1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1993 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
8 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
9 * Numa awareness, Christoph Lameter, SGI, June 2005
12 #include <linux/vmalloc.h>
14 #include <linux/module.h>
15 #include <linux/highmem.h>
16 #include <linux/sched/signal.h>
17 #include <linux/slab.h>
18 #include <linux/spinlock.h>
19 #include <linux/interrupt.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/set_memory.h>
23 #include <linux/debugobjects.h>
24 #include <linux/kallsyms.h>
25 #include <linux/list.h>
26 #include <linux/notifier.h>
27 #include <linux/rbtree.h>
28 #include <linux/radix-tree.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/llist.h>
35 #include <linux/bitops.h>
36 #include <linux/rbtree_augmented.h>
38 #include <linux/uaccess.h>
39 #include <asm/tlbflush.h>
40 #include <asm/shmparam.h>
44 struct vfree_deferred {
45 struct llist_head list;
46 struct work_struct wq;
48 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
50 static void __vunmap(const void *, int);
52 static void free_work(struct work_struct *w)
54 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
55 struct llist_node *t, *llnode;
57 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
58 __vunmap((void *)llnode, 1);
61 /*** Page table manipulation functions ***/
63 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
67 pte = pte_offset_kernel(pmd, addr);
69 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
70 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
71 } while (pte++, addr += PAGE_SIZE, addr != end);
74 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
79 pmd = pmd_offset(pud, addr);
81 next = pmd_addr_end(addr, end);
82 if (pmd_clear_huge(pmd))
84 if (pmd_none_or_clear_bad(pmd))
86 vunmap_pte_range(pmd, addr, next);
87 } while (pmd++, addr = next, addr != end);
90 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
95 pud = pud_offset(p4d, addr);
97 next = pud_addr_end(addr, end);
98 if (pud_clear_huge(pud))
100 if (pud_none_or_clear_bad(pud))
102 vunmap_pmd_range(pud, addr, next);
103 } while (pud++, addr = next, addr != end);
106 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
111 p4d = p4d_offset(pgd, addr);
113 next = p4d_addr_end(addr, end);
114 if (p4d_clear_huge(p4d))
116 if (p4d_none_or_clear_bad(p4d))
118 vunmap_pud_range(p4d, addr, next);
119 } while (p4d++, addr = next, addr != end);
122 static void vunmap_page_range(unsigned long addr, unsigned long end)
128 pgd = pgd_offset_k(addr);
130 next = pgd_addr_end(addr, end);
131 if (pgd_none_or_clear_bad(pgd))
133 vunmap_p4d_range(pgd, addr, next);
134 } while (pgd++, addr = next, addr != end);
137 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
138 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
143 * nr is a running index into the array which helps higher level
144 * callers keep track of where we're up to.
147 pte = pte_alloc_kernel(pmd, addr);
151 struct page *page = pages[*nr];
153 if (WARN_ON(!pte_none(*pte)))
157 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
159 } while (pte++, addr += PAGE_SIZE, addr != end);
163 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
169 pmd = pmd_alloc(&init_mm, pud, addr);
173 next = pmd_addr_end(addr, end);
174 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
176 } while (pmd++, addr = next, addr != end);
180 static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
181 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
186 pud = pud_alloc(&init_mm, p4d, addr);
190 next = pud_addr_end(addr, end);
191 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
193 } while (pud++, addr = next, addr != end);
197 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
198 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
203 p4d = p4d_alloc(&init_mm, pgd, addr);
207 next = p4d_addr_end(addr, end);
208 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
210 } while (p4d++, addr = next, addr != end);
215 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
216 * will have pfns corresponding to the "pages" array.
218 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
220 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
221 pgprot_t prot, struct page **pages)
225 unsigned long addr = start;
230 pgd = pgd_offset_k(addr);
232 next = pgd_addr_end(addr, end);
233 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
236 } while (pgd++, addr = next, addr != end);
241 static int vmap_page_range(unsigned long start, unsigned long end,
242 pgprot_t prot, struct page **pages)
246 ret = vmap_page_range_noflush(start, end, prot, pages);
247 flush_cache_vmap(start, end);
251 int is_vmalloc_or_module_addr(const void *x)
254 * ARM, x86-64 and sparc64 put modules in a special place,
255 * and fall back on vmalloc() if that fails. Others
256 * just put it in the vmalloc space.
258 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
259 unsigned long addr = (unsigned long)x;
260 if (addr >= MODULES_VADDR && addr < MODULES_END)
263 return is_vmalloc_addr(x);
267 * Walk a vmap address to the struct page it maps.
269 struct page *vmalloc_to_page(const void *vmalloc_addr)
271 unsigned long addr = (unsigned long) vmalloc_addr;
272 struct page *page = NULL;
273 pgd_t *pgd = pgd_offset_k(addr);
280 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
281 * architectures that do not vmalloc module space
283 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
287 p4d = p4d_offset(pgd, addr);
290 pud = pud_offset(p4d, addr);
293 * Don't dereference bad PUD or PMD (below) entries. This will also
294 * identify huge mappings, which we may encounter on architectures
295 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
296 * identified as vmalloc addresses by is_vmalloc_addr(), but are
297 * not [unambiguously] associated with a struct page, so there is
298 * no correct value to return for them.
300 WARN_ON_ONCE(pud_bad(*pud));
301 if (pud_none(*pud) || pud_bad(*pud))
303 pmd = pmd_offset(pud, addr);
304 WARN_ON_ONCE(pmd_bad(*pmd));
305 if (pmd_none(*pmd) || pmd_bad(*pmd))
308 ptep = pte_offset_map(pmd, addr);
310 if (pte_present(pte))
311 page = pte_page(pte);
315 EXPORT_SYMBOL(vmalloc_to_page);
318 * Map a vmalloc()-space virtual address to the physical page frame number.
320 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
322 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
324 EXPORT_SYMBOL(vmalloc_to_pfn);
327 /*** Global kva allocator ***/
329 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
330 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
333 static DEFINE_SPINLOCK(vmap_area_lock);
334 /* Export for kexec only */
335 LIST_HEAD(vmap_area_list);
336 static LLIST_HEAD(vmap_purge_list);
337 static struct rb_root vmap_area_root = RB_ROOT;
338 static bool vmap_initialized __read_mostly;
341 * This kmem_cache is used for vmap_area objects. Instead of
342 * allocating from slab we reuse an object from this cache to
343 * make things faster. Especially in "no edge" splitting of
346 static struct kmem_cache *vmap_area_cachep;
349 * This linked list is used in pair with free_vmap_area_root.
350 * It gives O(1) access to prev/next to perform fast coalescing.
352 static LIST_HEAD(free_vmap_area_list);
355 * This augment red-black tree represents the free vmap space.
356 * All vmap_area objects in this tree are sorted by va->va_start
357 * address. It is used for allocation and merging when a vmap
358 * object is released.
360 * Each vmap_area node contains a maximum available free block
361 * of its sub-tree, right or left. Therefore it is possible to
362 * find a lowest match of free area.
364 static struct rb_root free_vmap_area_root = RB_ROOT;
367 * Preload a CPU with one object for "no edge" split case. The
368 * aim is to get rid of allocations from the atomic context, thus
369 * to use more permissive allocation masks.
371 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
373 static __always_inline unsigned long
374 va_size(struct vmap_area *va)
376 return (va->va_end - va->va_start);
379 static __always_inline unsigned long
380 get_subtree_max_size(struct rb_node *node)
382 struct vmap_area *va;
384 va = rb_entry_safe(node, struct vmap_area, rb_node);
385 return va ? va->subtree_max_size : 0;
389 * Gets called when remove the node and rotate.
391 static __always_inline unsigned long
392 compute_subtree_max_size(struct vmap_area *va)
394 return max3(va_size(va),
395 get_subtree_max_size(va->rb_node.rb_left),
396 get_subtree_max_size(va->rb_node.rb_right));
399 RB_DECLARE_CALLBACKS(static, free_vmap_area_rb_augment_cb,
400 struct vmap_area, rb_node, unsigned long, subtree_max_size,
401 compute_subtree_max_size)
403 static void purge_vmap_area_lazy(void);
404 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
405 static unsigned long lazy_max_pages(void);
407 static atomic_long_t nr_vmalloc_pages;
409 unsigned long vmalloc_nr_pages(void)
411 return atomic_long_read(&nr_vmalloc_pages);
414 static struct vmap_area *__find_vmap_area(unsigned long addr)
416 struct rb_node *n = vmap_area_root.rb_node;
419 struct vmap_area *va;
421 va = rb_entry(n, struct vmap_area, rb_node);
422 if (addr < va->va_start)
424 else if (addr >= va->va_end)
434 * This function returns back addresses of parent node
435 * and its left or right link for further processing.
437 static __always_inline struct rb_node **
438 find_va_links(struct vmap_area *va,
439 struct rb_root *root, struct rb_node *from,
440 struct rb_node **parent)
442 struct vmap_area *tmp_va;
443 struct rb_node **link;
446 link = &root->rb_node;
447 if (unlikely(!*link)) {
456 * Go to the bottom of the tree. When we hit the last point
457 * we end up with parent rb_node and correct direction, i name
458 * it link, where the new va->rb_node will be attached to.
461 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
464 * During the traversal we also do some sanity check.
465 * Trigger the BUG() if there are sides(left/right)
468 if (va->va_start < tmp_va->va_end &&
469 va->va_end <= tmp_va->va_start)
470 link = &(*link)->rb_left;
471 else if (va->va_end > tmp_va->va_start &&
472 va->va_start >= tmp_va->va_end)
473 link = &(*link)->rb_right;
478 *parent = &tmp_va->rb_node;
482 static __always_inline struct list_head *
483 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
485 struct list_head *list;
487 if (unlikely(!parent))
489 * The red-black tree where we try to find VA neighbors
490 * before merging or inserting is empty, i.e. it means
491 * there is no free vmap space. Normally it does not
492 * happen but we handle this case anyway.
