1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/vmalloc.h>
50 #include "pgalloc-track.h"
52 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53 static unsigned int __ro_after_init ioremap_max_page_shift
= BITS_PER_LONG
- 1;
55 static int __init
set_nohugeiomap(char *str
)
57 ioremap_max_page_shift
= PAGE_SHIFT
;
60 early_param("nohugeiomap", set_nohugeiomap
);
61 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 static const unsigned int ioremap_max_page_shift
= PAGE_SHIFT
;
63 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66 static bool __ro_after_init vmap_allow_huge
= true;
68 static int __init
set_nohugevmalloc(char *str
)
70 vmap_allow_huge
= false;
73 early_param("nohugevmalloc", set_nohugevmalloc
);
74 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 static const bool vmap_allow_huge
= false;
76 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78 bool is_vmalloc_addr(const void *x
)
80 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
82 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
84 EXPORT_SYMBOL(is_vmalloc_addr
);
86 struct vfree_deferred
{
87 struct llist_head list
;
88 struct work_struct wq
;
90 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
92 /*** Page table manipulation functions ***/
93 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
94 phys_addr_t phys_addr
, pgprot_t prot
,
95 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
99 unsigned long size
= PAGE_SIZE
;
101 pfn
= phys_addr
>> PAGE_SHIFT
;
102 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
106 BUG_ON(!pte_none(*pte
));
108 #ifdef CONFIG_HUGETLB_PAGE
109 size
= arch_vmap_pte_range_map_size(addr
, end
, pfn
, max_page_shift
);
110 if (size
!= PAGE_SIZE
) {
111 pte_t entry
= pfn_pte(pfn
, prot
);
113 entry
= arch_make_huge_pte(entry
, ilog2(size
), 0);
114 set_huge_pte_at(&init_mm
, addr
, pte
, entry
);
115 pfn
+= PFN_DOWN(size
);
119 set_pte_at(&init_mm
, addr
, pte
, pfn_pte(pfn
, prot
));
121 } while (pte
+= PFN_DOWN(size
), addr
+= size
, addr
!= end
);
122 *mask
|= PGTBL_PTE_MODIFIED
;
126 static int vmap_try_huge_pmd(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
127 phys_addr_t phys_addr
, pgprot_t prot
,
128 unsigned int max_page_shift
)
130 if (max_page_shift
< PMD_SHIFT
)
133 if (!arch_vmap_pmd_supported(prot
))
136 if ((end
- addr
) != PMD_SIZE
)
139 if (!IS_ALIGNED(addr
, PMD_SIZE
))
142 if (!IS_ALIGNED(phys_addr
, PMD_SIZE
))
145 if (pmd_present(*pmd
) && !pmd_free_pte_page(pmd
, addr
))
148 return pmd_set_huge(pmd
, phys_addr
, prot
);
151 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
152 phys_addr_t phys_addr
, pgprot_t prot
,
153 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
158 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
162 next
= pmd_addr_end(addr
, end
);
164 if (vmap_try_huge_pmd(pmd
, addr
, next
, phys_addr
, prot
,
166 *mask
|= PGTBL_PMD_MODIFIED
;
170 if (vmap_pte_range(pmd
, addr
, next
, phys_addr
, prot
, max_page_shift
, mask
))
172 } while (pmd
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
176 static int vmap_try_huge_pud(pud_t
*pud
, unsigned long addr
, unsigned long end
,
177 phys_addr_t phys_addr
, pgprot_t prot
,
178 unsigned int max_page_shift
)
180 if (max_page_shift
< PUD_SHIFT
)
183 if (!arch_vmap_pud_supported(prot
))
186 if ((end
- addr
) != PUD_SIZE
)
189 if (!IS_ALIGNED(addr
, PUD_SIZE
))
192 if (!IS_ALIGNED(phys_addr
, PUD_SIZE
))
195 if (pud_present(*pud
) && !pud_free_pmd_page(pud
, addr
))
198 return pud_set_huge(pud
, phys_addr
, prot
);
201 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
202 phys_addr_t phys_addr
, pgprot_t prot
,
203 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
208 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
212 next
= pud_addr_end(addr
, end
);
214 if (vmap_try_huge_pud(pud
, addr
, next
, phys_addr
, prot
,
216 *mask
|= PGTBL_PUD_MODIFIED
;
220 if (vmap_pmd_range(pud
, addr
, next
, phys_addr
, prot
,
221 max_page_shift
, mask
))
223 } while (pud
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
227 static int vmap_try_huge_p4d(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
228 phys_addr_t phys_addr
, pgprot_t prot
,
229 unsigned int max_page_shift
)
231 if (max_page_shift
< P4D_SHIFT
)
234 if (!arch_vmap_p4d_supported(prot
))
237 if ((end
- addr
) != P4D_SIZE
)
240 if (!IS_ALIGNED(addr
, P4D_SIZE
))
243 if (!IS_ALIGNED(phys_addr
, P4D_SIZE
))
246 if (p4d_present(*p4d
) && !p4d_free_pud_page(p4d
, addr
))
249 return p4d_set_huge(p4d
, phys_addr
, prot
);
252 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
253 phys_addr_t phys_addr
, pgprot_t prot
,
254 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
259 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
263 next
= p4d_addr_end(addr
, end
);
265 if (vmap_try_huge_p4d(p4d
, addr
, next
, phys_addr
, prot
,
267 *mask
|= PGTBL_P4D_MODIFIED
;
271 if (vmap_pud_range(p4d
, addr
, next
, phys_addr
, prot
,
272 max_page_shift
, mask
))
274 } while (p4d
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
278 static int vmap_range_noflush(unsigned long addr
, unsigned long end
,
279 phys_addr_t phys_addr
, pgprot_t prot
,
280 unsigned int max_page_shift
)
286 pgtbl_mod_mask mask
= 0;
292 pgd
= pgd_offset_k(addr
);
294 next
= pgd_addr_end(addr
, end
);
295 err
= vmap_p4d_range(pgd
, addr
, next
, phys_addr
, prot
,
296 max_page_shift
, &mask
);
299 } while (pgd
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
301 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
302 arch_sync_kernel_mappings(start
, end
);
307 int ioremap_page_range(unsigned long addr
, unsigned long end
,
308 phys_addr_t phys_addr
, pgprot_t prot
)
312 err
= vmap_range_noflush(addr
, end
, phys_addr
, pgprot_nx(prot
),
313 ioremap_max_page_shift
);
314 flush_cache_vmap(addr
, end
);
316 err
= kmsan_ioremap_page_range(addr
, end
, phys_addr
, prot
,
317 ioremap_max_page_shift
);
321 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
322 pgtbl_mod_mask
*mask
)
326 pte
= pte_offset_kernel(pmd
, addr
);
328 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
329 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
330 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
331 *mask
|= PGTBL_PTE_MODIFIED
;
334 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
335 pgtbl_mod_mask
*mask
)
341 pmd
= pmd_offset(pud
, addr
);
343 next
= pmd_addr_end(addr
, end
);
345 cleared
= pmd_clear_huge(pmd
);
346 if (cleared
|| pmd_bad(*pmd
))
347 *mask
|= PGTBL_PMD_MODIFIED
;
351 if (pmd_none_or_clear_bad(pmd
))
353 vunmap_pte_range(pmd
, addr
, next
, mask
);
356 } while (pmd
++, addr
= next
, addr
!= end
);
359 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
360 pgtbl_mod_mask
*mask
)
366 pud
= pud_offset(p4d
, addr
);
368 next
= pud_addr_end(addr
, end
);
370 cleared
= pud_clear_huge(pud
);
371 if (cleared
|| pud_bad(*pud
))
372 *mask
|= PGTBL_PUD_MODIFIED
;
376 if (pud_none_or_clear_bad(pud
))
378 vunmap_pmd_range(pud
, addr
, next
, mask
);
379 } while (pud
++, addr
= next
, addr
!= end
);
382 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
383 pgtbl_mod_mask
*mask
)
388 p4d
= p4d_offset(pgd
, addr
);
390 next
= p4d_addr_end(addr
, end
);
394 *mask
|= PGTBL_P4D_MODIFIED
;
396 if (p4d_none_or_clear_bad(p4d
))
398 vunmap_pud_range(p4d
, addr
, next
, mask
);
399 } while (p4d
++, addr
= next
, addr
!= end
);
403 * vunmap_range_noflush is similar to vunmap_range, but does not
404 * flush caches or TLBs.
406 * The caller is responsible for calling flush_cache_vmap() before calling
407 * this function, and flush_tlb_kernel_range after it has returned
408 * successfully (and before the addresses are expected to cause a page fault
409 * or be re-mapped for something else, if TLB flushes are being delayed or
412 * This is an internal function only. Do not use outside mm/.
414 void __vunmap_range_noflush(unsigned long start
, unsigned long end
)
418 unsigned long addr
= start
;
419 pgtbl_mod_mask mask
= 0;
422 pgd
= pgd_offset_k(addr
);
424 next
= pgd_addr_end(addr
, end
);
426 mask
|= PGTBL_PGD_MODIFIED
;
427 if (pgd_none_or_clear_bad(pgd
))
429 vunmap_p4d_range(pgd
, addr
, next
, &mask
);
430 } while (pgd
++, addr
= next
, addr
!= end
);
432 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
433 arch_sync_kernel_mappings(start
, end
);
436 void vunmap_range_noflush(unsigned long start
, unsigned long end
)
438 kmsan_vunmap_range_noflush(start
, end
);
439 __vunmap_range_noflush(start
, end
);
443 * vunmap_range - unmap kernel virtual addresses
444 * @addr: start of the VM area to unmap
445 * @end: end of the VM area to unmap (non-inclusive)
447 * Clears any present PTEs in the virtual address range, flushes TLBs and
448 * caches. Any subsequent access to the address before it has been re-mapped
451 void vunmap_range(unsigned long addr
, unsigned long end
)
453 flush_cache_vunmap(addr
, end
);
454 vunmap_range_noflush(addr
, end
);
455 flush_tlb_kernel_range(addr
, end
);
458 static int vmap_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
459 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
460 pgtbl_mod_mask
*mask
)
465 * nr is a running index into the array which helps higher level
466 * callers keep track of where we're up to.
469 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
473 struct page
*page
= pages
[*nr
];
475 if (WARN_ON(!pte_none(*pte
)))
479 if (WARN_ON(!pfn_valid(page_to_pfn(page
))))
482 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
484 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
485 *mask
|= PGTBL_PTE_MODIFIED
;
489 static int vmap_pages_pmd_range(pud_t
*pud
, unsigned long addr
,
490 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
491 pgtbl_mod_mask
*mask
)
496 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
500 next
= pmd_addr_end(addr
, end
);
501 if (vmap_pages_pte_range(pmd
, addr
, next
, prot
, pages
, nr
, mask
))
503 } while (pmd
++, addr
= next
, addr
!= end
);
507 static int vmap_pages_pud_range(p4d_t
*p4d
, unsigned long addr
,
508 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
509 pgtbl_mod_mask
*mask
)
514 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
518 next
= pud_addr_end(addr
, end
);
519 if (vmap_pages_pmd_range(pud
, addr
, next
, prot
, pages
, nr
, mask
))
521 } while (pud
++, addr
= next
, addr
!= end
);
525 static int vmap_pages_p4d_range(pgd_t
*pgd
, unsigned long addr
,
526 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
527 pgtbl_mod_mask
*mask
)
532 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
536 next
= p4d_addr_end(addr
, end
);
537 if (vmap_pages_pud_range(p4d
, addr
, next
, prot
, pages
, nr
, mask
))
539 } while (p4d
++, addr
= next
, addr
!= end
);
543 static int vmap_small_pages_range_noflush(unsigned long addr
, unsigned long end
,
544 pgprot_t prot
, struct page
**pages
)
546 unsigned long start
= addr
;
551 pgtbl_mod_mask mask
= 0;
554 pgd
= pgd_offset_k(addr
);
556 next
= pgd_addr_end(addr
, end
);
558 mask
|= PGTBL_PGD_MODIFIED
;
559 err
= vmap_pages_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
, &mask
);
562 } while (pgd
++, addr
= next
, addr
!= end
);
564 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
565 arch_sync_kernel_mappings(start
, end
);
571 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
574 * The caller is responsible for calling flush_cache_vmap() after this
575 * function returns successfully and before the addresses are accessed.
577 * This is an internal function only. Do not use outside mm/.
579 int __vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
580 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
582 unsigned int i
, nr
= (end
- addr
) >> PAGE_SHIFT
;
584 WARN_ON(page_shift
< PAGE_SHIFT
);
586 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC
) ||
587 page_shift
== PAGE_SHIFT
)
588 return vmap_small_pages_range_noflush(addr
, end
, prot
, pages
);
590 for (i
= 0; i
< nr
; i
+= 1U << (page_shift
- PAGE_SHIFT
)) {
593 err
= vmap_range_noflush(addr
, addr
+ (1UL << page_shift
),
594 page_to_phys(pages
[i
]), prot
,
599 addr
+= 1UL << page_shift
;
605 int vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
606 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
608 int ret
= kmsan_vmap_pages_range_noflush(addr
, end
, prot
, pages
,
613 return __vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
617 * vmap_pages_range - map pages to a kernel virtual address
618 * @addr: start of the VM area to map
619 * @end: end of the VM area to map (non-inclusive)
620 * @prot: page protection flags to use
621 * @pages: pages to map (always PAGE_SIZE pages)
622 * @page_shift: maximum shift that the pages may be mapped with, @pages must
623 * be aligned and contiguous up to at least this shift.
