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;
327 BUG_ON(size & ~PAGE_MASK);
329 va = kmalloc_node(sizeof(struct vmap_area),
330 gfp_mask & GFP_RECLAIM_MASK, node);
332 return ERR_PTR(-ENOMEM);
335 addr = ALIGN(vstart, align);
337 spin_lock(&vmap_area_lock);
338 if (addr + size - 1 < addr)
341 /* XXX: could have a last_hole cache */
342 n = vmap_area_root.rb_node;
344 struct vmap_area *first = NULL;
347 struct vmap_area *tmp;
348 tmp = rb_entry(n, struct vmap_area, rb_node);
349 if (tmp->va_end >= addr) {
350 if (!first && tmp->va_start < addr + size)
362 if (first->va_end < addr) {
363 n = rb_next(&first->rb_node);
365 first = rb_entry(n, struct vmap_area, rb_node);
370 while (addr + size > first->va_start && addr + size <= vend) {
371 addr = ALIGN(first->va_end + PAGE_SIZE, align);
372 if (addr + size - 1 < addr)
375 n = rb_next(&first->rb_node);
377 first = rb_entry(n, struct vmap_area, rb_node);
383 if (addr + size > vend) {
385 spin_unlock(&vmap_area_lock);
387 purge_vmap_area_lazy();
391 if (printk_ratelimit())
393 "vmap allocation for size %lu failed: "
394 "use vmalloc=<size> to increase size.\n", size);
395 return ERR_PTR(-EBUSY);
398 BUG_ON(addr & (align-1));
401 va->va_end = addr + size;
403 __insert_vmap_area(va);
404 spin_unlock(&vmap_area_lock);
409 static void rcu_free_va(struct rcu_head *head)
411 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
416 static void __free_vmap_area(struct vmap_area *va)
418 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
419 rb_erase(&va->rb_node, &vmap_area_root);
420 RB_CLEAR_NODE(&va->rb_node);
421 list_del_rcu(&va->list);
423 call_rcu(&va->rcu_head, rcu_free_va);
427 * Free a region of KVA allocated by alloc_vmap_area
429 static void free_vmap_area(struct vmap_area *va)
431 spin_lock(&vmap_area_lock);
432 __free_vmap_area(va);
433 spin_unlock(&vmap_area_lock);
437 * Clear the pagetable entries of a given vmap_area
439 static void unmap_vmap_area(struct vmap_area *va)
441 vunmap_page_range(va->va_start, va->va_end);
444 static void vmap_debug_free_range(unsigned long start, unsigned long end)
447 * Unmap page tables and force a TLB flush immediately if
448 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
449 * bugs similarly to those in linear kernel virtual address
450 * space after a page has been freed.
452 * All the lazy freeing logic is still retained, in order to
453 * minimise intrusiveness of this debugging feature.
455 * This is going to be *slow* (linear kernel virtual address
456 * debugging doesn't do a broadcast TLB flush so it is a lot
459 #ifdef CONFIG_DEBUG_PAGEALLOC
460 vunmap_page_range(start, end);
461 flush_tlb_kernel_range(start, end);
466 * lazy_max_pages is the maximum amount of virtual address space we gather up
467 * before attempting to purge with a TLB flush.
469 * There is a tradeoff here: a larger number will cover more kernel page tables
470 * and take slightly longer to purge, but it will linearly reduce the number of
471 * global TLB flushes that must be performed. It would seem natural to scale
472 * this number up linearly with the number of CPUs (because vmapping activity
473 * could also scale linearly with the number of CPUs), however it is likely
474 * that in practice, workloads might be constrained in other ways that mean
475 * vmap activity will not scale linearly with CPUs. Also, I want to be
476 * conservative and not introduce a big latency on huge systems, so go with
477 * a less aggressive log scale. It will still be an improvement over the old
478 * code, and it will be simple to change the scale factor if we find that it
479 * becomes a problem on bigger systems.
481 static unsigned long lazy_max_pages(void)
485 log = fls(num_online_cpus());
487 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
490 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
493 * Purges all lazily-freed vmap areas.
495 * If sync is 0 then don't purge if there is already a purge in progress.
