4 * Copyright (C) 1993 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8 * Numa awareness, Christoph Lameter, SGI, June 2005
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/radix-tree.h>
28 #include <linux/rcupdate.h>
29 #include <linux/pfn.h>
30 #include <linux/kmemleak.h>
31 #include <linux/atomic.h>
32 #include <linux/compiler.h>
33 #include <linux/llist.h>
34 #include <linux/bitops.h>
35 #include <linux/rbtree_augmented.h>
37 #include <linux/uaccess.h>
38 #include <asm/tlbflush.h>
39 #include <asm/shmparam.h>
43 struct vfree_deferred {
44 struct llist_head list;
45 struct work_struct wq;
47 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
49 static void __vunmap(const void *, int);
51 static void free_work(struct work_struct *w)
53 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
54 struct llist_node *t, *llnode;
56 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
57 __vunmap((void *)llnode, 1);
60 /*** Page table manipulation functions ***/
62 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
66 pte = pte_offset_kernel(pmd, addr);
68 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
69 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
70 } while (pte++, addr += PAGE_SIZE, addr != end);
73 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
78 pmd = pmd_offset(pud, addr);
80 next = pmd_addr_end(addr, end);
81 if (pmd_clear_huge(pmd))
83 if (pmd_none_or_clear_bad(pmd))
85 vunmap_pte_range(pmd, addr, next);
86 } while (pmd++, addr = next, addr != end);
89 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
94 pud = pud_offset(p4d, addr);
96 next = pud_addr_end(addr, end);
97 if (pud_clear_huge(pud))
99 if (pud_none_or_clear_bad(pud))
101 vunmap_pmd_range(pud, addr, next);
102 } while (pud++, addr = next, addr != end);
105 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
110 p4d = p4d_offset(pgd, addr);
112 next = p4d_addr_end(addr, end);
113 if (p4d_clear_huge(p4d))
115 if (p4d_none_or_clear_bad(p4d))
117 vunmap_pud_range(p4d, addr, next);
118 } while (p4d++, addr = next, addr != end);
121 static void vunmap_page_range(unsigned long addr, unsigned long end)
127 pgd = pgd_offset_k(addr);
129 next = pgd_addr_end(addr, end);
130 if (pgd_none_or_clear_bad(pgd))
132 vunmap_p4d_range(pgd, addr, next);
133 } while (pgd++, addr = next, addr != end);
136 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
137 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
142 * nr is a running index into the array which helps higher level
143 * callers keep track of where we're up to.
146 pte = pte_alloc_kernel(pmd, addr);
150 struct page *page = pages[*nr];
152 if (WARN_ON(!pte_none(*pte)))
156 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
158 } while (pte++, addr += PAGE_SIZE, addr != end);
162 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
163 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
168 pmd = pmd_alloc(&init_mm, pud, addr);
172 next = pmd_addr_end(addr, end);
173 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
175 } while (pmd++, addr = next, addr != end);
179 static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
180 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
185 pud = pud_alloc(&init_mm, p4d, addr);
189 next = pud_addr_end(addr, end);
190 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
192 } while (pud++, addr = next, addr != end);
196 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
197 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
202 p4d = p4d_alloc(&init_mm, pgd, addr);
206 next = p4d_addr_end(addr, end);
207 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
209 } while (p4d++, addr = next, addr != end);
214 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
215 * will have pfns corresponding to the "pages" array.
217 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
219 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
220 pgprot_t prot, struct page **pages)
224 unsigned long addr = start;
229 pgd = pgd_offset_k(addr);
231 next = pgd_addr_end(addr, end);
232 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
235 } while (pgd++, addr = next, addr != end);
240 static int vmap_page_range(unsigned long start, unsigned long end,
241 pgprot_t prot, struct page **pages)
245 ret = vmap_page_range_noflush(start, end, prot, pages);
246 flush_cache_vmap(start, end);
250 int is_vmalloc_or_module_addr(const void *x)
253 * ARM, x86-64 and sparc64 put modules in a special place,
254 * and fall back on vmalloc() if that fails. Others
255 * just put it in the vmalloc space.
257 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
258 unsigned long addr = (unsigned long)x;
259 if (addr >= MODULES_VADDR && addr < MODULES_END)
262 return is_vmalloc_addr(x);
266 * Walk a vmap address to the struct page it maps.
268 struct page *vmalloc_to_page(const void *vmalloc_addr)
270 unsigned long addr = (unsigned long) vmalloc_addr;
271 struct page *page = NULL;
272 pgd_t *pgd = pgd_offset_k(addr);
279 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
280 * architectures that do not vmalloc module space
282 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
286 p4d = p4d_offset(pgd, addr);
289 pud = pud_offset(p4d, addr);
292 * Don't dereference bad PUD or PMD (below) entries. This will also
293 * identify huge mappings, which we may encounter on architectures
294 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
295 * identified as vmalloc addresses by is_vmalloc_addr(), but are
296 * not [unambiguously] associated with a struct page, so there is
297 * no correct value to return for them.
299 WARN_ON_ONCE(pud_bad(*pud));
300 if (pud_none(*pud) || pud_bad(*pud))
302 pmd = pmd_offset(pud, addr);
303 WARN_ON_ONCE(pmd_bad(*pmd));
304 if (pmd_none(*pmd) || pmd_bad(*pmd))
307 ptep = pte_offset_map(pmd, addr);
309 if (pte_present(pte))
310 page = pte_page(pte);
314 EXPORT_SYMBOL(vmalloc_to_page);
317 * Map a vmalloc()-space virtual address to the physical page frame number.
319 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
321 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
323 EXPORT_SYMBOL(vmalloc_to_pfn);
326 /*** Global kva allocator ***/
328 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
329 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
331 #define VM_LAZY_FREE 0x02
332 #define VM_VM_AREA 0x04
334 static DEFINE_SPINLOCK(vmap_area_lock);
335 /* Export for kexec only */
336 LIST_HEAD(vmap_area_list);
337 static LLIST_HEAD(vmap_purge_list);
338 static struct rb_root vmap_area_root = RB_ROOT;
339 static bool vmap_initialized __read_mostly;
342 * This kmem_cache is used for vmap_area objects. Instead of
343 * allocating from slab we reuse an object from this cache to
344 * make things faster. Especially in "no edge" splitting of
347 static struct kmem_cache *vmap_area_cachep;
350 * This linked list is used in pair with free_vmap_area_root.
351 * It gives O(1) access to prev/next to perform fast coalescing.
353 static LIST_HEAD(free_vmap_area_list);
356 * This augment red-black tree represents the free vmap space.
357 * All vmap_area objects in this tree are sorted by va->va_start
358 * address. It is used for allocation and merging when a vmap
359 * object is released.
361 * Each vmap_area node contains a maximum available free block
362 * of its sub-tree, right or left. Therefore it is possible to
363 * find a lowest match of free area.
365 static struct rb_root free_vmap_area_root = RB_ROOT;
367 static __always_inline unsigned long
368 va_size(struct vmap_area *va)
370 return (va->va_end - va->va_start);
373 static __always_inline unsigned long
374 get_subtree_max_size(struct rb_node *node)
376 struct vmap_area *va;
378 va = rb_entry_safe(node, struct vmap_area, rb_node);
379 return va ? va->subtree_max_size : 0;
383 * Gets called when remove the node and rotate.
385 static __always_inline unsigned long
386 compute_subtree_max_size(struct vmap_area *va)
388 return max3(va_size(va),
389 get_subtree_max_size(va->rb_node.rb_left),
390 get_subtree_max_size(va->rb_node.rb_right));
393 RB_DECLARE_CALLBACKS(static, free_vmap_area_rb_augment_cb,
394 struct vmap_area, rb_node, unsigned long, subtree_max_size,
395 compute_subtree_max_size)
397 static void purge_vmap_area_lazy(void);
398 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
399 static unsigned long lazy_max_pages(void);
401 static struct vmap_area *__find_vmap_area(unsigned long addr)
403 struct rb_node *n = vmap_area_root.rb_node;
406 struct vmap_area *va;
408 va = rb_entry(n, struct vmap_area, rb_node);
409 if (addr < va->va_start)
411 else if (addr >= va->va_end)
421 * This function returns back addresses of parent node
422 * and its left or right link for further processing.
424 static __always_inline struct rb_node **
425 find_va_links(struct vmap_area *va,
426 struct rb_root *root, struct rb_node *from,
427 struct rb_node **parent)
429 struct vmap_area *tmp_va;
430 struct rb_node **link;
433 link = &root->rb_node;
434 if (unlikely(!*link)) {
443 * Go to the bottom of the tree. When we hit the last point
444 * we end up with parent rb_node and correct direction, i name
445 * it link, where the new va->rb_node will be attached to.
448 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
451 * During the traversal we also do some sanity check.
452 * Trigger the BUG() if there are sides(left/right)
455 if (va->va_start < tmp_va->va_end &&
456 va->va_end <= tmp_va->va_start)
457 link = &(*link)->rb_left;
458 else if (va->va_end > tmp_va->va_start &&
459 va->va_start >= tmp_va->va_end)
460 link = &(*link)->rb_right;
465 *parent = &tmp_va->rb_node;
469 static __always_inline struct list_head *
470 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
472 struct list_head *list;
474 if (unlikely(!parent))
476 * The red-black tree where we try to find VA neighbors
477 * before merging or inserting is empty, i.e. it means
478 * there is no free vmap space. Normally it does not
479 * happen but we handle this case anyway.
