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>
34 /*** Page table manipulation functions ***/
36 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
40 pte = pte_offset_kernel(pmd, addr);
42 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
43 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
44 } while (pte++, addr += PAGE_SIZE, addr != end);
47 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
52 pmd = pmd_offset(pud, addr);
54 next = pmd_addr_end(addr, end);
55 if (pmd_none_or_clear_bad(pmd))
57 vunmap_pte_range(pmd, addr, next);
58 } while (pmd++, addr = next, addr != end);
61 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
66 pud = pud_offset(pgd, addr);
68 next = pud_addr_end(addr, end);
69 if (pud_none_or_clear_bad(pud))
71 vunmap_pmd_range(pud, addr, next);
72 } while (pud++, addr = next, addr != end);
75 static void vunmap_page_range(unsigned long addr, unsigned long end)
81 pgd = pgd_offset_k(addr);
83 next = pgd_addr_end(addr, end);
84 if (pgd_none_or_clear_bad(pgd))
86 vunmap_pud_range(pgd, addr, next);
87 } while (pgd++, addr = next, addr != end);
90 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
91 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
96 * nr is a running index into the array which helps higher level
97 * callers keep track of where we're up to.
100 pte = pte_alloc_kernel(pmd, addr);
104 struct page *page = pages[*nr];
106 if (WARN_ON(!pte_none(*pte)))
110 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
112 } while (pte++, addr += PAGE_SIZE, addr != end);
116 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
117 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
122 pmd = pmd_alloc(&init_mm, pud, addr);
126 next = pmd_addr_end(addr, end);
127 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
129 } while (pmd++, addr = next, addr != end);
133 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
134 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
139 pud = pud_alloc(&init_mm, pgd, addr);
143 next = pud_addr_end(addr, end);
144 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
146 } while (pud++, addr = next, addr != end);
151 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
152 * will have pfns corresponding to the "pages" array.
154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
156 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
157 pgprot_t prot, struct page **pages)
161 unsigned long addr = start;
166 pgd = pgd_offset_k(addr);
168 next = pgd_addr_end(addr, end);
169 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
172 } while (pgd++, addr = next, addr != end);
177 static int vmap_page_range(unsigned long start, unsigned long end,
178 pgprot_t prot, struct page **pages)
182 ret = vmap_page_range_noflush(start, end, prot, pages);
183 flush_cache_vmap(start, end);
187 int is_vmalloc_or_module_addr(const void *x)
190 * ARM, x86-64 and sparc64 put modules in a special place,
191 * and fall back on vmalloc() if that fails. Others
192 * just put it in the vmalloc space.
194 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
195 unsigned long addr = (unsigned long)x;
196 if (addr >= MODULES_VADDR && addr < MODULES_END)
199 return is_vmalloc_addr(x);
203 * Walk a vmap address to the struct page it maps.
205 struct page *vmalloc_to_page(const void *vmalloc_addr)
207 unsigned long addr = (unsigned long) vmalloc_addr;
208 struct page *page = NULL;
209 pgd_t *pgd = pgd_offset_k(addr);
212 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
213 * architectures that do not vmalloc module space
215 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
217 if (!pgd_none(*pgd)) {
218 pud_t *pud = pud_offset(pgd, addr);
219 if (!pud_none(*pud)) {
220 pmd_t *pmd = pmd_offset(pud, addr);
221 if (!pmd_none(*pmd)) {
224 ptep = pte_offset_map(pmd, addr);
226 if (pte_present(pte))
227 page = pte_page(pte);
234 EXPORT_SYMBOL(vmalloc_to_page);
237 * Map a vmalloc()-space virtual address to the physical page frame number.
239 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
241 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
243 EXPORT_SYMBOL(vmalloc_to_pfn);
246 /*** Global kva allocator ***/
248 #define VM_LAZY_FREE 0x01
249 #define VM_LAZY_FREEING 0x02
250 #define VM_VM_AREA 0x04
253 unsigned long va_start;
254 unsigned long va_end;
256 struct rb_node rb_node; /* address sorted rbtree */
257 struct list_head list; /* address sorted list */
258 struct list_head purge_list; /* "lazy purge" list */
260 struct rcu_head rcu_head;
263 static DEFINE_SPINLOCK(vmap_area_lock);
264 static LIST_HEAD(vmap_area_list);
265 static struct rb_root vmap_area_root = RB_ROOT;
267 /* The vmap cache globals are protected by vmap_area_lock */
268 static struct rb_node *free_vmap_cache;
269 static unsigned long cached_hole_size;
270 static unsigned long cached_vstart;
271 static unsigned long cached_align;
273 static unsigned long vmap_area_pcpu_hole;
275 static struct vmap_area *__find_vmap_area(unsigned long addr)
277 struct rb_node *n = vmap_area_root.rb_node;
280 struct vmap_area *va;
282 va = rb_entry(n, struct vmap_area, rb_node);
283 if (addr < va->va_start)
285 else if (addr > va->va_start)
294 static void __insert_vmap_area(struct vmap_area *va)
296 struct rb_node **p = &vmap_area_root.rb_node;
297 struct rb_node *parent = NULL;
301 struct vmap_area *tmp_va;
304 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
305 if (va->va_start < tmp_va->va_end)
307 else if (va->va_end > tmp_va->va_start)
313 rb_link_node(&va->rb_node, parent, p);
314 rb_insert_color(&va->rb_node, &vmap_area_root);
316 /* address-sort this list so it is usable like the vmlist */
317 tmp = rb_prev(&va->rb_node);
319 struct vmap_area *prev;
320 prev = rb_entry(tmp, struct vmap_area, rb_node);
321 list_add_rcu(&va->list, &prev->list);
323 list_add_rcu(&va->list, &vmap_area_list);
326 static void purge_vmap_area_lazy(void);
329 * Allocate a region of KVA of the specified size and alignment, within the
332 static struct vmap_area *alloc_vmap_area(unsigned long size,
334 unsigned long vstart, unsigned long vend,
335 int node, gfp_t gfp_mask)
337 struct vmap_area *va;
341 struct vmap_area *first;
344 BUG_ON(size & ~PAGE_MASK);
345 BUG_ON(!is_power_of_2(align));
347 va = kmalloc_node(sizeof(struct vmap_area),
348 gfp_mask & GFP_RECLAIM_MASK, node);
350 return ERR_PTR(-ENOMEM);
353 spin_lock(&vmap_area_lock);
355 * Invalidate cache if we have more permissive parameters.
356 * cached_hole_size notes the largest hole noticed _below_
357 * the vmap_area cached in free_vmap_cache: if size fits
358 * into that hole, we want to scan from vstart to reuse
359 * the hole instead of allocating above free_vmap_cache.
360 * Note that __free_vmap_area may update free_vmap_cache
361 * without updating cached_hole_size or cached_align.
