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 err = 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 int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
621 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
625 * vmap_pages_range - map pages to a kernel virtual address
626 * @addr: start of the VM area to map
627 * @end: end of the VM area to map (non-inclusive)
628 * @prot: page protection flags to use
629 * @pages: pages to map (always PAGE_SIZE pages)
630 * @page_shift: maximum shift that the pages may be mapped with, @pages must
631 * be aligned and contiguous up to at least this shift.
634 * 0 on success, -errno on failure.
636 static int vmap_pages_range(unsigned long addr, unsigned long end,
637 pgprot_t prot, struct page **pages, unsigned int page_shift)
641 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
642 flush_cache_vmap(addr, end);
646 int is_vmalloc_or_module_addr(const void *x)
649 * ARM, x86-64 and sparc64 put modules in a special place,
650 * and fall back on vmalloc() if that fails. Others
651 * just put it in the vmalloc space.
653 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
654 unsigned long addr = (unsigned long)kasan_reset_tag(x);
655 if (addr >= MODULES_VADDR && addr < MODULES_END)
658 return is_vmalloc_addr(x);
662 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
663 * return the tail page that corresponds to the base page address, which
664 * matches small vmap mappings.
666 struct page *vmalloc_to_page(const void *vmalloc_addr)
668 unsigned long addr = (unsigned long) vmalloc_addr;
669 struct page *page = NULL;
670 pgd_t *pgd = pgd_offset_k(addr);
677 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
678 * architectures that do not vmalloc module space
680 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
684 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
685 return NULL; /* XXX: no allowance for huge pgd */
686 if (WARN_ON_ONCE(pgd_bad(*pgd)))
689 p4d = p4d_offset(pgd, addr);
693 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
694 if (WARN_ON_ONCE(p4d_bad(*p4d)))
697 pud = pud_offset(p4d, addr);
701 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
702 if (WARN_ON_ONCE(pud_bad(*pud)))
705 pmd = pmd_offset(pud, addr);
709 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
710 if (WARN_ON_ONCE(pmd_bad(*pmd)))
713 ptep = pte_offset_map(pmd, addr);
715 if (pte_present(pte))
716 page = pte_page(pte);
721 EXPORT_SYMBOL(vmalloc_to_page);
724 * Map a vmalloc()-space virtual address to the physical page frame number.
726 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
728 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
730 EXPORT_SYMBOL(vmalloc_to_pfn);
733 /*** Global kva allocator ***/
735 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
736 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
739 static DEFINE_SPINLOCK(vmap_area_lock);
740 static DEFINE_SPINLOCK(free_vmap_area_lock);
741 /* Export for kexec only */
742 LIST_HEAD(vmap_area_list);
743 static struct rb_root vmap_area_root = RB_ROOT;
744 static bool vmap_initialized __read_mostly;
746 static struct rb_root purge_vmap_area_root = RB_ROOT;
747 static LIST_HEAD(purge_vmap_area_list);
748 static DEFINE_SPINLOCK(purge_vmap_area_lock);
751 * This kmem_cache is used for vmap_area objects. Instead of
752 * allocating from slab we reuse an object from this cache to
753 * make things faster. Especially in "no edge" splitting of
756 static struct kmem_cache *vmap_area_cachep;
759 * This linked list is used in pair with free_vmap_area_root.
760 * It gives O(1) access to prev/next to perform fast coalescing.
762 static LIST_HEAD(free_vmap_area_list);
765 * This augment red-black tree represents the free vmap space.
766 * All vmap_area objects in this tree are sorted by va->va_start
767 * address. It is used for allocation and merging when a vmap
768 * object is released.
770 * Each vmap_area node contains a maximum available free block
771 * of its sub-tree, right or left. Therefore it is possible to
772 * find a lowest match of free area.
774 static struct rb_root free_vmap_area_root = RB_ROOT;
777 * Preload a CPU with one object for "no edge" split case. The
778 * aim is to get rid of allocations from the atomic context, thus
779 * to use more permissive allocation masks.
781 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
783 static __always_inline unsigned long
784 va_size(struct vmap_area *va)
786 return (va->va_end - va->va_start);
789 static __always_inline unsigned long
790 get_subtree_max_size(struct rb_node *node)
792 struct vmap_area *va;
794 va = rb_entry_safe(node, struct vmap_area, rb_node);
795 return va ? va->subtree_max_size : 0;
798 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
799 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
801 static void purge_vmap_area_lazy(void);
802 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
803 static void drain_vmap_area_work(struct work_struct *work);
804 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
806 static atomic_long_t nr_vmalloc_pages;
808 unsigned long vmalloc_nr_pages(void)
810 return atomic_long_read(&nr_vmalloc_pages);
813 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
814 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
816 struct vmap_area *va = NULL;
817 struct rb_node *n = vmap_area_root.rb_node;
819 addr = (unsigned long)kasan_reset_tag((void *)addr);
822 struct vmap_area *tmp;
824 tmp = rb_entry(n, struct vmap_area, rb_node);
825 if (tmp->va_end > addr) {
827 if (tmp->va_start <= addr)
838 static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
840 struct rb_node *n = root->rb_node;
842 addr = (unsigned long)kasan_reset_tag((void *)addr);
845 struct vmap_area *va;
847 va = rb_entry(n, struct vmap_area, rb_node);
848 if (addr < va->va_start)
850 else if (addr >= va->va_end)
860 * This function returns back addresses of parent node
861 * and its left or right link for further processing.
863 * Otherwise NULL is returned. In that case all further
864 * steps regarding inserting of conflicting overlap range
865 * have to be declined and actually considered as a bug.
867 static __always_inline struct rb_node **
868 find_va_links(struct vmap_area *va,
869 struct rb_root *root, struct rb_node *from,
870 struct rb_node **parent)
872 struct vmap_area *tmp_va;
873 struct rb_node **link;
876 link = &root->rb_node;
877 if (unlikely(!*link)) {
886 * Go to the bottom of the tree. When we hit the last point
887 * we end up with parent rb_node and correct direction, i name
888 * it link, where the new va->rb_node will be attached to.
891 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
894 * During the traversal we also do some sanity check.
895 * Trigger the BUG() if there are sides(left/right)
898 if (va->va_end <= tmp_va->va_start)
899 link = &(*link)->rb_left;
900 else if (va->va_start >= tmp_va->va_end)
901 link = &(*link)->rb_right;
903 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
904 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
910 *parent = &tmp_va->rb_node;
914 static __always_inline struct list_head *
915 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
917 struct list_head *list;
919 if (unlikely(!parent))
921 * The red-black tree where we try to find VA neighbors
922 * before merging or inserting is empty, i.e. it means
923 * there is no free vmap space. Normally it does not
924 * happen but we handle this case anyway.
928 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
929 return (&parent->rb_right == link ? list->next : list);
932 static __always_inline void
933 __link_va(struct vmap_area *va, struct rb_root *root,
934 struct rb_node *parent, struct rb_node **link,
935 struct list_head *head, bool augment)
938 * VA is still not in the list, but we can
939 * identify its future previous list_head node.
941 if (likely(parent)) {
942 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
943 if (&parent->rb_right != link)
947 /* Insert to the rb-tree */
948 rb_link_node(&va->rb_node, parent, link);
951 * Some explanation here. Just perform simple insertion
952 * to the tree. We do not set va->subtree_max_size to
953 * its current size before calling rb_insert_augmented().
954 * It is because we populate the tree from the bottom
955 * to parent levels when the node _is_ in the tree.
957 * Therefore we set subtree_max_size to zero after insertion,
958 * to let __augment_tree_propagate_from() puts everything to
959 * the correct order later on.
961 rb_insert_augmented(&va->rb_node,
962 root, &free_vmap_area_rb_augment_cb);
963 va->subtree_max_size = 0;
965 rb_insert_color(&va->rb_node, root);
968 /* Address-sort this list */
969 list_add(&va->list, head);
972 static __always_inline void
973 link_va(struct vmap_area *va, struct rb_root *root,
974 struct rb_node *parent, struct rb_node **link,
975 struct list_head *head)
977 __link_va(va, root, parent, link, head, false);
980 static __always_inline void
981 link_va_augment(struct vmap_area *va, struct rb_root *root,
982 struct rb_node *parent, struct rb_node **link,
983 struct list_head *head)
985 __link_va(va, root, parent, link, head, true);
988 static __always_inline void
989 __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
991 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
995 rb_erase_augmented(&va->rb_node,
996 root, &free_vmap_area_rb_augment_cb);
998 rb_erase(&va->rb_node, root);
1000 list_del_init(&va->list);
1001 RB_CLEAR_NODE(&va->rb_node);
1004 static __always_inline void
1005 unlink_va(struct vmap_area *va, struct rb_root *root)
1007 __unlink_va(va, root, false);
1010 static __always_inline void
1011 unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1013 __unlink_va(va, root, true);
1016 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1018 * Gets called when remove the node and rotate.
1020 static __always_inline unsigned long
1021 compute_subtree_max_size(struct vmap_area *va)
1023 return max3(va_size(va),
1024 get_subtree_max_size(va->rb_node.rb_left),
1025 get_subtree_max_size(va->rb_node.rb_right));
1029 augment_tree_propagate_check(void)
1031 struct vmap_area *va;
1032 unsigned long computed_size;
1034 list_for_each_entry(va, &free_vmap_area_list, list) {
1035 computed_size = compute_subtree_max_size(va);
1036 if (computed_size != va->subtree_max_size)
1037 pr_emerg("tree is corrupted: %lu, %lu\n",
1038 va_size(va), va->subtree_max_size);
1044 * This function populates subtree_max_size from bottom to upper
1045 * levels starting from VA point. The propagation must be done
1046 * when VA size is modified by changing its va_start/va_end. Or
1047 * in case of newly inserting of VA to the tree.
1049 * It means that __augment_tree_propagate_from() must be called:
1050 * - After VA has been inserted to the tree(free path);
1051 * - After VA has been shrunk(allocation path);
1052 * - After VA has been increased(merging path).
1054 * Please note that, it does not mean that upper parent nodes
1055 * and their subtree_max_size are recalculated all the time up
1064 * For example if we modify the node 4, shrinking it to 2, then
1065 * no any modification is required. If we shrink the node 2 to 1
1066 * its subtree_max_size is updated only, and set to 1. If we shrink
1067 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1068 * node becomes 4--6.