496 * If force_flush is 1, then flush kernel TLBs between *start and *end even
497 * if we found no lazy vmap areas to unmap (callers can use this to optimise
498 * their own TLB flushing).
499 * Returns with *start = min(*start, lowest purged address)
500 * *end = max(*end, highest purged address)
502 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
503 int sync, int force_flush)
505 static DEFINE_SPINLOCK(purge_lock);
507 struct vmap_area *va;
511 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
512 * should not expect such behaviour. This just simplifies locking for
513 * the case that isn't actually used at the moment anyway.
515 if (!sync && !force_flush) {
516 if (!spin_trylock(&purge_lock))
519 spin_lock(&purge_lock);
522 list_for_each_entry_rcu(va, &vmap_area_list, list) {
523 if (va->flags & VM_LAZY_FREE) {
524 if (va->va_start < *start)
525 *start = va->va_start;
526 if (va->va_end > *end)
528 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
530 list_add_tail(&va->purge_list, &valist);
531 va->flags |= VM_LAZY_FREEING;
532 va->flags &= ~VM_LAZY_FREE;
538 BUG_ON(nr > atomic_read(&vmap_lazy_nr));
539 atomic_sub(nr, &vmap_lazy_nr);
542 if (nr || force_flush)
543 flush_tlb_kernel_range(*start, *end);
546 spin_lock(&vmap_area_lock);
547 list_for_each_entry(va, &valist, purge_list)
548 __free_vmap_area(va);
549 spin_unlock(&vmap_area_lock);
551 spin_unlock(&purge_lock);
555 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
556 * is already purging.
558 static void try_purge_vmap_area_lazy(void)
560 unsigned long start = ULONG_MAX, end = 0;
562 __purge_vmap_area_lazy(&start, &end, 0, 0);
566 * Kick off a purge of the outstanding lazy areas.
568 static void purge_vmap_area_lazy(void)
570 unsigned long start = ULONG_MAX, end = 0;
572 __purge_vmap_area_lazy(&start, &end, 1, 0);
576 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
577 * called for the correct range previously.
579 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
581 va->flags |= VM_LAZY_FREE;
582 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
583 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
584 try_purge_vmap_area_lazy();
588 * Free and unmap a vmap area
590 static void free_unmap_vmap_area(struct vmap_area *va)
592 flush_cache_vunmap(va->va_start, va->va_end);
593 free_unmap_vmap_area_noflush(va);
596 static struct vmap_area *find_vmap_area(unsigned long addr)
598 struct vmap_area *va;
600 spin_lock(&vmap_area_lock);
601 va = __find_vmap_area(addr);
602 spin_unlock(&vmap_area_lock);
607 static void free_unmap_vmap_area_addr(unsigned long addr)
609 struct vmap_area *va;
611 va = find_vmap_area(addr);
613 free_unmap_vmap_area(va);
617 /*** Per cpu kva allocator ***/
620 * vmap space is limited especially on 32 bit architectures. Ensure there is
621 * room for at least 16 percpu vmap blocks per CPU.
624 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
625 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
626 * instead (we just need a rough idea)
628 #if BITS_PER_LONG == 32
629 #define VMALLOC_SPACE (128UL*1024*1024)
631 #define VMALLOC_SPACE (128UL*1024*1024*1024)
634 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
635 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
636 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
637 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
638 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
639 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
640 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
641 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
642 VMALLOC_PAGES / NR_CPUS / 16))
644 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
646 static bool vmap_initialized __read_mostly = false;
648 struct vmap_block_queue {
650 struct list_head free;
651 struct list_head dirty;
652 unsigned int nr_dirty;
657 struct vmap_area *va;
658 struct vmap_block_queue *vbq;
659 unsigned long free, dirty;
660 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
661 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
664 struct list_head free_list;
665 struct list_head dirty_list;
667 struct rcu_head rcu_head;
671 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
672 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
675 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
676 * in the free path. Could get rid of this if we change the API to return a
677 * "cookie" from alloc, to be passed to free. But no big deal yet.
