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>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
47 #include "pgalloc-track.h"
49 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
50 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
52 static int __init set_nohugeiomap(char *str)
54 ioremap_max_page_shift = PAGE_SHIFT;
57 early_param("nohugeiomap", set_nohugeiomap);
58 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
59 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
60 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
63 static bool __ro_after_init vmap_allow_huge = true;
65 static int __init set_nohugevmalloc(char *str)
67 vmap_allow_huge = false;
70 early_param("nohugevmalloc", set_nohugevmalloc);
71 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
72 static const bool vmap_allow_huge = false;
73 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 bool is_vmalloc_addr(const void *x)
77 unsigned long addr = (unsigned long)kasan_reset_tag(x);
79 return addr >= VMALLOC_START && addr < VMALLOC_END;
81 EXPORT_SYMBOL(is_vmalloc_addr);
83 struct vfree_deferred {
84 struct llist_head list;
85 struct work_struct wq;
87 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
89 static void __vunmap(const void *, int);
91 static void free_work(struct work_struct *w)
93 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
94 struct llist_node *t, *llnode;
96 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
97 __vunmap((void *)llnode, 1);
100 /*** Page table manipulation functions ***/
101 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
102 phys_addr_t phys_addr, pgprot_t prot,
103 unsigned int max_page_shift, pgtbl_mod_mask *mask)
107 unsigned long size = PAGE_SIZE;
109 pfn = phys_addr >> PAGE_SHIFT;
110 pte = pte_alloc_kernel_track(pmd, addr, mask);
114 BUG_ON(!pte_none(*pte));
116 #ifdef CONFIG_HUGETLB_PAGE
117 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
118 if (size != PAGE_SIZE) {
119 pte_t entry = pfn_pte(pfn, prot);
121 entry = arch_make_huge_pte(entry, ilog2(size), 0);
122 set_huge_pte_at(&init_mm, addr, pte, entry);
123 pfn += PFN_DOWN(size);
127 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
129 } while (pte += PFN_DOWN(size), addr += size, addr != end);
130 *mask |= PGTBL_PTE_MODIFIED;
134 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
135 phys_addr_t phys_addr, pgprot_t prot,
136 unsigned int max_page_shift)
138 if (max_page_shift < PMD_SHIFT)
141 if (!arch_vmap_pmd_supported(prot))
144 if ((end - addr) != PMD_SIZE)
147 if (!IS_ALIGNED(addr, PMD_SIZE))
150 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
153 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
156 return pmd_set_huge(pmd, phys_addr, prot);
159 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
160 phys_addr_t phys_addr, pgprot_t prot,
161 unsigned int max_page_shift, pgtbl_mod_mask *mask)
166 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
170 next = pmd_addr_end(addr, end);
172 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
174 *mask |= PGTBL_PMD_MODIFIED;
178 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
180 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
184 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
185 phys_addr_t phys_addr, pgprot_t prot,
186 unsigned int max_page_shift)
188 if (max_page_shift < PUD_SHIFT)
191 if (!arch_vmap_pud_supported(prot))
194 if ((end - addr) != PUD_SIZE)
197 if (!IS_ALIGNED(addr, PUD_SIZE))
200 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
203 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
206 return pud_set_huge(pud, phys_addr, prot);
209 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
210 phys_addr_t phys_addr, pgprot_t prot,
211 unsigned int max_page_shift, pgtbl_mod_mask *mask)
216 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
220 next = pud_addr_end(addr, end);
222 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
224 *mask |= PGTBL_PUD_MODIFIED;
228 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
229 max_page_shift, mask))
231 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
235 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
236 phys_addr_t phys_addr, pgprot_t prot,
237 unsigned int max_page_shift)
239 if (max_page_shift < P4D_SHIFT)
242 if (!arch_vmap_p4d_supported(prot))
245 if ((end - addr) != P4D_SIZE)
248 if (!IS_ALIGNED(addr, P4D_SIZE))
251 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
254 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
257 return p4d_set_huge(p4d, phys_addr, prot);
260 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
261 phys_addr_t phys_addr, pgprot_t prot,
262 unsigned int max_page_shift, pgtbl_mod_mask *mask)
267 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
271 next = p4d_addr_end(addr, end);
273 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
275 *mask |= PGTBL_P4D_MODIFIED;
279 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
280 max_page_shift, mask))
282 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
286 static int vmap_range_noflush(unsigned long addr, unsigned long end,
287 phys_addr_t phys_addr, pgprot_t prot,
288 unsigned int max_page_shift)
294 pgtbl_mod_mask mask = 0;
300 pgd = pgd_offset_k(addr);
302 next = pgd_addr_end(addr, end);
303 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
304 max_page_shift, &mask);
307 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
309 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
310 arch_sync_kernel_mappings(start, end);
315 int ioremap_page_range(unsigned long addr, unsigned long end,
316 phys_addr_t phys_addr, pgprot_t prot)
320 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
321 ioremap_max_page_shift);
322 flush_cache_vmap(addr, end);
324 kmsan_ioremap_page_range(addr, end, phys_addr, prot,
325 ioremap_max_page_shift);
329 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
330 pgtbl_mod_mask *mask)
334 pte = pte_offset_kernel(pmd, addr);
336 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
337 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
338 } while (pte++, addr += PAGE_SIZE, addr != end);
339 *mask |= PGTBL_PTE_MODIFIED;
342 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
343 pgtbl_mod_mask *mask)
349 pmd = pmd_offset(pud, addr);
351 next = pmd_addr_end(addr, end);
353 cleared = pmd_clear_huge(pmd);
354 if (cleared || pmd_bad(*pmd))
355 *mask |= PGTBL_PMD_MODIFIED;
359 if (pmd_none_or_clear_bad(pmd))
361 vunmap_pte_range(pmd, addr, next, mask);
364 } while (pmd++, addr = next, addr != end);
367 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
368 pgtbl_mod_mask *mask)
374 pud = pud_offset(p4d, addr);
376 next = pud_addr_end(addr, end);
378 cleared = pud_clear_huge(pud);
379 if (cleared || pud_bad(*pud))
380 *mask |= PGTBL_PUD_MODIFIED;
384 if (pud_none_or_clear_bad(pud))
386 vunmap_pmd_range(pud, addr, next, mask);
387 } while (pud++, addr = next, addr != end);
390 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
391 pgtbl_mod_mask *mask)
396 p4d = p4d_offset(pgd, addr);
398 next = p4d_addr_end(addr, end);
402 *mask |= PGTBL_P4D_MODIFIED;
404 if (p4d_none_or_clear_bad(p4d))
406 vunmap_pud_range(p4d, addr, next, mask);
407 } while (p4d++, addr = next, addr != end);
411 * vunmap_range_noflush is similar to vunmap_range, but does not
412 * flush caches or TLBs.
414 * The caller is responsible for calling flush_cache_vmap() before calling
415 * this function, and flush_tlb_kernel_range after it has returned
416 * successfully (and before the addresses are expected to cause a page fault
417 * or be re-mapped for something else, if TLB flushes are being delayed or
420 * This is an internal function only. Do not use outside mm/.
422 void __vunmap_range_noflush(unsigned long start, unsigned long end)
426 unsigned long addr = start;
427 pgtbl_mod_mask mask = 0;
430 pgd = pgd_offset_k(addr);
432 next = pgd_addr_end(addr, end);
434 mask |= PGTBL_PGD_MODIFIED;
435 if (pgd_none_or_clear_bad(pgd))
437 vunmap_p4d_range(pgd, addr, next, &mask);
438 } while (pgd++, addr = next, addr != end);
440 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
441 arch_sync_kernel_mappings(start, end);
444 void vunmap_range_noflush(unsigned long start, unsigned long end)
446 kmsan_vunmap_range_noflush(start, end);
447 __vunmap_range_noflush(start, end);
451 * vunmap_range - unmap kernel virtual addresses
452 * @addr: start of the VM area to unmap
453 * @end: end of the VM area to unmap (non-inclusive)
455 * Clears any present PTEs in the virtual address range, flushes TLBs and
456 * caches. Any subsequent access to the address before it has been re-mapped
459 void vunmap_range(unsigned long addr, unsigned long end)
461 flush_cache_vunmap(addr, end);
462 vunmap_range_noflush(addr, end);
463 flush_tlb_kernel_range(addr, end);
466 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
467 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
468 pgtbl_mod_mask *mask)
473 * nr is a running index into the array which helps higher level
474 * callers keep track of where we're up to.
477 pte = pte_alloc_kernel_track(pmd, addr, mask);
481 struct page *page = pages[*nr];
483 if (WARN_ON(!pte_none(*pte)))
487 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
490 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
492 } while (pte++, addr += PAGE_SIZE, addr != end);
493 *mask |= PGTBL_PTE_MODIFIED;
497 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
498 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
499 pgtbl_mod_mask *mask)
504 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
508 next = pmd_addr_end(addr, end);
509 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
511 } while (pmd++, addr = next, addr != end);
515 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
516 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
517 pgtbl_mod_mask *mask)
522 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
526 next = pud_addr_end(addr, end);
527 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
529 } while (pud++, addr = next, addr != end);
533 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
534 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
535 pgtbl_mod_mask *mask)
540 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
544 next = p4d_addr_end(addr, end);
545 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
547 } while (p4d++, addr = next, addr != end);
551 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
552 pgprot_t prot, struct page **pages)
554 unsigned long start = addr;
559 pgtbl_mod_mask mask = 0;
562 pgd = pgd_offset_k(addr);
564 next = pgd_addr_end(addr, end);
566 mask |= PGTBL_PGD_MODIFIED;
567 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
570 } while (pgd++, addr = next, addr != end);
572 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
573 arch_sync_kernel_mappings(start, end);
579 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
582 * The caller is responsible for calling flush_cache_vmap() after this
583 * function returns successfully and before the addresses are accessed.
585 * This is an internal function only. Do not use outside mm/.
587 int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
588 pgprot_t prot, struct page **pages, unsigned int page_shift)
590 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
592 WARN_ON(page_shift < PAGE_SHIFT);
594 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
595 page_shift == PAGE_SHIFT)
596 return vmap_small_pages_range_noflush(addr, end, prot, pages);
598 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
601 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
602 page_to_phys(pages[i]), prot,
607 addr += 1UL << page_shift;
613 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
614 pgprot_t prot, struct page **pages, unsigned int page_shift)
616 kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
617 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
621 * vmap_pages_range - map pages to a kernel virtual address
622 * @addr: start of the VM area to map
623 * @end: end of the VM area to map (non-inclusive)
624 * @prot: page protection flags to use
625 * @pages: pages to map (always PAGE_SIZE pages)
626 * @page_shift: maximum shift that the pages may be mapped with, @pages must
627 * be aligned and contiguous up to at least this shift.
630 * 0 on success, -errno on failure.
632 static int vmap_pages_range(unsigned long addr, unsigned long end,
633 pgprot_t prot, struct page **pages, unsigned int page_shift)
637 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
638 flush_cache_vmap(addr, end);
642 int is_vmalloc_or_module_addr(const void *x)
645 * ARM, x86-64 and sparc64 put modules in a special place,
646 * and fall back on vmalloc() if that fails. Others
647 * just put it in the vmalloc space.
649 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
650 unsigned long addr = (unsigned long)kasan_reset_tag(x);
651 if (addr >= MODULES_VADDR && addr < MODULES_END)
654 return is_vmalloc_addr(x);
658 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
659 * return the tail page that corresponds to the base page address, which
660 * matches small vmap mappings.
