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/llist.h>
35 #include <linux/bitops.h>
36 #include <linux/rbtree_augmented.h>
37 #include <linux/overflow.h>
38 #include <linux/pgtable.h>
39 #include <linux/uaccess.h>
40 #include <linux/hugetlb.h>
41 #include <asm/tlbflush.h>
42 #include <asm/shmparam.h>
45 #include "pgalloc-track.h"
47 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
48 static bool __ro_after_init vmap_allow_huge = true;
50 static int __init set_nohugevmalloc(char *str)
52 vmap_allow_huge = false;
55 early_param("nohugevmalloc", set_nohugevmalloc);
56 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
57 static const bool vmap_allow_huge = false;
58 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
60 bool is_vmalloc_addr(const void *x)
62 unsigned long addr = (unsigned long)x;
64 return addr >= VMALLOC_START && addr < VMALLOC_END;
66 EXPORT_SYMBOL(is_vmalloc_addr);
68 struct vfree_deferred {
69 struct llist_head list;
70 struct work_struct wq;
72 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
74 static void __vunmap(const void *, int);
76 static void free_work(struct work_struct *w)
78 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
79 struct llist_node *t, *llnode;
81 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
82 __vunmap((void *)llnode, 1);
85 /*** Page table manipulation functions ***/
86 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
87 phys_addr_t phys_addr, pgprot_t prot,
88 unsigned int max_page_shift, pgtbl_mod_mask *mask)
92 unsigned long size = PAGE_SIZE;
94 pfn = phys_addr >> PAGE_SHIFT;
95 pte = pte_alloc_kernel_track(pmd, addr, mask);
99 BUG_ON(!pte_none(*pte));
101 #ifdef CONFIG_HUGETLB_PAGE
102 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
103 if (size != PAGE_SIZE) {
104 pte_t entry = pfn_pte(pfn, prot);
106 entry = pte_mkhuge(entry);
107 entry = arch_make_huge_pte(entry, ilog2(size), 0);
108 set_huge_pte_at(&init_mm, addr, pte, entry);
109 pfn += PFN_DOWN(size);
113 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
115 } while (pte += PFN_DOWN(size), addr += size, addr != end);
116 *mask |= PGTBL_PTE_MODIFIED;
120 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
121 phys_addr_t phys_addr, pgprot_t prot,
122 unsigned int max_page_shift)
124 if (max_page_shift < PMD_SHIFT)
127 if (!arch_vmap_pmd_supported(prot))
130 if ((end - addr) != PMD_SIZE)
133 if (!IS_ALIGNED(addr, PMD_SIZE))
136 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
139 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
142 return pmd_set_huge(pmd, phys_addr, prot);
145 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
146 phys_addr_t phys_addr, pgprot_t prot,
147 unsigned int max_page_shift, pgtbl_mod_mask *mask)
152 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
156 next = pmd_addr_end(addr, end);
158 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
160 *mask |= PGTBL_PMD_MODIFIED;
164 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
166 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
170 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
171 phys_addr_t phys_addr, pgprot_t prot,
172 unsigned int max_page_shift)
174 if (max_page_shift < PUD_SHIFT)
177 if (!arch_vmap_pud_supported(prot))
180 if ((end - addr) != PUD_SIZE)
183 if (!IS_ALIGNED(addr, PUD_SIZE))
186 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
189 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
192 return pud_set_huge(pud, phys_addr, prot);
195 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
196 phys_addr_t phys_addr, pgprot_t prot,
197 unsigned int max_page_shift, pgtbl_mod_mask *mask)
202 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
206 next = pud_addr_end(addr, end);
208 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
210 *mask |= PGTBL_PUD_MODIFIED;
214 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
215 max_page_shift, mask))
217 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
221 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
222 phys_addr_t phys_addr, pgprot_t prot,
223 unsigned int max_page_shift)
225 if (max_page_shift < P4D_SHIFT)
228 if (!arch_vmap_p4d_supported(prot))
231 if ((end - addr) != P4D_SIZE)
234 if (!IS_ALIGNED(addr, P4D_SIZE))
237 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
240 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
243 return p4d_set_huge(p4d, phys_addr, prot);
246 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
247 phys_addr_t phys_addr, pgprot_t prot,
248 unsigned int max_page_shift, pgtbl_mod_mask *mask)
253 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
257 next = p4d_addr_end(addr, end);
259 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
261 *mask |= PGTBL_P4D_MODIFIED;
265 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
266 max_page_shift, mask))
268 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
272 static int vmap_range_noflush(unsigned long addr, unsigned long end,
273 phys_addr_t phys_addr, pgprot_t prot,
274 unsigned int max_page_shift)
280 pgtbl_mod_mask mask = 0;
286 pgd = pgd_offset_k(addr);
288 next = pgd_addr_end(addr, end);
289 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
290 max_page_shift, &mask);
293 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
295 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
296 arch_sync_kernel_mappings(start, end);
301 int vmap_range(unsigned long addr, unsigned long end,
302 phys_addr_t phys_addr, pgprot_t prot,
303 unsigned int max_page_shift)
307 err = vmap_range_noflush(addr, end, phys_addr, prot, max_page_shift);
308 flush_cache_vmap(addr, end);
313 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
314 pgtbl_mod_mask *mask)
318 pte = pte_offset_kernel(pmd, addr);
320 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
321 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
322 } while (pte++, addr += PAGE_SIZE, addr != end);
323 *mask |= PGTBL_PTE_MODIFIED;
326 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
327 pgtbl_mod_mask *mask)
333 pmd = pmd_offset(pud, addr);
335 next = pmd_addr_end(addr, end);
337 cleared = pmd_clear_huge(pmd);
338 if (cleared || pmd_bad(*pmd))
339 *mask |= PGTBL_PMD_MODIFIED;
343 if (pmd_none_or_clear_bad(pmd))
345 vunmap_pte_range(pmd, addr, next, mask);
348 } while (pmd++, addr = next, addr != end);
351 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
352 pgtbl_mod_mask *mask)
358 pud = pud_offset(p4d, addr);
360 next = pud_addr_end(addr, end);
362 cleared = pud_clear_huge(pud);
363 if (cleared || pud_bad(*pud))
364 *mask |= PGTBL_PUD_MODIFIED;
368 if (pud_none_or_clear_bad(pud))
370 vunmap_pmd_range(pud, addr, next, mask);
371 } while (pud++, addr = next, addr != end);
374 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
375 pgtbl_mod_mask *mask)
381 p4d = p4d_offset(pgd, addr);
383 next = p4d_addr_end(addr, end);
385 cleared = p4d_clear_huge(p4d);
386 if (cleared || p4d_bad(*p4d))
387 *mask |= PGTBL_P4D_MODIFIED;
391 if (p4d_none_or_clear_bad(p4d))
393 vunmap_pud_range(p4d, addr, next, mask);
394 } while (p4d++, addr = next, addr != end);
398 * vunmap_range_noflush is similar to vunmap_range, but does not
399 * flush caches or TLBs.
401 * The caller is responsible for calling flush_cache_vmap() before calling
402 * this function, and flush_tlb_kernel_range after it has returned
403 * successfully (and before the addresses are expected to cause a page fault
404 * or be re-mapped for something else, if TLB flushes are being delayed or
407 * This is an internal function only. Do not use outside mm/.
409 void vunmap_range_noflush(unsigned long start, unsigned long end)
413 unsigned long addr = start;
414 pgtbl_mod_mask mask = 0;
417 pgd = pgd_offset_k(addr);
419 next = pgd_addr_end(addr, end);
421 mask |= PGTBL_PGD_MODIFIED;
422 if (pgd_none_or_clear_bad(pgd))
424 vunmap_p4d_range(pgd, addr, next, &mask);
425 } while (pgd++, addr = next, addr != end);
427 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
428 arch_sync_kernel_mappings(start, end);
432 * vunmap_range - unmap kernel virtual addresses
433 * @addr: start of the VM area to unmap
434 * @end: end of the VM area to unmap (non-inclusive)
436 * Clears any present PTEs in the virtual address range, flushes TLBs and
437 * caches. Any subsequent access to the address before it has been re-mapped
440 void vunmap_range(unsigned long addr, unsigned long end)
442 flush_cache_vunmap(addr, end);
443 vunmap_range_noflush(addr, end);
444 flush_tlb_kernel_range(addr, end);
447 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
448 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
449 pgtbl_mod_mask *mask)
454 * nr is a running index into the array which helps higher level
455 * callers keep track of where we're up to.
458 pte = pte_alloc_kernel_track(pmd, addr, mask);
462 struct page *page = pages[*nr];
464 if (WARN_ON(!pte_none(*pte)))
468 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
470 } while (pte++, addr += PAGE_SIZE, addr != end);
471 *mask |= PGTBL_PTE_MODIFIED;
475 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
476 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
477 pgtbl_mod_mask *mask)
482 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
486 next = pmd_addr_end(addr, end);
487 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
489 } while (pmd++, addr = next, addr != end);
493 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
494 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
495 pgtbl_mod_mask *mask)
500 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
504 next = pud_addr_end(addr, end);
505 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
507 } while (pud++, addr = next, addr != end);
511 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
512 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
513 pgtbl_mod_mask *mask)
518 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
522 next = p4d_addr_end(addr, end);
523 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
525 } while (p4d++, addr = next, addr != end);
529 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
530 pgprot_t prot, struct page **pages)
532 unsigned long start = addr;
537 pgtbl_mod_mask mask = 0;
540 pgd = pgd_offset_k(addr);
542 next = pgd_addr_end(addr, end);
544 mask |= PGTBL_PGD_MODIFIED;
545 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
548 } while (pgd++, addr = next, addr != end);
550 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
551 arch_sync_kernel_mappings(start, end);
557 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
560 * The caller is responsible for calling flush_cache_vmap() after this
561 * function returns successfully and before the addresses are accessed.
563 * This is an internal function only. Do not use outside mm/.
565 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
566 pgprot_t prot, struct page **pages, unsigned int page_shift)
568 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
570 WARN_ON(page_shift < PAGE_SHIFT);
572 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
573 page_shift == PAGE_SHIFT)
574 return vmap_small_pages_range_noflush(addr, end, prot, pages);
576 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
579 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
580 __pa(page_address(pages[i])), prot,
585 addr += 1UL << page_shift;
592 * vmap_pages_range - map pages to a kernel virtual address
593 * @addr: start of the VM area to map
594 * @end: end of the VM area to map (non-inclusive)
595 * @prot: page protection flags to use
596 * @pages: pages to map (always PAGE_SIZE pages)
597 * @page_shift: maximum shift that the pages may be mapped with, @pages must
598 * be aligned and contiguous up to at least this shift.
601 * 0 on success, -errno on failure.