679 static DEFINE_SPINLOCK(vmap_block_tree_lock);
680 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
683 * We should probably have a fallback mechanism to allocate virtual memory
684 * out of partially filled vmap blocks. However vmap block sizing should be
685 * fairly reasonable according to the vmalloc size, so it shouldn't be a
689 static unsigned long addr_to_vb_idx(unsigned long addr)
691 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
692 addr /= VMAP_BLOCK_SIZE;
696 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
698 struct vmap_block_queue *vbq;
699 struct vmap_block *vb;
700 struct vmap_area *va;
701 unsigned long vb_idx;
704 node = numa_node_id();
706 vb = kmalloc_node(sizeof(struct vmap_block),
707 gfp_mask & GFP_RECLAIM_MASK, node);
709 return ERR_PTR(-ENOMEM);
711 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
712 VMALLOC_START, VMALLOC_END,
714 if (unlikely(IS_ERR(va))) {
716 return ERR_PTR(PTR_ERR(va));
719 err = radix_tree_preload(gfp_mask);
726 spin_lock_init(&vb->lock);
728 vb->free = VMAP_BBMAP_BITS;
730 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
731 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
732 INIT_LIST_HEAD(&vb->free_list);
733 INIT_LIST_HEAD(&vb->dirty_list);
735 vb_idx = addr_to_vb_idx(va->va_start);
736 spin_lock(&vmap_block_tree_lock);
737 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
738 spin_unlock(&vmap_block_tree_lock);
740 radix_tree_preload_end();
742 vbq = &get_cpu_var(vmap_block_queue);
744 spin_lock(&vbq->lock);
745 list_add(&vb->free_list, &vbq->free);
746 spin_unlock(&vbq->lock);
747 put_cpu_var(vmap_cpu_blocks);
752 static void rcu_free_vb(struct rcu_head *head)
754 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
759 static void free_vmap_block(struct vmap_block *vb)
761 struct vmap_block *tmp;
762 unsigned long vb_idx;
764 spin_lock(&vb->vbq->lock);
765 if (!list_empty(&vb->free_list))
766 list_del(&vb->free_list);
767 if (!list_empty(&vb->dirty_list))
768 list_del(&vb->dirty_list);
769 spin_unlock(&vb->vbq->lock);
771 vb_idx = addr_to_vb_idx(vb->va->va_start);
772 spin_lock(&vmap_block_tree_lock);
773 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
774 spin_unlock(&vmap_block_tree_lock);
777 free_unmap_vmap_area_noflush(vb->va);
778 call_rcu(&vb->rcu_head, rcu_free_vb);
781 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
783 struct vmap_block_queue *vbq;
784 struct vmap_block *vb;
785 unsigned long addr = 0;
788 BUG_ON(size & ~PAGE_MASK);
789 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
790 order = get_order(size);
794 vbq = &get_cpu_var(vmap_block_queue);
795 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
798 spin_lock(&vb->lock);
799 i = bitmap_find_free_region(vb->alloc_map,
800 VMAP_BBMAP_BITS, order);
803 addr = vb->va->va_start + (i << PAGE_SHIFT);
804 BUG_ON(addr_to_vb_idx(addr) !=
805 addr_to_vb_idx(vb->va->va_start));
806 vb->free -= 1UL << order;
808 spin_lock(&vbq->lock);
809 list_del_init(&vb->free_list);
810 spin_unlock(&vbq->lock);
812 spin_unlock(&vb->lock);
815 spin_unlock(&vb->lock);
817 put_cpu_var(vmap_cpu_blocks);
821 vb = new_vmap_block(gfp_mask);
830 static void vb_free(const void *addr, unsigned long size)
832 unsigned long offset;
833 unsigned long vb_idx;
835 struct vmap_block *vb;
837 BUG_ON(size & ~PAGE_MASK);
838 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
840 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
842 order = get_order(size);
844 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
846 vb_idx = addr_to_vb_idx((unsigned long)addr);
848 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
852 spin_lock(&vb->lock);
853 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
855 spin_lock(&vb->vbq->lock);
856 list_add(&vb->dirty_list, &vb->vbq->dirty);
857 spin_unlock(&vb->vbq->lock);
859 vb->dirty += 1UL << order;
860 if (vb->dirty == VMAP_BBMAP_BITS) {
861 BUG_ON(vb->free || !list_empty(&vb->free_list));
862 spin_unlock(&vb->lock);
865 spin_unlock(&vb->lock);
869 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
871 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
872 * to amortize TLB flushing overheads. What this means is that any page you
873 * have now, may, in a former life, have been mapped into kernel virtual
874 * address by the vmap layer and so there might be some CPUs with TLB entries
875 * still referencing that page (additional to the regular 1:1 kernel mapping).
