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>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/vmalloc.h>
50 #include "pgalloc-track.h"
52 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
55 static int __init set_nohugeiomap(char *str)
57 ioremap_max_page_shift = PAGE_SHIFT;
60 early_param("nohugeiomap", set_nohugeiomap);
61 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
63 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66 static bool __ro_after_init vmap_allow_huge = true;
68 static int __init set_nohugevmalloc(char *str)
70 vmap_allow_huge = false;
73 early_param("nohugevmalloc", set_nohugevmalloc);
74 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 static const bool vmap_allow_huge = false;
76 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78 bool is_vmalloc_addr(const void *x)
80 unsigned long addr = (unsigned long)kasan_reset_tag(x);
82 return addr >= VMALLOC_START && addr < VMALLOC_END;
84 EXPORT_SYMBOL(is_vmalloc_addr);
86 struct vfree_deferred {
87 struct llist_head list;
88 struct work_struct wq;
90 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
92 static void __vunmap(const void *, int);
94 static void free_work(struct work_struct *w)
96 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
97 struct llist_node *t, *llnode;
99 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
100 __vunmap((void *)llnode, 1);
103 /*** Page table manipulation functions ***/
104 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
105 phys_addr_t phys_addr, pgprot_t prot,
106 unsigned int max_page_shift, pgtbl_mod_mask *mask)
110 unsigned long size = PAGE_SIZE;
112 pfn = phys_addr >> PAGE_SHIFT;
113 pte = pte_alloc_kernel_track(pmd, addr, mask);
117 BUG_ON(!pte_none(*pte));
119 #ifdef CONFIG_HUGETLB_PAGE
120 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
121 if (size != PAGE_SIZE) {
122 pte_t entry = pfn_pte(pfn, prot);
124 entry = arch_make_huge_pte(entry, ilog2(size), 0);
125 set_huge_pte_at(&init_mm, addr, pte, entry);
126 pfn += PFN_DOWN(size);
130 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
132 } while (pte += PFN_DOWN(size), addr += size, addr != end);
133 *mask |= PGTBL_PTE_MODIFIED;
137 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
138 phys_addr_t phys_addr, pgprot_t prot,
139 unsigned int max_page_shift)
141 if (max_page_shift < PMD_SHIFT)
144 if (!arch_vmap_pmd_supported(prot))
147 if ((end - addr) != PMD_SIZE)
150 if (!IS_ALIGNED(addr, PMD_SIZE))
153 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
156 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
159 return pmd_set_huge(pmd, phys_addr, prot);
162 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
163 phys_addr_t phys_addr, pgprot_t prot,
164 unsigned int max_page_shift, pgtbl_mod_mask *mask)
169 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
173 next = pmd_addr_end(addr, end);
175 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
177 *mask |= PGTBL_PMD_MODIFIED;
181 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
183 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
187 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
188 phys_addr_t phys_addr, pgprot_t prot,
189 unsigned int max_page_shift)
191 if (max_page_shift < PUD_SHIFT)
194 if (!arch_vmap_pud_supported(prot))
197 if ((end - addr) != PUD_SIZE)
200 if (!IS_ALIGNED(addr, PUD_SIZE))
203 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
206 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
209 return pud_set_huge(pud, phys_addr, prot);
212 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
213 phys_addr_t phys_addr, pgprot_t prot,
214 unsigned int max_page_shift, pgtbl_mod_mask *mask)
219 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
223 next = pud_addr_end(addr, end);
225 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
227 *mask |= PGTBL_PUD_MODIFIED;
231 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
232 max_page_shift, mask))
234 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
238 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
239 phys_addr_t phys_addr, pgprot_t prot,
240 unsigned int max_page_shift)
242 if (max_page_shift < P4D_SHIFT)
245 if (!arch_vmap_p4d_supported(prot))
248 if ((end - addr) != P4D_SIZE)
251 if (!IS_ALIGNED(addr, P4D_SIZE))
254 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
257 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
260 return p4d_set_huge(p4d, phys_addr, prot);
263 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
264 phys_addr_t phys_addr, pgprot_t prot,
265 unsigned int max_page_shift, pgtbl_mod_mask *mask)
270 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
274 next = p4d_addr_end(addr, end);
276 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
278 *mask |= PGTBL_P4D_MODIFIED;
282 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
283 max_page_shift, mask))
285 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
289 static int vmap_range_noflush(unsigned long addr, unsigned long end,
290 phys_addr_t phys_addr, pgprot_t prot,
291 unsigned int max_page_shift)
297 pgtbl_mod_mask mask = 0;
303 pgd = pgd_offset_k(addr);
305 next = pgd_addr_end(addr, end);
306 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
307 max_page_shift, &mask);
310 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
312 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
313 arch_sync_kernel_mappings(start, end);
318 int ioremap_page_range(unsigned long addr, unsigned long end,
319 phys_addr_t phys_addr, pgprot_t prot)
323 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
324 ioremap_max_page_shift);
325 flush_cache_vmap(addr, end);
327 kmsan_ioremap_page_range(addr, end, phys_addr, prot,
328 ioremap_max_page_shift);
332 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
333 pgtbl_mod_mask *mask)
337 pte = pte_offset_kernel(pmd, addr);
339 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
340 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
341 } while (pte++, addr += PAGE_SIZE, addr != end);
342 *mask |= PGTBL_PTE_MODIFIED;
345 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
346 pgtbl_mod_mask *mask)
352 pmd = pmd_offset(pud, addr);
354 next = pmd_addr_end(addr, end);
356 cleared = pmd_clear_huge(pmd);
357 if (cleared || pmd_bad(*pmd))
358 *mask |= PGTBL_PMD_MODIFIED;
362 if (pmd_none_or_clear_bad(pmd))
364 vunmap_pte_range(pmd, addr, next, mask);
367 } while (pmd++, addr = next, addr != end);
370 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
371 pgtbl_mod_mask *mask)
377 pud = pud_offset(p4d, addr);
379 next = pud_addr_end(addr, end);
381 cleared = pud_clear_huge(pud);
382 if (cleared || pud_bad(*pud))
383 *mask |= PGTBL_PUD_MODIFIED;
387 if (pud_none_or_clear_bad(pud))
389 vunmap_pmd_range(pud, addr, next, mask);
390 } while (pud++, addr = next, addr != end);
393 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
394 pgtbl_mod_mask *mask)
399 p4d = p4d_offset(pgd, addr);
401 next = p4d_addr_end(addr, end);
405 *mask |= PGTBL_P4D_MODIFIED;
407 if (p4d_none_or_clear_bad(p4d))
409 vunmap_pud_range(p4d, addr, next, mask);
410 } while (p4d++, addr = next, addr != end);
414 * vunmap_range_noflush is similar to vunmap_range, but does not
415 * flush caches or TLBs.
417 * The caller is responsible for calling flush_cache_vmap() before calling
418 * this function, and flush_tlb_kernel_range after it has returned
419 * successfully (and before the addresses are expected to cause a page fault
420 * or be re-mapped for something else, if TLB flushes are being delayed or
423 * This is an internal function only. Do not use outside mm/.
425 void __vunmap_range_noflush(unsigned long start, unsigned long end)
429 unsigned long addr = start;
430 pgtbl_mod_mask mask = 0;
433 pgd = pgd_offset_k(addr);
435 next = pgd_addr_end(addr, end);
437 mask |= PGTBL_PGD_MODIFIED;
438 if (pgd_none_or_clear_bad(pgd))
440 vunmap_p4d_range(pgd, addr, next, &mask);
441 } while (pgd++, addr = next, addr != end);
443 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
444 arch_sync_kernel_mappings(start, end);
447 void vunmap_range_noflush(unsigned long start, unsigned long end)
449 kmsan_vunmap_range_noflush(start, end);
450 __vunmap_range_noflush(start, end);
454 * vunmap_range - unmap kernel virtual addresses
455 * @addr: start of the VM area to unmap
456 * @end: end of the VM area to unmap (non-inclusive)
458 * Clears any present PTEs in the virtual address range, flushes TLBs and
459 * caches. Any subsequent access to the address before it has been re-mapped
462 void vunmap_range(unsigned long addr, unsigned long end)
464 flush_cache_vunmap(addr, end);
465 vunmap_range_noflush(addr, end);
466 flush_tlb_kernel_range(addr, end);
469 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
470 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
471 pgtbl_mod_mask *mask)
476 * nr is a running index into the array which helps higher level
477 * callers keep track of where we're up to.
480 pte = pte_alloc_kernel_track(pmd, addr, mask);
484 struct page *page = pages[*nr];
486 if (WARN_ON(!pte_none(*pte)))
490 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
493 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
495 } while (pte++, addr += PAGE_SIZE, addr != end);
496 *mask |= PGTBL_PTE_MODIFIED;
500 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
501 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
502 pgtbl_mod_mask *mask)
507 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
511 next = pmd_addr_end(addr, end);
512 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
514 } while (pmd++, addr = next, addr != end);
518 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
519 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
520 pgtbl_mod_mask *mask)
525 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
529 next = pud_addr_end(addr, end);
530 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
532 } while (pud++, addr = next, addr != end);
536 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
537 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
538 pgtbl_mod_mask *mask)
543 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
547 next = p4d_addr_end(addr, end);
548 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
550 } while (p4d++, addr = next, addr != end);
554 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
555 pgprot_t prot, struct page **pages)
557 unsigned long start = addr;
562 pgtbl_mod_mask mask = 0;
565 pgd = pgd_offset_k(addr);
567 next = pgd_addr_end(addr, end);
569 mask |= PGTBL_PGD_MODIFIED;
570 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
573 } while (pgd++, addr = next, addr != end);
575 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
576 arch_sync_kernel_mappings(start, end);
582 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
585 * The caller is responsible for calling flush_cache_vmap() after this
586 * function returns successfully and before the addresses are accessed.
588 * This is an internal function only. Do not use outside mm/.
590 int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
591 pgprot_t prot, struct page **pages, unsigned int page_shift)
593 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
595 WARN_ON(page_shift < PAGE_SHIFT);
597 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
598 page_shift == PAGE_SHIFT)
599 return vmap_small_pages_range_noflush(addr, end, prot, pages);
601 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
604 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
605 page_to_phys(pages[i]), prot,
610 addr += 1UL << page_shift;
616 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
617 pgprot_t prot, struct page **pages, unsigned int page_shift)
619 kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
620 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
624 * vmap_pages_range - map pages to a kernel virtual address
625 * @addr: start of the VM area to map
626 * @end: end of the VM area to map (non-inclusive)
627 * @prot: page protection flags to use
628 * @pages: pages to map (always PAGE_SIZE pages)
629 * @page_shift: maximum shift that the pages may be mapped with, @pages must
630 * be aligned and contiguous up to at least this shift.
633 * 0 on success, -errno on failure.
635 static int vmap_pages_range(unsigned long addr, unsigned long end,
636 pgprot_t prot, struct page **pages, unsigned int page_shift)
640 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
641 flush_cache_vmap(addr, end);
645 int is_vmalloc_or_module_addr(const void *x)
648 * ARM, x86-64 and sparc64 put modules in a special place,
649 * and fall back on vmalloc() if that fails. Others
650 * just put it in the vmalloc space.
652 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
653 unsigned long addr = (unsigned long)kasan_reset_tag(x);
654 if (addr >= MODULES_VADDR && addr < MODULES_END)
657 return is_vmalloc_addr(x);
661 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
662 * return the tail page that corresponds to the base page address, which
663 * matches small vmap mappings.
