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_exceed_addr(unsigned long addr)
804 struct vmap_area *va = NULL;
805 struct rb_node *n = vmap_area_root.rb_node;
808 struct vmap_area *tmp;
810 tmp = rb_entry(n, struct vmap_area, rb_node);
811 if (tmp->va_end > addr) {
813 if (tmp->va_start <= addr)
824 static struct vmap_area *__find_vmap_area(unsigned long addr)
826 struct rb_node *n = vmap_area_root.rb_node;
829 struct vmap_area *va;
831 va = rb_entry(n, struct vmap_area, rb_node);
832 if (addr < va->va_start)
834 else if (addr >= va->va_end)
844 * This function returns back addresses of parent node
845 * and its left or right link for further processing.
847 * Otherwise NULL is returned. In that case all further
848 * steps regarding inserting of conflicting overlap range
849 * have to be declined and actually considered as a bug.
851 static __always_inline struct rb_node **
852 find_va_links(struct vmap_area *va,
853 struct rb_root *root, struct rb_node *from,
854 struct rb_node **parent)
856 struct vmap_area *tmp_va;
857 struct rb_node **link;
860 link = &root->rb_node;
861 if (unlikely(!*link)) {
870 * Go to the bottom of the tree. When we hit the last point
871 * we end up with parent rb_node and correct direction, i name
872 * it link, where the new va->rb_node will be attached to.
875 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
878 * During the traversal we also do some sanity check.
879 * Trigger the BUG() if there are sides(left/right)
882 if (va->va_start < tmp_va->va_end &&
883 va->va_end <= tmp_va->va_start)
884 link = &(*link)->rb_left;
885 else if (va->va_end > tmp_va->va_start &&
886 va->va_start >= tmp_va->va_end)
887 link = &(*link)->rb_right;
889 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
890 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
896 *parent = &tmp_va->rb_node;
900 static __always_inline struct list_head *
901 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
903 struct list_head *list;
905 if (unlikely(!parent))
907 * The red-black tree where we try to find VA neighbors
908 * before merging or inserting is empty, i.e. it means
909 * there is no free vmap space. Normally it does not
910 * happen but we handle this case anyway.
914 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
915 return (&parent->rb_right == link ? list->next : list);
918 static __always_inline void
919 link_va(struct vmap_area *va, struct rb_root *root,
920 struct rb_node *parent, struct rb_node **link, struct list_head *head)
923 * VA is still not in the list, but we can
924 * identify its future previous list_head node.
926 if (likely(parent)) {
927 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
928 if (&parent->rb_right != link)
932 /* Insert to the rb-tree */
933 rb_link_node(&va->rb_node, parent, link);
934 if (root == &free_vmap_area_root) {
936 * Some explanation here. Just perform simple insertion
937 * to the tree. We do not set va->subtree_max_size to
938 * its current size before calling rb_insert_augmented().
939 * It is because of we populate the tree from the bottom
940 * to parent levels when the node _is_ in the tree.
942 * Therefore we set subtree_max_size to zero after insertion,
943 * to let __augment_tree_propagate_from() puts everything to
944 * the correct order later on.
946 rb_insert_augmented(&va->rb_node,
947 root, &free_vmap_area_rb_augment_cb);
948 va->subtree_max_size = 0;
950 rb_insert_color(&va->rb_node, root);
953 /* Address-sort this list */
954 list_add(&va->list, head);
957 static __always_inline void
958 unlink_va(struct vmap_area *va, struct rb_root *root)
960 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
963 if (root == &free_vmap_area_root)
964 rb_erase_augmented(&va->rb_node,
965 root, &free_vmap_area_rb_augment_cb);
967 rb_erase(&va->rb_node, root);
970 RB_CLEAR_NODE(&va->rb_node);
973 #if DEBUG_AUGMENT_PROPAGATE_CHECK
975 augment_tree_propagate_check(void)
977 struct vmap_area *va;
978 unsigned long computed_size;
980 list_for_each_entry(va, &free_vmap_area_list, list) {
981 computed_size = compute_subtree_max_size(va);
982 if (computed_size != va->subtree_max_size)
983 pr_emerg("tree is corrupted: %lu, %lu\n",
984 va_size(va), va->subtree_max_size);
990 * This function populates subtree_max_size from bottom to upper
991 * levels starting from VA point. The propagation must be done
992 * when VA size is modified by changing its va_start/va_end. Or
993 * in case of newly inserting of VA to the tree.
995 * It means that __augment_tree_propagate_from() must be called:
996 * - After VA has been inserted to the tree(free path);
997 * - After VA has been shrunk(allocation path);
998 * - After VA has been increased(merging path).
1000 * Please note that, it does not mean that upper parent nodes
1001 * and their subtree_max_size are recalculated all the time up
1010 * For example if we modify the node 4, shrinking it to 2, then
1011 * no any modification is required. If we shrink the node 2 to 1
1012 * its subtree_max_size is updated only, and set to 1. If we shrink
1013 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1014 * node becomes 4--6.
1016 static __always_inline void
1017 augment_tree_propagate_from(struct vmap_area *va)
1020 * Populate the tree from bottom towards the root until
1021 * the calculated maximum available size of checked node
1022 * is equal to its current one.
1024 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1026 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1027 augment_tree_propagate_check();
1032 insert_vmap_area(struct vmap_area *va,
1033 struct rb_root *root, struct list_head *head)
1035 struct rb_node **link;
1036 struct rb_node *parent;
1038 link = find_va_links(va, root, NULL, &parent);
1040 link_va(va, root, parent, link, head);
1044 insert_vmap_area_augment(struct vmap_area *va,
1045 struct rb_node *from, struct rb_root *root,
1046 struct list_head *head)
1048 struct rb_node **link;
1049 struct rb_node *parent;
1052 link = find_va_links(va, NULL, from, &parent);
1054 link = find_va_links(va, root, NULL, &parent);
1057 link_va(va, root, parent, link, head);
1058 augment_tree_propagate_from(va);
1063 * Merge de-allocated chunk of VA memory with previous
1064 * and next free blocks. If coalesce is not done a new
1065 * free area is inserted. If VA has been merged, it is
1068 * Please note, it can return NULL in case of overlap
1069 * ranges, followed by WARN() report. Despite it is a
1070 * buggy behaviour, a system can be alive and keep
1073 static __always_inline struct vmap_area *
1074 merge_or_add_vmap_area(struct vmap_area *va,
1075 struct rb_root *root, struct list_head *head)
1077 struct vmap_area *sibling;
1078 struct list_head *next;
1079 struct rb_node **link;
1080 struct rb_node *parent;
1081 bool merged = false;
1084 * Find a place in the tree where VA potentially will be
1085 * inserted, unless it is merged with its sibling/siblings.
1087 link = find_va_links(va, root, NULL, &parent);
1092 * Get next node of VA to check if merging can be done.
1094 next = get_va_next_sibling(parent, link);
1095 if (unlikely(next == NULL))
1101 * |<------VA------>|<-----Next----->|
1106 sibling = list_entry(next, struct vmap_area, list);
1107 if (sibling->va_start == va->va_end) {
1108 sibling->va_start = va->va_start;
1110 /* Free vmap_area object. */
1111 kmem_cache_free(vmap_area_cachep, va);
1113 /* Point to the new merged area. */
1122 * |<-----Prev----->|<------VA------>|
1126 if (next->prev != head) {
1127 sibling = list_entry(next->prev, struct vmap_area, list);
1128 if (sibling->va_end == va->va_start) {
1130 * If both neighbors are coalesced, it is important
1131 * to unlink the "next" node first, followed by merging
1132 * with "previous" one. Otherwise the tree might not be
1133 * fully populated if a sibling's augmented value is
1134 * "normalized" because of rotation operations.
1137 unlink_va(va, root);
1139 sibling->va_end = va->va_end;
1141 /* Free vmap_area object. */
1142 kmem_cache_free(vmap_area_cachep, va);
1144 /* Point to the new merged area. */
1152 link_va(va, root, parent, link, head);
1157 static __always_inline struct vmap_area *
1158 merge_or_add_vmap_area_augment(struct vmap_area *va,
1159 struct rb_root *root, struct list_head *head)
1161 va = merge_or_add_vmap_area(va, root, head);
1163 augment_tree_propagate_from(va);
1168 static __always_inline bool
1169 is_within_this_va(struct vmap_area *va, unsigned long size,
1170 unsigned long align, unsigned long vstart)
1172 unsigned long nva_start_addr;
1174 if (va->va_start > vstart)
1175 nva_start_addr = ALIGN(va->va_start, align);
1177 nva_start_addr = ALIGN(vstart, align);
1179 /* Can be overflowed due to big size or alignment. */
1180 if (nva_start_addr + size < nva_start_addr ||
1181 nva_start_addr < vstart)
1184 return (nva_start_addr + size <= va->va_end);
1188 * Find the first free block(lowest start address) in the tree,
1189 * that will accomplish the request corresponding to passing
1192 static __always_inline struct vmap_area *
1193 find_vmap_lowest_match(unsigned long size,
1194 unsigned long align, unsigned long vstart)
1196 struct vmap_area *va;
1197 struct rb_node *node;
1198 unsigned long length;
1200 /* Start from the root. */
1201 node = free_vmap_area_root.rb_node;
1203 /* Adjust the search size for alignment overhead. */
1204 length = size + align - 1;
1207 va = rb_entry(node, struct vmap_area, rb_node);
1209 if (get_subtree_max_size(node->rb_left) >= length &&
1210 vstart < va->va_start) {
1211 node = node->rb_left;
1213 if (is_within_this_va(va, size, align, vstart))
1217 * Does not make sense to go deeper towards the right
1218 * sub-tree if it does not have a free block that is
1219 * equal or bigger to the requested search length.