877 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
878 * be sure that none of the pages we have control over will have any aliases
879 * from the vmap layer.
881 void vm_unmap_aliases(void)
883 unsigned long start = ULONG_MAX, end = 0;
887 if (unlikely(!vmap_initialized))
890 for_each_possible_cpu(cpu) {
891 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
892 struct vmap_block *vb;
895 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
898 spin_lock(&vb->lock);
899 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
900 while (i < VMAP_BBMAP_BITS) {
903 j = find_next_zero_bit(vb->dirty_map,
906 s = vb->va->va_start + (i << PAGE_SHIFT);
907 e = vb->va->va_start + (j << PAGE_SHIFT);
908 vunmap_page_range(s, e);
917 i = find_next_bit(vb->dirty_map,
920 spin_unlock(&vb->lock);
925 __purge_vmap_area_lazy(&start, &end, 1, flush);
927 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
930 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
931 * @mem: the pointer returned by vm_map_ram
932 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
934 void vm_unmap_ram(const void *mem, unsigned int count)
936 unsigned long size = count << PAGE_SHIFT;
937 unsigned long addr = (unsigned long)mem;
940 BUG_ON(addr < VMALLOC_START);
941 BUG_ON(addr > VMALLOC_END);
942 BUG_ON(addr & (PAGE_SIZE-1));
944 debug_check_no_locks_freed(mem, size);
945 vmap_debug_free_range(addr, addr+size);
947 if (likely(count <= VMAP_MAX_ALLOC))
950 free_unmap_vmap_area_addr(addr);
952 EXPORT_SYMBOL(vm_unmap_ram);
955 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
956 * @pages: an array of pointers to the pages to be mapped
957 * @count: number of pages
958 * @node: prefer to allocate data structures on this node
959 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
961 * Returns: a pointer to the address that has been mapped, or %NULL on failure
963 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
965 unsigned long size = count << PAGE_SHIFT;
969 if (likely(count <= VMAP_MAX_ALLOC)) {
970 mem = vb_alloc(size, GFP_KERNEL);
973 addr = (unsigned long)mem;
975 struct vmap_area *va;
976 va = alloc_vmap_area(size, PAGE_SIZE,
977 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
984 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
985 vm_unmap_ram(mem, count);
990 EXPORT_SYMBOL(vm_map_ram);
992 void __init vmalloc_init(void)
994 struct vmap_area *va;
995 struct vm_struct *tmp;
998 for_each_possible_cpu(i) {
999 struct vmap_block_queue *vbq;
1001 vbq = &per_cpu(vmap_block_queue, i);
1002 spin_lock_init(&vbq->lock);
1003 INIT_LIST_HEAD(&vbq->free);
1004 INIT_LIST_HEAD(&vbq->dirty);
1008 /* Import existing vmlist entries. */
1009 for (tmp = vmlist; tmp; tmp = tmp->next) {
1010 va = alloc_bootmem(sizeof(struct vmap_area));
1011 va->flags = tmp->flags | VM_VM_AREA;
1012 va->va_start = (unsigned long)tmp->addr;
1013 va->va_end = va->va_start + tmp->size;
1014 __insert_vmap_area(va);
1016 vmap_initialized = true;
1019 void unmap_kernel_range(unsigned long addr, unsigned long size)
1021 unsigned long end = addr + size;
1023 flush_cache_vunmap(addr, end);
1024 vunmap_page_range(addr, end);
1025 flush_tlb_kernel_range(addr, end);
1028 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1030 unsigned long addr = (unsigned long)area->addr;
1031 unsigned long end = addr + area->size - PAGE_SIZE;
1034 err = vmap_page_range(addr, end, prot, *pages);
1042 EXPORT_SYMBOL_GPL(map_vm_area);
1044 /*** Old vmalloc interfaces ***/
1045 DEFINE_RWLOCK(vmlist_lock);
1046 struct vm_struct *vmlist;
1048 static struct vm_struct *__get_vm_area_node(unsigned long size,
1049 unsigned long flags, unsigned long start, unsigned long end,
1050 int node, gfp_t gfp_mask, void *caller)
1052 static struct vmap_area *va;
1053 struct vm_struct *area;
1054 struct vm_struct *tmp, **p;
1055 unsigned long align = 1;
1057 BUG_ON(in_interrupt());
1058 if (flags & VM_IOREMAP) {
1059 int bit = fls(size);
1061 if (bit > IOREMAP_MAX_ORDER)
1062 bit = IOREMAP_MAX_ORDER;
1063 else if (bit < PAGE_SHIFT)
1069 size = PAGE_ALIGN(size);
1070 if (unlikely(!size))
1073 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1074 if (unlikely(!area))
1078 * We always allocate a guard page.
