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/seq_file.h>
19 #include <linux/debugobjects.h>
20 #include <linux/kallsyms.h>
21 #include <linux/list.h>
22 #include <linux/rbtree.h>
23 #include <linux/radix-tree.h>
24 #include <linux/rcupdate.h>
26 #include <asm/atomic.h>
27 #include <asm/uaccess.h>
28 #include <asm/tlbflush.h>
31 /*** Page table manipulation functions ***/
33 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
37 pte = pte_offset_kernel(pmd, addr);
39 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
40 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
41 } while (pte++, addr += PAGE_SIZE, addr != end);
44 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
49 pmd = pmd_offset(pud, addr);
51 next = pmd_addr_end(addr, end);
52 if (pmd_none_or_clear_bad(pmd))
54 vunmap_pte_range(pmd, addr, next);
55 } while (pmd++, addr = next, addr != end);
58 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
63 pud = pud_offset(pgd, addr);
65 next = pud_addr_end(addr, end);
66 if (pud_none_or_clear_bad(pud))
68 vunmap_pmd_range(pud, addr, next);
69 } while (pud++, addr = next, addr != end);
72 static void vunmap_page_range(unsigned long addr, unsigned long end)
78 pgd = pgd_offset_k(addr);
79 flush_cache_vunmap(addr, end);
81 next = pgd_addr_end(addr, end);
82 if (pgd_none_or_clear_bad(pgd))
84 vunmap_pud_range(pgd, addr, next);
85 } while (pgd++, addr = next, addr != end);
88 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
89 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
94 * nr is a running index into the array which helps higher level
95 * callers keep track of where we're up to.
98 pte = pte_alloc_kernel(pmd, addr);
102 struct page *page = pages[*nr];
104 if (WARN_ON(!pte_none(*pte)))
108 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
110 } while (pte++, addr += PAGE_SIZE, addr != end);
114 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
115 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
120 pmd = pmd_alloc(&init_mm, pud, addr);
124 next = pmd_addr_end(addr, end);
125 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
127 } while (pmd++, addr = next, addr != end);
131 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
132 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
137 pud = pud_alloc(&init_mm, pgd, addr);
141 next = pud_addr_end(addr, end);
142 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
144 } while (pud++, addr = next, addr != end);
149 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
150 * will have pfns corresponding to the "pages" array.
152 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
154 static int vmap_page_range(unsigned long addr, unsigned long end,
155 pgprot_t prot, struct page **pages)
163 pgd = pgd_offset_k(addr);
165 next = pgd_addr_end(addr, end);
166 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
169 } while (pgd++, addr = next, addr != end);
170 flush_cache_vmap(addr, end);
177 static inline int is_vmalloc_or_module_addr(const void *x)
180 * x86-64 and sparc64 put modules in a special place,
181 * and fall back on vmalloc() if that fails. Others
182 * just put it in the vmalloc space.
184 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
185 unsigned long addr = (unsigned long)x;
186 if (addr >= MODULES_VADDR && addr < MODULES_END)
189 return is_vmalloc_addr(x);
193 * Walk a vmap address to the struct page it maps.
195 struct page *vmalloc_to_page(const void *vmalloc_addr)
197 unsigned long addr = (unsigned long) vmalloc_addr;
198 struct page *page = NULL;
199 pgd_t *pgd = pgd_offset_k(addr);
202 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
203 * architectures that do not vmalloc module space
205 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
207 if (!pgd_none(*pgd)) {
208 pud_t *pud = pud_offset(pgd, addr);
209 if (!pud_none(*pud)) {
210 pmd_t *pmd = pmd_offset(pud, addr);
211 if (!pmd_none(*pmd)) {
214 ptep = pte_offset_map(pmd, addr);
216 if (pte_present(pte))
217 page = pte_page(pte);
224 EXPORT_SYMBOL(vmalloc_to_page);
227 * Map a vmalloc()-space virtual address to the physical page frame number.
