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 <linux/atomic.h>
30#include <asm/uaccess.h>
31#include <asm/tlbflush.h>
32#include <asm/shmparam.h>
33
34/*** Page table manipulation functions ***/
35
36static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
37{
38    pte_t *pte;
39
40    pte = pte_offset_kernel(pmd, addr);
41    do {
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);
45}
46
47static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
48{
49    pmd_t *pmd;
50    unsigned long next;
51
52    pmd = pmd_offset(pud, addr);
53    do {
54        next = pmd_addr_end(addr, end);
55        if (pmd_none_or_clear_bad(pmd))
56            continue;
57        vunmap_pte_range(pmd, addr, next);
58    } while (pmd++, addr = next, addr != end);
59}
60
61static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
62{
63    pud_t *pud;
64    unsigned long next;
65
66    pud = pud_offset(pgd, addr);
67    do {
68        next = pud_addr_end(addr, end);
69        if (pud_none_or_clear_bad(pud))
70            continue;
71        vunmap_pmd_range(pud, addr, next);
72    } while (pud++, addr = next, addr != end);
73}
74
75static void vunmap_page_range(unsigned long addr, unsigned long end)
76{
77    pgd_t *pgd;
78    unsigned long next;
79
80    BUG_ON(addr >= end);
81    pgd = pgd_offset_k(addr);
82    do {
83        next = pgd_addr_end(addr, end);
84        if (pgd_none_or_clear_bad(pgd))
85            continue;
86        vunmap_pud_range(pgd, addr, next);
87    } while (pgd++, addr = next, addr != end);
88}
89
90static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
91        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
92{
93    pte_t *pte;
94
95    /*
96     * nr is a running index into the array which helps higher level
97     * callers keep track of where we're up to.
98     */
99
100    pte = pte_alloc_kernel(pmd, addr);
101    if (!pte)
102        return -ENOMEM;
103    do {
104        struct page *page = pages[*nr];
105
106        if (WARN_ON(!pte_none(*pte)))
107            return -EBUSY;
108        if (WARN_ON(!page))
109            return -ENOMEM;
110        set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
111        (*nr)++;
112    } while (pte++, addr += PAGE_SIZE, addr != end);
113    return 0;
114}
115
116static int vmap_pmd_range(pud_t *pud, unsigned long addr,
117        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
118{
119    pmd_t *pmd;
120    unsigned long next;
121
122    pmd = pmd_alloc(&init_mm, pud, addr);
123    if (!pmd)
124        return -ENOMEM;
125    do {
126        next = pmd_addr_end(addr, end);
127        if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
128            return -ENOMEM;
129    } while (pmd++, addr = next, addr != end);
130    return 0;
131}
132
133static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
134        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
135{
136    pud_t *pud;
137    unsigned long next;
138
139    pud = pud_alloc(&init_mm, pgd, addr);
140    if (!pud)
141        return -ENOMEM;
142    do {
143        next = pud_addr_end(addr, end);
144        if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
145            return -ENOMEM;
146    } while (pud++, addr = next, addr != end);
147    return 0;
148}
149
150/*
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.
153 *
154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
155 */
156static int vmap_page_range_noflush(unsigned long start, unsigned long end,
157                   pgprot_t prot, struct page **pages)
158{
159    pgd_t *pgd;
160    unsigned long next;
161    unsigned long addr = start;
162    int err = 0;
163    int nr = 0;
164
165    BUG_ON(addr >= end);
166    pgd = pgd_offset_k(addr);
167    do {
168        next = pgd_addr_end(addr, end);
169        err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
170        if (err)
171            return err;
172    } while (pgd++, addr = next, addr != end);
173
174    return nr;
175}
176
177static int vmap_page_range(unsigned long start, unsigned long end,
178               pgprot_t prot, struct page **pages)
179{
180    int ret;
181
182    ret = vmap_page_range_noflush(start, end, prot, pages);
183    flush_cache_vmap(start, end);
184    return ret;
185}
186
187int is_vmalloc_or_module_addr(const void *x)
188{
189    /*
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.
193     */
194#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
195    unsigned long addr = (unsigned long)x;
196    if (addr >= MODULES_VADDR && addr < MODULES_END)
197        return 1;
198#endif
199    return is_vmalloc_addr(x);
200}
201
202/*
203 * Walk a vmap address to the struct page it maps.
204 */
205struct page *vmalloc_to_page(const void *vmalloc_addr)
206{
207    unsigned long addr = (unsigned long) vmalloc_addr;
208    struct page *page = NULL;
209    pgd_t *pgd = pgd_offset_k(addr);
210
211    /*
212     * XXX we might need to change this if we add VIRTUAL_BUG_ON for
213     * architectures that do not vmalloc module space
214     */
215    VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
216
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)) {
222                pte_t *ptep, pte;
223
224                ptep = pte_offset_map(pmd, addr);
225                pte = *ptep;
226                if (pte_present(pte))
227                    page = pte_page(pte);
228                pte_unmap(ptep);
229            }
230        }
231    }
232    return page;
233}
234EXPORT_SYMBOL(vmalloc_to_page);
235
236/*
237 * Map a vmalloc()-space virtual address to the physical page frame number.
238 */
239unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
240{
241    return page_to_pfn(vmalloc_to_page(vmalloc_addr));
242}
243EXPORT_SYMBOL(vmalloc_to_pfn);
244
245
246/*** Global kva allocator ***/
247
248#define VM_LAZY_FREE 0x01
249#define VM_LAZY_FREEING 0x02
250#define VM_VM_AREA 0x04
251
252struct vmap_area {
253    unsigned long va_start;
254    unsigned long va_end;
255    unsigned long flags;
256    struct rb_node rb_node; /* address sorted rbtree */
257    struct list_head list; /* address sorted list */
258    struct list_head purge_list; /* "lazy purge" list */
259    struct vm_struct *vm;
260    struct rcu_head rcu_head;
261};
262
263static DEFINE_SPINLOCK(vmap_area_lock);
264static LIST_HEAD(vmap_area_list);
265static struct rb_root vmap_area_root = RB_ROOT;
266
267/* The vmap cache globals are protected by vmap_area_lock */
268static struct rb_node *free_vmap_cache;
269static unsigned long cached_hole_size;
270static unsigned long cached_vstart;
271static unsigned long cached_align;
272
273static unsigned long vmap_area_pcpu_hole;
274
275static struct vmap_area *__find_vmap_area(unsigned long addr)
276{
277    struct rb_node *n = vmap_area_root.rb_node;
278
279    while (n) {
280        struct vmap_area *va;
281
282        va = rb_entry(n, struct vmap_area, rb_node);
283        if (addr < va->va_start)
284            n = n->rb_left;
285        else if (addr > va->va_start)
286            n = n->rb_right;
287        else
288            return va;
289    }
290
291    return NULL;
292}
293
294static void __insert_vmap_area(struct vmap_area *va)
295{
296    struct rb_node **p = &vmap_area_root.rb_node;
297    struct rb_node *parent = NULL;
298    struct rb_node *tmp;
299
300    while (*p) {
301        struct vmap_area *tmp_va;
302
303        parent = *p;
304        tmp_va = rb_entry(parent, struct vmap_area, rb_node);
305        if (va->va_start < tmp_va->va_end)
306            p = &(*p)->rb_left;
307        else if (va->va_end > tmp_va->va_start)
308            p = &(*p)->rb_right;
309        else
310            BUG();
311    }
312
313    rb_link_node(&va->rb_node, parent, p);
314    rb_insert_color(&va->rb_node, &vmap_area_root);
315
316    /* address-sort this list so it is usable like the vmlist */
317    tmp = rb_prev(&va->rb_node);
318    if (tmp) {
319        struct vmap_area *prev;
320        prev = rb_entry(tmp, struct vmap_area, rb_node);
321        list_add_rcu(&va->list, &prev->list);
322    } else
323        list_add_rcu(&va->list, &vmap_area_list);
324}
325
326static void purge_vmap_area_lazy(void);
327
328/*
329 * Allocate a region of KVA of the specified size and alignment, within the
330 * vstart and vend.
331 */
332static struct vmap_area *alloc_vmap_area(unsigned long size,
333                unsigned long align,
334                unsigned long vstart, unsigned long vend,
335                int node, gfp_t gfp_mask)
336{
337    struct vmap_area *va;
338    struct rb_node *n;
339    unsigned long addr;
340    int purged = 0;
341    struct vmap_area *first;
342
343    BUG_ON(!size);
344    BUG_ON(size & ~PAGE_MASK);
345    BUG_ON(!is_power_of_2(align));
346
347    va = kmalloc_node(sizeof(struct vmap_area),
348            gfp_mask & GFP_RECLAIM_MASK, node);
349    if (unlikely(!va))
350        return ERR_PTR(-ENOMEM);
351
352retry:
353    spin_lock(&vmap_area_lock);
354    /*
355     * Invalidate cache if we have more permissive parameters.
356     * cached_hole_size notes the largest hole noticed _below_
357     * the vmap_area cached in free_vmap_cache: if size fits
358     * into that hole, we want to scan from vstart to reuse
359     * the hole instead of allocating above free_vmap_cache.
360     * Note that __free_vmap_area may update free_vmap_cache
361     * without updating cached_hole_size or cached_align.
