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
34bool vmap_lazy_unmap __read_mostly = true;
35
36/*** Page table manipulation functions ***/
37
38static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
39{
40    pte_t *pte;
41
42    pte = pte_offset_kernel(pmd, addr);
43    do {
44        pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
45        WARN_ON(!pte_none(ptent) && !pte_present(ptent));
46    } while (pte++, addr += PAGE_SIZE, addr != end);
47}
48
49static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
50{
51    pmd_t *pmd;
52    unsigned long next;
53
54    pmd = pmd_offset(pud, addr);
55    do {
56        next = pmd_addr_end(addr, end);
57        if (pmd_none_or_clear_bad(pmd))
58            continue;
59        vunmap_pte_range(pmd, addr, next);
60    } while (pmd++, addr = next, addr != end);
61}
62
63static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
64{
65    pud_t *pud;
66    unsigned long next;
67
68    pud = pud_offset(pgd, addr);
69    do {
70        next = pud_addr_end(addr, end);
71        if (pud_none_or_clear_bad(pud))
72            continue;
73        vunmap_pmd_range(pud, addr, next);
74    } while (pud++, addr = next, addr != end);
75}
76
77static void vunmap_page_range(unsigned long addr, unsigned long end)
78{
79    pgd_t *pgd;
80    unsigned long next;
81
82    BUG_ON(addr >= end);
83    pgd = pgd_offset_k(addr);
84    do {
85        next = pgd_addr_end(addr, end);
86        if (pgd_none_or_clear_bad(pgd))
87            continue;
88        vunmap_pud_range(pgd, addr, next);
89    } while (pgd++, addr = next, addr != end);
90}
91
92static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
93        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
94{
95    pte_t *pte;
96
97    /*
98     * nr is a running index into the array which helps higher level
99     * callers keep track of where we're up to.
100     */
101
102    pte = pte_alloc_kernel(pmd, addr);
103    if (!pte)
104        return -ENOMEM;
105    do {
106        struct page *page = pages[*nr];
107
108        if (WARN_ON(!pte_none(*pte)))
109            return -EBUSY;
110        if (WARN_ON(!page))
111            return -ENOMEM;
112        set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
113        (*nr)++;
114    } while (pte++, addr += PAGE_SIZE, addr != end);
115    return 0;
116}
117
118static int vmap_pmd_range(pud_t *pud, unsigned long addr,
119        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
120{
121    pmd_t *pmd;
122    unsigned long next;
123
124    pmd = pmd_alloc(&init_mm, pud, addr);
125    if (!pmd)
126        return -ENOMEM;
127    do {
128        next = pmd_addr_end(addr, end);
129        if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
130            return -ENOMEM;
131    } while (pmd++, addr = next, addr != end);
132    return 0;
133}
134
135static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
136        unsigned long end, pgprot_t prot, struct page **pages, int *nr)
137{
138    pud_t *pud;
139    unsigned long next;
140
141    pud = pud_alloc(&init_mm, pgd, addr);
142    if (!pud)
143        return -ENOMEM;
144    do {
145        next = pud_addr_end(addr, end);
146        if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
147            return -ENOMEM;
148    } while (pud++, addr = next, addr != end);
149    return 0;
150}
151
152/*
153 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
154 * will have pfns corresponding to the "pages" array.
155 *
156 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
157 */
158static int vmap_page_range_noflush(unsigned long start, unsigned long end,
159                   pgprot_t prot, struct page **pages)
160{
161    pgd_t *pgd;
162    unsigned long next;
163    unsigned long addr = start;
164    int err = 0;
165    int nr = 0;
166
167    BUG_ON(addr >= end);
168    pgd = pgd_offset_k(addr);
169    do {
170        next = pgd_addr_end(addr, end);
171        err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
172        if (err)
173            return err;
174    } while (pgd++, addr = next, addr != end);
175
176    return nr;
177}
178
179static int vmap_page_range(unsigned long start, unsigned long end,
180               pgprot_t prot, struct page **pages)
181{
182    int ret;
183
184    ret = vmap_page_range_noflush(start, end, prot, pages);
185    flush_cache_vmap(start, end);
186    return ret;
187}
188
189int is_vmalloc_or_module_addr(const void *x)
190{
191    /*
192     * ARM, x86-64 and sparc64 put modules in a special place,
193     * and fall back on vmalloc() if that fails. Others
194     * just put it in the vmalloc space.
195     */
196#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
197    unsigned long addr = (unsigned long)x;
198    if (addr >= MODULES_VADDR && addr < MODULES_END)
199        return 1;
200#endif
201    return is_vmalloc_addr(x);
202}
203
204/*
205 * Walk a vmap address to the struct page it maps.
206 */
207struct page *vmalloc_to_page(const void *vmalloc_addr)
208{
209    unsigned long addr = (unsigned long) vmalloc_addr;
210    struct page *page = NULL;
211    pgd_t *pgd = pgd_offset_k(addr);
212
213    /*
214     * XXX we might need to change this if we add VIRTUAL_BUG_ON for
215     * architectures that do not vmalloc module space
216     */
217    VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
218
219    if (!pgd_none(*pgd)) {
220        pud_t *pud = pud_offset(pgd, addr);
221        if (!pud_none(*pud)) {
222            pmd_t *pmd = pmd_offset(pud, addr);
223            if (!pmd_none(*pmd)) {
224                pte_t *ptep, pte;
225
226                ptep = pte_offset_map(pmd, addr);
227                pte = *ptep;
228                if (pte_present(pte))
229                    page = pte_page(pte);
230                pte_unmap(ptep);
231            }
232        }
233    }
234    return page;
235}
236EXPORT_SYMBOL(vmalloc_to_page);
237
238/*
239 * Map a vmalloc()-space virtual address to the physical page frame number.
240 */
241unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
242{
243    return page_to_pfn(vmalloc_to_page(vmalloc_addr));
244}
245EXPORT_SYMBOL(vmalloc_to_pfn);
246
247
248/*** Global kva allocator ***/
249
250#define VM_LAZY_FREE 0x01
251#define VM_LAZY_FREEING 0x02
252#define VM_VM_AREA 0x04
253
254struct vmap_area {
255    unsigned long va_start;
256    unsigned long va_end;
257    unsigned long flags;
258    struct rb_node rb_node; /* address sorted rbtree */
259    struct list_head list; /* address sorted list */
260    struct list_head purge_list; /* "lazy purge" list */
261    void *private;
262    struct rcu_head rcu_head;
263};
264
265static DEFINE_SPINLOCK(vmap_area_lock);
266static struct rb_root vmap_area_root = RB_ROOT;
267static LIST_HEAD(vmap_area_list);
268static unsigned long vmap_area_pcpu_hole;
269
270static struct vmap_area *__find_vmap_area(unsigned long addr)
271{
272    struct rb_node *n = vmap_area_root.rb_node;
273
274    while (n) {
275        struct vmap_area *va;
276
277        va = rb_entry(n, struct vmap_area, rb_node);
278        if (addr < va->va_start)
279            n = n->rb_left;
280        else if (addr > va->va_start)
281            n = n->rb_right;
282        else
283            return va;
284    }
285
286    return NULL;
287}
288
289static void __insert_vmap_area(struct vmap_area *va)
290{
291    struct rb_node **p = &vmap_area_root.rb_node;
292    struct rb_node *parent = NULL;
293    struct rb_node *tmp;
294
295    while (*p) {
296        struct vmap_area *tmp;
297
298        parent = *p;
299        tmp = rb_entry(parent, struct vmap_area, rb_node);
300        if (va->va_start < tmp->va_end)
301            p = &(*p)->rb_left;
302        else if (va->va_end > tmp->va_start)
303            p = &(*p)->rb_right;
304        else
305            BUG();
306    }
307
308    rb_link_node(&va->rb_node, parent, p);
309    rb_insert_color(&va->rb_node, &vmap_area_root);
310
311    /* address-sort this list so it is usable like the vmlist */
312    tmp = rb_prev(&va->rb_node);
313    if (tmp) {
314        struct vmap_area *prev;
315        prev = rb_entry(tmp, struct vmap_area, rb_node);
316        list_add_rcu(&va->list, &prev->list);
317    } else
318        list_add_rcu(&va->list, &vmap_area_list);
319}
320
321static void purge_vmap_area_lazy(void);
322
323/*
324 * Allocate a region of KVA of the specified size and alignment, within the
325 * vstart and vend.
