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.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/rbtree.h>
25 #include <linux/radix-tree.h>
26 #include <linux/rcupdate.h>
27 #include <linux/pfn.h>
28 #include <linux/kmemleak.h>
29 #include <asm/atomic.h>
30 #include <asm/uaccess.h>
31 #include <asm/tlbflush.h>
32 #include <asm/shmparam.h>
35 /*** Page table manipulation functions ***/
37 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
)
41 pte
= pte_offset_kernel(pmd
, addr
);
43 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
44 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
45 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
48 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
)
53 pmd
= pmd_offset(pud
, addr
);
55 next
= pmd_addr_end(addr
, end
);
56 if (pmd_none_or_clear_bad(pmd
))
58 vunmap_pte_range(pmd
, addr
, next
);
59 } while (pmd
++, addr
= next
, addr
!= end
);
62 static void vunmap_pud_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
)
67 pud
= pud_offset(pgd
, addr
);
69 next
= pud_addr_end(addr
, end
);
70 if (pud_none_or_clear_bad(pud
))
72 vunmap_pmd_range(pud
, addr
, next
);
73 } while (pud
++, addr
= next
, addr
!= end
);
76 static void vunmap_page_range(unsigned long addr
, unsigned long end
)
82 pgd
= pgd_offset_k(addr
);
84 next
= pgd_addr_end(addr
, end
);
85 if (pgd_none_or_clear_bad(pgd
))
87 vunmap_pud_range(pgd
, addr
, next
);
88 } while (pgd
++, addr
= next
, addr
!= end
);
91 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
92 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
97 * nr is a running index into the array which helps higher level
98 * callers keep track of where we're up to.
101 pte
= pte_alloc_kernel(pmd
, addr
);
105 struct page
*page
= pages
[*nr
];
107 if (WARN_ON(!pte_none(*pte
)))
111 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
113 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
117 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
118 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
123 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
127 next
= pmd_addr_end(addr
, end
);
128 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
130 } while (pmd
++, addr
= next
, addr
!= end
);
134 static int vmap_pud_range(pgd_t
*pgd
, unsigned long addr
,
135 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
140 pud
= pud_alloc(&init_mm
, pgd
, addr
);
144 next
= pud_addr_end(addr
, end
);
145 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
147 } while (pud
++, addr
= next
, addr
!= end
);
152 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
153 * will have pfns corresponding to the "pages" array.
155 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
157 static int vmap_page_range_noflush(unsigned long start
, unsigned long end
,
158 pgprot_t prot
, struct page
**pages
)
162 unsigned long addr
= start
;
167 pgd
= pgd_offset_k(addr
);
169 next
= pgd_addr_end(addr
, end
);
170 err
= vmap_pud_range(pgd
, addr
, next
, prot
, pages
, &nr
);
173 } while (pgd
++, addr
= next
, addr
!= end
);
178 static int vmap_page_range(unsigned long start
, unsigned long end
,
179 pgprot_t prot
, struct page
**pages
)
183 ret
= vmap_page_range_noflush(start
, end
, prot
, pages
);
184 flush_cache_vmap(start
, end
);
188 int is_vmalloc_or_module_addr(const void *x
)
191 * ARM, x86-64 and sparc64 put modules in a special place,
192 * and fall back on vmalloc() if that fails. Others
193 * just put it in the vmalloc space.
195 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
196 unsigned long addr
= (unsigned long)x
;
197 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
200 return is_vmalloc_addr(x
);
204 * Walk a vmap address to the struct page it maps.
206 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
208 unsigned long addr
= (unsigned long) vmalloc_addr
;
209 struct page
*page
= NULL
;
210 pgd_t
*pgd
= pgd_offset_k(addr
);
213 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
214 * architectures that do not vmalloc module space
216 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
218 if (!pgd_none(*pgd
)) {
219 pud_t
*pud
= pud_offset(pgd
, addr
);
220 if (!pud_none(*pud
)) {
221 pmd_t
*pmd
= pmd_offset(pud
, addr
);
222 if (!pmd_none(*pmd
)) {
225 ptep
= pte_offset_map(pmd
, addr
);
227 if (pte_present(pte
))
228 page
= pte_page(pte
);
235 EXPORT_SYMBOL(vmalloc_to_page
);
238 * Map a vmalloc()-space virtual address to the physical page frame number.
240 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
242 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
244 EXPORT_SYMBOL(vmalloc_to_pfn
);
247 /*** Global kva allocator ***/
249 #define VM_LAZY_FREE 0x01
250 #define VM_LAZY_FREEING 0x02
251 #define VM_VM_AREA 0x04
254 unsigned long va_start
;
255 unsigned long va_end
;
257 struct rb_node rb_node
; /* address sorted rbtree */
258 struct list_head list
; /* address sorted list */
259 struct list_head purge_list
; /* "lazy purge" list */
261 struct rcu_head rcu_head
;
264 static DEFINE_SPINLOCK(vmap_area_lock
);
265 static struct rb_root vmap_area_root
= RB_ROOT
;
266 static LIST_HEAD(vmap_area_list
);
267 static unsigned long vmap_area_pcpu_hole
;
269 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
271 struct rb_node
*n
= vmap_area_root
.rb_node
;
274 struct vmap_area
*va
;
276 va
= rb_entry(n
, struct vmap_area
, rb_node
);
277 if (addr
< va
->va_start
)
279 else if (addr
> va
->va_start
)
288 static void __insert_vmap_area(struct vmap_area
*va
)
290 struct rb_node
**p
= &vmap_area_root
.rb_node
;
291 struct rb_node
*parent
= NULL
;
295 struct vmap_area
*tmp
;
298 tmp
= rb_entry(parent
, struct vmap_area
, rb_node
);
299 if (va
->va_start
< tmp
->va_end
)
301 else if (va
->va_end
> tmp
->va_start
)
307 rb_link_node(&va
->rb_node
, parent
, p
);
308 rb_insert_color(&va
->rb_node
, &vmap_area_root
);
310 /* address-sort this list so it is usable like the vmlist */
311 tmp
= rb_prev(&va
->rb_node
);
313 struct vmap_area
*prev
;
314 prev
= rb_entry(tmp
, struct vmap_area
, rb_node
);
315 list_add_rcu(&va
->list
, &prev
->list
);
317 list_add_rcu(&va
->list
, &vmap_area_list
);
320 static void purge_vmap_area_lazy(void);
323 * Allocate a region of KVA of the specified size and alignment, within the
326 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
328 unsigned long vstart
, unsigned long vend
,
329 int node
, gfp_t gfp_mask
)
331 struct vmap_area
*va
;
337 BUG_ON(size
& ~PAGE_MASK
);
339 va
= kmalloc_node(sizeof(struct vmap_area
),
340 gfp_mask
& GFP_RECLAIM_MASK
, node
);
342 return ERR_PTR(-ENOMEM
);
345 addr
= ALIGN(vstart
, align
);
347 spin_lock(&vmap_area_lock
);
348 if (addr
+ size
- 1 < addr
)
351 /* XXX: could have a last_hole cache */
352 n
= vmap_area_root
.rb_node
;
354 struct vmap_area
*first
= NULL
;
357 struct vmap_area
*tmp
;
358 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
359 if (tmp
->va_end
>= addr
) {
360 if (!first
&& tmp
->va_start
< addr
+ size
)
372 if (first
->va_end
< addr
) {
373 n
= rb_next(&first
->rb_node
);
375 first
= rb_entry(n
, struct vmap_area
, rb_node
);
380 while (addr
+ size
> first
->va_start
&& addr
+ size
<= vend
) {
381 addr
= ALIGN(first
->va_end
+ PAGE_SIZE
, align
);
382 if (addr
+ size
- 1 < addr
)
385 n
= rb_next(&first
->rb_node
);
387 first
= rb_entry(n
, struct vmap_area
, rb_node
);
393 if (addr
+ size
> vend
) {
395 spin_unlock(&vmap_area_lock
);
397 purge_vmap_area_lazy();
401 if (printk_ratelimit())
403 "vmap allocation for size %lu failed: "
404 "use vmalloc=<size> to increase size.\n", size
);
406 return ERR_PTR(-EBUSY
);
409 BUG_ON(addr
& (align
-1));
412 va
->va_end
= addr
+ size
;
414 __insert_vmap_area(va
);
415 spin_unlock(&vmap_area_lock
);
420 static void rcu_free_va(struct rcu_head
*head
)
422 struct vmap_area
*va
= container_of(head
, struct vmap_area
, rcu_head
);
427 static void __free_vmap_area(struct vmap_area
*va
)
429 BUG_ON(RB_EMPTY_NODE(&va
->rb_node
));
430 rb_erase(&va
->rb_node
, &vmap_area_root
);
431 RB_CLEAR_NODE(&va
->rb_node
);
432 list_del_rcu(&va
->list
);
435 * Track the highest possible candidate for pcpu area
436 * allocation. Areas outside of vmalloc area can be returned
437 * here too, consider only end addresses which fall inside
438 * vmalloc area proper.