483 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
484 return (&parent->rb_right == link ? list->next : list);
487 static __always_inline void
488 link_va(struct vmap_area *va, struct rb_root *root,
489 struct rb_node *parent, struct rb_node **link, struct list_head *head)
492 * VA is still not in the list, but we can
493 * identify its future previous list_head node.
495 if (likely(parent)) {
496 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
497 if (&parent->rb_right != link)
501 /* Insert to the rb-tree */
502 rb_link_node(&va->rb_node, parent, link);
503 if (root == &free_vmap_area_root) {
505 * Some explanation here. Just perform simple insertion
506 * to the tree. We do not set va->subtree_max_size to
507 * its current size before calling rb_insert_augmented().
508 * It is because of we populate the tree from the bottom
509 * to parent levels when the node _is_ in the tree.
511 * Therefore we set subtree_max_size to zero after insertion,
512 * to let __augment_tree_propagate_from() puts everything to
513 * the correct order later on.
515 rb_insert_augmented(&va->rb_node,
516 root, &free_vmap_area_rb_augment_cb);
517 va->subtree_max_size = 0;
519 rb_insert_color(&va->rb_node, root);
522 /* Address-sort this list */
523 list_add(&va->list, head);
526 static __always_inline void
527 unlink_va(struct vmap_area *va, struct rb_root *root)
530 * During merging a VA node can be empty, therefore
531 * not linked with the tree nor list. Just check it.
533 if (!RB_EMPTY_NODE(&va->rb_node)) {
534 if (root == &free_vmap_area_root)
535 rb_erase_augmented(&va->rb_node,
536 root, &free_vmap_area_rb_augment_cb);
538 rb_erase(&va->rb_node, root);
541 RB_CLEAR_NODE(&va->rb_node);
545 #if DEBUG_AUGMENT_PROPAGATE_CHECK
547 augment_tree_propagate_check(struct rb_node *n)
549 struct vmap_area *va;
550 struct rb_node *node;
557 va = rb_entry(n, struct vmap_area, rb_node);
558 size = va->subtree_max_size;
562 va = rb_entry(node, struct vmap_area, rb_node);
564 if (get_subtree_max_size(node->rb_left) == size) {
565 node = node->rb_left;
567 if (va_size(va) == size) {
572 node = node->rb_right;
577 va = rb_entry(n, struct vmap_area, rb_node);
578 pr_emerg("tree is corrupted: %lu, %lu\n",
579 va_size(va), va->subtree_max_size);
582 augment_tree_propagate_check(n->rb_left);
583 augment_tree_propagate_check(n->rb_right);
588 * This function populates subtree_max_size from bottom to upper
589 * levels starting from VA point. The propagation must be done
590 * when VA size is modified by changing its va_start/va_end. Or
591 * in case of newly inserting of VA to the tree.
593 * It means that __augment_tree_propagate_from() must be called:
594 * - After VA has been inserted to the tree(free path);
595 * - After VA has been shrunk(allocation path);
596 * - After VA has been increased(merging path).
598 * Please note that, it does not mean that upper parent nodes
599 * and their subtree_max_size are recalculated all the time up
608 * For example if we modify the node 4, shrinking it to 2, then
609 * no any modification is required. If we shrink the node 2 to 1
610 * its subtree_max_size is updated only, and set to 1. If we shrink
611 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
614 static __always_inline void
615 augment_tree_propagate_from(struct vmap_area *va)
617 struct rb_node *node = &va->rb_node;
618 unsigned long new_va_sub_max_size;
621 va = rb_entry(node, struct vmap_area, rb_node);
622 new_va_sub_max_size = compute_subtree_max_size(va);
625 * If the newly calculated maximum available size of the
626 * subtree is equal to the current one, then it means that
627 * the tree is propagated correctly. So we have to stop at
628 * this point to save cycles.
630 if (va->subtree_max_size == new_va_sub_max_size)
633 va->subtree_max_size = new_va_sub_max_size;
634 node = rb_parent(&va->rb_node);
637 #if DEBUG_AUGMENT_PROPAGATE_CHECK
638 augment_tree_propagate_check(free_vmap_area_root.rb_node);
643 insert_vmap_area(struct vmap_area *va,
644 struct rb_root *root, struct list_head *head)
646 struct rb_node **link;
647 struct rb_node *parent;
649 link = find_va_links(va, root, NULL, &parent);
650 link_va(va, root, parent, link, head);
654 insert_vmap_area_augment(struct vmap_area *va,
655 struct rb_node *from, struct rb_root *root,
656 struct list_head *head)
658 struct rb_node **link;
659 struct rb_node *parent;
662 link = find_va_links(va, NULL, from, &parent);
664 link = find_va_links(va, root, NULL, &parent);
666 link_va(va, root, parent, link, head);
667 augment_tree_propagate_from(va);
671 * Merge de-allocated chunk of VA memory with previous
672 * and next free blocks. If coalesce is not done a new
673 * free area is inserted. If VA has been merged, it is
676 static __always_inline void
677 merge_or_add_vmap_area(struct vmap_area *va,
678 struct rb_root *root, struct list_head *head)
680 struct vmap_area *sibling;
681 struct list_head *next;
682 struct rb_node **link;
683 struct rb_node *parent;
687 * Find a place in the tree where VA potentially will be
688 * inserted, unless it is merged with its sibling/siblings.
690 link = find_va_links(va, root, NULL, &parent);
693 * Get next node of VA to check if merging can be done.
695 next = get_va_next_sibling(parent, link);
696 if (unlikely(next == NULL))
702 * |<------VA------>|<-----Next----->|
707 sibling = list_entry(next, struct vmap_area, list);
708 if (sibling->va_start == va->va_end) {
709 sibling->va_start = va->va_start;
711 /* Check and update the tree if needed. */
712 augment_tree_propagate_from(sibling);
714 /* Remove this VA, it has been merged. */
717 /* Free vmap_area object. */
718 kmem_cache_free(vmap_area_cachep, va);
720 /* Point to the new merged area. */
729 * |<-----Prev----->|<------VA------>|
733 if (next->prev != head) {
734 sibling = list_entry(next->prev, struct vmap_area, list);
735 if (sibling->va_end == va->va_start) {
736 sibling->va_end = va->va_end;
738 /* Check and update the tree if needed. */
739 augment_tree_propagate_from(sibling);
741 /* Remove this VA, it has been merged. */
744 /* Free vmap_area object. */
745 kmem_cache_free(vmap_area_cachep, va);
753 link_va(va, root, parent, link, head);
754 augment_tree_propagate_from(va);
758 static __always_inline bool
759 is_within_this_va(struct vmap_area *va, unsigned long size,
760 unsigned long align, unsigned long vstart)
762 unsigned long nva_start_addr;
764 if (va->va_start > vstart)
765 nva_start_addr = ALIGN(va->va_start, align);
767 nva_start_addr = ALIGN(vstart, align);
769 /* Can be overflowed due to big size or alignment. */
770 if (nva_start_addr + size < nva_start_addr ||
771 nva_start_addr < vstart)
774 return (nva_start_addr + size <= va->va_end);
778 * Find the first free block(lowest start address) in the tree,
779 * that will accomplish the request corresponding to passing
782 static __always_inline struct vmap_area *
783 find_vmap_lowest_match(unsigned long size,
784 unsigned long align, unsigned long vstart)
786 struct vmap_area *va;
787 struct rb_node *node;
788 unsigned long length;
790 /* Start from the root. */
791 node = free_vmap_area_root.rb_node;
793 /* Adjust the search size for alignment overhead. */
794 length = size + align - 1;
797 va = rb_entry(node, struct vmap_area, rb_node);
799 if (get_subtree_max_size(node->rb_left) >= length &&
800 vstart < va->va_start) {
801 node = node->rb_left;
803 if (is_within_this_va(va, size, align, vstart))
807 * Does not make sense to go deeper towards the right
808 * sub-tree if it does not have a free block that is
809 * equal or bigger to the requested search length.
811 if (get_subtree_max_size(node->rb_right) >= length) {
812 node = node->rb_right;
817 * OK. We roll back and find the fist right sub-tree,
818 * that will satisfy the search criteria. It can happen
819 * only once due to "vstart" restriction.
821 while ((node = rb_parent(node))) {
822 va = rb_entry(node, struct vmap_area, rb_node);
823 if (is_within_this_va(va, size, align, vstart))
826 if (get_subtree_max_size(node->rb_right) >= length &&
827 vstart <= va->va_start) {
828 node = node->rb_right;
838 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
839 #include <linux/random.h>
841 static struct vmap_area *
842 find_vmap_lowest_linear_match(unsigned long size,
843 unsigned long align, unsigned long vstart)
845 struct vmap_area *va;
847 list_for_each_entry(va, &free_vmap_area_list, list) {
848 if (!is_within_this_va(va, size, align, vstart))
858 find_vmap_lowest_match_check(unsigned long size)
860 struct vmap_area *va_1, *va_2;
861 unsigned long vstart;
864 get_random_bytes(&rnd, sizeof(rnd));
865 vstart = VMALLOC_START + rnd;
867 va_1 = find_vmap_lowest_match(size, 1, vstart);
868 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
871 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
878 FL_FIT_TYPE = 1, /* full fit */
879 LE_FIT_TYPE = 2, /* left edge fit */
880 RE_FIT_TYPE = 3, /* right edge fit */
881 NE_FIT_TYPE = 4 /* no edge fit */
884 static __always_inline enum fit_type
885 classify_va_fit_type(struct vmap_area *va,
886 unsigned long nva_start_addr, unsigned long size)
890 /* Check if it is within VA. */
891 if (nva_start_addr < va->va_start ||
892 nva_start_addr + size > va->va_end)
896 if (va->va_start == nva_start_addr) {
897 if (va->va_end == nva_start_addr + size)
901 } else if (va->va_end == nva_start_addr + size) {
910 static __always_inline int
911 adjust_va_to_fit_type(struct vmap_area *va,
912 unsigned long nva_start_addr, unsigned long size,
915 struct vmap_area *lva;
917 if (type == FL_FIT_TYPE) {
919 * No need to split VA, it fully fits.