665 struct page *vmalloc_to_page(const void *vmalloc_addr)
667 unsigned long addr = (unsigned long) vmalloc_addr;
668 struct page *page = NULL;
669 pgd_t *pgd = pgd_offset_k(addr);
676 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
677 * architectures that do not vmalloc module space
679 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
683 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
684 return NULL; /* XXX: no allowance for huge pgd */
685 if (WARN_ON_ONCE(pgd_bad(*pgd)))
688 p4d = p4d_offset(pgd, addr);
692 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
693 if (WARN_ON_ONCE(p4d_bad(*p4d)))
696 pud = pud_offset(p4d, addr);
700 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
701 if (WARN_ON_ONCE(pud_bad(*pud)))
704 pmd = pmd_offset(pud, addr);
708 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
709 if (WARN_ON_ONCE(pmd_bad(*pmd)))
712 ptep = pte_offset_map(pmd, addr);
714 if (pte_present(pte))
715 page = pte_page(pte);
720 EXPORT_SYMBOL(vmalloc_to_page);
723 * Map a vmalloc()-space virtual address to the physical page frame number.
725 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
727 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
729 EXPORT_SYMBOL(vmalloc_to_pfn);
732 /*** Global kva allocator ***/
734 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
735 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
738 static DEFINE_SPINLOCK(vmap_area_lock);
739 static DEFINE_SPINLOCK(free_vmap_area_lock);
740 /* Export for kexec only */
741 LIST_HEAD(vmap_area_list);
742 static struct rb_root vmap_area_root = RB_ROOT;
743 static bool vmap_initialized __read_mostly;
745 static struct rb_root purge_vmap_area_root = RB_ROOT;
746 static LIST_HEAD(purge_vmap_area_list);
747 static DEFINE_SPINLOCK(purge_vmap_area_lock);
750 * This kmem_cache is used for vmap_area objects. Instead of
751 * allocating from slab we reuse an object from this cache to
752 * make things faster. Especially in "no edge" splitting of
755 static struct kmem_cache *vmap_area_cachep;
758 * This linked list is used in pair with free_vmap_area_root.
759 * It gives O(1) access to prev/next to perform fast coalescing.
761 static LIST_HEAD(free_vmap_area_list);
764 * This augment red-black tree represents the free vmap space.
765 * All vmap_area objects in this tree are sorted by va->va_start
766 * address. It is used for allocation and merging when a vmap
767 * object is released.
769 * Each vmap_area node contains a maximum available free block
770 * of its sub-tree, right or left. Therefore it is possible to
771 * find a lowest match of free area.
773 static struct rb_root free_vmap_area_root = RB_ROOT;
776 * Preload a CPU with one object for "no edge" split case. The
777 * aim is to get rid of allocations from the atomic context, thus
778 * to use more permissive allocation masks.
780 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
782 static __always_inline unsigned long
783 va_size(struct vmap_area *va)
785 return (va->va_end - va->va_start);
788 static __always_inline unsigned long
789 get_subtree_max_size(struct rb_node *node)
791 struct vmap_area *va;
793 va = rb_entry_safe(node, struct vmap_area, rb_node);
794 return va ? va->subtree_max_size : 0;
797 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
798 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
800 static void purge_vmap_area_lazy(void);
801 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
802 static void drain_vmap_area_work(struct work_struct *work);
803 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
805 static atomic_long_t nr_vmalloc_pages;
807 unsigned long vmalloc_nr_pages(void)
809 return atomic_long_read(&nr_vmalloc_pages);
812 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
813 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
815 struct vmap_area *va = NULL;
816 struct rb_node *n = vmap_area_root.rb_node;
818 addr = (unsigned long)kasan_reset_tag((void *)addr);
821 struct vmap_area *tmp;
823 tmp = rb_entry(n, struct vmap_area, rb_node);
824 if (tmp->va_end > addr) {
826 if (tmp->va_start <= addr)
837 static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
839 struct rb_node *n = root->rb_node;
841 addr = (unsigned long)kasan_reset_tag((void *)addr);
844 struct vmap_area *va;
846 va = rb_entry(n, struct vmap_area, rb_node);
847 if (addr < va->va_start)
849 else if (addr >= va->va_end)
859 * This function returns back addresses of parent node
860 * and its left or right link for further processing.
862 * Otherwise NULL is returned. In that case all further
863 * steps regarding inserting of conflicting overlap range
864 * have to be declined and actually considered as a bug.
866 static __always_inline struct rb_node **
867 find_va_links(struct vmap_area *va,
868 struct rb_root *root, struct rb_node *from,
869 struct rb_node **parent)
871 struct vmap_area *tmp_va;
872 struct rb_node **link;
875 link = &root->rb_node;
876 if (unlikely(!*link)) {
885 * Go to the bottom of the tree. When we hit the last point
886 * we end up with parent rb_node and correct direction, i name
887 * it link, where the new va->rb_node will be attached to.
890 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
893 * During the traversal we also do some sanity check.
894 * Trigger the BUG() if there are sides(left/right)
897 if (va->va_end <= tmp_va->va_start)
898 link = &(*link)->rb_left;
899 else if (va->va_start >= tmp_va->va_end)
900 link = &(*link)->rb_right;
902 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
903 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
909 *parent = &tmp_va->rb_node;
913 static __always_inline struct list_head *
914 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
916 struct list_head *list;
918 if (unlikely(!parent))
920 * The red-black tree where we try to find VA neighbors
921 * before merging or inserting is empty, i.e. it means
922 * there is no free vmap space. Normally it does not
923 * happen but we handle this case anyway.
927 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
928 return (&parent->rb_right == link ? list->next : list);
931 static __always_inline void
932 __link_va(struct vmap_area *va, struct rb_root *root,
933 struct rb_node *parent, struct rb_node **link,
934 struct list_head *head, bool augment)
937 * VA is still not in the list, but we can
938 * identify its future previous list_head node.
940 if (likely(parent)) {
941 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
942 if (&parent->rb_right != link)
946 /* Insert to the rb-tree */
947 rb_link_node(&va->rb_node, parent, link);
950 * Some explanation here. Just perform simple insertion
951 * to the tree. We do not set va->subtree_max_size to
952 * its current size before calling rb_insert_augmented().
953 * It is because we populate the tree from the bottom
954 * to parent levels when the node _is_ in the tree.
956 * Therefore we set subtree_max_size to zero after insertion,
957 * to let __augment_tree_propagate_from() puts everything to
958 * the correct order later on.
960 rb_insert_augmented(&va->rb_node,
961 root, &free_vmap_area_rb_augment_cb);
962 va->subtree_max_size = 0;
964 rb_insert_color(&va->rb_node, root);
967 /* Address-sort this list */
968 list_add(&va->list, head);
971 static __always_inline void
972 link_va(struct vmap_area *va, struct rb_root *root,
973 struct rb_node *parent, struct rb_node **link,
974 struct list_head *head)
976 __link_va(va, root, parent, link, head, false);
979 static __always_inline void
980 link_va_augment(struct vmap_area *va, struct rb_root *root,
981 struct rb_node *parent, struct rb_node **link,
982 struct list_head *head)
984 __link_va(va, root, parent, link, head, true);
987 static __always_inline void
988 __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
990 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
994 rb_erase_augmented(&va->rb_node,
995 root, &free_vmap_area_rb_augment_cb);
997 rb_erase(&va->rb_node, root);
999 list_del_init(&va->list);
1000 RB_CLEAR_NODE(&va->rb_node);
1003 static __always_inline void
1004 unlink_va(struct vmap_area *va, struct rb_root *root)
1006 __unlink_va(va, root, false);
1009 static __always_inline void
1010 unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1012 __unlink_va(va, root, true);
1015 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1017 * Gets called when remove the node and rotate.
1019 static __always_inline unsigned long
1020 compute_subtree_max_size(struct vmap_area *va)
1022 return max3(va_size(va),
1023 get_subtree_max_size(va->rb_node.rb_left),
1024 get_subtree_max_size(va->rb_node.rb_right));
1028 augment_tree_propagate_check(void)
1030 struct vmap_area *va;
1031 unsigned long computed_size;
1033 list_for_each_entry(va, &free_vmap_area_list, list) {
1034 computed_size = compute_subtree_max_size(va);
1035 if (computed_size != va->subtree_max_size)
1036 pr_emerg("tree is corrupted: %lu, %lu\n",
1037 va_size(va), va->subtree_max_size);
1043 * This function populates subtree_max_size from bottom to upper
1044 * levels starting from VA point. The propagation must be done
1045 * when VA size is modified by changing its va_start/va_end. Or
1046 * in case of newly inserting of VA to the tree.
1048 * It means that __augment_tree_propagate_from() must be called:
1049 * - After VA has been inserted to the tree(free path);
1050 * - After VA has been shrunk(allocation path);
1051 * - After VA has been increased(merging path).
1053 * Please note that, it does not mean that upper parent nodes
1054 * and their subtree_max_size are recalculated all the time up
1063 * For example if we modify the node 4, shrinking it to 2, then
1064 * no any modification is required. If we shrink the node 2 to 1
1065 * its subtree_max_size is updated only, and set to 1. If we shrink
1066 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1067 * node becomes 4--6.
1069 static __always_inline void
1070 augment_tree_propagate_from(struct vmap_area *va)
1073 * Populate the tree from bottom towards the root until
1074 * the calculated maximum available size of checked node
1075 * is equal to its current one.
1077 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1079 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1080 augment_tree_propagate_check();
1085 insert_vmap_area(struct vmap_area *va,
1086 struct rb_root *root, struct list_head *head)
1088 struct rb_node **link;
1089 struct rb_node *parent;
1091 link = find_va_links(va, root, NULL, &parent);
1093 link_va(va, root, parent, link, head);
1097 insert_vmap_area_augment(struct vmap_area *va,
1098 struct rb_node *from, struct rb_root *root,
1099 struct list_head *head)
1101 struct rb_node **link;
1102 struct rb_node *parent;
1105 link = find_va_links(va, NULL, from, &parent);
1107 link = find_va_links(va, root, NULL, &parent);
1110 link_va_augment(va, root, parent, link, head);
1111 augment_tree_propagate_from(va);
1116 * Merge de-allocated chunk of VA memory with previous
1117 * and next free blocks. If coalesce is not done a new
1118 * free area is inserted. If VA has been merged, it is
1121 * Please note, it can return NULL in case of overlap
1122 * ranges, followed by WARN() report. Despite it is a
1123 * buggy behaviour, a system can be alive and keep
1126 static __always_inline struct vmap_area *
1127 __merge_or_add_vmap_area(struct vmap_area *va,
1128 struct rb_root *root, struct list_head *head, bool augment)
1130 struct vmap_area *sibling;
1131 struct list_head *next;
1132 struct rb_node **link;
1133 struct rb_node *parent;
1134 bool merged = false;
1137 * Find a place in the tree where VA potentially will be
1138 * inserted, unless it is merged with its sibling/siblings.
1140 link = find_va_links(va, root, NULL, &parent);
1145 * Get next node of VA to check if merging can be done.
