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

Archive Download this file



interactive