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/slab.h>
16 #include <linux/spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/debugobjects.h>
21 #include <linux/kallsyms.h>
22 #include <linux/list.h>
23 #include <linux/rbtree.h>
24 #include <linux/radix-tree.h>
25 #include <linux/rcupdate.h>
26 #include <linux/bootmem.h>
28 #include <asm/atomic.h>
29 #include <asm/uaccess.h>
30 #include <asm/tlbflush.h>
33 /*** Page table manipulation functions ***/
35 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
39 pte = pte_offset_kernel(pmd, addr);
41 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
42 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
43 } while (pte++, addr += PAGE_SIZE, addr != end);
46 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
51 pmd = pmd_offset(pud, addr);
53 next = pmd_addr_end(addr, end);
54 if (pmd_none_or_clear_bad(pmd))
56 vunmap_pte_range(pmd, addr, next);
57 } while (pmd++, addr = next, addr != end);
60 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
65 pud = pud_offset(pgd, addr);
67 next = pud_addr_end(addr, end);
68 if (pud_none_or_clear_bad(pud))
70 vunmap_pmd_range(pud, addr, next);
71 } while (pud++, addr = next, addr != end);
74 static void vunmap_page_range(unsigned long addr, unsigned long end)
80 pgd = pgd_offset_k(addr);
82 next = pgd_addr_end(addr, end);
83 if (pgd_none_or_clear_bad(pgd))
85 vunmap_pud_range(pgd, addr, next);
86 } while (pgd++, addr = next, addr != end);
89 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
90 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
95 * nr is a running index into the array which helps higher level
96 * callers keep track of where we're up to.
99 pte = pte_alloc_kernel(pmd, addr);
103 struct page *page = pages[*nr];
105 if (WARN_ON(!pte_none(*pte)))
109 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
111 } while (pte++, addr += PAGE_SIZE, addr != end);
115 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
116 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
121 pmd = pmd_alloc(&init_mm, pud, addr);
125 next = pmd_addr_end(addr, end);
126 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
128 } while (pmd++, addr = next, addr != end);
132 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
133 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
138 pud = pud_alloc(&init_mm, pgd, addr);
142 next = pud_addr_end(addr, end);
143 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
145 } while (pud++, addr = next, addr != end);
150 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
151 * will have pfns corresponding to the "pages" array.
153 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
155 static int vmap_page_range(unsigned long start, unsigned long end,
156 pgprot_t prot, struct page **pages)
160 unsigned long addr = start;
165 pgd = pgd_offset_k(addr);
167 next = pgd_addr_end(addr, end);
168 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
171 } while (pgd++, addr = next, addr != end);
172 flush_cache_vmap(start, end);
179 static inline int is_vmalloc_or_module_addr(const void *x)
182 * ARM, x86-64 and sparc64 put modules in a special place,
183 * and fall back on vmalloc() if that fails. Others
184 * just put it in the vmalloc space.
186 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
187 unsigned long addr = (unsigned long)x;
188 if (addr >= MODULES_VADDR && addr < MODULES_END)
191 return is_vmalloc_addr(x);
195 * Walk a vmap address to the struct page it maps.
197 struct page *vmalloc_to_page(const void *vmalloc_addr)
199 unsigned long addr = (unsigned long) vmalloc_addr;
200 struct page *page = NULL;
201 pgd_t *pgd = pgd_offset_k(addr);
204 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
205 * architectures that do not vmalloc module space
207 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
209 if (!pgd_none(*pgd)) {
210 pud_t *pud = pud_offset(pgd, addr);
211 if (!pud_none(*pud)) {
212 pmd_t *pmd = pmd_offset(pud, addr);
213 if (!pmd_none(*pmd)) {
216 ptep = pte_offset_map(pmd, addr);
218 if (pte_present(pte))
219 page = pte_page(pte);
226 EXPORT_SYMBOL(vmalloc_to_page);
229 * Map a vmalloc()-space virtual address to the physical page frame number.