1070 static __always_inline void
1071 augment_tree_propagate_from(struct vmap_area *va)
1074 * Populate the tree from bottom towards the root until
1075 * the calculated maximum available size of checked node
1076 * is equal to its current one.
1078 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1080 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1081 augment_tree_propagate_check();
1086 insert_vmap_area(struct vmap_area *va,
1087 struct rb_root *root, struct list_head *head)
1089 struct rb_node **link;
1090 struct rb_node *parent;
1092 link = find_va_links(va, root, NULL, &parent);
1094 link_va(va, root, parent, link, head);
1098 insert_vmap_area_augment(struct vmap_area *va,
1099 struct rb_node *from, struct rb_root *root,
1100 struct list_head *head)
1102 struct rb_node **link;
1103 struct rb_node *parent;
1106 link = find_va_links(va, NULL, from, &parent);
1108 link = find_va_links(va, root, NULL, &parent);
1111 link_va_augment(va, root, parent, link, head);
1112 augment_tree_propagate_from(va);
1117 * Merge de-allocated chunk of VA memory with previous
1118 * and next free blocks. If coalesce is not done a new
1119 * free area is inserted. If VA has been merged, it is
1122 * Please note, it can return NULL in case of overlap
1123 * ranges, followed by WARN() report. Despite it is a
1124 * buggy behaviour, a system can be alive and keep
1127 static __always_inline struct vmap_area *
1128 __merge_or_add_vmap_area(struct vmap_area *va,
1129 struct rb_root *root, struct list_head *head, bool augment)
1131 struct vmap_area *sibling;
1132 struct list_head *next;
1133 struct rb_node **link;
1134 struct rb_node *parent;
1135 bool merged = false;
1138 * Find a place in the tree where VA potentially will be
1139 * inserted, unless it is merged with its sibling/siblings.
1141 link = find_va_links(va, root, NULL, &parent);
1146 * Get next node of VA to check if merging can be done.
1148 next = get_va_next_sibling(parent, link);
1149 if (unlikely(next == NULL))
1155 * |<------VA------>|<-----Next----->|
1160 sibling = list_entry(next, struct vmap_area, list);
1161 if (sibling->va_start == va->va_end) {
1162 sibling->va_start = va->va_start;
1164 /* Free vmap_area object. */
1165 kmem_cache_free(vmap_area_cachep, va);
1167 /* Point to the new merged area. */
1176 * |<-----Prev----->|<------VA------>|
1180 if (next->prev != head) {
1181 sibling = list_entry(next->prev, struct vmap_area, list);
1182 if (sibling->va_end == va->va_start) {
1184 * If both neighbors are coalesced, it is important
1185 * to unlink the "next" node first, followed by merging
1186 * with "previous" one. Otherwise the tree might not be
1187 * fully populated if a sibling's augmented value is
1188 * "normalized" because of rotation operations.
1191 __unlink_va(va, root, augment);
1193 sibling->va_end = va->va_end;
1195 /* Free vmap_area object. */
1196 kmem_cache_free(vmap_area_cachep, va);
1198 /* Point to the new merged area. */
1206 __link_va(va, root, parent, link, head, augment);
1211 static __always_inline struct vmap_area *
1212 merge_or_add_vmap_area(struct vmap_area *va,
1213 struct rb_root *root, struct list_head *head)
1215 return __merge_or_add_vmap_area(va, root, head, false);
1218 static __always_inline struct vmap_area *
1219 merge_or_add_vmap_area_augment(struct vmap_area *va,
1220 struct rb_root *root, struct list_head *head)
1222 va = __merge_or_add_vmap_area(va, root, head, true);
1224 augment_tree_propagate_from(va);
1229 static __always_inline bool
1230 is_within_this_va(struct vmap_area *va, unsigned long size,
1231 unsigned long align, unsigned long vstart)
1233 unsigned long nva_start_addr;
1235 if (va->va_start > vstart)
1236 nva_start_addr = ALIGN(va->va_start, align);
1238 nva_start_addr = ALIGN(vstart, align);
1240 /* Can be overflowed due to big size or alignment. */
1241 if (nva_start_addr + size < nva_start_addr ||
1242 nva_start_addr < vstart)
1245 return (nva_start_addr + size <= va->va_end);
1249 * Find the first free block(lowest start address) in the tree,
1250 * that will accomplish the request corresponding to passing
1251 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1252 * a search length is adjusted to account for worst case alignment
1255 static __always_inline struct vmap_area *
1256 find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1257 unsigned long align, unsigned long vstart, bool adjust_search_size)
1259 struct vmap_area *va;
1260 struct rb_node *node;
1261 unsigned long length;
1263 /* Start from the root. */
1264 node = root->rb_node;
1266 /* Adjust the search size for alignment overhead. */
1267 length = adjust_search_size ? size + align - 1 : size;
1270 va = rb_entry(node, struct vmap_area, rb_node);
1272 if (get_subtree_max_size(node->rb_left) >= length &&
1273 vstart < va->va_start) {
1274 node = node->rb_left;
1276 if (is_within_this_va(va, size, align, vstart))
1280 * Does not make sense to go deeper towards the right
1281 * sub-tree if it does not have a free block that is
1282 * equal or bigger to the requested search length.
1284 if (get_subtree_max_size(node->rb_right) >= length) {
1285 node = node->rb_right;
1290 * OK. We roll back and find the first right sub-tree,
1291 * that will satisfy the search criteria. It can happen
1292 * due to "vstart" restriction or an alignment overhead
1293 * that is bigger then PAGE_SIZE.
1295 while ((node = rb_parent(node))) {
1296 va = rb_entry(node, struct vmap_area, rb_node);
1297 if (is_within_this_va(va, size, align, vstart))
1300 if (get_subtree_max_size(node->rb_right) >= length &&
1301 vstart <= va->va_start) {
1303 * Shift the vstart forward. Please note, we update it with
1304 * parent's start address adding "1" because we do not want
1305 * to enter same sub-tree after it has already been checked
1306 * and no suitable free block found there.
1308 vstart = va->va_start + 1;
1309 node = node->rb_right;
1319 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1320 #include <linux/random.h>
1322 static struct vmap_area *
1323 find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1324 unsigned long align, unsigned long vstart)
1326 struct vmap_area *va;
1328 list_for_each_entry(va, head, list) {
1329 if (!is_within_this_va(va, size, align, vstart))
1339 find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1340 unsigned long size, unsigned long align)
1342 struct vmap_area *va_1, *va_2;
1343 unsigned long vstart;
1346 get_random_bytes(&rnd, sizeof(rnd));
1347 vstart = VMALLOC_START + rnd;
1349 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1350 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1353 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1354 va_1, va_2, vstart);
1360 FL_FIT_TYPE = 1, /* full fit */
1361 LE_FIT_TYPE = 2, /* left edge fit */
1362 RE_FIT_TYPE = 3, /* right edge fit */
1363 NE_FIT_TYPE = 4 /* no edge fit */
1366 static __always_inline enum fit_type
1367 classify_va_fit_type(struct vmap_area *va,
1368 unsigned long nva_start_addr, unsigned long size)
1372 /* Check if it is within VA. */
1373 if (nva_start_addr < va->va_start ||
1374 nva_start_addr + size > va->va_end)
1378 if (va->va_start == nva_start_addr) {
1379 if (va->va_end == nva_start_addr + size)
1383 } else if (va->va_end == nva_start_addr + size) {
1392 static __always_inline int
1393 adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1394 struct vmap_area *va, unsigned long nva_start_addr,
1397 struct vmap_area *lva = NULL;
1398 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1400 if (type == FL_FIT_TYPE) {
1402 * No need to split VA, it fully fits.
1408 unlink_va_augment(va, root);
1409 kmem_cache_free(vmap_area_cachep, va);
1410 } else if (type == LE_FIT_TYPE) {
1412 * Split left edge of fit VA.
1418 va->va_start += size;
1419 } else if (type == RE_FIT_TYPE) {
1421 * Split right edge of fit VA.
1427 va->va_end = nva_start_addr;
1428 } else if (type == NE_FIT_TYPE) {
1430 * Split no edge of fit VA.
1436 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1437 if (unlikely(!lva)) {
1439 * For percpu allocator we do not do any pre-allocation
1440 * and leave it as it is. The reason is it most likely
1441 * never ends up with NE_FIT_TYPE splitting. In case of
1442 * percpu allocations offsets and sizes are aligned to
1443 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1444 * are its main fitting cases.
1446 * There are a few exceptions though, as an example it is
1447 * a first allocation (early boot up) when we have "one"
1448 * big free space that has to be split.
1450 * Also we can hit this path in case of regular "vmap"
1451 * allocations, if "this" current CPU was not preloaded.
1452 * See the comment in alloc_vmap_area() why. If so, then
1453 * GFP_NOWAIT is used instead to get an extra object for
1454 * split purpose. That is rare and most time does not
1457 * What happens if an allocation gets failed. Basically,
1458 * an "overflow" path is triggered to purge lazily freed
1459 * areas to free some memory, then, the "retry" path is
1460 * triggered to repeat one more time. See more details
1461 * in alloc_vmap_area() function.
1463 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1469 * Build the remainder.
1471 lva->va_start = va->va_start;
1472 lva->va_end = nva_start_addr;
1475 * Shrink this VA to remaining size.
1477 va->va_start = nva_start_addr + size;
1482 if (type != FL_FIT_TYPE) {
1483 augment_tree_propagate_from(va);
1485 if (lva) /* type == NE_FIT_TYPE */
1486 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1493 * Returns a start address of the newly allocated area, if success.
1494 * Otherwise a vend is returned that indicates failure.
1496 static __always_inline unsigned long
1497 __alloc_vmap_area(struct rb_root *root, struct list_head *head,
1498 unsigned long size, unsigned long align,
1499 unsigned long vstart, unsigned long vend)
1501 bool adjust_search_size = true;
1502 unsigned long nva_start_addr;
1503 struct vmap_area *va;
1507 * Do not adjust when:
1508 * a) align <= PAGE_SIZE, because it does not make any sense.
