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/uio.h>
37 #include <linux/bitops.h>
38 #include <linux/rbtree_augmented.h>
39 #include <linux/overflow.h>
40 #include <linux/pgtable.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(ptep_get(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(ptep_get(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_kernel(pmd
, addr
);
707 pte
= ptep_get(ptep
);
708 if (pte_present(pte
))
709 page
= pte_page(pte
);
713 EXPORT_SYMBOL(vmalloc_to_page
);
716 * Map a vmalloc()-space virtual address to the physical page frame number.
718 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
720 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
722 EXPORT_SYMBOL(vmalloc_to_pfn
);
725 /*** Global kva allocator ***/
727 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
728 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
731 static DEFINE_SPINLOCK(vmap_area_lock
);
732 static DEFINE_SPINLOCK(free_vmap_area_lock
);
733 /* Export for kexec only */
734 LIST_HEAD(vmap_area_list
);
735 static struct rb_root vmap_area_root
= RB_ROOT
;
736 static bool vmap_initialized __read_mostly
;
738 static struct rb_root purge_vmap_area_root
= RB_ROOT
;
739 static LIST_HEAD(purge_vmap_area_list
);
740 static DEFINE_SPINLOCK(purge_vmap_area_lock
);
743 * This kmem_cache is used for vmap_area objects. Instead of
744 * allocating from slab we reuse an object from this cache to
745 * make things faster. Especially in "no edge" splitting of
748 static struct kmem_cache
*vmap_area_cachep
;
751 * This linked list is used in pair with free_vmap_area_root.
752 * It gives O(1) access to prev/next to perform fast coalescing.
754 static LIST_HEAD(free_vmap_area_list
);
757 * This augment red-black tree represents the free vmap space.
758 * All vmap_area objects in this tree are sorted by va->va_start
759 * address. It is used for allocation and merging when a vmap
760 * object is released.
762 * Each vmap_area node contains a maximum available free block
763 * of its sub-tree, right or left. Therefore it is possible to
764 * find a lowest match of free area.
766 static struct rb_root free_vmap_area_root
= RB_ROOT
;
769 * Preload a CPU with one object for "no edge" split case. The
770 * aim is to get rid of allocations from the atomic context, thus
771 * to use more permissive allocation masks.
773 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
775 static __always_inline
unsigned long
776 va_size(struct vmap_area
*va
)
778 return (va
->va_end
- va
->va_start
);
781 static __always_inline
unsigned long
782 get_subtree_max_size(struct rb_node
*node
)
784 struct vmap_area
*va
;
786 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
787 return va
? va
->subtree_max_size
: 0;
790 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
791 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
793 static void reclaim_and_purge_vmap_areas(void);
794 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
795 static void drain_vmap_area_work(struct work_struct
*work
);
796 static DECLARE_WORK(drain_vmap_work
, drain_vmap_area_work
);
798 static atomic_long_t nr_vmalloc_pages
;
800 unsigned long vmalloc_nr_pages(void)
802 return atomic_long_read(&nr_vmalloc_pages
);
805 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
806 static struct vmap_area
*find_vmap_area_exceed_addr(unsigned long addr
)
808 struct vmap_area
*va
= NULL
;
809 struct rb_node
*n
= vmap_area_root
.rb_node
;
811 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
814 struct vmap_area
*tmp
;
816 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
817 if (tmp
->va_end
> addr
) {
819 if (tmp
->va_start
<= addr
)
830 static struct vmap_area
*__find_vmap_area(unsigned long addr
, struct rb_root
*root
)
832 struct rb_node
*n
= root
->rb_node
;
834 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
837 struct vmap_area
*va
;
839 va
= rb_entry(n
, struct vmap_area
, rb_node
);
840 if (addr
< va
->va_start
)
842 else if (addr
>= va
->va_end
)
852 * This function returns back addresses of parent node
853 * and its left or right link for further processing.
855 * Otherwise NULL is returned. In that case all further
856 * steps regarding inserting of conflicting overlap range
857 * have to be declined and actually considered as a bug.
859 static __always_inline
struct rb_node
**
860 find_va_links(struct vmap_area
*va
,
861 struct rb_root
*root
, struct rb_node
*from
,
862 struct rb_node
**parent
)
864 struct vmap_area
*tmp_va
;
865 struct rb_node
**link
;
868 link
= &root
->rb_node
;
869 if (unlikely(!*link
)) {
878 * Go to the bottom of the tree. When we hit the last point
879 * we end up with parent rb_node and correct direction, i name
880 * it link, where the new va->rb_node will be attached to.
883 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
886 * During the traversal we also do some sanity check.
887 * Trigger the BUG() if there are sides(left/right)
890 if (va
->va_end
<= tmp_va
->va_start
)
891 link
= &(*link
)->rb_left
;
892 else if (va
->va_start
>= tmp_va
->va_end
)
893 link
= &(*link
)->rb_right
;
895 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
896 va
->va_start
, va
->va_end
, tmp_va
->va_start
, tmp_va
->va_end
);
902 *parent
= &tmp_va
->rb_node
;
906 static __always_inline
struct list_head
*
907 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
909 struct list_head
*list
;
911 if (unlikely(!parent
))
913 * The red-black tree where we try to find VA neighbors
914 * before merging or inserting is empty, i.e. it means
915 * there is no free vmap space. Normally it does not
916 * happen but we handle this case anyway.
920 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
921 return (&parent
->rb_right
== link
? list
->next
: list
);
924 static __always_inline
void
925 __link_va(struct vmap_area
*va
, struct rb_root
*root
,
926 struct rb_node
*parent
, struct rb_node
**link
,
927 struct list_head
*head
, bool augment
)
930 * VA is still not in the list, but we can
931 * identify its future previous list_head node.
933 if (likely(parent
)) {
934 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
935 if (&parent
->rb_right
!= link
)
939 /* Insert to the rb-tree */
940 rb_link_node(&va
->rb_node
, parent
, link
);
943 * Some explanation here. Just perform simple insertion
944 * to the tree. We do not set va->subtree_max_size to
945 * its current size before calling rb_insert_augmented().
946 * It is because we populate the tree from the bottom
947 * to parent levels when the node _is_ in the tree.
949 * Therefore we set subtree_max_size to zero after insertion,
950 * to let __augment_tree_propagate_from() puts everything to
951 * the correct order later on.
953 rb_insert_augmented(&va
->rb_node
,
954 root
, &free_vmap_area_rb_augment_cb
);
955 va
->subtree_max_size
= 0;
957 rb_insert_color(&va
->rb_node
, root
);
960 /* Address-sort this list */
961 list_add(&va
->list
, head
);
964 static __always_inline
void
965 link_va(struct vmap_area
*va
, struct rb_root
*root
,
966 struct rb_node
*parent
, struct rb_node
**link
,
967 struct list_head
*head
)
969 __link_va(va
, root
, parent
, link
, head
, false);
972 static __always_inline
void
973 link_va_augment(struct vmap_area
*va
, struct rb_root
*root
,
974 struct rb_node
*parent
, struct rb_node
**link
,
975 struct list_head
*head
)
977 __link_va(va
, root
, parent
, link
, head
, true);
980 static __always_inline
void
981 __unlink_va(struct vmap_area
*va
, struct rb_root
*root
, bool augment
)
983 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
987 rb_erase_augmented(&va
->rb_node
,
988 root
, &free_vmap_area_rb_augment_cb
);
990 rb_erase(&va
->rb_node
, root
);
992 list_del_init(&va
->list
);
993 RB_CLEAR_NODE(&va
->rb_node
);
996 static __always_inline
void
997 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
999 __unlink_va(va
, root
, false);
1002 static __always_inline
void
1003 unlink_va_augment(struct vmap_area
*va
, struct rb_root
*root
)
1005 __unlink_va(va
, root
, true);
1008 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1010 * Gets called when remove the node and rotate.
1012 static __always_inline
unsigned long
1013 compute_subtree_max_size(struct vmap_area
*va
)
1015 return max3(va_size(va
),
1016 get_subtree_max_size(va
->rb_node
.rb_left
),
1017 get_subtree_max_size(va
->rb_node
.rb_right
));
1021 augment_tree_propagate_check(void)
1023 struct vmap_area
*va
;
1024 unsigned long computed_size
;
1026 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
1027 computed_size
= compute_subtree_max_size(va
);
1028 if (computed_size
!= va
->subtree_max_size
)
1029 pr_emerg("tree is corrupted: %lu, %lu\n",
1030 va_size(va
), va
->subtree_max_size
);
1036 * This function populates subtree_max_size from bottom to upper
1037 * levels starting from VA point. The propagation must be done
1038 * when VA size is modified by changing its va_start/va_end. Or
1039 * in case of newly inserting of VA to the tree.
1041 * It means that __augment_tree_propagate_from() must be called:
1042 * - After VA has been inserted to the tree(free path);
1043 * - After VA has been shrunk(allocation path);
1044 * - After VA has been increased(merging path).
1046 * Please note that, it does not mean that upper parent nodes
1047 * and their subtree_max_size are recalculated all the time up
1056 * For example if we modify the node 4, shrinking it to 2, then
1057 * no any modification is required. If we shrink the node 2 to 1
1058 * its subtree_max_size is updated only, and set to 1. If we shrink
1059 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1060 * node becomes 4--6.
1062 static __always_inline
void
1063 augment_tree_propagate_from(struct vmap_area
*va
)
1066 * Populate the tree from bottom towards the root until
1067 * the calculated maximum available size of checked node
1068 * is equal to its current one.
1070 free_vmap_area_rb_augment_cb_propagate(&va
->rb_node
, NULL
);
1072 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1073 augment_tree_propagate_check();
1078 insert_vmap_area(struct vmap_area
*va
,
1079 struct rb_root
*root
, struct list_head
*head
)
1081 struct rb_node
**link
;
1082 struct rb_node
*parent
;
1084 link
= find_va_links(va
, root
, NULL
, &parent
);
1086 link_va(va
, root
, parent
, link
, head
);
1090 insert_vmap_area_augment(struct vmap_area
*va
,
1091 struct rb_node
*from
, struct rb_root
*root
,
1092 struct list_head
*head
)
1094 struct rb_node
**link
;
1095 struct rb_node
*parent
;
1098 link
= find_va_links(va
, NULL
, from
, &parent
);
1100 link
= find_va_links(va
, root
, NULL
, &parent
);
1103 link_va_augment(va
, root
, parent
, link
, head
);
1104 augment_tree_propagate_from(va
);
1109 * Merge de-allocated chunk of VA memory with previous
1110 * and next free blocks. If coalesce is not done a new
1111 * free area is inserted. If VA has been merged, it is
1114 * Please note, it can return NULL in case of overlap
1115 * ranges, followed by WARN() report. Despite it is a
1116 * buggy behaviour, a system can be alive and keep
1119 static __always_inline
struct vmap_area
*
1120 __merge_or_add_vmap_area(struct vmap_area
*va
,
1121 struct rb_root
*root
, struct list_head
*head
, bool augment
)
1123 struct vmap_area
*sibling
;
1124 struct list_head
*next
;
1125 struct rb_node
**link
;
1126 struct rb_node
*parent
;
1127 bool merged
= false;
1130 * Find a place in the tree where VA potentially will be
1131 * inserted, unless it is merged with its sibling/siblings.
1133 link
= find_va_links(va
, root
, NULL
, &parent
);
1138 * Get next node of VA to check if merging can be done.
