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
*t
, *llnode
;
54 llist_for_each_safe(llnode
, t
, llist_del_all(&p
->list
))
55 __vunmap((void *)llnode
, 1);
58 /*** Page table manipulation functions ***/
60 static void vunmap_pte_range(pmd_t
*pmd
, unsigned long addr
, unsigned long end
)
64 pte
= pte_offset_kernel(pmd
, addr
);
66 pte_t ptent
= ptep_get_and_clear(&init_mm
, addr
, pte
);
67 WARN_ON(!pte_none(ptent
) && !pte_present(ptent
));
68 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
71 static void vunmap_pmd_range(pud_t
*pud
, unsigned long addr
, unsigned long end
)
76 pmd
= pmd_offset(pud
, addr
);
78 next
= pmd_addr_end(addr
, end
);
79 if (pmd_clear_huge(pmd
))
81 if (pmd_none_or_clear_bad(pmd
))
83 vunmap_pte_range(pmd
, addr
, next
);
84 } while (pmd
++, addr
= next
, addr
!= end
);
87 static void vunmap_pud_range(p4d_t
*p4d
, unsigned long addr
, unsigned long end
)
92 pud
= pud_offset(p4d
, addr
);
94 next
= pud_addr_end(addr
, end
);
95 if (pud_clear_huge(pud
))
97 if (pud_none_or_clear_bad(pud
))
99 vunmap_pmd_range(pud
, addr
, next
);
100 } while (pud
++, addr
= next
, addr
!= end
);
103 static void vunmap_p4d_range(pgd_t
*pgd
, unsigned long addr
, unsigned long end
)
108 p4d
= p4d_offset(pgd
, addr
);
110 next
= p4d_addr_end(addr
, end
);
111 if (p4d_clear_huge(p4d
))
113 if (p4d_none_or_clear_bad(p4d
))
115 vunmap_pud_range(p4d
, addr
, next
);
116 } while (p4d
++, addr
= next
, addr
!= end
);
119 static void vunmap_page_range(unsigned long addr
, unsigned long end
)
125 pgd
= pgd_offset_k(addr
);
127 next
= pgd_addr_end(addr
, end
);
128 if (pgd_none_or_clear_bad(pgd
))
130 vunmap_p4d_range(pgd
, addr
, next
);
131 } while (pgd
++, addr
= next
, addr
!= end
);
134 static int vmap_pte_range(pmd_t
*pmd
, unsigned long addr
,
135 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
140 * nr is a running index into the array which helps higher level
141 * callers keep track of where we're up to.
144 pte
= pte_alloc_kernel(pmd
, addr
);
148 struct page
*page
= pages
[*nr
];
150 if (WARN_ON(!pte_none(*pte
)))
154 set_pte_at(&init_mm
, addr
, pte
, mk_pte(page
, prot
));
156 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
160 static int vmap_pmd_range(pud_t
*pud
, unsigned long addr
,
161 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
166 pmd
= pmd_alloc(&init_mm
, pud
, addr
);
170 next
= pmd_addr_end(addr
, end
);
171 if (vmap_pte_range(pmd
, addr
, next
, prot
, pages
, nr
))
173 } while (pmd
++, addr
= next
, addr
!= end
);
177 static int vmap_pud_range(p4d_t
*p4d
, unsigned long addr
,
178 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
183 pud
= pud_alloc(&init_mm
, p4d
, addr
);
187 next
= pud_addr_end(addr
, end
);
188 if (vmap_pmd_range(pud
, addr
, next
, prot
, pages
, nr
))
190 } while (pud
++, addr
= next
, addr
!= end
);
194 static int vmap_p4d_range(pgd_t
*pgd
, unsigned long addr
,
195 unsigned long end
, pgprot_t prot
, struct page
**pages
, int *nr
)
200 p4d
= p4d_alloc(&init_mm
, pgd
, addr
);
204 next
= p4d_addr_end(addr
, end
);
205 if (vmap_pud_range(p4d
, addr
, next
, prot
, pages
, nr
))
207 } while (p4d
++, addr
= next
, addr
!= end
);
212 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
213 * will have pfns corresponding to the "pages" array.
215 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
217 static int vmap_page_range_noflush(unsigned long start
, unsigned long end
,
218 pgprot_t prot
, struct page
**pages
)
222 unsigned long addr
= start
;
227 pgd
= pgd_offset_k(addr
);
229 next
= pgd_addr_end(addr
, end
);
230 err
= vmap_p4d_range(pgd
, addr
, next
, prot
, pages
, &nr
);
233 } while (pgd
++, addr
= next
, addr
!= end
);
238 static int vmap_page_range(unsigned long start
, unsigned long end
,
239 pgprot_t prot
, struct page
**pages
)
243 ret
= vmap_page_range_noflush(start
, end
, prot
, pages
);
244 flush_cache_vmap(start
, end
);
248 int is_vmalloc_or_module_addr(const void *x
)
251 * ARM, x86-64 and sparc64 put modules in a special place,
252 * and fall back on vmalloc() if that fails. Others
253 * just put it in the vmalloc space.
255 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
256 unsigned long addr
= (unsigned long)x
;
257 if (addr
>= MODULES_VADDR
&& addr
< MODULES_END
)
260 return is_vmalloc_addr(x
);
264 * Walk a vmap address to the struct page it maps.
266 struct page
*vmalloc_to_page(const void *vmalloc_addr
)
268 unsigned long addr
= (unsigned long) vmalloc_addr
;
269 struct page
*page
= NULL
;
270 pgd_t
*pgd
= pgd_offset_k(addr
);
277 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
278 * architectures that do not vmalloc module space
280 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr
));
284 p4d
= p4d_offset(pgd
, addr
);
287 pud
= pud_offset(p4d
, addr
);
290 * Don't dereference bad PUD or PMD (below) entries. This will also
291 * identify huge mappings, which we may encounter on architectures
292 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
293 * identified as vmalloc addresses by is_vmalloc_addr(), but are
294 * not [unambiguously] associated with a struct page, so there is
295 * no correct value to return for them.
297 WARN_ON_ONCE(pud_bad(*pud
));
298 if (pud_none(*pud
) || pud_bad(*pud
))
300 pmd
= pmd_offset(pud
, addr
);
301 WARN_ON_ONCE(pmd_bad(*pmd
));
302 if (pmd_none(*pmd
) || pmd_bad(*pmd
))
305 ptep
= pte_offset_map(pmd
, addr
);
307 if (pte_present(pte
))
308 page
= pte_page(pte
);
312 EXPORT_SYMBOL(vmalloc_to_page
);
315 * Map a vmalloc()-space virtual address to the physical page frame number.
317 unsigned long vmalloc_to_pfn(const void *vmalloc_addr
)
319 return page_to_pfn(vmalloc_to_page(vmalloc_addr
));
321 EXPORT_SYMBOL(vmalloc_to_pfn
);
324 /*** Global kva allocator ***/
326 #define VM_LAZY_FREE 0x02
327 #define VM_VM_AREA 0x04
329 static DEFINE_SPINLOCK(vmap_area_lock
);
330 /* Export for kexec only */
331 LIST_HEAD(vmap_area_list
);
332 static LLIST_HEAD(vmap_purge_list
);
333 static struct rb_root vmap_area_root
= RB_ROOT
;
335 /* The vmap cache globals are protected by vmap_area_lock */
336 static struct rb_node
*free_vmap_cache
;
337 static unsigned long cached_hole_size
;
338 static unsigned long cached_vstart
;
339 static unsigned long cached_align
;
341 static unsigned long vmap_area_pcpu_hole
;
343 static struct vmap_area
*__find_vmap_area(unsigned long addr
)
345 struct rb_node
*n
= vmap_area_root
.rb_node
;
348 struct vmap_area
*va
;
350 va
= rb_entry(n
, struct vmap_area
, rb_node
);
351 if (addr
< va
->va_start
)
353 else if (addr
>= va
->va_end
)
362 static void __insert_vmap_area(struct vmap_area
*va
)
364 struct rb_node
**p
= &vmap_area_root
.rb_node
;
365 struct rb_node
*parent
= NULL
;
369 struct vmap_area
*tmp_va
;
372 tmp_va
= rb_entry(parent
, struct vmap_area
, rb_node
);
373 if (va
->va_start
< tmp_va
->va_end
)
375 else if (va
->va_end
> tmp_va
->va_start
)
381 rb_link_node(&va
->rb_node
, parent
, p
);
382 rb_insert_color(&va
->rb_node
, &vmap_area_root
);
384 /* address-sort this list */
385 tmp
= rb_prev(&va
->rb_node
);
387 struct vmap_area
*prev
;
388 prev
= rb_entry(tmp
, struct vmap_area
, rb_node
);
389 list_add_rcu(&va
->list
, &prev
->list
);
391 list_add_rcu(&va
->list
, &vmap_area_list
);
394 static void purge_vmap_area_lazy(void);
396 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list
);
399 * Allocate a region of KVA of the specified size and alignment, within the
402 static struct vmap_area
*alloc_vmap_area(unsigned long size
,
404 unsigned long vstart
, unsigned long vend
,
405 int node
, gfp_t gfp_mask
)
407 struct vmap_area
*va
;
411 struct vmap_area
*first
;
414 BUG_ON(offset_in_page(size
));
415 BUG_ON(!is_power_of_2(align
));
419 va
= kmalloc_node(sizeof(struct vmap_area
),
420 gfp_mask
& GFP_RECLAIM_MASK
, node
);
422 return ERR_PTR(-ENOMEM
);
425 * Only scan the relevant parts containing pointers to other objects
426 * to avoid false negatives.
