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 struct vmap_block_queue {
597 struct list_head free;
598 struct list_head dirty;
599 unsigned int nr_dirty;
604 struct vmap_area *va;
605 struct vmap_block_queue *vbq;
606 unsigned long free, dirty;
607 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
608 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
611 struct list_head free_list;
612 struct list_head dirty_list;
614 struct rcu_head rcu_head;
618 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
619 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
622 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
623 * in the free path. Could get rid of this if we change the API to return a
624 * "cookie" from alloc, to be passed to free. But no big deal yet.
626 static DEFINE_SPINLOCK(vmap_block_tree_lock);
627 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
630 * We should probably have a fallback mechanism to allocate virtual memory
631 * out of partially filled vmap blocks. However vmap block sizing should be
632 * fairly reasonable according to the vmalloc size, so it shouldn't be a
636 static unsigned long addr_to_vb_idx(unsigned long addr)
638 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
639 addr /= VMAP_BLOCK_SIZE;
643 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
645 struct vmap_block_queue *vbq;
646 struct vmap_block *vb;
647 struct vmap_area *va;
648 unsigned long vb_idx;
651 node = numa_node_id();
653 vb = kmalloc_node(sizeof(struct vmap_block),
654 gfp_mask & GFP_RECLAIM_MASK, node);
656 return ERR_PTR(-ENOMEM);
658 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
659 VMALLOC_START, VMALLOC_END,
661 if (unlikely(IS_ERR(va))) {
663 return ERR_PTR(PTR_ERR(va));
666 err = radix_tree_preload(gfp_mask);
673 spin_lock_init(&vb->lock);
675 vb->free = VMAP_BBMAP_BITS;
677 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
678 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
679 INIT_LIST_HEAD(&vb->free_list);
680 INIT_LIST_HEAD(&vb->dirty_list);
682 vb_idx = addr_to_vb_idx(va->va_start);
683 spin_lock(&vmap_block_tree_lock);
684 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
685 spin_unlock(&vmap_block_tree_lock);
687 radix_tree_preload_end();
689 vbq = &get_cpu_var(vmap_block_queue);
691 spin_lock(&vbq->lock);
692 list_add(&vb->free_list, &vbq->free);
693 spin_unlock(&vbq->lock);
694 put_cpu_var(vmap_cpu_blocks);
699 static void rcu_free_vb(struct rcu_head *head)
701 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
706 static void free_vmap_block(struct vmap_block *vb)
708 struct vmap_block *tmp;
709 unsigned long vb_idx;
711 spin_lock(&vb->vbq->lock);
712 if (!list_empty(&vb->free_list))
713 list_del(&vb->free_list);
714 if (!list_empty(&vb->dirty_list))
715 list_del(&vb->dirty_list);
716 spin_unlock(&vb->vbq->lock);
718 vb_idx = addr_to_vb_idx(vb->va->va_start);
719 spin_lock(&vmap_block_tree_lock);
720 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
721 spin_unlock(&vmap_block_tree_lock);
724 free_unmap_vmap_area(vb->va);
725 call_rcu(&vb->rcu_head, rcu_free_vb);
728 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
730 struct vmap_block_queue *vbq;
731 struct vmap_block *vb;
732 unsigned long addr = 0;
735 BUG_ON(size & ~PAGE_MASK);
736 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
737 order = get_order(size);
741 vbq = &get_cpu_var(vmap_block_queue);
742 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
745 spin_lock(&vb->lock);
746 i = bitmap_find_free_region(vb->alloc_map,
747 VMAP_BBMAP_BITS, order);
750 addr = vb->va->va_start + (i << PAGE_SHIFT);
751 BUG_ON(addr_to_vb_idx(addr) !=
752 addr_to_vb_idx(vb->va->va_start));
753 vb->free -= 1UL << order;
755 spin_lock(&vbq->lock);
756 list_del_init(&vb->free_list);
757 spin_unlock(&vbq->lock);
759 spin_unlock(&vb->lock);
762 spin_unlock(&vb->lock);
764 put_cpu_var(vmap_cpu_blocks);
768 vb = new_vmap_block(gfp_mask);
777 static void vb_free(const void *addr, unsigned long size)
779 unsigned long offset;
780 unsigned long vb_idx;
782 struct vmap_block *vb;
784 BUG_ON(size & ~PAGE_MASK);
785 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
786 order = get_order(size);
788 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
790 vb_idx = addr_to_vb_idx((unsigned long)addr);
792 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
796 spin_lock(&vb->lock);