362     */
363    if (!free_vmap_cache ||
364            size < cached_hole_size ||
365            vstart < cached_vstart ||
366            align < cached_align) {
367nocache:
368        cached_hole_size = 0;
369        free_vmap_cache = NULL;
370    }
371    /* record if we encounter less permissive parameters */
372    cached_vstart = vstart;
373    cached_align = align;
374
375    /* find starting point for our search */
376    if (free_vmap_cache) {
377        first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
378        addr = ALIGN(first->va_end, align);
379        if (addr < vstart)
380            goto nocache;
381        if (addr + size - 1 < addr)
382            goto overflow;
383
384    } else {
385        addr = ALIGN(vstart, align);
386        if (addr + size - 1 < addr)
387            goto overflow;
388
389        n = vmap_area_root.rb_node;
390        first = NULL;
391
392        while (n) {
393            struct vmap_area *tmp;
394            tmp = rb_entry(n, struct vmap_area, rb_node);
395            if (tmp->va_end >= addr) {
396                first = tmp;
397                if (tmp->va_start <= addr)
398                    break;
399                n = n->rb_left;
400            } else
401                n = n->rb_right;
402        }
403
404        if (!first)
405            goto found;
406    }
407
408    /* from the starting point, walk areas until a suitable hole is found */
409    while (addr + size > first->va_start && addr + size <= vend) {
410        if (addr + cached_hole_size < first->va_start)
411            cached_hole_size = first->va_start - addr;
412        addr = ALIGN(first->va_end, align);
413        if (addr + size - 1 < addr)
414            goto overflow;
415
416        n = rb_next(&first->rb_node);
417        if (n)
418            first = rb_entry(n, struct vmap_area, rb_node);
419        else
420            goto found;
421    }
422
423found:
424    if (addr + size > vend)
425        goto overflow;
426
427    va->va_start = addr;
428    va->va_end = addr + size;
429    va->flags = 0;
430    __insert_vmap_area(va);
431    free_vmap_cache = &va->rb_node;
432    spin_unlock(&vmap_area_lock);
433
434    BUG_ON(va->va_start & (align-1));
435    BUG_ON(va->va_start < vstart);
436    BUG_ON(va->va_end > vend);
437
438    return va;
439
440overflow:
441    spin_unlock(&vmap_area_lock);
442    if (!purged) {
443        purge_vmap_area_lazy();
444        purged = 1;
445        goto retry;
446    }
447    if (printk_ratelimit())
448        printk(KERN_WARNING
449            "vmap allocation for size %lu failed: "
450            "use vmalloc=<size> to increase size.\n", size);
451    kfree(va);
452    return ERR_PTR(-EBUSY);
453}
454
455static void __free_vmap_area(struct vmap_area *va)
456{
457    BUG_ON(RB_EMPTY_NODE(&va->rb_node));
458
459    if (free_vmap_cache) {
460        if (va->va_end < cached_vstart) {
461            free_vmap_cache = NULL;
462        } else {
463            struct vmap_area *cache;
464            cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
465            if (va->va_start <= cache->va_start) {
466                free_vmap_cache = rb_prev(&va->rb_node);
467                /*
468                 * We don't try to update cached_hole_size or
469                 * cached_align, but it won't go very wrong.
470                 */
471            }
472        }
473    }
474    rb_erase(&va->rb_node, &vmap_area_root);
475    RB_CLEAR_NODE(&va->rb_node);
476    list_del_rcu(&va->list);
477
478    /*
479     * Track the highest possible candidate for pcpu area
480     * allocation. Areas outside of vmalloc area can be returned
481     * here too, consider only end addresses which fall inside
482     * vmalloc area proper.
483     */
484    if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
485        vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
486
487    kfree_rcu(va, rcu_head);
488}
489
490/*
491 * Free a region of KVA allocated by alloc_vmap_area
492 */
493static void free_vmap_area(struct vmap_area *va)
494{
495    spin_lock(&vmap_area_lock);
496    __free_vmap_area(va);
497    spin_unlock(&vmap_area_lock);
498}
499
500/*
501 * Clear the pagetable entries of a given vmap_area
502 */
503static void unmap_vmap_area(struct vmap_area *va)
504{
505    vunmap_page_range(va->va_start, va->va_end);
506}
507
508static void vmap_debug_free_range(unsigned long start, unsigned long end)
509{
510    /*
511     * Unmap page tables and force a TLB flush immediately if
512     * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
513     * bugs similarly to those in linear kernel virtual address
514     * space after a page has been freed.
515     *
516     * All the lazy freeing logic is still retained, in order to
517     * minimise intrusiveness of this debugging feature.
518     *
519     * This is going to be *slow* (linear kernel virtual address
520     * debugging doesn't do a broadcast TLB flush so it is a lot
521     * faster).
522     */
523#ifdef CONFIG_DEBUG_PAGEALLOC
524    vunmap_page_range(start, end);
525    flush_tlb_kernel_range(start, end);
526#endif
527}
528
529/*
530 * lazy_max_pages is the maximum amount of virtual address space we gather up
531 * before attempting to purge with a TLB flush.
532 *
533 * There is a tradeoff here: a larger number will cover more kernel page tables
534 * and take slightly longer to purge, but it will linearly reduce the number of
535 * global TLB flushes that must be performed. It would seem natural to scale
536 * this number up linearly with the number of CPUs (because vmapping activity
537 * could also scale linearly with the number of CPUs), however it is likely
538 * that in practice, workloads might be constrained in other ways that mean
539 * vmap activity will not scale linearly with CPUs. Also, I want to be
540 * conservative and not introduce a big latency on huge systems, so go with
541 * a less aggressive log scale. It will still be an improvement over the old
542 * code, and it will be simple to change the scale factor if we find that it
543 * becomes a problem on bigger systems.
544 */
545static unsigned long lazy_max_pages(void)
546{
547    unsigned int log;
548
549    log = fls(num_online_cpus());
550
551    return log * (32UL * 1024 * 1024 / PAGE_SIZE);
552}
553
554static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
555
556/* for per-CPU blocks */
557static void purge_fragmented_blocks_allcpus(void);
558
559/*
560 * called before a call to iounmap() if the caller wants vm_area_struct's
561 * immediately freed.
562 */
563void set_iounmap_nonlazy(void)
564{
565    atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
566}
567
568/*
569 * Purges all lazily-freed vmap areas.
570 *
571 * If sync is 0 then don't purge if there is already a purge in progress.
572 * If force_flush is 1, then flush kernel TLBs between *start and *end even
573 * if we found no lazy vmap areas to unmap (callers can use this to optimise
574 * their own TLB flushing).
575 * Returns with *start = min(*start, lowest purged address)
576 * *end = max(*end, highest purged address)
577 */
578static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
579                    int sync, int force_flush)
580{
581    static DEFINE_SPINLOCK(purge_lock);
582    LIST_HEAD(valist);
583    struct vmap_area *va;
584    struct vmap_area *n_va;
585    int nr = 0;
586
587    /*
588     * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
589     * should not expect such behaviour. This just simplifies locking for
590     * the case that isn't actually used at the moment anyway.
591     */
592    if (!sync && !force_flush) {
593        if (!spin_trylock(&purge_lock))
594            return;
595    } else
596        spin_lock(&purge_lock);
597
598    if (sync)
599        purge_fragmented_blocks_allcpus();
600
601    rcu_read_lock();
602    list_for_each_entry_rcu(va, &vmap_area_list, list) {
603        if (va->flags & VM_LAZY_FREE) {
604            if (va->va_start < *start)
605                *start = va->va_start;
606            if (va->va_end > *end)
607                *end = va->va_end;
608            nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
609            list_add_tail(&va->purge_list, &valist);
610            va->flags |= VM_LAZY_FREEING;
611            va->flags &= ~VM_LAZY_FREE;
612        }
613    }
614    rcu_read_unlock();
615
616    if (nr)
617        atomic_sub(nr, &vmap_lazy_nr);
618
619    if (nr || force_flush)
620        flush_tlb_kernel_range(*start, *end);
621
622    if (nr) {
623        spin_lock(&vmap_area_lock);
624        list_for_each_entry_safe(va, n_va, &valist, purge_list)
625            __free_vmap_area(va);
626        spin_unlock(&vmap_area_lock);
627    }
628    spin_unlock(&purge_lock);
629}
630
631/*
632 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
633 * is already purging.
634 */
635static void try_purge_vmap_area_lazy(void)
636{
637    unsigned long start = ULONG_MAX, end = 0;
638
639    __purge_vmap_area_lazy(&start, &end, 0, 0);
640}
641
642/*
643 * Kick off a purge of the outstanding lazy areas.
644 */
645static void purge_vmap_area_lazy(void)
646{
647    unsigned long start = ULONG_MAX, end = 0;
648
649    __purge_vmap_area_lazy(&start, &end, 1, 0);
650}
651
652/*
653 * Free a vmap area, caller ensuring that the area has been unmapped
654 * and flush_cache_vunmap had been called for the correct range
655 * previously.
656 */
657static void free_vmap_area_noflush(struct vmap_area *va)
658{
659    va->flags |= VM_LAZY_FREE;
660    atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
661    if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
662        try_purge_vmap_area_lazy();
663}
664
665/*
666 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
667 * called for the correct range previously.
