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)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)
394 p4d = p4d_offset(pgd, addr);
396 next = p4d_addr_end(addr, end);
398 cleared = p4d_clear_huge(p4d);
399 if (cleared || p4d_bad(*p4d))
400 *mask |= PGTBL_P4D_MODIFIED;
404 if (p4d_none_or_clear_bad(p4d))
406 vunmap_pud_range(p4d, addr, next, mask);
407 } while (p4d++, addr = next, addr != end);
411 * vunmap_range_noflush is similar to vunmap_range, but does not
412 * flush caches or TLBs.
414 * The caller is responsible for calling flush_cache_vmap() before calling
415 * this function, and flush_tlb_kernel_range after it has returned
416 * successfully (and before the addresses are expected to cause a page fault
417 * or be re-mapped for something else, if TLB flushes are being delayed or
420 * This is an internal function only. Do not use outside mm/.
422 void vunmap_range_noflush(unsigned long start, unsigned long end)
426 unsigned long addr = start;
427 pgtbl_mod_mask mask = 0;
430 pgd = pgd_offset_k(addr);
432 next = pgd_addr_end(addr, end);
434 mask |= PGTBL_PGD_MODIFIED;
435 if (pgd_none_or_clear_bad(pgd))
437 vunmap_p4d_range(pgd, addr, next, &mask);
438 } while (pgd++, addr = next, addr != end);
440 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
441 arch_sync_kernel_mappings(start, end);
445 * vunmap_range - unmap kernel virtual addresses
446 * @addr: start of the VM area to unmap
447 * @end: end of the VM area to unmap (non-inclusive)
449 * Clears any present PTEs in the virtual address range, flushes TLBs and
450 * caches. Any subsequent access to the address before it has been re-mapped
453 void vunmap_range(unsigned long addr, unsigned long end)
455 flush_cache_vunmap(addr, end);
456 vunmap_range_noflush(addr, end);
457 flush_tlb_kernel_range(addr, end);
460 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
461 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
462 pgtbl_mod_mask *mask)
467 * nr is a running index into the array which helps higher level
468 * callers keep track of where we're up to.
471 pte = pte_alloc_kernel_track(pmd, addr, mask);
475 struct page *page = pages[*nr];
477 if (WARN_ON(!pte_none(*pte)))
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)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;
779 * Gets called when remove the node and rotate.
781 static __always_inline unsigned long
782 compute_subtree_max_size(struct vmap_area *va)
784 return max3(va_size(va),
785 get_subtree_max_size(va->rb_node.rb_left),
786 get_subtree_max_size(va->rb_node.rb_right));
789 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
790 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
792 static void purge_vmap_area_lazy(void);
793 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
794 static void drain_vmap_area_work(struct work_struct *work);
795 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
797 static atomic_long_t nr_vmalloc_pages;
799 unsigned long vmalloc_nr_pages(void)
801 return atomic_long_read(&nr_vmalloc_pages);
804 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
806 struct vmap_area *va = NULL;
807 struct rb_node *n = vmap_area_root.rb_node;
810 struct vmap_area *tmp;
812 tmp = rb_entry(n, struct vmap_area, rb_node);
813 if (tmp->va_end > addr) {
815 if (tmp->va_start <= addr)
826 static struct vmap_area *__find_vmap_area(unsigned long addr)
828 struct rb_node *n = vmap_area_root.rb_node;
831 struct vmap_area *va;
833 va = rb_entry(n, struct vmap_area, rb_node);
834 if (addr < va->va_start)
836 else if (addr >= va->va_end)
846 * This function returns back addresses of parent node
847 * and its left or right link for further processing.
849 * Otherwise NULL is returned. In that case all further
850 * steps regarding inserting of conflicting overlap range
851 * have to be declined and actually considered as a bug.
853 static __always_inline struct rb_node **
854 find_va_links(struct vmap_area *va,
855 struct rb_root *root, struct rb_node *from,
856 struct rb_node **parent)
858 struct vmap_area *tmp_va;
859 struct rb_node **link;
862 link = &root->rb_node;
863 if (unlikely(!*link)) {
872 * Go to the bottom of the tree. When we hit the last point
873 * we end up with parent rb_node and correct direction, i name
874 * it link, where the new va->rb_node will be attached to.
877 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
880 * During the traversal we also do some sanity check.
881 * Trigger the BUG() if there are sides(left/right)
884 if (va->va_start < tmp_va->va_end &&
885 va->va_end <= tmp_va->va_start)
886 link = &(*link)->rb_left;
887 else if (va->va_end > tmp_va->va_start &&
888 va->va_start >= tmp_va->va_end)
889 link = &(*link)->rb_right;
891 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
892 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
898 *parent = &tmp_va->rb_node;
902 static __always_inline struct list_head *
903 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
905 struct list_head *list;
907 if (unlikely(!parent))
909 * The red-black tree where we try to find VA neighbors
910 * before merging or inserting is empty, i.e. it means
911 * there is no free vmap space. Normally it does not
912 * happen but we handle this case anyway.
916 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
917 return (&parent->rb_right == link ? list->next : list);
920 static __always_inline void
921 link_va(struct vmap_area *va, struct rb_root *root,
922 struct rb_node *parent, struct rb_node **link, struct list_head *head)
925 * VA is still not in the list, but we can
926 * identify its future previous list_head node.
928 if (likely(parent)) {
929 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
930 if (&parent->rb_right != link)
934 /* Insert to the rb-tree */
935 rb_link_node(&va->rb_node, parent, link);
936 if (root == &free_vmap_area_root) {
938 * Some explanation here. Just perform simple insertion
939 * to the tree. We do not set va->subtree_max_size to
940 * its current size before calling rb_insert_augmented().
941 * It is because of we populate the tree from the bottom
942 * to parent levels when the node _is_ in the tree.
944 * Therefore we set subtree_max_size to zero after insertion,
945 * to let __augment_tree_propagate_from() puts everything to
946 * the correct order later on.
948 rb_insert_augmented(&va->rb_node,
949 root, &free_vmap_area_rb_augment_cb);
950 va->subtree_max_size = 0;
952 rb_insert_color(&va->rb_node, root);
955 /* Address-sort this list */
956 list_add(&va->list, head);
959 static __always_inline void
960 unlink_va(struct vmap_area *va, struct rb_root *root)
962 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
965 if (root == &free_vmap_area_root)
966 rb_erase_augmented(&va->rb_node,
967 root, &free_vmap_area_rb_augment_cb);
969 rb_erase(&va->rb_node, root);
972 RB_CLEAR_NODE(&va->rb_node);
975 #if DEBUG_AUGMENT_PROPAGATE_CHECK
977 augment_tree_propagate_check(void)
979 struct vmap_area *va;
980 unsigned long computed_size;
982 list_for_each_entry(va, &free_vmap_area_list, list) {
983 computed_size = compute_subtree_max_size(va);
984 if (computed_size != va->subtree_max_size)
985 pr_emerg("tree is corrupted: %lu, %lu\n",
986 va_size(va), va->subtree_max_size);
992 * This function populates subtree_max_size from bottom to upper
993 * levels starting from VA point. The propagation must be done
994 * when VA size is modified by changing its va_start/va_end. Or
995 * in case of newly inserting of VA to the tree.
997 * It means that __augment_tree_propagate_from() must be called:
998 * - After VA has been inserted to the tree(free path);
999 * - After VA has been shrunk(allocation path);
1000 * - After VA has been increased(merging path).
1002 * Please note that, it does not mean that upper parent nodes
1003 * and their subtree_max_size are recalculated all the time up
1012 * For example if we modify the node 4, shrinking it to 2, then
1013 * no any modification is required. If we shrink the node 2 to 1
1014 * its subtree_max_size is updated only, and set to 1. If we shrink
1015 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1016 * node becomes 4--6.
1018 static __always_inline void
1019 augment_tree_propagate_from(struct vmap_area *va)
1022 * Populate the tree from bottom towards the root until
1023 * the calculated maximum available size of checked node
1024 * is equal to its current one.
1026 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1028 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1029 augment_tree_propagate_check();
1034 insert_vmap_area(struct vmap_area *va,
1035 struct rb_root *root, struct list_head *head)
1037 struct rb_node **link;
1038 struct rb_node *parent;
1040 link = find_va_links(va, root, NULL, &parent);
1042 link_va(va, root, parent, link, head);
1046 insert_vmap_area_augment(struct vmap_area *va,
1047 struct rb_node *from, struct rb_root *root,
1048 struct list_head *head)
1050 struct rb_node **link;
1051 struct rb_node *parent;
1054 link = find_va_links(va, NULL, from, &parent);
1056 link = find_va_links(va, root, NULL, &parent);
1059 link_va(va, root, parent, link, head);
1060 augment_tree_propagate_from(va);
1065 * Merge de-allocated chunk of VA memory with previous
1066 * and next free blocks. If coalesce is not done a new
1067 * free area is inserted. If VA has been merged, it is
1070 * Please note, it can return NULL in case of overlap
1071 * ranges, followed by WARN() report. Despite it is a
1072 * buggy behaviour, a system can be alive and keep
1075 static __always_inline struct vmap_area *
1076 merge_or_add_vmap_area(struct vmap_area *va,
1077 struct rb_root *root, struct list_head *head)
1079 struct vmap_area *sibling;
1080 struct list_head *next;
1081 struct rb_node **link;
1082 struct rb_node *parent;
1083 bool merged = false;
1086 * Find a place in the tree where VA potentially will be
1087 * inserted, unless it is merged with its sibling/siblings.
1089 link = find_va_links(va, root, NULL, &parent);
1094 * Get next node of VA to check if merging can be done.