797 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
799 spin_lock(&vb->vbq->lock);
800 list_add(&vb->dirty_list, &vb->vbq->dirty);
801 spin_unlock(&vb->vbq->lock);
803 vb->dirty += 1UL << order;
804 if (vb->dirty == VMAP_BBMAP_BITS) {
805 BUG_ON(vb->free || !list_empty(&vb->free_list));
806 spin_unlock(&vb->lock);
809 spin_unlock(&vb->lock);
813 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
815 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
816 * to amortize TLB flushing overheads. What this means is that any page you
817 * have now, may, in a former life, have been mapped into kernel virtual
818 * address by the vmap layer and so there might be some CPUs with TLB entries
819 * still referencing that page (additional to the regular 1:1 kernel mapping).
821 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
822 * be sure that none of the pages we have control over will have any aliases
823 * from the vmap layer.
825 void vm_unmap_aliases(void)
827 unsigned long start = ULONG_MAX, end = 0;
831 for_each_possible_cpu(cpu) {
832 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
833 struct vmap_block *vb;
836 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
839 spin_lock(&vb->lock);
840 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
841 while (i < VMAP_BBMAP_BITS) {
844 j = find_next_zero_bit(vb->dirty_map,
847 s = vb->va->va_start + (i << PAGE_SHIFT);
848 e = vb->va->va_start + (j << PAGE_SHIFT);
849 vunmap_page_range(s, e);
858 i = find_next_bit(vb->dirty_map,
861 spin_unlock(&vb->lock);
866 __purge_vmap_area_lazy(&start, &end, 1, flush);
868 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
871 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
872 * @mem: the pointer returned by vm_map_ram
873 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
875 void vm_unmap_ram(const void *mem, unsigned int count)
877 unsigned long size = count << PAGE_SHIFT;
878 unsigned long addr = (unsigned long)mem;
881 BUG_ON(addr < VMALLOC_START);
882 BUG_ON(addr > VMALLOC_END);
883 BUG_ON(addr & (PAGE_SIZE-1));
885 debug_check_no_locks_freed(mem, size);
887 if (likely(count <= VMAP_MAX_ALLOC))
890 free_unmap_vmap_area_addr(addr);
892 EXPORT_SYMBOL(vm_unmap_ram);
895 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
896 * @pages: an array of pointers to the pages to be mapped
897 * @count: number of pages
898 * @node: prefer to allocate data structures on this node
899 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
901 * Returns: a pointer to the address that has been mapped, or %NULL on failure
903 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
905 unsigned long size = count << PAGE_SHIFT;
909 if (likely(count <= VMAP_MAX_ALLOC)) {
910 mem = vb_alloc(size, GFP_KERNEL);
913 addr = (unsigned long)mem;
915 struct vmap_area *va;
916 va = alloc_vmap_area(size, PAGE_SIZE,
917 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
924 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
925 vm_unmap_ram(mem, count);
930 EXPORT_SYMBOL(vm_map_ram);
932 void __init vmalloc_init(void)
936 for_each_possible_cpu(i) {
937 struct vmap_block_queue *vbq;
939 vbq = &per_cpu(vmap_block_queue, i);
940 spin_lock_init(&vbq->lock);
941 INIT_LIST_HEAD(&vbq->free);
942 INIT_LIST_HEAD(&vbq->dirty);
947 void unmap_kernel_range(unsigned long addr, unsigned long size)
949 unsigned long end = addr + size;
950 vunmap_page_range(addr, end);
951 flush_tlb_kernel_range(addr, end);
954 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
956 unsigned long addr = (unsigned long)area->addr;
957 unsigned long end = addr + area->size - PAGE_SIZE;
960 err = vmap_page_range(addr, end, prot, *pages);
968 EXPORT_SYMBOL_GPL(map_vm_area);
970 /*** Old vmalloc interfaces ***/
971 DEFINE_RWLOCK(vmlist_lock);
972 struct vm_struct *vmlist;
974 static struct vm_struct *__get_vm_area_node(unsigned long size,
975 unsigned long flags, unsigned long start, unsigned long end,
976 int node, gfp_t gfp_mask, void *caller)
978 static struct vmap_area *va;
979 struct vm_struct *area;
980 struct vm_struct *tmp, **p;
981 unsigned long align = 1;
983 BUG_ON(in_interrupt());
984 if (flags & VM_IOREMAP) {
987 if (bit > IOREMAP_MAX_ORDER)
988 bit = IOREMAP_MAX_ORDER;
989 else if (bit < PAGE_SHIFT)
995 size = PAGE_ALIGN(size);
999 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1000 if (unlikely(!area))
1004 * We always allocate a guard page.
