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. Don't bother if somebody
526 * is already purging.
528 static void try_purge_vmap_area_lazy(void)
530 unsigned long start = ULONG_MAX, end = 0;
532 __purge_vmap_area_lazy(&start, &end, 0, 0);
536 * Kick off a purge of the outstanding lazy areas.
538 static void purge_vmap_area_lazy(void)
540 unsigned long start = ULONG_MAX, end = 0;
542 __purge_vmap_area_lazy(&start, &end, 1, 0);
546 * Free and unmap a vmap area
548 static void free_unmap_vmap_area(struct vmap_area *va)
550 va->flags |= VM_LAZY_FREE;
551 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
552 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
553 try_purge_vmap_area_lazy();
556 static struct vmap_area *find_vmap_area(unsigned long addr)
558 struct vmap_area *va;
560 spin_lock(&vmap_area_lock);
561 va = __find_vmap_area(addr);
562 spin_unlock(&vmap_area_lock);
567 static void free_unmap_vmap_area_addr(unsigned long addr)
569 struct vmap_area *va;
571 va = find_vmap_area(addr);
573 free_unmap_vmap_area(va);
577 /*** Per cpu kva allocator ***/
580 * vmap space is limited especially on 32 bit architectures. Ensure there is
581 * room for at least 16 percpu vmap blocks per CPU.
584 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
585 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
586 * instead (we just need a rough idea)
588 #if BITS_PER_LONG == 32
589 #define VMALLOC_SPACE (128UL*1024*1024)
591 #define VMALLOC_SPACE (128UL*1024*1024*1024)
594 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
595 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
596 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
597 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
598 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
599 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
600 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
601 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
602 VMALLOC_PAGES / NR_CPUS / 16))
604 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
606 static bool vmap_initialized __read_mostly = false;
608 struct vmap_block_queue {
610 struct list_head free;
611 struct list_head dirty;
612 unsigned int nr_dirty;
617 struct vmap_area *va;
618 struct vmap_block_queue *vbq;
619 unsigned long free, dirty;
620 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
621 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
624 struct list_head free_list;
625 struct list_head dirty_list;
627 struct rcu_head rcu_head;
631 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
632 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
635 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
636 * in the free path. Could get rid of this if we change the API to return a
637 * "cookie" from alloc, to be passed to free. But no big deal yet.
639 static DEFINE_SPINLOCK(vmap_block_tree_lock);
640 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
643 * We should probably have a fallback mechanism to allocate virtual memory
644 * out of partially filled vmap blocks. However vmap block sizing should be
645 * fairly reasonable according to the vmalloc size, so it shouldn't be a
649 static unsigned long addr_to_vb_idx(unsigned long addr)
651 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
652 addr /= VMAP_BLOCK_SIZE;
656 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
658 struct vmap_block_queue *vbq;
659 struct vmap_block *vb;
660 struct vmap_area *va;
661 unsigned long vb_idx;
664 node = numa_node_id();
666 vb = kmalloc_node(sizeof(struct vmap_block),
667 gfp_mask & GFP_RECLAIM_MASK, node);
669 return ERR_PTR(-ENOMEM);
671 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
672 VMALLOC_START, VMALLOC_END,
674 if (unlikely(IS_ERR(va))) {
676 return ERR_PTR(PTR_ERR(va));
679 err = radix_tree_preload(gfp_mask);
686 spin_lock_init(&vb->lock);
688 vb->free = VMAP_BBMAP_BITS;
690 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
691 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
692 INIT_LIST_HEAD(&vb->free_list);
693 INIT_LIST_HEAD(&vb->dirty_list);
695 vb_idx = addr_to_vb_idx(va->va_start);
696 spin_lock(&vmap_block_tree_lock);
697 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
698 spin_unlock(&vmap_block_tree_lock);
700 radix_tree_preload_end();
