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/xarray.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>
44 #include "pgalloc-track.h"
46 bool is_vmalloc_addr(const void *x)
48 unsigned long addr = (unsigned long)x;
50 return addr >= VMALLOC_START && addr < VMALLOC_END;
52 EXPORT_SYMBOL(is_vmalloc_addr);
54 struct vfree_deferred {
55 struct llist_head list;
56 struct work_struct wq;
58 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
60 static void __vunmap(const void *, int);
62 static void free_work(struct work_struct *w)
64 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
65 struct llist_node *t, *llnode;
67 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
68 __vunmap((void *)llnode, 1);
71 /*** Page table manipulation functions ***/
73 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
78 pte = pte_offset_kernel(pmd, addr);
80 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
81 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
82 } while (pte++, addr += PAGE_SIZE, addr != end);
83 *mask |= PGTBL_PTE_MODIFIED;
86 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
93 pmd = pmd_offset(pud, addr);
95 next = pmd_addr_end(addr, end);
97 cleared = pmd_clear_huge(pmd);
98 if (cleared || pmd_bad(*pmd))
99 *mask |= PGTBL_PMD_MODIFIED;
103 if (pmd_none_or_clear_bad(pmd))
105 vunmap_pte_range(pmd, addr, next, mask);
106 } while (pmd++, addr = next, addr != end);
109 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
110 pgtbl_mod_mask *mask)
116 pud = pud_offset(p4d, addr);
118 next = pud_addr_end(addr, end);
120 cleared = pud_clear_huge(pud);
121 if (cleared || pud_bad(*pud))
122 *mask |= PGTBL_PUD_MODIFIED;
126 if (pud_none_or_clear_bad(pud))
128 vunmap_pmd_range(pud, addr, next, mask);
129 } while (pud++, addr = next, addr != end);
132 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
133 pgtbl_mod_mask *mask)
139 p4d = p4d_offset(pgd, addr);
141 next = p4d_addr_end(addr, end);
143 cleared = p4d_clear_huge(p4d);
144 if (cleared || p4d_bad(*p4d))
145 *mask |= PGTBL_P4D_MODIFIED;
149 if (p4d_none_or_clear_bad(p4d))
151 vunmap_pud_range(p4d, addr, next, mask);
152 } while (p4d++, addr = next, addr != end);
156 * unmap_kernel_range_noflush - unmap kernel VM area
157 * @start: start of the VM area to unmap
158 * @size: size of the VM area to unmap
160 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
161 * should have been allocated using get_vm_area() and its friends.
164 * This function does NOT do any cache flushing. The caller is responsible
165 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
166 * function and flush_tlb_kernel_range() after.
168 void unmap_kernel_range_noflush(unsigned long start, unsigned long size)
170 unsigned long end = start + size;
173 unsigned long addr = start;
174 pgtbl_mod_mask mask = 0;
178 pgd = pgd_offset_k(addr);
180 next = pgd_addr_end(addr, end);
182 mask |= PGTBL_PGD_MODIFIED;
183 if (pgd_none_or_clear_bad(pgd))
185 vunmap_p4d_range(pgd, addr, next, &mask);
186 } while (pgd++, addr = next, addr != end);
188 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
189 arch_sync_kernel_mappings(start, end);
192 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
193 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
194 pgtbl_mod_mask *mask)
199 * nr is a running index into the array which helps higher level
200 * callers keep track of where we're up to.
203 pte = pte_alloc_kernel_track(pmd, addr, mask);
207 struct page *page = pages[*nr];
209 if (WARN_ON(!pte_none(*pte)))
213 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
215 } while (pte++, addr += PAGE_SIZE, addr != end);
216 *mask |= PGTBL_PTE_MODIFIED;
220 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
221 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
222 pgtbl_mod_mask *mask)
227 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
231 next = pmd_addr_end(addr, end);
232 if (vmap_pte_range(pmd, addr, next, prot, pages, nr, mask))
234 } while (pmd++, addr = next, addr != end);
238 static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
239 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
240 pgtbl_mod_mask *mask)
245 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
249 next = pud_addr_end(addr, end);
250 if (vmap_pmd_range(pud, addr, next, prot, pages, nr, mask))
252 } while (pud++, addr = next, addr != end);
256 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
257 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
258 pgtbl_mod_mask *mask)
263 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
267 next = p4d_addr_end(addr, end);
268 if (vmap_pud_range(p4d, addr, next, prot, pages, nr, mask))
270 } while (p4d++, addr = next, addr != end);
275 * map_kernel_range_noflush - map kernel VM area with the specified pages
276 * @addr: start of the VM area to map
277 * @size: size of the VM area to map
278 * @prot: page protection flags to use
279 * @pages: pages to map
281 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
282 * have been allocated using get_vm_area() and its friends.
285 * This function does NOT do any cache flushing. The caller is responsible for
286 * calling flush_cache_vmap() on to-be-mapped areas before calling this
290 * 0 on success, -errno on failure.
292 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
293 pgprot_t prot, struct page **pages)
295 unsigned long start = addr;
296 unsigned long end = addr + size;
301 pgtbl_mod_mask mask = 0;
304 pgd = pgd_offset_k(addr);
306 next = pgd_addr_end(addr, end);
308 mask |= PGTBL_PGD_MODIFIED;
309 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
312 } while (pgd++, addr = next, addr != end);
314 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
315 arch_sync_kernel_mappings(start, end);
320 int map_kernel_range(unsigned long start, unsigned long size, pgprot_t prot,
325 ret = map_kernel_range_noflush(start, size, prot, pages);
326 flush_cache_vmap(start, start + size);
330 int is_vmalloc_or_module_addr(const void *x)
333 * ARM, x86-64 and sparc64 put modules in a special place,
334 * and fall back on vmalloc() if that fails. Others
335 * just put it in the vmalloc space.
337 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
338 unsigned long addr = (unsigned long)x;
339 if (addr >= MODULES_VADDR && addr < MODULES_END)
342 return is_vmalloc_addr(x);
346 * Walk a vmap address to the struct page it maps.
348 struct page *vmalloc_to_page(const void *vmalloc_addr)
350 unsigned long addr = (unsigned long) vmalloc_addr;
351 struct page *page = NULL;
352 pgd_t *pgd = pgd_offset_k(addr);
359 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
360 * architectures that do not vmalloc module space
362 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
366 p4d = p4d_offset(pgd, addr);
369 pud = pud_offset(p4d, addr);
372 * Don't dereference bad PUD or PMD (below) entries. This will also
373 * identify huge mappings, which we may encounter on architectures
374 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
375 * identified as vmalloc addresses by is_vmalloc_addr(), but are
376 * not [unambiguously] associated with a struct page, so there is
377 * no correct value to return for them.
379 WARN_ON_ONCE(pud_bad(*pud));
380 if (pud_none(*pud) || pud_bad(*pud))
382 pmd = pmd_offset(pud, addr);
383 WARN_ON_ONCE(pmd_bad(*pmd));
384 if (pmd_none(*pmd) || pmd_bad(*pmd))
387 ptep = pte_offset_map(pmd, addr);
389 if (pte_present(pte))
390 page = pte_page(pte);
394 EXPORT_SYMBOL(vmalloc_to_page);
397 * Map a vmalloc()-space virtual address to the physical page frame number.
399 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
401 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
403 EXPORT_SYMBOL(vmalloc_to_pfn);
406 /*** Global kva allocator ***/
408 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
409 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
412 static DEFINE_SPINLOCK(vmap_area_lock);
413 static DEFINE_SPINLOCK(free_vmap_area_lock);
414 /* Export for kexec only */
415 LIST_HEAD(vmap_area_list);
416 static LLIST_HEAD(vmap_purge_list);
417 static struct rb_root vmap_area_root = RB_ROOT;
418 static bool vmap_initialized __read_mostly;
421 * This kmem_cache is used for vmap_area objects. Instead of
422 * allocating from slab we reuse an object from this cache to
423 * make things faster. Especially in "no edge" splitting of
426 static struct kmem_cache *vmap_area_cachep;
429 * This linked list is used in pair with free_vmap_area_root.
430 * It gives O(1) access to prev/next to perform fast coalescing.
432 static LIST_HEAD(free_vmap_area_list);
435 * This augment red-black tree represents the free vmap space.
436 * All vmap_area objects in this tree are sorted by va->va_start
437 * address. It is used for allocation and merging when a vmap
438 * object is released.
440 * Each vmap_area node contains a maximum available free block
441 * of its sub-tree, right or left. Therefore it is possible to
442 * find a lowest match of free area.
444 static struct rb_root free_vmap_area_root = RB_ROOT;
447 * Preload a CPU with one object for "no edge" split case. The
448 * aim is to get rid of allocations from the atomic context, thus
449 * to use more permissive allocation masks.
451 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
453 static __always_inline unsigned long
454 va_size(struct vmap_area *va)
456 return (va->va_end - va->va_start);
459 static __always_inline unsigned long
460 get_subtree_max_size(struct rb_node *node)
462 struct vmap_area *va;
464 va = rb_entry_safe(node, struct vmap_area, rb_node);
465 return va ? va->subtree_max_size : 0;
469 * Gets called when remove the node and rotate.
471 static __always_inline unsigned long
472 compute_subtree_max_size(struct vmap_area *va)
474 return max3(va_size(va),
475 get_subtree_max_size(va->rb_node.rb_left),
476 get_subtree_max_size(va->rb_node.rb_right));
479 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
480 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
482 static void purge_vmap_area_lazy(void);
483 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
484 static unsigned long lazy_max_pages(void);
486 static atomic_long_t nr_vmalloc_pages;
488 unsigned long vmalloc_nr_pages(void)
490 return atomic_long_read(&nr_vmalloc_pages);
493 static struct vmap_area *__find_vmap_area(unsigned long addr)
495 struct rb_node *n = vmap_area_root.rb_node;
498 struct vmap_area *va;
500 va = rb_entry(n, struct vmap_area, rb_node);
501 if (addr < va->va_start)
503 else if (addr >= va->va_end)
513 * This function returns back addresses of parent node
514 * and its left or right link for further processing.
