4 * Copyright (C) 1993 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8 * Numa awareness, Christoph Lameter, SGI, June 2005
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/debugobjects.h>
22 #include <linux/kallsyms.h>
23 #include <linux/list.h>
24 #include <linux/notifier.h>
25 #include <linux/rbtree.h>
26 #include <linux/radix-tree.h>
27 #include <linux/rcupdate.h>
28 #include <linux/pfn.h>
29 #include <linux/kmemleak.h>
30 #include <linux/atomic.h>
31 #include <linux/compiler.h>
32 #include <linux/llist.h>
33 #include <linux/bitops.h>
35 #include <linux/uaccess.h>
36 #include <asm/tlbflush.h>
37 #include <asm/shmparam.h>
41 struct vfree_deferred
{
42 struct llist_head list
;
43 struct work_struct wq
;
45 static DEFINE_PER_CPU(struct vfree_deferred
, vfree_deferred
);
47 static void __vunmap(const void *, int);
49 static void free_work(struct work_struct
*w
)
51 struct vfree_deferred
*p
= container_of(w
, struct vfree_deferred
, wq
);
52 struct llist_node
*llnode
= llist_del_all(&p
->list
);
55 llnode
= llist_next(llnode
);
60 /*** Page table manipulation functions ***/
62 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
)
66 pte
= pte_offset_kernel(pmd
, addr
);
68 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
69 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
70 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
73 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
)
78 pmd
= pmd_offset(pud
, addr
);
80 next
= pmd_addr_end(addr
, end
);
81 if (pmd_clear_huge(pmd
))
83 if (pmd_none_or_clear_bad(pmd
))
85 vunmap_pte_range(pmd
, addr
, next
);
86 } while (pmd
++, addr
= next
, addr
!= end
);
89 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
)
94 pud
= pud_offset(p4d
, addr
);
96 next
= pud_addr_end(addr
, end
);
97 if (pud_clear_huge(pud
))
99 if (pud_none_or_clear_bad(pud
))
101 vunmap_pmd_range(pud
, addr
, next
);
102 } while (pud
++, addr
= next
, addr
!= end
);
105 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
)
110 p4d
= p4d_offset(pgd
, addr
);
112 next
= p4d_addr_end(addr
, end
);
113 if (p4d_clear_huge(p4d
))
115 if (p4d_none_or_clear_bad(p4d
))
117 vunmap_pud_range(p4d
, addr
, next
);
118 } while (p4d
++, addr
= next
, addr
!= end
);
121 static void vunmap_page_range(unsigned long addr
, unsigned long end
)
127 pgd
= pgd_offset_k(addr
);
129 next
= pgd_addr_end(addr
, end
);
130 if (pgd_none_or_clear_bad(pgd
))
132 vunmap_p4d_range(pgd
, addr
, next
);
133 } while (pgd
++, addr
= next
, addr
!= end
);
136 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
137 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
142 * nr is a running index into the array which helps higher level
143 * callers keep track of where we're up to.
146 pte
= pte_alloc_kernel(pmd
, addr
);
150 struct page
*page
= pages
[*nr
];
152 if (WARN_ON(!pte_none(*pte
)))
156 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
158 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
162 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
163 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
168 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
172 next
= pmd_addr_end(addr
, end
);
173 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
175 } while (pmd
++, addr
= next
, addr
!= end
);
179 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
180 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
185 pud
= pud_alloc(&init_mm
, p4d
, addr
);
189 next
= pud_addr_end(addr
, end
);
190 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
192 } while (pud
++, addr
= next
, addr
!= end
);
196 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
197 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
202 p4d
= p4d_alloc(&init_mm
, pgd
, addr
);
206 next
= p4d_addr_end(addr
, end
);
207 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
))
209 } while (p4d
++, addr
= next
, addr
!= end
);
214 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
215 * will have pfns corresponding to the "pages" array.
217 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
219 static int vmap_page_range_noflush(unsigned long start
, unsigned long end
,
220 pgprot_t prot
, struct page
**pages
)
224 unsigned long addr
= start
;
229 pgd
= pgd_offset_k(addr
);
231 next
= pgd_addr_end(addr
, end
);
232 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
);
235 } while (pgd
++, addr
= next
, addr
!= end
);
240 static int vmap_page_range(unsigned long start
, unsigned long end
,
241 pgprot_t prot
, struct page
**pages
)
245 ret
= vmap_page_range_noflush(start
, end
, prot
, pages
);
246 flush_cache_vmap(start
, end
);
250 int is_vmalloc_or_module_addr(const void *x
)
253 * ARM, x86-64 and sparc64 put modules in a special place,
254 * and fall back on vmalloc() if that fails. Others
255 * just put it in the vmalloc space.
257 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
258 unsigned long addr
= (unsigned long)x
;
259 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
262 return is_vmalloc_addr(x
);
266 * Walk a vmap address to the struct page it maps.
268 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
270 unsigned long addr
= (unsigned long) vmalloc_addr
;
271 struct page
*page
= NULL
;
272 pgd_t
*pgd
= pgd_offset_k(addr
);
279 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
280 * architectures that do not vmalloc module space
282 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
286 p4d
= p4d_offset(pgd
, addr
);
289 pud
= pud_offset(p4d
, addr
);
292 * Don't dereference bad PUD or PMD (below) entries. This will also
293 * identify huge mappings, which we may encounter on architectures
294 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
295 * identified as vmalloc addresses by is_vmalloc_addr(), but are
296 * not [unambiguously] associated with a struct page, so there is
297 * no correct value to return for them.
299 WARN_ON_ONCE(pud_bad(*pud
));
300 if (pud_none(*pud
) || pud_bad(*pud
))
302 pmd
= pmd_offset(pud
, addr
);
303 WARN_ON_ONCE(pmd_bad(*pmd
));
304 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
307 ptep
= pte_offset_map(pmd
, addr
);
309 if (pte_present(pte
))
310 page
= pte_page(pte
);
314 EXPORT_SYMBOL(vmalloc_to_page
);
317 * Map a vmalloc()-space virtual address to the physical page frame number.
319 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
321 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
323 EXPORT_SYMBOL(vmalloc_to_pfn
);
326 /*** Global kva allocator ***/
328 #define VM_VM_AREA 0x04
330 static DEFINE_SPINLOCK(vmap_area_lock
);
331 /* Export for kexec only */
332 LIST_HEAD(vmap_area_list
);
333 static LLIST_HEAD(vmap_purge_list
);
334 static struct rb_root vmap_area_root
= RB_ROOT
;
336 /* The vmap cache globals are protected by vmap_area_lock */
337 static struct rb_node
*free_vmap_cache
;
338 static unsigned long cached_hole_size
;
339 static unsigned long cached_vstart
;
340 static unsigned long cached_align
;
342 static unsigned long vmap_area_pcpu_hole
;
344 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
346 struct rb_node
*n
= vmap_area_root
.rb_node
;
349 struct vmap_area
*va
;
351 va
= rb_entry(n
, struct vmap_area
, rb_node
);
352 if (addr
< va
->va_start
)
354 else if (addr
>= va
->va_end
)
363 static void __insert_vmap_area(struct vmap_area
*va
)
365 struct rb_node
**p
= &vmap_area_root
.rb_node
;
366 struct rb_node
*parent
= NULL
;
370 struct vmap_area
*tmp_va
;
373 tmp_va
= rb_entry(parent
, struct vmap_area
, rb_node
);
374 if (va
->va_start
< tmp_va
->va_end
)
376 else if (va
->va_end
> tmp_va
->va_start
)
382 rb_link_node(&va
->rb_node
, parent
, p
);
383 rb_insert_color(&va
->rb_node
, &vmap_area_root
);
385 /* address-sort this list */
386 tmp
= rb_prev(&va
->rb_node
);
388 struct vmap_area
*prev
;
389 prev
= rb_entry(tmp
, struct vmap_area
, rb_node
);
390 list_add_rcu(&va
->list
, &prev
->list
);
392 list_add_rcu(&va
->list
, &vmap_area_list
);
395 static void purge_vmap_area_lazy(void);
397 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
400 * Allocate a region of KVA of the specified size and alignment, within the
403 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
405 unsigned long vstart
, unsigned long vend
,
406 int node
, gfp_t gfp_mask
)
408 struct vmap_area
*va
;
412 struct vmap_area
*first
;
415 BUG_ON(offset_in_page(size
));
416 BUG_ON(!is_power_of_2(align
));
420 va
= kmalloc_node(sizeof(struct vmap_area
),
421 gfp_mask
& GFP_RECLAIM_MASK
, node
);
423 return ERR_PTR(-ENOMEM
);
426 * Only scan the relevant parts containing pointers to other objects
427 * to avoid false negatives.
