Root/mm/vmalloc.c

1/*
2 * linux/mm/vmalloc.c
3 *
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
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.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>
33
34
35/*** Page table manipulation functions ***/
36
37static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
38{
39    pte_t *pte;
40
41    pte = pte_offset_kernel(pmd, addr);
42    do {
43        pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
44        WARN_ON(!pte_none(ptent) && !pte_present(ptent));
45    } while (pte++, addr += PAGE_SIZE, addr != end);
46}
47
48static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
49{
50    pmd_t *pmd;
51    unsigned long next;
52
53    pmd = pmd_offset(pud, addr);
54    do {
55        next = pmd_addr_end(addr, end);
56        if (pmd_none_or_clear_bad(pmd))
57            continue;
58        vunmap_pte_range(pmd, addr, next);
59    } while (pmd++, addr = next, addr != end);
60}
61
62static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
63{
64    pud_t *pud;
65    unsigned long next;
66
67    pud = pud_offset(pgd, addr);
68    do {
69        next = pud_addr_end(addr, end);
70        if (pud_none_or_clear_bad(pud))
71            continue;
72        vunmap_pmd_range(pud, addr, next);
73    } while (pud++, addr = next, addr != end);
74}
75
76static void vunmap_page_range(unsigned long addr, unsigned long end)
77{
78    pgd_t *pgd;
79    unsigned long next;
80
81    BUG_ON(addr >= end);
82    pgd = pgd_offset_k(addr);
83    do {
84        next = pgd_addr_end(addr, end);
85        if (pgd_none_or_clear_bad(pgd))
86            continue;
87        vunmap_pud_range(pgd, addr, next);
88    } while (pgd++, addr = next, addr != end);
89}
90
91static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
92        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
93{
94    pte_t *pte;
95
96    /*
97     * nr is a running index into the array which helps higher level
98     * callers keep track of where we're up to.
99     */
100
101    pte = pte_alloc_kernel(pmd, addr);
102    if (!pte)
103        return -ENOMEM;
104    do {
105        struct page *page = pages[*nr];
106
107        if (WARN_ON(!pte_none(*pte)))
108            return -EBUSY;
109        if (WARN_ON(!page))
110            return -ENOMEM;
111        set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
112        (*nr)++;
113    } while (pte++, addr += PAGE_SIZE, addr != end);
114    return 0;
115}
116
117static int vmap_pmd_range(pud_t *pud, unsigned long addr,
118        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
119{
120    pmd_t *pmd;
121    unsigned long next;
122
123    pmd = pmd_alloc(&init_mm, pud, addr);
124    if (!pmd)
125        return -ENOMEM;
126    do {
127        next = pmd_addr_end(addr, end);
128        if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
129            return -ENOMEM;
130    } while (pmd++, addr = next, addr != end);
131    return 0;
132}
133
134static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
135        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
136{
137    pud_t *pud;
138    unsigned long next;
139
140    pud = pud_alloc(&init_mm, pgd, addr);
141    if (!pud)
142        return -ENOMEM;
143    do {
144        next = pud_addr_end(addr, end);
145        if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
146            return -ENOMEM;
147    } while (pud++, addr = next, addr != end);
148    return 0;
149}
150
151/*
152 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
153 * will have pfns corresponding to the "pages" array.
154 *
155 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
156 */
157static int vmap_page_range_noflush(unsigned long start, unsigned long end,
158                   pgprot_t prot, struct page **pages)
159{
160    pgd_t *pgd;
161    unsigned long next;
162    unsigned long addr = start;
163    int err = 0;
164    int nr = 0;
165
166    BUG_ON(addr >= end);
167    pgd = pgd_offset_k(addr);
168    do {
169        next = pgd_addr_end(addr, end);
170        err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
171        if (err)
172            return err;
173    } while (pgd++, addr = next, addr != end);
174
175    return nr;
176}
177
178static int vmap_page_range(unsigned long start, unsigned long end,
179               pgprot_t prot, struct page **pages)
180{
181    int ret;
182
183    ret = vmap_page_range_noflush(start, end, prot, pages);
184    flush_cache_vmap(start, end);
185    return ret;
186}
187
188int is_vmalloc_or_module_addr(const void *x)
189{
190    /*
191     * ARM, x86-64 and sparc64 put modules in a special place,
192     * and fall back on vmalloc() if that fails. Others
193     * just put it in the vmalloc space.
194     */
195#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
196    unsigned long addr = (unsigned long)x;
197    if (addr >= MODULES_VADDR && addr < MODULES_END)
198        return 1;
199#endif
200    return is_vmalloc_addr(x);
201}
202
203/*
204 * Walk a vmap address to the struct page it maps.
205 */
206struct page *vmalloc_to_page(const void *vmalloc_addr)
207{
208    unsigned long addr = (unsigned long) vmalloc_addr;
209    struct page *page = NULL;
210    pgd_t *pgd = pgd_offset_k(addr);
211
212    /*
213     * XXX we might need to change this if we add VIRTUAL_BUG_ON for
214     * architectures that do not vmalloc module space
215     */
216    VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
217
218    if (!pgd_none(*pgd)) {
219        pud_t *pud = pud_offset(pgd, addr);
220        if (!pud_none(*pud)) {
221            pmd_t *pmd = pmd_offset(pud, addr);
222            if (!pmd_none(*pmd)) {
223                pte_t *ptep, pte;
224
225                ptep = pte_offset_map(pmd, addr);
226                pte = *ptep;
227                if (pte_present(pte))
228                    page = pte_page(pte);
229                pte_unmap(ptep);
230            }
231        }
232    }
233    return page;
234}
235EXPORT_SYMBOL(vmalloc_to_page);
236
237/*
238 * Map a vmalloc()-space virtual address to the physical page frame number.
239 */
240unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
241{
242    return page_to_pfn(vmalloc_to_page(vmalloc_addr));
243}
244EXPORT_SYMBOL(vmalloc_to_pfn);
245
246
247/*** Global kva allocator ***/
248
249#define VM_LAZY_FREE 0x01
250#define VM_LAZY_FREEING 0x02
251#define VM_VM_AREA 0x04
252
253struct vmap_area {
254    unsigned long va_start;
255    unsigned long va_end;
256    unsigned long flags;
257    struct rb_node rb_node; /* address sorted rbtree */
258    struct list_head list; /* address sorted list */
259    struct list_head purge_list; /* "lazy purge" list */
260    void *private;
261    struct rcu_head rcu_head;
262};
263
264static DEFINE_SPINLOCK(vmap_area_lock);
265static struct rb_root vmap_area_root = RB_ROOT;
266static LIST_HEAD(vmap_area_list);
267static unsigned long vmap_area_pcpu_hole;
268
269static struct vmap_area *__find_vmap_area(unsigned long addr)
270{
271    struct rb_node *n = vmap_area_root.rb_node;
272
273    while (n) {
274        struct vmap_area *va;
275
276        va = rb_entry(n, struct vmap_area, rb_node);
277        if (addr < va->va_start)
278            n = n->rb_left;
279        else if (addr > va->va_start)
280            n = n->rb_right;
281        else
282            return va;
283    }
284
285    return NULL;
286}
287
288static void __insert_vmap_area(struct vmap_area *va)
289{
290    struct rb_node **p = &vmap_area_root.rb_node;
291    struct rb_node *parent = NULL;
292    struct rb_node *tmp;
293
294    while (*p) {
295        struct vmap_area *tmp;
296
297        parent = *p;
298        tmp = rb_entry(parent, struct vmap_area, rb_node);
299        if (va->va_start < tmp->va_end)
300            p = &(*p)->rb_left;
301        else if (va->va_end > tmp->va_start)
302            p = &(*p)->rb_right;
303        else
304            BUG();
305    }
306
307    rb_link_node(&va->rb_node, parent, p);
308    rb_insert_color(&va->rb_node, &vmap_area_root);
309
310    /* address-sort this list so it is usable like the vmlist */
311    tmp = rb_prev(&va->rb_node);
312    if (tmp) {
313        struct vmap_area *prev;
314        prev = rb_entry(tmp, struct vmap_area, rb_node);
315        list_add_rcu(&va->list, &prev->list);
316    } else
317        list_add_rcu(&va->list, &vmap_area_list);
318}
319
320static void purge_vmap_area_lazy(void);
321
322/*
323 * Allocate a region of KVA of the specified size and alignment, within the
324 * vstart and vend.