516 static __always_inline struct rb_node **
517 find_va_links(struct vmap_area *va,
518 struct rb_root *root, struct rb_node *from,
519 struct rb_node **parent)
521 struct vmap_area *tmp_va;
522 struct rb_node **link;
525 link = &root->rb_node;
526 if (unlikely(!*link)) {
535 * Go to the bottom of the tree. When we hit the last point
536 * we end up with parent rb_node and correct direction, i name
537 * it link, where the new va->rb_node will be attached to.
540 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
543 * During the traversal we also do some sanity check.
544 * Trigger the BUG() if there are sides(left/right)
547 if (va->va_start < tmp_va->va_end &&
548 va->va_end <= tmp_va->va_start)
549 link = &(*link)->rb_left;
550 else if (va->va_end > tmp_va->va_start &&
551 va->va_start >= tmp_va->va_end)
552 link = &(*link)->rb_right;
557 *parent = &tmp_va->rb_node;
561 static __always_inline struct list_head *
562 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
564 struct list_head *list;
566 if (unlikely(!parent))
568 * The red-black tree where we try to find VA neighbors
569 * before merging or inserting is empty, i.e. it means
570 * there is no free vmap space. Normally it does not
571 * happen but we handle this case anyway.
575 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
576 return (&parent->rb_right == link ? list->next : list);
579 static __always_inline void
580 link_va(struct vmap_area *va, struct rb_root *root,
581 struct rb_node *parent, struct rb_node **link, struct list_head *head)
584 * VA is still not in the list, but we can
585 * identify its future previous list_head node.
587 if (likely(parent)) {
588 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
589 if (&parent->rb_right != link)
593 /* Insert to the rb-tree */
594 rb_link_node(&va->rb_node, parent, link);
595 if (root == &free_vmap_area_root) {
597 * Some explanation here. Just perform simple insertion
598 * to the tree. We do not set va->subtree_max_size to
599 * its current size before calling rb_insert_augmented().
600 * It is because of we populate the tree from the bottom
601 * to parent levels when the node _is_ in the tree.
603 * Therefore we set subtree_max_size to zero after insertion,
604 * to let __augment_tree_propagate_from() puts everything to
605 * the correct order later on.
607 rb_insert_augmented(&va->rb_node,
608 root, &free_vmap_area_rb_augment_cb);
609 va->subtree_max_size = 0;
611 rb_insert_color(&va->rb_node, root);
614 /* Address-sort this list */
615 list_add(&va->list, head);
618 static __always_inline void
619 unlink_va(struct vmap_area *va, struct rb_root *root)
621 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
624 if (root == &free_vmap_area_root)
625 rb_erase_augmented(&va->rb_node,
626 root, &free_vmap_area_rb_augment_cb);
628 rb_erase(&va->rb_node, root);
631 RB_CLEAR_NODE(&va->rb_node);
634 #if DEBUG_AUGMENT_PROPAGATE_CHECK
636 augment_tree_propagate_check(void)
638 struct vmap_area *va;
639 unsigned long computed_size;
641 list_for_each_entry(va, &free_vmap_area_list, list) {
642 computed_size = compute_subtree_max_size(va);
643 if (computed_size != va->subtree_max_size)
644 pr_emerg("tree is corrupted: %lu, %lu\n",
645 va_size(va), va->subtree_max_size);
651 * This function populates subtree_max_size from bottom to upper
652 * levels starting from VA point. The propagation must be done
653 * when VA size is modified by changing its va_start/va_end. Or
654 * in case of newly inserting of VA to the tree.
656 * It means that __augment_tree_propagate_from() must be called:
657 * - After VA has been inserted to the tree(free path);
658 * - After VA has been shrunk(allocation path);
659 * - After VA has been increased(merging path).
661 * Please note that, it does not mean that upper parent nodes
662 * and their subtree_max_size are recalculated all the time up
671 * For example if we modify the node 4, shrinking it to 2, then
672 * no any modification is required. If we shrink the node 2 to 1
673 * its subtree_max_size is updated only, and set to 1. If we shrink
674 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
677 static __always_inline void
678 augment_tree_propagate_from(struct vmap_area *va)
681 * Populate the tree from bottom towards the root until
682 * the calculated maximum available size of checked node
683 * is equal to its current one.
685 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
687 #if DEBUG_AUGMENT_PROPAGATE_CHECK
688 augment_tree_propagate_check();
693 insert_vmap_area(struct vmap_area *va,
694 struct rb_root *root, struct list_head *head)
696 struct rb_node **link;
697 struct rb_node *parent;
699 link = find_va_links(va, root, NULL, &parent);
700 link_va(va, root, parent, link, head);
704 insert_vmap_area_augment(struct vmap_area *va,
705 struct rb_node *from, struct rb_root *root,
706 struct list_head *head)
708 struct rb_node **link;
709 struct rb_node *parent;
712 link = find_va_links(va, NULL, from, &parent);
714 link = find_va_links(va, root, NULL, &parent);
716 link_va(va, root, parent, link, head);
717 augment_tree_propagate_from(va);
721 * Merge de-allocated chunk of VA memory with previous
722 * and next free blocks. If coalesce is not done a new
723 * free area is inserted. If VA has been merged, it is
726 static __always_inline struct vmap_area *
727 merge_or_add_vmap_area(struct vmap_area *va,
728 struct rb_root *root, struct list_head *head)
730 struct vmap_area *sibling;
731 struct list_head *next;
732 struct rb_node **link;
733 struct rb_node *parent;
737 * Find a place in the tree where VA potentially will be
738 * inserted, unless it is merged with its sibling/siblings.
740 link = find_va_links(va, root, NULL, &parent);
743 * Get next node of VA to check if merging can be done.
745 next = get_va_next_sibling(parent, link);
746 if (unlikely(next == NULL))
752 * |<------VA------>|<-----Next----->|
757 sibling = list_entry(next, struct vmap_area, list);
758 if (sibling->va_start == va->va_end) {
759 sibling->va_start = va->va_start;
761 /* Free vmap_area object. */
762 kmem_cache_free(vmap_area_cachep, va);
764 /* Point to the new merged area. */
773 * |<-----Prev----->|<------VA------>|
777 if (next->prev != head) {
778 sibling = list_entry(next->prev, struct vmap_area, list);
779 if (sibling->va_end == va->va_start) {
781 * If both neighbors are coalesced, it is important
782 * to unlink the "next" node first, followed by merging
783 * with "previous" one. Otherwise the tree might not be
784 * fully populated if a sibling's augmented value is
785 * "normalized" because of rotation operations.
790 sibling->va_end = va->va_end;
792 /* Free vmap_area object. */
793 kmem_cache_free(vmap_area_cachep, va);
795 /* Point to the new merged area. */
803 link_va(va, root, parent, link, head);
806 * Last step is to check and update the tree.
808 augment_tree_propagate_from(va);
812 static __always_inline bool
813 is_within_this_va(struct vmap_area *va, unsigned long size,
814 unsigned long align, unsigned long vstart)
816 unsigned long nva_start_addr;
818 if (va->va_start > vstart)
819 nva_start_addr = ALIGN(va->va_start, align);
821 nva_start_addr = ALIGN(vstart, align);
823 /* Can be overflowed due to big size or alignment. */
824 if (nva_start_addr + size < nva_start_addr ||
825 nva_start_addr < vstart)
828 return (nva_start_addr + size <= va->va_end);
832 * Find the first free block(lowest start address) in the tree,
833 * that will accomplish the request corresponding to passing
836 static __always_inline struct vmap_area *
837 find_vmap_lowest_match(unsigned long size,
838 unsigned long align, unsigned long vstart)
840 struct vmap_area *va;
841 struct rb_node *node;
842 unsigned long length;
844 /* Start from the root. */
845 node = free_vmap_area_root.rb_node;
847 /* Adjust the search size for alignment overhead. */
848 length = size + align - 1;
851 va = rb_entry(node, struct vmap_area, rb_node);
853 if (get_subtree_max_size(node->rb_left) >= length &&
854 vstart < va->va_start) {
855 node = node->rb_left;
857 if (is_within_this_va(va, size, align, vstart))
861 * Does not make sense to go deeper towards the right
862 * sub-tree if it does not have a free block that is
863 * equal or bigger to the requested search length.
865 if (get_subtree_max_size(node->rb_right) >= length) {
866 node = node->rb_right;
871 * OK. We roll back and find the first right sub-tree,
872 * that will satisfy the search criteria. It can happen
873 * only once due to "vstart" restriction.
875 while ((node = rb_parent(node))) {
876 va = rb_entry(node, struct vmap_area, rb_node);
877 if (is_within_this_va(va, size, align, vstart))
880 if (get_subtree_max_size(node->rb_right) >= length &&
881 vstart <= va->va_start) {
882 node = node->rb_right;
892 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
893 #include <linux/random.h>
895 static struct vmap_area *
896 find_vmap_lowest_linear_match(unsigned long size,
897 unsigned long align, unsigned long vstart)
899 struct vmap_area *va;
901 list_for_each_entry(va, &free_vmap_area_list, list) {
902 if (!is_within_this_va(va, size, align, vstart))
912 find_vmap_lowest_match_check(unsigned long size)
914 struct vmap_area *va_1, *va_2;
915 unsigned long vstart;
918 get_random_bytes(&rnd, sizeof(rnd));
919 vstart = VMALLOC_START + rnd;
921 va_1 = find_vmap_lowest_match(size, 1, vstart);
922 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
925 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
932 FL_FIT_TYPE = 1, /* full fit */
933 LE_FIT_TYPE = 2, /* left edge fit */
934 RE_FIT_TYPE = 3, /* right edge fit */
935 NE_FIT_TYPE = 4 /* no edge fit */
938 static __always_inline enum fit_type
939 classify_va_fit_type(struct vmap_area *va,
940 unsigned long nva_start_addr, unsigned long size)
944 /* Check if it is within VA. */
945 if (nva_start_addr < va->va_start ||
946 nva_start_addr + size > va->va_end)
950 if (va->va_start == nva_start_addr) {
951 if (va->va_end == nva_start_addr + size)
955 } else if (va->va_end == nva_start_addr + size) {
964 static __always_inline int
965 adjust_va_to_fit_type(struct vmap_area *va,
966 unsigned long nva_start_addr, unsigned long size,
969 struct vmap_area *lva = NULL;
971 if (type == FL_FIT_TYPE) {
973 * No need to split VA, it fully fits.