662 struct page *vmalloc_to_page(const void *vmalloc_addr)
664 unsigned long addr = (unsigned long) vmalloc_addr;
665 struct page *page = NULL;
666 pgd_t *pgd = pgd_offset_k(addr);
673 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
674 * architectures that do not vmalloc module space
676 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
680 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
681 return NULL; /* XXX: no allowance for huge pgd */
682 if (WARN_ON_ONCE(pgd_bad(*pgd)))
685 p4d = p4d_offset(pgd, addr);
689 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
690 if (WARN_ON_ONCE(p4d_bad(*p4d)))
693 pud = pud_offset(p4d, addr);
697 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
698 if (WARN_ON_ONCE(pud_bad(*pud)))
701 pmd = pmd_offset(pud, addr);
705 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
706 if (WARN_ON_ONCE(pmd_bad(*pmd)))
709 ptep = pte_offset_map(pmd, addr);
711 if (pte_present(pte))
712 page = pte_page(pte);
717 EXPORT_SYMBOL(vmalloc_to_page);
720 * Map a vmalloc()-space virtual address to the physical page frame number.
722 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
724 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
726 EXPORT_SYMBOL(vmalloc_to_pfn);
729 /*** Global kva allocator ***/
731 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
732 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
735 static DEFINE_SPINLOCK(vmap_area_lock);
736 static DEFINE_SPINLOCK(free_vmap_area_lock);
737 /* Export for kexec only */
738 LIST_HEAD(vmap_area_list);
739 static struct rb_root vmap_area_root = RB_ROOT;
740 static bool vmap_initialized __read_mostly;
742 static struct rb_root purge_vmap_area_root = RB_ROOT;
743 static LIST_HEAD(purge_vmap_area_list);
744 static DEFINE_SPINLOCK(purge_vmap_area_lock);
747 * This kmem_cache is used for vmap_area objects. Instead of
748 * allocating from slab we reuse an object from this cache to
749 * make things faster. Especially in "no edge" splitting of
752 static struct kmem_cache *vmap_area_cachep;
755 * This linked list is used in pair with free_vmap_area_root.
756 * It gives O(1) access to prev/next to perform fast coalescing.
758 static LIST_HEAD(free_vmap_area_list);
761 * This augment red-black tree represents the free vmap space.
762 * All vmap_area objects in this tree are sorted by va->va_start
763 * address. It is used for allocation and merging when a vmap
764 * object is released.
766 * Each vmap_area node contains a maximum available free block
767 * of its sub-tree, right or left. Therefore it is possible to
768 * find a lowest match of free area.
770 static struct rb_root free_vmap_area_root = RB_ROOT;
773 * Preload a CPU with one object for "no edge" split case. The
774 * aim is to get rid of allocations from the atomic context, thus
775 * to use more permissive allocation masks.
777 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
779 static __always_inline unsigned long
780 va_size(struct vmap_area *va)
782 return (va->va_end - va->va_start);
785 static __always_inline unsigned long
786 get_subtree_max_size(struct rb_node *node)
788 struct vmap_area *va;
790 va = rb_entry_safe(node, struct vmap_area, rb_node);
791 return va ? va->subtree_max_size : 0;
794 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
795 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
797 static void purge_vmap_area_lazy(void);
798 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
799 static void drain_vmap_area_work(struct work_struct *work);
800 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
802 static atomic_long_t nr_vmalloc_pages;
804 unsigned long vmalloc_nr_pages(void)
806 return atomic_long_read(&nr_vmalloc_pages);
809 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
810 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
812 struct vmap_area *va = NULL;
813 struct rb_node *n = vmap_area_root.rb_node;
815 addr = (unsigned long)kasan_reset_tag((void *)addr);
818 struct vmap_area *tmp;
820 tmp = rb_entry(n, struct vmap_area, rb_node);
821 if (tmp->va_end > addr) {
823 if (tmp->va_start <= addr)
834 static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
836 struct rb_node *n = root->rb_node;
838 addr = (unsigned long)kasan_reset_tag((void *)addr);
841 struct vmap_area *va;
843 va = rb_entry(n, struct vmap_area, rb_node);
844 if (addr < va->va_start)
846 else if (addr >= va->va_end)
856 * This function returns back addresses of parent node
857 * and its left or right link for further processing.
859 * Otherwise NULL is returned. In that case all further
860 * steps regarding inserting of conflicting overlap range
861 * have to be declined and actually considered as a bug.
863 static __always_inline struct rb_node **
864 find_va_links(struct vmap_area *va,
865 struct rb_root *root, struct rb_node *from,
866 struct rb_node **parent)
868 struct vmap_area *tmp_va;
869 struct rb_node **link;
872 link = &root->rb_node;
873 if (unlikely(!*link)) {
882 * Go to the bottom of the tree. When we hit the last point
883 * we end up with parent rb_node and correct direction, i name
884 * it link, where the new va->rb_node will be attached to.
887 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
890 * During the traversal we also do some sanity check.
891 * Trigger the BUG() if there are sides(left/right)
894 if (va->va_end <= tmp_va->va_start)
895 link = &(*link)->rb_left;
896 else if (va->va_start >= tmp_va->va_end)
897 link = &(*link)->rb_right;
899 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
900 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
906 *parent = &tmp_va->rb_node;
910 static __always_inline struct list_head *
911 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
913 struct list_head *list;
915 if (unlikely(!parent))
917 * The red-black tree where we try to find VA neighbors
918 * before merging or inserting is empty, i.e. it means
919 * there is no free vmap space. Normally it does not
920 * happen but we handle this case anyway.
924 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
925 return (&parent->rb_right == link ? list->next : list);
928 static __always_inline void
929 __link_va(struct vmap_area *va, struct rb_root *root,
930 struct rb_node *parent, struct rb_node **link,
931 struct list_head *head, bool augment)
934 * VA is still not in the list, but we can
935 * identify its future previous list_head node.
937 if (likely(parent)) {
938 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
939 if (&parent->rb_right != link)
943 /* Insert to the rb-tree */
944 rb_link_node(&va->rb_node, parent, link);
947 * Some explanation here. Just perform simple insertion
948 * to the tree. We do not set va->subtree_max_size to
949 * its current size before calling rb_insert_augmented().
950 * It is because we populate the tree from the bottom
951 * to parent levels when the node _is_ in the tree.
953 * Therefore we set subtree_max_size to zero after insertion,
954 * to let __augment_tree_propagate_from() puts everything to
955 * the correct order later on.
957 rb_insert_augmented(&va->rb_node,
958 root, &free_vmap_area_rb_augment_cb);
959 va->subtree_max_size = 0;
961 rb_insert_color(&va->rb_node, root);
964 /* Address-sort this list */
965 list_add(&va->list, head);
968 static __always_inline void
969 link_va(struct vmap_area *va, struct rb_root *root,
970 struct rb_node *parent, struct rb_node **link,
971 struct list_head *head)
973 __link_va(va, root, parent, link, head, false);
976 static __always_inline void
977 link_va_augment(struct vmap_area *va, struct rb_root *root,
978 struct rb_node *parent, struct rb_node **link,
979 struct list_head *head)
981 __link_va(va, root, parent, link, head, true);
984 static __always_inline void
985 __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
987 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
991 rb_erase_augmented(&va->rb_node,
992 root, &free_vmap_area_rb_augment_cb);
994 rb_erase(&va->rb_node, root);
996 list_del_init(&va->list);
997 RB_CLEAR_NODE(&va->rb_node);
1000 static __always_inline void
1001 unlink_va(struct vmap_area *va, struct rb_root *root)
1003 __unlink_va(va, root, false);
1006 static __always_inline void
1007 unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1009 __unlink_va(va, root, true);
1012 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1014 * Gets called when remove the node and rotate.
1016 static __always_inline unsigned long
1017 compute_subtree_max_size(struct vmap_area *va)
1019 return max3(va_size(va),
1020 get_subtree_max_size(va->rb_node.rb_left),
1021 get_subtree_max_size(va->rb_node.rb_right));
1025 augment_tree_propagate_check(void)
1027 struct vmap_area *va;
1028 unsigned long computed_size;
1030 list_for_each_entry(va, &free_vmap_area_list, list) {
1031 computed_size = compute_subtree_max_size(va);
1032 if (computed_size != va->subtree_max_size)
1033 pr_emerg("tree is corrupted: %lu, %lu\n",
1034 va_size(va), va->subtree_max_size);
1040 * This function populates subtree_max_size from bottom to upper
1041 * levels starting from VA point. The propagation must be done
1042 * when VA size is modified by changing its va_start/va_end. Or
1043 * in case of newly inserting of VA to the tree.
1045 * It means that __augment_tree_propagate_from() must be called:
1046 * - After VA has been inserted to the tree(free path);
1047 * - After VA has been shrunk(allocation path);
1048 * - After VA has been increased(merging path).
1050 * Please note that, it does not mean that upper parent nodes
1051 * and their subtree_max_size are recalculated all the time up
1060 * For example if we modify the node 4, shrinking it to 2, then
1061 * no any modification is required. If we shrink the node 2 to 1
1062 * its subtree_max_size is updated only, and set to 1. If we shrink
1063 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1064 * node becomes 4--6.
1066 static __always_inline void
1067 augment_tree_propagate_from(struct vmap_area *va)
1070 * Populate the tree from bottom towards the root until
1071 * the calculated maximum available size of checked node
1072 * is equal to its current one.
1074 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1076 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1077 augment_tree_propagate_check();
1082 insert_vmap_area(struct vmap_area *va,
1083 struct rb_root *root, struct list_head *head)
1085 struct rb_node **link;
1086 struct rb_node *parent;
1088 link = find_va_links(va, root, NULL, &parent);
1090 link_va(va, root, parent, link, head);
1094 insert_vmap_area_augment(struct vmap_area *va,
1095 struct rb_node *from, struct rb_root *root,
1096 struct list_head *head)
1098 struct rb_node **link;
1099 struct rb_node *parent;
1102 link = find_va_links(va, NULL, from, &parent);
1104 link = find_va_links(va, root, NULL, &parent);
1107 link_va_augment(va, root, parent, link, head);
1108 augment_tree_propagate_from(va);
1113 * Merge de-allocated chunk of VA memory with previous
1114 * and next free blocks. If coalesce is not done a new
1115 * free area is inserted. If VA has been merged, it is
1118 * Please note, it can return NULL in case of overlap
1119 * ranges, followed by WARN() report. Despite it is a
1120 * buggy behaviour, a system can be alive and keep
1123 static __always_inline struct vmap_area *
1124 __merge_or_add_vmap_area(struct vmap_area *va,
1125 struct rb_root *root, struct list_head *head, bool augment)
1127 struct vmap_area *sibling;
1128 struct list_head *next;
1129 struct rb_node **link;
1130 struct rb_node *parent;
1131 bool merged = false;
1134 * Find a place in the tree where VA potentially will be
1135 * inserted, unless it is merged with its sibling/siblings.
1137 link = find_va_links(va, root, NULL, &parent);
1142 * Get next node of VA to check if merging can be done.