626 * 0 on success, -errno on failure.
628 static int vmap_pages_range(unsigned long addr
, unsigned long end
,
629 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
633 err
= vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
634 flush_cache_vmap(addr
, end
);
638 int is_vmalloc_or_module_addr(const void *x
)
641 * ARM, x86-64 and sparc64 put modules in a special place,
642 * and fall back on vmalloc() if that fails. Others
643 * just put it in the vmalloc space.
645 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
646 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
647 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
650 return is_vmalloc_addr(x
);
652 EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr
);
655 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
656 * return the tail page that corresponds to the base page address, which
657 * matches small vmap mappings.
659 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
661 unsigned long addr
= (unsigned long) vmalloc_addr
;
662 struct page
*page
= NULL
;
663 pgd_t
*pgd
= pgd_offset_k(addr
);
670 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
671 * architectures that do not vmalloc module space
673 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
677 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
678 return NULL
; /* XXX: no allowance for huge pgd */
679 if (WARN_ON_ONCE(pgd_bad(*pgd
)))
682 p4d
= p4d_offset(pgd
, addr
);
686 return p4d_page(*p4d
) + ((addr
& ~P4D_MASK
) >> PAGE_SHIFT
);
687 if (WARN_ON_ONCE(p4d_bad(*p4d
)))
690 pud
= pud_offset(p4d
, addr
);
694 return pud_page(*pud
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
695 if (WARN_ON_ONCE(pud_bad(*pud
)))
698 pmd
= pmd_offset(pud
, addr
);
702 return pmd_page(*pmd
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
703 if (WARN_ON_ONCE(pmd_bad(*pmd
)))
706 ptep
= pte_offset_map(pmd
, addr
);
708 if (pte_present(pte
))
709 page
= pte_page(pte
);
714 EXPORT_SYMBOL(vmalloc_to_page
);
717 * Map a vmalloc()-space virtual address to the physical page frame number.
719 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
721 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
723 EXPORT_SYMBOL(vmalloc_to_pfn
);
726 /*** Global kva allocator ***/
728 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
729 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
732 static DEFINE_SPINLOCK(vmap_area_lock
);
733 static DEFINE_SPINLOCK(free_vmap_area_lock
);
734 /* Export for kexec only */
735 LIST_HEAD(vmap_area_list
);
736 static struct rb_root vmap_area_root
= RB_ROOT
;
737 static bool vmap_initialized __read_mostly
;
739 static struct rb_root purge_vmap_area_root
= RB_ROOT
;
740 static LIST_HEAD(purge_vmap_area_list
);
741 static DEFINE_SPINLOCK(purge_vmap_area_lock
);
744 * This kmem_cache is used for vmap_area objects. Instead of
745 * allocating from slab we reuse an object from this cache to
746 * make things faster. Especially in "no edge" splitting of
749 static struct kmem_cache
*vmap_area_cachep
;
752 * This linked list is used in pair with free_vmap_area_root.
753 * It gives O(1) access to prev/next to perform fast coalescing.
755 static LIST_HEAD(free_vmap_area_list
);
758 * This augment red-black tree represents the free vmap space.
759 * All vmap_area objects in this tree are sorted by va->va_start
760 * address. It is used for allocation and merging when a vmap
761 * object is released.
763 * Each vmap_area node contains a maximum available free block
764 * of its sub-tree, right or left. Therefore it is possible to
765 * find a lowest match of free area.
767 static struct rb_root free_vmap_area_root
= RB_ROOT
;
770 * Preload a CPU with one object for "no edge" split case. The
771 * aim is to get rid of allocations from the atomic context, thus
772 * to use more permissive allocation masks.
774 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
776 static __always_inline
unsigned long
777 va_size(struct vmap_area
*va
)
779 return (va
->va_end
- va
->va_start
);
782 static __always_inline
unsigned long
783 get_subtree_max_size(struct rb_node
*node
)
785 struct vmap_area
*va
;
787 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
788 return va
? va
->subtree_max_size
: 0;
791 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
792 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
794 static void purge_vmap_area_lazy(void);
795 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
796 static void drain_vmap_area_work(struct work_struct
*work
);
797 static DECLARE_WORK(drain_vmap_work
, drain_vmap_area_work
);
799 static atomic_long_t nr_vmalloc_pages
;
801 unsigned long vmalloc_nr_pages(void)
803 return atomic_long_read(&nr_vmalloc_pages
);
806 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
807 static struct vmap_area
*find_vmap_area_exceed_addr(unsigned long addr
)
809 struct vmap_area
*va
= NULL
;
810 struct rb_node
*n
= vmap_area_root
.rb_node
;
812 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
815 struct vmap_area
*tmp
;
817 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
818 if (tmp
->va_end
> addr
) {
820 if (tmp
->va_start
<= addr
)
831 static struct vmap_area
*__find_vmap_area(unsigned long addr
, struct rb_root
*root
)
833 struct rb_node
*n
= root
->rb_node
;
835 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
838 struct vmap_area
*va
;
840 va
= rb_entry(n
, struct vmap_area
, rb_node
);
841 if (addr
< va
->va_start
)
843 else if (addr
>= va
->va_end
)
853 * This function returns back addresses of parent node
854 * and its left or right link for further processing.
856 * Otherwise NULL is returned. In that case all further
857 * steps regarding inserting of conflicting overlap range
858 * have to be declined and actually considered as a bug.
860 static __always_inline
struct rb_node
**
861 find_va_links(struct vmap_area
*va
,
862 struct rb_root
*root
, struct rb_node
*from
,
863 struct rb_node
**parent
)
865 struct vmap_area
*tmp_va
;
866 struct rb_node
**link
;
869 link
= &root
->rb_node
;
870 if (unlikely(!*link
)) {
879 * Go to the bottom of the tree. When we hit the last point
880 * we end up with parent rb_node and correct direction, i name
881 * it link, where the new va->rb_node will be attached to.
884 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
887 * During the traversal we also do some sanity check.
888 * Trigger the BUG() if there are sides(left/right)
891 if (va
->va_end
<= tmp_va
->va_start
)
892 link
= &(*link
)->rb_left
;
893 else if (va
->va_start
>= tmp_va
->va_end
)
894 link
= &(*link
)->rb_right
;
896 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
897 va
->va_start
, va
->va_end
, tmp_va
->va_start
, tmp_va
->va_end
);
903 *parent
= &tmp_va
->rb_node
;
907 static __always_inline
struct list_head
*
908 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
910 struct list_head
*list
;
912 if (unlikely(!parent
))
914 * The red-black tree where we try to find VA neighbors
915 * before merging or inserting is empty, i.e. it means
916 * there is no free vmap space. Normally it does not
917 * happen but we handle this case anyway.
921 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
922 return (&parent
->rb_right
== link
? list
->next
: list
);
925 static __always_inline
void
926 __link_va(struct vmap_area
*va
, struct rb_root
*root
,
927 struct rb_node
*parent
, struct rb_node
**link
,
928 struct list_head
*head
, bool augment
)
931 * VA is still not in the list, but we can
932 * identify its future previous list_head node.
934 if (likely(parent
)) {
935 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
936 if (&parent
->rb_right
!= link
)
940 /* Insert to the rb-tree */
941 rb_link_node(&va
->rb_node
, parent
, link
);
944 * Some explanation here. Just perform simple insertion
945 * to the tree. We do not set va->subtree_max_size to
946 * its current size before calling rb_insert_augmented().
947 * It is because we populate the tree from the bottom
948 * to parent levels when the node _is_ in the tree.
950 * Therefore we set subtree_max_size to zero after insertion,
951 * to let __augment_tree_propagate_from() puts everything to
952 * the correct order later on.
954 rb_insert_augmented(&va
->rb_node
,
955 root
, &free_vmap_area_rb_augment_cb
);
956 va
->subtree_max_size
= 0;
958 rb_insert_color(&va
->rb_node
, root
);
961 /* Address-sort this list */
962 list_add(&va
->list
, head
);
965 static __always_inline
void
966 link_va(struct vmap_area
*va
, struct rb_root
*root
,
967 struct rb_node
*parent
, struct rb_node
**link
,
968 struct list_head
*head
)
970 __link_va(va
, root
, parent
, link
, head
, false);
973 static __always_inline
void
974 link_va_augment(struct vmap_area
*va
, struct rb_root
*root
,
975 struct rb_node
*parent
, struct rb_node
**link
,
976 struct list_head
*head
)
978 __link_va(va
, root
, parent
, link
, head
, true);
981 static __always_inline
void
982 __unlink_va(struct vmap_area
*va
, struct rb_root
*root
, bool augment
)
984 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
988 rb_erase_augmented(&va
->rb_node
,
989 root
, &free_vmap_area_rb_augment_cb
);
991 rb_erase(&va
->rb_node
, root
);
993 list_del_init(&va
->list
);
994 RB_CLEAR_NODE(&va
->rb_node
);
997 static __always_inline
void
998 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
1000 __unlink_va(va
, root
, false);
1003 static __always_inline
void
1004 unlink_va_augment(struct vmap_area
*va
, struct rb_root
*root
)
1006 __unlink_va(va
, root
, true);
1009 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1011 * Gets called when remove the node and rotate.
1013 static __always_inline
unsigned long
1014 compute_subtree_max_size(struct vmap_area
*va
)
1016 return max3(va_size(va
),
1017 get_subtree_max_size(va
->rb_node
.rb_left
),
1018 get_subtree_max_size(va
->rb_node
.rb_right
));
1022 augment_tree_propagate_check(void)
1024 struct vmap_area
*va
;
1025 unsigned long computed_size
;
1027 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
1028 computed_size
= compute_subtree_max_size(va
);
1029 if (computed_size
!= va
->subtree_max_size
)
1030 pr_emerg("tree is corrupted: %lu, %lu\n",
1031 va_size(va
), va
->subtree_max_size
);
1037 * This function populates subtree_max_size from bottom to upper
1038 * levels starting from VA point. The propagation must be done
1039 * when VA size is modified by changing its va_start/va_end. Or
1040 * in case of newly inserting of VA to the tree.
1042 * It means that __augment_tree_propagate_from() must be called:
1043 * - After VA has been inserted to the tree(free path);
1044 * - After VA has been shrunk(allocation path);
1045 * - After VA has been increased(merging path).
1047 * Please note that, it does not mean that upper parent nodes
1048 * and their subtree_max_size are recalculated all the time up
1057 * For example if we modify the node 4, shrinking it to 2, then
1058 * no any modification is required. If we shrink the node 2 to 1
1059 * its subtree_max_size is updated only, and set to 1. If we shrink
1060 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1061 * node becomes 4--6.
1063 static __always_inline
void
1064 augment_tree_propagate_from(struct vmap_area
*va
)
1067 * Populate the tree from bottom towards the root until
1068 * the calculated maximum available size of checked node
1069 * is equal to its current one.
1071 free_vmap_area_rb_augment_cb_propagate(&va
->rb_node
, NULL
);
1073 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1074 augment_tree_propagate_check();
1079 insert_vmap_area(struct vmap_area
*va
,
1080 struct rb_root
*root
, struct list_head
*head
)
1082 struct rb_node
**link
;
1083 struct rb_node
*parent
;
1085 link
= find_va_links(va
, root
, NULL
, &parent
);
1087 link_va(va
, root
, parent
, link
, head
);
1091 insert_vmap_area_augment(struct vmap_area
*va
,
1092 struct rb_node
*from
, struct rb_root
*root
,
1093 struct list_head
*head
)
1095 struct rb_node
**link
;
1096 struct rb_node
*parent
;
1099 link
= find_va_links(va
, NULL
, from
, &parent
);
1101 link
= find_va_links(va
, root
, NULL
, &parent
);
1104 link_va_augment(va
, root
, parent
, link
, head
);
1105 augment_tree_propagate_from(va
);
1110 * Merge de-allocated chunk of VA memory with previous
1111 * and next free blocks. If coalesce is not done a new
1112 * free area is inserted. If VA has been merged, it is
1115 * Please note, it can return NULL in case of overlap
1116 * ranges, followed by WARN() report. Despite it is a
1117 * buggy behaviour, a system can be alive and keep
1120 static __always_inline
struct vmap_area
*
1121 __merge_or_add_vmap_area(struct vmap_area
*va
,
1122 struct rb_root
*root
, struct list_head
*head
, bool augment
)
1124 struct vmap_area
*sibling
;
1125 struct list_head
*next
;
1126 struct rb_node
**link
;
1127 struct rb_node
*parent
;
1128 bool merged
= false;
1131 * Find a place in the tree where VA potentially will be
1132 * inserted, unless it is merged with its sibling/siblings.
1134 link
= find_va_links(va
, root
, NULL
, &parent
);
1139 * Get next node of VA to check if merging can be done.