429 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
& GFP_RECLAIM_MASK
);
432 spin_lock(&vmap_area_lock
);
434 * Invalidate cache if we have more permissive parameters.
435 * cached_hole_size notes the largest hole noticed _below_
436 * the vmap_area cached in free_vmap_cache: if size fits
437 * into that hole, we want to scan from vstart to reuse
438 * the hole instead of allocating above free_vmap_cache.
439 * Note that __free_vmap_area may update free_vmap_cache
440 * without updating cached_hole_size or cached_align.
442 if (!free_vmap_cache
||
443 size
< cached_hole_size
||
444 vstart
< cached_vstart
||
445 align
< cached_align
) {
447 cached_hole_size
= 0;
448 free_vmap_cache
= NULL
;
450 /* record if we encounter less permissive parameters */
451 cached_vstart
= vstart
;
452 cached_align
= align
;
454 /* find starting point for our search */
455 if (free_vmap_cache
) {
456 first
= rb_entry(free_vmap_cache
, struct vmap_area
, rb_node
);
457 addr
= ALIGN(first
->va_end
, align
);
460 if (addr
+ size
< addr
)
464 addr
= ALIGN(vstart
, align
);
465 if (addr
+ size
< addr
)
468 n
= vmap_area_root
.rb_node
;
472 struct vmap_area
*tmp
;
473 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
474 if (tmp
->va_end
>= addr
) {
476 if (tmp
->va_start
<= addr
)
487 /* from the starting point, walk areas until a suitable hole is found */
488 while (addr
+ size
> first
->va_start
&& addr
+ size
<= vend
) {
489 if (addr
+ cached_hole_size
< first
->va_start
)
490 cached_hole_size
= first
->va_start
- addr
;
491 addr
= ALIGN(first
->va_end
, align
);
492 if (addr
+ size
< addr
)
495 if (list_is_last(&first
->list
, &vmap_area_list
))
498 first
= list_next_entry(first
, list
);
502 if (addr
+ size
> vend
)
506 va
->va_end
= addr
+ size
;
508 __insert_vmap_area(va
);
509 free_vmap_cache
= &va
->rb_node
;
510 spin_unlock(&vmap_area_lock
);
512 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
513 BUG_ON(va
->va_start
< vstart
);
514 BUG_ON(va
->va_end
> vend
);
519 spin_unlock(&vmap_area_lock
);
521 purge_vmap_area_lazy();
526 if (gfpflags_allow_blocking(gfp_mask
)) {
527 unsigned long freed
= 0;
528 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
535 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
536 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
539 return ERR_PTR(-EBUSY
);
542 int register_vmap_purge_notifier(struct notifier_block
*nb
)
544 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
546 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
548 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
550 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
552 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
554 static void __free_vmap_area(struct vmap_area
*va
)
556 BUG_ON(RB_EMPTY_NODE(&va
->rb_node
));
558 if (free_vmap_cache
) {
559 if (va
->va_end
< cached_vstart
) {
560 free_vmap_cache
= NULL
;
562 struct vmap_area
*cache
;
563 cache
= rb_entry(free_vmap_cache
, struct vmap_area
, rb_node
);
564 if (va
->va_start
<= cache
->va_start
) {
565 free_vmap_cache
= rb_prev(&va
->rb_node
);
567 * We don't try to update cached_hole_size or
568 * cached_align, but it won't go very wrong.
573 rb_erase(&va
->rb_node
, &vmap_area_root
);
574 RB_CLEAR_NODE(&va
->rb_node
);
575 list_del_rcu(&va
->list
);
578 * Track the highest possible candidate for pcpu area
579 * allocation. Areas outside of vmalloc area can be returned
580 * here too, consider only end addresses which fall inside
581 * vmalloc area proper.
583 if (va
->va_end
> VMALLOC_START
&& va
->va_end
<= VMALLOC_END
)
584 vmap_area_pcpu_hole
= max(vmap_area_pcpu_hole
, va
->va_end
);
586 kfree_rcu(va
, rcu_head
);
590 * Free a region of KVA allocated by alloc_vmap_area
592 static void free_vmap_area(struct vmap_area
*va
)
594 spin_lock(&vmap_area_lock
);
595 __free_vmap_area(va
);
596 spin_unlock(&vmap_area_lock
);
600 * Clear the pagetable entries of a given vmap_area
602 static void unmap_vmap_area(struct vmap_area
*va
)
604 vunmap_page_range(va
->va_start
, va
->va_end
);
607 static void vmap_debug_free_range(unsigned long start
, unsigned long end
)
610 * Unmap page tables and force a TLB flush immediately if pagealloc
611 * debugging is enabled. This catches use after free bugs similarly to
612 * those in linear kernel virtual address space after a page has been
615 * All the lazy freeing logic is still retained, in order to minimise
616 * intrusiveness of this debugging feature.
618 * This is going to be *slow* (linear kernel virtual address debugging
619 * doesn't do a broadcast TLB flush so it is a lot faster).
621 if (debug_pagealloc_enabled()) {
622 vunmap_page_range(start
, end
);
623 flush_tlb_kernel_range(start
, end
);
628 * lazy_max_pages is the maximum amount of virtual address space we gather up
629 * before attempting to purge with a TLB flush.
631 * There is a tradeoff here: a larger number will cover more kernel page tables
632 * and take slightly longer to purge, but it will linearly reduce the number of
633 * global TLB flushes that must be performed. It would seem natural to scale
634 * this number up linearly with the number of CPUs (because vmapping activity
635 * could also scale linearly with the number of CPUs), however it is likely
636 * that in practice, workloads might be constrained in other ways that mean
637 * vmap activity will not scale linearly with CPUs. Also, I want to be
638 * conservative and not introduce a big latency on huge systems, so go with
639 * a less aggressive log scale. It will still be an improvement over the old
640 * code, and it will be simple to change the scale factor if we find that it
641 * becomes a problem on bigger systems.
643 static unsigned long lazy_max_pages(void)
647 log
= fls(num_online_cpus());
649 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
652 static atomic_t vmap_lazy_nr
= ATOMIC_INIT(0);
655 * Serialize vmap purging. There is no actual criticial section protected
656 * by this look, but we want to avoid concurrent calls for performance
657 * reasons and to make the pcpu_get_vm_areas more deterministic.
659 static DEFINE_MUTEX(vmap_purge_lock
);
661 /* for per-CPU blocks */
662 static void purge_fragmented_blocks_allcpus(void);
665 * called before a call to iounmap() if the caller wants vm_area_struct's
668 void set_iounmap_nonlazy(void)
670 atomic_set(&vmap_lazy_nr
, lazy_max_pages()+1);
674 * Purges all lazily-freed vmap areas.
676 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
678 struct llist_node
*valist
;
679 struct vmap_area
*va
;
680 struct vmap_area
*n_va
;
681 bool do_free
= false;
683 lockdep_assert_held(&vmap_purge_lock
);
685 valist
= llist_del_all(&vmap_purge_list
);
686 llist_for_each_entry(va
, valist
, purge_list
) {
687 if (va
->va_start
< start
)
688 start
= va
->va_start
;
689 if (va
->va_end
> end
)
697 flush_tlb_kernel_range(start
, end
);
699 spin_lock(&vmap_area_lock
);
700 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
701 int nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
703 __free_vmap_area(va
);
704 atomic_sub(nr
, &vmap_lazy_nr
);
705 cond_resched_lock(&vmap_area_lock
);
707 spin_unlock(&vmap_area_lock
);
712 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
713 * is already purging.
