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/sched.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.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>
27 #include <linux/pfn.h>
28 #include <linux/kmemleak.h>
29 #include <linux/atomic.h>
30 #include <asm/uaccess.h>
31 #include <asm/tlbflush.h>
32 #include <asm/shmparam.h>
34 /*** Page table manipulation functions ***/
36 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
40 pte = pte_offset_kernel(pmd, addr);
42 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
43 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
44 } while (pte++, addr += PAGE_SIZE, addr != end);
47 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
52 pmd = pmd_offset(pud, addr);
54 next = pmd_addr_end(addr, end);
55 if (pmd_none_or_clear_bad(pmd))
57 vunmap_pte_range(pmd, addr, next);
58 } while (pmd++, addr = next, addr != end);
61 static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
66 pud = pud_offset(pgd, addr);
68 next = pud_addr_end(addr, end);
69 if (pud_none_or_clear_bad(pud))
71 vunmap_pmd_range(pud, addr, next);
72 } while (pud++, addr = next, addr != end);
75 static void vunmap_page_range(unsigned long addr, unsigned long end)
81 pgd = pgd_offset_k(addr);
83 next = pgd_addr_end(addr, end);
84 if (pgd_none_or_clear_bad(pgd))
86 vunmap_pud_range(pgd, addr, next);
87 } while (pgd++, addr = next, addr != end);
90 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
91 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
96 * nr is a running index into the array which helps higher level
97 * callers keep track of where we're up to.
100 pte = pte_alloc_kernel(pmd, addr);
104 struct page *page = pages[*nr];
106 if (WARN_ON(!pte_none(*pte)))
110 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
112 } while (pte++, addr += PAGE_SIZE, addr != end);
116 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
117 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
122 pmd = pmd_alloc(&init_mm, pud, addr);
126 next = pmd_addr_end(addr, end);
127 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
129 } while (pmd++, addr = next, addr != end);
133 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
134 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
139 pud = pud_alloc(&init_mm, pgd, addr);
143 next = pud_addr_end(addr, end);
144 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
146 } while (pud++, addr = next, addr != end);
151 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
152 * will have pfns corresponding to the "pages" array.
154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
156 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
157 pgprot_t prot, struct page **pages)
161 unsigned long addr = start;
166 pgd = pgd_offset_k(addr);
168 next = pgd_addr_end(addr, end);
169 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
172 } while (pgd++, addr = next, addr != end);
177 static int vmap_page_range(unsigned long start, unsigned long end,
178 pgprot_t prot, struct page **pages)
182 ret = vmap_page_range_noflush(start, end, prot, pages);
183 flush_cache_vmap(start, end);
187 int is_vmalloc_or_module_addr(const void *x)
190 * ARM, x86-64 and sparc64 put modules in a special place,
191 * and fall back on vmalloc() if that fails. Others
192 * just put it in the vmalloc space.
194 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
195 unsigned long addr = (unsigned long)x;
196 if (addr >= MODULES_VADDR && addr < MODULES_END)
199 return is_vmalloc_addr(x);
203 * Walk a vmap address to the struct page it maps.
205 struct page *vmalloc_to_page(const void *vmalloc_addr)
207 unsigned long addr = (unsigned long) vmalloc_addr;
208 struct page *page = NULL;
209 pgd_t *pgd = pgd_offset_k(addr);
212 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
213 * architectures that do not vmalloc module space
215 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
217 if (!pgd_none(*pgd)) {
218 pud_t *pud = pud_offset(pgd, addr);
219 if (!pud_none(*pud)) {
220 pmd_t *pmd = pmd_offset(pud, addr);
221 if (!pmd_none(*pmd)) {
224 ptep = pte_offset_map(pmd, addr);
226 if (pte_present(pte))
227 page = pte_page(pte);
234 EXPORT_SYMBOL(vmalloc_to_page);
237 * Map a vmalloc()-space virtual address to the physical page frame number.
239 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
241 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
243 EXPORT_SYMBOL(vmalloc_to_pfn);
246 /*** Global kva allocator ***/
248 #define VM_LAZY_FREE 0x01
249 #define VM_LAZY_FREEING 0x02
250 #define VM_VM_AREA 0x04
253 unsigned long va_start;
254 unsigned long va_end;
256 struct rb_node rb_node; /* address sorted rbtree */
257 struct list_head list; /* address sorted list */
258 struct list_head purge_list; /* "lazy purge" list */
259 struct vm_struct *vm;
260 struct rcu_head rcu_head;
263 static DEFINE_SPINLOCK(vmap_area_lock);
264 static LIST_HEAD(vmap_area_list);
265 static struct rb_root vmap_area_root = RB_ROOT;
267 /* The vmap cache globals are protected by vmap_area_lock */
268 static struct rb_node *free_vmap_cache;
269 static unsigned long cached_hole_size;
270 static unsigned long cached_vstart;
271 static unsigned long cached_align;
273 static unsigned long vmap_area_pcpu_hole;
275 static struct vmap_area *__find_vmap_area(unsigned long addr)
277 struct rb_node *n = vmap_area_root.rb_node;
280 struct vmap_area *va;
282 va = rb_entry(n, struct vmap_area, rb_node);
283 if (addr < va->va_start)
285 else if (addr > va->va_start)
294 static void __insert_vmap_area(struct vmap_area *va)
296 struct rb_node **p = &vmap_area_root.rb_node;
297 struct rb_node *parent = NULL;
301 struct vmap_area *tmp_va;
304 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
305 if (va->va_start < tmp_va->va_end)
307 else if (va->va_end > tmp_va->va_start)
313 rb_link_node(&va->rb_node, parent, p);
314 rb_insert_color(&va->rb_node, &vmap_area_root);
316 /* address-sort this list so it is usable like the vmlist */
317 tmp = rb_prev(&va->rb_node);
319 struct vmap_area *prev;
320 prev = rb_entry(tmp, struct vmap_area, rb_node);
321 list_add_rcu(&va->list, &prev->list);
323 list_add_rcu(&va->list, &vmap_area_list);
326 static void purge_vmap_area_lazy(void);
329 * Allocate a region of KVA of the specified size and alignment, within the
332 static struct vmap_area *alloc_vmap_area(unsigned long size,
334 unsigned long vstart, unsigned long vend,
335 int node, gfp_t gfp_mask)
337 struct vmap_area *va;
341 struct vmap_area *first;
344 BUG_ON(size & ~PAGE_MASK);
345 BUG_ON(!is_power_of_2(align));
347 va = kmalloc_node(sizeof(struct vmap_area),
348 gfp_mask & GFP_RECLAIM_MASK, node);
350 return ERR_PTR(-ENOMEM);
353 * Only scan the relevant parts containing pointers to other objects
354 * to avoid false negatives.
356 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
359 spin_lock(&vmap_area_lock);
361 * Invalidate cache if we have more permissive parameters.
362 * cached_hole_size notes the largest hole noticed _below_
363 * the vmap_area cached in free_vmap_cache: if size fits
364 * into that hole, we want to scan from vstart to reuse
365 * the hole instead of allocating above free_vmap_cache.
