Loading...
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/signal.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/notifier.h>
25#include <linux/rbtree.h>
26#include <linux/radix-tree.h>
27#include <linux/rcupdate.h>
28#include <linux/pfn.h>
29#include <linux/kmemleak.h>
30#include <linux/atomic.h>
31#include <linux/compiler.h>
32#include <linux/llist.h>
33#include <linux/bitops.h>
34
35#include <linux/uaccess.h>
36#include <asm/tlbflush.h>
37#include <asm/shmparam.h>
38
39#include "internal.h"
40
41struct vfree_deferred {
42 struct llist_head list;
43 struct work_struct wq;
44};
45static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
46
47static void __vunmap(const void *, int);
48
49static void free_work(struct work_struct *w)
50{
51 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
52 struct llist_node *t, *llnode;
53
54 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
55 __vunmap((void *)llnode, 1);
56}
57
58/*** Page table manipulation functions ***/
59
60static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
61{
62 pte_t *pte;
63
64 pte = pte_offset_kernel(pmd, addr);
65 do {
66 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
67 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
68 } while (pte++, addr += PAGE_SIZE, addr != end);
69}
70
71static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
72{
73 pmd_t *pmd;
74 unsigned long next;
75
76 pmd = pmd_offset(pud, addr);
77 do {
78 next = pmd_addr_end(addr, end);
79 if (pmd_clear_huge(pmd))
80 continue;
81 if (pmd_none_or_clear_bad(pmd))
82 continue;
83 vunmap_pte_range(pmd, addr, next);
84 } while (pmd++, addr = next, addr != end);
85}
86
87static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
88{
89 pud_t *pud;
90 unsigned long next;
91
92 pud = pud_offset(p4d, addr);
93 do {
94 next = pud_addr_end(addr, end);
95 if (pud_clear_huge(pud))
96 continue;
97 if (pud_none_or_clear_bad(pud))
98 continue;
99 vunmap_pmd_range(pud, addr, next);
100 } while (pud++, addr = next, addr != end);
101}
102
103static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
104{
105 p4d_t *p4d;
106 unsigned long next;
107
108 p4d = p4d_offset(pgd, addr);
109 do {
110 next = p4d_addr_end(addr, end);
111 if (p4d_clear_huge(p4d))
112 continue;
113 if (p4d_none_or_clear_bad(p4d))
114 continue;
115 vunmap_pud_range(p4d, addr, next);
116 } while (p4d++, addr = next, addr != end);
117}
118
119static void vunmap_page_range(unsigned long addr, unsigned long end)
120{
121 pgd_t *pgd;
122 unsigned long next;
123
124 BUG_ON(addr >= end);
125 pgd = pgd_offset_k(addr);
126 do {
127 next = pgd_addr_end(addr, end);
128 if (pgd_none_or_clear_bad(pgd))
129 continue;
130 vunmap_p4d_range(pgd, addr, next);
131 } while (pgd++, addr = next, addr != end);
132}
133
134static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
135 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
136{
137 pte_t *pte;
138
139 /*
140 * nr is a running index into the array which helps higher level
141 * callers keep track of where we're up to.
142 */
143
144 pte = pte_alloc_kernel(pmd, addr);
145 if (!pte)
146 return -ENOMEM;
147 do {
148 struct page *page = pages[*nr];
149
150 if (WARN_ON(!pte_none(*pte)))
151 return -EBUSY;
152 if (WARN_ON(!page))
153 return -ENOMEM;
154 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
155 (*nr)++;
156 } while (pte++, addr += PAGE_SIZE, addr != end);
157 return 0;
158}
159
160static int vmap_pmd_range(pud_t *pud, unsigned long addr,
161 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
162{
163 pmd_t *pmd;
164 unsigned long next;
165
166 pmd = pmd_alloc(&init_mm, pud, addr);
167 if (!pmd)
168 return -ENOMEM;
169 do {
170 next = pmd_addr_end(addr, end);
171 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
172 return -ENOMEM;
173 } while (pmd++, addr = next, addr != end);
174 return 0;
175}
176
177static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
178 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
179{
180 pud_t *pud;
181 unsigned long next;
182
183 pud = pud_alloc(&init_mm, p4d, addr);
184 if (!pud)
185 return -ENOMEM;
186 do {
187 next = pud_addr_end(addr, end);
188 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
189 return -ENOMEM;
190 } while (pud++, addr = next, addr != end);
191 return 0;
192}
193
194static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
195 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
196{
197 p4d_t *p4d;
198 unsigned long next;
199
200 p4d = p4d_alloc(&init_mm, pgd, addr);
201 if (!p4d)
202 return -ENOMEM;
203 do {
204 next = p4d_addr_end(addr, end);
205 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
206 return -ENOMEM;
207 } while (p4d++, addr = next, addr != end);
208 return 0;
209}
210
211/*
212 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
213 * will have pfns corresponding to the "pages" array.
