1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/llist.h>
35 #include <linux/bitops.h>
36 #include <linux/rbtree_augmented.h>
37 #include <linux/overflow.h>
38 #include <linux/pgtable.h>
39 #include <linux/uaccess.h>
40 #include <linux/hugetlb.h>
41 #include <asm/tlbflush.h>
42 #include <asm/shmparam.h>
45 #include "pgalloc-track.h"
47 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
48 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
50 static int __init set_nohugeiomap(char *str)
52 ioremap_max_page_shift = PAGE_SHIFT;
55 early_param("nohugeiomap", set_nohugeiomap);
56 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
57 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
58 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
60 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
61 static bool __ro_after_init vmap_allow_huge = true;
63 static int __init set_nohugevmalloc(char *str)
65 vmap_allow_huge = false;
68 early_param("nohugevmalloc", set_nohugevmalloc);
69 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
70 static const bool vmap_allow_huge = false;
71 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
73 bool is_vmalloc_addr(const void *x)
75 unsigned long addr = (unsigned long)x;
77 return addr >= VMALLOC_START && addr < VMALLOC_END;
79 EXPORT_SYMBOL(is_vmalloc_addr);
81 struct vfree_deferred {
82 struct llist_head list;
83 struct work_struct wq;
85 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
87 static void __vunmap(const void *, int);
89 static void free_work(struct work_struct *w)
91 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
92 struct llist_node *t, *llnode;
94 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
95 __vunmap((void *)llnode, 1);
98 /*** Page table manipulation functions ***/
99 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
100 phys_addr_t phys_addr, pgprot_t prot,
101 unsigned int max_page_shift, pgtbl_mod_mask *mask)
105 unsigned long size = PAGE_SIZE;
107 pfn = phys_addr >> PAGE_SHIFT;
108 pte = pte_alloc_kernel_track(pmd, addr, mask);
112 BUG_ON(!pte_none(*pte));
114 #ifdef CONFIG_HUGETLB_PAGE
115 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
116 if (size != PAGE_SIZE) {
117 pte_t entry = pfn_pte(pfn, prot);
119 entry = pte_mkhuge(entry);
120 entry = arch_make_huge_pte(entry, ilog2(size), 0);
121 set_huge_pte_at(&init_mm, addr, pte, entry);
122 pfn += PFN_DOWN(size);
126 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
128 } while (pte += PFN_DOWN(size), addr += size, addr != end);
129 *mask |= PGTBL_PTE_MODIFIED;
133 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
134 phys_addr_t phys_addr, pgprot_t prot,
135 unsigned int max_page_shift)
137 if (max_page_shift < PMD_SHIFT)
140 if (!arch_vmap_pmd_supported(prot))
143 if ((end - addr) != PMD_SIZE)
146 if (!IS_ALIGNED(addr, PMD_SIZE))
149 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
152 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
155 return pmd_set_huge(pmd, phys_addr, prot);
158 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
159 phys_addr_t phys_addr, pgprot_t prot,
160 unsigned int max_page_shift, pgtbl_mod_mask *mask)
165 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
169 next = pmd_addr_end(addr, end);
171 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
173 *mask |= PGTBL_PMD_MODIFIED;
177 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
179 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
183 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
184 phys_addr_t phys_addr, pgprot_t prot,
185 unsigned int max_page_shift)
187 if (max_page_shift < PUD_SHIFT)
190 if (!arch_vmap_pud_supported(prot))
193 if ((end - addr) != PUD_SIZE)
196 if (!IS_ALIGNED(addr, PUD_SIZE))
199 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
202 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
205 return pud_set_huge(pud, phys_addr, prot);
208 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
209 phys_addr_t phys_addr, pgprot_t prot,
210 unsigned int max_page_shift, pgtbl_mod_mask *mask)
215 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
219 next = pud_addr_end(addr, end);
221 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
223 *mask |= PGTBL_PUD_MODIFIED;
227 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
228 max_page_shift, mask))
230 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
234 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
235 phys_addr_t phys_addr, pgprot_t prot,
236 unsigned int max_page_shift)
238 if (max_page_shift < P4D_SHIFT)
241 if (!arch_vmap_p4d_supported(prot))
244 if ((end - addr) != P4D_SIZE)
247 if (!IS_ALIGNED(addr, P4D_SIZE))
250 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
253 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
256 return p4d_set_huge(p4d, phys_addr, prot);
259 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
260 phys_addr_t phys_addr, pgprot_t prot,
261 unsigned int max_page_shift, pgtbl_mod_mask *mask)
266 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
270 next = p4d_addr_end(addr, end);
272 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
274 *mask |= PGTBL_P4D_MODIFIED;
278 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
279 max_page_shift, mask))
281 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
285 static int vmap_range_noflush(unsigned long addr, unsigned long end,
286 phys_addr_t phys_addr, pgprot_t prot,
287 unsigned int max_page_shift)
293 pgtbl_mod_mask mask = 0;
299 pgd = pgd_offset_k(addr);
301 next = pgd_addr_end(addr, end);
302 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
303 max_page_shift, &mask);
306 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
308 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
309 arch_sync_kernel_mappings(start, end);
314 int ioremap_page_range(unsigned long addr, unsigned long end,
315 phys_addr_t phys_addr, pgprot_t prot)
319 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
320 ioremap_max_page_shift);
321 flush_cache_vmap(addr, end);
325 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
326 pgtbl_mod_mask *mask)
330 pte = pte_offset_kernel(pmd, addr);
332 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
333 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
334 } while (pte++, addr += PAGE_SIZE, addr != end);
335 *mask |= PGTBL_PTE_MODIFIED;
338 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
339 pgtbl_mod_mask *mask)
345 pmd = pmd_offset(pud, addr);
347 next = pmd_addr_end(addr, end);
349 cleared = pmd_clear_huge(pmd);
350 if (cleared || pmd_bad(*pmd))
351 *mask |= PGTBL_PMD_MODIFIED;
355 if (pmd_none_or_clear_bad(pmd))
357 vunmap_pte_range(pmd, addr, next, mask);
360 } while (pmd++, addr = next, addr != end);
363 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
364 pgtbl_mod_mask *mask)
370 pud = pud_offset(p4d, addr);
372 next = pud_addr_end(addr, end);
374 cleared = pud_clear_huge(pud);
375 if (cleared || pud_bad(*pud))
376 *mask |= PGTBL_PUD_MODIFIED;
380 if (pud_none_or_clear_bad(pud))
382 vunmap_pmd_range(pud, addr, next, mask);
383 } while (pud++, addr = next, addr != end);
386 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
387 pgtbl_mod_mask *mask)
393 p4d = p4d_offset(pgd, addr);
395 next = p4d_addr_end(addr, end);
397 cleared = p4d_clear_huge(p4d);
398 if (cleared || p4d_bad(*p4d))
399 *mask |= PGTBL_P4D_MODIFIED;
403 if (p4d_none_or_clear_bad(p4d))
405 vunmap_pud_range(p4d, addr, next, mask);
406 } while (p4d++, addr = next, addr != end);
410 * vunmap_range_noflush is similar to vunmap_range, but does not
411 * flush caches or TLBs.
413 * The caller is responsible for calling flush_cache_vmap() before calling
414 * this function, and flush_tlb_kernel_range after it has returned
415 * successfully (and before the addresses are expected to cause a page fault
416 * or be re-mapped for something else, if TLB flushes are being delayed or
419 * This is an internal function only. Do not use outside mm/.
421 void vunmap_range_noflush(unsigned long start, unsigned long end)
425 unsigned long addr = start;
426 pgtbl_mod_mask mask = 0;
429 pgd = pgd_offset_k(addr);
431 next = pgd_addr_end(addr, end);
433 mask |= PGTBL_PGD_MODIFIED;
434 if (pgd_none_or_clear_bad(pgd))
436 vunmap_p4d_range(pgd, addr, next, &mask);
437 } while (pgd++, addr = next, addr != end);
439 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
440 arch_sync_kernel_mappings(start, end);
444 * vunmap_range - unmap kernel virtual addresses
445 * @addr: start of the VM area to unmap
446 * @end: end of the VM area to unmap (non-inclusive)
448 * Clears any present PTEs in the virtual address range, flushes TLBs and
449 * caches. Any subsequent access to the address before it has been re-mapped
452 void vunmap_range(unsigned long addr, unsigned long end)
454 flush_cache_vunmap(addr, end);
455 vunmap_range_noflush(addr, end);
456 flush_tlb_kernel_range(addr, end);
459 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
460 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
461 pgtbl_mod_mask *mask)
466 * nr is a running index into the array which helps higher level
467 * callers keep track of where we're up to.
470 pte = pte_alloc_kernel_track(pmd, addr, mask);
474 struct page *page = pages[*nr];
476 if (WARN_ON(!pte_none(*pte)))
480 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
482 } while (pte++, addr += PAGE_SIZE, addr != end);
483 *mask |= PGTBL_PTE_MODIFIED;
487 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
488 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
489 pgtbl_mod_mask *mask)
494 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
498 next = pmd_addr_end(addr, end);
499 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
501 } while (pmd++, addr = next, addr != end);
505 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
506 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
507 pgtbl_mod_mask *mask)
512 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
516 next = pud_addr_end(addr, end);
517 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
519 } while (pud++, addr = next, addr != end);
523 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
524 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
525 pgtbl_mod_mask *mask)
530 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
534 next = p4d_addr_end(addr, end);
535 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
537 } while (p4d++, addr = next, addr != end);
541 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
542 pgprot_t prot, struct page **pages)
544 unsigned long start = addr;
549 pgtbl_mod_mask mask = 0;
552 pgd = pgd_offset_k(addr);
554 next = pgd_addr_end(addr, end);
556 mask |= PGTBL_PGD_MODIFIED;
557 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
560 } while (pgd++, addr = next, addr != end);
562 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
563 arch_sync_kernel_mappings(start, end);
569 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
572 * The caller is responsible for calling flush_cache_vmap() after this
573 * function returns successfully and before the addresses are accessed.
575 * This is an internal function only. Do not use outside mm/.
577 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
578 pgprot_t prot, struct page **pages, unsigned int page_shift)
580 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
582 WARN_ON(page_shift < PAGE_SHIFT);
584 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
585 page_shift == PAGE_SHIFT)
586 return vmap_small_pages_range_noflush(addr, end, prot, pages);
588 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
591 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
592 __pa(page_address(pages[i])), prot,
597 addr += 1UL << page_shift;
604 * vmap_pages_range - map pages to a kernel virtual address
605 * @addr: start of the VM area to map
606 * @end: end of the VM area to map (non-inclusive)
607 * @prot: page protection flags to use
608 * @pages: pages to map (always PAGE_SIZE pages)
609 * @page_shift: maximum shift that the pages may be mapped with, @pages must
610 * be aligned and contiguous up to at least this shift.