325 */
326static struct vmap_area *alloc_vmap_area(unsigned long size,
327                unsigned long align,
328                unsigned long vstart, unsigned long vend,
329                int node, gfp_t gfp_mask)
330{
331    struct vmap_area *va;
332    struct rb_node *n;
333    unsigned long addr;
334    int purged = 0;
335
336    BUG_ON(!size);
337    BUG_ON(size & ~PAGE_MASK);
338
339    va = kmalloc_node(sizeof(struct vmap_area),
340            gfp_mask & GFP_RECLAIM_MASK, node);
341    if (unlikely(!va))
342        return ERR_PTR(-ENOMEM);
343
344retry:
345    addr = ALIGN(vstart, align);
346
347    spin_lock(&vmap_area_lock);
348    if (addr + size - 1 < addr)
349        goto overflow;
350
351    /* XXX: could have a last_hole cache */
352    n = vmap_area_root.rb_node;
353    if (n) {
354        struct vmap_area *first = NULL;
355
356        do {
357            struct vmap_area *tmp;
358            tmp = rb_entry(n, struct vmap_area, rb_node);
359            if (tmp->va_end >= addr) {
360                if (!first && tmp->va_start < addr + size)
361                    first = tmp;
362                n = n->rb_left;
363            } else {
364                first = tmp;
365                n = n->rb_right;
366            }
367        } while (n);
368
369        if (!first)
370            goto found;
371
372        if (first->va_end < addr) {
373            n = rb_next(&first->rb_node);
374            if (n)
375                first = rb_entry(n, struct vmap_area, rb_node);
376            else
377                goto found;
378        }
379
380        while (addr + size > first->va_start && addr + size <= vend) {
381            addr = ALIGN(first->va_end + PAGE_SIZE, align);
382            if (addr + size - 1 < addr)
383                goto overflow;
384
385            n = rb_next(&first->rb_node);
386            if (n)
387                first = rb_entry(n, struct vmap_area, rb_node);
388            else
389                goto found;
390        }
391    }
392found:
393    if (addr + size > vend) {
394overflow:
395        spin_unlock(&vmap_area_lock);
396        if (!purged) {
397            purge_vmap_area_lazy();
398            purged = 1;
399            goto retry;
400        }
401        if (printk_ratelimit())
402            printk(KERN_WARNING
403                "vmap allocation for size %lu failed: "
404                "use vmalloc=<size> to increase size.\n", size);
405        kfree(va);
406        return ERR_PTR(-EBUSY);
407    }
408
409    BUG_ON(addr & (align-1));
410
411    va->va_start = addr;
412    va->va_end = addr + size;
413    va->flags = 0;
414    __insert_vmap_area(va);
415    spin_unlock(&vmap_area_lock);
416
417    return va;
418}
419
420static void rcu_free_va(struct rcu_head *head)
421{
422    struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
423
424    kfree(va);
425}
426
427static void __free_vmap_area(struct vmap_area *va)
428{
429    BUG_ON(RB_EMPTY_NODE(&va->rb_node));
430    rb_erase(&va->rb_node, &vmap_area_root);
431    RB_CLEAR_NODE(&va->rb_node);
432    list_del_rcu(&va->list);
433
434    /*
435     * Track the highest possible candidate for pcpu area
436     * allocation. Areas outside of vmalloc area can be returned
437     * here too, consider only end addresses which fall inside
438     * vmalloc area proper.
439     */
440    if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
441        vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
442
443    call_rcu(&va->rcu_head, rcu_free_va);
444}
445
446/*
447 * Free a region of KVA allocated by alloc_vmap_area
448 */
449static void free_vmap_area(struct vmap_area *va)
450{
451    spin_lock(&vmap_area_lock);
452    __free_vmap_area(va);
453    spin_unlock(&vmap_area_lock);
454}
455
456/*
457 * Clear the pagetable entries of a given vmap_area
458 */
459static void unmap_vmap_area(struct vmap_area *va)
460{
461    vunmap_page_range(va->va_start, va->va_end);
462}
463
464static void vmap_debug_free_range(unsigned long start, unsigned long end)
465{
466    /*
467     * Unmap page tables and force a TLB flush immediately if
468     * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
469     * bugs similarly to those in linear kernel virtual address
470     * space after a page has been freed.
471     *
472     * All the lazy freeing logic is still retained, in order to
473     * minimise intrusiveness of this debugging feature.
474     *
475     * This is going to be *slow* (linear kernel virtual address
476     * debugging doesn't do a broadcast TLB flush so it is a lot
477     * faster).
478     */
479#ifdef CONFIG_DEBUG_PAGEALLOC
480    vunmap_page_range(start, end);
481    flush_tlb_kernel_range(start, end);
482#endif
483}
484
485/*
486 * lazy_max_pages is the maximum amount of virtual address space we gather up
487 * before attempting to purge with a TLB flush.
488 *
489 * There is a tradeoff here: a larger number will cover more kernel page tables
490 * and take slightly longer to purge, but it will linearly reduce the number of
491 * global TLB flushes that must be performed. It would seem natural to scale
492 * this number up linearly with the number of CPUs (because vmapping activity
493 * could also scale linearly with the number of CPUs), however it is likely
494 * that in practice, workloads might be constrained in other ways that mean
495 * vmap activity will not scale linearly with CPUs. Also, I want to be
496 * conservative and not introduce a big latency on huge systems, so go with
497 * a less aggressive log scale. It will still be an improvement over the old
498 * code, and it will be simple to change the scale factor if we find that it
499 * becomes a problem on bigger systems.
500 */
501static unsigned long lazy_max_pages(void)
502{
503    unsigned int log;
504
505    log = fls(num_online_cpus());
506
507    return log * (32UL * 1024 * 1024 / PAGE_SIZE);
508}
509
510static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
511
512/* for per-CPU blocks */
513static void purge_fragmented_blocks_allcpus(void);
514
515/*
516 * Purges all lazily-freed vmap areas.
517 *
518 * If sync is 0 then don't purge if there is already a purge in progress.
519 * If force_flush is 1, then flush kernel TLBs between *start and *end even
520 * if we found no lazy vmap areas to unmap (callers can use this to optimise
521 * their own TLB flushing).
522 * Returns with *start = min(*start, lowest purged address)
523 * *end = max(*end, highest purged address)
524 */
525static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
526                    int sync, int force_flush)
527{
528    static DEFINE_SPINLOCK(purge_lock);
529    LIST_HEAD(valist);
530    struct vmap_area *va;
531    struct vmap_area *n_va;
532    int nr = 0;
533
534    /*
535     * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
536     * should not expect such behaviour. This just simplifies locking for
537     * the case that isn't actually used at the moment anyway.
538     */
539    if (!sync && !force_flush) {
540        if (!spin_trylock(&purge_lock))
541            return;
542    } else
543        spin_lock(&purge_lock);
544
545    if (sync)
546        purge_fragmented_blocks_allcpus();
547
548    rcu_read_lock();
549    list_for_each_entry_rcu(va, &vmap_area_list, list) {
550        if (va->flags & VM_LAZY_FREE) {
551            if (va->va_start < *start)
552                *start = va->va_start;
553            if (va->va_end > *end)
554                *end = va->va_end;
555            nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
556            unmap_vmap_area(va);
557            list_add_tail(&va->purge_list, &valist);
558            va->flags |= VM_LAZY_FREEING;
559            va->flags &= ~VM_LAZY_FREE;
560        }
561    }
562    rcu_read_unlock();
563
564    if (nr)
565        atomic_sub(nr, &vmap_lazy_nr);
566
567    if (nr || force_flush)
568        flush_tlb_kernel_range(*start, *end);
569
570    if (nr) {
571        spin_lock(&vmap_area_lock);
572        list_for_each_entry_safe(va, n_va, &valist, purge_list)
573            __free_vmap_area(va);
574        spin_unlock(&vmap_area_lock);
575    }
576    spin_unlock(&purge_lock);
577}
578
579/*
580 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
581 * is already purging.
582 */
583static void try_purge_vmap_area_lazy(void)
584{
585    unsigned long start = ULONG_MAX, end = 0;
586
587    __purge_vmap_area_lazy(&start, &end, 0, 0);
588}
589
590/*
591 * Kick off a purge of the outstanding lazy areas.
592 */
593static void purge_vmap_area_lazy(void)
594{
595    unsigned long start = ULONG_MAX, end = 0;
596
597    __purge_vmap_area_lazy(&start, &end, 1, 0);
598}
599
600/*
601 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
602 * called for the correct range previously.
603 */
604static void free_unmap_vmap_area_noflush(struct vmap_area *va)
605{
606    va->flags |= VM_LAZY_FREE;
607    atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
608    if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
609        try_purge_vmap_area_lazy();
610}
611
612/*
613 * Free and unmap a vmap area
614 */
615static void free_unmap_vmap_area(struct vmap_area *va)
616{
617    flush_cache_vunmap(va->va_start, va->va_end);
618    free_unmap_vmap_area_noflush(va);
619}
620
621static struct vmap_area *find_vmap_area(unsigned long addr)
622{
623    struct vmap_area *va;
624
625    spin_lock(&vmap_area_lock);
626    va = __find_vmap_area(addr);
627    spin_unlock(&vmap_area_lock);
628
629    return va;
630}
631
632static void free_unmap_vmap_area_addr(unsigned long addr)
633{
634    struct vmap_area *va;
635
636    va = find_vmap_area(addr);
637    BUG_ON(!va);
638    free_unmap_vmap_area(va);
639}
640
641
642/*** Per cpu kva allocator ***/
643
644/*
645 * vmap space is limited especially on 32 bit architectures. Ensure there is
646 * room for at least 16 percpu vmap blocks per CPU.