428 kmemleak_scan_area(&va
->rb_node
, SIZE_MAX
, gfp_mask
& GFP_RECLAIM_MASK
);
431 spin_lock(&vmap_area_lock
);
433 * Invalidate cache if we have more permissive parameters.
434 * cached_hole_size notes the largest hole noticed _below_
435 * the vmap_area cached in free_vmap_cache: if size fits
436 * into that hole, we want to scan from vstart to reuse
437 * the hole instead of allocating above free_vmap_cache.
438 * Note that __free_vmap_area may update free_vmap_cache
439 * without updating cached_hole_size or cached_align.
441 if (!free_vmap_cache
||
442 size
< cached_hole_size
||
443 vstart
< cached_vstart
||
444 align
< cached_align
) {
446 cached_hole_size
= 0;
447 free_vmap_cache
= NULL
;
449 /* record if we encounter less permissive parameters */
450 cached_vstart
= vstart
;
451 cached_align
= align
;
453 /* find starting point for our search */
454 if (free_vmap_cache
) {
455 first
= rb_entry(free_vmap_cache
, struct vmap_area
, rb_node
);
456 addr
= ALIGN(first
->va_end
, align
);
459 if (addr
+ size
< addr
)
463 addr
= ALIGN(vstart
, align
);
464 if (addr
+ size
< addr
)
467 n
= vmap_area_root
.rb_node
;
471 struct vmap_area
*tmp
;
472 tmp
= rb_entry(n
, struct vmap_area
, rb_node
);
473 if (tmp
->va_end
>= addr
) {
475 if (tmp
->va_start
<= addr
)
486 /* from the starting point, walk areas until a suitable hole is found */
487 while (addr
+ size
> first
->va_start
&& addr
+ size
<= vend
) {
488 if (addr
+ cached_hole_size
< first
->va_start
)
489 cached_hole_size
= first
->va_start
- addr
;
490 addr
= ALIGN(first
->va_end
, align
);
491 if (addr
+ size
< addr
)
494 if (list_is_last(&first
->list
, &vmap_area_list
))
497 first
= list_next_entry(first
, list
);
502 * Check also calculated address against the vstart,
503 * because it can be 0 because of big align request.
505 if (addr
+ size
> vend
|| addr
< vstart
)
509 va
->va_end
= addr
+ size
;
511 __insert_vmap_area(va
);
512 free_vmap_cache
= &va
->rb_node
;
513 spin_unlock(&vmap_area_lock
);
515 BUG_ON(!IS_ALIGNED(va
->va_start
, align
));
516 BUG_ON(va
->va_start
< vstart
);
517 BUG_ON(va
->va_end
> vend
);
522 spin_unlock(&vmap_area_lock
);
524 purge_vmap_area_lazy();
529 if (gfpflags_allow_blocking(gfp_mask
)) {
530 unsigned long freed
= 0;
531 blocking_notifier_call_chain(&vmap_notify_list
, 0, &freed
);
538 if (!(gfp_mask
& __GFP_NOWARN
) && printk_ratelimit())
539 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
542 return ERR_PTR(-EBUSY
);
545 int register_vmap_purge_notifier(struct notifier_block
*nb
)
547 return blocking_notifier_chain_register(&vmap_notify_list
, nb
);
549 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier
);
551 int unregister_vmap_purge_notifier(struct notifier_block
*nb
)
553 return blocking_notifier_chain_unregister(&vmap_notify_list
, nb
);
555 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier
);
557 static void __free_vmap_area(struct vmap_area
*va
)
559 BUG_ON(RB_EMPTY_NODE(&va
->rb_node
));
561 if (free_vmap_cache
) {
562 if (va
->va_end
< cached_vstart
) {
563 free_vmap_cache
= NULL
;
565 struct vmap_area
*cache
;
566 cache
= rb_entry(free_vmap_cache
, struct vmap_area
, rb_node
);
567 if (va
->va_start
<= cache
->va_start
) {
568 free_vmap_cache
= rb_prev(&va
->rb_node
);
570 * We don't try to update cached_hole_size or
571 * cached_align, but it won't go very wrong.
576 rb_erase(&va
->rb_node
, &vmap_area_root
);
577 RB_CLEAR_NODE(&va
->rb_node
);
578 list_del_rcu(&va
->list
);
581 * Track the highest possible candidate for pcpu area
582 * allocation. Areas outside of vmalloc area can be returned
583 * here too, consider only end addresses which fall inside
584 * vmalloc area proper.
586 if (va
->va_end
> VMALLOC_START
&& va
->va_end
<= VMALLOC_END
)
587 vmap_area_pcpu_hole
= max(vmap_area_pcpu_hole
, va
->va_end
);
589 kfree_rcu(va
, rcu_head
);
593 * Free a region of KVA allocated by alloc_vmap_area
595 static void free_vmap_area(struct vmap_area
*va
)
597 spin_lock(&vmap_area_lock
);
598 __free_vmap_area(va
);
599 spin_unlock(&vmap_area_lock
);
603 * Clear the pagetable entries of a given vmap_area
605 static void unmap_vmap_area(struct vmap_area
*va
)
607 vunmap_page_range(va
->va_start
, va
->va_end
);
610 static void vmap_debug_free_range(unsigned long start
, unsigned long end
)
613 * Unmap page tables and force a TLB flush immediately if pagealloc
614 * debugging is enabled. This catches use after free bugs similarly to
615 * those in linear kernel virtual address space after a page has been
618 * All the lazy freeing logic is still retained, in order to minimise
619 * intrusiveness of this debugging feature.
621 * This is going to be *slow* (linear kernel virtual address debugging
622 * doesn't do a broadcast TLB flush so it is a lot faster).
624 if (debug_pagealloc_enabled()) {
625 vunmap_page_range(start
, end
);
626 flush_tlb_kernel_range(start
, end
);
631 * lazy_max_pages is the maximum amount of virtual address space we gather up
632 * before attempting to purge with a TLB flush.
634 * There is a tradeoff here: a larger number will cover more kernel page tables
635 * and take slightly longer to purge, but it will linearly reduce the number of
636 * global TLB flushes that must be performed. It would seem natural to scale
637 * this number up linearly with the number of CPUs (because vmapping activity
638 * could also scale linearly with the number of CPUs), however it is likely
639 * that in practice, workloads might be constrained in other ways that mean
640 * vmap activity will not scale linearly with CPUs. Also, I want to be
641 * conservative and not introduce a big latency on huge systems, so go with
642 * a less aggressive log scale. It will still be an improvement over the old
643 * code, and it will be simple to change the scale factor if we find that it
644 * becomes a problem on bigger systems.
646 static unsigned long lazy_max_pages(void)
650 log
= fls(num_online_cpus());
652 return log
* (32UL * 1024 * 1024 / PAGE_SIZE
);
655 static atomic_t vmap_lazy_nr
= ATOMIC_INIT(0);
658 * Serialize vmap purging. There is no actual criticial section protected
659 * by this look, but we want to avoid concurrent calls for performance
660 * reasons and to make the pcpu_get_vm_areas more deterministic.
662 static DEFINE_MUTEX(vmap_purge_lock
);
664 /* for per-CPU blocks */
665 static void purge_fragmented_blocks_allcpus(void);
668 * called before a call to iounmap() if the caller wants vm_area_struct's
671 void set_iounmap_nonlazy(void)
673 atomic_set(&vmap_lazy_nr
, lazy_max_pages()+1);
677 * Purges all lazily-freed vmap areas.
679 static bool __purge_vmap_area_lazy(unsigned long start
, unsigned long end
)
681 struct llist_node
*valist
;
682 struct vmap_area
*va
;
683 struct vmap_area
*n_va
;
684 int resched_threshold
;
686 lockdep_assert_held(&vmap_purge_lock
);
688 valist
= llist_del_all(&vmap_purge_list
);
689 if (unlikely(valist
== NULL
))
693 * First make sure the mappings are removed from all page-tables
694 * before they are freed.
699 * TODO: to calculate a flush range without looping.
700 * The list can be up to lazy_max_pages() elements.
702 llist_for_each_entry(va
, valist
, purge_list
) {
703 if (va
->va_start
< start
)
704 start
= va
->va_start
;
705 if (va
->va_end
> end
)
709 flush_tlb_kernel_range(start
, end
);
710 resched_threshold
= (int) lazy_max_pages() << 1;
712 spin_lock(&vmap_area_lock
);
713 llist_for_each_entry_safe(va
, n_va
, valist
, purge_list
) {
714 int nr
= (va
->va_end
- va
->va_start
) >> PAGE_SHIFT
;
716 __free_vmap_area(va
);
717 atomic_sub(nr
, &vmap_lazy_nr
);
719 if (atomic_read(&vmap_lazy_nr
) < resched_threshold
)
720 cond_resched_lock(&vmap_area_lock
);
722 spin_unlock(&vmap_area_lock
);
727 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
728 * is already purging.
730 static void try_purge_vmap_area_lazy(void)
732 if (mutex_trylock(&vmap_purge_lock
)) {
733 __purge_vmap_area_lazy(ULONG_MAX
, 0);
734 mutex_unlock(&vmap_purge_lock
);
739 * Kick off a purge of the outstanding lazy areas.
