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>
27 #include <asm/atomic.h>
28 #include <asm/uaccess.h>
29 #include <asm/tlbflush.h>
32 /*** Page table manipulation functions ***/
34 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
38 pte = pte_offset_kernel(pmd, addr);
40 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
41 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
42 } while (pte++, addr += PAGE_SIZE, addr != end);
45 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
50 pmd = pmd_offset(pud, addr);
52 next = pmd_addr_end(addr, end);
53 if (pmd_none_or_clear_bad(pmd))
55 vunmap_pte_range(pmd, addr, next);
56 } while (pmd++, addr = next, addr != end);
59 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
64 pud = pud_offset(pgd, addr);
66 next = pud_addr_end(addr, end);
67 if (pud_none_or_clear_bad(pud))
69 vunmap_pmd_range(pud, addr, next);
70 } while (pud++, addr = next, addr != end);
73 static void vunmap_page_range(unsigned long addr, unsigned long end)
79 pgd = pgd_offset_k(addr);
80 flush_cache_vunmap(addr, end);
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 addr, unsigned long end,
156 pgprot_t prot, struct page **pages)
164 pgd = pgd_offset_k(addr);
166 next = pgd_addr_end(addr, end);
167 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
170 } while (pgd++, addr = next, addr != end);
171 flush_cache_vmap(addr, end);
178 static inline int is_vmalloc_or_module_addr(const void *x)
181 * ARM, x86-64 and sparc64 put modules in a special place,
182 * and fall back on vmalloc() if that fails. Others
183 * just put it in the vmalloc space.
185 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
186 unsigned long addr = (unsigned long)x;
187 if (addr >= MODULES_VADDR && addr < MODULES_END)
190 return is_vmalloc_addr(x);
194 * Walk a vmap address to the struct page it maps.
196 struct page *vmalloc_to_page(const void *vmalloc_addr)
198 unsigned long addr = (unsigned long) vmalloc_addr;
199 struct page *page = NULL;
200 pgd_t *pgd = pgd_offset_k(addr);
203 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
204 * architectures that do not vmalloc module space
206 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
208 if (!pgd_none(*pgd)) {
209 pud_t *pud = pud_offset(pgd, addr);
210 if (!pud_none(*pud)) {
211 pmd_t *pmd = pmd_offset(pud, addr);
212 if (!pmd_none(*pmd)) {
215 ptep = pte_offset_map(pmd, addr);
217 if (pte_present(pte))
218 page = pte_page(pte);
225 EXPORT_SYMBOL(vmalloc_to_page);
228 * Map a vmalloc()-space virtual address to the physical page frame number.
230 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
232 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
234 EXPORT_SYMBOL(vmalloc_to_pfn);
237 /*** Global kva allocator ***/
239 #define VM_LAZY_FREE 0x01
240 #define VM_LAZY_FREEING 0x02
241 #define VM_VM_AREA 0x04
244 unsigned long va_start;
245 unsigned long va_end;
247 struct rb_node rb_node; /* address sorted rbtree */
248 struct list_head list; /* address sorted list */
249 struct list_head purge_list; /* "lazy purge" list */
251 struct rcu_head rcu_head;
254 static DEFINE_SPINLOCK(vmap_area_lock);
255 static struct rb_root vmap_area_root = RB_ROOT;
256 static LIST_HEAD(vmap_area_list);
258 static struct vmap_area *__find_vmap_area(unsigned long addr)
260 struct rb_node *n = vmap_area_root.rb_node;
263 struct vmap_area *va;
265 va = rb_entry(n, struct vmap_area, rb_node);
266 if (addr < va->va_start)
268 else if (addr > va->va_start)
277 static void __insert_vmap_area(struct vmap_area *va)
279 struct rb_node **p = &vmap_area_root.rb_node;
280 struct rb_node *parent = NULL;
284 struct vmap_area *tmp;
287 tmp = rb_entry(parent, struct vmap_area, rb_node);
288 if (va->va_start < tmp->va_end)
290 else if (va->va_end > tmp->va_start)
296 rb_link_node(&va->rb_node, parent, p);
297 rb_insert_color(&va->rb_node, &vmap_area_root);
299 /* address-sort this list so it is usable like the vmlist */
300 tmp = rb_prev(&va->rb_node);
302 struct vmap_area *prev;
303 prev = rb_entry(tmp, struct vmap_area, rb_node);
304 list_add_rcu(&va->list, &prev->list);
306 list_add_rcu(&va->list, &vmap_area_list);
