1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 #include <linux/vmalloc.h>
13 #include <linux/module.h>
14 #include <linux/highmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/set_memory.h>
22 #include <linux/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/list.h>
25 #include <linux/notifier.h>
26 #include <linux/rbtree.h>
27 #include <linux/xarray.h>
29 #include <linux/rcupdate.h>
30 #include <linux/pfn.h>
31 #include <linux/kmemleak.h>
32 #include <linux/atomic.h>
33 #include <linux/compiler.h>
34 #include <linux/memcontrol.h>
35 #include <linux/llist.h>
36 #include <linux/uio.h>
37 #include <linux/bitops.h>
38 #include <linux/rbtree_augmented.h>
39 #include <linux/overflow.h>
40 #include <linux/pgtable.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/vmalloc.h>
50 #include "pgalloc-track.h"
52 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53 static unsigned int __ro_after_init ioremap_max_page_shift
= BITS_PER_LONG
- 1;
55 static int __init
set_nohugeiomap(char *str
)
57 ioremap_max_page_shift
= PAGE_SHIFT
;
60 early_param("nohugeiomap", set_nohugeiomap
);
61 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62 static const unsigned int ioremap_max_page_shift
= PAGE_SHIFT
;
63 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66 static bool __ro_after_init vmap_allow_huge
= true;
68 static int __init
set_nohugevmalloc(char *str
)
70 vmap_allow_huge
= false;
73 early_param("nohugevmalloc", set_nohugevmalloc
);
74 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75 static const bool vmap_allow_huge
= false;
76 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78 bool is_vmalloc_addr(const void *x
)
80 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
82 return addr
>= VMALLOC_START
&& addr
< VMALLOC_END
;
84 EXPORT_SYMBOL(is_vmalloc_addr
);
86 struct vfree_deferred
{
87 struct llist_head list
;
88 struct work_struct wq
;
90 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
92 /*** Page table manipulation functions ***/
93 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
94 phys_addr_t phys_addr
, pgprot_t prot
,
95 unsigned int max_page_shift
, pgtbl_mod_mask
*mask
)
99 unsigned long size
= PAGE_SIZE
;
101 pfn
= phys_addr
>> PAGE_SHIFT
;
102 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
106 BUG_ON(!pte_none(ptep_get(pte
)));
108 #ifdef CONFIG_HUGETLB_PAGE
109 size
= arch_vmap_pte_range_map_size(addr
, end
, pfn
, max_page_shift
);
110 if (size
!= PAGE_SIZE
) {
111 pte_t entry
= pfn_pte(pfn
, prot
);
113 entry
= arch_make_huge_pte(entry
, ilog2(size
), 0);
114 set_huge_pte_at(&init_mm
, addr
, pte
, entry
, size
);
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 vmap_page_range(unsigned long addr
, unsigned long end
,
308 phys_addr_t phys_addr
, pgprot_t prot
)
312 err
= vmap_range_noflush(addr
, end
, phys_addr
, pgprot_nx(prot
),
313 ioremap_max_page_shift
);
314 flush_cache_vmap(addr
, end
);
316 err
= kmsan_ioremap_page_range(addr
, end
, phys_addr
, prot
,
317 ioremap_max_page_shift
);
321 int ioremap_page_range(unsigned long addr
, unsigned long end
,
322 phys_addr_t phys_addr
, pgprot_t prot
)
324 struct vm_struct
*area
;
326 area
= find_vm_area((void *)addr
);
327 if (!area
|| !(area
->flags
& VM_IOREMAP
)) {
328 WARN_ONCE(1, "vm_area at addr %lx is not marked as VM_IOREMAP\n", addr
);
331 if (addr
!= (unsigned long)area
->addr
||
332 (void *)end
!= area
->addr
+ get_vm_area_size(area
)) {
333 WARN_ONCE(1, "ioremap request [%lx,%lx) doesn't match vm_area [%lx, %lx)\n",
334 addr
, end
, (long)area
->addr
,
335 (long)area
->addr
+ get_vm_area_size(area
));
338 return vmap_page_range(addr
, end
, phys_addr
, prot
);
341 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
,
342 pgtbl_mod_mask
*mask
)
346 pte
= pte_offset_kernel(pmd
, addr
);
348 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
349 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
350 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
351 *mask
|= PGTBL_PTE_MODIFIED
;
354 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
,
355 pgtbl_mod_mask
*mask
)
361 pmd
= pmd_offset(pud
, addr
);
363 next
= pmd_addr_end(addr
, end
);
365 cleared
= pmd_clear_huge(pmd
);
366 if (cleared
|| pmd_bad(*pmd
))
367 *mask
|= PGTBL_PMD_MODIFIED
;
371 if (pmd_none_or_clear_bad(pmd
))
373 vunmap_pte_range(pmd
, addr
, next
, mask
);
376 } while (pmd
++, addr
= next
, addr
!= end
);
379 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
,
380 pgtbl_mod_mask
*mask
)
386 pud
= pud_offset(p4d
, addr
);
388 next
= pud_addr_end(addr
, end
);
390 cleared
= pud_clear_huge(pud
);
391 if (cleared
|| pud_bad(*pud
))
392 *mask
|= PGTBL_PUD_MODIFIED
;
396 if (pud_none_or_clear_bad(pud
))
398 vunmap_pmd_range(pud
, addr
, next
, mask
);
399 } while (pud
++, addr
= next
, addr
!= end
);
402 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
,
403 pgtbl_mod_mask
*mask
)
408 p4d
= p4d_offset(pgd
, addr
);
410 next
= p4d_addr_end(addr
, end
);
414 *mask
|= PGTBL_P4D_MODIFIED
;
416 if (p4d_none_or_clear_bad(p4d
))
418 vunmap_pud_range(p4d
, addr
, next
, mask
);
419 } while (p4d
++, addr
= next
, addr
!= end
);
423 * vunmap_range_noflush is similar to vunmap_range, but does not
424 * flush caches or TLBs.
426 * The caller is responsible for calling flush_cache_vmap() before calling
427 * this function, and flush_tlb_kernel_range after it has returned
428 * successfully (and before the addresses are expected to cause a page fault
429 * or be re-mapped for something else, if TLB flushes are being delayed or
432 * This is an internal function only. Do not use outside mm/.
434 void __vunmap_range_noflush(unsigned long start
, unsigned long end
)
438 unsigned long addr
= start
;
439 pgtbl_mod_mask mask
= 0;
442 pgd
= pgd_offset_k(addr
);
444 next
= pgd_addr_end(addr
, end
);
446 mask
|= PGTBL_PGD_MODIFIED
;
447 if (pgd_none_or_clear_bad(pgd
))
449 vunmap_p4d_range(pgd
, addr
, next
, &mask
);
450 } while (pgd
++, addr
= next
, addr
!= end
);
452 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
453 arch_sync_kernel_mappings(start
, end
);
456 void vunmap_range_noflush(unsigned long start
, unsigned long end
)
458 kmsan_vunmap_range_noflush(start
, end
);
459 __vunmap_range_noflush(start
, end
);
463 * vunmap_range - unmap kernel virtual addresses
464 * @addr: start of the VM area to unmap
465 * @end: end of the VM area to unmap (non-inclusive)
467 * Clears any present PTEs in the virtual address range, flushes TLBs and
468 * caches. Any subsequent access to the address before it has been re-mapped
471 void vunmap_range(unsigned long addr
, unsigned long end
)
473 flush_cache_vunmap(addr
, end
);
474 vunmap_range_noflush(addr
, end
);
475 flush_tlb_kernel_range(addr
, end
);
478 static int vmap_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
479 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
480 pgtbl_mod_mask
*mask
)
485 * nr is a running index into the array which helps higher level
486 * callers keep track of where we're up to.
489 pte
= pte_alloc_kernel_track(pmd
, addr
, mask
);
493 struct page
*page
= pages
[*nr
];
495 if (WARN_ON(!pte_none(ptep_get(pte
))))
499 if (WARN_ON(!pfn_valid(page_to_pfn(page
))))
502 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
504 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
505 *mask
|= PGTBL_PTE_MODIFIED
;
509 static int vmap_pages_pmd_range(pud_t
*pud
, unsigned long addr
,
510 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
511 pgtbl_mod_mask
*mask
)
516 pmd
= pmd_alloc_track(&init_mm
, pud
, addr
, mask
);
520 next
= pmd_addr_end(addr
, end
);
521 if (vmap_pages_pte_range(pmd
, addr
, next
, prot
, pages
, nr
, mask
))
523 } while (pmd
++, addr
= next
, addr
!= end
);
527 static int vmap_pages_pud_range(p4d_t
*p4d
, unsigned long addr
,
528 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
529 pgtbl_mod_mask
*mask
)
534 pud
= pud_alloc_track(&init_mm
, p4d
, addr
, mask
);
538 next
= pud_addr_end(addr
, end
);
539 if (vmap_pages_pmd_range(pud
, addr
, next
, prot
, pages
, nr
, mask
))
541 } while (pud
++, addr
= next
, addr
!= end
);
545 static int vmap_pages_p4d_range(pgd_t
*pgd
, unsigned long addr
,
546 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
,
547 pgtbl_mod_mask
*mask
)
552 p4d
= p4d_alloc_track(&init_mm
, pgd
, addr
, mask
);
556 next
= p4d_addr_end(addr
, end
);
557 if (vmap_pages_pud_range(p4d
, addr
, next
, prot
, pages
, nr
, mask
))
559 } while (p4d
++, addr
= next
, addr
!= end
);
563 static int vmap_small_pages_range_noflush(unsigned long addr
, unsigned long end
,
564 pgprot_t prot
, struct page
**pages
)
566 unsigned long start
= addr
;
571 pgtbl_mod_mask mask
= 0;
574 pgd
= pgd_offset_k(addr
);
576 next
= pgd_addr_end(addr
, end
);
578 mask
|= PGTBL_PGD_MODIFIED
;
579 err
= vmap_pages_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
, &mask
);
582 } while (pgd
++, addr
= next
, addr
!= end
);
584 if (mask
& ARCH_PAGE_TABLE_SYNC_MASK
)
585 arch_sync_kernel_mappings(start
, end
);
591 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
594 * The caller is responsible for calling flush_cache_vmap() after this
595 * function returns successfully and before the addresses are accessed.
597 * This is an internal function only. Do not use outside mm/.
599 int __vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
600 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
602 unsigned int i
, nr
= (end
- addr
) >> PAGE_SHIFT
;
604 WARN_ON(page_shift
< PAGE_SHIFT
);
606 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC
) ||
607 page_shift
== PAGE_SHIFT
)
608 return vmap_small_pages_range_noflush(addr
, end
, prot
, pages
);
610 for (i
= 0; i
< nr
; i
+= 1U << (page_shift
- PAGE_SHIFT
)) {
613 err
= vmap_range_noflush(addr
, addr
+ (1UL << page_shift
),
614 page_to_phys(pages
[i
]), prot
,
619 addr
+= 1UL << page_shift
;
625 int vmap_pages_range_noflush(unsigned long addr
, unsigned long end
,
626 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
628 int ret
= kmsan_vmap_pages_range_noflush(addr
, end
, prot
, pages
,
633 return __vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
637 * vmap_pages_range - map pages to a kernel virtual address
638 * @addr: start of the VM area to map
639 * @end: end of the VM area to map (non-inclusive)
640 * @prot: page protection flags to use
641 * @pages: pages to map (always PAGE_SIZE pages)
642 * @page_shift: maximum shift that the pages may be mapped with, @pages must
643 * be aligned and contiguous up to at least this shift.
646 * 0 on success, -errno on failure.
648 static int vmap_pages_range(unsigned long addr
, unsigned long end
,
649 pgprot_t prot
, struct page
**pages
, unsigned int page_shift
)
653 err
= vmap_pages_range_noflush(addr
, end
, prot
, pages
, page_shift
);
654 flush_cache_vmap(addr
, end
);
658 static int check_sparse_vm_area(struct vm_struct
*area
, unsigned long start
,
662 if (WARN_ON_ONCE(area
->flags
& VM_FLUSH_RESET_PERMS
))
664 if (WARN_ON_ONCE(area
->flags
& VM_NO_GUARD
))
666 if (WARN_ON_ONCE(!(area
->flags
& VM_SPARSE
)))
668 if ((end
- start
) >> PAGE_SHIFT
> totalram_pages())
670 if (start
< (unsigned long)area
->addr
||
671 (void *)end
> area
->addr
+ get_vm_area_size(area
))
677 * vm_area_map_pages - map pages inside given sparse vm_area
679 * @start: start address inside vm_area
680 * @end: end address inside vm_area
681 * @pages: pages to map (always PAGE_SIZE pages)
683 int vm_area_map_pages(struct vm_struct
*area
, unsigned long start
,
684 unsigned long end
, struct page
**pages
)
688 err
= check_sparse_vm_area(area
, start
, end
);
692 return vmap_pages_range(start
, end
, PAGE_KERNEL
, pages
, PAGE_SHIFT
);
696 * vm_area_unmap_pages - unmap pages inside given sparse vm_area
698 * @start: start address inside vm_area
699 * @end: end address inside vm_area
701 void vm_area_unmap_pages(struct vm_struct
*area
, unsigned long start
,
704 if (check_sparse_vm_area(area
, start
, end
))
707 vunmap_range(start
, end
);
710 int is_vmalloc_or_module_addr(const void *x
)
713 * ARM, x86-64 and sparc64 put modules in a special place,
714 * and fall back on vmalloc() if that fails. Others
715 * just put it in the vmalloc space.
717 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
718 unsigned long addr
= (unsigned long)kasan_reset_tag(x
);
719 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
722 return is_vmalloc_addr(x
);
724 EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr
);
727 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
728 * return the tail page that corresponds to the base page address, which
729 * matches small vmap mappings.
731 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
733 unsigned long addr
= (unsigned long) vmalloc_addr
;
734 struct page
*page
= NULL
;
735 pgd_t
*pgd
= pgd_offset_k(addr
);
742 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
743 * architectures that do not vmalloc module space
745 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
749 if (WARN_ON_ONCE(pgd_leaf(*pgd
)))
750 return NULL
; /* XXX: no allowance for huge pgd */
751 if (WARN_ON_ONCE(pgd_bad(*pgd
)))
754 p4d
= p4d_offset(pgd
, addr
);
758 return p4d_page(*p4d
) + ((addr
& ~P4D_MASK
) >> PAGE_SHIFT
);
759 if (WARN_ON_ONCE(p4d_bad(*p4d
)))
762 pud
= pud_offset(p4d
, addr
);
766 return pud_page(*pud
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
767 if (WARN_ON_ONCE(pud_bad(*pud
)))
770 pmd
= pmd_offset(pud
, addr
);
774 return pmd_page(*pmd
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
775 if (WARN_ON_ONCE(pmd_bad(*pmd
)))
778 ptep
= pte_offset_kernel(pmd
, addr
);
779 pte
= ptep_get(ptep
);
780 if (pte_present(pte
))
781 page
= pte_page(pte
);
785 EXPORT_SYMBOL(vmalloc_to_page
);
788 * Map a vmalloc()-space virtual address to the physical page frame number.
790 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
792 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
794 EXPORT_SYMBOL(vmalloc_to_pfn
);
797 /*** Global kva allocator ***/
799 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
800 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
803 static DEFINE_SPINLOCK(free_vmap_area_lock
);
804 static bool vmap_initialized __read_mostly
;
807 * This kmem_cache is used for vmap_area objects. Instead of
808 * allocating from slab we reuse an object from this cache to
809 * make things faster. Especially in "no edge" splitting of
812 static struct kmem_cache
*vmap_area_cachep
;
815 * This linked list is used in pair with free_vmap_area_root.
816 * It gives O(1) access to prev/next to perform fast coalescing.
818 static LIST_HEAD(free_vmap_area_list
);
821 * This augment red-black tree represents the free vmap space.
822 * All vmap_area objects in this tree are sorted by va->va_start
823 * address. It is used for allocation and merging when a vmap
824 * object is released.
826 * Each vmap_area node contains a maximum available free block
827 * of its sub-tree, right or left. Therefore it is possible to
828 * find a lowest match of free area.
830 static struct rb_root free_vmap_area_root
= RB_ROOT
;
833 * Preload a CPU with one object for "no edge" split case. The
834 * aim is to get rid of allocations from the atomic context, thus
835 * to use more permissive allocation masks.
837 static DEFINE_PER_CPU(struct vmap_area
*, ne_fit_preload_node
);
840 * This structure defines a single, solid model where a list and
841 * rb-tree are part of one entity protected by the lock. Nodes are
842 * sorted in ascending order, thus for O(1) access to left/right
843 * neighbors a list is used as well as for sequential traversal.
847 struct list_head head
;
852 * A fast size storage contains VAs up to 1M size. A pool consists
853 * of linked between each other ready to go VAs of certain sizes.
