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/mutex.h>
18 #include <linux/interrupt.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/debugobjects.h>
22 #include <linux/kallsyms.h>
23 #include <linux/list.h>
24 #include <linux/rbtree.h>
25 #include <linux/radix-tree.h>
26 #include <linux/rcupdate.h>
28 #include <asm/atomic.h>
29 #include <asm/uaccess.h>
30 #include <asm/tlbflush.h>
33 /*** Page table manipulation functions ***/
35 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
39 pte = pte_offset_kernel(pmd, addr);
41 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
42 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
43 } while (pte++, addr += PAGE_SIZE, addr != end);
46 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
51 pmd = pmd_offset(pud, addr);
53 next = pmd_addr_end(addr, end);
54 if (pmd_none_or_clear_bad(pmd))
56 vunmap_pte_range(pmd, addr, next);
57 } while (pmd++, addr = next, addr != end);
60 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
65 pud = pud_offset(pgd, addr);
67 next = pud_addr_end(addr, end);
68 if (pud_none_or_clear_bad(pud))
70 vunmap_pmd_range(pud, addr, next);
71 } while (pud++, addr = next, addr != end);
74 static void vunmap_page_range(unsigned long addr, unsigned long end)
80 pgd = pgd_offset_k(addr);
82 next = pgd_addr_end(addr, end);
83 if (pgd_none_or_clear_bad(pgd))
85 vunmap_pud_range(pgd, addr, next);
86 } while (pgd++, addr = next, addr != end);
89 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
90 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
95 * nr is a running index into the array which helps higher level
96 * callers keep track of where we're up to.
99 pte = pte_alloc_kernel(pmd, addr);
103 struct page *page = pages[*nr];
105 if (WARN_ON(!pte_none(*pte)))
109 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
111 } while (pte++, addr += PAGE_SIZE, addr != end);
115 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
116 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
121 pmd = pmd_alloc(&init_mm, pud, addr);
125 next = pmd_addr_end(addr, end);
126 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
128 } while (pmd++, addr = next, addr != end);
132 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
133 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
138 pud = pud_alloc(&init_mm, pgd, addr);
142 next = pud_addr_end(addr, end);
143 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
145 } while (pud++, addr = next, addr != end);
150 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
151 * will have pfns corresponding to the "pages" array.
153 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
155 static int vmap_page_range(unsigned long start, unsigned long end,
156 pgprot_t prot, struct page **pages)
160 unsigned long addr = start;
165 pgd = pgd_offset_k(addr);
167 next = pgd_addr_end(addr, end);
168 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
171 } while (pgd++, addr = next, addr != end);
172 flush_cache_vmap(start, end);
179 static inline int is_vmalloc_or_module_addr(const void *x)
182 * ARM, x86-64 and sparc64 put modules in a special place,
183 * and fall back on vmalloc() if that fails. Others
184 * just put it in the vmalloc space.
186 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
187 unsigned long addr = (unsigned long)x;
188 if (addr >= MODULES_VADDR && addr < MODULES_END)
191 return is_vmalloc_addr(x);
195 * Walk a vmap address to the struct page it maps.
197 struct page *vmalloc_to_page(const void *vmalloc_addr)
199 unsigned long addr = (unsigned long) vmalloc_addr;
200 struct page *page = NULL;
201 pgd_t *pgd = pgd_offset_k(addr);
204 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
205 * architectures that do not vmalloc module space
207 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
209 if (!pgd_none(*pgd)) {
210 pud_t *pud = pud_offset(pgd, addr);
211 if (!pud_none(*pud)) {
212 pmd_t *pmd = pmd_offset(pud, addr);
213 if (!pmd_none(*pmd)) {
216 ptep = pte_offset_map(pmd, addr);
218 if (pte_present(pte))
219 page = pte_page(pte);
226 EXPORT_SYMBOL(vmalloc_to_page);
229 * Map a vmalloc()-space virtual address to the physical page frame number.