1144 next = get_va_next_sibling(parent, link);
1145 if (unlikely(next == NULL))
1151 * |<------VA------>|<-----Next----->|
1156 sibling = list_entry(next, struct vmap_area, list);
1157 if (sibling->va_start == va->va_end) {
1158 sibling->va_start = va->va_start;
1160 /* Free vmap_area object. */
1161 kmem_cache_free(vmap_area_cachep, va);
1163 /* Point to the new merged area. */
1172 * |<-----Prev----->|<------VA------>|
1176 if (next->prev != head) {
1177 sibling = list_entry(next->prev, struct vmap_area, list);
1178 if (sibling->va_end == va->va_start) {
1180 * If both neighbors are coalesced, it is important
1181 * to unlink the "next" node first, followed by merging
1182 * with "previous" one. Otherwise the tree might not be
1183 * fully populated if a sibling's augmented value is
1184 * "normalized" because of rotation operations.
1187 __unlink_va(va, root, augment);
1189 sibling->va_end = va->va_end;
1191 /* Free vmap_area object. */
1192 kmem_cache_free(vmap_area_cachep, va);
1194 /* Point to the new merged area. */
1202 __link_va(va, root, parent, link, head, augment);
1207 static __always_inline struct vmap_area *
1208 merge_or_add_vmap_area(struct vmap_area *va,
1209 struct rb_root *root, struct list_head *head)
1211 return __merge_or_add_vmap_area(va, root, head, false);
1214 static __always_inline struct vmap_area *
1215 merge_or_add_vmap_area_augment(struct vmap_area *va,
1216 struct rb_root *root, struct list_head *head)
1218 va = __merge_or_add_vmap_area(va, root, head, true);
1220 augment_tree_propagate_from(va);
1225 static __always_inline bool
1226 is_within_this_va(struct vmap_area *va, unsigned long size,
1227 unsigned long align, unsigned long vstart)
1229 unsigned long nva_start_addr;
1231 if (va->va_start > vstart)
1232 nva_start_addr = ALIGN(va->va_start, align);
1234 nva_start_addr = ALIGN(vstart, align);
1236 /* Can be overflowed due to big size or alignment. */
1237 if (nva_start_addr + size < nva_start_addr ||
1238 nva_start_addr < vstart)
1241 return (nva_start_addr + size <= va->va_end);
1245 * Find the first free block(lowest start address) in the tree,
1246 * that will accomplish the request corresponding to passing
1247 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1248 * a search length is adjusted to account for worst case alignment
1251 static __always_inline struct vmap_area *
1252 find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1253 unsigned long align, unsigned long vstart, bool adjust_search_size)
1255 struct vmap_area *va;
1256 struct rb_node *node;
1257 unsigned long length;
1259 /* Start from the root. */
1260 node = root->rb_node;
1262 /* Adjust the search size for alignment overhead. */
1263 length = adjust_search_size ? size + align - 1 : size;
1266 va = rb_entry(node, struct vmap_area, rb_node);
1268 if (get_subtree_max_size(node->rb_left) >= length &&
1269 vstart < va->va_start) {
1270 node = node->rb_left;
1272 if (is_within_this_va(va, size, align, vstart))
1276 * Does not make sense to go deeper towards the right
1277 * sub-tree if it does not have a free block that is
1278 * equal or bigger to the requested search length.
1280 if (get_subtree_max_size(node->rb_right) >= length) {
1281 node = node->rb_right;
1286 * OK. We roll back and find the first right sub-tree,
1287 * that will satisfy the search criteria. It can happen
1288 * due to "vstart" restriction or an alignment overhead
1289 * that is bigger then PAGE_SIZE.
1291 while ((node = rb_parent(node))) {
1292 va = rb_entry(node, struct vmap_area, rb_node);
1293 if (is_within_this_va(va, size, align, vstart))
1296 if (get_subtree_max_size(node->rb_right) >= length &&
1297 vstart <= va->va_start) {
1299 * Shift the vstart forward. Please note, we update it with
1300 * parent's start address adding "1" because we do not want
1301 * to enter same sub-tree after it has already been checked
1302 * and no suitable free block found there.
1304 vstart = va->va_start + 1;
1305 node = node->rb_right;
1315 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1316 #include <linux/random.h>
1318 static struct vmap_area *
1319 find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1320 unsigned long align, unsigned long vstart)
1322 struct vmap_area *va;
1324 list_for_each_entry(va, head, list) {
1325 if (!is_within_this_va(va, size, align, vstart))
1335 find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1336 unsigned long size, unsigned long align)
1338 struct vmap_area *va_1, *va_2;
1339 unsigned long vstart;
1342 get_random_bytes(&rnd, sizeof(rnd));
1343 vstart = VMALLOC_START + rnd;
1345 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1346 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1349 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1350 va_1, va_2, vstart);
1356 FL_FIT_TYPE = 1, /* full fit */
1357 LE_FIT_TYPE = 2, /* left edge fit */
1358 RE_FIT_TYPE = 3, /* right edge fit */
1359 NE_FIT_TYPE = 4 /* no edge fit */
1362 static __always_inline enum fit_type
1363 classify_va_fit_type(struct vmap_area *va,
1364 unsigned long nva_start_addr, unsigned long size)
1368 /* Check if it is within VA. */
1369 if (nva_start_addr < va->va_start ||
1370 nva_start_addr + size > va->va_end)
1374 if (va->va_start == nva_start_addr) {
1375 if (va->va_end == nva_start_addr + size)
1379 } else if (va->va_end == nva_start_addr + size) {
1388 static __always_inline int
1389 adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1390 struct vmap_area *va, unsigned long nva_start_addr,
1393 struct vmap_area *lva = NULL;
1394 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1396 if (type == FL_FIT_TYPE) {
1398 * No need to split VA, it fully fits.
1404 unlink_va_augment(va, root);
1405 kmem_cache_free(vmap_area_cachep, va);
1406 } else if (type == LE_FIT_TYPE) {
1408 * Split left edge of fit VA.
1414 va->va_start += size;
1415 } else if (type == RE_FIT_TYPE) {
1417 * Split right edge of fit VA.
1423 va->va_end = nva_start_addr;
1424 } else if (type == NE_FIT_TYPE) {
1426 * Split no edge of fit VA.
1432 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1433 if (unlikely(!lva)) {
1435 * For percpu allocator we do not do any pre-allocation
1436 * and leave it as it is. The reason is it most likely
1437 * never ends up with NE_FIT_TYPE splitting. In case of
1438 * percpu allocations offsets and sizes are aligned to
1439 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1440 * are its main fitting cases.
1442 * There are a few exceptions though, as an example it is
1443 * a first allocation (early boot up) when we have "one"
1444 * big free space that has to be split.
1446 * Also we can hit this path in case of regular "vmap"
1447 * allocations, if "this" current CPU was not preloaded.
1448 * See the comment in alloc_vmap_area() why. If so, then
1449 * GFP_NOWAIT is used instead to get an extra object for
1450 * split purpose. That is rare and most time does not
1453 * What happens if an allocation gets failed. Basically,
1454 * an "overflow" path is triggered to purge lazily freed
1455 * areas to free some memory, then, the "retry" path is
1456 * triggered to repeat one more time. See more details
1457 * in alloc_vmap_area() function.
1459 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1465 * Build the remainder.
1467 lva->va_start = va->va_start;
1468 lva->va_end = nva_start_addr;
1471 * Shrink this VA to remaining size.
1473 va->va_start = nva_start_addr + size;
1478 if (type != FL_FIT_TYPE) {
1479 augment_tree_propagate_from(va);
1481 if (lva) /* type == NE_FIT_TYPE */
1482 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1489 * Returns a start address of the newly allocated area, if success.
1490 * Otherwise a vend is returned that indicates failure.
1492 static __always_inline unsigned long
1493 __alloc_vmap_area(struct rb_root *root, struct list_head *head,
1494 unsigned long size, unsigned long align,
1495 unsigned long vstart, unsigned long vend)
1497 bool adjust_search_size = true;
1498 unsigned long nva_start_addr;
1499 struct vmap_area *va;
1503 * Do not adjust when:
1504 * a) align <= PAGE_SIZE, because it does not make any sense.
1505 * All blocks(their start addresses) are at least PAGE_SIZE
1507 * b) a short range where a requested size corresponds to exactly
1508 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1509 * With adjusted search length an allocation would not succeed.
1511 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1512 adjust_search_size = false;
1514 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1518 if (va->va_start > vstart)
1519 nva_start_addr = ALIGN(va->va_start, align);
1521 nva_start_addr = ALIGN(vstart, align);
1523 /* Check the "vend" restriction. */
1524 if (nva_start_addr + size > vend)
1527 /* Update the free vmap_area. */
1528 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1529 if (WARN_ON_ONCE(ret))
1532 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1533 find_vmap_lowest_match_check(root, head, size, align);
1536 return nva_start_addr;
1540 * Free a region of KVA allocated by alloc_vmap_area
1542 static void free_vmap_area(struct vmap_area *va)
1545 * Remove from the busy tree/list.
1547 spin_lock(&vmap_area_lock);
1548 unlink_va(va, &vmap_area_root);
1549 spin_unlock(&vmap_area_lock);
1552 * Insert/Merge it back to the free tree/list.
1554 spin_lock(&free_vmap_area_lock);
1555 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1556 spin_unlock(&free_vmap_area_lock);
1560 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1562 struct vmap_area *va = NULL;
1565 * Preload this CPU with one extra vmap_area object. It is used
1566 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1567 * a CPU that does an allocation is preloaded.
1569 * We do it in non-atomic context, thus it allows us to use more
1570 * permissive allocation masks to be more stable under low memory
1571 * condition and high memory pressure.
1573 if (!this_cpu_read(ne_fit_preload_node))
1574 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1578 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1579 kmem_cache_free(vmap_area_cachep, va);
1583 * Allocate a region of KVA of the specified size and alignment, within the
1586 static struct vmap_area *alloc_vmap_area(unsigned long size,
1587 unsigned long align,
1588 unsigned long vstart, unsigned long vend,
1589 int node, gfp_t gfp_mask)
1591 struct vmap_area *va;
1592 unsigned long freed;
1598 BUG_ON(offset_in_page(size));
1599 BUG_ON(!is_power_of_2(align));
1601 if (unlikely(!vmap_initialized))
1602 return ERR_PTR(-EBUSY);
1605 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1607 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1609 return ERR_PTR(-ENOMEM);
1612 * Only scan the relevant parts containing pointers to other objects
1613 * to avoid false negatives.
1615 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1618 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1619 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1620 size, align, vstart, vend);
1621 spin_unlock(&free_vmap_area_lock);
1624 * If an allocation fails, the "vend" address is
1625 * returned. Therefore trigger the overflow path.
1627 if (unlikely(addr == vend))
1630 va->va_start = addr;
1631 va->va_end = addr + size;
1634 spin_lock(&vmap_area_lock);
1635 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1636 spin_unlock(&vmap_area_lock);
1638 BUG_ON(!IS_ALIGNED(va->va_start, align));
1639 BUG_ON(va->va_start < vstart);
1640 BUG_ON(va->va_end > vend);
1642 ret = kasan_populate_vmalloc(addr, size);
1645 return ERR_PTR(ret);
1652 purge_vmap_area_lazy();
1658 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1665 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1666 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1669 kmem_cache_free(vmap_area_cachep, va);
1670 return ERR_PTR(-EBUSY);
1673 int register_vmap_purge_notifier(struct notifier_block *nb)
1675 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1677 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1679 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1681 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1683 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1686 * lazy_max_pages is the maximum amount of virtual address space we gather up
1687 * before attempting to purge with a TLB flush.