613 * 0 on success, -errno on failure.
615 static int vmap_pages_range(unsigned long addr, unsigned long end,
616 pgprot_t prot, struct page **pages, unsigned int page_shift)
620 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
621 flush_cache_vmap(addr, end);
625 int is_vmalloc_or_module_addr(const void *x)
628 * ARM, x86-64 and sparc64 put modules in a special place,
629 * and fall back on vmalloc() if that fails. Others
630 * just put it in the vmalloc space.
632 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
633 unsigned long addr = (unsigned long)x;
634 if (addr >= MODULES_VADDR && addr < MODULES_END)
637 return is_vmalloc_addr(x);
641 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
642 * return the tail page that corresponds to the base page address, which
643 * matches small vmap mappings.
645 struct page *vmalloc_to_page(const void *vmalloc_addr)
647 unsigned long addr = (unsigned long) vmalloc_addr;
648 struct page *page = NULL;
649 pgd_t *pgd = pgd_offset_k(addr);
656 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
657 * architectures that do not vmalloc module space
659 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
663 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
664 return NULL; /* XXX: no allowance for huge pgd */
665 if (WARN_ON_ONCE(pgd_bad(*pgd)))
668 p4d = p4d_offset(pgd, addr);
672 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
673 if (WARN_ON_ONCE(p4d_bad(*p4d)))
676 pud = pud_offset(p4d, addr);
680 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
681 if (WARN_ON_ONCE(pud_bad(*pud)))
684 pmd = pmd_offset(pud, addr);
688 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
689 if (WARN_ON_ONCE(pmd_bad(*pmd)))
692 ptep = pte_offset_map(pmd, addr);
694 if (pte_present(pte))
695 page = pte_page(pte);
700 EXPORT_SYMBOL(vmalloc_to_page);
703 * Map a vmalloc()-space virtual address to the physical page frame number.
705 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
707 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
709 EXPORT_SYMBOL(vmalloc_to_pfn);
712 /*** Global kva allocator ***/
714 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
715 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
718 static DEFINE_SPINLOCK(vmap_area_lock);
719 static DEFINE_SPINLOCK(free_vmap_area_lock);
720 /* Export for kexec only */
721 LIST_HEAD(vmap_area_list);
722 static struct rb_root vmap_area_root = RB_ROOT;
723 static bool vmap_initialized __read_mostly;
725 static struct rb_root purge_vmap_area_root = RB_ROOT;
726 static LIST_HEAD(purge_vmap_area_list);
727 static DEFINE_SPINLOCK(purge_vmap_area_lock);
730 * This kmem_cache is used for vmap_area objects. Instead of
731 * allocating from slab we reuse an object from this cache to
732 * make things faster. Especially in "no edge" splitting of
735 static struct kmem_cache *vmap_area_cachep;
738 * This linked list is used in pair with free_vmap_area_root.
739 * It gives O(1) access to prev/next to perform fast coalescing.
741 static LIST_HEAD(free_vmap_area_list);
744 * This augment red-black tree represents the free vmap space.
745 * All vmap_area objects in this tree are sorted by va->va_start
746 * address. It is used for allocation and merging when a vmap
747 * object is released.
749 * Each vmap_area node contains a maximum available free block
750 * of its sub-tree, right or left. Therefore it is possible to
751 * find a lowest match of free area.
753 static struct rb_root free_vmap_area_root = RB_ROOT;
756 * Preload a CPU with one object for "no edge" split case. The
757 * aim is to get rid of allocations from the atomic context, thus
758 * to use more permissive allocation masks.
760 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
762 static __always_inline unsigned long
763 va_size(struct vmap_area *va)
765 return (va->va_end - va->va_start);
768 static __always_inline unsigned long
769 get_subtree_max_size(struct rb_node *node)
771 struct vmap_area *va;
773 va = rb_entry_safe(node, struct vmap_area, rb_node);
774 return va ? va->subtree_max_size : 0;
778 * Gets called when remove the node and rotate.
780 static __always_inline unsigned long
781 compute_subtree_max_size(struct vmap_area *va)
783 return max3(va_size(va),
784 get_subtree_max_size(va->rb_node.rb_left),
785 get_subtree_max_size(va->rb_node.rb_right));
788 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
789 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
791 static void purge_vmap_area_lazy(void);
792 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
793 static unsigned long lazy_max_pages(void);
795 static atomic_long_t nr_vmalloc_pages;
797 unsigned long vmalloc_nr_pages(void)
799 return atomic_long_read(&nr_vmalloc_pages);
802 static struct vmap_area *__find_vmap_area(unsigned long addr)
804 struct rb_node *n = vmap_area_root.rb_node;
807 struct vmap_area *va;
809 va = rb_entry(n, struct vmap_area, rb_node);
810 if (addr < va->va_start)
812 else if (addr >= va->va_end)
822 * This function returns back addresses of parent node
823 * and its left or right link for further processing.
825 * Otherwise NULL is returned. In that case all further
826 * steps regarding inserting of conflicting overlap range
827 * have to be declined and actually considered as a bug.
829 static __always_inline struct rb_node **
830 find_va_links(struct vmap_area *va,
831 struct rb_root *root, struct rb_node *from,
832 struct rb_node **parent)
834 struct vmap_area *tmp_va;
835 struct rb_node **link;
838 link = &root->rb_node;
839 if (unlikely(!*link)) {
848 * Go to the bottom of the tree. When we hit the last point
849 * we end up with parent rb_node and correct direction, i name
850 * it link, where the new va->rb_node will be attached to.
853 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
856 * During the traversal we also do some sanity check.
857 * Trigger the BUG() if there are sides(left/right)
860 if (va->va_start < tmp_va->va_end &&
861 va->va_end <= tmp_va->va_start)
862 link = &(*link)->rb_left;
863 else if (va->va_end > tmp_va->va_start &&
864 va->va_start >= tmp_va->va_end)
865 link = &(*link)->rb_right;
867 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
868 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
874 *parent = &tmp_va->rb_node;
878 static __always_inline struct list_head *
879 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
881 struct list_head *list;
883 if (unlikely(!parent))
885 * The red-black tree where we try to find VA neighbors
886 * before merging or inserting is empty, i.e. it means
887 * there is no free vmap space. Normally it does not
888 * happen but we handle this case anyway.
892 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
893 return (&parent->rb_right == link ? list->next : list);
896 static __always_inline void
897 link_va(struct vmap_area *va, struct rb_root *root,
898 struct rb_node *parent, struct rb_node **link, struct list_head *head)
901 * VA is still not in the list, but we can
902 * identify its future previous list_head node.
904 if (likely(parent)) {
905 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
906 if (&parent->rb_right != link)
910 /* Insert to the rb-tree */
911 rb_link_node(&va->rb_node, parent, link);
912 if (root == &free_vmap_area_root) {
914 * Some explanation here. Just perform simple insertion
915 * to the tree. We do not set va->subtree_max_size to
916 * its current size before calling rb_insert_augmented().
917 * It is because of we populate the tree from the bottom
918 * to parent levels when the node _is_ in the tree.
920 * Therefore we set subtree_max_size to zero after insertion,
921 * to let __augment_tree_propagate_from() puts everything to
922 * the correct order later on.
924 rb_insert_augmented(&va->rb_node,
925 root, &free_vmap_area_rb_augment_cb);
926 va->subtree_max_size = 0;
928 rb_insert_color(&va->rb_node, root);
931 /* Address-sort this list */
932 list_add(&va->list, head);
935 static __always_inline void
936 unlink_va(struct vmap_area *va, struct rb_root *root)
938 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
941 if (root == &free_vmap_area_root)
942 rb_erase_augmented(&va->rb_node,
943 root, &free_vmap_area_rb_augment_cb);
945 rb_erase(&va->rb_node, root);
948 RB_CLEAR_NODE(&va->rb_node);
951 #if DEBUG_AUGMENT_PROPAGATE_CHECK
953 augment_tree_propagate_check(void)
955 struct vmap_area *va;
956 unsigned long computed_size;
958 list_for_each_entry(va, &free_vmap_area_list, list) {
959 computed_size = compute_subtree_max_size(va);
960 if (computed_size != va->subtree_max_size)
961 pr_emerg("tree is corrupted: %lu, %lu\n",
962 va_size(va), va->subtree_max_size);
968 * This function populates subtree_max_size from bottom to upper
969 * levels starting from VA point. The propagation must be done
970 * when VA size is modified by changing its va_start/va_end. Or
971 * in case of newly inserting of VA to the tree.
973 * It means that __augment_tree_propagate_from() must be called:
974 * - After VA has been inserted to the tree(free path);
975 * - After VA has been shrunk(allocation path);
976 * - After VA has been increased(merging path).
978 * Please note that, it does not mean that upper parent nodes
979 * and their subtree_max_size are recalculated all the time up
988 * For example if we modify the node 4, shrinking it to 2, then
989 * no any modification is required. If we shrink the node 2 to 1
990 * its subtree_max_size is updated only, and set to 1. If we shrink
991 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
994 static __always_inline void
995 augment_tree_propagate_from(struct vmap_area *va)
998 * Populate the tree from bottom towards the root until
999 * the calculated maximum available size of checked node
1000 * is equal to its current one.
1002 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1004 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1005 augment_tree_propagate_check();
1010 insert_vmap_area(struct vmap_area *va,
1011 struct rb_root *root, struct list_head *head)
1013 struct rb_node **link;
1014 struct rb_node *parent;
1016 link = find_va_links(va, root, NULL, &parent);
1018 link_va(va, root, parent, link, head);
1022 insert_vmap_area_augment(struct vmap_area *va,
1023 struct rb_node *from, struct rb_root *root,
1024 struct list_head *head)
1026 struct rb_node **link;
1027 struct rb_node *parent;
1030 link = find_va_links(va, NULL, from, &parent);
1032 link = find_va_links(va, root, NULL, &parent);
1035 link_va(va, root, parent, link, head);
1036 augment_tree_propagate_from(va);
1041 * Merge de-allocated chunk of VA memory with previous
1042 * and next free blocks. If coalesce is not done a new
1043 * free area is inserted. If VA has been merged, it is
1046 * Please note, it can return NULL in case of overlap
1047 * ranges, followed by WARN() report. Despite it is a
1048 * buggy behaviour, a system can be alive and keep
1051 static __always_inline struct vmap_area *
1052 merge_or_add_vmap_area(struct vmap_area *va,
1053 struct rb_root *root, struct list_head *head)
1055 struct vmap_area *sibling;
1056 struct list_head *next;
1057 struct rb_node **link;
1058 struct rb_node *parent;
1059 bool merged = false;
1062 * Find a place in the tree where VA potentially will be
1063 * inserted, unless it is merged with its sibling/siblings.