440 if (va
->va_end
> VMALLOC_START
&& va
->va_end
<= VMALLOC_END
)
441 vmap_area_pcpu_hole
= max(vmap_area_pcpu_hole
, va
->va_end
);
443 call_rcu(&va
->rcu_head
, rcu_free_va
);
447 * Free a region of KVA allocated by alloc_vmap_area
449 static void free_vmap_area(struct vmap_area
*va
)
451 spin_lock(&vmap_area_lock
);
452 __free_vmap_area(va
);
453 spin_unlock(&vmap_area_lock
);
457 * Clear the pagetable entries of a given vmap_area
459 static void unmap_vmap_area(struct vmap_area
*va
)
461 vunmap_page_range(va
->va_start
, va
->va_end
);
464 static void vmap_debug_free_range(unsigned long start
, unsigned long end
)
467 * Unmap page tables and force a TLB flush immediately if
468 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
469 * bugs similarly to those in linear kernel virtual address
470 * space after a page has been freed.
472 * All the lazy freeing logic is still retained, in order to
473 * minimise intrusiveness of this debugging feature.
475 * This is going to be *slow* (linear kernel virtual address
476 * debugging doesn't do a broadcast TLB flush so it is a lot
479 #ifdef CONFIG_DEBUG_PAGEALLOC
480 vunmap_page_range(start
, end
);
481 flush_tlb_kernel_range(start
, end
);
486 * lazy_max_pages is the maximum amount of virtual address space we gather up
487 * before attempting to purge with a TLB flush.
489 * There is a tradeoff here: a larger number will cover more kernel page tables
490 * and take slightly longer to purge, but it will linearly reduce the number of
491 * global TLB flushes that must be performed. It would seem natural to scale
492 * this number up linearly with the number of CPUs (because vmapping activity
493 * could also scale linearly with the number of CPUs), however it is likely
494 * that in practice, workloads might be constrained in other ways that mean
495 * vmap activity will not scale linearly with CPUs. Also, I want to be
496 * conservative and not introduce a big latency on huge systems, so go with
497 * a less aggressive log scale. It will still be an improvement over the old
498 * code, and it will be simple to change the scale factor if we find that it
499 * becomes a problem on bigger systems.
501 static unsigned long lazy_max_pages(void)
505 log
= fls(num_online_cpus());
507 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
510 static atomic_t vmap_lazy_nr
= ATOMIC_INIT(0);
512 /* for per-CPU blocks */
513 static void purge_fragmented_blocks_allcpus(void);
516 * called before a call to iounmap() if the caller wants vm_area_struct's
519 void set_iounmap_nonlazy(void)
521 atomic_set(&vmap_lazy_nr
, lazy_max_pages()+1);
525 * Purges all lazily-freed vmap areas.
527 * If sync is 0 then don't purge if there is already a purge in progress.
528 * If force_flush is 1, then flush kernel TLBs between *start and *end even
529 * if we found no lazy vmap areas to unmap (callers can use this to optimise
530 * their own TLB flushing).
531 * Returns with *start = min(*start, lowest purged address)
532 * *end = max(*end, highest purged address)
534 static void __purge_vmap_area_lazy(unsigned long *start
, unsigned long *end
,
535 int sync
, int force_flush
)
537 static DEFINE_SPINLOCK(purge_lock
);
539 struct vmap_area
*va
;
540 struct vmap_area
*n_va
;
544 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
545 * should not expect such behaviour. This just simplifies locking for
546 * the case that isn't actually used at the moment anyway.
548 if (!sync
&& !force_flush
) {
549 if (!spin_trylock(&purge_lock
))
552 spin_lock(&purge_lock
);
555 purge_fragmented_blocks_allcpus();
558 list_for_each_entry_rcu(va
, &vmap_area_list
, list
) {
559 if (va
->flags
& VM_LAZY_FREE
) {
560 if (va
->va_start
< *start
)
561 *start
= va
->va_start
;
562 if (va
->va_end
> *end
)
564 nr
+= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
566 list_add_tail(&va
->purge_list
, &valist
);
567 va
->flags
|= VM_LAZY_FREEING
;
568 va
->flags
&= ~VM_LAZY_FREE
;
574 atomic_sub(nr
, &vmap_lazy_nr
);
576 if (nr
|| force_flush
)
577 flush_tlb_kernel_range(*start
, *end
);
580 spin_lock(&vmap_area_lock
);
581 list_for_each_entry_safe(va
, n_va
, &valist
, purge_list
)
582 __free_vmap_area(va
);
583 spin_unlock(&vmap_area_lock
);
585 spin_unlock(&purge_lock
);
589 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
590 * is already purging.
592 static void try_purge_vmap_area_lazy(void)
594 unsigned long start
= ULONG_MAX
, end
= 0;
596 __purge_vmap_area_lazy(&start
, &end
, 0, 0);
600 * Kick off a purge of the outstanding lazy areas.
602 static void purge_vmap_area_lazy(void)
604 unsigned long start
= ULONG_MAX
, end
= 0;
606 __purge_vmap_area_lazy(&start
, &end
, 1, 0);
610 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
611 * called for the correct range previously.
613 static void free_unmap_vmap_area_noflush(struct vmap_area
*va
)
615 va
->flags
|= VM_LAZY_FREE
;
616 atomic_add((va
->va_end
- va
->va_start
) >> PAGE_SHIFT
, &vmap_lazy_nr
);
617 if (unlikely(atomic_read(&vmap_lazy_nr
) > lazy_max_pages()))
618 try_purge_vmap_area_lazy();
622 * Free and unmap a vmap area
624 static void free_unmap_vmap_area(struct vmap_area
*va
)
626 flush_cache_vunmap(va
->va_start
, va
->va_end
);
627 free_unmap_vmap_area_noflush(va
);
630 static struct vmap_area
*find_vmap_area(unsigned long addr
)
632 struct vmap_area
*va
;
634 spin_lock(&vmap_area_lock
);
635 va
= __find_vmap_area(addr
);
636 spin_unlock(&vmap_area_lock
);
641 static void free_unmap_vmap_area_addr(unsigned long addr
)
643 struct vmap_area
*va
;
645 va
= find_vmap_area(addr
);
647 free_unmap_vmap_area(va
);
651 /*** Per cpu kva allocator ***/
654 * vmap space is limited especially on 32 bit architectures. Ensure there is
655 * room for at least 16 percpu vmap blocks per CPU.