1096 next = get_va_next_sibling(parent, link);
1097 if (unlikely(next == NULL))
1103 * |<------VA------>|<-----Next----->|
1108 sibling = list_entry(next, struct vmap_area, list);
1109 if (sibling->va_start == va->va_end) {
1110 sibling->va_start = va->va_start;
1112 /* Free vmap_area object. */
1113 kmem_cache_free(vmap_area_cachep, va);
1115 /* Point to the new merged area. */
1124 * |<-----Prev----->|<------VA------>|
1128 if (next->prev != head) {
1129 sibling = list_entry(next->prev, struct vmap_area, list);
1130 if (sibling->va_end == va->va_start) {
1132 * If both neighbors are coalesced, it is important
1133 * to unlink the "next" node first, followed by merging
1134 * with "previous" one. Otherwise the tree might not be
1135 * fully populated if a sibling's augmented value is
1136 * "normalized" because of rotation operations.
1139 unlink_va(va, root);
1141 sibling->va_end = va->va_end;
1143 /* Free vmap_area object. */
1144 kmem_cache_free(vmap_area_cachep, va);
1146 /* Point to the new merged area. */
1154 link_va(va, root, parent, link, head);
1159 static __always_inline struct vmap_area *
1160 merge_or_add_vmap_area_augment(struct vmap_area *va,
1161 struct rb_root *root, struct list_head *head)
1163 va = merge_or_add_vmap_area(va, root, head);
1165 augment_tree_propagate_from(va);
1170 static __always_inline bool
1171 is_within_this_va(struct vmap_area *va, unsigned long size,
1172 unsigned long align, unsigned long vstart)
1174 unsigned long nva_start_addr;
1176 if (va->va_start > vstart)
1177 nva_start_addr = ALIGN(va->va_start, align);
1179 nva_start_addr = ALIGN(vstart, align);
1181 /* Can be overflowed due to big size or alignment. */
1182 if (nva_start_addr + size < nva_start_addr ||
1183 nva_start_addr < vstart)
1186 return (nva_start_addr + size <= va->va_end);
1190 * Find the first free block(lowest start address) in the tree,
1191 * that will accomplish the request corresponding to passing
1192 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1193 * a search length is adjusted to account for worst case alignment
1196 static __always_inline struct vmap_area *
1197 find_vmap_lowest_match(unsigned long size, unsigned long align,
1198 unsigned long vstart, bool adjust_search_size)
1200 struct vmap_area *va;
1201 struct rb_node *node;
1202 unsigned long length;
1204 /* Start from the root. */
1205 node = free_vmap_area_root.rb_node;
1207 /* Adjust the search size for alignment overhead. */
1208 length = adjust_search_size ? size + align - 1 : size;
1211 va = rb_entry(node, struct vmap_area, rb_node);
1213 if (get_subtree_max_size(node->rb_left) >= length &&
1214 vstart < va->va_start) {
1215 node = node->rb_left;
1217 if (is_within_this_va(va, size, align, vstart))
1221 * Does not make sense to go deeper towards the right
1222 * sub-tree if it does not have a free block that is
1223 * equal or bigger to the requested search length.
1225 if (get_subtree_max_size(node->rb_right) >= length) {
1226 node = node->rb_right;
1231 * OK. We roll back and find the first right sub-tree,
1232 * that will satisfy the search criteria. It can happen
1233 * due to "vstart" restriction or an alignment overhead
1234 * that is bigger then PAGE_SIZE.
1236 while ((node = rb_parent(node))) {
1237 va = rb_entry(node, struct vmap_area, rb_node);
1238 if (is_within_this_va(va, size, align, vstart))
1241 if (get_subtree_max_size(node->rb_right) >= length &&
1242 vstart <= va->va_start) {
1244 * Shift the vstart forward. Please note, we update it with
1245 * parent's start address adding "1" because we do not want
1246 * to enter same sub-tree after it has already been checked
1247 * and no suitable free block found there.
1249 vstart = va->va_start + 1;
1250 node = node->rb_right;
1260 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1261 #include <linux/random.h>
1263 static struct vmap_area *
1264 find_vmap_lowest_linear_match(unsigned long size,
1265 unsigned long align, unsigned long vstart)
1267 struct vmap_area *va;
1269 list_for_each_entry(va, &free_vmap_area_list, list) {
1270 if (!is_within_this_va(va, size, align, vstart))
1280 find_vmap_lowest_match_check(unsigned long size, unsigned long align)
1282 struct vmap_area *va_1, *va_2;
1283 unsigned long vstart;
1286 get_random_bytes(&rnd, sizeof(rnd));
1287 vstart = VMALLOC_START + rnd;
1289 va_1 = find_vmap_lowest_match(size, align, vstart, false);
1290 va_2 = find_vmap_lowest_linear_match(size, align, vstart);
1293 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1294 va_1, va_2, vstart);
1300 FL_FIT_TYPE = 1, /* full fit */
1301 LE_FIT_TYPE = 2, /* left edge fit */
1302 RE_FIT_TYPE = 3, /* right edge fit */
1303 NE_FIT_TYPE = 4 /* no edge fit */
1306 static __always_inline enum fit_type
1307 classify_va_fit_type(struct vmap_area *va,
1308 unsigned long nva_start_addr, unsigned long size)
1312 /* Check if it is within VA. */
1313 if (nva_start_addr < va->va_start ||
1314 nva_start_addr + size > va->va_end)
1318 if (va->va_start == nva_start_addr) {
1319 if (va->va_end == nva_start_addr + size)
1323 } else if (va->va_end == nva_start_addr + size) {
1332 static __always_inline int
1333 adjust_va_to_fit_type(struct vmap_area *va,
1334 unsigned long nva_start_addr, unsigned long size,
1337 struct vmap_area *lva = NULL;
1339 if (type == FL_FIT_TYPE) {
1341 * No need to split VA, it fully fits.
1347 unlink_va(va, &free_vmap_area_root);
1348 kmem_cache_free(vmap_area_cachep, va);
1349 } else if (type == LE_FIT_TYPE) {
1351 * Split left edge of fit VA.
1357 va->va_start += size;
1358 } else if (type == RE_FIT_TYPE) {
1360 * Split right edge of fit VA.
1366 va->va_end = nva_start_addr;
1367 } else if (type == NE_FIT_TYPE) {
1369 * Split no edge of fit VA.
1375 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1376 if (unlikely(!lva)) {
1378 * For percpu allocator we do not do any pre-allocation
1379 * and leave it as it is. The reason is it most likely
1380 * never ends up with NE_FIT_TYPE splitting. In case of
1381 * percpu allocations offsets and sizes are aligned to
1382 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1383 * are its main fitting cases.
1385 * There are a few exceptions though, as an example it is
1386 * a first allocation (early boot up) when we have "one"
1387 * big free space that has to be split.
1389 * Also we can hit this path in case of regular "vmap"
1390 * allocations, if "this" current CPU was not preloaded.
1391 * See the comment in alloc_vmap_area() why. If so, then
1392 * GFP_NOWAIT is used instead to get an extra object for
1393 * split purpose. That is rare and most time does not
1396 * What happens if an allocation gets failed. Basically,
1397 * an "overflow" path is triggered to purge lazily freed
1398 * areas to free some memory, then, the "retry" path is
1399 * triggered to repeat one more time. See more details
1400 * in alloc_vmap_area() function.
1402 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1408 * Build the remainder.
1410 lva->va_start = va->va_start;
1411 lva->va_end = nva_start_addr;
1414 * Shrink this VA to remaining size.
1416 va->va_start = nva_start_addr + size;
1421 if (type != FL_FIT_TYPE) {
1422 augment_tree_propagate_from(va);
1424 if (lva) /* type == NE_FIT_TYPE */
1425 insert_vmap_area_augment(lva, &va->rb_node,
1426 &free_vmap_area_root, &free_vmap_area_list);
1433 * Returns a start address of the newly allocated area, if success.
1434 * Otherwise a vend is returned that indicates failure.
1436 static __always_inline unsigned long
1437 __alloc_vmap_area(unsigned long size, unsigned long align,
1438 unsigned long vstart, unsigned long vend)
1440 bool adjust_search_size = true;
1441 unsigned long nva_start_addr;
1442 struct vmap_area *va;
1447 * Do not adjust when:
1448 * a) align <= PAGE_SIZE, because it does not make any sense.
1449 * All blocks(their start addresses) are at least PAGE_SIZE
1451 * b) a short range where a requested size corresponds to exactly
1452 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1453 * With adjusted search length an allocation would not succeed.
1455 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1456 adjust_search_size = false;
1458 va = find_vmap_lowest_match(size, align, vstart, adjust_search_size);
1462 if (va->va_start > vstart)
1463 nva_start_addr = ALIGN(va->va_start, align);
1465 nva_start_addr = ALIGN(vstart, align);
1467 /* Check the "vend" restriction. */
1468 if (nva_start_addr + size > vend)
1471 /* Classify what we have found. */
1472 type = classify_va_fit_type(va, nva_start_addr, size);
1473 if (WARN_ON_ONCE(type == NOTHING_FIT))
1476 /* Update the free vmap_area. */
1477 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1481 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1482 find_vmap_lowest_match_check(size, align);
1485 return nva_start_addr;
1489 * Free a region of KVA allocated by alloc_vmap_area
1491 static void free_vmap_area(struct vmap_area *va)
1494 * Remove from the busy tree/list.
1496 spin_lock(&vmap_area_lock);
1497 unlink_va(va, &vmap_area_root);
1498 spin_unlock(&vmap_area_lock);
1501 * Insert/Merge it back to the free tree/list.