1065 link = find_va_links(va, root, NULL, &parent);
1070 * Get next node of VA to check if merging can be done.
1072 next = get_va_next_sibling(parent, link);
1073 if (unlikely(next == NULL))
1079 * |<------VA------>|<-----Next----->|
1084 sibling = list_entry(next, struct vmap_area, list);
1085 if (sibling->va_start == va->va_end) {
1086 sibling->va_start = va->va_start;
1088 /* Free vmap_area object. */
1089 kmem_cache_free(vmap_area_cachep, va);
1091 /* Point to the new merged area. */
1100 * |<-----Prev----->|<------VA------>|
1104 if (next->prev != head) {
1105 sibling = list_entry(next->prev, struct vmap_area, list);
1106 if (sibling->va_end == va->va_start) {
1108 * If both neighbors are coalesced, it is important
1109 * to unlink the "next" node first, followed by merging
1110 * with "previous" one. Otherwise the tree might not be
1111 * fully populated if a sibling's augmented value is
1112 * "normalized" because of rotation operations.
1115 unlink_va(va, root);
1117 sibling->va_end = va->va_end;
1119 /* Free vmap_area object. */
1120 kmem_cache_free(vmap_area_cachep, va);
1122 /* Point to the new merged area. */
1130 link_va(va, root, parent, link, head);
1135 static __always_inline struct vmap_area *
1136 merge_or_add_vmap_area_augment(struct vmap_area *va,
1137 struct rb_root *root, struct list_head *head)
1139 va = merge_or_add_vmap_area(va, root, head);
1141 augment_tree_propagate_from(va);
1146 static __always_inline bool
1147 is_within_this_va(struct vmap_area *va, unsigned long size,
1148 unsigned long align, unsigned long vstart)
1150 unsigned long nva_start_addr;
1152 if (va->va_start > vstart)
1153 nva_start_addr = ALIGN(va->va_start, align);
1155 nva_start_addr = ALIGN(vstart, align);
1157 /* Can be overflowed due to big size or alignment. */
1158 if (nva_start_addr + size < nva_start_addr ||
1159 nva_start_addr < vstart)
1162 return (nva_start_addr + size <= va->va_end);
1166 * Find the first free block(lowest start address) in the tree,
1167 * that will accomplish the request corresponding to passing
1170 static __always_inline struct vmap_area *
1171 find_vmap_lowest_match(unsigned long size,
1172 unsigned long align, unsigned long vstart)
1174 struct vmap_area *va;
1175 struct rb_node *node;
1176 unsigned long length;
1178 /* Start from the root. */
1179 node = free_vmap_area_root.rb_node;
1181 /* Adjust the search size for alignment overhead. */
1182 length = size + align - 1;
1185 va = rb_entry(node, struct vmap_area, rb_node);
1187 if (get_subtree_max_size(node->rb_left) >= length &&
1188 vstart < va->va_start) {
1189 node = node->rb_left;
1191 if (is_within_this_va(va, size, align, vstart))
1195 * Does not make sense to go deeper towards the right
1196 * sub-tree if it does not have a free block that is
1197 * equal or bigger to the requested search length.
1199 if (get_subtree_max_size(node->rb_right) >= length) {
1200 node = node->rb_right;
1205 * OK. We roll back and find the first right sub-tree,
1206 * that will satisfy the search criteria. It can happen
1207 * only once due to "vstart" restriction.
1209 while ((node = rb_parent(node))) {
1210 va = rb_entry(node, struct vmap_area, rb_node);
1211 if (is_within_this_va(va, size, align, vstart))
1214 if (get_subtree_max_size(node->rb_right) >= length &&
1215 vstart <= va->va_start) {
1216 node = node->rb_right;
1226 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1227 #include <linux/random.h>
1229 static struct vmap_area *
1230 find_vmap_lowest_linear_match(unsigned long size,
1231 unsigned long align, unsigned long vstart)
1233 struct vmap_area *va;
1235 list_for_each_entry(va, &free_vmap_area_list, list) {
1236 if (!is_within_this_va(va, size, align, vstart))
1246 find_vmap_lowest_match_check(unsigned long size)
1248 struct vmap_area *va_1, *va_2;
1249 unsigned long vstart;
1252 get_random_bytes(&rnd, sizeof(rnd));
1253 vstart = VMALLOC_START + rnd;
1255 va_1 = find_vmap_lowest_match(size, 1, vstart);
1256 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
1259 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1260 va_1, va_2, vstart);
1266 FL_FIT_TYPE = 1, /* full fit */
1267 LE_FIT_TYPE = 2, /* left edge fit */
1268 RE_FIT_TYPE = 3, /* right edge fit */
1269 NE_FIT_TYPE = 4 /* no edge fit */
1272 static __always_inline enum fit_type
1273 classify_va_fit_type(struct vmap_area *va,
1274 unsigned long nva_start_addr, unsigned long size)
1278 /* Check if it is within VA. */
1279 if (nva_start_addr < va->va_start ||
1280 nva_start_addr + size > va->va_end)
1284 if (va->va_start == nva_start_addr) {
1285 if (va->va_end == nva_start_addr + size)
1289 } else if (va->va_end == nva_start_addr + size) {
1298 static __always_inline int
1299 adjust_va_to_fit_type(struct vmap_area *va,
1300 unsigned long nva_start_addr, unsigned long size,
1303 struct vmap_area *lva = NULL;
1305 if (type == FL_FIT_TYPE) {
1307 * No need to split VA, it fully fits.
1313 unlink_va(va, &free_vmap_area_root);
1314 kmem_cache_free(vmap_area_cachep, va);
1315 } else if (type == LE_FIT_TYPE) {
1317 * Split left edge of fit VA.
1323 va->va_start += size;
1324 } else if (type == RE_FIT_TYPE) {
1326 * Split right edge of fit VA.
1332 va->va_end = nva_start_addr;
1333 } else if (type == NE_FIT_TYPE) {
1335 * Split no edge of fit VA.
1341 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1342 if (unlikely(!lva)) {
1344 * For percpu allocator we do not do any pre-allocation
1345 * and leave it as it is. The reason is it most likely
1346 * never ends up with NE_FIT_TYPE splitting. In case of
1347 * percpu allocations offsets and sizes are aligned to
1348 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1349 * are its main fitting cases.
1351 * There are a few exceptions though, as an example it is
1352 * a first allocation (early boot up) when we have "one"
1353 * big free space that has to be split.
1355 * Also we can hit this path in case of regular "vmap"
1356 * allocations, if "this" current CPU was not preloaded.
1357 * See the comment in alloc_vmap_area() why. If so, then
1358 * GFP_NOWAIT is used instead to get an extra object for
1359 * split purpose. That is rare and most time does not
1362 * What happens if an allocation gets failed. Basically,
1363 * an "overflow" path is triggered to purge lazily freed
1364 * areas to free some memory, then, the "retry" path is
1365 * triggered to repeat one more time. See more details
1366 * in alloc_vmap_area() function.
1368 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1374 * Build the remainder.
1376 lva->va_start = va->va_start;
1377 lva->va_end = nva_start_addr;
1380 * Shrink this VA to remaining size.
1382 va->va_start = nva_start_addr + size;
1387 if (type != FL_FIT_TYPE) {
1388 augment_tree_propagate_from(va);
1390 if (lva) /* type == NE_FIT_TYPE */
1391 insert_vmap_area_augment(lva, &va->rb_node,
1392 &free_vmap_area_root, &free_vmap_area_list);
1399 * Returns a start address of the newly allocated area, if success.
1400 * Otherwise a vend is returned that indicates failure.
1402 static __always_inline unsigned long
1403 __alloc_vmap_area(unsigned long size, unsigned long align,
1404 unsigned long vstart, unsigned long vend)
1406 unsigned long nva_start_addr;
1407 struct vmap_area *va;
1411 va = find_vmap_lowest_match(size, align, vstart);
1415 if (va->va_start > vstart)
1416 nva_start_addr = ALIGN(va->va_start, align);
1418 nva_start_addr = ALIGN(vstart, align);
1420 /* Check the "vend" restriction. */
1421 if (nva_start_addr + size > vend)
1424 /* Classify what we have found. */
1425 type = classify_va_fit_type(va, nva_start_addr, size);
1426 if (WARN_ON_ONCE(type == NOTHING_FIT))
1429 /* Update the free vmap_area. */
1430 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1434 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1435 find_vmap_lowest_match_check(size);
1438 return nva_start_addr;
1442 * Free a region of KVA allocated by alloc_vmap_area
1444 static void free_vmap_area(struct vmap_area *va)
1447 * Remove from the busy tree/list.
1449 spin_lock(&vmap_area_lock);
1450 unlink_va(va, &vmap_area_root);
1451 spin_unlock(&vmap_area_lock);
1454 * Insert/Merge it back to the free tree/list.
1456 spin_lock(&free_vmap_area_lock);
1457 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1458 spin_unlock(&free_vmap_area_lock);
1462 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1464 struct vmap_area *va = NULL;
1467 * Preload this CPU with one extra vmap_area object. It is used
1468 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1469 * a CPU that does an allocation is preloaded.
1471 * We do it in non-atomic context, thus it allows us to use more
1472 * permissive allocation masks to be more stable under low memory
1473 * condition and high memory pressure.
1475 if (!this_cpu_read(ne_fit_preload_node))
1476 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1480 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1481 kmem_cache_free(vmap_area_cachep, va);
1485 * Allocate a region of KVA of the specified size and alignment, within the
1488 static struct vmap_area *alloc_vmap_area(unsigned long size,
1489 unsigned long align,
1490 unsigned long vstart, unsigned long vend,
1491 int node, gfp_t gfp_mask)
1493 struct vmap_area *va;
1499 BUG_ON(offset_in_page(size));
1500 BUG_ON(!is_power_of_2(align));
1502 if (unlikely(!vmap_initialized))
1503 return ERR_PTR(-EBUSY);
1506 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1508 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1510 return ERR_PTR(-ENOMEM);
1513 * Only scan the relevant parts containing pointers to other objects
1514 * to avoid false negatives.
1516 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1519 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1520 addr = __alloc_vmap_area(size, align, vstart, vend);
1521 spin_unlock(&free_vmap_area_lock);
1524 * If an allocation fails, the "vend" address is
1525 * returned. Therefore trigger the overflow path.
