1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
47 #include "pgalloc-track.h"
49 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
50 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
52 static int __init set_nohugeiomap(char *str)
54 ioremap_max_page_shift = PAGE_SHIFT;
57 early_param("nohugeiomap", set_nohugeiomap);
58 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
59 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
60 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
63 static bool __ro_after_init vmap_allow_huge = true;
65 static int __init set_nohugevmalloc(char *str)
67 vmap_allow_huge = false;
70 early_param("nohugevmalloc", set_nohugevmalloc);
71 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
72 static const bool vmap_allow_huge = false;
73 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 bool is_vmalloc_addr(const void *x)
77 unsigned long addr = (unsigned long)kasan_reset_tag(x);
79 return addr >= VMALLOC_START && addr < VMALLOC_END;
81 EXPORT_SYMBOL(is_vmalloc_addr);
83 struct vfree_deferred {
84 struct llist_head list;
85 struct work_struct wq;
87 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
89 static void __vunmap(const void *, int);
91 static void free_work(struct work_struct *w)
93 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
94 struct llist_node *t, *llnode;
96 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
97 __vunmap((void *)llnode, 1);
100 /*** Page table manipulation functions ***/
101 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
102 phys_addr_t phys_addr, pgprot_t prot,
103 unsigned int max_page_shift, pgtbl_mod_mask *mask)
107 unsigned long size = PAGE_SIZE;
109 pfn = phys_addr >> PAGE_SHIFT;
110 pte = pte_alloc_kernel_track(pmd, addr, mask);
114 BUG_ON(!pte_none(*pte));
116 #ifdef CONFIG_HUGETLB_PAGE
117 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
118 if (size != PAGE_SIZE) {
119 pte_t entry = pfn_pte(pfn, prot);
121 entry = arch_make_huge_pte(entry, ilog2(size), 0);
122 set_huge_pte_at(&init_mm, addr, pte, entry);
123 pfn += PFN_DOWN(size);
127 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
129 } while (pte += PFN_DOWN(size), addr += size, addr != end);
130 *mask |= PGTBL_PTE_MODIFIED;
134 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
135 phys_addr_t phys_addr, pgprot_t prot,
136 unsigned int max_page_shift)
138 if (max_page_shift < PMD_SHIFT)
141 if (!arch_vmap_pmd_supported(prot))
144 if ((end - addr) != PMD_SIZE)
147 if (!IS_ALIGNED(addr, PMD_SIZE))
150 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
153 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
156 return pmd_set_huge(pmd, phys_addr, prot);
159 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
160 phys_addr_t phys_addr, pgprot_t prot,
161 unsigned int max_page_shift, pgtbl_mod_mask *mask)
166 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
170 next = pmd_addr_end(addr, end);
172 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
174 *mask |= PGTBL_PMD_MODIFIED;
178 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
180 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
184 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
185 phys_addr_t phys_addr, pgprot_t prot,
186 unsigned int max_page_shift)
188 if (max_page_shift < PUD_SHIFT)
191 if (!arch_vmap_pud_supported(prot))
194 if ((end - addr) != PUD_SIZE)
197 if (!IS_ALIGNED(addr, PUD_SIZE))
200 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
203 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
206 return pud_set_huge(pud, phys_addr, prot);
209 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
210 phys_addr_t phys_addr, pgprot_t prot,
211 unsigned int max_page_shift, pgtbl_mod_mask *mask)
216 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
220 next = pud_addr_end(addr, end);
222 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
224 *mask |= PGTBL_PUD_MODIFIED;
228 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
229 max_page_shift, mask))
231 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
235 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
236 phys_addr_t phys_addr, pgprot_t prot,
237 unsigned int max_page_shift)
239 if (max_page_shift < P4D_SHIFT)
242 if (!arch_vmap_p4d_supported(prot))
245 if ((end - addr) != P4D_SIZE)
248 if (!IS_ALIGNED(addr, P4D_SIZE))
251 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
254 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
257 return p4d_set_huge(p4d, phys_addr, prot);
260 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
261 phys_addr_t phys_addr, pgprot_t prot,
262 unsigned int max_page_shift, pgtbl_mod_mask *mask)
267 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
271 next = p4d_addr_end(addr, end);
273 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
275 *mask |= PGTBL_P4D_MODIFIED;
279 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
280 max_page_shift, mask))
282 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
286 static int vmap_range_noflush(unsigned long addr, unsigned long end,
287 phys_addr_t phys_addr, pgprot_t prot,
288 unsigned int max_page_shift)
294 pgtbl_mod_mask mask = 0;
300 pgd = pgd_offset_k(addr);
302 next = pgd_addr_end(addr, end);
303 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
304 max_page_shift, &mask);
307 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
309 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
310 arch_sync_kernel_mappings(start, end);
315 int ioremap_page_range(unsigned long addr, unsigned long end,
316 phys_addr_t phys_addr, pgprot_t prot)
320 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
321 ioremap_max_page_shift);
322 flush_cache_vmap(addr, end);
326 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
327 pgtbl_mod_mask *mask)
331 pte = pte_offset_kernel(pmd, addr);
333 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
334 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
335 } while (pte++, addr += PAGE_SIZE, addr != end);
336 *mask |= PGTBL_PTE_MODIFIED;
339 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
340 pgtbl_mod_mask *mask)
346 pmd = pmd_offset(pud, addr);
348 next = pmd_addr_end(addr, end);
350 cleared = pmd_clear_huge(pmd);
351 if (cleared || pmd_bad(*pmd))
352 *mask |= PGTBL_PMD_MODIFIED;
356 if (pmd_none_or_clear_bad(pmd))
358 vunmap_pte_range(pmd, addr, next, mask);
361 } while (pmd++, addr = next, addr != end);
364 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
365 pgtbl_mod_mask *mask)
371 pud = pud_offset(p4d, addr);
373 next = pud_addr_end(addr, end);
375 cleared = pud_clear_huge(pud);
376 if (cleared || pud_bad(*pud))
377 *mask |= PGTBL_PUD_MODIFIED;
381 if (pud_none_or_clear_bad(pud))
383 vunmap_pmd_range(pud, addr, next, mask);
384 } while (pud++, addr = next, addr != end);
387 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
388 pgtbl_mod_mask *mask)
393 p4d = p4d_offset(pgd, addr);
395 next = p4d_addr_end(addr, end);
399 *mask |= PGTBL_P4D_MODIFIED;
401 if (p4d_none_or_clear_bad(p4d))
403 vunmap_pud_range(p4d, addr, next, mask);
404 } while (p4d++, addr = next, addr != end);
408 * vunmap_range_noflush is similar to vunmap_range, but does not
409 * flush caches or TLBs.
411 * The caller is responsible for calling flush_cache_vmap() before calling
412 * this function, and flush_tlb_kernel_range after it has returned
413 * successfully (and before the addresses are expected to cause a page fault
414 * or be re-mapped for something else, if TLB flushes are being delayed or
417 * This is an internal function only. Do not use outside mm/.
419 void vunmap_range_noflush(unsigned long start, unsigned long end)
423 unsigned long addr = start;
424 pgtbl_mod_mask mask = 0;
427 pgd = pgd_offset_k(addr);
429 next = pgd_addr_end(addr, end);
431 mask |= PGTBL_PGD_MODIFIED;
432 if (pgd_none_or_clear_bad(pgd))
434 vunmap_p4d_range(pgd, addr, next, &mask);
435 } while (pgd++, addr = next, addr != end);
437 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
438 arch_sync_kernel_mappings(start, end);
442 * vunmap_range - unmap kernel virtual addresses
443 * @addr: start of the VM area to unmap
444 * @end: end of the VM area to unmap (non-inclusive)
446 * Clears any present PTEs in the virtual address range, flushes TLBs and
447 * caches. Any subsequent access to the address before it has been re-mapped
450 void vunmap_range(unsigned long addr, unsigned long end)
452 flush_cache_vunmap(addr, end);
453 vunmap_range_noflush(addr, end);
454 flush_tlb_kernel_range(addr, end);
457 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
458 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
459 pgtbl_mod_mask *mask)
464 * nr is a running index into the array which helps higher level
465 * callers keep track of where we're up to.
468 pte = pte_alloc_kernel_track(pmd, addr, mask);
472 struct page *page = pages[*nr];
474 if (WARN_ON(!pte_none(*pte)))
478 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
481 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
483 } while (pte++, addr += PAGE_SIZE, addr != end);
484 *mask |= PGTBL_PTE_MODIFIED;
488 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
489 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
490 pgtbl_mod_mask *mask)
495 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
499 next = pmd_addr_end(addr, end);
500 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
502 } while (pmd++, addr = next, addr != end);
506 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
507 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
508 pgtbl_mod_mask *mask)
513 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
517 next = pud_addr_end(addr, end);
518 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
520 } while (pud++, addr = next, addr != end);
524 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
525 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
526 pgtbl_mod_mask *mask)
531 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
535 next = p4d_addr_end(addr, end);
536 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
538 } while (p4d++, addr = next, addr != end);
542 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
543 pgprot_t prot, struct page **pages)
545 unsigned long start = addr;
550 pgtbl_mod_mask mask = 0;
553 pgd = pgd_offset_k(addr);
555 next = pgd_addr_end(addr, end);
557 mask |= PGTBL_PGD_MODIFIED;
558 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
561 } while (pgd++, addr = next, addr != end);
563 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
564 arch_sync_kernel_mappings(start, end);
570 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
573 * The caller is responsible for calling flush_cache_vmap() after this
574 * function returns successfully and before the addresses are accessed.
576 * This is an internal function only. Do not use outside mm/.
578 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
579 pgprot_t prot, struct page **pages, unsigned int page_shift)
581 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
583 WARN_ON(page_shift < PAGE_SHIFT);
585 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
586 page_shift == PAGE_SHIFT)
587 return vmap_small_pages_range_noflush(addr, end, prot, pages);
589 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
592 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
593 __pa(page_address(pages[i])), prot,
598 addr += 1UL << page_shift;
605 * vmap_pages_range - map pages to a kernel virtual address
606 * @addr: start of the VM area to map
607 * @end: end of the VM area to map (non-inclusive)
608 * @prot: page protection flags to use
609 * @pages: pages to map (always PAGE_SIZE pages)
610 * @page_shift: maximum shift that the pages may be mapped with, @pages must
611 * be aligned and contiguous up to at least this shift.
614 * 0 on success, -errno on failure.
616 static int vmap_pages_range(unsigned long addr, unsigned long end,
617 pgprot_t prot, struct page **pages, unsigned int page_shift)
621 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
622 flush_cache_vmap(addr, end);
626 int is_vmalloc_or_module_addr(const void *x)
629 * ARM, x86-64 and sparc64 put modules in a special place,
630 * and fall back on vmalloc() if that fails. Others
631 * just put it in the vmalloc space.
633 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
634 unsigned long addr = (unsigned long)kasan_reset_tag(x);
635 if (addr >= MODULES_VADDR && addr < MODULES_END)
638 return is_vmalloc_addr(x);
642 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
643 * return the tail page that corresponds to the base page address, which
644 * matches small vmap mappings.