1008 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1014 area->flags = flags;
1015 area->addr = (void *)va->va_start;
1019 area->phys_addr = 0;
1020 area->caller = caller;
1022 va->flags |= VM_VM_AREA;
1024 write_lock(&vmlist_lock);
1025 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1026 if (tmp->addr >= area->addr)
1031 write_unlock(&vmlist_lock);
1036 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1037 unsigned long start, unsigned long end)
1039 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1040 __builtin_return_address(0));
1042 EXPORT_SYMBOL_GPL(__get_vm_area);
1045 * get_vm_area - reserve a contiguous kernel virtual area
1046 * @size: size of the area
1047 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1049 * Search an area of @size in the kernel virtual mapping area,
1050 * and reserved it for out purposes. Returns the area descriptor
1051 * on success or %NULL on failure.
1053 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1055 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1056 -1, GFP_KERNEL, __builtin_return_address(0));
1059 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1062 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1063 -1, GFP_KERNEL, caller);
1066 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1067 int node, gfp_t gfp_mask)
1069 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1070 gfp_mask, __builtin_return_address(0));
1073 static struct vm_struct *find_vm_area(const void *addr)
1075 struct vmap_area *va;
1077 va = find_vmap_area((unsigned long)addr);
1078 if (va && va->flags & VM_VM_AREA)
1085 * remove_vm_area - find and remove a continuous kernel virtual area
1086 * @addr: base address
1088 * Search for the kernel VM area starting at @addr, and remove it.
1089 * This function returns the found VM area, but using it is NOT safe
1090 * on SMP machines, except for its size or flags.
1092 struct vm_struct *remove_vm_area(const void *addr)
1094 struct vmap_area *va;
1096 va = find_vmap_area((unsigned long)addr);
1097 if (va && va->flags & VM_VM_AREA) {
1098 struct vm_struct *vm = va->private;
1099 struct vm_struct *tmp, **p;
1100 free_unmap_vmap_area(va);
1101 vm->size -= PAGE_SIZE;
1103 write_lock(&vmlist_lock);
1104 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1107 write_unlock(&vmlist_lock);
1114 static void __vunmap(const void *addr, int deallocate_pages)
1116 struct vm_struct *area;
1121 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1122 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1126 area = remove_vm_area(addr);
1127 if (unlikely(!area)) {
1128 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1133 debug_check_no_locks_freed(addr, area->size);
1134 debug_check_no_obj_freed(addr, area->size);
1136 if (deallocate_pages) {
1139 for (i = 0; i < area->nr_pages; i++) {
1140 struct page *page = area->pages[i];
1146 if (area->flags & VM_VPAGES)
1157 * vfree - release memory allocated by vmalloc()
1158 * @addr: memory base address
1160 * Free the virtually continuous memory area starting at @addr, as
1161 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1162 * NULL, no operation is performed.