231 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
233 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
235 EXPORT_SYMBOL(vmalloc_to_pfn);
238 /*** Global kva allocator ***/
240 #define VM_LAZY_FREE 0x01
241 #define VM_LAZY_FREEING 0x02
242 #define VM_VM_AREA 0x04
245 unsigned long va_start;
246 unsigned long va_end;
248 struct rb_node rb_node; /* address sorted rbtree */
249 struct list_head list; /* address sorted list */
250 struct list_head purge_list; /* "lazy purge" list */
252 struct rcu_head rcu_head;
255 static DEFINE_SPINLOCK(vmap_area_lock);
256 static struct rb_root vmap_area_root = RB_ROOT;
257 static LIST_HEAD(vmap_area_list);
259 static struct vmap_area *__find_vmap_area(unsigned long addr)
261 struct rb_node *n = vmap_area_root.rb_node;
264 struct vmap_area *va;
266 va = rb_entry(n, struct vmap_area, rb_node);
267 if (addr < va->va_start)
269 else if (addr > va->va_start)
278 static void __insert_vmap_area(struct vmap_area *va)
280 struct rb_node **p = &vmap_area_root.rb_node;
281 struct rb_node *parent = NULL;
285 struct vmap_area *tmp;
288 tmp = rb_entry(parent, struct vmap_area, rb_node);
289 if (va->va_start < tmp->va_end)
291 else if (va->va_end > tmp->va_start)
297 rb_link_node(&va->rb_node, parent, p);
298 rb_insert_color(&va->rb_node, &vmap_area_root);
300 /* address-sort this list so it is usable like the vmlist */
301 tmp = rb_prev(&va->rb_node);
303 struct vmap_area *prev;
304 prev = rb_entry(tmp, struct vmap_area, rb_node);
305 list_add_rcu(&va->list, &prev->list);
307 list_add_rcu(&va->list, &vmap_area_list);
310 static void purge_vmap_area_lazy(void);
313 * Allocate a region of KVA of the specified size and alignment, within the
316 static struct vmap_area *alloc_vmap_area(unsigned long size,
318 unsigned long vstart, unsigned long vend,
319 int node, gfp_t gfp_mask)
321 struct vmap_area *va;
326 BUG_ON(size & ~PAGE_MASK);
328 va = kmalloc_node(sizeof(struct vmap_area),
329 gfp_mask & GFP_RECLAIM_MASK, node);
331 return ERR_PTR(-ENOMEM);
334 addr = ALIGN(vstart, align);
336 spin_lock(&vmap_area_lock);
337 /* XXX: could have a last_hole cache */
338 n = vmap_area_root.rb_node;
340 struct vmap_area *first = NULL;
343 struct vmap_area *tmp;
344 tmp = rb_entry(n, struct vmap_area, rb_node);
345 if (tmp->va_end >= addr) {
346 if (!first && tmp->va_start < addr + size)
358 if (first->va_end < addr) {
359 n = rb_next(&first->rb_node);
361 first = rb_entry(n, struct vmap_area, rb_node);
366 while (addr + size > first->va_start && addr + size <= vend) {
367 addr = ALIGN(first->va_end + PAGE_SIZE, align);
369 n = rb_next(&first->rb_node);
371 first = rb_entry(n, struct vmap_area, rb_node);
377 if (addr + size > vend) {
378 spin_unlock(&vmap_area_lock);
380 purge_vmap_area_lazy();
384 if (printk_ratelimit())
386 "vmap allocation for size %lu failed: "
387 "use vmalloc=<size> to increase size.\n", size);
388 return ERR_PTR(-EBUSY);
391 BUG_ON(addr & (align-1));
394 va->va_end = addr + size;
396 __insert_vmap_area(va);
397 spin_unlock(&vmap_area_lock);
402 static void rcu_free_va(struct rcu_head *head)
404 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
409 static void __free_vmap_area(struct vmap_area *va)
411 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
412 rb_erase(&va->rb_node, &vmap_area_root);
413 RB_CLEAR_NODE(&va->rb_node);
414 list_del_rcu(&va->list);
416 call_rcu(&va->rcu_head, rcu_free_va);
420 * Free a region of KVA allocated by alloc_vmap_area
422 static void free_vmap_area(struct vmap_area *va)
424 spin_lock(&vmap_area_lock);
425 __free_vmap_area(va);
426 spin_unlock(&vmap_area_lock);
430 * Clear the pagetable entries of a given vmap_area
432 static void unmap_vmap_area(struct vmap_area *va)
434 vunmap_page_range(va->va_start, va->va_end);
437 static void vmap_debug_free_range(unsigned long start, unsigned long end)
440 * Unmap page tables and force a TLB flush immediately if
441 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
442 * bugs similarly to those in linear kernel virtual address
443 * space after a page has been freed.
445 * All the lazy freeing logic is still retained, in order to
446 * minimise intrusiveness of this debugging feature.
448 * This is going to be *slow* (linear kernel virtual address
449 * debugging doesn't do a broadcast TLB flush so it is a lot
452 #ifdef CONFIG_DEBUG_PAGEALLOC
453 vunmap_page_range(start, end);
454 flush_tlb_kernel_range(start, end);
459 * lazy_max_pages is the maximum amount of virtual address space we gather up
460 * before attempting to purge with a TLB flush.
462 * There is a tradeoff here: a larger number will cover more kernel page tables
463 * and take slightly longer to purge, but it will linearly reduce the number of
464 * global TLB flushes that must be performed. It would seem natural to scale
465 * this number up linearly with the number of CPUs (because vmapping activity
466 * could also scale linearly with the number of CPUs), however it is likely
467 * that in practice, workloads might be constrained in other ways that mean
468 * vmap activity will not scale linearly with CPUs. Also, I want to be
469 * conservative and not introduce a big latency on huge systems, so go with
470 * a less aggressive log scale. It will still be an improvement over the old
471 * code, and it will be simple to change the scale factor if we find that it
472 * becomes a problem on bigger systems.
