1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/vmalloc.h>
50 #include "pgalloc-track.h"
52 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
55 static int __init set_nohugeiomap(char *str)
57 ioremap_max_page_shift = PAGE_SHIFT;
60 early_param("nohugeiomap", set_nohugeiomap);
61 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
63 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66 static bool __ro_after_init vmap_allow_huge = true;
68 static int __init set_nohugevmalloc(char *str)
70 vmap_allow_huge = false;
73 early_param("nohugevmalloc", set_nohugevmalloc);
74 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 static const bool vmap_allow_huge = false;
76 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78 bool is_vmalloc_addr(const void *x)
80 unsigned long addr = (unsigned long)kasan_reset_tag(x);
82 return addr >= VMALLOC_START && addr < VMALLOC_END;
84 EXPORT_SYMBOL(is_vmalloc_addr);
86 struct vfree_deferred {
87 struct llist_head list;
88 struct work_struct wq;
90 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
92 static void __vunmap(const void *, int);
94 static void free_work(struct work_struct *w)
96 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
97 struct llist_node *t, *llnode;
99 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
100 __vunmap((void *)llnode, 1);
103 /*** Page table manipulation functions ***/
104 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
105 phys_addr_t phys_addr, pgprot_t prot,
106 unsigned int max_page_shift, pgtbl_mod_mask *mask)
110 unsigned long size = PAGE_SIZE;
112 pfn = phys_addr >> PAGE_SHIFT;
113 pte = pte_alloc_kernel_track(pmd, addr, mask);
117 BUG_ON(!pte_none(*pte));
119 #ifdef CONFIG_HUGETLB_PAGE
120 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
121 if (size != PAGE_SIZE) {
122 pte_t entry = pfn_pte(pfn, prot);
124 entry = arch_make_huge_pte(entry, ilog2(size), 0);
125 set_huge_pte_at(&init_mm, addr, pte, entry);
126 pfn += PFN_DOWN(size);
130 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
132 } while (pte += PFN_DOWN(size), addr += size, addr != end);
133 *mask |= PGTBL_PTE_MODIFIED;
137 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
138 phys_addr_t phys_addr, pgprot_t prot,
139 unsigned int max_page_shift)
141 if (max_page_shift < PMD_SHIFT)
144 if (!arch_vmap_pmd_supported(prot))
147 if ((end - addr) != PMD_SIZE)
150 if (!IS_ALIGNED(addr, PMD_SIZE))
153 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
156 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
159 return pmd_set_huge(pmd, phys_addr, prot);
162 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
163 phys_addr_t phys_addr, pgprot_t prot,
164 unsigned int max_page_shift, pgtbl_mod_mask *mask)
169 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
173 next = pmd_addr_end(addr, end);
175 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
177 *mask |= PGTBL_PMD_MODIFIED;
181 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
183 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
187 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
188 phys_addr_t phys_addr, pgprot_t prot,
189 unsigned int max_page_shift)
191 if (max_page_shift < PUD_SHIFT)
194 if (!arch_vmap_pud_supported(prot))
197 if ((end - addr) != PUD_SIZE)
200 if (!IS_ALIGNED(addr, PUD_SIZE))
203 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
206 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
209 return pud_set_huge(pud, phys_addr, prot);
212 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
213 phys_addr_t phys_addr, pgprot_t prot,
214 unsigned int max_page_shift, pgtbl_mod_mask *mask)
219 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
223 next = pud_addr_end(addr, end);
225 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
227 *mask |= PGTBL_PUD_MODIFIED;
231 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
232 max_page_shift, mask))
234 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
238 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
239 phys_addr_t phys_addr, pgprot_t prot,
240 unsigned int max_page_shift)
242 if (max_page_shift < P4D_SHIFT)
245 if (!arch_vmap_p4d_supported(prot))
248 if ((end - addr) != P4D_SIZE)
251 if (!IS_ALIGNED(addr, P4D_SIZE))
254 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
257 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
260 return p4d_set_huge(p4d, phys_addr, prot);
263 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
264 phys_addr_t phys_addr, pgprot_t prot,
265 unsigned int max_page_shift, pgtbl_mod_mask *mask)
270 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
274 next = p4d_addr_end(addr, end);
276 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
278 *mask |= PGTBL_P4D_MODIFIED;
282 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
283 max_page_shift, mask))
285 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
289 static int vmap_range_noflush(unsigned long addr, unsigned long end,
290 phys_addr_t phys_addr, pgprot_t prot,
291 unsigned int max_page_shift)
297 pgtbl_mod_mask mask = 0;
303 pgd = pgd_offset_k(addr);
305 next = pgd_addr_end(addr, end);
306 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
307 max_page_shift, &mask);
310 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
312 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
313 arch_sync_kernel_mappings(start, end);
318 int ioremap_page_range(unsigned long addr, unsigned long end,
319 phys_addr_t phys_addr, pgprot_t prot)
323 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
324 ioremap_max_page_shift);
325 flush_cache_vmap(addr, end);
327 kmsan_ioremap_page_range(addr, end, phys_addr, prot,
328 ioremap_max_page_shift);
332 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
333 pgtbl_mod_mask *mask)
337 pte = pte_offset_kernel(pmd, addr);
339 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
340 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
341 } while (pte++, addr += PAGE_SIZE, addr != end);
342 *mask |= PGTBL_PTE_MODIFIED;
345 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
346 pgtbl_mod_mask *mask)
352 pmd = pmd_offset(pud, addr);
354 next = pmd_addr_end(addr, end);
356 cleared = pmd_clear_huge(pmd);
357 if (cleared || pmd_bad(*pmd))
358 *mask |= PGTBL_PMD_MODIFIED;
362 if (pmd_none_or_clear_bad(pmd))
364 vunmap_pte_range(pmd, addr, next, mask);
367 } while (pmd++, addr = next, addr != end);
370 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
371 pgtbl_mod_mask *mask)
377 pud = pud_offset(p4d, addr);
379 next = pud_addr_end(addr, end);
381 cleared = pud_clear_huge(pud);
382 if (cleared || pud_bad(*pud))
383 *mask |= PGTBL_PUD_MODIFIED;
387 if (pud_none_or_clear_bad(pud))
389 vunmap_pmd_range(pud, addr, next, mask);
390 } while (pud++, addr = next, addr != end);
393 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
394 pgtbl_mod_mask *mask)
399 p4d = p4d_offset(pgd, addr);
401 next = p4d_addr_end(addr, end);
405 *mask |= PGTBL_P4D_MODIFIED;
407 if (p4d_none_or_clear_bad(p4d))
409 vunmap_pud_range(p4d, addr, next, mask);
410 } while (p4d++, addr = next, addr != end);
414 * vunmap_range_noflush is similar to vunmap_range, but does not
415 * flush caches or TLBs.
417 * The caller is responsible for calling flush_cache_vmap() before calling
418 * this function, and flush_tlb_kernel_range after it has returned
419 * successfully (and before the addresses are expected to cause a page fault
420 * or be re-mapped for something else, if TLB flushes are being delayed or
423 * This is an internal function only. Do not use outside mm/.
425 void __vunmap_range_noflush(unsigned long start, unsigned long end)
429 unsigned long addr = start;
430 pgtbl_mod_mask mask = 0;
433 pgd = pgd_offset_k(addr);
435 next = pgd_addr_end(addr, end);
437 mask |= PGTBL_PGD_MODIFIED;
438 if (pgd_none_or_clear_bad(pgd))
440 vunmap_p4d_range(pgd, addr, next, &mask);
441 } while (pgd++, addr = next, addr != end);
443 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
444 arch_sync_kernel_mappings(start, end);
447 void vunmap_range_noflush(unsigned long start, unsigned long end)
449 kmsan_vunmap_range_noflush(start, end);
450 __vunmap_range_noflush(start, end);
454 * vunmap_range - unmap kernel virtual addresses
455 * @addr: start of the VM area to unmap
456 * @end: end of the VM area to unmap (non-inclusive)
458 * Clears any present PTEs in the virtual address range, flushes TLBs and
459 * caches. Any subsequent access to the address before it has been re-mapped
462 void vunmap_range(unsigned long addr, unsigned long end)
464 flush_cache_vunmap(addr, end);
465 vunmap_range_noflush(addr, end);
466 flush_tlb_kernel_range(addr, end);
469 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
470 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
471 pgtbl_mod_mask *mask)
476 * nr is a running index into the array which helps higher level
477 * callers keep track of where we're up to.
480 pte = pte_alloc_kernel_track(pmd, addr, mask);
484 struct page *page = pages[*nr];
486 if (WARN_ON(!pte_none(*pte)))
490 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
493 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
495 } while (pte++, addr += PAGE_SIZE, addr != end);
496 *mask |= PGTBL_PTE_MODIFIED;
500 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
501 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
502 pgtbl_mod_mask *mask)
507 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
511 next = pmd_addr_end(addr, end);
512 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
514 } while (pmd++, addr = next, addr != end);
518 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
519 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
520 pgtbl_mod_mask *mask)
525 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
529 next = pud_addr_end(addr, end);
530 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
532 } while (pud++, addr = next, addr != end);
536 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
537 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
538 pgtbl_mod_mask *mask)
543 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
547 next = p4d_addr_end(addr, end);
548 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
550 } while (p4d++, addr = next, addr != end);
554 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
555 pgprot_t prot, struct page **pages)
557 unsigned long start = addr;
562 pgtbl_mod_mask mask = 0;
565 pgd = pgd_offset_k(addr);
567 next = pgd_addr_end(addr, end);
569 mask |= PGTBL_PGD_MODIFIED;
570 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
573 } while (pgd++, addr = next, addr != end);
575 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
576 arch_sync_kernel_mappings(start, end);
582 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
585 * The caller is responsible for calling flush_cache_vmap() after this
586 * function returns successfully and before the addresses are accessed.
588 * This is an internal function only. Do not use outside mm/.
590 int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
591 pgprot_t prot, struct page **pages, unsigned int page_shift)
593 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
595 WARN_ON(page_shift < PAGE_SHIFT);
597 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
598 page_shift == PAGE_SHIFT)
599 return vmap_small_pages_range_noflush(addr, end, prot, pages);
601 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
604 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
605 page_to_phys(pages[i]), prot,
610 addr += 1UL << page_shift;
616 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
617 pgprot_t prot, struct page **pages, unsigned int page_shift)
619 kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
620 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
624 * vmap_pages_range - map pages to a kernel virtual address
625 * @addr: start of the VM area to map
626 * @end: end of the VM area to map (non-inclusive)
627 * @prot: page protection flags to use
628 * @pages: pages to map (always PAGE_SIZE pages)
629 * @page_shift: maximum shift that the pages may be mapped with, @pages must
630 * be aligned and contiguous up to at least this shift.
633 * 0 on success, -errno on failure.
635 static int vmap_pages_range(unsigned long addr, unsigned long end,
636 pgprot_t prot, struct page **pages, unsigned int page_shift)
640 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
641 flush_cache_vmap(addr, end);
645 int is_vmalloc_or_module_addr(const void *x)
648 * ARM, x86-64 and sparc64 put modules in a special place,
649 * and fall back on vmalloc() if that fails. Others
650 * just put it in the vmalloc space.
652 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
653 unsigned long addr = (unsigned long)kasan_reset_tag(x);
654 if (addr >= MODULES_VADDR && addr < MODULES_END)
657 return is_vmalloc_addr(x);
659 EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
662 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
663 * return the tail page that corresponds to the base page address, which
664 * matches small vmap mappings.
