maximum amount of virtual address space we gather up
1235 * before attempting to purge with a TLB flush.
1237 * There is a tradeoff here: a larger number will cover more kernel page tables
1238 * and take slightly longer to purge, but it will linearly reduce the number of
1239 * global TLB flushes that must be performed. It would seem natural to scale
1240 * this number up linearly with the number of CPUs (because vmapping activity
1241 * could also scale linearly with the number of CPUs), however it is likely
1242 * that in practice, workloads might be constrained in other ways that mean
1243 * vmap activity will not scale linearly with CPUs. Also, I want to be
1244 * conservative and not introduce a big latency on huge systems, so go with
1245 * a less aggressive log scale. It will still be an improvement over the old
1246 * code, and it will be simple to change the scale factor if we find that it
1247 * becomes a problem on bigger systems.
1249 static unsigned long lazy_max_pages(void)
1253 log
= fls(num_online_cpus());
1255 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1258 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1261 * Serialize vmap purging. There is no actual criticial section protected
1262 * by this look, but we want to avoid concurrent calls for performance
1263 * reasons and to make the pcpu_get_vm_areas more deterministic.
1265 static DEFINE_MUTEX(vmap_purge_lock
);
1267 /* for per-CPU blocks */
1268 static void purge_fragmented_blocks_allcpus(void);
1271 * called before a call to iounmap() if the caller wants vm_area_struct's
1272 * immediately freed.
1274 void set_iounmap_nonlazy(void)
1276 atomic_long_set(&vmap_lazy_nr
, lazy_max_pages()+1);
1280 * Purges all lazily-freed vmap areas.
1282 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1284 unsigned long resched_threshold
;
1285 struct llist_node
*valist
;
1286 struct vmap_area
*va
;
1287 struct vmap_area
*n_va
;
1289 lockdep_assert_held(&vmap_purge_lock
);
1291 valist
= llist_del_all(&vmap_purge_list
);
1292 if (unlikely(valist
== NULL
))
1296 * First make sure the mappings are removed from all page-tables
1297 * before they are freed.
1299 vmalloc_sync_unmappings();
1302 * TODO: to calculate a flush range without looping.
1303 * The list can be up to lazy_max_pages() elements.
1305 llist_for_each_entry(va
, valist
, purge_list
) {
1306 if (va
->va_start
< start
)
1307 start
= va
->va_start
;
1308 if (va
->va_end
> end
)
1312 flush_tlb_kernel_range(start
, end
);
1313 resched_threshold
= lazy_max_pages() << 1;
1315 spin_lock(&free_vmap_area_lock
);
1316 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
1317 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1318 unsigned long orig_start
= va
->va_start
;
1319 unsigned long orig_end
= va
->va_end
;
1322 * Finally insert or merge lazily-freed area. It is
1323 * detached and there is no need to "unlink" it from
1326 va
= merge_or_add_vmap_area(va
, &free_vmap_area_root
,
1327 &free_vmap_area_list
);
1329 if (is_vmalloc_or_module_addr((void *)orig_start
))
1330 kasan_release_vmalloc(orig_start
, orig_end
,
1331 va
->va_start
, va
->va_end
);
1333 atomic_long_sub(nr
, &vmap_lazy_nr
);
1335 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1336 cond_resched_lock(&free_vmap_area_lock
);
1338 spin_unlock(&free_vmap_area_lock
);
1343 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1344 * is already purging.
1346 static void try_purge_vmap_area_lazy(void)
1348 if (mutex_trylock(&vmap_purge_lock
)) {
1349 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1350 mutex_unlock(&vmap_purge_lock
);
1355 * Kick off a purge of the outstanding lazy areas.
1357 static void purge_vmap_area_lazy(void)
1359 mutex_lock(&vmap_purge_lock
);
1360 purge_fragmented_blocks_allcpus();
1361 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1362 mutex_unlock(&vmap_purge_lock
);
1366 * Free a vmap area, caller ensuring that the area has been unmapped
1367 * and flush_cache_vunmap had been called for the correct range
1370 static void free_vmap_area_noflush(struct vmap_area
*va
)
1372 unsigned long nr_lazy
;
1374 spin_lock(&vmap_area_lock
);
1375 unlink_va(va
, &vmap_area_root
);
1376 spin_unlock(&vmap_area_lock
);
1378 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1379 PAGE_SHIFT
, &vmap_lazy_nr
);
1381 /* After this point, we may free va at any time */
1382 llist_add(&va
->purge_list
, &vmap_purge_list
);
1384 if (unlikely(nr_lazy
> lazy_max_pages()))
1385 try_purge_vmap_area_lazy();
1389 * Free and unmap a vmap area
1391 static void free_unmap_vmap_area(struct vmap_area
*va
)
1393 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1394 unmap_vmap_area(va
);
1395 if (debug_pagealloc_enabled_static())
1396 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1398 free_vmap_area_noflush(va
);
1401 static struct vmap_area
*find_vmap_area(unsigned long addr
)
1403 struct vmap_area
*va
;
1405 spin_lock(&vmap_area_lock
);
1406 va
= __find_vmap_area(addr
);
1407 spin_unlock(&vmap_area_lock
);
1412 /*** Per cpu kva allocator ***/
1415 * vmap space is limited especially on 32 bit architectures. Ensure there is
1416 * room for at least 16 percpu vmap blocks per CPU.
1419 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1420 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1421 * instead (we just need a rough idea)
1423 #if BITS_PER_LONG == 32
1424 #define VMALLOC_SPACE (128UL*1024*1024)
1426 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1429 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1430 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1431 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1432 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1433 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1434 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1435 #define VMAP_BBMAP_BITS \
1436 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1437 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1438 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1440 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1442 struct vmap_block_queue
{
1444 struct list_head free
;
1449 struct vmap_area
*va
;
1450 unsigned long free
, dirty
;
1451 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1452 struct list_head free_list
;
1453 struct rcu_head rcu_head
;
1454 struct list_head purge
;
1457 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1458 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1461 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1462 * in the free path. Could get rid of this if we change the API to return a
1463 * "cookie" from alloc, to be passed to free. But no big deal yet.
