1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/bitops.h>
37 #include <linux/rbtree_augmented.h>
38 #include <linux/overflow.h>
39 #include <linux/pgtable.h>
40 #include <linux/uaccess.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/vmalloc.h>
50 #include "pgalloc-track.h"
52 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53 static unsigned int __ro_after_init ioremap_max_page_shift
= BITS_PER_LONG
- 1;
55 static int __init
set_nohugeiomap(char *str
)
57 ioremap_max_page_shift
= PAGE_SHIFT
;
60 early_param("nohugeiomap", set_nohugeiomap
);
61 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 static const unsigned int ioremap_max_page_shift
= PAGE_SHIFT
;
63 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66 static bool __ro_after_init vmap_allow_huge
= true;
68 static int __init
set_nohugevmalloc(char *str
)
70 vmap_allow_huge
= false;
73 early_param("nohugevmalloc", set_nohugevmalloc
);
74 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 static const bool vmap_allow_huge
= false;
76 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78 bool is_vmalloc_addr(const void *x
)
80 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
82 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
84 EXPORT_SYMBOL(is_vmalloc_addr
);
86 struct vfree_deferred
{
87 struct llist_head list
;
88 struct work_struct wq
;
90 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
92 /*** Page table manipulation functions ***/
93 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
94 phys_addr_t phys_addr
, pgprot_t prot
,
95 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
99 unsigned long size
= PAGE_SIZE
;
101 pfn
= phys_addr
>> PAGE_SHIFT
;
102 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
106 BUG_ON(!pte_none(*pte
));
108 #ifdef CONFIG_HUGETLB_PAGE
109 size
= arch_vmap_pte_range_map_size(addr
, end
, pfn
, max_page_shift
);
110 if (size
!= PAGE_SIZE
) {
111 pte_t entry
= pfn_pte(pfn
, prot
);
113 entry
= arch_make_huge_pte(entry
, ilog2(size
), 0);
114 set_huge_pte_at(&init_mm
, addr
, pte
, entry
);
115 pfn
+= PFN_DOWN(size
);
119 set_pte_at(&init_mm
, addr
, pte
, pfn_pte(pfn
, prot
));
121 } while (pte
+= PFN_DOWN(size
), addr
+= size
, addr
!= end
);
122 *mask
|= PGTBL_PTE_MODIFIED
;
126 static int vmap_try_huge_pmd(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
127 phys_addr_t phys_addr
, pgprot_t prot
,
128 unsigned int max_page_shift
)
130 if (max_page_shift
< PMD_SHIFT
)
133 if (!arch_vmap_pmd_supported(prot
))
136 if ((end
- addr
) != PMD_SIZE
)
139 if (!IS_ALIGNED(addr
, PMD_SIZE
))
142 if (!IS_ALIGNED(phys_addr
, PMD_SIZE
))
145 if (pmd_present(*pmd
) && !pmd_free_pte_page(pmd
, addr
))
148 return pmd_set_huge(pmd
, phys_addr
, prot
);
151 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
152 phys_addr_t phys_addr
, pgprot_t prot
,
153 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
158 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
162 next
= pmd_addr_end(addr
, end
);
164 if (vmap_try_huge_pmd(pmd
, addr
, next
, phys_addr
, prot
,
166 *mask
|= PGTBL_PMD_MODIFIED
;
170 if (vmap_pte_range(pmd
, addr
, next
, phys_addr
, prot
, max_page_shift
, mask
))
172 } while (pmd
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
176 static int vmap_try_huge_pud(pud_t
*pud
, unsigned long addr
, unsigned long end
,
177 phys_addr_t phys_addr
, pgprot_t prot
,
178 unsigned int max_page_shift
)
180 if (max_page_shift
< PUD_SHIFT
)
183 if (!arch_vmap_pud_supported(prot
))
186 if ((end
- addr
) != PUD_SIZE
)
189 if (!IS_ALIGNED(addr
, PUD_SIZE
))
192 if (!IS_ALIGNED(phys_addr
, PUD_SIZE
))
195 if (pud_present(*pud
) && !pud_free_pmd_page(pud
, addr
))
198 return pud_set_huge(pud
, phys_addr
, prot
);
201 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
202 phys_addr_t phys_addr
, pgprot_t prot
,
203 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
208 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
212 next
= pud_addr_end(addr
, end
);
214 if (vmap_try_huge_pud(pud
, addr
, next
, phys_addr
, prot
,
216 *mask
|= PGTBL_PUD_MODIFIED
;
220 if (vmap_pmd_range(pud
, addr
, next
, phys_addr
, prot
,
221 max_page_shift
, mask
))
223 } while (pud
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
227 static int vmap_try_huge_p4d(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
228 phys_addr_t phys_addr
, pgprot_t prot
,
229 unsigned int max_page_shift
)
231 if (max_page_shift
< P4D_SHIFT
)
234 if (!arch_vmap_p4d_supported(prot
))
237 if ((end
- addr
) != P4D_SIZE
)
240 if (!IS_ALIGNED(addr
, P4D_SIZE
))
243 if (!IS_ALIGNED(phys_addr
, P4D_SIZE
))
246 if (p4d_present(*p4d
) && !p4d_free_pud_page(p4d
, addr
))
249 return p4d_set_huge(p4d
, phys_addr
, prot
);
252 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
253 phys_addr_t phys_addr
, pgprot_t prot
,
254 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
259 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
263 next
= p4d_addr_end(addr
, end
);
265 if (vmap_try_huge_p4d(p4d
, addr
, next
, phys_addr
, prot
,
267 *mask
|= PGTBL_P4D_MODIFIED
;
271 if (vmap_pud_range(p4d
, addr
, next
, phys_addr
, prot
,
272 max_page_shift
, mask
))
274 } while (p4d
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
278 static int vmap_range_noflush(unsigned long addr
, unsigned long end
,
279 phys_addr_t phys_addr
, pgprot_t prot
,
280 unsigned int max_page_shift
)
286 pgtbl_mod_mask mask
= 0;
292 pgd
= pgd_offset_k(addr
);
294 next
= pgd_addr_end(addr
, end
);
295 err
= vmap_p4d_range(pgd
, addr
, next
, phys_addr
, prot
,
296 max_page_shift
, &mask
);
299 } while (pgd
++, phys_addr
+= (next
- addr
), addr
= next
, addr
!= end
);
301 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
302 arch_sync_kernel_mappings(start
, end
);
307 int ioremap_page_range(unsigned long addr
, unsigned long end
,
308 phys_addr_t phys_addr
, pgprot_t prot
)
312 err
= vmap_range_noflush(addr
, end
, phys_addr
, pgprot_nx(prot
),
313 ioremap_max_page_shift
);
314 flush_cache_vmap(addr
, end
);
316 kmsan_ioremap_page_range(addr
, end
, phys_addr
, prot
,
317 ioremap_max_page_shift
);
321 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
322 pgtbl_mod_mask
*mask
)
326 pte
= pte_offset_kernel(pmd
, addr
);
328 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
329 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
330 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
331 *mask
|= PGTBL_PTE_MODIFIED
;
334 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
335 pgtbl_mod_mask
*mask
)
341 pmd
= pmd_offset(pud
, addr
);
343 next
= pmd_addr_end(addr
, end
);
345 cleared
= pmd_clear_huge(pmd
);
346 if (cleared
|| pmd_bad(*pmd
))
347 *mask
|= PGTBL_PMD_MODIFIED
;
351 if (pmd_none_or_clear_bad(pmd
))
353 vunmap_pte_range(pmd
, addr
, next
, mask
);
356 } while (pmd
++, addr
= next
, addr
!= end
);
359 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
360 pgtbl_mod_mask
*mask
)
366 pud
= pud_offset(p4d
, addr
);
368 next
= pud_addr_end(addr
, end
);
370 cleared
= pud_clear_huge(pud
);
371 if (cleared
|| pud_bad(*pud
))
372 *mask
|= PGTBL_PUD_MODIFIED
;
376 if (pud_none_or_clear_bad(pud
))
378 vunmap_pmd_range(pud
, addr
, next
, mask
);
379 } while (pud
++, addr
= next
, addr
!= end
);
382 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
383 pgtbl_mod_mask
*mask
)
388 p4d
= p4d_offset(pgd
, addr
);
390 next
= p4d_addr_end(addr
, end
);
394 *mask
|= PGTBL_P4D_MODIFIED
;
396 if (p4d_none_or_clear_bad(p4d
))
398 vunmap_pud_range(p4d
, addr
, next
, mask
);
399 } while (p4d
++, addr
= next
, addr
!= end
);
403 * vunmap_range_noflush is similar to vunmap_range, but does not
404 * flush caches or TLBs.
406 * The caller is responsible for calling flush_cache_vmap() before calling
407 * this function, and flush_tlb_kernel_range after it has returned
408 * successfully (and before the addresses are expected to cause a page fault
409 * or be re-mapped for something else, if TLB flushes are being delayed or
412 * This is an internal function only. Do not use outside mm/.
414 void __vunmap_range_noflush(unsigned long start
, unsigned long end
)
418 unsigned long addr
= start
;
419 pgtbl_mod_mask mask
= 0;
422 pgd
= pgd_offset_k(addr
);
424 next
= pgd_addr_end(addr
, end
);
426 mask
|= PGTBL_PGD_MODIFIED
;
427 if (pgd_none_or_clear_bad(pgd
))
429 vunmap_p4d_range(pgd
, addr
, next
, &mask
);
430 } while (pgd
++, addr
= next
, addr
!= end
);
432 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
433 arch_sync_kernel_mappings(start
, end
);
436 void vunmap_range_noflush(unsigned long start
, unsigned long end
)
438 kmsan_vunmap_range_noflush(start
, end
);
439 __vunmap_range_noflush(start
, end
);
443 * vunmap_range - unmap kernel virtual addresses
444 * @addr: start of the VM area to unmap
445 * @end: end of the VM area to unmap (non-inclusive)
447 * Clears any present PTEs in the virtual address range, flushes TLBs and
448 * caches. Any subsequent access to the address before it has been re-mapped
451 void vunmap_range(unsigned long addr
, unsigned long end
)
453 flush_cache_vunmap(addr
, end
);
454 vunmap_range_noflush(addr
, end
);
455 flush_tlb_kernel_range(addr
, end
);
458 static int vmap_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
459 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
460 pgtbl_mod_mask
*mask
)
465 * nr is a running index into the array which helps higher level
466 * callers keep track of where we're up to.
469 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
473 struct page
*page
= pages
[*nr
];
475 if (WARN_ON(!pte_none(*pte
)))
479 if (WARN_ON(!pfn_valid(page_to_pfn(page
))))
482 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
484 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
485 *mask
|= PGTBL_PTE_MODIFIED
;
489 static int vmap_pages_pmd_range(pud_t
*pud
, unsigned long addr
,
490 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
491 pgtbl_mod_mask
*mask
)
496 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
500 next
= pmd_addr_end(addr
, end
);
501 if (vmap_pages_pte_range(pmd
, addr
, next
, prot
, pages
, nr
, mask
))
503 } while (pmd
++, addr
= next
, addr
!= end
);
507 static int vmap_pages_pud_range(p4d_t
*p4d
, unsigned long addr
,
508 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
509 pgtbl_mod_mask
*mask
)
514 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
518 next
= pud_addr_end(addr
, end
);
519 if (vmap_pages_pmd_range(pud
, addr
, next
, prot
, pages
, nr
, mask
))
521 } while (pud
++, addr
= next
, addr
!= end
);
525 static int vmap_pages_p4d_range(pgd_t
*pgd
, unsigned long addr
,
526 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
527 pgtbl_mod_mask
*mask
)
532 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
536 next
= p4d_addr_end(addr
, end
);
537 if (vmap_pages_pud_range(p4d
, addr
, next
, prot
, pages
, nr
, mask
))
539 } while (p4d
++, addr
= next
, addr
!= end
);
543 static int vmap_small_pages_range_noflush(unsigned long addr
, unsigned long end
,
544 pgprot_t prot
, struct page
**pages
)
546 unsigned long start
= addr
;
551 pgtbl_mod_mask mask
= 0;
554 pgd
= pgd_offset_k(addr
);
556 next
= pgd_addr_end(addr
, end
);
558 mask
|= PGTBL_PGD_MODIFIED
;
559 err
= vmap_pages_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
, &mask
);
562 } while (pgd
++, addr
= next
, addr
!= end
);
564 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
565 arch_sync_kernel_mappings(start
, end
);
571 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
574 * The caller is responsible for calling flush_cache_vmap() after this
575 * function returns successfully and before the addresses are accessed.
577 * This is an internal function only. Do not use outside mm/.
579 int __vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
580 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
582 unsigned int i
, nr
= (end
- addr
) >> PAGE_SHIFT
;
584 WARN_ON(page_shift
< PAGE_SHIFT
);
586 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC
) ||
587 page_shift
== PAGE_SHIFT
)
588 return vmap_small_pages_range_noflush(addr
, end
, prot
, pages
);
590 for (i
= 0; i
< nr
; i
+= 1U << (page_shift
- PAGE_SHIFT
)) {
593 err
= vmap_range_noflush(addr
, addr
+ (1UL << page_shift
),
594 page_to_phys(pages
[i
]), prot
,
599 addr
+= 1UL << page_shift
;
605 int vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
606 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
608 kmsan_vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
609 return __vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
613 * vmap_pages_range - map pages to a kernel virtual address
614 * @addr: start of the VM area to map
615 * @end: end of the VM area to map (non-inclusive)
616 * @prot: page protection flags to use
617 * @pages: pages to map (always PAGE_SIZE pages)
618 * @page_shift: maximum shift that the pages may be mapped with, @pages must
619 * be aligned and contiguous up to at least this shift.