666 struct page *vmalloc_to_page(const void *vmalloc_addr)
668 unsigned long addr = (unsigned long) vmalloc_addr;
669 struct page *page = NULL;
670 pgd_t *pgd = pgd_offset_k(addr);
677 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
678 * architectures that do not vmalloc module space
680 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
684 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
685 return NULL; /* XXX: no allowance for huge pgd */
686 if (WARN_ON_ONCE(pgd_bad(*pgd)))
689 p4d = p4d_offset(pgd, addr);
693 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
694 if (WARN_ON_ONCE(p4d_bad(*p4d)))
697 pud = pud_offset(p4d, addr);
701 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
702 if (WARN_ON_ONCE(pud_bad(*pud)))
705 pmd = pmd_offset(pud, addr);
709 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
710 if (WARN_ON_ONCE(pmd_bad(*pmd)))
713 ptep = pte_offset_map(pmd, addr);
715 if (pte_present(pte))
716 page = pte_page(pte);
721 EXPORT_SYMBOL(vmalloc_to_page);
724 * Map a vmalloc()-space virtual address to the physical page frame number.
726 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
728 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
730 EXPORT_SYMBOL(vmalloc_to_pfn);
733 /*** Global kva allocator ***/
735 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
736 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
739 static DEFINE_SPINLOCK(vmap_area_lock);
740 static DEFINE_SPINLOCK(free_vmap_area_lock);
741 /* Export for kexec only */
742 LIST_HEAD(vmap_area_list);
743 static struct rb_root vmap_area_root = RB_ROOT;
744 static bool vmap_initialized __read_mostly;
746 static struct rb_root purge_vmap_area_root = RB_ROOT;
747 static LIST_HEAD(purge_vmap_area_list);
748 static DEFINE_SPINLOCK(purge_vmap_area_lock);
751 * This kmem_cache is used for vmap_area objects. Instead of
752 * allocating from slab we reuse an object from this cache to
753 * make things faster. Especially in "no edge" splitting of
756 static struct kmem_cache *vmap_area_cachep;
759 * This linked list is used in pair with free_vmap_area_root.
760 * It gives O(1) access to prev/next to perform fast coalescing.
762 static LIST_HEAD(free_vmap_area_list);
765 * This augment red-black tree represents the free vmap space.
766 * All vmap_area objects in this tree are sorted by va->va_start
767 * address. It is used for allocation and merging when a vmap
768 * object is released.
770 * Each vmap_area node contains a maximum available free block
771 * of its sub-tree, right or left. Therefore it is possible to
772 * find a lowest match of free area.
774 static struct rb_root free_vmap_area_root = RB_ROOT;
777 * Preload a CPU with one object for "no edge" split case. The
778 * aim is to get rid of allocations from the atomic context, thus
779 * to use more permissive allocation masks.
781 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
783 static __always_inline unsigned long
784 va_size(struct vmap_area *va)
786 return (va->va_end - va->va_start);
789 static __always_inline unsigned long
790 get_subtree_max_size(struct rb_node *node)
792 struct vmap_area *va;
794 va = rb_entry_safe(node, struct vmap_area, rb_node);
795 return va ? va->subtree_max_size : 0;
798 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
799 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
801 static void purge_vmap_area_lazy(void);
802 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
803 static void drain_vmap_area_work(struct work_struct *work);
804 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
806 static atomic_long_t nr_vmalloc_pages;
808 unsigned long vmalloc_nr_pages(void)
810 return atomic_long_read(&nr_vmalloc_pages);
813 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
814 static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
816 struct vmap_area *va = NULL;
817 struct rb_node *n = vmap_area_root.rb_node;
819 addr = (unsigned long)kasan_reset_tag((void *)addr);
822 struct vmap_area *tmp;
824 tmp = rb_entry(n, struct vmap_area, rb_node);
825 if (tmp->va_end > addr) {
827 if (tmp->va_start <= addr)
838 static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
840 struct rb_node *n = root->rb_node;
842 addr = (unsigned long)kasan_reset_tag((void *)addr);
845 struct vmap_area *va;
847 va = rb_entry(n, struct vmap_area, rb_node);
848 if (addr < va->va_start)
850 else if (addr >= va->va_end)
860 * This function returns back addresses of parent node
861 * and its left or right link for further processing.
863 * Otherwise NULL is returned. In that case all further
864 * steps regarding inserting of conflicting overlap range
865 * have to be declined and actually considered as a bug.
867 static __always_inline struct rb_node **
868 find_va_links(struct vmap_area *va,
869 struct rb_root *root, struct rb_node *from,
870 struct rb_node **parent)
872 struct vmap_area *tmp_va;
873 struct rb_node **link;
876 link = &root->rb_node;
877 if (unlikely(!*link)) {
886 * Go to the bottom of the tree. When we hit the last point
887 * we end up with parent rb_node and correct direction, i name
888 * it link, where the new va->rb_node will be attached to.
891 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
894 * During the traversal we also do some sanity check.
895 * Trigger the BUG() if there are sides(left/right)
898 if (va->va_end <= tmp_va->va_start)
899 link = &(*link)->rb_left;
900 else if (va->va_start >= tmp_va->va_end)
901 link = &(*link)->rb_right;
903 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
904 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
910 *parent = &tmp_va->rb_node;
914 static __always_inline struct list_head *
915 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
917 struct list_head *list;
919 if (unlikely(!parent))
921 * The red-black tree where we try to find VA neighbors
922 * before merging or inserting is empty, i.e. it means
923 * there is no free vmap space. Normally it does not
924 * happen but we handle this case anyway.
928 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
929 return (&parent->rb_right == link ? list->next : list);
932 static __always_inline void
933 __link_va(struct vmap_area *va, struct rb_root *root,
934 struct rb_node *parent, struct rb_node **link,
935 struct list_head *head, bool augment)
938 * VA is still not in the list, but we can
939 * identify its future previous list_head node.
941 if (likely(parent)) {
942 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
943 if (&parent->rb_right != link)
947 /* Insert to the rb-tree */
948 rb_link_node(&va->rb_node, parent, link);
951 * Some explanation here. Just perform simple insertion
952 * to the tree. We do not set va->subtree_max_size to
953 * its current size before calling rb_insert_augmented().
954 * It is because we populate the tree from the bottom
955 * to parent levels when the node _is_ in the tree.
957 * Therefore we set subtree_max_size to zero after insertion,
958 * to let __augment_tree_propagate_from() puts everything to
959 * the correct order later on.
961 rb_insert_augmented(&va->rb_node,
962 root, &free_vmap_area_rb_augment_cb);
963 va->subtree_max_size = 0;
965 rb_insert_color(&va->rb_node, root);
968 /* Address-sort this list */
969 list_add(&va->list, head);
972 static __always_inline void
973 link_va(struct vmap_area *va, struct rb_root *root,
974 struct rb_node *parent, struct rb_node **link,
975 struct list_head *head)
977 __link_va(va, root, parent, link, head, false);
980 static __always_inline void
981 link_va_augment(struct vmap_area *va, struct rb_root *root,
982 struct rb_node *parent, struct rb_node **link,
983 struct list_head *head)
985 __link_va(va, root, parent, link, head, true);
988 static __always_inline void
989 __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
991 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
995 rb_erase_augmented(&va->rb_node,
996 root, &free_vmap_area_rb_augment_cb);
998 rb_erase(&va->rb_node, root);
1000 list_del_init(&va->list);
1001 RB_CLEAR_NODE(&va->rb_node);
1004 static __always_inline void
1005 unlink_va(struct vmap_area *va, struct rb_root *root)
1007 __unlink_va(va, root, false);
1010 static __always_inline void
1011 unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1013 __unlink_va(va, root, true);
1016 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1018 * Gets called when remove the node and rotate.
1020 static __always_inline unsigned long
1021 compute_subtree_max_size(struct vmap_area *va)
1023 return max3(va_size(va),
1024 get_subtree_max_size(va->rb_node.rb_left),
1025 get_subtree_max_size(va->rb_node.rb_right));
1029 augment_tree_propagate_check(void)
1031 struct vmap_area *va;
1032 unsigned long computed_size;
1034 list_for_each_entry(va, &free_vmap_area_list, list) {
1035 computed_size = compute_subtree_max_size(va);
1036 if (computed_size != va->subtree_max_size)
1037 pr_emerg("tree is corrupted: %lu, %lu\n",
1038 va_size(va), va->subtree_max_size);
1044 * This function populates subtree_max_size from bottom to upper
1045 * levels starting from VA point. The propagation must be done
1046 * when VA size is modified by changing its va_start/va_end. Or
1047 * in case of newly inserting of VA to the tree.
1049 * It means that __augment_tree_propagate_from() must be called:
1050 * - After VA has been inserted to the tree(free path);
1051 * - After VA has been shrunk(allocation path);
1052 * - After VA has been increased(merging path).
1054 * Please note that, it does not mean that upper parent nodes
1055 * and their subtree_max_size are recalculated all the time up
1064 * For example if we modify the node 4, shrinking it to 2, then
1065 * no any modification is required. If we shrink the node 2 to 1
1066 * its subtree_max_size is updated only, and set to 1. If we shrink
1067 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1068 * node becomes 4--6.
1070 static __always_inline void
1071 augment_tree_propagate_from(struct vmap_area *va)
1074 * Populate the tree from bottom towards the root until
1075 * the calculated maximum available size of checked node
1076 * is equal to its current one.
1078 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1080 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1081 augment_tree_propagate_check();
1086 insert_vmap_area(struct vmap_area *va,
1087 struct rb_root *root, struct list_head *head)
1089 struct rb_node **link;
1090 struct rb_node *parent;
1092 link = find_va_links(va, root, NULL, &parent);
1094 link_va(va, root, parent, link, head);
1098 insert_vmap_area_augment(struct vmap_area *va,
1099 struct rb_node *from, struct rb_root *root,
1100 struct list_head *head)
1102 struct rb_node **link;
1103 struct rb_node *parent;
1106 link = find_va_links(va, NULL, from, &parent);
1108 link = find_va_links(va, root, NULL, &parent);
1111 link_va_augment(va, root, parent, link, head);
1112 augment_tree_propagate_from(va);
1117 * Merge de-allocated chunk of VA memory with previous
1118 * and next free blocks. If coalesce is not done a new
1119 * free area is inserted. If VA has been merged, it is
1122 * Please note, it can return NULL in case of overlap
1123 * ranges, followed by WARN() report. Despite it is a
1124 * buggy behaviour, a system can be alive and keep
1127 static __always_inline struct vmap_area *
1128 __merge_or_add_vmap_area(struct vmap_area *va,
1129 struct rb_root *root, struct list_head *head, bool augment)
1131 struct vmap_area *sibling;
1132 struct list_head *next;
1133 struct rb_node **link;
1134 struct rb_node *parent;
1135 bool merged = false;
1138 * Find a place in the tree where VA potentially will be
1139 * inserted, unless it is merged with its sibling/siblings.
1141 link = find_va_links(va, root, NULL, &parent);
1146 * Get next node of VA to check if merging can be done.