646 struct page *vmalloc_to_page(const void *vmalloc_addr)
648 unsigned long addr = (unsigned long) vmalloc_addr;
649 struct page *page = NULL;
650 pgd_t *pgd = pgd_offset_k(addr);
657 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
658 * architectures that do not vmalloc module space
660 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
664 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
665 return NULL; /* XXX: no allowance for huge pgd */
666 if (WARN_ON_ONCE(pgd_bad(*pgd)))
669 p4d = p4d_offset(pgd, addr);
673 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
674 if (WARN_ON_ONCE(p4d_bad(*p4d)))
677 pud = pud_offset(p4d, addr);
681 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
682 if (WARN_ON_ONCE(pud_bad(*pud)))
685 pmd = pmd_offset(pud, addr);
689 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
690 if (WARN_ON_ONCE(pmd_bad(*pmd)))
693 ptep = pte_offset_map(pmd, addr);
695 if (pte_present(pte))
696 page = pte_page(pte);
701 EXPORT_SYMBOL(vmalloc_to_page);
704 * Map a vmalloc()-space virtual address to the physical page frame number.
706 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
708 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
710 EXPORT_SYMBOL(vmalloc_to_pfn);
713 /*** Global kva allocator ***/
715 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
716 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
719 static DEFINE_SPINLOCK(vmap_area_lock);
720 static DEFINE_SPINLOCK(free_vmap_area_lock);
721 /* Export for kexec only */
722 LIST_HEAD(vmap_area_list);
723 static struct rb_root vmap_area_root = RB_ROOT;
724 static bool vmap_initialized __read_mostly;
726 static struct rb_root purge_vmap_area_root = RB_ROOT;
727 static LIST_HEAD(purge_vmap_area_list);
728 static DEFINE_SPINLOCK(purge_vmap_area_lock);
731 * This kmem_cache is used for vmap_area objects. Instead of
732 * allocating from slab we reuse an object from this cache to
733 * make things faster. Especially in "no edge" splitting of
736 static struct kmem_cache *vmap_area_cachep;
739 * This linked list is used in pair with free_vmap_area_root.
740 * It gives O(1) access to prev/next to perform fast coalescing.
742 static LIST_HEAD(free_vmap_area_list);
745 * This augment red-black tree represents the free vmap space.
746 * All vmap_area objects in this tree are sorted by va->va_start
747 * address. It is used for allocation and merging when a vmap
748 * object is released.
750 * Each vmap_area node contains a maximum available free block
751 * of its sub-tree, right or left. Therefore it is possible to
752 * find a lowest match of free area.
754 static struct rb_root free_vmap_area_root = RB_ROOT;
757 * Preload a CPU with one object for "no edge" split case. The
758 * aim is to get rid of allocations from the atomic context, thus
759 * to use more permissive allocation masks.
761 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
763 static __always_inline unsigned long
764 va_size(struct vmap_area *va)
766 return (va->va_end - va->va_start);
769 static __always_inline unsigned long
770 get_subtree_max_size(struct rb_node *node)
772 struct vmap_area *va;
774 va = rb_entry_safe(node, struct vmap_area, rb_node);
775 return va ? va->subtree_max_size : 0;
778 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
779 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
781 static void purge_vmap_area_lazy(void);
782 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
783 static void drain_vmap_area_work(struct work_struct *work);
784 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
786 static atomic_long_t nr_vmalloc_pages;
788 unsigned long vmalloc_nr_pages(void)
790 return atomic_long_read(&nr_vmalloc_pages);
793 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
795 struct vmap_area *va = NULL;
796 struct rb_node *n = vmap_area_root.rb_node;
798 addr = (unsigned long)kasan_reset_tag((void *)addr);
801 struct vmap_area *tmp;
803 tmp = rb_entry(n, struct vmap_area, rb_node);
804 if (tmp->va_end > addr) {
806 if (tmp->va_start <= addr)
817 static struct vmap_area *__find_vmap_area(unsigned long addr)
819 struct rb_node *n = vmap_area_root.rb_node;
821 addr = (unsigned long)kasan_reset_tag((void *)addr);
824 struct vmap_area *va;
826 va = rb_entry(n, struct vmap_area, rb_node);
827 if (addr < va->va_start)
829 else if (addr >= va->va_end)
839 * This function returns back addresses of parent node
840 * and its left or right link for further processing.
842 * Otherwise NULL is returned. In that case all further
843 * steps regarding inserting of conflicting overlap range
844 * have to be declined and actually considered as a bug.
846 static __always_inline struct rb_node **
847 find_va_links(struct vmap_area *va,
848 struct rb_root *root, struct rb_node *from,
849 struct rb_node **parent)
851 struct vmap_area *tmp_va;
852 struct rb_node **link;
855 link = &root->rb_node;
856 if (unlikely(!*link)) {
865 * Go to the bottom of the tree. When we hit the last point
866 * we end up with parent rb_node and correct direction, i name
867 * it link, where the new va->rb_node will be attached to.
870 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
873 * During the traversal we also do some sanity check.
874 * Trigger the BUG() if there are sides(left/right)
877 if (va->va_start < tmp_va->va_end &&
878 va->va_end <= tmp_va->va_start)
879 link = &(*link)->rb_left;
880 else if (va->va_end > tmp_va->va_start &&
881 va->va_start >= tmp_va->va_end)
882 link = &(*link)->rb_right;
884 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
885 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
891 *parent = &tmp_va->rb_node;
895 static __always_inline struct list_head *
896 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
898 struct list_head *list;
900 if (unlikely(!parent))
902 * The red-black tree where we try to find VA neighbors
903 * before merging or inserting is empty, i.e. it means
904 * there is no free vmap space. Normally it does not
905 * happen but we handle this case anyway.
909 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
910 return (&parent->rb_right == link ? list->next : list);
913 static __always_inline void
914 link_va(struct vmap_area *va, struct rb_root *root,
915 struct rb_node *parent, struct rb_node **link, struct list_head *head)
918 * VA is still not in the list, but we can
919 * identify its future previous list_head node.
921 if (likely(parent)) {
922 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
923 if (&parent->rb_right != link)
927 /* Insert to the rb-tree */
928 rb_link_node(&va->rb_node, parent, link);
929 if (root == &free_vmap_area_root) {
931 * Some explanation here. Just perform simple insertion
932 * to the tree. We do not set va->subtree_max_size to
933 * its current size before calling rb_insert_augmented().
934 * It is because of we populate the tree from the bottom
935 * to parent levels when the node _is_ in the tree.
937 * Therefore we set subtree_max_size to zero after insertion,
938 * to let __augment_tree_propagate_from() puts everything to
939 * the correct order later on.
941 rb_insert_augmented(&va->rb_node,
942 root, &free_vmap_area_rb_augment_cb);
943 va->subtree_max_size = 0;
945 rb_insert_color(&va->rb_node, root);
948 /* Address-sort this list */
949 list_add(&va->list, head);
952 static __always_inline void
953 unlink_va(struct vmap_area *va, struct rb_root *root)
955 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
958 if (root == &free_vmap_area_root)
959 rb_erase_augmented(&va->rb_node,
960 root, &free_vmap_area_rb_augment_cb);
962 rb_erase(&va->rb_node, root);
965 RB_CLEAR_NODE(&va->rb_node);
968 #if DEBUG_AUGMENT_PROPAGATE_CHECK
970 * Gets called when remove the node and rotate.
972 static __always_inline unsigned long
973 compute_subtree_max_size(struct vmap_area *va)
975 return max3(va_size(va),
976 get_subtree_max_size(va->rb_node.rb_left),
977 get_subtree_max_size(va->rb_node.rb_right));
981 augment_tree_propagate_check(void)
983 struct vmap_area *va;
984 unsigned long computed_size;
986 list_for_each_entry(va, &free_vmap_area_list, list) {
987 computed_size = compute_subtree_max_size(va);
988 if (computed_size != va->subtree_max_size)
989 pr_emerg("tree is corrupted: %lu, %lu\n",
990 va_size(va), va->subtree_max_size);
996 * This function populates subtree_max_size from bottom to upper
997 * levels starting from VA point. The propagation must be done
998 * when VA size is modified by changing its va_start/va_end. Or
999 * in case of newly inserting of VA to the tree.
1001 * It means that __augment_tree_propagate_from() must be called:
1002 * - After VA has been inserted to the tree(free path);
1003 * - After VA has been shrunk(allocation path);
1004 * - After VA has been increased(merging path).
1006 * Please note that, it does not mean that upper parent nodes
1007 * and their subtree_max_size are recalculated all the time up
1016 * For example if we modify the node 4, shrinking it to 2, then
1017 * no any modification is required. If we shrink the node 2 to 1
1018 * its subtree_max_size is updated only, and set to 1. If we shrink
1019 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1020 * node becomes 4--6.
1022 static __always_inline void
1023 augment_tree_propagate_from(struct vmap_area *va)
1026 * Populate the tree from bottom towards the root until
1027 * the calculated maximum available size of checked node
1028 * is equal to its current one.
1030 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1032 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1033 augment_tree_propagate_check();
1038 insert_vmap_area(struct vmap_area *va,
1039 struct rb_root *root, struct list_head *head)
1041 struct rb_node **link;
1042 struct rb_node *parent;
1044 link = find_va_links(va, root, NULL, &parent);
1046 link_va(va, root, parent, link, head);
1050 insert_vmap_area_augment(struct vmap_area *va,
1051 struct rb_node *from, struct rb_root *root,
1052 struct list_head *head)
1054 struct rb_node **link;
1055 struct rb_node *parent;
1058 link = find_va_links(va, NULL, from, &parent);
1060 link = find_va_links(va, root, NULL, &parent);
1063 link_va(va, root, parent, link, head);
1064 augment_tree_propagate_from(va);
1069 * Merge de-allocated chunk of VA memory with previous
1070 * and next free blocks. If coalesce is not done a new
1071 * free area is inserted. If VA has been merged, it is
1074 * Please note, it can return NULL in case of overlap
1075 * ranges, followed by WARN() report. Despite it is a
1076 * buggy behaviour, a system can be alive and keep
1079 static __always_inline struct vmap_area *
1080 merge_or_add_vmap_area(struct vmap_area *va,
1081 struct rb_root *root, struct list_head *head)
1083 struct vmap_area *sibling;
1084 struct list_head *next;
1085 struct rb_node **link;
1086 struct rb_node *parent;
1087 bool merged = false;
1090 * Find a place in the tree where VA potentially will be
1091 * inserted, unless it is merged with its sibling/siblings.
1093 link = find_va_links(va, root, NULL, &parent);
1098 * Get next node of VA to check if merging can be done.
1100 next = get_va_next_sibling(parent, link);
1101 if (unlikely(next == NULL))
1107 * |<------VA------>|<-----Next----->|
1112 sibling = list_entry(next, struct vmap_area, list);
1113 if (sibling->va_start == va->va_end) {
1114 sibling->va_start = va->va_start;
1116 /* Free vmap_area object. */
1117 kmem_cache_free(vmap_area_cachep, va);
1119 /* Point to the new merged area. */
1128 * |<-----Prev----->|<------VA------>|
1132 if (next->prev != head) {
1133 sibling = list_entry(next->prev, struct vmap_area, list);
1134 if (sibling->va_end == va->va_start) {
1136 * If both neighbors are coalesced, it is important
1137 * to unlink the "next" node first, followed by merging
1138 * with "previous" one. Otherwise the tree might not be
1139 * fully populated if a sibling's augmented value is
1140 * "normalized" because of rotation operations.