925 unlink_va(va, &free_vmap_area_root);
926 kmem_cache_free(vmap_area_cachep, va);
927 } else if (type == LE_FIT_TYPE) {
929 * Split left edge of fit VA.
935 va->va_start += size;
936 } else if (type == RE_FIT_TYPE) {
938 * Split right edge of fit VA.
944 va->va_end = nva_start_addr;
945 } else if (type == NE_FIT_TYPE) {
947 * Split no edge of fit VA.
953 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
958 * Build the remainder.
960 lva->va_start = va->va_start;
961 lva->va_end = nva_start_addr;
964 * Shrink this VA to remaining size.
966 va->va_start = nva_start_addr + size;
971 if (type != FL_FIT_TYPE) {
972 augment_tree_propagate_from(va);
974 if (type == NE_FIT_TYPE)
975 insert_vmap_area_augment(lva, &va->rb_node,
976 &free_vmap_area_root, &free_vmap_area_list);
983 * Returns a start address of the newly allocated area, if success.
984 * Otherwise a vend is returned that indicates failure.
986 static __always_inline unsigned long
987 __alloc_vmap_area(unsigned long size, unsigned long align,
988 unsigned long vstart, unsigned long vend, int node)
990 unsigned long nva_start_addr;
991 struct vmap_area *va;
995 va = find_vmap_lowest_match(size, align, vstart);
999 if (va->va_start > vstart)
1000 nva_start_addr = ALIGN(va->va_start, align);
1002 nva_start_addr = ALIGN(vstart, align);
1004 /* Check the "vend" restriction. */
1005 if (nva_start_addr + size > vend)
1008 /* Classify what we have found. */
1009 type = classify_va_fit_type(va, nva_start_addr, size);
1010 if (WARN_ON_ONCE(type == NOTHING_FIT))
1013 /* Update the free vmap_area. */
1014 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1018 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1019 find_vmap_lowest_match_check(size);
1022 return nva_start_addr;
1026 * Allocate a region of KVA of the specified size and alignment, within the
1029 static struct vmap_area *alloc_vmap_area(unsigned long size,
1030 unsigned long align,
1031 unsigned long vstart, unsigned long vend,
1032 int node, gfp_t gfp_mask)
1034 struct vmap_area *va;
1039 BUG_ON(offset_in_page(size));
1040 BUG_ON(!is_power_of_2(align));
1042 if (unlikely(!vmap_initialized))
1043 return ERR_PTR(-EBUSY);
1047 va = kmem_cache_alloc_node(vmap_area_cachep,
1048 gfp_mask & GFP_RECLAIM_MASK, node);
1050 return ERR_PTR(-ENOMEM);
1053 * Only scan the relevant parts containing pointers to other objects
1054 * to avoid false negatives.
1056 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
1059 spin_lock(&vmap_area_lock);
1062 * If an allocation fails, the "vend" address is
1063 * returned. Therefore trigger the overflow path.
1065 addr = __alloc_vmap_area(size, align, vstart, vend, node);
1066 if (unlikely(addr == vend))
1069 va->va_start = addr;
1070 va->va_end = addr + size;
1072 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1074 spin_unlock(&vmap_area_lock);
1076 BUG_ON(!IS_ALIGNED(va->va_start, align));
1077 BUG_ON(va->va_start < vstart);
1078 BUG_ON(va->va_end > vend);
1083 spin_unlock(&vmap_area_lock);
1085 purge_vmap_area_lazy();
1090 if (gfpflags_allow_blocking(gfp_mask)) {
1091 unsigned long freed = 0;
1092 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1099 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1100 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1103 kmem_cache_free(vmap_area_cachep, va);
1104 return ERR_PTR(-EBUSY);
1107 int register_vmap_purge_notifier(struct notifier_block *nb)
1109 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1111 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1113 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1115 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1117 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1119 static void __free_vmap_area(struct vmap_area *va)
1121 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
1124 * Remove from the busy tree/list.
1126 unlink_va(va, &vmap_area_root);
1129 * Merge VA with its neighbors, otherwise just add it.
1131 merge_or_add_vmap_area(va,
1132 &free_vmap_area_root, &free_vmap_area_list);
1136 * Free a region of KVA allocated by alloc_vmap_area
1138 static void free_vmap_area(struct vmap_area *va)
1140 spin_lock(&vmap_area_lock);
1141 __free_vmap_area(va);
1142 spin_unlock(&vmap_area_lock);
1146 * Clear the pagetable entries of a given vmap_area
1148 static void unmap_vmap_area(struct vmap_area *va)
1150 vunmap_page_range(va->va_start, va->va_end);
1154 * lazy_max_pages is the maximum amount of virtual address space we gather up
1155 * before attempting to purge with a TLB flush.
1157 * There is a tradeoff here: a larger number will cover more kernel page tables
1158 * and take slightly longer to purge, but it will linearly reduce the number of
1159 * global TLB flushes that must be performed. It would seem natural to scale
1160 * this number up linearly with the number of CPUs (because vmapping activity
1161 * could also scale linearly with the number of CPUs), however it is likely
1162 * that in practice, workloads might be constrained in other ways that mean
1163 * vmap activity will not scale linearly with CPUs. Also, I want to be
1164 * conservative and not introduce a big latency on huge systems, so go with
1165 * a less aggressive log scale. It will still be an improvement over the old
1166 * code, and it will be simple to change the scale factor if we find that it
1167 * becomes a problem on bigger systems.
1169 static unsigned long lazy_max_pages(void)
1173 log = fls(num_online_cpus());
1175 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1178 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1181 * Serialize vmap purging. There is no actual criticial section protected
1182 * by this look, but we want to avoid concurrent calls for performance
1183 * reasons and to make the pcpu_get_vm_areas more deterministic.
1185 static DEFINE_MUTEX(vmap_purge_lock);
1187 /* for per-CPU blocks */
1188 static void purge_fragmented_blocks_allcpus(void);
1191 * called before a call to iounmap() if the caller wants vm_area_struct's
1192 * immediately freed.
1194 void set_iounmap_nonlazy(void)
1196 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1200 * Purges all lazily-freed vmap areas.
1202 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1204 unsigned long resched_threshold;
1205 struct llist_node *valist;
1206 struct vmap_area *va;
1207 struct vmap_area *n_va;
1209 lockdep_assert_held(&vmap_purge_lock);
1211 valist = llist_del_all(&vmap_purge_list);
1212 if (unlikely(valist == NULL))
1216 * TODO: to calculate a flush range without looping.
1217 * The list can be up to lazy_max_pages() elements.
1219 llist_for_each_entry(va, valist, purge_list) {
1220 if (va->va_start < start)
1221 start = va->va_start;
1222 if (va->va_end > end)
1226 flush_tlb_kernel_range(start, end);
1227 resched_threshold = lazy_max_pages() << 1;
1229 spin_lock(&vmap_area_lock);
1230 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1231 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1233 __free_vmap_area(va);
1234 atomic_long_sub(nr, &vmap_lazy_nr);
1236 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1237 cond_resched_lock(&vmap_area_lock);
1239 spin_unlock(&vmap_area_lock);
1244 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1245 * is already purging.
1247 static void try_purge_vmap_area_lazy(void)
1249 if (mutex_trylock(&vmap_purge_lock)) {
1250 __purge_vmap_area_lazy(ULONG_MAX, 0);
1251 mutex_unlock(&vmap_purge_lock);
1256 * Kick off a purge of the outstanding lazy areas.
1258 static void purge_vmap_area_lazy(void)
1260 mutex_lock(&vmap_purge_lock);
1261 purge_fragmented_blocks_allcpus();
1262 __purge_vmap_area_lazy(ULONG_MAX, 0);
1263 mutex_unlock(&vmap_purge_lock);
1267 * Free a vmap area, caller ensuring that the area has been unmapped
1268 * and flush_cache_vunmap had been called for the correct range
1271 static void free_vmap_area_noflush(struct vmap_area *va)
1273 unsigned long nr_lazy;
1275 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1276 PAGE_SHIFT, &vmap_lazy_nr);
1278 /* After this point, we may free va at any time */
1279 llist_add(&va->purge_list, &vmap_purge_list);
1281 if (unlikely(nr_lazy > lazy_max_pages()))
1282 try_purge_vmap_area_lazy();
1286 * Free and unmap a vmap area
1288 static void free_unmap_vmap_area(struct vmap_area *va)
1290 flush_cache_vunmap(va->va_start, va->va_end);
1291 unmap_vmap_area(va);
1292 if (debug_pagealloc_enabled())
1293 flush_tlb_kernel_range(va->va_start, va->va_end);
1295 free_vmap_area_noflush(va);
1298 static struct vmap_area *find_vmap_area(unsigned long addr)
1300 struct vmap_area *va;
1302 spin_lock(&vmap_area_lock);
1303 va = __find_vmap_area(addr);
1304 spin_unlock(&vmap_area_lock);
1309 /*** Per cpu kva allocator ***/
1312 * vmap space is limited especially on 32 bit architectures. Ensure there is
1313 * room for at least 16 percpu vmap blocks per CPU.