363 if (!free_vmap_cache ||
364 size < cached_hole_size ||
365 vstart < cached_vstart ||
366 align < cached_align) {
368 cached_hole_size = 0;
369 free_vmap_cache = NULL;
371 /* record if we encounter less permissive parameters */
372 cached_vstart = vstart;
373 cached_align = align;
375 /* find starting point for our search */
376 if (free_vmap_cache) {
377 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
378 addr = ALIGN(first->va_end, align);
381 if (addr + size - 1 < addr)
385 addr = ALIGN(vstart, align);
386 if (addr + size - 1 < addr)
389 n = vmap_area_root.rb_node;
393 struct vmap_area *tmp;
394 tmp = rb_entry(n, struct vmap_area, rb_node);
395 if (tmp->va_end >= addr) {
397 if (tmp->va_start <= addr)
408 /* from the starting point, walk areas until a suitable hole is found */
409 while (addr + size > first->va_start && addr + size <= vend) {
410 if (addr + cached_hole_size < first->va_start)
411 cached_hole_size = first->va_start - addr;
412 addr = ALIGN(first->va_end, align);
413 if (addr + size - 1 < addr)
416 n = rb_next(&first->rb_node);
418 first = rb_entry(n, struct vmap_area, rb_node);
424 if (addr + size > vend)
428 va->va_end = addr + size;
430 __insert_vmap_area(va);
431 free_vmap_cache = &va->rb_node;
432 spin_unlock(&vmap_area_lock);
434 BUG_ON(va->va_start & (align-1));
435 BUG_ON(va->va_start < vstart);
436 BUG_ON(va->va_end > vend);
441 spin_unlock(&vmap_area_lock);
443 purge_vmap_area_lazy();
447 if (printk_ratelimit())
449 "vmap allocation for size %lu failed: "
450 "use vmalloc=<size> to increase size.\n", size);
452 return ERR_PTR(-EBUSY);
455 static void rcu_free_va(struct rcu_head *head)
457 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
462 static void __free_vmap_area(struct vmap_area *va)
464 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
466 if (free_vmap_cache) {
467 if (va->va_end < cached_vstart) {
468 free_vmap_cache = NULL;
470 struct vmap_area *cache;
471 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
472 if (va->va_start <= cache->va_start) {
473 free_vmap_cache = rb_prev(&va->rb_node);
475 * We don't try to update cached_hole_size or
476 * cached_align, but it won't go very wrong.
481 rb_erase(&va->rb_node, &vmap_area_root);
482 RB_CLEAR_NODE(&va->rb_node);
483 list_del_rcu(&va->list);
486 * Track the highest possible candidate for pcpu area
487 * allocation. Areas outside of vmalloc area can be returned
488 * here too, consider only end addresses which fall inside
489 * vmalloc area proper.
491 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
492 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
494 call_rcu(&va->rcu_head, rcu_free_va);
498 * Free a region of KVA allocated by alloc_vmap_area
500 static void free_vmap_area(struct vmap_area *va)
502 spin_lock(&vmap_area_lock);
503 __free_vmap_area(va);
504 spin_unlock(&vmap_area_lock);
508 * Clear the pagetable entries of a given vmap_area
510 static void unmap_vmap_area(struct vmap_area *va)
512 vunmap_page_range(va->va_start, va->va_end);
515 static void vmap_debug_free_range(unsigned long start, unsigned long end)
518 * Unmap page tables and force a TLB flush immediately if
519 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
520 * bugs similarly to those in linear kernel virtual address
521 * space after a page has been freed.
523 * All the lazy freeing logic is still retained, in order to
524 * minimise intrusiveness of this debugging feature.
526 * This is going to be *slow* (linear kernel virtual address
527 * debugging doesn't do a broadcast TLB flush so it is a lot
530 #ifdef CONFIG_DEBUG_PAGEALLOC
531 vunmap_page_range(start, end);
532 flush_tlb_kernel_range(start, end);
537 * lazy_max_pages is the maximum amount of virtual address space we gather up
538 * before attempting to purge with a TLB flush.
540 * There is a tradeoff here: a larger number will cover more kernel page tables
541 * and take slightly longer to purge, but it will linearly reduce the number of
542 * global TLB flushes that must be performed. It would seem natural to scale
543 * this number up linearly with the number of CPUs (because vmapping activity
544 * could also scale linearly with the number of CPUs), however it is likely
545 * that in practice, workloads might be constrained in other ways that mean
546 * vmap activity will not scale linearly with CPUs. Also, I want to be
547 * conservative and not introduce a big latency on huge systems, so go with
548 * a less aggressive log scale. It will still be an improvement over the old
549 * code, and it will be simple to change the scale factor if we find that it
550 * becomes a problem on bigger systems.
552 static unsigned long lazy_max_pages(void)
556 log = fls(num_online_cpus());
558 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
561 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
563 /* for per-CPU blocks */
564 static void purge_fragmented_blocks_allcpus(void);
567 * called before a call to iounmap() if the caller wants vm_area_struct's
570 void set_iounmap_nonlazy(void)
572 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
576 * Purges all lazily-freed vmap areas.
578 * If sync is 0 then don't purge if there is already a purge in progress.
579 * If force_flush is 1, then flush kernel TLBs between *start and *end even
580 * if we found no lazy vmap areas to unmap (callers can use this to optimise
581 * their own TLB flushing).
582 * Returns with *start = min(*start, lowest purged address)
583 * *end = max(*end, highest purged address)
585 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
586 int sync, int force_flush)
588 static DEFINE_SPINLOCK(purge_lock);
590 struct vmap_area *va;
591 struct vmap_area *n_va;
595 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
596 * should not expect such behaviour. This just simplifies locking for
597 * the case that isn't actually used at the moment anyway.
599 if (!sync && !force_flush) {
600 if (!spin_trylock(&purge_lock))
603 spin_lock(&purge_lock);
606 purge_fragmented_blocks_allcpus();
609 list_for_each_entry_rcu(va, &vmap_area_list, list) {
610 if (va->flags & VM_LAZY_FREE) {
611 if (va->va_start < *start)
612 *start = va->va_start;
613 if (va->va_end > *end)
615 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
616 list_add_tail(&va->purge_list, &valist);
617 va->flags |= VM_LAZY_FREEING;
618 va->flags &= ~VM_LAZY_FREE;
624 atomic_sub(nr, &vmap_lazy_nr);
626 if (nr || force_flush)
627 flush_tlb_kernel_range(*start, *end);
630 spin_lock(&vmap_area_lock);
631 list_for_each_entry_safe(va, n_va, &valist, purge_list)
632 __free_vmap_area(va);
633 spin_unlock(&vmap_area_lock);
635 spin_unlock(&purge_lock);
639 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
640 * is already purging.
642 static void try_purge_vmap_area_lazy(void)
644 unsigned long start = ULONG_MAX, end = 0;
646 __purge_vmap_area_lazy(&start, &end, 0, 0);
650 * Kick off a purge of the outstanding lazy areas.
