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>
38 #include <linux/uaccess.h>
39 #include <asm/tlbflush.h>
40 #include <asm/shmparam.h>
44 bool is_vmalloc_addr(const void *x
)
46 unsigned long addr
= (unsigned long)x
;
48 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
50 EXPORT_SYMBOL(is_vmalloc_addr
);
52 struct vfree_deferred
{
53 struct llist_head list
;
54 struct work_struct wq
;
56 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
58 static void __vunmap(const void *, int);
60 static void free_work(struct work_struct
*w
)
62 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
63 struct llist_node
*t
, *llnode
;
65 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
66 __vunmap((void *)llnode
, 1);
69 /*** Page table manipulation functions ***/
71 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
)
75 pte
= pte_offset_kernel(pmd
, addr
);
77 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
78 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
79 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
82 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
)
87 pmd
= pmd_offset(pud
, addr
);
89 next
= pmd_addr_end(addr
, end
);
90 if (pmd_clear_huge(pmd
))
92 if (pmd_none_or_clear_bad(pmd
))
94 vunmap_pte_range(pmd
, addr
, next
);
95 } while (pmd
++, addr
= next
, addr
!= end
);
98 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
)
103 pud
= pud_offset(p4d
, addr
);
105 next
= pud_addr_end(addr
, end
);
106 if (pud_clear_huge(pud
))
108 if (pud_none_or_clear_bad(pud
))
110 vunmap_pmd_range(pud
, addr
, next
);
111 } while (pud
++, addr
= next
, addr
!= end
);
114 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
)
119 p4d
= p4d_offset(pgd
, addr
);
121 next
= p4d_addr_end(addr
, end
);
122 if (p4d_clear_huge(p4d
))
124 if (p4d_none_or_clear_bad(p4d
))
126 vunmap_pud_range(p4d
, addr
, next
);
127 } while (p4d
++, addr
= next
, addr
!= end
);
130 static void vunmap_page_range(unsigned long addr
, unsigned long end
)
136 pgd
= pgd_offset_k(addr
);
138 next
= pgd_addr_end(addr
, end
);
139 if (pgd_none_or_clear_bad(pgd
))
141 vunmap_p4d_range(pgd
, addr
, next
);
142 } while (pgd
++, addr
= next
, addr
!= end
);
145 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
146 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
151 * nr is a running index into the array which helps higher level
152 * callers keep track of where we're up to.
155 pte
= pte_alloc_kernel(pmd
, addr
);
159 struct page
*page
= pages
[*nr
];
161 if (WARN_ON(!pte_none(*pte
)))
165 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
167 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
171 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
172 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
177 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
181 next
= pmd_addr_end(addr
, end
);
182 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
184 } while (pmd
++, addr
= next
, addr
!= end
);
188 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
189 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
194 pud
= pud_alloc(&init_mm
, p4d
, addr
);
198 next
= pud_addr_end(addr
, end
);
199 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
201 } while (pud
++, addr
= next
, addr
!= end
);
205 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
206 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
211 p4d
= p4d_alloc(&init_mm
, pgd
, addr
);
215 next
= p4d_addr_end(addr
, end
);
216 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
))
218 } while (p4d
++, addr
= next
, addr
!= end
);
223 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
224 * will have pfns corresponding to the "pages" array.
226 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
228 static int vmap_page_range_noflush(unsigned long start
, unsigned long end
,
229 pgprot_t prot
, struct page
**pages
)
233 unsigned long addr
= start
;
238 pgd
= pgd_offset_k(addr
);
240 next
= pgd_addr_end(addr
, end
);
241 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
);
244 } while (pgd
++, addr
= next
, addr
!= end
);
249 static int vmap_page_range(unsigned long start
, unsigned long end
,
250 pgprot_t prot
, struct page
**pages
)
254 ret
= vmap_page_range_noflush(start
, end
, prot
, pages
);
255 flush_cache_vmap(start
, end
);
259 int is_vmalloc_or_module_addr(const void *x
)
262 * ARM, x86-64 and sparc64 put modules in a special place,
263 * and fall back on vmalloc() if that fails. Others
264 * just put it in the vmalloc space.
266 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
267 unsigned long addr
= (unsigned long)x
;
268 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
271 return is_vmalloc_addr(x
);
275 * Walk a vmap address to the struct page it maps.
277 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
279 unsigned long addr
= (unsigned long) vmalloc_addr
;
280 struct page
*page
= NULL
;
281 pgd_t
*pgd
= pgd_offset_k(addr
);
288 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
289 * architectures that do not vmalloc module space
291 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
295 p4d
= p4d_offset(pgd
, addr
);
298 pud
= pud_offset(p4d
, addr
);
301 * Don't dereference bad PUD or PMD (below) entries. This will also
302 * identify huge mappings, which we may encounter on architectures
303 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
304 * identified as vmalloc addresses by is_vmalloc_addr(), but are
305 * not [unambiguously] associated with a struct page, so there is
306 * no correct value to return for them.
308 WARN_ON_ONCE(pud_bad(*pud
));
309 if (pud_none(*pud
) || pud_bad(*pud
))
311 pmd
= pmd_offset(pud
, addr
);
312 WARN_ON_ONCE(pmd_bad(*pmd
));
313 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
316 ptep
= pte_offset_map(pmd
, addr
);
318 if (pte_present(pte
))
319 page
= pte_page(pte
);
323 EXPORT_SYMBOL(vmalloc_to_page
);
326 * Map a vmalloc()-space virtual address to the physical page frame number.
328 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
330 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
332 EXPORT_SYMBOL(vmalloc_to_pfn
);
335 /*** Global kva allocator ***/
337 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
338 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
341 static DEFINE_SPINLOCK(vmap_area_lock
);
342 static DEFINE_SPINLOCK(free_vmap_area_lock
);
343 /* Export for kexec only */
344 LIST_HEAD(vmap_area_list
);
345 static LLIST_HEAD(vmap_purge_list
);
346 static struct rb_root vmap_area_root
= RB_ROOT
;
347 static bool vmap_initialized __read_mostly
;
350 * This kmem_cache is used for vmap_area objects. Instead of
351 * allocating from slab we reuse an object from this cache to
352 * make things faster. Especially in "no edge" splitting of
355 static struct kmem_cache
*vmap_area_cachep
;
358 * This linked list is used in pair with free_vmap_area_root.
359 * It gives O(1) access to prev/next to perform fast coalescing.
361 static LIST_HEAD(free_vmap_area_list
);
364 * This augment red-black tree represents the free vmap space.
365 * All vmap_area objects in this tree are sorted by va->va_start
366 * address. It is used for allocation and merging when a vmap
367 * object is released.
369 * Each vmap_area node contains a maximum available free block
370 * of its sub-tree, right or left. Therefore it is possible to
371 * find a lowest match of free area.
373 static struct rb_root free_vmap_area_root
= RB_ROOT
;
376 * Preload a CPU with one object for "no edge" split case. The
377 * aim is to get rid of allocations from the atomic context, thus
378 * to use more permissive allocation masks.
380 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
382 static __always_inline
unsigned long
383 va_size(struct vmap_area
*va
)
385 return (va
->va_end
- va
->va_start
);
388 static __always_inline
unsigned long
389 get_subtree_max_size(struct rb_node
*node
)
391 struct vmap_area
*va
;
393 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
394 return va
? va
->subtree_max_size
: 0;
398 * Gets called when remove the node and rotate.
400 static __always_inline
unsigned long
401 compute_subtree_max_size(struct vmap_area
*va
)
403 return max3(va_size(va
),
404 get_subtree_max_size(va
->rb_node
.rb_left
),
405 get_subtree_max_size(va
->rb_node
.rb_right
));
408 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
409 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
411 static void purge_vmap_area_lazy(void);
412 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
413 static unsigned long lazy_max_pages(void);
415 static atomic_long_t nr_vmalloc_pages
;
417 unsigned long vmalloc_nr_pages(void)
419 return atomic_long_read(&nr_vmalloc_pages
);
422 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
424 struct rb_node
*n
= vmap_area_root
.rb_node
;
427 struct vmap_area
*va
;
429 va
= rb_entry(n
, struct vmap_area
, rb_node
);
430 if (addr
< va
->va_start
)
432 else if (addr
>= va
->va_end
)
442 * This function returns back addresses of parent node
443 * and its left or right link for further processing.
445 static __always_inline
struct rb_node
**
446 find_va_links(struct vmap_area
*va
,
447 struct rb_root
*root
, struct rb_node
*from
,
448 struct rb_node
**parent
)
450 struct vmap_area
*tmp_va
;
451 struct rb_node
**link
;
454 link
= &root
->rb_node
;
455 if (unlikely(!*link
)) {
464 * Go to the bottom of the tree. When we hit the last point
465 * we end up with parent rb_node and correct direction, i name
466 * it link, where the new va->rb_node will be attached to.
469 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
472 * During the traversal we also do some sanity check.
473 * Trigger the BUG() if there are sides(left/right)
476 if (va
->va_start
< tmp_va
->va_end
&&
477 va
->va_end
<= tmp_va
->va_start
)
478 link
= &(*link
)->rb_left
;
479 else if (va
->va_end
> tmp_va
->va_start
&&
480 va
->va_start
>= tmp_va
->va_end
)
481 link
= &(*link
)->rb_right
;
486 *parent
= &tmp_va
->rb_node
;
490 static __always_inline
struct list_head
*
491 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
493 struct list_head
*list
;
495 if (unlikely(!parent
))
497 * The red-black tree where we try to find VA neighbors
498 * before merging or inserting is empty, i.e. it means
499 * there is no free vmap space. Normally it does not
500 * happen but we handle this case anyway.
504 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
505 return (&parent
->rb_right
== link
? list
->next
: list
);
508 static __always_inline
void
509 link_va(struct vmap_area
*va
, struct rb_root
*root
,
510 struct rb_node
*parent
, struct rb_node
**link
, struct list_head
*head
)
513 * VA is still not in the list, but we can
514 * identify its future previous list_head node.
516 if (likely(parent
)) {
517 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
518 if (&parent
->rb_right
!= link
)
522 /* Insert to the rb-tree */
523 rb_link_node(&va
->rb_node
, parent
, link
);
524 if (root
== &free_vmap_area_root
) {
526 * Some explanation here. Just perform simple insertion
527 * to the tree. We do not set va->subtree_max_size to
528 * its current size before calling rb_insert_augmented().