1141 next
= get_va_next_sibling(parent
, link
);
1142 if (unlikely(next
== NULL
))
1148 * |<------VA------>|<-----Next----->|
1153 sibling
= list_entry(next
, struct vmap_area
, list
);
1154 if (sibling
->va_start
== va
->va_end
) {
1155 sibling
->va_start
= va
->va_start
;
1157 /* Free vmap_area object. */
1158 kmem_cache_free(vmap_area_cachep
, va
);
1160 /* Point to the new merged area. */
1169 * |<-----Prev----->|<------VA------>|
1173 if (next
->prev
!= head
) {
1174 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
1175 if (sibling
->va_end
== va
->va_start
) {
1177 * If both neighbors are coalesced, it is important
1178 * to unlink the "next" node first, followed by merging
1179 * with "previous" one. Otherwise the tree might not be
1180 * fully populated if a sibling's augmented value is
1181 * "normalized" because of rotation operations.
1184 __unlink_va(va
, root
, augment
);
1186 sibling
->va_end
= va
->va_end
;
1188 /* Free vmap_area object. */
1189 kmem_cache_free(vmap_area_cachep
, va
);
1191 /* Point to the new merged area. */
1199 __link_va(va
, root
, parent
, link
, head
, augment
);
1204 static __always_inline
struct vmap_area
*
1205 merge_or_add_vmap_area(struct vmap_area
*va
,
1206 struct rb_root
*root
, struct list_head
*head
)
1208 return __merge_or_add_vmap_area(va
, root
, head
, false);
1211 static __always_inline
struct vmap_area
*
1212 merge_or_add_vmap_area_augment(struct vmap_area
*va
,
1213 struct rb_root
*root
, struct list_head
*head
)
1215 va
= __merge_or_add_vmap_area(va
, root
, head
, true);
1217 augment_tree_propagate_from(va
);
1222 static __always_inline
bool
1223 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
1224 unsigned long align
, unsigned long vstart
)
1226 unsigned long nva_start_addr
;
1228 if (va
->va_start
> vstart
)
1229 nva_start_addr
= ALIGN(va
->va_start
, align
);
1231 nva_start_addr
= ALIGN(vstart
, align
);
1233 /* Can be overflowed due to big size or alignment. */
1234 if (nva_start_addr
+ size
< nva_start_addr
||
1235 nva_start_addr
< vstart
)
1238 return (nva_start_addr
+ size
<= va
->va_end
);
1242 * Find the first free block(lowest start address) in the tree,
1243 * that will accomplish the request corresponding to passing
1244 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1245 * a search length is adjusted to account for worst case alignment
1248 static __always_inline
struct vmap_area
*
1249 find_vmap_lowest_match(struct rb_root
*root
, unsigned long size
,
1250 unsigned long align
, unsigned long vstart
, bool adjust_search_size
)
1252 struct vmap_area
*va
;
1253 struct rb_node
*node
;
1254 unsigned long length
;
1256 /* Start from the root. */
1257 node
= root
->rb_node
;
1259 /* Adjust the search size for alignment overhead. */
1260 length
= adjust_search_size
? size
+ align
- 1 : size
;
1263 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1265 if (get_subtree_max_size(node
->rb_left
) >= length
&&
1266 vstart
< va
->va_start
) {
1267 node
= node
->rb_left
;
1269 if (is_within_this_va(va
, size
, align
, vstart
))
1273 * Does not make sense to go deeper towards the right
1274 * sub-tree if it does not have a free block that is
1275 * equal or bigger to the requested search length.
1277 if (get_subtree_max_size(node
->rb_right
) >= length
) {
1278 node
= node
->rb_right
;
1283 * OK. We roll back and find the first right sub-tree,
1284 * that will satisfy the search criteria. It can happen
1285 * due to "vstart" restriction or an alignment overhead
1286 * that is bigger then PAGE_SIZE.
1288 while ((node
= rb_parent(node
))) {
1289 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1290 if (is_within_this_va(va
, size
, align
, vstart
))
1293 if (get_subtree_max_size(node
->rb_right
) >= length
&&
1294 vstart
<= va
->va_start
) {
1296 * Shift the vstart forward. Please note, we update it with
1297 * parent's start address adding "1" because we do not want
1298 * to enter same sub-tree after it has already been checked
1299 * and no suitable free block found there.
1301 vstart
= va
->va_start
+ 1;
1302 node
= node
->rb_right
;
1312 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1313 #include <linux/random.h>
1315 static struct vmap_area
*
1316 find_vmap_lowest_linear_match(struct list_head
*head
, unsigned long size
,
1317 unsigned long align
, unsigned long vstart
)
1319 struct vmap_area
*va
;
1321 list_for_each_entry(va
, head
, list
) {
1322 if (!is_within_this_va(va
, size
, align
, vstart
))
1332 find_vmap_lowest_match_check(struct rb_root
*root
, struct list_head
*head
,
1333 unsigned long size
, unsigned long align
)
1335 struct vmap_area
*va_1
, *va_2
;
1336 unsigned long vstart
;
1339 get_random_bytes(&rnd
, sizeof(rnd
));
1340 vstart
= VMALLOC_START
+ rnd
;
1342 va_1
= find_vmap_lowest_match(root
, size
, align
, vstart
, false);
1343 va_2
= find_vmap_lowest_linear_match(head
, size
, align
, vstart
);
1346 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1347 va_1
, va_2
, vstart
);
1353 FL_FIT_TYPE
= 1, /* full fit */
1354 LE_FIT_TYPE
= 2, /* left edge fit */
1355 RE_FIT_TYPE
= 3, /* right edge fit */
1356 NE_FIT_TYPE
= 4 /* no edge fit */
1359 static __always_inline
enum fit_type
1360 classify_va_fit_type(struct vmap_area
*va
,
1361 unsigned long nva_start_addr
, unsigned long size
)
1365 /* Check if it is within VA. */
1366 if (nva_start_addr
< va
->va_start
||
1367 nva_start_addr
+ size
> va
->va_end
)
1371 if (va
->va_start
== nva_start_addr
) {
1372 if (va
->va_end
== nva_start_addr
+ size
)
1376 } else if (va
->va_end
== nva_start_addr
+ size
) {
1385 static __always_inline
int
1386 adjust_va_to_fit_type(struct rb_root
*root
, struct list_head
*head
,
1387 struct vmap_area
*va
, unsigned long nva_start_addr
,
1390 struct vmap_area
*lva
= NULL
;
1391 enum fit_type type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1393 if (type
== FL_FIT_TYPE
) {
1395 * No need to split VA, it fully fits.
1401 unlink_va_augment(va
, root
);
1402 kmem_cache_free(vmap_area_cachep
, va
);
1403 } else if (type
== LE_FIT_TYPE
) {
1405 * Split left edge of fit VA.
1411 va
->va_start
+= size
;
1412 } else if (type
== RE_FIT_TYPE
) {
1414 * Split right edge of fit VA.
1420 va
->va_end
= nva_start_addr
;
1421 } else if (type
== NE_FIT_TYPE
) {
1423 * Split no edge of fit VA.
1429 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1430 if (unlikely(!lva
)) {
1432 * For percpu allocator we do not do any pre-allocation
1433 * and leave it as it is. The reason is it most likely
1434 * never ends up with NE_FIT_TYPE splitting. In case of
1435 * percpu allocations offsets and sizes are aligned to
1436 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1437 * are its main fitting cases.
1439 * There are a few exceptions though, as an example it is
1440 * a first allocation (early boot up) when we have "one"
1441 * big free space that has to be split.
1443 * Also we can hit this path in case of regular "vmap"
1444 * allocations, if "this" current CPU was not preloaded.
1445 * See the comment in alloc_vmap_area() why. If so, then
1446 * GFP_NOWAIT is used instead to get an extra object for
1447 * split purpose. That is rare and most time does not
1450 * What happens if an allocation gets failed. Basically,
1451 * an "overflow" path is triggered to purge lazily freed
1452 * areas to free some memory, then, the "retry" path is
1453 * triggered to repeat one more time. See more details
1454 * in alloc_vmap_area() function.
1456 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1462 * Build the remainder.
1464 lva
->va_start
= va
->va_start
;
1465 lva
->va_end
= nva_start_addr
;
1468 * Shrink this VA to remaining size.
1470 va
->va_start
= nva_start_addr
+ size
;
1475 if (type
!= FL_FIT_TYPE
) {
1476 augment_tree_propagate_from(va
);
1478 if (lva
) /* type == NE_FIT_TYPE */
1479 insert_vmap_area_augment(lva
, &va
->rb_node
, root
, head
);
1486 * Returns a start address of the newly allocated area, if success.
1487 * Otherwise a vend is returned that indicates failure.
1489 static __always_inline
unsigned long
1490 __alloc_vmap_area(struct rb_root
*root
, struct list_head
*head
,
1491 unsigned long size
, unsigned long align
,
1492 unsigned long vstart
, unsigned long vend
)
1494 bool adjust_search_size
= true;
1495 unsigned long nva_start_addr
;
1496 struct vmap_area
*va
;
1500 * Do not adjust when:
1501 * a) align <= PAGE_SIZE, because it does not make any sense.
1502 * All blocks(their start addresses) are at least PAGE_SIZE
1504 * b) a short range where a requested size corresponds to exactly
1505 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1506 * With adjusted search length an allocation would not succeed.
1508 if (align
<= PAGE_SIZE
|| (align
> PAGE_SIZE
&& (vend
- vstart
) == size
))
1509 adjust_search_size
= false;
1511 va
= find_vmap_lowest_match(root
, size
, align
, vstart
, adjust_search_size
);
1515 if (va
->va_start
> vstart
)
1516 nva_start_addr
= ALIGN(va
->va_start
, align
);
1518 nva_start_addr
= ALIGN(vstart
, align
);
1520 /* Check the "vend" restriction. */
1521 if (nva_start_addr
+ size
> vend
)
1524 /* Update the free vmap_area. */
1525 ret
= adjust_va_to_fit_type(root
, head
, va
, nva_start_addr
, size
);
1526 if (WARN_ON_ONCE(ret
))
1529 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1530 find_vmap_lowest_match_check(root
, head
, size
, align
);
1533 return nva_start_addr
;
1537 * Free a region of KVA allocated by alloc_vmap_area
1539 static void free_vmap_area(struct vmap_area
*va
)
1542 * Remove from the busy tree/list.
1544 spin_lock(&vmap_area_lock
);
1545 unlink_va(va
, &vmap_area_root
);
1546 spin_unlock(&vmap_area_lock
);
1549 * Insert/Merge it back to the free tree/list.
1551 spin_lock(&free_vmap_area_lock
);
1552 merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1553 spin_unlock(&free_vmap_area_lock
);
1557 preload_this_cpu_lock(spinlock_t
*lock
, gfp_t gfp_mask
, int node
)
1559 struct vmap_area
*va
= NULL
;
1562 * Preload this CPU with one extra vmap_area object. It is used
1563 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1564 * a CPU that does an allocation is preloaded.
1566 * We do it in non-atomic context, thus it allows us to use more
1567 * permissive allocation masks to be more stable under low memory
1568 * condition and high memory pressure.
1570 if (!this_cpu_read(ne_fit_preload_node
))
1571 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1575 if (va
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, va
))
1576 kmem_cache_free(vmap_area_cachep
, va
);
1580 * Allocate a region of KVA of the specified size and alignment, within the
1583 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1584 unsigned long align
,
1585 unsigned long vstart
, unsigned long vend
,
1586 int node
, gfp_t gfp_mask
,
1587 unsigned long va_flags
)
1589 struct vmap_area
*va
;
1590 unsigned long freed
;
1595 if (unlikely(!size
|| offset_in_page(size
) || !is_power_of_2(align
)))
1596 return ERR_PTR(-EINVAL
);
1598 if (unlikely(!vmap_initialized
))
1599 return ERR_PTR(-EBUSY
);
1602 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1604 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1606 return ERR_PTR(-ENOMEM
);
1609 * Only scan the relevant parts containing pointers to other objects
1610 * to avoid false negatives.
1612 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1615 preload_this_cpu_lock(&free_vmap_area_lock
, gfp_mask
, node
);
1616 addr
= __alloc_vmap_area(&free_vmap_area_root
, &free_vmap_area_list
,
1617 size
, align
, vstart
, vend
);
1618 spin_unlock(&free_vmap_area_lock
);
1620 trace_alloc_vmap_area(addr
, size
, align
, vstart
, vend
, addr
== vend
);
1623 * If an allocation fails, the "vend" address is
1624 * returned. Therefore trigger the overflow path.
1626 if (unlikely(addr
== vend
))
1629 va
->va_start
= addr
;
1630 va
->va_end
= addr
+ size
;
1632 va
->flags
= va_flags
;
1634 spin_lock(&vmap_area_lock
);
1635 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1636 spin_unlock(&vmap_area_lock
);
1638 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1639 BUG_ON(va
->va_start
< vstart
);
1640 BUG_ON(va
->va_end
> vend
);
1642 ret
= kasan_populate_vmalloc(addr
, size
);
1645 return ERR_PTR(ret
);
1652 purge_vmap_area_lazy();
1658 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1665 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1666 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1669 kmem_cache_free(vmap_area_cachep
, va
);
1670 return ERR_PTR(-EBUSY
);
1673 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1675 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1677 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1679 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1681 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1683 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1686 * lazy_max_pages is the maximum amount of virtual address space we gather up
1687 * before attempting to purge with a TLB flush.
