1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
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/xarray.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>
36 #include <linux/overflow.h>
37 #include <linux/pgtable.h>
38 #include <linux/uaccess.h>
39 #include <asm/tlbflush.h>
40 #include <asm/shmparam.h>
43 #include "pgalloc-track.h"
45 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
46 static bool __ro_after_init vmap_allow_huge = true;
48 static int __init set_nohugevmalloc(char *str)
50 vmap_allow_huge = false;
53 early_param("nohugevmalloc", set_nohugevmalloc);
54 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
55 static const bool vmap_allow_huge = false;
56 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
58 bool is_vmalloc_addr(const void *x)
60 unsigned long addr = (unsigned long)x;
62 return addr >= VMALLOC_START && addr < VMALLOC_END;
64 EXPORT_SYMBOL(is_vmalloc_addr);
66 struct vfree_deferred {
67 struct llist_head list;
68 struct work_struct wq;
70 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
72 static void __vunmap(const void *, int);
74 static void free_work(struct work_struct *w)
76 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
77 struct llist_node *t, *llnode;
79 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
80 __vunmap((void *)llnode, 1);
83 /*** Page table manipulation functions ***/
84 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
85 phys_addr_t phys_addr, pgprot_t prot,
91 pfn = phys_addr >> PAGE_SHIFT;
92 pte = pte_alloc_kernel_track(pmd, addr, mask);
96 BUG_ON(!pte_none(*pte));
97 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
99 } while (pte++, addr += PAGE_SIZE, addr != end);
100 *mask |= PGTBL_PTE_MODIFIED;
104 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
105 phys_addr_t phys_addr, pgprot_t prot,
106 unsigned int max_page_shift)
108 if (max_page_shift < PMD_SHIFT)
111 if (!arch_vmap_pmd_supported(prot))
114 if ((end - addr) != PMD_SIZE)
117 if (!IS_ALIGNED(addr, PMD_SIZE))
120 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
123 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
126 return pmd_set_huge(pmd, phys_addr, prot);
129 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
130 phys_addr_t phys_addr, pgprot_t prot,
131 unsigned int max_page_shift, pgtbl_mod_mask *mask)
136 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
140 next = pmd_addr_end(addr, end);
142 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
144 *mask |= PGTBL_PMD_MODIFIED;
148 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, mask))
150 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
154 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
155 phys_addr_t phys_addr, pgprot_t prot,
156 unsigned int max_page_shift)
158 if (max_page_shift < PUD_SHIFT)
161 if (!arch_vmap_pud_supported(prot))
164 if ((end - addr) != PUD_SIZE)
167 if (!IS_ALIGNED(addr, PUD_SIZE))
170 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
173 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
176 return pud_set_huge(pud, phys_addr, prot);
179 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
180 phys_addr_t phys_addr, pgprot_t prot,
181 unsigned int max_page_shift, pgtbl_mod_mask *mask)
186 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
190 next = pud_addr_end(addr, end);
192 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
194 *mask |= PGTBL_PUD_MODIFIED;
198 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
199 max_page_shift, mask))
201 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
205 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
206 phys_addr_t phys_addr, pgprot_t prot,
207 unsigned int max_page_shift)
209 if (max_page_shift < P4D_SHIFT)
212 if (!arch_vmap_p4d_supported(prot))
215 if ((end - addr) != P4D_SIZE)
218 if (!IS_ALIGNED(addr, P4D_SIZE))
221 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
224 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
227 return p4d_set_huge(p4d, phys_addr, prot);
230 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
231 phys_addr_t phys_addr, pgprot_t prot,
232 unsigned int max_page_shift, pgtbl_mod_mask *mask)
237 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
241 next = p4d_addr_end(addr, end);
243 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
245 *mask |= PGTBL_P4D_MODIFIED;
249 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
250 max_page_shift, mask))
252 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
256 static int vmap_range_noflush(unsigned long addr, unsigned long end,
257 phys_addr_t phys_addr, pgprot_t prot,
258 unsigned int max_page_shift)
264 pgtbl_mod_mask mask = 0;
270 pgd = pgd_offset_k(addr);
272 next = pgd_addr_end(addr, end);
273 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
274 max_page_shift, &mask);
277 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
279 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
280 arch_sync_kernel_mappings(start, end);
285 int vmap_range(unsigned long addr, unsigned long end,
286 phys_addr_t phys_addr, pgprot_t prot,
287 unsigned int max_page_shift)
291 err = vmap_range_noflush(addr, end, phys_addr, prot, max_page_shift);
292 flush_cache_vmap(addr, end);
297 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
298 pgtbl_mod_mask *mask)
302 pte = pte_offset_kernel(pmd, addr);
304 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
305 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
306 } while (pte++, addr += PAGE_SIZE, addr != end);
307 *mask |= PGTBL_PTE_MODIFIED;
310 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
311 pgtbl_mod_mask *mask)
317 pmd = pmd_offset(pud, addr);
319 next = pmd_addr_end(addr, end);
321 cleared = pmd_clear_huge(pmd);
322 if (cleared || pmd_bad(*pmd))
323 *mask |= PGTBL_PMD_MODIFIED;
327 if (pmd_none_or_clear_bad(pmd))
329 vunmap_pte_range(pmd, addr, next, mask);
332 } while (pmd++, addr = next, addr != end);
335 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
336 pgtbl_mod_mask *mask)
342 pud = pud_offset(p4d, addr);
344 next = pud_addr_end(addr, end);
346 cleared = pud_clear_huge(pud);
347 if (cleared || pud_bad(*pud))
348 *mask |= PGTBL_PUD_MODIFIED;
352 if (pud_none_or_clear_bad(pud))
354 vunmap_pmd_range(pud, addr, next, mask);
355 } while (pud++, addr = next, addr != end);
358 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
359 pgtbl_mod_mask *mask)
365 p4d = p4d_offset(pgd, addr);
367 next = p4d_addr_end(addr, end);
369 cleared = p4d_clear_huge(p4d);
370 if (cleared || p4d_bad(*p4d))
371 *mask |= PGTBL_P4D_MODIFIED;
375 if (p4d_none_or_clear_bad(p4d))
377 vunmap_pud_range(p4d, addr, next, mask);
378 } while (p4d++, addr = next, addr != end);
382 * vunmap_range_noflush is similar to vunmap_range, but does not
383 * flush caches or TLBs.
385 * The caller is responsible for calling flush_cache_vmap() before calling
386 * this function, and flush_tlb_kernel_range after it has returned
387 * successfully (and before the addresses are expected to cause a page fault
388 * or be re-mapped for something else, if TLB flushes are being delayed or
391 * This is an internal function only. Do not use outside mm/.
393 void vunmap_range_noflush(unsigned long start, unsigned long end)
397 unsigned long addr = start;
398 pgtbl_mod_mask mask = 0;
401 pgd = pgd_offset_k(addr);
403 next = pgd_addr_end(addr, end);
405 mask |= PGTBL_PGD_MODIFIED;
406 if (pgd_none_or_clear_bad(pgd))
408 vunmap_p4d_range(pgd, addr, next, &mask);
409 } while (pgd++, addr = next, addr != end);
411 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
412 arch_sync_kernel_mappings(start, end);
416 * vunmap_range - unmap kernel virtual addresses
417 * @addr: start of the VM area to unmap
418 * @end: end of the VM area to unmap (non-inclusive)
420 * Clears any present PTEs in the virtual address range, flushes TLBs and
421 * caches. Any subsequent access to the address before it has been re-mapped
424 void vunmap_range(unsigned long addr, unsigned long end)
426 flush_cache_vunmap(addr, end);
427 vunmap_range_noflush(addr, end);
428 flush_tlb_kernel_range(addr, end);
431 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
432 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
433 pgtbl_mod_mask *mask)
438 * nr is a running index into the array which helps higher level
439 * callers keep track of where we're up to.
442 pte = pte_alloc_kernel_track(pmd, addr, mask);
446 struct page *page = pages[*nr];
448 if (WARN_ON(!pte_none(*pte)))
452 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
454 } while (pte++, addr += PAGE_SIZE, addr != end);
455 *mask |= PGTBL_PTE_MODIFIED;
459 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
460 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
461 pgtbl_mod_mask *mask)
466 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
470 next = pmd_addr_end(addr, end);
471 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
473 } while (pmd++, addr = next, addr != end);
477 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
478 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
479 pgtbl_mod_mask *mask)
484 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
488 next = pud_addr_end(addr, end);
489 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
491 } while (pud++, addr = next, addr != end);
495 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
496 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
497 pgtbl_mod_mask *mask)
502 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
506 next = p4d_addr_end(addr, end);
507 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
509 } while (p4d++, addr = next, addr != end);
513 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
514 pgprot_t prot, struct page **pages)
516 unsigned long start = addr;
521 pgtbl_mod_mask mask = 0;
524 pgd = pgd_offset_k(addr);
526 next = pgd_addr_end(addr, end);
528 mask |= PGTBL_PGD_MODIFIED;
529 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
532 } while (pgd++, addr = next, addr != end);
534 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
535 arch_sync_kernel_mappings(start, end);
541 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
544 * The caller is responsible for calling flush_cache_vmap() after this
545 * function returns successfully and before the addresses are accessed.
547 * This is an internal function only. Do not use outside mm/.
549 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
550 pgprot_t prot, struct page **pages, unsigned int page_shift)
552 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
554 WARN_ON(page_shift < PAGE_SHIFT);
556 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
557 page_shift == PAGE_SHIFT)
558 return vmap_small_pages_range_noflush(addr, end, prot, pages);
560 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
563 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
564 __pa(page_address(pages[i])), prot,
569 addr += 1UL << page_shift;
576 * vmap_pages_range - map pages to a kernel virtual address
577 * @addr: start of the VM area to map
578 * @end: end of the VM area to map (non-inclusive)
579 * @prot: page protection flags to use
580 * @pages: pages to map (always PAGE_SIZE pages)
581 * @page_shift: maximum shift that the pages may be mapped with, @pages must
582 * be aligned and contiguous up to at least this shift.
585 * 0 on success, -errno on failure.
587 static int vmap_pages_range(unsigned long addr, unsigned long end,
588 pgprot_t prot, struct page **pages, unsigned int page_shift)
592 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
593 flush_cache_vmap(addr, end);
597 int is_vmalloc_or_module_addr(const void *x)
600 * ARM, x86-64 and sparc64 put modules in a special place,
601 * and fall back on vmalloc() if that fails. Others
602 * just put it in the vmalloc space.