529 * It is because of we populate the tree from the bottom
530 * to parent levels when the node _is_ in the tree.
532 * Therefore we set subtree_max_size to zero after insertion,
533 * to let __augment_tree_propagate_from() puts everything to
534 * the correct order later on.
536 rb_insert_augmented(&va
->rb_node
,
537 root
, &free_vmap_area_rb_augment_cb
);
538 va
->subtree_max_size
= 0;
540 rb_insert_color(&va
->rb_node
, root
);
543 /* Address-sort this list */
544 list_add(&va
->list
, head
);
547 static __always_inline
void
548 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
550 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
553 if (root
== &free_vmap_area_root
)
554 rb_erase_augmented(&va
->rb_node
,
555 root
, &free_vmap_area_rb_augment_cb
);
557 rb_erase(&va
->rb_node
, root
);
560 RB_CLEAR_NODE(&va
->rb_node
);
563 #if DEBUG_AUGMENT_PROPAGATE_CHECK
565 augment_tree_propagate_check(struct rb_node
*n
)
567 struct vmap_area
*va
;
568 struct rb_node
*node
;
575 va
= rb_entry(n
, struct vmap_area
, rb_node
);
576 size
= va
->subtree_max_size
;
580 va
= rb_entry(node
, struct vmap_area
, rb_node
);
582 if (get_subtree_max_size(node
->rb_left
) == size
) {
583 node
= node
->rb_left
;
585 if (va_size(va
) == size
) {
590 node
= node
->rb_right
;
595 va
= rb_entry(n
, struct vmap_area
, rb_node
);
596 pr_emerg("tree is corrupted: %lu, %lu\n",
597 va_size(va
), va
->subtree_max_size
);
600 augment_tree_propagate_check(n
->rb_left
);
601 augment_tree_propagate_check(n
->rb_right
);
606 * This function populates subtree_max_size from bottom to upper
607 * levels starting from VA point. The propagation must be done
608 * when VA size is modified by changing its va_start/va_end. Or
609 * in case of newly inserting of VA to the tree.
611 * It means that __augment_tree_propagate_from() must be called:
612 * - After VA has been inserted to the tree(free path);
613 * - After VA has been shrunk(allocation path);
614 * - After VA has been increased(merging path).
616 * Please note that, it does not mean that upper parent nodes
617 * and their subtree_max_size are recalculated all the time up
626 * For example if we modify the node 4, shrinking it to 2, then
627 * no any modification is required. If we shrink the node 2 to 1
628 * its subtree_max_size is updated only, and set to 1. If we shrink
629 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
632 static __always_inline
void
633 augment_tree_propagate_from(struct vmap_area
*va
)
635 struct rb_node
*node
= &va
->rb_node
;
636 unsigned long new_va_sub_max_size
;
639 va
= rb_entry(node
, struct vmap_area
, rb_node
);
640 new_va_sub_max_size
= compute_subtree_max_size(va
);
643 * If the newly calculated maximum available size of the
644 * subtree is equal to the current one, then it means that
645 * the tree is propagated correctly. So we have to stop at
646 * this point to save cycles.
648 if (va
->subtree_max_size
== new_va_sub_max_size
)
651 va
->subtree_max_size
= new_va_sub_max_size
;
652 node
= rb_parent(&va
->rb_node
);
655 #if DEBUG_AUGMENT_PROPAGATE_CHECK
656 augment_tree_propagate_check(free_vmap_area_root
.rb_node
);
661 insert_vmap_area(struct vmap_area
*va
,
662 struct rb_root
*root
, struct list_head
*head
)
664 struct rb_node
**link
;
665 struct rb_node
*parent
;
667 link
= find_va_links(va
, root
, NULL
, &parent
);
668 link_va(va
, root
, parent
, link
, head
);
672 insert_vmap_area_augment(struct vmap_area
*va
,
673 struct rb_node
*from
, struct rb_root
*root
,
674 struct list_head
*head
)
676 struct rb_node
**link
;
677 struct rb_node
*parent
;
680 link
= find_va_links(va
, NULL
, from
, &parent
);
682 link
= find_va_links(va
, root
, NULL
, &parent
);
684 link_va(va
, root
, parent
, link
, head
);
685 augment_tree_propagate_from(va
);
689 * Merge de-allocated chunk of VA memory with previous
690 * and next free blocks. If coalesce is not done a new
691 * free area is inserted. If VA has been merged, it is
694 static __always_inline
struct vmap_area
*
695 merge_or_add_vmap_area(struct vmap_area
*va
,
696 struct rb_root
*root
, struct list_head
*head
)
698 struct vmap_area
*sibling
;
699 struct list_head
*next
;
700 struct rb_node
**link
;
701 struct rb_node
*parent
;
705 * Find a place in the tree where VA potentially will be
706 * inserted, unless it is merged with its sibling/siblings.
708 link
= find_va_links(va
, root
, NULL
, &parent
);
711 * Get next node of VA to check if merging can be done.
713 next
= get_va_next_sibling(parent
, link
);
714 if (unlikely(next
== NULL
))
720 * |<------VA------>|<-----Next----->|
725 sibling
= list_entry(next
, struct vmap_area
, list
);
726 if (sibling
->va_start
== va
->va_end
) {
727 sibling
->va_start
= va
->va_start
;
729 /* Check and update the tree if needed. */
730 augment_tree_propagate_from(sibling
);
732 /* Free vmap_area object. */
733 kmem_cache_free(vmap_area_cachep
, va
);
735 /* Point to the new merged area. */
744 * |<-----Prev----->|<------VA------>|
748 if (next
->prev
!= head
) {
749 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
750 if (sibling
->va_end
== va
->va_start
) {
751 sibling
->va_end
= va
->va_end
;
753 /* Check and update the tree if needed. */
754 augment_tree_propagate_from(sibling
);
759 /* Free vmap_area object. */
760 kmem_cache_free(vmap_area_cachep
, va
);
762 /* Point to the new merged area. */
770 link_va(va
, root
, parent
, link
, head
);
771 augment_tree_propagate_from(va
);
777 static __always_inline
bool
778 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
779 unsigned long align
, unsigned long vstart
)
781 unsigned long nva_start_addr
;
783 if (va
->va_start
> vstart
)
784 nva_start_addr
= ALIGN(va
->va_start
, align
);
786 nva_start_addr
= ALIGN(vstart
, align
);
788 /* Can be overflowed due to big size or alignment. */
789 if (nva_start_addr
+ size
< nva_start_addr
||
790 nva_start_addr
< vstart
)
793 return (nva_start_addr
+ size
<= va
->va_end
);
797 * Find the first free block(lowest start address) in the tree,
798 * that will accomplish the request corresponding to passing
801 static __always_inline
struct vmap_area
*
802 find_vmap_lowest_match(unsigned long size
,
803 unsigned long align
, unsigned long vstart
)
805 struct vmap_area
*va
;
806 struct rb_node
*node
;
807 unsigned long length
;
809 /* Start from the root. */
810 node
= free_vmap_area_root
.rb_node
;
812 /* Adjust the search size for alignment overhead. */
813 length
= size
+ align
- 1;
816 va
= rb_entry(node
, struct vmap_area
, rb_node
);
818 if (get_subtree_max_size(node
->rb_left
) >= length
&&
819 vstart
< va
->va_start
) {
820 node
= node
->rb_left
;
822 if (is_within_this_va(va
, size
, align
, vstart
))
826 * Does not make sense to go deeper towards the right
827 * sub-tree if it does not have a free block that is
828 * equal or bigger to the requested search length.
830 if (get_subtree_max_size(node
->rb_right
) >= length
) {
831 node
= node
->rb_right
;
836 * OK. We roll back and find the first right sub-tree,
837 * that will satisfy the search criteria. It can happen
838 * only once due to "vstart" restriction.
840 while ((node
= rb_parent(node
))) {
841 va
= rb_entry(node
, struct vmap_area
, rb_node
);
842 if (is_within_this_va(va
, size
, align
, vstart
))
845 if (get_subtree_max_size(node
->rb_right
) >= length
&&
846 vstart
<= va
->va_start
) {
847 node
= node
->rb_right
;
857 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
858 #include <linux/random.h>
860 static struct vmap_area
*
861 find_vmap_lowest_linear_match(unsigned long size
,
862 unsigned long align
, unsigned long vstart
)
864 struct vmap_area
*va
;
866 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
867 if (!is_within_this_va(va
, size
, align
, vstart
))
877 find_vmap_lowest_match_check(unsigned long size
)
879 struct vmap_area
*va_1
, *va_2
;
880 unsigned long vstart
;
883 get_random_bytes(&rnd
, sizeof(rnd
));
884 vstart
= VMALLOC_START
+ rnd
;
886 va_1
= find_vmap_lowest_match(size
, 1, vstart
);
887 va_2
= find_vmap_lowest_linear_match(size
, 1, vstart
);
890 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
897 FL_FIT_TYPE
= 1, /* full fit */
898 LE_FIT_TYPE
= 2, /* left edge fit */
899 RE_FIT_TYPE
= 3, /* right edge fit */
900 NE_FIT_TYPE
= 4 /* no edge fit */
903 static __always_inline
enum fit_type
904 classify_va_fit_type(struct vmap_area
*va
,
905 unsigned long nva_start_addr
, unsigned long size
)
909 /* Check if it is within VA. */
910 if (nva_start_addr
< va
->va_start
||
911 nva_start_addr
+ size
> va
->va_end
)
915 if (va
->va_start
== nva_start_addr
) {
916 if (va
->va_end
== nva_start_addr
+ size
)
920 } else if (va
->va_end
== nva_start_addr
+ size
) {
929 static __always_inline
int
930 adjust_va_to_fit_type(struct vmap_area
*va
,
931 unsigned long nva_start_addr
, unsigned long size
,
934 struct vmap_area
*lva
= NULL
;
936 if (type
== FL_FIT_TYPE
) {
938 * No need to split VA, it fully fits.
944 unlink_va(va
, &free_vmap_area_root
);
945 kmem_cache_free(vmap_area_cachep
, va
);
946 } else if (type
== LE_FIT_TYPE
) {
948 * Split left edge of fit VA.
954 va
->va_start
+= size
;
955 } else if (type
== RE_FIT_TYPE
) {
957 * Split right edge of fit VA.
963 va
->va_end
= nva_start_addr
;
964 } else if (type
== NE_FIT_TYPE
) {
966 * Split no edge of fit VA.