715 static void try_purge_vmap_area_lazy(void)
717 if (mutex_trylock(&vmap_purge_lock
)) {
718 __purge_vmap_area_lazy(ULONG_MAX
, 0);
719 mutex_unlock(&vmap_purge_lock
);
724 * Kick off a purge of the outstanding lazy areas.
726 static void purge_vmap_area_lazy(void)
728 mutex_lock(&vmap_purge_lock
);
729 purge_fragmented_blocks_allcpus();
730 __purge_vmap_area_lazy(ULONG_MAX
, 0);
731 mutex_unlock(&vmap_purge_lock
);
735 * Free a vmap area, caller ensuring that the area has been unmapped
736 * and flush_cache_vunmap had been called for the correct range
739 static void free_vmap_area_noflush(struct vmap_area
*va
)
743 nr_lazy
= atomic_add_return((va
->va_end
- va
->va_start
) >> PAGE_SHIFT
,
746 /* After this point, we may free va at any time */
747 llist_add(&va
->purge_list
, &vmap_purge_list
);
749 if (unlikely(nr_lazy
> lazy_max_pages()))
750 try_purge_vmap_area_lazy();
754 * Free and unmap a vmap area
756 static void free_unmap_vmap_area(struct vmap_area
*va
)
758 flush_cache_vunmap(va
->va_start
, va
->va_end
);
760 free_vmap_area_noflush(va
);
763 static struct vmap_area
*find_vmap_area(unsigned long addr
)
765 struct vmap_area
*va
;
767 spin_lock(&vmap_area_lock
);
768 va
= __find_vmap_area(addr
);
769 spin_unlock(&vmap_area_lock
);
774 /*** Per cpu kva allocator ***/
777 * vmap space is limited especially on 32 bit architectures. Ensure there is
778 * room for at least 16 percpu vmap blocks per CPU.
781 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
782 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
783 * instead (we just need a rough idea)
785 #if BITS_PER_LONG == 32
786 #define VMALLOC_SPACE (128UL*1024*1024)
788 #define VMALLOC_SPACE (128UL*1024*1024*1024)
791 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
792 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
793 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
794 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
795 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
796 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
797 #define VMAP_BBMAP_BITS \
798 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
799 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
800 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
802 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
804 static bool vmap_initialized __read_mostly
= false;
806 struct vmap_block_queue
{
808 struct list_head free
;
813 struct vmap_area
*va
;
814 unsigned long free
, dirty
;
815 unsigned long dirty_min
, dirty_max
; /*< dirty range */
816 struct list_head free_list
;
817 struct rcu_head rcu_head
;
818 struct list_head purge
;
821 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
822 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
825 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
826 * in the free path. Could get rid of this if we change the API to return a
827 * "cookie" from alloc, to be passed to free. But no big deal yet.
829 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
830 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
833 * We should probably have a fallback mechanism to allocate virtual memory
834 * out of partially filled vmap blocks. However vmap block sizing should be
835 * fairly reasonable according to the vmalloc size, so it shouldn't be a
839 static unsigned long addr_to_vb_idx(unsigned long addr
)
841 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
842 addr
/= VMAP_BLOCK_SIZE
;
846 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
850 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
851 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
856 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
857 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
858 * @order: how many 2^order pages should be occupied in newly allocated block
859 * @gfp_mask: flags for the page level allocator
861 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
863 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
865 struct vmap_block_queue
*vbq
;
866 struct vmap_block
*vb
;
867 struct vmap_area
*va
;
868 unsigned long vb_idx
;
872 node
= numa_node_id();
874 vb
= kmalloc_node(sizeof(struct vmap_block
),
875 gfp_mask
& GFP_RECLAIM_MASK
, node
);
877 return ERR_PTR(-ENOMEM
);
879 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
880 VMALLOC_START
, VMALLOC_END
,
887 err
= radix_tree_preload(gfp_mask
);
894 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
895 spin_lock_init(&vb
->lock
);
897 /* At least something should be left free */
898 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
899 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
901 vb
->dirty_min
= VMAP_BBMAP_BITS
;
903 INIT_LIST_HEAD(&vb
->free_list
);
905 vb_idx
= addr_to_vb_idx(va
->va_start
);
906 spin_lock(&vmap_block_tree_lock
);
907 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
908 spin_unlock(&vmap_block_tree_lock
);
910 radix_tree_preload_end();
912 vbq
= &get_cpu_var(vmap_block_queue
);
913 spin_lock(&vbq
->lock
);
914 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
915 spin_unlock(&vbq
->lock
);
916 put_cpu_var(vmap_block_queue
);
921 static void free_vmap_block(struct vmap_block
*vb
)
923 struct vmap_block
*tmp
;
924 unsigned long vb_idx
;
926 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
927 spin_lock(&vmap_block_tree_lock
);
928 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
929 spin_unlock(&vmap_block_tree_lock
);
932 free_vmap_area_noflush(vb
->va
);
933 kfree_rcu(vb
, rcu_head
);
936 static void purge_fragmented_blocks(int cpu
)
939 struct vmap_block
*vb
;
940 struct vmap_block
*n_vb
;
941 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
944 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
946 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
949 spin_lock(&vb
->lock
);
950 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
951 vb
->free
= 0; /* prevent further allocs after releasing lock */
952 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
954 vb
->dirty_max
= VMAP_BBMAP_BITS
;
955 spin_lock(&vbq
->lock
);
956 list_del_rcu(&vb
->free_list
);
957 spin_unlock(&vbq
->lock
);
958 spin_unlock(&vb
->lock
);
959 list_add_tail(&vb
->purge
, &purge
);
961 spin_unlock(&vb
->lock
);
965 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
966 list_del(&vb
->purge
);
971 static void purge_fragmented_blocks_allcpus(void)
975 for_each_possible_cpu(cpu
)
976 purge_fragmented_blocks(cpu
);
979 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
981 struct vmap_block_queue
*vbq
;
982 struct vmap_block
*vb
;
986 BUG_ON(offset_in_page(size
));
987 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
988 if (WARN_ON(size
== 0)) {
990 * Allocating 0 bytes isn't what caller wants since
991 * get_order(0) returns funny result. Just warn and terminate
996 order
= get_order(size
);
999 vbq
= &get_cpu_var(vmap_block_queue
);
1000 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1001 unsigned long pages_off
;
1003 spin_lock(&vb
->lock
);
1004 if (vb
->free
< (1UL << order
)) {
1005 spin_unlock(&vb
->lock
);
1009 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1010 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1011 vb
->free
-= 1UL << order
;
1012 if (vb
->free
== 0) {
1013 spin_lock(&vbq
->lock
);
1014 list_del_rcu(&vb
->free_list
);
1015 spin_unlock(&vbq
->lock
);
1018 spin_unlock(&vb
->lock
);
1022 put_cpu_var(vmap_block_queue
);
1025 /* Allocate new block if nothing was found */
1027 vaddr
= new_vmap_block(order
, gfp_mask
);
1032 static void vb_free(const void *addr
, unsigned long size
)
1034 unsigned long offset
;
1035 unsigned long vb_idx
;
1037 struct vmap_block
*vb
;
1039 BUG_ON(offset_in_page(size
));
1040 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1042 flush_cache_vunmap((unsigned long)addr
, (unsigned long)addr
+ size
);
1044 order
= get_order(size
);
1046 offset
= (unsigned long)addr
& (VMAP_BLOCK_SIZE
- 1);
1047 offset
>>= PAGE_SHIFT
;
1049 vb_idx
= addr_to_vb_idx((unsigned long)addr
);
1051 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1055 vunmap_page_range((unsigned long)addr
, (unsigned long)addr
+ size
);
1057 spin_lock(&vb
->lock
);
1059 /* Expand dirty range */
1060 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1061 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1063 vb
->dirty
+= 1UL << order
;
1064 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1066 spin_unlock(&vb
->lock
);
1067 free_vmap_block(vb
);
1069 spin_unlock(&vb
->lock
);
1073 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1075 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1076 * to amortize TLB flushing overheads. What this means is that any page you
1077 * have now, may, in a former life, have been mapped into kernel virtual
1078 * address by the vmap layer and so there might be some CPUs with TLB entries
1079 * still referencing that page (additional to the regular 1:1 kernel mapping).