1316 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1317 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1318 * instead (we just need a rough idea)
1320 #if BITS_PER_LONG == 32
1321 #define VMALLOC_SPACE (128UL*1024*1024)
1323 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1326 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1327 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1328 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1329 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1330 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1331 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1332 #define VMAP_BBMAP_BITS \
1333 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1334 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1335 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1337 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1339 struct vmap_block_queue {
1341 struct list_head free;
1346 struct vmap_area *va;
1347 unsigned long free, dirty;
1348 unsigned long dirty_min, dirty_max; /*< dirty range */
1349 struct list_head free_list;
1350 struct rcu_head rcu_head;
1351 struct list_head purge;
1354 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1355 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1358 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1359 * in the free path. Could get rid of this if we change the API to return a
1360 * "cookie" from alloc, to be passed to free. But no big deal yet.
1362 static DEFINE_SPINLOCK(vmap_block_tree_lock);
1363 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
1366 * We should probably have a fallback mechanism to allocate virtual memory
1367 * out of partially filled vmap blocks. However vmap block sizing should be
1368 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1372 static unsigned long addr_to_vb_idx(unsigned long addr)
1374 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1375 addr /= VMAP_BLOCK_SIZE;
1379 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1383 addr = va_start + (pages_off << PAGE_SHIFT);
1384 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1385 return (void *)addr;
1389 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1390 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1391 * @order: how many 2^order pages should be occupied in newly allocated block
1392 * @gfp_mask: flags for the page level allocator
1394 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1396 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1398 struct vmap_block_queue *vbq;
1399 struct vmap_block *vb;
1400 struct vmap_area *va;
1401 unsigned long vb_idx;
1405 node = numa_node_id();
1407 vb = kmalloc_node(sizeof(struct vmap_block),
1408 gfp_mask & GFP_RECLAIM_MASK, node);
1410 return ERR_PTR(-ENOMEM);
1412 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1413 VMALLOC_START, VMALLOC_END,
1417 return ERR_CAST(va);
1420 err = radix_tree_preload(gfp_mask);
1421 if (unlikely(err)) {
1424 return ERR_PTR(err);
1427 vaddr = vmap_block_vaddr(va->va_start, 0);
1428 spin_lock_init(&vb->lock);
1430 /* At least something should be left free */
1431 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1432 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1434 vb->dirty_min = VMAP_BBMAP_BITS;
1436 INIT_LIST_HEAD(&vb->free_list);
1438 vb_idx = addr_to_vb_idx(va->va_start);
1439 spin_lock(&vmap_block_tree_lock);
1440 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
1441 spin_unlock(&vmap_block_tree_lock);
1443 radix_tree_preload_end();
1445 vbq = &get_cpu_var(vmap_block_queue);
1446 spin_lock(&vbq->lock);
1447 list_add_tail_rcu(&vb->free_list, &vbq->free);
1448 spin_unlock(&vbq->lock);
1449 put_cpu_var(vmap_block_queue);
1454 static void free_vmap_block(struct vmap_block *vb)
1456 struct vmap_block *tmp;
1457 unsigned long vb_idx;
1459 vb_idx = addr_to_vb_idx(vb->va->va_start);
1460 spin_lock(&vmap_block_tree_lock);
1461 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
1462 spin_unlock(&vmap_block_tree_lock);
1465 free_vmap_area_noflush(vb->va);
1466 kfree_rcu(vb, rcu_head);
1469 static void purge_fragmented_blocks(int cpu)
1472 struct vmap_block *vb;
1473 struct vmap_block *n_vb;
1474 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1477 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1479 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1482 spin_lock(&vb->lock);
1483 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1484 vb->free = 0; /* prevent further allocs after releasing lock */
1485 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1487 vb->dirty_max = VMAP_BBMAP_BITS;
1488 spin_lock(&vbq->lock);
1489 list_del_rcu(&vb->free_list);
1490 spin_unlock(&vbq->lock);
1491 spin_unlock(&vb->lock);
1492 list_add_tail(&vb->purge, &purge);
1494 spin_unlock(&vb->lock);
1498 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1499 list_del(&vb->purge);
1500 free_vmap_block(vb);
1504 static void purge_fragmented_blocks_allcpus(void)
1508 for_each_possible_cpu(cpu)
1509 purge_fragmented_blocks(cpu);
1512 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1514 struct vmap_block_queue *vbq;
1515 struct vmap_block *vb;
1519 BUG_ON(offset_in_page(size));
1520 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1521 if (WARN_ON(size == 0)) {
1523 * Allocating 0 bytes isn't what caller wants since
1524 * get_order(0) returns funny result. Just warn and terminate
1529 order = get_order(size);
1532 vbq = &get_cpu_var(vmap_block_queue);
1533 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1534 unsigned long pages_off;
1536 spin_lock(&vb->lock);
1537 if (vb->free < (1UL << order)) {
1538 spin_unlock(&vb->lock);
1542 pages_off = VMAP_BBMAP_BITS - vb->free;
1543 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1544 vb->free -= 1UL << order;
1545 if (vb->free == 0) {
1546 spin_lock(&vbq->lock);
1547 list_del_rcu(&vb->free_list);
1548 spin_unlock(&vbq->lock);
1551 spin_unlock(&vb->lock);
1555 put_cpu_var(vmap_block_queue);
1558 /* Allocate new block if nothing was found */
1560 vaddr = new_vmap_block(order, gfp_mask);
1565 static void vb_free(const void *addr, unsigned long size)
1567 unsigned long offset;
1568 unsigned long vb_idx;
1570 struct vmap_block *vb;
1572 BUG_ON(offset_in_page(size));
1573 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1575 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1577 order = get_order(size);
1579 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1580 offset >>= PAGE_SHIFT;
1582 vb_idx = addr_to_vb_idx((unsigned long)addr);
1584 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1588 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1590 if (debug_pagealloc_enabled())
1591 flush_tlb_kernel_range((unsigned long)addr,
1592 (unsigned long)addr + size);
1594 spin_lock(&vb->lock);
1596 /* Expand dirty range */
1597 vb->dirty_min = min(vb->dirty_min, offset);
1598 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1600 vb->dirty += 1UL << order;
1601 if (vb->dirty == VMAP_BBMAP_BITS) {
1603 spin_unlock(&vb->lock);
1604 free_vmap_block(vb);
1606 spin_unlock(&vb->lock);
1609 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1613 if (unlikely(!vmap_initialized))
1618 for_each_possible_cpu(cpu) {
1619 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1620 struct vmap_block *vb;
1623 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1624 spin_lock(&vb->lock);
1626 unsigned long va_start = vb->va->va_start;
1629 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1630 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1632 start = min(s, start);
1637 spin_unlock(&vb->lock);
1642 mutex_lock(&vmap_purge_lock);
1643 purge_fragmented_blocks_allcpus();
1644 if (!__purge_vmap_area_lazy(start, end) && flush)
1645 flush_tlb_kernel_range(start, end);
1646 mutex_unlock(&vmap_purge_lock);
1650 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1652 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1653 * to amortize TLB flushing overheads. What this means is that any page you
1654 * have now, may, in a former life, have been mapped into kernel virtual
1655 * address by the vmap layer and so there might be some CPUs with TLB entries
1656 * still referencing that page (additional to the regular 1:1 kernel mapping).
1658 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1659 * be sure that none of the pages we have control over will have any aliases
1660 * from the vmap layer.
1662 void vm_unmap_aliases(void)
1664 unsigned long start = ULONG_MAX, end = 0;
1667 _vm_unmap_aliases(start, end, flush);
1669 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1672 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1673 * @mem: the pointer returned by vm_map_ram
1674 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1676 void vm_unmap_ram(const void *mem, unsigned int count)
1678 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1679 unsigned long addr = (unsigned long)mem;
1680 struct vmap_area *va;
1684 BUG_ON(addr < VMALLOC_START);
1685 BUG_ON(addr > VMALLOC_END);
1686 BUG_ON(!PAGE_ALIGNED(addr));
1688 if (likely(count <= VMAP_MAX_ALLOC)) {
1689 debug_check_no_locks_freed(mem, size);
1694 va = find_vmap_area(addr);
1696 debug_check_no_locks_freed((void *)va->va_start,
1697 (va->va_end - va->va_start));
1698 free_unmap_vmap_area(va);
1700 EXPORT_SYMBOL(vm_unmap_ram);
1703 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1704 * @pages: an array of pointers to the pages to be mapped
1705 * @count: number of pages
1706 * @node: prefer to allocate data structures on this node
1707 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1709 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1710 * faster than vmap so it's good. But if you mix long-life and short-life
1711 * objects with vm_map_ram(), it could consume lots of address space through
1712 * fragmentation (especially on a 32bit machine). You could see failures in
1713 * the end. Please use this function for short-lived objects.
1715 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1717 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1719 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1723 if (likely(count <= VMAP_MAX_ALLOC)) {
1724 mem = vb_alloc(size, GFP_KERNEL);
1727 addr = (unsigned long)mem;
1729 struct vmap_area *va;
1730 va = alloc_vmap_area(size, PAGE_SIZE,
1731 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1735 addr = va->va_start;
1738 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1739 vm_unmap_ram(mem, count);
1744 EXPORT_SYMBOL(vm_map_ram);
1746 static struct vm_struct *vmlist __initdata;
1749 * vm_area_add_early - add vmap area early during boot
1750 * @vm: vm_struct to add
1752 * This function is used to add fixed kernel vm area to vmlist before
1753 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1754 * should contain proper values and the other fields should be zero.