972 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
973 if (unlikely(!lva
)) {
975 * For percpu allocator we do not do any pre-allocation
976 * and leave it as it is. The reason is it most likely
977 * never ends up with NE_FIT_TYPE splitting. In case of
978 * percpu allocations offsets and sizes are aligned to
979 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
980 * are its main fitting cases.
982 * There are a few exceptions though, as an example it is
983 * a first allocation (early boot up) when we have "one"
984 * big free space that has to be split.
986 * Also we can hit this path in case of regular "vmap"
987 * allocations, if "this" current CPU was not preloaded.
988 * See the comment in alloc_vmap_area() why. If so, then
989 * GFP_NOWAIT is used instead to get an extra object for
990 * split purpose. That is rare and most time does not
993 * What happens if an allocation gets failed. Basically,
994 * an "overflow" path is triggered to purge lazily freed
995 * areas to free some memory, then, the "retry" path is
996 * triggered to repeat one more time. See more details
997 * in alloc_vmap_area() function.
999 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1005 * Build the remainder.
1007 lva
->va_start
= va
->va_start
;
1008 lva
->va_end
= nva_start_addr
;
1011 * Shrink this VA to remaining size.
1013 va
->va_start
= nva_start_addr
+ size
;
1018 if (type
!= FL_FIT_TYPE
) {
1019 augment_tree_propagate_from(va
);
1021 if (lva
) /* type == NE_FIT_TYPE */
1022 insert_vmap_area_augment(lva
, &va
->rb_node
,
1023 &free_vmap_area_root
, &free_vmap_area_list
);
1030 * Returns a start address of the newly allocated area, if success.
1031 * Otherwise a vend is returned that indicates failure.
1033 static __always_inline
unsigned long
1034 __alloc_vmap_area(unsigned long size
, unsigned long align
,
1035 unsigned long vstart
, unsigned long vend
)
1037 unsigned long nva_start_addr
;
1038 struct vmap_area
*va
;
1042 va
= find_vmap_lowest_match(size
, align
, vstart
);
1046 if (va
->va_start
> vstart
)
1047 nva_start_addr
= ALIGN(va
->va_start
, align
);
1049 nva_start_addr
= ALIGN(vstart
, align
);
1051 /* Check the "vend" restriction. */
1052 if (nva_start_addr
+ size
> vend
)
1055 /* Classify what we have found. */
1056 type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1057 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
1060 /* Update the free vmap_area. */
1061 ret
= adjust_va_to_fit_type(va
, nva_start_addr
, size
, type
);
1065 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1066 find_vmap_lowest_match_check(size
);
1069 return nva_start_addr
;
1073 * Allocate a region of KVA of the specified size and alignment, within the
1076 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1077 unsigned long align
,
1078 unsigned long vstart
, unsigned long vend
,
1079 int node
, gfp_t gfp_mask
)
1081 struct vmap_area
*va
, *pva
;
1086 BUG_ON(offset_in_page(size
));
1087 BUG_ON(!is_power_of_2(align
));
1089 if (unlikely(!vmap_initialized
))
1090 return ERR_PTR(-EBUSY
);
1093 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1095 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1097 return ERR_PTR(-ENOMEM
);
1100 * Only scan the relevant parts containing pointers to other objects
1101 * to avoid false negatives.
1103 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1107 * Preload this CPU with one extra vmap_area object. It is used
1108 * when fit type of free area is NE_FIT_TYPE. Please note, it
1109 * does not guarantee that an allocation occurs on a CPU that
1110 * is preloaded, instead we minimize the case when it is not.
1111 * It can happen because of cpu migration, because there is a
1112 * race until the below spinlock is taken.
1114 * The preload is done in non-atomic context, thus it allows us
1115 * to use more permissive allocation masks to be more stable under
1116 * low memory condition and high memory pressure. In rare case,
1117 * if not preloaded, GFP_NOWAIT is used.
1119 * Set "pva" to NULL here, because of "retry" path.
1123 if (!this_cpu_read(ne_fit_preload_node
))
1125 * Even if it fails we do not really care about that.
1126 * Just proceed as it is. If needed "overflow" path
1127 * will refill the cache we allocate from.
1129 pva
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1131 spin_lock(&free_vmap_area_lock
);
1133 if (pva
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, pva
))
1134 kmem_cache_free(vmap_area_cachep
, pva
);
1137 * If an allocation fails, the "vend" address is
1138 * returned. Therefore trigger the overflow path.
1140 addr
= __alloc_vmap_area(size
, align
, vstart
, vend
);
1141 spin_unlock(&free_vmap_area_lock
);
1143 if (unlikely(addr
== vend
))
1146 va
->va_start
= addr
;
1147 va
->va_end
= addr
+ size
;
1150 spin_lock(&vmap_area_lock
);
1151 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1152 spin_unlock(&vmap_area_lock
);
1154 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1155 BUG_ON(va
->va_start
< vstart
);
1156 BUG_ON(va
->va_end
> vend
);
1162 purge_vmap_area_lazy();
1167 if (gfpflags_allow_blocking(gfp_mask
)) {
1168 unsigned long freed
= 0;
1169 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1176 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1177 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1180 kmem_cache_free(vmap_area_cachep
, va
);
1181 return ERR_PTR(-EBUSY
);
1184 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1186 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1188 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1190 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1192 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1194 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1197 * Free a region of KVA allocated by alloc_vmap_area
1199 static void free_vmap_area(struct vmap_area
*va
)
1202 * Remove from the busy tree/list.
1204 spin_lock(&vmap_area_lock
);
1205 unlink_va(va
, &vmap_area_root
);
1206 spin_unlock(&vmap_area_lock
);
1209 * Insert/Merge it back to the free tree/list.
1211 spin_lock(&free_vmap_area_lock
);
1212 merge_or_add_vmap_area(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1213 spin_unlock(&free_vmap_area_lock
);
1217 * Clear the pagetable entries of a given vmap_area
1219 static void unmap_vmap_area(struct vmap_area
*va
)
1221 vunmap_page_range(va
->va_start
, va
->va_end
);
1225 * lazy_max_pages is the maximum amount of virtual address space we gather up
1226 * before attempting to purge with a TLB flush.
1228 * There is a tradeoff here: a larger number will cover more kernel page tables
1229 * and take slightly longer to purge, but it will linearly reduce the number of
1230 * global TLB flushes that must be performed. It would seem natural to scale
1231 * this number up linearly with the number of CPUs (because vmapping activity
1232 * could also scale linearly with the number of CPUs), however it is likely
1233 * that in practice, workloads might be constrained in other ways that mean
1234 * vmap activity will not scale linearly with CPUs. Also, I want to be
1235 * conservative and not introduce a big latency on huge systems, so go with
1236 * a less aggressive log scale. It will still be an improvement over the old
1237 * code, and it will be simple to change the scale factor if we find that it
1238 * becomes a problem on bigger systems.
1240 static unsigned long lazy_max_pages(void)
1244 log
= fls(num_online_cpus());
1246 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1249 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1252 * Serialize vmap purging. There is no actual criticial section protected
1253 * by this look, but we want to avoid concurrent calls for performance
1254 * reasons and to make the pcpu_get_vm_areas more deterministic.
1256 static DEFINE_MUTEX(vmap_purge_lock
);
1258 /* for per-CPU blocks */
1259 static void purge_fragmented_blocks_allcpus(void);
1262 * called before a call to iounmap() if the caller wants vm_area_struct's
1263 * immediately freed.
1265 void set_iounmap_nonlazy(void)
1267 atomic_long_set(&vmap_lazy_nr
, lazy_max_pages()+1);
1271 * Purges all lazily-freed vmap areas.
1273 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1275 unsigned long resched_threshold
;
1276 struct llist_node
*valist
;
1277 struct vmap_area
*va
;
1278 struct vmap_area
*n_va
;
1280 lockdep_assert_held(&vmap_purge_lock
);
1282 valist
= llist_del_all(&vmap_purge_list
);
1283 if (unlikely(valist
== NULL
))
1287 * First make sure the mappings are removed from all page-tables
1288 * before they are freed.
1293 * TODO: to calculate a flush range without looping.
1294 * The list can be up to lazy_max_pages() elements.
1296 llist_for_each_entry(va
, valist
, purge_list
) {
1297 if (va
->va_start
< start
)
1298 start
= va
->va_start
;
1299 if (va
->va_end
> end
)
1303 flush_tlb_kernel_range(start
, end
);
1304 resched_threshold
= lazy_max_pages() << 1;
1306 spin_lock(&free_vmap_area_lock
);
1307 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
1308 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1309 unsigned long orig_start
= va
->va_start
;
1310 unsigned long orig_end
= va
->va_end
;
1313 * Finally insert or merge lazily-freed area. It is
1314 * detached and there is no need to "unlink" it from
1317 va
= merge_or_add_vmap_area(va
, &free_vmap_area_root
,
1318 &free_vmap_area_list
);
1320 if (is_vmalloc_or_module_addr((void *)orig_start
))
1321 kasan_release_vmalloc(orig_start
, orig_end
,
1322 va
->va_start
, va
->va_end
);
1324 atomic_long_sub(nr
, &vmap_lazy_nr
);
1326 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1327 cond_resched_lock(&free_vmap_area_lock
);
1329 spin_unlock(&free_vmap_area_lock
);
1334 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1335 * is already purging.
1337 static void try_purge_vmap_area_lazy(void)
1339 if (mutex_trylock(&vmap_purge_lock
)) {
1340 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1341 mutex_unlock(&vmap_purge_lock
);
1346 * Kick off a purge of the outstanding lazy areas.
1348 static void purge_vmap_area_lazy(void)
1350 mutex_lock(&vmap_purge_lock
);
1351 purge_fragmented_blocks_allcpus();
1352 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1353 mutex_unlock(&vmap_purge_lock
);
1357 * Free a vmap area, caller ensuring that the area has been unmapped
1358 * and flush_cache_vunmap had been called for the correct range
1361 static void free_vmap_area_noflush(struct vmap_area
*va
)
1363 unsigned long nr_lazy
;
1365 spin_lock(&vmap_area_lock
);
1366 unlink_va(va
, &vmap_area_root
);
1367 spin_unlock(&vmap_area_lock
);
1369 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1370 PAGE_SHIFT
, &vmap_lazy_nr
);
1372 /* After this point, we may free va at any time */
1373 llist_add(&va
->purge_list
, &vmap_purge_list
);
1375 if (unlikely(nr_lazy
> lazy_max_pages()))
1376 try_purge_vmap_area_lazy();
1380 * Free and unmap a vmap area
1382 static void free_unmap_vmap_area(struct vmap_area
*va
)
1384 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1385 unmap_vmap_area(va
);
1386 if (debug_pagealloc_enabled())
1387 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1389 free_vmap_area_noflush(va
);
1392 static struct vmap_area
*find_vmap_area(unsigned long addr
)
1394 struct vmap_area
*va
;
1396 spin_lock(&vmap_area_lock
);
1397 va
= __find_vmap_area(addr
);
1398 spin_unlock(&vmap_area_lock
);
1403 /*** Per cpu kva allocator ***/
1406 * vmap space is limited especially on 32 bit architectures. Ensure there is
1407 * room for at least 16 percpu vmap blocks per CPU.