1503 spin_lock(&free_vmap_area_lock);
1504 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1505 spin_unlock(&free_vmap_area_lock);
1509 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1511 struct vmap_area *va = NULL;
1514 * Preload this CPU with one extra vmap_area object. It is used
1515 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1516 * a CPU that does an allocation is preloaded.
1518 * We do it in non-atomic context, thus it allows us to use more
1519 * permissive allocation masks to be more stable under low memory
1520 * condition and high memory pressure.
1522 if (!this_cpu_read(ne_fit_preload_node))
1523 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1527 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1528 kmem_cache_free(vmap_area_cachep, va);
1532 * Allocate a region of KVA of the specified size and alignment, within the
1535 static struct vmap_area *alloc_vmap_area(unsigned long size,
1536 unsigned long align,
1537 unsigned long vstart, unsigned long vend,
1538 int node, gfp_t gfp_mask)
1540 struct vmap_area *va;
1541 unsigned long freed;
1547 BUG_ON(offset_in_page(size));
1548 BUG_ON(!is_power_of_2(align));
1550 if (unlikely(!vmap_initialized))
1551 return ERR_PTR(-EBUSY);
1554 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1556 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1558 return ERR_PTR(-ENOMEM);
1561 * Only scan the relevant parts containing pointers to other objects
1562 * to avoid false negatives.
1564 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1567 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1568 addr = __alloc_vmap_area(size, align, vstart, vend);
1569 spin_unlock(&free_vmap_area_lock);
1572 * If an allocation fails, the "vend" address is
1573 * returned. Therefore trigger the overflow path.
1575 if (unlikely(addr == vend))
1578 va->va_start = addr;
1579 va->va_end = addr + size;
1582 spin_lock(&vmap_area_lock);
1583 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1584 spin_unlock(&vmap_area_lock);
1586 BUG_ON(!IS_ALIGNED(va->va_start, align));
1587 BUG_ON(va->va_start < vstart);
1588 BUG_ON(va->va_end > vend);
1590 ret = kasan_populate_vmalloc(addr, size);
1593 return ERR_PTR(ret);
1600 purge_vmap_area_lazy();
1606 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1613 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1614 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1617 kmem_cache_free(vmap_area_cachep, va);
1618 return ERR_PTR(-EBUSY);
1621 int register_vmap_purge_notifier(struct notifier_block *nb)
1623 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1625 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1627 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1629 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1631 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1634 * lazy_max_pages is the maximum amount of virtual address space we gather up
1635 * before attempting to purge with a TLB flush.
1637 * There is a tradeoff here: a larger number will cover more kernel page tables
1638 * and take slightly longer to purge, but it will linearly reduce the number of
1639 * global TLB flushes that must be performed. It would seem natural to scale
1640 * this number up linearly with the number of CPUs (because vmapping activity
1641 * could also scale linearly with the number of CPUs), however it is likely
1642 * that in practice, workloads might be constrained in other ways that mean
1643 * vmap activity will not scale linearly with CPUs. Also, I want to be
1644 * conservative and not introduce a big latency on huge systems, so go with
1645 * a less aggressive log scale. It will still be an improvement over the old
1646 * code, and it will be simple to change the scale factor if we find that it
1647 * becomes a problem on bigger systems.
1649 static unsigned long lazy_max_pages(void)
1653 log = fls(num_online_cpus());
1655 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1658 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1661 * Serialize vmap purging. There is no actual critical section protected
1662 * by this look, but we want to avoid concurrent calls for performance
1663 * reasons and to make the pcpu_get_vm_areas more deterministic.
1665 static DEFINE_MUTEX(vmap_purge_lock);
1667 /* for per-CPU blocks */
1668 static void purge_fragmented_blocks_allcpus(void);
1670 #ifdef CONFIG_X86_64
1672 * called before a call to iounmap() if the caller wants vm_area_struct's
1673 * immediately freed.
1675 void set_iounmap_nonlazy(void)
1677 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1679 #endif /* CONFIG_X86_64 */
1682 * Purges all lazily-freed vmap areas.
1684 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1686 unsigned long resched_threshold;
1687 struct list_head local_pure_list;
1688 struct vmap_area *va, *n_va;
1690 lockdep_assert_held(&vmap_purge_lock);
1692 spin_lock(&purge_vmap_area_lock);
1693 purge_vmap_area_root = RB_ROOT;
1694 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1695 spin_unlock(&purge_vmap_area_lock);
1697 if (unlikely(list_empty(&local_pure_list)))
1701 list_first_entry(&local_pure_list,
1702 struct vmap_area, list)->va_start);
1705 list_last_entry(&local_pure_list,
1706 struct vmap_area, list)->va_end);
1708 flush_tlb_kernel_range(start, end);
1709 resched_threshold = lazy_max_pages() << 1;
1711 spin_lock(&free_vmap_area_lock);
1712 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1713 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1714 unsigned long orig_start = va->va_start;
1715 unsigned long orig_end = va->va_end;
1718 * Finally insert or merge lazily-freed area. It is
1719 * detached and there is no need to "unlink" it from
1722 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1723 &free_vmap_area_list);
1728 if (is_vmalloc_or_module_addr((void *)orig_start))
1729 kasan_release_vmalloc(orig_start, orig_end,
1730 va->va_start, va->va_end);
1732 atomic_long_sub(nr, &vmap_lazy_nr);
1734 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1735 cond_resched_lock(&free_vmap_area_lock);
1737 spin_unlock(&free_vmap_area_lock);
1742 * Kick off a purge of the outstanding lazy areas.
1744 static void purge_vmap_area_lazy(void)
1746 mutex_lock(&vmap_purge_lock);
1747 purge_fragmented_blocks_allcpus();
1748 __purge_vmap_area_lazy(ULONG_MAX, 0);
1749 mutex_unlock(&vmap_purge_lock);
1752 static void drain_vmap_area_work(struct work_struct *work)
1754 unsigned long nr_lazy;
1757 mutex_lock(&vmap_purge_lock);
1758 __purge_vmap_area_lazy(ULONG_MAX, 0);
1759 mutex_unlock(&vmap_purge_lock);
1761 /* Recheck if further work is required. */
1762 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1763 } while (nr_lazy > lazy_max_pages());
1767 * Free a vmap area, caller ensuring that the area has been unmapped
1768 * and flush_cache_vunmap had been called for the correct range
1771 static void free_vmap_area_noflush(struct vmap_area *va)
1773 unsigned long nr_lazy;
1775 spin_lock(&vmap_area_lock);
1776 unlink_va(va, &vmap_area_root);
1777 spin_unlock(&vmap_area_lock);
1779 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1780 PAGE_SHIFT, &vmap_lazy_nr);
1783 * Merge or place it to the purge tree/list.
1785 spin_lock(&purge_vmap_area_lock);
1786 merge_or_add_vmap_area(va,
1787 &purge_vmap_area_root, &purge_vmap_area_list);
1788 spin_unlock(&purge_vmap_area_lock);
1790 /* After this point, we may free va at any time */
1791 if (unlikely(nr_lazy > lazy_max_pages()))
1792 schedule_work(&drain_vmap_work);
1796 * Free and unmap a vmap area
1798 static void free_unmap_vmap_area(struct vmap_area *va)
1800 flush_cache_vunmap(va->va_start, va->va_end);
1801 vunmap_range_noflush(va->va_start, va->va_end);
1802 if (debug_pagealloc_enabled_static())
1803 flush_tlb_kernel_range(va->va_start, va->va_end);
1805 free_vmap_area_noflush(va);
1808 static struct vmap_area *find_vmap_area(unsigned long addr)
1810 struct vmap_area *va;
1812 spin_lock(&vmap_area_lock);
1813 va = __find_vmap_area(addr);
1814 spin_unlock(&vmap_area_lock);
1819 /*** Per cpu kva allocator ***/
1822 * vmap space is limited especially on 32 bit architectures. Ensure there is
1823 * room for at least 16 percpu vmap blocks per CPU.
1826 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1827 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1828 * instead (we just need a rough idea)
1830 #if BITS_PER_LONG == 32
1831 #define VMALLOC_SPACE (128UL*1024*1024)
1833 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1836 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1837 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1838 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1839 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1840 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1841 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1842 #define VMAP_BBMAP_BITS \
1843 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1844 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1845 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1847 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1849 struct vmap_block_queue {
1851 struct list_head free;
1856 struct vmap_area *va;
1857 unsigned long free, dirty;
1858 unsigned long dirty_min, dirty_max; /*< dirty range */
1859 struct list_head free_list;
1860 struct rcu_head rcu_head;
1861 struct list_head purge;
1864 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1865 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1868 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1869 * in the free path. Could get rid of this if we change the API to return a
1870 * "cookie" from alloc, to be passed to free. But no big deal yet.