229 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
231 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
233 EXPORT_SYMBOL(vmalloc_to_pfn);
236 /*** Global kva allocator ***/
238 #define VM_LAZY_FREE 0x01
239 #define VM_LAZY_FREEING 0x02
240 #define VM_VM_AREA 0x04
243 unsigned long va_start;
244 unsigned long va_end;
246 struct rb_node rb_node; /* address sorted rbtree */
247 struct list_head list; /* address sorted list */
248 struct list_head purge_list; /* "lazy purge" list */
250 struct rcu_head rcu_head;
253 static DEFINE_SPINLOCK(vmap_area_lock);
254 static struct rb_root vmap_area_root = RB_ROOT;
255 static LIST_HEAD(vmap_area_list);
257 static struct vmap_area *__find_vmap_area(unsigned long addr)
259 struct rb_node *n = vmap_area_root.rb_node;
262 struct vmap_area *va;
264 va = rb_entry(n, struct vmap_area, rb_node);
265 if (addr < va->va_start)
267 else if (addr > va->va_start)
276 static void __insert_vmap_area(struct vmap_area *va)
278 struct rb_node **p = &vmap_area_root.rb_node;
279 struct rb_node *parent = NULL;
283 struct vmap_area *tmp;
286 tmp = rb_entry(parent, struct vmap_area, rb_node);
287 if (va->va_start < tmp->va_end)
289 else if (va->va_end > tmp->va_start)
295 rb_link_node(&va->rb_node, parent, p);
296 rb_insert_color(&va->rb_node, &vmap_area_root);
298 /* address-sort this list so it is usable like the vmlist */
299 tmp = rb_prev(&va->rb_node);
301 struct vmap_area *prev;
302 prev = rb_entry(tmp, struct vmap_area, rb_node);
303 list_add_rcu(&va->list, &prev->list);
305 list_add_rcu(&va->list, &vmap_area_list);
308 static void purge_vmap_area_lazy(void);
311 * Allocate a region of KVA of the specified size and alignment, within the
314 static struct vmap_area *alloc_vmap_area(unsigned long size,
316 unsigned long vstart, unsigned long vend,
317 int node, gfp_t gfp_mask)
319 struct vmap_area *va;
324 BUG_ON(size & ~PAGE_MASK);
326 addr = ALIGN(vstart, align);
328 va = kmalloc_node(sizeof(struct vmap_area),
329 gfp_mask & GFP_RECLAIM_MASK, node);
331 return ERR_PTR(-ENOMEM);
334 spin_lock(&vmap_area_lock);
335 /* XXX: could have a last_hole cache */
336 n = vmap_area_root.rb_node;
338 struct vmap_area *first = NULL;
341 struct vmap_area *tmp;
342 tmp = rb_entry(n, struct vmap_area, rb_node);
343 if (tmp->va_end >= addr) {
344 if (!first && tmp->va_start < addr + size)
356 if (first->va_end < addr) {
357 n = rb_next(&first->rb_node);
359 first = rb_entry(n, struct vmap_area, rb_node);
364 while (addr + size >= first->va_start && addr + size <= vend) {
365 addr = ALIGN(first->va_end + PAGE_SIZE, align);
367 n = rb_next(&first->rb_node);
369 first = rb_entry(n, struct vmap_area, rb_node);
375 if (addr + size > vend) {
376 spin_unlock(&vmap_area_lock);
378 purge_vmap_area_lazy();
382 if (printk_ratelimit())
383 printk(KERN_WARNING "vmap allocation failed: "
384 "use vmalloc=<size> to increase size.\n");
385 return ERR_PTR(-EBUSY);
388 BUG_ON(addr & (align-1));
391 va->va_end = addr + size;
393 __insert_vmap_area(va);
394 spin_unlock(&vmap_area_lock);
399 static void rcu_free_va(struct rcu_head *head)
401 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
406 static void __free_vmap_area(struct vmap_area *va)
408 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
409 rb_erase(&va->rb_node, &vmap_area_root);
410 RB_CLEAR_NODE(&va->rb_node);
411 list_del_rcu(&va->list);
413 call_rcu(&va->rcu_head, rcu_free_va);
417 * Free a region of KVA allocated by alloc_vmap_area
419 static void free_vmap_area(struct vmap_area *va)
421 spin_lock(&vmap_area_lock);
422 __free_vmap_area(va);
423 spin_unlock(&vmap_area_lock);
427 * Clear the pagetable entries of a given vmap_area
429 static void unmap_vmap_area(struct vmap_area *va)
431 vunmap_page_range(va->va_start, va->va_end);
435 * lazy_max_pages is the maximum amount of virtual address space we gather up
436 * before attempting to purge with a TLB flush.
438 * There is a tradeoff here: a larger number will cover more kernel page tables
439 * and take slightly longer to purge, but it will linearly reduce the number of
440 * global TLB flushes that must be performed. It would seem natural to scale
441 * this number up linearly with the number of CPUs (because vmapping activity
442 * could also scale linearly with the number of CPUs), however it is likely
443 * that in practice, workloads might be constrained in other ways that mean
444 * vmap activity will not scale linearly with CPUs. Also, I want to be
445 * conservative and not introduce a big latency on huge systems, so go with
446 * a less aggressive log scale. It will still be an improvement over the old
447 * code, and it will be simple to change the scale factor if we find that it
448 * becomes a problem on bigger systems.