603 static int vmap_pages_range(unsigned long addr, unsigned long end,
604 pgprot_t prot, struct page **pages, unsigned int page_shift)
608 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
609 flush_cache_vmap(addr, end);
613 int is_vmalloc_or_module_addr(const void *x)
616 * ARM, x86-64 and sparc64 put modules in a special place,
617 * and fall back on vmalloc() if that fails. Others
618 * just put it in the vmalloc space.
620 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
621 unsigned long addr = (unsigned long)x;
622 if (addr >= MODULES_VADDR && addr < MODULES_END)
625 return is_vmalloc_addr(x);
629 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
630 * return the tail page that corresponds to the base page address, which
631 * matches small vmap mappings.
633 struct page *vmalloc_to_page(const void *vmalloc_addr)
635 unsigned long addr = (unsigned long) vmalloc_addr;
636 struct page *page = NULL;
637 pgd_t *pgd = pgd_offset_k(addr);
644 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
645 * architectures that do not vmalloc module space
647 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
651 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
652 return NULL; /* XXX: no allowance for huge pgd */
653 if (WARN_ON_ONCE(pgd_bad(*pgd)))
656 p4d = p4d_offset(pgd, addr);
660 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
661 if (WARN_ON_ONCE(p4d_bad(*p4d)))
664 pud = pud_offset(p4d, addr);
668 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
669 if (WARN_ON_ONCE(pud_bad(*pud)))
672 pmd = pmd_offset(pud, addr);
676 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
677 if (WARN_ON_ONCE(pmd_bad(*pmd)))
680 ptep = pte_offset_map(pmd, addr);
682 if (pte_present(pte))
683 page = pte_page(pte);
688 EXPORT_SYMBOL(vmalloc_to_page);
691 * Map a vmalloc()-space virtual address to the physical page frame number.
693 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
695 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
697 EXPORT_SYMBOL(vmalloc_to_pfn);
700 /*** Global kva allocator ***/
702 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
703 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
706 static DEFINE_SPINLOCK(vmap_area_lock);
707 static DEFINE_SPINLOCK(free_vmap_area_lock);
708 /* Export for kexec only */
709 LIST_HEAD(vmap_area_list);
710 static struct rb_root vmap_area_root = RB_ROOT;
711 static bool vmap_initialized __read_mostly;
713 static struct rb_root purge_vmap_area_root = RB_ROOT;
714 static LIST_HEAD(purge_vmap_area_list);
715 static DEFINE_SPINLOCK(purge_vmap_area_lock);
718 * This kmem_cache is used for vmap_area objects. Instead of
719 * allocating from slab we reuse an object from this cache to
720 * make things faster. Especially in "no edge" splitting of
723 static struct kmem_cache *vmap_area_cachep;
726 * This linked list is used in pair with free_vmap_area_root.
727 * It gives O(1) access to prev/next to perform fast coalescing.
729 static LIST_HEAD(free_vmap_area_list);
732 * This augment red-black tree represents the free vmap space.
733 * All vmap_area objects in this tree are sorted by va->va_start
734 * address. It is used for allocation and merging when a vmap
735 * object is released.
737 * Each vmap_area node contains a maximum available free block
738 * of its sub-tree, right or left. Therefore it is possible to
739 * find a lowest match of free area.
741 static struct rb_root free_vmap_area_root = RB_ROOT;
744 * Preload a CPU with one object for "no edge" split case. The
745 * aim is to get rid of allocations from the atomic context, thus
746 * to use more permissive allocation masks.
748 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
750 static __always_inline unsigned long
751 va_size(struct vmap_area *va)
753 return (va->va_end - va->va_start);
756 static __always_inline unsigned long
757 get_subtree_max_size(struct rb_node *node)
759 struct vmap_area *va;
761 va = rb_entry_safe(node, struct vmap_area, rb_node);
762 return va ? va->subtree_max_size : 0;
766 * Gets called when remove the node and rotate.
768 static __always_inline unsigned long
769 compute_subtree_max_size(struct vmap_area *va)
771 return max3(va_size(va),
772 get_subtree_max_size(va->rb_node.rb_left),
773 get_subtree_max_size(va->rb_node.rb_right));
776 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
777 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
779 static void purge_vmap_area_lazy(void);
780 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
781 static unsigned long lazy_max_pages(void);
783 static atomic_long_t nr_vmalloc_pages;
785 unsigned long vmalloc_nr_pages(void)
787 return atomic_long_read(&nr_vmalloc_pages);
790 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
792 struct vmap_area *va = NULL;
793 struct rb_node *n = vmap_area_root.rb_node;
796 struct vmap_area *tmp;
798 tmp = rb_entry(n, struct vmap_area, rb_node);
799 if (tmp->va_end > addr) {
801 if (tmp->va_start <= addr)
812 static struct vmap_area *__find_vmap_area(unsigned long addr)
814 struct rb_node *n = vmap_area_root.rb_node;
817 struct vmap_area *va;
819 va = rb_entry(n, struct vmap_area, rb_node);
820 if (addr < va->va_start)
822 else if (addr >= va->va_end)
832 * This function returns back addresses of parent node
833 * and its left or right link for further processing.
835 * Otherwise NULL is returned. In that case all further
836 * steps regarding inserting of conflicting overlap range
837 * have to be declined and actually considered as a bug.
839 static __always_inline struct rb_node **
840 find_va_links(struct vmap_area *va,
841 struct rb_root *root, struct rb_node *from,
842 struct rb_node **parent)
844 struct vmap_area *tmp_va;
845 struct rb_node **link;
848 link = &root->rb_node;
849 if (unlikely(!*link)) {
858 * Go to the bottom of the tree. When we hit the last point
859 * we end up with parent rb_node and correct direction, i name
860 * it link, where the new va->rb_node will be attached to.
863 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
866 * During the traversal we also do some sanity check.
867 * Trigger the BUG() if there are sides(left/right)
870 if (va->va_start < tmp_va->va_end &&
871 va->va_end <= tmp_va->va_start)
872 link = &(*link)->rb_left;
873 else if (va->va_end > tmp_va->va_start &&
874 va->va_start >= tmp_va->va_end)
875 link = &(*link)->rb_right;
877 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
878 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
884 *parent = &tmp_va->rb_node;
888 static __always_inline struct list_head *
889 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
891 struct list_head *list;
893 if (unlikely(!parent))
895 * The red-black tree where we try to find VA neighbors
896 * before merging or inserting is empty, i.e. it means
897 * there is no free vmap space. Normally it does not
898 * happen but we handle this case anyway.
902 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
903 return (&parent->rb_right == link ? list->next : list);
906 static __always_inline void
907 link_va(struct vmap_area *va, struct rb_root *root,
908 struct rb_node *parent, struct rb_node **link, struct list_head *head)
911 * VA is still not in the list, but we can
912 * identify its future previous list_head node.
914 if (likely(parent)) {
915 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
916 if (&parent->rb_right != link)
920 /* Insert to the rb-tree */
921 rb_link_node(&va->rb_node, parent, link);
922 if (root == &free_vmap_area_root) {
924 * Some explanation here. Just perform simple insertion
925 * to the tree. We do not set va->subtree_max_size to
926 * its current size before calling rb_insert_augmented().
927 * It is because of we populate the tree from the bottom
928 * to parent levels when the node _is_ in the tree.
930 * Therefore we set subtree_max_size to zero after insertion,
931 * to let __augment_tree_propagate_from() puts everything to
932 * the correct order later on.
934 rb_insert_augmented(&va->rb_node,
935 root, &free_vmap_area_rb_augment_cb);
936 va->subtree_max_size = 0;
938 rb_insert_color(&va->rb_node, root);
941 /* Address-sort this list */
942 list_add(&va->list, head);
945 static __always_inline void
946 unlink_va(struct vmap_area *va, struct rb_root *root)
948 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
951 if (root == &free_vmap_area_root)
952 rb_erase_augmented(&va->rb_node,
953 root, &free_vmap_area_rb_augment_cb);
955 rb_erase(&va->rb_node, root);
958 RB_CLEAR_NODE(&va->rb_node);
961 #if DEBUG_AUGMENT_PROPAGATE_CHECK
963 augment_tree_propagate_check(void)
965 struct vmap_area *va;
966 unsigned long computed_size;
968 list_for_each_entry(va, &free_vmap_area_list, list) {
969 computed_size = compute_subtree_max_size(va);
970 if (computed_size != va->subtree_max_size)
971 pr_emerg("tree is corrupted: %lu, %lu\n",
972 va_size(va), va->subtree_max_size);
978 * This function populates subtree_max_size from bottom to upper
979 * levels starting from VA point. The propagation must be done
980 * when VA size is modified by changing its va_start/va_end. Or
981 * in case of newly inserting of VA to the tree.
983 * It means that __augment_tree_propagate_from() must be called:
984 * - After VA has been inserted to the tree(free path);
985 * - After VA has been shrunk(allocation path);
986 * - After VA has been increased(merging path).
988 * Please note that, it does not mean that upper parent nodes
989 * and their subtree_max_size are recalculated all the time up
998 * For example if we modify the node 4, shrinking it to 2, then
999 * no any modification is required. If we shrink the node 2 to 1
1000 * its subtree_max_size is updated only, and set to 1. If we shrink
1001 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1002 * node becomes 4--6.
1004 static __always_inline void
1005 augment_tree_propagate_from(struct vmap_area *va)
1008 * Populate the tree from bottom towards the root until
1009 * the calculated maximum available size of checked node
1010 * is equal to its current one.
1012 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1014 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1015 augment_tree_propagate_check();
1020 insert_vmap_area(struct vmap_area *va,
1021 struct rb_root *root, struct list_head *head)
1023 struct rb_node **link;
1024 struct rb_node *parent;
1026 link = find_va_links(va, root, NULL, &parent);
1028 link_va(va, root, parent, link, head);
1032 insert_vmap_area_augment(struct vmap_area *va,
1033 struct rb_node *from, struct rb_root *root,
1034 struct list_head *head)
1036 struct rb_node **link;
1037 struct rb_node *parent;
1040 link = find_va_links(va, NULL, from, &parent);
1042 link = find_va_links(va, root, NULL, &parent);
1045 link_va(va, root, parent, link, head);
1046 augment_tree_propagate_from(va);
1051 * Merge de-allocated chunk of VA memory with previous
1052 * and next free blocks. If coalesce is not done a new
1053 * free area is inserted. If VA has been merged, it is
1056 * Please note, it can return NULL in case of overlap
1057 * ranges, followed by WARN() report. Despite it is a
1058 * buggy behaviour, a system can be alive and keep
1061 static __always_inline struct vmap_area *
1062 merge_or_add_vmap_area(struct vmap_area *va,
1063 struct rb_root *root, struct list_head *head)
1065 struct vmap_area *sibling;
1066 struct list_head *next;
1067 struct rb_node **link;
1068 struct rb_node *parent;
1069 bool merged = false;
1072 * Find a place in the tree where VA potentially will be
1073 * inserted, unless it is merged with its sibling/siblings.