647 */
648/*
649 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
650 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
651 * instead (we just need a rough idea)
652 */
653#if BITS_PER_LONG == 32
654#define VMALLOC_SPACE (128UL*1024*1024)
655#else
656#define VMALLOC_SPACE (128UL*1024*1024*1024)
657#endif
658
659#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
660#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
661#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
662#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
663#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
664#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
665#define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
666                    VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
667                        VMALLOC_PAGES / NR_CPUS / 16))
668
669#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
670
671static bool vmap_initialized __read_mostly = false;
672
673struct vmap_block_queue {
674    spinlock_t lock;
675    struct list_head free;
676};
677
678struct vmap_block {
679    spinlock_t lock;
680    struct vmap_area *va;
681    struct vmap_block_queue *vbq;
682    unsigned long free, dirty;
683    DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
684    DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
685    struct list_head free_list;
686    struct rcu_head rcu_head;
687    struct list_head purge;
688};
689
690/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
691static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
692
693/*
694 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
695 * in the free path. Could get rid of this if we change the API to return a
696 * "cookie" from alloc, to be passed to free. But no big deal yet.
697 */
698static DEFINE_SPINLOCK(vmap_block_tree_lock);
699static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
700
701/*
702 * We should probably have a fallback mechanism to allocate virtual memory
703 * out of partially filled vmap blocks. However vmap block sizing should be
704 * fairly reasonable according to the vmalloc size, so it shouldn't be a
705 * big problem.
706 */
707
708static unsigned long addr_to_vb_idx(unsigned long addr)
709{
710    addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
711    addr /= VMAP_BLOCK_SIZE;
712    return addr;
713}
714
715static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
716{
717    struct vmap_block_queue *vbq;
718    struct vmap_block *vb;
719    struct vmap_area *va;
720    unsigned long vb_idx;
721    int node, err;
722
723    node = numa_node_id();
724
725    vb = kmalloc_node(sizeof(struct vmap_block),
726            gfp_mask & GFP_RECLAIM_MASK, node);
727    if (unlikely(!vb))
728        return ERR_PTR(-ENOMEM);
729
730    va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
731                    VMALLOC_START, VMALLOC_END,
732                    node, gfp_mask);
733    if (unlikely(IS_ERR(va))) {
734        kfree(vb);
735        return ERR_PTR(PTR_ERR(va));
736    }
737
738    err = radix_tree_preload(gfp_mask);
739    if (unlikely(err)) {
740        kfree(vb);
741        free_vmap_area(va);
742        return ERR_PTR(err);
743    }
744
745    spin_lock_init(&vb->lock);
746    vb->va = va;
747    vb->free = VMAP_BBMAP_BITS;
748    vb->dirty = 0;
749    bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
750    bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
751    INIT_LIST_HEAD(&vb->free_list);
752
753    vb_idx = addr_to_vb_idx(va->va_start);
754    spin_lock(&vmap_block_tree_lock);
755    err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
756    spin_unlock(&vmap_block_tree_lock);
757    BUG_ON(err);
758    radix_tree_preload_end();
759
760    vbq = &get_cpu_var(vmap_block_queue);
761    vb->vbq = vbq;
762    spin_lock(&vbq->lock);
763    list_add_rcu(&vb->free_list, &vbq->free);
764    spin_unlock(&vbq->lock);
765    put_cpu_var(vmap_block_queue);
766
767    return vb;
768}
769
770static void rcu_free_vb(struct rcu_head *head)
771{
772    struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
773
774    kfree(vb);
775}
776
777static void free_vmap_block(struct vmap_block *vb)
778{
779    struct vmap_block *tmp;
780    unsigned long vb_idx;
781
782    vb_idx = addr_to_vb_idx(vb->va->va_start);
783    spin_lock(&vmap_block_tree_lock);
784    tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
785    spin_unlock(&vmap_block_tree_lock);
786    BUG_ON(tmp != vb);
787
788    free_unmap_vmap_area_noflush(vb->va);
789    call_rcu(&vb->rcu_head, rcu_free_vb);
790}
791
792static void purge_fragmented_blocks(int cpu)
793{
794    LIST_HEAD(purge);
795    struct vmap_block *vb;
796    struct vmap_block *n_vb;
797    struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
798
799    rcu_read_lock();
800    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
801
802        if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
803            continue;
804
805        spin_lock(&vb->lock);
806        if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
807            vb->free = 0; /* prevent further allocs after releasing lock */
808            vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
809            bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
810            bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
811            spin_lock(&vbq->lock);
812            list_del_rcu(&vb->free_list);
813            spin_unlock(&vbq->lock);
814            spin_unlock(&vb->lock);
815            list_add_tail(&vb->purge, &purge);
816        } else
817            spin_unlock(&vb->lock);
818    }
819    rcu_read_unlock();
820
821    list_for_each_entry_safe(vb, n_vb, &purge, purge) {
822        list_del(&vb->purge);
823        free_vmap_block(vb);
824    }
825}
826
827static void purge_fragmented_blocks_thiscpu(void)
828{
829    purge_fragmented_blocks(smp_processor_id());
830}
831
832static void purge_fragmented_blocks_allcpus(void)
833{
834    int cpu;
835
836    for_each_possible_cpu(cpu)
837        purge_fragmented_blocks(cpu);
838}
839
840static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
841{
842    struct vmap_block_queue *vbq;
843    struct vmap_block *vb;
844    unsigned long addr = 0;
845    unsigned int order;
846    int purge = 0;
847
848    BUG_ON(size & ~PAGE_MASK);
849    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
850    order = get_order(size);
851
852again:
853    rcu_read_lock();
854    vbq = &get_cpu_var(vmap_block_queue);
855    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
856        int i;
857
858        spin_lock(&vb->lock);
859        if (vb->free < 1UL << order)
860            goto next;
861
862        i = bitmap_find_free_region(vb->alloc_map,
863                        VMAP_BBMAP_BITS, order);
864
865        if (i < 0) {
866            if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
867                /* fragmented and no outstanding allocations */
868                BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
869                purge = 1;
870            }
871            goto next;
872        }
873        addr = vb->va->va_start + (i << PAGE_SHIFT);
874        BUG_ON(addr_to_vb_idx(addr) !=
875                addr_to_vb_idx(vb->va->va_start));
876        vb->free -= 1UL << order;
877        if (vb->free == 0) {
878            spin_lock(&vbq->lock);
879            list_del_rcu(&vb->free_list);
880            spin_unlock(&vbq->lock);
881        }
882        spin_unlock(&vb->lock);
883        break;
884next:
885        spin_unlock(&vb->lock);
886    }
887
888    if (purge)
889        purge_fragmented_blocks_thiscpu();
890
891    put_cpu_var(vmap_block_queue);
892    rcu_read_unlock();
893
894    if (!addr) {
895        vb = new_vmap_block(gfp_mask);
896        if (IS_ERR(vb))
897            return vb;
898        goto again;
899    }
900
901    return (void *)addr;
902}
903
904static void vb_free(const void *addr, unsigned long size)
905{
906    unsigned long offset;
907    unsigned long vb_idx;
908    unsigned int order;
909    struct vmap_block *vb;
910
911    BUG_ON(size & ~PAGE_MASK);
912    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
913
914    flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
915
916    order = get_order(size);
917
918    offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
919
920    vb_idx = addr_to_vb_idx((unsigned long)addr);
921    rcu_read_lock();
922    vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
923    rcu_read_unlock();
924    BUG_ON(!vb);
925
926    spin_lock(&vb->lock);
927    BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
928
929    vb->dirty += 1UL << order;
930    if (vb->dirty == VMAP_BBMAP_BITS) {
931        BUG_ON(vb->free);
932        spin_unlock(&vb->lock);
933        free_vmap_block(vb);
934    } else
935        spin_unlock(&vb->lock);
936}
937
938/**
939 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
940 *
941 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
942 * to amortize TLB flushing overheads. What this means is that any page you
943 * have now, may, in a former life, have been mapped into kernel virtual
944 * address by the vmap layer and so there might be some CPUs with TLB entries
945 * still referencing that page (additional to the regular 1:1 kernel mapping).
946 *
947 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
948 * be sure that none of the pages we have control over will have any aliases
949 * from the vmap layer.
