1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
47 #include "pgalloc-track.h"
49 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
50 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
52 static int __init set_nohugeiomap(char *str)
54 ioremap_max_page_shift = PAGE_SHIFT;
57 early_param("nohugeiomap", set_nohugeiomap);
58 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
59 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
60 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
63 static bool __ro_after_init vmap_allow_huge = true;
65 static int __init set_nohugevmalloc(char *str)
67 vmap_allow_huge = false;
70 early_param("nohugevmalloc", set_nohugevmalloc);
71 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
72 static const bool vmap_allow_huge = false;
73 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 bool is_vmalloc_addr(const void *x)
77 unsigned long addr = (unsigned long)x;
79 return addr >= VMALLOC_START && addr < VMALLOC_END;
81 EXPORT_SYMBOL(is_vmalloc_addr);
83 struct vfree_deferred {
84 struct llist_head list;
85 struct work_struct wq;
87 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
89 static void __vunmap(const void *, int);
91 static void free_work(struct work_struct *w)
93 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
94 struct llist_node *t, *llnode;
96 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
97 __vunmap((void *)llnode, 1);
100 /*** Page table manipulation functions ***/
101 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
102 phys_addr_t phys_addr, pgprot_t prot,
103 unsigned int max_page_shift, pgtbl_mod_mask *mask)
107 unsigned long size = PAGE_SIZE;
109 pfn = phys_addr >> PAGE_SHIFT;
110 pte = pte_alloc_kernel_track(pmd, addr, mask);
114 BUG_ON(!pte_none(*pte));
116 #ifdef CONFIG_HUGETLB_PAGE
117 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
118 if (size != PAGE_SIZE) {
119 pte_t entry = pfn_pte(pfn, prot);
121 entry = pte_mkhuge(entry);
122 entry = arch_make_huge_pte(entry, ilog2(size), 0);
123 set_huge_pte_at(&init_mm, addr, pte, entry);
124 pfn += PFN_DOWN(size);
128 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
130 } while (pte += PFN_DOWN(size), addr += size, addr != end);
131 *mask |= PGTBL_PTE_MODIFIED;
135 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
136 phys_addr_t phys_addr, pgprot_t prot,
137 unsigned int max_page_shift)
139 if (max_page_shift < PMD_SHIFT)
142 if (!arch_vmap_pmd_supported(prot))
145 if ((end - addr) != PMD_SIZE)
148 if (!IS_ALIGNED(addr, PMD_SIZE))
151 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
154 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
157 return pmd_set_huge(pmd, phys_addr, prot);
160 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
161 phys_addr_t phys_addr, pgprot_t prot,
162 unsigned int max_page_shift, pgtbl_mod_mask *mask)
167 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
171 next = pmd_addr_end(addr, end);
173 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
175 *mask |= PGTBL_PMD_MODIFIED;
179 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
181 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
185 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
186 phys_addr_t phys_addr, pgprot_t prot,
187 unsigned int max_page_shift)
189 if (max_page_shift < PUD_SHIFT)
192 if (!arch_vmap_pud_supported(prot))
195 if ((end - addr) != PUD_SIZE)
198 if (!IS_ALIGNED(addr, PUD_SIZE))
201 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
204 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
207 return pud_set_huge(pud, phys_addr, prot);
210 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
211 phys_addr_t phys_addr, pgprot_t prot,
212 unsigned int max_page_shift, pgtbl_mod_mask *mask)
217 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
221 next = pud_addr_end(addr, end);
223 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
225 *mask |= PGTBL_PUD_MODIFIED;
229 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
230 max_page_shift, mask))
232 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
236 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
237 phys_addr_t phys_addr, pgprot_t prot,
238 unsigned int max_page_shift)
240 if (max_page_shift < P4D_SHIFT)
243 if (!arch_vmap_p4d_supported(prot))
246 if ((end - addr) != P4D_SIZE)
249 if (!IS_ALIGNED(addr, P4D_SIZE))
252 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
255 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
258 return p4d_set_huge(p4d, phys_addr, prot);
261 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
262 phys_addr_t phys_addr, pgprot_t prot,
263 unsigned int max_page_shift, pgtbl_mod_mask *mask)
268 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
272 next = p4d_addr_end(addr, end);
274 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
276 *mask |= PGTBL_P4D_MODIFIED;
280 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
281 max_page_shift, mask))
283 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
287 static int vmap_range_noflush(unsigned long addr, unsigned long end,
288 phys_addr_t phys_addr, pgprot_t prot,
289 unsigned int max_page_shift)
295 pgtbl_mod_mask mask = 0;
301 pgd = pgd_offset_k(addr);
303 next = pgd_addr_end(addr, end);
304 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
305 max_page_shift, &mask);
308 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
310 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
311 arch_sync_kernel_mappings(start, end);
316 int ioremap_page_range(unsigned long addr, unsigned long end,
317 phys_addr_t phys_addr, pgprot_t prot)
321 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
322 ioremap_max_page_shift);
323 flush_cache_vmap(addr, end);
327 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
328 pgtbl_mod_mask *mask)
332 pte = pte_offset_kernel(pmd, addr);
334 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
335 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
336 } while (pte++, addr += PAGE_SIZE, addr != end);
337 *mask |= PGTBL_PTE_MODIFIED;
340 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
341 pgtbl_mod_mask *mask)
347 pmd = pmd_offset(pud, addr);
349 next = pmd_addr_end(addr, end);
351 cleared = pmd_clear_huge(pmd);
352 if (cleared || pmd_bad(*pmd))
353 *mask |= PGTBL_PMD_MODIFIED;
357 if (pmd_none_or_clear_bad(pmd))
359 vunmap_pte_range(pmd, addr, next, mask);
362 } while (pmd++, addr = next, addr != end);
365 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
366 pgtbl_mod_mask *mask)
372 pud = pud_offset(p4d, addr);
374 next = pud_addr_end(addr, end);
376 cleared = pud_clear_huge(pud);
377 if (cleared || pud_bad(*pud))
378 *mask |= PGTBL_PUD_MODIFIED;
382 if (pud_none_or_clear_bad(pud))
384 vunmap_pmd_range(pud, addr, next, mask);
385 } while (pud++, addr = next, addr != end);
388 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
389 pgtbl_mod_mask *mask)
395 p4d = p4d_offset(pgd, addr);
397 next = p4d_addr_end(addr, end);
399 cleared = p4d_clear_huge(p4d);
400 if (cleared || p4d_bad(*p4d))
401 *mask |= PGTBL_P4D_MODIFIED;
405 if (p4d_none_or_clear_bad(p4d))
407 vunmap_pud_range(p4d, addr, next, mask);
408 } while (p4d++, addr = next, addr != end);
412 * vunmap_range_noflush is similar to vunmap_range, but does not
413 * flush caches or TLBs.
415 * The caller is responsible for calling flush_cache_vmap() before calling
416 * this function, and flush_tlb_kernel_range after it has returned
417 * successfully (and before the addresses are expected to cause a page fault
418 * or be re-mapped for something else, if TLB flushes are being delayed or
421 * This is an internal function only. Do not use outside mm/.
423 void vunmap_range_noflush(unsigned long start, unsigned long end)
427 unsigned long addr = start;
428 pgtbl_mod_mask mask = 0;
431 pgd = pgd_offset_k(addr);
433 next = pgd_addr_end(addr, end);
435 mask |= PGTBL_PGD_MODIFIED;
436 if (pgd_none_or_clear_bad(pgd))
438 vunmap_p4d_range(pgd, addr, next, &mask);
439 } while (pgd++, addr = next, addr != end);
441 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
442 arch_sync_kernel_mappings(start, end);
446 * vunmap_range - unmap kernel virtual addresses
447 * @addr: start of the VM area to unmap
448 * @end: end of the VM area to unmap (non-inclusive)
450 * Clears any present PTEs in the virtual address range, flushes TLBs and
451 * caches. Any subsequent access to the address before it has been re-mapped
454 void vunmap_range(unsigned long addr, unsigned long end)
456 flush_cache_vunmap(addr, end);
457 vunmap_range_noflush(addr, end);
458 flush_tlb_kernel_range(addr, end);
461 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
462 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
463 pgtbl_mod_mask *mask)
468 * nr is a running index into the array which helps higher level
469 * callers keep track of where we're up to.
472 pte = pte_alloc_kernel_track(pmd, addr, mask);
476 struct page *page = pages[*nr];
478 if (WARN_ON(!pte_none(*pte)))
482 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
484 } while (pte++, addr += PAGE_SIZE, addr != end);
485 *mask |= PGTBL_PTE_MODIFIED;
489 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
490 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
491 pgtbl_mod_mask *mask)
496 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
500 next = pmd_addr_end(addr, end);
501 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
503 } while (pmd++, addr = next, addr != end);
507 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
508 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
509 pgtbl_mod_mask *mask)
514 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
518 next = pud_addr_end(addr, end);
519 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
521 } while (pud++, addr = next, addr != end);
525 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
526 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
527 pgtbl_mod_mask *mask)
532 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
536 next = p4d_addr_end(addr, end);
537 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
539 } while (p4d++, addr = next, addr != end);
543 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
544 pgprot_t prot, struct page **pages)
546 unsigned long start = addr;
551 pgtbl_mod_mask mask = 0;
554 pgd = pgd_offset_k(addr);
556 next = pgd_addr_end(addr, end);
558 mask |= PGTBL_PGD_MODIFIED;
559 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
562 } while (pgd++, addr = next, addr != end);
564 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
565 arch_sync_kernel_mappings(start, end);
571 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
574 * The caller is responsible for calling flush_cache_vmap() after this
575 * function returns successfully and before the addresses are accessed.
577 * This is an internal function only. Do not use outside mm/.
579 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
580 pgprot_t prot, struct page **pages, unsigned int page_shift)
582 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
584 WARN_ON(page_shift < PAGE_SHIFT);
586 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
587 page_shift == PAGE_SHIFT)
588 return vmap_small_pages_range_noflush(addr, end, prot, pages);
590 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
593 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
594 __pa(page_address(pages[i])), prot,
599 addr += 1UL << page_shift;
606 * vmap_pages_range - map pages to a kernel virtual address
607 * @addr: start of the VM area to map
608 * @end: end of the VM area to map (non-inclusive)
609 * @prot: page protection flags to use
610 * @pages: pages to map (always PAGE_SIZE pages)
611 * @page_shift: maximum shift that the pages may be mapped with, @pages must
612 * be aligned and contiguous up to at least this shift.
615 * 0 on success, -errno on failure.
617 static int vmap_pages_range(unsigned long addr, unsigned long end,
618 pgprot_t prot, struct page **pages, unsigned int page_shift)
622 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
623 flush_cache_vmap(addr, end);
627 int is_vmalloc_or_module_addr(const void *x)
630 * ARM, x86-64 and sparc64 put modules in a special place,
631 * and fall back on vmalloc() if that fails. Others
632 * just put it in the vmalloc space.