366 * Note that __free_vmap_area may update free_vmap_cache
367 * without updating cached_hole_size or cached_align.
369 if (!free_vmap_cache ||
370 size < cached_hole_size ||
371 vstart < cached_vstart ||
372 align < cached_align) {
374 cached_hole_size = 0;
375 free_vmap_cache = NULL;
377 /* record if we encounter less permissive parameters */
378 cached_vstart = vstart;
379 cached_align = align;
381 /* find starting point for our search */
382 if (free_vmap_cache) {
383 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
384 addr = ALIGN(first->va_end, align);
387 if (addr + size - 1 < addr)
391 addr = ALIGN(vstart, align);
392 if (addr + size - 1 < addr)
395 n = vmap_area_root.rb_node;
399 struct vmap_area *tmp;
400 tmp = rb_entry(n, struct vmap_area, rb_node);
401 if (tmp->va_end >= addr) {
403 if (tmp->va_start <= addr)
414 /* from the starting point, walk areas until a suitable hole is found */
415 while (addr + size > first->va_start && addr + size <= vend) {
416 if (addr + cached_hole_size < first->va_start)
417 cached_hole_size = first->va_start - addr;
418 addr = ALIGN(first->va_end, align);
419 if (addr + size - 1 < addr)
422 n = rb_next(&first->rb_node);
424 first = rb_entry(n, struct vmap_area, rb_node);
430 if (addr + size > vend)
434 va->va_end = addr + size;
436 __insert_vmap_area(va);
437 free_vmap_cache = &va->rb_node;
438 spin_unlock(&vmap_area_lock);
440 BUG_ON(va->va_start & (align-1));
441 BUG_ON(va->va_start < vstart);
442 BUG_ON(va->va_end > vend);
447 spin_unlock(&vmap_area_lock);
449 purge_vmap_area_lazy();
453 if (printk_ratelimit())
455 "vmap allocation for size %lu failed: "
456 "use vmalloc=<size> to increase size.\n", size);
458 return ERR_PTR(-EBUSY);
461 static void __free_vmap_area(struct vmap_area *va)
463 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
465 if (free_vmap_cache) {
466 if (va->va_end < cached_vstart) {
467 free_vmap_cache = NULL;
469 struct vmap_area *cache;
470 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
471 if (va->va_start <= cache->va_start) {
472 free_vmap_cache = rb_prev(&va->rb_node);
474 * We don't try to update cached_hole_size or
475 * cached_align, but it won't go very wrong.
480 rb_erase(&va->rb_node, &vmap_area_root);
481 RB_CLEAR_NODE(&va->rb_node);
482 list_del_rcu(&va->list);
485 * Track the highest possible candidate for pcpu area
486 * allocation. Areas outside of vmalloc area can be returned
487 * here too, consider only end addresses which fall inside
488 * vmalloc area proper.
490 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
491 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
493 kfree_rcu(va, rcu_head);
497 * Free a region of KVA allocated by alloc_vmap_area
499 static void free_vmap_area(struct vmap_area *va)
501 spin_lock(&vmap_area_lock);
502 __free_vmap_area(va);
503 spin_unlock(&vmap_area_lock);
507 * Clear the pagetable entries of a given vmap_area
509 static void unmap_vmap_area(struct vmap_area *va)
511 vunmap_page_range(va->va_start, va->va_end);
514 static void vmap_debug_free_range(unsigned long start, unsigned long end)
517 * Unmap page tables and force a TLB flush immediately if
518 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
519 * bugs similarly to those in linear kernel virtual address
520 * space after a page has been freed.
522 * All the lazy freeing logic is still retained, in order to
523 * minimise intrusiveness of this debugging feature.
525 * This is going to be *slow* (linear kernel virtual address
526 * debugging doesn't do a broadcast TLB flush so it is a lot
529 #ifdef CONFIG_DEBUG_PAGEALLOC
530 vunmap_page_range(start, end);
531 flush_tlb_kernel_range(start, end);
536 * lazy_max_pages is the maximum amount of virtual address space we gather up
537 * before attempting to purge with a TLB flush.
539 * There is a tradeoff here: a larger number will cover more kernel page tables
540 * and take slightly longer to purge, but it will linearly reduce the number of
541 * global TLB flushes that must be performed. It would seem natural to scale
542 * this number up linearly with the number of CPUs (because vmapping activity
543 * could also scale linearly with the number of CPUs), however it is likely
544 * that in practice, workloads might be constrained in other ways that mean
545 * vmap activity will not scale linearly with CPUs. Also, I want to be
546 * conservative and not introduce a big latency on huge systems, so go with
547 * a less aggressive log scale. It will still be an improvement over the old
548 * code, and it will be simple to change the scale factor if we find that it
549 * becomes a problem on bigger systems.
551 static unsigned long lazy_max_pages(void)
555 log = fls(num_online_cpus());
557 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
560 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
562 /* for per-CPU blocks */
563 static void purge_fragmented_blocks_allcpus(void);
566 * called before a call to iounmap() if the caller wants vm_area_struct's
569 void set_iounmap_nonlazy(void)
571 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
575 * Purges all lazily-freed vmap areas.
577 * If sync is 0 then don't purge if there is already a purge in progress.
578 * If force_flush is 1, then flush kernel TLBs between *start and *end even
579 * if we found no lazy vmap areas to unmap (callers can use this to optimise
580 * their own TLB flushing).
581 * Returns with *start = min(*start, lowest purged address)
582 * *end = max(*end, highest purged address)
584 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
585 int sync, int force_flush)
587 static DEFINE_SPINLOCK(purge_lock);
589 struct vmap_area *va;
590 struct vmap_area *n_va;
594 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
595 * should not expect such behaviour. This just simplifies locking for
596 * the case that isn't actually used at the moment anyway.
598 if (!sync && !force_flush) {
599 if (!spin_trylock(&purge_lock))
602 spin_lock(&purge_lock);
605 purge_fragmented_blocks_allcpus();
608 list_for_each_entry_rcu(va, &vmap_area_list, list) {
609 if (va->flags & VM_LAZY_FREE) {
610 if (va->va_start < *start)
611 *start = va->va_start;
612 if (va->va_end > *end)
614 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
615 list_add_tail(&va->purge_list, &valist);
616 va->flags |= VM_LAZY_FREEING;
617 va->flags &= ~VM_LAZY_FREE;
623 atomic_sub(nr, &vmap_lazy_nr);
625 if (nr || force_flush)
626 flush_tlb_kernel_range(*start, *end);
629 spin_lock(&vmap_area_lock);
630 list_for_each_entry_safe(va, n_va, &valist, purge_list)
631 __free_vmap_area(va);
632 spin_unlock(&vmap_area_lock);
634 spin_unlock(&purge_lock);
638 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
639 * is already purging.
641 static void try_purge_vmap_area_lazy(void)
643 unsigned long start = ULONG_MAX, end = 0;
645 __purge_vmap_area_lazy(&start, &end, 0, 0);
649 * Kick off a purge of the outstanding lazy areas.