1147 next = get_va_next_sibling(parent, link);
1148 if (unlikely(next == NULL))
1154 * |<------VA------>|<-----Next----->|
1159 sibling = list_entry(next, struct vmap_area, list);
1160 if (sibling->va_start == va->va_end) {
1161 sibling->va_start = va->va_start;
1163 /* Free vmap_area object. */
1164 kmem_cache_free(vmap_area_cachep, va);
1166 /* Point to the new merged area. */
1175 * |<-----Prev----->|<------VA------>|
1179 if (next->prev != head) {
1180 sibling = list_entry(next->prev, struct vmap_area, list);
1181 if (sibling->va_end == va->va_start) {
1183 * If both neighbors are coalesced, it is important
1184 * to unlink the "next" node first, followed by merging
1185 * with "previous" one. Otherwise the tree might not be
1186 * fully populated if a sibling's augmented value is
1187 * "normalized" because of rotation operations.
1190 __unlink_va(va, root, augment);
1192 sibling->va_end = va->va_end;
1194 /* Free vmap_area object. */
1195 kmem_cache_free(vmap_area_cachep, va);
1197 /* Point to the new merged area. */
1205 __link_va(va, root, parent, link, head, augment);
1210 static __always_inline struct vmap_area *
1211 merge_or_add_vmap_area(struct vmap_area *va,
1212 struct rb_root *root, struct list_head *head)
1214 return __merge_or_add_vmap_area(va, root, head, false);
1217 static __always_inline struct vmap_area *
1218 merge_or_add_vmap_area_augment(struct vmap_area *va,
1219 struct rb_root *root, struct list_head *head)
1221 va = __merge_or_add_vmap_area(va, root, head, true);
1223 augment_tree_propagate_from(va);
1228 static __always_inline bool
1229 is_within_this_va(struct vmap_area *va, unsigned long size,
1230 unsigned long align, unsigned long vstart)
1232 unsigned long nva_start_addr;
1234 if (va->va_start > vstart)
1235 nva_start_addr = ALIGN(va->va_start, align);
1237 nva_start_addr = ALIGN(vstart, align);
1239 /* Can be overflowed due to big size or alignment. */
1240 if (nva_start_addr + size < nva_start_addr ||
1241 nva_start_addr < vstart)
1244 return (nva_start_addr + size <= va->va_end);
1248 * Find the first free block(lowest start address) in the tree,
1249 * that will accomplish the request corresponding to passing
1250 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1251 * a search length is adjusted to account for worst case alignment
1254 static __always_inline struct vmap_area *
1255 find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1256 unsigned long align, unsigned long vstart, bool adjust_search_size)
1258 struct vmap_area *va;
1259 struct rb_node *node;
1260 unsigned long length;
1262 /* Start from the root. */
1263 node = root->rb_node;
1265 /* Adjust the search size for alignment overhead. */
1266 length = adjust_search_size ? size + align - 1 : size;
1269 va = rb_entry(node, struct vmap_area, rb_node);
1271 if (get_subtree_max_size(node->rb_left) >= length &&
1272 vstart < va->va_start) {
1273 node = node->rb_left;
1275 if (is_within_this_va(va, size, align, vstart))
1279 * Does not make sense to go deeper towards the right
1280 * sub-tree if it does not have a free block that is
1281 * equal or bigger to the requested search length.
1283 if (get_subtree_max_size(node->rb_right) >= length) {
1284 node = node->rb_right;
1289 * OK. We roll back and find the first right sub-tree,
1290 * that will satisfy the search criteria. It can happen
1291 * due to "vstart" restriction or an alignment overhead
1292 * that is bigger then PAGE_SIZE.
1294 while ((node = rb_parent(node))) {
1295 va = rb_entry(node, struct vmap_area, rb_node);
1296 if (is_within_this_va(va, size, align, vstart))
1299 if (get_subtree_max_size(node->rb_right) >= length &&
1300 vstart <= va->va_start) {
1302 * Shift the vstart forward. Please note, we update it with
1303 * parent's start address adding "1" because we do not want
1304 * to enter same sub-tree after it has already been checked
1305 * and no suitable free block found there.
1307 vstart = va->va_start + 1;
1308 node = node->rb_right;
1318 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1319 #include <linux/random.h>
1321 static struct vmap_area *
1322 find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1323 unsigned long align, unsigned long vstart)
1325 struct vmap_area *va;
1327 list_for_each_entry(va, head, list) {
1328 if (!is_within_this_va(va, size, align, vstart))
1338 find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1339 unsigned long size, unsigned long align)
1341 struct vmap_area *va_1, *va_2;
1342 unsigned long vstart;
1345 get_random_bytes(&rnd, sizeof(rnd));
1346 vstart = VMALLOC_START + rnd;
1348 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1349 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1352 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1353 va_1, va_2, vstart);
1359 FL_FIT_TYPE = 1, /* full fit */
1360 LE_FIT_TYPE = 2, /* left edge fit */
1361 RE_FIT_TYPE = 3, /* right edge fit */
1362 NE_FIT_TYPE = 4 /* no edge fit */
1365 static __always_inline enum fit_type
1366 classify_va_fit_type(struct vmap_area *va,
1367 unsigned long nva_start_addr, unsigned long size)
1371 /* Check if it is within VA. */
1372 if (nva_start_addr < va->va_start ||
1373 nva_start_addr + size > va->va_end)
1377 if (va->va_start == nva_start_addr) {
1378 if (va->va_end == nva_start_addr + size)
1382 } else if (va->va_end == nva_start_addr + size) {
1391 static __always_inline int
1392 adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1393 struct vmap_area *va, unsigned long nva_start_addr,
1396 struct vmap_area *lva = NULL;
1397 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1399 if (type == FL_FIT_TYPE) {
1401 * No need to split VA, it fully fits.
1407 unlink_va_augment(va, root);
1408 kmem_cache_free(vmap_area_cachep, va);
1409 } else if (type == LE_FIT_TYPE) {
1411 * Split left edge of fit VA.
1417 va->va_start += size;
1418 } else if (type == RE_FIT_TYPE) {
1420 * Split right edge of fit VA.
1426 va->va_end = nva_start_addr;
1427 } else if (type == NE_FIT_TYPE) {
1429 * Split no edge of fit VA.
1435 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1436 if (unlikely(!lva)) {
1438 * For percpu allocator we do not do any pre-allocation
1439 * and leave it as it is. The reason is it most likely
1440 * never ends up with NE_FIT_TYPE splitting. In case of
1441 * percpu allocations offsets and sizes are aligned to
1442 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1443 * are its main fitting cases.
1445 * There are a few exceptions though, as an example it is
1446 * a first allocation (early boot up) when we have "one"
1447 * big free space that has to be split.
1449 * Also we can hit this path in case of regular "vmap"
1450 * allocations, if "this" current CPU was not preloaded.
1451 * See the comment in alloc_vmap_area() why. If so, then
1452 * GFP_NOWAIT is used instead to get an extra object for
1453 * split purpose. That is rare and most time does not
1456 * What happens if an allocation gets failed. Basically,
1457 * an "overflow" path is triggered to purge lazily freed
1458 * areas to free some memory, then, the "retry" path is
1459 * triggered to repeat one more time. See more details
1460 * in alloc_vmap_area() function.
1462 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1468 * Build the remainder.
1470 lva->va_start = va->va_start;
1471 lva->va_end = nva_start_addr;
1474 * Shrink this VA to remaining size.
1476 va->va_start = nva_start_addr + size;
1481 if (type != FL_FIT_TYPE) {
1482 augment_tree_propagate_from(va);
1484 if (lva) /* type == NE_FIT_TYPE */
1485 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1492 * Returns a start address of the newly allocated area, if success.
1493 * Otherwise a vend is returned that indicates failure.
1495 static __always_inline unsigned long
1496 __alloc_vmap_area(struct rb_root *root, struct list_head *head,
1497 unsigned long size, unsigned long align,
1498 unsigned long vstart, unsigned long vend)
1500 bool adjust_search_size = true;
1501 unsigned long nva_start_addr;
1502 struct vmap_area *va;
1506 * Do not adjust when:
1507 * a) align <= PAGE_SIZE, because it does not make any sense.
1508 * All blocks(their start addresses) are at least PAGE_SIZE
1510 * b) a short range where a requested size corresponds to exactly
1511 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1512 * With adjusted search length an allocation would not succeed.
1514 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1515 adjust_search_size = false;
1517 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1521 if (va->va_start > vstart)
1522 nva_start_addr = ALIGN(va->va_start, align);
1524 nva_start_addr = ALIGN(vstart, align);
1526 /* Check the "vend" restriction. */
1527 if (nva_start_addr + size > vend)
1530 /* Update the free vmap_area. */
1531 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1532 if (WARN_ON_ONCE(ret))
1535 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1536 find_vmap_lowest_match_check(root, head, size, align);
1539 return nva_start_addr;
1543 * Free a region of KVA allocated by alloc_vmap_area
1545 static void free_vmap_area(struct vmap_area *va)
1548 * Remove from the busy tree/list.
1550 spin_lock(&vmap_area_lock);
1551 unlink_va(va, &vmap_area_root);
1552 spin_unlock(&vmap_area_lock);
1555 * Insert/Merge it back to the free tree/list.
1557 spin_lock(&free_vmap_area_lock);
1558 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1559 spin_unlock(&free_vmap_area_lock);
1563 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1565 struct vmap_area *va = NULL;
1568 * Preload this CPU with one extra vmap_area object. It is used
1569 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1570 * a CPU that does an allocation is preloaded.
1572 * We do it in non-atomic context, thus it allows us to use more
1573 * permissive allocation masks to be more stable under low memory
1574 * condition and high memory pressure.
1576 if (!this_cpu_read(ne_fit_preload_node))
1577 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1581 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1582 kmem_cache_free(vmap_area_cachep, va);
1586 * Allocate a region of KVA of the specified size and alignment, within the
1589 static struct vmap_area *alloc_vmap_area(unsigned long size,
1590 unsigned long align,
1591 unsigned long vstart, unsigned long vend,
1592 int node, gfp_t gfp_mask)
1594 struct vmap_area *va;
1595 unsigned long freed;
1601 BUG_ON(offset_in_page(size));
1602 BUG_ON(!is_power_of_2(align));
1604 if (unlikely(!vmap_initialized))
1605 return ERR_PTR(-EBUSY);
1608 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1610 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1612 return ERR_PTR(-ENOMEM);
1615 * Only scan the relevant parts containing pointers to other objects
1616 * to avoid false negatives.
1618 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1621 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1622 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1623 size, align, vstart, vend);
1624 spin_unlock(&free_vmap_area_lock);
1626 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1629 * If an allocation fails, the "vend" address is
1630 * returned. Therefore trigger the overflow path.
1632 if (unlikely(addr == vend))
1635 va->va_start = addr;
1636 va->va_end = addr + size;
1639 spin_lock(&vmap_area_lock);
1640 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1641 spin_unlock(&vmap_area_lock);
1643 BUG_ON(!IS_ALIGNED(va->va_start, align));
1644 BUG_ON(va->va_start < vstart);
1645 BUG_ON(va->va_end > vend);
1647 ret = kasan_populate_vmalloc(addr, size);
1650 return ERR_PTR(ret);
1657 purge_vmap_area_lazy();
1663 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1670 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1671 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1674 kmem_cache_free(vmap_area_cachep, va);
1675 return ERR_PTR(-EBUSY);
1678 int register_vmap_purge_notifier(struct notifier_block *nb)
1680 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1682 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1684 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1686 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1688 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1691 * lazy_max_pages is the maximum amount of virtual address space we gather up
1692 * before attempting to purge with a TLB flush.