1689 * There is a tradeoff here: a larger number will cover more kernel page tables
1690 * and take slightly longer to purge, but it will linearly reduce the number of
1691 * global TLB flushes that must be performed. It would seem natural to scale
1692 * this number up linearly with the number of CPUs (because vmapping activity
1693 * could also scale linearly with the number of CPUs), however it is likely
1694 * that in practice, workloads might be constrained in other ways that mean
1695 * vmap activity will not scale linearly with CPUs. Also, I want to be
1696 * conservative and not introduce a big latency on huge systems, so go with
1697 * a less aggressive log scale. It will still be an improvement over the old
1698 * code, and it will be simple to change the scale factor if we find that it
1699 * becomes a problem on bigger systems.
1701 static unsigned long lazy_max_pages(void)
1705 log
= fls(num_online_cpus());
1707 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1710 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1713 * Serialize vmap purging. There is no actual critical section protected
1714 * by this lock, but we want to avoid concurrent calls for performance
1715 * reasons and to make the pcpu_get_vm_areas more deterministic.
1717 static DEFINE_MUTEX(vmap_purge_lock
);
1719 /* for per-CPU blocks */
1720 static void purge_fragmented_blocks_allcpus(void);
1723 * Purges all lazily-freed vmap areas.
1725 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1727 unsigned long resched_threshold
;
1728 unsigned int num_purged_areas
= 0;
1729 struct list_head local_purge_list
;
1730 struct vmap_area
*va
, *n_va
;
1732 lockdep_assert_held(&vmap_purge_lock
);
1734 spin_lock(&purge_vmap_area_lock
);
1735 purge_vmap_area_root
= RB_ROOT
;
1736 list_replace_init(&purge_vmap_area_list
, &local_purge_list
);
1737 spin_unlock(&purge_vmap_area_lock
);
1739 if (unlikely(list_empty(&local_purge_list
)))
1743 list_first_entry(&local_purge_list
,
1744 struct vmap_area
, list
)->va_start
);
1747 list_last_entry(&local_purge_list
,
1748 struct vmap_area
, list
)->va_end
);
1750 flush_tlb_kernel_range(start
, end
);
1751 resched_threshold
= lazy_max_pages() << 1;
1753 spin_lock(&free_vmap_area_lock
);
1754 list_for_each_entry_safe(va
, n_va
, &local_purge_list
, list
) {
1755 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1756 unsigned long orig_start
= va
->va_start
;
1757 unsigned long orig_end
= va
->va_end
;
1760 * Finally insert or merge lazily-freed area. It is
1761 * detached and there is no need to "unlink" it from
1764 va
= merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
,
1765 &free_vmap_area_list
);
1770 if (is_vmalloc_or_module_addr((void *)orig_start
))
1771 kasan_release_vmalloc(orig_start
, orig_end
,
1772 va
->va_start
, va
->va_end
);
1774 atomic_long_sub(nr
, &vmap_lazy_nr
);
1777 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1778 cond_resched_lock(&free_vmap_area_lock
);
1780 spin_unlock(&free_vmap_area_lock
);
1783 trace_purge_vmap_area_lazy(start
, end
, num_purged_areas
);
1784 return num_purged_areas
> 0;
1788 * Kick off a purge of the outstanding lazy areas.
1790 static void purge_vmap_area_lazy(void)
1792 mutex_lock(&vmap_purge_lock
);
1793 purge_fragmented_blocks_allcpus();
1794 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1795 mutex_unlock(&vmap_purge_lock
);
1798 static void drain_vmap_area_work(struct work_struct
*work
)
1800 unsigned long nr_lazy
;
1803 mutex_lock(&vmap_purge_lock
);
1804 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1805 mutex_unlock(&vmap_purge_lock
);
1807 /* Recheck if further work is required. */
1808 nr_lazy
= atomic_long_read(&vmap_lazy_nr
);
1809 } while (nr_lazy
> lazy_max_pages());
1813 * Free a vmap area, caller ensuring that the area has been unmapped,
1814 * unlinked and flush_cache_vunmap had been called for the correct
1817 static void free_vmap_area_noflush(struct vmap_area
*va
)
1819 unsigned long nr_lazy_max
= lazy_max_pages();
1820 unsigned long va_start
= va
->va_start
;
1821 unsigned long nr_lazy
;
1823 if (WARN_ON_ONCE(!list_empty(&va
->list
)))
1826 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1827 PAGE_SHIFT
, &vmap_lazy_nr
);
1830 * Merge or place it to the purge tree/list.
1832 spin_lock(&purge_vmap_area_lock
);
1833 merge_or_add_vmap_area(va
,
1834 &purge_vmap_area_root
, &purge_vmap_area_list
);
1835 spin_unlock(&purge_vmap_area_lock
);
1837 trace_free_vmap_area_noflush(va_start
, nr_lazy
, nr_lazy_max
);
1839 /* After this point, we may free va at any time */
1840 if (unlikely(nr_lazy
> nr_lazy_max
))
1841 schedule_work(&drain_vmap_work
);
1845 * Free and unmap a vmap area
1847 static void free_unmap_vmap_area(struct vmap_area
*va
)
1849 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1850 vunmap_range_noflush(va
->va_start
, va
->va_end
);
1851 if (debug_pagealloc_enabled_static())
1852 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1854 free_vmap_area_noflush(va
);
1857 struct vmap_area
*find_vmap_area(unsigned long addr
)
1859 struct vmap_area
*va
;
1861 spin_lock(&vmap_area_lock
);
1862 va
= __find_vmap_area(addr
, &vmap_area_root
);
1863 spin_unlock(&vmap_area_lock
);
1868 static struct vmap_area
*find_unlink_vmap_area(unsigned long addr
)
1870 struct vmap_area
*va
;
1872 spin_lock(&vmap_area_lock
);
1873 va
= __find_vmap_area(addr
, &vmap_area_root
);
1875 unlink_va(va
, &vmap_area_root
);
1876 spin_unlock(&vmap_area_lock
);
1881 /*** Per cpu kva allocator ***/
1884 * vmap space is limited especially on 32 bit architectures. Ensure there is
1885 * room for at least 16 percpu vmap blocks per CPU.
1888 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1889 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1890 * instead (we just need a rough idea)
1892 #if BITS_PER_LONG == 32
1893 #define VMALLOC_SPACE (128UL*1024*1024)
1895 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1898 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1899 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1900 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1901 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1902 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1903 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1904 #define VMAP_BBMAP_BITS \
1905 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1906 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1907 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1909 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1911 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
1912 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
1913 #define VMAP_FLAGS_MASK 0x3
1915 struct vmap_block_queue
{
1917 struct list_head free
;
1922 struct vmap_area
*va
;
1923 unsigned long free
, dirty
;
1924 DECLARE_BITMAP(used_map
, VMAP_BBMAP_BITS
);
1925 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1926 struct list_head free_list
;
1927 struct rcu_head rcu_head
;
1928 struct list_head purge
;
1931 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1932 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1935 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1936 * in the free path. Could get rid of this if we change the API to return a
1937 * "cookie" from alloc, to be passed to free. But no big deal yet.
1939 static DEFINE_XARRAY(vmap_blocks
);
1942 * We should probably have a fallback mechanism to allocate virtual memory
1943 * out of partially filled vmap blocks. However vmap block sizing should be
1944 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1948 static unsigned long addr_to_vb_idx(unsigned long addr
)
1950 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1951 addr
/= VMAP_BLOCK_SIZE
;
1955 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1959 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1960 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1961 return (void *)addr
;
1965 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1966 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1967 * @order: how many 2^order pages should be occupied in newly allocated block
1968 * @gfp_mask: flags for the page level allocator
1970 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1972 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1974 struct vmap_block_queue
*vbq
;
1975 struct vmap_block
*vb
;
1976 struct vmap_area
*va
;
1977 unsigned long vb_idx
;
1981 node
= numa_node_id();
1983 vb
= kmalloc_node(sizeof(struct vmap_block
),
1984 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1986 return ERR_PTR(-ENOMEM
);
1988 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1989 VMALLOC_START
, VMALLOC_END
,
1991 VMAP_RAM
|VMAP_BLOCK
);
1994 return ERR_CAST(va
);
1997 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1998 spin_lock_init(&vb
->lock
);
2000 /* At least something should be left free */
2001 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
2002 bitmap_zero(vb
->used_map
, VMAP_BBMAP_BITS
);
2003 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
2005 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2007 bitmap_set(vb
->used_map
, 0, (1UL << order
));
2008 INIT_LIST_HEAD(&vb
->free_list
);
2010 vb_idx
= addr_to_vb_idx(va
->va_start
);
2011 err
= xa_insert(&vmap_blocks
, vb_idx
, vb
, gfp_mask
);
2015 return ERR_PTR(err
);
2018 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2019 spin_lock(&vbq
->lock
);
2020 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
2021 spin_unlock(&vbq
->lock
);
2026 static void free_vmap_block(struct vmap_block
*vb
)
2028 struct vmap_block
*tmp
;
2030 tmp
= xa_erase(&vmap_blocks
, addr_to_vb_idx(vb
->va
->va_start
));
2033 spin_lock(&vmap_area_lock
);
2034 unlink_va(vb
->va
, &vmap_area_root
);
2035 spin_unlock(&vmap_area_lock
);
2037 free_vmap_area_noflush(vb
->va
);
2038 kfree_rcu(vb
, rcu_head
);
2041 static void purge_fragmented_blocks(int cpu
)
2044 struct vmap_block
*vb
;
2045 struct vmap_block
*n_vb
;
2046 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2049 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2051 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
2054 spin_lock(&vb
->lock
);
2055 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2056 vb
->free
= 0; /* prevent further allocs after releasing lock */
2057 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
2059 vb
->dirty_max
= VMAP_BBMAP_BITS
;
2060 spin_lock(&vbq
->lock
);
2061 list_del_rcu(&vb
->free_list
);
2062 spin_unlock(&vbq
->lock
);
2063 spin_unlock(&vb
->lock
);
2064 list_add_tail(&vb
->purge
, &purge
);
2066 spin_unlock(&vb
->lock
);
2070 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
2071 list_del(&vb
->purge
);
2072 free_vmap_block(vb
);
2076 static void purge_fragmented_blocks_allcpus(void)
2080 for_each_possible_cpu(cpu
)
2081 purge_fragmented_blocks(cpu
);
2084 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
2086 struct vmap_block_queue
*vbq
;
2087 struct vmap_block
*vb
;
2091 BUG_ON(offset_in_page(size
));
2092 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2093 if (WARN_ON(size
== 0)) {
2095 * Allocating 0 bytes isn't what caller wants since
2096 * get_order(0) returns funny result. Just warn and terminate
2101 order
= get_order(size
);
2104 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2105 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2106 unsigned long pages_off
;
2108 spin_lock(&vb
->lock
);
2109 if (vb
->free
< (1UL << order
)) {
2110 spin_unlock(&vb
->lock
);
2114 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
2115 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
2116 vb
->free
-= 1UL << order
;
2117 bitmap_set(vb
->used_map
, pages_off
, (1UL << order
));
2118 if (vb
->free
== 0) {
2119 spin_lock(&vbq
->lock
);
2120 list_del_rcu(&vb
->free_list
);
2121 spin_unlock(&vbq
->lock
);
2124 spin_unlock(&vb
->lock
);
2130 /* Allocate new block if nothing was found */
2132 vaddr
= new_vmap_block(order
, gfp_mask
);
2137 static void vb_free(unsigned long addr
, unsigned long size
)
2139 unsigned long offset
;
2141 struct vmap_block
*vb
;
2143 BUG_ON(offset_in_page(size
));
2144 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2146 flush_cache_vunmap(addr
, addr
+ size
);
2148 order
= get_order(size
);
2149 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
2150 vb
= xa_load(&vmap_blocks
, addr_to_vb_idx(addr
));
2151 spin_lock(&vb
->lock
);
2152 bitmap_clear(vb
->used_map
, offset
, (1UL << order
));
2153 spin_unlock(&vb
->lock
);
2155 vunmap_range_noflush(addr
, addr
+ size
);
2157 if (debug_pagealloc_enabled_static())
2158 flush_tlb_kernel_range(addr
, addr
+ size
);
2160 spin_lock(&vb
->lock
);
2162 /* Expand dirty range */
2163 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
2164 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
2166 vb
->dirty
+= 1UL << order
;
2167 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
2169 spin_unlock(&vb
->lock
);
2170 free_vmap_block(vb
);
2172 spin_unlock(&vb
->lock
);
2175 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
2179 if (unlikely(!vmap_initialized
))
2184 for_each_possible_cpu(cpu
) {
2185 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2186 struct vmap_block
*vb
;
2189 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2190 spin_lock(&vb
->lock
);
2191 if (vb
->dirty
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2192 unsigned long va_start
= vb
->va
->va_start
;
2195 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
2196 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
2198 start
= min(s
, start
);
2203 spin_unlock(&vb
->lock
);
2208 mutex_lock(&vmap_purge_lock
);
2209 purge_fragmented_blocks_allcpus();
2210 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
2211 flush_tlb_kernel_range(start
, end
);
2212 mutex_unlock(&vmap_purge_lock
);
2216 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2218 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2219 * to amortize TLB flushing overheads. What this means is that any page you
2220 * have now, may, in a former life, have been mapped into kernel virtual
2221 * address by the vmap layer and so there might be some CPUs with TLB entries
2222 * still referencing that page (additional to the regular 1:1 kernel mapping).