741 static void purge_vmap_area_lazy(void)
743 mutex_lock(&vmap_purge_lock
);
744 purge_fragmented_blocks_allcpus();
745 __purge_vmap_area_lazy(ULONG_MAX
, 0);
746 mutex_unlock(&vmap_purge_lock
);
750 * Free a vmap area, caller ensuring that the area has been unmapped
751 * and flush_cache_vunmap had been called for the correct range
754 static void free_vmap_area_noflush(struct vmap_area
*va
)
758 nr_lazy
= atomic_add_return((va
->va_end
- va
->va_start
) >> PAGE_SHIFT
,
761 /* After this point, we may free va at any time */
762 llist_add(&va
->purge_list
, &vmap_purge_list
);
764 if (unlikely(nr_lazy
> lazy_max_pages()))
765 try_purge_vmap_area_lazy();
769 * Free and unmap a vmap area
771 static void free_unmap_vmap_area(struct vmap_area
*va
)
773 flush_cache_vunmap(va
->va_start
, va
->va_end
);
775 free_vmap_area_noflush(va
);
778 static struct vmap_area
*find_vmap_area(unsigned long addr
)
780 struct vmap_area
*va
;
782 spin_lock(&vmap_area_lock
);
783 va
= __find_vmap_area(addr
);
784 spin_unlock(&vmap_area_lock
);
789 /*** Per cpu kva allocator ***/
792 * vmap space is limited especially on 32 bit architectures. Ensure there is
793 * room for at least 16 percpu vmap blocks per CPU.
796 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
797 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
798 * instead (we just need a rough idea)
800 #if BITS_PER_LONG == 32
801 #define VMALLOC_SPACE (128UL*1024*1024)
803 #define VMALLOC_SPACE (128UL*1024*1024*1024)
806 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
807 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
808 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
809 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
810 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
811 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
812 #define VMAP_BBMAP_BITS \
813 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
814 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
815 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
817 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
819 static bool vmap_initialized __read_mostly
= false;
821 struct vmap_block_queue
{
823 struct list_head free
;
828 struct vmap_area
*va
;
829 unsigned long free
, dirty
;
830 unsigned long dirty_min
, dirty_max
; /*< dirty range */
831 struct list_head free_list
;
832 struct rcu_head rcu_head
;
833 struct list_head purge
;
836 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
837 static DEFINE_PER_CPU(struct vmap_block_queue
, vmap_block_queue
);
840 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
841 * in the free path. Could get rid of this if we change the API to return a
842 * "cookie" from alloc, to be passed to free. But no big deal yet.
844 static DEFINE_SPINLOCK(vmap_block_tree_lock
);
845 static RADIX_TREE(vmap_block_tree
, GFP_ATOMIC
);
848 * We should probably have a fallback mechanism to allocate virtual memory
849 * out of partially filled vmap blocks. However vmap block sizing should be
850 * fairly reasonable according to the vmalloc size, so it shouldn't be a
854 static unsigned long addr_to_vb_idx(unsigned long addr
)
856 addr
-= VMALLOC_START
& ~(VMAP_BLOCK_SIZE
-1);
857 addr
/= VMAP_BLOCK_SIZE
;
861 static void *vmap_block_vaddr(unsigned long va_start
, unsigned long pages_off
)
865 addr
= va_start
+ (pages_off
<< PAGE_SHIFT
);
866 BUG_ON(addr_to_vb_idx(addr
) != addr_to_vb_idx(va_start
));
871 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
872 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
873 * @order: how many 2^order pages should be occupied in newly allocated block
874 * @gfp_mask: flags for the page level allocator
876 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
878 static void *new_vmap_block(unsigned int order
, gfp_t gfp_mask
)
880 struct vmap_block_queue
*vbq
;
881 struct vmap_block
*vb
;
882 struct vmap_area
*va
;
883 unsigned long vb_idx
;
887 node
= numa_node_id();
889 vb
= kmalloc_node(sizeof(struct vmap_block
),
890 gfp_mask
& GFP_RECLAIM_MASK
, node
);
892 return ERR_PTR(-ENOMEM
);
894 va
= alloc_vmap_area(VMAP_BLOCK_SIZE
, VMAP_BLOCK_SIZE
,
895 VMALLOC_START
, VMALLOC_END
,
902 err
= radix_tree_preload(gfp_mask
);
909 vaddr
= vmap_block_vaddr(va
->va_start
, 0);
910 spin_lock_init(&vb
->lock
);
912 /* At least something should be left free */
913 BUG_ON(VMAP_BBMAP_BITS
<= (1UL << order
));
914 vb
->free
= VMAP_BBMAP_BITS
- (1UL << order
);
916 vb
->dirty_min
= VMAP_BBMAP_BITS
;
918 INIT_LIST_HEAD(&vb
->free_list
);
920 vb_idx
= addr_to_vb_idx(va
->va_start
);
921 spin_lock(&vmap_block_tree_lock
);
922 err
= radix_tree_insert(&vmap_block_tree
, vb_idx
, vb
);
923 spin_unlock(&vmap_block_tree_lock
);
925 radix_tree_preload_end();
927 vbq
= &get_cpu_var(vmap_block_queue
);
928 spin_lock(&vbq
->lock
);
929 list_add_tail_rcu(&vb
->free_list
, &vbq
->free
);
930 spin_unlock(&vbq
->lock
);
931 put_cpu_var(vmap_block_queue
);
936 static void free_vmap_block(struct vmap_block
*vb
)
938 struct vmap_block
*tmp
;
939 unsigned long vb_idx
;
941 vb_idx
= addr_to_vb_idx(vb
->va
->va_start
);
942 spin_lock(&vmap_block_tree_lock
);
943 tmp
= radix_tree_delete(&vmap_block_tree
, vb_idx
);
944 spin_unlock(&vmap_block_tree_lock
);
947 free_vmap_area_noflush(vb
->va
);
948 kfree_rcu(vb
, rcu_head
);
951 static void purge_fragmented_blocks(int cpu
)
954 struct vmap_block
*vb
;
955 struct vmap_block
*n_vb
;
956 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
959 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
961 if (!(vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
))
964 spin_lock(&vb
->lock
);
965 if (vb
->free
+ vb
->dirty
== VMAP_BBMAP_BITS
&& vb
->dirty
!= VMAP_BBMAP_BITS
) {
966 vb
->free
= 0; /* prevent further allocs after releasing lock */
967 vb
->dirty
= VMAP_BBMAP_BITS
; /* prevent purging it again */
969 vb
->dirty_max
= VMAP_BBMAP_BITS
;
970 spin_lock(&vbq
->lock
);
971 list_del_rcu(&vb
->free_list
);
972 spin_unlock(&vbq
->lock
);
973 spin_unlock(&vb
->lock
);
974 list_add_tail(&vb
->purge
, &purge
);
976 spin_unlock(&vb
->lock
);
980 list_for_each_entry_safe(vb
, n_vb
, &purge
, purge
) {
981 list_del(&vb
->purge
);
986 static void purge_fragmented_blocks_allcpus(void)
990 for_each_possible_cpu(cpu
)
991 purge_fragmented_blocks(cpu
);
994 static void *vb_alloc(unsigned long size
, gfp_t gfp_mask
)
996 struct vmap_block_queue
*vbq
;
997 struct vmap_block
*vb
;
1001 BUG_ON(offset_in_page(size
));
1002 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1003 if (WARN_ON(size
== 0)) {
1005 * Allocating 0 bytes isn't what caller wants since
1006 * get_order(0) returns funny result. Just warn and terminate
1011 order
= get_order(size
);
1014 vbq
= &get_cpu_var(vmap_block_queue
);
1015 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1016 unsigned long pages_off
;
1018 spin_lock(&vb
->lock
);
1019 if (vb
->free
< (1UL << order
)) {
1020 spin_unlock(&vb
->lock
);
1024 pages_off
= VMAP_BBMAP_BITS
- vb
->free
;
1025 vaddr
= vmap_block_vaddr(vb
->va
->va_start
, pages_off
);
1026 vb
->free
-= 1UL << order
;
1027 if (vb
->free
== 0) {
1028 spin_lock(&vbq
->lock
);
1029 list_del_rcu(&vb
->free_list
);
1030 spin_unlock(&vbq
->lock
);
1033 spin_unlock(&vb
->lock
);
1037 put_cpu_var(vmap_block_queue
);
1040 /* Allocate new block if nothing was found */
1042 vaddr
= new_vmap_block(order
, gfp_mask
);
1047 static void vb_free(const void *addr
, unsigned long size
)
1049 unsigned long offset
;
1050 unsigned long vb_idx
;
1052 struct vmap_block
*vb
;
1054 BUG_ON(offset_in_page(size
));
1055 BUG_ON(size
> PAGE_SIZE
*VMAP_MAX_ALLOC
);
1057 flush_cache_vunmap((unsigned long)addr
, (unsigned long)addr
+ size
);
1059 order
= get_order(size
);
1061 offset
= (unsigned long)addr
& (VMAP_BLOCK_SIZE
- 1);
1062 offset
>>= PAGE_SHIFT
;
1064 vb_idx
= addr_to_vb_idx((unsigned long)addr
);
1066 vb
= radix_tree_lookup(&vmap_block_tree
, vb_idx
);
1070 vunmap_page_range((unsigned long)addr
, (unsigned long)addr
+ size
);
1072 spin_lock(&vb
->lock
);
1074 /* Expand dirty range */
1075 vb
->dirty_min
= min(vb
->dirty_min
, offset
);
1076 vb
->dirty_max
= max(vb
->dirty_max
, offset
+ (1UL << order
));
1078 vb
->dirty
+= 1UL << order
;
1079 if (vb
->dirty
== VMAP_BBMAP_BITS
) {
1081 spin_unlock(&vb
->lock
);
1082 free_vmap_block(vb
);
1084 spin_unlock(&vb
->lock
);
1088 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1090 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1091 * to amortize TLB flushing overheads. What this means is that any page you
1092 * have now, may, in a former life, have been mapped into kernel virtual
1093 * address by the vmap layer and so there might be some CPUs with TLB entries
1094 * still referencing that page (additional to the regular 1:1 kernel mapping).