668 */
669static void free_unmap_vmap_area_noflush(struct vmap_area *va)
670{
671    unmap_vmap_area(va);
672    free_vmap_area_noflush(va);
673}
674
675/*
676 * Free and unmap a vmap area
677 */
678static void free_unmap_vmap_area(struct vmap_area *va)
679{
680    flush_cache_vunmap(va->va_start, va->va_end);
681    free_unmap_vmap_area_noflush(va);
682}
683
684static struct vmap_area *find_vmap_area(unsigned long addr)
685{
686    struct vmap_area *va;
687
688    spin_lock(&vmap_area_lock);
689    va = __find_vmap_area(addr);
690    spin_unlock(&vmap_area_lock);
691
692    return va;
693}
694
695static void free_unmap_vmap_area_addr(unsigned long addr)
696{
697    struct vmap_area *va;
698
699    va = find_vmap_area(addr);
700    BUG_ON(!va);
701    free_unmap_vmap_area(va);
702}
703
704
705/*** Per cpu kva allocator ***/
706
707/*
708 * vmap space is limited especially on 32 bit architectures. Ensure there is
709 * room for at least 16 percpu vmap blocks per CPU.
710 */
711/*
712 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
713 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
714 * instead (we just need a rough idea)
715 */
716#if BITS_PER_LONG == 32
717#define VMALLOC_SPACE (128UL*1024*1024)
718#else
719#define VMALLOC_SPACE (128UL*1024*1024*1024)
720#endif
721
722#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
723#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
724#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
725#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
726#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
727#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
728#define VMAP_BBMAP_BITS \
729        VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
730        VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
731            VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
732
733#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
734
735static bool vmap_initialized __read_mostly = false;
736
737struct vmap_block_queue {
738    spinlock_t lock;
739    struct list_head free;
740};
741
742struct vmap_block {
743    spinlock_t lock;
744    struct vmap_area *va;
745    struct vmap_block_queue *vbq;
746    unsigned long free, dirty;
747    DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
748    DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
749    struct list_head free_list;
750    struct rcu_head rcu_head;
751    struct list_head purge;
752};
753
754/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
755static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
756
757/*
758 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
759 * in the free path. Could get rid of this if we change the API to return a
760 * "cookie" from alloc, to be passed to free. But no big deal yet.
761 */
762static DEFINE_SPINLOCK(vmap_block_tree_lock);
763static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
764
765/*
766 * We should probably have a fallback mechanism to allocate virtual memory
767 * out of partially filled vmap blocks. However vmap block sizing should be
768 * fairly reasonable according to the vmalloc size, so it shouldn't be a
769 * big problem.
770 */
771
772static unsigned long addr_to_vb_idx(unsigned long addr)
773{
774    addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
775    addr /= VMAP_BLOCK_SIZE;
776    return addr;
777}
778
779static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
780{
781    struct vmap_block_queue *vbq;
782    struct vmap_block *vb;
783    struct vmap_area *va;
784    unsigned long vb_idx;
785    int node, err;
786
787    node = numa_node_id();
788
789    vb = kmalloc_node(sizeof(struct vmap_block),
790            gfp_mask & GFP_RECLAIM_MASK, node);
791    if (unlikely(!vb))
792        return ERR_PTR(-ENOMEM);
793
794    va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
795                    VMALLOC_START, VMALLOC_END,
796                    node, gfp_mask);
797    if (IS_ERR(va)) {
798        kfree(vb);
799        return ERR_CAST(va);
800    }
801
802    err = radix_tree_preload(gfp_mask);
803    if (unlikely(err)) {
804        kfree(vb);
805        free_vmap_area(va);
806        return ERR_PTR(err);
807    }
808
809    spin_lock_init(&vb->lock);
810    vb->va = va;
811    vb->free = VMAP_BBMAP_BITS;
812    vb->dirty = 0;
813    bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
814    bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
815    INIT_LIST_HEAD(&vb->free_list);
816
817    vb_idx = addr_to_vb_idx(va->va_start);
818    spin_lock(&vmap_block_tree_lock);
819    err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
820    spin_unlock(&vmap_block_tree_lock);
821    BUG_ON(err);
822    radix_tree_preload_end();
823
824    vbq = &get_cpu_var(vmap_block_queue);
825    vb->vbq = vbq;
826    spin_lock(&vbq->lock);
827    list_add_rcu(&vb->free_list, &vbq->free);
828    spin_unlock(&vbq->lock);
829    put_cpu_var(vmap_block_queue);
830
831    return vb;
832}
833
834static void free_vmap_block(struct vmap_block *vb)
835{
836    struct vmap_block *tmp;
837    unsigned long vb_idx;
838
839    vb_idx = addr_to_vb_idx(vb->va->va_start);
840    spin_lock(&vmap_block_tree_lock);
841    tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
842    spin_unlock(&vmap_block_tree_lock);
843    BUG_ON(tmp != vb);
844
845    free_vmap_area_noflush(vb->va);
846    kfree_rcu(vb, rcu_head);
847}
848
849static void purge_fragmented_blocks(int cpu)
850{
851    LIST_HEAD(purge);
852    struct vmap_block *vb;
853    struct vmap_block *n_vb;
854    struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
855
856    rcu_read_lock();
857    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
858
859        if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
860            continue;
861
862        spin_lock(&vb->lock);
863        if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
864            vb->free = 0; /* prevent further allocs after releasing lock */
865            vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
866            bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
867            bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
868            spin_lock(&vbq->lock);
869            list_del_rcu(&vb->free_list);
870            spin_unlock(&vbq->lock);
871            spin_unlock(&vb->lock);
872            list_add_tail(&vb->purge, &purge);
873        } else
874            spin_unlock(&vb->lock);
875    }
876    rcu_read_unlock();
877
878    list_for_each_entry_safe(vb, n_vb, &purge, purge) {
879        list_del(&vb->purge);
880        free_vmap_block(vb);
881    }
882}
883
884static void purge_fragmented_blocks_thiscpu(void)
885{
886    purge_fragmented_blocks(smp_processor_id());
887}
888
889static void purge_fragmented_blocks_allcpus(void)
890{
891    int cpu;
892
893    for_each_possible_cpu(cpu)
894        purge_fragmented_blocks(cpu);
895}
896
897static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
898{
899    struct vmap_block_queue *vbq;
900    struct vmap_block *vb;
901    unsigned long addr = 0;
902    unsigned int order;
903    int purge = 0;
904
905    BUG_ON(size & ~PAGE_MASK);
906    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
907    order = get_order(size);
908
909again:
910    rcu_read_lock();
911    vbq = &get_cpu_var(vmap_block_queue);
912    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
913        int i;
914
915        spin_lock(&vb->lock);
916        if (vb->free < 1UL << order)
917            goto next;
918
919        i = bitmap_find_free_region(vb->alloc_map,
920                        VMAP_BBMAP_BITS, order);
921
922        if (i < 0) {
923            if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
924                /* fragmented and no outstanding allocations */
925                BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
926                purge = 1;
927            }
928            goto next;
929        }
930        addr = vb->va->va_start + (i << PAGE_SHIFT);
931        BUG_ON(addr_to_vb_idx(addr) !=
932                addr_to_vb_idx(vb->va->va_start));
933        vb->free -= 1UL << order;
934        if (vb->free == 0) {
935            spin_lock(&vbq->lock);
936            list_del_rcu(&vb->free_list);
937            spin_unlock(&vbq->lock);
938        }
939        spin_unlock(&vb->lock);
940        break;
941next:
942        spin_unlock(&vb->lock);
943    }
944
945    if (purge)
946        purge_fragmented_blocks_thiscpu();
947
948    put_cpu_var(vmap_block_queue);
949    rcu_read_unlock();
950
951    if (!addr) {
952        vb = new_vmap_block(gfp_mask);
953        if (IS_ERR(vb))
954            return vb;
955        goto again;
956    }
957
958    return (void *)addr;
959}
960
961static void vb_free(const void *addr, unsigned long size)
962{
963    unsigned long offset;
964    unsigned long vb_idx;
965    unsigned int order;
966    struct vmap_block *vb;
967
968    BUG_ON(size & ~PAGE_MASK);
969    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
970
971    flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
972
973    order = get_order(size);
974
975    offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
976
977    vb_idx = addr_to_vb_idx((unsigned long)addr);
978    rcu_read_lock();
979    vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
980    rcu_read_unlock();
981    BUG_ON(!vb);
982
983    vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
984
985    spin_lock(&vb->lock);
986    BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
987
988    vb->dirty += 1UL << order;
989    if (vb->dirty == VMAP_BBMAP_BITS) {
990        BUG_ON(vb->free);
991        spin_unlock(&vb->lock);
992        free_vmap_block(vb);
993    } else
994        spin_unlock(&vb->lock);
995}
996
997/**
998 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
999 *
1000 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1001 * to amortize TLB flushing overheads. What this means is that any page you
1002 * have now, may, in a former life, have been mapped into kernel virtual
1003 * address by the vmap layer and so there might be some CPUs with TLB entries
1004 * still referencing that page (additional to the regular 1:1 kernel mapping).
1005 *
1006 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1007 * be sure that none of the pages we have control over will have any aliases
1008 * from the vmap layer.