1410 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1411 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1412 * instead (we just need a rough idea)
1414 #if BITS_PER_LONG == 32
1415 #define VMALLOC_SPACE (128UL*1024*1024)
1417 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1420 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1421 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1422 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1423 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1424 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1425 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1426 #define VMAP_BBMAP_BITS \
1427 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1428 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1429 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1431 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1433 struct vmap_block_queue
{
1435 struct list_head free
;
1440 struct vmap_area
*va
;
1441 unsigned long free
, dirty
;
1442 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1443 struct list_head free_list
;
1444 struct rcu_head rcu_head
;
1445 struct list_head purge
;
1448 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1449 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1452 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1453 * in the free path. Could get rid of this if we change the API to return a
1454 * "cookie" from alloc, to be passed to free. But no big deal yet.
1456 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
1457 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
1460 * We should probably have a fallback mechanism to allocate virtual memory
1461 * out of partially filled vmap blocks. However vmap block sizing should be
1462 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1466 static unsigned long addr_to_vb_idx(unsigned long addr
)
1468 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1469 addr
/= VMAP_BLOCK_SIZE
;
1473 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1477 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1478 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1479 return (void *)addr
;
1483 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1484 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1485 * @order: how many 2^order pages should be occupied in newly allocated block
1486 * @gfp_mask: flags for the page level allocator
1488 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1490 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1492 struct vmap_block_queue
*vbq
;
1493 struct vmap_block
*vb
;
1494 struct vmap_area
*va
;
1495 unsigned long vb_idx
;
1499 node
= numa_node_id();
1501 vb
= kmalloc_node(sizeof(struct vmap_block
),
1502 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1504 return ERR_PTR(-ENOMEM
);
1506 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1507 VMALLOC_START
, VMALLOC_END
,
1511 return ERR_CAST(va
);
1514 err
= radix_tree_preload(gfp_mask
);
1515 if (unlikely(err
)) {
1518 return ERR_PTR(err
);
1521 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1522 spin_lock_init(&vb
->lock
);
1524 /* At least something should be left free */
1525 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1526 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
1528 vb
->dirty_min
= VMAP_BBMAP_BITS
;
1530 INIT_LIST_HEAD(&vb
->free_list
);
1532 vb_idx
= addr_to_vb_idx(va
->va_start
);
1533 spin_lock(&vmap_block_tree_lock
);
1534 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
1535 spin_unlock(&vmap_block_tree_lock
);
1537 radix_tree_preload_end();
1539 vbq
= &get_cpu_var(vmap_block_queue
);
1540 spin_lock(&vbq
->lock
);
1541 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
1542 spin_unlock(&vbq
->lock
);
1543 put_cpu_var(vmap_block_queue
);
1548 static void free_vmap_block(struct vmap_block
*vb
)
1550 struct vmap_block
*tmp
;
1551 unsigned long vb_idx
;
1553 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
1554 spin_lock(&vmap_block_tree_lock
);
1555 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
1556 spin_unlock(&vmap_block_tree_lock
);
1559 free_vmap_area_noflush(vb
->va
);
1560 kfree_rcu(vb
, rcu_head
);
1563 static void purge_fragmented_blocks(int cpu
)
1566 struct vmap_block
*vb
;
1567 struct vmap_block
*n_vb
;
1568 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1571 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1573 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
1576 spin_lock(&vb
->lock
);
1577 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
1578 vb
->free
= 0; /* prevent further allocs after releasing lock */
1579 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
1581 vb
->dirty_max
= VMAP_BBMAP_BITS
;
1582 spin_lock(&vbq
->lock
);
1583 list_del_rcu(&vb
->free_list
);
1584 spin_unlock(&vbq
->lock
);
1585 spin_unlock(&vb
->lock
);
1586 list_add_tail(&vb
->purge
, &purge
);
1588 spin_unlock(&vb
->lock
);
1592 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
1593 list_del(&vb
->purge
);
1594 free_vmap_block(vb
);
1598 static void purge_fragmented_blocks_allcpus(void)
1602 for_each_possible_cpu(cpu
)
1603 purge_fragmented_blocks(cpu
);
1606 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
1608 struct vmap_block_queue
*vbq
;
1609 struct vmap_block
*vb
;
1613 BUG_ON(offset_in_page(size
));
1614 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1615 if (WARN_ON(size
== 0)) {
1617 * Allocating 0 bytes isn't what caller wants since
1618 * get_order(0) returns funny result. Just warn and terminate
1623 order
= get_order(size
);
1626 vbq
= &get_cpu_var(vmap_block_queue
);
1627 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1628 unsigned long pages_off
;
1630 spin_lock(&vb
->lock
);
1631 if (vb
->free
< (1UL << order
)) {
1632 spin_unlock(&vb
->lock
);
1636 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1637 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1638 vb
->free
-= 1UL << order
;
1639 if (vb
->free
== 0) {
1640 spin_lock(&vbq
->lock
);
1641 list_del_rcu(&vb
->free_list
);
1642 spin_unlock(&vbq
->lock
);
1645 spin_unlock(&vb
->lock
);
1649 put_cpu_var(vmap_block_queue
);
1652 /* Allocate new block if nothing was found */
1654 vaddr
= new_vmap_block(order
, gfp_mask
);
1659 static void vb_free(const void *addr
, unsigned long size
)
1661 unsigned long offset
;
1662 unsigned long vb_idx
;
1664 struct vmap_block
*vb
;
1666 BUG_ON(offset_in_page(size
));
1667 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1669 flush_cache_vunmap((unsigned long)addr
, (unsigned long)addr
+ size
);
1671 order
= get_order(size
);
1673 offset
= (unsigned long)addr
& (VMAP_BLOCK_SIZE
- 1);
1674 offset
>>= PAGE_SHIFT
;
1676 vb_idx
= addr_to_vb_idx((unsigned long)addr
);
1678 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1682 vunmap_page_range((unsigned long)addr
, (unsigned long)addr
+ size
);
1684 if (debug_pagealloc_enabled())
1685 flush_tlb_kernel_range((unsigned long)addr
,
1686 (unsigned long)addr
+ size
);
1688 spin_lock(&vb
->lock
);
1690 /* Expand dirty range */
1691 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1692 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1694 vb
->dirty
+= 1UL << order
;
1695 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1697 spin_unlock(&vb
->lock
);
1698 free_vmap_block(vb
);
1700 spin_unlock(&vb
->lock
);
1703 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
1707 if (unlikely(!vmap_initialized
))
1712 for_each_possible_cpu(cpu
) {
1713 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1714 struct vmap_block
*vb
;
1717 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1718 spin_lock(&vb
->lock
);
1720 unsigned long va_start
= vb
->va
->va_start
;
1723 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1724 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1726 start
= min(s
, start
);
1731 spin_unlock(&vb
->lock
);
1736 mutex_lock(&vmap_purge_lock
);
1737 purge_fragmented_blocks_allcpus();
1738 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1739 flush_tlb_kernel_range(start
, end
);
1740 mutex_unlock(&vmap_purge_lock
);
1744 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1746 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1747 * to amortize TLB flushing overheads. What this means is that any page you
1748 * have now, may, in a former life, have been mapped into kernel virtual
1749 * address by the vmap layer and so there might be some CPUs with TLB entries
1750 * still referencing that page (additional to the regular 1:1 kernel mapping).
1752 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1753 * be sure that none of the pages we have control over will have any aliases
1754 * from the vmap layer.
1756 void vm_unmap_aliases(void)
1758 unsigned long start
= ULONG_MAX
, end
= 0;
1761 _vm_unmap_aliases(start
, end
, flush
);
1763 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1766 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1767 * @mem: the pointer returned by vm_map_ram
1768 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1770 void vm_unmap_ram(const void *mem
, unsigned int count
)
1772 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1773 unsigned long addr
= (unsigned long)mem
;
1774 struct vmap_area
*va
;
1778 BUG_ON(addr
< VMALLOC_START
);
1779 BUG_ON(addr
> VMALLOC_END
);
1780 BUG_ON(!PAGE_ALIGNED(addr
));
1782 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1783 debug_check_no_locks_freed(mem
, size
);
1788 va
= find_vmap_area(addr
);
1790 debug_check_no_locks_freed((void *)va
->va_start
,
1791 (va
->va_end
- va
->va_start
));
1792 free_unmap_vmap_area(va
);
1794 EXPORT_SYMBOL(vm_unmap_ram
);
1797 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1798 * @pages: an array of pointers to the pages to be mapped
1799 * @count: number of pages
1800 * @node: prefer to allocate data structures on this node
1801 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1803 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1804 * faster than vmap so it's good. But if you mix long-life and short-life
1805 * objects with vm_map_ram(), it could consume lots of address space through
1806 * fragmentation (especially on a 32bit machine). You could see failures in
1807 * the end. Please use this function for short-lived objects.
1809 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1811 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
, pgprot_t prot
)
1813 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1817 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1818 mem
= vb_alloc(size
, GFP_KERNEL
);
1821 addr
= (unsigned long)mem
;
1823 struct vmap_area
*va
;
1824 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1825 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1829 addr
= va
->va_start
;
1832 if (vmap_page_range(addr
, addr
+ size
, prot
, pages
) < 0) {
1833 vm_unmap_ram(mem
, count
);
1838 EXPORT_SYMBOL(vm_map_ram
);
1840 static struct vm_struct
*vmlist __initdata
;
1843 * vm_area_add_early - add vmap area early during boot
1844 * @vm: vm_struct to add
1846 * This function is used to add fixed kernel vm area to vmlist before
1847 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1848 * should contain proper values and the other fields should be zero.