622 * 0 on success, -errno on failure.
624 static int vmap_pages_range(unsigned long addr
, unsigned long end
,
625 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
629 err
= vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
630 flush_cache_vmap(addr
, end
);
634 int is_vmalloc_or_module_addr(const void *x
)
637 * ARM, x86-64 and sparc64 put modules in a special place,
638 * and fall back on vmalloc() if that fails. Others
639 * just put it in the vmalloc space.
641 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
642 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
643 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
646 return is_vmalloc_addr(x
);
648 EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr
);
651 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
652 * return the tail page that corresponds to the base page address, which
653 * matches small vmap mappings.
655 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
657 unsigned long addr
= (unsigned long) vmalloc_addr
;
658 struct page
*page
= NULL
;
659 pgd_t
*pgd
= pgd_offset_k(addr
);
666 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
667 * architectures that do not vmalloc module space
669 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
673 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
674 return NULL
; /* XXX: no allowance for huge pgd */
675 if (WARN_ON_ONCE(pgd_bad(*pgd
)))
678 p4d
= p4d_offset(pgd
, addr
);
682 return p4d_page(*p4d
) + ((addr
& ~P4D_MASK
) >> PAGE_SHIFT
);
683 if (WARN_ON_ONCE(p4d_bad(*p4d
)))
686 pud
= pud_offset(p4d
, addr
);
690 return pud_page(*pud
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
691 if (WARN_ON_ONCE(pud_bad(*pud
)))
694 pmd
= pmd_offset(pud
, addr
);
698 return pmd_page(*pmd
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
699 if (WARN_ON_ONCE(pmd_bad(*pmd
)))
702 ptep
= pte_offset_map(pmd
, addr
);
704 if (pte_present(pte
))
705 page
= pte_page(pte
);
710 EXPORT_SYMBOL(vmalloc_to_page
);
713 * Map a vmalloc()-space virtual address to the physical page frame number.
715 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
717 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
719 EXPORT_SYMBOL(vmalloc_to_pfn
);
722 /*** Global kva allocator ***/
724 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
725 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
728 static DEFINE_SPINLOCK(vmap_area_lock
);
729 static DEFINE_SPINLOCK(free_vmap_area_lock
);
730 /* Export for kexec only */
731 LIST_HEAD(vmap_area_list
);
732 static struct rb_root vmap_area_root
= RB_ROOT
;
733 static bool vmap_initialized __read_mostly
;
735 static struct rb_root purge_vmap_area_root
= RB_ROOT
;
736 static LIST_HEAD(purge_vmap_area_list
);
737 static DEFINE_SPINLOCK(purge_vmap_area_lock
);
740 * This kmem_cache is used for vmap_area objects. Instead of
741 * allocating from slab we reuse an object from this cache to
742 * make things faster. Especially in "no edge" splitting of
745 static struct kmem_cache
*vmap_area_cachep
;
748 * This linked list is used in pair with free_vmap_area_root.
749 * It gives O(1) access to prev/next to perform fast coalescing.
751 static LIST_HEAD(free_vmap_area_list
);
754 * This augment red-black tree represents the free vmap space.
755 * All vmap_area objects in this tree are sorted by va->va_start
756 * address. It is used for allocation and merging when a vmap
757 * object is released.
759 * Each vmap_area node contains a maximum available free block
760 * of its sub-tree, right or left. Therefore it is possible to
761 * find a lowest match of free area.
763 static struct rb_root free_vmap_area_root
= RB_ROOT
;
766 * Preload a CPU with one object for "no edge" split case. The
767 * aim is to get rid of allocations from the atomic context, thus
768 * to use more permissive allocation masks.
770 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
772 static __always_inline
unsigned long
773 va_size(struct vmap_area
*va
)
775 return (va
->va_end
- va
->va_start
);
778 static __always_inline
unsigned long
779 get_subtree_max_size(struct rb_node
*node
)
781 struct vmap_area
*va
;
783 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
784 return va
? va
->subtree_max_size
: 0;
787 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
788 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
790 static void purge_vmap_area_lazy(void);
791 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
792 static void drain_vmap_area_work(struct work_struct
*work
);
793 static DECLARE_WORK(drain_vmap_work
, drain_vmap_area_work
);
795 static atomic_long_t nr_vmalloc_pages
;
797 unsigned long vmalloc_nr_pages(void)
799 return atomic_long_read(&nr_vmalloc_pages
);
802 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
803 static struct vmap_area
*find_vmap_area_exceed_addr(unsigned long addr
)
805 struct vmap_area
*va
= NULL
;
806 struct rb_node
*n
= vmap_area_root
.rb_node
;
808 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
811 struct vmap_area
*tmp
;
813 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
814 if (tmp
->va_end
> addr
) {
816 if (tmp
->va_start
<= addr
)
827 static struct vmap_area
*__find_vmap_area(unsigned long addr
, struct rb_root
*root
)
829 struct rb_node
*n
= root
->rb_node
;
831 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
834 struct vmap_area
*va
;
836 va
= rb_entry(n
, struct vmap_area
, rb_node
);
837 if (addr
< va
->va_start
)
839 else if (addr
>= va
->va_end
)
849 * This function returns back addresses of parent node
850 * and its left or right link for further processing.
852 * Otherwise NULL is returned. In that case all further
853 * steps regarding inserting of conflicting overlap range
854 * have to be declined and actually considered as a bug.
856 static __always_inline
struct rb_node
**
857 find_va_links(struct vmap_area
*va
,
858 struct rb_root
*root
, struct rb_node
*from
,
859 struct rb_node
**parent
)
861 struct vmap_area
*tmp_va
;
862 struct rb_node
**link
;
865 link
= &root
->rb_node
;
866 if (unlikely(!*link
)) {
875 * Go to the bottom of the tree. When we hit the last point
876 * we end up with parent rb_node and correct direction, i name
877 * it link, where the new va->rb_node will be attached to.
880 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
883 * During the traversal we also do some sanity check.
884 * Trigger the BUG() if there are sides(left/right)
887 if (va
->va_end
<= tmp_va
->va_start
)
888 link
= &(*link
)->rb_left
;
889 else if (va
->va_start
>= tmp_va
->va_end
)
890 link
= &(*link
)->rb_right
;
892 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
893 va
->va_start
, va
->va_end
, tmp_va
->va_start
, tmp_va
->va_end
);
899 *parent
= &tmp_va
->rb_node
;
903 static __always_inline
struct list_head
*
904 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
906 struct list_head
*list
;
908 if (unlikely(!parent
))
910 * The red-black tree where we try to find VA neighbors
911 * before merging or inserting is empty, i.e. it means
912 * there is no free vmap space. Normally it does not
913 * happen but we handle this case anyway.
917 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
918 return (&parent
->rb_right
== link
? list
->next
: list
);
921 static __always_inline
void
922 __link_va(struct vmap_area
*va
, struct rb_root
*root
,
923 struct rb_node
*parent
, struct rb_node
**link
,
924 struct list_head
*head
, bool augment
)
927 * VA is still not in the list, but we can
928 * identify its future previous list_head node.
930 if (likely(parent
)) {
931 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
932 if (&parent
->rb_right
!= link
)
936 /* Insert to the rb-tree */
937 rb_link_node(&va
->rb_node
, parent
, link
);
940 * Some explanation here. Just perform simple insertion
941 * to the tree. We do not set va->subtree_max_size to
942 * its current size before calling rb_insert_augmented().
943 * It is because we populate the tree from the bottom
944 * to parent levels when the node _is_ in the tree.
946 * Therefore we set subtree_max_size to zero after insertion,
947 * to let __augment_tree_propagate_from() puts everything to
948 * the correct order later on.
950 rb_insert_augmented(&va
->rb_node
,
951 root
, &free_vmap_area_rb_augment_cb
);
952 va
->subtree_max_size
= 0;
954 rb_insert_color(&va
->rb_node
, root
);
957 /* Address-sort this list */
958 list_add(&va
->list
, head
);
961 static __always_inline
void
962 link_va(struct vmap_area
*va
, struct rb_root
*root
,
963 struct rb_node
*parent
, struct rb_node
**link
,
964 struct list_head
*head
)
966 __link_va(va
, root
, parent
, link
, head
, false);
969 static __always_inline
void
970 link_va_augment(struct vmap_area
*va
, struct rb_root
*root
,
971 struct rb_node
*parent
, struct rb_node
**link
,
972 struct list_head
*head
)
974 __link_va(va
, root
, parent
, link
, head
, true);
977 static __always_inline
void
978 __unlink_va(struct vmap_area
*va
, struct rb_root
*root
, bool augment
)
980 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
984 rb_erase_augmented(&va
->rb_node
,
985 root
, &free_vmap_area_rb_augment_cb
);
987 rb_erase(&va
->rb_node
, root
);
989 list_del_init(&va
->list
);
990 RB_CLEAR_NODE(&va
->rb_node
);
993 static __always_inline
void
994 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
996 __unlink_va(va
, root
, false);
999 static __always_inline
void
1000 unlink_va_augment(struct vmap_area
*va
, struct rb_root
*root
)
1002 __unlink_va(va
, root
, true);
1005 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1007 * Gets called when remove the node and rotate.
1009 static __always_inline
unsigned long
1010 compute_subtree_max_size(struct vmap_area
*va
)
1012 return max3(va_size(va
),
1013 get_subtree_max_size(va
->rb_node
.rb_left
),
1014 get_subtree_max_size(va
->rb_node
.rb_right
));
1018 augment_tree_propagate_check(void)
1020 struct vmap_area
*va
;
1021 unsigned long computed_size
;
1023 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
1024 computed_size
= compute_subtree_max_size(va
);
1025 if (computed_size
!= va
->subtree_max_size
)
1026 pr_emerg("tree is corrupted: %lu, %lu\n",
1027 va_size(va
), va
->subtree_max_size
);
1033 * This function populates subtree_max_size from bottom to upper
1034 * levels starting from VA point. The propagation must be done
1035 * when VA size is modified by changing its va_start/va_end. Or
1036 * in case of newly inserting of VA to the tree.
1038 * It means that __augment_tree_propagate_from() must be called:
1039 * - After VA has been inserted to the tree(free path);
1040 * - After VA has been shrunk(allocation path);
1041 * - After VA has been increased(merging path).
1043 * Please note that, it does not mean that upper parent nodes
1044 * and their subtree_max_size are recalculated all the time up
1053 * For example if we modify the node 4, shrinking it to 2, then
1054 * no any modification is required. If we shrink the node 2 to 1
1055 * its subtree_max_size is updated only, and set to 1. If we shrink
1056 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1057 * node becomes 4--6.
1059 static __always_inline
void
1060 augment_tree_propagate_from(struct vmap_area
*va
)
1063 * Populate the tree from bottom towards the root until
1064 * the calculated maximum available size of checked node
1065 * is equal to its current one.
1067 free_vmap_area_rb_augment_cb_propagate(&va
->rb_node
, NULL
);
1069 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1070 augment_tree_propagate_check();
1075 insert_vmap_area(struct vmap_area
*va
,
1076 struct rb_root
*root
, struct list_head
*head
)
1078 struct rb_node
**link
;
1079 struct rb_node
*parent
;
1081 link
= find_va_links(va
, root
, NULL
, &parent
);
1083 link_va(va
, root
, parent
, link
, head
);
1087 insert_vmap_area_augment(struct vmap_area
*va
,
1088 struct rb_node
*from
, struct rb_root
*root
,
1089 struct list_head
*head
)
1091 struct rb_node
**link
;
1092 struct rb_node
*parent
;
1095 link
= find_va_links(va
, NULL
, from
, &parent
);
1097 link
= find_va_links(va
, root
, NULL
, &parent
);
1100 link_va_augment(va
, root
, parent
, link
, head
);
1101 augment_tree_propagate_from(va
);
1106 * Merge de-allocated chunk of VA memory with previous
1107 * and next free blocks. If coalesce is not done a new
1108 * free area is inserted. If VA has been merged, it is
1111 * Please note, it can return NULL in case of overlap
1112 * ranges, followed by WARN() report. Despite it is a
1113 * buggy behaviour, a system can be alive and keep
1116 static __always_inline
struct vmap_area
*
1117 __merge_or_add_vmap_area(struct vmap_area
*va
,
1118 struct rb_root
*root
, struct list_head
*head
, bool augment
)
1120 struct vmap_area
*sibling
;
1121 struct list_head
*next
;
1122 struct rb_node
**link
;
1123 struct rb_node
*parent
;
1124 bool merged
= false;
1127 * Find a place in the tree where VA potentially will be
1128 * inserted, unless it is merged with its sibling/siblings.
1130 link
= find_va_links(va
, root
, NULL
, &parent
);
1135 * Get next node of VA to check if merging can be done.