1689 * There is a tradeoff here: a larger number will cover more kernel page tables
1690 * and take slightly longer to purge, but it will linearly reduce the number of
1691 * global TLB flushes that must be performed. It would seem natural to scale
1692 * this number up linearly with the number of CPUs (because vmapping activity
1693 * could also scale linearly with the number of CPUs), however it is likely
1694 * that in practice, workloads might be constrained in other ways that mean
1695 * vmap activity will not scale linearly with CPUs. Also, I want to be
1696 * conservative and not introduce a big latency on huge systems, so go with
1697 * a less aggressive log scale. It will still be an improvement over the old
1698 * code, and it will be simple to change the scale factor if we find that it
1699 * becomes a problem on bigger systems.
1701 static unsigned long lazy_max_pages(void)
1705 log = fls(num_online_cpus());
1707 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1710 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1713 * Serialize vmap purging. There is no actual critical section protected
1714 * by this lock, but we want to avoid concurrent calls for performance
1715 * reasons and to make the pcpu_get_vm_areas more deterministic.
1717 static DEFINE_MUTEX(vmap_purge_lock);
1719 /* for per-CPU blocks */
1720 static void purge_fragmented_blocks_allcpus(void);
1723 * Purges all lazily-freed vmap areas.
1725 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1727 unsigned long resched_threshold;
1728 struct list_head local_purge_list;
1729 struct vmap_area *va, *n_va;
1731 lockdep_assert_held(&vmap_purge_lock);
1733 spin_lock(&purge_vmap_area_lock);
1734 purge_vmap_area_root = RB_ROOT;
1735 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1736 spin_unlock(&purge_vmap_area_lock);
1738 if (unlikely(list_empty(&local_purge_list)))
1742 list_first_entry(&local_purge_list,
1743 struct vmap_area, list)->va_start);
1746 list_last_entry(&local_purge_list,
1747 struct vmap_area, list)->va_end);
1749 flush_tlb_kernel_range(start, end);
1750 resched_threshold = lazy_max_pages() << 1;
1752 spin_lock(&free_vmap_area_lock);
1753 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1754 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1755 unsigned long orig_start = va->va_start;
1756 unsigned long orig_end = va->va_end;
1759 * Finally insert or merge lazily-freed area. It is
1760 * detached and there is no need to "unlink" it from
1763 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1764 &free_vmap_area_list);
1769 if (is_vmalloc_or_module_addr((void *)orig_start))
1770 kasan_release_vmalloc(orig_start, orig_end,
1771 va->va_start, va->va_end);
1773 atomic_long_sub(nr, &vmap_lazy_nr);
1775 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1776 cond_resched_lock(&free_vmap_area_lock);
1778 spin_unlock(&free_vmap_area_lock);
1783 * Kick off a purge of the outstanding lazy areas.
1785 static void purge_vmap_area_lazy(void)
1787 mutex_lock(&vmap_purge_lock);
1788 purge_fragmented_blocks_allcpus();
1789 __purge_vmap_area_lazy(ULONG_MAX, 0);
1790 mutex_unlock(&vmap_purge_lock);
1793 static void drain_vmap_area_work(struct work_struct *work)
1795 unsigned long nr_lazy;
1798 mutex_lock(&vmap_purge_lock);
1799 __purge_vmap_area_lazy(ULONG_MAX, 0);
1800 mutex_unlock(&vmap_purge_lock);
1802 /* Recheck if further work is required. */
1803 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1804 } while (nr_lazy > lazy_max_pages());
1808 * Free a vmap area, caller ensuring that the area has been unmapped
1809 * and flush_cache_vunmap had been called for the correct range
1812 static void free_vmap_area_noflush(struct vmap_area *va)
1814 unsigned long nr_lazy;
1816 spin_lock(&vmap_area_lock);
1817 unlink_va(va, &vmap_area_root);
1818 spin_unlock(&vmap_area_lock);
1820 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1821 PAGE_SHIFT, &vmap_lazy_nr);
1824 * Merge or place it to the purge tree/list.
1826 spin_lock(&purge_vmap_area_lock);
1827 merge_or_add_vmap_area(va,
1828 &purge_vmap_area_root, &purge_vmap_area_list);
1829 spin_unlock(&purge_vmap_area_lock);
1831 /* After this point, we may free va at any time */
1832 if (unlikely(nr_lazy > lazy_max_pages()))
1833 schedule_work(&drain_vmap_work);
1837 * Free and unmap a vmap area
1839 static void free_unmap_vmap_area(struct vmap_area *va)
1841 flush_cache_vunmap(va->va_start, va->va_end);
1842 vunmap_range_noflush(va->va_start, va->va_end);
1843 if (debug_pagealloc_enabled_static())
1844 flush_tlb_kernel_range(va->va_start, va->va_end);
1846 free_vmap_area_noflush(va);
1849 struct vmap_area *find_vmap_area(unsigned long addr)
1851 struct vmap_area *va;
1853 spin_lock(&vmap_area_lock);
1854 va = __find_vmap_area(addr, &vmap_area_root);
1855 spin_unlock(&vmap_area_lock);
1860 /*** Per cpu kva allocator ***/
1863 * vmap space is limited especially on 32 bit architectures. Ensure there is
1864 * room for at least 16 percpu vmap blocks per CPU.
1867 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1868 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1869 * instead (we just need a rough idea)
1871 #if BITS_PER_LONG == 32
1872 #define VMALLOC_SPACE (128UL*1024*1024)
1874 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1877 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1878 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1879 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1880 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1881 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1882 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1883 #define VMAP_BBMAP_BITS \
1884 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1885 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1886 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1888 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1890 struct vmap_block_queue {
1892 struct list_head free;
1897 struct vmap_area *va;
1898 unsigned long free, dirty;
1899 unsigned long dirty_min, dirty_max; /*< dirty range */
1900 struct list_head free_list;
1901 struct rcu_head rcu_head;
1902 struct list_head purge;
1905 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1906 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1909 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1910 * in the free path. Could get rid of this if we change the API to return a
1911 * "cookie" from alloc, to be passed to free. But no big deal yet.
1913 static DEFINE_XARRAY(vmap_blocks);
1916 * We should probably have a fallback mechanism to allocate virtual memory
1917 * out of partially filled vmap blocks. However vmap block sizing should be
1918 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1922 static unsigned long addr_to_vb_idx(unsigned long addr)
1924 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1925 addr /= VMAP_BLOCK_SIZE;
1929 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1933 addr = va_start + (pages_off << PAGE_SHIFT);
1934 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1935 return (void *)addr;
1939 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1940 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1941 * @order: how many 2^order pages should be occupied in newly allocated block
1942 * @gfp_mask: flags for the page level allocator
1944 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1946 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1948 struct vmap_block_queue *vbq;
1949 struct vmap_block *vb;
1950 struct vmap_area *va;
1951 unsigned long vb_idx;
1955 node = numa_node_id();
1957 vb = kmalloc_node(sizeof(struct vmap_block),
1958 gfp_mask & GFP_RECLAIM_MASK, node);
1960 return ERR_PTR(-ENOMEM);
1962 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1963 VMALLOC_START, VMALLOC_END,
1967 return ERR_CAST(va);
1970 vaddr = vmap_block_vaddr(va->va_start, 0);
1971 spin_lock_init(&vb->lock);
1973 /* At least something should be left free */
1974 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1975 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1977 vb->dirty_min = VMAP_BBMAP_BITS;
1979 INIT_LIST_HEAD(&vb->free_list);
1981 vb_idx = addr_to_vb_idx(va->va_start);
1982 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1986 return ERR_PTR(err);
1989 vbq = raw_cpu_ptr(&vmap_block_queue);
1990 spin_lock(&vbq->lock);
1991 list_add_tail_rcu(&vb->free_list, &vbq->free);
1992 spin_unlock(&vbq->lock);
1997 static void free_vmap_block(struct vmap_block *vb)
1999 struct vmap_block *tmp;
2001 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2004 free_vmap_area_noflush(vb->va);
2005 kfree_rcu(vb, rcu_head);
2008 static void purge_fragmented_blocks(int cpu)
2011 struct vmap_block *vb;
2012 struct vmap_block *n_vb;
2013 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2016 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2018 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2021 spin_lock(&vb->lock);
2022 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2023 vb->free = 0; /* prevent further allocs after releasing lock */
2024 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2026 vb->dirty_max = VMAP_BBMAP_BITS;
2027 spin_lock(&vbq->lock);
2028 list_del_rcu(&vb->free_list);
2029 spin_unlock(&vbq->lock);
2030 spin_unlock(&vb->lock);
2031 list_add_tail(&vb->purge, &purge);
2033 spin_unlock(&vb->lock);
2037 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2038 list_del(&vb->purge);
2039 free_vmap_block(vb);
2043 static void purge_fragmented_blocks_allcpus(void)
2047 for_each_possible_cpu(cpu)
2048 purge_fragmented_blocks(cpu);
2051 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2053 struct vmap_block_queue *vbq;
2054 struct vmap_block *vb;
2058 BUG_ON(offset_in_page(size));
2059 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2060 if (WARN_ON(size == 0)) {
2062 * Allocating 0 bytes isn't what caller wants since
2063 * get_order(0) returns funny result. Just warn and terminate
2068 order = get_order(size);
2071 vbq = raw_cpu_ptr(&vmap_block_queue);
2072 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2073 unsigned long pages_off;
2075 spin_lock(&vb->lock);
2076 if (vb->free < (1UL << order)) {
2077 spin_unlock(&vb->lock);
2081 pages_off = VMAP_BBMAP_BITS - vb->free;
2082 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2083 vb->free -= 1UL << order;
2084 if (vb->free == 0) {
2085 spin_lock(&vbq->lock);
2086 list_del_rcu(&vb->free_list);
2087 spin_unlock(&vbq->lock);
2090 spin_unlock(&vb->lock);
2096 /* Allocate new block if nothing was found */
2098 vaddr = new_vmap_block(order, gfp_mask);
2103 static void vb_free(unsigned long addr, unsigned long size)
2105 unsigned long offset;
2107 struct vmap_block *vb;
2109 BUG_ON(offset_in_page(size));
2110 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2112 flush_cache_vunmap(addr, addr + size);
2114 order = get_order(size);
2115 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2116 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2118 vunmap_range_noflush(addr, addr + size);
2120 if (debug_pagealloc_enabled_static())
2121 flush_tlb_kernel_range(addr, addr + size);
2123 spin_lock(&vb->lock);
2125 /* Expand dirty range */
2126 vb->dirty_min = min(vb->dirty_min, offset);
2127 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2129 vb->dirty += 1UL << order;
2130 if (vb->dirty == VMAP_BBMAP_BITS) {
2132 spin_unlock(&vb->lock);
2133 free_vmap_block(vb);
2135 spin_unlock(&vb->lock);
2138 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2142 if (unlikely(!vmap_initialized))
2147 for_each_possible_cpu(cpu) {
2148 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2149 struct vmap_block *vb;
2152 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2153 spin_lock(&vb->lock);
2154 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2155 unsigned long va_start = vb->va->va_start;
2158 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2159 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2161 start = min(s, start);
2166 spin_unlock(&vb->lock);
2171 mutex_lock(&vmap_purge_lock);
2172 purge_fragmented_blocks_allcpus();
2173 if (!__purge_vmap_area_lazy(start, end) && flush)
2174 flush_tlb_kernel_range(start, end);
2175 mutex_unlock(&vmap_purge_lock);
2179 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2181 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2182 * to amortize TLB flushing overheads. What this means is that any page you
2183 * have now, may, in a former life, have been mapped into kernel virtual
2184 * address by the vmap layer and so there might be some CPUs with TLB entries
2185 * still referencing that page (additional to the regular 1:1 kernel mapping).
