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
10 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
13 #include <linux/vmalloc.h>
15 #include <linux/module.h>
16 #include <linux/highmem.h>
17 #include <linux/sched/signal.h>
18 #include <linux/slab.h>
19 #include <linux/spinlock.h>
20 #include <linux/interrupt.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/set_memory.h>
24 #include <linux/debugobjects.h>
25 #include <linux/kallsyms.h>
26 #include <linux/list.h>
27 #include <linux/notifier.h>
28 #include <linux/rbtree.h>
29 #include <linux/xarray.h>
30 #include <linux/rcupdate.h>
31 #include <linux/pfn.h>
32 #include <linux/kmemleak.h>
33 #include <linux/atomic.h>
34 #include <linux/compiler.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
40 #include <linux/uaccess.h>
41 #include <asm/tlbflush.h>
42 #include <asm/shmparam.h>
45 #include "pgalloc-track.h"
47 bool is_vmalloc_addr(const void *x
)
49 unsigned long addr
= (unsigned long)x
;
51 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
53 EXPORT_SYMBOL(is_vmalloc_addr
);
55 struct vfree_deferred
{
56 struct llist_head list
;
57 struct work_struct wq
;
59 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
61 static void __vunmap(const void *, int);
63 static void free_work(struct work_struct
*w
)
65 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
66 struct llist_node
*t
, *llnode
;
68 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
69 __vunmap((void *)llnode
, 1);
72 /*** Page table manipulation functions ***/
74 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
79 pte
= pte_offset_kernel(pmd
, addr
);
81 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
82 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
83 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
84 *mask
|= PGTBL_PTE_MODIFIED
;
87 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
94 pmd
= pmd_offset(pud
, addr
);
96 next
= pmd_addr_end(addr
, end
);
98 cleared
= pmd_clear_huge(pmd
);
99 if (cleared
|| pmd_bad(*pmd
))
100 *mask
|= PGTBL_PMD_MODIFIED
;
104 if (pmd_none_or_clear_bad(pmd
))
106 vunmap_pte_range(pmd
, addr
, next
, mask
);
109 } while (pmd
++, addr
= next
, addr
!= end
);
112 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
113 pgtbl_mod_mask
*mask
)
119 pud
= pud_offset(p4d
, addr
);
121 next
= pud_addr_end(addr
, end
);
123 cleared
= pud_clear_huge(pud
);
124 if (cleared
|| pud_bad(*pud
))
125 *mask
|= PGTBL_PUD_MODIFIED
;
129 if (pud_none_or_clear_bad(pud
))
131 vunmap_pmd_range(pud
, addr
, next
, mask
);
132 } while (pud
++, addr
= next
, addr
!= end
);
135 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
136 pgtbl_mod_mask
*mask
)
142 p4d
= p4d_offset(pgd
, addr
);
144 next
= p4d_addr_end(addr
, end
);
146 cleared
= p4d_clear_huge(p4d
);
147 if (cleared
|| p4d_bad(*p4d
))
148 *mask
|= PGTBL_P4D_MODIFIED
;
152 if (p4d_none_or_clear_bad(p4d
))
154 vunmap_pud_range(p4d
, addr
, next
, mask
);
155 } while (p4d
++, addr
= next
, addr
!= end
);
159 * unmap_kernel_range_noflush - unmap kernel VM area
160 * @start: start of the VM area to unmap
161 * @size: size of the VM area to unmap
163 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
164 * should have been allocated using get_vm_area() and its friends.
167 * This function does NOT do any cache flushing. The caller is responsible
168 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
169 * function and flush_tlb_kernel_range() after.
171 void unmap_kernel_range_noflush(unsigned long start
, unsigned long size
)
173 unsigned long end
= start
+ size
;
176 unsigned long addr
= start
;
177 pgtbl_mod_mask mask
= 0;
180 pgd
= pgd_offset_k(addr
);
182 next
= pgd_addr_end(addr
, end
);
184 mask
|= PGTBL_PGD_MODIFIED
;
185 if (pgd_none_or_clear_bad(pgd
))
187 vunmap_p4d_range(pgd
, addr
, next
, &mask
);
188 } while (pgd
++, addr
= next
, addr
!= end
);
190 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
191 arch_sync_kernel_mappings(start
, end
);
194 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
195 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
196 pgtbl_mod_mask
*mask
)
201 * nr is a running index into the array which helps higher level
202 * callers keep track of where we're up to.
205 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
209 struct page
*page
= pages
[*nr
];
211 if (WARN_ON(!pte_none(*pte
)))
215 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
217 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
218 *mask
|= PGTBL_PTE_MODIFIED
;
222 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
223 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
224 pgtbl_mod_mask
*mask
)
229 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
233 next
= pmd_addr_end(addr
, end
);
234 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
, mask
))
236 } while (pmd
++, addr
= next
, addr
!= end
);
240 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
241 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
242 pgtbl_mod_mask
*mask
)
247 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
251 next
= pud_addr_end(addr
, end
);
252 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
, mask
))
254 } while (pud
++, addr
= next
, addr
!= end
);
258 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
259 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
260 pgtbl_mod_mask
*mask
)
265 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
269 next
= p4d_addr_end(addr
, end
);
270 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
, mask
))
272 } while (p4d
++, addr
= next
, addr
!= end
);
277 * map_kernel_range_noflush - map kernel VM area with the specified pages
278 * @addr: start of the VM area to map
279 * @size: size of the VM area to map
280 * @prot: page protection flags to use
281 * @pages: pages to map
283 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
284 * have been allocated using get_vm_area() and its friends.
287 * This function does NOT do any cache flushing. The caller is responsible for
288 * calling flush_cache_vmap() on to-be-mapped areas before calling this
292 * 0 on success, -errno on failure.
294 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
295 pgprot_t prot
, struct page
**pages
)
297 unsigned long start
= addr
;
298 unsigned long end
= addr
+ size
;
303 pgtbl_mod_mask mask
= 0;
306 pgd
= pgd_offset_k(addr
);
308 next
= pgd_addr_end(addr
, end
);
310 mask
|= PGTBL_PGD_MODIFIED
;
311 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
, &mask
);
314 } while (pgd
++, addr
= next
, addr
!= end
);
316 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
317 arch_sync_kernel_mappings(start
, end
);
322 int map_kernel_range(unsigned long start
, unsigned long size
, pgprot_t prot
,
327 ret
= map_kernel_range_noflush(start
, size
, prot
, pages
);
328 flush_cache_vmap(start
, start
+ size
);
332 int is_vmalloc_or_module_addr(const void *x
)
335 * ARM, x86-64 and sparc64 put modules in a special place,
336 * and fall back on vmalloc() if that fails. Others
337 * just put it in the vmalloc space.
339 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
340 unsigned long addr
= (unsigned long)x
;
341 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
344 return is_vmalloc_addr(x
);
348 * Walk a vmap address to the struct page it maps.
350 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
352 unsigned long addr
= (unsigned long) vmalloc_addr
;
353 struct page
*page
= NULL
;
354 pgd_t
*pgd
= pgd_offset_k(addr
);
361 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
362 * architectures that do not vmalloc module space
364 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
368 p4d
= p4d_offset(pgd
, addr
);
371 pud
= pud_offset(p4d
, addr
);
374 * Don't dereference bad PUD or PMD (below) entries. This will also
375 * identify huge mappings, which we may encounter on architectures
376 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
377 * identified as vmalloc addresses by is_vmalloc_addr(), but are
378 * not [unambiguously] associated with a struct page, so there is
379 * no correct value to return for them.
381 WARN_ON_ONCE(pud_bad(*pud
));
382 if (pud_none(*pud
) || pud_bad(*pud
))
384 pmd
= pmd_offset(pud
, addr
);
385 WARN_ON_ONCE(pmd_bad(*pmd
));
386 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
389 ptep
= pte_offset_map(pmd
, addr
);
391 if (pte_present(pte
))
392 page
= pte_page(pte
);
396 EXPORT_SYMBOL(vmalloc_to_page
);
399 * Map a vmalloc()-space virtual address to the physical page frame number.
401 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
403 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
405 EXPORT_SYMBOL(vmalloc_to_pfn
);
408 /*** Global kva allocator ***/
410 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
411 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
414 static DEFINE_SPINLOCK(vmap_area_lock
);
415 static DEFINE_SPINLOCK(free_vmap_area_lock
);
416 /* Export for kexec only */
417 LIST_HEAD(vmap_area_list
);
418 static LLIST_HEAD(vmap_purge_list
);
419 static struct rb_root vmap_area_root
= RB_ROOT
;
420 static bool vmap_initialized __read_mostly
;
423 * This kmem_cache is used for vmap_area objects. Instead of
424 * allocating from slab we reuse an object from this cache to
425 * make things faster. Especially in "no edge" splitting of
428 static struct kmem_cache
*vmap_area_cachep
;
431 * This linked list is used in pair with free_vmap_area_root.
432 * It gives O(1) access to prev/next to perform fast coalescing.
434 static LIST_HEAD(free_vmap_area_list
);
437 * This augment red-black tree represents the free vmap space.
438 * All vmap_area objects in this tree are sorted by va->va_start
439 * address. It is used for allocation and merging when a vmap
440 * object is released.
442 * Each vmap_area node contains a maximum available free block
443 * of its sub-tree, right or left. Therefore it is possible to
444 * find a lowest match of free area.
446 static struct rb_root free_vmap_area_root
= RB_ROOT
;
449 * Preload a CPU with one object for "no edge" split case. The
450 * aim is to get rid of allocations from the atomic context, thus
451 * to use more permissive allocation masks.
453 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
455 static __always_inline
unsigned long
456 va_size(struct vmap_area
*va
)
458 return (va
->va_end
- va
->va_start
);
461 static __always_inline
unsigned long
462 get_subtree_max_size(struct rb_node
*node
)
464 struct vmap_area
*va
;
466 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
467 return va
? va
->subtree_max_size
: 0;
471 * Gets called when remove the node and rotate.