1137 next
= get_va_next_sibling(parent
, link
);
1138 if (unlikely(next
== NULL
))
1144 * |<------VA------>|<-----Next----->|
1149 sibling
= list_entry(next
, struct vmap_area
, list
);
1150 if (sibling
->va_start
== va
->va_end
) {
1151 sibling
->va_start
= va
->va_start
;
1153 /* Free vmap_area object. */
1154 kmem_cache_free(vmap_area_cachep
, va
);
1156 /* Point to the new merged area. */
1165 * |<-----Prev----->|<------VA------>|
1169 if (next
->prev
!= head
) {
1170 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
1171 if (sibling
->va_end
== va
->va_start
) {
1173 * If both neighbors are coalesced, it is important
1174 * to unlink the "next" node first, followed by merging
1175 * with "previous" one. Otherwise the tree might not be
1176 * fully populated if a sibling's augmented value is
1177 * "normalized" because of rotation operations.
1180 __unlink_va(va
, root
, augment
);
1182 sibling
->va_end
= va
->va_end
;
1184 /* Free vmap_area object. */
1185 kmem_cache_free(vmap_area_cachep
, va
);
1187 /* Point to the new merged area. */
1195 __link_va(va
, root
, parent
, link
, head
, augment
);
1200 static __always_inline
struct vmap_area
*
1201 merge_or_add_vmap_area(struct vmap_area
*va
,
1202 struct rb_root
*root
, struct list_head
*head
)
1204 return __merge_or_add_vmap_area(va
, root
, head
, false);
1207 static __always_inline
struct vmap_area
*
1208 merge_or_add_vmap_area_augment(struct vmap_area
*va
,
1209 struct rb_root
*root
, struct list_head
*head
)
1211 va
= __merge_or_add_vmap_area(va
, root
, head
, true);
1213 augment_tree_propagate_from(va
);
1218 static __always_inline
bool
1219 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
1220 unsigned long align
, unsigned long vstart
)
1222 unsigned long nva_start_addr
;
1224 if (va
->va_start
> vstart
)
1225 nva_start_addr
= ALIGN(va
->va_start
, align
);
1227 nva_start_addr
= ALIGN(vstart
, align
);
1229 /* Can be overflowed due to big size or alignment. */
1230 if (nva_start_addr
+ size
< nva_start_addr
||
1231 nva_start_addr
< vstart
)
1234 return (nva_start_addr
+ size
<= va
->va_end
);
1238 * Find the first free block(lowest start address) in the tree,
1239 * that will accomplish the request corresponding to passing
1240 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1241 * a search length is adjusted to account for worst case alignment
1244 static __always_inline
struct vmap_area
*
1245 find_vmap_lowest_match(struct rb_root
*root
, unsigned long size
,
1246 unsigned long align
, unsigned long vstart
, bool adjust_search_size
)
1248 struct vmap_area
*va
;
1249 struct rb_node
*node
;
1250 unsigned long length
;
1252 /* Start from the root. */
1253 node
= root
->rb_node
;
1255 /* Adjust the search size for alignment overhead. */
1256 length
= adjust_search_size
? size
+ align
- 1 : size
;
1259 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1261 if (get_subtree_max_size(node
->rb_left
) >= length
&&
1262 vstart
< va
->va_start
) {
1263 node
= node
->rb_left
;
1265 if (is_within_this_va(va
, size
, align
, vstart
))
1269 * Does not make sense to go deeper towards the right
1270 * sub-tree if it does not have a free block that is
1271 * equal or bigger to the requested search length.
1273 if (get_subtree_max_size(node
->rb_right
) >= length
) {
1274 node
= node
->rb_right
;
1279 * OK. We roll back and find the first right sub-tree,
1280 * that will satisfy the search criteria. It can happen
1281 * due to "vstart" restriction or an alignment overhead
1282 * that is bigger then PAGE_SIZE.
1284 while ((node
= rb_parent(node
))) {
1285 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1286 if (is_within_this_va(va
, size
, align
, vstart
))
1289 if (get_subtree_max_size(node
->rb_right
) >= length
&&
1290 vstart
<= va
->va_start
) {
1292 * Shift the vstart forward. Please note, we update it with
1293 * parent's start address adding "1" because we do not want
1294 * to enter same sub-tree after it has already been checked
1295 * and no suitable free block found there.
1297 vstart
= va
->va_start
+ 1;
1298 node
= node
->rb_right
;
1308 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1309 #include <linux/random.h>
1311 static struct vmap_area
*
1312 find_vmap_lowest_linear_match(struct list_head
*head
, unsigned long size
,
1313 unsigned long align
, unsigned long vstart
)
1315 struct vmap_area
*va
;
1317 list_for_each_entry(va
, head
, list
) {
1318 if (!is_within_this_va(va
, size
, align
, vstart
))
1328 find_vmap_lowest_match_check(struct rb_root
*root
, struct list_head
*head
,
1329 unsigned long size
, unsigned long align
)
1331 struct vmap_area
*va_1
, *va_2
;
1332 unsigned long vstart
;
1335 get_random_bytes(&rnd
, sizeof(rnd
));
1336 vstart
= VMALLOC_START
+ rnd
;
1338 va_1
= find_vmap_lowest_match(root
, size
, align
, vstart
, false);
1339 va_2
= find_vmap_lowest_linear_match(head
, size
, align
, vstart
);
1342 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1343 va_1
, va_2
, vstart
);
1349 FL_FIT_TYPE
= 1, /* full fit */
1350 LE_FIT_TYPE
= 2, /* left edge fit */
1351 RE_FIT_TYPE
= 3, /* right edge fit */
1352 NE_FIT_TYPE
= 4 /* no edge fit */
1355 static __always_inline
enum fit_type
1356 classify_va_fit_type(struct vmap_area
*va
,
1357 unsigned long nva_start_addr
, unsigned long size
)
1361 /* Check if it is within VA. */
1362 if (nva_start_addr
< va
->va_start
||
1363 nva_start_addr
+ size
> va
->va_end
)
1367 if (va
->va_start
== nva_start_addr
) {
1368 if (va
->va_end
== nva_start_addr
+ size
)
1372 } else if (va
->va_end
== nva_start_addr
+ size
) {
1381 static __always_inline
int
1382 adjust_va_to_fit_type(struct rb_root
*root
, struct list_head
*head
,
1383 struct vmap_area
*va
, unsigned long nva_start_addr
,
1386 struct vmap_area
*lva
= NULL
;
1387 enum fit_type type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1389 if (type
== FL_FIT_TYPE
) {
1391 * No need to split VA, it fully fits.
1397 unlink_va_augment(va
, root
);
1398 kmem_cache_free(vmap_area_cachep
, va
);
1399 } else if (type
== LE_FIT_TYPE
) {
1401 * Split left edge of fit VA.
1407 va
->va_start
+= size
;
1408 } else if (type
== RE_FIT_TYPE
) {
1410 * Split right edge of fit VA.
1416 va
->va_end
= nva_start_addr
;
1417 } else if (type
== NE_FIT_TYPE
) {
1419 * Split no edge of fit VA.
1425 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1426 if (unlikely(!lva
)) {
1428 * For percpu allocator we do not do any pre-allocation
1429 * and leave it as it is. The reason is it most likely
1430 * never ends up with NE_FIT_TYPE splitting. In case of
1431 * percpu allocations offsets and sizes are aligned to
1432 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1433 * are its main fitting cases.
1435 * There are a few exceptions though, as an example it is
1436 * a first allocation (early boot up) when we have "one"
1437 * big free space that has to be split.
1439 * Also we can hit this path in case of regular "vmap"
1440 * allocations, if "this" current CPU was not preloaded.
1441 * See the comment in alloc_vmap_area() why. If so, then
1442 * GFP_NOWAIT is used instead to get an extra object for
1443 * split purpose. That is rare and most time does not
1446 * What happens if an allocation gets failed. Basically,
1447 * an "overflow" path is triggered to purge lazily freed
1448 * areas to free some memory, then, the "retry" path is
1449 * triggered to repeat one more time. See more details
1450 * in alloc_vmap_area() function.
1452 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1458 * Build the remainder.
1460 lva
->va_start
= va
->va_start
;
1461 lva
->va_end
= nva_start_addr
;
1464 * Shrink this VA to remaining size.
1466 va
->va_start
= nva_start_addr
+ size
;
1471 if (type
!= FL_FIT_TYPE
) {
1472 augment_tree_propagate_from(va
);
1474 if (lva
) /* type == NE_FIT_TYPE */
1475 insert_vmap_area_augment(lva
, &va
->rb_node
, root
, head
);
1482 * Returns a start address of the newly allocated area, if success.
1483 * Otherwise a vend is returned that indicates failure.
1485 static __always_inline
unsigned long
1486 __alloc_vmap_area(struct rb_root
*root
, struct list_head
*head
,
1487 unsigned long size
, unsigned long align
,
1488 unsigned long vstart
, unsigned long vend
)
1490 bool adjust_search_size
= true;
1491 unsigned long nva_start_addr
;
1492 struct vmap_area
*va
;
1496 * Do not adjust when:
1497 * a) align <= PAGE_SIZE, because it does not make any sense.
1498 * All blocks(their start addresses) are at least PAGE_SIZE
1500 * b) a short range where a requested size corresponds to exactly
1501 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1502 * With adjusted search length an allocation would not succeed.
1504 if (align
<= PAGE_SIZE
|| (align
> PAGE_SIZE
&& (vend
- vstart
) == size
))
1505 adjust_search_size
= false;
1507 va
= find_vmap_lowest_match(root
, size
, align
, vstart
, adjust_search_size
);
1511 if (va
->va_start
> vstart
)
1512 nva_start_addr
= ALIGN(va
->va_start
, align
);
1514 nva_start_addr
= ALIGN(vstart
, align
);
1516 /* Check the "vend" restriction. */
1517 if (nva_start_addr
+ size
> vend
)
1520 /* Update the free vmap_area. */
1521 ret
= adjust_va_to_fit_type(root
, head
, va
, nva_start_addr
, size
);
1522 if (WARN_ON_ONCE(ret
))
1525 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1526 find_vmap_lowest_match_check(root
, head
, size
, align
);
1529 return nva_start_addr
;
1533 * Free a region of KVA allocated by alloc_vmap_area
1535 static void free_vmap_area(struct vmap_area
*va
)
1538 * Remove from the busy tree/list.
1540 spin_lock(&vmap_area_lock
);
1541 unlink_va(va
, &vmap_area_root
);
1542 spin_unlock(&vmap_area_lock
);
1545 * Insert/Merge it back to the free tree/list.
1547 spin_lock(&free_vmap_area_lock
);
1548 merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1549 spin_unlock(&free_vmap_area_lock
);
1553 preload_this_cpu_lock(spinlock_t
*lock
, gfp_t gfp_mask
, int node
)
1555 struct vmap_area
*va
= NULL
;
1558 * Preload this CPU with one extra vmap_area object. It is used
1559 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1560 * a CPU that does an allocation is preloaded.
1562 * We do it in non-atomic context, thus it allows us to use more
1563 * permissive allocation masks to be more stable under low memory
1564 * condition and high memory pressure.
1566 if (!this_cpu_read(ne_fit_preload_node
))
1567 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1571 if (va
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, va
))
1572 kmem_cache_free(vmap_area_cachep
, va
);
1576 * Allocate a region of KVA of the specified size and alignment, within the
1579 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1580 unsigned long align
,
1581 unsigned long vstart
, unsigned long vend
,
1582 int node
, gfp_t gfp_mask
,
1583 unsigned long va_flags
)
1585 struct vmap_area
*va
;
1586 unsigned long freed
;
1591 if (unlikely(!size
|| offset_in_page(size
) || !is_power_of_2(align
)))
1592 return ERR_PTR(-EINVAL
);
1594 if (unlikely(!vmap_initialized
))
1595 return ERR_PTR(-EBUSY
);
1598 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1600 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1602 return ERR_PTR(-ENOMEM
);
1605 * Only scan the relevant parts containing pointers to other objects
1606 * to avoid false negatives.
1608 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1611 preload_this_cpu_lock(&free_vmap_area_lock
, gfp_mask
, node
);
1612 addr
= __alloc_vmap_area(&free_vmap_area_root
, &free_vmap_area_list
,
1613 size
, align
, vstart
, vend
);
1614 spin_unlock(&free_vmap_area_lock
);
1616 trace_alloc_vmap_area(addr
, size
, align
, vstart
, vend
, addr
== vend
);
1619 * If an allocation fails, the "vend" address is
1620 * returned. Therefore trigger the overflow path.
1622 if (unlikely(addr
== vend
))
1625 va
->va_start
= addr
;
1626 va
->va_end
= addr
+ size
;
1628 va
->flags
= va_flags
;
1630 spin_lock(&vmap_area_lock
);
1631 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
1632 spin_unlock(&vmap_area_lock
);
1634 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1635 BUG_ON(va
->va_start
< vstart
);
1636 BUG_ON(va
->va_end
> vend
);
1638 ret
= kasan_populate_vmalloc(addr
, size
);
1641 return ERR_PTR(ret
);
1648 purge_vmap_area_lazy();
1654 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
1661 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
1662 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1665 kmem_cache_free(vmap_area_cachep
, va
);
1666 return ERR_PTR(-EBUSY
);
1669 int register_vmap_purge_notifier(struct notifier_block
*nb
)
1671 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
1673 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
1675 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
1677 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
1679 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
1682 * lazy_max_pages is the maximum amount of virtual address space we gather up
1683 * before attempting to purge with a TLB flush.