2187 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2188 * be sure that none of the pages we have control over will have any aliases
2189 * from the vmap layer.
2191 void vm_unmap_aliases(void)
2193 unsigned long start = ULONG_MAX, end = 0;
2196 _vm_unmap_aliases(start, end, flush);
2198 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2201 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2202 * @mem: the pointer returned by vm_map_ram
2203 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2205 void vm_unmap_ram(const void *mem, unsigned int count)
2207 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2208 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2209 struct vmap_area *va;
2213 BUG_ON(addr < VMALLOC_START);
2214 BUG_ON(addr > VMALLOC_END);
2215 BUG_ON(!PAGE_ALIGNED(addr));
2217 kasan_poison_vmalloc(mem, size);
2219 if (likely(count <= VMAP_MAX_ALLOC)) {
2220 debug_check_no_locks_freed(mem, size);
2221 vb_free(addr, size);
2225 va = find_vmap_area(addr);
2227 debug_check_no_locks_freed((void *)va->va_start,
2228 (va->va_end - va->va_start));
2229 free_unmap_vmap_area(va);
2231 EXPORT_SYMBOL(vm_unmap_ram);
2234 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2235 * @pages: an array of pointers to the pages to be mapped
2236 * @count: number of pages
2237 * @node: prefer to allocate data structures on this node
2239 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2240 * faster than vmap so it's good. But if you mix long-life and short-life
2241 * objects with vm_map_ram(), it could consume lots of address space through
2242 * fragmentation (especially on a 32bit machine). You could see failures in
2243 * the end. Please use this function for short-lived objects.
2245 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2247 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2249 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2253 if (likely(count <= VMAP_MAX_ALLOC)) {
2254 mem = vb_alloc(size, GFP_KERNEL);
2257 addr = (unsigned long)mem;
2259 struct vmap_area *va;
2260 va = alloc_vmap_area(size, PAGE_SIZE,
2261 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2265 addr = va->va_start;
2269 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2270 pages, PAGE_SHIFT) < 0) {
2271 vm_unmap_ram(mem, count);
2276 * Mark the pages as accessible, now that they are mapped.
2277 * With hardware tag-based KASAN, marking is skipped for
2278 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2280 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2284 EXPORT_SYMBOL(vm_map_ram);
2286 static struct vm_struct *vmlist __initdata;
2288 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2290 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2291 return vm->page_order;
2297 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2299 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2300 vm->page_order = order;
2307 * vm_area_add_early - add vmap area early during boot
2308 * @vm: vm_struct to add
2310 * This function is used to add fixed kernel vm area to vmlist before
2311 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2312 * should contain proper values and the other fields should be zero.
2314 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2316 void __init vm_area_add_early(struct vm_struct *vm)
2318 struct vm_struct *tmp, **p;
2320 BUG_ON(vmap_initialized);
2321 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2322 if (tmp->addr >= vm->addr) {
2323 BUG_ON(tmp->addr < vm->addr + vm->size);
2326 BUG_ON(tmp->addr + tmp->size > vm->addr);
2333 * vm_area_register_early - register vmap area early during boot
2334 * @vm: vm_struct to register
2335 * @align: requested alignment
2337 * This function is used to register kernel vm area before
2338 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2339 * proper values on entry and other fields should be zero. On return,
2340 * vm->addr contains the allocated address.
2342 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2344 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2346 unsigned long addr = ALIGN(VMALLOC_START, align);
2347 struct vm_struct *cur, **p;
2349 BUG_ON(vmap_initialized);
2351 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2352 if ((unsigned long)cur->addr - addr >= vm->size)
2354 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2357 BUG_ON(addr > VMALLOC_END - vm->size);
2358 vm->addr = (void *)addr;
2361 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2364 static void vmap_init_free_space(void)
2366 unsigned long vmap_start = 1;
2367 const unsigned long vmap_end = ULONG_MAX;
2368 struct vmap_area *busy, *free;
2372 * -|-----|.....|-----|-----|-----|.....|-
2374 * |<--------------------------------->|
2376 list_for_each_entry(busy, &vmap_area_list, list) {
2377 if (busy->va_start - vmap_start > 0) {
2378 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2379 if (!WARN_ON_ONCE(!free)) {
2380 free->va_start = vmap_start;
2381 free->va_end = busy->va_start;
2383 insert_vmap_area_augment(free, NULL,
2384 &free_vmap_area_root,
2385 &free_vmap_area_list);
2389 vmap_start = busy->va_end;
2392 if (vmap_end - vmap_start > 0) {
2393 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2394 if (!WARN_ON_ONCE(!free)) {
2395 free->va_start = vmap_start;
2396 free->va_end = vmap_end;
2398 insert_vmap_area_augment(free, NULL,
2399 &free_vmap_area_root,
2400 &free_vmap_area_list);
2405 void __init vmalloc_init(void)
2407 struct vmap_area *va;
2408 struct vm_struct *tmp;
2412 * Create the cache for vmap_area objects.
2414 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2416 for_each_possible_cpu(i) {
2417 struct vmap_block_queue *vbq;
2418 struct vfree_deferred *p;
2420 vbq = &per_cpu(vmap_block_queue, i);
2421 spin_lock_init(&vbq->lock);
2422 INIT_LIST_HEAD(&vbq->free);
2423 p = &per_cpu(vfree_deferred, i);
2424 init_llist_head(&p->list);
2425 INIT_WORK(&p->wq, free_work);
2428 /* Import existing vmlist entries. */
2429 for (tmp = vmlist; tmp; tmp = tmp->next) {
2430 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2431 if (WARN_ON_ONCE(!va))
2434 va->va_start = (unsigned long)tmp->addr;
2435 va->va_end = va->va_start + tmp->size;
2437 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2441 * Now we can initialize a free vmap space.
2443 vmap_init_free_space();
2444 vmap_initialized = true;
2447 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2448 struct vmap_area *va, unsigned long flags, const void *caller)
2451 vm->addr = (void *)va->va_start;
2452 vm->size = va->va_end - va->va_start;
2453 vm->caller = caller;
2457 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2458 unsigned long flags, const void *caller)
2460 spin_lock(&vmap_area_lock);
2461 setup_vmalloc_vm_locked(vm, va, flags, caller);
2462 spin_unlock(&vmap_area_lock);
2465 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2468 * Before removing VM_UNINITIALIZED,
2469 * we should make sure that vm has proper values.
2470 * Pair with smp_rmb() in show_numa_info().
2473 vm->flags &= ~VM_UNINITIALIZED;
2476 static struct vm_struct *__get_vm_area_node(unsigned long size,
2477 unsigned long align, unsigned long shift, unsigned long flags,
2478 unsigned long start, unsigned long end, int node,
2479 gfp_t gfp_mask, const void *caller)
2481 struct vmap_area *va;
2482 struct vm_struct *area;
2483 unsigned long requested_size = size;
2485 BUG_ON(in_interrupt());
2486 size = ALIGN(size, 1ul << shift);
2487 if (unlikely(!size))
2490 if (flags & VM_IOREMAP)
2491 align = 1ul << clamp_t(int, get_count_order_long(size),
2492 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2494 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2495 if (unlikely(!area))
2498 if (!(flags & VM_NO_GUARD))
2501 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2507 setup_vmalloc_vm(area, va, flags, caller);
2510 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2511 * best-effort approach, as they can be mapped outside of vmalloc code.
2512 * For VM_ALLOC mappings, the pages are marked as accessible after
2513 * getting mapped in __vmalloc_node_range().
2514 * With hardware tag-based KASAN, marking is skipped for
2515 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2517 if (!(flags & VM_ALLOC))
2518 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2519 KASAN_VMALLOC_PROT_NORMAL);
2524 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2525 unsigned long start, unsigned long end,
2528 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2529 NUMA_NO_NODE, GFP_KERNEL, caller);
2533 * get_vm_area - reserve a contiguous kernel virtual area
2534 * @size: size of the area
2535 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2537 * Search an area of @size in the kernel virtual mapping area,
2538 * and reserved it for out purposes. Returns the area descriptor
2539 * on success or %NULL on failure.
2541 * Return: the area descriptor on success or %NULL on failure.
2543 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2545 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2546 VMALLOC_START, VMALLOC_END,
2547 NUMA_NO_NODE, GFP_KERNEL,
2548 __builtin_return_address(0));
2551 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2554 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2555 VMALLOC_START, VMALLOC_END,
2556 NUMA_NO_NODE, GFP_KERNEL, caller);
2560 * find_vm_area - find a continuous kernel virtual area
2561 * @addr: base address
2563 * Search for the kernel VM area starting at @addr, and return it.
2564 * It is up to the caller to do all required locking to keep the returned
2567 * Return: the area descriptor on success or %NULL on failure.
2569 struct vm_struct *find_vm_area(const void *addr)
2571 struct vmap_area *va;
2573 va = find_vmap_area((unsigned long)addr);
2581 * remove_vm_area - find and remove a continuous kernel virtual area
2582 * @addr: base address
2584 * Search for the kernel VM area starting at @addr, and remove it.
2585 * This function returns the found VM area, but using it is NOT safe
2586 * on SMP machines, except for its size or flags.
2588 * Return: the area descriptor on success or %NULL on failure.
2590 struct vm_struct *remove_vm_area(const void *addr)
2592 struct vmap_area *va;
2596 spin_lock(&vmap_area_lock);
2597 va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2599 struct vm_struct *vm = va->vm;
2602 spin_unlock(&vmap_area_lock);
2604 kasan_free_module_shadow(vm);
2605 free_unmap_vmap_area(va);
2610 spin_unlock(&vmap_area_lock);
2614 static inline void set_area_direct_map(const struct vm_struct *area,
2615 int (*set_direct_map)(struct page *page))
2619 /* HUGE_VMALLOC passes small pages to set_direct_map */
2620 for (i = 0; i < area->nr_pages; i++)
2621 if (page_address(area->pages[i]))
2622 set_direct_map(area->pages[i]);
2625 /* Handle removing and resetting vm mappings related to the vm_struct. */
2626 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2628 unsigned long start = ULONG_MAX, end = 0;
2629 unsigned int page_order = vm_area_page_order(area);
2630 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2634 remove_vm_area(area->addr);
2636 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2641 * If not deallocating pages, just do the flush of the VM area and
2644 if (!deallocate_pages) {
2650 * If execution gets here, flush the vm mapping and reset the direct
2651 * map. Find the start and end range of the direct mappings to make sure
2652 * the vm_unmap_aliases() flush includes the direct map.
2654 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2655 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2657 unsigned long page_size;
2659 page_size = PAGE_SIZE << page_order;
2660 start = min(addr, start);
2661 end = max(addr + page_size, end);
2667 * Set direct map to something invalid so that it won't be cached if
2668 * there are any accesses after the TLB flush, then flush the TLB and
2669 * reset the direct map permissions to the default.