1143 unlink_va(va, root);
1145 sibling->va_end = va->va_end;
1147 /* Free vmap_area object. */
1148 kmem_cache_free(vmap_area_cachep, va);
1150 /* Point to the new merged area. */
1158 link_va(va, root, parent, link, head);
1163 static __always_inline struct vmap_area *
1164 merge_or_add_vmap_area_augment(struct vmap_area *va,
1165 struct rb_root *root, struct list_head *head)
1167 va = merge_or_add_vmap_area(va, root, head);
1169 augment_tree_propagate_from(va);
1174 static __always_inline bool
1175 is_within_this_va(struct vmap_area *va, unsigned long size,
1176 unsigned long align, unsigned long vstart)
1178 unsigned long nva_start_addr;
1180 if (va->va_start > vstart)
1181 nva_start_addr = ALIGN(va->va_start, align);
1183 nva_start_addr = ALIGN(vstart, align);
1185 /* Can be overflowed due to big size or alignment. */
1186 if (nva_start_addr + size < nva_start_addr ||
1187 nva_start_addr < vstart)
1190 return (nva_start_addr + size <= va->va_end);
1194 * Find the first free block(lowest start address) in the tree,
1195 * that will accomplish the request corresponding to passing
1196 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1197 * a search length is adjusted to account for worst case alignment
1200 static __always_inline struct vmap_area *
1201 find_vmap_lowest_match(unsigned long size, unsigned long align,
1202 unsigned long vstart, bool adjust_search_size)
1204 struct vmap_area *va;
1205 struct rb_node *node;
1206 unsigned long length;
1208 /* Start from the root. */
1209 node = free_vmap_area_root.rb_node;
1211 /* Adjust the search size for alignment overhead. */
1212 length = adjust_search_size ? size + align - 1 : size;
1215 va = rb_entry(node, struct vmap_area, rb_node);
1217 if (get_subtree_max_size(node->rb_left) >= length &&
1218 vstart < va->va_start) {
1219 node = node->rb_left;
1221 if (is_within_this_va(va, size, align, vstart))
1225 * Does not make sense to go deeper towards the right
1226 * sub-tree if it does not have a free block that is
1227 * equal or bigger to the requested search length.
1229 if (get_subtree_max_size(node->rb_right) >= length) {
1230 node = node->rb_right;
1235 * OK. We roll back and find the first right sub-tree,
1236 * that will satisfy the search criteria. It can happen
1237 * due to "vstart" restriction or an alignment overhead
1238 * that is bigger then PAGE_SIZE.
1240 while ((node = rb_parent(node))) {
1241 va = rb_entry(node, struct vmap_area, rb_node);
1242 if (is_within_this_va(va, size, align, vstart))
1245 if (get_subtree_max_size(node->rb_right) >= length &&
1246 vstart <= va->va_start) {
1248 * Shift the vstart forward. Please note, we update it with
1249 * parent's start address adding "1" because we do not want
1250 * to enter same sub-tree after it has already been checked
1251 * and no suitable free block found there.
1253 vstart = va->va_start + 1;
1254 node = node->rb_right;
1264 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1265 #include <linux/random.h>
1267 static struct vmap_area *
1268 find_vmap_lowest_linear_match(unsigned long size,
1269 unsigned long align, unsigned long vstart)
1271 struct vmap_area *va;
1273 list_for_each_entry(va, &free_vmap_area_list, list) {
1274 if (!is_within_this_va(va, size, align, vstart))
1284 find_vmap_lowest_match_check(unsigned long size, unsigned long align)
1286 struct vmap_area *va_1, *va_2;
1287 unsigned long vstart;
1290 get_random_bytes(&rnd, sizeof(rnd));
1291 vstart = VMALLOC_START + rnd;
1293 va_1 = find_vmap_lowest_match(size, align, vstart, false);
1294 va_2 = find_vmap_lowest_linear_match(size, align, vstart);
1297 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1298 va_1, va_2, vstart);
1304 FL_FIT_TYPE = 1, /* full fit */
1305 LE_FIT_TYPE = 2, /* left edge fit */
1306 RE_FIT_TYPE = 3, /* right edge fit */
1307 NE_FIT_TYPE = 4 /* no edge fit */
1310 static __always_inline enum fit_type
1311 classify_va_fit_type(struct vmap_area *va,
1312 unsigned long nva_start_addr, unsigned long size)
1316 /* Check if it is within VA. */
1317 if (nva_start_addr < va->va_start ||
1318 nva_start_addr + size > va->va_end)
1322 if (va->va_start == nva_start_addr) {
1323 if (va->va_end == nva_start_addr + size)
1327 } else if (va->va_end == nva_start_addr + size) {
1336 static __always_inline int
1337 adjust_va_to_fit_type(struct vmap_area *va,
1338 unsigned long nva_start_addr, unsigned long size,
1341 struct vmap_area *lva = NULL;
1343 if (type == FL_FIT_TYPE) {
1345 * No need to split VA, it fully fits.
1351 unlink_va(va, &free_vmap_area_root);
1352 kmem_cache_free(vmap_area_cachep, va);
1353 } else if (type == LE_FIT_TYPE) {
1355 * Split left edge of fit VA.
1361 va->va_start += size;
1362 } else if (type == RE_FIT_TYPE) {
1364 * Split right edge of fit VA.
1370 va->va_end = nva_start_addr;
1371 } else if (type == NE_FIT_TYPE) {
1373 * Split no edge of fit VA.
1379 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1380 if (unlikely(!lva)) {
1382 * For percpu allocator we do not do any pre-allocation
1383 * and leave it as it is. The reason is it most likely
1384 * never ends up with NE_FIT_TYPE splitting. In case of
1385 * percpu allocations offsets and sizes are aligned to
1386 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1387 * are its main fitting cases.
1389 * There are a few exceptions though, as an example it is
1390 * a first allocation (early boot up) when we have "one"
1391 * big free space that has to be split.
1393 * Also we can hit this path in case of regular "vmap"
1394 * allocations, if "this" current CPU was not preloaded.
1395 * See the comment in alloc_vmap_area() why. If so, then
1396 * GFP_NOWAIT is used instead to get an extra object for
1397 * split purpose. That is rare and most time does not
1400 * What happens if an allocation gets failed. Basically,
1401 * an "overflow" path is triggered to purge lazily freed
1402 * areas to free some memory, then, the "retry" path is
1403 * triggered to repeat one more time. See more details
1404 * in alloc_vmap_area() function.
1406 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1412 * Build the remainder.
1414 lva->va_start = va->va_start;
1415 lva->va_end = nva_start_addr;
1418 * Shrink this VA to remaining size.
1420 va->va_start = nva_start_addr + size;
1425 if (type != FL_FIT_TYPE) {
1426 augment_tree_propagate_from(va);
1428 if (lva) /* type == NE_FIT_TYPE */
1429 insert_vmap_area_augment(lva, &va->rb_node,
1430 &free_vmap_area_root, &free_vmap_area_list);
1437 * Returns a start address of the newly allocated area, if success.
1438 * Otherwise a vend is returned that indicates failure.
1440 static __always_inline unsigned long
1441 __alloc_vmap_area(unsigned long size, unsigned long align,
1442 unsigned long vstart, unsigned long vend)
1444 bool adjust_search_size = true;
1445 unsigned long nva_start_addr;
1446 struct vmap_area *va;
1451 * Do not adjust when:
1452 * a) align <= PAGE_SIZE, because it does not make any sense.
1453 * All blocks(their start addresses) are at least PAGE_SIZE
1455 * b) a short range where a requested size corresponds to exactly
1456 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1457 * With adjusted search length an allocation would not succeed.
1459 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1460 adjust_search_size = false;
1462 va = find_vmap_lowest_match(size, align, vstart, adjust_search_size);
1466 if (va->va_start > vstart)
1467 nva_start_addr = ALIGN(va->va_start, align);
1469 nva_start_addr = ALIGN(vstart, align);
1471 /* Check the "vend" restriction. */
1472 if (nva_start_addr + size > vend)
1475 /* Classify what we have found. */
1476 type = classify_va_fit_type(va, nva_start_addr, size);
1477 if (WARN_ON_ONCE(type == NOTHING_FIT))
1480 /* Update the free vmap_area. */
1481 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1485 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1486 find_vmap_lowest_match_check(size, align);
1489 return nva_start_addr;
1493 * Free a region of KVA allocated by alloc_vmap_area
1495 static void free_vmap_area(struct vmap_area *va)
1498 * Remove from the busy tree/list.
1500 spin_lock(&vmap_area_lock);
1501 unlink_va(va, &vmap_area_root);
1502 spin_unlock(&vmap_area_lock);
1505 * Insert/Merge it back to the free tree/list.
1507 spin_lock(&free_vmap_area_lock);
1508 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1509 spin_unlock(&free_vmap_area_lock);
1513 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1515 struct vmap_area *va = NULL;
1518 * Preload this CPU with one extra vmap_area object. It is used
1519 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1520 * a CPU that does an allocation is preloaded.
1522 * We do it in non-atomic context, thus it allows us to use more
1523 * permissive allocation masks to be more stable under low memory
1524 * condition and high memory pressure.
1526 if (!this_cpu_read(ne_fit_preload_node))
1527 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1531 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1532 kmem_cache_free(vmap_area_cachep, va);
1536 * Allocate a region of KVA of the specified size and alignment, within the
1539 static struct vmap_area *alloc_vmap_area(unsigned long size,
1540 unsigned long align,
1541 unsigned long vstart, unsigned long vend,
1542 int node, gfp_t gfp_mask)
1544 struct vmap_area *va;
1545 unsigned long freed;
1551 BUG_ON(offset_in_page(size));
1552 BUG_ON(!is_power_of_2(align));
1554 if (unlikely(!vmap_initialized))
1555 return ERR_PTR(-EBUSY);
1558 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1560 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1562 return ERR_PTR(-ENOMEM);
1565 * Only scan the relevant parts containing pointers to other objects
1566 * to avoid false negatives.
1568 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1571 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1572 addr = __alloc_vmap_area(size, align, vstart, vend);
1573 spin_unlock(&free_vmap_area_lock);
1576 * If an allocation fails, the "vend" address is
1577 * returned. Therefore trigger the overflow path.
1579 if (unlikely(addr == vend))
1582 va->va_start = addr;
1583 va->va_end = addr + size;
1586 spin_lock(&vmap_area_lock);
1587 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1588 spin_unlock(&vmap_area_lock);
1590 BUG_ON(!IS_ALIGNED(va->va_start, align));
1591 BUG_ON(va->va_start < vstart);
1592 BUG_ON(va->va_end > vend);
1594 ret = kasan_populate_vmalloc(addr, size);
1597 return ERR_PTR(ret);
1604 purge_vmap_area_lazy();
1610 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1617 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1618 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1621 kmem_cache_free(vmap_area_cachep, va);
1622 return ERR_PTR(-EBUSY);
1625 int register_vmap_purge_notifier(struct notifier_block *nb)
1627 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1629 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1631 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1633 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1635 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1638 * lazy_max_pages is the maximum amount of virtual address space we gather up
1639 * before attempting to purge with a TLB flush.