1527 if (unlikely(addr == vend))
1530 va->va_start = addr;
1531 va->va_end = addr + size;
1534 spin_lock(&vmap_area_lock);
1535 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1536 spin_unlock(&vmap_area_lock);
1538 BUG_ON(!IS_ALIGNED(va->va_start, align));
1539 BUG_ON(va->va_start < vstart);
1540 BUG_ON(va->va_end > vend);
1542 ret = kasan_populate_vmalloc(addr, size);
1545 return ERR_PTR(ret);
1552 purge_vmap_area_lazy();
1557 if (gfpflags_allow_blocking(gfp_mask)) {
1558 unsigned long freed = 0;
1559 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1566 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1567 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1570 kmem_cache_free(vmap_area_cachep, va);
1571 return ERR_PTR(-EBUSY);
1574 int register_vmap_purge_notifier(struct notifier_block *nb)
1576 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1578 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1580 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1582 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1584 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1587 * lazy_max_pages is the maximum amount of virtual address space we gather up
1588 * before attempting to purge with a TLB flush.
1590 * There is a tradeoff here: a larger number will cover more kernel page tables
1591 * and take slightly longer to purge, but it will linearly reduce the number of
1592 * global TLB flushes that must be performed. It would seem natural to scale
1593 * this number up linearly with the number of CPUs (because vmapping activity
1594 * could also scale linearly with the number of CPUs), however it is likely
1595 * that in practice, workloads might be constrained in other ways that mean
1596 * vmap activity will not scale linearly with CPUs. Also, I want to be
1597 * conservative and not introduce a big latency on huge systems, so go with
1598 * a less aggressive log scale. It will still be an improvement over the old
1599 * code, and it will be simple to change the scale factor if we find that it
1600 * becomes a problem on bigger systems.
1602 static unsigned long lazy_max_pages(void)
1606 log = fls(num_online_cpus());
1608 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1611 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1614 * Serialize vmap purging. There is no actual critical section protected
1615 * by this look, but we want to avoid concurrent calls for performance
1616 * reasons and to make the pcpu_get_vm_areas more deterministic.
1618 static DEFINE_MUTEX(vmap_purge_lock);
1620 /* for per-CPU blocks */
1621 static void purge_fragmented_blocks_allcpus(void);
1623 #ifdef CONFIG_X86_64
1625 * called before a call to iounmap() if the caller wants vm_area_struct's
1626 * immediately freed.
1628 void set_iounmap_nonlazy(void)
1630 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1632 #endif /* CONFIG_X86_64 */
1635 * Purges all lazily-freed vmap areas.
1637 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1639 unsigned long resched_threshold;
1640 struct list_head local_pure_list;
1641 struct vmap_area *va, *n_va;
1643 lockdep_assert_held(&vmap_purge_lock);
1645 spin_lock(&purge_vmap_area_lock);
1646 purge_vmap_area_root = RB_ROOT;
1647 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1648 spin_unlock(&purge_vmap_area_lock);
1650 if (unlikely(list_empty(&local_pure_list)))
1654 list_first_entry(&local_pure_list,
1655 struct vmap_area, list)->va_start);
1658 list_last_entry(&local_pure_list,
1659 struct vmap_area, list)->va_end);
1661 flush_tlb_kernel_range(start, end);
1662 resched_threshold = lazy_max_pages() << 1;
1664 spin_lock(&free_vmap_area_lock);
1665 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1666 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1667 unsigned long orig_start = va->va_start;
1668 unsigned long orig_end = va->va_end;
1671 * Finally insert or merge lazily-freed area. It is
1672 * detached and there is no need to "unlink" it from
1675 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1676 &free_vmap_area_list);
1681 if (is_vmalloc_or_module_addr((void *)orig_start))
1682 kasan_release_vmalloc(orig_start, orig_end,
1683 va->va_start, va->va_end);
1685 atomic_long_sub(nr, &vmap_lazy_nr);
1687 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1688 cond_resched_lock(&free_vmap_area_lock);
1690 spin_unlock(&free_vmap_area_lock);
1695 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1696 * is already purging.
1698 static void try_purge_vmap_area_lazy(void)
1700 if (mutex_trylock(&vmap_purge_lock)) {
1701 __purge_vmap_area_lazy(ULONG_MAX, 0);
1702 mutex_unlock(&vmap_purge_lock);
1707 * Kick off a purge of the outstanding lazy areas.
1709 static void purge_vmap_area_lazy(void)
1711 mutex_lock(&vmap_purge_lock);
1712 purge_fragmented_blocks_allcpus();
1713 __purge_vmap_area_lazy(ULONG_MAX, 0);
1714 mutex_unlock(&vmap_purge_lock);
1718 * Free a vmap area, caller ensuring that the area has been unmapped
1719 * and flush_cache_vunmap had been called for the correct range
1722 static void free_vmap_area_noflush(struct vmap_area *va)
1724 unsigned long nr_lazy;
1726 spin_lock(&vmap_area_lock);
1727 unlink_va(va, &vmap_area_root);
1728 spin_unlock(&vmap_area_lock);
1730 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1731 PAGE_SHIFT, &vmap_lazy_nr);
1734 * Merge or place it to the purge tree/list.
1736 spin_lock(&purge_vmap_area_lock);
1737 merge_or_add_vmap_area(va,
1738 &purge_vmap_area_root, &purge_vmap_area_list);
1739 spin_unlock(&purge_vmap_area_lock);
1741 /* After this point, we may free va at any time */
1742 if (unlikely(nr_lazy > lazy_max_pages()))
1743 try_purge_vmap_area_lazy();
1747 * Free and unmap a vmap area
1749 static void free_unmap_vmap_area(struct vmap_area *va)
1751 flush_cache_vunmap(va->va_start, va->va_end);
1752 vunmap_range_noflush(va->va_start, va->va_end);
1753 if (debug_pagealloc_enabled_static())
1754 flush_tlb_kernel_range(va->va_start, va->va_end);
1756 free_vmap_area_noflush(va);
1759 static struct vmap_area *find_vmap_area(unsigned long addr)
1761 struct vmap_area *va;
1763 spin_lock(&vmap_area_lock);
1764 va = __find_vmap_area(addr);
1765 spin_unlock(&vmap_area_lock);
1770 /*** Per cpu kva allocator ***/
1773 * vmap space is limited especially on 32 bit architectures. Ensure there is
1774 * room for at least 16 percpu vmap blocks per CPU.
1777 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1778 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1779 * instead (we just need a rough idea)
1781 #if BITS_PER_LONG == 32
1782 #define VMALLOC_SPACE (128UL*1024*1024)
1784 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1787 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1788 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1789 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1790 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1791 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1792 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1793 #define VMAP_BBMAP_BITS \
1794 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1795 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1796 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1798 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1800 struct vmap_block_queue {
1802 struct list_head free;
1807 struct vmap_area *va;
1808 unsigned long free, dirty;
1809 unsigned long dirty_min, dirty_max; /*< dirty range */
1810 struct list_head free_list;
1811 struct rcu_head rcu_head;
1812 struct list_head purge;
1815 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1816 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1819 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1820 * in the free path. Could get rid of this if we change the API to return a
1821 * "cookie" from alloc, to be passed to free. But no big deal yet.
1823 static DEFINE_XARRAY(vmap_blocks);
1826 * We should probably have a fallback mechanism to allocate virtual memory
1827 * out of partially filled vmap blocks. However vmap block sizing should be
1828 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1832 static unsigned long addr_to_vb_idx(unsigned long addr)
1834 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1835 addr /= VMAP_BLOCK_SIZE;
1839 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1843 addr = va_start + (pages_off << PAGE_SHIFT);
1844 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1845 return (void *)addr;
1849 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1850 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1851 * @order: how many 2^order pages should be occupied in newly allocated block
1852 * @gfp_mask: flags for the page level allocator
1854 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1856 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1858 struct vmap_block_queue *vbq;
1859 struct vmap_block *vb;
1860 struct vmap_area *va;
1861 unsigned long vb_idx;
1865 node = numa_node_id();
1867 vb = kmalloc_node(sizeof(struct vmap_block),
1868 gfp_mask & GFP_RECLAIM_MASK, node);
1870 return ERR_PTR(-ENOMEM);
1872 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1873 VMALLOC_START, VMALLOC_END,
1877 return ERR_CAST(va);
1880 vaddr = vmap_block_vaddr(va->va_start, 0);
1881 spin_lock_init(&vb->lock);
1883 /* At least something should be left free */
1884 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1885 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1887 vb->dirty_min = VMAP_BBMAP_BITS;
1889 INIT_LIST_HEAD(&vb->free_list);
1891 vb_idx = addr_to_vb_idx(va->va_start);
1892 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1896 return ERR_PTR(err);
1899 vbq = &get_cpu_var(vmap_block_queue);
1900 spin_lock(&vbq->lock);
1901 list_add_tail_rcu(&vb->free_list, &vbq->free);
1902 spin_unlock(&vbq->lock);
1903 put_cpu_var(vmap_block_queue);
1908 static void free_vmap_block(struct vmap_block *vb)
1910 struct vmap_block *tmp;
1912 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1915 free_vmap_area_noflush(vb->va);
1916 kfree_rcu(vb, rcu_head);
1919 static void purge_fragmented_blocks(int cpu)
1922 struct vmap_block *vb;
1923 struct vmap_block *n_vb;
1924 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1927 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1929 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1932 spin_lock(&vb->lock);
1933 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1934 vb->free = 0; /* prevent further allocs after releasing lock */
1935 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1937 vb->dirty_max = VMAP_BBMAP_BITS;
1938 spin_lock(&vbq->lock);
1939 list_del_rcu(&vb->free_list);
1940 spin_unlock(&vbq->lock);
1941 spin_unlock(&vb->lock);
1942 list_add_tail(&vb->purge, &purge);
1944 spin_unlock(&vb->lock);
1948 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1949 list_del(&vb->purge);
1950 free_vmap_block(vb);
1954 static void purge_fragmented_blocks_allcpus(void)
1958 for_each_possible_cpu(cpu)
1959 purge_fragmented_blocks(cpu);
1962 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1964 struct vmap_block_queue *vbq;
1965 struct vmap_block *vb;
1969 BUG_ON(offset_in_page(size));
1970 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1971 if (WARN_ON(size == 0)) {
1973 * Allocating 0 bytes isn't what caller wants since
1974 * get_order(0) returns funny result. Just warn and terminate
1979 order = get_order(size);
1982 vbq = &get_cpu_var(vmap_block_queue);
1983 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1984 unsigned long pages_off;
1986 spin_lock(&vb->lock);
1987 if (vb->free < (1UL << order)) {
1988 spin_unlock(&vb->lock);
1992 pages_off = VMAP_BBMAP_BITS - vb->free;
1993 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1994 vb->free -= 1UL << order;
1995 if (vb->free == 0) {
1996 spin_lock(&vbq->lock);
1997 list_del_rcu(&vb->free_list);
1998 spin_unlock(&vbq->lock);
2001 spin_unlock(&vb->lock);
2005 put_cpu_var(vmap_block_queue);
2008 /* Allocate new block if nothing was found */
2010 vaddr = new_vmap_block(order, gfp_mask);
2015 static void vb_free(unsigned long addr, unsigned long size)
2017 unsigned long offset;
2019 struct vmap_block *vb;
2021 BUG_ON(offset_in_page(size));
2022 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2024 flush_cache_vunmap(addr, addr + size);
2026 order = get_order(size);
2027 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2028 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2030 vunmap_range_noflush(addr, addr + size);
2032 if (debug_pagealloc_enabled_static())
2033 flush_tlb_kernel_range(addr, addr + size);
2035 spin_lock(&vb->lock);
2037 /* Expand dirty range */
2038 vb->dirty_min = min(vb->dirty_min, offset);
2039 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2041 vb->dirty += 1UL << order;
2042 if (vb->dirty == VMAP_BBMAP_BITS) {
2044 spin_unlock(&vb->lock);
2045 free_vmap_block(vb);
2047 spin_unlock(&vb->lock);
2050 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2054 if (unlikely(!vmap_initialized))
2059 for_each_possible_cpu(cpu) {
2060 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2061 struct vmap_block *vb;
2064 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2065 spin_lock(&vb->lock);
2066 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2067 unsigned long va_start = vb->va->va_start;
2070 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2071 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2073 start = min(s, start);
2078 spin_unlock(&vb->lock);
2083 mutex_lock(&vmap_purge_lock);
2084 purge_fragmented_blocks_allcpus();
2085 if (!__purge_vmap_area_lazy(start, end) && flush)
2086 flush_tlb_kernel_range(start, end);
2087 mutex_unlock(&vmap_purge_lock);
2091 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2093 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2094 * to amortize TLB flushing overheads. What this means is that any page you
2095 * have now, may, in a former life, have been mapped into kernel virtual
2096 * address by the vmap layer and so there might be some CPUs with TLB entries
2097 * still referencing that page (additional to the regular 1:1 kernel mapping).