1509 * All blocks(their start addresses) are at least PAGE_SIZE
1511 * b) a short range where a requested size corresponds to exactly
1512 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1513 * With adjusted search length an allocation would not succeed.
1515 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1516 adjust_search_size = false;
1518 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1522 if (va->va_start > vstart)
1523 nva_start_addr = ALIGN(va->va_start, align);
1525 nva_start_addr = ALIGN(vstart, align);
1527 /* Check the "vend" restriction. */
1528 if (nva_start_addr + size > vend)
1531 /* Update the free vmap_area. */
1532 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1533 if (WARN_ON_ONCE(ret))
1536 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1537 find_vmap_lowest_match_check(root, head, size, align);
1540 return nva_start_addr;
1544 * Free a region of KVA allocated by alloc_vmap_area
1546 static void free_vmap_area(struct vmap_area *va)
1549 * Remove from the busy tree/list.
1551 spin_lock(&vmap_area_lock);
1552 unlink_va(va, &vmap_area_root);
1553 spin_unlock(&vmap_area_lock);
1556 * Insert/Merge it back to the free tree/list.
1558 spin_lock(&free_vmap_area_lock);
1559 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1560 spin_unlock(&free_vmap_area_lock);
1564 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1566 struct vmap_area *va = NULL;
1569 * Preload this CPU with one extra vmap_area object. It is used
1570 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1571 * a CPU that does an allocation is preloaded.
1573 * We do it in non-atomic context, thus it allows us to use more
1574 * permissive allocation masks to be more stable under low memory
1575 * condition and high memory pressure.
1577 if (!this_cpu_read(ne_fit_preload_node))
1578 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1582 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1583 kmem_cache_free(vmap_area_cachep, va);
1587 * Allocate a region of KVA of the specified size and alignment, within the
1590 static struct vmap_area *alloc_vmap_area(unsigned long size,
1591 unsigned long align,
1592 unsigned long vstart, unsigned long vend,
1593 int node, gfp_t gfp_mask)
1595 struct vmap_area *va;
1596 unsigned long freed;
1602 BUG_ON(offset_in_page(size));
1603 BUG_ON(!is_power_of_2(align));
1605 if (unlikely(!vmap_initialized))
1606 return ERR_PTR(-EBUSY);
1609 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1611 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1613 return ERR_PTR(-ENOMEM);
1616 * Only scan the relevant parts containing pointers to other objects
1617 * to avoid false negatives.
1619 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1622 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1623 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1624 size, align, vstart, vend);
1625 spin_unlock(&free_vmap_area_lock);
1628 * If an allocation fails, the "vend" address is
1629 * returned. Therefore trigger the overflow path.
1631 if (unlikely(addr == vend))
1634 va->va_start = addr;
1635 va->va_end = addr + size;
1638 spin_lock(&vmap_area_lock);
1639 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1640 spin_unlock(&vmap_area_lock);
1642 BUG_ON(!IS_ALIGNED(va->va_start, align));
1643 BUG_ON(va->va_start < vstart);
1644 BUG_ON(va->va_end > vend);
1646 ret = kasan_populate_vmalloc(addr, size);
1649 return ERR_PTR(ret);
1656 purge_vmap_area_lazy();
1662 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1669 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1670 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1673 kmem_cache_free(vmap_area_cachep, va);
1674 return ERR_PTR(-EBUSY);
1677 int register_vmap_purge_notifier(struct notifier_block *nb)
1679 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1681 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1683 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1685 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1687 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1690 * lazy_max_pages is the maximum amount of virtual address space we gather up
1691 * before attempting to purge with a TLB flush.
1693 * There is a tradeoff here: a larger number will cover more kernel page tables
1694 * and take slightly longer to purge, but it will linearly reduce the number of
1695 * global TLB flushes that must be performed. It would seem natural to scale
1696 * this number up linearly with the number of CPUs (because vmapping activity
1697 * could also scale linearly with the number of CPUs), however it is likely
1698 * that in practice, workloads might be constrained in other ways that mean
1699 * vmap activity will not scale linearly with CPUs. Also, I want to be
1700 * conservative and not introduce a big latency on huge systems, so go with
1701 * a less aggressive log scale. It will still be an improvement over the old
1702 * code, and it will be simple to change the scale factor if we find that it
1703 * becomes a problem on bigger systems.
1705 static unsigned long lazy_max_pages(void)
1709 log = fls(num_online_cpus());
1711 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1714 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1717 * Serialize vmap purging. There is no actual critical section protected
1718 * by this lock, but we want to avoid concurrent calls for performance
1719 * reasons and to make the pcpu_get_vm_areas more deterministic.
1721 static DEFINE_MUTEX(vmap_purge_lock);
1723 /* for per-CPU blocks */
1724 static void purge_fragmented_blocks_allcpus(void);
1727 * Purges all lazily-freed vmap areas.
1729 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1731 unsigned long resched_threshold;
1732 struct list_head local_purge_list;
1733 struct vmap_area *va, *n_va;
1735 lockdep_assert_held(&vmap_purge_lock);
1737 spin_lock(&purge_vmap_area_lock);
1738 purge_vmap_area_root = RB_ROOT;
1739 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1740 spin_unlock(&purge_vmap_area_lock);
1742 if (unlikely(list_empty(&local_purge_list)))
1746 list_first_entry(&local_purge_list,
1747 struct vmap_area, list)->va_start);
1750 list_last_entry(&local_purge_list,
1751 struct vmap_area, list)->va_end);
1753 flush_tlb_kernel_range(start, end);
1754 resched_threshold = lazy_max_pages() << 1;
1756 spin_lock(&free_vmap_area_lock);
1757 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1758 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1759 unsigned long orig_start = va->va_start;
1760 unsigned long orig_end = va->va_end;
1763 * Finally insert or merge lazily-freed area. It is
1764 * detached and there is no need to "unlink" it from
1767 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1768 &free_vmap_area_list);
1773 if (is_vmalloc_or_module_addr((void *)orig_start))
1774 kasan_release_vmalloc(orig_start, orig_end,
1775 va->va_start, va->va_end);
1777 atomic_long_sub(nr, &vmap_lazy_nr);
1779 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1780 cond_resched_lock(&free_vmap_area_lock);
1782 spin_unlock(&free_vmap_area_lock);
1787 * Kick off a purge of the outstanding lazy areas.
1789 static void purge_vmap_area_lazy(void)
1791 mutex_lock(&vmap_purge_lock);
1792 purge_fragmented_blocks_allcpus();
1793 __purge_vmap_area_lazy(ULONG_MAX, 0);
1794 mutex_unlock(&vmap_purge_lock);
1797 static void drain_vmap_area_work(struct work_struct *work)
1799 unsigned long nr_lazy;
1802 mutex_lock(&vmap_purge_lock);
1803 __purge_vmap_area_lazy(ULONG_MAX, 0);
1804 mutex_unlock(&vmap_purge_lock);
1806 /* Recheck if further work is required. */
1807 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1808 } while (nr_lazy > lazy_max_pages());
1812 * Free a vmap area, caller ensuring that the area has been unmapped
1813 * and flush_cache_vunmap had been called for the correct range
1816 static void free_vmap_area_noflush(struct vmap_area *va)
1818 unsigned long nr_lazy;
1820 spin_lock(&vmap_area_lock);
1821 unlink_va(va, &vmap_area_root);
1822 spin_unlock(&vmap_area_lock);
1824 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1825 PAGE_SHIFT, &vmap_lazy_nr);
1828 * Merge or place it to the purge tree/list.
1830 spin_lock(&purge_vmap_area_lock);
1831 merge_or_add_vmap_area(va,
1832 &purge_vmap_area_root, &purge_vmap_area_list);
1833 spin_unlock(&purge_vmap_area_lock);
1835 /* After this point, we may free va at any time */
1836 if (unlikely(nr_lazy > lazy_max_pages()))
1837 schedule_work(&drain_vmap_work);
1841 * Free and unmap a vmap area
1843 static void free_unmap_vmap_area(struct vmap_area *va)
1845 flush_cache_vunmap(va->va_start, va->va_end);
1846 vunmap_range_noflush(va->va_start, va->va_end);
1847 if (debug_pagealloc_enabled_static())
1848 flush_tlb_kernel_range(va->va_start, va->va_end);
1850 free_vmap_area_noflush(va);
1853 struct vmap_area *find_vmap_area(unsigned long addr)
1855 struct vmap_area *va;
1857 spin_lock(&vmap_area_lock);
1858 va = __find_vmap_area(addr, &vmap_area_root);
1859 spin_unlock(&vmap_area_lock);
1864 /*** Per cpu kva allocator ***/
1867 * vmap space is limited especially on 32 bit architectures. Ensure there is
1868 * room for at least 16 percpu vmap blocks per CPU.
1871 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1872 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1873 * instead (we just need a rough idea)
1875 #if BITS_PER_LONG == 32
1876 #define VMALLOC_SPACE (128UL*1024*1024)
1878 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1881 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1882 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1883 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1884 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1885 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1886 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1887 #define VMAP_BBMAP_BITS \
1888 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1889 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1890 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1892 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1894 struct vmap_block_queue {
1896 struct list_head free;
1901 struct vmap_area *va;
1902 unsigned long free, dirty;
1903 unsigned long dirty_min, dirty_max; /*< dirty range */
1904 struct list_head free_list;
1905 struct rcu_head rcu_head;
1906 struct list_head purge;
1909 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1910 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1913 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1914 * in the free path. Could get rid of this if we change the API to return a
1915 * "cookie" from alloc, to be passed to free. But no big deal yet.