1082 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1088 area->flags = flags;
1089 area->addr = (void *)va->va_start;
1093 area->phys_addr = 0;
1094 area->caller = caller;
1096 va->flags |= VM_VM_AREA;
1098 write_lock(&vmlist_lock);
1099 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1100 if (tmp->addr >= area->addr)
1105 write_unlock(&vmlist_lock);
1110 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1111 unsigned long start, unsigned long end)
1113 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1114 __builtin_return_address(0));
1116 EXPORT_SYMBOL_GPL(__get_vm_area);
1118 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1119 unsigned long start, unsigned long end,
1122 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1127 * get_vm_area - reserve a contiguous kernel virtual area
1128 * @size: size of the area
1129 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1131 * Search an area of @size in the kernel virtual mapping area,
1132 * and reserved it for out purposes. Returns the area descriptor
1133 * on success or %NULL on failure.
1135 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1137 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1138 -1, GFP_KERNEL, __builtin_return_address(0));
1141 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1144 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1145 -1, GFP_KERNEL, caller);
1148 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1149 int node, gfp_t gfp_mask)
1151 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1152 gfp_mask, __builtin_return_address(0));
1155 static struct vm_struct *find_vm_area(const void *addr)
1157 struct vmap_area *va;
1159 va = find_vmap_area((unsigned long)addr);
1160 if (va && va->flags & VM_VM_AREA)
1167 * remove_vm_area - find and remove a continuous kernel virtual area
1168 * @addr: base address
1170 * Search for the kernel VM area starting at @addr, and remove it.
1171 * This function returns the found VM area, but using it is NOT safe
1172 * on SMP machines, except for its size or flags.
1174 struct vm_struct *remove_vm_area(const void *addr)
1176 struct vmap_area *va;
1178 va = find_vmap_area((unsigned long)addr);
1179 if (va && va->flags & VM_VM_AREA) {
1180 struct vm_struct *vm = va->private;
1181 struct vm_struct *tmp, **p;
1183 vmap_debug_free_range(va->va_start, va->va_end);
1184 free_unmap_vmap_area(va);
1185 vm->size -= PAGE_SIZE;
1187 write_lock(&vmlist_lock);
1188 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1191 write_unlock(&vmlist_lock);
1198 static void __vunmap(const void *addr, int deallocate_pages)
1200 struct vm_struct *area;
1205 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1206 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1210 area = remove_vm_area(addr);
1211 if (unlikely(!area)) {
1212 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1217 debug_check_no_locks_freed(addr, area->size);
1218 debug_check_no_obj_freed(addr, area->size);
1220 if (deallocate_pages) {
1223 for (i = 0; i < area->nr_pages; i++) {
1224 struct page *page = area->pages[i];
1230 if (area->flags & VM_VPAGES)
1241 * vfree - release memory allocated by vmalloc()
1242 * @addr: memory base address
1244 * Free the virtually continuous memory area starting at @addr, as
1245 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1246 * NULL, no operation is performed.