1221 if (get_subtree_max_size(node->rb_right) >= length) {
1222 node = node->rb_right;
1227 * OK. We roll back and find the first right sub-tree,
1228 * that will satisfy the search criteria. It can happen
1229 * only once due to "vstart" restriction.
1231 while ((node = rb_parent(node))) {
1232 va = rb_entry(node, struct vmap_area, rb_node);
1233 if (is_within_this_va(va, size, align, vstart))
1236 if (get_subtree_max_size(node->rb_right) >= length &&
1237 vstart <= va->va_start) {
1238 node = node->rb_right;
1248 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1249 #include <linux/random.h>
1251 static struct vmap_area *
1252 find_vmap_lowest_linear_match(unsigned long size,
1253 unsigned long align, unsigned long vstart)
1255 struct vmap_area *va;
1257 list_for_each_entry(va, &free_vmap_area_list, list) {
1258 if (!is_within_this_va(va, size, align, vstart))
1268 find_vmap_lowest_match_check(unsigned long size)
1270 struct vmap_area *va_1, *va_2;
1271 unsigned long vstart;
1274 get_random_bytes(&rnd, sizeof(rnd));
1275 vstart = VMALLOC_START + rnd;
1277 va_1 = find_vmap_lowest_match(size, 1, vstart);
1278 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
1281 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1282 va_1, va_2, vstart);
1288 FL_FIT_TYPE = 1, /* full fit */
1289 LE_FIT_TYPE = 2, /* left edge fit */
1290 RE_FIT_TYPE = 3, /* right edge fit */
1291 NE_FIT_TYPE = 4 /* no edge fit */
1294 static __always_inline enum fit_type
1295 classify_va_fit_type(struct vmap_area *va,
1296 unsigned long nva_start_addr, unsigned long size)
1300 /* Check if it is within VA. */
1301 if (nva_start_addr < va->va_start ||
1302 nva_start_addr + size > va->va_end)
1306 if (va->va_start == nva_start_addr) {
1307 if (va->va_end == nva_start_addr + size)
1311 } else if (va->va_end == nva_start_addr + size) {
1320 static __always_inline int
1321 adjust_va_to_fit_type(struct vmap_area *va,
1322 unsigned long nva_start_addr, unsigned long size,
1325 struct vmap_area *lva = NULL;
1327 if (type == FL_FIT_TYPE) {
1329 * No need to split VA, it fully fits.
1335 unlink_va(va, &free_vmap_area_root);
1336 kmem_cache_free(vmap_area_cachep, va);
1337 } else if (type == LE_FIT_TYPE) {
1339 * Split left edge of fit VA.
1345 va->va_start += size;
1346 } else if (type == RE_FIT_TYPE) {
1348 * Split right edge of fit VA.
1354 va->va_end = nva_start_addr;
1355 } else if (type == NE_FIT_TYPE) {
1357 * Split no edge of fit VA.
1363 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1364 if (unlikely(!lva)) {
1366 * For percpu allocator we do not do any pre-allocation
1367 * and leave it as it is. The reason is it most likely
1368 * never ends up with NE_FIT_TYPE splitting. In case of
1369 * percpu allocations offsets and sizes are aligned to
1370 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1371 * are its main fitting cases.
1373 * There are a few exceptions though, as an example it is
1374 * a first allocation (early boot up) when we have "one"
1375 * big free space that has to be split.
1377 * Also we can hit this path in case of regular "vmap"
1378 * allocations, if "this" current CPU was not preloaded.
1379 * See the comment in alloc_vmap_area() why. If so, then
1380 * GFP_NOWAIT is used instead to get an extra object for
1381 * split purpose. That is rare and most time does not
1384 * What happens if an allocation gets failed. Basically,
1385 * an "overflow" path is triggered to purge lazily freed
1386 * areas to free some memory, then, the "retry" path is
1387 * triggered to repeat one more time. See more details
1388 * in alloc_vmap_area() function.
1390 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1396 * Build the remainder.
1398 lva->va_start = va->va_start;
1399 lva->va_end = nva_start_addr;
1402 * Shrink this VA to remaining size.
1404 va->va_start = nva_start_addr + size;
1409 if (type != FL_FIT_TYPE) {
1410 augment_tree_propagate_from(va);
1412 if (lva) /* type == NE_FIT_TYPE */
1413 insert_vmap_area_augment(lva, &va->rb_node,
1414 &free_vmap_area_root, &free_vmap_area_list);
1421 * Returns a start address of the newly allocated area, if success.
1422 * Otherwise a vend is returned that indicates failure.
1424 static __always_inline unsigned long
1425 __alloc_vmap_area(unsigned long size, unsigned long align,
1426 unsigned long vstart, unsigned long vend)
1428 unsigned long nva_start_addr;
1429 struct vmap_area *va;
1433 va = find_vmap_lowest_match(size, align, vstart);
1437 if (va->va_start > vstart)
1438 nva_start_addr = ALIGN(va->va_start, align);
1440 nva_start_addr = ALIGN(vstart, align);
1442 /* Check the "vend" restriction. */
1443 if (nva_start_addr + size > vend)
1446 /* Classify what we have found. */
1447 type = classify_va_fit_type(va, nva_start_addr, size);
1448 if (WARN_ON_ONCE(type == NOTHING_FIT))
1451 /* Update the free vmap_area. */
1452 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1456 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1457 find_vmap_lowest_match_check(size);
1460 return nva_start_addr;
1464 * Free a region of KVA allocated by alloc_vmap_area
1466 static void free_vmap_area(struct vmap_area *va)
1469 * Remove from the busy tree/list.
1471 spin_lock(&vmap_area_lock);
1472 unlink_va(va, &vmap_area_root);
1473 spin_unlock(&vmap_area_lock);
1476 * Insert/Merge it back to the free tree/list.
1478 spin_lock(&free_vmap_area_lock);
1479 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1480 spin_unlock(&free_vmap_area_lock);
1484 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1486 struct vmap_area *va = NULL;
1489 * Preload this CPU with one extra vmap_area object. It is used
1490 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1491 * a CPU that does an allocation is preloaded.
1493 * We do it in non-atomic context, thus it allows us to use more
1494 * permissive allocation masks to be more stable under low memory
1495 * condition and high memory pressure.
1497 if (!this_cpu_read(ne_fit_preload_node))
1498 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1502 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1503 kmem_cache_free(vmap_area_cachep, va);
1507 * Allocate a region of KVA of the specified size and alignment, within the
1510 static struct vmap_area *alloc_vmap_area(unsigned long size,
1511 unsigned long align,
1512 unsigned long vstart, unsigned long vend,
1513 int node, gfp_t gfp_mask)
1515 struct vmap_area *va;
1516 unsigned long freed;
1522 BUG_ON(offset_in_page(size));
1523 BUG_ON(!is_power_of_2(align));
1525 if (unlikely(!vmap_initialized))
1526 return ERR_PTR(-EBUSY);
1529 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1531 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1533 return ERR_PTR(-ENOMEM);
1536 * Only scan the relevant parts containing pointers to other objects
1537 * to avoid false negatives.
1539 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1542 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1543 addr = __alloc_vmap_area(size, align, vstart, vend);
1544 spin_unlock(&free_vmap_area_lock);
1547 * If an allocation fails, the "vend" address is
1548 * returned. Therefore trigger the overflow path.
1550 if (unlikely(addr == vend))
1553 va->va_start = addr;
1554 va->va_end = addr + size;
1557 spin_lock(&vmap_area_lock);
1558 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1559 spin_unlock(&vmap_area_lock);
1561 BUG_ON(!IS_ALIGNED(va->va_start, align));
1562 BUG_ON(va->va_start < vstart);
1563 BUG_ON(va->va_end > vend);
1565 ret = kasan_populate_vmalloc(addr, size);
1568 return ERR_PTR(ret);
1575 purge_vmap_area_lazy();
1581 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1588 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1589 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1592 kmem_cache_free(vmap_area_cachep, va);
1593 return ERR_PTR(-EBUSY);
1596 int register_vmap_purge_notifier(struct notifier_block *nb)
1598 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1600 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1602 int unregister_vmap_purge_notifier(struct notifier_block *nb)
1604 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1606 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1609 * lazy_max_pages is the maximum amount of virtual address space we gather up
1610 * before attempting to purge with a TLB flush.