231 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
233 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
235 EXPORT_SYMBOL(vmalloc_to_pfn);
238 /*** Global kva allocator ***/
240 #define VM_LAZY_FREE 0x01
241 #define VM_LAZY_FREEING 0x02
242 #define VM_VM_AREA 0x04
245 unsigned long va_start;
246 unsigned long va_end;
248 struct rb_node rb_node; /* address sorted rbtree */
249 struct list_head list; /* address sorted list */
250 struct list_head purge_list; /* "lazy purge" list */
252 struct rcu_head rcu_head;
255 static DEFINE_SPINLOCK(vmap_area_lock);
256 static struct rb_root vmap_area_root = RB_ROOT;
257 static LIST_HEAD(vmap_area_list);
259 static struct vmap_area *__find_vmap_area(unsigned long addr)
261 struct rb_node *n = vmap_area_root.rb_node;
264 struct vmap_area *va;
266 va = rb_entry(n, struct vmap_area, rb_node);
267 if (addr < va->va_start)
269 else if (addr > va->va_start)
278 static void __insert_vmap_area(struct vmap_area *va)
280 struct rb_node **p = &vmap_area_root.rb_node;
281 struct rb_node *parent = NULL;
285 struct vmap_area *tmp;
288 tmp = rb_entry(parent, struct vmap_area, rb_node);
289 if (va->va_start < tmp->va_end)
291 else if (va->va_end > tmp->va_start)
297 rb_link_node(&va->rb_node, parent, p);
298 rb_insert_color(&va->rb_node, &vmap_area_root);
300 /* address-sort this list so it is usable like the vmlist */
301 tmp = rb_prev(&va->rb_node);
303 struct vmap_area *prev;
304 prev = rb_entry(tmp, struct vmap_area, rb_node);
305 list_add_rcu(&va->list, &prev->list);
307 list_add_rcu(&va->list, &vmap_area_list);
310 static void purge_vmap_area_lazy(void);
313 * Allocate a region of KVA of the specified size and alignment, within the
316 static struct vmap_area *alloc_vmap_area(unsigned long size,
318 unsigned long vstart, unsigned long vend,
319 int node, gfp_t gfp_mask)
321 struct vmap_area *va;
326 BUG_ON(size & ~PAGE_MASK);
328 va = kmalloc_node(sizeof(struct vmap_area),
329 gfp_mask & GFP_RECLAIM_MASK, node);
331 return ERR_PTR(-ENOMEM);
334 addr = ALIGN(vstart, align);
336 spin_lock(&vmap_area_lock);
337 /* XXX: could have a last_hole cache */
338 n = vmap_area_root.rb_node;
340 struct vmap_area *first = NULL;
343 struct vmap_area *tmp;
344 tmp = rb_entry(n, struct vmap_area, rb_node);
345 if (tmp->va_end >= addr) {
346 if (!first && tmp->va_start < addr + size)
358 if (first->va_end < addr) {
359 n = rb_next(&first->rb_node);
361 first = rb_entry(n, struct vmap_area, rb_node);
366 while (addr + size > first->va_start && addr + size <= vend) {
367 addr = ALIGN(first->va_end + PAGE_SIZE, align);
369 n = rb_next(&first->rb_node);
371 first = rb_entry(n, struct vmap_area, rb_node);
377 if (addr + size > vend) {
378 spin_unlock(&vmap_area_lock);
380 purge_vmap_area_lazy();
384 if (printk_ratelimit())
386 "vmap allocation for size %lu failed: "
387 "use vmalloc=<size> to increase size.\n", size);
388 return ERR_PTR(-EBUSY);
391 BUG_ON(addr & (align-1));
394 va->va_end = addr + size;
396 __insert_vmap_area(va);
397 spin_unlock(&vmap_area_lock);
402 static void rcu_free_va(struct rcu_head *head)
404 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
409 static void __free_vmap_area(struct vmap_area *va)
411 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
412 rb_erase(&va->rb_node, &vmap_area_root);
413 RB_CLEAR_NODE(&va->rb_node);
414 list_del_rcu(&va->list);
416 call_rcu(&va->rcu_head, rcu_free_va);
420 * Free a region of KVA allocated by alloc_vmap_area
422 static void free_vmap_area(struct vmap_area *va)
424 spin_lock(&vmap_area_lock);
425 __free_vmap_area(va);
426 spin_unlock(&vmap_area_lock);
430 * Clear the pagetable entries of a given vmap_area
432 static void unmap_vmap_area(struct vmap_area *va)
434 vunmap_page_range(va->va_start, va->va_end);
438 * lazy_max_pages is the maximum amount of virtual address space we gather up
439 * before attempting to purge with a TLB flush.
441 * There is a tradeoff here: a larger number will cover more kernel page tables
442 * and take slightly longer to purge, but it will linearly reduce the number of
443 * global TLB flushes that must be performed. It would seem natural to scale
444 * this number up linearly with the number of CPUs (because vmapping activity
445 * could also scale linearly with the number of CPUs), however it is likely
446 * that in practice, workloads might be constrained in other ways that mean
447 * vmap activity will not scale linearly with CPUs. Also, I want to be
448 * conservative and not introduce a big latency on huge systems, so go with
449 * a less aggressive log scale. It will still be an improvement over the old
450 * code, and it will be simple to change the scale factor if we find that it
451 * becomes a problem on bigger systems.
453 static unsigned long lazy_max_pages(void)
457 log = fls(num_online_cpus());
459 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
462 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
465 * Purges all lazily-freed vmap areas.