473 static __always_inline
unsigned long
474 compute_subtree_max_size(struct vmap_area
*va
)
476 return max3(va_size(va
),
477 get_subtree_max_size(va
->rb_node
.rb_left
),
478 get_subtree_max_size(va
->rb_node
.rb_right
));
481 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
482 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
484 static void purge_vmap_area_lazy(void);
485 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
486 static unsigned long lazy_max_pages(void);
488 static atomic_long_t nr_vmalloc_pages
;
490 unsigned long vmalloc_nr_pages(void)
492 return atomic_long_read(&nr_vmalloc_pages
);
495 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
497 struct rb_node
*n
= vmap_area_root
.rb_node
;
500 struct vmap_area
*va
;
502 va
= rb_entry(n
, struct vmap_area
, rb_node
);
503 if (addr
< va
->va_start
)
505 else if (addr
>= va
->va_end
)
515 * This function returns back addresses of parent node
516 * and its left or right link for further processing.
518 * Otherwise NULL is returned. In that case all further
519 * steps regarding inserting of conflicting overlap range
520 * have to be declined and actually considered as a bug.
522 static __always_inline
struct rb_node
**
523 find_va_links(struct vmap_area
*va
,
524 struct rb_root
*root
, struct rb_node
*from
,
525 struct rb_node
**parent
)
527 struct vmap_area
*tmp_va
;
528 struct rb_node
**link
;
531 link
= &root
->rb_node
;
532 if (unlikely(!*link
)) {
541 * Go to the bottom of the tree. When we hit the last point
542 * we end up with parent rb_node and correct direction, i name
543 * it link, where the new va->rb_node will be attached to.
546 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
549 * During the traversal we also do some sanity check.
550 * Trigger the BUG() if there are sides(left/right)
553 if (va
->va_start
< tmp_va
->va_end
&&
554 va
->va_end
<= tmp_va
->va_start
)
555 link
= &(*link
)->rb_left
;
556 else if (va
->va_end
> tmp_va
->va_start
&&
557 va
->va_start
>= tmp_va
->va_end
)
558 link
= &(*link
)->rb_right
;
560 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
561 va
->va_start
, va
->va_end
, tmp_va
->va_start
, tmp_va
->va_end
);
567 *parent
= &tmp_va
->rb_node
;
571 static __always_inline
struct list_head
*
572 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
574 struct list_head
*list
;
576 if (unlikely(!parent
))
578 * The red-black tree where we try to find VA neighbors
579 * before merging or inserting is empty, i.e. it means
580 * there is no free vmap space. Normally it does not
581 * happen but we handle this case anyway.
585 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
586 return (&parent
->rb_right
== link
? list
->next
: list
);
589 static __always_inline
void
590 link_va(struct vmap_area
*va
, struct rb_root
*root
,
591 struct rb_node
*parent
, struct rb_node
**link
, struct list_head
*head
)
594 * VA is still not in the list, but we can
595 * identify its future previous list_head node.
597 if (likely(parent
)) {
598 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
599 if (&parent
->rb_right
!= link
)
603 /* Insert to the rb-tree */
604 rb_link_node(&va
->rb_node
, parent
, link
);
605 if (root
== &free_vmap_area_root
) {
607 * Some explanation here. Just perform simple insertion
608 * to the tree. We do not set va->subtree_max_size to
609 * its current size before calling rb_insert_augmented().
610 * It is because of we populate the tree from the bottom
611 * to parent levels when the node _is_ in the tree.
613 * Therefore we set subtree_max_size to zero after insertion,
614 * to let __augment_tree_propagate_from() puts everything to
615 * the correct order later on.
617 rb_insert_augmented(&va
->rb_node
,
618 root
, &free_vmap_area_rb_augment_cb
);
619 va
->subtree_max_size
= 0;
621 rb_insert_color(&va
->rb_node
, root
);
624 /* Address-sort this list */
625 list_add(&va
->list
, head
);
628 static __always_inline
void
629 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
631 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
634 if (root
== &free_vmap_area_root
)
635 rb_erase_augmented(&va
->rb_node
,
636 root
, &free_vmap_area_rb_augment_cb
);
638 rb_erase(&va
->rb_node
, root
);
641 RB_CLEAR_NODE(&va
->rb_node
);
644 #if DEBUG_AUGMENT_PROPAGATE_CHECK
646 augment_tree_propagate_check(void)
648 struct vmap_area
*va
;
649 unsigned long computed_size
;
651 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
652 computed_size
= compute_subtree_max_size(va
);
653 if (computed_size
!= va
->subtree_max_size
)
654 pr_emerg("tree is corrupted: %lu, %lu\n",
655 va_size(va
), va
->subtree_max_size
);
661 * This function populates subtree_max_size from bottom to upper
662 * levels starting from VA point. The propagation must be done
663 * when VA size is modified by changing its va_start/va_end. Or
664 * in case of newly inserting of VA to the tree.
666 * It means that __augment_tree_propagate_from() must be called:
667 * - After VA has been inserted to the tree(free path);
668 * - After VA has been shrunk(allocation path);
669 * - After VA has been increased(merging path).
671 * Please note that, it does not mean that upper parent nodes
672 * and their subtree_max_size are recalculated all the time up
681 * For example if we modify the node 4, shrinking it to 2, then
682 * no any modification is required. If we shrink the node 2 to 1
683 * its subtree_max_size is updated only, and set to 1. If we shrink
684 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
687 static __always_inline
void
688 augment_tree_propagate_from(struct vmap_area
*va
)
691 * Populate the tree from bottom towards the root until
692 * the calculated maximum available size of checked node
693 * is equal to its current one.
695 free_vmap_area_rb_augment_cb_propagate(&va
->rb_node
, NULL
);
697 #if DEBUG_AUGMENT_PROPAGATE_CHECK
698 augment_tree_propagate_check();
703 insert_vmap_area(struct vmap_area
*va
,
704 struct rb_root
*root
, struct list_head
*head
)
706 struct rb_node
**link
;
707 struct rb_node
*parent
;
709 link
= find_va_links(va
, root
, NULL
, &parent
);
711 link_va(va
, root
, parent
, link
, head
);
715 insert_vmap_area_augment(struct vmap_area
*va
,
716 struct rb_node
*from
, struct rb_root
*root
,
717 struct list_head
*head
)
719 struct rb_node
**link
;
720 struct rb_node
*parent
;
723 link
= find_va_links(va
, NULL
, from
, &parent
);
725 link
= find_va_links(va
, root
, NULL
, &parent
);
728 link_va(va
, root
, parent
, link
, head
);
729 augment_tree_propagate_from(va
);
734 * Merge de-allocated chunk of VA memory with previous
735 * and next free blocks. If coalesce is not done a new
736 * free area is inserted. If VA has been merged, it is
739 * Please note, it can return NULL in case of overlap
740 * ranges, followed by WARN() report. Despite it is a
741 * buggy behaviour, a system can be alive and keep
744 static __always_inline
struct vmap_area
*
745 merge_or_add_vmap_area(struct vmap_area
*va
,
746 struct rb_root
*root
, struct list_head
*head
)
748 struct vmap_area
*sibling
;
749 struct list_head
*next
;
750 struct rb_node
**link
;
751 struct rb_node
*parent
;
755 * Find a place in the tree where VA potentially will be
756 * inserted, unless it is merged with its sibling/siblings.
758 link
= find_va_links(va
, root
, NULL
, &parent
);
763 * Get next node of VA to check if merging can be done.
765 next
= get_va_next_sibling(parent
, link
);
766 if (unlikely(next
== NULL
))
772 * |<------VA------>|<-----Next----->|
777 sibling
= list_entry(next
, struct vmap_area
, list
);
778 if (sibling
->va_start
== va
->va_end
) {
779 sibling
->va_start
= va
->va_start
;
781 /* Free vmap_area object. */
782 kmem_cache_free(vmap_area_cachep
, va
);
784 /* Point to the new merged area. */
793 * |<-----Prev----->|<------VA------>|
797 if (next
->prev
!= head
) {
798 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
799 if (sibling
->va_end
== va
->va_start
) {
801 * If both neighbors are coalesced, it is important
802 * to unlink the "next" node first, followed by merging
803 * with "previous" one. Otherwise the tree might not be
804 * fully populated if a sibling's augmented value is
805 * "normalized" because of rotation operations.
810 sibling
->va_end
= va
->va_end
;
812 /* Free vmap_area object. */
813 kmem_cache_free(vmap_area_cachep
, va
);
815 /* Point to the new merged area. */
823 link_va(va
, root
, parent
, link
, head
);
826 * Last step is to check and update the tree.
828 augment_tree_propagate_from(va
);
832 static __always_inline
bool
833 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
834 unsigned long align
, unsigned long vstart
)
836 unsigned long nva_start_addr
;
838 if (va
->va_start
> vstart
)
839 nva_start_addr
= ALIGN(va
->va_start
, align
);
841 nva_start_addr
= ALIGN(vstart
, align
);
843 /* Can be overflowed due to big size or alignment. */
844 if (nva_start_addr
+ size
< nva_start_addr
||
845 nva_start_addr
< vstart
)
848 return (nva_start_addr
+ size
<= va
->va_end
);
852 * Find the first free block(lowest start address) in the tree,
853 * that will accomplish the request corresponding to passing
856 static __always_inline
struct vmap_area
*
857 find_vmap_lowest_match(unsigned long size
,
858 unsigned long align
, unsigned long vstart
)
860 struct vmap_area
*va
;
861 struct rb_node
*node
;
862 unsigned long length
;
864 /* Start from the root. */
865 node
= free_vmap_area_root
.rb_node
;
867 /* Adjust the search size for alignment overhead. */
868 length
= size
+ align
- 1;
871 va
= rb_entry(node
, struct vmap_area
, rb_node
);
873 if (get_subtree_max_size(node
->rb_left
) >= length
&&
874 vstart
< va
->va_start
) {
875 node
= node
->rb_left
;
877 if (is_within_this_va(va
, size
, align
, vstart
))
881 * Does not make sense to go deeper towards the right
882 * sub-tree if it does not have a free block that is
883 * equal or bigger to the requested search length.
885 if (get_subtree_max_size(node
->rb_right
) >= length
) {
886 node
= node
->rb_right
;
891 * OK. We roll back and find the first right sub-tree,
892 * that will satisfy the search criteria. It can happen
893 * only once due to "vstart" restriction.