2671 set_area_direct_map(area, set_direct_map_invalid_noflush);
2672 _vm_unmap_aliases(start, end, flush_dmap);
2673 set_area_direct_map(area, set_direct_map_default_noflush);
2676 static void __vunmap(const void *addr, int deallocate_pages)
2678 struct vm_struct *area;
2683 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2687 area = find_vm_area(addr);
2688 if (unlikely(!area)) {
2689 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2694 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2695 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2697 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2699 vm_remove_mappings(area, deallocate_pages);
2701 if (deallocate_pages) {
2704 for (i = 0; i < area->nr_pages; i++) {
2705 struct page *page = area->pages[i];
2708 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2710 * High-order allocs for huge vmallocs are split, so
2711 * can be freed as an array of order-0 allocations
2713 __free_pages(page, 0);
2716 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2718 kvfree(area->pages);
2724 static inline void __vfree_deferred(const void *addr)
2727 * Use raw_cpu_ptr() because this can be called from preemptible
2728 * context. Preemption is absolutely fine here, because the llist_add()
2729 * implementation is lockless, so it works even if we are adding to
2730 * another cpu's list. schedule_work() should be fine with this too.
2732 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2734 if (llist_add((struct llist_node *)addr, &p->list))
2735 schedule_work(&p->wq);
2739 * vfree_atomic - release memory allocated by vmalloc()
2740 * @addr: memory base address
2742 * This one is just like vfree() but can be called in any atomic context
2745 void vfree_atomic(const void *addr)
2749 kmemleak_free(addr);
2753 __vfree_deferred(addr);
2756 static void __vfree(const void *addr)
2758 if (unlikely(in_interrupt()))
2759 __vfree_deferred(addr);
2765 * vfree - Release memory allocated by vmalloc()
2766 * @addr: Memory base address
2768 * Free the virtually continuous memory area starting at @addr, as obtained
2769 * from one of the vmalloc() family of APIs. This will usually also free the
2770 * physical memory underlying the virtual allocation, but that memory is
2771 * reference counted, so it will not be freed until the last user goes away.
2773 * If @addr is NULL, no operation is performed.
2776 * May sleep if called *not* from interrupt context.
2777 * Must not be called in NMI context (strictly speaking, it could be
2778 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2779 * conventions for vfree() arch-dependent would be a really bad idea).
2781 void vfree(const void *addr)
2785 kmemleak_free(addr);
2787 might_sleep_if(!in_interrupt());
2794 EXPORT_SYMBOL(vfree);
2797 * vunmap - release virtual mapping obtained by vmap()
2798 * @addr: memory base address
2800 * Free the virtually contiguous memory area starting at @addr,
2801 * which was created from the page array passed to vmap().
2803 * Must not be called in interrupt context.
2805 void vunmap(const void *addr)
2807 BUG_ON(in_interrupt());
2812 EXPORT_SYMBOL(vunmap);
2815 * vmap - map an array of pages into virtually contiguous space
2816 * @pages: array of page pointers
2817 * @count: number of pages to map
2818 * @flags: vm_area->flags
2819 * @prot: page protection for the mapping
2821 * Maps @count pages from @pages into contiguous kernel virtual space.
2822 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2823 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2824 * are transferred from the caller to vmap(), and will be freed / dropped when
2825 * vfree() is called on the return value.
2827 * Return: the address of the area or %NULL on failure
2829 void *vmap(struct page **pages, unsigned int count,
2830 unsigned long flags, pgprot_t prot)
2832 struct vm_struct *area;
2834 unsigned long size; /* In bytes */
2839 * Your top guard is someone else's bottom guard. Not having a top
2840 * guard compromises someone else's mappings too.
2842 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2843 flags &= ~VM_NO_GUARD;
2845 if (count > totalram_pages())
2848 size = (unsigned long)count << PAGE_SHIFT;
2849 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2853 addr = (unsigned long)area->addr;
2854 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2855 pages, PAGE_SHIFT) < 0) {
2860 if (flags & VM_MAP_PUT_PAGES) {
2861 area->pages = pages;
2862 area->nr_pages = count;
2866 EXPORT_SYMBOL(vmap);
2868 #ifdef CONFIG_VMAP_PFN
2869 struct vmap_pfn_data {
2870 unsigned long *pfns;
2875 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2877 struct vmap_pfn_data *data = private;
2879 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2881 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2886 * vmap_pfn - map an array of PFNs into virtually contiguous space
2887 * @pfns: array of PFNs
2888 * @count: number of pages to map
2889 * @prot: page protection for the mapping
2891 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2892 * the start address of the mapping.
2894 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2896 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2897 struct vm_struct *area;
2899 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2900 __builtin_return_address(0));
2903 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2904 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2910 EXPORT_SYMBOL_GPL(vmap_pfn);
2911 #endif /* CONFIG_VMAP_PFN */
2913 static inline unsigned int
2914 vm_area_alloc_pages(gfp_t gfp, int nid,
2915 unsigned int order, unsigned int nr_pages, struct page **pages)
2917 unsigned int nr_allocated = 0;
2922 * For order-0 pages we make use of bulk allocator, if
2923 * the page array is partly or not at all populated due
2924 * to fails, fallback to a single page allocator that is
2928 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2930 while (nr_allocated < nr_pages) {
2931 unsigned int nr, nr_pages_request;
2934 * A maximum allowed request is hard-coded and is 100
2935 * pages per call. That is done in order to prevent a
2936 * long preemption off scenario in the bulk-allocator
2937 * so the range is [1:100].
2939 nr_pages_request = min(100U, nr_pages - nr_allocated);
2941 /* memory allocation should consider mempolicy, we can't
2942 * wrongly use nearest node when nid == NUMA_NO_NODE,
2943 * otherwise memory may be allocated in only one node,
2944 * but mempolicy wants to alloc memory by interleaving.
2946 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2947 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2949 pages + nr_allocated);
2952 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2954 pages + nr_allocated);
2960 * If zero or pages were obtained partly,
2961 * fallback to a single page allocator.
2963 if (nr != nr_pages_request)
2968 /* High-order pages or fallback path if "bulk" fails. */
2970 while (nr_allocated < nr_pages) {
2971 if (fatal_signal_pending(current))
2974 if (nid == NUMA_NO_NODE)
2975 page = alloc_pages(gfp, order);
2977 page = alloc_pages_node(nid, gfp, order);
2978 if (unlikely(!page))
2981 * Higher order allocations must be able to be treated as
2982 * indepdenent small pages by callers (as they can with
2983 * small-page vmallocs). Some drivers do their own refcounting
2984 * on vmalloc_to_page() pages, some use page->mapping,
2988 split_page(page, order);
2991 * Careful, we allocate and map page-order pages, but
2992 * tracking is done per PAGE_SIZE page so as to keep the
2993 * vm_struct APIs independent of the physical/mapped size.
2995 for (i = 0; i < (1U << order); i++)
2996 pages[nr_allocated + i] = page + i;
2999 nr_allocated += 1U << order;
3002 return nr_allocated;
3005 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3006 pgprot_t prot, unsigned int page_shift,
3009 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3010 bool nofail = gfp_mask & __GFP_NOFAIL;
3011 unsigned long addr = (unsigned long)area->addr;
3012 unsigned long size = get_vm_area_size(area);
3013 unsigned long array_size;
3014 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3015 unsigned int page_order;
3019 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3020 gfp_mask |= __GFP_NOWARN;
3021 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3022 gfp_mask |= __GFP_HIGHMEM;
3024 /* Please note that the recursion is strictly bounded. */
3025 if (array_size > PAGE_SIZE) {
3026 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3029 area->pages = kmalloc_node(array_size, nested_gfp, node);
3033 warn_alloc(gfp_mask, NULL,
3034 "vmalloc error: size %lu, failed to allocated page array size %lu",
3035 nr_small_pages * PAGE_SIZE, array_size);
3040 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3041 page_order = vm_area_page_order(area);
3043 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3044 node, page_order, nr_small_pages, area->pages);
3046 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3047 if (gfp_mask & __GFP_ACCOUNT) {
3050 for (i = 0; i < area->nr_pages; i++)
3051 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3055 * If not enough pages were obtained to accomplish an
3056 * allocation request, free them via __vfree() if any.
3058 if (area->nr_pages != nr_small_pages) {
3059 warn_alloc(gfp_mask, NULL,
3060 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3061 area->nr_pages * PAGE_SIZE, page_order);
3066 * page tables allocations ignore external gfp mask, enforce it
3069 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3070 flags = memalloc_nofs_save();
3071 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3072 flags = memalloc_noio_save();
3075 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3077 if (nofail && (ret < 0))
3078 schedule_timeout_uninterruptible(1);
3079 } while (nofail && (ret < 0));
3081 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3082 memalloc_nofs_restore(flags);
3083 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3084 memalloc_noio_restore(flags);
3087 warn_alloc(gfp_mask, NULL,
3088 "vmalloc error: size %lu, failed to map pages",
3089 area->nr_pages * PAGE_SIZE);
3096 __vfree(area->addr);
3101 * __vmalloc_node_range - allocate virtually contiguous memory
3102 * @size: allocation size
3103 * @align: desired alignment
3104 * @start: vm area range start
3105 * @end: vm area range end
3106 * @gfp_mask: flags for the page level allocator
3107 * @prot: protection mask for the allocated pages
3108 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3109 * @node: node to use for allocation or NUMA_NO_NODE
3110 * @caller: caller's return address
3112 * Allocate enough pages to cover @size from the page level
3113 * allocator with @gfp_mask flags. Please note that the full set of gfp
3114 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3116 * Zone modifiers are not supported. From the reclaim modifiers
3117 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3118 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3119 * __GFP_RETRY_MAYFAIL are not supported).
3121 * __GFP_NOWARN can be used to suppress failures messages.
3123 * Map them into contiguous kernel virtual space, using a pagetable
3124 * protection of @prot.
3126 * Return: the address of the area or %NULL on failure
3128 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3129 unsigned long start, unsigned long end, gfp_t gfp_mask,
3130 pgprot_t prot, unsigned long vm_flags, int node,
3133 struct vm_struct *area;
3135 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3136 unsigned long real_size = size;
3137 unsigned long real_align = align;
3138 unsigned int shift = PAGE_SHIFT;
3140 if (WARN_ON_ONCE(!size))
3143 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3144 warn_alloc(gfp_mask, NULL,
3145 "vmalloc error: size %lu, exceeds total pages",
3150 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3151 unsigned long size_per_node;
3154 * Try huge pages. Only try for PAGE_KERNEL allocations,
3155 * others like modules don't yet expect huge pages in
3156 * their allocations due to apply_to_page_range not
3160 size_per_node = size;
3161 if (node == NUMA_NO_NODE)
3162 size_per_node /= num_online_nodes();
3163 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3166 shift = arch_vmap_pte_supported_shift(size_per_node);
3168 align = max(real_align, 1UL << shift);
3169 size = ALIGN(real_size, 1UL << shift);
3173 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3174 VM_UNINITIALIZED | vm_flags, start, end, node,
3177 bool nofail = gfp_mask & __GFP_NOFAIL;
3178 warn_alloc(gfp_mask, NULL,
3179 "vmalloc error: size %lu, vm_struct allocation failed%s",
3180 real_size, (nofail) ? ". Retrying." : "");
3182 schedule_timeout_uninterruptible(1);
3189 * Prepare arguments for __vmalloc_area_node() and
3190 * kasan_unpoison_vmalloc().
3192 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3193 if (kasan_hw_tags_enabled()) {
3195 * Modify protection bits to allow tagging.
3196 * This must be done before mapping.