1872 static DEFINE_XARRAY(vmap_blocks);
1875 * We should probably have a fallback mechanism to allocate virtual memory
1876 * out of partially filled vmap blocks. However vmap block sizing should be
1877 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1881 static unsigned long addr_to_vb_idx(unsigned long addr)
1883 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1884 addr /= VMAP_BLOCK_SIZE;
1888 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1892 addr = va_start + (pages_off << PAGE_SHIFT);
1893 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1894 return (void *)addr;
1898 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1899 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1900 * @order: how many 2^order pages should be occupied in newly allocated block
1901 * @gfp_mask: flags for the page level allocator
1903 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1905 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1907 struct vmap_block_queue *vbq;
1908 struct vmap_block *vb;
1909 struct vmap_area *va;
1910 unsigned long vb_idx;
1914 node = numa_node_id();
1916 vb = kmalloc_node(sizeof(struct vmap_block),
1917 gfp_mask & GFP_RECLAIM_MASK, node);
1919 return ERR_PTR(-ENOMEM);
1921 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1922 VMALLOC_START, VMALLOC_END,
1926 return ERR_CAST(va);
1929 vaddr = vmap_block_vaddr(va->va_start, 0);
1930 spin_lock_init(&vb->lock);
1932 /* At least something should be left free */
1933 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1934 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1936 vb->dirty_min = VMAP_BBMAP_BITS;
1938 INIT_LIST_HEAD(&vb->free_list);
1940 vb_idx = addr_to_vb_idx(va->va_start);
1941 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1945 return ERR_PTR(err);
1948 vbq = &get_cpu_var(vmap_block_queue);
1949 spin_lock(&vbq->lock);
1950 list_add_tail_rcu(&vb->free_list, &vbq->free);
1951 spin_unlock(&vbq->lock);
1952 put_cpu_var(vmap_block_queue);
1957 static void free_vmap_block(struct vmap_block *vb)
1959 struct vmap_block *tmp;
1961 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1964 free_vmap_area_noflush(vb->va);
1965 kfree_rcu(vb, rcu_head);
1968 static void purge_fragmented_blocks(int cpu)
1971 struct vmap_block *vb;
1972 struct vmap_block *n_vb;
1973 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1976 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1978 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1981 spin_lock(&vb->lock);
1982 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1983 vb->free = 0; /* prevent further allocs after releasing lock */
1984 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1986 vb->dirty_max = VMAP_BBMAP_BITS;
1987 spin_lock(&vbq->lock);
1988 list_del_rcu(&vb->free_list);
1989 spin_unlock(&vbq->lock);
1990 spin_unlock(&vb->lock);
1991 list_add_tail(&vb->purge, &purge);
1993 spin_unlock(&vb->lock);
1997 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1998 list_del(&vb->purge);
1999 free_vmap_block(vb);
2003 static void purge_fragmented_blocks_allcpus(void)
2007 for_each_possible_cpu(cpu)
2008 purge_fragmented_blocks(cpu);
2011 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2013 struct vmap_block_queue *vbq;
2014 struct vmap_block *vb;
2018 BUG_ON(offset_in_page(size));
2019 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2020 if (WARN_ON(size == 0)) {
2022 * Allocating 0 bytes isn't what caller wants since
2023 * get_order(0) returns funny result. Just warn and terminate
2028 order = get_order(size);
2031 vbq = &get_cpu_var(vmap_block_queue);
2032 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2033 unsigned long pages_off;
2035 spin_lock(&vb->lock);
2036 if (vb->free < (1UL << order)) {
2037 spin_unlock(&vb->lock);
2041 pages_off = VMAP_BBMAP_BITS - vb->free;
2042 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2043 vb->free -= 1UL << order;
2044 if (vb->free == 0) {
2045 spin_lock(&vbq->lock);
2046 list_del_rcu(&vb->free_list);
2047 spin_unlock(&vbq->lock);
2050 spin_unlock(&vb->lock);
2054 put_cpu_var(vmap_block_queue);
2057 /* Allocate new block if nothing was found */
2059 vaddr = new_vmap_block(order, gfp_mask);
2064 static void vb_free(unsigned long addr, unsigned long size)
2066 unsigned long offset;
2068 struct vmap_block *vb;
2070 BUG_ON(offset_in_page(size));
2071 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2073 flush_cache_vunmap(addr, addr + size);
2075 order = get_order(size);
2076 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2077 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2079 vunmap_range_noflush(addr, addr + size);
2081 if (debug_pagealloc_enabled_static())
2082 flush_tlb_kernel_range(addr, addr + size);
2084 spin_lock(&vb->lock);
2086 /* Expand dirty range */
2087 vb->dirty_min = min(vb->dirty_min, offset);
2088 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2090 vb->dirty += 1UL << order;
2091 if (vb->dirty == VMAP_BBMAP_BITS) {
2093 spin_unlock(&vb->lock);
2094 free_vmap_block(vb);
2096 spin_unlock(&vb->lock);
2099 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2103 if (unlikely(!vmap_initialized))
2108 for_each_possible_cpu(cpu) {
2109 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2110 struct vmap_block *vb;
2113 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2114 spin_lock(&vb->lock);
2115 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2116 unsigned long va_start = vb->va->va_start;
2119 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2120 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2122 start = min(s, start);
2127 spin_unlock(&vb->lock);
2132 mutex_lock(&vmap_purge_lock);
2133 purge_fragmented_blocks_allcpus();
2134 if (!__purge_vmap_area_lazy(start, end) && flush)
2135 flush_tlb_kernel_range(start, end);
2136 mutex_unlock(&vmap_purge_lock);
2140 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2142 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2143 * to amortize TLB flushing overheads. What this means is that any page you
2144 * have now, may, in a former life, have been mapped into kernel virtual
2145 * address by the vmap layer and so there might be some CPUs with TLB entries
2146 * still referencing that page (additional to the regular 1:1 kernel mapping).
2148 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2149 * be sure that none of the pages we have control over will have any aliases
2150 * from the vmap layer.
2152 void vm_unmap_aliases(void)
2154 unsigned long start = ULONG_MAX, end = 0;
2157 _vm_unmap_aliases(start, end, flush);
2159 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2162 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2163 * @mem: the pointer returned by vm_map_ram
2164 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2166 void vm_unmap_ram(const void *mem, unsigned int count)
2168 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2169 unsigned long addr = (unsigned long)mem;
2170 struct vmap_area *va;
2174 BUG_ON(addr < VMALLOC_START);
2175 BUG_ON(addr > VMALLOC_END);
2176 BUG_ON(!PAGE_ALIGNED(addr));
2178 kasan_poison_vmalloc(mem, size);
2180 if (likely(count <= VMAP_MAX_ALLOC)) {
2181 debug_check_no_locks_freed(mem, size);
2182 vb_free(addr, size);
2186 va = find_vmap_area(addr);
2188 debug_check_no_locks_freed((void *)va->va_start,
2189 (va->va_end - va->va_start));
2190 free_unmap_vmap_area(va);
2192 EXPORT_SYMBOL(vm_unmap_ram);
2195 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2196 * @pages: an array of pointers to the pages to be mapped
2197 * @count: number of pages
2198 * @node: prefer to allocate data structures on this node
2200 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2201 * faster than vmap so it's good. But if you mix long-life and short-life
2202 * objects with vm_map_ram(), it could consume lots of address space through
2203 * fragmentation (especially on a 32bit machine). You could see failures in
2204 * the end. Please use this function for short-lived objects.
2206 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2208 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2210 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2214 if (likely(count <= VMAP_MAX_ALLOC)) {
2215 mem = vb_alloc(size, GFP_KERNEL);
2218 addr = (unsigned long)mem;
2220 struct vmap_area *va;
2221 va = alloc_vmap_area(size, PAGE_SIZE,
2222 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2226 addr = va->va_start;
2230 kasan_unpoison_vmalloc(mem, size);
2232 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2233 pages, PAGE_SHIFT) < 0) {
2234 vm_unmap_ram(mem, count);
2240 EXPORT_SYMBOL(vm_map_ram);
2242 static struct vm_struct *vmlist __initdata;
2244 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2246 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2247 return vm->page_order;
2253 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2255 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2256 vm->page_order = order;
2263 * vm_area_add_early - add vmap area early during boot
2264 * @vm: vm_struct to add
2266 * This function is used to add fixed kernel vm area to vmlist before
2267 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2268 * should contain proper values and the other fields should be zero.
2270 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2272 void __init vm_area_add_early(struct vm_struct *vm)
2274 struct vm_struct *tmp, **p;
2276 BUG_ON(vmap_initialized);
2277 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2278 if (tmp->addr >= vm->addr) {
2279 BUG_ON(tmp->addr < vm->addr + vm->size);
2282 BUG_ON(tmp->addr + tmp->size > vm->addr);
2289 * vm_area_register_early - register vmap area early during boot
2290 * @vm: vm_struct to register
2291 * @align: requested alignment
2293 * This function is used to register kernel vm area before
2294 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2295 * proper values on entry and other fields should be zero. On return,
2296 * vm->addr contains the allocated address.
2298 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2300 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2302 unsigned long addr = ALIGN(VMALLOC_START, align);
2303 struct vm_struct *cur, **p;
2305 BUG_ON(vmap_initialized);
2307 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2308 if ((unsigned long)cur->addr - addr >= vm->size)
2310 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2313 BUG_ON(addr > VMALLOC_END - vm->size);
2314 vm->addr = (void *)addr;
2317 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2320 static void vmap_init_free_space(void)
2322 unsigned long vmap_start = 1;
2323 const unsigned long vmap_end = ULONG_MAX;
2324 struct vmap_area *busy, *free;
2328 * -|-----|.....|-----|-----|-----|.....|-
2330 * |<--------------------------------->|
2332 list_for_each_entry(busy, &vmap_area_list, list) {
2333 if (busy->va_start - vmap_start > 0) {
2334 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2335 if (!WARN_ON_ONCE(!free)) {
2336 free->va_start = vmap_start;
2337 free->va_end = busy->va_start;
2339 insert_vmap_area_augment(free, NULL,
2340 &free_vmap_area_root,
2341 &free_vmap_area_list);
2345 vmap_start = busy->va_end;
2348 if (vmap_end - vmap_start > 0) {
2349 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2350 if (!WARN_ON_ONCE(!free)) {
2351 free->va_start = vmap_start;
2352 free->va_end = vmap_end;
2354 insert_vmap_area_augment(free, NULL,
2355 &free_vmap_area_root,
2356 &free_vmap_area_list);
2361 void __init vmalloc_init(void)
2363 struct vmap_area *va;
2364 struct vm_struct *tmp;
2368 * Create the cache for vmap_area objects.