702 vbq = &get_cpu_var(vmap_block_queue);
704 spin_lock(&vbq->lock);
705 list_add(&vb->free_list, &vbq->free);
706 spin_unlock(&vbq->lock);
707 put_cpu_var(vmap_cpu_blocks);
712 static void rcu_free_vb(struct rcu_head *head)
714 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
719 static void free_vmap_block(struct vmap_block *vb)
721 struct vmap_block *tmp;
722 unsigned long vb_idx;
724 spin_lock(&vb->vbq->lock);
725 if (!list_empty(&vb->free_list))
726 list_del(&vb->free_list);
727 if (!list_empty(&vb->dirty_list))
728 list_del(&vb->dirty_list);
729 spin_unlock(&vb->vbq->lock);
731 vb_idx = addr_to_vb_idx(vb->va->va_start);
732 spin_lock(&vmap_block_tree_lock);
733 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
734 spin_unlock(&vmap_block_tree_lock);
737 free_unmap_vmap_area(vb->va);
738 call_rcu(&vb->rcu_head, rcu_free_vb);
741 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
743 struct vmap_block_queue *vbq;
744 struct vmap_block *vb;
745 unsigned long addr = 0;
748 BUG_ON(size & ~PAGE_MASK);
749 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
750 order = get_order(size);
754 vbq = &get_cpu_var(vmap_block_queue);
755 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
758 spin_lock(&vb->lock);
759 i = bitmap_find_free_region(vb->alloc_map,
760 VMAP_BBMAP_BITS, order);
763 addr = vb->va->va_start + (i << PAGE_SHIFT);
764 BUG_ON(addr_to_vb_idx(addr) !=
765 addr_to_vb_idx(vb->va->va_start));
766 vb->free -= 1UL << order;
768 spin_lock(&vbq->lock);
769 list_del_init(&vb->free_list);
770 spin_unlock(&vbq->lock);
772 spin_unlock(&vb->lock);
775 spin_unlock(&vb->lock);
777 put_cpu_var(vmap_cpu_blocks);
781 vb = new_vmap_block(gfp_mask);
790 static void vb_free(const void *addr, unsigned long size)
792 unsigned long offset;
793 unsigned long vb_idx;
795 struct vmap_block *vb;
797 BUG_ON(size & ~PAGE_MASK);
798 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
799 order = get_order(size);
801 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
803 vb_idx = addr_to_vb_idx((unsigned long)addr);
805 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
809 spin_lock(&vb->lock);
810 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
812 spin_lock(&vb->vbq->lock);
813 list_add(&vb->dirty_list, &vb->vbq->dirty);
814 spin_unlock(&vb->vbq->lock);
816 vb->dirty += 1UL << order;
817 if (vb->dirty == VMAP_BBMAP_BITS) {
818 BUG_ON(vb->free || !list_empty(&vb->free_list));
819 spin_unlock(&vb->lock);
822 spin_unlock(&vb->lock);
826 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
828 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
829 * to amortize TLB flushing overheads. What this means is that any page you
830 * have now, may, in a former life, have been mapped into kernel virtual
831 * address by the vmap layer and so there might be some CPUs with TLB entries
832 * still referencing that page (additional to the regular 1:1 kernel mapping).
834 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
835 * be sure that none of the pages we have control over will have any aliases
836 * from the vmap layer.
838 void vm_unmap_aliases(void)
840 unsigned long start = ULONG_MAX, end = 0;
844 if (unlikely(!vmap_initialized))
847 for_each_possible_cpu(cpu) {
848 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
849 struct vmap_block *vb;
852 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
855 spin_lock(&vb->lock);
856 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
857 while (i < VMAP_BBMAP_BITS) {
860 j = find_next_zero_bit(vb->dirty_map,
863 s = vb->va->va_start + (i << PAGE_SHIFT);
864 e = vb->va->va_start + (j << PAGE_SHIFT);
865 vunmap_page_range(s, e);
874 i = find_next_bit(vb->dirty_map,
877 spin_unlock(&vb->lock);
882 __purge_vmap_area_lazy(&start, &end, 1, flush);
884 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
887 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
888 * @mem: the pointer returned by vm_map_ram
889 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
891 void vm_unmap_ram(const void *mem, unsigned int count)
893 unsigned long size = count << PAGE_SHIFT;
894 unsigned long addr = (unsigned long)mem;
897 BUG_ON(addr < VMALLOC_START);
898 BUG_ON(addr > VMALLOC_END);
899 BUG_ON(addr & (PAGE_SIZE-1));
901 debug_check_no_locks_freed(mem, size);