3198 prot = arch_vmap_pgprot_tagged(prot);
3201 * Skip page_alloc poisoning and zeroing for physical
3202 * pages backing VM_ALLOC mapping. Memory is instead
3203 * poisoned and zeroed by kasan_unpoison_vmalloc().
3205 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3208 /* Take note that the mapping is PAGE_KERNEL. */
3209 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3212 /* Allocate physical pages and map them into vmalloc space. */
3213 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3218 * Mark the pages as accessible, now that they are mapped.
3219 * The condition for setting KASAN_VMALLOC_INIT should complement the
3220 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3221 * to make sure that memory is initialized under the same conditions.
3222 * Tag-based KASAN modes only assign tags to normal non-executable
3223 * allocations, see __kasan_unpoison_vmalloc().
3225 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3226 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3227 (gfp_mask & __GFP_SKIP_ZERO))
3228 kasan_flags |= KASAN_VMALLOC_INIT;
3229 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3230 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3233 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3234 * flag. It means that vm_struct is not fully initialized.
3235 * Now, it is fully initialized, so remove this flag here.
3237 clear_vm_uninitialized_flag(area);
3239 size = PAGE_ALIGN(size);
3240 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3241 kmemleak_vmalloc(area, size, gfp_mask);
3246 if (shift > PAGE_SHIFT) {
3257 * __vmalloc_node - allocate virtually contiguous memory
3258 * @size: allocation size
3259 * @align: desired alignment
3260 * @gfp_mask: flags for the page level allocator
3261 * @node: node to use for allocation or NUMA_NO_NODE
3262 * @caller: caller's return address
3264 * Allocate enough pages to cover @size from the page level allocator with
3265 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3267 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3268 * and __GFP_NOFAIL are not supported
3270 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3273 * Return: pointer to the allocated memory or %NULL on error
3275 void *__vmalloc_node(unsigned long size, unsigned long align,
3276 gfp_t gfp_mask, int node, const void *caller)
3278 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3279 gfp_mask, PAGE_KERNEL, 0, node, caller);
3282 * This is only for performance analysis of vmalloc and stress purpose.
3283 * It is required by vmalloc test module, therefore do not use it other
3286 #ifdef CONFIG_TEST_VMALLOC_MODULE
3287 EXPORT_SYMBOL_GPL(__vmalloc_node);
3290 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3292 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3293 __builtin_return_address(0));
3295 EXPORT_SYMBOL(__vmalloc);
3298 * vmalloc - allocate virtually contiguous memory
3299 * @size: allocation size
3301 * Allocate enough pages to cover @size from the page level
3302 * allocator and map them into contiguous kernel virtual space.
3304 * For tight control over page level allocator and protection flags
3305 * use __vmalloc() instead.
3307 * Return: pointer to the allocated memory or %NULL on error
3309 void *vmalloc(unsigned long size)
3311 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3312 __builtin_return_address(0));
3314 EXPORT_SYMBOL(vmalloc);
3317 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3318 * @size: allocation size
3319 * @gfp_mask: flags for the page level allocator
3321 * Allocate enough pages to cover @size from the page level
3322 * allocator and map them into contiguous kernel virtual space.
3323 * If @size is greater than or equal to PMD_SIZE, allow using
3324 * huge pages for the memory
3326 * Return: pointer to the allocated memory or %NULL on error
3328 void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3330 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3331 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3332 NUMA_NO_NODE, __builtin_return_address(0));
3334 EXPORT_SYMBOL_GPL(vmalloc_huge);
3337 * vzalloc - allocate virtually contiguous memory with zero fill
3338 * @size: allocation size
3340 * Allocate enough pages to cover @size from the page level
3341 * allocator and map them into contiguous kernel virtual space.
3342 * The memory allocated is set to zero.
3344 * For tight control over page level allocator and protection flags
3345 * use __vmalloc() instead.
3347 * Return: pointer to the allocated memory or %NULL on error
3349 void *vzalloc(unsigned long size)
3351 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3352 __builtin_return_address(0));
3354 EXPORT_SYMBOL(vzalloc);
3357 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3358 * @size: allocation size
3360 * The resulting memory area is zeroed so it can be mapped to userspace
3361 * without leaking data.
3363 * Return: pointer to the allocated memory or %NULL on error
3365 void *vmalloc_user(unsigned long size)
3367 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3368 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3369 VM_USERMAP, NUMA_NO_NODE,
3370 __builtin_return_address(0));
3372 EXPORT_SYMBOL(vmalloc_user);
3375 * vmalloc_node - allocate memory on a specific node
3376 * @size: allocation size
3379 * Allocate enough pages to cover @size from the page level
3380 * allocator and map them into contiguous kernel virtual space.
3382 * For tight control over page level allocator and protection flags
3383 * use __vmalloc() instead.
3385 * Return: pointer to the allocated memory or %NULL on error
3387 void *vmalloc_node(unsigned long size, int node)
3389 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3390 __builtin_return_address(0));
3392 EXPORT_SYMBOL(vmalloc_node);
3395 * vzalloc_node - allocate memory on a specific node with zero fill
3396 * @size: allocation size
3399 * Allocate enough pages to cover @size from the page level
3400 * allocator and map them into contiguous kernel virtual space.
3401 * The memory allocated is set to zero.
3403 * Return: pointer to the allocated memory or %NULL on error
3405 void *vzalloc_node(unsigned long size, int node)
3407 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3408 __builtin_return_address(0));
3410 EXPORT_SYMBOL(vzalloc_node);
3412 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3413 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3414 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3415 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3418 * 64b systems should always have either DMA or DMA32 zones. For others
3419 * GFP_DMA32 should do the right thing and use the normal zone.
3421 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3425 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3426 * @size: allocation size
3428 * Allocate enough 32bit PA addressable pages to cover @size from the
3429 * page level allocator and map them into contiguous kernel virtual space.
3431 * Return: pointer to the allocated memory or %NULL on error
3433 void *vmalloc_32(unsigned long size)
3435 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3436 __builtin_return_address(0));
3438 EXPORT_SYMBOL(vmalloc_32);
3441 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3442 * @size: allocation size
3444 * The resulting memory area is 32bit addressable and zeroed so it can be
3445 * mapped to userspace without leaking data.
3447 * Return: pointer to the allocated memory or %NULL on error
3449 void *vmalloc_32_user(unsigned long size)
3451 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3452 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3453 VM_USERMAP, NUMA_NO_NODE,
3454 __builtin_return_address(0));
3456 EXPORT_SYMBOL(vmalloc_32_user);
3459 * small helper routine , copy contents to buf from addr.
3460 * If the page is not present, fill zero.
3463 static int aligned_vread(char *buf, char *addr, unsigned long count)
3469 unsigned long offset, length;
3471 offset = offset_in_page(addr);
3472 length = PAGE_SIZE - offset;
3475 p = vmalloc_to_page(addr);
3477 * To do safe access to this _mapped_ area, we need
3478 * lock. But adding lock here means that we need to add
3479 * overhead of vmalloc()/vfree() calls for this _debug_
3480 * interface, rarely used. Instead of that, we'll use
3481 * kmap() and get small overhead in this access function.
3484 /* We can expect USER0 is not used -- see vread() */
3485 void *map = kmap_atomic(p);
3486 memcpy(buf, map + offset, length);
3489 memset(buf, 0, length);
3500 * vread() - read vmalloc area in a safe way.
3501 * @buf: buffer for reading data
3502 * @addr: vm address.
3503 * @count: number of bytes to be read.
3505 * This function checks that addr is a valid vmalloc'ed area, and
3506 * copy data from that area to a given buffer. If the given memory range
3507 * of [addr...addr+count) includes some valid address, data is copied to
3508 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3509 * IOREMAP area is treated as memory hole and no copy is done.
3511 * If [addr...addr+count) doesn't includes any intersects with alive
3512 * vm_struct area, returns 0. @buf should be kernel's buffer.
3514 * Note: In usual ops, vread() is never necessary because the caller
3515 * should know vmalloc() area is valid and can use memcpy().
3516 * This is for routines which have to access vmalloc area without
3517 * any information, as /proc/kcore.
3519 * Return: number of bytes for which addr and buf should be increased
3520 * (same number as @count) or %0 if [addr...addr+count) doesn't
3521 * include any intersection with valid vmalloc area
3523 long vread(char *buf, char *addr, unsigned long count)
3525 struct vmap_area *va;
3526 struct vm_struct *vm;
3527 char *vaddr, *buf_start = buf;
3528 unsigned long buflen = count;
3531 addr = kasan_reset_tag(addr);
3533 /* Don't allow overflow */
3534 if ((unsigned long) addr + count < count)
3535 count = -(unsigned long) addr;
3537 spin_lock(&vmap_area_lock);
3538 va = find_vmap_area_exceed_addr((unsigned long)addr);
3542 /* no intersects with alive vmap_area */
3543 if ((unsigned long)addr + count <= va->va_start)
3546 list_for_each_entry_from(va, &vmap_area_list, list) {
3554 vaddr = (char *) vm->addr;
3555 if (addr >= vaddr + get_vm_area_size(vm))
3557 while (addr < vaddr) {
3565 n = vaddr + get_vm_area_size(vm) - addr;
3568 if (!(vm->flags & VM_IOREMAP))
3569 aligned_vread(buf, addr, n);
3570 else /* IOREMAP area is treated as memory hole */
3577 spin_unlock(&vmap_area_lock);
3579 if (buf == buf_start)
3581 /* zero-fill memory holes */
3582 if (buf != buf_start + buflen)
3583 memset(buf, 0, buflen - (buf - buf_start));
3589 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3590 * @vma: vma to cover
3591 * @uaddr: target user address to start at
3592 * @kaddr: virtual address of vmalloc kernel memory
3593 * @pgoff: offset from @kaddr to start at
3594 * @size: size of map area
3596 * Returns: 0 for success, -Exxx on failure
3598 * This function checks that @kaddr is a valid vmalloc'ed area,
3599 * and that it is big enough to cover the range starting at
3600 * @uaddr in @vma. Will return failure if that criteria isn't
3603 * Similar to remap_pfn_range() (see mm/memory.c)
3605 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3606 void *kaddr, unsigned long pgoff,
3609 struct vm_struct *area;
3611 unsigned long end_index;
3613 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3616 size = PAGE_ALIGN(size);
3618 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3621 area = find_vm_area(kaddr);
3625 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3628 if (check_add_overflow(size, off, &end_index) ||
3629 end_index > get_vm_area_size(area))
3634 struct page *page = vmalloc_to_page(kaddr);
3637 ret = vm_insert_page(vma, uaddr, page);
3646 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3652 * remap_vmalloc_range - map vmalloc pages to userspace
3653 * @vma: vma to cover (map full range of vma)
3654 * @addr: vmalloc memory
3655 * @pgoff: number of pages into addr before first page to map
3657 * Returns: 0 for success, -Exxx on failure
3659 * This function checks that addr is a valid vmalloc'ed area, and
3660 * that it is big enough to cover the vma. Will return failure if
3661 * that criteria isn't met.