634 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
635 unsigned long addr = (unsigned long)x;
636 if (addr >= MODULES_VADDR && addr < MODULES_END)
639 return is_vmalloc_addr(x);
643 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
644 * return the tail page that corresponds to the base page address, which
645 * matches small vmap mappings.
647 struct page *vmalloc_to_page(const void *vmalloc_addr)
649 unsigned long addr = (unsigned long) vmalloc_addr;
650 struct page *page = NULL;
651 pgd_t *pgd = pgd_offset_k(addr);
658 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
659 * architectures that do not vmalloc module space
661 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
665 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
666 return NULL; /* XXX: no allowance for huge pgd */
667 if (WARN_ON_ONCE(pgd_bad(*pgd)))
670 p4d = p4d_offset(pgd, addr);
674 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
675 if (WARN_ON_ONCE(p4d_bad(*p4d)))
678 pud = pud_offset(p4d, addr);
682 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
683 if (WARN_ON_ONCE(pud_bad(*pud)))
686 pmd = pmd_offset(pud, addr);
690 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
691 if (WARN_ON_ONCE(pmd_bad(*pmd)))
694 ptep = pte_offset_map(pmd, addr);
696 if (pte_present(pte))
697 page = pte_page(pte);
702 EXPORT_SYMBOL(vmalloc_to_page);
705 * Map a vmalloc()-space virtual address to the physical page frame number.
707 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
709 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
711 EXPORT_SYMBOL(vmalloc_to_pfn);
714 /*** Global kva allocator ***/
716 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
717 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
720 static DEFINE_SPINLOCK(vmap_area_lock);
721 static DEFINE_SPINLOCK(free_vmap_area_lock);
722 /* Export for kexec only */
723 LIST_HEAD(vmap_area_list);
724 static struct rb_root vmap_area_root = RB_ROOT;
725 static bool vmap_initialized __read_mostly;
727 static struct rb_root purge_vmap_area_root = RB_ROOT;
728 static LIST_HEAD(purge_vmap_area_list);
729 static DEFINE_SPINLOCK(purge_vmap_area_lock);
732 * This kmem_cache is used for vmap_area objects. Instead of
733 * allocating from slab we reuse an object from this cache to
734 * make things faster. Especially in "no edge" splitting of
737 static struct kmem_cache *vmap_area_cachep;
740 * This linked list is used in pair with free_vmap_area_root.
741 * It gives O(1) access to prev/next to perform fast coalescing.
743 static LIST_HEAD(free_vmap_area_list);
746 * This augment red-black tree represents the free vmap space.
747 * All vmap_area objects in this tree are sorted by va->va_start
748 * address. It is used for allocation and merging when a vmap
749 * object is released.
751 * Each vmap_area node contains a maximum available free block
752 * of its sub-tree, right or left. Therefore it is possible to
753 * find a lowest match of free area.
755 static struct rb_root free_vmap_area_root = RB_ROOT;
758 * Preload a CPU with one object for "no edge" split case. The
759 * aim is to get rid of allocations from the atomic context, thus
760 * to use more permissive allocation masks.
762 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
764 static __always_inline unsigned long
765 va_size(struct vmap_area *va)
767 return (va->va_end - va->va_start);
770 static __always_inline unsigned long
771 get_subtree_max_size(struct rb_node *node)
773 struct vmap_area *va;
775 va = rb_entry_safe(node, struct vmap_area, rb_node);
776 return va ? va->subtree_max_size : 0;
780 * Gets called when remove the node and rotate.
782 static __always_inline unsigned long
783 compute_subtree_max_size(struct vmap_area *va)
785 return max3(va_size(va),
786 get_subtree_max_size(va->rb_node.rb_left),
787 get_subtree_max_size(va->rb_node.rb_right));
790 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
791 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
793 static void purge_vmap_area_lazy(void);
794 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
795 static unsigned long lazy_max_pages(void);
797 static atomic_long_t nr_vmalloc_pages;
799 unsigned long vmalloc_nr_pages(void)
801 return atomic_long_read(&nr_vmalloc_pages);
804 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
806 struct vmap_area *va = NULL;
807 struct rb_node *n = vmap_area_root.rb_node;
810 struct vmap_area *tmp;
812 tmp = rb_entry(n, struct vmap_area, rb_node);
813 if (tmp->va_end > addr) {
815 if (tmp->va_start <= addr)
826 static struct vmap_area *__find_vmap_area(unsigned long addr)
828 struct rb_node *n = vmap_area_root.rb_node;
831 struct vmap_area *va;
833 va = rb_entry(n, struct vmap_area, rb_node);
834 if (addr < va->va_start)
836 else if (addr >= va->va_end)
846 * This function returns back addresses of parent node
847 * and its left or right link for further processing.
849 * Otherwise NULL is returned. In that case all further
850 * steps regarding inserting of conflicting overlap range
851 * have to be declined and actually considered as a bug.
853 static __always_inline struct rb_node **
854 find_va_links(struct vmap_area *va,
855 struct rb_root *root, struct rb_node *from,
856 struct rb_node **parent)
858 struct vmap_area *tmp_va;
859 struct rb_node **link;
862 link = &root->rb_node;
863 if (unlikely(!*link)) {
872 * Go to the bottom of the tree. When we hit the last point
873 * we end up with parent rb_node and correct direction, i name
874 * it link, where the new va->rb_node will be attached to.
877 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
880 * During the traversal we also do some sanity check.
881 * Trigger the BUG() if there are sides(left/right)
884 if (va->va_start < tmp_va->va_end &&
885 va->va_end <= tmp_va->va_start)
886 link = &(*link)->rb_left;
887 else if (va->va_end > tmp_va->va_start &&
888 va->va_start >= tmp_va->va_end)
889 link = &(*link)->rb_right;
891 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
892 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
898 *parent = &tmp_va->rb_node;
902 static __always_inline struct list_head *
903 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
905 struct list_head *list;
907 if (unlikely(!parent))
909 * The red-black tree where we try to find VA neighbors
910 * before merging or inserting is empty, i.e. it means
911 * there is no free vmap space. Normally it does not
912 * happen but we handle this case anyway.
916 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
917 return (&parent->rb_right == link ? list->next : list);
920 static __always_inline void
921 link_va(struct vmap_area *va, struct rb_root *root,
922 struct rb_node *parent, struct rb_node **link, struct list_head *head)
925 * VA is still not in the list, but we can
926 * identify its future previous list_head node.
928 if (likely(parent)) {
929 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
930 if (&parent->rb_right != link)
934 /* Insert to the rb-tree */
935 rb_link_node(&va->rb_node, parent, link);
936 if (root == &free_vmap_area_root) {
938 * Some explanation here. Just perform simple insertion
939 * to the tree. We do not set va->subtree_max_size to
940 * its current size before calling rb_insert_augmented().
941 * It is because of we populate the tree from the bottom
942 * to parent levels when the node _is_ in the tree.
944 * Therefore we set subtree_max_size to zero after insertion,
945 * to let __augment_tree_propagate_from() puts everything to
946 * the correct order later on.
948 rb_insert_augmented(&va->rb_node,
949 root, &free_vmap_area_rb_augment_cb);
950 va->subtree_max_size = 0;
952 rb_insert_color(&va->rb_node, root);
955 /* Address-sort this list */
956 list_add(&va->list, head);
959 static __always_inline void
960 unlink_va(struct vmap_area *va, struct rb_root *root)
962 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
965 if (root == &free_vmap_area_root)
966 rb_erase_augmented(&va->rb_node,
967 root, &free_vmap_area_rb_augment_cb);
969 rb_erase(&va->rb_node, root);
972 RB_CLEAR_NODE(&va->rb_node);
975 #if DEBUG_AUGMENT_PROPAGATE_CHECK
977 augment_tree_propagate_check(void)
979 struct vmap_area *va;
980 unsigned long computed_size;
982 list_for_each_entry(va, &free_vmap_area_list, list) {
983 computed_size = compute_subtree_max_size(va);
984 if (computed_size != va->subtree_max_size)
985 pr_emerg("tree is corrupted: %lu, %lu\n",
986 va_size(va), va->subtree_max_size);
992 * This function populates subtree_max_size from bottom to upper
993 * levels starting from VA point. The propagation must be done
994 * when VA size is modified by changing its va_start/va_end. Or
995 * in case of newly inserting of VA to the tree.
997 * It means that __augment_tree_propagate_from() must be called:
998 * - After VA has been inserted to the tree(free path);
999 * - After VA has been shrunk(allocation path);
1000 * - After VA has been increased(merging path).
1002 * Please note that, it does not mean that upper parent nodes
1003 * and their subtree_max_size are recalculated all the time up
1012 * For example if we modify the node 4, shrinking it to 2, then
1013 * no any modification is required. If we shrink the node 2 to 1
1014 * its subtree_max_size is updated only, and set to 1. If we shrink
1015 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1016 * node becomes 4--6.
1018 static __always_inline void
1019 augment_tree_propagate_from(struct vmap_area *va)
1022 * Populate the tree from bottom towards the root until
1023 * the calculated maximum available size of checked node
1024 * is equal to its current one.
1026 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1028 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1029 augment_tree_propagate_check();
1034 insert_vmap_area(struct vmap_area *va,
1035 struct rb_root *root, struct list_head *head)
1037 struct rb_node **link;
1038 struct rb_node *parent;
1040 link = find_va_links(va, root, NULL, &parent);
1042 link_va(va, root, parent, link, head);
1046 insert_vmap_area_augment(struct vmap_area *va,
1047 struct rb_node *from, struct rb_root *root,
1048 struct list_head *head)
1050 struct rb_node **link;
1051 struct rb_node *parent;
1054 link = find_va_links(va, NULL, from, &parent);
1056 link = find_va_links(va, root, NULL, &parent);
1059 link_va(va, root, parent, link, head);
1060 augment_tree_propagate_from(va);
1065 * Merge de-allocated chunk of VA memory with previous
1066 * and next free blocks. If coalesce is not done a new
1067 * free area is inserted. If VA has been merged, it is
1070 * Please note, it can return NULL in case of overlap
1071 * ranges, followed by WARN() report. Despite it is a
1072 * buggy behaviour, a system can be alive and keep
1075 static __always_inline struct vmap_area *
1076 merge_or_add_vmap_area(struct vmap_area *va,
1077 struct rb_root *root, struct list_head *head)
1079 struct vmap_area *sibling;
1080 struct list_head *next;
1081 struct rb_node **link;
1082 struct rb_node *parent;
1083 bool merged = false;
1086 * Find a place in the tree where VA potentially will be
1087 * inserted, unless it is merged with its sibling/siblings.
1089 link = find_va_links(va, root, NULL, &parent);
1094 * Get next node of VA to check if merging can be done.
1096 next = get_va_next_sibling(parent, link);
1097 if (unlikely(next == NULL))
1103 * |<------VA------>|<-----Next----->|
1108 sibling = list_entry(next, struct vmap_area, list);
1109 if (sibling->va_start == va->va_end) {
1110 sibling->va_start = va->va_start;
1112 /* Free vmap_area object. */
1113 kmem_cache_free(vmap_area_cachep, va);
1115 /* Point to the new merged area. */
1124 * |<-----Prev----->|<------VA------>|
1128 if (next->prev != head) {
1129 sibling = list_entry(next->prev, struct vmap_area, list);
1130 if (sibling->va_end == va->va_start) {
1132 * If both neighbors are coalesced, it is important
1133 * to unlink the "next" node first, followed by merging
1134 * with "previous" one. Otherwise the tree might not be
1135 * fully populated if a sibling's augmented value is
1136 * "normalized" because of rotation operations.