1140 next
= get_va_next_sibling(parent
, link
);
1141 if (unlikely(next
== NULL
))
1147 * |<------VA------>|<-----Next----->|
1152 sibling
= list_entry(next
, struct vmap_area
, list
);
1153 if (sibling
->va_start
== va
->va_end
) {
1154 sibling
->va_start
= va
->va_start
;
1156 /* Free vmap_area object. */
1157 kmem_cache_free(vmap_area_cachep
, va
);
1159 /* Point to the new merged area. */
1168 * |<-----Prev----->|<------VA------>|
1172 if (next
->prev
!= head
) {
1173 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
1174 if (sibling
->va_end
== va
->va_start
) {
1176 * If both neighbors are coalesced, it is important
1177 * to unlink the "next" node first, followed by merging
1178 * with "previous" one. Otherwise the tree might not be
1179 * fully populated if a sibling's augmented value is
1180 * "normalized" because of rotation operations.
1183 __unlink_va(va
, root
, augment
);
1185 sibling
->va_end
= va
->va_end
;
1187 /* Free vmap_area object. */
1188 kmem_cache_free(vmap_area_cachep
, va
);
1190 /* Point to the new merged area. */
1198 __link_va(va
, root
, parent
, link
, head
, augment
);
1203 static __always_inline
struct vmap_area
*
1204 merge_or_add_vmap_area(struct vmap_area
*va
,
1205 struct rb_root
*root
, struct list_head
*head
)
1207 return __merge_or_add_vmap_area(va
, root
, head
, false);
1210 static __always_inline
struct vmap_area
*
1211 merge_or_add_vmap_area_augment(struct vmap_area
*va
,
1212 struct rb_root
*root
, struct list_head
*head
)
1214 va
= __merge_or_add_vmap_area(va
, root
, head
, true);
1216 augment_tree_propagate_from(va
);
1221 static __always_inline
bool
1222 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
1223 unsigned long align
, unsigned long vstart
)
1225 unsigned long nva_start_addr
;
1227 if (va
->va_start
> vstart
)
1228 nva_start_addr
= ALIGN(va
->va_start
, align
);
1230 nva_start_addr
= ALIGN(vstart
, align
);
1232 /* Can be overflowed due to big size or alignment. */
1233 if (nva_start_addr
+ size
< nva_start_addr
||
1234 nva_start_addr
< vstart
)
1237 return (nva_start_addr
+ size
<= va
->va_end
);
1241 * Find the first free block(lowest start address) in the tree,
1242 * that will accomplish the request corresponding to passing
1243 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1244 * a search length is adjusted to account for worst case alignment
1247 static __always_inline
struct vmap_area
*
1248 find_vmap_lowest_match(struct rb_root
*root
, unsigned long size
,
1249 unsigned long align
, unsigned long vstart
, bool adjust_search_size
)
1251 struct vmap_area
*va
;
1252 struct rb_node
*node
;
1253 unsigned long length
;
1255 /* Start from the root. */
1256 node
= root
->rb_node
;
1258 /* Adjust the search size for alignment overhead. */
1259 length
= adjust_search_size
? size
+ align
- 1 : size
;
1262 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1264 if (get_subtree_max_size(node
->rb_left
) >= length
&&
1265 vstart
< va
->va_start
) {
1266 node
= node
->rb_left
;
1268 if (is_within_this_va(va
, size
, align
, vstart
))
1272 * Does not make sense to go deeper towards the right
1273 * sub-tree if it does not have a free block that is
1274 * equal or bigger to the requested search length.
1276 if (get_subtree_max_size(node
->rb_right
) >= length
) {
1277 node
= node
->rb_right
;
1282 * OK. We roll back and find the first right sub-tree,
1283 * that will satisfy the search criteria. It can happen
1284 * due to "vstart" restriction or an alignment overhead
1285 * that is bigger then PAGE_SIZE.
1287 while ((node
= rb_parent(node
))) {
1288 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1289 if (is_within_this_va(va
, size
, align
, vstart
))
1292 if (get_subtree_max_size(node
->rb_right
) >= length
&&
1293 vstart
<= va
->va_start
) {
1295 * Shift the vstart forward. Please note, we update it with
1296 * parent's start address adding "1" because we do not want
1297 * to enter same sub-tree after it has already been checked
1298 * and no suitable free block found there.
1300 vstart
= va
->va_start
+ 1;
1301 node
= node
->rb_right
;
1311 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1312 #include <linux/random.h>
1314 static struct vmap_area
*
1315 find_vmap_lowest_linear_match(struct list_head
*head
, unsigned long size
,
1316 unsigned long align
, unsigned long vstart
)
1318 struct vmap_area
*va
;
1320 list_for_each_entry(va
, head
, list
) {
1321 if (!is_within_this_va(va
, size
, align
, vstart
))
1331 find_vmap_lowest_match_check(struct rb_root
*root
, struct list_head
*head
,
1332 unsigned long size
, unsigned long align
)
1334 struct vmap_area
*va_1
, *va_2
;
1335 unsigned long vstart
;
1338 get_random_bytes(&rnd
, sizeof(rnd
));
1339 vstart
= VMALLOC_START
+ rnd
;
1341 va_1
= find_vmap_lowest_match(root
, size
, align
, vstart
, false);
1342 va_2
= find_vmap_lowest_linear_match(head
, size
, align
, vstart
);
1345 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1346 va_1
, va_2
, vstart
);
1352 FL_FIT_TYPE
= 1, /* full fit */
1353 LE_FIT_TYPE
= 2, /* left edge fit */
1354 RE_FIT_TYPE
= 3, /* right edge fit */
1355 NE_FIT_TYPE
= 4 /* no edge fit */
1358 static __always_inline
enum fit_type
1359 classify_va_fit_type(struct vmap_area
*va
,
1360 unsigned long nva_start_addr
, unsigned long size
)
1364 /* Check if it is within VA. */
1365 if (nva_start_addr
< va
->va_start
||
1366 nva_start_addr
+ size
> va
->va_end
)
1370 if (va
->va_start
== nva_start_addr
) {
1371 if (va
->va_end
== nva_start_addr
+ size
)
1375 } else if (va
->va_end
== nva_start_addr
+ size
) {
1384 static __always_inline
int
1385 adjust_va_to_fit_type(struct rb_root
*root
, struct list_head
*head
,
1386 struct vmap_area
*va
, unsigned long nva_start_addr
,
1389 struct vmap_area
*lva
= NULL
;
1390 enum fit_type type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1392 if (type
== FL_FIT_TYPE
) {
1394 * No need to split VA, it fully fits.
1400 unlink_va_augment(va
, root
);
1401 kmem_cache_free(vmap_area_cachep
, va
);
1402 } else if (type
== LE_FIT_TYPE
) {
1404 * Split left edge of fit VA.
1410 va
->va_start
+= size
;
1411 } else if (type
== RE_FIT_TYPE
) {
1413 * Split right edge of fit VA.
1419 va
->va_end
= nva_start_addr
;
1420 } else if (type
== NE_FIT_TYPE
) {
1422 * Split no edge of fit VA.
1428 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1429 if (unlikely(!lva
)) {
1431 * For percpu allocator we do not do any pre-allocation
1432 * and leave it as it is. The reason is it most likely
1433 * never ends up with NE_FIT_TYPE splitting. In case of
1434 * percpu allocations offsets and sizes are aligned to
1435 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1436 * are its main fitting cases.
1438 * There are a few exceptions though, as an example it is
1439 * a first allocation (early boot up) when we have "one"
1440 * big free space that has to be split.
1442 * Also we can hit this path in case of regular "vmap"
1443 * allocations, if "this" current CPU was not preloaded.
1444 * See the comment in alloc_vmap_area() why. If so, then
1445 * GFP_NOWAIT is used instead to get an extra object for
1446 * split purpose. That is rare and most time does not
1449 * What happens if an allocation gets failed. Basically,
1450 * an "overflow" path is triggered to purge lazily freed
1451 * areas to free some memory, then, the "retry" path is
1452 * triggered to repeat one more time. See more details
1453 * in alloc_vmap_area() function.
1455 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1461 * Build the remainder.
1463 lva
->va_start
= va
->va_start
;
1464 lva
->va_end
= nva_start_addr
;
1467 * Shrink this VA to remaining size.
1469 va
->va_start
= nva_start_addr
+ size
;
1474 if (type
!= FL_FIT_TYPE
) {
1475 augment_tree_propagate_from(va
);
1477 if (lva
) /* type == NE_FIT_TYPE */
1478 insert_vmap_area_augment(lva
, &va
->rb_node
, root
, head
);
1485 * Returns a start address of the newly allocated area, if success.
1486 * Otherwise a vend is returned that indicates failure.
1488 static __always_inline
unsigned long
1489 __alloc_vmap_area(struct rb_root
*root
, struct list_head
*head
,
1490 unsigned long size
, unsigned long align
,
1491 unsigned long vstart
, unsigned long vend
)
1493 bool adjust_search_size
= true;
1494 unsigned long nva_start_addr
;
1495 struct vmap_area
*va
;
1499 * Do not adjust when:
1500 * a) align <= PAGE_SIZE, because it does not make any sense.
1501 * All blocks(their start addresses) are at least PAGE_SIZE
1503 * b) a short range where a requested size corresponds to exactly
1504 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1505 * With adjusted search length an allocation would not succeed.
1507 if (align
<= PAGE_SIZE
|| (align
> PAGE_SIZE
&& (vend
- vstart
) == size
))
1508 adjust_search_size
= false;
1510 va
= find_vmap_lowest_match(root
, size
, align
, vstart
, adjust_search_size
);
1514 if (va
->va_start
> vstart
)
1515 nva_start_addr
= ALIGN(va
->va_start
, align
);
1517 nva_start_addr
= ALIGN(vstart
, align
);
1519 /* Check the "vend" restriction. */
1520 if (nva_start_addr
+ size
> vend
)
1523 /* Update the free vmap_area. */
1524 ret
= adjust_va_to_fit_type(root
, head
, va
, nva_start_addr
, size
);
1525 if (WARN_ON_ONCE(ret
))
1528 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1529 find_vmap_lowest_match_check(root
, head
, size
, align
);
1532 return nva_start_addr
;
1536 * Free a region of KVA allocated by alloc_vmap_area
1538 static void free_vmap_area(struct vmap_area
*va
)
1541 * Remove from the busy tree/list.
1543 spin_lock(&vmap_area_lock
);
1544 unlink_va(va
, &vmap_area_root
);
1545 spin_unlock(&vmap_area_lock
);
1548 * Insert/Merge it back to the free tree/list.
1550 spin_lock(&free_vmap_area_lock
);
1551 merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1552 spin_unlock(&free_vmap_area_lock
);
1556 preload_this_cpu_lock(spinlock_t
*lock
, gfp_t gfp_mask
, int node
)
1558 struct vmap_area
*va
= NULL
;
1561 * Preload this CPU with one extra vmap_area object. It is used
1562 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1563 * a CPU that does an allocation is preloaded.
1565 * We do it in non-atomic context, thus it allows us to use more
1566 * permissive allocation masks to be more stable under low memory
1567 * condition and high memory pressure.
1569 if (!this_cpu_read(ne_fit_preload_node
))
1570 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1574 if (va
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, va
))
1575 kmem_cache_free(vmap_area_cachep
, va
);
1579 * Allocate a region of KVA of the specified size and alignment, within the
1582 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1583 unsigned long align
,
1584 unsigned long vstart
, unsigned long vend
,
1585 int node
, gfp_t gfp_mask
,
1586 unsigned long va_flags
)
1588 struct vmap_area
*va
;
1589 unsigned long freed
;
1594 if (unlikely(!size
|| offset_in_page(size
) || !is_power_of_2(align
)))
1595 return ERR_PTR(-EINVAL
);
1597 if (unlikely(!vmap_initialized
))
1598 return ERR_PTR(-EBUSY
);
1601 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1603 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1605 return ERR_PTR(-ENOMEM
);
1608 * Only scan the relevant parts containing pointers to other objects
1609 * to avoid false negatives.