474 static unsigned long lazy_max_pages(void)
478 log = fls(num_online_cpus());
480 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
483 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
486 * Purges all lazily-freed vmap areas.
488 * If sync is 0 then don't purge if there is already a purge in progress.
489 * If force_flush is 1, then flush kernel TLBs between *start and *end even
490 * if we found no lazy vmap areas to unmap (callers can use this to optimise
491 * their own TLB flushing).
492 * Returns with *start = min(*start, lowest purged address)
493 * *end = max(*end, highest purged address)
495 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
496 int sync, int force_flush)
498 static DEFINE_SPINLOCK(purge_lock);
500 struct vmap_area *va;
504 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
505 * should not expect such behaviour. This just simplifies locking for
506 * the case that isn't actually used at the moment anyway.
508 if (!sync && !force_flush) {
509 if (!spin_trylock(&purge_lock))
512 spin_lock(&purge_lock);
515 list_for_each_entry_rcu(va, &vmap_area_list, list) {
516 if (va->flags & VM_LAZY_FREE) {
517 if (va->va_start < *start)
518 *start = va->va_start;
519 if (va->va_end > *end)
521 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
523 list_add_tail(&va->purge_list, &valist);
524 va->flags |= VM_LAZY_FREEING;
525 va->flags &= ~VM_LAZY_FREE;
531 BUG_ON(nr > atomic_read(&vmap_lazy_nr));
532 atomic_sub(nr, &vmap_lazy_nr);
535 if (nr || force_flush)
536 flush_tlb_kernel_range(*start, *end);
539 spin_lock(&vmap_area_lock);
540 list_for_each_entry(va, &valist, purge_list)
541 __free_vmap_area(va);
542 spin_unlock(&vmap_area_lock);
544 spin_unlock(&purge_lock);
548 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
549 * is already purging.
551 static void try_purge_vmap_area_lazy(void)
553 unsigned long start = ULONG_MAX, end = 0;
555 __purge_vmap_area_lazy(&start, &end, 0, 0);
559 * Kick off a purge of the outstanding lazy areas.
561 static void purge_vmap_area_lazy(void)
563 unsigned long start = ULONG_MAX, end = 0;
565 __purge_vmap_area_lazy(&start, &end, 1, 0);
569 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
570 * called for the correct range previously.
572 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
574 va->flags |= VM_LAZY_FREE;
575 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
576 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
577 try_purge_vmap_area_lazy();
581 * Free and unmap a vmap area
583 static void free_unmap_vmap_area(struct vmap_area *va)
585 flush_cache_vunmap(va->va_start, va->va_end);
586 free_unmap_vmap_area_noflush(va);
589 static struct vmap_area *find_vmap_area(unsigned long addr)
591 struct vmap_area *va;
593 spin_lock(&vmap_area_lock);
594 va = __find_vmap_area(addr);
595 spin_unlock(&vmap_area_lock);
600 static void free_unmap_vmap_area_addr(unsigned long addr)
602 struct vmap_area *va;
604 va = find_vmap_area(addr);
606 free_unmap_vmap_area(va);
610 /*** Per cpu kva allocator ***/
613 * vmap space is limited especially on 32 bit architectures. Ensure there is
614 * room for at least 16 percpu vmap blocks per CPU.
617 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
618 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
619 * instead (we just need a rough idea)
621 #if BITS_PER_LONG == 32
622 #define VMALLOC_SPACE (128UL*1024*1024)
624 #define VMALLOC_SPACE (128UL*1024*1024*1024)
627 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
628 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
629 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
630 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
631 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
632 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
633 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
634 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
635 VMALLOC_PAGES / NR_CPUS / 16))
637 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
639 static bool vmap_initialized __read_mostly = false;
641 struct vmap_block_queue {
643 struct list_head free;
644 struct list_head dirty;
645 unsigned int nr_dirty;
650 struct vmap_area *va;
651 struct vmap_block_queue *vbq;
652 unsigned long free, dirty;
653 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
654 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
657 struct list_head free_list;
658 struct list_head dirty_list;
660 struct rcu_head rcu_head;
664 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
665 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
668 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
669 * in the free path. Could get rid of this if we change the API to return a
670 * "cookie" from alloc, to be passed to free. But no big deal yet.