1139 unlink_va(va, root);
1141 sibling->va_end = va->va_end;
1143 /* Free vmap_area object. */
1144 kmem_cache_free(vmap_area_cachep, va);
1146 /* Point to the new merged area. */
1154 link_va(va, root, parent, link, head);
1159 static __always_inline struct vmap_area *
1160 merge_or_add_vmap_area_augment(struct vmap_area *va,
1161 struct rb_root *root, struct list_head *head)
1163 va = merge_or_add_vmap_area(va, root, head);
1165 augment_tree_propagate_from(va);
1170 static __always_inline bool
1171 is_within_this_va(struct vmap_area *va, unsigned long size,
1172 unsigned long align, unsigned long vstart)
1174 unsigned long nva_start_addr;
1176 if (va->va_start > vstart)
1177 nva_start_addr = ALIGN(va->va_start, align);
1179 nva_start_addr = ALIGN(vstart, align);
1181 /* Can be overflowed due to big size or alignment. */
1182 if (nva_start_addr + size < nva_start_addr ||
1183 nva_start_addr < vstart)
1186 return (nva_start_addr + size <= va->va_end);
1190 * Find the first free block(lowest start address) in the tree,
1191 * that will accomplish the request corresponding to passing
1194 static __always_inline struct vmap_area *
1195 find_vmap_lowest_match(unsigned long size,
1196 unsigned long align, unsigned long vstart)
1198 struct vmap_area *va;
1199 struct rb_node *node;
1201 /* Start from the root. */
1202 node = free_vmap_area_root.rb_node;
1205 va = rb_entry(node, struct vmap_area, rb_node);
1207 if (get_subtree_max_size(node->rb_left) >= size &&
1208 vstart < va->va_start) {
1209 node = node->rb_left;
1211 if (is_within_this_va(va, size, align, vstart))
1215 * Does not make sense to go deeper towards the right
1216 * sub-tree if it does not have a free block that is
1217 * equal or bigger to the requested search size.
1219 if (get_subtree_max_size(node->rb_right) >= size) {
1220 node = node->rb_right;
1225 * OK. We roll back and find the first right sub-tree,
1226 * that will satisfy the search criteria. It can happen
1227 * due to "vstart" restriction or an alignment overhead
1228 * that is bigger then PAGE_SIZE.
1230 while ((node = rb_parent(node))) {
1231 va = rb_entry(node, struct vmap_area, rb_node);
1232 if (is_within_this_va(va, size, align, vstart))
1235 if (get_subtree_max_size(node->rb_right) >= size &&
1236 vstart <= va->va_start) {
1238 * Shift the vstart forward. Please note, we update it with
1239 * parent's start address adding "1" because we do not want
1240 * to enter same sub-tree after it has already been checked
1241 * and no suitable free block found there.
1243 vstart = va->va_start + 1;
1244 node = node->rb_right;
1254 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1255 #include <linux/random.h>
1257 static struct vmap_area *
1258 find_vmap_lowest_linear_match(unsigned long size,
1259 unsigned long align, unsigned long vstart)
1261 struct vmap_area *va;
1263 list_for_each_entry(va, &free_vmap_area_list, list) {
1264 if (!is_within_this_va(va, size, align, vstart))
1274 find_vmap_lowest_match_check(unsigned long size, unsigned long align)
1276 struct vmap_area *va_1, *va_2;
1277 unsigned long vstart;
1280 get_random_bytes(&rnd, sizeof(rnd));
1281 vstart = VMALLOC_START + rnd;
1283 va_1 = find_vmap_lowest_match(size, align, vstart);
1284 va_2 = find_vmap_lowest_linear_match(size, align, vstart);
1287 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1288 va_1, va_2, vstart);
1294 FL_FIT_TYPE = 1, /* full fit */
1295 LE_FIT_TYPE = 2, /* left edge fit */
1296 RE_FIT_TYPE = 3, /* right edge fit */
1297 NE_FIT_TYPE = 4 /* no edge fit */
1300 static __always_inline enum fit_type
1301 classify_va_fit_type(struct vmap_area *va,
1302 unsigned long nva_start_addr, unsigned long size)
1306 /* Check if it is within VA. */
1307 if (nva_start_addr < va->va_start ||
1308 nva_start_addr + size > va->va_end)
1312 if (va->va_start == nva_start_addr) {
1313 if (va->va_end == nva_start_addr + size)
1317 } else if (va->va_end == nva_start_addr + size) {
1326 static __always_inline int
1327 adjust_va_to_fit_type(struct vmap_area *va,
1328 unsigned long nva_start_addr, unsigned long size,
1331 struct vmap_area *lva = NULL;
1333 if (type == FL_FIT_TYPE) {
1335 * No need to split VA, it fully fits.
1341 unlink_va(va, &free_vmap_area_root);
1342 kmem_cache_free(vmap_area_cachep, va);
1343 } else if (type == LE_FIT_TYPE) {
1345 * Split left edge of fit VA.
1351 va->va_start += size;
1352 } else if (type == RE_FIT_TYPE) {
1354 * Split right edge of fit VA.
1360 va->va_end = nva_start_addr;
1361 } else if (type == NE_FIT_TYPE) {
1363 * Split no edge of fit VA.
1369 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1370 if (unlikely(!lva)) {
1372 * For percpu allocator we do not do any pre-allocation
1373 * and leave it as it is. The reason is it most likely
1374 * never ends up with NE_FIT_TYPE splitting. In case of
1375 * percpu allocations offsets and sizes are aligned to
1376 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1377 * are its main fitting cases.
1379 * There are a few exceptions though, as an example it is
1380 * a first allocation (early boot up) when we have "one"
1381 * big free space that has to be split.
1383 * Also we can hit this path in case of regular "vmap"
1384 * allocations, if "this" current CPU was not preloaded.
1385 * See the comment in alloc_vmap_area() why. If so, then
1386 * GFP_NOWAIT is used instead to get an extra object for
1387 * split purpose. That is rare and most time does not
1390 * What happens if an allocation gets failed. Basically,
1391 * an "overflow" path is triggered to purge lazily freed
1392 * areas to free some memory, then, the "retry" path is
1393 * triggered to repeat one more time. See more details
1394 * in alloc_vmap_area() function.
1396 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1402 * Build the remainder.
1404 lva->va_start = va->va_start;
1405 lva->va_end = nva_start_addr;
1408 * Shrink this VA to remaining size.
1410 va->va_start = nva_start_addr + size;
1415 if (type != FL_FIT_TYPE) {
1416 augment_tree_propagate_from(va);
1418 if (lva) /* type == NE_FIT_TYPE */
1419 insert_vmap_area_augment(lva, &va->rb_node,
1420 &free_vmap_area_root, &free_vmap_area_list);
1427 * Returns a start address of the newly allocated area, if success.
1428 * Otherwise a vend is returned that indicates failure.
1430 static __always_inline unsigned long
1431 __alloc_vmap_area(unsigned long size, unsigned long align,
1432 unsigned long vstart, unsigned long vend)
1434 unsigned long nva_start_addr;
1435 struct vmap_area *va;
1439 va = find_vmap_lowest_match(size, align, vstart);
1443 if (va->va_start > vstart)
1444 nva_start_addr = ALIGN(va->va_start, align);
1446 nva_start_addr = ALIGN(vstart, align);
1448 /* Check the "vend" restriction. */
1449 if (nva_start_addr + size > vend)
1452 /* Classify what we have found. */
1453 type = classify_va_fit_type(va, nva_start_addr, size);
1454 if (WARN_ON_ONCE(type == NOTHING_FIT))
1457 /* Update the free vmap_area. */
1458 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1462 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1463 find_vmap_lowest_match_check(size, align);
1466 return nva_start_addr;
1470 * Free a region of KVA allocated by alloc_vmap_area
1472 static void free_vmap_area(struct vmap_area *va)
1475 * Remove from the busy tree/list.
1477 spin_lock(&vmap_area_lock);
1478 unlink_va(va, &vmap_area_root);
1479 spin_unlock(&vmap_area_lock);
1482 * Insert/Merge it back to the free tree/list.
1484 spin_lock(&free_vmap_area_lock);
1485 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1486 spin_unlock(&free_vmap_area_lock);
1490 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1492 struct vmap_area *va = NULL;
1495 * Preload this CPU with one extra vmap_area object. It is used
1496 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1497 * a CPU that does an allocation is preloaded.
1499 * We do it in non-atomic context, thus it allows us to use more
1500 * permissive allocation masks to be more stable under low memory
1501 * condition and high memory pressure.
1503 if (!this_cpu_read(ne_fit_preload_node))
1504 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1508 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1509 kmem_cache_free(vmap_area_cachep, va);
1513 * Allocate a region of KVA of the specified size and alignment, within the
1516 static struct vmap_area *alloc_vmap_area(unsigned long size,
1517 unsigned long align,
1518 unsigned long vstart, unsigned long vend,
1519 int node, gfp_t gfp_mask)
1521 struct vmap_area *va;
1522 unsigned long freed;
1528 BUG_ON(offset_in_page(size));
1529 BUG_ON(!is_power_of_2(align));
1531 if (unlikely(!vmap_initialized))
1532 return ERR_PTR(-EBUSY);
1535 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1537 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1539 return ERR_PTR(-ENOMEM);
1542 * Only scan the relevant parts containing pointers to other objects
1543 * to avoid false negatives.
1545 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1548 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1549 addr = __alloc_vmap_area(size, align, vstart, vend);
1550 spin_unlock(&free_vmap_area_lock);
1553 * If an allocation fails, the "vend" address is
1554 * returned. Therefore trigger the overflow path.
1556 if (unlikely(addr == vend))
1559 va->va_start = addr;
1560 va->va_end = addr + size;
1563 spin_lock(&vmap_area_lock);
1564 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1565 spin_unlock(&vmap_area_lock);
1567 BUG_ON(!IS_ALIGNED(va->va_start, align));
1568 BUG_ON(va->va_start < vstart);
1569 BUG_ON(va->va_end > vend);
1571 ret = kasan_populate_vmalloc(addr, size);
1574 return ERR_PTR(ret);
1581 purge_vmap_area_lazy();
1587 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1594 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1595 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1598 kmem_cache_free(vmap_area_cachep, va);
1599 return ERR_PTR(-EBUSY);
1602 int register_vmap_purge_notifier(struct notifier_block *nb)
1604 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1606 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1608 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1610 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1612 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1615 * lazy_max_pages is the maximum amount of virtual address space we gather up
1616 * before attempting to purge with a TLB flush.