496 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
497 return (&parent->rb_right == link ? list->next : list);
500 static __always_inline void
501 link_va(struct vmap_area *va, struct rb_root *root,
502 struct rb_node *parent, struct rb_node **link, struct list_head *head)
505 * VA is still not in the list, but we can
506 * identify its future previous list_head node.
508 if (likely(parent)) {
509 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
510 if (&parent->rb_right != link)
514 /* Insert to the rb-tree */
515 rb_link_node(&va->rb_node, parent, link);
516 if (root == &free_vmap_area_root) {
518 * Some explanation here. Just perform simple insertion
519 * to the tree. We do not set va->subtree_max_size to
520 * its current size before calling rb_insert_augmented().
521 * It is because of we populate the tree from the bottom
522 * to parent levels when the node _is_ in the tree.
524 * Therefore we set subtree_max_size to zero after insertion,
525 * to let __augment_tree_propagate_from() puts everything to
526 * the correct order later on.
528 rb_insert_augmented(&va->rb_node,
529 root, &free_vmap_area_rb_augment_cb);
530 va->subtree_max_size = 0;
532 rb_insert_color(&va->rb_node, root);
535 /* Address-sort this list */
536 list_add(&va->list, head);
539 static __always_inline void
540 unlink_va(struct vmap_area *va, struct rb_root *root)
542 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
545 if (root == &free_vmap_area_root)
546 rb_erase_augmented(&va->rb_node,
547 root, &free_vmap_area_rb_augment_cb);
549 rb_erase(&va->rb_node, root);
552 RB_CLEAR_NODE(&va->rb_node);
555 #if DEBUG_AUGMENT_PROPAGATE_CHECK
557 augment_tree_propagate_check(struct rb_node *n)
559 struct vmap_area *va;
560 struct rb_node *node;
567 va = rb_entry(n, struct vmap_area, rb_node);
568 size = va->subtree_max_size;
572 va = rb_entry(node, struct vmap_area, rb_node);
574 if (get_subtree_max_size(node->rb_left) == size) {
575 node = node->rb_left;
577 if (va_size(va) == size) {
582 node = node->rb_right;
587 va = rb_entry(n, struct vmap_area, rb_node);
588 pr_emerg("tree is corrupted: %lu, %lu\n",
589 va_size(va), va->subtree_max_size);
592 augment_tree_propagate_check(n->rb_left);
593 augment_tree_propagate_check(n->rb_right);
598 * This function populates subtree_max_size from bottom to upper
599 * levels starting from VA point. The propagation must be done
600 * when VA size is modified by changing its va_start/va_end. Or
601 * in case of newly inserting of VA to the tree.
603 * It means that __augment_tree_propagate_from() must be called:
604 * - After VA has been inserted to the tree(free path);
605 * - After VA has been shrunk(allocation path);
606 * - After VA has been increased(merging path).
608 * Please note that, it does not mean that upper parent nodes
609 * and their subtree_max_size are recalculated all the time up
618 * For example if we modify the node 4, shrinking it to 2, then
619 * no any modification is required. If we shrink the node 2 to 1
620 * its subtree_max_size is updated only, and set to 1. If we shrink
621 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
624 static __always_inline void
625 augment_tree_propagate_from(struct vmap_area *va)
627 struct rb_node *node = &va->rb_node;
628 unsigned long new_va_sub_max_size;
631 va = rb_entry(node, struct vmap_area, rb_node);
632 new_va_sub_max_size = compute_subtree_max_size(va);
635 * If the newly calculated maximum available size of the
636 * subtree is equal to the current one, then it means that
637 * the tree is propagated correctly. So we have to stop at
638 * this point to save cycles.
640 if (va->subtree_max_size == new_va_sub_max_size)
643 va->subtree_max_size = new_va_sub_max_size;
644 node = rb_parent(&va->rb_node);
647 #if DEBUG_AUGMENT_PROPAGATE_CHECK
648 augment_tree_propagate_check(free_vmap_area_root.rb_node);
653 insert_vmap_area(struct vmap_area *va,
654 struct rb_root *root, struct list_head *head)
656 struct rb_node **link;
657 struct rb_node *parent;
659 link = find_va_links(va, root, NULL, &parent);
660 link_va(va, root, parent, link, head);
664 insert_vmap_area_augment(struct vmap_area *va,
665 struct rb_node *from, struct rb_root *root,
666 struct list_head *head)
668 struct rb_node **link;
669 struct rb_node *parent;
672 link = find_va_links(va, NULL, from, &parent);
674 link = find_va_links(va, root, NULL, &parent);
676 link_va(va, root, parent, link, head);
677 augment_tree_propagate_from(va);
681 * Merge de-allocated chunk of VA memory with previous
682 * and next free blocks. If coalesce is not done a new
683 * free area is inserted. If VA has been merged, it is
686 static __always_inline void
687 merge_or_add_vmap_area(struct vmap_area *va,
688 struct rb_root *root, struct list_head *head)
690 struct vmap_area *sibling;
691 struct list_head *next;
692 struct rb_node **link;
693 struct rb_node *parent;
697 * Find a place in the tree where VA potentially will be
698 * inserted, unless it is merged with its sibling/siblings.
700 link = find_va_links(va, root, NULL, &parent);
703 * Get next node of VA to check if merging can be done.
705 next = get_va_next_sibling(parent, link);
706 if (unlikely(next == NULL))
712 * |<------VA------>|<-----Next----->|
717 sibling = list_entry(next, struct vmap_area, list);
718 if (sibling->va_start == va->va_end) {
719 sibling->va_start = va->va_start;
721 /* Check and update the tree if needed. */
722 augment_tree_propagate_from(sibling);
724 /* Free vmap_area object. */
725 kmem_cache_free(vmap_area_cachep, va);
727 /* Point to the new merged area. */
736 * |<-----Prev----->|<------VA------>|
740 if (next->prev != head) {
741 sibling = list_entry(next->prev, struct vmap_area, list);
742 if (sibling->va_end == va->va_start) {
743 sibling->va_end = va->va_end;
745 /* Check and update the tree if needed. */
746 augment_tree_propagate_from(sibling);
751 /* Free vmap_area object. */
752 kmem_cache_free(vmap_area_cachep, va);
759 link_va(va, root, parent, link, head);
760 augment_tree_propagate_from(va);
764 static __always_inline bool
765 is_within_this_va(struct vmap_area *va, unsigned long size,
766 unsigned long align, unsigned long vstart)
768 unsigned long nva_start_addr;
770 if (va->va_start > vstart)
771 nva_start_addr = ALIGN(va->va_start, align);
773 nva_start_addr = ALIGN(vstart, align);
775 /* Can be overflowed due to big size or alignment. */
776 if (nva_start_addr + size < nva_start_addr ||
777 nva_start_addr < vstart)
780 return (nva_start_addr + size <= va->va_end);
784 * Find the first free block(lowest start address) in the tree,
785 * that will accomplish the request corresponding to passing
788 static __always_inline struct vmap_area *
789 find_vmap_lowest_match(unsigned long size,
790 unsigned long align, unsigned long vstart)
792 struct vmap_area *va;
793 struct rb_node *node;
794 unsigned long length;
796 /* Start from the root. */
797 node = free_vmap_area_root.rb_node;
799 /* Adjust the search size for alignment overhead. */
800 length = size + align - 1;
803 va = rb_entry(node, struct vmap_area, rb_node);
805 if (get_subtree_max_size(node->rb_left) >= length &&
806 vstart < va->va_start) {
807 node = node->rb_left;
809 if (is_within_this_va(va, size, align, vstart))
813 * Does not make sense to go deeper towards the right
814 * sub-tree if it does not have a free block that is
815 * equal or bigger to the requested search length.
817 if (get_subtree_max_size(node->rb_right) >= length) {
818 node = node->rb_right;
823 * OK. We roll back and find the first right sub-tree,
824 * that will satisfy the search criteria. It can happen
825 * only once due to "vstart" restriction.
827 while ((node = rb_parent(node))) {
828 va = rb_entry(node, struct vmap_area, rb_node);
829 if (is_within_this_va(va, size, align, vstart))
832 if (get_subtree_max_size(node->rb_right) >= length &&
833 vstart <= va->va_start) {
834 node = node->rb_right;
844 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
845 #include <linux/random.h>
847 static struct vmap_area *
848 find_vmap_lowest_linear_match(unsigned long size,
849 unsigned long align, unsigned long vstart)
851 struct vmap_area *va;
853 list_for_each_entry(va, &free_vmap_area_list, list) {
854 if (!is_within_this_va(va, size, align, vstart))
864 find_vmap_lowest_match_check(unsigned long size)
866 struct vmap_area *va_1, *va_2;
867 unsigned long vstart;
870 get_random_bytes(&rnd, sizeof(rnd));
871 vstart = VMALLOC_START + rnd;
873 va_1 = find_vmap_lowest_match(size, 1, vstart);
874 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
877 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
884 FL_FIT_TYPE = 1, /* full fit */
885 LE_FIT_TYPE = 2, /* left edge fit */
886 RE_FIT_TYPE = 3, /* right edge fit */
887 NE_FIT_TYPE = 4 /* no edge fit */
890 static __always_inline enum fit_type
891 classify_va_fit_type(struct vmap_area *va,
892 unsigned long nva_start_addr, unsigned long size)
896 /* Check if it is within VA. */
897 if (nva_start_addr < va->va_start ||
898 nva_start_addr + size > va->va_end)
902 if (va->va_start == nva_start_addr) {
903 if (va->va_end == nva_start_addr + size)
907 } else if (va->va_end == nva_start_addr + size) {
916 static __always_inline int
917 adjust_va_to_fit_type(struct vmap_area *va,
918 unsigned long nva_start_addr, unsigned long size,
921 struct vmap_area *lva = NULL;
923 if (type == FL_FIT_TYPE) {
925 * No need to split VA, it fully fits.