1009 */
1010void vm_unmap_aliases(void)
1011{
1012    unsigned long start = ULONG_MAX, end = 0;
1013    int cpu;
1014    int flush = 0;
1015
1016    if (unlikely(!vmap_initialized))
1017        return;
1018
1019    for_each_possible_cpu(cpu) {
1020        struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1021        struct vmap_block *vb;
1022
1023        rcu_read_lock();
1024        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1025            int i;
1026
1027            spin_lock(&vb->lock);
1028            i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1029            while (i < VMAP_BBMAP_BITS) {
1030                unsigned long s, e;
1031                int j;
1032                j = find_next_zero_bit(vb->dirty_map,
1033                    VMAP_BBMAP_BITS, i);
1034
1035                s = vb->va->va_start + (i << PAGE_SHIFT);
1036                e = vb->va->va_start + (j << PAGE_SHIFT);
1037                flush = 1;
1038
1039                if (s < start)
1040                    start = s;
1041                if (e > end)
1042                    end = e;
1043
1044                i = j;
1045                i = find_next_bit(vb->dirty_map,
1046                            VMAP_BBMAP_BITS, i);
1047            }
1048            spin_unlock(&vb->lock);
1049        }
1050        rcu_read_unlock();
1051    }
1052
1053    __purge_vmap_area_lazy(&start, &end, 1, flush);
1054}
1055EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1056
1057/**
1058 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1059 * @mem: the pointer returned by vm_map_ram
1060 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1061 */
1062void vm_unmap_ram(const void *mem, unsigned int count)
1063{
1064    unsigned long size = count << PAGE_SHIFT;
1065    unsigned long addr = (unsigned long)mem;
1066
1067    BUG_ON(!addr);
1068    BUG_ON(addr < VMALLOC_START);
1069    BUG_ON(addr > VMALLOC_END);
1070    BUG_ON(addr & (PAGE_SIZE-1));
1071
1072    debug_check_no_locks_freed(mem, size);
1073    vmap_debug_free_range(addr, addr+size);
1074
1075    if (likely(count <= VMAP_MAX_ALLOC))
1076        vb_free(mem, size);
1077    else
1078        free_unmap_vmap_area_addr(addr);
1079}
1080EXPORT_SYMBOL(vm_unmap_ram);
1081
1082/**
1083 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1084 * @pages: an array of pointers to the pages to be mapped
1085 * @count: number of pages
1086 * @node: prefer to allocate data structures on this node
1087 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1088 *
1089 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1090 */
1091void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1092{
1093    unsigned long size = count << PAGE_SHIFT;
1094    unsigned long addr;
1095    void *mem;
1096
1097    if (likely(count <= VMAP_MAX_ALLOC)) {
1098        mem = vb_alloc(size, GFP_KERNEL);
1099        if (IS_ERR(mem))
1100            return NULL;
1101        addr = (unsigned long)mem;
1102    } else {
1103        struct vmap_area *va;
1104        va = alloc_vmap_area(size, PAGE_SIZE,
1105                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1106        if (IS_ERR(va))
1107            return NULL;
1108
1109        addr = va->va_start;
1110        mem = (void *)addr;
1111    }
1112    if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1113        vm_unmap_ram(mem, count);
1114        return NULL;
1115    }
1116    return mem;
1117}
1118EXPORT_SYMBOL(vm_map_ram);
1119
1120/**
1121 * vm_area_add_early - add vmap area early during boot
1122 * @vm: vm_struct to add
1123 *
1124 * This function is used to add fixed kernel vm area to vmlist before
1125 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1126 * should contain proper values and the other fields should be zero.
1127 *
1128 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1129 */
1130void __init vm_area_add_early(struct vm_struct *vm)
1131{
1132    struct vm_struct *tmp, **p;
1133
1134    BUG_ON(vmap_initialized);
1135    for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1136        if (tmp->addr >= vm->addr) {
1137            BUG_ON(tmp->addr < vm->addr + vm->size);
1138            break;
1139        } else
1140            BUG_ON(tmp->addr + tmp->size > vm->addr);
1141    }
1142    vm->next = *p;
1143    *p = vm;
1144}
1145
1146/**
1147 * vm_area_register_early - register vmap area early during boot
1148 * @vm: vm_struct to register
1149 * @align: requested alignment
1150 *
1151 * This function is used to register kernel vm area before
1152 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1153 * proper values on entry and other fields should be zero. On return,
1154 * vm->addr contains the allocated address.
1155 *
1156 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1157 */
1158void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1159{
1160    static size_t vm_init_off __initdata;
1161    unsigned long addr;
1162
1163    addr = ALIGN(VMALLOC_START + vm_init_off, align);
1164    vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1165
1166    vm->addr = (void *)addr;
1167
1168    vm_area_add_early(vm);
1169}
1170
1171void __init vmalloc_init(void)
1172{
1173    struct vmap_area *va;
1174    struct vm_struct *tmp;
1175    int i;
1176
1177    for_each_possible_cpu(i) {
1178        struct vmap_block_queue *vbq;
1179
1180        vbq = &per_cpu(vmap_block_queue, i);
1181        spin_lock_init(&vbq->lock);
1182        INIT_LIST_HEAD(&vbq->free);
1183    }
1184
1185    /* Import existing vmlist entries. */
1186    for (tmp = vmlist; tmp; tmp = tmp->next) {
1187        va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1188        va->flags = tmp->flags | VM_VM_AREA;
1189        va->va_start = (unsigned long)tmp->addr;
1190        va->va_end = va->va_start + tmp->size;
1191        __insert_vmap_area(va);
1192    }
1193
1194    vmap_area_pcpu_hole = VMALLOC_END;
1195
1196    vmap_initialized = true;
1197}
1198
1199/**
1200 * map_kernel_range_noflush - map kernel VM area with the specified pages
1201 * @addr: start of the VM area to map
1202 * @size: size of the VM area to map
1203 * @prot: page protection flags to use
1204 * @pages: pages to map
1205 *
1206 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1207 * specify should have been allocated using get_vm_area() and its
1208 * friends.
1209 *
1210 * NOTE:
1211 * This function does NOT do any cache flushing. The caller is
1212 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1213 * before calling this function.
1214 *
1215 * RETURNS:
1216 * The number of pages mapped on success, -errno on failure.
1217 */
1218int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1219                 pgprot_t prot, struct page **pages)
1220{
1221    return vmap_page_range_noflush(addr, addr + size, prot, pages);
1222}
1223
1224/**
1225 * unmap_kernel_range_noflush - unmap kernel VM area
1226 * @addr: start of the VM area to unmap
1227 * @size: size of the VM area to unmap
1228 *
1229 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1230 * specify should have been allocated using get_vm_area() and its
1231 * friends.
1232 *
1233 * NOTE:
1234 * This function does NOT do any cache flushing. The caller is
1235 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1236 * before calling this function and flush_tlb_kernel_range() after.
1237 */
1238void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1239{
1240    vunmap_page_range(addr, addr + size);
1241}
1242EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1243
1244/**
1245 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1246 * @addr: start of the VM area to unmap
1247 * @size: size of the VM area to unmap
1248 *
1249 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1250 * the unmapping and tlb after.
1251 */
1252void unmap_kernel_range(unsigned long addr, unsigned long size)
1253{
1254    unsigned long end = addr + size;
1255
1256    flush_cache_vunmap(addr, end);
1257    vunmap_page_range(addr, end);
1258    flush_tlb_kernel_range(addr, end);
1259}
1260
1261int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1262{
1263    unsigned long addr = (unsigned long)area->addr;
1264    unsigned long end = addr + area->size - PAGE_SIZE;
1265    int err;
1266
1267    err = vmap_page_range(addr, end, prot, *pages);
1268    if (err > 0) {
1269        *pages += err;
1270        err = 0;
1271    }
1272
1273    return err;
1274}
1275EXPORT_SYMBOL_GPL(map_vm_area);
1276
1277/*** Old vmalloc interfaces ***/
1278DEFINE_RWLOCK(vmlist_lock);
1279struct vm_struct *vmlist;
1280
1281static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1282                  unsigned long flags, void *caller)
1283{
1284    vm->flags = flags;
1285    vm->addr = (void *)va->va_start;
1286    vm->size = va->va_end - va->va_start;
1287    vm->caller = caller;
1288    va->vm = vm;
1289    va->flags |= VM_VM_AREA;
1290}
1291
1292static void insert_vmalloc_vmlist(struct vm_struct *vm)
1293{
1294    struct vm_struct *tmp, **p;
1295
1296    vm->flags &= ~VM_UNLIST;
1297    write_lock(&vmlist_lock);
1298    for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1299        if (tmp->addr >= vm->addr)
1300            break;
1301    }
1302    vm->next = *p;
1303    *p = vm;
1304    write_unlock(&vmlist_lock);
1305}
1306
1307static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1308                  unsigned long flags, void *caller)
1309{
1310    setup_vmalloc_vm(vm, va, flags, caller);
1311    insert_vmalloc_vmlist(vm);
1312}
1313
1314static struct vm_struct *__get_vm_area_node(unsigned long size,
1315        unsigned long align, unsigned long flags, unsigned long start,
1316        unsigned long end, int node, gfp_t gfp_mask, void *caller)
1317{
1318    struct vmap_area *va;
1319    struct vm_struct *area;
1320
1321    BUG_ON(in_interrupt());
1322    if (flags & VM_IOREMAP) {
1323        int bit = fls(size);
1324
1325        if (bit > IOREMAP_MAX_ORDER)
1326            bit = IOREMAP_MAX_ORDER;
1327        else if (bit < PAGE_SHIFT)
1328            bit = PAGE_SHIFT;
1329
1330        align = 1ul << bit;
1331    }
1332
1333    size = PAGE_ALIGN(size);
1334    if (unlikely(!size))
1335        return NULL;
1336
1337    area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1338    if (unlikely(!area))
1339        return NULL;
1340
1341    /*
1342     * We always allocate a guard page.
1343     */
1344    size += PAGE_SIZE;
1345
1346    va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1347    if (IS_ERR(va)) {
1348        kfree(area);
1349        return NULL;
1350    }
1351
1352    /*
1353     * When this function is called from __vmalloc_node_range,
1354     * we do not add vm_struct to vmlist here to avoid
1355     * accessing uninitialized members of vm_struct such as
1356     * pages and nr_pages fields. They will be set later.