979 unlink_va(va, &free_vmap_area_root);
980 kmem_cache_free(vmap_area_cachep, va);
981 } else if (type == LE_FIT_TYPE) {
983 * Split left edge of fit VA.
989 va->va_start += size;
990 } else if (type == RE_FIT_TYPE) {
992 * Split right edge of fit VA.
998 va->va_end = nva_start_addr;
999 } else if (type == NE_FIT_TYPE) {
1001 * Split no edge of fit VA.
1007 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1008 if (unlikely(!lva)) {
1010 * For percpu allocator we do not do any pre-allocation
1011 * and leave it as it is. The reason is it most likely
1012 * never ends up with NE_FIT_TYPE splitting. In case of
1013 * percpu allocations offsets and sizes are aligned to
1014 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1015 * are its main fitting cases.
1017 * There are a few exceptions though, as an example it is
1018 * a first allocation (early boot up) when we have "one"
1019 * big free space that has to be split.
1021 * Also we can hit this path in case of regular "vmap"
1022 * allocations, if "this" current CPU was not preloaded.
1023 * See the comment in alloc_vmap_area() why. If so, then
1024 * GFP_NOWAIT is used instead to get an extra object for
1025 * split purpose. That is rare and most time does not
1028 * What happens if an allocation gets failed. Basically,
1029 * an "overflow" path is triggered to purge lazily freed
1030 * areas to free some memory, then, the "retry" path is
1031 * triggered to repeat one more time. See more details
1032 * in alloc_vmap_area() function.
1034 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1040 * Build the remainder.
1042 lva->va_start = va->va_start;
1043 lva->va_end = nva_start_addr;
1046 * Shrink this VA to remaining size.
1048 va->va_start = nva_start_addr + size;
1053 if (type != FL_FIT_TYPE) {
1054 augment_tree_propagate_from(va);
1056 if (lva) /* type == NE_FIT_TYPE */
1057 insert_vmap_area_augment(lva, &va->rb_node,
1058 &free_vmap_area_root, &free_vmap_area_list);
1065 * Returns a start address of the newly allocated area, if success.
1066 * Otherwise a vend is returned that indicates failure.
1068 static __always_inline unsigned long
1069 __alloc_vmap_area(unsigned long size, unsigned long align,
1070 unsigned long vstart, unsigned long vend)
1072 unsigned long nva_start_addr;
1073 struct vmap_area *va;
1077 va = find_vmap_lowest_match(size, align, vstart);
1081 if (va->va_start > vstart)
1082 nva_start_addr = ALIGN(va->va_start, align);
1084 nva_start_addr = ALIGN(vstart, align);
1086 /* Check the "vend" restriction. */
1087 if (nva_start_addr + size > vend)
1090 /* Classify what we have found. */
1091 type = classify_va_fit_type(va, nva_start_addr, size);
1092 if (WARN_ON_ONCE(type == NOTHING_FIT))
1095 /* Update the free vmap_area. */
1096 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1100 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1101 find_vmap_lowest_match_check(size);
1104 return nva_start_addr;
1108 * Free a region of KVA allocated by alloc_vmap_area
1110 static void free_vmap_area(struct vmap_area *va)
1113 * Remove from the busy tree/list.
1115 spin_lock(&vmap_area_lock);
1116 unlink_va(va, &vmap_area_root);
1117 spin_unlock(&vmap_area_lock);
1120 * Insert/Merge it back to the free tree/list.
1122 spin_lock(&free_vmap_area_lock);
1123 merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list);
1124 spin_unlock(&free_vmap_area_lock);
1128 * Allocate a region of KVA of the specified size and alignment, within the
1131 static struct vmap_area *alloc_vmap_area(unsigned long size,
1132 unsigned long align,
1133 unsigned long vstart, unsigned long vend,
1134 int node, gfp_t gfp_mask)
1136 struct vmap_area *va, *pva;
1142 BUG_ON(offset_in_page(size));
1143 BUG_ON(!is_power_of_2(align));
1145 if (unlikely(!vmap_initialized))
1146 return ERR_PTR(-EBUSY);
1149 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1151 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1153 return ERR_PTR(-ENOMEM);
1156 * Only scan the relevant parts containing pointers to other objects
1157 * to avoid false negatives.
1159 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1163 * Preload this CPU with one extra vmap_area object. It is used
1164 * when fit type of free area is NE_FIT_TYPE. Please note, it
1165 * does not guarantee that an allocation occurs on a CPU that
1166 * is preloaded, instead we minimize the case when it is not.
1167 * It can happen because of cpu migration, because there is a
1168 * race until the below spinlock is taken.
1170 * The preload is done in non-atomic context, thus it allows us
1171 * to use more permissive allocation masks to be more stable under
1172 * low memory condition and high memory pressure. In rare case,
1173 * if not preloaded, GFP_NOWAIT is used.
1175 * Set "pva" to NULL here, because of "retry" path.
1179 if (!this_cpu_read(ne_fit_preload_node))
1181 * Even if it fails we do not really care about that.
1182 * Just proceed as it is. If needed "overflow" path
1183 * will refill the cache we allocate from.
1185 pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1187 spin_lock(&free_vmap_area_lock);
1189 if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva))
1190 kmem_cache_free(vmap_area_cachep, pva);
1193 * If an allocation fails, the "vend" address is
1194 * returned. Therefore trigger the overflow path.
1196 addr = __alloc_vmap_area(size, align, vstart, vend);
1197 spin_unlock(&free_vmap_area_lock);
1199 if (unlikely(addr == vend))
1202 va->va_start = addr;
1203 va->va_end = addr + size;
1207 spin_lock(&vmap_area_lock);
1208 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1209 spin_unlock(&vmap_area_lock);
1211 BUG_ON(!IS_ALIGNED(va->va_start, align));
1212 BUG_ON(va->va_start < vstart);
1213 BUG_ON(va->va_end > vend);
1215 ret = kasan_populate_vmalloc(addr, size);
1218 return ERR_PTR(ret);
1225 purge_vmap_area_lazy();
1230 if (gfpflags_allow_blocking(gfp_mask)) {
1231 unsigned long freed = 0;
1232 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1239 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1240 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1243 kmem_cache_free(vmap_area_cachep, va);
1244 return ERR_PTR(-EBUSY);
1247 int register_vmap_purge_notifier(struct notifier_block *nb)
1249 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1251 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1253 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1255 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1257 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1260 * lazy_max_pages is the maximum amount of virtual address space we gather up
1261 * before attempting to purge with a TLB flush.
1263 * There is a tradeoff here: a larger number will cover more kernel page tables
1264 * and take slightly longer to purge, but it will linearly reduce the number of
1265 * global TLB flushes that must be performed. It would seem natural to scale
1266 * this number up linearly with the number of CPUs (because vmapping activity
1267 * could also scale linearly with the number of CPUs), however it is likely
1268 * that in practice, workloads might be constrained in other ways that mean
1269 * vmap activity will not scale linearly with CPUs. Also, I want to be
1270 * conservative and not introduce a big latency on huge systems, so go with
1271 * a less aggressive log scale. It will still be an improvement over the old
1272 * code, and it will be simple to change the scale factor if we find that it
1273 * becomes a problem on bigger systems.
1275 static unsigned long lazy_max_pages(void)
1279 log = fls(num_online_cpus());
1281 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1284 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1287 * Serialize vmap purging. There is no actual criticial section protected
1288 * by this look, but we want to avoid concurrent calls for performance
1289 * reasons and to make the pcpu_get_vm_areas more deterministic.
1291 static DEFINE_MUTEX(vmap_purge_lock);
1293 /* for per-CPU blocks */
1294 static void purge_fragmented_blocks_allcpus(void);
1297 * called before a call to iounmap() if the caller wants vm_area_struct's
1298 * immediately freed.
1300 void set_iounmap_nonlazy(void)
1302 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1306 * Purges all lazily-freed vmap areas.
1308 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1310 unsigned long resched_threshold;
1311 struct llist_node *valist;
1312 struct vmap_area *va;
1313 struct vmap_area *n_va;
1315 lockdep_assert_held(&vmap_purge_lock);
1317 valist = llist_del_all(&vmap_purge_list);
1318 if (unlikely(valist == NULL))
1322 * TODO: to calculate a flush range without looping.
1323 * The list can be up to lazy_max_pages() elements.
1325 llist_for_each_entry(va, valist, purge_list) {
1326 if (va->va_start < start)
1327 start = va->va_start;
1328 if (va->va_end > end)
1332 flush_tlb_kernel_range(start, end);
1333 resched_threshold = lazy_max_pages() << 1;
1335 spin_lock(&free_vmap_area_lock);
1336 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1337 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1338 unsigned long orig_start = va->va_start;
1339 unsigned long orig_end = va->va_end;
1342 * Finally insert or merge lazily-freed area. It is
1343 * detached and there is no need to "unlink" it from
1346 va = merge_or_add_vmap_area(va, &free_vmap_area_root,
1347 &free_vmap_area_list);
1349 if (is_vmalloc_or_module_addr((void *)orig_start))
1350 kasan_release_vmalloc(orig_start, orig_end,
1351 va->va_start, va->va_end);
1353 atomic_long_sub(nr, &vmap_lazy_nr);
1355 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1356 cond_resched_lock(&free_vmap_area_lock);
1358 spin_unlock(&free_vmap_area_lock);
1363 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1364 * is already purging.