1694 * There is a tradeoff here: a larger number will cover more kernel page tables
1695 * and take slightly longer to purge, but it will linearly reduce the number of
1696 * global TLB flushes that must be performed. It would seem natural to scale
1697 * this number up linearly with the number of CPUs (because vmapping activity
1698 * could also scale linearly with the number of CPUs), however it is likely
1699 * that in practice, workloads might be constrained in other ways that mean
1700 * vmap activity will not scale linearly with CPUs. Also, I want to be
1701 * conservative and not introduce a big latency on huge systems, so go with
1702 * a less aggressive log scale. It will still be an improvement over the old
1703 * code, and it will be simple to change the scale factor if we find that it
1704 * becomes a problem on bigger systems.
1706 static unsigned long lazy_max_pages(void)
1710 log = fls(num_online_cpus());
1712 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1715 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1718 * Serialize vmap purging. There is no actual critical section protected
1719 * by this lock, but we want to avoid concurrent calls for performance
1720 * reasons and to make the pcpu_get_vm_areas more deterministic.
1722 static DEFINE_MUTEX(vmap_purge_lock);
1724 /* for per-CPU blocks */
1725 static void purge_fragmented_blocks_allcpus(void);
1728 * Purges all lazily-freed vmap areas.
1730 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1732 unsigned long resched_threshold;
1733 struct list_head local_purge_list;
1734 struct vmap_area *va, *n_va;
1736 lockdep_assert_held(&vmap_purge_lock);
1738 spin_lock(&purge_vmap_area_lock);
1739 purge_vmap_area_root = RB_ROOT;
1740 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1741 spin_unlock(&purge_vmap_area_lock);
1743 if (unlikely(list_empty(&local_purge_list)))
1747 list_first_entry(&local_purge_list,
1748 struct vmap_area, list)->va_start);
1751 list_last_entry(&local_purge_list,
1752 struct vmap_area, list)->va_end);
1754 flush_tlb_kernel_range(start, end);
1755 resched_threshold = lazy_max_pages() << 1;
1757 spin_lock(&free_vmap_area_lock);
1758 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1759 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1760 unsigned long orig_start = va->va_start;
1761 unsigned long orig_end = va->va_end;
1764 * Finally insert or merge lazily-freed area. It is
1765 * detached and there is no need to "unlink" it from
1768 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1769 &free_vmap_area_list);
1774 if (is_vmalloc_or_module_addr((void *)orig_start))
1775 kasan_release_vmalloc(orig_start, orig_end,
1776 va->va_start, va->va_end);
1778 atomic_long_sub(nr, &vmap_lazy_nr);
1780 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1781 cond_resched_lock(&free_vmap_area_lock);
1783 spin_unlock(&free_vmap_area_lock);
1788 * Kick off a purge of the outstanding lazy areas.
1790 static void purge_vmap_area_lazy(void)
1792 mutex_lock(&vmap_purge_lock);
1793 purge_fragmented_blocks_allcpus();
1794 __purge_vmap_area_lazy(ULONG_MAX, 0);
1795 mutex_unlock(&vmap_purge_lock);
1798 static void drain_vmap_area_work(struct work_struct *work)
1800 unsigned long nr_lazy;
1803 mutex_lock(&vmap_purge_lock);
1804 __purge_vmap_area_lazy(ULONG_MAX, 0);
1805 mutex_unlock(&vmap_purge_lock);
1807 /* Recheck if further work is required. */
1808 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1809 } while (nr_lazy > lazy_max_pages());
1813 * Free a vmap area, caller ensuring that the area has been unmapped
1814 * and flush_cache_vunmap had been called for the correct range
1817 static void free_vmap_area_noflush(struct vmap_area *va)
1819 unsigned long nr_lazy;
1821 spin_lock(&vmap_area_lock);
1822 unlink_va(va, &vmap_area_root);
1823 spin_unlock(&vmap_area_lock);
1825 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1826 PAGE_SHIFT, &vmap_lazy_nr);
1829 * Merge or place it to the purge tree/list.
1831 spin_lock(&purge_vmap_area_lock);
1832 merge_or_add_vmap_area(va,
1833 &purge_vmap_area_root, &purge_vmap_area_list);
1834 spin_unlock(&purge_vmap_area_lock);
1836 /* After this point, we may free va at any time */
1837 if (unlikely(nr_lazy > lazy_max_pages()))
1838 schedule_work(&drain_vmap_work);
1842 * Free and unmap a vmap area
1844 static void free_unmap_vmap_area(struct vmap_area *va)
1846 flush_cache_vunmap(va->va_start, va->va_end);
1847 vunmap_range_noflush(va->va_start, va->va_end);
1848 if (debug_pagealloc_enabled_static())
1849 flush_tlb_kernel_range(va->va_start, va->va_end);
1851 free_vmap_area_noflush(va);
1854 struct vmap_area *find_vmap_area(unsigned long addr)
1856 struct vmap_area *va;
1858 spin_lock(&vmap_area_lock);
1859 va = __find_vmap_area(addr, &vmap_area_root);
1860 spin_unlock(&vmap_area_lock);
1865 /*** Per cpu kva allocator ***/
1868 * vmap space is limited especially on 32 bit architectures. Ensure there is
1869 * room for at least 16 percpu vmap blocks per CPU.
1872 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1873 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1874 * instead (we just need a rough idea)
1876 #if BITS_PER_LONG == 32
1877 #define VMALLOC_SPACE (128UL*1024*1024)
1879 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1882 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1883 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1884 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1885 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1886 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1887 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1888 #define VMAP_BBMAP_BITS \
1889 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1890 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1891 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1893 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1895 struct vmap_block_queue {
1897 struct list_head free;
1902 struct vmap_area *va;
1903 unsigned long free, dirty;
1904 unsigned long dirty_min, dirty_max; /*< dirty range */
1905 struct list_head free_list;
1906 struct rcu_head rcu_head;
1907 struct list_head purge;
1910 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1911 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1914 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1915 * in the free path. Could get rid of this if we change the API to return a
1916 * "cookie" from alloc, to be passed to free. But no big deal yet.
1918 static DEFINE_XARRAY(vmap_blocks);
1921 * We should probably have a fallback mechanism to allocate virtual memory
1922 * out of partially filled vmap blocks. However vmap block sizing should be
1923 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1927 static unsigned long addr_to_vb_idx(unsigned long addr)
1929 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1930 addr /= VMAP_BLOCK_SIZE;
1934 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1938 addr = va_start + (pages_off << PAGE_SHIFT);
1939 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1940 return (void *)addr;
1944 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1945 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1946 * @order: how many 2^order pages should be occupied in newly allocated block
1947 * @gfp_mask: flags for the page level allocator
1949 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1951 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1953 struct vmap_block_queue *vbq;
1954 struct vmap_block *vb;
1955 struct vmap_area *va;
1956 unsigned long vb_idx;
1960 node = numa_node_id();
1962 vb = kmalloc_node(sizeof(struct vmap_block),
1963 gfp_mask & GFP_RECLAIM_MASK, node);
1965 return ERR_PTR(-ENOMEM);
1967 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1968 VMALLOC_START, VMALLOC_END,
1972 return ERR_CAST(va);
1975 vaddr = vmap_block_vaddr(va->va_start, 0);
1976 spin_lock_init(&vb->lock);
1978 /* At least something should be left free */
1979 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1980 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1982 vb->dirty_min = VMAP_BBMAP_BITS;
1984 INIT_LIST_HEAD(&vb->free_list);
1986 vb_idx = addr_to_vb_idx(va->va_start);
1987 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1991 return ERR_PTR(err);
1994 vbq = raw_cpu_ptr(&vmap_block_queue);
1995 spin_lock(&vbq->lock);
1996 list_add_tail_rcu(&vb->free_list, &vbq->free);
1997 spin_unlock(&vbq->lock);
2002 static void free_vmap_block(struct vmap_block *vb)
2004 struct vmap_block *tmp;
2006 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2009 free_vmap_area_noflush(vb->va);
2010 kfree_rcu(vb, rcu_head);
2013 static void purge_fragmented_blocks(int cpu)
2016 struct vmap_block *vb;
2017 struct vmap_block *n_vb;
2018 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2021 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2023 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2026 spin_lock(&vb->lock);
2027 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2028 vb->free = 0; /* prevent further allocs after releasing lock */
2029 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2031 vb->dirty_max = VMAP_BBMAP_BITS;
2032 spin_lock(&vbq->lock);
2033 list_del_rcu(&vb->free_list);
2034 spin_unlock(&vbq->lock);
2035 spin_unlock(&vb->lock);
2036 list_add_tail(&vb->purge, &purge);
2038 spin_unlock(&vb->lock);
2042 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2043 list_del(&vb->purge);
2044 free_vmap_block(vb);
2048 static void purge_fragmented_blocks_allcpus(void)
2052 for_each_possible_cpu(cpu)
2053 purge_fragmented_blocks(cpu);
2056 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2058 struct vmap_block_queue *vbq;
2059 struct vmap_block *vb;
2063 BUG_ON(offset_in_page(size));
2064 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2065 if (WARN_ON(size == 0)) {
2067 * Allocating 0 bytes isn't what caller wants since
2068 * get_order(0) returns funny result. Just warn and terminate
2073 order = get_order(size);
2076 vbq = raw_cpu_ptr(&vmap_block_queue);
2077 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2078 unsigned long pages_off;
2080 spin_lock(&vb->lock);
2081 if (vb->free < (1UL << order)) {
2082 spin_unlock(&vb->lock);
2086 pages_off = VMAP_BBMAP_BITS - vb->free;
2087 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2088 vb->free -= 1UL << order;
2089 if (vb->free == 0) {
2090 spin_lock(&vbq->lock);
2091 list_del_rcu(&vb->free_list);
2092 spin_unlock(&vbq->lock);
2095 spin_unlock(&vb->lock);
2101 /* Allocate new block if nothing was found */
2103 vaddr = new_vmap_block(order, gfp_mask);
2108 static void vb_free(unsigned long addr, unsigned long size)
2110 unsigned long offset;
2112 struct vmap_block *vb;
2114 BUG_ON(offset_in_page(size));
2115 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2117 flush_cache_vunmap(addr, addr + size);
2119 order = get_order(size);
2120 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2121 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2123 vunmap_range_noflush(addr, addr + size);
2125 if (debug_pagealloc_enabled_static())
2126 flush_tlb_kernel_range(addr, addr + size);
2128 spin_lock(&vb->lock);
2130 /* Expand dirty range */
2131 vb->dirty_min = min(vb->dirty_min, offset);
2132 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2134 vb->dirty += 1UL << order;
2135 if (vb->dirty == VMAP_BBMAP_BITS) {
2137 spin_unlock(&vb->lock);
2138 free_vmap_block(vb);
2140 spin_unlock(&vb->lock);
2143 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2147 if (unlikely(!vmap_initialized))
2152 for_each_possible_cpu(cpu) {
2153 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2154 struct vmap_block *vb;
2157 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2158 spin_lock(&vb->lock);
2159 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2160 unsigned long va_start = vb->va->va_start;
2163 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2164 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2166 start = min(s, start);
2171 spin_unlock(&vb->lock);
2176 mutex_lock(&vmap_purge_lock);
2177 purge_fragmented_blocks_allcpus();
2178 if (!__purge_vmap_area_lazy(start, end) && flush)
2179 flush_tlb_kernel_range(start, end);
2180 mutex_unlock(&vmap_purge_lock);
2184 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2186 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2187 * to amortize TLB flushing overheads. What this means is that any page you
2188 * have now, may, in a former life, have been mapped into kernel virtual
2189 * address by the vmap layer and so there might be some CPUs with TLB entries
2190 * still referencing that page (additional to the regular 1:1 kernel mapping).