854 * An index in the pool-array corresponds to number of pages + 1.
856 #define MAX_VA_SIZE_PAGES 256
859 struct list_head head
;
864 * An effective vmap-node logic. Users make use of nodes instead
865 * of a global heap. It allows to balance an access and mitigate
868 static struct vmap_node
{
869 /* Simple size segregated storage. */
870 struct vmap_pool pool
[MAX_VA_SIZE_PAGES
];
871 spinlock_t pool_lock
;
874 /* Bookkeeping data of this node. */
879 * Ready-to-free areas.
881 struct list_head purge_list
;
882 struct work_struct purge_work
;
883 unsigned long nr_purged
;
887 * Initial setup consists of one single node, i.e. a balancing
888 * is fully disabled. Later on, after vmap is initialized these
889 * parameters are updated based on a system capacity.
891 static struct vmap_node
*vmap_nodes
= &single
;
892 static __read_mostly
unsigned int nr_vmap_nodes
= 1;
893 static __read_mostly
unsigned int vmap_zone_size
= 1;
895 static inline unsigned int
896 addr_to_node_id(unsigned long addr
)
898 return (addr
/ vmap_zone_size
) % nr_vmap_nodes
;
901 static inline struct vmap_node
*
902 addr_to_node(unsigned long addr
)
904 return &vmap_nodes
[addr_to_node_id(addr
)];
907 static inline struct vmap_node
*
908 id_to_node(unsigned int id
)
910 return &vmap_nodes
[id
% nr_vmap_nodes
];
914 * We use the value 0 to represent "no node", that is why
915 * an encoded value will be the node-id incremented by 1.
916 * It is always greater then 0. A valid node_id which can
917 * be encoded is [0:nr_vmap_nodes - 1]. If a passed node_id
918 * is not valid 0 is returned.
921 encode_vn_id(unsigned int node_id
)
923 /* Can store U8_MAX [0:254] nodes. */
924 if (node_id
< nr_vmap_nodes
)
925 return (node_id
+ 1) << BITS_PER_BYTE
;
927 /* Warn and no node encoded. */
928 WARN_ONCE(1, "Encode wrong node id (%u)\n", node_id
);
933 * Returns an encoded node-id, the valid range is within
934 * [0:nr_vmap_nodes-1] values. Otherwise nr_vmap_nodes is
935 * returned if extracted data is wrong.
938 decode_vn_id(unsigned int val
)
940 unsigned int node_id
= (val
>> BITS_PER_BYTE
) - 1;
942 /* Can store U8_MAX [0:254] nodes. */
943 if (node_id
< nr_vmap_nodes
)
946 /* If it was _not_ zero, warn. */
947 WARN_ONCE(node_id
!= UINT_MAX
,
948 "Decode wrong node id (%d)\n", node_id
);
950 return nr_vmap_nodes
;
954 is_vn_id_valid(unsigned int node_id
)
956 if (node_id
< nr_vmap_nodes
)
962 static __always_inline
unsigned long
963 va_size(struct vmap_area
*va
)
965 return (va
->va_end
- va
->va_start
);
968 static __always_inline
unsigned long
969 get_subtree_max_size(struct rb_node
*node
)
971 struct vmap_area
*va
;
973 va
= rb_entry_safe(node
, struct vmap_area
, rb_node
);
974 return va
? va
->subtree_max_size
: 0;
977 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb
,
978 struct vmap_area
, rb_node
, unsigned long, subtree_max_size
, va_size
)
980 static void reclaim_and_purge_vmap_areas(void);
981 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
982 static void drain_vmap_area_work(struct work_struct
*work
);
983 static DECLARE_WORK(drain_vmap_work
, drain_vmap_area_work
);
985 static atomic_long_t nr_vmalloc_pages
;
987 unsigned long vmalloc_nr_pages(void)
989 return atomic_long_read(&nr_vmalloc_pages
);
992 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
993 static struct vmap_area
*
994 __find_vmap_area_exceed_addr(unsigned long addr
, struct rb_root
*root
)
996 struct vmap_area
*va
= NULL
;
997 struct rb_node
*n
= root
->rb_node
;
999 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
1002 struct vmap_area
*tmp
;
1004 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
1005 if (tmp
->va_end
> addr
) {
1007 if (tmp
->va_start
<= addr
)
1019 * Returns a node where a first VA, that satisfies addr < va_end, resides.
1020 * If success, a node is locked. A user is responsible to unlock it when a
1021 * VA is no longer needed to be accessed.
1023 * Returns NULL if nothing found.
1025 static struct vmap_node
*
1026 find_vmap_area_exceed_addr_lock(unsigned long addr
, struct vmap_area
**va
)
1028 struct vmap_node
*vn
, *va_node
= NULL
;
1029 struct vmap_area
*va_lowest
;
1032 for (i
= 0; i
< nr_vmap_nodes
; i
++) {
1033 vn
= &vmap_nodes
[i
];
1035 spin_lock(&vn
->busy
.lock
);
1036 va_lowest
= __find_vmap_area_exceed_addr(addr
, &vn
->busy
.root
);
1038 if (!va_node
|| va_lowest
->va_start
< (*va
)->va_start
) {
1040 spin_unlock(&va_node
->busy
.lock
);
1047 spin_unlock(&vn
->busy
.lock
);
1053 static struct vmap_area
*__find_vmap_area(unsigned long addr
, struct rb_root
*root
)
1055 struct rb_node
*n
= root
->rb_node
;
1057 addr
= (unsigned long)kasan_reset_tag((void *)addr
);
1060 struct vmap_area
*va
;
1062 va
= rb_entry(n
, struct vmap_area
, rb_node
);
1063 if (addr
< va
->va_start
)
1065 else if (addr
>= va
->va_end
)
1075 * This function returns back addresses of parent node
1076 * and its left or right link for further processing.
1078 * Otherwise NULL is returned. In that case all further
1079 * steps regarding inserting of conflicting overlap range
1080 * have to be declined and actually considered as a bug.
1082 static __always_inline
struct rb_node
**
1083 find_va_links(struct vmap_area
*va
,
1084 struct rb_root
*root
, struct rb_node
*from
,
1085 struct rb_node
**parent
)
1087 struct vmap_area
*tmp_va
;
1088 struct rb_node
**link
;
1091 link
= &root
->rb_node
;
1092 if (unlikely(!*link
)) {
1101 * Go to the bottom of the tree. When we hit the last point
1102 * we end up with parent rb_node and correct direction, i name
1103 * it link, where the new va->rb_node will be attached to.
1106 tmp_va
= rb_entry(*link
, struct vmap_area
, rb_node
);
1109 * During the traversal we also do some sanity check.
1110 * Trigger the BUG() if there are sides(left/right)
1113 if (va
->va_end
<= tmp_va
->va_start
)
1114 link
= &(*link
)->rb_left
;
1115 else if (va
->va_start
>= tmp_va
->va_end
)
1116 link
= &(*link
)->rb_right
;
1118 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
1119 va
->va_start
, va
->va_end
, tmp_va
->va_start
, tmp_va
->va_end
);
1125 *parent
= &tmp_va
->rb_node
;
1129 static __always_inline
struct list_head
*
1130 get_va_next_sibling(struct rb_node
*parent
, struct rb_node
**link
)
1132 struct list_head
*list
;
1134 if (unlikely(!parent
))
1136 * The red-black tree where we try to find VA neighbors
1137 * before merging or inserting is empty, i.e. it means
1138 * there is no free vmap space. Normally it does not
1139 * happen but we handle this case anyway.
1143 list
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
1144 return (&parent
->rb_right
== link
? list
->next
: list
);
1147 static __always_inline
void
1148 __link_va(struct vmap_area
*va
, struct rb_root
*root
,
1149 struct rb_node
*parent
, struct rb_node
**link
,
1150 struct list_head
*head
, bool augment
)
1153 * VA is still not in the list, but we can
1154 * identify its future previous list_head node.
1156 if (likely(parent
)) {
1157 head
= &rb_entry(parent
, struct vmap_area
, rb_node
)->list
;
1158 if (&parent
->rb_right
!= link
)
1162 /* Insert to the rb-tree */
1163 rb_link_node(&va
->rb_node
, parent
, link
);
1166 * Some explanation here. Just perform simple insertion
1167 * to the tree. We do not set va->subtree_max_size to
1168 * its current size before calling rb_insert_augmented().
1169 * It is because we populate the tree from the bottom
1170 * to parent levels when the node _is_ in the tree.
1172 * Therefore we set subtree_max_size to zero after insertion,
1173 * to let __augment_tree_propagate_from() puts everything to
1174 * the correct order later on.
1176 rb_insert_augmented(&va
->rb_node
,
1177 root
, &free_vmap_area_rb_augment_cb
);
1178 va
->subtree_max_size
= 0;
1180 rb_insert_color(&va
->rb_node
, root
);
1183 /* Address-sort this list */
1184 list_add(&va
->list
, head
);
1187 static __always_inline
void
1188 link_va(struct vmap_area
*va
, struct rb_root
*root
,
1189 struct rb_node
*parent
, struct rb_node
**link
,
1190 struct list_head
*head
)
1192 __link_va(va
, root
, parent
, link
, head
, false);
1195 static __always_inline
void
1196 link_va_augment(struct vmap_area
*va
, struct rb_root
*root
,
1197 struct rb_node
*parent
, struct rb_node
**link
,
1198 struct list_head
*head
)
1200 __link_va(va
, root
, parent
, link
, head
, true);
1203 static __always_inline
void
1204 __unlink_va(struct vmap_area
*va
, struct rb_root
*root
, bool augment
)
1206 if (WARN_ON(RB_EMPTY_NODE(&va
->rb_node
)))
1210 rb_erase_augmented(&va
->rb_node
,
1211 root
, &free_vmap_area_rb_augment_cb
);
1213 rb_erase(&va
->rb_node
, root
);
1215 list_del_init(&va
->list
);
1216 RB_CLEAR_NODE(&va
->rb_node
);
1219 static __always_inline
void
1220 unlink_va(struct vmap_area
*va
, struct rb_root
*root
)
1222 __unlink_va(va
, root
, false);
1225 static __always_inline
void
1226 unlink_va_augment(struct vmap_area
*va
, struct rb_root
*root
)
1228 __unlink_va(va
, root
, true);
1231 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1233 * Gets called when remove the node and rotate.
1235 static __always_inline
unsigned long
1236 compute_subtree_max_size(struct vmap_area
*va
)
1238 return max3(va_size(va
),
1239 get_subtree_max_size(va
->rb_node
.rb_left
),
1240 get_subtree_max_size(va
->rb_node
.rb_right
));
1244 augment_tree_propagate_check(void)
1246 struct vmap_area
*va
;
1247 unsigned long computed_size
;
1249 list_for_each_entry(va
, &free_vmap_area_list
, list
) {
1250 computed_size
= compute_subtree_max_size(va
);
1251 if (computed_size
!= va
->subtree_max_size
)
1252 pr_emerg("tree is corrupted: %lu, %lu\n",
1253 va_size(va
), va
->subtree_max_size
);
1259 * This function populates subtree_max_size from bottom to upper
1260 * levels starting from VA point. The propagation must be done
1261 * when VA size is modified by changing its va_start/va_end. Or
1262 * in case of newly inserting of VA to the tree.
1264 * It means that __augment_tree_propagate_from() must be called:
1265 * - After VA has been inserted to the tree(free path);
1266 * - After VA has been shrunk(allocation path);
1267 * - After VA has been increased(merging path).
1269 * Please note that, it does not mean that upper parent nodes
1270 * and their subtree_max_size are recalculated all the time up
1279 * For example if we modify the node 4, shrinking it to 2, then
1280 * no any modification is required. If we shrink the node 2 to 1
1281 * its subtree_max_size is updated only, and set to 1. If we shrink
1282 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1283 * node becomes 4--6.
1285 static __always_inline
void
1286 augment_tree_propagate_from(struct vmap_area
*va
)
1289 * Populate the tree from bottom towards the root until
1290 * the calculated maximum available size of checked node
1291 * is equal to its current one.
1293 free_vmap_area_rb_augment_cb_propagate(&va
->rb_node
, NULL
);
1295 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1296 augment_tree_propagate_check();
1301 insert_vmap_area(struct vmap_area
*va
,
1302 struct rb_root
*root
, struct list_head
*head
)
1304 struct rb_node
**link
;
1305 struct rb_node
*parent
;
1307 link
= find_va_links(va
, root
, NULL
, &parent
);
1309 link_va(va
, root
, parent
, link
, head
);
1313 insert_vmap_area_augment(struct vmap_area
*va
,
1314 struct rb_node
*from
, struct rb_root
*root
,
1315 struct list_head
*head
)
1317 struct rb_node
**link
;
1318 struct rb_node
*parent
;
1321 link
= find_va_links(va
, NULL
, from
, &parent
);
1323 link
= find_va_links(va
, root
, NULL
, &parent
);
1326 link_va_augment(va
, root
, parent
, link
, head
);
1327 augment_tree_propagate_from(va
);
1332 * Merge de-allocated chunk of VA memory with previous
1333 * and next free blocks. If coalesce is not done a new
1334 * free area is inserted. If VA has been merged, it is
1337 * Please note, it can return NULL in case of overlap
1338 * ranges, followed by WARN() report. Despite it is a
1339 * buggy behaviour, a system can be alive and keep
1342 static __always_inline
struct vmap_area
*
1343 __merge_or_add_vmap_area(struct vmap_area
*va
,
1344 struct rb_root
*root
, struct list_head
*head
, bool augment
)
1346 struct vmap_area
*sibling
;
1347 struct list_head
*next
;
1348 struct rb_node
**link
;
1349 struct rb_node
*parent
;
1350 bool merged
= false;
1353 * Find a place in the tree where VA potentially will be
1354 * inserted, unless it is merged with its sibling/siblings.
1356 link
= find_va_links(va
, root
, NULL
, &parent
);
1361 * Get next node of VA to check if merging can be done.
1363 next
= get_va_next_sibling(parent
, link
);
1364 if (unlikely(next
== NULL
))
1370 * |<------VA------>|<-----Next----->|
1375 sibling
= list_entry(next
, struct vmap_area
, list
);
1376 if (sibling
->va_start
== va
->va_end
) {
1377 sibling
->va_start
= va
->va_start
;
1379 /* Free vmap_area object. */
1380 kmem_cache_free(vmap_area_cachep
, va
);
1382 /* Point to the new merged area. */
1391 * |<-----Prev----->|<------VA------>|
1395 if (next
->prev
!= head
) {
1396 sibling
= list_entry(next
->prev
, struct vmap_area
, list
);
1397 if (sibling
->va_end
== va
->va_start
) {
1399 * If both neighbors are coalesced, it is important
1400 * to unlink the "next" node first, followed by merging
1401 * with "previous" one. Otherwise the tree might not be
1402 * fully populated if a sibling's augmented value is
1403 * "normalized" because of rotation operations.
1406 __unlink_va(va
, root
, augment
);
1408 sibling
->va_end
= va
->va_end
;
1410 /* Free vmap_area object. */
1411 kmem_cache_free(vmap_area_cachep
, va
);
1413 /* Point to the new merged area. */
1421 __link_va(va
, root
, parent
, link
, head
, augment
);
1426 static __always_inline
struct vmap_area
*
1427 merge_or_add_vmap_area(struct vmap_area
*va
,
1428 struct rb_root
*root
, struct list_head
*head
)
1430 return __merge_or_add_vmap_area(va
, root
, head
, false);
1433 static __always_inline
struct vmap_area
*
1434 merge_or_add_vmap_area_augment(struct vmap_area
*va
,
1435 struct rb_root
*root
, struct list_head
*head
)
1437 va
= __merge_or_add_vmap_area(va
, root
, head
, true);
1439 augment_tree_propagate_from(va
);
1444 static __always_inline
bool
1445 is_within_this_va(struct vmap_area
*va
, unsigned long size
,
1446 unsigned long align
, unsigned long vstart
)
1448 unsigned long nva_start_addr
;
1450 if (va
->va_start
> vstart
)
1451 nva_start_addr
= ALIGN(va
->va_start
, align
);
1453 nva_start_addr
= ALIGN(vstart
, align
);
1455 /* Can be overflowed due to big size or alignment. */
1456 if (nva_start_addr
+ size
< nva_start_addr
||
1457 nva_start_addr
< vstart
)
1460 return (nva_start_addr
+ size
<= va
->va_end
);
1464 * Find the first free block(lowest start address) in the tree,
1465 * that will accomplish the request corresponding to passing
1466 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1467 * a search length is adjusted to account for worst case alignment
1470 static __always_inline
struct vmap_area
*
1471 find_vmap_lowest_match(struct rb_root
*root
, unsigned long size
,
1472 unsigned long align
, unsigned long vstart
, bool adjust_search_size
)
1474 struct vmap_area
*va
;
1475 struct rb_node
*node
;
1476 unsigned long length
;
1478 /* Start from the root. */
1479 node
= root
->rb_node
;
1481 /* Adjust the search size for alignment overhead. */
1482 length
= adjust_search_size
? size
+ align
- 1 : size
;
1485 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1487 if (get_subtree_max_size(node
->rb_left
) >= length
&&
1488 vstart
< va
->va_start
) {
1489 node
= node
->rb_left
;
1491 if (is_within_this_va(va
, size
, align
, vstart
))
1495 * Does not make sense to go deeper towards the right
1496 * sub-tree if it does not have a free block that is
1497 * equal or bigger to the requested search length.