1641 * There is a tradeoff here: a larger number will cover more kernel page tables
1642 * and take slightly longer to purge, but it will linearly reduce the number of
1643 * global TLB flushes that must be performed. It would seem natural to scale
1644 * this number up linearly with the number of CPUs (because vmapping activity
1645 * could also scale linearly with the number of CPUs), however it is likely
1646 * that in practice, workloads might be constrained in other ways that mean
1647 * vmap activity will not scale linearly with CPUs. Also, I want to be
1648 * conservative and not introduce a big latency on huge systems, so go with
1649 * a less aggressive log scale. It will still be an improvement over the old
1650 * code, and it will be simple to change the scale factor if we find that it
1651 * becomes a problem on bigger systems.
1653 static unsigned long lazy_max_pages(void)
1657 log = fls(num_online_cpus());
1659 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1662 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1665 * Serialize vmap purging. There is no actual critical section protected
1666 * by this look, but we want to avoid concurrent calls for performance
1667 * reasons and to make the pcpu_get_vm_areas more deterministic.
1669 static DEFINE_MUTEX(vmap_purge_lock);
1671 /* for per-CPU blocks */
1672 static void purge_fragmented_blocks_allcpus(void);
1675 * Purges all lazily-freed vmap areas.
1677 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1679 unsigned long resched_threshold;
1680 struct list_head local_pure_list;
1681 struct vmap_area *va, *n_va;
1683 lockdep_assert_held(&vmap_purge_lock);
1685 spin_lock(&purge_vmap_area_lock);
1686 purge_vmap_area_root = RB_ROOT;
1687 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1688 spin_unlock(&purge_vmap_area_lock);
1690 if (unlikely(list_empty(&local_pure_list)))
1694 list_first_entry(&local_pure_list,
1695 struct vmap_area, list)->va_start);
1698 list_last_entry(&local_pure_list,
1699 struct vmap_area, list)->va_end);
1701 flush_tlb_kernel_range(start, end);
1702 resched_threshold = lazy_max_pages() << 1;
1704 spin_lock(&free_vmap_area_lock);
1705 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1706 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1707 unsigned long orig_start = va->va_start;
1708 unsigned long orig_end = va->va_end;
1711 * Finally insert or merge lazily-freed area. It is
1712 * detached and there is no need to "unlink" it from
1715 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1716 &free_vmap_area_list);
1721 if (is_vmalloc_or_module_addr((void *)orig_start))
1722 kasan_release_vmalloc(orig_start, orig_end,
1723 va->va_start, va->va_end);
1725 atomic_long_sub(nr, &vmap_lazy_nr);
1727 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1728 cond_resched_lock(&free_vmap_area_lock);
1730 spin_unlock(&free_vmap_area_lock);
1735 * Kick off a purge of the outstanding lazy areas.
1737 static void purge_vmap_area_lazy(void)
1739 mutex_lock(&vmap_purge_lock);
1740 purge_fragmented_blocks_allcpus();
1741 __purge_vmap_area_lazy(ULONG_MAX, 0);
1742 mutex_unlock(&vmap_purge_lock);
1745 static void drain_vmap_area_work(struct work_struct *work)
1747 unsigned long nr_lazy;
1750 mutex_lock(&vmap_purge_lock);
1751 __purge_vmap_area_lazy(ULONG_MAX, 0);
1752 mutex_unlock(&vmap_purge_lock);
1754 /* Recheck if further work is required. */
1755 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1756 } while (nr_lazy > lazy_max_pages());
1760 * Free a vmap area, caller ensuring that the area has been unmapped
1761 * and flush_cache_vunmap had been called for the correct range
1764 static void free_vmap_area_noflush(struct vmap_area *va)
1766 unsigned long nr_lazy;
1768 spin_lock(&vmap_area_lock);
1769 unlink_va(va, &vmap_area_root);
1770 spin_unlock(&vmap_area_lock);
1772 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1773 PAGE_SHIFT, &vmap_lazy_nr);
1776 * Merge or place it to the purge tree/list.
1778 spin_lock(&purge_vmap_area_lock);
1779 merge_or_add_vmap_area(va,
1780 &purge_vmap_area_root, &purge_vmap_area_list);
1781 spin_unlock(&purge_vmap_area_lock);
1783 /* After this point, we may free va at any time */
1784 if (unlikely(nr_lazy > lazy_max_pages()))
1785 schedule_work(&drain_vmap_work);
1789 * Free and unmap a vmap area
1791 static void free_unmap_vmap_area(struct vmap_area *va)
1793 flush_cache_vunmap(va->va_start, va->va_end);
1794 vunmap_range_noflush(va->va_start, va->va_end);
1795 if (debug_pagealloc_enabled_static())
1796 flush_tlb_kernel_range(va->va_start, va->va_end);
1798 free_vmap_area_noflush(va);
1801 struct vmap_area *find_vmap_area(unsigned long addr)
1803 struct vmap_area *va;
1805 spin_lock(&vmap_area_lock);
1806 va = __find_vmap_area(addr);
1807 spin_unlock(&vmap_area_lock);
1812 /*** Per cpu kva allocator ***/
1815 * vmap space is limited especially on 32 bit architectures. Ensure there is
1816 * room for at least 16 percpu vmap blocks per CPU.
1819 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1820 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1821 * instead (we just need a rough idea)
1823 #if BITS_PER_LONG == 32
1824 #define VMALLOC_SPACE (128UL*1024*1024)
1826 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1829 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1830 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1831 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1832 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1833 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1834 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1835 #define VMAP_BBMAP_BITS \
1836 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1837 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1838 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1840 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1842 struct vmap_block_queue {
1844 struct list_head free;
1849 struct vmap_area *va;
1850 unsigned long free, dirty;
1851 unsigned long dirty_min, dirty_max; /*< dirty range */
1852 struct list_head free_list;
1853 struct rcu_head rcu_head;
1854 struct list_head purge;
1857 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1858 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1861 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1862 * in the free path. Could get rid of this if we change the API to return a
1863 * "cookie" from alloc, to be passed to free. But no big deal yet.
1865 static DEFINE_XARRAY(vmap_blocks);
1868 * We should probably have a fallback mechanism to allocate virtual memory
1869 * out of partially filled vmap blocks. However vmap block sizing should be
1870 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1874 static unsigned long addr_to_vb_idx(unsigned long addr)
1876 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1877 addr /= VMAP_BLOCK_SIZE;
1881 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1885 addr = va_start + (pages_off << PAGE_SHIFT);
1886 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1887 return (void *)addr;
1891 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1892 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1893 * @order: how many 2^order pages should be occupied in newly allocated block
1894 * @gfp_mask: flags for the page level allocator
1896 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1898 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1900 struct vmap_block_queue *vbq;
1901 struct vmap_block *vb;
1902 struct vmap_area *va;
1903 unsigned long vb_idx;
1907 node = numa_node_id();
1909 vb = kmalloc_node(sizeof(struct vmap_block),
1910 gfp_mask & GFP_RECLAIM_MASK, node);
1912 return ERR_PTR(-ENOMEM);
1914 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1915 VMALLOC_START, VMALLOC_END,
1919 return ERR_CAST(va);
1922 vaddr = vmap_block_vaddr(va->va_start, 0);
1923 spin_lock_init(&vb->lock);
1925 /* At least something should be left free */
1926 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1927 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1929 vb->dirty_min = VMAP_BBMAP_BITS;
1931 INIT_LIST_HEAD(&vb->free_list);
1933 vb_idx = addr_to_vb_idx(va->va_start);
1934 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1938 return ERR_PTR(err);
1941 vbq = raw_cpu_ptr(&vmap_block_queue);
1942 spin_lock(&vbq->lock);
1943 list_add_tail_rcu(&vb->free_list, &vbq->free);
1944 spin_unlock(&vbq->lock);
1949 static void free_vmap_block(struct vmap_block *vb)
1951 struct vmap_block *tmp;
1953 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1956 free_vmap_area_noflush(vb->va);
1957 kfree_rcu(vb, rcu_head);
1960 static void purge_fragmented_blocks(int cpu)
1963 struct vmap_block *vb;
1964 struct vmap_block *n_vb;
1965 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1968 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1970 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1973 spin_lock(&vb->lock);
1974 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1975 vb->free = 0; /* prevent further allocs after releasing lock */
1976 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1978 vb->dirty_max = VMAP_BBMAP_BITS;
1979 spin_lock(&vbq->lock);
1980 list_del_rcu(&vb->free_list);
1981 spin_unlock(&vbq->lock);
1982 spin_unlock(&vb->lock);
1983 list_add_tail(&vb->purge, &purge);
1985 spin_unlock(&vb->lock);
1989 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1990 list_del(&vb->purge);
1991 free_vmap_block(vb);
1995 static void purge_fragmented_blocks_allcpus(void)
1999 for_each_possible_cpu(cpu)
2000 purge_fragmented_blocks(cpu);
2003 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2005 struct vmap_block_queue *vbq;
2006 struct vmap_block *vb;
2010 BUG_ON(offset_in_page(size));
2011 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2012 if (WARN_ON(size == 0)) {
2014 * Allocating 0 bytes isn't what caller wants since
2015 * get_order(0) returns funny result. Just warn and terminate
2020 order = get_order(size);
2023 vbq = raw_cpu_ptr(&vmap_block_queue);
2024 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2025 unsigned long pages_off;
2027 spin_lock(&vb->lock);
2028 if (vb->free < (1UL << order)) {
2029 spin_unlock(&vb->lock);
2033 pages_off = VMAP_BBMAP_BITS - vb->free;
2034 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2035 vb->free -= 1UL << order;
2036 if (vb->free == 0) {
2037 spin_lock(&vbq->lock);
2038 list_del_rcu(&vb->free_list);
2039 spin_unlock(&vbq->lock);
2042 spin_unlock(&vb->lock);
2048 /* Allocate new block if nothing was found */
2050 vaddr = new_vmap_block(order, gfp_mask);
2055 static void vb_free(unsigned long addr, unsigned long size)
2057 unsigned long offset;
2059 struct vmap_block *vb;
2061 BUG_ON(offset_in_page(size));
2062 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2064 flush_cache_vunmap(addr, addr + size);
2066 order = get_order(size);
2067 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2068 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2070 vunmap_range_noflush(addr, addr + size);
2072 if (debug_pagealloc_enabled_static())
2073 flush_tlb_kernel_range(addr, addr + size);
2075 spin_lock(&vb->lock);
2077 /* Expand dirty range */
2078 vb->dirty_min = min(vb->dirty_min, offset);
2079 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2081 vb->dirty += 1UL << order;
2082 if (vb->dirty == VMAP_BBMAP_BITS) {
2084 spin_unlock(&vb->lock);
2085 free_vmap_block(vb);
2087 spin_unlock(&vb->lock);
2090 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2094 if (unlikely(!vmap_initialized))
2099 for_each_possible_cpu(cpu) {
2100 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2101 struct vmap_block *vb;
2104 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2105 spin_lock(&vb->lock);
2106 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2107 unsigned long va_start = vb->va->va_start;
2110 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2111 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2113 start = min(s, start);
2118 spin_unlock(&vb->lock);
2123 mutex_lock(&vmap_purge_lock);
2124 purge_fragmented_blocks_allcpus();
2125 if (!__purge_vmap_area_lazy(start, end) && flush)
2126 flush_tlb_kernel_range(start, end);
2127 mutex_unlock(&vmap_purge_lock);
2131 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2133 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2134 * to amortize TLB flushing overheads. What this means is that any page you
2135 * have now, may, in a former life, have been mapped into kernel virtual
2136 * address by the vmap layer and so there might be some CPUs with TLB entries
2137 * still referencing that page (additional to the regular 1:1 kernel mapping).