672 static DEFINE_SPINLOCK(vmap_block_tree_lock);
673 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
676 * We should probably have a fallback mechanism to allocate virtual memory
677 * out of partially filled vmap blocks. However vmap block sizing should be
678 * fairly reasonable according to the vmalloc size, so it shouldn't be a
682 static unsigned long addr_to_vb_idx(unsigned long addr)
684 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
685 addr /= VMAP_BLOCK_SIZE;
689 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
691 struct vmap_block_queue *vbq;
692 struct vmap_block *vb;
693 struct vmap_area *va;
694 unsigned long vb_idx;
697 node = numa_node_id();
699 vb = kmalloc_node(sizeof(struct vmap_block),
700 gfp_mask & GFP_RECLAIM_MASK, node);
702 return ERR_PTR(-ENOMEM);
704 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
705 VMALLOC_START, VMALLOC_END,
707 if (unlikely(IS_ERR(va))) {
709 return ERR_PTR(PTR_ERR(va));
712 err = radix_tree_preload(gfp_mask);
719 spin_lock_init(&vb->lock);
721 vb->free = VMAP_BBMAP_BITS;
723 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
724 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
725 INIT_LIST_HEAD(&vb->free_list);
726 INIT_LIST_HEAD(&vb->dirty_list);
728 vb_idx = addr_to_vb_idx(va->va_start);
729 spin_lock(&vmap_block_tree_lock);
730 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
731 spin_unlock(&vmap_block_tree_lock);
733 radix_tree_preload_end();
735 vbq = &get_cpu_var(vmap_block_queue);
737 spin_lock(&vbq->lock);
738 list_add(&vb->free_list, &vbq->free);
739 spin_unlock(&vbq->lock);
740 put_cpu_var(vmap_cpu_blocks);
745 static void rcu_free_vb(struct rcu_head *head)
747 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
752 static void free_vmap_block(struct vmap_block *vb)
754 struct vmap_block *tmp;
755 unsigned long vb_idx;
757 spin_lock(&vb->vbq->lock);
758 if (!list_empty(&vb->free_list))
759 list_del(&vb->free_list);
760 if (!list_empty(&vb->dirty_list))
761 list_del(&vb->dirty_list);
762 spin_unlock(&vb->vbq->lock);
764 vb_idx = addr_to_vb_idx(vb->va->va_start);
765 spin_lock(&vmap_block_tree_lock);
766 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
767 spin_unlock(&vmap_block_tree_lock);
770 free_unmap_vmap_area_noflush(vb->va);
771 call_rcu(&vb->rcu_head, rcu_free_vb);
774 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
776 struct vmap_block_queue *vbq;
777 struct vmap_block *vb;
778 unsigned long addr = 0;
781 BUG_ON(size & ~PAGE_MASK);
782 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
783 order = get_order(size);
787 vbq = &get_cpu_var(vmap_block_queue);
788 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
791 spin_lock(&vb->lock);
792 i = bitmap_find_free_region(vb->alloc_map,
793 VMAP_BBMAP_BITS, order);
796 addr = vb->va->va_start + (i << PAGE_SHIFT);
797 BUG_ON(addr_to_vb_idx(addr) !=
798 addr_to_vb_idx(vb->va->va_start));
799 vb->free -= 1UL << order;
801 spin_lock(&vbq->lock);
802 list_del_init(&vb->free_list);
803 spin_unlock(&vbq->lock);
805 spin_unlock(&vb->lock);
808 spin_unlock(&vb->lock);
810 put_cpu_var(vmap_cpu_blocks);
814 vb = new_vmap_block(gfp_mask);
823 static void vb_free(const void *addr, unsigned long size)
825 unsigned long offset;
826 unsigned long vb_idx;
828 struct vmap_block *vb;
830 BUG_ON(size & ~PAGE_MASK);
831 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
833 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
835 order = get_order(size);
837 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
839 vb_idx = addr_to_vb_idx((unsigned long)addr);
841 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
845 spin_lock(&vb->lock);
846 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
848 spin_lock(&vb->vbq->lock);
849 list_add(&vb->dirty_list, &vb->vbq->dirty);
850 spin_unlock(&vb->vbq->lock);
852 vb->dirty += 1UL << order;
853 if (vb->dirty == VMAP_BBMAP_BITS) {
854 BUG_ON(vb->free || !list_empty(&vb->free_list));
855 spin_unlock(&vb->lock);
858 spin_unlock(&vb->lock);
862 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
864 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
865 * to amortize TLB flushing overheads. What this means is that any page you
866 * have now, may, in a former life, have been mapped into kernel virtual
867 * address by the vmap layer and so there might be some CPUs with TLB entries
868 * still referencing that page (additional to the regular 1:1 kernel mapping).
870 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
871 * be sure that none of the pages we have control over will have any aliases
872 * from the vmap layer.