931 unlink_va(va, &free_vmap_area_root);
932 kmem_cache_free(vmap_area_cachep, va);
933 } else if (type == LE_FIT_TYPE) {
935 * Split left edge of fit VA.
941 va->va_start += size;
942 } else if (type == RE_FIT_TYPE) {
944 * Split right edge of fit VA.
950 va->va_end = nva_start_addr;
951 } else if (type == NE_FIT_TYPE) {
953 * Split no edge of fit VA.
959 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
960 if (unlikely(!lva)) {
962 * For percpu allocator we do not do any pre-allocation
963 * and leave it as it is. The reason is it most likely
964 * never ends up with NE_FIT_TYPE splitting. In case of
965 * percpu allocations offsets and sizes are aligned to
966 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
967 * are its main fitting cases.
969 * There are a few exceptions though, as an example it is
970 * a first allocation (early boot up) when we have "one"
971 * big free space that has to be split.
973 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
979 * Build the remainder.
981 lva->va_start = va->va_start;
982 lva->va_end = nva_start_addr;
985 * Shrink this VA to remaining size.
987 va->va_start = nva_start_addr + size;
992 if (type != FL_FIT_TYPE) {
993 augment_tree_propagate_from(va);
995 if (lva) /* type == NE_FIT_TYPE */
996 insert_vmap_area_augment(lva, &va->rb_node,
997 &free_vmap_area_root, &free_vmap_area_list);
1004 * Returns a start address of the newly allocated area, if success.
1005 * Otherwise a vend is returned that indicates failure.
1007 static __always_inline unsigned long
1008 __alloc_vmap_area(unsigned long size, unsigned long align,
1009 unsigned long vstart, unsigned long vend)
1011 unsigned long nva_start_addr;
1012 struct vmap_area *va;
1016 va = find_vmap_lowest_match(size, align, vstart);
1020 if (va->va_start > vstart)
1021 nva_start_addr = ALIGN(va->va_start, align);
1023 nva_start_addr = ALIGN(vstart, align);
1025 /* Check the "vend" restriction. */
1026 if (nva_start_addr + size > vend)
1029 /* Classify what we have found. */
1030 type = classify_va_fit_type(va, nva_start_addr, size);
1031 if (WARN_ON_ONCE(type == NOTHING_FIT))
1034 /* Update the free vmap_area. */
1035 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1039 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1040 find_vmap_lowest_match_check(size);
1043 return nva_start_addr;
1047 * Allocate a region of KVA of the specified size and alignment, within the
1050 static struct vmap_area *alloc_vmap_area(unsigned long size,
1051 unsigned long align,
1052 unsigned long vstart, unsigned long vend,
1053 int node, gfp_t gfp_mask)
1055 struct vmap_area *va, *pva;
1060 BUG_ON(offset_in_page(size));
1061 BUG_ON(!is_power_of_2(align));
1063 if (unlikely(!vmap_initialized))
1064 return ERR_PTR(-EBUSY);
1068 va = kmem_cache_alloc_node(vmap_area_cachep,
1069 gfp_mask & GFP_RECLAIM_MASK, node);
1071 return ERR_PTR(-ENOMEM);
1074 * Only scan the relevant parts containing pointers to other objects
1075 * to avoid false negatives.
1077 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
1081 * Preload this CPU with one extra vmap_area object to ensure
1082 * that we have it available when fit type of free area is
1085 * The preload is done in non-atomic context, thus it allows us
1086 * to use more permissive allocation masks to be more stable under
1087 * low memory condition and high memory pressure.
1089 * Even if it fails we do not really care about that. Just proceed
1090 * as it is. "overflow" path will refill the cache we allocate from.
1093 if (!__this_cpu_read(ne_fit_preload_node)) {
1095 pva = kmem_cache_alloc_node(vmap_area_cachep, GFP_KERNEL, node);
1098 if (__this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva)) {
1100 kmem_cache_free(vmap_area_cachep, pva);
1104 spin_lock(&vmap_area_lock);
1108 * If an allocation fails, the "vend" address is
1109 * returned. Therefore trigger the overflow path.
1111 addr = __alloc_vmap_area(size, align, vstart, vend);
1112 if (unlikely(addr == vend))
1115 va->va_start = addr;
1116 va->va_end = addr + size;
1118 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1120 spin_unlock(&vmap_area_lock);
1122 BUG_ON(!IS_ALIGNED(va->va_start, align));
1123 BUG_ON(va->va_start < vstart);
1124 BUG_ON(va->va_end > vend);
1129 spin_unlock(&vmap_area_lock);
1131 purge_vmap_area_lazy();
1136 if (gfpflags_allow_blocking(gfp_mask)) {
1137 unsigned long freed = 0;
1138 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1145 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1146 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1149 kmem_cache_free(vmap_area_cachep, va);
1150 return ERR_PTR(-EBUSY);
1153 int register_vmap_purge_notifier(struct notifier_block *nb)
1155 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1157 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1159 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1161 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1163 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1165 static void __free_vmap_area(struct vmap_area *va)
1168 * Remove from the busy tree/list.
1170 unlink_va(va, &vmap_area_root);
1173 * Merge VA with its neighbors, otherwise just add it.
1175 merge_or_add_vmap_area(va,
1176 &free_vmap_area_root, &free_vmap_area_list);
1180 * Free a region of KVA allocated by alloc_vmap_area
1182 static void free_vmap_area(struct vmap_area *va)
1184 spin_lock(&vmap_area_lock);
1185 __free_vmap_area(va);
1186 spin_unlock(&vmap_area_lock);
1190 * Clear the pagetable entries of a given vmap_area
1192 static void unmap_vmap_area(struct vmap_area *va)
1194 vunmap_page_range(va->va_start, va->va_end);
1198 * lazy_max_pages is the maximum amount of virtual address space we gather up
1199 * before attempting to purge with a TLB flush.
1201 * There is a tradeoff here: a larger number will cover more kernel page tables
1202 * and take slightly longer to purge, but it will linearly reduce the number of
1203 * global TLB flushes that must be performed. It would seem natural to scale
1204 * this number up linearly with the number of CPUs (because vmapping activity
1205 * could also scale linearly with the number of CPUs), however it is likely
1206 * that in practice, workloads might be constrained in other ways that mean
1207 * vmap activity will not scale linearly with CPUs. Also, I want to be
1208 * conservative and not introduce a big latency on huge systems, so go with
1209 * a less aggressive log scale. It will still be an improvement over the old
1210 * code, and it will be simple to change the scale factor if we find that it
1211 * becomes a problem on bigger systems.
1213 static unsigned long lazy_max_pages(void)
1217 log = fls(num_online_cpus());
1219 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1222 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1225 * Serialize vmap purging. There is no actual criticial section protected
1226 * by this look, but we want to avoid concurrent calls for performance
1227 * reasons and to make the pcpu_get_vm_areas more deterministic.
1229 static DEFINE_MUTEX(vmap_purge_lock);
1231 /* for per-CPU blocks */
1232 static void purge_fragmented_blocks_allcpus(void);
1235 * called before a call to iounmap() if the caller wants vm_area_struct's
1236 * immediately freed.
1238 void set_iounmap_nonlazy(void)
1240 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1244 * Purges all lazily-freed vmap areas.
1246 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1248 unsigned long resched_threshold;
1249 struct llist_node *valist;
1250 struct vmap_area *va;
1251 struct vmap_area *n_va;
1253 lockdep_assert_held(&vmap_purge_lock);
1255 valist = llist_del_all(&vmap_purge_list);
1256 if (unlikely(valist == NULL))
1260 * First make sure the mappings are removed from all page-tables
1261 * before they are freed.
1266 * TODO: to calculate a flush range without looping.
1267 * The list can be up to lazy_max_pages() elements.
1269 llist_for_each_entry(va, valist, purge_list) {
1270 if (va->va_start < start)
1271 start = va->va_start;
1272 if (va->va_end > end)
1276 flush_tlb_kernel_range(start, end);
1277 resched_threshold = lazy_max_pages() << 1;
1279 spin_lock(&vmap_area_lock);
1280 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1281 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1284 * Finally insert or merge lazily-freed area. It is
1285 * detached and there is no need to "unlink" it from
1288 merge_or_add_vmap_area(va,
1289 &free_vmap_area_root, &free_vmap_area_list);
1291 atomic_long_sub(nr, &vmap_lazy_nr);
1293 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1294 cond_resched_lock(&vmap_area_lock);
1296 spin_unlock(&vmap_area_lock);
1301 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1302 * is already purging.
1304 static void try_purge_vmap_area_lazy(void)
1306 if (mutex_trylock(&vmap_purge_lock)) {
1307 __purge_vmap_area_lazy(ULONG_MAX, 0);
1308 mutex_unlock(&vmap_purge_lock);
1313 * Kick off a purge of the outstanding lazy areas.