604 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
605 unsigned long addr = (unsigned long)x;
606 if (addr >= MODULES_VADDR && addr < MODULES_END)
609 return is_vmalloc_addr(x);
613 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
614 * return the tail page that corresponds to the base page address, which
615 * matches small vmap mappings.
617 struct page *vmalloc_to_page(const void *vmalloc_addr)
619 unsigned long addr = (unsigned long) vmalloc_addr;
620 struct page *page = NULL;
621 pgd_t *pgd = pgd_offset_k(addr);
628 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
629 * architectures that do not vmalloc module space
631 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
635 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
636 return NULL; /* XXX: no allowance for huge pgd */
637 if (WARN_ON_ONCE(pgd_bad(*pgd)))
640 p4d = p4d_offset(pgd, addr);
644 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
645 if (WARN_ON_ONCE(p4d_bad(*p4d)))
648 pud = pud_offset(p4d, addr);
652 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
653 if (WARN_ON_ONCE(pud_bad(*pud)))
656 pmd = pmd_offset(pud, addr);
660 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
661 if (WARN_ON_ONCE(pmd_bad(*pmd)))
664 ptep = pte_offset_map(pmd, addr);
666 if (pte_present(pte))
667 page = pte_page(pte);
672 EXPORT_SYMBOL(vmalloc_to_page);
675 * Map a vmalloc()-space virtual address to the physical page frame number.
677 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
679 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
681 EXPORT_SYMBOL(vmalloc_to_pfn);
684 /*** Global kva allocator ***/
686 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
687 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
690 static DEFINE_SPINLOCK(vmap_area_lock);
691 static DEFINE_SPINLOCK(free_vmap_area_lock);
692 /* Export for kexec only */
693 LIST_HEAD(vmap_area_list);
694 static struct rb_root vmap_area_root = RB_ROOT;
695 static bool vmap_initialized __read_mostly;
697 static struct rb_root purge_vmap_area_root = RB_ROOT;
698 static LIST_HEAD(purge_vmap_area_list);
699 static DEFINE_SPINLOCK(purge_vmap_area_lock);
702 * This kmem_cache is used for vmap_area objects. Instead of
703 * allocating from slab we reuse an object from this cache to
704 * make things faster. Especially in "no edge" splitting of
707 static struct kmem_cache *vmap_area_cachep;
710 * This linked list is used in pair with free_vmap_area_root.
711 * It gives O(1) access to prev/next to perform fast coalescing.
713 static LIST_HEAD(free_vmap_area_list);
716 * This augment red-black tree represents the free vmap space.
717 * All vmap_area objects in this tree are sorted by va->va_start
718 * address. It is used for allocation and merging when a vmap
719 * object is released.
721 * Each vmap_area node contains a maximum available free block
722 * of its sub-tree, right or left. Therefore it is possible to
723 * find a lowest match of free area.
725 static struct rb_root free_vmap_area_root = RB_ROOT;
728 * Preload a CPU with one object for "no edge" split case. The
729 * aim is to get rid of allocations from the atomic context, thus
730 * to use more permissive allocation masks.
732 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
734 static __always_inline unsigned long
735 va_size(struct vmap_area *va)
737 return (va->va_end - va->va_start);
740 static __always_inline unsigned long
741 get_subtree_max_size(struct rb_node *node)
743 struct vmap_area *va;
745 va = rb_entry_safe(node, struct vmap_area, rb_node);
746 return va ? va->subtree_max_size : 0;
750 * Gets called when remove the node and rotate.
752 static __always_inline unsigned long
753 compute_subtree_max_size(struct vmap_area *va)
755 return max3(va_size(va),
756 get_subtree_max_size(va->rb_node.rb_left),
757 get_subtree_max_size(va->rb_node.rb_right));
760 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
761 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
763 static void purge_vmap_area_lazy(void);
764 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
765 static unsigned long lazy_max_pages(void);
767 static atomic_long_t nr_vmalloc_pages;
769 unsigned long vmalloc_nr_pages(void)
771 return atomic_long_read(&nr_vmalloc_pages);
774 static struct vmap_area *__find_vmap_area(unsigned long addr)
776 struct rb_node *n = vmap_area_root.rb_node;
779 struct vmap_area *va;
781 va = rb_entry(n, struct vmap_area, rb_node);
782 if (addr < va->va_start)
784 else if (addr >= va->va_end)
794 * This function returns back addresses of parent node
795 * and its left or right link for further processing.
797 * Otherwise NULL is returned. In that case all further
798 * steps regarding inserting of conflicting overlap range
799 * have to be declined and actually considered as a bug.
801 static __always_inline struct rb_node **
802 find_va_links(struct vmap_area *va,
803 struct rb_root *root, struct rb_node *from,
804 struct rb_node **parent)
806 struct vmap_area *tmp_va;
807 struct rb_node **link;
810 link = &root->rb_node;
811 if (unlikely(!*link)) {
820 * Go to the bottom of the tree. When we hit the last point
821 * we end up with parent rb_node and correct direction, i name
822 * it link, where the new va->rb_node will be attached to.
825 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
828 * During the traversal we also do some sanity check.
829 * Trigger the BUG() if there are sides(left/right)
832 if (va->va_start < tmp_va->va_end &&
833 va->va_end <= tmp_va->va_start)
834 link = &(*link)->rb_left;
835 else if (va->va_end > tmp_va->va_start &&
836 va->va_start >= tmp_va->va_end)
837 link = &(*link)->rb_right;
839 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
840 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
846 *parent = &tmp_va->rb_node;
850 static __always_inline struct list_head *
851 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
853 struct list_head *list;
855 if (unlikely(!parent))
857 * The red-black tree where we try to find VA neighbors
858 * before merging or inserting is empty, i.e. it means
859 * there is no free vmap space. Normally it does not
860 * happen but we handle this case anyway.
864 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
865 return (&parent->rb_right == link ? list->next : list);
868 static __always_inline void
869 link_va(struct vmap_area *va, struct rb_root *root,
870 struct rb_node *parent, struct rb_node **link, struct list_head *head)
873 * VA is still not in the list, but we can
874 * identify its future previous list_head node.
876 if (likely(parent)) {
877 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
878 if (&parent->rb_right != link)
882 /* Insert to the rb-tree */
883 rb_link_node(&va->rb_node, parent, link);
884 if (root == &free_vmap_area_root) {
886 * Some explanation here. Just perform simple insertion
887 * to the tree. We do not set va->subtree_max_size to
888 * its current size before calling rb_insert_augmented().
889 * It is because of we populate the tree from the bottom
890 * to parent levels when the node _is_ in the tree.
892 * Therefore we set subtree_max_size to zero after insertion,
893 * to let __augment_tree_propagate_from() puts everything to
894 * the correct order later on.
896 rb_insert_augmented(&va->rb_node,
897 root, &free_vmap_area_rb_augment_cb);
898 va->subtree_max_size = 0;
900 rb_insert_color(&va->rb_node, root);
903 /* Address-sort this list */
904 list_add(&va->list, head);
907 static __always_inline void
908 unlink_va(struct vmap_area *va, struct rb_root *root)
910 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
913 if (root == &free_vmap_area_root)
914 rb_erase_augmented(&va->rb_node,
915 root, &free_vmap_area_rb_augment_cb);
917 rb_erase(&va->rb_node, root);
920 RB_CLEAR_NODE(&va->rb_node);
923 #if DEBUG_AUGMENT_PROPAGATE_CHECK
925 augment_tree_propagate_check(void)
927 struct vmap_area *va;
928 unsigned long computed_size;
930 list_for_each_entry(va, &free_vmap_area_list, list) {
931 computed_size = compute_subtree_max_size(va);
932 if (computed_size != va->subtree_max_size)
933 pr_emerg("tree is corrupted: %lu, %lu\n",
934 va_size(va), va->subtree_max_size);
940 * This function populates subtree_max_size from bottom to upper
941 * levels starting from VA point. The propagation must be done
942 * when VA size is modified by changing its va_start/va_end. Or
943 * in case of newly inserting of VA to the tree.
945 * It means that __augment_tree_propagate_from() must be called:
946 * - After VA has been inserted to the tree(free path);
947 * - After VA has been shrunk(allocation path);
948 * - After VA has been increased(merging path).
950 * Please note that, it does not mean that upper parent nodes
951 * and their subtree_max_size are recalculated all the time up
960 * For example if we modify the node 4, shrinking it to 2, then
961 * no any modification is required. If we shrink the node 2 to 1
962 * its subtree_max_size is updated only, and set to 1. If we shrink
963 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
966 static __always_inline void
967 augment_tree_propagate_from(struct vmap_area *va)
970 * Populate the tree from bottom towards the root until
971 * the calculated maximum available size of checked node
972 * is equal to its current one.
974 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
976 #if DEBUG_AUGMENT_PROPAGATE_CHECK
977 augment_tree_propagate_check();
982 insert_vmap_area(struct vmap_area *va,
983 struct rb_root *root, struct list_head *head)
985 struct rb_node **link;
986 struct rb_node *parent;
988 link = find_va_links(va, root, NULL, &parent);
990 link_va(va, root, parent, link, head);
994 insert_vmap_area_augment(struct vmap_area *va,
995 struct rb_node *from, struct rb_root *root,
996 struct list_head *head)
998 struct rb_node **link;
999 struct rb_node *parent;
1002 link = find_va_links(va, NULL, from, &parent);
1004 link = find_va_links(va, root, NULL, &parent);
1007 link_va(va, root, parent, link, head);
1008 augment_tree_propagate_from(va);
1013 * Merge de-allocated chunk of VA memory with previous
1014 * and next free blocks. If coalesce is not done a new
1015 * free area is inserted. If VA has been merged, it is
1018 * Please note, it can return NULL in case of overlap
1019 * ranges, followed by WARN() report. Despite it is a
1020 * buggy behaviour, a system can be alive and keep
1023 static __always_inline struct vmap_area *
1024 merge_or_add_vmap_area(struct vmap_area *va,
1025 struct rb_root *root, struct list_head *head)
1027 struct vmap_area *sibling;
1028 struct list_head *next;
1029 struct rb_node **link;
1030 struct rb_node *parent;
1031 bool merged = false;
1034 * Find a place in the tree where VA potentially will be
1035 * inserted, unless it is merged with its sibling/siblings.
1037 link = find_va_links(va, root, NULL, &parent);
1042 * Get next node of VA to check if merging can be done.