1366 static void try_purge_vmap_area_lazy(void)
1368 if (mutex_trylock(&vmap_purge_lock)) {
1369 __purge_vmap_area_lazy(ULONG_MAX, 0);
1370 mutex_unlock(&vmap_purge_lock);
1375 * Kick off a purge of the outstanding lazy areas.
1377 static void purge_vmap_area_lazy(void)
1379 mutex_lock(&vmap_purge_lock);
1380 purge_fragmented_blocks_allcpus();
1381 __purge_vmap_area_lazy(ULONG_MAX, 0);
1382 mutex_unlock(&vmap_purge_lock);
1386 * Free a vmap area, caller ensuring that the area has been unmapped
1387 * and flush_cache_vunmap had been called for the correct range
1390 static void free_vmap_area_noflush(struct vmap_area *va)
1392 unsigned long nr_lazy;
1394 spin_lock(&vmap_area_lock);
1395 unlink_va(va, &vmap_area_root);
1396 spin_unlock(&vmap_area_lock);
1398 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1399 PAGE_SHIFT, &vmap_lazy_nr);
1401 /* After this point, we may free va at any time */
1402 llist_add(&va->purge_list, &vmap_purge_list);
1404 if (unlikely(nr_lazy > lazy_max_pages()))
1405 try_purge_vmap_area_lazy();
1409 * Free and unmap a vmap area
1411 static void free_unmap_vmap_area(struct vmap_area *va)
1413 flush_cache_vunmap(va->va_start, va->va_end);
1414 unmap_kernel_range_noflush(va->va_start, va->va_end - va->va_start);
1415 if (debug_pagealloc_enabled_static())
1416 flush_tlb_kernel_range(va->va_start, va->va_end);
1418 free_vmap_area_noflush(va);
1421 static struct vmap_area *find_vmap_area(unsigned long addr)
1423 struct vmap_area *va;
1425 spin_lock(&vmap_area_lock);
1426 va = __find_vmap_area(addr);
1427 spin_unlock(&vmap_area_lock);
1432 /*** Per cpu kva allocator ***/
1435 * vmap space is limited especially on 32 bit architectures. Ensure there is
1436 * room for at least 16 percpu vmap blocks per CPU.
1439 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1440 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1441 * instead (we just need a rough idea)
1443 #if BITS_PER_LONG == 32
1444 #define VMALLOC_SPACE (128UL*1024*1024)
1446 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1449 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1450 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1451 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1452 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1453 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1454 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1455 #define VMAP_BBMAP_BITS \
1456 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1457 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1458 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1460 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1462 struct vmap_block_queue {
1464 struct list_head free;
1469 struct vmap_area *va;
1470 unsigned long free, dirty;
1471 unsigned long dirty_min, dirty_max; /*< dirty range */
1472 struct list_head free_list;
1473 struct rcu_head rcu_head;
1474 struct list_head purge;
1477 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1478 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1481 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1482 * in the free path. Could get rid of this if we change the API to return a
1483 * "cookie" from alloc, to be passed to free. But no big deal yet.
1485 static DEFINE_XARRAY(vmap_blocks);
1488 * We should probably have a fallback mechanism to allocate virtual memory
1489 * out of partially filled vmap blocks. However vmap block sizing should be
1490 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1494 static unsigned long addr_to_vb_idx(unsigned long addr)
1496 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1497 addr /= VMAP_BLOCK_SIZE;
1501 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1505 addr = va_start + (pages_off << PAGE_SHIFT);
1506 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1507 return (void *)addr;
1511 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1512 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1513 * @order: how many 2^order pages should be occupied in newly allocated block
1514 * @gfp_mask: flags for the page level allocator
1516 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1518 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1520 struct vmap_block_queue *vbq;
1521 struct vmap_block *vb;
1522 struct vmap_area *va;
1523 unsigned long vb_idx;
1527 node = numa_node_id();
1529 vb = kmalloc_node(sizeof(struct vmap_block),
1530 gfp_mask & GFP_RECLAIM_MASK, node);
1532 return ERR_PTR(-ENOMEM);
1534 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1535 VMALLOC_START, VMALLOC_END,
1539 return ERR_CAST(va);
1542 vaddr = vmap_block_vaddr(va->va_start, 0);
1543 spin_lock_init(&vb->lock);
1545 /* At least something should be left free */
1546 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1547 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1549 vb->dirty_min = VMAP_BBMAP_BITS;
1551 INIT_LIST_HEAD(&vb->free_list);
1553 vb_idx = addr_to_vb_idx(va->va_start);
1554 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1558 return ERR_PTR(err);
1561 vbq = &get_cpu_var(vmap_block_queue);
1562 spin_lock(&vbq->lock);
1563 list_add_tail_rcu(&vb->free_list, &vbq->free);
1564 spin_unlock(&vbq->lock);
1565 put_cpu_var(vmap_block_queue);
1570 static void free_vmap_block(struct vmap_block *vb)
1572 struct vmap_block *tmp;
1574 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1577 free_vmap_area_noflush(vb->va);
1578 kfree_rcu(vb, rcu_head);
1581 static void purge_fragmented_blocks(int cpu)
1584 struct vmap_block *vb;
1585 struct vmap_block *n_vb;
1586 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1589 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1591 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1594 spin_lock(&vb->lock);
1595 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1596 vb->free = 0; /* prevent further allocs after releasing lock */
1597 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1599 vb->dirty_max = VMAP_BBMAP_BITS;
1600 spin_lock(&vbq->lock);
1601 list_del_rcu(&vb->free_list);
1602 spin_unlock(&vbq->lock);
1603 spin_unlock(&vb->lock);
1604 list_add_tail(&vb->purge, &purge);
1606 spin_unlock(&vb->lock);
1610 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1611 list_del(&vb->purge);
1612 free_vmap_block(vb);
1616 static void purge_fragmented_blocks_allcpus(void)
1620 for_each_possible_cpu(cpu)
1621 purge_fragmented_blocks(cpu);
1624 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1626 struct vmap_block_queue *vbq;
1627 struct vmap_block *vb;
1631 BUG_ON(offset_in_page(size));
1632 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1633 if (WARN_ON(size == 0)) {
1635 * Allocating 0 bytes isn't what caller wants since
1636 * get_order(0) returns funny result. Just warn and terminate
1641 order = get_order(size);
1644 vbq = &get_cpu_var(vmap_block_queue);
1645 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1646 unsigned long pages_off;
1648 spin_lock(&vb->lock);
1649 if (vb->free < (1UL << order)) {
1650 spin_unlock(&vb->lock);
1654 pages_off = VMAP_BBMAP_BITS - vb->free;
1655 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1656 vb->free -= 1UL << order;
1657 if (vb->free == 0) {
1658 spin_lock(&vbq->lock);
1659 list_del_rcu(&vb->free_list);
1660 spin_unlock(&vbq->lock);
1663 spin_unlock(&vb->lock);
1667 put_cpu_var(vmap_block_queue);
1670 /* Allocate new block if nothing was found */
1672 vaddr = new_vmap_block(order, gfp_mask);
1677 static void vb_free(unsigned long addr, unsigned long size)
1679 unsigned long offset;
1681 struct vmap_block *vb;
1683 BUG_ON(offset_in_page(size));
1684 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1686 flush_cache_vunmap(addr, addr + size);
1688 order = get_order(size);
1689 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
1690 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
1692 unmap_kernel_range_noflush(addr, size);
1694 if (debug_pagealloc_enabled_static())
1695 flush_tlb_kernel_range(addr, addr + size);
1697 spin_lock(&vb->lock);
1699 /* Expand dirty range */
1700 vb->dirty_min = min(vb->dirty_min, offset);
1701 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1703 vb->dirty += 1UL << order;
1704 if (vb->dirty == VMAP_BBMAP_BITS) {
1706 spin_unlock(&vb->lock);
1707 free_vmap_block(vb);
1709 spin_unlock(&vb->lock);
1712 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1716 if (unlikely(!vmap_initialized))
1721 for_each_possible_cpu(cpu) {
1722 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1723 struct vmap_block *vb;
1726 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1727 spin_lock(&vb->lock);
1729 unsigned long va_start = vb->va->va_start;
1732 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1733 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1735 start = min(s, start);
1740 spin_unlock(&vb->lock);
1745 mutex_lock(&vmap_purge_lock);
1746 purge_fragmented_blocks_allcpus();
1747 if (!__purge_vmap_area_lazy(start, end) && flush)
1748 flush_tlb_kernel_range(start, end);
1749 mutex_unlock(&vmap_purge_lock);
1753 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1755 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1756 * to amortize TLB flushing overheads. What this means is that any page you
1757 * have now, may, in a former life, have been mapped into kernel virtual
1758 * address by the vmap layer and so there might be some CPUs with TLB entries
1759 * still referencing that page (additional to the regular 1:1 kernel mapping).
1761 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1762 * be sure that none of the pages we have control over will have any aliases
1763 * from the vmap layer.