1499 if (get_subtree_max_size(node
->rb_right
) >= length
) {
1500 node
= node
->rb_right
;
1505 * OK. We roll back and find the first right sub-tree,
1506 * that will satisfy the search criteria. It can happen
1507 * due to "vstart" restriction or an alignment overhead
1508 * that is bigger then PAGE_SIZE.
1510 while ((node
= rb_parent(node
))) {
1511 va
= rb_entry(node
, struct vmap_area
, rb_node
);
1512 if (is_within_this_va(va
, size
, align
, vstart
))
1515 if (get_subtree_max_size(node
->rb_right
) >= length
&&
1516 vstart
<= va
->va_start
) {
1518 * Shift the vstart forward. Please note, we update it with
1519 * parent's start address adding "1" because we do not want
1520 * to enter same sub-tree after it has already been checked
1521 * and no suitable free block found there.
1523 vstart
= va
->va_start
+ 1;
1524 node
= node
->rb_right
;
1534 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1535 #include <linux/random.h>
1537 static struct vmap_area
*
1538 find_vmap_lowest_linear_match(struct list_head
*head
, unsigned long size
,
1539 unsigned long align
, unsigned long vstart
)
1541 struct vmap_area
*va
;
1543 list_for_each_entry(va
, head
, list
) {
1544 if (!is_within_this_va(va
, size
, align
, vstart
))
1554 find_vmap_lowest_match_check(struct rb_root
*root
, struct list_head
*head
,
1555 unsigned long size
, unsigned long align
)
1557 struct vmap_area
*va_1
, *va_2
;
1558 unsigned long vstart
;
1561 get_random_bytes(&rnd
, sizeof(rnd
));
1562 vstart
= VMALLOC_START
+ rnd
;
1564 va_1
= find_vmap_lowest_match(root
, size
, align
, vstart
, false);
1565 va_2
= find_vmap_lowest_linear_match(head
, size
, align
, vstart
);
1568 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1569 va_1
, va_2
, vstart
);
1575 FL_FIT_TYPE
= 1, /* full fit */
1576 LE_FIT_TYPE
= 2, /* left edge fit */
1577 RE_FIT_TYPE
= 3, /* right edge fit */
1578 NE_FIT_TYPE
= 4 /* no edge fit */
1581 static __always_inline
enum fit_type
1582 classify_va_fit_type(struct vmap_area
*va
,
1583 unsigned long nva_start_addr
, unsigned long size
)
1587 /* Check if it is within VA. */
1588 if (nva_start_addr
< va
->va_start
||
1589 nva_start_addr
+ size
> va
->va_end
)
1593 if (va
->va_start
== nva_start_addr
) {
1594 if (va
->va_end
== nva_start_addr
+ size
)
1598 } else if (va
->va_end
== nva_start_addr
+ size
) {
1607 static __always_inline
int
1608 va_clip(struct rb_root
*root
, struct list_head
*head
,
1609 struct vmap_area
*va
, unsigned long nva_start_addr
,
1612 struct vmap_area
*lva
= NULL
;
1613 enum fit_type type
= classify_va_fit_type(va
, nva_start_addr
, size
);
1615 if (type
== FL_FIT_TYPE
) {
1617 * No need to split VA, it fully fits.
1623 unlink_va_augment(va
, root
);
1624 kmem_cache_free(vmap_area_cachep
, va
);
1625 } else if (type
== LE_FIT_TYPE
) {
1627 * Split left edge of fit VA.
1633 va
->va_start
+= size
;
1634 } else if (type
== RE_FIT_TYPE
) {
1636 * Split right edge of fit VA.
1642 va
->va_end
= nva_start_addr
;
1643 } else if (type
== NE_FIT_TYPE
) {
1645 * Split no edge of fit VA.
1651 lva
= __this_cpu_xchg(ne_fit_preload_node
, NULL
);
1652 if (unlikely(!lva
)) {
1654 * For percpu allocator we do not do any pre-allocation
1655 * and leave it as it is. The reason is it most likely
1656 * never ends up with NE_FIT_TYPE splitting. In case of
1657 * percpu allocations offsets and sizes are aligned to
1658 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1659 * are its main fitting cases.
1661 * There are a few exceptions though, as an example it is
1662 * a first allocation (early boot up) when we have "one"
1663 * big free space that has to be split.
1665 * Also we can hit this path in case of regular "vmap"
1666 * allocations, if "this" current CPU was not preloaded.
1667 * See the comment in alloc_vmap_area() why. If so, then
1668 * GFP_NOWAIT is used instead to get an extra object for
1669 * split purpose. That is rare and most time does not
1672 * What happens if an allocation gets failed. Basically,
1673 * an "overflow" path is triggered to purge lazily freed
1674 * areas to free some memory, then, the "retry" path is
1675 * triggered to repeat one more time. See more details
1676 * in alloc_vmap_area() function.
1678 lva
= kmem_cache_alloc(vmap_area_cachep
, GFP_NOWAIT
);
1684 * Build the remainder.
1686 lva
->va_start
= va
->va_start
;
1687 lva
->va_end
= nva_start_addr
;
1690 * Shrink this VA to remaining size.
1692 va
->va_start
= nva_start_addr
+ size
;
1697 if (type
!= FL_FIT_TYPE
) {
1698 augment_tree_propagate_from(va
);
1700 if (lva
) /* type == NE_FIT_TYPE */
1701 insert_vmap_area_augment(lva
, &va
->rb_node
, root
, head
);
1707 static unsigned long
1708 va_alloc(struct vmap_area
*va
,
1709 struct rb_root
*root
, struct list_head
*head
,
1710 unsigned long size
, unsigned long align
,
1711 unsigned long vstart
, unsigned long vend
)
1713 unsigned long nva_start_addr
;
1716 if (va
->va_start
> vstart
)
1717 nva_start_addr
= ALIGN(va
->va_start
, align
);
1719 nva_start_addr
= ALIGN(vstart
, align
);
1721 /* Check the "vend" restriction. */
1722 if (nva_start_addr
+ size
> vend
)
1725 /* Update the free vmap_area. */
1726 ret
= va_clip(root
, head
, va
, nva_start_addr
, size
);
1727 if (WARN_ON_ONCE(ret
))
1730 return nva_start_addr
;
1734 * Returns a start address of the newly allocated area, if success.
1735 * Otherwise a vend is returned that indicates failure.
1737 static __always_inline
unsigned long
1738 __alloc_vmap_area(struct rb_root
*root
, struct list_head
*head
,
1739 unsigned long size
, unsigned long align
,
1740 unsigned long vstart
, unsigned long vend
)
1742 bool adjust_search_size
= true;
1743 unsigned long nva_start_addr
;
1744 struct vmap_area
*va
;
1747 * Do not adjust when:
1748 * a) align <= PAGE_SIZE, because it does not make any sense.
1749 * All blocks(their start addresses) are at least PAGE_SIZE
1751 * b) a short range where a requested size corresponds to exactly
1752 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1753 * With adjusted search length an allocation would not succeed.
1755 if (align
<= PAGE_SIZE
|| (align
> PAGE_SIZE
&& (vend
- vstart
) == size
))
1756 adjust_search_size
= false;
1758 va
= find_vmap_lowest_match(root
, size
, align
, vstart
, adjust_search_size
);
1762 nva_start_addr
= va_alloc(va
, root
, head
, size
, align
, vstart
, vend
);
1763 if (nva_start_addr
== vend
)
1766 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1767 find_vmap_lowest_match_check(root
, head
, size
, align
);
1770 return nva_start_addr
;
1774 * Free a region of KVA allocated by alloc_vmap_area
1776 static void free_vmap_area(struct vmap_area
*va
)
1778 struct vmap_node
*vn
= addr_to_node(va
->va_start
);
1781 * Remove from the busy tree/list.
1783 spin_lock(&vn
->busy
.lock
);
1784 unlink_va(va
, &vn
->busy
.root
);
1785 spin_unlock(&vn
->busy
.lock
);
1788 * Insert/Merge it back to the free tree/list.
1790 spin_lock(&free_vmap_area_lock
);
1791 merge_or_add_vmap_area_augment(va
, &free_vmap_area_root
, &free_vmap_area_list
);
1792 spin_unlock(&free_vmap_area_lock
);
1796 preload_this_cpu_lock(spinlock_t
*lock
, gfp_t gfp_mask
, int node
)
1798 struct vmap_area
*va
= NULL
;
1801 * Preload this CPU with one extra vmap_area object. It is used
1802 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1803 * a CPU that does an allocation is preloaded.
1805 * We do it in non-atomic context, thus it allows us to use more
1806 * permissive allocation masks to be more stable under low memory
1807 * condition and high memory pressure.
1809 if (!this_cpu_read(ne_fit_preload_node
))
1810 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1814 if (va
&& __this_cpu_cmpxchg(ne_fit_preload_node
, NULL
, va
))
1815 kmem_cache_free(vmap_area_cachep
, va
);
1818 static struct vmap_pool
*
1819 size_to_va_pool(struct vmap_node
*vn
, unsigned long size
)
1821 unsigned int idx
= (size
- 1) / PAGE_SIZE
;
1823 if (idx
< MAX_VA_SIZE_PAGES
)
1824 return &vn
->pool
[idx
];
1830 node_pool_add_va(struct vmap_node
*n
, struct vmap_area
*va
)
1832 struct vmap_pool
*vp
;
1834 vp
= size_to_va_pool(n
, va_size(va
));
1838 spin_lock(&n
->pool_lock
);
1839 list_add(&va
->list
, &vp
->head
);
1840 WRITE_ONCE(vp
->len
, vp
->len
+ 1);
1841 spin_unlock(&n
->pool_lock
);
1846 static struct vmap_area
*
1847 node_pool_del_va(struct vmap_node
*vn
, unsigned long size
,
1848 unsigned long align
, unsigned long vstart
,
1851 struct vmap_area
*va
= NULL
;
1852 struct vmap_pool
*vp
;
1855 vp
= size_to_va_pool(vn
, size
);
1856 if (!vp
|| list_empty(&vp
->head
))
1859 spin_lock(&vn
->pool_lock
);
1860 if (!list_empty(&vp
->head
)) {
1861 va
= list_first_entry(&vp
->head
, struct vmap_area
, list
);
1863 if (IS_ALIGNED(va
->va_start
, align
)) {
1865 * Do some sanity check and emit a warning
1866 * if one of below checks detects an error.
1868 err
|= (va_size(va
) != size
);
1869 err
|= (va
->va_start
< vstart
);
1870 err
|= (va
->va_end
> vend
);
1872 if (!WARN_ON_ONCE(err
)) {
1873 list_del_init(&va
->list
);
1874 WRITE_ONCE(vp
->len
, vp
->len
- 1);
1879 list_move_tail(&va
->list
, &vp
->head
);
1883 spin_unlock(&vn
->pool_lock
);
1888 static struct vmap_area
*
1889 node_alloc(unsigned long size
, unsigned long align
,
1890 unsigned long vstart
, unsigned long vend
,
1891 unsigned long *addr
, unsigned int *vn_id
)
1893 struct vmap_area
*va
;
1899 * Fallback to a global heap if not vmalloc or there
1902 if (vstart
!= VMALLOC_START
|| vend
!= VMALLOC_END
||
1906 *vn_id
= raw_smp_processor_id() % nr_vmap_nodes
;
1907 va
= node_pool_del_va(id_to_node(*vn_id
), size
, align
, vstart
, vend
);
1908 *vn_id
= encode_vn_id(*vn_id
);
1911 *addr
= va
->va_start
;
1917 * Allocate a region of KVA of the specified size and alignment, within the
1920 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
1921 unsigned long align
,
1922 unsigned long vstart
, unsigned long vend
,
1923 int node
, gfp_t gfp_mask
,
1924 unsigned long va_flags
)
1926 struct vmap_node
*vn
;
1927 struct vmap_area
*va
;
1928 unsigned long freed
;
1934 if (unlikely(!size
|| offset_in_page(size
) || !is_power_of_2(align
)))
1935 return ERR_PTR(-EINVAL
);
1937 if (unlikely(!vmap_initialized
))
1938 return ERR_PTR(-EBUSY
);
1943 * If a VA is obtained from a global heap(if it fails here)
1944 * it is anyway marked with this "vn_id" so it is returned
1945 * to this pool's node later. Such way gives a possibility
1946 * to populate pools based on users demand.
1948 * On success a ready to go VA is returned.
1950 va
= node_alloc(size
, align
, vstart
, vend
, &addr
, &vn_id
);
1952 gfp_mask
= gfp_mask
& GFP_RECLAIM_MASK
;
1954 va
= kmem_cache_alloc_node(vmap_area_cachep
, gfp_mask
, node
);
1956 return ERR_PTR(-ENOMEM
);
1959 * Only scan the relevant parts containing pointers to other objects
1960 * to avoid false negatives.
1962 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
);
1967 preload_this_cpu_lock(&free_vmap_area_lock
, gfp_mask
, node
);
1968 addr
= __alloc_vmap_area(&free_vmap_area_root
, &free_vmap_area_list
,
1969 size
, align
, vstart
, vend
);
1970 spin_unlock(&free_vmap_area_lock
);
1973 trace_alloc_vmap_area(addr
, size
, align
, vstart
, vend
, addr
== vend
);
1976 * If an allocation fails, the "vend" address is
1977 * returned. Therefore trigger the overflow path.
1979 if (unlikely(addr
== vend
))
1982 va
->va_start
= addr
;
1983 va
->va_end
= addr
+ size
;
1985 va
->flags
= (va_flags
| vn_id
);
1987 vn
= addr_to_node(va
->va_start
);
1989 spin_lock(&vn
->busy
.lock
);
1990 insert_vmap_area(va
, &vn
->busy
.root
, &vn
->busy
.head
);
1991 spin_unlock(&vn
->busy
.lock
);
1993 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
1994 BUG_ON(va
->va_start
< vstart
);
1995 BUG_ON(va
->va_end
> vend
);
1997 ret
= kasan_populate_vmalloc(addr
, size
);
2000 return ERR_PTR(ret
);
2007 reclaim_and_purge_vmap_areas();
2013 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
2020 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
2021 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
2024 kmem_cache_free(vmap_area_cachep
, va
);
2025 return ERR_PTR(-EBUSY
);
2028 int register_vmap_purge_notifier(struct notifier_block
*nb
)
2030 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
2032 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
2034 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
2036 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
2038 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
2041 * lazy_max_pages is the maximum amount of virtual address space we gather up
2042 * before attempting to purge with a TLB flush.
2044 * There is a tradeoff here: a larger number will cover more kernel page tables
2045 * and take slightly longer to purge, but it will linearly reduce the number of
2046 * global TLB flushes that must be performed. It would seem natural to scale
2047 * this number up linearly with the number of CPUs (because vmapping activity
2048 * could also scale linearly with the number of CPUs), however it is likely
2049 * that in practice, workloads might be constrained in other ways that mean
2050 * vmap activity will not scale linearly with CPUs. Also, I want to be
2051 * conservative and not introduce a big latency on huge systems, so go with
2052 * a less aggressive log scale. It will still be an improvement over the old
2053 * code, and it will be simple to change the scale factor if we find that it
2054 * becomes a problem on bigger systems.
2056 static unsigned long lazy_max_pages(void)
2060 log
= fls(num_online_cpus());
2062 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
2065 static atomic_long_t vmap_lazy_nr
= ATOMIC_LONG_INIT(0);
2068 * Serialize vmap purging. There is no actual critical section protected
2069 * by this lock, but we want to avoid concurrent calls for performance
2070 * reasons and to make the pcpu_get_vm_areas more deterministic.