1075 link = find_va_links(va, root, NULL, &parent);
1080 * Get next node of VA to check if merging can be done.
1082 next = get_va_next_sibling(parent, link);
1083 if (unlikely(next == NULL))
1089 * |<------VA------>|<-----Next----->|
1094 sibling = list_entry(next, struct vmap_area, list);
1095 if (sibling->va_start == va->va_end) {
1096 sibling->va_start = va->va_start;
1098 /* Free vmap_area object. */
1099 kmem_cache_free(vmap_area_cachep, va);
1101 /* Point to the new merged area. */
1110 * |<-----Prev----->|<------VA------>|
1114 if (next->prev != head) {
1115 sibling = list_entry(next->prev, struct vmap_area, list);
1116 if (sibling->va_end == va->va_start) {
1118 * If both neighbors are coalesced, it is important
1119 * to unlink the "next" node first, followed by merging
1120 * with "previous" one. Otherwise the tree might not be
1121 * fully populated if a sibling's augmented value is
1122 * "normalized" because of rotation operations.
1125 unlink_va(va, root);
1127 sibling->va_end = va->va_end;
1129 /* Free vmap_area object. */
1130 kmem_cache_free(vmap_area_cachep, va);
1132 /* Point to the new merged area. */
1140 link_va(va, root, parent, link, head);
1145 static __always_inline struct vmap_area *
1146 merge_or_add_vmap_area_augment(struct vmap_area *va,
1147 struct rb_root *root, struct list_head *head)
1149 va = merge_or_add_vmap_area(va, root, head);
1151 augment_tree_propagate_from(va);
1156 static __always_inline bool
1157 is_within_this_va(struct vmap_area *va, unsigned long size,
1158 unsigned long align, unsigned long vstart)
1160 unsigned long nva_start_addr;
1162 if (va->va_start > vstart)
1163 nva_start_addr = ALIGN(va->va_start, align);
1165 nva_start_addr = ALIGN(vstart, align);
1167 /* Can be overflowed due to big size or alignment. */
1168 if (nva_start_addr + size < nva_start_addr ||
1169 nva_start_addr < vstart)
1172 return (nva_start_addr + size <= va->va_end);
1176 * Find the first free block(lowest start address) in the tree,
1177 * that will accomplish the request corresponding to passing
1180 static __always_inline struct vmap_area *
1181 find_vmap_lowest_match(unsigned long size,
1182 unsigned long align, unsigned long vstart)
1184 struct vmap_area *va;
1185 struct rb_node *node;
1186 unsigned long length;
1188 /* Start from the root. */
1189 node = free_vmap_area_root.rb_node;
1191 /* Adjust the search size for alignment overhead. */
1192 length = size + align - 1;
1195 va = rb_entry(node, struct vmap_area, rb_node);
1197 if (get_subtree_max_size(node->rb_left) >= length &&
1198 vstart < va->va_start) {
1199 node = node->rb_left;
1201 if (is_within_this_va(va, size, align, vstart))
1205 * Does not make sense to go deeper towards the right
1206 * sub-tree if it does not have a free block that is
1207 * equal or bigger to the requested search length.
1209 if (get_subtree_max_size(node->rb_right) >= length) {
1210 node = node->rb_right;
1215 * OK. We roll back and find the first right sub-tree,
1216 * that will satisfy the search criteria. It can happen
1217 * only once due to "vstart" restriction.
1219 while ((node = rb_parent(node))) {
1220 va = rb_entry(node, struct vmap_area, rb_node);
1221 if (is_within_this_va(va, size, align, vstart))
1224 if (get_subtree_max_size(node->rb_right) >= length &&
1225 vstart <= va->va_start) {
1226 node = node->rb_right;
1236 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1237 #include <linux/random.h>
1239 static struct vmap_area *
1240 find_vmap_lowest_linear_match(unsigned long size,
1241 unsigned long align, unsigned long vstart)
1243 struct vmap_area *va;
1245 list_for_each_entry(va, &free_vmap_area_list, list) {
1246 if (!is_within_this_va(va, size, align, vstart))
1256 find_vmap_lowest_match_check(unsigned long size)
1258 struct vmap_area *va_1, *va_2;
1259 unsigned long vstart;
1262 get_random_bytes(&rnd, sizeof(rnd));
1263 vstart = VMALLOC_START + rnd;
1265 va_1 = find_vmap_lowest_match(size, 1, vstart);
1266 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
1269 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1270 va_1, va_2, vstart);
1276 FL_FIT_TYPE = 1, /* full fit */
1277 LE_FIT_TYPE = 2, /* left edge fit */
1278 RE_FIT_TYPE = 3, /* right edge fit */
1279 NE_FIT_TYPE = 4 /* no edge fit */
1282 static __always_inline enum fit_type
1283 classify_va_fit_type(struct vmap_area *va,
1284 unsigned long nva_start_addr, unsigned long size)
1288 /* Check if it is within VA. */
1289 if (nva_start_addr < va->va_start ||
1290 nva_start_addr + size > va->va_end)
1294 if (va->va_start == nva_start_addr) {
1295 if (va->va_end == nva_start_addr + size)
1299 } else if (va->va_end == nva_start_addr + size) {
1308 static __always_inline int
1309 adjust_va_to_fit_type(struct vmap_area *va,
1310 unsigned long nva_start_addr, unsigned long size,
1313 struct vmap_area *lva = NULL;
1315 if (type == FL_FIT_TYPE) {
1317 * No need to split VA, it fully fits.
1323 unlink_va(va, &free_vmap_area_root);
1324 kmem_cache_free(vmap_area_cachep, va);
1325 } else if (type == LE_FIT_TYPE) {
1327 * Split left edge of fit VA.
1333 va->va_start += size;
1334 } else if (type == RE_FIT_TYPE) {
1336 * Split right edge of fit VA.
1342 va->va_end = nva_start_addr;
1343 } else if (type == NE_FIT_TYPE) {
1345 * Split no edge of fit VA.
1351 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1352 if (unlikely(!lva)) {
1354 * For percpu allocator we do not do any pre-allocation
1355 * and leave it as it is. The reason is it most likely
1356 * never ends up with NE_FIT_TYPE splitting. In case of
1357 * percpu allocations offsets and sizes are aligned to
1358 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1359 * are its main fitting cases.
1361 * There are a few exceptions though, as an example it is
1362 * a first allocation (early boot up) when we have "one"
1363 * big free space that has to be split.
1365 * Also we can hit this path in case of regular "vmap"
1366 * allocations, if "this" current CPU was not preloaded.
1367 * See the comment in alloc_vmap_area() why. If so, then
1368 * GFP_NOWAIT is used instead to get an extra object for
1369 * split purpose. That is rare and most time does not
1372 * What happens if an allocation gets failed. Basically,
1373 * an "overflow" path is triggered to purge lazily freed
1374 * areas to free some memory, then, the "retry" path is
1375 * triggered to repeat one more time. See more details
1376 * in alloc_vmap_area() function.
1378 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1384 * Build the remainder.
1386 lva->va_start = va->va_start;
1387 lva->va_end = nva_start_addr;
1390 * Shrink this VA to remaining size.
1392 va->va_start = nva_start_addr + size;
1397 if (type != FL_FIT_TYPE) {
1398 augment_tree_propagate_from(va);
1400 if (lva) /* type == NE_FIT_TYPE */
1401 insert_vmap_area_augment(lva, &va->rb_node,
1402 &free_vmap_area_root, &free_vmap_area_list);
1409 * Returns a start address of the newly allocated area, if success.
1410 * Otherwise a vend is returned that indicates failure.
1412 static __always_inline unsigned long
1413 __alloc_vmap_area(unsigned long size, unsigned long align,
1414 unsigned long vstart, unsigned long vend)
1416 unsigned long nva_start_addr;
1417 struct vmap_area *va;
1421 va = find_vmap_lowest_match(size, align, vstart);
1425 if (va->va_start > vstart)
1426 nva_start_addr = ALIGN(va->va_start, align);
1428 nva_start_addr = ALIGN(vstart, align);
1430 /* Check the "vend" restriction. */
1431 if (nva_start_addr + size > vend)
1434 /* Classify what we have found. */
1435 type = classify_va_fit_type(va, nva_start_addr, size);
1436 if (WARN_ON_ONCE(type == NOTHING_FIT))
1439 /* Update the free vmap_area. */
1440 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1444 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1445 find_vmap_lowest_match_check(size);
1448 return nva_start_addr;
1452 * Free a region of KVA allocated by alloc_vmap_area
1454 static void free_vmap_area(struct vmap_area *va)
1457 * Remove from the busy tree/list.
1459 spin_lock(&vmap_area_lock);
1460 unlink_va(va, &vmap_area_root);
1461 spin_unlock(&vmap_area_lock);
1464 * Insert/Merge it back to the free tree/list.
1466 spin_lock(&free_vmap_area_lock);
1467 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1468 spin_unlock(&free_vmap_area_lock);
1472 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1474 struct vmap_area *va = NULL;
1477 * Preload this CPU with one extra vmap_area object. It is used
1478 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1479 * a CPU that does an allocation is preloaded.
1481 * We do it in non-atomic context, thus it allows us to use more
1482 * permissive allocation masks to be more stable under low memory
1483 * condition and high memory pressure.
1485 if (!this_cpu_read(ne_fit_preload_node))
1486 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1490 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1491 kmem_cache_free(vmap_area_cachep, va);
1495 * Allocate a region of KVA of the specified size and alignment, within the
1498 static struct vmap_area *alloc_vmap_area(unsigned long size,
1499 unsigned long align,
1500 unsigned long vstart, unsigned long vend,
1501 int node, gfp_t gfp_mask)
1503 struct vmap_area *va;
1504 unsigned long freed;
1510 BUG_ON(offset_in_page(size));
1511 BUG_ON(!is_power_of_2(align));
1513 if (unlikely(!vmap_initialized))
1514 return ERR_PTR(-EBUSY);
1517 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1519 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1521 return ERR_PTR(-ENOMEM);
1524 * Only scan the relevant parts containing pointers to other objects
1525 * to avoid false negatives.
1527 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1530 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1531 addr = __alloc_vmap_area(size, align, vstart, vend);
1532 spin_unlock(&free_vmap_area_lock);
1535 * If an allocation fails, the "vend" address is
1536 * returned. Therefore trigger the overflow path.
