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
);
132 * unmap_kernel_range_noflush - unmap kernel VM area
133 * @addr: start of the VM area to unmap
134 * @size: size of the VM area to unmap
136 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
137 * should have been allocated using get_vm_area() and its friends.
140 * This function does NOT do any cache flushing. The caller is responsible
141 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
142 * function and flush_tlb_kernel_range() after.
144 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
146 unsigned long end
= addr
+ size
;
151 pgd
= pgd_offset_k(addr
);
153 next
= pgd_addr_end(addr
, end
);
154 if (pgd_none_or_clear_bad(pgd
))
156 vunmap_p4d_range(pgd
, addr
, next
);
157 } while (pgd
++, addr
= next
, addr
!= end
);
160 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
161 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
166 * nr is a running index into the array which helps higher level
167 * callers keep track of where we're up to.
170 pte
= pte_alloc_kernel(pmd
, addr
);
174 struct page
*page
= pages
[*nr
];
176 if (WARN_ON(!pte_none(*pte
)))
180 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
182 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
186 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
187 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
192 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
196 next
= pmd_addr_end(addr
, end
);
197 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
199 } while (pmd
++, addr
= next
, addr
!= end
);
203 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
204 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
209 pud
= pud_alloc(&init_mm
, p4d
, addr
);
213 next
= pud_addr_end(addr
, end
);
214 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
216 } while (pud
++, addr
= next
, addr
!= end
);
220 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
221 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
226 p4d
= p4d_alloc(&init_mm
, pgd
, addr
);
230 next
= p4d_addr_end(addr
, end
);
231 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
))
233 } while (p4d
++, addr
= next
, addr
!= end
);
238 * map_kernel_range_noflush - map kernel VM area with the specified pages
239 * @addr: start of the VM area to map
240 * @size: size of the VM area to map
241 * @prot: page protection flags to use
242 * @pages: pages to map
244 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
245 * have been allocated using get_vm_area() and its friends.
248 * This function does NOT do any cache flushing. The caller is responsible for
249 * calling flush_cache_vmap() on to-be-mapped areas before calling this
253 * 0 on success, -errno on failure.
255 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
256 pgprot_t prot
, struct page
**pages
)
258 unsigned long end
= addr
+ size
;
265 pgd
= pgd_offset_k(addr
);
267 next
= pgd_addr_end(addr
, end
);
268 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
);
271 } while (pgd
++, addr
= next
, addr
!= end
);
276 int map_kernel_range(unsigned long start
, unsigned long size
, pgprot_t prot
,
281 ret
= map_kernel_range_noflush(start
, size
, prot
, pages
);
282 flush_cache_vmap(start
, start
+ size
);
286 int is_vmalloc_or_module_addr(const void *x
)
289 * ARM, x86-64 and sparc64 put modules in a special place,
290 * and fall back on vmalloc() if that fails. Others
291 * just put it in the vmalloc space.
293 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
294 unsigned long addr
= (unsigned long)x
;
295 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
298 return is_vmalloc_addr(x
);
302 * Walk a vmap address to the struct page it maps.
304 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
306 unsigned long addr
= (unsigned long) vmalloc_addr
;
307 struct page
*page
= NULL
;
308 pgd_t
*pgd
= pgd_offset_k(addr
);
315 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
316 * architectures that do not vmalloc module space
318 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
322 p4d
= p4d_offset(pgd
, addr
);
325 pud
= pud_offset(p4d
, addr
);
328 * Don't dereference bad PUD or PMD (below) entries. This will also
329 * identify huge mappings, which we may encounter on architectures
330 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
331 * identified as vmalloc addresses by is_vmalloc_addr(), but are
332 * not [unambiguously] associated with a struct page, so there is
333 * no correct value to return for them.
335 WARN_ON_ONCE(pud_bad(*pud
));
336 if (pud_none(*pud
) || pud_bad(*pud
))
338 pmd
= pmd_offset(pud
, addr
);
339 WARN_ON_ONCE(pmd_bad(*pmd
));
340 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
343 ptep
= pte_offset_map(pmd
, addr
);
345 if (pte_present(pte
))
346 page
= pte_page(pte
);
350 EXPORT_SYMBOL(vmalloc_to_page
);
353 * Map a vmalloc()-space virtual address to the physical page frame number.
355 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
357 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
359 EXPORT_SYMBOL(vmalloc_to_pfn
);
362 /*** Global kva allocator ***/
364 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
365 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
368 static DEFINE_SPINLOCK(vmap_area_lock
);
369 static DEFINE_SPINLOCK(free_vmap_area_lock
);
370 /* Export for kexec only */
371 LIST_HEAD(vmap_area_list
);
372 static LLIST_HEAD(vmap_purge_list
);
373 static struct rb_root vmap_area_root
= RB_ROOT
;
374 static bool vmap_initialized __read_mostly
;
377 * This kmem_cache is used for vmap_area objects. Instead of
378 * allocating from slab we reuse an object from this cache to
379 * make things faster. Especially in "no edge" splitting of
382 static struct kmem_cache
*vmap_area_cachep
;
385 * This linked list is used in pair with free_vmap_area_root.
386 * It gives O(1) access to prev/next to perform fast coalescing.
388 static LIST_HEAD(free_vmap_area_list
);
391 * This augment red-black tree represents the free vmap space.
392 * All vmap_area objects in this tree are sorted by va->va_start
393 * address. It is used for allocation and merging when a vmap
394 * object is released.
396 * Each vmap_area node contains a maximum available free block
397 * of its sub-tree, right or left. Therefore it is possible to
398 * find a lowest match of free area.
400 static struct rb_root free_vmap_area_root
= RB_ROOT
;
403 * Preload a CPU with one object for "no edge" split case. The
404 * aim is to get rid of allocations from the atomic context, thus
405 * to use more permissive allocation masks.
407 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
409 static __always_inline
unsigned long
410 va_size(struct vmap_area
*va
)
412 return (va
->va_end
- va
->va_start
);
415 static __always_inline
unsigned long
416 get_subtree_max_size(struct rb_node
*node
)
418 struct vmap_area
*va
;
420 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
421 return va
? va
->subtree_max_size
: 0;
425 * Gets called when remove the node and rotate.
427 static __always_inline
unsigned long
428 compute_subtree_max_size(struct vmap_area
*va
)
430 return max3(va_size(va
),
431 get_subtree_max_size(va
->rb_node
.rb_left
),
432 get_subtree_max_size(va
->rb_node
.rb_right
));
435 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
436 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
438 static void purge_vmap_area_lazy(void);
439 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
440 static unsigned long lazy_max_pages(void);
442 static atomic_long_t nr_vmalloc_pages
;
444 unsigned long vmalloc_nr_pages(void)
446 return atomic_long_read(&nr_vmalloc_pages
);
449 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
451 struct rb_node
*n
= vmap_area_root
.rb_node
;
454 struct vmap_area
*va
;
456 va
= rb_entry(n
, struct vmap_area
, rb_node
);
457 if (addr
< va
->va_start
)
459 else if (addr
>= va
->va_end
)
469 * This function returns back addresses of parent node
470 * and its left or right link for further processing.
472 static __always_inline
struct rb_node
**
473 find_va_links(struct vmap_area
*va
,
474 struct rb_root
*root
, struct rb_node
*from
,
475 struct rb_node
**parent
)
477 struct vmap_area
*tmp_va
;
478 struct rb_node
**link
;
481 link
= &root
->rb_node
;
482 if (unlikely(!*link
)) {
491 * Go to the bottom of the tree. When we hit the last point
492 * we end up with parent rb_node and correct direction, i name
493 * it link, where the new va->rb_node will be attached to.
496 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
499 * During the traversal we also do some sanity check.
500 * Trigger the BUG() if there are sides(left/right)
503 if (va
->va_start
< tmp_va
->va_end
&&
504 va
->va_end
<= tmp_va
->va_start
)
505 link
= &(*link
)->rb_left
;
506 else if (va
->va_end
> tmp_va
->va_start
&&
507 va
->va_start
>= tmp_va
->va_end
)
508 link
= &(*link
)->rb_right
;
513 *parent
= &tmp_va
->rb_node
;
517 static __always_inline
struct list_head
*
518 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
520 struct list_head
*list
;
522 if (unlikely(!parent
))
524 * The red-black tree where we try to find VA neighbors
525 * before merging or inserting is empty, i.e. it means
526 * there is no free vmap space. Normally it does not
527 * happen but we handle this case anyway.
531 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
532 return (&parent
->rb_right
== link
? list
->next
: list
);
535 static __always_inline
void
536 link_va(struct vmap_area
*va
, struct rb_root
*root
,
537 struct rb_node
*parent
, struct rb_node
**link
, struct list_head
*head
)
540 * VA is still not in the list, but we can
541 * identify its future previous list_head node.
543 if (likely(parent
)) {
544 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
545 if (&parent
->rb_right
!= link
)
549 /* Insert to the rb-tree */
550 rb_link_node(&va
->rb_node
, parent
, link
);
551 if (root
== &free_vmap_area_root
) {
553 * Some explanation here. Just perform simple insertion
554 * to the tree. We do not set va->subtree_max_size to
555 * its current size before calling rb_insert_augmented().
556 * It is because of we populate the tree from the bottom
557 * to parent levels when the node _is_ in the tree.
559 * Therefore we set subtree_max_size to zero after insertion,
560 * to let __augment_tree_propagate_from() puts everything to
561 * the correct order later on.