1612 * There is a tradeoff here: a larger number will cover more kernel page tables
1613 * and take slightly longer to purge, but it will linearly reduce the number of
1614 * global TLB flushes that must be performed. It would seem natural to scale
1615 * this number up linearly with the number of CPUs (because vmapping activity
1616 * could also scale linearly with the number of CPUs), however it is likely
1617 * that in practice, workloads might be constrained in other ways that mean
1618 * vmap activity will not scale linearly with CPUs. Also, I want to be
1619 * conservative and not introduce a big latency on huge systems, so go with
1620 * a less aggressive log scale. It will still be an improvement over the old
1621 * code, and it will be simple to change the scale factor if we find that it
1622 * becomes a problem on bigger systems.
1624 static unsigned long lazy_max_pages(void)
1628 log = fls(num_online_cpus());
1630 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1633 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1636 * Serialize vmap purging. There is no actual critical section protected
1637 * by this look, but we want to avoid concurrent calls for performance
1638 * reasons and to make the pcpu_get_vm_areas more deterministic.
1640 static DEFINE_MUTEX(vmap_purge_lock);
1642 /* for per-CPU blocks */
1643 static void purge_fragmented_blocks_allcpus(void);
1645 #ifdef CONFIG_X86_64
1647 * called before a call to iounmap() if the caller wants vm_area_struct's
1648 * immediately freed.
1650 void set_iounmap_nonlazy(void)
1652 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1654 #endif /* CONFIG_X86_64 */
1657 * Purges all lazily-freed vmap areas.
1659 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1661 unsigned long resched_threshold;
1662 struct list_head local_pure_list;
1663 struct vmap_area *va, *n_va;
1665 lockdep_assert_held(&vmap_purge_lock);
1667 spin_lock(&purge_vmap_area_lock);
1668 purge_vmap_area_root = RB_ROOT;
1669 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1670 spin_unlock(&purge_vmap_area_lock);
1672 if (unlikely(list_empty(&local_pure_list)))
1676 list_first_entry(&local_pure_list,
1677 struct vmap_area, list)->va_start);
1680 list_last_entry(&local_pure_list,
1681 struct vmap_area, list)->va_end);
1683 flush_tlb_kernel_range(start, end);
1684 resched_threshold = lazy_max_pages() << 1;
1686 spin_lock(&free_vmap_area_lock);
1687 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1688 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1689 unsigned long orig_start = va->va_start;
1690 unsigned long orig_end = va->va_end;
1693 * Finally insert or merge lazily-freed area. It is
1694 * detached and there is no need to "unlink" it from
1697 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1698 &free_vmap_area_list);
1703 if (is_vmalloc_or_module_addr((void *)orig_start))
1704 kasan_release_vmalloc(orig_start, orig_end,
1705 va->va_start, va->va_end);
1707 atomic_long_sub(nr, &vmap_lazy_nr);
1709 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1710 cond_resched_lock(&free_vmap_area_lock);
1712 spin_unlock(&free_vmap_area_lock);
1717 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1718 * is already purging.
1720 static void try_purge_vmap_area_lazy(void)
1722 if (mutex_trylock(&vmap_purge_lock)) {
1723 __purge_vmap_area_lazy(ULONG_MAX, 0);
1724 mutex_unlock(&vmap_purge_lock);
1729 * Kick off a purge of the outstanding lazy areas.
1731 static void purge_vmap_area_lazy(void)
1733 mutex_lock(&vmap_purge_lock);
1734 purge_fragmented_blocks_allcpus();
1735 __purge_vmap_area_lazy(ULONG_MAX, 0);
1736 mutex_unlock(&vmap_purge_lock);
1740 * Free a vmap area, caller ensuring that the area has been unmapped
1741 * and flush_cache_vunmap had been called for the correct range
1744 static void free_vmap_area_noflush(struct vmap_area *va)
1746 unsigned long nr_lazy;
1748 spin_lock(&vmap_area_lock);
1749 unlink_va(va, &vmap_area_root);
1750 spin_unlock(&vmap_area_lock);
1752 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1753 PAGE_SHIFT, &vmap_lazy_nr);
1756 * Merge or place it to the purge tree/list.
1758 spin_lock(&purge_vmap_area_lock);
1759 merge_or_add_vmap_area(va,
1760 &purge_vmap_area_root, &purge_vmap_area_list);
1761 spin_unlock(&purge_vmap_area_lock);
1763 /* After this point, we may free va at any time */
1764 if (unlikely(nr_lazy > lazy_max_pages()))
1765 try_purge_vmap_area_lazy();
1769 * Free and unmap a vmap area
1771 static void free_unmap_vmap_area(struct vmap_area *va)
1773 flush_cache_vunmap(va->va_start, va->va_end);
1774 vunmap_range_noflush(va->va_start, va->va_end);
1775 if (debug_pagealloc_enabled_static())
1776 flush_tlb_kernel_range(va->va_start, va->va_end);
1778 free_vmap_area_noflush(va);
1781 static struct vmap_area *find_vmap_area(unsigned long addr)
1783 struct vmap_area *va;
1785 spin_lock(&vmap_area_lock);
1786 va = __find_vmap_area(addr);
1787 spin_unlock(&vmap_area_lock);
1792 /*** Per cpu kva allocator ***/
1795 * vmap space is limited especially on 32 bit architectures. Ensure there is
1796 * room for at least 16 percpu vmap blocks per CPU.
1799 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1800 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1801 * instead (we just need a rough idea)
1803 #if BITS_PER_LONG == 32
1804 #define VMALLOC_SPACE (128UL*1024*1024)
1806 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1809 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1810 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1811 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1812 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1813 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1814 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1815 #define VMAP_BBMAP_BITS \
1816 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1817 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1818 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1820 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1822 struct vmap_block_queue {
1824 struct list_head free;
1829 struct vmap_area *va;
1830 unsigned long free, dirty;
1831 unsigned long dirty_min, dirty_max; /*< dirty range */
1832 struct list_head free_list;
1833 struct rcu_head rcu_head;
1834 struct list_head purge;
1837 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1838 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1841 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1842 * in the free path. Could get rid of this if we change the API to return a
1843 * "cookie" from alloc, to be passed to free. But no big deal yet.