2139 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2140 * be sure that none of the pages we have control over will have any aliases
2141 * from the vmap layer.
2143 void vm_unmap_aliases(void)
2145 unsigned long start = ULONG_MAX, end = 0;
2148 _vm_unmap_aliases(start, end, flush);
2150 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2153 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2154 * @mem: the pointer returned by vm_map_ram
2155 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2157 void vm_unmap_ram(const void *mem, unsigned int count)
2159 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2160 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2161 struct vmap_area *va;
2165 BUG_ON(addr < VMALLOC_START);
2166 BUG_ON(addr > VMALLOC_END);
2167 BUG_ON(!PAGE_ALIGNED(addr));
2169 kasan_poison_vmalloc(mem, size);
2171 if (likely(count <= VMAP_MAX_ALLOC)) {
2172 debug_check_no_locks_freed(mem, size);
2173 vb_free(addr, size);
2177 va = find_vmap_area(addr);
2179 debug_check_no_locks_freed((void *)va->va_start,
2180 (va->va_end - va->va_start));
2181 free_unmap_vmap_area(va);
2183 EXPORT_SYMBOL(vm_unmap_ram);
2186 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2187 * @pages: an array of pointers to the pages to be mapped
2188 * @count: number of pages
2189 * @node: prefer to allocate data structures on this node
2191 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2192 * faster than vmap so it's good. But if you mix long-life and short-life
2193 * objects with vm_map_ram(), it could consume lots of address space through
2194 * fragmentation (especially on a 32bit machine). You could see failures in
2195 * the end. Please use this function for short-lived objects.
2197 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2199 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2201 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2205 if (likely(count <= VMAP_MAX_ALLOC)) {
2206 mem = vb_alloc(size, GFP_KERNEL);
2209 addr = (unsigned long)mem;
2211 struct vmap_area *va;
2212 va = alloc_vmap_area(size, PAGE_SIZE,
2213 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2217 addr = va->va_start;
2221 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2222 pages, PAGE_SHIFT) < 0) {
2223 vm_unmap_ram(mem, count);
2228 * Mark the pages as accessible, now that they are mapped.
2229 * With hardware tag-based KASAN, marking is skipped for
2230 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2232 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2236 EXPORT_SYMBOL(vm_map_ram);
2238 static struct vm_struct *vmlist __initdata;
2240 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2242 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2243 return vm->page_order;
2249 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2251 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2252 vm->page_order = order;
2259 * vm_area_add_early - add vmap area early during boot
2260 * @vm: vm_struct to add
2262 * This function is used to add fixed kernel vm area to vmlist before
2263 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2264 * should contain proper values and the other fields should be zero.
2266 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2268 void __init vm_area_add_early(struct vm_struct *vm)
2270 struct vm_struct *tmp, **p;
2272 BUG_ON(vmap_initialized);
2273 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2274 if (tmp->addr >= vm->addr) {
2275 BUG_ON(tmp->addr < vm->addr + vm->size);
2278 BUG_ON(tmp->addr + tmp->size > vm->addr);
2285 * vm_area_register_early - register vmap area early during boot
2286 * @vm: vm_struct to register
2287 * @align: requested alignment
2289 * This function is used to register kernel vm area before
2290 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2291 * proper values on entry and other fields should be zero. On return,
2292 * vm->addr contains the allocated address.
2294 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2296 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2298 unsigned long addr = ALIGN(VMALLOC_START, align);
2299 struct vm_struct *cur, **p;
2301 BUG_ON(vmap_initialized);
2303 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2304 if ((unsigned long)cur->addr - addr >= vm->size)
2306 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2309 BUG_ON(addr > VMALLOC_END - vm->size);
2310 vm->addr = (void *)addr;
2313 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2316 static void vmap_init_free_space(void)
2318 unsigned long vmap_start = 1;
2319 const unsigned long vmap_end = ULONG_MAX;
2320 struct vmap_area *busy, *free;
2324 * -|-----|.....|-----|-----|-----|.....|-
2326 * |<--------------------------------->|
2328 list_for_each_entry(busy, &vmap_area_list, list) {
2329 if (busy->va_start - vmap_start > 0) {
2330 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2331 if (!WARN_ON_ONCE(!free)) {
2332 free->va_start = vmap_start;
2333 free->va_end = busy->va_start;
2335 insert_vmap_area_augment(free, NULL,
2336 &free_vmap_area_root,
2337 &free_vmap_area_list);
2341 vmap_start = busy->va_end;
2344 if (vmap_end - vmap_start > 0) {
2345 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2346 if (!WARN_ON_ONCE(!free)) {
2347 free->va_start = vmap_start;
2348 free->va_end = vmap_end;
2350 insert_vmap_area_augment(free, NULL,
2351 &free_vmap_area_root,
2352 &free_vmap_area_list);
2357 void __init vmalloc_init(void)
2359 struct vmap_area *va;
2360 struct vm_struct *tmp;
2364 * Create the cache for vmap_area objects.
2366 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2368 for_each_possible_cpu(i) {
2369 struct vmap_block_queue *vbq;
2370 struct vfree_deferred *p;
2372 vbq = &per_cpu(vmap_block_queue, i);
2373 spin_lock_init(&vbq->lock);
2374 INIT_LIST_HEAD(&vbq->free);
2375 p = &per_cpu(vfree_deferred, i);
2376 init_llist_head(&p->list);
2377 INIT_WORK(&p->wq, free_work);
2380 /* Import existing vmlist entries. */
2381 for (tmp = vmlist; tmp; tmp = tmp->next) {
2382 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2383 if (WARN_ON_ONCE(!va))
2386 va->va_start = (unsigned long)tmp->addr;
2387 va->va_end = va->va_start + tmp->size;
2389 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2393 * Now we can initialize a free vmap space.
2395 vmap_init_free_space();
2396 vmap_initialized = true;
2399 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2400 struct vmap_area *va, unsigned long flags, const void *caller)
2403 vm->addr = (void *)va->va_start;
2404 vm->size = va->va_end - va->va_start;
2405 vm->caller = caller;
2409 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2410 unsigned long flags, const void *caller)
2412 spin_lock(&vmap_area_lock);
2413 setup_vmalloc_vm_locked(vm, va, flags, caller);
2414 spin_unlock(&vmap_area_lock);
2417 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2420 * Before removing VM_UNINITIALIZED,
2421 * we should make sure that vm has proper values.
2422 * Pair with smp_rmb() in show_numa_info().
2425 vm->flags &= ~VM_UNINITIALIZED;
2428 static struct vm_struct *__get_vm_area_node(unsigned long size,
2429 unsigned long align, unsigned long shift, unsigned long flags,
2430 unsigned long start, unsigned long end, int node,
2431 gfp_t gfp_mask, const void *caller)
2433 struct vmap_area *va;
2434 struct vm_struct *area;
2435 unsigned long requested_size = size;
2437 BUG_ON(in_interrupt());
2438 size = ALIGN(size, 1ul << shift);
2439 if (unlikely(!size))
2442 if (flags & VM_IOREMAP)
2443 align = 1ul << clamp_t(int, get_count_order_long(size),
2444 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2446 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2447 if (unlikely(!area))
2450 if (!(flags & VM_NO_GUARD))
2453 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2459 setup_vmalloc_vm(area, va, flags, caller);
2462 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2463 * best-effort approach, as they can be mapped outside of vmalloc code.
2464 * For VM_ALLOC mappings, the pages are marked as accessible after
2465 * getting mapped in __vmalloc_node_range().
2466 * With hardware tag-based KASAN, marking is skipped for
2467 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2469 if (!(flags & VM_ALLOC))
2470 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2471 KASAN_VMALLOC_PROT_NORMAL);
2476 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2477 unsigned long start, unsigned long end,
2480 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2481 NUMA_NO_NODE, GFP_KERNEL, caller);
2485 * get_vm_area - reserve a contiguous kernel virtual area
2486 * @size: size of the area
2487 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2489 * Search an area of @size in the kernel virtual mapping area,
2490 * and reserved it for out purposes. Returns the area descriptor
2491 * on success or %NULL on failure.
2493 * Return: the area descriptor on success or %NULL on failure.
2495 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2497 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2498 VMALLOC_START, VMALLOC_END,
2499 NUMA_NO_NODE, GFP_KERNEL,
2500 __builtin_return_address(0));
2503 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2506 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2507 VMALLOC_START, VMALLOC_END,
2508 NUMA_NO_NODE, GFP_KERNEL, caller);
2512 * find_vm_area - find a continuous kernel virtual area
2513 * @addr: base address
2515 * Search for the kernel VM area starting at @addr, and return it.
2516 * It is up to the caller to do all required locking to keep the returned
2519 * Return: the area descriptor on success or %NULL on failure.
2521 struct vm_struct *find_vm_area(const void *addr)
2523 struct vmap_area *va;
2525 va = find_vmap_area((unsigned long)addr);
2533 * remove_vm_area - find and remove a continuous kernel virtual area
2534 * @addr: base address
2536 * Search for the kernel VM area starting at @addr, and remove it.
2537 * This function returns the found VM area, but using it is NOT safe
2538 * on SMP machines, except for its size or flags.
2540 * Return: the area descriptor on success or %NULL on failure.
2542 struct vm_struct *remove_vm_area(const void *addr)
2544 struct vmap_area *va;
2548 spin_lock(&vmap_area_lock);
2549 va = __find_vmap_area((unsigned long)addr);
2551 struct vm_struct *vm = va->vm;
2554 spin_unlock(&vmap_area_lock);
2556 kasan_free_module_shadow(vm);
2557 free_unmap_vmap_area(va);
2562 spin_unlock(&vmap_area_lock);
2566 static inline void set_area_direct_map(const struct vm_struct *area,
2567 int (*set_direct_map)(struct page *page))
2571 /* HUGE_VMALLOC passes small pages to set_direct_map */
2572 for (i = 0; i < area->nr_pages; i++)
2573 if (page_address(area->pages[i]))
2574 set_direct_map(area->pages[i]);
2577 /* Handle removing and resetting vm mappings related to the vm_struct. */
2578 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2580 unsigned long start = ULONG_MAX, end = 0;
2581 unsigned int page_order = vm_area_page_order(area);
2582 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2586 remove_vm_area(area->addr);
2588 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2593 * If not deallocating pages, just do the flush of the VM area and
2596 if (!deallocate_pages) {
2602 * If execution gets here, flush the vm mapping and reset the direct
2603 * map. Find the start and end range of the direct mappings to make sure
2604 * the vm_unmap_aliases() flush includes the direct map.
2606 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2607 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2609 unsigned long page_size;
2611 page_size = PAGE_SIZE << page_order;
2612 start = min(addr, start);
2613 end = max(addr + page_size, end);
2619 * Set direct map to something invalid so that it won't be cached if
2620 * there are any accesses after the TLB flush, then flush the TLB and
2621 * reset the direct map permissions to the default.