326 */
327static struct vmap_area *alloc_vmap_area(unsigned long size,
328                unsigned long align,
329                unsigned long vstart, unsigned long vend,
330                int node, gfp_t gfp_mask)
331{
332    struct vmap_area *va;
333    struct rb_node *n;
334    unsigned long addr;
335    int purged = 0;
336
337    BUG_ON(!size);
338    BUG_ON(size & ~PAGE_MASK);
339
340    va = kmalloc_node(sizeof(struct vmap_area),
341            gfp_mask & GFP_RECLAIM_MASK, node);
342    if (unlikely(!va))
343        return ERR_PTR(-ENOMEM);
344
345retry:
346    addr = ALIGN(vstart, align);
347
348    spin_lock(&vmap_area_lock);
349    if (addr + size - 1 < addr)
350        goto overflow;
351
352    /* XXX: could have a last_hole cache */
353    n = vmap_area_root.rb_node;
354    if (n) {
355        struct vmap_area *first = NULL;
356
357        do {
358            struct vmap_area *tmp;
359            tmp = rb_entry(n, struct vmap_area, rb_node);
360            if (tmp->va_end >= addr) {
361                if (!first && tmp->va_start < addr + size)
362                    first = tmp;
363                n = n->rb_left;
364            } else {
365                first = tmp;
366                n = n->rb_right;
367            }
368        } while (n);
369
370        if (!first)
371            goto found;
372
373        if (first->va_end < addr) {
374            n = rb_next(&first->rb_node);
375            if (n)
376                first = rb_entry(n, struct vmap_area, rb_node);
377            else
378                goto found;
379        }
380
381        while (addr + size > first->va_start && addr + size <= vend) {
382            addr = ALIGN(first->va_end + PAGE_SIZE, align);
383            if (addr + size - 1 < addr)
384                goto overflow;
385
386            n = rb_next(&first->rb_node);
387            if (n)
388                first = rb_entry(n, struct vmap_area, rb_node);
389            else
390                goto found;
391        }
392    }
393found:
394    if (addr + size > vend) {
395overflow:
396        spin_unlock(&vmap_area_lock);
397        if (!purged) {
398            purge_vmap_area_lazy();
399            purged = 1;
400            goto retry;
401        }
402        if (printk_ratelimit())
403            printk(KERN_WARNING
404                "vmap allocation for size %lu failed: "
405                "use vmalloc=<size> to increase size.\n", size);
406        kfree(va);
407        return ERR_PTR(-EBUSY);
408    }
409
410    BUG_ON(addr & (align-1));
411
412    va->va_start = addr;
413    va->va_end = addr + size;
414    va->flags = 0;
415    __insert_vmap_area(va);
416    spin_unlock(&vmap_area_lock);
417
418    return va;
419}
420
421static void rcu_free_va(struct rcu_head *head)
422{
423    struct vmap_area *va = container_of(head, struct vmap_area, rcu_head);
424
425    kfree(va);
426}
427
428static void __free_vmap_area(struct vmap_area *va)
429{
430    BUG_ON(RB_EMPTY_NODE(&va->rb_node));
431    rb_erase(&va->rb_node, &vmap_area_root);
432    RB_CLEAR_NODE(&va->rb_node);
433    list_del_rcu(&va->list);
434
435    /*
436     * Track the highest possible candidate for pcpu area
437     * allocation. Areas outside of vmalloc area can be returned
438     * here too, consider only end addresses which fall inside
439     * vmalloc area proper.
440     */
441    if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
442        vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
443
444    call_rcu(&va->rcu_head, rcu_free_va);
445}
446
447/*
448 * Free a region of KVA allocated by alloc_vmap_area
449 */
450static void free_vmap_area(struct vmap_area *va)
451{
452    spin_lock(&vmap_area_lock);
453    __free_vmap_area(va);
454    spin_unlock(&vmap_area_lock);
455}
456
457/*
458 * Clear the pagetable entries of a given vmap_area
459 */
460static void unmap_vmap_area(struct vmap_area *va)
461{
462    vunmap_page_range(va->va_start, va->va_end);
463}
464
465static void vmap_debug_free_range(unsigned long start, unsigned long end)
466{
467    /*
468     * Unmap page tables and force a TLB flush immediately if
469     * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
470     * bugs similarly to those in linear kernel virtual address
471     * space after a page has been freed.
472     *
473     * All the lazy freeing logic is still retained, in order to
474     * minimise intrusiveness of this debugging feature.
475     *
476     * This is going to be *slow* (linear kernel virtual address
477     * debugging doesn't do a broadcast TLB flush so it is a lot
478     * faster).
479     */
480#ifdef CONFIG_DEBUG_PAGEALLOC
481    vunmap_page_range(start, end);
482    flush_tlb_kernel_range(start, end);
483#endif
484}
485
486/*
487 * lazy_max_pages is the maximum amount of virtual address space we gather up
488 * before attempting to purge with a TLB flush.
489 *
490 * There is a tradeoff here: a larger number will cover more kernel page tables
491 * and take slightly longer to purge, but it will linearly reduce the number of
492 * global TLB flushes that must be performed. It would seem natural to scale
493 * this number up linearly with the number of CPUs (because vmapping activity
494 * could also scale linearly with the number of CPUs), however it is likely
495 * that in practice, workloads might be constrained in other ways that mean
496 * vmap activity will not scale linearly with CPUs. Also, I want to be
497 * conservative and not introduce a big latency on huge systems, so go with
498 * a less aggressive log scale. It will still be an improvement over the old
499 * code, and it will be simple to change the scale factor if we find that it
500 * becomes a problem on bigger systems.
501 */
502static unsigned long lazy_max_pages(void)
503{
504    unsigned int log;
505
506    if (!vmap_lazy_unmap)
507        return 0;
508
509    log = fls(num_online_cpus());
510
511    return log * (32UL * 1024 * 1024 / PAGE_SIZE);
512}
513
514static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
515
516/* for per-CPU blocks */
517static void purge_fragmented_blocks_allcpus(void);
518
519/*
520 * Purges all lazily-freed vmap areas.
521 *
522 * If sync is 0 then don't purge if there is already a purge in progress.
523 * If force_flush is 1, then flush kernel TLBs between *start and *end even
524 * if we found no lazy vmap areas to unmap (callers can use this to optimise
525 * their own TLB flushing).
526 * Returns with *start = min(*start, lowest purged address)
527 * *end = max(*end, highest purged address)
528 */
529static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
530                    int sync, int force_flush)
531{
532    static DEFINE_SPINLOCK(purge_lock);
533    LIST_HEAD(valist);
534    struct vmap_area *va;
535    struct vmap_area *n_va;
536    int nr = 0;
537
538    /*
539     * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
540     * should not expect such behaviour. This just simplifies locking for
541     * the case that isn't actually used at the moment anyway.
542     */
543    if (!sync && !force_flush) {
544        if (!spin_trylock(&purge_lock))
545            return;
546    } else
547        spin_lock(&purge_lock);
548
549    if (sync)
550        purge_fragmented_blocks_allcpus();
551
552    rcu_read_lock();
553    list_for_each_entry_rcu(va, &vmap_area_list, list) {
554        if (va->flags & VM_LAZY_FREE) {
555            if (va->va_start < *start)
556                *start = va->va_start;
557            if (va->va_end > *end)
558                *end = va->va_end;
559            nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
560            unmap_vmap_area(va);
561            list_add_tail(&va->purge_list, &valist);
562            va->flags |= VM_LAZY_FREEING;
563            va->flags &= ~VM_LAZY_FREE;
564        }
565    }
566    rcu_read_unlock();
567
568    if (nr)
569        atomic_sub(nr, &vmap_lazy_nr);
570
571    if (nr || force_flush)
572        flush_tlb_kernel_range(*start, *end);
573
574    if (nr) {
575        spin_lock(&vmap_area_lock);
576        list_for_each_entry_safe(va, n_va, &valist, purge_list)
577            __free_vmap_area(va);
578        spin_unlock(&vmap_area_lock);
579    }
580    spin_unlock(&purge_lock);
581}
582
583/*
584 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
585 * is already purging.
586 */
587static void try_purge_vmap_area_lazy(void)
588{
589    unsigned long start = ULONG_MAX, end = 0;
590
591    __purge_vmap_area_lazy(&start, &end, 0, 0);
592}
593
594/*
595 * Kick off a purge of the outstanding lazy areas.
596 */
597static void purge_vmap_area_lazy(void)
598{
599    unsigned long start = ULONG_MAX, end = 0;
600
601    __purge_vmap_area_lazy(&start, &end, 1, 0);
602}
603
604/*
605 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
606 * called for the correct range previously.
607 */
608static void free_unmap_vmap_area_noflush(struct vmap_area *va)
609{
610    va->flags |= VM_LAZY_FREE;
611    atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
612    if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
613        try_purge_vmap_area_lazy();
614}
615
616/*
617 * Free and unmap a vmap area
618 */
619static void free_unmap_vmap_area(struct vmap_area *va)
620{
621    flush_cache_vunmap(va->va_start, va->va_end);
622    free_unmap_vmap_area_noflush(va);
623}
624
625static struct vmap_area *find_vmap_area(unsigned long addr)
626{
627    struct vmap_area *va;
628
629    spin_lock(&vmap_area_lock);
630    va = __find_vmap_area(addr);
631    spin_unlock(&vmap_area_lock);
632
633    return va;
634}
635
636static void free_unmap_vmap_area_addr(unsigned long addr)
637{
638    struct vmap_area *va;
639
640    va = find_vmap_area(addr);
641    BUG_ON(!va);
642    free_unmap_vmap_area(va);
643}
644
645
646/*** Per cpu kva allocator ***/
647
648/*
649 * vmap space is limited especially on 32 bit architectures. Ensure there is
650 * room for at least 16 percpu vmap blocks per CPU.
