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 <asm/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 */
260 struct rcu_head rcu_head;
263 static DEFINE_SPINLOCK(vmap_area_lock);
264 static struct rb_root vmap_area_root = RB_ROOT;
265 static LIST_HEAD(vmap_area_list);
266 static unsigned long vmap_area_pcpu_hole;
268 static struct vmap_area *__find_vmap_area(unsigned long addr)
270 struct rb_node *n = vmap_area_root.rb_node;
273 struct vmap_area *va;
275 va = rb_entry(n, struct vmap_area, rb_node);
276 if (addr < va->va_start)
278 else if (addr > va->va_start)
287 static void __insert_vmap_area(struct vmap_area *va)
289 struct rb_node **p = &vmap_area_root.rb_node;
290 struct rb_node *parent = NULL;
294 struct vmap_area *tmp_va;
297 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
298 if (va->va_start < tmp_va->va_end)
300 else if (va->va_end > tmp_va->va_start)
306 rb_link_node(&va->rb_node, parent, p);
307 rb_insert_color(&va->rb_node, &vmap_area_root);
309 /* address-sort this list so it is usable like the vmlist */
310 tmp = rb_prev(&va->rb_node);
312 struct vmap_area *prev;
313 prev = rb_entry(tmp, struct vmap_area, rb_node);
314 list_add_rcu(&va->list, &prev->list);
316 list_add_rcu(&va->list, &vmap_area_list);
319 static void purge_vmap_area_lazy(void);
322 * Allocate a region of KVA of the specified size and alignment, within the
325 static struct vmap_area *alloc_vmap_area(unsigned long size,
327 unsigned long vstart, unsigned long vend,
328 int node, gfp_t gfp_mask)
330 struct vmap_area *va;
336 BUG_ON(size & ~PAGE_MASK);
338 va = kmalloc_node(sizeof(struct vmap_area),
339 gfp_mask & GFP_RECLAIM_MASK, node);
341 return ERR_PTR(-ENOMEM);
344 addr = ALIGN(vstart, align);
346 spin_lock(&vmap_area_lock);
347 if (addr + size - 1 < addr)
350 /* XXX: could have a last_hole cache */
351 n = vmap_area_root.rb_node;
353 struct vmap_area *first = NULL;
356 struct vmap_area *tmp;
357 tmp = rb_entry(n, struct vmap_area, rb_node);
358 if (tmp->va_end >= addr) {
359 if (!first && tmp->va_start < addr + size)
371 if (first->va_end < addr) {
372 n = rb_next(&first->rb_node);
374 first = rb_entry(n, struct vmap_area, rb_node);
379 while (addr + size > first->va_start && addr + size <= vend) {
380 addr = ALIGN(first->va_end + PAGE_SIZE, align);
381 if (addr + size - 1 < addr)
384 n = rb_next(&first->rb_node);
386 first = rb_entry(n, struct vmap_area, rb_node);
392 if (addr + size > vend) {
394 spin_unlock(&vmap_area_lock);
396 purge_vmap_area_lazy();
400 if (printk_ratelimit())
402 "vmap allocation for size %lu failed: "
403 "use vmalloc=<size> to increase size.\n", size);
405 return ERR_PTR(-EBUSY);
408 BUG_ON(addr & (align-1));
411 va->va_end = addr + size;
413 __insert_vmap_area(va);
414 spin_unlock(&vmap_area_lock);
419 static void rcu_free_va(struct rcu_head *head)
421 struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
426 static void __free_vmap_area(struct vmap_area *va)
428 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
429 rb_erase(&va->rb_node, &vmap_area_root);
430 RB_CLEAR_NODE(&va->rb_node);
431 list_del_rcu(&va->list);
434 * Track the highest possible candidate for pcpu area
435 * allocation. Areas outside of vmalloc area can be returned
436 * here too, consider only end addresses which fall inside
437 * vmalloc area proper.
439 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
440 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
442 call_rcu(&va->rcu_head, rcu_free_va);
446 * Free a region of KVA allocated by alloc_vmap_area
448 static void free_vmap_area(struct vmap_area *va)
450 spin_lock(&vmap_area_lock);
451 __free_vmap_area(va);
452 spin_unlock(&vmap_area_lock);
456 * Clear the pagetable entries of a given vmap_area
458 static void unmap_vmap_area(struct vmap_area *va)
460 vunmap_page_range(va->va_start, va->va_end);
463 static void vmap_debug_free_range(unsigned long start, unsigned long end)
466 * Unmap page tables and force a TLB flush immediately if
467 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
468 * bugs similarly to those in linear kernel virtual address
469 * space after a page has been freed.
471 * All the lazy freeing logic is still retained, in order to
472 * minimise intrusiveness of this debugging feature.
474 * This is going to be *slow* (linear kernel virtual address
475 * debugging doesn't do a broadcast TLB flush so it is a lot
478 #ifdef CONFIG_DEBUG_PAGEALLOC
479 vunmap_page_range(start, end);
480 flush_tlb_kernel_range(start, end);
485 * lazy_max_pages is the maximum amount of virtual address space we gather up
486 * before attempting to purge with a TLB flush.
488 * There is a tradeoff here: a larger number will cover more kernel page tables
489 * and take slightly longer to purge, but it will linearly reduce the number of
490 * global TLB flushes that must be performed. It would seem natural to scale
491 * this number up linearly with the number of CPUs (because vmapping activity
492 * could also scale linearly with the number of CPUs), however it is likely
493 * that in practice, workloads might be constrained in other ways that mean
494 * vmap activity will not scale linearly with CPUs. Also, I want to be
495 * conservative and not introduce a big latency on huge systems, so go with
496 * a less aggressive log scale. It will still be an improvement over the old
497 * code, and it will be simple to change the scale factor if we find that it
498 * becomes a problem on bigger systems.
500 static unsigned long lazy_max_pages(void)
504 log = fls(num_online_cpus());
506 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
509 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
511 /* for per-CPU blocks */
512 static void purge_fragmented_blocks_allcpus(void);
515 * called before a call to iounmap() if the caller wants vm_area_struct's
518 void set_iounmap_nonlazy(void)
520 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
524 * Purges all lazily-freed vmap areas.