1756 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1758 void __init vm_area_add_early(struct vm_struct *vm)
1760 struct vm_struct *tmp, **p;
1762 BUG_ON(vmap_initialized);
1763 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1764 if (tmp->addr >= vm->addr) {
1765 BUG_ON(tmp->addr < vm->addr + vm->size);
1768 BUG_ON(tmp->addr + tmp->size > vm->addr);
1775 * vm_area_register_early - register vmap area early during boot
1776 * @vm: vm_struct to register
1777 * @align: requested alignment
1779 * This function is used to register kernel vm area before
1780 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1781 * proper values on entry and other fields should be zero. On return,
1782 * vm->addr contains the allocated address.
1784 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1786 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1788 static size_t vm_init_off __initdata;
1791 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1792 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1794 vm->addr = (void *)addr;
1796 vm_area_add_early(vm);
1799 static void vmap_init_free_space(void)
1801 unsigned long vmap_start = 1;
1802 const unsigned long vmap_end = ULONG_MAX;
1803 struct vmap_area *busy, *free;
1807 * -|-----|.....|-----|-----|-----|.....|-
1809 * |<--------------------------------->|
1811 list_for_each_entry(busy, &vmap_area_list, list) {
1812 if (busy->va_start - vmap_start > 0) {
1813 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1814 if (!WARN_ON_ONCE(!free)) {
1815 free->va_start = vmap_start;
1816 free->va_end = busy->va_start;
1818 insert_vmap_area_augment(free, NULL,
1819 &free_vmap_area_root,
1820 &free_vmap_area_list);
1824 vmap_start = busy->va_end;
1827 if (vmap_end - vmap_start > 0) {
1828 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1829 if (!WARN_ON_ONCE(!free)) {
1830 free->va_start = vmap_start;
1831 free->va_end = vmap_end;
1833 insert_vmap_area_augment(free, NULL,
1834 &free_vmap_area_root,
1835 &free_vmap_area_list);
1840 void __init vmalloc_init(void)
1842 struct vmap_area *va;
1843 struct vm_struct *tmp;
1847 * Create the cache for vmap_area objects.
1849 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1851 for_each_possible_cpu(i) {
1852 struct vmap_block_queue *vbq;
1853 struct vfree_deferred *p;
1855 vbq = &per_cpu(vmap_block_queue, i);
1856 spin_lock_init(&vbq->lock);
1857 INIT_LIST_HEAD(&vbq->free);
1858 p = &per_cpu(vfree_deferred, i);
1859 init_llist_head(&p->list);
1860 INIT_WORK(&p->wq, free_work);
1863 /* Import existing vmlist entries. */
1864 for (tmp = vmlist; tmp; tmp = tmp->next) {
1865 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1866 if (WARN_ON_ONCE(!va))
1869 va->flags = VM_VM_AREA;
1870 va->va_start = (unsigned long)tmp->addr;
1871 va->va_end = va->va_start + tmp->size;
1873 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1877 * Now we can initialize a free vmap space.
1879 vmap_init_free_space();
1880 vmap_initialized = true;
1884 * map_kernel_range_noflush - map kernel VM area with the specified pages
1885 * @addr: start of the VM area to map
1886 * @size: size of the VM area to map
1887 * @prot: page protection flags to use
1888 * @pages: pages to map
1890 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1891 * specify should have been allocated using get_vm_area() and its
1895 * This function does NOT do any cache flushing. The caller is
1896 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1897 * before calling this function.
1900 * The number of pages mapped on success, -errno on failure.
1902 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1903 pgprot_t prot, struct page **pages)
1905 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1909 * unmap_kernel_range_noflush - unmap kernel VM area
1910 * @addr: start of the VM area to unmap
1911 * @size: size of the VM area to unmap
1913 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1914 * specify should have been allocated using get_vm_area() and its
1918 * This function does NOT do any cache flushing. The caller is
1919 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1920 * before calling this function and flush_tlb_kernel_range() after.
1922 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1924 vunmap_page_range(addr, addr + size);
1926 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1929 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1930 * @addr: start of the VM area to unmap
1931 * @size: size of the VM area to unmap
1933 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1934 * the unmapping and tlb after.
1936 void unmap_kernel_range(unsigned long addr, unsigned long size)
1938 unsigned long end = addr + size;
1940 flush_cache_vunmap(addr, end);
1941 vunmap_page_range(addr, end);
1942 flush_tlb_kernel_range(addr, end);
1944 EXPORT_SYMBOL_GPL(unmap_kernel_range);
1946 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1948 unsigned long addr = (unsigned long)area->addr;
1949 unsigned long end = addr + get_vm_area_size(area);
1952 err = vmap_page_range(addr, end, prot, pages);
1954 return err > 0 ? 0 : err;
1956 EXPORT_SYMBOL_GPL(map_vm_area);
1958 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1959 unsigned long flags, const void *caller)
1961 spin_lock(&vmap_area_lock);
1963 vm->addr = (void *)va->va_start;
1964 vm->size = va->va_end - va->va_start;
1965 vm->caller = caller;
1967 va->flags |= VM_VM_AREA;
1968 spin_unlock(&vmap_area_lock);
1971 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1974 * Before removing VM_UNINITIALIZED,
1975 * we should make sure that vm has proper values.
1976 * Pair with smp_rmb() in show_numa_info().
1979 vm->flags &= ~VM_UNINITIALIZED;
1982 static struct vm_struct *__get_vm_area_node(unsigned long size,
1983 unsigned long align, unsigned long flags, unsigned long start,
1984 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1986 struct vmap_area *va;
1987 struct vm_struct *area;
1989 BUG_ON(in_interrupt());
1990 size = PAGE_ALIGN(size);
1991 if (unlikely(!size))
1994 if (flags & VM_IOREMAP)
1995 align = 1ul << clamp_t(int, get_count_order_long(size),
1996 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1998 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1999 if (unlikely(!area))
2002 if (!(flags & VM_NO_GUARD))
2005 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2011 setup_vmalloc_vm(area, va, flags, caller);
2016 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
2017 unsigned long start, unsigned long end)
2019 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2020 GFP_KERNEL, __builtin_return_address(0));
2022 EXPORT_SYMBOL_GPL(__get_vm_area);
2024 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2025 unsigned long start, unsigned long end,
2028 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2029 GFP_KERNEL, caller);
2033 * get_vm_area - reserve a contiguous kernel virtual area
2034 * @size: size of the area
2035 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2037 * Search an area of @size in the kernel virtual mapping area,
2038 * and reserved it for out purposes. Returns the area descriptor
2039 * on success or %NULL on failure.
2041 * Return: the area descriptor on success or %NULL on failure.
2043 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2045 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2046 NUMA_NO_NODE, GFP_KERNEL,
2047 __builtin_return_address(0));
2050 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2053 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2054 NUMA_NO_NODE, GFP_KERNEL, caller);
2058 * find_vm_area - find a continuous kernel virtual area
2059 * @addr: base address
2061 * Search for the kernel VM area starting at @addr, and return it.
2062 * It is up to the caller to do all required locking to keep the returned
2065 * Return: pointer to the found area or %NULL on faulure
2067 struct vm_struct *find_vm_area(const void *addr)
2069 struct vmap_area *va;
2071 va = find_vmap_area((unsigned long)addr);
2072 if (va && va->flags & VM_VM_AREA)
2079 * remove_vm_area - find and remove a continuous kernel virtual area
2080 * @addr: base address
2082 * Search for the kernel VM area starting at @addr, and remove it.
2083 * This function returns the found VM area, but using it is NOT safe
2084 * on SMP machines, except for its size or flags.
2086 * Return: pointer to the found area or %NULL on faulure
2088 struct vm_struct *remove_vm_area(const void *addr)
2090 struct vmap_area *va;
2094 va = find_vmap_area((unsigned long)addr);
2095 if (va && va->flags & VM_VM_AREA) {
2096 struct vm_struct *vm = va->vm;
2098 spin_lock(&vmap_area_lock);
2100 va->flags &= ~VM_VM_AREA;
2101 va->flags |= VM_LAZY_FREE;
2102 spin_unlock(&vmap_area_lock);
2104 kasan_free_shadow(vm);
2105 free_unmap_vmap_area(va);
2112 static inline void set_area_direct_map(const struct vm_struct *area,
2113 int (*set_direct_map)(struct page *page))
2117 for (i = 0; i < area->nr_pages; i++)
2118 if (page_address(area->pages[i]))
2119 set_direct_map(area->pages[i]);
2122 /* Handle removing and resetting vm mappings related to the vm_struct. */
2123 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2125 unsigned long addr = (unsigned long)area->addr;
2126 unsigned long start = ULONG_MAX, end = 0;
2127 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2131 * The below block can be removed when all architectures that have
2132 * direct map permissions also have set_direct_map_() implementations.
2133 * This is concerned with resetting the direct map any an vm alias with
2134 * execute permissions, without leaving a RW+X window.
2136 if (flush_reset && !IS_ENABLED(CONFIG_ARCH_HAS_SET_DIRECT_MAP)) {
2137 set_memory_nx(addr, area->nr_pages);
2138 set_memory_rw(addr, area->nr_pages);
2141 remove_vm_area(area->addr);
2143 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2148 * If not deallocating pages, just do the flush of the VM area and
2151 if (!deallocate_pages) {
2157 * If execution gets here, flush the vm mapping and reset the direct
2158 * map. Find the start and end range of the direct mappings to make sure
2159 * the vm_unmap_aliases() flush includes the direct map.