1618 * There is a tradeoff here: a larger number will cover more kernel page tables
1619 * and take slightly longer to purge, but it will linearly reduce the number of
1620 * global TLB flushes that must be performed. It would seem natural to scale
1621 * this number up linearly with the number of CPUs (because vmapping activity
1622 * could also scale linearly with the number of CPUs), however it is likely
1623 * that in practice, workloads might be constrained in other ways that mean
1624 * vmap activity will not scale linearly with CPUs. Also, I want to be
1625 * conservative and not introduce a big latency on huge systems, so go with
1626 * a less aggressive log scale. It will still be an improvement over the old
1627 * code, and it will be simple to change the scale factor if we find that it
1628 * becomes a problem on bigger systems.
1630 static unsigned long lazy_max_pages(void)
1634 log = fls(num_online_cpus());
1636 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1639 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1642 * Serialize vmap purging. There is no actual critical section protected
1643 * by this look, but we want to avoid concurrent calls for performance
1644 * reasons and to make the pcpu_get_vm_areas more deterministic.
1646 static DEFINE_MUTEX(vmap_purge_lock);
1648 /* for per-CPU blocks */
1649 static void purge_fragmented_blocks_allcpus(void);
1651 #ifdef CONFIG_X86_64
1653 * called before a call to iounmap() if the caller wants vm_area_struct's
1654 * immediately freed.
1656 void set_iounmap_nonlazy(void)
1658 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1660 #endif /* CONFIG_X86_64 */
1663 * Purges all lazily-freed vmap areas.
1665 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1667 unsigned long resched_threshold;
1668 struct list_head local_pure_list;
1669 struct vmap_area *va, *n_va;
1671 lockdep_assert_held(&vmap_purge_lock);
1673 spin_lock(&purge_vmap_area_lock);
1674 purge_vmap_area_root = RB_ROOT;
1675 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1676 spin_unlock(&purge_vmap_area_lock);
1678 if (unlikely(list_empty(&local_pure_list)))
1682 list_first_entry(&local_pure_list,
1683 struct vmap_area, list)->va_start);
1686 list_last_entry(&local_pure_list,
1687 struct vmap_area, list)->va_end);
1689 flush_tlb_kernel_range(start, end);
1690 resched_threshold = lazy_max_pages() << 1;
1692 spin_lock(&free_vmap_area_lock);
1693 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1694 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1695 unsigned long orig_start = va->va_start;
1696 unsigned long orig_end = va->va_end;
1699 * Finally insert or merge lazily-freed area. It is
1700 * detached and there is no need to "unlink" it from
1703 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1704 &free_vmap_area_list);
1709 if (is_vmalloc_or_module_addr((void *)orig_start))
1710 kasan_release_vmalloc(orig_start, orig_end,
1711 va->va_start, va->va_end);
1713 atomic_long_sub(nr, &vmap_lazy_nr);
1715 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1716 cond_resched_lock(&free_vmap_area_lock);
1718 spin_unlock(&free_vmap_area_lock);
1723 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1724 * is already purging.
1726 static void try_purge_vmap_area_lazy(void)
1728 if (mutex_trylock(&vmap_purge_lock)) {
1729 __purge_vmap_area_lazy(ULONG_MAX, 0);
1730 mutex_unlock(&vmap_purge_lock);
1735 * Kick off a purge of the outstanding lazy areas.
1737 static void purge_vmap_area_lazy(void)
1739 mutex_lock(&vmap_purge_lock);
1740 purge_fragmented_blocks_allcpus();
1741 __purge_vmap_area_lazy(ULONG_MAX, 0);
1742 mutex_unlock(&vmap_purge_lock);
1746 * Free a vmap area, caller ensuring that the area has been unmapped
1747 * and flush_cache_vunmap had been called for the correct range
1750 static void free_vmap_area_noflush(struct vmap_area *va)
1752 unsigned long nr_lazy;
1754 spin_lock(&vmap_area_lock);
1755 unlink_va(va, &vmap_area_root);
1756 spin_unlock(&vmap_area_lock);
1758 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1759 PAGE_SHIFT, &vmap_lazy_nr);
1762 * Merge or place it to the purge tree/list.
1764 spin_lock(&purge_vmap_area_lock);
1765 merge_or_add_vmap_area(va,
1766 &purge_vmap_area_root, &purge_vmap_area_list);
1767 spin_unlock(&purge_vmap_area_lock);
1769 /* After this point, we may free va at any time */
1770 if (unlikely(nr_lazy > lazy_max_pages()))
1771 try_purge_vmap_area_lazy();
1775 * Free and unmap a vmap area
1777 static void free_unmap_vmap_area(struct vmap_area *va)
1779 flush_cache_vunmap(va->va_start, va->va_end);
1780 vunmap_range_noflush(va->va_start, va->va_end);
1781 if (debug_pagealloc_enabled_static())
1782 flush_tlb_kernel_range(va->va_start, va->va_end);
1784 free_vmap_area_noflush(va);
1787 static struct vmap_area *find_vmap_area(unsigned long addr)
1789 struct vmap_area *va;
1791 spin_lock(&vmap_area_lock);
1792 va = __find_vmap_area(addr);
1793 spin_unlock(&vmap_area_lock);
1798 /*** Per cpu kva allocator ***/
1801 * vmap space is limited especially on 32 bit architectures. Ensure there is
1802 * room for at least 16 percpu vmap blocks per CPU.
1805 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1806 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1807 * instead (we just need a rough idea)
1809 #if BITS_PER_LONG == 32
1810 #define VMALLOC_SPACE (128UL*1024*1024)
1812 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1815 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1816 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1817 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1818 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1819 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1820 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1821 #define VMAP_BBMAP_BITS \
1822 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1823 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1824 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1826 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1828 struct vmap_block_queue {
1830 struct list_head free;
1835 struct vmap_area *va;
1836 unsigned long free, dirty;
1837 unsigned long dirty_min, dirty_max; /*< dirty range */
1838 struct list_head free_list;
1839 struct rcu_head rcu_head;
1840 struct list_head purge;
1843 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1844 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1847 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1848 * in the free path. Could get rid of this if we change the API to return a
1849 * "cookie" from alloc, to be passed to free. But no big deal yet.
1851 static DEFINE_XARRAY(vmap_blocks);
1854 * We should probably have a fallback mechanism to allocate virtual memory
1855 * out of partially filled vmap blocks. However vmap block sizing should be
1856 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1860 static unsigned long addr_to_vb_idx(unsigned long addr)
1862 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1863 addr /= VMAP_BLOCK_SIZE;
1867 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1871 addr = va_start + (pages_off << PAGE_SHIFT);
1872 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1873 return (void *)addr;
1877 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1878 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1879 * @order: how many 2^order pages should be occupied in newly allocated block
1880 * @gfp_mask: flags for the page level allocator
1882 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1884 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1886 struct vmap_block_queue *vbq;
1887 struct vmap_block *vb;
1888 struct vmap_area *va;
1889 unsigned long vb_idx;
1893 node = numa_node_id();
1895 vb = kmalloc_node(sizeof(struct vmap_block),
1896 gfp_mask & GFP_RECLAIM_MASK, node);
1898 return ERR_PTR(-ENOMEM);
1900 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1901 VMALLOC_START, VMALLOC_END,
1905 return ERR_CAST(va);
1908 vaddr = vmap_block_vaddr(va->va_start, 0);
1909 spin_lock_init(&vb->lock);
1911 /* At least something should be left free */
1912 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1913 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1915 vb->dirty_min = VMAP_BBMAP_BITS;
1917 INIT_LIST_HEAD(&vb->free_list);
1919 vb_idx = addr_to_vb_idx(va->va_start);
1920 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1924 return ERR_PTR(err);
1927 vbq = &get_cpu_var(vmap_block_queue);
1928 spin_lock(&vbq->lock);
1929 list_add_tail_rcu(&vb->free_list, &vbq->free);
1930 spin_unlock(&vbq->lock);
1931 put_cpu_var(vmap_block_queue);
1936 static void free_vmap_block(struct vmap_block *vb)
1938 struct vmap_block *tmp;
1940 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1943 free_vmap_area_noflush(vb->va);
1944 kfree_rcu(vb, rcu_head);
1947 static void purge_fragmented_blocks(int cpu)
1950 struct vmap_block *vb;
1951 struct vmap_block *n_vb;
1952 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1955 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1957 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1960 spin_lock(&vb->lock);
1961 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1962 vb->free = 0; /* prevent further allocs after releasing lock */
1963 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1965 vb->dirty_max = VMAP_BBMAP_BITS;
1966 spin_lock(&vbq->lock);
1967 list_del_rcu(&vb->free_list);
1968 spin_unlock(&vbq->lock);
1969 spin_unlock(&vb->lock);
1970 list_add_tail(&vb->purge, &purge);
1972 spin_unlock(&vb->lock);
1976 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1977 list_del(&vb->purge);
1978 free_vmap_block(vb);
1982 static void purge_fragmented_blocks_allcpus(void)
1986 for_each_possible_cpu(cpu)
1987 purge_fragmented_blocks(cpu);
1990 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1992 struct vmap_block_queue *vbq;
1993 struct vmap_block *vb;
1997 BUG_ON(offset_in_page(size));
1998 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1999 if (WARN_ON(size == 0)) {
2001 * Allocating 0 bytes isn't what caller wants since
2002 * get_order(0) returns funny result. Just warn and terminate
2007 order = get_order(size);
2010 vbq = &get_cpu_var(vmap_block_queue);
2011 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2012 unsigned long pages_off;
2014 spin_lock(&vb->lock);
2015 if (vb->free < (1UL << order)) {
2016 spin_unlock(&vb->lock);
2020 pages_off = VMAP_BBMAP_BITS - vb->free;
2021 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2022 vb->free -= 1UL << order;
2023 if (vb->free == 0) {
2024 spin_lock(&vbq->lock);
2025 list_del_rcu(&vb->free_list);
2026 spin_unlock(&vbq->lock);
2029 spin_unlock(&vb->lock);
2033 put_cpu_var(vmap_block_queue);
2036 /* Allocate new block if nothing was found */
2038 vaddr = new_vmap_block(order, gfp_mask);
2043 static void vb_free(unsigned long addr, unsigned long size)
2045 unsigned long offset;
2047 struct vmap_block *vb;
2049 BUG_ON(offset_in_page(size));
2050 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2052 flush_cache_vunmap(addr, addr + size);
2054 order = get_order(size);
2055 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2056 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2058 vunmap_range_noflush(addr, addr + size);
2060 if (debug_pagealloc_enabled_static())
2061 flush_tlb_kernel_range(addr, addr + size);
2063 spin_lock(&vb->lock);
2065 /* Expand dirty range */
2066 vb->dirty_min = min(vb->dirty_min, offset);
2067 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2069 vb->dirty += 1UL << order;
2070 if (vb->dirty == VMAP_BBMAP_BITS) {
2072 spin_unlock(&vb->lock);
2073 free_vmap_block(vb);
2075 spin_unlock(&vb->lock);
2078 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2082 if (unlikely(!vmap_initialized))
2087 for_each_possible_cpu(cpu) {
2088 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2089 struct vmap_block *vb;
2092 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2093 spin_lock(&vb->lock);
2094 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2095 unsigned long va_start = vb->va->va_start;
2098 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2099 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2101 start = min(s, start);
2106 spin_unlock(&vb->lock);
2111 mutex_lock(&vmap_purge_lock);
2112 purge_fragmented_blocks_allcpus();
2113 if (!__purge_vmap_area_lazy(start, end) && flush)
2114 flush_tlb_kernel_range(start, end);
2115 mutex_unlock(&vmap_purge_lock);
2119 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2121 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2122 * to amortize TLB flushing overheads. What this means is that any page you
2123 * have now, may, in a former life, have been mapped into kernel virtual
2124 * address by the vmap layer and so there might be some CPUs with TLB entries
2125 * still referencing that page (additional to the regular 1:1 kernel mapping).