1465 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
1466 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
1469 * We should probably have a fallback mechanism to allocate virtual memory
1470 * out of partially filled vmap blocks. However vmap block sizing should be
1471 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1475 static unsigned long addr_to_vb_idx(unsigned long addr
)
1477 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1478 addr
/= VMAP_BLOCK_SIZE
;
1482 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1486 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1487 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1488 return (void *)addr
;
1492 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1493 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1494 * @order: how many 2^order pages should be occupied in newly allocated block
1495 * @gfp_mask: flags for the page level allocator
1497 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1499 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1501 struct vmap_block_queue
*vbq
;
1502 struct vmap_block
*vb
;
1503 struct vmap_area
*va
;
1504 unsigned long vb_idx
;
1508 node
= numa_node_id();
1510 vb
= kmalloc_node(sizeof(struct vmap_block
),
1511 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1513 return ERR_PTR(-ENOMEM
);
1515 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1516 VMALLOC_START
, VMALLOC_END
,
1520 return ERR_CAST(va
);
1523 err
= radix_tree_preload(gfp_mask
);
1524 if (unlikely(err
)) {
1527 return ERR_PTR(err
);
1530 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1531 spin_lock_init(&vb
->lock
);
1533 /* At least something should be left free */
1534 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1535 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
1537 vb
->dirty_min
= VMAP_BBMAP_BITS
;
1539 INIT_LIST_HEAD(&vb
->free_list
);
1541 vb_idx
= addr_to_vb_idx(va
->va_start
);
1542 spin_lock(&vmap_block_tree_lock
);
1543 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
1544 spin_unlock(&vmap_block_tree_lock
);
1546 radix_tree_preload_end();
1548 vbq
= &get_cpu_var(vmap_block_queue
);
1549 spin_lock(&vbq
->lock
);
1550 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
1551 spin_unlock(&vbq
->lock
);
1552 put_cpu_var(vmap_block_queue
);
1557 static void free_vmap_block(struct vmap_block
*vb
)
1559 struct vmap_block
*tmp
;
1560 unsigned long vb_idx
;
1562 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
1563 spin_lock(&vmap_block_tree_lock
);
1564 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
1565 spin_unlock(&vmap_block_tree_lock
);
1568 free_vmap_area_noflush(vb
->va
);
1569 kfree_rcu(vb
, rcu_head
);
1572 static void purge_fragmented_blocks(int cpu
)
1575 struct vmap_block
*vb
;
1576 struct vmap_block
*n_vb
;
1577 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1580 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1582 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
1585 spin_lock(&vb
->lock
);
1586 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
1587 vb
->free
= 0; /* prevent further allocs after releasing lock */
1588 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
1590 vb
->dirty_max
= VMAP_BBMAP_BITS
;
1591 spin_lock(&vbq
->lock
);
1592 list_del_rcu(&vb
->free_list
);
1593 spin_unlock(&vbq
->lock
);
1594 spin_unlock(&vb
->lock
);
1595 list_add_tail(&vb
->purge
, &purge
);
1597 spin_unlock(&vb
->lock
);
1601 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
1602 list_del(&vb
->purge
);
1603 free_vmap_block(vb
);
1607 static void purge_fragmented_blocks_allcpus(void)
1611 for_each_possible_cpu(cpu
)
1612 purge_fragmented_blocks(cpu
);
1615 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
1617 struct vmap_block_queue
*vbq
;
1618 struct vmap_block
*vb
;
1622 BUG_ON(offset_in_page(size
));
1623 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1624 if (WARN_ON(size
== 0)) {
1626 * Allocating 0 bytes isn't what caller wants since
1627 * get_order(0) returns funny result. Just warn and terminate
1632 order
= get_order(size
);
1635 vbq
= &get_cpu_var(vmap_block_queue
);
1636 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1637 unsigned long pages_off
;
1639 spin_lock(&vb
->lock
);
1640 if (vb
->free
< (1UL << order
)) {
1641 spin_unlock(&vb
->lock
);
1645 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1646 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1647 vb
->free
-= 1UL << order
;
1648 if (vb
->free
== 0) {
1649 spin_lock(&vbq
->lock
);
1650 list_del_rcu(&vb
->free_list
);
1651 spin_unlock(&vbq
->lock
);
1654 spin_unlock(&vb
->lock
);
1658 put_cpu_var(vmap_block_queue
);
1661 /* Allocate new block if nothing was found */
1663 vaddr
= new_vmap_block(order
, gfp_mask
);
1668 static void vb_free(const void *addr
, unsigned long size
)
1670 unsigned long offset
;
1671 unsigned long vb_idx
;
1673 struct vmap_block
*vb
;
1675 BUG_ON(offset_in_page(size
));
1676 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1678 flush_cache_vunmap((unsigned long)addr
, (unsigned long)addr
+ size
);
1680 order
= get_order(size
);
1682 offset
= (unsigned long)addr
& (VMAP_BLOCK_SIZE
- 1);
1683 offset
>>= PAGE_SHIFT
;
1685 vb_idx
= addr_to_vb_idx((unsigned long)addr
);
1687 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1691 vunmap_page_range((unsigned long)addr
, (unsigned long)addr
+ size
);
1693 if (debug_pagealloc_enabled_static())
1694 flush_tlb_kernel_range((unsigned long)addr
,
1695 (unsigned long)addr
+ size
);
1697 spin_lock(&vb
->lock
);
1699 /* Expand dirty range */
1700 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1701 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1703 vb
->dirty
+= 1UL << order
;
1704 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1706 spin_unlock(&vb
->lock
);
1707 free_vmap_block(vb
);
1709 spin_unlock(&vb
->lock
);
1712 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
1716 if (unlikely(!vmap_initialized
))
1721 for_each_possible_cpu(cpu
) {
1722 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1723 struct vmap_block
*vb
;
1726 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1727 spin_lock(&vb
->lock
);
1729 unsigned long va_start
= vb
->va
->va_start
;
1732 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1733 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1735 start
= min(s
, start
);
1740 spin_unlock(&vb
->lock
);
1745 mutex_lock(&vmap_purge_lock
);
1746 purge_fragmented_blocks_allcpus();
1747 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1748 flush_tlb_kernel_range(start
, end
);
1749 mutex_unlock(&vmap_purge_lock
);
1753 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1755 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1756 * to amortize TLB flushing overheads. What this means is that any page you
1757 * have now, may, in a former life, have been mapped into kernel virtual
1758 * address by the vmap layer and so there might be some CPUs with TLB entries
1759 * still referencing that page (additional to the regular 1:1 kernel mapping).
1761 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1762 * be sure that none of the pages we have control over will have any aliases
1763 * from the vmap layer.
1765 void vm_unmap_aliases(void)
1767 unsigned long start
= ULONG_MAX
, end
= 0;
1770 _vm_unmap_aliases(start
, end
, flush
);
1772 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1775 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1776 * @mem: the pointer returned by vm_map_ram
1777 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1779 void vm_unmap_ram(const void *mem
, unsigned int count
)
1781 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1782 unsigned long addr
= (unsigned long)mem
;
1783 struct vmap_area
*va
;
1787 BUG_ON(addr
< VMALLOC_START
);
1788 BUG_ON(addr
> VMALLOC_END
);
1789 BUG_ON(!PAGE_ALIGNED(addr
));
1791 kasan_poison_vmalloc(mem
, size
);
1793 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1794 debug_check_no_locks_freed(mem
, size
);
1799 va
= find_vmap_area(addr
);
1801 debug_check_no_locks_freed((void *)va
->va_start
,
1802 (va
->va_end
- va
->va_start
));
1803 free_unmap_vmap_area(va
);
1805 EXPORT_SYMBOL(vm_unmap_ram
);
1808 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1809 * @pages: an array of pointers to the pages to be mapped
1810 * @count: number of pages
1811 * @node: prefer to allocate data structures on this node
1812 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1814 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1815 * faster than vmap so it's good. But if you mix long-life and short-life
1816 * objects with vm_map_ram(), it could consume lots of address space through
1817 * fragmentation (especially on a 32bit machine). You could see failures in
1818 * the end. Please use this function for short-lived objects.