895 while ((node
= rb_parent(node
))) {
896 va
= rb_entry(node
, struct vmap_area
, rb_node
);
897 if (is_within_this_va(va
, size
, align
, vstart
))
900 if (get_subtree_max_size(node
->rb_right
) >= length
&&
901 vstart
<= va
->va_start
) {
902 node
= node
->rb_right
;
912 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
913 #include <linux/random.h>
915 static struct vmap_area
*
916 find_vmap_lowest_linear_match(unsigned long size
,
917 unsigned long align
, unsigned long vstart
)
919 struct vmap_area
*va
;
921 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
922 if (!is_within_this_va(va
, size
, align
, vstart
))
932 find_vmap_lowest_match_check(unsigned long size
)
934 struct vmap_area
*va_1
, *va_2
;
935 unsigned long vstart
;
938 get_random_bytes(&rnd
, sizeof(rnd
));
939 vstart
= VMALLOC_START
+ rnd
;
941 va_1
= find_vmap_lowest_match(size
, 1, vstart
);
942 va_2
= find_vmap_lowest_linear_match(size
, 1, vstart
);
945 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
952 FL_FIT_TYPE
= 1, /* full fit */
953 LE_FIT_TYPE
= 2, /* left edge fit */
954 RE_FIT_TYPE
= 3, /* right edge fit */
955 NE_FIT_TYPE
= 4 /* no edge fit */
958 static __always_inline
enum fit_type
959 classify_va_fit_type(struct vmap_area
*va
,
960 unsigned long nva_start_addr
, unsigned long size
)
964 /* Check if it is within VA. */
965 if (nva_start_addr
< va
->va_start
||
966 nva_start_addr
+ size
> va
->va_end
)
970 if (va
->va_start
== nva_start_addr
) {
971 if (va
->va_end
== nva_start_addr
+ size
)
975 } else if (va
->va_end
== nva_start_addr
+ size
) {
984 static __always_inline
int
985 adjust_va_to_fit_type(struct vmap_area
*va
,
986 unsigned long nva_start_addr
, unsigned long size
,
989 struct vmap_area
*lva
= NULL
;
991 if (type
== FL_FIT_TYPE
) {
993 * No need to split VA, it fully fits.
999 unlink_va(va
, &free_vmap_area_root
);
1000 kmem_cache_free(vmap_area_cachep
, va
);
1001 } else if (type
== LE_FIT_TYPE
) {
1003 * Split left edge of fit VA.
1009 va
->va_start
+= size
;
1010 } else if (type
== RE_FIT_TYPE
) {
1012 * Split right edge of fit VA.
1018 va
->va_end
= nva_start_addr
;
1019 } else if (type
== NE_FIT_TYPE
) {
1021 * Split no edge of fit VA.
1027 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1028 if (unlikely(!lva
)) {
1030 * For percpu allocator we do not do any pre-allocation
1031 * and leave it as it is. The reason is it most likely
1032 * never ends up with NE_FIT_TYPE splitting. In case of
1033 * percpu allocations offsets and sizes are aligned to
1034 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1035 * are its main fitting cases.
1037 * There are a few exceptions though, as an example it is
1038 * a first allocation (early boot up) when we have "one"
1039 * big free space that has to be split.
1041 * Also we can hit this path in case of regular "vmap"
1042 * allocations, if "this" current CPU was not preloaded.
1043 * See the comment in alloc_vmap_area() why. If so, then
1044 * GFP_NOWAIT is used instead to get an extra object for
1045 * split purpose. That is rare and most time does not
1048 * What happens if an allocation gets failed. Basically,
1049 * an "overflow" path is triggered to purge lazily freed
1050 * areas to free some memory, then, the "retry" path is
1051 * triggered to repeat one more time. See more details
1052 * in alloc_vmap_area() function.
1054 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1060 * Build the remainder.
1062 lva
->va_start
= va
->va_start
;
1063 lva
->va_end
= nva_start_addr
;
1066 * Shrink this VA to remaining size.
1068 va
->va_start
= nva_start_addr
+ size
;
1073 if (type
!= FL_FIT_TYPE
) {
1074 augment_tree_propagate_from(va
);
1076 if (lva
) /* type == NE_FIT_TYPE */
1077 insert_vmap_area_augment(lva
, &va
->rb_node
,
1078 &free_vmap_area_root
, &free_vmap_area_list
);
1085 * Returns a start address of the newly allocated area, if success.
1086 * Otherwise a vend is returned that indicates failure.
1088 static __always_inline
unsigned long
1089 __alloc_vmap_area(unsigned long size
, unsigned long align
,
1090 unsigned long vstart
, unsigned long vend
)
1092 unsigned long nva_start_addr
;
1093 struct vmap_area
*va
;
1097 va
= find_vmap_lowest_match(size
, align
, vstart
);
1101 if (va
->va_start
> vstart
)
1102 nva_start_addr
= ALIGN(va
->va_start
, align
);
1104 nva_start_addr
= ALIGN(vstart
, align
);
1106 /* Check the "vend" restriction. */
1107 if (nva_start_addr
+ size
> vend
)
1110 /* Classify what we have found. */
1111 type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1112 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
1115 /* Update the free vmap_area. */
1116 ret
= adjust_va_to_fit_type(va
, nva_start_addr
, size
, type
);
1120 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1121 find_vmap_lowest_match_check(size
);
1124 return nva_start_addr
;
1128 * Free a region of KVA allocated by alloc_vmap_area
1130 static void free_vmap_area(struct vmap_area
*va
)
1133 * Remove from the busy tree/list.
1135 spin_lock(&vmap_area_lock
);
1136 unlink_va(va
, &vmap_area_root
);
1137 spin_unlock(&vmap_area_lock
);
1140 * Insert/Merge it back to the free tree/list.
1142 spin_lock(&free_vmap_area_lock
);
1143 merge_or_add_vmap_area(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1144 spin_unlock(&free_vmap_area_lock
);
1148 * Allocate a region of KVA of the specified size and alignment, within the
1151 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1152 unsigned long align
,
1153 unsigned long vstart
, unsigned long vend
,
1154 int node
, gfp_t gfp_mask
)
1156 struct vmap_area
*va
, *pva
;
1162 BUG_ON(offset_in_page(size
));
1163 BUG_ON(!is_power_of_2(align
));
1165 if (unlikely(!vmap_initialized
))
1166 return ERR_PTR(-EBUSY
);
1169 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1171 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1173 return ERR_PTR(-ENOMEM
);
1176 * Only scan the relevant parts containing pointers to other objects
1177 * to avoid false negatives.
1179 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1183 * Preload this CPU with one extra vmap_area object. It is used
1184 * when fit type of free area is NE_FIT_TYPE. Please note, it
1185 * does not guarantee that an allocation occurs on a CPU that
1186 * is preloaded, instead we minimize the case when it is not.
1187 * It can happen because of cpu migration, because there is a
1188 * race until the below spinlock is taken.
1190 * The preload is done in non-atomic context, thus it allows us
1191 * to use more permissive allocation masks to be more stable under
1192 * low memory condition and high memory pressure. In rare case,
1193 * if not preloaded, GFP_NOWAIT is used.
1195 * Set "pva" to NULL here, because of "retry" path.
1199 if (!this_cpu_read(ne_fit_preload_node
))
1201 * Even if it fails we do not really care about that.
1202 * Just proceed as it is. If needed "overflow" path
1203 * will refill the cache we allocate from.
1205 pva
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1207 spin_lock(&free_vmap_area_lock
);
1209 if (pva
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, pva
))
1210 kmem_cache_free(vmap_area_cachep
, pva
);
1213 * If an allocation fails, the "vend" address is
1214 * returned. Therefore trigger the overflow path.
1216 addr
= __alloc_vmap_area(size
, align
, vstart
, vend
);
1217 spin_unlock(&free_vmap_area_lock
);
1219 if (unlikely(addr
== vend
))
1222 va
->va_start
= addr
;
1223 va
->va_end
= addr
+ size
;
1227 spin_lock(&vmap_area_lock
);
1228 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1229 spin_unlock(&vmap_area_lock
);
1231 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1232 BUG_ON(va
->va_start
< vstart
);
1233 BUG_ON(va
->va_end
> vend
);
1235 ret
= kasan_populate_vmalloc(addr
, size
);
1238 return ERR_PTR(ret
);
1245 purge_vmap_area_lazy();
1250 if (gfpflags_allow_blocking(gfp_mask
)) {
1251 unsigned long freed
= 0;
1252 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1259 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1260 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1263 kmem_cache_free(vmap_area_cachep
, va
);
1264 return ERR_PTR(-EBUSY
);
1267 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1269 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1271 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1273 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1275 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1277 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1280 * lazy_max_pages is the maximum amount of virtual address space we gather up
1281 * before attempting to purge with a TLB flush.
1283 * There is a tradeoff here: a larger number will cover more kernel page tables
1284 * and take slightly longer to purge, but it will linearly reduce the number of
1285 * global TLB flushes that must be performed. It would seem natural to scale
1286 * this number up linearly with the number of CPUs (because vmapping activity
1287 * could also scale linearly with the number of CPUs), however it is likely
1288 * that in practice, workloads might be constrained in other ways that mean
1289 * vmap activity will not scale linearly with CPUs. Also, I want to be
1290 * conservative and not introduce a big latency on huge systems, so go with
1291 * a less aggressive log scale. It will still be an improvement over the old
1292 * code, and it will be simple to change the scale factor if we find that it
1293 * becomes a problem on bigger systems.
1295 static unsigned long lazy_max_pages(void)
1299 log
= fls(num_online_cpus());
1301 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1304 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1307 * Serialize vmap purging. There is no actual criticial section protected
1308 * by this look, but we want to avoid concurrent calls for performance
1309 * reasons and to make the pcpu_get_vm_areas more deterministic.
1311 static DEFINE_MUTEX(vmap_purge_lock
);
1313 /* for per-CPU blocks */
1314 static void purge_fragmented_blocks_allcpus(void);
1317 * called before a call to iounmap() if the caller wants vm_area_struct's
1318 * immediately freed.
1320 void set_iounmap_nonlazy(void)
1322 atomic_long_set(&vmap_lazy_nr
, lazy_max_pages()+1);
1326 * Purges all lazily-freed vmap areas.
