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
,
77 pte
= pte_offset_kernel(pmd
, addr
);
79 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
80 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
81 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
82 *mask
|= PGTBL_PTE_MODIFIED
;
85 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
92 pmd
= pmd_offset(pud
, addr
);
94 next
= pmd_addr_end(addr
, end
);
96 cleared
= pmd_clear_huge(pmd
);
97 if (cleared
|| pmd_bad(*pmd
))
98 *mask
|= PGTBL_PMD_MODIFIED
;
102 if (pmd_none_or_clear_bad(pmd
))
104 vunmap_pte_range(pmd
, addr
, next
, mask
);
105 } while (pmd
++, addr
= next
, addr
!= end
);
108 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
109 pgtbl_mod_mask
*mask
)
115 pud
= pud_offset(p4d
, addr
);
117 next
= pud_addr_end(addr
, end
);
119 cleared
= pud_clear_huge(pud
);
120 if (cleared
|| pud_bad(*pud
))
121 *mask
|= PGTBL_PUD_MODIFIED
;
125 if (pud_none_or_clear_bad(pud
))
127 vunmap_pmd_range(pud
, addr
, next
, mask
);
128 } while (pud
++, addr
= next
, addr
!= end
);
131 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
132 pgtbl_mod_mask
*mask
)
138 p4d
= p4d_offset(pgd
, addr
);
140 next
= p4d_addr_end(addr
, end
);
142 cleared
= p4d_clear_huge(p4d
);
143 if (cleared
|| p4d_bad(*p4d
))
144 *mask
|= PGTBL_P4D_MODIFIED
;
148 if (p4d_none_or_clear_bad(p4d
))
150 vunmap_pud_range(p4d
, addr
, next
, mask
);
151 } while (p4d
++, addr
= next
, addr
!= end
);
155 * unmap_kernel_range_noflush - unmap kernel VM area
156 * @start: start of the VM area to unmap
157 * @size: size of the VM area to unmap
159 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
160 * should have been allocated using get_vm_area() and its friends.
163 * This function does NOT do any cache flushing. The caller is responsible
164 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
165 * function and flush_tlb_kernel_range() after.
167 void unmap_kernel_range_noflush(unsigned long start
, unsigned long size
)
169 unsigned long end
= start
+ size
;
172 unsigned long addr
= start
;
173 pgtbl_mod_mask mask
= 0;
177 pgd
= pgd_offset_k(addr
);
179 next
= pgd_addr_end(addr
, end
);
181 mask
|= PGTBL_PGD_MODIFIED
;
182 if (pgd_none_or_clear_bad(pgd
))
184 vunmap_p4d_range(pgd
, addr
, next
, &mask
);
185 } while (pgd
++, addr
= next
, addr
!= end
);
187 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
188 arch_sync_kernel_mappings(start
, end
);
191 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
192 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
193 pgtbl_mod_mask
*mask
)
198 * nr is a running index into the array which helps higher level
199 * callers keep track of where we're up to.
202 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
206 struct page
*page
= pages
[*nr
];
208 if (WARN_ON(!pte_none(*pte
)))
212 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
214 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
215 *mask
|= PGTBL_PTE_MODIFIED
;
219 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
220 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
221 pgtbl_mod_mask
*mask
)
226 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
230 next
= pmd_addr_end(addr
, end
);
231 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
, mask
))
233 } while (pmd
++, addr
= next
, addr
!= end
);
237 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
238 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
239 pgtbl_mod_mask
*mask
)
244 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
248 next
= pud_addr_end(addr
, end
);
249 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
, mask
))
251 } while (pud
++, addr
= next
, addr
!= end
);
255 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
256 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
257 pgtbl_mod_mask
*mask
)
262 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
266 next
= p4d_addr_end(addr
, end
);
267 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
, mask
))
269 } while (p4d
++, addr
= next
, addr
!= end
);
274 * map_kernel_range_noflush - map kernel VM area with the specified pages
275 * @addr: start of the VM area to map
276 * @size: size of the VM area to map
277 * @prot: page protection flags to use
278 * @pages: pages to map
280 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
281 * have been allocated using get_vm_area() and its friends.
284 * This function does NOT do any cache flushing. The caller is responsible for
285 * calling flush_cache_vmap() on to-be-mapped areas before calling this
289 * 0 on success, -errno on failure.
291 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
292 pgprot_t prot
, struct page
**pages
)
294 unsigned long start
= addr
;
295 unsigned long end
= addr
+ size
;
300 pgtbl_mod_mask mask
= 0;
303 pgd
= pgd_offset_k(addr
);
305 next
= pgd_addr_end(addr
, end
);
307 mask
|= PGTBL_PGD_MODIFIED
;
308 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
, &mask
);
311 } while (pgd
++, addr
= next
, addr
!= end
);
313 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
314 arch_sync_kernel_mappings(start
, end
);
319 int map_kernel_range(unsigned long start
, unsigned long size
, pgprot_t prot
,
324 ret
= map_kernel_range_noflush(start
, size
, prot
, pages
);
325 flush_cache_vmap(start
, start
+ size
);
329 int is_vmalloc_or_module_addr(const void *x
)
332 * ARM, x86-64 and sparc64 put modules in a special place,
333 * and fall back on vmalloc() if that fails. Others
334 * just put it in the vmalloc space.
336 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
337 unsigned long addr
= (unsigned long)x
;
338 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
341 return is_vmalloc_addr(x
);
345 * Walk a vmap address to the struct page it maps.
347 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
349 unsigned long addr
= (unsigned long) vmalloc_addr
;
350 struct page
*page
= NULL
;
351 pgd_t
*pgd
= pgd_offset_k(addr
);
358 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
359 * architectures that do not vmalloc module space
361 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
365 p4d
= p4d_offset(pgd
, addr
);
368 pud
= pud_offset(p4d
, addr
);
371 * Don't dereference bad PUD or PMD (below) entries. This will also
372 * identify huge mappings, which we may encounter on architectures
373 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
374 * identified as vmalloc addresses by is_vmalloc_addr(), but are
375 * not [unambiguously] associated with a struct page, so there is
376 * no correct value to return for them.
378 WARN_ON_ONCE(pud_bad(*pud
));
379 if (pud_none(*pud
) || pud_bad(*pud
))
381 pmd
= pmd_offset(pud
, addr
);
382 WARN_ON_ONCE(pmd_bad(*pmd
));
383 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
386 ptep
= pte_offset_map(pmd
, addr
);
388 if (pte_present(pte
))
389 page
= pte_page(pte
);
393 EXPORT_SYMBOL(vmalloc_to_page
);
396 * Map a vmalloc()-space virtual address to the physical page frame number.
398 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
400 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
402 EXPORT_SYMBOL(vmalloc_to_pfn
);
405 /*** Global kva allocator ***/
407 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
408 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
411 static DEFINE_SPINLOCK(vmap_area_lock
);
412 static DEFINE_SPINLOCK(free_vmap_area_lock
);
413 /* Export for kexec only */
414 LIST_HEAD(vmap_area_list
);
415 static LLIST_HEAD(vmap_purge_list
);
416 static struct rb_root vmap_area_root
= RB_ROOT
;
417 static bool vmap_initialized __read_mostly
;
420 * This kmem_cache is used for vmap_area objects. Instead of
421 * allocating from slab we reuse an object from this cache to
422 * make things faster. Especially in "no edge" splitting of
425 static struct kmem_cache
*vmap_area_cachep
;
428 * This linked list is used in pair with free_vmap_area_root.
429 * It gives O(1) access to prev/next to perform fast coalescing.
431 static LIST_HEAD(free_vmap_area_list
);
434 * This augment red-black tree represents the free vmap space.
435 * All vmap_area objects in this tree are sorted by va->va_start
436 * address. It is used for allocation and merging when a vmap
437 * object is released.
439 * Each vmap_area node contains a maximum available free block
440 * of its sub-tree, right or left. Therefore it is possible to
441 * find a lowest match of free area.
443 static struct rb_root free_vmap_area_root
= RB_ROOT
;
446 * Preload a CPU with one object for "no edge" split case. The
447 * aim is to get rid of allocations from the atomic context, thus
448 * to use more permissive allocation masks.
450 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
452 static __always_inline
unsigned long
453 va_size(struct vmap_area
*va
)
455 return (va
->va_end
- va
->va_start
);
458 static __always_inline
unsigned long
459 get_subtree_max_size(struct rb_node
*node
)
461 struct vmap_area
*va
;
463 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
464 return va
? va
->subtree_max_size
: 0;
468 * Gets called when remove the node and rotate.
470 static __always_inline
unsigned long
471 compute_subtree_max_size(struct vmap_area
*va
)
473 return max3(va_size(va
),
474 get_subtree_max_size(va
->rb_node
.rb_left
),
475 get_subtree_max_size(va
->rb_node
.rb_right
));
478 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
479 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
481 static void purge_vmap_area_lazy(void);
482 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
483 static unsigned long lazy_max_pages(void);
485 static atomic_long_t nr_vmalloc_pages
;
487 unsigned long vmalloc_nr_pages(void)
489 return atomic_long_read(&nr_vmalloc_pages
);
492 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
494 struct rb_node
*n
= vmap_area_root
.rb_node
;
497 struct vmap_area
*va
;
499 va
= rb_entry(n
, struct vmap_area
, rb_node
);
500 if (addr
< va
->va_start
)
502 else if (addr
>= va
->va_end
)
512 * This function returns back addresses of parent node
513 * and its left or right link for further processing.
515 static __always_inline
struct rb_node
**
516 find_va_links(struct vmap_area
*va
,
517 struct rb_root
*root
, struct rb_node
*from
,
518 struct rb_node
**parent
)
520 struct vmap_area
*tmp_va
;
521 struct rb_node
**link
;
524 link
= &root
->rb_node
;
525 if (unlikely(!*link
)) {
534 * Go to the bottom of the tree. When we hit the last point
535 * we end up with parent rb_node and correct direction, i name
536 * it link, where the new va->rb_node will be attached to.
539 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
542 * During the traversal we also do some sanity check.