1096 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1097 * be sure that none of the pages we have control over will have any aliases
1098 * from the vmap layer.
1100 void vm_unmap_aliases(void)
1102 unsigned long start
= ULONG_MAX
, end
= 0;
1106 if (unlikely(!vmap_initialized
))
1111 for_each_possible_cpu(cpu
) {
1112 struct vmap_block_queue
*vbq
= &per_cpu(vmap_block_queue
, cpu
);
1113 struct vmap_block
*vb
;
1116 list_for_each_entry_rcu(vb
, &vbq
->free
, free_list
) {
1117 spin_lock(&vb
->lock
);
1119 unsigned long va_start
= vb
->va
->va_start
;
1122 s
= va_start
+ (vb
->dirty_min
<< PAGE_SHIFT
);
1123 e
= va_start
+ (vb
->dirty_max
<< PAGE_SHIFT
);
1125 start
= min(s
, start
);
1130 spin_unlock(&vb
->lock
);
1135 mutex_lock(&vmap_purge_lock
);
1136 purge_fragmented_blocks_allcpus();
1137 if (!__purge_vmap_area_lazy(start
, end
) && flush
)
1138 flush_tlb_kernel_range(start
, end
);
1139 mutex_unlock(&vmap_purge_lock
);
1141 EXPORT_SYMBOL_GPL(vm_unmap_aliases
);
1144 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1145 * @mem: the pointer returned by vm_map_ram
1146 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1148 void vm_unmap_ram(const void *mem
, unsigned int count
)
1150 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1151 unsigned long addr
= (unsigned long)mem
;
1152 struct vmap_area
*va
;
1156 BUG_ON(addr
< VMALLOC_START
);
1157 BUG_ON(addr
> VMALLOC_END
);
1158 BUG_ON(!PAGE_ALIGNED(addr
));
1160 debug_check_no_locks_freed(mem
, size
);
1161 vmap_debug_free_range(addr
, addr
+size
);
1163 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1168 va
= find_vmap_area(addr
);
1170 free_unmap_vmap_area(va
);
1172 EXPORT_SYMBOL(vm_unmap_ram
);
1175 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1176 * @pages: an array of pointers to the pages to be mapped
1177 * @count: number of pages
1178 * @node: prefer to allocate data structures on this node
1179 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1181 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1182 * faster than vmap so it's good. But if you mix long-life and short-life
1183 * objects with vm_map_ram(), it could consume lots of address space through
1184 * fragmentation (especially on a 32bit machine). You could see failures in
1185 * the end. Please use this function for short-lived objects.
1187 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1189 void *vm_map_ram(struct page
**pages
, unsigned int count
, int node
, pgprot_t prot
)
1191 unsigned long size
= (unsigned long)count
<< PAGE_SHIFT
;
1195 if (likely(count
<= VMAP_MAX_ALLOC
)) {
1196 mem
= vb_alloc(size
, GFP_KERNEL
);
1199 addr
= (unsigned long)mem
;
1201 struct vmap_area
*va
;
1202 va
= alloc_vmap_area(size
, PAGE_SIZE
,
1203 VMALLOC_START
, VMALLOC_END
, node
, GFP_KERNEL
);
1207 addr
= va
->va_start
;
1210 if (vmap_page_range(addr
, addr
+ size
, prot
, pages
) < 0) {
1211 vm_unmap_ram(mem
, count
);
1216 EXPORT_SYMBOL(vm_map_ram
);
1218 static struct vm_struct
*vmlist __initdata
;
1220 * vm_area_add_early - add vmap area early during boot
1221 * @vm: vm_struct to add
1223 * This function is used to add fixed kernel vm area to vmlist before
1224 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1225 * should contain proper values and the other fields should be zero.
1227 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1229 void __init
vm_area_add_early(struct vm_struct
*vm
)
1231 struct vm_struct
*tmp
, **p
;
1233 BUG_ON(vmap_initialized
);
1234 for (p
= &vmlist
; (tmp
= *p
) != NULL
; p
= &tmp
->next
) {
1235 if (tmp
->addr
>= vm
->addr
) {
1236 BUG_ON(tmp
->addr
< vm
->addr
+ vm
->size
);
1239 BUG_ON(tmp
->addr
+ tmp
->size
> vm
->addr
);
1246 * vm_area_register_early - register vmap area early during boot
1247 * @vm: vm_struct to register
1248 * @align: requested alignment
1250 * This function is used to register kernel vm area before
1251 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1252 * proper values on entry and other fields should be zero. On return,
1253 * vm->addr contains the allocated address.
1255 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1257 void __init
vm_area_register_early(struct vm_struct
*vm
, size_t align
)
1259 static size_t vm_init_off __initdata
;
1262 addr
= ALIGN(VMALLOC_START
+ vm_init_off
, align
);
1263 vm_init_off
= PFN_ALIGN(addr
+ vm
->size
) - VMALLOC_START
;
1265 vm
->addr
= (void *)addr
;
1267 vm_area_add_early(vm
);
1270 void __init
vmalloc_init(void)
1272 struct vmap_area
*va
;
1273 struct vm_struct
*tmp
;
1276 for_each_possible_cpu(i
) {
1277 struct vmap_block_queue
*vbq
;
1278 struct vfree_deferred
*p
;
1280 vbq
= &per_cpu(vmap_block_queue
, i
);
1281 spin_lock_init(&vbq
->lock
);
1282 INIT_LIST_HEAD(&vbq
->free
);
1283 p
= &per_cpu(vfree_deferred
, i
);
1284 init_llist_head(&p
->list
);
1285 INIT_WORK(&p
->wq
, free_work
);
1288 /* Import existing vmlist entries. */
1289 for (tmp
= vmlist
; tmp
; tmp
= tmp
->next
) {
1290 va
= kzalloc(sizeof(struct vmap_area
), GFP_NOWAIT
);
1291 va
->flags
= VM_VM_AREA
;
1292 va
->va_start
= (unsigned long)tmp
->addr
;
1293 va
->va_end
= va
->va_start
+ tmp
->size
;
1295 __insert_vmap_area(va
);
1298 vmap_area_pcpu_hole
= VMALLOC_END
;
1300 vmap_initialized
= true;
1304 * map_kernel_range_noflush - map kernel VM area with the specified pages
1305 * @addr: start of the VM area to map
1306 * @size: size of the VM area to map
1307 * @prot: page protection flags to use
1308 * @pages: pages to map
1310 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1311 * specify should have been allocated using get_vm_area() and its
1315 * This function does NOT do any cache flushing. The caller is
1316 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1317 * before calling this function.
1320 * The number of pages mapped on success, -errno on failure.
1322 int map_kernel_range_noflush(unsigned long addr
, unsigned long size
,
1323 pgprot_t prot
, struct page
**pages
)
1325 return vmap_page_range_noflush(addr
, addr
+ size
, prot
, pages
);
1329 * unmap_kernel_range_noflush - unmap kernel VM area
1330 * @addr: start of the VM area to unmap
1331 * @size: size of the VM area to unmap
1333 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1334 * specify should have been allocated using get_vm_area() and its
1338 * This function does NOT do any cache flushing. The caller is
1339 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1340 * before calling this function and flush_tlb_kernel_range() after.
1342 void unmap_kernel_range_noflush(unsigned long addr
, unsigned long size
)
1344 vunmap_page_range(addr
, addr
+ size
);
1346 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush
);
1349 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1350 * @addr: start of the VM area to unmap
1351 * @size: size of the VM area to unmap
1353 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1354 * the unmapping and tlb after.
1356 void unmap_kernel_range(unsigned long addr
, unsigned long size
)
1358 unsigned long end
= addr
+ size
;
1360 flush_cache_vunmap(addr
, end
);
1361 vunmap_page_range(addr
, end
);
1362 flush_tlb_kernel_range(addr
, end
);
1364 EXPORT_SYMBOL_GPL(unmap_kernel_range
);
1366 int map_vm_area(struct vm_struct
*area
, pgprot_t prot
, struct page
**pages
)
1368 unsigned long addr
= (unsigned long)area
->addr
;
1369 unsigned long end
= addr
+ get_vm_area_size(area
);
1372 err
= vmap_page_range(addr
, end
, prot
, pages
);
1374 return err
> 0 ? 0 : err
;
1376 EXPORT_SYMBOL_GPL(map_vm_area
);
1378 static void setup_vmalloc_vm(struct vm_struct
*vm
, struct vmap_area
*va
,
1379 unsigned long flags
, const void *caller
)
1381 spin_lock(&vmap_area_lock
);
1383 vm
->addr
= (void *)va
->va_start
;
1384 vm
->size
= va
->va_end
- va
->va_start
;
1385 vm
->caller
= caller
;
1387 va
->flags
|= VM_VM_AREA
;
1388 spin_unlock(&vmap_area_lock
);
1391 static void clear_vm_uninitialized_flag(struct vm_struct
*vm
)
1394 * Before removing VM_UNINITIALIZED,
1395 * we should make sure that vm has proper values.
1396 * Pair with smp_rmb() in show_numa_info().