651 static void purge_vmap_area_lazy(void)
653 unsigned long start = ULONG_MAX, end = 0;
655 __purge_vmap_area_lazy(&start, &end, 1, 0);
659 * Free a vmap area, caller ensuring that the area has been unmapped
660 * and flush_cache_vunmap had been called for the correct range
663 static void free_vmap_area_noflush(struct vmap_area *va)
665 va->flags |= VM_LAZY_FREE;
666 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
667 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
668 try_purge_vmap_area_lazy();
672 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
673 * called for the correct range previously.
675 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
678 free_vmap_area_noflush(va);
682 * Free and unmap a vmap area
684 static void free_unmap_vmap_area(struct vmap_area *va)
686 flush_cache_vunmap(va->va_start, va->va_end);
687 free_unmap_vmap_area_noflush(va);
690 static struct vmap_area *find_vmap_area(unsigned long addr)
692 struct vmap_area *va;
694 spin_lock(&vmap_area_lock);
695 va = __find_vmap_area(addr);
696 spin_unlock(&vmap_area_lock);
701 static void free_unmap_vmap_area_addr(unsigned long addr)
703 struct vmap_area *va;
705 va = find_vmap_area(addr);
707 free_unmap_vmap_area(va);
711 /*** Per cpu kva allocator ***/
714 * vmap space is limited especially on 32 bit architectures. Ensure there is
715 * room for at least 16 percpu vmap blocks per CPU.
718 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
719 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
720 * instead (we just need a rough idea)
722 #if BITS_PER_LONG == 32
723 #define VMALLOC_SPACE (128UL*1024*1024)
725 #define VMALLOC_SPACE (128UL*1024*1024*1024)
728 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
729 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
730 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
731 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
732 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
733 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
734 #define VMAP_BBMAP_BITS \
735 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
736 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
737 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
739 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
741 static bool vmap_initialized __read_mostly = false;
743 struct vmap_block_queue {
745 struct list_head free;
750 struct vmap_area *va;
751 struct vmap_block_queue *vbq;
752 unsigned long free, dirty;
753 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
754 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
755 struct list_head free_list;
756 struct rcu_head rcu_head;
757 struct list_head purge;
760 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
761 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
764 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
765 * in the free path. Could get rid of this if we change the API to return a
766 * "cookie" from alloc, to be passed to free. But no big deal yet.
768 static DEFINE_SPINLOCK(vmap_block_tree_lock);
769 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
772 * We should probably have a fallback mechanism to allocate virtual memory
773 * out of partially filled vmap blocks. However vmap block sizing should be
774 * fairly reasonable according to the vmalloc size, so it shouldn't be a
778 static unsigned long addr_to_vb_idx(unsigned long addr)
780 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
781 addr /= VMAP_BLOCK_SIZE;
785 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
787 struct vmap_block_queue *vbq;
788 struct vmap_block *vb;
789 struct vmap_area *va;
790 unsigned long vb_idx;
793 node = numa_node_id();
795 vb = kmalloc_node(sizeof(struct vmap_block),
796 gfp_mask & GFP_RECLAIM_MASK, node);
798 return ERR_PTR(-ENOMEM);
800 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
801 VMALLOC_START, VMALLOC_END,
808 err = radix_tree_preload(gfp_mask);
815 spin_lock_init(&vb->lock);
817 vb->free = VMAP_BBMAP_BITS;
819 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
820 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
821 INIT_LIST_HEAD(&vb->free_list);
823 vb_idx = addr_to_vb_idx(va->va_start);
824 spin_lock(&vmap_block_tree_lock);
825 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
826 spin_unlock(&vmap_block_tree_lock);
828 radix_tree_preload_end();
830 vbq = &get_cpu_var(vmap_block_queue);
832 spin_lock(&vbq->lock);
833 list_add_rcu(&vb->free_list, &vbq->free);
834 spin_unlock(&vbq->lock);
835 put_cpu_var(vmap_block_queue);
840 static void free_vmap_block(struct vmap_block *vb)
842 struct vmap_block *tmp;
843 unsigned long vb_idx;
845 vb_idx = addr_to_vb_idx(vb->va->va_start);
846 spin_lock(&vmap_block_tree_lock);
847 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
848 spin_unlock(&vmap_block_tree_lock);
851 free_vmap_area_noflush(vb->va);
852 kfree_rcu(vb, rcu_head);
855 static void purge_fragmented_blocks(int cpu)
858 struct vmap_block *vb;
859 struct vmap_block *n_vb;
860 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
863 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
865 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
868 spin_lock(&vb->lock);
869 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
870 vb->free = 0; /* prevent further allocs after releasing lock */
871 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
872 bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
873 bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
874 spin_lock(&vbq->lock);
875 list_del_rcu(&vb->free_list);
876 spin_unlock(&vbq->lock);
877 spin_unlock(&vb->lock);
878 list_add_tail(&vb->purge, &purge);
880 spin_unlock(&vb->lock);
884 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
885 list_del(&vb->purge);
890 static void purge_fragmented_blocks_thiscpu(void)
892 purge_fragmented_blocks(smp_processor_id());
895 static void purge_fragmented_blocks_allcpus(void)
899 for_each_possible_cpu(cpu)
900 purge_fragmented_blocks(cpu);
903 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
905 struct vmap_block_queue *vbq;
906 struct vmap_block *vb;
907 unsigned long addr = 0;
911 BUG_ON(size & ~PAGE_MASK);
912 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
913 order = get_order(size);
917 vbq = &get_cpu_var(vmap_block_queue);
918 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
921 spin_lock(&vb->lock);
922 if (vb->free < 1UL << order)
925 i = bitmap_find_free_region(vb->alloc_map,
926 VMAP_BBMAP_BITS, order);
929 if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
930 /* fragmented and no outstanding allocations */
931 BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
936 addr = vb->va->va_start + (i << PAGE_SHIFT);
937 BUG_ON(addr_to_vb_idx(addr) !=
938 addr_to_vb_idx(vb->va->va_start));
939 vb->free -= 1UL << order;
941 spin_lock(&vbq->lock);
942 list_del_rcu(&vb->free_list);
943 spin_unlock(&vbq->lock);
945 spin_unlock(&vb->lock);
948 spin_unlock(&vb->lock);
952 purge_fragmented_blocks_thiscpu();
954 put_cpu_var(vmap_block_queue);
958 vb = new_vmap_block(gfp_mask);
967 static void vb_free(const void *addr, unsigned long size)
969 unsigned long offset;
970 unsigned long vb_idx;
972 struct vmap_block *vb;
974 BUG_ON(size & ~PAGE_MASK);
975 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
977 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
979 order = get_order(size);
981 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
983 vb_idx = addr_to_vb_idx((unsigned long)addr);
985 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
989 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
991 spin_lock(&vb->lock);
992 BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
994 vb->dirty += 1UL << order;
995 if (vb->dirty == VMAP_BBMAP_BITS) {
997 spin_unlock(&vb->lock);
1000 spin_unlock(&vb->lock);
1004 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1006 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1007 * to amortize TLB flushing overheads. What this means is that any page you
1008 * have now, may, in a former life, have been mapped into kernel virtual
1009 * address by the vmap layer and so there might be some CPUs with TLB entries
1010 * still referencing that page (additional to the regular 1:1 kernel mapping).