526 * If sync is 0 then don't purge if there is already a purge in progress.
527 * If force_flush is 1, then flush kernel TLBs between *start and *end even
528 * if we found no lazy vmap areas to unmap (callers can use this to optimise
529 * their own TLB flushing).
530 * Returns with *start = min(*start, lowest purged address)
531 * *end = max(*end, highest purged address)
533 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
534 int sync, int force_flush)
536 static DEFINE_SPINLOCK(purge_lock);
538 struct vmap_area *va;
539 struct vmap_area *n_va;
543 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
544 * should not expect such behaviour. This just simplifies locking for
545 * the case that isn't actually used at the moment anyway.
547 if (!sync && !force_flush) {
548 if (!spin_trylock(&purge_lock))
551 spin_lock(&purge_lock);
554 purge_fragmented_blocks_allcpus();
557 list_for_each_entry_rcu(va, &vmap_area_list, list) {
558 if (va->flags & VM_LAZY_FREE) {
559 if (va->va_start < *start)
560 *start = va->va_start;
561 if (va->va_end > *end)
563 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
564 list_add_tail(&va->purge_list, &valist);
565 va->flags |= VM_LAZY_FREEING;
566 va->flags &= ~VM_LAZY_FREE;
572 atomic_sub(nr, &vmap_lazy_nr);
574 if (nr || force_flush)
575 flush_tlb_kernel_range(*start, *end);
578 spin_lock(&vmap_area_lock);
579 list_for_each_entry_safe(va, n_va, &valist, purge_list)
580 __free_vmap_area(va);
581 spin_unlock(&vmap_area_lock);
583 spin_unlock(&purge_lock);
587 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
588 * is already purging.
590 static void try_purge_vmap_area_lazy(void)
592 unsigned long start = ULONG_MAX, end = 0;
594 __purge_vmap_area_lazy(&start, &end, 0, 0);
598 * Kick off a purge of the outstanding lazy areas.
600 static void purge_vmap_area_lazy(void)
602 unsigned long start = ULONG_MAX, end = 0;
604 __purge_vmap_area_lazy(&start, &end, 1, 0);
608 * Free a vmap area, caller ensuring that the area has been unmapped
609 * and flush_cache_vunmap had been called for the correct range
612 static void free_vmap_area_noflush(struct vmap_area *va)
614 va->flags |= VM_LAZY_FREE;
615 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
616 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
617 try_purge_vmap_area_lazy();
621 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
622 * called for the correct range previously.
624 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
627 free_vmap_area_noflush(va);
631 * Free and unmap a vmap area
633 static void free_unmap_vmap_area(struct vmap_area *va)
635 flush_cache_vunmap(va->va_start, va->va_end);
636 free_unmap_vmap_area_noflush(va);
639 static struct vmap_area *find_vmap_area(unsigned long addr)
641 struct vmap_area *va;
643 spin_lock(&vmap_area_lock);
644 va = __find_vmap_area(addr);
645 spin_unlock(&vmap_area_lock);
650 static void free_unmap_vmap_area_addr(unsigned long addr)
652 struct vmap_area *va;
654 va = find_vmap_area(addr);
656 free_unmap_vmap_area(va);
660 /*** Per cpu kva allocator ***/
663 * vmap space is limited especially on 32 bit architectures. Ensure there is
664 * room for at least 16 percpu vmap blocks per CPU.
667 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
668 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
669 * instead (we just need a rough idea)
671 #if BITS_PER_LONG == 32
672 #define VMALLOC_SPACE (128UL*1024*1024)
674 #define VMALLOC_SPACE (128UL*1024*1024*1024)
677 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
678 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
679 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
680 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
681 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
682 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
683 #define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
684 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
685 VMALLOC_PAGES / NR_CPUS / 16))
687 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
689 static bool vmap_initialized __read_mostly = false;
691 struct vmap_block_queue {
693 struct list_head free;
698 struct vmap_area *va;
699 struct vmap_block_queue *vbq;
700 unsigned long free, dirty;
701 DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
702 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
703 struct list_head free_list;
704 struct rcu_head rcu_head;
705 struct list_head purge;
708 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
709 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
712 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
713 * in the free path. Could get rid of this if we change the API to return a
714 * "cookie" from alloc, to be passed to free. But no big deal yet.
716 static DEFINE_SPINLOCK(vmap_block_tree_lock);
717 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
720 * We should probably have a fallback mechanism to allocate virtual memory
721 * out of partially filled vmap blocks. However vmap block sizing should be
722 * fairly reasonable according to the vmalloc size, so it shouldn't be a
726 static unsigned long addr_to_vb_idx(unsigned long addr)
728 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
729 addr /= VMAP_BLOCK_SIZE;
733 static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
735 struct vmap_block_queue *vbq;
736 struct vmap_block *vb;
737 struct vmap_area *va;
738 unsigned long vb_idx;
741 node = numa_node_id();
743 vb = kmalloc_node(sizeof(struct vmap_block),
744 gfp_mask & GFP_RECLAIM_MASK, node);
746 return ERR_PTR(-ENOMEM);
748 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
749 VMALLOC_START, VMALLOC_END,
751 if (unlikely(IS_ERR(va))) {
756 err = radix_tree_preload(gfp_mask);
763 spin_lock_init(&vb->lock);
765 vb->free = VMAP_BBMAP_BITS;
767 bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
768 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
769 INIT_LIST_HEAD(&vb->free_list);
771 vb_idx = addr_to_vb_idx(va->va_start);
772 spin_lock(&vmap_block_tree_lock);
773 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
774 spin_unlock(&vmap_block_tree_lock);
776 radix_tree_preload_end();
778 vbq = &get_cpu_var(vmap_block_queue);
780 spin_lock(&vbq->lock);
781 list_add_rcu(&vb->free_list, &vbq->free);
782 spin_unlock(&vbq->lock);
783 put_cpu_var(vmap_block_queue);
788 static void rcu_free_vb(struct rcu_head *head)
790 struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