563 rb_insert_augmented(&va
->rb_node
,
564 root
, &free_vmap_area_rb_augment_cb
);
565 va
->subtree_max_size
= 0;
567 rb_insert_color(&va
->rb_node
, root
);
570 /* Address-sort this list */
571 list_add(&va
->list
, head
);
574 static __always_inline
void
575 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
577 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
580 if (root
== &free_vmap_area_root
)
581 rb_erase_augmented(&va
->rb_node
,
582 root
, &free_vmap_area_rb_augment_cb
);
584 rb_erase(&va
->rb_node
, root
);
587 RB_CLEAR_NODE(&va
->rb_node
);
590 #if DEBUG_AUGMENT_PROPAGATE_CHECK
592 augment_tree_propagate_check(struct rb_node
*n
)
594 struct vmap_area
*va
;
595 struct rb_node
*node
;
602 va
= rb_entry(n
, struct vmap_area
, rb_node
);
603 size
= va
->subtree_max_size
;
607 va
= rb_entry(node
, struct vmap_area
, rb_node
);
609 if (get_subtree_max_size(node
->rb_left
) == size
) {
610 node
= node
->rb_left
;
612 if (va_size(va
) == size
) {
617 node
= node
->rb_right
;
622 va
= rb_entry(n
, struct vmap_area
, rb_node
);
623 pr_emerg("tree is corrupted: %lu, %lu\n",
624 va_size(va
), va
->subtree_max_size
);
627 augment_tree_propagate_check(n
->rb_left
);
628 augment_tree_propagate_check(n
->rb_right
);
633 * This function populates subtree_max_size from bottom to upper
634 * levels starting from VA point. The propagation must be done
635 * when VA size is modified by changing its va_start/va_end. Or
636 * in case of newly inserting of VA to the tree.
638 * It means that __augment_tree_propagate_from() must be called:
639 * - After VA has been inserted to the tree(free path);
640 * - After VA has been shrunk(allocation path);
641 * - After VA has been increased(merging path).
643 * Please note that, it does not mean that upper parent nodes
644 * and their subtree_max_size are recalculated all the time up
653 * For example if we modify the node 4, shrinking it to 2, then
654 * no any modification is required. If we shrink the node 2 to 1
655 * its subtree_max_size is updated only, and set to 1. If we shrink
656 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
659 static __always_inline
void
660 augment_tree_propagate_from(struct vmap_area
*va
)
662 struct rb_node
*node
= &va
->rb_node
;
663 unsigned long new_va_sub_max_size
;
666 va
= rb_entry(node
, struct vmap_area
, rb_node
);
667 new_va_sub_max_size
= compute_subtree_max_size(va
);
670 * If the newly calculated maximum available size of the
671 * subtree is equal to the current one, then it means that
672 * the tree is propagated correctly. So we have to stop at
673 * this point to save cycles.
675 if (va
->subtree_max_size
== new_va_sub_max_size
)
678 va
->subtree_max_size
= new_va_sub_max_size
;
679 node
= rb_parent(&va
->rb_node
);
682 #if DEBUG_AUGMENT_PROPAGATE_CHECK
683 augment_tree_propagate_check(free_vmap_area_root
.rb_node
);
688 insert_vmap_area(struct vmap_area
*va
,
689 struct rb_root
*root
, struct list_head
*head
)
691 struct rb_node
**link
;
692 struct rb_node
*parent
;
694 link
= find_va_links(va
, root
, NULL
, &parent
);
695 link_va(va
, root
, parent
, link
, head
);
699 insert_vmap_area_augment(struct vmap_area
*va
,
700 struct rb_node
*from
, struct rb_root
*root
,
701 struct list_head
*head
)
703 struct rb_node
**link
;
704 struct rb_node
*parent
;
707 link
= find_va_links(va
, NULL
, from
, &parent
);
709 link
= find_va_links(va
, root
, NULL
, &parent
);
711 link_va(va
, root
, parent
, link
, head
);
712 augment_tree_propagate_from(va
);
716 * Merge de-allocated chunk of VA memory with previous
717 * and next free blocks. If coalesce is not done a new
718 * free area is inserted. If VA has been merged, it is
721 static __always_inline
struct vmap_area
*
722 merge_or_add_vmap_area(struct vmap_area
*va
,
723 struct rb_root
*root
, struct list_head
*head
)
725 struct vmap_area
*sibling
;
726 struct list_head
*next
;
727 struct rb_node
**link
;
728 struct rb_node
*parent
;
732 * Find a place in the tree where VA potentially will be
733 * inserted, unless it is merged with its sibling/siblings.
735 link
= find_va_links(va
, root
, NULL
, &parent
);
738 * Get next node of VA to check if merging can be done.
740 next
= get_va_next_sibling(parent
, link
);
741 if (unlikely(next
== NULL
))
747 * |<------VA------>|<-----Next----->|
752 sibling
= list_entry(next
, struct vmap_area
, list
);
753 if (sibling
->va_start
== va
->va_end
) {
754 sibling
->va_start
= va
->va_start
;
756 /* Check and update the tree if needed. */
757 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. */
771 * |<-----Prev----->|<------VA------>|
775 if (next
->prev
!= head
) {
776 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
777 if (sibling
->va_end
== va
->va_start
) {
778 sibling
->va_end
= va
->va_end
;
780 /* Check and update the tree if needed. */
781 augment_tree_propagate_from(sibling
);
786 /* Free vmap_area object. */
787 kmem_cache_free(vmap_area_cachep
, va
);
789 /* Point to the new merged area. */
797 link_va(va
, root
, parent
, link
, head
);
798 augment_tree_propagate_from(va
);
804 static __always_inline
bool
805 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
806 unsigned long align
, unsigned long vstart
)
808 unsigned long nva_start_addr
;
810 if (va
->va_start
> vstart
)
811 nva_start_addr
= ALIGN(va
->va_start
, align
);
813 nva_start_addr
= ALIGN(vstart
, align
);
815 /* Can be overflowed due to big size or alignment. */
816 if (nva_start_addr
+ size
< nva_start_addr
||
817 nva_start_addr
< vstart
)
820 return (nva_start_addr
+ size
<= va
->va_end
);
824 * Find the first free block(lowest start address) in the tree,
825 * that will accomplish the request corresponding to passing
828 static __always_inline
struct vmap_area
*
829 find_vmap_lowest_match(unsigned long size
,
830 unsigned long align
, unsigned long vstart
)
832 struct vmap_area
*va
;
833 struct rb_node
*node
;
834 unsigned long length
;
836 /* Start from the root. */
837 node
= free_vmap_area_root
.rb_node
;
839 /* Adjust the search size for alignment overhead. */
840 length
= size
+ align
- 1;
843 va
= rb_entry(node
, struct vmap_area
, rb_node
);
845 if (get_subtree_max_size(node
->rb_left
) >= length
&&
846 vstart
< va
->va_start
) {
847 node
= node
->rb_left
;
849 if (is_within_this_va(va
, size
, align
, vstart
))
853 * Does not make sense to go deeper towards the right
854 * sub-tree if it does not have a free block that is
855 * equal or bigger to the requested search length.
857 if (get_subtree_max_size(node
->rb_right
) >= length
) {
858 node
= node
->rb_right
;
863 * OK. We roll back and find the first right sub-tree,
864 * that will satisfy the search criteria. It can happen
865 * only once due to "vstart" restriction.
867 while ((node
= rb_parent(node
))) {
868 va
= rb_entry(node
, struct vmap_area
, rb_node
);
869 if (is_within_this_va(va
, size
, align
, vstart
))
872 if (get_subtree_max_size(node
->rb_right
) >= length
&&
873 vstart
<= va
->va_start
) {
874 node
= node
->rb_right
;
884 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
885 #include <linux/random.h>
887 static struct vmap_area
*
888 find_vmap_lowest_linear_match(unsigned long size
,
889 unsigned long align
, unsigned long vstart
)
891 struct vmap_area
*va
;
893 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
894 if (!is_within_this_va(va
, size
, align
, vstart
))
904 find_vmap_lowest_match_check(unsigned long size
)
906 struct vmap_area
*va_1
, *va_2
;
907 unsigned long vstart
;
910 get_random_bytes(&rnd
, sizeof(rnd
));
911 vstart
= VMALLOC_START
+ rnd
;
913 va_1
= find_vmap_lowest_match(size
, 1, vstart
);
914 va_2
= find_vmap_lowest_linear_match(size
, 1, vstart
);
917 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
924 FL_FIT_TYPE
= 1, /* full fit */
925 LE_FIT_TYPE
= 2, /* left edge fit */
926 RE_FIT_TYPE
= 3, /* right edge fit */
927 NE_FIT_TYPE
= 4 /* no edge fit */
930 static __always_inline
enum fit_type
931 classify_va_fit_type(struct vmap_area
*va
,
932 unsigned long nva_start_addr
, unsigned long size
)
936 /* Check if it is within VA. */
937 if (nva_start_addr
< va
->va_start
||
938 nva_start_addr
+ size
> va
->va_end
)
942 if (va
->va_start
== nva_start_addr
) {
943 if (va
->va_end
== nva_start_addr
+ size
)
947 } else if (va
->va_end
== nva_start_addr
+ size
) {
956 static __always_inline
int
957 adjust_va_to_fit_type(struct vmap_area
*va
,
958 unsigned long nva_start_addr
, unsigned long size
,
961 struct vmap_area
*lva
= NULL
;
963 if (type
== FL_FIT_TYPE
) {
965 * No need to split VA, it fully fits.
971 unlink_va(va
, &free_vmap_area_root
);
972 kmem_cache_free(vmap_area_cachep
, va
);
973 } else if (type
== LE_FIT_TYPE
) {
975 * Split left edge of fit VA.
981 va
->va_start
+= size
;
982 } else if (type
== RE_FIT_TYPE
) {
984 * Split right edge of fit VA.
990 va
->va_end
= nva_start_addr
;
991 } else if (type
== NE_FIT_TYPE
) {
993 * Split no edge of fit VA.
999 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1000 if (unlikely(!lva
)) {
1002 * For percpu allocator we do not do any pre-allocation
1003 * and leave it as it is. The reason is it most likely
1004 * never ends up with NE_FIT_TYPE splitting. In case of
1005 * percpu allocations offsets and sizes are aligned to
1006 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1007 * are its main fitting cases.