1081 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1082 * be sure that none of the pages we have control over will have any aliases
1083 * from the vmap layer.
1085 void vm_unmap_aliases(void)
1087 unsigned long start
= ULONG_MAX
, end
= 0;
1091 if (unlikely(!vmap_initialized
))
1096 for_each_possible_cpu(cpu
) {
1097 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1098 struct vmap_block
*vb
;
1101 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1102 spin_lock(&vb
->lock
);
1104 unsigned long va_start
= vb
->va
->va_start
;
1107 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1108 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1110 start
= min(s
, start
);
1115 spin_unlock(&vb
->lock
);
1120 mutex_lock(&vmap_purge_lock
);
1121 purge_fragmented_blocks_allcpus();
1122 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1123 flush_tlb_kernel_range(start
, end
);
1124 mutex_unlock(&vmap_purge_lock
);
1126 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1129 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1130 * @mem: the pointer returned by vm_map_ram
1131 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1133 void vm_unmap_ram(const void *mem
, unsigned int count
)
1135 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1136 unsigned long addr
= (unsigned long)mem
;
1137 struct vmap_area
*va
;
1141 BUG_ON(addr
< VMALLOC_START
);
1142 BUG_ON(addr
> VMALLOC_END
);
1143 BUG_ON(!PAGE_ALIGNED(addr
));
1145 debug_check_no_locks_freed(mem
, size
);
1146 vmap_debug_free_range(addr
, addr
+size
);
1148 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1153 va
= find_vmap_area(addr
);
1155 free_unmap_vmap_area(va
);
1157 EXPORT_SYMBOL(vm_unmap_ram
);
1160 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1161 * @pages: an array of pointers to the pages to be mapped
1162 * @count: number of pages
1163 * @node: prefer to allocate data structures on this node
1164 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1166 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1167 * faster than vmap so it's good. But if you mix long-life and short-life
1168 * objects with vm_map_ram(), it could consume lots of address space through
1169 * fragmentation (especially on a 32bit machine). You could see failures in
1170 * the end. Please use this function for short-lived objects.
1172 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1174 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
, pgprot_t prot
)
1176 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1180 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1181 mem
= vb_alloc(size
, GFP_KERNEL
);
1184 addr
= (unsigned long)mem
;
1186 struct vmap_area
*va
;
1187 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1188 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1192 addr
= va
->va_start
;
1195 if (vmap_page_range(addr
, addr
+ size
, prot
, pages
) < 0) {
1196 vm_unmap_ram(mem
, count
);
1201 EXPORT_SYMBOL(vm_map_ram
);
1203 static struct vm_struct
*vmlist __initdata
;
1205 * vm_area_add_early - add vmap area early during boot
1206 * @vm: vm_struct to add
1208 * This function is used to add fixed kernel vm area to vmlist before
1209 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1210 * should contain proper values and the other fields should be zero.
1212 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1214 void __init
vm_area_add_early(struct vm_struct
*vm
)
1216 struct vm_struct
*tmp
, **p
;
1218 BUG_ON(vmap_initialized
);
1219 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1220 if (tmp
->addr
>= vm
->addr
) {
1221 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1224 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1231 * vm_area_register_early - register vmap area early during boot
1232 * @vm: vm_struct to register
1233 * @align: requested alignment
1235 * This function is used to register kernel vm area before
1236 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1237 * proper values on entry and other fields should be zero. On return,
1238 * vm->addr contains the allocated address.
1240 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1242 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1244 static size_t vm_init_off __initdata
;
1247 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1248 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1250 vm
->addr
= (void *)addr
;
1252 vm_area_add_early(vm
);
1255 void __init
vmalloc_init(void)
1257 struct vmap_area
*va
;
1258 struct vm_struct
*tmp
;
1261 for_each_possible_cpu(i
) {
1262 struct vmap_block_queue
*vbq
;
1263 struct vfree_deferred
*p
;
1265 vbq
= &per_cpu(vmap_block_queue
, i
);
1266 spin_lock_init(&vbq
->lock
);
1267 INIT_LIST_HEAD(&vbq
->free
);
1268 p
= &per_cpu(vfree_deferred
, i
);
1269 init_llist_head(&p
->list
);
1270 INIT_WORK(&p
->wq
, free_work
);
1273 /* Import existing vmlist entries. */
1274 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1275 va
= kzalloc(sizeof(struct vmap_area
), GFP_NOWAIT
);
1276 va
->flags
= VM_VM_AREA
;
1277 va
->va_start
= (unsigned long)tmp
->addr
;
1278 va
->va_end
= va
->va_start
+ tmp
->size
;
1280 __insert_vmap_area(va
);
1283 vmap_area_pcpu_hole
= VMALLOC_END
;
1285 vmap_initialized
= true;
1289 * map_kernel_range_noflush - map kernel VM area with the specified pages
1290 * @addr: start of the VM area to map
1291 * @size: size of the VM area to map
1292 * @prot: page protection flags to use
1293 * @pages: pages to map
1295 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1296 * specify should have been allocated using get_vm_area() and its
1300 * This function does NOT do any cache flushing. The caller is
1301 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1302 * before calling this function.
1305 * The number of pages mapped on success, -errno on failure.
1307 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
1308 pgprot_t prot
, struct page
**pages
)
1310 return vmap_page_range_noflush(addr
, addr
+ size
, prot
, pages
);
1314 * unmap_kernel_range_noflush - unmap kernel VM area
1315 * @addr: start of the VM area to unmap
1316 * @size: size of the VM area to unmap
1318 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1319 * specify should have been allocated using get_vm_area() and its
1323 * This function does NOT do any cache flushing. The caller is
1324 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1325 * before calling this function and flush_tlb_kernel_range() after.
1327 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
1329 vunmap_page_range(addr
, addr
+ size
);
1331 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush
);
1334 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1335 * @addr: start of the VM area to unmap
1336 * @size: size of the VM area to unmap
1338 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1339 * the unmapping and tlb after.
1341 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
1343 unsigned long end
= addr
+ size
;
1345 flush_cache_vunmap(addr
, end
);
1346 vunmap_page_range(addr
, end
);
1347 flush_tlb_kernel_range(addr
, end
);
1349 EXPORT_SYMBOL_GPL(unmap_kernel_range
);
1351 int map_vm_area(struct vm_struct
*area
, pgprot_t prot
, struct page
**pages
)
1353 unsigned long addr
= (unsigned long)area
->addr
;
1354 unsigned long end
= addr
+ get_vm_area_size(area
);
1357 err
= vmap_page_range(addr
, end
, prot
, pages
);
1359 return err
> 0 ? 0 : err
;
1361 EXPORT_SYMBOL_GPL(map_vm_area
);
1363 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
1364 unsigned long flags
, const void *caller
)
1366 spin_lock(&vmap_area_lock
);
1368 vm
->addr
= (void *)va
->va_start
;
1369 vm
->size
= va
->va_end
- va
->va_start
;
1370 vm
->caller
= caller
;
1372 va
->flags
|= VM_VM_AREA
;
1373 spin_unlock(&vmap_area_lock
);
1376 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
1379 * Before removing VM_UNINITIALIZED,
1380 * we should make sure that vm has proper values.
1381 * Pair with smp_rmb() in show_numa_info().