2099 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2100 * be sure that none of the pages we have control over will have any aliases
2101 * from the vmap layer.
2103 void vm_unmap_aliases(void)
2105 unsigned long start = ULONG_MAX, end = 0;
2108 _vm_unmap_aliases(start, end, flush);
2110 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2113 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2114 * @mem: the pointer returned by vm_map_ram
2115 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2117 void vm_unmap_ram(const void *mem, unsigned int count)
2119 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2120 unsigned long addr = (unsigned long)mem;
2121 struct vmap_area *va;
2125 BUG_ON(addr < VMALLOC_START);
2126 BUG_ON(addr > VMALLOC_END);
2127 BUG_ON(!PAGE_ALIGNED(addr));
2129 kasan_poison_vmalloc(mem, size);
2131 if (likely(count <= VMAP_MAX_ALLOC)) {
2132 debug_check_no_locks_freed(mem, size);
2133 vb_free(addr, size);
2137 va = find_vmap_area(addr);
2139 debug_check_no_locks_freed((void *)va->va_start,
2140 (va->va_end - va->va_start));
2141 free_unmap_vmap_area(va);
2143 EXPORT_SYMBOL(vm_unmap_ram);
2146 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2147 * @pages: an array of pointers to the pages to be mapped
2148 * @count: number of pages
2149 * @node: prefer to allocate data structures on this node
2151 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2152 * faster than vmap so it's good. But if you mix long-life and short-life
2153 * objects with vm_map_ram(), it could consume lots of address space through
2154 * fragmentation (especially on a 32bit machine). You could see failures in
2155 * the end. Please use this function for short-lived objects.
2157 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2159 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2161 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2165 if (likely(count <= VMAP_MAX_ALLOC)) {
2166 mem = vb_alloc(size, GFP_KERNEL);
2169 addr = (unsigned long)mem;
2171 struct vmap_area *va;
2172 va = alloc_vmap_area(size, PAGE_SIZE,
2173 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2177 addr = va->va_start;
2181 kasan_unpoison_vmalloc(mem, size);
2183 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2184 pages, PAGE_SHIFT) < 0) {
2185 vm_unmap_ram(mem, count);
2191 EXPORT_SYMBOL(vm_map_ram);
2193 static struct vm_struct *vmlist __initdata;
2195 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2197 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2198 return vm->page_order;
2204 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2206 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2207 vm->page_order = order;
2214 * vm_area_add_early - add vmap area early during boot
2215 * @vm: vm_struct to add
2217 * This function is used to add fixed kernel vm area to vmlist before
2218 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2219 * should contain proper values and the other fields should be zero.
2221 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2223 void __init vm_area_add_early(struct vm_struct *vm)
2225 struct vm_struct *tmp, **p;
2227 BUG_ON(vmap_initialized);
2228 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2229 if (tmp->addr >= vm->addr) {
2230 BUG_ON(tmp->addr < vm->addr + vm->size);
2233 BUG_ON(tmp->addr + tmp->size > vm->addr);
2240 * vm_area_register_early - register vmap area early during boot
2241 * @vm: vm_struct to register
2242 * @align: requested alignment
2244 * This function is used to register kernel vm area before
2245 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2246 * proper values on entry and other fields should be zero. On return,
2247 * vm->addr contains the allocated address.
2249 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2251 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2253 static size_t vm_init_off __initdata;
2256 addr = ALIGN(VMALLOC_START + vm_init_off, align);
2257 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
2259 vm->addr = (void *)addr;
2261 vm_area_add_early(vm);
2264 static void vmap_init_free_space(void)
2266 unsigned long vmap_start = 1;
2267 const unsigned long vmap_end = ULONG_MAX;
2268 struct vmap_area *busy, *free;
2272 * -|-----|.....|-----|-----|-----|.....|-
2274 * |<--------------------------------->|
2276 list_for_each_entry(busy, &vmap_area_list, list) {
2277 if (busy->va_start - vmap_start > 0) {
2278 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2279 if (!WARN_ON_ONCE(!free)) {
2280 free->va_start = vmap_start;
2281 free->va_end = busy->va_start;
2283 insert_vmap_area_augment(free, NULL,
2284 &free_vmap_area_root,
2285 &free_vmap_area_list);
2289 vmap_start = busy->va_end;
2292 if (vmap_end - vmap_start > 0) {
2293 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2294 if (!WARN_ON_ONCE(!free)) {
2295 free->va_start = vmap_start;
2296 free->va_end = vmap_end;
2298 insert_vmap_area_augment(free, NULL,
2299 &free_vmap_area_root,
2300 &free_vmap_area_list);
2305 void __init vmalloc_init(void)
2307 struct vmap_area *va;
2308 struct vm_struct *tmp;
2312 * Create the cache for vmap_area objects.
2314 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2316 for_each_possible_cpu(i) {
2317 struct vmap_block_queue *vbq;
2318 struct vfree_deferred *p;
2320 vbq = &per_cpu(vmap_block_queue, i);
2321 spin_lock_init(&vbq->lock);
2322 INIT_LIST_HEAD(&vbq->free);
2323 p = &per_cpu(vfree_deferred, i);
2324 init_llist_head(&p->list);
2325 INIT_WORK(&p->wq, free_work);
2328 /* Import existing vmlist entries. */
2329 for (tmp = vmlist; tmp; tmp = tmp->next) {
2330 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2331 if (WARN_ON_ONCE(!va))
2334 va->va_start = (unsigned long)tmp->addr;
2335 va->va_end = va->va_start + tmp->size;
2337 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2341 * Now we can initialize a free vmap space.
2343 vmap_init_free_space();
2344 vmap_initialized = true;
2347 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2348 struct vmap_area *va, unsigned long flags, const void *caller)
2351 vm->addr = (void *)va->va_start;
2352 vm->size = va->va_end - va->va_start;
2353 vm->caller = caller;
2357 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2358 unsigned long flags, const void *caller)
2360 spin_lock(&vmap_area_lock);
2361 setup_vmalloc_vm_locked(vm, va, flags, caller);
2362 spin_unlock(&vmap_area_lock);
2365 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2368 * Before removing VM_UNINITIALIZED,
2369 * we should make sure that vm has proper values.
2370 * Pair with smp_rmb() in show_numa_info().
2373 vm->flags &= ~VM_UNINITIALIZED;
2376 static struct vm_struct *__get_vm_area_node(unsigned long size,
2377 unsigned long align, unsigned long shift, unsigned long flags,
2378 unsigned long start, unsigned long end, int node,
2379 gfp_t gfp_mask, const void *caller)
2381 struct vmap_area *va;
2382 struct vm_struct *area;
2383 unsigned long requested_size = size;
2385 BUG_ON(in_interrupt());
2386 size = ALIGN(size, 1ul << shift);
2387 if (unlikely(!size))
2390 if (flags & VM_IOREMAP)
2391 align = 1ul << clamp_t(int, get_count_order_long(size),
2392 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2394 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2395 if (unlikely(!area))
2398 if (!(flags & VM_NO_GUARD))
2401 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2407 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2409 setup_vmalloc_vm(area, va, flags, caller);
2414 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2415 unsigned long start, unsigned long end,
2418 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2419 NUMA_NO_NODE, GFP_KERNEL, caller);
2423 * get_vm_area - reserve a contiguous kernel virtual area
2424 * @size: size of the area
2425 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2427 * Search an area of @size in the kernel virtual mapping area,
2428 * and reserved it for out purposes. Returns the area descriptor
2429 * on success or %NULL on failure.
2431 * Return: the area descriptor on success or %NULL on failure.
2433 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2435 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2436 VMALLOC_START, VMALLOC_END,
2437 NUMA_NO_NODE, GFP_KERNEL,
2438 __builtin_return_address(0));
2441 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2444 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2445 VMALLOC_START, VMALLOC_END,
2446 NUMA_NO_NODE, GFP_KERNEL, caller);
2450 * find_vm_area - find a continuous kernel virtual area
2451 * @addr: base address
2453 * Search for the kernel VM area starting at @addr, and return it.
2454 * It is up to the caller to do all required locking to keep the returned
2457 * Return: the area descriptor on success or %NULL on failure.
2459 struct vm_struct *find_vm_area(const void *addr)
2461 struct vmap_area *va;
2463 va = find_vmap_area((unsigned long)addr);
2471 * remove_vm_area - find and remove a continuous kernel virtual area
2472 * @addr: base address
2474 * Search for the kernel VM area starting at @addr, and remove it.
2475 * This function returns the found VM area, but using it is NOT safe
2476 * on SMP machines, except for its size or flags.
2478 * Return: the area descriptor on success or %NULL on failure.