1012 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1013 * be sure that none of the pages we have control over will have any aliases
1014 * from the vmap layer.
1016 void vm_unmap_aliases(void)
1018 unsigned long start = ULONG_MAX, end = 0;
1022 if (unlikely(!vmap_initialized))
1025 for_each_possible_cpu(cpu) {
1026 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1027 struct vmap_block *vb;
1030 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1033 spin_lock(&vb->lock);
1034 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1035 while (i < VMAP_BBMAP_BITS) {
1038 j = find_next_zero_bit(vb->dirty_map,
1039 VMAP_BBMAP_BITS, i);
1041 s = vb->va->va_start + (i << PAGE_SHIFT);
1042 e = vb->va->va_start + (j << PAGE_SHIFT);
1051 i = find_next_bit(vb->dirty_map,
1052 VMAP_BBMAP_BITS, i);
1054 spin_unlock(&vb->lock);
1059 __purge_vmap_area_lazy(&start, &end, 1, flush);
1061 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1064 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1065 * @mem: the pointer returned by vm_map_ram
1066 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1068 void vm_unmap_ram(const void *mem, unsigned int count)
1070 unsigned long size = count << PAGE_SHIFT;
1071 unsigned long addr = (unsigned long)mem;
1074 BUG_ON(addr < VMALLOC_START);
1075 BUG_ON(addr > VMALLOC_END);
1076 BUG_ON(addr & (PAGE_SIZE-1));
1078 debug_check_no_locks_freed(mem, size);
1079 vmap_debug_free_range(addr, addr+size);
1081 if (likely(count <= VMAP_MAX_ALLOC))
1084 free_unmap_vmap_area_addr(addr);
1086 EXPORT_SYMBOL(vm_unmap_ram);
1089 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1090 * @pages: an array of pointers to the pages to be mapped
1091 * @count: number of pages
1092 * @node: prefer to allocate data structures on this node
1093 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1095 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1097 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1099 unsigned long size = count << PAGE_SHIFT;
1103 if (likely(count <= VMAP_MAX_ALLOC)) {
1104 mem = vb_alloc(size, GFP_KERNEL);
1107 addr = (unsigned long)mem;
1109 struct vmap_area *va;
1110 va = alloc_vmap_area(size, PAGE_SIZE,
1111 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1115 addr = va->va_start;
1118 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1119 vm_unmap_ram(mem, count);
1124 EXPORT_SYMBOL(vm_map_ram);
1127 * vm_area_register_early - register vmap area early during boot
1128 * @vm: vm_struct to register
1129 * @align: requested alignment
1131 * This function is used to register kernel vm area before
1132 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1133 * proper values on entry and other fields should be zero. On return,
1134 * vm->addr contains the allocated address.
1136 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1138 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1140 static size_t vm_init_off __initdata;
1143 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1144 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1146 vm->addr = (void *)addr;
1152 void __init vmalloc_init(void)
1154 struct vmap_area *va;
1155 struct vm_struct *tmp;
1158 for_each_possible_cpu(i) {
1159 struct vmap_block_queue *vbq;
1161 vbq = &per_cpu(vmap_block_queue, i);
1162 spin_lock_init(&vbq->lock);
1163 INIT_LIST_HEAD(&vbq->free);
1166 /* Import existing vmlist entries. */
1167 for (tmp = vmlist; tmp; tmp = tmp->next) {
1168 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1169 va->flags = VM_VM_AREA;
1170 va->va_start = (unsigned long)tmp->addr;
1171 va->va_end = va->va_start + tmp->size;
1173 __insert_vmap_area(va);
1176 vmap_area_pcpu_hole = VMALLOC_END;
1178 vmap_initialized = true;
1182 * map_kernel_range_noflush - map kernel VM area with the specified pages
1183 * @addr: start of the VM area to map
1184 * @size: size of the VM area to map
1185 * @prot: page protection flags to use
1186 * @pages: pages to map
1188 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1189 * specify should have been allocated using get_vm_area() and its
1193 * This function does NOT do any cache flushing. The caller is
1194 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1195 * before calling this function.
1198 * The number of pages mapped on success, -errno on failure.
1200 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1201 pgprot_t prot, struct page **pages)
1203 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1207 * unmap_kernel_range_noflush - unmap kernel VM area
1208 * @addr: start of the VM area to unmap
1209 * @size: size of the VM area to unmap
1211 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1212 * specify should have been allocated using get_vm_area() and its
1216 * This function does NOT do any cache flushing. The caller is
1217 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1218 * before calling this function and flush_tlb_kernel_range() after.
1220 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1222 vunmap_page_range(addr, addr + size);
1224 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1227 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1228 * @addr: start of the VM area to unmap
1229 * @size: size of the VM area to unmap
1231 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1232 * the unmapping and tlb after.
1234 void unmap_kernel_range(unsigned long addr, unsigned long size)
1236 unsigned long end = addr + size;
1238 flush_cache_vunmap(addr, end);
1239 vunmap_page_range(addr, end);
1240 flush_tlb_kernel_range(addr, end);
1243 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1245 unsigned long addr = (unsigned long)area->addr;
1246 unsigned long end = addr + area->size - PAGE_SIZE;
1249 err = vmap_page_range(addr, end, prot, *pages);
1257 EXPORT_SYMBOL_GPL(map_vm_area);
1259 /*** Old vmalloc interfaces ***/
1260 DEFINE_RWLOCK(vmlist_lock);
1261 struct vm_struct *vmlist;
1263 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1264 unsigned long flags, void *caller)
1267 vm->addr = (void *)va->va_start;
1268 vm->size = va->va_end - va->va_start;
1269 vm->caller = caller;
1271 va->flags |= VM_VM_AREA;
1274 static void insert_vmalloc_vmlist(struct vm_struct *vm)
1276 struct vm_struct *tmp, **p;
1278 vm->flags &= ~VM_UNLIST;
1279 write_lock(&vmlist_lock);
1280 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1281 if (tmp->addr >= vm->addr)
1286 write_unlock(&vmlist_lock);
1289 static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1290 unsigned long flags, void *caller)
1292 setup_vmalloc_vm(vm, va, flags, caller);
1293 insert_vmalloc_vmlist(vm);
1296 static struct vm_struct *__get_vm_area_node(unsigned long size,
1297 unsigned long align, unsigned long flags, unsigned long start,
1298 unsigned long end, int node, gfp_t gfp_mask, void *caller)
1300 struct vmap_area *va;
1301 struct vm_struct *area;
1303 BUG_ON(in_interrupt());
1304 if (flags & VM_IOREMAP) {
1305 int bit = fls(size);
1307 if (bit > IOREMAP_MAX_ORDER)
1308 bit = IOREMAP_MAX_ORDER;
1309 else if (bit < PAGE_SHIFT)
1315 size = PAGE_ALIGN(size);
1316 if (unlikely(!size))
1319 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1320 if (unlikely(!area))
1324 * We always allocate a guard page.