658 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
659 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
660 * instead (we just need a rough idea)
662 #if BITS_PER_LONG == 32
663 #define VMALLOC_SPACE (128UL*1024*1024)
665 #define VMALLOC_SPACE (128UL*1024*1024*1024)
668 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
669 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
670 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
671 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
672 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
673 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
674 #define VMAP_BBMAP_BITS \
675 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
676 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
677 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
679 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
681 static bool vmap_initialized __read_mostly
= false;
683 struct vmap_block_queue
{
685 struct list_head free
;
690 struct vmap_area
*va
;
691 struct vmap_block_queue
*vbq
;
692 unsigned long free
, dirty
;
693 DECLARE_BITMAP(alloc_map
, VMAP_BBMAP_BITS
);
694 DECLARE_BITMAP(dirty_map
, VMAP_BBMAP_BITS
);
695 struct list_head free_list
;
696 struct rcu_head rcu_head
;
697 struct list_head purge
;
700 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
701 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
704 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
705 * in the free path. Could get rid of this if we change the API to return a
706 * "cookie" from alloc, to be passed to free. But no big deal yet.
708 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
709 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
712 * We should probably have a fallback mechanism to allocate virtual memory
713 * out of partially filled vmap blocks. However vmap block sizing should be
714 * fairly reasonable according to the vmalloc size, so it shouldn't be a
718 static unsigned long addr_to_vb_idx(unsigned long addr
)
720 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
721 addr
/= VMAP_BLOCK_SIZE
;
725 static struct vmap_block
*new_vmap_block(gfp_t gfp_mask
)
727 struct vmap_block_queue
*vbq
;
728 struct vmap_block
*vb
;
729 struct vmap_area
*va
;
730 unsigned long vb_idx
;
733 node
= numa_node_id();
735 vb
= kmalloc_node(sizeof(struct vmap_block
),
736 gfp_mask
& GFP_RECLAIM_MASK
, node
);
738 return ERR_PTR(-ENOMEM
);
740 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
741 VMALLOC_START
, VMALLOC_END
,
743 if (unlikely(IS_ERR(va
))) {
745 return ERR_PTR(PTR_ERR(va
));
748 err
= radix_tree_preload(gfp_mask
);
755 spin_lock_init(&vb
->lock
);
757 vb
->free
= VMAP_BBMAP_BITS
;
759 bitmap_zero(vb
->alloc_map
, VMAP_BBMAP_BITS
);
760 bitmap_zero(vb
->dirty_map
, VMAP_BBMAP_BITS
);
761 INIT_LIST_HEAD(&vb
->free_list
);
763 vb_idx
= addr_to_vb_idx(va
->va_start
);
764 spin_lock(&vmap_block_tree_lock
);
765 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
766 spin_unlock(&vmap_block_tree_lock
);
768 radix_tree_preload_end();
770 vbq
= &get_cpu_var(vmap_block_queue
);
772 spin_lock(&vbq
->lock
);
773 list_add_rcu(&vb
->free_list
, &vbq
->free
);
774 spin_unlock(&vbq
->lock
);
775 put_cpu_var(vmap_block_queue
);
780 static void rcu_free_vb(struct rcu_head
*head
)
782 struct vmap_block
*vb
= container_of(head
, struct vmap_block
, rcu_head
);
787 static void free_vmap_block(struct vmap_block
*vb
)
789 struct vmap_block
*tmp
;
790 unsigned long vb_idx
;
792 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
793 spin_lock(&vmap_block_tree_lock
);
794 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
795 spin_unlock(&vmap_block_tree_lock
);
798 free_unmap_vmap_area_noflush(vb
->va
);
799 call_rcu(&vb
->rcu_head
, rcu_free_vb
);
802 static void purge_fragmented_blocks(int cpu
)
805 struct vmap_block
*vb
;
806 struct vmap_block
*n_vb
;
807 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
810 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
812 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
815 spin_lock(&vb
->lock
);
816 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
817 vb
->free
= 0; /* prevent further allocs after releasing lock */
818 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
819 bitmap_fill(vb
->alloc_map
, VMAP_BBMAP_BITS
);
820 bitmap_fill(vb
->dirty_map
, VMAP_BBMAP_BITS
);
821 spin_lock(&vbq
->lock
);
822 list_del_rcu(&vb
->free_list
);
823 spin_unlock(&vbq
->lock
);
824 spin_unlock(&vb
->lock
);
825 list_add_tail(&vb
->purge
, &purge
);
827 spin_unlock(&vb
->lock
);
831 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
832 list_del(&vb
->purge
);
837 static void purge_fragmented_blocks_thiscpu(void)
839 purge_fragmented_blocks(smp_processor_id());
842 static void purge_fragmented_blocks_allcpus(void)
846 for_each_possible_cpu(cpu
)
847 purge_fragmented_blocks(cpu
);
850 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
852 struct vmap_block_queue
*vbq
;
853 struct vmap_block
*vb
;
854 unsigned long addr
= 0;
858 BUG_ON(size
& ~PAGE_MASK
);
859 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
860 order
= get_order(size
);
864 vbq
= &get_cpu_var(vmap_block_queue
);
865 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
868 spin_lock(&vb
->lock
);
869 if (vb
->free
< 1UL << order
)
872 i
= bitmap_find_free_region(vb
->alloc_map
,
873 VMAP_BBMAP_BITS
, order
);
876 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
) {
877 /* fragmented and no outstanding allocations */
878 BUG_ON(vb
->dirty
!= VMAP_BBMAP_BITS
);
883 addr
= vb
->va
->va_start
+ (i
<< PAGE_SHIFT
);
884 BUG_ON(addr_to_vb_idx(addr
) !=
885 addr_to_vb_idx(vb
->va
->va_start
));
886 vb
->free
-= 1UL << order
;
888 spin_lock(&vbq
->lock
);
889 list_del_rcu(&vb
->free_list
);
890 spin_unlock(&vbq
->lock
);
892 spin_unlock(&vb
->lock
);
895 spin_unlock(&vb
->lock
);
899 purge_fragmented_blocks_thiscpu();
901 put_cpu_var(vmap_block_queue
);
905 vb
= new_vmap_block(gfp_mask
);
914 static void vb_free(const void *addr
, unsigned long size
)
916 unsigned long offset
;
917 unsigned long vb_idx
;
919 struct vmap_block
*vb
;
921 BUG_ON(size
& ~PAGE_MASK
);
922 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
924 flush_cache_vunmap((unsigned long)addr
, (unsigned long)addr
+ size
);
926 order
= get_order(size
);
928 offset
= (unsigned long)addr
& (VMAP_BLOCK_SIZE
- 1);
930 vb_idx
= addr_to_vb_idx((unsigned long)addr
);
932 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
936 spin_lock(&vb
->lock
);
937 BUG_ON(bitmap_allocate_region(vb
->dirty_map
, offset
>> PAGE_SHIFT
, order
));
939 vb
->dirty
+= 1UL << order
;
940 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
942 spin_unlock(&vb
->lock
);
945 spin_unlock(&vb
->lock
);
949 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
951 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
952 * to amortize TLB flushing overheads. What this means is that any page you
953 * have now, may, in a former life, have been mapped into kernel virtual
954 * address by the vmap layer and so there might be some CPUs with TLB entries
955 * still referencing that page (additional to the regular 1:1 kernel mapping).
957 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
958 * be sure that none of the pages we have control over will have any aliases
959 * from the vmap layer.