652 static void purge_vmap_area_lazy(void)
654 unsigned long start = ULONG_MAX, end = 0;
656 __purge_vmap_area_lazy(&start, &end, 1, 0);
660 * Free a vmap area, caller ensuring that the area has been unmapped
661 * and flush_cache_vunmap had been called for the correct range
664 static void free_vmap_area_noflush(struct vmap_area *va)
666 va->flags |= VM_LAZY_FREE;
667 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
668 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
669 try_purge_vmap_area_lazy();
673 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
674 * called for the correct range previously.
676 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
679 free_vmap_area_noflush(va);
683 * Free and unmap a vmap area
685 static void free_unmap_vmap_area(struct vmap_area *va)
687 flush_cache_vunmap(va->va_start, va->va_end);
688 free_unmap_vmap_area_noflush(va);
691 static struct vmap_area *find_vmap_area(unsigned long addr)
693 struct vmap_area *va;
695 spin_lock(&vmap_area_lock);
696 va = __find_vmap_area(addr);
697 spin_unlock(&vmap_area_lock);
702 static void free_unmap_vmap_area_addr(unsigned long addr)
704 struct vmap_area *va;
706 va = find_vmap_area(addr);
708 free_unmap_vmap_area(va);
712 /*** Per cpu kva allocator ***/
715 * vmap space is limited especially on 32 bit architectures. Ensure there is
716 * room for at least 16 percpu vmap blocks per CPU.
719 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
720 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
721 * instead (we just need a rough idea)
723 #if BITS_PER_LONG == 32
724 #define VMALLOC_SPACE (128UL*1024*1024)
726 #define VMALLOC_SPACE (128UL*1024*1024*1024)
729 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
730 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
731 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
732 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
733 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
734 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
735 #define VMAP_BBMAP_BITS \
736 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
737 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
738 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
740 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
742 static bool vmap_initialized __read_mostly = false;
744 struct vmap_block_queue {
746 struct list_head free;
751 struct vmap_area *va;
752 struct vmap_block_queue *vbq;
753 unsigned long free, dirty;
754 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
755 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
756 struct list_head free_list;
757 struct rcu_head rcu_head;
758 struct list_head purge;
761 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
762 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
765 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
766 * in the free path. Could get rid of this if we change the API to return a
767 * "cookie" from alloc, to be passed to free. But no big deal yet.
769 static DEFINE_SPINLOCK(vmap_block_tree_lock);
770 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
773 * We should probably have a fallback mechanism to allocate virtual memory
774 * out of partially filled vmap blocks. However vmap block sizing should be
775 * fairly reasonable according to the vmalloc size, so it shouldn't be a
779 static unsigned long addr_to_vb_idx(unsigned long addr)
781 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
782 addr /= VMAP_BLOCK_SIZE;
786 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
788 struct vmap_block_queue *vbq;
789 struct vmap_block *vb;
790 struct vmap_area *va;
791 unsigned long vb_idx;
794 node = numa_node_id();
796 vb = kmalloc_node(sizeof(struct vmap_block),
797 gfp_mask & GFP_RECLAIM_MASK, node);
799 return ERR_PTR(-ENOMEM);
801 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
802 VMALLOC_START, VMALLOC_END,
809 err = radix_tree_preload(gfp_mask);
816 spin_lock_init(&vb->lock);
818 vb->free = VMAP_BBMAP_BITS;
820 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
821 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
822 INIT_LIST_HEAD(&vb->free_list);
824 vb_idx = addr_to_vb_idx(va->va_start);
825 spin_lock(&vmap_block_tree_lock);
826 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
827 spin_unlock(&vmap_block_tree_lock);
829 radix_tree_preload_end();
831 vbq = &get_cpu_var(vmap_block_queue);
833 spin_lock(&vbq->lock);
834 list_add_rcu(&vb->free_list, &vbq->free);
835 spin_unlock(&vbq->lock);
836 put_cpu_var(vmap_block_queue);
841 static void rcu_free_vb(struct rcu_head *head)
843 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
848 static void free_vmap_block(struct vmap_block *vb)
850 struct vmap_block *tmp;
851 unsigned long vb_idx;
853 vb_idx = addr_to_vb_idx(vb->va->va_start);
854 spin_lock(&vmap_block_tree_lock);
855 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
856 spin_unlock(&vmap_block_tree_lock);
859 free_vmap_area_noflush(vb->va);
860 call_rcu(&vb->rcu_head, rcu_free_vb);
863 static void purge_fragmented_blocks(int cpu)
866 struct vmap_block *vb;
867 struct vmap_block *n_vb;
868 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
871 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
873 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
876 spin_lock(&vb->lock);
877 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
878 vb->free = 0; /* prevent further allocs after releasing lock */
879 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
880 bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
881 bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
882 spin_lock(&vbq->lock);
883 list_del_rcu(&vb->free_list);
884 spin_unlock(&vbq->lock);
885 spin_unlock(&vb->lock);
886 list_add_tail(&vb->purge, &purge);
888 spin_unlock(&vb->lock);
892 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
893 list_del(&vb->purge);
898 static void purge_fragmented_blocks_thiscpu(void)
900 purge_fragmented_blocks(smp_processor_id());
903 static void purge_fragmented_blocks_allcpus(void)
907 for_each_possible_cpu(cpu)
908 purge_fragmented_blocks(cpu);
911 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
913 struct vmap_block_queue *vbq;
914 struct vmap_block *vb;
915 unsigned long addr = 0;
919 BUG_ON(size & ~PAGE_MASK);
920 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
921 order = get_order(size);
925 vbq = &get_cpu_var(vmap_block_queue);
926 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
929 spin_lock(&vb->lock);
930 if (vb->free < 1UL << order)
933 i = bitmap_find_free_region(vb->alloc_map,
934 VMAP_BBMAP_BITS, order);
937 if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
938 /* fragmented and no outstanding allocations */
939 BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
944 addr = vb->va->va_start + (i << PAGE_SHIFT);
945 BUG_ON(addr_to_vb_idx(addr) !=
946 addr_to_vb_idx(vb->va->va_start));
947 vb->free -= 1UL << order;
949 spin_lock(&vbq->lock);
950 list_del_rcu(&vb->free_list);
951 spin_unlock(&vbq->lock);
953 spin_unlock(&vb->lock);
956 spin_unlock(&vb->lock);
960 purge_fragmented_blocks_thiscpu();
962 put_cpu_var(vmap_block_queue);
966 vb = new_vmap_block(gfp_mask);
975 static void vb_free(const void *addr, unsigned long size)
977 unsigned long offset;
978 unsigned long vb_idx;
980 struct vmap_block *vb;
982 BUG_ON(size & ~PAGE_MASK);
983 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
985 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
987 order = get_order(size);
989 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
991 vb_idx = addr_to_vb_idx((unsigned long)addr);
993 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
997 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
999 spin_lock(&vb->lock);
1000 BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
1002 vb->dirty += 1UL << order;
1003 if (vb->dirty == VMAP_BBMAP_BITS) {
1005 spin_unlock(&vb->lock);
1006 free_vmap_block(vb);
1008 spin_unlock(&vb->lock);
1012 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1014 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1015 * to amortize TLB flushing overheads. What this means is that any page you
1016 * have now, may, in a former life, have been mapped into kernel virtual
1017 * address by the vmap layer and so there might be some CPUs with TLB entries
1018 * still referencing that page (additional to the regular 1:1 kernel mapping).