1685 * There is a tradeoff here: a larger number will cover more kernel page tables
1686 * and take slightly longer to purge, but it will linearly reduce the number of
1687 * global TLB flushes that must be performed. It would seem natural to scale
1688 * this number up linearly with the number of CPUs (because vmapping activity
1689 * could also scale linearly with the number of CPUs), however it is likely
1690 * that in practice, workloads might be constrained in other ways that mean
1691 * vmap activity will not scale linearly with CPUs. Also, I want to be
1692 * conservative and not introduce a big latency on huge systems, so go with
1693 * a less aggressive log scale. It will still be an improvement over the old
1694 * code, and it will be simple to change the scale factor if we find that it
1695 * becomes a problem on bigger systems.
1697 static unsigned long lazy_max_pages(void)
1701 log
= fls(num_online_cpus());
1703 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
1706 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
1709 * Serialize vmap purging. There is no actual critical section protected
1710 * by this lock, but we want to avoid concurrent calls for performance
1711 * reasons and to make the pcpu_get_vm_areas more deterministic.
1713 static DEFINE_MUTEX(vmap_purge_lock
);
1715 /* for per-CPU blocks */
1716 static void purge_fragmented_blocks_allcpus(void);
1719 * Purges all lazily-freed vmap areas.
1721 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
1723 unsigned long resched_threshold
;
1724 unsigned int num_purged_areas
= 0;
1725 struct list_head local_purge_list
;
1726 struct vmap_area
*va
, *n_va
;
1728 lockdep_assert_held(&vmap_purge_lock
);
1730 spin_lock(&purge_vmap_area_lock
);
1731 purge_vmap_area_root
= RB_ROOT
;
1732 list_replace_init(&purge_vmap_area_list
, &local_purge_list
);
1733 spin_unlock(&purge_vmap_area_lock
);
1735 if (unlikely(list_empty(&local_purge_list
)))
1739 list_first_entry(&local_purge_list
,
1740 struct vmap_area
, list
)->va_start
);
1743 list_last_entry(&local_purge_list
,
1744 struct vmap_area
, list
)->va_end
);
1746 flush_tlb_kernel_range(start
, end
);
1747 resched_threshold
= lazy_max_pages() << 1;
1749 spin_lock(&free_vmap_area_lock
);
1750 list_for_each_entry_safe(va
, n_va
, &local_purge_list
, list
) {
1751 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
1752 unsigned long orig_start
= va
->va_start
;
1753 unsigned long orig_end
= va
->va_end
;
1756 * Finally insert or merge lazily-freed area. It is
1757 * detached and there is no need to "unlink" it from
1760 va
= merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
,
1761 &free_vmap_area_list
);
1766 if (is_vmalloc_or_module_addr((void *)orig_start
))
1767 kasan_release_vmalloc(orig_start
, orig_end
,
1768 va
->va_start
, va
->va_end
);
1770 atomic_long_sub(nr
, &vmap_lazy_nr
);
1773 if (atomic_long_read(&vmap_lazy_nr
) < resched_threshold
)
1774 cond_resched_lock(&free_vmap_area_lock
);
1776 spin_unlock(&free_vmap_area_lock
);
1779 trace_purge_vmap_area_lazy(start
, end
, num_purged_areas
);
1780 return num_purged_areas
> 0;
1784 * Kick off a purge of the outstanding lazy areas.
1786 static void purge_vmap_area_lazy(void)
1788 mutex_lock(&vmap_purge_lock
);
1789 purge_fragmented_blocks_allcpus();
1790 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1791 mutex_unlock(&vmap_purge_lock
);
1794 static void drain_vmap_area_work(struct work_struct
*work
)
1796 unsigned long nr_lazy
;
1799 mutex_lock(&vmap_purge_lock
);
1800 __purge_vmap_area_lazy(ULONG_MAX
, 0);
1801 mutex_unlock(&vmap_purge_lock
);
1803 /* Recheck if further work is required. */
1804 nr_lazy
= atomic_long_read(&vmap_lazy_nr
);
1805 } while (nr_lazy
> lazy_max_pages());
1809 * Free a vmap area, caller ensuring that the area has been unmapped,
1810 * unlinked and flush_cache_vunmap had been called for the correct
1813 static void free_vmap_area_noflush(struct vmap_area
*va
)
1815 unsigned long nr_lazy_max
= lazy_max_pages();
1816 unsigned long va_start
= va
->va_start
;
1817 unsigned long nr_lazy
;
1819 if (WARN_ON_ONCE(!list_empty(&va
->list
)))
1822 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
1823 PAGE_SHIFT
, &vmap_lazy_nr
);
1826 * Merge or place it to the purge tree/list.
1828 spin_lock(&purge_vmap_area_lock
);
1829 merge_or_add_vmap_area(va
,
1830 &purge_vmap_area_root
, &purge_vmap_area_list
);
1831 spin_unlock(&purge_vmap_area_lock
);
1833 trace_free_vmap_area_noflush(va_start
, nr_lazy
, nr_lazy_max
);
1835 /* After this point, we may free va at any time */
1836 if (unlikely(nr_lazy
> nr_lazy_max
))
1837 schedule_work(&drain_vmap_work
);
1841 * Free and unmap a vmap area
1843 static void free_unmap_vmap_area(struct vmap_area
*va
)
1845 flush_cache_vunmap(va
->va_start
, va
->va_end
);
1846 vunmap_range_noflush(va
->va_start
, va
->va_end
);
1847 if (debug_pagealloc_enabled_static())
1848 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
1850 free_vmap_area_noflush(va
);
1853 struct vmap_area
*find_vmap_area(unsigned long addr
)
1855 struct vmap_area
*va
;
1857 spin_lock(&vmap_area_lock
);
1858 va
= __find_vmap_area(addr
, &vmap_area_root
);
1859 spin_unlock(&vmap_area_lock
);
1864 static struct vmap_area
*find_unlink_vmap_area(unsigned long addr
)
1866 struct vmap_area
*va
;
1868 spin_lock(&vmap_area_lock
);
1869 va
= __find_vmap_area(addr
, &vmap_area_root
);
1871 unlink_va(va
, &vmap_area_root
);
1872 spin_unlock(&vmap_area_lock
);
1877 /*** Per cpu kva allocator ***/
1880 * vmap space is limited especially on 32 bit architectures. Ensure there is
1881 * room for at least 16 percpu vmap blocks per CPU.
1884 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1885 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1886 * instead (we just need a rough idea)
1888 #if BITS_PER_LONG == 32
1889 #define VMALLOC_SPACE (128UL*1024*1024)
1891 #define VMALLOC_SPACE (128UL*1024*1024*1024)
1894 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1895 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1896 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1897 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1898 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1899 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1900 #define VMAP_BBMAP_BITS \
1901 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1902 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1903 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1905 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1907 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
1908 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
1909 #define VMAP_FLAGS_MASK 0x3
1911 struct vmap_block_queue
{
1913 struct list_head free
;
1918 struct vmap_area
*va
;
1919 unsigned long free
, dirty
;
1920 DECLARE_BITMAP(used_map
, VMAP_BBMAP_BITS
);
1921 unsigned long dirty_min
, dirty_max
; /*< dirty range */
1922 struct list_head free_list
;
1923 struct rcu_head rcu_head
;
1924 struct list_head purge
;
1927 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1928 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
1931 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1932 * in the free path. Could get rid of this if we change the API to return a
1933 * "cookie" from alloc, to be passed to free. But no big deal yet.
1935 static DEFINE_XARRAY(vmap_blocks
);
1938 * We should probably have a fallback mechanism to allocate virtual memory
1939 * out of partially filled vmap blocks. However vmap block sizing should be
1940 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1944 static unsigned long addr_to_vb_idx(unsigned long addr
)
1946 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
1947 addr
/= VMAP_BLOCK_SIZE
;
1951 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
1955 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
1956 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
1957 return (void *)addr
;
1961 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1962 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1963 * @order: how many 2^order pages should be occupied in newly allocated block
1964 * @gfp_mask: flags for the page level allocator
1966 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1968 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
1970 struct vmap_block_queue
*vbq
;
1971 struct vmap_block
*vb
;
1972 struct vmap_area
*va
;
1973 unsigned long vb_idx
;
1977 node
= numa_node_id();
1979 vb
= kmalloc_node(sizeof(struct vmap_block
),
1980 gfp_mask
& GFP_RECLAIM_MASK
, node
);
1982 return ERR_PTR(-ENOMEM
);
1984 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
1985 VMALLOC_START
, VMALLOC_END
,
1987 VMAP_RAM
|VMAP_BLOCK
);
1990 return ERR_CAST(va
);
1993 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
1994 spin_lock_init(&vb
->lock
);
1996 /* At least something should be left free */
1997 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
1998 bitmap_zero(vb
->used_map
, VMAP_BBMAP_BITS
);
1999 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
2001 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2003 bitmap_set(vb
->used_map
, 0, (1UL << order
));
2004 INIT_LIST_HEAD(&vb
->free_list
);
2006 vb_idx
= addr_to_vb_idx(va
->va_start
);
2007 err
= xa_insert(&vmap_blocks
, vb_idx
, vb
, gfp_mask
);
2011 return ERR_PTR(err
);
2014 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2015 spin_lock(&vbq
->lock
);
2016 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
2017 spin_unlock(&vbq
->lock
);
2022 static void free_vmap_block(struct vmap_block
*vb
)
2024 struct vmap_block
*tmp
;
2026 tmp
= xa_erase(&vmap_blocks
, addr_to_vb_idx(vb
->va
->va_start
));
2029 spin_lock(&vmap_area_lock
);
2030 unlink_va(vb
->va
, &vmap_area_root
);
2031 spin_unlock(&vmap_area_lock
);
2033 free_vmap_area_noflush(vb
->va
);
2034 kfree_rcu(vb
, rcu_head
);
2037 static void purge_fragmented_blocks(int cpu
)
2040 struct vmap_block
*vb
;
2041 struct vmap_block
*n_vb
;
2042 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2045 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2047 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
2050 spin_lock(&vb
->lock
);
2051 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2052 vb
->free
= 0; /* prevent further allocs after releasing lock */
2053 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
2055 vb
->dirty_max
= VMAP_BBMAP_BITS
;
2056 spin_lock(&vbq
->lock
);
2057 list_del_rcu(&vb
->free_list
);
2058 spin_unlock(&vbq
->lock
);
2059 spin_unlock(&vb
->lock
);
2060 list_add_tail(&vb
->purge
, &purge
);
2062 spin_unlock(&vb
->lock
);
2066 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
2067 list_del(&vb
->purge
);
2068 free_vmap_block(vb
);
2072 static void purge_fragmented_blocks_allcpus(void)
2076 for_each_possible_cpu(cpu
)
2077 purge_fragmented_blocks(cpu
);
2080 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
2082 struct vmap_block_queue
*vbq
;
2083 struct vmap_block
*vb
;
2087 BUG_ON(offset_in_page(size
));
2088 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2089 if (WARN_ON(size
== 0)) {
2091 * Allocating 0 bytes isn't what caller wants since
2092 * get_order(0) returns funny result. Just warn and terminate
2097 order
= get_order(size
);
2100 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2101 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2102 unsigned long pages_off
;
2104 spin_lock(&vb
->lock
);
2105 if (vb
->free
< (1UL << order
)) {
2106 spin_unlock(&vb
->lock
);
2110 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
2111 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
2112 vb
->free
-= 1UL << order
;
2113 bitmap_set(vb
->used_map
, pages_off
, (1UL << order
));
2114 if (vb
->free
== 0) {
2115 spin_lock(&vbq
->lock
);
2116 list_del_rcu(&vb
->free_list
);
2117 spin_unlock(&vbq
->lock
);
2120 spin_unlock(&vb
->lock
);
2126 /* Allocate new block if nothing was found */
2128 vaddr
= new_vmap_block(order
, gfp_mask
);
2133 static void vb_free(unsigned long addr
, unsigned long size
)
2135 unsigned long offset
;
2137 struct vmap_block
*vb
;
2139 BUG_ON(offset_in_page(size
));
2140 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2142 flush_cache_vunmap(addr
, addr
+ size
);
2144 order
= get_order(size
);
2145 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
2146 vb
= xa_load(&vmap_blocks
, addr_to_vb_idx(addr
));
2147 spin_lock(&vb
->lock
);
2148 bitmap_clear(vb
->used_map
, offset
, (1UL << order
));
2149 spin_unlock(&vb
->lock
);
2151 vunmap_range_noflush(addr
, addr
+ size
);
2153 if (debug_pagealloc_enabled_static())
2154 flush_tlb_kernel_range(addr
, addr
+ size
);
2156 spin_lock(&vb
->lock
);
2158 /* Expand dirty range */
2159 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
2160 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
2162 vb
->dirty
+= 1UL << order
;
2163 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
2165 spin_unlock(&vb
->lock
);
2166 free_vmap_block(vb
);
2168 spin_unlock(&vb
->lock
);
2171 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
2175 if (unlikely(!vmap_initialized
))
2180 for_each_possible_cpu(cpu
) {
2181 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2182 struct vmap_block
*vb
;
2185 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2186 spin_lock(&vb
->lock
);
2187 if (vb
->dirty
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2188 unsigned long va_start
= vb
->va
->va_start
;
2191 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
2192 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
2194 start
= min(s
, start
);
2199 spin_unlock(&vb
->lock
);
2204 mutex_lock(&vmap_purge_lock
);
2205 purge_fragmented_blocks_allcpus();
2206 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
2207 flush_tlb_kernel_range(start
, end
);
2208 mutex_unlock(&vmap_purge_lock
);
2212 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2214 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2215 * to amortize TLB flushing overheads. What this means is that any page you
2216 * have now, may, in a former life, have been mapped into kernel virtual
2217 * address by the vmap layer and so there might be some CPUs with TLB entries
2218 * still referencing that page (additional to the regular 1:1 kernel mapping).