1765 void vm_unmap_aliases(void)
1767 unsigned long start = ULONG_MAX, end = 0;
1770 _vm_unmap_aliases(start, end, flush);
1772 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1775 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1776 * @mem: the pointer returned by vm_map_ram
1777 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1779 void vm_unmap_ram(const void *mem, unsigned int count)
1781 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1782 unsigned long addr = (unsigned long)mem;
1783 struct vmap_area *va;
1787 BUG_ON(addr < VMALLOC_START);
1788 BUG_ON(addr > VMALLOC_END);
1789 BUG_ON(!PAGE_ALIGNED(addr));
1791 kasan_poison_vmalloc(mem, size);
1793 if (likely(count <= VMAP_MAX_ALLOC)) {
1794 debug_check_no_locks_freed(mem, size);
1795 vb_free(addr, size);
1799 va = find_vmap_area(addr);
1801 debug_check_no_locks_freed((void *)va->va_start,
1802 (va->va_end - va->va_start));
1803 free_unmap_vmap_area(va);
1805 EXPORT_SYMBOL(vm_unmap_ram);
1808 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1809 * @pages: an array of pointers to the pages to be mapped
1810 * @count: number of pages
1811 * @node: prefer to allocate data structures on this node
1813 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1814 * faster than vmap so it's good. But if you mix long-life and short-life
1815 * objects with vm_map_ram(), it could consume lots of address space through
1816 * fragmentation (especially on a 32bit machine). You could see failures in
1817 * the end. Please use this function for short-lived objects.
1819 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1821 void *vm_map_ram(struct page **pages, unsigned int count, int node)
1823 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1827 if (likely(count <= VMAP_MAX_ALLOC)) {
1828 mem = vb_alloc(size, GFP_KERNEL);
1831 addr = (unsigned long)mem;
1833 struct vmap_area *va;
1834 va = alloc_vmap_area(size, PAGE_SIZE,
1835 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1839 addr = va->va_start;
1843 kasan_unpoison_vmalloc(mem, size);
1845 if (map_kernel_range(addr, size, PAGE_KERNEL, pages) < 0) {
1846 vm_unmap_ram(mem, count);
1851 EXPORT_SYMBOL(vm_map_ram);
1853 static struct vm_struct *vmlist __initdata;
1856 * vm_area_add_early - add vmap area early during boot
1857 * @vm: vm_struct to add
1859 * This function is used to add fixed kernel vm area to vmlist before
1860 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1861 * should contain proper values and the other fields should be zero.
1863 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1865 void __init vm_area_add_early(struct vm_struct *vm)
1867 struct vm_struct *tmp, **p;
1869 BUG_ON(vmap_initialized);
1870 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1871 if (tmp->addr >= vm->addr) {
1872 BUG_ON(tmp->addr < vm->addr + vm->size);
1875 BUG_ON(tmp->addr + tmp->size > vm->addr);
1882 * vm_area_register_early - register vmap area early during boot
1883 * @vm: vm_struct to register
1884 * @align: requested alignment
1886 * This function is used to register kernel vm area before
1887 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1888 * proper values on entry and other fields should be zero. On return,
1889 * vm->addr contains the allocated address.
1891 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1893 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1895 static size_t vm_init_off __initdata;
1898 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1899 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1901 vm->addr = (void *)addr;
1903 vm_area_add_early(vm);
1906 static void vmap_init_free_space(void)
1908 unsigned long vmap_start = 1;
1909 const unsigned long vmap_end = ULONG_MAX;
1910 struct vmap_area *busy, *free;
1914 * -|-----|.....|-----|-----|-----|.....|-
1916 * |<--------------------------------->|
1918 list_for_each_entry(busy, &vmap_area_list, list) {
1919 if (busy->va_start - vmap_start > 0) {
1920 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1921 if (!WARN_ON_ONCE(!free)) {
1922 free->va_start = vmap_start;
1923 free->va_end = busy->va_start;
1925 insert_vmap_area_augment(free, NULL,
1926 &free_vmap_area_root,
1927 &free_vmap_area_list);
1931 vmap_start = busy->va_end;
1934 if (vmap_end - vmap_start > 0) {
1935 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1936 if (!WARN_ON_ONCE(!free)) {
1937 free->va_start = vmap_start;
1938 free->va_end = vmap_end;
1940 insert_vmap_area_augment(free, NULL,
1941 &free_vmap_area_root,
1942 &free_vmap_area_list);
1947 void __init vmalloc_init(void)
1949 struct vmap_area *va;
1950 struct vm_struct *tmp;
1954 * Create the cache for vmap_area objects.
1956 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1958 for_each_possible_cpu(i) {
1959 struct vmap_block_queue *vbq;
1960 struct vfree_deferred *p;
1962 vbq = &per_cpu(vmap_block_queue, i);
1963 spin_lock_init(&vbq->lock);
1964 INIT_LIST_HEAD(&vbq->free);
1965 p = &per_cpu(vfree_deferred, i);
1966 init_llist_head(&p->list);
1967 INIT_WORK(&p->wq, free_work);
1970 /* Import existing vmlist entries. */
1971 for (tmp = vmlist; tmp; tmp = tmp->next) {
1972 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1973 if (WARN_ON_ONCE(!va))
1976 va->va_start = (unsigned long)tmp->addr;
1977 va->va_end = va->va_start + tmp->size;
1979 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1983 * Now we can initialize a free vmap space.
1985 vmap_init_free_space();
1986 vmap_initialized = true;
1990 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1991 * @addr: start of the VM area to unmap
1992 * @size: size of the VM area to unmap
1994 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1995 * the unmapping and tlb after.
1997 void unmap_kernel_range(unsigned long addr, unsigned long size)
1999 unsigned long end = addr + size;
2001 flush_cache_vunmap(addr, end);
2002 unmap_kernel_range_noflush(addr, size);
2003 flush_tlb_kernel_range(addr, end);
2006 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2007 struct vmap_area *va, unsigned long flags, const void *caller)
2010 vm->addr = (void *)va->va_start;
2011 vm->size = va->va_end - va->va_start;
2012 vm->caller = caller;
2016 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2017 unsigned long flags, const void *caller)
2019 spin_lock(&vmap_area_lock);
2020 setup_vmalloc_vm_locked(vm, va, flags, caller);
2021 spin_unlock(&vmap_area_lock);
2024 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2027 * Before removing VM_UNINITIALIZED,
2028 * we should make sure that vm has proper values.
2029 * Pair with smp_rmb() in show_numa_info().
2032 vm->flags &= ~VM_UNINITIALIZED;
2035 static struct vm_struct *__get_vm_area_node(unsigned long size,
2036 unsigned long align, unsigned long flags, unsigned long start,
2037 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2039 struct vmap_area *va;
2040 struct vm_struct *area;
2041 unsigned long requested_size = size;
2043 BUG_ON(in_interrupt());
2044 size = PAGE_ALIGN(size);
2045 if (unlikely(!size))
2048 if (flags & VM_IOREMAP)
2049 align = 1ul << clamp_t(int, get_count_order_long(size),
2050 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2052 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2053 if (unlikely(!area))
2056 if (!(flags & VM_NO_GUARD))
2059 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2065 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2067 setup_vmalloc_vm(area, va, flags, caller);
2072 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2073 unsigned long start, unsigned long end,
2076 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2077 GFP_KERNEL, caller);
2081 * get_vm_area - reserve a contiguous kernel virtual area
2082 * @size: size of the area
2083 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2085 * Search an area of @size in the kernel virtual mapping area,
2086 * and reserved it for out purposes. Returns the area descriptor
2087 * on success or %NULL on failure.
2089 * Return: the area descriptor on success or %NULL on failure.
2091 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2093 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2094 NUMA_NO_NODE, GFP_KERNEL,
2095 __builtin_return_address(0));
2098 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2101 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2102 NUMA_NO_NODE, GFP_KERNEL, caller);
2106 * find_vm_area - find a continuous kernel virtual area
2107 * @addr: base address
2109 * Search for the kernel VM area starting at @addr, and return it.
2110 * It is up to the caller to do all required locking to keep the returned
2113 * Return: pointer to the found area or %NULL on faulure
2115 struct vm_struct *find_vm_area(const void *addr)
2117 struct vmap_area *va;
2119 va = find_vmap_area((unsigned long)addr);
2127 * remove_vm_area - find and remove a continuous kernel virtual area
2128 * @addr: base address
2130 * Search for the kernel VM area starting at @addr, and remove it.
2131 * This function returns the found VM area, but using it is NOT safe
2132 * on SMP machines, except for its size or flags.
2134 * Return: pointer to the found area or %NULL on faulure
2136 struct vm_struct *remove_vm_area(const void *addr)
2138 struct vmap_area *va;
2142 spin_lock(&vmap_area_lock);
2143 va = __find_vmap_area((unsigned long)addr);
2145 struct vm_struct *vm = va->vm;
2148 spin_unlock(&vmap_area_lock);
2150 kasan_free_shadow(vm);
2151 free_unmap_vmap_area(va);
2156 spin_unlock(&vmap_area_lock);
2160 static inline void set_area_direct_map(const struct vm_struct *area,
2161 int (*set_direct_map)(struct page *page))
2165 for (i = 0; i < area->nr_pages; i++)
2166 if (page_address(area->pages[i]))
2167 set_direct_map(area->pages[i]);
2170 /* Handle removing and resetting vm mappings related to the vm_struct. */
2171 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2173 unsigned long start = ULONG_MAX, end = 0;
2174 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2178 remove_vm_area(area->addr);
2180 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2185 * If not deallocating pages, just do the flush of the VM area and
2188 if (!deallocate_pages) {
2194 * If execution gets here, flush the vm mapping and reset the direct
2195 * map. Find the start and end range of the direct mappings to make sure
2196 * the vm_unmap_aliases() flush includes the direct map.
2198 for (i = 0; i < area->nr_pages; i++) {
2199 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2201 start = min(addr, start);
2202 end = max(addr + PAGE_SIZE, end);
2208 * Set direct map to something invalid so that it won't be cached if
2209 * there are any accesses after the TLB flush, then flush the TLB and
2210 * reset the direct map permissions to the default.