1845 static DEFINE_XARRAY(vmap_blocks);
1848 * We should probably have a fallback mechanism to allocate virtual memory
1849 * out of partially filled vmap blocks. However vmap block sizing should be
1850 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1854 static unsigned long addr_to_vb_idx(unsigned long addr)
1856 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1857 addr /= VMAP_BLOCK_SIZE;
1861 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1865 addr = va_start + (pages_off << PAGE_SHIFT);
1866 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1867 return (void *)addr;
1871 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1872 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1873 * @order: how many 2^order pages should be occupied in newly allocated block
1874 * @gfp_mask: flags for the page level allocator
1876 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1878 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1880 struct vmap_block_queue *vbq;
1881 struct vmap_block *vb;
1882 struct vmap_area *va;
1883 unsigned long vb_idx;
1887 node = numa_node_id();
1889 vb = kmalloc_node(sizeof(struct vmap_block),
1890 gfp_mask & GFP_RECLAIM_MASK, node);
1892 return ERR_PTR(-ENOMEM);
1894 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1895 VMALLOC_START, VMALLOC_END,
1899 return ERR_CAST(va);
1902 vaddr = vmap_block_vaddr(va->va_start, 0);
1903 spin_lock_init(&vb->lock);
1905 /* At least something should be left free */
1906 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1907 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1909 vb->dirty_min = VMAP_BBMAP_BITS;
1911 INIT_LIST_HEAD(&vb->free_list);
1913 vb_idx = addr_to_vb_idx(va->va_start);
1914 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1918 return ERR_PTR(err);
1921 vbq = &get_cpu_var(vmap_block_queue);
1922 spin_lock(&vbq->lock);
1923 list_add_tail_rcu(&vb->free_list, &vbq->free);
1924 spin_unlock(&vbq->lock);
1925 put_cpu_var(vmap_block_queue);
1930 static void free_vmap_block(struct vmap_block *vb)
1932 struct vmap_block *tmp;
1934 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1937 free_vmap_area_noflush(vb->va);
1938 kfree_rcu(vb, rcu_head);
1941 static void purge_fragmented_blocks(int cpu)
1944 struct vmap_block *vb;
1945 struct vmap_block *n_vb;
1946 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1949 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1951 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1954 spin_lock(&vb->lock);
1955 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1956 vb->free = 0; /* prevent further allocs after releasing lock */
1957 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1959 vb->dirty_max = VMAP_BBMAP_BITS;
1960 spin_lock(&vbq->lock);
1961 list_del_rcu(&vb->free_list);
1962 spin_unlock(&vbq->lock);
1963 spin_unlock(&vb->lock);
1964 list_add_tail(&vb->purge, &purge);
1966 spin_unlock(&vb->lock);
1970 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1971 list_del(&vb->purge);
1972 free_vmap_block(vb);
1976 static void purge_fragmented_blocks_allcpus(void)
1980 for_each_possible_cpu(cpu)
1981 purge_fragmented_blocks(cpu);
1984 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1986 struct vmap_block_queue *vbq;
1987 struct vmap_block *vb;
1991 BUG_ON(offset_in_page(size));
1992 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1993 if (WARN_ON(size == 0)) {
1995 * Allocating 0 bytes isn't what caller wants since
1996 * get_order(0) returns funny result. Just warn and terminate
2001 order = get_order(size);
2004 vbq = &get_cpu_var(vmap_block_queue);
2005 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2006 unsigned long pages_off;
2008 spin_lock(&vb->lock);
2009 if (vb->free < (1UL << order)) {
2010 spin_unlock(&vb->lock);
2014 pages_off = VMAP_BBMAP_BITS - vb->free;
2015 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2016 vb->free -= 1UL << order;
2017 if (vb->free == 0) {
2018 spin_lock(&vbq->lock);
2019 list_del_rcu(&vb->free_list);
2020 spin_unlock(&vbq->lock);
2023 spin_unlock(&vb->lock);
2027 put_cpu_var(vmap_block_queue);
2030 /* Allocate new block if nothing was found */
2032 vaddr = new_vmap_block(order, gfp_mask);
2037 static void vb_free(unsigned long addr, unsigned long size)
2039 unsigned long offset;
2041 struct vmap_block *vb;
2043 BUG_ON(offset_in_page(size));
2044 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2046 flush_cache_vunmap(addr, addr + size);
2048 order = get_order(size);
2049 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2050 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2052 vunmap_range_noflush(addr, addr + size);
2054 if (debug_pagealloc_enabled_static())
2055 flush_tlb_kernel_range(addr, addr + size);
2057 spin_lock(&vb->lock);
2059 /* Expand dirty range */
2060 vb->dirty_min = min(vb->dirty_min, offset);
2061 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2063 vb->dirty += 1UL << order;
2064 if (vb->dirty == VMAP_BBMAP_BITS) {
2066 spin_unlock(&vb->lock);
2067 free_vmap_block(vb);
2069 spin_unlock(&vb->lock);
2072 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2076 if (unlikely(!vmap_initialized))
2081 for_each_possible_cpu(cpu) {
2082 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2083 struct vmap_block *vb;
2086 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2087 spin_lock(&vb->lock);
2088 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2089 unsigned long va_start = vb->va->va_start;
2092 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2093 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2095 start = min(s, start);
2100 spin_unlock(&vb->lock);
2105 mutex_lock(&vmap_purge_lock);
2106 purge_fragmented_blocks_allcpus();
2107 if (!__purge_vmap_area_lazy(start, end) && flush)
2108 flush_tlb_kernel_range(start, end);
2109 mutex_unlock(&vmap_purge_lock);
2113 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2115 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2116 * to amortize TLB flushing overheads. What this means is that any page you
2117 * have now, may, in a former life, have been mapped into kernel virtual
2118 * address by the vmap layer and so there might be some CPUs with TLB entries
2119 * still referencing that page (additional to the regular 1:1 kernel mapping).
2121 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2122 * be sure that none of the pages we have control over will have any aliases
2123 * from the vmap layer.
2125 void vm_unmap_aliases(void)
2127 unsigned long start = ULONG_MAX, end = 0;
2130 _vm_unmap_aliases(start, end, flush);
2132 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2135 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2136 * @mem: the pointer returned by vm_map_ram
2137 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2139 void vm_unmap_ram(const void *mem, unsigned int count)
2141 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2142 unsigned long addr = (unsigned long)mem;
2143 struct vmap_area *va;
2147 BUG_ON(addr < VMALLOC_START);
2148 BUG_ON(addr > VMALLOC_END);
2149 BUG_ON(!PAGE_ALIGNED(addr));
2151 kasan_poison_vmalloc(mem, size);
2153 if (likely(count <= VMAP_MAX_ALLOC)) {
2154 debug_check_no_locks_freed(mem, size);
2155 vb_free(addr, size);
2159 va = find_vmap_area(addr);
2161 debug_check_no_locks_freed((void *)va->va_start,
2162 (va->va_end - va->va_start));
2163 free_unmap_vmap_area(va);
2165 EXPORT_SYMBOL(vm_unmap_ram);
2168 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2169 * @pages: an array of pointers to the pages to be mapped
2170 * @count: number of pages
2171 * @node: prefer to allocate data structures on this node
2173 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2174 * faster than vmap so it's good. But if you mix long-life and short-life
2175 * objects with vm_map_ram(), it could consume lots of address space through
2176 * fragmentation (especially on a 32bit machine). You could see failures in
2177 * the end. Please use this function for short-lived objects.
2179 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2181 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2183 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2187 if (likely(count <= VMAP_MAX_ALLOC)) {
2188 mem = vb_alloc(size, GFP_KERNEL);
2191 addr = (unsigned long)mem;
2193 struct vmap_area *va;
2194 va = alloc_vmap_area(size, PAGE_SIZE,
2195 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2199 addr = va->va_start;
2203 kasan_unpoison_vmalloc(mem, size);
2205 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2206 pages, PAGE_SHIFT) < 0) {
2207 vm_unmap_ram(mem, count);
2213 EXPORT_SYMBOL(vm_map_ram);
2215 static struct vm_struct *vmlist __initdata;
2217 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2219 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2220 return vm->page_order;
2226 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2228 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2229 vm->page_order = order;
2236 * vm_area_add_early - add vmap area early during boot
2237 * @vm: vm_struct to add
2239 * This function is used to add fixed kernel vm area to vmlist before
2240 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2241 * should contain proper values and the other fields should be zero.
2243 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2245 void __init vm_area_add_early(struct vm_struct *vm)
2247 struct vm_struct *tmp, **p;
2249 BUG_ON(vmap_initialized);
2250 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2251 if (tmp->addr >= vm->addr) {
2252 BUG_ON(tmp->addr < vm->addr + vm->size);
2255 BUG_ON(tmp->addr + tmp->size > vm->addr);
2262 * vm_area_register_early - register vmap area early during boot
2263 * @vm: vm_struct to register
2264 * @align: requested alignment
2266 * This function is used to register kernel vm area before
2267 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2268 * proper values on entry and other fields should be zero. On return,
2269 * vm->addr contains the allocated address.
2271 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2273 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2275 static size_t vm_init_off __initdata;
2278 addr = ALIGN(VMALLOC_START + vm_init_off, align);
2279 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
2281 vm->addr = (void *)addr;
2283 vm_area_add_early(vm);
2286 static void vmap_init_free_space(void)
2288 unsigned long vmap_start = 1;
2289 const unsigned long vmap_end = ULONG_MAX;
2290 struct vmap_area *busy, *free;
2294 * -|-----|.....|-----|-----|-----|.....|-
2296 * |<--------------------------------->|
2298 list_for_each_entry(busy, &vmap_area_list, list) {
2299 if (busy->va_start - vmap_start > 0) {
2300 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2301 if (!WARN_ON_ONCE(!free)) {
2302 free->va_start = vmap_start;
2303 free->va_end = busy->va_start;
2305 insert_vmap_area_augment(free, NULL,
2306 &free_vmap_area_root,
2307 &free_vmap_area_list);
2311 vmap_start = busy->va_end;
2314 if (vmap_end - vmap_start > 0) {
2315 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2316 if (!WARN_ON_ONCE(!free)) {
2317 free->va_start = vmap_start;
2318 free->va_end = vmap_end;
2320 insert_vmap_area_augment(free, NULL,
2321 &free_vmap_area_root,
2322 &free_vmap_area_list);
2327 void __init vmalloc_init(void)
2329 struct vmap_area *va;
2330 struct vm_struct *tmp;
2334 * Create the cache for vmap_area objects.
2336 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2338 for_each_possible_cpu(i) {
2339 struct vmap_block_queue *vbq;
2340 struct vfree_deferred *p;
2342 vbq = &per_cpu(vmap_block_queue, i);
2343 spin_lock_init(&vbq->lock);
2344 INIT_LIST_HEAD(&vbq->free);
2345 p = &per_cpu(vfree_deferred, i);
2346 init_llist_head(&p->list);
2347 INIT_WORK(&p->wq, free_work);
2350 /* Import existing vmlist entries. */
2351 for (tmp = vmlist; tmp; tmp = tmp->next) {
2352 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2353 if (WARN_ON_ONCE(!va))
2356 va->va_start = (unsigned long)tmp->addr;
2357 va->va_end = va->va_start + tmp->size;
2359 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2363 * Now we can initialize a free vmap space.