1820 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1822 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
, pgprot_t prot
)
1824 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1828 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1829 mem
= vb_alloc(size
, GFP_KERNEL
);
1832 addr
= (unsigned long)mem
;
1834 struct vmap_area
*va
;
1835 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1836 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1840 addr
= va
->va_start
;
1844 kasan_unpoison_vmalloc(mem
, size
);
1846 if (vmap_page_range(addr
, addr
+ size
, prot
, pages
) < 0) {
1847 vm_unmap_ram(mem
, count
);
1852 EXPORT_SYMBOL(vm_map_ram
);
1854 static struct vm_struct
*vmlist __initdata
;
1857 * vm_area_add_early - add vmap area early during boot
1858 * @vm: vm_struct to add
1860 * This function is used to add fixed kernel vm area to vmlist before
1861 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1862 * should contain proper values and the other fields should be zero.
1864 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1866 void __init
vm_area_add_early(struct vm_struct
*vm
)
1868 struct vm_struct
*tmp
, **p
;
1870 BUG_ON(vmap_initialized
);
1871 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1872 if (tmp
->addr
>= vm
->addr
) {
1873 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1876 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1883 * vm_area_register_early - register vmap area early during boot
1884 * @vm: vm_struct to register
1885 * @align: requested alignment
1887 * This function is used to register kernel vm area before
1888 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1889 * proper values on entry and other fields should be zero. On return,
1890 * vm->addr contains the allocated address.
1892 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1894 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1896 static size_t vm_init_off __initdata
;
1899 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1900 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1902 vm
->addr
= (void *)addr
;
1904 vm_area_add_early(vm
);
1907 static void vmap_init_free_space(void)
1909 unsigned long vmap_start
= 1;
1910 const unsigned long vmap_end
= ULONG_MAX
;
1911 struct vmap_area
*busy
, *free
;
1915 * -|-----|.....|-----|-----|-----|.....|-
1917 * |<--------------------------------->|
1919 list_for_each_entry(busy
, &vmap_area_list
, list
) {
1920 if (busy
->va_start
- vmap_start
> 0) {
1921 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1922 if (!WARN_ON_ONCE(!free
)) {
1923 free
->va_start
= vmap_start
;
1924 free
->va_end
= busy
->va_start
;
1926 insert_vmap_area_augment(free
, NULL
,
1927 &free_vmap_area_root
,
1928 &free_vmap_area_list
);
1932 vmap_start
= busy
->va_end
;
1935 if (vmap_end
- vmap_start
> 0) {
1936 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1937 if (!WARN_ON_ONCE(!free
)) {
1938 free
->va_start
= vmap_start
;
1939 free
->va_end
= vmap_end
;
1941 insert_vmap_area_augment(free
, NULL
,
1942 &free_vmap_area_root
,
1943 &free_vmap_area_list
);
1948 void __init
vmalloc_init(void)
1950 struct vmap_area
*va
;
1951 struct vm_struct
*tmp
;
1955 * Create the cache for vmap_area objects.
1957 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
1959 for_each_possible_cpu(i
) {
1960 struct vmap_block_queue
*vbq
;
1961 struct vfree_deferred
*p
;
1963 vbq
= &per_cpu(vmap_block_queue
, i
);
1964 spin_lock_init(&vbq
->lock
);
1965 INIT_LIST_HEAD(&vbq
->free
);
1966 p
= &per_cpu(vfree_deferred
, i
);
1967 init_llist_head(&p
->list
);
1968 INIT_WORK(&p
->wq
, free_work
);
1971 /* Import existing vmlist entries. */
1972 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1973 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1974 if (WARN_ON_ONCE(!va
))
1977 va
->va_start
= (unsigned long)tmp
->addr
;
1978 va
->va_end
= va
->va_start
+ tmp
->size
;
1980 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1984 * Now we can initialize a free vmap space.
1986 vmap_init_free_space();
1987 vmap_initialized
= true;
1991 * map_kernel_range_noflush - map kernel VM area with the specified pages
1992 * @addr: start of the VM area to map
1993 * @size: size of the VM area to map
1994 * @prot: page protection flags to use
1995 * @pages: pages to map
1997 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1998 * specify should have been allocated using get_vm_area() and its
2002 * This function does NOT do any cache flushing. The caller is
2003 * responsible for calling flush_cache_vmap() on to-be-mapped areas
2004 * before calling this function.
2007 * The number of pages mapped on success, -errno on failure.
2009 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
2010 pgprot_t prot
, struct page
**pages
)
2012 return vmap_page_range_noflush(addr
, addr
+ size
, prot
, pages
);
2016 * unmap_kernel_range_noflush - unmap kernel VM area
2017 * @addr: start of the VM area to unmap
2018 * @size: size of the VM area to unmap
2020 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
2021 * specify should have been allocated using get_vm_area() and its
2025 * This function does NOT do any cache flushing. The caller is
2026 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
2027 * before calling this function and flush_tlb_kernel_range() after.
2029 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
2031 vunmap_page_range(addr
, addr
+ size
);
2033 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush
);
2036 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2037 * @addr: start of the VM area to unmap
2038 * @size: size of the VM area to unmap
2040 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2041 * the unmapping and tlb after.
2043 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
2045 unsigned long end
= addr
+ size
;
2047 flush_cache_vunmap(addr
, end
);
2048 vunmap_page_range(addr
, end
);
2049 flush_tlb_kernel_range(addr
, end
);
2051 EXPORT_SYMBOL_GPL(unmap_kernel_range
);
2053 int map_vm_area(struct vm_struct
*area
, pgprot_t prot
, struct page
**pages
)
2055 unsigned long addr
= (unsigned long)area
->addr
;
2056 unsigned long end
= addr
+ get_vm_area_size(area
);
2059 err
= vmap_page_range(addr
, end
, prot
, pages
);
2061 return err
> 0 ? 0 : err
;
2063 EXPORT_SYMBOL_GPL(map_vm_area
);
2065 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2066 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2069 vm
->addr
= (void *)va
->va_start
;
2070 vm
->size
= va
->va_end
- va
->va_start
;
2071 vm
->caller
= caller
;
2075 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2076 unsigned long flags
, const void *caller
)
2078 spin_lock(&vmap_area_lock
);
2079 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2080 spin_unlock(&vmap_area_lock
);
2083 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2086 * Before removing VM_UNINITIALIZED,
2087 * we should make sure that vm has proper values.
2088 * Pair with smp_rmb() in show_numa_info().
2091 vm
->flags
&= ~VM_UNINITIALIZED
;
2094 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2095 unsigned long align
, unsigned long flags
, unsigned long start
,
2096 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
2098 struct vmap_area
*va
;
2099 struct vm_struct
*area
;
2100 unsigned long requested_size
= size
;
2102 BUG_ON(in_interrupt());
2103 size
= PAGE_ALIGN(size
);
2104 if (unlikely(!size
))
2107 if (flags
& VM_IOREMAP
)
2108 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2109 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2111 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2112 if (unlikely(!area
))
2115 if (!(flags
& VM_NO_GUARD
))
2118 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
2124 kasan_unpoison_vmalloc((void *)va
->va_start
, requested_size
);
2126 setup_vmalloc_vm(area
, va
, flags
, caller
);
2131 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2132 unsigned long start
, unsigned long end
,
2135 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2136 GFP_KERNEL
, caller
);
2140 * get_vm_area - reserve a contiguous kernel virtual area
2141 * @size: size of the area
2142 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2144 * Search an area of @size in the kernel virtual mapping area,
2145 * and reserved it for out purposes. Returns the area descriptor
2146 * on success or %NULL on failure.