309 static void purge_vmap_area_lazy(void);
312 * Allocate a region of KVA of the specified size and alignment, within the
315 static struct vmap_area *alloc_vmap_area(unsigned long size,
317 unsigned long vstart, unsigned long vend,
318 int node, gfp_t gfp_mask)
320 struct vmap_area *va;
325 BUG_ON(size & ~PAGE_MASK);
327 addr = ALIGN(vstart, align);
329 va = kmalloc_node(sizeof(struct vmap_area),
330 gfp_mask & GFP_RECLAIM_MASK, node);
332 return ERR_PTR(-ENOMEM);
335 spin_lock(&vmap_area_lock);
336 /* XXX: could have a last_hole cache */
337 n = vmap_area_root.rb_node;
339 struct vmap_area *first = NULL;
342 struct vmap_area *tmp;
343 tmp = rb_entry(n, struct vmap_area, rb_node);
344 if (tmp->va_end >= addr) {
345 if (!first && tmp->va_start < addr + size)
357 if (first->va_end < addr) {
358 n = rb_next(&first->rb_node);
360 first = rb_entry(n, struct vmap_area, rb_node);
365 while (addr + size >= first->va_start && addr + size <= vend) {
366 addr = ALIGN(first->va_end + PAGE_SIZE, align);
368 n = rb_next(&first->rb_node);
370 first = rb_entry(n, struct vmap_area, rb_node);
376 if (addr + size > vend) {
377 spin_unlock(&vmap_area_lock);
379 purge_vmap_area_lazy();
383 if (printk_ratelimit())
384 printk(KERN_WARNING "vmap allocation failed: "
385 "use vmalloc=<size> to increase size.\n");
386 return ERR_PTR(-EBUSY);
389 BUG_ON(addr & (align-1));
392 va->va_end = addr + size;
394 __insert_vmap_area(va);
395 spin_unlock(&vmap_area_lock);
400 static void rcu_free_va(struct rcu_head *head)
402 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
407 static void __free_vmap_area(struct vmap_area *va)
409 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
410 rb_erase(&va->rb_node, &vmap_area_root);
411 RB_CLEAR_NODE(&va->rb_node);
412 list_del_rcu(&va->list);
414 call_rcu(&va->rcu_head, rcu_free_va);
418 * Free a region of KVA allocated by alloc_vmap_area
420 static void free_vmap_area(struct vmap_area *va)
422 spin_lock(&vmap_area_lock);
423 __free_vmap_area(va);
424 spin_unlock(&vmap_area_lock);
428 * Clear the pagetable entries of a given vmap_area
430 static void unmap_vmap_area(struct vmap_area *va)
432 vunmap_page_range(va->va_start, va->va_end);
436 * lazy_max_pages is the maximum amount of virtual address space we gather up
437 * before attempting to purge with a TLB flush.
439 * There is a tradeoff here: a larger number will cover more kernel page tables
440 * and take slightly longer to purge, but it will linearly reduce the number of
441 * global TLB flushes that must be performed. It would seem natural to scale
442 * this number up linearly with the number of CPUs (because vmapping activity
443 * could also scale linearly with the number of CPUs), however it is likely
444 * that in practice, workloads might be constrained in other ways that mean
445 * vmap activity will not scale linearly with CPUs. Also, I want to be
446 * conservative and not introduce a big latency on huge systems, so go with
447 * a less aggressive log scale. It will still be an improvement over the old
448 * code, and it will be simple to change the scale factor if we find that it
449 * becomes a problem on bigger systems.
451 static unsigned long lazy_max_pages(void)
455 log = fls(num_online_cpus());
457 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
460 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
463 * Purges all lazily-freed vmap areas.
465 * If sync is 0 then don't purge if there is already a purge in progress.
466 * If force_flush is 1, then flush kernel TLBs between *start and *end even
467 * if we found no lazy vmap areas to unmap (callers can use this to optimise
468 * their own TLB flushing).
469 * Returns with *start = min(*start, lowest purged address)
470 * *end = max(*end, highest purged address)
472 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
473 int sync, int force_flush)
475 static DEFINE_SPINLOCK(purge_lock);
477 struct vmap_area *va;
481 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
482 * should not expect such behaviour. This just simplifies locking for
483 * the case that isn't actually used at the moment anyway.