1917 static DEFINE_XARRAY(vmap_blocks);
1920 * We should probably have a fallback mechanism to allocate virtual memory
1921 * out of partially filled vmap blocks. However vmap block sizing should be
1922 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1926 static unsigned long addr_to_vb_idx(unsigned long addr)
1928 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1929 addr /= VMAP_BLOCK_SIZE;
1933 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1937 addr = va_start + (pages_off << PAGE_SHIFT);
1938 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1939 return (void *)addr;
1943 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1944 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1945 * @order: how many 2^order pages should be occupied in newly allocated block
1946 * @gfp_mask: flags for the page level allocator
1948 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1950 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1952 struct vmap_block_queue *vbq;
1953 struct vmap_block *vb;
1954 struct vmap_area *va;
1955 unsigned long vb_idx;
1959 node = numa_node_id();
1961 vb = kmalloc_node(sizeof(struct vmap_block),
1962 gfp_mask & GFP_RECLAIM_MASK, node);
1964 return ERR_PTR(-ENOMEM);
1966 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1967 VMALLOC_START, VMALLOC_END,
1971 return ERR_CAST(va);
1974 vaddr = vmap_block_vaddr(va->va_start, 0);
1975 spin_lock_init(&vb->lock);
1977 /* At least something should be left free */
1978 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1979 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1981 vb->dirty_min = VMAP_BBMAP_BITS;
1983 INIT_LIST_HEAD(&vb->free_list);
1985 vb_idx = addr_to_vb_idx(va->va_start);
1986 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1990 return ERR_PTR(err);
1993 vbq = raw_cpu_ptr(&vmap_block_queue);
1994 spin_lock(&vbq->lock);
1995 list_add_tail_rcu(&vb->free_list, &vbq->free);
1996 spin_unlock(&vbq->lock);
2001 static void free_vmap_block(struct vmap_block *vb)
2003 struct vmap_block *tmp;
2005 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2008 free_vmap_area_noflush(vb->va);
2009 kfree_rcu(vb, rcu_head);
2012 static void purge_fragmented_blocks(int cpu)
2015 struct vmap_block *vb;
2016 struct vmap_block *n_vb;
2017 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2020 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2022 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2025 spin_lock(&vb->lock);
2026 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2027 vb->free = 0; /* prevent further allocs after releasing lock */
2028 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2030 vb->dirty_max = VMAP_BBMAP_BITS;
2031 spin_lock(&vbq->lock);
2032 list_del_rcu(&vb->free_list);
2033 spin_unlock(&vbq->lock);
2034 spin_unlock(&vb->lock);
2035 list_add_tail(&vb->purge, &purge);
2037 spin_unlock(&vb->lock);
2041 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2042 list_del(&vb->purge);
2043 free_vmap_block(vb);
2047 static void purge_fragmented_blocks_allcpus(void)
2051 for_each_possible_cpu(cpu)
2052 purge_fragmented_blocks(cpu);
2055 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2057 struct vmap_block_queue *vbq;
2058 struct vmap_block *vb;
2062 BUG_ON(offset_in_page(size));
2063 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2064 if (WARN_ON(size == 0)) {
2066 * Allocating 0 bytes isn't what caller wants since
2067 * get_order(0) returns funny result. Just warn and terminate
2072 order = get_order(size);
2075 vbq = raw_cpu_ptr(&vmap_block_queue);
2076 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2077 unsigned long pages_off;
2079 spin_lock(&vb->lock);
2080 if (vb->free < (1UL << order)) {
2081 spin_unlock(&vb->lock);
2085 pages_off = VMAP_BBMAP_BITS - vb->free;
2086 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2087 vb->free -= 1UL << order;
2088 if (vb->free == 0) {
2089 spin_lock(&vbq->lock);
2090 list_del_rcu(&vb->free_list);
2091 spin_unlock(&vbq->lock);
2094 spin_unlock(&vb->lock);
2100 /* Allocate new block if nothing was found */
2102 vaddr = new_vmap_block(order, gfp_mask);
2107 static void vb_free(unsigned long addr, unsigned long size)
2109 unsigned long offset;
2111 struct vmap_block *vb;
2113 BUG_ON(offset_in_page(size));
2114 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2116 flush_cache_vunmap(addr, addr + size);
2118 order = get_order(size);
2119 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2120 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2122 vunmap_range_noflush(addr, addr + size);
2124 if (debug_pagealloc_enabled_static())
2125 flush_tlb_kernel_range(addr, addr + size);
2127 spin_lock(&vb->lock);
2129 /* Expand dirty range */
2130 vb->dirty_min = min(vb->dirty_min, offset);
2131 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2133 vb->dirty += 1UL << order;
2134 if (vb->dirty == VMAP_BBMAP_BITS) {
2136 spin_unlock(&vb->lock);
2137 free_vmap_block(vb);
2139 spin_unlock(&vb->lock);
2142 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2146 if (unlikely(!vmap_initialized))
2151 for_each_possible_cpu(cpu) {
2152 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2153 struct vmap_block *vb;
2156 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2157 spin_lock(&vb->lock);
2158 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2159 unsigned long va_start = vb->va->va_start;
2162 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2163 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2165 start = min(s, start);
2170 spin_unlock(&vb->lock);
2175 mutex_lock(&vmap_purge_lock);
2176 purge_fragmented_blocks_allcpus();
2177 if (!__purge_vmap_area_lazy(start, end) && flush)
2178 flush_tlb_kernel_range(start, end);
2179 mutex_unlock(&vmap_purge_lock);
2183 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2185 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2186 * to amortize TLB flushing overheads. What this means is that any page you
2187 * have now, may, in a former life, have been mapped into kernel virtual
2188 * address by the vmap layer and so there might be some CPUs with TLB entries
2189 * still referencing that page (additional to the regular 1:1 kernel mapping).
2191 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2192 * be sure that none of the pages we have control over will have any aliases
2193 * from the vmap layer.
2195 void vm_unmap_aliases(void)
2197 unsigned long start = ULONG_MAX, end = 0;
2200 _vm_unmap_aliases(start, end, flush);
2202 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2205 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2206 * @mem: the pointer returned by vm_map_ram
2207 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2209 void vm_unmap_ram(const void *mem, unsigned int count)
2211 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2212 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2213 struct vmap_area *va;
2217 BUG_ON(addr < VMALLOC_START);
2218 BUG_ON(addr > VMALLOC_END);
2219 BUG_ON(!PAGE_ALIGNED(addr));
2221 kasan_poison_vmalloc(mem, size);
2223 if (likely(count <= VMAP_MAX_ALLOC)) {
2224 debug_check_no_locks_freed(mem, size);
2225 vb_free(addr, size);
2229 va = find_vmap_area(addr);
2231 debug_check_no_locks_freed((void *)va->va_start,
2232 (va->va_end - va->va_start));
2233 free_unmap_vmap_area(va);
2235 EXPORT_SYMBOL(vm_unmap_ram);
2238 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2239 * @pages: an array of pointers to the pages to be mapped
2240 * @count: number of pages
2241 * @node: prefer to allocate data structures on this node
2243 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2244 * faster than vmap so it's good. But if you mix long-life and short-life
2245 * objects with vm_map_ram(), it could consume lots of address space through
2246 * fragmentation (especially on a 32bit machine). You could see failures in
2247 * the end. Please use this function for short-lived objects.
2249 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2251 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2253 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2257 if (likely(count <= VMAP_MAX_ALLOC)) {
2258 mem = vb_alloc(size, GFP_KERNEL);
2261 addr = (unsigned long)mem;
2263 struct vmap_area *va;
2264 va = alloc_vmap_area(size, PAGE_SIZE,
2265 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2269 addr = va->va_start;
2273 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2274 pages, PAGE_SHIFT) < 0) {
2275 vm_unmap_ram(mem, count);
2280 * Mark the pages as accessible, now that they are mapped.
2281 * With hardware tag-based KASAN, marking is skipped for
2282 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2284 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2288 EXPORT_SYMBOL(vm_map_ram);
2290 static struct vm_struct *vmlist __initdata;
2292 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2294 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2295 return vm->page_order;
2301 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2303 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2304 vm->page_order = order;
2311 * vm_area_add_early - add vmap area early during boot
2312 * @vm: vm_struct to add
2314 * This function is used to add fixed kernel vm area to vmlist before
2315 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2316 * should contain proper values and the other fields should be zero.
2318 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2320 void __init vm_area_add_early(struct vm_struct *vm)
2322 struct vm_struct *tmp, **p;
2324 BUG_ON(vmap_initialized);
2325 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2326 if (tmp->addr >= vm->addr) {
2327 BUG_ON(tmp->addr < vm->addr + vm->size);
2330 BUG_ON(tmp->addr + tmp->size > vm->addr);
2337 * vm_area_register_early - register vmap area early during boot
2338 * @vm: vm_struct to register
2339 * @align: requested alignment
2341 * This function is used to register kernel vm area before
2342 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2343 * proper values on entry and other fields should be zero. On return,
2344 * vm->addr contains the allocated address.
2346 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2348 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2350 unsigned long addr = ALIGN(VMALLOC_START, align);
2351 struct vm_struct *cur, **p;
2353 BUG_ON(vmap_initialized);
2355 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2356 if ((unsigned long)cur->addr - addr >= vm->size)
2358 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2361 BUG_ON(addr > VMALLOC_END - vm->size);
2362 vm->addr = (void *)addr;
2365 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2368 static void vmap_init_free_space(void)
2370 unsigned long vmap_start = 1;
2371 const unsigned long vmap_end = ULONG_MAX;
2372 struct vmap_area *busy, *free;
2376 * -|-----|.....|-----|-----|-----|.....|-
2378 * |<--------------------------------->|
2380 list_for_each_entry(busy, &vmap_area_list, list) {
2381 if (busy->va_start - vmap_start > 0) {
2382 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2383 if (!WARN_ON_ONCE(!free)) {
2384 free->va_start = vmap_start;
2385 free->va_end = busy->va_start;
2387 insert_vmap_area_augment(free, NULL,
2388 &free_vmap_area_root,
2389 &free_vmap_area_list);
2393 vmap_start = busy->va_end;
2396 if (vmap_end - vmap_start > 0) {
2397 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2398 if (!WARN_ON_ONCE(!free)) {
2399 free->va_start = vmap_start;
2400 free->va_end = vmap_end;
2402 insert_vmap_area_augment(free, NULL,
2403 &free_vmap_area_root,
2404 &free_vmap_area_list);
2409 void __init vmalloc_init(void)
2411 struct vmap_area *va;
2412 struct vm_struct *tmp;
2416 * Create the cache for vmap_area objects.