1009 * There are a few exceptions though, as an example it is
1010 * a first allocation (early boot up) when we have "one"
1011 * big free space that has to be split.
1013 * Also we can hit this path in case of regular "vmap"
1014 * allocations, if "this" current CPU was not preloaded.
1015 * See the comment in alloc_vmap_area() why. If so, then
1016 * GFP_NOWAIT is used instead to get an extra object for
1017 * split purpose. That is rare and most time does not
1020 * What happens if an allocation gets failed. Basically,
1021 * an "overflow" path is triggered to purge lazily freed
1022 * areas to free some memory, then, the "retry" path is
1023 * triggered to repeat one more time. See more details
1024 * in alloc_vmap_area() function.
1026 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1032 * Build the remainder.
1034 lva
->va_start
= va
->va_start
;
1035 lva
->va_end
= nva_start_addr
;
1038 * Shrink this VA to remaining size.
1040 va
->va_start
= nva_start_addr
+ size
;
1045 if (type
!= FL_FIT_TYPE
) {
1046 augment_tree_propagate_from(va
);
1048 if (lva
) /* type == NE_FIT_TYPE */
1049 insert_vmap_area_augment(lva
, &va
->rb_node
,
1050 &free_vmap_area_root
, &free_vmap_area_list
);
1057 * Returns a start address of the newly allocated area, if success.
1058 * Otherwise a vend is returned that indicates failure.
1060 static __always_inline
unsigned long
1061 __alloc_vmap_area(unsigned long size
, unsigned long align
,
1062 unsigned long vstart
, unsigned long vend
)
1064 unsigned long nva_start_addr
;
1065 struct vmap_area
*va
;
1069 va
= find_vmap_lowest_match(size
, align
, vstart
);
1073 if (va
->va_start
> vstart
)
1074 nva_start_addr
= ALIGN(va
->va_start
, align
);
1076 nva_start_addr
= ALIGN(vstart
, align
);
1078 /* Check the "vend" restriction. */
1079 if (nva_start_addr
+ size
> vend
)
1082 /* Classify what we have found. */
1083 type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1084 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
1087 /* Update the free vmap_area. */
1088 ret
= adjust_va_to_fit_type(va
, nva_start_addr
, size
, type
);
1092 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1093 find_vmap_lowest_match_check(size
);
1096 return nva_start_addr
;
1100 * Free a region of KVA allocated by alloc_vmap_area
1102 static void free_vmap_area(struct vmap_area
*va
)
1105 * Remove from the busy tree/list.
1107 spin_lock(&vmap_area_lock
);
1108 unlink_va(va
, &vmap_area_root
);
1109 spin_unlock(&vmap_area_lock
);
1112 * Insert/Merge it back to the free tree/list.
1114 spin_lock(&free_vmap_area_lock
);
1115 merge_or_add_vmap_area(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1116 spin_unlock(&free_vmap_area_lock
);
1120 * Allocate a region of KVA of the specified size and alignment, within the
1123 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1124 unsigned long align
,
1125 unsigned long vstart
, unsigned long vend
,
1126 int node
, gfp_t gfp_mask
)
1128 struct vmap_area
*va
, *pva
;
1134 BUG_ON(offset_in_page(size
));
1135 BUG_ON(!is_power_of_2(align
));
1137 if (unlikely(!vmap_initialized
))
1138 return ERR_PTR(-EBUSY
);
1141 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1143 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1145 return ERR_PTR(-ENOMEM
);
1148 * Only scan the relevant parts containing pointers to other objects
1149 * to avoid false negatives.
1151 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1155 * Preload this CPU with one extra vmap_area object. It is used
1156 * when fit type of free area is NE_FIT_TYPE. Please note, it
1157 * does not guarantee that an allocation occurs on a CPU that
1158 * is preloaded, instead we minimize the case when it is not.
1159 * It can happen because of cpu migration, because there is a
1160 * race until the below spinlock is taken.
1162 * The preload is done in non-atomic context, thus it allows us
1163 * to use more permissive allocation masks to be more stable under
1164 * low memory condition and high memory pressure. In rare case,
1165 * if not preloaded, GFP_NOWAIT is used.
1167 * Set "pva" to NULL here, because of "retry" path.
1171 if (!this_cpu_read(ne_fit_preload_node
))
1173 * Even if it fails we do not really care about that.
1174 * Just proceed as it is. If needed "overflow" path
1175 * will refill the cache we allocate from.
1177 pva
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1179 spin_lock(&free_vmap_area_lock
);
1181 if (pva
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, pva
))
1182 kmem_cache_free(vmap_area_cachep
, pva
);
1185 * If an allocation fails, the "vend" address is
1186 * returned. Therefore trigger the overflow path.
1188 addr
= __alloc_vmap_area(size
, align
, vstart
, vend
);
1189 spin_unlock(&free_vmap_area_lock
);
1191 if (unlikely(addr
== vend
))
1194 va
->va_start
= addr
;
1195 va
->va_end
= addr
+ size
;
1199 spin_lock(&vmap_area_lock
);
1200 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1201 spin_unlock(&vmap_area_lock
);
1203 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1204 BUG_ON(va
->va_start
< vstart
);
1205 BUG_ON(va
->va_end
> vend
);
1207 ret
= kasan_populate_vmalloc(addr
, size
);
1210 return ERR_PTR(ret
);
1217 purge_vmap_area_lazy();
1222 if (gfpflags_allow_blocking(gfp_mask
)) {
1223 unsigned long freed
= 0;
1224 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1231 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1232 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1235 kmem_cache_free(vmap_area_cachep
, va
);
1236 return ERR_PTR(-EBUSY
);
1239 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1241 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1243 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1245 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1247 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1249 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1252 * lazy_max_pages is the maximum amount of virtual address space we gather up
1253 * before attempting to purge with a TLB flush.
1255 * There is a tradeoff here: a larger number will cover more kernel page tables
1256 * and take slightly longer to purge, but it will linearly reduce the number of
1257 * global TLB flushes that must be performed. It would seem natural to scale
1258 * this number up linearly with the number of CPUs (because vmapping activity
1259 * could also scale linearly with the number of CPUs), however it is likely
1260 * that in practice, workloads might be constrained in other ways that mean
1261 * vmap activity will not scale linearly with CPUs. Also, I want to be
1262 * conservative and not introduce a big latency on huge systems, so go with
1263 * a less aggressive log scale. It will still be an improvement over the old
1264 * code, and it will be simple to change the scale factor if we find that it
1265 * becomes a problem on bigger systems.
1267 static unsigned long lazy_max_pages(void)
1271 log
= fls(num_online_cpus());
1273 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1276 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1279 * Serialize vmap purging. There is no actual criticial section protected
1280 * by this look, but we want to avoid concurrent calls for performance
1281 * reasons and to make the pcpu_get_vm_areas more deterministic.
1283 static DEFINE_MUTEX(vmap_purge_lock
);
1285 /* for per-CPU blocks */
1286 static void purge_fragmented_blocks_allcpus(void);
1289 * called before a call to iounmap() if the caller wants vm_area_struct's
1290 * immediately freed.
1292 void set_iounmap_nonlazy(void)
1294 atomic_long_set(&vmap_lazy_nr
, lazy_max_pages()+1);
1298 * Purges all lazily-freed vmap areas.
1300 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1302 unsigned long resched_threshold
;
1303 struct llist_node
*valist
;
1304 struct vmap_area
*va
;
1305 struct vmap_area
*n_va
;
1307 lockdep_assert_held(&vmap_purge_lock
);
1309 valist
= llist_del_all(&vmap_purge_list
);
1310 if (unlikely(valist
== NULL
))
1314 * First make sure the mappings are removed from all page-tables
1315 * before they are freed.
1317 vmalloc_sync_unmappings();
1320 * TODO: to calculate a flush range without looping.
1321 * The list can be up to lazy_max_pages() elements.
1323 llist_for_each_entry(va
, valist
, purge_list
) {
1324 if (va
->va_start
< start
)
1325 start
= va
->va_start
;
1326 if (va
->va_end
> end
)
1330 flush_tlb_kernel_range(start
, end
);
1331 resched_threshold
= lazy_max_pages() << 1;
1333 spin_lock(&free_vmap_area_lock
);
1334 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
1335 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1336 unsigned long orig_start
= va
->va_start
;
1337 unsigned long orig_end
= va
->va_end
;
1340 * Finally insert or merge lazily-freed area. It is
1341 * detached and there is no need to "unlink" it from
1344 va
= merge_or_add_vmap_area(va
, &free_vmap_area_root
,
1345 &free_vmap_area_list
);
1347 if (is_vmalloc_or_module_addr((void *)orig_start
))
1348 kasan_release_vmalloc(orig_start
, orig_end
,
1349 va
->va_start
, va
->va_end
);
1351 atomic_long_sub(nr
, &vmap_lazy_nr
);
1353 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1354 cond_resched_lock(&free_vmap_area_lock
);
1356 spin_unlock(&free_vmap_area_lock
);
1361 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1362 * is already purging.
1364 static void try_purge_vmap_area_lazy(void)
1366 if (mutex_trylock(&vmap_purge_lock
)) {
1367 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1368 mutex_unlock(&vmap_purge_lock
);
1373 * Kick off a purge of the outstanding lazy areas.