2212 set_area_direct_map(area, set_direct_map_invalid_noflush);
2213 _vm_unmap_aliases(start, end, flush_dmap);
2214 set_area_direct_map(area, set_direct_map_default_noflush);
2217 static void __vunmap(const void *addr, int deallocate_pages)
2219 struct vm_struct *area;
2224 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2228 area = find_vm_area(addr);
2229 if (unlikely(!area)) {
2230 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2235 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2236 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2238 kasan_poison_vmalloc(area->addr, area->size);
2240 vm_remove_mappings(area, deallocate_pages);
2242 if (deallocate_pages) {
2245 for (i = 0; i < area->nr_pages; i++) {
2246 struct page *page = area->pages[i];
2249 __free_pages(page, 0);
2251 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2253 kvfree(area->pages);
2260 static inline void __vfree_deferred(const void *addr)
2263 * Use raw_cpu_ptr() because this can be called from preemptible
2264 * context. Preemption is absolutely fine here, because the llist_add()
2265 * implementation is lockless, so it works even if we are adding to
2266 * another cpu's list. schedule_work() should be fine with this too.
2268 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2270 if (llist_add((struct llist_node *)addr, &p->list))
2271 schedule_work(&p->wq);
2275 * vfree_atomic - release memory allocated by vmalloc()
2276 * @addr: memory base address
2278 * This one is just like vfree() but can be called in any atomic context
2281 void vfree_atomic(const void *addr)
2285 kmemleak_free(addr);
2289 __vfree_deferred(addr);
2292 static void __vfree(const void *addr)
2294 if (unlikely(in_interrupt()))
2295 __vfree_deferred(addr);
2301 * vfree - release memory allocated by vmalloc()
2302 * @addr: memory base address
2304 * Free the virtually continuous memory area starting at @addr, as
2305 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2306 * NULL, no operation is performed.
2308 * Must not be called in NMI context (strictly speaking, only if we don't
2309 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2310 * conventions for vfree() arch-depenedent would be a really bad idea)
2312 * May sleep if called *not* from interrupt context.
2314 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2316 void vfree(const void *addr)
2320 kmemleak_free(addr);
2322 might_sleep_if(!in_interrupt());
2329 EXPORT_SYMBOL(vfree);
2332 * vunmap - release virtual mapping obtained by vmap()
2333 * @addr: memory base address
2335 * Free the virtually contiguous memory area starting at @addr,
2336 * which was created from the page array passed to vmap().
2338 * Must not be called in interrupt context.
2340 void vunmap(const void *addr)
2342 BUG_ON(in_interrupt());
2347 EXPORT_SYMBOL(vunmap);
2350 * vmap - map an array of pages into virtually contiguous space
2351 * @pages: array of page pointers
2352 * @count: number of pages to map
2353 * @flags: vm_area->flags
2354 * @prot: page protection for the mapping
2356 * Maps @count pages from @pages into contiguous kernel virtual
2359 * Return: the address of the area or %NULL on failure
2361 void *vmap(struct page **pages, unsigned int count,
2362 unsigned long flags, pgprot_t prot)
2364 struct vm_struct *area;
2365 unsigned long size; /* In bytes */
2369 if (count > totalram_pages())
2372 size = (unsigned long)count << PAGE_SHIFT;
2373 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2377 if (map_kernel_range((unsigned long)area->addr, size, pgprot_nx(prot),
2385 EXPORT_SYMBOL(vmap);
2387 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2388 pgprot_t prot, int node)
2390 struct page **pages;
2391 unsigned int nr_pages, array_size, i;
2392 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2393 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2394 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2398 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2399 array_size = (nr_pages * sizeof(struct page *));
2401 /* Please note that the recursion is strictly bounded. */
2402 if (array_size > PAGE_SIZE) {
2403 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2404 node, area->caller);
2406 pages = kmalloc_node(array_size, nested_gfp, node);
2410 remove_vm_area(area->addr);
2415 area->pages = pages;
2416 area->nr_pages = nr_pages;
2418 for (i = 0; i < area->nr_pages; i++) {
2421 if (node == NUMA_NO_NODE)
2422 page = alloc_page(alloc_mask|highmem_mask);
2424 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2426 if (unlikely(!page)) {
2427 /* Successfully allocated i pages, free them in __vunmap() */
2429 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2432 area->pages[i] = page;
2433 if (gfpflags_allow_blocking(gfp_mask))
2436 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2438 if (map_kernel_range((unsigned long)area->addr, get_vm_area_size(area),
2445 warn_alloc(gfp_mask, NULL,
2446 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2447 (area->nr_pages*PAGE_SIZE), area->size);
2448 __vfree(area->addr);
2453 * __vmalloc_node_range - allocate virtually contiguous memory
2454 * @size: allocation size
2455 * @align: desired alignment
2456 * @start: vm area range start
2457 * @end: vm area range end
2458 * @gfp_mask: flags for the page level allocator
2459 * @prot: protection mask for the allocated pages
2460 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2461 * @node: node to use for allocation or NUMA_NO_NODE
2462 * @caller: caller's return address
2464 * Allocate enough pages to cover @size from the page level
2465 * allocator with @gfp_mask flags. Map them into contiguous
2466 * kernel virtual space, using a pagetable protection of @prot.
2468 * Return: the address of the area or %NULL on failure
2470 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2471 unsigned long start, unsigned long end, gfp_t gfp_mask,
2472 pgprot_t prot, unsigned long vm_flags, int node,
2475 struct vm_struct *area;
2477 unsigned long real_size = size;
2479 size = PAGE_ALIGN(size);
2480 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2483 area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED |
2484 vm_flags, start, end, node, gfp_mask, caller);
2488 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2493 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2494 * flag. It means that vm_struct is not fully initialized.
2495 * Now, it is fully initialized, so remove this flag here.
2497 clear_vm_uninitialized_flag(area);
2499 kmemleak_vmalloc(area, size, gfp_mask);
2504 warn_alloc(gfp_mask, NULL,
2505 "vmalloc: allocation failure: %lu bytes", real_size);
2510 * __vmalloc_node - allocate virtually contiguous memory
2511 * @size: allocation size
2512 * @align: desired alignment
2513 * @gfp_mask: flags for the page level allocator
2514 * @node: node to use for allocation or NUMA_NO_NODE
2515 * @caller: caller's return address
2517 * Allocate enough pages to cover @size from the page level allocator with
2518 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2520 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2521 * and __GFP_NOFAIL are not supported
2523 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2526 * Return: pointer to the allocated memory or %NULL on error
2528 void *__vmalloc_node(unsigned long size, unsigned long align,
2529 gfp_t gfp_mask, int node, const void *caller)
2531 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2532 gfp_mask, PAGE_KERNEL, 0, node, caller);
2535 * This is only for performance analysis of vmalloc and stress purpose.
2536 * It is required by vmalloc test module, therefore do not use it other
2539 #ifdef CONFIG_TEST_VMALLOC_MODULE
2540 EXPORT_SYMBOL_GPL(__vmalloc_node);
2543 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
2545 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
2546 __builtin_return_address(0));
2548 EXPORT_SYMBOL(__vmalloc);
2551 * vmalloc - allocate virtually contiguous memory
2552 * @size: allocation size
2554 * Allocate enough pages to cover @size from the page level
2555 * allocator and map them into contiguous kernel virtual space.
2557 * For tight control over page level allocator and protection flags
2558 * use __vmalloc() instead.
2560 * Return: pointer to the allocated memory or %NULL on error
2562 void *vmalloc(unsigned long size)
2564 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
2565 __builtin_return_address(0));
2567 EXPORT_SYMBOL(vmalloc);
2570 * vzalloc - allocate virtually contiguous memory with zero fill
2571 * @size: allocation size
2573 * Allocate enough pages to cover @size from the page level
2574 * allocator and map them into contiguous kernel virtual space.
2575 * The memory allocated is set to zero.
2577 * For tight control over page level allocator and protection flags
2578 * use __vmalloc() instead.
2580 * Return: pointer to the allocated memory or %NULL on error
2582 void *vzalloc(unsigned long size)
2584 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
2585 __builtin_return_address(0));
2587 EXPORT_SYMBOL(vzalloc);
2590 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2591 * @size: allocation size
2593 * The resulting memory area is zeroed so it can be mapped to userspace
2594 * without leaking data.
2596 * Return: pointer to the allocated memory or %NULL on error
2598 void *vmalloc_user(unsigned long size)
2600 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2601 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2602 VM_USERMAP, NUMA_NO_NODE,
2603 __builtin_return_address(0));
2605 EXPORT_SYMBOL(vmalloc_user);
2608 * vmalloc_node - allocate memory on a specific node
2609 * @size: allocation size
2612 * Allocate enough pages to cover @size from the page level
2613 * allocator and map them into contiguous kernel virtual space.
2615 * For tight control over page level allocator and protection flags
2616 * use __vmalloc() instead.
2618 * Return: pointer to the allocated memory or %NULL on error
2620 void *vmalloc_node(unsigned long size, int node)
2622 return __vmalloc_node(size, 1, GFP_KERNEL, node,
2623 __builtin_return_address(0));
2625 EXPORT_SYMBOL(vmalloc_node);
2628 * vzalloc_node - allocate memory on a specific node with zero fill
2629 * @size: allocation size
2632 * Allocate enough pages to cover @size from the page level
2633 * allocator and map them into contiguous kernel virtual space.
2634 * The memory allocated is set to zero.
2636 * Return: pointer to the allocated memory or %NULL on error
2638 void *vzalloc_node(unsigned long size, int node)
2640 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
2641 __builtin_return_address(0));
2643 EXPORT_SYMBOL(vzalloc_node);
2645 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2646 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2647 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2648 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2651 * 64b systems should always have either DMA or DMA32 zones. For others
2652 * GFP_DMA32 should do the right thing and use the normal zone.
2654 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2658 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2659 * @size: allocation size
2661 * Allocate enough 32bit PA addressable pages to cover @size from the
2662 * page level allocator and map them into contiguous kernel virtual space.