2418 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2420 for_each_possible_cpu(i) {
2421 struct vmap_block_queue *vbq;
2422 struct vfree_deferred *p;
2424 vbq = &per_cpu(vmap_block_queue, i);
2425 spin_lock_init(&vbq->lock);
2426 INIT_LIST_HEAD(&vbq->free);
2427 p = &per_cpu(vfree_deferred, i);
2428 init_llist_head(&p->list);
2429 INIT_WORK(&p->wq, free_work);
2432 /* Import existing vmlist entries. */
2433 for (tmp = vmlist; tmp; tmp = tmp->next) {
2434 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2435 if (WARN_ON_ONCE(!va))
2438 va->va_start = (unsigned long)tmp->addr;
2439 va->va_end = va->va_start + tmp->size;
2441 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2445 * Now we can initialize a free vmap space.
2447 vmap_init_free_space();
2448 vmap_initialized = true;
2451 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2452 struct vmap_area *va, unsigned long flags, const void *caller)
2455 vm->addr = (void *)va->va_start;
2456 vm->size = va->va_end - va->va_start;
2457 vm->caller = caller;
2461 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2462 unsigned long flags, const void *caller)
2464 spin_lock(&vmap_area_lock);
2465 setup_vmalloc_vm_locked(vm, va, flags, caller);
2466 spin_unlock(&vmap_area_lock);
2469 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2472 * Before removing VM_UNINITIALIZED,
2473 * we should make sure that vm has proper values.
2474 * Pair with smp_rmb() in show_numa_info().
2477 vm->flags &= ~VM_UNINITIALIZED;
2480 static struct vm_struct *__get_vm_area_node(unsigned long size,
2481 unsigned long align, unsigned long shift, unsigned long flags,
2482 unsigned long start, unsigned long end, int node,
2483 gfp_t gfp_mask, const void *caller)
2485 struct vmap_area *va;
2486 struct vm_struct *area;
2487 unsigned long requested_size = size;
2489 BUG_ON(in_interrupt());
2490 size = ALIGN(size, 1ul << shift);
2491 if (unlikely(!size))
2494 if (flags & VM_IOREMAP)
2495 align = 1ul << clamp_t(int, get_count_order_long(size),
2496 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2498 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2499 if (unlikely(!area))
2502 if (!(flags & VM_NO_GUARD))
2505 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2511 setup_vmalloc_vm(area, va, flags, caller);
2514 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2515 * best-effort approach, as they can be mapped outside of vmalloc code.
2516 * For VM_ALLOC mappings, the pages are marked as accessible after
2517 * getting mapped in __vmalloc_node_range().
2518 * With hardware tag-based KASAN, marking is skipped for
2519 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2521 if (!(flags & VM_ALLOC))
2522 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2523 KASAN_VMALLOC_PROT_NORMAL);
2528 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2529 unsigned long start, unsigned long end,
2532 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2533 NUMA_NO_NODE, GFP_KERNEL, caller);
2537 * get_vm_area - reserve a contiguous kernel virtual area
2538 * @size: size of the area
2539 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2541 * Search an area of @size in the kernel virtual mapping area,
2542 * and reserved it for out purposes. Returns the area descriptor
2543 * on success or %NULL on failure.
2545 * Return: the area descriptor on success or %NULL on failure.
2547 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2549 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2550 VMALLOC_START, VMALLOC_END,
2551 NUMA_NO_NODE, GFP_KERNEL,
2552 __builtin_return_address(0));
2555 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2558 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2559 VMALLOC_START, VMALLOC_END,
2560 NUMA_NO_NODE, GFP_KERNEL, caller);
2564 * find_vm_area - find a continuous kernel virtual area
2565 * @addr: base address
2567 * Search for the kernel VM area starting at @addr, and return it.
2568 * It is up to the caller to do all required locking to keep the returned
2571 * Return: the area descriptor on success or %NULL on failure.
2573 struct vm_struct *find_vm_area(const void *addr)
2575 struct vmap_area *va;
2577 va = find_vmap_area((unsigned long)addr);
2585 * remove_vm_area - find and remove a continuous kernel virtual area
2586 * @addr: base address
2588 * Search for the kernel VM area starting at @addr, and remove it.
2589 * This function returns the found VM area, but using it is NOT safe
2590 * on SMP machines, except for its size or flags.
2592 * Return: the area descriptor on success or %NULL on failure.
2594 struct vm_struct *remove_vm_area(const void *addr)
2596 struct vmap_area *va;
2600 spin_lock(&vmap_area_lock);
2601 va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2603 struct vm_struct *vm = va->vm;
2606 spin_unlock(&vmap_area_lock);
2608 kasan_free_module_shadow(vm);
2609 free_unmap_vmap_area(va);
2614 spin_unlock(&vmap_area_lock);
2618 static inline void set_area_direct_map(const struct vm_struct *area,
2619 int (*set_direct_map)(struct page *page))
2623 /* HUGE_VMALLOC passes small pages to set_direct_map */
2624 for (i = 0; i < area->nr_pages; i++)
2625 if (page_address(area->pages[i]))
2626 set_direct_map(area->pages[i]);
2629 /* Handle removing and resetting vm mappings related to the vm_struct. */
2630 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2632 unsigned long start = ULONG_MAX, end = 0;
2633 unsigned int page_order = vm_area_page_order(area);
2634 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2638 remove_vm_area(area->addr);
2640 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2645 * If not deallocating pages, just do the flush of the VM area and
2648 if (!deallocate_pages) {
2654 * If execution gets here, flush the vm mapping and reset the direct
2655 * map. Find the start and end range of the direct mappings to make sure
2656 * the vm_unmap_aliases() flush includes the direct map.
2658 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2659 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2661 unsigned long page_size;
2663 page_size = PAGE_SIZE << page_order;
2664 start = min(addr, start);
2665 end = max(addr + page_size, end);
2671 * Set direct map to something invalid so that it won't be cached if
2672 * there are any accesses after the TLB flush, then flush the TLB and
2673 * reset the direct map permissions to the default.
2675 set_area_direct_map(area, set_direct_map_invalid_noflush);
2676 _vm_unmap_aliases(start, end, flush_dmap);
2677 set_area_direct_map(area, set_direct_map_default_noflush);
2680 static void __vunmap(const void *addr, int deallocate_pages)
2682 struct vm_struct *area;
2687 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2691 area = find_vm_area(addr);
2692 if (unlikely(!area)) {
2693 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2698 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2699 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2701 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2703 vm_remove_mappings(area, deallocate_pages);
2705 if (deallocate_pages) {
2708 for (i = 0; i < area->nr_pages; i++) {
2709 struct page *page = area->pages[i];
2712 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2714 * High-order allocs for huge vmallocs are split, so
2715 * can be freed as an array of order-0 allocations
2717 __free_pages(page, 0);
2720 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2722 kvfree(area->pages);
2728 static inline void __vfree_deferred(const void *addr)
2731 * Use raw_cpu_ptr() because this can be called from preemptible
2732 * context. Preemption is absolutely fine here, because the llist_add()
2733 * implementation is lockless, so it works even if we are adding to
2734 * another cpu's list. schedule_work() should be fine with this too.
2736 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2738 if (llist_add((struct llist_node *)addr, &p->list))
2739 schedule_work(&p->wq);
2743 * vfree_atomic - release memory allocated by vmalloc()
2744 * @addr: memory base address
2746 * This one is just like vfree() but can be called in any atomic context
2749 void vfree_atomic(const void *addr)
2753 kmemleak_free(addr);
2757 __vfree_deferred(addr);
2760 static void __vfree(const void *addr)
2762 if (unlikely(in_interrupt()))
2763 __vfree_deferred(addr);
2769 * vfree - Release memory allocated by vmalloc()
2770 * @addr: Memory base address
2772 * Free the virtually continuous memory area starting at @addr, as obtained
2773 * from one of the vmalloc() family of APIs. This will usually also free the
2774 * physical memory underlying the virtual allocation, but that memory is
2775 * reference counted, so it will not be freed until the last user goes away.
2777 * If @addr is NULL, no operation is performed.
2780 * May sleep if called *not* from interrupt context.
2781 * Must not be called in NMI context (strictly speaking, it could be
2782 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2783 * conventions for vfree() arch-dependent would be a really bad idea).
2785 void vfree(const void *addr)
2789 kmemleak_free(addr);
2791 might_sleep_if(!in_interrupt());
2798 EXPORT_SYMBOL(vfree);
2801 * vunmap - release virtual mapping obtained by vmap()
2802 * @addr: memory base address
2804 * Free the virtually contiguous memory area starting at @addr,
2805 * which was created from the page array passed to vmap().
2807 * Must not be called in interrupt context.
2809 void vunmap(const void *addr)
2811 BUG_ON(in_interrupt());
2816 EXPORT_SYMBOL(vunmap);
2819 * vmap - map an array of pages into virtually contiguous space
2820 * @pages: array of page pointers
2821 * @count: number of pages to map
2822 * @flags: vm_area->flags
2823 * @prot: page protection for the mapping
2825 * Maps @count pages from @pages into contiguous kernel virtual space.
2826 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2827 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2828 * are transferred from the caller to vmap(), and will be freed / dropped when
2829 * vfree() is called on the return value.
2831 * Return: the address of the area or %NULL on failure
2833 void *vmap(struct page **pages, unsigned int count,
2834 unsigned long flags, pgprot_t prot)
2836 struct vm_struct *area;
2838 unsigned long size; /* In bytes */
2843 * Your top guard is someone else's bottom guard. Not having a top
2844 * guard compromises someone else's mappings too.
2846 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2847 flags &= ~VM_NO_GUARD;
2849 if (count > totalram_pages())
2852 size = (unsigned long)count << PAGE_SHIFT;
2853 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2857 addr = (unsigned long)area->addr;
2858 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2859 pages, PAGE_SHIFT) < 0) {
2864 if (flags & VM_MAP_PUT_PAGES) {
2865 area->pages = pages;
2866 area->nr_pages = count;
2870 EXPORT_SYMBOL(vmap);
2872 #ifdef CONFIG_VMAP_PFN
2873 struct vmap_pfn_data {
2874 unsigned long *pfns;
2879 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2881 struct vmap_pfn_data *data = private;
2883 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2885 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2890 * vmap_pfn - map an array of PFNs into virtually contiguous space
2891 * @pfns: array of PFNs
2892 * @count: number of pages to map
2893 * @prot: page protection for the mapping
2895 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2896 * the start address of the mapping.