1375 static void purge_vmap_area_lazy(void)
1377 mutex_lock(&vmap_purge_lock
);
1378 purge_fragmented_blocks_allcpus();
1379 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1380 mutex_unlock(&vmap_purge_lock
);
1384 * Free a vmap area, caller ensuring that the area has been unmapped
1385 * and flush_cache_vunmap had been called for the correct range
1388 static void free_vmap_area_noflush(struct vmap_area
*va
)
1390 unsigned long nr_lazy
;
1392 spin_lock(&vmap_area_lock
);
1393 unlink_va(va
, &vmap_area_root
);
1394 spin_unlock(&vmap_area_lock
);
1396 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1397 PAGE_SHIFT
, &vmap_lazy_nr
);
1399 /* After this point, we may free va at any time */
1400 llist_add(&va
->purge_list
, &vmap_purge_list
);
1402 if (unlikely(nr_lazy
> lazy_max_pages()))
1403 try_purge_vmap_area_lazy();
1407 * Free and unmap a vmap area
1409 static void free_unmap_vmap_area(struct vmap_area
*va
)
1411 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1412 unmap_kernel_range_noflush(va
->va_start
, va
->va_end
- va
->va_start
);
1413 if (debug_pagealloc_enabled_static())
1414 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1416 free_vmap_area_noflush(va
);
1419 static struct vmap_area
*find_vmap_area(unsigned long addr
)
1421 struct vmap_area
*va
;
1423 spin_lock(&vmap_area_lock
);
1424 va
= __find_vmap_area(addr
);
1425 spin_unlock(&vmap_area_lock
);
1430 /*** Per cpu kva allocator ***/
1433 * vmap space is limited especially on 32 bit architectures. Ensure there is
1434 * room for at least 16 percpu vmap blocks per CPU.
1437 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1438 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1439 * instead (we just need a rough idea)
1441 #if BITS_PER_LONG == 32
1442 #define VMALLOC_SPACE (128UL*1024*1024)
1444 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1447 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1448 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1449 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1450 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1451 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1452 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1453 #define VMAP_BBMAP_BITS \
1454 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1455 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1456 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1458 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1460 struct vmap_block_queue
{
1462 struct list_head free
;
1467 struct vmap_area
*va
;
1468 unsigned long free
, dirty
;
1469 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1470 struct list_head free_list
;
1471 struct rcu_head rcu_head
;
1472 struct list_head purge
;
1475 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1476 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1479 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1480 * in the free path. Could get rid of this if we change the API to return a
1481 * "cookie" from alloc, to be passed to free. But no big deal yet.
1483 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
1484 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
1487 * We should probably have a fallback mechanism to allocate virtual memory
1488 * out of partially filled vmap blocks. However vmap block sizing should be
1489 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1493 static unsigned long addr_to_vb_idx(unsigned long addr
)
1495 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1496 addr
/= VMAP_BLOCK_SIZE
;
1500 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1504 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1505 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1506 return (void *)addr
;
1510 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1511 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1512 * @order: how many 2^order pages should be occupied in newly allocated block
1513 * @gfp_mask: flags for the page level allocator
1515 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1517 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1519 struct vmap_block_queue
*vbq
;
1520 struct vmap_block
*vb
;
1521 struct vmap_area
*va
;
1522 unsigned long vb_idx
;
1526 node
= numa_node_id();
1528 vb
= kmalloc_node(sizeof(struct vmap_block
),
1529 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1531 return ERR_PTR(-ENOMEM
);
1533 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1534 VMALLOC_START
, VMALLOC_END
,
1538 return ERR_CAST(va
);
1541 err
= radix_tree_preload(gfp_mask
);
1542 if (unlikely(err
)) {
1545 return ERR_PTR(err
);
1548 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1549 spin_lock_init(&vb
->lock
);
1551 /* At least something should be left free */
1552 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1553 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
1555 vb
->dirty_min
= VMAP_BBMAP_BITS
;
1557 INIT_LIST_HEAD(&vb
->free_list
);
1559 vb_idx
= addr_to_vb_idx(va
->va_start
);
1560 spin_lock(&vmap_block_tree_lock
);
1561 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
1562 spin_unlock(&vmap_block_tree_lock
);
1564 radix_tree_preload_end();
1566 vbq
= &get_cpu_var(vmap_block_queue
);
1567 spin_lock(&vbq
->lock
);
1568 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
1569 spin_unlock(&vbq
->lock
);
1570 put_cpu_var(vmap_block_queue
);
1575 static void free_vmap_block(struct vmap_block
*vb
)
1577 struct vmap_block
*tmp
;
1578 unsigned long vb_idx
;
1580 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
1581 spin_lock(&vmap_block_tree_lock
);
1582 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
1583 spin_unlock(&vmap_block_tree_lock
);
1586 free_vmap_area_noflush(vb
->va
);
1587 kfree_rcu(vb
, rcu_head
);
1590 static void purge_fragmented_blocks(int cpu
)
1593 struct vmap_block
*vb
;
1594 struct vmap_block
*n_vb
;
1595 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1598 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1600 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
1603 spin_lock(&vb
->lock
);
1604 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
1605 vb
->free
= 0; /* prevent further allocs after releasing lock */
1606 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
1608 vb
->dirty_max
= VMAP_BBMAP_BITS
;
1609 spin_lock(&vbq
->lock
);
1610 list_del_rcu(&vb
->free_list
);
1611 spin_unlock(&vbq
->lock
);
1612 spin_unlock(&vb
->lock
);
1613 list_add_tail(&vb
->purge
, &purge
);
1615 spin_unlock(&vb
->lock
);
1619 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
1620 list_del(&vb
->purge
);
1621 free_vmap_block(vb
);
1625 static void purge_fragmented_blocks_allcpus(void)
1629 for_each_possible_cpu(cpu
)
1630 purge_fragmented_blocks(cpu
);
1633 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
1635 struct vmap_block_queue
*vbq
;
1636 struct vmap_block
*vb
;
1640 BUG_ON(offset_in_page(size
));
1641 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1642 if (WARN_ON(size
== 0)) {
1644 * Allocating 0 bytes isn't what caller wants since
1645 * get_order(0) returns funny result. Just warn and terminate
1650 order
= get_order(size
);
1653 vbq
= &get_cpu_var(vmap_block_queue
);
1654 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1655 unsigned long pages_off
;
1657 spin_lock(&vb
->lock
);
1658 if (vb
->free
< (1UL << order
)) {
1659 spin_unlock(&vb
->lock
);
1663 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1664 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1665 vb
->free
-= 1UL << order
;
1666 if (vb
->free
== 0) {
1667 spin_lock(&vbq
->lock
);
1668 list_del_rcu(&vb
->free_list
);
1669 spin_unlock(&vbq
->lock
);
1672 spin_unlock(&vb
->lock
);
1676 put_cpu_var(vmap_block_queue
);
1679 /* Allocate new block if nothing was found */
1681 vaddr
= new_vmap_block(order
, gfp_mask
);
1686 static void vb_free(unsigned long addr
, unsigned long size
)
1688 unsigned long offset
;
1689 unsigned long vb_idx
;
1691 struct vmap_block
*vb
;
1693 BUG_ON(offset_in_page(size
));
1694 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1696 flush_cache_vunmap(addr
, addr
+ size
);
1698 order
= get_order(size
);
1700 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
1702 vb_idx
= addr_to_vb_idx(addr
);
1704 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1708 unmap_kernel_range_noflush(addr
, size
);
1710 if (debug_pagealloc_enabled_static())
1711 flush_tlb_kernel_range(addr
, addr
+ size
);
1713 spin_lock(&vb
->lock
);
1715 /* Expand dirty range */
1716 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1717 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1719 vb
->dirty
+= 1UL << order
;
1720 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1722 spin_unlock(&vb
->lock
);
1723 free_vmap_block(vb
);
1725 spin_unlock(&vb
->lock
);
1728 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
1732 if (unlikely(!vmap_initialized
))
1737 for_each_possible_cpu(cpu
) {
1738 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1739 struct vmap_block
*vb
;
1742 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1743 spin_lock(&vb
->lock
);
1745 unsigned long va_start
= vb
->va
->va_start
;
1748 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1749 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1751 start
= min(s
, start
);
1756 spin_unlock(&vb
->lock
);
1761 mutex_lock(&vmap_purge_lock
);
1762 purge_fragmented_blocks_allcpus();
1763 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1764 flush_tlb_kernel_range(start
, end
);
1765 mutex_unlock(&vmap_purge_lock
);
1769 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1771 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1772 * to amortize TLB flushing overheads. What this means is that any page you
1773 * have now, may, in a former life, have been mapped into kernel virtual
1774 * address by the vmap layer and so there might be some CPUs with TLB entries
1775 * still referencing that page (additional to the regular 1:1 kernel mapping).
1777 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1778 * be sure that none of the pages we have control over will have any aliases
1779 * from the vmap layer.
1781 void vm_unmap_aliases(void)
1783 unsigned long start
= ULONG_MAX
, end
= 0;
1786 _vm_unmap_aliases(start
, end
, flush
);
1788 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1791 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1792 * @mem: the pointer returned by vm_map_ram
1793 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1795 void vm_unmap_ram(const void *mem
, unsigned int count
)
1797 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1798 unsigned long addr
= (unsigned long)mem
;
1799 struct vmap_area
*va
;
1803 BUG_ON(addr
< VMALLOC_START
);
1804 BUG_ON(addr
> VMALLOC_END
);
1805 BUG_ON(!PAGE_ALIGNED(addr
));
1807 kasan_poison_vmalloc(mem
, size
);
1809 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1810 debug_check_no_locks_freed(mem
, size
);
1811 vb_free(addr
, size
);
1815 va
= find_vmap_area(addr
);
1817 debug_check_no_locks_freed((void *)va
->va_start
,
1818 (va
->va_end
- va
->va_start
));
1819 free_unmap_vmap_area(va
);
1821 EXPORT_SYMBOL(vm_unmap_ram
);
1824 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1825 * @pages: an array of pointers to the pages to be mapped
1826 * @count: number of pages
1827 * @node: prefer to allocate data structures on this node
1828 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1830 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1831 * faster than vmap so it's good. But if you mix long-life and short-life
1832 * objects with vm_map_ram(), it could consume lots of address space through
1833 * fragmentation (especially on a 32bit machine). You could see failures in
1834 * the end. Please use this function for short-lived objects.