1020 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1021 * be sure that none of the pages we have control over will have any aliases
1022 * from the vmap layer.
1024 void vm_unmap_aliases(void)
1026 unsigned long start = ULONG_MAX, end = 0;
1030 if (unlikely(!vmap_initialized))
1033 for_each_possible_cpu(cpu) {
1034 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1035 struct vmap_block *vb;
1038 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1041 spin_lock(&vb->lock);
1042 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1043 while (i < VMAP_BBMAP_BITS) {
1046 j = find_next_zero_bit(vb->dirty_map,
1047 VMAP_BBMAP_BITS, i);
1049 s = vb->va->va_start + (i << PAGE_SHIFT);
1050 e = vb->va->va_start + (j << PAGE_SHIFT);
1059 i = find_next_bit(vb->dirty_map,
1060 VMAP_BBMAP_BITS, i);
1062 spin_unlock(&vb->lock);
1067 __purge_vmap_area_lazy(&start, &end, 1, flush);
1069 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1072 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1073 * @mem: the pointer returned by vm_map_ram
1074 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1076 void vm_unmap_ram(const void *mem, unsigned int count)
1078 unsigned long size = count << PAGE_SHIFT;
1079 unsigned long addr = (unsigned long)mem;
1082 BUG_ON(addr < VMALLOC_START);
1083 BUG_ON(addr > VMALLOC_END);
1084 BUG_ON(addr & (PAGE_SIZE-1));
1086 debug_check_no_locks_freed(mem, size);
1087 vmap_debug_free_range(addr, addr+size);
1089 if (likely(count <= VMAP_MAX_ALLOC))
1092 free_unmap_vmap_area_addr(addr);
1094 EXPORT_SYMBOL(vm_unmap_ram);
1097 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1098 * @pages: an array of pointers to the pages to be mapped
1099 * @count: number of pages
1100 * @node: prefer to allocate data structures on this node
1101 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1103 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1105 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1107 unsigned long size = count << PAGE_SHIFT;
1111 if (likely(count <= VMAP_MAX_ALLOC)) {
1112 mem = vb_alloc(size, GFP_KERNEL);
1115 addr = (unsigned long)mem;
1117 struct vmap_area *va;
1118 va = alloc_vmap_area(size, PAGE_SIZE,
1119 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1123 addr = va->va_start;
1126 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1127 vm_unmap_ram(mem, count);
1132 EXPORT_SYMBOL(vm_map_ram);
1135 * vm_area_register_early - register vmap area early during boot
1136 * @vm: vm_struct to register
1137 * @align: requested alignment
1139 * This function is used to register kernel vm area before
1140 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1141 * proper values on entry and other fields should be zero. On return,
1142 * vm->addr contains the allocated address.
1144 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1146 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1148 static size_t vm_init_off __initdata;
1151 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1152 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1154 vm->addr = (void *)addr;
1160 void __init vmalloc_init(void)
1162 struct vmap_area *va;
1163 struct vm_struct *tmp;
1166 for_each_possible_cpu(i) {
1167 struct vmap_block_queue *vbq;
1169 vbq = &per_cpu(vmap_block_queue, i);
1170 spin_lock_init(&vbq->lock);
1171 INIT_LIST_HEAD(&vbq->free);
1174 /* Import existing vmlist entries. */
1175 for (tmp = vmlist; tmp; tmp = tmp->next) {
1176 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1177 va->flags = tmp->flags | VM_VM_AREA;
1178 va->va_start = (unsigned long)tmp->addr;
1179 va->va_end = va->va_start + tmp->size;
1180 __insert_vmap_area(va);
1183 vmap_area_pcpu_hole = VMALLOC_END;
1185 vmap_initialized = true;
1189 * map_kernel_range_noflush - map kernel VM area with the specified pages
1190 * @addr: start of the VM area to map
1191 * @size: size of the VM area to map
1192 * @prot: page protection flags to use
1193 * @pages: pages to map
1195 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1196 * specify should have been allocated using get_vm_area() and its
1200 * This function does NOT do any cache flushing. The caller is
1201 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1202 * before calling this function.
1205 * The number of pages mapped on success, -errno on failure.
1207 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1208 pgprot_t prot, struct page **pages)
1210 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1214 * unmap_kernel_range_noflush - unmap kernel VM area
1215 * @addr: start of the VM area to unmap
1216 * @size: size of the VM area to unmap
1218 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1219 * specify should have been allocated using get_vm_area() and its
1223 * This function does NOT do any cache flushing. The caller is
1224 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1225 * before calling this function and flush_tlb_kernel_range() after.
1227 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1229 vunmap_page_range(addr, addr + size);
1231 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1234 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1235 * @addr: start of the VM area to unmap
1236 * @size: size of the VM area to unmap
1238 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1239 * the unmapping and tlb after.
1241 void unmap_kernel_range(unsigned long addr, unsigned long size)
1243 unsigned long end = addr + size;
1245 flush_cache_vunmap(addr, end);
1246 vunmap_page_range(addr, end);
1247 flush_tlb_kernel_range(addr, end);
1250 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1252 unsigned long addr = (unsigned long)area->addr;
1253 unsigned long end = addr + area->size - PAGE_SIZE;
1256 err = vmap_page_range(addr, end, prot, *pages);
1264 EXPORT_SYMBOL_GPL(map_vm_area);
1266 /*** Old vmalloc interfaces ***/
1267 DEFINE_RWLOCK(vmlist_lock);
1268 struct vm_struct *vmlist;
1270 static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1271 unsigned long flags, void *caller)
1273 struct vm_struct *tmp, **p;
1276 vm->addr = (void *)va->va_start;
1277 vm->size = va->va_end - va->va_start;
1278 vm->caller = caller;
1280 va->flags |= VM_VM_AREA;
1282 write_lock(&vmlist_lock);
1283 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1284 if (tmp->addr >= vm->addr)
1289 write_unlock(&vmlist_lock);
1292 static struct vm_struct *__get_vm_area_node(unsigned long size,
1293 unsigned long align, unsigned long flags, unsigned long start,
1294 unsigned long end, int node, gfp_t gfp_mask, void *caller)
1296 static struct vmap_area *va;
1297 struct vm_struct *area;
1299 BUG_ON(in_interrupt());
1300 if (flags & VM_IOREMAP) {
1301 int bit = fls(size);
1303 if (bit > IOREMAP_MAX_ORDER)
1304 bit = IOREMAP_MAX_ORDER;
1305 else if (bit < PAGE_SHIFT)
1311 size = PAGE_ALIGN(size);
1312 if (unlikely(!size))
1315 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1316 if (unlikely(!area))
1320 * We always allocate a guard page.