903 if (likely(count <= VMAP_MAX_ALLOC))
906 free_unmap_vmap_area_addr(addr);
908 EXPORT_SYMBOL(vm_unmap_ram);
911 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
912 * @pages: an array of pointers to the pages to be mapped
913 * @count: number of pages
914 * @node: prefer to allocate data structures on this node
915 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
917 * Returns: a pointer to the address that has been mapped, or %NULL on failure
919 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
921 unsigned long size = count << PAGE_SHIFT;
925 if (likely(count <= VMAP_MAX_ALLOC)) {
926 mem = vb_alloc(size, GFP_KERNEL);
929 addr = (unsigned long)mem;
931 struct vmap_area *va;
932 va = alloc_vmap_area(size, PAGE_SIZE,
933 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
940 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
941 vm_unmap_ram(mem, count);
946 EXPORT_SYMBOL(vm_map_ram);
948 void __init vmalloc_init(void)
952 for_each_possible_cpu(i) {
953 struct vmap_block_queue *vbq;
955 vbq = &per_cpu(vmap_block_queue, i);
956 spin_lock_init(&vbq->lock);
957 INIT_LIST_HEAD(&vbq->free);
958 INIT_LIST_HEAD(&vbq->dirty);
962 vmap_initialized = true;
965 void unmap_kernel_range(unsigned long addr, unsigned long size)
967 unsigned long end = addr + size;
968 vunmap_page_range(addr, end);
969 flush_tlb_kernel_range(addr, end);
972 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
974 unsigned long addr = (unsigned long)area->addr;
975 unsigned long end = addr + area->size - PAGE_SIZE;
978 err = vmap_page_range(addr, end, prot, *pages);
986 EXPORT_SYMBOL_GPL(map_vm_area);
988 /*** Old vmalloc interfaces ***/
989 DEFINE_RWLOCK(vmlist_lock);
990 struct vm_struct *vmlist;
992 static struct vm_struct *__get_vm_area_node(unsigned long size,
993 unsigned long flags, unsigned long start, unsigned long end,
994 int node, gfp_t gfp_mask, void *caller)
996 static struct vmap_area *va;
997 struct vm_struct *area;
998 struct vm_struct *tmp, **p;
999 unsigned long align = 1;
1001 BUG_ON(in_interrupt());
1002 if (flags & VM_IOREMAP) {
1003 int bit = fls(size);
1005 if (bit > IOREMAP_MAX_ORDER)
1006 bit = IOREMAP_MAX_ORDER;
1007 else if (bit < PAGE_SHIFT)
1013 size = PAGE_ALIGN(size);
1014 if (unlikely(!size))
1017 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1018 if (unlikely(!area))
1022 * We always allocate a guard page.
1026 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1032 area->flags = flags;
1033 area->addr = (void *)va->va_start;
1037 area->phys_addr = 0;
1038 area->caller = caller;
1040 va->flags |= VM_VM_AREA;
1042 write_lock(&vmlist_lock);
1043 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1044 if (tmp->addr >= area->addr)
1049 write_unlock(&vmlist_lock);
1054 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1055 unsigned long start, unsigned long end)
1057 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1058 __builtin_return_address(0));
1060 EXPORT_SYMBOL_GPL(__get_vm_area);
1063 * get_vm_area - reserve a contiguous kernel virtual area
1064 * @size: size of the area
1065 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1067 * Search an area of @size in the kernel virtual mapping area,
1068 * and reserved it for out purposes. Returns the area descriptor
1069 * on success or %NULL on failure.
1071 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1073 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1074 -1, GFP_KERNEL, __builtin_return_address(0));
1077 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1080 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1081 -1, GFP_KERNEL, caller);
1084 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1085 int node, gfp_t gfp_mask)
1087 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1088 gfp_mask, __builtin_return_address(0));
1091 static struct vm_struct *find_vm_area(const void *addr)
1093 struct vmap_area *va;
1095 va = find_vmap_area((unsigned long)addr);
1096 if (va && va->flags & VM_VM_AREA)
1103 * remove_vm_area - find and remove a continuous kernel virtual area
1104 * @addr: base address
1106 * Search for the kernel VM area starting at @addr, and remove it.
1107 * This function returns the found VM area, but using it is NOT safe
1108 * on SMP machines, except for its size or flags.