1611 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1614 preload_this_cpu_lock(&free_vmap_area_lock
, gfp_mask
, node
);
1615 addr
= __alloc_vmap_area(&free_vmap_area_root
, &free_vmap_area_list
,
1616 size
, align
, vstart
, vend
);
1617 spin_unlock(&free_vmap_area_lock
);
1619 trace_alloc_vmap_area(addr
, size
, align
, vstart
, vend
, addr
== vend
);
1622 * If an allocation fails, the "vend" address is
1623 * returned. Therefore trigger the overflow path.
1625 if (unlikely(addr
== vend
))
1628 va
->va_start
= addr
;
1629 va
->va_end
= addr
+ size
;
1631 va
->flags
= va_flags
;
1633 spin_lock(&vmap_area_lock
);
1634 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1635 spin_unlock(&vmap_area_lock
);
1637 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1638 BUG_ON(va
->va_start
< vstart
);
1639 BUG_ON(va
->va_end
> vend
);
1641 ret
= kasan_populate_vmalloc(addr
, size
);
1644 return ERR_PTR(ret
);
1651 reclaim_and_purge_vmap_areas();
1657 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1664 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1665 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1668 kmem_cache_free(vmap_area_cachep
, va
);
1669 return ERR_PTR(-EBUSY
);
1672 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1674 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1676 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1678 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1680 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1682 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1685 * lazy_max_pages is the maximum amount of virtual address space we gather up
1686 * before attempting to purge with a TLB flush.
1688 * There is a tradeoff here: a larger number will cover more kernel page tables
1689 * and take slightly longer to purge, but it will linearly reduce the number of
1690 * global TLB flushes that must be performed. It would seem natural to scale
1691 * this number up linearly with the number of CPUs (because vmapping activity
1692 * could also scale linearly with the number of CPUs), however it is likely
1693 * that in practice, workloads might be constrained in other ways that mean
1694 * vmap activity will not scale linearly with CPUs. Also, I want to be
1695 * conservative and not introduce a big latency on huge systems, so go with
1696 * a less aggressive log scale. It will still be an improvement over the old
1697 * code, and it will be simple to change the scale factor if we find that it
1698 * becomes a problem on bigger systems.
1700 static unsigned long lazy_max_pages(void)
1704 log
= fls(num_online_cpus());
1706 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1709 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1712 * Serialize vmap purging. There is no actual critical section protected
1713 * by this lock, but we want to avoid concurrent calls for performance
1714 * reasons and to make the pcpu_get_vm_areas more deterministic.
1716 static DEFINE_MUTEX(vmap_purge_lock
);
1718 /* for per-CPU blocks */
1719 static void purge_fragmented_blocks_allcpus(void);
1722 * Purges all lazily-freed vmap areas.
1724 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1726 unsigned long resched_threshold
;
1727 unsigned int num_purged_areas
= 0;
1728 struct list_head local_purge_list
;
1729 struct vmap_area
*va
, *n_va
;
1731 lockdep_assert_held(&vmap_purge_lock
);
1733 spin_lock(&purge_vmap_area_lock
);
1734 purge_vmap_area_root
= RB_ROOT
;
1735 list_replace_init(&purge_vmap_area_list
, &local_purge_list
);
1736 spin_unlock(&purge_vmap_area_lock
);
1738 if (unlikely(list_empty(&local_purge_list
)))
1742 list_first_entry(&local_purge_list
,
1743 struct vmap_area
, list
)->va_start
);
1746 list_last_entry(&local_purge_list
,
1747 struct vmap_area
, list
)->va_end
);
1749 flush_tlb_kernel_range(start
, end
);
1750 resched_threshold
= lazy_max_pages() << 1;
1752 spin_lock(&free_vmap_area_lock
);
1753 list_for_each_entry_safe(va
, n_va
, &local_purge_list
, list
) {
1754 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1755 unsigned long orig_start
= va
->va_start
;
1756 unsigned long orig_end
= va
->va_end
;
1759 * Finally insert or merge lazily-freed area. It is
1760 * detached and there is no need to "unlink" it from
1763 va
= merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
,
1764 &free_vmap_area_list
);
1769 if (is_vmalloc_or_module_addr((void *)orig_start
))
1770 kasan_release_vmalloc(orig_start
, orig_end
,
1771 va
->va_start
, va
->va_end
);
1773 atomic_long_sub(nr
, &vmap_lazy_nr
);
1776 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1777 cond_resched_lock(&free_vmap_area_lock
);
1779 spin_unlock(&free_vmap_area_lock
);
1782 trace_purge_vmap_area_lazy(start
, end
, num_purged_areas
);
1783 return num_purged_areas
> 0;
1787 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
1789 static void reclaim_and_purge_vmap_areas(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)
1912 * Purge threshold to prevent overeager purging of fragmented blocks for
1913 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
1915 #define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
1917 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
1918 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
1919 #define VMAP_FLAGS_MASK 0x3
1921 struct vmap_block_queue
{
1923 struct list_head free
;
1926 * An xarray requires an extra memory dynamically to
1927 * be allocated. If it is an issue, we can use rb-tree
1930 struct xarray vmap_blocks
;
1935 struct vmap_area
*va
;
1936 unsigned long free
, dirty
;
1937 DECLARE_BITMAP(used_map
, VMAP_BBMAP_BITS
);
1938 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1939 struct list_head free_list
;
1940 struct rcu_head rcu_head
;
1941 struct list_head purge
;
1944 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1945 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1948 * In order to fast access to any "vmap_block" associated with a
1949 * specific address, we use a hash.
1951 * A per-cpu vmap_block_queue is used in both ways, to serialize
1952 * an access to free block chains among CPUs(alloc path) and it
1953 * also acts as a vmap_block hash(alloc/free paths). It means we
1954 * overload it, since we already have the per-cpu array which is
1955 * used as a hash table. When used as a hash a 'cpu' passed to
1956 * per_cpu() is not actually a CPU but rather a hash index.
1958 * A hash function is addr_to_vb_xa() which hashes any address
1959 * to a specific index(in a hash) it belongs to. This then uses a
1960 * per_cpu() macro to access an array with generated index.
1967 * 0 10 20 30 40 50 60
1968 * |------|------|------|------|------|------|...<vmap address space>
1969 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
1971 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
1972 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
1974 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
1975 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
1977 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
1978 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
1980 * This technique almost always avoids lock contention on insert/remove,
1981 * however xarray spinlocks protect against any contention that remains.
1983 static struct xarray
*
1984 addr_to_vb_xa(unsigned long addr
)
1986 int index
= (addr
/ VMAP_BLOCK_SIZE
) % num_possible_cpus();
1988 return &per_cpu(vmap_block_queue
, index
).vmap_blocks
;
1992 * We should probably have a fallback mechanism to allocate virtual memory
1993 * out of partially filled vmap blocks. However vmap block sizing should be
1994 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1998 static unsigned long addr_to_vb_idx(unsigned long addr
)
2000 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
2001 addr
/= VMAP_BLOCK_SIZE
;
2005 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
2009 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
2010 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
2011 return (void *)addr
;
2015 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2016 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2017 * @order: how many 2^order pages should be occupied in newly allocated block
2018 * @gfp_mask: flags for the page level allocator
2020 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2022 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
2024 struct vmap_block_queue
*vbq
;
2025 struct vmap_block
*vb
;
2026 struct vmap_area
*va
;
2028 unsigned long vb_idx
;
2032 node
= numa_node_id();
2034 vb
= kmalloc_node(sizeof(struct vmap_block
),
2035 gfp_mask
& GFP_RECLAIM_MASK
, node
);
2037 return ERR_PTR(-ENOMEM
);
2039 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
2040 VMALLOC_START
, VMALLOC_END
,
2042 VMAP_RAM
|VMAP_BLOCK
);
2045 return ERR_CAST(va
);
2048 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
2049 spin_lock_init(&vb
->lock
);
2051 /* At least something should be left free */
2052 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
2053 bitmap_zero(vb
->used_map
, VMAP_BBMAP_BITS
);
2054 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
2056 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2058 bitmap_set(vb
->used_map
, 0, (1UL << order
));
2059 INIT_LIST_HEAD(&vb
->free_list
);
2061 xa
= addr_to_vb_xa(va
->va_start
);
2062 vb_idx
= addr_to_vb_idx(va
->va_start
);
2063 err
= xa_insert(xa
, vb_idx
, vb
, gfp_mask
);
2067 return ERR_PTR(err
);
2070 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2071 spin_lock(&vbq
->lock
);
2072 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
2073 spin_unlock(&vbq
->lock
);
2078 static void free_vmap_block(struct vmap_block
*vb
)
2080 struct vmap_block
*tmp
;
2083 xa
= addr_to_vb_xa(vb
->va
->va_start
);
2084 tmp
= xa_erase(xa
, addr_to_vb_idx(vb
->va
->va_start
));
2087 spin_lock(&vmap_area_lock
);
2088 unlink_va(vb
->va
, &vmap_area_root
);
2089 spin_unlock(&vmap_area_lock
);
2091 free_vmap_area_noflush(vb
->va
);
2092 kfree_rcu(vb
, rcu_head
);
2095 static bool purge_fragmented_block(struct vmap_block
*vb
,
2096 struct vmap_block_queue
*vbq
, struct list_head
*purge_list
,
2099 if (vb
->free
+ vb
->dirty
!= VMAP_BBMAP_BITS
||
2100 vb
->dirty
== VMAP_BBMAP_BITS
)
2103 /* Don't overeagerly purge usable blocks unless requested */
2104 if (!(force_purge
|| vb
->free
< VMAP_PURGE_THRESHOLD
))
2107 /* prevent further allocs after releasing lock */
2108 WRITE_ONCE(vb
->free
, 0);
2109 /* prevent purging it again */
2110 WRITE_ONCE(vb
->dirty
, VMAP_BBMAP_BITS
);
2112 vb
->dirty_max
= VMAP_BBMAP_BITS
;
2113 spin_lock(&vbq
->lock
);
2114 list_del_rcu(&vb
->free_list
);
2115 spin_unlock(&vbq
->lock
);
2116 list_add_tail(&vb
->purge
, purge_list
);
2120 static void free_purged_blocks(struct list_head
*purge_list
)
2122 struct vmap_block
*vb
, *n_vb
;
2124 list_for_each_entry_safe(vb
, n_vb
, purge_list
, purge
) {
2125 list_del(&vb
->purge
);
2126 free_vmap_block(vb
);
2130 static void purge_fragmented_blocks(int cpu
)
2133 struct vmap_block
*vb
;
2134 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2137 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2138 unsigned long free
= READ_ONCE(vb
->free
);
2139 unsigned long dirty
= READ_ONCE(vb
->dirty
);
2141 if (free
+ dirty
!= VMAP_BBMAP_BITS
||
2142 dirty
== VMAP_BBMAP_BITS
)
2145 spin_lock(&vb
->lock
);
2146 purge_fragmented_block(vb
, vbq
, &purge
, true);
2147 spin_unlock(&vb
->lock
);
2150 free_purged_blocks(&purge
);
2153 static void purge_fragmented_blocks_allcpus(void)
2157 for_each_possible_cpu(cpu
)
2158 purge_fragmented_blocks(cpu
);
2161 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
2163 struct vmap_block_queue
*vbq
;
2164 struct vmap_block
*vb
;
2168 BUG_ON(offset_in_page(size
));
2169 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2170 if (WARN_ON(size
== 0)) {
2172 * Allocating 0 bytes isn't what caller wants since
2173 * get_order(0) returns funny result. Just warn and terminate
2178 order
= get_order(size
);
2181 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2182 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2183 unsigned long pages_off
;
2185 if (READ_ONCE(vb
->free
) < (1UL << order
))
2188 spin_lock(&vb
->lock
);
2189 if (vb
->free
< (1UL << order
)) {
2190 spin_unlock(&vb
->lock
);
2194 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
2195 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
2196 WRITE_ONCE(vb
->free
, vb
->free
- (1UL << order
));
2197 bitmap_set(vb
->used_map
, pages_off
, (1UL << order
));
2198 if (vb
->free
== 0) {
2199 spin_lock(&vbq
->lock
);
2200 list_del_rcu(&vb
->free_list
);
2201 spin_unlock(&vbq
->lock
);
2204 spin_unlock(&vb
->lock
);
2210 /* Allocate new block if nothing was found */
2212 vaddr
= new_vmap_block(order
, gfp_mask
);
2217 static void vb_free(unsigned long addr
, unsigned long size
)
2219 unsigned long offset
;
2221 struct vmap_block
*vb
;
2224 BUG_ON(offset_in_page(size
));
2225 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2227 flush_cache_vunmap(addr
, addr
+ size
);
2229 order
= get_order(size
);
2230 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
2232 xa
= addr_to_vb_xa(addr
);
2233 vb
= xa_load(xa
, addr_to_vb_idx(addr
));
2235 spin_lock(&vb
->lock
);
2236 bitmap_clear(vb
->used_map
, offset
, (1UL << order
));
2237 spin_unlock(&vb
->lock
);
2239 vunmap_range_noflush(addr
, addr
+ size
);
2241 if (debug_pagealloc_enabled_static())
2242 flush_tlb_kernel_range(addr
, addr
+ size
);
2244 spin_lock(&vb
->lock
);
2246 /* Expand the not yet TLB flushed dirty range */
2247 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
2248 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
2250 WRITE_ONCE(vb
->dirty
, vb
->dirty
+ (1UL << order
));
2251 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
2253 spin_unlock(&vb
->lock
);
2254 free_vmap_block(vb
);
2256 spin_unlock(&vb
->lock
);
2259 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
2261 LIST_HEAD(purge_list
);
2264 if (unlikely(!vmap_initialized
))
2267 mutex_lock(&vmap_purge_lock
);
2269 for_each_possible_cpu(cpu
) {
2270 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2271 struct vmap_block
*vb
;
2275 xa_for_each(&vbq
->vmap_blocks
, idx
, vb
) {
2276 spin_lock(&vb
->lock
);
2279 * Try to purge a fragmented block first. If it's
2280 * not purgeable, check whether there is dirty
2281 * space to be flushed.