874 void vm_unmap_aliases(void)
876 unsigned long start = ULONG_MAX, end = 0;
880 if (unlikely(!vmap_initialized))
883 for_each_possible_cpu(cpu) {
884 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
885 struct vmap_block *vb;
888 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
891 spin_lock(&vb->lock);
892 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
893 while (i < VMAP_BBMAP_BITS) {
896 j = find_next_zero_bit(vb->dirty_map,
899 s = vb->va->va_start + (i << PAGE_SHIFT);
900 e = vb->va->va_start + (j << PAGE_SHIFT);
901 vunmap_page_range(s, e);
910 i = find_next_bit(vb->dirty_map,
913 spin_unlock(&vb->lock);
918 __purge_vmap_area_lazy(&start, &end, 1, flush);
920 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
923 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
924 * @mem: the pointer returned by vm_map_ram
925 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
927 void vm_unmap_ram(const void *mem, unsigned int count)
929 unsigned long size = count << PAGE_SHIFT;
930 unsigned long addr = (unsigned long)mem;
933 BUG_ON(addr < VMALLOC_START);
934 BUG_ON(addr > VMALLOC_END);
935 BUG_ON(addr & (PAGE_SIZE-1));
937 debug_check_no_locks_freed(mem, size);
938 vmap_debug_free_range(addr, addr+size);
940 if (likely(count <= VMAP_MAX_ALLOC))
943 free_unmap_vmap_area_addr(addr);
945 EXPORT_SYMBOL(vm_unmap_ram);
948 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
949 * @pages: an array of pointers to the pages to be mapped
950 * @count: number of pages
951 * @node: prefer to allocate data structures on this node
952 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
954 * Returns: a pointer to the address that has been mapped, or %NULL on failure
956 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
958 unsigned long size = count << PAGE_SHIFT;
962 if (likely(count <= VMAP_MAX_ALLOC)) {
963 mem = vb_alloc(size, GFP_KERNEL);
966 addr = (unsigned long)mem;
968 struct vmap_area *va;
969 va = alloc_vmap_area(size, PAGE_SIZE,
970 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
977 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
978 vm_unmap_ram(mem, count);
983 EXPORT_SYMBOL(vm_map_ram);
985 void __init vmalloc_init(void)
987 struct vmap_area *va;
988 struct vm_struct *tmp;
991 for_each_possible_cpu(i) {
992 struct vmap_block_queue *vbq;
994 vbq = &per_cpu(vmap_block_queue, i);
995 spin_lock_init(&vbq->lock);
996 INIT_LIST_HEAD(&vbq->free);
997 INIT_LIST_HEAD(&vbq->dirty);
1001 /* Import existing vmlist entries. */
1002 for (tmp = vmlist; tmp; tmp = tmp->next) {
1003 va = alloc_bootmem(sizeof(struct vmap_area));
1004 va->flags = tmp->flags | VM_VM_AREA;
1005 va->va_start = (unsigned long)tmp->addr;
1006 va->va_end = va->va_start + tmp->size;
1007 __insert_vmap_area(va);
1009 vmap_initialized = true;
1012 void unmap_kernel_range(unsigned long addr, unsigned long size)
1014 unsigned long end = addr + size;
1015 vunmap_page_range(addr, end);
1016 flush_tlb_kernel_range(addr, end);
1019 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1021 unsigned long addr = (unsigned long)area->addr;
1022 unsigned long end = addr + area->size - PAGE_SIZE;
1025 err = vmap_page_range(addr, end, prot, *pages);
1033 EXPORT_SYMBOL_GPL(map_vm_area);
1035 /*** Old vmalloc interfaces ***/
1036 DEFINE_RWLOCK(vmlist_lock);
1037 struct vm_struct *vmlist;
1039 static struct vm_struct *__get_vm_area_node(unsigned long size,
1040 unsigned long flags, unsigned long start, unsigned long end,
1041 int node, gfp_t gfp_mask, void *caller)
1043 static struct vmap_area *va;
1044 struct vm_struct *area;
1045 struct vm_struct *tmp, **p;
1046 unsigned long align = 1;
1048 BUG_ON(in_interrupt());
1049 if (flags & VM_IOREMAP) {
1050 int bit = fls(size);
1052 if (bit > IOREMAP_MAX_ORDER)
1053 bit = IOREMAP_MAX_ORDER;
1054 else if (bit < PAGE_SHIFT)
1060 size = PAGE_ALIGN(size);
1061 if (unlikely(!size))
1064 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1065 if (unlikely(!area))
1069 * We always allocate a guard page.
1073 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1079 area->flags = flags;
1080 area->addr = (void *)va->va_start;
1084 area->phys_addr = 0;
1085 area->caller = caller;
1087 va->flags |= VM_VM_AREA;
1089 write_lock(&vmlist_lock);
1090 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1091 if (tmp->addr >= area->addr)
1096 write_unlock(&vmlist_lock);
1101 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1102 unsigned long start, unsigned long end)
1104 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1105 __builtin_return_address(0));
1107 EXPORT_SYMBOL_GPL(__get_vm_area);
1110 * get_vm_area - reserve a contiguous kernel virtual area
1111 * @size: size of the area
1112 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1114 * Search an area of @size in the kernel virtual mapping area,
1115 * and reserved it for out purposes. Returns the area descriptor
1116 * on success or %NULL on failure.