1357     * To distinguish it from others, we use a VM_UNLIST flag.
1358     */
1359    if (flags & VM_UNLIST)
1360        setup_vmalloc_vm(area, va, flags, caller);
1361    else
1362        insert_vmalloc_vm(area, va, flags, caller);
1363
1364    return area;
1365}
1366
1367struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1368                unsigned long start, unsigned long end)
1369{
1370    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1371                        __builtin_return_address(0));
1372}
1373EXPORT_SYMBOL_GPL(__get_vm_area);
1374
1375struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1376                       unsigned long start, unsigned long end,
1377                       void *caller)
1378{
1379    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1380                  caller);
1381}
1382
1383/**
1384 * get_vm_area - reserve a contiguous kernel virtual area
1385 * @size: size of the area
1386 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1387 *
1388 * Search an area of @size in the kernel virtual mapping area,
1389 * and reserved it for out purposes. Returns the area descriptor
1390 * on success or %NULL on failure.
1391 */
1392struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1393{
1394    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1395                -1, GFP_KERNEL, __builtin_return_address(0));
1396}
1397
1398struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1399                void *caller)
1400{
1401    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1402                        -1, GFP_KERNEL, caller);
1403}
1404
1405static struct vm_struct *find_vm_area(const void *addr)
1406{
1407    struct vmap_area *va;
1408
1409    va = find_vmap_area((unsigned long)addr);
1410    if (va && va->flags & VM_VM_AREA)
1411        return va->vm;
1412
1413    return NULL;
1414}
1415
1416/**
1417 * remove_vm_area - find and remove a continuous kernel virtual area
1418 * @addr: base address
1419 *
1420 * Search for the kernel VM area starting at @addr, and remove it.
1421 * This function returns the found VM area, but using it is NOT safe
1422 * on SMP machines, except for its size or flags.
1423 */
1424struct vm_struct *remove_vm_area(const void *addr)
1425{
1426    struct vmap_area *va;
1427
1428    va = find_vmap_area((unsigned long)addr);
1429    if (va && va->flags & VM_VM_AREA) {
1430        struct vm_struct *vm = va->vm;
1431
1432        if (!(vm->flags & VM_UNLIST)) {
1433            struct vm_struct *tmp, **p;
1434            /*
1435             * remove from list and disallow access to
1436             * this vm_struct before unmap. (address range
1437             * confliction is maintained by vmap.)
1438             */
1439            write_lock(&vmlist_lock);
1440            for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1441                ;
1442            *p = tmp->next;
1443            write_unlock(&vmlist_lock);
1444        }
1445
1446        vmap_debug_free_range(va->va_start, va->va_end);
1447        free_unmap_vmap_area(va);
1448        vm->size -= PAGE_SIZE;
1449
1450        return vm;
1451    }
1452    return NULL;
1453}
1454
1455static void __vunmap(const void *addr, int deallocate_pages)
1456{
1457    struct vm_struct *area;
1458
1459    if (!addr)
1460        return;
1461
1462    if ((PAGE_SIZE-1) & (unsigned long)addr) {
1463        WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1464        return;
1465    }
1466
1467    area = remove_vm_area(addr);
1468    if (unlikely(!area)) {
1469        WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1470                addr);
1471        return;
1472    }
1473
1474    debug_check_no_locks_freed(addr, area->size);
1475    debug_check_no_obj_freed(addr, area->size);
1476
1477    if (deallocate_pages) {
1478        int i;
1479
1480        for (i = 0; i < area->nr_pages; i++) {
1481            struct page *page = area->pages[i];
1482
1483            BUG_ON(!page);
1484            __free_page(page);
1485        }
1486
1487        if (area->flags & VM_VPAGES)
1488            vfree(area->pages);
1489        else
1490            kfree(area->pages);
1491    }
1492
1493    kfree(area);
1494    return;
1495}
1496
1497/**
1498 * vfree - release memory allocated by vmalloc()
1499 * @addr: memory base address
1500 *
1501 * Free the virtually continuous memory area starting at @addr, as
1502 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1503 * NULL, no operation is performed.
1504 *
1505 * Must not be called in interrupt context.
1506 */
1507void vfree(const void *addr)
1508{
1509    BUG_ON(in_interrupt());
1510
1511    kmemleak_free(addr);
1512
1513    __vunmap(addr, 1);
1514}
1515EXPORT_SYMBOL(vfree);
1516
1517/**
1518 * vunmap - release virtual mapping obtained by vmap()
1519 * @addr: memory base address
1520 *
1521 * Free the virtually contiguous memory area starting at @addr,
1522 * which was created from the page array passed to vmap().
1523 *
1524 * Must not be called in interrupt context.
1525 */
1526void vunmap(const void *addr)
1527{
1528    BUG_ON(in_interrupt());
1529    might_sleep();
1530    __vunmap(addr, 0);
1531}
1532EXPORT_SYMBOL(vunmap);
1533
1534/**
1535 * vmap - map an array of pages into virtually contiguous space
1536 * @pages: array of page pointers
1537 * @count: number of pages to map
1538 * @flags: vm_area->flags
1539 * @prot: page protection for the mapping
1540 *
1541 * Maps @count pages from @pages into contiguous kernel virtual
1542 * space.
1543 */
1544void *vmap(struct page **pages, unsigned int count,
1545        unsigned long flags, pgprot_t prot)
1546{
1547    struct vm_struct *area;
1548
1549    might_sleep();
1550
1551    if (count > totalram_pages)
1552        return NULL;
1553
1554    area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1555                    __builtin_return_address(0));
1556    if (!area)
1557        return NULL;
1558
1559    if (map_vm_area(area, prot, &pages)) {
1560        vunmap(area->addr);
1561        return NULL;
1562    }
1563
1564    return area->addr;
1565}
1566EXPORT_SYMBOL(vmap);
1567
1568static void *__vmalloc_node(unsigned long size, unsigned long align,
1569                gfp_t gfp_mask, pgprot_t prot,
1570                int node, void *caller);
1571static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1572                 pgprot_t prot, int node, void *caller)
1573{
1574    const int order = 0;
1575    struct page **pages;
1576    unsigned int nr_pages, array_size, i;
1577    gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1578
1579    nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1580    array_size = (nr_pages * sizeof(struct page *));
1581
1582    area->nr_pages = nr_pages;
1583    /* Please note that the recursion is strictly bounded. */
1584    if (array_size > PAGE_SIZE) {
1585        pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1586                PAGE_KERNEL, node, caller);
1587        area->flags |= VM_VPAGES;
1588    } else {
1589        pages = kmalloc_node(array_size, nested_gfp, node);
1590    }
1591    area->pages = pages;
1592    area->caller = caller;
1593    if (!area->pages) {
1594        remove_vm_area(area->addr);
1595        kfree(area);
1596        return NULL;
1597    }
1598
1599    for (i = 0; i < area->nr_pages; i++) {
1600        struct page *page;
1601        gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1602
1603        if (node < 0)
1604            page = alloc_page(tmp_mask);
1605        else
1606            page = alloc_pages_node(node, tmp_mask, order);
1607
1608        if (unlikely(!page)) {
1609            /* Successfully allocated i pages, free them in __vunmap() */
1610            area->nr_pages = i;
1611            goto fail;
1612        }
1613        area->pages[i] = page;
1614    }
1615
1616    if (map_vm_area(area, prot, &pages))
1617        goto fail;
1618    return area->addr;
1619
1620fail:
1621    warn_alloc_failed(gfp_mask, order,
1622              "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1623              (area->nr_pages*PAGE_SIZE), area->size);
1624    vfree(area->addr);
1625    return NULL;
1626}
1627
1628/**
1629 * __vmalloc_node_range - allocate virtually contiguous memory
1630 * @size: allocation size
1631 * @align: desired alignment
1632 * @start: vm area range start
1633 * @end: vm area range end
1634 * @gfp_mask: flags for the page level allocator
1635 * @prot: protection mask for the allocated pages
1636 * @node: node to use for allocation or -1
1637 * @caller: caller's return address
1638 *
1639 * Allocate enough pages to cover @size from the page level
1640 * allocator with @gfp_mask flags. Map them into contiguous
1641 * kernel virtual space, using a pagetable protection of @prot.
1642 */
1643void *__vmalloc_node_range(unsigned long size, unsigned long align,
1644            unsigned long start, unsigned long end, gfp_t gfp_mask,
1645            pgprot_t prot, int node, void *caller)
1646{
1647    struct vm_struct *area;
1648    void *addr;
1649    unsigned long real_size = size;
1650
1651    size = PAGE_ALIGN(size);
1652    if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1653        goto fail;
1654
1655    area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1656                  start, end, node, gfp_mask, caller);
1657    if (!area)
1658        goto fail;
1659
1660    addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1661    if (!addr)
1662        return NULL;
1663
1664    /*
1665     * In this function, newly allocated vm_struct is not added
1666     * to vmlist at __get_vm_area_node(). so, it is added here.
1667     */
1668    insert_vmalloc_vmlist(area);
1669
1670    /*
1671     * A ref_count = 3 is needed because the vm_struct and vmap_area
1672     * structures allocated in the __get_vm_area_node() function contain
1673     * references to the virtual address of the vmalloc'ed block.