1164 * Must not be called in interrupt context.
1166 void vfree(const void *addr)
1168 BUG_ON(in_interrupt());
1171 EXPORT_SYMBOL(vfree);
1174 * vunmap - release virtual mapping obtained by vmap()
1175 * @addr: memory base address
1177 * Free the virtually contiguous memory area starting at @addr,
1178 * which was created from the page array passed to vmap().
1180 * Must not be called in interrupt context.
1182 void vunmap(const void *addr)
1184 BUG_ON(in_interrupt());
1187 EXPORT_SYMBOL(vunmap);
1190 * vmap - map an array of pages into virtually contiguous space
1191 * @pages: array of page pointers
1192 * @count: number of pages to map
1193 * @flags: vm_area->flags
1194 * @prot: page protection for the mapping
1196 * Maps @count pages from @pages into contiguous kernel virtual
1199 void *vmap(struct page **pages, unsigned int count,
1200 unsigned long flags, pgprot_t prot)
1202 struct vm_struct *area;
1204 if (count > num_physpages)
1207 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1208 __builtin_return_address(0));
1212 if (map_vm_area(area, prot, &pages)) {
1219 EXPORT_SYMBOL(vmap);
1221 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1222 int node, void *caller);
1223 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1224 pgprot_t prot, int node, void *caller)
1226 struct page **pages;
1227 unsigned int nr_pages, array_size, i;
1229 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1230 array_size = (nr_pages * sizeof(struct page *));
1232 area->nr_pages = nr_pages;
1233 /* Please note that the recursion is strictly bounded. */
1234 if (array_size > PAGE_SIZE) {
1235 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1236 PAGE_KERNEL, node, caller);
1237 area->flags |= VM_VPAGES;
1239 pages = kmalloc_node(array_size,
1240 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1243 area->pages = pages;
1244 area->caller = caller;
1246 remove_vm_area(area->addr);
1251 for (i = 0; i < area->nr_pages; i++) {
1255 page = alloc_page(gfp_mask);
1257 page = alloc_pages_node(node, gfp_mask, 0);
1259 if (unlikely(!page)) {
1260 /* Successfully allocated i pages, free them in __vunmap() */
1264 area->pages[i] = page;
1267 if (map_vm_area(area, prot, &pages))
1276 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1278 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1279 __builtin_return_address(0));
1283 * __vmalloc_node - allocate virtually contiguous memory
1284 * @size: allocation size
1285 * @gfp_mask: flags for the page level allocator
1286 * @prot: protection mask for the allocated pages
1287 * @node: node to use for allocation or -1
1288 * @caller: caller's return address
1290 * Allocate enough pages to cover @size from the page level
1291 * allocator with @gfp_mask flags. Map them into contiguous
1292 * kernel virtual space, using a pagetable protection of @prot.
1294 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1295 int node, void *caller)
1297 struct vm_struct *area;
1299 size = PAGE_ALIGN(size);
1300 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1303 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1304 node, gfp_mask, caller);
1309 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1312 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1314 return __vmalloc_node(size, gfp_mask, prot, -1,
1315 __builtin_return_address(0));
1317 EXPORT_SYMBOL(__vmalloc);
1320 * vmalloc - allocate virtually contiguous memory
1321 * @size: allocation size
1322 * Allocate enough pages to cover @size from the page level
1323 * allocator and map them into contiguous kernel virtual space.
1325 * For tight control over page level allocator and protection flags
1326 * use __vmalloc() instead.
1328 void *vmalloc(unsigned long size)
1330 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1331 -1, __builtin_return_address(0));
1333 EXPORT_SYMBOL(vmalloc);
1336 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1337 * @size: allocation size
1339 * The resulting memory area is zeroed so it can be mapped to userspace
1340 * without leaking data.
1342 void *vmalloc_user(unsigned long size)
1344 struct vm_struct *area;
1347 ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
1349 area = find_vm_area(ret);
1350 area->flags |= VM_USERMAP;
1354 EXPORT_SYMBOL(vmalloc_user);
1357 * vmalloc_node - allocate memory on a specific node
1358 * @size: allocation size
1361 * Allocate enough pages to cover @size from the page level
1362 * allocator and map them into contiguous kernel virtual space.