2623 set_area_direct_map(area, set_direct_map_invalid_noflush);
2624 _vm_unmap_aliases(start, end, flush_dmap);
2625 set_area_direct_map(area, set_direct_map_default_noflush);
2628 static void __vunmap(const void *addr, int deallocate_pages)
2630 struct vm_struct *area;
2635 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2639 area = find_vm_area(addr);
2640 if (unlikely(!area)) {
2641 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2646 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2647 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2649 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2651 vm_remove_mappings(area, deallocate_pages);
2653 if (deallocate_pages) {
2656 for (i = 0; i < area->nr_pages; i++) {
2657 struct page *page = area->pages[i];
2660 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2662 * High-order allocs for huge vmallocs are split, so
2663 * can be freed as an array of order-0 allocations
2665 __free_pages(page, 0);
2668 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2670 kvfree(area->pages);
2676 static inline void __vfree_deferred(const void *addr)
2679 * Use raw_cpu_ptr() because this can be called from preemptible
2680 * context. Preemption is absolutely fine here, because the llist_add()
2681 * implementation is lockless, so it works even if we are adding to
2682 * another cpu's list. schedule_work() should be fine with this too.
2684 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2686 if (llist_add((struct llist_node *)addr, &p->list))
2687 schedule_work(&p->wq);
2691 * vfree_atomic - release memory allocated by vmalloc()
2692 * @addr: memory base address
2694 * This one is just like vfree() but can be called in any atomic context
2697 void vfree_atomic(const void *addr)
2701 kmemleak_free(addr);
2705 __vfree_deferred(addr);
2708 static void __vfree(const void *addr)
2710 if (unlikely(in_interrupt()))
2711 __vfree_deferred(addr);
2717 * vfree - Release memory allocated by vmalloc()
2718 * @addr: Memory base address
2720 * Free the virtually continuous memory area starting at @addr, as obtained
2721 * from one of the vmalloc() family of APIs. This will usually also free the
2722 * physical memory underlying the virtual allocation, but that memory is
2723 * reference counted, so it will not be freed until the last user goes away.
2725 * If @addr is NULL, no operation is performed.
2728 * May sleep if called *not* from interrupt context.
2729 * Must not be called in NMI context (strictly speaking, it could be
2730 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2731 * conventions for vfree() arch-dependent would be a really bad idea).
2733 void vfree(const void *addr)
2737 kmemleak_free(addr);
2739 might_sleep_if(!in_interrupt());
2746 EXPORT_SYMBOL(vfree);
2749 * vunmap - release virtual mapping obtained by vmap()
2750 * @addr: memory base address
2752 * Free the virtually contiguous memory area starting at @addr,
2753 * which was created from the page array passed to vmap().
2755 * Must not be called in interrupt context.
2757 void vunmap(const void *addr)
2759 BUG_ON(in_interrupt());
2764 EXPORT_SYMBOL(vunmap);
2767 * vmap - map an array of pages into virtually contiguous space
2768 * @pages: array of page pointers
2769 * @count: number of pages to map
2770 * @flags: vm_area->flags
2771 * @prot: page protection for the mapping
2773 * Maps @count pages from @pages into contiguous kernel virtual space.
2774 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2775 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2776 * are transferred from the caller to vmap(), and will be freed / dropped when
2777 * vfree() is called on the return value.
2779 * Return: the address of the area or %NULL on failure
2781 void *vmap(struct page **pages, unsigned int count,
2782 unsigned long flags, pgprot_t prot)
2784 struct vm_struct *area;
2786 unsigned long size; /* In bytes */
2791 * Your top guard is someone else's bottom guard. Not having a top
2792 * guard compromises someone else's mappings too.
2794 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2795 flags &= ~VM_NO_GUARD;
2797 if (count > totalram_pages())
2800 size = (unsigned long)count << PAGE_SHIFT;
2801 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2805 addr = (unsigned long)area->addr;
2806 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2807 pages, PAGE_SHIFT) < 0) {
2812 if (flags & VM_MAP_PUT_PAGES) {
2813 area->pages = pages;
2814 area->nr_pages = count;
2818 EXPORT_SYMBOL(vmap);
2820 #ifdef CONFIG_VMAP_PFN
2821 struct vmap_pfn_data {
2822 unsigned long *pfns;
2827 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2829 struct vmap_pfn_data *data = private;
2831 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2833 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2838 * vmap_pfn - map an array of PFNs into virtually contiguous space
2839 * @pfns: array of PFNs
2840 * @count: number of pages to map
2841 * @prot: page protection for the mapping
2843 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2844 * the start address of the mapping.
2846 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2848 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2849 struct vm_struct *area;
2851 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2852 __builtin_return_address(0));
2855 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2856 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2862 EXPORT_SYMBOL_GPL(vmap_pfn);
2863 #endif /* CONFIG_VMAP_PFN */
2865 static inline unsigned int
2866 vm_area_alloc_pages(gfp_t gfp, int nid,
2867 unsigned int order, unsigned int nr_pages, struct page **pages)
2869 unsigned int nr_allocated = 0;
2874 * For order-0 pages we make use of bulk allocator, if
2875 * the page array is partly or not at all populated due
2876 * to fails, fallback to a single page allocator that is
2880 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2882 while (nr_allocated < nr_pages) {
2883 unsigned int nr, nr_pages_request;
2886 * A maximum allowed request is hard-coded and is 100
2887 * pages per call. That is done in order to prevent a
2888 * long preemption off scenario in the bulk-allocator
2889 * so the range is [1:100].
2891 nr_pages_request = min(100U, nr_pages - nr_allocated);
2893 /* memory allocation should consider mempolicy, we can't
2894 * wrongly use nearest node when nid == NUMA_NO_NODE,
2895 * otherwise memory may be allocated in only one node,
2896 * but mempolicy wants to alloc memory by interleaving.
2898 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2899 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2901 pages + nr_allocated);
2904 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2906 pages + nr_allocated);
2912 * If zero or pages were obtained partly,
2913 * fallback to a single page allocator.
2915 if (nr != nr_pages_request)
2920 /* High-order pages or fallback path if "bulk" fails. */
2922 while (nr_allocated < nr_pages) {
2923 if (fatal_signal_pending(current))
2926 if (nid == NUMA_NO_NODE)
2927 page = alloc_pages(gfp, order);
2929 page = alloc_pages_node(nid, gfp, order);
2930 if (unlikely(!page))
2933 * Higher order allocations must be able to be treated as
2934 * indepdenent small pages by callers (as they can with
2935 * small-page vmallocs). Some drivers do their own refcounting
2936 * on vmalloc_to_page() pages, some use page->mapping,
2940 split_page(page, order);
2943 * Careful, we allocate and map page-order pages, but
2944 * tracking is done per PAGE_SIZE page so as to keep the
2945 * vm_struct APIs independent of the physical/mapped size.
2947 for (i = 0; i < (1U << order); i++)
2948 pages[nr_allocated + i] = page + i;
2951 nr_allocated += 1U << order;
2954 return nr_allocated;
2957 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2958 pgprot_t prot, unsigned int page_shift,
2961 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2962 bool nofail = gfp_mask & __GFP_NOFAIL;
2963 unsigned long addr = (unsigned long)area->addr;
2964 unsigned long size = get_vm_area_size(area);
2965 unsigned long array_size;
2966 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2967 unsigned int page_order;
2971 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2972 gfp_mask |= __GFP_NOWARN;
2973 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2974 gfp_mask |= __GFP_HIGHMEM;
2976 /* Please note that the recursion is strictly bounded. */
2977 if (array_size > PAGE_SIZE) {
2978 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2981 area->pages = kmalloc_node(array_size, nested_gfp, node);
2985 warn_alloc(gfp_mask, NULL,
2986 "vmalloc error: size %lu, failed to allocated page array size %lu",
2987 nr_small_pages * PAGE_SIZE, array_size);
2992 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2993 page_order = vm_area_page_order(area);
2995 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
2996 node, page_order, nr_small_pages, area->pages);
2998 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2999 if (gfp_mask & __GFP_ACCOUNT) {
3002 for (i = 0; i < area->nr_pages; i++)
3003 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3007 * If not enough pages were obtained to accomplish an
3008 * allocation request, free them via __vfree() if any.
3010 if (area->nr_pages != nr_small_pages) {
3011 warn_alloc(gfp_mask, NULL,
3012 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3013 area->nr_pages * PAGE_SIZE, page_order);
3018 * page tables allocations ignore external gfp mask, enforce it
3021 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3022 flags = memalloc_nofs_save();
3023 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3024 flags = memalloc_noio_save();
3027 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3029 if (nofail && (ret < 0))
3030 schedule_timeout_uninterruptible(1);
3031 } while (nofail && (ret < 0));
3033 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3034 memalloc_nofs_restore(flags);
3035 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3036 memalloc_noio_restore(flags);
3039 warn_alloc(gfp_mask, NULL,
3040 "vmalloc error: size %lu, failed to map pages",
3041 area->nr_pages * PAGE_SIZE);
3048 __vfree(area->addr);
3053 * __vmalloc_node_range - allocate virtually contiguous memory
3054 * @size: allocation size
3055 * @align: desired alignment
3056 * @start: vm area range start
3057 * @end: vm area range end
3058 * @gfp_mask: flags for the page level allocator
3059 * @prot: protection mask for the allocated pages
3060 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3061 * @node: node to use for allocation or NUMA_NO_NODE
3062 * @caller: caller's return address
3064 * Allocate enough pages to cover @size from the page level
3065 * allocator with @gfp_mask flags. Please note that the full set of gfp
3066 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3068 * Zone modifiers are not supported. From the reclaim modifiers
3069 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3070 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3071 * __GFP_RETRY_MAYFAIL are not supported).
3073 * __GFP_NOWARN can be used to suppress failures messages.
3075 * Map them into contiguous kernel virtual space, using a pagetable
3076 * protection of @prot.
3078 * Return: the address of the area or %NULL on failure
3080 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3081 unsigned long start, unsigned long end, gfp_t gfp_mask,
3082 pgprot_t prot, unsigned long vm_flags, int node,
3085 struct vm_struct *area;
3087 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3088 unsigned long real_size = size;
3089 unsigned long real_align = align;
3090 unsigned int shift = PAGE_SHIFT;
3092 if (WARN_ON_ONCE(!size))
3095 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3096 warn_alloc(gfp_mask, NULL,
3097 "vmalloc error: size %lu, exceeds total pages",
3102 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3103 unsigned long size_per_node;
3106 * Try huge pages. Only try for PAGE_KERNEL allocations,
3107 * others like modules don't yet expect huge pages in
3108 * their allocations due to apply_to_page_range not
3112 size_per_node = size;
3113 if (node == NUMA_NO_NODE)
3114 size_per_node /= num_online_nodes();
3115 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3118 shift = arch_vmap_pte_supported_shift(size_per_node);
3120 align = max(real_align, 1UL << shift);
3121 size = ALIGN(real_size, 1UL << shift);
3125 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3126 VM_UNINITIALIZED | vm_flags, start, end, node,
3129 bool nofail = gfp_mask & __GFP_NOFAIL;
3130 warn_alloc(gfp_mask, NULL,
3131 "vmalloc error: size %lu, vm_struct allocation failed%s",
3132 real_size, (nofail) ? ". Retrying." : "");
3134 schedule_timeout_uninterruptible(1);
3141 * Prepare arguments for __vmalloc_area_node() and
3142 * kasan_unpoison_vmalloc().
3144 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3145 if (kasan_hw_tags_enabled()) {
3147 * Modify protection bits to allow tagging.
3148 * This must be done before mapping.