543 * Trigger the BUG() if there are sides(left/right)
546 if (va
->va_start
< tmp_va
->va_end
&&
547 va
->va_end
<= tmp_va
->va_start
)
548 link
= &(*link
)->rb_left
;
549 else if (va
->va_end
> tmp_va
->va_start
&&
550 va
->va_start
>= tmp_va
->va_end
)
551 link
= &(*link
)->rb_right
;
556 *parent
= &tmp_va
->rb_node
;
560 static __always_inline
struct list_head
*
561 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
563 struct list_head
*list
;
565 if (unlikely(!parent
))
567 * The red-black tree where we try to find VA neighbors
568 * before merging or inserting is empty, i.e. it means
569 * there is no free vmap space. Normally it does not
570 * happen but we handle this case anyway.
574 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
575 return (&parent
->rb_right
== link
? list
->next
: list
);
578 static __always_inline
void
579 link_va(struct vmap_area
*va
, struct rb_root
*root
,
580 struct rb_node
*parent
, struct rb_node
**link
, struct list_head
*head
)
583 * VA is still not in the list, but we can
584 * identify its future previous list_head node.
586 if (likely(parent
)) {
587 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
588 if (&parent
->rb_right
!= link
)
592 /* Insert to the rb-tree */
593 rb_link_node(&va
->rb_node
, parent
, link
);
594 if (root
== &free_vmap_area_root
) {
596 * Some explanation here. Just perform simple insertion
597 * to the tree. We do not set va->subtree_max_size to
598 * its current size before calling rb_insert_augmented().
599 * It is because of we populate the tree from the bottom
600 * to parent levels when the node _is_ in the tree.
602 * Therefore we set subtree_max_size to zero after insertion,
603 * to let __augment_tree_propagate_from() puts everything to
604 * the correct order later on.
606 rb_insert_augmented(&va
->rb_node
,
607 root
, &free_vmap_area_rb_augment_cb
);
608 va
->subtree_max_size
= 0;
610 rb_insert_color(&va
->rb_node
, root
);
613 /* Address-sort this list */
614 list_add(&va
->list
, head
);
617 static __always_inline
void
618 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
620 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
623 if (root
== &free_vmap_area_root
)
624 rb_erase_augmented(&va
->rb_node
,
625 root
, &free_vmap_area_rb_augment_cb
);
627 rb_erase(&va
->rb_node
, root
);
630 RB_CLEAR_NODE(&va
->rb_node
);
633 #if DEBUG_AUGMENT_PROPAGATE_CHECK
635 augment_tree_propagate_check(struct rb_node
*n
)
637 struct vmap_area
*va
;
638 struct rb_node
*node
;
645 va
= rb_entry(n
, struct vmap_area
, rb_node
);
646 size
= va
->subtree_max_size
;
650 va
= rb_entry(node
, struct vmap_area
, rb_node
);
652 if (get_subtree_max_size(node
->rb_left
) == size
) {
653 node
= node
->rb_left
;
655 if (va_size(va
) == size
) {
660 node
= node
->rb_right
;
665 va
= rb_entry(n
, struct vmap_area
, rb_node
);
666 pr_emerg("tree is corrupted: %lu, %lu\n",
667 va_size(va
), va
->subtree_max_size
);
670 augment_tree_propagate_check(n
->rb_left
);
671 augment_tree_propagate_check(n
->rb_right
);
676 * This function populates subtree_max_size from bottom to upper
677 * levels starting from VA point. The propagation must be done
678 * when VA size is modified by changing its va_start/va_end. Or
679 * in case of newly inserting of VA to the tree.
681 * It means that __augment_tree_propagate_from() must be called:
682 * - After VA has been inserted to the tree(free path);
683 * - After VA has been shrunk(allocation path);
684 * - After VA has been increased(merging path).
686 * Please note that, it does not mean that upper parent nodes
687 * and their subtree_max_size are recalculated all the time up
696 * For example if we modify the node 4, shrinking it to 2, then
697 * no any modification is required. If we shrink the node 2 to 1
698 * its subtree_max_size is updated only, and set to 1. If we shrink
699 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
702 static __always_inline
void
703 augment_tree_propagate_from(struct vmap_area
*va
)
705 struct rb_node
*node
= &va
->rb_node
;
706 unsigned long new_va_sub_max_size
;
709 va
= rb_entry(node
, struct vmap_area
, rb_node
);
710 new_va_sub_max_size
= compute_subtree_max_size(va
);
713 * If the newly calculated maximum available size of the
714 * subtree is equal to the current one, then it means that
715 * the tree is propagated correctly. So we have to stop at
716 * this point to save cycles.
718 if (va
->subtree_max_size
== new_va_sub_max_size
)
721 va
->subtree_max_size
= new_va_sub_max_size
;
722 node
= rb_parent(&va
->rb_node
);
725 #if DEBUG_AUGMENT_PROPAGATE_CHECK
726 augment_tree_propagate_check(free_vmap_area_root
.rb_node
);
731 insert_vmap_area(struct vmap_area
*va
,
732 struct rb_root
*root
, struct list_head
*head
)
734 struct rb_node
**link
;
735 struct rb_node
*parent
;
737 link
= find_va_links(va
, root
, NULL
, &parent
);
738 link_va(va
, root
, parent
, link
, head
);
742 insert_vmap_area_augment(struct vmap_area
*va
,
743 struct rb_node
*from
, struct rb_root
*root
,
744 struct list_head
*head
)
746 struct rb_node
**link
;
747 struct rb_node
*parent
;
750 link
= find_va_links(va
, NULL
, from
, &parent
);
752 link
= find_va_links(va
, root
, NULL
, &parent
);
754 link_va(va
, root
, parent
, link
, head
);
755 augment_tree_propagate_from(va
);
759 * Merge de-allocated chunk of VA memory with previous
760 * and next free blocks. If coalesce is not done a new
761 * free area is inserted. If VA has been merged, it is
764 static __always_inline
struct vmap_area
*
765 merge_or_add_vmap_area(struct vmap_area
*va
,
766 struct rb_root
*root
, struct list_head
*head
)
768 struct vmap_area
*sibling
;
769 struct list_head
*next
;
770 struct rb_node
**link
;
771 struct rb_node
*parent
;
775 * Find a place in the tree where VA potentially will be
776 * inserted, unless it is merged with its sibling/siblings.
778 link
= find_va_links(va
, root
, NULL
, &parent
);
781 * Get next node of VA to check if merging can be done.
783 next
= get_va_next_sibling(parent
, link
);
784 if (unlikely(next
== NULL
))
790 * |<------VA------>|<-----Next----->|
795 sibling
= list_entry(next
, struct vmap_area
, list
);
796 if (sibling
->va_start
== va
->va_end
) {
797 sibling
->va_start
= va
->va_start
;
799 /* Check and update the tree if needed. */
800 augment_tree_propagate_from(sibling
);
802 /* Free vmap_area object. */
803 kmem_cache_free(vmap_area_cachep
, va
);
805 /* Point to the new merged area. */
814 * |<-----Prev----->|<------VA------>|
818 if (next
->prev
!= head
) {
819 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
820 if (sibling
->va_end
== va
->va_start
) {
821 sibling
->va_end
= va
->va_end
;
823 /* Check and update the tree if needed. */
824 augment_tree_propagate_from(sibling
);
829 /* Free vmap_area object. */
830 kmem_cache_free(vmap_area_cachep
, va
);
832 /* Point to the new merged area. */
840 link_va(va
, root
, parent
, link
, head
);
841 augment_tree_propagate_from(va
);
847 static __always_inline
bool
848 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
849 unsigned long align
, unsigned long vstart
)
851 unsigned long nva_start_addr
;
853 if (va
->va_start
> vstart
)
854 nva_start_addr
= ALIGN(va
->va_start
, align
);
856 nva_start_addr
= ALIGN(vstart
, align
);
858 /* Can be overflowed due to big size or alignment. */
859 if (nva_start_addr
+ size
< nva_start_addr
||
860 nva_start_addr
< vstart
)
863 return (nva_start_addr
+ size
<= va
->va_end
);
867 * Find the first free block(lowest start address) in the tree,
868 * that will accomplish the request corresponding to passing
871 static __always_inline
struct vmap_area
*
872 find_vmap_lowest_match(unsigned long size
,
873 unsigned long align
, unsigned long vstart
)
875 struct vmap_area
*va
;
876 struct rb_node
*node
;
877 unsigned long length
;
879 /* Start from the root. */
880 node
= free_vmap_area_root
.rb_node
;
882 /* Adjust the search size for alignment overhead. */
883 length
= size
+ align
- 1;
886 va
= rb_entry(node
, struct vmap_area
, rb_node
);
888 if (get_subtree_max_size(node
->rb_left
) >= length
&&
889 vstart
< va
->va_start
) {
890 node
= node
->rb_left
;
892 if (is_within_this_va(va
, size
, align
, vstart
))
896 * Does not make sense to go deeper towards the right
897 * sub-tree if it does not have a free block that is
898 * equal or bigger to the requested search length.
900 if (get_subtree_max_size(node
->rb_right
) >= length
) {
901 node
= node
->rb_right
;
906 * OK. We roll back and find the first right sub-tree,
907 * that will satisfy the search criteria. It can happen
908 * only once due to "vstart" restriction.