1044 next = get_va_next_sibling(parent, link);
1045 if (unlikely(next == NULL))
1051 * |<------VA------>|<-----Next----->|
1056 sibling = list_entry(next, struct vmap_area, list);
1057 if (sibling->va_start == va->va_end) {
1058 sibling->va_start = va->va_start;
1060 /* Free vmap_area object. */
1061 kmem_cache_free(vmap_area_cachep, va);
1063 /* Point to the new merged area. */
1072 * |<-----Prev----->|<------VA------>|
1076 if (next->prev != head) {
1077 sibling = list_entry(next->prev, struct vmap_area, list);
1078 if (sibling->va_end == va->va_start) {
1080 * If both neighbors are coalesced, it is important
1081 * to unlink the "next" node first, followed by merging
1082 * with "previous" one. Otherwise the tree might not be
1083 * fully populated if a sibling's augmented value is
1084 * "normalized" because of rotation operations.
1087 unlink_va(va, root);
1089 sibling->va_end = va->va_end;
1091 /* Free vmap_area object. */
1092 kmem_cache_free(vmap_area_cachep, va);
1094 /* Point to the new merged area. */
1102 link_va(va, root, parent, link, head);
1107 static __always_inline struct vmap_area *
1108 merge_or_add_vmap_area_augment(struct vmap_area *va,
1109 struct rb_root *root, struct list_head *head)
1111 va = merge_or_add_vmap_area(va, root, head);
1113 augment_tree_propagate_from(va);
1118 static __always_inline bool
1119 is_within_this_va(struct vmap_area *va, unsigned long size,
1120 unsigned long align, unsigned long vstart)
1122 unsigned long nva_start_addr;
1124 if (va->va_start > vstart)
1125 nva_start_addr = ALIGN(va->va_start, align);
1127 nva_start_addr = ALIGN(vstart, align);
1129 /* Can be overflowed due to big size or alignment. */
1130 if (nva_start_addr + size < nva_start_addr ||
1131 nva_start_addr < vstart)
1134 return (nva_start_addr + size <= va->va_end);
1138 * Find the first free block(lowest start address) in the tree,
1139 * that will accomplish the request corresponding to passing
1142 static __always_inline struct vmap_area *
1143 find_vmap_lowest_match(unsigned long size,
1144 unsigned long align, unsigned long vstart)
1146 struct vmap_area *va;
1147 struct rb_node *node;
1148 unsigned long length;
1150 /* Start from the root. */
1151 node = free_vmap_area_root.rb_node;
1153 /* Adjust the search size for alignment overhead. */
1154 length = size + align - 1;
1157 va = rb_entry(node, struct vmap_area, rb_node);
1159 if (get_subtree_max_size(node->rb_left) >= length &&
1160 vstart < va->va_start) {
1161 node = node->rb_left;
1163 if (is_within_this_va(va, size, align, vstart))
1167 * Does not make sense to go deeper towards the right
1168 * sub-tree if it does not have a free block that is
1169 * equal or bigger to the requested search length.
1171 if (get_subtree_max_size(node->rb_right) >= length) {
1172 node = node->rb_right;
1177 * OK. We roll back and find the first right sub-tree,
1178 * that will satisfy the search criteria. It can happen
1179 * only once due to "vstart" restriction.
1181 while ((node = rb_parent(node))) {
1182 va = rb_entry(node, struct vmap_area, rb_node);
1183 if (is_within_this_va(va, size, align, vstart))
1186 if (get_subtree_max_size(node->rb_right) >= length &&
1187 vstart <= va->va_start) {
1188 node = node->rb_right;
1198 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1199 #include <linux/random.h>
1201 static struct vmap_area *
1202 find_vmap_lowest_linear_match(unsigned long size,
1203 unsigned long align, unsigned long vstart)
1205 struct vmap_area *va;
1207 list_for_each_entry(va, &free_vmap_area_list, list) {
1208 if (!is_within_this_va(va, size, align, vstart))
1218 find_vmap_lowest_match_check(unsigned long size)
1220 struct vmap_area *va_1, *va_2;
1221 unsigned long vstart;
1224 get_random_bytes(&rnd, sizeof(rnd));
1225 vstart = VMALLOC_START + rnd;
1227 va_1 = find_vmap_lowest_match(size, 1, vstart);
1228 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
1231 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1232 va_1, va_2, vstart);
1238 FL_FIT_TYPE = 1, /* full fit */
1239 LE_FIT_TYPE = 2, /* left edge fit */
1240 RE_FIT_TYPE = 3, /* right edge fit */
1241 NE_FIT_TYPE = 4 /* no edge fit */
1244 static __always_inline enum fit_type
1245 classify_va_fit_type(struct vmap_area *va,
1246 unsigned long nva_start_addr, unsigned long size)
1250 /* Check if it is within VA. */
1251 if (nva_start_addr < va->va_start ||
1252 nva_start_addr + size > va->va_end)
1256 if (va->va_start == nva_start_addr) {
1257 if (va->va_end == nva_start_addr + size)
1261 } else if (va->va_end == nva_start_addr + size) {
1270 static __always_inline int
1271 adjust_va_to_fit_type(struct vmap_area *va,
1272 unsigned long nva_start_addr, unsigned long size,
1275 struct vmap_area *lva = NULL;
1277 if (type == FL_FIT_TYPE) {
1279 * No need to split VA, it fully fits.
1285 unlink_va(va, &free_vmap_area_root);
1286 kmem_cache_free(vmap_area_cachep, va);
1287 } else if (type == LE_FIT_TYPE) {
1289 * Split left edge of fit VA.
1295 va->va_start += size;
1296 } else if (type == RE_FIT_TYPE) {
1298 * Split right edge of fit VA.
1304 va->va_end = nva_start_addr;
1305 } else if (type == NE_FIT_TYPE) {
1307 * Split no edge of fit VA.
1313 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1314 if (unlikely(!lva)) {
1316 * For percpu allocator we do not do any pre-allocation
1317 * and leave it as it is. The reason is it most likely
1318 * never ends up with NE_FIT_TYPE splitting. In case of
1319 * percpu allocations offsets and sizes are aligned to
1320 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1321 * are its main fitting cases.
1323 * There are a few exceptions though, as an example it is
1324 * a first allocation (early boot up) when we have "one"
1325 * big free space that has to be split.
1327 * Also we can hit this path in case of regular "vmap"
1328 * allocations, if "this" current CPU was not preloaded.
1329 * See the comment in alloc_vmap_area() why. If so, then
1330 * GFP_NOWAIT is used instead to get an extra object for
1331 * split purpose. That is rare and most time does not
1334 * What happens if an allocation gets failed. Basically,
1335 * an "overflow" path is triggered to purge lazily freed
1336 * areas to free some memory, then, the "retry" path is
1337 * triggered to repeat one more time. See more details
1338 * in alloc_vmap_area() function.
1340 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1346 * Build the remainder.
1348 lva->va_start = va->va_start;
1349 lva->va_end = nva_start_addr;
1352 * Shrink this VA to remaining size.
1354 va->va_start = nva_start_addr + size;
1359 if (type != FL_FIT_TYPE) {
1360 augment_tree_propagate_from(va);
1362 if (lva) /* type == NE_FIT_TYPE */
1363 insert_vmap_area_augment(lva, &va->rb_node,
1364 &free_vmap_area_root, &free_vmap_area_list);
1371 * Returns a start address of the newly allocated area, if success.
1372 * Otherwise a vend is returned that indicates failure.
1374 static __always_inline unsigned long
1375 __alloc_vmap_area(unsigned long size, unsigned long align,
1376 unsigned long vstart, unsigned long vend)
1378 unsigned long nva_start_addr;
1379 struct vmap_area *va;
1383 va = find_vmap_lowest_match(size, align, vstart);
1387 if (va->va_start > vstart)
1388 nva_start_addr = ALIGN(va->va_start, align);
1390 nva_start_addr = ALIGN(vstart, align);
1392 /* Check the "vend" restriction. */
1393 if (nva_start_addr + size > vend)
1396 /* Classify what we have found. */
1397 type = classify_va_fit_type(va, nva_start_addr, size);
1398 if (WARN_ON_ONCE(type == NOTHING_FIT))
1401 /* Update the free vmap_area. */
1402 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1406 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1407 find_vmap_lowest_match_check(size);
1410 return nva_start_addr;
1414 * Free a region of KVA allocated by alloc_vmap_area
1416 static void free_vmap_area(struct vmap_area *va)
1419 * Remove from the busy tree/list.
1421 spin_lock(&vmap_area_lock);
1422 unlink_va(va, &vmap_area_root);
1423 spin_unlock(&vmap_area_lock);
1426 * Insert/Merge it back to the free tree/list.
1428 spin_lock(&free_vmap_area_lock);
1429 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1430 spin_unlock(&free_vmap_area_lock);
1434 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1436 struct vmap_area *va = NULL;
1439 * Preload this CPU with one extra vmap_area object. It is used
1440 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1441 * a CPU that does an allocation is preloaded.
1443 * We do it in non-atomic context, thus it allows us to use more
1444 * permissive allocation masks to be more stable under low memory
1445 * condition and high memory pressure.
1447 if (!this_cpu_read(ne_fit_preload_node))
1448 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1452 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1453 kmem_cache_free(vmap_area_cachep, va);
1457 * Allocate a region of KVA of the specified size and alignment, within the
1460 static struct vmap_area *alloc_vmap_area(unsigned long size,
1461 unsigned long align,
1462 unsigned long vstart, unsigned long vend,
1463 int node, gfp_t gfp_mask)
1465 struct vmap_area *va;
1471 BUG_ON(offset_in_page(size));
1472 BUG_ON(!is_power_of_2(align));
1474 if (unlikely(!vmap_initialized))
1475 return ERR_PTR(-EBUSY);
1478 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1480 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1482 return ERR_PTR(-ENOMEM);
1485 * Only scan the relevant parts containing pointers to other objects
1486 * to avoid false negatives.
1488 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1491 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1492 addr = __alloc_vmap_area(size, align, vstart, vend);
1493 spin_unlock(&free_vmap_area_lock);
1496 * If an allocation fails, the "vend" address is
1497 * returned. Therefore trigger the overflow path.
1499 if (unlikely(addr == vend))
1502 va->va_start = addr;
1503 va->va_end = addr + size;
1506 spin_lock(&vmap_area_lock);
1507 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1508 spin_unlock(&vmap_area_lock);
1510 BUG_ON(!IS_ALIGNED(va->va_start, align));
1511 BUG_ON(va->va_start < vstart);
1512 BUG_ON(va->va_end > vend);
1514 ret = kasan_populate_vmalloc(addr, size);
1517 return ERR_PTR(ret);
1524 purge_vmap_area_lazy();
1529 if (gfpflags_allow_blocking(gfp_mask)) {
1530 unsigned long freed = 0;
1531 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1538 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1539 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1542 kmem_cache_free(vmap_area_cachep, va);
1543 return ERR_PTR(-EBUSY);
1546 int register_vmap_purge_notifier(struct notifier_block *nb)
1548 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1550 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1552 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1554 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1556 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1559 * lazy_max_pages is the maximum amount of virtual address space we gather up
1560 * before attempting to purge with a TLB flush.