1315 static void purge_vmap_area_lazy(void)
1317 mutex_lock(&vmap_purge_lock);
1318 purge_fragmented_blocks_allcpus();
1319 __purge_vmap_area_lazy(ULONG_MAX, 0);
1320 mutex_unlock(&vmap_purge_lock);
1324 * Free a vmap area, caller ensuring that the area has been unmapped
1325 * and flush_cache_vunmap had been called for the correct range
1328 static void free_vmap_area_noflush(struct vmap_area *va)
1330 unsigned long nr_lazy;
1332 spin_lock(&vmap_area_lock);
1333 unlink_va(va, &vmap_area_root);
1334 spin_unlock(&vmap_area_lock);
1336 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1337 PAGE_SHIFT, &vmap_lazy_nr);
1339 /* After this point, we may free va at any time */
1340 llist_add(&va->purge_list, &vmap_purge_list);
1342 if (unlikely(nr_lazy > lazy_max_pages()))
1343 try_purge_vmap_area_lazy();
1347 * Free and unmap a vmap area
1349 static void free_unmap_vmap_area(struct vmap_area *va)
1351 flush_cache_vunmap(va->va_start, va->va_end);
1352 unmap_vmap_area(va);
1353 if (debug_pagealloc_enabled())
1354 flush_tlb_kernel_range(va->va_start, va->va_end);
1356 free_vmap_area_noflush(va);
1359 static struct vmap_area *find_vmap_area(unsigned long addr)
1361 struct vmap_area *va;
1363 spin_lock(&vmap_area_lock);
1364 va = __find_vmap_area(addr);
1365 spin_unlock(&vmap_area_lock);
1370 /*** Per cpu kva allocator ***/
1373 * vmap space is limited especially on 32 bit architectures. Ensure there is
1374 * room for at least 16 percpu vmap blocks per CPU.
1377 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1378 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1379 * instead (we just need a rough idea)
1381 #if BITS_PER_LONG == 32
1382 #define VMALLOC_SPACE (128UL*1024*1024)
1384 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1387 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1388 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1389 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1390 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1391 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1392 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1393 #define VMAP_BBMAP_BITS \
1394 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1395 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1396 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1398 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1400 struct vmap_block_queue {
1402 struct list_head free;
1407 struct vmap_area *va;
1408 unsigned long free, dirty;
1409 unsigned long dirty_min, dirty_max; /*< dirty range */
1410 struct list_head free_list;
1411 struct rcu_head rcu_head;
1412 struct list_head purge;
1415 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1416 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1419 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1420 * in the free path. Could get rid of this if we change the API to return a
1421 * "cookie" from alloc, to be passed to free. But no big deal yet.
1423 static DEFINE_SPINLOCK(vmap_block_tree_lock);
1424 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
1427 * We should probably have a fallback mechanism to allocate virtual memory
1428 * out of partially filled vmap blocks. However vmap block sizing should be
1429 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1433 static unsigned long addr_to_vb_idx(unsigned long addr)
1435 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1436 addr /= VMAP_BLOCK_SIZE;
1440 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1444 addr = va_start + (pages_off << PAGE_SHIFT);
1445 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1446 return (void *)addr;
1450 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1451 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1452 * @order: how many 2^order pages should be occupied in newly allocated block
1453 * @gfp_mask: flags for the page level allocator
1455 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1457 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1459 struct vmap_block_queue *vbq;
1460 struct vmap_block *vb;
1461 struct vmap_area *va;
1462 unsigned long vb_idx;
1466 node = numa_node_id();
1468 vb = kmalloc_node(sizeof(struct vmap_block),
1469 gfp_mask & GFP_RECLAIM_MASK, node);
1471 return ERR_PTR(-ENOMEM);
1473 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1474 VMALLOC_START, VMALLOC_END,
1478 return ERR_CAST(va);
1481 err = radix_tree_preload(gfp_mask);
1482 if (unlikely(err)) {
1485 return ERR_PTR(err);
1488 vaddr = vmap_block_vaddr(va->va_start, 0);
1489 spin_lock_init(&vb->lock);
1491 /* At least something should be left free */
1492 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1493 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1495 vb->dirty_min = VMAP_BBMAP_BITS;
1497 INIT_LIST_HEAD(&vb->free_list);
1499 vb_idx = addr_to_vb_idx(va->va_start);
1500 spin_lock(&vmap_block_tree_lock);
1501 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
1502 spin_unlock(&vmap_block_tree_lock);
1504 radix_tree_preload_end();
1506 vbq = &get_cpu_var(vmap_block_queue);
1507 spin_lock(&vbq->lock);
1508 list_add_tail_rcu(&vb->free_list, &vbq->free);
1509 spin_unlock(&vbq->lock);
1510 put_cpu_var(vmap_block_queue);
1515 static void free_vmap_block(struct vmap_block *vb)
1517 struct vmap_block *tmp;
1518 unsigned long vb_idx;
1520 vb_idx = addr_to_vb_idx(vb->va->va_start);
1521 spin_lock(&vmap_block_tree_lock);
1522 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
1523 spin_unlock(&vmap_block_tree_lock);
1526 free_vmap_area_noflush(vb->va);
1527 kfree_rcu(vb, rcu_head);
1530 static void purge_fragmented_blocks(int cpu)
1533 struct vmap_block *vb;
1534 struct vmap_block *n_vb;
1535 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1538 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1540 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1543 spin_lock(&vb->lock);
1544 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1545 vb->free = 0; /* prevent further allocs after releasing lock */
1546 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1548 vb->dirty_max = VMAP_BBMAP_BITS;
1549 spin_lock(&vbq->lock);
1550 list_del_rcu(&vb->free_list);
1551 spin_unlock(&vbq->lock);
1552 spin_unlock(&vb->lock);
1553 list_add_tail(&vb->purge, &purge);
1555 spin_unlock(&vb->lock);
1559 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1560 list_del(&vb->purge);
1561 free_vmap_block(vb);
1565 static void purge_fragmented_blocks_allcpus(void)
1569 for_each_possible_cpu(cpu)
1570 purge_fragmented_blocks(cpu);
1573 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1575 struct vmap_block_queue *vbq;
1576 struct vmap_block *vb;
1580 BUG_ON(offset_in_page(size));
1581 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1582 if (WARN_ON(size == 0)) {
1584 * Allocating 0 bytes isn't what caller wants since
1585 * get_order(0) returns funny result. Just warn and terminate
1590 order = get_order(size);
1593 vbq = &get_cpu_var(vmap_block_queue);
1594 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1595 unsigned long pages_off;
1597 spin_lock(&vb->lock);
1598 if (vb->free < (1UL << order)) {
1599 spin_unlock(&vb->lock);
1603 pages_off = VMAP_BBMAP_BITS - vb->free;
1604 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1605 vb->free -= 1UL << order;
1606 if (vb->free == 0) {
1607 spin_lock(&vbq->lock);
1608 list_del_rcu(&vb->free_list);
1609 spin_unlock(&vbq->lock);
1612 spin_unlock(&vb->lock);
1616 put_cpu_var(vmap_block_queue);
1619 /* Allocate new block if nothing was found */
1621 vaddr = new_vmap_block(order, gfp_mask);
1626 static void vb_free(const void *addr, unsigned long size)
1628 unsigned long offset;
1629 unsigned long vb_idx;
1631 struct vmap_block *vb;
1633 BUG_ON(offset_in_page(size));
1634 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1636 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1638 order = get_order(size);
1640 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1641 offset >>= PAGE_SHIFT;
1643 vb_idx = addr_to_vb_idx((unsigned long)addr);
1645 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1649 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1651 if (debug_pagealloc_enabled())
1652 flush_tlb_kernel_range((unsigned long)addr,
1653 (unsigned long)addr + size);
1655 spin_lock(&vb->lock);
1657 /* Expand dirty range */
1658 vb->dirty_min = min(vb->dirty_min, offset);
1659 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1661 vb->dirty += 1UL << order;
1662 if (vb->dirty == VMAP_BBMAP_BITS) {
1664 spin_unlock(&vb->lock);
1665 free_vmap_block(vb);
1667 spin_unlock(&vb->lock);
1670 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1674 if (unlikely(!vmap_initialized))
1679 for_each_possible_cpu(cpu) {
1680 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1681 struct vmap_block *vb;
1684 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1685 spin_lock(&vb->lock);
1687 unsigned long va_start = vb->va->va_start;
1690 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1691 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1693 start = min(s, start);
1698 spin_unlock(&vb->lock);
1703 mutex_lock(&vmap_purge_lock);
1704 purge_fragmented_blocks_allcpus();
1705 if (!__purge_vmap_area_lazy(start, end) && flush)
1706 flush_tlb_kernel_range(start, end);
1707 mutex_unlock(&vmap_purge_lock);
1711 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1713 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1714 * to amortize TLB flushing overheads. What this means is that any page you
1715 * have now, may, in a former life, have been mapped into kernel virtual
1716 * address by the vmap layer and so there might be some CPUs with TLB entries
1717 * still referencing that page (additional to the regular 1:1 kernel mapping).
1719 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1720 * be sure that none of the pages we have control over will have any aliases
1721 * from the vmap layer.
1723 void vm_unmap_aliases(void)
1725 unsigned long start = ULONG_MAX, end = 0;
1728 _vm_unmap_aliases(start, end, flush);
1730 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1733 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1734 * @mem: the pointer returned by vm_map_ram
1735 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1737 void vm_unmap_ram(const void *mem, unsigned int count)
1739 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1740 unsigned long addr = (unsigned long)mem;
1741 struct vmap_area *va;
1745 BUG_ON(addr < VMALLOC_START);
1746 BUG_ON(addr > VMALLOC_END);
1747 BUG_ON(!PAGE_ALIGNED(addr));
1749 if (likely(count <= VMAP_MAX_ALLOC)) {
1750 debug_check_no_locks_freed(mem, size);
1755 va = find_vmap_area(addr);
1757 debug_check_no_locks_freed((void *)va->va_start,
1758 (va->va_end - va->va_start));
1759 free_unmap_vmap_area(va);
1761 EXPORT_SYMBOL(vm_unmap_ram);
1764 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1765 * @pages: an array of pointers to the pages to be mapped
1766 * @count: number of pages
1767 * @node: prefer to allocate data structures on this node
1768 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1770 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1771 * faster than vmap so it's good. But if you mix long-life and short-life
1772 * objects with vm_map_ram(), it could consume lots of address space through
1773 * fragmentation (especially on a 32bit machine). You could see failures in
1774 * the end. Please use this function for short-lived objects.