3150 prot = arch_vmap_pgprot_tagged(prot);
3153 * Skip page_alloc poisoning and zeroing for physical
3154 * pages backing VM_ALLOC mapping. Memory is instead
3155 * poisoned and zeroed by kasan_unpoison_vmalloc().
3157 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3160 /* Take note that the mapping is PAGE_KERNEL. */
3161 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3164 /* Allocate physical pages and map them into vmalloc space. */
3165 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3170 * Mark the pages as accessible, now that they are mapped.
3171 * The init condition should match the one in post_alloc_hook()
3172 * (except for the should_skip_init() check) to make sure that memory
3173 * is initialized under the same conditions regardless of the enabled
3175 * Tag-based KASAN modes only assign tags to normal non-executable
3176 * allocations, see __kasan_unpoison_vmalloc().
3178 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3179 if (!want_init_on_free() && want_init_on_alloc(gfp_mask))
3180 kasan_flags |= KASAN_VMALLOC_INIT;
3181 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3182 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3185 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3186 * flag. It means that vm_struct is not fully initialized.
3187 * Now, it is fully initialized, so remove this flag here.
3189 clear_vm_uninitialized_flag(area);
3191 size = PAGE_ALIGN(size);
3192 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3193 kmemleak_vmalloc(area, size, gfp_mask);
3198 if (shift > PAGE_SHIFT) {
3209 * __vmalloc_node - allocate virtually contiguous memory
3210 * @size: allocation size
3211 * @align: desired alignment
3212 * @gfp_mask: flags for the page level allocator
3213 * @node: node to use for allocation or NUMA_NO_NODE
3214 * @caller: caller's return address
3216 * Allocate enough pages to cover @size from the page level allocator with
3217 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3219 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3220 * and __GFP_NOFAIL are not supported
3222 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3225 * Return: pointer to the allocated memory or %NULL on error
3227 void *__vmalloc_node(unsigned long size, unsigned long align,
3228 gfp_t gfp_mask, int node, const void *caller)
3230 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3231 gfp_mask, PAGE_KERNEL, 0, node, caller);
3234 * This is only for performance analysis of vmalloc and stress purpose.
3235 * It is required by vmalloc test module, therefore do not use it other
3238 #ifdef CONFIG_TEST_VMALLOC_MODULE
3239 EXPORT_SYMBOL_GPL(__vmalloc_node);
3242 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3244 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3245 __builtin_return_address(0));
3247 EXPORT_SYMBOL(__vmalloc);
3250 * vmalloc - allocate virtually contiguous memory
3251 * @size: allocation size
3253 * Allocate enough pages to cover @size from the page level
3254 * allocator and map them into contiguous kernel virtual space.
3256 * For tight control over page level allocator and protection flags
3257 * use __vmalloc() instead.
3259 * Return: pointer to the allocated memory or %NULL on error
3261 void *vmalloc(unsigned long size)
3263 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3264 __builtin_return_address(0));
3266 EXPORT_SYMBOL(vmalloc);
3269 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3270 * @size: allocation size
3271 * @gfp_mask: flags for the page level allocator
3273 * Allocate enough pages to cover @size from the page level
3274 * allocator and map them into contiguous kernel virtual space.
3275 * If @size is greater than or equal to PMD_SIZE, allow using
3276 * huge pages for the memory
3278 * Return: pointer to the allocated memory or %NULL on error
3280 void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3282 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3283 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3284 NUMA_NO_NODE, __builtin_return_address(0));
3286 EXPORT_SYMBOL_GPL(vmalloc_huge);
3289 * vzalloc - allocate virtually contiguous memory with zero fill
3290 * @size: allocation size
3292 * Allocate enough pages to cover @size from the page level
3293 * allocator and map them into contiguous kernel virtual space.
3294 * The memory allocated is set to zero.
3296 * For tight control over page level allocator and protection flags
3297 * use __vmalloc() instead.
3299 * Return: pointer to the allocated memory or %NULL on error
3301 void *vzalloc(unsigned long size)
3303 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3304 __builtin_return_address(0));
3306 EXPORT_SYMBOL(vzalloc);
3309 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3310 * @size: allocation size
3312 * The resulting memory area is zeroed so it can be mapped to userspace
3313 * without leaking data.
3315 * Return: pointer to the allocated memory or %NULL on error
3317 void *vmalloc_user(unsigned long size)
3319 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3320 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3321 VM_USERMAP, NUMA_NO_NODE,
3322 __builtin_return_address(0));
3324 EXPORT_SYMBOL(vmalloc_user);
3327 * vmalloc_node - allocate memory on a specific node
3328 * @size: allocation size
3331 * Allocate enough pages to cover @size from the page level
3332 * allocator and map them into contiguous kernel virtual space.
3334 * For tight control over page level allocator and protection flags
3335 * use __vmalloc() instead.
3337 * Return: pointer to the allocated memory or %NULL on error
3339 void *vmalloc_node(unsigned long size, int node)
3341 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3342 __builtin_return_address(0));
3344 EXPORT_SYMBOL(vmalloc_node);
3347 * vzalloc_node - allocate memory on a specific node with zero fill
3348 * @size: allocation size
3351 * Allocate enough pages to cover @size from the page level
3352 * allocator and map them into contiguous kernel virtual space.
3353 * The memory allocated is set to zero.
3355 * Return: pointer to the allocated memory or %NULL on error
3357 void *vzalloc_node(unsigned long size, int node)
3359 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3360 __builtin_return_address(0));
3362 EXPORT_SYMBOL(vzalloc_node);
3364 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3365 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3366 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3367 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3370 * 64b systems should always have either DMA or DMA32 zones. For others
3371 * GFP_DMA32 should do the right thing and use the normal zone.
3373 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3377 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3378 * @size: allocation size
3380 * Allocate enough 32bit PA addressable pages to cover @size from the
3381 * page level allocator and map them into contiguous kernel virtual space.
3383 * Return: pointer to the allocated memory or %NULL on error
3385 void *vmalloc_32(unsigned long size)
3387 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3388 __builtin_return_address(0));
3390 EXPORT_SYMBOL(vmalloc_32);
3393 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3394 * @size: allocation size
3396 * The resulting memory area is 32bit addressable and zeroed so it can be
3397 * mapped to userspace without leaking data.
3399 * Return: pointer to the allocated memory or %NULL on error
3401 void *vmalloc_32_user(unsigned long size)
3403 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3404 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3405 VM_USERMAP, NUMA_NO_NODE,
3406 __builtin_return_address(0));
3408 EXPORT_SYMBOL(vmalloc_32_user);
3411 * small helper routine , copy contents to buf from addr.
3412 * If the page is not present, fill zero.
3415 static int aligned_vread(char *buf, char *addr, unsigned long count)
3421 unsigned long offset, length;
3423 offset = offset_in_page(addr);
3424 length = PAGE_SIZE - offset;
3427 p = vmalloc_to_page(addr);
3429 * To do safe access to this _mapped_ area, we need
3430 * lock. But adding lock here means that we need to add
3431 * overhead of vmalloc()/vfree() calls for this _debug_
3432 * interface, rarely used. Instead of that, we'll use
3433 * kmap() and get small overhead in this access function.
3436 /* We can expect USER0 is not used -- see vread() */
3437 void *map = kmap_atomic(p);
3438 memcpy(buf, map + offset, length);
3441 memset(buf, 0, length);
3452 * vread() - read vmalloc area in a safe way.
3453 * @buf: buffer for reading data
3454 * @addr: vm address.
3455 * @count: number of bytes to be read.
3457 * This function checks that addr is a valid vmalloc'ed area, and
3458 * copy data from that area to a given buffer. If the given memory range
3459 * of [addr...addr+count) includes some valid address, data is copied to
3460 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3461 * IOREMAP area is treated as memory hole and no copy is done.
3463 * If [addr...addr+count) doesn't includes any intersects with alive
3464 * vm_struct area, returns 0. @buf should be kernel's buffer.
3466 * Note: In usual ops, vread() is never necessary because the caller
3467 * should know vmalloc() area is valid and can use memcpy().
3468 * This is for routines which have to access vmalloc area without
3469 * any information, as /proc/kcore.
3471 * Return: number of bytes for which addr and buf should be increased
3472 * (same number as @count) or %0 if [addr...addr+count) doesn't
3473 * include any intersection with valid vmalloc area
3475 long vread(char *buf, char *addr, unsigned long count)
3477 struct vmap_area *va;
3478 struct vm_struct *vm;
3479 char *vaddr, *buf_start = buf;
3480 unsigned long buflen = count;
3483 addr = kasan_reset_tag(addr);
3485 /* Don't allow overflow */
3486 if ((unsigned long) addr + count < count)
3487 count = -(unsigned long) addr;
3489 spin_lock(&vmap_area_lock);
3490 va = find_vmap_area_exceed_addr((unsigned long)addr);
3494 /* no intersects with alive vmap_area */
3495 if ((unsigned long)addr + count <= va->va_start)
3498 list_for_each_entry_from(va, &vmap_area_list, list) {
3506 vaddr = (char *) vm->addr;
3507 if (addr >= vaddr + get_vm_area_size(vm))
3509 while (addr < vaddr) {
3517 n = vaddr + get_vm_area_size(vm) - addr;
3520 if (!(vm->flags & VM_IOREMAP))
3521 aligned_vread(buf, addr, n);
3522 else /* IOREMAP area is treated as memory hole */
3529 spin_unlock(&vmap_area_lock);
3531 if (buf == buf_start)
3533 /* zero-fill memory holes */
3534 if (buf != buf_start + buflen)
3535 memset(buf, 0, buflen - (buf - buf_start));
3541 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3542 * @vma: vma to cover
3543 * @uaddr: target user address to start at
3544 * @kaddr: virtual address of vmalloc kernel memory
3545 * @pgoff: offset from @kaddr to start at
3546 * @size: size of map area
3548 * Returns: 0 for success, -Exxx on failure
3550 * This function checks that @kaddr is a valid vmalloc'ed area,
3551 * and that it is big enough to cover the range starting at
3552 * @uaddr in @vma. Will return failure if that criteria isn't
3555 * Similar to remap_pfn_range() (see mm/memory.c)
3557 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3558 void *kaddr, unsigned long pgoff,
3561 struct vm_struct *area;
3563 unsigned long end_index;
3565 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3568 size = PAGE_ALIGN(size);
3570 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3573 area = find_vm_area(kaddr);
3577 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3580 if (check_add_overflow(size, off, &end_index) ||
3581 end_index > get_vm_area_size(area))
3586 struct page *page = vmalloc_to_page(kaddr);
3589 ret = vm_insert_page(vma, uaddr, page);
3598 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3604 * remap_vmalloc_range - map vmalloc pages to userspace
3605 * @vma: vma to cover (map full range of vma)
3606 * @addr: vmalloc memory
3607 * @pgoff: number of pages into addr before first page to map
3609 * Returns: 0 for success, -Exxx on failure
3611 * This function checks that addr is a valid vmalloc'ed area, and
3612 * that it is big enough to cover the vma. Will return failure if
3613 * that criteria isn't met.