2161 for (i = 0; i < area->nr_pages; i++) {
2162 if (page_address(area->pages[i])) {
2163 start = min(addr, start);
2164 end = max(addr, end);
2169 * Set direct map to something invalid so that it won't be cached if
2170 * there are any accesses after the TLB flush, then flush the TLB and
2171 * reset the direct map permissions to the default.
2173 set_area_direct_map(area, set_direct_map_invalid_noflush);
2174 _vm_unmap_aliases(start, end, 1);
2175 set_area_direct_map(area, set_direct_map_default_noflush);
2178 static void __vunmap(const void *addr, int deallocate_pages)
2180 struct vm_struct *area;
2185 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2189 area = find_vm_area(addr);
2190 if (unlikely(!area)) {
2191 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2196 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2197 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2199 vm_remove_mappings(area, deallocate_pages);
2201 if (deallocate_pages) {
2204 for (i = 0; i < area->nr_pages; i++) {
2205 struct page *page = area->pages[i];
2208 __free_pages(page, 0);
2211 kvfree(area->pages);
2218 static inline void __vfree_deferred(const void *addr)
2221 * Use raw_cpu_ptr() because this can be called from preemptible
2222 * context. Preemption is absolutely fine here, because the llist_add()
2223 * implementation is lockless, so it works even if we are adding to
2224 * nother cpu's list. schedule_work() should be fine with this too.
2226 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2228 if (llist_add((struct llist_node *)addr, &p->list))
2229 schedule_work(&p->wq);
2233 * vfree_atomic - release memory allocated by vmalloc()
2234 * @addr: memory base address
2236 * This one is just like vfree() but can be called in any atomic context
2239 void vfree_atomic(const void *addr)
2243 kmemleak_free(addr);
2247 __vfree_deferred(addr);
2250 static void __vfree(const void *addr)
2252 if (unlikely(in_interrupt()))
2253 __vfree_deferred(addr);
2259 * vfree - release memory allocated by vmalloc()
2260 * @addr: memory base address
2262 * Free the virtually continuous memory area starting at @addr, as
2263 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2264 * NULL, no operation is performed.
2266 * Must not be called in NMI context (strictly speaking, only if we don't
2267 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2268 * conventions for vfree() arch-depenedent would be a really bad idea)
2270 * May sleep if called *not* from interrupt context.
2272 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2274 void vfree(const void *addr)
2278 kmemleak_free(addr);
2280 might_sleep_if(!in_interrupt());
2287 EXPORT_SYMBOL(vfree);
2290 * vunmap - release virtual mapping obtained by vmap()
2291 * @addr: memory base address
2293 * Free the virtually contiguous memory area starting at @addr,
2294 * which was created from the page array passed to vmap().
2296 * Must not be called in interrupt context.
2298 void vunmap(const void *addr)
2300 BUG_ON(in_interrupt());
2305 EXPORT_SYMBOL(vunmap);
2308 * vmap - map an array of pages into virtually contiguous space
2309 * @pages: array of page pointers
2310 * @count: number of pages to map
2311 * @flags: vm_area->flags
2312 * @prot: page protection for the mapping
2314 * Maps @count pages from @pages into contiguous kernel virtual
2317 * Return: the address of the area or %NULL on failure
2319 void *vmap(struct page **pages, unsigned int count,
2320 unsigned long flags, pgprot_t prot)
2322 struct vm_struct *area;
2323 unsigned long size; /* In bytes */
2327 if (count > totalram_pages())
2330 size = (unsigned long)count << PAGE_SHIFT;
2331 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2335 if (map_vm_area(area, prot, pages)) {
2342 EXPORT_SYMBOL(vmap);
2344 static void *__vmalloc_node(unsigned long size, unsigned long align,
2345 gfp_t gfp_mask, pgprot_t prot,
2346 int node, const void *caller);
2347 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2348 pgprot_t prot, int node)
2350 struct page **pages;
2351 unsigned int nr_pages, array_size, i;
2352 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2353 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2354 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2358 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2359 array_size = (nr_pages * sizeof(struct page *));
2361 area->nr_pages = nr_pages;
2362 /* Please note that the recursion is strictly bounded. */
2363 if (array_size > PAGE_SIZE) {
2364 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2365 PAGE_KERNEL, node, area->caller);
2367 pages = kmalloc_node(array_size, nested_gfp, node);
2369 area->pages = pages;
2371 remove_vm_area(area->addr);
2376 for (i = 0; i < area->nr_pages; i++) {
2379 if (node == NUMA_NO_NODE)
2380 page = alloc_page(alloc_mask|highmem_mask);
2382 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2384 if (unlikely(!page)) {
2385 /* Successfully allocated i pages, free them in __vunmap() */
2389 area->pages[i] = page;
2390 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
2394 if (map_vm_area(area, prot, pages))
2399 warn_alloc(gfp_mask, NULL,
2400 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2401 (area->nr_pages*PAGE_SIZE), area->size);
2402 __vfree(area->addr);
2407 * __vmalloc_node_range - allocate virtually contiguous memory
2408 * @size: allocation size
2409 * @align: desired alignment
2410 * @start: vm area range start
2411 * @end: vm area range end
2412 * @gfp_mask: flags for the page level allocator
2413 * @prot: protection mask for the allocated pages
2414 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2415 * @node: node to use for allocation or NUMA_NO_NODE
2416 * @caller: caller's return address
2418 * Allocate enough pages to cover @size from the page level
2419 * allocator with @gfp_mask flags. Map them into contiguous
2420 * kernel virtual space, using a pagetable protection of @prot.
2422 * Return: the address of the area or %NULL on failure
2424 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2425 unsigned long start, unsigned long end, gfp_t gfp_mask,
2426 pgprot_t prot, unsigned long vm_flags, int node,
2429 struct vm_struct *area;
2431 unsigned long real_size = size;
2433 size = PAGE_ALIGN(size);
2434 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2437 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
2438 vm_flags, start, end, node, gfp_mask, caller);
2442 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2447 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2448 * flag. It means that vm_struct is not fully initialized.
2449 * Now, it is fully initialized, so remove this flag here.
2451 clear_vm_uninitialized_flag(area);
2453 kmemleak_vmalloc(area, size, gfp_mask);
2458 warn_alloc(gfp_mask, NULL,
2459 "vmalloc: allocation failure: %lu bytes", real_size);
2464 * This is only for performance analysis of vmalloc and stress purpose.
2465 * It is required by vmalloc test module, therefore do not use it other
2468 #ifdef CONFIG_TEST_VMALLOC_MODULE
2469 EXPORT_SYMBOL_GPL(__vmalloc_node_range);
2473 * __vmalloc_node - allocate virtually contiguous memory
2474 * @size: allocation size
2475 * @align: desired alignment
2476 * @gfp_mask: flags for the page level allocator
2477 * @prot: protection mask for the allocated pages
2478 * @node: node to use for allocation or NUMA_NO_NODE
2479 * @caller: caller's return address
2481 * Allocate enough pages to cover @size from the page level
2482 * allocator with @gfp_mask flags. Map them into contiguous
2483 * kernel virtual space, using a pagetable protection of @prot.
2485 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2486 * and __GFP_NOFAIL are not supported
2488 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2491 * Return: pointer to the allocated memory or %NULL on error
2493 static void *__vmalloc_node(unsigned long size, unsigned long align,
2494 gfp_t gfp_mask, pgprot_t prot,
2495 int node, const void *caller)
2497 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2498 gfp_mask, prot, 0, node, caller);
2501 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
2503 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
2504 __builtin_return_address(0));
2506 EXPORT_SYMBOL(__vmalloc);
2508 static inline void *__vmalloc_node_flags(unsigned long size,
2509 int node, gfp_t flags)
2511 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
2512 node, __builtin_return_address(0));
2516 void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
2519 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
2523 * vmalloc - allocate virtually contiguous memory
2524 * @size: allocation size
2526 * Allocate enough pages to cover @size from the page level
2527 * allocator and map them into contiguous kernel virtual space.
2529 * For tight control over page level allocator and protection flags
2530 * use __vmalloc() instead.
2532 * Return: pointer to the allocated memory or %NULL on error
2534 void *vmalloc(unsigned long size)
2536 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2539 EXPORT_SYMBOL(vmalloc);
2542 * vzalloc - allocate virtually contiguous memory with zero fill
2543 * @size: allocation size
2545 * Allocate enough pages to cover @size from the page level
2546 * allocator and map them into contiguous kernel virtual space.
2547 * The memory allocated is set to zero.
2549 * For tight control over page level allocator and protection flags
2550 * use __vmalloc() instead.
2552 * Return: pointer to the allocated memory or %NULL on error
2554 void *vzalloc(unsigned long size)
2556 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2557 GFP_KERNEL | __GFP_ZERO);
2559 EXPORT_SYMBOL(vzalloc);
2562 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2563 * @size: allocation size
2565 * The resulting memory area is zeroed so it can be mapped to userspace
2566 * without leaking data.
2568 * Return: pointer to the allocated memory or %NULL on error
2570 void *vmalloc_user(unsigned long size)
2572 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2573 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2574 VM_USERMAP, NUMA_NO_NODE,
2575 __builtin_return_address(0));
2577 EXPORT_SYMBOL(vmalloc_user);
2580 * vmalloc_node - allocate memory on a specific node
2581 * @size: allocation size
2584 * Allocate enough pages to cover @size from the page level
2585 * allocator and map them into contiguous kernel virtual space.
2587 * For tight control over page level allocator and protection flags
2588 * use __vmalloc() instead.