2127 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2128 * be sure that none of the pages we have control over will have any aliases
2129 * from the vmap layer.
2131 void vm_unmap_aliases(void)
2133 unsigned long start = ULONG_MAX, end = 0;
2136 _vm_unmap_aliases(start, end, flush);
2138 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2141 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2142 * @mem: the pointer returned by vm_map_ram
2143 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2145 void vm_unmap_ram(const void *mem, unsigned int count)
2147 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2148 unsigned long addr = (unsigned long)mem;
2149 struct vmap_area *va;
2153 BUG_ON(addr < VMALLOC_START);
2154 BUG_ON(addr > VMALLOC_END);
2155 BUG_ON(!PAGE_ALIGNED(addr));
2157 kasan_poison_vmalloc(mem, size);
2159 if (likely(count <= VMAP_MAX_ALLOC)) {
2160 debug_check_no_locks_freed(mem, size);
2161 vb_free(addr, size);
2165 va = find_vmap_area(addr);
2167 debug_check_no_locks_freed((void *)va->va_start,
2168 (va->va_end - va->va_start));
2169 free_unmap_vmap_area(va);
2171 EXPORT_SYMBOL(vm_unmap_ram);
2174 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2175 * @pages: an array of pointers to the pages to be mapped
2176 * @count: number of pages
2177 * @node: prefer to allocate data structures on this node
2179 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2180 * faster than vmap so it's good. But if you mix long-life and short-life
2181 * objects with vm_map_ram(), it could consume lots of address space through
2182 * fragmentation (especially on a 32bit machine). You could see failures in
2183 * the end. Please use this function for short-lived objects.
2185 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2187 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2189 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2193 if (likely(count <= VMAP_MAX_ALLOC)) {
2194 mem = vb_alloc(size, GFP_KERNEL);
2197 addr = (unsigned long)mem;
2199 struct vmap_area *va;
2200 va = alloc_vmap_area(size, PAGE_SIZE,
2201 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2205 addr = va->va_start;
2209 kasan_unpoison_vmalloc(mem, size);
2211 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2212 pages, PAGE_SHIFT) < 0) {
2213 vm_unmap_ram(mem, count);
2219 EXPORT_SYMBOL(vm_map_ram);
2221 static struct vm_struct *vmlist __initdata;
2223 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2225 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2226 return vm->page_order;
2232 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2234 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2235 vm->page_order = order;
2242 * vm_area_add_early - add vmap area early during boot
2243 * @vm: vm_struct to add
2245 * This function is used to add fixed kernel vm area to vmlist before
2246 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2247 * should contain proper values and the other fields should be zero.
2249 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2251 void __init vm_area_add_early(struct vm_struct *vm)
2253 struct vm_struct *tmp, **p;
2255 BUG_ON(vmap_initialized);
2256 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2257 if (tmp->addr >= vm->addr) {
2258 BUG_ON(tmp->addr < vm->addr + vm->size);
2261 BUG_ON(tmp->addr + tmp->size > vm->addr);
2268 * vm_area_register_early - register vmap area early during boot
2269 * @vm: vm_struct to register
2270 * @align: requested alignment
2272 * This function is used to register kernel vm area before
2273 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2274 * proper values on entry and other fields should be zero. On return,
2275 * vm->addr contains the allocated address.
2277 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2279 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2281 unsigned long addr = ALIGN(VMALLOC_START, align);
2282 struct vm_struct *cur, **p;
2284 BUG_ON(vmap_initialized);
2286 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2287 if ((unsigned long)cur->addr - addr >= vm->size)
2289 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2292 BUG_ON(addr > VMALLOC_END - vm->size);
2293 vm->addr = (void *)addr;
2296 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2299 static void vmap_init_free_space(void)
2301 unsigned long vmap_start = 1;
2302 const unsigned long vmap_end = ULONG_MAX;
2303 struct vmap_area *busy, *free;
2307 * -|-----|.....|-----|-----|-----|.....|-
2309 * |<--------------------------------->|
2311 list_for_each_entry(busy, &vmap_area_list, list) {
2312 if (busy->va_start - vmap_start > 0) {
2313 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2314 if (!WARN_ON_ONCE(!free)) {
2315 free->va_start = vmap_start;
2316 free->va_end = busy->va_start;
2318 insert_vmap_area_augment(free, NULL,
2319 &free_vmap_area_root,
2320 &free_vmap_area_list);
2324 vmap_start = busy->va_end;
2327 if (vmap_end - vmap_start > 0) {
2328 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2329 if (!WARN_ON_ONCE(!free)) {
2330 free->va_start = vmap_start;
2331 free->va_end = vmap_end;
2333 insert_vmap_area_augment(free, NULL,
2334 &free_vmap_area_root,
2335 &free_vmap_area_list);
2340 void __init vmalloc_init(void)
2342 struct vmap_area *va;
2343 struct vm_struct *tmp;
2347 * Create the cache for vmap_area objects.
2349 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2351 for_each_possible_cpu(i) {
2352 struct vmap_block_queue *vbq;
2353 struct vfree_deferred *p;
2355 vbq = &per_cpu(vmap_block_queue, i);
2356 spin_lock_init(&vbq->lock);
2357 INIT_LIST_HEAD(&vbq->free);
2358 p = &per_cpu(vfree_deferred, i);
2359 init_llist_head(&p->list);
2360 INIT_WORK(&p->wq, free_work);
2363 /* Import existing vmlist entries. */
2364 for (tmp = vmlist; tmp; tmp = tmp->next) {
2365 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2366 if (WARN_ON_ONCE(!va))
2369 va->va_start = (unsigned long)tmp->addr;
2370 va->va_end = va->va_start + tmp->size;
2372 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2376 * Now we can initialize a free vmap space.
2378 vmap_init_free_space();
2379 vmap_initialized = true;
2382 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2383 struct vmap_area *va, unsigned long flags, const void *caller)
2386 vm->addr = (void *)va->va_start;
2387 vm->size = va->va_end - va->va_start;
2388 vm->caller = caller;
2392 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2393 unsigned long flags, const void *caller)
2395 spin_lock(&vmap_area_lock);
2396 setup_vmalloc_vm_locked(vm, va, flags, caller);
2397 spin_unlock(&vmap_area_lock);
2400 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2403 * Before removing VM_UNINITIALIZED,
2404 * we should make sure that vm has proper values.
2405 * Pair with smp_rmb() in show_numa_info().
2408 vm->flags &= ~VM_UNINITIALIZED;
2411 static struct vm_struct *__get_vm_area_node(unsigned long size,
2412 unsigned long align, unsigned long shift, unsigned long flags,
2413 unsigned long start, unsigned long end, int node,
2414 gfp_t gfp_mask, const void *caller)
2416 struct vmap_area *va;
2417 struct vm_struct *area;
2418 unsigned long requested_size = size;
2420 BUG_ON(in_interrupt());
2421 size = ALIGN(size, 1ul << shift);
2422 if (unlikely(!size))
2425 if (flags & VM_IOREMAP)
2426 align = 1ul << clamp_t(int, get_count_order_long(size),
2427 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2429 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2430 if (unlikely(!area))
2433 if (!(flags & VM_NO_GUARD))
2436 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2442 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2444 setup_vmalloc_vm(area, va, flags, caller);
2449 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2450 unsigned long start, unsigned long end,
2453 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2454 NUMA_NO_NODE, GFP_KERNEL, caller);
2458 * get_vm_area - reserve a contiguous kernel virtual area
2459 * @size: size of the area
2460 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2462 * Search an area of @size in the kernel virtual mapping area,
2463 * and reserved it for out purposes. Returns the area descriptor
2464 * on success or %NULL on failure.
2466 * Return: the area descriptor on success or %NULL on failure.
2468 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2470 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2471 VMALLOC_START, VMALLOC_END,
2472 NUMA_NO_NODE, GFP_KERNEL,
2473 __builtin_return_address(0));
2476 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2479 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2480 VMALLOC_START, VMALLOC_END,
2481 NUMA_NO_NODE, GFP_KERNEL, caller);
2485 * find_vm_area - find a continuous kernel virtual area
2486 * @addr: base address
2488 * Search for the kernel VM area starting at @addr, and return it.
2489 * It is up to the caller to do all required locking to keep the returned
2492 * Return: the area descriptor on success or %NULL on failure.
2494 struct vm_struct *find_vm_area(const void *addr)
2496 struct vmap_area *va;
2498 va = find_vmap_area((unsigned long)addr);
2506 * remove_vm_area - find and remove a continuous kernel virtual area
2507 * @addr: base address
2509 * Search for the kernel VM area starting at @addr, and remove it.
2510 * This function returns the found VM area, but using it is NOT safe
2511 * on SMP machines, except for its size or flags.
2513 * Return: the area descriptor on success or %NULL on failure.
2515 struct vm_struct *remove_vm_area(const void *addr)
2517 struct vmap_area *va;
2521 spin_lock(&vmap_area_lock);
2522 va = __find_vmap_area((unsigned long)addr);
2524 struct vm_struct *vm = va->vm;
2527 spin_unlock(&vmap_area_lock);
2529 kasan_free_shadow(vm);
2530 free_unmap_vmap_area(va);
2535 spin_unlock(&vmap_area_lock);
2539 static inline void set_area_direct_map(const struct vm_struct *area,
2540 int (*set_direct_map)(struct page *page))
2544 /* HUGE_VMALLOC passes small pages to set_direct_map */
2545 for (i = 0; i < area->nr_pages; i++)
2546 if (page_address(area->pages[i]))
2547 set_direct_map(area->pages[i]);
2550 /* Handle removing and resetting vm mappings related to the vm_struct. */
2551 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2553 unsigned long start = ULONG_MAX, end = 0;
2554 unsigned int page_order = vm_area_page_order(area);
2555 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2559 remove_vm_area(area->addr);
2561 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2566 * If not deallocating pages, just do the flush of the VM area and
2569 if (!deallocate_pages) {
2575 * If execution gets here, flush the vm mapping and reset the direct
2576 * map. Find the start and end range of the direct mappings to make sure
2577 * the vm_unmap_aliases() flush includes the direct map.