2283 if (!purge_fragmented_block(vb
, vbq
, &purge_list
, false) &&
2284 vb
->dirty_max
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2285 unsigned long va_start
= vb
->va
->va_start
;
2288 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
2289 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
2291 start
= min(s
, start
);
2294 /* Prevent that this is flushed again */
2295 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2300 spin_unlock(&vb
->lock
);
2304 free_purged_blocks(&purge_list
);
2306 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
2307 flush_tlb_kernel_range(start
, end
);
2308 mutex_unlock(&vmap_purge_lock
);
2312 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2314 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2315 * to amortize TLB flushing overheads. What this means is that any page you
2316 * have now, may, in a former life, have been mapped into kernel virtual
2317 * address by the vmap layer and so there might be some CPUs with TLB entries
2318 * still referencing that page (additional to the regular 1:1 kernel mapping).
2320 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2321 * be sure that none of the pages we have control over will have any aliases
2322 * from the vmap layer.
2324 void vm_unmap_aliases(void)
2326 unsigned long start
= ULONG_MAX
, end
= 0;
2329 _vm_unmap_aliases(start
, end
, flush
);
2331 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
2334 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2335 * @mem: the pointer returned by vm_map_ram
2336 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2338 void vm_unmap_ram(const void *mem
, unsigned int count
)
2340 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2341 unsigned long addr
= (unsigned long)kasan_reset_tag(mem
);
2342 struct vmap_area
*va
;
2346 BUG_ON(addr
< VMALLOC_START
);
2347 BUG_ON(addr
> VMALLOC_END
);
2348 BUG_ON(!PAGE_ALIGNED(addr
));
2350 kasan_poison_vmalloc(mem
, size
);
2352 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2353 debug_check_no_locks_freed(mem
, size
);
2354 vb_free(addr
, size
);
2358 va
= find_unlink_vmap_area(addr
);
2359 if (WARN_ON_ONCE(!va
))
2362 debug_check_no_locks_freed((void *)va
->va_start
,
2363 (va
->va_end
- va
->va_start
));
2364 free_unmap_vmap_area(va
);
2366 EXPORT_SYMBOL(vm_unmap_ram
);
2369 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2370 * @pages: an array of pointers to the pages to be mapped
2371 * @count: number of pages
2372 * @node: prefer to allocate data structures on this node
2374 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2375 * faster than vmap so it's good. But if you mix long-life and short-life
2376 * objects with vm_map_ram(), it could consume lots of address space through
2377 * fragmentation (especially on a 32bit machine). You could see failures in
2378 * the end. Please use this function for short-lived objects.
2380 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2382 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
2384 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2388 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2389 mem
= vb_alloc(size
, GFP_KERNEL
);
2392 addr
= (unsigned long)mem
;
2394 struct vmap_area
*va
;
2395 va
= alloc_vmap_area(size
, PAGE_SIZE
,
2396 VMALLOC_START
, VMALLOC_END
,
2397 node
, GFP_KERNEL
, VMAP_RAM
);
2401 addr
= va
->va_start
;
2405 if (vmap_pages_range(addr
, addr
+ size
, PAGE_KERNEL
,
2406 pages
, PAGE_SHIFT
) < 0) {
2407 vm_unmap_ram(mem
, count
);
2412 * Mark the pages as accessible, now that they are mapped.
2413 * With hardware tag-based KASAN, marking is skipped for
2414 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2416 mem
= kasan_unpoison_vmalloc(mem
, size
, KASAN_VMALLOC_PROT_NORMAL
);
2420 EXPORT_SYMBOL(vm_map_ram
);
2422 static struct vm_struct
*vmlist __initdata
;
2424 static inline unsigned int vm_area_page_order(struct vm_struct
*vm
)
2426 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2427 return vm
->page_order
;
2433 static inline void set_vm_area_page_order(struct vm_struct
*vm
, unsigned int order
)
2435 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2436 vm
->page_order
= order
;
2443 * vm_area_add_early - add vmap area early during boot
2444 * @vm: vm_struct to add
2446 * This function is used to add fixed kernel vm area to vmlist before
2447 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2448 * should contain proper values and the other fields should be zero.
2450 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2452 void __init
vm_area_add_early(struct vm_struct
*vm
)
2454 struct vm_struct
*tmp
, **p
;
2456 BUG_ON(vmap_initialized
);
2457 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
2458 if (tmp
->addr
>= vm
->addr
) {
2459 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
2462 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
2469 * vm_area_register_early - register vmap area early during boot
2470 * @vm: vm_struct to register
2471 * @align: requested alignment
2473 * This function is used to register kernel vm area before
2474 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2475 * proper values on entry and other fields should be zero. On return,
2476 * vm->addr contains the allocated address.
2478 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2480 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
2482 unsigned long addr
= ALIGN(VMALLOC_START
, align
);
2483 struct vm_struct
*cur
, **p
;
2485 BUG_ON(vmap_initialized
);
2487 for (p
= &vmlist
; (cur
= *p
) != NULL
; p
= &cur
->next
) {
2488 if ((unsigned long)cur
->addr
- addr
>= vm
->size
)
2490 addr
= ALIGN((unsigned long)cur
->addr
+ cur
->size
, align
);
2493 BUG_ON(addr
> VMALLOC_END
- vm
->size
);
2494 vm
->addr
= (void *)addr
;
2497 kasan_populate_early_vm_area_shadow(vm
->addr
, vm
->size
);
2500 static void vmap_init_free_space(void)
2502 unsigned long vmap_start
= 1;
2503 const unsigned long vmap_end
= ULONG_MAX
;
2504 struct vmap_area
*busy
, *free
;
2508 * -|-----|.....|-----|-----|-----|.....|-
2510 * |<--------------------------------->|
2512 list_for_each_entry(busy
, &vmap_area_list
, list
) {
2513 if (busy
->va_start
- vmap_start
> 0) {
2514 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2515 if (!WARN_ON_ONCE(!free
)) {
2516 free
->va_start
= vmap_start
;
2517 free
->va_end
= busy
->va_start
;
2519 insert_vmap_area_augment(free
, NULL
,
2520 &free_vmap_area_root
,
2521 &free_vmap_area_list
);
2525 vmap_start
= busy
->va_end
;
2528 if (vmap_end
- vmap_start
> 0) {
2529 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2530 if (!WARN_ON_ONCE(!free
)) {
2531 free
->va_start
= vmap_start
;
2532 free
->va_end
= vmap_end
;
2534 insert_vmap_area_augment(free
, NULL
,
2535 &free_vmap_area_root
,
2536 &free_vmap_area_list
);
2541 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2542 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2545 vm
->addr
= (void *)va
->va_start
;
2546 vm
->size
= va
->va_end
- va
->va_start
;
2547 vm
->caller
= caller
;
2551 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2552 unsigned long flags
, const void *caller
)
2554 spin_lock(&vmap_area_lock
);
2555 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2556 spin_unlock(&vmap_area_lock
);
2559 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2562 * Before removing VM_UNINITIALIZED,
2563 * we should make sure that vm has proper values.
2564 * Pair with smp_rmb() in show_numa_info().
2567 vm
->flags
&= ~VM_UNINITIALIZED
;
2570 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2571 unsigned long align
, unsigned long shift
, unsigned long flags
,
2572 unsigned long start
, unsigned long end
, int node
,
2573 gfp_t gfp_mask
, const void *caller
)
2575 struct vmap_area
*va
;
2576 struct vm_struct
*area
;
2577 unsigned long requested_size
= size
;
2579 BUG_ON(in_interrupt());
2580 size
= ALIGN(size
, 1ul << shift
);
2581 if (unlikely(!size
))
2584 if (flags
& VM_IOREMAP
)
2585 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2586 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2588 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2589 if (unlikely(!area
))
2592 if (!(flags
& VM_NO_GUARD
))
2595 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
, 0);
2601 setup_vmalloc_vm(area
, va
, flags
, caller
);
2604 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2605 * best-effort approach, as they can be mapped outside of vmalloc code.
2606 * For VM_ALLOC mappings, the pages are marked as accessible after
2607 * getting mapped in __vmalloc_node_range().
2608 * With hardware tag-based KASAN, marking is skipped for
2609 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2611 if (!(flags
& VM_ALLOC
))
2612 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, requested_size
,
2613 KASAN_VMALLOC_PROT_NORMAL
);
2618 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2619 unsigned long start
, unsigned long end
,
2622 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
, start
, end
,
2623 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2627 * get_vm_area - reserve a contiguous kernel virtual area
2628 * @size: size of the area
2629 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2631 * Search an area of @size in the kernel virtual mapping area,
2632 * and reserved it for out purposes. Returns the area descriptor
2633 * on success or %NULL on failure.
2635 * Return: the area descriptor on success or %NULL on failure.
2637 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2639 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2640 VMALLOC_START
, VMALLOC_END
,
2641 NUMA_NO_NODE
, GFP_KERNEL
,
2642 __builtin_return_address(0));
2645 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2648 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2649 VMALLOC_START
, VMALLOC_END
,
2650 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2654 * find_vm_area - find a continuous kernel virtual area
2655 * @addr: base address
2657 * Search for the kernel VM area starting at @addr, and return it.