485 if (!sync && !force_flush) {
486 if (!spin_trylock(&purge_lock))
489 spin_lock(&purge_lock);
492 list_for_each_entry_rcu(va, &vmap_area_list, list) {
493 if (va->flags & VM_LAZY_FREE) {
494 if (va->va_start < *start)
495 *start = va->va_start;
496 if (va->va_end > *end)
498 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
500 list_add_tail(&va->purge_list, &valist);
501 va->flags |= VM_LAZY_FREEING;
502 va->flags &= ~VM_LAZY_FREE;
508 BUG_ON(nr > atomic_read(&vmap_lazy_nr));
509 atomic_sub(nr, &vmap_lazy_nr);
512 if (nr || force_flush)
513 flush_tlb_kernel_range(*start, *end);
516 spin_lock(&vmap_area_lock);
517 list_for_each_entry(va, &valist, purge_list)
518 __free_vmap_area(va);
519 spin_unlock(&vmap_area_lock);
521 spin_unlock(&purge_lock);
525 * Kick off a purge of the outstanding lazy areas.
527 static void purge_vmap_area_lazy(void)
529 unsigned long start = ULONG_MAX, end = 0;
531 __purge_vmap_area_lazy(&start, &end, 0, 0);
535 * Free and unmap a vmap area
537 static void free_unmap_vmap_area(struct vmap_area *va)
539 va->flags |= VM_LAZY_FREE;
540 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
541 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
542 purge_vmap_area_lazy();
545 static struct vmap_area *find_vmap_area(unsigned long addr)
547 struct vmap_area *va;
549 spin_lock(&vmap_area_lock);
550 va = __find_vmap_area(addr);
551 spin_unlock(&vmap_area_lock);
556 static void free_unmap_vmap_area_addr(unsigned long addr)
558 struct vmap_area *va;
560 va = find_vmap_area(addr);
562 free_unmap_vmap_area(va);
566 /*** Per cpu kva allocator ***/
569 * vmap space is limited especially on 32 bit architectures. Ensure there is
570 * room for at least 16 percpu vmap blocks per CPU.
573 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
574 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
575 * instead (we just need a rough idea)
577 #if BITS_PER_LONG == 32
578 #define VMALLOC_SPACE (128UL*1024*1024)
580 #define VMALLOC_SPACE (128UL*1024*1024*1024)
583 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
584 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
585 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
586 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
587 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
588 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
589 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
590 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
591 VMALLOC_PAGES / NR_CPUS / 16))
593 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
595 static bool vmap_initialized __read_mostly = false;
597 struct vmap_block_queue {
599 struct list_head free;
600 struct list_head dirty;
601 unsigned int nr_dirty;
606 struct vmap_area *va;
607 struct vmap_block_queue *vbq;
608 unsigned long free, dirty;
609 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
610 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
613 struct list_head free_list;
614 struct list_head dirty_list;
616 struct rcu_head rcu_head;
620 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
621 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
624 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
625 * in the free path. Could get rid of this if we change the API to return a
626 * "cookie" from alloc, to be passed to free. But no big deal yet.
628 static DEFINE_SPINLOCK(vmap_block_tree_lock);
629 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
632 * We should probably have a fallback mechanism to allocate virtual memory
633 * out of partially filled vmap blocks. However vmap block sizing should be
634 * fairly reasonable according to the vmalloc size, so it shouldn't be a
638 static unsigned long addr_to_vb_idx(unsigned long addr)
640 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
641 addr /= VMAP_BLOCK_SIZE;
645 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
647 struct vmap_block_queue *vbq;