1148 next = get_va_next_sibling(parent, link);
1149 if (unlikely(next == NULL))
1155 * |<------VA------>|<-----Next----->|
1160 sibling = list_entry(next, struct vmap_area, list);
1161 if (sibling->va_start == va->va_end) {
1162 sibling->va_start = va->va_start;
1164 /* Free vmap_area object. */
1165 kmem_cache_free(vmap_area_cachep, va);
1167 /* Point to the new merged area. */
1176 * |<-----Prev----->|<------VA------>|
1180 if (next->prev != head) {
1181 sibling = list_entry(next->prev, struct vmap_area, list);
1182 if (sibling->va_end == va->va_start) {
1184 * If both neighbors are coalesced, it is important
1185 * to unlink the "next" node first, followed by merging
1186 * with "previous" one. Otherwise the tree might not be
1187 * fully populated if a sibling's augmented value is
1188 * "normalized" because of rotation operations.
1191 __unlink_va(va, root, augment);
1193 sibling->va_end = va->va_end;
1195 /* Free vmap_area object. */
1196 kmem_cache_free(vmap_area_cachep, va);
1198 /* Point to the new merged area. */
1206 __link_va(va, root, parent, link, head, augment);
1211 static __always_inline struct vmap_area *
1212 merge_or_add_vmap_area(struct vmap_area *va,
1213 struct rb_root *root, struct list_head *head)
1215 return __merge_or_add_vmap_area(va, root, head, false);
1218 static __always_inline struct vmap_area *
1219 merge_or_add_vmap_area_augment(struct vmap_area *va,
1220 struct rb_root *root, struct list_head *head)
1222 va = __merge_or_add_vmap_area(va, root, head, true);
1224 augment_tree_propagate_from(va);
1229 static __always_inline bool
1230 is_within_this_va(struct vmap_area *va, unsigned long size,
1231 unsigned long align, unsigned long vstart)
1233 unsigned long nva_start_addr;
1235 if (va->va_start > vstart)
1236 nva_start_addr = ALIGN(va->va_start, align);
1238 nva_start_addr = ALIGN(vstart, align);
1240 /* Can be overflowed due to big size or alignment. */
1241 if (nva_start_addr + size < nva_start_addr ||
1242 nva_start_addr < vstart)
1245 return (nva_start_addr + size <= va->va_end);
1249 * Find the first free block(lowest start address) in the tree,
1250 * that will accomplish the request corresponding to passing
1251 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1252 * a search length is adjusted to account for worst case alignment
1255 static __always_inline struct vmap_area *
1256 find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1257 unsigned long align, unsigned long vstart, bool adjust_search_size)
1259 struct vmap_area *va;
1260 struct rb_node *node;
1261 unsigned long length;
1263 /* Start from the root. */
1264 node = root->rb_node;
1266 /* Adjust the search size for alignment overhead. */
1267 length = adjust_search_size ? size + align - 1 : size;
1270 va = rb_entry(node, struct vmap_area, rb_node);
1272 if (get_subtree_max_size(node->rb_left) >= length &&
1273 vstart < va->va_start) {
1274 node = node->rb_left;
1276 if (is_within_this_va(va, size, align, vstart))
1280 * Does not make sense to go deeper towards the right
1281 * sub-tree if it does not have a free block that is
1282 * equal or bigger to the requested search length.
1284 if (get_subtree_max_size(node->rb_right) >= length) {
1285 node = node->rb_right;
1290 * OK. We roll back and find the first right sub-tree,
1291 * that will satisfy the search criteria. It can happen
1292 * due to "vstart" restriction or an alignment overhead
1293 * that is bigger then PAGE_SIZE.
1295 while ((node = rb_parent(node))) {
1296 va = rb_entry(node, struct vmap_area, rb_node);
1297 if (is_within_this_va(va, size, align, vstart))
1300 if (get_subtree_max_size(node->rb_right) >= length &&
1301 vstart <= va->va_start) {
1303 * Shift the vstart forward. Please note, we update it with
1304 * parent's start address adding "1" because we do not want
1305 * to enter same sub-tree after it has already been checked
1306 * and no suitable free block found there.
1308 vstart = va->va_start + 1;
1309 node = node->rb_right;
1319 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1320 #include <linux/random.h>
1322 static struct vmap_area *
1323 find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1324 unsigned long align, unsigned long vstart)
1326 struct vmap_area *va;
1328 list_for_each_entry(va, head, list) {
1329 if (!is_within_this_va(va, size, align, vstart))
1339 find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1340 unsigned long size, unsigned long align)
1342 struct vmap_area *va_1, *va_2;
1343 unsigned long vstart;
1346 get_random_bytes(&rnd, sizeof(rnd));
1347 vstart = VMALLOC_START + rnd;
1349 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1350 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1353 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1354 va_1, va_2, vstart);
1360 FL_FIT_TYPE = 1, /* full fit */
1361 LE_FIT_TYPE = 2, /* left edge fit */
1362 RE_FIT_TYPE = 3, /* right edge fit */
1363 NE_FIT_TYPE = 4 /* no edge fit */
1366 static __always_inline enum fit_type
1367 classify_va_fit_type(struct vmap_area *va,
1368 unsigned long nva_start_addr, unsigned long size)
1372 /* Check if it is within VA. */
1373 if (nva_start_addr < va->va_start ||
1374 nva_start_addr + size > va->va_end)
1378 if (va->va_start == nva_start_addr) {
1379 if (va->va_end == nva_start_addr + size)
1383 } else if (va->va_end == nva_start_addr + size) {
1392 static __always_inline int
1393 adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1394 struct vmap_area *va, unsigned long nva_start_addr,
1397 struct vmap_area *lva = NULL;
1398 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1400 if (type == FL_FIT_TYPE) {
1402 * No need to split VA, it fully fits.
1408 unlink_va_augment(va, root);
1409 kmem_cache_free(vmap_area_cachep, va);
1410 } else if (type == LE_FIT_TYPE) {
1412 * Split left edge of fit VA.
1418 va->va_start += size;
1419 } else if (type == RE_FIT_TYPE) {
1421 * Split right edge of fit VA.
1427 va->va_end = nva_start_addr;
1428 } else if (type == NE_FIT_TYPE) {
1430 * Split no edge of fit VA.
1436 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1437 if (unlikely(!lva)) {
1439 * For percpu allocator we do not do any pre-allocation
1440 * and leave it as it is. The reason is it most likely
1441 * never ends up with NE_FIT_TYPE splitting. In case of
1442 * percpu allocations offsets and sizes are aligned to
1443 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1444 * are its main fitting cases.
1446 * There are a few exceptions though, as an example it is
1447 * a first allocation (early boot up) when we have "one"
1448 * big free space that has to be split.
1450 * Also we can hit this path in case of regular "vmap"
1451 * allocations, if "this" current CPU was not preloaded.
1452 * See the comment in alloc_vmap_area() why. If so, then
1453 * GFP_NOWAIT is used instead to get an extra object for
1454 * split purpose. That is rare and most time does not
1457 * What happens if an allocation gets failed. Basically,
1458 * an "overflow" path is triggered to purge lazily freed
1459 * areas to free some memory, then, the "retry" path is
1460 * triggered to repeat one more time. See more details
1461 * in alloc_vmap_area() function.
1463 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1469 * Build the remainder.
1471 lva->va_start = va->va_start;
1472 lva->va_end = nva_start_addr;
1475 * Shrink this VA to remaining size.
1477 va->va_start = nva_start_addr + size;
1482 if (type != FL_FIT_TYPE) {
1483 augment_tree_propagate_from(va);
1485 if (lva) /* type == NE_FIT_TYPE */
1486 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1493 * Returns a start address of the newly allocated area, if success.
1494 * Otherwise a vend is returned that indicates failure.
1496 static __always_inline unsigned long
1497 __alloc_vmap_area(struct rb_root *root, struct list_head *head,
1498 unsigned long size, unsigned long align,
1499 unsigned long vstart, unsigned long vend)
1501 bool adjust_search_size = true;
1502 unsigned long nva_start_addr;
1503 struct vmap_area *va;
1507 * Do not adjust when:
1508 * a) align <= PAGE_SIZE, because it does not make any sense.
1509 * All blocks(their start addresses) are at least PAGE_SIZE
1511 * b) a short range where a requested size corresponds to exactly
1512 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1513 * With adjusted search length an allocation would not succeed.
1515 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1516 adjust_search_size = false;
1518 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1522 if (va->va_start > vstart)
1523 nva_start_addr = ALIGN(va->va_start, align);
1525 nva_start_addr = ALIGN(vstart, align);
1527 /* Check the "vend" restriction. */
1528 if (nva_start_addr + size > vend)
1531 /* Update the free vmap_area. */
1532 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1533 if (WARN_ON_ONCE(ret))
1536 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1537 find_vmap_lowest_match_check(root, head, size, align);
1540 return nva_start_addr;
1544 * Free a region of KVA allocated by alloc_vmap_area
1546 static void free_vmap_area(struct vmap_area *va)
1549 * Remove from the busy tree/list.
1551 spin_lock(&vmap_area_lock);
1552 unlink_va(va, &vmap_area_root);
1553 spin_unlock(&vmap_area_lock);
1556 * Insert/Merge it back to the free tree/list.
1558 spin_lock(&free_vmap_area_lock);
1559 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1560 spin_unlock(&free_vmap_area_lock);
1564 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1566 struct vmap_area *va = NULL;
1569 * Preload this CPU with one extra vmap_area object. It is used
1570 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1571 * a CPU that does an allocation is preloaded.
1573 * We do it in non-atomic context, thus it allows us to use more
1574 * permissive allocation masks to be more stable under low memory
1575 * condition and high memory pressure.
1577 if (!this_cpu_read(ne_fit_preload_node))
1578 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1582 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1583 kmem_cache_free(vmap_area_cachep, va);
1587 * Allocate a region of KVA of the specified size and alignment, within the
1590 static struct vmap_area *alloc_vmap_area(unsigned long size,
1591 unsigned long align,
1592 unsigned long vstart, unsigned long vend,
1593 int node, gfp_t gfp_mask)
1595 struct vmap_area *va;
1596 unsigned long freed;
1602 BUG_ON(offset_in_page(size));
1603 BUG_ON(!is_power_of_2(align));
1605 if (unlikely(!vmap_initialized))
1606 return ERR_PTR(-EBUSY);
1609 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1611 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1613 return ERR_PTR(-ENOMEM);
1616 * Only scan the relevant parts containing pointers to other objects
1617 * to avoid false negatives.
1619 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1622 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1623 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1624 size, align, vstart, vend);
1625 spin_unlock(&free_vmap_area_lock);
1627 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1630 * If an allocation fails, the "vend" address is
1631 * returned. Therefore trigger the overflow path.
1633 if (unlikely(addr == vend))
1636 va->va_start = addr;
1637 va->va_end = addr + size;
1640 spin_lock(&vmap_area_lock);
1641 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1642 spin_unlock(&vmap_area_lock);
1644 BUG_ON(!IS_ALIGNED(va->va_start, align));
1645 BUG_ON(va->va_start < vstart);
1646 BUG_ON(va->va_end > vend);
1648 ret = kasan_populate_vmalloc(addr, size);
1651 return ERR_PTR(ret);
1658 purge_vmap_area_lazy();
1664 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1671 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1672 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1675 kmem_cache_free(vmap_area_cachep, va);
1676 return ERR_PTR(-EBUSY);
1679 int register_vmap_purge_notifier(struct notifier_block *nb)
1681 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1683 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1685 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1687 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1689 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1692 * lazy_max_pages is the maximum amount of virtual address space we gather up
1693 * before attempting to purge with a TLB flush.