2220 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2221 * be sure that none of the pages we have control over will have any aliases
2222 * from the vmap layer.
2224 void vm_unmap_aliases(void)
2226 unsigned long start
= ULONG_MAX
, end
= 0;
2229 _vm_unmap_aliases(start
, end
, flush
);
2231 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
2234 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2235 * @mem: the pointer returned by vm_map_ram
2236 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2238 void vm_unmap_ram(const void *mem
, unsigned int count
)
2240 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2241 unsigned long addr
= (unsigned long)kasan_reset_tag(mem
);
2242 struct vmap_area
*va
;
2246 BUG_ON(addr
< VMALLOC_START
);
2247 BUG_ON(addr
> VMALLOC_END
);
2248 BUG_ON(!PAGE_ALIGNED(addr
));
2250 kasan_poison_vmalloc(mem
, size
);
2252 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2253 debug_check_no_locks_freed(mem
, size
);
2254 vb_free(addr
, size
);
2258 va
= find_unlink_vmap_area(addr
);
2259 if (WARN_ON_ONCE(!va
))
2262 debug_check_no_locks_freed((void *)va
->va_start
,
2263 (va
->va_end
- va
->va_start
));
2264 free_unmap_vmap_area(va
);
2266 EXPORT_SYMBOL(vm_unmap_ram
);
2269 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2270 * @pages: an array of pointers to the pages to be mapped
2271 * @count: number of pages
2272 * @node: prefer to allocate data structures on this node
2274 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2275 * faster than vmap so it's good. But if you mix long-life and short-life
2276 * objects with vm_map_ram(), it could consume lots of address space through
2277 * fragmentation (especially on a 32bit machine). You could see failures in
2278 * the end. Please use this function for short-lived objects.
2280 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2282 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
2284 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2288 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2289 mem
= vb_alloc(size
, GFP_KERNEL
);
2292 addr
= (unsigned long)mem
;
2294 struct vmap_area
*va
;
2295 va
= alloc_vmap_area(size
, PAGE_SIZE
,
2296 VMALLOC_START
, VMALLOC_END
,
2297 node
, GFP_KERNEL
, VMAP_RAM
);
2301 addr
= va
->va_start
;
2305 if (vmap_pages_range(addr
, addr
+ size
, PAGE_KERNEL
,
2306 pages
, PAGE_SHIFT
) < 0) {
2307 vm_unmap_ram(mem
, count
);
2312 * Mark the pages as accessible, now that they are mapped.
2313 * With hardware tag-based KASAN, marking is skipped for
2314 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2316 mem
= kasan_unpoison_vmalloc(mem
, size
, KASAN_VMALLOC_PROT_NORMAL
);
2320 EXPORT_SYMBOL(vm_map_ram
);
2322 static struct vm_struct
*vmlist __initdata
;
2324 static inline unsigned int vm_area_page_order(struct vm_struct
*vm
)
2326 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2327 return vm
->page_order
;
2333 static inline void set_vm_area_page_order(struct vm_struct
*vm
, unsigned int order
)
2335 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2336 vm
->page_order
= order
;
2343 * vm_area_add_early - add vmap area early during boot
2344 * @vm: vm_struct to add
2346 * This function is used to add fixed kernel vm area to vmlist before
2347 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2348 * should contain proper values and the other fields should be zero.
2350 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2352 void __init
vm_area_add_early(struct vm_struct
*vm
)
2354 struct vm_struct
*tmp
, **p
;
2356 BUG_ON(vmap_initialized
);
2357 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
2358 if (tmp
->addr
>= vm
->addr
) {
2359 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
2362 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
2369 * vm_area_register_early - register vmap area early during boot
2370 * @vm: vm_struct to register
2371 * @align: requested alignment
2373 * This function is used to register kernel vm area before
2374 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2375 * proper values on entry and other fields should be zero. On return,
2376 * vm->addr contains the allocated address.
2378 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2380 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
2382 unsigned long addr
= ALIGN(VMALLOC_START
, align
);
2383 struct vm_struct
*cur
, **p
;
2385 BUG_ON(vmap_initialized
);
2387 for (p
= &vmlist
; (cur
= *p
) != NULL
; p
= &cur
->next
) {
2388 if ((unsigned long)cur
->addr
- addr
>= vm
->size
)
2390 addr
= ALIGN((unsigned long)cur
->addr
+ cur
->size
, align
);
2393 BUG_ON(addr
> VMALLOC_END
- vm
->size
);
2394 vm
->addr
= (void *)addr
;
2397 kasan_populate_early_vm_area_shadow(vm
->addr
, vm
->size
);
2400 static void vmap_init_free_space(void)
2402 unsigned long vmap_start
= 1;
2403 const unsigned long vmap_end
= ULONG_MAX
;
2404 struct vmap_area
*busy
, *free
;
2408 * -|-----|.....|-----|-----|-----|.....|-
2410 * |<--------------------------------->|
2412 list_for_each_entry(busy
, &vmap_area_list
, list
) {
2413 if (busy
->va_start
- vmap_start
> 0) {
2414 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2415 if (!WARN_ON_ONCE(!free
)) {
2416 free
->va_start
= vmap_start
;
2417 free
->va_end
= busy
->va_start
;
2419 insert_vmap_area_augment(free
, NULL
,
2420 &free_vmap_area_root
,
2421 &free_vmap_area_list
);
2425 vmap_start
= busy
->va_end
;
2428 if (vmap_end
- vmap_start
> 0) {
2429 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
2430 if (!WARN_ON_ONCE(!free
)) {
2431 free
->va_start
= vmap_start
;
2432 free
->va_end
= vmap_end
;
2434 insert_vmap_area_augment(free
, NULL
,
2435 &free_vmap_area_root
,
2436 &free_vmap_area_list
);
2441 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
2442 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
2445 vm
->addr
= (void *)va
->va_start
;
2446 vm
->size
= va
->va_end
- va
->va_start
;
2447 vm
->caller
= caller
;
2451 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
2452 unsigned long flags
, const void *caller
)
2454 spin_lock(&vmap_area_lock
);
2455 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
2456 spin_unlock(&vmap_area_lock
);
2459 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
2462 * Before removing VM_UNINITIALIZED,
2463 * we should make sure that vm has proper values.
2464 * Pair with smp_rmb() in show_numa_info().
2467 vm
->flags
&= ~VM_UNINITIALIZED
;
2470 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
2471 unsigned long align
, unsigned long shift
, unsigned long flags
,
2472 unsigned long start
, unsigned long end
, int node
,
2473 gfp_t gfp_mask
, const void *caller
)
2475 struct vmap_area
*va
;
2476 struct vm_struct
*area
;
2477 unsigned long requested_size
= size
;
2479 BUG_ON(in_interrupt());
2480 size
= ALIGN(size
, 1ul << shift
);
2481 if (unlikely(!size
))
2484 if (flags
& VM_IOREMAP
)
2485 align
= 1ul << clamp_t(int, get_count_order_long(size
),
2486 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
2488 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
2489 if (unlikely(!area
))
2492 if (!(flags
& VM_NO_GUARD
))
2495 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
, 0);
2501 setup_vmalloc_vm(area
, va
, flags
, caller
);
2504 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2505 * best-effort approach, as they can be mapped outside of vmalloc code.
2506 * For VM_ALLOC mappings, the pages are marked as accessible after
2507 * getting mapped in __vmalloc_node_range().
2508 * With hardware tag-based KASAN, marking is skipped for
2509 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2511 if (!(flags
& VM_ALLOC
))
2512 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, requested_size
,
2513 KASAN_VMALLOC_PROT_NORMAL
);
2518 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
2519 unsigned long start
, unsigned long end
,
2522 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
, start
, end
,
2523 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2527 * get_vm_area - reserve a contiguous kernel virtual area
2528 * @size: size of the area
2529 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2531 * Search an area of @size in the kernel virtual mapping area,
2532 * and reserved it for out purposes. Returns the area descriptor
2533 * on success or %NULL on failure.
2535 * Return: the area descriptor on success or %NULL on failure.
2537 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
2539 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2540 VMALLOC_START
, VMALLOC_END
,
2541 NUMA_NO_NODE
, GFP_KERNEL
,
2542 __builtin_return_address(0));
2545 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
2548 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
2549 VMALLOC_START
, VMALLOC_END
,
2550 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
2554 * find_vm_area - find a continuous kernel virtual area
2555 * @addr: base address
2557 * Search for the kernel VM area starting at @addr, and return it.
2558 * It is up to the caller to do all required locking to keep the returned
2561 * Return: the area descriptor on success or %NULL on failure.
2563 struct vm_struct
*find_vm_area(const void *addr
)
2565 struct vmap_area
*va
;
2567 va
= find_vmap_area((unsigned long)addr
);
2575 * remove_vm_area - find and remove a continuous kernel virtual area
2576 * @addr: base address
2578 * Search for the kernel VM area starting at @addr, and remove it.
2579 * This function returns the found VM area, but using it is NOT safe
2580 * on SMP machines, except for its size or flags.
2582 * Return: the area descriptor on success or %NULL on failure.
2584 struct vm_struct
*remove_vm_area(const void *addr
)
2586 struct vmap_area
*va
;
2587 struct vm_struct
*vm
;
2591 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
2595 va
= find_unlink_vmap_area((unsigned long)addr
);
2600 debug_check_no_locks_freed(vm
->addr
, get_vm_area_size(vm
));
2601 debug_check_no_obj_freed(vm
->addr
, get_vm_area_size(vm
));
2602 kasan_free_module_shadow(vm
);
2603 kasan_poison_vmalloc(vm
->addr
, get_vm_area_size(vm
));
2605 free_unmap_vmap_area(va
);
2609 static inline void set_area_direct_map(const struct vm_struct
*area
,
2610 int (*set_direct_map
)(struct page
*page
))
2614 /* HUGE_VMALLOC passes small pages to set_direct_map */
2615 for (i
= 0; i
< area
->nr_pages
; i
++)
2616 if (page_address(area
->pages
[i
]))
2617 set_direct_map(area
->pages
[i
]);
2621 * Flush the vm mapping and reset the direct map.
2623 static void vm_reset_perms(struct vm_struct
*area
)
2625 unsigned long start
= ULONG_MAX
, end
= 0;
2626 unsigned int page_order
= vm_area_page_order(area
);
2631 * Find the start and end range of the direct mappings to make sure that
2632 * the vm_unmap_aliases() flush includes the direct map.
2634 for (i
= 0; i
< area
->nr_pages
; i
+= 1U << page_order
) {
2635 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
2638 unsigned long page_size
;
2640 page_size
= PAGE_SIZE
<< page_order
;
2641 start
= min(addr
, start
);
2642 end
= max(addr
+ page_size
, end
);
2648 * Set direct map to something invalid so that it won't be cached if
2649 * there are any accesses after the TLB flush, then flush the TLB and
2650 * reset the direct map permissions to the default.
2652 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
2653 _vm_unmap_aliases(start
, end
, flush_dmap
);
2654 set_area_direct_map(area
, set_direct_map_default_noflush
);
2657 static void delayed_vfree_work(struct work_struct
*w
)
2659 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
2660 struct llist_node
*t
, *llnode
;
2662 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
2667 * vfree_atomic - release memory allocated by vmalloc()
2668 * @addr: memory base address
2670 * This one is just like vfree() but can be called in any atomic context
2673 void vfree_atomic(const void *addr
)
2675 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
2678 kmemleak_free(addr
);
2681 * Use raw_cpu_ptr() because this can be called from preemptible
2682 * context. Preemption is absolutely fine here, because the llist_add()
2683 * implementation is lockless, so it works even if we are adding to
2684 * another cpu's list. schedule_work() should be fine with this too.
2686 if (addr
&& llist_add((struct llist_node
*)addr
, &p
->list
))
2687 schedule_work(&p
->wq
);
2691 * vfree - Release memory allocated by vmalloc()
2692 * @addr: Memory base address
2694 * Free the virtually continuous memory area starting at @addr, as obtained
2695 * from one of the vmalloc() family of APIs. This will usually also free the
2696 * physical memory underlying the virtual allocation, but that memory is
2697 * reference counted, so it will not be freed until the last user goes away.
2699 * If @addr is NULL, no operation is performed.
2702 * May sleep if called *not* from interrupt context.
2703 * Must not be called in NMI context (strictly speaking, it could be
2704 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2705 * conventions for vfree() arch-dependent would be a really bad idea).