1836 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1838 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
1840 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1844 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1845 mem
= vb_alloc(size
, GFP_KERNEL
);
1848 addr
= (unsigned long)mem
;
1850 struct vmap_area
*va
;
1851 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1852 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1856 addr
= va
->va_start
;
1860 kasan_unpoison_vmalloc(mem
, size
);
1862 if (map_kernel_range(addr
, size
, PAGE_KERNEL
, pages
) < 0) {
1863 vm_unmap_ram(mem
, count
);
1868 EXPORT_SYMBOL(vm_map_ram
);
1870 static struct vm_struct
*vmlist __initdata
;
1873 * vm_area_add_early - add vmap area early during boot
1874 * @vm: vm_struct to add
1876 * This function is used to add fixed kernel vm area to vmlist before
1877 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1878 * should contain proper values and the other fields should be zero.
1880 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1882 void __init
vm_area_add_early(struct vm_struct
*vm
)
1884 struct vm_struct
*tmp
, **p
;
1886 BUG_ON(vmap_initialized
);
1887 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1888 if (tmp
->addr
>= vm
->addr
) {
1889 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1892 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1899 * vm_area_register_early - register vmap area early during boot
1900 * @vm: vm_struct to register
1901 * @align: requested alignment
1903 * This function is used to register kernel vm area before
1904 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1905 * proper values on entry and other fields should be zero. On return,
1906 * vm->addr contains the allocated address.
1908 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1910 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1912 static size_t vm_init_off __initdata
;
1915 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1916 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1918 vm
->addr
= (void *)addr
;
1920 vm_area_add_early(vm
);
1923 static void vmap_init_free_space(void)
1925 unsigned long vmap_start
= 1;
1926 const unsigned long vmap_end
= ULONG_MAX
;
1927 struct vmap_area
*busy
, *free
;
1931 * -|-----|.....|-----|-----|-----|.....|-
1933 * |<--------------------------------->|
1935 list_for_each_entry(busy
, &vmap_area_list
, list
) {
1936 if (busy
->va_start
- vmap_start
> 0) {
1937 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1938 if (!WARN_ON_ONCE(!free
)) {
1939 free
->va_start
= vmap_start
;
1940 free
->va_end
= busy
->va_start
;
1942 insert_vmap_area_augment(free
, NULL
,
1943 &free_vmap_area_root
,
1944 &free_vmap_area_list
);
1948 vmap_start
= busy
->va_end
;
1951 if (vmap_end
- vmap_start
> 0) {
1952 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1953 if (!WARN_ON_ONCE(!free
)) {
1954 free
->va_start
= vmap_start
;
1955 free
->va_end
= vmap_end
;
1957 insert_vmap_area_augment(free
, NULL
,
1958 &free_vmap_area_root
,
1959 &free_vmap_area_list
);
1964 void __init
vmalloc_init(void)
1966 struct vmap_area
*va
;
1967 struct vm_struct
*tmp
;
1971 * Create the cache for vmap_area objects.
1973 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
1975 for_each_possible_cpu(i
) {
1976 struct vmap_block_queue
*vbq
;
1977 struct vfree_deferred
*p
;
1979 vbq
= &per_cpu(vmap_block_queue
, i
);
1980 spin_lock_init(&vbq
->lock
);
1981 INIT_LIST_HEAD(&vbq
->free
);
1982 p
= &per_cpu(vfree_deferred
, i
);
1983 init_llist_head(&p
->list
);
1984 INIT_WORK(&p
->wq
, free_work
);
1987 /* Import existing vmlist entries. */
1988 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1989 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1990 if (WARN_ON_ONCE(!va
))
1993 va
->va_start
= (unsigned long)tmp
->addr
;
1994 va
->va_end
= va
->va_start
+ tmp
->size
;
1996 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
2000 * Now we can initialize a free vmap space.
2002 vmap_init_free_space();
2003 vmap_initialized
= true;
2007 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2008 * @addr: start of the VM area to unmap
2009 * @size: size of the VM area to unmap
2011 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2012 * the unmapping and tlb after.
2014 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
2016 unsigned long end
= addr
+ size
;
2018 flush_cache_vunmap(addr
, end
);
2019 unmap_kernel_range_noflush(addr
, size
);
2020 flush_tlb_kernel_range(addr
, end
);
2023 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2024 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2027 vm
->addr
= (void *)va
->va_start
;
2028 vm
->size
= va
->va_end
- va
->va_start
;
2029 vm
->caller
= caller
;
2033 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2034 unsigned long flags
, const void *caller
)
2036 spin_lock(&vmap_area_lock
);
2037 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2038 spin_unlock(&vmap_area_lock
);
2041 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2044 * Before removing VM_UNINITIALIZED,
2045 * we should make sure that vm has proper values.
2046 * Pair with smp_rmb() in show_numa_info().
2049 vm
->flags
&= ~VM_UNINITIALIZED
;
2052 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2053 unsigned long align
, unsigned long flags
, unsigned long start
,
2054 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
2056 struct vmap_area
*va
;
2057 struct vm_struct
*area
;
2058 unsigned long requested_size
= size
;
2060 BUG_ON(in_interrupt());
2061 size
= PAGE_ALIGN(size
);
2062 if (unlikely(!size
))
2065 if (flags
& VM_IOREMAP
)
2066 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2067 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2069 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2070 if (unlikely(!area
))
2073 if (!(flags
& VM_NO_GUARD
))
2076 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
2082 kasan_unpoison_vmalloc((void *)va
->va_start
, requested_size
);
2084 setup_vmalloc_vm(area
, va
, flags
, caller
);
2089 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2090 unsigned long start
, unsigned long end
,
2093 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2094 GFP_KERNEL
, caller
);
2098 * get_vm_area - reserve a contiguous kernel virtual area
2099 * @size: size of the area
2100 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2102 * Search an area of @size in the kernel virtual mapping area,
2103 * and reserved it for out purposes. Returns the area descriptor
2104 * on success or %NULL on failure.
2106 * Return: the area descriptor on success or %NULL on failure.
2108 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2110 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2111 NUMA_NO_NODE
, GFP_KERNEL
,
2112 __builtin_return_address(0));
2115 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2118 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2119 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2123 * find_vm_area - find a continuous kernel virtual area
2124 * @addr: base address
2126 * Search for the kernel VM area starting at @addr, and return it.
2127 * It is up to the caller to do all required locking to keep the returned
2130 * Return: pointer to the found area or %NULL on faulure
2132 struct vm_struct
*find_vm_area(const void *addr
)
2134 struct vmap_area
*va
;
2136 va
= find_vmap_area((unsigned long)addr
);
2144 * remove_vm_area - find and remove a continuous kernel virtual area
2145 * @addr: base address
2147 * Search for the kernel VM area starting at @addr, and remove it.
2148 * This function returns the found VM area, but using it is NOT safe
2149 * on SMP machines, except for its size or flags.
2151 * Return: pointer to the found area or %NULL on faulure
2153 struct vm_struct
*remove_vm_area(const void *addr
)
2155 struct vmap_area
*va
;
2159 spin_lock(&vmap_area_lock
);
2160 va
= __find_vmap_area((unsigned long)addr
);
2162 struct vm_struct
*vm
= va
->vm
;
2165 spin_unlock(&vmap_area_lock
);
2167 kasan_free_shadow(vm
);
2168 free_unmap_vmap_area(va
);
2173 spin_unlock(&vmap_area_lock
);
2177 static inline void set_area_direct_map(const struct vm_struct
*area
,
2178 int (*set_direct_map
)(struct page
*page
))
2182 for (i
= 0; i
< area
->nr_pages
; i
++)
2183 if (page_address(area
->pages
[i
]))
2184 set_direct_map(area
->pages
[i
]);
2187 /* Handle removing and resetting vm mappings related to the vm_struct. */
2188 static void vm_remove_mappings(struct vm_struct
*area
, int deallocate_pages
)
2190 unsigned long start
= ULONG_MAX
, end
= 0;
2191 int flush_reset
= area
->flags
& VM_FLUSH_RESET_PERMS
;
2195 remove_vm_area(area
->addr
);
2197 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2202 * If not deallocating pages, just do the flush of the VM area and
2205 if (!deallocate_pages
) {
2211 * If execution gets here, flush the vm mapping and reset the direct
2212 * map. Find the start and end range of the direct mappings to make sure
2213 * the vm_unmap_aliases() flush includes the direct map.
2215 for (i
= 0; i
< area
->nr_pages
; i
++) {
2216 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2218 start
= min(addr
, start
);
2219 end
= max(addr
+ PAGE_SIZE
, end
);
2225 * Set direct map to something invalid so that it won't be cached if
2226 * there are any accesses after the TLB flush, then flush the TLB and
2227 * reset the direct map permissions to the default.
2229 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2230 _vm_unmap_aliases(start
, end
, flush_dmap
);
2231 set_area_direct_map(area
, set_direct_map_default_noflush
);
2234 static void __vunmap(const void *addr
, int deallocate_pages
)
2236 struct vm_struct
*area
;
2241 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2245 area
= find_vm_area(addr
);
2246 if (unlikely(!area
)) {
2247 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2252 debug_check_no_locks_freed(area
->addr
, get_vm_area_size(area
));
2253 debug_check_no_obj_freed(area
->addr
, get_vm_area_size(area
));
2255 kasan_poison_vmalloc(area
->addr
, area
->size
);
2257 vm_remove_mappings(area
, deallocate_pages
);
2259 if (deallocate_pages
) {
2262 for (i
= 0; i
< area
->nr_pages
; i
++) {
2263 struct page
*page
= area
->pages
[i
];
2266 __free_pages(page
, 0);
2268 atomic_long_sub(area
->nr_pages
, &nr_vmalloc_pages
);
2270 kvfree(area
->pages
);
2277 static inline void __vfree_deferred(const void *addr
)
2280 * Use raw_cpu_ptr() because this can be called from preemptible
2281 * context. Preemption is absolutely fine here, because the llist_add()
2282 * implementation is lockless, so it works even if we are adding to
2283 * nother cpu's list. schedule_work() should be fine with this too.