2148 * Return: the area descriptor on success or %NULL on failure.
2150 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2152 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2153 NUMA_NO_NODE
, GFP_KERNEL
,
2154 __builtin_return_address(0));
2157 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2160 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2161 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2165 * find_vm_area - find a continuous kernel virtual area
2166 * @addr: base address
2168 * Search for the kernel VM area starting at @addr, and return it.
2169 * It is up to the caller to do all required locking to keep the returned
2172 * Return: pointer to the found area or %NULL on faulure
2174 struct vm_struct
*find_vm_area(const void *addr
)
2176 struct vmap_area
*va
;
2178 va
= find_vmap_area((unsigned long)addr
);
2186 * remove_vm_area - find and remove a continuous kernel virtual area
2187 * @addr: base address
2189 * Search for the kernel VM area starting at @addr, and remove it.
2190 * This function returns the found VM area, but using it is NOT safe
2191 * on SMP machines, except for its size or flags.
2193 * Return: pointer to the found area or %NULL on faulure
2195 struct vm_struct
*remove_vm_area(const void *addr
)
2197 struct vmap_area
*va
;
2201 spin_lock(&vmap_area_lock
);
2202 va
= __find_vmap_area((unsigned long)addr
);
2204 struct vm_struct
*vm
= va
->vm
;
2207 spin_unlock(&vmap_area_lock
);
2209 kasan_free_shadow(vm
);
2210 free_unmap_vmap_area(va
);
2215 spin_unlock(&vmap_area_lock
);
2219 static inline void set_area_direct_map(const struct vm_struct
*area
,
2220 int (*set_direct_map
)(struct page
*page
))
2224 for (i
= 0; i
< area
->nr_pages
; i
++)
2225 if (page_address(area
->pages
[i
]))
2226 set_direct_map(area
->pages
[i
]);
2229 /* Handle removing and resetting vm mappings related to the vm_struct. */
2230 static void vm_remove_mappings(struct vm_struct
*area
, int deallocate_pages
)
2232 unsigned long start
= ULONG_MAX
, end
= 0;
2233 int flush_reset
= area
->flags
& VM_FLUSH_RESET_PERMS
;
2237 remove_vm_area(area
->addr
);
2239 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2244 * If not deallocating pages, just do the flush of the VM area and
2247 if (!deallocate_pages
) {
2253 * If execution gets here, flush the vm mapping and reset the direct
2254 * map. Find the start and end range of the direct mappings to make sure
2255 * the vm_unmap_aliases() flush includes the direct map.
2257 for (i
= 0; i
< area
->nr_pages
; i
++) {
2258 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2260 start
= min(addr
, start
);
2261 end
= max(addr
+ PAGE_SIZE
, end
);
2267 * Set direct map to something invalid so that it won't be cached if
2268 * there are any accesses after the TLB flush, then flush the TLB and
2269 * reset the direct map permissions to the default.
2271 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2272 _vm_unmap_aliases(start
, end
, flush_dmap
);
2273 set_area_direct_map(area
, set_direct_map_default_noflush
);
2276 static void __vunmap(const void *addr
, int deallocate_pages
)
2278 struct vm_struct
*area
;
2283 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2287 area
= find_vm_area(addr
);
2288 if (unlikely(!area
)) {
2289 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2294 debug_check_no_locks_freed(area
->addr
, get_vm_area_size(area
));
2295 debug_check_no_obj_freed(area
->addr
, get_vm_area_size(area
));
2297 kasan_poison_vmalloc(area
->addr
, area
->size
);
2299 vm_remove_mappings(area
, deallocate_pages
);
2301 if (deallocate_pages
) {
2304 for (i
= 0; i
< area
->nr_pages
; i
++) {
2305 struct page
*page
= area
->pages
[i
];
2308 __free_pages(page
, 0);
2310 atomic_long_sub(area
->nr_pages
, &nr_vmalloc_pages
);
2312 kvfree(area
->pages
);
2319 static inline void __vfree_deferred(const void *addr
)
2322 * Use raw_cpu_ptr() because this can be called from preemptible
2323 * context. Preemption is absolutely fine here, because the llist_add()
2324 * implementation is lockless, so it works even if we are adding to
2325 * nother cpu's list. schedule_work() should be fine with this too.
2327 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2329 if (llist_add((struct llist_node
*)addr
, &p
->list
))
2330 schedule_work(&p
->wq
);
2334 * vfree_atomic - release memory allocated by vmalloc()
2335 * @addr: memory base address
2337 * This one is just like vfree() but can be called in any atomic context
2340 void vfree_atomic(const void *addr
)
2344 kmemleak_free(addr
);
2348 __vfree_deferred(addr
);
2351 static void __vfree(const void *addr
)
2353 if (unlikely(in_interrupt()))
2354 __vfree_deferred(addr
);
2360 * vfree - release memory allocated by vmalloc()
2361 * @addr: memory base address
2363 * Free the virtually continuous memory area starting at @addr, as
2364 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2365 * NULL, no operation is performed.
2367 * Must not be called in NMI context (strictly speaking, only if we don't
2368 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2369 * conventions for vfree() arch-depenedent would be a really bad idea)
2371 * May sleep if called *not* from interrupt context.
2373 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2375 void vfree(const void *addr
)
2379 kmemleak_free(addr
);
2381 might_sleep_if(!in_interrupt());
2388 EXPORT_SYMBOL(vfree
);
2391 * vunmap - release virtual mapping obtained by vmap()
2392 * @addr: memory base address
2394 * Free the virtually contiguous memory area starting at @addr,
2395 * which was created from the page array passed to vmap().
2397 * Must not be called in interrupt context.