1776 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1778 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1780 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1784 if (likely(count <= VMAP_MAX_ALLOC)) {
1785 mem = vb_alloc(size, GFP_KERNEL);
1788 addr = (unsigned long)mem;
1790 struct vmap_area *va;
1791 va = alloc_vmap_area(size, PAGE_SIZE,
1792 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1796 addr = va->va_start;
1799 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1800 vm_unmap_ram(mem, count);
1805 EXPORT_SYMBOL(vm_map_ram);
1807 static struct vm_struct *vmlist __initdata;
1810 * vm_area_add_early - add vmap area early during boot
1811 * @vm: vm_struct to add
1813 * This function is used to add fixed kernel vm area to vmlist before
1814 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1815 * should contain proper values and the other fields should be zero.
1817 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1819 void __init vm_area_add_early(struct vm_struct *vm)
1821 struct vm_struct *tmp, **p;
1823 BUG_ON(vmap_initialized);
1824 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1825 if (tmp->addr >= vm->addr) {
1826 BUG_ON(tmp->addr < vm->addr + vm->size);
1829 BUG_ON(tmp->addr + tmp->size > vm->addr);
1836 * vm_area_register_early - register vmap area early during boot
1837 * @vm: vm_struct to register
1838 * @align: requested alignment
1840 * This function is used to register kernel vm area before
1841 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1842 * proper values on entry and other fields should be zero. On return,
1843 * vm->addr contains the allocated address.
1845 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1847 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1849 static size_t vm_init_off __initdata;
1852 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1853 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1855 vm->addr = (void *)addr;
1857 vm_area_add_early(vm);
1860 static void vmap_init_free_space(void)
1862 unsigned long vmap_start = 1;
1863 const unsigned long vmap_end = ULONG_MAX;
1864 struct vmap_area *busy, *free;
1868 * -|-----|.....|-----|-----|-----|.....|-
1870 * |<--------------------------------->|
1872 list_for_each_entry(busy, &vmap_area_list, list) {
1873 if (busy->va_start - vmap_start > 0) {
1874 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1875 if (!WARN_ON_ONCE(!free)) {
1876 free->va_start = vmap_start;
1877 free->va_end = busy->va_start;
1879 insert_vmap_area_augment(free, NULL,
1880 &free_vmap_area_root,
1881 &free_vmap_area_list);
1885 vmap_start = busy->va_end;
1888 if (vmap_end - vmap_start > 0) {
1889 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1890 if (!WARN_ON_ONCE(!free)) {
1891 free->va_start = vmap_start;
1892 free->va_end = vmap_end;
1894 insert_vmap_area_augment(free, NULL,
1895 &free_vmap_area_root,
1896 &free_vmap_area_list);
1901 void __init vmalloc_init(void)
1903 struct vmap_area *va;
1904 struct vm_struct *tmp;
1908 * Create the cache for vmap_area objects.
1910 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1912 for_each_possible_cpu(i) {
1913 struct vmap_block_queue *vbq;
1914 struct vfree_deferred *p;
1916 vbq = &per_cpu(vmap_block_queue, i);
1917 spin_lock_init(&vbq->lock);
1918 INIT_LIST_HEAD(&vbq->free);
1919 p = &per_cpu(vfree_deferred, i);
1920 init_llist_head(&p->list);
1921 INIT_WORK(&p->wq, free_work);
1924 /* Import existing vmlist entries. */
1925 for (tmp = vmlist; tmp; tmp = tmp->next) {
1926 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1927 if (WARN_ON_ONCE(!va))
1930 va->va_start = (unsigned long)tmp->addr;
1931 va->va_end = va->va_start + tmp->size;
1933 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1937 * Now we can initialize a free vmap space.
1939 vmap_init_free_space();
1940 vmap_initialized = true;
1944 * map_kernel_range_noflush - map kernel VM area with the specified pages
1945 * @addr: start of the VM area to map
1946 * @size: size of the VM area to map
1947 * @prot: page protection flags to use
1948 * @pages: pages to map
1950 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1951 * specify should have been allocated using get_vm_area() and its
1955 * This function does NOT do any cache flushing. The caller is
1956 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1957 * before calling this function.
1960 * The number of pages mapped on success, -errno on failure.
1962 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1963 pgprot_t prot, struct page **pages)
1965 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1969 * unmap_kernel_range_noflush - unmap kernel VM area
1970 * @addr: start of the VM area to unmap
1971 * @size: size of the VM area to unmap
1973 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1974 * specify should have been allocated using get_vm_area() and its
1978 * This function does NOT do any cache flushing. The caller is
1979 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1980 * before calling this function and flush_tlb_kernel_range() after.
1982 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1984 vunmap_page_range(addr, addr + size);
1986 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1989 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1990 * @addr: start of the VM area to unmap
1991 * @size: size of the VM area to unmap
1993 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1994 * the unmapping and tlb after.
1996 void unmap_kernel_range(unsigned long addr, unsigned long size)
1998 unsigned long end = addr + size;
2000 flush_cache_vunmap(addr, end);
2001 vunmap_page_range(addr, end);
2002 flush_tlb_kernel_range(addr, end);
2004 EXPORT_SYMBOL_GPL(unmap_kernel_range);
2006 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
2008 unsigned long addr = (unsigned long)area->addr;
2009 unsigned long end = addr + get_vm_area_size(area);
2012 err = vmap_page_range(addr, end, prot, pages);
2014 return err > 0 ? 0 : err;
2016 EXPORT_SYMBOL_GPL(map_vm_area);
2018 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2019 unsigned long flags, const void *caller)
2021 spin_lock(&vmap_area_lock);
2023 vm->addr = (void *)va->va_start;
2024 vm->size = va->va_end - va->va_start;
2025 vm->caller = caller;
2027 spin_unlock(&vmap_area_lock);
2030 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2033 * Before removing VM_UNINITIALIZED,
2034 * we should make sure that vm has proper values.
2035 * Pair with smp_rmb() in show_numa_info().
2038 vm->flags &= ~VM_UNINITIALIZED;
2041 static struct vm_struct *__get_vm_area_node(unsigned long size,
2042 unsigned long align, unsigned long flags, unsigned long start,
2043 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2045 struct vmap_area *va;
2046 struct vm_struct *area;
2048 BUG_ON(in_interrupt());
2049 size = PAGE_ALIGN(size);
2050 if (unlikely(!size))
2053 if (flags & VM_IOREMAP)
2054 align = 1ul << clamp_t(int, get_count_order_long(size),
2055 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2057 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2058 if (unlikely(!area))
2061 if (!(flags & VM_NO_GUARD))
2064 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2070 setup_vmalloc_vm(area, va, flags, caller);
2075 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
2076 unsigned long start, unsigned long end)
2078 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2079 GFP_KERNEL, __builtin_return_address(0));
2081 EXPORT_SYMBOL_GPL(__get_vm_area);
2083 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2084 unsigned long start, unsigned long end,
2087 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2088 GFP_KERNEL, caller);
2092 * get_vm_area - reserve a contiguous kernel virtual area
2093 * @size: size of the area
2094 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2096 * Search an area of @size in the kernel virtual mapping area,
2097 * and reserved it for out purposes. Returns the area descriptor
2098 * on success or %NULL on failure.
2100 * Return: the area descriptor on success or %NULL on failure.
2102 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2104 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2105 NUMA_NO_NODE, GFP_KERNEL,
2106 __builtin_return_address(0));
2109 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2112 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2113 NUMA_NO_NODE, GFP_KERNEL, caller);
2117 * find_vm_area - find a continuous kernel virtual area
2118 * @addr: base address
2120 * Search for the kernel VM area starting at @addr, and return it.
2121 * It is up to the caller to do all required locking to keep the returned
2124 * Return: pointer to the found area or %NULL on faulure
2126 struct vm_struct *find_vm_area(const void *addr)
2128 struct vmap_area *va;
2130 va = find_vmap_area((unsigned long)addr);
2138 * remove_vm_area - find and remove a continuous kernel virtual area
2139 * @addr: base address
2141 * Search for the kernel VM area starting at @addr, and remove it.
2142 * This function returns the found VM area, but using it is NOT safe
2143 * on SMP machines, except for its size or flags.
2145 * Return: pointer to the found area or %NULL on faulure
2147 struct vm_struct *remove_vm_area(const void *addr)
2149 struct vmap_area *va;
2153 spin_lock(&vmap_area_lock);
2154 va = __find_vmap_area((unsigned long)addr);
2156 struct vm_struct *vm = va->vm;
2159 spin_unlock(&vmap_area_lock);
2161 kasan_free_shadow(vm);
2162 free_unmap_vmap_area(va);
2167 spin_unlock(&vmap_area_lock);
2171 static inline void set_area_direct_map(const struct vm_struct *area,
2172 int (*set_direct_map)(struct page *page))
2176 for (i = 0; i < area->nr_pages; i++)
2177 if (page_address(area->pages[i]))
2178 set_direct_map(area->pages[i]);
2181 /* Handle removing and resetting vm mappings related to the vm_struct. */
2182 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2184 unsigned long start = ULONG_MAX, end = 0;
2185 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2189 remove_vm_area(area->addr);
2191 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2196 * If not deallocating pages, just do the flush of the VM area and
2199 if (!deallocate_pages) {
2205 * If execution gets here, flush the vm mapping and reset the direct
2206 * map. Find the start and end range of the direct mappings to make sure
2207 * the vm_unmap_aliases() flush includes the direct map.
2209 for (i = 0; i < area->nr_pages; i++) {
2210 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2212 start = min(addr, start);
2213 end = max(addr + PAGE_SIZE, end);
2219 * Set direct map to something invalid so that it won't be cached if
2220 * there are any accesses after the TLB flush, then flush the TLB and
2221 * reset the direct map permissions to the default.