1538 if (unlikely(addr == vend))
1541 va->va_start = addr;
1542 va->va_end = addr + size;
1545 spin_lock(&vmap_area_lock);
1546 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1547 spin_unlock(&vmap_area_lock);
1549 BUG_ON(!IS_ALIGNED(va->va_start, align));
1550 BUG_ON(va->va_start < vstart);
1551 BUG_ON(va->va_end > vend);
1553 ret = kasan_populate_vmalloc(addr, size);
1556 return ERR_PTR(ret);
1563 purge_vmap_area_lazy();
1569 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1576 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1577 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1580 kmem_cache_free(vmap_area_cachep, va);
1581 return ERR_PTR(-EBUSY);
1584 int register_vmap_purge_notifier(struct notifier_block *nb)
1586 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1588 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1590 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1592 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1594 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1597 * lazy_max_pages is the maximum amount of virtual address space we gather up
1598 * before attempting to purge with a TLB flush.
1600 * There is a tradeoff here: a larger number will cover more kernel page tables
1601 * and take slightly longer to purge, but it will linearly reduce the number of
1602 * global TLB flushes that must be performed. It would seem natural to scale
1603 * this number up linearly with the number of CPUs (because vmapping activity
1604 * could also scale linearly with the number of CPUs), however it is likely
1605 * that in practice, workloads might be constrained in other ways that mean
1606 * vmap activity will not scale linearly with CPUs. Also, I want to be
1607 * conservative and not introduce a big latency on huge systems, so go with
1608 * a less aggressive log scale. It will still be an improvement over the old
1609 * code, and it will be simple to change the scale factor if we find that it
1610 * becomes a problem on bigger systems.
1612 static unsigned long lazy_max_pages(void)
1616 log = fls(num_online_cpus());
1618 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1621 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1624 * Serialize vmap purging. There is no actual critical section protected
1625 * by this look, but we want to avoid concurrent calls for performance
1626 * reasons and to make the pcpu_get_vm_areas more deterministic.
1628 static DEFINE_MUTEX(vmap_purge_lock);
1630 /* for per-CPU blocks */
1631 static void purge_fragmented_blocks_allcpus(void);
1633 #ifdef CONFIG_X86_64
1635 * called before a call to iounmap() if the caller wants vm_area_struct's
1636 * immediately freed.
1638 void set_iounmap_nonlazy(void)
1640 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1642 #endif /* CONFIG_X86_64 */
1645 * Purges all lazily-freed vmap areas.
1647 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1649 unsigned long resched_threshold;
1650 struct list_head local_pure_list;
1651 struct vmap_area *va, *n_va;
1653 lockdep_assert_held(&vmap_purge_lock);
1655 spin_lock(&purge_vmap_area_lock);
1656 purge_vmap_area_root = RB_ROOT;
1657 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1658 spin_unlock(&purge_vmap_area_lock);
1660 if (unlikely(list_empty(&local_pure_list)))
1664 list_first_entry(&local_pure_list,
1665 struct vmap_area, list)->va_start);
1668 list_last_entry(&local_pure_list,
1669 struct vmap_area, list)->va_end);
1671 flush_tlb_kernel_range(start, end);
1672 resched_threshold = lazy_max_pages() << 1;
1674 spin_lock(&free_vmap_area_lock);
1675 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1676 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1677 unsigned long orig_start = va->va_start;
1678 unsigned long orig_end = va->va_end;
1681 * Finally insert or merge lazily-freed area. It is
1682 * detached and there is no need to "unlink" it from
1685 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1686 &free_vmap_area_list);
1691 if (is_vmalloc_or_module_addr((void *)orig_start))
1692 kasan_release_vmalloc(orig_start, orig_end,
1693 va->va_start, va->va_end);
1695 atomic_long_sub(nr, &vmap_lazy_nr);
1697 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1698 cond_resched_lock(&free_vmap_area_lock);
1700 spin_unlock(&free_vmap_area_lock);
1705 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1706 * is already purging.
1708 static void try_purge_vmap_area_lazy(void)
1710 if (mutex_trylock(&vmap_purge_lock)) {
1711 __purge_vmap_area_lazy(ULONG_MAX, 0);
1712 mutex_unlock(&vmap_purge_lock);
1717 * Kick off a purge of the outstanding lazy areas.
1719 static void purge_vmap_area_lazy(void)
1721 mutex_lock(&vmap_purge_lock);
1722 purge_fragmented_blocks_allcpus();
1723 __purge_vmap_area_lazy(ULONG_MAX, 0);
1724 mutex_unlock(&vmap_purge_lock);
1728 * Free a vmap area, caller ensuring that the area has been unmapped
1729 * and flush_cache_vunmap had been called for the correct range
1732 static void free_vmap_area_noflush(struct vmap_area *va)
1734 unsigned long nr_lazy;
1736 spin_lock(&vmap_area_lock);
1737 unlink_va(va, &vmap_area_root);
1738 spin_unlock(&vmap_area_lock);
1740 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1741 PAGE_SHIFT, &vmap_lazy_nr);
1744 * Merge or place it to the purge tree/list.
1746 spin_lock(&purge_vmap_area_lock);
1747 merge_or_add_vmap_area(va,
1748 &purge_vmap_area_root, &purge_vmap_area_list);
1749 spin_unlock(&purge_vmap_area_lock);
1751 /* After this point, we may free va at any time */
1752 if (unlikely(nr_lazy > lazy_max_pages()))
1753 try_purge_vmap_area_lazy();
1757 * Free and unmap a vmap area
1759 static void free_unmap_vmap_area(struct vmap_area *va)
1761 flush_cache_vunmap(va->va_start, va->va_end);
1762 vunmap_range_noflush(va->va_start, va->va_end);
1763 if (debug_pagealloc_enabled_static())
1764 flush_tlb_kernel_range(va->va_start, va->va_end);
1766 free_vmap_area_noflush(va);
1769 static struct vmap_area *find_vmap_area(unsigned long addr)
1771 struct vmap_area *va;
1773 spin_lock(&vmap_area_lock);
1774 va = __find_vmap_area(addr);
1775 spin_unlock(&vmap_area_lock);
1780 /*** Per cpu kva allocator ***/
1783 * vmap space is limited especially on 32 bit architectures. Ensure there is
1784 * room for at least 16 percpu vmap blocks per CPU.
1787 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1788 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1789 * instead (we just need a rough idea)
1791 #if BITS_PER_LONG == 32
1792 #define VMALLOC_SPACE (128UL*1024*1024)
1794 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1797 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1798 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1799 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1800 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1801 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1802 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1803 #define VMAP_BBMAP_BITS \
1804 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1805 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1806 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1808 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1810 struct vmap_block_queue {
1812 struct list_head free;
1817 struct vmap_area *va;
1818 unsigned long free, dirty;
1819 unsigned long dirty_min, dirty_max; /*< dirty range */
1820 struct list_head free_list;
1821 struct rcu_head rcu_head;
1822 struct list_head purge;
1825 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1826 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1829 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1830 * in the free path. Could get rid of this if we change the API to return a
1831 * "cookie" from alloc, to be passed to free. But no big deal yet.
1833 static DEFINE_XARRAY(vmap_blocks);
1836 * We should probably have a fallback mechanism to allocate virtual memory
1837 * out of partially filled vmap blocks. However vmap block sizing should be
1838 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1842 static unsigned long addr_to_vb_idx(unsigned long addr)
1844 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1845 addr /= VMAP_BLOCK_SIZE;
1849 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1853 addr = va_start + (pages_off << PAGE_SHIFT);
1854 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1855 return (void *)addr;
1859 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1860 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1861 * @order: how many 2^order pages should be occupied in newly allocated block
1862 * @gfp_mask: flags for the page level allocator
1864 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1866 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1868 struct vmap_block_queue *vbq;
1869 struct vmap_block *vb;
1870 struct vmap_area *va;
1871 unsigned long vb_idx;
1875 node = numa_node_id();
1877 vb = kmalloc_node(sizeof(struct vmap_block),
1878 gfp_mask & GFP_RECLAIM_MASK, node);
1880 return ERR_PTR(-ENOMEM);
1882 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1883 VMALLOC_START, VMALLOC_END,
1887 return ERR_CAST(va);
1890 vaddr = vmap_block_vaddr(va->va_start, 0);
1891 spin_lock_init(&vb->lock);
1893 /* At least something should be left free */
1894 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1895 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1897 vb->dirty_min = VMAP_BBMAP_BITS;
1899 INIT_LIST_HEAD(&vb->free_list);
1901 vb_idx = addr_to_vb_idx(va->va_start);
1902 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1906 return ERR_PTR(err);
1909 vbq = &get_cpu_var(vmap_block_queue);
1910 spin_lock(&vbq->lock);
1911 list_add_tail_rcu(&vb->free_list, &vbq->free);
1912 spin_unlock(&vbq->lock);
1913 put_cpu_var(vmap_block_queue);
1918 static void free_vmap_block(struct vmap_block *vb)
1920 struct vmap_block *tmp;
1922 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1925 free_vmap_area_noflush(vb->va);
1926 kfree_rcu(vb, rcu_head);
1929 static void purge_fragmented_blocks(int cpu)
1932 struct vmap_block *vb;
1933 struct vmap_block *n_vb;
1934 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1937 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1939 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1942 spin_lock(&vb->lock);
1943 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1944 vb->free = 0; /* prevent further allocs after releasing lock */
1945 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1947 vb->dirty_max = VMAP_BBMAP_BITS;
1948 spin_lock(&vbq->lock);
1949 list_del_rcu(&vb->free_list);
1950 spin_unlock(&vbq->lock);
1951 spin_unlock(&vb->lock);
1952 list_add_tail(&vb->purge, &purge);
1954 spin_unlock(&vb->lock);
1958 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1959 list_del(&vb->purge);
1960 free_vmap_block(vb);
1964 static void purge_fragmented_blocks_allcpus(void)
1968 for_each_possible_cpu(cpu)
1969 purge_fragmented_blocks(cpu);
1972 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1974 struct vmap_block_queue *vbq;
1975 struct vmap_block *vb;
1979 BUG_ON(offset_in_page(size));
1980 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1981 if (WARN_ON(size == 0)) {
1983 * Allocating 0 bytes isn't what caller wants since
1984 * get_order(0) returns funny result. Just warn and terminate
1989 order = get_order(size);
1992 vbq = &get_cpu_var(vmap_block_queue);
1993 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1994 unsigned long pages_off;
1996 spin_lock(&vb->lock);
1997 if (vb->free < (1UL << order)) {
1998 spin_unlock(&vb->lock);
2002 pages_off = VMAP_BBMAP_BITS - vb->free;
2003 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2004 vb->free -= 1UL << order;
2005 if (vb->free == 0) {
2006 spin_lock(&vbq->lock);
2007 list_del_rcu(&vb->free_list);
2008 spin_unlock(&vbq->lock);
2011 spin_unlock(&vb->lock);
2015 put_cpu_var(vmap_block_queue);
2018 /* Allocate new block if nothing was found */
2020 vaddr = new_vmap_block(order, gfp_mask);
2025 static void vb_free(unsigned long addr, unsigned long size)
2027 unsigned long offset;
2029 struct vmap_block *vb;
2031 BUG_ON(offset_in_page(size));
2032 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2034 flush_cache_vunmap(addr, addr + size);
2036 order = get_order(size);
2037 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2038 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2040 vunmap_range_noflush(addr, addr + size);
2042 if (debug_pagealloc_enabled_static())
2043 flush_tlb_kernel_range(addr, addr + size);
2045 spin_lock(&vb->lock);
2047 /* Expand dirty range */
2048 vb->dirty_min = min(vb->dirty_min, offset);
2049 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2051 vb->dirty += 1UL << order;
2052 if (vb->dirty == VMAP_BBMAP_BITS) {
2054 spin_unlock(&vb->lock);
2055 free_vmap_block(vb);
2057 spin_unlock(&vb->lock);
2060 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2064 if (unlikely(!vmap_initialized))
2069 for_each_possible_cpu(cpu) {
2070 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2071 struct vmap_block *vb;
2074 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2075 spin_lock(&vb->lock);
2076 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2077 unsigned long va_start = vb->va->va_start;
2080 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2081 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2083 start = min(s, start);
2088 spin_unlock(&vb->lock);
2093 mutex_lock(&vmap_purge_lock);
2094 purge_fragmented_blocks_allcpus();
2095 if (!__purge_vmap_area_lazy(start, end) && flush)
2096 flush_tlb_kernel_range(start, end);
2097 mutex_unlock(&vmap_purge_lock);
2101 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2103 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2104 * to amortize TLB flushing overheads. What this means is that any page you
2105 * have now, may, in a former life, have been mapped into kernel virtual
2106 * address by the vmap layer and so there might be some CPUs with TLB entries
2107 * still referencing that page (additional to the regular 1:1 kernel mapping).