2399 void vunmap(const void *addr
)
2401 BUG_ON(in_interrupt());
2406 EXPORT_SYMBOL(vunmap
);
2409 * vmap - map an array of pages into virtually contiguous space
2410 * @pages: array of page pointers
2411 * @count: number of pages to map
2412 * @flags: vm_area->flags
2413 * @prot: page protection for the mapping
2415 * Maps @count pages from @pages into contiguous kernel virtual
2418 * Return: the address of the area or %NULL on failure
2420 void *vmap(struct page
**pages
, unsigned int count
,
2421 unsigned long flags
, pgprot_t prot
)
2423 struct vm_struct
*area
;
2424 unsigned long size
; /* In bytes */
2428 if (count
> totalram_pages())
2431 size
= (unsigned long)count
<< PAGE_SHIFT
;
2432 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2436 if (map_vm_area(area
, prot
, pages
)) {
2443 EXPORT_SYMBOL(vmap
);
2445 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
2446 gfp_t gfp_mask
, pgprot_t prot
,
2447 int node
, const void *caller
);
2448 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2449 pgprot_t prot
, int node
)
2451 struct page
**pages
;
2452 unsigned int nr_pages
, array_size
, i
;
2453 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2454 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
2455 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
2459 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
2460 array_size
= (nr_pages
* sizeof(struct page
*));
2462 /* Please note that the recursion is strictly bounded. */
2463 if (array_size
> PAGE_SIZE
) {
2464 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
2465 PAGE_KERNEL
, node
, area
->caller
);
2467 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
2471 remove_vm_area(area
->addr
);
2476 area
->pages
= pages
;
2477 area
->nr_pages
= nr_pages
;
2479 for (i
= 0; i
< area
->nr_pages
; i
++) {
2482 if (node
== NUMA_NO_NODE
)
2483 page
= alloc_page(alloc_mask
|highmem_mask
);
2485 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
2487 if (unlikely(!page
)) {
2488 /* Successfully allocated i pages, free them in __vunmap() */
2490 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2493 area
->pages
[i
] = page
;
2494 if (gfpflags_allow_blocking(gfp_mask
))
2497 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2499 if (map_vm_area(area
, prot
, pages
))
2504 warn_alloc(gfp_mask
, NULL
,
2505 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2506 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
2507 __vfree(area
->addr
);
2512 * __vmalloc_node_range - allocate virtually contiguous memory
2513 * @size: allocation size
2514 * @align: desired alignment
2515 * @start: vm area range start
2516 * @end: vm area range end
2517 * @gfp_mask: flags for the page level allocator
2518 * @prot: protection mask for the allocated pages
2519 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2520 * @node: node to use for allocation or NUMA_NO_NODE
2521 * @caller: caller's return address
2523 * Allocate enough pages to cover @size from the page level
2524 * allocator with @gfp_mask flags. Map them into contiguous
2525 * kernel virtual space, using a pagetable protection of @prot.
2527 * Return: the address of the area or %NULL on failure
2529 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
2530 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
2531 pgprot_t prot
, unsigned long vm_flags
, int node
,
2534 struct vm_struct
*area
;
2536 unsigned long real_size
= size
;
2538 size
= PAGE_ALIGN(size
);
2539 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages())
2542 area
= __get_vm_area_node(real_size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
2543 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
2547 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
2552 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2553 * flag. It means that vm_struct is not fully initialized.
2554 * Now, it is fully initialized, so remove this flag here.
2556 clear_vm_uninitialized_flag(area
);
2558 kmemleak_vmalloc(area
, size
, gfp_mask
);
2563 warn_alloc(gfp_mask
, NULL
,
2564 "vmalloc: allocation failure: %lu bytes", real_size
);
2569 * This is only for performance analysis of vmalloc and stress purpose.
2570 * It is required by vmalloc test module, therefore do not use it other
2573 #ifdef CONFIG_TEST_VMALLOC_MODULE
2574 EXPORT_SYMBOL_GPL(__vmalloc_node_range
);
2578 * __vmalloc_node - allocate virtually contiguous memory
2579 * @size: allocation size
2580 * @align: desired alignment
2581 * @gfp_mask: flags for the page level allocator
2582 * @prot: protection mask for the allocated pages
2583 * @node: node to use for allocation or NUMA_NO_NODE
2584 * @caller: caller's return address
2586 * Allocate enough pages to cover @size from the page level
2587 * allocator with @gfp_mask flags. Map them into contiguous
2588 * kernel virtual space, using a pagetable protection of @prot.
2590 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2591 * and __GFP_NOFAIL are not supported
2593 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2596 * Return: pointer to the allocated memory or %NULL on error
2598 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
2599 gfp_t gfp_mask
, pgprot_t prot
,
2600 int node
, const void *caller
)
2602 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
2603 gfp_mask
, prot
, 0, node
, caller
);
2606 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
, pgprot_t prot
)
2608 return __vmalloc_node(size
, 1, gfp_mask
, prot
, NUMA_NO_NODE
,
2609 __builtin_return_address(0));
2611 EXPORT_SYMBOL(__vmalloc
);
2613 static inline void *__vmalloc_node_flags(unsigned long size
,
2614 int node
, gfp_t flags
)
2616 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
,
2617 node
, __builtin_return_address(0));
2621 void *__vmalloc_node_flags_caller(unsigned long size
, int node
, gfp_t flags
,
2624 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
, node
, caller
);
2628 * vmalloc - allocate virtually contiguous memory
2629 * @size: allocation size
2631 * Allocate enough pages to cover @size from the page level
2632 * allocator and map them into contiguous kernel virtual space.
2634 * For tight control over page level allocator and protection flags
2635 * use __vmalloc() instead.
2637 * Return: pointer to the allocated memory or %NULL on error
2639 void *vmalloc(unsigned long size
)
2641 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
2644 EXPORT_SYMBOL(vmalloc
);
2647 * vzalloc - allocate virtually contiguous memory with zero fill
2648 * @size: allocation size
2650 * Allocate enough pages to cover @size from the page level
2651 * allocator and map them into contiguous kernel virtual space.
2652 * The memory allocated is set to zero.
2654 * For tight control over page level allocator and protection flags
2655 * use __vmalloc() instead.
2657 * Return: pointer to the allocated memory or %NULL on error
2659 void *vzalloc(unsigned long size
)
2661 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
2662 GFP_KERNEL
| __GFP_ZERO
);
2664 EXPORT_SYMBOL(vzalloc
);
2667 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2668 * @size: allocation size
2670 * The resulting memory area is zeroed so it can be mapped to userspace
2671 * without leaking data.
2673 * Return: pointer to the allocated memory or %NULL on error
2675 void *vmalloc_user(unsigned long size
)
2677 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2678 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
2679 VM_USERMAP
, NUMA_NO_NODE
,
2680 __builtin_return_address(0));
2682 EXPORT_SYMBOL(vmalloc_user
);
2685 * vmalloc_node - allocate memory on a specific node
2686 * @size: allocation size
2689 * Allocate enough pages to cover @size from the page level
2690 * allocator and map them into contiguous kernel virtual space.
2692 * For tight control over page level allocator and protection flags
2693 * use __vmalloc() instead.
2695 * Return: pointer to the allocated memory or %NULL on error
2697 void *vmalloc_node(unsigned long size
, int node
)
2699 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL
,
2700 node
, __builtin_return_address(0));
2702 EXPORT_SYMBOL(vmalloc_node
);
2705 * vzalloc_node - allocate memory on a specific node with zero fill
2706 * @size: allocation size
2709 * Allocate enough pages to cover @size from the page level
2710 * allocator and map them into contiguous kernel virtual space.
2711 * The memory allocated is set to zero.
2713 * For tight control over page level allocator and protection flags
2714 * use __vmalloc_node() instead.
2716 * Return: pointer to the allocated memory or %NULL on error
2718 void *vzalloc_node(unsigned long size
, int node
)
2720 return __vmalloc_node_flags(size
, node
,
2721 GFP_KERNEL
| __GFP_ZERO
);
2723 EXPORT_SYMBOL(vzalloc_node
);
2726 * vmalloc_user_node_flags - allocate memory for userspace on a specific node
2727 * @size: allocation size
2729 * @flags: flags for the page level allocator
2731 * The resulting memory area is zeroed so it can be mapped to userspace
2732 * without leaking data.