2192 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2193 * be sure that none of the pages we have control over will have any aliases
2194 * from the vmap layer.
2196 void vm_unmap_aliases(void)
2198 unsigned long start = ULONG_MAX, end = 0;
2201 _vm_unmap_aliases(start, end, flush);
2203 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2206 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2207 * @mem: the pointer returned by vm_map_ram
2208 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2210 void vm_unmap_ram(const void *mem, unsigned int count)
2212 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2213 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2214 struct vmap_area *va;
2218 BUG_ON(addr < VMALLOC_START);
2219 BUG_ON(addr > VMALLOC_END);
2220 BUG_ON(!PAGE_ALIGNED(addr));
2222 kasan_poison_vmalloc(mem, size);
2224 if (likely(count <= VMAP_MAX_ALLOC)) {
2225 debug_check_no_locks_freed(mem, size);
2226 vb_free(addr, size);
2230 va = find_vmap_area(addr);
2232 debug_check_no_locks_freed((void *)va->va_start,
2233 (va->va_end - va->va_start));
2234 free_unmap_vmap_area(va);
2236 EXPORT_SYMBOL(vm_unmap_ram);
2239 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2240 * @pages: an array of pointers to the pages to be mapped
2241 * @count: number of pages
2242 * @node: prefer to allocate data structures on this node
2244 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2245 * faster than vmap so it's good. But if you mix long-life and short-life
2246 * objects with vm_map_ram(), it could consume lots of address space through
2247 * fragmentation (especially on a 32bit machine). You could see failures in
2248 * the end. Please use this function for short-lived objects.
2250 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2252 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2254 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2258 if (likely(count <= VMAP_MAX_ALLOC)) {
2259 mem = vb_alloc(size, GFP_KERNEL);
2262 addr = (unsigned long)mem;
2264 struct vmap_area *va;
2265 va = alloc_vmap_area(size, PAGE_SIZE,
2266 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2270 addr = va->va_start;
2274 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2275 pages, PAGE_SHIFT) < 0) {
2276 vm_unmap_ram(mem, count);
2281 * Mark the pages as accessible, now that they are mapped.
2282 * With hardware tag-based KASAN, marking is skipped for
2283 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2285 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2289 EXPORT_SYMBOL(vm_map_ram);
2291 static struct vm_struct *vmlist __initdata;
2293 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2295 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2296 return vm->page_order;
2302 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2304 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2305 vm->page_order = order;
2312 * vm_area_add_early - add vmap area early during boot
2313 * @vm: vm_struct to add
2315 * This function is used to add fixed kernel vm area to vmlist before
2316 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2317 * should contain proper values and the other fields should be zero.
2319 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2321 void __init vm_area_add_early(struct vm_struct *vm)
2323 struct vm_struct *tmp, **p;
2325 BUG_ON(vmap_initialized);
2326 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2327 if (tmp->addr >= vm->addr) {
2328 BUG_ON(tmp->addr < vm->addr + vm->size);
2331 BUG_ON(tmp->addr + tmp->size > vm->addr);
2338 * vm_area_register_early - register vmap area early during boot
2339 * @vm: vm_struct to register
2340 * @align: requested alignment
2342 * This function is used to register kernel vm area before
2343 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2344 * proper values on entry and other fields should be zero. On return,
2345 * vm->addr contains the allocated address.
2347 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2349 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2351 unsigned long addr = ALIGN(VMALLOC_START, align);
2352 struct vm_struct *cur, **p;
2354 BUG_ON(vmap_initialized);
2356 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2357 if ((unsigned long)cur->addr - addr >= vm->size)
2359 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2362 BUG_ON(addr > VMALLOC_END - vm->size);
2363 vm->addr = (void *)addr;
2366 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2369 static void vmap_init_free_space(void)
2371 unsigned long vmap_start = 1;
2372 const unsigned long vmap_end = ULONG_MAX;
2373 struct vmap_area *busy, *free;
2377 * -|-----|.....|-----|-----|-----|.....|-
2379 * |<--------------------------------->|
2381 list_for_each_entry(busy, &vmap_area_list, list) {
2382 if (busy->va_start - vmap_start > 0) {
2383 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2384 if (!WARN_ON_ONCE(!free)) {
2385 free->va_start = vmap_start;
2386 free->va_end = busy->va_start;
2388 insert_vmap_area_augment(free, NULL,
2389 &free_vmap_area_root,
2390 &free_vmap_area_list);
2394 vmap_start = busy->va_end;
2397 if (vmap_end - vmap_start > 0) {
2398 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2399 if (!WARN_ON_ONCE(!free)) {
2400 free->va_start = vmap_start;
2401 free->va_end = vmap_end;
2403 insert_vmap_area_augment(free, NULL,
2404 &free_vmap_area_root,
2405 &free_vmap_area_list);
2410 void __init vmalloc_init(void)
2412 struct vmap_area *va;
2413 struct vm_struct *tmp;
2417 * Create the cache for vmap_area objects.
2419 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2421 for_each_possible_cpu(i) {
2422 struct vmap_block_queue *vbq;
2423 struct vfree_deferred *p;
2425 vbq = &per_cpu(vmap_block_queue, i);
2426 spin_lock_init(&vbq->lock);
2427 INIT_LIST_HEAD(&vbq->free);
2428 p = &per_cpu(vfree_deferred, i);
2429 init_llist_head(&p->list);
2430 INIT_WORK(&p->wq, free_work);
2433 /* Import existing vmlist entries. */
2434 for (tmp = vmlist; tmp; tmp = tmp->next) {
2435 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2436 if (WARN_ON_ONCE(!va))
2439 va->va_start = (unsigned long)tmp->addr;
2440 va->va_end = va->va_start + tmp->size;
2442 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2446 * Now we can initialize a free vmap space.
2448 vmap_init_free_space();
2449 vmap_initialized = true;
2452 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2453 struct vmap_area *va, unsigned long flags, const void *caller)
2456 vm->addr = (void *)va->va_start;
2457 vm->size = va->va_end - va->va_start;
2458 vm->caller = caller;
2462 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2463 unsigned long flags, const void *caller)
2465 spin_lock(&vmap_area_lock);
2466 setup_vmalloc_vm_locked(vm, va, flags, caller);
2467 spin_unlock(&vmap_area_lock);
2470 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2473 * Before removing VM_UNINITIALIZED,
2474 * we should make sure that vm has proper values.
2475 * Pair with smp_rmb() in show_numa_info().
2478 vm->flags &= ~VM_UNINITIALIZED;
2481 static struct vm_struct *__get_vm_area_node(unsigned long size,
2482 unsigned long align, unsigned long shift, unsigned long flags,
2483 unsigned long start, unsigned long end, int node,
2484 gfp_t gfp_mask, const void *caller)
2486 struct vmap_area *va;
2487 struct vm_struct *area;
2488 unsigned long requested_size = size;
2490 BUG_ON(in_interrupt());
2491 size = ALIGN(size, 1ul << shift);
2492 if (unlikely(!size))
2495 if (flags & VM_IOREMAP)
2496 align = 1ul << clamp_t(int, get_count_order_long(size),
2497 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2499 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2500 if (unlikely(!area))
2503 if (!(flags & VM_NO_GUARD))
2506 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2512 setup_vmalloc_vm(area, va, flags, caller);
2515 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2516 * best-effort approach, as they can be mapped outside of vmalloc code.
2517 * For VM_ALLOC mappings, the pages are marked as accessible after
2518 * getting mapped in __vmalloc_node_range().
2519 * With hardware tag-based KASAN, marking is skipped for
2520 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2522 if (!(flags & VM_ALLOC))
2523 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2524 KASAN_VMALLOC_PROT_NORMAL);
2529 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2530 unsigned long start, unsigned long end,
2533 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2534 NUMA_NO_NODE, GFP_KERNEL, caller);
2538 * get_vm_area - reserve a contiguous kernel virtual area
2539 * @size: size of the area
2540 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2542 * Search an area of @size in the kernel virtual mapping area,
2543 * and reserved it for out purposes. Returns the area descriptor
2544 * on success or %NULL on failure.
2546 * Return: the area descriptor on success or %NULL on failure.
2548 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2550 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2551 VMALLOC_START, VMALLOC_END,
2552 NUMA_NO_NODE, GFP_KERNEL,
2553 __builtin_return_address(0));
2556 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2559 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2560 VMALLOC_START, VMALLOC_END,
2561 NUMA_NO_NODE, GFP_KERNEL, caller);
2565 * find_vm_area - find a continuous kernel virtual area
2566 * @addr: base address
2568 * Search for the kernel VM area starting at @addr, and return it.
2569 * It is up to the caller to do all required locking to keep the returned
2572 * Return: the area descriptor on success or %NULL on failure.
2574 struct vm_struct *find_vm_area(const void *addr)
2576 struct vmap_area *va;
2578 va = find_vmap_area((unsigned long)addr);
2586 * remove_vm_area - find and remove a continuous kernel virtual area
2587 * @addr: base address
2589 * Search for the kernel VM area starting at @addr, and remove it.
2590 * This function returns the found VM area, but using it is NOT safe
2591 * on SMP machines, except for its size or flags.
2593 * Return: the area descriptor on success or %NULL on failure.
2595 struct vm_struct *remove_vm_area(const void *addr)
2597 struct vmap_area *va;
2601 spin_lock(&vmap_area_lock);
2602 va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2604 struct vm_struct *vm = va->vm;
2607 spin_unlock(&vmap_area_lock);
2609 kasan_free_module_shadow(vm);
2610 free_unmap_vmap_area(va);
2615 spin_unlock(&vmap_area_lock);
2619 static inline void set_area_direct_map(const struct vm_struct *area,
2620 int (*set_direct_map)(struct page *page))
2624 /* HUGE_VMALLOC passes small pages to set_direct_map */
2625 for (i = 0; i < area->nr_pages; i++)
2626 if (page_address(area->pages[i]))
2627 set_direct_map(area->pages[i]);
2630 /* Handle removing and resetting vm mappings related to the vm_struct. */
2631 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2633 unsigned long start = ULONG_MAX, end = 0;
2634 unsigned int page_order = vm_area_page_order(area);
2635 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2639 remove_vm_area(area->addr);
2641 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2646 * If not deallocating pages, just do the flush of the VM area and
2649 if (!deallocate_pages) {
2655 * If execution gets here, flush the vm mapping and reset the direct
2656 * map. Find the start and end range of the direct mappings to make sure
2657 * the vm_unmap_aliases() flush includes the direct map.
2659 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2660 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2662 unsigned long page_size;
2664 page_size = PAGE_SIZE << page_order;
2665 start = min(addr, start);
2666 end = max(addr + page_size, end);
2672 * Set direct map to something invalid so that it won't be cached if
2673 * there are any accesses after the TLB flush, then flush the TLB and
2674 * reset the direct map permissions to the default.