1328 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1335 * When this function is called from __vmalloc_node_range,
1336 * we do not add vm_struct to vmlist here to avoid
1337 * accessing uninitialized members of vm_struct such as
1338 * pages and nr_pages fields. They will be set later.
1339 * To distinguish it from others, we use a VM_UNLIST flag.
1341 if (flags & VM_UNLIST)
1342 setup_vmalloc_vm(area, va, flags, caller);
1344 insert_vmalloc_vm(area, va, flags, caller);
1349 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1350 unsigned long start, unsigned long end)
1352 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1353 __builtin_return_address(0));
1355 EXPORT_SYMBOL_GPL(__get_vm_area);
1357 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1358 unsigned long start, unsigned long end,
1361 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1366 * get_vm_area - reserve a contiguous kernel virtual area
1367 * @size: size of the area
1368 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1370 * Search an area of @size in the kernel virtual mapping area,
1371 * and reserved it for out purposes. Returns the area descriptor
1372 * on success or %NULL on failure.
1374 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1376 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1377 -1, GFP_KERNEL, __builtin_return_address(0));
1380 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1383 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1384 -1, GFP_KERNEL, caller);
1387 static struct vm_struct *find_vm_area(const void *addr)
1389 struct vmap_area *va;
1391 va = find_vmap_area((unsigned long)addr);
1392 if (va && va->flags & VM_VM_AREA)
1399 * remove_vm_area - find and remove a continuous kernel virtual area
1400 * @addr: base address
1402 * Search for the kernel VM area starting at @addr, and remove it.
1403 * This function returns the found VM area, but using it is NOT safe
1404 * on SMP machines, except for its size or flags.
1406 struct vm_struct *remove_vm_area(const void *addr)
1408 struct vmap_area *va;
1410 va = find_vmap_area((unsigned long)addr);
1411 if (va && va->flags & VM_VM_AREA) {
1412 struct vm_struct *vm = va->vm;
1414 if (!(vm->flags & VM_UNLIST)) {
1415 struct vm_struct *tmp, **p;
1417 * remove from list and disallow access to
1418 * this vm_struct before unmap. (address range
1419 * confliction is maintained by vmap.)
1421 write_lock(&vmlist_lock);
1422 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1425 write_unlock(&vmlist_lock);
1428 vmap_debug_free_range(va->va_start, va->va_end);
1429 free_unmap_vmap_area(va);
1430 vm->size -= PAGE_SIZE;
1437 static void __vunmap(const void *addr, int deallocate_pages)
1439 struct vm_struct *area;
1444 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1445 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1449 area = remove_vm_area(addr);
1450 if (unlikely(!area)) {
1451 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1456 debug_check_no_locks_freed(addr, area->size);
1457 debug_check_no_obj_freed(addr, area->size);
1459 if (deallocate_pages) {
1462 for (i = 0; i < area->nr_pages; i++) {
1463 struct page *page = area->pages[i];
1469 if (area->flags & VM_VPAGES)
1480 * vfree - release memory allocated by vmalloc()
1481 * @addr: memory base address
1483 * Free the virtually continuous memory area starting at @addr, as
1484 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1485 * NULL, no operation is performed.
1487 * Must not be called in interrupt context.
1489 void vfree(const void *addr)
1491 BUG_ON(in_interrupt());
1493 kmemleak_free(addr);
1497 EXPORT_SYMBOL(vfree);
1500 * vunmap - release virtual mapping obtained by vmap()
1501 * @addr: memory base address
1503 * Free the virtually contiguous memory area starting at @addr,
1504 * which was created from the page array passed to vmap().
1506 * Must not be called in interrupt context.
1508 void vunmap(const void *addr)
1510 BUG_ON(in_interrupt());
1514 EXPORT_SYMBOL(vunmap);
1517 * vmap - map an array of pages into virtually contiguous space
1518 * @pages: array of page pointers
1519 * @count: number of pages to map
1520 * @flags: vm_area->flags
1521 * @prot: page protection for the mapping
1523 * Maps @count pages from @pages into contiguous kernel virtual
1526 void *vmap(struct page **pages, unsigned int count,
1527 unsigned long flags, pgprot_t prot)
1529 struct vm_struct *area;
1533 if (count > totalram_pages)
1536 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1537 __builtin_return_address(0));
1541 if (map_vm_area(area, prot, &pages)) {
1548 EXPORT_SYMBOL(vmap);
1550 static void *__vmalloc_node(unsigned long size, unsigned long align,
1551 gfp_t gfp_mask, pgprot_t prot,
1552 int node, void *caller);
1553 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1554 pgprot_t prot, int node, void *caller)
1556 const int order = 0;
1557 struct page **pages;
1558 unsigned int nr_pages, array_size, i;
1559 gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1561 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1562 array_size = (nr_pages * sizeof(struct page *));
1564 area->nr_pages = nr_pages;
1565 /* Please note that the recursion is strictly bounded. */
1566 if (array_size > PAGE_SIZE) {
1567 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1568 PAGE_KERNEL, node, caller);
1569 area->flags |= VM_VPAGES;
1571 pages = kmalloc_node(array_size, nested_gfp, node);
1573 area->pages = pages;
1574 area->caller = caller;
1576 remove_vm_area(area->addr);
1581 for (i = 0; i < area->nr_pages; i++) {
1583 gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1586 page = alloc_page(tmp_mask);
1588 page = alloc_pages_node(node, tmp_mask, order);
1590 if (unlikely(!page)) {
1591 /* Successfully allocated i pages, free them in __vunmap() */
1595 area->pages[i] = page;
1598 if (map_vm_area(area, prot, &pages))
1603 warn_alloc_failed(gfp_mask, order,
1604 "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1605 (area->nr_pages*PAGE_SIZE), area->size);
1611 * __vmalloc_node_range - allocate virtually contiguous memory
1612 * @size: allocation size
1613 * @align: desired alignment
1614 * @start: vm area range start
1615 * @end: vm area range end
1616 * @gfp_mask: flags for the page level allocator
1617 * @prot: protection mask for the allocated pages
1618 * @node: node to use for allocation or -1
1619 * @caller: caller's return address
1621 * Allocate enough pages to cover @size from the page level
1622 * allocator with @gfp_mask flags. Map them into contiguous
1623 * kernel virtual space, using a pagetable protection of @prot.
1625 void *__vmalloc_node_range(unsigned long size, unsigned long align,
1626 unsigned long start, unsigned long end, gfp_t gfp_mask,
1627 pgprot_t prot, int node, void *caller)
1629 struct vm_struct *area;
1631 unsigned long real_size = size;
1633 size = PAGE_ALIGN(size);
1634 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1637 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1638 start, end, node, gfp_mask, caller);
1642 addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1647 * In this function, newly allocated vm_struct is not added
1648 * to vmlist at __get_vm_area_node(). so, it is added here.
1650 insert_vmalloc_vmlist(area);
1653 * A ref_count = 2 is needed because vm_struct allocated in
1654 * __get_vm_area_node() contains a reference to the virtual address of
1655 * the vmalloc'ed block.