795 static void free_vmap_block(struct vmap_block *vb)
797 struct vmap_block *tmp;
798 unsigned long vb_idx;
800 vb_idx = addr_to_vb_idx(vb->va->va_start);
801 spin_lock(&vmap_block_tree_lock);
802 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
803 spin_unlock(&vmap_block_tree_lock);
806 free_vmap_area_noflush(vb->va);
807 call_rcu(&vb->rcu_head, rcu_free_vb);
810 static void purge_fragmented_blocks(int cpu)
813 struct vmap_block *vb;
814 struct vmap_block *n_vb;
815 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
818 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
820 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
823 spin_lock(&vb->lock);
824 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
825 vb->free = 0; /* prevent further allocs after releasing lock */
826 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
827 bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
828 bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
829 spin_lock(&vbq->lock);
830 list_del_rcu(&vb->free_list);
831 spin_unlock(&vbq->lock);
832 spin_unlock(&vb->lock);
833 list_add_tail(&vb->purge, &purge);
835 spin_unlock(&vb->lock);
839 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
840 list_del(&vb->purge);
845 static void purge_fragmented_blocks_thiscpu(void)
847 purge_fragmented_blocks(smp_processor_id());
850 static void purge_fragmented_blocks_allcpus(void)
854 for_each_possible_cpu(cpu)
855 purge_fragmented_blocks(cpu);
858 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
860 struct vmap_block_queue *vbq;
861 struct vmap_block *vb;
862 unsigned long addr = 0;
866 BUG_ON(size & ~PAGE_MASK);
867 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
868 order = get_order(size);
872 vbq = &get_cpu_var(vmap_block_queue);
873 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
876 spin_lock(&vb->lock);
877 if (vb->free < 1UL << order)
880 i = bitmap_find_free_region(vb->alloc_map,
881 VMAP_BBMAP_BITS, order);
884 if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
885 /* fragmented and no outstanding allocations */
886 BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
891 addr = vb->va->va_start + (i << PAGE_SHIFT);
892 BUG_ON(addr_to_vb_idx(addr) !=
893 addr_to_vb_idx(vb->va->va_start));
894 vb->free -= 1UL << order;
896 spin_lock(&vbq->lock);
897 list_del_rcu(&vb->free_list);
898 spin_unlock(&vbq->lock);
900 spin_unlock(&vb->lock);
903 spin_unlock(&vb->lock);
907 purge_fragmented_blocks_thiscpu();
909 put_cpu_var(vmap_block_queue);
913 vb = new_vmap_block(gfp_mask);
922 static void vb_free(const void *addr, unsigned long size)
924 unsigned long offset;
925 unsigned long vb_idx;
927 struct vmap_block *vb;
929 BUG_ON(size & ~PAGE_MASK);
930 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
932 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
934 order = get_order(size);
936 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
938 vb_idx = addr_to_vb_idx((unsigned long)addr);
940 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
944 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
946 spin_lock(&vb->lock);
947 BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
949 vb->dirty += 1UL << order;
950 if (vb->dirty == VMAP_BBMAP_BITS) {
952 spin_unlock(&vb->lock);
955 spin_unlock(&vb->lock);
959 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
961 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
962 * to amortize TLB flushing overheads. What this means is that any page you
963 * have now, may, in a former life, have been mapped into kernel virtual
964 * address by the vmap layer and so there might be some CPUs with TLB entries
965 * still referencing that page (additional to the regular 1:1 kernel mapping).
967 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
968 * be sure that none of the pages we have control over will have any aliases
969 * from the vmap layer.
971 void vm_unmap_aliases(void)
973 unsigned long start = ULONG_MAX, end = 0;
977 if (unlikely(!vmap_initialized))
980 for_each_possible_cpu(cpu) {
981 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
982 struct vmap_block *vb;
985 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
988 spin_lock(&vb->lock);
989 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
990 while (i < VMAP_BBMAP_BITS) {
993 j = find_next_zero_bit(vb->dirty_map,
996 s = vb->va->va_start + (i << PAGE_SHIFT);
997 e = vb->va->va_start + (j << PAGE_SHIFT);
1006 i = find_next_bit(vb->dirty_map,
1007 VMAP_BBMAP_BITS, i);
1009 spin_unlock(&vb->lock);
1014 __purge_vmap_area_lazy(&start, &end, 1, flush);
1016 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1019 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1020 * @mem: the pointer returned by vm_map_ram
1021 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1023 void vm_unmap_ram(const void *mem, unsigned int count)
1025 unsigned long size = count << PAGE_SHIFT;
1026 unsigned long addr = (unsigned long)mem;
1029 BUG_ON(addr < VMALLOC_START);
1030 BUG_ON(addr > VMALLOC_END);
1031 BUG_ON(addr & (PAGE_SIZE-1));
1033 debug_check_no_locks_freed(mem, size);
1034 vmap_debug_free_range(addr, addr+size);
1036 if (likely(count <= VMAP_MAX_ALLOC))
1039 free_unmap_vmap_area_addr(addr);
1041 EXPORT_SYMBOL(vm_unmap_ram);
1044 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1045 * @pages: an array of pointers to the pages to be mapped
1046 * @count: number of pages
1047 * @node: prefer to allocate data structures on this node
1048 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1050 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1052 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1054 unsigned long size = count << PAGE_SHIFT;
1058 if (likely(count <= VMAP_MAX_ALLOC)) {
1059 mem = vb_alloc(size, GFP_KERNEL);
1062 addr = (unsigned long)mem;
1064 struct vmap_area *va;
1065 va = alloc_vmap_area(size, PAGE_SIZE,
1066 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1070 addr = va->va_start;
1073 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1074 vm_unmap_ram(mem, count);
1079 EXPORT_SYMBOL(vm_map_ram);
1082 * vm_area_register_early - register vmap area early during boot
1083 * @vm: vm_struct to register
1084 * @align: requested alignment
1086 * This function is used to register kernel vm area before
1087 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1088 * proper values on entry and other fields should be zero. On return,
1089 * vm->addr contains the allocated address.