2370 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2372 for_each_possible_cpu(i) {
2373 struct vmap_block_queue *vbq;
2374 struct vfree_deferred *p;
2376 vbq = &per_cpu(vmap_block_queue, i);
2377 spin_lock_init(&vbq->lock);
2378 INIT_LIST_HEAD(&vbq->free);
2379 p = &per_cpu(vfree_deferred, i);
2380 init_llist_head(&p->list);
2381 INIT_WORK(&p->wq, free_work);
2384 /* Import existing vmlist entries. */
2385 for (tmp = vmlist; tmp; tmp = tmp->next) {
2386 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2387 if (WARN_ON_ONCE(!va))
2390 va->va_start = (unsigned long)tmp->addr;
2391 va->va_end = va->va_start + tmp->size;
2393 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2397 * Now we can initialize a free vmap space.
2399 vmap_init_free_space();
2400 vmap_initialized = true;
2403 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2404 struct vmap_area *va, unsigned long flags, const void *caller)
2407 vm->addr = (void *)va->va_start;
2408 vm->size = va->va_end - va->va_start;
2409 vm->caller = caller;
2413 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2414 unsigned long flags, const void *caller)
2416 spin_lock(&vmap_area_lock);
2417 setup_vmalloc_vm_locked(vm, va, flags, caller);
2418 spin_unlock(&vmap_area_lock);
2421 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2424 * Before removing VM_UNINITIALIZED,
2425 * we should make sure that vm has proper values.
2426 * Pair with smp_rmb() in show_numa_info().
2429 vm->flags &= ~VM_UNINITIALIZED;
2432 static struct vm_struct *__get_vm_area_node(unsigned long size,
2433 unsigned long align, unsigned long shift, unsigned long flags,
2434 unsigned long start, unsigned long end, int node,
2435 gfp_t gfp_mask, const void *caller)
2437 struct vmap_area *va;
2438 struct vm_struct *area;
2439 unsigned long requested_size = size;
2441 BUG_ON(in_interrupt());
2442 size = ALIGN(size, 1ul << shift);
2443 if (unlikely(!size))
2446 if (flags & VM_IOREMAP)
2447 align = 1ul << clamp_t(int, get_count_order_long(size),
2448 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2450 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2451 if (unlikely(!area))
2454 if (!(flags & VM_NO_GUARD))
2457 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2463 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2465 setup_vmalloc_vm(area, va, flags, caller);
2470 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2471 unsigned long start, unsigned long end,
2474 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2475 NUMA_NO_NODE, GFP_KERNEL, caller);
2479 * get_vm_area - reserve a contiguous kernel virtual area
2480 * @size: size of the area
2481 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2483 * Search an area of @size in the kernel virtual mapping area,
2484 * and reserved it for out purposes. Returns the area descriptor
2485 * on success or %NULL on failure.
2487 * Return: the area descriptor on success or %NULL on failure.
2489 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2491 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2492 VMALLOC_START, VMALLOC_END,
2493 NUMA_NO_NODE, GFP_KERNEL,
2494 __builtin_return_address(0));
2497 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2500 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2501 VMALLOC_START, VMALLOC_END,
2502 NUMA_NO_NODE, GFP_KERNEL, caller);
2506 * find_vm_area - find a continuous kernel virtual area
2507 * @addr: base address
2509 * Search for the kernel VM area starting at @addr, and return it.
2510 * It is up to the caller to do all required locking to keep the returned
2513 * Return: the area descriptor on success or %NULL on failure.
2515 struct vm_struct *find_vm_area(const void *addr)
2517 struct vmap_area *va;
2519 va = find_vmap_area((unsigned long)addr);
2527 * remove_vm_area - find and remove a continuous kernel virtual area
2528 * @addr: base address
2530 * Search for the kernel VM area starting at @addr, and remove it.
2531 * This function returns the found VM area, but using it is NOT safe
2532 * on SMP machines, except for its size or flags.
2534 * Return: the area descriptor on success or %NULL on failure.
2536 struct vm_struct *remove_vm_area(const void *addr)
2538 struct vmap_area *va;
2542 spin_lock(&vmap_area_lock);
2543 va = __find_vmap_area((unsigned long)addr);
2545 struct vm_struct *vm = va->vm;
2548 spin_unlock(&vmap_area_lock);
2550 kasan_free_shadow(vm);
2551 free_unmap_vmap_area(va);
2556 spin_unlock(&vmap_area_lock);
2560 static inline void set_area_direct_map(const struct vm_struct *area,
2561 int (*set_direct_map)(struct page *page))
2565 /* HUGE_VMALLOC passes small pages to set_direct_map */
2566 for (i = 0; i < area->nr_pages; i++)
2567 if (page_address(area->pages[i]))
2568 set_direct_map(area->pages[i]);
2571 /* Handle removing and resetting vm mappings related to the vm_struct. */
2572 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2574 unsigned long start = ULONG_MAX, end = 0;
2575 unsigned int page_order = vm_area_page_order(area);
2576 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2580 remove_vm_area(area->addr);
2582 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2587 * If not deallocating pages, just do the flush of the VM area and
2590 if (!deallocate_pages) {
2596 * If execution gets here, flush the vm mapping and reset the direct
2597 * map. Find the start and end range of the direct mappings to make sure
2598 * the vm_unmap_aliases() flush includes the direct map.
2600 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2601 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2603 unsigned long page_size;
2605 page_size = PAGE_SIZE << page_order;
2606 start = min(addr, start);
2607 end = max(addr + page_size, end);
2613 * Set direct map to something invalid so that it won't be cached if
2614 * there are any accesses after the TLB flush, then flush the TLB and
2615 * reset the direct map permissions to the default.
2617 set_area_direct_map(area, set_direct_map_invalid_noflush);
2618 _vm_unmap_aliases(start, end, flush_dmap);
2619 set_area_direct_map(area, set_direct_map_default_noflush);
2622 static void __vunmap(const void *addr, int deallocate_pages)
2624 struct vm_struct *area;
2629 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2633 area = find_vm_area(addr);
2634 if (unlikely(!area)) {
2635 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2640 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2641 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2643 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2645 vm_remove_mappings(area, deallocate_pages);
2647 if (deallocate_pages) {
2648 unsigned int page_order = vm_area_page_order(area);
2649 int i, step = 1U << page_order;
2651 for (i = 0; i < area->nr_pages; i += step) {
2652 struct page *page = area->pages[i];
2655 mod_memcg_page_state(page, MEMCG_VMALLOC, -step);
2656 __free_pages(page, page_order);
2659 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2661 kvfree(area->pages);
2667 static inline void __vfree_deferred(const void *addr)
2670 * Use raw_cpu_ptr() because this can be called from preemptible
2671 * context. Preemption is absolutely fine here, because the llist_add()
2672 * implementation is lockless, so it works even if we are adding to
2673 * another cpu's list. schedule_work() should be fine with this too.
2675 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2677 if (llist_add((struct llist_node *)addr, &p->list))
2678 schedule_work(&p->wq);
2682 * vfree_atomic - release memory allocated by vmalloc()
2683 * @addr: memory base address
2685 * This one is just like vfree() but can be called in any atomic context
2688 void vfree_atomic(const void *addr)
2692 kmemleak_free(addr);
2696 __vfree_deferred(addr);
2699 static void __vfree(const void *addr)
2701 if (unlikely(in_interrupt()))
2702 __vfree_deferred(addr);
2708 * vfree - Release memory allocated by vmalloc()
2709 * @addr: Memory base address
2711 * Free the virtually continuous memory area starting at @addr, as obtained
2712 * from one of the vmalloc() family of APIs. This will usually also free the
2713 * physical memory underlying the virtual allocation, but that memory is
2714 * reference counted, so it will not be freed until the last user goes away.
2716 * If @addr is NULL, no operation is performed.
2719 * May sleep if called *not* from interrupt context.
2720 * Must not be called in NMI context (strictly speaking, it could be
2721 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2722 * conventions for vfree() arch-dependent would be a really bad idea).
2724 void vfree(const void *addr)
2728 kmemleak_free(addr);
2730 might_sleep_if(!in_interrupt());
2737 EXPORT_SYMBOL(vfree);
2740 * vunmap - release virtual mapping obtained by vmap()
2741 * @addr: memory base address
2743 * Free the virtually contiguous memory area starting at @addr,
2744 * which was created from the page array passed to vmap().
2746 * Must not be called in interrupt context.
2748 void vunmap(const void *addr)
2750 BUG_ON(in_interrupt());
2755 EXPORT_SYMBOL(vunmap);
2758 * vmap - map an array of pages into virtually contiguous space
2759 * @pages: array of page pointers
2760 * @count: number of pages to map
2761 * @flags: vm_area->flags
2762 * @prot: page protection for the mapping
2764 * Maps @count pages from @pages into contiguous kernel virtual space.
2765 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2766 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2767 * are transferred from the caller to vmap(), and will be freed / dropped when
2768 * vfree() is called on the return value.
2770 * Return: the address of the area or %NULL on failure
2772 void *vmap(struct page **pages, unsigned int count,
2773 unsigned long flags, pgprot_t prot)
2775 struct vm_struct *area;
2777 unsigned long size; /* In bytes */
2782 * Your top guard is someone else's bottom guard. Not having a top
2783 * guard compromises someone else's mappings too.