2707 void vfree(const void *addr
)
2709 struct vm_struct
*vm
;
2712 if (unlikely(in_interrupt())) {
2718 kmemleak_free(addr
);
2724 vm
= remove_vm_area(addr
);
2725 if (unlikely(!vm
)) {
2726 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
2731 if (unlikely(vm
->flags
& VM_FLUSH_RESET_PERMS
))
2733 for (i
= 0; i
< vm
->nr_pages
; i
++) {
2734 struct page
*page
= vm
->pages
[i
];
2737 mod_memcg_page_state(page
, MEMCG_VMALLOC
, -1);
2739 * High-order allocs for huge vmallocs are split, so
2740 * can be freed as an array of order-0 allocations
2742 __free_pages(page
, 0);
2745 atomic_long_sub(vm
->nr_pages
, &nr_vmalloc_pages
);
2749 EXPORT_SYMBOL(vfree
);
2752 * vunmap - release virtual mapping obtained by vmap()
2753 * @addr: memory base address
2755 * Free the virtually contiguous memory area starting at @addr,
2756 * which was created from the page array passed to vmap().
2758 * Must not be called in interrupt context.
2760 void vunmap(const void *addr
)
2762 struct vm_struct
*vm
;
2764 BUG_ON(in_interrupt());
2769 vm
= remove_vm_area(addr
);
2770 if (unlikely(!vm
)) {
2771 WARN(1, KERN_ERR
"Trying to vunmap() nonexistent vm area (%p)\n",
2777 EXPORT_SYMBOL(vunmap
);
2780 * vmap - map an array of pages into virtually contiguous space
2781 * @pages: array of page pointers
2782 * @count: number of pages to map
2783 * @flags: vm_area->flags
2784 * @prot: page protection for the mapping
2786 * Maps @count pages from @pages into contiguous kernel virtual space.
2787 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2788 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2789 * are transferred from the caller to vmap(), and will be freed / dropped when
2790 * vfree() is called on the return value.
2792 * Return: the address of the area or %NULL on failure
2794 void *vmap(struct page
**pages
, unsigned int count
,
2795 unsigned long flags
, pgprot_t prot
)
2797 struct vm_struct
*area
;
2799 unsigned long size
; /* In bytes */
2803 if (WARN_ON_ONCE(flags
& VM_FLUSH_RESET_PERMS
))
2807 * Your top guard is someone else's bottom guard. Not having a top
2808 * guard compromises someone else's mappings too.
2810 if (WARN_ON_ONCE(flags
& VM_NO_GUARD
))
2811 flags
&= ~VM_NO_GUARD
;
2813 if (count
> totalram_pages())
2816 size
= (unsigned long)count
<< PAGE_SHIFT
;
2817 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
2821 addr
= (unsigned long)area
->addr
;
2822 if (vmap_pages_range(addr
, addr
+ size
, pgprot_nx(prot
),
2823 pages
, PAGE_SHIFT
) < 0) {
2828 if (flags
& VM_MAP_PUT_PAGES
) {
2829 area
->pages
= pages
;
2830 area
->nr_pages
= count
;
2834 EXPORT_SYMBOL(vmap
);
2836 #ifdef CONFIG_VMAP_PFN
2837 struct vmap_pfn_data
{
2838 unsigned long *pfns
;
2843 static int vmap_pfn_apply(pte_t
*pte
, unsigned long addr
, void *private)
2845 struct vmap_pfn_data
*data
= private;
2847 if (WARN_ON_ONCE(pfn_valid(data
->pfns
[data
->idx
])))
2849 *pte
= pte_mkspecial(pfn_pte(data
->pfns
[data
->idx
++], data
->prot
));
2854 * vmap_pfn - map an array of PFNs into virtually contiguous space
2855 * @pfns: array of PFNs
2856 * @count: number of pages to map
2857 * @prot: page protection for the mapping
2859 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2860 * the start address of the mapping.
2862 void *vmap_pfn(unsigned long *pfns
, unsigned int count
, pgprot_t prot
)
2864 struct vmap_pfn_data data
= { .pfns
= pfns
, .prot
= pgprot_nx(prot
) };
2865 struct vm_struct
*area
;
2867 area
= get_vm_area_caller(count
* PAGE_SIZE
, VM_IOREMAP
,
2868 __builtin_return_address(0));
2871 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2872 count
* PAGE_SIZE
, vmap_pfn_apply
, &data
)) {
2878 EXPORT_SYMBOL_GPL(vmap_pfn
);
2879 #endif /* CONFIG_VMAP_PFN */
2881 static inline unsigned int
2882 vm_area_alloc_pages(gfp_t gfp
, int nid
,
2883 unsigned int order
, unsigned int nr_pages
, struct page
**pages
)
2885 unsigned int nr_allocated
= 0;
2890 * For order-0 pages we make use of bulk allocator, if
2891 * the page array is partly or not at all populated due
2892 * to fails, fallback to a single page allocator that is
2896 gfp_t bulk_gfp
= gfp
& ~__GFP_NOFAIL
;
2898 while (nr_allocated
< nr_pages
) {
2899 unsigned int nr
, nr_pages_request
;
2902 * A maximum allowed request is hard-coded and is 100
2903 * pages per call. That is done in order to prevent a
2904 * long preemption off scenario in the bulk-allocator
2905 * so the range is [1:100].
2907 nr_pages_request
= min(100U, nr_pages
- nr_allocated
);
2909 /* memory allocation should consider mempolicy, we can't
2910 * wrongly use nearest node when nid == NUMA_NO_NODE,
2911 * otherwise memory may be allocated in only one node,
2912 * but mempolicy wants to alloc memory by interleaving.
2914 if (IS_ENABLED(CONFIG_NUMA
) && nid
== NUMA_NO_NODE
)
2915 nr
= alloc_pages_bulk_array_mempolicy(bulk_gfp
,
2917 pages
+ nr_allocated
);
2920 nr
= alloc_pages_bulk_array_node(bulk_gfp
, nid
,
2922 pages
+ nr_allocated
);
2928 * If zero or pages were obtained partly,
2929 * fallback to a single page allocator.
2931 if (nr
!= nr_pages_request
)
2936 /* High-order pages or fallback path if "bulk" fails. */
2938 while (nr_allocated
< nr_pages
) {
2939 if (fatal_signal_pending(current
))
2942 if (nid
== NUMA_NO_NODE
)
2943 page
= alloc_pages(gfp
, order
);
2945 page
= alloc_pages_node(nid
, gfp
, order
);
2946 if (unlikely(!page
))
2949 * Higher order allocations must be able to be treated as
2950 * indepdenent small pages by callers (as they can with
2951 * small-page vmallocs). Some drivers do their own refcounting
2952 * on vmalloc_to_page() pages, some use page->mapping,
2956 split_page(page
, order
);
2959 * Careful, we allocate and map page-order pages, but
2960 * tracking is done per PAGE_SIZE page so as to keep the
2961 * vm_struct APIs independent of the physical/mapped size.
2963 for (i
= 0; i
< (1U << order
); i
++)
2964 pages
[nr_allocated
+ i
] = page
+ i
;
2967 nr_allocated
+= 1U << order
;
2970 return nr_allocated
;
2973 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
2974 pgprot_t prot
, unsigned int page_shift
,
2977 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
2978 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
2979 unsigned long addr
= (unsigned long)area
->addr
;
2980 unsigned long size
= get_vm_area_size(area
);
2981 unsigned long array_size
;
2982 unsigned int nr_small_pages
= size
>> PAGE_SHIFT
;
2983 unsigned int page_order
;
2987 array_size
= (unsigned long)nr_small_pages
* sizeof(struct page
*);
2989 if (!(gfp_mask
& (GFP_DMA
| GFP_DMA32
)))
2990 gfp_mask
|= __GFP_HIGHMEM
;
2992 /* Please note that the recursion is strictly bounded. */
2993 if (array_size
> PAGE_SIZE
) {
2994 area
->pages
= __vmalloc_node(array_size
, 1, nested_gfp
, node
,
2997 area
->pages
= kmalloc_node(array_size
, nested_gfp
, node
);
3001 warn_alloc(gfp_mask
, NULL
,
3002 "vmalloc error: size %lu, failed to allocated page array size %lu",
3003 nr_small_pages
* PAGE_SIZE
, array_size
);
3008 set_vm_area_page_order(area
, page_shift
- PAGE_SHIFT
);
3009 page_order
= vm_area_page_order(area
);
3011 area
->nr_pages
= vm_area_alloc_pages(gfp_mask
| __GFP_NOWARN
,
3012 node
, page_order
, nr_small_pages
, area
->pages
);
3014 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
3015 if (gfp_mask
& __GFP_ACCOUNT
) {
3018 for (i
= 0; i
< area
->nr_pages
; i
++)
3019 mod_memcg_page_state(area
->pages
[i
], MEMCG_VMALLOC
, 1);
3023 * If not enough pages were obtained to accomplish an
3024 * allocation request, free them via vfree() if any.
3026 if (area
->nr_pages
!= nr_small_pages
) {
3027 warn_alloc(gfp_mask
, NULL
,
3028 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3029 area
->nr_pages
* PAGE_SIZE
, page_order
);
3034 * page tables allocations ignore external gfp mask, enforce it
3037 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3038 flags
= memalloc_nofs_save();
3039 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3040 flags
= memalloc_noio_save();
3043 ret
= vmap_pages_range(addr
, addr
+ size
, prot
, area
->pages
,
3045 if (nofail
&& (ret
< 0))
3046 schedule_timeout_uninterruptible(1);
3047 } while (nofail
&& (ret
< 0));
3049 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3050 memalloc_nofs_restore(flags
);
3051 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3052 memalloc_noio_restore(flags
);
3055 warn_alloc(gfp_mask
, NULL
,
3056 "vmalloc error: size %lu, failed to map pages",
3057 area
->nr_pages
* PAGE_SIZE
);
3069 * __vmalloc_node_range - allocate virtually contiguous memory
3070 * @size: allocation size
3071 * @align: desired alignment
3072 * @start: vm area range start
3073 * @end: vm area range end
3074 * @gfp_mask: flags for the page level allocator
3075 * @prot: protection mask for the allocated pages
3076 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3077 * @node: node to use for allocation or NUMA_NO_NODE
3078 * @caller: caller's return address
3080 * Allocate enough pages to cover @size from the page level
3081 * allocator with @gfp_mask flags. Please note that the full set of gfp
3082 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3084 * Zone modifiers are not supported. From the reclaim modifiers
3085 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3086 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3087 * __GFP_RETRY_MAYFAIL are not supported).
3089 * __GFP_NOWARN can be used to suppress failures messages.
3091 * Map them into contiguous kernel virtual space, using a pagetable
3092 * protection of @prot.
3094 * Return: the address of the area or %NULL on failure
3096 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
3097 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
3098 pgprot_t prot
, unsigned long vm_flags
, int node
,
3101 struct vm_struct
*area
;
3103 kasan_vmalloc_flags_t kasan_flags
= KASAN_VMALLOC_NONE
;
3104 unsigned long real_size
= size
;
3105 unsigned long real_align
= align
;
3106 unsigned int shift
= PAGE_SHIFT
;
3108 if (WARN_ON_ONCE(!size
))
3111 if ((size
>> PAGE_SHIFT
) > totalram_pages()) {
3112 warn_alloc(gfp_mask
, NULL
,
3113 "vmalloc error: size %lu, exceeds total pages",
3118 if (vmap_allow_huge
&& (vm_flags
& VM_ALLOW_HUGE_VMAP
)) {
3119 unsigned long size_per_node
;
3122 * Try huge pages. Only try for PAGE_KERNEL allocations,
3123 * others like modules don't yet expect huge pages in
3124 * their allocations due to apply_to_page_range not
3128 size_per_node
= size
;
3129 if (node
== NUMA_NO_NODE
)
3130 size_per_node
/= num_online_nodes();
3131 if (arch_vmap_pmd_supported(prot
) && size_per_node
>= PMD_SIZE
)
3134 shift
= arch_vmap_pte_supported_shift(size_per_node
);
3136 align
= max(real_align
, 1UL << shift
);
3137 size
= ALIGN(real_size
, 1UL << shift
);
3141 area
= __get_vm_area_node(real_size
, align
, shift
, VM_ALLOC
|
3142 VM_UNINITIALIZED
| vm_flags
, start
, end
, node
,
3145 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3146 warn_alloc(gfp_mask
, NULL
,
3147 "vmalloc error: size %lu, vm_struct allocation failed%s",
3148 real_size
, (nofail
) ? ". Retrying." : "");
3150 schedule_timeout_uninterruptible(1);
3157 * Prepare arguments for __vmalloc_area_node() and
3158 * kasan_unpoison_vmalloc().
3160 if (pgprot_val(prot
) == pgprot_val(PAGE_KERNEL
)) {
3161 if (kasan_hw_tags_enabled()) {
3163 * Modify protection bits to allow tagging.
3164 * This must be done before mapping.
3166 prot
= arch_vmap_pgprot_tagged(prot
);
3169 * Skip page_alloc poisoning and zeroing for physical
3170 * pages backing VM_ALLOC mapping. Memory is instead
3171 * poisoned and zeroed by kasan_unpoison_vmalloc().
3173 gfp_mask
|= __GFP_SKIP_KASAN_UNPOISON
| __GFP_SKIP_ZERO
;
3176 /* Take note that the mapping is PAGE_KERNEL. */
3177 kasan_flags
|= KASAN_VMALLOC_PROT_NORMAL
;
3180 /* Allocate physical pages and map them into vmalloc space. */
3181 ret
= __vmalloc_area_node(area
, gfp_mask
, prot
, shift
, node
);
3186 * Mark the pages as accessible, now that they are mapped.
3187 * The condition for setting KASAN_VMALLOC_INIT should complement the
3188 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3189 * to make sure that memory is initialized under the same conditions.