1850 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1852 void __init
vm_area_add_early(struct vm_struct
*vm
)
1854 struct vm_struct
*tmp
, **p
;
1856 BUG_ON(vmap_initialized
);
1857 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1858 if (tmp
->addr
>= vm
->addr
) {
1859 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1862 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1869 * vm_area_register_early - register vmap area early during boot
1870 * @vm: vm_struct to register
1871 * @align: requested alignment
1873 * This function is used to register kernel vm area before
1874 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1875 * proper values on entry and other fields should be zero. On return,
1876 * vm->addr contains the allocated address.
1878 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1880 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1882 static size_t vm_init_off __initdata
;
1885 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1886 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1888 vm
->addr
= (void *)addr
;
1890 vm_area_add_early(vm
);
1893 static void vmap_init_free_space(void)
1895 unsigned long vmap_start
= 1;
1896 const unsigned long vmap_end
= ULONG_MAX
;
1897 struct vmap_area
*busy
, *free
;
1901 * -|-----|.....|-----|-----|-----|.....|-
1903 * |<--------------------------------->|
1905 list_for_each_entry(busy
, &vmap_area_list
, list
) {
1906 if (busy
->va_start
- vmap_start
> 0) {
1907 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1908 if (!WARN_ON_ONCE(!free
)) {
1909 free
->va_start
= vmap_start
;
1910 free
->va_end
= busy
->va_start
;
1912 insert_vmap_area_augment(free
, NULL
,
1913 &free_vmap_area_root
,
1914 &free_vmap_area_list
);
1918 vmap_start
= busy
->va_end
;
1921 if (vmap_end
- vmap_start
> 0) {
1922 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1923 if (!WARN_ON_ONCE(!free
)) {
1924 free
->va_start
= vmap_start
;
1925 free
->va_end
= vmap_end
;
1927 insert_vmap_area_augment(free
, NULL
,
1928 &free_vmap_area_root
,
1929 &free_vmap_area_list
);
1934 void __init
vmalloc_init(void)
1936 struct vmap_area
*va
;
1937 struct vm_struct
*tmp
;
1941 * Create the cache for vmap_area objects.
1943 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
1945 for_each_possible_cpu(i
) {
1946 struct vmap_block_queue
*vbq
;
1947 struct vfree_deferred
*p
;
1949 vbq
= &per_cpu(vmap_block_queue
, i
);
1950 spin_lock_init(&vbq
->lock
);
1951 INIT_LIST_HEAD(&vbq
->free
);
1952 p
= &per_cpu(vfree_deferred
, i
);
1953 init_llist_head(&p
->list
);
1954 INIT_WORK(&p
->wq
, free_work
);
1957 /* Import existing vmlist entries. */
1958 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1959 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1960 if (WARN_ON_ONCE(!va
))
1963 va
->va_start
= (unsigned long)tmp
->addr
;
1964 va
->va_end
= va
->va_start
+ tmp
->size
;
1966 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1970 * Now we can initialize a free vmap space.
1972 vmap_init_free_space();
1973 vmap_initialized
= true;
1977 * map_kernel_range_noflush - map kernel VM area with the specified pages
1978 * @addr: start of the VM area to map
1979 * @size: size of the VM area to map
1980 * @prot: page protection flags to use
1981 * @pages: pages to map
1983 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1984 * specify should have been allocated using get_vm_area() and its
1988 * This function does NOT do any cache flushing. The caller is
1989 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1990 * before calling this function.
1993 * The number of pages mapped on success, -errno on failure.
1995 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
1996 pgprot_t prot
, struct page
**pages
)
1998 return vmap_page_range_noflush(addr
, addr
+ size
, prot
, pages
);
2002 * unmap_kernel_range_noflush - unmap kernel VM area
2003 * @addr: start of the VM area to unmap
2004 * @size: size of the VM area to unmap
2006 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
2007 * specify should have been allocated using get_vm_area() and its
2011 * This function does NOT do any cache flushing. The caller is
2012 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
2013 * before calling this function and flush_tlb_kernel_range() after.
2015 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
2017 vunmap_page_range(addr
, addr
+ size
);
2019 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush
);
2022 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2023 * @addr: start of the VM area to unmap
2024 * @size: size of the VM area to unmap
2026 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2027 * the unmapping and tlb after.
2029 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
2031 unsigned long end
= addr
+ size
;
2033 flush_cache_vunmap(addr
, end
);
2034 vunmap_page_range(addr
, end
);
2035 flush_tlb_kernel_range(addr
, end
);
2037 EXPORT_SYMBOL_GPL(unmap_kernel_range
);
2039 int map_vm_area(struct vm_struct
*area
, pgprot_t prot
, struct page
**pages
)
2041 unsigned long addr
= (unsigned long)area
->addr
;
2042 unsigned long end
= addr
+ get_vm_area_size(area
);
2045 err
= vmap_page_range(addr
, end
, prot
, pages
);
2047 return err
> 0 ? 0 : err
;
2049 EXPORT_SYMBOL_GPL(map_vm_area
);
2051 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2052 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2055 vm
->addr
= (void *)va
->va_start
;
2056 vm
->size
= va
->va_end
- va
->va_start
;
2057 vm
->caller
= caller
;
2061 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2062 unsigned long flags
, const void *caller
)
2064 spin_lock(&vmap_area_lock
);
2065 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2066 spin_unlock(&vmap_area_lock
);
2069 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2072 * Before removing VM_UNINITIALIZED,
2073 * we should make sure that vm has proper values.
2074 * Pair with smp_rmb() in show_numa_info().
2077 vm
->flags
&= ~VM_UNINITIALIZED
;
2080 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2081 unsigned long align
, unsigned long flags
, unsigned long start
,
2082 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
2084 struct vmap_area
*va
;
2085 struct vm_struct
*area
;
2087 BUG_ON(in_interrupt());
2088 size
= PAGE_ALIGN(size
);
2089 if (unlikely(!size
))
2092 if (flags
& VM_IOREMAP
)
2093 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2094 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2096 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2097 if (unlikely(!area
))
2100 if (!(flags
& VM_NO_GUARD
))
2103 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
2109 setup_vmalloc_vm(area
, va
, flags
, caller
);
2112 * For KASAN, if we are in vmalloc space, we need to cover the shadow
2113 * area with real memory. If we come here through VM_ALLOC, this is
2114 * done by a higher level function that has access to the true size,
2115 * which might not be a full page.
2117 * We assume module space comes via VM_ALLOC path.
2119 if (is_vmalloc_addr(area
->addr
) && !(area
->flags
& VM_ALLOC
)) {
2120 if (kasan_populate_vmalloc(area
->size
, area
)) {
2121 unmap_vmap_area(va
);
2130 struct vm_struct
*__get_vm_area(unsigned long size
, unsigned long flags
,
2131 unsigned long start
, unsigned long end
)
2133 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2134 GFP_KERNEL
, __builtin_return_address(0));
2136 EXPORT_SYMBOL_GPL(__get_vm_area
);
2138 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2139 unsigned long start
, unsigned long end
,
2142 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2143 GFP_KERNEL
, caller
);
2147 * get_vm_area - reserve a contiguous kernel virtual area
2148 * @size: size of the area
2149 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2151 * Search an area of @size in the kernel virtual mapping area,
2152 * and reserved it for out purposes. Returns the area descriptor
2153 * on success or %NULL on failure.
2155 * Return: the area descriptor on success or %NULL on failure.
2157 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2159 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2160 NUMA_NO_NODE
, GFP_KERNEL
,
2161 __builtin_return_address(0));
2164 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2167 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2168 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2172 * find_vm_area - find a continuous kernel virtual area
2173 * @addr: base address
2175 * Search for the kernel VM area starting at @addr, and return it.
2176 * It is up to the caller to do all required locking to keep the returned
2179 * Return: pointer to the found area or %NULL on faulure
2181 struct vm_struct
*find_vm_area(const void *addr
)
2183 struct vmap_area
*va
;
2185 va
= find_vmap_area((unsigned long)addr
);
2193 * remove_vm_area - find and remove a continuous kernel virtual area
2194 * @addr: base address
2196 * Search for the kernel VM area starting at @addr, and remove it.
2197 * This function returns the found VM area, but using it is NOT safe
2198 * on SMP machines, except for its size or flags.
2200 * Return: pointer to the found area or %NULL on faulure
2202 struct vm_struct
*remove_vm_area(const void *addr
)
2204 struct vmap_area
*va
;
2208 spin_lock(&vmap_area_lock
);
2209 va
= __find_vmap_area((unsigned long)addr
);
2211 struct vm_struct
*vm
= va
->vm
;
2214 spin_unlock(&vmap_area_lock
);
2216 kasan_free_shadow(vm
);
2217 free_unmap_vmap_area(va
);
2222 spin_unlock(&vmap_area_lock
);
2226 static inline void set_area_direct_map(const struct vm_struct
*area
,
2227 int (*set_direct_map
)(struct page
*page
))
2231 for (i
= 0; i
< area
->nr_pages
; i
++)
2232 if (page_address(area
->pages
[i
]))
2233 set_direct_map(area
->pages
[i
]);
2236 /* Handle removing and resetting vm mappings related to the vm_struct. */
2237 static void vm_remove_mappings(struct vm_struct
*area
, int deallocate_pages
)
2239 unsigned long start
= ULONG_MAX
, end
= 0;
2240 int flush_reset
= area
->flags
& VM_FLUSH_RESET_PERMS
;
2244 remove_vm_area(area
->addr
);
2246 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2251 * If not deallocating pages, just do the flush of the VM area and
2254 if (!deallocate_pages
) {
2260 * If execution gets here, flush the vm mapping and reset the direct
2261 * map. Find the start and end range of the direct mappings to make sure
2262 * the vm_unmap_aliases() flush includes the direct map.
2264 for (i
= 0; i
< area
->nr_pages
; i
++) {
2265 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2267 start
= min(addr
, start
);
2268 end
= max(addr
+ PAGE_SIZE
, end
);
2274 * Set direct map to something invalid so that it won't be cached if
2275 * there are any accesses after the TLB flush, then flush the TLB and
2276 * reset the direct map permissions to the default.