1674     */
1675    kmemleak_alloc(addr, real_size, 3, gfp_mask);
1676
1677    return addr;
1678
1679fail:
1680    warn_alloc_failed(gfp_mask, 0,
1681              "vmalloc: allocation failure: %lu bytes\n",
1682              real_size);
1683    return NULL;
1684}
1685
1686/**
1687 * __vmalloc_node - allocate virtually contiguous memory
1688 * @size: allocation size
1689 * @align: desired alignment
1690 * @gfp_mask: flags for the page level allocator
1691 * @prot: protection mask for the allocated pages
1692 * @node: node to use for allocation or -1
1693 * @caller: caller's return address
1694 *
1695 * Allocate enough pages to cover @size from the page level
1696 * allocator with @gfp_mask flags. Map them into contiguous
1697 * kernel virtual space, using a pagetable protection of @prot.
1698 */
1699static void *__vmalloc_node(unsigned long size, unsigned long align,
1700                gfp_t gfp_mask, pgprot_t prot,
1701                int node, void *caller)
1702{
1703    return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1704                gfp_mask, prot, node, caller);
1705}
1706
1707void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1708{
1709    return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1710                __builtin_return_address(0));
1711}
1712EXPORT_SYMBOL(__vmalloc);
1713
1714static inline void *__vmalloc_node_flags(unsigned long size,
1715                    int node, gfp_t flags)
1716{
1717    return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1718                    node, __builtin_return_address(0));
1719}
1720
1721/**
1722 * vmalloc - allocate virtually contiguous memory
1723 * @size: allocation size
1724 * Allocate enough pages to cover @size from the page level
1725 * allocator and map them into contiguous kernel virtual space.
1726 *
1727 * For tight control over page level allocator and protection flags
1728 * use __vmalloc() instead.
1729 */
1730void *vmalloc(unsigned long size)
1731{
1732    return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1733}
1734EXPORT_SYMBOL(vmalloc);
1735
1736/**
1737 * vzalloc - allocate virtually contiguous memory with zero fill
1738 * @size: allocation size
1739 * Allocate enough pages to cover @size from the page level
1740 * allocator and map them into contiguous kernel virtual space.
1741 * The memory allocated is set to zero.
1742 *
1743 * For tight control over page level allocator and protection flags
1744 * use __vmalloc() instead.
1745 */
1746void *vzalloc(unsigned long size)
1747{
1748    return __vmalloc_node_flags(size, -1,
1749                GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1750}
1751EXPORT_SYMBOL(vzalloc);
1752
1753/**
1754 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1755 * @size: allocation size
1756 *
1757 * The resulting memory area is zeroed so it can be mapped to userspace
1758 * without leaking data.
1759 */
1760void *vmalloc_user(unsigned long size)
1761{
1762    struct vm_struct *area;
1763    void *ret;
1764
1765    ret = __vmalloc_node(size, SHMLBA,
1766                 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1767                 PAGE_KERNEL, -1, __builtin_return_address(0));
1768    if (ret) {
1769        area = find_vm_area(ret);
1770        area->flags |= VM_USERMAP;
1771    }
1772    return ret;
1773}
1774EXPORT_SYMBOL(vmalloc_user);
1775
1776/**
1777 * vmalloc_node - allocate memory on a specific node
1778 * @size: allocation size
1779 * @node: numa node
1780 *
1781 * Allocate enough pages to cover @size from the page level
1782 * allocator and map them into contiguous kernel virtual space.
1783 *
1784 * For tight control over page level allocator and protection flags
1785 * use __vmalloc() instead.
1786 */
1787void *vmalloc_node(unsigned long size, int node)
1788{
1789    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1790                    node, __builtin_return_address(0));
1791}
1792EXPORT_SYMBOL(vmalloc_node);
1793
1794/**
1795 * vzalloc_node - allocate memory on a specific node with zero fill
1796 * @size: allocation size
1797 * @node: numa node
1798 *
1799 * Allocate enough pages to cover @size from the page level
1800 * allocator and map them into contiguous kernel virtual space.
1801 * The memory allocated is set to zero.
1802 *
1803 * For tight control over page level allocator and protection flags
1804 * use __vmalloc_node() instead.
1805 */
1806void *vzalloc_node(unsigned long size, int node)
1807{
1808    return __vmalloc_node_flags(size, node,
1809             GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1810}
1811EXPORT_SYMBOL(vzalloc_node);
1812
1813#ifndef PAGE_KERNEL_EXEC
1814# define PAGE_KERNEL_EXEC PAGE_KERNEL
1815#endif
1816
1817/**
1818 * vmalloc_exec - allocate virtually contiguous, executable memory
1819 * @size: allocation size
1820 *
1821 * Kernel-internal function to allocate enough pages to cover @size
1822 * the page level allocator and map them into contiguous and
1823 * executable kernel virtual space.
1824 *
1825 * For tight control over page level allocator and protection flags
1826 * use __vmalloc() instead.
1827 */
1828
1829void *vmalloc_exec(unsigned long size)
1830{
1831    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1832                  -1, __builtin_return_address(0));
1833}
1834
1835#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1836#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1837#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1838#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1839#else
1840#define GFP_VMALLOC32 GFP_KERNEL
1841#endif
1842
1843/**
1844 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1845 * @size: allocation size
1846 *
1847 * Allocate enough 32bit PA addressable pages to cover @size from the
1848 * page level allocator and map them into contiguous kernel virtual space.
1849 */
1850void *vmalloc_32(unsigned long size)
1851{
1852    return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1853                  -1, __builtin_return_address(0));
1854}
1855EXPORT_SYMBOL(vmalloc_32);
1856
1857/**
1858 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1859 * @size: allocation size
1860 *
1861 * The resulting memory area is 32bit addressable and zeroed so it can be
1862 * mapped to userspace without leaking data.
1863 */
1864void *vmalloc_32_user(unsigned long size)
1865{
1866    struct vm_struct *area;
1867    void *ret;
1868
1869    ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1870                 -1, __builtin_return_address(0));
1871    if (ret) {
1872        area = find_vm_area(ret);
1873        area->flags |= VM_USERMAP;
1874    }
1875    return ret;
1876}
1877EXPORT_SYMBOL(vmalloc_32_user);
1878
1879/*
1880 * small helper routine , copy contents to buf from addr.
1881 * If the page is not present, fill zero.
1882 */
1883
1884static int aligned_vread(char *buf, char *addr, unsigned long count)
1885{
1886    struct page *p;
1887    int copied = 0;
1888
1889    while (count) {
1890        unsigned long offset, length;
1891
1892        offset = (unsigned long)addr & ~PAGE_MASK;
1893        length = PAGE_SIZE - offset;
1894        if (length > count)
1895            length = count;
1896        p = vmalloc_to_page(addr);
1897        /*
1898         * To do safe access to this _mapped_ area, we need
1899         * lock. But adding lock here means that we need to add
1900         * overhead of vmalloc()/vfree() calles for this _debug_
1901         * interface, rarely used. Instead of that, we'll use
1902         * kmap() and get small overhead in this access function.
1903         */
1904        if (p) {
1905            /*
1906             * we can expect USER0 is not used (see vread/vwrite's
1907             * function description)
1908             */
1909            void *map = kmap_atomic(p);
1910            memcpy(buf, map + offset, length);
1911            kunmap_atomic(map);
1912        } else
1913            memset(buf, 0, length);
1914
1915        addr += length;
1916        buf += length;
1917        copied += length;
1918        count -= length;
1919    }
1920    return copied;
1921}
1922
1923static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1924{
1925    struct page *p;
1926    int copied = 0;
1927
1928    while (count) {
1929        unsigned long offset, length;
1930
1931        offset = (unsigned long)addr & ~PAGE_MASK;
1932        length = PAGE_SIZE - offset;
1933        if (length > count)
1934            length = count;
1935        p = vmalloc_to_page(addr);
1936        /*
1937         * To do safe access to this _mapped_ area, we need
1938         * lock. But adding lock here means that we need to add
1939         * overhead of vmalloc()/vfree() calles for this _debug_
1940         * interface, rarely used. Instead of that, we'll use
1941         * kmap() and get small overhead in this access function.
1942         */
1943        if (p) {
1944            /*
1945             * we can expect USER0 is not used (see vread/vwrite's
1946             * function description)
1947             */
1948            void *map = kmap_atomic(p);
1949            memcpy(map + offset, buf, length);
1950            kunmap_atomic(map);
1951        }
1952        addr += length;
1953        buf += length;
1954        copied += length;
1955        count -= length;
1956    }
1957    return copied;
1958}
1959
1960/**
1961 * vread() - read vmalloc area in a safe way.
1962 * @buf: buffer for reading data
1963 * @addr: vm address.
1964 * @count: number of bytes to be read.
1965 *
1966 * Returns # of bytes which addr and buf should be increased.
1967 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1968 * includes any intersect with alive vmalloc area.
1969 *
1970 * This function checks that addr is a valid vmalloc'ed area, and
1971 * copy data from that area to a given buffer. If the given memory range
1972 * of [addr...addr+count) includes some valid address, data is copied to
1973 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1974 * IOREMAP area is treated as memory hole and no copy is done.
1975 *
1976 * If [addr...addr+count) doesn't includes any intersects with alive
1977 * vm_struct area, returns 0.
1978 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1979 * the caller should guarantee KM_USER0 is not used.
1980 *
1981 * Note: In usual ops, vread() is never necessary because the caller
1982 * should know vmalloc() area is valid and can use memcpy().
1983 * This is for routines which have to access vmalloc area without
1984 * any informaion, as /dev/kmem.