2658 * It is up to the caller to do all required locking to keep the returned
2661 * Return: the area descriptor on success or %NULL on failure.
2663 struct vm_struct
*find_vm_area(const void *addr
)
2665 struct vmap_area
*va
;
2667 va
= find_vmap_area((unsigned long)addr
);
2675 * remove_vm_area - find and remove a continuous kernel virtual area
2676 * @addr: base address
2678 * Search for the kernel VM area starting at @addr, and remove it.
2679 * This function returns the found VM area, but using it is NOT safe
2680 * on SMP machines, except for its size or flags.
2682 * Return: the area descriptor on success or %NULL on failure.
2684 struct vm_struct
*remove_vm_area(const void *addr
)
2686 struct vmap_area
*va
;
2687 struct vm_struct
*vm
;
2691 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2695 va
= find_unlink_vmap_area((unsigned long)addr
);
2700 debug_check_no_locks_freed(vm
->addr
, get_vm_area_size(vm
));
2701 debug_check_no_obj_freed(vm
->addr
, get_vm_area_size(vm
));
2702 kasan_free_module_shadow(vm
);
2703 kasan_poison_vmalloc(vm
->addr
, get_vm_area_size(vm
));
2705 free_unmap_vmap_area(va
);
2709 static inline void set_area_direct_map(const struct vm_struct
*area
,
2710 int (*set_direct_map
)(struct page
*page
))
2714 /* HUGE_VMALLOC passes small pages to set_direct_map */
2715 for (i
= 0; i
< area
->nr_pages
; i
++)
2716 if (page_address(area
->pages
[i
]))
2717 set_direct_map(area
->pages
[i
]);
2721 * Flush the vm mapping and reset the direct map.
2723 static void vm_reset_perms(struct vm_struct
*area
)
2725 unsigned long start
= ULONG_MAX
, end
= 0;
2726 unsigned int page_order
= vm_area_page_order(area
);
2731 * Find the start and end range of the direct mappings to make sure that
2732 * the vm_unmap_aliases() flush includes the direct map.
2734 for (i
= 0; i
< area
->nr_pages
; i
+= 1U << page_order
) {
2735 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2738 unsigned long page_size
;
2740 page_size
= PAGE_SIZE
<< page_order
;
2741 start
= min(addr
, start
);
2742 end
= max(addr
+ page_size
, end
);
2748 * Set direct map to something invalid so that it won't be cached if
2749 * there are any accesses after the TLB flush, then flush the TLB and
2750 * reset the direct map permissions to the default.
2752 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2753 _vm_unmap_aliases(start
, end
, flush_dmap
);
2754 set_area_direct_map(area
, set_direct_map_default_noflush
);
2757 static void delayed_vfree_work(struct work_struct
*w
)
2759 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
2760 struct llist_node
*t
, *llnode
;
2762 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
2767 * vfree_atomic - release memory allocated by vmalloc()
2768 * @addr: memory base address
2770 * This one is just like vfree() but can be called in any atomic context
2773 void vfree_atomic(const void *addr
)
2775 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2778 kmemleak_free(addr
);
2781 * Use raw_cpu_ptr() because this can be called from preemptible
2782 * context. Preemption is absolutely fine here, because the llist_add()
2783 * implementation is lockless, so it works even if we are adding to
2784 * another cpu's list. schedule_work() should be fine with this too.
2786 if (addr
&& llist_add((struct llist_node
*)addr
, &p
->list
))
2787 schedule_work(&p
->wq
);
2791 * vfree - Release memory allocated by vmalloc()
2792 * @addr: Memory base address
2794 * Free the virtually continuous memory area starting at @addr, as obtained
2795 * from one of the vmalloc() family of APIs. This will usually also free the
2796 * physical memory underlying the virtual allocation, but that memory is
2797 * reference counted, so it will not be freed until the last user goes away.
2799 * If @addr is NULL, no operation is performed.
2802 * May sleep if called *not* from interrupt context.
2803 * Must not be called in NMI context (strictly speaking, it could be
2804 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2805 * conventions for vfree() arch-dependent would be a really bad idea).
2807 void vfree(const void *addr
)
2809 struct vm_struct
*vm
;
2812 if (unlikely(in_interrupt())) {
2818 kmemleak_free(addr
);
2824 vm
= remove_vm_area(addr
);
2825 if (unlikely(!vm
)) {
2826 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2831 if (unlikely(vm
->flags
& VM_FLUSH_RESET_PERMS
))
2833 for (i
= 0; i
< vm
->nr_pages
; i
++) {
2834 struct page
*page
= vm
->pages
[i
];
2837 mod_memcg_page_state(page
, MEMCG_VMALLOC
, -1);
2839 * High-order allocs for huge vmallocs are split, so
2840 * can be freed as an array of order-0 allocations
2845 atomic_long_sub(vm
->nr_pages
, &nr_vmalloc_pages
);
2849 EXPORT_SYMBOL(vfree
);
2852 * vunmap - release virtual mapping obtained by vmap()
2853 * @addr: memory base address
2855 * Free the virtually contiguous memory area starting at @addr,
2856 * which was created from the page array passed to vmap().
2858 * Must not be called in interrupt context.
2860 void vunmap(const void *addr
)
2862 struct vm_struct
*vm
;
2864 BUG_ON(in_interrupt());
2869 vm
= remove_vm_area(addr
);
2870 if (unlikely(!vm
)) {
2871 WARN(1, KERN_ERR
"Trying to vunmap() nonexistent vm area (%p)\n",
2877 EXPORT_SYMBOL(vunmap
);
2880 * vmap - map an array of pages into virtually contiguous space
2881 * @pages: array of page pointers
2882 * @count: number of pages to map
2883 * @flags: vm_area->flags
2884 * @prot: page protection for the mapping
2886 * Maps @count pages from @pages into contiguous kernel virtual space.
2887 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2888 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2889 * are transferred from the caller to vmap(), and will be freed / dropped when
2890 * vfree() is called on the return value.
2892 * Return: the address of the area or %NULL on failure
2894 void *vmap(struct page
**pages
, unsigned int count
,
2895 unsigned long flags
, pgprot_t prot
)
2897 struct vm_struct
*area
;
2899 unsigned long size
; /* In bytes */
2903 if (WARN_ON_ONCE(flags
& VM_FLUSH_RESET_PERMS
))
2907 * Your top guard is someone else's bottom guard. Not having a top
2908 * guard compromises someone else's mappings too.
2910 if (WARN_ON_ONCE(flags
& VM_NO_GUARD
))
2911 flags
&= ~VM_NO_GUARD
;
2913 if (count
> totalram_pages())
2916 size
= (unsigned long)count
<< PAGE_SHIFT
;
2917 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2921 addr
= (unsigned long)area
->addr
;
2922 if (vmap_pages_range(addr
, addr
+ size
, pgprot_nx(prot
),
2923 pages
, PAGE_SHIFT
) < 0) {
2928 if (flags
& VM_MAP_PUT_PAGES
) {
2929 area
->pages
= pages
;
2930 area
->nr_pages
= count
;
2934 EXPORT_SYMBOL(vmap
);
2936 #ifdef CONFIG_VMAP_PFN
2937 struct vmap_pfn_data
{
2938 unsigned long *pfns
;
2943 static int vmap_pfn_apply(pte_t
*pte
, unsigned long addr
, void *private)
2945 struct vmap_pfn_data
*data
= private;
2946 unsigned long pfn
= data
->pfns
[data
->idx
];
2949 if (WARN_ON_ONCE(pfn_valid(pfn
)))
2952 ptent
= pte_mkspecial(pfn_pte(pfn
, data
->prot
));
2953 set_pte_at(&init_mm
, addr
, pte
, ptent
);
2960 * vmap_pfn - map an array of PFNs into virtually contiguous space
2961 * @pfns: array of PFNs
2962 * @count: number of pages to map
2963 * @prot: page protection for the mapping
2965 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2966 * the start address of the mapping.
2968 void *vmap_pfn(unsigned long *pfns
, unsigned int count
, pgprot_t prot
)
2970 struct vmap_pfn_data data
= { .pfns
= pfns
, .prot
= pgprot_nx(prot
) };
2971 struct vm_struct
*area
;
2973 area
= get_vm_area_caller(count
* PAGE_SIZE
, VM_IOREMAP
,
2974 __builtin_return_address(0));
2977 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2978 count
* PAGE_SIZE
, vmap_pfn_apply
, &data
)) {
2984 EXPORT_SYMBOL_GPL(vmap_pfn
);
2985 #endif /* CONFIG_VMAP_PFN */
2987 static inline unsigned int
2988 vm_area_alloc_pages(gfp_t gfp
, int nid
,
2989 unsigned int order
, unsigned int nr_pages
, struct page
**pages
)
2991 unsigned int nr_allocated
= 0;
2992 gfp_t alloc_gfp
= gfp
;
2993 bool nofail
= false;
2998 * For order-0 pages we make use of bulk allocator, if
2999 * the page array is partly or not at all populated due
3000 * to fails, fallback to a single page allocator that is
3004 /* bulk allocator doesn't support nofail req. officially */
3005 gfp_t bulk_gfp
= gfp
& ~__GFP_NOFAIL
;
3007 while (nr_allocated
< nr_pages
) {
3008 unsigned int nr
, nr_pages_request
;
3011 * A maximum allowed request is hard-coded and is 100
3012 * pages per call. That is done in order to prevent a
3013 * long preemption off scenario in the bulk-allocator
3014 * so the range is [1:100].
3016 nr_pages_request
= min(100U, nr_pages
- nr_allocated
);
3018 /* memory allocation should consider mempolicy, we can't
3019 * wrongly use nearest node when nid == NUMA_NO_NODE,
3020 * otherwise memory may be allocated in only one node,
3021 * but mempolicy wants to alloc memory by interleaving.
3023 if (IS_ENABLED(CONFIG_NUMA
) && nid
== NUMA_NO_NODE
)
3024 nr
= alloc_pages_bulk_array_mempolicy(bulk_gfp
,
3026 pages
+ nr_allocated
);
3029 nr
= alloc_pages_bulk_array_node(bulk_gfp
, nid
,
3031 pages
+ nr_allocated
);
3037 * If zero or pages were obtained partly,
3038 * fallback to a single page allocator.
3040 if (nr
!= nr_pages_request
)
3043 } else if (gfp
& __GFP_NOFAIL
) {
3045 * Higher order nofail allocations are really expensive and
3046 * potentially dangerous (pre-mature OOM, disruptive reclaim
3047 * and compaction etc.
3049 alloc_gfp
&= ~__GFP_NOFAIL
;
3053 /* High-order pages or fallback path if "bulk" fails. */
3054 while (nr_allocated
< nr_pages
) {
3055 if (fatal_signal_pending(current
))
3058 if (nid
== NUMA_NO_NODE
)
3059 page
= alloc_pages(alloc_gfp
, order
);
3061 page
= alloc_pages_node(nid
, alloc_gfp
, order
);
3062 if (unlikely(!page
)) {
3066 /* fall back to the zero order allocations */
3067 alloc_gfp
|= __GFP_NOFAIL
;
3073 * Higher order allocations must be able to be treated as
3074 * indepdenent small pages by callers (as they can with
3075 * small-page vmallocs). Some drivers do their own refcounting
3076 * on vmalloc_to_page() pages, some use page->mapping,
3080 split_page(page
, order
);
3083 * Careful, we allocate and map page-order pages, but
3084 * tracking is done per PAGE_SIZE page so as to keep the
3085 * vm_struct APIs independent of the physical/mapped size.