2278 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2279 _vm_unmap_aliases(start
, end
, flush_dmap
);
2280 set_area_direct_map(area
, set_direct_map_default_noflush
);
2283 static void __vunmap(const void *addr
, int deallocate_pages
)
2285 struct vm_struct
*area
;
2290 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2294 area
= find_vm_area(addr
);
2295 if (unlikely(!area
)) {
2296 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2301 debug_check_no_locks_freed(area
->addr
, get_vm_area_size(area
));
2302 debug_check_no_obj_freed(area
->addr
, get_vm_area_size(area
));
2304 if (area
->flags
& VM_KASAN
)
2305 kasan_poison_vmalloc(area
->addr
, area
->size
);
2307 vm_remove_mappings(area
, deallocate_pages
);
2309 if (deallocate_pages
) {
2312 for (i
= 0; i
< area
->nr_pages
; i
++) {
2313 struct page
*page
= area
->pages
[i
];
2316 __free_pages(page
, 0);
2318 atomic_long_sub(area
->nr_pages
, &nr_vmalloc_pages
);
2320 kvfree(area
->pages
);
2327 static inline void __vfree_deferred(const void *addr
)
2330 * Use raw_cpu_ptr() because this can be called from preemptible
2331 * context. Preemption is absolutely fine here, because the llist_add()
2332 * implementation is lockless, so it works even if we are adding to
2333 * nother cpu's list. schedule_work() should be fine with this too.
2335 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2337 if (llist_add((struct llist_node
*)addr
, &p
->list
))
2338 schedule_work(&p
->wq
);
2342 * vfree_atomic - release memory allocated by vmalloc()
2343 * @addr: memory base address
2345 * This one is just like vfree() but can be called in any atomic context
2348 void vfree_atomic(const void *addr
)
2352 kmemleak_free(addr
);
2356 __vfree_deferred(addr
);
2359 static void __vfree(const void *addr
)
2361 if (unlikely(in_interrupt()))
2362 __vfree_deferred(addr
);
2368 * vfree - release memory allocated by vmalloc()
2369 * @addr: memory base address
2371 * Free the virtually continuous memory area starting at @addr, as
2372 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2373 * NULL, no operation is performed.
2375 * Must not be called in NMI context (strictly speaking, only if we don't
2376 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2377 * conventions for vfree() arch-depenedent would be a really bad idea)
2379 * May sleep if called *not* from interrupt context.
2381 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2383 void vfree(const void *addr
)
2387 kmemleak_free(addr
);
2389 might_sleep_if(!in_interrupt());
2396 EXPORT_SYMBOL(vfree
);
2399 * vunmap - release virtual mapping obtained by vmap()
2400 * @addr: memory base address
2402 * Free the virtually contiguous memory area starting at @addr,
2403 * which was created from the page array passed to vmap().
2405 * Must not be called in interrupt context.
2407 void vunmap(const void *addr
)
2409 BUG_ON(in_interrupt());
2414 EXPORT_SYMBOL(vunmap
);
2417 * vmap - map an array of pages into virtually contiguous space
2418 * @pages: array of page pointers
2419 * @count: number of pages to map
2420 * @flags: vm_area->flags
2421 * @prot: page protection for the mapping
2423 * Maps @count pages from @pages into contiguous kernel virtual
2426 * Return: the address of the area or %NULL on failure
2428 void *vmap(struct page
**pages
, unsigned int count
,
2429 unsigned long flags
, pgprot_t prot
)
2431 struct vm_struct
*area
;
2432 unsigned long size
; /* In bytes */
2436 if (count
> totalram_pages())
2439 size
= (unsigned long)count
<< PAGE_SHIFT
;
2440 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2444 if (map_vm_area(area
, prot
, pages
)) {
2451 EXPORT_SYMBOL(vmap
);
2453 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
2454 gfp_t gfp_mask
, pgprot_t prot
,
2455 int node
, const void *caller
);
2456 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2457 pgprot_t prot
, int node
)
2459 struct page
**pages
;
2460 unsigned int nr_pages
, array_size
, i
;
2461 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2462 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
2463 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
2467 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
2468 array_size
= (nr_pages
* sizeof(struct page
*));
2470 /* Please note that the recursion is strictly bounded. */
2471 if (array_size
> PAGE_SIZE
) {
2472 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
2473 PAGE_KERNEL
, node
, area
->caller
);
2475 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
2479 remove_vm_area(area
->addr
);
2484 area
->pages
= pages
;
2485 area
->nr_pages
= nr_pages
;
2487 for (i
= 0; i
< area
->nr_pages
; i
++) {
2490 if (node
== NUMA_NO_NODE
)
2491 page
= alloc_page(alloc_mask
|highmem_mask
);
2493 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
2495 if (unlikely(!page
)) {
2496 /* Successfully allocated i pages, free them in __vunmap() */
2498 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2501 area
->pages
[i
] = page
;
2502 if (gfpflags_allow_blocking(gfp_mask
))
2505 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2507 if (map_vm_area(area
, prot
, pages
))
2512 warn_alloc(gfp_mask
, NULL
,
2513 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2514 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
2515 __vfree(area
->addr
);
2520 * __vmalloc_node_range - allocate virtually contiguous memory
2521 * @size: allocation size
2522 * @align: desired alignment
2523 * @start: vm area range start
2524 * @end: vm area range end
2525 * @gfp_mask: flags for the page level allocator
2526 * @prot: protection mask for the allocated pages
2527 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2528 * @node: node to use for allocation or NUMA_NO_NODE
2529 * @caller: caller's return address
2531 * Allocate enough pages to cover @size from the page level
2532 * allocator with @gfp_mask flags. Map them into contiguous
2533 * kernel virtual space, using a pagetable protection of @prot.
2535 * Return: the address of the area or %NULL on failure
2537 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
2538 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
2539 pgprot_t prot
, unsigned long vm_flags
, int node
,
2542 struct vm_struct
*area
;
2544 unsigned long real_size
= size
;
2546 size
= PAGE_ALIGN(size
);
2547 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages())
2550 area
= __get_vm_area_node(size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
2551 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
2555 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
2559 if (is_vmalloc_or_module_addr(area
->addr
)) {
2560 if (kasan_populate_vmalloc(real_size
, area
))
2565 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2566 * flag. It means that vm_struct is not fully initialized.
2567 * Now, it is fully initialized, so remove this flag here.
2569 clear_vm_uninitialized_flag(area
);
2571 kmemleak_vmalloc(area
, size
, gfp_mask
);
2576 warn_alloc(gfp_mask
, NULL
,
2577 "vmalloc: allocation failure: %lu bytes", real_size
);
2582 * This is only for performance analysis of vmalloc and stress purpose.
2583 * It is required by vmalloc test module, therefore do not use it other
2586 #ifdef CONFIG_TEST_VMALLOC_MODULE
2587 EXPORT_SYMBOL_GPL(__vmalloc_node_range
);
2591 * __vmalloc_node - allocate virtually contiguous memory
2592 * @size: allocation size
2593 * @align: desired alignment
2594 * @gfp_mask: flags for the page level allocator
2595 * @prot: protection mask for the allocated pages
2596 * @node: node to use for allocation or NUMA_NO_NODE
2597 * @caller: caller's return address
2599 * Allocate enough pages to cover @size from the page level
2600 * allocator with @gfp_mask flags. Map them into contiguous
2601 * kernel virtual space, using a pagetable protection of @prot.
2603 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2604 * and __GFP_NOFAIL are not supported
2606 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2609 * Return: pointer to the allocated memory or %NULL on error
2611 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
2612 gfp_t gfp_mask
, pgprot_t prot
,
2613 int node
, const void *caller
)
2615 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
2616 gfp_mask
, prot
, 0, node
, caller
);
2619 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
, pgprot_t prot
)
2621 return __vmalloc_node(size
, 1, gfp_mask
, prot
, NUMA_NO_NODE
,
2622 __builtin_return_address(0));
2624 EXPORT_SYMBOL(__vmalloc
);
2626 static inline void *__vmalloc_node_flags(unsigned long size
,
2627 int node
, gfp_t flags
)
2629 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
,
2630 node
, __builtin_return_address(0));
2634 void *__vmalloc_node_flags_caller(unsigned long size
, int node
, gfp_t flags
,
2637 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
, node
, caller
);
2641 * vmalloc - allocate virtually contiguous memory
2642 * @size: allocation size
2644 * Allocate enough pages to cover @size from the page level
2645 * allocator and map them into contiguous kernel virtual space.
2647 * For tight control over page level allocator and protection flags
2648 * use __vmalloc() instead.
2650 * Return: pointer to the allocated memory or %NULL on error
2652 void *vmalloc(unsigned long size
)
2654 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
2657 EXPORT_SYMBOL(vmalloc
);
2660 * vzalloc - allocate virtually contiguous memory with zero fill
2661 * @size: allocation size
2663 * Allocate enough pages to cover @size from the page level
2664 * allocator and map them into contiguous kernel virtual space.
2665 * The memory allocated is set to zero.
2667 * For tight control over page level allocator and protection flags
2668 * use __vmalloc() instead.
2670 * Return: pointer to the allocated memory or %NULL on error
2672 void *vzalloc(unsigned long size
)
2674 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
2675 GFP_KERNEL
| __GFP_ZERO
);
2677 EXPORT_SYMBOL(vzalloc
);
2680 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2681 * @size: allocation size
2683 * The resulting memory area is zeroed so it can be mapped to userspace
2684 * without leaking data.
2686 * Return: pointer to the allocated memory or %NULL on error
2688 void *vmalloc_user(unsigned long size
)
2690 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2691 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
2692 VM_USERMAP
, NUMA_NO_NODE
,
2693 __builtin_return_address(0));
2695 EXPORT_SYMBOL(vmalloc_user
);
2698 * vmalloc_node - allocate memory on a specific node
2699 * @size: allocation size
2702 * Allocate enough pages to cover @size from the page level
2703 * allocator and map them into contiguous kernel virtual space.
2705 * For tight control over page level allocator and protection flags
2706 * use __vmalloc() instead.
2708 * Return: pointer to the allocated memory or %NULL on error
2710 void *vmalloc_node(unsigned long size
, int node
)
2712 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL
,
2713 node
, __builtin_return_address(0));
2715 EXPORT_SYMBOL(vmalloc_node
);
2718 * vzalloc_node - allocate memory on a specific node with zero fill
2719 * @size: allocation size
2722 * Allocate enough pages to cover @size from the page level
2723 * allocator and map them into contiguous kernel virtual space.
2724 * The memory allocated is set to zero.
2726 * For tight control over page level allocator and protection flags
2727 * use __vmalloc_node() instead.