2734 * Return: pointer to the allocated memory or %NULL on error
2736 void *vmalloc_user_node_flags(unsigned long size
, int node
, gfp_t flags
)
2738 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2739 flags
| __GFP_ZERO
, PAGE_KERNEL
,
2741 __builtin_return_address(0));
2743 EXPORT_SYMBOL(vmalloc_user_node_flags
);
2746 * vmalloc_exec - allocate virtually contiguous, executable memory
2747 * @size: allocation size
2749 * Kernel-internal function to allocate enough pages to cover @size
2750 * the page level allocator and map them into contiguous and
2751 * executable kernel virtual space.
2753 * For tight control over page level allocator and protection flags
2754 * use __vmalloc() instead.
2756 * Return: pointer to the allocated memory or %NULL on error
2758 void *vmalloc_exec(unsigned long size
)
2760 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
2761 GFP_KERNEL
, PAGE_KERNEL_EXEC
, VM_FLUSH_RESET_PERMS
,
2762 NUMA_NO_NODE
, __builtin_return_address(0));
2765 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2766 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2767 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2768 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2771 * 64b systems should always have either DMA or DMA32 zones. For others
2772 * GFP_DMA32 should do the right thing and use the normal zone.
2774 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2778 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2779 * @size: allocation size
2781 * Allocate enough 32bit PA addressable pages to cover @size from the
2782 * page level allocator and map them into contiguous kernel virtual space.
2784 * Return: pointer to the allocated memory or %NULL on error
2786 void *vmalloc_32(unsigned long size
)
2788 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, PAGE_KERNEL
,
2789 NUMA_NO_NODE
, __builtin_return_address(0));
2791 EXPORT_SYMBOL(vmalloc_32
);
2794 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2795 * @size: allocation size
2797 * The resulting memory area is 32bit addressable and zeroed so it can be
2798 * mapped to userspace without leaking data.
2800 * Return: pointer to the allocated memory or %NULL on error
2802 void *vmalloc_32_user(unsigned long size
)
2804 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2805 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2806 VM_USERMAP
, NUMA_NO_NODE
,
2807 __builtin_return_address(0));
2809 EXPORT_SYMBOL(vmalloc_32_user
);
2812 * small helper routine , copy contents to buf from addr.
2813 * If the page is not present, fill zero.
2816 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2822 unsigned long offset
, length
;
2824 offset
= offset_in_page(addr
);
2825 length
= PAGE_SIZE
- offset
;
2828 p
= vmalloc_to_page(addr
);
2830 * To do safe access to this _mapped_ area, we need
2831 * lock. But adding lock here means that we need to add
2832 * overhead of vmalloc()/vfree() calles for this _debug_
2833 * interface, rarely used. Instead of that, we'll use
2834 * kmap() and get small overhead in this access function.
2838 * we can expect USER0 is not used (see vread/vwrite's
2839 * function description)
2841 void *map
= kmap_atomic(p
);
2842 memcpy(buf
, map
+ offset
, length
);
2845 memset(buf
, 0, length
);
2855 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2861 unsigned long offset
, length
;
2863 offset
= offset_in_page(addr
);
2864 length
= PAGE_SIZE
- offset
;
2867 p
= vmalloc_to_page(addr
);
2869 * To do safe access to this _mapped_ area, we need
2870 * lock. But adding lock here means that we need to add
2871 * overhead of vmalloc()/vfree() calles for this _debug_
2872 * interface, rarely used. Instead of that, we'll use
2873 * kmap() and get small overhead in this access function.
2877 * we can expect USER0 is not used (see vread/vwrite's
2878 * function description)
2880 void *map
= kmap_atomic(p
);
2881 memcpy(map
+ offset
, buf
, length
);
2893 * vread() - read vmalloc area in a safe way.
2894 * @buf: buffer for reading data
2895 * @addr: vm address.
2896 * @count: number of bytes to be read.
2898 * This function checks that addr is a valid vmalloc'ed area, and
2899 * copy data from that area to a given buffer. If the given memory range
2900 * of [addr...addr+count) includes some valid address, data is copied to
2901 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2902 * IOREMAP area is treated as memory hole and no copy is done.
2904 * If [addr...addr+count) doesn't includes any intersects with alive
2905 * vm_struct area, returns 0. @buf should be kernel's buffer.
2907 * Note: In usual ops, vread() is never necessary because the caller
2908 * should know vmalloc() area is valid and can use memcpy().
2909 * This is for routines which have to access vmalloc area without
2910 * any information, as /dev/kmem.
2912 * Return: number of bytes for which addr and buf should be increased
2913 * (same number as @count) or %0 if [addr...addr+count) doesn't
2914 * include any intersection with valid vmalloc area
2916 long vread(char *buf
, char *addr
, unsigned long count
)
2918 struct vmap_area
*va
;
2919 struct vm_struct
*vm
;
2920 char *vaddr
, *buf_start
= buf
;
2921 unsigned long buflen
= count
;
2924 /* Don't allow overflow */
2925 if ((unsigned long) addr
+ count
< count
)
2926 count
= -(unsigned long) addr
;
2928 spin_lock(&vmap_area_lock
);
2929 list_for_each_entry(va
, &vmap_area_list
, list
) {
2937 vaddr
= (char *) vm
->addr
;
2938 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2940 while (addr
< vaddr
) {
2948 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2951 if (!(vm
->flags
& VM_IOREMAP
))
2952 aligned_vread(buf
, addr
, n
);
2953 else /* IOREMAP area is treated as memory hole */
2960 spin_unlock(&vmap_area_lock
);
2962 if (buf
== buf_start
)
2964 /* zero-fill memory holes */
2965 if (buf
!= buf_start
+ buflen
)
2966 memset(buf
, 0, buflen
- (buf
- buf_start
));
2972 * vwrite() - write vmalloc area in a safe way.
2973 * @buf: buffer for source data
2974 * @addr: vm address.
2975 * @count: number of bytes to be read.
2977 * This function checks that addr is a valid vmalloc'ed area, and
2978 * copy data from a buffer to the given addr. If specified range of
2979 * [addr...addr+count) includes some valid address, data is copied from
2980 * proper area of @buf. If there are memory holes, no copy to hole.
2981 * IOREMAP area is treated as memory hole and no copy is done.
2983 * If [addr...addr+count) doesn't includes any intersects with alive
2984 * vm_struct area, returns 0. @buf should be kernel's buffer.
2986 * Note: In usual ops, vwrite() is never necessary because the caller
2987 * should know vmalloc() area is valid and can use memcpy().
2988 * This is for routines which have to access vmalloc area without
2989 * any information, as /dev/kmem.