1248 * Must not be called in interrupt context.
1250 void vfree(const void *addr)
1252 BUG_ON(in_interrupt());
1255 EXPORT_SYMBOL(vfree);
1258 * vunmap - release virtual mapping obtained by vmap()
1259 * @addr: memory base address
1261 * Free the virtually contiguous memory area starting at @addr,
1262 * which was created from the page array passed to vmap().
1264 * Must not be called in interrupt context.
1266 void vunmap(const void *addr)
1268 BUG_ON(in_interrupt());
1271 EXPORT_SYMBOL(vunmap);
1274 * vmap - map an array of pages into virtually contiguous space
1275 * @pages: array of page pointers
1276 * @count: number of pages to map
1277 * @flags: vm_area->flags
1278 * @prot: page protection for the mapping
1280 * Maps @count pages from @pages into contiguous kernel virtual
1283 void *vmap(struct page **pages, unsigned int count,
1284 unsigned long flags, pgprot_t prot)
1286 struct vm_struct *area;
1288 if (count > num_physpages)
1291 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1292 __builtin_return_address(0));
1296 if (map_vm_area(area, prot, &pages)) {
1303 EXPORT_SYMBOL(vmap);
1305 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1306 int node, void *caller);
1307 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1308 pgprot_t prot, int node, void *caller)
1310 struct page **pages;
1311 unsigned int nr_pages, array_size, i;
1313 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1314 array_size = (nr_pages * sizeof(struct page *));
1316 area->nr_pages = nr_pages;
1317 /* Please note that the recursion is strictly bounded. */
1318 if (array_size > PAGE_SIZE) {
1319 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1320 PAGE_KERNEL, node, caller);
1321 area->flags |= VM_VPAGES;
1323 pages = kmalloc_node(array_size,
1324 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1327 area->pages = pages;
1328 area->caller = caller;
1330 remove_vm_area(area->addr);
1335 for (i = 0; i < area->nr_pages; i++) {
1339 page = alloc_page(gfp_mask);
1341 page = alloc_pages_node(node, gfp_mask, 0);
1343 if (unlikely(!page)) {
1344 /* Successfully allocated i pages, free them in __vunmap() */
1348 area->pages[i] = page;
1351 if (map_vm_area(area, prot, &pages))
1360 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1362 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1363 __builtin_return_address(0));
1367 * __vmalloc_node - allocate virtually contiguous memory
1368 * @size: allocation size
1369 * @gfp_mask: flags for the page level allocator
1370 * @prot: protection mask for the allocated pages
1371 * @node: node to use for allocation or -1
1372 * @caller: caller's return address
1374 * Allocate enough pages to cover @size from the page level
1375 * allocator with @gfp_mask flags. Map them into contiguous
1376 * kernel virtual space, using a pagetable protection of @prot.
1378 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1379 int node, void *caller)
1381 struct vm_struct *area;
1383 size = PAGE_ALIGN(size);
1384 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1387 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1388 node, gfp_mask, caller);
1393 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1396 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1398 return __vmalloc_node(size, gfp_mask, prot, -1,
1399 __builtin_return_address(0));
1401 EXPORT_SYMBOL(__vmalloc);
1404 * vmalloc - allocate virtually contiguous memory
1405 * @size: allocation size
1406 * Allocate enough pages to cover @size from the page level
1407 * allocator and map them into contiguous kernel virtual space.
1409 * For tight control over page level allocator and protection flags
1410 * use __vmalloc() instead.
1412 void *vmalloc(unsigned long size)
1414 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1415 -1, __builtin_return_address(0));
1417 EXPORT_SYMBOL(vmalloc);
1420 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1421 * @size: allocation size
1423 * The resulting memory area is zeroed so it can be mapped to userspace
1424 * without leaking data.
1426 void *vmalloc_user(unsigned long size)
1428 struct vm_struct *area;
1431 ret = __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1432 PAGE_KERNEL, -1, __builtin_return_address(0));
1434 area = find_vm_area(ret);
1435 area->flags |= VM_USERMAP;
1439 EXPORT_SYMBOL(vmalloc_user);
1442 * vmalloc_node - allocate memory on a specific node
1443 * @size: allocation size
1446 * Allocate enough pages to cover @size from the page level
1447 * allocator and map them into contiguous kernel virtual space.