2285 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2287 if (llist_add((struct llist_node
*)addr
, &p
->list
))
2288 schedule_work(&p
->wq
);
2292 * vfree_atomic - release memory allocated by vmalloc()
2293 * @addr: memory base address
2295 * This one is just like vfree() but can be called in any atomic context
2298 void vfree_atomic(const void *addr
)
2302 kmemleak_free(addr
);
2306 __vfree_deferred(addr
);
2309 static void __vfree(const void *addr
)
2311 if (unlikely(in_interrupt()))
2312 __vfree_deferred(addr
);
2318 * vfree - release memory allocated by vmalloc()
2319 * @addr: memory base address
2321 * Free the virtually continuous memory area starting at @addr, as
2322 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2323 * NULL, no operation is performed.
2325 * Must not be called in NMI context (strictly speaking, only if we don't
2326 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2327 * conventions for vfree() arch-depenedent would be a really bad idea)
2329 * May sleep if called *not* from interrupt context.
2331 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2333 void vfree(const void *addr
)
2337 kmemleak_free(addr
);
2339 might_sleep_if(!in_interrupt());
2346 EXPORT_SYMBOL(vfree
);
2349 * vunmap - release virtual mapping obtained by vmap()
2350 * @addr: memory base address
2352 * Free the virtually contiguous memory area starting at @addr,
2353 * which was created from the page array passed to vmap().
2355 * Must not be called in interrupt context.
2357 void vunmap(const void *addr
)
2359 BUG_ON(in_interrupt());
2364 EXPORT_SYMBOL(vunmap
);
2367 * vmap - map an array of pages into virtually contiguous space
2368 * @pages: array of page pointers
2369 * @count: number of pages to map
2370 * @flags: vm_area->flags
2371 * @prot: page protection for the mapping
2373 * Maps @count pages from @pages into contiguous kernel virtual
2376 * Return: the address of the area or %NULL on failure
2378 void *vmap(struct page
**pages
, unsigned int count
,
2379 unsigned long flags
, pgprot_t prot
)
2381 struct vm_struct
*area
;
2382 unsigned long size
; /* In bytes */
2386 if (count
> totalram_pages())
2389 size
= (unsigned long)count
<< PAGE_SHIFT
;
2390 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2394 if (map_kernel_range((unsigned long)area
->addr
, size
, pgprot_nx(prot
),
2402 EXPORT_SYMBOL(vmap
);
2404 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2405 pgprot_t prot
, int node
)
2407 struct page
**pages
;
2408 unsigned int nr_pages
, array_size
, i
;
2409 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2410 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
2411 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
2415 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
2416 array_size
= (nr_pages
* sizeof(struct page
*));
2418 /* Please note that the recursion is strictly bounded. */
2419 if (array_size
> PAGE_SIZE
) {
2420 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
2421 node
, area
->caller
);
2423 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
2427 remove_vm_area(area
->addr
);
2432 area
->pages
= pages
;
2433 area
->nr_pages
= nr_pages
;
2435 for (i
= 0; i
< area
->nr_pages
; i
++) {
2438 if (node
== NUMA_NO_NODE
)
2439 page
= alloc_page(alloc_mask
|highmem_mask
);
2441 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
2443 if (unlikely(!page
)) {
2444 /* Successfully allocated i pages, free them in __vunmap() */
2446 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2449 area
->pages
[i
] = page
;
2450 if (gfpflags_allow_blocking(gfp_mask
))
2453 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2455 if (map_kernel_range((unsigned long)area
->addr
, get_vm_area_size(area
),
2462 warn_alloc(gfp_mask
, NULL
,
2463 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2464 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
2465 __vfree(area
->addr
);
2470 * __vmalloc_node_range - allocate virtually contiguous memory
2471 * @size: allocation size
2472 * @align: desired alignment
2473 * @start: vm area range start
2474 * @end: vm area range end
2475 * @gfp_mask: flags for the page level allocator
2476 * @prot: protection mask for the allocated pages
2477 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2478 * @node: node to use for allocation or NUMA_NO_NODE
2479 * @caller: caller's return address
2481 * Allocate enough pages to cover @size from the page level
2482 * allocator with @gfp_mask flags. Map them into contiguous
2483 * kernel virtual space, using a pagetable protection of @prot.
2485 * Return: the address of the area or %NULL on failure
2487 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
2488 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
2489 pgprot_t prot
, unsigned long vm_flags
, int node
,
2492 struct vm_struct
*area
;
2494 unsigned long real_size
= size
;
2496 size
= PAGE_ALIGN(size
);
2497 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages())
2500 area
= __get_vm_area_node(real_size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
2501 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
2505 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
2510 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2511 * flag. It means that vm_struct is not fully initialized.
2512 * Now, it is fully initialized, so remove this flag here.
2514 clear_vm_uninitialized_flag(area
);
2516 kmemleak_vmalloc(area
, size
, gfp_mask
);
2521 warn_alloc(gfp_mask
, NULL
,
2522 "vmalloc: allocation failure: %lu bytes", real_size
);
2527 * __vmalloc_node - allocate virtually contiguous memory
2528 * @size: allocation size
2529 * @align: desired alignment
2530 * @gfp_mask: flags for the page level allocator
2531 * @node: node to use for allocation or NUMA_NO_NODE
2532 * @caller: caller's return address
2534 * Allocate enough pages to cover @size from the page level allocator with
2535 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2537 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2538 * and __GFP_NOFAIL are not supported
2540 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2543 * Return: pointer to the allocated memory or %NULL on error
2545 void *__vmalloc_node(unsigned long size
, unsigned long align
,
2546 gfp_t gfp_mask
, int node
, const void *caller
)
2548 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
2549 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
2552 * This is only for performance analysis of vmalloc and stress purpose.
2553 * It is required by vmalloc test module, therefore do not use it other
2556 #ifdef CONFIG_TEST_VMALLOC_MODULE
2557 EXPORT_SYMBOL_GPL(__vmalloc_node
);
2560 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
2562 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
2563 __builtin_return_address(0));
2565 EXPORT_SYMBOL(__vmalloc
);
2568 * vmalloc - allocate virtually contiguous memory
2569 * @size: allocation size
2571 * Allocate enough pages to cover @size from the page level
2572 * allocator and map them into contiguous kernel virtual space.
2574 * For tight control over page level allocator and protection flags
2575 * use __vmalloc() instead.
2577 * Return: pointer to the allocated memory or %NULL on error
2579 void *vmalloc(unsigned long size
)
2581 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
2582 __builtin_return_address(0));
2584 EXPORT_SYMBOL(vmalloc
);
2587 * vzalloc - allocate virtually contiguous memory with zero fill
2588 * @size: allocation size
2590 * Allocate enough pages to cover @size from the page level
2591 * allocator and map them into contiguous kernel virtual space.
2592 * The memory allocated is set to zero.
2594 * For tight control over page level allocator and protection flags
2595 * use __vmalloc() instead.
2597 * Return: pointer to the allocated memory or %NULL on error
2599 void *vzalloc(unsigned long size
)
2601 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
2602 __builtin_return_address(0));
2604 EXPORT_SYMBOL(vzalloc
);
2607 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2608 * @size: allocation size
2610 * The resulting memory area is zeroed so it can be mapped to userspace
2611 * without leaking data.
2613 * Return: pointer to the allocated memory or %NULL on error
2615 void *vmalloc_user(unsigned long size
)
2617 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2618 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
2619 VM_USERMAP
, NUMA_NO_NODE
,
2620 __builtin_return_address(0));
2622 EXPORT_SYMBOL(vmalloc_user
);
2625 * vmalloc_node - allocate memory on a specific node
2626 * @size: allocation size
2629 * Allocate enough pages to cover @size from the page level
2630 * allocator and map them into contiguous kernel virtual space.
2632 * For tight control over page level allocator and protection flags
2633 * use __vmalloc() instead.
2635 * Return: pointer to the allocated memory or %NULL on error
2637 void *vmalloc_node(unsigned long size
, int node
)
2639 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
2640 __builtin_return_address(0));
2642 EXPORT_SYMBOL(vmalloc_node
);
2645 * vzalloc_node - allocate memory on a specific node with zero fill
2646 * @size: allocation size
2649 * Allocate enough pages to cover @size from the page level
2650 * allocator and map them into contiguous kernel virtual space.
2651 * The memory allocated is set to zero.
2653 * Return: pointer to the allocated memory or %NULL on error
2655 void *vzalloc_node(unsigned long size
, int node
)
2657 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
2658 __builtin_return_address(0));
2660 EXPORT_SYMBOL(vzalloc_node
);
2663 * vmalloc_exec - allocate virtually contiguous, executable memory
2664 * @size: allocation size
2666 * Kernel-internal function to allocate enough pages to cover @size
2667 * the page level allocator and map them into contiguous and
2668 * executable kernel virtual space.
2670 * For tight control over page level allocator and protection flags
2671 * use __vmalloc() instead.
2673 * Return: pointer to the allocated memory or %NULL on error
2675 void *vmalloc_exec(unsigned long size
)
2677 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
2678 GFP_KERNEL
, PAGE_KERNEL_EXEC
, VM_FLUSH_RESET_PERMS
,
2679 NUMA_NO_NODE
, __builtin_return_address(0));
2682 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2683 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2684 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2685 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2688 * 64b systems should always have either DMA or DMA32 zones. For others
2689 * GFP_DMA32 should do the right thing and use the normal zone.
2691 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2695 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2696 * @size: allocation size
2698 * Allocate enough 32bit PA addressable pages to cover @size from the
2699 * page level allocator and map them into contiguous kernel virtual space.
2701 * Return: pointer to the allocated memory or %NULL on error
2703 void *vmalloc_32(unsigned long size
)
2705 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
2706 __builtin_return_address(0));
2708 EXPORT_SYMBOL(vmalloc_32
);
2711 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2712 * @size: allocation size
2714 * The resulting memory area is 32bit addressable and zeroed so it can be
2715 * mapped to userspace without leaking data.