1110 struct vm_struct *remove_vm_area(const void *addr)
1112 struct vmap_area *va;
1114 va = find_vmap_area((unsigned long)addr);
1115 if (va && va->flags & VM_VM_AREA) {
1116 struct vm_struct *vm = va->private;
1117 struct vm_struct *tmp, **p;
1118 free_unmap_vmap_area(va);
1119 vm->size -= PAGE_SIZE;
1121 write_lock(&vmlist_lock);
1122 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1125 write_unlock(&vmlist_lock);
1132 static void __vunmap(const void *addr, int deallocate_pages)
1134 struct vm_struct *area;
1139 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1140 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1144 area = remove_vm_area(addr);
1145 if (unlikely(!area)) {
1146 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1151 debug_check_no_locks_freed(addr, area->size);
1152 debug_check_no_obj_freed(addr, area->size);
1154 if (deallocate_pages) {
1157 for (i = 0; i < area->nr_pages; i++) {
1158 struct page *page = area->pages[i];
1164 if (area->flags & VM_VPAGES)
1175 * vfree - release memory allocated by vmalloc()
1176 * @addr: memory base address
1178 * Free the virtually continuous memory area starting at @addr, as
1179 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1180 * NULL, no operation is performed.
1182 * Must not be called in interrupt context.
1184 void vfree(const void *addr)
1186 BUG_ON(in_interrupt());
1189 EXPORT_SYMBOL(vfree);
1192 * vunmap - release virtual mapping obtained by vmap()
1193 * @addr: memory base address
1195 * Free the virtually contiguous memory area starting at @addr,
1196 * which was created from the page array passed to vmap().
1198 * Must not be called in interrupt context.
1200 void vunmap(const void *addr)
1202 BUG_ON(in_interrupt());
1205 EXPORT_SYMBOL(vunmap);
1208 * vmap - map an array of pages into virtually contiguous space
1209 * @pages: array of page pointers
1210 * @count: number of pages to map
1211 * @flags: vm_area->flags
1212 * @prot: page protection for the mapping
1214 * Maps @count pages from @pages into contiguous kernel virtual
1217 void *vmap(struct page **pages, unsigned int count,
1218 unsigned long flags, pgprot_t prot)
1220 struct vm_struct *area;
1222 if (count > num_physpages)
1225 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1226 __builtin_return_address(0));
1230 if (map_vm_area(area, prot, &pages)) {
1237 EXPORT_SYMBOL(vmap);
1239 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1240 int node, void *caller);
1241 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1242 pgprot_t prot, int node, void *caller)
1244 struct page **pages;
1245 unsigned int nr_pages, array_size, i;
1247 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1248 array_size = (nr_pages * sizeof(struct page *));
1250 area->nr_pages = nr_pages;
1251 /* Please note that the recursion is strictly bounded. */
1252 if (array_size > PAGE_SIZE) {
1253 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1254 PAGE_KERNEL, node, caller);
1255 area->flags |= VM_VPAGES;
1257 pages = kmalloc_node(array_size,
1258 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1261 area->pages = pages;
1262 area->caller = caller;
1264 remove_vm_area(area->addr);
1269 for (i = 0; i < area->nr_pages; i++) {
1273 page = alloc_page(gfp_mask);
1275 page = alloc_pages_node(node, gfp_mask, 0);
1277 if (unlikely(!page)) {
1278 /* Successfully allocated i pages, free them in __vunmap() */
1282 area->pages[i] = page;
1285 if (map_vm_area(area, prot, &pages))
1294 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1296 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1297 __builtin_return_address(0));
1301 * __vmalloc_node - allocate virtually contiguous memory
1302 * @size: allocation size
1303 * @gfp_mask: flags for the page level allocator
1304 * @prot: protection mask for the allocated pages
1305 * @node: node to use for allocation or -1
1306 * @caller: caller's return address
1308 * Allocate enough pages to cover @size from the page level
1309 * allocator with @gfp_mask flags. Map them into contiguous
1310 * kernel virtual space, using a pagetable protection of @prot.
1312 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1313 int node, void *caller)
1315 struct vm_struct *area;
1317 size = PAGE_ALIGN(size);
1318 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1321 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1322 node, gfp_mask, caller);
1327 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1330 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1332 return __vmalloc_node(size, gfp_mask, prot, -1,
1333 __builtin_return_address(0));
1335 EXPORT_SYMBOL(__vmalloc);
1338 * vmalloc - allocate virtually contiguous memory
1339 * @size: allocation size
1340 * Allocate enough pages to cover @size from the page level
1341 * allocator and map them into contiguous kernel virtual space.