648 struct vmap_block *vb;
649 struct vmap_area *va;
650 unsigned long vb_idx;
653 node = numa_node_id();
655 vb = kmalloc_node(sizeof(struct vmap_block),
656 gfp_mask & GFP_RECLAIM_MASK, node);
658 return ERR_PTR(-ENOMEM);
660 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
661 VMALLOC_START, VMALLOC_END,
663 if (unlikely(IS_ERR(va))) {
665 return ERR_PTR(PTR_ERR(va));
668 err = radix_tree_preload(gfp_mask);
675 spin_lock_init(&vb->lock);
677 vb->free = VMAP_BBMAP_BITS;
679 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
680 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
681 INIT_LIST_HEAD(&vb->free_list);
682 INIT_LIST_HEAD(&vb->dirty_list);
684 vb_idx = addr_to_vb_idx(va->va_start);
685 spin_lock(&vmap_block_tree_lock);
686 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
687 spin_unlock(&vmap_block_tree_lock);
689 radix_tree_preload_end();
691 vbq = &get_cpu_var(vmap_block_queue);
693 spin_lock(&vbq->lock);
694 list_add(&vb->free_list, &vbq->free);
695 spin_unlock(&vbq->lock);
696 put_cpu_var(vmap_cpu_blocks);
701 static void rcu_free_vb(struct rcu_head *head)
703 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
708 static void free_vmap_block(struct vmap_block *vb)
710 struct vmap_block *tmp;
711 unsigned long vb_idx;
713 spin_lock(&vb->vbq->lock);
714 if (!list_empty(&vb->free_list))
715 list_del(&vb->free_list);
716 if (!list_empty(&vb->dirty_list))
717 list_del(&vb->dirty_list);
718 spin_unlock(&vb->vbq->lock);
720 vb_idx = addr_to_vb_idx(vb->va->va_start);
721 spin_lock(&vmap_block_tree_lock);
722 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
723 spin_unlock(&vmap_block_tree_lock);
726 free_unmap_vmap_area(vb->va);
727 call_rcu(&vb->rcu_head, rcu_free_vb);
730 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
732 struct vmap_block_queue *vbq;
733 struct vmap_block *vb;
734 unsigned long addr = 0;
737 BUG_ON(size & ~PAGE_MASK);
738 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
739 order = get_order(size);
743 vbq = &get_cpu_var(vmap_block_queue);
744 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
747 spin_lock(&vb->lock);
748 i = bitmap_find_free_region(vb->alloc_map,
749 VMAP_BBMAP_BITS, order);
752 addr = vb->va->va_start + (i << PAGE_SHIFT);
753 BUG_ON(addr_to_vb_idx(addr) !=
754 addr_to_vb_idx(vb->va->va_start));
755 vb->free -= 1UL << order;
757 spin_lock(&vbq->lock);
758 list_del_init(&vb->free_list);
759 spin_unlock(&vbq->lock);
761 spin_unlock(&vb->lock);
764 spin_unlock(&vb->lock);
766 put_cpu_var(vmap_cpu_blocks);
770 vb = new_vmap_block(gfp_mask);
779 static void vb_free(const void *addr, unsigned long size)
781 unsigned long offset;
782 unsigned long vb_idx;
784 struct vmap_block *vb;
786 BUG_ON(size & ~PAGE_MASK);
787 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
788 order = get_order(size);
790 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
792 vb_idx = addr_to_vb_idx((unsigned long)addr);
794 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
798 spin_lock(&vb->lock);
799 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
801 spin_lock(&vb->vbq->lock);
802 list_add(&vb->dirty_list, &vb->vbq->dirty);
803 spin_unlock(&vb->vbq->lock);
805 vb->dirty += 1UL << order;
806 if (vb->dirty == VMAP_BBMAP_BITS) {
807 BUG_ON(vb->free || !list_empty(&vb->free_list));
808 spin_unlock(&vb->lock);
811 spin_unlock(&vb->lock);
815 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
817 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
818 * to amortize TLB flushing overheads. What this means is that any page you
819 * have now, may, in a former life, have been mapped into kernel virtual
820 * address by the vmap layer and so there might be some CPUs with TLB entries
821 * still referencing that page (additional to the regular 1:1 kernel mapping).
823 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
824 * be sure that none of the pages we have control over will have any aliases
825 * from the vmap layer.