961 void vm_unmap_aliases(void)
963 unsigned long start
= ULONG_MAX
, end
= 0;
967 if (unlikely(!vmap_initialized
))
970 for_each_possible_cpu(cpu
) {
971 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
972 struct vmap_block
*vb
;
975 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
978 spin_lock(&vb
->lock
);
979 i
= find_first_bit(vb
->dirty_map
, VMAP_BBMAP_BITS
);
980 while (i
< VMAP_BBMAP_BITS
) {
983 j
= find_next_zero_bit(vb
->dirty_map
,
986 s
= vb
->va
->va_start
+ (i
<< PAGE_SHIFT
);
987 e
= vb
->va
->va_start
+ (j
<< PAGE_SHIFT
);
988 vunmap_page_range(s
, e
);
997 i
= find_next_bit(vb
->dirty_map
,
1000 spin_unlock(&vb
->lock
);
1005 __purge_vmap_area_lazy(&start
, &end
, 1, flush
);
1007 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1010 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1011 * @mem: the pointer returned by vm_map_ram
1012 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1014 void vm_unmap_ram(const void *mem
, unsigned int count
)
1016 unsigned long size
= count
<< PAGE_SHIFT
;
1017 unsigned long addr
= (unsigned long)mem
;
1020 BUG_ON(addr
< VMALLOC_START
);
1021 BUG_ON(addr
> VMALLOC_END
);
1022 BUG_ON(addr
& (PAGE_SIZE
-1));
1024 debug_check_no_locks_freed(mem
, size
);
1025 vmap_debug_free_range(addr
, addr
+size
);
1027 if (likely(count
<= VMAP_MAX_ALLOC
))
1030 free_unmap_vmap_area_addr(addr
);
1032 EXPORT_SYMBOL(vm_unmap_ram
);
1035 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1036 * @pages: an array of pointers to the pages to be mapped
1037 * @count: number of pages
1038 * @node: prefer to allocate data structures on this node
1039 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1041 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1043 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
, pgprot_t prot
)
1045 unsigned long size
= count
<< PAGE_SHIFT
;
1049 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1050 mem
= vb_alloc(size
, GFP_KERNEL
);
1053 addr
= (unsigned long)mem
;
1055 struct vmap_area
*va
;
1056 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1057 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1061 addr
= va
->va_start
;
1064 if (vmap_page_range(addr
, addr
+ size
, prot
, pages
) < 0) {
1065 vm_unmap_ram(mem
, count
);
1070 EXPORT_SYMBOL(vm_map_ram
);
1073 * vm_area_register_early - register vmap area early during boot
1074 * @vm: vm_struct to register
1075 * @align: requested alignment
1077 * This function is used to register kernel vm area before
1078 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1079 * proper values on entry and other fields should be zero. On return,
1080 * vm->addr contains the allocated address.
1082 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1084 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1086 static size_t vm_init_off __initdata
;
1089 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1090 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1092 vm
->addr
= (void *)addr
;
1098 void __init
vmalloc_init(void)
1100 struct vmap_area
*va
;
1101 struct vm_struct
*tmp
;
1104 for_each_possible_cpu(i
) {
1105 struct vmap_block_queue
*vbq
;
1107 vbq
= &per_cpu(vmap_block_queue
, i
);
1108 spin_lock_init(&vbq
->lock
);
1109 INIT_LIST_HEAD(&vbq
->free
);
1112 /* Import existing vmlist entries. */
1113 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1114 va
= kzalloc(sizeof(struct vmap_area
), GFP_NOWAIT
);
1115 va
->flags
= tmp
->flags
| VM_VM_AREA
;
1116 va
->va_start
= (unsigned long)tmp
->addr
;
1117 va
->va_end
= va
->va_start
+ tmp
->size
;
1118 __insert_vmap_area(va
);
1121 vmap_area_pcpu_hole
= VMALLOC_END
;
1123 vmap_initialized
= true;
1127 * map_kernel_range_noflush - map kernel VM area with the specified pages
1128 * @addr: start of the VM area to map
1129 * @size: size of the VM area to map
1130 * @prot: page protection flags to use
1131 * @pages: pages to map
1133 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1134 * specify should have been allocated using get_vm_area() and its
1138 * This function does NOT do any cache flushing. The caller is
1139 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1140 * before calling this function.
1143 * The number of pages mapped on success, -errno on failure.
1145 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
1146 pgprot_t prot
, struct page
**pages
)
1148 return vmap_page_range_noflush(addr
, addr
+ size
, prot
, pages
);
1152 * unmap_kernel_range_noflush - unmap kernel VM area
1153 * @addr: start of the VM area to unmap
1154 * @size: size of the VM area to unmap
1156 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1157 * specify should have been allocated using get_vm_area() and its
1161 * This function does NOT do any cache flushing. The caller is
1162 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1163 * before calling this function and flush_tlb_kernel_range() after.
1165 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
1167 vunmap_page_range(addr
, addr
+ size
);
1171 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1172 * @addr: start of the VM area to unmap
1173 * @size: size of the VM area to unmap
1175 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1176 * the unmapping and tlb after.
1178 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
1180 unsigned long end
= addr
+ size
;
1182 flush_cache_vunmap(addr
, end
);
1183 vunmap_page_range(addr
, end
);
1184 flush_tlb_kernel_range(addr
, end
);
1187 int map_vm_area(struct vm_struct
*area
, pgprot_t prot
, struct page
***pages
)
1189 unsigned long addr
= (unsigned long)area
->addr
;
1190 unsigned long end
= addr
+ area
->size
- PAGE_SIZE
;
1193 err
= vmap_page_range(addr
, end
, prot
, *pages
);
1201 EXPORT_SYMBOL_GPL(map_vm_area
);
1203 /*** Old vmalloc interfaces ***/
1204 DEFINE_RWLOCK(vmlist_lock
);
1205 struct vm_struct
*vmlist
;
1207 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
1208 unsigned long flags
, void *caller
)
1211 vm
->addr
= (void *)va
->va_start
;
1212 vm
->size
= va
->va_end
- va
->va_start
;
1213 vm
->caller
= caller
;
1215 va
->flags
|= VM_VM_AREA
;
1218 static void insert_vmalloc_vmlist(struct vm_struct
*vm
)
1220 struct vm_struct
*tmp
, **p
;
1222 vm
->flags
&= ~VM_UNLIST
;
1223 write_lock(&vmlist_lock
);
1224 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1225 if (tmp
->addr
>= vm
->addr
)
1230 write_unlock(&vmlist_lock
);
1233 static void insert_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
1234 unsigned long flags
, void *caller
)
1236 setup_vmalloc_vm(vm
, va
, flags
, caller
);
1237 insert_vmalloc_vmlist(vm
);
1240 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
1241 unsigned long align
, unsigned long flags
, unsigned long start
,
1242 unsigned long end
, int node
, gfp_t gfp_mask
, void *caller
)
1244 static struct vmap_area
*va
;
1245 struct vm_struct
*area
;
1247 BUG_ON(in_interrupt());
1248 if (flags
& VM_IOREMAP
) {
1249 int bit
= fls(size
);
1251 if (bit
> IOREMAP_MAX_ORDER
)
1252 bit
= IOREMAP_MAX_ORDER
;
1253 else if (bit
< PAGE_SHIFT
)
1259 size
= PAGE_ALIGN(size
);
1260 if (unlikely(!size
))
1263 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
1264 if (unlikely(!area
))
1268 * We always allocate a guard page.
1272 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
1279 * When this function is called from __vmalloc_node_range,
1280 * we do not add vm_struct to vmlist here to avoid
1281 * accessing uninitialized members of vm_struct such as
1282 * pages and nr_pages fields. They will be set later.
1283 * To distinguish it from others, we use a VM_UNLIST flag.