2991 * Return: number of bytes for which addr and buf should be
2992 * increased (same number as @count) or %0 if [addr...addr+count)
2993 * doesn't include any intersection with valid vmalloc area
2995 long vwrite(char *buf
, char *addr
, unsigned long count
)
2997 struct vmap_area
*va
;
2998 struct vm_struct
*vm
;
3000 unsigned long n
, buflen
;
3003 /* Don't allow overflow */
3004 if ((unsigned long) addr
+ count
< count
)
3005 count
= -(unsigned long) addr
;
3008 spin_lock(&vmap_area_lock
);
3009 list_for_each_entry(va
, &vmap_area_list
, list
) {
3017 vaddr
= (char *) vm
->addr
;
3018 if (addr
>= vaddr
+ get_vm_area_size(vm
))
3020 while (addr
< vaddr
) {
3027 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
3030 if (!(vm
->flags
& VM_IOREMAP
)) {
3031 aligned_vwrite(buf
, addr
, n
);
3039 spin_unlock(&vmap_area_lock
);
3046 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3047 * @vma: vma to cover
3048 * @uaddr: target user address to start at
3049 * @kaddr: virtual address of vmalloc kernel memory
3050 * @pgoff: offset from @kaddr to start at
3051 * @size: size of map area
3053 * Returns: 0 for success, -Exxx on failure
3055 * This function checks that @kaddr is a valid vmalloc'ed area,
3056 * and that it is big enough to cover the range starting at
3057 * @uaddr in @vma. Will return failure if that criteria isn't
3060 * Similar to remap_pfn_range() (see mm/memory.c)
3062 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3063 void *kaddr
, unsigned long pgoff
,
3066 struct vm_struct
*area
;
3068 unsigned long end_index
;
3070 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
3073 size
= PAGE_ALIGN(size
);
3075 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3078 area
= find_vm_area(kaddr
);
3082 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3085 if (check_add_overflow(size
, off
, &end_index
) ||
3086 end_index
> get_vm_area_size(area
))
3091 struct page
*page
= vmalloc_to_page(kaddr
);
3094 ret
= vm_insert_page(vma
, uaddr
, page
);
3103 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3107 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
3110 * remap_vmalloc_range - map vmalloc pages to userspace
3111 * @vma: vma to cover (map full range of vma)
3112 * @addr: vmalloc memory
3113 * @pgoff: number of pages into addr before first page to map
3115 * Returns: 0 for success, -Exxx on failure
3117 * This function checks that addr is a valid vmalloc'ed area, and
3118 * that it is big enough to cover the vma. Will return failure if
3119 * that criteria isn't met.
3121 * Similar to remap_pfn_range() (see mm/memory.c)
3123 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3124 unsigned long pgoff
)
3126 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3128 vma
->vm_end
- vma
->vm_start
);
3130 EXPORT_SYMBOL(remap_vmalloc_range
);
3133 * Implement stubs for vmalloc_sync_[un]mappings () if the architecture chose
3136 * The purpose of this function is to make sure the vmalloc area
3137 * mappings are identical in all page-tables in the system.
3139 void __weak
vmalloc_sync_mappings(void)
3143 void __weak
vmalloc_sync_unmappings(void)
3147 static int f(pte_t
*pte
, unsigned long addr
, void *data
)
3159 * alloc_vm_area - allocate a range of kernel address space
3160 * @size: size of the area
3161 * @ptes: returns the PTEs for the address space
3163 * Returns: NULL on failure, vm_struct on success
3165 * This function reserves a range of kernel address space, and
3166 * allocates pagetables to map that range. No actual mappings
3169 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3170 * allocated for the VM area are returned.
3172 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
3174 struct vm_struct
*area
;
3176 area
= get_vm_area_caller(size
, VM_IOREMAP
,
3177 __builtin_return_address(0));
3182 * This ensures that page tables are constructed for this region
3183 * of kernel virtual address space and mapped into init_mm.
3185 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3186 size
, f
, ptes
? &ptes
: NULL
)) {
3193 EXPORT_SYMBOL_GPL(alloc_vm_area
);
3195 void free_vm_area(struct vm_struct
*area
)
3197 struct vm_struct
*ret
;
3198 ret
= remove_vm_area(area
->addr
);
3199 BUG_ON(ret
!= area
);
3202 EXPORT_SYMBOL_GPL(free_vm_area
);
3205 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3207 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3211 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3212 * @addr: target address
3214 * Returns: vmap_area if it is found. If there is no such area
3215 * the first highest(reverse order) vmap_area is returned
3216 * i.e. va->va_start < addr && va->va_end < addr or NULL
3217 * if there are no any areas before @addr.
3219 static struct vmap_area
*
3220 pvm_find_va_enclose_addr(unsigned long addr
)
3222 struct vmap_area
*va
, *tmp
;
3225 n
= free_vmap_area_root
.rb_node
;
3229 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3230 if (tmp
->va_start
<= addr
) {
3232 if (tmp
->va_end
>= addr
)
3245 * pvm_determine_end_from_reverse - find the highest aligned address
3246 * of free block below VMALLOC_END
3248 * in - the VA we start the search(reverse order);
3249 * out - the VA with the highest aligned end address.
3251 * Returns: determined end address within vmap_area
3253 static unsigned long
3254 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3256 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3260 list_for_each_entry_from_reverse((*va
),
3261 &free_vmap_area_list
, list
) {
3262 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3263 if ((*va
)->va_start
< addr
)
3272 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3273 * @offsets: array containing offset of each area
3274 * @sizes: array containing size of each area
3275 * @nr_vms: the number of areas to allocate
3276 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3278 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3279 * vm_structs on success, %NULL on failure
3281 * Percpu allocator wants to use congruent vm areas so that it can
3282 * maintain the offsets among percpu areas. This function allocates
3283 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3284 * be scattered pretty far, distance between two areas easily going up
3285 * to gigabytes. To avoid interacting with regular vmallocs, these
3286 * areas are allocated from top.
3288 * Despite its complicated look, this allocator is rather simple. It
3289 * does everything top-down and scans free blocks from the end looking
3290 * for matching base. While scanning, if any of the areas do not fit the
3291 * base address is pulled down to fit the area. Scanning is repeated till
3292 * all the areas fit and then all necessary data structures are inserted
3293 * and the result is returned.
3295 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3296 const size_t *sizes
, int nr_vms
,
3299 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3300 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3301 struct vmap_area
**vas
, *va
;
3302 struct vm_struct
**vms
;
3303 int area
, area2
, last_area
, term_area
;
3304 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3305 bool purged
= false;
3308 /* verify parameters and allocate data structures */
3309 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3310 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3311 start
= offsets
[area
];
3312 end
= start
+ sizes
[area
];
3314 /* is everything aligned properly? */
3315 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3316 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3318 /* detect the area with the highest address */
3319 if (start
> offsets
[last_area
])
3322 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3323 unsigned long start2
= offsets
[area2
];
3324 unsigned long end2
= start2
+ sizes
[area2
];
3326 BUG_ON(start2
< end
&& start
< end2
);
3329 last_end
= offsets
[last_area
] + sizes
[last_area
];
3331 if (vmalloc_end
- vmalloc_start
< last_end
) {
3336 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3337 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3341 for (area
= 0; area
< nr_vms
; area
++) {
3342 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3343 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3344 if (!vas
[area
] || !vms
[area
])
3348 spin_lock(&free_vmap_area_lock
);
3350 /* start scanning - we scan from the top, begin with the last area */
3351 area
= term_area
= last_area
;
3352 start
= offsets
[area
];
3353 end
= start
+ sizes
[area
];
3355 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3356 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3360 * base might have underflowed, add last_end before
3363 if (base
+ last_end
< vmalloc_start
+ last_end
)
3367 * Fitting base has not been found.
3373 * If required width exceeds current VA block, move
3374 * base downwards and then recheck.