1657 kmemleak_alloc(addr, real_size, 2, gfp_mask);
1662 warn_alloc_failed(gfp_mask, 0,
1663 "vmalloc: allocation failure: %lu bytes\n",
1669 * __vmalloc_node - allocate virtually contiguous memory
1670 * @size: allocation size
1671 * @align: desired alignment
1672 * @gfp_mask: flags for the page level allocator
1673 * @prot: protection mask for the allocated pages
1674 * @node: node to use for allocation or -1
1675 * @caller: caller's return address
1677 * Allocate enough pages to cover @size from the page level
1678 * allocator with @gfp_mask flags. Map them into contiguous
1679 * kernel virtual space, using a pagetable protection of @prot.
1681 static void *__vmalloc_node(unsigned long size, unsigned long align,
1682 gfp_t gfp_mask, pgprot_t prot,
1683 int node, void *caller)
1685 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1686 gfp_mask, prot, node, caller);
1689 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1691 return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1692 __builtin_return_address(0));
1694 EXPORT_SYMBOL(__vmalloc);
1696 static inline void *__vmalloc_node_flags(unsigned long size,
1697 int node, gfp_t flags)
1699 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1700 node, __builtin_return_address(0));
1704 * vmalloc - allocate virtually contiguous memory
1705 * @size: allocation size
1706 * Allocate enough pages to cover @size from the page level
1707 * allocator and map them into contiguous kernel virtual space.
1709 * For tight control over page level allocator and protection flags
1710 * use __vmalloc() instead.
1712 void *vmalloc(unsigned long size)
1714 return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1716 EXPORT_SYMBOL(vmalloc);
1719 * vzalloc - allocate virtually contiguous memory with zero fill
1720 * @size: allocation size
1721 * Allocate enough pages to cover @size from the page level
1722 * allocator and map them into contiguous kernel virtual space.
1723 * The memory allocated is set to zero.
1725 * For tight control over page level allocator and protection flags
1726 * use __vmalloc() instead.
1728 void *vzalloc(unsigned long size)
1730 return __vmalloc_node_flags(size, -1,
1731 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1733 EXPORT_SYMBOL(vzalloc);
1736 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1737 * @size: allocation size
1739 * The resulting memory area is zeroed so it can be mapped to userspace
1740 * without leaking data.
1742 void *vmalloc_user(unsigned long size)
1744 struct vm_struct *area;
1747 ret = __vmalloc_node(size, SHMLBA,
1748 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1749 PAGE_KERNEL, -1, __builtin_return_address(0));
1751 area = find_vm_area(ret);
1752 area->flags |= VM_USERMAP;
1756 EXPORT_SYMBOL(vmalloc_user);
1759 * vmalloc_node - allocate memory on a specific node
1760 * @size: allocation size
1763 * Allocate enough pages to cover @size from the page level
1764 * allocator and map them into contiguous kernel virtual space.
1766 * For tight control over page level allocator and protection flags
1767 * use __vmalloc() instead.
1769 void *vmalloc_node(unsigned long size, int node)
1771 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1772 node, __builtin_return_address(0));
1774 EXPORT_SYMBOL(vmalloc_node);
1777 * vzalloc_node - allocate memory on a specific node with zero fill
1778 * @size: allocation size
1781 * Allocate enough pages to cover @size from the page level
1782 * allocator and map them into contiguous kernel virtual space.
1783 * The memory allocated is set to zero.
1785 * For tight control over page level allocator and protection flags
1786 * use __vmalloc_node() instead.
1788 void *vzalloc_node(unsigned long size, int node)
1790 return __vmalloc_node_flags(size, node,
1791 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1793 EXPORT_SYMBOL(vzalloc_node);
1795 #ifndef PAGE_KERNEL_EXEC
1796 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1800 * vmalloc_exec - allocate virtually contiguous, executable memory
1801 * @size: allocation size
1803 * Kernel-internal function to allocate enough pages to cover @size
1804 * the page level allocator and map them into contiguous and
1805 * executable kernel virtual space.
1807 * For tight control over page level allocator and protection flags
1808 * use __vmalloc() instead.
1811 void *vmalloc_exec(unsigned long size)
1813 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1814 -1, __builtin_return_address(0));
1817 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1818 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1819 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1820 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1822 #define GFP_VMALLOC32 GFP_KERNEL
1826 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1827 * @size: allocation size
1829 * Allocate enough 32bit PA addressable pages to cover @size from the
1830 * page level allocator and map them into contiguous kernel virtual space.
1832 void *vmalloc_32(unsigned long size)
1834 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1835 -1, __builtin_return_address(0));
1837 EXPORT_SYMBOL(vmalloc_32);
1840 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1841 * @size: allocation size
1843 * The resulting memory area is 32bit addressable and zeroed so it can be
1844 * mapped to userspace without leaking data.
1846 void *vmalloc_32_user(unsigned long size)
1848 struct vm_struct *area;
1851 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1852 -1, __builtin_return_address(0));
1854 area = find_vm_area(ret);
1855 area->flags |= VM_USERMAP;
1859 EXPORT_SYMBOL(vmalloc_32_user);
1862 * small helper routine , copy contents to buf from addr.
1863 * If the page is not present, fill zero.
1866 static int aligned_vread(char *buf, char *addr, unsigned long count)
1872 unsigned long offset, length;
1874 offset = (unsigned long)addr & ~PAGE_MASK;
1875 length = PAGE_SIZE - offset;
1878 p = vmalloc_to_page(addr);
1880 * To do safe access to this _mapped_ area, we need
1881 * lock. But adding lock here means that we need to add
1882 * overhead of vmalloc()/vfree() calles for this _debug_
1883 * interface, rarely used. Instead of that, we'll use
1884 * kmap() and get small overhead in this access function.
1888 * we can expect USER0 is not used (see vread/vwrite's
1889 * function description)
1891 void *map = kmap_atomic(p, KM_USER0);
1892 memcpy(buf, map + offset, length);
1893 kunmap_atomic(map, KM_USER0);
1895 memset(buf, 0, length);
1905 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1911 unsigned long offset, length;
1913 offset = (unsigned long)addr & ~PAGE_MASK;
1914 length = PAGE_SIZE - offset;
1917 p = vmalloc_to_page(addr);
1919 * To do safe access to this _mapped_ area, we need
1920 * lock. But adding lock here means that we need to add
1921 * overhead of vmalloc()/vfree() calles for this _debug_
1922 * interface, rarely used. Instead of that, we'll use
1923 * kmap() and get small overhead in this access function.
1927 * we can expect USER0 is not used (see vread/vwrite's
1928 * function description)
1930 void *map = kmap_atomic(p, KM_USER0);
1931 memcpy(map + offset, buf, length);
1932 kunmap_atomic(map, KM_USER0);
1943 * vread() - read vmalloc area in a safe way.
1944 * @buf: buffer for reading data
1945 * @addr: vm address.
1946 * @count: number of bytes to be read.
1948 * Returns # of bytes which addr and buf should be increased.
1949 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1950 * includes any intersect with alive vmalloc area.