1562 * There is a tradeoff here: a larger number will cover more kernel page tables
1563 * and take slightly longer to purge, but it will linearly reduce the number of
1564 * global TLB flushes that must be performed. It would seem natural to scale
1565 * this number up linearly with the number of CPUs (because vmapping activity
1566 * could also scale linearly with the number of CPUs), however it is likely
1567 * that in practice, workloads might be constrained in other ways that mean
1568 * vmap activity will not scale linearly with CPUs. Also, I want to be
1569 * conservative and not introduce a big latency on huge systems, so go with
1570 * a less aggressive log scale. It will still be an improvement over the old
1571 * code, and it will be simple to change the scale factor if we find that it
1572 * becomes a problem on bigger systems.
1574 static unsigned long lazy_max_pages(void)
1578 log = fls(num_online_cpus());
1580 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1583 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1586 * Serialize vmap purging. There is no actual critical section protected
1587 * by this look, but we want to avoid concurrent calls for performance
1588 * reasons and to make the pcpu_get_vm_areas more deterministic.
1590 static DEFINE_MUTEX(vmap_purge_lock);
1592 /* for per-CPU blocks */
1593 static void purge_fragmented_blocks_allcpus(void);
1596 * called before a call to iounmap() if the caller wants vm_area_struct's
1597 * immediately freed.
1599 void set_iounmap_nonlazy(void)
1601 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1605 * Purges all lazily-freed vmap areas.
1607 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1609 unsigned long resched_threshold;
1610 struct list_head local_pure_list;
1611 struct vmap_area *va, *n_va;
1613 lockdep_assert_held(&vmap_purge_lock);
1615 spin_lock(&purge_vmap_area_lock);
1616 purge_vmap_area_root = RB_ROOT;
1617 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1618 spin_unlock(&purge_vmap_area_lock);
1620 if (unlikely(list_empty(&local_pure_list)))
1624 list_first_entry(&local_pure_list,
1625 struct vmap_area, list)->va_start);
1628 list_last_entry(&local_pure_list,
1629 struct vmap_area, list)->va_end);
1631 flush_tlb_kernel_range(start, end);
1632 resched_threshold = lazy_max_pages() << 1;
1634 spin_lock(&free_vmap_area_lock);
1635 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1636 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1637 unsigned long orig_start = va->va_start;
1638 unsigned long orig_end = va->va_end;
1641 * Finally insert or merge lazily-freed area. It is
1642 * detached and there is no need to "unlink" it from
1645 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1646 &free_vmap_area_list);
1651 if (is_vmalloc_or_module_addr((void *)orig_start))
1652 kasan_release_vmalloc(orig_start, orig_end,
1653 va->va_start, va->va_end);
1655 atomic_long_sub(nr, &vmap_lazy_nr);
1657 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1658 cond_resched_lock(&free_vmap_area_lock);
1660 spin_unlock(&free_vmap_area_lock);
1665 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1666 * is already purging.
1668 static void try_purge_vmap_area_lazy(void)
1670 if (mutex_trylock(&vmap_purge_lock)) {
1671 __purge_vmap_area_lazy(ULONG_MAX, 0);
1672 mutex_unlock(&vmap_purge_lock);
1677 * Kick off a purge of the outstanding lazy areas.
1679 static void purge_vmap_area_lazy(void)
1681 mutex_lock(&vmap_purge_lock);
1682 purge_fragmented_blocks_allcpus();
1683 __purge_vmap_area_lazy(ULONG_MAX, 0);
1684 mutex_unlock(&vmap_purge_lock);
1688 * Free a vmap area, caller ensuring that the area has been unmapped
1689 * and flush_cache_vunmap had been called for the correct range
1692 static void free_vmap_area_noflush(struct vmap_area *va)
1694 unsigned long nr_lazy;
1696 spin_lock(&vmap_area_lock);
1697 unlink_va(va, &vmap_area_root);
1698 spin_unlock(&vmap_area_lock);
1700 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1701 PAGE_SHIFT, &vmap_lazy_nr);
1704 * Merge or place it to the purge tree/list.
1706 spin_lock(&purge_vmap_area_lock);
1707 merge_or_add_vmap_area(va,
1708 &purge_vmap_area_root, &purge_vmap_area_list);
1709 spin_unlock(&purge_vmap_area_lock);
1711 /* After this point, we may free va at any time */
1712 if (unlikely(nr_lazy > lazy_max_pages()))
1713 try_purge_vmap_area_lazy();
1717 * Free and unmap a vmap area
1719 static void free_unmap_vmap_area(struct vmap_area *va)
1721 flush_cache_vunmap(va->va_start, va->va_end);
1722 vunmap_range_noflush(va->va_start, va->va_end);
1723 if (debug_pagealloc_enabled_static())
1724 flush_tlb_kernel_range(va->va_start, va->va_end);
1726 free_vmap_area_noflush(va);
1729 static struct vmap_area *find_vmap_area(unsigned long addr)
1731 struct vmap_area *va;
1733 spin_lock(&vmap_area_lock);
1734 va = __find_vmap_area(addr);
1735 spin_unlock(&vmap_area_lock);
1740 /*** Per cpu kva allocator ***/
1743 * vmap space is limited especially on 32 bit architectures. Ensure there is
1744 * room for at least 16 percpu vmap blocks per CPU.
1747 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1748 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1749 * instead (we just need a rough idea)
1751 #if BITS_PER_LONG == 32
1752 #define VMALLOC_SPACE (128UL*1024*1024)
1754 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1757 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1758 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1759 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1760 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1761 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1762 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1763 #define VMAP_BBMAP_BITS \
1764 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1765 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1766 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1768 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1770 struct vmap_block_queue {
1772 struct list_head free;
1777 struct vmap_area *va;
1778 unsigned long free, dirty;
1779 unsigned long dirty_min, dirty_max; /*< dirty range */
1780 struct list_head free_list;
1781 struct rcu_head rcu_head;
1782 struct list_head purge;
1785 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1786 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1789 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1790 * in the free path. Could get rid of this if we change the API to return a
1791 * "cookie" from alloc, to be passed to free. But no big deal yet.
1793 static DEFINE_XARRAY(vmap_blocks);
1796 * We should probably have a fallback mechanism to allocate virtual memory
1797 * out of partially filled vmap blocks. However vmap block sizing should be
1798 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1802 static unsigned long addr_to_vb_idx(unsigned long addr)
1804 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1805 addr /= VMAP_BLOCK_SIZE;
1809 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1813 addr = va_start + (pages_off << PAGE_SHIFT);
1814 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1815 return (void *)addr;
1819 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1820 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1821 * @order: how many 2^order pages should be occupied in newly allocated block
1822 * @gfp_mask: flags for the page level allocator
1824 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1826 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1828 struct vmap_block_queue *vbq;
1829 struct vmap_block *vb;
1830 struct vmap_area *va;
1831 unsigned long vb_idx;
1835 node = numa_node_id();
1837 vb = kmalloc_node(sizeof(struct vmap_block),
1838 gfp_mask & GFP_RECLAIM_MASK, node);
1840 return ERR_PTR(-ENOMEM);
1842 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1843 VMALLOC_START, VMALLOC_END,
1847 return ERR_CAST(va);
1850 vaddr = vmap_block_vaddr(va->va_start, 0);
1851 spin_lock_init(&vb->lock);
1853 /* At least something should be left free */
1854 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1855 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1857 vb->dirty_min = VMAP_BBMAP_BITS;
1859 INIT_LIST_HEAD(&vb->free_list);
1861 vb_idx = addr_to_vb_idx(va->va_start);
1862 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1866 return ERR_PTR(err);
1869 vbq = &get_cpu_var(vmap_block_queue);
1870 spin_lock(&vbq->lock);
1871 list_add_tail_rcu(&vb->free_list, &vbq->free);
1872 spin_unlock(&vbq->lock);
1873 put_cpu_var(vmap_block_queue);
1878 static void free_vmap_block(struct vmap_block *vb)
1880 struct vmap_block *tmp;
1882 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1885 free_vmap_area_noflush(vb->va);
1886 kfree_rcu(vb, rcu_head);
1889 static void purge_fragmented_blocks(int cpu)
1892 struct vmap_block *vb;
1893 struct vmap_block *n_vb;
1894 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1897 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1899 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1902 spin_lock(&vb->lock);
1903 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1904 vb->free = 0; /* prevent further allocs after releasing lock */
1905 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1907 vb->dirty_max = VMAP_BBMAP_BITS;
1908 spin_lock(&vbq->lock);
1909 list_del_rcu(&vb->free_list);
1910 spin_unlock(&vbq->lock);
1911 spin_unlock(&vb->lock);
1912 list_add_tail(&vb->purge, &purge);
1914 spin_unlock(&vb->lock);
1918 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1919 list_del(&vb->purge);
1920 free_vmap_block(vb);
1924 static void purge_fragmented_blocks_allcpus(void)
1928 for_each_possible_cpu(cpu)
1929 purge_fragmented_blocks(cpu);
1932 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1934 struct vmap_block_queue *vbq;
1935 struct vmap_block *vb;
1939 BUG_ON(offset_in_page(size));
1940 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1941 if (WARN_ON(size == 0)) {
1943 * Allocating 0 bytes isn't what caller wants since
1944 * get_order(0) returns funny result. Just warn and terminate
1949 order = get_order(size);
1952 vbq = &get_cpu_var(vmap_block_queue);
1953 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1954 unsigned long pages_off;
1956 spin_lock(&vb->lock);
1957 if (vb->free < (1UL << order)) {
1958 spin_unlock(&vb->lock);
1962 pages_off = VMAP_BBMAP_BITS - vb->free;
1963 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1964 vb->free -= 1UL << order;
1965 if (vb->free == 0) {
1966 spin_lock(&vbq->lock);
1967 list_del_rcu(&vb->free_list);
1968 spin_unlock(&vbq->lock);
1971 spin_unlock(&vb->lock);
1975 put_cpu_var(vmap_block_queue);
1978 /* Allocate new block if nothing was found */
1980 vaddr = new_vmap_block(order, gfp_mask);
1985 static void vb_free(unsigned long addr, unsigned long size)
1987 unsigned long offset;
1989 struct vmap_block *vb;
1991 BUG_ON(offset_in_page(size));
1992 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1994 flush_cache_vunmap(addr, addr + size);
1996 order = get_order(size);
1997 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
1998 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2000 vunmap_range_noflush(addr, addr + size);
2002 if (debug_pagealloc_enabled_static())
2003 flush_tlb_kernel_range(addr, addr + size);
2005 spin_lock(&vb->lock);
2007 /* Expand dirty range */
2008 vb->dirty_min = min(vb->dirty_min, offset);
2009 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2011 vb->dirty += 1UL << order;
2012 if (vb->dirty == VMAP_BBMAP_BITS) {
2014 spin_unlock(&vb->lock);
2015 free_vmap_block(vb);
2017 spin_unlock(&vb->lock);
2020 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2024 if (unlikely(!vmap_initialized))
2029 for_each_possible_cpu(cpu) {
2030 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2031 struct vmap_block *vb;
2034 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2035 spin_lock(&vb->lock);
2036 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2037 unsigned long va_start = vb->va->va_start;
2040 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2041 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2043 start = min(s, start);
2048 spin_unlock(&vb->lock);
2053 mutex_lock(&vmap_purge_lock);
2054 purge_fragmented_blocks_allcpus();
2055 if (!__purge_vmap_area_lazy(start, end) && flush)
2056 flush_tlb_kernel_range(start, end);
2057 mutex_unlock(&vmap_purge_lock);
2061 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2063 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2064 * to amortize TLB flushing overheads. What this means is that any page you
2065 * have now, may, in a former life, have been mapped into kernel virtual
2066 * address by the vmap layer and so there might be some CPUs with TLB entries
2067 * still referencing that page (additional to the regular 1:1 kernel mapping).