450 static unsigned long lazy_max_pages(void)
454 log = fls(num_online_cpus());
456 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
459 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
462 * Purges all lazily-freed vmap areas.
464 * If sync is 0 then don't purge if there is already a purge in progress.
465 * If force_flush is 1, then flush kernel TLBs between *start and *end even
466 * if we found no lazy vmap areas to unmap (callers can use this to optimise
467 * their own TLB flushing).
468 * Returns with *start = min(*start, lowest purged address)
469 * *end = max(*end, highest purged address)
471 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
472 int sync, int force_flush)
474 static DEFINE_SPINLOCK(purge_lock);
476 struct vmap_area *va;
480 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
481 * should not expect such behaviour. This just simplifies locking for
482 * the case that isn't actually used at the moment anyway.
484 if (!sync && !force_flush) {
485 if (!spin_trylock(&purge_lock))
488 spin_lock(&purge_lock);
491 list_for_each_entry_rcu(va, &vmap_area_list, list) {
492 if (va->flags & VM_LAZY_FREE) {
493 if (va->va_start < *start)
494 *start = va->va_start;
495 if (va->va_end > *end)
497 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
499 list_add_tail(&va->purge_list, &valist);
500 va->flags |= VM_LAZY_FREEING;
501 va->flags &= ~VM_LAZY_FREE;
507 BUG_ON(nr > atomic_read(&vmap_lazy_nr));
508 atomic_sub(nr, &vmap_lazy_nr);
511 if (nr || force_flush)
512 flush_tlb_kernel_range(*start, *end);
515 spin_lock(&vmap_area_lock);
516 list_for_each_entry(va, &valist, purge_list)
517 __free_vmap_area(va);
518 spin_unlock(&vmap_area_lock);
520 spin_unlock(&purge_lock);
524 * Kick off a purge of the outstanding lazy areas.
526 static void purge_vmap_area_lazy(void)
528 unsigned long start = ULONG_MAX, end = 0;
530 __purge_vmap_area_lazy(&start, &end, 0, 0);
534 * Free and unmap a vmap area
536 static void free_unmap_vmap_area(struct vmap_area *va)
538 va->flags |= VM_LAZY_FREE;
539 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
540 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
541 purge_vmap_area_lazy();
544 static struct vmap_area *find_vmap_area(unsigned long addr)
546 struct vmap_area *va;
548 spin_lock(&vmap_area_lock);
549 va = __find_vmap_area(addr);
550 spin_unlock(&vmap_area_lock);
555 static void free_unmap_vmap_area_addr(unsigned long addr)
557 struct vmap_area *va;
559 va = find_vmap_area(addr);
561 free_unmap_vmap_area(va);
565 /*** Per cpu kva allocator ***/
568 * vmap space is limited especially on 32 bit architectures. Ensure there is
569 * room for at least 16 percpu vmap blocks per CPU.
572 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
573 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
574 * instead (we just need a rough idea)
576 #if BITS_PER_LONG == 32
577 #define VMALLOC_SPACE (128UL*1024*1024)
579 #define VMALLOC_SPACE (128UL*1024*1024*1024)
582 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
583 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
584 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
585 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
586 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
587 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
588 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
589 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
590 VMALLOC_PAGES / NR_CPUS / 16))
592 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
594 struct vmap_block_queue {
596 struct list_head free;
597 struct list_head dirty;
598 unsigned int nr_dirty;
603 struct vmap_area *va;
604 struct vmap_block_queue *vbq;
605 unsigned long free, dirty;
606 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
607 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
610 struct list_head free_list;
611 struct list_head dirty_list;
613 struct rcu_head rcu_head;
617 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
618 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
621 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
622 * in the free path. Could get rid of this if we change the API to return a
623 * "cookie" from alloc, to be passed to free. But no big deal yet.