2480 struct vm_struct *remove_vm_area(const void *addr)
2482 struct vmap_area *va;
2486 spin_lock(&vmap_area_lock);
2487 va = __find_vmap_area((unsigned long)addr);
2489 struct vm_struct *vm = va->vm;
2492 spin_unlock(&vmap_area_lock);
2494 kasan_free_shadow(vm);
2495 free_unmap_vmap_area(va);
2500 spin_unlock(&vmap_area_lock);
2504 static inline void set_area_direct_map(const struct vm_struct *area,
2505 int (*set_direct_map)(struct page *page))
2509 /* HUGE_VMALLOC passes small pages to set_direct_map */
2510 for (i = 0; i < area->nr_pages; i++)
2511 if (page_address(area->pages[i]))
2512 set_direct_map(area->pages[i]);
2515 /* Handle removing and resetting vm mappings related to the vm_struct. */
2516 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2518 unsigned long start = ULONG_MAX, end = 0;
2519 unsigned int page_order = vm_area_page_order(area);
2520 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2524 remove_vm_area(area->addr);
2526 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2531 * If not deallocating pages, just do the flush of the VM area and
2534 if (!deallocate_pages) {
2540 * If execution gets here, flush the vm mapping and reset the direct
2541 * map. Find the start and end range of the direct mappings to make sure
2542 * the vm_unmap_aliases() flush includes the direct map.
2544 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2545 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2547 unsigned long page_size;
2549 page_size = PAGE_SIZE << page_order;
2550 start = min(addr, start);
2551 end = max(addr + page_size, end);
2557 * Set direct map to something invalid so that it won't be cached if
2558 * there are any accesses after the TLB flush, then flush the TLB and
2559 * reset the direct map permissions to the default.
2561 set_area_direct_map(area, set_direct_map_invalid_noflush);
2562 _vm_unmap_aliases(start, end, flush_dmap);
2563 set_area_direct_map(area, set_direct_map_default_noflush);
2566 static void __vunmap(const void *addr, int deallocate_pages)
2568 struct vm_struct *area;
2573 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2577 area = find_vm_area(addr);
2578 if (unlikely(!area)) {
2579 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2584 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2585 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2587 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2589 vm_remove_mappings(area, deallocate_pages);
2591 if (deallocate_pages) {
2592 unsigned int page_order = vm_area_page_order(area);
2595 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2596 struct page *page = area->pages[i];
2599 __free_pages(page, page_order);
2602 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2604 kvfree(area->pages);
2610 static inline void __vfree_deferred(const void *addr)
2613 * Use raw_cpu_ptr() because this can be called from preemptible
2614 * context. Preemption is absolutely fine here, because the llist_add()
2615 * implementation is lockless, so it works even if we are adding to
2616 * another cpu's list. schedule_work() should be fine with this too.
2618 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2620 if (llist_add((struct llist_node *)addr, &p->list))
2621 schedule_work(&p->wq);
2625 * vfree_atomic - release memory allocated by vmalloc()
2626 * @addr: memory base address
2628 * This one is just like vfree() but can be called in any atomic context
2631 void vfree_atomic(const void *addr)
2635 kmemleak_free(addr);
2639 __vfree_deferred(addr);
2642 static void __vfree(const void *addr)
2644 if (unlikely(in_interrupt()))
2645 __vfree_deferred(addr);
2651 * vfree - Release memory allocated by vmalloc()
2652 * @addr: Memory base address
2654 * Free the virtually continuous memory area starting at @addr, as obtained
2655 * from one of the vmalloc() family of APIs. This will usually also free the
2656 * physical memory underlying the virtual allocation, but that memory is
2657 * reference counted, so it will not be freed until the last user goes away.
2659 * If @addr is NULL, no operation is performed.
2662 * May sleep if called *not* from interrupt context.
2663 * Must not be called in NMI context (strictly speaking, it could be
2664 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2665 * conventions for vfree() arch-dependent would be a really bad idea).
2667 void vfree(const void *addr)
2671 kmemleak_free(addr);
2673 might_sleep_if(!in_interrupt());
2680 EXPORT_SYMBOL(vfree);
2683 * vunmap - release virtual mapping obtained by vmap()
2684 * @addr: memory base address
2686 * Free the virtually contiguous memory area starting at @addr,
2687 * which was created from the page array passed to vmap().
2689 * Must not be called in interrupt context.
2691 void vunmap(const void *addr)
2693 BUG_ON(in_interrupt());
2698 EXPORT_SYMBOL(vunmap);
2701 * vmap - map an array of pages into virtually contiguous space
2702 * @pages: array of page pointers
2703 * @count: number of pages to map
2704 * @flags: vm_area->flags
2705 * @prot: page protection for the mapping
2707 * Maps @count pages from @pages into contiguous kernel virtual space.
2708 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2709 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2710 * are transferred from the caller to vmap(), and will be freed / dropped when
2711 * vfree() is called on the return value.
2713 * Return: the address of the area or %NULL on failure
2715 void *vmap(struct page **pages, unsigned int count,
2716 unsigned long flags, pgprot_t prot)
2718 struct vm_struct *area;
2720 unsigned long size; /* In bytes */
2724 if (count > totalram_pages())
2727 size = (unsigned long)count << PAGE_SHIFT;
2728 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2732 addr = (unsigned long)area->addr;
2733 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2734 pages, PAGE_SHIFT) < 0) {
2739 if (flags & VM_MAP_PUT_PAGES) {
2740 area->pages = pages;
2741 area->nr_pages = count;
2745 EXPORT_SYMBOL(vmap);
2747 #ifdef CONFIG_VMAP_PFN
2748 struct vmap_pfn_data {
2749 unsigned long *pfns;
2754 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2756 struct vmap_pfn_data *data = private;
2758 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2760 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2765 * vmap_pfn - map an array of PFNs into virtually contiguous space
2766 * @pfns: array of PFNs
2767 * @count: number of pages to map
2768 * @prot: page protection for the mapping
2770 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2771 * the start address of the mapping.
2773 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2775 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2776 struct vm_struct *area;
2778 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2779 __builtin_return_address(0));
2782 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2783 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2789 EXPORT_SYMBOL_GPL(vmap_pfn);
2790 #endif /* CONFIG_VMAP_PFN */
2792 static inline unsigned int
2793 vm_area_alloc_pages(gfp_t gfp, int nid,
2794 unsigned int order, unsigned long nr_pages, struct page **pages)
2796 unsigned int nr_allocated = 0;
2799 * For order-0 pages we make use of bulk allocator, if
2800 * the page array is partly or not at all populated due
2801 * to fails, fallback to a single page allocator that is
2805 nr_allocated = alloc_pages_bulk_array_node(
2806 gfp, nid, nr_pages, pages);
2809 * Compound pages required for remap_vmalloc_page if
2814 /* High-order pages or fallback path if "bulk" fails. */
2815 while (nr_allocated < nr_pages) {
2819 page = alloc_pages_node(nid, gfp, order);
2820 if (unlikely(!page))
2824 * Careful, we allocate and map page-order pages, but
2825 * tracking is done per PAGE_SIZE page so as to keep the
2826 * vm_struct APIs independent of the physical/mapped size.
2828 for (i = 0; i < (1U << order); i++)
2829 pages[nr_allocated + i] = page + i;
2831 if (gfpflags_allow_blocking(gfp))
2834 nr_allocated += 1U << order;
2837 return nr_allocated;
2840 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2841 pgprot_t prot, unsigned int page_shift,
2844 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2845 unsigned long addr = (unsigned long)area->addr;
2846 unsigned long size = get_vm_area_size(area);
2847 unsigned long array_size;
2848 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2849 unsigned int page_order;
2851 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2852 gfp_mask |= __GFP_NOWARN;
2853 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2854 gfp_mask |= __GFP_HIGHMEM;
2856 /* Please note that the recursion is strictly bounded. */
2857 if (array_size > PAGE_SIZE) {
2858 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2861 area->pages = kmalloc_node(array_size, nested_gfp, node);
2865 warn_alloc(gfp_mask, NULL,
2866 "vmalloc error: size %lu, failed to allocated page array size %lu",
2867 nr_small_pages * PAGE_SIZE, array_size);
2872 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2873 page_order = vm_area_page_order(area);
2875 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2876 page_order, nr_small_pages, area->pages);
2878 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2881 * If not enough pages were obtained to accomplish an
2882 * allocation request, free them via __vfree() if any.
2884 if (area->nr_pages != nr_small_pages) {
2885 warn_alloc(gfp_mask, NULL,
2886 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2887 area->nr_pages * PAGE_SIZE, page_order);
2891 if (vmap_pages_range(addr, addr + size, prot, area->pages,
2893 warn_alloc(gfp_mask, NULL,
2894 "vmalloc error: size %lu, failed to map pages",
2895 area->nr_pages * PAGE_SIZE);
2902 __vfree(area->addr);
2907 * __vmalloc_node_range - allocate virtually contiguous memory
2908 * @size: allocation size
2909 * @align: desired alignment
2910 * @start: vm area range start
2911 * @end: vm area range end
2912 * @gfp_mask: flags for the page level allocator
2913 * @prot: protection mask for the allocated pages
2914 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2915 * @node: node to use for allocation or NUMA_NO_NODE
2916 * @caller: caller's return address
2918 * Allocate enough pages to cover @size from the page level
2919 * allocator with @gfp_mask flags. Map them into contiguous
2920 * kernel virtual space, using a pagetable protection of @prot.
2922 * Return: the address of the area or %NULL on failure
2924 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2925 unsigned long start, unsigned long end, gfp_t gfp_mask,
2926 pgprot_t prot, unsigned long vm_flags, int node,
2929 struct vm_struct *area;
2931 unsigned long real_size = size;
2932 unsigned long real_align = align;
2933 unsigned int shift = PAGE_SHIFT;
2935 if (WARN_ON_ONCE(!size))
2938 if ((size >> PAGE_SHIFT) > totalram_pages()) {
2939 warn_alloc(gfp_mask, NULL,
2940 "vmalloc error: size %lu, exceeds total pages",
2945 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
2946 unsigned long size_per_node;
2949 * Try huge pages. Only try for PAGE_KERNEL allocations,
2950 * others like modules don't yet expect huge pages in
2951 * their allocations due to apply_to_page_range not
2955 size_per_node = size;
2956 if (node == NUMA_NO_NODE)
2957 size_per_node /= num_online_nodes();
2958 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
2961 shift = arch_vmap_pte_supported_shift(size_per_node);
2963 align = max(real_align, 1UL << shift);
2964 size = ALIGN(real_size, 1UL << shift);
2968 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
2969 VM_UNINITIALIZED | vm_flags, start, end, node,
2972 warn_alloc(gfp_mask, NULL,
2973 "vmalloc error: size %lu, vm_struct allocation failed",
2978 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
2983 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2984 * flag. It means that vm_struct is not fully initialized.