2069 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2070 * be sure that none of the pages we have control over will have any aliases
2071 * from the vmap layer.
2073 void vm_unmap_aliases(void)
2075 unsigned long start = ULONG_MAX, end = 0;
2078 _vm_unmap_aliases(start, end, flush);
2080 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2083 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2084 * @mem: the pointer returned by vm_map_ram
2085 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2087 void vm_unmap_ram(const void *mem, unsigned int count)
2089 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2090 unsigned long addr = (unsigned long)mem;
2091 struct vmap_area *va;
2095 BUG_ON(addr < VMALLOC_START);
2096 BUG_ON(addr > VMALLOC_END);
2097 BUG_ON(!PAGE_ALIGNED(addr));
2099 kasan_poison_vmalloc(mem, size);
2101 if (likely(count <= VMAP_MAX_ALLOC)) {
2102 debug_check_no_locks_freed(mem, size);
2103 vb_free(addr, size);
2107 va = find_vmap_area(addr);
2109 debug_check_no_locks_freed((void *)va->va_start,
2110 (va->va_end - va->va_start));
2111 free_unmap_vmap_area(va);
2113 EXPORT_SYMBOL(vm_unmap_ram);
2116 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2117 * @pages: an array of pointers to the pages to be mapped
2118 * @count: number of pages
2119 * @node: prefer to allocate data structures on this node
2121 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2122 * faster than vmap so it's good. But if you mix long-life and short-life
2123 * objects with vm_map_ram(), it could consume lots of address space through
2124 * fragmentation (especially on a 32bit machine). You could see failures in
2125 * the end. Please use this function for short-lived objects.
2127 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2129 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2131 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2135 if (likely(count <= VMAP_MAX_ALLOC)) {
2136 mem = vb_alloc(size, GFP_KERNEL);
2139 addr = (unsigned long)mem;
2141 struct vmap_area *va;
2142 va = alloc_vmap_area(size, PAGE_SIZE,
2143 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2147 addr = va->va_start;
2151 kasan_unpoison_vmalloc(mem, size);
2153 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2154 pages, PAGE_SHIFT) < 0) {
2155 vm_unmap_ram(mem, count);
2161 EXPORT_SYMBOL(vm_map_ram);
2163 static struct vm_struct *vmlist __initdata;
2165 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2167 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2168 return vm->page_order;
2174 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2176 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2177 vm->page_order = order;
2184 * vm_area_add_early - add vmap area early during boot
2185 * @vm: vm_struct to add
2187 * This function is used to add fixed kernel vm area to vmlist before
2188 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2189 * should contain proper values and the other fields should be zero.
2191 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2193 void __init vm_area_add_early(struct vm_struct *vm)
2195 struct vm_struct *tmp, **p;
2197 BUG_ON(vmap_initialized);
2198 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2199 if (tmp->addr >= vm->addr) {
2200 BUG_ON(tmp->addr < vm->addr + vm->size);
2203 BUG_ON(tmp->addr + tmp->size > vm->addr);
2210 * vm_area_register_early - register vmap area early during boot
2211 * @vm: vm_struct to register
2212 * @align: requested alignment
2214 * This function is used to register kernel vm area before
2215 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2216 * proper values on entry and other fields should be zero. On return,
2217 * vm->addr contains the allocated address.
2219 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2221 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2223 static size_t vm_init_off __initdata;
2226 addr = ALIGN(VMALLOC_START + vm_init_off, align);
2227 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
2229 vm->addr = (void *)addr;
2231 vm_area_add_early(vm);
2234 static void vmap_init_free_space(void)
2236 unsigned long vmap_start = 1;
2237 const unsigned long vmap_end = ULONG_MAX;
2238 struct vmap_area *busy, *free;
2242 * -|-----|.....|-----|-----|-----|.....|-
2244 * |<--------------------------------->|
2246 list_for_each_entry(busy, &vmap_area_list, list) {
2247 if (busy->va_start - vmap_start > 0) {
2248 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2249 if (!WARN_ON_ONCE(!free)) {
2250 free->va_start = vmap_start;
2251 free->va_end = busy->va_start;
2253 insert_vmap_area_augment(free, NULL,
2254 &free_vmap_area_root,
2255 &free_vmap_area_list);
2259 vmap_start = busy->va_end;
2262 if (vmap_end - vmap_start > 0) {
2263 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2264 if (!WARN_ON_ONCE(!free)) {
2265 free->va_start = vmap_start;
2266 free->va_end = vmap_end;
2268 insert_vmap_area_augment(free, NULL,
2269 &free_vmap_area_root,
2270 &free_vmap_area_list);
2275 void __init vmalloc_init(void)
2277 struct vmap_area *va;
2278 struct vm_struct *tmp;
2282 * Create the cache for vmap_area objects.
2284 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2286 for_each_possible_cpu(i) {
2287 struct vmap_block_queue *vbq;
2288 struct vfree_deferred *p;
2290 vbq = &per_cpu(vmap_block_queue, i);
2291 spin_lock_init(&vbq->lock);
2292 INIT_LIST_HEAD(&vbq->free);
2293 p = &per_cpu(vfree_deferred, i);
2294 init_llist_head(&p->list);
2295 INIT_WORK(&p->wq, free_work);
2298 /* Import existing vmlist entries. */
2299 for (tmp = vmlist; tmp; tmp = tmp->next) {
2300 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2301 if (WARN_ON_ONCE(!va))
2304 va->va_start = (unsigned long)tmp->addr;
2305 va->va_end = va->va_start + tmp->size;
2307 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2311 * Now we can initialize a free vmap space.
2313 vmap_init_free_space();
2314 vmap_initialized = true;
2317 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2318 struct vmap_area *va, unsigned long flags, const void *caller)
2321 vm->addr = (void *)va->va_start;
2322 vm->size = va->va_end - va->va_start;
2323 vm->caller = caller;
2327 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2328 unsigned long flags, const void *caller)
2330 spin_lock(&vmap_area_lock);
2331 setup_vmalloc_vm_locked(vm, va, flags, caller);
2332 spin_unlock(&vmap_area_lock);
2335 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2338 * Before removing VM_UNINITIALIZED,
2339 * we should make sure that vm has proper values.
2340 * Pair with smp_rmb() in show_numa_info().
2343 vm->flags &= ~VM_UNINITIALIZED;
2346 static struct vm_struct *__get_vm_area_node(unsigned long size,
2347 unsigned long align, unsigned long shift, unsigned long flags,
2348 unsigned long start, unsigned long end, int node,
2349 gfp_t gfp_mask, const void *caller)
2351 struct vmap_area *va;
2352 struct vm_struct *area;
2353 unsigned long requested_size = size;
2355 BUG_ON(in_interrupt());
2356 size = ALIGN(size, 1ul << shift);
2357 if (unlikely(!size))
2360 if (flags & VM_IOREMAP)
2361 align = 1ul << clamp_t(int, get_count_order_long(size),
2362 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2364 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2365 if (unlikely(!area))
2368 if (!(flags & VM_NO_GUARD))
2371 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2377 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2379 setup_vmalloc_vm(area, va, flags, caller);
2384 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2385 unsigned long start, unsigned long end,
2388 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2389 NUMA_NO_NODE, GFP_KERNEL, caller);
2393 * get_vm_area - reserve a contiguous kernel virtual area
2394 * @size: size of the area
2395 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2397 * Search an area of @size in the kernel virtual mapping area,
2398 * and reserved it for out purposes. Returns the area descriptor
2399 * on success or %NULL on failure.
2401 * Return: the area descriptor on success or %NULL on failure.
2403 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2405 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2406 VMALLOC_START, VMALLOC_END,
2407 NUMA_NO_NODE, GFP_KERNEL,
2408 __builtin_return_address(0));
2411 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2414 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2415 VMALLOC_START, VMALLOC_END,
2416 NUMA_NO_NODE, GFP_KERNEL, caller);
2420 * find_vm_area - find a continuous kernel virtual area
2421 * @addr: base address
2423 * Search for the kernel VM area starting at @addr, and return it.
2424 * It is up to the caller to do all required locking to keep the returned
2427 * Return: the area descriptor on success or %NULL on failure.
2429 struct vm_struct *find_vm_area(const void *addr)
2431 struct vmap_area *va;
2433 va = find_vmap_area((unsigned long)addr);
2441 * remove_vm_area - find and remove a continuous kernel virtual area
2442 * @addr: base address
2444 * Search for the kernel VM area starting at @addr, and remove it.
2445 * This function returns the found VM area, but using it is NOT safe
2446 * on SMP machines, except for its size or flags.
2448 * Return: the area descriptor on success or %NULL on failure.