1449 * For tight control over page level allocator and protection flags
1450 * use __vmalloc() instead.
1452 void *vmalloc_node(unsigned long size, int node)
1454 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1455 node, __builtin_return_address(0));
1457 EXPORT_SYMBOL(vmalloc_node);
1459 #ifndef PAGE_KERNEL_EXEC
1460 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1464 * vmalloc_exec - allocate virtually contiguous, executable memory
1465 * @size: allocation size
1467 * Kernel-internal function to allocate enough pages to cover @size
1468 * the page level allocator and map them into contiguous and
1469 * executable kernel virtual space.
1471 * For tight control over page level allocator and protection flags
1472 * use __vmalloc() instead.
1475 void *vmalloc_exec(unsigned long size)
1477 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1478 -1, __builtin_return_address(0));
1481 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1482 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1483 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1484 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1486 #define GFP_VMALLOC32 GFP_KERNEL
1490 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1491 * @size: allocation size
1493 * Allocate enough 32bit PA addressable pages to cover @size from the
1494 * page level allocator and map them into contiguous kernel virtual space.
1496 void *vmalloc_32(unsigned long size)
1498 return __vmalloc_node(size, GFP_VMALLOC32, PAGE_KERNEL,
1499 -1, __builtin_return_address(0));
1501 EXPORT_SYMBOL(vmalloc_32);
1504 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1505 * @size: allocation size
1507 * The resulting memory area is 32bit addressable and zeroed so it can be
1508 * mapped to userspace without leaking data.
1510 void *vmalloc_32_user(unsigned long size)
1512 struct vm_struct *area;
1515 ret = __vmalloc_node(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1516 -1, __builtin_return_address(0));
1518 area = find_vm_area(ret);
1519 area->flags |= VM_USERMAP;
1523 EXPORT_SYMBOL(vmalloc_32_user);
1525 long vread(char *buf, char *addr, unsigned long count)
1527 struct vm_struct *tmp;
1528 char *vaddr, *buf_start = buf;
1531 /* Don't allow overflow */
1532 if ((unsigned long) addr + count < count)
1533 count = -(unsigned long) addr;
1535 read_lock(&vmlist_lock);
1536 for (tmp = vmlist; tmp; tmp = tmp->next) {
1537 vaddr = (char *) tmp->addr;
1538 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1540 while (addr < vaddr) {
1548 n = vaddr + tmp->size - PAGE_SIZE - addr;
1559 read_unlock(&vmlist_lock);
1560 return buf - buf_start;
1563 long vwrite(char *buf, char *addr, unsigned long count)
1565 struct vm_struct *tmp;
1566 char *vaddr, *buf_start = buf;
1569 /* Don't allow overflow */
1570 if ((unsigned long) addr + count < count)
1571 count = -(unsigned long) addr;
1573 read_lock(&vmlist_lock);
1574 for (tmp = vmlist; tmp; tmp = tmp->next) {
1575 vaddr = (char *) tmp->addr;
1576 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1578 while (addr < vaddr) {
1585 n = vaddr + tmp->size - PAGE_SIZE - addr;
1596 read_unlock(&vmlist_lock);
1597 return buf - buf_start;
1601 * remap_vmalloc_range - map vmalloc pages to userspace
1602 * @vma: vma to cover (map full range of vma)
1603 * @addr: vmalloc memory
1604 * @pgoff: number of pages into addr before first page to map
1606 * Returns: 0 for success, -Exxx on failure
1608 * This function checks that addr is a valid vmalloc'ed area, and
1609 * that it is big enough to cover the vma. Will return failure if
1610 * that criteria isn't met.