1324 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1330 insert_vmalloc_vm(area, va, flags, caller);
1334 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1335 unsigned long start, unsigned long end)
1337 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1338 __builtin_return_address(0));
1340 EXPORT_SYMBOL_GPL(__get_vm_area);
1342 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1343 unsigned long start, unsigned long end,
1346 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1351 * get_vm_area - reserve a contiguous kernel virtual area
1352 * @size: size of the area
1353 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1355 * Search an area of @size in the kernel virtual mapping area,
1356 * and reserved it for out purposes. Returns the area descriptor
1357 * on success or %NULL on failure.
1359 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1361 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1362 -1, GFP_KERNEL, __builtin_return_address(0));
1365 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1368 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1369 -1, GFP_KERNEL, caller);
1372 static struct vm_struct *find_vm_area(const void *addr)
1374 struct vmap_area *va;
1376 va = find_vmap_area((unsigned long)addr);
1377 if (va && va->flags & VM_VM_AREA)
1384 * remove_vm_area - find and remove a continuous kernel virtual area
1385 * @addr: base address
1387 * Search for the kernel VM area starting at @addr, and remove it.
1388 * This function returns the found VM area, but using it is NOT safe
1389 * on SMP machines, except for its size or flags.
1391 struct vm_struct *remove_vm_area(const void *addr)
1393 struct vmap_area *va;
1395 va = find_vmap_area((unsigned long)addr);
1396 if (va && va->flags & VM_VM_AREA) {
1397 struct vm_struct *vm = va->private;
1398 struct vm_struct *tmp, **p;
1400 * remove from list and disallow access to this vm_struct
1401 * before unmap. (address range confliction is maintained by
1404 write_lock(&vmlist_lock);
1405 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1408 write_unlock(&vmlist_lock);
1410 vmap_debug_free_range(va->va_start, va->va_end);
1411 free_unmap_vmap_area(va);
1412 vm->size -= PAGE_SIZE;
1419 static void __vunmap(const void *addr, int deallocate_pages)
1421 struct vm_struct *area;
1426 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1427 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1431 area = remove_vm_area(addr);
1432 if (unlikely(!area)) {
1433 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1438 debug_check_no_locks_freed(addr, area->size);
1439 debug_check_no_obj_freed(addr, area->size);
1441 if (deallocate_pages) {
1444 for (i = 0; i < area->nr_pages; i++) {
1445 struct page *page = area->pages[i];
1451 if (area->flags & VM_VPAGES)
1462 * vfree - release memory allocated by vmalloc()
1463 * @addr: memory base address
1465 * Free the virtually continuous memory area starting at @addr, as
1466 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1467 * NULL, no operation is performed.
1469 * Must not be called in interrupt context.
1471 void vfree(const void *addr)
1473 BUG_ON(in_interrupt());
1475 kmemleak_free(addr);
1479 EXPORT_SYMBOL(vfree);
1482 * vunmap - release virtual mapping obtained by vmap()
1483 * @addr: memory base address
1485 * Free the virtually contiguous memory area starting at @addr,
1486 * which was created from the page array passed to vmap().
1488 * Must not be called in interrupt context.
1490 void vunmap(const void *addr)
1492 BUG_ON(in_interrupt());
1496 EXPORT_SYMBOL(vunmap);
1499 * vmap - map an array of pages into virtually contiguous space
1500 * @pages: array of page pointers
1501 * @count: number of pages to map
1502 * @flags: vm_area->flags
1503 * @prot: page protection for the mapping
1505 * Maps @count pages from @pages into contiguous kernel virtual
1508 void *vmap(struct page **pages, unsigned int count,
1509 unsigned long flags, pgprot_t prot)
1511 struct vm_struct *area;
1515 if (count > totalram_pages)
1518 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1519 __builtin_return_address(0));
1523 if (map_vm_area(area, prot, &pages)) {
1530 EXPORT_SYMBOL(vmap);
1532 static void *__vmalloc_node(unsigned long size, unsigned long align,
1533 gfp_t gfp_mask, pgprot_t prot,
1534 int node, void *caller);
1535 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1536 pgprot_t prot, int node, void *caller)
1538 const int order = 0;
1539 struct page **pages;
1540 unsigned int nr_pages, array_size, i;
1541 gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1543 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1544 array_size = (nr_pages * sizeof(struct page *));
1546 area->nr_pages = nr_pages;
1547 /* Please note that the recursion is strictly bounded. */
1548 if (array_size > PAGE_SIZE) {
1549 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1550 PAGE_KERNEL, node, caller);
1551 area->flags |= VM_VPAGES;
1553 pages = kmalloc_node(array_size, nested_gfp, node);
1555 area->pages = pages;
1556 area->caller = caller;
1558 remove_vm_area(area->addr);
1563 for (i = 0; i < area->nr_pages; i++) {
1565 gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1568 page = alloc_page(tmp_mask);
1570 page = alloc_pages_node(node, tmp_mask, order);
1572 if (unlikely(!page)) {
1573 /* Successfully allocated i pages, free them in __vunmap() */
1577 area->pages[i] = page;
1580 if (map_vm_area(area, prot, &pages))
1585 warn_alloc_failed(gfp_mask, order, "vmalloc: allocation failure, "
1586 "allocated %ld of %ld bytes\n",
1587 (area->nr_pages*PAGE_SIZE), area->size);
1593 * __vmalloc_node_range - allocate virtually contiguous memory
1594 * @size: allocation size
1595 * @align: desired alignment
1596 * @start: vm area range start
1597 * @end: vm area range end
1598 * @gfp_mask: flags for the page level allocator
1599 * @prot: protection mask for the allocated pages
1600 * @node: node to use for allocation or -1
1601 * @caller: caller's return address
1603 * Allocate enough pages to cover @size from the page level
1604 * allocator with @gfp_mask flags. Map them into contiguous
1605 * kernel virtual space, using a pagetable protection of @prot.
1607 void *__vmalloc_node_range(unsigned long size, unsigned long align,
1608 unsigned long start, unsigned long end, gfp_t gfp_mask,
1609 pgprot_t prot, int node, void *caller)
1611 struct vm_struct *area;
1613 unsigned long real_size = size;
1615 size = PAGE_ALIGN(size);
1616 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1619 area = __get_vm_area_node(size, align, VM_ALLOC, start, end, node,
1625 addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1628 * A ref_count = 3 is needed because the vm_struct and vmap_area
1629 * structures allocated in the __get_vm_area_node() function contain
1630 * references to the virtual address of the vmalloc'ed block.
1632 kmemleak_alloc(addr, real_size, 3, gfp_mask);
1638 * __vmalloc_node - allocate virtually contiguous memory
1639 * @size: allocation size
1640 * @align: desired alignment
1641 * @gfp_mask: flags for the page level allocator
1642 * @prot: protection mask for the allocated pages
1643 * @node: node to use for allocation or -1
1644 * @caller: caller's return address
1646 * Allocate enough pages to cover @size from the page level
1647 * allocator with @gfp_mask flags. Map them into contiguous
1648 * kernel virtual space, using a pagetable protection of @prot.