2579 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2580 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2582 unsigned long page_size;
2584 page_size = PAGE_SIZE << page_order;
2585 start = min(addr, start);
2586 end = max(addr + page_size, end);
2592 * Set direct map to something invalid so that it won't be cached if
2593 * there are any accesses after the TLB flush, then flush the TLB and
2594 * reset the direct map permissions to the default.
2596 set_area_direct_map(area, set_direct_map_invalid_noflush);
2597 _vm_unmap_aliases(start, end, flush_dmap);
2598 set_area_direct_map(area, set_direct_map_default_noflush);
2601 static void __vunmap(const void *addr, int deallocate_pages)
2603 struct vm_struct *area;
2608 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2612 area = find_vm_area(addr);
2613 if (unlikely(!area)) {
2614 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2619 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2620 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2622 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2624 vm_remove_mappings(area, deallocate_pages);
2626 if (deallocate_pages) {
2627 unsigned int page_order = vm_area_page_order(area);
2628 int i, step = 1U << page_order;
2630 for (i = 0; i < area->nr_pages; i += step) {
2631 struct page *page = area->pages[i];
2634 mod_memcg_page_state(page, MEMCG_VMALLOC, -step);
2635 __free_pages(page, page_order);
2638 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2640 kvfree(area->pages);
2646 static inline void __vfree_deferred(const void *addr)
2649 * Use raw_cpu_ptr() because this can be called from preemptible
2650 * context. Preemption is absolutely fine here, because the llist_add()
2651 * implementation is lockless, so it works even if we are adding to
2652 * another cpu's list. schedule_work() should be fine with this too.
2654 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2656 if (llist_add((struct llist_node *)addr, &p->list))
2657 schedule_work(&p->wq);
2661 * vfree_atomic - release memory allocated by vmalloc()
2662 * @addr: memory base address
2664 * This one is just like vfree() but can be called in any atomic context
2667 void vfree_atomic(const void *addr)
2671 kmemleak_free(addr);
2675 __vfree_deferred(addr);
2678 static void __vfree(const void *addr)
2680 if (unlikely(in_interrupt()))
2681 __vfree_deferred(addr);
2687 * vfree - Release memory allocated by vmalloc()
2688 * @addr: Memory base address
2690 * Free the virtually continuous memory area starting at @addr, as obtained
2691 * from one of the vmalloc() family of APIs. This will usually also free the
2692 * physical memory underlying the virtual allocation, but that memory is
2693 * reference counted, so it will not be freed until the last user goes away.
2695 * If @addr is NULL, no operation is performed.
2698 * May sleep if called *not* from interrupt context.
2699 * Must not be called in NMI context (strictly speaking, it could be
2700 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2701 * conventions for vfree() arch-dependent would be a really bad idea).
2703 void vfree(const void *addr)
2707 kmemleak_free(addr);
2709 might_sleep_if(!in_interrupt());
2716 EXPORT_SYMBOL(vfree);
2719 * vunmap - release virtual mapping obtained by vmap()
2720 * @addr: memory base address
2722 * Free the virtually contiguous memory area starting at @addr,
2723 * which was created from the page array passed to vmap().
2725 * Must not be called in interrupt context.
2727 void vunmap(const void *addr)
2729 BUG_ON(in_interrupt());
2734 EXPORT_SYMBOL(vunmap);
2737 * vmap - map an array of pages into virtually contiguous space
2738 * @pages: array of page pointers
2739 * @count: number of pages to map
2740 * @flags: vm_area->flags
2741 * @prot: page protection for the mapping
2743 * Maps @count pages from @pages into contiguous kernel virtual space.
2744 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2745 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2746 * are transferred from the caller to vmap(), and will be freed / dropped when
2747 * vfree() is called on the return value.
2749 * Return: the address of the area or %NULL on failure
2751 void *vmap(struct page **pages, unsigned int count,
2752 unsigned long flags, pgprot_t prot)
2754 struct vm_struct *area;
2756 unsigned long size; /* In bytes */
2761 * Your top guard is someone else's bottom guard. Not having a top
2762 * guard compromises someone else's mappings too.
2764 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2765 flags &= ~VM_NO_GUARD;
2767 if (count > totalram_pages())
2770 size = (unsigned long)count << PAGE_SHIFT;
2771 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2775 addr = (unsigned long)area->addr;
2776 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2777 pages, PAGE_SHIFT) < 0) {
2782 if (flags & VM_MAP_PUT_PAGES) {
2783 area->pages = pages;
2784 area->nr_pages = count;
2788 EXPORT_SYMBOL(vmap);
2790 #ifdef CONFIG_VMAP_PFN
2791 struct vmap_pfn_data {
2792 unsigned long *pfns;
2797 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2799 struct vmap_pfn_data *data = private;
2801 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2803 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2808 * vmap_pfn - map an array of PFNs into virtually contiguous space
2809 * @pfns: array of PFNs
2810 * @count: number of pages to map
2811 * @prot: page protection for the mapping
2813 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2814 * the start address of the mapping.
2816 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2818 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2819 struct vm_struct *area;
2821 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2822 __builtin_return_address(0));
2825 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2826 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2832 EXPORT_SYMBOL_GPL(vmap_pfn);
2833 #endif /* CONFIG_VMAP_PFN */
2835 static inline unsigned int
2836 vm_area_alloc_pages(gfp_t gfp, int nid,
2837 unsigned int order, unsigned int nr_pages, struct page **pages)
2839 unsigned int nr_allocated = 0;
2844 * For order-0 pages we make use of bulk allocator, if
2845 * the page array is partly or not at all populated due
2846 * to fails, fallback to a single page allocator that is
2850 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2852 while (nr_allocated < nr_pages) {
2853 unsigned int nr, nr_pages_request;
2856 * A maximum allowed request is hard-coded and is 100
2857 * pages per call. That is done in order to prevent a
2858 * long preemption off scenario in the bulk-allocator
2859 * so the range is [1:100].
2861 nr_pages_request = min(100U, nr_pages - nr_allocated);
2863 /* memory allocation should consider mempolicy, we can't
2864 * wrongly use nearest node when nid == NUMA_NO_NODE,
2865 * otherwise memory may be allocated in only one node,
2866 * but mempolcy want to alloc memory by interleaving.
2868 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2869 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2871 pages + nr_allocated);
2874 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2876 pages + nr_allocated);
2882 * If zero or pages were obtained partly,
2883 * fallback to a single page allocator.
2885 if (nr != nr_pages_request)
2890 * Compound pages required for remap_vmalloc_page if
2895 /* High-order pages or fallback path if "bulk" fails. */
2897 while (nr_allocated < nr_pages) {
2898 if (fatal_signal_pending(current))
2901 if (nid == NUMA_NO_NODE)
2902 page = alloc_pages(gfp, order);
2904 page = alloc_pages_node(nid, gfp, order);
2905 if (unlikely(!page))
2909 * Careful, we allocate and map page-order pages, but
2910 * tracking is done per PAGE_SIZE page so as to keep the
2911 * vm_struct APIs independent of the physical/mapped size.
2913 for (i = 0; i < (1U << order); i++)
2914 pages[nr_allocated + i] = page + i;
2917 nr_allocated += 1U << order;
2920 return nr_allocated;
2923 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2924 pgprot_t prot, unsigned int page_shift,
2927 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2928 const gfp_t orig_gfp_mask = gfp_mask;
2929 bool nofail = gfp_mask & __GFP_NOFAIL;
2930 unsigned long addr = (unsigned long)area->addr;
2931 unsigned long size = get_vm_area_size(area);
2932 unsigned long array_size;
2933 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2934 unsigned int page_order;
2938 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2939 gfp_mask |= __GFP_NOWARN;
2940 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2941 gfp_mask |= __GFP_HIGHMEM;
2943 /* Please note that the recursion is strictly bounded. */
2944 if (array_size > PAGE_SIZE) {
2945 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2948 area->pages = kmalloc_node(array_size, nested_gfp, node);
2952 warn_alloc(orig_gfp_mask, NULL,
2953 "vmalloc error: size %lu, failed to allocated page array size %lu",
2954 nr_small_pages * PAGE_SIZE, array_size);
2959 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2960 page_order = vm_area_page_order(area);
2962 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2963 page_order, nr_small_pages, area->pages);
2965 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2966 if (gfp_mask & __GFP_ACCOUNT) {
2967 int i, step = 1U << page_order;
2969 for (i = 0; i < area->nr_pages; i += step)
2970 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC,
2975 * If not enough pages were obtained to accomplish an
2976 * allocation request, free them via __vfree() if any.
2978 if (area->nr_pages != nr_small_pages) {
2979 warn_alloc(orig_gfp_mask, NULL,
2980 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2981 area->nr_pages * PAGE_SIZE, page_order);
2986 * page tables allocations ignore external gfp mask, enforce it
2989 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
2990 flags = memalloc_nofs_save();
2991 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
2992 flags = memalloc_noio_save();
2995 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
2997 if (nofail && (ret < 0))
2998 schedule_timeout_uninterruptible(1);
2999 } while (nofail && (ret < 0));
3001 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3002 memalloc_nofs_restore(flags);
3003 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3004 memalloc_noio_restore(flags);
3007 warn_alloc(orig_gfp_mask, NULL,
3008 "vmalloc error: size %lu, failed to map pages",
3009 area->nr_pages * PAGE_SIZE);
3016 __vfree(area->addr);
3021 * __vmalloc_node_range - allocate virtually contiguous memory
3022 * @size: allocation size
3023 * @align: desired alignment
3024 * @start: vm area range start
3025 * @end: vm area range end
3026 * @gfp_mask: flags for the page level allocator
3027 * @prot: protection mask for the allocated pages
3028 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3029 * @node: node to use for allocation or NUMA_NO_NODE
3030 * @caller: caller's return address
3032 * Allocate enough pages to cover @size from the page level
3033 * allocator with @gfp_mask flags. Please note that the full set of gfp
3034 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3036 * Zone modifiers are not supported. From the reclaim modifiers
3037 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3038 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3039 * __GFP_RETRY_MAYFAIL are not supported).
3041 * __GFP_NOWARN can be used to suppress failures messages.
3043 * Map them into contiguous kernel virtual space, using a pagetable
3044 * protection of @prot.
3046 * Return: the address of the area or %NULL on failure
3048 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3049 unsigned long start, unsigned long end, gfp_t gfp_mask,
3050 pgprot_t prot, unsigned long vm_flags, int node,
3053 struct vm_struct *area;
3055 unsigned long real_size = size;
3056 unsigned long real_align = align;
3057 unsigned int shift = PAGE_SHIFT;
3059 if (WARN_ON_ONCE(!size))
3062 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3063 warn_alloc(gfp_mask, NULL,
3064 "vmalloc error: size %lu, exceeds total pages",
3069 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
3070 unsigned long size_per_node;
3073 * Try huge pages. Only try for PAGE_KERNEL allocations,
3074 * others like modules don't yet expect huge pages in
3075 * their allocations due to apply_to_page_range not
3079 size_per_node = size;
3080 if (node == NUMA_NO_NODE)
3081 size_per_node /= num_online_nodes();
3082 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3085 shift = arch_vmap_pte_supported_shift(size_per_node);
3087 align = max(real_align, 1UL << shift);
3088 size = ALIGN(real_size, 1UL << shift);
3092 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3093 VM_UNINITIALIZED | vm_flags, start, end, node,
3096 bool nofail = gfp_mask & __GFP_NOFAIL;
3097 warn_alloc(gfp_mask, NULL,
3098 "vmalloc error: size %lu, vm_struct allocation failed%s",
3099 real_size, (nofail) ? ". Retrying." : "");
3101 schedule_timeout_uninterruptible(1);
3107 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3112 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3113 * flag. It means that vm_struct is not fully initialized.