2450 struct vm_struct *remove_vm_area(const void *addr)
2452 struct vmap_area *va;
2456 spin_lock(&vmap_area_lock);
2457 va = __find_vmap_area((unsigned long)addr);
2459 struct vm_struct *vm = va->vm;
2462 spin_unlock(&vmap_area_lock);
2464 kasan_free_shadow(vm);
2465 free_unmap_vmap_area(va);
2470 spin_unlock(&vmap_area_lock);
2474 static inline void set_area_direct_map(const struct vm_struct *area,
2475 int (*set_direct_map)(struct page *page))
2479 /* HUGE_VMALLOC passes small pages to set_direct_map */
2480 for (i = 0; i < area->nr_pages; i++)
2481 if (page_address(area->pages[i]))
2482 set_direct_map(area->pages[i]);
2485 /* Handle removing and resetting vm mappings related to the vm_struct. */
2486 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2488 unsigned long start = ULONG_MAX, end = 0;
2489 unsigned int page_order = vm_area_page_order(area);
2490 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2494 remove_vm_area(area->addr);
2496 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2501 * If not deallocating pages, just do the flush of the VM area and
2504 if (!deallocate_pages) {
2510 * If execution gets here, flush the vm mapping and reset the direct
2511 * map. Find the start and end range of the direct mappings to make sure
2512 * the vm_unmap_aliases() flush includes the direct map.
2514 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2515 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2517 unsigned long page_size;
2519 page_size = PAGE_SIZE << page_order;
2520 start = min(addr, start);
2521 end = max(addr + page_size, end);
2527 * Set direct map to something invalid so that it won't be cached if
2528 * there are any accesses after the TLB flush, then flush the TLB and
2529 * reset the direct map permissions to the default.
2531 set_area_direct_map(area, set_direct_map_invalid_noflush);
2532 _vm_unmap_aliases(start, end, flush_dmap);
2533 set_area_direct_map(area, set_direct_map_default_noflush);
2536 static void __vunmap(const void *addr, int deallocate_pages)
2538 struct vm_struct *area;
2543 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2547 area = find_vm_area(addr);
2548 if (unlikely(!area)) {
2549 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2554 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2555 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2557 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2559 vm_remove_mappings(area, deallocate_pages);
2561 if (deallocate_pages) {
2562 unsigned int page_order = vm_area_page_order(area);
2565 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2566 struct page *page = area->pages[i];
2569 __free_pages(page, page_order);
2571 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2573 kvfree(area->pages);
2579 static inline void __vfree_deferred(const void *addr)
2582 * Use raw_cpu_ptr() because this can be called from preemptible
2583 * context. Preemption is absolutely fine here, because the llist_add()
2584 * implementation is lockless, so it works even if we are adding to
2585 * another cpu's list. schedule_work() should be fine with this too.
2587 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2589 if (llist_add((struct llist_node *)addr, &p->list))
2590 schedule_work(&p->wq);
2594 * vfree_atomic - release memory allocated by vmalloc()
2595 * @addr: memory base address
2597 * This one is just like vfree() but can be called in any atomic context
2600 void vfree_atomic(const void *addr)
2604 kmemleak_free(addr);
2608 __vfree_deferred(addr);
2611 static void __vfree(const void *addr)
2613 if (unlikely(in_interrupt()))
2614 __vfree_deferred(addr);
2620 * vfree - Release memory allocated by vmalloc()
2621 * @addr: Memory base address
2623 * Free the virtually continuous memory area starting at @addr, as obtained
2624 * from one of the vmalloc() family of APIs. This will usually also free the
2625 * physical memory underlying the virtual allocation, but that memory is
2626 * reference counted, so it will not be freed until the last user goes away.
2628 * If @addr is NULL, no operation is performed.
2631 * May sleep if called *not* from interrupt context.
2632 * Must not be called in NMI context (strictly speaking, it could be
2633 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2634 * conventions for vfree() arch-dependent would be a really bad idea).
2636 void vfree(const void *addr)
2640 kmemleak_free(addr);
2642 might_sleep_if(!in_interrupt());
2649 EXPORT_SYMBOL(vfree);
2652 * vunmap - release virtual mapping obtained by vmap()
2653 * @addr: memory base address
2655 * Free the virtually contiguous memory area starting at @addr,
2656 * which was created from the page array passed to vmap().
2658 * Must not be called in interrupt context.
2660 void vunmap(const void *addr)
2662 BUG_ON(in_interrupt());
2667 EXPORT_SYMBOL(vunmap);
2670 * vmap - map an array of pages into virtually contiguous space
2671 * @pages: array of page pointers
2672 * @count: number of pages to map
2673 * @flags: vm_area->flags
2674 * @prot: page protection for the mapping
2676 * Maps @count pages from @pages into contiguous kernel virtual space.
2677 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2678 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2679 * are transferred from the caller to vmap(), and will be freed / dropped when
2680 * vfree() is called on the return value.
2682 * Return: the address of the area or %NULL on failure
2684 void *vmap(struct page **pages, unsigned int count,
2685 unsigned long flags, pgprot_t prot)
2687 struct vm_struct *area;
2689 unsigned long size; /* In bytes */
2693 if (count > totalram_pages())
2696 size = (unsigned long)count << PAGE_SHIFT;
2697 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2701 addr = (unsigned long)area->addr;
2702 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2703 pages, PAGE_SHIFT) < 0) {
2708 if (flags & VM_MAP_PUT_PAGES) {
2709 area->pages = pages;
2710 area->nr_pages = count;
2714 EXPORT_SYMBOL(vmap);
2716 #ifdef CONFIG_VMAP_PFN
2717 struct vmap_pfn_data {
2718 unsigned long *pfns;
2723 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2725 struct vmap_pfn_data *data = private;
2727 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2729 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2734 * vmap_pfn - map an array of PFNs into virtually contiguous space
2735 * @pfns: array of PFNs
2736 * @count: number of pages to map
2737 * @prot: page protection for the mapping
2739 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2740 * the start address of the mapping.
2742 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2744 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2745 struct vm_struct *area;
2747 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2748 __builtin_return_address(0));
2751 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2752 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2758 EXPORT_SYMBOL_GPL(vmap_pfn);
2759 #endif /* CONFIG_VMAP_PFN */
2761 static inline unsigned int
2762 vm_area_alloc_pages(gfp_t gfp, int nid,
2763 unsigned int order, unsigned long nr_pages, struct page **pages)
2765 unsigned int nr_allocated = 0;
2768 * For order-0 pages we make use of bulk allocator, if
2769 * the page array is partly or not at all populated due
2770 * to fails, fallback to a single page allocator that is
2774 nr_allocated = alloc_pages_bulk_array_node(
2775 gfp, nid, nr_pages, pages);
2778 * Compound pages required for remap_vmalloc_page if
2783 /* High-order pages or fallback path if "bulk" fails. */
2784 while (nr_allocated < nr_pages) {
2788 page = alloc_pages_node(nid, gfp, order);
2789 if (unlikely(!page))
2793 * Careful, we allocate and map page-order pages, but
2794 * tracking is done per PAGE_SIZE page so as to keep the
2795 * vm_struct APIs independent of the physical/mapped size.
2797 for (i = 0; i < (1U << order); i++)
2798 pages[nr_allocated + i] = page + i;
2800 if (gfpflags_allow_blocking(gfp))
2803 nr_allocated += 1U << order;
2806 return nr_allocated;
2809 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2810 pgprot_t prot, unsigned int page_shift,
2813 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2814 unsigned long addr = (unsigned long)area->addr;
2815 unsigned long size = get_vm_area_size(area);
2816 unsigned long array_size;
2817 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2818 unsigned int page_order;
2820 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2821 gfp_mask |= __GFP_NOWARN;
2822 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2823 gfp_mask |= __GFP_HIGHMEM;
2825 /* Please note that the recursion is strictly bounded. */
2826 if (array_size > PAGE_SIZE) {
2827 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2830 area->pages = kmalloc_node(array_size, nested_gfp, node);
2834 warn_alloc(gfp_mask, NULL,
2835 "vmalloc error: size %lu, failed to allocated page array size %lu",
2836 nr_small_pages * PAGE_SIZE, array_size);
2841 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2842 page_order = vm_area_page_order(area);
2844 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2845 page_order, nr_small_pages, area->pages);
2847 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2850 * If not enough pages were obtained to accomplish an
2851 * allocation request, free them via __vfree() if any.
2853 if (area->nr_pages != nr_small_pages) {
2854 warn_alloc(gfp_mask, NULL,
2855 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2856 area->nr_pages * PAGE_SIZE, page_order);
2860 if (vmap_pages_range(addr, addr + size, prot, area->pages,
2862 warn_alloc(gfp_mask, NULL,
2863 "vmalloc error: size %lu, failed to map pages",
2864 area->nr_pages * PAGE_SIZE);
2871 __vfree(area->addr);
2876 * __vmalloc_node_range - allocate virtually contiguous memory
2877 * @size: allocation size
2878 * @align: desired alignment
2879 * @start: vm area range start
2880 * @end: vm area range end
2881 * @gfp_mask: flags for the page level allocator
2882 * @prot: protection mask for the allocated pages
2883 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2884 * @node: node to use for allocation or NUMA_NO_NODE
2885 * @caller: caller's return address
2887 * Allocate enough pages to cover @size from the page level
2888 * allocator with @gfp_mask flags. Map them into contiguous
2889 * kernel virtual space, using a pagetable protection of @prot.
2891 * Return: the address of the area or %NULL on failure
2893 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2894 unsigned long start, unsigned long end, gfp_t gfp_mask,
2895 pgprot_t prot, unsigned long vm_flags, int node,
2898 struct vm_struct *area;
2900 unsigned long real_size = size;
2901 unsigned long real_align = align;
2902 unsigned int shift = PAGE_SHIFT;
2904 if (WARN_ON_ONCE(!size))
2907 if ((size >> PAGE_SHIFT) > totalram_pages()) {
2908 warn_alloc(gfp_mask, NULL,
2909 "vmalloc error: size %lu, exceeds total pages",
2914 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP) &&
2915 arch_vmap_pmd_supported(prot)) {
2916 unsigned long size_per_node;
2919 * Try huge pages. Only try for PAGE_KERNEL allocations,
2920 * others like modules don't yet expect huge pages in
2921 * their allocations due to apply_to_page_range not
2925 size_per_node = size;
2926 if (node == NUMA_NO_NODE)
2927 size_per_node /= num_online_nodes();
2928 if (size_per_node >= PMD_SIZE) {
2930 align = max(real_align, 1UL << shift);
2931 size = ALIGN(real_size, 1UL << shift);
2936 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
2937 VM_UNINITIALIZED | vm_flags, start, end, node,
2940 warn_alloc(gfp_mask, NULL,
2941 "vmalloc error: size %lu, vm_struct allocation failed",
2946 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
2951 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2952 * flag. It means that vm_struct is not fully initialized.
2953 * Now, it is fully initialized, so remove this flag here.