1650 static void *__vmalloc_node(unsigned long size, unsigned long align,
1651 gfp_t gfp_mask, pgprot_t prot,
1652 int node, void *caller)
1654 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1655 gfp_mask, prot, node, caller);
1658 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1660 return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1661 __builtin_return_address(0));
1663 EXPORT_SYMBOL(__vmalloc);
1665 static inline void *__vmalloc_node_flags(unsigned long size,
1666 int node, gfp_t flags)
1668 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1669 node, __builtin_return_address(0));
1673 * vmalloc - allocate virtually contiguous memory
1674 * @size: allocation size
1675 * Allocate enough pages to cover @size from the page level
1676 * allocator and map them into contiguous kernel virtual space.
1678 * For tight control over page level allocator and protection flags
1679 * use __vmalloc() instead.
1681 void *vmalloc(unsigned long size)
1683 return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1685 EXPORT_SYMBOL(vmalloc);
1688 * vzalloc - allocate virtually contiguous memory with zero fill
1689 * @size: allocation size
1690 * Allocate enough pages to cover @size from the page level
1691 * allocator and map them into contiguous kernel virtual space.
1692 * The memory allocated is set to zero.
1694 * For tight control over page level allocator and protection flags
1695 * use __vmalloc() instead.
1697 void *vzalloc(unsigned long size)
1699 return __vmalloc_node_flags(size, -1,
1700 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1702 EXPORT_SYMBOL(vzalloc);
1705 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1706 * @size: allocation size
1708 * The resulting memory area is zeroed so it can be mapped to userspace
1709 * without leaking data.
1711 void *vmalloc_user(unsigned long size)
1713 struct vm_struct *area;
1716 ret = __vmalloc_node(size, SHMLBA,
1717 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1718 PAGE_KERNEL, -1, __builtin_return_address(0));
1720 area = find_vm_area(ret);
1721 area->flags |= VM_USERMAP;
1725 EXPORT_SYMBOL(vmalloc_user);
1728 * vmalloc_node - allocate memory on a specific node
1729 * @size: allocation size
1732 * Allocate enough pages to cover @size from the page level
1733 * allocator and map them into contiguous kernel virtual space.
1735 * For tight control over page level allocator and protection flags
1736 * use __vmalloc() instead.
1738 void *vmalloc_node(unsigned long size, int node)
1740 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1741 node, __builtin_return_address(0));
1743 EXPORT_SYMBOL(vmalloc_node);
1746 * vzalloc_node - allocate memory on a specific node with zero fill
1747 * @size: allocation size
1750 * Allocate enough pages to cover @size from the page level
1751 * allocator and map them into contiguous kernel virtual space.
1752 * The memory allocated is set to zero.
1754 * For tight control over page level allocator and protection flags
1755 * use __vmalloc_node() instead.
1757 void *vzalloc_node(unsigned long size, int node)
1759 return __vmalloc_node_flags(size, node,
1760 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1762 EXPORT_SYMBOL(vzalloc_node);
1764 #ifndef PAGE_KERNEL_EXEC
1765 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1769 * vmalloc_exec - allocate virtually contiguous, executable memory
1770 * @size: allocation size
1772 * Kernel-internal function to allocate enough pages to cover @size
1773 * the page level allocator and map them into contiguous and
1774 * executable kernel virtual space.
1776 * For tight control over page level allocator and protection flags
1777 * use __vmalloc() instead.
1780 void *vmalloc_exec(unsigned long size)
1782 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1783 -1, __builtin_return_address(0));
1786 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1787 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1788 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1789 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1791 #define GFP_VMALLOC32 GFP_KERNEL
1795 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1796 * @size: allocation size
1798 * Allocate enough 32bit PA addressable pages to cover @size from the
1799 * page level allocator and map them into contiguous kernel virtual space.
1801 void *vmalloc_32(unsigned long size)
1803 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1804 -1, __builtin_return_address(0));
1806 EXPORT_SYMBOL(vmalloc_32);
1809 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1810 * @size: allocation size
1812 * The resulting memory area is 32bit addressable and zeroed so it can be
1813 * mapped to userspace without leaking data.
1815 void *vmalloc_32_user(unsigned long size)
1817 struct vm_struct *area;
1820 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1821 -1, __builtin_return_address(0));
1823 area = find_vm_area(ret);
1824 area->flags |= VM_USERMAP;
1828 EXPORT_SYMBOL(vmalloc_32_user);
1831 * small helper routine , copy contents to buf from addr.
1832 * If the page is not present, fill zero.
1835 static int aligned_vread(char *buf, char *addr, unsigned long count)
1841 unsigned long offset, length;
1843 offset = (unsigned long)addr & ~PAGE_MASK;
1844 length = PAGE_SIZE - offset;
1847 p = vmalloc_to_page(addr);
1849 * To do safe access to this _mapped_ area, we need
1850 * lock. But adding lock here means that we need to add
1851 * overhead of vmalloc()/vfree() calles for this _debug_
1852 * interface, rarely used. Instead of that, we'll use
1853 * kmap() and get small overhead in this access function.
1857 * we can expect USER0 is not used (see vread/vwrite's
1858 * function description)
1860 void *map = kmap_atomic(p, KM_USER0);
1861 memcpy(buf, map + offset, length);
1862 kunmap_atomic(map, KM_USER0);
1864 memset(buf, 0, length);
1874 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1880 unsigned long offset, length;
1882 offset = (unsigned long)addr & ~PAGE_MASK;
1883 length = PAGE_SIZE - offset;
1886 p = vmalloc_to_page(addr);
1888 * To do safe access to this _mapped_ area, we need
1889 * lock. But adding lock here means that we need to add
1890 * overhead of vmalloc()/vfree() calles for this _debug_
1891 * interface, rarely used. Instead of that, we'll use
1892 * kmap() and get small overhead in this access function.
1896 * we can expect USER0 is not used (see vread/vwrite's
1897 * function description)
1899 void *map = kmap_atomic(p, KM_USER0);
1900 memcpy(map + offset, buf, length);
1901 kunmap_atomic(map, KM_USER0);
1912 * vread() - read vmalloc area in a safe way.
1913 * @buf: buffer for reading data
1914 * @addr: vm address.
1915 * @count: number of bytes to be read.
1917 * Returns # of bytes which addr and buf should be increased.
1918 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1919 * includes any intersect with alive vmalloc area.
1921 * This function checks that addr is a valid vmalloc'ed area, and
1922 * copy data from that area to a given buffer. If the given memory range
1923 * of [addr...addr+count) includes some valid address, data is copied to
1924 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1925 * IOREMAP area is treated as memory hole and no copy is done.
1927 * If [addr...addr+count) doesn't includes any intersects with alive
1928 * vm_struct area, returns 0.
1929 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1930 * the caller should guarantee KM_USER0 is not used.