3114 * Now, it is fully initialized, so remove this flag here.
3116 clear_vm_uninitialized_flag(area);
3118 size = PAGE_ALIGN(size);
3119 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3120 kmemleak_vmalloc(area, size, gfp_mask);
3125 if (shift > PAGE_SHIFT) {
3136 * __vmalloc_node - allocate virtually contiguous memory
3137 * @size: allocation size
3138 * @align: desired alignment
3139 * @gfp_mask: flags for the page level allocator
3140 * @node: node to use for allocation or NUMA_NO_NODE
3141 * @caller: caller's return address
3143 * Allocate enough pages to cover @size from the page level allocator with
3144 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3146 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3147 * and __GFP_NOFAIL are not supported
3149 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3152 * Return: pointer to the allocated memory or %NULL on error
3154 void *__vmalloc_node(unsigned long size, unsigned long align,
3155 gfp_t gfp_mask, int node, const void *caller)
3157 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3158 gfp_mask, PAGE_KERNEL, 0, node, caller);
3161 * This is only for performance analysis of vmalloc and stress purpose.
3162 * It is required by vmalloc test module, therefore do not use it other
3165 #ifdef CONFIG_TEST_VMALLOC_MODULE
3166 EXPORT_SYMBOL_GPL(__vmalloc_node);
3169 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3171 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3172 __builtin_return_address(0));
3174 EXPORT_SYMBOL(__vmalloc);
3177 * vmalloc - allocate virtually contiguous memory
3178 * @size: allocation size
3180 * Allocate enough pages to cover @size from the page level
3181 * allocator and map them into contiguous kernel virtual space.
3183 * For tight control over page level allocator and protection flags
3184 * use __vmalloc() instead.
3186 * Return: pointer to the allocated memory or %NULL on error
3188 void *vmalloc(unsigned long size)
3190 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3191 __builtin_return_address(0));
3193 EXPORT_SYMBOL(vmalloc);
3196 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3197 * @size: allocation size
3199 * Allocate enough non-huge pages to cover @size from the page level
3200 * allocator and map them into contiguous kernel virtual space.
3202 * Return: pointer to the allocated memory or %NULL on error
3204 void *vmalloc_no_huge(unsigned long size)
3206 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3207 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3208 NUMA_NO_NODE, __builtin_return_address(0));
3210 EXPORT_SYMBOL(vmalloc_no_huge);
3213 * vzalloc - allocate virtually contiguous memory with zero fill
3214 * @size: allocation size
3216 * Allocate enough pages to cover @size from the page level
3217 * allocator and map them into contiguous kernel virtual space.
3218 * The memory allocated is set to zero.
3220 * For tight control over page level allocator and protection flags
3221 * use __vmalloc() instead.
3223 * Return: pointer to the allocated memory or %NULL on error
3225 void *vzalloc(unsigned long size)
3227 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3228 __builtin_return_address(0));
3230 EXPORT_SYMBOL(vzalloc);
3233 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3234 * @size: allocation size
3236 * The resulting memory area is zeroed so it can be mapped to userspace
3237 * without leaking data.
3239 * Return: pointer to the allocated memory or %NULL on error
3241 void *vmalloc_user(unsigned long size)
3243 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3244 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3245 VM_USERMAP, NUMA_NO_NODE,
3246 __builtin_return_address(0));
3248 EXPORT_SYMBOL(vmalloc_user);
3251 * vmalloc_node - allocate memory on a specific node
3252 * @size: allocation size
3255 * Allocate enough pages to cover @size from the page level
3256 * allocator and map them into contiguous kernel virtual space.
3258 * For tight control over page level allocator and protection flags
3259 * use __vmalloc() instead.
3261 * Return: pointer to the allocated memory or %NULL on error
3263 void *vmalloc_node(unsigned long size, int node)
3265 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3266 __builtin_return_address(0));
3268 EXPORT_SYMBOL(vmalloc_node);
3271 * vzalloc_node - allocate memory on a specific node with zero fill
3272 * @size: allocation size
3275 * Allocate enough pages to cover @size from the page level
3276 * allocator and map them into contiguous kernel virtual space.
3277 * The memory allocated is set to zero.
3279 * Return: pointer to the allocated memory or %NULL on error
3281 void *vzalloc_node(unsigned long size, int node)
3283 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3284 __builtin_return_address(0));
3286 EXPORT_SYMBOL(vzalloc_node);
3288 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3289 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3290 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3291 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3294 * 64b systems should always have either DMA or DMA32 zones. For others
3295 * GFP_DMA32 should do the right thing and use the normal zone.
3297 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3301 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3302 * @size: allocation size
3304 * Allocate enough 32bit PA addressable pages to cover @size from the
3305 * page level allocator and map them into contiguous kernel virtual space.
3307 * Return: pointer to the allocated memory or %NULL on error
3309 void *vmalloc_32(unsigned long size)
3311 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3312 __builtin_return_address(0));
3314 EXPORT_SYMBOL(vmalloc_32);
3317 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3318 * @size: allocation size
3320 * The resulting memory area is 32bit addressable and zeroed so it can be
3321 * mapped to userspace without leaking data.
3323 * Return: pointer to the allocated memory or %NULL on error
3325 void *vmalloc_32_user(unsigned long size)
3327 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3328 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3329 VM_USERMAP, NUMA_NO_NODE,
3330 __builtin_return_address(0));
3332 EXPORT_SYMBOL(vmalloc_32_user);
3335 * small helper routine , copy contents to buf from addr.
3336 * If the page is not present, fill zero.
3339 static int aligned_vread(char *buf, char *addr, unsigned long count)
3345 unsigned long offset, length;
3347 offset = offset_in_page(addr);
3348 length = PAGE_SIZE - offset;
3351 p = vmalloc_to_page(addr);
3353 * To do safe access to this _mapped_ area, we need
3354 * lock. But adding lock here means that we need to add
3355 * overhead of vmalloc()/vfree() calls for this _debug_
3356 * interface, rarely used. Instead of that, we'll use
3357 * kmap() and get small overhead in this access function.
3360 /* We can expect USER0 is not used -- see vread() */
3361 void *map = kmap_atomic(p);
3362 memcpy(buf, map + offset, length);
3365 memset(buf, 0, length);
3376 * vread() - read vmalloc area in a safe way.
3377 * @buf: buffer for reading data
3378 * @addr: vm address.
3379 * @count: number of bytes to be read.
3381 * This function checks that addr is a valid vmalloc'ed area, and
3382 * copy data from that area to a given buffer. If the given memory range
3383 * of [addr...addr+count) includes some valid address, data is copied to
3384 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3385 * IOREMAP area is treated as memory hole and no copy is done.
3387 * If [addr...addr+count) doesn't includes any intersects with alive
3388 * vm_struct area, returns 0. @buf should be kernel's buffer.
3390 * Note: In usual ops, vread() is never necessary because the caller
3391 * should know vmalloc() area is valid and can use memcpy().
3392 * This is for routines which have to access vmalloc area without
3393 * any information, as /proc/kcore.
3395 * Return: number of bytes for which addr and buf should be increased
3396 * (same number as @count) or %0 if [addr...addr+count) doesn't
3397 * include any intersection with valid vmalloc area
3399 long vread(char *buf, char *addr, unsigned long count)
3401 struct vmap_area *va;
3402 struct vm_struct *vm;
3403 char *vaddr, *buf_start = buf;
3404 unsigned long buflen = count;
3407 /* Don't allow overflow */
3408 if ((unsigned long) addr + count < count)
3409 count = -(unsigned long) addr;
3411 spin_lock(&vmap_area_lock);
3412 va = find_vmap_area_exceed_addr((unsigned long)addr);
3416 /* no intersects with alive vmap_area */
3417 if ((unsigned long)addr + count <= va->va_start)
3420 list_for_each_entry_from(va, &vmap_area_list, list) {
3428 vaddr = (char *) vm->addr;
3429 if (addr >= vaddr + get_vm_area_size(vm))
3431 while (addr < vaddr) {
3439 n = vaddr + get_vm_area_size(vm) - addr;
3442 if (!(vm->flags & VM_IOREMAP))
3443 aligned_vread(buf, addr, n);
3444 else /* IOREMAP area is treated as memory hole */
3451 spin_unlock(&vmap_area_lock);
3453 if (buf == buf_start)
3455 /* zero-fill memory holes */
3456 if (buf != buf_start + buflen)
3457 memset(buf, 0, buflen - (buf - buf_start));
3463 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3464 * @vma: vma to cover
3465 * @uaddr: target user address to start at
3466 * @kaddr: virtual address of vmalloc kernel memory
3467 * @pgoff: offset from @kaddr to start at
3468 * @size: size of map area
3470 * Returns: 0 for success, -Exxx on failure
3472 * This function checks that @kaddr is a valid vmalloc'ed area,
3473 * and that it is big enough to cover the range starting at
3474 * @uaddr in @vma. Will return failure if that criteria isn't
3477 * Similar to remap_pfn_range() (see mm/memory.c)
3479 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3480 void *kaddr, unsigned long pgoff,
3483 struct vm_struct *area;
3485 unsigned long end_index;
3487 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3490 size = PAGE_ALIGN(size);
3492 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3495 area = find_vm_area(kaddr);
3499 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3502 if (check_add_overflow(size, off, &end_index) ||
3503 end_index > get_vm_area_size(area))
3508 struct page *page = vmalloc_to_page(kaddr);
3511 ret = vm_insert_page(vma, uaddr, page);
3520 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3526 * remap_vmalloc_range - map vmalloc pages to userspace
3527 * @vma: vma to cover (map full range of vma)
3528 * @addr: vmalloc memory
3529 * @pgoff: number of pages into addr before first page to map
3531 * Returns: 0 for success, -Exxx on failure
3533 * This function checks that addr is a valid vmalloc'ed area, and
3534 * that it is big enough to cover the vma. Will return failure if
3535 * that criteria isn't met.
3537 * Similar to remap_pfn_range() (see mm/memory.c)
3539 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3540 unsigned long pgoff)
3542 return remap_vmalloc_range_partial(vma, vma->vm_start,
3544 vma->vm_end - vma->vm_start);
3546 EXPORT_SYMBOL(remap_vmalloc_range);
3548 void free_vm_area(struct vm_struct *area)
3550 struct vm_struct *ret;
3551 ret = remove_vm_area(area->addr);
3552 BUG_ON(ret != area);
3555 EXPORT_SYMBOL_GPL(free_vm_area);
3558 static struct vmap_area *node_to_va(struct rb_node *n)
3560 return rb_entry_safe(n, struct vmap_area, rb_node);
3564 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3565 * @addr: target address
3567 * Returns: vmap_area if it is found. If there is no such area
3568 * the first highest(reverse order) vmap_area is returned
3569 * i.e. va->va_start < addr && va->va_end < addr or NULL
3570 * if there are no any areas before @addr.