625 static DEFINE_SPINLOCK(vmap_block_tree_lock);
626 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
629 * We should probably have a fallback mechanism to allocate virtual memory
630 * out of partially filled vmap blocks. However vmap block sizing should be
631 * fairly reasonable according to the vmalloc size, so it shouldn't be a
635 static unsigned long addr_to_vb_idx(unsigned long addr)
637 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
638 addr /= VMAP_BLOCK_SIZE;
642 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
644 struct vmap_block_queue *vbq;
645 struct vmap_block *vb;
646 struct vmap_area *va;
647 unsigned long vb_idx;
650 node = numa_node_id();
652 vb = kmalloc_node(sizeof(struct vmap_block),
653 gfp_mask & GFP_RECLAIM_MASK, node);
655 return ERR_PTR(-ENOMEM);
657 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
658 VMALLOC_START, VMALLOC_END,
660 if (unlikely(IS_ERR(va))) {
662 return ERR_PTR(PTR_ERR(va));
665 err = radix_tree_preload(gfp_mask);
672 spin_lock_init(&vb->lock);
674 vb->free = VMAP_BBMAP_BITS;
676 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
677 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
678 INIT_LIST_HEAD(&vb->free_list);
679 INIT_LIST_HEAD(&vb->dirty_list);
681 vb_idx = addr_to_vb_idx(va->va_start);
682 spin_lock(&vmap_block_tree_lock);
683 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
684 spin_unlock(&vmap_block_tree_lock);
686 radix_tree_preload_end();
688 vbq = &get_cpu_var(vmap_block_queue);
690 spin_lock(&vbq->lock);
691 list_add(&vb->free_list, &vbq->free);
692 spin_unlock(&vbq->lock);
693 put_cpu_var(vmap_cpu_blocks);
698 static void rcu_free_vb(struct rcu_head *head)
700 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
705 static void free_vmap_block(struct vmap_block *vb)
707 struct vmap_block *tmp;
708 unsigned long vb_idx;
710 spin_lock(&vb->vbq->lock);
711 if (!list_empty(&vb->free_list))
712 list_del(&vb->free_list);
713 if (!list_empty(&vb->dirty_list))
714 list_del(&vb->dirty_list);
715 spin_unlock(&vb->vbq->lock);
717 vb_idx = addr_to_vb_idx(vb->va->va_start);
718 spin_lock(&vmap_block_tree_lock);
719 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
720 spin_unlock(&vmap_block_tree_lock);
723 free_unmap_vmap_area(vb->va);
724 call_rcu(&vb->rcu_head, rcu_free_vb);
727 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
729 struct vmap_block_queue *vbq;
730 struct vmap_block *vb;
731 unsigned long addr = 0;
734 BUG_ON(size & ~PAGE_MASK);
735 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
736 order = get_order(size);
740 vbq = &get_cpu_var(vmap_block_queue);
741 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
744 spin_lock(&vb->lock);
745 i = bitmap_find_free_region(vb->alloc_map,
746 VMAP_BBMAP_BITS, order);
749 addr = vb->va->va_start + (i << PAGE_SHIFT);
750 BUG_ON(addr_to_vb_idx(addr) !=
751 addr_to_vb_idx(vb->va->va_start));
752 vb->free -= 1UL << order;
754 spin_lock(&vbq->lock);
755 list_del_init(&vb->free_list);
756 spin_unlock(&vbq->lock);
758 spin_unlock(&vb->lock);
761 spin_unlock(&vb->lock);
763 put_cpu_var(vmap_cpu_blocks);
767 vb = new_vmap_block(gfp_mask);
776 static void vb_free(const void *addr, unsigned long size)
778 unsigned long offset;
779 unsigned long vb_idx;
781 struct vmap_block *vb;
783 BUG_ON(size & ~PAGE_MASK);
784 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
785 order = get_order(size);
787 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
789 vb_idx = addr_to_vb_idx((unsigned long)addr);
791 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
795 spin_lock(&vb->lock);
796 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
798 spin_lock(&vb->vbq->lock);
799 list_add(&vb->dirty_list, &vb->vbq->dirty);
800 spin_unlock(&vb->vbq->lock);
802 vb->dirty += 1UL << order;
803 if (vb->dirty == VMAP_BBMAP_BITS) {
804 BUG_ON(vb->free || !list_empty(&vb->free_list));
805 spin_unlock(&vb->lock);
808 spin_unlock(&vb->lock);
812 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
814 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
815 * to amortize TLB flushing overheads. What this means is that any page you
816 * have now, may, in a former life, have been mapped into kernel virtual
817 * address by the vmap layer and so there might be some CPUs with TLB entries
818 * still referencing that page (additional to the regular 1:1 kernel mapping).
820 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
821 * be sure that none of the pages we have control over will have any aliases
822 * from the vmap layer.