2109 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2110 * be sure that none of the pages we have control over will have any aliases
2111 * from the vmap layer.
2113 void vm_unmap_aliases(void)
2115 unsigned long start = ULONG_MAX, end = 0;
2118 _vm_unmap_aliases(start, end, flush);
2120 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2123 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2124 * @mem: the pointer returned by vm_map_ram
2125 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2127 void vm_unmap_ram(const void *mem, unsigned int count)
2129 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2130 unsigned long addr = (unsigned long)mem;
2131 struct vmap_area *va;
2135 BUG_ON(addr < VMALLOC_START);
2136 BUG_ON(addr > VMALLOC_END);
2137 BUG_ON(!PAGE_ALIGNED(addr));
2139 kasan_poison_vmalloc(mem, size);
2141 if (likely(count <= VMAP_MAX_ALLOC)) {
2142 debug_check_no_locks_freed(mem, size);
2143 vb_free(addr, size);
2147 va = find_vmap_area(addr);
2149 debug_check_no_locks_freed((void *)va->va_start,
2150 (va->va_end - va->va_start));
2151 free_unmap_vmap_area(va);
2153 EXPORT_SYMBOL(vm_unmap_ram);
2156 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2157 * @pages: an array of pointers to the pages to be mapped
2158 * @count: number of pages
2159 * @node: prefer to allocate data structures on this node
2161 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2162 * faster than vmap so it's good. But if you mix long-life and short-life
2163 * objects with vm_map_ram(), it could consume lots of address space through
2164 * fragmentation (especially on a 32bit machine). You could see failures in
2165 * the end. Please use this function for short-lived objects.
2167 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2169 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2171 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2175 if (likely(count <= VMAP_MAX_ALLOC)) {
2176 mem = vb_alloc(size, GFP_KERNEL);
2179 addr = (unsigned long)mem;
2181 struct vmap_area *va;
2182 va = alloc_vmap_area(size, PAGE_SIZE,
2183 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2187 addr = va->va_start;
2191 kasan_unpoison_vmalloc(mem, size);
2193 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2194 pages, PAGE_SHIFT) < 0) {
2195 vm_unmap_ram(mem, count);
2201 EXPORT_SYMBOL(vm_map_ram);
2203 static struct vm_struct *vmlist __initdata;
2205 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2207 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2208 return vm->page_order;
2214 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2216 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2217 vm->page_order = order;
2224 * vm_area_add_early - add vmap area early during boot
2225 * @vm: vm_struct to add
2227 * This function is used to add fixed kernel vm area to vmlist before
2228 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2229 * should contain proper values and the other fields should be zero.
2231 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2233 void __init vm_area_add_early(struct vm_struct *vm)
2235 struct vm_struct *tmp, **p;
2237 BUG_ON(vmap_initialized);
2238 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2239 if (tmp->addr >= vm->addr) {
2240 BUG_ON(tmp->addr < vm->addr + vm->size);
2243 BUG_ON(tmp->addr + tmp->size > vm->addr);
2250 * vm_area_register_early - register vmap area early during boot
2251 * @vm: vm_struct to register
2252 * @align: requested alignment
2254 * This function is used to register kernel vm area before
2255 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2256 * proper values on entry and other fields should be zero. On return,
2257 * vm->addr contains the allocated address.
2259 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2261 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2263 static size_t vm_init_off __initdata;
2266 addr = ALIGN(VMALLOC_START + vm_init_off, align);
2267 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
2269 vm->addr = (void *)addr;
2271 vm_area_add_early(vm);
2274 static void vmap_init_free_space(void)
2276 unsigned long vmap_start = 1;
2277 const unsigned long vmap_end = ULONG_MAX;
2278 struct vmap_area *busy, *free;
2282 * -|-----|.....|-----|-----|-----|.....|-
2284 * |<--------------------------------->|
2286 list_for_each_entry(busy, &vmap_area_list, list) {
2287 if (busy->va_start - vmap_start > 0) {
2288 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2289 if (!WARN_ON_ONCE(!free)) {
2290 free->va_start = vmap_start;
2291 free->va_end = busy->va_start;
2293 insert_vmap_area_augment(free, NULL,
2294 &free_vmap_area_root,
2295 &free_vmap_area_list);
2299 vmap_start = busy->va_end;
2302 if (vmap_end - vmap_start > 0) {
2303 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2304 if (!WARN_ON_ONCE(!free)) {
2305 free->va_start = vmap_start;
2306 free->va_end = vmap_end;
2308 insert_vmap_area_augment(free, NULL,
2309 &free_vmap_area_root,
2310 &free_vmap_area_list);
2315 void __init vmalloc_init(void)
2317 struct vmap_area *va;
2318 struct vm_struct *tmp;
2322 * Create the cache for vmap_area objects.
2324 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2326 for_each_possible_cpu(i) {
2327 struct vmap_block_queue *vbq;
2328 struct vfree_deferred *p;
2330 vbq = &per_cpu(vmap_block_queue, i);
2331 spin_lock_init(&vbq->lock);
2332 INIT_LIST_HEAD(&vbq->free);
2333 p = &per_cpu(vfree_deferred, i);
2334 init_llist_head(&p->list);
2335 INIT_WORK(&p->wq, free_work);
2338 /* Import existing vmlist entries. */
2339 for (tmp = vmlist; tmp; tmp = tmp->next) {
2340 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2341 if (WARN_ON_ONCE(!va))
2344 va->va_start = (unsigned long)tmp->addr;
2345 va->va_end = va->va_start + tmp->size;
2347 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2351 * Now we can initialize a free vmap space.
2353 vmap_init_free_space();
2354 vmap_initialized = true;
2357 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2358 struct vmap_area *va, unsigned long flags, const void *caller)
2361 vm->addr = (void *)va->va_start;
2362 vm->size = va->va_end - va->va_start;
2363 vm->caller = caller;
2367 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2368 unsigned long flags, const void *caller)
2370 spin_lock(&vmap_area_lock);
2371 setup_vmalloc_vm_locked(vm, va, flags, caller);
2372 spin_unlock(&vmap_area_lock);
2375 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2378 * Before removing VM_UNINITIALIZED,
2379 * we should make sure that vm has proper values.
2380 * Pair with smp_rmb() in show_numa_info().
2383 vm->flags &= ~VM_UNINITIALIZED;
2386 static struct vm_struct *__get_vm_area_node(unsigned long size,
2387 unsigned long align, unsigned long shift, unsigned long flags,
2388 unsigned long start, unsigned long end, int node,
2389 gfp_t gfp_mask, const void *caller)
2391 struct vmap_area *va;
2392 struct vm_struct *area;
2393 unsigned long requested_size = size;
2395 BUG_ON(in_interrupt());
2396 size = ALIGN(size, 1ul << shift);
2397 if (unlikely(!size))
2400 if (flags & VM_IOREMAP)
2401 align = 1ul << clamp_t(int, get_count_order_long(size),
2402 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2404 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2405 if (unlikely(!area))
2408 if (!(flags & VM_NO_GUARD))
2411 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2417 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2419 setup_vmalloc_vm(area, va, flags, caller);
2424 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2425 unsigned long start, unsigned long end,
2428 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2429 NUMA_NO_NODE, GFP_KERNEL, caller);
2433 * get_vm_area - reserve a contiguous kernel virtual area
2434 * @size: size of the area
2435 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2437 * Search an area of @size in the kernel virtual mapping area,
2438 * and reserved it for out purposes. Returns the area descriptor
2439 * on success or %NULL on failure.
2441 * Return: the area descriptor on success or %NULL on failure.
2443 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2445 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2446 VMALLOC_START, VMALLOC_END,
2447 NUMA_NO_NODE, GFP_KERNEL,
2448 __builtin_return_address(0));
2451 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2454 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2455 VMALLOC_START, VMALLOC_END,
2456 NUMA_NO_NODE, GFP_KERNEL, caller);
2460 * find_vm_area - find a continuous kernel virtual area
2461 * @addr: base address
2463 * Search for the kernel VM area starting at @addr, and return it.
2464 * It is up to the caller to do all required locking to keep the returned
2467 * Return: the area descriptor on success or %NULL on failure.
2469 struct vm_struct *find_vm_area(const void *addr)
2471 struct vmap_area *va;
2473 va = find_vmap_area((unsigned long)addr);
2481 * remove_vm_area - find and remove a continuous kernel virtual area
2482 * @addr: base address
2484 * Search for the kernel VM area starting at @addr, and remove it.
2485 * This function returns the found VM area, but using it is NOT safe
2486 * on SMP machines, except for its size or flags.
2488 * Return: the area descriptor on success or %NULL on failure.