2365 vmap_init_free_space();
2366 vmap_initialized = true;
2369 static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2370 struct vmap_area *va, unsigned long flags, const void *caller)
2373 vm->addr = (void *)va->va_start;
2374 vm->size = va->va_end - va->va_start;
2375 vm->caller = caller;
2379 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2380 unsigned long flags, const void *caller)
2382 spin_lock(&vmap_area_lock);
2383 setup_vmalloc_vm_locked(vm, va, flags, caller);
2384 spin_unlock(&vmap_area_lock);
2387 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2390 * Before removing VM_UNINITIALIZED,
2391 * we should make sure that vm has proper values.
2392 * Pair with smp_rmb() in show_numa_info().
2395 vm->flags &= ~VM_UNINITIALIZED;
2398 static struct vm_struct *__get_vm_area_node(unsigned long size,
2399 unsigned long align, unsigned long shift, unsigned long flags,
2400 unsigned long start, unsigned long end, int node,
2401 gfp_t gfp_mask, const void *caller)
2403 struct vmap_area *va;
2404 struct vm_struct *area;
2405 unsigned long requested_size = size;
2407 BUG_ON(in_interrupt());
2408 size = ALIGN(size, 1ul << shift);
2409 if (unlikely(!size))
2412 if (flags & VM_IOREMAP)
2413 align = 1ul << clamp_t(int, get_count_order_long(size),
2414 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2416 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2417 if (unlikely(!area))
2420 if (!(flags & VM_NO_GUARD))
2423 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2429 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2431 setup_vmalloc_vm(area, va, flags, caller);
2436 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2437 unsigned long start, unsigned long end,
2440 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2441 NUMA_NO_NODE, GFP_KERNEL, caller);
2445 * get_vm_area - reserve a contiguous kernel virtual area
2446 * @size: size of the area
2447 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2449 * Search an area of @size in the kernel virtual mapping area,
2450 * and reserved it for out purposes. Returns the area descriptor
2451 * on success or %NULL on failure.
2453 * Return: the area descriptor on success or %NULL on failure.
2455 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2457 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2458 VMALLOC_START, VMALLOC_END,
2459 NUMA_NO_NODE, GFP_KERNEL,
2460 __builtin_return_address(0));
2463 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2466 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2467 VMALLOC_START, VMALLOC_END,
2468 NUMA_NO_NODE, GFP_KERNEL, caller);
2472 * find_vm_area - find a continuous kernel virtual area
2473 * @addr: base address
2475 * Search for the kernel VM area starting at @addr, and return it.
2476 * It is up to the caller to do all required locking to keep the returned
2479 * Return: the area descriptor on success or %NULL on failure.
2481 struct vm_struct *find_vm_area(const void *addr)
2483 struct vmap_area *va;
2485 va = find_vmap_area((unsigned long)addr);
2493 * remove_vm_area - find and remove a continuous kernel virtual area
2494 * @addr: base address
2496 * Search for the kernel VM area starting at @addr, and remove it.
2497 * This function returns the found VM area, but using it is NOT safe
2498 * on SMP machines, except for its size or flags.
2500 * Return: the area descriptor on success or %NULL on failure.
2502 struct vm_struct *remove_vm_area(const void *addr)
2504 struct vmap_area *va;
2508 spin_lock(&vmap_area_lock);
2509 va = __find_vmap_area((unsigned long)addr);
2511 struct vm_struct *vm = va->vm;
2514 spin_unlock(&vmap_area_lock);
2516 kasan_free_shadow(vm);
2517 free_unmap_vmap_area(va);
2522 spin_unlock(&vmap_area_lock);
2526 static inline void set_area_direct_map(const struct vm_struct *area,
2527 int (*set_direct_map)(struct page *page))
2531 /* HUGE_VMALLOC passes small pages to set_direct_map */
2532 for (i = 0; i < area->nr_pages; i++)
2533 if (page_address(area->pages[i]))
2534 set_direct_map(area->pages[i]);
2537 /* Handle removing and resetting vm mappings related to the vm_struct. */
2538 static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2540 unsigned long start = ULONG_MAX, end = 0;
2541 unsigned int page_order = vm_area_page_order(area);
2542 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2546 remove_vm_area(area->addr);
2548 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2553 * If not deallocating pages, just do the flush of the VM area and
2556 if (!deallocate_pages) {
2562 * If execution gets here, flush the vm mapping and reset the direct
2563 * map. Find the start and end range of the direct mappings to make sure
2564 * the vm_unmap_aliases() flush includes the direct map.
2566 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2567 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2569 unsigned long page_size;
2571 page_size = PAGE_SIZE << page_order;
2572 start = min(addr, start);
2573 end = max(addr + page_size, end);
2579 * Set direct map to something invalid so that it won't be cached if
2580 * there are any accesses after the TLB flush, then flush the TLB and
2581 * reset the direct map permissions to the default.
2583 set_area_direct_map(area, set_direct_map_invalid_noflush);
2584 _vm_unmap_aliases(start, end, flush_dmap);
2585 set_area_direct_map(area, set_direct_map_default_noflush);
2588 static void __vunmap(const void *addr, int deallocate_pages)
2590 struct vm_struct *area;
2595 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2599 area = find_vm_area(addr);
2600 if (unlikely(!area)) {
2601 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2606 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2607 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2609 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2611 vm_remove_mappings(area, deallocate_pages);
2613 if (deallocate_pages) {
2614 unsigned int page_order = vm_area_page_order(area);
2617 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2618 struct page *page = area->pages[i];
2621 __free_pages(page, page_order);
2624 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2626 kvfree(area->pages);
2632 static inline void __vfree_deferred(const void *addr)
2635 * Use raw_cpu_ptr() because this can be called from preemptible
2636 * context. Preemption is absolutely fine here, because the llist_add()
2637 * implementation is lockless, so it works even if we are adding to
2638 * another cpu's list. schedule_work() should be fine with this too.
2640 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2642 if (llist_add((struct llist_node *)addr, &p->list))
2643 schedule_work(&p->wq);
2647 * vfree_atomic - release memory allocated by vmalloc()
2648 * @addr: memory base address
2650 * This one is just like vfree() but can be called in any atomic context
2653 void vfree_atomic(const void *addr)
2657 kmemleak_free(addr);
2661 __vfree_deferred(addr);
2664 static void __vfree(const void *addr)
2666 if (unlikely(in_interrupt()))
2667 __vfree_deferred(addr);
2673 * vfree - Release memory allocated by vmalloc()
2674 * @addr: Memory base address
2676 * Free the virtually continuous memory area starting at @addr, as obtained
2677 * from one of the vmalloc() family of APIs. This will usually also free the
2678 * physical memory underlying the virtual allocation, but that memory is
2679 * reference counted, so it will not be freed until the last user goes away.
2681 * If @addr is NULL, no operation is performed.
2684 * May sleep if called *not* from interrupt context.
2685 * Must not be called in NMI context (strictly speaking, it could be
2686 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2687 * conventions for vfree() arch-dependent would be a really bad idea).
2689 void vfree(const void *addr)
2693 kmemleak_free(addr);
2695 might_sleep_if(!in_interrupt());
2702 EXPORT_SYMBOL(vfree);
2705 * vunmap - release virtual mapping obtained by vmap()
2706 * @addr: memory base address
2708 * Free the virtually contiguous memory area starting at @addr,
2709 * which was created from the page array passed to vmap().
2711 * Must not be called in interrupt context.
2713 void vunmap(const void *addr)
2715 BUG_ON(in_interrupt());
2720 EXPORT_SYMBOL(vunmap);
2723 * vmap - map an array of pages into virtually contiguous space
2724 * @pages: array of page pointers
2725 * @count: number of pages to map
2726 * @flags: vm_area->flags
2727 * @prot: page protection for the mapping
2729 * Maps @count pages from @pages into contiguous kernel virtual space.
2730 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2731 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2732 * are transferred from the caller to vmap(), and will be freed / dropped when
2733 * vfree() is called on the return value.
2735 * Return: the address of the area or %NULL on failure
2737 void *vmap(struct page **pages, unsigned int count,
2738 unsigned long flags, pgprot_t prot)
2740 struct vm_struct *area;
2742 unsigned long size; /* In bytes */
2746 if (count > totalram_pages())
2749 size = (unsigned long)count << PAGE_SHIFT;
2750 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2754 addr = (unsigned long)area->addr;
2755 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2756 pages, PAGE_SHIFT) < 0) {
2761 if (flags & VM_MAP_PUT_PAGES) {
2762 area->pages = pages;
2763 area->nr_pages = count;
2767 EXPORT_SYMBOL(vmap);
2769 #ifdef CONFIG_VMAP_PFN
2770 struct vmap_pfn_data {
2771 unsigned long *pfns;
2776 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2778 struct vmap_pfn_data *data = private;
2780 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2782 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2787 * vmap_pfn - map an array of PFNs into virtually contiguous space
2788 * @pfns: array of PFNs
2789 * @count: number of pages to map
2790 * @prot: page protection for the mapping
2792 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2793 * the start address of the mapping.