1285 if (flags
& VM_UNLIST
)
1286 setup_vmalloc_vm(area
, va
, flags
, caller
);
1288 insert_vmalloc_vm(area
, va
, flags
, caller
);
1293 struct vm_struct
*__get_vm_area(unsigned long size
, unsigned long flags
,
1294 unsigned long start
, unsigned long end
)
1296 return __get_vm_area_node(size
, 1, flags
, start
, end
, -1, GFP_KERNEL
,
1297 __builtin_return_address(0));
1299 EXPORT_SYMBOL_GPL(__get_vm_area
);
1301 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
1302 unsigned long start
, unsigned long end
,
1305 return __get_vm_area_node(size
, 1, flags
, start
, end
, -1, GFP_KERNEL
,
1310 * get_vm_area - reserve a contiguous kernel virtual area
1311 * @size: size of the area
1312 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1314 * Search an area of @size in the kernel virtual mapping area,
1315 * and reserved it for out purposes. Returns the area descriptor
1316 * on success or %NULL on failure.
1318 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
1320 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1321 -1, GFP_KERNEL
, __builtin_return_address(0));
1324 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
1327 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1328 -1, GFP_KERNEL
, caller
);
1331 struct vm_struct
*get_vm_area_node(unsigned long size
, unsigned long flags
,
1332 int node
, gfp_t gfp_mask
)
1334 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1335 node
, gfp_mask
, __builtin_return_address(0));
1338 static struct vm_struct
*find_vm_area(const void *addr
)
1340 struct vmap_area
*va
;
1342 va
= find_vmap_area((unsigned long)addr
);
1343 if (va
&& va
->flags
& VM_VM_AREA
)
1350 * remove_vm_area - find and remove a continuous kernel virtual area
1351 * @addr: base address
1353 * Search for the kernel VM area starting at @addr, and remove it.
1354 * This function returns the found VM area, but using it is NOT safe
1355 * on SMP machines, except for its size or flags.
1357 struct vm_struct
*remove_vm_area(const void *addr
)
1359 struct vmap_area
*va
;
1361 va
= find_vmap_area((unsigned long)addr
);
1362 if (va
&& va
->flags
& VM_VM_AREA
) {
1363 struct vm_struct
*vm
= va
->private;
1365 if (!(vm
->flags
& VM_UNLIST
)) {
1366 struct vm_struct
*tmp
, **p
;
1368 * remove from list and disallow access to
1369 * this vm_struct before unmap. (address range
1370 * confliction is maintained by vmap.)
1372 write_lock(&vmlist_lock
);
1373 for (p
= &vmlist
; (tmp
= *p
) != vm
; p
= &tmp
->next
)
1376 write_unlock(&vmlist_lock
);
1379 vmap_debug_free_range(va
->va_start
, va
->va_end
);
1380 free_unmap_vmap_area(va
);
1381 vm
->size
-= PAGE_SIZE
;
1388 static void __vunmap(const void *addr
, int deallocate_pages
)
1390 struct vm_struct
*area
;
1395 if ((PAGE_SIZE
-1) & (unsigned long)addr
) {
1396 WARN(1, KERN_ERR
"Trying to vfree() bad address (%p)\n", addr
);
1400 area
= remove_vm_area(addr
);
1401 if (unlikely(!area
)) {
1402 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
1407 debug_check_no_locks_freed(addr
, area
->size
);
1408 debug_check_no_obj_freed(addr
, area
->size
);
1410 if (deallocate_pages
) {
1413 for (i
= 0; i
< area
->nr_pages
; i
++) {
1414 struct page
*page
= area
->pages
[i
];
1420 if (area
->flags
& VM_VPAGES
)
1431 * vfree - release memory allocated by vmalloc()
1432 * @addr: memory base address
1434 * Free the virtually continuous memory area starting at @addr, as
1435 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1436 * NULL, no operation is performed.
1438 * Must not be called in interrupt context.
1440 void vfree(const void *addr
)
1442 BUG_ON(in_interrupt());
1444 kmemleak_free(addr
);
1448 EXPORT_SYMBOL(vfree
);
1451 * vunmap - release virtual mapping obtained by vmap()
1452 * @addr: memory base address
1454 * Free the virtually contiguous memory area starting at @addr,
1455 * which was created from the page array passed to vmap().
1457 * Must not be called in interrupt context.
1459 void vunmap(const void *addr
)
1461 BUG_ON(in_interrupt());
1465 EXPORT_SYMBOL(vunmap
);
1468 * vmap - map an array of pages into virtually contiguous space
1469 * @pages: array of page pointers
1470 * @count: number of pages to map
1471 * @flags: vm_area->flags
1472 * @prot: page protection for the mapping
1474 * Maps @count pages from @pages into contiguous kernel virtual
1477 void *vmap(struct page
**pages
, unsigned int count
,
1478 unsigned long flags
, pgprot_t prot
)
1480 struct vm_struct
*area
;
1484 if (count
> totalram_pages
)
1487 area
= get_vm_area_caller((count
<< PAGE_SHIFT
), flags
,
1488 __builtin_return_address(0));
1492 if (map_vm_area(area
, prot
, &pages
)) {
1499 EXPORT_SYMBOL(vmap
);
1501 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1502 gfp_t gfp_mask
, pgprot_t prot
,
1503 int node
, void *caller
);
1504 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
1505 pgprot_t prot
, int node
, void *caller
)
1507 struct page
**pages
;
1508 unsigned int nr_pages
, array_size
, i
;
1509 gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
1511 nr_pages
= (area
->size
- PAGE_SIZE
) >> PAGE_SHIFT
;
1512 array_size
= (nr_pages
* sizeof(struct page
*));
1514 area
->nr_pages
= nr_pages
;
1515 /* Please note that the recursion is strictly bounded. */
1516 if (array_size
> PAGE_SIZE
) {
1517 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|__GFP_HIGHMEM
,
1518 PAGE_KERNEL
, node
, caller
);
1519 area
->flags
|= VM_VPAGES
;
1521 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
1523 area
->pages
= pages
;
1524 area
->caller
= caller
;
1526 remove_vm_area(area
->addr
);
1531 for (i
= 0; i
< area
->nr_pages
; i
++) {
1535 page
= alloc_page(gfp_mask
);
1537 page
= alloc_pages_node(node
, gfp_mask
, 0);
1539 if (unlikely(!page
)) {
1540 /* Successfully allocated i pages, free them in __vunmap() */
1544 area
->pages
[i
] = page
;
1547 if (map_vm_area(area
, prot
, &pages
))
1556 void *__vmalloc_area(struct vm_struct
*area
, gfp_t gfp_mask
, pgprot_t prot
)
1558 void *addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, -1,
1559 __builtin_return_address(0));
1562 * A ref_count = 3 is needed because the vm_struct and vmap_area
1563 * structures allocated in the __get_vm_area_node() function contain
1564 * references to the virtual address of the vmalloc'ed block.
1566 kmemleak_alloc(addr
, area
->size
- PAGE_SIZE
, 3, gfp_mask
);
1572 * __vmalloc_node - allocate virtually contiguous memory
1573 * @size: allocation size
1574 * @align: desired alignment
1575 * @gfp_mask: flags for the page level allocator
1576 * @prot: protection mask for the allocated pages
1577 * @node: node to use for allocation or -1
1578 * @caller: caller's return address
1580 * Allocate enough pages to cover @size from the page level
1581 * allocator with @gfp_mask flags. Map them into contiguous
1582 * kernel virtual space, using a pagetable protection of @prot.