827 void vm_unmap_aliases(void)
829 unsigned long start = ULONG_MAX, end = 0;
833 if (unlikely(!vmap_initialized))
836 for_each_possible_cpu(cpu) {
837 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
838 struct vmap_block *vb;
841 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
844 spin_lock(&vb->lock);
845 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
846 while (i < VMAP_BBMAP_BITS) {
849 j = find_next_zero_bit(vb->dirty_map,
852 s = vb->va->va_start + (i << PAGE_SHIFT);
853 e = vb->va->va_start + (j << PAGE_SHIFT);
854 vunmap_page_range(s, e);
863 i = find_next_bit(vb->dirty_map,
866 spin_unlock(&vb->lock);
871 __purge_vmap_area_lazy(&start, &end, 1, flush);
873 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
876 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
877 * @mem: the pointer returned by vm_map_ram
878 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
880 void vm_unmap_ram(const void *mem, unsigned int count)
882 unsigned long size = count << PAGE_SHIFT;
883 unsigned long addr = (unsigned long)mem;
886 BUG_ON(addr < VMALLOC_START);
887 BUG_ON(addr > VMALLOC_END);
888 BUG_ON(addr & (PAGE_SIZE-1));
890 debug_check_no_locks_freed(mem, size);
892 if (likely(count <= VMAP_MAX_ALLOC))
895 free_unmap_vmap_area_addr(addr);
897 EXPORT_SYMBOL(vm_unmap_ram);
900 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
901 * @pages: an array of pointers to the pages to be mapped
902 * @count: number of pages
903 * @node: prefer to allocate data structures on this node
904 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
906 * Returns: a pointer to the address that has been mapped, or %NULL on failure
908 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
910 unsigned long size = count << PAGE_SHIFT;
914 if (likely(count <= VMAP_MAX_ALLOC)) {
915 mem = vb_alloc(size, GFP_KERNEL);
918 addr = (unsigned long)mem;
920 struct vmap_area *va;
921 va = alloc_vmap_area(size, PAGE_SIZE,
922 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
929 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
930 vm_unmap_ram(mem, count);
935 EXPORT_SYMBOL(vm_map_ram);
937 void __init vmalloc_init(void)
941 for_each_possible_cpu(i) {
942 struct vmap_block_queue *vbq;
944 vbq = &per_cpu(vmap_block_queue, i);
945 spin_lock_init(&vbq->lock);
946 INIT_LIST_HEAD(&vbq->free);
947 INIT_LIST_HEAD(&vbq->dirty);
951 vmap_initialized = true;
954 void unmap_kernel_range(unsigned long addr, unsigned long size)
956 unsigned long end = addr + size;
957 vunmap_page_range(addr, end);
958 flush_tlb_kernel_range(addr, end);
961 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
963 unsigned long addr = (unsigned long)area->addr;
964 unsigned long end = addr + area->size - PAGE_SIZE;
967 err = vmap_page_range(addr, end, prot, *pages);
975 EXPORT_SYMBOL_GPL(map_vm_area);
977 /*** Old vmalloc interfaces ***/
978 DEFINE_RWLOCK(vmlist_lock);
979 struct vm_struct *vmlist;
981 static struct vm_struct *__get_vm_area_node(unsigned long size,
982 unsigned long flags, unsigned long start, unsigned long end,
983 int node, gfp_t gfp_mask, void *caller)
985 static struct vmap_area *va;
986 struct vm_struct *area;
987 struct vm_struct *tmp, **p;
988 unsigned long align = 1;
990 BUG_ON(in_interrupt());
991 if (flags & VM_IOREMAP) {
994 if (bit > IOREMAP_MAX_ORDER)
995 bit = IOREMAP_MAX_ORDER;
996 else if (bit < PAGE_SHIFT)
1002 size = PAGE_ALIGN(size);
1003 if (unlikely(!size))
1006 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1007 if (unlikely(!area))
1011 * We always allocate a guard page.
1015 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1021 area->flags = flags;
1022 area->addr = (void *)va->va_start;
1026 area->phys_addr = 0;
1027 area->caller = caller;
1029 va->flags |= VM_VM_AREA;
1031 write_lock(&vmlist_lock);
1032 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1033 if (tmp->addr >= area->addr)
1038 write_unlock(&vmlist_lock);
1043 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1044 unsigned long start, unsigned long end)
1046 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1047 __builtin_return_address(0));
1049 EXPORT_SYMBOL_GPL(__get_vm_area);
1052 * get_vm_area - reserve a contiguous kernel virtual area
1053 * @size: size of the area
1054 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1056 * Search an area of @size in the kernel virtual mapping area,
1057 * and reserved it for out purposes. Returns the area descriptor
1058 * on success or %NULL on failure.
1060 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1062 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1063 -1, GFP_KERNEL, __builtin_return_address(0));
1066 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1069 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1070 -1, GFP_KERNEL, caller);
1073 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1074 int node, gfp_t gfp_mask)
1076 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1077 gfp_mask, __builtin_return_address(0));
1080 static struct vm_struct *find_vm_area(const void *addr)
1082 struct vmap_area *va;
1084 va = find_vmap_area((unsigned long)addr);
1085 if (va && va->flags & VM_VM_AREA)
1092 * remove_vm_area - find and remove a continuous kernel virtual area
1093 * @addr: base address
1095 * Search for the kernel VM area starting at @addr, and remove it.
1096 * This function returns the found VM area, but using it is NOT safe
1097 * on SMP machines, except for its size or flags.