3572 static struct vmap_area *
3573 pvm_find_va_enclose_addr(unsigned long addr)
3575 struct vmap_area *va, *tmp;
3578 n = free_vmap_area_root.rb_node;
3582 tmp = rb_entry(n, struct vmap_area, rb_node);
3583 if (tmp->va_start <= addr) {
3585 if (tmp->va_end >= addr)
3598 * pvm_determine_end_from_reverse - find the highest aligned address
3599 * of free block below VMALLOC_END
3601 * in - the VA we start the search(reverse order);
3602 * out - the VA with the highest aligned end address.
3603 * @align: alignment for required highest address
3605 * Returns: determined end address within vmap_area
3607 static unsigned long
3608 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3610 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3614 list_for_each_entry_from_reverse((*va),
3615 &free_vmap_area_list, list) {
3616 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3617 if ((*va)->va_start < addr)
3626 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3627 * @offsets: array containing offset of each area
3628 * @sizes: array containing size of each area
3629 * @nr_vms: the number of areas to allocate
3630 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3632 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3633 * vm_structs on success, %NULL on failure
3635 * Percpu allocator wants to use congruent vm areas so that it can
3636 * maintain the offsets among percpu areas. This function allocates
3637 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3638 * be scattered pretty far, distance between two areas easily going up
3639 * to gigabytes. To avoid interacting with regular vmallocs, these
3640 * areas are allocated from top.
3642 * Despite its complicated look, this allocator is rather simple. It
3643 * does everything top-down and scans free blocks from the end looking
3644 * for matching base. While scanning, if any of the areas do not fit the
3645 * base address is pulled down to fit the area. Scanning is repeated till
3646 * all the areas fit and then all necessary data structures are inserted
3647 * and the result is returned.
3649 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3650 const size_t *sizes, int nr_vms,
3653 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3654 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3655 struct vmap_area **vas, *va;
3656 struct vm_struct **vms;
3657 int area, area2, last_area, term_area;
3658 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3659 bool purged = false;
3662 /* verify parameters and allocate data structures */
3663 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3664 for (last_area = 0, area = 0; area < nr_vms; area++) {
3665 start = offsets[area];
3666 end = start + sizes[area];
3668 /* is everything aligned properly? */
3669 BUG_ON(!IS_ALIGNED(offsets[area], align));
3670 BUG_ON(!IS_ALIGNED(sizes[area], align));
3672 /* detect the area with the highest address */
3673 if (start > offsets[last_area])
3676 for (area2 = area + 1; area2 < nr_vms; area2++) {
3677 unsigned long start2 = offsets[area2];
3678 unsigned long end2 = start2 + sizes[area2];
3680 BUG_ON(start2 < end && start < end2);
3683 last_end = offsets[last_area] + sizes[last_area];
3685 if (vmalloc_end - vmalloc_start < last_end) {
3690 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3691 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3695 for (area = 0; area < nr_vms; area++) {
3696 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3697 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3698 if (!vas[area] || !vms[area])
3702 spin_lock(&free_vmap_area_lock);
3704 /* start scanning - we scan from the top, begin with the last area */
3705 area = term_area = last_area;
3706 start = offsets[area];
3707 end = start + sizes[area];
3709 va = pvm_find_va_enclose_addr(vmalloc_end);
3710 base = pvm_determine_end_from_reverse(&va, align) - end;
3714 * base might have underflowed, add last_end before
3717 if (base + last_end < vmalloc_start + last_end)
3721 * Fitting base has not been found.
3727 * If required width exceeds current VA block, move
3728 * base downwards and then recheck.
3730 if (base + end > va->va_end) {
3731 base = pvm_determine_end_from_reverse(&va, align) - end;
3737 * If this VA does not fit, move base downwards and recheck.
3739 if (base + start < va->va_start) {
3740 va = node_to_va(rb_prev(&va->rb_node));
3741 base = pvm_determine_end_from_reverse(&va, align) - end;
3747 * This area fits, move on to the previous one. If
3748 * the previous one is the terminal one, we're done.
3750 area = (area + nr_vms - 1) % nr_vms;
3751 if (area == term_area)
3754 start = offsets[area];
3755 end = start + sizes[area];
3756 va = pvm_find_va_enclose_addr(base + end);
3759 /* we've found a fitting base, insert all va's */
3760 for (area = 0; area < nr_vms; area++) {
3763 start = base + offsets[area];
3766 va = pvm_find_va_enclose_addr(start);
3767 if (WARN_ON_ONCE(va == NULL))
3768 /* It is a BUG(), but trigger recovery instead. */
3771 type = classify_va_fit_type(va, start, size);
3772 if (WARN_ON_ONCE(type == NOTHING_FIT))
3773 /* It is a BUG(), but trigger recovery instead. */
3776 ret = adjust_va_to_fit_type(va, start, size, type);
3780 /* Allocated area. */
3782 va->va_start = start;
3783 va->va_end = start + size;
3786 spin_unlock(&free_vmap_area_lock);
3788 /* populate the kasan shadow space */
3789 for (area = 0; area < nr_vms; area++) {
3790 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3791 goto err_free_shadow;
3793 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3797 /* insert all vm's */
3798 spin_lock(&vmap_area_lock);
3799 for (area = 0; area < nr_vms; area++) {
3800 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3802 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3805 spin_unlock(&vmap_area_lock);
3812 * Remove previously allocated areas. There is no
3813 * need in removing these areas from the busy tree,
3814 * because they are inserted only on the final step
3815 * and when pcpu_get_vm_areas() is success.
3818 orig_start = vas[area]->va_start;
3819 orig_end = vas[area]->va_end;
3820 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3821 &free_vmap_area_list);
3823 kasan_release_vmalloc(orig_start, orig_end,
3824 va->va_start, va->va_end);
3829 spin_unlock(&free_vmap_area_lock);
3831 purge_vmap_area_lazy();
3834 /* Before "retry", check if we recover. */
3835 for (area = 0; area < nr_vms; area++) {
3839 vas[area] = kmem_cache_zalloc(
3840 vmap_area_cachep, GFP_KERNEL);
3849 for (area = 0; area < nr_vms; area++) {
3851 kmem_cache_free(vmap_area_cachep, vas[area]);
3861 spin_lock(&free_vmap_area_lock);
3863 * We release all the vmalloc shadows, even the ones for regions that
3864 * hadn't been successfully added. This relies on kasan_release_vmalloc
3865 * being able to tolerate this case.
3867 for (area = 0; area < nr_vms; area++) {
3868 orig_start = vas[area]->va_start;
3869 orig_end = vas[area]->va_end;
3870 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3871 &free_vmap_area_list);
3873 kasan_release_vmalloc(orig_start, orig_end,
3874 va->va_start, va->va_end);
3878 spin_unlock(&free_vmap_area_lock);
3885 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3886 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3887 * @nr_vms: the number of allocated areas
3889 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3891 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3895 for (i = 0; i < nr_vms; i++)
3896 free_vm_area(vms[i]);
3899 #endif /* CONFIG_SMP */
3901 #ifdef CONFIG_PRINTK
3902 bool vmalloc_dump_obj(void *object)
3904 struct vm_struct *vm;
3905 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3907 vm = find_vm_area(objp);
3910 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3911 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3916 #ifdef CONFIG_PROC_FS
3917 static void *s_start(struct seq_file *m, loff_t *pos)
3918 __acquires(&vmap_purge_lock)
3919 __acquires(&vmap_area_lock)
3921 mutex_lock(&vmap_purge_lock);
3922 spin_lock(&vmap_area_lock);
3924 return seq_list_start(&vmap_area_list, *pos);
3927 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3929 return seq_list_next(p, &vmap_area_list, pos);
3932 static void s_stop(struct seq_file *m, void *p)
3933 __releases(&vmap_area_lock)
3934 __releases(&vmap_purge_lock)
3936 spin_unlock(&vmap_area_lock);
3937 mutex_unlock(&vmap_purge_lock);
3940 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3942 if (IS_ENABLED(CONFIG_NUMA)) {
3943 unsigned int nr, *counters = m->private;
3944 unsigned int step = 1U << vm_area_page_order(v);
3949 if (v->flags & VM_UNINITIALIZED)
3951 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3954 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3956 for (nr = 0; nr < v->nr_pages; nr += step)
3957 counters[page_to_nid(v->pages[nr])] += step;
3958 for_each_node_state(nr, N_HIGH_MEMORY)
3960 seq_printf(m, " N%u=%u", nr, counters[nr]);
3964 static void show_purge_info(struct seq_file *m)
3966 struct vmap_area *va;
3968 spin_lock(&purge_vmap_area_lock);
3969 list_for_each_entry(va, &purge_vmap_area_list, list) {
3970 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3971 (void *)va->va_start, (void *)va->va_end,
3972 va->va_end - va->va_start);
3974 spin_unlock(&purge_vmap_area_lock);
3977 static int s_show(struct seq_file *m, void *p)
3979 struct vmap_area *va;
3980 struct vm_struct *v;
3982 va = list_entry(p, struct vmap_area, list);
3985 * s_show can encounter race with remove_vm_area, !vm on behalf
3986 * of vmap area is being tear down or vm_map_ram allocation.
3989 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3990 (void *)va->va_start, (void *)va->va_end,
3991 va->va_end - va->va_start);
3998 seq_printf(m, "0x%pK-0x%pK %7ld",
3999 v->addr, v->addr + v->size, v->size);
4002 seq_printf(m, " %pS", v->caller);
4005 seq_printf(m, " pages=%d", v->nr_pages);
4008 seq_printf(m, " phys=%pa", &v->phys_addr);
4010 if (v->flags & VM_IOREMAP)
4011 seq_puts(m, " ioremap");
4013 if (v->flags & VM_ALLOC)
4014 seq_puts(m, " vmalloc");
4016 if (v->flags & VM_MAP)
4017 seq_puts(m, " vmap");
4019 if (v->flags & VM_USERMAP)
4020 seq_puts(m, " user");
4022 if (v->flags & VM_DMA_COHERENT)
4023 seq_puts(m, " dma-coherent");
4025 if (is_vmalloc_addr(v->pages))
4026 seq_puts(m, " vpages");
4028 show_numa_info(m, v);
4032 * As a final step, dump "unpurged" areas.
4035 if (list_is_last(&va->list, &vmap_area_list))
4041 static const struct seq_operations vmalloc_op = {
4048 static int __init proc_vmalloc_init(void)
4050 if (IS_ENABLED(CONFIG_NUMA))
4051 proc_create_seq_private("vmallocinfo", 0400, NULL,
4053 nr_node_ids * sizeof(unsigned int), NULL);
4055 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4058 module_init(proc_vmalloc_init);