2955 clear_vm_uninitialized_flag(area);
2957 size = PAGE_ALIGN(size);
2958 kmemleak_vmalloc(area, size, gfp_mask);
2963 if (shift > PAGE_SHIFT) {
2974 * __vmalloc_node - allocate virtually contiguous memory
2975 * @size: allocation size
2976 * @align: desired alignment
2977 * @gfp_mask: flags for the page level allocator
2978 * @node: node to use for allocation or NUMA_NO_NODE
2979 * @caller: caller's return address
2981 * Allocate enough pages to cover @size from the page level allocator with
2982 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2984 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2985 * and __GFP_NOFAIL are not supported
2987 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2990 * Return: pointer to the allocated memory or %NULL on error
2992 void *__vmalloc_node(unsigned long size, unsigned long align,
2993 gfp_t gfp_mask, int node, const void *caller)
2995 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2996 gfp_mask, PAGE_KERNEL, 0, node, caller);
2999 * This is only for performance analysis of vmalloc and stress purpose.
3000 * It is required by vmalloc test module, therefore do not use it other
3003 #ifdef CONFIG_TEST_VMALLOC_MODULE
3004 EXPORT_SYMBOL_GPL(__vmalloc_node);
3007 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3009 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3010 __builtin_return_address(0));
3012 EXPORT_SYMBOL(__vmalloc);
3015 * vmalloc - allocate virtually contiguous memory
3016 * @size: allocation size
3018 * Allocate enough pages to cover @size from the page level
3019 * allocator and map them into contiguous kernel virtual space.
3021 * For tight control over page level allocator and protection flags
3022 * use __vmalloc() instead.
3024 * Return: pointer to the allocated memory or %NULL on error
3026 void *vmalloc(unsigned long size)
3028 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3029 __builtin_return_address(0));
3031 EXPORT_SYMBOL(vmalloc);
3034 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3035 * @size: allocation size
3037 * Allocate enough non-huge pages to cover @size from the page level
3038 * allocator and map them into contiguous kernel virtual space.
3040 * Return: pointer to the allocated memory or %NULL on error
3042 void *vmalloc_no_huge(unsigned long size)
3044 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3045 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3046 NUMA_NO_NODE, __builtin_return_address(0));
3048 EXPORT_SYMBOL(vmalloc_no_huge);
3051 * vzalloc - allocate virtually contiguous memory with zero fill
3052 * @size: allocation size
3054 * Allocate enough pages to cover @size from the page level
3055 * allocator and map them into contiguous kernel virtual space.
3056 * The memory allocated is set to zero.
3058 * For tight control over page level allocator and protection flags
3059 * use __vmalloc() instead.
3061 * Return: pointer to the allocated memory or %NULL on error
3063 void *vzalloc(unsigned long size)
3065 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3066 __builtin_return_address(0));
3068 EXPORT_SYMBOL(vzalloc);
3071 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3072 * @size: allocation size
3074 * The resulting memory area is zeroed so it can be mapped to userspace
3075 * without leaking data.
3077 * Return: pointer to the allocated memory or %NULL on error
3079 void *vmalloc_user(unsigned long size)
3081 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3082 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3083 VM_USERMAP, NUMA_NO_NODE,
3084 __builtin_return_address(0));
3086 EXPORT_SYMBOL(vmalloc_user);
3089 * vmalloc_node - allocate memory on a specific node
3090 * @size: allocation size
3093 * Allocate enough pages to cover @size from the page level
3094 * allocator and map them into contiguous kernel virtual space.
3096 * For tight control over page level allocator and protection flags
3097 * use __vmalloc() instead.
3099 * Return: pointer to the allocated memory or %NULL on error
3101 void *vmalloc_node(unsigned long size, int node)
3103 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3104 __builtin_return_address(0));
3106 EXPORT_SYMBOL(vmalloc_node);
3109 * vzalloc_node - allocate memory on a specific node with zero fill
3110 * @size: allocation size
3113 * Allocate enough pages to cover @size from the page level
3114 * allocator and map them into contiguous kernel virtual space.
3115 * The memory allocated is set to zero.
3117 * Return: pointer to the allocated memory or %NULL on error
3119 void *vzalloc_node(unsigned long size, int node)
3121 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3122 __builtin_return_address(0));
3124 EXPORT_SYMBOL(vzalloc_node);
3126 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3127 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3128 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3129 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3132 * 64b systems should always have either DMA or DMA32 zones. For others
3133 * GFP_DMA32 should do the right thing and use the normal zone.
3135 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3139 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3140 * @size: allocation size
3142 * Allocate enough 32bit PA addressable pages to cover @size from the
3143 * page level allocator and map them into contiguous kernel virtual space.
3145 * Return: pointer to the allocated memory or %NULL on error
3147 void *vmalloc_32(unsigned long size)
3149 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3150 __builtin_return_address(0));
3152 EXPORT_SYMBOL(vmalloc_32);
3155 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3156 * @size: allocation size
3158 * The resulting memory area is 32bit addressable and zeroed so it can be
3159 * mapped to userspace without leaking data.
3161 * Return: pointer to the allocated memory or %NULL on error
3163 void *vmalloc_32_user(unsigned long size)
3165 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3166 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3167 VM_USERMAP, NUMA_NO_NODE,
3168 __builtin_return_address(0));
3170 EXPORT_SYMBOL(vmalloc_32_user);
3173 * small helper routine , copy contents to buf from addr.
3174 * If the page is not present, fill zero.
3177 static int aligned_vread(char *buf, char *addr, unsigned long count)
3183 unsigned long offset, length;
3185 offset = offset_in_page(addr);
3186 length = PAGE_SIZE - offset;
3189 p = vmalloc_to_page(addr);
3191 * To do safe access to this _mapped_ area, we need
3192 * lock. But adding lock here means that we need to add
3193 * overhead of vmalloc()/vfree() calls for this _debug_
3194 * interface, rarely used. Instead of that, we'll use
3195 * kmap() and get small overhead in this access function.
3198 /* We can expect USER0 is not used -- see vread() */
3199 void *map = kmap_atomic(p);
3200 memcpy(buf, map + offset, length);
3203 memset(buf, 0, length);
3214 * vread() - read vmalloc area in a safe way.
3215 * @buf: buffer for reading data
3216 * @addr: vm address.
3217 * @count: number of bytes to be read.
3219 * This function checks that addr is a valid vmalloc'ed area, and
3220 * copy data from that area to a given buffer. If the given memory range
3221 * of [addr...addr+count) includes some valid address, data is copied to
3222 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3223 * IOREMAP area is treated as memory hole and no copy is done.
3225 * If [addr...addr+count) doesn't includes any intersects with alive
3226 * vm_struct area, returns 0. @buf should be kernel's buffer.
3228 * Note: In usual ops, vread() is never necessary because the caller
3229 * should know vmalloc() area is valid and can use memcpy().
3230 * This is for routines which have to access vmalloc area without
3231 * any information, as /proc/kcore.
3233 * Return: number of bytes for which addr and buf should be increased
3234 * (same number as @count) or %0 if [addr...addr+count) doesn't
3235 * include any intersection with valid vmalloc area
3237 long vread(char *buf, char *addr, unsigned long count)
3239 struct vmap_area *va;
3240 struct vm_struct *vm;
3241 char *vaddr, *buf_start = buf;
3242 unsigned long buflen = count;
3245 /* Don't allow overflow */
3246 if ((unsigned long) addr + count < count)
3247 count = -(unsigned long) addr;
3249 spin_lock(&vmap_area_lock);
3250 va = __find_vmap_area((unsigned long)addr);
3253 list_for_each_entry_from(va, &vmap_area_list, list) {
3261 vaddr = (char *) vm->addr;
3262 if (addr >= vaddr + get_vm_area_size(vm))
3264 while (addr < vaddr) {
3272 n = vaddr + get_vm_area_size(vm) - addr;
3275 if (!(vm->flags & VM_IOREMAP))
3276 aligned_vread(buf, addr, n);
3277 else /* IOREMAP area is treated as memory hole */
3284 spin_unlock(&vmap_area_lock);
3286 if (buf == buf_start)
3288 /* zero-fill memory holes */
3289 if (buf != buf_start + buflen)
3290 memset(buf, 0, buflen - (buf - buf_start));
3296 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3297 * @vma: vma to cover
3298 * @uaddr: target user address to start at
3299 * @kaddr: virtual address of vmalloc kernel memory
3300 * @pgoff: offset from @kaddr to start at
3301 * @size: size of map area
3303 * Returns: 0 for success, -Exxx on failure
3305 * This function checks that @kaddr is a valid vmalloc'ed area,
3306 * and that it is big enough to cover the range starting at
3307 * @uaddr in @vma. Will return failure if that criteria isn't
3310 * Similar to remap_pfn_range() (see mm/memory.c)
3312 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3313 void *kaddr, unsigned long pgoff,
3316 struct vm_struct *area;
3318 unsigned long end_index;
3320 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3323 size = PAGE_ALIGN(size);
3325 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3328 area = find_vm_area(kaddr);
3332 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3335 if (check_add_overflow(size, off, &end_index) ||
3336 end_index > get_vm_area_size(area))
3341 struct page *page = vmalloc_to_page(kaddr);
3344 ret = vm_insert_page(vma, uaddr, page);
3353 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3359 * remap_vmalloc_range - map vmalloc pages to userspace
3360 * @vma: vma to cover (map full range of vma)
3361 * @addr: vmalloc memory
3362 * @pgoff: number of pages into addr before first page to map
3364 * Returns: 0 for success, -Exxx on failure
3366 * This function checks that addr is a valid vmalloc'ed area, and
3367 * that it is big enough to cover the vma. Will return failure if
3368 * that criteria isn't met.
3370 * Similar to remap_pfn_range() (see mm/memory.c)
3372 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3373 unsigned long pgoff)
3375 return remap_vmalloc_range_partial(vma, vma->vm_start,
3377 vma->vm_end - vma->vm_start);
3379 EXPORT_SYMBOL(remap_vmalloc_range);
3381 void free_vm_area(struct vm_struct *area)
3383 struct vm_struct *ret;
3384 ret = remove_vm_area(area->addr);
3385 BUG_ON(ret != area);
3388 EXPORT_SYMBOL_GPL(free_vm_area);
3391 static struct vmap_area *node_to_va(struct rb_node *n)
3393 return rb_entry_safe(n, struct vmap_area, rb_node);
3397 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3398 * @addr: target address
3400 * Returns: vmap_area if it is found. If there is no such area
3401 * the first highest(reverse order) vmap_area is returned
3402 * i.e. va->va_start < addr && va->va_end < addr or NULL
3403 * if there are no any areas before @addr.