1364 * For tight control over page level allocator and protection flags
1365 * use __vmalloc() instead.
1367 void *vmalloc_node(unsigned long size, int node)
1369 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1370 node, __builtin_return_address(0));
1372 EXPORT_SYMBOL(vmalloc_node);
1374 #ifndef PAGE_KERNEL_EXEC
1375 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1379 * vmalloc_exec - allocate virtually contiguous, executable memory
1380 * @size: allocation size
1382 * Kernel-internal function to allocate enough pages to cover @size
1383 * the page level allocator and map them into contiguous and
1384 * executable kernel virtual space.
1386 * For tight control over page level allocator and protection flags
1387 * use __vmalloc() instead.
1390 void *vmalloc_exec(unsigned long size)
1392 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
1395 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1396 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1397 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1398 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1400 #define GFP_VMALLOC32 GFP_KERNEL
1404 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1405 * @size: allocation size
1407 * Allocate enough 32bit PA addressable pages to cover @size from the
1408 * page level allocator and map them into contiguous kernel virtual space.
1410 void *vmalloc_32(unsigned long size)
1412 return __vmalloc(size, GFP_VMALLOC32, PAGE_KERNEL);
1414 EXPORT_SYMBOL(vmalloc_32);
1417 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1418 * @size: allocation size
1420 * The resulting memory area is 32bit addressable and zeroed so it can be
1421 * mapped to userspace without leaking data.
1423 void *vmalloc_32_user(unsigned long size)
1425 struct vm_struct *area;
1428 ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
1430 area = find_vm_area(ret);
1431 area->flags |= VM_USERMAP;
1435 EXPORT_SYMBOL(vmalloc_32_user);
1437 long vread(char *buf, char *addr, unsigned long count)
1439 struct vm_struct *tmp;
1440 char *vaddr, *buf_start = buf;
1443 /* Don't allow overflow */
1444 if ((unsigned long) addr + count < count)
1445 count = -(unsigned long) addr;
1447 read_lock(&vmlist_lock);
1448 for (tmp = vmlist; tmp; tmp = tmp->next) {
1449 vaddr = (char *) tmp->addr;
1450 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1452 while (addr < vaddr) {
1460 n = vaddr + tmp->size - PAGE_SIZE - addr;
1471 read_unlock(&vmlist_lock);
1472 return buf - buf_start;
1475 long vwrite(char *buf, char *addr, unsigned long count)
1477 struct vm_struct *tmp;
1478 char *vaddr, *buf_start = buf;
1481 /* Don't allow overflow */
1482 if ((unsigned long) addr + count < count)
1483 count = -(unsigned long) addr;
1485 read_lock(&vmlist_lock);
1486 for (tmp = vmlist; tmp; tmp = tmp->next) {
1487 vaddr = (char *) tmp->addr;
1488 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1490 while (addr < vaddr) {
1497 n = vaddr + tmp->size - PAGE_SIZE - addr;
1508 read_unlock(&vmlist_lock);
1509 return buf - buf_start;
1513 * remap_vmalloc_range - map vmalloc pages to userspace
1514 * @vma: vma to cover (map full range of vma)
1515 * @addr: vmalloc memory
1516 * @pgoff: number of pages into addr before first page to map
1518 * Returns: 0 for success, -Exxx on failure
1520 * This function checks that addr is a valid vmalloc'ed area, and
1521 * that it is big enough to cover the vma. Will return failure if
1522 * that criteria isn't met.