1932 * Note: In usual ops, vread() is never necessary because the caller
1933 * should know vmalloc() area is valid and can use memcpy().
1934 * This is for routines which have to access vmalloc area without
1935 * any informaion, as /dev/kmem.
1939 long vread(char *buf, char *addr, unsigned long count)
1941 struct vm_struct *tmp;
1942 char *vaddr, *buf_start = buf;
1943 unsigned long buflen = count;
1946 /* Don't allow overflow */
1947 if ((unsigned long) addr + count < count)
1948 count = -(unsigned long) addr;
1950 read_lock(&vmlist_lock);
1951 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1952 vaddr = (char *) tmp->addr;
1953 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1955 while (addr < vaddr) {
1963 n = vaddr + tmp->size - PAGE_SIZE - addr;
1966 if (!(tmp->flags & VM_IOREMAP))
1967 aligned_vread(buf, addr, n);
1968 else /* IOREMAP area is treated as memory hole */
1975 read_unlock(&vmlist_lock);
1977 if (buf == buf_start)
1979 /* zero-fill memory holes */
1980 if (buf != buf_start + buflen)
1981 memset(buf, 0, buflen - (buf - buf_start));
1987 * vwrite() - write vmalloc area in a safe way.
1988 * @buf: buffer for source data
1989 * @addr: vm address.
1990 * @count: number of bytes to be read.
1992 * Returns # of bytes which addr and buf should be incresed.
1993 * (same number to @count).
1994 * If [addr...addr+count) doesn't includes any intersect with valid
1995 * vmalloc area, returns 0.
1997 * This function checks that addr is a valid vmalloc'ed area, and
1998 * copy data from a buffer to the given addr. If specified range of
1999 * [addr...addr+count) includes some valid address, data is copied from
2000 * proper area of @buf. If there are memory holes, no copy to hole.
2001 * IOREMAP area is treated as memory hole and no copy is done.
2003 * If [addr...addr+count) doesn't includes any intersects with alive
2004 * vm_struct area, returns 0.
2005 * @buf should be kernel's buffer. Because this function uses KM_USER0,
2006 * the caller should guarantee KM_USER0 is not used.
2008 * Note: In usual ops, vwrite() is never necessary because the caller
2009 * should know vmalloc() area is valid and can use memcpy().
2010 * This is for routines which have to access vmalloc area without
2011 * any informaion, as /dev/kmem.
2014 long vwrite(char *buf, char *addr, unsigned long count)
2016 struct vm_struct *tmp;
2018 unsigned long n, buflen;
2021 /* Don't allow overflow */
2022 if ((unsigned long) addr + count < count)
2023 count = -(unsigned long) addr;
2026 read_lock(&vmlist_lock);
2027 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2028 vaddr = (char *) tmp->addr;
2029 if (addr >= vaddr + tmp->size - PAGE_SIZE)
2031 while (addr < vaddr) {
2038 n = vaddr + tmp->size - PAGE_SIZE - addr;
2041 if (!(tmp->flags & VM_IOREMAP)) {
2042 aligned_vwrite(buf, addr, n);
2050 read_unlock(&vmlist_lock);
2057 * remap_vmalloc_range - map vmalloc pages to userspace
2058 * @vma: vma to cover (map full range of vma)
2059 * @addr: vmalloc memory
2060 * @pgoff: number of pages into addr before first page to map
2062 * Returns: 0 for success, -Exxx on failure
2064 * This function checks that addr is a valid vmalloc'ed area, and
2065 * that it is big enough to cover the vma. Will return failure if
2066 * that criteria isn't met.
2068 * Similar to remap_pfn_range() (see mm/memory.c)
2070 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2071 unsigned long pgoff)
2073 struct vm_struct *area;
2074 unsigned long uaddr = vma->vm_start;
2075 unsigned long usize = vma->vm_end - vma->vm_start;
2077 if ((PAGE_SIZE-1) & (unsigned long)addr)
2080 area = find_vm_area(addr);
2084 if (!(area->flags & VM_USERMAP))
2087 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2090 addr += pgoff << PAGE_SHIFT;
2092 struct page *page = vmalloc_to_page(addr);
2095 ret = vm_insert_page(vma, uaddr, page);
2102 } while (usize > 0);
2104 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2105 vma->vm_flags |= VM_RESERVED;
2109 EXPORT_SYMBOL(remap_vmalloc_range);
2112 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2115 void __attribute__((weak)) vmalloc_sync_all(void)
2120 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2122 /* apply_to_page_range() does all the hard work. */
2127 * alloc_vm_area - allocate a range of kernel address space
2128 * @size: size of the area
2130 * Returns: NULL on failure, vm_struct on success
2132 * This function reserves a range of kernel address space, and
2133 * allocates pagetables to map that range. No actual mappings
2134 * are created. If the kernel address space is not shared
2135 * between processes, it syncs the pagetable across all
2138 struct vm_struct *alloc_vm_area(size_t size)
2140 struct vm_struct *area;
2142 area = get_vm_area_caller(size, VM_IOREMAP,
2143 __builtin_return_address(0));
2148 * This ensures that page tables are constructed for this region
2149 * of kernel virtual address space and mapped into init_mm.
2151 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2152 area->size, f, NULL)) {
2158 * If the allocated address space is passed to a hypercall
2159 * before being used then we cannot rely on a page fault to
2160 * trigger an update of the page tables. So sync all the page
2167 EXPORT_SYMBOL_GPL(alloc_vm_area);
2169 void free_vm_area(struct vm_struct *area)
2171 struct vm_struct *ret;
2172 ret = remove_vm_area(area->addr);
2173 BUG_ON(ret != area);
2176 EXPORT_SYMBOL_GPL(free_vm_area);
2179 static struct vmap_area *node_to_va(struct rb_node *n)
2181 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2185 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2186 * @end: target address
2187 * @pnext: out arg for the next vmap_area
2188 * @pprev: out arg for the previous vmap_area
2190 * Returns: %true if either or both of next and prev are found,
2191 * %false if no vmap_area exists
2193 * Find vmap_areas end addresses of which enclose @end. ie. if not
2194 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2196 static bool pvm_find_next_prev(unsigned long end,
2197 struct vmap_area **pnext,
2198 struct vmap_area **pprev)
2200 struct rb_node *n = vmap_area_root.rb_node;
2201 struct vmap_area *va = NULL;
2204 va = rb_entry(n, struct vmap_area, rb_node);
2205 if (end < va->va_end)
2207 else if (end > va->va_end)
2216 if (va->va_end > end) {
2218 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2221 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2227 * pvm_determine_end - find the highest aligned address between two vmap_areas
2228 * @pnext: in/out arg for the next vmap_area
2229 * @pprev: in/out arg for the previous vmap_area
2232 * Returns: determined end address
2234 * Find the highest aligned address between *@pnext and *@pprev below
2235 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2236 * down address is between the end addresses of the two vmap_areas.