1099 struct vm_struct *remove_vm_area(const void *addr)
1101 struct vmap_area *va;
1103 va = find_vmap_area((unsigned long)addr);
1104 if (va && va->flags & VM_VM_AREA) {
1105 struct vm_struct *vm = va->private;
1106 struct vm_struct *tmp, **p;
1107 free_unmap_vmap_area(va);
1108 vm->size -= PAGE_SIZE;
1110 write_lock(&vmlist_lock);
1111 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1114 write_unlock(&vmlist_lock);
1121 static void __vunmap(const void *addr, int deallocate_pages)
1123 struct vm_struct *area;
1128 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1129 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1133 area = remove_vm_area(addr);
1134 if (unlikely(!area)) {
1135 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1140 debug_check_no_locks_freed(addr, area->size);
1141 debug_check_no_obj_freed(addr, area->size);
1143 if (deallocate_pages) {
1146 for (i = 0; i < area->nr_pages; i++) {
1147 struct page *page = area->pages[i];
1153 if (area->flags & VM_VPAGES)
1164 * vfree - release memory allocated by vmalloc()
1165 * @addr: memory base address
1167 * Free the virtually continuous memory area starting at @addr, as
1168 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1169 * NULL, no operation is performed.
1171 * Must not be called in interrupt context.
1173 void vfree(const void *addr)
1175 BUG_ON(in_interrupt());
1178 EXPORT_SYMBOL(vfree);
1181 * vunmap - release virtual mapping obtained by vmap()
1182 * @addr: memory base address
1184 * Free the virtually contiguous memory area starting at @addr,
1185 * which was created from the page array passed to vmap().
1187 * Must not be called in interrupt context.
1189 void vunmap(const void *addr)
1191 BUG_ON(in_interrupt());
1194 EXPORT_SYMBOL(vunmap);
1197 * vmap - map an array of pages into virtually contiguous space
1198 * @pages: array of page pointers
1199 * @count: number of pages to map
1200 * @flags: vm_area->flags
1201 * @prot: page protection for the mapping
1203 * Maps @count pages from @pages into contiguous kernel virtual
1206 void *vmap(struct page **pages, unsigned int count,
1207 unsigned long flags, pgprot_t prot)
1209 struct vm_struct *area;
1211 if (count > num_physpages)
1214 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1215 __builtin_return_address(0));
1219 if (map_vm_area(area, prot, &pages)) {
1226 EXPORT_SYMBOL(vmap);
1228 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1229 int node, void *caller);
1230 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1231 pgprot_t prot, int node, void *caller)
1233 struct page **pages;
1234 unsigned int nr_pages, array_size, i;
1236 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1237 array_size = (nr_pages * sizeof(struct page *));
1239 area->nr_pages = nr_pages;
1240 /* Please note that the recursion is strictly bounded. */
1241 if (array_size > PAGE_SIZE) {
1242 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1243 PAGE_KERNEL, node, caller);
1244 area->flags |= VM_VPAGES;
1246 pages = kmalloc_node(array_size,
1247 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1250 area->pages = pages;
1251 area->caller = caller;
1253 remove_vm_area(area->addr);
1258 for (i = 0; i < area->nr_pages; i++) {
1262 page = alloc_page(gfp_mask);
1264 page = alloc_pages_node(node, gfp_mask, 0);
1266 if (unlikely(!page)) {
1267 /* Successfully allocated i pages, free them in __vunmap() */
1271 area->pages[i] = page;
1274 if (map_vm_area(area, prot, &pages))
1283 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1285 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1286 __builtin_return_address(0));
1290 * __vmalloc_node - allocate virtually contiguous memory
1291 * @size: allocation size
1292 * @gfp_mask: flags for the page level allocator
1293 * @prot: protection mask for the allocated pages
1294 * @node: node to use for allocation or -1
1295 * @caller: caller's return address
1297 * Allocate enough pages to cover @size from the page level
1298 * allocator with @gfp_mask flags. Map them into contiguous
1299 * kernel virtual space, using a pagetable protection of @prot.
1301 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1302 int node, void *caller)
1304 struct vm_struct *area;
1306 size = PAGE_ALIGN(size);
1307 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1310 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1311 node, gfp_mask, caller);
1316 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1319 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1321 return __vmalloc_node(size, gfp_mask, prot, -1,
1322 __builtin_return_address(0));
1324 EXPORT_SYMBOL(__vmalloc);
1327 * vmalloc - allocate virtually contiguous memory
1328 * @size: allocation size
1329 * Allocate enough pages to cover @size from the page level
1330 * allocator and map them into contiguous kernel virtual space.