1524 * Similar to remap_pfn_range() (see mm/memory.c)
1526 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1527 unsigned long pgoff)
1529 struct vm_struct *area;
1530 unsigned long uaddr = vma->vm_start;
1531 unsigned long usize = vma->vm_end - vma->vm_start;
1533 if ((PAGE_SIZE-1) & (unsigned long)addr)
1536 area = find_vm_area(addr);
1540 if (!(area->flags & VM_USERMAP))
1543 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1546 addr += pgoff << PAGE_SHIFT;
1548 struct page *page = vmalloc_to_page(addr);
1551 ret = vm_insert_page(vma, uaddr, page);
1558 } while (usize > 0);
1560 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1561 vma->vm_flags |= VM_RESERVED;
1565 EXPORT_SYMBOL(remap_vmalloc_range);
1568 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1571 void __attribute__((weak)) vmalloc_sync_all(void)
1576 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1578 /* apply_to_page_range() does all the hard work. */
1583 * alloc_vm_area - allocate a range of kernel address space
1584 * @size: size of the area
1586 * Returns: NULL on failure, vm_struct on success
1588 * This function reserves a range of kernel address space, and
1589 * allocates pagetables to map that range. No actual mappings
1590 * are created. If the kernel address space is not shared
1591 * between processes, it syncs the pagetable across all
1594 struct vm_struct *alloc_vm_area(size_t size)
1596 struct vm_struct *area;
1598 area = get_vm_area_caller(size, VM_IOREMAP,
1599 __builtin_return_address(0));
1604 * This ensures that page tables are constructed for this region
1605 * of kernel virtual address space and mapped into init_mm.
1607 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1608 area->size, f, NULL)) {
1613 /* Make sure the pagetables are constructed in process kernel
1619 EXPORT_SYMBOL_GPL(alloc_vm_area);
1621 void free_vm_area(struct vm_struct *area)
1623 struct vm_struct *ret;
1624 ret = remove_vm_area(area->addr);
1625 BUG_ON(ret != area);
1628 EXPORT_SYMBOL_GPL(free_vm_area);
1631 #ifdef CONFIG_PROC_FS
1632 static void *s_start(struct seq_file *m, loff_t *pos)
1635 struct vm_struct *v;
1637 read_lock(&vmlist_lock);
1639 while (n > 0 && v) {
1650 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1652 struct vm_struct *v = p;
1658 static void s_stop(struct seq_file *m, void *p)
1660 read_unlock(&vmlist_lock);
1663 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1666 unsigned int nr, *counters = m->private;
1671 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1673 for (nr = 0; nr < v->nr_pages; nr++)
1674 counters[page_to_nid(v->pages[nr])]++;
1676 for_each_node_state(nr, N_HIGH_MEMORY)
1678 seq_printf(m, " N%u=%u", nr, counters[nr]);
1682 static int s_show(struct seq_file *m, void *p)
1684 struct vm_struct *v = p;
1686 seq_printf(m, "0x%p-0x%p %7ld",
1687 v->addr, v->addr + v->size, v->size);
1690 char buff[2 * KSYM_NAME_LEN];
1693 sprint_symbol(buff, (unsigned long)v->caller);
1698 seq_printf(m, " pages=%d", v->nr_pages);
1701 seq_printf(m, " phys=%lx", v->phys_addr);
1703 if (v->flags & VM_IOREMAP)
1704 seq_printf(m, " ioremap");
1706 if (v->flags & VM_ALLOC)
1707 seq_printf(m, " vmalloc");
1709 if (v->flags & VM_MAP)
1710 seq_printf(m, " vmap");
1712 if (v->flags & VM_USERMAP)
1713 seq_printf(m, " user");
1715 if (v->flags & VM_VPAGES)
1716 seq_printf(m, " vpages");
1718 show_numa_info(m, v);
1723 static const struct seq_operations vmalloc_op = {
1730 static int vmalloc_open(struct inode *inode, struct file *file)
1732 unsigned int *ptr = NULL;
1736 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1737 ret = seq_open(file, &vmalloc_op);
1739 struct seq_file *m = file->private_data;
1746 static const struct file_operations proc_vmalloc_operations = {
1747 .open = vmalloc_open,
1749 .llseek = seq_lseek,
1750 .release = seq_release_private,
1753 static int __init proc_vmalloc_init(void)
1755 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1758 module_init(proc_vmalloc_init);