1695 * There is a tradeoff here: a larger number will cover more kernel page tables
1696 * and take slightly longer to purge, but it will linearly reduce the number of
1697 * global TLB flushes that must be performed. It would seem natural to scale
1698 * this number up linearly with the number of CPUs (because vmapping activity
1699 * could also scale linearly with the number of CPUs), however it is likely
1700 * that in practice, workloads might be constrained in other ways that mean
1701 * vmap activity will not scale linearly with CPUs. Also, I want to be
1702 * conservative and not introduce a big latency on huge systems, so go with
1703 * a less aggressive log scale. It will still be an improvement over the old
1704 * code, and it will be simple to change the scale factor if we find that it
1705 * becomes a problem on bigger systems.
1707 static unsigned long lazy_max_pages(void)
1711 log = fls(num_online_cpus());
1713 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1716 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1719 * Serialize vmap purging. There is no actual critical section protected
1720 * by this lock, but we want to avoid concurrent calls for performance
1721 * reasons and to make the pcpu_get_vm_areas more deterministic.
1723 static DEFINE_MUTEX(vmap_purge_lock);
1725 /* for per-CPU blocks */
1726 static void purge_fragmented_blocks_allcpus(void);
1729 * Purges all lazily-freed vmap areas.
1731 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1733 unsigned long resched_threshold;
1734 unsigned int num_purged_areas = 0;
1735 struct list_head local_purge_list;
1736 struct vmap_area *va, *n_va;
1738 lockdep_assert_held(&vmap_purge_lock);
1740 spin_lock(&purge_vmap_area_lock);
1741 purge_vmap_area_root = RB_ROOT;
1742 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1743 spin_unlock(&purge_vmap_area_lock);
1745 if (unlikely(list_empty(&local_purge_list)))
1749 list_first_entry(&local_purge_list,
1750 struct vmap_area, list)->va_start);
1753 list_last_entry(&local_purge_list,
1754 struct vmap_area, list)->va_end);
1756 flush_tlb_kernel_range(start, end);
1757 resched_threshold = lazy_max_pages() << 1;
1759 spin_lock(&free_vmap_area_lock);
1760 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1761 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1762 unsigned long orig_start = va->va_start;
1763 unsigned long orig_end = va->va_end;
1766 * Finally insert or merge lazily-freed area. It is
1767 * detached and there is no need to "unlink" it from
1770 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1771 &free_vmap_area_list);
1776 if (is_vmalloc_or_module_addr((void *)orig_start))
1777 kasan_release_vmalloc(orig_start, orig_end,
1778 va->va_start, va->va_end);
1780 atomic_long_sub(nr, &vmap_lazy_nr);
1783 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1784 cond_resched_lock(&free_vmap_area_lock);
1786 spin_unlock(&free_vmap_area_lock);
1789 trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1790 return num_purged_areas > 0;
1794 * Kick off a purge of the outstanding lazy areas.
1796 static void purge_vmap_area_lazy(void)
1798 mutex_lock(&vmap_purge_lock);
1799 purge_fragmented_blocks_allcpus();
1800 __purge_vmap_area_lazy(ULONG_MAX, 0);
1801 mutex_unlock(&vmap_purge_lock);
1804 static void drain_vmap_area_work(struct work_struct *work)
1806 unsigned long nr_lazy;
1809 mutex_lock(&vmap_purge_lock);
1810 __purge_vmap_area_lazy(ULONG_MAX, 0);
1811 mutex_unlock(&vmap_purge_lock);
1813 /* Recheck if further work is required. */
1814 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1815 } while (nr_lazy > lazy_max_pages());
1819 * Free a vmap area, caller ensuring that the area has been unmapped
1820 * and flush_cache_vunmap had been called for the correct range
1823 static void free_vmap_area_noflush(struct vmap_area *va)
1825 unsigned long nr_lazy_max = lazy_max_pages();
1826 unsigned long va_start = va->va_start;
1827 unsigned long nr_lazy;
1829 spin_lock(&vmap_area_lock);
1830 unlink_va(va, &vmap_area_root);
1831 spin_unlock(&vmap_area_lock);
1833 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1834 PAGE_SHIFT, &vmap_lazy_nr);
1837 * Merge or place it to the purge tree/list.
1839 spin_lock(&purge_vmap_area_lock);
1840 merge_or_add_vmap_area(va,
1841 &purge_vmap_area_root, &purge_vmap_area_list);
1842 spin_unlock(&purge_vmap_area_lock);
1844 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1846 /* After this point, we may free va at any time */
1847 if (unlikely(nr_lazy > nr_lazy_max))
1848 schedule_work(&drain_vmap_work);
1852 * Free and unmap a vmap area
1854 static void free_unmap_vmap_area(struct vmap_area *va)
1856 flush_cache_vunmap(va->va_start, va->va_end);
1857 vunmap_range_noflush(va->va_start, va->va_end);
1858 if (debug_pagealloc_enabled_static())
1859 flush_tlb_kernel_range(va->va_start, va->va_end);
1861 free_vmap_area_noflush(va);
1864 struct vmap_area *find_vmap_area(unsigned long addr)
1866 struct vmap_area *va;
1868 spin_lock(&vmap_area_lock);
1869 va = __find_vmap_area(addr, &vmap_area_root);
1870 spin_unlock(&vmap_area_lock);
1875 /*** Per cpu kva allocator ***/
1878 * vmap space is limited especially on 32 bit architectures. Ensure there is
1879 * room for at least 16 percpu vmap blocks per CPU.
1882 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1883 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1884 * instead (we just need a rough idea)
1886 #if BITS_PER_LONG == 32
1887 #define VMALLOC_SPACE (128UL*1024*1024)
1889 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1892 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1893 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1894 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1895 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1896 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1897 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1898 #define VMAP_BBMAP_BITS \
1899 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1900 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1901 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1903 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1905 struct vmap_block_queue {
1907 struct list_head free;
1912 struct vmap_area *va;
1913 unsigned long free, dirty;
1914 unsigned long dirty_min, dirty_max; /*< dirty range */
1915 struct list_head free_list;
1916 struct rcu_head rcu_head;
1917 struct list_head purge;
1920 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1921 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1924 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1925 * in the free path. Could get rid of this if we change the API to return a
1926 * "cookie" from alloc, to be passed to free. But no big deal yet.
1928 static DEFINE_XARRAY(vmap_blocks);
1931 * We should probably have a fallback mechanism to allocate virtual memory
1932 * out of partially filled vmap blocks. However vmap block sizing should be
1933 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1937 static unsigned long addr_to_vb_idx(unsigned long addr)
1939 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1940 addr /= VMAP_BLOCK_SIZE;
1944 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1948 addr = va_start + (pages_off << PAGE_SHIFT);
1949 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1950 return (void *)addr;
1954 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1955 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1956 * @order: how many 2^order pages should be occupied in newly allocated block
1957 * @gfp_mask: flags for the page level allocator
1959 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1961 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1963 struct vmap_block_queue *vbq;
1964 struct vmap_block *vb;
1965 struct vmap_area *va;
1966 unsigned long vb_idx;
1970 node = numa_node_id();
1972 vb = kmalloc_node(sizeof(struct vmap_block),
1973 gfp_mask & GFP_RECLAIM_MASK, node);
1975 return ERR_PTR(-ENOMEM);
1977 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1978 VMALLOC_START, VMALLOC_END,
1982 return ERR_CAST(va);
1985 vaddr = vmap_block_vaddr(va->va_start, 0);
1986 spin_lock_init(&vb->lock);
1988 /* At least something should be left free */
1989 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1990 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1992 vb->dirty_min = VMAP_BBMAP_BITS;
1994 INIT_LIST_HEAD(&vb->free_list);
1996 vb_idx = addr_to_vb_idx(va->va_start);
1997 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
2001 return ERR_PTR(err);
2004 vbq = raw_cpu_ptr(&vmap_block_queue);
2005 spin_lock(&vbq->lock);
2006 list_add_tail_rcu(&vb->free_list, &vbq->free);
2007 spin_unlock(&vbq->lock);
2012 static void free_vmap_block(struct vmap_block *vb)
2014 struct vmap_block *tmp;
2016 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2019 free_vmap_area_noflush(vb->va);
2020 kfree_rcu(vb, rcu_head);
2023 static void purge_fragmented_blocks(int cpu)
2026 struct vmap_block *vb;
2027 struct vmap_block *n_vb;
2028 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2031 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2033 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2036 spin_lock(&vb->lock);
2037 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2038 vb->free = 0; /* prevent further allocs after releasing lock */
2039 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2041 vb->dirty_max = VMAP_BBMAP_BITS;
2042 spin_lock(&vbq->lock);
2043 list_del_rcu(&vb->free_list);
2044 spin_unlock(&vbq->lock);
2045 spin_unlock(&vb->lock);
2046 list_add_tail(&vb->purge, &purge);
2048 spin_unlock(&vb->lock);
2052 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2053 list_del(&vb->purge);
2054 free_vmap_block(vb);
2058 static void purge_fragmented_blocks_allcpus(void)
2062 for_each_possible_cpu(cpu)
2063 purge_fragmented_blocks(cpu);
2066 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2068 struct vmap_block_queue *vbq;
2069 struct vmap_block *vb;
2073 BUG_ON(offset_in_page(size));
2074 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2075 if (WARN_ON(size == 0)) {
2077 * Allocating 0 bytes isn't what caller wants since
2078 * get_order(0) returns funny result. Just warn and terminate
2083 order = get_order(size);
2086 vbq = raw_cpu_ptr(&vmap_block_queue);
2087 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2088 unsigned long pages_off;
2090 spin_lock(&vb->lock);
2091 if (vb->free < (1UL << order)) {
2092 spin_unlock(&vb->lock);
2096 pages_off = VMAP_BBMAP_BITS - vb->free;
2097 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2098 vb->free -= 1UL << order;
2099 if (vb->free == 0) {
2100 spin_lock(&vbq->lock);
2101 list_del_rcu(&vb->free_list);
2102 spin_unlock(&vbq->lock);
2105 spin_unlock(&vb->lock);
2111 /* Allocate new block if nothing was found */
2113 vaddr = new_vmap_block(order, gfp_mask);
2118 static void vb_free(unsigned long addr, unsigned long size)
2120 unsigned long offset;
2122 struct vmap_block *vb;
2124 BUG_ON(offset_in_page(size));
2125 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2127 flush_cache_vunmap(addr, addr + size);
2129 order = get_order(size);
2130 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2131 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2133 vunmap_range_noflush(addr, addr + size);
2135 if (debug_pagealloc_enabled_static())
2136 flush_tlb_kernel_range(addr, addr + size);
2138 spin_lock(&vb->lock);
2140 /* Expand dirty range */
2141 vb->dirty_min = min(vb->dirty_min, offset);
2142 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2144 vb->dirty += 1UL << order;
2145 if (vb->dirty == VMAP_BBMAP_BITS) {
2147 spin_unlock(&vb->lock);
2148 free_vmap_block(vb);
2150 spin_unlock(&vb->lock);
2153 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2157 if (unlikely(!vmap_initialized))
2162 for_each_possible_cpu(cpu) {
2163 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2164 struct vmap_block *vb;
2167 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2168 spin_lock(&vb->lock);
2169 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2170 unsigned long va_start = vb->va->va_start;
2173 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2174 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2176 start = min(s, start);
2181 spin_unlock(&vb->lock);
2186 mutex_lock(&vmap_purge_lock);
2187 purge_fragmented_blocks_allcpus();
2188 if (!__purge_vmap_area_lazy(start, end) && flush)
2189 flush_tlb_kernel_range(start, end);
2190 mutex_unlock(&vmap_purge_lock);
2194 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2196 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2197 * to amortize TLB flushing overheads. What this means is that any page you
2198 * have now, may, in a former life, have been mapped into kernel virtual
2199 * address by the vmap layer and so there might be some CPUs with TLB entries
2200 * still referencing that page (additional to the regular 1:1 kernel mapping).