2795 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2797 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2798 struct vm_struct *area;
2800 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2801 __builtin_return_address(0));
2804 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2805 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2811 EXPORT_SYMBOL_GPL(vmap_pfn);
2812 #endif /* CONFIG_VMAP_PFN */
2814 static inline unsigned int
2815 vm_area_alloc_pages(gfp_t gfp, int nid,
2816 unsigned int order, unsigned int nr_pages, struct page **pages)
2818 unsigned int nr_allocated = 0;
2821 * For order-0 pages we make use of bulk allocator, if
2822 * the page array is partly or not at all populated due
2823 * to fails, fallback to a single page allocator that is
2827 while (nr_allocated < nr_pages) {
2828 unsigned int nr, nr_pages_request;
2831 * A maximum allowed request is hard-coded and is 100
2832 * pages per call. That is done in order to prevent a
2833 * long preemption off scenario in the bulk-allocator
2834 * so the range is [1:100].
2836 nr_pages_request = min(100U, nr_pages - nr_allocated);
2838 nr = alloc_pages_bulk_array_node(gfp, nid,
2839 nr_pages_request, pages + nr_allocated);
2845 * If zero or pages were obtained partly,
2846 * fallback to a single page allocator.
2848 if (nr != nr_pages_request)
2853 * Compound pages required for remap_vmalloc_page if
2858 /* High-order pages or fallback path if "bulk" fails. */
2859 while (nr_allocated < nr_pages) {
2863 page = alloc_pages_node(nid, gfp, order);
2864 if (unlikely(!page))
2868 * Careful, we allocate and map page-order pages, but
2869 * tracking is done per PAGE_SIZE page so as to keep the
2870 * vm_struct APIs independent of the physical/mapped size.
2872 for (i = 0; i < (1U << order); i++)
2873 pages[nr_allocated + i] = page + i;
2876 nr_allocated += 1U << order;
2879 return nr_allocated;
2882 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2883 pgprot_t prot, unsigned int page_shift,
2886 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2887 unsigned long addr = (unsigned long)area->addr;
2888 unsigned long size = get_vm_area_size(area);
2889 unsigned long array_size;
2890 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2891 unsigned int page_order;
2893 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2894 gfp_mask |= __GFP_NOWARN;
2895 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2896 gfp_mask |= __GFP_HIGHMEM;
2898 /* Please note that the recursion is strictly bounded. */
2899 if (array_size > PAGE_SIZE) {
2900 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2903 area->pages = kmalloc_node(array_size, nested_gfp, node);
2907 warn_alloc(gfp_mask, NULL,
2908 "vmalloc error: size %lu, failed to allocated page array size %lu",
2909 nr_small_pages * PAGE_SIZE, array_size);
2914 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2915 page_order = vm_area_page_order(area);
2917 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2918 page_order, nr_small_pages, area->pages);
2920 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2923 * If not enough pages were obtained to accomplish an
2924 * allocation request, free them via __vfree() if any.
2926 if (area->nr_pages != nr_small_pages) {
2927 warn_alloc(gfp_mask, NULL,
2928 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2929 area->nr_pages * PAGE_SIZE, page_order);
2933 if (vmap_pages_range(addr, addr + size, prot, area->pages,
2935 warn_alloc(gfp_mask, NULL,
2936 "vmalloc error: size %lu, failed to map pages",
2937 area->nr_pages * PAGE_SIZE);
2944 __vfree(area->addr);
2949 * __vmalloc_node_range - allocate virtually contiguous memory
2950 * @size: allocation size
2951 * @align: desired alignment
2952 * @start: vm area range start
2953 * @end: vm area range end
2954 * @gfp_mask: flags for the page level allocator
2955 * @prot: protection mask for the allocated pages
2956 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2957 * @node: node to use for allocation or NUMA_NO_NODE
2958 * @caller: caller's return address
2960 * Allocate enough pages to cover @size from the page level
2961 * allocator with @gfp_mask flags. Map them into contiguous
2962 * kernel virtual space, using a pagetable protection of @prot.
2964 * Return: the address of the area or %NULL on failure
2966 void *__vmalloc_node_range(unsigned long size, unsigned long align,
2967 unsigned long start, unsigned long end, gfp_t gfp_mask,
2968 pgprot_t prot, unsigned long vm_flags, int node,
2971 struct vm_struct *area;
2973 unsigned long real_size = size;
2974 unsigned long real_align = align;
2975 unsigned int shift = PAGE_SHIFT;
2977 if (WARN_ON_ONCE(!size))
2980 if ((size >> PAGE_SHIFT) > totalram_pages()) {
2981 warn_alloc(gfp_mask, NULL,
2982 "vmalloc error: size %lu, exceeds total pages",
2987 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
2988 unsigned long size_per_node;
2991 * Try huge pages. Only try for PAGE_KERNEL allocations,
2992 * others like modules don't yet expect huge pages in
2993 * their allocations due to apply_to_page_range not
2997 size_per_node = size;
2998 if (node == NUMA_NO_NODE)
2999 size_per_node /= num_online_nodes();
3000 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3003 shift = arch_vmap_pte_supported_shift(size_per_node);
3005 align = max(real_align, 1UL << shift);
3006 size = ALIGN(real_size, 1UL << shift);
3010 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3011 VM_UNINITIALIZED | vm_flags, start, end, node,
3014 warn_alloc(gfp_mask, NULL,
3015 "vmalloc error: size %lu, vm_struct allocation failed",
3020 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3025 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3026 * flag. It means that vm_struct is not fully initialized.
3027 * Now, it is fully initialized, so remove this flag here.
3029 clear_vm_uninitialized_flag(area);
3031 size = PAGE_ALIGN(size);
3032 kmemleak_vmalloc(area, size, gfp_mask);
3037 if (shift > PAGE_SHIFT) {
3048 * __vmalloc_node - allocate virtually contiguous memory
3049 * @size: allocation size
3050 * @align: desired alignment
3051 * @gfp_mask: flags for the page level allocator
3052 * @node: node to use for allocation or NUMA_NO_NODE
3053 * @caller: caller's return address
3055 * Allocate enough pages to cover @size from the page level allocator with
3056 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3058 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3059 * and __GFP_NOFAIL are not supported
3061 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3064 * Return: pointer to the allocated memory or %NULL on error
3066 void *__vmalloc_node(unsigned long size, unsigned long align,
3067 gfp_t gfp_mask, int node, const void *caller)
3069 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3070 gfp_mask, PAGE_KERNEL, 0, node, caller);
3073 * This is only for performance analysis of vmalloc and stress purpose.
3074 * It is required by vmalloc test module, therefore do not use it other
3077 #ifdef CONFIG_TEST_VMALLOC_MODULE
3078 EXPORT_SYMBOL_GPL(__vmalloc_node);
3081 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3083 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3084 __builtin_return_address(0));
3086 EXPORT_SYMBOL(__vmalloc);
3089 * vmalloc - allocate virtually contiguous memory
3090 * @size: allocation size
3092 * Allocate enough pages to cover @size from the page level
3093 * allocator and map them into contiguous kernel virtual space.
3095 * For tight control over page level allocator and protection flags
3096 * use __vmalloc() instead.
3098 * Return: pointer to the allocated memory or %NULL on error
3100 void *vmalloc(unsigned long size)
3102 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3103 __builtin_return_address(0));
3105 EXPORT_SYMBOL(vmalloc);
3108 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3109 * @size: allocation size
3111 * Allocate enough non-huge pages to cover @size from the page level
3112 * allocator and map them into contiguous kernel virtual space.
3114 * Return: pointer to the allocated memory or %NULL on error
3116 void *vmalloc_no_huge(unsigned long size)
3118 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3119 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3120 NUMA_NO_NODE, __builtin_return_address(0));
3122 EXPORT_SYMBOL(vmalloc_no_huge);
3125 * vzalloc - allocate virtually contiguous memory with zero fill
3126 * @size: allocation size
3128 * Allocate enough pages to cover @size from the page level
3129 * allocator and map them into contiguous kernel virtual space.
3130 * The memory allocated is set to zero.
3132 * For tight control over page level allocator and protection flags
3133 * use __vmalloc() instead.
3135 * Return: pointer to the allocated memory or %NULL on error
3137 void *vzalloc(unsigned long size)
3139 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3140 __builtin_return_address(0));
3142 EXPORT_SYMBOL(vzalloc);
3145 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3146 * @size: allocation size
3148 * The resulting memory area is zeroed so it can be mapped to userspace
3149 * without leaking data.
3151 * Return: pointer to the allocated memory or %NULL on error
3153 void *vmalloc_user(unsigned long size)
3155 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3156 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3157 VM_USERMAP, NUMA_NO_NODE,
3158 __builtin_return_address(0));
3160 EXPORT_SYMBOL(vmalloc_user);
3163 * vmalloc_node - allocate memory on a specific node
3164 * @size: allocation size
3167 * Allocate enough pages to cover @size from the page level
3168 * allocator and map them into contiguous kernel virtual space.
3170 * For tight control over page level allocator and protection flags
3171 * use __vmalloc() instead.