910 while ((node
= rb_parent(node
))) {
911 va
= rb_entry(node
, struct vmap_area
, rb_node
);
912 if (is_within_this_va(va
, size
, align
, vstart
))
915 if (get_subtree_max_size(node
->rb_right
) >= length
&&
916 vstart
<= va
->va_start
) {
917 node
= node
->rb_right
;
927 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
928 #include <linux/random.h>
930 static struct vmap_area
*
931 find_vmap_lowest_linear_match(unsigned long size
,
932 unsigned long align
, unsigned long vstart
)
934 struct vmap_area
*va
;
936 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
937 if (!is_within_this_va(va
, size
, align
, vstart
))
947 find_vmap_lowest_match_check(unsigned long size
)
949 struct vmap_area
*va_1
, *va_2
;
950 unsigned long vstart
;
953 get_random_bytes(&rnd
, sizeof(rnd
));
954 vstart
= VMALLOC_START
+ rnd
;
956 va_1
= find_vmap_lowest_match(size
, 1, vstart
);
957 va_2
= find_vmap_lowest_linear_match(size
, 1, vstart
);
960 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
967 FL_FIT_TYPE
= 1, /* full fit */
968 LE_FIT_TYPE
= 2, /* left edge fit */
969 RE_FIT_TYPE
= 3, /* right edge fit */
970 NE_FIT_TYPE
= 4 /* no edge fit */
973 static __always_inline
enum fit_type
974 classify_va_fit_type(struct vmap_area
*va
,
975 unsigned long nva_start_addr
, unsigned long size
)
979 /* Check if it is within VA. */
980 if (nva_start_addr
< va
->va_start
||
981 nva_start_addr
+ size
> va
->va_end
)
985 if (va
->va_start
== nva_start_addr
) {
986 if (va
->va_end
== nva_start_addr
+ size
)
990 } else if (va
->va_end
== nva_start_addr
+ size
) {
999 static __always_inline
int
1000 adjust_va_to_fit_type(struct vmap_area
*va
,
1001 unsigned long nva_start_addr
, unsigned long size
,
1004 struct vmap_area
*lva
= NULL
;
1006 if (type
== FL_FIT_TYPE
) {
1008 * No need to split VA, it fully fits.
1014 unlink_va(va
, &free_vmap_area_root
);
1015 kmem_cache_free(vmap_area_cachep
, va
);
1016 } else if (type
== LE_FIT_TYPE
) {
1018 * Split left edge of fit VA.
1024 va
->va_start
+= size
;
1025 } else if (type
== RE_FIT_TYPE
) {
1027 * Split right edge of fit VA.
1033 va
->va_end
= nva_start_addr
;
1034 } else if (type
== NE_FIT_TYPE
) {
1036 * Split no edge of fit VA.
1042 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1043 if (unlikely(!lva
)) {
1045 * For percpu allocator we do not do any pre-allocation
1046 * and leave it as it is. The reason is it most likely
1047 * never ends up with NE_FIT_TYPE splitting. In case of
1048 * percpu allocations offsets and sizes are aligned to
1049 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1050 * are its main fitting cases.
1052 * There are a few exceptions though, as an example it is
1053 * a first allocation (early boot up) when we have "one"
1054 * big free space that has to be split.
1056 * Also we can hit this path in case of regular "vmap"
1057 * allocations, if "this" current CPU was not preloaded.
1058 * See the comment in alloc_vmap_area() why. If so, then
1059 * GFP_NOWAIT is used instead to get an extra object for
1060 * split purpose. That is rare and most time does not
1063 * What happens if an allocation gets failed. Basically,
1064 * an "overflow" path is triggered to purge lazily freed
1065 * areas to free some memory, then, the "retry" path is
1066 * triggered to repeat one more time. See more details
1067 * in alloc_vmap_area() function.
1069 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1075 * Build the remainder.
1077 lva
->va_start
= va
->va_start
;
1078 lva
->va_end
= nva_start_addr
;
1081 * Shrink this VA to remaining size.
1083 va
->va_start
= nva_start_addr
+ size
;
1088 if (type
!= FL_FIT_TYPE
) {
1089 augment_tree_propagate_from(va
);
1091 if (lva
) /* type == NE_FIT_TYPE */
1092 insert_vmap_area_augment(lva
, &va
->rb_node
,
1093 &free_vmap_area_root
, &free_vmap_area_list
);
1100 * Returns a start address of the newly allocated area, if success.
1101 * Otherwise a vend is returned that indicates failure.
1103 static __always_inline
unsigned long
1104 __alloc_vmap_area(unsigned long size
, unsigned long align
,
1105 unsigned long vstart
, unsigned long vend
)
1107 unsigned long nva_start_addr
;
1108 struct vmap_area
*va
;
1112 va
= find_vmap_lowest_match(size
, align
, vstart
);
1116 if (va
->va_start
> vstart
)
1117 nva_start_addr
= ALIGN(va
->va_start
, align
);
1119 nva_start_addr
= ALIGN(vstart
, align
);
1121 /* Check the "vend" restriction. */
1122 if (nva_start_addr
+ size
> vend
)
1125 /* Classify what we have found. */
1126 type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1127 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
1130 /* Update the free vmap_area. */
1131 ret
= adjust_va_to_fit_type(va
, nva_start_addr
, size
, type
);
1135 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1136 find_vmap_lowest_match_check(size
);
1139 return nva_start_addr
;
1143 * Free a region of KVA allocated by alloc_vmap_area
1145 static void free_vmap_area(struct vmap_area
*va
)
1148 * Remove from the busy tree/list.
1150 spin_lock(&vmap_area_lock
);
1151 unlink_va(va
, &vmap_area_root
);
1152 spin_unlock(&vmap_area_lock
);
1155 * Insert/Merge it back to the free tree/list.
1157 spin_lock(&free_vmap_area_lock
);
1158 merge_or_add_vmap_area(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1159 spin_unlock(&free_vmap_area_lock
);
1163 * Allocate a region of KVA of the specified size and alignment, within the
1166 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1167 unsigned long align
,
1168 unsigned long vstart
, unsigned long vend
,
1169 int node
, gfp_t gfp_mask
)
1171 struct vmap_area
*va
, *pva
;
1177 BUG_ON(offset_in_page(size
));
1178 BUG_ON(!is_power_of_2(align
));
1180 if (unlikely(!vmap_initialized
))
1181 return ERR_PTR(-EBUSY
);
1184 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1186 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1188 return ERR_PTR(-ENOMEM
);
1191 * Only scan the relevant parts containing pointers to other objects
1192 * to avoid false negatives.
1194 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1198 * Preload this CPU with one extra vmap_area object. It is used
1199 * when fit type of free area is NE_FIT_TYPE. Please note, it
1200 * does not guarantee that an allocation occurs on a CPU that
1201 * is preloaded, instead we minimize the case when it is not.
1202 * It can happen because of cpu migration, because there is a
1203 * race until the below spinlock is taken.
1205 * The preload is done in non-atomic context, thus it allows us
1206 * to use more permissive allocation masks to be more stable under
1207 * low memory condition and high memory pressure. In rare case,
1208 * if not preloaded, GFP_NOWAIT is used.
1210 * Set "pva" to NULL here, because of "retry" path.
1214 if (!this_cpu_read(ne_fit_preload_node
))
1216 * Even if it fails we do not really care about that.
1217 * Just proceed as it is. If needed "overflow" path
1218 * will refill the cache we allocate from.
1220 pva
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1222 spin_lock(&free_vmap_area_lock
);
1224 if (pva
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, pva
))
1225 kmem_cache_free(vmap_area_cachep
, pva
);
1228 * If an allocation fails, the "vend" address is
1229 * returned. Therefore trigger the overflow path.
1231 addr
= __alloc_vmap_area(size
, align
, vstart
, vend
);
1232 spin_unlock(&free_vmap_area_lock
);
1234 if (unlikely(addr
== vend
))
1237 va
->va_start
= addr
;
1238 va
->va_end
= addr
+ size
;
1242 spin_lock(&vmap_area_lock
);
1243 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1244 spin_unlock(&vmap_area_lock
);
1246 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1247 BUG_ON(va
->va_start
< vstart
);
1248 BUG_ON(va
->va_end
> vend
);
1250 ret
= kasan_populate_vmalloc(addr
, size
);
1253 return ERR_PTR(ret
);
1260 purge_vmap_area_lazy();
1265 if (gfpflags_allow_blocking(gfp_mask
)) {
1266 unsigned long freed
= 0;
1267 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1274 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1275 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1278 kmem_cache_free(vmap_area_cachep
, va
);
1279 return ERR_PTR(-EBUSY
);
1282 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1284 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1286 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1288 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1290 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1292 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1295 * lazy_max_pages is the maximum amount of virtual address space we gather up
1296 * before attempting to purge with a TLB flush.
1298 * There is a tradeoff here: a larger number will cover more kernel page tables
1299 * and take slightly longer to purge, but it will linearly reduce the number of
1300 * global TLB flushes that must be performed. It would seem natural to scale
1301 * this number up linearly with the number of CPUs (because vmapping activity
1302 * could also scale linearly with the number of CPUs), however it is likely
1303 * that in practice, workloads might be constrained in other ways that mean
1304 * vmap activity will not scale linearly with CPUs. Also, I want to be
1305 * conservative and not introduce a big latency on huge systems, so go with
1306 * a less aggressive log scale. It will still be an improvement over the old
1307 * code, and it will be simple to change the scale factor if we find that it
1308 * becomes a problem on bigger systems.
1310 static unsigned long lazy_max_pages(void)
1314 log
= fls(num_online_cpus());
1316 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1319 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1322 * Serialize vmap purging. There is no actual criticial section protected
1323 * by this look, but we want to avoid concurrent calls for performance
1324 * reasons and to make the pcpu_get_vm_areas more deterministic.
1326 static DEFINE_MUTEX(vmap_purge_lock
);
1328 /* for per-CPU blocks */
1329 static void purge_fragmented_blocks_allcpus(void);
1332 * called before a call to iounmap() if the caller wants vm_area_struct's
1333 * immediately freed.
1335 void set_iounmap_nonlazy(void)
1337 atomic_long_set(&vmap_lazy_nr
, lazy_max_pages()+1);
1341 * Purges all lazily-freed vmap areas.
1343 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1345 unsigned long resched_threshold
;
1346 struct llist_node
*valist
;
1347 struct vmap_area
*va
;
1348 struct vmap_area
*n_va
;
1350 lockdep_assert_held(&vmap_purge_lock
);
1352 valist
= llist_del_all(&vmap_purge_list
);
1353 if (unlikely(valist
== NULL
))
1357 * TODO: to calculate a flush range without looping.
1358 * The list can be up to lazy_max_pages() elements.