1118 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1120 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1121 -1, GFP_KERNEL, __builtin_return_address(0));
1124 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1127 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1128 -1, GFP_KERNEL, caller);
1131 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1132 int node, gfp_t gfp_mask)
1134 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1135 gfp_mask, __builtin_return_address(0));
1138 static struct vm_struct *find_vm_area(const void *addr)
1140 struct vmap_area *va;
1142 va = find_vmap_area((unsigned long)addr);
1143 if (va && va->flags & VM_VM_AREA)
1150 * remove_vm_area - find and remove a continuous kernel virtual area
1151 * @addr: base address
1153 * Search for the kernel VM area starting at @addr, and remove it.
1154 * This function returns the found VM area, but using it is NOT safe
1155 * on SMP machines, except for its size or flags.
1157 struct vm_struct *remove_vm_area(const void *addr)
1159 struct vmap_area *va;
1161 va = find_vmap_area((unsigned long)addr);
1162 if (va && va->flags & VM_VM_AREA) {
1163 struct vm_struct *vm = va->private;
1164 struct vm_struct *tmp, **p;
1166 vmap_debug_free_range(va->va_start, va->va_end);
1167 free_unmap_vmap_area(va);
1168 vm->size -= PAGE_SIZE;
1170 write_lock(&vmlist_lock);
1171 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1174 write_unlock(&vmlist_lock);
1181 static void __vunmap(const void *addr, int deallocate_pages)
1183 struct vm_struct *area;
1188 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1189 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1193 area = remove_vm_area(addr);
1194 if (unlikely(!area)) {
1195 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1200 debug_check_no_locks_freed(addr, area->size);
1201 debug_check_no_obj_freed(addr, area->size);
1203 if (deallocate_pages) {
1206 for (i = 0; i < area->nr_pages; i++) {
1207 struct page *page = area->pages[i];
1213 if (area->flags & VM_VPAGES)
1224 * vfree - release memory allocated by vmalloc()
1225 * @addr: memory base address
1227 * Free the virtually continuous memory area starting at @addr, as
1228 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1229 * NULL, no operation is performed.
1231 * Must not be called in interrupt context.
1233 void vfree(const void *addr)
1235 BUG_ON(in_interrupt());
1238 EXPORT_SYMBOL(vfree);
1241 * vunmap - release virtual mapping obtained by vmap()
1242 * @addr: memory base address
1244 * Free the virtually contiguous memory area starting at @addr,
1245 * which was created from the page array passed to vmap().
1247 * Must not be called in interrupt context.
1249 void vunmap(const void *addr)
1251 BUG_ON(in_interrupt());
1254 EXPORT_SYMBOL(vunmap);
1257 * vmap - map an array of pages into virtually contiguous space
1258 * @pages: array of page pointers
1259 * @count: number of pages to map
1260 * @flags: vm_area->flags
1261 * @prot: page protection for the mapping
1263 * Maps @count pages from @pages into contiguous kernel virtual
1266 void *vmap(struct page **pages, unsigned int count,
1267 unsigned long flags, pgprot_t prot)
1269 struct vm_struct *area;
1271 if (count > num_physpages)
1274 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1275 __builtin_return_address(0));
1279 if (map_vm_area(area, prot, &pages)) {
1286 EXPORT_SYMBOL(vmap);
1288 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1289 int node, void *caller);
1290 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1291 pgprot_t prot, int node, void *caller)
1293 struct page **pages;
1294 unsigned int nr_pages, array_size, i;
1296 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1297 array_size = (nr_pages * sizeof(struct page *));
1299 area->nr_pages = nr_pages;
1300 /* Please note that the recursion is strictly bounded. */
1301 if (array_size > PAGE_SIZE) {
1302 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1303 PAGE_KERNEL, node, caller);
1304 area->flags |= VM_VPAGES;
1306 pages = kmalloc_node(array_size,
1307 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1310 area->pages = pages;
1311 area->caller = caller;
1313 remove_vm_area(area->addr);
1318 for (i = 0; i < area->nr_pages; i++) {
1322 page = alloc_page(gfp_mask);
1324 page = alloc_pages_node(node, gfp_mask, 0);
1326 if (unlikely(!page)) {
1327 /* Successfully allocated i pages, free them in __vunmap() */
1331 area->pages[i] = page;
1334 if (map_vm_area(area, prot, &pages))
1343 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1345 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1346 __builtin_return_address(0));
1350 * __vmalloc_node - allocate virtually contiguous memory
1351 * @size: allocation size
1352 * @gfp_mask: flags for the page level allocator
1353 * @prot: protection mask for the allocated pages
1354 * @node: node to use for allocation or -1
1355 * @caller: caller's return address
1357 * Allocate enough pages to cover @size from the page level
1358 * allocator with @gfp_mask flags. Map them into contiguous
1359 * kernel virtual space, using a pagetable protection of @prot.