2224 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2225 * be sure that none of the pages we have control over will have any aliases
2226 * from the vmap layer.
2228 void vm_unmap_aliases(void)
2230 unsigned long start
= ULONG_MAX
, end
= 0;
2233 _vm_unmap_aliases(start
, end
, flush
);
2235 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
2238 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2239 * @mem: the pointer returned by vm_map_ram
2240 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2242 void vm_unmap_ram(const void *mem
, unsigned int count
)
2244 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2245 unsigned long addr
= (unsigned long)kasan_reset_tag(mem
);
2246 struct vmap_area
*va
;
2250 BUG_ON(addr
< VMALLOC_START
);
2251 BUG_ON(addr
> VMALLOC_END
);
2252 BUG_ON(!PAGE_ALIGNED(addr
));
2254 kasan_poison_vmalloc(mem
, size
);
2256 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2257 debug_check_no_locks_freed(mem
, size
);
2258 vb_free(addr
, size
);
2262 va
= find_unlink_vmap_area(addr
);
2263 if (WARN_ON_ONCE(!va
))
2266 debug_check_no_locks_freed((void *)va
->va_start
,
2267 (va
->va_end
- va
->va_start
));
2268 free_unmap_vmap_area(va
);
2270 EXPORT_SYMBOL(vm_unmap_ram
);
2273 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2274 * @pages: an array of pointers to the pages to be mapped
2275 * @count: number of pages
2276 * @node: prefer to allocate data structures on this node
2278 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2279 * faster than vmap so it's good. But if you mix long-life and short-life
2280 * objects with vm_map_ram(), it could consume lots of address space through
2281 * fragmentation (especially on a 32bit machine). You could see failures in
2282 * the end. Please use this function for short-lived objects.
2284 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2286 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
2288 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2292 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2293 mem
= vb_alloc(size
, GFP_KERNEL
);
2296 addr
= (unsigned long)mem
;
2298 struct vmap_area
*va
;
2299 va
= alloc_vmap_area(size
, PAGE_SIZE
,
2300 VMALLOC_START
, VMALLOC_END
,
2301 node
, GFP_KERNEL
, VMAP_RAM
);
2305 addr
= va
->va_start
;
2309 if (vmap_pages_range(addr
, addr
+ size
, PAGE_KERNEL
,
2310 pages
, PAGE_SHIFT
) < 0) {
2311 vm_unmap_ram(mem
, count
);
2316 * Mark the pages as accessible, now that they are mapped.
2317 * With hardware tag-based KASAN, marking is skipped for
2318 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2320 mem
= kasan_unpoison_vmalloc(mem
, size
, KASAN_VMALLOC_PROT_NORMAL
);
2324 EXPORT_SYMBOL(vm_map_ram
);
2326 static struct vm_struct
*vmlist __initdata
;
2328 static inline unsigned int vm_area_page_order(struct vm_struct
*vm
)
2330 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2331 return vm
->page_order
;
2337 static inline void set_vm_area_page_order(struct vm_struct
*vm
, unsigned int order
)
2339 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2340 vm
->page_order
= order
;
2347 * vm_area_add_early - add vmap area early during boot
2348 * @vm: vm_struct to add
2350 * This function is used to add fixed kernel vm area to vmlist before
2351 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2352 * should contain proper values and the other fields should be zero.
2354 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2356 void __init
vm_area_add_early(struct vm_struct
*vm
)
2358 struct vm_struct
*tmp
, **p
;
2360 BUG_ON(vmap_initialized
);
2361 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
2362 if (tmp
->addr
>= vm
->addr
) {
2363 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
2366 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
2373 * vm_area_register_early - register vmap area early during boot
2374 * @vm: vm_struct to register
2375 * @align: requested alignment
2377 * This function is used to register kernel vm area before
2378 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2379 * proper values on entry and other fields should be zero. On return,
2380 * vm->addr contains the allocated address.
2382 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2384 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
2386 unsigned long addr
= ALIGN(VMALLOC_START
, align
);
2387 struct vm_struct
*cur
, **p
;
2389 BUG_ON(vmap_initialized
);
2391 for (p
= &vmlist
; (cur
= *p
) != NULL
; p
= &cur
->next
) {
2392 if ((unsigned long)cur
->addr
- addr
>= vm
->size
)
2394 addr
= ALIGN((unsigned long)cur
->addr
+ cur
->size
, align
);
2397 BUG_ON(addr
> VMALLOC_END
- vm
->size
);
2398 vm
->addr
= (void *)addr
;
2401 kasan_populate_early_vm_area_shadow(vm
->addr
, vm
->size
);
2404 static void vmap_init_free_space(void)
2406 unsigned long vmap_start
= 1;
2407 const unsigned long vmap_end
= ULONG_MAX
;
2408 struct vmap_area
*busy
, *free
;
2412 * -|-----|.....|-----|-----|-----|.....|-
2414 * |<--------------------------------->|
2416 list_for_each_entry(busy
, &vmap_area_list
, list
) {
2417 if (busy
->va_start
- vmap_start
> 0) {
2418 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2419 if (!WARN_ON_ONCE(!free
)) {
2420 free
->va_start
= vmap_start
;
2421 free
->va_end
= busy
->va_start
;
2423 insert_vmap_area_augment(free
, NULL
,
2424 &free_vmap_area_root
,
2425 &free_vmap_area_list
);
2429 vmap_start
= busy
->va_end
;
2432 if (vmap_end
- vmap_start
> 0) {
2433 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2434 if (!WARN_ON_ONCE(!free
)) {
2435 free
->va_start
= vmap_start
;
2436 free
->va_end
= vmap_end
;
2438 insert_vmap_area_augment(free
, NULL
,
2439 &free_vmap_area_root
,
2440 &free_vmap_area_list
);
2445 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2446 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2449 vm
->addr
= (void *)va
->va_start
;
2450 vm
->size
= va
->va_end
- va
->va_start
;
2451 vm
->caller
= caller
;
2455 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2456 unsigned long flags
, const void *caller
)
2458 spin_lock(&vmap_area_lock
);
2459 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2460 spin_unlock(&vmap_area_lock
);
2463 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2466 * Before removing VM_UNINITIALIZED,
2467 * we should make sure that vm has proper values.
2468 * Pair with smp_rmb() in show_numa_info().
2471 vm
->flags
&= ~VM_UNINITIALIZED
;
2474 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2475 unsigned long align
, unsigned long shift
, unsigned long flags
,
2476 unsigned long start
, unsigned long end
, int node
,
2477 gfp_t gfp_mask
, const void *caller
)
2479 struct vmap_area
*va
;
2480 struct vm_struct
*area
;
2481 unsigned long requested_size
= size
;
2483 BUG_ON(in_interrupt());
2484 size
= ALIGN(size
, 1ul << shift
);
2485 if (unlikely(!size
))
2488 if (flags
& VM_IOREMAP
)
2489 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2490 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2492 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2493 if (unlikely(!area
))
2496 if (!(flags
& VM_NO_GUARD
))
2499 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
, 0);
2505 setup_vmalloc_vm(area
, va
, flags
, caller
);
2508 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2509 * best-effort approach, as they can be mapped outside of vmalloc code.
2510 * For VM_ALLOC mappings, the pages are marked as accessible after
2511 * getting mapped in __vmalloc_node_range().
2512 * With hardware tag-based KASAN, marking is skipped for
2513 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2515 if (!(flags
& VM_ALLOC
))
2516 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, requested_size
,
2517 KASAN_VMALLOC_PROT_NORMAL
);
2522 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2523 unsigned long start
, unsigned long end
,
2526 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
, start
, end
,
2527 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2531 * get_vm_area - reserve a contiguous kernel virtual area
2532 * @size: size of the area
2533 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2535 * Search an area of @size in the kernel virtual mapping area,
2536 * and reserved it for out purposes. Returns the area descriptor
2537 * on success or %NULL on failure.
2539 * Return: the area descriptor on success or %NULL on failure.
2541 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2543 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2544 VMALLOC_START
, VMALLOC_END
,
2545 NUMA_NO_NODE
, GFP_KERNEL
,
2546 __builtin_return_address(0));
2549 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2552 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2553 VMALLOC_START
, VMALLOC_END
,
2554 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2558 * find_vm_area - find a continuous kernel virtual area
2559 * @addr: base address
2561 * Search for the kernel VM area starting at @addr, and return it.
2562 * It is up to the caller to do all required locking to keep the returned
2565 * Return: the area descriptor on success or %NULL on failure.
2567 struct vm_struct
*find_vm_area(const void *addr
)
2569 struct vmap_area
*va
;
2571 va
= find_vmap_area((unsigned long)addr
);
2579 * remove_vm_area - find and remove a continuous kernel virtual area
2580 * @addr: base address
2582 * Search for the kernel VM area starting at @addr, and remove it.
2583 * This function returns the found VM area, but using it is NOT safe
2584 * on SMP machines, except for its size or flags.
2586 * Return: the area descriptor on success or %NULL on failure.
2588 struct vm_struct
*remove_vm_area(const void *addr
)
2590 struct vmap_area
*va
;
2591 struct vm_struct
*vm
;
2595 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2599 va
= find_unlink_vmap_area((unsigned long)addr
);
2604 debug_check_no_locks_freed(vm
->addr
, get_vm_area_size(vm
));
2605 debug_check_no_obj_freed(vm
->addr
, get_vm_area_size(vm
));
2606 kasan_free_module_shadow(vm
);
2607 kasan_poison_vmalloc(vm
->addr
, get_vm_area_size(vm
));
2609 free_unmap_vmap_area(va
);
2613 static inline void set_area_direct_map(const struct vm_struct
*area
,
2614 int (*set_direct_map
)(struct page
*page
))
2618 /* HUGE_VMALLOC passes small pages to set_direct_map */
2619 for (i
= 0; i
< area
->nr_pages
; i
++)
2620 if (page_address(area
->pages
[i
]))
2621 set_direct_map(area
->pages
[i
]);
2625 * Flush the vm mapping and reset the direct map.
2627 static void vm_reset_perms(struct vm_struct
*area
)
2629 unsigned long start
= ULONG_MAX
, end
= 0;
2630 unsigned int page_order
= vm_area_page_order(area
);
2635 * Find the start and end range of the direct mappings to make sure that
2636 * the vm_unmap_aliases() flush includes the direct map.
2638 for (i
= 0; i
< area
->nr_pages
; i
+= 1U << page_order
) {
2639 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2642 unsigned long page_size
;
2644 page_size
= PAGE_SIZE
<< page_order
;
2645 start
= min(addr
, start
);
2646 end
= max(addr
+ page_size
, end
);
2652 * Set direct map to something invalid so that it won't be cached if
2653 * there are any accesses after the TLB flush, then flush the TLB and
2654 * reset the direct map permissions to the default.
2656 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2657 _vm_unmap_aliases(start
, end
, flush_dmap
);
2658 set_area_direct_map(area
, set_direct_map_default_noflush
);
2661 static void delayed_vfree_work(struct work_struct
*w
)
2663 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
2664 struct llist_node
*t
, *llnode
;
2666 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
2671 * vfree_atomic - release memory allocated by vmalloc()
2672 * @addr: memory base address
2674 * This one is just like vfree() but can be called in any atomic context
2677 void vfree_atomic(const void *addr
)
2679 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2682 kmemleak_free(addr
);
2685 * Use raw_cpu_ptr() because this can be called from preemptible
2686 * context. Preemption is absolutely fine here, because the llist_add()
2687 * implementation is lockless, so it works even if we are adding to
2688 * another cpu's list. schedule_work() should be fine with this too.
2690 if (addr
&& llist_add((struct llist_node
*)addr
, &p
->list
))
2691 schedule_work(&p
->wq
);
2695 * vfree - Release memory allocated by vmalloc()
2696 * @addr: Memory base address
2698 * Free the virtually continuous memory area starting at @addr, as obtained
2699 * from one of the vmalloc() family of APIs. This will usually also free the
2700 * physical memory underlying the virtual allocation, but that memory is
2701 * reference counted, so it will not be freed until the last user goes away.
2703 * If @addr is NULL, no operation is performed.
2706 * May sleep if called *not* from interrupt context.
2707 * Must not be called in NMI context (strictly speaking, it could be
2708 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2709 * conventions for vfree() arch-dependent would be a really bad idea).
2711 void vfree(const void *addr
)
2713 struct vm_struct
*vm
;
2716 if (unlikely(in_interrupt())) {
2722 kmemleak_free(addr
);
2728 vm
= remove_vm_area(addr
);
2729 if (unlikely(!vm
)) {
2730 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2735 if (unlikely(vm
->flags
& VM_FLUSH_RESET_PERMS
))
2737 for (i
= 0; i
< vm
->nr_pages
; i
++) {
2738 struct page
*page
= vm
->pages
[i
];
2741 mod_memcg_page_state(page
, MEMCG_VMALLOC
, -1);
2743 * High-order allocs for huge vmallocs are split, so
2744 * can be freed as an array of order-0 allocations
2746 __free_pages(page
, 0);
2749 atomic_long_sub(vm
->nr_pages
, &nr_vmalloc_pages
);
2753 EXPORT_SYMBOL(vfree
);
2756 * vunmap - release virtual mapping obtained by vmap()
2757 * @addr: memory base address
2759 * Free the virtually contiguous memory area starting at @addr,
2760 * which was created from the page array passed to vmap().