467 * If sync is 0 then don't purge if there is already a purge in progress.
468 * If force_flush is 1, then flush kernel TLBs between *start and *end even
469 * if we found no lazy vmap areas to unmap (callers can use this to optimise
470 * their own TLB flushing).
471 * Returns with *start = min(*start, lowest purged address)
472 * *end = max(*end, highest purged address)
474 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
475 int sync, int force_flush)
477 static DEFINE_MUTEX(purge_lock);
479 struct vmap_area *va;
483 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
484 * should not expect such behaviour. This just simplifies locking for
485 * the case that isn't actually used at the moment anyway.
487 if (!sync && !force_flush) {
488 if (!mutex_trylock(&purge_lock))
491 mutex_lock(&purge_lock);
494 list_for_each_entry_rcu(va, &vmap_area_list, list) {
495 if (va->flags & VM_LAZY_FREE) {
496 if (va->va_start < *start)
497 *start = va->va_start;
498 if (va->va_end > *end)
500 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
502 list_add_tail(&va->purge_list, &valist);
503 va->flags |= VM_LAZY_FREEING;
504 va->flags &= ~VM_LAZY_FREE;
510 BUG_ON(nr > atomic_read(&vmap_lazy_nr));
511 atomic_sub(nr, &vmap_lazy_nr);
514 if (nr || force_flush)
515 flush_tlb_kernel_range(*start, *end);
518 spin_lock(&vmap_area_lock);
519 list_for_each_entry(va, &valist, purge_list)
520 __free_vmap_area(va);
521 spin_unlock(&vmap_area_lock);
523 mutex_unlock(&purge_lock);
527 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
528 * is already purging.
530 static void try_purge_vmap_area_lazy(void)
532 unsigned long start = ULONG_MAX, end = 0;
534 __purge_vmap_area_lazy(&start, &end, 0, 0);
538 * Kick off a purge of the outstanding lazy areas.
540 static void purge_vmap_area_lazy(void)
542 unsigned long start = ULONG_MAX, end = 0;
544 __purge_vmap_area_lazy(&start, &end, 1, 0);
548 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
549 * called for the correct range previously.
551 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
553 va->flags |= VM_LAZY_FREE;
554 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
555 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
556 try_purge_vmap_area_lazy();
560 * Free and unmap a vmap area
562 static void free_unmap_vmap_area(struct vmap_area *va)
564 flush_cache_vunmap(va->va_start, va->va_end);
565 free_unmap_vmap_area_noflush(va);
568 static struct vmap_area *find_vmap_area(unsigned long addr)
570 struct vmap_area *va;
572 spin_lock(&vmap_area_lock);
573 va = __find_vmap_area(addr);
574 spin_unlock(&vmap_area_lock);
579 static void free_unmap_vmap_area_addr(unsigned long addr)
581 struct vmap_area *va;
583 va = find_vmap_area(addr);
585 free_unmap_vmap_area(va);
589 /*** Per cpu kva allocator ***/
592 * vmap space is limited especially on 32 bit architectures. Ensure there is
593 * room for at least 16 percpu vmap blocks per CPU.
596 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
597 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
598 * instead (we just need a rough idea)
600 #if BITS_PER_LONG == 32
601 #define VMALLOC_SPACE (128UL*1024*1024)
603 #define VMALLOC_SPACE (128UL*1024*1024*1024)
606 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
607 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
608 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
609 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
610 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
611 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
612 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
613 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
614 VMALLOC_PAGES / NR_CPUS / 16))
616 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
618 static bool vmap_initialized __read_mostly = false;
620 struct vmap_block_queue {
622 struct list_head free;
623 struct list_head dirty;
624 unsigned int nr_dirty;
629 struct vmap_area *va;
630 struct vmap_block_queue *vbq;
631 unsigned long free, dirty;
632 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
633 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
636 struct list_head free_list;
637 struct list_head dirty_list;
639 struct rcu_head rcu_head;
643 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
644 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
647 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
648 * in the free path. Could get rid of this if we change the API to return a
649 * "cookie" from alloc, to be passed to free. But no big deal yet.