2223 set_area_direct_map(area, set_direct_map_invalid_noflush);
2224 _vm_unmap_aliases(start, end, flush_dmap);
2225 set_area_direct_map(area, set_direct_map_default_noflush);
2228 static void __vunmap(const void *addr, int deallocate_pages)
2230 struct vm_struct *area;
2235 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2239 area = find_vm_area(addr);
2240 if (unlikely(!area)) {
2241 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2246 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2247 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2249 vm_remove_mappings(area, deallocate_pages);
2251 if (deallocate_pages) {
2254 for (i = 0; i < area->nr_pages; i++) {
2255 struct page *page = area->pages[i];
2258 __free_pages(page, 0);
2260 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2262 kvfree(area->pages);
2269 static inline void __vfree_deferred(const void *addr)
2272 * Use raw_cpu_ptr() because this can be called from preemptible
2273 * context. Preemption is absolutely fine here, because the llist_add()
2274 * implementation is lockless, so it works even if we are adding to
2275 * nother cpu's list. schedule_work() should be fine with this too.
2277 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2279 if (llist_add((struct llist_node *)addr, &p->list))
2280 schedule_work(&p->wq);
2284 * vfree_atomic - release memory allocated by vmalloc()
2285 * @addr: memory base address
2287 * This one is just like vfree() but can be called in any atomic context
2290 void vfree_atomic(const void *addr)
2294 kmemleak_free(addr);
2298 __vfree_deferred(addr);
2301 static void __vfree(const void *addr)
2303 if (unlikely(in_interrupt()))
2304 __vfree_deferred(addr);
2310 * vfree - release memory allocated by vmalloc()
2311 * @addr: memory base address
2313 * Free the virtually continuous memory area starting at @addr, as
2314 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2315 * NULL, no operation is performed.
2317 * Must not be called in NMI context (strictly speaking, only if we don't
2318 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2319 * conventions for vfree() arch-depenedent would be a really bad idea)
2321 * May sleep if called *not* from interrupt context.
2323 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2325 void vfree(const void *addr)
2329 kmemleak_free(addr);
2331 might_sleep_if(!in_interrupt());
2338 EXPORT_SYMBOL(vfree);
2341 * vunmap - release virtual mapping obtained by vmap()
2342 * @addr: memory base address
2344 * Free the virtually contiguous memory area starting at @addr,
2345 * which was created from the page array passed to vmap().
2347 * Must not be called in interrupt context.
2349 void vunmap(const void *addr)
2351 BUG_ON(in_interrupt());
2356 EXPORT_SYMBOL(vunmap);
2359 * vmap - map an array of pages into virtually contiguous space
2360 * @pages: array of page pointers
2361 * @count: number of pages to map
2362 * @flags: vm_area->flags
2363 * @prot: page protection for the mapping
2365 * Maps @count pages from @pages into contiguous kernel virtual
2368 * Return: the address of the area or %NULL on failure
2370 void *vmap(struct page **pages, unsigned int count,
2371 unsigned long flags, pgprot_t prot)
2373 struct vm_struct *area;
2374 unsigned long size; /* In bytes */
2378 if (count > totalram_pages())
2381 size = (unsigned long)count << PAGE_SHIFT;
2382 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2386 if (map_vm_area(area, prot, pages)) {
2393 EXPORT_SYMBOL(vmap);
2395 static void *__vmalloc_node(unsigned long size, unsigned long align,
2396 gfp_t gfp_mask, pgprot_t prot,
2397 int node, const void *caller);
2398 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2399 pgprot_t prot, int node)
2401 struct page **pages;
2402 unsigned int nr_pages, array_size, i;
2403 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2404 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2405 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2409 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2410 array_size = (nr_pages * sizeof(struct page *));
2412 area->nr_pages = nr_pages;
2413 /* Please note that the recursion is strictly bounded. */
2414 if (array_size > PAGE_SIZE) {
2415 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2416 PAGE_KERNEL, node, area->caller);
2418 pages = kmalloc_node(array_size, nested_gfp, node);
2420 area->pages = pages;
2422 remove_vm_area(area->addr);
2427 for (i = 0; i < area->nr_pages; i++) {
2430 if (node == NUMA_NO_NODE)
2431 page = alloc_page(alloc_mask|highmem_mask);
2433 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2435 if (unlikely(!page)) {
2436 /* Successfully allocated i pages, free them in __vunmap() */
2438 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2441 area->pages[i] = page;
2442 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
2445 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2447 if (map_vm_area(area, prot, pages))
2452 warn_alloc(gfp_mask, NULL,
2453 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2454 (area->nr_pages*PAGE_SIZE), area->size);
2455 __vfree(area->addr);
2460 * __vmalloc_node_range - allocate virtually contiguous memory
2461 * @size: allocation size
2462 * @align: desired alignment
2463 * @start: vm area range start
2464 * @end: vm area range end
2465 * @gfp_mask: flags for the page level allocator
2466 * @prot: protection mask for the allocated pages
2467 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2468 * @node: node to use for allocation or NUMA_NO_NODE
2469 * @caller: caller's return address
2471 * Allocate enough pages to cover @size from the page level
2472 * allocator with @gfp_mask flags. Map them into contiguous
2473 * kernel virtual space, using a pagetable protection of @prot.
2475 * Return: the address of the area or %NULL on failure
2477 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2478 unsigned long start, unsigned long end, gfp_t gfp_mask,
2479 pgprot_t prot, unsigned long vm_flags, int node,
2482 struct vm_struct *area;
2484 unsigned long real_size = size;
2486 size = PAGE_ALIGN(size);
2487 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2490 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
2491 vm_flags, start, end, node, gfp_mask, caller);
2495 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2500 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2501 * flag. It means that vm_struct is not fully initialized.
2502 * Now, it is fully initialized, so remove this flag here.
2504 clear_vm_uninitialized_flag(area);
2506 kmemleak_vmalloc(area, size, gfp_mask);
2511 warn_alloc(gfp_mask, NULL,
2512 "vmalloc: allocation failure: %lu bytes", real_size);
2517 * This is only for performance analysis of vmalloc and stress purpose.
2518 * It is required by vmalloc test module, therefore do not use it other
2521 #ifdef CONFIG_TEST_VMALLOC_MODULE
2522 EXPORT_SYMBOL_GPL(__vmalloc_node_range);
2526 * __vmalloc_node - allocate virtually contiguous memory
2527 * @size: allocation size
2528 * @align: desired alignment
2529 * @gfp_mask: flags for the page level allocator
2530 * @prot: protection mask for the allocated pages
2531 * @node: node to use for allocation or NUMA_NO_NODE
2532 * @caller: caller's return address
2534 * Allocate enough pages to cover @size from the page level
2535 * allocator with @gfp_mask flags. Map them into contiguous
2536 * kernel virtual space, using a pagetable protection of @prot.
2538 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2539 * and __GFP_NOFAIL are not supported
2541 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2544 * Return: pointer to the allocated memory or %NULL on error
2546 static void *__vmalloc_node(unsigned long size, unsigned long align,
2547 gfp_t gfp_mask, pgprot_t prot,
2548 int node, const void *caller)
2550 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2551 gfp_mask, prot, 0, node, caller);
2554 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
2556 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
2557 __builtin_return_address(0));
2559 EXPORT_SYMBOL(__vmalloc);
2561 static inline void *__vmalloc_node_flags(unsigned long size,
2562 int node, gfp_t flags)
2564 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
2565 node, __builtin_return_address(0));
2569 void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
2572 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
2576 * vmalloc - allocate virtually contiguous memory
2577 * @size: allocation size
2579 * Allocate enough pages to cover @size from the page level
2580 * allocator and map them into contiguous kernel virtual space.
2582 * For tight control over page level allocator and protection flags
2583 * use __vmalloc() instead.
2585 * Return: pointer to the allocated memory or %NULL on error
2587 void *vmalloc(unsigned long size)
2589 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2592 EXPORT_SYMBOL(vmalloc);
2595 * vzalloc - allocate virtually contiguous memory with zero fill
2596 * @size: allocation size
2598 * Allocate enough pages to cover @size from the page level
2599 * allocator and map them into contiguous kernel virtual space.
2600 * The memory allocated is set to zero.
2602 * For tight control over page level allocator and protection flags
2603 * use __vmalloc() instead.
2605 * Return: pointer to the allocated memory or %NULL on error
2607 void *vzalloc(unsigned long size)
2609 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2610 GFP_KERNEL | __GFP_ZERO);
2612 EXPORT_SYMBOL(vzalloc);
2615 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2616 * @size: allocation size
2618 * The resulting memory area is zeroed so it can be mapped to userspace
2619 * without leaking data.
2621 * Return: pointer to the allocated memory or %NULL on error
2623 void *vmalloc_user(unsigned long size)
2625 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2626 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2627 VM_USERMAP, NUMA_NO_NODE,
2628 __builtin_return_address(0));
2630 EXPORT_SYMBOL(vmalloc_user);
2633 * vmalloc_node - allocate memory on a specific node
2634 * @size: allocation size
2637 * Allocate enough pages to cover @size from the page level
2638 * allocator and map them into contiguous kernel virtual space.
2640 * For tight control over page level allocator and protection flags
2641 * use __vmalloc() instead.
2643 * Return: pointer to the allocated memory or %NULL on error
2645 void *vmalloc_node(unsigned long size, int node)
2647 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
2648 node, __builtin_return_address(0));
2650 EXPORT_SYMBOL(vmalloc_node);
2653 * vzalloc_node - allocate memory on a specific node with zero fill
2654 * @size: allocation size
2657 * Allocate enough pages to cover @size from the page level
2658 * allocator and map them into contiguous kernel virtual space.
2659 * The memory allocated is set to zero.
2661 * For tight control over page level allocator and protection flags
2662 * use __vmalloc_node() instead.