1343 * For tight control over page level allocator and protection flags
1344 * use __vmalloc() instead.
1346 void *vmalloc(unsigned long size)
1348 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1349 -1, __builtin_return_address(0));
1351 EXPORT_SYMBOL(vmalloc);
1354 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1355 * @size: allocation size
1357 * The resulting memory area is zeroed so it can be mapped to userspace
1358 * without leaking data.
1360 void *vmalloc_user(unsigned long size)
1362 struct vm_struct *area;
1365 ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
1367 area = find_vm_area(ret);
1368 area->flags |= VM_USERMAP;
1372 EXPORT_SYMBOL(vmalloc_user);
1375 * vmalloc_node - allocate memory on a specific node
1376 * @size: allocation size
1379 * Allocate enough pages to cover @size from the page level
1380 * allocator and map them into contiguous kernel virtual space.
1382 * For tight control over page level allocator and protection flags
1383 * use __vmalloc() instead.
1385 void *vmalloc_node(unsigned long size, int node)
1387 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1388 node, __builtin_return_address(0));
1390 EXPORT_SYMBOL(vmalloc_node);
1392 #ifndef PAGE_KERNEL_EXEC
1393 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1397 * vmalloc_exec - allocate virtually contiguous, executable memory
1398 * @size: allocation size
1400 * Kernel-internal function to allocate enough pages to cover @size
1401 * the page level allocator and map them into contiguous and
1402 * executable kernel virtual space.
1404 * For tight control over page level allocator and protection flags
1405 * use __vmalloc() instead.
1408 void *vmalloc_exec(unsigned long size)
1410 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
1413 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1414 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1415 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1416 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1418 #define GFP_VMALLOC32 GFP_KERNEL
1422 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1423 * @size: allocation size
1425 * Allocate enough 32bit PA addressable pages to cover @size from the
1426 * page level allocator and map them into contiguous kernel virtual space.
1428 void *vmalloc_32(unsigned long size)
1430 return __vmalloc(size, GFP_VMALLOC32, PAGE_KERNEL);
1432 EXPORT_SYMBOL(vmalloc_32);
1435 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1436 * @size: allocation size
1438 * The resulting memory area is 32bit addressable and zeroed so it can be
1439 * mapped to userspace without leaking data.
1441 void *vmalloc_32_user(unsigned long size)
1443 struct vm_struct *area;
1446 ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
1448 area = find_vm_area(ret);
1449 area->flags |= VM_USERMAP;
1453 EXPORT_SYMBOL(vmalloc_32_user);
1455 long vread(char *buf, char *addr, unsigned long count)
1457 struct vm_struct *tmp;
1458 char *vaddr, *buf_start = buf;
1461 /* Don't allow overflow */
1462 if ((unsigned long) addr + count < count)
1463 count = -(unsigned long) addr;
1465 read_lock(&vmlist_lock);
1466 for (tmp = vmlist; tmp; tmp = tmp->next) {
1467 vaddr = (char *) tmp->addr;
1468 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1470 while (addr < vaddr) {
1478 n = vaddr + tmp->size - PAGE_SIZE - addr;
1489 read_unlock(&vmlist_lock);
1490 return buf - buf_start;
1493 long vwrite(char *buf, char *addr, unsigned long count)
1495 struct vm_struct *tmp;
1496 char *vaddr, *buf_start = buf;
1499 /* Don't allow overflow */
1500 if ((unsigned long) addr + count < count)
1501 count = -(unsigned long) addr;
1503 read_lock(&vmlist_lock);
1504 for (tmp = vmlist; tmp; tmp = tmp->next) {
1505 vaddr = (char *) tmp->addr;
1506 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1508 while (addr < vaddr) {
1515 n = vaddr + tmp->size - PAGE_SIZE - addr;
1526 read_unlock(&vmlist_lock);
1527 return buf - buf_start;
1531 * remap_vmalloc_range - map vmalloc pages to userspace
1532 * @vma: vma to cover (map full range of vma)
1533 * @addr: vmalloc memory
1534 * @pgoff: number of pages into addr before first page to map
1536 * Returns: 0 for success, -Exxx on failure
1538 * This function checks that addr is a valid vmalloc'ed area, and
1539 * that it is big enough to cover the vma. Will return failure if
1540 * that criteria isn't met.