2202 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2203 * be sure that none of the pages we have control over will have any aliases
2204 * from the vmap layer.
2206 void vm_unmap_aliases(void)
2208 unsigned long start = ULONG_MAX, end = 0;
2211 _vm_unmap_aliases(start, end, flush);
2213 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2216 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2217 * @mem: the pointer returned by vm_map_ram
2218 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2220 void vm_unmap_ram(const void *mem, unsigned int count)
2222 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2223 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2224 struct vmap_area *va;
2228 BUG_ON(addr < VMALLOC_START);
2229 BUG_ON(addr > VMALLOC_END);
2230 BUG_ON(!PAGE_ALIGNED(addr));
2232 kasan_poison_vmalloc(mem, size);
2234 if (likely(count <= VMAP_MAX_ALLOC)) {
2235 debug_check_no_locks_freed(mem, size);
2236 vb_free(addr, size);
2240 va = find_vmap_area(addr);
2242 debug_check_no_locks_freed((void *)va->va_start,
2243 (va->va_end - va->va_start));
2244 free_unmap_vmap_area(va);
2246 EXPORT_SYMBOL(vm_unmap_ram);
2249 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2250 * @pages: an array of pointers to the pages to be mapped
2251 * @count: number of pages
2252 * @node: prefer to allocate data structures on this node
2254 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2255 * faster than vmap so it's good. But if you mix long-life and short-life
2256 * objects with vm_map_ram(), it could consume lots of address space through
2257 * fragmentation (especially on a 32bit machine). You could see failures in
2258 * the end. Please use this function for short-lived objects.
2260 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2262 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2264 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2268 if (likely(count <= VMAP_MAX_ALLOC)) {
2269 mem = vb_alloc(size, GFP_KERNEL);
2272 addr = (unsigned long)mem;
2274 struct vmap_area *va;
2275 va = alloc_vmap_area(size, PAGE_SIZE,
2276 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2280 addr = va->va_start;
2284 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2285 pages, PAGE_SHIFT) < 0) {
2286 vm_unmap_ram(mem, count);
2291 * Mark the pages as accessible, now that they are mapped.
2292 * With hardware tag-based KASAN, marking is skipped for
2293 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2295 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2299 EXPORT_SYMBOL(vm_map_ram);
2301 static struct vm_struct *vmlist __initdata;
2303 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2305 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2306 return vm->page_order;
2312 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2314 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2315 vm->page_order = order;
2322 * vm_area_add_early - add vmap area early during boot
2323 * @vm: vm_struct to add
2325 * This function is used to add fixed kernel vm area to vmlist before
2326 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2327 * should contain proper values and the other fields should be zero.
2329 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2331 void __init vm_area_add_early(struct vm_struct *vm)
2333 struct vm_struct *tmp, **p;
2335 BUG_ON(vmap_initialized);
2336 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2337 if (tmp->addr >= vm->addr) {
2338 BUG_ON(tmp->addr < vm->addr + vm->size);
2341 BUG_ON(tmp->addr + tmp->size > vm->addr);
2348 * vm_area_register_early - register vmap area early during boot
2349 * @vm: vm_struct to register
2350 * @align: requested alignment
2352 * This function is used to register kernel vm area before
2353 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2354 * proper values on entry and other fields should be zero. On return,
2355 * vm->addr contains the allocated address.
2357 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2359 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2361 unsigned long addr = ALIGN(VMALLOC_START, align);
2362 struct vm_struct *cur, **p;
2364 BUG_ON(vmap_initialized);
2366 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2367 if ((unsigned long)cur->addr - addr >= vm->size)
2369 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2372 BUG_ON(addr > VMALLOC_END - vm->size);
2373 vm->addr = (void *)addr;
2376 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2379 static void vmap_init_free_space(void)
2381 unsigned long vmap_start = 1;
2382 const unsigned long vmap_end = ULONG_MAX;
2383 struct vmap_area *busy, *free;
2387 * -|-----|.....|-----|-----|-----|.....|-
2389 * |<--------------------------------->|
2391 list_for_each_entry(busy, &vmap_area_list, list) {
2392 if (busy->va_start - vmap_start > 0) {
2393 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2394 if (!WARN_ON_ONCE(!free)) {
2395 free->va_start = vmap_start;
2396 free->va_end = busy->va_start;
2398 insert_vmap_area_augment(free, NULL,
2399 &free_vmap_area_root,
2400 &free_vmap_area_list);
2404 vmap_start = busy->va_end;
2407 if (vmap_end - vmap_start > 0) {
2408 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2409 if (!WARN_ON_ONCE(!free)) {
2410 free->va_start = vmap_start;
2411 free->va_end = vmap_end;
2413 insert_vmap_area_augment(free, NULL,
2414 &free_vmap_area_root,
2415 &free_vmap_area_list);
2420 void __init vmalloc_init(void)
2422 struct vmap_area *va;
2423 struct vm_struct *tmp;
2427 * Create the cache for vmap_area objects.
2429 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2431 for_each_possible_cpu(i) {
2432 struct vmap_block_queue *vbq;
2433 struct vfree_deferred *p;
2435 vbq = &per_cpu(vmap_block_queue, i);
2436 spin_lock_init(&vbq->lock);
2437 INIT_LIST_HEAD(&vbq->free);
2438 p = &per_cpu(vfree_deferred, i);
2439 init_llist_head(&p->list);
2440 INIT_WORK(&p->wq, free_work);
2443 /* Import existing vmlist entries. */
2444 for (tmp = vmlist; tmp; tmp = tmp->next) {
2445 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2446 if (WARN_ON_ONCE(!va))
2449 va->va_start = (unsigned long)tmp->addr;
2450 va->va_end = va->va_start + tmp->size;
2452 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2456 * Now we can initialize a free vmap space.
2458 vmap_init_free_space();
2459 vmap_initialized = true;
2462 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2463 struct vmap_area *va, unsigned long flags, const void *caller)
2466 vm->addr = (void *)va->va_start;
2467 vm->size = va->va_end - va->va_start;
2468 vm->caller = caller;
2472 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2473 unsigned long flags, const void *caller)
2475 spin_lock(&vmap_area_lock);
2476 setup_vmalloc_vm_locked(vm, va, flags, caller);
2477 spin_unlock(&vmap_area_lock);
2480 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2483 * Before removing VM_UNINITIALIZED,
2484 * we should make sure that vm has proper values.
2485 * Pair with smp_rmb() in show_numa_info().
2488 vm->flags &= ~VM_UNINITIALIZED;
2491 static struct vm_struct *__get_vm_area_node(unsigned long size,
2492 unsigned long align, unsigned long shift, unsigned long flags,
2493 unsigned long start, unsigned long end, int node,
2494 gfp_t gfp_mask, const void *caller)
2496 struct vmap_area *va;
2497 struct vm_struct *area;
2498 unsigned long requested_size = size;
2500 BUG_ON(in_interrupt());
2501 size = ALIGN(size, 1ul << shift);
2502 if (unlikely(!size))
2505 if (flags & VM_IOREMAP)
2506 align = 1ul << clamp_t(int, get_count_order_long(size),
2507 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2509 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2510 if (unlikely(!area))
2513 if (!(flags & VM_NO_GUARD))
2516 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2522 setup_vmalloc_vm(area, va, flags, caller);
2525 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2526 * best-effort approach, as they can be mapped outside of vmalloc code.
2527 * For VM_ALLOC mappings, the pages are marked as accessible after
2528 * getting mapped in __vmalloc_node_range().
2529 * With hardware tag-based KASAN, marking is skipped for
2530 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2532 if (!(flags & VM_ALLOC))
2533 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2534 KASAN_VMALLOC_PROT_NORMAL);
2539 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2540 unsigned long start, unsigned long end,
2543 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2544 NUMA_NO_NODE, GFP_KERNEL, caller);
2548 * get_vm_area - reserve a contiguous kernel virtual area
2549 * @size: size of the area
2550 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2552 * Search an area of @size in the kernel virtual mapping area,
2553 * and reserved it for out purposes. Returns the area descriptor
2554 * on success or %NULL on failure.
2556 * Return: the area descriptor on success or %NULL on failure.
2558 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2560 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2561 VMALLOC_START, VMALLOC_END,
2562 NUMA_NO_NODE, GFP_KERNEL,
2563 __builtin_return_address(0));
2566 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2569 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2570 VMALLOC_START, VMALLOC_END,
2571 NUMA_NO_NODE, GFP_KERNEL, caller);
2575 * find_vm_area - find a continuous kernel virtual area
2576 * @addr: base address
2578 * Search for the kernel VM area starting at @addr, and return it.
2579 * It is up to the caller to do all required locking to keep the returned
2582 * Return: the area descriptor on success or %NULL on failure.
2584 struct vm_struct *find_vm_area(const void *addr)
2586 struct vmap_area *va;
2588 va = find_vmap_area((unsigned long)addr);
2596 * remove_vm_area - find and remove a continuous kernel virtual area
2597 * @addr: base address
2599 * Search for the kernel VM area starting at @addr, and remove it.
2600 * This function returns the found VM area, but using it is NOT safe
2601 * on SMP machines, except for its size or flags.
2603 * Return: the area descriptor on success or %NULL on failure.
2605 struct vm_struct *remove_vm_area(const void *addr)
2607 struct vmap_area *va;
2611 spin_lock(&vmap_area_lock);
2612 va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2614 struct vm_struct *vm = va->vm;
2617 spin_unlock(&vmap_area_lock);
2619 kasan_free_module_shadow(vm);
2620 free_unmap_vmap_area(va);
2625 spin_unlock(&vmap_area_lock);
2629 static inline void set_area_direct_map(const struct vm_struct *area,
2630 int (*set_direct_map)(struct page *page))
2634 /* HUGE_VMALLOC passes small pages to set_direct_map */
2635 for (i = 0; i < area->nr_pages; i++)
2636 if (page_address(area->pages[i]))
2637 set_direct_map(area->pages[i]);
2640 /* Handle removing and resetting vm mappings related to the vm_struct. */
2641 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2643 unsigned long start = ULONG_MAX, end = 0;
2644 unsigned int page_order = vm_area_page_order(area);
2645 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2649 remove_vm_area(area->addr);
2651 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2656 * If not deallocating pages, just do the flush of the VM area and
2659 if (!deallocate_pages) {
2665 * If execution gets here, flush the vm mapping and reset the direct
2666 * map. Find the start and end range of the direct mappings to make sure
2667 * the vm_unmap_aliases() flush includes the direct map.
2669 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2670 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2672 unsigned long page_size;
2674 page_size = PAGE_SIZE << page_order;
2675 start = min(addr, start);
2676 end = max(addr + page_size, end);
2682 * Set direct map to something invalid so that it won't be cached if
2683 * there are any accesses after the TLB flush, then flush the TLB and
2684 * reset the direct map permissions to the default.