651 */
652/*
653 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
654 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
655 * instead (we just need a rough idea)
656 */
657#if BITS_PER_LONG == 32
658#define VMALLOC_SPACE (128UL*1024*1024)
659#else
660#define VMALLOC_SPACE (128UL*1024*1024*1024)
661#endif
662
663#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
664#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
665#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
666#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
667#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
668#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
669#define VMAP_BBMAP_BITS VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
670                    VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
671                        VMALLOC_PAGES / NR_CPUS / 16))
672
673#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
674
675static bool vmap_initialized __read_mostly = false;
676
677struct vmap_block_queue {
678    spinlock_t lock;
679    struct list_head free;
680};
681
682struct vmap_block {
683    spinlock_t lock;
684    struct vmap_area *va;
685    struct vmap_block_queue *vbq;
686    unsigned long free, dirty;
687    DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
688    DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
689    struct list_head free_list;
690    struct rcu_head rcu_head;
691    struct list_head purge;
692};
693
694/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
695static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
696
697/*
698 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
699 * in the free path. Could get rid of this if we change the API to return a
700 * "cookie" from alloc, to be passed to free. But no big deal yet.
701 */
702static DEFINE_SPINLOCK(vmap_block_tree_lock);
703static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
704
705/*
706 * We should probably have a fallback mechanism to allocate virtual memory
707 * out of partially filled vmap blocks. However vmap block sizing should be
708 * fairly reasonable according to the vmalloc size, so it shouldn't be a
709 * big problem.
710 */
711
712static unsigned long addr_to_vb_idx(unsigned long addr)
713{
714    addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
715    addr /= VMAP_BLOCK_SIZE;
716    return addr;
717}
718
719static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
720{
721    struct vmap_block_queue *vbq;
722    struct vmap_block *vb;
723    struct vmap_area *va;
724    unsigned long vb_idx;
725    int node, err;
726
727    node = numa_node_id();
728
729    vb = kmalloc_node(sizeof(struct vmap_block),
730            gfp_mask & GFP_RECLAIM_MASK, node);
731    if (unlikely(!vb))
732        return ERR_PTR(-ENOMEM);
733
734    va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
735                    VMALLOC_START, VMALLOC_END,
736                    node, gfp_mask);
737    if (unlikely(IS_ERR(va))) {
738        kfree(vb);
739        return ERR_CAST(va);
740    }
741
742    err = radix_tree_preload(gfp_mask);
743    if (unlikely(err)) {
744        kfree(vb);
745        free_vmap_area(va);
746        return ERR_PTR(err);
747    }
748
749    spin_lock_init(&vb->lock);
750    vb->va = va;
751    vb->free = VMAP_BBMAP_BITS;
752    vb->dirty = 0;
753    bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
754    bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
755    INIT_LIST_HEAD(&vb->free_list);
756
757    vb_idx = addr_to_vb_idx(va->va_start);
758    spin_lock(&vmap_block_tree_lock);
759    err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
760    spin_unlock(&vmap_block_tree_lock);
761    BUG_ON(err);
762    radix_tree_preload_end();
763
764    vbq = &get_cpu_var(vmap_block_queue);
765    vb->vbq = vbq;
766    spin_lock(&vbq->lock);
767    list_add_rcu(&vb->free_list, &vbq->free);
768    spin_unlock(&vbq->lock);
769    put_cpu_var(vmap_block_queue);
770
771    return vb;
772}
773
774static void rcu_free_vb(struct rcu_head *head)
775{
776    struct vmap_block *vb = container_of(head, struct vmap_block, rcu_head);
777
778    kfree(vb);
779}
780
781static void free_vmap_block(struct vmap_block *vb)
782{
783    struct vmap_block *tmp;
784    unsigned long vb_idx;
785
786    vb_idx = addr_to_vb_idx(vb->va->va_start);
787    spin_lock(&vmap_block_tree_lock);
788    tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
789    spin_unlock(&vmap_block_tree_lock);
790    BUG_ON(tmp != vb);
791
792    free_unmap_vmap_area_noflush(vb->va);
793    call_rcu(&vb->rcu_head, rcu_free_vb);
794}
795
796static void purge_fragmented_blocks(int cpu)
797{
798    LIST_HEAD(purge);
799    struct vmap_block *vb;
800    struct vmap_block *n_vb;
801    struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
802
803    rcu_read_lock();
804    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
805
806        if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
807            continue;
808
809        spin_lock(&vb->lock);
810        if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
811            vb->free = 0; /* prevent further allocs after releasing lock */
812            vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
813            bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
814            bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
815            spin_lock(&vbq->lock);
816            list_del_rcu(&vb->free_list);
817            spin_unlock(&vbq->lock);
818            spin_unlock(&vb->lock);
819            list_add_tail(&vb->purge, &purge);
820        } else
821            spin_unlock(&vb->lock);
822    }
823    rcu_read_unlock();
824
825    list_for_each_entry_safe(vb, n_vb, &purge, purge) {
826        list_del(&vb->purge);
827        free_vmap_block(vb);
828    }
829}
830
831static void purge_fragmented_blocks_thiscpu(void)
832{
833    purge_fragmented_blocks(smp_processor_id());
834}
835
836static void purge_fragmented_blocks_allcpus(void)
837{
838    int cpu;
839
840    for_each_possible_cpu(cpu)
841        purge_fragmented_blocks(cpu);
842}
843
844static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
845{
846    struct vmap_block_queue *vbq;
847    struct vmap_block *vb;
848    unsigned long addr = 0;
849    unsigned int order;
850    int purge = 0;
851
852    BUG_ON(size & ~PAGE_MASK);
853    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
854    order = get_order(size);
855
856again:
857    rcu_read_lock();
858    vbq = &get_cpu_var(vmap_block_queue);
859    list_for_each_entry_rcu(vb, &vbq->free, free_list) {
860        int i;
861
862        spin_lock(&vb->lock);
863        if (vb->free < 1UL << order)
864            goto next;
865
866        i = bitmap_find_free_region(vb->alloc_map,
867                        VMAP_BBMAP_BITS, order);
868
869        if (i < 0) {
870            if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
871                /* fragmented and no outstanding allocations */
872                BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
873                purge = 1;
874            }
875            goto next;
876        }
877        addr = vb->va->va_start + (i << PAGE_SHIFT);
878        BUG_ON(addr_to_vb_idx(addr) !=
879                addr_to_vb_idx(vb->va->va_start));
880        vb->free -= 1UL << order;
881        if (vb->free == 0) {
882            spin_lock(&vbq->lock);
883            list_del_rcu(&vb->free_list);
884            spin_unlock(&vbq->lock);
885        }
886        spin_unlock(&vb->lock);
887        break;
888next:
889        spin_unlock(&vb->lock);
890    }
891
892    if (purge)
893        purge_fragmented_blocks_thiscpu();
894
895    put_cpu_var(vmap_block_queue);
896    rcu_read_unlock();
897
898    if (!addr) {
899        vb = new_vmap_block(gfp_mask);
900        if (IS_ERR(vb))
901            return vb;
902        goto again;
903    }
904
905    return (void *)addr;
906}
907
908static void vb_free(const void *addr, unsigned long size)
909{
910    unsigned long offset;
911    unsigned long vb_idx;
912    unsigned int order;
913    struct vmap_block *vb;
914
915    BUG_ON(size & ~PAGE_MASK);
916    BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
917
918    flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
919
920    order = get_order(size);
921
922    offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
923
924    vb_idx = addr_to_vb_idx((unsigned long)addr);
925    rcu_read_lock();
926    vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
927    rcu_read_unlock();
928    BUG_ON(!vb);
929
930    spin_lock(&vb->lock);
931    BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
932
933    vb->dirty += 1UL << order;
934    if (vb->dirty == VMAP_BBMAP_BITS) {
935        BUG_ON(vb->free);
936        spin_unlock(&vb->lock);
937        free_vmap_block(vb);
938    } else
939        spin_unlock(&vb->lock);
940}
941
942/**
943 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
944 *
945 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
946 * to amortize TLB flushing overheads. What this means is that any page you
947 * have now, may, in a former life, have been mapped into kernel virtual
948 * address by the vmap layer and so there might be some CPUs with TLB entries
949 * still referencing that page (additional to the regular 1:1 kernel mapping).
950 *
951 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
952 * be sure that none of the pages we have control over will have any aliases
953 * from the vmap layer.