3190 * Tag-based KASAN modes only assign tags to normal non-executable
3191 * allocations, see __kasan_unpoison_vmalloc().
3193 kasan_flags
|= KASAN_VMALLOC_VM_ALLOC
;
3194 if (!want_init_on_free() && want_init_on_alloc(gfp_mask
) &&
3195 (gfp_mask
& __GFP_SKIP_ZERO
))
3196 kasan_flags
|= KASAN_VMALLOC_INIT
;
3197 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3198 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, real_size
, kasan_flags
);
3201 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3202 * flag. It means that vm_struct is not fully initialized.
3203 * Now, it is fully initialized, so remove this flag here.
3205 clear_vm_uninitialized_flag(area
);
3207 size
= PAGE_ALIGN(size
);
3208 if (!(vm_flags
& VM_DEFER_KMEMLEAK
))
3209 kmemleak_vmalloc(area
, size
, gfp_mask
);
3214 if (shift
> PAGE_SHIFT
) {
3225 * __vmalloc_node - allocate virtually contiguous memory
3226 * @size: allocation size
3227 * @align: desired alignment
3228 * @gfp_mask: flags for the page level allocator
3229 * @node: node to use for allocation or NUMA_NO_NODE
3230 * @caller: caller's return address
3232 * Allocate enough pages to cover @size from the page level allocator with
3233 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3235 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3236 * and __GFP_NOFAIL are not supported
3238 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3241 * Return: pointer to the allocated memory or %NULL on error
3243 void *__vmalloc_node(unsigned long size
, unsigned long align
,
3244 gfp_t gfp_mask
, int node
, const void *caller
)
3246 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
3247 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
3250 * This is only for performance analysis of vmalloc and stress purpose.
3251 * It is required by vmalloc test module, therefore do not use it other
3254 #ifdef CONFIG_TEST_VMALLOC_MODULE
3255 EXPORT_SYMBOL_GPL(__vmalloc_node
);
3258 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
3260 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
3261 __builtin_return_address(0));
3263 EXPORT_SYMBOL(__vmalloc
);
3266 * vmalloc - allocate virtually contiguous memory
3267 * @size: allocation size
3269 * Allocate enough pages to cover @size from the page level
3270 * allocator and map them into contiguous kernel virtual space.
3272 * For tight control over page level allocator and protection flags
3273 * use __vmalloc() instead.
3275 * Return: pointer to the allocated memory or %NULL on error
3277 void *vmalloc(unsigned long size
)
3279 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
3280 __builtin_return_address(0));
3282 EXPORT_SYMBOL(vmalloc
);
3285 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3286 * @size: allocation size
3287 * @gfp_mask: flags for the page level allocator
3289 * Allocate enough pages to cover @size from the page level
3290 * allocator and map them into contiguous kernel virtual space.
3291 * If @size is greater than or equal to PMD_SIZE, allow using
3292 * huge pages for the memory
3294 * Return: pointer to the allocated memory or %NULL on error
3296 void *vmalloc_huge(unsigned long size
, gfp_t gfp_mask
)
3298 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
3299 gfp_mask
, PAGE_KERNEL
, VM_ALLOW_HUGE_VMAP
,
3300 NUMA_NO_NODE
, __builtin_return_address(0));
3302 EXPORT_SYMBOL_GPL(vmalloc_huge
);
3305 * vzalloc - allocate virtually contiguous memory with zero fill
3306 * @size: allocation size
3308 * Allocate enough pages to cover @size from the page level
3309 * allocator and map them into contiguous kernel virtual space.
3310 * The memory allocated is set to zero.
3312 * For tight control over page level allocator and protection flags
3313 * use __vmalloc() instead.
3315 * Return: pointer to the allocated memory or %NULL on error
3317 void *vzalloc(unsigned long size
)
3319 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
3320 __builtin_return_address(0));
3322 EXPORT_SYMBOL(vzalloc
);
3325 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3326 * @size: allocation size
3328 * The resulting memory area is zeroed so it can be mapped to userspace
3329 * without leaking data.
3331 * Return: pointer to the allocated memory or %NULL on error
3333 void *vmalloc_user(unsigned long size
)
3335 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3336 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
3337 VM_USERMAP
, NUMA_NO_NODE
,
3338 __builtin_return_address(0));
3340 EXPORT_SYMBOL(vmalloc_user
);
3343 * vmalloc_node - allocate memory on a specific node
3344 * @size: allocation size
3347 * Allocate enough pages to cover @size from the page level
3348 * allocator and map them into contiguous kernel virtual space.
3350 * For tight control over page level allocator and protection flags
3351 * use __vmalloc() instead.
3353 * Return: pointer to the allocated memory or %NULL on error
3355 void *vmalloc_node(unsigned long size
, int node
)
3357 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
3358 __builtin_return_address(0));
3360 EXPORT_SYMBOL(vmalloc_node
);
3363 * vzalloc_node - allocate memory on a specific node with zero fill
3364 * @size: allocation size
3367 * Allocate enough pages to cover @size from the page level
3368 * allocator and map them into contiguous kernel virtual space.
3369 * The memory allocated is set to zero.
3371 * Return: pointer to the allocated memory or %NULL on error
3373 void *vzalloc_node(unsigned long size
, int node
)
3375 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
3376 __builtin_return_address(0));
3378 EXPORT_SYMBOL(vzalloc_node
);
3380 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3381 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3382 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3383 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3386 * 64b systems should always have either DMA or DMA32 zones. For others
3387 * GFP_DMA32 should do the right thing and use the normal zone.
3389 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3393 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3394 * @size: allocation size
3396 * Allocate enough 32bit PA addressable pages to cover @size from the
3397 * page level allocator and map them into contiguous kernel virtual space.
3399 * Return: pointer to the allocated memory or %NULL on error
3401 void *vmalloc_32(unsigned long size
)
3403 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
3404 __builtin_return_address(0));
3406 EXPORT_SYMBOL(vmalloc_32
);
3409 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3410 * @size: allocation size
3412 * The resulting memory area is 32bit addressable and zeroed so it can be
3413 * mapped to userspace without leaking data.
3415 * Return: pointer to the allocated memory or %NULL on error
3417 void *vmalloc_32_user(unsigned long size
)
3419 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3420 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
3421 VM_USERMAP
, NUMA_NO_NODE
,
3422 __builtin_return_address(0));
3424 EXPORT_SYMBOL(vmalloc_32_user
);
3427 * small helper routine , copy contents to buf from addr.
3428 * If the page is not present, fill zero.
3431 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
3437 unsigned long offset
, length
;
3439 offset
= offset_in_page(addr
);
3440 length
= PAGE_SIZE
- offset
;
3443 p
= vmalloc_to_page(addr
);
3445 * To do safe access to this _mapped_ area, we need
3446 * lock. But adding lock here means that we need to add
3447 * overhead of vmalloc()/vfree() calls for this _debug_
3448 * interface, rarely used. Instead of that, we'll use
3449 * kmap() and get small overhead in this access function.
3452 /* We can expect USER0 is not used -- see vread() */
3453 void *map
= kmap_atomic(p
);
3454 memcpy(buf
, map
+ offset
, length
);
3457 memset(buf
, 0, length
);
3467 static void vmap_ram_vread(char *buf
, char *addr
, int count
, unsigned long flags
)
3470 struct vmap_block
*vb
;
3471 unsigned long offset
;
3472 unsigned int rs
, re
, n
;
3475 * If it's area created by vm_map_ram() interface directly, but
3476 * not further subdividing and delegating management to vmap_block,
3479 if (!(flags
& VMAP_BLOCK
)) {
3480 aligned_vread(buf
, addr
, count
);
3485 * Area is split into regions and tracked with vmap_block, read out
3486 * each region and zero fill the hole between regions.
3488 vb
= xa_load(&vmap_blocks
, addr_to_vb_idx((unsigned long)addr
));
3492 spin_lock(&vb
->lock
);
3493 if (bitmap_empty(vb
->used_map
, VMAP_BBMAP_BITS
)) {
3494 spin_unlock(&vb
->lock
);
3497 for_each_set_bitrange(rs
, re
, vb
->used_map
, VMAP_BBMAP_BITS
) {
3500 start
= vmap_block_vaddr(vb
->va
->va_start
, rs
);
3501 while (addr
< start
) {
3509 /*it could start reading from the middle of used region*/
3510 offset
= offset_in_page(addr
);
3511 n
= ((re
- rs
+ 1) << PAGE_SHIFT
) - offset
;
3514 aligned_vread(buf
, start
+offset
, n
);
3521 spin_unlock(&vb
->lock
);
3524 /* zero-fill the left dirty or free regions */
3526 memset(buf
, 0, count
);
3530 * vread() - read vmalloc area in a safe way.
3531 * @buf: buffer for reading data
3532 * @addr: vm address.
3533 * @count: number of bytes to be read.
3535 * This function checks that addr is a valid vmalloc'ed area, and
3536 * copy data from that area to a given buffer. If the given memory range
3537 * of [addr...addr+count) includes some valid address, data is copied to
3538 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3539 * IOREMAP area is treated as memory hole and no copy is done.
3541 * If [addr...addr+count) doesn't includes any intersects with alive
3542 * vm_struct area, returns 0. @buf should be kernel's buffer.
3544 * Note: In usual ops, vread() is never necessary because the caller
3545 * should know vmalloc() area is valid and can use memcpy().
3546 * This is for routines which have to access vmalloc area without
3547 * any information, as /proc/kcore.
3549 * Return: number of bytes for which addr and buf should be increased
3550 * (same number as @count) or %0 if [addr...addr+count) doesn't
3551 * include any intersection with valid vmalloc area
3553 long vread(char *buf
, char *addr
, unsigned long count
)
3555 struct vmap_area
*va
;
3556 struct vm_struct
*vm
;
3557 char *vaddr
, *buf_start
= buf
;
3558 unsigned long buflen
= count
;
3559 unsigned long n
, size
, flags
;
3561 addr
= kasan_reset_tag(addr
);
3563 /* Don't allow overflow */
3564 if ((unsigned long) addr
+ count
< count
)
3565 count
= -(unsigned long) addr
;
3567 spin_lock(&vmap_area_lock
);
3568 va
= find_vmap_area_exceed_addr((unsigned long)addr
);
3572 /* no intersects with alive vmap_area */
3573 if ((unsigned long)addr
+ count
<= va
->va_start
)
3576 list_for_each_entry_from(va
, &vmap_area_list
, list
) {
3581 flags
= va
->flags
& VMAP_FLAGS_MASK
;
3583 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
3584 * be set together with VMAP_RAM.
3586 WARN_ON(flags
== VMAP_BLOCK
);
3591 if (vm
&& (vm
->flags
& VM_UNINITIALIZED
))
3593 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3596 vaddr
= (char *) va
->va_start
;
3597 size
= vm
? get_vm_area_size(vm
) : va_size(va
);
3599 if (addr
>= vaddr
+ size
)
3601 while (addr
< vaddr
) {
3609 n
= vaddr
+ size
- addr
;
3613 if (flags
& VMAP_RAM
)
3614 vmap_ram_vread(buf
, addr
, n
, flags
);
3615 else if (!(vm
->flags
& VM_IOREMAP
))
3616 aligned_vread(buf
, addr
, n
);
3617 else /* IOREMAP area is treated as memory hole */
3624 spin_unlock(&vmap_area_lock
);
3626 if (buf
== buf_start
)
3628 /* zero-fill memory holes */
3629 if (buf
!= buf_start
+ buflen
)
3630 memset(buf
, 0, buflen
- (buf
- buf_start
));
3636 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3637 * @vma: vma to cover
3638 * @uaddr: target user address to start at
3639 * @kaddr: virtual address of vmalloc kernel memory
3640 * @pgoff: offset from @kaddr to start at
3641 * @size: size of map area
3643 * Returns: 0 for success, -Exxx on failure
3645 * This function checks that @kaddr is a valid vmalloc'ed area,
3646 * and that it is big enough to cover the range starting at
3647 * @uaddr in @vma. Will return failure if that criteria isn't
3650 * Similar to remap_pfn_range() (see mm/memory.c)
3652 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
3653 void *kaddr
, unsigned long pgoff
,
3656 struct vm_struct
*area
;
3658 unsigned long end_index
;
3660 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
3663 size
= PAGE_ALIGN(size
);
3665 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
3668 area
= find_vm_area(kaddr
);
3672 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
3675 if (check_add_overflow(size
, off
, &end_index
) ||
3676 end_index
> get_vm_area_size(area
))
3681 struct page
*page
= vmalloc_to_page(kaddr
);
3684 ret
= vm_insert_page(vma
, uaddr
, page
);
3693 vm_flags_set(vma
, VM_DONTEXPAND
| VM_DONTDUMP
);
3699 * remap_vmalloc_range - map vmalloc pages to userspace
3700 * @vma: vma to cover (map full range of vma)
3701 * @addr: vmalloc memory
3702 * @pgoff: number of pages into addr before first page to map
3704 * Returns: 0 for success, -Exxx on failure
3706 * This function checks that addr is a valid vmalloc'ed area, and
3707 * that it is big enough to cover the vma. Will return failure if
3708 * that criteria isn't met.