2686 set_area_direct_map(area, set_direct_map_invalid_noflush);
2687 _vm_unmap_aliases(start, end, flush_dmap);
2688 set_area_direct_map(area, set_direct_map_default_noflush);
2691 static void __vunmap(const void *addr, int deallocate_pages)
2693 struct vm_struct *area;
2698 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2702 area = find_vm_area(addr);
2703 if (unlikely(!area)) {
2704 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2709 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2710 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2712 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2714 vm_remove_mappings(area, deallocate_pages);
2716 if (deallocate_pages) {
2719 for (i = 0; i < area->nr_pages; i++) {
2720 struct page *page = area->pages[i];
2723 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2725 * High-order allocs for huge vmallocs are split, so
2726 * can be freed as an array of order-0 allocations
2728 __free_pages(page, 0);
2731 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2733 kvfree(area->pages);
2739 static inline void __vfree_deferred(const void *addr)
2742 * Use raw_cpu_ptr() because this can be called from preemptible
2743 * context. Preemption is absolutely fine here, because the llist_add()
2744 * implementation is lockless, so it works even if we are adding to
2745 * another cpu's list. schedule_work() should be fine with this too.
2747 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2749 if (llist_add((struct llist_node *)addr, &p->list))
2750 schedule_work(&p->wq);
2754 * vfree_atomic - release memory allocated by vmalloc()
2755 * @addr: memory base address
2757 * This one is just like vfree() but can be called in any atomic context
2760 void vfree_atomic(const void *addr)
2764 kmemleak_free(addr);
2768 __vfree_deferred(addr);
2771 static void __vfree(const void *addr)
2773 if (unlikely(in_interrupt()))
2774 __vfree_deferred(addr);
2780 * vfree - Release memory allocated by vmalloc()
2781 * @addr: Memory base address
2783 * Free the virtually continuous memory area starting at @addr, as obtained
2784 * from one of the vmalloc() family of APIs. This will usually also free the
2785 * physical memory underlying the virtual allocation, but that memory is
2786 * reference counted, so it will not be freed until the last user goes away.
2788 * If @addr is NULL, no operation is performed.
2791 * May sleep if called *not* from interrupt context.
2792 * Must not be called in NMI context (strictly speaking, it could be
2793 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2794 * conventions for vfree() arch-dependent would be a really bad idea).
2796 void vfree(const void *addr)
2800 kmemleak_free(addr);
2802 might_sleep_if(!in_interrupt());
2809 EXPORT_SYMBOL(vfree);
2812 * vunmap - release virtual mapping obtained by vmap()
2813 * @addr: memory base address
2815 * Free the virtually contiguous memory area starting at @addr,
2816 * which was created from the page array passed to vmap().
2818 * Must not be called in interrupt context.
2820 void vunmap(const void *addr)
2822 BUG_ON(in_interrupt());
2827 EXPORT_SYMBOL(vunmap);
2830 * vmap - map an array of pages into virtually contiguous space
2831 * @pages: array of page pointers
2832 * @count: number of pages to map
2833 * @flags: vm_area->flags
2834 * @prot: page protection for the mapping
2836 * Maps @count pages from @pages into contiguous kernel virtual space.
2837 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2838 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2839 * are transferred from the caller to vmap(), and will be freed / dropped when
2840 * vfree() is called on the return value.
2842 * Return: the address of the area or %NULL on failure
2844 void *vmap(struct page **pages, unsigned int count,
2845 unsigned long flags, pgprot_t prot)
2847 struct vm_struct *area;
2849 unsigned long size; /* In bytes */
2854 * Your top guard is someone else's bottom guard. Not having a top
2855 * guard compromises someone else's mappings too.
2857 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2858 flags &= ~VM_NO_GUARD;
2860 if (count > totalram_pages())
2863 size = (unsigned long)count << PAGE_SHIFT;
2864 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2868 addr = (unsigned long)area->addr;
2869 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2870 pages, PAGE_SHIFT) < 0) {
2875 if (flags & VM_MAP_PUT_PAGES) {
2876 area->pages = pages;
2877 area->nr_pages = count;
2881 EXPORT_SYMBOL(vmap);
2883 #ifdef CONFIG_VMAP_PFN
2884 struct vmap_pfn_data {
2885 unsigned long *pfns;
2890 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2892 struct vmap_pfn_data *data = private;
2894 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2896 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2901 * vmap_pfn - map an array of PFNs into virtually contiguous space
2902 * @pfns: array of PFNs
2903 * @count: number of pages to map
2904 * @prot: page protection for the mapping
2906 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2907 * the start address of the mapping.
2909 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2911 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2912 struct vm_struct *area;
2914 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2915 __builtin_return_address(0));
2918 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2919 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2925 EXPORT_SYMBOL_GPL(vmap_pfn);
2926 #endif /* CONFIG_VMAP_PFN */
2928 static inline unsigned int
2929 vm_area_alloc_pages(gfp_t gfp, int nid,
2930 unsigned int order, unsigned int nr_pages, struct page **pages)
2932 unsigned int nr_allocated = 0;
2937 * For order-0 pages we make use of bulk allocator, if
2938 * the page array is partly or not at all populated due
2939 * to fails, fallback to a single page allocator that is
2943 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2945 while (nr_allocated < nr_pages) {
2946 unsigned int nr, nr_pages_request;
2949 * A maximum allowed request is hard-coded and is 100
2950 * pages per call. That is done in order to prevent a
2951 * long preemption off scenario in the bulk-allocator
2952 * so the range is [1:100].
2954 nr_pages_request = min(100U, nr_pages - nr_allocated);
2956 /* memory allocation should consider mempolicy, we can't
2957 * wrongly use nearest node when nid == NUMA_NO_NODE,
2958 * otherwise memory may be allocated in only one node,
2959 * but mempolicy wants to alloc memory by interleaving.
2961 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2962 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2964 pages + nr_allocated);
2967 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2969 pages + nr_allocated);
2975 * If zero or pages were obtained partly,
2976 * fallback to a single page allocator.
2978 if (nr != nr_pages_request)
2983 /* High-order pages or fallback path if "bulk" fails. */
2985 while (nr_allocated < nr_pages) {
2986 if (fatal_signal_pending(current))
2989 if (nid == NUMA_NO_NODE)
2990 page = alloc_pages(gfp, order);
2992 page = alloc_pages_node(nid, gfp, order);
2993 if (unlikely(!page))
2996 * Higher order allocations must be able to be treated as
2997 * indepdenent small pages by callers (as they can with
2998 * small-page vmallocs). Some drivers do their own refcounting
2999 * on vmalloc_to_page() pages, some use page->mapping,
3003 split_page(page, order);
3006 * Careful, we allocate and map page-order pages, but
3007 * tracking is done per PAGE_SIZE page so as to keep the
3008 * vm_struct APIs independent of the physical/mapped size.
3010 for (i = 0; i < (1U << order); i++)
3011 pages[nr_allocated + i] = page + i;
3014 nr_allocated += 1U << order;
3017 return nr_allocated;
3020 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3021 pgprot_t prot, unsigned int page_shift,
3024 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3025 bool nofail = gfp_mask & __GFP_NOFAIL;
3026 unsigned long addr = (unsigned long)area->addr;
3027 unsigned long size = get_vm_area_size(area);
3028 unsigned long array_size;
3029 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3030 unsigned int page_order;
3034 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3035 gfp_mask |= __GFP_NOWARN;
3036 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3037 gfp_mask |= __GFP_HIGHMEM;
3039 /* Please note that the recursion is strictly bounded. */
3040 if (array_size > PAGE_SIZE) {
3041 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3044 area->pages = kmalloc_node(array_size, nested_gfp, node);
3048 warn_alloc(gfp_mask, NULL,
3049 "vmalloc error: size %lu, failed to allocated page array size %lu",
3050 nr_small_pages * PAGE_SIZE, array_size);
3055 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3056 page_order = vm_area_page_order(area);
3058 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3059 node, page_order, nr_small_pages, area->pages);
3061 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3062 if (gfp_mask & __GFP_ACCOUNT) {
3065 for (i = 0; i < area->nr_pages; i++)
3066 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3070 * If not enough pages were obtained to accomplish an
3071 * allocation request, free them via __vfree() if any.
3073 if (area->nr_pages != nr_small_pages) {
3074 warn_alloc(gfp_mask, NULL,
3075 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3076 area->nr_pages * PAGE_SIZE, page_order);
3081 * page tables allocations ignore external gfp mask, enforce it
3084 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3085 flags = memalloc_nofs_save();
3086 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3087 flags = memalloc_noio_save();
3090 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3092 if (nofail && (ret < 0))
3093 schedule_timeout_uninterruptible(1);
3094 } while (nofail && (ret < 0));
3096 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3097 memalloc_nofs_restore(flags);
3098 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3099 memalloc_noio_restore(flags);
3102 warn_alloc(gfp_mask, NULL,
3103 "vmalloc error: size %lu, failed to map pages",
3104 area->nr_pages * PAGE_SIZE);
3111 __vfree(area->addr);
3116 * __vmalloc_node_range - allocate virtually contiguous memory
3117 * @size: allocation size
3118 * @align: desired alignment
3119 * @start: vm area range start
3120 * @end: vm area range end
3121 * @gfp_mask: flags for the page level allocator
3122 * @prot: protection mask for the allocated pages
3123 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3124 * @node: node to use for allocation or NUMA_NO_NODE
3125 * @caller: caller's return address
3127 * Allocate enough pages to cover @size from the page level
3128 * allocator with @gfp_mask flags. Please note that the full set of gfp
3129 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3131 * Zone modifiers are not supported. From the reclaim modifiers
3132 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3133 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3134 * __GFP_RETRY_MAYFAIL are not supported).
3136 * __GFP_NOWARN can be used to suppress failures messages.
3138 * Map them into contiguous kernel virtual space, using a pagetable
3139 * protection of @prot.
3141 * Return: the address of the area or %NULL on failure
3143 void *__vmalloc_node_range(unsigned long size, unsigned long align,
3144 unsigned long start, unsigned long end, gfp_t gfp_mask,
3145 pgprot_t prot, unsigned long vm_flags, int node,
3148 struct vm_struct *area;
3150 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3151 unsigned long real_size = size;
3152 unsigned long real_align = align;
3153 unsigned int shift = PAGE_SHIFT;
3155 if (WARN_ON_ONCE(!size))
3158 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3159 warn_alloc(gfp_mask, NULL,
3160 "vmalloc error: size %lu, exceeds total pages",
3165 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3166 unsigned long size_per_node;
3169 * Try huge pages. Only try for PAGE_KERNEL allocations,
3170 * others like modules don't yet expect huge pages in
3171 * their allocations due to apply_to_page_range not
3175 size_per_node = size;
3176 if (node == NUMA_NO_NODE)
3177 size_per_node /= num_online_nodes();
3178 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3181 shift = arch_vmap_pte_supported_shift(size_per_node);
3183 align = max(real_align, 1UL << shift);
3184 size = ALIGN(real_size, 1UL << shift);
3188 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3189 VM_UNINITIALIZED | vm_flags, start, end, node,
3192 bool nofail = gfp_mask & __GFP_NOFAIL;
3193 warn_alloc(gfp_mask, NULL,
3194 "vmalloc error: size %lu, vm_struct allocation failed%s",
3195 real_size, (nofail) ? ". Retrying." : "");
3197 schedule_timeout_uninterruptible(1);
3204 * Prepare arguments for __vmalloc_area_node() and
3205 * kasan_unpoison_vmalloc().
3207 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3208 if (kasan_hw_tags_enabled()) {
3210 * Modify protection bits to allow tagging.
3211 * This must be done before mapping.