950 */
951void vm_unmap_aliases(void)
952{
953    unsigned long start = ULONG_MAX, end = 0;
954    int cpu;
955    int flush = 0;
956
957    if (unlikely(!vmap_initialized))
958        return;
959
960    for_each_possible_cpu(cpu) {
961        struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
962        struct vmap_block *vb;
963
964        rcu_read_lock();
965        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
966            int i;
967
968            spin_lock(&vb->lock);
969            i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
970            while (i < VMAP_BBMAP_BITS) {
971                unsigned long s, e;
972                int j;
973                j = find_next_zero_bit(vb->dirty_map,
974                    VMAP_BBMAP_BITS, i);
975
976                s = vb->va->va_start + (i << PAGE_SHIFT);
977                e = vb->va->va_start + (j << PAGE_SHIFT);
978                vunmap_page_range(s, e);
979                flush = 1;
980
981                if (s < start)
982                    start = s;
983                if (e > end)
984                    end = e;
985
986                i = j;
987                i = find_next_bit(vb->dirty_map,
988                            VMAP_BBMAP_BITS, i);
989            }
990            spin_unlock(&vb->lock);
991        }
992        rcu_read_unlock();
993    }
994
995    __purge_vmap_area_lazy(&start, &end, 1, flush);
996}
997EXPORT_SYMBOL_GPL(vm_unmap_aliases);
998
999/**
1000 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1001 * @mem: the pointer returned by vm_map_ram
1002 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1003 */
1004void vm_unmap_ram(const void *mem, unsigned int count)
1005{
1006    unsigned long size = count << PAGE_SHIFT;
1007    unsigned long addr = (unsigned long)mem;
1008
1009    BUG_ON(!addr);
1010    BUG_ON(addr < VMALLOC_START);
1011    BUG_ON(addr > VMALLOC_END);
1012    BUG_ON(addr & (PAGE_SIZE-1));
1013
1014    debug_check_no_locks_freed(mem, size);
1015    vmap_debug_free_range(addr, addr+size);
1016
1017    if (likely(count <= VMAP_MAX_ALLOC))
1018        vb_free(mem, size);
1019    else
1020        free_unmap_vmap_area_addr(addr);
1021}
1022EXPORT_SYMBOL(vm_unmap_ram);
1023
1024/**
1025 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1026 * @pages: an array of pointers to the pages to be mapped
1027 * @count: number of pages
1028 * @node: prefer to allocate data structures on this node
1029 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1030 *
1031 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1032 */
1033void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1034{
1035    unsigned long size = count << PAGE_SHIFT;
1036    unsigned long addr;
1037    void *mem;
1038
1039    if (likely(count <= VMAP_MAX_ALLOC)) {
1040        mem = vb_alloc(size, GFP_KERNEL);
1041        if (IS_ERR(mem))
1042            return NULL;
1043        addr = (unsigned long)mem;
1044    } else {
1045        struct vmap_area *va;
1046        va = alloc_vmap_area(size, PAGE_SIZE,
1047                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1048        if (IS_ERR(va))
1049            return NULL;
1050
1051        addr = va->va_start;
1052        mem = (void *)addr;
1053    }
1054    if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1055        vm_unmap_ram(mem, count);
1056        return NULL;
1057    }
1058    return mem;
1059}
1060EXPORT_SYMBOL(vm_map_ram);
1061
1062/**
1063 * vm_area_register_early - register vmap area early during boot
1064 * @vm: vm_struct to register
1065 * @align: requested alignment
1066 *
1067 * This function is used to register kernel vm area before
1068 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1069 * proper values on entry and other fields should be zero. On return,
1070 * vm->addr contains the allocated address.
1071 *
1072 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1073 */
1074void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1075{
1076    static size_t vm_init_off __initdata;
1077    unsigned long addr;
1078
1079    addr = ALIGN(VMALLOC_START + vm_init_off, align);
1080    vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1081
1082    vm->addr = (void *)addr;
1083
1084    vm->next = vmlist;
1085    vmlist = vm;
1086}
1087
1088void __init vmalloc_init(void)
1089{
1090    struct vmap_area *va;
1091    struct vm_struct *tmp;
1092    int i;
1093
1094    for_each_possible_cpu(i) {
1095        struct vmap_block_queue *vbq;
1096
1097        vbq = &per_cpu(vmap_block_queue, i);
1098        spin_lock_init(&vbq->lock);
1099        INIT_LIST_HEAD(&vbq->free);
1100    }
1101
1102    /* Import existing vmlist entries. */
1103    for (tmp = vmlist; tmp; tmp = tmp->next) {
1104        va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1105        va->flags = tmp->flags | VM_VM_AREA;
1106        va->va_start = (unsigned long)tmp->addr;
1107        va->va_end = va->va_start + tmp->size;
1108        __insert_vmap_area(va);
1109    }
1110
1111    vmap_area_pcpu_hole = VMALLOC_END;
1112
1113    vmap_initialized = true;
1114}
1115
1116/**
1117 * map_kernel_range_noflush - map kernel VM area with the specified pages
1118 * @addr: start of the VM area to map
1119 * @size: size of the VM area to map
1120 * @prot: page protection flags to use
1121 * @pages: pages to map
1122 *
1123 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1124 * specify should have been allocated using get_vm_area() and its
1125 * friends.
1126 *
1127 * NOTE:
1128 * This function does NOT do any cache flushing. The caller is
1129 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1130 * before calling this function.
1131 *
1132 * RETURNS:
1133 * The number of pages mapped on success, -errno on failure.
1134 */
1135int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1136                 pgprot_t prot, struct page **pages)
1137{
1138    return vmap_page_range_noflush(addr, addr + size, prot, pages);
1139}
1140
1141/**
1142 * unmap_kernel_range_noflush - unmap kernel VM area
1143 * @addr: start of the VM area to unmap
1144 * @size: size of the VM area to unmap
1145 *
1146 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1147 * specify should have been allocated using get_vm_area() and its
1148 * friends.
1149 *
1150 * NOTE:
1151 * This function does NOT do any cache flushing. The caller is
1152 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1153 * before calling this function and flush_tlb_kernel_range() after.
1154 */
1155void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1156{
1157    vunmap_page_range(addr, addr + size);
1158}
1159
1160/**
1161 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1162 * @addr: start of the VM area to unmap
1163 * @size: size of the VM area to unmap
1164 *
1165 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1166 * the unmapping and tlb after.
1167 */
1168void unmap_kernel_range(unsigned long addr, unsigned long size)
1169{
1170    unsigned long end = addr + size;
1171
1172    flush_cache_vunmap(addr, end);
1173    vunmap_page_range(addr, end);
1174    flush_tlb_kernel_range(addr, end);
1175}
1176
1177int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1178{
1179    unsigned long addr = (unsigned long)area->addr;
1180    unsigned long end = addr + area->size - PAGE_SIZE;
1181    int err;
1182
1183    err = vmap_page_range(addr, end, prot, *pages);
1184    if (err > 0) {
1185        *pages += err;
1186        err = 0;
1187    }
1188
1189    return err;
1190}
1191EXPORT_SYMBOL_GPL(map_vm_area);
1192
1193/*** Old vmalloc interfaces ***/
1194DEFINE_RWLOCK(vmlist_lock);
1195struct vm_struct *vmlist;
1196
1197static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1198                  unsigned long flags, void *caller)
1199{
1200    struct vm_struct *tmp, **p;
1201
1202    vm->flags = flags;
1203    vm->addr = (void *)va->va_start;
1204    vm->size = va->va_end - va->va_start;
1205    vm->caller = caller;
1206    va->private = vm;
1207    va->flags |= VM_VM_AREA;
1208
1209    write_lock(&vmlist_lock);
1210    for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1211        if (tmp->addr >= vm->addr)
1212            break;
1213    }
1214    vm->next = *p;
1215    *p = vm;
1216    write_unlock(&vmlist_lock);
1217}
1218
1219static struct vm_struct *__get_vm_area_node(unsigned long size,
1220        unsigned long align, unsigned long flags, unsigned long start,
1221        unsigned long end, int node, gfp_t gfp_mask, void *caller)
1222{
1223    static struct vmap_area *va;
1224    struct vm_struct *area;
1225
1226    BUG_ON(in_interrupt());
1227    if (flags & VM_IOREMAP) {
1228        int bit = fls(size);
1229
1230        if (bit > IOREMAP_MAX_ORDER)
1231            bit = IOREMAP_MAX_ORDER;
1232        else if (bit < PAGE_SHIFT)
1233            bit = PAGE_SHIFT;
1234
1235        align = 1ul << bit;
1236    }
1237
1238    size = PAGE_ALIGN(size);
1239    if (unlikely(!size))
1240        return NULL;
1241
1242    area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1243    if (unlikely(!area))
1244        return NULL;
1245
1246    /*
1247     * We always allocate a guard page.