651 static DEFINE_SPINLOCK(vmap_block_tree_lock);
652 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
655 * We should probably have a fallback mechanism to allocate virtual memory
656 * out of partially filled vmap blocks. However vmap block sizing should be
657 * fairly reasonable according to the vmalloc size, so it shouldn't be a
661 static unsigned long addr_to_vb_idx(unsigned long addr)
663 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
664 addr /= VMAP_BLOCK_SIZE;
668 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
670 struct vmap_block_queue *vbq;
671 struct vmap_block *vb;
672 struct vmap_area *va;
673 unsigned long vb_idx;
676 node = numa_node_id();
678 vb = kmalloc_node(sizeof(struct vmap_block),
679 gfp_mask & GFP_RECLAIM_MASK, node);
681 return ERR_PTR(-ENOMEM);
683 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
684 VMALLOC_START, VMALLOC_END,
686 if (unlikely(IS_ERR(va))) {
688 return ERR_PTR(PTR_ERR(va));
691 err = radix_tree_preload(gfp_mask);
698 spin_lock_init(&vb->lock);
700 vb->free = VMAP_BBMAP_BITS;
702 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
703 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
704 INIT_LIST_HEAD(&vb->free_list);
705 INIT_LIST_HEAD(&vb->dirty_list);
707 vb_idx = addr_to_vb_idx(va->va_start);
708 spin_lock(&vmap_block_tree_lock);
709 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
710 spin_unlock(&vmap_block_tree_lock);
712 radix_tree_preload_end();
714 vbq = &get_cpu_var(vmap_block_queue);
716 spin_lock(&vbq->lock);
717 list_add(&vb->free_list, &vbq->free);
718 spin_unlock(&vbq->lock);
719 put_cpu_var(vmap_cpu_blocks);
724 static void rcu_free_vb(struct rcu_head *head)
726 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
731 static void free_vmap_block(struct vmap_block *vb)
733 struct vmap_block *tmp;
734 unsigned long vb_idx;
736 spin_lock(&vb->vbq->lock);
737 if (!list_empty(&vb->free_list))
738 list_del(&vb->free_list);
739 if (!list_empty(&vb->dirty_list))
740 list_del(&vb->dirty_list);
741 spin_unlock(&vb->vbq->lock);
743 vb_idx = addr_to_vb_idx(vb->va->va_start);
744 spin_lock(&vmap_block_tree_lock);
745 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
746 spin_unlock(&vmap_block_tree_lock);
749 free_unmap_vmap_area_noflush(vb->va);
750 call_rcu(&vb->rcu_head, rcu_free_vb);
753 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
755 struct vmap_block_queue *vbq;
756 struct vmap_block *vb;
757 unsigned long addr = 0;
760 BUG_ON(size & ~PAGE_MASK);
761 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
762 order = get_order(size);
766 vbq = &get_cpu_var(vmap_block_queue);
767 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
770 spin_lock(&vb->lock);
771 i = bitmap_find_free_region(vb->alloc_map,
772 VMAP_BBMAP_BITS, order);
775 addr = vb->va->va_start + (i << PAGE_SHIFT);
776 BUG_ON(addr_to_vb_idx(addr) !=
777 addr_to_vb_idx(vb->va->va_start));
778 vb->free -= 1UL << order;
780 spin_lock(&vbq->lock);
781 list_del_init(&vb->free_list);
782 spin_unlock(&vbq->lock);
784 spin_unlock(&vb->lock);
787 spin_unlock(&vb->lock);
789 put_cpu_var(vmap_cpu_blocks);
793 vb = new_vmap_block(gfp_mask);
802 static void vb_free(const void *addr, unsigned long size)
804 unsigned long offset;
805 unsigned long vb_idx;
807 struct vmap_block *vb;
809 BUG_ON(size & ~PAGE_MASK);
810 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
812 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
814 order = get_order(size);
816 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
818 vb_idx = addr_to_vb_idx((unsigned long)addr);
820 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
824 spin_lock(&vb->lock);
825 bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order);
827 spin_lock(&vb->vbq->lock);
828 list_add(&vb->dirty_list, &vb->vbq->dirty);
829 spin_unlock(&vb->vbq->lock);
831 vb->dirty += 1UL << order;
832 if (vb->dirty == VMAP_BBMAP_BITS) {
833 BUG_ON(vb->free || !list_empty(&vb->free_list));
834 spin_unlock(&vb->lock);
837 spin_unlock(&vb->lock);
841 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
843 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
844 * to amortize TLB flushing overheads. What this means is that any page you
845 * have now, may, in a former life, have been mapped into kernel virtual
846 * address by the vmap layer and so there might be some CPUs with TLB entries
847 * still referencing that page (additional to the regular 1:1 kernel mapping).
849 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
850 * be sure that none of the pages we have control over will have any aliases
851 * from the vmap layer.