2072 static DEFINE_MUTEX(vmap_purge_lock
);
2074 /* for per-CPU blocks */
2075 static void purge_fragmented_blocks_allcpus(void);
2076 static cpumask_t purge_nodes
;
2079 reclaim_list_global(struct list_head
*head
)
2081 struct vmap_area
*va
, *n
;
2083 if (list_empty(head
))
2086 spin_lock(&free_vmap_area_lock
);
2087 list_for_each_entry_safe(va
, n
, head
, list
)
2088 merge_or_add_vmap_area_augment(va
,
2089 &free_vmap_area_root
, &free_vmap_area_list
);
2090 spin_unlock(&free_vmap_area_lock
);
2094 decay_va_pool_node(struct vmap_node
*vn
, bool full_decay
)
2096 struct vmap_area
*va
, *nva
;
2097 struct list_head decay_list
;
2098 struct rb_root decay_root
;
2099 unsigned long n_decay
;
2102 decay_root
= RB_ROOT
;
2103 INIT_LIST_HEAD(&decay_list
);
2105 for (i
= 0; i
< MAX_VA_SIZE_PAGES
; i
++) {
2106 struct list_head tmp_list
;
2108 if (list_empty(&vn
->pool
[i
].head
))
2111 INIT_LIST_HEAD(&tmp_list
);
2113 /* Detach the pool, so no-one can access it. */
2114 spin_lock(&vn
->pool_lock
);
2115 list_replace_init(&vn
->pool
[i
].head
, &tmp_list
);
2116 spin_unlock(&vn
->pool_lock
);
2119 WRITE_ONCE(vn
->pool
[i
].len
, 0);
2121 /* Decay a pool by ~25% out of left objects. */
2122 n_decay
= vn
->pool
[i
].len
>> 2;
2124 list_for_each_entry_safe(va
, nva
, &tmp_list
, list
) {
2125 list_del_init(&va
->list
);
2126 merge_or_add_vmap_area(va
, &decay_root
, &decay_list
);
2129 WRITE_ONCE(vn
->pool
[i
].len
, vn
->pool
[i
].len
- 1);
2137 * Attach the pool back if it has been partly decayed.
2138 * Please note, it is supposed that nobody(other contexts)
2139 * can populate the pool therefore a simple list replace
2140 * operation takes place here.
2142 if (!full_decay
&& !list_empty(&tmp_list
)) {
2143 spin_lock(&vn
->pool_lock
);
2144 list_replace_init(&tmp_list
, &vn
->pool
[i
].head
);
2145 spin_unlock(&vn
->pool_lock
);
2149 reclaim_list_global(&decay_list
);
2152 static void purge_vmap_node(struct work_struct
*work
)
2154 struct vmap_node
*vn
= container_of(work
,
2155 struct vmap_node
, purge_work
);
2156 struct vmap_area
*va
, *n_va
;
2157 LIST_HEAD(local_list
);
2161 list_for_each_entry_safe(va
, n_va
, &vn
->purge_list
, list
) {
2162 unsigned long nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
2163 unsigned long orig_start
= va
->va_start
;
2164 unsigned long orig_end
= va
->va_end
;
2165 unsigned int vn_id
= decode_vn_id(va
->flags
);
2167 list_del_init(&va
->list
);
2169 if (is_vmalloc_or_module_addr((void *)orig_start
))
2170 kasan_release_vmalloc(orig_start
, orig_end
,
2171 va
->va_start
, va
->va_end
);
2173 atomic_long_sub(nr
, &vmap_lazy_nr
);
2176 if (is_vn_id_valid(vn_id
) && !vn
->skip_populate
)
2177 if (node_pool_add_va(vn
, va
))
2180 /* Go back to global. */
2181 list_add(&va
->list
, &local_list
);
2184 reclaim_list_global(&local_list
);
2188 * Purges all lazily-freed vmap areas.
2190 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
,
2191 bool full_pool_decay
)
2193 unsigned long nr_purged_areas
= 0;
2194 unsigned int nr_purge_helpers
;
2195 unsigned int nr_purge_nodes
;
2196 struct vmap_node
*vn
;
2199 lockdep_assert_held(&vmap_purge_lock
);
2202 * Use cpumask to mark which node has to be processed.
2204 purge_nodes
= CPU_MASK_NONE
;
2206 for (i
= 0; i
< nr_vmap_nodes
; i
++) {
2207 vn
= &vmap_nodes
[i
];
2209 INIT_LIST_HEAD(&vn
->purge_list
);
2210 vn
->skip_populate
= full_pool_decay
;
2211 decay_va_pool_node(vn
, full_pool_decay
);
2213 if (RB_EMPTY_ROOT(&vn
->lazy
.root
))
2216 spin_lock(&vn
->lazy
.lock
);
2217 WRITE_ONCE(vn
->lazy
.root
.rb_node
, NULL
);
2218 list_replace_init(&vn
->lazy
.head
, &vn
->purge_list
);
2219 spin_unlock(&vn
->lazy
.lock
);
2221 start
= min(start
, list_first_entry(&vn
->purge_list
,
2222 struct vmap_area
, list
)->va_start
);
2224 end
= max(end
, list_last_entry(&vn
->purge_list
,
2225 struct vmap_area
, list
)->va_end
);
2227 cpumask_set_cpu(i
, &purge_nodes
);
2230 nr_purge_nodes
= cpumask_weight(&purge_nodes
);
2231 if (nr_purge_nodes
> 0) {
2232 flush_tlb_kernel_range(start
, end
);
2234 /* One extra worker is per a lazy_max_pages() full set minus one. */
2235 nr_purge_helpers
= atomic_long_read(&vmap_lazy_nr
) / lazy_max_pages();
2236 nr_purge_helpers
= clamp(nr_purge_helpers
, 1U, nr_purge_nodes
) - 1;
2238 for_each_cpu(i
, &purge_nodes
) {
2239 vn
= &vmap_nodes
[i
];
2241 if (nr_purge_helpers
> 0) {
2242 INIT_WORK(&vn
->purge_work
, purge_vmap_node
);
2244 if (cpumask_test_cpu(i
, cpu_online_mask
))
2245 schedule_work_on(i
, &vn
->purge_work
);
2247 schedule_work(&vn
->purge_work
);
2251 vn
->purge_work
.func
= NULL
;
2252 purge_vmap_node(&vn
->purge_work
);
2253 nr_purged_areas
+= vn
->nr_purged
;
2257 for_each_cpu(i
, &purge_nodes
) {
2258 vn
= &vmap_nodes
[i
];
2260 if (vn
->purge_work
.func
) {
2261 flush_work(&vn
->purge_work
);
2262 nr_purged_areas
+= vn
->nr_purged
;
2267 trace_purge_vmap_area_lazy(start
, end
, nr_purged_areas
);
2268 return nr_purged_areas
> 0;
2272 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
2274 static void reclaim_and_purge_vmap_areas(void)
2277 mutex_lock(&vmap_purge_lock
);
2278 purge_fragmented_blocks_allcpus();
2279 __purge_vmap_area_lazy(ULONG_MAX
, 0, true);
2280 mutex_unlock(&vmap_purge_lock
);
2283 static void drain_vmap_area_work(struct work_struct
*work
)
2285 mutex_lock(&vmap_purge_lock
);
2286 __purge_vmap_area_lazy(ULONG_MAX
, 0, false);
2287 mutex_unlock(&vmap_purge_lock
);
2291 * Free a vmap area, caller ensuring that the area has been unmapped,
2292 * unlinked and flush_cache_vunmap had been called for the correct
2295 static void free_vmap_area_noflush(struct vmap_area
*va
)
2297 unsigned long nr_lazy_max
= lazy_max_pages();
2298 unsigned long va_start
= va
->va_start
;
2299 unsigned int vn_id
= decode_vn_id(va
->flags
);
2300 struct vmap_node
*vn
;
2301 unsigned long nr_lazy
;
2303 if (WARN_ON_ONCE(!list_empty(&va
->list
)))
2306 nr_lazy
= atomic_long_add_return((va
->va_end
- va
->va_start
) >>
2307 PAGE_SHIFT
, &vmap_lazy_nr
);
2310 * If it was request by a certain node we would like to
2311 * return it to that node, i.e. its pool for later reuse.
2313 vn
= is_vn_id_valid(vn_id
) ?
2314 id_to_node(vn_id
):addr_to_node(va
->va_start
);
2316 spin_lock(&vn
->lazy
.lock
);
2317 insert_vmap_area(va
, &vn
->lazy
.root
, &vn
->lazy
.head
);
2318 spin_unlock(&vn
->lazy
.lock
);
2320 trace_free_vmap_area_noflush(va_start
, nr_lazy
, nr_lazy_max
);
2322 /* After this point, we may free va at any time */
2323 if (unlikely(nr_lazy
> nr_lazy_max
))
2324 schedule_work(&drain_vmap_work
);
2328 * Free and unmap a vmap area
2330 static void free_unmap_vmap_area(struct vmap_area
*va
)
2332 flush_cache_vunmap(va
->va_start
, va
->va_end
);
2333 vunmap_range_noflush(va
->va_start
, va
->va_end
);
2334 if (debug_pagealloc_enabled_static())
2335 flush_tlb_kernel_range(va
->va_start
, va
->va_end
);
2337 free_vmap_area_noflush(va
);
2340 struct vmap_area
*find_vmap_area(unsigned long addr
)
2342 struct vmap_node
*vn
;
2343 struct vmap_area
*va
;
2347 * An addr_to_node_id(addr) converts an address to a node index
2348 * where a VA is located. If VA spans several zones and passed
2349 * addr is not the same as va->va_start, what is not common, we
2350 * may need to scan extra nodes. See an example:
2353 * -|-----|-----|-----|-----|-
2356 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
2357 * addr is within 2 or 0 nodes we should do extra work.
2359 i
= j
= addr_to_node_id(addr
);
2361 vn
= &vmap_nodes
[i
];
2363 spin_lock(&vn
->busy
.lock
);
2364 va
= __find_vmap_area(addr
, &vn
->busy
.root
);
2365 spin_unlock(&vn
->busy
.lock
);
2369 } while ((i
= (i
+ 1) % nr_vmap_nodes
) != j
);
2374 static struct vmap_area
*find_unlink_vmap_area(unsigned long addr
)
2376 struct vmap_node
*vn
;
2377 struct vmap_area
*va
;
2381 * Check the comment in the find_vmap_area() about the loop.
2383 i
= j
= addr_to_node_id(addr
);
2385 vn
= &vmap_nodes
[i
];
2387 spin_lock(&vn
->busy
.lock
);
2388 va
= __find_vmap_area(addr
, &vn
->busy
.root
);
2390 unlink_va(va
, &vn
->busy
.root
);
2391 spin_unlock(&vn
->busy
.lock
);
2395 } while ((i
= (i
+ 1) % nr_vmap_nodes
) != j
);
2400 /*** Per cpu kva allocator ***/
2403 * vmap space is limited especially on 32 bit architectures. Ensure there is
2404 * room for at least 16 percpu vmap blocks per CPU.
2407 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
2408 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
2409 * instead (we just need a rough idea)
2411 #if BITS_PER_LONG == 32
2412 #define VMALLOC_SPACE (128UL*1024*1024)
2414 #define VMALLOC_SPACE (128UL*1024*1024*1024)
2417 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
2418 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
2419 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
2420 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
2421 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
2422 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
2423 #define VMAP_BBMAP_BITS \
2424 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
2425 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
2426 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
2428 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
2431 * Purge threshold to prevent overeager purging of fragmented blocks for
2432 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
2434 #define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
2436 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
2437 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
2438 #define VMAP_FLAGS_MASK 0x3
2440 struct vmap_block_queue
{
2442 struct list_head free
;
2445 * An xarray requires an extra memory dynamically to
2446 * be allocated. If it is an issue, we can use rb-tree
2449 struct xarray vmap_blocks
;
2454 struct vmap_area
*va
;
2455 unsigned long free
, dirty
;
2456 DECLARE_BITMAP(used_map
, VMAP_BBMAP_BITS
);
2457 unsigned long dirty_min
, dirty_max
; /*< dirty range */
2458 struct list_head free_list
;
2459 struct rcu_head rcu_head
;
2460 struct list_head purge
;
2463 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
2464 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
2467 * In order to fast access to any "vmap_block" associated with a
2468 * specific address, we use a hash.
2470 * A per-cpu vmap_block_queue is used in both ways, to serialize
2471 * an access to free block chains among CPUs(alloc path) and it
2472 * also acts as a vmap_block hash(alloc/free paths). It means we
2473 * overload it, since we already have the per-cpu array which is
2474 * used as a hash table. When used as a hash a 'cpu' passed to
2475 * per_cpu() is not actually a CPU but rather a hash index.
2477 * A hash function is addr_to_vb_xa() which hashes any address
2478 * to a specific index(in a hash) it belongs to. This then uses a
2479 * per_cpu() macro to access an array with generated index.
2486 * 0 10 20 30 40 50 60
2487 * |------|------|------|------|------|------|...<vmap address space>
2488 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
2490 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
2491 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
2493 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
2494 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
2496 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
2497 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
2499 * This technique almost always avoids lock contention on insert/remove,
2500 * however xarray spinlocks protect against any contention that remains.
2502 static struct xarray
*
2503 addr_to_vb_xa(unsigned long addr
)
2505 int index
= (addr
/ VMAP_BLOCK_SIZE
) % num_possible_cpus();
2507 return &per_cpu(vmap_block_queue
, index
).vmap_blocks
;
2511 * We should probably have a fallback mechanism to allocate virtual memory
2512 * out of partially filled vmap blocks. However vmap block sizing should be
2513 * fairly reasonable according to the vmalloc size, so it shouldn't be a
2517 static unsigned long addr_to_vb_idx(unsigned long addr
)
2519 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
2520 addr
/= VMAP_BLOCK_SIZE
;
2524 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
2528 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
2529 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
2530 return (void *)addr
;
2534 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2535 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2536 * @order: how many 2^order pages should be occupied in newly allocated block
2537 * @gfp_mask: flags for the page level allocator
2539 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2541 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
2543 struct vmap_block_queue
*vbq
;
2544 struct vmap_block
*vb
;
2545 struct vmap_area
*va
;
2547 unsigned long vb_idx
;
2551 node
= numa_node_id();
2553 vb
= kmalloc_node(sizeof(struct vmap_block
),
2554 gfp_mask
& GFP_RECLAIM_MASK
, node
);
2556 return ERR_PTR(-ENOMEM
);
2558 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
2559 VMALLOC_START
, VMALLOC_END
,
2561 VMAP_RAM
|VMAP_BLOCK
);
2564 return ERR_CAST(va
);
2567 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
2568 spin_lock_init(&vb
->lock
);
2570 /* At least something should be left free */
2571 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
2572 bitmap_zero(vb
->used_map
, VMAP_BBMAP_BITS
);
2573 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
2575 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2577 bitmap_set(vb
->used_map
, 0, (1UL << order
));
2578 INIT_LIST_HEAD(&vb
->free_list
);
2580 xa
= addr_to_vb_xa(va
->va_start
);
2581 vb_idx
= addr_to_vb_idx(va
->va_start
);
2582 err
= xa_insert(xa
, vb_idx
, vb
, gfp_mask
);
2586 return ERR_PTR(err
);
2589 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2590 spin_lock(&vbq
->lock
);
2591 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
2592 spin_unlock(&vbq
->lock
);
2597 static void free_vmap_block(struct vmap_block
*vb
)
2599 struct vmap_node
*vn
;
2600 struct vmap_block
*tmp
;
2603 xa
= addr_to_vb_xa(vb
->va
->va_start
);
2604 tmp
= xa_erase(xa
, addr_to_vb_idx(vb
->va
->va_start
));
2607 vn
= addr_to_node(vb
->va
->va_start
);
2608 spin_lock(&vn
->busy
.lock
);
2609 unlink_va(vb
->va
, &vn
->busy
.root
);
2610 spin_unlock(&vn
->busy
.lock
);
2612 free_vmap_area_noflush(vb
->va
);
2613 kfree_rcu(vb
, rcu_head
);
2616 static bool purge_fragmented_block(struct vmap_block
*vb
,
2617 struct vmap_block_queue
*vbq
, struct list_head
*purge_list
,
2620 if (vb
->free
+ vb
->dirty
!= VMAP_BBMAP_BITS
||
2621 vb
->dirty
== VMAP_BBMAP_BITS
)
2624 /* Don't overeagerly purge usable blocks unless requested */
2625 if (!(force_purge
|| vb
->free
< VMAP_PURGE_THRESHOLD
))
2628 /* prevent further allocs after releasing lock */
2629 WRITE_ONCE(vb
->free
, 0);
2630 /* prevent purging it again */
2631 WRITE_ONCE(vb
->dirty
, VMAP_BBMAP_BITS
);
2633 vb
->dirty_max
= VMAP_BBMAP_BITS
;
2634 spin_lock(&vbq
->lock
);
2635 list_del_rcu(&vb
->free_list
);
2636 spin_unlock(&vbq
->lock
);
2637 list_add_tail(&vb
->purge
, purge_list
);
2641 static void free_purged_blocks(struct list_head
*purge_list
)
2643 struct vmap_block
*vb
, *n_vb
;
2645 list_for_each_entry_safe(vb
, n_vb
, purge_list
, purge
) {
2646 list_del(&vb
->purge
);
2647 free_vmap_block(vb
);
2651 static void purge_fragmented_blocks(int cpu
)
2654 struct vmap_block
*vb
;
2655 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2658 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2659 unsigned long free
= READ_ONCE(vb
->free
);
2660 unsigned long dirty
= READ_ONCE(vb
->dirty
);
2662 if (free
+ dirty
!= VMAP_BBMAP_BITS
||
2663 dirty
== VMAP_BBMAP_BITS
)
2666 spin_lock(&vb
->lock
);
2667 purge_fragmented_block(vb
, vbq
, &purge
, true);
2668 spin_unlock(&vb
->lock
);
2671 free_purged_blocks(&purge
);
2674 static void purge_fragmented_blocks_allcpus(void)
2678 for_each_possible_cpu(cpu
)
2679 purge_fragmented_blocks(cpu
);
2682 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
2684 struct vmap_block_queue
*vbq
;
2685 struct vmap_block
*vb
;
2689 BUG_ON(offset_in_page(size
));
2690 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2691 if (WARN_ON(size
== 0)) {
2693 * Allocating 0 bytes isn't what caller wants since
2694 * get_order(0) returns funny result. Just warn and terminate
2699 order
= get_order(size
);
2702 vbq
= raw_cpu_ptr(&vmap_block_queue
);
2703 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
2704 unsigned long pages_off
;
2706 if (READ_ONCE(vb
->free
) < (1UL << order
))
2709 spin_lock(&vb
->lock
);
2710 if (vb
->free
< (1UL << order
)) {
2711 spin_unlock(&vb
->lock
);
2715 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
2716 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
2717 WRITE_ONCE(vb
->free
, vb
->free
- (1UL << order
));
2718 bitmap_set(vb
->used_map
, pages_off
, (1UL << order
));
2719 if (vb
->free
== 0) {
2720 spin_lock(&vbq
->lock
);
2721 list_del_rcu(&vb
->free_list
);
2722 spin_unlock(&vbq
->lock
);
2725 spin_unlock(&vb
->lock
);
2731 /* Allocate new block if nothing was found */
2733 vaddr
= new_vmap_block(order
, gfp_mask
);
2738 static void vb_free(unsigned long addr
, unsigned long size
)
2740 unsigned long offset
;
2742 struct vmap_block
*vb
;
2745 BUG_ON(offset_in_page(size
));
2746 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
2748 flush_cache_vunmap(addr
, addr
+ size
);
2750 order
= get_order(size
);
2751 offset
= (addr
& (VMAP_BLOCK_SIZE
- 1)) >> PAGE_SHIFT
;
2753 xa
= addr_to_vb_xa(addr
);
2754 vb
= xa_load(xa
, addr_to_vb_idx(addr
));
2756 spin_lock(&vb
->lock
);
2757 bitmap_clear(vb
->used_map
, offset
, (1UL << order
));
2758 spin_unlock(&vb
->lock
);
2760 vunmap_range_noflush(addr
, addr
+ size
);
2762 if (debug_pagealloc_enabled_static())
2763 flush_tlb_kernel_range(addr
, addr
+ size
);
2765 spin_lock(&vb
->lock
);
2767 /* Expand the not yet TLB flushed dirty range */
2768 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
2769 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
2771 WRITE_ONCE(vb
->dirty
, vb
->dirty
+ (1UL << order
));
2772 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
2774 spin_unlock(&vb
->lock
);
2775 free_vmap_block(vb
);
2777 spin_unlock(&vb
->lock
);
2780 static void _vm_unmap_aliases(unsigned long start
, unsigned long end
, int flush
)
2782 LIST_HEAD(purge_list
);
2785 if (unlikely(!vmap_initialized
))
2788 mutex_lock(&vmap_purge_lock
);
2790 for_each_possible_cpu(cpu
) {
2791 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
2792 struct vmap_block
*vb
;
2796 xa_for_each(&vbq
->vmap_blocks
, idx
, vb
) {
2797 spin_lock(&vb
->lock
);
2800 * Try to purge a fragmented block first. If it's
2801 * not purgeable, check whether there is dirty
2802 * space to be flushed.