2898 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2900 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2901 struct vm_struct *area;
2903 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2904 __builtin_return_address(0));
2907 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2908 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2913 flush_cache_vmap((unsigned long)area->addr,
2914 (unsigned long)area->addr + count * PAGE_SIZE);
2918 EXPORT_SYMBOL_GPL(vmap_pfn);
2919 #endif /* CONFIG_VMAP_PFN */
2921 static inline unsigned int
2922 vm_area_alloc_pages(gfp_t gfp, int nid,
2923 unsigned int order, unsigned int nr_pages, struct page **pages)
2925 unsigned int nr_allocated = 0;
2930 * For order-0 pages we make use of bulk allocator, if
2931 * the page array is partly or not at all populated due
2932 * to fails, fallback to a single page allocator that is
2936 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2938 while (nr_allocated < nr_pages) {
2939 unsigned int nr, nr_pages_request;
2942 * A maximum allowed request is hard-coded and is 100
2943 * pages per call. That is done in order to prevent a
2944 * long preemption off scenario in the bulk-allocator
2945 * so the range is [1:100].
2947 nr_pages_request = min(100U, nr_pages - nr_allocated);
2949 /* memory allocation should consider mempolicy, we can't
2950 * wrongly use nearest node when nid == NUMA_NO_NODE,
2951 * otherwise memory may be allocated in only one node,
2952 * but mempolicy wants to alloc memory by interleaving.
2954 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2955 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2957 pages + nr_allocated);
2960 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2962 pages + nr_allocated);
2968 * If zero or pages were obtained partly,
2969 * fallback to a single page allocator.
2971 if (nr != nr_pages_request)
2976 /* High-order pages or fallback path if "bulk" fails. */
2978 while (nr_allocated < nr_pages) {
2979 if (fatal_signal_pending(current))
2982 if (nid == NUMA_NO_NODE)
2983 page = alloc_pages(gfp, order);
2985 page = alloc_pages_node(nid, gfp, order);
2986 if (unlikely(!page))
2989 * Higher order allocations must be able to be treated as
2990 * indepdenent small pages by callers (as they can with
2991 * small-page vmallocs). Some drivers do their own refcounting
2992 * on vmalloc_to_page() pages, some use page->mapping,
2996 split_page(page, order);
2999 * Careful, we allocate and map page-order pages, but
3000 * tracking is done per PAGE_SIZE page so as to keep the
3001 * vm_struct APIs independent of the physical/mapped size.
3003 for (i = 0; i < (1U << order); i++)
3004 pages[nr_allocated + i] = page + i;
3007 nr_allocated += 1U << order;
3010 return nr_allocated;
3013 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3014 pgprot_t prot, unsigned int page_shift,
3017 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3018 bool nofail = gfp_mask & __GFP_NOFAIL;
3019 unsigned long addr = (unsigned long)area->addr;
3020 unsigned long size = get_vm_area_size(area);
3021 unsigned long array_size;
3022 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3023 unsigned int page_order;
3027 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3028 gfp_mask |= __GFP_NOWARN;
3029 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3030 gfp_mask |= __GFP_HIGHMEM;
3032 /* Please note that the recursion is strictly bounded. */
3033 if (array_size > PAGE_SIZE) {
3034 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3037 area->pages = kmalloc_node(array_size, nested_gfp, node);
3041 warn_alloc(gfp_mask, NULL,
3042 "vmalloc error: size %lu, failed to allocated page array size %lu",
3043 nr_small_pages * PAGE_SIZE, array_size);
3048 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3049 page_order = vm_area_page_order(area);
3051 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3052 node, page_order, nr_small_pages, area->pages);
3054 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3055 if (gfp_mask & __GFP_ACCOUNT) {
3058 for (i = 0; i < area->nr_pages; i++)
3059 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3063 * If not enough pages were obtained to accomplish an
3064 * allocation request, free them via __vfree() if any.
3066 if (area->nr_pages != nr_small_pages) {
3067 /* vm_area_alloc_pages() can also fail due to a fatal signal */
3068 if (!fatal_signal_pending(current))
3069 warn_alloc(gfp_mask, NULL,
3070 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3071 area->nr_pages * PAGE_SIZE, page_order);
3076 * page tables allocations ignore external gfp mask, enforce it
3079 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3080 flags = memalloc_nofs_save();
3081 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3082 flags = memalloc_noio_save();
3085 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3087 if (nofail && (ret < 0))
3088 schedule_timeout_uninterruptible(1);
3089 } while (nofail && (ret < 0));
3091 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3092 memalloc_nofs_restore(flags);
3093 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3094 memalloc_noio_restore(flags);
3097 warn_alloc(gfp_mask, NULL,
3098 "vmalloc error: size %lu, failed to map pages",
3099 area->nr_pages * PAGE_SIZE);
3106 __vfree(area->addr);
3111 * __vmalloc_node_range - allocate virtually contiguous memory
3112 * @size: allocation size
3113 * @align: desired alignment
3114 * @start: vm area range start
3115 * @end: vm area range end
3116 * @gfp_mask: flags for the page level allocator
3117 * @prot: protection mask for the allocated pages
3118 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3119 * @node: node to use for allocation or NUMA_NO_NODE
3120 * @caller: caller's return address
3122 * Allocate enough pages to cover @size from the page level
3123 * allocator with @gfp_mask flags. Please note that the full set of gfp
3124 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3126 * Zone modifiers are not supported. From the reclaim modifiers
3127 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3128 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3129 * __GFP_RETRY_MAYFAIL are not supported).
3131 * __GFP_NOWARN can be used to suppress failures messages.
3133 * Map them into contiguous kernel virtual space, using a pagetable
3134 * protection of @prot.
3136 * Return: the address of the area or %NULL on failure
3138 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3139 unsigned long start, unsigned long end, gfp_t gfp_mask,
3140 pgprot_t prot, unsigned long vm_flags, int node,
3143 struct vm_struct *area;
3145 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3146 unsigned long real_size = size;
3147 unsigned long real_align = align;
3148 unsigned int shift = PAGE_SHIFT;
3150 if (WARN_ON_ONCE(!size))
3153 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3154 warn_alloc(gfp_mask, NULL,
3155 "vmalloc error: size %lu, exceeds total pages",
3160 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3161 unsigned long size_per_node;
3164 * Try huge pages. Only try for PAGE_KERNEL allocations,
3165 * others like modules don't yet expect huge pages in
3166 * their allocations due to apply_to_page_range not
3170 size_per_node = size;
3171 if (node == NUMA_NO_NODE)
3172 size_per_node /= num_online_nodes();
3173 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3176 shift = arch_vmap_pte_supported_shift(size_per_node);
3178 align = max(real_align, 1UL << shift);
3179 size = ALIGN(real_size, 1UL << shift);
3183 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3184 VM_UNINITIALIZED | vm_flags, start, end, node,
3187 bool nofail = gfp_mask & __GFP_NOFAIL;
3188 warn_alloc(gfp_mask, NULL,
3189 "vmalloc error: size %lu, vm_struct allocation failed%s",
3190 real_size, (nofail) ? ". Retrying." : "");
3192 schedule_timeout_uninterruptible(1);
3199 * Prepare arguments for __vmalloc_area_node() and
3200 * kasan_unpoison_vmalloc().
3202 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3203 if (kasan_hw_tags_enabled()) {
3205 * Modify protection bits to allow tagging.
3206 * This must be done before mapping.
3208 prot = arch_vmap_pgprot_tagged(prot);
3211 * Skip page_alloc poisoning and zeroing for physical
3212 * pages backing VM_ALLOC mapping. Memory is instead
3213 * poisoned and zeroed by kasan_unpoison_vmalloc().
3215 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3218 /* Take note that the mapping is PAGE_KERNEL. */
3219 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3222 /* Allocate physical pages and map them into vmalloc space. */
3223 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3228 * Mark the pages as accessible, now that they are mapped.
3229 * The condition for setting KASAN_VMALLOC_INIT should complement the
3230 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3231 * to make sure that memory is initialized under the same conditions.
3232 * Tag-based KASAN modes only assign tags to normal non-executable
3233 * allocations, see __kasan_unpoison_vmalloc().
3235 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3236 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3237 (gfp_mask & __GFP_SKIP_ZERO))
3238 kasan_flags |= KASAN_VMALLOC_INIT;
3239 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3240 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3243 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3244 * flag. It means that vm_struct is not fully initialized.
3245 * Now, it is fully initialized, so remove this flag here.
3247 clear_vm_uninitialized_flag(area);
3249 size = PAGE_ALIGN(size);
3250 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3251 kmemleak_vmalloc(area, size, gfp_mask);
3256 if (shift > PAGE_SHIFT) {
3267 * __vmalloc_node - allocate virtually contiguous memory
3268 * @size: allocation size
3269 * @align: desired alignment
3270 * @gfp_mask: flags for the page level allocator
3271 * @node: node to use for allocation or NUMA_NO_NODE
3272 * @caller: caller's return address
3274 * Allocate enough pages to cover @size from the page level allocator with
3275 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3277 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3278 * and __GFP_NOFAIL are not supported
3280 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3283 * Return: pointer to the allocated memory or %NULL on error
3285 void *__vmalloc_node(unsigned long size, unsigned long align,
3286 gfp_t gfp_mask, int node, const void *caller)
3288 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3289 gfp_mask, PAGE_KERNEL, 0, node, caller);
3292 * This is only for performance analysis of vmalloc and stress purpose.
3293 * It is required by vmalloc test module, therefore do not use it other
3296 #ifdef CONFIG_TEST_VMALLOC_MODULE
3297 EXPORT_SYMBOL_GPL(__vmalloc_node);
3300 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3302 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3303 __builtin_return_address(0));
3305 EXPORT_SYMBOL(__vmalloc);
3308 * vmalloc - allocate virtually contiguous memory
3309 * @size: allocation size
3311 * Allocate enough pages to cover @size from the page level
3312 * allocator and map them into contiguous kernel virtual space.
3314 * For tight control over page level allocator and protection flags
3315 * use __vmalloc() instead.
3317 * Return: pointer to the allocated memory or %NULL on error
3319 void *vmalloc(unsigned long size)
3321 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3322 __builtin_return_address(0));
3324 EXPORT_SYMBOL(vmalloc);
3327 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3328 * @size: allocation size
3329 * @gfp_mask: flags for the page level allocator
3331 * Allocate enough pages to cover @size from the page level
3332 * allocator and map them into contiguous kernel virtual space.