853 void vm_unmap_aliases(void)
855 unsigned long start = ULONG_MAX, end = 0;
859 if (unlikely(!vmap_initialized))
862 for_each_possible_cpu(cpu) {
863 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
864 struct vmap_block *vb;
867 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
870 spin_lock(&vb->lock);
871 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
872 while (i < VMAP_BBMAP_BITS) {
875 j = find_next_zero_bit(vb->dirty_map,
878 s = vb->va->va_start + (i << PAGE_SHIFT);
879 e = vb->va->va_start + (j << PAGE_SHIFT);
880 vunmap_page_range(s, e);
889 i = find_next_bit(vb->dirty_map,
892 spin_unlock(&vb->lock);
897 __purge_vmap_area_lazy(&start, &end, 1, flush);
899 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
902 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
903 * @mem: the pointer returned by vm_map_ram
904 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
906 void vm_unmap_ram(const void *mem, unsigned int count)
908 unsigned long size = count << PAGE_SHIFT;
909 unsigned long addr = (unsigned long)mem;
912 BUG_ON(addr < VMALLOC_START);
913 BUG_ON(addr > VMALLOC_END);
914 BUG_ON(addr & (PAGE_SIZE-1));
916 debug_check_no_locks_freed(mem, size);
918 if (likely(count <= VMAP_MAX_ALLOC))
921 free_unmap_vmap_area_addr(addr);
923 EXPORT_SYMBOL(vm_unmap_ram);
926 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
927 * @pages: an array of pointers to the pages to be mapped
928 * @count: number of pages
929 * @node: prefer to allocate data structures on this node
930 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
932 * Returns: a pointer to the address that has been mapped, or %NULL on failure
934 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
936 unsigned long size = count << PAGE_SHIFT;
940 if (likely(count <= VMAP_MAX_ALLOC)) {
941 mem = vb_alloc(size, GFP_KERNEL);
944 addr = (unsigned long)mem;
946 struct vmap_area *va;
947 va = alloc_vmap_area(size, PAGE_SIZE,
948 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
955 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
956 vm_unmap_ram(mem, count);
961 EXPORT_SYMBOL(vm_map_ram);
963 void __init vmalloc_init(void)
967 for_each_possible_cpu(i) {
968 struct vmap_block_queue *vbq;
970 vbq = &per_cpu(vmap_block_queue, i);
971 spin_lock_init(&vbq->lock);
972 INIT_LIST_HEAD(&vbq->free);
973 INIT_LIST_HEAD(&vbq->dirty);
977 vmap_initialized = true;
980 void unmap_kernel_range(unsigned long addr, unsigned long size)
982 unsigned long end = addr + size;
983 vunmap_page_range(addr, end);
984 flush_tlb_kernel_range(addr, end);
987 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
989 unsigned long addr = (unsigned long)area->addr;
990 unsigned long end = addr + area->size - PAGE_SIZE;
993 err = vmap_page_range(addr, end, prot, *pages);
1001 EXPORT_SYMBOL_GPL(map_vm_area);
1003 /*** Old vmalloc interfaces ***/
1004 DEFINE_RWLOCK(vmlist_lock);
1005 struct vm_struct *vmlist;
1007 static struct vm_struct *__get_vm_area_node(unsigned long size,
1008 unsigned long flags, unsigned long start, unsigned long end,
1009 int node, gfp_t gfp_mask, void *caller)
1011 static struct vmap_area *va;
1012 struct vm_struct *area;
1013 struct vm_struct *tmp, **p;
1014 unsigned long align = 1;
1016 BUG_ON(in_interrupt());
1017 if (flags & VM_IOREMAP) {
1018 int bit = fls(size);
1020 if (bit > IOREMAP_MAX_ORDER)
1021 bit = IOREMAP_MAX_ORDER;
1022 else if (bit < PAGE_SHIFT)
1028 size = PAGE_ALIGN(size);
1029 if (unlikely(!size))
1032 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1033 if (unlikely(!area))
1037 * We always allocate a guard page.
1041 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1047 area->flags = flags;
1048 area->addr = (void *)va->va_start;
1052 area->phys_addr = 0;
1053 area->caller = caller;
1055 va->flags |= VM_VM_AREA;
1057 write_lock(&vmlist_lock);
1058 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1059 if (tmp->addr >= area->addr)
1064 write_unlock(&vmlist_lock);
1069 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1070 unsigned long start, unsigned long end)
1072 return __get_vm_area_node(size, flags, start, end, -1, GFP_KERNEL,
1073 __builtin_return_address(0));
1075 EXPORT_SYMBOL_GPL(__get_vm_area);
1078 * get_vm_area - reserve a contiguous kernel virtual area
1079 * @size: size of the area
1080 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1082 * Search an area of @size in the kernel virtual mapping area,
1083 * and reserved it for out purposes. Returns the area descriptor
1084 * on success or %NULL on failure.
1086 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1088 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1089 -1, GFP_KERNEL, __builtin_return_address(0));
1092 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1095 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END,
1096 -1, GFP_KERNEL, caller);
1099 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1100 int node, gfp_t gfp_mask)
1102 return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node,
1103 gfp_mask, __builtin_return_address(0));
1106 static struct vm_struct *find_vm_area(const void *addr)
1108 struct vmap_area *va;
1110 va = find_vmap_area((unsigned long)addr);
1111 if (va && va->flags & VM_VM_AREA)
1118 * remove_vm_area - find and remove a continuous kernel virtual area
1119 * @addr: base address
1121 * Search for the kernel VM area starting at @addr, and remove it.