3087 for (i
= 0; i
< (1U << order
); i
++)
3088 pages
[nr_allocated
+ i
] = page
+ i
;
3091 nr_allocated
+= 1U << order
;
3094 return nr_allocated
;
3097 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
3098 pgprot_t prot
, unsigned int page_shift
,
3101 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
3102 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3103 unsigned long addr
= (unsigned long)area
->addr
;
3104 unsigned long size
= get_vm_area_size(area
);
3105 unsigned long array_size
;
3106 unsigned int nr_small_pages
= size
>> PAGE_SHIFT
;
3107 unsigned int page_order
;
3111 array_size
= (unsigned long)nr_small_pages
* sizeof(struct page
*);
3113 if (!(gfp_mask
& (GFP_DMA
| GFP_DMA32
)))
3114 gfp_mask
|= __GFP_HIGHMEM
;
3116 /* Please note that the recursion is strictly bounded. */
3117 if (array_size
> PAGE_SIZE
) {
3118 area
->pages
= __vmalloc_node(array_size
, 1, nested_gfp
, node
,
3121 area
->pages
= kmalloc_node(array_size
, nested_gfp
, node
);
3125 warn_alloc(gfp_mask
, NULL
,
3126 "vmalloc error: size %lu, failed to allocated page array size %lu",
3127 nr_small_pages
* PAGE_SIZE
, array_size
);
3132 set_vm_area_page_order(area
, page_shift
- PAGE_SHIFT
);
3133 page_order
= vm_area_page_order(area
);
3135 area
->nr_pages
= vm_area_alloc_pages(gfp_mask
| __GFP_NOWARN
,
3136 node
, page_order
, nr_small_pages
, area
->pages
);
3138 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
3139 if (gfp_mask
& __GFP_ACCOUNT
) {
3142 for (i
= 0; i
< area
->nr_pages
; i
++)
3143 mod_memcg_page_state(area
->pages
[i
], MEMCG_VMALLOC
, 1);
3147 * If not enough pages were obtained to accomplish an
3148 * allocation request, free them via vfree() if any.
3150 if (area
->nr_pages
!= nr_small_pages
) {
3152 * vm_area_alloc_pages() can fail due to insufficient memory but
3155 * - a pending fatal signal
3156 * - insufficient huge page-order pages
3158 * Since we always retry allocations at order-0 in the huge page
3159 * case a warning for either is spurious.
3161 if (!fatal_signal_pending(current
) && page_order
== 0)
3162 warn_alloc(gfp_mask
, NULL
,
3163 "vmalloc error: size %lu, failed to allocate pages",
3164 area
->nr_pages
* PAGE_SIZE
);
3169 * page tables allocations ignore external gfp mask, enforce it
3172 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3173 flags
= memalloc_nofs_save();
3174 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3175 flags
= memalloc_noio_save();
3178 ret
= vmap_pages_range(addr
, addr
+ size
, prot
, area
->pages
,
3180 if (nofail
&& (ret
< 0))
3181 schedule_timeout_uninterruptible(1);
3182 } while (nofail
&& (ret
< 0));
3184 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3185 memalloc_nofs_restore(flags
);
3186 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3187 memalloc_noio_restore(flags
);
3190 warn_alloc(gfp_mask
, NULL
,
3191 "vmalloc error: size %lu, failed to map pages",
3192 area
->nr_pages
* PAGE_SIZE
);
3204 * __vmalloc_node_range - allocate virtually contiguous memory
3205 * @size: allocation size
3206 * @align: desired alignment
3207 * @start: vm area range start
3208 * @end: vm area range end
3209 * @gfp_mask: flags for the page level allocator
3210 * @prot: protection mask for the allocated pages
3211 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3212 * @node: node to use for allocation or NUMA_NO_NODE
3213 * @caller: caller's return address
3215 * Allocate enough pages to cover @size from the page level
3216 * allocator with @gfp_mask flags. Please note that the full set of gfp
3217 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3219 * Zone modifiers are not supported. From the reclaim modifiers
3220 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3221 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3222 * __GFP_RETRY_MAYFAIL are not supported).
3224 * __GFP_NOWARN can be used to suppress failures messages.
3226 * Map them into contiguous kernel virtual space, using a pagetable
3227 * protection of @prot.
3229 * Return: the address of the area or %NULL on failure
3231 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
3232 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
3233 pgprot_t prot
, unsigned long vm_flags
, int node
,
3236 struct vm_struct
*area
;
3238 kasan_vmalloc_flags_t kasan_flags
= KASAN_VMALLOC_NONE
;
3239 unsigned long real_size
= size
;
3240 unsigned long real_align
= align
;
3241 unsigned int shift
= PAGE_SHIFT
;
3243 if (WARN_ON_ONCE(!size
))
3246 if ((size
>> PAGE_SHIFT
) > totalram_pages()) {
3247 warn_alloc(gfp_mask
, NULL
,
3248 "vmalloc error: size %lu, exceeds total pages",
3253 if (vmap_allow_huge
&& (vm_flags
& VM_ALLOW_HUGE_VMAP
)) {
3254 unsigned long size_per_node
;
3257 * Try huge pages. Only try for PAGE_KERNEL allocations,
3258 * others like modules don't yet expect huge pages in
3259 * their allocations due to apply_to_page_range not
3263 size_per_node
= size
;
3264 if (node
== NUMA_NO_NODE
)
3265 size_per_node
/= num_online_nodes();
3266 if (arch_vmap_pmd_supported(prot
) && size_per_node
>= PMD_SIZE
)
3269 shift
= arch_vmap_pte_supported_shift(size_per_node
);
3271 align
= max(real_align
, 1UL << shift
);
3272 size
= ALIGN(real_size
, 1UL << shift
);
3276 area
= __get_vm_area_node(real_size
, align
, shift
, VM_ALLOC
|
3277 VM_UNINITIALIZED
| vm_flags
, start
, end
, node
,
3280 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3281 warn_alloc(gfp_mask
, NULL
,
3282 "vmalloc error: size %lu, vm_struct allocation failed%s",
3283 real_size
, (nofail
) ? ". Retrying." : "");
3285 schedule_timeout_uninterruptible(1);
3292 * Prepare arguments for __vmalloc_area_node() and
3293 * kasan_unpoison_vmalloc().
3295 if (pgprot_val(prot
) == pgprot_val(PAGE_KERNEL
)) {
3296 if (kasan_hw_tags_enabled()) {
3298 * Modify protection bits to allow tagging.
3299 * This must be done before mapping.
3301 prot
= arch_vmap_pgprot_tagged(prot
);
3304 * Skip page_alloc poisoning and zeroing for physical
3305 * pages backing VM_ALLOC mapping. Memory is instead
3306 * poisoned and zeroed by kasan_unpoison_vmalloc().
3308 gfp_mask
|= __GFP_SKIP_KASAN
| __GFP_SKIP_ZERO
;
3311 /* Take note that the mapping is PAGE_KERNEL. */
3312 kasan_flags
|= KASAN_VMALLOC_PROT_NORMAL
;
3315 /* Allocate physical pages and map them into vmalloc space. */
3316 ret
= __vmalloc_area_node(area
, gfp_mask
, prot
, shift
, node
);
3321 * Mark the pages as accessible, now that they are mapped.
3322 * The condition for setting KASAN_VMALLOC_INIT should complement the
3323 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3324 * to make sure that memory is initialized under the same conditions.
3325 * Tag-based KASAN modes only assign tags to normal non-executable
3326 * allocations, see __kasan_unpoison_vmalloc().
3328 kasan_flags
|= KASAN_VMALLOC_VM_ALLOC
;
3329 if (!want_init_on_free() && want_init_on_alloc(gfp_mask
) &&
3330 (gfp_mask
& __GFP_SKIP_ZERO
))
3331 kasan_flags
|= KASAN_VMALLOC_INIT
;
3332 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3333 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, real_size
, kasan_flags
);
3336 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3337 * flag. It means that vm_struct is not fully initialized.
3338 * Now, it is fully initialized, so remove this flag here.
3340 clear_vm_uninitialized_flag(area
);
3342 size
= PAGE_ALIGN(size
);
3343 if (!(vm_flags
& VM_DEFER_KMEMLEAK
))
3344 kmemleak_vmalloc(area
, size
, gfp_mask
);
3349 if (shift
> PAGE_SHIFT
) {
3360 * __vmalloc_node - allocate virtually contiguous memory
3361 * @size: allocation size
3362 * @align: desired alignment
3363 * @gfp_mask: flags for the page level allocator
3364 * @node: node to use for allocation or NUMA_NO_NODE
3365 * @caller: caller's return address
3367 * Allocate enough pages to cover @size from the page level allocator with
3368 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3370 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3371 * and __GFP_NOFAIL are not supported
3373 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3376 * Return: pointer to the allocated memory or %NULL on error
3378 void *__vmalloc_node(unsigned long size
, unsigned long align
,
3379 gfp_t gfp_mask
, int node
, const void *caller
)
3381 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
3382 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
3385 * This is only for performance analysis of vmalloc and stress purpose.
3386 * It is required by vmalloc test module, therefore do not use it other
3389 #ifdef CONFIG_TEST_VMALLOC_MODULE
3390 EXPORT_SYMBOL_GPL(__vmalloc_node
);
3393 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
3395 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
3396 __builtin_return_address(0));
3398 EXPORT_SYMBOL(__vmalloc
);
3401 * vmalloc - allocate virtually contiguous memory
3402 * @size: allocation size
3404 * Allocate enough pages to cover @size from the page level
3405 * allocator and map them into contiguous kernel virtual space.
3407 * For tight control over page level allocator and protection flags
3408 * use __vmalloc() instead.
3410 * Return: pointer to the allocated memory or %NULL on error
3412 void *vmalloc(unsigned long size
)
3414 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
3415 __builtin_return_address(0));
3417 EXPORT_SYMBOL(vmalloc
);
3420 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3421 * @size: allocation size
3422 * @gfp_mask: flags for the page level allocator
3424 * Allocate enough pages to cover @size from the page level
3425 * allocator and map them into contiguous kernel virtual space.
3426 * If @size is greater than or equal to PMD_SIZE, allow using
3427 * huge pages for the memory
3429 * Return: pointer to the allocated memory or %NULL on error
3431 void *vmalloc_huge(unsigned long size
, gfp_t gfp_mask
)
3433 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
3434 gfp_mask
, PAGE_KERNEL
, VM_ALLOW_HUGE_VMAP
,
3435 NUMA_NO_NODE
, __builtin_return_address(0));
3437 EXPORT_SYMBOL_GPL(vmalloc_huge
);
3440 * vzalloc - allocate virtually contiguous memory with zero fill
3441 * @size: allocation size
3443 * Allocate enough pages to cover @size from the page level
3444 * allocator and map them into contiguous kernel virtual space.
3445 * The memory allocated is set to zero.
3447 * For tight control over page level allocator and protection flags
3448 * use __vmalloc() instead.
3450 * Return: pointer to the allocated memory or %NULL on error
3452 void *vzalloc(unsigned long size
)
3454 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
3455 __builtin_return_address(0));
3457 EXPORT_SYMBOL(vzalloc
);
3460 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3461 * @size: allocation size
3463 * The resulting memory area is zeroed so it can be mapped to userspace
3464 * without leaking data.
3466 * Return: pointer to the allocated memory or %NULL on error
3468 void *vmalloc_user(unsigned long size
)
3470 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3471 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
3472 VM_USERMAP
, NUMA_NO_NODE
,
3473 __builtin_return_address(0));
3475 EXPORT_SYMBOL(vmalloc_user
);
3478 * vmalloc_node - allocate memory on a specific node
3479 * @size: allocation size
3482 * Allocate enough pages to cover @size from the page level
3483 * allocator and map them into contiguous kernel virtual space.
3485 * For tight control over page level allocator and protection flags
3486 * use __vmalloc() instead.