3333 * If @size is greater than or equal to PMD_SIZE, allow using
3334 * huge pages for the memory
3336 * Return: pointer to the allocated memory or %NULL on error
3338 void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3340 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3341 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3342 NUMA_NO_NODE, __builtin_return_address(0));
3344 EXPORT_SYMBOL_GPL(vmalloc_huge);
3347 * vzalloc - allocate virtually contiguous memory with zero fill
3348 * @size: allocation size
3350 * Allocate enough pages to cover @size from the page level
3351 * allocator and map them into contiguous kernel virtual space.
3352 * The memory allocated is set to zero.
3354 * For tight control over page level allocator and protection flags
3355 * use __vmalloc() instead.
3357 * Return: pointer to the allocated memory or %NULL on error
3359 void *vzalloc(unsigned long size)
3361 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3362 __builtin_return_address(0));
3364 EXPORT_SYMBOL(vzalloc);
3367 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3368 * @size: allocation size
3370 * The resulting memory area is zeroed so it can be mapped to userspace
3371 * without leaking data.
3373 * Return: pointer to the allocated memory or %NULL on error
3375 void *vmalloc_user(unsigned long size)
3377 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3378 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3379 VM_USERMAP, NUMA_NO_NODE,
3380 __builtin_return_address(0));
3382 EXPORT_SYMBOL(vmalloc_user);
3385 * vmalloc_node - allocate memory on a specific node
3386 * @size: allocation size
3389 * Allocate enough pages to cover @size from the page level
3390 * allocator and map them into contiguous kernel virtual space.
3392 * For tight control over page level allocator and protection flags
3393 * use __vmalloc() instead.
3395 * Return: pointer to the allocated memory or %NULL on error
3397 void *vmalloc_node(unsigned long size, int node)
3399 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3400 __builtin_return_address(0));
3402 EXPORT_SYMBOL(vmalloc_node);
3405 * vzalloc_node - allocate memory on a specific node with zero fill
3406 * @size: allocation size
3409 * Allocate enough pages to cover @size from the page level
3410 * allocator and map them into contiguous kernel virtual space.
3411 * The memory allocated is set to zero.
3413 * Return: pointer to the allocated memory or %NULL on error
3415 void *vzalloc_node(unsigned long size, int node)
3417 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3418 __builtin_return_address(0));
3420 EXPORT_SYMBOL(vzalloc_node);
3422 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3423 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3424 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3425 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3428 * 64b systems should always have either DMA or DMA32 zones. For others
3429 * GFP_DMA32 should do the right thing and use the normal zone.
3431 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3435 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3436 * @size: allocation size
3438 * Allocate enough 32bit PA addressable pages to cover @size from the
3439 * page level allocator and map them into contiguous kernel virtual space.
3441 * Return: pointer to the allocated memory or %NULL on error
3443 void *vmalloc_32(unsigned long size)
3445 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3446 __builtin_return_address(0));
3448 EXPORT_SYMBOL(vmalloc_32);
3451 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3452 * @size: allocation size
3454 * The resulting memory area is 32bit addressable and zeroed so it can be
3455 * mapped to userspace without leaking data.
3457 * Return: pointer to the allocated memory or %NULL on error
3459 void *vmalloc_32_user(unsigned long size)
3461 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3462 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3463 VM_USERMAP, NUMA_NO_NODE,
3464 __builtin_return_address(0));
3466 EXPORT_SYMBOL(vmalloc_32_user);
3469 * small helper routine , copy contents to buf from addr.
3470 * If the page is not present, fill zero.
3473 static int aligned_vread(char *buf, char *addr, unsigned long count)
3479 unsigned long offset, length;
3481 offset = offset_in_page(addr);
3482 length = PAGE_SIZE - offset;
3485 p = vmalloc_to_page(addr);
3487 * To do safe access to this _mapped_ area, we need
3488 * lock. But adding lock here means that we need to add
3489 * overhead of vmalloc()/vfree() calls for this _debug_
3490 * interface, rarely used. Instead of that, we'll use
3491 * kmap() and get small overhead in this access function.
3494 /* We can expect USER0 is not used -- see vread() */
3495 void *map = kmap_atomic(p);
3496 memcpy(buf, map + offset, length);
3499 memset(buf, 0, length);
3510 * vread() - read vmalloc area in a safe way.
3511 * @buf: buffer for reading data
3512 * @addr: vm address.
3513 * @count: number of bytes to be read.
3515 * This function checks that addr is a valid vmalloc'ed area, and
3516 * copy data from that area to a given buffer. If the given memory range
3517 * of [addr...addr+count) includes some valid address, data is copied to
3518 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3519 * IOREMAP area is treated as memory hole and no copy is done.
3521 * If [addr...addr+count) doesn't includes any intersects with alive
3522 * vm_struct area, returns 0. @buf should be kernel's buffer.
3524 * Note: In usual ops, vread() is never necessary because the caller
3525 * should know vmalloc() area is valid and can use memcpy().
3526 * This is for routines which have to access vmalloc area without
3527 * any information, as /proc/kcore.
3529 * Return: number of bytes for which addr and buf should be increased
3530 * (same number as @count) or %0 if [addr...addr+count) doesn't
3531 * include any intersection with valid vmalloc area
3533 long vread(char *buf, char *addr, unsigned long count)
3535 struct vmap_area *va;
3536 struct vm_struct *vm;
3537 char *vaddr, *buf_start = buf;
3538 unsigned long buflen = count;
3541 addr = kasan_reset_tag(addr);
3543 /* Don't allow overflow */
3544 if ((unsigned long) addr + count < count)
3545 count = -(unsigned long) addr;
3547 spin_lock(&vmap_area_lock);
3548 va = find_vmap_area_exceed_addr((unsigned long)addr);
3552 /* no intersects with alive vmap_area */
3553 if ((unsigned long)addr + count <= va->va_start)
3556 list_for_each_entry_from(va, &vmap_area_list, list) {
3564 vaddr = (char *) vm->addr;
3565 if (addr >= vaddr + get_vm_area_size(vm))
3567 while (addr < vaddr) {
3575 n = vaddr + get_vm_area_size(vm) - addr;
3578 if (!(vm->flags & VM_IOREMAP))
3579 aligned_vread(buf, addr, n);
3580 else /* IOREMAP area is treated as memory hole */
3587 spin_unlock(&vmap_area_lock);
3589 if (buf == buf_start)
3591 /* zero-fill memory holes */
3592 if (buf != buf_start + buflen)
3593 memset(buf, 0, buflen - (buf - buf_start));
3599 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3600 * @vma: vma to cover
3601 * @uaddr: target user address to start at
3602 * @kaddr: virtual address of vmalloc kernel memory
3603 * @pgoff: offset from @kaddr to start at
3604 * @size: size of map area
3606 * Returns: 0 for success, -Exxx on failure
3608 * This function checks that @kaddr is a valid vmalloc'ed area,
3609 * and that it is big enough to cover the range starting at
3610 * @uaddr in @vma. Will return failure if that criteria isn't
3613 * Similar to remap_pfn_range() (see mm/memory.c)
3615 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3616 void *kaddr, unsigned long pgoff,
3619 struct vm_struct *area;
3621 unsigned long end_index;
3623 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3626 size = PAGE_ALIGN(size);
3628 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3631 area = find_vm_area(kaddr);
3635 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3638 if (check_add_overflow(size, off, &end_index) ||
3639 end_index > get_vm_area_size(area))
3644 struct page *page = vmalloc_to_page(kaddr);
3647 ret = vm_insert_page(vma, uaddr, page);
3656 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3662 * remap_vmalloc_range - map vmalloc pages to userspace
3663 * @vma: vma to cover (map full range of vma)
3664 * @addr: vmalloc memory
3665 * @pgoff: number of pages into addr before first page to map
3667 * Returns: 0 for success, -Exxx on failure
3669 * This function checks that addr is a valid vmalloc'ed area, and
3670 * that it is big enough to cover the vma. Will return failure if
3671 * that criteria isn't met.
3673 * Similar to remap_pfn_range() (see mm/memory.c)
3675 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3676 unsigned long pgoff)
3678 return remap_vmalloc_range_partial(vma, vma->vm_start,
3680 vma->vm_end - vma->vm_start);
3682 EXPORT_SYMBOL(remap_vmalloc_range);
3684 void free_vm_area(struct vm_struct *area)
3686 struct vm_struct *ret;
3687 ret = remove_vm_area(area->addr);
3688 BUG_ON(ret != area);
3691 EXPORT_SYMBOL_GPL(free_vm_area);
3694 static struct vmap_area *node_to_va(struct rb_node *n)
3696 return rb_entry_safe(n, struct vmap_area, rb_node);
3700 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3701 * @addr: target address
3703 * Returns: vmap_area if it is found. If there is no such area
3704 * the first highest(reverse order) vmap_area is returned
3705 * i.e. va->va_start < addr && va->va_end < addr or NULL
3706 * if there are no any areas before @addr.
3708 static struct vmap_area *
3709 pvm_find_va_enclose_addr(unsigned long addr)
3711 struct vmap_area *va, *tmp;
3714 n = free_vmap_area_root.rb_node;
3718 tmp = rb_entry(n, struct vmap_area, rb_node);
3719 if (tmp->va_start <= addr) {
3721 if (tmp->va_end >= addr)
3734 * pvm_determine_end_from_reverse - find the highest aligned address
3735 * of free block below VMALLOC_END
3737 * in - the VA we start the search(reverse order);
3738 * out - the VA with the highest aligned end address.
3739 * @align: alignment for required highest address
3741 * Returns: determined end address within vmap_area
3743 static unsigned long
3744 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3746 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3750 list_for_each_entry_from_reverse((*va),
3751 &free_vmap_area_list, list) {
3752 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3753 if ((*va)->va_start < addr)
3762 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3763 * @offsets: array containing offset of each area
3764 * @sizes: array containing size of each area
3765 * @nr_vms: the number of areas to allocate
3766 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3768 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3769 * vm_structs on success, %NULL on failure
3771 * Percpu allocator wants to use congruent vm areas so that it can
3772 * maintain the offsets among percpu areas. This function allocates
3773 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3774 * be scattered pretty far, distance between two areas easily going up
3775 * to gigabytes. To avoid interacting with regular vmallocs, these
3776 * areas are allocated from top.