2985 * Now, it is fully initialized, so remove this flag here.
2987 clear_vm_uninitialized_flag(area);
2989 size = PAGE_ALIGN(size);
2990 kmemleak_vmalloc(area, size, gfp_mask);
2995 if (shift > PAGE_SHIFT) {
3006 * __vmalloc_node - allocate virtually contiguous memory
3007 * @size: allocation size
3008 * @align: desired alignment
3009 * @gfp_mask: flags for the page level allocator
3010 * @node: node to use for allocation or NUMA_NO_NODE
3011 * @caller: caller's return address
3013 * Allocate enough pages to cover @size from the page level allocator with
3014 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3016 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3017 * and __GFP_NOFAIL are not supported
3019 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3022 * Return: pointer to the allocated memory or %NULL on error
3024 void *__vmalloc_node(unsigned long size, unsigned long align,
3025 gfp_t gfp_mask, int node, const void *caller)
3027 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3028 gfp_mask, PAGE_KERNEL, 0, node, caller);
3031 * This is only for performance analysis of vmalloc and stress purpose.
3032 * It is required by vmalloc test module, therefore do not use it other
3035 #ifdef CONFIG_TEST_VMALLOC_MODULE
3036 EXPORT_SYMBOL_GPL(__vmalloc_node);
3039 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3041 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3042 __builtin_return_address(0));
3044 EXPORT_SYMBOL(__vmalloc);
3047 * vmalloc - allocate virtually contiguous memory
3048 * @size: allocation size
3050 * Allocate enough pages to cover @size from the page level
3051 * allocator and map them into contiguous kernel virtual space.
3053 * For tight control over page level allocator and protection flags
3054 * use __vmalloc() instead.
3056 * Return: pointer to the allocated memory or %NULL on error
3058 void *vmalloc(unsigned long size)
3060 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3061 __builtin_return_address(0));
3063 EXPORT_SYMBOL(vmalloc);
3066 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3067 * @size: allocation size
3069 * Allocate enough non-huge pages to cover @size from the page level
3070 * allocator and map them into contiguous kernel virtual space.
3072 * Return: pointer to the allocated memory or %NULL on error
3074 void *vmalloc_no_huge(unsigned long size)
3076 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3077 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3078 NUMA_NO_NODE, __builtin_return_address(0));
3080 EXPORT_SYMBOL(vmalloc_no_huge);
3083 * vzalloc - allocate virtually contiguous memory with zero fill
3084 * @size: allocation size
3086 * Allocate enough pages to cover @size from the page level
3087 * allocator and map them into contiguous kernel virtual space.
3088 * The memory allocated is set to zero.
3090 * For tight control over page level allocator and protection flags
3091 * use __vmalloc() instead.
3093 * Return: pointer to the allocated memory or %NULL on error
3095 void *vzalloc(unsigned long size)
3097 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3098 __builtin_return_address(0));
3100 EXPORT_SYMBOL(vzalloc);
3103 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3104 * @size: allocation size
3106 * The resulting memory area is zeroed so it can be mapped to userspace
3107 * without leaking data.
3109 * Return: pointer to the allocated memory or %NULL on error
3111 void *vmalloc_user(unsigned long size)
3113 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3114 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3115 VM_USERMAP, NUMA_NO_NODE,
3116 __builtin_return_address(0));
3118 EXPORT_SYMBOL(vmalloc_user);
3121 * vmalloc_node - allocate memory on a specific node
3122 * @size: allocation size
3125 * Allocate enough pages to cover @size from the page level
3126 * allocator and map them into contiguous kernel virtual space.
3128 * For tight control over page level allocator and protection flags
3129 * use __vmalloc() instead.
3131 * Return: pointer to the allocated memory or %NULL on error
3133 void *vmalloc_node(unsigned long size, int node)
3135 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3136 __builtin_return_address(0));
3138 EXPORT_SYMBOL(vmalloc_node);
3141 * vzalloc_node - allocate memory on a specific node with zero fill
3142 * @size: allocation size
3145 * Allocate enough pages to cover @size from the page level
3146 * allocator and map them into contiguous kernel virtual space.
3147 * The memory allocated is set to zero.
3149 * Return: pointer to the allocated memory or %NULL on error
3151 void *vzalloc_node(unsigned long size, int node)
3153 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3154 __builtin_return_address(0));
3156 EXPORT_SYMBOL(vzalloc_node);
3158 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3159 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3160 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3161 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3164 * 64b systems should always have either DMA or DMA32 zones. For others
3165 * GFP_DMA32 should do the right thing and use the normal zone.
3167 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3171 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3172 * @size: allocation size
3174 * Allocate enough 32bit PA addressable pages to cover @size from the
3175 * page level allocator and map them into contiguous kernel virtual space.
3177 * Return: pointer to the allocated memory or %NULL on error
3179 void *vmalloc_32(unsigned long size)
3181 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3182 __builtin_return_address(0));
3184 EXPORT_SYMBOL(vmalloc_32);
3187 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3188 * @size: allocation size
3190 * The resulting memory area is 32bit addressable and zeroed so it can be
3191 * mapped to userspace without leaking data.
3193 * Return: pointer to the allocated memory or %NULL on error
3195 void *vmalloc_32_user(unsigned long size)
3197 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3198 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3199 VM_USERMAP, NUMA_NO_NODE,
3200 __builtin_return_address(0));
3202 EXPORT_SYMBOL(vmalloc_32_user);
3205 * small helper routine , copy contents to buf from addr.
3206 * If the page is not present, fill zero.
3209 static int aligned_vread(char *buf, char *addr, unsigned long count)
3215 unsigned long offset, length;
3217 offset = offset_in_page(addr);
3218 length = PAGE_SIZE - offset;
3221 p = vmalloc_to_page(addr);
3223 * To do safe access to this _mapped_ area, we need
3224 * lock. But adding lock here means that we need to add
3225 * overhead of vmalloc()/vfree() calls for this _debug_
3226 * interface, rarely used. Instead of that, we'll use
3227 * kmap() and get small overhead in this access function.
3230 /* We can expect USER0 is not used -- see vread() */
3231 void *map = kmap_atomic(p);
3232 memcpy(buf, map + offset, length);
3235 memset(buf, 0, length);
3246 * vread() - read vmalloc area in a safe way.
3247 * @buf: buffer for reading data
3248 * @addr: vm address.
3249 * @count: number of bytes to be read.
3251 * This function checks that addr is a valid vmalloc'ed area, and
3252 * copy data from that area to a given buffer. If the given memory range
3253 * of [addr...addr+count) includes some valid address, data is copied to
3254 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3255 * IOREMAP area is treated as memory hole and no copy is done.
3257 * If [addr...addr+count) doesn't includes any intersects with alive
3258 * vm_struct area, returns 0. @buf should be kernel's buffer.
3260 * Note: In usual ops, vread() is never necessary because the caller
3261 * should know vmalloc() area is valid and can use memcpy().
3262 * This is for routines which have to access vmalloc area without
3263 * any information, as /proc/kcore.
3265 * Return: number of bytes for which addr and buf should be increased
3266 * (same number as @count) or %0 if [addr...addr+count) doesn't
3267 * include any intersection with valid vmalloc area
3269 long vread(char *buf, char *addr, unsigned long count)
3271 struct vmap_area *va;
3272 struct vm_struct *vm;
3273 char *vaddr, *buf_start = buf;
3274 unsigned long buflen = count;
3277 /* Don't allow overflow */
3278 if ((unsigned long) addr + count < count)
3279 count = -(unsigned long) addr;
3281 spin_lock(&vmap_area_lock);
3282 va = __find_vmap_area((unsigned long)addr);
3285 list_for_each_entry_from(va, &vmap_area_list, list) {
3293 vaddr = (char *) vm->addr;
3294 if (addr >= vaddr + get_vm_area_size(vm))
3296 while (addr < vaddr) {
3304 n = vaddr + get_vm_area_size(vm) - addr;
3307 if (!(vm->flags & VM_IOREMAP))
3308 aligned_vread(buf, addr, n);
3309 else /* IOREMAP area is treated as memory hole */
3316 spin_unlock(&vmap_area_lock);
3318 if (buf == buf_start)
3320 /* zero-fill memory holes */
3321 if (buf != buf_start + buflen)
3322 memset(buf, 0, buflen - (buf - buf_start));
3328 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3329 * @vma: vma to cover
3330 * @uaddr: target user address to start at
3331 * @kaddr: virtual address of vmalloc kernel memory
3332 * @pgoff: offset from @kaddr to start at
3333 * @size: size of map area
3335 * Returns: 0 for success, -Exxx on failure
3337 * This function checks that @kaddr is a valid vmalloc'ed area,
3338 * and that it is big enough to cover the range starting at
3339 * @uaddr in @vma. Will return failure if that criteria isn't
3342 * Similar to remap_pfn_range() (see mm/memory.c)
3344 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3345 void *kaddr, unsigned long pgoff,
3348 struct vm_struct *area;
3350 unsigned long end_index;
3352 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3355 size = PAGE_ALIGN(size);
3357 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3360 area = find_vm_area(kaddr);
3364 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3367 if (check_add_overflow(size, off, &end_index) ||
3368 end_index > get_vm_area_size(area))
3373 struct page *page = vmalloc_to_page(kaddr);
3376 ret = vm_insert_page(vma, uaddr, page);
3385 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3391 * remap_vmalloc_range - map vmalloc pages to userspace
3392 * @vma: vma to cover (map full range of vma)
3393 * @addr: vmalloc memory
3394 * @pgoff: number of pages into addr before first page to map
3396 * Returns: 0 for success, -Exxx on failure
3398 * This function checks that addr is a valid vmalloc'ed area, and
3399 * that it is big enough to cover the vma. Will return failure if
3400 * that criteria isn't met.
3402 * Similar to remap_pfn_range() (see mm/memory.c)
3404 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3405 unsigned long pgoff)
3407 return remap_vmalloc_range_partial(vma, vma->vm_start,
3409 vma->vm_end - vma->vm_start);
3411 EXPORT_SYMBOL(remap_vmalloc_range);
3413 void free_vm_area(struct vm_struct *area)
3415 struct vm_struct *ret;
3416 ret = remove_vm_area(area->addr);
3417 BUG_ON(ret != area);
3420 EXPORT_SYMBOL_GPL(free_vm_area);
3423 static struct vmap_area *node_to_va(struct rb_node *n)
3425 return rb_entry_safe(n, struct vmap_area, rb_node);
3429 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3430 * @addr: target address
3432 * Returns: vmap_area if it is found. If there is no such area
3433 * the first highest(reverse order) vmap_area is returned
3434 * i.e. va->va_start < addr && va->va_end < addr or NULL
3435 * if there are no any areas before @addr.