1360 llist_for_each_entry(va
, valist
, purge_list
) {
1361 if (va
->va_start
< start
)
1362 start
= va
->va_start
;
1363 if (va
->va_end
> end
)
1367 flush_tlb_kernel_range(start
, end
);
1368 resched_threshold
= lazy_max_pages() << 1;
1370 spin_lock(&free_vmap_area_lock
);
1371 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
1372 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1373 unsigned long orig_start
= va
->va_start
;
1374 unsigned long orig_end
= va
->va_end
;
1377 * Finally insert or merge lazily-freed area. It is
1378 * detached and there is no need to "unlink" it from
1381 va
= merge_or_add_vmap_area(va
, &free_vmap_area_root
,
1382 &free_vmap_area_list
);
1384 if (is_vmalloc_or_module_addr((void *)orig_start
))
1385 kasan_release_vmalloc(orig_start
, orig_end
,
1386 va
->va_start
, va
->va_end
);
1388 atomic_long_sub(nr
, &vmap_lazy_nr
);
1390 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1391 cond_resched_lock(&free_vmap_area_lock
);
1393 spin_unlock(&free_vmap_area_lock
);
1398 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1399 * is already purging.
1401 static void try_purge_vmap_area_lazy(void)
1403 if (mutex_trylock(&vmap_purge_lock
)) {
1404 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1405 mutex_unlock(&vmap_purge_lock
);
1410 * Kick off a purge of the outstanding lazy areas.
1412 static void purge_vmap_area_lazy(void)
1414 mutex_lock(&vmap_purge_lock
);
1415 purge_fragmented_blocks_allcpus();
1416 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1417 mutex_unlock(&vmap_purge_lock
);
1421 * Free a vmap area, caller ensuring that the area has been unmapped
1422 * and flush_cache_vunmap had been called for the correct range
1425 static void free_vmap_area_noflush(struct vmap_area
*va
)
1427 unsigned long nr_lazy
;
1429 spin_lock(&vmap_area_lock
);
1430 unlink_va(va
, &vmap_area_root
);
1431 spin_unlock(&vmap_area_lock
);
1433 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1434 PAGE_SHIFT
, &vmap_lazy_nr
);
1436 /* After this point, we may free va at any time */
1437 llist_add(&va
->purge_list
, &vmap_purge_list
);
1439 if (unlikely(nr_lazy
> lazy_max_pages()))
1440 try_purge_vmap_area_lazy();
1444 * Free and unmap a vmap area
1446 static void free_unmap_vmap_area(struct vmap_area
*va
)
1448 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1449 unmap_kernel_range_noflush(va
->va_start
, va
->va_end
- va
->va_start
);
1450 if (debug_pagealloc_enabled_static())
1451 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1453 free_vmap_area_noflush(va
);
1456 static struct vmap_area
*find_vmap_area(unsigned long addr
)
1458 struct vmap_area
*va
;
1460 spin_lock(&vmap_area_lock
);
1461 va
= __find_vmap_area(addr
);
1462 spin_unlock(&vmap_area_lock
);
1467 /*** Per cpu kva allocator ***/
1470 * vmap space is limited especially on 32 bit architectures. Ensure there is
1471 * room for at least 16 percpu vmap blocks per CPU.
1474 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1475 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1476 * instead (we just need a rough idea)
1478 #if BITS_PER_LONG == 32
1479 #define VMALLOC_SPACE (128UL*1024*1024)
1481 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1484 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1485 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1486 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1487 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1488 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1489 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1490 #define VMAP_BBMAP_BITS \
1491 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1492 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1493 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1495 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1497 struct vmap_block_queue
{
1499 struct list_head free
;
1504 struct vmap_area
*va
;
1505 unsigned long free
, dirty
;
1506 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1507 struct list_head free_list
;
1508 struct rcu_head rcu_head
;
1509 struct list_head purge
;
1512 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1513 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1516 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1517 * in the free path. Could get rid of this if we change the API to return a
1518 * "cookie" from alloc, to be passed to free. But no big deal yet.
1520 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
1521 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
1524 * We should probably have a fallback mechanism to allocate virtual memory
1525 * out of partially filled vmap blocks. However vmap block sizing should be
1526 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1530 static unsigned long addr_to_vb_idx(unsigned long addr
)
1532 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1533 addr
/= VMAP_BLOCK_SIZE
;
1537 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1541 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1542 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1543 return (void *)addr
;
1547 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1548 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1549 * @order: how many 2^order pages should be occupied in newly allocated block
1550 * @gfp_mask: flags for the page level allocator
1552 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1554 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1556 struct vmap_block_queue
*vbq
;
1557 struct vmap_block
*vb
;
1558 struct vmap_area
*va
;
1559 unsigned long vb_idx
;
1563 node
= numa_node_id();
1565 vb
= kmalloc_node(sizeof(struct vmap_block
),
1566 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1568 return ERR_PTR(-ENOMEM
);
1570 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1571 VMALLOC_START
, VMALLOC_END
,
1575 return ERR_CAST(va
);
1578 err
= radix_tree_preload(gfp_mask
);
1579 if (unlikely(err
)) {
1582 return ERR_PTR(err
);
1585 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1586 spin_lock_init(&vb
->lock
);
1588 /* At least something should be left free */
1589 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1590 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
1592 vb
->dirty_min
= VMAP_BBMAP_BITS
;
1594 INIT_LIST_HEAD(&vb
->free_list
);
1596 vb_idx
= addr_to_vb_idx(va
->va_start
);
1597 spin_lock(&vmap_block_tree_lock
);
1598 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
1599 spin_unlock(&vmap_block_tree_lock
);
1601 radix_tree_preload_end();
1603 vbq
= &get_cpu_var(vmap_block_queue
);
1604 spin_lock(&vbq
->lock
);
1605 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
1606 spin_unlock(&vbq
->lock
);
1607 put_cpu_var(vmap_block_queue
);
1612 static void free_vmap_block(struct vmap_block
*vb
)
1614 struct vmap_block
*tmp
;
1615 unsigned long vb_idx
;
1617 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
1618 spin_lock(&vmap_block_tree_lock
);
1619 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
1620 spin_unlock(&vmap_block_tree_lock
);
1623 free_vmap_area_noflush(vb
->va
);
1624 kfree_rcu(vb
, rcu_head
);
1627 static void purge_fragmented_blocks(int cpu
)
1630 struct vmap_block
*vb
;
1631 struct vmap_block
*n_vb
;
1632 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1635 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1637 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
1640 spin_lock(&vb
->lock
);
1641 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
1642 vb
->free
= 0; /* prevent further allocs after releasing lock */
1643 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
1645 vb
->dirty_max
= VMAP_BBMAP_BITS
;
1646 spin_lock(&vbq
->lock
);
1647 list_del_rcu(&vb
->free_list
);
1648 spin_unlock(&vbq
->lock
);
1649 spin_unlock(&vb
->lock
);
1650 list_add_tail(&vb
->purge
, &purge
);
1652 spin_unlock(&vb
->lock
);
1656 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
1657 list_del(&vb
->purge
);
1658 free_vmap_block(vb
);
1662 static void purge_fragmented_blocks_allcpus(void)
1666 for_each_possible_cpu(cpu
)
1667 purge_fragmented_blocks(cpu
);
1670 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
1672 struct vmap_block_queue
*vbq
;
1673 struct vmap_block
*vb
;
1677 BUG_ON(offset_in_page(size
));
1678 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1679 if (WARN_ON(size
== 0)) {
1681 * Allocating 0 bytes isn't what caller wants since
1682 * get_order(0) returns funny result. Just warn and terminate
1687 order
= get_order(size
);
1690 vbq
= &get_cpu_var(vmap_block_queue
);
1691 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1692 unsigned long pages_off
;
1694 spin_lock(&vb
->lock
);
1695 if (vb
->free
< (1UL << order
)) {
1696 spin_unlock(&vb
->lock
);
1700 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1701 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1702 vb
->free
-= 1UL << order
;
1703 if (vb
->free
== 0) {
1704 spin_lock(&vbq
->lock
);
1705 list_del_rcu(&vb
->free_list
);
1706 spin_unlock(&vbq
->lock
);
1709 spin_unlock(&vb
->lock
);
1713 put_cpu_var(vmap_block_queue
);
1716 /* Allocate new block if nothing was found */
1718 vaddr
= new_vmap_block(order
, gfp_mask
);
1723 static void vb_free(unsigned long addr
, unsigned long size
)
1725 unsigned long offset
;
1726 unsigned long vb_idx
;
1728 struct vmap_block
*vb
;
1730 BUG_ON(offset_in_page(size
));
1731 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1733 flush_cache_vunmap(addr
, addr
+ size
);
1735 order
= get_order(size
);
1737 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
1739 vb_idx
= addr_to_vb_idx(addr
);
1741 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1745 unmap_kernel_range_noflush(addr
, size
);
1747 if (debug_pagealloc_enabled_static())
1748 flush_tlb_kernel_range(addr
, addr
+ size
);
1750 spin_lock(&vb
->lock
);
1752 /* Expand dirty range */
1753 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1754 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1756 vb
->dirty
+= 1UL << order
;
1757 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1759 spin_unlock(&vb
->lock
);
1760 free_vmap_block(vb
);
1762 spin_unlock(&vb
->lock
);
1765 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
1769 if (unlikely(!vmap_initialized
))
1774 for_each_possible_cpu(cpu
) {
1775 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1776 struct vmap_block
*vb
;
1779 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1780 spin_lock(&vb
->lock
);
1782 unsigned long va_start
= vb
->va
->va_start
;
1785 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1786 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1788 start
= min(s
, start
);
1793 spin_unlock(&vb
->lock
);
1798 mutex_lock(&vmap_purge_lock
);
1799 purge_fragmented_blocks_allcpus();
1800 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1801 flush_tlb_kernel_range(start
, end
);
1802 mutex_unlock(&vmap_purge_lock
);
1806 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1808 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1809 * to amortize TLB flushing overheads. What this means is that any page you
1810 * have now, may, in a former life, have been mapped into kernel virtual
1811 * address by the vmap layer and so there might be some CPUs with TLB entries
1812 * still referencing that page (additional to the regular 1:1 kernel mapping).
1814 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1815 * be sure that none of the pages we have control over will have any aliases
1816 * from the vmap layer.