2804 if (!purge_fragmented_block(vb
, vbq
, &purge_list
, false) &&
2805 vb
->dirty_max
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
2806 unsigned long va_start
= vb
->va
->va_start
;
2809 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
2810 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
2812 start
= min(s
, start
);
2815 /* Prevent that this is flushed again */
2816 vb
->dirty_min
= VMAP_BBMAP_BITS
;
2821 spin_unlock(&vb
->lock
);
2825 free_purged_blocks(&purge_list
);
2827 if (!__purge_vmap_area_lazy(start
, end
, false) && flush
)
2828 flush_tlb_kernel_range(start
, end
);
2829 mutex_unlock(&vmap_purge_lock
);
2833 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2835 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2836 * to amortize TLB flushing overheads. What this means is that any page you
2837 * have now, may, in a former life, have been mapped into kernel virtual
2838 * address by the vmap layer and so there might be some CPUs with TLB entries
2839 * still referencing that page (additional to the regular 1:1 kernel mapping).
2841 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2842 * be sure that none of the pages we have control over will have any aliases
2843 * from the vmap layer.
2845 void vm_unmap_aliases(void)
2847 unsigned long start
= ULONG_MAX
, end
= 0;
2850 _vm_unmap_aliases(start
, end
, flush
);
2852 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
2855 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2856 * @mem: the pointer returned by vm_map_ram
2857 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2859 void vm_unmap_ram(const void *mem
, unsigned int count
)
2861 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2862 unsigned long addr
= (unsigned long)kasan_reset_tag(mem
);
2863 struct vmap_area
*va
;
2867 BUG_ON(addr
< VMALLOC_START
);
2868 BUG_ON(addr
> VMALLOC_END
);
2869 BUG_ON(!PAGE_ALIGNED(addr
));
2871 kasan_poison_vmalloc(mem
, size
);
2873 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2874 debug_check_no_locks_freed(mem
, size
);
2875 vb_free(addr
, size
);
2879 va
= find_unlink_vmap_area(addr
);
2880 if (WARN_ON_ONCE(!va
))
2883 debug_check_no_locks_freed((void *)va
->va_start
,
2884 (va
->va_end
- va
->va_start
));
2885 free_unmap_vmap_area(va
);
2887 EXPORT_SYMBOL(vm_unmap_ram
);
2890 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2891 * @pages: an array of pointers to the pages to be mapped
2892 * @count: number of pages
2893 * @node: prefer to allocate data structures on this node
2895 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2896 * faster than vmap so it's good. But if you mix long-life and short-life
2897 * objects with vm_map_ram(), it could consume lots of address space through
2898 * fragmentation (especially on a 32bit machine). You could see failures in
2899 * the end. Please use this function for short-lived objects.
2901 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2903 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
)
2905 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
2909 if (likely(count
<= VMAP_MAX_ALLOC
)) {
2910 mem
= vb_alloc(size
, GFP_KERNEL
);
2913 addr
= (unsigned long)mem
;
2915 struct vmap_area
*va
;
2916 va
= alloc_vmap_area(size
, PAGE_SIZE
,
2917 VMALLOC_START
, VMALLOC_END
,
2918 node
, GFP_KERNEL
, VMAP_RAM
);
2922 addr
= va
->va_start
;
2926 if (vmap_pages_range(addr
, addr
+ size
, PAGE_KERNEL
,
2927 pages
, PAGE_SHIFT
) < 0) {
2928 vm_unmap_ram(mem
, count
);
2933 * Mark the pages as accessible, now that they are mapped.
2934 * With hardware tag-based KASAN, marking is skipped for
2935 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2937 mem
= kasan_unpoison_vmalloc(mem
, size
, KASAN_VMALLOC_PROT_NORMAL
);
2941 EXPORT_SYMBOL(vm_map_ram
);
2943 static struct vm_struct
*vmlist __initdata
;
2945 static inline unsigned int vm_area_page_order(struct vm_struct
*vm
)
2947 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2948 return vm
->page_order
;
2954 static inline void set_vm_area_page_order(struct vm_struct
*vm
, unsigned int order
)
2956 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2957 vm
->page_order
= order
;
2964 * vm_area_add_early - add vmap area early during boot
2965 * @vm: vm_struct to add
2967 * This function is used to add fixed kernel vm area to vmlist before
2968 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2969 * should contain proper values and the other fields should be zero.
2971 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2973 void __init
vm_area_add_early(struct vm_struct
*vm
)
2975 struct vm_struct
*tmp
, **p
;
2977 BUG_ON(vmap_initialized
);
2978 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
2979 if (tmp
->addr
>= vm
->addr
) {
2980 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
2983 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
2990 * vm_area_register_early - register vmap area early during boot
2991 * @vm: vm_struct to register
2992 * @align: requested alignment
2994 * This function is used to register kernel vm area before
2995 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2996 * proper values on entry and other fields should be zero. On return,
2997 * vm->addr contains the allocated address.
2999 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3001 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
3003 unsigned long addr
= ALIGN(VMALLOC_START
, align
);
3004 struct vm_struct
*cur
, **p
;
3006 BUG_ON(vmap_initialized
);
3008 for (p
= &vmlist
; (cur
= *p
) != NULL
; p
= &cur
->next
) {
3009 if ((unsigned long)cur
->addr
- addr
>= vm
->size
)
3011 addr
= ALIGN((unsigned long)cur
->addr
+ cur
->size
, align
);
3014 BUG_ON(addr
> VMALLOC_END
- vm
->size
);
3015 vm
->addr
= (void *)addr
;
3018 kasan_populate_early_vm_area_shadow(vm
->addr
, vm
->size
);
3021 static inline void setup_vmalloc_vm_locked(struct vm_struct
*vm
,
3022 struct vmap_area
*va
, unsigned long flags
, const void *caller
)
3025 vm
->addr
= (void *)va
->va_start
;
3026 vm
->size
= va
->va_end
- va
->va_start
;
3027 vm
->caller
= caller
;
3031 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
3032 unsigned long flags
, const void *caller
)
3034 struct vmap_node
*vn
= addr_to_node(va
->va_start
);
3036 spin_lock(&vn
->busy
.lock
);
3037 setup_vmalloc_vm_locked(vm
, va
, flags
, caller
);
3038 spin_unlock(&vn
->busy
.lock
);
3041 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
3044 * Before removing VM_UNINITIALIZED,
3045 * we should make sure that vm has proper values.
3046 * Pair with smp_rmb() in show_numa_info().
3049 vm
->flags
&= ~VM_UNINITIALIZED
;
3052 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
3053 unsigned long align
, unsigned long shift
, unsigned long flags
,
3054 unsigned long start
, unsigned long end
, int node
,
3055 gfp_t gfp_mask
, const void *caller
)
3057 struct vmap_area
*va
;
3058 struct vm_struct
*area
;
3059 unsigned long requested_size
= size
;
3061 BUG_ON(in_interrupt());
3062 size
= ALIGN(size
, 1ul << shift
);
3063 if (unlikely(!size
))
3066 if (flags
& VM_IOREMAP
)
3067 align
= 1ul << clamp_t(int, get_count_order_long(size
),
3068 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
3070 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
3071 if (unlikely(!area
))
3074 if (!(flags
& VM_NO_GUARD
))
3077 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
, 0);
3083 setup_vmalloc_vm(area
, va
, flags
, caller
);
3086 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
3087 * best-effort approach, as they can be mapped outside of vmalloc code.
3088 * For VM_ALLOC mappings, the pages are marked as accessible after
3089 * getting mapped in __vmalloc_node_range().
3090 * With hardware tag-based KASAN, marking is skipped for
3091 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3093 if (!(flags
& VM_ALLOC
))
3094 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, requested_size
,
3095 KASAN_VMALLOC_PROT_NORMAL
);
3100 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
3101 unsigned long start
, unsigned long end
,
3104 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
, start
, end
,
3105 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
3109 * get_vm_area - reserve a contiguous kernel virtual area
3110 * @size: size of the area
3111 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
3113 * Search an area of @size in the kernel virtual mapping area,
3114 * and reserved it for out purposes. Returns the area descriptor
3115 * on success or %NULL on failure.
3117 * Return: the area descriptor on success or %NULL on failure.
3119 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
3121 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
3122 VMALLOC_START
, VMALLOC_END
,
3123 NUMA_NO_NODE
, GFP_KERNEL
,
3124 __builtin_return_address(0));
3127 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
3130 return __get_vm_area_node(size
, 1, PAGE_SHIFT
, flags
,
3131 VMALLOC_START
, VMALLOC_END
,
3132 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
3136 * find_vm_area - find a continuous kernel virtual area
3137 * @addr: base address
3139 * Search for the kernel VM area starting at @addr, and return it.
3140 * It is up to the caller to do all required locking to keep the returned
3143 * Return: the area descriptor on success or %NULL on failure.
3145 struct vm_struct
*find_vm_area(const void *addr
)
3147 struct vmap_area
*va
;
3149 va
= find_vmap_area((unsigned long)addr
);
3157 * remove_vm_area - find and remove a continuous kernel virtual area
3158 * @addr: base address
3160 * Search for the kernel VM area starting at @addr, and remove it.
3161 * This function returns the found VM area, but using it is NOT safe
3162 * on SMP machines, except for its size or flags.
3164 * Return: the area descriptor on success or %NULL on failure.
3166 struct vm_struct
*remove_vm_area(const void *addr
)
3168 struct vmap_area
*va
;
3169 struct vm_struct
*vm
;
3173 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
3177 va
= find_unlink_vmap_area((unsigned long)addr
);
3182 debug_check_no_locks_freed(vm
->addr
, get_vm_area_size(vm
));
3183 debug_check_no_obj_freed(vm
->addr
, get_vm_area_size(vm
));
3184 kasan_free_module_shadow(vm
);
3185 kasan_poison_vmalloc(vm
->addr
, get_vm_area_size(vm
));
3187 free_unmap_vmap_area(va
);
3191 static inline void set_area_direct_map(const struct vm_struct
*area
,
3192 int (*set_direct_map
)(struct page
*page
))
3196 /* HUGE_VMALLOC passes small pages to set_direct_map */
3197 for (i
= 0; i
< area
->nr_pages
; i
++)
3198 if (page_address(area
->pages
[i
]))
3199 set_direct_map(area
->pages
[i
]);
3203 * Flush the vm mapping and reset the direct map.
3205 static void vm_reset_perms(struct vm_struct
*area
)
3207 unsigned long start
= ULONG_MAX
, end
= 0;
3208 unsigned int page_order
= vm_area_page_order(area
);
3213 * Find the start and end range of the direct mappings to make sure that
3214 * the vm_unmap_aliases() flush includes the direct map.
3216 for (i
= 0; i
< area
->nr_pages
; i
+= 1U << page_order
) {
3217 unsigned long addr
= (unsigned long)page_address(area
->pages
[i
]);
3220 unsigned long page_size
;
3222 page_size
= PAGE_SIZE
<< page_order
;
3223 start
= min(addr
, start
);
3224 end
= max(addr
+ page_size
, end
);
3230 * Set direct map to something invalid so that it won't be cached if
3231 * there are any accesses after the TLB flush, then flush the TLB and
3232 * reset the direct map permissions to the default.
3234 set_area_direct_map(area
, set_direct_map_invalid_noflush
);
3235 _vm_unmap_aliases(start
, end
, flush_dmap
);
3236 set_area_direct_map(area
, set_direct_map_default_noflush
);
3239 static void delayed_vfree_work(struct work_struct
*w
)
3241 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
3242 struct llist_node
*t
, *llnode
;
3244 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
3249 * vfree_atomic - release memory allocated by vmalloc()
3250 * @addr: memory base address
3252 * This one is just like vfree() but can be called in any atomic context
3255 void vfree_atomic(const void *addr
)
3257 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
3260 kmemleak_free(addr
);
3263 * Use raw_cpu_ptr() because this can be called from preemptible
3264 * context. Preemption is absolutely fine here, because the llist_add()
3265 * implementation is lockless, so it works even if we are adding to
3266 * another cpu's list. schedule_work() should be fine with this too.
3268 if (addr
&& llist_add((struct llist_node
*)addr
, &p
->list
))
3269 schedule_work(&p
->wq
);
3273 * vfree - Release memory allocated by vmalloc()
3274 * @addr: Memory base address
3276 * Free the virtually continuous memory area starting at @addr, as obtained
3277 * from one of the vmalloc() family of APIs. This will usually also free the
3278 * physical memory underlying the virtual allocation, but that memory is
3279 * reference counted, so it will not be freed until the last user goes away.
3281 * If @addr is NULL, no operation is performed.
3284 * May sleep if called *not* from interrupt context.
3285 * Must not be called in NMI context (strictly speaking, it could be
3286 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
3287 * conventions for vfree() arch-dependent would be a really bad idea).