954 */
955void vm_unmap_aliases(void)
956{
957    unsigned long start = ULONG_MAX, end = 0;
958    int cpu;
959    int flush = 0;
960
961    if (unlikely(!vmap_initialized))
962        return;
963
964    for_each_possible_cpu(cpu) {
965        struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
966        struct vmap_block *vb;
967
968        rcu_read_lock();
969        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
970            int i;
971
972            spin_lock(&vb->lock);
973            i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
974            while (i < VMAP_BBMAP_BITS) {
975                unsigned long s, e;
976                int j;
977                j = find_next_zero_bit(vb->dirty_map,
978                    VMAP_BBMAP_BITS, i);
979
980                s = vb->va->va_start + (i << PAGE_SHIFT);
981                e = vb->va->va_start + (j << PAGE_SHIFT);
982                vunmap_page_range(s, e);
983                flush = 1;
984
985                if (s < start)
986                    start = s;
987                if (e > end)
988                    end = e;
989
990                i = j;
991                i = find_next_bit(vb->dirty_map,
992                            VMAP_BBMAP_BITS, i);
993            }
994            spin_unlock(&vb->lock);
995        }
996        rcu_read_unlock();
997    }
998
999    __purge_vmap_area_lazy(&start, &end, 1, flush);
1000}
1001EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1002
1003/**
1004 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1005 * @mem: the pointer returned by vm_map_ram
1006 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1007 */
1008void vm_unmap_ram(const void *mem, unsigned int count)
1009{
1010    unsigned long size = count << PAGE_SHIFT;
1011    unsigned long addr = (unsigned long)mem;
1012
1013    BUG_ON(!addr);
1014    BUG_ON(addr < VMALLOC_START);
1015    BUG_ON(addr > VMALLOC_END);
1016    BUG_ON(addr & (PAGE_SIZE-1));
1017
1018    debug_check_no_locks_freed(mem, size);
1019    vmap_debug_free_range(addr, addr+size);
1020
1021    if (likely(count <= VMAP_MAX_ALLOC))
1022        vb_free(mem, size);
1023    else
1024        free_unmap_vmap_area_addr(addr);
1025}
1026EXPORT_SYMBOL(vm_unmap_ram);
1027
1028/**
1029 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1030 * @pages: an array of pointers to the pages to be mapped
1031 * @count: number of pages
1032 * @node: prefer to allocate data structures on this node
1033 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1034 *
1035 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1036 */
1037void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1038{
1039    unsigned long size = count << PAGE_SHIFT;
1040    unsigned long addr;
1041    void *mem;
1042
1043    if (likely(count <= VMAP_MAX_ALLOC)) {
1044        mem = vb_alloc(size, GFP_KERNEL);
1045        if (IS_ERR(mem))
1046            return NULL;
1047        addr = (unsigned long)mem;
1048    } else {
1049        struct vmap_area *va;
1050        va = alloc_vmap_area(size, PAGE_SIZE,
1051                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1052        if (IS_ERR(va))
1053            return NULL;
1054
1055        addr = va->va_start;
1056        mem = (void *)addr;
1057    }
1058    if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1059        vm_unmap_ram(mem, count);
1060        return NULL;
1061    }
1062    return mem;
1063}
1064EXPORT_SYMBOL(vm_map_ram);
1065
1066/**
1067 * vm_area_register_early - register vmap area early during boot
1068 * @vm: vm_struct to register
1069 * @align: requested alignment
1070 *
1071 * This function is used to register kernel vm area before
1072 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1073 * proper values on entry and other fields should be zero. On return,
1074 * vm->addr contains the allocated address.
1075 *
1076 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1077 */
1078void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1079{
1080    static size_t vm_init_off __initdata;
1081    unsigned long addr;
1082
1083    addr = ALIGN(VMALLOC_START + vm_init_off, align);
1084    vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1085
1086    vm->addr = (void *)addr;
1087
1088    vm->next = vmlist;
1089    vmlist = vm;
1090}
1091
1092void __init vmalloc_init(void)
1093{
1094    struct vmap_area *va;
1095    struct vm_struct *tmp;
1096    int i;
1097
1098    for_each_possible_cpu(i) {
1099        struct vmap_block_queue *vbq;
1100
1101        vbq = &per_cpu(vmap_block_queue, i);
1102        spin_lock_init(&vbq->lock);
1103        INIT_LIST_HEAD(&vbq->free);
1104    }
1105
1106    /* Import existing vmlist entries. */
1107    for (tmp = vmlist; tmp; tmp = tmp->next) {
1108        va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1109        va->flags = tmp->flags | VM_VM_AREA;
1110        va->va_start = (unsigned long)tmp->addr;
1111        va->va_end = va->va_start + tmp->size;
1112        __insert_vmap_area(va);
1113    }
1114
1115    vmap_area_pcpu_hole = VMALLOC_END;
1116
1117    vmap_initialized = true;
1118}
1119
1120/**
1121 * map_kernel_range_noflush - map kernel VM area with the specified pages
1122 * @addr: start of the VM area to map
1123 * @size: size of the VM area to map
1124 * @prot: page protection flags to use
1125 * @pages: pages to map
1126 *
1127 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1128 * specify should have been allocated using get_vm_area() and its
1129 * friends.
1130 *
1131 * NOTE:
1132 * This function does NOT do any cache flushing. The caller is
1133 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1134 * before calling this function.
1135 *
1136 * RETURNS:
1137 * The number of pages mapped on success, -errno on failure.
1138 */
1139int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1140                 pgprot_t prot, struct page **pages)
1141{
1142    return vmap_page_range_noflush(addr, addr + size, prot, pages);
1143}
1144
1145/**
1146 * unmap_kernel_range_noflush - unmap kernel VM area
1147 * @addr: start of the VM area to unmap
1148 * @size: size of the VM area to unmap
1149 *
1150 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1151 * specify should have been allocated using get_vm_area() and its
1152 * friends.
1153 *
1154 * NOTE:
1155 * This function does NOT do any cache flushing. The caller is
1156 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1157 * before calling this function and flush_tlb_kernel_range() after.
1158 */
1159void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1160{
1161    vunmap_page_range(addr, addr + size);
1162}
1163
1164/**
1165 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1166 * @addr: start of the VM area to unmap
1167 * @size: size of the VM area to unmap
1168 *
1169 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1170 * the unmapping and tlb after.
1171 */
1172void unmap_kernel_range(unsigned long addr, unsigned long size)
1173{
1174    unsigned long end = addr + size;
1175
1176    flush_cache_vunmap(addr, end);
1177    vunmap_page_range(addr, end);
1178    flush_tlb_kernel_range(addr, end);
1179}
1180
1181int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1182{
1183    unsigned long addr = (unsigned long)area->addr;
1184    unsigned long end = addr + area->size - PAGE_SIZE;
1185    int err;
1186
1187    err = vmap_page_range(addr, end, prot, *pages);
1188    if (err > 0) {
1189        *pages += err;
1190        err = 0;
1191    }
1192
1193    return err;
1194}
1195EXPORT_SYMBOL_GPL(map_vm_area);
1196
1197/*** Old vmalloc interfaces ***/
1198DEFINE_RWLOCK(vmlist_lock);
1199struct vm_struct *vmlist;
1200
1201static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1202                  unsigned long flags, void *caller)
1203{
1204    struct vm_struct *tmp, **p;
1205
1206    vm->flags = flags;
1207    vm->addr = (void *)va->va_start;
1208    vm->size = va->va_end - va->va_start;
1209    vm->caller = caller;
1210    va->private = vm;
1211    va->flags |= VM_VM_AREA;
1212
1213    write_lock(&vmlist_lock);
1214    for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1215        if (tmp->addr >= vm->addr)
1216            break;
1217    }
1218    vm->next = *p;
1219    *p = vm;
1220    write_unlock(&vmlist_lock);
1221}
1222
1223static struct vm_struct *__get_vm_area_node(unsigned long size,
1224        unsigned long align, unsigned long flags, unsigned long start,
1225        unsigned long end, int node, gfp_t gfp_mask, void *caller)
1226{
1227    static struct vmap_area *va;
1228    struct vm_struct *area;
1229
1230    BUG_ON(in_interrupt());
1231    if (flags & VM_IOREMAP) {
1232        int bit = fls(size);
1233
1234        if (bit > IOREMAP_MAX_ORDER)
1235            bit = IOREMAP_MAX_ORDER;
1236        else if (bit < PAGE_SHIFT)
1237            bit = PAGE_SHIFT;
1238
1239        align = 1ul << bit;
1240    }
1241
1242    size = PAGE_ALIGN(size);
1243    if (unlikely(!size))
1244        return NULL;
1245
1246    area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1247    if (unlikely(!area))
1248        return NULL;
1249
1250    /*
1251     * We always allocate a guard page.
1252     */
1253    size += PAGE_SIZE;
1254
1255    va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1256    if (IS_ERR(va)) {
1257        kfree(area);
1258        return NULL;
1259    }
1260
1261    insert_vmalloc_vm(area, va, flags, caller);
1262    return area;
1263}
1264
1265struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1266                unsigned long start, unsigned long end)
1267{
1268    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1269                        __builtin_return_address(0));
1270}
1271EXPORT_SYMBOL_GPL(__get_vm_area);
1272
1273struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1274                       unsigned long start, unsigned long end,
1275                       void *caller)
1276{
1277    return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1278                  caller);
1279}
1280
1281/**
1282 * get_vm_area - reserve a contiguous kernel virtual area
1283 * @size: size of the area
1284 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1285 *
1286 * Search an area of @size in the kernel virtual mapping area,
1287 * and reserved it for out purposes. Returns the area descriptor
1288 * on success or %NULL on failure.