3488 * Return: pointer to the allocated memory or %NULL on error
3490 void *vmalloc_node(unsigned long size
, int node
)
3492 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
3493 __builtin_return_address(0));
3495 EXPORT_SYMBOL(vmalloc_node
);
3498 * vzalloc_node - allocate memory on a specific node with zero fill
3499 * @size: allocation size
3502 * Allocate enough pages to cover @size from the page level
3503 * allocator and map them into contiguous kernel virtual space.
3504 * The memory allocated is set to zero.
3506 * Return: pointer to the allocated memory or %NULL on error
3508 void *vzalloc_node(unsigned long size
, int node
)
3510 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
3511 __builtin_return_address(0));
3513 EXPORT_SYMBOL(vzalloc_node
);
3515 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3516 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3517 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3518 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3521 * 64b systems should always have either DMA or DMA32 zones. For others
3522 * GFP_DMA32 should do the right thing and use the normal zone.
3524 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3528 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3529 * @size: allocation size
3531 * Allocate enough 32bit PA addressable pages to cover @size from the
3532 * page level allocator and map them into contiguous kernel virtual space.
3534 * Return: pointer to the allocated memory or %NULL on error
3536 void *vmalloc_32(unsigned long size
)
3538 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
3539 __builtin_return_address(0));
3541 EXPORT_SYMBOL(vmalloc_32
);
3544 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3545 * @size: allocation size
3547 * The resulting memory area is 32bit addressable and zeroed so it can be
3548 * mapped to userspace without leaking data.
3550 * Return: pointer to the allocated memory or %NULL on error
3552 void *vmalloc_32_user(unsigned long size
)
3554 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3555 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
3556 VM_USERMAP
, NUMA_NO_NODE
,
3557 __builtin_return_address(0));
3559 EXPORT_SYMBOL(vmalloc_32_user
);
3562 * Atomically zero bytes in the iterator.
3564 * Returns the number of zeroed bytes.
3566 static size_t zero_iter(struct iov_iter
*iter
, size_t count
)
3568 size_t remains
= count
;
3570 while (remains
> 0) {
3573 num
= min_t(size_t, remains
, PAGE_SIZE
);
3574 copied
= copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num
, iter
);
3581 return count
- remains
;
3585 * small helper routine, copy contents to iter from addr.
3586 * If the page is not present, fill zero.
3588 * Returns the number of copied bytes.
3590 static size_t aligned_vread_iter(struct iov_iter
*iter
,
3591 const char *addr
, size_t count
)
3593 size_t remains
= count
;
3596 while (remains
> 0) {
3597 unsigned long offset
, length
;
3600 offset
= offset_in_page(addr
);
3601 length
= PAGE_SIZE
- offset
;
3602 if (length
> remains
)
3604 page
= vmalloc_to_page(addr
);
3606 * To do safe access to this _mapped_ area, we need lock. But
3607 * adding lock here means that we need to add overhead of
3608 * vmalloc()/vfree() calls for this _debug_ interface, rarely
3609 * used. Instead of that, we'll use an local mapping via
3610 * copy_page_to_iter_nofault() and accept a small overhead in
3611 * this access function.
3614 copied
= copy_page_to_iter_nofault(page
, offset
,
3617 copied
= zero_iter(iter
, length
);
3622 if (copied
!= length
)
3626 return count
- remains
;
3630 * Read from a vm_map_ram region of memory.
3632 * Returns the number of copied bytes.
3634 static size_t vmap_ram_vread_iter(struct iov_iter
*iter
, const char *addr
,
3635 size_t count
, unsigned long flags
)
3638 struct vmap_block
*vb
;
3640 unsigned long offset
;
3641 unsigned int rs
, re
;
3645 * If it's area created by vm_map_ram() interface directly, but
3646 * not further subdividing and delegating management to vmap_block,
3649 if (!(flags
& VMAP_BLOCK
))
3650 return aligned_vread_iter(iter
, addr
, count
);
3655 * Area is split into regions and tracked with vmap_block, read out
3656 * each region and zero fill the hole between regions.
3658 xa
= addr_to_vb_xa((unsigned long) addr
);
3659 vb
= xa_load(xa
, addr_to_vb_idx((unsigned long)addr
));
3663 spin_lock(&vb
->lock
);
3664 if (bitmap_empty(vb
->used_map
, VMAP_BBMAP_BITS
)) {
3665 spin_unlock(&vb
->lock
);
3669 for_each_set_bitrange(rs
, re
, vb
->used_map
, VMAP_BBMAP_BITS
) {
3675 start
= vmap_block_vaddr(vb
->va
->va_start
, rs
);
3678 size_t to_zero
= min_t(size_t, start
- addr
, remains
);
3679 size_t zeroed
= zero_iter(iter
, to_zero
);
3684 if (remains
== 0 || zeroed
!= to_zero
)
3688 /*it could start reading from the middle of used region*/
3689 offset
= offset_in_page(addr
);
3690 n
= ((re
- rs
+ 1) << PAGE_SHIFT
) - offset
;
3694 copied
= aligned_vread_iter(iter
, start
+ offset
, n
);
3703 spin_unlock(&vb
->lock
);
3706 /* zero-fill the left dirty or free regions */
3707 return count
- remains
+ zero_iter(iter
, remains
);
3709 /* We couldn't copy/zero everything */
3710 spin_unlock(&vb
->lock
);
3711 return count
- remains
;
3715 * vread_iter() - read vmalloc area in a safe way to an iterator.
3716 * @iter: the iterator to which data should be written.
3717 * @addr: vm address.
3718 * @count: number of bytes to be read.
3720 * This function checks that addr is a valid vmalloc'ed area, and
3721 * copy data from that area to a given buffer. If the given memory range
3722 * of [addr...addr+count) includes some valid address, data is copied to
3723 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3724 * IOREMAP area is treated as memory hole and no copy is done.
3726 * If [addr...addr+count) doesn't includes any intersects with alive
3727 * vm_struct area, returns 0. @buf should be kernel's buffer.
3729 * Note: In usual ops, vread() is never necessary because the caller
3730 * should know vmalloc() area is valid and can use memcpy().
3731 * This is for routines which have to access vmalloc area without
3732 * any information, as /proc/kcore.
3734 * Return: number of bytes for which addr and buf should be increased
3735 * (same number as @count) or %0 if [addr...addr+count) doesn't
3736 * include any intersection with valid vmalloc area
3738 long vread_iter(struct iov_iter
*iter
, const char *addr
, size_t count
)
3740 struct vmap_area
*va
;
3741 struct vm_struct
*vm
;
3743 size_t n
, size
, flags
, remains
;
3745 addr
= kasan_reset_tag(addr
);
3747 /* Don't allow overflow */
3748 if ((unsigned long) addr
+ count
< count
)
3749 count
= -(unsigned long) addr
;
3753 spin_lock(&vmap_area_lock
);
3754 va
= find_vmap_area_exceed_addr((unsigned long)addr
);
3758 /* no intersects with alive vmap_area */
3759 if ((unsigned long)addr
+ remains
<= va
->va_start
)
3762 list_for_each_entry_from(va
, &vmap_area_list
, list
) {
3769 flags
= va
->flags
& VMAP_FLAGS_MASK
;
3771 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
3772 * be set together with VMAP_RAM.
3774 WARN_ON(flags
== VMAP_BLOCK
);
3779 if (vm
&& (vm
->flags
& VM_UNINITIALIZED
))
3782 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3785 vaddr
= (char *) va
->va_start
;
3786 size
= vm
? get_vm_area_size(vm
) : va_size(va
);
3788 if (addr
>= vaddr
+ size
)
3792 size_t to_zero
= min_t(size_t, vaddr
- addr
, remains
);
3793 size_t zeroed
= zero_iter(iter
, to_zero
);
3798 if (remains
== 0 || zeroed
!= to_zero
)
3802 n
= vaddr
+ size
- addr
;
3806 if (flags
& VMAP_RAM
)
3807 copied
= vmap_ram_vread_iter(iter
, addr
, n
, flags
);
3808 else if (!(vm
->flags
& VM_IOREMAP
))
3809 copied
= aligned_vread_iter(iter
, addr
, n
);
3810 else /* IOREMAP area is treated as memory hole */
3811 copied
= zero_iter(iter
, n
);
3821 spin_unlock(&vmap_area_lock
);
3822 /* zero-fill memory holes */
3823 return count
- remains
+ zero_iter(iter
, remains
);
3825 /* Nothing remains, or We couldn't copy/zero everything. */
3826 spin_unlock(&vmap_area_lock
);
3828 return count
- remains
;
3832 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3833 * @vma: vma to cover
3834 * @uaddr: target user address to start at
3835 * @kaddr: virtual address of vmalloc kernel memory
3836 * @pgoff: offset from @kaddr to start at
3837 * @size: size of map area
3839 * Returns: 0 for success, -Exxx on failure
3841 * This function checks that @kaddr is a valid vmalloc'ed area,
3842 * and that it is big enough to cover the range starting at
3843 * @uaddr in @vma. Will return failure if that criteria isn't
3846 * Similar to remap_pfn_range() (see mm/memory.c)
3848 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3849 void *kaddr
, unsigned long pgoff
,
3852 struct vm_struct
*area
;
3854 unsigned long end_index
;
3856 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
3859 size
= PAGE_ALIGN(size
);
3861 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3864 area
= find_vm_area(kaddr
);
3868 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3871 if (check_add_overflow(size
, off
, &end_index
) ||
3872 end_index
> get_vm_area_size(area
))
3877 struct page
*page
= vmalloc_to_page(kaddr
);
3880 ret
= vm_insert_page(vma
, uaddr
, page
);
3889 vm_flags_set(vma
, VM_DONTEXPAND
| VM_DONTDUMP
);
3895 * remap_vmalloc_range - map vmalloc pages to userspace
3896 * @vma: vma to cover (map full range of vma)
3897 * @addr: vmalloc memory
3898 * @pgoff: number of pages into addr before first page to map
3900 * Returns: 0 for success, -Exxx on failure
3902 * This function checks that addr is a valid vmalloc'ed area, and
3903 * that it is big enough to cover the vma. Will return failure if
3904 * that criteria isn't met.
3906 * Similar to remap_pfn_range() (see mm/memory.c)
3908 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3909 unsigned long pgoff
)
3911 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3913 vma
->vm_end
- vma
->vm_start
);
3915 EXPORT_SYMBOL(remap_vmalloc_range
);
3917 void free_vm_area(struct vm_struct
*area
)
3919 struct vm_struct
*ret
;
3920 ret
= remove_vm_area(area
->addr
);
3921 BUG_ON(ret
!= area
);
3924 EXPORT_SYMBOL_GPL(free_vm_area
);
3927 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3929 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3933 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3934 * @addr: target address
3936 * Returns: vmap_area if it is found. If there is no such area
3937 * the first highest(reverse order) vmap_area is returned
3938 * i.e. va->va_start < addr && va->va_end < addr or NULL
3939 * if there are no any areas before @addr.
3941 static struct vmap_area
*
3942 pvm_find_va_enclose_addr(unsigned long addr
)
3944 struct vmap_area
*va
, *tmp
;
3947 n
= free_vmap_area_root
.rb_node
;
3951 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3952 if (tmp
->va_start
<= addr
) {
3954 if (tmp
->va_end
>= addr
)
3967 * pvm_determine_end_from_reverse - find the highest aligned address
3968 * of free block below VMALLOC_END
3970 * in - the VA we start the search(reverse order);
3971 * out - the VA with the highest aligned end address.