1248     */
1249    size += PAGE_SIZE;
1250
1251    va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1252    if (IS_ERR(va)) {
1253        kfree(area);
1254        return NULL;
1255    }
1256
1257    insert_vmalloc_vm(area, va, flags, caller);
1258    return area;
1259}
1260
1261struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1262                unsigned long start, unsigned long end)
1263{
1264    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1265                        __builtin_return_address(0));
1266}
1267EXPORT_SYMBOL_GPL(__get_vm_area);
1268
1269struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1270                       unsigned long start, unsigned long end,
1271                       void *caller)
1272{
1273    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1274                  caller);
1275}
1276
1277/**
1278 * get_vm_area - reserve a contiguous kernel virtual area
1279 * @size: size of the area
1280 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1281 *
1282 * Search an area of @size in the kernel virtual mapping area,
1283 * and reserved it for out purposes. Returns the area descriptor
1284 * on success or %NULL on failure.
1285 */
1286struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1287{
1288    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1289                -1, GFP_KERNEL, __builtin_return_address(0));
1290}
1291
1292struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1293                void *caller)
1294{
1295    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1296                        -1, GFP_KERNEL, caller);
1297}
1298
1299struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1300                   int node, gfp_t gfp_mask)
1301{
1302    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1303                  node, gfp_mask, __builtin_return_address(0));
1304}
1305
1306static struct vm_struct *find_vm_area(const void *addr)
1307{
1308    struct vmap_area *va;
1309
1310    va = find_vmap_area((unsigned long)addr);
1311    if (va && va->flags & VM_VM_AREA)
1312        return va->private;
1313
1314    return NULL;
1315}
1316
1317/**
1318 * remove_vm_area - find and remove a continuous kernel virtual area
1319 * @addr: base address
1320 *
1321 * Search for the kernel VM area starting at @addr, and remove it.
1322 * This function returns the found VM area, but using it is NOT safe
1323 * on SMP machines, except for its size or flags.
1324 */
1325struct vm_struct *remove_vm_area(const void *addr)
1326{
1327    struct vmap_area *va;
1328
1329    va = find_vmap_area((unsigned long)addr);
1330    if (va && va->flags & VM_VM_AREA) {
1331        struct vm_struct *vm = va->private;
1332        struct vm_struct *tmp, **p;
1333        /*
1334         * remove from list and disallow access to this vm_struct
1335         * before unmap. (address range confliction is maintained by
1336         * vmap.)
1337         */
1338        write_lock(&vmlist_lock);
1339        for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1340            ;
1341        *p = tmp->next;
1342        write_unlock(&vmlist_lock);
1343
1344        vmap_debug_free_range(va->va_start, va->va_end);
1345        free_unmap_vmap_area(va);
1346        vm->size -= PAGE_SIZE;
1347
1348        return vm;
1349    }
1350    return NULL;
1351}
1352
1353static void __vunmap(const void *addr, int deallocate_pages)
1354{
1355    struct vm_struct *area;
1356
1357    if (!addr)
1358        return;
1359
1360    if ((PAGE_SIZE-1) & (unsigned long)addr) {
1361        WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1362        return;
1363    }
1364
1365    area = remove_vm_area(addr);
1366    if (unlikely(!area)) {
1367        WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1368                addr);
1369        return;
1370    }
1371
1372    debug_check_no_locks_freed(addr, area->size);
1373    debug_check_no_obj_freed(addr, area->size);
1374
1375    if (deallocate_pages) {
1376        int i;
1377
1378        for (i = 0; i < area->nr_pages; i++) {
1379            struct page *page = area->pages[i];
1380
1381            BUG_ON(!page);
1382            __free_page(page);
1383        }
1384
1385        if (area->flags & VM_VPAGES)
1386            vfree(area->pages);
1387        else
1388            kfree(area->pages);
1389    }
1390
1391    kfree(area);
1392    return;
1393}
1394
1395/**
1396 * vfree - release memory allocated by vmalloc()
1397 * @addr: memory base address
1398 *
1399 * Free the virtually continuous memory area starting at @addr, as
1400 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1401 * NULL, no operation is performed.
1402 *
1403 * Must not be called in interrupt context.
1404 */
1405void vfree(const void *addr)
1406{
1407    BUG_ON(in_interrupt());
1408
1409    kmemleak_free(addr);
1410
1411    __vunmap(addr, 1);
1412}
1413EXPORT_SYMBOL(vfree);
1414
1415/**
1416 * vunmap - release virtual mapping obtained by vmap()
1417 * @addr: memory base address
1418 *
1419 * Free the virtually contiguous memory area starting at @addr,
1420 * which was created from the page array passed to vmap().
1421 *
1422 * Must not be called in interrupt context.
1423 */
1424void vunmap(const void *addr)
1425{
1426    BUG_ON(in_interrupt());
1427    might_sleep();
1428    __vunmap(addr, 0);
1429}
1430EXPORT_SYMBOL(vunmap);
1431
1432/**
1433 * vmap - map an array of pages into virtually contiguous space
1434 * @pages: array of page pointers
1435 * @count: number of pages to map
1436 * @flags: vm_area->flags
1437 * @prot: page protection for the mapping
1438 *
1439 * Maps @count pages from @pages into contiguous kernel virtual
1440 * space.
1441 */
1442void *vmap(struct page **pages, unsigned int count,
1443        unsigned long flags, pgprot_t prot)
1444{
1445    struct vm_struct *area;
1446
1447    might_sleep();
1448
1449    if (count > totalram_pages)
1450        return NULL;
1451
1452    area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1453                    __builtin_return_address(0));
1454    if (!area)
1455        return NULL;
1456
1457    if (map_vm_area(area, prot, &pages)) {
1458        vunmap(area->addr);
1459        return NULL;
1460    }
1461
1462    return area->addr;
1463}
1464EXPORT_SYMBOL(vmap);
1465
1466static void *__vmalloc_node(unsigned long size, unsigned long align,
1467                gfp_t gfp_mask, pgprot_t prot,
1468                int node, void *caller);
1469static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1470                 pgprot_t prot, int node, void *caller)
1471{
1472    struct page **pages;
1473    unsigned int nr_pages, array_size, i;
1474    gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1475
1476    nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1477    array_size = (nr_pages * sizeof(struct page *));
1478
1479    area->nr_pages = nr_pages;
1480    /* Please note that the recursion is strictly bounded. */
1481    if (array_size > PAGE_SIZE) {
1482        pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1483                PAGE_KERNEL, node, caller);
1484        area->flags |= VM_VPAGES;
1485    } else {
1486        pages = kmalloc_node(array_size, nested_gfp, node);
1487    }
1488    area->pages = pages;
1489    area->caller = caller;
1490    if (!area->pages) {
1491        remove_vm_area(area->addr);
1492        kfree(area);
1493        return NULL;
1494    }
1495
1496    for (i = 0; i < area->nr_pages; i++) {
1497        struct page *page;
1498
1499        if (node < 0)
1500            page = alloc_page(gfp_mask);
1501        else
1502            page = alloc_pages_node(node, gfp_mask, 0);
1503
1504        if (unlikely(!page)) {
1505            /* Successfully allocated i pages, free them in __vunmap() */
1506            area->nr_pages = i;
1507            goto fail;
1508        }
1509        area->pages[i] = page;
1510    }
1511
1512    if (map_vm_area(area, prot, &pages))
1513        goto fail;
1514    return area->addr;
1515
1516fail:
1517    vfree(area->addr);
1518    return NULL;
1519}
1520
1521void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1522{
1523    void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1524                     __builtin_return_address(0));
1525
1526    /*
1527     * A ref_count = 3 is needed because the vm_struct and vmap_area
1528     * structures allocated in the __get_vm_area_node() function contain
1529     * references to the virtual address of the vmalloc'ed block.
1530     */
1531    kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1532
1533    return addr;
1534}
1535
1536/**
1537 * __vmalloc_node - allocate virtually contiguous memory
1538 * @size: allocation size
1539 * @align: desired alignment
1540 * @gfp_mask: flags for the page level allocator
1541 * @prot: protection mask for the allocated pages
1542 * @node: node to use for allocation or -1
1543 * @caller: caller's return address
1544 *
1545 * Allocate enough pages to cover @size from the page level
1546 * allocator with @gfp_mask flags. Map them into contiguous
1547 * kernel virtual space, using a pagetable protection of @prot.
1548 */
1549static void *__vmalloc_node(unsigned long size, unsigned long align,
1550                gfp_t gfp_mask, pgprot_t prot,
1551                int node, void *caller)
1552{
1553    struct vm_struct *area;
1554    void *addr;
1555    unsigned long real_size = size;
1556
1557    size = PAGE_ALIGN(size);
1558    if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1559        return NULL;
1560
1561    area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
1562                  VMALLOC_END, node, gfp_mask, caller);
1563
1564    if (!area)
1565        return NULL;
1566
1567    addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1568
1569    /*
1570     * A ref_count = 3 is needed because the vm_struct and vmap_area
1571     * structures allocated in the __get_vm_area_node() function contain
1572     * references to the virtual address of the vmalloc'ed block.