1399 vm
->flags
&= ~VM_UNINITIALIZED
;
1402 static struct vm_struct
*__get_vm_area_node(unsigned long size
,
1403 unsigned long align
, unsigned long flags
, unsigned long start
,
1404 unsigned long end
, int node
, gfp_t gfp_mask
, const void *caller
)
1406 struct vmap_area
*va
;
1407 struct vm_struct
*area
;
1409 BUG_ON(in_interrupt());
1410 size
= PAGE_ALIGN(size
);
1411 if (unlikely(!size
))
1414 if (flags
& VM_IOREMAP
)
1415 align
= 1ul << clamp_t(int, get_count_order_long(size
),
1416 PAGE_SHIFT
, IOREMAP_MAX_ORDER
);
1418 area
= kzalloc_node(sizeof(*area
), gfp_mask
& GFP_RECLAIM_MASK
, node
);
1419 if (unlikely(!area
))
1422 if (!(flags
& VM_NO_GUARD
))
1425 va
= alloc_vmap_area(size
, align
, start
, end
, node
, gfp_mask
);
1431 setup_vmalloc_vm(area
, va
, flags
, caller
);
1436 struct vm_struct
*__get_vm_area(unsigned long size
, unsigned long flags
,
1437 unsigned long start
, unsigned long end
)
1439 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
1440 GFP_KERNEL
, __builtin_return_address(0));
1442 EXPORT_SYMBOL_GPL(__get_vm_area
);
1444 struct vm_struct
*__get_vm_area_caller(unsigned long size
, unsigned long flags
,
1445 unsigned long start
, unsigned long end
,
1448 return __get_vm_area_node(size
, 1, flags
, start
, end
, NUMA_NO_NODE
,
1449 GFP_KERNEL
, caller
);
1453 * get_vm_area - reserve a contiguous kernel virtual area
1454 * @size: size of the area
1455 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1457 * Search an area of @size in the kernel virtual mapping area,
1458 * and reserved it for out purposes. Returns the area descriptor
1459 * on success or %NULL on failure.
1461 struct vm_struct
*get_vm_area(unsigned long size
, unsigned long flags
)
1463 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1464 NUMA_NO_NODE
, GFP_KERNEL
,
1465 __builtin_return_address(0));
1468 struct vm_struct
*get_vm_area_caller(unsigned long size
, unsigned long flags
,
1471 return __get_vm_area_node(size
, 1, flags
, VMALLOC_START
, VMALLOC_END
,
1472 NUMA_NO_NODE
, GFP_KERNEL
, caller
);
1476 * find_vm_area - find a continuous kernel virtual area
1477 * @addr: base address
1479 * Search for the kernel VM area starting at @addr, and return it.
1480 * It is up to the caller to do all required locking to keep the returned
1483 struct vm_struct
*find_vm_area(const void *addr
)
1485 struct vmap_area
*va
;
1487 va
= find_vmap_area((unsigned long)addr
);
1488 if (va
&& va
->flags
& VM_VM_AREA
)
1495 * remove_vm_area - find and remove a continuous kernel virtual area
1496 * @addr: base address
1498 * Search for the kernel VM area starting at @addr, and remove it.
1499 * This function returns the found VM area, but using it is NOT safe
1500 * on SMP machines, except for its size or flags.
1502 struct vm_struct
*remove_vm_area(const void *addr
)
1504 struct vmap_area
*va
;
1508 va
= find_vmap_area((unsigned long)addr
);
1509 if (va
&& va
->flags
& VM_VM_AREA
) {
1510 struct vm_struct
*vm
= va
->vm
;
1512 spin_lock(&vmap_area_lock
);
1514 va
->flags
&= ~VM_VM_AREA
;
1515 va
->flags
|= VM_LAZY_FREE
;
1516 spin_unlock(&vmap_area_lock
);
1518 vmap_debug_free_range(va
->va_start
, va
->va_end
);
1519 kasan_free_shadow(vm
);
1520 free_unmap_vmap_area(va
);
1527 static void __vunmap(const void *addr
, int deallocate_pages
)
1529 struct vm_struct
*area
;
1534 if (WARN(!PAGE_ALIGNED(addr
), "Trying to vfree() bad address (%p)\n",
1538 area
= find_vmap_area((unsigned long)addr
)->vm
;
1539 if (unlikely(!area
)) {
1540 WARN(1, KERN_ERR
"Trying to vfree() nonexistent vm area (%p)\n",
1545 debug_check_no_locks_freed(addr
, get_vm_area_size(area
));
1546 debug_check_no_obj_freed(addr
, get_vm_area_size(area
));
1548 remove_vm_area(addr
);
1549 if (deallocate_pages
) {
1552 for (i
= 0; i
< area
->nr_pages
; i
++) {
1553 struct page
*page
= area
->pages
[i
];
1556 __free_pages(page
, 0);
1559 kvfree(area
->pages
);
1566 static inline void __vfree_deferred(const void *addr
)
1569 * Use raw_cpu_ptr() because this can be called from preemptible
1570 * context. Preemption is absolutely fine here, because the llist_add()
1571 * implementation is lockless, so it works even if we are adding to
1572 * nother cpu's list. schedule_work() should be fine with this too.
1574 struct vfree_deferred
*p
= raw_cpu_ptr(&vfree_deferred
);
1576 if (llist_add((struct llist_node
*)addr
, &p
->list
))
1577 schedule_work(&p
->wq
);
1581 * vfree_atomic - release memory allocated by vmalloc()
1582 * @addr: memory base address
1584 * This one is just like vfree() but can be called in any atomic context
1587 void vfree_atomic(const void *addr
)
1591 kmemleak_free(addr
);
1595 __vfree_deferred(addr
);
1599 * vfree - release memory allocated by vmalloc()
1600 * @addr: memory base address
1602 * Free the virtually continuous memory area starting at @addr, as
1603 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1604 * NULL, no operation is performed.
1606 * Must not be called in NMI context (strictly speaking, only if we don't
1607 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1608 * conventions for vfree() arch-depenedent would be a really bad idea)
1610 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1612 void vfree(const void *addr
)
1616 kmemleak_free(addr
);
1620 if (unlikely(in_interrupt()))
1621 __vfree_deferred(addr
);
1625 EXPORT_SYMBOL(vfree
);
1628 * vunmap - release virtual mapping obtained by vmap()
1629 * @addr: memory base address
1631 * Free the virtually contiguous memory area starting at @addr,
1632 * which was created from the page array passed to vmap().
1634 * Must not be called in interrupt context.
1636 void vunmap(const void *addr
)
1638 BUG_ON(in_interrupt());
1643 EXPORT_SYMBOL(vunmap
);
1646 * vmap - map an array of pages into virtually contiguous space
1647 * @pages: array of page pointers
1648 * @count: number of pages to map
1649 * @flags: vm_area->flags
1650 * @prot: page protection for the mapping
1652 * Maps @count pages from @pages into contiguous kernel virtual
1655 void *vmap(struct page
**pages
, unsigned int count
,
1656 unsigned long flags
, pgprot_t prot
)
1658 struct vm_struct
*area
;
1659 unsigned long size
; /* In bytes */
1663 if (count
> totalram_pages
)
1666 size
= (unsigned long)count
<< PAGE_SHIFT
;
1667 area
= get_vm_area_caller(size
, flags
, __builtin_return_address(0));
1671 if (map_vm_area(area
, prot
, pages
)) {
1678 EXPORT_SYMBOL(vmap
);
1680 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1681 gfp_t gfp_mask
, pgprot_t prot
,
1682 int node
, const void *caller
);
1683 static void *__vmalloc_area_node(struct vm_struct
*area
, gfp_t gfp_mask
,
1684 pgprot_t prot
, int node
)
1686 struct page
**pages
;
1687 unsigned int nr_pages
, array_size
, i
;
1688 const gfp_t nested_gfp
= (gfp_mask
& GFP_RECLAIM_MASK
) | __GFP_ZERO
;
1689 const gfp_t alloc_mask
= gfp_mask
| __GFP_NOWARN
;
1690 const gfp_t highmem_mask
= (gfp_mask
& (GFP_DMA
| GFP_DMA32
)) ?
1694 nr_pages
= get_vm_area_size(area
) >> PAGE_SHIFT
;
1695 array_size
= (nr_pages
* sizeof(struct page
*));
1697 area
->nr_pages
= nr_pages
;
1698 /* Please note that the recursion is strictly bounded. */
1699 if (array_size
> PAGE_SIZE
) {
1700 pages
= __vmalloc_node(array_size
, 1, nested_gfp
|highmem_mask
,
1701 PAGE_KERNEL
, node
, area
->caller
);
1703 pages
= kmalloc_node(array_size
, nested_gfp
, node
);
1705 area
->pages
= pages
;
1707 remove_vm_area(area
->addr
);
1712 for (i
= 0; i
< area
->nr_pages
; i
++) {
1715 if (node
== NUMA_NO_NODE
)
1716 page
= alloc_page(alloc_mask
|highmem_mask
);
1718 page
= alloc_pages_node(node
, alloc_mask
|highmem_mask
, 0);
1720 if (unlikely(!page
)) {
1721 /* Successfully allocated i pages, free them in __vunmap() */
1725 area
->pages
[i
] = page
;
1726 if (gfpflags_allow_blocking(gfp_mask
|highmem_mask
))
1730 if (map_vm_area(area
, prot
, pages
))
1735 warn_alloc(gfp_mask
, NULL
,
1736 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1737 (area
->nr_pages
*PAGE_SIZE
), area
->size
);
1743 * __vmalloc_node_range - allocate virtually contiguous memory
1744 * @size: allocation size
1745 * @align: desired alignment
1746 * @start: vm area range start
1747 * @end: vm area range end
1748 * @gfp_mask: flags for the page level allocator
1749 * @prot: protection mask for the allocated pages
1750 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1751 * @node: node to use for allocation or NUMA_NO_NODE
1752 * @caller: caller's return address
1754 * Allocate enough pages to cover @size from the page level
1755 * allocator with @gfp_mask flags. Map them into contiguous
1756 * kernel virtual space, using a pagetable protection of @prot.