3213 prot = arch_vmap_pgprot_tagged(prot);
3216 * Skip page_alloc poisoning and zeroing for physical
3217 * pages backing VM_ALLOC mapping. Memory is instead
3218 * poisoned and zeroed by kasan_unpoison_vmalloc().
3220 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3223 /* Take note that the mapping is PAGE_KERNEL. */
3224 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3227 /* Allocate physical pages and map them into vmalloc space. */
3228 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3233 * Mark the pages as accessible, now that they are mapped.
3234 * The condition for setting KASAN_VMALLOC_INIT should complement the
3235 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3236 * to make sure that memory is initialized under the same conditions.
3237 * Tag-based KASAN modes only assign tags to normal non-executable
3238 * allocations, see __kasan_unpoison_vmalloc().
3240 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3241 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3242 (gfp_mask & __GFP_SKIP_ZERO))
3243 kasan_flags |= KASAN_VMALLOC_INIT;
3244 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3245 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3248 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3249 * flag. It means that vm_struct is not fully initialized.
3250 * Now, it is fully initialized, so remove this flag here.
3252 clear_vm_uninitialized_flag(area);
3254 size = PAGE_ALIGN(size);
3255 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3256 kmemleak_vmalloc(area, size, gfp_mask);
3261 if (shift > PAGE_SHIFT) {
3272 * __vmalloc_node - allocate virtually contiguous memory
3273 * @size: allocation size
3274 * @align: desired alignment
3275 * @gfp_mask: flags for the page level allocator
3276 * @node: node to use for allocation or NUMA_NO_NODE
3277 * @caller: caller's return address
3279 * Allocate enough pages to cover @size from the page level allocator with
3280 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3282 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3283 * and __GFP_NOFAIL are not supported
3285 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3288 * Return: pointer to the allocated memory or %NULL on error
3290 void *__vmalloc_node(unsigned long size, unsigned long align,
3291 gfp_t gfp_mask, int node, const void *caller)
3293 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3294 gfp_mask, PAGE_KERNEL, 0, node, caller);
3297 * This is only for performance analysis of vmalloc and stress purpose.
3298 * It is required by vmalloc test module, therefore do not use it other
3301 #ifdef CONFIG_TEST_VMALLOC_MODULE
3302 EXPORT_SYMBOL_GPL(__vmalloc_node);
3305 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3307 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3308 __builtin_return_address(0));
3310 EXPORT_SYMBOL(__vmalloc);
3313 * vmalloc - allocate virtually contiguous memory
3314 * @size: allocation size
3316 * Allocate enough pages to cover @size from the page level
3317 * allocator and map them into contiguous kernel virtual space.
3319 * For tight control over page level allocator and protection flags
3320 * use __vmalloc() instead.
3322 * Return: pointer to the allocated memory or %NULL on error
3324 void *vmalloc(unsigned long size)
3326 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3327 __builtin_return_address(0));
3329 EXPORT_SYMBOL(vmalloc);
3332 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3333 * @size: allocation size
3334 * @gfp_mask: flags for the page level allocator
3336 * Allocate enough pages to cover @size from the page level
3337 * allocator and map them into contiguous kernel virtual space.
3338 * If @size is greater than or equal to PMD_SIZE, allow using
3339 * huge pages for the memory
3341 * Return: pointer to the allocated memory or %NULL on error
3343 void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3345 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3346 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3347 NUMA_NO_NODE, __builtin_return_address(0));
3349 EXPORT_SYMBOL_GPL(vmalloc_huge);
3352 * vzalloc - allocate virtually contiguous memory with zero fill
3353 * @size: allocation size
3355 * Allocate enough pages to cover @size from the page level
3356 * allocator and map them into contiguous kernel virtual space.
3357 * The memory allocated is set to zero.
3359 * For tight control over page level allocator and protection flags
3360 * use __vmalloc() instead.
3362 * Return: pointer to the allocated memory or %NULL on error
3364 void *vzalloc(unsigned long size)
3366 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3367 __builtin_return_address(0));
3369 EXPORT_SYMBOL(vzalloc);
3372 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3373 * @size: allocation size
3375 * The resulting memory area is zeroed so it can be mapped to userspace
3376 * without leaking data.
3378 * Return: pointer to the allocated memory or %NULL on error
3380 void *vmalloc_user(unsigned long size)
3382 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3383 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3384 VM_USERMAP, NUMA_NO_NODE,
3385 __builtin_return_address(0));
3387 EXPORT_SYMBOL(vmalloc_user);
3390 * vmalloc_node - allocate memory on a specific node
3391 * @size: allocation size
3394 * Allocate enough pages to cover @size from the page level
3395 * allocator and map them into contiguous kernel virtual space.
3397 * For tight control over page level allocator and protection flags
3398 * use __vmalloc() instead.
3400 * Return: pointer to the allocated memory or %NULL on error
3402 void *vmalloc_node(unsigned long size, int node)
3404 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3405 __builtin_return_address(0));
3407 EXPORT_SYMBOL(vmalloc_node);
3410 * vzalloc_node - allocate memory on a specific node with zero fill
3411 * @size: allocation size
3414 * Allocate enough pages to cover @size from the page level
3415 * allocator and map them into contiguous kernel virtual space.
3416 * The memory allocated is set to zero.
3418 * Return: pointer to the allocated memory or %NULL on error
3420 void *vzalloc_node(unsigned long size, int node)
3422 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3423 __builtin_return_address(0));
3425 EXPORT_SYMBOL(vzalloc_node);
3427 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3428 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3429 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3430 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3433 * 64b systems should always have either DMA or DMA32 zones. For others
3434 * GFP_DMA32 should do the right thing and use the normal zone.
3436 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3440 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3441 * @size: allocation size
3443 * Allocate enough 32bit PA addressable pages to cover @size from the
3444 * page level allocator and map them into contiguous kernel virtual space.
3446 * Return: pointer to the allocated memory or %NULL on error
3448 void *vmalloc_32(unsigned long size)
3450 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3451 __builtin_return_address(0));
3453 EXPORT_SYMBOL(vmalloc_32);
3456 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3457 * @size: allocation size
3459 * The resulting memory area is 32bit addressable and zeroed so it can be
3460 * mapped to userspace without leaking data.
3462 * Return: pointer to the allocated memory or %NULL on error
3464 void *vmalloc_32_user(unsigned long size)
3466 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3467 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3468 VM_USERMAP, NUMA_NO_NODE,
3469 __builtin_return_address(0));
3471 EXPORT_SYMBOL(vmalloc_32_user);
3474 * small helper routine , copy contents to buf from addr.
3475 * If the page is not present, fill zero.
3478 static int aligned_vread(char *buf, char *addr, unsigned long count)
3484 unsigned long offset, length;
3486 offset = offset_in_page(addr);
3487 length = PAGE_SIZE - offset;
3490 p = vmalloc_to_page(addr);
3492 * To do safe access to this _mapped_ area, we need
3493 * lock. But adding lock here means that we need to add
3494 * overhead of vmalloc()/vfree() calls for this _debug_
3495 * interface, rarely used. Instead of that, we'll use
3496 * kmap() and get small overhead in this access function.
3499 /* We can expect USER0 is not used -- see vread() */
3500 void *map = kmap_atomic(p);
3501 memcpy(buf, map + offset, length);
3504 memset(buf, 0, length);
3515 * vread() - read vmalloc area in a safe way.
3516 * @buf: buffer for reading data
3517 * @addr: vm address.
3518 * @count: number of bytes to be read.
3520 * This function checks that addr is a valid vmalloc'ed area, and
3521 * copy data from that area to a given buffer. If the given memory range
3522 * of [addr...addr+count) includes some valid address, data is copied to
3523 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3524 * IOREMAP area is treated as memory hole and no copy is done.
3526 * If [addr...addr+count) doesn't includes any intersects with alive
3527 * vm_struct area, returns 0. @buf should be kernel's buffer.
3529 * Note: In usual ops, vread() is never necessary because the caller
3530 * should know vmalloc() area is valid and can use memcpy().
3531 * This is for routines which have to access vmalloc area without
3532 * any information, as /proc/kcore.
3534 * Return: number of bytes for which addr and buf should be increased
3535 * (same number as @count) or %0 if [addr...addr+count) doesn't
3536 * include any intersection with valid vmalloc area
3538 long vread(char *buf, char *addr, unsigned long count)
3540 struct vmap_area *va;
3541 struct vm_struct *vm;
3542 char *vaddr, *buf_start = buf;
3543 unsigned long buflen = count;
3546 addr = kasan_reset_tag(addr);
3548 /* Don't allow overflow */
3549 if ((unsigned long) addr + count < count)
3550 count = -(unsigned long) addr;
3552 spin_lock(&vmap_area_lock);
3553 va = find_vmap_area_exceed_addr((unsigned long)addr);
3557 /* no intersects with alive vmap_area */
3558 if ((unsigned long)addr + count <= va->va_start)
3561 list_for_each_entry_from(va, &vmap_area_list, list) {
3569 vaddr = (char *) vm->addr;
3570 if (addr >= vaddr + get_vm_area_size(vm))
3572 while (addr < vaddr) {
3580 n = vaddr + get_vm_area_size(vm) - addr;
3583 if (!(vm->flags & VM_IOREMAP))
3584 aligned_vread(buf, addr, n);
3585 else /* IOREMAP area is treated as memory hole */
3592 spin_unlock(&vmap_area_lock);
3594 if (buf == buf_start)
3596 /* zero-fill memory holes */
3597 if (buf != buf_start + buflen)
3598 memset(buf, 0, buflen - (buf - buf_start));
3604 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3605 * @vma: vma to cover
3606 * @uaddr: target user address to start at
3607 * @kaddr: virtual address of vmalloc kernel memory
3608 * @pgoff: offset from @kaddr to start at
3609 * @size: size of map area
3611 * Returns: 0 for success, -Exxx on failure
3613 * This function checks that @kaddr is a valid vmalloc'ed area,
3614 * and that it is big enough to cover the range starting at
3615 * @uaddr in @vma. Will return failure if that criteria isn't
3618 * Similar to remap_pfn_range() (see mm/memory.c)
3620 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3621 void *kaddr, unsigned long pgoff,
3624 struct vm_struct *area;
3626 unsigned long end_index;
3628 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3631 size = PAGE_ALIGN(size);
3633 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3636 area = find_vm_area(kaddr);
3640 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3643 if (check_add_overflow(size, off, &end_index) ||
3644 end_index > get_vm_area_size(area))
3649 struct page *page = vmalloc_to_page(kaddr);
3652 ret = vm_insert_page(vma, uaddr, page);
3661 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3667 * remap_vmalloc_range - map vmalloc pages to userspace
3668 * @vma: vma to cover (map full range of vma)
3669 * @addr: vmalloc memory
3670 * @pgoff: number of pages into addr before first page to map
3672 * Returns: 0 for success, -Exxx on failure
3674 * This function checks that addr is a valid vmalloc'ed area, and
3675 * that it is big enough to cover the vma. Will return failure if
3676 * that criteria isn't met.