2717 * Return: pointer to the allocated memory or %NULL on error
2719 void *vmalloc_32_user(unsigned long size
)
2721 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2722 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2723 VM_USERMAP
, NUMA_NO_NODE
,
2724 __builtin_return_address(0));
2726 EXPORT_SYMBOL(vmalloc_32_user
);
2729 * small helper routine , copy contents to buf from addr.
2730 * If the page is not present, fill zero.
2733 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2739 unsigned long offset
, length
;
2741 offset
= offset_in_page(addr
);
2742 length
= PAGE_SIZE
- offset
;
2745 p
= vmalloc_to_page(addr
);
2747 * To do safe access to this _mapped_ area, we need
2748 * lock. But adding lock here means that we need to add
2749 * overhead of vmalloc()/vfree() calles for this _debug_
2750 * interface, rarely used. Instead of that, we'll use
2751 * kmap() and get small overhead in this access function.
2755 * we can expect USER0 is not used (see vread/vwrite's
2756 * function description)
2758 void *map
= kmap_atomic(p
);
2759 memcpy(buf
, map
+ offset
, length
);
2762 memset(buf
, 0, length
);
2772 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2778 unsigned long offset
, length
;
2780 offset
= offset_in_page(addr
);
2781 length
= PAGE_SIZE
- offset
;
2784 p
= vmalloc_to_page(addr
);
2786 * To do safe access to this _mapped_ area, we need
2787 * lock. But adding lock here means that we need to add
2788 * overhead of vmalloc()/vfree() calles for this _debug_
2789 * interface, rarely used. Instead of that, we'll use
2790 * kmap() and get small overhead in this access function.
2794 * we can expect USER0 is not used (see vread/vwrite's
2795 * function description)
2797 void *map
= kmap_atomic(p
);
2798 memcpy(map
+ offset
, buf
, length
);
2810 * vread() - read vmalloc area in a safe way.
2811 * @buf: buffer for reading data
2812 * @addr: vm address.
2813 * @count: number of bytes to be read.
2815 * This function checks that addr is a valid vmalloc'ed area, and
2816 * copy data from that area to a given buffer. If the given memory range
2817 * of [addr...addr+count) includes some valid address, data is copied to
2818 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2819 * IOREMAP area is treated as memory hole and no copy is done.
2821 * If [addr...addr+count) doesn't includes any intersects with alive
2822 * vm_struct area, returns 0. @buf should be kernel's buffer.
2824 * Note: In usual ops, vread() is never necessary because the caller
2825 * should know vmalloc() area is valid and can use memcpy().
2826 * This is for routines which have to access vmalloc area without
2827 * any information, as /dev/kmem.
2829 * Return: number of bytes for which addr and buf should be increased
2830 * (same number as @count) or %0 if [addr...addr+count) doesn't
2831 * include any intersection with valid vmalloc area
2833 long vread(char *buf
, char *addr
, unsigned long count
)
2835 struct vmap_area
*va
;
2836 struct vm_struct
*vm
;
2837 char *vaddr
, *buf_start
= buf
;
2838 unsigned long buflen
= count
;
2841 /* Don't allow overflow */
2842 if ((unsigned long) addr
+ count
< count
)
2843 count
= -(unsigned long) addr
;
2845 spin_lock(&vmap_area_lock
);
2846 list_for_each_entry(va
, &vmap_area_list
, list
) {
2854 vaddr
= (char *) vm
->addr
;
2855 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2857 while (addr
< vaddr
) {
2865 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2868 if (!(vm
->flags
& VM_IOREMAP
))
2869 aligned_vread(buf
, addr
, n
);
2870 else /* IOREMAP area is treated as memory hole */
2877 spin_unlock(&vmap_area_lock
);
2879 if (buf
== buf_start
)
2881 /* zero-fill memory holes */
2882 if (buf
!= buf_start
+ buflen
)
2883 memset(buf
, 0, buflen
- (buf
- buf_start
));
2889 * vwrite() - write vmalloc area in a safe way.
2890 * @buf: buffer for source data
2891 * @addr: vm address.
2892 * @count: number of bytes to be read.
2894 * This function checks that addr is a valid vmalloc'ed area, and
2895 * copy data from a buffer to the given addr. If specified range of
2896 * [addr...addr+count) includes some valid address, data is copied from
2897 * proper area of @buf. If there are memory holes, no copy to hole.
2898 * IOREMAP area is treated as memory hole and no copy is done.
2900 * If [addr...addr+count) doesn't includes any intersects with alive
2901 * vm_struct area, returns 0. @buf should be kernel's buffer.
2903 * Note: In usual ops, vwrite() is never necessary because the caller
2904 * should know vmalloc() area is valid and can use memcpy().
2905 * This is for routines which have to access vmalloc area without
2906 * any information, as /dev/kmem.
2908 * Return: number of bytes for which addr and buf should be
2909 * increased (same number as @count) or %0 if [addr...addr+count)
2910 * doesn't include any intersection with valid vmalloc area
2912 long vwrite(char *buf
, char *addr
, unsigned long count
)
2914 struct vmap_area
*va
;
2915 struct vm_struct
*vm
;
2917 unsigned long n
, buflen
;
2920 /* Don't allow overflow */
2921 if ((unsigned long) addr
+ count
< count
)
2922 count
= -(unsigned long) addr
;
2925 spin_lock(&vmap_area_lock
);
2926 list_for_each_entry(va
, &vmap_area_list
, list
) {
2934 vaddr
= (char *) vm
->addr
;
2935 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2937 while (addr
< vaddr
) {
2944 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2947 if (!(vm
->flags
& VM_IOREMAP
)) {
2948 aligned_vwrite(buf
, addr
, n
);
2956 spin_unlock(&vmap_area_lock
);
2963 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2964 * @vma: vma to cover
2965 * @uaddr: target user address to start at
2966 * @kaddr: virtual address of vmalloc kernel memory
2967 * @pgoff: offset from @kaddr to start at
2968 * @size: size of map area
2970 * Returns: 0 for success, -Exxx on failure
2972 * This function checks that @kaddr is a valid vmalloc'ed area,
2973 * and that it is big enough to cover the range starting at
2974 * @uaddr in @vma. Will return failure if that criteria isn't
2977 * Similar to remap_pfn_range() (see mm/memory.c)
2979 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
2980 void *kaddr
, unsigned long pgoff
,
2983 struct vm_struct
*area
;
2985 unsigned long end_index
;
2987 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
2990 size
= PAGE_ALIGN(size
);
2992 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
2995 area
= find_vm_area(kaddr
);
2999 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3002 if (check_add_overflow(size
, off
, &end_index
) ||
3003 end_index
> get_vm_area_size(area
))
3008 struct page
*page
= vmalloc_to_page(kaddr
);
3011 ret
= vm_insert_page(vma
, uaddr
, page
);
3020 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3024 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
3027 * remap_vmalloc_range - map vmalloc pages to userspace
3028 * @vma: vma to cover (map full range of vma)
3029 * @addr: vmalloc memory
3030 * @pgoff: number of pages into addr before first page to map
3032 * Returns: 0 for success, -Exxx on failure
3034 * This function checks that addr is a valid vmalloc'ed area, and
3035 * that it is big enough to cover the vma. Will return failure if
3036 * that criteria isn't met.
3038 * Similar to remap_pfn_range() (see mm/memory.c)
3040 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3041 unsigned long pgoff
)
3043 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3045 vma
->vm_end
- vma
->vm_start
);
3047 EXPORT_SYMBOL(remap_vmalloc_range
);
3050 * Implement stubs for vmalloc_sync_[un]mappings () if the architecture chose
3053 * The purpose of this function is to make sure the vmalloc area
3054 * mappings are identical in all page-tables in the system.
3056 void __weak
vmalloc_sync_mappings(void)
3060 void __weak
vmalloc_sync_unmappings(void)
3064 static int f(pte_t
*pte
, unsigned long addr
, void *data
)
3076 * alloc_vm_area - allocate a range of kernel address space
3077 * @size: size of the area
3078 * @ptes: returns the PTEs for the address space
3080 * Returns: NULL on failure, vm_struct on success
3082 * This function reserves a range of kernel address space, and
3083 * allocates pagetables to map that range. No actual mappings
3086 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3087 * allocated for the VM area are returned.
3089 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
3091 struct vm_struct
*area
;
3093 area
= get_vm_area_caller(size
, VM_IOREMAP
,
3094 __builtin_return_address(0));
3099 * This ensures that page tables are constructed for this region
3100 * of kernel virtual address space and mapped into init_mm.
3102 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3103 size
, f
, ptes
? &ptes
: NULL
)) {
3110 EXPORT_SYMBOL_GPL(alloc_vm_area
);
3112 void free_vm_area(struct vm_struct
*area
)
3114 struct vm_struct
*ret
;
3115 ret
= remove_vm_area(area
->addr
);
3116 BUG_ON(ret
!= area
);
3119 EXPORT_SYMBOL_GPL(free_vm_area
);
3122 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3124 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3128 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3129 * @addr: target address
3131 * Returns: vmap_area if it is found. If there is no such area
3132 * the first highest(reverse order) vmap_area is returned
3133 * i.e. va->va_start < addr && va->va_end < addr or NULL
3134 * if there are no any areas before @addr.
3136 static struct vmap_area
*
3137 pvm_find_va_enclose_addr(unsigned long addr
)
3139 struct vmap_area
*va
, *tmp
;
3142 n
= free_vmap_area_root
.rb_node
;
3146 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3147 if (tmp
->va_start
<= addr
) {
3149 if (tmp
->va_end
>= addr
)
3162 * pvm_determine_end_from_reverse - find the highest aligned address
3163 * of free block below VMALLOC_END
3165 * in - the VA we start the search(reverse order);
3166 * out - the VA with the highest aligned end address.