1332 * For tight control over page level allocator and protection flags
1333 * use __vmalloc() instead.
1335 void *vmalloc(unsigned long size)
1337 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1338 -1, __builtin_return_address(0));
1340 EXPORT_SYMBOL(vmalloc);
1343 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1344 * @size: allocation size
1346 * The resulting memory area is zeroed so it can be mapped to userspace
1347 * without leaking data.
1349 void *vmalloc_user(unsigned long size)
1351 struct vm_struct *area;
1354 ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
1356 area = find_vm_area(ret);
1357 area->flags |= VM_USERMAP;
1361 EXPORT_SYMBOL(vmalloc_user);
1364 * vmalloc_node - allocate memory on a specific node
1365 * @size: allocation size
1368 * Allocate enough pages to cover @size from the page level
1369 * allocator and map them into contiguous kernel virtual space.
1371 * For tight control over page level allocator and protection flags
1372 * use __vmalloc() instead.
1374 void *vmalloc_node(unsigned long size, int node)
1376 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1377 node, __builtin_return_address(0));
1379 EXPORT_SYMBOL(vmalloc_node);
1381 #ifndef PAGE_KERNEL_EXEC
1382 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1386 * vmalloc_exec - allocate virtually contiguous, executable memory
1387 * @size: allocation size
1389 * Kernel-internal function to allocate enough pages to cover @size
1390 * the page level allocator and map them into contiguous and
1391 * executable kernel virtual space.
1393 * For tight control over page level allocator and protection flags
1394 * use __vmalloc() instead.
1397 void *vmalloc_exec(unsigned long size)
1399 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
1402 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1403 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1404 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1405 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1407 #define GFP_VMALLOC32 GFP_KERNEL
1411 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1412 * @size: allocation size
1414 * Allocate enough 32bit PA addressable pages to cover @size from the
1415 * page level allocator and map them into contiguous kernel virtual space.
1417 void *vmalloc_32(unsigned long size)
1419 return __vmalloc(size, GFP_VMALLOC32, PAGE_KERNEL);
1421 EXPORT_SYMBOL(vmalloc_32);
1424 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1425 * @size: allocation size
1427 * The resulting memory area is 32bit addressable and zeroed so it can be
1428 * mapped to userspace without leaking data.
1430 void *vmalloc_32_user(unsigned long size)
1432 struct vm_struct *area;
1435 ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
1437 area = find_vm_area(ret);
1438 area->flags |= VM_USERMAP;
1442 EXPORT_SYMBOL(vmalloc_32_user);
1444 long vread(char *buf, char *addr, unsigned long count)
1446 struct vm_struct *tmp;
1447 char *vaddr, *buf_start = buf;
1450 /* Don't allow overflow */
1451 if ((unsigned long) addr + count < count)
1452 count = -(unsigned long) addr;
1454 read_lock(&vmlist_lock);
1455 for (tmp = vmlist; tmp; tmp = tmp->next) {
1456 vaddr = (char *) tmp->addr;
1457 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1459 while (addr < vaddr) {
1467 n = vaddr + tmp->size - PAGE_SIZE - addr;
1478 read_unlock(&vmlist_lock);
1479 return buf - buf_start;
1482 long vwrite(char *buf, char *addr, unsigned long count)
1484 struct vm_struct *tmp;
1485 char *vaddr, *buf_start = buf;
1488 /* Don't allow overflow */
1489 if ((unsigned long) addr + count < count)
1490 count = -(unsigned long) addr;
1492 read_lock(&vmlist_lock);
1493 for (tmp = vmlist; tmp; tmp = tmp->next) {
1494 vaddr = (char *) tmp->addr;
1495 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1497 while (addr < vaddr) {
1504 n = vaddr + tmp->size - PAGE_SIZE - addr;
1515 read_unlock(&vmlist_lock);
1516 return buf - buf_start;
1520 * remap_vmalloc_range - map vmalloc pages to userspace
1521 * @vma: vma to cover (map full range of vma)
1522 * @addr: vmalloc memory
1523 * @pgoff: number of pages into addr before first page to map
1525 * Returns: 0 for success, -Exxx on failure
1527 * This function checks that addr is a valid vmalloc'ed area, and
1528 * that it is big enough to cover the vma. Will return failure if
1529 * that criteria isn't met.