1122 * This function returns the found VM area, but using it is NOT safe
1123 * on SMP machines, except for its size or flags.
1125 struct vm_struct *remove_vm_area(const void *addr)
1127 struct vmap_area *va;
1129 va = find_vmap_area((unsigned long)addr);
1130 if (va && va->flags & VM_VM_AREA) {
1131 struct vm_struct *vm = va->private;
1132 struct vm_struct *tmp, **p;
1133 free_unmap_vmap_area(va);
1134 vm->size -= PAGE_SIZE;
1136 write_lock(&vmlist_lock);
1137 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1140 write_unlock(&vmlist_lock);
1147 static void __vunmap(const void *addr, int deallocate_pages)
1149 struct vm_struct *area;
1154 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1155 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1159 area = remove_vm_area(addr);
1160 if (unlikely(!area)) {
1161 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1166 debug_check_no_locks_freed(addr, area->size);
1167 debug_check_no_obj_freed(addr, area->size);
1169 if (deallocate_pages) {
1172 for (i = 0; i < area->nr_pages; i++) {
1173 struct page *page = area->pages[i];
1179 if (area->flags & VM_VPAGES)
1190 * vfree - release memory allocated by vmalloc()
1191 * @addr: memory base address
1193 * Free the virtually continuous memory area starting at @addr, as
1194 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1195 * NULL, no operation is performed.
1197 * Must not be called in interrupt context.
1199 void vfree(const void *addr)
1201 BUG_ON(in_interrupt());
1204 EXPORT_SYMBOL(vfree);
1207 * vunmap - release virtual mapping obtained by vmap()
1208 * @addr: memory base address
1210 * Free the virtually contiguous memory area starting at @addr,
1211 * which was created from the page array passed to vmap().
1213 * Must not be called in interrupt context.
1215 void vunmap(const void *addr)
1217 BUG_ON(in_interrupt());
1220 EXPORT_SYMBOL(vunmap);
1223 * vmap - map an array of pages into virtually contiguous space
1224 * @pages: array of page pointers
1225 * @count: number of pages to map
1226 * @flags: vm_area->flags
1227 * @prot: page protection for the mapping
1229 * Maps @count pages from @pages into contiguous kernel virtual
1232 void *vmap(struct page **pages, unsigned int count,
1233 unsigned long flags, pgprot_t prot)
1235 struct vm_struct *area;
1237 if (count > num_physpages)
1240 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1241 __builtin_return_address(0));
1245 if (map_vm_area(area, prot, &pages)) {
1252 EXPORT_SYMBOL(vmap);
1254 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1255 int node, void *caller);
1256 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1257 pgprot_t prot, int node, void *caller)
1259 struct page **pages;
1260 unsigned int nr_pages, array_size, i;
1262 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1263 array_size = (nr_pages * sizeof(struct page *));
1265 area->nr_pages = nr_pages;
1266 /* Please note that the recursion is strictly bounded. */
1267 if (array_size > PAGE_SIZE) {
1268 pages = __vmalloc_node(array_size, gfp_mask | __GFP_ZERO,
1269 PAGE_KERNEL, node, caller);
1270 area->flags |= VM_VPAGES;
1272 pages = kmalloc_node(array_size,
1273 (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO,
1276 area->pages = pages;
1277 area->caller = caller;
1279 remove_vm_area(area->addr);
1284 for (i = 0; i < area->nr_pages; i++) {
1288 page = alloc_page(gfp_mask);
1290 page = alloc_pages_node(node, gfp_mask, 0);
1292 if (unlikely(!page)) {
1293 /* Successfully allocated i pages, free them in __vunmap() */
1297 area->pages[i] = page;
1300 if (map_vm_area(area, prot, &pages))
1309 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1311 return __vmalloc_area_node(area, gfp_mask, prot, -1,
1312 __builtin_return_address(0));
1316 * __vmalloc_node - allocate virtually contiguous memory
1317 * @size: allocation size
1318 * @gfp_mask: flags for the page level allocator
1319 * @prot: protection mask for the allocated pages
1320 * @node: node to use for allocation or -1
1321 * @caller: caller's return address
1323 * Allocate enough pages to cover @size from the page level
1324 * allocator with @gfp_mask flags. Map them into contiguous
1325 * kernel virtual space, using a pagetable protection of @prot.