3289 void vfree(const void *addr
)
3291 struct vm_struct
*vm
;
3294 if (unlikely(in_interrupt())) {
3300 kmemleak_free(addr
);
3306 vm
= remove_vm_area(addr
);
3307 if (unlikely(!vm
)) {
3308 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
3313 if (unlikely(vm
->flags
& VM_FLUSH_RESET_PERMS
))
3315 for (i
= 0; i
< vm
->nr_pages
; i
++) {
3316 struct page
*page
= vm
->pages
[i
];
3319 mod_memcg_page_state(page
, MEMCG_VMALLOC
, -1);
3321 * High-order allocs for huge vmallocs are split, so
3322 * can be freed as an array of order-0 allocations
3327 atomic_long_sub(vm
->nr_pages
, &nr_vmalloc_pages
);
3331 EXPORT_SYMBOL(vfree
);
3334 * vunmap - release virtual mapping obtained by vmap()
3335 * @addr: memory base address
3337 * Free the virtually contiguous memory area starting at @addr,
3338 * which was created from the page array passed to vmap().
3340 * Must not be called in interrupt context.
3342 void vunmap(const void *addr
)
3344 struct vm_struct
*vm
;
3346 BUG_ON(in_interrupt());
3351 vm
= remove_vm_area(addr
);
3352 if (unlikely(!vm
)) {
3353 WARN(1, KERN_ERR
"Trying to vunmap() nonexistent vm area (%p)\n",
3359 EXPORT_SYMBOL(vunmap
);
3362 * vmap - map an array of pages into virtually contiguous space
3363 * @pages: array of page pointers
3364 * @count: number of pages to map
3365 * @flags: vm_area->flags
3366 * @prot: page protection for the mapping
3368 * Maps @count pages from @pages into contiguous kernel virtual space.
3369 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
3370 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
3371 * are transferred from the caller to vmap(), and will be freed / dropped when
3372 * vfree() is called on the return value.
3374 * Return: the address of the area or %NULL on failure
3376 void *vmap(struct page
**pages
, unsigned int count
,
3377 unsigned long flags
, pgprot_t prot
)
3379 struct vm_struct
*area
;
3381 unsigned long size
; /* In bytes */
3385 if (WARN_ON_ONCE(flags
& VM_FLUSH_RESET_PERMS
))
3389 * Your top guard is someone else's bottom guard. Not having a top
3390 * guard compromises someone else's mappings too.
3392 if (WARN_ON_ONCE(flags
& VM_NO_GUARD
))
3393 flags
&= ~VM_NO_GUARD
;
3395 if (count
> totalram_pages())
3398 size
= (unsigned long)count
<< PAGE_SHIFT
;
3399 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
3403 addr
= (unsigned long)area
->addr
;
3404 if (vmap_pages_range(addr
, addr
+ size
, pgprot_nx(prot
),
3405 pages
, PAGE_SHIFT
) < 0) {
3410 if (flags
& VM_MAP_PUT_PAGES
) {
3411 area
->pages
= pages
;
3412 area
->nr_pages
= count
;
3416 EXPORT_SYMBOL(vmap
);
3418 #ifdef CONFIG_VMAP_PFN
3419 struct vmap_pfn_data
{
3420 unsigned long *pfns
;
3425 static int vmap_pfn_apply(pte_t
*pte
, unsigned long addr
, void *private)
3427 struct vmap_pfn_data
*data
= private;
3428 unsigned long pfn
= data
->pfns
[data
->idx
];
3431 if (WARN_ON_ONCE(pfn_valid(pfn
)))
3434 ptent
= pte_mkspecial(pfn_pte(pfn
, data
->prot
));
3435 set_pte_at(&init_mm
, addr
, pte
, ptent
);
3442 * vmap_pfn - map an array of PFNs into virtually contiguous space
3443 * @pfns: array of PFNs
3444 * @count: number of pages to map
3445 * @prot: page protection for the mapping
3447 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
3448 * the start address of the mapping.
3450 void *vmap_pfn(unsigned long *pfns
, unsigned int count
, pgprot_t prot
)
3452 struct vmap_pfn_data data
= { .pfns
= pfns
, .prot
= pgprot_nx(prot
) };
3453 struct vm_struct
*area
;
3455 area
= get_vm_area_caller(count
* PAGE_SIZE
, VM_IOREMAP
,
3456 __builtin_return_address(0));
3459 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
3460 count
* PAGE_SIZE
, vmap_pfn_apply
, &data
)) {
3465 flush_cache_vmap((unsigned long)area
->addr
,
3466 (unsigned long)area
->addr
+ count
* PAGE_SIZE
);
3470 EXPORT_SYMBOL_GPL(vmap_pfn
);
3471 #endif /* CONFIG_VMAP_PFN */
3473 static inline unsigned int
3474 vm_area_alloc_pages(gfp_t gfp
, int nid
,
3475 unsigned int order
, unsigned int nr_pages
, struct page
**pages
)
3477 unsigned int nr_allocated
= 0;
3478 gfp_t alloc_gfp
= gfp
;
3479 bool nofail
= false;
3484 * For order-0 pages we make use of bulk allocator, if
3485 * the page array is partly or not at all populated due
3486 * to fails, fallback to a single page allocator that is
3490 /* bulk allocator doesn't support nofail req. officially */
3491 gfp_t bulk_gfp
= gfp
& ~__GFP_NOFAIL
;
3493 while (nr_allocated
< nr_pages
) {
3494 unsigned int nr
, nr_pages_request
;
3497 * A maximum allowed request is hard-coded and is 100
3498 * pages per call. That is done in order to prevent a
3499 * long preemption off scenario in the bulk-allocator
3500 * so the range is [1:100].
3502 nr_pages_request
= min(100U, nr_pages
- nr_allocated
);
3504 /* memory allocation should consider mempolicy, we can't
3505 * wrongly use nearest node when nid == NUMA_NO_NODE,
3506 * otherwise memory may be allocated in only one node,
3507 * but mempolicy wants to alloc memory by interleaving.
3509 if (IS_ENABLED(CONFIG_NUMA
) && nid
== NUMA_NO_NODE
)
3510 nr
= alloc_pages_bulk_array_mempolicy(bulk_gfp
,
3512 pages
+ nr_allocated
);
3515 nr
= alloc_pages_bulk_array_node(bulk_gfp
, nid
,
3517 pages
+ nr_allocated
);
3523 * If zero or pages were obtained partly,
3524 * fallback to a single page allocator.
3526 if (nr
!= nr_pages_request
)
3529 } else if (gfp
& __GFP_NOFAIL
) {
3531 * Higher order nofail allocations are really expensive and
3532 * potentially dangerous (pre-mature OOM, disruptive reclaim
3533 * and compaction etc.
3535 alloc_gfp
&= ~__GFP_NOFAIL
;
3539 /* High-order pages or fallback path if "bulk" fails. */
3540 while (nr_allocated
< nr_pages
) {
3541 if (fatal_signal_pending(current
))
3544 if (nid
== NUMA_NO_NODE
)
3545 page
= alloc_pages(alloc_gfp
, order
);
3547 page
= alloc_pages_node(nid
, alloc_gfp
, order
);
3548 if (unlikely(!page
)) {
3552 /* fall back to the zero order allocations */
3553 alloc_gfp
|= __GFP_NOFAIL
;
3559 * Higher order allocations must be able to be treated as
3560 * indepdenent small pages by callers (as they can with
3561 * small-page vmallocs). Some drivers do their own refcounting
3562 * on vmalloc_to_page() pages, some use page->mapping,
3566 split_page(page
, order
);
3569 * Careful, we allocate and map page-order pages, but
3570 * tracking is done per PAGE_SIZE page so as to keep the
3571 * vm_struct APIs independent of the physical/mapped size.
3573 for (i
= 0; i
< (1U << order
); i
++)
3574 pages
[nr_allocated
+ i
] = page
+ i
;
3577 nr_allocated
+= 1U << order
;
3580 return nr_allocated
;
3583 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
3584 pgprot_t prot
, unsigned int page_shift
,
3587 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
3588 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3589 unsigned long addr
= (unsigned long)area
->addr
;
3590 unsigned long size
= get_vm_area_size(area
);
3591 unsigned long array_size
;
3592 unsigned int nr_small_pages
= size
>> PAGE_SHIFT
;
3593 unsigned int page_order
;
3597 array_size
= (unsigned long)nr_small_pages
* sizeof(struct page
*);
3599 if (!(gfp_mask
& (GFP_DMA
| GFP_DMA32
)))
3600 gfp_mask
|= __GFP_HIGHMEM
;
3602 /* Please note that the recursion is strictly bounded. */
3603 if (array_size
> PAGE_SIZE
) {
3604 area
->pages
= __vmalloc_node(array_size
, 1, nested_gfp
, node
,
3607 area
->pages
= kmalloc_node(array_size
, nested_gfp
, node
);
3611 warn_alloc(gfp_mask
, NULL
,
3612 "vmalloc error: size %lu, failed to allocated page array size %lu",
3613 nr_small_pages
* PAGE_SIZE
, array_size
);
3618 set_vm_area_page_order(area
, page_shift
- PAGE_SHIFT
);
3619 page_order
= vm_area_page_order(area
);
3621 area
->nr_pages
= vm_area_alloc_pages(gfp_mask
| __GFP_NOWARN
,
3622 node
, page_order
, nr_small_pages
, area
->pages
);
3624 atomic_long_add(area
->nr_pages
, &nr_vmalloc_pages
);
3625 if (gfp_mask
& __GFP_ACCOUNT
) {
3628 for (i
= 0; i
< area
->nr_pages
; i
++)
3629 mod_memcg_page_state(area
->pages
[i
], MEMCG_VMALLOC
, 1);
3633 * If not enough pages were obtained to accomplish an
3634 * allocation request, free them via vfree() if any.
3636 if (area
->nr_pages
!= nr_small_pages
) {
3638 * vm_area_alloc_pages() can fail due to insufficient memory but
3641 * - a pending fatal signal
3642 * - insufficient huge page-order pages
3644 * Since we always retry allocations at order-0 in the huge page
3645 * case a warning for either is spurious.
3647 if (!fatal_signal_pending(current
) && page_order
== 0)
3648 warn_alloc(gfp_mask
, NULL
,
3649 "vmalloc error: size %lu, failed to allocate pages",
3650 area
->nr_pages
* PAGE_SIZE
);
3655 * page tables allocations ignore external gfp mask, enforce it
3658 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3659 flags
= memalloc_nofs_save();
3660 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3661 flags
= memalloc_noio_save();
3664 ret
= vmap_pages_range(addr
, addr
+ size
, prot
, area
->pages
,
3666 if (nofail
&& (ret
< 0))
3667 schedule_timeout_uninterruptible(1);
3668 } while (nofail
&& (ret
< 0));
3670 if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == __GFP_IO
)
3671 memalloc_nofs_restore(flags
);
3672 else if ((gfp_mask
& (__GFP_FS
| __GFP_IO
)) == 0)
3673 memalloc_noio_restore(flags
);
3676 warn_alloc(gfp_mask
, NULL
,
3677 "vmalloc error: size %lu, failed to map pages",
3678 area
->nr_pages
* PAGE_SIZE
);
3690 * __vmalloc_node_range - allocate virtually contiguous memory
3691 * @size: allocation size
3692 * @align: desired alignment
3693 * @start: vm area range start
3694 * @end: vm area range end
3695 * @gfp_mask: flags for the page level allocator
3696 * @prot: protection mask for the allocated pages
3697 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3698 * @node: node to use for allocation or NUMA_NO_NODE
3699 * @caller: caller's return address
3701 * Allocate enough pages to cover @size from the page level
3702 * allocator with @gfp_mask flags. Please note that the full set of gfp
3703 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3705 * Zone modifiers are not supported. From the reclaim modifiers
3706 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3707 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3708 * __GFP_RETRY_MAYFAIL are not supported).
3710 * __GFP_NOWARN can be used to suppress failures messages.
3712 * Map them into contiguous kernel virtual space, using a pagetable
3713 * protection of @prot.
3715 * Return: the address of the area or %NULL on failure
3717 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
3718 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
3719 pgprot_t prot
, unsigned long vm_flags
, int node
,
3722 struct vm_struct
*area
;
3724 kasan_vmalloc_flags_t kasan_flags
= KASAN_VMALLOC_NONE
;
3725 unsigned long real_size
= size
;
3726 unsigned long real_align
= align
;
3727 unsigned int shift
= PAGE_SHIFT
;
3729 if (WARN_ON_ONCE(!size
))
3732 if ((size
>> PAGE_SHIFT
) > totalram_pages()) {
3733 warn_alloc(gfp_mask
, NULL
,
3734 "vmalloc error: size %lu, exceeds total pages",
3739 if (vmap_allow_huge
&& (vm_flags
& VM_ALLOW_HUGE_VMAP
)) {
3740 unsigned long size_per_node
;
3743 * Try huge pages. Only try for PAGE_KERNEL allocations,
3744 * others like modules don't yet expect huge pages in
3745 * their allocations due to apply_to_page_range not
3749 size_per_node
= size
;
3750 if (node
== NUMA_NO_NODE
)
3751 size_per_node
/= num_online_nodes();
3752 if (arch_vmap_pmd_supported(prot
) && size_per_node
>= PMD_SIZE
)
3755 shift
= arch_vmap_pte_supported_shift(size_per_node
);
3757 align
= max(real_align
, 1UL << shift
);
3758 size
= ALIGN(real_size
, 1UL << shift
);
3762 area
= __get_vm_area_node(real_size
, align
, shift
, VM_ALLOC
|
3763 VM_UNINITIALIZED
| vm_flags
, start
, end
, node
,
3766 bool nofail
= gfp_mask
& __GFP_NOFAIL
;
3767 warn_alloc(gfp_mask
, NULL
,
3768 "vmalloc error: size %lu, vm_struct allocation failed%s",
3769 real_size
, (nofail
) ? ". Retrying." : "");
3771 schedule_timeout_uninterruptible(1);
3778 * Prepare arguments for __vmalloc_area_node() and
3779 * kasan_unpoison_vmalloc().
3781 if (pgprot_val(prot
) == pgprot_val(PAGE_KERNEL
)) {
3782 if (kasan_hw_tags_enabled()) {
3784 * Modify protection bits to allow tagging.
3785 * This must be done before mapping.
3787 prot
= arch_vmap_pgprot_tagged(prot
);
3790 * Skip page_alloc poisoning and zeroing for physical
3791 * pages backing VM_ALLOC mapping. Memory is instead
3792 * poisoned and zeroed by kasan_unpoison_vmalloc().
3794 gfp_mask
|= __GFP_SKIP_KASAN
| __GFP_SKIP_ZERO
;
3797 /* Take note that the mapping is PAGE_KERNEL. */
3798 kasan_flags
|= KASAN_VMALLOC_PROT_NORMAL
;
3801 /* Allocate physical pages and map them into vmalloc space. */
3802 ret
= __vmalloc_area_node(area
, gfp_mask
, prot
, shift
, node
);
3807 * Mark the pages as accessible, now that they are mapped.
3808 * The condition for setting KASAN_VMALLOC_INIT should complement the
3809 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3810 * to make sure that memory is initialized under the same conditions.
3811 * Tag-based KASAN modes only assign tags to normal non-executable
3812 * allocations, see __kasan_unpoison_vmalloc().
3814 kasan_flags
|= KASAN_VMALLOC_VM_ALLOC
;
3815 if (!want_init_on_free() && want_init_on_alloc(gfp_mask
) &&
3816 (gfp_mask
& __GFP_SKIP_ZERO
))
3817 kasan_flags
|= KASAN_VMALLOC_INIT
;
3818 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3819 area
->addr
= kasan_unpoison_vmalloc(area
->addr
, real_size
, kasan_flags
);
3822 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3823 * flag. It means that vm_struct is not fully initialized.
3824 * Now, it is fully initialized, so remove this flag here.
3826 clear_vm_uninitialized_flag(area
);
3828 size
= PAGE_ALIGN(size
);
3829 if (!(vm_flags
& VM_DEFER_KMEMLEAK
))
3830 kmemleak_vmalloc(area
, size
, gfp_mask
);
3835 if (shift
> PAGE_SHIFT
) {
3846 * __vmalloc_node - allocate virtually contiguous memory
3847 * @size: allocation size
3848 * @align: desired alignment
3849 * @gfp_mask: flags for the page level allocator
3850 * @node: node to use for allocation or NUMA_NO_NODE
3851 * @caller: caller's return address
3853 * Allocate enough pages to cover @size from the page level allocator with
3854 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3856 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3857 * and __GFP_NOFAIL are not supported
3859 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3862 * Return: pointer to the allocated memory or %NULL on error
3864 void *__vmalloc_node(unsigned long size
, unsigned long align
,
3865 gfp_t gfp_mask
, int node
, const void *caller
)
3867 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
3868 gfp_mask
, PAGE_KERNEL
, 0, node
, caller
);
3871 * This is only for performance analysis of vmalloc and stress purpose.