1818 void vm_unmap_aliases(void)
1820 unsigned long start
= ULONG_MAX
, end
= 0;
1823 _vm_unmap_aliases(start
, end
, flush
);
1825 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1828 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1829 * @mem: the pointer returned by vm_map_ram
1830 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1832 void vm_unmap_ram(const void *mem
, unsigned int count
)
1834 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1835 unsigned long addr
= (unsigned long)mem
;
1836 struct vmap_area
*va
;
1840 BUG_ON(addr
< VMALLOC_START
);
1841 BUG_ON(addr
> VMALLOC_END
);
1842 BUG_ON(!PAGE_ALIGNED(addr
));
1844 kasan_poison_vmalloc(mem
, size
);
1846 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1847 debug_check_no_locks_freed(mem
, size
);
1848 vb_free(addr
, size
);
1852 va
= find_vmap_area(addr
);
1854 debug_check_no_locks_freed((void *)va
->va_start
,
1855 (va
->va_end
- va
->va_start
));
1856 free_unmap_vmap_area(va
);
1858 EXPORT_SYMBOL(vm_unmap_ram
);
1861 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1862 * @pages: an array of pointers to the pages to be mapped
1863 * @count: number of pages
1864 * @node: prefer to allocate data structures on this node
1865 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1867 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1868 * faster than vmap so it's good. But if you mix long-life and short-life
1869 * objects with vm_map_ram(), it could consume lots of address space through
1870 * fragmentation (especially on a 32bit machine). You could see failures in
1871 * the end. Please use this function for short-lived objects.
1873 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1875 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
1877 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1881 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1882 mem
= vb_alloc(size
, GFP_KERNEL
);
1885 addr
= (unsigned long)mem
;
1887 struct vmap_area
*va
;
1888 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1889 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1893 addr
= va
->va_start
;
1897 kasan_unpoison_vmalloc(mem
, size
);
1899 if (map_kernel_range(addr
, size
, PAGE_KERNEL
, pages
) < 0) {
1900 vm_unmap_ram(mem
, count
);
1905 EXPORT_SYMBOL(vm_map_ram
);
1907 static struct vm_struct
*vmlist __initdata
;
1910 * vm_area_add_early - add vmap area early during boot
1911 * @vm: vm_struct to add
1913 * This function is used to add fixed kernel vm area to vmlist before
1914 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1915 * should contain proper values and the other fields should be zero.
1917 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1919 void __init
vm_area_add_early(struct vm_struct
*vm
)
1921 struct vm_struct
*tmp
, **p
;
1923 BUG_ON(vmap_initialized
);
1924 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1925 if (tmp
->addr
>= vm
->addr
) {
1926 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1929 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1936 * vm_area_register_early - register vmap area early during boot
1937 * @vm: vm_struct to register
1938 * @align: requested alignment
1940 * This function is used to register kernel vm area before
1941 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1942 * proper values on entry and other fields should be zero. On return,
1943 * vm->addr contains the allocated address.
1945 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1947 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1949 static size_t vm_init_off __initdata
;
1952 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1953 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1955 vm
->addr
= (void *)addr
;
1957 vm_area_add_early(vm
);
1960 static void vmap_init_free_space(void)
1962 unsigned long vmap_start
= 1;
1963 const unsigned long vmap_end
= ULONG_MAX
;
1964 struct vmap_area
*busy
, *free
;
1968 * -|-----|.....|-----|-----|-----|.....|-
1970 * |<--------------------------------->|
1972 list_for_each_entry(busy
, &vmap_area_list
, list
) {
1973 if (busy
->va_start
- vmap_start
> 0) {
1974 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1975 if (!WARN_ON_ONCE(!free
)) {
1976 free
->va_start
= vmap_start
;
1977 free
->va_end
= busy
->va_start
;
1979 insert_vmap_area_augment(free
, NULL
,
1980 &free_vmap_area_root
,
1981 &free_vmap_area_list
);
1985 vmap_start
= busy
->va_end
;
1988 if (vmap_end
- vmap_start
> 0) {
1989 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
1990 if (!WARN_ON_ONCE(!free
)) {
1991 free
->va_start
= vmap_start
;
1992 free
->va_end
= vmap_end
;
1994 insert_vmap_area_augment(free
, NULL
,
1995 &free_vmap_area_root
,
1996 &free_vmap_area_list
);
2001 void __init
vmalloc_init(void)
2003 struct vmap_area
*va
;
2004 struct vm_struct
*tmp
;
2008 * Create the cache for vmap_area objects.
2010 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
2012 for_each_possible_cpu(i
) {
2013 struct vmap_block_queue
*vbq
;
2014 struct vfree_deferred
*p
;
2016 vbq
= &per_cpu(vmap_block_queue
, i
);
2017 spin_lock_init(&vbq
->lock
);
2018 INIT_LIST_HEAD(&vbq
->free
);
2019 p
= &per_cpu(vfree_deferred
, i
);
2020 init_llist_head(&p
->list
);
2021 INIT_WORK(&p
->wq
, free_work
);
2024 /* Import existing vmlist entries. */
2025 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
2026 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2027 if (WARN_ON_ONCE(!va
))
2030 va
->va_start
= (unsigned long)tmp
->addr
;
2031 va
->va_end
= va
->va_start
+ tmp
->size
;
2033 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
2037 * Now we can initialize a free vmap space.
2039 vmap_init_free_space();
2040 vmap_initialized
= true;
2044 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2045 * @addr: start of the VM area to unmap
2046 * @size: size of the VM area to unmap
2048 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2049 * the unmapping and tlb after.
2051 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
2053 unsigned long end
= addr
+ size
;
2055 flush_cache_vunmap(addr
, end
);
2056 unmap_kernel_range_noflush(addr
, size
);
2057 flush_tlb_kernel_range(addr
, end
);
2060 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2061 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2064 vm
->addr
= (void *)va
->va_start
;
2065 vm
->size
= va
->va_end
- va
->va_start
;
2066 vm
->caller
= caller
;
2070 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2071 unsigned long flags
, const void *caller
)
2073 spin_lock(&vmap_area_lock
);
2074 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2075 spin_unlock(&vmap_area_lock
);
2078 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2081 * Before removing VM_UNINITIALIZED,
2082 * we should make sure that vm has proper values.
2083 * Pair with smp_rmb() in show_numa_info().
2086 vm
->flags
&= ~VM_UNINITIALIZED
;
2089 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2090 unsigned long align
, unsigned long flags
, unsigned long start
,
2091 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
2093 struct vmap_area
*va
;
2094 struct vm_struct
*area
;
2095 unsigned long requested_size
= size
;
2097 BUG_ON(in_interrupt());
2098 size
= PAGE_ALIGN(size
);
2099 if (unlikely(!size
))
2102 if (flags
& VM_IOREMAP
)
2103 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2104 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2106 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2107 if (unlikely(!area
))
2110 if (!(flags
& VM_NO_GUARD
))
2113 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
2119 kasan_unpoison_vmalloc((void *)va
->va_start
, requested_size
);
2121 setup_vmalloc_vm(area
, va
, flags
, caller
);
2126 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2127 unsigned long start
, unsigned long end
,
2130 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
2131 GFP_KERNEL
, caller
);
2135 * get_vm_area - reserve a contiguous kernel virtual area
2136 * @size: size of the area
2137 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2139 * Search an area of @size in the kernel virtual mapping area,
2140 * and reserved it for out purposes. Returns the area descriptor
2141 * on success or %NULL on failure.
2143 * Return: the area descriptor on success or %NULL on failure.
2145 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2147 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2148 NUMA_NO_NODE
, GFP_KERNEL
,
2149 __builtin_return_address(0));
2152 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2155 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
2156 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2160 * find_vm_area - find a continuous kernel virtual area
2161 * @addr: base address
2163 * Search for the kernel VM area starting at @addr, and return it.
2164 * It is up to the caller to do all required locking to keep the returned
2167 * Return: pointer to the found area or %NULL on faulure
2169 struct vm_struct
*find_vm_area(const void *addr
)
2171 struct vmap_area
*va
;
2173 va
= find_vmap_area((unsigned long)addr
);
2181 * remove_vm_area - find and remove a continuous kernel virtual area
2182 * @addr: base address
2184 * Search for the kernel VM area starting at @addr, and remove it.
2185 * This function returns the found VM area, but using it is NOT safe
2186 * on SMP machines, except for its size or flags.
2188 * Return: pointer to the found area or %NULL on faulure
2190 struct vm_struct
*remove_vm_area(const void *addr
)
2192 struct vmap_area
*va
;
2196 spin_lock(&vmap_area_lock
);
2197 va
= __find_vmap_area((unsigned long)addr
);
2199 struct vm_struct
*vm
= va
->vm
;
2202 spin_unlock(&vmap_area_lock
);
2204 kasan_free_shadow(vm
);
2205 free_unmap_vmap_area(va
);
2210 spin_unlock(&vmap_area_lock
);
2214 static inline void set_area_direct_map(const struct vm_struct
*area
,
2215 int (*set_direct_map
)(struct page
*page
))
2219 for (i
= 0; i
< area
->nr_pages
; i
++)
2220 if (page_address(area
->pages
[i
]))
2221 set_direct_map(area
->pages
[i
]);
2224 /* Handle removing and resetting vm mappings related to the vm_struct. */
2225 static void vm_remove_mappings(struct vm_struct
*area
, int deallocate_pages
)
2227 unsigned long start
= ULONG_MAX
, end
= 0;
2228 int flush_reset
= area
->flags
& VM_FLUSH_RESET_PERMS
;
2232 remove_vm_area(area
->addr
);
2234 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2239 * If not deallocating pages, just do the flush of the VM area and
2242 if (!deallocate_pages
) {
2248 * If execution gets here, flush the vm mapping and reset the direct
2249 * map. Find the start and end range of the direct mappings to make sure
2250 * the vm_unmap_aliases() flush includes the direct map.
2252 for (i
= 0; i
< area
->nr_pages
; i
++) {
2253 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2255 start
= min(addr
, start
);
2256 end
= max(addr
+ PAGE_SIZE
, end
);
2262 * Set direct map to something invalid so that it won't be cached if
2263 * there are any accesses after the TLB flush, then flush the TLB and
2264 * reset the direct map permissions to the default.