1361 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1362 int node, void *caller)
1364 struct vm_struct *area;
1366 size = PAGE_ALIGN(size);
1367 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1370 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1371 node, gfp_mask, caller);
1376 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1379 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1381 return __vmalloc_node(size, gfp_mask, prot, -1,
1382 __builtin_return_address(0));
1384 EXPORT_SYMBOL(__vmalloc);
1387 * vmalloc - allocate virtually contiguous memory
1388 * @size: allocation size
1389 * Allocate enough pages to cover @size from the page level
1390 * allocator and map them into contiguous kernel virtual space.
1392 * For tight control over page level allocator and protection flags
1393 * use __vmalloc() instead.
1395 void *vmalloc(unsigned long size)
1397 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1398 -1, __builtin_return_address(0));
1400 EXPORT_SYMBOL(vmalloc);
1403 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1404 * @size: allocation size
1406 * The resulting memory area is zeroed so it can be mapped to userspace
1407 * without leaking data.
1409 void *vmalloc_user(unsigned long size)
1411 struct vm_struct *area;
1414 ret = __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1415 PAGE_KERNEL, -1, __builtin_return_address(0));
1417 area = find_vm_area(ret);
1418 area->flags |= VM_USERMAP;
1422 EXPORT_SYMBOL(vmalloc_user);
1425 * vmalloc_node - allocate memory on a specific node
1426 * @size: allocation size
1429 * Allocate enough pages to cover @size from the page level
1430 * allocator and map them into contiguous kernel virtual space.
1432 * For tight control over page level allocator and protection flags
1433 * use __vmalloc() instead.
1435 void *vmalloc_node(unsigned long size, int node)
1437 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1438 node, __builtin_return_address(0));
1440 EXPORT_SYMBOL(vmalloc_node);
1442 #ifndef PAGE_KERNEL_EXEC
1443 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1447 * vmalloc_exec - allocate virtually contiguous, executable memory
1448 * @size: allocation size
1450 * Kernel-internal function to allocate enough pages to cover @size
1451 * the page level allocator and map them into contiguous and
1452 * executable kernel virtual space.
1454 * For tight control over page level allocator and protection flags
1455 * use __vmalloc() instead.
1458 void *vmalloc_exec(unsigned long size)
1460 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1461 -1, __builtin_return_address(0));
1464 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1465 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1466 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1467 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1469 #define GFP_VMALLOC32 GFP_KERNEL
1473 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1474 * @size: allocation size
1476 * Allocate enough 32bit PA addressable pages to cover @size from the
1477 * page level allocator and map them into contiguous kernel virtual space.
1479 void *vmalloc_32(unsigned long size)
1481 return __vmalloc_node(size, GFP_VMALLOC32, PAGE_KERNEL,
1482 -1, __builtin_return_address(0));
1484 EXPORT_SYMBOL(vmalloc_32);
1487 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1488 * @size: allocation size
1490 * The resulting memory area is 32bit addressable and zeroed so it can be
1491 * mapped to userspace without leaking data.
1493 void *vmalloc_32_user(unsigned long size)
1495 struct vm_struct *area;
1498 ret = __vmalloc_node(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1499 -1, __builtin_return_address(0));
1501 area = find_vm_area(ret);
1502 area->flags |= VM_USERMAP;
1506 EXPORT_SYMBOL(vmalloc_32_user);
1508 long vread(char *buf, char *addr, unsigned long count)
1510 struct vm_struct *tmp;
1511 char *vaddr, *buf_start = buf;
1514 /* Don't allow overflow */
1515 if ((unsigned long) addr + count < count)
1516 count = -(unsigned long) addr;
1518 read_lock(&vmlist_lock);
1519 for (tmp = vmlist; tmp; tmp = tmp->next) {
1520 vaddr = (char *) tmp->addr;
1521 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1523 while (addr < vaddr) {
1531 n = vaddr + tmp->size - PAGE_SIZE - addr;
1542 read_unlock(&vmlist_lock);
1543 return buf - buf_start;
1546 long vwrite(char *buf, char *addr, unsigned long count)
1548 struct vm_struct *tmp;
1549 char *vaddr, *buf_start = buf;
1552 /* Don't allow overflow */
1553 if ((unsigned long) addr + count < count)
1554 count = -(unsigned long) addr;
1556 read_lock(&vmlist_lock);
1557 for (tmp = vmlist; tmp; tmp = tmp->next) {
1558 vaddr = (char *) tmp->addr;
1559 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1561 while (addr < vaddr) {
1568 n = vaddr + tmp->size - PAGE_SIZE - addr;
1579 read_unlock(&vmlist_lock);
1580 return buf - buf_start;
1584 * remap_vmalloc_range - map vmalloc pages to userspace
1585 * @vma: vma to cover (map full range of vma)
1586 * @addr: vmalloc memory
1587 * @pgoff: number of pages into addr before first page to map
1589 * Returns: 0 for success, -Exxx on failure
1591 * This function checks that addr is a valid vmalloc'ed area, and
1592 * that it is big enough to cover the vma. Will return failure if
1593 * that criteria isn't met.