3872 * It is required by vmalloc test module, therefore do not use it other
3875 #ifdef CONFIG_TEST_VMALLOC_MODULE
3876 EXPORT_SYMBOL_GPL(__vmalloc_node
);
3879 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
)
3881 return __vmalloc_node(size
, 1, gfp_mask
, NUMA_NO_NODE
,
3882 __builtin_return_address(0));
3884 EXPORT_SYMBOL(__vmalloc
);
3887 * vmalloc - allocate virtually contiguous memory
3888 * @size: allocation size
3890 * Allocate enough pages to cover @size from the page level
3891 * allocator and map them into contiguous kernel virtual space.
3893 * For tight control over page level allocator and protection flags
3894 * use __vmalloc() instead.
3896 * Return: pointer to the allocated memory or %NULL on error
3898 void *vmalloc(unsigned long size
)
3900 return __vmalloc_node(size
, 1, GFP_KERNEL
, NUMA_NO_NODE
,
3901 __builtin_return_address(0));
3903 EXPORT_SYMBOL(vmalloc
);
3906 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3907 * @size: allocation size
3908 * @gfp_mask: flags for the page level allocator
3910 * Allocate enough pages to cover @size from the page level
3911 * allocator and map them into contiguous kernel virtual space.
3912 * If @size is greater than or equal to PMD_SIZE, allow using
3913 * huge pages for the memory
3915 * Return: pointer to the allocated memory or %NULL on error
3917 void *vmalloc_huge(unsigned long size
, gfp_t gfp_mask
)
3919 return __vmalloc_node_range(size
, 1, VMALLOC_START
, VMALLOC_END
,
3920 gfp_mask
, PAGE_KERNEL
, VM_ALLOW_HUGE_VMAP
,
3921 NUMA_NO_NODE
, __builtin_return_address(0));
3923 EXPORT_SYMBOL_GPL(vmalloc_huge
);
3926 * vzalloc - allocate virtually contiguous memory with zero fill
3927 * @size: allocation size
3929 * Allocate enough pages to cover @size from the page level
3930 * allocator and map them into contiguous kernel virtual space.
3931 * The memory allocated is set to zero.
3933 * For tight control over page level allocator and protection flags
3934 * use __vmalloc() instead.
3936 * Return: pointer to the allocated memory or %NULL on error
3938 void *vzalloc(unsigned long size
)
3940 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, NUMA_NO_NODE
,
3941 __builtin_return_address(0));
3943 EXPORT_SYMBOL(vzalloc
);
3946 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3947 * @size: allocation size
3949 * The resulting memory area is zeroed so it can be mapped to userspace
3950 * without leaking data.
3952 * Return: pointer to the allocated memory or %NULL on error
3954 void *vmalloc_user(unsigned long size
)
3956 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
3957 GFP_KERNEL
| __GFP_ZERO
, PAGE_KERNEL
,
3958 VM_USERMAP
, NUMA_NO_NODE
,
3959 __builtin_return_address(0));
3961 EXPORT_SYMBOL(vmalloc_user
);
3964 * vmalloc_node - allocate memory on a specific node
3965 * @size: allocation size
3968 * Allocate enough pages to cover @size from the page level
3969 * allocator and map them into contiguous kernel virtual space.
3971 * For tight control over page level allocator and protection flags
3972 * use __vmalloc() instead.
3974 * Return: pointer to the allocated memory or %NULL on error
3976 void *vmalloc_node(unsigned long size
, int node
)
3978 return __vmalloc_node(size
, 1, GFP_KERNEL
, node
,
3979 __builtin_return_address(0));
3981 EXPORT_SYMBOL(vmalloc_node
);
3984 * vzalloc_node - allocate memory on a specific node with zero fill
3985 * @size: allocation size
3988 * Allocate enough pages to cover @size from the page level
3989 * allocator and map them into contiguous kernel virtual space.
3990 * The memory allocated is set to zero.
3992 * Return: pointer to the allocated memory or %NULL on error
3994 void *vzalloc_node(unsigned long size
, int node
)
3996 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_ZERO
, node
,
3997 __builtin_return_address(0));
3999 EXPORT_SYMBOL(vzalloc_node
);
4001 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
4002 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4003 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
4004 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
4007 * 64b systems should always have either DMA or DMA32 zones. For others
4008 * GFP_DMA32 should do the right thing and use the normal zone.
4010 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4014 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
4015 * @size: allocation size
4017 * Allocate enough 32bit PA addressable pages to cover @size from the
4018 * page level allocator and map them into contiguous kernel virtual space.
4020 * Return: pointer to the allocated memory or %NULL on error
4022 void *vmalloc_32(unsigned long size
)
4024 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, NUMA_NO_NODE
,
4025 __builtin_return_address(0));
4027 EXPORT_SYMBOL(vmalloc_32
);
4030 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
4031 * @size: allocation size
4033 * The resulting memory area is 32bit addressable and zeroed so it can be
4034 * mapped to userspace without leaking data.
4036 * Return: pointer to the allocated memory or %NULL on error
4038 void *vmalloc_32_user(unsigned long size
)
4040 return __vmalloc_node_range(size
, SHMLBA
, VMALLOC_START
, VMALLOC_END
,
4041 GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
4042 VM_USERMAP
, NUMA_NO_NODE
,
4043 __builtin_return_address(0));
4045 EXPORT_SYMBOL(vmalloc_32_user
);
4048 * Atomically zero bytes in the iterator.
4050 * Returns the number of zeroed bytes.
4052 static size_t zero_iter(struct iov_iter
*iter
, size_t count
)
4054 size_t remains
= count
;
4056 while (remains
> 0) {
4059 num
= min_t(size_t, remains
, PAGE_SIZE
);
4060 copied
= copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num
, iter
);
4067 return count
- remains
;
4071 * small helper routine, copy contents to iter from addr.
4072 * If the page is not present, fill zero.
4074 * Returns the number of copied bytes.
4076 static size_t aligned_vread_iter(struct iov_iter
*iter
,
4077 const char *addr
, size_t count
)
4079 size_t remains
= count
;
4082 while (remains
> 0) {
4083 unsigned long offset
, length
;
4086 offset
= offset_in_page(addr
);
4087 length
= PAGE_SIZE
- offset
;
4088 if (length
> remains
)
4090 page
= vmalloc_to_page(addr
);
4092 * To do safe access to this _mapped_ area, we need lock. But
4093 * adding lock here means that we need to add overhead of
4094 * vmalloc()/vfree() calls for this _debug_ interface, rarely
4095 * used. Instead of that, we'll use an local mapping via
4096 * copy_page_to_iter_nofault() and accept a small overhead in
4097 * this access function.
4100 copied
= copy_page_to_iter_nofault(page
, offset
,
4103 copied
= zero_iter(iter
, length
);
4108 if (copied
!= length
)
4112 return count
- remains
;
4116 * Read from a vm_map_ram region of memory.
4118 * Returns the number of copied bytes.
4120 static size_t vmap_ram_vread_iter(struct iov_iter
*iter
, const char *addr
,
4121 size_t count
, unsigned long flags
)
4124 struct vmap_block
*vb
;
4126 unsigned long offset
;
4127 unsigned int rs
, re
;
4131 * If it's area created by vm_map_ram() interface directly, but
4132 * not further subdividing and delegating management to vmap_block,
4135 if (!(flags
& VMAP_BLOCK
))
4136 return aligned_vread_iter(iter
, addr
, count
);
4141 * Area is split into regions and tracked with vmap_block, read out
4142 * each region and zero fill the hole between regions.
4144 xa
= addr_to_vb_xa((unsigned long) addr
);
4145 vb
= xa_load(xa
, addr_to_vb_idx((unsigned long)addr
));
4149 spin_lock(&vb
->lock
);
4150 if (bitmap_empty(vb
->used_map
, VMAP_BBMAP_BITS
)) {
4151 spin_unlock(&vb
->lock
);
4155 for_each_set_bitrange(rs
, re
, vb
->used_map
, VMAP_BBMAP_BITS
) {
4161 start
= vmap_block_vaddr(vb
->va
->va_start
, rs
);
4164 size_t to_zero
= min_t(size_t, start
- addr
, remains
);
4165 size_t zeroed
= zero_iter(iter
, to_zero
);
4170 if (remains
== 0 || zeroed
!= to_zero
)
4174 /*it could start reading from the middle of used region*/
4175 offset
= offset_in_page(addr
);
4176 n
= ((re
- rs
+ 1) << PAGE_SHIFT
) - offset
;
4180 copied
= aligned_vread_iter(iter
, start
+ offset
, n
);
4189 spin_unlock(&vb
->lock
);
4192 /* zero-fill the left dirty or free regions */
4193 return count
- remains
+ zero_iter(iter
, remains
);
4195 /* We couldn't copy/zero everything */
4196 spin_unlock(&vb
->lock
);
4197 return count
- remains
;
4201 * vread_iter() - read vmalloc area in a safe way to an iterator.
4202 * @iter: the iterator to which data should be written.
4203 * @addr: vm address.
4204 * @count: number of bytes to be read.
4206 * This function checks that addr is a valid vmalloc'ed area, and
4207 * copy data from that area to a given buffer. If the given memory range
4208 * of [addr...addr+count) includes some valid address, data is copied to
4209 * proper area of @buf. If there are memory holes, they'll be zero-filled.
4210 * IOREMAP area is treated as memory hole and no copy is done.
4212 * If [addr...addr+count) doesn't includes any intersects with alive
4213 * vm_struct area, returns 0. @buf should be kernel's buffer.
4215 * Note: In usual ops, vread() is never necessary because the caller
4216 * should know vmalloc() area is valid and can use memcpy().
4217 * This is for routines which have to access vmalloc area without
4218 * any information, as /proc/kcore.
4220 * Return: number of bytes for which addr and buf should be increased
4221 * (same number as @count) or %0 if [addr...addr+count) doesn't
4222 * include any intersection with valid vmalloc area
4224 long vread_iter(struct iov_iter
*iter
, const char *addr
, size_t count
)
4226 struct vmap_node
*vn
;
4227 struct vmap_area
*va
;
4228 struct vm_struct
*vm
;
4230 size_t n
, size
, flags
, remains
;
4233 addr
= kasan_reset_tag(addr
);
4235 /* Don't allow overflow */
4236 if ((unsigned long) addr
+ count
< count
)
4237 count
= -(unsigned long) addr
;
4241 vn
= find_vmap_area_exceed_addr_lock((unsigned long) addr
, &va
);
4245 /* no intersects with alive vmap_area */
4246 if ((unsigned long)addr
+ remains
<= va
->va_start
)
4256 flags
= va
->flags
& VMAP_FLAGS_MASK
;
4258 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
4259 * be set together with VMAP_RAM.
4261 WARN_ON(flags
== VMAP_BLOCK
);
4266 if (vm
&& (vm
->flags
& VM_UNINITIALIZED
))
4269 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4272 vaddr
= (char *) va
->va_start
;
4273 size
= vm
? get_vm_area_size(vm
) : va_size(va
);
4275 if (addr
>= vaddr
+ size
)
4279 size_t to_zero
= min_t(size_t, vaddr
- addr
, remains
);
4280 size_t zeroed
= zero_iter(iter
, to_zero
);
4285 if (remains
== 0 || zeroed
!= to_zero
)
4289 n
= vaddr
+ size
- addr
;
4293 if (flags
& VMAP_RAM
)
4294 copied
= vmap_ram_vread_iter(iter
, addr
, n
, flags
);
4295 else if (!(vm
&& (vm
->flags
& (VM_IOREMAP
| VM_SPARSE
))))
4296 copied
= aligned_vread_iter(iter
, addr
, n
);
4297 else /* IOREMAP | SPARSE area is treated as memory hole */
4298 copied
= zero_iter(iter
, n
);
4308 spin_unlock(&vn
->busy
.lock
);
4309 } while ((vn
= find_vmap_area_exceed_addr_lock(next
, &va
)));
4313 spin_unlock(&vn
->busy
.lock
);
4315 /* zero-fill memory holes */
4316 return count
- remains
+ zero_iter(iter
, remains
);
4318 /* Nothing remains, or We couldn't copy/zero everything. */
4320 spin_unlock(&vn
->busy
.lock
);
4322 return count
- remains
;
4326 * remap_vmalloc_range_partial - map vmalloc pages to userspace
4327 * @vma: vma to cover
4328 * @uaddr: target user address to start at
4329 * @kaddr: virtual address of vmalloc kernel memory
4330 * @pgoff: offset from @kaddr to start at
4331 * @size: size of map area
4333 * Returns: 0 for success, -Exxx on failure
4335 * This function checks that @kaddr is a valid vmalloc'ed area,
4336 * and that it is big enough to cover the range starting at
4337 * @uaddr in @vma. Will return failure if that criteria isn't
4340 * Similar to remap_pfn_range() (see mm/memory.c)
4342 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
4343 void *kaddr
, unsigned long pgoff
,
4346 struct vm_struct
*area
;
4348 unsigned long end_index
;
4350 if (check_shl_overflow(pgoff
, PAGE_SHIFT
, &off
))
4353 size
= PAGE_ALIGN(size
);
4355 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
4358 area
= find_vm_area(kaddr
);
4362 if (!(area
->flags
& (VM_USERMAP
| VM_DMA_COHERENT
)))
4365 if (check_add_overflow(size
, off
, &end_index
) ||
4366 end_index
> get_vm_area_size(area
))
4371 struct page
*page
= vmalloc_to_page(kaddr
);
4374 ret
= vm_insert_page(vma
, uaddr
, page
);
4383 vm_flags_set(vma
, VM_DONTEXPAND
| VM_DONTDUMP
);
4389 * remap_vmalloc_range - map vmalloc pages to userspace
4390 * @vma: vma to cover (map full range of vma)
4391 * @addr: vmalloc memory
4392 * @pgoff: number of pages into addr before first page to map
4394 * Returns: 0 for success, -Exxx on failure
4396 * This function checks that addr is a valid vmalloc'ed area, and
4397 * that it is big enough to cover the vma. Will return failure if
4398 * that criteria isn't met.
4400 * Similar to remap_pfn_range() (see mm/memory.c)
4402 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
4403 unsigned long pgoff
)
4405 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
4407 vma
->vm_end
- vma
->vm_start
);
4409 EXPORT_SYMBOL(remap_vmalloc_range
);
4411 void free_vm_area(struct vm_struct
*area
)
4413 struct vm_struct
*ret
;
4414 ret
= remove_vm_area(area
->addr
);
4415 BUG_ON(ret
!= area
);
4418 EXPORT_SYMBOL_GPL(free_vm_area
);
4421 static struct vmap_area
*node_to_va(struct rb_node
*n
)
4423 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
4427 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
4428 * @addr: target address
4430 * Returns: vmap_area if it is found. If there is no such area
4431 * the first highest(reverse order) vmap_area is returned
4432 * i.e. va->va_start < addr && va->va_end < addr or NULL
4433 * if there are no any areas before @addr.
4435 static struct vmap_area
*
4436 pvm_find_va_enclose_addr(unsigned long addr
)
4438 struct vmap_area
*va
, *tmp
;
4441 n
= free_vmap_area_root
.rb_node
;
4445 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
4446 if (tmp
->va_start
<= addr
) {
4448 if (tmp
->va_end
>= addr
)
4461 * pvm_determine_end_from_reverse - find the highest aligned address
4462 * of free block below VMALLOC_END
4464 * in - the VA we start the search(reverse order);
4465 * out - the VA with the highest aligned end address.
4466 * @align: alignment for required highest address
4468 * Returns: determined end address within vmap_area
4470 static unsigned long
4471 pvm_determine_end_from_reverse(struct vmap_area
**va
, unsigned long align
)
4473 unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
4477 list_for_each_entry_from_reverse((*va
),
4478 &free_vmap_area_list
, list
) {
4479 addr
= min((*va
)->va_end
& ~(align
- 1), vmalloc_end
);
4480 if ((*va
)->va_start
< addr
)
4489 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4490 * @offsets: array containing offset of each area
4491 * @sizes: array containing size of each area
4492 * @nr_vms: the number of areas to allocate
4493 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4495 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4496 * vm_structs on success, %NULL on failure
4498 * Percpu allocator wants to use congruent vm areas so that it can
4499 * maintain the offsets among percpu areas. This function allocates
4500 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4501 * be scattered pretty far, distance between two areas easily going up
4502 * to gigabytes. To avoid interacting with regular vmallocs, these
4503 * areas are allocated from top.
4505 * Despite its complicated look, this allocator is rather simple. It
4506 * does everything top-down and scans free blocks from the end looking
4507 * for matching base. While scanning, if any of the areas do not fit the
4508 * base address is pulled down to fit the area. Scanning is repeated till
4509 * all the areas fit and then all necessary data structures are inserted
4510 * and the result is returned.