3778 * Despite its complicated look, this allocator is rather simple. It
3779 * does everything top-down and scans free blocks from the end looking
3780 * for matching base. While scanning, if any of the areas do not fit the
3781 * base address is pulled down to fit the area. Scanning is repeated till
3782 * all the areas fit and then all necessary data structures are inserted
3783 * and the result is returned.
3785 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3786 const size_t *sizes, int nr_vms,
3789 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3790 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3791 struct vmap_area **vas, *va;
3792 struct vm_struct **vms;
3793 int area, area2, last_area, term_area;
3794 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3795 bool purged = false;
3797 /* verify parameters and allocate data structures */
3798 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3799 for (last_area = 0, area = 0; area < nr_vms; area++) {
3800 start = offsets[area];
3801 end = start + sizes[area];
3803 /* is everything aligned properly? */
3804 BUG_ON(!IS_ALIGNED(offsets[area], align));
3805 BUG_ON(!IS_ALIGNED(sizes[area], align));
3807 /* detect the area with the highest address */
3808 if (start > offsets[last_area])
3811 for (area2 = area + 1; area2 < nr_vms; area2++) {
3812 unsigned long start2 = offsets[area2];
3813 unsigned long end2 = start2 + sizes[area2];
3815 BUG_ON(start2 < end && start < end2);
3818 last_end = offsets[last_area] + sizes[last_area];
3820 if (vmalloc_end - vmalloc_start < last_end) {
3825 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3826 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3830 for (area = 0; area < nr_vms; area++) {
3831 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3832 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3833 if (!vas[area] || !vms[area])
3837 spin_lock(&free_vmap_area_lock);
3839 /* start scanning - we scan from the top, begin with the last area */
3840 area = term_area = last_area;
3841 start = offsets[area];
3842 end = start + sizes[area];
3844 va = pvm_find_va_enclose_addr(vmalloc_end);
3845 base = pvm_determine_end_from_reverse(&va, align) - end;
3849 * base might have underflowed, add last_end before
3852 if (base + last_end < vmalloc_start + last_end)
3856 * Fitting base has not been found.
3862 * If required width exceeds current VA block, move
3863 * base downwards and then recheck.
3865 if (base + end > va->va_end) {
3866 base = pvm_determine_end_from_reverse(&va, align) - end;
3872 * If this VA does not fit, move base downwards and recheck.
3874 if (base + start < va->va_start) {
3875 va = node_to_va(rb_prev(&va->rb_node));
3876 base = pvm_determine_end_from_reverse(&va, align) - end;
3882 * This area fits, move on to the previous one. If
3883 * the previous one is the terminal one, we're done.
3885 area = (area + nr_vms - 1) % nr_vms;
3886 if (area == term_area)
3889 start = offsets[area];
3890 end = start + sizes[area];
3891 va = pvm_find_va_enclose_addr(base + end);
3894 /* we've found a fitting base, insert all va's */
3895 for (area = 0; area < nr_vms; area++) {
3898 start = base + offsets[area];
3901 va = pvm_find_va_enclose_addr(start);
3902 if (WARN_ON_ONCE(va == NULL))
3903 /* It is a BUG(), but trigger recovery instead. */
3906 ret = adjust_va_to_fit_type(&free_vmap_area_root,
3907 &free_vmap_area_list,
3909 if (WARN_ON_ONCE(unlikely(ret)))
3910 /* It is a BUG(), but trigger recovery instead. */
3913 /* Allocated area. */
3915 va->va_start = start;
3916 va->va_end = start + size;
3919 spin_unlock(&free_vmap_area_lock);
3921 /* populate the kasan shadow space */
3922 for (area = 0; area < nr_vms; area++) {
3923 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3924 goto err_free_shadow;
3927 /* insert all vm's */
3928 spin_lock(&vmap_area_lock);
3929 for (area = 0; area < nr_vms; area++) {
3930 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3932 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3935 spin_unlock(&vmap_area_lock);
3938 * Mark allocated areas as accessible. Do it now as a best-effort
3939 * approach, as they can be mapped outside of vmalloc code.
3940 * With hardware tag-based KASAN, marking is skipped for
3941 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3943 for (area = 0; area < nr_vms; area++)
3944 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3945 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3952 * Remove previously allocated areas. There is no
3953 * need in removing these areas from the busy tree,
3954 * because they are inserted only on the final step
3955 * and when pcpu_get_vm_areas() is success.
3958 orig_start = vas[area]->va_start;
3959 orig_end = vas[area]->va_end;
3960 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3961 &free_vmap_area_list);
3963 kasan_release_vmalloc(orig_start, orig_end,
3964 va->va_start, va->va_end);
3969 spin_unlock(&free_vmap_area_lock);
3971 purge_vmap_area_lazy();
3974 /* Before "retry", check if we recover. */
3975 for (area = 0; area < nr_vms; area++) {
3979 vas[area] = kmem_cache_zalloc(
3980 vmap_area_cachep, GFP_KERNEL);
3989 for (area = 0; area < nr_vms; area++) {
3991 kmem_cache_free(vmap_area_cachep, vas[area]);
4001 spin_lock(&free_vmap_area_lock);
4003 * We release all the vmalloc shadows, even the ones for regions that
4004 * hadn't been successfully added. This relies on kasan_release_vmalloc
4005 * being able to tolerate this case.
4007 for (area = 0; area < nr_vms; area++) {
4008 orig_start = vas[area]->va_start;
4009 orig_end = vas[area]->va_end;
4010 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4011 &free_vmap_area_list);
4013 kasan_release_vmalloc(orig_start, orig_end,
4014 va->va_start, va->va_end);
4018 spin_unlock(&free_vmap_area_lock);
4025 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4026 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4027 * @nr_vms: the number of allocated areas
4029 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4031 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4035 for (i = 0; i < nr_vms; i++)
4036 free_vm_area(vms[i]);
4039 #endif /* CONFIG_SMP */
4041 #ifdef CONFIG_PRINTK
4042 bool vmalloc_dump_obj(void *object)
4044 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4046 struct vm_struct *vm;
4047 struct vmap_area *va;
4049 unsigned int nr_pages;
4051 if (!spin_trylock(&vmap_area_lock))
4053 va = __find_vmap_area((unsigned long)objp, &vmap_area_root);
4055 spin_unlock(&vmap_area_lock);
4061 spin_unlock(&vmap_area_lock);
4064 addr = (unsigned long)vm->addr;
4065 caller = vm->caller;
4066 nr_pages = vm->nr_pages;
4067 spin_unlock(&vmap_area_lock);
4068 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4069 nr_pages, addr, caller);
4074 #ifdef CONFIG_PROC_FS
4075 static void *s_start(struct seq_file *m, loff_t *pos)
4076 __acquires(&vmap_purge_lock)
4077 __acquires(&vmap_area_lock)
4079 mutex_lock(&vmap_purge_lock);
4080 spin_lock(&vmap_area_lock);
4082 return seq_list_start(&vmap_area_list, *pos);
4085 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4087 return seq_list_next(p, &vmap_area_list, pos);
4090 static void s_stop(struct seq_file *m, void *p)
4091 __releases(&vmap_area_lock)
4092 __releases(&vmap_purge_lock)
4094 spin_unlock(&vmap_area_lock);
4095 mutex_unlock(&vmap_purge_lock);
4098 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4100 if (IS_ENABLED(CONFIG_NUMA)) {
4101 unsigned int nr, *counters = m->private;
4102 unsigned int step = 1U << vm_area_page_order(v);
4107 if (v->flags & VM_UNINITIALIZED)
4109 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4112 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4114 for (nr = 0; nr < v->nr_pages; nr += step)
4115 counters[page_to_nid(v->pages[nr])] += step;
4116 for_each_node_state(nr, N_HIGH_MEMORY)
4118 seq_printf(m, " N%u=%u", nr, counters[nr]);
4122 static void show_purge_info(struct seq_file *m)
4124 struct vmap_area *va;
4126 spin_lock(&purge_vmap_area_lock);
4127 list_for_each_entry(va, &purge_vmap_area_list, list) {
4128 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4129 (void *)va->va_start, (void *)va->va_end,
4130 va->va_end - va->va_start);
4132 spin_unlock(&purge_vmap_area_lock);
4135 static int s_show(struct seq_file *m, void *p)
4137 struct vmap_area *va;
4138 struct vm_struct *v;
4140 va = list_entry(p, struct vmap_area, list);
4143 * s_show can encounter race with remove_vm_area, !vm on behalf
4144 * of vmap area is being tear down or vm_map_ram allocation.
4147 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4148 (void *)va->va_start, (void *)va->va_end,
4149 va->va_end - va->va_start);
4156 seq_printf(m, "0x%pK-0x%pK %7ld",
4157 v->addr, v->addr + v->size, v->size);
4160 seq_printf(m, " %pS", v->caller);
4163 seq_printf(m, " pages=%d", v->nr_pages);
4166 seq_printf(m, " phys=%pa", &v->phys_addr);
4168 if (v->flags & VM_IOREMAP)
4169 seq_puts(m, " ioremap");
4171 if (v->flags & VM_ALLOC)
4172 seq_puts(m, " vmalloc");
4174 if (v->flags & VM_MAP)
4175 seq_puts(m, " vmap");
4177 if (v->flags & VM_USERMAP)
4178 seq_puts(m, " user");
4180 if (v->flags & VM_DMA_COHERENT)
4181 seq_puts(m, " dma-coherent");
4183 if (is_vmalloc_addr(v->pages))
4184 seq_puts(m, " vpages");
4186 show_numa_info(m, v);
4190 * As a final step, dump "unpurged" areas.
4193 if (list_is_last(&va->list, &vmap_area_list))
4199 static const struct seq_operations vmalloc_op = {
4206 static int __init proc_vmalloc_init(void)
4208 if (IS_ENABLED(CONFIG_NUMA))
4209 proc_create_seq_private("vmallocinfo", 0400, NULL,
4211 nr_node_ids * sizeof(unsigned int), NULL);
4213 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4216 module_init(proc_vmalloc_init);