2785 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2786 flags &= ~VM_NO_GUARD;
2788 if (count > totalram_pages())
2791 size = (unsigned long)count << PAGE_SHIFT;
2792 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2796 addr = (unsigned long)area->addr;
2797 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2798 pages, PAGE_SHIFT) < 0) {
2803 if (flags & VM_MAP_PUT_PAGES) {
2804 area->pages = pages;
2805 area->nr_pages = count;
2809 EXPORT_SYMBOL(vmap);
2811 #ifdef CONFIG_VMAP_PFN
2812 struct vmap_pfn_data {
2813 unsigned long *pfns;
2818 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2820 struct vmap_pfn_data *data = private;
2822 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2824 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2829 * vmap_pfn - map an array of PFNs into virtually contiguous space
2830 * @pfns: array of PFNs
2831 * @count: number of pages to map
2832 * @prot: page protection for the mapping
2834 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2835 * the start address of the mapping.
2837 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2839 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2840 struct vm_struct *area;
2842 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2843 __builtin_return_address(0));
2846 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2847 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2853 EXPORT_SYMBOL_GPL(vmap_pfn);
2854 #endif /* CONFIG_VMAP_PFN */
2856 static inline unsigned int
2857 vm_area_alloc_pages(gfp_t gfp, int nid,
2858 unsigned int order, unsigned int nr_pages, struct page **pages)
2860 unsigned int nr_allocated = 0;
2865 * For order-0 pages we make use of bulk allocator, if
2866 * the page array is partly or not at all populated due
2867 * to fails, fallback to a single page allocator that is
2871 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2873 while (nr_allocated < nr_pages) {
2874 unsigned int nr, nr_pages_request;
2877 * A maximum allowed request is hard-coded and is 100
2878 * pages per call. That is done in order to prevent a
2879 * long preemption off scenario in the bulk-allocator
2880 * so the range is [1:100].
2882 nr_pages_request = min(100U, nr_pages - nr_allocated);
2884 /* memory allocation should consider mempolicy, we can't
2885 * wrongly use nearest node when nid == NUMA_NO_NODE,
2886 * otherwise memory may be allocated in only one node,
2887 * but mempolcy want to alloc memory by interleaving.
2889 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2890 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2892 pages + nr_allocated);
2895 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2897 pages + nr_allocated);
2903 * If zero or pages were obtained partly,
2904 * fallback to a single page allocator.
2906 if (nr != nr_pages_request)
2911 * Compound pages required for remap_vmalloc_page if
2916 /* High-order pages or fallback path if "bulk" fails. */
2918 while (nr_allocated < nr_pages) {
2919 if (fatal_signal_pending(current))
2922 if (nid == NUMA_NO_NODE)
2923 page = alloc_pages(gfp, order);
2925 page = alloc_pages_node(nid, gfp, order);
2926 if (unlikely(!page))
2930 * Careful, we allocate and map page-order pages, but
2931 * tracking is done per PAGE_SIZE page so as to keep the
2932 * vm_struct APIs independent of the physical/mapped size.
2934 for (i = 0; i < (1U << order); i++)
2935 pages[nr_allocated + i] = page + i;
2938 nr_allocated += 1U << order;
2941 return nr_allocated;
2944 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2945 pgprot_t prot, unsigned int page_shift,
2948 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2949 const gfp_t orig_gfp_mask = gfp_mask;
2950 bool nofail = gfp_mask & __GFP_NOFAIL;
2951 unsigned long addr = (unsigned long)area->addr;
2952 unsigned long size = get_vm_area_size(area);
2953 unsigned long array_size;
2954 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2955 unsigned int page_order;
2959 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2960 gfp_mask |= __GFP_NOWARN;
2961 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2962 gfp_mask |= __GFP_HIGHMEM;
2964 /* Please note that the recursion is strictly bounded. */
2965 if (array_size > PAGE_SIZE) {
2966 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2969 area->pages = kmalloc_node(array_size, nested_gfp, node);
2973 warn_alloc(orig_gfp_mask, NULL,
2974 "vmalloc error: size %lu, failed to allocated page array size %lu",
2975 nr_small_pages * PAGE_SIZE, array_size);
2980 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2981 page_order = vm_area_page_order(area);
2983 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2984 page_order, nr_small_pages, area->pages);
2986 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2987 if (gfp_mask & __GFP_ACCOUNT) {
2988 int i, step = 1U << page_order;
2990 for (i = 0; i < area->nr_pages; i += step)
2991 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC,
2996 * If not enough pages were obtained to accomplish an
2997 * allocation request, free them via __vfree() if any.
2999 if (area->nr_pages != nr_small_pages) {
3000 warn_alloc(orig_gfp_mask, NULL,
3001 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3002 area->nr_pages * PAGE_SIZE, page_order);
3007 * page tables allocations ignore external gfp mask, enforce it
3010 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3011 flags = memalloc_nofs_save();
3012 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3013 flags = memalloc_noio_save();
3016 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3018 if (nofail && (ret < 0))
3019 schedule_timeout_uninterruptible(1);
3020 } while (nofail && (ret < 0));
3022 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3023 memalloc_nofs_restore(flags);
3024 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3025 memalloc_noio_restore(flags);
3028 warn_alloc(orig_gfp_mask, NULL,
3029 "vmalloc error: size %lu, failed to map pages",
3030 area->nr_pages * PAGE_SIZE);
3037 __vfree(area->addr);
3042 * __vmalloc_node_range - allocate virtually contiguous memory
3043 * @size: allocation size
3044 * @align: desired alignment
3045 * @start: vm area range start
3046 * @end: vm area range end
3047 * @gfp_mask: flags for the page level allocator
3048 * @prot: protection mask for the allocated pages
3049 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3050 * @node: node to use for allocation or NUMA_NO_NODE
3051 * @caller: caller's return address
3053 * Allocate enough pages to cover @size from the page level
3054 * allocator with @gfp_mask flags. Please note that the full set of gfp
3055 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3057 * Zone modifiers are not supported. From the reclaim modifiers
3058 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3059 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3060 * __GFP_RETRY_MAYFAIL are not supported).
3062 * __GFP_NOWARN can be used to suppress failures messages.
3064 * Map them into contiguous kernel virtual space, using a pagetable
3065 * protection of @prot.
3067 * Return: the address of the area or %NULL on failure
3069 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3070 unsigned long start, unsigned long end, gfp_t gfp_mask,
3071 pgprot_t prot, unsigned long vm_flags, int node,
3074 struct vm_struct *area;
3076 unsigned long real_size = size;
3077 unsigned long real_align = align;
3078 unsigned int shift = PAGE_SHIFT;
3080 if (WARN_ON_ONCE(!size))
3083 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3084 warn_alloc(gfp_mask, NULL,
3085 "vmalloc error: size %lu, exceeds total pages",
3090 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
3091 unsigned long size_per_node;
3094 * Try huge pages. Only try for PAGE_KERNEL allocations,
3095 * others like modules don't yet expect huge pages in
3096 * their allocations due to apply_to_page_range not
3100 size_per_node = size;
3101 if (node == NUMA_NO_NODE)
3102 size_per_node /= num_online_nodes();
3103 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3106 shift = arch_vmap_pte_supported_shift(size_per_node);
3108 align = max(real_align, 1UL << shift);
3109 size = ALIGN(real_size, 1UL << shift);
3113 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3114 VM_UNINITIALIZED | vm_flags, start, end, node,
3117 bool nofail = gfp_mask & __GFP_NOFAIL;
3118 warn_alloc(gfp_mask, NULL,
3119 "vmalloc error: size %lu, vm_struct allocation failed%s",
3120 real_size, (nofail) ? ". Retrying." : "");
3122 schedule_timeout_uninterruptible(1);
3128 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3133 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3134 * flag. It means that vm_struct is not fully initialized.
3135 * Now, it is fully initialized, so remove this flag here.
3137 clear_vm_uninitialized_flag(area);
3139 size = PAGE_ALIGN(size);
3140 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3141 kmemleak_vmalloc(area, size, gfp_mask);
3146 if (shift > PAGE_SHIFT) {
3157 * __vmalloc_node - allocate virtually contiguous memory
3158 * @size: allocation size
3159 * @align: desired alignment
3160 * @gfp_mask: flags for the page level allocator
3161 * @node: node to use for allocation or NUMA_NO_NODE
3162 * @caller: caller's return address
3164 * Allocate enough pages to cover @size from the page level allocator with
3165 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3167 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3168 * and __GFP_NOFAIL are not supported
3170 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3173 * Return: pointer to the allocated memory or %NULL on error
3175 void *__vmalloc_node(unsigned long size, unsigned long align,
3176 gfp_t gfp_mask, int node, const void *caller)
3178 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3179 gfp_mask, PAGE_KERNEL, 0, node, caller);
3182 * This is only for performance analysis of vmalloc and stress purpose.
3183 * It is required by vmalloc test module, therefore do not use it other
3186 #ifdef CONFIG_TEST_VMALLOC_MODULE
3187 EXPORT_SYMBOL_GPL(__vmalloc_node);
3190 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3192 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3193 __builtin_return_address(0));
3195 EXPORT_SYMBOL(__vmalloc);
3198 * vmalloc - allocate virtually contiguous memory
3199 * @size: allocation size
3201 * Allocate enough pages to cover @size from the page level
3202 * allocator and map them into contiguous kernel virtual space.
3204 * For tight control over page level allocator and protection flags
3205 * use __vmalloc() instead.
3207 * Return: pointer to the allocated memory or %NULL on error
3209 void *vmalloc(unsigned long size)
3211 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3212 __builtin_return_address(0));
3214 EXPORT_SYMBOL(vmalloc);
3217 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3218 * @size: allocation size
3220 * Allocate enough non-huge pages to cover @size from the page level
3221 * allocator and map them into contiguous kernel virtual space.