1091 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1093 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1095 static size_t vm_init_off __initdata;
1098 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1099 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1101 vm->addr = (void *)addr;
1107 void __init vmalloc_init(void)
1109 struct vmap_area *va;
1110 struct vm_struct *tmp;
1113 for_each_possible_cpu(i) {
1114 struct vmap_block_queue *vbq;
1116 vbq = &per_cpu(vmap_block_queue, i);
1117 spin_lock_init(&vbq->lock);
1118 INIT_LIST_HEAD(&vbq->free);
1121 /* Import existing vmlist entries. */
1122 for (tmp = vmlist; tmp; tmp = tmp->next) {
1123 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1124 va->flags = tmp->flags | VM_VM_AREA;
1125 va->va_start = (unsigned long)tmp->addr;
1126 va->va_end = va->va_start + tmp->size;
1127 __insert_vmap_area(va);
1130 vmap_area_pcpu_hole = VMALLOC_END;
1132 vmap_initialized = true;
1136 * map_kernel_range_noflush - map kernel VM area with the specified pages
1137 * @addr: start of the VM area to map
1138 * @size: size of the VM area to map
1139 * @prot: page protection flags to use
1140 * @pages: pages to map
1142 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1143 * specify should have been allocated using get_vm_area() and its
1147 * This function does NOT do any cache flushing. The caller is
1148 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1149 * before calling this function.
1152 * The number of pages mapped on success, -errno on failure.
1154 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1155 pgprot_t prot, struct page **pages)
1157 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1161 * unmap_kernel_range_noflush - unmap kernel VM area
1162 * @addr: start of the VM area to unmap
1163 * @size: size of the VM area to unmap
1165 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1166 * specify should have been allocated using get_vm_area() and its
1170 * This function does NOT do any cache flushing. The caller is
1171 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1172 * before calling this function and flush_tlb_kernel_range() after.
1174 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1176 vunmap_page_range(addr, addr + size);
1180 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1181 * @addr: start of the VM area to unmap
1182 * @size: size of the VM area to unmap
1184 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1185 * the unmapping and tlb after.
1187 void unmap_kernel_range(unsigned long addr, unsigned long size)
1189 unsigned long end = addr + size;
1191 flush_cache_vunmap(addr, end);
1192 vunmap_page_range(addr, end);
1193 flush_tlb_kernel_range(addr, end);
1196 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1198 unsigned long addr = (unsigned long)area->addr;
1199 unsigned long end = addr + area->size - PAGE_SIZE;
1202 err = vmap_page_range(addr, end, prot, *pages);
1210 EXPORT_SYMBOL_GPL(map_vm_area);
1212 /*** Old vmalloc interfaces ***/
1213 DEFINE_RWLOCK(vmlist_lock);
1214 struct vm_struct *vmlist;
1216 static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1217 unsigned long flags, void *caller)
1219 struct vm_struct *tmp, **p;
1222 vm->addr = (void *)va->va_start;
1223 vm->size = va->va_end - va->va_start;
1224 vm->caller = caller;
1226 va->flags |= VM_VM_AREA;
1228 write_lock(&vmlist_lock);
1229 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1230 if (tmp->addr >= vm->addr)
1235 write_unlock(&vmlist_lock);
1238 static struct vm_struct *__get_vm_area_node(unsigned long size,
1239 unsigned long align, unsigned long flags, unsigned long start,
1240 unsigned long end, int node, gfp_t gfp_mask, void *caller)
1242 static struct vmap_area *va;
1243 struct vm_struct *area;
1245 BUG_ON(in_interrupt());
1246 if (flags & VM_IOREMAP) {
1247 int bit = fls(size);
1249 if (bit > IOREMAP_MAX_ORDER)
1250 bit = IOREMAP_MAX_ORDER;
1251 else if (bit < PAGE_SHIFT)
1257 size = PAGE_ALIGN(size);
1258 if (unlikely(!size))
1261 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1262 if (unlikely(!area))
1266 * We always allocate a guard page.
1270 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1276 insert_vmalloc_vm(area, va, flags, caller);
1280 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1281 unsigned long start, unsigned long end)
1283 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1284 __builtin_return_address(0));
1286 EXPORT_SYMBOL_GPL(__get_vm_area);
1288 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1289 unsigned long start, unsigned long end,
1292 return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1297 * get_vm_area - reserve a contiguous kernel virtual area
1298 * @size: size of the area
1299 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1301 * Search an area of @size in the kernel virtual mapping area,
1302 * and reserved it for out purposes. Returns the area descriptor
1303 * on success or %NULL on failure.
1305 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1307 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1308 -1, GFP_KERNEL, __builtin_return_address(0));
1311 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1314 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1315 -1, GFP_KERNEL, caller);
1318 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1319 int node, gfp_t gfp_mask)
1321 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1322 node, gfp_mask, __builtin_return_address(0));
1325 static struct vm_struct *find_vm_area(const void *addr)
1327 struct vmap_area *va;
1329 va = find_vmap_area((unsigned long)addr);
1330 if (va && va->flags & VM_VM_AREA)
1337 * remove_vm_area - find and remove a continuous kernel virtual area
1338 * @addr: base address
1340 * Search for the kernel VM area starting at @addr, and remove it.
1341 * This function returns the found VM area, but using it is NOT safe
1342 * on SMP machines, except for its size or flags.
1344 struct vm_struct *remove_vm_area(const void *addr)
1346 struct vmap_area *va;
1348 va = find_vmap_area((unsigned long)addr);
1349 if (va && va->flags & VM_VM_AREA) {
1350 struct vm_struct *vm = va->private;
1351 struct vm_struct *tmp, **p;
1353 * remove from list and disallow access to this vm_struct
1354 * before unmap. (address range confliction is maintained by
1357 write_lock(&vmlist_lock);
1358 for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1361 write_unlock(&vmlist_lock);
1363 vmap_debug_free_range(va->va_start, va->va_end);
1364 free_unmap_vmap_area(va);
1365 vm->size -= PAGE_SIZE;
1372 static void __vunmap(const void *addr, int deallocate_pages)
1374 struct vm_struct *area;
1379 if ((PAGE_SIZE-1) & (unsigned long)addr) {
1380 WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1384 area = remove_vm_area(addr);
1385 if (unlikely(!area)) {
1386 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1391 debug_check_no_locks_freed(addr, area->size);
1392 debug_check_no_obj_freed(addr, area->size);
1394 if (deallocate_pages) {
1397 for (i = 0; i < area->nr_pages; i++) {
1398 struct page *page = area->pages[i];
1404 if (area->flags & VM_VPAGES)
1415 * vfree - release memory allocated by vmalloc()
1416 * @addr: memory base address
1418 * Free the virtually continuous memory area starting at @addr, as
1419 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1420 * NULL, no operation is performed.