1327 static void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
1328 int node, void *caller)
1330 struct vm_struct *area;
1332 size = PAGE_ALIGN(size);
1333 if (!size || (size >> PAGE_SHIFT) > num_physpages)
1336 area = __get_vm_area_node(size, VM_ALLOC, VMALLOC_START, VMALLOC_END,
1337 node, gfp_mask, caller);
1342 return __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1345 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1347 return __vmalloc_node(size, gfp_mask, prot, -1,
1348 __builtin_return_address(0));
1350 EXPORT_SYMBOL(__vmalloc);
1353 * vmalloc - allocate virtually contiguous memory
1354 * @size: allocation size
1355 * Allocate enough pages to cover @size from the page level
1356 * allocator and map them into contiguous kernel virtual space.
1358 * For tight control over page level allocator and protection flags
1359 * use __vmalloc() instead.
1361 void *vmalloc(unsigned long size)
1363 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1364 -1, __builtin_return_address(0));
1366 EXPORT_SYMBOL(vmalloc);
1369 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1370 * @size: allocation size
1372 * The resulting memory area is zeroed so it can be mapped to userspace
1373 * without leaking data.
1375 void *vmalloc_user(unsigned long size)
1377 struct vm_struct *area;
1380 ret = __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1381 PAGE_KERNEL, -1, __builtin_return_address(0));
1383 area = find_vm_area(ret);
1384 area->flags |= VM_USERMAP;
1388 EXPORT_SYMBOL(vmalloc_user);
1391 * vmalloc_node - allocate memory on a specific node
1392 * @size: allocation size
1395 * Allocate enough pages to cover @size from the page level
1396 * allocator and map them into contiguous kernel virtual space.
1398 * For tight control over page level allocator and protection flags
1399 * use __vmalloc() instead.
1401 void *vmalloc_node(unsigned long size, int node)
1403 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1404 node, __builtin_return_address(0));
1406 EXPORT_SYMBOL(vmalloc_node);
1408 #ifndef PAGE_KERNEL_EXEC
1409 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1413 * vmalloc_exec - allocate virtually contiguous, executable memory
1414 * @size: allocation size
1416 * Kernel-internal function to allocate enough pages to cover @size
1417 * the page level allocator and map them into contiguous and
1418 * executable kernel virtual space.
1420 * For tight control over page level allocator and protection flags
1421 * use __vmalloc() instead.
1424 void *vmalloc_exec(unsigned long size)
1426 return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1427 -1, __builtin_return_address(0));
1430 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1431 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1432 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1433 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1435 #define GFP_VMALLOC32 GFP_KERNEL
1439 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1440 * @size: allocation size
1442 * Allocate enough 32bit PA addressable pages to cover @size from the
1443 * page level allocator and map them into contiguous kernel virtual space.
1445 void *vmalloc_32(unsigned long size)
1447 return __vmalloc_node(size, GFP_VMALLOC32, PAGE_KERNEL,
1448 -1, __builtin_return_address(0));
1450 EXPORT_SYMBOL(vmalloc_32);
1453 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1454 * @size: allocation size
1456 * The resulting memory area is 32bit addressable and zeroed so it can be
1457 * mapped to userspace without leaking data.
1459 void *vmalloc_32_user(unsigned long size)
1461 struct vm_struct *area;
1464 ret = __vmalloc_node(size, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1465 -1, __builtin_return_address(0));
1467 area = find_vm_area(ret);
1468 area->flags |= VM_USERMAP;
1472 EXPORT_SYMBOL(vmalloc_32_user);
1474 long vread(char *buf, char *addr, unsigned long count)
1476 struct vm_struct *tmp;
1477 char *vaddr, *buf_start = buf;
1480 /* Don't allow overflow */
1481 if ((unsigned long) addr + count < count)
1482 count = -(unsigned long) addr;
1484 read_lock(&vmlist_lock);
1485 for (tmp = vmlist; tmp; tmp = tmp->next) {
1486 vaddr = (char *) tmp->addr;
1487 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1489 while (addr < vaddr) {
1497 n = vaddr + tmp->size - PAGE_SIZE - addr;
1508 read_unlock(&vmlist_lock);
1509 return buf - buf_start;
1512 long vwrite(char *buf, char *addr, unsigned long count)
1514 struct vm_struct *tmp;
1515 char *vaddr, *buf_start = buf;
1518 /* Don't allow overflow */
1519 if ((unsigned long) addr + count < count)
1520 count = -(unsigned long) addr;
1522 read_lock(&vmlist_lock);
1523 for (tmp = vmlist; tmp; tmp = tmp->next) {
1524 vaddr = (char *) tmp->addr;
1525 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1527 while (addr < vaddr) {
1534 n = vaddr + tmp->size - PAGE_SIZE - addr;
1545 read_unlock(&vmlist_lock);
1546 return buf - buf_start;
1550 * remap_vmalloc_range - map vmalloc pages to userspace
1551 * @vma: vma to cover (map full range of vma)
1552 * @addr: vmalloc memory
1553 * @pgoff: number of pages into addr before first page to map
1555 * Returns: 0 for success, -Exxx on failure
1557 * This function checks that addr is a valid vmalloc'ed area, and
1558 * that it is big enough to cover the vma. Will return failure if
1559 * that criteria isn't met.