2664 * Return: pointer to the allocated memory or %NULL on error
2666 void *vmalloc_32(unsigned long size)
2668 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
2669 __builtin_return_address(0));
2671 EXPORT_SYMBOL(vmalloc_32);
2674 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2675 * @size: allocation size
2677 * The resulting memory area is 32bit addressable and zeroed so it can be
2678 * mapped to userspace without leaking data.
2680 * Return: pointer to the allocated memory or %NULL on error
2682 void *vmalloc_32_user(unsigned long size)
2684 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2685 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2686 VM_USERMAP, NUMA_NO_NODE,
2687 __builtin_return_address(0));
2689 EXPORT_SYMBOL(vmalloc_32_user);
2692 * small helper routine , copy contents to buf from addr.
2693 * If the page is not present, fill zero.
2696 static int aligned_vread(char *buf, char *addr, unsigned long count)
2702 unsigned long offset, length;
2704 offset = offset_in_page(addr);
2705 length = PAGE_SIZE - offset;
2708 p = vmalloc_to_page(addr);
2710 * To do safe access to this _mapped_ area, we need
2711 * lock. But adding lock here means that we need to add
2712 * overhead of vmalloc()/vfree() calles for this _debug_
2713 * interface, rarely used. Instead of that, we'll use
2714 * kmap() and get small overhead in this access function.
2718 * we can expect USER0 is not used (see vread/vwrite's
2719 * function description)
2721 void *map = kmap_atomic(p);
2722 memcpy(buf, map + offset, length);
2725 memset(buf, 0, length);
2735 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2741 unsigned long offset, length;
2743 offset = offset_in_page(addr);
2744 length = PAGE_SIZE - offset;
2747 p = vmalloc_to_page(addr);
2749 * To do safe access to this _mapped_ area, we need
2750 * lock. But adding lock here means that we need to add
2751 * overhead of vmalloc()/vfree() calles for this _debug_
2752 * interface, rarely used. Instead of that, we'll use
2753 * kmap() and get small overhead in this access function.
2757 * we can expect USER0 is not used (see vread/vwrite's
2758 * function description)
2760 void *map = kmap_atomic(p);
2761 memcpy(map + offset, buf, length);
2773 * vread() - read vmalloc area in a safe way.
2774 * @buf: buffer for reading data
2775 * @addr: vm address.
2776 * @count: number of bytes to be read.
2778 * This function checks that addr is a valid vmalloc'ed area, and
2779 * copy data from that area to a given buffer. If the given memory range
2780 * of [addr...addr+count) includes some valid address, data is copied to
2781 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2782 * IOREMAP area is treated as memory hole and no copy is done.
2784 * If [addr...addr+count) doesn't includes any intersects with alive
2785 * vm_struct area, returns 0. @buf should be kernel's buffer.
2787 * Note: In usual ops, vread() is never necessary because the caller
2788 * should know vmalloc() area is valid and can use memcpy().
2789 * This is for routines which have to access vmalloc area without
2790 * any information, as /dev/kmem.
2792 * Return: number of bytes for which addr and buf should be increased
2793 * (same number as @count) or %0 if [addr...addr+count) doesn't
2794 * include any intersection with valid vmalloc area
2796 long vread(char *buf, char *addr, unsigned long count)
2798 struct vmap_area *va;
2799 struct vm_struct *vm;
2800 char *vaddr, *buf_start = buf;
2801 unsigned long buflen = count;
2804 /* Don't allow overflow */
2805 if ((unsigned long) addr + count < count)
2806 count = -(unsigned long) addr;
2808 spin_lock(&vmap_area_lock);
2809 list_for_each_entry(va, &vmap_area_list, list) {
2817 vaddr = (char *) vm->addr;
2818 if (addr >= vaddr + get_vm_area_size(vm))
2820 while (addr < vaddr) {
2828 n = vaddr + get_vm_area_size(vm) - addr;
2831 if (!(vm->flags & VM_IOREMAP))
2832 aligned_vread(buf, addr, n);
2833 else /* IOREMAP area is treated as memory hole */
2840 spin_unlock(&vmap_area_lock);
2842 if (buf == buf_start)
2844 /* zero-fill memory holes */
2845 if (buf != buf_start + buflen)
2846 memset(buf, 0, buflen - (buf - buf_start));
2852 * vwrite() - write vmalloc area in a safe way.
2853 * @buf: buffer for source data
2854 * @addr: vm address.
2855 * @count: number of bytes to be read.
2857 * This function checks that addr is a valid vmalloc'ed area, and
2858 * copy data from a buffer to the given addr. If specified range of
2859 * [addr...addr+count) includes some valid address, data is copied from
2860 * proper area of @buf. If there are memory holes, no copy to hole.
2861 * IOREMAP area is treated as memory hole and no copy is done.
2863 * If [addr...addr+count) doesn't includes any intersects with alive
2864 * vm_struct area, returns 0. @buf should be kernel's buffer.
2866 * Note: In usual ops, vwrite() is never necessary because the caller
2867 * should know vmalloc() area is valid and can use memcpy().
2868 * This is for routines which have to access vmalloc area without
2869 * any information, as /dev/kmem.
2871 * Return: number of bytes for which addr and buf should be
2872 * increased (same number as @count) or %0 if [addr...addr+count)
2873 * doesn't include any intersection with valid vmalloc area
2875 long vwrite(char *buf, char *addr, unsigned long count)
2877 struct vmap_area *va;
2878 struct vm_struct *vm;
2880 unsigned long n, buflen;
2883 /* Don't allow overflow */
2884 if ((unsigned long) addr + count < count)
2885 count = -(unsigned long) addr;
2888 spin_lock(&vmap_area_lock);
2889 list_for_each_entry(va, &vmap_area_list, list) {
2897 vaddr = (char *) vm->addr;
2898 if (addr >= vaddr + get_vm_area_size(vm))
2900 while (addr < vaddr) {
2907 n = vaddr + get_vm_area_size(vm) - addr;
2910 if (!(vm->flags & VM_IOREMAP)) {
2911 aligned_vwrite(buf, addr, n);
2919 spin_unlock(&vmap_area_lock);
2926 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2927 * @vma: vma to cover
2928 * @uaddr: target user address to start at
2929 * @kaddr: virtual address of vmalloc kernel memory
2930 * @pgoff: offset from @kaddr to start at
2931 * @size: size of map area
2933 * Returns: 0 for success, -Exxx on failure
2935 * This function checks that @kaddr is a valid vmalloc'ed area,
2936 * and that it is big enough to cover the range starting at
2937 * @uaddr in @vma. Will return failure if that criteria isn't
2940 * Similar to remap_pfn_range() (see mm/memory.c)
2942 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2943 void *kaddr, unsigned long pgoff,
2946 struct vm_struct *area;
2948 unsigned long end_index;
2950 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
2953 size = PAGE_ALIGN(size);
2955 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2958 area = find_vm_area(kaddr);
2962 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
2965 if (check_add_overflow(size, off, &end_index) ||
2966 end_index > get_vm_area_size(area))
2971 struct page *page = vmalloc_to_page(kaddr);
2974 ret = vm_insert_page(vma, uaddr, page);
2983 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2987 EXPORT_SYMBOL(remap_vmalloc_range_partial);
2990 * remap_vmalloc_range - map vmalloc pages to userspace
2991 * @vma: vma to cover (map full range of vma)
2992 * @addr: vmalloc memory
2993 * @pgoff: number of pages into addr before first page to map
2995 * Returns: 0 for success, -Exxx on failure
2997 * This function checks that addr is a valid vmalloc'ed area, and
2998 * that it is big enough to cover the vma. Will return failure if
2999 * that criteria isn't met.
3001 * Similar to remap_pfn_range() (see mm/memory.c)
3003 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3004 unsigned long pgoff)
3006 return remap_vmalloc_range_partial(vma, vma->vm_start,
3008 vma->vm_end - vma->vm_start);
3010 EXPORT_SYMBOL(remap_vmalloc_range);
3012 static int f(pte_t *pte, unsigned long addr, void *data)
3024 * alloc_vm_area - allocate a range of kernel address space
3025 * @size: size of the area
3026 * @ptes: returns the PTEs for the address space
3028 * Returns: NULL on failure, vm_struct on success
3030 * This function reserves a range of kernel address space, and
3031 * allocates pagetables to map that range. No actual mappings
3034 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3035 * allocated for the VM area are returned.
3037 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3039 struct vm_struct *area;
3041 area = get_vm_area_caller(size, VM_IOREMAP,
3042 __builtin_return_address(0));
3047 * This ensures that page tables are constructed for this region
3048 * of kernel virtual address space and mapped into init_mm.
3050 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3051 size, f, ptes ? &ptes : NULL)) {
3058 EXPORT_SYMBOL_GPL(alloc_vm_area);
3060 void free_vm_area(struct vm_struct *area)
3062 struct vm_struct *ret;
3063 ret = remove_vm_area(area->addr);
3064 BUG_ON(ret != area);
3067 EXPORT_SYMBOL_GPL(free_vm_area);
3070 static struct vmap_area *node_to_va(struct rb_node *n)
3072 return rb_entry_safe(n, struct vmap_area, rb_node);
3076 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3077 * @addr: target address
3079 * Returns: vmap_area if it is found. If there is no such area
3080 * the first highest(reverse order) vmap_area is returned
3081 * i.e. va->va_start < addr && va->va_end < addr or NULL
3082 * if there are no any areas before @addr.
3084 static struct vmap_area *
3085 pvm_find_va_enclose_addr(unsigned long addr)
3087 struct vmap_area *va, *tmp;
3090 n = free_vmap_area_root.rb_node;
3094 tmp = rb_entry(n, struct vmap_area, rb_node);
3095 if (tmp->va_start <= addr) {
3097 if (tmp->va_end >= addr)
3110 * pvm_determine_end_from_reverse - find the highest aligned address
3111 * of free block below VMALLOC_END
3113 * in - the VA we start the search(reverse order);
3114 * out - the VA with the highest aligned end address.