3405 static struct vmap_area *
3406 pvm_find_va_enclose_addr(unsigned long addr)
3408 struct vmap_area *va, *tmp;
3411 n = free_vmap_area_root.rb_node;
3415 tmp = rb_entry(n, struct vmap_area, rb_node);
3416 if (tmp->va_start <= addr) {
3418 if (tmp->va_end >= addr)
3431 * pvm_determine_end_from_reverse - find the highest aligned address
3432 * of free block below VMALLOC_END
3434 * in - the VA we start the search(reverse order);
3435 * out - the VA with the highest aligned end address.
3436 * @align: alignment for required highest address
3438 * Returns: determined end address within vmap_area
3440 static unsigned long
3441 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3443 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3447 list_for_each_entry_from_reverse((*va),
3448 &free_vmap_area_list, list) {
3449 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3450 if ((*va)->va_start < addr)
3459 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3460 * @offsets: array containing offset of each area
3461 * @sizes: array containing size of each area
3462 * @nr_vms: the number of areas to allocate
3463 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3465 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3466 * vm_structs on success, %NULL on failure
3468 * Percpu allocator wants to use congruent vm areas so that it can
3469 * maintain the offsets among percpu areas. This function allocates
3470 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3471 * be scattered pretty far, distance between two areas easily going up
3472 * to gigabytes. To avoid interacting with regular vmallocs, these
3473 * areas are allocated from top.
3475 * Despite its complicated look, this allocator is rather simple. It
3476 * does everything top-down and scans free blocks from the end looking
3477 * for matching base. While scanning, if any of the areas do not fit the
3478 * base address is pulled down to fit the area. Scanning is repeated till
3479 * all the areas fit and then all necessary data structures are inserted
3480 * and the result is returned.
3482 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3483 const size_t *sizes, int nr_vms,
3486 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3487 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3488 struct vmap_area **vas, *va;
3489 struct vm_struct **vms;
3490 int area, area2, last_area, term_area;
3491 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3492 bool purged = false;
3495 /* verify parameters and allocate data structures */
3496 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3497 for (last_area = 0, area = 0; area < nr_vms; area++) {
3498 start = offsets[area];
3499 end = start + sizes[area];
3501 /* is everything aligned properly? */
3502 BUG_ON(!IS_ALIGNED(offsets[area], align));
3503 BUG_ON(!IS_ALIGNED(sizes[area], align));
3505 /* detect the area with the highest address */
3506 if (start > offsets[last_area])
3509 for (area2 = area + 1; area2 < nr_vms; area2++) {
3510 unsigned long start2 = offsets[area2];
3511 unsigned long end2 = start2 + sizes[area2];
3513 BUG_ON(start2 < end && start < end2);
3516 last_end = offsets[last_area] + sizes[last_area];
3518 if (vmalloc_end - vmalloc_start < last_end) {
3523 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3524 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3528 for (area = 0; area < nr_vms; area++) {
3529 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3530 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3531 if (!vas[area] || !vms[area])
3535 spin_lock(&free_vmap_area_lock);
3537 /* start scanning - we scan from the top, begin with the last area */
3538 area = term_area = last_area;
3539 start = offsets[area];
3540 end = start + sizes[area];
3542 va = pvm_find_va_enclose_addr(vmalloc_end);
3543 base = pvm_determine_end_from_reverse(&va, align) - end;
3547 * base might have underflowed, add last_end before
3550 if (base + last_end < vmalloc_start + last_end)
3554 * Fitting base has not been found.
3560 * If required width exceeds current VA block, move
3561 * base downwards and then recheck.
3563 if (base + end > va->va_end) {
3564 base = pvm_determine_end_from_reverse(&va, align) - end;
3570 * If this VA does not fit, move base downwards and recheck.
3572 if (base + start < va->va_start) {
3573 va = node_to_va(rb_prev(&va->rb_node));
3574 base = pvm_determine_end_from_reverse(&va, align) - end;
3580 * This area fits, move on to the previous one. If
3581 * the previous one is the terminal one, we're done.
3583 area = (area + nr_vms - 1) % nr_vms;
3584 if (area == term_area)
3587 start = offsets[area];
3588 end = start + sizes[area];
3589 va = pvm_find_va_enclose_addr(base + end);
3592 /* we've found a fitting base, insert all va's */
3593 for (area = 0; area < nr_vms; area++) {
3596 start = base + offsets[area];
3599 va = pvm_find_va_enclose_addr(start);
3600 if (WARN_ON_ONCE(va == NULL))
3601 /* It is a BUG(), but trigger recovery instead. */
3604 type = classify_va_fit_type(va, start, size);
3605 if (WARN_ON_ONCE(type == NOTHING_FIT))
3606 /* It is a BUG(), but trigger recovery instead. */
3609 ret = adjust_va_to_fit_type(va, start, size, type);
3613 /* Allocated area. */
3615 va->va_start = start;
3616 va->va_end = start + size;
3619 spin_unlock(&free_vmap_area_lock);
3621 /* populate the kasan shadow space */
3622 for (area = 0; area < nr_vms; area++) {
3623 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3624 goto err_free_shadow;
3626 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3630 /* insert all vm's */
3631 spin_lock(&vmap_area_lock);
3632 for (area = 0; area < nr_vms; area++) {
3633 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3635 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3638 spin_unlock(&vmap_area_lock);
3645 * Remove previously allocated areas. There is no
3646 * need in removing these areas from the busy tree,
3647 * because they are inserted only on the final step
3648 * and when pcpu_get_vm_areas() is success.
3651 orig_start = vas[area]->va_start;
3652 orig_end = vas[area]->va_end;
3653 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3654 &free_vmap_area_list);
3656 kasan_release_vmalloc(orig_start, orig_end,
3657 va->va_start, va->va_end);
3662 spin_unlock(&free_vmap_area_lock);
3664 purge_vmap_area_lazy();
3667 /* Before "retry", check if we recover. */
3668 for (area = 0; area < nr_vms; area++) {
3672 vas[area] = kmem_cache_zalloc(
3673 vmap_area_cachep, GFP_KERNEL);
3682 for (area = 0; area < nr_vms; area++) {
3684 kmem_cache_free(vmap_area_cachep, vas[area]);
3694 spin_lock(&free_vmap_area_lock);
3696 * We release all the vmalloc shadows, even the ones for regions that
3697 * hadn't been successfully added. This relies on kasan_release_vmalloc
3698 * being able to tolerate this case.
3700 for (area = 0; area < nr_vms; area++) {
3701 orig_start = vas[area]->va_start;
3702 orig_end = vas[area]->va_end;
3703 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3704 &free_vmap_area_list);
3706 kasan_release_vmalloc(orig_start, orig_end,
3707 va->va_start, va->va_end);
3711 spin_unlock(&free_vmap_area_lock);
3718 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3719 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3720 * @nr_vms: the number of allocated areas
3722 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3724 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3728 for (i = 0; i < nr_vms; i++)
3729 free_vm_area(vms[i]);
3732 #endif /* CONFIG_SMP */
3734 #ifdef CONFIG_PRINTK
3735 bool vmalloc_dump_obj(void *object)
3737 struct vm_struct *vm;
3738 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3740 vm = find_vm_area(objp);
3743 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3744 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3749 #ifdef CONFIG_PROC_FS
3750 static void *s_start(struct seq_file *m, loff_t *pos)
3751 __acquires(&vmap_purge_lock)
3752 __acquires(&vmap_area_lock)
3754 mutex_lock(&vmap_purge_lock);
3755 spin_lock(&vmap_area_lock);
3757 return seq_list_start(&vmap_area_list, *pos);
3760 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3762 return seq_list_next(p, &vmap_area_list, pos);
3765 static void s_stop(struct seq_file *m, void *p)
3766 __releases(&vmap_area_lock)
3767 __releases(&vmap_purge_lock)
3769 spin_unlock(&vmap_area_lock);
3770 mutex_unlock(&vmap_purge_lock);
3773 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3775 if (IS_ENABLED(CONFIG_NUMA)) {
3776 unsigned int nr, *counters = m->private;
3781 if (v->flags & VM_UNINITIALIZED)
3783 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3786 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3788 for (nr = 0; nr < v->nr_pages; nr++)
3789 counters[page_to_nid(v->pages[nr])]++;
3791 for_each_node_state(nr, N_HIGH_MEMORY)
3793 seq_printf(m, " N%u=%u", nr, counters[nr]);
3797 static void show_purge_info(struct seq_file *m)
3799 struct vmap_area *va;
3801 spin_lock(&purge_vmap_area_lock);
3802 list_for_each_entry(va, &purge_vmap_area_list, list) {
3803 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3804 (void *)va->va_start, (void *)va->va_end,
3805 va->va_end - va->va_start);
3807 spin_unlock(&purge_vmap_area_lock);
3810 static int s_show(struct seq_file *m, void *p)
3812 struct vmap_area *va;
3813 struct vm_struct *v;
3815 va = list_entry(p, struct vmap_area, list);
3818 * s_show can encounter race with remove_vm_area, !vm on behalf
3819 * of vmap area is being tear down or vm_map_ram allocation.
3822 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3823 (void *)va->va_start, (void *)va->va_end,
3824 va->va_end - va->va_start);
3831 seq_printf(m, "0x%pK-0x%pK %7ld",
3832 v->addr, v->addr + v->size, v->size);
3835 seq_printf(m, " %pS", v->caller);
3838 seq_printf(m, " pages=%d", v->nr_pages);
3841 seq_printf(m, " phys=%pa", &v->phys_addr);
3843 if (v->flags & VM_IOREMAP)
3844 seq_puts(m, " ioremap");
3846 if (v->flags & VM_ALLOC)
3847 seq_puts(m, " vmalloc");
3849 if (v->flags & VM_MAP)
3850 seq_puts(m, " vmap");
3852 if (v->flags & VM_USERMAP)
3853 seq_puts(m, " user");
3855 if (v->flags & VM_DMA_COHERENT)
3856 seq_puts(m, " dma-coherent");
3858 if (is_vmalloc_addr(v->pages))
3859 seq_puts(m, " vpages");
3861 show_numa_info(m, v);
3865 * As a final step, dump "unpurged" areas.
3867 if (list_is_last(&va->list, &vmap_area_list))
3873 static const struct seq_operations vmalloc_op = {
3880 static int __init proc_vmalloc_init(void)
3882 if (IS_ENABLED(CONFIG_NUMA))
3883 proc_create_seq_private("vmallocinfo", 0400, NULL,
3885 nr_node_ids * sizeof(unsigned int), NULL);
3887 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3890 module_init(proc_vmalloc_init);