2729 * Return: pointer to the allocated memory or %NULL on error
2731 void *vzalloc_node(unsigned long size
, int node
)
2733 return __vmalloc_node_flags(size
, node
,
2734 GFP_KERNEL
| __GFP_ZERO
);
2736 EXPORT_SYMBOL(vzalloc_node
);
2739 * vmalloc_user_node_flags - allocate memory for userspace on a specific node
2740 * @size: allocation size
2742 * @flags: flags for the page level allocator
2744 * The resulting memory area is zeroed so it can be mapped to userspace
2745 * without leaking data.
2747 * Return: pointer to the allocated memory or %NULL on error
2749 void *vmalloc_user_node_flags(unsigned long size
, int node
, gfp_t flags
)
2751 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2752 flags
| __GFP_ZERO
, PAGE_KERNEL
,
2754 __builtin_return_address(0));
2756 EXPORT_SYMBOL(vmalloc_user_node_flags
);
2759 * vmalloc_exec - allocate virtually contiguous, executable memory
2760 * @size: allocation size
2762 * Kernel-internal function to allocate enough pages to cover @size
2763 * the page level allocator and map them into contiguous and
2764 * executable kernel virtual space.
2766 * For tight control over page level allocator and protection flags
2767 * use __vmalloc() instead.
2769 * Return: pointer to the allocated memory or %NULL on error
2771 void *vmalloc_exec(unsigned long size
)
2773 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
2774 GFP_KERNEL
, PAGE_KERNEL_EXEC
, VM_FLUSH_RESET_PERMS
,
2775 NUMA_NO_NODE
, __builtin_return_address(0));
2778 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2779 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2780 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2781 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2784 * 64b systems should always have either DMA or DMA32 zones. For others
2785 * GFP_DMA32 should do the right thing and use the normal zone.
2787 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2791 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2792 * @size: allocation size
2794 * Allocate enough 32bit PA addressable pages to cover @size from the
2795 * page level allocator and map them into contiguous kernel virtual space.
2797 * Return: pointer to the allocated memory or %NULL on error
2799 void *vmalloc_32(unsigned long size
)
2801 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, PAGE_KERNEL
,
2802 NUMA_NO_NODE
, __builtin_return_address(0));
2804 EXPORT_SYMBOL(vmalloc_32
);
2807 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2808 * @size: allocation size
2810 * The resulting memory area is 32bit addressable and zeroed so it can be
2811 * mapped to userspace without leaking data.
2813 * Return: pointer to the allocated memory or %NULL on error
2815 void *vmalloc_32_user(unsigned long size
)
2817 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2818 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2819 VM_USERMAP
, NUMA_NO_NODE
,
2820 __builtin_return_address(0));
2822 EXPORT_SYMBOL(vmalloc_32_user
);
2825 * small helper routine , copy contents to buf from addr.
2826 * If the page is not present, fill zero.
2829 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2835 unsigned long offset
, length
;
2837 offset
= offset_in_page(addr
);
2838 length
= PAGE_SIZE
- offset
;
2841 p
= vmalloc_to_page(addr
);
2843 * To do safe access to this _mapped_ area, we need
2844 * lock. But adding lock here means that we need to add
2845 * overhead of vmalloc()/vfree() calles for this _debug_
2846 * interface, rarely used. Instead of that, we'll use
2847 * kmap() and get small overhead in this access function.
2851 * we can expect USER0 is not used (see vread/vwrite's
2852 * function description)
2854 void *map
= kmap_atomic(p
);
2855 memcpy(buf
, map
+ offset
, length
);
2858 memset(buf
, 0, length
);
2868 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2874 unsigned long offset
, length
;
2876 offset
= offset_in_page(addr
);
2877 length
= PAGE_SIZE
- offset
;
2880 p
= vmalloc_to_page(addr
);
2882 * To do safe access to this _mapped_ area, we need
2883 * lock. But adding lock here means that we need to add
2884 * overhead of vmalloc()/vfree() calles for this _debug_
2885 * interface, rarely used. Instead of that, we'll use
2886 * kmap() and get small overhead in this access function.
2890 * we can expect USER0 is not used (see vread/vwrite's
2891 * function description)
2893 void *map
= kmap_atomic(p
);
2894 memcpy(map
+ offset
, buf
, length
);
2906 * vread() - read vmalloc area in a safe way.
2907 * @buf: buffer for reading data
2908 * @addr: vm address.
2909 * @count: number of bytes to be read.
2911 * This function checks that addr is a valid vmalloc'ed area, and
2912 * copy data from that area to a given buffer. If the given memory range
2913 * of [addr...addr+count) includes some valid address, data is copied to
2914 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2915 * IOREMAP area is treated as memory hole and no copy is done.
2917 * If [addr...addr+count) doesn't includes any intersects with alive
2918 * vm_struct area, returns 0. @buf should be kernel's buffer.
2920 * Note: In usual ops, vread() is never necessary because the caller
2921 * should know vmalloc() area is valid and can use memcpy().
2922 * This is for routines which have to access vmalloc area without
2923 * any information, as /dev/kmem.
2925 * Return: number of bytes for which addr and buf should be increased
2926 * (same number as @count) or %0 if [addr...addr+count) doesn't
2927 * include any intersection with valid vmalloc area
2929 long vread(char *buf
, char *addr
, unsigned long count
)
2931 struct vmap_area
*va
;
2932 struct vm_struct
*vm
;
2933 char *vaddr
, *buf_start
= buf
;
2934 unsigned long buflen
= count
;
2937 /* Don't allow overflow */
2938 if ((unsigned long) addr
+ count
< count
)
2939 count
= -(unsigned long) addr
;
2941 spin_lock(&vmap_area_lock
);
2942 list_for_each_entry(va
, &vmap_area_list
, list
) {
2950 vaddr
= (char *) vm
->addr
;
2951 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2953 while (addr
< vaddr
) {
2961 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2964 if (!(vm
->flags
& VM_IOREMAP
))
2965 aligned_vread(buf
, addr
, n
);
2966 else /* IOREMAP area is treated as memory hole */
2973 spin_unlock(&vmap_area_lock
);
2975 if (buf
== buf_start
)
2977 /* zero-fill memory holes */
2978 if (buf
!= buf_start
+ buflen
)
2979 memset(buf
, 0, buflen
- (buf
- buf_start
));
2985 * vwrite() - write vmalloc area in a safe way.
2986 * @buf: buffer for source data
2987 * @addr: vm address.
2988 * @count: number of bytes to be read.
2990 * This function checks that addr is a valid vmalloc'ed area, and
2991 * copy data from a buffer to the given addr. If specified range of
2992 * [addr...addr+count) includes some valid address, data is copied from
2993 * proper area of @buf. If there are memory holes, no copy to hole.
2994 * IOREMAP area is treated as memory hole and no copy is done.
2996 * If [addr...addr+count) doesn't includes any intersects with alive
2997 * vm_struct area, returns 0. @buf should be kernel's buffer.
2999 * Note: In usual ops, vwrite() is never necessary because the caller
3000 * should know vmalloc() area is valid and can use memcpy().
3001 * This is for routines which have to access vmalloc area without
3002 * any information, as /dev/kmem.
3004 * Return: number of bytes for which addr and buf should be
3005 * increased (same number as @count) or %0 if [addr...addr+count)
3006 * doesn't include any intersection with valid vmalloc area
3008 long vwrite(char *buf
, char *addr
, unsigned long count
)
3010 struct vmap_area
*va
;
3011 struct vm_struct
*vm
;
3013 unsigned long n
, buflen
;
3016 /* Don't allow overflow */
3017 if ((unsigned long) addr
+ count
< count
)
3018 count
= -(unsigned long) addr
;
3021 spin_lock(&vmap_area_lock
);
3022 list_for_each_entry(va
, &vmap_area_list
, list
) {
3030 vaddr
= (char *) vm
->addr
;
3031 if (addr
>= vaddr
+ get_vm_area_size(vm
))
3033 while (addr
< vaddr
) {
3040 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
3043 if (!(vm
->flags
& VM_IOREMAP
)) {
3044 aligned_vwrite(buf
, addr
, n
);
3052 spin_unlock(&vmap_area_lock
);
3059 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3060 * @vma: vma to cover
3061 * @uaddr: target user address to start at
3062 * @kaddr: virtual address of vmalloc kernel memory
3063 * @size: size of map area
3065 * Returns: 0 for success, -Exxx on failure
3067 * This function checks that @kaddr is a valid vmalloc'ed area,
3068 * and that it is big enough to cover the range starting at
3069 * @uaddr in @vma. Will return failure if that criteria isn't
3072 * Similar to remap_pfn_range() (see mm/memory.c)
3074 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3075 void *kaddr
, unsigned long size
)
3077 struct vm_struct
*area
;
3079 size
= PAGE_ALIGN(size
);
3081 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3084 area
= find_vm_area(kaddr
);
3088 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3091 if (kaddr
+ size
> area
->addr
+ get_vm_area_size(area
))
3095 struct page
*page
= vmalloc_to_page(kaddr
);
3098 ret
= vm_insert_page(vma
, uaddr
, page
);
3107 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3111 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
3114 * remap_vmalloc_range - map vmalloc pages to userspace
3115 * @vma: vma to cover (map full range of vma)
3116 * @addr: vmalloc memory
3117 * @pgoff: number of pages into addr before first page to map
3119 * Returns: 0 for success, -Exxx on failure
3121 * This function checks that addr is a valid vmalloc'ed area, and
3122 * that it is big enough to cover the vma. Will return failure if
3123 * that criteria isn't met.
3125 * Similar to remap_pfn_range() (see mm/memory.c)
3127 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3128 unsigned long pgoff
)
3130 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3131 addr
+ (pgoff
<< PAGE_SHIFT
),
3132 vma
->vm_end
- vma
->vm_start
);
3134 EXPORT_SYMBOL(remap_vmalloc_range
);
3137 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3140 * The purpose of this function is to make sure the vmalloc area
3141 * mappings are identical in all page-tables in the system.
3143 void __weak
vmalloc_sync_all(void)
3148 static int f(pte_t
*pte
, unsigned long addr
, void *data
)
3160 * alloc_vm_area - allocate a range of kernel address space
3161 * @size: size of the area
3162 * @ptes: returns the PTEs for the address space
3164 * Returns: NULL on failure, vm_struct on success
3166 * This function reserves a range of kernel address space, and
3167 * allocates pagetables to map that range. No actual mappings
3170 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3171 * allocated for the VM area are returned.
3173 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
3175 struct vm_struct
*area
;
3177 area
= get_vm_area_caller(size
, VM_IOREMAP
,
3178 __builtin_return_address(0));
3183 * This ensures that page tables are constructed for this region
3184 * of kernel virtual address space and mapped into init_mm.