2238 * Please note that the address returned by this function may fall
2239 * inside *@pnext vmap_area. The caller is responsible for checking
2242 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2243 struct vmap_area **pprev,
2244 unsigned long align)
2246 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2250 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2254 while (*pprev && (*pprev)->va_end > addr) {
2256 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2263 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2264 * @offsets: array containing offset of each area
2265 * @sizes: array containing size of each area
2266 * @nr_vms: the number of areas to allocate
2267 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2269 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2270 * vm_structs on success, %NULL on failure
2272 * Percpu allocator wants to use congruent vm areas so that it can
2273 * maintain the offsets among percpu areas. This function allocates
2274 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2275 * be scattered pretty far, distance between two areas easily going up
2276 * to gigabytes. To avoid interacting with regular vmallocs, these
2277 * areas are allocated from top.
2279 * Despite its complicated look, this allocator is rather simple. It
2280 * does everything top-down and scans areas from the end looking for
2281 * matching slot. While scanning, if any of the areas overlaps with
2282 * existing vmap_area, the base address is pulled down to fit the
2283 * area. Scanning is repeated till all the areas fit and then all
2284 * necessary data structres are inserted and the result is returned.
2286 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2287 const size_t *sizes, int nr_vms,
2290 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2291 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2292 struct vmap_area **vas, *prev, *next;
2293 struct vm_struct **vms;
2294 int area, area2, last_area, term_area;
2295 unsigned long base, start, end, last_end;
2296 bool purged = false;
2298 /* verify parameters and allocate data structures */
2299 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2300 for (last_area = 0, area = 0; area < nr_vms; area++) {
2301 start = offsets[area];
2302 end = start + sizes[area];
2304 /* is everything aligned properly? */
2305 BUG_ON(!IS_ALIGNED(offsets[area], align));
2306 BUG_ON(!IS_ALIGNED(sizes[area], align));
2308 /* detect the area with the highest address */
2309 if (start > offsets[last_area])
2312 for (area2 = 0; area2 < nr_vms; area2++) {
2313 unsigned long start2 = offsets[area2];
2314 unsigned long end2 = start2 + sizes[area2];
2319 BUG_ON(start2 >= start && start2 < end);
2320 BUG_ON(end2 <= end && end2 > start);
2323 last_end = offsets[last_area] + sizes[last_area];
2325 if (vmalloc_end - vmalloc_start < last_end) {
2330 vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
2331 vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
2335 for (area = 0; area < nr_vms; area++) {
2336 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2337 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2338 if (!vas[area] || !vms[area])
2342 spin_lock(&vmap_area_lock);
2344 /* start scanning - we scan from the top, begin with the last area */
2345 area = term_area = last_area;
2346 start = offsets[area];
2347 end = start + sizes[area];
2349 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2350 base = vmalloc_end - last_end;
2353 base = pvm_determine_end(&next, &prev, align) - end;
2356 BUG_ON(next && next->va_end <= base + end);
2357 BUG_ON(prev && prev->va_end > base + end);
2360 * base might have underflowed, add last_end before
2363 if (base + last_end < vmalloc_start + last_end) {
2364 spin_unlock(&vmap_area_lock);
2366 purge_vmap_area_lazy();
2374 * If next overlaps, move base downwards so that it's
2375 * right below next and then recheck.
2377 if (next && next->va_start < base + end) {
2378 base = pvm_determine_end(&next, &prev, align) - end;
2384 * If prev overlaps, shift down next and prev and move
2385 * base so that it's right below new next and then
2388 if (prev && prev->va_end > base + start) {
2390 prev = node_to_va(rb_prev(&next->rb_node));
2391 base = pvm_determine_end(&next, &prev, align) - end;
2397 * This area fits, move on to the previous one. If
2398 * the previous one is the terminal one, we're done.
2400 area = (area + nr_vms - 1) % nr_vms;
2401 if (area == term_area)
2403 start = offsets[area];
2404 end = start + sizes[area];
2405 pvm_find_next_prev(base + end, &next, &prev);
2408 /* we've found a fitting base, insert all va's */
2409 for (area = 0; area < nr_vms; area++) {
2410 struct vmap_area *va = vas[area];
2412 va->va_start = base + offsets[area];
2413 va->va_end = va->va_start + sizes[area];
2414 __insert_vmap_area(va);
2417 vmap_area_pcpu_hole = base + offsets[last_area];
2419 spin_unlock(&vmap_area_lock);
2421 /* insert all vm's */
2422 for (area = 0; area < nr_vms; area++)
2423 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2430 for (area = 0; area < nr_vms; area++) {
2442 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2443 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2444 * @nr_vms: the number of allocated areas
2446 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2448 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2452 for (i = 0; i < nr_vms; i++)
2453 free_vm_area(vms[i]);
2456 #endif /* CONFIG_SMP */
2458 #ifdef CONFIG_PROC_FS
2459 static void *s_start(struct seq_file *m, loff_t *pos)
2460 __acquires(&vmlist_lock)
2463 struct vm_struct *v;
2465 read_lock(&vmlist_lock);
2467 while (n > 0 && v) {
2478 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2480 struct vm_struct *v = p;
2486 static void s_stop(struct seq_file *m, void *p)
2487 __releases(&vmlist_lock)
2489 read_unlock(&vmlist_lock);
2492 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2495 unsigned int nr, *counters = m->private;
2500 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2502 for (nr = 0; nr < v->nr_pages; nr++)
2503 counters[page_to_nid(v->pages[nr])]++;
2505 for_each_node_state(nr, N_HIGH_MEMORY)
2507 seq_printf(m, " N%u=%u", nr, counters[nr]);
2511 static int s_show(struct seq_file *m, void *p)
2513 struct vm_struct *v = p;
2515 seq_printf(m, "0x%p-0x%p %7ld",
2516 v->addr, v->addr + v->size, v->size);
2519 seq_printf(m, " %pS", v->caller);
2522 seq_printf(m, " pages=%d", v->nr_pages);
2525 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2527 if (v->flags & VM_IOREMAP)
2528 seq_printf(m, " ioremap");
2530 if (v->flags & VM_ALLOC)
2531 seq_printf(m, " vmalloc");
2533 if (v->flags & VM_MAP)
2534 seq_printf(m, " vmap");
2536 if (v->flags & VM_USERMAP)
2537 seq_printf(m, " user");
2539 if (v->flags & VM_VPAGES)
2540 seq_printf(m, " vpages");
2542 show_numa_info(m, v);
2547 static const struct seq_operations vmalloc_op = {
2554 static int vmalloc_open(struct inode *inode, struct file *file)
2556 unsigned int *ptr = NULL;
2560 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2564 ret = seq_open(file, &vmalloc_op);
2566 struct seq_file *m = file->private_data;
2573 static const struct file_operations proc_vmalloc_operations = {
2574 .open = vmalloc_open,
2576 .llseek = seq_lseek,
2577 .release = seq_release_private,
2580 static int __init proc_vmalloc_init(void)
2582 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2585 module_init(proc_vmalloc_init);