3173 * Return: pointer to the allocated memory or %NULL on error
3175 void *vmalloc_node(unsigned long size, int node)
3177 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3178 __builtin_return_address(0));
3180 EXPORT_SYMBOL(vmalloc_node);
3183 * vzalloc_node - allocate memory on a specific node with zero fill
3184 * @size: allocation size
3187 * Allocate enough pages to cover @size from the page level
3188 * allocator and map them into contiguous kernel virtual space.
3189 * The memory allocated is set to zero.
3191 * Return: pointer to the allocated memory or %NULL on error
3193 void *vzalloc_node(unsigned long size, int node)
3195 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3196 __builtin_return_address(0));
3198 EXPORT_SYMBOL(vzalloc_node);
3200 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3201 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3202 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3203 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3206 * 64b systems should always have either DMA or DMA32 zones. For others
3207 * GFP_DMA32 should do the right thing and use the normal zone.
3209 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3213 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3214 * @size: allocation size
3216 * Allocate enough 32bit PA addressable pages to cover @size from the
3217 * page level allocator and map them into contiguous kernel virtual space.
3219 * Return: pointer to the allocated memory or %NULL on error
3221 void *vmalloc_32(unsigned long size)
3223 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3224 __builtin_return_address(0));
3226 EXPORT_SYMBOL(vmalloc_32);
3229 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3230 * @size: allocation size
3232 * The resulting memory area is 32bit addressable and zeroed so it can be
3233 * mapped to userspace without leaking data.
3235 * Return: pointer to the allocated memory or %NULL on error
3237 void *vmalloc_32_user(unsigned long size)
3239 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3240 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3241 VM_USERMAP, NUMA_NO_NODE,
3242 __builtin_return_address(0));
3244 EXPORT_SYMBOL(vmalloc_32_user);
3247 * small helper routine , copy contents to buf from addr.
3248 * If the page is not present, fill zero.
3251 static int aligned_vread(char *buf, char *addr, unsigned long count)
3257 unsigned long offset, length;
3259 offset = offset_in_page(addr);
3260 length = PAGE_SIZE - offset;
3263 p = vmalloc_to_page(addr);
3265 * To do safe access to this _mapped_ area, we need
3266 * lock. But adding lock here means that we need to add
3267 * overhead of vmalloc()/vfree() calls for this _debug_
3268 * interface, rarely used. Instead of that, we'll use
3269 * kmap() and get small overhead in this access function.
3272 /* We can expect USER0 is not used -- see vread() */
3273 void *map = kmap_atomic(p);
3274 memcpy(buf, map + offset, length);
3277 memset(buf, 0, length);
3288 * vread() - read vmalloc area in a safe way.
3289 * @buf: buffer for reading data
3290 * @addr: vm address.
3291 * @count: number of bytes to be read.
3293 * This function checks that addr is a valid vmalloc'ed area, and
3294 * copy data from that area to a given buffer. If the given memory range
3295 * of [addr...addr+count) includes some valid address, data is copied to
3296 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3297 * IOREMAP area is treated as memory hole and no copy is done.
3299 * If [addr...addr+count) doesn't includes any intersects with alive
3300 * vm_struct area, returns 0. @buf should be kernel's buffer.
3302 * Note: In usual ops, vread() is never necessary because the caller
3303 * should know vmalloc() area is valid and can use memcpy().
3304 * This is for routines which have to access vmalloc area without
3305 * any information, as /proc/kcore.
3307 * Return: number of bytes for which addr and buf should be increased
3308 * (same number as @count) or %0 if [addr...addr+count) doesn't
3309 * include any intersection with valid vmalloc area
3311 long vread(char *buf, char *addr, unsigned long count)
3313 struct vmap_area *va;
3314 struct vm_struct *vm;
3315 char *vaddr, *buf_start = buf;
3316 unsigned long buflen = count;
3319 /* Don't allow overflow */
3320 if ((unsigned long) addr + count < count)
3321 count = -(unsigned long) addr;
3323 spin_lock(&vmap_area_lock);
3324 va = find_vmap_area_exceed_addr((unsigned long)addr);
3328 /* no intersects with alive vmap_area */
3329 if ((unsigned long)addr + count <= va->va_start)
3332 list_for_each_entry_from(va, &vmap_area_list, list) {
3340 vaddr = (char *) vm->addr;
3341 if (addr >= vaddr + get_vm_area_size(vm))
3343 while (addr < vaddr) {
3351 n = vaddr + get_vm_area_size(vm) - addr;
3354 if (!(vm->flags & VM_IOREMAP))
3355 aligned_vread(buf, addr, n);
3356 else /* IOREMAP area is treated as memory hole */
3363 spin_unlock(&vmap_area_lock);
3365 if (buf == buf_start)
3367 /* zero-fill memory holes */
3368 if (buf != buf_start + buflen)
3369 memset(buf, 0, buflen - (buf - buf_start));
3375 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3376 * @vma: vma to cover
3377 * @uaddr: target user address to start at
3378 * @kaddr: virtual address of vmalloc kernel memory
3379 * @pgoff: offset from @kaddr to start at
3380 * @size: size of map area
3382 * Returns: 0 for success, -Exxx on failure
3384 * This function checks that @kaddr is a valid vmalloc'ed area,
3385 * and that it is big enough to cover the range starting at
3386 * @uaddr in @vma. Will return failure if that criteria isn't
3389 * Similar to remap_pfn_range() (see mm/memory.c)
3391 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3392 void *kaddr, unsigned long pgoff,
3395 struct vm_struct *area;
3397 unsigned long end_index;
3399 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3402 size = PAGE_ALIGN(size);
3404 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3407 area = find_vm_area(kaddr);
3411 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3414 if (check_add_overflow(size, off, &end_index) ||
3415 end_index > get_vm_area_size(area))
3420 struct page *page = vmalloc_to_page(kaddr);
3423 ret = vm_insert_page(vma, uaddr, page);
3432 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3438 * remap_vmalloc_range - map vmalloc pages to userspace
3439 * @vma: vma to cover (map full range of vma)
3440 * @addr: vmalloc memory
3441 * @pgoff: number of pages into addr before first page to map
3443 * Returns: 0 for success, -Exxx on failure
3445 * This function checks that addr is a valid vmalloc'ed area, and
3446 * that it is big enough to cover the vma. Will return failure if
3447 * that criteria isn't met.
3449 * Similar to remap_pfn_range() (see mm/memory.c)
3451 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3452 unsigned long pgoff)
3454 return remap_vmalloc_range_partial(vma, vma->vm_start,
3456 vma->vm_end - vma->vm_start);
3458 EXPORT_SYMBOL(remap_vmalloc_range);
3460 void free_vm_area(struct vm_struct *area)
3462 struct vm_struct *ret;
3463 ret = remove_vm_area(area->addr);
3464 BUG_ON(ret != area);
3467 EXPORT_SYMBOL_GPL(free_vm_area);
3470 static struct vmap_area *node_to_va(struct rb_node *n)
3472 return rb_entry_safe(n, struct vmap_area, rb_node);
3476 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3477 * @addr: target address
3479 * Returns: vmap_area if it is found. If there is no such area
3480 * the first highest(reverse order) vmap_area is returned
3481 * i.e. va->va_start < addr && va->va_end < addr or NULL
3482 * if there are no any areas before @addr.
3484 static struct vmap_area *
3485 pvm_find_va_enclose_addr(unsigned long addr)
3487 struct vmap_area *va, *tmp;
3490 n = free_vmap_area_root.rb_node;
3494 tmp = rb_entry(n, struct vmap_area, rb_node);
3495 if (tmp->va_start <= addr) {
3497 if (tmp->va_end >= addr)
3510 * pvm_determine_end_from_reverse - find the highest aligned address
3511 * of free block below VMALLOC_END
3513 * in - the VA we start the search(reverse order);
3514 * out - the VA with the highest aligned end address.
3515 * @align: alignment for required highest address
3517 * Returns: determined end address within vmap_area
3519 static unsigned long
3520 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3522 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3526 list_for_each_entry_from_reverse((*va),
3527 &free_vmap_area_list, list) {
3528 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3529 if ((*va)->va_start < addr)
3538 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3539 * @offsets: array containing offset of each area
3540 * @sizes: array containing size of each area
3541 * @nr_vms: the number of areas to allocate
3542 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3544 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3545 * vm_structs on success, %NULL on failure
3547 * Percpu allocator wants to use congruent vm areas so that it can
3548 * maintain the offsets among percpu areas. This function allocates
3549 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3550 * be scattered pretty far, distance between two areas easily going up
3551 * to gigabytes. To avoid interacting with regular vmallocs, these
3552 * areas are allocated from top.
3554 * Despite its complicated look, this allocator is rather simple. It
3555 * does everything top-down and scans free blocks from the end looking
3556 * for matching base. While scanning, if any of the areas do not fit the
3557 * base address is pulled down to fit the area. Scanning is repeated till
3558 * all the areas fit and then all necessary data structures are inserted
3559 * and the result is returned.