2676 set_area_direct_map(area, set_direct_map_invalid_noflush);
2677 _vm_unmap_aliases(start, end, flush_dmap);
2678 set_area_direct_map(area, set_direct_map_default_noflush);
2681 static void __vunmap(const void *addr, int deallocate_pages)
2683 struct vm_struct *area;
2688 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2692 area = find_vm_area(addr);
2693 if (unlikely(!area)) {
2694 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2699 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2700 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2702 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2704 vm_remove_mappings(area, deallocate_pages);
2706 if (deallocate_pages) {
2709 for (i = 0; i < area->nr_pages; i++) {
2710 struct page *page = area->pages[i];
2713 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2715 * High-order allocs for huge vmallocs are split, so
2716 * can be freed as an array of order-0 allocations
2718 __free_pages(page, 0);
2721 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2723 kvfree(area->pages);
2729 static inline void __vfree_deferred(const void *addr)
2732 * Use raw_cpu_ptr() because this can be called from preemptible
2733 * context. Preemption is absolutely fine here, because the llist_add()
2734 * implementation is lockless, so it works even if we are adding to
2735 * another cpu's list. schedule_work() should be fine with this too.
2737 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2739 if (llist_add((struct llist_node *)addr, &p->list))
2740 schedule_work(&p->wq);
2744 * vfree_atomic - release memory allocated by vmalloc()
2745 * @addr: memory base address
2747 * This one is just like vfree() but can be called in any atomic context
2750 void vfree_atomic(const void *addr)
2754 kmemleak_free(addr);
2758 __vfree_deferred(addr);
2761 static void __vfree(const void *addr)
2763 if (unlikely(in_interrupt()))
2764 __vfree_deferred(addr);
2770 * vfree - Release memory allocated by vmalloc()
2771 * @addr: Memory base address
2773 * Free the virtually continuous memory area starting at @addr, as obtained
2774 * from one of the vmalloc() family of APIs. This will usually also free the
2775 * physical memory underlying the virtual allocation, but that memory is
2776 * reference counted, so it will not be freed until the last user goes away.
2778 * If @addr is NULL, no operation is performed.
2781 * May sleep if called *not* from interrupt context.
2782 * Must not be called in NMI context (strictly speaking, it could be
2783 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2784 * conventions for vfree() arch-dependent would be a really bad idea).
2786 void vfree(const void *addr)
2790 kmemleak_free(addr);
2792 might_sleep_if(!in_interrupt());
2799 EXPORT_SYMBOL(vfree);
2802 * vunmap - release virtual mapping obtained by vmap()
2803 * @addr: memory base address
2805 * Free the virtually contiguous memory area starting at @addr,
2806 * which was created from the page array passed to vmap().
2808 * Must not be called in interrupt context.
2810 void vunmap(const void *addr)
2812 BUG_ON(in_interrupt());
2817 EXPORT_SYMBOL(vunmap);
2820 * vmap - map an array of pages into virtually contiguous space
2821 * @pages: array of page pointers
2822 * @count: number of pages to map
2823 * @flags: vm_area->flags
2824 * @prot: page protection for the mapping
2826 * Maps @count pages from @pages into contiguous kernel virtual space.
2827 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2828 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2829 * are transferred from the caller to vmap(), and will be freed / dropped when
2830 * vfree() is called on the return value.
2832 * Return: the address of the area or %NULL on failure
2834 void *vmap(struct page **pages, unsigned int count,
2835 unsigned long flags, pgprot_t prot)
2837 struct vm_struct *area;
2839 unsigned long size; /* In bytes */
2844 * Your top guard is someone else's bottom guard. Not having a top
2845 * guard compromises someone else's mappings too.
2847 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2848 flags &= ~VM_NO_GUARD;
2850 if (count > totalram_pages())
2853 size = (unsigned long)count << PAGE_SHIFT;
2854 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2858 addr = (unsigned long)area->addr;
2859 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2860 pages, PAGE_SHIFT) < 0) {
2865 if (flags & VM_MAP_PUT_PAGES) {
2866 area->pages = pages;
2867 area->nr_pages = count;
2871 EXPORT_SYMBOL(vmap);
2873 #ifdef CONFIG_VMAP_PFN
2874 struct vmap_pfn_data {
2875 unsigned long *pfns;
2880 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2882 struct vmap_pfn_data *data = private;
2884 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2886 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2891 * vmap_pfn - map an array of PFNs into virtually contiguous space
2892 * @pfns: array of PFNs
2893 * @count: number of pages to map
2894 * @prot: page protection for the mapping
2896 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2897 * the start address of the mapping.
2899 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2901 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2902 struct vm_struct *area;
2904 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2905 __builtin_return_address(0));
2908 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2909 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2915 EXPORT_SYMBOL_GPL(vmap_pfn);
2916 #endif /* CONFIG_VMAP_PFN */
2918 static inline unsigned int
2919 vm_area_alloc_pages(gfp_t gfp, int nid,
2920 unsigned int order, unsigned int nr_pages, struct page **pages)
2922 unsigned int nr_allocated = 0;
2927 * For order-0 pages we make use of bulk allocator, if
2928 * the page array is partly or not at all populated due
2929 * to fails, fallback to a single page allocator that is
2933 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2935 while (nr_allocated < nr_pages) {
2936 unsigned int nr, nr_pages_request;
2939 * A maximum allowed request is hard-coded and is 100
2940 * pages per call. That is done in order to prevent a
2941 * long preemption off scenario in the bulk-allocator
2942 * so the range is [1:100].
2944 nr_pages_request = min(100U, nr_pages - nr_allocated);
2946 /* memory allocation should consider mempolicy, we can't
2947 * wrongly use nearest node when nid == NUMA_NO_NODE,
2948 * otherwise memory may be allocated in only one node,
2949 * but mempolicy wants to alloc memory by interleaving.
2951 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2952 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2954 pages + nr_allocated);
2957 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2959 pages + nr_allocated);
2965 * If zero or pages were obtained partly,
2966 * fallback to a single page allocator.
2968 if (nr != nr_pages_request)
2973 /* High-order pages or fallback path if "bulk" fails. */
2975 while (nr_allocated < nr_pages) {
2976 if (fatal_signal_pending(current))
2979 if (nid == NUMA_NO_NODE)
2980 page = alloc_pages(gfp, order);
2982 page = alloc_pages_node(nid, gfp, order);
2983 if (unlikely(!page))
2986 * Higher order allocations must be able to be treated as
2987 * indepdenent small pages by callers (as they can with
2988 * small-page vmallocs). Some drivers do their own refcounting
2989 * on vmalloc_to_page() pages, some use page->mapping,
2993 split_page(page, order);
2996 * Careful, we allocate and map page-order pages, but
2997 * tracking is done per PAGE_SIZE page so as to keep the
2998 * vm_struct APIs independent of the physical/mapped size.
3000 for (i = 0; i < (1U << order); i++)
3001 pages[nr_allocated + i] = page + i;
3004 nr_allocated += 1U << order;
3007 return nr_allocated;
3010 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3011 pgprot_t prot, unsigned int page_shift,
3014 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3015 bool nofail = gfp_mask & __GFP_NOFAIL;
3016 unsigned long addr = (unsigned long)area->addr;
3017 unsigned long size = get_vm_area_size(area);
3018 unsigned long array_size;
3019 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3020 unsigned int page_order;
3024 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3025 gfp_mask |= __GFP_NOWARN;
3026 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3027 gfp_mask |= __GFP_HIGHMEM;
3029 /* Please note that the recursion is strictly bounded. */
3030 if (array_size > PAGE_SIZE) {
3031 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3034 area->pages = kmalloc_node(array_size, nested_gfp, node);
3038 warn_alloc(gfp_mask, NULL,
3039 "vmalloc error: size %lu, failed to allocated page array size %lu",
3040 nr_small_pages * PAGE_SIZE, array_size);
3045 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3046 page_order = vm_area_page_order(area);
3048 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3049 node, page_order, nr_small_pages, area->pages);
3051 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3052 if (gfp_mask & __GFP_ACCOUNT) {
3055 for (i = 0; i < area->nr_pages; i++)
3056 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3060 * If not enough pages were obtained to accomplish an
3061 * allocation request, free them via __vfree() if any.
3063 if (area->nr_pages != nr_small_pages) {
3064 warn_alloc(gfp_mask, NULL,
3065 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3066 area->nr_pages * PAGE_SIZE, page_order);
3071 * page tables allocations ignore external gfp mask, enforce it
3074 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3075 flags = memalloc_nofs_save();
3076 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3077 flags = memalloc_noio_save();
3080 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3082 if (nofail && (ret < 0))
3083 schedule_timeout_uninterruptible(1);
3084 } while (nofail && (ret < 0));
3086 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3087 memalloc_nofs_restore(flags);
3088 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3089 memalloc_noio_restore(flags);
3092 warn_alloc(gfp_mask, NULL,
3093 "vmalloc error: size %lu, failed to map pages",
3094 area->nr_pages * PAGE_SIZE);
3101 __vfree(area->addr);
3106 * __vmalloc_node_range - allocate virtually contiguous memory
3107 * @size: allocation size
3108 * @align: desired alignment
3109 * @start: vm area range start
3110 * @end: vm area range end
3111 * @gfp_mask: flags for the page level allocator
3112 * @prot: protection mask for the allocated pages
3113 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3114 * @node: node to use for allocation or NUMA_NO_NODE
3115 * @caller: caller's return address
3117 * Allocate enough pages to cover @size from the page level
3118 * allocator with @gfp_mask flags. Please note that the full set of gfp
3119 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3121 * Zone modifiers are not supported. From the reclaim modifiers
3122 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3123 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3124 * __GFP_RETRY_MAYFAIL are not supported).
3126 * __GFP_NOWARN can be used to suppress failures messages.
3128 * Map them into contiguous kernel virtual space, using a pagetable
3129 * protection of @prot.
3131 * Return: the address of the area or %NULL on failure
3133 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3134 unsigned long start, unsigned long end, gfp_t gfp_mask,
3135 pgprot_t prot, unsigned long vm_flags, int node,
3138 struct vm_struct *area;
3140 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3141 unsigned long real_size = size;
3142 unsigned long real_align = align;
3143 unsigned int shift = PAGE_SHIFT;
3145 if (WARN_ON_ONCE(!size))
3148 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3149 warn_alloc(gfp_mask, NULL,
3150 "vmalloc error: size %lu, exceeds total pages",
3155 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3156 unsigned long size_per_node;
3159 * Try huge pages. Only try for PAGE_KERNEL allocations,
3160 * others like modules don't yet expect huge pages in
3161 * their allocations due to apply_to_page_range not
3165 size_per_node = size;
3166 if (node == NUMA_NO_NODE)
3167 size_per_node /= num_online_nodes();
3168 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3171 shift = arch_vmap_pte_supported_shift(size_per_node);
3173 align = max(real_align, 1UL << shift);
3174 size = ALIGN(real_size, 1UL << shift);
3178 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3179 VM_UNINITIALIZED | vm_flags, start, end, node,
3182 bool nofail = gfp_mask & __GFP_NOFAIL;
3183 warn_alloc(gfp_mask, NULL,
3184 "vmalloc error: size %lu, vm_struct allocation failed%s",
3185 real_size, (nofail) ? ". Retrying." : "");
3187 schedule_timeout_uninterruptible(1);
3194 * Prepare arguments for __vmalloc_area_node() and
3195 * kasan_unpoison_vmalloc().
3197 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3198 if (kasan_hw_tags_enabled()) {
3200 * Modify protection bits to allow tagging.
3201 * This must be done before mapping.