1384 vm
->flags
&= ~VM_UNINITIALIZED
;
1387 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
1388 unsigned long align
, unsigned long flags
, unsigned long start
,
1389 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
1391 struct vmap_area
*va
;
1392 struct vm_struct
*area
;
1394 BUG_ON(in_interrupt());
1395 size
= PAGE_ALIGN(size
);
1396 if (unlikely(!size
))
1399 if (flags
& VM_IOREMAP
)
1400 align
= 1ul << clamp_t(int, get_count_order_long(size
),
1401 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
1403 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
1404 if (unlikely(!area
))
1407 if (!(flags
& VM_NO_GUARD
))
1410 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
1416 setup_vmalloc_vm(area
, va
, flags
, caller
);
1421 struct vm_struct
*__get_vm_area(unsigned long size
, unsigned long flags
,
1422 unsigned long start
, unsigned long end
)
1424 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
1425 GFP_KERNEL
, __builtin_return_address(0));
1427 EXPORT_SYMBOL_GPL(__get_vm_area
);
1429 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
1430 unsigned long start
, unsigned long end
,
1433 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
1434 GFP_KERNEL
, caller
);
1438 * get_vm_area - reserve a contiguous kernel virtual area
1439 * @size: size of the area
1440 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1442 * Search an area of @size in the kernel virtual mapping area,
1443 * and reserved it for out purposes. Returns the area descriptor
1444 * on success or %NULL on failure.
1446 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
1448 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1449 NUMA_NO_NODE
, GFP_KERNEL
,
1450 __builtin_return_address(0));
1453 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
1456 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1457 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
1461 * find_vm_area - find a continuous kernel virtual area
1462 * @addr: base address
1464 * Search for the kernel VM area starting at @addr, and return it.
1465 * It is up to the caller to do all required locking to keep the returned
1468 struct vm_struct
*find_vm_area(const void *addr
)
1470 struct vmap_area
*va
;
1472 va
= find_vmap_area((unsigned long)addr
);
1473 if (va
&& va
->flags
& VM_VM_AREA
)
1480 * remove_vm_area - find and remove a continuous kernel virtual area
1481 * @addr: base address
1483 * Search for the kernel VM area starting at @addr, and remove it.
1484 * This function returns the found VM area, but using it is NOT safe
1485 * on SMP machines, except for its size or flags.
1487 struct vm_struct
*remove_vm_area(const void *addr
)
1489 struct vmap_area
*va
;
1493 va
= find_vmap_area((unsigned long)addr
);
1494 if (va
&& va
->flags
& VM_VM_AREA
) {
1495 struct vm_struct
*vm
= va
->vm
;
1497 spin_lock(&vmap_area_lock
);
1499 va
->flags
&= ~VM_VM_AREA
;
1500 spin_unlock(&vmap_area_lock
);
1502 vmap_debug_free_range(va
->va_start
, va
->va_end
);
1503 kasan_free_shadow(vm
);
1504 free_unmap_vmap_area(va
);
1511 static void __vunmap(const void *addr
, int deallocate_pages
)
1513 struct vm_struct
*area
;
1518 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
1522 area
= remove_vm_area(addr
);
1523 if (unlikely(!area
)) {
1524 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
1529 debug_check_no_locks_freed(addr
, get_vm_area_size(area
));
1530 debug_check_no_obj_freed(addr
, get_vm_area_size(area
));
1532 if (deallocate_pages
) {
1535 for (i
= 0; i
< area
->nr_pages
; i
++) {
1536 struct page
*page
= area
->pages
[i
];
1539 __free_pages(page
, 0);
1542 kvfree(area
->pages
);
1549 static inline void __vfree_deferred(const void *addr
)
1552 * Use raw_cpu_ptr() because this can be called from preemptible
1553 * context. Preemption is absolutely fine here, because the llist_add()
1554 * implementation is lockless, so it works even if we are adding to
1555 * nother cpu's list. schedule_work() should be fine with this too.
1557 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
1559 if (llist_add((struct llist_node
*)addr
, &p
->list
))
1560 schedule_work(&p
->wq
);
1564 * vfree_atomic - release memory allocated by vmalloc()
1565 * @addr: memory base address
1567 * This one is just like vfree() but can be called in any atomic context
1570 void vfree_atomic(const void *addr
)
1574 kmemleak_free(addr
);
1578 __vfree_deferred(addr
);
1582 * vfree - release memory allocated by vmalloc()
1583 * @addr: memory base address
1585 * Free the virtually continuous memory area starting at @addr, as
1586 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1587 * NULL, no operation is performed.
1589 * Must not be called in NMI context (strictly speaking, only if we don't
1590 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1591 * conventions for vfree() arch-depenedent would be a really bad idea)
1593 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1595 void vfree(const void *addr
)
1599 kmemleak_free(addr
);
1603 if (unlikely(in_interrupt()))
1604 __vfree_deferred(addr
);
1608 EXPORT_SYMBOL(vfree
);
1611 * vunmap - release virtual mapping obtained by vmap()
1612 * @addr: memory base address
1614 * Free the virtually contiguous memory area starting at @addr,
1615 * which was created from the page array passed to vmap().
1617 * Must not be called in interrupt context.
1619 void vunmap(const void *addr
)
1621 BUG_ON(in_interrupt());
1626 EXPORT_SYMBOL(vunmap
);
1629 * vmap - map an array of pages into virtually contiguous space
1630 * @pages: array of page pointers
1631 * @count: number of pages to map
1632 * @flags: vm_area->flags
1633 * @prot: page protection for the mapping
1635 * Maps @count pages from @pages into contiguous kernel virtual
1638 void *vmap(struct page
**pages
, unsigned int count
,
1639 unsigned long flags
, pgprot_t prot
)
1641 struct vm_struct
*area
;
1642 unsigned long size
; /* In bytes */
1646 if (count
> totalram_pages
)
1649 size
= (unsigned long)count
<< PAGE_SHIFT
;
1650 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
1654 if (map_vm_area(area
, prot
, pages
)) {
1661 EXPORT_SYMBOL(vmap
);
1663 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1664 gfp_t gfp_mask
, pgprot_t prot
,
1665 int node
, const void *caller
);
1666 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
1667 pgprot_t prot
, int node
)
1669 struct page
**pages
;
1670 unsigned int nr_pages
, array_size
, i
;
1671 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
1672 const gfp_t alloc_mask
= gfp_mask
| __GFP_HIGHMEM
| __GFP_NOWARN
;
1674 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
1675 array_size
= (nr_pages
* sizeof(struct page
*));
1677 area
->nr_pages
= nr_pages
;
1678 /* Please note that the recursion is strictly bounded. */
1679 if (array_size
> PAGE_SIZE
) {
1680 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|__GFP_HIGHMEM
,
1681 PAGE_KERNEL
, node
, area
->caller
);
1683 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
1685 area
->pages
= pages
;
1687 remove_vm_area(area
->addr
);
1692 for (i
= 0; i
< area
->nr_pages
; i
++) {
1695 if (fatal_signal_pending(current
)) {
1700 if (node
== NUMA_NO_NODE
)
1701 page
= alloc_page(alloc_mask
);
1703 page
= alloc_pages_node(node
, alloc_mask
, 0);
1705 if (unlikely(!page
)) {
1706 /* Successfully allocated i pages, free them in __vunmap() */
1710 area
->pages
[i
] = page
;
1711 if (gfpflags_allow_blocking(gfp_mask
))
1715 if (map_vm_area(area
, prot
, pages
))
1720 warn_alloc(gfp_mask
, NULL
,
1721 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1722 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
1729 * __vmalloc_node_range - allocate virtually contiguous memory
1730 * @size: allocation size
1731 * @align: desired alignment
1732 * @start: vm area range start
1733 * @end: vm area range end
1734 * @gfp_mask: flags for the page level allocator
1735 * @prot: protection mask for the allocated pages
1736 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1737 * @node: node to use for allocation or NUMA_NO_NODE
1738 * @caller: caller's return address
1740 * Allocate enough pages to cover @size from the page level
1741 * allocator with @gfp_mask flags. Map them into contiguous
1742 * kernel virtual space, using a pagetable protection of @prot.
1744 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
1745 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
1746 pgprot_t prot
, unsigned long vm_flags
, int node
,
1749 struct vm_struct
*area
;
1751 unsigned long real_size
= size
;
1753 size
= PAGE_ALIGN(size
);
1754 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages
)
1757 area
= __get_vm_area_node(size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
1758 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
1762 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
1767 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1768 * flag. It means that vm_struct is not fully initialized.
1769 * Now, it is fully initialized, so remove this flag here.