2664 * Return: pointer to the allocated memory or %NULL on error
2666 void *vzalloc_node(unsigned long size, int node)
2668 return __vmalloc_node_flags(size, node,
2669 GFP_KERNEL | __GFP_ZERO);
2671 EXPORT_SYMBOL(vzalloc_node);
2674 * vmalloc_exec - allocate virtually contiguous, executable memory
2675 * @size: allocation size
2677 * Kernel-internal function to allocate enough pages to cover @size
2678 * the page level allocator and map them into contiguous and
2679 * executable kernel virtual space.
2681 * For tight control over page level allocator and protection flags
2682 * use __vmalloc() instead.
2684 * Return: pointer to the allocated memory or %NULL on error
2686 void *vmalloc_exec(unsigned long size)
2688 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
2689 GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
2690 NUMA_NO_NODE, __builtin_return_address(0));
2693 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2694 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2695 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2696 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2699 * 64b systems should always have either DMA or DMA32 zones. For others
2700 * GFP_DMA32 should do the right thing and use the normal zone.
2702 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2706 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2707 * @size: allocation size
2709 * Allocate enough 32bit PA addressable pages to cover @size from the
2710 * page level allocator and map them into contiguous kernel virtual space.
2712 * Return: pointer to the allocated memory or %NULL on error
2714 void *vmalloc_32(unsigned long size)
2716 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
2717 NUMA_NO_NODE, __builtin_return_address(0));
2719 EXPORT_SYMBOL(vmalloc_32);
2722 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2723 * @size: allocation size
2725 * The resulting memory area is 32bit addressable and zeroed so it can be
2726 * mapped to userspace without leaking data.
2728 * Return: pointer to the allocated memory or %NULL on error
2730 void *vmalloc_32_user(unsigned long size)
2732 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2733 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2734 VM_USERMAP, NUMA_NO_NODE,
2735 __builtin_return_address(0));
2737 EXPORT_SYMBOL(vmalloc_32_user);
2740 * small helper routine , copy contents to buf from addr.
2741 * If the page is not present, fill zero.
2744 static int aligned_vread(char *buf, char *addr, unsigned long count)
2750 unsigned long offset, length;
2752 offset = offset_in_page(addr);
2753 length = PAGE_SIZE - offset;
2756 p = vmalloc_to_page(addr);
2758 * To do safe access to this _mapped_ area, we need
2759 * lock. But adding lock here means that we need to add
2760 * overhead of vmalloc()/vfree() calles for this _debug_
2761 * interface, rarely used. Instead of that, we'll use
2762 * kmap() and get small overhead in this access function.
2766 * we can expect USER0 is not used (see vread/vwrite's
2767 * function description)
2769 void *map = kmap_atomic(p);
2770 memcpy(buf, map + offset, length);
2773 memset(buf, 0, length);
2783 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2789 unsigned long offset, length;
2791 offset = offset_in_page(addr);
2792 length = PAGE_SIZE - offset;
2795 p = vmalloc_to_page(addr);
2797 * To do safe access to this _mapped_ area, we need
2798 * lock. But adding lock here means that we need to add
2799 * overhead of vmalloc()/vfree() calles for this _debug_
2800 * interface, rarely used. Instead of that, we'll use
2801 * kmap() and get small overhead in this access function.
2805 * we can expect USER0 is not used (see vread/vwrite's
2806 * function description)
2808 void *map = kmap_atomic(p);
2809 memcpy(map + offset, buf, length);
2821 * vread() - read vmalloc area in a safe way.
2822 * @buf: buffer for reading data
2823 * @addr: vm address.
2824 * @count: number of bytes to be read.
2826 * This function checks that addr is a valid vmalloc'ed area, and
2827 * copy data from that area to a given buffer. If the given memory range
2828 * of [addr...addr+count) includes some valid address, data is copied to
2829 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2830 * IOREMAP area is treated as memory hole and no copy is done.
2832 * If [addr...addr+count) doesn't includes any intersects with alive
2833 * vm_struct area, returns 0. @buf should be kernel's buffer.
2835 * Note: In usual ops, vread() is never necessary because the caller
2836 * should know vmalloc() area is valid and can use memcpy().
2837 * This is for routines which have to access vmalloc area without
2838 * any information, as /dev/kmem.
2840 * Return: number of bytes for which addr and buf should be increased
2841 * (same number as @count) or %0 if [addr...addr+count) doesn't
2842 * include any intersection with valid vmalloc area
2844 long vread(char *buf, char *addr, unsigned long count)
2846 struct vmap_area *va;
2847 struct vm_struct *vm;
2848 char *vaddr, *buf_start = buf;
2849 unsigned long buflen = count;
2852 /* Don't allow overflow */
2853 if ((unsigned long) addr + count < count)
2854 count = -(unsigned long) addr;
2856 spin_lock(&vmap_area_lock);
2857 list_for_each_entry(va, &vmap_area_list, list) {
2865 vaddr = (char *) vm->addr;
2866 if (addr >= vaddr + get_vm_area_size(vm))
2868 while (addr < vaddr) {
2876 n = vaddr + get_vm_area_size(vm) - addr;
2879 if (!(vm->flags & VM_IOREMAP))
2880 aligned_vread(buf, addr, n);
2881 else /* IOREMAP area is treated as memory hole */
2888 spin_unlock(&vmap_area_lock);
2890 if (buf == buf_start)
2892 /* zero-fill memory holes */
2893 if (buf != buf_start + buflen)
2894 memset(buf, 0, buflen - (buf - buf_start));
2900 * vwrite() - write vmalloc area in a safe way.
2901 * @buf: buffer for source data
2902 * @addr: vm address.
2903 * @count: number of bytes to be read.
2905 * This function checks that addr is a valid vmalloc'ed area, and
2906 * copy data from a buffer to the given addr. If specified range of
2907 * [addr...addr+count) includes some valid address, data is copied from
2908 * proper area of @buf. If there are memory holes, no copy to hole.
2909 * IOREMAP area is treated as memory hole and no copy is done.
2911 * If [addr...addr+count) doesn't includes any intersects with alive
2912 * vm_struct area, returns 0. @buf should be kernel's buffer.
2914 * Note: In usual ops, vwrite() is never necessary because the caller
2915 * should know vmalloc() area is valid and can use memcpy().
2916 * This is for routines which have to access vmalloc area without
2917 * any information, as /dev/kmem.
2919 * Return: number of bytes for which addr and buf should be
2920 * increased (same number as @count) or %0 if [addr...addr+count)
2921 * doesn't include any intersection with valid vmalloc area
2923 long vwrite(char *buf, char *addr, unsigned long count)
2925 struct vmap_area *va;
2926 struct vm_struct *vm;
2928 unsigned long n, buflen;
2931 /* Don't allow overflow */
2932 if ((unsigned long) addr + count < count)
2933 count = -(unsigned long) addr;
2936 spin_lock(&vmap_area_lock);
2937 list_for_each_entry(va, &vmap_area_list, list) {
2945 vaddr = (char *) vm->addr;
2946 if (addr >= vaddr + get_vm_area_size(vm))
2948 while (addr < vaddr) {
2955 n = vaddr + get_vm_area_size(vm) - addr;
2958 if (!(vm->flags & VM_IOREMAP)) {
2959 aligned_vwrite(buf, addr, n);
2967 spin_unlock(&vmap_area_lock);
2974 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2975 * @vma: vma to cover
2976 * @uaddr: target user address to start at
2977 * @kaddr: virtual address of vmalloc kernel memory
2978 * @size: size of map area
2980 * Returns: 0 for success, -Exxx on failure
2982 * This function checks that @kaddr is a valid vmalloc'ed area,
2983 * and that it is big enough to cover the range starting at
2984 * @uaddr in @vma. Will return failure if that criteria isn't
2987 * Similar to remap_pfn_range() (see mm/memory.c)
2989 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2990 void *kaddr, unsigned long size)
2992 struct vm_struct *area;
2994 size = PAGE_ALIGN(size);
2996 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2999 area = find_vm_area(kaddr);
3003 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3006 if (kaddr + size > area->addr + get_vm_area_size(area))
3010 struct page *page = vmalloc_to_page(kaddr);
3013 ret = vm_insert_page(vma, uaddr, page);
3022 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3026 EXPORT_SYMBOL(remap_vmalloc_range_partial);
3029 * remap_vmalloc_range - map vmalloc pages to userspace
3030 * @vma: vma to cover (map full range of vma)
3031 * @addr: vmalloc memory
3032 * @pgoff: number of pages into addr before first page to map
3034 * Returns: 0 for success, -Exxx on failure
3036 * This function checks that addr is a valid vmalloc'ed area, and
3037 * that it is big enough to cover the vma. Will return failure if
3038 * that criteria isn't met.
3040 * Similar to remap_pfn_range() (see mm/memory.c)
3042 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3043 unsigned long pgoff)
3045 return remap_vmalloc_range_partial(vma, vma->vm_start,
3046 addr + (pgoff << PAGE_SHIFT),
3047 vma->vm_end - vma->vm_start);
3049 EXPORT_SYMBOL(remap_vmalloc_range);
3052 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3055 * The purpose of this function is to make sure the vmalloc area
3056 * mappings are identical in all page-tables in the system.
3058 void __weak vmalloc_sync_all(void)
3063 static int f(pte_t *pte, unsigned long addr, void *data)
3075 * alloc_vm_area - allocate a range of kernel address space
3076 * @size: size of the area
3077 * @ptes: returns the PTEs for the address space
3079 * Returns: NULL on failure, vm_struct on success
3081 * This function reserves a range of kernel address space, and
3082 * allocates pagetables to map that range. No actual mappings
3085 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3086 * allocated for the VM area are returned.
3088 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3090 struct vm_struct *area;
3092 area = get_vm_area_caller(size, VM_IOREMAP,
3093 __builtin_return_address(0));
3098 * This ensures that page tables are constructed for this region
3099 * of kernel virtual address space and mapped into init_mm.