1595 * Similar to remap_pfn_range() (see mm/memory.c)
1597 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1598 unsigned long pgoff)
1600 struct vm_struct *area;
1601 unsigned long uaddr = vma->vm_start;
1602 unsigned long usize = vma->vm_end - vma->vm_start;
1604 if ((PAGE_SIZE-1) & (unsigned long)addr)
1607 area = find_vm_area(addr);
1611 if (!(area->flags & VM_USERMAP))
1614 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1617 addr += pgoff << PAGE_SHIFT;
1619 struct page *page = vmalloc_to_page(addr);
1622 ret = vm_insert_page(vma, uaddr, page);
1629 } while (usize > 0);
1631 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1632 vma->vm_flags |= VM_RESERVED;
1636 EXPORT_SYMBOL(remap_vmalloc_range);
1639 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1642 void __attribute__((weak)) vmalloc_sync_all(void)
1647 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1649 /* apply_to_page_range() does all the hard work. */
1654 * alloc_vm_area - allocate a range of kernel address space
1655 * @size: size of the area
1657 * Returns: NULL on failure, vm_struct on success
1659 * This function reserves a range of kernel address space, and
1660 * allocates pagetables to map that range. No actual mappings
1661 * are created. If the kernel address space is not shared
1662 * between processes, it syncs the pagetable across all
1665 struct vm_struct *alloc_vm_area(size_t size)
1667 struct vm_struct *area;
1669 area = get_vm_area_caller(size, VM_IOREMAP,
1670 __builtin_return_address(0));
1675 * This ensures that page tables are constructed for this region
1676 * of kernel virtual address space and mapped into init_mm.
1678 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1679 area->size, f, NULL)) {
1684 /* Make sure the pagetables are constructed in process kernel
1690 EXPORT_SYMBOL_GPL(alloc_vm_area);
1692 void free_vm_area(struct vm_struct *area)
1694 struct vm_struct *ret;
1695 ret = remove_vm_area(area->addr);
1696 BUG_ON(ret != area);
1699 EXPORT_SYMBOL_GPL(free_vm_area);
1702 #ifdef CONFIG_PROC_FS
1703 static void *s_start(struct seq_file *m, loff_t *pos)
1706 struct vm_struct *v;
1708 read_lock(&vmlist_lock);
1710 while (n > 0 && v) {
1721 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1723 struct vm_struct *v = p;
1729 static void s_stop(struct seq_file *m, void *p)
1731 read_unlock(&vmlist_lock);
1734 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1737 unsigned int nr, *counters = m->private;
1742 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1744 for (nr = 0; nr < v->nr_pages; nr++)
1745 counters[page_to_nid(v->pages[nr])]++;
1747 for_each_node_state(nr, N_HIGH_MEMORY)
1749 seq_printf(m, " N%u=%u", nr, counters[nr]);
1753 static int s_show(struct seq_file *m, void *p)
1755 struct vm_struct *v = p;
1757 seq_printf(m, "0x%p-0x%p %7ld",
1758 v->addr, v->addr + v->size, v->size);
1761 char buff[KSYM_SYMBOL_LEN];
1764 sprint_symbol(buff, (unsigned long)v->caller);
1769 seq_printf(m, " pages=%d", v->nr_pages);
1772 seq_printf(m, " phys=%lx", v->phys_addr);
1774 if (v->flags & VM_IOREMAP)
1775 seq_printf(m, " ioremap");
1777 if (v->flags & VM_ALLOC)
1778 seq_printf(m, " vmalloc");
1780 if (v->flags & VM_MAP)
1781 seq_printf(m, " vmap");
1783 if (v->flags & VM_USERMAP)
1784 seq_printf(m, " user");
1786 if (v->flags & VM_VPAGES)
1787 seq_printf(m, " vpages");
1789 show_numa_info(m, v);
1794 static const struct seq_operations vmalloc_op = {
1801 static int vmalloc_open(struct inode *inode, struct file *file)
1803 unsigned int *ptr = NULL;
1807 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1808 ret = seq_open(file, &vmalloc_op);
1810 struct seq_file *m = file->private_data;
1817 static const struct file_operations proc_vmalloc_operations = {
1818 .open = vmalloc_open,
1820 .llseek = seq_lseek,
1821 .release = seq_release_private,
1824 static int __init proc_vmalloc_init(void)
1826 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1829 module_init(proc_vmalloc_init);