1328 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1330 unsigned long resched_threshold
;
1331 struct llist_node
*valist
;
1332 struct vmap_area
*va
;
1333 struct vmap_area
*n_va
;
1335 lockdep_assert_held(&vmap_purge_lock
);
1337 valist
= llist_del_all(&vmap_purge_list
);
1338 if (unlikely(valist
== NULL
))
1342 * TODO: to calculate a flush range without looping.
1343 * The list can be up to lazy_max_pages() elements.
1345 llist_for_each_entry(va
, valist
, purge_list
) {
1346 if (va
->va_start
< start
)
1347 start
= va
->va_start
;
1348 if (va
->va_end
> end
)
1352 flush_tlb_kernel_range(start
, end
);
1353 resched_threshold
= lazy_max_pages() << 1;
1355 spin_lock(&free_vmap_area_lock
);
1356 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
1357 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1358 unsigned long orig_start
= va
->va_start
;
1359 unsigned long orig_end
= va
->va_end
;
1362 * Finally insert or merge lazily-freed area. It is
1363 * detached and there is no need to "unlink" it from
1366 va
= merge_or_add_vmap_area(va
, &free_vmap_area_root
,
1367 &free_vmap_area_list
);
1372 if (is_vmalloc_or_module_addr((void *)orig_start
))
1373 kasan_release_vmalloc(orig_start
, orig_end
,
1374 va
->va_start
, va
->va_end
);
1376 atomic_long_sub(nr
, &vmap_lazy_nr
);
1378 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1379 cond_resched_lock(&free_vmap_area_lock
);
1381 spin_unlock(&free_vmap_area_lock
);
1386 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1387 * is already purging.
1389 static void try_purge_vmap_area_lazy(void)
1391 if (mutex_trylock(&vmap_purge_lock
)) {
1392 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1393 mutex_unlock(&vmap_purge_lock
);
1398 * Kick off a purge of the outstanding lazy areas.
1400 static void purge_vmap_area_lazy(void)
1402 mutex_lock(&vmap_purge_lock
);
1403 purge_fragmented_blocks_allcpus();
1404 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1405 mutex_unlock(&vmap_purge_lock
);
1409 * Free a vmap area, caller ensuring that the area has been unmapped
1410 * and flush_cache_vunmap had been called for the correct range
1413 static void free_vmap_area_noflush(struct vmap_area
*va
)
1415 unsigned long nr_lazy
;
1417 spin_lock(&vmap_area_lock
);
1418 unlink_va(va
, &vmap_area_root
);
1419 spin_unlock(&vmap_area_lock
);
1421 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1422 PAGE_SHIFT
, &vmap_lazy_nr
);
1424 /* After this point, we may free va at any time */
1425 llist_add(&va
->purge_list
, &vmap_purge_list
);
1427 if (unlikely(nr_lazy
> lazy_max_pages()))
1428 try_purge_vmap_area_lazy();
1432 * Free and unmap a vmap area
1434 static void free_unmap_vmap_area(struct vmap_area
*va
)
1436 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1437 unmap_kernel_range_noflush(va
->va_start
, va
->va_end
- va
->va_start
);
1438 if (debug_pagealloc_enabled_static())
1439 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1441 free_vmap_area_noflush(va
);
1444 static struct vmap_area
*find_vmap_area(unsigned long addr
)
1446 struct vmap_area
*va
;
1448 spin_lock(&vmap_area_lock
);
1449 va
= __find_vmap_area(addr
);
1450 spin_unlock(&vmap_area_lock
);
1455 /*** Per cpu kva allocator ***/
1458 * vmap space is limited especially on 32 bit architectures. Ensure there is
1459 * room for at least 16 percpu vmap blocks per CPU.
1462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1463 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1464 * instead (we just need a rough idea)
1466 #if BITS_PER_LONG == 32
1467 #define VMALLOC_SPACE (128UL*1024*1024)
1469 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1472 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1473 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1474 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1475 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1476 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1477 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1478 #define VMAP_BBMAP_BITS \
1479 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1480 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1481 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1483 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1485 struct vmap_block_queue
{
1487 struct list_head free
;
1492 struct vmap_area
*va
;
1493 unsigned long free
, dirty
;
1494 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1495 struct list_head free_list
;
1496 struct rcu_head rcu_head
;
1497 struct list_head purge
;
1500 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1501 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1504 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1505 * in the free path. Could get rid of this if we change the API to return a
1506 * "cookie" from alloc, to be passed to free. But no big deal yet.
1508 static DEFINE_XARRAY(vmap_blocks
);
1511 * We should probably have a fallback mechanism to allocate virtual memory
1512 * out of partially filled vmap blocks. However vmap block sizing should be
1513 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1517 static unsigned long addr_to_vb_idx(unsigned long addr
)
1519 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1520 addr
/= VMAP_BLOCK_SIZE
;
1524 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1528 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1529 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1530 return (void *)addr
;
1534 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1535 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1536 * @order: how many 2^order pages should be occupied in newly allocated block
1537 * @gfp_mask: flags for the page level allocator
1539 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1541 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1543 struct vmap_block_queue
*vbq
;
1544 struct vmap_block
*vb
;
1545 struct vmap_area
*va
;
1546 unsigned long vb_idx
;
1550 node
= numa_node_id();
1552 vb
= kmalloc_node(sizeof(struct vmap_block
),
1553 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1555 return ERR_PTR(-ENOMEM
);
1557 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1558 VMALLOC_START
, VMALLOC_END
,
1562 return ERR_CAST(va
);
1565 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1566 spin_lock_init(&vb
->lock
);
1568 /* At least something should be left free */
1569 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1570 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
1572 vb
->dirty_min
= VMAP_BBMAP_BITS
;
1574 INIT_LIST_HEAD(&vb
->free_list
);
1576 vb_idx
= addr_to_vb_idx(va
->va_start
);
1577 err
= xa_insert(&vmap_blocks
, vb_idx
, vb
, gfp_mask
);
1581 return ERR_PTR(err
);
1584 vbq
= &get_cpu_var(vmap_block_queue
);
1585 spin_lock(&vbq
->lock
);
1586 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
1587 spin_unlock(&vbq
->lock
);
1588 put_cpu_var(vmap_block_queue
);
1593 static void free_vmap_block(struct vmap_block
*vb
)
1595 struct vmap_block
*tmp
;
1597 tmp
= xa_erase(&vmap_blocks
, addr_to_vb_idx(vb
->va
->va_start
));
1600 free_vmap_area_noflush(vb
->va
);
1601 kfree_rcu(vb
, rcu_head
);
1604 static void purge_fragmented_blocks(int cpu
)
1607 struct vmap_block
*vb
;
1608 struct vmap_block
*n_vb
;
1609 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1612 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1614 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
1617 spin_lock(&vb
->lock
);
1618 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
1619 vb
->free
= 0; /* prevent further allocs after releasing lock */
1620 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
1622 vb
->dirty_max
= VMAP_BBMAP_BITS
;
1623 spin_lock(&vbq
->lock
);
1624 list_del_rcu(&vb
->free_list
);
1625 spin_unlock(&vbq
->lock
);
1626 spin_unlock(&vb
->lock
);
1627 list_add_tail(&vb
->purge
, &purge
);
1629 spin_unlock(&vb
->lock
);
1633 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
1634 list_del(&vb
->purge
);
1635 free_vmap_block(vb
);
1639 static void purge_fragmented_blocks_allcpus(void)
1643 for_each_possible_cpu(cpu
)
1644 purge_fragmented_blocks(cpu
);
1647 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
1649 struct vmap_block_queue
*vbq
;
1650 struct vmap_block
*vb
;
1654 BUG_ON(offset_in_page(size
));
1655 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1656 if (WARN_ON(size
== 0)) {
1658 * Allocating 0 bytes isn't what caller wants since
1659 * get_order(0) returns funny result. Just warn and terminate
1664 order
= get_order(size
);
1667 vbq
= &get_cpu_var(vmap_block_queue
);
1668 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1669 unsigned long pages_off
;
1671 spin_lock(&vb
->lock
);
1672 if (vb
->free
< (1UL << order
)) {
1673 spin_unlock(&vb
->lock
);
1677 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1678 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1679 vb
->free
-= 1UL << order
;
1680 if (vb
->free
== 0) {
1681 spin_lock(&vbq
->lock
);
1682 list_del_rcu(&vb
->free_list
);
1683 spin_unlock(&vbq
->lock
);
1686 spin_unlock(&vb
->lock
);
1690 put_cpu_var(vmap_block_queue
);
1693 /* Allocate new block if nothing was found */
1695 vaddr
= new_vmap_block(order
, gfp_mask
);
1700 static void vb_free(unsigned long addr
, unsigned long size
)
1702 unsigned long offset
;
1704 struct vmap_block
*vb
;
1706 BUG_ON(offset_in_page(size
));
1707 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1709 flush_cache_vunmap(addr
, addr
+ size
);
1711 order
= get_order(size
);
1712 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
1713 vb
= xa_load(&vmap_blocks
, addr_to_vb_idx(addr
));
1715 unmap_kernel_range_noflush(addr
, size
);
1717 if (debug_pagealloc_enabled_static())
1718 flush_tlb_kernel_range(addr
, addr
+ size
);
1720 spin_lock(&vb
->lock
);
1722 /* Expand dirty range */
1723 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1724 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1726 vb
->dirty
+= 1UL << order
;
1727 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1729 spin_unlock(&vb
->lock
);
1730 free_vmap_block(vb
);
1732 spin_unlock(&vb
->lock
);
1735 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
1739 if (unlikely(!vmap_initialized
))
1744 for_each_possible_cpu(cpu
) {
1745 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1746 struct vmap_block
*vb
;
1749 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1750 spin_lock(&vb
->lock
);
1752 unsigned long va_start
= vb
->va
->va_start
;
1755 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1756 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1758 start
= min(s
, start
);
1763 spin_unlock(&vb
->lock
);
1768 mutex_lock(&vmap_purge_lock
);
1769 purge_fragmented_blocks_allcpus();
1770 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1771 flush_tlb_kernel_range(start
, end
);
1772 mutex_unlock(&vmap_purge_lock
);
1776 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1778 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1779 * to amortize TLB flushing overheads. What this means is that any page you
1780 * have now, may, in a former life, have been mapped into kernel virtual
1781 * address by the vmap layer and so there might be some CPUs with TLB entries
1782 * still referencing that page (additional to the regular 1:1 kernel mapping).