3437 static struct vmap_area *
3438 pvm_find_va_enclose_addr(unsigned long addr)
3440 struct vmap_area *va, *tmp;
3443 n = free_vmap_area_root.rb_node;
3447 tmp = rb_entry(n, struct vmap_area, rb_node);
3448 if (tmp->va_start <= addr) {
3450 if (tmp->va_end >= addr)
3463 * pvm_determine_end_from_reverse - find the highest aligned address
3464 * of free block below VMALLOC_END
3466 * in - the VA we start the search(reverse order);
3467 * out - the VA with the highest aligned end address.
3468 * @align: alignment for required highest address
3470 * Returns: determined end address within vmap_area
3472 static unsigned long
3473 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3475 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3479 list_for_each_entry_from_reverse((*va),
3480 &free_vmap_area_list, list) {
3481 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3482 if ((*va)->va_start < addr)
3491 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3492 * @offsets: array containing offset of each area
3493 * @sizes: array containing size of each area
3494 * @nr_vms: the number of areas to allocate
3495 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3497 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3498 * vm_structs on success, %NULL on failure
3500 * Percpu allocator wants to use congruent vm areas so that it can
3501 * maintain the offsets among percpu areas. This function allocates
3502 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3503 * be scattered pretty far, distance between two areas easily going up
3504 * to gigabytes. To avoid interacting with regular vmallocs, these
3505 * areas are allocated from top.
3507 * Despite its complicated look, this allocator is rather simple. It
3508 * does everything top-down and scans free blocks from the end looking
3509 * for matching base. While scanning, if any of the areas do not fit the
3510 * base address is pulled down to fit the area. Scanning is repeated till
3511 * all the areas fit and then all necessary data structures are inserted
3512 * and the result is returned.
3514 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3515 const size_t *sizes, int nr_vms,
3518 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3519 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3520 struct vmap_area **vas, *va;
3521 struct vm_struct **vms;
3522 int area, area2, last_area, term_area;
3523 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3524 bool purged = false;
3527 /* verify parameters and allocate data structures */
3528 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3529 for (last_area = 0, area = 0; area < nr_vms; area++) {
3530 start = offsets[area];
3531 end = start + sizes[area];
3533 /* is everything aligned properly? */
3534 BUG_ON(!IS_ALIGNED(offsets[area], align));
3535 BUG_ON(!IS_ALIGNED(sizes[area], align));
3537 /* detect the area with the highest address */
3538 if (start > offsets[last_area])
3541 for (area2 = area + 1; area2 < nr_vms; area2++) {
3542 unsigned long start2 = offsets[area2];
3543 unsigned long end2 = start2 + sizes[area2];
3545 BUG_ON(start2 < end && start < end2);
3548 last_end = offsets[last_area] + sizes[last_area];
3550 if (vmalloc_end - vmalloc_start < last_end) {
3555 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3556 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3560 for (area = 0; area < nr_vms; area++) {
3561 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3562 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3563 if (!vas[area] || !vms[area])
3567 spin_lock(&free_vmap_area_lock);
3569 /* start scanning - we scan from the top, begin with the last area */
3570 area = term_area = last_area;
3571 start = offsets[area];
3572 end = start + sizes[area];
3574 va = pvm_find_va_enclose_addr(vmalloc_end);
3575 base = pvm_determine_end_from_reverse(&va, align) - end;
3579 * base might have underflowed, add last_end before
3582 if (base + last_end < vmalloc_start + last_end)
3586 * Fitting base has not been found.
3592 * If required width exceeds current VA block, move
3593 * base downwards and then recheck.
3595 if (base + end > va->va_end) {
3596 base = pvm_determine_end_from_reverse(&va, align) - end;
3602 * If this VA does not fit, move base downwards and recheck.
3604 if (base + start < va->va_start) {
3605 va = node_to_va(rb_prev(&va->rb_node));
3606 base = pvm_determine_end_from_reverse(&va, align) - end;
3612 * This area fits, move on to the previous one. If
3613 * the previous one is the terminal one, we're done.
3615 area = (area + nr_vms - 1) % nr_vms;
3616 if (area == term_area)
3619 start = offsets[area];
3620 end = start + sizes[area];
3621 va = pvm_find_va_enclose_addr(base + end);
3624 /* we've found a fitting base, insert all va's */
3625 for (area = 0; area < nr_vms; area++) {
3628 start = base + offsets[area];
3631 va = pvm_find_va_enclose_addr(start);
3632 if (WARN_ON_ONCE(va == NULL))
3633 /* It is a BUG(), but trigger recovery instead. */
3636 type = classify_va_fit_type(va, start, size);
3637 if (WARN_ON_ONCE(type == NOTHING_FIT))
3638 /* It is a BUG(), but trigger recovery instead. */
3641 ret = adjust_va_to_fit_type(va, start, size, type);
3645 /* Allocated area. */
3647 va->va_start = start;
3648 va->va_end = start + size;
3651 spin_unlock(&free_vmap_area_lock);
3653 /* populate the kasan shadow space */
3654 for (area = 0; area < nr_vms; area++) {
3655 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3656 goto err_free_shadow;
3658 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3662 /* insert all vm's */
3663 spin_lock(&vmap_area_lock);
3664 for (area = 0; area < nr_vms; area++) {
3665 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3667 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3670 spin_unlock(&vmap_area_lock);
3677 * Remove previously allocated areas. There is no
3678 * need in removing these areas from the busy tree,
3679 * because they are inserted only on the final step
3680 * and when pcpu_get_vm_areas() is success.
3683 orig_start = vas[area]->va_start;
3684 orig_end = vas[area]->va_end;
3685 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3686 &free_vmap_area_list);
3688 kasan_release_vmalloc(orig_start, orig_end,
3689 va->va_start, va->va_end);
3694 spin_unlock(&free_vmap_area_lock);
3696 purge_vmap_area_lazy();
3699 /* Before "retry", check if we recover. */
3700 for (area = 0; area < nr_vms; area++) {
3704 vas[area] = kmem_cache_zalloc(
3705 vmap_area_cachep, GFP_KERNEL);
3714 for (area = 0; area < nr_vms; area++) {
3716 kmem_cache_free(vmap_area_cachep, vas[area]);
3726 spin_lock(&free_vmap_area_lock);
3728 * We release all the vmalloc shadows, even the ones for regions that
3729 * hadn't been successfully added. This relies on kasan_release_vmalloc
3730 * being able to tolerate this case.
3732 for (area = 0; area < nr_vms; area++) {
3733 orig_start = vas[area]->va_start;
3734 orig_end = vas[area]->va_end;
3735 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3736 &free_vmap_area_list);
3738 kasan_release_vmalloc(orig_start, orig_end,
3739 va->va_start, va->va_end);
3743 spin_unlock(&free_vmap_area_lock);
3750 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3751 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3752 * @nr_vms: the number of allocated areas
3754 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3756 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3760 for (i = 0; i < nr_vms; i++)
3761 free_vm_area(vms[i]);
3764 #endif /* CONFIG_SMP */
3766 #ifdef CONFIG_PRINTK
3767 bool vmalloc_dump_obj(void *object)
3769 struct vm_struct *vm;
3770 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3772 vm = find_vm_area(objp);
3775 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3776 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3781 #ifdef CONFIG_PROC_FS
3782 static void *s_start(struct seq_file *m, loff_t *pos)
3783 __acquires(&vmap_purge_lock)
3784 __acquires(&vmap_area_lock)
3786 mutex_lock(&vmap_purge_lock);
3787 spin_lock(&vmap_area_lock);
3789 return seq_list_start(&vmap_area_list, *pos);
3792 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3794 return seq_list_next(p, &vmap_area_list, pos);
3797 static void s_stop(struct seq_file *m, void *p)
3798 __releases(&vmap_area_lock)
3799 __releases(&vmap_purge_lock)
3801 spin_unlock(&vmap_area_lock);
3802 mutex_unlock(&vmap_purge_lock);
3805 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3807 if (IS_ENABLED(CONFIG_NUMA)) {
3808 unsigned int nr, *counters = m->private;
3813 if (v->flags & VM_UNINITIALIZED)
3815 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3818 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3820 for (nr = 0; nr < v->nr_pages; nr++)
3821 counters[page_to_nid(v->pages[nr])]++;
3823 for_each_node_state(nr, N_HIGH_MEMORY)
3825 seq_printf(m, " N%u=%u", nr, counters[nr]);
3829 static void show_purge_info(struct seq_file *m)
3831 struct vmap_area *va;
3833 spin_lock(&purge_vmap_area_lock);
3834 list_for_each_entry(va, &purge_vmap_area_list, list) {
3835 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3836 (void *)va->va_start, (void *)va->va_end,
3837 va->va_end - va->va_start);
3839 spin_unlock(&purge_vmap_area_lock);
3842 static int s_show(struct seq_file *m, void *p)
3844 struct vmap_area *va;
3845 struct vm_struct *v;
3847 va = list_entry(p, struct vmap_area, list);
3850 * s_show can encounter race with remove_vm_area, !vm on behalf
3851 * of vmap area is being tear down or vm_map_ram allocation.
3854 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3855 (void *)va->va_start, (void *)va->va_end,
3856 va->va_end - va->va_start);
3863 seq_printf(m, "0x%pK-0x%pK %7ld",
3864 v->addr, v->addr + v->size, v->size);
3867 seq_printf(m, " %pS", v->caller);
3870 seq_printf(m, " pages=%d", v->nr_pages);
3873 seq_printf(m, " phys=%pa", &v->phys_addr);
3875 if (v->flags & VM_IOREMAP)
3876 seq_puts(m, " ioremap");
3878 if (v->flags & VM_ALLOC)
3879 seq_puts(m, " vmalloc");
3881 if (v->flags & VM_MAP)
3882 seq_puts(m, " vmap");
3884 if (v->flags & VM_USERMAP)
3885 seq_puts(m, " user");
3887 if (v->flags & VM_DMA_COHERENT)
3888 seq_puts(m, " dma-coherent");
3890 if (is_vmalloc_addr(v->pages))
3891 seq_puts(m, " vpages");
3893 show_numa_info(m, v);
3897 * As a final step, dump "unpurged" areas.
3899 if (list_is_last(&va->list, &vmap_area_list))
3905 static const struct seq_operations vmalloc_op = {
3912 static int __init proc_vmalloc_init(void)
3914 if (IS_ENABLED(CONFIG_NUMA))
3915 proc_create_seq_private("vmallocinfo", 0400, NULL,
3917 nr_node_ids * sizeof(unsigned int), NULL);
3919 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3922 module_init(proc_vmalloc_init);