1771 clear_vm_uninitialized_flag(area
);
1774 * A ref_count = 2 is needed because vm_struct allocated in
1775 * __get_vm_area_node() contains a reference to the virtual address of
1776 * the vmalloc'ed block.
1778 kmemleak_alloc(addr
, real_size
, 2, gfp_mask
);
1783 warn_alloc(gfp_mask
, NULL
,
1784 "vmalloc: allocation failure: %lu bytes", real_size
);
1789 * __vmalloc_node - allocate virtually contiguous memory
1790 * @size: allocation size
1791 * @align: desired alignment
1792 * @gfp_mask: flags for the page level allocator
1793 * @prot: protection mask for the allocated pages
1794 * @node: node to use for allocation or NUMA_NO_NODE
1795 * @caller: caller's return address
1797 * Allocate enough pages to cover @size from the page level
1798 * allocator with @gfp_mask flags. Map them into contiguous
1799 * kernel virtual space, using a pagetable protection of @prot.
1801 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_REPEAT
1802 * and __GFP_NOFAIL are not supported
1804 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1808 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1809 gfp_t gfp_mask
, pgprot_t prot
,
1810 int node
, const void *caller
)
1812 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
1813 gfp_mask
, prot
, 0, node
, caller
);
1816 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
, pgprot_t prot
)
1818 return __vmalloc_node(size
, 1, gfp_mask
, prot
, NUMA_NO_NODE
,
1819 __builtin_return_address(0));
1821 EXPORT_SYMBOL(__vmalloc
);
1823 static inline void *__vmalloc_node_flags(unsigned long size
,
1824 int node
, gfp_t flags
)
1826 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
,
1827 node
, __builtin_return_address(0));
1831 void *__vmalloc_node_flags_caller(unsigned long size
, int node
, gfp_t flags
,
1834 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
, node
, caller
);
1838 * vmalloc - allocate virtually contiguous memory
1839 * @size: allocation size
1840 * Allocate enough pages to cover @size from the page level
1841 * allocator and map them into contiguous kernel virtual space.
1843 * For tight control over page level allocator and protection flags
1844 * use __vmalloc() instead.
1846 void *vmalloc(unsigned long size
)
1848 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
1851 EXPORT_SYMBOL(vmalloc
);
1854 * vzalloc - allocate virtually contiguous memory with zero fill
1855 * @size: allocation size
1856 * Allocate enough pages to cover @size from the page level
1857 * allocator and map them into contiguous kernel virtual space.
1858 * The memory allocated is set to zero.
1860 * For tight control over page level allocator and protection flags
1861 * use __vmalloc() instead.
1863 void *vzalloc(unsigned long size
)
1865 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
1866 GFP_KERNEL
| __GFP_ZERO
);
1868 EXPORT_SYMBOL(vzalloc
);
1871 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1872 * @size: allocation size
1874 * The resulting memory area is zeroed so it can be mapped to userspace
1875 * without leaking data.
1877 void *vmalloc_user(unsigned long size
)
1879 struct vm_struct
*area
;
1882 ret
= __vmalloc_node(size
, SHMLBA
,
1883 GFP_KERNEL
| __GFP_ZERO
,
1884 PAGE_KERNEL
, NUMA_NO_NODE
,
1885 __builtin_return_address(0));
1887 area
= find_vm_area(ret
);
1888 area
->flags
|= VM_USERMAP
;
1892 EXPORT_SYMBOL(vmalloc_user
);
1895 * vmalloc_node - allocate memory on a specific node
1896 * @size: allocation size
1899 * Allocate enough pages to cover @size from the page level
1900 * allocator and map them into contiguous kernel virtual space.
1902 * For tight control over page level allocator and protection flags
1903 * use __vmalloc() instead.
1905 void *vmalloc_node(unsigned long size
, int node
)
1907 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL
,
1908 node
, __builtin_return_address(0));
1910 EXPORT_SYMBOL(vmalloc_node
);
1913 * vzalloc_node - allocate memory on a specific node with zero fill
1914 * @size: allocation size
1917 * Allocate enough pages to cover @size from the page level
1918 * allocator and map them into contiguous kernel virtual space.
1919 * The memory allocated is set to zero.
1921 * For tight control over page level allocator and protection flags
1922 * use __vmalloc_node() instead.
1924 void *vzalloc_node(unsigned long size
, int node
)
1926 return __vmalloc_node_flags(size
, node
,
1927 GFP_KERNEL
| __GFP_ZERO
);
1929 EXPORT_SYMBOL(vzalloc_node
);
1931 #ifndef PAGE_KERNEL_EXEC
1932 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1936 * vmalloc_exec - allocate virtually contiguous, executable memory
1937 * @size: allocation size
1939 * Kernel-internal function to allocate enough pages to cover @size
1940 * the page level allocator and map them into contiguous and
1941 * executable kernel virtual space.
1943 * For tight control over page level allocator and protection flags
1944 * use __vmalloc() instead.
1947 void *vmalloc_exec(unsigned long size
)
1949 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL_EXEC
,
1950 NUMA_NO_NODE
, __builtin_return_address(0));
1953 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1954 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1955 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1956 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1958 #define GFP_VMALLOC32 GFP_KERNEL
1962 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1963 * @size: allocation size
1965 * Allocate enough 32bit PA addressable pages to cover @size from the
1966 * page level allocator and map them into contiguous kernel virtual space.
1968 void *vmalloc_32(unsigned long size
)
1970 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, PAGE_KERNEL
,
1971 NUMA_NO_NODE
, __builtin_return_address(0));
1973 EXPORT_SYMBOL(vmalloc_32
);
1976 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1977 * @size: allocation size
1979 * The resulting memory area is 32bit addressable and zeroed so it can be
1980 * mapped to userspace without leaking data.
1982 void *vmalloc_32_user(unsigned long size
)
1984 struct vm_struct
*area
;
1987 ret
= __vmalloc_node(size
, 1, GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
1988 NUMA_NO_NODE
, __builtin_return_address(0));
1990 area
= find_vm_area(ret
);
1991 area
->flags
|= VM_USERMAP
;
1995 EXPORT_SYMBOL(vmalloc_32_user
);
1998 * small helper routine , copy contents to buf from addr.
1999 * If the page is not present, fill zero.
2002 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2008 unsigned long offset
, length
;
2010 offset
= offset_in_page(addr
);
2011 length
= PAGE_SIZE
- offset
;
2014 p
= vmalloc_to_page(addr
);
2016 * To do safe access to this _mapped_ area, we need
2017 * lock. But adding lock here means that we need to add
2018 * overhead of vmalloc()/vfree() calles for this _debug_
2019 * interface, rarely used. Instead of that, we'll use
2020 * kmap() and get small overhead in this access function.
2024 * we can expect USER0 is not used (see vread/vwrite's
2025 * function description)
2027 void *map
= kmap_atomic(p
);
2028 memcpy(buf
, map
+ offset
, length
);
2031 memset(buf
, 0, length
);
2041 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2047 unsigned long offset
, length
;
2049 offset
= offset_in_page(addr
);
2050 length
= PAGE_SIZE
- offset
;
2053 p
= vmalloc_to_page(addr
);
2055 * To do safe access to this _mapped_ area, we need
2056 * lock. But adding lock here means that we need to add
2057 * overhead of vmalloc()/vfree() calles for this _debug_
2058 * interface, rarely used. Instead of that, we'll use
2059 * kmap() and get small overhead in this access function.
2063 * we can expect USER0 is not used (see vread/vwrite's
2064 * function description)
2066 void *map
= kmap_atomic(p
);
2067 memcpy(map
+ offset
, buf
, length
);
2079 * vread() - read vmalloc area in a safe way.
2080 * @buf: buffer for reading data
2081 * @addr: vm address.
2082 * @count: number of bytes to be read.
2084 * Returns # of bytes which addr and buf should be increased.
2085 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2086 * includes any intersect with alive vmalloc area.
2088 * This function checks that addr is a valid vmalloc'ed area, and
2089 * copy data from that area to a given buffer. If the given memory range
2090 * of [addr...addr+count) includes some valid address, data is copied to
2091 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2092 * IOREMAP area is treated as memory hole and no copy is done.