1584 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1585 gfp_t gfp_mask
, pgprot_t prot
,
1586 int node
, void *caller
)
1588 struct vm_struct
*area
;
1590 unsigned long real_size
= size
;
1592 size
= PAGE_ALIGN(size
);
1593 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages
)
1596 area
= __get_vm_area_node(size
, align
, VM_ALLOC
| VM_UNLIST
,
1597 VMALLOC_START
, VMALLOC_END
, node
,
1603 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
, caller
);
1608 * In this function, newly allocated vm_struct is not added
1609 * to vmlist at __get_vm_area_node(). so, it is added here.
1611 insert_vmalloc_vmlist(area
);
1614 * A ref_count = 3 is needed because the vm_struct and vmap_area
1615 * structures allocated in the __get_vm_area_node() function contain
1616 * references to the virtual address of the vmalloc'ed block.
1618 kmemleak_alloc(addr
, real_size
, 3, gfp_mask
);
1623 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
, pgprot_t prot
)
1625 return __vmalloc_node(size
, 1, gfp_mask
, prot
, -1,
1626 __builtin_return_address(0));
1628 EXPORT_SYMBOL(__vmalloc
);
1631 * vmalloc - allocate virtually contiguous memory
1632 * @size: allocation size
1633 * Allocate enough pages to cover @size from the page level
1634 * allocator and map them into contiguous kernel virtual space.
1636 * For tight control over page level allocator and protection flags
1637 * use __vmalloc() instead.
1639 void *vmalloc(unsigned long size
)
1641 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_HIGHMEM
, PAGE_KERNEL
,
1642 -1, __builtin_return_address(0));
1644 EXPORT_SYMBOL(vmalloc
);
1647 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1648 * @size: allocation size
1650 * The resulting memory area is zeroed so it can be mapped to userspace
1651 * without leaking data.
1653 void *vmalloc_user(unsigned long size
)
1655 struct vm_struct
*area
;
1658 ret
= __vmalloc_node(size
, SHMLBA
,
1659 GFP_KERNEL
| __GFP_HIGHMEM
| __GFP_ZERO
,
1660 PAGE_KERNEL
, -1, __builtin_return_address(0));
1662 area
= find_vm_area(ret
);
1663 area
->flags
|= VM_USERMAP
;
1667 EXPORT_SYMBOL(vmalloc_user
);
1670 * vmalloc_node - allocate memory on a specific node
1671 * @size: allocation size
1674 * Allocate enough pages to cover @size from the page level
1675 * allocator and map them into contiguous kernel virtual space.
1677 * For tight control over page level allocator and protection flags
1678 * use __vmalloc() instead.
1680 void *vmalloc_node(unsigned long size
, int node
)
1682 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_HIGHMEM
, PAGE_KERNEL
,
1683 node
, __builtin_return_address(0));
1685 EXPORT_SYMBOL(vmalloc_node
);
1687 #ifndef PAGE_KERNEL_EXEC
1688 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1692 * vmalloc_exec - allocate virtually contiguous, executable memory
1693 * @size: allocation size
1695 * Kernel-internal function to allocate enough pages to cover @size
1696 * the page level allocator and map them into contiguous and
1697 * executable kernel virtual space.
1699 * For tight control over page level allocator and protection flags
1700 * use __vmalloc() instead.
1703 void *vmalloc_exec(unsigned long size
)
1705 return __vmalloc_node(size
, 1, GFP_KERNEL
| __GFP_HIGHMEM
, PAGE_KERNEL_EXEC
,
1706 -1, __builtin_return_address(0));
1709 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1710 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1711 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1712 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1714 #define GFP_VMALLOC32 GFP_KERNEL
1718 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1719 * @size: allocation size
1721 * Allocate enough 32bit PA addressable pages to cover @size from the
1722 * page level allocator and map them into contiguous kernel virtual space.
1724 void *vmalloc_32(unsigned long size
)
1726 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, PAGE_KERNEL
,
1727 -1, __builtin_return_address(0));
1729 EXPORT_SYMBOL(vmalloc_32
);
1732 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1733 * @size: allocation size
1735 * The resulting memory area is 32bit addressable and zeroed so it can be
1736 * mapped to userspace without leaking data.
1738 void *vmalloc_32_user(unsigned long size
)
1740 struct vm_struct
*area
;
1743 ret
= __vmalloc_node(size
, 1, GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
1744 -1, __builtin_return_address(0));
1746 area
= find_vm_area(ret
);
1747 area
->flags
|= VM_USERMAP
;
1751 EXPORT_SYMBOL(vmalloc_32_user
);
1754 * small helper routine , copy contents to buf from addr.
1755 * If the page is not present, fill zero.
1758 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
1764 unsigned long offset
, length
;
1766 offset
= (unsigned long)addr
& ~PAGE_MASK
;
1767 length
= PAGE_SIZE
- offset
;
1770 p
= vmalloc_to_page(addr
);
1772 * To do safe access to this _mapped_ area, we need
1773 * lock. But adding lock here means that we need to add
1774 * overhead of vmalloc()/vfree() calles for this _debug_
1775 * interface, rarely used. Instead of that, we'll use
1776 * kmap() and get small overhead in this access function.
1780 * we can expect USER0 is not used (see vread/vwrite's
1781 * function description)
1783 void *map
= kmap_atomic(p
, KM_USER0
);
1784 memcpy(buf
, map
+ offset
, length
);
1785 kunmap_atomic(map
, KM_USER0
);
1787 memset(buf
, 0, length
);
1797 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
1803 unsigned long offset
, length
;
1805 offset
= (unsigned long)addr
& ~PAGE_MASK
;
1806 length
= PAGE_SIZE
- offset
;
1809 p
= vmalloc_to_page(addr
);
1811 * To do safe access to this _mapped_ area, we need
1812 * lock. But adding lock here means that we need to add
1813 * overhead of vmalloc()/vfree() calles for this _debug_
1814 * interface, rarely used. Instead of that, we'll use
1815 * kmap() and get small overhead in this access function.
1819 * we can expect USER0 is not used (see vread/vwrite's
1820 * function description)
1822 void *map
= kmap_atomic(p
, KM_USER0
);
1823 memcpy(map
+ offset
, buf
, length
);
1824 kunmap_atomic(map
, KM_USER0
);
1835 * vread() - read vmalloc area in a safe way.
1836 * @buf: buffer for reading data
1837 * @addr: vm address.
1838 * @count: number of bytes to be read.
1840 * Returns # of bytes which addr and buf should be increased.
1841 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1842 * includes any intersect with alive vmalloc area.
1844 * This function checks that addr is a valid vmalloc'ed area, and
1845 * copy data from that area to a given buffer. If the given memory range
1846 * of [addr...addr+count) includes some valid address, data is copied to
1847 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1848 * IOREMAP area is treated as memory hole and no copy is done.
1850 * If [addr...addr+count) doesn't includes any intersects with alive
1851 * vm_struct area, returns 0.
1852 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1853 * the caller should guarantee KM_USER0 is not used.
1855 * Note: In usual ops, vread() is never necessary because the caller
1856 * should know vmalloc() area is valid and can use memcpy().
1857 * This is for routines which have to access vmalloc area without
1858 * any informaion, as /dev/kmem.
1862 long vread(char *buf
, char *addr
, unsigned long count
)
1864 struct vm_struct
*tmp
;
1865 char *vaddr
, *buf_start
= buf
;
1866 unsigned long buflen
= count
;
1869 /* Don't allow overflow */
1870 if ((unsigned long) addr
+ count
< count
)
1871 count
= -(unsigned long) addr
;
1873 read_lock(&vmlist_lock
);
1874 for (tmp
= vmlist
; count
&& tmp
; tmp
= tmp
->next
) {
1875 vaddr
= (char *) tmp
->addr
;
1876 if (addr
>= vaddr
+ tmp
->size
- PAGE_SIZE
)
1878 while (addr
< vaddr
) {
1886 n
= vaddr
+ tmp
->size
- PAGE_SIZE
- addr
;
1889 if (!(tmp
->flags
& VM_IOREMAP
))
1890 aligned_vread(buf
, addr
, n
);
1891 else /* IOREMAP area is treated as memory hole */
1898 read_unlock(&vmlist_lock
);
1900 if (buf
== buf_start
)
1902 /* zero-fill memory holes */
1903 if (buf
!= buf_start
+ buflen
)
1904 memset(buf
, 0, buflen
- (buf
- buf_start
));
1910 * vwrite() - write vmalloc area in a safe way.