824 void vm_unmap_aliases(void)
826 unsigned long start = ULONG_MAX, end = 0;
830 for_each_possible_cpu(cpu) {
831 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
832 struct vmap_block *vb;
835 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
838 spin_lock(&vb->lock);
839 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
840 while (i < VMAP_BBMAP_BITS) {
843 j = find_next_zero_bit(vb->dirty_map,
846 s = vb->va->va_start + (i << PAGE_SHIFT);
847 e = vb->va->va_start + (j << PAGE_SHIFT);
848 vunmap_page_range(s, e);
857 i = find_next_bit(vb->dirty_map,
860 spin_unlock(&vb->lock);
865 __purge_vmap_area_lazy(&start, &end, 1, flush);
867 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
870 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
871 * @mem: the pointer returned by vm_map_ram
872 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
874 void vm_unmap_ram(const void *mem, unsigned int count)
876 unsigned long size = count << PAGE_SHIFT;
877 unsigned long addr = (unsigned long)mem;
880 BUG_ON(addr < VMALLOC_START);
881 BUG_ON(addr > VMALLOC_END);
882 BUG_ON(addr & (PAGE_SIZE-1));
884 debug_check_no_locks_freed(mem, size);
886 if (likely(count <= VMAP_MAX_ALLOC))
889 free_unmap_vmap_area_addr(addr);
891 EXPORT_SYMBOL(vm_unmap_ram);
894 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
895 * @pages: an array of pointers to the pages to be mapped
896 * @count: number of pages
897 * @node: prefer to allocate data structures on this node
898 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
899 * @returns: a pointer to the address that has been mapped, or NULL on failure
901 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
903 unsigned long size = count << PAGE_SHIFT;
907 if (likely(count <= VMAP_MAX_ALLOC)) {
908 mem = vb_alloc(size, GFP_KERNEL);
911 addr = (unsigned long)mem;
913 struct vmap_area *va;
914 va = alloc_vmap_area(size, PAGE_SIZE,
915 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
922 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
923 vm_unmap_ram(mem, count);
928 EXPORT_SYMBOL(vm_map_ram);
930 void __init vmalloc_init(void)
934 for_each_possible_cpu(i) {
935 struct vmap_block_queue *vbq;
937 vbq = &per_cpu(vmap_block_queue, i);
938 spin_lock_init(&vbq->lock);
939 INIT_LIST_HEAD(&vbq->free);
940 INIT_LIST_HEAD(&vbq->dirty);
945 void unmap_kernel_range(unsigned long addr, unsigned long size)
947 unsigned long end = addr + size;
948 vunmap_page_range(addr, end);
949 flush_tlb_kernel_range(addr, end);
952 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
954 unsigned long addr = (unsigned long)area->addr;
955 unsigned long end = addr + area->size - PAGE_SIZE;
958 err = vmap_page_range(addr, end, prot, *pages);
966 EXPORT_SYMBOL_GPL(map_vm_area);
968 /*** Old vmalloc interfaces ***/
969 DEFINE_RWLOCK(vmlist_lock);
970 struct vm_struct *vmlist;
972 static struct vm_struct *__get_vm_area_node(unsigned long size,
973 unsigned long flags, unsigned long start, unsigned long end,
974 int node, gfp_t gfp_mask, void *caller)
976 static struct vmap_area *va;
977 struct vm_struct *area;
978 struct vm_struct *tmp, **p;
979 unsigned long align = 1;
981 BUG_ON(in_interrupt());
982 if (flags & VM_IOREMAP) {
985 if (bit > IOREMAP_MAX_ORDER)
986 bit = IOREMAP_MAX_ORDER;
987 else if (bit < PAGE_SHIFT)
993 size = PAGE_ALIGN(size);
997 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1002 * We always allocate a guard page.
1006 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1012 area->flags = flags;
1013 area->addr = (void *)va->va_start;
1017 area->phys_addr = 0;
1018 area->caller = caller;
1020 va->flags |= VM_VM_AREA;
1022 write_lock(&vmlist_lock);
1023 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1024 if (tmp->addr >= area->addr)
1029 write_unlock(&vmlist_lock);
1034 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1035 unsigned long start, unsigned long end)
1037 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1038 __builtin_return_address(0));
1040 EXPORT_SYMBOL_GPL(__get_vm_area);
1043 * get_vm_area - reserve a contiguous kernel virtual area
1044 * @size: size of the area
1045 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1047 * Search an area of @size in the kernel virtual mapping area,
1048 * and reserved it for out purposes. Returns the area descriptor
1049 * on success or %NULL on failure.