1985 *
1986 */
1987
1988long vread(char *buf, char *addr, unsigned long count)
1989{
1990    struct vm_struct *tmp;
1991    char *vaddr, *buf_start = buf;
1992    unsigned long buflen = count;
1993    unsigned long n;
1994
1995    /* Don't allow overflow */
1996    if ((unsigned long) addr + count < count)
1997        count = -(unsigned long) addr;
1998
1999    read_lock(&vmlist_lock);
2000    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2001        vaddr = (char *) tmp->addr;
2002        if (addr >= vaddr + tmp->size - PAGE_SIZE)
2003            continue;
2004        while (addr < vaddr) {
2005            if (count == 0)
2006                goto finished;
2007            *buf = '\0';
2008            buf++;
2009            addr++;
2010            count--;
2011        }
2012        n = vaddr + tmp->size - PAGE_SIZE - addr;
2013        if (n > count)
2014            n = count;
2015        if (!(tmp->flags & VM_IOREMAP))
2016            aligned_vread(buf, addr, n);
2017        else /* IOREMAP area is treated as memory hole */
2018            memset(buf, 0, n);
2019        buf += n;
2020        addr += n;
2021        count -= n;
2022    }
2023finished:
2024    read_unlock(&vmlist_lock);
2025
2026    if (buf == buf_start)
2027        return 0;
2028    /* zero-fill memory holes */
2029    if (buf != buf_start + buflen)
2030        memset(buf, 0, buflen - (buf - buf_start));
2031
2032    return buflen;
2033}
2034
2035/**
2036 * vwrite() - write vmalloc area in a safe way.
2037 * @buf: buffer for source data
2038 * @addr: vm address.
2039 * @count: number of bytes to be read.
2040 *
2041 * Returns # of bytes which addr and buf should be incresed.
2042 * (same number to @count).
2043 * If [addr...addr+count) doesn't includes any intersect with valid
2044 * vmalloc area, returns 0.
2045 *
2046 * This function checks that addr is a valid vmalloc'ed area, and
2047 * copy data from a buffer to the given addr. If specified range of
2048 * [addr...addr+count) includes some valid address, data is copied from
2049 * proper area of @buf. If there are memory holes, no copy to hole.
2050 * IOREMAP area is treated as memory hole and no copy is done.
2051 *
2052 * If [addr...addr+count) doesn't includes any intersects with alive
2053 * vm_struct area, returns 0.
2054 * @buf should be kernel's buffer. Because this function uses KM_USER0,
2055 * the caller should guarantee KM_USER0 is not used.
2056 *
2057 * Note: In usual ops, vwrite() is never necessary because the caller
2058 * should know vmalloc() area is valid and can use memcpy().
2059 * This is for routines which have to access vmalloc area without
2060 * any informaion, as /dev/kmem.
2061 */
2062
2063long vwrite(char *buf, char *addr, unsigned long count)
2064{
2065    struct vm_struct *tmp;
2066    char *vaddr;
2067    unsigned long n, buflen;
2068    int copied = 0;
2069
2070    /* Don't allow overflow */
2071    if ((unsigned long) addr + count < count)
2072        count = -(unsigned long) addr;
2073    buflen = count;
2074
2075    read_lock(&vmlist_lock);
2076    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2077        vaddr = (char *) tmp->addr;
2078        if (addr >= vaddr + tmp->size - PAGE_SIZE)
2079            continue;
2080        while (addr < vaddr) {
2081            if (count == 0)
2082                goto finished;
2083            buf++;
2084            addr++;
2085            count--;
2086        }
2087        n = vaddr + tmp->size - PAGE_SIZE - addr;
2088        if (n > count)
2089            n = count;
2090        if (!(tmp->flags & VM_IOREMAP)) {
2091            aligned_vwrite(buf, addr, n);
2092            copied++;
2093        }
2094        buf += n;
2095        addr += n;
2096        count -= n;
2097    }
2098finished:
2099    read_unlock(&vmlist_lock);
2100    if (!copied)
2101        return 0;
2102    return buflen;
2103}
2104
2105/**
2106 * remap_vmalloc_range - map vmalloc pages to userspace
2107 * @vma: vma to cover (map full range of vma)
2108 * @addr: vmalloc memory
2109 * @pgoff: number of pages into addr before first page to map
2110 *
2111 * Returns: 0 for success, -Exxx on failure
2112 *
2113 * This function checks that addr is a valid vmalloc'ed area, and
2114 * that it is big enough to cover the vma. Will return failure if
2115 * that criteria isn't met.
2116 *
2117 * Similar to remap_pfn_range() (see mm/memory.c)
2118 */
2119int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2120                        unsigned long pgoff)
2121{
2122    struct vm_struct *area;
2123    unsigned long uaddr = vma->vm_start;
2124    unsigned long usize = vma->vm_end - vma->vm_start;
2125
2126    if ((PAGE_SIZE-1) & (unsigned long)addr)
2127        return -EINVAL;
2128
2129    area = find_vm_area(addr);
2130    if (!area)
2131        return -EINVAL;
2132
2133    if (!(area->flags & VM_USERMAP))
2134        return -EINVAL;
2135
2136    if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2137        return -EINVAL;
2138
2139    addr += pgoff << PAGE_SHIFT;
2140    do {
2141        struct page *page = vmalloc_to_page(addr);
2142        int ret;
2143
2144        ret = vm_insert_page(vma, uaddr, page);
2145        if (ret)
2146            return ret;
2147
2148        uaddr += PAGE_SIZE;
2149        addr += PAGE_SIZE;
2150        usize -= PAGE_SIZE;
2151    } while (usize > 0);
2152
2153    /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2154    vma->vm_flags |= VM_RESERVED;
2155
2156    return 0;
2157}
2158EXPORT_SYMBOL(remap_vmalloc_range);
2159
2160/*
2161 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2162 * have one.
2163 */
2164void __attribute__((weak)) vmalloc_sync_all(void)
2165{
2166}
2167
2168
2169static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2170{
2171    pte_t ***p = data;
2172
2173    if (p) {
2174        *(*p) = pte;
2175        (*p)++;
2176    }
2177    return 0;
2178}
2179
2180/**
2181 * alloc_vm_area - allocate a range of kernel address space
2182 * @size: size of the area
2183 * @ptes: returns the PTEs for the address space
2184 *
2185 * Returns: NULL on failure, vm_struct on success
2186 *
2187 * This function reserves a range of kernel address space, and
2188 * allocates pagetables to map that range. No actual mappings
2189 * are created.
2190 *
2191 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2192 * allocated for the VM area are returned.
2193 */
2194struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2195{
2196    struct vm_struct *area;
2197
2198    area = get_vm_area_caller(size, VM_IOREMAP,
2199                __builtin_return_address(0));
2200    if (area == NULL)
2201        return NULL;
2202
2203    /*
2204     * This ensures that page tables are constructed for this region
2205     * of kernel virtual address space and mapped into init_mm.
2206     */
2207    if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2208                size, f, ptes ? &ptes : NULL)) {
2209        free_vm_area(area);
2210        return NULL;
2211    }
2212
2213    return area;
2214}
2215EXPORT_SYMBOL_GPL(alloc_vm_area);
2216
2217void free_vm_area(struct vm_struct *area)
2218{
2219    struct vm_struct *ret;
2220    ret = remove_vm_area(area->addr);
2221    BUG_ON(ret != area);
2222    kfree(area);
2223}
2224EXPORT_SYMBOL_GPL(free_vm_area);
2225
2226#ifdef CONFIG_SMP
2227static struct vmap_area *node_to_va(struct rb_node *n)
2228{
2229    return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2230}
2231
2232/**
2233 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2234 * @end: target address
2235 * @pnext: out arg for the next vmap_area
2236 * @pprev: out arg for the previous vmap_area
2237 *
2238 * Returns: %true if either or both of next and prev are found,
2239 * %false if no vmap_area exists
2240 *
2241 * Find vmap_areas end addresses of which enclose @end. ie. if not
2242 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2243 */
2244static bool pvm_find_next_prev(unsigned long end,
2245                   struct vmap_area **pnext,
2246                   struct vmap_area **pprev)
2247{
2248    struct rb_node *n = vmap_area_root.rb_node;
2249    struct vmap_area *va = NULL;
2250
2251    while (n) {
2252        va = rb_entry(n, struct vmap_area, rb_node);
2253        if (end < va->va_end)
2254            n = n->rb_left;
2255        else if (end > va->va_end)
2256            n = n->rb_right;
2257        else
2258            break;
2259    }
2260
2261    if (!va)
2262        return false;
2263
2264    if (va->va_end > end) {
2265        *pnext = va;
2266        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2267    } else {
2268        *pprev = va;
2269        *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2270    }
2271    return true;
2272}
2273
2274/**
2275 * pvm_determine_end - find the highest aligned address between two vmap_areas
2276 * @pnext: in/out arg for the next vmap_area
2277 * @pprev: in/out arg for the previous vmap_area
2278 * @align: alignment
2279 *
2280 * Returns: determined end address
2281 *
2282 * Find the highest aligned address between *@pnext and *@pprev below
2283 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2284 * down address is between the end addresses of the two vmap_areas.
2285 *
2286 * Please note that the address returned by this function may fall
2287 * inside *@pnext vmap_area. The caller is responsible for checking
2288 * that.