3186 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3187 size
, f
, ptes
? &ptes
: NULL
)) {
3194 EXPORT_SYMBOL_GPL(alloc_vm_area
);
3196 void free_vm_area(struct vm_struct
*area
)
3198 struct vm_struct
*ret
;
3199 ret
= remove_vm_area(area
->addr
);
3200 BUG_ON(ret
!= area
);
3203 EXPORT_SYMBOL_GPL(free_vm_area
);
3206 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3208 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3212 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3213 * @addr: target address
3215 * Returns: vmap_area if it is found. If there is no such area
3216 * the first highest(reverse order) vmap_area is returned
3217 * i.e. va->va_start < addr && va->va_end < addr or NULL
3218 * if there are no any areas before @addr.
3220 static struct vmap_area
*
3221 pvm_find_va_enclose_addr(unsigned long addr
)
3223 struct vmap_area
*va
, *tmp
;
3226 n
= free_vmap_area_root
.rb_node
;
3230 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3231 if (tmp
->va_start
<= addr
) {
3233 if (tmp
->va_end
>= addr
)
3246 * pvm_determine_end_from_reverse - find the highest aligned address
3247 * of free block below VMALLOC_END
3249 * in - the VA we start the search(reverse order);
3250 * out - the VA with the highest aligned end address.
3252 * Returns: determined end address within vmap_area
3254 static unsigned long
3255 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3257 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3261 list_for_each_entry_from_reverse((*va
),
3262 &free_vmap_area_list
, list
) {
3263 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3264 if ((*va
)->va_start
< addr
)
3273 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3274 * @offsets: array containing offset of each area
3275 * @sizes: array containing size of each area
3276 * @nr_vms: the number of areas to allocate
3277 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3279 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3280 * vm_structs on success, %NULL on failure
3282 * Percpu allocator wants to use congruent vm areas so that it can
3283 * maintain the offsets among percpu areas. This function allocates
3284 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3285 * be scattered pretty far, distance between two areas easily going up
3286 * to gigabytes. To avoid interacting with regular vmallocs, these
3287 * areas are allocated from top.
3289 * Despite its complicated look, this allocator is rather simple. It
3290 * does everything top-down and scans free blocks from the end looking
3291 * for matching base. While scanning, if any of the areas do not fit the
3292 * base address is pulled down to fit the area. Scanning is repeated till
3293 * all the areas fit and then all necessary data structures are inserted
3294 * and the result is returned.
3296 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3297 const size_t *sizes
, int nr_vms
,
3300 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3301 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3302 struct vmap_area
**vas
, *va
;
3303 struct vm_struct
**vms
;
3304 int area
, area2
, last_area
, term_area
;
3305 unsigned long base
, start
, size
, end
, last_end
;
3306 bool purged
= false;
3309 /* verify parameters and allocate data structures */
3310 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3311 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3312 start
= offsets
[area
];
3313 end
= start
+ sizes
[area
];
3315 /* is everything aligned properly? */
3316 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3317 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3319 /* detect the area with the highest address */
3320 if (start
> offsets
[last_area
])
3323 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3324 unsigned long start2
= offsets
[area2
];
3325 unsigned long end2
= start2
+ sizes
[area2
];
3327 BUG_ON(start2
< end
&& start
< end2
);
3330 last_end
= offsets
[last_area
] + sizes
[last_area
];
3332 if (vmalloc_end
- vmalloc_start
< last_end
) {
3337 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3338 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3342 for (area
= 0; area
< nr_vms
; area
++) {
3343 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3344 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3345 if (!vas
[area
] || !vms
[area
])
3349 spin_lock(&free_vmap_area_lock
);
3351 /* start scanning - we scan from the top, begin with the last area */
3352 area
= term_area
= last_area
;
3353 start
= offsets
[area
];
3354 end
= start
+ sizes
[area
];
3356 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3357 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3361 * base might have underflowed, add last_end before
3364 if (base
+ last_end
< vmalloc_start
+ last_end
)
3368 * Fitting base has not been found.
3374 * If required width exeeds current VA block, move
3375 * base downwards and then recheck.
3377 if (base
+ end
> va
->va_end
) {
3378 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3384 * If this VA does not fit, move base downwards and recheck.
3386 if (base
+ start
< va
->va_start
) {
3387 va
= node_to_va(rb_prev(&va
->rb_node
));
3388 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3394 * This area fits, move on to the previous one. If
3395 * the previous one is the terminal one, we're done.
3397 area
= (area
+ nr_vms
- 1) % nr_vms
;
3398 if (area
== term_area
)
3401 start
= offsets
[area
];
3402 end
= start
+ sizes
[area
];
3403 va
= pvm_find_va_enclose_addr(base
+ end
);
3406 /* we've found a fitting base, insert all va's */
3407 for (area
= 0; area
< nr_vms
; area
++) {
3410 start
= base
+ offsets
[area
];
3413 va
= pvm_find_va_enclose_addr(start
);
3414 if (WARN_ON_ONCE(va
== NULL
))
3415 /* It is a BUG(), but trigger recovery instead. */
3418 type
= classify_va_fit_type(va
, start
, size
);
3419 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
3420 /* It is a BUG(), but trigger recovery instead. */
3423 ret
= adjust_va_to_fit_type(va
, start
, size
, type
);
3427 /* Allocated area. */
3429 va
->va_start
= start
;
3430 va
->va_end
= start
+ size
;
3433 spin_unlock(&free_vmap_area_lock
);
3435 /* insert all vm's */
3436 spin_lock(&vmap_area_lock
);
3437 for (area
= 0; area
< nr_vms
; area
++) {
3438 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3440 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3443 spin_unlock(&vmap_area_lock
);
3445 /* populate the shadow space outside of the lock */
3446 for (area
= 0; area
< nr_vms
; area
++) {
3447 /* assume success here */
3448 kasan_populate_vmalloc(sizes
[area
], vms
[area
]);
3456 * Remove previously allocated areas. There is no
3457 * need in removing these areas from the busy tree,
3458 * because they are inserted only on the final step
3459 * and when pcpu_get_vm_areas() is success.
3462 merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3463 &free_vmap_area_list
);
3468 spin_unlock(&free_vmap_area_lock
);
3470 purge_vmap_area_lazy();
3473 /* Before "retry", check if we recover. */
3474 for (area
= 0; area
< nr_vms
; area
++) {
3478 vas
[area
] = kmem_cache_zalloc(
3479 vmap_area_cachep
, GFP_KERNEL
);
3488 for (area
= 0; area
< nr_vms
; area
++) {
3490 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
3501 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3502 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3503 * @nr_vms: the number of allocated areas
3505 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3507 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
3511 for (i
= 0; i
< nr_vms
; i
++)
3512 free_vm_area(vms
[i
]);
3515 #endif /* CONFIG_SMP */
3517 #ifdef CONFIG_PROC_FS
3518 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3519 __acquires(&vmap_purge_lock
)
3520 __acquires(&vmap_area_lock
)
3522 mutex_lock(&vmap_purge_lock
);
3523 spin_lock(&vmap_area_lock
);
3525 return seq_list_start(&vmap_area_list
, *pos
);
3528 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3530 return seq_list_next(p
, &vmap_area_list
, pos
);
3533 static void s_stop(struct seq_file
*m
, void *p
)
3534 __releases(&vmap_purge_lock
)
3535 __releases(&vmap_area_lock
)
3537 mutex_unlock(&vmap_purge_lock
);
3538 spin_unlock(&vmap_area_lock
);
3541 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
3543 if (IS_ENABLED(CONFIG_NUMA
)) {
3544 unsigned int nr
, *counters
= m
->private;
3549 if (v
->flags
& VM_UNINITIALIZED
)
3551 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3554 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
3556 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
3557 counters
[page_to_nid(v
->pages
[nr
])]++;
3559 for_each_node_state(nr
, N_HIGH_MEMORY
)
3561 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
3565 static void show_purge_info(struct seq_file
*m
)
3567 struct llist_node
*head
;
3568 struct vmap_area
*va
;
3570 head
= READ_ONCE(vmap_purge_list
.first
);
3574 llist_for_each_entry(va
, head
, purge_list
) {
3575 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3576 (void *)va
->va_start
, (void *)va
->va_end
,
3577 va
->va_end
- va
->va_start
);
3581 static int s_show(struct seq_file
*m
, void *p
)
3583 struct vmap_area
*va
;
3584 struct vm_struct
*v
;
3586 va
= list_entry(p
, struct vmap_area
, list
);
3589 * s_show can encounter race with remove_vm_area, !vm on behalf
3590 * of vmap area is being tear down or vm_map_ram allocation.
3593 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
3594 (void *)va
->va_start
, (void *)va
->va_end
,
3595 va
->va_end
- va
->va_start
);
3602 seq_printf(m
, "0x%pK-0x%pK %7ld",
3603 v
->addr
, v
->addr
+ v
->size
, v
->size
);
3606 seq_printf(m
, " %pS", v
->caller
);
3609 seq_printf(m
, " pages=%d", v
->nr_pages
);
3612 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
3614 if (v
->flags
& VM_IOREMAP
)
3615 seq_puts(m
, " ioremap");
3617 if (v
->flags
& VM_ALLOC
)
3618 seq_puts(m
, " vmalloc");
3620 if (v
->flags
& VM_MAP
)
3621 seq_puts(m
, " vmap");
3623 if (v
->flags
& VM_USERMAP
)
3624 seq_puts(m
, " user");
3626 if (v
->flags
& VM_DMA_COHERENT
)
3627 seq_puts(m
, " dma-coherent");
3629 if (is_vmalloc_addr(v
->pages
))
3630 seq_puts(m
, " vpages");
3632 show_numa_info(m
, v
);
3636 * As a final step, dump "unpurged" areas. Note,
3637 * that entire "/proc/vmallocinfo" output will not
3638 * be address sorted, because the purge list is not
3641 if (list_is_last(&va
->list
, &vmap_area_list
))
3647 static const struct seq_operations vmalloc_op
= {
3654 static int __init
proc_vmalloc_init(void)
3656 if (IS_ENABLED(CONFIG_NUMA
))
3657 proc_create_seq_private("vmallocinfo", 0400, NULL
,
3659 nr_node_ids
* sizeof(unsigned int), NULL
);
3661 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
3664 module_init(proc_vmalloc_init
);