1051 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1053 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1054 -1, GFP_KERNEL, __builtin_return_address(0));
1057 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1060 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1061 -1, GFP_KERNEL, caller);
1064 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1065 int node, gfp_t gfp_mask)
1067 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1068 gfp_mask, __builtin_return_address(0));
1071 static struct vm_struct *find_vm_area(const void *addr)
1073 struct vmap_area *va;
1075 va = find_vmap_area((unsigned long)addr);
1076 if (va && va->flags & VM_VM_AREA)
1083 * remove_vm_area - find and remove a continuous kernel virtual area
1084 * @addr: base address
1086 * Search for the kernel VM area starting at @addr, and remove it.
1087 * This function returns the found VM area, but using it is NOT safe
1088 * on SMP machines, except for its size or flags.
1090 struct vm_struct *remove_vm_area(const void *addr)
1092 struct vmap_area *va;
1094 va = find_vmap_area((unsigned long)addr);
1095 if (va && va->flags & VM_VM_AREA) {
1096 struct vm_struct *vm = va->private;
1097 struct vm_struct *tmp, **p;
1098 free_unmap_vmap_area(va);
1099 vm->size -= PAGE_SIZE;
1101 write_lock(&vmlist_lock);
1102 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1105 write_unlock(&vmlist_lock);
1112 static void __vunmap(const void *addr, int deallocate_pages)
1114 struct vm_struct *area;
1119 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1120 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1124 area = remove_vm_area(addr);
1125 if (unlikely(!area)) {
1126 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1131 debug_check_no_locks_freed(addr, area->size);
1132 debug_check_no_obj_freed(addr, area->size);
1134 if (deallocate_pages) {
1137 for (i = 0; i < area->nr_pages; i++) {
1138 struct page *page = area->pages[i];
1144 if (area->flags & VM_VPAGES)
1155 * vfree - release memory allocated by vmalloc()
1156 * @addr: memory base address
1158 * Free the virtually continuous memory area starting at @addr, as
1159 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1160 * NULL, no operation is performed.
1162 * Must not be called in interrupt context.
1164 void vfree(const void *addr)
1166 BUG_ON(in_interrupt());
1169 EXPORT_SYMBOL(vfree);
1172 * vunmap - release virtual mapping obtained by vmap()
1173 * @addr: memory base address
1175 * Free the virtually contiguous memory area starting at @addr,
1176 * which was created from the page array passed to vmap().
1178 * Must not be called in interrupt context.
1180 void vunmap(const void *addr)
1182 BUG_ON(in_interrupt());
1185 EXPORT_SYMBOL(vunmap);
1188 * vmap - map an array of pages into virtually contiguous space
1189 * @pages: array of page pointers
1190 * @count: number of pages to map
1191 * @flags: vm_area->flags
1192 * @prot: page protection for the mapping
1194 * Maps @count pages from @pages into contiguous kernel virtual
1197 void *vmap(struct page **pages, unsigned int count,
1198 unsigned long flags, pgprot_t prot)
1200 struct vm_struct *area;
1202 if (count > num_physpages)
1205 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1206 __builtin_return_address(0));
1210 if (map_vm_area(area, prot, &pages)) {
1217 EXPORT_SYMBOL(vmap);
1219 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1220 int node, void *caller);
1221 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1222 pgprot_t prot, int node, void *caller)
1224 struct page **pages;
1225 unsigned int nr_pages, array_size, i;
1227 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1228 array_size = (nr_pages * sizeof(struct page *));
1230 area->nr_pages = nr_pages;
1231 /* Please note that the recursion is strictly bounded. */
1232 if (array_size > PAGE_SIZE) {
1233 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1234 PAGE_KERNEL, node, caller);
1235 area->flags |= VM_VPAGES;
1237 pages = kmalloc_node(array_size,
1238 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1241 area->pages = pages;
1242 area->caller = caller;
1244 remove_vm_area(area->addr);
1249 for (i = 0; i < area->nr_pages; i++) {
1253 page = alloc_page(gfp_mask);
1255 page = alloc_pages_node(node, gfp_mask, 0);
1257 if (unlikely(!page)) {
1258 /* Successfully allocated i pages, free them in __vunmap() */
1262 area->pages[i] = page;
1265 if (map_vm_area(area, prot, &pages))
1274 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1276 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1277 __builtin_return_address(0));
1281 * __vmalloc_node - allocate virtually contiguous memory
1282 * @size: allocation size
1283 * @gfp_mask: flags for the page level allocator
1284 * @prot: protection mask for the allocated pages
1285 * @node: node to use for allocation or -1
1286 * @caller: caller's return address
1288 * Allocate enough pages to cover @size from the page level
1289 * allocator with @gfp_mask flags. Map them into contiguous
1290 * kernel virtual space, using a pagetable protection of @prot.