1911 * @buf: buffer for source data
1912 * @addr: vm address.
1913 * @count: number of bytes to be read.
1915 * Returns # of bytes which addr and buf should be incresed.
1916 * (same number to @count).
1917 * If [addr...addr+count) doesn't includes any intersect with valid
1918 * vmalloc area, returns 0.
1920 * This function checks that addr is a valid vmalloc'ed area, and
1921 * copy data from a buffer to the given addr. If specified range of
1922 * [addr...addr+count) includes some valid address, data is copied from
1923 * proper area of @buf. If there are memory holes, no copy to hole.
1924 * IOREMAP area is treated as memory hole and no copy is done.
1926 * If [addr...addr+count) doesn't includes any intersects with alive
1927 * vm_struct area, returns 0.
1928 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1929 * the caller should guarantee KM_USER0 is not used.
1931 * Note: In usual ops, vwrite() is never necessary because the caller
1932 * should know vmalloc() area is valid and can use memcpy().
1933 * This is for routines which have to access vmalloc area without
1934 * any informaion, as /dev/kmem.
1936 * The caller should guarantee KM_USER1 is not used.
1939 long vwrite(char *buf
, char *addr
, unsigned long count
)
1941 struct vm_struct
*tmp
;
1943 unsigned long n
, buflen
;
1946 /* Don't allow overflow */
1947 if ((unsigned long) addr
+ count
< count
)
1948 count
= -(unsigned long) addr
;
1951 read_lock(&vmlist_lock
);
1952 for (tmp
= vmlist
; count
&& tmp
; tmp
= tmp
->next
) {
1953 vaddr
= (char *) tmp
->addr
;
1954 if (addr
>= vaddr
+ tmp
->size
- PAGE_SIZE
)
1956 while (addr
< vaddr
) {
1963 n
= vaddr
+ tmp
->size
- PAGE_SIZE
- addr
;
1966 if (!(tmp
->flags
& VM_IOREMAP
)) {
1967 aligned_vwrite(buf
, addr
, n
);
1975 read_unlock(&vmlist_lock
);
1982 * remap_vmalloc_range - map vmalloc pages to userspace
1983 * @vma: vma to cover (map full range of vma)
1984 * @addr: vmalloc memory
1985 * @pgoff: number of pages into addr before first page to map
1987 * Returns: 0 for success, -Exxx on failure
1989 * This function checks that addr is a valid vmalloc'ed area, and
1990 * that it is big enough to cover the vma. Will return failure if
1991 * that criteria isn't met.
1993 * Similar to remap_pfn_range() (see mm/memory.c)
1995 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
1996 unsigned long pgoff
)
1998 struct vm_struct
*area
;
1999 unsigned long uaddr
= vma
->vm_start
;
2000 unsigned long usize
= vma
->vm_end
- vma
->vm_start
;
2002 if ((PAGE_SIZE
-1) & (unsigned long)addr
)
2005 area
= find_vm_area(addr
);
2009 if (!(area
->flags
& VM_USERMAP
))
2012 if (usize
+ (pgoff
<< PAGE_SHIFT
) > area
->size
- PAGE_SIZE
)
2015 addr
+= pgoff
<< PAGE_SHIFT
;
2017 struct page
*page
= vmalloc_to_page(addr
);
2020 ret
= vm_insert_page(vma
, uaddr
, page
);
2027 } while (usize
> 0);
2029 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2030 vma
->vm_flags
|= VM_RESERVED
;
2034 EXPORT_SYMBOL(remap_vmalloc_range
);
2037 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2040 void __attribute__((weak
)) vmalloc_sync_all(void)
2045 static int f(pte_t
*pte
, pgtable_t table
, unsigned long addr
, void *data
)
2047 /* apply_to_page_range() does all the hard work. */
2052 * alloc_vm_area - allocate a range of kernel address space
2053 * @size: size of the area
2055 * Returns: NULL on failure, vm_struct on success
2057 * This function reserves a range of kernel address space, and
2058 * allocates pagetables to map that range. No actual mappings
2059 * are created. If the kernel address space is not shared
2060 * between processes, it syncs the pagetable across all
2063 struct vm_struct
*alloc_vm_area(size_t size
)
2065 struct vm_struct
*area
;
2067 area
= get_vm_area_caller(size
, VM_IOREMAP
,
2068 __builtin_return_address(0));
2073 * This ensures that page tables are constructed for this region
2074 * of kernel virtual address space and mapped into init_mm.
2076 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2077 area
->size
, f
, NULL
)) {
2082 /* Make sure the pagetables are constructed in process kernel
2088 EXPORT_SYMBOL_GPL(alloc_vm_area
);
2090 void free_vm_area(struct vm_struct
*area
)
2092 struct vm_struct
*ret
;
2093 ret
= remove_vm_area(area
->addr
);
2094 BUG_ON(ret
!= area
);
2097 EXPORT_SYMBOL_GPL(free_vm_area
);
2099 static struct vmap_area
*node_to_va(struct rb_node
*n
)
2101 return n
? rb_entry(n
, struct vmap_area
, rb_node
) : NULL
;
2105 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2106 * @end: target address
2107 * @pnext: out arg for the next vmap_area
2108 * @pprev: out arg for the previous vmap_area
2110 * Returns: %true if either or both of next and prev are found,
2111 * %false if no vmap_area exists
2113 * Find vmap_areas end addresses of which enclose @end. ie. if not
2114 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2116 static bool pvm_find_next_prev(unsigned long end
,
2117 struct vmap_area
**pnext
,
2118 struct vmap_area
**pprev
)
2120 struct rb_node
*n
= vmap_area_root
.rb_node
;
2121 struct vmap_area
*va
= NULL
;
2124 va
= rb_entry(n
, struct vmap_area
, rb_node
);
2125 if (end
< va
->va_end
)
2127 else if (end
> va
->va_end
)
2136 if (va
->va_end
> end
) {
2138 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2141 *pnext
= node_to_va(rb_next(&(*pprev
)->rb_node
));
2147 * pvm_determine_end - find the highest aligned address between two vmap_areas
2148 * @pnext: in/out arg for the next vmap_area
2149 * @pprev: in/out arg for the previous vmap_area
2152 * Returns: determined end address
2154 * Find the highest aligned address between *@pnext and *@pprev below
2155 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2156 * down address is between the end addresses of the two vmap_areas.
2158 * Please note that the address returned by this function may fall
2159 * inside *@pnext vmap_area. The caller is responsible for checking
2162 static unsigned long pvm_determine_end(struct vmap_area
**pnext
,
2163 struct vmap_area
**pprev
,
2164 unsigned long align
)
2166 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2170 addr
= min((*pnext
)->va_start
& ~(align
- 1), vmalloc_end
);
2174 while (*pprev
&& (*pprev
)->va_end
> addr
) {
2176 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2183 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2184 * @offsets: array containing offset of each area
2185 * @sizes: array containing size of each area
2186 * @nr_vms: the number of areas to allocate
2187 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2188 * @gfp_mask: allocation mask
2190 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2191 * vm_structs on success, %NULL on failure
2193 * Percpu allocator wants to use congruent vm areas so that it can
2194 * maintain the offsets among percpu areas. This function allocates
2195 * congruent vmalloc areas for it. These areas tend to be scattered
2196 * pretty far, distance between two areas easily going up to
2197 * gigabytes. To avoid interacting with regular vmallocs, these areas
2198 * are allocated from top.