3678 * Similar to remap_pfn_range() (see mm/memory.c)
3680 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3681 unsigned long pgoff)
3683 return remap_vmalloc_range_partial(vma, vma->vm_start,
3685 vma->vm_end - vma->vm_start);
3687 EXPORT_SYMBOL(remap_vmalloc_range);
3689 void free_vm_area(struct vm_struct *area)
3691 struct vm_struct *ret;
3692 ret = remove_vm_area(area->addr);
3693 BUG_ON(ret != area);
3696 EXPORT_SYMBOL_GPL(free_vm_area);
3699 static struct vmap_area *node_to_va(struct rb_node *n)
3701 return rb_entry_safe(n, struct vmap_area, rb_node);
3705 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3706 * @addr: target address
3708 * Returns: vmap_area if it is found. If there is no such area
3709 * the first highest(reverse order) vmap_area is returned
3710 * i.e. va->va_start < addr && va->va_end < addr or NULL
3711 * if there are no any areas before @addr.
3713 static struct vmap_area *
3714 pvm_find_va_enclose_addr(unsigned long addr)
3716 struct vmap_area *va, *tmp;
3719 n = free_vmap_area_root.rb_node;
3723 tmp = rb_entry(n, struct vmap_area, rb_node);
3724 if (tmp->va_start <= addr) {
3726 if (tmp->va_end >= addr)
3739 * pvm_determine_end_from_reverse - find the highest aligned address
3740 * of free block below VMALLOC_END
3742 * in - the VA we start the search(reverse order);
3743 * out - the VA with the highest aligned end address.
3744 * @align: alignment for required highest address
3746 * Returns: determined end address within vmap_area
3748 static unsigned long
3749 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3751 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3755 list_for_each_entry_from_reverse((*va),
3756 &free_vmap_area_list, list) {
3757 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3758 if ((*va)->va_start < addr)
3767 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3768 * @offsets: array containing offset of each area
3769 * @sizes: array containing size of each area
3770 * @nr_vms: the number of areas to allocate
3771 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3773 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3774 * vm_structs on success, %NULL on failure
3776 * Percpu allocator wants to use congruent vm areas so that it can
3777 * maintain the offsets among percpu areas. This function allocates
3778 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3779 * be scattered pretty far, distance between two areas easily going up
3780 * to gigabytes. To avoid interacting with regular vmallocs, these
3781 * areas are allocated from top.
3783 * Despite its complicated look, this allocator is rather simple. It
3784 * does everything top-down and scans free blocks from the end looking
3785 * for matching base. While scanning, if any of the areas do not fit the
3786 * base address is pulled down to fit the area. Scanning is repeated till
3787 * all the areas fit and then all necessary data structures are inserted
3788 * and the result is returned.
3790 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3791 const size_t *sizes, int nr_vms,
3794 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3795 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3796 struct vmap_area **vas, *va;
3797 struct vm_struct **vms;
3798 int area, area2, last_area, term_area;
3799 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3800 bool purged = false;
3802 /* verify parameters and allocate data structures */
3803 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3804 for (last_area = 0, area = 0; area < nr_vms; area++) {
3805 start = offsets[area];
3806 end = start + sizes[area];
3808 /* is everything aligned properly? */
3809 BUG_ON(!IS_ALIGNED(offsets[area], align));
3810 BUG_ON(!IS_ALIGNED(sizes[area], align));
3812 /* detect the area with the highest address */
3813 if (start > offsets[last_area])
3816 for (area2 = area + 1; area2 < nr_vms; area2++) {
3817 unsigned long start2 = offsets[area2];
3818 unsigned long end2 = start2 + sizes[area2];
3820 BUG_ON(start2 < end && start < end2);
3823 last_end = offsets[last_area] + sizes[last_area];
3825 if (vmalloc_end - vmalloc_start < last_end) {
3830 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3831 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3835 for (area = 0; area < nr_vms; area++) {
3836 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3837 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3838 if (!vas[area] || !vms[area])
3842 spin_lock(&free_vmap_area_lock);
3844 /* start scanning - we scan from the top, begin with the last area */
3845 area = term_area = last_area;
3846 start = offsets[area];
3847 end = start + sizes[area];
3849 va = pvm_find_va_enclose_addr(vmalloc_end);
3850 base = pvm_determine_end_from_reverse(&va, align) - end;
3854 * base might have underflowed, add last_end before
3857 if (base + last_end < vmalloc_start + last_end)
3861 * Fitting base has not been found.
3867 * If required width exceeds current VA block, move
3868 * base downwards and then recheck.
3870 if (base + end > va->va_end) {
3871 base = pvm_determine_end_from_reverse(&va, align) - end;
3877 * If this VA does not fit, move base downwards and recheck.
3879 if (base + start < va->va_start) {
3880 va = node_to_va(rb_prev(&va->rb_node));
3881 base = pvm_determine_end_from_reverse(&va, align) - end;
3887 * This area fits, move on to the previous one. If
3888 * the previous one is the terminal one, we're done.
3890 area = (area + nr_vms - 1) % nr_vms;
3891 if (area == term_area)
3894 start = offsets[area];
3895 end = start + sizes[area];
3896 va = pvm_find_va_enclose_addr(base + end);
3899 /* we've found a fitting base, insert all va's */
3900 for (area = 0; area < nr_vms; area++) {
3903 start = base + offsets[area];
3906 va = pvm_find_va_enclose_addr(start);
3907 if (WARN_ON_ONCE(va == NULL))
3908 /* It is a BUG(), but trigger recovery instead. */
3911 ret = adjust_va_to_fit_type(&free_vmap_area_root,
3912 &free_vmap_area_list,
3914 if (WARN_ON_ONCE(unlikely(ret)))
3915 /* It is a BUG(), but trigger recovery instead. */
3918 /* Allocated area. */
3920 va->va_start = start;
3921 va->va_end = start + size;
3924 spin_unlock(&free_vmap_area_lock);
3926 /* populate the kasan shadow space */
3927 for (area = 0; area < nr_vms; area++) {
3928 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3929 goto err_free_shadow;
3932 /* insert all vm's */
3933 spin_lock(&vmap_area_lock);
3934 for (area = 0; area < nr_vms; area++) {
3935 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3937 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3940 spin_unlock(&vmap_area_lock);
3943 * Mark allocated areas as accessible. Do it now as a best-effort
3944 * approach, as they can be mapped outside of vmalloc code.
3945 * With hardware tag-based KASAN, marking is skipped for
3946 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3948 for (area = 0; area < nr_vms; area++)
3949 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3950 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3957 * Remove previously allocated areas. There is no
3958 * need in removing these areas from the busy tree,
3959 * because they are inserted only on the final step
3960 * and when pcpu_get_vm_areas() is success.
3963 orig_start = vas[area]->va_start;
3964 orig_end = vas[area]->va_end;
3965 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3966 &free_vmap_area_list);
3968 kasan_release_vmalloc(orig_start, orig_end,
3969 va->va_start, va->va_end);
3974 spin_unlock(&free_vmap_area_lock);
3976 purge_vmap_area_lazy();
3979 /* Before "retry", check if we recover. */
3980 for (area = 0; area < nr_vms; area++) {
3984 vas[area] = kmem_cache_zalloc(
3985 vmap_area_cachep, GFP_KERNEL);
3994 for (area = 0; area < nr_vms; area++) {
3996 kmem_cache_free(vmap_area_cachep, vas[area]);
4006 spin_lock(&free_vmap_area_lock);
4008 * We release all the vmalloc shadows, even the ones for regions that
4009 * hadn't been successfully added. This relies on kasan_release_vmalloc
4010 * being able to tolerate this case.
4012 for (area = 0; area < nr_vms; area++) {
4013 orig_start = vas[area]->va_start;
4014 orig_end = vas[area]->va_end;
4015 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4016 &free_vmap_area_list);
4018 kasan_release_vmalloc(orig_start, orig_end,
4019 va->va_start, va->va_end);
4023 spin_unlock(&free_vmap_area_lock);
4030 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4031 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4032 * @nr_vms: the number of allocated areas
4034 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4036 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4040 for (i = 0; i < nr_vms; i++)
4041 free_vm_area(vms[i]);
4044 #endif /* CONFIG_SMP */
4046 #ifdef CONFIG_PRINTK
4047 bool vmalloc_dump_obj(void *object)
4049 struct vm_struct *vm;
4050 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4052 vm = find_vm_area(objp);
4055 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4056 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4061 #ifdef CONFIG_PROC_FS
4062 static void *s_start(struct seq_file *m, loff_t *pos)
4063 __acquires(&vmap_purge_lock)
4064 __acquires(&vmap_area_lock)
4066 mutex_lock(&vmap_purge_lock);
4067 spin_lock(&vmap_area_lock);
4069 return seq_list_start(&vmap_area_list, *pos);
4072 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4074 return seq_list_next(p, &vmap_area_list, pos);
4077 static void s_stop(struct seq_file *m, void *p)
4078 __releases(&vmap_area_lock)
4079 __releases(&vmap_purge_lock)
4081 spin_unlock(&vmap_area_lock);
4082 mutex_unlock(&vmap_purge_lock);
4085 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4087 if (IS_ENABLED(CONFIG_NUMA)) {
4088 unsigned int nr, *counters = m->private;
4089 unsigned int step = 1U << vm_area_page_order(v);
4094 if (v->flags & VM_UNINITIALIZED)
4096 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4099 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4101 for (nr = 0; nr < v->nr_pages; nr += step)
4102 counters[page_to_nid(v->pages[nr])] += step;
4103 for_each_node_state(nr, N_HIGH_MEMORY)
4105 seq_printf(m, " N%u=%u", nr, counters[nr]);
4109 static void show_purge_info(struct seq_file *m)
4111 struct vmap_area *va;
4113 spin_lock(&purge_vmap_area_lock);
4114 list_for_each_entry(va, &purge_vmap_area_list, list) {
4115 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4116 (void *)va->va_start, (void *)va->va_end,
4117 va->va_end - va->va_start);
4119 spin_unlock(&purge_vmap_area_lock);
4122 static int s_show(struct seq_file *m, void *p)
4124 struct vmap_area *va;
4125 struct vm_struct *v;
4127 va = list_entry(p, struct vmap_area, list);
4130 * s_show can encounter race with remove_vm_area, !vm on behalf
4131 * of vmap area is being tear down or vm_map_ram allocation.
4134 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4135 (void *)va->va_start, (void *)va->va_end,
4136 va->va_end - va->va_start);
4143 seq_printf(m, "0x%pK-0x%pK %7ld",
4144 v->addr, v->addr + v->size, v->size);
4147 seq_printf(m, " %pS", v->caller);
4150 seq_printf(m, " pages=%d", v->nr_pages);
4153 seq_printf(m, " phys=%pa", &v->phys_addr);
4155 if (v->flags & VM_IOREMAP)
4156 seq_puts(m, " ioremap");
4158 if (v->flags & VM_ALLOC)
4159 seq_puts(m, " vmalloc");
4161 if (v->flags & VM_MAP)
4162 seq_puts(m, " vmap");
4164 if (v->flags & VM_USERMAP)
4165 seq_puts(m, " user");
4167 if (v->flags & VM_DMA_COHERENT)
4168 seq_puts(m, " dma-coherent");
4170 if (is_vmalloc_addr(v->pages))
4171 seq_puts(m, " vpages");
4173 show_numa_info(m, v);
4177 * As a final step, dump "unpurged" areas.
4180 if (list_is_last(&va->list, &vmap_area_list))
4186 static const struct seq_operations vmalloc_op = {
4193 static int __init proc_vmalloc_init(void)
4195 if (IS_ENABLED(CONFIG_NUMA))
4196 proc_create_seq_private("vmallocinfo", 0400, NULL,
4198 nr_node_ids * sizeof(unsigned int), NULL);
4200 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4203 module_init(proc_vmalloc_init);