1573     */
1574    kmemleak_alloc(addr, real_size, 3, gfp_mask);
1575
1576    return addr;
1577}
1578
1579void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1580{
1581    return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1582                __builtin_return_address(0));
1583}
1584EXPORT_SYMBOL(__vmalloc);
1585
1586/**
1587 * vmalloc - allocate virtually contiguous memory
1588 * @size: allocation size
1589 * Allocate enough pages to cover @size from the page level
1590 * allocator and map them into contiguous kernel virtual space.
1591 *
1592 * For tight control over page level allocator and protection flags
1593 * use __vmalloc() instead.
1594 */
1595void *vmalloc(unsigned long size)
1596{
1597    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1598                    -1, __builtin_return_address(0));
1599}
1600EXPORT_SYMBOL(vmalloc);
1601
1602/**
1603 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1604 * @size: allocation size
1605 *
1606 * The resulting memory area is zeroed so it can be mapped to userspace
1607 * without leaking data.
1608 */
1609void *vmalloc_user(unsigned long size)
1610{
1611    struct vm_struct *area;
1612    void *ret;
1613
1614    ret = __vmalloc_node(size, SHMLBA,
1615                 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1616                 PAGE_KERNEL, -1, __builtin_return_address(0));
1617    if (ret) {
1618        area = find_vm_area(ret);
1619        area->flags |= VM_USERMAP;
1620    }
1621    return ret;
1622}
1623EXPORT_SYMBOL(vmalloc_user);
1624
1625/**
1626 * vmalloc_node - allocate memory on a specific node
1627 * @size: allocation size
1628 * @node: numa node
1629 *
1630 * Allocate enough pages to cover @size from the page level
1631 * allocator and map them into contiguous kernel virtual space.
1632 *
1633 * For tight control over page level allocator and protection flags
1634 * use __vmalloc() instead.
1635 */
1636void *vmalloc_node(unsigned long size, int node)
1637{
1638    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1639                    node, __builtin_return_address(0));
1640}
1641EXPORT_SYMBOL(vmalloc_node);
1642
1643#ifndef PAGE_KERNEL_EXEC
1644# define PAGE_KERNEL_EXEC PAGE_KERNEL
1645#endif
1646
1647/**
1648 * vmalloc_exec - allocate virtually contiguous, executable memory
1649 * @size: allocation size
1650 *
1651 * Kernel-internal function to allocate enough pages to cover @size
1652 * the page level allocator and map them into contiguous and
1653 * executable kernel virtual space.
1654 *
1655 * For tight control over page level allocator and protection flags
1656 * use __vmalloc() instead.
1657 */
1658
1659void *vmalloc_exec(unsigned long size)
1660{
1661    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1662                  -1, __builtin_return_address(0));
1663}
1664
1665#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1666#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1667#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1668#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1669#else
1670#define GFP_VMALLOC32 GFP_KERNEL
1671#endif
1672
1673/**
1674 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1675 * @size: allocation size
1676 *
1677 * Allocate enough 32bit PA addressable pages to cover @size from the
1678 * page level allocator and map them into contiguous kernel virtual space.
1679 */
1680void *vmalloc_32(unsigned long size)
1681{
1682    return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1683                  -1, __builtin_return_address(0));
1684}
1685EXPORT_SYMBOL(vmalloc_32);
1686
1687/**
1688 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1689 * @size: allocation size
1690 *
1691 * The resulting memory area is 32bit addressable and zeroed so it can be
1692 * mapped to userspace without leaking data.
1693 */
1694void *vmalloc_32_user(unsigned long size)
1695{
1696    struct vm_struct *area;
1697    void *ret;
1698
1699    ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1700                 -1, __builtin_return_address(0));
1701    if (ret) {
1702        area = find_vm_area(ret);
1703        area->flags |= VM_USERMAP;
1704    }
1705    return ret;
1706}
1707EXPORT_SYMBOL(vmalloc_32_user);
1708
1709/*
1710 * small helper routine , copy contents to buf from addr.
1711 * If the page is not present, fill zero.
1712 */
1713
1714static int aligned_vread(char *buf, char *addr, unsigned long count)
1715{
1716    struct page *p;
1717    int copied = 0;
1718
1719    while (count) {
1720        unsigned long offset, length;
1721
1722        offset = (unsigned long)addr & ~PAGE_MASK;
1723        length = PAGE_SIZE - offset;
1724        if (length > count)
1725            length = count;
1726        p = vmalloc_to_page(addr);
1727        /*
1728         * To do safe access to this _mapped_ area, we need
1729         * lock. But adding lock here means that we need to add
1730         * overhead of vmalloc()/vfree() calles for this _debug_
1731         * interface, rarely used. Instead of that, we'll use
1732         * kmap() and get small overhead in this access function.
1733         */
1734        if (p) {
1735            /*
1736             * we can expect USER0 is not used (see vread/vwrite's
1737             * function description)
1738             */
1739            void *map = kmap_atomic(p, KM_USER0);
1740            memcpy(buf, map + offset, length);
1741            kunmap_atomic(map, KM_USER0);
1742        } else
1743            memset(buf, 0, length);
1744
1745        addr += length;
1746        buf += length;
1747        copied += length;
1748        count -= length;
1749    }
1750    return copied;
1751}
1752
1753static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1754{
1755    struct page *p;
1756    int copied = 0;
1757
1758    while (count) {
1759        unsigned long offset, length;
1760
1761        offset = (unsigned long)addr & ~PAGE_MASK;
1762        length = PAGE_SIZE - offset;
1763        if (length > count)
1764            length = count;
1765        p = vmalloc_to_page(addr);
1766        /*
1767         * To do safe access to this _mapped_ area, we need
1768         * lock. But adding lock here means that we need to add
1769         * overhead of vmalloc()/vfree() calles for this _debug_
1770         * interface, rarely used. Instead of that, we'll use
1771         * kmap() and get small overhead in this access function.
1772         */
1773        if (p) {
1774            /*
1775             * we can expect USER0 is not used (see vread/vwrite's
1776             * function description)
1777             */
1778            void *map = kmap_atomic(p, KM_USER0);
1779            memcpy(map + offset, buf, length);
1780            kunmap_atomic(map, KM_USER0);
1781        }
1782        addr += length;
1783        buf += length;
1784        copied += length;
1785        count -= length;
1786    }
1787    return copied;
1788}
1789
1790/**
1791 * vread() - read vmalloc area in a safe way.
1792 * @buf: buffer for reading data
1793 * @addr: vm address.
1794 * @count: number of bytes to be read.
1795 *
1796 * Returns # of bytes which addr and buf should be increased.
1797 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1798 * includes any intersect with alive vmalloc area.
1799 *
1800 * This function checks that addr is a valid vmalloc'ed area, and
1801 * copy data from that area to a given buffer. If the given memory range
1802 * of [addr...addr+count) includes some valid address, data is copied to
1803 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1804 * IOREMAP area is treated as memory hole and no copy is done.
1805 *
1806 * If [addr...addr+count) doesn't includes any intersects with alive
1807 * vm_struct area, returns 0.
1808 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1809 * the caller should guarantee KM_USER0 is not used.
1810 *
1811 * Note: In usual ops, vread() is never necessary because the caller
1812 * should know vmalloc() area is valid and can use memcpy().
1813 * This is for routines which have to access vmalloc area without
1814 * any informaion, as /dev/kmem.
1815 *
1816 */
1817
1818long vread(char *buf, char *addr, unsigned long count)
1819{
1820    struct vm_struct *tmp;
1821    char *vaddr, *buf_start = buf;
1822    unsigned long buflen = count;
1823    unsigned long n;
1824
1825    /* Don't allow overflow */
1826    if ((unsigned long) addr + count < count)
1827        count = -(unsigned long) addr;
1828
1829    read_lock(&vmlist_lock);
1830    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1831        vaddr = (char *) tmp->addr;
1832        if (addr >= vaddr + tmp->size - PAGE_SIZE)
1833            continue;
1834        while (addr < vaddr) {
1835            if (count == 0)
1836                goto finished;
1837            *buf = '\0';
1838            buf++;
1839            addr++;
1840            count--;
1841        }
1842        n = vaddr + tmp->size - PAGE_SIZE - addr;
1843        if (n > count)
1844            n = count;
1845        if (!(tmp->flags & VM_IOREMAP))
1846            aligned_vread(buf, addr, n);
1847        else /* IOREMAP area is treated as memory hole */
1848            memset(buf, 0, n);
1849        buf += n;
1850        addr += n;
1851        count -= n;
1852    }
1853finished:
1854    read_unlock(&vmlist_lock);
1855
1856    if (buf == buf_start)
1857        return 0;
1858    /* zero-fill memory holes */
1859    if (buf != buf_start + buflen)
1860        memset(buf, 0, buflen - (buf - buf_start));
1861
1862    return buflen;
1863}
1864
1865/**
1866 * vwrite() - write vmalloc area in a safe way.
1867 * @buf: buffer for source data
1868 * @addr: vm address.
1869 * @count: number of bytes to be read.
1870 *
1871 * Returns # of bytes which addr and buf should be incresed.
1872 * (same number to @count).