1542 * Similar to remap_pfn_range() (see mm/memory.c)
1544 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1545 unsigned long pgoff)
1547 struct vm_struct *area;
1548 unsigned long uaddr = vma->vm_start;
1549 unsigned long usize = vma->vm_end - vma->vm_start;
1551 if ((PAGE_SIZE-1) & (unsigned long)addr)
1554 area = find_vm_area(addr);
1558 if (!(area->flags & VM_USERMAP))
1561 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1564 addr += pgoff << PAGE_SHIFT;
1566 struct page *page = vmalloc_to_page(addr);
1569 ret = vm_insert_page(vma, uaddr, page);
1576 } while (usize > 0);
1578 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1579 vma->vm_flags |= VM_RESERVED;
1583 EXPORT_SYMBOL(remap_vmalloc_range);
1586 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1589 void __attribute__((weak)) vmalloc_sync_all(void)
1594 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1596 /* apply_to_page_range() does all the hard work. */
1601 * alloc_vm_area - allocate a range of kernel address space
1602 * @size: size of the area
1604 * Returns: NULL on failure, vm_struct on success
1606 * This function reserves a range of kernel address space, and
1607 * allocates pagetables to map that range. No actual mappings
1608 * are created. If the kernel address space is not shared
1609 * between processes, it syncs the pagetable across all
1612 struct vm_struct *alloc_vm_area(size_t size)
1614 struct vm_struct *area;
1616 area = get_vm_area_caller(size, VM_IOREMAP,
1617 __builtin_return_address(0));
1622 * This ensures that page tables are constructed for this region
1623 * of kernel virtual address space and mapped into init_mm.
1625 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1626 area->size, f, NULL)) {
1631 /* Make sure the pagetables are constructed in process kernel
1637 EXPORT_SYMBOL_GPL(alloc_vm_area);
1639 void free_vm_area(struct vm_struct *area)
1641 struct vm_struct *ret;
1642 ret = remove_vm_area(area->addr);
1643 BUG_ON(ret != area);
1646 EXPORT_SYMBOL_GPL(free_vm_area);
1649 #ifdef CONFIG_PROC_FS
1650 static void *s_start(struct seq_file *m, loff_t *pos)
1653 struct vm_struct *v;
1655 read_lock(&vmlist_lock);
1657 while (n > 0 && v) {
1668 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1670 struct vm_struct *v = p;
1676 static void s_stop(struct seq_file *m, void *p)
1678 read_unlock(&vmlist_lock);
1681 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1684 unsigned int nr, *counters = m->private;
1689 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1691 for (nr = 0; nr < v->nr_pages; nr++)
1692 counters[page_to_nid(v->pages[nr])]++;
1694 for_each_node_state(nr, N_HIGH_MEMORY)
1696 seq_printf(m, " N%u=%u", nr, counters[nr]);
1700 static int s_show(struct seq_file *m, void *p)
1702 struct vm_struct *v = p;
1704 seq_printf(m, "0x%p-0x%p %7ld",
1705 v->addr, v->addr + v->size, v->size);
1708 char buff[2 * KSYM_NAME_LEN];
1711 sprint_symbol(buff, (unsigned long)v->caller);
1716 seq_printf(m, " pages=%d", v->nr_pages);
1719 seq_printf(m, " phys=%lx", v->phys_addr);
1721 if (v->flags & VM_IOREMAP)
1722 seq_printf(m, " ioremap");
1724 if (v->flags & VM_ALLOC)
1725 seq_printf(m, " vmalloc");
1727 if (v->flags & VM_MAP)
1728 seq_printf(m, " vmap");
1730 if (v->flags & VM_USERMAP)
1731 seq_printf(m, " user");
1733 if (v->flags & VM_VPAGES)
1734 seq_printf(m, " vpages");
1736 show_numa_info(m, v);
1741 static const struct seq_operations vmalloc_op = {
1748 static int vmalloc_open(struct inode *inode, struct file *file)
1750 unsigned int *ptr = NULL;
1754 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1755 ret = seq_open(file, &vmalloc_op);
1757 struct seq_file *m = file->private_data;
1764 static const struct file_operations proc_vmalloc_operations = {
1765 .open = vmalloc_open,
1767 .llseek = seq_lseek,
1768 .release = seq_release_private,
1771 static int __init proc_vmalloc_init(void)
1773 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1776 module_init(proc_vmalloc_init);