1758 void *__vmalloc_node_range(unsigned long size
, unsigned long align
,
1759 unsigned long start
, unsigned long end
, gfp_t gfp_mask
,
1760 pgprot_t prot
, unsigned long vm_flags
, int node
,
1763 struct vm_struct
*area
;
1765 unsigned long real_size
= size
;
1767 size
= PAGE_ALIGN(size
);
1768 if (!size
|| (size
>> PAGE_SHIFT
) > totalram_pages
)
1771 area
= __get_vm_area_node(size
, align
, VM_ALLOC
| VM_UNINITIALIZED
|
1772 vm_flags
, start
, end
, node
, gfp_mask
, caller
);
1776 addr
= __vmalloc_area_node(area
, gfp_mask
, prot
, node
);
1781 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1782 * flag. It means that vm_struct is not fully initialized.
1783 * Now, it is fully initialized, so remove this flag here.
1785 clear_vm_uninitialized_flag(area
);
1787 kmemleak_vmalloc(area
, size
, gfp_mask
);
1792 warn_alloc(gfp_mask
, NULL
,
1793 "vmalloc: allocation failure: %lu bytes", real_size
);
1798 * __vmalloc_node - allocate virtually contiguous memory
1799 * @size: allocation size
1800 * @align: desired alignment
1801 * @gfp_mask: flags for the page level allocator
1802 * @prot: protection mask for the allocated pages
1803 * @node: node to use for allocation or NUMA_NO_NODE
1804 * @caller: caller's return address
1806 * Allocate enough pages to cover @size from the page level
1807 * allocator with @gfp_mask flags. Map them into contiguous
1808 * kernel virtual space, using a pagetable protection of @prot.
1810 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1811 * and __GFP_NOFAIL are not supported
1813 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1817 static void *__vmalloc_node(unsigned long size
, unsigned long align
,
1818 gfp_t gfp_mask
, pgprot_t prot
,
1819 int node
, const void *caller
)
1821 return __vmalloc_node_range(size
, align
, VMALLOC_START
, VMALLOC_END
,
1822 gfp_mask
, prot
, 0, node
, caller
);
1825 void *__vmalloc(unsigned long size
, gfp_t gfp_mask
, pgprot_t prot
)
1827 return __vmalloc_node(size
, 1, gfp_mask
, prot
, NUMA_NO_NODE
,
1828 __builtin_return_address(0));
1830 EXPORT_SYMBOL(__vmalloc
);
1832 static inline void *__vmalloc_node_flags(unsigned long size
,
1833 int node
, gfp_t flags
)
1835 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
,
1836 node
, __builtin_return_address(0));
1840 void *__vmalloc_node_flags_caller(unsigned long size
, int node
, gfp_t flags
,
1843 return __vmalloc_node(size
, 1, flags
, PAGE_KERNEL
, node
, caller
);
1847 * vmalloc - allocate virtually contiguous memory
1848 * @size: allocation size
1849 * Allocate enough pages to cover @size from the page level
1850 * allocator and map them into contiguous kernel virtual space.
1852 * For tight control over page level allocator and protection flags
1853 * use __vmalloc() instead.
1855 void *vmalloc(unsigned long size
)
1857 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
1860 EXPORT_SYMBOL(vmalloc
);
1863 * vzalloc - allocate virtually contiguous memory with zero fill
1864 * @size: allocation size
1865 * Allocate enough pages to cover @size from the page level
1866 * allocator and map them into contiguous kernel virtual space.
1867 * The memory allocated is set to zero.
1869 * For tight control over page level allocator and protection flags
1870 * use __vmalloc() instead.
1872 void *vzalloc(unsigned long size
)
1874 return __vmalloc_node_flags(size
, NUMA_NO_NODE
,
1875 GFP_KERNEL
| __GFP_ZERO
);
1877 EXPORT_SYMBOL(vzalloc
);
1880 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1881 * @size: allocation size
1883 * The resulting memory area is zeroed so it can be mapped to userspace
1884 * without leaking data.
1886 void *vmalloc_user(unsigned long size
)
1888 struct vm_struct
*area
;
1891 ret
= __vmalloc_node(size
, SHMLBA
,
1892 GFP_KERNEL
| __GFP_ZERO
,
1893 PAGE_KERNEL
, NUMA_NO_NODE
,
1894 __builtin_return_address(0));
1896 area
= find_vm_area(ret
);
1897 area
->flags
|= VM_USERMAP
;
1901 EXPORT_SYMBOL(vmalloc_user
);
1904 * vmalloc_node - allocate memory on a specific node
1905 * @size: allocation size
1908 * Allocate enough pages to cover @size from the page level
1909 * allocator and map them into contiguous kernel virtual space.
1911 * For tight control over page level allocator and protection flags
1912 * use __vmalloc() instead.
1914 void *vmalloc_node(unsigned long size
, int node
)
1916 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL
,
1917 node
, __builtin_return_address(0));
1919 EXPORT_SYMBOL(vmalloc_node
);
1922 * vzalloc_node - allocate memory on a specific node with zero fill
1923 * @size: allocation size
1926 * Allocate enough pages to cover @size from the page level
1927 * allocator and map them into contiguous kernel virtual space.
1928 * The memory allocated is set to zero.
1930 * For tight control over page level allocator and protection flags
1931 * use __vmalloc_node() instead.
1933 void *vzalloc_node(unsigned long size
, int node
)
1935 return __vmalloc_node_flags(size
, node
,
1936 GFP_KERNEL
| __GFP_ZERO
);
1938 EXPORT_SYMBOL(vzalloc_node
);
1940 #ifndef PAGE_KERNEL_EXEC
1941 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1945 * vmalloc_exec - allocate virtually contiguous, executable memory
1946 * @size: allocation size
1948 * Kernel-internal function to allocate enough pages to cover @size
1949 * the page level allocator and map them into contiguous and
1950 * executable kernel virtual space.
1952 * For tight control over page level allocator and protection flags
1953 * use __vmalloc() instead.
1956 void *vmalloc_exec(unsigned long size
)
1958 return __vmalloc_node(size
, 1, GFP_KERNEL
, PAGE_KERNEL_EXEC
,
1959 NUMA_NO_NODE
, __builtin_return_address(0));
1962 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1963 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
1964 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1965 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
1968 * 64b systems should always have either DMA or DMA32 zones. For others
1969 * GFP_DMA32 should do the right thing and use the normal zone.
1971 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1975 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1976 * @size: allocation size
1978 * Allocate enough 32bit PA addressable pages to cover @size from the
1979 * page level allocator and map them into contiguous kernel virtual space.
1981 void *vmalloc_32(unsigned long size
)
1983 return __vmalloc_node(size
, 1, GFP_VMALLOC32
, PAGE_KERNEL
,
1984 NUMA_NO_NODE
, __builtin_return_address(0));
1986 EXPORT_SYMBOL(vmalloc_32
);
1989 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1990 * @size: allocation size
1992 * The resulting memory area is 32bit addressable and zeroed so it can be
1993 * mapped to userspace without leaking data.
1995 void *vmalloc_32_user(unsigned long size
)
1997 struct vm_struct
*area
;
2000 ret
= __vmalloc_node(size
, 1, GFP_VMALLOC32
| __GFP_ZERO
, PAGE_KERNEL
,
2001 NUMA_NO_NODE
, __builtin_return_address(0));
2003 area
= find_vm_area(ret
);
2004 area
->flags
|= VM_USERMAP
;
2008 EXPORT_SYMBOL(vmalloc_32_user
);
2011 * small helper routine , copy contents to buf from addr.
2012 * If the page is not present, fill zero.
2015 static int aligned_vread(char *buf
, char *addr
, unsigned long count
)
2021 unsigned long offset
, length
;
2023 offset
= offset_in_page(addr
);
2024 length
= PAGE_SIZE
- offset
;
2027 p
= vmalloc_to_page(addr
);
2029 * To do safe access to this _mapped_ area, we need
2030 * lock. But adding lock here means that we need to add
2031 * overhead of vmalloc()/vfree() calles for this _debug_
2032 * interface, rarely used. Instead of that, we'll use
2033 * kmap() and get small overhead in this access function.
2037 * we can expect USER0 is not used (see vread/vwrite's
2038 * function description)
2040 void *map
= kmap_atomic(p
);
2041 memcpy(buf
, map
+ offset
, length
);
2044 memset(buf
, 0, length
);
2054 static int aligned_vwrite(char *buf
, char *addr
, unsigned long count
)
2060 unsigned long offset
, length
;
2062 offset
= offset_in_page(addr
);
2063 length
= PAGE_SIZE
- offset
;
2066 p
= vmalloc_to_page(addr
);
2068 * To do safe access to this _mapped_ area, we need
2069 * lock. But adding lock here means that we need to add
2070 * overhead of vmalloc()/vfree() calles for this _debug_
2071 * interface, rarely used. Instead of that, we'll use
2072 * kmap() and get small overhead in this access function.
2076 * we can expect USER0 is not used (see vread/vwrite's
2077 * function description)
2079 void *map
= kmap_atomic(p
);
2080 memcpy(map
+ offset
, buf
, length
);
2092 * vread() - read vmalloc area in a safe way.