1784 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1785 * be sure that none of the pages we have control over will have any aliases
1786 * from the vmap layer.
1788 void vm_unmap_aliases(void)
1790 unsigned long start
= ULONG_MAX
, end
= 0;
1793 _vm_unmap_aliases(start
, end
, flush
);
1795 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1798 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1799 * @mem: the pointer returned by vm_map_ram
1800 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1802 void vm_unmap_ram(const void *mem
, unsigned int count
)
1804 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1805 unsigned long addr
= (unsigned long)mem
;
1806 struct vmap_area
*va
;
1810 BUG_ON(addr
< VMALLOC_START
);
1811 BUG_ON(addr
> VMALLOC_END
);
1812 BUG_ON(!PAGE_ALIGNED(addr
));
1814 kasan_poison_vmalloc(mem
, size
);
1816 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1817 debug_check_no_locks_freed(mem
, size
);
1818 vb_free(addr
, size
);
1822 va
= find_vmap_area(addr
);
1824 debug_check_no_locks_freed((void *)va
->va_start
,
1825 (va
->va_end
- va
->va_start
));
1826 free_unmap_vmap_area(va
);
1828 EXPORT_SYMBOL(vm_unmap_ram
);
1831 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1832 * @pages: an array of pointers to the pages to be mapped
1833 * @count: number of pages
1834 * @node: prefer to allocate data structures on this node
1836 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1837 * faster than vmap so it's good. But if you mix long-life and short-life
1838 * objects with vm_map_ram(), it could consume lots of address space through
1839 * fragmentation (especially on a 32bit machine). You could see failures in
1840 * the end. Please use this function for short-lived objects.
1842 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1844 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
1846 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1850 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1851 mem
= vb_alloc(size
, GFP_KERNEL
);
1854 addr
= (unsigned long)mem
;
1856 struct vmap_area
*va
;
1857 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1858 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1862 addr
= va
->va_start
;
1866 kasan_unpoison_vmalloc(mem
, size
);
1868 if (map_kernel_range(addr
, size
, PAGE_KERNEL
, pages
) < 0) {
1869 vm_unmap_ram(mem
, count
);
1874 EXPORT_SYMBOL(vm_map_ram
);
1876 static struct vm_struct
*vmlist __initdata
;
1879 * vm_area_add_early - add vmap area early during boot
1880 * @vm: vm_struct to add
1882 * This function is used to add fixed kernel vm area to vmlist before
1883 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1884 * should contain proper values and the other fields should be zero.
1886 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1888 void __init
vm_area_add_early(struct vm_struct
*vm
)
1890 struct vm_struct
*tmp
, **p
;
1892 BUG_ON(vmap_initialized
);
1893 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1894 if (tmp
->addr
>= vm
->addr
) {
1895 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1898 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1905 * vm_area_register_early - register vmap area early during boot
1906 * @vm: vm_struct to register
1907 * @align: requested alignment
1909 * This function is used to register kernel vm area before
1910 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1911 * proper values on entry and other fields should be zero. On return,
1912 * vm->addr contains the allocated address.
1914 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1916 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1918 static size_t vm_init_off __initdata
;
1921 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1922 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1924 vm
->addr
= (void *)addr
;
1926 vm_area_add_early(vm
);
1929 static void vmap_init_free_space(void)
1931 unsigned long vmap_start
= 1;
1932 const unsigned long vmap_end
= ULONG_MAX
;
1933 struct vmap_area
*busy
, *free
;
1937 * -|-----|.....|-----|-----|-----|.....|-
1939 * |<--------------------------------->|
1941 list_for_each_entry(busy
, &vmap_area_list
, list
) {
1942 if (busy
->va_start
- vmap_start
> 0) {
1943 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1944 if (!WARN_ON_ONCE(!free
)) {
1945 free
->va_start
= vmap_start
;
1946 free
->va_end
= busy
->va_start
;
1948 insert_vmap_area_augment(free
, NULL
,
1949 &free_vmap_area_root
,
1950 &free_vmap_area_list
);
1954 vmap_start
= busy
->va_end
;
1957 if (vmap_end
- vmap_start
> 0) {
1958 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1959 if (!WARN_ON_ONCE(!free
)) {
1960 free
->va_start
= vmap_start
;
1961 free
->va_end
= vmap_end
;
1963 insert_vmap_area_augment(free
, NULL
,
1964 &free_vmap_area_root
,
1965 &free_vmap_area_list
);
1970 void __init
vmalloc_init(void)
1972 struct vmap_area
*va
;
1973 struct vm_struct
*tmp
;
1977 * Create the cache for vmap_area objects.
1979 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
1981 for_each_possible_cpu(i
) {
1982 struct vmap_block_queue
*vbq
;
1983 struct vfree_deferred
*p
;
1985 vbq
= &per_cpu(vmap_block_queue
, i
);
1986 spin_lock_init(&vbq
->lock
);
1987 INIT_LIST_HEAD(&vbq
->free
);
1988 p
= &per_cpu(vfree_deferred
, i
);
1989 init_llist_head(&p
->list
);
1990 INIT_WORK(&p
->wq
, free_work
);
1993 /* Import existing vmlist entries. */
1994 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1995 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1996 if (WARN_ON_ONCE(!va
))
1999 va
->va_start
= (unsigned long)tmp
->addr
;
2000 va
->va_end
= va
->va_start
+ tmp
->size
;
2002 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
2006 * Now we can initialize a free vmap space.
2008 vmap_init_free_space();
2009 vmap_initialized
= true;
2013 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2014 * @addr: start of the VM area to unmap
2015 * @size: size of the VM area to unmap
2017 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2018 * the unmapping and tlb after.
2020 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
2022 unsigned long end
= addr
+ size
;
2024 flush_cache_vunmap(addr
, end
);
2025 unmap_kernel_range_noflush(addr
, size
);
2026 flush_tlb_kernel_range(addr
, end
);
2029 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2030 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2033 vm
->addr
= (void *)va
->va_start
;
2034 vm
->size
= va
->va_end
- va
->va_start
;
2035 vm
->caller
= caller
;
2039 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2040 unsigned long flags
, const void *caller
)
2042 spin_lock(&vmap_area_lock
);
2043 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2044 spin_unlock(&vmap_area_lock
);
2047 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2050 * Before removing VM_UNINITIALIZED,
2051 * we should make sure that vm has proper values.
2052 * Pair with smp_rmb() in show_numa_info().
2055 vm
->flags
&= ~VM_UNINITIALIZED
;
2058 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2059 unsigned long align
, unsigned long flags
, unsigned long start
,
2060 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
2062 struct vmap_area
*va
;
2063 struct vm_struct
*area
;
2064 unsigned long requested_size
= size
;
2066 BUG_ON(in_interrupt());
2067 size
= PAGE_ALIGN(size
);
2068 if (unlikely(!size
))
2071 if (flags
& VM_IOREMAP
)
2072 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2073 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2075 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2076 if (unlikely(!area
))
2079 if (!(flags
& VM_NO_GUARD
))
2082 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
2088 kasan_unpoison_vmalloc((void *)va
->va_start
, requested_size
);
2090 setup_vmalloc_vm(area
, va
, flags
, caller
);
2095 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2096 unsigned long start
, unsigned long end
,
2099 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2100 GFP_KERNEL
, caller
);
2104 * get_vm_area - reserve a contiguous kernel virtual area
2105 * @size: size of the area
2106 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2108 * Search an area of @size in the kernel virtual mapping area,
2109 * and reserved it for out purposes. Returns the area descriptor
2110 * on success or %NULL on failure.
2112 * Return: the area descriptor on success or %NULL on failure.
2114 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2116 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2117 NUMA_NO_NODE
, GFP_KERNEL
,
2118 __builtin_return_address(0));
2121 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2124 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2125 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2129 * find_vm_area - find a continuous kernel virtual area
2130 * @addr: base address
2132 * Search for the kernel VM area starting at @addr, and return it.
2133 * It is up to the caller to do all required locking to keep the returned
2136 * Return: pointer to the found area or %NULL on faulure
2138 struct vm_struct
*find_vm_area(const void *addr
)
2140 struct vmap_area
*va
;
2142 va
= find_vmap_area((unsigned long)addr
);
2150 * remove_vm_area - find and remove a continuous kernel virtual area
2151 * @addr: base address
2153 * Search for the kernel VM area starting at @addr, and remove it.
2154 * This function returns the found VM area, but using it is NOT safe
2155 * on SMP machines, except for its size or flags.
2157 * Return: pointer to the found area or %NULL on faulure
2159 struct vm_struct
*remove_vm_area(const void *addr
)
2161 struct vmap_area
*va
;
2165 spin_lock(&vmap_area_lock
);
2166 va
= __find_vmap_area((unsigned long)addr
);
2168 struct vm_struct
*vm
= va
->vm
;
2171 spin_unlock(&vmap_area_lock
);
2173 kasan_free_shadow(vm
);
2174 free_unmap_vmap_area(va
);
2179 spin_unlock(&vmap_area_lock
);
2183 static inline void set_area_direct_map(const struct vm_struct
*area
,
2184 int (*set_direct_map
)(struct page
*page
))
2188 for (i
= 0; i
< area
->nr_pages
; i
++)
2189 if (page_address(area
->pages
[i
]))
2190 set_direct_map(area
->pages
[i
]);
2193 /* Handle removing and resetting vm mappings related to the vm_struct. */
2194 static void vm_remove_mappings(struct vm_struct
*area
, int deallocate_pages
)
2196 unsigned long start
= ULONG_MAX
, end
= 0;
2197 int flush_reset
= area
->flags
& VM_FLUSH_RESET_PERMS
;
2201 remove_vm_area(area
->addr
);
2203 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2208 * If not deallocating pages, just do the flush of the VM area and
2211 if (!deallocate_pages
) {
2217 * If execution gets here, flush the vm mapping and reset the direct
2218 * map. Find the start and end range of the direct mappings to make sure
2219 * the vm_unmap_aliases() flush includes the direct map.