4512 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
4513 const size_t *sizes
, int nr_vms
,
4516 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
4517 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
4518 struct vmap_area
**vas
, *va
;
4519 struct vm_struct
**vms
;
4520 int area
, area2
, last_area
, term_area
;
4521 unsigned long base
, start
, size
, end
, last_end
, orig_start
, orig_end
;
4522 bool purged
= false;
4524 /* verify parameters and allocate data structures */
4525 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
4526 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
4527 start
= offsets
[area
];
4528 end
= start
+ sizes
[area
];
4530 /* is everything aligned properly? */
4531 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
4532 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
4534 /* detect the area with the highest address */
4535 if (start
> offsets
[last_area
])
4538 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
4539 unsigned long start2
= offsets
[area2
];
4540 unsigned long end2
= start2
+ sizes
[area2
];
4542 BUG_ON(start2
< end
&& start
< end2
);
4545 last_end
= offsets
[last_area
] + sizes
[last_area
];
4547 if (vmalloc_end
- vmalloc_start
< last_end
) {
4552 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
4553 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
4557 for (area
= 0; area
< nr_vms
; area
++) {
4558 vas
[area
] = kmem_cache_zalloc(vmap_area_cachep
, GFP_KERNEL
);
4559 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
4560 if (!vas
[area
] || !vms
[area
])
4564 spin_lock(&free_vmap_area_lock
);
4566 /* start scanning - we scan from the top, begin with the last area */
4567 area
= term_area
= last_area
;
4568 start
= offsets
[area
];
4569 end
= start
+ sizes
[area
];
4571 va
= pvm_find_va_enclose_addr(vmalloc_end
);
4572 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4576 * base might have underflowed, add last_end before
4579 if (base
+ last_end
< vmalloc_start
+ last_end
)
4583 * Fitting base has not been found.
4589 * If required width exceeds current VA block, move
4590 * base downwards and then recheck.
4592 if (base
+ end
> va
->va_end
) {
4593 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4599 * If this VA does not fit, move base downwards and recheck.
4601 if (base
+ start
< va
->va_start
) {
4602 va
= node_to_va(rb_prev(&va
->rb_node
));
4603 base
= pvm_determine_end_from_reverse(&va
, align
) - end
;
4609 * This area fits, move on to the previous one. If
4610 * the previous one is the terminal one, we're done.
4612 area
= (area
+ nr_vms
- 1) % nr_vms
;
4613 if (area
== term_area
)
4616 start
= offsets
[area
];
4617 end
= start
+ sizes
[area
];
4618 va
= pvm_find_va_enclose_addr(base
+ end
);
4621 /* we've found a fitting base, insert all va's */
4622 for (area
= 0; area
< nr_vms
; area
++) {
4625 start
= base
+ offsets
[area
];
4628 va
= pvm_find_va_enclose_addr(start
);
4629 if (WARN_ON_ONCE(va
== NULL
))
4630 /* It is a BUG(), but trigger recovery instead. */
4633 ret
= va_clip(&free_vmap_area_root
,
4634 &free_vmap_area_list
, va
, start
, size
);
4635 if (WARN_ON_ONCE(unlikely(ret
)))
4636 /* It is a BUG(), but trigger recovery instead. */
4639 /* Allocated area. */
4641 va
->va_start
= start
;
4642 va
->va_end
= start
+ size
;
4645 spin_unlock(&free_vmap_area_lock
);
4647 /* populate the kasan shadow space */
4648 for (area
= 0; area
< nr_vms
; area
++) {
4649 if (kasan_populate_vmalloc(vas
[area
]->va_start
, sizes
[area
]))
4650 goto err_free_shadow
;
4653 /* insert all vm's */
4654 for (area
= 0; area
< nr_vms
; area
++) {
4655 struct vmap_node
*vn
= addr_to_node(vas
[area
]->va_start
);
4657 spin_lock(&vn
->busy
.lock
);
4658 insert_vmap_area(vas
[area
], &vn
->busy
.root
, &vn
->busy
.head
);
4659 setup_vmalloc_vm_locked(vms
[area
], vas
[area
], VM_ALLOC
,
4661 spin_unlock(&vn
->busy
.lock
);
4665 * Mark allocated areas as accessible. Do it now as a best-effort
4666 * approach, as they can be mapped outside of vmalloc code.
4667 * With hardware tag-based KASAN, marking is skipped for
4668 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4670 for (area
= 0; area
< nr_vms
; area
++)
4671 vms
[area
]->addr
= kasan_unpoison_vmalloc(vms
[area
]->addr
,
4672 vms
[area
]->size
, KASAN_VMALLOC_PROT_NORMAL
);
4679 * Remove previously allocated areas. There is no
4680 * need in removing these areas from the busy tree,
4681 * because they are inserted only on the final step
4682 * and when pcpu_get_vm_areas() is success.
4685 orig_start
= vas
[area
]->va_start
;
4686 orig_end
= vas
[area
]->va_end
;
4687 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4688 &free_vmap_area_list
);
4690 kasan_release_vmalloc(orig_start
, orig_end
,
4691 va
->va_start
, va
->va_end
);
4696 spin_unlock(&free_vmap_area_lock
);
4698 reclaim_and_purge_vmap_areas();
4701 /* Before "retry", check if we recover. */
4702 for (area
= 0; area
< nr_vms
; area
++) {
4706 vas
[area
] = kmem_cache_zalloc(
4707 vmap_area_cachep
, GFP_KERNEL
);
4716 for (area
= 0; area
< nr_vms
; area
++) {
4718 kmem_cache_free(vmap_area_cachep
, vas
[area
]);
4728 spin_lock(&free_vmap_area_lock
);
4730 * We release all the vmalloc shadows, even the ones for regions that
4731 * hadn't been successfully added. This relies on kasan_release_vmalloc
4732 * being able to tolerate this case.
4734 for (area
= 0; area
< nr_vms
; area
++) {
4735 orig_start
= vas
[area
]->va_start
;
4736 orig_end
= vas
[area
]->va_end
;
4737 va
= merge_or_add_vmap_area_augment(vas
[area
], &free_vmap_area_root
,
4738 &free_vmap_area_list
);
4740 kasan_release_vmalloc(orig_start
, orig_end
,
4741 va
->va_start
, va
->va_end
);
4745 spin_unlock(&free_vmap_area_lock
);
4752 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4753 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4754 * @nr_vms: the number of allocated areas
4756 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4758 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
4762 for (i
= 0; i
< nr_vms
; i
++)
4763 free_vm_area(vms
[i
]);
4766 #endif /* CONFIG_SMP */
4768 #ifdef CONFIG_PRINTK
4769 bool vmalloc_dump_obj(void *object
)
4772 struct vm_struct
*vm
;
4773 struct vmap_area
*va
;
4774 struct vmap_node
*vn
;
4776 unsigned int nr_pages
;
4778 addr
= PAGE_ALIGN((unsigned long) object
);
4779 vn
= addr_to_node(addr
);
4781 if (!spin_trylock(&vn
->busy
.lock
))
4784 va
= __find_vmap_area(addr
, &vn
->busy
.root
);
4785 if (!va
|| !va
->vm
) {
4786 spin_unlock(&vn
->busy
.lock
);
4791 addr
= (unsigned long) vm
->addr
;
4792 caller
= vm
->caller
;
4793 nr_pages
= vm
->nr_pages
;
4794 spin_unlock(&vn
->busy
.lock
);
4796 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4797 nr_pages
, addr
, caller
);
4803 #ifdef CONFIG_PROC_FS
4804 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
4806 if (IS_ENABLED(CONFIG_NUMA
)) {
4807 unsigned int nr
, *counters
= m
->private;
4808 unsigned int step
= 1U << vm_area_page_order(v
);
4813 if (v
->flags
& VM_UNINITIALIZED
)
4815 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4818 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
4820 for (nr
= 0; nr
< v
->nr_pages
; nr
+= step
)
4821 counters
[page_to_nid(v
->pages
[nr
])] += step
;
4822 for_each_node_state(nr
, N_HIGH_MEMORY
)
4824 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
4828 static void show_purge_info(struct seq_file
*m
)
4830 struct vmap_node
*vn
;
4831 struct vmap_area
*va
;
4834 for (i
= 0; i
< nr_vmap_nodes
; i
++) {
4835 vn
= &vmap_nodes
[i
];
4837 spin_lock(&vn
->lazy
.lock
);
4838 list_for_each_entry(va
, &vn
->lazy
.head
, list
) {
4839 seq_printf(m
, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4840 (void *)va
->va_start
, (void *)va
->va_end
,
4841 va
->va_end
- va
->va_start
);
4843 spin_unlock(&vn
->lazy
.lock
);
4847 static int vmalloc_info_show(struct seq_file
*m
, void *p
)
4849 struct vmap_node
*vn
;
4850 struct vmap_area
*va
;
4851 struct vm_struct
*v
;
4854 for (i
= 0; i
< nr_vmap_nodes
; i
++) {
4855 vn
= &vmap_nodes
[i
];
4857 spin_lock(&vn
->busy
.lock
);
4858 list_for_each_entry(va
, &vn
->busy
.head
, list
) {
4860 if (va
->flags
& VMAP_RAM
)
4861 seq_printf(m
, "0x%pK-0x%pK %7ld vm_map_ram\n",
4862 (void *)va
->va_start
, (void *)va
->va_end
,
4863 va
->va_end
- va
->va_start
);
4870 seq_printf(m
, "0x%pK-0x%pK %7ld",
4871 v
->addr
, v
->addr
+ v
->size
, v
->size
);
4874 seq_printf(m
, " %pS", v
->caller
);
4877 seq_printf(m
, " pages=%d", v
->nr_pages
);
4880 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
4882 if (v
->flags
& VM_IOREMAP
)
4883 seq_puts(m
, " ioremap");
4885 if (v
->flags
& VM_SPARSE
)
4886 seq_puts(m
, " sparse");
4888 if (v
->flags
& VM_ALLOC
)
4889 seq_puts(m
, " vmalloc");
4891 if (v
->flags
& VM_MAP
)
4892 seq_puts(m
, " vmap");
4894 if (v
->flags
& VM_USERMAP
)
4895 seq_puts(m
, " user");
4897 if (v
->flags
& VM_DMA_COHERENT
)
4898 seq_puts(m
, " dma-coherent");
4900 if (is_vmalloc_addr(v
->pages
))
4901 seq_puts(m
, " vpages");
4903 show_numa_info(m
, v
);
4906 spin_unlock(&vn
->busy
.lock
);
4910 * As a final step, dump "unpurged" areas.
4916 static int __init
proc_vmalloc_init(void)
4918 void *priv_data
= NULL
;
4920 if (IS_ENABLED(CONFIG_NUMA
))
4921 priv_data
= kmalloc(nr_node_ids
* sizeof(unsigned int), GFP_KERNEL
);
4923 proc_create_single_data("vmallocinfo",
4924 0400, NULL
, vmalloc_info_show
, priv_data
);
4928 module_init(proc_vmalloc_init
);
4932 static void __init
vmap_init_free_space(void)
4934 unsigned long vmap_start
= 1;
4935 const unsigned long vmap_end
= ULONG_MAX
;
4936 struct vmap_area
*free
;
4937 struct vm_struct
*busy
;
4941 * -|-----|.....|-----|-----|-----|.....|-
4943 * |<--------------------------------->|
4945 for (busy
= vmlist
; busy
; busy
= busy
->next
) {
4946 if ((unsigned long) busy
->addr
- vmap_start
> 0) {
4947 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
4948 if (!WARN_ON_ONCE(!free
)) {
4949 free
->va_start
= vmap_start
;
4950 free
->va_end
= (unsigned long) busy
->addr
;
4952 insert_vmap_area_augment(free
, NULL
,
4953 &free_vmap_area_root
,
4954 &free_vmap_area_list
);
4958 vmap_start
= (unsigned long) busy
->addr
+ busy
->size
;
4961 if (vmap_end
- vmap_start
> 0) {
4962 free
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
4963 if (!WARN_ON_ONCE(!free
)) {
4964 free
->va_start
= vmap_start
;
4965 free
->va_end
= vmap_end
;
4967 insert_vmap_area_augment(free
, NULL
,
4968 &free_vmap_area_root
,
4969 &free_vmap_area_list
);
4974 static void vmap_init_nodes(void)
4976 struct vmap_node
*vn
;
4979 #if BITS_PER_LONG == 64
4981 * A high threshold of max nodes is fixed and bound to 128,
4982 * thus a scale factor is 1 for systems where number of cores
4983 * are less or equal to specified threshold.
4985 * As for NUMA-aware notes. For bigger systems, for example
4986 * NUMA with multi-sockets, where we can end-up with thousands
4987 * of cores in total, a "sub-numa-clustering" should be added.
4989 * In this case a NUMA domain is considered as a single entity
4990 * with dedicated sub-nodes in it which describe one group or
4991 * set of cores. Therefore a per-domain purging is supposed to
4992 * be added as well as a per-domain balancing.
4994 n
= clamp_t(unsigned int, num_possible_cpus(), 1, 128);
4997 vn
= kmalloc_array(n
, sizeof(*vn
), GFP_NOWAIT
| __GFP_NOWARN
);
4999 /* Node partition is 16 pages. */
5000 vmap_zone_size
= (1 << 4) * PAGE_SIZE
;
5004 pr_err("Failed to allocate an array. Disable a node layer\n");
5009 for (n
= 0; n
< nr_vmap_nodes
; n
++) {
5010 vn
= &vmap_nodes
[n
];
5011 vn
->busy
.root
= RB_ROOT
;
5012 INIT_LIST_HEAD(&vn
->busy
.head
);
5013 spin_lock_init(&vn
->busy
.lock
);
5015 vn
->lazy
.root
= RB_ROOT
;
5016 INIT_LIST_HEAD(&vn
->lazy
.head
);
5017 spin_lock_init(&vn
->lazy
.lock
);
5019 for (i
= 0; i
< MAX_VA_SIZE_PAGES
; i
++) {
5020 INIT_LIST_HEAD(&vn
->pool
[i
].head
);
5021 WRITE_ONCE(vn
->pool
[i
].len
, 0);
5024 spin_lock_init(&vn
->pool_lock
);
5028 static unsigned long
5029 vmap_node_shrink_count(struct shrinker
*shrink
, struct shrink_control
*sc
)
5031 unsigned long count
;
5032 struct vmap_node
*vn
;
5035 for (count
= 0, i
= 0; i
< nr_vmap_nodes
; i
++) {
5036 vn
= &vmap_nodes
[i
];
5038 for (j
= 0; j
< MAX_VA_SIZE_PAGES
; j
++)
5039 count
+= READ_ONCE(vn
->pool
[j
].len
);
5042 return count
? count
: SHRINK_EMPTY
;
5045 static unsigned long
5046 vmap_node_shrink_scan(struct shrinker
*shrink
, struct shrink_control
*sc
)
5050 for (i
= 0; i
< nr_vmap_nodes
; i
++)
5051 decay_va_pool_node(&vmap_nodes
[i
], true);
5056 void __init
vmalloc_init(void)
5058 struct shrinker
*vmap_node_shrinker
;
5059 struct vmap_area
*va
;
5060 struct vmap_node
*vn
;
5061 struct vm_struct
*tmp
;
5065 * Create the cache for vmap_area objects.
5067 vmap_area_cachep
= KMEM_CACHE(vmap_area
, SLAB_PANIC
);
5069 for_each_possible_cpu(i
) {
5070 struct vmap_block_queue
*vbq
;
5071 struct vfree_deferred
*p
;
5073 vbq
= &per_cpu(vmap_block_queue
, i
);
5074 spin_lock_init(&vbq
->lock
);
5075 INIT_LIST_HEAD(&vbq
->free
);
5076 p
= &per_cpu(vfree_deferred
, i
);
5077 init_llist_head(&p
->list
);
5078 INIT_WORK(&p
->wq
, delayed_vfree_work
);
5079 xa_init(&vbq
->vmap_blocks
);
5083 * Setup nodes before importing vmlist.
5087 /* Import existing vmlist entries. */
5088 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
5089 va
= kmem_cache_zalloc(vmap_area_cachep
, GFP_NOWAIT
);
5090 if (WARN_ON_ONCE(!va
))
5093 va
->va_start
= (unsigned long)tmp
->addr
;
5094 va
->va_end
= va
->va_start
+ tmp
->size
;
5097 vn
= addr_to_node(va
->va_start
);
5098 insert_vmap_area(va
, &vn
->busy
.root
, &vn
->busy
.head
);
5102 * Now we can initialize a free vmap space.
5104 vmap_init_free_space();
5105 vmap_initialized
= true;
5107 vmap_node_shrinker
= shrinker_alloc(0, "vmap-node");
5108 if (!vmap_node_shrinker
) {
5109 pr_err("Failed to allocate vmap-node shrinker!\n");
5113 vmap_node_shrinker
->count_objects
= vmap_node_shrink_count
;
5114 vmap_node_shrinker
->scan_objects
= vmap_node_shrink_scan
;
5115 shrinker_register(vmap_node_shrinker
);