1292 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1293 int node, void *caller)
1295 struct vm_struct *area;
1297 size = PAGE_ALIGN(size);
1298 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1301 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1302 node, gfp_mask, caller);
1307 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1310 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1312 return __vmalloc_node(size, gfp_mask, prot, -1,
1313 __builtin_return_address(0));
1315 EXPORT_SYMBOL(__vmalloc);
1318 * vmalloc - allocate virtually contiguous memory
1319 * @size: allocation size
1320 * Allocate enough pages to cover @size from the page level
1321 * allocator and map them into contiguous kernel virtual space.
1323 * For tight control over page level allocator and protection flags
1324 * use __vmalloc() instead.
1326 void *vmalloc(unsigned long size)
1328 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1329 -1, __builtin_return_address(0));
1331 EXPORT_SYMBOL(vmalloc);
1334 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1335 * @size: allocation size
1337 * The resulting memory area is zeroed so it can be mapped to userspace
1338 * without leaking data.
1340 void *vmalloc_user(unsigned long size)
1342 struct vm_struct *area;
1345 ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO, PAGE_KERNEL);
1347 area = find_vm_area(ret);
1348 area->flags |= VM_USERMAP;
1352 EXPORT_SYMBOL(vmalloc_user);
1355 * vmalloc_node - allocate memory on a specific node
1356 * @size: allocation size
1359 * Allocate enough pages to cover @size from the page level
1360 * allocator and map them into contiguous kernel virtual space.
1362 * For tight control over page level allocator and protection flags
1363 * use __vmalloc() instead.
1365 void *vmalloc_node(unsigned long size, int node)
1367 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1368 node, __builtin_return_address(0));
1370 EXPORT_SYMBOL(vmalloc_node);
1372 #ifndef PAGE_KERNEL_EXEC
1373 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1377 * vmalloc_exec - allocate virtually contiguous, executable memory
1378 * @size: allocation size
1380 * Kernel-internal function to allocate enough pages to cover @size
1381 * the page level allocator and map them into contiguous and
1382 * executable kernel virtual space.
1384 * For tight control over page level allocator and protection flags
1385 * use __vmalloc() instead.
1388 void *vmalloc_exec(unsigned long size)
1390 return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
1393 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1394 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1395 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1396 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1398 #define GFP_VMALLOC32 GFP_KERNEL
1402 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1403 * @size: allocation size
1405 * Allocate enough 32bit PA addressable pages to cover @size from the
1406 * page level allocator and map them into contiguous kernel virtual space.
1408 void *vmalloc_32(unsigned long size)
1410 return __vmalloc(size, GFP_VMALLOC32, PAGE_KERNEL);
1412 EXPORT_SYMBOL(vmalloc_32);
1415 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1416 * @size: allocation size
1418 * The resulting memory area is 32bit addressable and zeroed so it can be
1419 * mapped to userspace without leaking data.
1421 void *vmalloc_32_user(unsigned long size)
1423 struct vm_struct *area;
1426 ret = __vmalloc(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL);
1428 area = find_vm_area(ret);
1429 area->flags |= VM_USERMAP;
1433 EXPORT_SYMBOL(vmalloc_32_user);
1435 long vread(char *buf, char *addr, unsigned long count)
1437 struct vm_struct *tmp;
1438 char *vaddr, *buf_start = buf;
1441 /* Don't allow overflow */
1442 if ((unsigned long) addr + count < count)
1443 count = -(unsigned long) addr;
1445 read_lock(&vmlist_lock);
1446 for (tmp = vmlist; tmp; tmp = tmp->next) {
1447 vaddr = (char *) tmp->addr;
1448 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1450 while (addr < vaddr) {
1458 n = vaddr + tmp->size - PAGE_SIZE - addr;
1469 read_unlock(&vmlist_lock);
1470 return buf - buf_start;
1473 long vwrite(char *buf, char *addr, unsigned long count)
1475 struct vm_struct *tmp;
1476 char *vaddr, *buf_start = buf;
1479 /* Don't allow overflow */
1480 if ((unsigned long) addr + count < count)
1481 count = -(unsigned long) addr;
1483 read_lock(&vmlist_lock);
1484 for (tmp = vmlist; tmp; tmp = tmp->next) {
1485 vaddr = (char *) tmp->addr;
1486 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1488 while (addr < vaddr) {
1495 n = vaddr + tmp->size - PAGE_SIZE - addr;
1506 read_unlock(&vmlist_lock);
1507 return buf - buf_start;
1511 * remap_vmalloc_range - map vmalloc pages to userspace
1512 * @vma: vma to cover (map full range of vma)
1513 * @addr: vmalloc memory
1514 * @pgoff: number of pages into addr before first page to map
1516 * Returns: 0 for success, -Exxx on failure
1518 * This function checks that addr is a valid vmalloc'ed area, and
1519 * that it is big enough to cover the vma. Will return failure if
1520 * that criteria isn't met.