3561 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3562 const size_t *sizes, int nr_vms,
3565 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3566 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3567 struct vmap_area **vas, *va;
3568 struct vm_struct **vms;
3569 int area, area2, last_area, term_area;
3570 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3571 bool purged = false;
3574 /* verify parameters and allocate data structures */
3575 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3576 for (last_area = 0, area = 0; area < nr_vms; area++) {
3577 start = offsets[area];
3578 end = start + sizes[area];
3580 /* is everything aligned properly? */
3581 BUG_ON(!IS_ALIGNED(offsets[area], align));
3582 BUG_ON(!IS_ALIGNED(sizes[area], align));
3584 /* detect the area with the highest address */
3585 if (start > offsets[last_area])
3588 for (area2 = area + 1; area2 < nr_vms; area2++) {
3589 unsigned long start2 = offsets[area2];
3590 unsigned long end2 = start2 + sizes[area2];
3592 BUG_ON(start2 < end && start < end2);
3595 last_end = offsets[last_area] + sizes[last_area];
3597 if (vmalloc_end - vmalloc_start < last_end) {
3602 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3603 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3607 for (area = 0; area < nr_vms; area++) {
3608 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3609 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3610 if (!vas[area] || !vms[area])
3614 spin_lock(&free_vmap_area_lock);
3616 /* start scanning - we scan from the top, begin with the last area */
3617 area = term_area = last_area;
3618 start = offsets[area];
3619 end = start + sizes[area];
3621 va = pvm_find_va_enclose_addr(vmalloc_end);
3622 base = pvm_determine_end_from_reverse(&va, align) - end;
3626 * base might have underflowed, add last_end before
3629 if (base + last_end < vmalloc_start + last_end)
3633 * Fitting base has not been found.
3639 * If required width exceeds current VA block, move
3640 * base downwards and then recheck.
3642 if (base + end > va->va_end) {
3643 base = pvm_determine_end_from_reverse(&va, align) - end;
3649 * If this VA does not fit, move base downwards and recheck.
3651 if (base + start < va->va_start) {
3652 va = node_to_va(rb_prev(&va->rb_node));
3653 base = pvm_determine_end_from_reverse(&va, align) - end;
3659 * This area fits, move on to the previous one. If
3660 * the previous one is the terminal one, we're done.
3662 area = (area + nr_vms - 1) % nr_vms;
3663 if (area == term_area)
3666 start = offsets[area];
3667 end = start + sizes[area];
3668 va = pvm_find_va_enclose_addr(base + end);
3671 /* we've found a fitting base, insert all va's */
3672 for (area = 0; area < nr_vms; area++) {
3675 start = base + offsets[area];
3678 va = pvm_find_va_enclose_addr(start);
3679 if (WARN_ON_ONCE(va == NULL))
3680 /* It is a BUG(), but trigger recovery instead. */
3683 type = classify_va_fit_type(va, start, size);
3684 if (WARN_ON_ONCE(type == NOTHING_FIT))
3685 /* It is a BUG(), but trigger recovery instead. */
3688 ret = adjust_va_to_fit_type(va, start, size, type);
3692 /* Allocated area. */
3694 va->va_start = start;
3695 va->va_end = start + size;
3698 spin_unlock(&free_vmap_area_lock);
3700 /* populate the kasan shadow space */
3701 for (area = 0; area < nr_vms; area++) {
3702 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3703 goto err_free_shadow;
3705 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3709 /* insert all vm's */
3710 spin_lock(&vmap_area_lock);
3711 for (area = 0; area < nr_vms; area++) {
3712 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3714 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3717 spin_unlock(&vmap_area_lock);
3724 * Remove previously allocated areas. There is no
3725 * need in removing these areas from the busy tree,
3726 * because they are inserted only on the final step
3727 * and when pcpu_get_vm_areas() is success.
3730 orig_start = vas[area]->va_start;
3731 orig_end = vas[area]->va_end;
3732 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3733 &free_vmap_area_list);
3735 kasan_release_vmalloc(orig_start, orig_end,
3736 va->va_start, va->va_end);
3741 spin_unlock(&free_vmap_area_lock);
3743 purge_vmap_area_lazy();
3746 /* Before "retry", check if we recover. */
3747 for (area = 0; area < nr_vms; area++) {
3751 vas[area] = kmem_cache_zalloc(
3752 vmap_area_cachep, GFP_KERNEL);
3761 for (area = 0; area < nr_vms; area++) {
3763 kmem_cache_free(vmap_area_cachep, vas[area]);
3773 spin_lock(&free_vmap_area_lock);
3775 * We release all the vmalloc shadows, even the ones for regions that
3776 * hadn't been successfully added. This relies on kasan_release_vmalloc
3777 * being able to tolerate this case.
3779 for (area = 0; area < nr_vms; area++) {
3780 orig_start = vas[area]->va_start;
3781 orig_end = vas[area]->va_end;
3782 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3783 &free_vmap_area_list);
3785 kasan_release_vmalloc(orig_start, orig_end,
3786 va->va_start, va->va_end);
3790 spin_unlock(&free_vmap_area_lock);
3797 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3798 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3799 * @nr_vms: the number of allocated areas
3801 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3803 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3807 for (i = 0; i < nr_vms; i++)
3808 free_vm_area(vms[i]);
3811 #endif /* CONFIG_SMP */
3813 #ifdef CONFIG_PRINTK
3814 bool vmalloc_dump_obj(void *object)
3816 struct vm_struct *vm;
3817 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3819 vm = find_vm_area(objp);
3822 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3823 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3828 #ifdef CONFIG_PROC_FS
3829 static void *s_start(struct seq_file *m, loff_t *pos)
3830 __acquires(&vmap_purge_lock)
3831 __acquires(&vmap_area_lock)
3833 mutex_lock(&vmap_purge_lock);
3834 spin_lock(&vmap_area_lock);
3836 return seq_list_start(&vmap_area_list, *pos);
3839 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3841 return seq_list_next(p, &vmap_area_list, pos);
3844 static void s_stop(struct seq_file *m, void *p)
3845 __releases(&vmap_area_lock)
3846 __releases(&vmap_purge_lock)
3848 spin_unlock(&vmap_area_lock);
3849 mutex_unlock(&vmap_purge_lock);
3852 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3854 if (IS_ENABLED(CONFIG_NUMA)) {
3855 unsigned int nr, *counters = m->private;
3860 if (v->flags & VM_UNINITIALIZED)
3862 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3865 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3867 for (nr = 0; nr < v->nr_pages; nr++)
3868 counters[page_to_nid(v->pages[nr])]++;
3870 for_each_node_state(nr, N_HIGH_MEMORY)
3872 seq_printf(m, " N%u=%u", nr, counters[nr]);
3876 static void show_purge_info(struct seq_file *m)
3878 struct vmap_area *va;
3880 spin_lock(&purge_vmap_area_lock);
3881 list_for_each_entry(va, &purge_vmap_area_list, list) {
3882 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3883 (void *)va->va_start, (void *)va->va_end,
3884 va->va_end - va->va_start);
3886 spin_unlock(&purge_vmap_area_lock);
3889 static int s_show(struct seq_file *m, void *p)
3891 struct vmap_area *va;
3892 struct vm_struct *v;
3894 va = list_entry(p, struct vmap_area, list);
3897 * s_show can encounter race with remove_vm_area, !vm on behalf
3898 * of vmap area is being tear down or vm_map_ram allocation.
3901 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3902 (void *)va->va_start, (void *)va->va_end,
3903 va->va_end - va->va_start);
3910 seq_printf(m, "0x%pK-0x%pK %7ld",
3911 v->addr, v->addr + v->size, v->size);
3914 seq_printf(m, " %pS", v->caller);
3917 seq_printf(m, " pages=%d", v->nr_pages);
3920 seq_printf(m, " phys=%pa", &v->phys_addr);
3922 if (v->flags & VM_IOREMAP)
3923 seq_puts(m, " ioremap");
3925 if (v->flags & VM_ALLOC)
3926 seq_puts(m, " vmalloc");
3928 if (v->flags & VM_MAP)
3929 seq_puts(m, " vmap");
3931 if (v->flags & VM_USERMAP)
3932 seq_puts(m, " user");
3934 if (v->flags & VM_DMA_COHERENT)
3935 seq_puts(m, " dma-coherent");
3937 if (is_vmalloc_addr(v->pages))
3938 seq_puts(m, " vpages");
3940 show_numa_info(m, v);
3944 * As a final step, dump "unpurged" areas.
3946 if (list_is_last(&va->list, &vmap_area_list))
3952 static const struct seq_operations vmalloc_op = {
3959 static int __init proc_vmalloc_init(void)
3961 if (IS_ENABLED(CONFIG_NUMA))
3962 proc_create_seq_private("vmallocinfo", 0400, NULL,
3964 nr_node_ids * sizeof(unsigned int), NULL);
3966 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3969 module_init(proc_vmalloc_init);