1422 * Must not be called in interrupt context.
1424 void vfree(const void *addr)
1426 BUG_ON(in_interrupt());
1428 kmemleak_free(addr);
1432 EXPORT_SYMBOL(vfree);
1435 * vunmap - release virtual mapping obtained by vmap()
1436 * @addr: memory base address
1438 * Free the virtually contiguous memory area starting at @addr,
1439 * which was created from the page array passed to vmap().
1441 * Must not be called in interrupt context.
1443 void vunmap(const void *addr)
1445 BUG_ON(in_interrupt());
1449 EXPORT_SYMBOL(vunmap);
1452 * vmap - map an array of pages into virtually contiguous space
1453 * @pages: array of page pointers
1454 * @count: number of pages to map
1455 * @flags: vm_area->flags
1456 * @prot: page protection for the mapping
1458 * Maps @count pages from @pages into contiguous kernel virtual
1461 void *vmap(struct page **pages, unsigned int count,
1462 unsigned long flags, pgprot_t prot)
1464 struct vm_struct *area;
1468 if (count > totalram_pages)
1471 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1472 __builtin_return_address(0));
1476 if (map_vm_area(area, prot, &pages)) {
1483 EXPORT_SYMBOL(vmap);
1485 static void *__vmalloc_node(unsigned long size, unsigned long align,
1486 gfp_t gfp_mask, pgprot_t prot,
1487 int node, void *caller);
1488 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1489 pgprot_t prot, int node, void *caller)
1491 struct page **pages;
1492 unsigned int nr_pages, array_size, i;
1493 gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1495 nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1496 array_size = (nr_pages * sizeof(struct page *));
1498 area->nr_pages = nr_pages;
1499 /* Please note that the recursion is strictly bounded. */
1500 if (array_size > PAGE_SIZE) {
1501 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1502 PAGE_KERNEL, node, caller);
1503 area->flags |= VM_VPAGES;
1505 pages = kmalloc_node(array_size, nested_gfp, node);
1507 area->pages = pages;
1508 area->caller = caller;
1510 remove_vm_area(area->addr);
1515 for (i = 0; i < area->nr_pages; i++) {
1519 page = alloc_page(gfp_mask);
1521 page = alloc_pages_node(node, gfp_mask, 0);
1523 if (unlikely(!page)) {
1524 /* Successfully allocated i pages, free them in __vunmap() */
1528 area->pages[i] = page;
1531 if (map_vm_area(area, prot, &pages))
1540 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1542 void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1543 __builtin_return_address(0));
1546 * A ref_count = 3 is needed because the vm_struct and vmap_area
1547 * structures allocated in the __get_vm_area_node() function contain
1548 * references to the virtual address of the vmalloc'ed block.
1550 kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1556 * __vmalloc_node - allocate virtually contiguous memory
1557 * @size: allocation size
1558 * @align: desired alignment
1559 * @gfp_mask: flags for the page level allocator
1560 * @prot: protection mask for the allocated pages
1561 * @node: node to use for allocation or -1
1562 * @caller: caller's return address
1564 * Allocate enough pages to cover @size from the page level
1565 * allocator with @gfp_mask flags. Map them into contiguous
1566 * kernel virtual space, using a pagetable protection of @prot.
1568 static void *__vmalloc_node(unsigned long size, unsigned long align,
1569 gfp_t gfp_mask, pgprot_t prot,
1570 int node, void *caller)
1572 struct vm_struct *area;
1574 unsigned long real_size = size;
1576 size = PAGE_ALIGN(size);
1577 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1580 area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
1581 VMALLOC_END, node, gfp_mask, caller);
1586 addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1589 * A ref_count = 3 is needed because the vm_struct and vmap_area
1590 * structures allocated in the __get_vm_area_node() function contain
1591 * references to the virtual address of the vmalloc'ed block.
1593 kmemleak_alloc(addr, real_size, 3, gfp_mask);
1598 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1600 return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1601 __builtin_return_address(0));
1603 EXPORT_SYMBOL(__vmalloc);
1605 static inline void *__vmalloc_node_flags(unsigned long size,
1606 int node, gfp_t flags)
1608 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1609 node, __builtin_return_address(0));
1613 * vmalloc - allocate virtually contiguous memory
1614 * @size: allocation size
1615 * Allocate enough pages to cover @size from the page level
1616 * allocator and map them into contiguous kernel virtual space.
1618 * For tight control over page level allocator and protection flags
1619 * use __vmalloc() instead.
1621 void *vmalloc(unsigned long size)
1623 return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1625 EXPORT_SYMBOL(vmalloc);
1628 * vzalloc - allocate virtually contiguous memory with zero fill
1629 * @size: allocation size
1630 * Allocate enough pages to cover @size from the page level
1631 * allocator and map them into contiguous kernel virtual space.
1632 * The memory allocated is set to zero.
1634 * For tight control over page level allocator and protection flags
1635 * use __vmalloc() instead.
1637 void *vzalloc(unsigned long size)
1639 return __vmalloc_node_flags(size, -1,
1640 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1642 EXPORT_SYMBOL(vzalloc);
1645 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1646 * @size: allocation size
1648 * The resulting memory area is zeroed so it can be mapped to userspace
1649 * without leaking data.
1651 void *vmalloc_user(unsigned long size)
1653 struct vm_struct *area;
1656 ret = __vmalloc_node(size, SHMLBA,
1657 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1658 PAGE_KERNEL, -1, __builtin_return_address(0));
1660 area = find_vm_area(ret);
1661 area->flags |= VM_USERMAP;
1665 EXPORT_SYMBOL(vmalloc_user);
1668 * vmalloc_node - allocate memory on a specific node
1669 * @size: allocation size
1672 * Allocate enough pages to cover @size from the page level
1673 * allocator and map them into contiguous kernel virtual space.
1675 * For tight control over page level allocator and protection flags
1676 * use __vmalloc() instead.