3223 * Return: pointer to the allocated memory or %NULL on error
3225 void *vmalloc_no_huge(unsigned long size)
3227 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3228 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3229 NUMA_NO_NODE, __builtin_return_address(0));
3231 EXPORT_SYMBOL(vmalloc_no_huge);
3234 * vzalloc - allocate virtually contiguous memory with zero fill
3235 * @size: allocation size
3237 * Allocate enough pages to cover @size from the page level
3238 * allocator and map them into contiguous kernel virtual space.
3239 * The memory allocated is set to zero.
3241 * For tight control over page level allocator and protection flags
3242 * use __vmalloc() instead.
3244 * Return: pointer to the allocated memory or %NULL on error
3246 void *vzalloc(unsigned long size)
3248 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3249 __builtin_return_address(0));
3251 EXPORT_SYMBOL(vzalloc);
3254 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3255 * @size: allocation size
3257 * The resulting memory area is zeroed so it can be mapped to userspace
3258 * without leaking data.
3260 * Return: pointer to the allocated memory or %NULL on error
3262 void *vmalloc_user(unsigned long size)
3264 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3265 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3266 VM_USERMAP, NUMA_NO_NODE,
3267 __builtin_return_address(0));
3269 EXPORT_SYMBOL(vmalloc_user);
3272 * vmalloc_node - allocate memory on a specific node
3273 * @size: allocation size
3276 * Allocate enough pages to cover @size from the page level
3277 * allocator and map them into contiguous kernel virtual space.
3279 * For tight control over page level allocator and protection flags
3280 * use __vmalloc() instead.
3282 * Return: pointer to the allocated memory or %NULL on error
3284 void *vmalloc_node(unsigned long size, int node)
3286 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3287 __builtin_return_address(0));
3289 EXPORT_SYMBOL(vmalloc_node);
3292 * vzalloc_node - allocate memory on a specific node with zero fill
3293 * @size: allocation size
3296 * Allocate enough pages to cover @size from the page level
3297 * allocator and map them into contiguous kernel virtual space.
3298 * The memory allocated is set to zero.
3300 * Return: pointer to the allocated memory or %NULL on error
3302 void *vzalloc_node(unsigned long size, int node)
3304 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3305 __builtin_return_address(0));
3307 EXPORT_SYMBOL(vzalloc_node);
3309 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3310 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3311 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3312 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3315 * 64b systems should always have either DMA or DMA32 zones. For others
3316 * GFP_DMA32 should do the right thing and use the normal zone.
3318 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3322 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3323 * @size: allocation size
3325 * Allocate enough 32bit PA addressable pages to cover @size from the
3326 * page level allocator and map them into contiguous kernel virtual space.
3328 * Return: pointer to the allocated memory or %NULL on error
3330 void *vmalloc_32(unsigned long size)
3332 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3333 __builtin_return_address(0));
3335 EXPORT_SYMBOL(vmalloc_32);
3338 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3339 * @size: allocation size
3341 * The resulting memory area is 32bit addressable and zeroed so it can be
3342 * mapped to userspace without leaking data.
3344 * Return: pointer to the allocated memory or %NULL on error
3346 void *vmalloc_32_user(unsigned long size)
3348 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3349 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3350 VM_USERMAP, NUMA_NO_NODE,
3351 __builtin_return_address(0));
3353 EXPORT_SYMBOL(vmalloc_32_user);
3356 * small helper routine , copy contents to buf from addr.
3357 * If the page is not present, fill zero.
3360 static int aligned_vread(char *buf, char *addr, unsigned long count)
3366 unsigned long offset, length;
3368 offset = offset_in_page(addr);
3369 length = PAGE_SIZE - offset;
3372 p = vmalloc_to_page(addr);
3374 * To do safe access to this _mapped_ area, we need
3375 * lock. But adding lock here means that we need to add
3376 * overhead of vmalloc()/vfree() calls for this _debug_
3377 * interface, rarely used. Instead of that, we'll use
3378 * kmap() and get small overhead in this access function.
3381 /* We can expect USER0 is not used -- see vread() */
3382 void *map = kmap_atomic(p);
3383 memcpy(buf, map + offset, length);
3386 memset(buf, 0, length);
3397 * vread() - read vmalloc area in a safe way.
3398 * @buf: buffer for reading data
3399 * @addr: vm address.
3400 * @count: number of bytes to be read.
3402 * This function checks that addr is a valid vmalloc'ed area, and
3403 * copy data from that area to a given buffer. If the given memory range
3404 * of [addr...addr+count) includes some valid address, data is copied to
3405 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3406 * IOREMAP area is treated as memory hole and no copy is done.
3408 * If [addr...addr+count) doesn't includes any intersects with alive
3409 * vm_struct area, returns 0. @buf should be kernel's buffer.
3411 * Note: In usual ops, vread() is never necessary because the caller
3412 * should know vmalloc() area is valid and can use memcpy().
3413 * This is for routines which have to access vmalloc area without
3414 * any information, as /proc/kcore.
3416 * Return: number of bytes for which addr and buf should be increased
3417 * (same number as @count) or %0 if [addr...addr+count) doesn't
3418 * include any intersection with valid vmalloc area
3420 long vread(char *buf, char *addr, unsigned long count)
3422 struct vmap_area *va;
3423 struct vm_struct *vm;
3424 char *vaddr, *buf_start = buf;
3425 unsigned long buflen = count;
3428 /* Don't allow overflow */
3429 if ((unsigned long) addr + count < count)
3430 count = -(unsigned long) addr;
3432 spin_lock(&vmap_area_lock);
3433 va = find_vmap_area_exceed_addr((unsigned long)addr);
3437 /* no intersects with alive vmap_area */
3438 if ((unsigned long)addr + count <= va->va_start)
3441 list_for_each_entry_from(va, &vmap_area_list, list) {
3449 vaddr = (char *) vm->addr;
3450 if (addr >= vaddr + get_vm_area_size(vm))
3452 while (addr < vaddr) {
3460 n = vaddr + get_vm_area_size(vm) - addr;
3463 if (!(vm->flags & VM_IOREMAP))
3464 aligned_vread(buf, addr, n);
3465 else /* IOREMAP area is treated as memory hole */
3472 spin_unlock(&vmap_area_lock);
3474 if (buf == buf_start)
3476 /* zero-fill memory holes */
3477 if (buf != buf_start + buflen)
3478 memset(buf, 0, buflen - (buf - buf_start));
3484 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3485 * @vma: vma to cover
3486 * @uaddr: target user address to start at
3487 * @kaddr: virtual address of vmalloc kernel memory
3488 * @pgoff: offset from @kaddr to start at
3489 * @size: size of map area
3491 * Returns: 0 for success, -Exxx on failure
3493 * This function checks that @kaddr is a valid vmalloc'ed area,
3494 * and that it is big enough to cover the range starting at
3495 * @uaddr in @vma. Will return failure if that criteria isn't
3498 * Similar to remap_pfn_range() (see mm/memory.c)
3500 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3501 void *kaddr, unsigned long pgoff,
3504 struct vm_struct *area;
3506 unsigned long end_index;
3508 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3511 size = PAGE_ALIGN(size);
3513 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3516 area = find_vm_area(kaddr);
3520 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3523 if (check_add_overflow(size, off, &end_index) ||
3524 end_index > get_vm_area_size(area))
3529 struct page *page = vmalloc_to_page(kaddr);
3532 ret = vm_insert_page(vma, uaddr, page);
3541 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3547 * remap_vmalloc_range - map vmalloc pages to userspace
3548 * @vma: vma to cover (map full range of vma)
3549 * @addr: vmalloc memory
3550 * @pgoff: number of pages into addr before first page to map
3552 * Returns: 0 for success, -Exxx on failure
3554 * This function checks that addr is a valid vmalloc'ed area, and
3555 * that it is big enough to cover the vma. Will return failure if
3556 * that criteria isn't met.
3558 * Similar to remap_pfn_range() (see mm/memory.c)
3560 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3561 unsigned long pgoff)
3563 return remap_vmalloc_range_partial(vma, vma->vm_start,
3565 vma->vm_end - vma->vm_start);
3567 EXPORT_SYMBOL(remap_vmalloc_range);
3569 void free_vm_area(struct vm_struct *area)
3571 struct vm_struct *ret;
3572 ret = remove_vm_area(area->addr);
3573 BUG_ON(ret != area);
3576 EXPORT_SYMBOL_GPL(free_vm_area);
3579 static struct vmap_area *node_to_va(struct rb_node *n)
3581 return rb_entry_safe(n, struct vmap_area, rb_node);
3585 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3586 * @addr: target address
3588 * Returns: vmap_area if it is found. If there is no such area
3589 * the first highest(reverse order) vmap_area is returned
3590 * i.e. va->va_start < addr && va->va_end < addr or NULL
3591 * if there are no any areas before @addr.
3593 static struct vmap_area *
3594 pvm_find_va_enclose_addr(unsigned long addr)
3596 struct vmap_area *va, *tmp;
3599 n = free_vmap_area_root.rb_node;
3603 tmp = rb_entry(n, struct vmap_area, rb_node);
3604 if (tmp->va_start <= addr) {
3606 if (tmp->va_end >= addr)
3619 * pvm_determine_end_from_reverse - find the highest aligned address
3620 * of free block below VMALLOC_END
3622 * in - the VA we start the search(reverse order);
3623 * out - the VA with the highest aligned end address.
3624 * @align: alignment for required highest address
3626 * Returns: determined end address within vmap_area
3628 static unsigned long
3629 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3631 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3635 list_for_each_entry_from_reverse((*va),
3636 &free_vmap_area_list, list) {
3637 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3638 if ((*va)->va_start < addr)
3647 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3648 * @offsets: array containing offset of each area
3649 * @sizes: array containing size of each area
3650 * @nr_vms: the number of areas to allocate
3651 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3653 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3654 * vm_structs on success, %NULL on failure
3656 * Percpu allocator wants to use congruent vm areas so that it can
3657 * maintain the offsets among percpu areas. This function allocates
3658 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3659 * be scattered pretty far, distance between two areas easily going up
3660 * to gigabytes. To avoid interacting with regular vmallocs, these
3661 * areas are allocated from top.