3203 prot = arch_vmap_pgprot_tagged(prot);
3206 * Skip page_alloc poisoning and zeroing for physical
3207 * pages backing VM_ALLOC mapping. Memory is instead
3208 * poisoned and zeroed by kasan_unpoison_vmalloc().
3210 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3213 /* Take note that the mapping is PAGE_KERNEL. */
3214 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3217 /* Allocate physical pages and map them into vmalloc space. */
3218 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3223 * Mark the pages as accessible, now that they are mapped.
3224 * The condition for setting KASAN_VMALLOC_INIT should complement the
3225 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3226 * to make sure that memory is initialized under the same conditions.
3227 * Tag-based KASAN modes only assign tags to normal non-executable
3228 * allocations, see __kasan_unpoison_vmalloc().
3230 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3231 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3232 (gfp_mask & __GFP_SKIP_ZERO))
3233 kasan_flags |= KASAN_VMALLOC_INIT;
3234 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3235 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3238 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3239 * flag. It means that vm_struct is not fully initialized.
3240 * Now, it is fully initialized, so remove this flag here.
3242 clear_vm_uninitialized_flag(area);
3244 size = PAGE_ALIGN(size);
3245 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3246 kmemleak_vmalloc(area, size, gfp_mask);
3251 if (shift > PAGE_SHIFT) {
3262 * __vmalloc_node - allocate virtually contiguous memory
3263 * @size: allocation size
3264 * @align: desired alignment
3265 * @gfp_mask: flags for the page level allocator
3266 * @node: node to use for allocation or NUMA_NO_NODE
3267 * @caller: caller's return address
3269 * Allocate enough pages to cover @size from the page level allocator with
3270 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3272 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3273 * and __GFP_NOFAIL are not supported
3275 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3278 * Return: pointer to the allocated memory or %NULL on error
3280 void *__vmalloc_node(unsigned long size, unsigned long align,
3281 gfp_t gfp_mask, int node, const void *caller)
3283 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3284 gfp_mask, PAGE_KERNEL, 0, node, caller);
3287 * This is only for performance analysis of vmalloc and stress purpose.
3288 * It is required by vmalloc test module, therefore do not use it other
3291 #ifdef CONFIG_TEST_VMALLOC_MODULE
3292 EXPORT_SYMBOL_GPL(__vmalloc_node);
3295 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3297 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3298 __builtin_return_address(0));
3300 EXPORT_SYMBOL(__vmalloc);
3303 * vmalloc - allocate virtually contiguous memory
3304 * @size: allocation size
3306 * Allocate enough pages to cover @size from the page level
3307 * allocator and map them into contiguous kernel virtual space.
3309 * For tight control over page level allocator and protection flags
3310 * use __vmalloc() instead.
3312 * Return: pointer to the allocated memory or %NULL on error
3314 void *vmalloc(unsigned long size)
3316 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3317 __builtin_return_address(0));
3319 EXPORT_SYMBOL(vmalloc);
3322 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3323 * @size: allocation size
3324 * @gfp_mask: flags for the page level allocator
3326 * Allocate enough pages to cover @size from the page level
3327 * allocator and map them into contiguous kernel virtual space.
3328 * If @size is greater than or equal to PMD_SIZE, allow using
3329 * huge pages for the memory
3331 * Return: pointer to the allocated memory or %NULL on error
3333 void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3335 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3336 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3337 NUMA_NO_NODE, __builtin_return_address(0));
3339 EXPORT_SYMBOL_GPL(vmalloc_huge);
3342 * vzalloc - allocate virtually contiguous memory with zero fill
3343 * @size: allocation size
3345 * Allocate enough pages to cover @size from the page level
3346 * allocator and map them into contiguous kernel virtual space.
3347 * The memory allocated is set to zero.
3349 * For tight control over page level allocator and protection flags
3350 * use __vmalloc() instead.
3352 * Return: pointer to the allocated memory or %NULL on error
3354 void *vzalloc(unsigned long size)
3356 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3357 __builtin_return_address(0));
3359 EXPORT_SYMBOL(vzalloc);
3362 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3363 * @size: allocation size
3365 * The resulting memory area is zeroed so it can be mapped to userspace
3366 * without leaking data.
3368 * Return: pointer to the allocated memory or %NULL on error
3370 void *vmalloc_user(unsigned long size)
3372 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3373 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3374 VM_USERMAP, NUMA_NO_NODE,
3375 __builtin_return_address(0));
3377 EXPORT_SYMBOL(vmalloc_user);
3380 * vmalloc_node - allocate memory on a specific node
3381 * @size: allocation size
3384 * Allocate enough pages to cover @size from the page level
3385 * allocator and map them into contiguous kernel virtual space.
3387 * For tight control over page level allocator and protection flags
3388 * use __vmalloc() instead.
3390 * Return: pointer to the allocated memory or %NULL on error
3392 void *vmalloc_node(unsigned long size, int node)
3394 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3395 __builtin_return_address(0));
3397 EXPORT_SYMBOL(vmalloc_node);
3400 * vzalloc_node - allocate memory on a specific node with zero fill
3401 * @size: allocation size
3404 * Allocate enough pages to cover @size from the page level
3405 * allocator and map them into contiguous kernel virtual space.
3406 * The memory allocated is set to zero.
3408 * Return: pointer to the allocated memory or %NULL on error
3410 void *vzalloc_node(unsigned long size, int node)
3412 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3413 __builtin_return_address(0));
3415 EXPORT_SYMBOL(vzalloc_node);
3417 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3418 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3419 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3420 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3423 * 64b systems should always have either DMA or DMA32 zones. For others
3424 * GFP_DMA32 should do the right thing and use the normal zone.
3426 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3430 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3431 * @size: allocation size
3433 * Allocate enough 32bit PA addressable pages to cover @size from the
3434 * page level allocator and map them into contiguous kernel virtual space.
3436 * Return: pointer to the allocated memory or %NULL on error
3438 void *vmalloc_32(unsigned long size)
3440 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3441 __builtin_return_address(0));
3443 EXPORT_SYMBOL(vmalloc_32);
3446 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3447 * @size: allocation size
3449 * The resulting memory area is 32bit addressable and zeroed so it can be
3450 * mapped to userspace without leaking data.
3452 * Return: pointer to the allocated memory or %NULL on error
3454 void *vmalloc_32_user(unsigned long size)
3456 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3457 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3458 VM_USERMAP, NUMA_NO_NODE,
3459 __builtin_return_address(0));
3461 EXPORT_SYMBOL(vmalloc_32_user);
3464 * small helper routine , copy contents to buf from addr.
3465 * If the page is not present, fill zero.
3468 static int aligned_vread(char *buf, char *addr, unsigned long count)
3474 unsigned long offset, length;
3476 offset = offset_in_page(addr);
3477 length = PAGE_SIZE - offset;
3480 p = vmalloc_to_page(addr);
3482 * To do safe access to this _mapped_ area, we need
3483 * lock. But adding lock here means that we need to add
3484 * overhead of vmalloc()/vfree() calls for this _debug_
3485 * interface, rarely used. Instead of that, we'll use
3486 * kmap() and get small overhead in this access function.
3489 /* We can expect USER0 is not used -- see vread() */
3490 void *map = kmap_atomic(p);
3491 memcpy(buf, map + offset, length);
3494 memset(buf, 0, length);
3505 * vread() - read vmalloc area in a safe way.
3506 * @buf: buffer for reading data
3507 * @addr: vm address.
3508 * @count: number of bytes to be read.
3510 * This function checks that addr is a valid vmalloc'ed area, and
3511 * copy data from that area to a given buffer. If the given memory range
3512 * of [addr...addr+count) includes some valid address, data is copied to
3513 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3514 * IOREMAP area is treated as memory hole and no copy is done.
3516 * If [addr...addr+count) doesn't includes any intersects with alive
3517 * vm_struct area, returns 0. @buf should be kernel's buffer.
3519 * Note: In usual ops, vread() is never necessary because the caller
3520 * should know vmalloc() area is valid and can use memcpy().
3521 * This is for routines which have to access vmalloc area without
3522 * any information, as /proc/kcore.
3524 * Return: number of bytes for which addr and buf should be increased
3525 * (same number as @count) or %0 if [addr...addr+count) doesn't
3526 * include any intersection with valid vmalloc area
3528 long vread(char *buf, char *addr, unsigned long count)
3530 struct vmap_area *va;
3531 struct vm_struct *vm;
3532 char *vaddr, *buf_start = buf;
3533 unsigned long buflen = count;
3536 addr = kasan_reset_tag(addr);
3538 /* Don't allow overflow */
3539 if ((unsigned long) addr + count < count)
3540 count = -(unsigned long) addr;
3542 spin_lock(&vmap_area_lock);
3543 va = find_vmap_area_exceed_addr((unsigned long)addr);
3547 /* no intersects with alive vmap_area */
3548 if ((unsigned long)addr + count <= va->va_start)
3551 list_for_each_entry_from(va, &vmap_area_list, list) {
3559 vaddr = (char *) vm->addr;
3560 if (addr >= vaddr + get_vm_area_size(vm))
3562 while (addr < vaddr) {
3570 n = vaddr + get_vm_area_size(vm) - addr;
3573 if (!(vm->flags & VM_IOREMAP))
3574 aligned_vread(buf, addr, n);
3575 else /* IOREMAP area is treated as memory hole */
3582 spin_unlock(&vmap_area_lock);
3584 if (buf == buf_start)
3586 /* zero-fill memory holes */
3587 if (buf != buf_start + buflen)
3588 memset(buf, 0, buflen - (buf - buf_start));
3594 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3595 * @vma: vma to cover
3596 * @uaddr: target user address to start at
3597 * @kaddr: virtual address of vmalloc kernel memory
3598 * @pgoff: offset from @kaddr to start at
3599 * @size: size of map area
3601 * Returns: 0 for success, -Exxx on failure
3603 * This function checks that @kaddr is a valid vmalloc'ed area,
3604 * and that it is big enough to cover the range starting at
3605 * @uaddr in @vma. Will return failure if that criteria isn't
3608 * Similar to remap_pfn_range() (see mm/memory.c)
3610 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3611 void *kaddr, unsigned long pgoff,
3614 struct vm_struct *area;
3616 unsigned long end_index;
3618 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3621 size = PAGE_ALIGN(size);
3623 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3626 area = find_vm_area(kaddr);
3630 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3633 if (check_add_overflow(size, off, &end_index) ||
3634 end_index > get_vm_area_size(area))
3639 struct page *page = vmalloc_to_page(kaddr);
3642 ret = vm_insert_page(vma, uaddr, page);
3651 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3657 * remap_vmalloc_range - map vmalloc pages to userspace
3658 * @vma: vma to cover (map full range of vma)
3659 * @addr: vmalloc memory
3660 * @pgoff: number of pages into addr before first page to map
3662 * Returns: 0 for success, -Exxx on failure
3664 * This function checks that addr is a valid vmalloc'ed area, and
3665 * that it is big enough to cover the vma. Will return failure if
3666 * that criteria isn't met.