1678 void *vmalloc_node(unsigned long size, int node)
1680 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1681 node, __builtin_return_address(0));
1683 EXPORT_SYMBOL(vmalloc_node);
1686 * vzalloc_node - allocate memory on a specific node with zero fill
1687 * @size: allocation size
1690 * Allocate enough pages to cover @size from the page level
1691 * allocator and map them into contiguous kernel virtual space.
1692 * The memory allocated is set to zero.
1694 * For tight control over page level allocator and protection flags
1695 * use __vmalloc_node() instead.
1697 void *vzalloc_node(unsigned long size, int node)
1699 return __vmalloc_node_flags(size, node,
1700 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1702 EXPORT_SYMBOL(vzalloc_node);
1704 #ifndef PAGE_KERNEL_EXEC
1705 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1709 * vmalloc_exec - allocate virtually contiguous, executable memory
1710 * @size: allocation size
1712 * Kernel-internal function to allocate enough pages to cover @size
1713 * the page level allocator and map them into contiguous and
1714 * executable kernel virtual space.
1716 * For tight control over page level allocator and protection flags
1717 * use __vmalloc() instead.
1720 void *vmalloc_exec(unsigned long size)
1722 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1723 -1, __builtin_return_address(0));
1726 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1727 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1728 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1729 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1731 #define GFP_VMALLOC32 GFP_KERNEL
1735 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1736 * @size: allocation size
1738 * Allocate enough 32bit PA addressable pages to cover @size from the
1739 * page level allocator and map them into contiguous kernel virtual space.
1741 void *vmalloc_32(unsigned long size)
1743 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1744 -1, __builtin_return_address(0));
1746 EXPORT_SYMBOL(vmalloc_32);
1749 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1750 * @size: allocation size
1752 * The resulting memory area is 32bit addressable and zeroed so it can be
1753 * mapped to userspace without leaking data.
1755 void *vmalloc_32_user(unsigned long size)
1757 struct vm_struct *area;
1760 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1761 -1, __builtin_return_address(0));
1763 area = find_vm_area(ret);
1764 area->flags |= VM_USERMAP;
1768 EXPORT_SYMBOL(vmalloc_32_user);
1771 * small helper routine , copy contents to buf from addr.
1772 * If the page is not present, fill zero.
1775 static int aligned_vread(char *buf, char *addr, unsigned long count)
1781 unsigned long offset, length;
1783 offset = (unsigned long)addr & ~PAGE_MASK;
1784 length = PAGE_SIZE - offset;
1787 p = vmalloc_to_page(addr);
1789 * To do safe access to this _mapped_ area, we need
1790 * lock. But adding lock here means that we need to add
1791 * overhead of vmalloc()/vfree() calles for this _debug_
1792 * interface, rarely used. Instead of that, we'll use
1793 * kmap() and get small overhead in this access function.
1797 * we can expect USER0 is not used (see vread/vwrite's
1798 * function description)
1800 void *map = kmap_atomic(p, KM_USER0);
1801 memcpy(buf, map + offset, length);
1802 kunmap_atomic(map, KM_USER0);
1804 memset(buf, 0, length);
1814 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1820 unsigned long offset, length;
1822 offset = (unsigned long)addr & ~PAGE_MASK;
1823 length = PAGE_SIZE - offset;
1826 p = vmalloc_to_page(addr);
1828 * To do safe access to this _mapped_ area, we need
1829 * lock. But adding lock here means that we need to add
1830 * overhead of vmalloc()/vfree() calles for this _debug_
1831 * interface, rarely used. Instead of that, we'll use
1832 * kmap() and get small overhead in this access function.
1836 * we can expect USER0 is not used (see vread/vwrite's
1837 * function description)
1839 void *map = kmap_atomic(p, KM_USER0);
1840 memcpy(map + offset, buf, length);
1841 kunmap_atomic(map, KM_USER0);
1852 * vread() - read vmalloc area in a safe way.
1853 * @buf: buffer for reading data
1854 * @addr: vm address.
1855 * @count: number of bytes to be read.
1857 * Returns # of bytes which addr and buf should be increased.
1858 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1859 * includes any intersect with alive vmalloc area.
1861 * This function checks that addr is a valid vmalloc'ed area, and
1862 * copy data from that area to a given buffer. If the given memory range
1863 * of [addr...addr+count) includes some valid address, data is copied to
1864 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1865 * IOREMAP area is treated as memory hole and no copy is done.
1867 * If [addr...addr+count) doesn't includes any intersects with alive
1868 * vm_struct area, returns 0.
1869 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1870 * the caller should guarantee KM_USER0 is not used.
1872 * Note: In usual ops, vread() is never necessary because the caller
1873 * should know vmalloc() area is valid and can use memcpy().
1874 * This is for routines which have to access vmalloc area without
1875 * any informaion, as /dev/kmem.
1879 long vread(char *buf, char *addr, unsigned long count)
1881 struct vm_struct *tmp;
1882 char *vaddr, *buf_start = buf;
1883 unsigned long buflen = count;
1886 /* Don't allow overflow */
1887 if ((unsigned long) addr + count < count)
1888 count = -(unsigned long) addr;
1890 read_lock(&vmlist_lock);
1891 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1892 vaddr = (char *) tmp->addr;
1893 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1895 while (addr < vaddr) {
1903 n = vaddr + tmp->size - PAGE_SIZE - addr;
1906 if (!(tmp->flags & VM_IOREMAP))
1907 aligned_vread(buf, addr, n);
1908 else /* IOREMAP area is treated as memory hole */
1915 read_unlock(&vmlist_lock);
1917 if (buf == buf_start)
1919 /* zero-fill memory holes */
1920 if (buf != buf_start + buflen)
1921 memset(buf, 0, buflen - (buf - buf_start));
1927 * vwrite() - write vmalloc area in a safe way.
1928 * @buf: buffer for source data
1929 * @addr: vm address.
1930 * @count: number of bytes to be read.
1932 * Returns # of bytes which addr and buf should be incresed.
1933 * (same number to @count).
1934 * If [addr...addr+count) doesn't includes any intersect with valid
1935 * vmalloc area, returns 0.
1937 * This function checks that addr is a valid vmalloc'ed area, and
1938 * copy data from a buffer to the given addr. If specified range of
1939 * [addr...addr+count) includes some valid address, data is copied from
1940 * proper area of @buf. If there are memory holes, no copy to hole.