1561 * Similar to remap_pfn_range() (see mm/memory.c)
1563 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1564 unsigned long pgoff)
1566 struct vm_struct *area;
1567 unsigned long uaddr = vma->vm_start;
1568 unsigned long usize = vma->vm_end - vma->vm_start;
1570 if ((PAGE_SIZE-1) & (unsigned long)addr)
1573 area = find_vm_area(addr);
1577 if (!(area->flags & VM_USERMAP))
1580 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1583 addr += pgoff << PAGE_SHIFT;
1585 struct page *page = vmalloc_to_page(addr);
1588 ret = vm_insert_page(vma, uaddr, page);
1595 } while (usize > 0);
1597 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1598 vma->vm_flags |= VM_RESERVED;
1602 EXPORT_SYMBOL(remap_vmalloc_range);
1605 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1608 void __attribute__((weak)) vmalloc_sync_all(void)
1613 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
1615 /* apply_to_page_range() does all the hard work. */
1620 * alloc_vm_area - allocate a range of kernel address space
1621 * @size: size of the area
1623 * Returns: NULL on failure, vm_struct on success
1625 * This function reserves a range of kernel address space, and
1626 * allocates pagetables to map that range. No actual mappings
1627 * are created. If the kernel address space is not shared
1628 * between processes, it syncs the pagetable across all
1631 struct vm_struct *alloc_vm_area(size_t size)
1633 struct vm_struct *area;
1635 area = get_vm_area_caller(size, VM_IOREMAP,
1636 __builtin_return_address(0));
1641 * This ensures that page tables are constructed for this region
1642 * of kernel virtual address space and mapped into init_mm.
1644 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
1645 area->size, f, NULL)) {
1650 /* Make sure the pagetables are constructed in process kernel
1656 EXPORT_SYMBOL_GPL(alloc_vm_area);
1658 void free_vm_area(struct vm_struct *area)
1660 struct vm_struct *ret;
1661 ret = remove_vm_area(area->addr);
1662 BUG_ON(ret != area);
1665 EXPORT_SYMBOL_GPL(free_vm_area);
1668 #ifdef CONFIG_PROC_FS
1669 static void *s_start(struct seq_file *m, loff_t *pos)
1672 struct vm_struct *v;
1674 read_lock(&vmlist_lock);
1676 while (n > 0 && v) {
1687 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
1689 struct vm_struct *v = p;
1695 static void s_stop(struct seq_file *m, void *p)
1697 read_unlock(&vmlist_lock);
1700 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
1703 unsigned int nr, *counters = m->private;
1708 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
1710 for (nr = 0; nr < v->nr_pages; nr++)
1711 counters[page_to_nid(v->pages[nr])]++;
1713 for_each_node_state(nr, N_HIGH_MEMORY)
1715 seq_printf(m, " N%u=%u", nr, counters[nr]);
1719 static int s_show(struct seq_file *m, void *p)
1721 struct vm_struct *v = p;
1723 seq_printf(m, "0x%p-0x%p %7ld",
1724 v->addr, v->addr + v->size, v->size);
1727 char buff[KSYM_SYMBOL_LEN];
1730 sprint_symbol(buff, (unsigned long)v->caller);
1735 seq_printf(m, " pages=%d", v->nr_pages);
1738 seq_printf(m, " phys=%lx", v->phys_addr);
1740 if (v->flags & VM_IOREMAP)
1741 seq_printf(m, " ioremap");
1743 if (v->flags & VM_ALLOC)
1744 seq_printf(m, " vmalloc");
1746 if (v->flags & VM_MAP)
1747 seq_printf(m, " vmap");
1749 if (v->flags & VM_USERMAP)
1750 seq_printf(m, " user");
1752 if (v->flags & VM_VPAGES)
1753 seq_printf(m, " vpages");
1755 show_numa_info(m, v);
1760 static const struct seq_operations vmalloc_op = {
1767 static int vmalloc_open(struct inode *inode, struct file *file)
1769 unsigned int *ptr = NULL;
1773 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
1774 ret = seq_open(file, &vmalloc_op);
1776 struct seq_file *m = file->private_data;
1783 static const struct file_operations proc_vmalloc_operations = {
1784 .open = vmalloc_open,
1786 .llseek = seq_lseek,
1787 .release = seq_release_private,
1790 static int __init proc_vmalloc_init(void)
1792 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
1795 module_init(proc_vmalloc_init);