1522 * Similar to remap_pfn_range() (see mm/memory.c)
1524 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1525 unsigned long pgoff)
1527 struct vm_struct *area;
1528 unsigned long uaddr = vma->vm_start;
1529 unsigned long usize = vma->vm_end - vma->vm_start;
1531 if ((PAGE_SIZE-1) & (unsigned long)addr)
1534 area = find_vm_area(addr);
1538 if (!(area->flags & VM_USERMAP))
1541 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1544 addr += pgoff << PAGE_SHIFT;
1546 struct page *page = vmalloc_to_page(addr);
1549 ret = vm_insert_page(vma, uaddr, page);
1556 } while (usize > 0);
1558 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1559 vma->vm_flags |= VM_RESERVED;
1563 EXPORT_SYMBOL(remap_vmalloc_range);
1566 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1569 void __attribute__((weak)) vmalloc_sync_all(void)
1574 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1576 /* apply_to_page_range() does all the hard work. */
1581 * alloc_vm_area - allocate a range of kernel address space
1582 * @size: size of the area
1584 * Returns: NULL on failure, vm_struct on success
1586 * This function reserves a range of kernel address space, and
1587 * allocates pagetables to map that range. No actual mappings
1588 * are created. If the kernel address space is not shared
1589 * between processes, it syncs the pagetable across all
1592 struct vm_struct *alloc_vm_area(size_t size)
1594 struct vm_struct *area;
1596 area = get_vm_area_caller(size, VM_IOREMAP,
1597 __builtin_return_address(0));
1602 * This ensures that page tables are constructed for this region
1603 * of kernel virtual address space and mapped into init_mm.
1605 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1606 area->size, f, NULL)) {
1611 /* Make sure the pagetables are constructed in process kernel
1617 EXPORT_SYMBOL_GPL(alloc_vm_area);
1619 void free_vm_area(struct vm_struct *area)
1621 struct vm_struct *ret;
1622 ret = remove_vm_area(area->addr);
1623 BUG_ON(ret != area);
1626 EXPORT_SYMBOL_GPL(free_vm_area);
1629 #ifdef CONFIG_PROC_FS
1630 static void *s_start(struct seq_file *m, loff_t *pos)
1633 struct vm_struct *v;
1635 read_lock(&vmlist_lock);
1637 while (n > 0 && v) {
1648 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1650 struct vm_struct *v = p;
1656 static void s_stop(struct seq_file *m, void *p)
1658 read_unlock(&vmlist_lock);
1661 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1664 unsigned int nr, *counters = m->private;
1669 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1671 for (nr = 0; nr < v->nr_pages; nr++)
1672 counters[page_to_nid(v->pages[nr])]++;
1674 for_each_node_state(nr, N_HIGH_MEMORY)
1676 seq_printf(m, " N%u=%u", nr, counters[nr]);
1680 static int s_show(struct seq_file *m, void *p)
1682 struct vm_struct *v = p;
1684 seq_printf(m, "0x%p-0x%p %7ld",
1685 v->addr, v->addr + v->size, v->size);
1688 char buff[2 * KSYM_NAME_LEN];
1691 sprint_symbol(buff, (unsigned long)v->caller);
1696 seq_printf(m, " pages=%d", v->nr_pages);
1699 seq_printf(m, " phys=%lx", v->phys_addr);
1701 if (v->flags & VM_IOREMAP)
1702 seq_printf(m, " ioremap");
1704 if (v->flags & VM_ALLOC)
1705 seq_printf(m, " vmalloc");
1707 if (v->flags & VM_MAP)
1708 seq_printf(m, " vmap");
1710 if (v->flags & VM_USERMAP)
1711 seq_printf(m, " user");
1713 if (v->flags & VM_VPAGES)
1714 seq_printf(m, " vpages");
1716 show_numa_info(m, v);
1721 const struct seq_operations vmalloc_op = {