2093 * @buf: buffer for reading data
2094 * @addr: vm address.
2095 * @count: number of bytes to be read.
2097 * Returns # of bytes which addr and buf should be increased.
2098 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2099 * includes any intersect with alive vmalloc area.
2101 * This function checks that addr is a valid vmalloc'ed area, and
2102 * copy data from that area to a given buffer. If the given memory range
2103 * of [addr...addr+count) includes some valid address, data is copied to
2104 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2105 * IOREMAP area is treated as memory hole and no copy is done.
2107 * If [addr...addr+count) doesn't includes any intersects with alive
2108 * vm_struct area, returns 0. @buf should be kernel's buffer.
2110 * Note: In usual ops, vread() is never necessary because the caller
2111 * should know vmalloc() area is valid and can use memcpy().
2112 * This is for routines which have to access vmalloc area without
2113 * any informaion, as /dev/kmem.
2117 long vread(char *buf
, char *addr
, unsigned long count
)
2119 struct vmap_area
*va
;
2120 struct vm_struct
*vm
;
2121 char *vaddr
, *buf_start
= buf
;
2122 unsigned long buflen
= count
;
2125 /* Don't allow overflow */
2126 if ((unsigned long) addr
+ count
< count
)
2127 count
= -(unsigned long) addr
;
2129 spin_lock(&vmap_area_lock
);
2130 list_for_each_entry(va
, &vmap_area_list
, list
) {
2134 if (!(va
->flags
& VM_VM_AREA
))
2138 vaddr
= (char *) vm
->addr
;
2139 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2141 while (addr
< vaddr
) {
2149 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2152 if (!(vm
->flags
& VM_IOREMAP
))
2153 aligned_vread(buf
, addr
, n
);
2154 else /* IOREMAP area is treated as memory hole */
2161 spin_unlock(&vmap_area_lock
);
2163 if (buf
== buf_start
)
2165 /* zero-fill memory holes */
2166 if (buf
!= buf_start
+ buflen
)
2167 memset(buf
, 0, buflen
- (buf
- buf_start
));
2173 * vwrite() - write vmalloc area in a safe way.
2174 * @buf: buffer for source data
2175 * @addr: vm address.
2176 * @count: number of bytes to be read.
2178 * Returns # of bytes which addr and buf should be incresed.
2179 * (same number to @count).
2180 * If [addr...addr+count) doesn't includes any intersect with valid
2181 * vmalloc area, returns 0.
2183 * This function checks that addr is a valid vmalloc'ed area, and
2184 * copy data from a buffer to the given addr. If specified range of
2185 * [addr...addr+count) includes some valid address, data is copied from
2186 * proper area of @buf. If there are memory holes, no copy to hole.
2187 * IOREMAP area is treated as memory hole and no copy is done.
2189 * If [addr...addr+count) doesn't includes any intersects with alive
2190 * vm_struct area, returns 0. @buf should be kernel's buffer.
2192 * Note: In usual ops, vwrite() is never necessary because the caller
2193 * should know vmalloc() area is valid and can use memcpy().
2194 * This is for routines which have to access vmalloc area without
2195 * any informaion, as /dev/kmem.
2198 long vwrite(char *buf
, char *addr
, unsigned long count
)
2200 struct vmap_area
*va
;
2201 struct vm_struct
*vm
;
2203 unsigned long n
, buflen
;
2206 /* Don't allow overflow */
2207 if ((unsigned long) addr
+ count
< count
)
2208 count
= -(unsigned long) addr
;
2211 spin_lock(&vmap_area_lock
);
2212 list_for_each_entry(va
, &vmap_area_list
, list
) {
2216 if (!(va
->flags
& VM_VM_AREA
))
2220 vaddr
= (char *) vm
->addr
;
2221 if (addr
>= vaddr
+ get_vm_area_size(vm
))
2223 while (addr
< vaddr
) {
2230 n
= vaddr
+ get_vm_area_size(vm
) - addr
;
2233 if (!(vm
->flags
& VM_IOREMAP
)) {
2234 aligned_vwrite(buf
, addr
, n
);
2242 spin_unlock(&vmap_area_lock
);
2249 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2250 * @vma: vma to cover
2251 * @uaddr: target user address to start at
2252 * @kaddr: virtual address of vmalloc kernel memory
2253 * @size: size of map area
2255 * Returns: 0 for success, -Exxx on failure
2257 * This function checks that @kaddr is a valid vmalloc'ed area,
2258 * and that it is big enough to cover the range starting at
2259 * @uaddr in @vma. Will return failure if that criteria isn't
2262 * Similar to remap_pfn_range() (see mm/memory.c)
2264 int remap_vmalloc_range_partial(struct vm_area_struct
*vma
, unsigned long uaddr
,
2265 void *kaddr
, unsigned long size
)
2267 struct vm_struct
*area
;
2269 size
= PAGE_ALIGN(size
);
2271 if (!PAGE_ALIGNED(uaddr
) || !PAGE_ALIGNED(kaddr
))
2274 area
= find_vm_area(kaddr
);
2278 if (!(area
->flags
& VM_USERMAP
))
2281 if (kaddr
+ size
> area
->addr
+ get_vm_area_size(area
))
2285 struct page
*page
= vmalloc_to_page(kaddr
);
2288 ret
= vm_insert_page(vma
, uaddr
, page
);
2297 vma
->vm_flags
|= VM_DONTEXPAND
| VM_DONTDUMP
;
2301 EXPORT_SYMBOL(remap_vmalloc_range_partial
);
2304 * remap_vmalloc_range - map vmalloc pages to userspace
2305 * @vma: vma to cover (map full range of vma)
2306 * @addr: vmalloc memory
2307 * @pgoff: number of pages into addr before first page to map
2309 * Returns: 0 for success, -Exxx on failure
2311 * This function checks that addr is a valid vmalloc'ed area, and
2312 * that it is big enough to cover the vma. Will return failure if
2313 * that criteria isn't met.
2315 * Similar to remap_pfn_range() (see mm/memory.c)
2317 int remap_vmalloc_range(struct vm_area_struct
*vma
, void *addr
,
2318 unsigned long pgoff
)
2320 return remap_vmalloc_range_partial(vma
, vma
->vm_start
,
2321 addr
+ (pgoff
<< PAGE_SHIFT
),
2322 vma
->vm_end
- vma
->vm_start
);
2324 EXPORT_SYMBOL(remap_vmalloc_range
);
2327 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2330 * The purpose of this function is to make sure the vmalloc area
2331 * mappings are identical in all page-tables in the system.
2333 void __weak
vmalloc_sync_all(void)
2338 static int f(pte_t
*pte
, pgtable_t table
, unsigned long addr
, void *data
)
2350 * alloc_vm_area - allocate a range of kernel address space
2351 * @size: size of the area
2352 * @ptes: returns the PTEs for the address space
2354 * Returns: NULL on failure, vm_struct on success
2356 * This function reserves a range of kernel address space, and
2357 * allocates pagetables to map that range. No actual mappings
2360 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2361 * allocated for the VM area are returned.
2363 struct vm_struct
*alloc_vm_area(size_t size
, pte_t
**ptes
)
2365 struct vm_struct
*area
;
2367 area
= get_vm_area_caller(size
, VM_IOREMAP
,
2368 __builtin_return_address(0));
2373 * This ensures that page tables are constructed for this region
2374 * of kernel virtual address space and mapped into init_mm.
2376 if (apply_to_page_range(&init_mm
, (unsigned long)area
->addr
,
2377 size
, f
, ptes
? &ptes
: NULL
)) {
2384 EXPORT_SYMBOL_GPL(alloc_vm_area
);
2386 void free_vm_area(struct vm_struct
*area
)
2388 struct vm_struct
*ret
;
2389 ret
= remove_vm_area(area
->addr
);
2390 BUG_ON(ret
!= area
);
2393 EXPORT_SYMBOL_GPL(free_vm_area
);
2396 static struct vmap_area
*node_to_va(struct rb_node
*n
)
2398 return rb_entry_safe(n
, struct vmap_area
, rb_node
);
2402 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2403 * @end: target address
2404 * @pnext: out arg for the next vmap_area
2405 * @pprev: out arg for the previous vmap_area
2407 * Returns: %true if either or both of next and prev are found,
2408 * %false if no vmap_area exists
2410 * Find vmap_areas end addresses of which enclose @end. ie. if not
2411 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2413 static bool pvm_find_next_prev(unsigned long end
,
2414 struct vmap_area
**pnext
,
2415 struct vmap_area
**pprev
)
2417 struct rb_node
*n
= vmap_area_root
.rb_node
;
2418 struct vmap_area
*va
= NULL
;
2421 va
= rb_entry(n
, struct vmap_area
, rb_node
);
2422 if (end
< va
->va_end
)
2424 else if (end
> va
->va_end
)
2433 if (va
->va_end
> end
) {
2435 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2438 *pnext
= node_to_va(rb_next(&(*pprev
)->rb_node
));
2444 * pvm_determine_end - find the highest aligned address between two vmap_areas
2445 * @pnext: in/out arg for the next vmap_area
2446 * @pprev: in/out arg for the previous vmap_area
2449 * Returns: determined end address
2451 * Find the highest aligned address between *@pnext and *@pprev below
2452 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2453 * down address is between the end addresses of the two vmap_areas.