3615 * Similar to remap_pfn_range() (see mm/memory.c)
3617 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3618 unsigned long pgoff)
3620 return remap_vmalloc_range_partial(vma, vma->vm_start,
3622 vma->vm_end - vma->vm_start);
3624 EXPORT_SYMBOL(remap_vmalloc_range);
3626 void free_vm_area(struct vm_struct *area)
3628 struct vm_struct *ret;
3629 ret = remove_vm_area(area->addr);
3630 BUG_ON(ret != area);
3633 EXPORT_SYMBOL_GPL(free_vm_area);
3636 static struct vmap_area *node_to_va(struct rb_node *n)
3638 return rb_entry_safe(n, struct vmap_area, rb_node);
3642 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3643 * @addr: target address
3645 * Returns: vmap_area if it is found. If there is no such area
3646 * the first highest(reverse order) vmap_area is returned
3647 * i.e. va->va_start < addr && va->va_end < addr or NULL
3648 * if there are no any areas before @addr.
3650 static struct vmap_area *
3651 pvm_find_va_enclose_addr(unsigned long addr)
3653 struct vmap_area *va, *tmp;
3656 n = free_vmap_area_root.rb_node;
3660 tmp = rb_entry(n, struct vmap_area, rb_node);
3661 if (tmp->va_start <= addr) {
3663 if (tmp->va_end >= addr)
3676 * pvm_determine_end_from_reverse - find the highest aligned address
3677 * of free block below VMALLOC_END
3679 * in - the VA we start the search(reverse order);
3680 * out - the VA with the highest aligned end address.
3681 * @align: alignment for required highest address
3683 * Returns: determined end address within vmap_area
3685 static unsigned long
3686 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3688 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3692 list_for_each_entry_from_reverse((*va),
3693 &free_vmap_area_list, list) {
3694 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3695 if ((*va)->va_start < addr)
3704 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3705 * @offsets: array containing offset of each area
3706 * @sizes: array containing size of each area
3707 * @nr_vms: the number of areas to allocate
3708 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3710 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3711 * vm_structs on success, %NULL on failure
3713 * Percpu allocator wants to use congruent vm areas so that it can
3714 * maintain the offsets among percpu areas. This function allocates
3715 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3716 * be scattered pretty far, distance between two areas easily going up
3717 * to gigabytes. To avoid interacting with regular vmallocs, these
3718 * areas are allocated from top.
3720 * Despite its complicated look, this allocator is rather simple. It
3721 * does everything top-down and scans free blocks from the end looking
3722 * for matching base. While scanning, if any of the areas do not fit the
3723 * base address is pulled down to fit the area. Scanning is repeated till
3724 * all the areas fit and then all necessary data structures are inserted
3725 * and the result is returned.
3727 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3728 const size_t *sizes, int nr_vms,
3731 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3732 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3733 struct vmap_area **vas, *va;
3734 struct vm_struct **vms;
3735 int area, area2, last_area, term_area;
3736 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3737 bool purged = false;
3740 /* verify parameters and allocate data structures */
3741 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3742 for (last_area = 0, area = 0; area < nr_vms; area++) {
3743 start = offsets[area];
3744 end = start + sizes[area];
3746 /* is everything aligned properly? */
3747 BUG_ON(!IS_ALIGNED(offsets[area], align));
3748 BUG_ON(!IS_ALIGNED(sizes[area], align));
3750 /* detect the area with the highest address */
3751 if (start > offsets[last_area])
3754 for (area2 = area + 1; area2 < nr_vms; area2++) {
3755 unsigned long start2 = offsets[area2];
3756 unsigned long end2 = start2 + sizes[area2];
3758 BUG_ON(start2 < end && start < end2);
3761 last_end = offsets[last_area] + sizes[last_area];
3763 if (vmalloc_end - vmalloc_start < last_end) {
3768 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3769 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3773 for (area = 0; area < nr_vms; area++) {
3774 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3775 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3776 if (!vas[area] || !vms[area])
3780 spin_lock(&free_vmap_area_lock);
3782 /* start scanning - we scan from the top, begin with the last area */
3783 area = term_area = last_area;
3784 start = offsets[area];
3785 end = start + sizes[area];
3787 va = pvm_find_va_enclose_addr(vmalloc_end);
3788 base = pvm_determine_end_from_reverse(&va, align) - end;
3792 * base might have underflowed, add last_end before
3795 if (base + last_end < vmalloc_start + last_end)
3799 * Fitting base has not been found.
3805 * If required width exceeds current VA block, move
3806 * base downwards and then recheck.
3808 if (base + end > va->va_end) {
3809 base = pvm_determine_end_from_reverse(&va, align) - end;
3815 * If this VA does not fit, move base downwards and recheck.
3817 if (base + start < va->va_start) {
3818 va = node_to_va(rb_prev(&va->rb_node));
3819 base = pvm_determine_end_from_reverse(&va, align) - end;
3825 * This area fits, move on to the previous one. If
3826 * the previous one is the terminal one, we're done.
3828 area = (area + nr_vms - 1) % nr_vms;
3829 if (area == term_area)
3832 start = offsets[area];
3833 end = start + sizes[area];
3834 va = pvm_find_va_enclose_addr(base + end);
3837 /* we've found a fitting base, insert all va's */
3838 for (area = 0; area < nr_vms; area++) {
3841 start = base + offsets[area];
3844 va = pvm_find_va_enclose_addr(start);
3845 if (WARN_ON_ONCE(va == NULL))
3846 /* It is a BUG(), but trigger recovery instead. */
3849 type = classify_va_fit_type(va, start, size);
3850 if (WARN_ON_ONCE(type == NOTHING_FIT))
3851 /* It is a BUG(), but trigger recovery instead. */
3854 ret = adjust_va_to_fit_type(va, start, size, type);
3858 /* Allocated area. */
3860 va->va_start = start;
3861 va->va_end = start + size;
3864 spin_unlock(&free_vmap_area_lock);
3866 /* populate the kasan shadow space */
3867 for (area = 0; area < nr_vms; area++) {
3868 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3869 goto err_free_shadow;
3872 /* insert all vm's */
3873 spin_lock(&vmap_area_lock);
3874 for (area = 0; area < nr_vms; area++) {
3875 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3877 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3880 spin_unlock(&vmap_area_lock);
3883 * Mark allocated areas as accessible. Do it now as a best-effort
3884 * approach, as they can be mapped outside of vmalloc code.
3885 * With hardware tag-based KASAN, marking is skipped for
3886 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3888 for (area = 0; area < nr_vms; area++)
3889 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3890 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3897 * Remove previously allocated areas. There is no
3898 * need in removing these areas from the busy tree,
3899 * because they are inserted only on the final step
3900 * and when pcpu_get_vm_areas() is success.
3903 orig_start = vas[area]->va_start;
3904 orig_end = vas[area]->va_end;
3905 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3906 &free_vmap_area_list);
3908 kasan_release_vmalloc(orig_start, orig_end,
3909 va->va_start, va->va_end);
3914 spin_unlock(&free_vmap_area_lock);
3916 purge_vmap_area_lazy();
3919 /* Before "retry", check if we recover. */
3920 for (area = 0; area < nr_vms; area++) {
3924 vas[area] = kmem_cache_zalloc(
3925 vmap_area_cachep, GFP_KERNEL);
3934 for (area = 0; area < nr_vms; area++) {
3936 kmem_cache_free(vmap_area_cachep, vas[area]);
3946 spin_lock(&free_vmap_area_lock);
3948 * We release all the vmalloc shadows, even the ones for regions that
3949 * hadn't been successfully added. This relies on kasan_release_vmalloc
3950 * being able to tolerate this case.
3952 for (area = 0; area < nr_vms; area++) {
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);
3963 spin_unlock(&free_vmap_area_lock);
3970 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3971 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3972 * @nr_vms: the number of allocated areas
3974 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3976 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3980 for (i = 0; i < nr_vms; i++)
3981 free_vm_area(vms[i]);
3984 #endif /* CONFIG_SMP */
3986 #ifdef CONFIG_PRINTK
3987 bool vmalloc_dump_obj(void *object)
3989 struct vm_struct *vm;
3990 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3992 vm = find_vm_area(objp);
3995 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3996 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4001 #ifdef CONFIG_PROC_FS
4002 static void *s_start(struct seq_file *m, loff_t *pos)
4003 __acquires(&vmap_purge_lock)
4004 __acquires(&vmap_area_lock)
4006 mutex_lock(&vmap_purge_lock);
4007 spin_lock(&vmap_area_lock);
4009 return seq_list_start(&vmap_area_list, *pos);
4012 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4014 return seq_list_next(p, &vmap_area_list, pos);
4017 static void s_stop(struct seq_file *m, void *p)
4018 __releases(&vmap_area_lock)
4019 __releases(&vmap_purge_lock)
4021 spin_unlock(&vmap_area_lock);
4022 mutex_unlock(&vmap_purge_lock);
4025 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4027 if (IS_ENABLED(CONFIG_NUMA)) {
4028 unsigned int nr, *counters = m->private;
4029 unsigned int step = 1U << vm_area_page_order(v);
4034 if (v->flags & VM_UNINITIALIZED)
4036 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4039 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4041 for (nr = 0; nr < v->nr_pages; nr += step)
4042 counters[page_to_nid(v->pages[nr])] += step;
4043 for_each_node_state(nr, N_HIGH_MEMORY)
4045 seq_printf(m, " N%u=%u", nr, counters[nr]);
4049 static void show_purge_info(struct seq_file *m)
4051 struct vmap_area *va;
4053 spin_lock(&purge_vmap_area_lock);
4054 list_for_each_entry(va, &purge_vmap_area_list, list) {
4055 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4056 (void *)va->va_start, (void *)va->va_end,
4057 va->va_end - va->va_start);
4059 spin_unlock(&purge_vmap_area_lock);
4062 static int s_show(struct seq_file *m, void *p)
4064 struct vmap_area *va;
4065 struct vm_struct *v;
4067 va = list_entry(p, struct vmap_area, list);
4070 * s_show can encounter race with remove_vm_area, !vm on behalf
4071 * of vmap area is being tear down or vm_map_ram allocation.
4074 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4075 (void *)va->va_start, (void *)va->va_end,
4076 va->va_end - va->va_start);
4083 seq_printf(m, "0x%pK-0x%pK %7ld",
4084 v->addr, v->addr + v->size, v->size);
4087 seq_printf(m, " %pS", v->caller);
4090 seq_printf(m, " pages=%d", v->nr_pages);
4093 seq_printf(m, " phys=%pa", &v->phys_addr);
4095 if (v->flags & VM_IOREMAP)
4096 seq_puts(m, " ioremap");
4098 if (v->flags & VM_ALLOC)
4099 seq_puts(m, " vmalloc");
4101 if (v->flags & VM_MAP)
4102 seq_puts(m, " vmap");
4104 if (v->flags & VM_USERMAP)
4105 seq_puts(m, " user");
4107 if (v->flags & VM_DMA_COHERENT)
4108 seq_puts(m, " dma-coherent");
4110 if (is_vmalloc_addr(v->pages))
4111 seq_puts(m, " vpages");
4113 show_numa_info(m, v);
4117 * As a final step, dump "unpurged" areas.
4120 if (list_is_last(&va->list, &vmap_area_list))
4126 static const struct seq_operations vmalloc_op = {
4133 static int __init proc_vmalloc_init(void)
4135 if (IS_ENABLED(CONFIG_NUMA))
4136 proc_create_seq_private("vmallocinfo", 0400, NULL,
4138 nr_node_ids * sizeof(unsigned int), NULL);
4140 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4143 module_init(proc_vmalloc_init);