3972 * @align: alignment for required highest address
3974 * Returns: determined end address within vmap_area
3976 static unsigned long
3977 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3979 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3983 list_for_each_entry_from_reverse((*va
),
3984 &free_vmap_area_list
, list
) {
3985 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3986 if ((*va
)->va_start
< addr
)
3995 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3996 * @offsets: array containing offset of each area
3997 * @sizes: array containing size of each area
3998 * @nr_vms: the number of areas to allocate
3999 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4001 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4002 * vm_structs on success, %NULL on failure
4004 * Percpu allocator wants to use congruent vm areas so that it can
4005 * maintain the offsets among percpu areas. This function allocates
4006 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4007 * be scattered pretty far, distance between two areas easily going up
4008 * to gigabytes. To avoid interacting with regular vmallocs, these
4009 * areas are allocated from top.
4011 * Despite its complicated look, this allocator is rather simple. It
4012 * does everything top-down and scans free blocks from the end looking
4013 * for matching base. While scanning, if any of the areas do not fit the
4014 * base address is pulled down to fit the area. Scanning is repeated till
4015 * all the areas fit and then all necessary data structures are inserted
4016 * and the result is returned.
4018 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
4019 const size_t *sizes
, int nr_vms
,
4022 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
4023 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
4024 struct vmap_area
**vas
, *va
;
4025 struct vm_struct
**vms
;
4026 int area
, area2
, last_area
, term_area
;
4027 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
4028 bool purged
= false;
4030 /* verify parameters and allocate data structures */
4031 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
4032 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
4033 start
= offsets
[area
];
4034 end
= start
+ sizes
[area
];
4036 /* is everything aligned properly? */
4037 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
4038 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
4040 /* detect the area with the highest address */
4041 if (start
> offsets
[last_area
])
4044 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
4045 unsigned long start2
= offsets
[area2
];
4046 unsigned long end2
= start2
+ sizes
[area2
];
4048 BUG_ON(start2
< end
&& start
< end2
);
4051 last_end
= offsets
[last_area
] + sizes
[last_area
];
4053 if (vmalloc_end
- vmalloc_start
< last_end
) {
4058 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
4059 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
4063 for (area
= 0; area
< nr_vms
; area
++) {
4064 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
4065 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
4066 if (!vas
[area
] || !vms
[area
])
4070 spin_lock(&free_vmap_area_lock
);
4072 /* start scanning - we scan from the top, begin with the last area */
4073 area
= term_area
= last_area
;
4074 start
= offsets
[area
];
4075 end
= start
+ sizes
[area
];
4077 va
= pvm_find_va_enclose_addr(vmalloc_end
);
4078 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4082 * base might have underflowed, add last_end before
4085 if (base
+ last_end
< vmalloc_start
+ last_end
)
4089 * Fitting base has not been found.
4095 * If required width exceeds current VA block, move
4096 * base downwards and then recheck.
4098 if (base
+ end
> va
->va_end
) {
4099 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4105 * If this VA does not fit, move base downwards and recheck.
4107 if (base
+ start
< va
->va_start
) {
4108 va
= node_to_va(rb_prev(&va
->rb_node
));
4109 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4115 * This area fits, move on to the previous one. If
4116 * the previous one is the terminal one, we're done.
4118 area
= (area
+ nr_vms
- 1) % nr_vms
;
4119 if (area
== term_area
)
4122 start
= offsets
[area
];
4123 end
= start
+ sizes
[area
];
4124 va
= pvm_find_va_enclose_addr(base
+ end
);
4127 /* we've found a fitting base, insert all va's */
4128 for (area
= 0; area
< nr_vms
; area
++) {
4131 start
= base
+ offsets
[area
];
4134 va
= pvm_find_va_enclose_addr(start
);
4135 if (WARN_ON_ONCE(va
== NULL
))
4136 /* It is a BUG(), but trigger recovery instead. */
4139 ret
= adjust_va_to_fit_type(&free_vmap_area_root
,
4140 &free_vmap_area_list
,
4142 if (WARN_ON_ONCE(unlikely(ret
)))
4143 /* It is a BUG(), but trigger recovery instead. */
4146 /* Allocated area. */
4148 va
->va_start
= start
;
4149 va
->va_end
= start
+ size
;
4152 spin_unlock(&free_vmap_area_lock
);
4154 /* populate the kasan shadow space */
4155 for (area
= 0; area
< nr_vms
; area
++) {
4156 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
4157 goto err_free_shadow
;
4160 /* insert all vm's */
4161 spin_lock(&vmap_area_lock
);
4162 for (area
= 0; area
< nr_vms
; area
++) {
4163 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
4165 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
4168 spin_unlock(&vmap_area_lock
);
4171 * Mark allocated areas as accessible. Do it now as a best-effort
4172 * approach, as they can be mapped outside of vmalloc code.
4173 * With hardware tag-based KASAN, marking is skipped for
4174 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4176 for (area
= 0; area
< nr_vms
; area
++)
4177 vms
[area
]->addr
= kasan_unpoison_vmalloc(vms
[area
]->addr
,
4178 vms
[area
]->size
, KASAN_VMALLOC_PROT_NORMAL
);
4185 * Remove previously allocated areas. There is no
4186 * need in removing these areas from the busy tree,
4187 * because they are inserted only on the final step
4188 * and when pcpu_get_vm_areas() is success.
4191 orig_start
= vas
[area
]->va_start
;
4192 orig_end
= vas
[area
]->va_end
;
4193 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4194 &free_vmap_area_list
);
4196 kasan_release_vmalloc(orig_start
, orig_end
,
4197 va
->va_start
, va
->va_end
);
4202 spin_unlock(&free_vmap_area_lock
);
4204 reclaim_and_purge_vmap_areas();
4207 /* Before "retry", check if we recover. */
4208 for (area
= 0; area
< nr_vms
; area
++) {
4212 vas
[area
] = kmem_cache_zalloc(
4213 vmap_area_cachep
, GFP_KERNEL
);
4222 for (area
= 0; area
< nr_vms
; area
++) {
4224 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
4234 spin_lock(&free_vmap_area_lock
);
4236 * We release all the vmalloc shadows, even the ones for regions that
4237 * hadn't been successfully added. This relies on kasan_release_vmalloc
4238 * being able to tolerate this case.
4240 for (area
= 0; area
< nr_vms
; area
++) {
4241 orig_start
= vas
[area
]->va_start
;
4242 orig_end
= vas
[area
]->va_end
;
4243 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4244 &free_vmap_area_list
);
4246 kasan_release_vmalloc(orig_start
, orig_end
,
4247 va
->va_start
, va
->va_end
);
4251 spin_unlock(&free_vmap_area_lock
);
4258 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4259 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4260 * @nr_vms: the number of allocated areas
4262 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4264 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
4268 for (i
= 0; i
< nr_vms
; i
++)
4269 free_vm_area(vms
[i
]);
4272 #endif /* CONFIG_SMP */
4274 #ifdef CONFIG_PRINTK
4275 bool vmalloc_dump_obj(void *object
)
4277 struct vm_struct
*vm
;
4278 void *objp
= (void *)PAGE_ALIGN((unsigned long)object
);
4280 vm
= find_vm_area(objp
);
4283 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4284 vm
->nr_pages
, (unsigned long)vm
->addr
, vm
->caller
);
4289 #ifdef CONFIG_PROC_FS
4290 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4291 __acquires(&vmap_purge_lock
)
4292 __acquires(&vmap_area_lock
)
4294 mutex_lock(&vmap_purge_lock
);
4295 spin_lock(&vmap_area_lock
);
4297 return seq_list_start(&vmap_area_list
, *pos
);
4300 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4302 return seq_list_next(p
, &vmap_area_list
, pos
);
4305 static void s_stop(struct seq_file
*m
, void *p
)
4306 __releases(&vmap_area_lock
)
4307 __releases(&vmap_purge_lock
)
4309 spin_unlock(&vmap_area_lock
);
4310 mutex_unlock(&vmap_purge_lock
);
4313 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
4315 if (IS_ENABLED(CONFIG_NUMA
)) {
4316 unsigned int nr
, *counters
= m
->private;
4317 unsigned int step
= 1U << vm_area_page_order(v
);
4322 if (v
->flags
& VM_UNINITIALIZED
)
4324 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4327 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
4329 for (nr
= 0; nr
< v
->nr_pages
; nr
+= step
)
4330 counters
[page_to_nid(v
->pages
[nr
])] += step
;
4331 for_each_node_state(nr
, N_HIGH_MEMORY
)
4333 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
4337 static void show_purge_info(struct seq_file
*m
)
4339 struct vmap_area
*va
;
4341 spin_lock(&purge_vmap_area_lock
);
4342 list_for_each_entry(va
, &purge_vmap_area_list
, list
) {
4343 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4344 (void *)va
->va_start
, (void *)va
->va_end
,
4345 va
->va_end
- va
->va_start
);
4347 spin_unlock(&purge_vmap_area_lock
);
4350 static int s_show(struct seq_file
*m
, void *p
)
4352 struct vmap_area
*va
;
4353 struct vm_struct
*v
;
4355 va
= list_entry(p
, struct vmap_area
, list
);
4358 if (va
->flags
& VMAP_RAM
)
4359 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
4360 (void *)va
->va_start
, (void *)va
->va_end
,
4361 va
->va_end
- va
->va_start
);
4368 seq_printf(m
, "0x%pK-0x%pK %7ld",
4369 v
->addr
, v
->addr
+ v
->size
, v
->size
);
4372 seq_printf(m
, " %pS", v
->caller
);
4375 seq_printf(m
, " pages=%d", v
->nr_pages
);
4378 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
4380 if (v
->flags
& VM_IOREMAP
)
4381 seq_puts(m
, " ioremap");
4383 if (v
->flags
& VM_ALLOC
)
4384 seq_puts(m
, " vmalloc");
4386 if (v
->flags
& VM_MAP
)
4387 seq_puts(m
, " vmap");
4389 if (v
->flags
& VM_USERMAP
)
4390 seq_puts(m
, " user");
4392 if (v
->flags
& VM_DMA_COHERENT
)
4393 seq_puts(m
, " dma-coherent");
4395 if (is_vmalloc_addr(v
->pages
))
4396 seq_puts(m
, " vpages");
4398 show_numa_info(m
, v
);
4402 * As a final step, dump "unpurged" areas.
4405 if (list_is_last(&va
->list
, &vmap_area_list
))
4411 static const struct seq_operations vmalloc_op
= {
4418 static int __init
proc_vmalloc_init(void)
4420 if (IS_ENABLED(CONFIG_NUMA
))
4421 proc_create_seq_private("vmallocinfo", 0400, NULL
,
4423 nr_node_ids
* sizeof(unsigned int), NULL
);
4425 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
4428 module_init(proc_vmalloc_init
);
4432 void __init
vmalloc_init(void)
4434 struct vmap_area
*va
;
4435 struct vm_struct
*tmp
;
4439 * Create the cache for vmap_area objects.
4441 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
4443 for_each_possible_cpu(i
) {
4444 struct vmap_block_queue
*vbq
;
4445 struct vfree_deferred
*p
;
4447 vbq
= &per_cpu(vmap_block_queue
, i
);
4448 spin_lock_init(&vbq
->lock
);
4449 INIT_LIST_HEAD(&vbq
->free
);
4450 p
= &per_cpu(vfree_deferred
, i
);
4451 init_llist_head(&p
->list
);
4452 INIT_WORK(&p
->wq
, delayed_vfree_work
);
4453 xa_init(&vbq
->vmap_blocks
);
4456 /* Import existing vmlist entries. */
4457 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
4458 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
4459 if (WARN_ON_ONCE(!va
))
4462 va
->va_start
= (unsigned long)tmp
->addr
;
4463 va
->va_end
= va
->va_start
+ tmp
->size
;
4465 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
4469 * Now we can initialize a free vmap space.
4471 vmap_init_free_space();
4472 vmap_initialized
= true;