1873 * If [addr...addr+count) doesn't includes any intersect with valid
1874 * vmalloc area, returns 0.
1875 *
1876 * This function checks that addr is a valid vmalloc'ed area, and
1877 * copy data from a buffer to the given addr. If specified range of
1878 * [addr...addr+count) includes some valid address, data is copied from
1879 * proper area of @buf. If there are memory holes, no copy to hole.
1880 * IOREMAP area is treated as memory hole and no copy is done.
1881 *
1882 * If [addr...addr+count) doesn't includes any intersects with alive
1883 * vm_struct area, returns 0.
1884 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1885 * the caller should guarantee KM_USER0 is not used.
1886 *
1887 * Note: In usual ops, vwrite() is never necessary because the caller
1888 * should know vmalloc() area is valid and can use memcpy().
1889 * This is for routines which have to access vmalloc area without
1890 * any informaion, as /dev/kmem.
1891 *
1892 * The caller should guarantee KM_USER1 is not used.
1893 */
1894
1895long vwrite(char *buf, char *addr, unsigned long count)
1896{
1897    struct vm_struct *tmp;
1898    char *vaddr;
1899    unsigned long n, buflen;
1900    int copied = 0;
1901
1902    /* Don't allow overflow */
1903    if ((unsigned long) addr + count < count)
1904        count = -(unsigned long) addr;
1905    buflen = count;
1906
1907    read_lock(&vmlist_lock);
1908    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1909        vaddr = (char *) tmp->addr;
1910        if (addr >= vaddr + tmp->size - PAGE_SIZE)
1911            continue;
1912        while (addr < vaddr) {
1913            if (count == 0)
1914                goto finished;
1915            buf++;
1916            addr++;
1917            count--;
1918        }
1919        n = vaddr + tmp->size - PAGE_SIZE - addr;
1920        if (n > count)
1921            n = count;
1922        if (!(tmp->flags & VM_IOREMAP)) {
1923            aligned_vwrite(buf, addr, n);
1924            copied++;
1925        }
1926        buf += n;
1927        addr += n;
1928        count -= n;
1929    }
1930finished:
1931    read_unlock(&vmlist_lock);
1932    if (!copied)
1933        return 0;
1934    return buflen;
1935}
1936
1937/**
1938 * remap_vmalloc_range - map vmalloc pages to userspace
1939 * @vma: vma to cover (map full range of vma)
1940 * @addr: vmalloc memory
1941 * @pgoff: number of pages into addr before first page to map
1942 *
1943 * Returns: 0 for success, -Exxx on failure
1944 *
1945 * This function checks that addr is a valid vmalloc'ed area, and
1946 * that it is big enough to cover the vma. Will return failure if
1947 * that criteria isn't met.
1948 *
1949 * Similar to remap_pfn_range() (see mm/memory.c)
1950 */
1951int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1952                        unsigned long pgoff)
1953{
1954    struct vm_struct *area;
1955    unsigned long uaddr = vma->vm_start;
1956    unsigned long usize = vma->vm_end - vma->vm_start;
1957
1958    if ((PAGE_SIZE-1) & (unsigned long)addr)
1959        return -EINVAL;
1960
1961    area = find_vm_area(addr);
1962    if (!area)
1963        return -EINVAL;
1964
1965    if (!(area->flags & VM_USERMAP))
1966        return -EINVAL;
1967
1968    if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1969        return -EINVAL;
1970
1971    addr += pgoff << PAGE_SHIFT;
1972    do {
1973        struct page *page = vmalloc_to_page(addr);
1974        int ret;
1975
1976        ret = vm_insert_page(vma, uaddr, page);
1977        if (ret)
1978            return ret;
1979
1980        uaddr += PAGE_SIZE;
1981        addr += PAGE_SIZE;
1982        usize -= PAGE_SIZE;
1983    } while (usize > 0);
1984
1985    /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1986    vma->vm_flags |= VM_RESERVED;
1987
1988    return 0;
1989}
1990EXPORT_SYMBOL(remap_vmalloc_range);
1991
1992/*
1993 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1994 * have one.
1995 */
1996void __attribute__((weak)) vmalloc_sync_all(void)
1997{
1998}
1999
2000
2001static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2002{
2003    /* apply_to_page_range() does all the hard work. */
2004    return 0;
2005}
2006
2007/**
2008 * alloc_vm_area - allocate a range of kernel address space
2009 * @size: size of the area
2010 *
2011 * Returns: NULL on failure, vm_struct on success
2012 *
2013 * This function reserves a range of kernel address space, and
2014 * allocates pagetables to map that range. No actual mappings
2015 * are created. If the kernel address space is not shared
2016 * between processes, it syncs the pagetable across all
2017 * processes.
2018 */
2019struct vm_struct *alloc_vm_area(size_t size)
2020{
2021    struct vm_struct *area;
2022
2023    area = get_vm_area_caller(size, VM_IOREMAP,
2024                __builtin_return_address(0));
2025    if (area == NULL)
2026        return NULL;
2027
2028    /*
2029     * This ensures that page tables are constructed for this region
2030     * of kernel virtual address space and mapped into init_mm.
2031     */
2032    if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2033                area->size, f, NULL)) {
2034        free_vm_area(area);
2035        return NULL;
2036    }
2037
2038    /* Make sure the pagetables are constructed in process kernel
2039       mappings */
2040    vmalloc_sync_all();
2041
2042    return area;
2043}
2044EXPORT_SYMBOL_GPL(alloc_vm_area);
2045
2046void free_vm_area(struct vm_struct *area)
2047{
2048    struct vm_struct *ret;
2049    ret = remove_vm_area(area->addr);
2050    BUG_ON(ret != area);
2051    kfree(area);
2052}
2053EXPORT_SYMBOL_GPL(free_vm_area);
2054
2055static struct vmap_area *node_to_va(struct rb_node *n)
2056{
2057    return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2058}
2059
2060/**
2061 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2062 * @end: target address
2063 * @pnext: out arg for the next vmap_area
2064 * @pprev: out arg for the previous vmap_area
2065 *
2066 * Returns: %true if either or both of next and prev are found,
2067 * %false if no vmap_area exists
2068 *
2069 * Find vmap_areas end addresses of which enclose @end. ie. if not
2070 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2071 */
2072static bool pvm_find_next_prev(unsigned long end,
2073                   struct vmap_area **pnext,
2074                   struct vmap_area **pprev)
2075{
2076    struct rb_node *n = vmap_area_root.rb_node;
2077    struct vmap_area *va = NULL;
2078
2079    while (n) {
2080        va = rb_entry(n, struct vmap_area, rb_node);
2081        if (end < va->va_end)
2082            n = n->rb_left;
2083        else if (end > va->va_end)
2084            n = n->rb_right;
2085        else
2086            break;
2087    }
2088
2089    if (!va)
2090        return false;
2091
2092    if (va->va_end > end) {
2093        *pnext = va;
2094        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2095    } else {
2096        *pprev = va;
2097        *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2098    }
2099    return true;
2100}
2101
2102/**
2103 * pvm_determine_end - find the highest aligned address between two vmap_areas
2104 * @pnext: in/out arg for the next vmap_area
2105 * @pprev: in/out arg for the previous vmap_area
2106 * @align: alignment
2107 *
2108 * Returns: determined end address
2109 *
2110 * Find the highest aligned address between *@pnext and *@pprev below
2111 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2112 * down address is between the end addresses of the two vmap_areas.
2113 *
2114 * Please note that the address returned by this function may fall
2115 * inside *@pnext vmap_area. The caller is responsible for checking
2116 * that.
2117 */
2118static unsigned long pvm_determine_end(struct vmap_area **pnext,
2119                       struct vmap_area **pprev,
2120                       unsigned long align)
2121{
2122    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2123    unsigned long addr;
2124
2125    if (*pnext)
2126        addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2127    else
2128        addr = vmalloc_end;
2129
2130    while (*pprev && (*pprev)->va_end > addr) {
2131        *pnext = *pprev;
2132        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2133    }
2134
2135    return addr;
2136}
2137
2138/**
2139 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2140 * @offsets: array containing offset of each area
2141 * @sizes: array containing size of each area
2142 * @nr_vms: the number of areas to allocate
2143 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2144 * @gfp_mask: allocation mask
2145 *
2146 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2147 * vm_structs on success, %NULL on failure
2148 *
2149 * Percpu allocator wants to use congruent vm areas so that it can
2150 * maintain the offsets among percpu areas. This function allocates
2151 * congruent vmalloc areas for it. These areas tend to be scattered
2152 * pretty far, distance between two areas easily going up to
2153 * gigabytes. To avoid interacting with regular vmallocs, these areas
2154 * are allocated from top.
2155 *
2156 * Despite its complicated look, this allocator is rather simple. It
2157 * does everything top-down and scans areas from the end looking for
2158 * matching slot. While scanning, if any of the areas overlaps with
2159 * existing vmap_area, the base address is pulled down to fit the
2160 * area. Scanning is repeated till all the areas fit and then all
2161 * necessary data structres are inserted and the result is returned.