2094 * If [addr...addr+count) doesn't includes any intersects with alive
2095 * vm_struct area, returns 0. @buf should be kernel's buffer.
2097 * Note: In usual ops, vread() is never necessary because the caller
2098 * should know vmalloc() area is valid and can use memcpy().
2099 * This is for routines which have to access vmalloc area without
2100 * any informaion, as /dev/kmem.
2104 long vread(char *buf
, char *addr
, unsigned long count
)
2106 struct vmap_area
*va
;
2107 struct vm_struct
*vm
;
2108 char *vaddr
, *buf_start
= buf
;
2109 unsigned long buflen
= count
;
2112 /* Don't allow overflow */
2113 if ((unsigned long) addr
+ count
< count
)
2114 count
= -(unsigned long) addr
;
2116 spin_lock(&vmap_area_lock
);
2117 list_for_each_entry(va
, &vmap_area_list
, list
) {
2121 if (!(va
->flags
& VM_VM_AREA
))
2125 vaddr
= (char *) vm
->addr
;
2126 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2128 while (addr
< vaddr
) {
2136 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2139 if (!(vm
->flags
& VM_IOREMAP
))
2140 aligned_vread(buf
, addr
, n
);
2141 else /* IOREMAP area is treated as memory hole */
2148 spin_unlock(&vmap_area_lock
);
2150 if (buf
== buf_start
)
2152 /* zero-fill memory holes */
2153 if (buf
!= buf_start
+ buflen
)
2154 memset(buf
, 0, buflen
- (buf
- buf_start
));
2160 * vwrite() - write vmalloc area in a safe way.
2161 * @buf: buffer for source data
2162 * @addr: vm address.
2163 * @count: number of bytes to be read.
2165 * Returns # of bytes which addr and buf should be incresed.
2166 * (same number to @count).
2167 * If [addr...addr+count) doesn't includes any intersect with valid
2168 * vmalloc area, returns 0.
2170 * This function checks that addr is a valid vmalloc'ed area, and
2171 * copy data from a buffer to the given addr. If specified range of
2172 * [addr...addr+count) includes some valid address, data is copied from
2173 * proper area of @buf. If there are memory holes, no copy to hole.
2174 * IOREMAP area is treated as memory hole and no copy is done.
2176 * If [addr...addr+count) doesn't includes any intersects with alive
2177 * vm_struct area, returns 0. @buf should be kernel's buffer.
2179 * Note: In usual ops, vwrite() is never necessary because the caller
2180 * should know vmalloc() area is valid and can use memcpy().
2181 * This is for routines which have to access vmalloc area without
2182 * any informaion, as /dev/kmem.
2185 long vwrite(char *buf
, char *addr
, unsigned long count
)
2187 struct vmap_area
*va
;
2188 struct vm_struct
*vm
;
2190 unsigned long n
, buflen
;
2193 /* Don't allow overflow */
2194 if ((unsigned long) addr
+ count
< count
)
2195 count
= -(unsigned long) addr
;
2198 spin_lock(&vmap_area_lock
);
2199 list_for_each_entry(va
, &vmap_area_list
, list
) {
2203 if (!(va
->flags
& VM_VM_AREA
))
2207 vaddr
= (char *) vm
->addr
;
2208 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2210 while (addr
< vaddr
) {
2217 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2220 if (!(vm
->flags
& VM_IOREMAP
)) {
2221 aligned_vwrite(buf
, addr
, n
);
2229 spin_unlock(&vmap_area_lock
);
2236 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2237 * @vma: vma to cover
2238 * @uaddr: target user address to start at
2239 * @kaddr: virtual address of vmalloc kernel memory
2240 * @size: size of map area
2242 * Returns: 0 for success, -Exxx on failure
2244 * This function checks that @kaddr is a valid vmalloc'ed area,
2245 * and that it is big enough to cover the range starting at
2246 * @uaddr in @vma. Will return failure if that criteria isn't
2249 * Similar to remap_pfn_range() (see mm/memory.c)
2251 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
2252 void *kaddr
, unsigned long size
)
2254 struct vm_struct
*area
;
2256 size
= PAGE_ALIGN(size
);
2258 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
2261 area
= find_vm_area(kaddr
);
2265 if (!(area
->flags
& VM_USERMAP
))
2268 if (kaddr
+ size
> area
->addr
+ area
->size
)
2272 struct page
*page
= vmalloc_to_page(kaddr
);
2275 ret
= vm_insert_page(vma
, uaddr
, page
);
2284 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
2288 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
2291 * remap_vmalloc_range - map vmalloc pages to userspace
2292 * @vma: vma to cover (map full range of vma)
2293 * @addr: vmalloc memory
2294 * @pgoff: number of pages into addr before first page to map
2296 * Returns: 0 for success, -Exxx on failure
2298 * This function checks that addr is a valid vmalloc'ed area, and
2299 * that it is big enough to cover the vma. Will return failure if
2300 * that criteria isn't met.
2302 * Similar to remap_pfn_range() (see mm/memory.c)
2304 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
2305 unsigned long pgoff
)
2307 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
2308 addr
+ (pgoff
<< PAGE_SHIFT
),
2309 vma
->vm_end
- vma
->vm_start
);
2311 EXPORT_SYMBOL(remap_vmalloc_range
);
2314 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2317 void __weak
vmalloc_sync_all(void)
2322 static int f(pte_t
*pte
, pgtable_t table
, unsigned long addr
, void *data
)
2334 * alloc_vm_area - allocate a range of kernel address space
2335 * @size: size of the area
2336 * @ptes: returns the PTEs for the address space
2338 * Returns: NULL on failure, vm_struct on success
2340 * This function reserves a range of kernel address space, and
2341 * allocates pagetables to map that range. No actual mappings
2344 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2345 * allocated for the VM area are returned.
2347 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
2349 struct vm_struct
*area
;
2351 area
= get_vm_area_caller(size
, VM_IOREMAP
,
2352 __builtin_return_address(0));
2357 * This ensures that page tables are constructed for this region
2358 * of kernel virtual address space and mapped into init_mm.
2360 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2361 size
, f
, ptes
? &ptes
: NULL
)) {
2368 EXPORT_SYMBOL_GPL(alloc_vm_area
);
2370 void free_vm_area(struct vm_struct
*area
)
2372 struct vm_struct
*ret
;
2373 ret
= remove_vm_area(area
->addr
);
2374 BUG_ON(ret
!= area
);
2377 EXPORT_SYMBOL_GPL(free_vm_area
);
2380 static struct vmap_area
*node_to_va(struct rb_node
*n
)
2382 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
2386 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2387 * @end: target address
2388 * @pnext: out arg for the next vmap_area
2389 * @pprev: out arg for the previous vmap_area
2391 * Returns: %true if either or both of next and prev are found,
2392 * %false if no vmap_area exists
2394 * Find vmap_areas end addresses of which enclose @end. ie. if not
2395 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2397 static bool pvm_find_next_prev(unsigned long end
,
2398 struct vmap_area
**pnext
,
2399 struct vmap_area
**pprev
)
2401 struct rb_node
*n
= vmap_area_root
.rb_node
;
2402 struct vmap_area
*va
= NULL
;
2405 va
= rb_entry(n
, struct vmap_area
, rb_node
);
2406 if (end
< va
->va_end
)
2408 else if (end
> va
->va_end
)
2417 if (va
->va_end
> end
) {
2419 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2422 *pnext
= node_to_va(rb_next(&(*pprev
)->rb_node
));
2428 * pvm_determine_end - find the highest aligned address between two vmap_areas
2429 * @pnext: in/out arg for the next vmap_area
2430 * @pprev: in/out arg for the previous vmap_area
2433 * Returns: determined end address
2435 * Find the highest aligned address between *@pnext and *@pprev below
2436 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2437 * down address is between the end addresses of the two vmap_areas.