2762 * Must not be called in interrupt context.
2764 void vunmap(const void *addr
)
2766 struct vm_struct
*vm
;
2768 BUG_ON(in_interrupt());
2773 vm
= remove_vm_area(addr
);
2774 if (unlikely(!vm
)) {
2775 WARN(1, KERN_ERR
"Trying to vunmap() nonexistent vm area (%p)\n",
2781 EXPORT_SYMBOL(vunmap
);
2784 * vmap - map an array of pages into virtually contiguous space
2785 * @pages: array of page pointers
2786 * @count: number of pages to map
2787 * @flags: vm_area->flags
2788 * @prot: page protection for the mapping
2790 * Maps @count pages from @pages into contiguous kernel virtual space.
2791 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2792 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2793 * are transferred from the caller to vmap(), and will be freed / dropped when
2794 * vfree() is called on the return value.
2796 * Return: the address of the area or %NULL on failure
2798 void *vmap(struct page
**pages
, unsigned int count
,
2799 unsigned long flags
, pgprot_t prot
)
2801 struct vm_struct
*area
;
2803 unsigned long size
; /* In bytes */
2807 if (WARN_ON_ONCE(flags
& VM_FLUSH_RESET_PERMS
))
2811 * Your top guard is someone else's bottom guard. Not having a top
2812 * guard compromises someone else's mappings too.
2814 if (WARN_ON_ONCE(flags
& VM_NO_GUARD
))
2815 flags
&= ~VM_NO_GUARD
;
2817 if (count
> totalram_pages())
2820 size
= (unsigned long)count
<< PAGE_SHIFT
;
2821 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2825 addr
= (unsigned long)area
->addr
;
2826 if (vmap_pages_range(addr
, addr
+ size
, pgprot_nx(prot
),
2827 pages
, PAGE_SHIFT
) < 0) {
2832 if (flags
& VM_MAP_PUT_PAGES
) {
2833 area
->pages
= pages
;
2834 area
->nr_pages
= count
;
2838 EXPORT_SYMBOL(vmap
);
2840 #ifdef CONFIG_VMAP_PFN
2841 struct vmap_pfn_data
{
2842 unsigned long *pfns
;
2847 static int vmap_pfn_apply(pte_t
*pte
, unsigned long addr
, void *private)
2849 struct vmap_pfn_data
*data
= private;
2851 if (WARN_ON_ONCE(pfn_valid(data
->pfns
[data
->idx
])))
2853 *pte
= pte_mkspecial(pfn_pte(data
->pfns
[data
->idx
++], data
->prot
));
2858 * vmap_pfn - map an array of PFNs into virtually contiguous space
2859 * @pfns: array of PFNs
2860 * @count: number of pages to map
2861 * @prot: page protection for the mapping
2863 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2864 * the start address of the mapping.
2866 void *vmap_pfn(unsigned long *pfns
, unsigned int count
, pgprot_t prot
)
2868 struct vmap_pfn_data data
= { .pfns
= pfns
, .prot
= pgprot_nx(prot
) };
2869 struct vm_struct
*area
;
2871 area
= get_vm_area_caller(count
* PAGE_SIZE
, VM_IOREMAP
,
2872 __builtin_return_address(0));
2875 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2876 count
* PAGE_SIZE
, vmap_pfn_apply
, &data
)) {
2882 EXPORT_SYMBOL_GPL(vmap_pfn
);
2883 #endif /* CONFIG_VMAP_PFN */
2885 static inline unsigned int
2886 vm_area_alloc_pages(gfp_t gfp
, int nid
,
2887 unsigned int order
, unsigned int nr_pages
, struct page
**pages
)
2889 unsigned int nr_allocated
= 0;
2890 gfp_t alloc_gfp
= gfp
;
2891 bool nofail
= false;
2896 * For order-0 pages we make use of bulk allocator, if
2897 * the page array is partly or not at all populated due
2898 * to fails, fallback to a single page allocator that is
2902 /* bulk allocator doesn't support nofail req. officially */
2903 gfp_t bulk_gfp
= gfp
& ~__GFP_NOFAIL
;
2905 while (nr_allocated
< nr_pages
) {
2906 unsigned int nr
, nr_pages_request
;
2909 * A maximum allowed request is hard-coded and is 100
2910 * pages per call. That is done in order to prevent a
2911 * long preemption off scenario in the bulk-allocator
2912 * so the range is [1:100].
2914 nr_pages_request
= min(100U, nr_pages
- nr_allocated
);
2916 /* memory allocation should consider mempolicy, we can't
2917 * wrongly use nearest node when nid == NUMA_NO_NODE,
2918 * otherwise memory may be allocated in only one node,
2919 * but mempolicy wants to alloc memory by interleaving.
2921 if (IS_ENABLED(CONFIG_NUMA
) && nid
== NUMA_NO_NODE
)
2922 nr
= alloc_pages_bulk_array_mempolicy(bulk_gfp
,
2924 pages
+ nr_allocated
);
2927 nr
= alloc_pages_bulk_array_node(bulk_gfp
, nid
,
2929 pages
+ nr_allocated
);
2935 * If zero or pages were obtained partly,
2936 * fallback to a single page allocator.
2938 if (nr
!= nr_pages_request
)
2941 } else if (gfp
& __GFP_NOFAIL
) {
2943 * Higher order nofail allocations are really expensive and
2944 * potentially dangerous (pre-mature OOM, disruptive reclaim
2945 * and compaction etc.
2947 alloc_gfp
&= ~__GFP_NOFAIL
;
2951 /* High-order pages or fallback path if "bulk" fails. */
2952 while (nr_allocated
< nr_pages
) {
2953 if (fatal_signal_pending(current
))
2956 if (nid
== NUMA_NO_NODE
)
2957 page
= alloc_pages(alloc_gfp
, order
);
2959 page
= alloc_pages_node(nid
, alloc_gfp
, order
);
2960 if (unlikely(!page
)) {
2964 /* fall back to the zero order allocations */
2965 alloc_gfp
|= __GFP_NOFAIL
;
2971 * Higher order allocations must be able to be treated as
2972 * indepdenent small pages by callers (as they can with
2973 * small-page vmallocs). Some drivers do their own refcounting
2974 * on vmalloc_to_page() pages, some use page->mapping,
2978 split_page(page
, order
);
2981 * Careful, we allocate and map page-order pages, but
2982 * tracking is done per PAGE_SIZE page so as to keep the
2983 * vm_struct APIs independent of the physical/mapped size.
2985 for (i
= 0; i
< (1U << order
); i
++)
2986 pages
[nr_allocated
+ i
] = page
+ i
;
2989 nr_allocated
+= 1U << order
;
2992 return nr_allocated
;
2995 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2996 pgprot_t prot
, unsigned int page_shift
,
2999 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
3000 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3001 unsigned long addr
= (unsigned long)area
->addr
;
3002 unsigned long size
= get_vm_area_size(area
);
3003 unsigned long array_size
;
3004 unsigned int nr_small_pages
= size
>> PAGE_SHIFT
;
3005 unsigned int page_order
;
3009 array_size
= (unsigned long)nr_small_pages
* sizeof(struct page
*);
3011 if (!(gfp_mask
& (GFP_DMA
| GFP_DMA32
)))
3012 gfp_mask
|= __GFP_HIGHMEM
;
3014 /* Please note that the recursion is strictly bounded. */
3015 if (array_size
> PAGE_SIZE
) {
3016 area
->pages
= __vmalloc_node(array_size
, 1, nested_gfp
, node
,
3019 area
->pages
= kmalloc_node(array_size
, nested_gfp
, node
);
3023 warn_alloc(gfp_mask
, NULL
,
3024 "vmalloc error: size %lu, failed to allocated page array size %lu",
3025 nr_small_pages
* PAGE_SIZE
, array_size
);
3030 set_vm_area_page_order(area
, page_shift
- PAGE_SHIFT
);
3031 page_order
= vm_area_page_order(area
);
3033 area
->nr_pages
= vm_area_alloc_pages(gfp_mask
| __GFP_NOWARN
,
3034 node
, page_order
, nr_small_pages
, area
->pages
);
3036 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
3037 if (gfp_mask
& __GFP_ACCOUNT
) {
3040 for (i
= 0; i
< area
->nr_pages
; i
++)
3041 mod_memcg_page_state(area
->pages
[i
], MEMCG_VMALLOC
, 1);
3045 * If not enough pages were obtained to accomplish an
3046 * allocation request, free them via vfree() if any.
3048 if (area
->nr_pages
!= nr_small_pages
) {
3049 /* vm_area_alloc_pages() can also fail due to a fatal signal */
3050 if (!fatal_signal_pending(current
))
3051 warn_alloc(gfp_mask
, NULL
,
3052 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3053 area
->nr_pages
* PAGE_SIZE
, page_order
);
3058 * page tables allocations ignore external gfp mask, enforce it
3061 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3062 flags
= memalloc_nofs_save();
3063 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3064 flags
= memalloc_noio_save();
3067 ret
= vmap_pages_range(addr
, addr
+ size
, prot
, area
->pages
,
3069 if (nofail
&& (ret
< 0))
3070 schedule_timeout_uninterruptible(1);
3071 } while (nofail
&& (ret
< 0));
3073 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3074 memalloc_nofs_restore(flags
);
3075 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3076 memalloc_noio_restore(flags
);
3079 warn_alloc(gfp_mask
, NULL
,
3080 "vmalloc error: size %lu, failed to map pages",
3081 area
->nr_pages
* PAGE_SIZE
);
3093 * __vmalloc_node_range - allocate virtually contiguous memory
3094 * @size: allocation size
3095 * @align: desired alignment
3096 * @start: vm area range start
3097 * @end: vm area range end
3098 * @gfp_mask: flags for the page level allocator
3099 * @prot: protection mask for the allocated pages
3100 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3101 * @node: node to use for allocation or NUMA_NO_NODE
3102 * @caller: caller's return address
3104 * Allocate enough pages to cover @size from the page level
3105 * allocator with @gfp_mask flags. Please note that the full set of gfp
3106 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3108 * Zone modifiers are not supported. From the reclaim modifiers
3109 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3110 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3111 * __GFP_RETRY_MAYFAIL are not supported).
3113 * __GFP_NOWARN can be used to suppress failures messages.
3115 * Map them into contiguous kernel virtual space, using a pagetable
3116 * protection of @prot.
3118 * Return: the address of the area or %NULL on failure
3120 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
3121 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
3122 pgprot_t prot
, unsigned long vm_flags
, int node
,
3125 struct vm_struct
*area
;
3127 kasan_vmalloc_flags_t kasan_flags
= KASAN_VMALLOC_NONE
;
3128 unsigned long real_size
= size
;
3129 unsigned long real_align
= align
;
3130 unsigned int shift
= PAGE_SHIFT
;
3132 if (WARN_ON_ONCE(!size
))
3135 if ((size
>> PAGE_SHIFT
) > totalram_pages()) {
3136 warn_alloc(gfp_mask
, NULL
,
3137 "vmalloc error: size %lu, exceeds total pages",
3142 if (vmap_allow_huge
&& (vm_flags
& VM_ALLOW_HUGE_VMAP
)) {
3143 unsigned long size_per_node
;
3146 * Try huge pages. Only try for PAGE_KERNEL allocations,
3147 * others like modules don't yet expect huge pages in
3148 * their allocations due to apply_to_page_range not
3152 size_per_node
= size
;
3153 if (node
== NUMA_NO_NODE
)
3154 size_per_node
/= num_online_nodes();
3155 if (arch_vmap_pmd_supported(prot
) && size_per_node
>= PMD_SIZE
)
3158 shift
= arch_vmap_pte_supported_shift(size_per_node
);
3160 align
= max(real_align
, 1UL << shift
);
3161 size
= ALIGN(real_size
, 1UL << shift
);
3165 area
= __get_vm_area_node(real_size
, align
, shift
, VM_ALLOC
|
3166 VM_UNINITIALIZED
| vm_flags
, start
, end
, node
,
3169 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3170 warn_alloc(gfp_mask
, NULL
,
3171 "vmalloc error: size %lu, vm_struct allocation failed%s",
3172 real_size
, (nofail
) ? ". Retrying." : "");
3174 schedule_timeout_uninterruptible(1);
3181 * Prepare arguments for __vmalloc_area_node() and
3182 * kasan_unpoison_vmalloc().
3184 if (pgprot_val(prot
) == pgprot_val(PAGE_KERNEL
)) {
3185 if (kasan_hw_tags_enabled()) {
3187 * Modify protection bits to allow tagging.
3188 * This must be done before mapping.
3190 prot
= arch_vmap_pgprot_tagged(prot
);
3193 * Skip page_alloc poisoning and zeroing for physical
3194 * pages backing VM_ALLOC mapping. Memory is instead
3195 * poisoned and zeroed by kasan_unpoison_vmalloc().
3197 gfp_mask
|= __GFP_SKIP_KASAN_UNPOISON
| __GFP_SKIP_ZERO
;
3200 /* Take note that the mapping is PAGE_KERNEL. */
3201 kasan_flags
|= KASAN_VMALLOC_PROT_NORMAL
;
3204 /* Allocate physical pages and map them into vmalloc space. */
3205 ret
= __vmalloc_area_node(area
, gfp_mask
, prot
, shift
, node
);
3210 * Mark the pages as accessible, now that they are mapped.