1531 * Similar to remap_pfn_range() (see mm/memory.c)
1533 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1534 unsigned long pgoff)
1536 struct vm_struct *area;
1537 unsigned long uaddr = vma->vm_start;
1538 unsigned long usize = vma->vm_end - vma->vm_start;
1540 if ((PAGE_SIZE-1) & (unsigned long)addr)
1543 area = find_vm_area(addr);
1547 if (!(area->flags & VM_USERMAP))
1550 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1553 addr += pgoff << PAGE_SHIFT;
1555 struct page *page = vmalloc_to_page(addr);
1558 ret = vm_insert_page(vma, uaddr, page);
1565 } while (usize > 0);
1567 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1568 vma->vm_flags |= VM_RESERVED;
1572 EXPORT_SYMBOL(remap_vmalloc_range);
1575 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1578 void __attribute__((weak)) vmalloc_sync_all(void)
1583 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1585 /* apply_to_page_range() does all the hard work. */
1590 * alloc_vm_area - allocate a range of kernel address space
1591 * @size: size of the area
1593 * Returns: NULL on failure, vm_struct on success
1595 * This function reserves a range of kernel address space, and
1596 * allocates pagetables to map that range. No actual mappings
1597 * are created. If the kernel address space is not shared
1598 * between processes, it syncs the pagetable across all
1601 struct vm_struct *alloc_vm_area(size_t size)
1603 struct vm_struct *area;
1605 area = get_vm_area_caller(size, VM_IOREMAP,
1606 __builtin_return_address(0));
1611 * This ensures that page tables are constructed for this region
1612 * of kernel virtual address space and mapped into init_mm.
1614 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1615 area->size, f, NULL)) {
1620 /* Make sure the pagetables are constructed in process kernel
1626 EXPORT_SYMBOL_GPL(alloc_vm_area);
1628 void free_vm_area(struct vm_struct *area)
1630 struct vm_struct *ret;
1631 ret = remove_vm_area(area->addr);
1632 BUG_ON(ret != area);
1635 EXPORT_SYMBOL_GPL(free_vm_area);
1638 #ifdef CONFIG_PROC_FS
1639 static void *s_start(struct seq_file *m, loff_t *pos)
1642 struct vm_struct *v;
1644 read_lock(&vmlist_lock);
1646 while (n > 0 && v) {
1657 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1659 struct vm_struct *v = p;
1665 static void s_stop(struct seq_file *m, void *p)
1667 read_unlock(&vmlist_lock);
1670 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1673 unsigned int nr, *counters = m->private;
1678 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1680 for (nr = 0; nr < v->nr_pages; nr++)
1681 counters[page_to_nid(v->pages[nr])]++;
1683 for_each_node_state(nr, N_HIGH_MEMORY)
1685 seq_printf(m, " N%u=%u", nr, counters[nr]);
1689 static int s_show(struct seq_file *m, void *p)
1691 struct vm_struct *v = p;
1693 seq_printf(m, "0x%p-0x%p %7ld",
1694 v->addr, v->addr + v->size, v->size);
1697 char buff[2 * KSYM_NAME_LEN];
1700 sprint_symbol(buff, (unsigned long)v->caller);
1705 seq_printf(m, " pages=%d", v->nr_pages);
1708 seq_printf(m, " phys=%lx", v->phys_addr);
1710 if (v->flags & VM_IOREMAP)
1711 seq_printf(m, " ioremap");
1713 if (v->flags & VM_ALLOC)
1714 seq_printf(m, " vmalloc");
1716 if (v->flags & VM_MAP)
1717 seq_printf(m, " vmap");
1719 if (v->flags & VM_USERMAP)
1720 seq_printf(m, " user");
1722 if (v->flags & VM_VPAGES)
1723 seq_printf(m, " vpages");
1725 show_numa_info(m, v);
1730 static const struct seq_operations vmalloc_op = {
1737 static int vmalloc_open(struct inode *inode, struct file *file)
1739 unsigned int *ptr = NULL;
1743 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1744 ret = seq_open(file, &vmalloc_op);
1746 struct seq_file *m = file->private_data;
1753 static const struct file_operations proc_vmalloc_operations = {
1754 .open = vmalloc_open,
1756 .llseek = seq_lseek,
1757 .release = seq_release_private,
1760 static int __init proc_vmalloc_init(void)
1762 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1765 module_init(proc_vmalloc_init);