3376 if (base
+ end
> va
->va_end
) {
3377 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3383 * If this VA does not fit, move base downwards and recheck.
3385 if (base
+ start
< va
->va_start
) {
3386 va
= node_to_va(rb_prev(&va
->rb_node
));
3387 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3393 * This area fits, move on to the previous one. If
3394 * the previous one is the terminal one, we're done.
3396 area
= (area
+ nr_vms
- 1) % nr_vms
;
3397 if (area
== term_area
)
3400 start
= offsets
[area
];
3401 end
= start
+ sizes
[area
];
3402 va
= pvm_find_va_enclose_addr(base
+ end
);
3405 /* we've found a fitting base, insert all va's */
3406 for (area
= 0; area
< nr_vms
; area
++) {
3409 start
= base
+ offsets
[area
];
3412 va
= pvm_find_va_enclose_addr(start
);
3413 if (WARN_ON_ONCE(va
== NULL
))
3414 /* It is a BUG(), but trigger recovery instead. */
3417 type
= classify_va_fit_type(va
, start
, size
);
3418 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
3419 /* It is a BUG(), but trigger recovery instead. */
3422 ret
= adjust_va_to_fit_type(va
, start
, size
, type
);
3426 /* Allocated area. */
3428 va
->va_start
= start
;
3429 va
->va_end
= start
+ size
;
3432 spin_unlock(&free_vmap_area_lock
);
3434 /* populate the kasan shadow space */
3435 for (area
= 0; area
< nr_vms
; area
++) {
3436 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3437 goto err_free_shadow
;
3439 kasan_unpoison_vmalloc((void *)vas
[area
]->va_start
,
3443 /* insert all vm's */
3444 spin_lock(&vmap_area_lock
);
3445 for (area
= 0; area
< nr_vms
; area
++) {
3446 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3448 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3451 spin_unlock(&vmap_area_lock
);
3458 * Remove previously allocated areas. There is no
3459 * need in removing these areas from the busy tree,
3460 * because they are inserted only on the final step
3461 * and when pcpu_get_vm_areas() is success.
3464 orig_start
= vas
[area
]->va_start
;
3465 orig_end
= vas
[area
]->va_end
;
3466 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3467 &free_vmap_area_list
);
3468 kasan_release_vmalloc(orig_start
, orig_end
,
3469 va
->va_start
, va
->va_end
);
3474 spin_unlock(&free_vmap_area_lock
);
3476 purge_vmap_area_lazy();
3479 /* Before "retry", check if we recover. */
3480 for (area
= 0; area
< nr_vms
; area
++) {
3484 vas
[area
] = kmem_cache_zalloc(
3485 vmap_area_cachep
, GFP_KERNEL
);
3494 for (area
= 0; area
< nr_vms
; area
++) {
3496 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
3506 spin_lock(&free_vmap_area_lock
);
3508 * We release all the vmalloc shadows, even the ones for regions that
3509 * hadn't been successfully added. This relies on kasan_release_vmalloc
3510 * being able to tolerate this case.
3512 for (area
= 0; area
< nr_vms
; area
++) {
3513 orig_start
= vas
[area
]->va_start
;
3514 orig_end
= vas
[area
]->va_end
;
3515 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3516 &free_vmap_area_list
);
3517 kasan_release_vmalloc(orig_start
, orig_end
,
3518 va
->va_start
, va
->va_end
);
3522 spin_unlock(&free_vmap_area_lock
);
3529 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3530 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3531 * @nr_vms: the number of allocated areas
3533 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3535 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
3539 for (i
= 0; i
< nr_vms
; i
++)
3540 free_vm_area(vms
[i
]);
3543 #endif /* CONFIG_SMP */
3545 #ifdef CONFIG_PROC_FS
3546 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3547 __acquires(&vmap_purge_lock
)
3548 __acquires(&vmap_area_lock
)
3550 mutex_lock(&vmap_purge_lock
);
3551 spin_lock(&vmap_area_lock
);
3553 return seq_list_start(&vmap_area_list
, *pos
);
3556 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3558 return seq_list_next(p
, &vmap_area_list
, pos
);
3561 static void s_stop(struct seq_file
*m
, void *p
)
3562 __releases(&vmap_purge_lock
)
3563 __releases(&vmap_area_lock
)
3565 mutex_unlock(&vmap_purge_lock
);
3566 spin_unlock(&vmap_area_lock
);
3569 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
3571 if (IS_ENABLED(CONFIG_NUMA
)) {
3572 unsigned int nr
, *counters
= m
->private;
3577 if (v
->flags
& VM_UNINITIALIZED
)
3579 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3582 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
3584 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
3585 counters
[page_to_nid(v
->pages
[nr
])]++;
3587 for_each_node_state(nr
, N_HIGH_MEMORY
)
3589 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
3593 static void show_purge_info(struct seq_file
*m
)
3595 struct llist_node
*head
;
3596 struct vmap_area
*va
;
3598 head
= READ_ONCE(vmap_purge_list
.first
);
3602 llist_for_each_entry(va
, head
, purge_list
) {
3603 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3604 (void *)va
->va_start
, (void *)va
->va_end
,
3605 va
->va_end
- va
->va_start
);
3609 static int s_show(struct seq_file
*m
, void *p
)
3611 struct vmap_area
*va
;
3612 struct vm_struct
*v
;
3614 va
= list_entry(p
, struct vmap_area
, list
);
3617 * s_show can encounter race with remove_vm_area, !vm on behalf
3618 * of vmap area is being tear down or vm_map_ram allocation.
3621 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
3622 (void *)va
->va_start
, (void *)va
->va_end
,
3623 va
->va_end
- va
->va_start
);
3630 seq_printf(m
, "0x%pK-0x%pK %7ld",
3631 v
->addr
, v
->addr
+ v
->size
, v
->size
);
3634 seq_printf(m
, " %pS", v
->caller
);
3637 seq_printf(m
, " pages=%d", v
->nr_pages
);
3640 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
3642 if (v
->flags
& VM_IOREMAP
)
3643 seq_puts(m
, " ioremap");
3645 if (v
->flags
& VM_ALLOC
)
3646 seq_puts(m
, " vmalloc");
3648 if (v
->flags
& VM_MAP
)
3649 seq_puts(m
, " vmap");
3651 if (v
->flags
& VM_USERMAP
)
3652 seq_puts(m
, " user");
3654 if (v
->flags
& VM_DMA_COHERENT
)
3655 seq_puts(m
, " dma-coherent");
3657 if (is_vmalloc_addr(v
->pages
))
3658 seq_puts(m
, " vpages");
3660 show_numa_info(m
, v
);
3664 * As a final step, dump "unpurged" areas. Note,
3665 * that entire "/proc/vmallocinfo" output will not
3666 * be address sorted, because the purge list is not
3669 if (list_is_last(&va
->list
, &vmap_area_list
))
3675 static const struct seq_operations vmalloc_op
= {
3682 static int __init
proc_vmalloc_init(void)
3684 if (IS_ENABLED(CONFIG_NUMA
))
3685 proc_create_seq_private("vmallocinfo", 0400, NULL
,
3687 nr_node_ids
* sizeof(unsigned int), NULL
);
3689 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
3692 module_init(proc_vmalloc_init
);