2221 for (i
= 0; i
< area
->nr_pages
; i
++) {
2222 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2224 start
= min(addr
, start
);
2225 end
= max(addr
+ PAGE_SIZE
, end
);
2231 * Set direct map to something invalid so that it won't be cached if
2232 * there are any accesses after the TLB flush, then flush the TLB and
2233 * reset the direct map permissions to the default.
2235 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2236 _vm_unmap_aliases(start
, end
, flush_dmap
);
2237 set_area_direct_map(area
, set_direct_map_default_noflush
);
2240 static void __vunmap(const void *addr
, int deallocate_pages
)
2242 struct vm_struct
*area
;
2247 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2251 area
= find_vm_area(addr
);
2252 if (unlikely(!area
)) {
2253 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2258 debug_check_no_locks_freed(area
->addr
, get_vm_area_size(area
));
2259 debug_check_no_obj_freed(area
->addr
, get_vm_area_size(area
));
2261 kasan_poison_vmalloc(area
->addr
, area
->size
);
2263 vm_remove_mappings(area
, deallocate_pages
);
2265 if (deallocate_pages
) {
2268 for (i
= 0; i
< area
->nr_pages
; i
++) {
2269 struct page
*page
= area
->pages
[i
];
2272 __free_pages(page
, 0);
2274 atomic_long_sub(area
->nr_pages
, &nr_vmalloc_pages
);
2276 kvfree(area
->pages
);
2283 static inline void __vfree_deferred(const void *addr
)
2286 * Use raw_cpu_ptr() because this can be called from preemptible
2287 * context. Preemption is absolutely fine here, because the llist_add()
2288 * implementation is lockless, so it works even if we are adding to
2289 * another cpu's list. schedule_work() should be fine with this too.
2291 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2293 if (llist_add((struct llist_node
*)addr
, &p
->list
))
2294 schedule_work(&p
->wq
);
2298 * vfree_atomic - release memory allocated by vmalloc()
2299 * @addr: memory base address
2301 * This one is just like vfree() but can be called in any atomic context
2304 void vfree_atomic(const void *addr
)
2308 kmemleak_free(addr
);
2312 __vfree_deferred(addr
);
2315 static void __vfree(const void *addr
)
2317 if (unlikely(in_interrupt()))
2318 __vfree_deferred(addr
);
2324 * vfree - release memory allocated by vmalloc()
2325 * @addr: memory base address
2327 * Free the virtually continuous memory area starting at @addr, as
2328 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2329 * NULL, no operation is performed.
2331 * Must not be called in NMI context (strictly speaking, only if we don't
2332 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2333 * conventions for vfree() arch-depenedent would be a really bad idea)
2335 * May sleep if called *not* from interrupt context.
2337 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2339 void vfree(const void *addr
)
2343 kmemleak_free(addr
);
2345 might_sleep_if(!in_interrupt());
2352 EXPORT_SYMBOL(vfree
);
2355 * vunmap - release virtual mapping obtained by vmap()
2356 * @addr: memory base address
2358 * Free the virtually contiguous memory area starting at @addr,
2359 * which was created from the page array passed to vmap().
2361 * Must not be called in interrupt context.
2363 void vunmap(const void *addr
)
2365 BUG_ON(in_interrupt());
2370 EXPORT_SYMBOL(vunmap
);
2373 * vmap - map an array of pages into virtually contiguous space
2374 * @pages: array of page pointers
2375 * @count: number of pages to map
2376 * @flags: vm_area->flags
2377 * @prot: page protection for the mapping
2379 * Maps @count pages from @pages into contiguous kernel virtual
2382 * Return: the address of the area or %NULL on failure
2384 void *vmap(struct page
**pages
, unsigned int count
,
2385 unsigned long flags
, pgprot_t prot
)
2387 struct vm_struct
*area
;
2388 unsigned long size
; /* In bytes */
2392 if (count
> totalram_pages())
2395 size
= (unsigned long)count
<< PAGE_SHIFT
;
2396 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2400 if (map_kernel_range((unsigned long)area
->addr
, size
, pgprot_nx(prot
),
2408 EXPORT_SYMBOL(vmap
);
2410 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2411 pgprot_t prot
, int node
)
2413 struct page
**pages
;
2414 unsigned int nr_pages
, array_size
, i
;
2415 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2416 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
2417 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
2421 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
2422 array_size
= (nr_pages
* sizeof(struct page
*));
2424 /* Please note that the recursion is strictly bounded. */
2425 if (array_size
> PAGE_SIZE
) {
2426 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
2427 node
, area
->caller
);
2429 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
2433 remove_vm_area(area
->addr
);
2438 area
->pages
= pages
;
2439 area
->nr_pages
= nr_pages
;
2441 for (i
= 0; i
< area
->nr_pages
; i
++) {
2444 if (node
== NUMA_NO_NODE
)
2445 page
= alloc_page(alloc_mask
|highmem_mask
);
2447 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
2449 if (unlikely(!page
)) {
2450 /* Successfully allocated i pages, free them in __vunmap() */
2452 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2455 area
->pages
[i
] = page
;
2456 if (gfpflags_allow_blocking(gfp_mask
))
2459 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2461 if (map_kernel_range((unsigned long)area
->addr
, get_vm_area_size(area
),
2468 warn_alloc(gfp_mask
, NULL
,
2469 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2470 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
2471 __vfree(area
->addr
);
2476 * __vmalloc_node_range - allocate virtually contiguous memory
2477 * @size: allocation size
2478 * @align: desired alignment
2479 * @start: vm area range start
2480 * @end: vm area range end
2481 * @gfp_mask: flags for the page level allocator
2482 * @prot: protection mask for the allocated pages
2483 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2484 * @node: node to use for allocation or NUMA_NO_NODE
2485 * @caller: caller's return address
2487 * Allocate enough pages to cover @size from the page level
2488 * allocator with @gfp_mask flags. Map them into contiguous
2489 * kernel virtual space, using a pagetable protection of @prot.
2491 * Return: the address of the area or %NULL on failure
2493 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
2494 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
2495 pgprot_t prot
, unsigned long vm_flags
, int node
,
2498 struct vm_struct
*area
;
2500 unsigned long real_size
= size
;
2502 size
= PAGE_ALIGN(size
);
2503 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages())
2506 area
= __get_vm_area_node(real_size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
2507 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
2511 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
2516 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2517 * flag. It means that vm_struct is not fully initialized.
2518 * Now, it is fully initialized, so remove this flag here.
2520 clear_vm_uninitialized_flag(area
);
2522 kmemleak_vmalloc(area
, size
, gfp_mask
);
2527 warn_alloc(gfp_mask
, NULL
,
2528 "vmalloc: allocation failure: %lu bytes", real_size
);
2533 * __vmalloc_node - allocate virtually contiguous memory
2534 * @size: allocation size
2535 * @align: desired alignment
2536 * @gfp_mask: flags for the page level allocator
2537 * @node: node to use for allocation or NUMA_NO_NODE
2538 * @caller: caller's return address
2540 * Allocate enough pages to cover @size from the page level allocator with
2541 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2543 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2544 * and __GFP_NOFAIL are not supported
2546 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2549 * Return: pointer to the allocated memory or %NULL on error
2551 void *__vmalloc_node(unsigned long size
, unsigned long align
,
2552 gfp_t gfp_mask
, int node
, const void *caller
)
2554 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
2555 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
2558 * This is only for performance analysis of vmalloc and stress purpose.
2559 * It is required by vmalloc test module, therefore do not use it other
2562 #ifdef CONFIG_TEST_VMALLOC_MODULE
2563 EXPORT_SYMBOL_GPL(__vmalloc_node
);
2566 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
2568 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
2569 __builtin_return_address(0));
2571 EXPORT_SYMBOL(__vmalloc
);
2574 * vmalloc - allocate virtually contiguous memory
2575 * @size: allocation size
2577 * Allocate enough pages to cover @size from the page level
2578 * allocator and map them into contiguous kernel virtual space.
2580 * For tight control over page level allocator and protection flags
2581 * use __vmalloc() instead.
2583 * Return: pointer to the allocated memory or %NULL on error
2585 void *vmalloc(unsigned long size
)
2587 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
2588 __builtin_return_address(0));
2590 EXPORT_SYMBOL(vmalloc
);
2593 * vzalloc - allocate virtually contiguous memory with zero fill
2594 * @size: allocation size
2596 * Allocate enough pages to cover @size from the page level
2597 * allocator and map them into contiguous kernel virtual space.
2598 * The memory allocated is set to zero.
2600 * For tight control over page level allocator and protection flags
2601 * use __vmalloc() instead.
2603 * Return: pointer to the allocated memory or %NULL on error
2605 void *vzalloc(unsigned long size
)
2607 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
2608 __builtin_return_address(0));
2610 EXPORT_SYMBOL(vzalloc
);
2613 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2614 * @size: allocation size
2616 * The resulting memory area is zeroed so it can be mapped to userspace
2617 * without leaking data.
2619 * Return: pointer to the allocated memory or %NULL on error
2621 void *vmalloc_user(unsigned long size
)
2623 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2624 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
2625 VM_USERMAP
, NUMA_NO_NODE
,
2626 __builtin_return_address(0));
2628 EXPORT_SYMBOL(vmalloc_user
);
2631 * vmalloc_node - allocate memory on a specific node
2632 * @size: allocation size
2635 * Allocate enough pages to cover @size from the page level
2636 * allocator and map them into contiguous kernel virtual space.
2638 * For tight control over page level allocator and protection flags
2639 * use __vmalloc() instead.
2641 * Return: pointer to the allocated memory or %NULL on error
2643 void *vmalloc_node(unsigned long size
, int node
)
2645 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
2646 __builtin_return_address(0));
2648 EXPORT_SYMBOL(vmalloc_node
);
2651 * vzalloc_node - allocate memory on a specific node with zero fill
2652 * @size: allocation size
2655 * Allocate enough pages to cover @size from the page level
2656 * allocator and map them into contiguous kernel virtual space.