2266 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2267 _vm_unmap_aliases(start
, end
, flush_dmap
);
2268 set_area_direct_map(area
, set_direct_map_default_noflush
);
2271 static void __vunmap(const void *addr
, int deallocate_pages
)
2273 struct vm_struct
*area
;
2278 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2282 area
= find_vm_area(addr
);
2283 if (unlikely(!area
)) {
2284 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2289 debug_check_no_locks_freed(area
->addr
, get_vm_area_size(area
));
2290 debug_check_no_obj_freed(area
->addr
, get_vm_area_size(area
));
2292 kasan_poison_vmalloc(area
->addr
, area
->size
);
2294 vm_remove_mappings(area
, deallocate_pages
);
2296 if (deallocate_pages
) {
2299 for (i
= 0; i
< area
->nr_pages
; i
++) {
2300 struct page
*page
= area
->pages
[i
];
2303 __free_pages(page
, 0);
2305 atomic_long_sub(area
->nr_pages
, &nr_vmalloc_pages
);
2307 kvfree(area
->pages
);
2314 static inline void __vfree_deferred(const void *addr
)
2317 * Use raw_cpu_ptr() because this can be called from preemptible
2318 * context. Preemption is absolutely fine here, because the llist_add()
2319 * implementation is lockless, so it works even if we are adding to
2320 * another cpu's list. schedule_work() should be fine with this too.
2322 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2324 if (llist_add((struct llist_node
*)addr
, &p
->list
))
2325 schedule_work(&p
->wq
);
2329 * vfree_atomic - release memory allocated by vmalloc()
2330 * @addr: memory base address
2332 * This one is just like vfree() but can be called in any atomic context
2335 void vfree_atomic(const void *addr
)
2339 kmemleak_free(addr
);
2343 __vfree_deferred(addr
);
2346 static void __vfree(const void *addr
)
2348 if (unlikely(in_interrupt()))
2349 __vfree_deferred(addr
);
2355 * vfree - release memory allocated by vmalloc()
2356 * @addr: memory base address
2358 * Free the virtually continuous memory area starting at @addr, as
2359 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2360 * NULL, no operation is performed.
2362 * Must not be called in NMI context (strictly speaking, only if we don't
2363 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2364 * conventions for vfree() arch-depenedent would be a really bad idea)
2366 * May sleep if called *not* from interrupt context.
2368 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2370 void vfree(const void *addr
)
2374 kmemleak_free(addr
);
2376 might_sleep_if(!in_interrupt());
2383 EXPORT_SYMBOL(vfree
);
2386 * vunmap - release virtual mapping obtained by vmap()
2387 * @addr: memory base address
2389 * Free the virtually contiguous memory area starting at @addr,
2390 * which was created from the page array passed to vmap().
2392 * Must not be called in interrupt context.
2394 void vunmap(const void *addr
)
2396 BUG_ON(in_interrupt());
2401 EXPORT_SYMBOL(vunmap
);
2404 * vmap - map an array of pages into virtually contiguous space
2405 * @pages: array of page pointers
2406 * @count: number of pages to map
2407 * @flags: vm_area->flags
2408 * @prot: page protection for the mapping
2410 * Maps @count pages from @pages into contiguous kernel virtual
2413 * Return: the address of the area or %NULL on failure
2415 void *vmap(struct page
**pages
, unsigned int count
,
2416 unsigned long flags
, pgprot_t prot
)
2418 struct vm_struct
*area
;
2419 unsigned long size
; /* In bytes */
2423 if (count
> totalram_pages())
2426 size
= (unsigned long)count
<< PAGE_SHIFT
;
2427 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2431 if (map_kernel_range((unsigned long)area
->addr
, size
, pgprot_nx(prot
),
2439 EXPORT_SYMBOL(vmap
);
2441 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2442 pgprot_t prot
, int node
)
2444 struct page
**pages
;
2445 unsigned int nr_pages
, array_size
, i
;
2446 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2447 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
2448 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
2452 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
2453 array_size
= (nr_pages
* sizeof(struct page
*));
2455 /* Please note that the recursion is strictly bounded. */
2456 if (array_size
> PAGE_SIZE
) {
2457 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
2458 node
, area
->caller
);
2460 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
2464 remove_vm_area(area
->addr
);
2469 area
->pages
= pages
;
2470 area
->nr_pages
= nr_pages
;
2472 for (i
= 0; i
< area
->nr_pages
; i
++) {
2475 if (node
== NUMA_NO_NODE
)
2476 page
= alloc_page(alloc_mask
|highmem_mask
);
2478 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
2480 if (unlikely(!page
)) {
2481 /* Successfully allocated i pages, free them in __vunmap() */
2483 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2486 area
->pages
[i
] = page
;
2487 if (gfpflags_allow_blocking(gfp_mask
))
2490 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
2492 if (map_kernel_range((unsigned long)area
->addr
, get_vm_area_size(area
),
2499 warn_alloc(gfp_mask
, NULL
,
2500 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2501 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
2502 __vfree(area
->addr
);
2507 * __vmalloc_node_range - allocate virtually contiguous memory
2508 * @size: allocation size
2509 * @align: desired alignment
2510 * @start: vm area range start
2511 * @end: vm area range end
2512 * @gfp_mask: flags for the page level allocator
2513 * @prot: protection mask for the allocated pages
2514 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2515 * @node: node to use for allocation or NUMA_NO_NODE
2516 * @caller: caller's return address
2518 * Allocate enough pages to cover @size from the page level
2519 * allocator with @gfp_mask flags. Map them into contiguous
2520 * kernel virtual space, using a pagetable protection of @prot.
2522 * Return: the address of the area or %NULL on failure
2524 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
2525 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
2526 pgprot_t prot
, unsigned long vm_flags
, int node
,
2529 struct vm_struct
*area
;
2531 unsigned long real_size
= size
;
2533 size
= PAGE_ALIGN(size
);
2534 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages())
2537 area
= __get_vm_area_node(real_size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
2538 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
2542 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
2547 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2548 * flag. It means that vm_struct is not fully initialized.
2549 * Now, it is fully initialized, so remove this flag here.
2551 clear_vm_uninitialized_flag(area
);
2553 kmemleak_vmalloc(area
, size
, gfp_mask
);
2558 warn_alloc(gfp_mask
, NULL
,
2559 "vmalloc: allocation failure: %lu bytes", real_size
);
2564 * __vmalloc_node - allocate virtually contiguous memory
2565 * @size: allocation size
2566 * @align: desired alignment
2567 * @gfp_mask: flags for the page level allocator
2568 * @node: node to use for allocation or NUMA_NO_NODE
2569 * @caller: caller's return address
2571 * Allocate enough pages to cover @size from the page level allocator with
2572 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2574 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2575 * and __GFP_NOFAIL are not supported
2577 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2580 * Return: pointer to the allocated memory or %NULL on error
2582 void *__vmalloc_node(unsigned long size
, unsigned long align
,
2583 gfp_t gfp_mask
, int node
, const void *caller
)
2585 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
2586 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
2589 * This is only for performance analysis of vmalloc and stress purpose.
2590 * It is required by vmalloc test module, therefore do not use it other
2593 #ifdef CONFIG_TEST_VMALLOC_MODULE
2594 EXPORT_SYMBOL_GPL(__vmalloc_node
);
2597 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
2599 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
2600 __builtin_return_address(0));
2602 EXPORT_SYMBOL(__vmalloc
);
2605 * vmalloc - allocate virtually contiguous memory
2606 * @size: allocation size
2608 * Allocate enough pages to cover @size from the page level
2609 * allocator and map them into contiguous kernel virtual space.
2611 * For tight control over page level allocator and protection flags
2612 * use __vmalloc() instead.
2614 * Return: pointer to the allocated memory or %NULL on error
2616 void *vmalloc(unsigned long size
)
2618 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
2619 __builtin_return_address(0));
2621 EXPORT_SYMBOL(vmalloc
);
2624 * vzalloc - allocate virtually contiguous memory with zero fill
2625 * @size: allocation size
2627 * Allocate enough pages to cover @size from the page level
2628 * allocator and map them into contiguous kernel virtual space.
2629 * The memory allocated is set to zero.
2631 * For tight control over page level allocator and protection flags
2632 * use __vmalloc() instead.
2634 * Return: pointer to the allocated memory or %NULL on error
2636 void *vzalloc(unsigned long size
)
2638 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
2639 __builtin_return_address(0));
2641 EXPORT_SYMBOL(vzalloc
);
2644 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2645 * @size: allocation size
2647 * The resulting memory area is zeroed so it can be mapped to userspace
2648 * without leaking data.
2650 * Return: pointer to the allocated memory or %NULL on error
2652 void *vmalloc_user(unsigned long size
)
2654 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2655 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
2656 VM_USERMAP
, NUMA_NO_NODE
,
2657 __builtin_return_address(0));
2659 EXPORT_SYMBOL(vmalloc_user
);
2662 * vmalloc_node - allocate memory on a specific node
2663 * @size: allocation size
2666 * Allocate enough pages to cover @size from the page level
2667 * allocator and map them into contiguous kernel virtual space.
2669 * For tight control over page level allocator and protection flags
2670 * use __vmalloc() instead.
2672 * Return: pointer to the allocated memory or %NULL on error
2674 void *vmalloc_node(unsigned long size
, int node
)
2676 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
2677 __builtin_return_address(0));
2679 EXPORT_SYMBOL(vmalloc_node
);
2682 * vzalloc_node - allocate memory on a specific node with zero fill
2683 * @size: allocation size
2686 * Allocate enough pages to cover @size from the page level
2687 * allocator and map them into contiguous kernel virtual space.
2688 * The memory allocated is set to zero.
2690 * Return: pointer to the allocated memory or %NULL on error
2692 void *vzalloc_node(unsigned long size
, int node
)
2694 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
2695 __builtin_return_address(0));
2697 EXPORT_SYMBOL(vzalloc_node
);
2700 * vmalloc_exec - allocate virtually contiguous, executable memory
2701 * @size: allocation size
2703 * Kernel-internal function to allocate enough pages to cover @size
2704 * the page level allocator and map them into contiguous and
2705 * executable kernel virtual space.
2707 * For tight control over page level allocator and protection flags
2708 * use __vmalloc() instead.