1952 * This function checks that addr is a valid vmalloc'ed area, and
1953 * copy data from that area to a given buffer. If the given memory range
1954 * of [addr...addr+count) includes some valid address, data is copied to
1955 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1956 * IOREMAP area is treated as memory hole and no copy is done.
1958 * If [addr...addr+count) doesn't includes any intersects with alive
1959 * vm_struct area, returns 0.
1960 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1961 * the caller should guarantee KM_USER0 is not used.
1963 * Note: In usual ops, vread() is never necessary because the caller
1964 * should know vmalloc() area is valid and can use memcpy().
1965 * This is for routines which have to access vmalloc area without
1966 * any informaion, as /dev/kmem.
1970 long vread(char *buf, char *addr, unsigned long count)
1972 struct vm_struct *tmp;
1973 char *vaddr, *buf_start = buf;
1974 unsigned long buflen = count;
1977 /* Don't allow overflow */
1978 if ((unsigned long) addr + count < count)
1979 count = -(unsigned long) addr;
1981 read_lock(&vmlist_lock);
1982 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1983 vaddr = (char *) tmp->addr;
1984 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1986 while (addr < vaddr) {
1994 n = vaddr + tmp->size - PAGE_SIZE - addr;
1997 if (!(tmp->flags & VM_IOREMAP))
1998 aligned_vread(buf, addr, n);
1999 else /* IOREMAP area is treated as memory hole */
2006 read_unlock(&vmlist_lock);
2008 if (buf == buf_start)
2010 /* zero-fill memory holes */
2011 if (buf != buf_start + buflen)
2012 memset(buf, 0, buflen - (buf - buf_start));
2018 * vwrite() - write vmalloc area in a safe way.
2019 * @buf: buffer for source data
2020 * @addr: vm address.
2021 * @count: number of bytes to be read.
2023 * Returns # of bytes which addr and buf should be incresed.
2024 * (same number to @count).
2025 * If [addr...addr+count) doesn't includes any intersect with valid
2026 * vmalloc area, returns 0.
2028 * This function checks that addr is a valid vmalloc'ed area, and
2029 * copy data from a buffer to the given addr. If specified range of
2030 * [addr...addr+count) includes some valid address, data is copied from
2031 * proper area of @buf. If there are memory holes, no copy to hole.
2032 * IOREMAP area is treated as memory hole and no copy is done.
2034 * If [addr...addr+count) doesn't includes any intersects with alive
2035 * vm_struct area, returns 0.
2036 * @buf should be kernel's buffer. Because this function uses KM_USER0,
2037 * the caller should guarantee KM_USER0 is not used.
2039 * Note: In usual ops, vwrite() is never necessary because the caller
2040 * should know vmalloc() area is valid and can use memcpy().
2041 * This is for routines which have to access vmalloc area without
2042 * any informaion, as /dev/kmem.
2045 long vwrite(char *buf, char *addr, unsigned long count)
2047 struct vm_struct *tmp;
2049 unsigned long n, buflen;
2052 /* Don't allow overflow */
2053 if ((unsigned long) addr + count < count)
2054 count = -(unsigned long) addr;
2057 read_lock(&vmlist_lock);
2058 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2059 vaddr = (char *) tmp->addr;
2060 if (addr >= vaddr + tmp->size - PAGE_SIZE)
2062 while (addr < vaddr) {
2069 n = vaddr + tmp->size - PAGE_SIZE - addr;
2072 if (!(tmp->flags & VM_IOREMAP)) {
2073 aligned_vwrite(buf, addr, n);
2081 read_unlock(&vmlist_lock);
2088 * remap_vmalloc_range - map vmalloc pages to userspace
2089 * @vma: vma to cover (map full range of vma)
2090 * @addr: vmalloc memory
2091 * @pgoff: number of pages into addr before first page to map
2093 * Returns: 0 for success, -Exxx on failure
2095 * This function checks that addr is a valid vmalloc'ed area, and
2096 * that it is big enough to cover the vma. Will return failure if
2097 * that criteria isn't met.
2099 * Similar to remap_pfn_range() (see mm/memory.c)
2101 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2102 unsigned long pgoff)
2104 struct vm_struct *area;
2105 unsigned long uaddr = vma->vm_start;
2106 unsigned long usize = vma->vm_end - vma->vm_start;
2108 if ((PAGE_SIZE-1) & (unsigned long)addr)
2111 area = find_vm_area(addr);
2115 if (!(area->flags & VM_USERMAP))
2118 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2121 addr += pgoff << PAGE_SHIFT;
2123 struct page *page = vmalloc_to_page(addr);
2126 ret = vm_insert_page(vma, uaddr, page);
2133 } while (usize > 0);
2135 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2136 vma->vm_flags |= VM_RESERVED;
2140 EXPORT_SYMBOL(remap_vmalloc_range);
2143 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2146 void __attribute__((weak)) vmalloc_sync_all(void)
2151 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2163 * alloc_vm_area - allocate a range of kernel address space
2164 * @size: size of the area
2165 * @ptes: returns the PTEs for the address space
2167 * Returns: NULL on failure, vm_struct on success
2169 * This function reserves a range of kernel address space, and
2170 * allocates pagetables to map that range. No actual mappings
2173 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2174 * allocated for the VM area are returned.
2176 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2178 struct vm_struct *area;
2180 area = get_vm_area_caller(size, VM_IOREMAP,
2181 __builtin_return_address(0));
2186 * This ensures that page tables are constructed for this region
2187 * of kernel virtual address space and mapped into init_mm.
2189 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2190 size, f, ptes ? &ptes : NULL)) {
2197 EXPORT_SYMBOL_GPL(alloc_vm_area);
2199 void free_vm_area(struct vm_struct *area)
2201 struct vm_struct *ret;
2202 ret = remove_vm_area(area->addr);
2203 BUG_ON(ret != area);
2206 EXPORT_SYMBOL_GPL(free_vm_area);
2209 static struct vmap_area *node_to_va(struct rb_node *n)
2211 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2215 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2216 * @end: target address
2217 * @pnext: out arg for the next vmap_area
2218 * @pprev: out arg for the previous vmap_area
2220 * Returns: %true if either or both of next and prev are found,
2221 * %false if no vmap_area exists
2223 * Find vmap_areas end addresses of which enclose @end. ie. if not
2224 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2226 static bool pvm_find_next_prev(unsigned long end,
2227 struct vmap_area **pnext,
2228 struct vmap_area **pprev)
2230 struct rb_node *n = vmap_area_root.rb_node;
2231 struct vmap_area *va = NULL;
2234 va = rb_entry(n, struct vmap_area, rb_node);
2235 if (end < va->va_end)
2237 else if (end > va->va_end)
2246 if (va->va_end > end) {
2248 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2251 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2257 * pvm_determine_end - find the highest aligned address between two vmap_areas
2258 * @pnext: in/out arg for the next vmap_area
2259 * @pprev: in/out arg for the previous vmap_area
2262 * Returns: determined end address
2264 * Find the highest aligned address between *@pnext and *@pprev below
2265 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2266 * down address is between the end addresses of the two vmap_areas.