1289 */
1290struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1291{
1292    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1293                -1, GFP_KERNEL, __builtin_return_address(0));
1294}
1295
1296struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1297                void *caller)
1298{
1299    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1300                        -1, GFP_KERNEL, caller);
1301}
1302
1303struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags,
1304                   int node, gfp_t gfp_mask)
1305{
1306    return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1307                  node, gfp_mask, __builtin_return_address(0));
1308}
1309
1310static struct vm_struct *find_vm_area(const void *addr)
1311{
1312    struct vmap_area *va;
1313
1314    va = find_vmap_area((unsigned long)addr);
1315    if (va && va->flags & VM_VM_AREA)
1316        return va->private;
1317
1318    return NULL;
1319}
1320
1321/**
1322 * remove_vm_area - find and remove a continuous kernel virtual area
1323 * @addr: base address
1324 *
1325 * Search for the kernel VM area starting at @addr, and remove it.
1326 * This function returns the found VM area, but using it is NOT safe
1327 * on SMP machines, except for its size or flags.
1328 */
1329struct vm_struct *remove_vm_area(const void *addr)
1330{
1331    struct vmap_area *va;
1332
1333    va = find_vmap_area((unsigned long)addr);
1334    if (va && va->flags & VM_VM_AREA) {
1335        struct vm_struct *vm = va->private;
1336        struct vm_struct *tmp, **p;
1337        /*
1338         * remove from list and disallow access to this vm_struct
1339         * before unmap. (address range confliction is maintained by
1340         * vmap.)
1341         */
1342        write_lock(&vmlist_lock);
1343        for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1344            ;
1345        *p = tmp->next;
1346        write_unlock(&vmlist_lock);
1347
1348        vmap_debug_free_range(va->va_start, va->va_end);
1349        free_unmap_vmap_area(va);
1350        vm->size -= PAGE_SIZE;
1351
1352        return vm;
1353    }
1354    return NULL;
1355}
1356
1357static void __vunmap(const void *addr, int deallocate_pages)
1358{
1359    struct vm_struct *area;
1360
1361    if (!addr)
1362        return;
1363
1364    if ((PAGE_SIZE-1) & (unsigned long)addr) {
1365        WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1366        return;
1367    }
1368
1369    area = remove_vm_area(addr);
1370    if (unlikely(!area)) {
1371        WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1372                addr);
1373        return;
1374    }
1375
1376    debug_check_no_locks_freed(addr, area->size);
1377    debug_check_no_obj_freed(addr, area->size);
1378
1379    if (deallocate_pages) {
1380        int i;
1381
1382        for (i = 0; i < area->nr_pages; i++) {
1383            struct page *page = area->pages[i];
1384
1385            BUG_ON(!page);
1386            __free_page(page);
1387        }
1388
1389        if (area->flags & VM_VPAGES)
1390            vfree(area->pages);
1391        else
1392            kfree(area->pages);
1393    }
1394
1395    kfree(area);
1396    return;
1397}
1398
1399/**
1400 * vfree - release memory allocated by vmalloc()
1401 * @addr: memory base address
1402 *
1403 * Free the virtually continuous memory area starting at @addr, as
1404 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1405 * NULL, no operation is performed.
1406 *
1407 * Must not be called in interrupt context.
1408 */
1409void vfree(const void *addr)
1410{
1411    BUG_ON(in_interrupt());
1412
1413    kmemleak_free(addr);
1414
1415    __vunmap(addr, 1);
1416}
1417EXPORT_SYMBOL(vfree);
1418
1419/**
1420 * vunmap - release virtual mapping obtained by vmap()
1421 * @addr: memory base address
1422 *
1423 * Free the virtually contiguous memory area starting at @addr,
1424 * which was created from the page array passed to vmap().
1425 *
1426 * Must not be called in interrupt context.
1427 */
1428void vunmap(const void *addr)
1429{
1430    BUG_ON(in_interrupt());
1431    might_sleep();
1432    __vunmap(addr, 0);
1433}
1434EXPORT_SYMBOL(vunmap);
1435
1436/**
1437 * vmap - map an array of pages into virtually contiguous space
1438 * @pages: array of page pointers
1439 * @count: number of pages to map
1440 * @flags: vm_area->flags
1441 * @prot: page protection for the mapping
1442 *
1443 * Maps @count pages from @pages into contiguous kernel virtual
1444 * space.
1445 */
1446void *vmap(struct page **pages, unsigned int count,
1447        unsigned long flags, pgprot_t prot)
1448{
1449    struct vm_struct *area;
1450
1451    might_sleep();
1452
1453    if (count > totalram_pages)
1454        return NULL;
1455
1456    area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1457                    __builtin_return_address(0));
1458    if (!area)
1459        return NULL;
1460
1461    if (map_vm_area(area, prot, &pages)) {
1462        vunmap(area->addr);
1463        return NULL;
1464    }
1465
1466    return area->addr;
1467}
1468EXPORT_SYMBOL(vmap);
1469
1470static void *__vmalloc_node(unsigned long size, unsigned long align,
1471                gfp_t gfp_mask, pgprot_t prot,
1472                int node, void *caller);
1473static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1474                 pgprot_t prot, int node, void *caller)
1475{
1476    struct page **pages;
1477    unsigned int nr_pages, array_size, i;
1478    gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1479
1480    nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1481    array_size = (nr_pages * sizeof(struct page *));
1482
1483    area->nr_pages = nr_pages;
1484    /* Please note that the recursion is strictly bounded. */
1485    if (array_size > PAGE_SIZE) {
1486        pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1487                PAGE_KERNEL, node, caller);
1488        area->flags |= VM_VPAGES;
1489    } else {
1490        pages = kmalloc_node(array_size, nested_gfp, node);
1491    }
1492    area->pages = pages;
1493    area->caller = caller;
1494    if (!area->pages) {
1495        remove_vm_area(area->addr);
1496        kfree(area);
1497        return NULL;
1498    }
1499
1500    for (i = 0; i < area->nr_pages; i++) {
1501        struct page *page;
1502
1503        if (node < 0)
1504            page = alloc_page(gfp_mask);
1505        else
1506            page = alloc_pages_node(node, gfp_mask, 0);
1507
1508        if (unlikely(!page)) {
1509            /* Successfully allocated i pages, free them in __vunmap() */
1510            area->nr_pages = i;
1511            goto fail;
1512        }
1513        area->pages[i] = page;
1514    }
1515
1516    if (map_vm_area(area, prot, &pages))
1517        goto fail;
1518    return area->addr;
1519
1520fail:
1521    vfree(area->addr);
1522    return NULL;
1523}
1524
1525void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
1526{
1527    void *addr = __vmalloc_area_node(area, gfp_mask, prot, -1,
1528                     __builtin_return_address(0));
1529
1530    /*
1531     * A ref_count = 3 is needed because the vm_struct and vmap_area
1532     * structures allocated in the __get_vm_area_node() function contain
1533     * references to the virtual address of the vmalloc'ed block.
1534     */
1535    kmemleak_alloc(addr, area->size - PAGE_SIZE, 3, gfp_mask);
1536
1537    return addr;
1538}
1539
1540/**
1541 * __vmalloc_node - allocate virtually contiguous memory
1542 * @size: allocation size
1543 * @align: desired alignment
1544 * @gfp_mask: flags for the page level allocator
1545 * @prot: protection mask for the allocated pages
1546 * @node: node to use for allocation or -1
1547 * @caller: caller's return address
1548 *
1549 * Allocate enough pages to cover @size from the page level
1550 * allocator with @gfp_mask flags. Map them into contiguous
1551 * kernel virtual space, using a pagetable protection of @prot.
1552 */
1553static void *__vmalloc_node(unsigned long size, unsigned long align,
1554                gfp_t gfp_mask, pgprot_t prot,
1555                int node, void *caller)
1556{
1557    struct vm_struct *area;
1558    void *addr;
1559    unsigned long real_size = size;
1560
1561    size = PAGE_ALIGN(size);
1562    if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1563        return NULL;
1564
1565    area = __get_vm_area_node(size, align, VM_ALLOC, VMALLOC_START,
1566                  VMALLOC_END, node, gfp_mask, caller);
1567
1568    if (!area)
1569        return NULL;
1570
1571    addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1572
1573    /*
1574     * A ref_count = 3 is needed because the vm_struct and vmap_area
1575     * structures allocated in the __get_vm_area_node() function contain
1576     * references to the virtual address of the vmalloc'ed block.