3101 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3102 size, f, ptes ? &ptes : NULL)) {
3109 EXPORT_SYMBOL_GPL(alloc_vm_area);
3111 void free_vm_area(struct vm_struct *area)
3113 struct vm_struct *ret;
3114 ret = remove_vm_area(area->addr);
3115 BUG_ON(ret != area);
3118 EXPORT_SYMBOL_GPL(free_vm_area);
3121 static struct vmap_area *node_to_va(struct rb_node *n)
3123 return rb_entry_safe(n, struct vmap_area, rb_node);
3127 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3128 * @addr: target address
3130 * Returns: vmap_area if it is found. If there is no such area
3131 * the first highest(reverse order) vmap_area is returned
3132 * i.e. va->va_start < addr && va->va_end < addr or NULL
3133 * if there are no any areas before @addr.
3135 static struct vmap_area *
3136 pvm_find_va_enclose_addr(unsigned long addr)
3138 struct vmap_area *va, *tmp;
3141 n = free_vmap_area_root.rb_node;
3145 tmp = rb_entry(n, struct vmap_area, rb_node);
3146 if (tmp->va_start <= addr) {
3148 if (tmp->va_end >= addr)
3161 * pvm_determine_end_from_reverse - find the highest aligned address
3162 * of free block below VMALLOC_END
3164 * in - the VA we start the search(reverse order);
3165 * out - the VA with the highest aligned end address.
3167 * Returns: determined end address within vmap_area
3169 static unsigned long
3170 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3172 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3176 list_for_each_entry_from_reverse((*va),
3177 &free_vmap_area_list, list) {
3178 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3179 if ((*va)->va_start < addr)
3188 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3189 * @offsets: array containing offset of each area
3190 * @sizes: array containing size of each area
3191 * @nr_vms: the number of areas to allocate
3192 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3194 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3195 * vm_structs on success, %NULL on failure
3197 * Percpu allocator wants to use congruent vm areas so that it can
3198 * maintain the offsets among percpu areas. This function allocates
3199 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3200 * be scattered pretty far, distance between two areas easily going up
3201 * to gigabytes. To avoid interacting with regular vmallocs, these
3202 * areas are allocated from top.
3204 * Despite its complicated look, this allocator is rather simple. It
3205 * does everything top-down and scans free blocks from the end looking
3206 * for matching base. While scanning, if any of the areas do not fit the
3207 * base address is pulled down to fit the area. Scanning is repeated till
3208 * all the areas fit and then all necessary data structures are inserted
3209 * and the result is returned.
3211 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3212 const size_t *sizes, int nr_vms,
3215 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3216 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3217 struct vmap_area **vas, *va;
3218 struct vm_struct **vms;
3219 int area, area2, last_area, term_area;
3220 unsigned long base, start, size, end, last_end;
3221 bool purged = false;
3224 /* verify parameters and allocate data structures */
3225 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3226 for (last_area = 0, area = 0; area < nr_vms; area++) {
3227 start = offsets[area];
3228 end = start + sizes[area];
3230 /* is everything aligned properly? */
3231 BUG_ON(!IS_ALIGNED(offsets[area], align));
3232 BUG_ON(!IS_ALIGNED(sizes[area], align));
3234 /* detect the area with the highest address */
3235 if (start > offsets[last_area])
3238 for (area2 = area + 1; area2 < nr_vms; area2++) {
3239 unsigned long start2 = offsets[area2];
3240 unsigned long end2 = start2 + sizes[area2];
3242 BUG_ON(start2 < end && start < end2);
3245 last_end = offsets[last_area] + sizes[last_area];
3247 if (vmalloc_end - vmalloc_start < last_end) {
3252 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3253 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3257 for (area = 0; area < nr_vms; area++) {
3258 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3259 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3260 if (!vas[area] || !vms[area])
3264 spin_lock(&vmap_area_lock);
3266 /* start scanning - we scan from the top, begin with the last area */
3267 area = term_area = last_area;
3268 start = offsets[area];
3269 end = start + sizes[area];
3271 va = pvm_find_va_enclose_addr(vmalloc_end);
3272 base = pvm_determine_end_from_reverse(&va, align) - end;
3276 * base might have underflowed, add last_end before
3279 if (base + last_end < vmalloc_start + last_end)
3283 * Fitting base has not been found.
3289 * If required width exeeds current VA block, move
3290 * base downwards and then recheck.
3292 if (base + end > va->va_end) {
3293 base = pvm_determine_end_from_reverse(&va, align) - end;
3299 * If this VA does not fit, move base downwards and recheck.
3301 if (base + start < va->va_start) {
3302 va = node_to_va(rb_prev(&va->rb_node));
3303 base = pvm_determine_end_from_reverse(&va, align) - end;
3309 * This area fits, move on to the previous one. If
3310 * the previous one is the terminal one, we're done.
3312 area = (area + nr_vms - 1) % nr_vms;
3313 if (area == term_area)
3316 start = offsets[area];
3317 end = start + sizes[area];
3318 va = pvm_find_va_enclose_addr(base + end);
3321 /* we've found a fitting base, insert all va's */
3322 for (area = 0; area < nr_vms; area++) {
3325 start = base + offsets[area];
3328 va = pvm_find_va_enclose_addr(start);
3329 if (WARN_ON_ONCE(va == NULL))
3330 /* It is a BUG(), but trigger recovery instead. */
3333 type = classify_va_fit_type(va, start, size);
3334 if (WARN_ON_ONCE(type == NOTHING_FIT))
3335 /* It is a BUG(), but trigger recovery instead. */
3338 ret = adjust_va_to_fit_type(va, start, size, type);
3342 /* Allocated area. */
3344 va->va_start = start;
3345 va->va_end = start + size;
3347 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
3350 spin_unlock(&vmap_area_lock);
3352 /* insert all vm's */
3353 for (area = 0; area < nr_vms; area++)
3354 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
3361 /* Remove previously inserted areas. */
3363 __free_vmap_area(vas[area]);
3368 spin_unlock(&vmap_area_lock);
3370 purge_vmap_area_lazy();
3373 /* Before "retry", check if we recover. */
3374 for (area = 0; area < nr_vms; area++) {
3378 vas[area] = kmem_cache_zalloc(
3379 vmap_area_cachep, GFP_KERNEL);
3388 for (area = 0; area < nr_vms; area++) {
3390 kmem_cache_free(vmap_area_cachep, vas[area]);
3401 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3402 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3403 * @nr_vms: the number of allocated areas
3405 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3407 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3411 for (i = 0; i < nr_vms; i++)
3412 free_vm_area(vms[i]);
3415 #endif /* CONFIG_SMP */
3417 #ifdef CONFIG_PROC_FS
3418 static void *s_start(struct seq_file *m, loff_t *pos)
3419 __acquires(&vmap_area_lock)
3421 spin_lock(&vmap_area_lock);
3422 return seq_list_start(&vmap_area_list, *pos);
3425 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3427 return seq_list_next(p, &vmap_area_list, pos);
3430 static void s_stop(struct seq_file *m, void *p)
3431 __releases(&vmap_area_lock)
3433 spin_unlock(&vmap_area_lock);
3436 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3438 if (IS_ENABLED(CONFIG_NUMA)) {
3439 unsigned int nr, *counters = m->private;
3444 if (v->flags & VM_UNINITIALIZED)
3446 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3449 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3451 for (nr = 0; nr < v->nr_pages; nr++)
3452 counters[page_to_nid(v->pages[nr])]++;
3454 for_each_node_state(nr, N_HIGH_MEMORY)
3456 seq_printf(m, " N%u=%u", nr, counters[nr]);
3460 static void show_purge_info(struct seq_file *m)
3462 struct llist_node *head;
3463 struct vmap_area *va;
3465 head = READ_ONCE(vmap_purge_list.first);
3469 llist_for_each_entry(va, head, purge_list) {
3470 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3471 (void *)va->va_start, (void *)va->va_end,
3472 va->va_end - va->va_start);
3476 static int s_show(struct seq_file *m, void *p)
3478 struct vmap_area *va;
3479 struct vm_struct *v;
3481 va = list_entry(p, struct vmap_area, list);
3484 * s_show can encounter race with remove_vm_area, !vm on behalf
3485 * of vmap area is being tear down or vm_map_ram allocation.
3488 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3489 (void *)va->va_start, (void *)va->va_end,
3490 va->va_end - va->va_start);
3497 seq_printf(m, "0x%pK-0x%pK %7ld",
3498 v->addr, v->addr + v->size, v->size);
3501 seq_printf(m, " %pS", v->caller);
3504 seq_printf(m, " pages=%d", v->nr_pages);
3507 seq_printf(m, " phys=%pa", &v->phys_addr);
3509 if (v->flags & VM_IOREMAP)
3510 seq_puts(m, " ioremap");
3512 if (v->flags & VM_ALLOC)
3513 seq_puts(m, " vmalloc");
3515 if (v->flags & VM_MAP)
3516 seq_puts(m, " vmap");
3518 if (v->flags & VM_USERMAP)
3519 seq_puts(m, " user");
3521 if (v->flags & VM_DMA_COHERENT)
3522 seq_puts(m, " dma-coherent");
3524 if (is_vmalloc_addr(v->pages))
3525 seq_puts(m, " vpages");
3527 show_numa_info(m, v);
3531 * As a final step, dump "unpurged" areas. Note,
3532 * that entire "/proc/vmallocinfo" output will not
3533 * be address sorted, because the purge list is not
3536 if (list_is_last(&va->list, &vmap_area_list))
3542 static const struct seq_operations vmalloc_op = {
3549 static int __init proc_vmalloc_init(void)
3551 if (IS_ENABLED(CONFIG_NUMA))
3552 proc_create_seq_private("vmallocinfo", 0400, NULL,
3554 nr_node_ids * sizeof(unsigned int), NULL);
3556 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3559 module_init(proc_vmalloc_init);