3211 * The condition for setting KASAN_VMALLOC_INIT should complement the
3212 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3213 * to make sure that memory is initialized under the same conditions.
3214 * Tag-based KASAN modes only assign tags to normal non-executable
3215 * allocations, see __kasan_unpoison_vmalloc().
3217 kasan_flags
|= KASAN_VMALLOC_VM_ALLOC
;
3218 if (!want_init_on_free() && want_init_on_alloc(gfp_mask
) &&
3219 (gfp_mask
& __GFP_SKIP_ZERO
))
3220 kasan_flags
|= KASAN_VMALLOC_INIT
;
3221 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3222 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, real_size
, kasan_flags
);
3225 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3226 * flag. It means that vm_struct is not fully initialized.
3227 * Now, it is fully initialized, so remove this flag here.
3229 clear_vm_uninitialized_flag(area
);
3231 size
= PAGE_ALIGN(size
);
3232 if (!(vm_flags
& VM_DEFER_KMEMLEAK
))
3233 kmemleak_vmalloc(area
, size
, gfp_mask
);
3238 if (shift
> PAGE_SHIFT
) {
3249 * __vmalloc_node - allocate virtually contiguous memory
3250 * @size: allocation size
3251 * @align: desired alignment
3252 * @gfp_mask: flags for the page level allocator
3253 * @node: node to use for allocation or NUMA_NO_NODE
3254 * @caller: caller's return address
3256 * Allocate enough pages to cover @size from the page level allocator with
3257 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3259 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3260 * and __GFP_NOFAIL are not supported
3262 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3265 * Return: pointer to the allocated memory or %NULL on error
3267 void *__vmalloc_node(unsigned long size
, unsigned long align
,
3268 gfp_t gfp_mask
, int node
, const void *caller
)
3270 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
3271 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
3274 * This is only for performance analysis of vmalloc and stress purpose.
3275 * It is required by vmalloc test module, therefore do not use it other
3278 #ifdef CONFIG_TEST_VMALLOC_MODULE
3279 EXPORT_SYMBOL_GPL(__vmalloc_node
);
3282 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
3284 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
3285 __builtin_return_address(0));
3287 EXPORT_SYMBOL(__vmalloc
);
3290 * vmalloc - allocate virtually contiguous memory
3291 * @size: allocation size
3293 * Allocate enough pages to cover @size from the page level
3294 * allocator and map them into contiguous kernel virtual space.
3296 * For tight control over page level allocator and protection flags
3297 * use __vmalloc() instead.
3299 * Return: pointer to the allocated memory or %NULL on error
3301 void *vmalloc(unsigned long size
)
3303 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
3304 __builtin_return_address(0));
3306 EXPORT_SYMBOL(vmalloc
);
3309 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3310 * @size: allocation size
3311 * @gfp_mask: flags for the page level allocator
3313 * Allocate enough pages to cover @size from the page level
3314 * allocator and map them into contiguous kernel virtual space.
3315 * If @size is greater than or equal to PMD_SIZE, allow using
3316 * huge pages for the memory
3318 * Return: pointer to the allocated memory or %NULL on error
3320 void *vmalloc_huge(unsigned long size
, gfp_t gfp_mask
)
3322 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
3323 gfp_mask
, PAGE_KERNEL
, VM_ALLOW_HUGE_VMAP
,
3324 NUMA_NO_NODE
, __builtin_return_address(0));
3326 EXPORT_SYMBOL_GPL(vmalloc_huge
);
3329 * vzalloc - allocate virtually contiguous memory with zero fill
3330 * @size: allocation size
3332 * Allocate enough pages to cover @size from the page level
3333 * allocator and map them into contiguous kernel virtual space.
3334 * The memory allocated is set to zero.
3336 * For tight control over page level allocator and protection flags
3337 * use __vmalloc() instead.
3339 * Return: pointer to the allocated memory or %NULL on error
3341 void *vzalloc(unsigned long size
)
3343 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
3344 __builtin_return_address(0));
3346 EXPORT_SYMBOL(vzalloc
);
3349 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3350 * @size: allocation size
3352 * The resulting memory area is zeroed so it can be mapped to userspace
3353 * without leaking data.
3355 * Return: pointer to the allocated memory or %NULL on error
3357 void *vmalloc_user(unsigned long size
)
3359 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3360 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
3361 VM_USERMAP
, NUMA_NO_NODE
,
3362 __builtin_return_address(0));
3364 EXPORT_SYMBOL(vmalloc_user
);
3367 * vmalloc_node - allocate memory on a specific node
3368 * @size: allocation size
3371 * Allocate enough pages to cover @size from the page level
3372 * allocator and map them into contiguous kernel virtual space.
3374 * For tight control over page level allocator and protection flags
3375 * use __vmalloc() instead.
3377 * Return: pointer to the allocated memory or %NULL on error
3379 void *vmalloc_node(unsigned long size
, int node
)
3381 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
3382 __builtin_return_address(0));
3384 EXPORT_SYMBOL(vmalloc_node
);
3387 * vzalloc_node - allocate memory on a specific node with zero fill
3388 * @size: allocation size
3391 * Allocate enough pages to cover @size from the page level
3392 * allocator and map them into contiguous kernel virtual space.
3393 * The memory allocated is set to zero.
3395 * Return: pointer to the allocated memory or %NULL on error
3397 void *vzalloc_node(unsigned long size
, int node
)
3399 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
3400 __builtin_return_address(0));
3402 EXPORT_SYMBOL(vzalloc_node
);
3404 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3405 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3406 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3407 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3410 * 64b systems should always have either DMA or DMA32 zones. For others
3411 * GFP_DMA32 should do the right thing and use the normal zone.
3413 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3417 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3418 * @size: allocation size
3420 * Allocate enough 32bit PA addressable pages to cover @size from the
3421 * page level allocator and map them into contiguous kernel virtual space.
3423 * Return: pointer to the allocated memory or %NULL on error
3425 void *vmalloc_32(unsigned long size
)
3427 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
3428 __builtin_return_address(0));
3430 EXPORT_SYMBOL(vmalloc_32
);
3433 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3434 * @size: allocation size
3436 * The resulting memory area is 32bit addressable and zeroed so it can be
3437 * mapped to userspace without leaking data.
3439 * Return: pointer to the allocated memory or %NULL on error
3441 void *vmalloc_32_user(unsigned long size
)
3443 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3444 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
3445 VM_USERMAP
, NUMA_NO_NODE
,
3446 __builtin_return_address(0));
3448 EXPORT_SYMBOL(vmalloc_32_user
);
3451 * small helper routine , copy contents to buf from addr.
3452 * If the page is not present, fill zero.
3455 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
3461 unsigned long offset
, length
;
3463 offset
= offset_in_page(addr
);
3464 length
= PAGE_SIZE
- offset
;
3467 p
= vmalloc_to_page(addr
);
3469 * To do safe access to this _mapped_ area, we need
3470 * lock. But adding lock here means that we need to add
3471 * overhead of vmalloc()/vfree() calls for this _debug_
3472 * interface, rarely used. Instead of that, we'll use
3473 * kmap() and get small overhead in this access function.
3476 /* We can expect USER0 is not used -- see vread() */
3477 void *map
= kmap_atomic(p
);
3478 memcpy(buf
, map
+ offset
, length
);
3481 memset(buf
, 0, length
);
3491 static void vmap_ram_vread(char *buf
, char *addr
, int count
, unsigned long flags
)
3494 struct vmap_block
*vb
;
3495 unsigned long offset
;
3496 unsigned int rs
, re
, n
;
3499 * If it's area created by vm_map_ram() interface directly, but
3500 * not further subdividing and delegating management to vmap_block,
3503 if (!(flags
& VMAP_BLOCK
)) {
3504 aligned_vread(buf
, addr
, count
);
3509 * Area is split into regions and tracked with vmap_block, read out
3510 * each region and zero fill the hole between regions.
3512 vb
= xa_load(&vmap_blocks
, addr_to_vb_idx((unsigned long)addr
));
3516 spin_lock(&vb
->lock
);
3517 if (bitmap_empty(vb
->used_map
, VMAP_BBMAP_BITS
)) {
3518 spin_unlock(&vb
->lock
);
3521 for_each_set_bitrange(rs
, re
, vb
->used_map
, VMAP_BBMAP_BITS
) {
3524 start
= vmap_block_vaddr(vb
->va
->va_start
, rs
);
3525 while (addr
< start
) {
3533 /*it could start reading from the middle of used region*/
3534 offset
= offset_in_page(addr
);
3535 n
= ((re
- rs
+ 1) << PAGE_SHIFT
) - offset
;
3538 aligned_vread(buf
, start
+offset
, n
);
3545 spin_unlock(&vb
->lock
);
3548 /* zero-fill the left dirty or free regions */
3550 memset(buf
, 0, count
);
3554 * vread() - read vmalloc area in a safe way.
3555 * @buf: buffer for reading data
3556 * @addr: vm address.
3557 * @count: number of bytes to be read.
3559 * This function checks that addr is a valid vmalloc'ed area, and
3560 * copy data from that area to a given buffer. If the given memory range
3561 * of [addr...addr+count) includes some valid address, data is copied to
3562 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3563 * IOREMAP area is treated as memory hole and no copy is done.
3565 * If [addr...addr+count) doesn't includes any intersects with alive
3566 * vm_struct area, returns 0. @buf should be kernel's buffer.
3568 * Note: In usual ops, vread() is never necessary because the caller
3569 * should know vmalloc() area is valid and can use memcpy().
3570 * This is for routines which have to access vmalloc area without
3571 * any information, as /proc/kcore.
3573 * Return: number of bytes for which addr and buf should be increased
3574 * (same number as @count) or %0 if [addr...addr+count) doesn't
3575 * include any intersection with valid vmalloc area
3577 long vread(char *buf
, char *addr
, unsigned long count
)
3579 struct vmap_area
*va
;
3580 struct vm_struct
*vm
;
3581 char *vaddr
, *buf_start
= buf
;
3582 unsigned long buflen
= count
;
3583 unsigned long n
, size
, flags
;
3585 addr
= kasan_reset_tag(addr
);
3587 /* Don't allow overflow */
3588 if ((unsigned long) addr
+ count
< count
)
3589 count
= -(unsigned long) addr
;
3591 spin_lock(&vmap_area_lock
);
3592 va
= find_vmap_area_exceed_addr((unsigned long)addr
);
3596 /* no intersects with alive vmap_area */
3597 if ((unsigned long)addr
+ count
<= va
->va_start
)
3600 list_for_each_entry_from(va
, &vmap_area_list
, list
) {
3605 flags
= va
->flags
& VMAP_FLAGS_MASK
;
3607 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
3608 * be set together with VMAP_RAM.
3610 WARN_ON(flags
== VMAP_BLOCK
);
3615 if (vm
&& (vm
->flags
& VM_UNINITIALIZED
))
3617 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3620 vaddr
= (char *) va
->va_start
;
3621 size
= vm
? get_vm_area_size(vm
) : va_size(va
);
3623 if (addr
>= vaddr
+ size
)
3625 while (addr
< vaddr
) {
3633 n
= vaddr
+ size
- addr
;
3637 if (flags
& VMAP_RAM
)
3638 vmap_ram_vread(buf
, addr
, n
, flags
);
3639 else if (!(vm
->flags
& VM_IOREMAP
))
3640 aligned_vread(buf
, addr
, n
);
3641 else /* IOREMAP area is treated as memory hole */
3648 spin_unlock(&vmap_area_lock
);
3650 if (buf
== buf_start
)
3652 /* zero-fill memory holes */
3653 if (buf
!= buf_start
+ buflen
)
3654 memset(buf
, 0, buflen
- (buf
- buf_start
));
3660 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3661 * @vma: vma to cover
3662 * @uaddr: target user address to start at
3663 * @kaddr: virtual address of vmalloc kernel memory
3664 * @pgoff: offset from @kaddr to start at
3665 * @size: size of map area
3667 * Returns: 0 for success, -Exxx on failure
3669 * This function checks that @kaddr is a valid vmalloc'ed area,
3670 * and that it is big enough to cover the range starting at
3671 * @uaddr in @vma. Will return failure if that criteria isn't
3674 * Similar to remap_pfn_range() (see mm/memory.c)
3676 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3677 void *kaddr
, unsigned long pgoff
,
3680 struct vm_struct
*area
;
3682 unsigned long end_index
;
3684 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
3687 size
= PAGE_ALIGN(size
);
3689 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3692 area
= find_vm_area(kaddr
);
3696 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3699 if (check_add_overflow(size
, off
, &end_index
) ||
3700 end_index
> get_vm_area_size(area
))
3705 struct page
*page
= vmalloc_to_page(kaddr
);
3708 ret
= vm_insert_page(vma
, uaddr
, page
);
3717 vm_flags_set(vma
, VM_DONTEXPAND
| VM_DONTDUMP
);
3723 * remap_vmalloc_range - map vmalloc pages to userspace
3724 * @vma: vma to cover (map full range of vma)
3725 * @addr: vmalloc memory
3726 * @pgoff: number of pages into addr before first page to map
3728 * Returns: 0 for success, -Exxx on failure
3730 * This function checks that addr is a valid vmalloc'ed area, and
3731 * that it is big enough to cover the vma. Will return failure if
3732 * that criteria isn't met.