2710 * Return: pointer to the allocated memory or %NULL on error
2712 void *vmalloc_exec(unsigned long size
)
2714 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
2715 GFP_KERNEL
, PAGE_KERNEL_EXEC
, VM_FLUSH_RESET_PERMS
,
2716 NUMA_NO_NODE
, __builtin_return_address(0));
2719 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2720 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2721 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2722 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2725 * 64b systems should always have either DMA or DMA32 zones. For others
2726 * GFP_DMA32 should do the right thing and use the normal zone.
2728 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2732 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2733 * @size: allocation size
2735 * Allocate enough 32bit PA addressable pages to cover @size from the
2736 * page level allocator and map them into contiguous kernel virtual space.
2738 * Return: pointer to the allocated memory or %NULL on error
2740 void *vmalloc_32(unsigned long size
)
2742 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
2743 __builtin_return_address(0));
2745 EXPORT_SYMBOL(vmalloc_32
);
2748 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2749 * @size: allocation size
2751 * The resulting memory area is 32bit addressable and zeroed so it can be
2752 * mapped to userspace without leaking data.
2754 * Return: pointer to the allocated memory or %NULL on error
2756 void *vmalloc_32_user(unsigned long size
)
2758 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
2759 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2760 VM_USERMAP
, NUMA_NO_NODE
,
2761 __builtin_return_address(0));
2763 EXPORT_SYMBOL(vmalloc_32_user
);
2766 * small helper routine , copy contents to buf from addr.
2767 * If the page is not present, fill zero.
2770 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2776 unsigned long offset
, length
;
2778 offset
= offset_in_page(addr
);
2779 length
= PAGE_SIZE
- offset
;
2782 p
= vmalloc_to_page(addr
);
2784 * To do safe access to this _mapped_ area, we need
2785 * lock. But adding lock here means that we need to add
2786 * overhead of vmalloc()/vfree() calles for this _debug_
2787 * interface, rarely used. Instead of that, we'll use
2788 * kmap() and get small overhead in this access function.
2792 * we can expect USER0 is not used (see vread/vwrite's
2793 * function description)
2795 void *map
= kmap_atomic(p
);
2796 memcpy(buf
, map
+ offset
, length
);
2799 memset(buf
, 0, length
);
2809 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2815 unsigned long offset
, length
;
2817 offset
= offset_in_page(addr
);
2818 length
= PAGE_SIZE
- offset
;
2821 p
= vmalloc_to_page(addr
);
2823 * To do safe access to this _mapped_ area, we need
2824 * lock. But adding lock here means that we need to add
2825 * overhead of vmalloc()/vfree() calles for this _debug_
2826 * interface, rarely used. Instead of that, we'll use
2827 * kmap() and get small overhead in this access function.
2831 * we can expect USER0 is not used (see vread/vwrite's
2832 * function description)
2834 void *map
= kmap_atomic(p
);
2835 memcpy(map
+ offset
, buf
, length
);
2847 * vread() - read vmalloc area in a safe way.
2848 * @buf: buffer for reading data
2849 * @addr: vm address.
2850 * @count: number of bytes to be read.
2852 * This function checks that addr is a valid vmalloc'ed area, and
2853 * copy data from that area to a given buffer. If the given memory range
2854 * of [addr...addr+count) includes some valid address, data is copied to
2855 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2856 * IOREMAP area is treated as memory hole and no copy is done.
2858 * If [addr...addr+count) doesn't includes any intersects with alive
2859 * vm_struct area, returns 0. @buf should be kernel's buffer.
2861 * Note: In usual ops, vread() is never necessary because the caller
2862 * should know vmalloc() area is valid and can use memcpy().
2863 * This is for routines which have to access vmalloc area without
2864 * any information, as /dev/kmem.
2866 * Return: number of bytes for which addr and buf should be increased
2867 * (same number as @count) or %0 if [addr...addr+count) doesn't
2868 * include any intersection with valid vmalloc area
2870 long vread(char *buf
, char *addr
, unsigned long count
)
2872 struct vmap_area
*va
;
2873 struct vm_struct
*vm
;
2874 char *vaddr
, *buf_start
= buf
;
2875 unsigned long buflen
= count
;
2878 /* Don't allow overflow */
2879 if ((unsigned long) addr
+ count
< count
)
2880 count
= -(unsigned long) addr
;
2882 spin_lock(&vmap_area_lock
);
2883 list_for_each_entry(va
, &vmap_area_list
, list
) {
2891 vaddr
= (char *) vm
->addr
;
2892 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2894 while (addr
< vaddr
) {
2902 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2905 if (!(vm
->flags
& VM_IOREMAP
))
2906 aligned_vread(buf
, addr
, n
);
2907 else /* IOREMAP area is treated as memory hole */
2914 spin_unlock(&vmap_area_lock
);
2916 if (buf
== buf_start
)
2918 /* zero-fill memory holes */
2919 if (buf
!= buf_start
+ buflen
)
2920 memset(buf
, 0, buflen
- (buf
- buf_start
));
2926 * vwrite() - write vmalloc area in a safe way.
2927 * @buf: buffer for source data
2928 * @addr: vm address.
2929 * @count: number of bytes to be read.
2931 * This function checks that addr is a valid vmalloc'ed area, and
2932 * copy data from a buffer to the given addr. If specified range of
2933 * [addr...addr+count) includes some valid address, data is copied from
2934 * proper area of @buf. If there are memory holes, no copy to hole.
2935 * IOREMAP area is treated as memory hole and no copy is done.
2937 * If [addr...addr+count) doesn't includes any intersects with alive
2938 * vm_struct area, returns 0. @buf should be kernel's buffer.
2940 * Note: In usual ops, vwrite() is never necessary because the caller
2941 * should know vmalloc() area is valid and can use memcpy().
2942 * This is for routines which have to access vmalloc area without
2943 * any information, as /dev/kmem.
2945 * Return: number of bytes for which addr and buf should be
2946 * increased (same number as @count) or %0 if [addr...addr+count)
2947 * doesn't include any intersection with valid vmalloc area
2949 long vwrite(char *buf
, char *addr
, unsigned long count
)
2951 struct vmap_area
*va
;
2952 struct vm_struct
*vm
;
2954 unsigned long n
, buflen
;
2957 /* Don't allow overflow */
2958 if ((unsigned long) addr
+ count
< count
)
2959 count
= -(unsigned long) addr
;
2962 spin_lock(&vmap_area_lock
);
2963 list_for_each_entry(va
, &vmap_area_list
, list
) {
2971 vaddr
= (char *) vm
->addr
;
2972 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2974 while (addr
< vaddr
) {
2981 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2984 if (!(vm
->flags
& VM_IOREMAP
)) {
2985 aligned_vwrite(buf
, addr
, n
);
2993 spin_unlock(&vmap_area_lock
);
3000 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3001 * @vma: vma to cover
3002 * @uaddr: target user address to start at
3003 * @kaddr: virtual address of vmalloc kernel memory
3004 * @pgoff: offset from @kaddr to start at
3005 * @size: size of map area
3007 * Returns: 0 for success, -Exxx on failure
3009 * This function checks that @kaddr is a valid vmalloc'ed area,
3010 * and that it is big enough to cover the range starting at
3011 * @uaddr in @vma. Will return failure if that criteria isn't
3014 * Similar to remap_pfn_range() (see mm/memory.c)
3016 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3017 void *kaddr
, unsigned long pgoff
,
3020 struct vm_struct
*area
;
3022 unsigned long end_index
;
3024 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
3027 size
= PAGE_ALIGN(size
);
3029 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3032 area
= find_vm_area(kaddr
);
3036 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3039 if (check_add_overflow(size
, off
, &end_index
) ||
3040 end_index
> get_vm_area_size(area
))
3045 struct page
*page
= vmalloc_to_page(kaddr
);
3048 ret
= vm_insert_page(vma
, uaddr
, page
);
3057 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
3061 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
3064 * remap_vmalloc_range - map vmalloc pages to userspace
3065 * @vma: vma to cover (map full range of vma)
3066 * @addr: vmalloc memory
3067 * @pgoff: number of pages into addr before first page to map
3069 * Returns: 0 for success, -Exxx on failure
3071 * This function checks that addr is a valid vmalloc'ed area, and
3072 * that it is big enough to cover the vma. Will return failure if
3073 * that criteria isn't met.
3075 * Similar to remap_pfn_range() (see mm/memory.c)
3077 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3078 unsigned long pgoff
)
3080 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3082 vma
->vm_end
- vma
->vm_start
);
3084 EXPORT_SYMBOL(remap_vmalloc_range
);
3086 static int f(pte_t
*pte
, unsigned long addr
, void *data
)
3098 * alloc_vm_area - allocate a range of kernel address space
3099 * @size: size of the area
3100 * @ptes: returns the PTEs for the address space
3102 * Returns: NULL on failure, vm_struct on success
3104 * This function reserves a range of kernel address space, and
3105 * allocates pagetables to map that range. No actual mappings
3108 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3109 * allocated for the VM area are returned.
3111 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
3113 struct vm_struct
*area
;
3115 area
= get_vm_area_caller(size
, VM_IOREMAP
,
3116 __builtin_return_address(0));
3121 * This ensures that page tables are constructed for this region
3122 * of kernel virtual address space and mapped into init_mm.
3124 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3125 size
, f
, ptes
? &ptes
: NULL
)) {
3132 EXPORT_SYMBOL_GPL(alloc_vm_area
);
3134 void free_vm_area(struct vm_struct
*area
)
3136 struct vm_struct
*ret
;
3137 ret
= remove_vm_area(area
->addr
);
3138 BUG_ON(ret
!= area
);
3141 EXPORT_SYMBOL_GPL(free_vm_area
);
3144 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3146 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3150 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3151 * @addr: target address
3153 * Returns: vmap_area if it is found. If there is no such area
3154 * the first highest(reverse order) vmap_area is returned
3155 * i.e. va->va_start < addr && va->va_end < addr or NULL
3156 * if there are no any areas before @addr.
3158 static struct vmap_area
*
3159 pvm_find_va_enclose_addr(unsigned long addr
)
3161 struct vmap_area
*va
, *tmp
;
3164 n
= free_vmap_area_root
.rb_node
;
3168 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3169 if (tmp
->va_start
<= addr
) {
3171 if (tmp
->va_end
>= addr
)
3184 * pvm_determine_end_from_reverse - find the highest aligned address
3185 * of free block below VMALLOC_END
3187 * in - the VA we start the search(reverse order);
3188 * out - the VA with the highest aligned end address.