2455 * Please note that the address returned by this function may fall
2456 * inside *@pnext vmap_area. The caller is responsible for checking
2459 static unsigned long pvm_determine_end(struct vmap_area
**pnext
,
2460 struct vmap_area
**pprev
,
2461 unsigned long align
)
2463 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2467 addr
= min((*pnext
)->va_start
& ~(align
- 1), vmalloc_end
);
2471 while (*pprev
&& (*pprev
)->va_end
> addr
) {
2473 *pprev
= node_to_va(rb_prev(&(*pnext
)->rb_node
));
2480 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2481 * @offsets: array containing offset of each area
2482 * @sizes: array containing size of each area
2483 * @nr_vms: the number of areas to allocate
2484 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2486 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2487 * vm_structs on success, %NULL on failure
2489 * Percpu allocator wants to use congruent vm areas so that it can
2490 * maintain the offsets among percpu areas. This function allocates
2491 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2492 * be scattered pretty far, distance between two areas easily going up
2493 * to gigabytes. To avoid interacting with regular vmallocs, these
2494 * areas are allocated from top.
2496 * Despite its complicated look, this allocator is rather simple. It
2497 * does everything top-down and scans areas from the end looking for
2498 * matching slot. While scanning, if any of the areas overlaps with
2499 * existing vmap_area, the base address is pulled down to fit the
2500 * area. Scanning is repeated till all the areas fit and then all
2501 * necessary data structures are inserted and the result is returned.
2503 struct vm_struct
**pcpu_get_vm_areas(const unsigned long *offsets
,
2504 const size_t *sizes
, int nr_vms
,
2507 const unsigned long vmalloc_start
= ALIGN(VMALLOC_START
, align
);
2508 const unsigned long vmalloc_end
= VMALLOC_END
& ~(align
- 1);
2509 struct vmap_area
**vas
, *prev
, *next
;
2510 struct vm_struct
**vms
;
2511 int area
, area2
, last_area
, term_area
;
2512 unsigned long base
, start
, end
, last_end
;
2513 bool purged
= false;
2515 /* verify parameters and allocate data structures */
2516 BUG_ON(offset_in_page(align
) || !is_power_of_2(align
));
2517 for (last_area
= 0, area
= 0; area
< nr_vms
; area
++) {
2518 start
= offsets
[area
];
2519 end
= start
+ sizes
[area
];
2521 /* is everything aligned properly? */
2522 BUG_ON(!IS_ALIGNED(offsets
[area
], align
));
2523 BUG_ON(!IS_ALIGNED(sizes
[area
], align
));
2525 /* detect the area with the highest address */
2526 if (start
> offsets
[last_area
])
2529 for (area2
= area
+ 1; area2
< nr_vms
; area2
++) {
2530 unsigned long start2
= offsets
[area2
];
2531 unsigned long end2
= start2
+ sizes
[area2
];
2533 BUG_ON(start2
< end
&& start
< end2
);
2536 last_end
= offsets
[last_area
] + sizes
[last_area
];
2538 if (vmalloc_end
- vmalloc_start
< last_end
) {
2543 vms
= kcalloc(nr_vms
, sizeof(vms
[0]), GFP_KERNEL
);
2544 vas
= kcalloc(nr_vms
, sizeof(vas
[0]), GFP_KERNEL
);
2548 for (area
= 0; area
< nr_vms
; area
++) {
2549 vas
[area
] = kzalloc(sizeof(struct vmap_area
), GFP_KERNEL
);
2550 vms
[area
] = kzalloc(sizeof(struct vm_struct
), GFP_KERNEL
);
2551 if (!vas
[area
] || !vms
[area
])
2555 spin_lock(&vmap_area_lock
);
2557 /* start scanning - we scan from the top, begin with the last area */
2558 area
= term_area
= last_area
;
2559 start
= offsets
[area
];
2560 end
= start
+ sizes
[area
];
2562 if (!pvm_find_next_prev(vmap_area_pcpu_hole
, &next
, &prev
)) {
2563 base
= vmalloc_end
- last_end
;
2566 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2569 BUG_ON(next
&& next
->va_end
<= base
+ end
);
2570 BUG_ON(prev
&& prev
->va_end
> base
+ end
);
2573 * base might have underflowed, add last_end before
2576 if (base
+ last_end
< vmalloc_start
+ last_end
) {
2577 spin_unlock(&vmap_area_lock
);
2579 purge_vmap_area_lazy();
2587 * If next overlaps, move base downwards so that it's
2588 * right below next and then recheck.
2590 if (next
&& next
->va_start
< base
+ end
) {
2591 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2597 * If prev overlaps, shift down next and prev and move
2598 * base so that it's right below new next and then
2601 if (prev
&& prev
->va_end
> base
+ start
) {
2603 prev
= node_to_va(rb_prev(&next
->rb_node
));
2604 base
= pvm_determine_end(&next
, &prev
, align
) - end
;
2610 * This area fits, move on to the previous one. If
2611 * the previous one is the terminal one, we're done.
2613 area
= (area
+ nr_vms
- 1) % nr_vms
;
2614 if (area
== term_area
)
2616 start
= offsets
[area
];
2617 end
= start
+ sizes
[area
];
2618 pvm_find_next_prev(base
+ end
, &next
, &prev
);
2621 /* we've found a fitting base, insert all va's */
2622 for (area
= 0; area
< nr_vms
; area
++) {
2623 struct vmap_area
*va
= vas
[area
];
2625 va
->va_start
= base
+ offsets
[area
];
2626 va
->va_end
= va
->va_start
+ sizes
[area
];
2627 __insert_vmap_area(va
);
2630 vmap_area_pcpu_hole
= base
+ offsets
[last_area
];
2632 spin_unlock(&vmap_area_lock
);
2634 /* insert all vm's */
2635 for (area
= 0; area
< nr_vms
; area
++)
2636 setup_vmalloc_vm(vms
[area
], vas
[area
], VM_ALLOC
,
2643 for (area
= 0; area
< nr_vms
; area
++) {
2654 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2655 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2656 * @nr_vms: the number of allocated areas
2658 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2660 void pcpu_free_vm_areas(struct vm_struct
**vms
, int nr_vms
)
2664 for (i
= 0; i
< nr_vms
; i
++)
2665 free_vm_area(vms
[i
]);
2668 #endif /* CONFIG_SMP */
2670 #ifdef CONFIG_PROC_FS
2671 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
2672 __acquires(&vmap_area_lock
)
2674 spin_lock(&vmap_area_lock
);
2675 return seq_list_start(&vmap_area_list
, *pos
);
2678 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
2680 return seq_list_next(p
, &vmap_area_list
, pos
);
2683 static void s_stop(struct seq_file
*m
, void *p
)
2684 __releases(&vmap_area_lock
)
2686 spin_unlock(&vmap_area_lock
);
2689 static void show_numa_info(struct seq_file
*m
, struct vm_struct
*v
)
2691 if (IS_ENABLED(CONFIG_NUMA
)) {
2692 unsigned int nr
, *counters
= m
->private;
2697 if (v
->flags
& VM_UNINITIALIZED
)
2699 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2702 memset(counters
, 0, nr_node_ids
* sizeof(unsigned int));
2704 for (nr
= 0; nr
< v
->nr_pages
; nr
++)
2705 counters
[page_to_nid(v
->pages
[nr
])]++;
2707 for_each_node_state(nr
, N_HIGH_MEMORY
)
2709 seq_printf(m
, " N%u=%u", nr
, counters
[nr
]);
2713 static int s_show(struct seq_file
*m
, void *p
)
2715 struct vmap_area
*va
;
2716 struct vm_struct
*v
;
2718 va
= list_entry(p
, struct vmap_area
, list
);
2721 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2722 * behalf of vmap area is being tear down or vm_map_ram allocation.
2724 if (!(va
->flags
& VM_VM_AREA
)) {
2725 seq_printf(m
, "0x%pK-0x%pK %7ld %s\n",
2726 (void *)va
->va_start
, (void *)va
->va_end
,
2727 va
->va_end
- va
->va_start
,
2728 va
->flags
& VM_LAZY_FREE
? "unpurged vm_area" : "vm_map_ram");
2735 seq_printf(m
, "0x%pK-0x%pK %7ld",
2736 v
->addr
, v
->addr
+ v
->size
, v
->size
);
2739 seq_printf(m
, " %pS", v
->caller
);
2742 seq_printf(m
, " pages=%d", v
->nr_pages
);
2745 seq_printf(m
, " phys=%pa", &v
->phys_addr
);
2747 if (v
->flags
& VM_IOREMAP
)
2748 seq_puts(m
, " ioremap");
2750 if (v
->flags
& VM_ALLOC
)
2751 seq_puts(m
, " vmalloc");
2753 if (v
->flags
& VM_MAP
)
2754 seq_puts(m
, " vmap");
2756 if (v
->flags
& VM_USERMAP
)
2757 seq_puts(m
, " user");
2759 if (is_vmalloc_addr(v
->pages
))
2760 seq_puts(m
, " vpages");
2762 show_numa_info(m
, v
);
2767 static const struct seq_operations vmalloc_op
= {
2774 static int vmalloc_open(struct inode
*inode
, struct file
*file
)
2776 if (IS_ENABLED(CONFIG_NUMA
))
2777 return seq_open_private(file
, &vmalloc_op
,
2778 nr_node_ids
* sizeof(unsigned int));
2780 return seq_open(file
, &vmalloc_op
);
2783 static const struct file_operations proc_vmalloc_operations
= {
2784 .open
= vmalloc_open
,
2786 .llseek
= seq_lseek
,
2787 .release
= seq_release_private
,
2790 static int __init
proc_vmalloc_init(void)
2792 proc_create("vmallocinfo", S_IRUSR
, NULL
, &proc_vmalloc_operations
);
2795 module_init(proc_vmalloc_init
);