2490 struct vm_struct *remove_vm_area(const void *addr)
2492 struct vmap_area *va;
2496 spin_lock(&vmap_area_lock);
2497 va = __find_vmap_area((unsigned long)addr);
2499 struct vm_struct *vm = va->vm;
2502 spin_unlock(&vmap_area_lock);
2504 kasan_free_shadow(vm);
2505 free_unmap_vmap_area(va);
2510 spin_unlock(&vmap_area_lock);
2514 static inline void set_area_direct_map(const struct vm_struct *area,
2515 int (*set_direct_map)(struct page *page))
2519 /* HUGE_VMALLOC passes small pages to set_direct_map */
2520 for (i = 0; i < area->nr_pages; i++)
2521 if (page_address(area->pages[i]))
2522 set_direct_map(area->pages[i]);
2525 /* Handle removing and resetting vm mappings related to the vm_struct. */
2526 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2528 unsigned long start = ULONG_MAX, end = 0;
2529 unsigned int page_order = vm_area_page_order(area);
2530 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2534 remove_vm_area(area->addr);
2536 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2541 * If not deallocating pages, just do the flush of the VM area and
2544 if (!deallocate_pages) {
2550 * If execution gets here, flush the vm mapping and reset the direct
2551 * map. Find the start and end range of the direct mappings to make sure
2552 * the vm_unmap_aliases() flush includes the direct map.
2554 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2555 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2557 unsigned long page_size;
2559 page_size = PAGE_SIZE << page_order;
2560 start = min(addr, start);
2561 end = max(addr + page_size, end);
2567 * Set direct map to something invalid so that it won't be cached if
2568 * there are any accesses after the TLB flush, then flush the TLB and
2569 * reset the direct map permissions to the default.
2571 set_area_direct_map(area, set_direct_map_invalid_noflush);
2572 _vm_unmap_aliases(start, end, flush_dmap);
2573 set_area_direct_map(area, set_direct_map_default_noflush);
2576 static void __vunmap(const void *addr, int deallocate_pages)
2578 struct vm_struct *area;
2583 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2587 area = find_vm_area(addr);
2588 if (unlikely(!area)) {
2589 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2594 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2595 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2597 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2599 vm_remove_mappings(area, deallocate_pages);
2601 if (deallocate_pages) {
2602 unsigned int page_order = vm_area_page_order(area);
2605 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2606 struct page *page = area->pages[i];
2609 __free_pages(page, page_order);
2612 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2614 kvfree(area->pages);
2620 static inline void __vfree_deferred(const void *addr)
2623 * Use raw_cpu_ptr() because this can be called from preemptible
2624 * context. Preemption is absolutely fine here, because the llist_add()
2625 * implementation is lockless, so it works even if we are adding to
2626 * another cpu's list. schedule_work() should be fine with this too.
2628 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2630 if (llist_add((struct llist_node *)addr, &p->list))
2631 schedule_work(&p->wq);
2635 * vfree_atomic - release memory allocated by vmalloc()
2636 * @addr: memory base address
2638 * This one is just like vfree() but can be called in any atomic context
2641 void vfree_atomic(const void *addr)
2645 kmemleak_free(addr);
2649 __vfree_deferred(addr);
2652 static void __vfree(const void *addr)
2654 if (unlikely(in_interrupt()))
2655 __vfree_deferred(addr);
2661 * vfree - Release memory allocated by vmalloc()
2662 * @addr: Memory base address
2664 * Free the virtually continuous memory area starting at @addr, as obtained
2665 * from one of the vmalloc() family of APIs. This will usually also free the
2666 * physical memory underlying the virtual allocation, but that memory is
2667 * reference counted, so it will not be freed until the last user goes away.
2669 * If @addr is NULL, no operation is performed.
2672 * May sleep if called *not* from interrupt context.
2673 * Must not be called in NMI context (strictly speaking, it could be
2674 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2675 * conventions for vfree() arch-dependent would be a really bad idea).
2677 void vfree(const void *addr)
2681 kmemleak_free(addr);
2683 might_sleep_if(!in_interrupt());
2690 EXPORT_SYMBOL(vfree);
2693 * vunmap - release virtual mapping obtained by vmap()
2694 * @addr: memory base address
2696 * Free the virtually contiguous memory area starting at @addr,
2697 * which was created from the page array passed to vmap().
2699 * Must not be called in interrupt context.
2701 void vunmap(const void *addr)
2703 BUG_ON(in_interrupt());
2708 EXPORT_SYMBOL(vunmap);
2711 * vmap - map an array of pages into virtually contiguous space
2712 * @pages: array of page pointers
2713 * @count: number of pages to map
2714 * @flags: vm_area->flags
2715 * @prot: page protection for the mapping
2717 * Maps @count pages from @pages into contiguous kernel virtual space.
2718 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2719 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2720 * are transferred from the caller to vmap(), and will be freed / dropped when
2721 * vfree() is called on the return value.
2723 * Return: the address of the area or %NULL on failure
2725 void *vmap(struct page **pages, unsigned int count,
2726 unsigned long flags, pgprot_t prot)
2728 struct vm_struct *area;
2730 unsigned long size; /* In bytes */
2734 if (count > totalram_pages())
2737 size = (unsigned long)count << PAGE_SHIFT;
2738 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2742 addr = (unsigned long)area->addr;
2743 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2744 pages, PAGE_SHIFT) < 0) {
2749 if (flags & VM_MAP_PUT_PAGES) {
2750 area->pages = pages;
2751 area->nr_pages = count;
2755 EXPORT_SYMBOL(vmap);
2757 #ifdef CONFIG_VMAP_PFN
2758 struct vmap_pfn_data {
2759 unsigned long *pfns;
2764 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2766 struct vmap_pfn_data *data = private;
2768 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2770 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2775 * vmap_pfn - map an array of PFNs into virtually contiguous space
2776 * @pfns: array of PFNs
2777 * @count: number of pages to map
2778 * @prot: page protection for the mapping
2780 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2781 * the start address of the mapping.
2783 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2785 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2786 struct vm_struct *area;
2788 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2789 __builtin_return_address(0));
2792 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2793 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2799 EXPORT_SYMBOL_GPL(vmap_pfn);
2800 #endif /* CONFIG_VMAP_PFN */
2802 static inline unsigned int
2803 vm_area_alloc_pages(gfp_t gfp, int nid,
2804 unsigned int order, unsigned int nr_pages, struct page **pages)
2806 unsigned int nr_allocated = 0;
2809 * For order-0 pages we make use of bulk allocator, if
2810 * the page array is partly or not at all populated due
2811 * to fails, fallback to a single page allocator that is
2815 while (nr_allocated < nr_pages) {
2816 unsigned int nr, nr_pages_request;
2819 * A maximum allowed request is hard-coded and is 100
2820 * pages per call. That is done in order to prevent a
2821 * long preemption off scenario in the bulk-allocator
2822 * so the range is [1:100].
2824 nr_pages_request = min(100U, nr_pages - nr_allocated);
2826 nr = alloc_pages_bulk_array_node(gfp, nid,
2827 nr_pages_request, pages + nr_allocated);
2833 * If zero or pages were obtained partly,
2834 * fallback to a single page allocator.
2836 if (nr != nr_pages_request)
2841 * Compound pages required for remap_vmalloc_page if
2846 /* High-order pages or fallback path if "bulk" fails. */
2847 while (nr_allocated < nr_pages) {
2851 page = alloc_pages_node(nid, gfp, order);
2852 if (unlikely(!page))
2856 * Careful, we allocate and map page-order pages, but
2857 * tracking is done per PAGE_SIZE page so as to keep the
2858 * vm_struct APIs independent of the physical/mapped size.
2860 for (i = 0; i < (1U << order); i++)
2861 pages[nr_allocated + i] = page + i;
2864 nr_allocated += 1U << order;
2867 return nr_allocated;
2870 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2871 pgprot_t prot, unsigned int page_shift,
2874 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2875 unsigned long addr = (unsigned long)area->addr;
2876 unsigned long size = get_vm_area_size(area);
2877 unsigned long array_size;
2878 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2879 unsigned int page_order;
2881 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2882 gfp_mask |= __GFP_NOWARN;
2883 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2884 gfp_mask |= __GFP_HIGHMEM;
2886 /* Please note that the recursion is strictly bounded. */
2887 if (array_size > PAGE_SIZE) {
2888 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2891 area->pages = kmalloc_node(array_size, nested_gfp, node);
2895 warn_alloc(gfp_mask, NULL,
2896 "vmalloc error: size %lu, failed to allocated page array size %lu",
2897 nr_small_pages * PAGE_SIZE, array_size);
2902 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2903 page_order = vm_area_page_order(area);
2905 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2906 page_order, nr_small_pages, area->pages);
2908 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2911 * If not enough pages were obtained to accomplish an
2912 * allocation request, free them via __vfree() if any.
2914 if (area->nr_pages != nr_small_pages) {
2915 warn_alloc(gfp_mask, NULL,
2916 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2917 area->nr_pages * PAGE_SIZE, page_order);
2921 if (vmap_pages_range(addr, addr + size, prot, area->pages,
2923 warn_alloc(gfp_mask, NULL,
2924 "vmalloc error: size %lu, failed to map pages",
2925 area->nr_pages * PAGE_SIZE);
2932 __vfree(area->addr);
2937 * __vmalloc_node_range - allocate virtually contiguous memory
2938 * @size: allocation size
2939 * @align: desired alignment
2940 * @start: vm area range start
2941 * @end: vm area range end
2942 * @gfp_mask: flags for the page level allocator
2943 * @prot: protection mask for the allocated pages
2944 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2945 * @node: node to use for allocation or NUMA_NO_NODE
2946 * @caller: caller's return address
2948 * Allocate enough pages to cover @size from the page level
2949 * allocator with @gfp_mask flags. Map them into contiguous
2950 * kernel virtual space, using a pagetable protection of @prot.
2952 * Return: the address of the area or %NULL on failure
2954 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2955 unsigned long start, unsigned long end, gfp_t gfp_mask,
2956 pgprot_t prot, unsigned long vm_flags, int node,
2959 struct vm_struct *area;
2961 unsigned long real_size = size;
2962 unsigned long real_align = align;
2963 unsigned int shift = PAGE_SHIFT;
2965 if (WARN_ON_ONCE(!size))
2968 if ((size >> PAGE_SHIFT) > totalram_pages()) {
2969 warn_alloc(gfp_mask, NULL,
2970 "vmalloc error: size %lu, exceeds total pages",
2975 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
2976 unsigned long size_per_node;
2979 * Try huge pages. Only try for PAGE_KERNEL allocations,
2980 * others like modules don't yet expect huge pages in
2981 * their allocations due to apply_to_page_range not
2985 size_per_node = size;
2986 if (node == NUMA_NO_NODE)
2987 size_per_node /= num_online_nodes();
2988 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
2991 shift = arch_vmap_pte_supported_shift(size_per_node);
2993 align = max(real_align, 1UL << shift);
2994 size = ALIGN(real_size, 1UL << shift);
2998 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
2999 VM_UNINITIALIZED | vm_flags, start, end, node,
3002 warn_alloc(gfp_mask, NULL,
3003 "vmalloc error: size %lu, vm_struct allocation failed",
3008 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3013 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3014 * flag. It means that vm_struct is not fully initialized.