3116 * Returns: determined end address within vmap_area
3118 static unsigned long
3119 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3121 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3125 list_for_each_entry_from_reverse((*va),
3126 &free_vmap_area_list, list) {
3127 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3128 if ((*va)->va_start < addr)
3137 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3138 * @offsets: array containing offset of each area
3139 * @sizes: array containing size of each area
3140 * @nr_vms: the number of areas to allocate
3141 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3143 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3144 * vm_structs on success, %NULL on failure
3146 * Percpu allocator wants to use congruent vm areas so that it can
3147 * maintain the offsets among percpu areas. This function allocates
3148 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3149 * be scattered pretty far, distance between two areas easily going up
3150 * to gigabytes. To avoid interacting with regular vmallocs, these
3151 * areas are allocated from top.
3153 * Despite its complicated look, this allocator is rather simple. It
3154 * does everything top-down and scans free blocks from the end looking
3155 * for matching base. While scanning, if any of the areas do not fit the
3156 * base address is pulled down to fit the area. Scanning is repeated till
3157 * all the areas fit and then all necessary data structures are inserted
3158 * and the result is returned.
3160 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3161 const size_t *sizes, int nr_vms,
3164 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3165 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3166 struct vmap_area **vas, *va;
3167 struct vm_struct **vms;
3168 int area, area2, last_area, term_area;
3169 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3170 bool purged = false;
3173 /* verify parameters and allocate data structures */
3174 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3175 for (last_area = 0, area = 0; area < nr_vms; area++) {
3176 start = offsets[area];
3177 end = start + sizes[area];
3179 /* is everything aligned properly? */
3180 BUG_ON(!IS_ALIGNED(offsets[area], align));
3181 BUG_ON(!IS_ALIGNED(sizes[area], align));
3183 /* detect the area with the highest address */
3184 if (start > offsets[last_area])
3187 for (area2 = area + 1; area2 < nr_vms; area2++) {
3188 unsigned long start2 = offsets[area2];
3189 unsigned long end2 = start2 + sizes[area2];
3191 BUG_ON(start2 < end && start < end2);
3194 last_end = offsets[last_area] + sizes[last_area];
3196 if (vmalloc_end - vmalloc_start < last_end) {
3201 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3202 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3206 for (area = 0; area < nr_vms; area++) {
3207 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3208 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3209 if (!vas[area] || !vms[area])
3213 spin_lock(&free_vmap_area_lock);
3215 /* start scanning - we scan from the top, begin with the last area */
3216 area = term_area = last_area;
3217 start = offsets[area];
3218 end = start + sizes[area];
3220 va = pvm_find_va_enclose_addr(vmalloc_end);
3221 base = pvm_determine_end_from_reverse(&va, align) - end;
3225 * base might have underflowed, add last_end before
3228 if (base + last_end < vmalloc_start + last_end)
3232 * Fitting base has not been found.
3238 * If required width exceeds current VA block, move
3239 * base downwards and then recheck.
3241 if (base + end > va->va_end) {
3242 base = pvm_determine_end_from_reverse(&va, align) - end;
3248 * If this VA does not fit, move base downwards and recheck.
3250 if (base + start < va->va_start) {
3251 va = node_to_va(rb_prev(&va->rb_node));
3252 base = pvm_determine_end_from_reverse(&va, align) - end;
3258 * This area fits, move on to the previous one. If
3259 * the previous one is the terminal one, we're done.
3261 area = (area + nr_vms - 1) % nr_vms;
3262 if (area == term_area)
3265 start = offsets[area];
3266 end = start + sizes[area];
3267 va = pvm_find_va_enclose_addr(base + end);
3270 /* we've found a fitting base, insert all va's */
3271 for (area = 0; area < nr_vms; area++) {
3274 start = base + offsets[area];
3277 va = pvm_find_va_enclose_addr(start);
3278 if (WARN_ON_ONCE(va == NULL))
3279 /* It is a BUG(), but trigger recovery instead. */
3282 type = classify_va_fit_type(va, start, size);
3283 if (WARN_ON_ONCE(type == NOTHING_FIT))
3284 /* It is a BUG(), but trigger recovery instead. */
3287 ret = adjust_va_to_fit_type(va, start, size, type);
3291 /* Allocated area. */
3293 va->va_start = start;
3294 va->va_end = start + size;
3297 spin_unlock(&free_vmap_area_lock);
3299 /* populate the kasan shadow space */
3300 for (area = 0; area < nr_vms; area++) {
3301 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3302 goto err_free_shadow;
3304 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3308 /* insert all vm's */
3309 spin_lock(&vmap_area_lock);
3310 for (area = 0; area < nr_vms; area++) {
3311 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3313 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3316 spin_unlock(&vmap_area_lock);
3323 * Remove previously allocated areas. There is no
3324 * need in removing these areas from the busy tree,
3325 * because they are inserted only on the final step
3326 * and when pcpu_get_vm_areas() is success.
3329 orig_start = vas[area]->va_start;
3330 orig_end = vas[area]->va_end;
3331 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3332 &free_vmap_area_list);
3333 kasan_release_vmalloc(orig_start, orig_end,
3334 va->va_start, va->va_end);
3339 spin_unlock(&free_vmap_area_lock);
3341 purge_vmap_area_lazy();
3344 /* Before "retry", check if we recover. */
3345 for (area = 0; area < nr_vms; area++) {
3349 vas[area] = kmem_cache_zalloc(
3350 vmap_area_cachep, GFP_KERNEL);
3359 for (area = 0; area < nr_vms; area++) {
3361 kmem_cache_free(vmap_area_cachep, vas[area]);
3371 spin_lock(&free_vmap_area_lock);
3373 * We release all the vmalloc shadows, even the ones for regions that
3374 * hadn't been successfully added. This relies on kasan_release_vmalloc
3375 * being able to tolerate this case.
3377 for (area = 0; area < nr_vms; area++) {
3378 orig_start = vas[area]->va_start;
3379 orig_end = vas[area]->va_end;
3380 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3381 &free_vmap_area_list);
3382 kasan_release_vmalloc(orig_start, orig_end,
3383 va->va_start, va->va_end);
3387 spin_unlock(&free_vmap_area_lock);
3394 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3395 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3396 * @nr_vms: the number of allocated areas
3398 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3400 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3404 for (i = 0; i < nr_vms; i++)
3405 free_vm_area(vms[i]);
3408 #endif /* CONFIG_SMP */
3410 #ifdef CONFIG_PROC_FS
3411 static void *s_start(struct seq_file *m, loff_t *pos)
3412 __acquires(&vmap_purge_lock)
3413 __acquires(&vmap_area_lock)
3415 mutex_lock(&vmap_purge_lock);
3416 spin_lock(&vmap_area_lock);
3418 return seq_list_start(&vmap_area_list, *pos);
3421 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3423 return seq_list_next(p, &vmap_area_list, pos);
3426 static void s_stop(struct seq_file *m, void *p)
3427 __releases(&vmap_purge_lock)
3428 __releases(&vmap_area_lock)
3430 mutex_unlock(&vmap_purge_lock);
3431 spin_unlock(&vmap_area_lock);
3434 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3436 if (IS_ENABLED(CONFIG_NUMA)) {
3437 unsigned int nr, *counters = m->private;
3442 if (v->flags & VM_UNINITIALIZED)
3444 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3447 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3449 for (nr = 0; nr < v->nr_pages; nr++)
3450 counters[page_to_nid(v->pages[nr])]++;
3452 for_each_node_state(nr, N_HIGH_MEMORY)
3454 seq_printf(m, " N%u=%u", nr, counters[nr]);
3458 static void show_purge_info(struct seq_file *m)
3460 struct llist_node *head;
3461 struct vmap_area *va;
3463 head = READ_ONCE(vmap_purge_list.first);
3467 llist_for_each_entry(va, head, purge_list) {
3468 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3469 (void *)va->va_start, (void *)va->va_end,
3470 va->va_end - va->va_start);
3474 static int s_show(struct seq_file *m, void *p)
3476 struct vmap_area *va;
3477 struct vm_struct *v;
3479 va = list_entry(p, struct vmap_area, list);
3482 * s_show can encounter race with remove_vm_area, !vm on behalf
3483 * of vmap area is being tear down or vm_map_ram allocation.
3486 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3487 (void *)va->va_start, (void *)va->va_end,
3488 va->va_end - va->va_start);
3495 seq_printf(m, "0x%pK-0x%pK %7ld",
3496 v->addr, v->addr + v->size, v->size);
3499 seq_printf(m, " %pS", v->caller);
3502 seq_printf(m, " pages=%d", v->nr_pages);
3505 seq_printf(m, " phys=%pa", &v->phys_addr);
3507 if (v->flags & VM_IOREMAP)
3508 seq_puts(m, " ioremap");
3510 if (v->flags & VM_ALLOC)
3511 seq_puts(m, " vmalloc");
3513 if (v->flags & VM_MAP)
3514 seq_puts(m, " vmap");
3516 if (v->flags & VM_USERMAP)
3517 seq_puts(m, " user");
3519 if (v->flags & VM_DMA_COHERENT)
3520 seq_puts(m, " dma-coherent");
3522 if (is_vmalloc_addr(v->pages))
3523 seq_puts(m, " vpages");
3525 show_numa_info(m, v);
3529 * As a final step, dump "unpurged" areas. Note,
3530 * that entire "/proc/vmallocinfo" output will not
3531 * be address sorted, because the purge list is not
3534 if (list_is_last(&va->list, &vmap_area_list))
3540 static const struct seq_operations vmalloc_op = {
3547 static int __init proc_vmalloc_init(void)
3549 if (IS_ENABLED(CONFIG_NUMA))
3550 proc_create_seq_private("vmallocinfo", 0400, NULL,
3552 nr_node_ids * sizeof(unsigned int), NULL);
3554 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3557 module_init(proc_vmalloc_init);