1577     */
1578    kmemleak_alloc(addr, real_size, 3, gfp_mask);
1579
1580    return addr;
1581}
1582
1583void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1584{
1585    return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1586                __builtin_return_address(0));
1587}
1588EXPORT_SYMBOL(__vmalloc);
1589
1590/**
1591 * vmalloc - allocate virtually contiguous memory
1592 * @size: allocation size
1593 * Allocate enough pages to cover @size from the page level
1594 * allocator and map them into contiguous kernel virtual space.
1595 *
1596 * For tight control over page level allocator and protection flags
1597 * use __vmalloc() instead.
1598 */
1599void *vmalloc(unsigned long size)
1600{
1601    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1602                    -1, __builtin_return_address(0));
1603}
1604EXPORT_SYMBOL(vmalloc);
1605
1606/**
1607 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1608 * @size: allocation size
1609 *
1610 * The resulting memory area is zeroed so it can be mapped to userspace
1611 * without leaking data.
1612 */
1613void *vmalloc_user(unsigned long size)
1614{
1615    struct vm_struct *area;
1616    void *ret;
1617
1618    ret = __vmalloc_node(size, SHMLBA,
1619                 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1620                 PAGE_KERNEL, -1, __builtin_return_address(0));
1621    if (ret) {
1622        area = find_vm_area(ret);
1623        area->flags |= VM_USERMAP;
1624    }
1625    return ret;
1626}
1627EXPORT_SYMBOL(vmalloc_user);
1628
1629/**
1630 * vmalloc_node - allocate memory on a specific node
1631 * @size: allocation size
1632 * @node: numa node
1633 *
1634 * Allocate enough pages to cover @size from the page level
1635 * allocator and map them into contiguous kernel virtual space.
1636 *
1637 * For tight control over page level allocator and protection flags
1638 * use __vmalloc() instead.
1639 */
1640void *vmalloc_node(unsigned long size, int node)
1641{
1642    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1643                    node, __builtin_return_address(0));
1644}
1645EXPORT_SYMBOL(vmalloc_node);
1646
1647#ifndef PAGE_KERNEL_EXEC
1648# define PAGE_KERNEL_EXEC PAGE_KERNEL
1649#endif
1650
1651/**
1652 * vmalloc_exec - allocate virtually contiguous, executable memory
1653 * @size: allocation size
1654 *
1655 * Kernel-internal function to allocate enough pages to cover @size
1656 * the page level allocator and map them into contiguous and
1657 * executable kernel virtual space.
1658 *
1659 * For tight control over page level allocator and protection flags
1660 * use __vmalloc() instead.
1661 */
1662
1663void *vmalloc_exec(unsigned long size)
1664{
1665    return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1666                  -1, __builtin_return_address(0));
1667}
1668
1669#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1670#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1671#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1672#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1673#else
1674#define GFP_VMALLOC32 GFP_KERNEL
1675#endif
1676
1677/**
1678 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1679 * @size: allocation size
1680 *
1681 * Allocate enough 32bit PA addressable pages to cover @size from the
1682 * page level allocator and map them into contiguous kernel virtual space.
1683 */
1684void *vmalloc_32(unsigned long size)
1685{
1686    return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1687                  -1, __builtin_return_address(0));
1688}
1689EXPORT_SYMBOL(vmalloc_32);
1690
1691/**
1692 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1693 * @size: allocation size
1694 *
1695 * The resulting memory area is 32bit addressable and zeroed so it can be
1696 * mapped to userspace without leaking data.
1697 */
1698void *vmalloc_32_user(unsigned long size)
1699{
1700    struct vm_struct *area;
1701    void *ret;
1702
1703    ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1704                 -1, __builtin_return_address(0));
1705    if (ret) {
1706        area = find_vm_area(ret);
1707        area->flags |= VM_USERMAP;
1708    }
1709    return ret;
1710}
1711EXPORT_SYMBOL(vmalloc_32_user);
1712
1713/*
1714 * small helper routine , copy contents to buf from addr.
1715 * If the page is not present, fill zero.
1716 */
1717
1718static int aligned_vread(char *buf, char *addr, unsigned long count)
1719{
1720    struct page *p;
1721    int copied = 0;
1722
1723    while (count) {
1724        unsigned long offset, length;
1725
1726        offset = (unsigned long)addr & ~PAGE_MASK;
1727        length = PAGE_SIZE - offset;
1728        if (length > count)
1729            length = count;
1730        p = vmalloc_to_page(addr);
1731        /*
1732         * To do safe access to this _mapped_ area, we need
1733         * lock. But adding lock here means that we need to add
1734         * overhead of vmalloc()/vfree() calles for this _debug_
1735         * interface, rarely used. Instead of that, we'll use
1736         * kmap() and get small overhead in this access function.
1737         */
1738        if (p) {
1739            /*
1740             * we can expect USER0 is not used (see vread/vwrite's
1741             * function description)
1742             */
1743            void *map = kmap_atomic(p, KM_USER0);
1744            memcpy(buf, map + offset, length);
1745            kunmap_atomic(map, KM_USER0);
1746        } else
1747            memset(buf, 0, length);
1748
1749        addr += length;
1750        buf += length;
1751        copied += length;
1752        count -= length;
1753    }
1754    return copied;
1755}
1756
1757static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1758{
1759    struct page *p;
1760    int copied = 0;
1761
1762    while (count) {
1763        unsigned long offset, length;
1764
1765        offset = (unsigned long)addr & ~PAGE_MASK;
1766        length = PAGE_SIZE - offset;
1767        if (length > count)
1768            length = count;
1769        p = vmalloc_to_page(addr);
1770        /*
1771         * To do safe access to this _mapped_ area, we need
1772         * lock. But adding lock here means that we need to add
1773         * overhead of vmalloc()/vfree() calles for this _debug_
1774         * interface, rarely used. Instead of that, we'll use
1775         * kmap() and get small overhead in this access function.
1776         */
1777        if (p) {
1778            /*
1779             * we can expect USER0 is not used (see vread/vwrite's
1780             * function description)
1781             */
1782            void *map = kmap_atomic(p, KM_USER0);
1783            memcpy(map + offset, buf, length);
1784            kunmap_atomic(map, KM_USER0);
1785        }
1786        addr += length;
1787        buf += length;
1788        copied += length;
1789        count -= length;
1790    }
1791    return copied;
1792}
1793
1794/**
1795 * vread() - read vmalloc area in a safe way.
1796 * @buf: buffer for reading data
1797 * @addr: vm address.
1798 * @count: number of bytes to be read.
1799 *
1800 * Returns # of bytes which addr and buf should be increased.
1801 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1802 * includes any intersect with alive vmalloc area.
1803 *
1804 * This function checks that addr is a valid vmalloc'ed area, and
1805 * copy data from that area to a given buffer. If the given memory range
1806 * of [addr...addr+count) includes some valid address, data is copied to
1807 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1808 * IOREMAP area is treated as memory hole and no copy is done.
1809 *
1810 * If [addr...addr+count) doesn't includes any intersects with alive
1811 * vm_struct area, returns 0.
1812 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1813 * the caller should guarantee KM_USER0 is not used.
1814 *
1815 * Note: In usual ops, vread() is never necessary because the caller
1816 * should know vmalloc() area is valid and can use memcpy().
1817 * This is for routines which have to access vmalloc area without
1818 * any informaion, as /dev/kmem.
1819 *
1820 */
1821
1822long vread(char *buf, char *addr, unsigned long count)
1823{
1824    struct vm_struct *tmp;
1825    char *vaddr, *buf_start = buf;
1826    unsigned long buflen = count;
1827    unsigned long n;
1828
1829    /* Don't allow overflow */
1830    if ((unsigned long) addr + count < count)
1831        count = -(unsigned long) addr;
1832
1833    read_lock(&vmlist_lock);
1834    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1835        vaddr = (char *) tmp->addr;
1836        if (addr >= vaddr + tmp->size - PAGE_SIZE)
1837            continue;
1838        while (addr < vaddr) {
1839            if (count == 0)
1840                goto finished;
1841            *buf = '\0';
1842            buf++;
1843            addr++;
1844            count--;
1845        }
1846        n = vaddr + tmp->size - PAGE_SIZE - addr;
1847        if (n > count)
1848            n = count;
1849        if (!(tmp->flags & VM_IOREMAP))
1850            aligned_vread(buf, addr, n);
1851        else /* IOREMAP area is treated as memory hole */
1852            memset(buf, 0, n);
1853        buf += n;
1854        addr += n;
1855        count -= n;
1856    }
1857finished:
1858    read_unlock(&vmlist_lock);
1859
1860    if (buf == buf_start)
1861        return 0;
1862    /* zero-fill memory holes */
1863    if (buf != buf_start + buflen)
1864        memset(buf, 0, buflen - (buf - buf_start));
1865
1866    return buflen;
1867}
1868
1869/**
1870 * vwrite() - write vmalloc area in a safe way.
1871 * @buf: buffer for source data
1872 * @addr: vm address.
1873 * @count: number of bytes to be read.
1874 *
1875 * Returns # of bytes which addr and buf should be incresed.
1876 * (same number to @count).