2590 * Return: pointer to the allocated memory or %NULL on error
2592 void *vmalloc_node(unsigned long size, int node)
2594 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
2595 node, __builtin_return_address(0));
2597 EXPORT_SYMBOL(vmalloc_node);
2600 * vzalloc_node - allocate memory on a specific node with zero fill
2601 * @size: allocation size
2604 * Allocate enough pages to cover @size from the page level
2605 * allocator and map them into contiguous kernel virtual space.
2606 * The memory allocated is set to zero.
2608 * For tight control over page level allocator and protection flags
2609 * use __vmalloc_node() instead.
2611 * Return: pointer to the allocated memory or %NULL on error
2613 void *vzalloc_node(unsigned long size, int node)
2615 return __vmalloc_node_flags(size, node,
2616 GFP_KERNEL | __GFP_ZERO);
2618 EXPORT_SYMBOL(vzalloc_node);
2621 * vmalloc_exec - allocate virtually contiguous, executable memory
2622 * @size: allocation size
2624 * Kernel-internal function to allocate enough pages to cover @size
2625 * the page level allocator and map them into contiguous and
2626 * executable kernel virtual space.
2628 * For tight control over page level allocator and protection flags
2629 * use __vmalloc() instead.
2631 * Return: pointer to the allocated memory or %NULL on error
2633 void *vmalloc_exec(unsigned long size)
2635 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
2636 GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
2637 NUMA_NO_NODE, __builtin_return_address(0));
2640 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2641 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2642 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2643 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2646 * 64b systems should always have either DMA or DMA32 zones. For others
2647 * GFP_DMA32 should do the right thing and use the normal zone.
2649 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2653 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2654 * @size: allocation size
2656 * Allocate enough 32bit PA addressable pages to cover @size from the
2657 * page level allocator and map them into contiguous kernel virtual space.
2659 * Return: pointer to the allocated memory or %NULL on error
2661 void *vmalloc_32(unsigned long size)
2663 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
2664 NUMA_NO_NODE, __builtin_return_address(0));
2666 EXPORT_SYMBOL(vmalloc_32);
2669 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2670 * @size: allocation size
2672 * The resulting memory area is 32bit addressable and zeroed so it can be
2673 * mapped to userspace without leaking data.
2675 * Return: pointer to the allocated memory or %NULL on error
2677 void *vmalloc_32_user(unsigned long size)
2679 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2680 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2681 VM_USERMAP, NUMA_NO_NODE,
2682 __builtin_return_address(0));
2684 EXPORT_SYMBOL(vmalloc_32_user);
2687 * small helper routine , copy contents to buf from addr.
2688 * If the page is not present, fill zero.
2691 static int aligned_vread(char *buf, char *addr, unsigned long count)
2697 unsigned long offset, length;
2699 offset = offset_in_page(addr);
2700 length = PAGE_SIZE - offset;
2703 p = vmalloc_to_page(addr);
2705 * To do safe access to this _mapped_ area, we need
2706 * lock. But adding lock here means that we need to add
2707 * overhead of vmalloc()/vfree() calles for this _debug_
2708 * interface, rarely used. Instead of that, we'll use
2709 * kmap() and get small overhead in this access function.
2713 * we can expect USER0 is not used (see vread/vwrite's
2714 * function description)
2716 void *map = kmap_atomic(p);
2717 memcpy(buf, map + offset, length);
2720 memset(buf, 0, length);
2730 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2736 unsigned long offset, length;
2738 offset = offset_in_page(addr);
2739 length = PAGE_SIZE - offset;
2742 p = vmalloc_to_page(addr);
2744 * To do safe access to this _mapped_ area, we need
2745 * lock. But adding lock here means that we need to add
2746 * overhead of vmalloc()/vfree() calles for this _debug_
2747 * interface, rarely used. Instead of that, we'll use
2748 * kmap() and get small overhead in this access function.
2752 * we can expect USER0 is not used (see vread/vwrite's
2753 * function description)
2755 void *map = kmap_atomic(p);
2756 memcpy(map + offset, buf, length);
2768 * vread() - read vmalloc area in a safe way.
2769 * @buf: buffer for reading data
2770 * @addr: vm address.
2771 * @count: number of bytes to be read.
2773 * This function checks that addr is a valid vmalloc'ed area, and
2774 * copy data from that area to a given buffer. If the given memory range
2775 * of [addr...addr+count) includes some valid address, data is copied to
2776 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2777 * IOREMAP area is treated as memory hole and no copy is done.
2779 * If [addr...addr+count) doesn't includes any intersects with alive
2780 * vm_struct area, returns 0. @buf should be kernel's buffer.
2782 * Note: In usual ops, vread() is never necessary because the caller
2783 * should know vmalloc() area is valid and can use memcpy().
2784 * This is for routines which have to access vmalloc area without
2785 * any informaion, as /dev/kmem.
2787 * Return: number of bytes for which addr and buf should be increased
2788 * (same number as @count) or %0 if [addr...addr+count) doesn't
2789 * include any intersection with valid vmalloc area
2791 long vread(char *buf, char *addr, unsigned long count)
2793 struct vmap_area *va;
2794 struct vm_struct *vm;
2795 char *vaddr, *buf_start = buf;
2796 unsigned long buflen = count;
2799 /* Don't allow overflow */
2800 if ((unsigned long) addr + count < count)
2801 count = -(unsigned long) addr;
2803 spin_lock(&vmap_area_lock);
2804 list_for_each_entry(va, &vmap_area_list, list) {
2808 if (!(va->flags & VM_VM_AREA))
2812 vaddr = (char *) vm->addr;
2813 if (addr >= vaddr + get_vm_area_size(vm))
2815 while (addr < vaddr) {
2823 n = vaddr + get_vm_area_size(vm) - addr;
2826 if (!(vm->flags & VM_IOREMAP))
2827 aligned_vread(buf, addr, n);
2828 else /* IOREMAP area is treated as memory hole */
2835 spin_unlock(&vmap_area_lock);
2837 if (buf == buf_start)
2839 /* zero-fill memory holes */
2840 if (buf != buf_start + buflen)
2841 memset(buf, 0, buflen - (buf - buf_start));
2847 * vwrite() - write vmalloc area in a safe way.
2848 * @buf: buffer for source 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 a buffer to the given addr. If specified range of
2854 * [addr...addr+count) includes some valid address, data is copied from
2855 * proper area of @buf. If there are memory holes, no copy to hole.
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, vwrite() 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 informaion, as /dev/kmem.
2866 * Return: number of bytes for which addr and buf should be
2867 * increased (same number as @count) or %0 if [addr...addr+count)
2868 * doesn't include any intersection with valid vmalloc area
2870 long vwrite(char *buf, char *addr, unsigned long count)
2872 struct vmap_area *va;
2873 struct vm_struct *vm;
2875 unsigned long n, buflen;
2878 /* Don't allow overflow */
2879 if ((unsigned long) addr + count < count)
2880 count = -(unsigned long) addr;
2883 spin_lock(&vmap_area_lock);
2884 list_for_each_entry(va, &vmap_area_list, list) {
2888 if (!(va->flags & VM_VM_AREA))
2892 vaddr = (char *) vm->addr;
2893 if (addr >= vaddr + get_vm_area_size(vm))
2895 while (addr < vaddr) {
2902 n = vaddr + get_vm_area_size(vm) - addr;
2905 if (!(vm->flags & VM_IOREMAP)) {
2906 aligned_vwrite(buf, addr, n);
2914 spin_unlock(&vmap_area_lock);
2921 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2922 * @vma: vma to cover
2923 * @uaddr: target user address to start at
2924 * @kaddr: virtual address of vmalloc kernel memory
2925 * @size: size of map area
2927 * Returns: 0 for success, -Exxx on failure
2929 * This function checks that @kaddr is a valid vmalloc'ed area,
2930 * and that it is big enough to cover the range starting at
2931 * @uaddr in @vma. Will return failure if that criteria isn't
2934 * Similar to remap_pfn_range() (see mm/memory.c)
2936 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2937 void *kaddr, unsigned long size)
2939 struct vm_struct *area;
2941 size = PAGE_ALIGN(size);
2943 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2946 area = find_vm_area(kaddr);
2950 if (!(area->flags & VM_USERMAP))
2953 if (kaddr + size > area->addr + get_vm_area_size(area))
2957 struct page *page = vmalloc_to_page(kaddr);
2960 ret = vm_insert_page(vma, uaddr, page);
2969 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2973 EXPORT_SYMBOL(remap_vmalloc_range_partial);
2976 * remap_vmalloc_range - map vmalloc pages to userspace
2977 * @vma: vma to cover (map full range of vma)
2978 * @addr: vmalloc memory
2979 * @pgoff: number of pages into addr before first page to map
2981 * Returns: 0 for success, -Exxx on failure
2983 * This function checks that addr is a valid vmalloc'ed area, and
2984 * that it is big enough to cover the vma. Will return failure if
2985 * that criteria isn't met.
2987 * Similar to remap_pfn_range() (see mm/memory.c)
2989 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2990 unsigned long pgoff)
2992 return remap_vmalloc_range_partial(vma, vma->vm_start,
2993 addr + (pgoff << PAGE_SHIFT),
2994 vma->vm_end - vma->vm_start);
2996 EXPORT_SYMBOL(remap_vmalloc_range);
2999 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3002 void __weak vmalloc_sync_all(void)
3007 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
3019 * alloc_vm_area - allocate a range of kernel address space
3020 * @size: size of the area
3021 * @ptes: returns the PTEs for the address space
3023 * Returns: NULL on failure, vm_struct on success
3025 * This function reserves a range of kernel address space, and
3026 * allocates pagetables to map that range. No actual mappings
3029 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3030 * allocated for the VM area are returned.