3015 * Now, it is fully initialized, so remove this flag here.
3017 clear_vm_uninitialized_flag(area);
3019 size = PAGE_ALIGN(size);
3020 kmemleak_vmalloc(area, size, gfp_mask);
3025 if (shift > PAGE_SHIFT) {
3036 * __vmalloc_node - allocate virtually contiguous memory
3037 * @size: allocation size
3038 * @align: desired alignment
3039 * @gfp_mask: flags for the page level allocator
3040 * @node: node to use for allocation or NUMA_NO_NODE
3041 * @caller: caller's return address
3043 * Allocate enough pages to cover @size from the page level allocator with
3044 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3046 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3047 * and __GFP_NOFAIL are not supported
3049 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3052 * Return: pointer to the allocated memory or %NULL on error
3054 void *__vmalloc_node(unsigned long size, unsigned long align,
3055 gfp_t gfp_mask, int node, const void *caller)
3057 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3058 gfp_mask, PAGE_KERNEL, 0, node, caller);
3061 * This is only for performance analysis of vmalloc and stress purpose.
3062 * It is required by vmalloc test module, therefore do not use it other
3065 #ifdef CONFIG_TEST_VMALLOC_MODULE
3066 EXPORT_SYMBOL_GPL(__vmalloc_node);
3069 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3071 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3072 __builtin_return_address(0));
3074 EXPORT_SYMBOL(__vmalloc);
3077 * vmalloc - allocate virtually contiguous memory
3078 * @size: allocation size
3080 * Allocate enough pages to cover @size from the page level
3081 * allocator and map them into contiguous kernel virtual space.
3083 * For tight control over page level allocator and protection flags
3084 * use __vmalloc() instead.
3086 * Return: pointer to the allocated memory or %NULL on error
3088 void *vmalloc(unsigned long size)
3090 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3091 __builtin_return_address(0));
3093 EXPORT_SYMBOL(vmalloc);
3096 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3097 * @size: allocation size
3099 * Allocate enough non-huge pages to cover @size from the page level
3100 * allocator and map them into contiguous kernel virtual space.
3102 * Return: pointer to the allocated memory or %NULL on error
3104 void *vmalloc_no_huge(unsigned long size)
3106 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3107 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3108 NUMA_NO_NODE, __builtin_return_address(0));
3110 EXPORT_SYMBOL(vmalloc_no_huge);
3113 * vzalloc - allocate virtually contiguous memory with zero fill
3114 * @size: allocation size
3116 * Allocate enough pages to cover @size from the page level
3117 * allocator and map them into contiguous kernel virtual space.
3118 * The memory allocated is set to zero.
3120 * For tight control over page level allocator and protection flags
3121 * use __vmalloc() instead.
3123 * Return: pointer to the allocated memory or %NULL on error
3125 void *vzalloc(unsigned long size)
3127 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3128 __builtin_return_address(0));
3130 EXPORT_SYMBOL(vzalloc);
3133 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3134 * @size: allocation size
3136 * The resulting memory area is zeroed so it can be mapped to userspace
3137 * without leaking data.
3139 * Return: pointer to the allocated memory or %NULL on error
3141 void *vmalloc_user(unsigned long size)
3143 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3144 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3145 VM_USERMAP, NUMA_NO_NODE,
3146 __builtin_return_address(0));
3148 EXPORT_SYMBOL(vmalloc_user);
3151 * vmalloc_node - allocate memory on a specific node
3152 * @size: allocation size
3155 * Allocate enough pages to cover @size from the page level
3156 * allocator and map them into contiguous kernel virtual space.
3158 * For tight control over page level allocator and protection flags
3159 * use __vmalloc() instead.
3161 * Return: pointer to the allocated memory or %NULL on error
3163 void *vmalloc_node(unsigned long size, int node)
3165 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3166 __builtin_return_address(0));
3168 EXPORT_SYMBOL(vmalloc_node);
3171 * vzalloc_node - allocate memory on a specific node with zero fill
3172 * @size: allocation size
3175 * Allocate enough pages to cover @size from the page level
3176 * allocator and map them into contiguous kernel virtual space.
3177 * The memory allocated is set to zero.
3179 * Return: pointer to the allocated memory or %NULL on error
3181 void *vzalloc_node(unsigned long size, int node)
3183 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3184 __builtin_return_address(0));
3186 EXPORT_SYMBOL(vzalloc_node);
3188 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3189 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3190 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3191 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3194 * 64b systems should always have either DMA or DMA32 zones. For others
3195 * GFP_DMA32 should do the right thing and use the normal zone.
3197 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3201 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3202 * @size: allocation size
3204 * Allocate enough 32bit PA addressable pages to cover @size from the
3205 * page level allocator and map them into contiguous kernel virtual space.
3207 * Return: pointer to the allocated memory or %NULL on error
3209 void *vmalloc_32(unsigned long size)
3211 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3212 __builtin_return_address(0));
3214 EXPORT_SYMBOL(vmalloc_32);
3217 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3218 * @size: allocation size
3220 * The resulting memory area is 32bit addressable and zeroed so it can be
3221 * mapped to userspace without leaking data.
3223 * Return: pointer to the allocated memory or %NULL on error
3225 void *vmalloc_32_user(unsigned long size)
3227 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3228 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3229 VM_USERMAP, NUMA_NO_NODE,
3230 __builtin_return_address(0));
3232 EXPORT_SYMBOL(vmalloc_32_user);
3235 * small helper routine , copy contents to buf from addr.
3236 * If the page is not present, fill zero.
3239 static int aligned_vread(char *buf, char *addr, unsigned long count)
3245 unsigned long offset, length;
3247 offset = offset_in_page(addr);
3248 length = PAGE_SIZE - offset;
3251 p = vmalloc_to_page(addr);
3253 * To do safe access to this _mapped_ area, we need
3254 * lock. But adding lock here means that we need to add
3255 * overhead of vmalloc()/vfree() calls for this _debug_
3256 * interface, rarely used. Instead of that, we'll use
3257 * kmap() and get small overhead in this access function.
3260 /* We can expect USER0 is not used -- see vread() */
3261 void *map = kmap_atomic(p);
3262 memcpy(buf, map + offset, length);
3265 memset(buf, 0, length);
3276 * vread() - read vmalloc area in a safe way.
3277 * @buf: buffer for reading data
3278 * @addr: vm address.
3279 * @count: number of bytes to be read.
3281 * This function checks that addr is a valid vmalloc'ed area, and
3282 * copy data from that area to a given buffer. If the given memory range
3283 * of [addr...addr+count) includes some valid address, data is copied to
3284 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3285 * IOREMAP area is treated as memory hole and no copy is done.
3287 * If [addr...addr+count) doesn't includes any intersects with alive
3288 * vm_struct area, returns 0. @buf should be kernel's buffer.
3290 * Note: In usual ops, vread() is never necessary because the caller
3291 * should know vmalloc() area is valid and can use memcpy().
3292 * This is for routines which have to access vmalloc area without
3293 * any information, as /proc/kcore.
3295 * Return: number of bytes for which addr and buf should be increased
3296 * (same number as @count) or %0 if [addr...addr+count) doesn't
3297 * include any intersection with valid vmalloc area
3299 long vread(char *buf, char *addr, unsigned long count)
3301 struct vmap_area *va;
3302 struct vm_struct *vm;
3303 char *vaddr, *buf_start = buf;
3304 unsigned long buflen = count;
3307 /* Don't allow overflow */
3308 if ((unsigned long) addr + count < count)
3309 count = -(unsigned long) addr;
3311 spin_lock(&vmap_area_lock);
3312 va = find_vmap_area_exceed_addr((unsigned long)addr);
3316 /* no intersects with alive vmap_area */
3317 if ((unsigned long)addr + count <= va->va_start)
3320 list_for_each_entry_from(va, &vmap_area_list, list) {
3328 vaddr = (char *) vm->addr;
3329 if (addr >= vaddr + get_vm_area_size(vm))
3331 while (addr < vaddr) {
3339 n = vaddr + get_vm_area_size(vm) - addr;
3342 if (!(vm->flags & VM_IOREMAP))
3343 aligned_vread(buf, addr, n);
3344 else /* IOREMAP area is treated as memory hole */
3351 spin_unlock(&vmap_area_lock);
3353 if (buf == buf_start)
3355 /* zero-fill memory holes */
3356 if (buf != buf_start + buflen)
3357 memset(buf, 0, buflen - (buf - buf_start));
3363 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3364 * @vma: vma to cover
3365 * @uaddr: target user address to start at
3366 * @kaddr: virtual address of vmalloc kernel memory
3367 * @pgoff: offset from @kaddr to start at
3368 * @size: size of map area
3370 * Returns: 0 for success, -Exxx on failure
3372 * This function checks that @kaddr is a valid vmalloc'ed area,
3373 * and that it is big enough to cover the range starting at
3374 * @uaddr in @vma. Will return failure if that criteria isn't
3377 * Similar to remap_pfn_range() (see mm/memory.c)
3379 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3380 void *kaddr, unsigned long pgoff,
3383 struct vm_struct *area;
3385 unsigned long end_index;
3387 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3390 size = PAGE_ALIGN(size);
3392 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3395 area = find_vm_area(kaddr);
3399 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3402 if (check_add_overflow(size, off, &end_index) ||
3403 end_index > get_vm_area_size(area))
3408 struct page *page = vmalloc_to_page(kaddr);
3411 ret = vm_insert_page(vma, uaddr, page);
3420 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3426 * remap_vmalloc_range - map vmalloc pages to userspace
3427 * @vma: vma to cover (map full range of vma)
3428 * @addr: vmalloc memory
3429 * @pgoff: number of pages into addr before first page to map
3431 * Returns: 0 for success, -Exxx on failure
3433 * This function checks that addr is a valid vmalloc'ed area, and
3434 * that it is big enough to cover the vma. Will return failure if
3435 * that criteria isn't met.
3437 * Similar to remap_pfn_range() (see mm/memory.c)
3439 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3440 unsigned long pgoff)
3442 return remap_vmalloc_range_partial(vma, vma->vm_start,
3444 vma->vm_end - vma->vm_start);
3446 EXPORT_SYMBOL(remap_vmalloc_range);
3448 void free_vm_area(struct vm_struct *area)
3450 struct vm_struct *ret;
3451 ret = remove_vm_area(area->addr);
3452 BUG_ON(ret != area);
3455 EXPORT_SYMBOL_GPL(free_vm_area);
3458 static struct vmap_area *node_to_va(struct rb_node *n)
3460 return rb_entry_safe(n, struct vmap_area, rb_node);
3464 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3465 * @addr: target address
3467 * Returns: vmap_area if it is found. If there is no such area
3468 * the first highest(reverse order) vmap_area is returned
3469 * i.e. va->va_start < addr && va->va_end < addr or NULL
3470 * if there are no any areas before @addr.