1941 * IOREMAP area is treated as memory hole and no copy is done.
1943 * If [addr...addr+count) doesn't includes any intersects with alive
1944 * vm_struct area, returns 0.
1945 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1946 * the caller should guarantee KM_USER0 is not used.
1948 * Note: In usual ops, vwrite() is never necessary because the caller
1949 * should know vmalloc() area is valid and can use memcpy().
1950 * This is for routines which have to access vmalloc area without
1951 * any informaion, as /dev/kmem.
1953 * The caller should guarantee KM_USER1 is not used.
1956 long vwrite(char *buf, char *addr, unsigned long count)
1958 struct vm_struct *tmp;
1960 unsigned long n, buflen;
1963 /* Don't allow overflow */
1964 if ((unsigned long) addr + count < count)
1965 count = -(unsigned long) addr;
1968 read_lock(&vmlist_lock);
1969 for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1970 vaddr = (char *) tmp->addr;
1971 if (addr >= vaddr + tmp->size - PAGE_SIZE)
1973 while (addr < vaddr) {
1980 n = vaddr + tmp->size - PAGE_SIZE - addr;
1983 if (!(tmp->flags & VM_IOREMAP)) {
1984 aligned_vwrite(buf, addr, n);
1992 read_unlock(&vmlist_lock);
1999 * remap_vmalloc_range - map vmalloc pages to userspace
2000 * @vma: vma to cover (map full range of vma)
2001 * @addr: vmalloc memory
2002 * @pgoff: number of pages into addr before first page to map
2004 * Returns: 0 for success, -Exxx on failure
2006 * This function checks that addr is a valid vmalloc'ed area, and
2007 * that it is big enough to cover the vma. Will return failure if
2008 * that criteria isn't met.
2010 * Similar to remap_pfn_range() (see mm/memory.c)
2012 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2013 unsigned long pgoff)
2015 struct vm_struct *area;
2016 unsigned long uaddr = vma->vm_start;
2017 unsigned long usize = vma->vm_end - vma->vm_start;
2019 if ((PAGE_SIZE-1) & (unsigned long)addr)
2022 area = find_vm_area(addr);
2026 if (!(area->flags & VM_USERMAP))
2029 if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2032 addr += pgoff << PAGE_SHIFT;
2034 struct page *page = vmalloc_to_page(addr);
2037 ret = vm_insert_page(vma, uaddr, page);
2044 } while (usize > 0);
2046 /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2047 vma->vm_flags |= VM_RESERVED;
2051 EXPORT_SYMBOL(remap_vmalloc_range);
2054 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2057 void __attribute__((weak)) vmalloc_sync_all(void)
2062 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2064 /* apply_to_page_range() does all the hard work. */
2069 * alloc_vm_area - allocate a range of kernel address space
2070 * @size: size of the area
2072 * Returns: NULL on failure, vm_struct on success
2074 * This function reserves a range of kernel address space, and
2075 * allocates pagetables to map that range. No actual mappings
2076 * are created. If the kernel address space is not shared
2077 * between processes, it syncs the pagetable across all
2080 struct vm_struct *alloc_vm_area(size_t size)
2082 struct vm_struct *area;
2084 area = get_vm_area_caller(size, VM_IOREMAP,
2085 __builtin_return_address(0));
2090 * This ensures that page tables are constructed for this region
2091 * of kernel virtual address space and mapped into init_mm.
2093 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2094 area->size, f, NULL)) {
2099 /* Make sure the pagetables are constructed in process kernel
2105 EXPORT_SYMBOL_GPL(alloc_vm_area);
2107 void free_vm_area(struct vm_struct *area)
2109 struct vm_struct *ret;
2110 ret = remove_vm_area(area->addr);
2111 BUG_ON(ret != area);
2114 EXPORT_SYMBOL_GPL(free_vm_area);
2117 static struct vmap_area *node_to_va(struct rb_node *n)
2119 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2123 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2124 * @end: target address
2125 * @pnext: out arg for the next vmap_area
2126 * @pprev: out arg for the previous vmap_area
2128 * Returns: %true if either or both of next and prev are found,
2129 * %false if no vmap_area exists
2131 * Find vmap_areas end addresses of which enclose @end. ie. if not
2132 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2134 static bool pvm_find_next_prev(unsigned long end,
2135 struct vmap_area **pnext,
2136 struct vmap_area **pprev)
2138 struct rb_node *n = vmap_area_root.rb_node;
2139 struct vmap_area *va = NULL;
2142 va = rb_entry(n, struct vmap_area, rb_node);
2143 if (end < va->va_end)
2145 else if (end > va->va_end)
2154 if (va->va_end > end) {
2156 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2159 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2165 * pvm_determine_end - find the highest aligned address between two vmap_areas
2166 * @pnext: in/out arg for the next vmap_area
2167 * @pprev: in/out arg for the previous vmap_area
2170 * Returns: determined end address
2172 * Find the highest aligned address between *@pnext and *@pprev below
2173 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2174 * down address is between the end addresses of the two vmap_areas.
2176 * Please note that the address returned by this function may fall
2177 * inside *@pnext vmap_area. The caller is responsible for checking
2180 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2181 struct vmap_area **pprev,
2182 unsigned long align)
2184 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2188 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2192 while (*pprev && (*pprev)->va_end > addr) {
2194 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2201 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2202 * @offsets: array containing offset of each area
2203 * @sizes: array containing size of each area
2204 * @nr_vms: the number of areas to allocate
2205 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2206 * @gfp_mask: allocation mask
2208 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2209 * vm_structs on success, %NULL on failure
2211 * Percpu allocator wants to use congruent vm areas so that it can
2212 * maintain the offsets among percpu areas. This function allocates
2213 * congruent vmalloc areas for it. These areas tend to be scattered
2214 * pretty far, distance between two areas easily going up to
2215 * gigabytes. To avoid interacting with regular vmallocs, these areas
2216 * are allocated from top.
2218 * Despite its complicated look, this allocator is rather simple. It
2219 * does everything top-down and scans areas from the end looking for
2220 * matching slot. While scanning, if any of the areas overlaps with
2221 * existing vmap_area, the base address is pulled down to fit the
2222 * area. Scanning is repeated till all the areas fit and then all
2223 * necessary data structres are inserted and the result is returned.