2439 * Please note that the address returned by this function may fall
2440 * inside *@pnext vmap_area. The caller is responsible for checking
2443 static unsigned long pvm_determine_end(struct vmap_area
**pnext
,
2444 struct vmap_area
**pprev
,
2445 unsigned long align
)
2447 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2451 addr
= min((*pnext
)->va_start
& ~(align
- 1), vmalloc_end
);
2455 while (*pprev
&& (*pprev
)->va_end
> addr
) {
2457 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2464 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2465 * @offsets: array containing offset of each area
2466 * @sizes: array containing size of each area
2467 * @nr_vms: the number of areas to allocate
2468 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2470 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2471 * vm_structs on success, %NULL on failure
2473 * Percpu allocator wants to use congruent vm areas so that it can
2474 * maintain the offsets among percpu areas. This function allocates
2475 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2476 * be scattered pretty far, distance between two areas easily going up
2477 * to gigabytes. To avoid interacting with regular vmallocs, these
2478 * areas are allocated from top.
2480 * Despite its complicated look, this allocator is rather simple. It
2481 * does everything top-down and scans areas from the end looking for
2482 * matching slot. While scanning, if any of the areas overlaps with
2483 * existing vmap_area, the base address is pulled down to fit the
2484 * area. Scanning is repeated till all the areas fit and then all
2485 * necessary data structres are inserted and the result is returned.
2487 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
2488 const size_t *sizes
, int nr_vms
,
2491 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
2492 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2493 struct vmap_area
**vas
, *prev
, *next
;
2494 struct vm_struct
**vms
;
2495 int area
, area2
, last_area
, term_area
;
2496 unsigned long base
, start
, end
, last_end
;
2497 bool purged
= false;
2499 /* verify parameters and allocate data structures */
2500 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
2501 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
2502 start
= offsets
[area
];
2503 end
= start
+ sizes
[area
];
2505 /* is everything aligned properly? */
2506 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
2507 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
2509 /* detect the area with the highest address */
2510 if (start
> offsets
[last_area
])
2513 for (area2
= 0; area2
< nr_vms
; area2
++) {
2514 unsigned long start2
= offsets
[area2
];
2515 unsigned long end2
= start2
+ sizes
[area2
];
2520 BUG_ON(start2
>= start
&& start2
< end
);
2521 BUG_ON(end2
<= end
&& end2
> start
);
2524 last_end
= offsets
[last_area
] + sizes
[last_area
];
2526 if (vmalloc_end
- vmalloc_start
< last_end
) {
2531 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
2532 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
2536 for (area
= 0; area
< nr_vms
; area
++) {
2537 vas
[area
] = kzalloc(sizeof(struct vmap_area
), GFP_KERNEL
);
2538 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
2539 if (!vas
[area
] || !vms
[area
])
2543 spin_lock(&vmap_area_lock
);
2545 /* start scanning - we scan from the top, begin with the last area */
2546 area
= term_area
= last_area
;
2547 start
= offsets
[area
];
2548 end
= start
+ sizes
[area
];
2550 if (!pvm_find_next_prev(vmap_area_pcpu_hole
, &next
, &prev
)) {
2551 base
= vmalloc_end
- last_end
;
2554 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2557 BUG_ON(next
&& next
->va_end
<= base
+ end
);
2558 BUG_ON(prev
&& prev
->va_end
> base
+ end
);
2561 * base might have underflowed, add last_end before
2564 if (base
+ last_end
< vmalloc_start
+ last_end
) {
2565 spin_unlock(&vmap_area_lock
);
2567 purge_vmap_area_lazy();
2575 * If next overlaps, move base downwards so that it's
2576 * right below next and then recheck.
2578 if (next
&& next
->va_start
< base
+ end
) {
2579 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2585 * If prev overlaps, shift down next and prev and move
2586 * base so that it's right below new next and then
2589 if (prev
&& prev
->va_end
> base
+ start
) {
2591 prev
= node_to_va(rb_prev(&next
->rb_node
));
2592 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2598 * This area fits, move on to the previous one. If
2599 * the previous one is the terminal one, we're done.
2601 area
= (area
+ nr_vms
- 1) % nr_vms
;
2602 if (area
== term_area
)
2604 start
= offsets
[area
];
2605 end
= start
+ sizes
[area
];
2606 pvm_find_next_prev(base
+ end
, &next
, &prev
);
2609 /* we've found a fitting base, insert all va's */
2610 for (area
= 0; area
< nr_vms
; area
++) {
2611 struct vmap_area
*va
= vas
[area
];
2613 va
->va_start
= base
+ offsets
[area
];
2614 va
->va_end
= va
->va_start
+ sizes
[area
];
2615 __insert_vmap_area(va
);
2618 vmap_area_pcpu_hole
= base
+ offsets
[last_area
];
2620 spin_unlock(&vmap_area_lock
);
2622 /* insert all vm's */
2623 for (area
= 0; area
< nr_vms
; area
++)
2624 setup_vmalloc_vm(vms
[area
], vas
[area
], VM_ALLOC
,
2631 for (area
= 0; area
< nr_vms
; area
++) {
2642 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2643 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2644 * @nr_vms: the number of allocated areas
2646 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2648 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
2652 for (i
= 0; i
< nr_vms
; i
++)
2653 free_vm_area(vms
[i
]);
2656 #endif /* CONFIG_SMP */
2658 #ifdef CONFIG_PROC_FS
2659 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
2660 __acquires(&vmap_area_lock
)
2662 spin_lock(&vmap_area_lock
);
2663 return seq_list_start(&vmap_area_list
, *pos
);
2666 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
2668 return seq_list_next(p
, &vmap_area_list
, pos
);
2671 static void s_stop(struct seq_file
*m
, void *p
)
2672 __releases(&vmap_area_lock
)
2674 spin_unlock(&vmap_area_lock
);
2677 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
2679 if (IS_ENABLED(CONFIG_NUMA
)) {
2680 unsigned int nr
, *counters
= m
->private;
2685 if (v
->flags
& VM_UNINITIALIZED
)
2687 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2690 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
2692 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
2693 counters
[page_to_nid(v
->pages
[nr
])]++;
2695 for_each_node_state(nr
, N_HIGH_MEMORY
)
2697 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
2701 static int s_show(struct seq_file
*m
, void *p
)
2703 struct vmap_area
*va
;
2704 struct vm_struct
*v
;
2706 va
= list_entry(p
, struct vmap_area
, list
);
2709 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2710 * behalf of vmap area is being tear down or vm_map_ram allocation.
2712 if (!(va
->flags
& VM_VM_AREA
))
2717 seq_printf(m
, "0x%pK-0x%pK %7ld",
2718 v
->addr
, v
->addr
+ v
->size
, v
->size
);
2721 seq_printf(m
, " %pS", v
->caller
);
2724 seq_printf(m
, " pages=%d", v
->nr_pages
);
2727 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
2729 if (v
->flags
& VM_IOREMAP
)
2730 seq_puts(m
, " ioremap");
2732 if (v
->flags
& VM_ALLOC
)
2733 seq_puts(m
, " vmalloc");
2735 if (v
->flags
& VM_MAP
)
2736 seq_puts(m
, " vmap");
2738 if (v
->flags
& VM_USERMAP
)
2739 seq_puts(m
, " user");
2741 if (is_vmalloc_addr(v
->pages
))
2742 seq_puts(m
, " vpages");
2744 show_numa_info(m
, v
);
2749 static const struct seq_operations vmalloc_op
= {
2756 static int vmalloc_open(struct inode
*inode
, struct file
*file
)
2758 if (IS_ENABLED(CONFIG_NUMA
))
2759 return seq_open_private(file
, &vmalloc_op
,
2760 nr_node_ids
* sizeof(unsigned int));
2762 return seq_open(file
, &vmalloc_op
);
2765 static const struct file_operations proc_vmalloc_operations
= {
2766 .open
= vmalloc_open
,
2768 .llseek
= seq_lseek
,
2769 .release
= seq_release_private
,
2772 static int __init
proc_vmalloc_init(void)
2774 proc_create("vmallocinfo", S_IRUSR
, NULL
, &proc_vmalloc_operations
);
2777 module_init(proc_vmalloc_init
);