1877 * If [addr...addr+count) doesn't includes any intersect with valid
1878 * vmalloc area, returns 0.
1879 *
1880 * This function checks that addr is a valid vmalloc'ed area, and
1881 * copy data from a buffer to the given addr. If specified range of
1882 * [addr...addr+count) includes some valid address, data is copied from
1883 * proper area of @buf. If there are memory holes, no copy to hole.
1884 * IOREMAP area is treated as memory hole and no copy is done.
1885 *
1886 * If [addr...addr+count) doesn't includes any intersects with alive
1887 * vm_struct area, returns 0.
1888 * @buf should be kernel's buffer. Because this function uses KM_USER0,
1889 * the caller should guarantee KM_USER0 is not used.
1890 *
1891 * Note: In usual ops, vwrite() is never necessary because the caller
1892 * should know vmalloc() area is valid and can use memcpy().
1893 * This is for routines which have to access vmalloc area without
1894 * any informaion, as /dev/kmem.
1895 *
1896 * The caller should guarantee KM_USER1 is not used.
1897 */
1898
1899long vwrite(char *buf, char *addr, unsigned long count)
1900{
1901    struct vm_struct *tmp;
1902    char *vaddr;
1903    unsigned long n, buflen;
1904    int copied = 0;
1905
1906    /* Don't allow overflow */
1907    if ((unsigned long) addr + count < count)
1908        count = -(unsigned long) addr;
1909    buflen = count;
1910
1911    read_lock(&vmlist_lock);
1912    for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1913        vaddr = (char *) tmp->addr;
1914        if (addr >= vaddr + tmp->size - PAGE_SIZE)
1915            continue;
1916        while (addr < vaddr) {
1917            if (count == 0)
1918                goto finished;
1919            buf++;
1920            addr++;
1921            count--;
1922        }
1923        n = vaddr + tmp->size - PAGE_SIZE - addr;
1924        if (n > count)
1925            n = count;
1926        if (!(tmp->flags & VM_IOREMAP)) {
1927            aligned_vwrite(buf, addr, n);
1928            copied++;
1929        }
1930        buf += n;
1931        addr += n;
1932        count -= n;
1933    }
1934finished:
1935    read_unlock(&vmlist_lock);
1936    if (!copied)
1937        return 0;
1938    return buflen;
1939}
1940
1941/**
1942 * remap_vmalloc_range - map vmalloc pages to userspace
1943 * @vma: vma to cover (map full range of vma)
1944 * @addr: vmalloc memory
1945 * @pgoff: number of pages into addr before first page to map
1946 *
1947 * Returns: 0 for success, -Exxx on failure
1948 *
1949 * This function checks that addr is a valid vmalloc'ed area, and
1950 * that it is big enough to cover the vma. Will return failure if
1951 * that criteria isn't met.
1952 *
1953 * Similar to remap_pfn_range() (see mm/memory.c)
1954 */
1955int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1956                        unsigned long pgoff)
1957{
1958    struct vm_struct *area;
1959    unsigned long uaddr = vma->vm_start;
1960    unsigned long usize = vma->vm_end - vma->vm_start;
1961
1962    if ((PAGE_SIZE-1) & (unsigned long)addr)
1963        return -EINVAL;
1964
1965    area = find_vm_area(addr);
1966    if (!area)
1967        return -EINVAL;
1968
1969    if (!(area->flags & VM_USERMAP))
1970        return -EINVAL;
1971
1972    if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
1973        return -EINVAL;
1974
1975    addr += pgoff << PAGE_SHIFT;
1976    do {
1977        struct page *page = vmalloc_to_page(addr);
1978        int ret;
1979
1980        ret = vm_insert_page(vma, uaddr, page);
1981        if (ret)
1982            return ret;
1983
1984        uaddr += PAGE_SIZE;
1985        addr += PAGE_SIZE;
1986        usize -= PAGE_SIZE;
1987    } while (usize > 0);
1988
1989    /* Prevent "things" like memory migration? VM_flags need a cleanup... */
1990    vma->vm_flags |= VM_RESERVED;
1991
1992    return 0;
1993}
1994EXPORT_SYMBOL(remap_vmalloc_range);
1995
1996/*
1997 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
1998 * have one.
1999 */
2000void __attribute__((weak)) vmalloc_sync_all(void)
2001{
2002}
2003
2004
2005static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2006{
2007    /* apply_to_page_range() does all the hard work. */
2008    return 0;
2009}
2010
2011/**
2012 * alloc_vm_area - allocate a range of kernel address space
2013 * @size: size of the area
2014 *
2015 * Returns: NULL on failure, vm_struct on success
2016 *
2017 * This function reserves a range of kernel address space, and
2018 * allocates pagetables to map that range. No actual mappings
2019 * are created. If the kernel address space is not shared
2020 * between processes, it syncs the pagetable across all
2021 * processes.
2022 */
2023struct vm_struct *alloc_vm_area(size_t size)
2024{
2025    struct vm_struct *area;
2026
2027    area = get_vm_area_caller(size, VM_IOREMAP,
2028                __builtin_return_address(0));
2029    if (area == NULL)
2030        return NULL;
2031
2032    /*
2033     * This ensures that page tables are constructed for this region
2034     * of kernel virtual address space and mapped into init_mm.
2035     */
2036    if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2037                area->size, f, NULL)) {
2038        free_vm_area(area);
2039        return NULL;
2040    }
2041
2042    /* Make sure the pagetables are constructed in process kernel
2043       mappings */
2044    vmalloc_sync_all();
2045
2046    return area;
2047}
2048EXPORT_SYMBOL_GPL(alloc_vm_area);
2049
2050void free_vm_area(struct vm_struct *area)
2051{
2052    struct vm_struct *ret;
2053    ret = remove_vm_area(area->addr);
2054    BUG_ON(ret != area);
2055    kfree(area);
2056}
2057EXPORT_SYMBOL_GPL(free_vm_area);
2058
2059static struct vmap_area *node_to_va(struct rb_node *n)
2060{
2061    return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2062}
2063
2064/**
2065 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2066 * @end: target address
2067 * @pnext: out arg for the next vmap_area
2068 * @pprev: out arg for the previous vmap_area
2069 *
2070 * Returns: %true if either or both of next and prev are found,
2071 * %false if no vmap_area exists
2072 *
2073 * Find vmap_areas end addresses of which enclose @end. ie. if not
2074 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2075 */
2076static bool pvm_find_next_prev(unsigned long end,
2077                   struct vmap_area **pnext,
2078                   struct vmap_area **pprev)
2079{
2080    struct rb_node *n = vmap_area_root.rb_node;
2081    struct vmap_area *va = NULL;
2082
2083    while (n) {
2084        va = rb_entry(n, struct vmap_area, rb_node);
2085        if (end < va->va_end)
2086            n = n->rb_left;
2087        else if (end > va->va_end)
2088            n = n->rb_right;
2089        else
2090            break;
2091    }
2092
2093    if (!va)
2094        return false;
2095
2096    if (va->va_end > end) {
2097        *pnext = va;
2098        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2099    } else {
2100        *pprev = va;
2101        *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2102    }
2103    return true;
2104}
2105
2106/**
2107 * pvm_determine_end - find the highest aligned address between two vmap_areas
2108 * @pnext: in/out arg for the next vmap_area
2109 * @pprev: in/out arg for the previous vmap_area
2110 * @align: alignment
2111 *
2112 * Returns: determined end address
2113 *
2114 * Find the highest aligned address between *@pnext and *@pprev below
2115 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2116 * down address is between the end addresses of the two vmap_areas.
2117 *
2118 * Please note that the address returned by this function may fall
2119 * inside *@pnext vmap_area. The caller is responsible for checking
2120 * that.
2121 */
2122static unsigned long pvm_determine_end(struct vmap_area **pnext,
2123                       struct vmap_area **pprev,
2124                       unsigned long align)
2125{
2126    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2127    unsigned long addr;
2128
2129    if (*pnext)
2130        addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2131    else
2132        addr = vmalloc_end;
2133
2134    while (*pprev && (*pprev)->va_end > addr) {
2135        *pnext = *pprev;
2136        *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2137    }
2138
2139    return addr;
2140}
2141
2142/**
2143 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2144 * @offsets: array containing offset of each area
2145 * @sizes: array containing size of each area
2146 * @nr_vms: the number of areas to allocate
2147 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2148 * @gfp_mask: allocation mask
2149 *
2150 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2151 * vm_structs on success, %NULL on failure
2152 *
2153 * Percpu allocator wants to use congruent vm areas so that it can
2154 * maintain the offsets among percpu areas. This function allocates
2155 * congruent vmalloc areas for it. These areas tend to be scattered
2156 * pretty far, distance between two areas easily going up to
2157 * gigabytes. To avoid interacting with regular vmallocs, these areas
2158 * are allocated from top.