3032 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3034 struct vm_struct *area;
3036 area = get_vm_area_caller(size, VM_IOREMAP,
3037 __builtin_return_address(0));
3042 * This ensures that page tables are constructed for this region
3043 * of kernel virtual address space and mapped into init_mm.
3045 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3046 size, f, ptes ? &ptes : NULL)) {
3053 EXPORT_SYMBOL_GPL(alloc_vm_area);
3055 void free_vm_area(struct vm_struct *area)
3057 struct vm_struct *ret;
3058 ret = remove_vm_area(area->addr);
3059 BUG_ON(ret != area);
3062 EXPORT_SYMBOL_GPL(free_vm_area);
3065 static struct vmap_area *node_to_va(struct rb_node *n)
3067 return rb_entry_safe(n, struct vmap_area, rb_node);
3071 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3072 * @addr: target address
3074 * Returns: vmap_area if it is found. If there is no such area
3075 * the first highest(reverse order) vmap_area is returned
3076 * i.e. va->va_start < addr && va->va_end < addr or NULL
3077 * if there are no any areas before @addr.
3079 static struct vmap_area *
3080 pvm_find_va_enclose_addr(unsigned long addr)
3082 struct vmap_area *va, *tmp;
3085 n = free_vmap_area_root.rb_node;
3089 tmp = rb_entry(n, struct vmap_area, rb_node);
3090 if (tmp->va_start <= addr) {
3092 if (tmp->va_end >= addr)
3105 * pvm_determine_end_from_reverse - find the highest aligned address
3106 * of free block below VMALLOC_END
3108 * in - the VA we start the search(reverse order);
3109 * out - the VA with the highest aligned end address.
3111 * Returns: determined end address within vmap_area
3113 static unsigned long
3114 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3116 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3120 list_for_each_entry_from_reverse((*va),
3121 &free_vmap_area_list, list) {
3122 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3123 if ((*va)->va_start < addr)
3132 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3133 * @offsets: array containing offset of each area
3134 * @sizes: array containing size of each area
3135 * @nr_vms: the number of areas to allocate
3136 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3138 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3139 * vm_structs on success, %NULL on failure
3141 * Percpu allocator wants to use congruent vm areas so that it can
3142 * maintain the offsets among percpu areas. This function allocates
3143 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3144 * be scattered pretty far, distance between two areas easily going up
3145 * to gigabytes. To avoid interacting with regular vmallocs, these
3146 * areas are allocated from top.
3148 * Despite its complicated look, this allocator is rather simple. It
3149 * does everything top-down and scans free blocks from the end looking
3150 * for matching base. While scanning, if any of the areas do not fit the
3151 * base address is pulled down to fit the area. Scanning is repeated till
3152 * all the areas fit and then all necessary data structures are inserted
3153 * and the result is returned.
3155 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3156 const size_t *sizes, int nr_vms,
3159 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3160 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3161 struct vmap_area **vas, *va;
3162 struct vm_struct **vms;
3163 int area, area2, last_area, term_area;
3164 unsigned long base, start, size, end, last_end;
3165 bool purged = false;
3168 /* verify parameters and allocate data structures */
3169 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3170 for (last_area = 0, area = 0; area < nr_vms; area++) {
3171 start = offsets[area];
3172 end = start + sizes[area];
3174 /* is everything aligned properly? */
3175 BUG_ON(!IS_ALIGNED(offsets[area], align));
3176 BUG_ON(!IS_ALIGNED(sizes[area], align));
3178 /* detect the area with the highest address */
3179 if (start > offsets[last_area])
3182 for (area2 = area + 1; area2 < nr_vms; area2++) {
3183 unsigned long start2 = offsets[area2];
3184 unsigned long end2 = start2 + sizes[area2];
3186 BUG_ON(start2 < end && start < end2);
3189 last_end = offsets[last_area] + sizes[last_area];
3191 if (vmalloc_end - vmalloc_start < last_end) {
3196 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3197 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3201 for (area = 0; area < nr_vms; area++) {
3202 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3203 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3204 if (!vas[area] || !vms[area])
3208 spin_lock(&vmap_area_lock);
3210 /* start scanning - we scan from the top, begin with the last area */
3211 area = term_area = last_area;
3212 start = offsets[area];
3213 end = start + sizes[area];
3215 va = pvm_find_va_enclose_addr(vmalloc_end);
3216 base = pvm_determine_end_from_reverse(&va, align) - end;
3220 * base might have underflowed, add last_end before
3223 if (base + last_end < vmalloc_start + last_end)
3227 * Fitting base has not been found.
3233 * If this VA does not fit, move base downwards and recheck.
3235 if (base + start < va->va_start || base + end > va->va_end) {
3236 va = node_to_va(rb_prev(&va->rb_node));
3237 base = pvm_determine_end_from_reverse(&va, align) - end;
3243 * This area fits, move on to the previous one. If
3244 * the previous one is the terminal one, we're done.
3246 area = (area + nr_vms - 1) % nr_vms;
3247 if (area == term_area)
3250 start = offsets[area];
3251 end = start + sizes[area];
3252 va = pvm_find_va_enclose_addr(base + end);
3255 /* we've found a fitting base, insert all va's */
3256 for (area = 0; area < nr_vms; area++) {
3259 start = base + offsets[area];
3262 va = pvm_find_va_enclose_addr(start);
3263 if (WARN_ON_ONCE(va == NULL))
3264 /* It is a BUG(), but trigger recovery instead. */
3267 type = classify_va_fit_type(va, start, size);
3268 if (WARN_ON_ONCE(type == NOTHING_FIT))
3269 /* It is a BUG(), but trigger recovery instead. */
3272 ret = adjust_va_to_fit_type(va, start, size, type);
3276 /* Allocated area. */
3278 va->va_start = start;
3279 va->va_end = start + size;
3281 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
3284 spin_unlock(&vmap_area_lock);
3286 /* insert all vm's */
3287 for (area = 0; area < nr_vms; area++)
3288 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
3295 /* Remove previously inserted areas. */
3297 __free_vmap_area(vas[area]);
3302 spin_unlock(&vmap_area_lock);
3304 purge_vmap_area_lazy();
3307 /* Before "retry", check if we recover. */
3308 for (area = 0; area < nr_vms; area++) {
3312 vas[area] = kmem_cache_zalloc(
3313 vmap_area_cachep, GFP_KERNEL);
3322 for (area = 0; area < nr_vms; area++) {
3324 kmem_cache_free(vmap_area_cachep, vas[area]);
3335 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3336 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3337 * @nr_vms: the number of allocated areas
3339 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3341 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3345 for (i = 0; i < nr_vms; i++)
3346 free_vm_area(vms[i]);
3349 #endif /* CONFIG_SMP */
3351 #ifdef CONFIG_PROC_FS
3352 static void *s_start(struct seq_file *m, loff_t *pos)
3353 __acquires(&vmap_area_lock)
3355 spin_lock(&vmap_area_lock);
3356 return seq_list_start(&vmap_area_list, *pos);
3359 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3361 return seq_list_next(p, &vmap_area_list, pos);
3364 static void s_stop(struct seq_file *m, void *p)
3365 __releases(&vmap_area_lock)
3367 spin_unlock(&vmap_area_lock);
3370 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3372 if (IS_ENABLED(CONFIG_NUMA)) {
3373 unsigned int nr, *counters = m->private;
3378 if (v->flags & VM_UNINITIALIZED)
3380 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3383 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3385 for (nr = 0; nr < v->nr_pages; nr++)
3386 counters[page_to_nid(v->pages[nr])]++;
3388 for_each_node_state(nr, N_HIGH_MEMORY)
3390 seq_printf(m, " N%u=%u", nr, counters[nr]);
3394 static int s_show(struct seq_file *m, void *p)
3396 struct vmap_area *va;
3397 struct vm_struct *v;
3399 va = list_entry(p, struct vmap_area, list);
3402 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
3403 * behalf of vmap area is being tear down or vm_map_ram allocation.
3405 if (!(va->flags & VM_VM_AREA)) {
3406 seq_printf(m, "0x%pK-0x%pK %7ld %s\n",
3407 (void *)va->va_start, (void *)va->va_end,
3408 va->va_end - va->va_start,
3409 va->flags & VM_LAZY_FREE ? "unpurged vm_area" : "vm_map_ram");
3416 seq_printf(m, "0x%pK-0x%pK %7ld",
3417 v->addr, v->addr + v->size, v->size);
3420 seq_printf(m, " %pS", v->caller);
3423 seq_printf(m, " pages=%d", v->nr_pages);
3426 seq_printf(m, " phys=%pa", &v->phys_addr);
3428 if (v->flags & VM_IOREMAP)
3429 seq_puts(m, " ioremap");
3431 if (v->flags & VM_ALLOC)
3432 seq_puts(m, " vmalloc");
3434 if (v->flags & VM_MAP)
3435 seq_puts(m, " vmap");
3437 if (v->flags & VM_USERMAP)
3438 seq_puts(m, " user");
3440 if (is_vmalloc_addr(v->pages))
3441 seq_puts(m, " vpages");
3443 show_numa_info(m, v);
3448 static const struct seq_operations vmalloc_op = {
3455 static int __init proc_vmalloc_init(void)
3457 if (IS_ENABLED(CONFIG_NUMA))
3458 proc_create_seq_private("vmallocinfo", 0400, NULL,
3460 nr_node_ids * sizeof(unsigned int), NULL);
3462 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3465 module_init(proc_vmalloc_init);