3710 * Similar to remap_pfn_range() (see mm/memory.c)
3712 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
3713 unsigned long pgoff
)
3715 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
3717 vma
->vm_end
- vma
->vm_start
);
3719 EXPORT_SYMBOL(remap_vmalloc_range
);
3721 void free_vm_area(struct vm_struct
*area
)
3723 struct vm_struct
*ret
;
3724 ret
= remove_vm_area(area
->addr
);
3725 BUG_ON(ret
!= area
);
3728 EXPORT_SYMBOL_GPL(free_vm_area
);
3731 static struct vmap_area
*node_to_va(struct rb_node
*n
)
3733 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
3737 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3738 * @addr: target address
3740 * Returns: vmap_area if it is found. If there is no such area
3741 * the first highest(reverse order) vmap_area is returned
3742 * i.e. va->va_start < addr && va->va_end < addr or NULL
3743 * if there are no any areas before @addr.
3745 static struct vmap_area
*
3746 pvm_find_va_enclose_addr(unsigned long addr
)
3748 struct vmap_area
*va
, *tmp
;
3751 n
= free_vmap_area_root
.rb_node
;
3755 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
3756 if (tmp
->va_start
<= addr
) {
3758 if (tmp
->va_end
>= addr
)
3771 * pvm_determine_end_from_reverse - find the highest aligned address
3772 * of free block below VMALLOC_END
3774 * in - the VA we start the search(reverse order);
3775 * out - the VA with the highest aligned end address.
3776 * @align: alignment for required highest address
3778 * Returns: determined end address within vmap_area
3780 static unsigned long
3781 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
3783 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3787 list_for_each_entry_from_reverse((*va
),
3788 &free_vmap_area_list
, list
) {
3789 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
3790 if ((*va
)->va_start
< addr
)
3799 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3800 * @offsets: array containing offset of each area
3801 * @sizes: array containing size of each area
3802 * @nr_vms: the number of areas to allocate
3803 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3805 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3806 * vm_structs on success, %NULL on failure
3808 * Percpu allocator wants to use congruent vm areas so that it can
3809 * maintain the offsets among percpu areas. This function allocates
3810 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3811 * be scattered pretty far, distance between two areas easily going up
3812 * to gigabytes. To avoid interacting with regular vmallocs, these
3813 * areas are allocated from top.
3815 * Despite its complicated look, this allocator is rather simple. It
3816 * does everything top-down and scans free blocks from the end looking
3817 * for matching base. While scanning, if any of the areas do not fit the
3818 * base address is pulled down to fit the area. Scanning is repeated till
3819 * all the areas fit and then all necessary data structures are inserted
3820 * and the result is returned.
3822 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
3823 const size_t *sizes
, int nr_vms
,
3826 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
3827 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
3828 struct vmap_area
**vas
, *va
;
3829 struct vm_struct
**vms
;
3830 int area
, area2
, last_area
, term_area
;
3831 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
3832 bool purged
= false;
3834 /* verify parameters and allocate data structures */
3835 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
3836 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
3837 start
= offsets
[area
];
3838 end
= start
+ sizes
[area
];
3840 /* is everything aligned properly? */
3841 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
3842 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
3844 /* detect the area with the highest address */
3845 if (start
> offsets
[last_area
])
3848 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
3849 unsigned long start2
= offsets
[area2
];
3850 unsigned long end2
= start2
+ sizes
[area2
];
3852 BUG_ON(start2
< end
&& start
< end2
);
3855 last_end
= offsets
[last_area
] + sizes
[last_area
];
3857 if (vmalloc_end
- vmalloc_start
< last_end
) {
3862 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
3863 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
3867 for (area
= 0; area
< nr_vms
; area
++) {
3868 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
3869 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
3870 if (!vas
[area
] || !vms
[area
])
3874 spin_lock(&free_vmap_area_lock
);
3876 /* start scanning - we scan from the top, begin with the last area */
3877 area
= term_area
= last_area
;
3878 start
= offsets
[area
];
3879 end
= start
+ sizes
[area
];
3881 va
= pvm_find_va_enclose_addr(vmalloc_end
);
3882 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3886 * base might have underflowed, add last_end before
3889 if (base
+ last_end
< vmalloc_start
+ last_end
)
3893 * Fitting base has not been found.
3899 * If required width exceeds current VA block, move
3900 * base downwards and then recheck.
3902 if (base
+ end
> va
->va_end
) {
3903 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3909 * If this VA does not fit, move base downwards and recheck.
3911 if (base
+ start
< va
->va_start
) {
3912 va
= node_to_va(rb_prev(&va
->rb_node
));
3913 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
3919 * This area fits, move on to the previous one. If
3920 * the previous one is the terminal one, we're done.
3922 area
= (area
+ nr_vms
- 1) % nr_vms
;
3923 if (area
== term_area
)
3926 start
= offsets
[area
];
3927 end
= start
+ sizes
[area
];
3928 va
= pvm_find_va_enclose_addr(base
+ end
);
3931 /* we've found a fitting base, insert all va's */
3932 for (area
= 0; area
< nr_vms
; area
++) {
3935 start
= base
+ offsets
[area
];
3938 va
= pvm_find_va_enclose_addr(start
);
3939 if (WARN_ON_ONCE(va
== NULL
))
3940 /* It is a BUG(), but trigger recovery instead. */
3943 ret
= adjust_va_to_fit_type(&free_vmap_area_root
,
3944 &free_vmap_area_list
,
3946 if (WARN_ON_ONCE(unlikely(ret
)))
3947 /* It is a BUG(), but trigger recovery instead. */
3950 /* Allocated area. */
3952 va
->va_start
= start
;
3953 va
->va_end
= start
+ size
;
3956 spin_unlock(&free_vmap_area_lock
);
3958 /* populate the kasan shadow space */
3959 for (area
= 0; area
< nr_vms
; area
++) {
3960 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
3961 goto err_free_shadow
;
3964 /* insert all vm's */
3965 spin_lock(&vmap_area_lock
);
3966 for (area
= 0; area
< nr_vms
; area
++) {
3967 insert_vmap_area(vas
[area
], &vmap_area_root
, &vmap_area_list
);
3969 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
3972 spin_unlock(&vmap_area_lock
);
3975 * Mark allocated areas as accessible. Do it now as a best-effort
3976 * approach, as they can be mapped outside of vmalloc code.
3977 * With hardware tag-based KASAN, marking is skipped for
3978 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3980 for (area
= 0; area
< nr_vms
; area
++)
3981 vms
[area
]->addr
= kasan_unpoison_vmalloc(vms
[area
]->addr
,
3982 vms
[area
]->size
, KASAN_VMALLOC_PROT_NORMAL
);
3989 * Remove previously allocated areas. There is no
3990 * need in removing these areas from the busy tree,
3991 * because they are inserted only on the final step
3992 * and when pcpu_get_vm_areas() is success.
3995 orig_start
= vas
[area
]->va_start
;
3996 orig_end
= vas
[area
]->va_end
;
3997 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
3998 &free_vmap_area_list
);
4000 kasan_release_vmalloc(orig_start
, orig_end
,
4001 va
->va_start
, va
->va_end
);
4006 spin_unlock(&free_vmap_area_lock
);
4008 purge_vmap_area_lazy();
4011 /* Before "retry", check if we recover. */
4012 for (area
= 0; area
< nr_vms
; area
++) {
4016 vas
[area
] = kmem_cache_zalloc(
4017 vmap_area_cachep
, GFP_KERNEL
);
4026 for (area
= 0; area
< nr_vms
; area
++) {
4028 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
4038 spin_lock(&free_vmap_area_lock
);
4040 * We release all the vmalloc shadows, even the ones for regions that
4041 * hadn't been successfully added. This relies on kasan_release_vmalloc
4042 * being able to tolerate this case.
4044 for (area
= 0; area
< nr_vms
; area
++) {
4045 orig_start
= vas
[area
]->va_start
;
4046 orig_end
= vas
[area
]->va_end
;
4047 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4048 &free_vmap_area_list
);
4050 kasan_release_vmalloc(orig_start
, orig_end
,
4051 va
->va_start
, va
->va_end
);
4055 spin_unlock(&free_vmap_area_lock
);
4062 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4063 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4064 * @nr_vms: the number of allocated areas
4066 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4068 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
4072 for (i
= 0; i
< nr_vms
; i
++)
4073 free_vm_area(vms
[i
]);
4076 #endif /* CONFIG_SMP */
4078 #ifdef CONFIG_PRINTK
4079 bool vmalloc_dump_obj(void *object
)
4081 struct vm_struct
*vm
;
4082 void *objp
= (void *)PAGE_ALIGN((unsigned long)object
);
4084 vm
= find_vm_area(objp
);
4087 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4088 vm
->nr_pages
, (unsigned long)vm
->addr
, vm
->caller
);
4093 #ifdef CONFIG_PROC_FS
4094 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4095 __acquires(&vmap_purge_lock
)
4096 __acquires(&vmap_area_lock
)
4098 mutex_lock(&vmap_purge_lock
);
4099 spin_lock(&vmap_area_lock
);
4101 return seq_list_start(&vmap_area_list
, *pos
);
4104 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4106 return seq_list_next(p
, &vmap_area_list
, pos
);
4109 static void s_stop(struct seq_file
*m
, void *p
)
4110 __releases(&vmap_area_lock
)
4111 __releases(&vmap_purge_lock
)
4113 spin_unlock(&vmap_area_lock
);
4114 mutex_unlock(&vmap_purge_lock
);
4117 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
4119 if (IS_ENABLED(CONFIG_NUMA
)) {
4120 unsigned int nr
, *counters
= m
->private;
4121 unsigned int step
= 1U << vm_area_page_order(v
);
4126 if (v
->flags
& VM_UNINITIALIZED
)
4128 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4131 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
4133 for (nr
= 0; nr
< v
->nr_pages
; nr
+= step
)
4134 counters
[page_to_nid(v
->pages
[nr
])] += step
;
4135 for_each_node_state(nr
, N_HIGH_MEMORY
)
4137 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
4141 static void show_purge_info(struct seq_file
*m
)
4143 struct vmap_area
*va
;
4145 spin_lock(&purge_vmap_area_lock
);
4146 list_for_each_entry(va
, &purge_vmap_area_list
, list
) {
4147 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4148 (void *)va
->va_start
, (void *)va
->va_end
,
4149 va
->va_end
- va
->va_start
);
4151 spin_unlock(&purge_vmap_area_lock
);
4154 static int s_show(struct seq_file
*m
, void *p
)
4156 struct vmap_area
*va
;
4157 struct vm_struct
*v
;
4159 va
= list_entry(p
, struct vmap_area
, list
);
4162 if (va
->flags
& VMAP_RAM
)
4163 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
4164 (void *)va
->va_start
, (void *)va
->va_end
,
4165 va
->va_end
- va
->va_start
);
4172 seq_printf(m
, "0x%pK-0x%pK %7ld",
4173 v
->addr
, v
->addr
+ v
->size
, v
->size
);
4176 seq_printf(m
, " %pS", v
->caller
);
4179 seq_printf(m
, " pages=%d", v
->nr_pages
);
4182 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
4184 if (v
->flags
& VM_IOREMAP
)
4185 seq_puts(m
, " ioremap");
4187 if (v
->flags
& VM_ALLOC
)
4188 seq_puts(m
, " vmalloc");
4190 if (v
->flags
& VM_MAP
)
4191 seq_puts(m
, " vmap");
4193 if (v
->flags
& VM_USERMAP
)
4194 seq_puts(m
, " user");
4196 if (v
->flags
& VM_DMA_COHERENT
)
4197 seq_puts(m
, " dma-coherent");
4199 if (is_vmalloc_addr(v
->pages
))
4200 seq_puts(m
, " vpages");
4202 show_numa_info(m
, v
);
4206 * As a final step, dump "unpurged" areas.
4209 if (list_is_last(&va
->list
, &vmap_area_list
))
4215 static const struct seq_operations vmalloc_op
= {
4222 static int __init
proc_vmalloc_init(void)
4224 if (IS_ENABLED(CONFIG_NUMA
))
4225 proc_create_seq_private("vmallocinfo", 0400, NULL
,
4227 nr_node_ids
* sizeof(unsigned int), NULL
);
4229 proc_create_seq("vmallocinfo", 0400, NULL
, &vmalloc_op
);
4232 module_init(proc_vmalloc_init
);
4236 void __init
vmalloc_init(void)
4238 struct vmap_area
*va
;
4239 struct vm_struct
*tmp
;
4243 * Create the cache for vmap_area objects.
4245 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
4247 for_each_possible_cpu(i
) {
4248 struct vmap_block_queue
*vbq
;
4249 struct vfree_deferred
*p
;
4251 vbq
= &per_cpu(vmap_block_queue
, i
);
4252 spin_lock_init(&vbq
->lock
);
4253 INIT_LIST_HEAD(&vbq
->free
);
4254 p
= &per_cpu(vfree_deferred
, i
);
4255 init_llist_head(&p
->list
);
4256 INIT_WORK(&p
->wq
, delayed_vfree_work
);
4259 /* Import existing vmlist entries. */
4260 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
4261 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
4262 if (WARN_ON_ONCE(!va
))
4265 va
->va_start
= (unsigned long)tmp
->addr
;
4266 va
->va_end
= va
->va_start
+ tmp
->size
;
4268 insert_vmap_area(va
, &vmap_area_root
, &vmap_area_list
);
4272 * Now we can initialize a free vmap space.
4274 vmap_init_free_space();
4275 vmap_initialized
= true;