2200 * Despite its complicated look, this allocator is rather simple. It
2201 * does everything top-down and scans areas from the end looking for
2202 * matching slot. While scanning, if any of the areas overlaps with
2203 * existing vmap_area, the base address is pulled down to fit the
2204 * area. Scanning is repeated till all the areas fit and then all
2205 * necessary data structres are inserted and the result is returned.
2207 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
2208 const size_t *sizes
, int nr_vms
,
2209 size_t align
, gfp_t gfp_mask
)
2211 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
2212 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2213 struct vmap_area
**vas
, *prev
, *next
;
2214 struct vm_struct
**vms
;
2215 int area
, area2
, last_area
, term_area
;
2216 unsigned long base
, start
, end
, last_end
;
2217 bool purged
= false;
2219 gfp_mask
&= GFP_RECLAIM_MASK
;
2221 /* verify parameters and allocate data structures */
2222 BUG_ON(align
& ~PAGE_MASK
|| !is_power_of_2(align
));
2223 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
2224 start
= offsets
[area
];
2225 end
= start
+ sizes
[area
];
2227 /* is everything aligned properly? */
2228 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
2229 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
2231 /* detect the area with the highest address */
2232 if (start
> offsets
[last_area
])
2235 for (area2
= 0; area2
< nr_vms
; area2
++) {
2236 unsigned long start2
= offsets
[area2
];
2237 unsigned long end2
= start2
+ sizes
[area2
];
2242 BUG_ON(start2
>= start
&& start2
< end
);
2243 BUG_ON(end2
<= end
&& end2
> start
);
2246 last_end
= offsets
[last_area
] + sizes
[last_area
];
2248 if (vmalloc_end
- vmalloc_start
< last_end
) {
2253 vms
= kzalloc(sizeof(vms
[0]) * nr_vms
, gfp_mask
);
2254 vas
= kzalloc(sizeof(vas
[0]) * nr_vms
, gfp_mask
);
2258 for (area
= 0; area
< nr_vms
; area
++) {
2259 vas
[area
] = kzalloc(sizeof(struct vmap_area
), gfp_mask
);
2260 vms
[area
] = kzalloc(sizeof(struct vm_struct
), gfp_mask
);
2261 if (!vas
[area
] || !vms
[area
])
2265 spin_lock(&vmap_area_lock
);
2267 /* start scanning - we scan from the top, begin with the last area */
2268 area
= term_area
= last_area
;
2269 start
= offsets
[area
];
2270 end
= start
+ sizes
[area
];
2272 if (!pvm_find_next_prev(vmap_area_pcpu_hole
, &next
, &prev
)) {
2273 base
= vmalloc_end
- last_end
;
2276 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2279 BUG_ON(next
&& next
->va_end
<= base
+ end
);
2280 BUG_ON(prev
&& prev
->va_end
> base
+ end
);
2283 * base might have underflowed, add last_end before
2286 if (base
+ last_end
< vmalloc_start
+ last_end
) {
2287 spin_unlock(&vmap_area_lock
);
2289 purge_vmap_area_lazy();
2297 * If next overlaps, move base downwards so that it's
2298 * right below next and then recheck.
2300 if (next
&& next
->va_start
< base
+ end
) {
2301 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2307 * If prev overlaps, shift down next and prev and move
2308 * base so that it's right below new next and then
2311 if (prev
&& prev
->va_end
> base
+ start
) {
2313 prev
= node_to_va(rb_prev(&next
->rb_node
));
2314 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2320 * This area fits, move on to the previous one. If
2321 * the previous one is the terminal one, we're done.
2323 area
= (area
+ nr_vms
- 1) % nr_vms
;
2324 if (area
== term_area
)
2326 start
= offsets
[area
];
2327 end
= start
+ sizes
[area
];
2328 pvm_find_next_prev(base
+ end
, &next
, &prev
);
2331 /* we've found a fitting base, insert all va's */
2332 for (area
= 0; area
< nr_vms
; area
++) {
2333 struct vmap_area
*va
= vas
[area
];
2335 va
->va_start
= base
+ offsets
[area
];
2336 va
->va_end
= va
->va_start
+ sizes
[area
];
2337 __insert_vmap_area(va
);
2340 vmap_area_pcpu_hole
= base
+ offsets
[last_area
];
2342 spin_unlock(&vmap_area_lock
);
2344 /* insert all vm's */
2345 for (area
= 0; area
< nr_vms
; area
++)
2346 insert_vmalloc_vm(vms
[area
], vas
[area
], VM_ALLOC
,
2353 for (area
= 0; area
< nr_vms
; area
++) {
2365 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2366 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2367 * @nr_vms: the number of allocated areas
2369 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2371 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
2375 for (i
= 0; i
< nr_vms
; i
++)
2376 free_vm_area(vms
[i
]);
2380 #ifdef CONFIG_PROC_FS
2381 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
2384 struct vm_struct
*v
;
2386 read_lock(&vmlist_lock
);
2388 while (n
> 0 && v
) {
2399 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
2401 struct vm_struct
*v
= p
;
2407 static void s_stop(struct seq_file
*m
, void *p
)
2409 read_unlock(&vmlist_lock
);
2412 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
2415 unsigned int nr
, *counters
= m
->private;
2420 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
2422 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
2423 counters
[page_to_nid(v
->pages
[nr
])]++;
2425 for_each_node_state(nr
, N_HIGH_MEMORY
)
2427 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
2431 static int s_show(struct seq_file
*m
, void *p
)
2433 struct vm_struct
*v
= p
;
2435 seq_printf(m
, "0x%p-0x%p %7ld",
2436 v
->addr
, v
->addr
+ v
->size
, v
->size
);
2439 char buff
[KSYM_SYMBOL_LEN
];
2442 sprint_symbol(buff
, (unsigned long)v
->caller
);
2447 seq_printf(m
, " pages=%d", v
->nr_pages
);
2450 seq_printf(m
, " phys=%lx", v
->phys_addr
);
2452 if (v
->flags
& VM_IOREMAP
)
2453 seq_printf(m
, " ioremap");
2455 if (v
->flags
& VM_ALLOC
)
2456 seq_printf(m
, " vmalloc");
2458 if (v
->flags
& VM_MAP
)
2459 seq_printf(m
, " vmap");
2461 if (v
->flags
& VM_USERMAP
)
2462 seq_printf(m
, " user");
2464 if (v
->flags
& VM_VPAGES
)
2465 seq_printf(m
, " vpages");
2467 show_numa_info(m
, v
);
2472 static const struct seq_operations vmalloc_op
= {
2479 static int vmalloc_open(struct inode
*inode
, struct file
*file
)
2481 unsigned int *ptr
= NULL
;
2485 ptr
= kmalloc(nr_node_ids
* sizeof(unsigned int), GFP_KERNEL
);
2486 ret
= seq_open(file
, &vmalloc_op
);
2488 struct seq_file
*m
= file
->private_data
;
2495 static const struct file_operations proc_vmalloc_operations
= {
2496 .open
= vmalloc_open
,
2498 .llseek
= seq_lseek
,
2499 .release
= seq_release_private
,
2502 static int __init
proc_vmalloc_init(void)
2504 proc_create("vmallocinfo", S_IRUSR
, NULL
, &proc_vmalloc_operations
);
2507 module_init(proc_vmalloc_init
);