3168 * Returns: determined end address within vmap_area
3170 static unsigned long
3171 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3173 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3177 list_for_each_entry_from_reverse((*va
),
3178 &free_vmap_area_list
, list
) {
3179 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3180 if ((*va
)->va_start
< addr
)
3189 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3190 * @offsets: array containing offset of each area
3191 * @sizes: array containing size of each area
3192 * @nr_vms: the number of areas to allocate
3193 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3195 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3196 * vm_structs on success, %NULL on failure
3198 * Percpu allocator wants to use congruent vm areas so that it can
3199 * maintain the offsets among percpu areas. This function allocates
3200 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3201 * be scattered pretty far, distance between two areas easily going up
3202 * to gigabytes. To avoid interacting with regular vmallocs, these
3203 * areas are allocated from top.
3205 * Despite its complicated look, this allocator is rather simple. It
3206 * does everything top-down and scans free blocks from the end looking
3207 * for matching base. While scanning, if any of the areas do not fit the
3208 * base address is pulled down to fit the area. Scanning is repeated till
3209 * all the areas fit and then all necessary data structures are inserted
3210 * and the result is returned.
3212 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3213 const size_t *sizes
, int nr_vms
,
3216 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3217 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3218 struct vmap_area
**vas
, *va
;
3219 struct vm_struct
**vms
;
3220 int area
, area2
, last_area
, term_area
;
3221 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3222 bool purged
= false;
3225 /* verify parameters and allocate data structures */
3226 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3227 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3228 start
= offsets
[area
];
3229 end
= start
+ sizes
[area
];
3231 /* is everything aligned properly? */
3232 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3233 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3235 /* detect the area with the highest address */
3236 if (start
> offsets
[last_area
])
3239 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3240 unsigned long start2
= offsets
[area2
];
3241 unsigned long end2
= start2
+ sizes
[area2
];
3243 BUG_ON(start2
< end
&& start
< end2
);
3246 last_end
= offsets
[last_area
] + sizes
[last_area
];
3248 if (vmalloc_end
- vmalloc_start
< last_end
) {
3253 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3254 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3258 for (area
= 0; area
< nr_vms
; area
++) {
3259 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3260 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3261 if (!vas
[area
] || !vms
[area
])
3265 spin_lock(&free_vmap_area_lock
);
3267 /* start scanning - we scan from the top, begin with the last area */
3268 area
= term_area
= last_area
;
3269 start
= offsets
[area
];
3270 end
= start
+ sizes
[area
];
3272 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3273 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3277 * base might have underflowed, add last_end before
3280 if (base
+ last_end
< vmalloc_start
+ last_end
)
3284 * Fitting base has not been found.
3290 * If required width exceeds current VA block, move
3291 * base downwards and then recheck.
3293 if (base
+ end
> va
->va_end
) {
3294 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3300 * If this VA does not fit, move base downwards and recheck.
3302 if (base
+ start
< va
->va_start
) {
3303 va
= node_to_va(rb_prev(&va
->rb_node
));
3304 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3310 * This area fits, move on to the previous one. If
3311 * the previous one is the terminal one, we're done.
3313 area
= (area
+ nr_vms
- 1) % nr_vms
;
3314 if (area
== term_area
)
3317 start
= offsets
[area
];
3318 end
= start
+ sizes
[area
];
3319 va
= pvm_find_va_enclose_addr(base
+ end
);
3322 /* we've found a fitting base, insert all va's */
3323 for (area
= 0; area
< nr_vms
; area
++) {
3326 start
= base
+ offsets
[area
];
3329 va
= pvm_find_va_enclose_addr(start
);
3330 if (WARN_ON_ONCE(va
== NULL
))
3331 /* It is a BUG(), but trigger recovery instead. */
3334 type
= classify_va_fit_type(va
, start
, size
);
3335 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
3336 /* It is a BUG(), but trigger recovery instead. */
3339 ret
= adjust_va_to_fit_type(va
, start
, size
, type
);
3343 /* Allocated area. */
3345 va
->va_start
= start
;
3346 va
->va_end
= start
+ size
;
3349 spin_unlock(&free_vmap_area_lock
);
3351 /* populate the kasan shadow space */
3352 for (area
= 0; area
< nr_vms
; area
++) {
3353 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3354 goto err_free_shadow
;
3356 kasan_unpoison_vmalloc((void *)vas
[area
]->va_start
,
3360 /* insert all vm's */
3361 spin_lock(&vmap_area_lock
);
3362 for (area
= 0; area
< nr_vms
; area
++) {
3363 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3365 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3368 spin_unlock(&vmap_area_lock
);
3375 * Remove previously allocated areas. There is no
3376 * need in removing these areas from the busy tree,
3377 * because they are inserted only on the final step
3378 * and when pcpu_get_vm_areas() is success.
3381 orig_start
= vas
[area
]->va_start
;
3382 orig_end
= vas
[area
]->va_end
;
3383 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3384 &free_vmap_area_list
);
3385 kasan_release_vmalloc(orig_start
, orig_end
,
3386 va
->va_start
, va
->va_end
);
3391 spin_unlock(&free_vmap_area_lock
);
3393 purge_vmap_area_lazy();
3396 /* Before "retry", check if we recover. */
3397 for (area
= 0; area
< nr_vms
; area
++) {
3401 vas
[area
] = kmem_cache_zalloc(
3402 vmap_area_cachep
, GFP_KERNEL
);
3411 for (area
= 0; area
< nr_vms
; area
++) {
3413 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
3423 spin_lock(&free_vmap_area_lock
);
3425 * We release all the vmalloc shadows, even the ones for regions that
3426 * hadn't been successfully added. This relies on kasan_release_vmalloc
3427 * being able to tolerate this case.
3429 for (area
= 0; area
< nr_vms
; area
++) {
3430 orig_start
= vas
[area
]->va_start
;
3431 orig_end
= vas
[area
]->va_end
;
3432 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3433 &free_vmap_area_list
);
3434 kasan_release_vmalloc(orig_start
, orig_end
,
3435 va
->va_start
, va
->va_end
);
3439 spin_unlock(&free_vmap_area_lock
);
3446 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3447 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3448 * @nr_vms: the number of allocated areas
3450 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3452 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
3456 for (i
= 0; i
< nr_vms
; i
++)
3457 free_vm_area(vms
[i
]);
3460 #endif /* CONFIG_SMP */
3462 #ifdef CONFIG_PROC_FS
3463 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3464 __acquires(&vmap_purge_lock
)
3465 __acquires(&vmap_area_lock
)
3467 mutex_lock(&vmap_purge_lock
);
3468 spin_lock(&vmap_area_lock
);
3470 return seq_list_start(&vmap_area_list
, *pos
);
3473 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3475 return seq_list_next(p
, &vmap_area_list
, pos
);
3478 static void s_stop(struct seq_file
*m
, void *p
)
3479 __releases(&vmap_purge_lock
)
3480 __releases(&vmap_area_lock
)
3482 mutex_unlock(&vmap_purge_lock
);
3483 spin_unlock(&vmap_area_lock
);
3486 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
3488 if (IS_ENABLED(CONFIG_NUMA
)) {
3489 unsigned int nr
, *counters
= m
->private;
3494 if (v
->flags
& VM_UNINITIALIZED
)
3496 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3499 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
3501 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
3502 counters
[page_to_nid(v
->pages
[nr
])]++;
3504 for_each_node_state(nr
, N_HIGH_MEMORY
)
3506 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
3510 static void show_purge_info(struct seq_file
*m
)
3512 struct llist_node
*head
;
3513 struct vmap_area
*va
;
3515 head
= READ_ONCE(vmap_purge_list
.first
);
3519 llist_for_each_entry(va
, head
, purge_list
) {
3520 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3521 (void *)va
->va_start
, (void *)va
->va_end
,
3522 va
->va_end
- va
->va_start
);
3526 static int s_show(struct seq_file
*m
, void *p
)
3528 struct vmap_area
*va
;
3529 struct vm_struct
*v
;
3531 va
= list_entry(p
, struct vmap_area
, list
);
3534 * s_show can encounter race with remove_vm_area, !vm on behalf
3535 * of vmap area is being tear down or vm_map_ram allocation.
3538 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
3539 (void *)va
->va_start
, (void *)va
->va_end
,
3540 va
->va_end
- va
->va_start
);
3547 seq_printf(m
, "0x%pK-0x%pK %7ld",
3548 v
->addr
, v
->addr
+ v
->size
, v
->size
);
3551 seq_printf(m
, " %pS", v
->caller
);
3554 seq_printf(m
, " pages=%d", v
->nr_pages
);
3557 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
3559 if (v
->flags
& VM_IOREMAP
)
3560 seq_puts(m
, " ioremap");
3562 if (v
->flags
& VM_ALLOC
)
3563 seq_puts(m
, " vmalloc");
3565 if (v
->flags
& VM_MAP
)
3566 seq_puts(m
, " vmap");
3568 if (v
->flags
& VM_USERMAP
)
3569 seq_puts(m
, " user");
3571 if (v
->flags
& VM_DMA_COHERENT
)
3572 seq_puts(m
, " dma-coherent");
3574 if (is_vmalloc_addr(v
->pages
))
3575 seq_puts(m
, " vpages");
3577 show_numa_info(m
, v
);
3581 * As a final step, dump "unpurged" areas. Note,
3582 * that entire "/proc/vmallocinfo" output will not
3583 * be address sorted, because the purge list is not
3586 if (list_is_last(&va
->list
, &vmap_area_list
))
3592 static const struct seq_operations vmalloc_op
= {
3599 static int __init
proc_vmalloc_init(void)
3601 if (IS_ENABLED(CONFIG_NUMA
))
3602 proc_create_seq_private("vmallocinfo", 0400, NULL
,
3604 nr_node_ids
* sizeof(unsigned int), NULL
);
3606 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
3609 module_init(proc_vmalloc_init
);