3734 * Similar to remap_pfn_range() (see mm/memory.c)
3736 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3737 unsigned long pgoff
)
3739 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3741 vma
->vm_end
- vma
->vm_start
);
3743 EXPORT_SYMBOL(remap_vmalloc_range
);
3745 void free_vm_area(struct vm_struct
*area
)
3747 struct vm_struct
*ret
;
3748 ret
= remove_vm_area(area
->addr
);
3749 BUG_ON(ret
!= area
);
3752 EXPORT_SYMBOL_GPL(free_vm_area
);
3755 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3757 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3761 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3762 * @addr: target address
3764 * Returns: vmap_area if it is found. If there is no such area
3765 * the first highest(reverse order) vmap_area is returned
3766 * i.e. va->va_start < addr && va->va_end < addr or NULL
3767 * if there are no any areas before @addr.
3769 static struct vmap_area
*
3770 pvm_find_va_enclose_addr(unsigned long addr
)
3772 struct vmap_area
*va
, *tmp
;
3775 n
= free_vmap_area_root
.rb_node
;
3779 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3780 if (tmp
->va_start
<= addr
) {
3782 if (tmp
->va_end
>= addr
)
3795 * pvm_determine_end_from_reverse - find the highest aligned address
3796 * of free block below VMALLOC_END
3798 * in - the VA we start the search(reverse order);
3799 * out - the VA with the highest aligned end address.
3800 * @align: alignment for required highest address
3802 * Returns: determined end address within vmap_area
3804 static unsigned long
3805 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3807 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3811 list_for_each_entry_from_reverse((*va
),
3812 &free_vmap_area_list
, list
) {
3813 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3814 if ((*va
)->va_start
< addr
)
3823 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3824 * @offsets: array containing offset of each area
3825 * @sizes: array containing size of each area
3826 * @nr_vms: the number of areas to allocate
3827 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3829 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3830 * vm_structs on success, %NULL on failure
3832 * Percpu allocator wants to use congruent vm areas so that it can
3833 * maintain the offsets among percpu areas. This function allocates
3834 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3835 * be scattered pretty far, distance between two areas easily going up
3836 * to gigabytes. To avoid interacting with regular vmallocs, these
3837 * areas are allocated from top.
3839 * Despite its complicated look, this allocator is rather simple. It
3840 * does everything top-down and scans free blocks from the end looking
3841 * for matching base. While scanning, if any of the areas do not fit the
3842 * base address is pulled down to fit the area. Scanning is repeated till
3843 * all the areas fit and then all necessary data structures are inserted
3844 * and the result is returned.
3846 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3847 const size_t *sizes
, int nr_vms
,
3850 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3851 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3852 struct vmap_area
**vas
, *va
;
3853 struct vm_struct
**vms
;
3854 int area
, area2
, last_area
, term_area
;
3855 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3856 bool purged
= false;
3858 /* verify parameters and allocate data structures */
3859 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3860 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3861 start
= offsets
[area
];
3862 end
= start
+ sizes
[area
];
3864 /* is everything aligned properly? */
3865 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3866 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3868 /* detect the area with the highest address */
3869 if (start
> offsets
[last_area
])
3872 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3873 unsigned long start2
= offsets
[area2
];
3874 unsigned long end2
= start2
+ sizes
[area2
];
3876 BUG_ON(start2
< end
&& start
< end2
);
3879 last_end
= offsets
[last_area
] + sizes
[last_area
];
3881 if (vmalloc_end
- vmalloc_start
< last_end
) {
3886 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3887 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3891 for (area
= 0; area
< nr_vms
; area
++) {
3892 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3893 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3894 if (!vas
[area
] || !vms
[area
])
3898 spin_lock(&free_vmap_area_lock
);
3900 /* start scanning - we scan from the top, begin with the last area */
3901 area
= term_area
= last_area
;
3902 start
= offsets
[area
];
3903 end
= start
+ sizes
[area
];
3905 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3906 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3910 * base might have underflowed, add last_end before
3913 if (base
+ last_end
< vmalloc_start
+ last_end
)
3917 * Fitting base has not been found.
3923 * If required width exceeds current VA block, move
3924 * base downwards and then recheck.
3926 if (base
+ end
> va
->va_end
) {
3927 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3933 * If this VA does not fit, move base downwards and recheck.
3935 if (base
+ start
< va
->va_start
) {
3936 va
= node_to_va(rb_prev(&va
->rb_node
));
3937 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3943 * This area fits, move on to the previous one. If
3944 * the previous one is the terminal one, we're done.
3946 area
= (area
+ nr_vms
- 1) % nr_vms
;
3947 if (area
== term_area
)
3950 start
= offsets
[area
];
3951 end
= start
+ sizes
[area
];
3952 va
= pvm_find_va_enclose_addr(base
+ end
);
3955 /* we've found a fitting base, insert all va's */
3956 for (area
= 0; area
< nr_vms
; area
++) {
3959 start
= base
+ offsets
[area
];
3962 va
= pvm_find_va_enclose_addr(start
);
3963 if (WARN_ON_ONCE(va
== NULL
))
3964 /* It is a BUG(), but trigger recovery instead. */
3967 ret
= adjust_va_to_fit_type(&free_vmap_area_root
,
3968 &free_vmap_area_list
,
3970 if (WARN_ON_ONCE(unlikely(ret
)))
3971 /* It is a BUG(), but trigger recovery instead. */
3974 /* Allocated area. */
3976 va
->va_start
= start
;
3977 va
->va_end
= start
+ size
;
3980 spin_unlock(&free_vmap_area_lock
);
3982 /* populate the kasan shadow space */
3983 for (area
= 0; area
< nr_vms
; area
++) {
3984 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3985 goto err_free_shadow
;
3988 /* insert all vm's */
3989 spin_lock(&vmap_area_lock
);
3990 for (area
= 0; area
< nr_vms
; area
++) {
3991 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3993 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3996 spin_unlock(&vmap_area_lock
);
3999 * Mark allocated areas as accessible. Do it now as a best-effort
4000 * approach, as they can be mapped outside of vmalloc code.
4001 * With hardware tag-based KASAN, marking is skipped for
4002 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4004 for (area
= 0; area
< nr_vms
; area
++)
4005 vms
[area
]->addr
= kasan_unpoison_vmalloc(vms
[area
]->addr
,
4006 vms
[area
]->size
, KASAN_VMALLOC_PROT_NORMAL
);
4013 * Remove previously allocated areas. There is no
4014 * need in removing these areas from the busy tree,
4015 * because they are inserted only on the final step
4016 * and when pcpu_get_vm_areas() is success.
4019 orig_start
= vas
[area
]->va_start
;
4020 orig_end
= vas
[area
]->va_end
;
4021 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4022 &free_vmap_area_list
);
4024 kasan_release_vmalloc(orig_start
, orig_end
,
4025 va
->va_start
, va
->va_end
);
4030 spin_unlock(&free_vmap_area_lock
);
4032 purge_vmap_area_lazy();
4035 /* Before "retry", check if we recover. */
4036 for (area
= 0; area
< nr_vms
; area
++) {
4040 vas
[area
] = kmem_cache_zalloc(
4041 vmap_area_cachep
, GFP_KERNEL
);
4050 for (area
= 0; area
< nr_vms
; area
++) {
4052 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
4062 spin_lock(&free_vmap_area_lock
);
4064 * We release all the vmalloc shadows, even the ones for regions that
4065 * hadn't been successfully added. This relies on kasan_release_vmalloc
4066 * being able to tolerate this case.
4068 for (area
= 0; area
< nr_vms
; area
++) {
4069 orig_start
= vas
[area
]->va_start
;
4070 orig_end
= vas
[area
]->va_end
;
4071 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4072 &free_vmap_area_list
);
4074 kasan_release_vmalloc(orig_start
, orig_end
,
4075 va
->va_start
, va
->va_end
);
4079 spin_unlock(&free_vmap_area_lock
);
4086 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4087 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4088 * @nr_vms: the number of allocated areas
4090 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4092 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
4096 for (i
= 0; i
< nr_vms
; i
++)
4097 free_vm_area(vms
[i
]);
4100 #endif /* CONFIG_SMP */
4102 #ifdef CONFIG_PRINTK
4103 bool vmalloc_dump_obj(void *object
)
4105 struct vm_struct
*vm
;
4106 void *objp
= (void *)PAGE_ALIGN((unsigned long)object
);
4108 vm
= find_vm_area(objp
);
4111 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4112 vm
->nr_pages
, (unsigned long)vm
->addr
, vm
->caller
);
4117 #ifdef CONFIG_PROC_FS
4118 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4119 __acquires(&vmap_purge_lock
)
4120 __acquires(&vmap_area_lock
)
4122 mutex_lock(&vmap_purge_lock
);
4123 spin_lock(&vmap_area_lock
);
4125 return seq_list_start(&vmap_area_list
, *pos
);
4128 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4130 return seq_list_next(p
, &vmap_area_list
, pos
);
4133 static void s_stop(struct seq_file
*m
, void *p
)
4134 __releases(&vmap_area_lock
)
4135 __releases(&vmap_purge_lock
)
4137 spin_unlock(&vmap_area_lock
);
4138 mutex_unlock(&vmap_purge_lock
);
4141 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
4143 if (IS_ENABLED(CONFIG_NUMA
)) {
4144 unsigned int nr
, *counters
= m
->private;
4145 unsigned int step
= 1U << vm_area_page_order(v
);
4150 if (v
->flags
& VM_UNINITIALIZED
)
4152 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4155 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
4157 for (nr
= 0; nr
< v
->nr_pages
; nr
+= step
)
4158 counters
[page_to_nid(v
->pages
[nr
])] += step
;
4159 for_each_node_state(nr
, N_HIGH_MEMORY
)
4161 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
4165 static void show_purge_info(struct seq_file
*m
)
4167 struct vmap_area
*va
;
4169 spin_lock(&purge_vmap_area_lock
);
4170 list_for_each_entry(va
, &purge_vmap_area_list
, list
) {
4171 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4172 (void *)va
->va_start
, (void *)va
->va_end
,
4173 va
->va_end
- va
->va_start
);
4175 spin_unlock(&purge_vmap_area_lock
);
4178 static int s_show(struct seq_file
*m
, void *p
)
4180 struct vmap_area
*va
;
4181 struct vm_struct
*v
;
4183 va
= list_entry(p
, struct vmap_area
, list
);
4186 if (va
->flags
& VMAP_RAM
)
4187 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
4188 (void *)va
->va_start
, (void *)va
->va_end
,
4189 va
->va_end
- va
->va_start
);
4196 seq_printf(m
, "0x%pK-0x%pK %7ld",
4197 v
->addr
, v
->addr
+ v
->size
, v
->size
);
4200 seq_printf(m
, " %pS", v
->caller
);
4203 seq_printf(m
, " pages=%d", v
->nr_pages
);
4206 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
4208 if (v
->flags
& VM_IOREMAP
)
4209 seq_puts(m
, " ioremap");
4211 if (v
->flags
& VM_ALLOC
)
4212 seq_puts(m
, " vmalloc");
4214 if (v
->flags
& VM_MAP
)
4215 seq_puts(m
, " vmap");
4217 if (v
->flags
& VM_USERMAP
)
4218 seq_puts(m
, " user");
4220 if (v
->flags
& VM_DMA_COHERENT
)
4221 seq_puts(m
, " dma-coherent");
4223 if (is_vmalloc_addr(v
->pages
))
4224 seq_puts(m
, " vpages");
4226 show_numa_info(m
, v
);
4230 * As a final step, dump "unpurged" areas.
4233 if (list_is_last(&va
->list
, &vmap_area_list
))
4239 static const struct seq_operations vmalloc_op
= {
4246 static int __init
proc_vmalloc_init(void)
4248 if (IS_ENABLED(CONFIG_NUMA
))
4249 proc_create_seq_private("vmallocinfo", 0400, NULL
,
4251 nr_node_ids
* sizeof(unsigned int), NULL
);
4253 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
4256 module_init(proc_vmalloc_init
);
4260 void __init
vmalloc_init(void)
4262 struct vmap_area
*va
;
4263 struct vm_struct
*tmp
;
4267 * Create the cache for vmap_area objects.
4269 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
4271 for_each_possible_cpu(i
) {
4272 struct vmap_block_queue
*vbq
;
4273 struct vfree_deferred
*p
;
4275 vbq
= &per_cpu(vmap_block_queue
, i
);
4276 spin_lock_init(&vbq
->lock
);
4277 INIT_LIST_HEAD(&vbq
->free
);
4278 p
= &per_cpu(vfree_deferred
, i
);
4279 init_llist_head(&p
->list
);
4280 INIT_WORK(&p
->wq
, delayed_vfree_work
);
4283 /* Import existing vmlist entries. */
4284 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
4285 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
4286 if (WARN_ON_ONCE(!va
))
4289 va
->va_start
= (unsigned long)tmp
->addr
;
4290 va
->va_end
= va
->va_start
+ tmp
->size
;
4292 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
4296 * Now we can initialize a free vmap space.
4298 vmap_init_free_space();
4299 vmap_initialized
= true;