3472 static struct vmap_area *
3473 pvm_find_va_enclose_addr(unsigned long addr)
3475 struct vmap_area *va, *tmp;
3478 n = free_vmap_area_root.rb_node;
3482 tmp = rb_entry(n, struct vmap_area, rb_node);
3483 if (tmp->va_start <= addr) {
3485 if (tmp->va_end >= addr)
3498 * pvm_determine_end_from_reverse - find the highest aligned address
3499 * of free block below VMALLOC_END
3501 * in - the VA we start the search(reverse order);
3502 * out - the VA with the highest aligned end address.
3503 * @align: alignment for required highest address
3505 * Returns: determined end address within vmap_area
3507 static unsigned long
3508 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3510 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3514 list_for_each_entry_from_reverse((*va),
3515 &free_vmap_area_list, list) {
3516 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3517 if ((*va)->va_start < addr)
3526 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3527 * @offsets: array containing offset of each area
3528 * @sizes: array containing size of each area
3529 * @nr_vms: the number of areas to allocate
3530 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3532 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3533 * vm_structs on success, %NULL on failure
3535 * Percpu allocator wants to use congruent vm areas so that it can
3536 * maintain the offsets among percpu areas. This function allocates
3537 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3538 * be scattered pretty far, distance between two areas easily going up
3539 * to gigabytes. To avoid interacting with regular vmallocs, these
3540 * areas are allocated from top.
3542 * Despite its complicated look, this allocator is rather simple. It
3543 * does everything top-down and scans free blocks from the end looking
3544 * for matching base. While scanning, if any of the areas do not fit the
3545 * base address is pulled down to fit the area. Scanning is repeated till
3546 * all the areas fit and then all necessary data structures are inserted
3547 * and the result is returned.
3549 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3550 const size_t *sizes, int nr_vms,
3553 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3554 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3555 struct vmap_area **vas, *va;
3556 struct vm_struct **vms;
3557 int area, area2, last_area, term_area;
3558 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3559 bool purged = false;
3562 /* verify parameters and allocate data structures */
3563 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3564 for (last_area = 0, area = 0; area < nr_vms; area++) {
3565 start = offsets[area];
3566 end = start + sizes[area];
3568 /* is everything aligned properly? */
3569 BUG_ON(!IS_ALIGNED(offsets[area], align));
3570 BUG_ON(!IS_ALIGNED(sizes[area], align));
3572 /* detect the area with the highest address */
3573 if (start > offsets[last_area])
3576 for (area2 = area + 1; area2 < nr_vms; area2++) {
3577 unsigned long start2 = offsets[area2];
3578 unsigned long end2 = start2 + sizes[area2];
3580 BUG_ON(start2 < end && start < end2);
3583 last_end = offsets[last_area] + sizes[last_area];
3585 if (vmalloc_end - vmalloc_start < last_end) {
3590 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3591 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3595 for (area = 0; area < nr_vms; area++) {
3596 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3597 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3598 if (!vas[area] || !vms[area])
3602 spin_lock(&free_vmap_area_lock);
3604 /* start scanning - we scan from the top, begin with the last area */
3605 area = term_area = last_area;
3606 start = offsets[area];
3607 end = start + sizes[area];
3609 va = pvm_find_va_enclose_addr(vmalloc_end);
3610 base = pvm_determine_end_from_reverse(&va, align) - end;
3614 * base might have underflowed, add last_end before
3617 if (base + last_end < vmalloc_start + last_end)
3621 * Fitting base has not been found.
3627 * If required width exceeds current VA block, move
3628 * base downwards and then recheck.
3630 if (base + end > va->va_end) {
3631 base = pvm_determine_end_from_reverse(&va, align) - end;
3637 * If this VA does not fit, move base downwards and recheck.
3639 if (base + start < va->va_start) {
3640 va = node_to_va(rb_prev(&va->rb_node));
3641 base = pvm_determine_end_from_reverse(&va, align) - end;
3647 * This area fits, move on to the previous one. If
3648 * the previous one is the terminal one, we're done.
3650 area = (area + nr_vms - 1) % nr_vms;
3651 if (area == term_area)
3654 start = offsets[area];
3655 end = start + sizes[area];
3656 va = pvm_find_va_enclose_addr(base + end);
3659 /* we've found a fitting base, insert all va's */
3660 for (area = 0; area < nr_vms; area++) {
3663 start = base + offsets[area];
3666 va = pvm_find_va_enclose_addr(start);
3667 if (WARN_ON_ONCE(va == NULL))
3668 /* It is a BUG(), but trigger recovery instead. */
3671 type = classify_va_fit_type(va, start, size);
3672 if (WARN_ON_ONCE(type == NOTHING_FIT))
3673 /* It is a BUG(), but trigger recovery instead. */
3676 ret = adjust_va_to_fit_type(va, start, size, type);
3680 /* Allocated area. */
3682 va->va_start = start;
3683 va->va_end = start + size;
3686 spin_unlock(&free_vmap_area_lock);
3688 /* populate the kasan shadow space */
3689 for (area = 0; area < nr_vms; area++) {
3690 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3691 goto err_free_shadow;
3693 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3697 /* insert all vm's */
3698 spin_lock(&vmap_area_lock);
3699 for (area = 0; area < nr_vms; area++) {
3700 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3702 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3705 spin_unlock(&vmap_area_lock);
3712 * Remove previously allocated areas. There is no
3713 * need in removing these areas from the busy tree,
3714 * because they are inserted only on the final step
3715 * and when pcpu_get_vm_areas() is success.
3718 orig_start = vas[area]->va_start;
3719 orig_end = vas[area]->va_end;
3720 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3721 &free_vmap_area_list);
3723 kasan_release_vmalloc(orig_start, orig_end,
3724 va->va_start, va->va_end);
3729 spin_unlock(&free_vmap_area_lock);
3731 purge_vmap_area_lazy();
3734 /* Before "retry", check if we recover. */
3735 for (area = 0; area < nr_vms; area++) {
3739 vas[area] = kmem_cache_zalloc(
3740 vmap_area_cachep, GFP_KERNEL);
3749 for (area = 0; area < nr_vms; area++) {
3751 kmem_cache_free(vmap_area_cachep, vas[area]);
3761 spin_lock(&free_vmap_area_lock);
3763 * We release all the vmalloc shadows, even the ones for regions that
3764 * hadn't been successfully added. This relies on kasan_release_vmalloc
3765 * being able to tolerate this case.
3767 for (area = 0; area < nr_vms; area++) {
3768 orig_start = vas[area]->va_start;
3769 orig_end = vas[area]->va_end;
3770 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3771 &free_vmap_area_list);
3773 kasan_release_vmalloc(orig_start, orig_end,
3774 va->va_start, va->va_end);
3778 spin_unlock(&free_vmap_area_lock);
3785 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3786 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3787 * @nr_vms: the number of allocated areas
3789 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3791 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3795 for (i = 0; i < nr_vms; i++)
3796 free_vm_area(vms[i]);
3799 #endif /* CONFIG_SMP */
3801 #ifdef CONFIG_PRINTK
3802 bool vmalloc_dump_obj(void *object)
3804 struct vm_struct *vm;
3805 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3807 vm = find_vm_area(objp);
3810 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3811 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3816 #ifdef CONFIG_PROC_FS
3817 static void *s_start(struct seq_file *m, loff_t *pos)
3818 __acquires(&vmap_purge_lock)
3819 __acquires(&vmap_area_lock)
3821 mutex_lock(&vmap_purge_lock);
3822 spin_lock(&vmap_area_lock);
3824 return seq_list_start(&vmap_area_list, *pos);
3827 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3829 return seq_list_next(p, &vmap_area_list, pos);
3832 static void s_stop(struct seq_file *m, void *p)
3833 __releases(&vmap_area_lock)
3834 __releases(&vmap_purge_lock)
3836 spin_unlock(&vmap_area_lock);
3837 mutex_unlock(&vmap_purge_lock);
3840 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3842 if (IS_ENABLED(CONFIG_NUMA)) {
3843 unsigned int nr, *counters = m->private;
3848 if (v->flags & VM_UNINITIALIZED)
3850 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3853 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3855 for (nr = 0; nr < v->nr_pages; nr++)
3856 counters[page_to_nid(v->pages[nr])]++;
3858 for_each_node_state(nr, N_HIGH_MEMORY)
3860 seq_printf(m, " N%u=%u", nr, counters[nr]);
3864 static void show_purge_info(struct seq_file *m)
3866 struct vmap_area *va;
3868 spin_lock(&purge_vmap_area_lock);
3869 list_for_each_entry(va, &purge_vmap_area_list, list) {
3870 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3871 (void *)va->va_start, (void *)va->va_end,
3872 va->va_end - va->va_start);
3874 spin_unlock(&purge_vmap_area_lock);
3877 static int s_show(struct seq_file *m, void *p)
3879 struct vmap_area *va;
3880 struct vm_struct *v;
3882 va = list_entry(p, struct vmap_area, list);
3885 * s_show can encounter race with remove_vm_area, !vm on behalf
3886 * of vmap area is being tear down or vm_map_ram allocation.
3889 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3890 (void *)va->va_start, (void *)va->va_end,
3891 va->va_end - va->va_start);
3898 seq_printf(m, "0x%pK-0x%pK %7ld",
3899 v->addr, v->addr + v->size, v->size);
3902 seq_printf(m, " %pS", v->caller);
3905 seq_printf(m, " pages=%d", v->nr_pages);
3908 seq_printf(m, " phys=%pa", &v->phys_addr);
3910 if (v->flags & VM_IOREMAP)
3911 seq_puts(m, " ioremap");
3913 if (v->flags & VM_ALLOC)
3914 seq_puts(m, " vmalloc");
3916 if (v->flags & VM_MAP)
3917 seq_puts(m, " vmap");
3919 if (v->flags & VM_USERMAP)
3920 seq_puts(m, " user");
3922 if (v->flags & VM_DMA_COHERENT)
3923 seq_puts(m, " dma-coherent");
3925 if (is_vmalloc_addr(v->pages))
3926 seq_puts(m, " vpages");
3928 show_numa_info(m, v);
3932 * As a final step, dump "unpurged" areas.
3934 if (list_is_last(&va->list, &vmap_area_list))
3940 static const struct seq_operations vmalloc_op = {
3947 static int __init proc_vmalloc_init(void)
3949 if (IS_ENABLED(CONFIG_NUMA))
3950 proc_create_seq_private("vmallocinfo", 0400, NULL,
3952 nr_node_ids * sizeof(unsigned int), NULL);
3954 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3957 module_init(proc_vmalloc_init);