1612 * Similar to remap_pfn_range() (see mm/memory.c)
1614 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1615 unsigned long pgoff)
1617 struct vm_struct *area;
1618 unsigned long uaddr = vma->vm_start;
1619 unsigned long usize = vma->vm_end - vma->vm_start;
1621 if ((PAGE_SIZE-1) & (unsigned long)addr)
1624 area = find_vm_area(addr);
1628 if (!(area->flags & VM_USERMAP))
1631 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1634 addr += pgoff << PAGE_SHIFT;
1636 struct page *page = vmalloc_to_page(addr);
1639 ret = vm_insert_page(vma, uaddr, page);
1646 } while (usize > 0);
1648 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1649 vma->vm_flags |= VM_RESERVED;
1653 EXPORT_SYMBOL(remap_vmalloc_range);
1656 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1659 void __attribute__((weak)) vmalloc_sync_all(void)
1664 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1666 /* apply_to_page_range() does all the hard work. */
1671 * alloc_vm_area - allocate a range of kernel address space
1672 * @size: size of the area
1674 * Returns: NULL on failure, vm_struct on success
1676 * This function reserves a range of kernel address space, and
1677 * allocates pagetables to map that range. No actual mappings
1678 * are created. If the kernel address space is not shared
1679 * between processes, it syncs the pagetable across all
1682 struct vm_struct *alloc_vm_area(size_t size)
1684 struct vm_struct *area;
1686 area = get_vm_area_caller(size, VM_IOREMAP,
1687 __builtin_return_address(0));
1692 * This ensures that page tables are constructed for this region
1693 * of kernel virtual address space and mapped into init_mm.
1695 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1696 area->size, f, NULL)) {
1701 /* Make sure the pagetables are constructed in process kernel
1707 EXPORT_SYMBOL_GPL(alloc_vm_area);
1709 void free_vm_area(struct vm_struct *area)
1711 struct vm_struct *ret;
1712 ret = remove_vm_area(area->addr);
1713 BUG_ON(ret != area);
1716 EXPORT_SYMBOL_GPL(free_vm_area);
1719 #ifdef CONFIG_PROC_FS
1720 static void *s_start(struct seq_file *m, loff_t *pos)
1723 struct vm_struct *v;
1725 read_lock(&vmlist_lock);
1727 while (n > 0 && v) {
1738 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1740 struct vm_struct *v = p;
1746 static void s_stop(struct seq_file *m, void *p)
1748 read_unlock(&vmlist_lock);
1751 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1754 unsigned int nr, *counters = m->private;
1759 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1761 for (nr = 0; nr < v->nr_pages; nr++)
1762 counters[page_to_nid(v->pages[nr])]++;
1764 for_each_node_state(nr, N_HIGH_MEMORY)
1766 seq_printf(m, " N%u=%u", nr, counters[nr]);
1770 static int s_show(struct seq_file *m, void *p)
1772 struct vm_struct *v = p;
1774 seq_printf(m, "0x%p-0x%p %7ld",
1775 v->addr, v->addr + v->size, v->size);
1778 char buff[KSYM_SYMBOL_LEN];
1781 sprint_symbol(buff, (unsigned long)v->caller);
1786 seq_printf(m, " pages=%d", v->nr_pages);
1789 seq_printf(m, " phys=%lx", v->phys_addr);
1791 if (v->flags & VM_IOREMAP)
1792 seq_printf(m, " ioremap");
1794 if (v->flags & VM_ALLOC)
1795 seq_printf(m, " vmalloc");
1797 if (v->flags & VM_MAP)
1798 seq_printf(m, " vmap");
1800 if (v->flags & VM_USERMAP)
1801 seq_printf(m, " user");
1803 if (v->flags & VM_VPAGES)
1804 seq_printf(m, " vpages");
1806 show_numa_info(m, v);
1811 static const struct seq_operations vmalloc_op = {
1818 static int vmalloc_open(struct inode *inode, struct file *file)
1820 unsigned int *ptr = NULL;
1824 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1825 ret = seq_open(file, &vmalloc_op);
1827 struct seq_file *m = file->private_data;
1834 static const struct file_operations proc_vmalloc_operations = {
1835 .open = vmalloc_open,
1837 .llseek = seq_lseek,
1838 .release = seq_release_private,
1841 static int __init proc_vmalloc_init(void)
1843 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1846 module_init(proc_vmalloc_init);