2657 * The memory allocated is set to zero.
2659 * Return: pointer to the allocated memory or %NULL on error
2661 void *vzalloc_node(unsigned long size
, int node
)
2663 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
2664 __builtin_return_address(0));
2666 EXPORT_SYMBOL(vzalloc_node
);
2668 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2669 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2670 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2671 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2674 * 64b systems should always have either DMA or DMA32 zones. For others
2675 * GFP_DMA32 should do the right thing and use the normal zone.
2677 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2681 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2682 * @size: allocation size
2684 * Allocate enough 32bit PA addressable pages to cover @size from the
2685 * page level allocator and map them into contiguous kernel virtual space.
2687 * Return: pointer to the allocated memory or %NULL on error
2689 void *vmalloc_32(unsigned long size
)
2691 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
2692 __builtin_return_address(0));
2694 EXPORT_SYMBOL(vmalloc_32
);
2697 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2698 * @size: allocation size
2700 * The resulting memory area is 32bit addressable and zeroed so it can be
2701 * mapped to userspace without leaking data.
2703 * Return: pointer to the allocated memory or %NULL on error
2705 void *vmalloc_32_user(unsigned long size
)
2707 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2708 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2709 VM_USERMAP
, NUMA_NO_NODE
,
2710 __builtin_return_address(0));
2712 EXPORT_SYMBOL(vmalloc_32_user
);
2715 * small helper routine , copy contents to buf from addr.
2716 * If the page is not present, fill zero.
2719 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2725 unsigned long offset
, length
;
2727 offset
= offset_in_page(addr
);
2728 length
= PAGE_SIZE
- offset
;
2731 p
= vmalloc_to_page(addr
);
2733 * To do safe access to this _mapped_ area, we need
2734 * lock. But adding lock here means that we need to add
2735 * overhead of vmalloc()/vfree() calles for this _debug_
2736 * interface, rarely used. Instead of that, we'll use
2737 * kmap() and get small overhead in this access function.
2741 * we can expect USER0 is not used (see vread/vwrite's
2742 * function description)
2744 void *map
= kmap_atomic(p
);
2745 memcpy(buf
, map
+ offset
, length
);
2748 memset(buf
, 0, length
);
2758 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2764 unsigned long offset
, length
;
2766 offset
= offset_in_page(addr
);
2767 length
= PAGE_SIZE
- offset
;
2770 p
= vmalloc_to_page(addr
);
2772 * To do safe access to this _mapped_ area, we need
2773 * lock. But adding lock here means that we need to add
2774 * overhead of vmalloc()/vfree() calles for this _debug_
2775 * interface, rarely used. Instead of that, we'll use
2776 * kmap() and get small overhead in this access function.
2780 * we can expect USER0 is not used (see vread/vwrite's
2781 * function description)
2783 void *map
= kmap_atomic(p
);
2784 memcpy(map
+ offset
, buf
, length
);
2796 * vread() - read vmalloc area in a safe way.
2797 * @buf: buffer for reading data
2798 * @addr: vm address.
2799 * @count: number of bytes to be read.
2801 * This function checks that addr is a valid vmalloc'ed area, and
2802 * copy data from that area to a given buffer. If the given memory range
2803 * of [addr...addr+count) includes some valid address, data is copied to
2804 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2805 * IOREMAP area is treated as memory hole and no copy is done.
2807 * If [addr...addr+count) doesn't includes any intersects with alive
2808 * vm_struct area, returns 0. @buf should be kernel's buffer.
2810 * Note: In usual ops, vread() is never necessary because the caller
2811 * should know vmalloc() area is valid and can use memcpy().
2812 * This is for routines which have to access vmalloc area without
2813 * any information, as /dev/kmem.
2815 * Return: number of bytes for which addr and buf should be increased
2816 * (same number as @count) or %0 if [addr...addr+count) doesn't
2817 * include any intersection with valid vmalloc area
2819 long vread(char *buf
, char *addr
, unsigned long count
)
2821 struct vmap_area
*va
;
2822 struct vm_struct
*vm
;
2823 char *vaddr
, *buf_start
= buf
;
2824 unsigned long buflen
= count
;
2827 /* Don't allow overflow */
2828 if ((unsigned long) addr
+ count
< count
)
2829 count
= -(unsigned long) addr
;
2831 spin_lock(&vmap_area_lock
);
2832 list_for_each_entry(va
, &vmap_area_list
, list
) {
2840 vaddr
= (char *) vm
->addr
;
2841 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2843 while (addr
< vaddr
) {
2851 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2854 if (!(vm
->flags
& VM_IOREMAP
))
2855 aligned_vread(buf
, addr
, n
);
2856 else /* IOREMAP area is treated as memory hole */
2863 spin_unlock(&vmap_area_lock
);
2865 if (buf
== buf_start
)
2867 /* zero-fill memory holes */
2868 if (buf
!= buf_start
+ buflen
)
2869 memset(buf
, 0, buflen
- (buf
- buf_start
));
2875 * vwrite() - write vmalloc area in a safe way.
2876 * @buf: buffer for source data
2877 * @addr: vm address.
2878 * @count: number of bytes to be read.
2880 * This function checks that addr is a valid vmalloc'ed area, and
2881 * copy data from a buffer to the given addr. If specified range of
2882 * [addr...addr+count) includes some valid address, data is copied from
2883 * proper area of @buf. If there are memory holes, no copy to hole.
2884 * IOREMAP area is treated as memory hole and no copy is done.
2886 * If [addr...addr+count) doesn't includes any intersects with alive
2887 * vm_struct area, returns 0. @buf should be kernel's buffer.
2889 * Note: In usual ops, vwrite() is never necessary because the caller
2890 * should know vmalloc() area is valid and can use memcpy().
2891 * This is for routines which have to access vmalloc area without
2892 * any information, as /dev/kmem.
2894 * Return: number of bytes for which addr and buf should be
2895 * increased (same number as @count) or %0 if [addr...addr+count)
2896 * doesn't include any intersection with valid vmalloc area
2898 long vwrite(char *buf
, char *addr
, unsigned long count
)
2900 struct vmap_area
*va
;
2901 struct vm_struct
*vm
;
2903 unsigned long n
, buflen
;
2906 /* Don't allow overflow */
2907 if ((unsigned long) addr
+ count
< count
)
2908 count
= -(unsigned long) addr
;
2911 spin_lock(&vmap_area_lock
);
2912 list_for_each_entry(va
, &vmap_area_list
, list
) {
2920 vaddr
= (char *) vm
->addr
;
2921 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2923 while (addr
< vaddr
) {
2930 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2933 if (!(vm
->flags
& VM_IOREMAP
)) {
2934 aligned_vwrite(buf
, addr
, n
);
2942 spin_unlock(&vmap_area_lock
);
2949 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2950 * @vma: vma to cover
2951 * @uaddr: target user address to start at
2952 * @kaddr: virtual address of vmalloc kernel memory
2953 * @pgoff: offset from @kaddr to start at
2954 * @size: size of map area
2956 * Returns: 0 for success, -Exxx on failure
2958 * This function checks that @kaddr is a valid vmalloc'ed area,
2959 * and that it is big enough to cover the range starting at
2960 * @uaddr in @vma. Will return failure if that criteria isn't
2963 * Similar to remap_pfn_range() (see mm/memory.c)
2965 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
2966 void *kaddr
, unsigned long pgoff
,
2969 struct vm_struct
*area
;
2971 unsigned long end_index
;
2973 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
2976 size
= PAGE_ALIGN(size
);
2978 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
2981 area
= find_vm_area(kaddr
);
2985 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
2988 if (check_add_overflow(size
, off
, &end_index
) ||
2989 end_index
> get_vm_area_size(area
))
2994 struct page
*page
= vmalloc_to_page(kaddr
);
2997 ret
= vm_insert_page(vma
, uaddr
, page
);
3006 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3010 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
3013 * remap_vmalloc_range - map vmalloc pages to userspace
3014 * @vma: vma to cover (map full range of vma)
3015 * @addr: vmalloc memory
3016 * @pgoff: number of pages into addr before first page to map
3018 * Returns: 0 for success, -Exxx on failure
3020 * This function checks that addr is a valid vmalloc'ed area, and
3021 * that it is big enough to cover the vma. Will return failure if
3022 * that criteria isn't met.
3024 * Similar to remap_pfn_range() (see mm/memory.c)
3026 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3027 unsigned long pgoff
)
3029 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3031 vma
->vm_end
- vma
->vm_start
);
3033 EXPORT_SYMBOL(remap_vmalloc_range
);
3035 static int f(pte_t
*pte
, unsigned long addr
, void *data
)
3047 * alloc_vm_area - allocate a range of kernel address space
3048 * @size: size of the area
3049 * @ptes: returns the PTEs for the address space
3051 * Returns: NULL on failure, vm_struct on success
3053 * This function reserves a range of kernel address space, and
3054 * allocates pagetables to map that range. No actual mappings
3057 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3058 * allocated for the VM area are returned.
3060 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
3062 struct vm_struct
*area
;
3064 area
= get_vm_area_caller(size
, VM_IOREMAP
,
3065 __builtin_return_address(0));
3070 * This ensures that page tables are constructed for this region
3071 * of kernel virtual address space and mapped into init_mm.
3073 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3074 size
, f
, ptes
? &ptes
: NULL
)) {
3081 EXPORT_SYMBOL_GPL(alloc_vm_area
);
3083 void free_vm_area(struct vm_struct
*area
)
3085 struct vm_struct
*ret
;
3086 ret
= remove_vm_area(area
->addr
);
3087 BUG_ON(ret
!= area
);
3090 EXPORT_SYMBOL_GPL(free_vm_area
);
3093 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3095 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3099 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3100 * @addr: target address
3102 * Returns: vmap_area if it is found. If there is no such area
3103 * the first highest(reverse order) vmap_area is returned
3104 * i.e. va->va_start < addr && va->va_end < addr or NULL
3105 * if there are no any areas before @addr.