2159 *
2160 * Despite its complicated look, this allocator is rather simple. It
2161 * does everything top-down and scans areas from the end looking for
2162 * matching slot. While scanning, if any of the areas overlaps with
2163 * existing vmap_area, the base address is pulled down to fit the
2164 * area. Scanning is repeated till all the areas fit and then all
2165 * necessary data structres are inserted and the result is returned.
2166 */
2167struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2168                     const size_t *sizes, int nr_vms,
2169                     size_t align, gfp_t gfp_mask)
2170{
2171    const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2172    const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2173    struct vmap_area **vas, *prev, *next;
2174    struct vm_struct **vms;
2175    int area, area2, last_area, term_area;
2176    unsigned long base, start, end, last_end;
2177    bool purged = false;
2178
2179    gfp_mask &= GFP_RECLAIM_MASK;
2180
2181    /* verify parameters and allocate data structures */
2182    BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2183    for (last_area = 0, area = 0; area < nr_vms; area++) {
2184        start = offsets[area];
2185        end = start + sizes[area];
2186
2187        /* is everything aligned properly? */
2188        BUG_ON(!IS_ALIGNED(offsets[area], align));
2189        BUG_ON(!IS_ALIGNED(sizes[area], align));
2190
2191        /* detect the area with the highest address */
2192        if (start > offsets[last_area])
2193            last_area = area;
2194
2195        for (area2 = 0; area2 < nr_vms; area2++) {
2196            unsigned long start2 = offsets[area2];
2197            unsigned long end2 = start2 + sizes[area2];
2198
2199            if (area2 == area)
2200                continue;
2201
2202            BUG_ON(start2 >= start && start2 < end);
2203            BUG_ON(end2 <= end && end2 > start);
2204        }
2205    }
2206    last_end = offsets[last_area] + sizes[last_area];
2207
2208    if (vmalloc_end - vmalloc_start < last_end) {
2209        WARN_ON(true);
2210        return NULL;
2211    }
2212
2213    vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
2214    vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
2215    if (!vas || !vms)
2216        goto err_free;
2217
2218    for (area = 0; area < nr_vms; area++) {
2219        vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
2220        vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
2221        if (!vas[area] || !vms[area])
2222            goto err_free;
2223    }
2224retry:
2225    spin_lock(&vmap_area_lock);
2226
2227    /* start scanning - we scan from the top, begin with the last area */
2228    area = term_area = last_area;
2229    start = offsets[area];
2230    end = start + sizes[area];
2231
2232    if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2233        base = vmalloc_end - last_end;
2234        goto found;
2235    }
2236    base = pvm_determine_end(&next, &prev, align) - end;
2237
2238    while (true) {
2239        BUG_ON(next && next->va_end <= base + end);
2240        BUG_ON(prev && prev->va_end > base + end);
2241
2242        /*
2243         * base might have underflowed, add last_end before
2244         * comparing.
2245         */
2246        if (base + last_end < vmalloc_start + last_end) {
2247            spin_unlock(&vmap_area_lock);
2248            if (!purged) {
2249                purge_vmap_area_lazy();
2250                purged = true;
2251                goto retry;
2252            }
2253            goto err_free;
2254        }
2255
2256        /*
2257         * If next overlaps, move base downwards so that it's
2258         * right below next and then recheck.
2259         */
2260        if (next && next->va_start < base + end) {
2261            base = pvm_determine_end(&next, &prev, align) - end;
2262            term_area = area;
2263            continue;
2264        }
2265
2266        /*
2267         * If prev overlaps, shift down next and prev and move
2268         * base so that it's right below new next and then
2269         * recheck.
2270         */
2271        if (prev && prev->va_end > base + start) {
2272            next = prev;
2273            prev = node_to_va(rb_prev(&next->rb_node));
2274            base = pvm_determine_end(&next, &prev, align) - end;
2275            term_area = area;
2276            continue;
2277        }
2278
2279        /*
2280         * This area fits, move on to the previous one. If
2281         * the previous one is the terminal one, we're done.
2282         */
2283        area = (area + nr_vms - 1) % nr_vms;
2284        if (area == term_area)
2285            break;
2286        start = offsets[area];
2287        end = start + sizes[area];
2288        pvm_find_next_prev(base + end, &next, &prev);
2289    }
2290found:
2291    /* we've found a fitting base, insert all va's */
2292    for (area = 0; area < nr_vms; area++) {
2293        struct vmap_area *va = vas[area];
2294
2295        va->va_start = base + offsets[area];
2296        va->va_end = va->va_start + sizes[area];
2297        __insert_vmap_area(va);
2298    }
2299
2300    vmap_area_pcpu_hole = base + offsets[last_area];
2301
2302    spin_unlock(&vmap_area_lock);
2303
2304    /* insert all vm's */
2305    for (area = 0; area < nr_vms; area++)
2306        insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2307                  pcpu_get_vm_areas);
2308
2309    kfree(vas);
2310    return vms;
2311
2312err_free:
2313    for (area = 0; area < nr_vms; area++) {
2314        if (vas)
2315            kfree(vas[area]);
2316        if (vms)
2317            kfree(vms[area]);
2318    }
2319    kfree(vas);
2320    kfree(vms);
2321    return NULL;
2322}
2323
2324/**
2325 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2326 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2327 * @nr_vms: the number of allocated areas
2328 *
2329 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2330 */
2331void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2332{
2333    int i;
2334
2335    for (i = 0; i < nr_vms; i++)
2336        free_vm_area(vms[i]);
2337    kfree(vms);
2338}
2339
2340#ifdef CONFIG_PROC_FS
2341static void *s_start(struct seq_file *m, loff_t *pos)
2342{
2343    loff_t n = *pos;
2344    struct vm_struct *v;
2345
2346    read_lock(&vmlist_lock);
2347    v = vmlist;
2348    while (n > 0 && v) {
2349        n--;
2350        v = v->next;
2351    }
2352    if (!n)
2353        return v;
2354
2355    return NULL;
2356
2357}
2358
2359static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2360{
2361    struct vm_struct *v = p;
2362
2363    ++*pos;
2364    return v->next;
2365}
2366
2367static void s_stop(struct seq_file *m, void *p)
2368{
2369    read_unlock(&vmlist_lock);
2370}
2371
2372static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2373{
2374    if (NUMA_BUILD) {
2375        unsigned int nr, *counters = m->private;
2376
2377        if (!counters)
2378            return;
2379
2380        memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2381
2382        for (nr = 0; nr < v->nr_pages; nr++)
2383            counters[page_to_nid(v->pages[nr])]++;
2384
2385        for_each_node_state(nr, N_HIGH_MEMORY)
2386            if (counters[nr])
2387                seq_printf(m, " N%u=%u", nr, counters[nr]);
2388    }
2389}
2390
2391static int s_show(struct seq_file *m, void *p)
2392{
2393    struct vm_struct *v = p;
2394
2395    seq_printf(m, "0x%p-0x%p %7ld",
2396        v->addr, v->addr + v->size, v->size);
2397
2398    if (v->caller) {
2399        char buff[KSYM_SYMBOL_LEN];
2400
2401        seq_putc(m, ' ');
2402        sprint_symbol(buff, (unsigned long)v->caller);
2403        seq_puts(m, buff);
2404    }
2405
2406    if (v->nr_pages)
2407        seq_printf(m, " pages=%d", v->nr_pages);
2408
2409    if (v->phys_addr)
2410        seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2411
2412    if (v->flags & VM_IOREMAP)
2413        seq_printf(m, " ioremap");
2414
2415    if (v->flags & VM_ALLOC)
2416        seq_printf(m, " vmalloc");
2417
2418    if (v->flags & VM_MAP)
2419        seq_printf(m, " vmap");
2420
2421    if (v->flags & VM_USERMAP)
2422        seq_printf(m, " user");
2423
2424    if (v->flags & VM_VPAGES)
2425        seq_printf(m, " vpages");
2426
2427    show_numa_info(m, v);
2428    seq_putc(m, '\n');
2429    return 0;
2430}
2431
2432static const struct seq_operations vmalloc_op = {
2433    .start = s_start,
2434    .next = s_next,
2435    .stop = s_stop,
2436    .show = s_show,
2437};
2438
2439static int vmalloc_open(struct inode *inode, struct file *file)
2440{
2441    unsigned int *ptr = NULL;
2442    int ret;
2443
2444    if (NUMA_BUILD) {
2445        ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2446        if (ptr == NULL)
2447            return -ENOMEM;
2448    }
2449    ret = seq_open(file, &vmalloc_op);
2450    if (!ret) {
2451        struct seq_file *m = file->private_data;
2452        m->private = ptr;
2453    } else
2454        kfree(ptr);
2455    return ret;
2456}
2457
2458static const struct file_operations proc_vmalloc_operations = {
2459    .open = vmalloc_open,
2460    .read = seq_read,
2461    .llseek = seq_lseek,
2462    .release = seq_release_private,
2463};
2464
2465static int __init proc_vmalloc_init(void)
2466{
2467    proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2468    return 0;
2469}
2470module_init(proc_vmalloc_init);
2471#endif
2472
2473

Archive Download this file



interactive