2268 * Please note that the address returned by this function may fall
2269 * inside *@pnext vmap_area. The caller is responsible for checking
2272 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2273 struct vmap_area **pprev,
2274 unsigned long align)
2276 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2280 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2284 while (*pprev && (*pprev)->va_end > addr) {
2286 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2293 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2294 * @offsets: array containing offset of each area
2295 * @sizes: array containing size of each area
2296 * @nr_vms: the number of areas to allocate
2297 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2299 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2300 * vm_structs on success, %NULL on failure
2302 * Percpu allocator wants to use congruent vm areas so that it can
2303 * maintain the offsets among percpu areas. This function allocates
2304 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2305 * be scattered pretty far, distance between two areas easily going up
2306 * to gigabytes. To avoid interacting with regular vmallocs, these
2307 * areas are allocated from top.
2309 * Despite its complicated look, this allocator is rather simple. It
2310 * does everything top-down and scans areas from the end looking for
2311 * matching slot. While scanning, if any of the areas overlaps with
2312 * existing vmap_area, the base address is pulled down to fit the
2313 * area. Scanning is repeated till all the areas fit and then all
2314 * necessary data structres are inserted and the result is returned.
2316 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2317 const size_t *sizes, int nr_vms,
2320 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2321 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2322 struct vmap_area **vas, *prev, *next;
2323 struct vm_struct **vms;
2324 int area, area2, last_area, term_area;
2325 unsigned long base, start, end, last_end;
2326 bool purged = false;
2328 /* verify parameters and allocate data structures */
2329 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2330 for (last_area = 0, area = 0; area < nr_vms; area++) {
2331 start = offsets[area];
2332 end = start + sizes[area];
2334 /* is everything aligned properly? */
2335 BUG_ON(!IS_ALIGNED(offsets[area], align));
2336 BUG_ON(!IS_ALIGNED(sizes[area], align));
2338 /* detect the area with the highest address */
2339 if (start > offsets[last_area])
2342 for (area2 = 0; area2 < nr_vms; area2++) {
2343 unsigned long start2 = offsets[area2];
2344 unsigned long end2 = start2 + sizes[area2];
2349 BUG_ON(start2 >= start && start2 < end);
2350 BUG_ON(end2 <= end && end2 > start);
2353 last_end = offsets[last_area] + sizes[last_area];
2355 if (vmalloc_end - vmalloc_start < last_end) {
2360 vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
2361 vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
2365 for (area = 0; area < nr_vms; area++) {
2366 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2367 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2368 if (!vas[area] || !vms[area])
2372 spin_lock(&vmap_area_lock);
2374 /* start scanning - we scan from the top, begin with the last area */
2375 area = term_area = last_area;
2376 start = offsets[area];
2377 end = start + sizes[area];
2379 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2380 base = vmalloc_end - last_end;
2383 base = pvm_determine_end(&next, &prev, align) - end;
2386 BUG_ON(next && next->va_end <= base + end);
2387 BUG_ON(prev && prev->va_end > base + end);
2390 * base might have underflowed, add last_end before
2393 if (base + last_end < vmalloc_start + last_end) {
2394 spin_unlock(&vmap_area_lock);
2396 purge_vmap_area_lazy();
2404 * If next overlaps, move base downwards so that it's
2405 * right below next and then recheck.
2407 if (next && next->va_start < base + end) {
2408 base = pvm_determine_end(&next, &prev, align) - end;
2414 * If prev overlaps, shift down next and prev and move
2415 * base so that it's right below new next and then
2418 if (prev && prev->va_end > base + start) {
2420 prev = node_to_va(rb_prev(&next->rb_node));
2421 base = pvm_determine_end(&next, &prev, align) - end;
2427 * This area fits, move on to the previous one. If
2428 * the previous one is the terminal one, we're done.
2430 area = (area + nr_vms - 1) % nr_vms;
2431 if (area == term_area)
2433 start = offsets[area];
2434 end = start + sizes[area];
2435 pvm_find_next_prev(base + end, &next, &prev);
2438 /* we've found a fitting base, insert all va's */
2439 for (area = 0; area < nr_vms; area++) {
2440 struct vmap_area *va = vas[area];
2442 va->va_start = base + offsets[area];
2443 va->va_end = va->va_start + sizes[area];
2444 __insert_vmap_area(va);
2447 vmap_area_pcpu_hole = base + offsets[last_area];
2449 spin_unlock(&vmap_area_lock);
2451 /* insert all vm's */
2452 for (area = 0; area < nr_vms; area++)
2453 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2460 for (area = 0; area < nr_vms; area++) {
2472 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2473 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2474 * @nr_vms: the number of allocated areas
2476 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2478 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2482 for (i = 0; i < nr_vms; i++)
2483 free_vm_area(vms[i]);
2486 #endif /* CONFIG_SMP */
2488 #ifdef CONFIG_PROC_FS
2489 static void *s_start(struct seq_file *m, loff_t *pos)
2490 __acquires(&vmlist_lock)
2493 struct vm_struct *v;
2495 read_lock(&vmlist_lock);
2497 while (n > 0 && v) {
2508 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2510 struct vm_struct *v = p;
2516 static void s_stop(struct seq_file *m, void *p)
2517 __releases(&vmlist_lock)
2519 read_unlock(&vmlist_lock);
2522 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2525 unsigned int nr, *counters = m->private;
2530 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2532 for (nr = 0; nr < v->nr_pages; nr++)
2533 counters[page_to_nid(v->pages[nr])]++;
2535 for_each_node_state(nr, N_HIGH_MEMORY)
2537 seq_printf(m, " N%u=%u", nr, counters[nr]);
2541 static int s_show(struct seq_file *m, void *p)
2543 struct vm_struct *v = p;
2545 seq_printf(m, "0x%p-0x%p %7ld",
2546 v->addr, v->addr + v->size, v->size);
2549 seq_printf(m, " %pS", v->caller);
2552 seq_printf(m, " pages=%d", v->nr_pages);
2555 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2557 if (v->flags & VM_IOREMAP)
2558 seq_printf(m, " ioremap");
2560 if (v->flags & VM_ALLOC)
2561 seq_printf(m, " vmalloc");
2563 if (v->flags & VM_MAP)
2564 seq_printf(m, " vmap");
2566 if (v->flags & VM_USERMAP)
2567 seq_printf(m, " user");
2569 if (v->flags & VM_VPAGES)
2570 seq_printf(m, " vpages");
2572 show_numa_info(m, v);
2577 static const struct seq_operations vmalloc_op = {
2584 static int vmalloc_open(struct inode *inode, struct file *file)
2586 unsigned int *ptr = NULL;
2590 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2594 ret = seq_open(file, &vmalloc_op);
2596 struct seq_file *m = file->private_data;
2603 static const struct file_operations proc_vmalloc_operations = {
2604 .open = vmalloc_open,
2606 .llseek = seq_lseek,
2607 .release = seq_release_private,
2610 static int __init proc_vmalloc_init(void)
2612 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2615 module_init(proc_vmalloc_init);