2162 */
2163struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2164                     const size_t *sizes, int nr_vms,
2165                     size_t align, gfp_t gfp_mask)
2166{
2167    const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2168    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2169    struct vmap_area **vas, *prev, *next;
2170    struct vm_struct **vms;
2171    int area, area2, last_area, term_area;
2172    unsigned long base, start, end, last_end;
2173    bool purged = false;
2174
2175    gfp_mask &= GFP_RECLAIM_MASK;
2176
2177    /* verify parameters and allocate data structures */
2178    BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2179    for (last_area = 0, area = 0; area < nr_vms; area++) {
2180        start = offsets[area];
2181        end = start + sizes[area];
2182
2183        /* is everything aligned properly? */
2184        BUG_ON(!IS_ALIGNED(offsets[area], align));
2185        BUG_ON(!IS_ALIGNED(sizes[area], align));
2186
2187        /* detect the area with the highest address */
2188        if (start > offsets[last_area])
2189            last_area = area;
2190
2191        for (area2 = 0; area2 < nr_vms; area2++) {
2192            unsigned long start2 = offsets[area2];
2193            unsigned long end2 = start2 + sizes[area2];
2194
2195            if (area2 == area)
2196                continue;
2197
2198            BUG_ON(start2 >= start && start2 < end);
2199            BUG_ON(end2 <= end && end2 > start);
2200        }
2201    }
2202    last_end = offsets[last_area] + sizes[last_area];
2203
2204    if (vmalloc_end - vmalloc_start < last_end) {
2205        WARN_ON(true);
2206        return NULL;
2207    }
2208
2209    vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2210    vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2211    if (!vas || !vms)
2212        goto err_free;
2213
2214    for (area = 0; area < nr_vms; area++) {
2215        vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2216        vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2217        if (!vas[area] || !vms[area])
2218            goto err_free;
2219    }
2220retry:
2221    spin_lock(&vmap_area_lock);
2222
2223    /* start scanning - we scan from the top, begin with the last area */
2224    area = term_area = last_area;
2225    start = offsets[area];
2226    end = start + sizes[area];
2227
2228    if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2229        base = vmalloc_end - last_end;
2230        goto found;
2231    }
2232    base = pvm_determine_end(&next, &prev, align) - end;
2233
2234    while (true) {
2235        BUG_ON(next && next->va_end <= base + end);
2236        BUG_ON(prev && prev->va_end > base + end);
2237
2238        /*
2239         * base might have underflowed, add last_end before
2240         * comparing.
2241         */
2242        if (base + last_end < vmalloc_start + last_end) {
2243            spin_unlock(&vmap_area_lock);
2244            if (!purged) {
2245                purge_vmap_area_lazy();
2246                purged = true;
2247                goto retry;
2248            }
2249            goto err_free;
2250        }
2251
2252        /*
2253         * If next overlaps, move base downwards so that it's
2254         * right below next and then recheck.
2255         */
2256        if (next && next->va_start < base + end) {
2257            base = pvm_determine_end(&next, &prev, align) - end;
2258            term_area = area;
2259            continue;
2260        }
2261
2262        /*
2263         * If prev overlaps, shift down next and prev and move
2264         * base so that it's right below new next and then
2265         * recheck.
2266         */
2267        if (prev && prev->va_end > base + start) {
2268            next = prev;
2269            prev = node_to_va(rb_prev(&next->rb_node));
2270            base = pvm_determine_end(&next, &prev, align) - end;
2271            term_area = area;
2272            continue;
2273        }
2274
2275        /*
2276         * This area fits, move on to the previous one. If
2277         * the previous one is the terminal one, we're done.
2278         */
2279        area = (area + nr_vms - 1) % nr_vms;
2280        if (area == term_area)
2281            break;
2282        start = offsets[area];
2283        end = start + sizes[area];
2284        pvm_find_next_prev(base + end, &next, &prev);
2285    }
2286found:
2287    /* we've found a fitting base, insert all va's */
2288    for (area = 0; area < nr_vms; area++) {
2289        struct vmap_area *va = vas[area];
2290
2291        va->va_start = base + offsets[area];
2292        va->va_end = va->va_start + sizes[area];
2293        __insert_vmap_area(va);
2294    }
2295
2296    vmap_area_pcpu_hole = base + offsets[last_area];
2297
2298    spin_unlock(&vmap_area_lock);
2299
2300    /* insert all vm's */
2301    for (area = 0; area < nr_vms; area++)
2302        insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2303                  pcpu_get_vm_areas);
2304
2305    kfree(vas);
2306    return vms;
2307
2308err_free:
2309    for (area = 0; area < nr_vms; area++) {
2310        if (vas)
2311            kfree(vas[area]);
2312        if (vms)
2313            kfree(vms[area]);
2314    }
2315    kfree(vas);
2316    kfree(vms);
2317    return NULL;
2318}
2319
2320/**
2321 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2322 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2323 * @nr_vms: the number of allocated areas
2324 *
2325 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2326 */
2327void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2328{
2329    int i;
2330
2331    for (i = 0; i < nr_vms; i++)
2332        free_vm_area(vms[i]);
2333    kfree(vms);
2334}
2335
2336#ifdef CONFIG_PROC_FS
2337static void *s_start(struct seq_file *m, loff_t *pos)
2338{
2339    loff_t n = *pos;
2340    struct vm_struct *v;
2341
2342    read_lock(&vmlist_lock);
2343    v = vmlist;
2344    while (n > 0 && v) {
2345        n--;
2346        v = v->next;
2347    }
2348    if (!n)
2349        return v;
2350
2351    return NULL;
2352
2353}
2354
2355static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2356{
2357    struct vm_struct *v = p;
2358
2359    ++*pos;
2360    return v->next;
2361}
2362
2363static void s_stop(struct seq_file *m, void *p)
2364{
2365    read_unlock(&vmlist_lock);
2366}
2367
2368static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2369{
2370    if (NUMA_BUILD) {
2371        unsigned int nr, *counters = m->private;
2372
2373        if (!counters)
2374            return;
2375
2376        memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2377
2378        for (nr = 0; nr < v->nr_pages; nr++)
2379            counters[page_to_nid(v->pages[nr])]++;
2380
2381        for_each_node_state(nr, N_HIGH_MEMORY)
2382            if (counters[nr])
2383                seq_printf(m, " N%u=%u", nr, counters[nr]);
2384    }
2385}
2386
2387static int s_show(struct seq_file *m, void *p)
2388{
2389    struct vm_struct *v = p;
2390
2391    seq_printf(m, "0x%p-0x%p %7ld",
2392        v->addr, v->addr + v->size, v->size);
2393
2394    if (v->caller) {
2395        char buff[KSYM_SYMBOL_LEN];
2396
2397        seq_putc(m, ' ');
2398        sprint_symbol(buff, (unsigned long)v->caller);
2399        seq_puts(m, buff);
2400    }
2401
2402    if (v->nr_pages)
2403        seq_printf(m, " pages=%d", v->nr_pages);
2404
2405    if (v->phys_addr)
2406        seq_printf(m, " phys=%lx", v->phys_addr);
2407
2408    if (v->flags & VM_IOREMAP)
2409        seq_printf(m, " ioremap");
2410
2411    if (v->flags & VM_ALLOC)
2412        seq_printf(m, " vmalloc");
2413
2414    if (v->flags & VM_MAP)
2415        seq_printf(m, " vmap");
2416
2417    if (v->flags & VM_USERMAP)
2418        seq_printf(m, " user");
2419
2420    if (v->flags & VM_VPAGES)
2421        seq_printf(m, " vpages");
2422
2423    show_numa_info(m, v);
2424    seq_putc(m, '\n');
2425    return 0;
2426}
2427
2428static const struct seq_operations vmalloc_op = {
2429    .start = s_start,
2430    .next = s_next,
2431    .stop = s_stop,
2432    .show = s_show,
2433};
2434
2435static int vmalloc_open(struct inode *inode, struct file *file)
2436{
2437    unsigned int *ptr = NULL;
2438    int ret;
2439
2440    if (NUMA_BUILD)
2441        ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2442    ret = seq_open(file, &vmalloc_op);
2443    if (!ret) {
2444        struct seq_file *m = file->private_data;
2445        m->private = ptr;
2446    } else
2447        kfree(ptr);
2448    return ret;
2449}
2450
2451static const struct file_operations proc_vmalloc_operations = {
2452    .open = vmalloc_open,
2453    .read = seq_read,
2454    .llseek = seq_lseek,
2455    .release = seq_release_private,
2456};
2457
2458static int __init proc_vmalloc_init(void)
2459{
2460    proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2461    return 0;
2462}
2463module_init(proc_vmalloc_init);
2464#endif
2465
2466

Archive Download this file



interactive