2289 */
2290static unsigned long pvm_determine_end(struct vmap_area **pnext,
2291                       struct vmap_area **pprev,
2292                       unsigned long align)
2293{
2294    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2295    unsigned long addr;
2296
2297    if (*pnext)
2298        addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2299    else
2300        addr = vmalloc_end;
2301
2302    while (*pprev && (*pprev)->va_end > addr) {
2303        *pnext = *pprev;
2304        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2305    }
2306
2307    return addr;
2308}
2309
2310/**
2311 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2312 * @offsets: array containing offset of each area
2313 * @sizes: array containing size of each area
2314 * @nr_vms: the number of areas to allocate
2315 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2316 *
2317 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2318 * vm_structs on success, %NULL on failure
2319 *
2320 * Percpu allocator wants to use congruent vm areas so that it can
2321 * maintain the offsets among percpu areas. This function allocates
2322 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2323 * be scattered pretty far, distance between two areas easily going up
2324 * to gigabytes. To avoid interacting with regular vmallocs, these
2325 * areas are allocated from top.
2326 *
2327 * Despite its complicated look, this allocator is rather simple. It
2328 * does everything top-down and scans areas from the end looking for
2329 * matching slot. While scanning, if any of the areas overlaps with
2330 * existing vmap_area, the base address is pulled down to fit the
2331 * area. Scanning is repeated till all the areas fit and then all
2332 * necessary data structres are inserted and the result is returned.
2333 */
2334struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2335                     const size_t *sizes, int nr_vms,
2336                     size_t align)
2337{
2338    const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2339    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2340    struct vmap_area **vas, *prev, *next;
2341    struct vm_struct **vms;
2342    int area, area2, last_area, term_area;
2343    unsigned long base, start, end, last_end;
2344    bool purged = false;
2345
2346    /* verify parameters and allocate data structures */
2347    BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2348    for (last_area = 0, area = 0; area < nr_vms; area++) {
2349        start = offsets[area];
2350        end = start + sizes[area];
2351
2352        /* is everything aligned properly? */
2353        BUG_ON(!IS_ALIGNED(offsets[area], align));
2354        BUG_ON(!IS_ALIGNED(sizes[area], align));
2355
2356        /* detect the area with the highest address */
2357        if (start > offsets[last_area])
2358            last_area = area;
2359
2360        for (area2 = 0; area2 < nr_vms; area2++) {
2361            unsigned long start2 = offsets[area2];
2362            unsigned long end2 = start2 + sizes[area2];
2363
2364            if (area2 == area)
2365                continue;
2366
2367            BUG_ON(start2 >= start && start2 < end);
2368            BUG_ON(end2 <= end && end2 > start);
2369        }
2370    }
2371    last_end = offsets[last_area] + sizes[last_area];
2372
2373    if (vmalloc_end - vmalloc_start < last_end) {
2374        WARN_ON(true);
2375        return NULL;
2376    }
2377
2378    vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
2379    vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
2380    if (!vas || !vms)
2381        goto err_free2;
2382
2383    for (area = 0; area < nr_vms; area++) {
2384        vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2385        vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2386        if (!vas[area] || !vms[area])
2387            goto err_free;
2388    }
2389retry:
2390    spin_lock(&vmap_area_lock);
2391
2392    /* start scanning - we scan from the top, begin with the last area */
2393    area = term_area = last_area;
2394    start = offsets[area];
2395    end = start + sizes[area];
2396
2397    if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2398        base = vmalloc_end - last_end;
2399        goto found;
2400    }
2401    base = pvm_determine_end(&next, &prev, align) - end;
2402
2403    while (true) {
2404        BUG_ON(next && next->va_end <= base + end);
2405        BUG_ON(prev && prev->va_end > base + end);
2406
2407        /*
2408         * base might have underflowed, add last_end before
2409         * comparing.
2410         */
2411        if (base + last_end < vmalloc_start + last_end) {
2412            spin_unlock(&vmap_area_lock);
2413            if (!purged) {
2414                purge_vmap_area_lazy();
2415                purged = true;
2416                goto retry;
2417            }
2418            goto err_free;
2419        }
2420
2421        /*
2422         * If next overlaps, move base downwards so that it's
2423         * right below next and then recheck.
2424         */
2425        if (next && next->va_start < base + end) {
2426            base = pvm_determine_end(&next, &prev, align) - end;
2427            term_area = area;
2428            continue;
2429        }
2430
2431        /*
2432         * If prev overlaps, shift down next and prev and move
2433         * base so that it's right below new next and then
2434         * recheck.
2435         */
2436        if (prev && prev->va_end > base + start) {
2437            next = prev;
2438            prev = node_to_va(rb_prev(&next->rb_node));
2439            base = pvm_determine_end(&next, &prev, align) - end;
2440            term_area = area;
2441            continue;
2442        }
2443
2444        /*
2445         * This area fits, move on to the previous one. If
2446         * the previous one is the terminal one, we're done.
2447         */
2448        area = (area + nr_vms - 1) % nr_vms;
2449        if (area == term_area)
2450            break;
2451        start = offsets[area];
2452        end = start + sizes[area];
2453        pvm_find_next_prev(base + end, &next, &prev);
2454    }
2455found:
2456    /* we've found a fitting base, insert all va's */
2457    for (area = 0; area < nr_vms; area++) {
2458        struct vmap_area *va = vas[area];
2459
2460        va->va_start = base + offsets[area];
2461        va->va_end = va->va_start + sizes[area];
2462        __insert_vmap_area(va);
2463    }
2464
2465    vmap_area_pcpu_hole = base + offsets[last_area];
2466
2467    spin_unlock(&vmap_area_lock);
2468
2469    /* insert all vm's */
2470    for (area = 0; area < nr_vms; area++)
2471        insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2472                  pcpu_get_vm_areas);
2473
2474    kfree(vas);
2475    return vms;
2476
2477err_free:
2478    for (area = 0; area < nr_vms; area++) {
2479        kfree(vas[area]);
2480        kfree(vms[area]);
2481    }
2482err_free2:
2483    kfree(vas);
2484    kfree(vms);
2485    return NULL;
2486}
2487
2488/**
2489 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2490 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2491 * @nr_vms: the number of allocated areas
2492 *
2493 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2494 */
2495void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2496{
2497    int i;
2498
2499    for (i = 0; i < nr_vms; i++)
2500        free_vm_area(vms[i]);
2501    kfree(vms);
2502}
2503#endif /* CONFIG_SMP */
2504
2505#ifdef CONFIG_PROC_FS
2506static void *s_start(struct seq_file *m, loff_t *pos)
2507    __acquires(&vmlist_lock)
2508{
2509    loff_t n = *pos;
2510    struct vm_struct *v;
2511
2512    read_lock(&vmlist_lock);
2513    v = vmlist;
2514    while (n > 0 && v) {
2515        n--;
2516        v = v->next;
2517    }
2518    if (!n)
2519        return v;
2520
2521    return NULL;
2522
2523}
2524
2525static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2526{
2527    struct vm_struct *v = p;
2528
2529    ++*pos;
2530    return v->next;
2531}
2532
2533static void s_stop(struct seq_file *m, void *p)
2534    __releases(&vmlist_lock)
2535{
2536    read_unlock(&vmlist_lock);
2537}
2538
2539static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2540{
2541    if (NUMA_BUILD) {
2542        unsigned int nr, *counters = m->private;
2543
2544        if (!counters)
2545            return;
2546
2547        memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2548
2549        for (nr = 0; nr < v->nr_pages; nr++)
2550            counters[page_to_nid(v->pages[nr])]++;
2551
2552        for_each_node_state(nr, N_HIGH_MEMORY)
2553            if (counters[nr])
2554                seq_printf(m, " N%u=%u", nr, counters[nr]);
2555    }
2556}
2557
2558static int s_show(struct seq_file *m, void *p)
2559{
2560    struct vm_struct *v = p;
2561
2562    seq_printf(m, "0x%p-0x%p %7ld",
2563        v->addr, v->addr + v->size, v->size);
2564
2565    if (v->caller)
2566        seq_printf(m, " %pS", v->caller);
2567
2568    if (v->nr_pages)
2569        seq_printf(m, " pages=%d", v->nr_pages);
2570
2571    if (v->phys_addr)
2572        seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2573
2574    if (v->flags & VM_IOREMAP)
2575        seq_printf(m, " ioremap");
2576
2577    if (v->flags & VM_ALLOC)
2578        seq_printf(m, " vmalloc");
2579
2580    if (v->flags & VM_MAP)
2581        seq_printf(m, " vmap");
2582
2583    if (v->flags & VM_USERMAP)
2584        seq_printf(m, " user");
2585
2586    if (v->flags & VM_VPAGES)
2587        seq_printf(m, " vpages");
2588
2589    show_numa_info(m, v);
2590    seq_putc(m, '\n');
2591    return 0;
2592}
2593
2594static const struct seq_operations vmalloc_op = {
2595    .start = s_start,
2596    .next = s_next,
2597    .stop = s_stop,
2598    .show = s_show,
2599};
2600
2601static int vmalloc_open(struct inode *inode, struct file *file)
2602{
2603    unsigned int *ptr = NULL;
2604    int ret;
2605
2606    if (NUMA_BUILD) {
2607        ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2608        if (ptr == NULL)
2609            return -ENOMEM;
2610    }
2611    ret = seq_open(file, &vmalloc_op);
2612    if (!ret) {
2613        struct seq_file *m = file->private_data;
2614        m->private = ptr;
2615    } else
2616        kfree(ptr);
2617    return ret;
2618}
2619
2620static const struct file_operations proc_vmalloc_operations = {
2621    .open = vmalloc_open,
2622    .read = seq_read,
2623    .llseek = seq_lseek,
2624    .release = seq_release_private,
2625};
2626
2627static int __init proc_vmalloc_init(void)
2628{
2629    proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2630    return 0;
2631}
2632module_init(proc_vmalloc_init);
2633#endif
2634
2635

Archive Download this file



interactive