3663 * Similar to remap_pfn_range() (see mm/memory.c)
3665 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3666 unsigned long pgoff)
3668 return remap_vmalloc_range_partial(vma, vma->vm_start,
3670 vma->vm_end - vma->vm_start);
3672 EXPORT_SYMBOL(remap_vmalloc_range);
3674 void free_vm_area(struct vm_struct *area)
3676 struct vm_struct *ret;
3677 ret = remove_vm_area(area->addr);
3678 BUG_ON(ret != area);
3681 EXPORT_SYMBOL_GPL(free_vm_area);
3684 static struct vmap_area *node_to_va(struct rb_node *n)
3686 return rb_entry_safe(n, struct vmap_area, rb_node);
3690 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3691 * @addr: target address
3693 * Returns: vmap_area if it is found. If there is no such area
3694 * the first highest(reverse order) vmap_area is returned
3695 * i.e. va->va_start < addr && va->va_end < addr or NULL
3696 * if there are no any areas before @addr.
3698 static struct vmap_area *
3699 pvm_find_va_enclose_addr(unsigned long addr)
3701 struct vmap_area *va, *tmp;
3704 n = free_vmap_area_root.rb_node;
3708 tmp = rb_entry(n, struct vmap_area, rb_node);
3709 if (tmp->va_start <= addr) {
3711 if (tmp->va_end >= addr)
3724 * pvm_determine_end_from_reverse - find the highest aligned address
3725 * of free block below VMALLOC_END
3727 * in - the VA we start the search(reverse order);
3728 * out - the VA with the highest aligned end address.
3729 * @align: alignment for required highest address
3731 * Returns: determined end address within vmap_area
3733 static unsigned long
3734 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3736 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3740 list_for_each_entry_from_reverse((*va),
3741 &free_vmap_area_list, list) {
3742 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3743 if ((*va)->va_start < addr)
3752 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3753 * @offsets: array containing offset of each area
3754 * @sizes: array containing size of each area
3755 * @nr_vms: the number of areas to allocate
3756 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3758 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3759 * vm_structs on success, %NULL on failure
3761 * Percpu allocator wants to use congruent vm areas so that it can
3762 * maintain the offsets among percpu areas. This function allocates
3763 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3764 * be scattered pretty far, distance between two areas easily going up
3765 * to gigabytes. To avoid interacting with regular vmallocs, these
3766 * areas are allocated from top.
3768 * Despite its complicated look, this allocator is rather simple. It
3769 * does everything top-down and scans free blocks from the end looking
3770 * for matching base. While scanning, if any of the areas do not fit the
3771 * base address is pulled down to fit the area. Scanning is repeated till
3772 * all the areas fit and then all necessary data structures are inserted
3773 * and the result is returned.
3775 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3776 const size_t *sizes, int nr_vms,
3779 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3780 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3781 struct vmap_area **vas, *va;
3782 struct vm_struct **vms;
3783 int area, area2, last_area, term_area;
3784 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3785 bool purged = false;
3787 /* verify parameters and allocate data structures */
3788 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3789 for (last_area = 0, area = 0; area < nr_vms; area++) {
3790 start = offsets[area];
3791 end = start + sizes[area];
3793 /* is everything aligned properly? */
3794 BUG_ON(!IS_ALIGNED(offsets[area], align));
3795 BUG_ON(!IS_ALIGNED(sizes[area], align));
3797 /* detect the area with the highest address */
3798 if (start > offsets[last_area])
3801 for (area2 = area + 1; area2 < nr_vms; area2++) {
3802 unsigned long start2 = offsets[area2];
3803 unsigned long end2 = start2 + sizes[area2];
3805 BUG_ON(start2 < end && start < end2);
3808 last_end = offsets[last_area] + sizes[last_area];
3810 if (vmalloc_end - vmalloc_start < last_end) {
3815 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3816 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3820 for (area = 0; area < nr_vms; area++) {
3821 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3822 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3823 if (!vas[area] || !vms[area])
3827 spin_lock(&free_vmap_area_lock);
3829 /* start scanning - we scan from the top, begin with the last area */
3830 area = term_area = last_area;
3831 start = offsets[area];
3832 end = start + sizes[area];
3834 va = pvm_find_va_enclose_addr(vmalloc_end);
3835 base = pvm_determine_end_from_reverse(&va, align) - end;
3839 * base might have underflowed, add last_end before
3842 if (base + last_end < vmalloc_start + last_end)
3846 * Fitting base has not been found.
3852 * If required width exceeds current VA block, move
3853 * base downwards and then recheck.
3855 if (base + end > va->va_end) {
3856 base = pvm_determine_end_from_reverse(&va, align) - end;
3862 * If this VA does not fit, move base downwards and recheck.
3864 if (base + start < va->va_start) {
3865 va = node_to_va(rb_prev(&va->rb_node));
3866 base = pvm_determine_end_from_reverse(&va, align) - end;
3872 * This area fits, move on to the previous one. If
3873 * the previous one is the terminal one, we're done.
3875 area = (area + nr_vms - 1) % nr_vms;
3876 if (area == term_area)
3879 start = offsets[area];
3880 end = start + sizes[area];
3881 va = pvm_find_va_enclose_addr(base + end);
3884 /* we've found a fitting base, insert all va's */
3885 for (area = 0; area < nr_vms; area++) {
3888 start = base + offsets[area];
3891 va = pvm_find_va_enclose_addr(start);
3892 if (WARN_ON_ONCE(va == NULL))
3893 /* It is a BUG(), but trigger recovery instead. */
3896 ret = adjust_va_to_fit_type(&free_vmap_area_root,
3897 &free_vmap_area_list,
3899 if (WARN_ON_ONCE(unlikely(ret)))
3900 /* It is a BUG(), but trigger recovery instead. */
3903 /* Allocated area. */
3905 va->va_start = start;
3906 va->va_end = start + size;
3909 spin_unlock(&free_vmap_area_lock);
3911 /* populate the kasan shadow space */
3912 for (area = 0; area < nr_vms; area++) {
3913 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3914 goto err_free_shadow;
3917 /* insert all vm's */
3918 spin_lock(&vmap_area_lock);
3919 for (area = 0; area < nr_vms; area++) {
3920 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3922 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3925 spin_unlock(&vmap_area_lock);
3928 * Mark allocated areas as accessible. Do it now as a best-effort
3929 * approach, as they can be mapped outside of vmalloc code.
3930 * With hardware tag-based KASAN, marking is skipped for
3931 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3933 for (area = 0; area < nr_vms; area++)
3934 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3935 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3942 * Remove previously allocated areas. There is no
3943 * need in removing these areas from the busy tree,
3944 * because they are inserted only on the final step
3945 * and when pcpu_get_vm_areas() is success.
3948 orig_start = vas[area]->va_start;
3949 orig_end = vas[area]->va_end;
3950 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3951 &free_vmap_area_list);
3953 kasan_release_vmalloc(orig_start, orig_end,
3954 va->va_start, va->va_end);
3959 spin_unlock(&free_vmap_area_lock);
3961 purge_vmap_area_lazy();
3964 /* Before "retry", check if we recover. */
3965 for (area = 0; area < nr_vms; area++) {
3969 vas[area] = kmem_cache_zalloc(
3970 vmap_area_cachep, GFP_KERNEL);
3979 for (area = 0; area < nr_vms; area++) {
3981 kmem_cache_free(vmap_area_cachep, vas[area]);
3991 spin_lock(&free_vmap_area_lock);
3993 * We release all the vmalloc shadows, even the ones for regions that
3994 * hadn't been successfully added. This relies on kasan_release_vmalloc
3995 * being able to tolerate this case.
3997 for (area = 0; area < nr_vms; area++) {
3998 orig_start = vas[area]->va_start;
3999 orig_end = vas[area]->va_end;
4000 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4001 &free_vmap_area_list);
4003 kasan_release_vmalloc(orig_start, orig_end,
4004 va->va_start, va->va_end);
4008 spin_unlock(&free_vmap_area_lock);
4015 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4016 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4017 * @nr_vms: the number of allocated areas
4019 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4021 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4025 for (i = 0; i < nr_vms; i++)
4026 free_vm_area(vms[i]);
4029 #endif /* CONFIG_SMP */
4031 #ifdef CONFIG_PRINTK
4032 bool vmalloc_dump_obj(void *object)
4034 struct vm_struct *vm;
4035 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4037 vm = find_vm_area(objp);
4040 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4041 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4046 #ifdef CONFIG_PROC_FS
4047 static void *s_start(struct seq_file *m, loff_t *pos)
4048 __acquires(&vmap_purge_lock)
4049 __acquires(&vmap_area_lock)
4051 mutex_lock(&vmap_purge_lock);
4052 spin_lock(&vmap_area_lock);
4054 return seq_list_start(&vmap_area_list, *pos);
4057 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4059 return seq_list_next(p, &vmap_area_list, pos);
4062 static void s_stop(struct seq_file *m, void *p)
4063 __releases(&vmap_area_lock)
4064 __releases(&vmap_purge_lock)
4066 spin_unlock(&vmap_area_lock);
4067 mutex_unlock(&vmap_purge_lock);
4070 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4072 if (IS_ENABLED(CONFIG_NUMA)) {
4073 unsigned int nr, *counters = m->private;
4074 unsigned int step = 1U << vm_area_page_order(v);
4079 if (v->flags & VM_UNINITIALIZED)
4081 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4084 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4086 for (nr = 0; nr < v->nr_pages; nr += step)
4087 counters[page_to_nid(v->pages[nr])] += step;
4088 for_each_node_state(nr, N_HIGH_MEMORY)
4090 seq_printf(m, " N%u=%u", nr, counters[nr]);
4094 static void show_purge_info(struct seq_file *m)
4096 struct vmap_area *va;
4098 spin_lock(&purge_vmap_area_lock);
4099 list_for_each_entry(va, &purge_vmap_area_list, list) {
4100 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4101 (void *)va->va_start, (void *)va->va_end,
4102 va->va_end - va->va_start);
4104 spin_unlock(&purge_vmap_area_lock);
4107 static int s_show(struct seq_file *m, void *p)
4109 struct vmap_area *va;
4110 struct vm_struct *v;
4112 va = list_entry(p, struct vmap_area, list);
4115 * s_show can encounter race with remove_vm_area, !vm on behalf
4116 * of vmap area is being tear down or vm_map_ram allocation.
4119 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4120 (void *)va->va_start, (void *)va->va_end,
4121 va->va_end - va->va_start);
4128 seq_printf(m, "0x%pK-0x%pK %7ld",
4129 v->addr, v->addr + v->size, v->size);
4132 seq_printf(m, " %pS", v->caller);
4135 seq_printf(m, " pages=%d", v->nr_pages);
4138 seq_printf(m, " phys=%pa", &v->phys_addr);
4140 if (v->flags & VM_IOREMAP)
4141 seq_puts(m, " ioremap");
4143 if (v->flags & VM_ALLOC)
4144 seq_puts(m, " vmalloc");
4146 if (v->flags & VM_MAP)
4147 seq_puts(m, " vmap");
4149 if (v->flags & VM_USERMAP)
4150 seq_puts(m, " user");
4152 if (v->flags & VM_DMA_COHERENT)
4153 seq_puts(m, " dma-coherent");
4155 if (is_vmalloc_addr(v->pages))
4156 seq_puts(m, " vpages");
4158 show_numa_info(m, v);
4162 * As a final step, dump "unpurged" areas.
4165 if (list_is_last(&va->list, &vmap_area_list))
4171 static const struct seq_operations vmalloc_op = {
4178 static int __init proc_vmalloc_init(void)
4180 if (IS_ENABLED(CONFIG_NUMA))
4181 proc_create_seq_private("vmallocinfo", 0400, NULL,
4183 nr_node_ids * sizeof(unsigned int), NULL);
4185 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4188 module_init(proc_vmalloc_init);