3190 * Returns: determined end address within vmap_area
3192 static unsigned long
3193 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3195 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3199 list_for_each_entry_from_reverse((*va
),
3200 &free_vmap_area_list
, list
) {
3201 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3202 if ((*va
)->va_start
< addr
)
3211 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3212 * @offsets: array containing offset of each area
3213 * @sizes: array containing size of each area
3214 * @nr_vms: the number of areas to allocate
3215 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3217 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3218 * vm_structs on success, %NULL on failure
3220 * Percpu allocator wants to use congruent vm areas so that it can
3221 * maintain the offsets among percpu areas. This function allocates
3222 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3223 * be scattered pretty far, distance between two areas easily going up
3224 * to gigabytes. To avoid interacting with regular vmallocs, these
3225 * areas are allocated from top.
3227 * Despite its complicated look, this allocator is rather simple. It
3228 * does everything top-down and scans free blocks from the end looking
3229 * for matching base. While scanning, if any of the areas do not fit the
3230 * base address is pulled down to fit the area. Scanning is repeated till
3231 * all the areas fit and then all necessary data structures are inserted
3232 * and the result is returned.
3234 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3235 const size_t *sizes
, int nr_vms
,
3238 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3239 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3240 struct vmap_area
**vas
, *va
;
3241 struct vm_struct
**vms
;
3242 int area
, area2
, last_area
, term_area
;
3243 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3244 bool purged
= false;
3247 /* verify parameters and allocate data structures */
3248 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3249 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3250 start
= offsets
[area
];
3251 end
= start
+ sizes
[area
];
3253 /* is everything aligned properly? */
3254 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3255 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3257 /* detect the area with the highest address */
3258 if (start
> offsets
[last_area
])
3261 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3262 unsigned long start2
= offsets
[area2
];
3263 unsigned long end2
= start2
+ sizes
[area2
];
3265 BUG_ON(start2
< end
&& start
< end2
);
3268 last_end
= offsets
[last_area
] + sizes
[last_area
];
3270 if (vmalloc_end
- vmalloc_start
< last_end
) {
3275 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3276 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3280 for (area
= 0; area
< nr_vms
; area
++) {
3281 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3282 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3283 if (!vas
[area
] || !vms
[area
])
3287 spin_lock(&free_vmap_area_lock
);
3289 /* start scanning - we scan from the top, begin with the last area */
3290 area
= term_area
= last_area
;
3291 start
= offsets
[area
];
3292 end
= start
+ sizes
[area
];
3294 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3295 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3299 * base might have underflowed, add last_end before
3302 if (base
+ last_end
< vmalloc_start
+ last_end
)
3306 * Fitting base has not been found.
3312 * If required width exceeds current VA block, move
3313 * base downwards and then recheck.
3315 if (base
+ end
> va
->va_end
) {
3316 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3322 * If this VA does not fit, move base downwards and recheck.
3324 if (base
+ start
< va
->va_start
) {
3325 va
= node_to_va(rb_prev(&va
->rb_node
));
3326 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3332 * This area fits, move on to the previous one. If
3333 * the previous one is the terminal one, we're done.
3335 area
= (area
+ nr_vms
- 1) % nr_vms
;
3336 if (area
== term_area
)
3339 start
= offsets
[area
];
3340 end
= start
+ sizes
[area
];
3341 va
= pvm_find_va_enclose_addr(base
+ end
);
3344 /* we've found a fitting base, insert all va's */
3345 for (area
= 0; area
< nr_vms
; area
++) {
3348 start
= base
+ offsets
[area
];
3351 va
= pvm_find_va_enclose_addr(start
);
3352 if (WARN_ON_ONCE(va
== NULL
))
3353 /* It is a BUG(), but trigger recovery instead. */
3356 type
= classify_va_fit_type(va
, start
, size
);
3357 if (WARN_ON_ONCE(type
== NOTHING_FIT
))
3358 /* It is a BUG(), but trigger recovery instead. */
3361 ret
= adjust_va_to_fit_type(va
, start
, size
, type
);
3365 /* Allocated area. */
3367 va
->va_start
= start
;
3368 va
->va_end
= start
+ size
;
3371 spin_unlock(&free_vmap_area_lock
);
3373 /* populate the kasan shadow space */
3374 for (area
= 0; area
< nr_vms
; area
++) {
3375 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3376 goto err_free_shadow
;
3378 kasan_unpoison_vmalloc((void *)vas
[area
]->va_start
,
3382 /* insert all vm's */
3383 spin_lock(&vmap_area_lock
);
3384 for (area
= 0; area
< nr_vms
; area
++) {
3385 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3387 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3390 spin_unlock(&vmap_area_lock
);
3397 * Remove previously allocated areas. There is no
3398 * need in removing these areas from the busy tree,
3399 * because they are inserted only on the final step
3400 * and when pcpu_get_vm_areas() is success.
3403 orig_start
= vas
[area
]->va_start
;
3404 orig_end
= vas
[area
]->va_end
;
3405 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3406 &free_vmap_area_list
);
3407 kasan_release_vmalloc(orig_start
, orig_end
,
3408 va
->va_start
, va
->va_end
);
3413 spin_unlock(&free_vmap_area_lock
);
3415 purge_vmap_area_lazy();
3418 /* Before "retry", check if we recover. */
3419 for (area
= 0; area
< nr_vms
; area
++) {
3423 vas
[area
] = kmem_cache_zalloc(
3424 vmap_area_cachep
, GFP_KERNEL
);
3433 for (area
= 0; area
< nr_vms
; area
++) {
3435 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
3445 spin_lock(&free_vmap_area_lock
);
3447 * We release all the vmalloc shadows, even the ones for regions that
3448 * hadn't been successfully added. This relies on kasan_release_vmalloc
3449 * being able to tolerate this case.
3451 for (area
= 0; area
< nr_vms
; area
++) {
3452 orig_start
= vas
[area
]->va_start
;
3453 orig_end
= vas
[area
]->va_end
;
3454 va
= merge_or_add_vmap_area(vas
[area
], &free_vmap_area_root
,
3455 &free_vmap_area_list
);
3456 kasan_release_vmalloc(orig_start
, orig_end
,
3457 va
->va_start
, va
->va_end
);
3461 spin_unlock(&free_vmap_area_lock
);
3468 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3469 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3470 * @nr_vms: the number of allocated areas
3472 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3474 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
3478 for (i
= 0; i
< nr_vms
; i
++)
3479 free_vm_area(vms
[i
]);
3482 #endif /* CONFIG_SMP */
3484 #ifdef CONFIG_PROC_FS
3485 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
3486 __acquires(&vmap_purge_lock
)
3487 __acquires(&vmap_area_lock
)
3489 mutex_lock(&vmap_purge_lock
);
3490 spin_lock(&vmap_area_lock
);
3492 return seq_list_start(&vmap_area_list
, *pos
);
3495 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
3497 return seq_list_next(p
, &vmap_area_list
, pos
);
3500 static void s_stop(struct seq_file
*m
, void *p
)
3501 __releases(&vmap_purge_lock
)
3502 __releases(&vmap_area_lock
)
3504 mutex_unlock(&vmap_purge_lock
);
3505 spin_unlock(&vmap_area_lock
);
3508 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
3510 if (IS_ENABLED(CONFIG_NUMA
)) {
3511 unsigned int nr
, *counters
= m
->private;
3516 if (v
->flags
& VM_UNINITIALIZED
)
3518 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3521 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
3523 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
3524 counters
[page_to_nid(v
->pages
[nr
])]++;
3526 for_each_node_state(nr
, N_HIGH_MEMORY
)
3528 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
3532 static void show_purge_info(struct seq_file
*m
)
3534 struct llist_node
*head
;
3535 struct vmap_area
*va
;
3537 head
= READ_ONCE(vmap_purge_list
.first
);
3541 llist_for_each_entry(va
, head
, purge_list
) {
3542 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3543 (void *)va
->va_start
, (void *)va
->va_end
,
3544 va
->va_end
- va
->va_start
);
3548 static int s_show(struct seq_file
*m
, void *p
)
3550 struct vmap_area
*va
;
3551 struct vm_struct
*v
;
3553 va
= list_entry(p
, struct vmap_area
, list
);
3556 * s_show can encounter race with remove_vm_area, !vm on behalf
3557 * of vmap area is being tear down or vm_map_ram allocation.
3560 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
3561 (void *)va
->va_start
, (void *)va
->va_end
,
3562 va
->va_end
- va
->va_start
);
3569 seq_printf(m
, "0x%pK-0x%pK %7ld",
3570 v
->addr
, v
->addr
+ v
->size
, v
->size
);
3573 seq_printf(m
, " %pS", v
->caller
);
3576 seq_printf(m
, " pages=%d", v
->nr_pages
);
3579 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
3581 if (v
->flags
& VM_IOREMAP
)
3582 seq_puts(m
, " ioremap");
3584 if (v
->flags
& VM_ALLOC
)
3585 seq_puts(m
, " vmalloc");
3587 if (v
->flags
& VM_MAP
)
3588 seq_puts(m
, " vmap");
3590 if (v
->flags
& VM_USERMAP
)
3591 seq_puts(m
, " user");
3593 if (v
->flags
& VM_DMA_COHERENT
)
3594 seq_puts(m
, " dma-coherent");
3596 if (is_vmalloc_addr(v
->pages
))
3597 seq_puts(m
, " vpages");
3599 show_numa_info(m
, v
);
3603 * As a final step, dump "unpurged" areas. Note,
3604 * that entire "/proc/vmallocinfo" output will not
3605 * be address sorted, because the purge list is not
3608 if (list_is_last(&va
->list
, &vmap_area_list
))
3614 static const struct seq_operations vmalloc_op
= {
3621 static int __init
proc_vmalloc_init(void)
3623 if (IS_ENABLED(CONFIG_NUMA
))
3624 proc_create_seq_private("vmallocinfo", 0400, NULL
,
3626 nr_node_ids
* sizeof(unsigned int), NULL
);
3628 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
3631 module_init(proc_vmalloc_init
);