3107 static struct vmap_area
*
3108 pvm_find_va_enclose_addr(unsigned long addr
)
3110 struct vmap_area
*va
, *tmp
;
3113 n
= free_vmap_area_root
.rb_node
;
3117 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3118 if (tmp
->va_start
<= addr
) {
3120 if (tmp
->va_end
>= addr
)
3133 * pvm_determine_end_from_reverse - find the highest aligned address
3134 * of free block below VMALLOC_END
3136 * in - the VA we start the search(reverse order);
3137 * out - the VA with the highest aligned end address.
3139 * Returns: determined end address within vmap_area
3141 static unsigned long
3142 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3144 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3148 list_for_each_entry_from_reverse((*va
),
3149 &free_vmap_area_list
, list
) {
3150 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3151 if ((*va
)->va_start
< addr
)
3160 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3161 * @offsets: array containing offset of each area
3162 * @sizes: array containing size of each area
3163 * @nr_vms: the number of areas to allocate
3164 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3166 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3167 * vm_structs on success, %NULL on failure
3169 * Percpu allocator wants to use congruent vm areas so that it can
3170 * maintain the offsets among percpu areas. This function allocates
3171 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3172 * be scattered pretty far, distance between two areas easily going up
3173 * to gigabytes. To avoid interacting with regular vmallocs, these
3174 * areas are allocated from top.
3176 * Despite its complicated look, this allocator is rather simple. It
3177 * does everything top-down and scans free blocks from the end looking
3178 * for matching base. While scanning, if any of the areas do not fit the
3179 * base address is pulled down to fit the area. Scanning is repeated till
3180 * all the areas fit and then all necessary data structures are inserted
3181 * and the result is returned.
3183 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3184 const size_t *sizes
, int nr_vms
,
3187 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3188 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3189 struct vmap_area
**vas
, *va
;
3190 struct vm_struct
**vms
;
3191 int area
, area2
, last_area
, term_area
;
3192 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3193 bool purged
= false;
3196 /* verify parameters and allocate data structures */
3197 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3198 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3199 start
= offsets
[area
];
3200 end
= start
+ sizes
[area
];
3202 /* is everything aligned properly? */
3203 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3204 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3206 /* detect the area with the highest address */
3207 if (start
> offsets
[last_area
])
3210 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3211 unsigned long start2
= offsets
[area2
];
3212 unsigned long end2
= start2
+ sizes
[area2
];
3214 BUG_ON(start2
< end
&& start
< end2
);
3217 last_end
= offsets
[last_area
] + sizes
[last_area
];
3219 if (vmalloc_end
- vmalloc_start
< last_end
) {
3224 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3225 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3229 for (area
= 0; area
< nr_vms
; area
++) {
3230 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3231 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3232 if (!vas
[area
] || !vms
[area
])
3236 spin_lock(&free_vmap_area_lock
);
3238 /* start scanning - we scan from the top, begin with the last area */
3239 area
= term_area
= last_area
;
3240 start
= offsets
[area
];
3241 end
= start
+ sizes
[area
];
3243 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3244 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3248 * base might have underflowed, add last_end before
3251 if (base
+ last_end
< vmalloc_start
+ last_end
)
3255 * Fitting base has not been found.
3261 * If required width exceeds current VA block, move
3262 * base downwards and then recheck.
3264 if (base
+ end
> va
->va_end
) {
3265 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3271 * If this VA does not fit, move base downwards and recheck.
3273 if (base
+ start
< va
->va_start
) {
3274 va
= node_to_va(rb_prev(&va
->rb_node
));
3275 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3281 * This area fits, move on to the previous one. If
3282 * the previous one is the terminal one, we're done.
3284 area
= (area
+ nr_vms
- 1) % nr_vms
;
3285 if (area
== term_area
)
3288 start
= offsets
[area
];
3289 end
= start
+ sizes
[area
];
3290 va
= pvm_find_va_enclose_addr(base
+ end
);
3293 /* we've found a fitting base, insert all va's */
3294 for (area
= 0; area
< nr_vms
; area
++) {
3297 start
= base
+ offsets
[area
];
3300 va
= pvm_find_va_enclose_addr(start
);
3301 if (WARN_ON_ONCE(va
== NULL
))
3302 /* It is a BUG(), but trigger recovery instead. */
3305 type
= classify_va_fit_type(va
, start
, size
);
3306 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
3307 /* It is a BUG(), but trigger recovery instead. */
3310 ret
= adjust_va_to_fit_type(va
, start
, size
, type
);
3314 /* Allocated area. */
3316 va
->va_start
= start
;
3317 va
->va_end
= start
+ size
;
3320 spin_unlock(&free_vmap_area_lock
);
3322 /* populate the kasan shadow space */
3323 for (area
= 0; area
< nr_vms
; area
++) {
3324 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3325 goto err_free_shadow
;
3327 kasan_unpoison_vmalloc((void *)vas
[area
]->va_start
,
3331 /* insert all vm's */
3332 spin_lock(&vmap_area_lock
);
3333 for (area
= 0; area
< nr_vms
; area
++) {
3334 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3336 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3339 spin_unlock(&vmap_area_lock
);
3346 * Remove previously allocated areas. There is no
3347 * need in removing these areas from the busy tree,
3348 * because they are inserted only on the final step
3349 * and when pcpu_get_vm_areas() is success.
3352 orig_start
= vas
[area
]->va_start
;
3353 orig_end
= vas
[area
]->va_end
;
3354 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3355 &free_vmap_area_list
);
3357 kasan_release_vmalloc(orig_start
, orig_end
,
3358 va
->va_start
, va
->va_end
);
3363 spin_unlock(&free_vmap_area_lock
);
3365 purge_vmap_area_lazy();
3368 /* Before "retry", check if we recover. */
3369 for (area
= 0; area
< nr_vms
; area
++) {
3373 vas
[area
] = kmem_cache_zalloc(
3374 vmap_area_cachep
, GFP_KERNEL
);
3383 for (area
= 0; area
< nr_vms
; area
++) {
3385 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
3395 spin_lock(&free_vmap_area_lock
);
3397 * We release all the vmalloc shadows, even the ones for regions that
3398 * hadn't been successfully added. This relies on kasan_release_vmalloc
3399 * being able to tolerate this case.
3401 for (area
= 0; area
< nr_vms
; area
++) {
3402 orig_start
= vas
[area
]->va_start
;
3403 orig_end
= vas
[area
]->va_end
;
3404 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3405 &free_vmap_area_list
);
3407 kasan_release_vmalloc(orig_start
, orig_end
,
3408 va
->va_start
, va
->va_end
);
3412 spin_unlock(&free_vmap_area_lock
);
3419 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3420 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3421 * @nr_vms: the number of allocated areas
3423 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3425 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
3429 for (i
= 0; i
< nr_vms
; i
++)
3430 free_vm_area(vms
[i
]);
3433 #endif /* CONFIG_SMP */
3435 #ifdef CONFIG_PROC_FS
3436 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3437 __acquires(&vmap_purge_lock
)
3438 __acquires(&vmap_area_lock
)
3440 mutex_lock(&vmap_purge_lock
);
3441 spin_lock(&vmap_area_lock
);
3443 return seq_list_start(&vmap_area_list
, *pos
);
3446 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3448 return seq_list_next(p
, &vmap_area_list
, pos
);
3451 static void s_stop(struct seq_file
*m
, void *p
)
3452 __releases(&vmap_purge_lock
)
3453 __releases(&vmap_area_lock
)
3455 mutex_unlock(&vmap_purge_lock
);
3456 spin_unlock(&vmap_area_lock
);
3459 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
3461 if (IS_ENABLED(CONFIG_NUMA
)) {
3462 unsigned int nr
, *counters
= m
->private;
3467 if (v
->flags
& VM_UNINITIALIZED
)
3469 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3472 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
3474 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
3475 counters
[page_to_nid(v
->pages
[nr
])]++;
3477 for_each_node_state(nr
, N_HIGH_MEMORY
)
3479 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
3483 static void show_purge_info(struct seq_file
*m
)
3485 struct llist_node
*head
;
3486 struct vmap_area
*va
;
3488 head
= READ_ONCE(vmap_purge_list
.first
);
3492 llist_for_each_entry(va
, head
, purge_list
) {
3493 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3494 (void *)va
->va_start
, (void *)va
->va_end
,
3495 va
->va_end
- va
->va_start
);
3499 static int s_show(struct seq_file
*m
, void *p
)
3501 struct vmap_area
*va
;
3502 struct vm_struct
*v
;
3504 va
= list_entry(p
, struct vmap_area
, list
);
3507 * s_show can encounter race with remove_vm_area, !vm on behalf
3508 * of vmap area is being tear down or vm_map_ram allocation.
3511 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
3512 (void *)va
->va_start
, (void *)va
->va_end
,
3513 va
->va_end
- va
->va_start
);
3520 seq_printf(m
, "0x%pK-0x%pK %7ld",
3521 v
->addr
, v
->addr
+ v
->size
, v
->size
);
3524 seq_printf(m
, " %pS", v
->caller
);
3527 seq_printf(m
, " pages=%d", v
->nr_pages
);
3530 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
3532 if (v
->flags
& VM_IOREMAP
)
3533 seq_puts(m
, " ioremap");
3535 if (v
->flags
& VM_ALLOC
)
3536 seq_puts(m
, " vmalloc");
3538 if (v
->flags
& VM_MAP
)
3539 seq_puts(m
, " vmap");
3541 if (v
->flags
& VM_USERMAP
)
3542 seq_puts(m
, " user");
3544 if (v
->flags
& VM_DMA_COHERENT
)
3545 seq_puts(m
, " dma-coherent");
3547 if (is_vmalloc_addr(v
->pages
))
3548 seq_puts(m
, " vpages");
3550 show_numa_info(m
, v
);
3554 * As a final step, dump "unpurged" areas. Note,
3555 * that entire "/proc/vmallocinfo" output will not
3556 * be address sorted, because the purge list is not
3559 if (list_is_last(&va
->list
, &vmap_area_list
))
3565 static const struct seq_operations vmalloc_op
= {
3572 static int __init
proc_vmalloc_init(void)
3574 if (IS_ENABLED(CONFIG_NUMA
))
3575 proc_create_seq_private("vmallocinfo", 0400, NULL
,
3577 nr_node_ids
* sizeof(unsigned int), NULL
);
3579 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
3582 module_init(proc_vmalloc_init
);