3668 * Similar to remap_pfn_range() (see mm/memory.c)
3670 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3671 unsigned long pgoff)
3673 return remap_vmalloc_range_partial(vma, vma->vm_start,
3675 vma->vm_end - vma->vm_start);
3677 EXPORT_SYMBOL(remap_vmalloc_range);
3679 void free_vm_area(struct vm_struct *area)
3681 struct vm_struct *ret;
3682 ret = remove_vm_area(area->addr);
3683 BUG_ON(ret != area);
3686 EXPORT_SYMBOL_GPL(free_vm_area);
3689 static struct vmap_area *node_to_va(struct rb_node *n)
3691 return rb_entry_safe(n, struct vmap_area, rb_node);
3695 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3696 * @addr: target address
3698 * Returns: vmap_area if it is found. If there is no such area
3699 * the first highest(reverse order) vmap_area is returned
3700 * i.e. va->va_start < addr && va->va_end < addr or NULL
3701 * if there are no any areas before @addr.
3703 static struct vmap_area *
3704 pvm_find_va_enclose_addr(unsigned long addr)
3706 struct vmap_area *va, *tmp;
3709 n = free_vmap_area_root.rb_node;
3713 tmp = rb_entry(n, struct vmap_area, rb_node);
3714 if (tmp->va_start <= addr) {
3716 if (tmp->va_end >= addr)
3729 * pvm_determine_end_from_reverse - find the highest aligned address
3730 * of free block below VMALLOC_END
3732 * in - the VA we start the search(reverse order);
3733 * out - the VA with the highest aligned end address.
3734 * @align: alignment for required highest address
3736 * Returns: determined end address within vmap_area
3738 static unsigned long
3739 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3741 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3745 list_for_each_entry_from_reverse((*va),
3746 &free_vmap_area_list, list) {
3747 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3748 if ((*va)->va_start < addr)
3757 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3758 * @offsets: array containing offset of each area
3759 * @sizes: array containing size of each area
3760 * @nr_vms: the number of areas to allocate
3761 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3763 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3764 * vm_structs on success, %NULL on failure
3766 * Percpu allocator wants to use congruent vm areas so that it can
3767 * maintain the offsets among percpu areas. This function allocates
3768 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3769 * be scattered pretty far, distance between two areas easily going up
3770 * to gigabytes. To avoid interacting with regular vmallocs, these
3771 * areas are allocated from top.
3773 * Despite its complicated look, this allocator is rather simple. It
3774 * does everything top-down and scans free blocks from the end looking
3775 * for matching base. While scanning, if any of the areas do not fit the
3776 * base address is pulled down to fit the area. Scanning is repeated till
3777 * all the areas fit and then all necessary data structures are inserted
3778 * and the result is returned.
3780 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3781 const size_t *sizes, int nr_vms,
3784 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3785 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3786 struct vmap_area **vas, *va;
3787 struct vm_struct **vms;
3788 int area, area2, last_area, term_area;
3789 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3790 bool purged = false;
3792 /* verify parameters and allocate data structures */
3793 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3794 for (last_area = 0, area = 0; area < nr_vms; area++) {
3795 start = offsets[area];
3796 end = start + sizes[area];
3798 /* is everything aligned properly? */
3799 BUG_ON(!IS_ALIGNED(offsets[area], align));
3800 BUG_ON(!IS_ALIGNED(sizes[area], align));
3802 /* detect the area with the highest address */
3803 if (start > offsets[last_area])
3806 for (area2 = area + 1; area2 < nr_vms; area2++) {
3807 unsigned long start2 = offsets[area2];
3808 unsigned long end2 = start2 + sizes[area2];
3810 BUG_ON(start2 < end && start < end2);
3813 last_end = offsets[last_area] + sizes[last_area];
3815 if (vmalloc_end - vmalloc_start < last_end) {
3820 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3821 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3825 for (area = 0; area < nr_vms; area++) {
3826 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3827 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3828 if (!vas[area] || !vms[area])
3832 spin_lock(&free_vmap_area_lock);
3834 /* start scanning - we scan from the top, begin with the last area */
3835 area = term_area = last_area;
3836 start = offsets[area];
3837 end = start + sizes[area];
3839 va = pvm_find_va_enclose_addr(vmalloc_end);
3840 base = pvm_determine_end_from_reverse(&va, align) - end;
3844 * base might have underflowed, add last_end before
3847 if (base + last_end < vmalloc_start + last_end)
3851 * Fitting base has not been found.
3857 * If required width exceeds current VA block, move
3858 * base downwards and then recheck.
3860 if (base + end > va->va_end) {
3861 base = pvm_determine_end_from_reverse(&va, align) - end;
3867 * If this VA does not fit, move base downwards and recheck.
3869 if (base + start < va->va_start) {
3870 va = node_to_va(rb_prev(&va->rb_node));
3871 base = pvm_determine_end_from_reverse(&va, align) - end;
3877 * This area fits, move on to the previous one. If
3878 * the previous one is the terminal one, we're done.
3880 area = (area + nr_vms - 1) % nr_vms;
3881 if (area == term_area)
3884 start = offsets[area];
3885 end = start + sizes[area];
3886 va = pvm_find_va_enclose_addr(base + end);
3889 /* we've found a fitting base, insert all va's */
3890 for (area = 0; area < nr_vms; area++) {
3893 start = base + offsets[area];
3896 va = pvm_find_va_enclose_addr(start);
3897 if (WARN_ON_ONCE(va == NULL))
3898 /* It is a BUG(), but trigger recovery instead. */
3901 ret = adjust_va_to_fit_type(&free_vmap_area_root,
3902 &free_vmap_area_list,
3904 if (WARN_ON_ONCE(unlikely(ret)))
3905 /* It is a BUG(), but trigger recovery instead. */
3908 /* Allocated area. */
3910 va->va_start = start;
3911 va->va_end = start + size;
3914 spin_unlock(&free_vmap_area_lock);
3916 /* populate the kasan shadow space */
3917 for (area = 0; area < nr_vms; area++) {
3918 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3919 goto err_free_shadow;
3922 /* insert all vm's */
3923 spin_lock(&vmap_area_lock);
3924 for (area = 0; area < nr_vms; area++) {
3925 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3927 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3930 spin_unlock(&vmap_area_lock);
3933 * Mark allocated areas as accessible. Do it now as a best-effort
3934 * approach, as they can be mapped outside of vmalloc code.
3935 * With hardware tag-based KASAN, marking is skipped for
3936 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3938 for (area = 0; area < nr_vms; area++)
3939 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3940 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3947 * Remove previously allocated areas. There is no
3948 * need in removing these areas from the busy tree,
3949 * because they are inserted only on the final step
3950 * and when pcpu_get_vm_areas() is success.
3953 orig_start = vas[area]->va_start;
3954 orig_end = vas[area]->va_end;
3955 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3956 &free_vmap_area_list);
3958 kasan_release_vmalloc(orig_start, orig_end,
3959 va->va_start, va->va_end);
3964 spin_unlock(&free_vmap_area_lock);
3966 purge_vmap_area_lazy();
3969 /* Before "retry", check if we recover. */
3970 for (area = 0; area < nr_vms; area++) {
3974 vas[area] = kmem_cache_zalloc(
3975 vmap_area_cachep, GFP_KERNEL);
3984 for (area = 0; area < nr_vms; area++) {
3986 kmem_cache_free(vmap_area_cachep, vas[area]);
3996 spin_lock(&free_vmap_area_lock);
3998 * We release all the vmalloc shadows, even the ones for regions that
3999 * hadn't been successfully added. This relies on kasan_release_vmalloc
4000 * being able to tolerate this case.
4002 for (area = 0; area < nr_vms; area++) {
4003 orig_start = vas[area]->va_start;
4004 orig_end = vas[area]->va_end;
4005 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4006 &free_vmap_area_list);
4008 kasan_release_vmalloc(orig_start, orig_end,
4009 va->va_start, va->va_end);
4013 spin_unlock(&free_vmap_area_lock);
4020 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4021 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4022 * @nr_vms: the number of allocated areas
4024 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4026 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4030 for (i = 0; i < nr_vms; i++)
4031 free_vm_area(vms[i]);
4034 #endif /* CONFIG_SMP */
4036 #ifdef CONFIG_PRINTK
4037 bool vmalloc_dump_obj(void *object)
4039 struct vm_struct *vm;
4040 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4042 vm = find_vm_area(objp);
4045 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4046 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4051 #ifdef CONFIG_PROC_FS
4052 static void *s_start(struct seq_file *m, loff_t *pos)
4053 __acquires(&vmap_purge_lock)
4054 __acquires(&vmap_area_lock)
4056 mutex_lock(&vmap_purge_lock);
4057 spin_lock(&vmap_area_lock);
4059 return seq_list_start(&vmap_area_list, *pos);
4062 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4064 return seq_list_next(p, &vmap_area_list, pos);
4067 static void s_stop(struct seq_file *m, void *p)
4068 __releases(&vmap_area_lock)
4069 __releases(&vmap_purge_lock)
4071 spin_unlock(&vmap_area_lock);
4072 mutex_unlock(&vmap_purge_lock);
4075 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4077 if (IS_ENABLED(CONFIG_NUMA)) {
4078 unsigned int nr, *counters = m->private;
4079 unsigned int step = 1U << vm_area_page_order(v);
4084 if (v->flags & VM_UNINITIALIZED)
4086 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4089 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4091 for (nr = 0; nr < v->nr_pages; nr += step)
4092 counters[page_to_nid(v->pages[nr])] += step;
4093 for_each_node_state(nr, N_HIGH_MEMORY)
4095 seq_printf(m, " N%u=%u", nr, counters[nr]);
4099 static void show_purge_info(struct seq_file *m)
4101 struct vmap_area *va;
4103 spin_lock(&purge_vmap_area_lock);
4104 list_for_each_entry(va, &purge_vmap_area_list, list) {
4105 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4106 (void *)va->va_start, (void *)va->va_end,
4107 va->va_end - va->va_start);
4109 spin_unlock(&purge_vmap_area_lock);
4112 static int s_show(struct seq_file *m, void *p)
4114 struct vmap_area *va;
4115 struct vm_struct *v;
4117 va = list_entry(p, struct vmap_area, list);
4120 * s_show can encounter race with remove_vm_area, !vm on behalf
4121 * of vmap area is being tear down or vm_map_ram allocation.
4124 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4125 (void *)va->va_start, (void *)va->va_end,
4126 va->va_end - va->va_start);
4133 seq_printf(m, "0x%pK-0x%pK %7ld",
4134 v->addr, v->addr + v->size, v->size);
4137 seq_printf(m, " %pS", v->caller);
4140 seq_printf(m, " pages=%d", v->nr_pages);
4143 seq_printf(m, " phys=%pa", &v->phys_addr);
4145 if (v->flags & VM_IOREMAP)
4146 seq_puts(m, " ioremap");
4148 if (v->flags & VM_ALLOC)
4149 seq_puts(m, " vmalloc");
4151 if (v->flags & VM_MAP)
4152 seq_puts(m, " vmap");
4154 if (v->flags & VM_USERMAP)
4155 seq_puts(m, " user");
4157 if (v->flags & VM_DMA_COHERENT)
4158 seq_puts(m, " dma-coherent");
4160 if (is_vmalloc_addr(v->pages))
4161 seq_puts(m, " vpages");
4163 show_numa_info(m, v);
4167 * As a final step, dump "unpurged" areas.
4170 if (list_is_last(&va->list, &vmap_area_list))
4176 static const struct seq_operations vmalloc_op = {
4183 static int __init proc_vmalloc_init(void)
4185 if (IS_ENABLED(CONFIG_NUMA))
4186 proc_create_seq_private("vmallocinfo", 0400, NULL,
4188 nr_node_ids * sizeof(unsigned int), NULL);
4190 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4193 module_init(proc_vmalloc_init);