3663 * Despite its complicated look, this allocator is rather simple. It
3664 * does everything top-down and scans free blocks from the end looking
3665 * for matching base. While scanning, if any of the areas do not fit the
3666 * base address is pulled down to fit the area. Scanning is repeated till
3667 * all the areas fit and then all necessary data structures are inserted
3668 * and the result is returned.
3670 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3671 const size_t *sizes, int nr_vms,
3674 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3675 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3676 struct vmap_area **vas, *va;
3677 struct vm_struct **vms;
3678 int area, area2, last_area, term_area;
3679 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3680 bool purged = false;
3683 /* verify parameters and allocate data structures */
3684 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3685 for (last_area = 0, area = 0; area < nr_vms; area++) {
3686 start = offsets[area];
3687 end = start + sizes[area];
3689 /* is everything aligned properly? */
3690 BUG_ON(!IS_ALIGNED(offsets[area], align));
3691 BUG_ON(!IS_ALIGNED(sizes[area], align));
3693 /* detect the area with the highest address */
3694 if (start > offsets[last_area])
3697 for (area2 = area + 1; area2 < nr_vms; area2++) {
3698 unsigned long start2 = offsets[area2];
3699 unsigned long end2 = start2 + sizes[area2];
3701 BUG_ON(start2 < end && start < end2);
3704 last_end = offsets[last_area] + sizes[last_area];
3706 if (vmalloc_end - vmalloc_start < last_end) {
3711 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3712 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3716 for (area = 0; area < nr_vms; area++) {
3717 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3718 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3719 if (!vas[area] || !vms[area])
3723 spin_lock(&free_vmap_area_lock);
3725 /* start scanning - we scan from the top, begin with the last area */
3726 area = term_area = last_area;
3727 start = offsets[area];
3728 end = start + sizes[area];
3730 va = pvm_find_va_enclose_addr(vmalloc_end);
3731 base = pvm_determine_end_from_reverse(&va, align) - end;
3735 * base might have underflowed, add last_end before
3738 if (base + last_end < vmalloc_start + last_end)
3742 * Fitting base has not been found.
3748 * If required width exceeds current VA block, move
3749 * base downwards and then recheck.
3751 if (base + end > va->va_end) {
3752 base = pvm_determine_end_from_reverse(&va, align) - end;
3758 * If this VA does not fit, move base downwards and recheck.
3760 if (base + start < va->va_start) {
3761 va = node_to_va(rb_prev(&va->rb_node));
3762 base = pvm_determine_end_from_reverse(&va, align) - end;
3768 * This area fits, move on to the previous one. If
3769 * the previous one is the terminal one, we're done.
3771 area = (area + nr_vms - 1) % nr_vms;
3772 if (area == term_area)
3775 start = offsets[area];
3776 end = start + sizes[area];
3777 va = pvm_find_va_enclose_addr(base + end);
3780 /* we've found a fitting base, insert all va's */
3781 for (area = 0; area < nr_vms; area++) {
3784 start = base + offsets[area];
3787 va = pvm_find_va_enclose_addr(start);
3788 if (WARN_ON_ONCE(va == NULL))
3789 /* It is a BUG(), but trigger recovery instead. */
3792 type = classify_va_fit_type(va, start, size);
3793 if (WARN_ON_ONCE(type == NOTHING_FIT))
3794 /* It is a BUG(), but trigger recovery instead. */
3797 ret = adjust_va_to_fit_type(va, start, size, type);
3801 /* Allocated area. */
3803 va->va_start = start;
3804 va->va_end = start + size;
3807 spin_unlock(&free_vmap_area_lock);
3809 /* populate the kasan shadow space */
3810 for (area = 0; area < nr_vms; area++) {
3811 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3812 goto err_free_shadow;
3814 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3818 /* insert all vm's */
3819 spin_lock(&vmap_area_lock);
3820 for (area = 0; area < nr_vms; area++) {
3821 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3823 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3826 spin_unlock(&vmap_area_lock);
3833 * Remove previously allocated areas. There is no
3834 * need in removing these areas from the busy tree,
3835 * because they are inserted only on the final step
3836 * and when pcpu_get_vm_areas() is success.
3839 orig_start = vas[area]->va_start;
3840 orig_end = vas[area]->va_end;
3841 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3842 &free_vmap_area_list);
3844 kasan_release_vmalloc(orig_start, orig_end,
3845 va->va_start, va->va_end);
3850 spin_unlock(&free_vmap_area_lock);
3852 purge_vmap_area_lazy();
3855 /* Before "retry", check if we recover. */
3856 for (area = 0; area < nr_vms; area++) {
3860 vas[area] = kmem_cache_zalloc(
3861 vmap_area_cachep, GFP_KERNEL);
3870 for (area = 0; area < nr_vms; area++) {
3872 kmem_cache_free(vmap_area_cachep, vas[area]);
3882 spin_lock(&free_vmap_area_lock);
3884 * We release all the vmalloc shadows, even the ones for regions that
3885 * hadn't been successfully added. This relies on kasan_release_vmalloc
3886 * being able to tolerate this case.
3888 for (area = 0; area < nr_vms; area++) {
3889 orig_start = vas[area]->va_start;
3890 orig_end = vas[area]->va_end;
3891 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3892 &free_vmap_area_list);
3894 kasan_release_vmalloc(orig_start, orig_end,
3895 va->va_start, va->va_end);
3899 spin_unlock(&free_vmap_area_lock);
3906 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3907 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3908 * @nr_vms: the number of allocated areas
3910 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3912 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3916 for (i = 0; i < nr_vms; i++)
3917 free_vm_area(vms[i]);
3920 #endif /* CONFIG_SMP */
3922 #ifdef CONFIG_PRINTK
3923 bool vmalloc_dump_obj(void *object)
3925 struct vm_struct *vm;
3926 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3928 vm = find_vm_area(objp);
3931 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3932 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3937 #ifdef CONFIG_PROC_FS
3938 static void *s_start(struct seq_file *m, loff_t *pos)
3939 __acquires(&vmap_purge_lock)
3940 __acquires(&vmap_area_lock)
3942 mutex_lock(&vmap_purge_lock);
3943 spin_lock(&vmap_area_lock);
3945 return seq_list_start(&vmap_area_list, *pos);
3948 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3950 return seq_list_next(p, &vmap_area_list, pos);
3953 static void s_stop(struct seq_file *m, void *p)
3954 __releases(&vmap_area_lock)
3955 __releases(&vmap_purge_lock)
3957 spin_unlock(&vmap_area_lock);
3958 mutex_unlock(&vmap_purge_lock);
3961 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3963 if (IS_ENABLED(CONFIG_NUMA)) {
3964 unsigned int nr, *counters = m->private;
3965 unsigned int step = 1U << vm_area_page_order(v);
3970 if (v->flags & VM_UNINITIALIZED)
3972 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3975 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3977 for (nr = 0; nr < v->nr_pages; nr += step)
3978 counters[page_to_nid(v->pages[nr])] += step;
3979 for_each_node_state(nr, N_HIGH_MEMORY)
3981 seq_printf(m, " N%u=%u", nr, counters[nr]);
3985 static void show_purge_info(struct seq_file *m)
3987 struct vmap_area *va;
3989 spin_lock(&purge_vmap_area_lock);
3990 list_for_each_entry(va, &purge_vmap_area_list, list) {
3991 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3992 (void *)va->va_start, (void *)va->va_end,
3993 va->va_end - va->va_start);
3995 spin_unlock(&purge_vmap_area_lock);
3998 static int s_show(struct seq_file *m, void *p)
4000 struct vmap_area *va;
4001 struct vm_struct *v;
4003 va = list_entry(p, struct vmap_area, list);
4006 * s_show can encounter race with remove_vm_area, !vm on behalf
4007 * of vmap area is being tear down or vm_map_ram allocation.
4010 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4011 (void *)va->va_start, (void *)va->va_end,
4012 va->va_end - va->va_start);
4019 seq_printf(m, "0x%pK-0x%pK %7ld",
4020 v->addr, v->addr + v->size, v->size);
4023 seq_printf(m, " %pS", v->caller);
4026 seq_printf(m, " pages=%d", v->nr_pages);
4029 seq_printf(m, " phys=%pa", &v->phys_addr);
4031 if (v->flags & VM_IOREMAP)
4032 seq_puts(m, " ioremap");
4034 if (v->flags & VM_ALLOC)
4035 seq_puts(m, " vmalloc");
4037 if (v->flags & VM_MAP)
4038 seq_puts(m, " vmap");
4040 if (v->flags & VM_USERMAP)
4041 seq_puts(m, " user");
4043 if (v->flags & VM_DMA_COHERENT)
4044 seq_puts(m, " dma-coherent");
4046 if (is_vmalloc_addr(v->pages))
4047 seq_puts(m, " vpages");
4049 show_numa_info(m, v);
4053 * As a final step, dump "unpurged" areas.
4056 if (list_is_last(&va->list, &vmap_area_list))
4062 static const struct seq_operations vmalloc_op = {
4069 static int __init proc_vmalloc_init(void)
4071 if (IS_ENABLED(CONFIG_NUMA))
4072 proc_create_seq_private("vmallocinfo", 0400, NULL,
4074 nr_node_ids * sizeof(unsigned int), NULL);
4076 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4079 module_init(proc_vmalloc_init);