2225 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2226 const size_t *sizes, int nr_vms,
2227 size_t align, gfp_t gfp_mask)
2229 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2230 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2231 struct vmap_area **vas, *prev, *next;
2232 struct vm_struct **vms;
2233 int area, area2, last_area, term_area;
2234 unsigned long base, start, end, last_end;
2235 bool purged = false;
2237 gfp_mask &= GFP_RECLAIM_MASK;
2239 /* verify parameters and allocate data structures */
2240 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2241 for (last_area = 0, area = 0; area < nr_vms; area++) {
2242 start = offsets[area];
2243 end = start + sizes[area];
2245 /* is everything aligned properly? */
2246 BUG_ON(!IS_ALIGNED(offsets[area], align));
2247 BUG_ON(!IS_ALIGNED(sizes[area], align));
2249 /* detect the area with the highest address */
2250 if (start > offsets[last_area])
2253 for (area2 = 0; area2 < nr_vms; area2++) {
2254 unsigned long start2 = offsets[area2];
2255 unsigned long end2 = start2 + sizes[area2];
2260 BUG_ON(start2 >= start && start2 < end);
2261 BUG_ON(end2 <= end && end2 > start);
2264 last_end = offsets[last_area] + sizes[last_area];
2266 if (vmalloc_end - vmalloc_start < last_end) {
2271 vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2272 vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2276 for (area = 0; area < nr_vms; area++) {
2277 vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2278 vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2279 if (!vas[area] || !vms[area])
2283 spin_lock(&vmap_area_lock);
2285 /* start scanning - we scan from the top, begin with the last area */
2286 area = term_area = last_area;
2287 start = offsets[area];
2288 end = start + sizes[area];
2290 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2291 base = vmalloc_end - last_end;
2294 base = pvm_determine_end(&next, &prev, align) - end;
2297 BUG_ON(next && next->va_end <= base + end);
2298 BUG_ON(prev && prev->va_end > base + end);
2301 * base might have underflowed, add last_end before
2304 if (base + last_end < vmalloc_start + last_end) {
2305 spin_unlock(&vmap_area_lock);
2307 purge_vmap_area_lazy();
2315 * If next overlaps, move base downwards so that it's
2316 * right below next and then recheck.
2318 if (next && next->va_start < base + end) {
2319 base = pvm_determine_end(&next, &prev, align) - end;
2325 * If prev overlaps, shift down next and prev and move
2326 * base so that it's right below new next and then
2329 if (prev && prev->va_end > base + start) {
2331 prev = node_to_va(rb_prev(&next->rb_node));
2332 base = pvm_determine_end(&next, &prev, align) - end;
2338 * This area fits, move on to the previous one. If
2339 * the previous one is the terminal one, we're done.
2341 area = (area + nr_vms - 1) % nr_vms;
2342 if (area == term_area)
2344 start = offsets[area];
2345 end = start + sizes[area];
2346 pvm_find_next_prev(base + end, &next, &prev);
2349 /* we've found a fitting base, insert all va's */
2350 for (area = 0; area < nr_vms; area++) {
2351 struct vmap_area *va = vas[area];
2353 va->va_start = base + offsets[area];
2354 va->va_end = va->va_start + sizes[area];
2355 __insert_vmap_area(va);
2358 vmap_area_pcpu_hole = base + offsets[last_area];
2360 spin_unlock(&vmap_area_lock);
2362 /* insert all vm's */
2363 for (area = 0; area < nr_vms; area++)
2364 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2371 for (area = 0; area < nr_vms; area++) {
2383 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2384 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2385 * @nr_vms: the number of allocated areas
2387 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2389 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2393 for (i = 0; i < nr_vms; i++)
2394 free_vm_area(vms[i]);
2397 #endif /* CONFIG_SMP */
2399 #ifdef CONFIG_PROC_FS
2400 static void *s_start(struct seq_file *m, loff_t *pos)
2401 __acquires(&vmlist_lock)
2404 struct vm_struct *v;
2406 read_lock(&vmlist_lock);
2408 while (n > 0 && v) {
2419 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2421 struct vm_struct *v = p;
2427 static void s_stop(struct seq_file *m, void *p)
2428 __releases(&vmlist_lock)
2430 read_unlock(&vmlist_lock);
2433 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2436 unsigned int nr, *counters = m->private;
2441 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2443 for (nr = 0; nr < v->nr_pages; nr++)
2444 counters[page_to_nid(v->pages[nr])]++;
2446 for_each_node_state(nr, N_HIGH_MEMORY)
2448 seq_printf(m, " N%u=%u", nr, counters[nr]);
2452 static int s_show(struct seq_file *m, void *p)
2454 struct vm_struct *v = p;
2456 seq_printf(m, "0x%p-0x%p %7ld",
2457 v->addr, v->addr + v->size, v->size);
2460 char buff[KSYM_SYMBOL_LEN];
2463 sprint_symbol(buff, (unsigned long)v->caller);
2468 seq_printf(m, " pages=%d", v->nr_pages);
2471 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2473 if (v->flags & VM_IOREMAP)
2474 seq_printf(m, " ioremap");
2476 if (v->flags & VM_ALLOC)
2477 seq_printf(m, " vmalloc");
2479 if (v->flags & VM_MAP)
2480 seq_printf(m, " vmap");
2482 if (v->flags & VM_USERMAP)
2483 seq_printf(m, " user");
2485 if (v->flags & VM_VPAGES)
2486 seq_printf(m, " vpages");
2488 show_numa_info(m, v);
2493 static const struct seq_operations vmalloc_op = {
2500 static int vmalloc_open(struct inode *inode, struct file *file)
2502 unsigned int *ptr = NULL;
2506 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2510 ret = seq_open(file, &vmalloc_op);
2512 struct seq_file *m = file->private_data;
2519 static const struct file_operations proc_vmalloc_operations = {
2520 .open = vmalloc_open,
2522 .llseek = seq_lseek,
2523 .release = seq_release_private,
2526 static int __init proc_vmalloc_init(void)
2528 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2531 module_init(proc_vmalloc_init);