214 *
215 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
216 */
217static int vmap_page_range_noflush(unsigned long start, unsigned long end,
218 pgprot_t prot, struct page **pages)
219{
220 pgd_t *pgd;
221 unsigned long next;
222 unsigned long addr = start;
223 int err = 0;
224 int nr = 0;
225
226 BUG_ON(addr >= end);
227 pgd = pgd_offset_k(addr);
228 do {
229 next = pgd_addr_end(addr, end);
230 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
231 if (err)
232 return err;
233 } while (pgd++, addr = next, addr != end);
234
235 return nr;
236}
237
238static int vmap_page_range(unsigned long start, unsigned long end,
239 pgprot_t prot, struct page **pages)
240{
241 int ret;
242
243 ret = vmap_page_range_noflush(start, end, prot, pages);
244 flush_cache_vmap(start, end);
245 return ret;
246}
247
248int is_vmalloc_or_module_addr(const void *x)
249{
250 /*
251 * ARM, x86-64 and sparc64 put modules in a special place,
252 * and fall back on vmalloc() if that fails. Others
253 * just put it in the vmalloc space.
254 */
255#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
256 unsigned long addr = (unsigned long)x;
257 if (addr >= MODULES_VADDR && addr < MODULES_END)
258 return 1;
259#endif
260 return is_vmalloc_addr(x);
261}
262
263/*
264 * Walk a vmap address to the struct page it maps.
265 */
266struct page *vmalloc_to_page(const void *vmalloc_addr)
267{
268 unsigned long addr = (unsigned long) vmalloc_addr;
269 struct page *page = NULL;
270 pgd_t *pgd = pgd_offset_k(addr);
271 p4d_t *p4d;
272 pud_t *pud;
273 pmd_t *pmd;
274 pte_t *ptep, pte;
275
276 /*
277 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
278 * architectures that do not vmalloc module space
279 */
280 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
281
282 if (pgd_none(*pgd))
283 return NULL;
284 p4d = p4d_offset(pgd, addr);
285 if (p4d_none(*p4d))
286 return NULL;
287 pud = pud_offset(p4d, addr);
288
289 /*
290 * Don't dereference bad PUD or PMD (below) entries. This will also
291 * identify huge mappings, which we may encounter on architectures
292 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
293 * identified as vmalloc addresses by is_vmalloc_addr(), but are
294 * not [unambiguously] associated with a struct page, so there is
295 * no correct value to return for them.
296 */
297 WARN_ON_ONCE(pud_bad(*pud));
298 if (pud_none(*pud) || pud_bad(*pud))
299 return NULL;
300 pmd = pmd_offset(pud, addr);
301 WARN_ON_ONCE(pmd_bad(*pmd));
302 if (pmd_none(*pmd) || pmd_bad(*pmd))
303 return NULL;
304
305 ptep = pte_offset_map(pmd, addr);
306 pte = *ptep;
307 if (pte_present(pte))
308 page = pte_page(pte);
309 pte_unmap(ptep);
310 return page;
311}
312EXPORT_SYMBOL(vmalloc_to_page);
313
314/*
315 * Map a vmalloc()-space virtual address to the physical page frame number.
316 */
317unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
318{
319 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
320}
321EXPORT_SYMBOL(vmalloc_to_pfn);
322
323
324/*** Global kva allocator ***/
325
326#define VM_LAZY_FREE 0x02
327#define VM_VM_AREA 0x04
328
329static DEFINE_SPINLOCK(vmap_area_lock);
330/* Export for kexec only */
331LIST_HEAD(vmap_area_list);
332static LLIST_HEAD(vmap_purge_list);
333static struct rb_root vmap_area_root = RB_ROOT;
334
335/* The vmap cache globals are protected by vmap_area_lock */
336static struct rb_node *free_vmap_cache;
337static unsigned long cached_hole_size;
338static unsigned long cached_vstart;
339static unsigned long cached_align;
340
341static unsigned long vmap_area_pcpu_hole;
342
343static struct vmap_area *__find_vmap_area(unsigned long addr)
344{
345 struct rb_node *n = vmap_area_root.rb_node;
346
347 while (n) {
348 struct vmap_area *va;
349
350 va = rb_entry(n, struct vmap_area, rb_node);
351 if (addr < va->va_start)
352 n = n->rb_left;
353 else if (addr >= va->va_end)
354 n = n->rb_right;
355 else
356 return va;
357 }
358
359 return NULL;
360}
361
362static void __insert_vmap_area(struct vmap_area *va)
363{
364 struct rb_node **p = &vmap_area_root.rb_node;
365 struct rb_node *parent = NULL;
366 struct rb_node *tmp;
367
368 while (*p) {
369 struct vmap_area *tmp_va;
370
371 parent = *p;
372 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
373 if (va->va_start < tmp_va->va_end)
374 p = &(*p)->rb_left;
375 else if (va->va_end > tmp_va->va_start)
376 p = &(*p)->rb_right;
377 else
378 BUG();
379 }
380
381 rb_link_node(&va->rb_node, parent, p);
382 rb_insert_color(&va->rb_node, &vmap_area_root);
383
384 /* address-sort this list */
385 tmp = rb_prev(&va->rb_node);
386 if (tmp) {
387 struct vmap_area *prev;
388 prev = rb_entry(tmp, struct vmap_area, rb_node);
389 list_add_rcu(&va->list, &prev->list);
390 } else
391 list_add_rcu(&va->list, &vmap_area_list);
392}
393
394static void purge_vmap_area_lazy(void);
395
396static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
397
398/*
399 * Allocate a region of KVA of the specified size and alignment, within the
400 * vstart and vend.
401 */
402static struct vmap_area *alloc_vmap_area(unsigned long size,
403 unsigned long align,
404 unsigned long vstart, unsigned long vend,
405 int node, gfp_t gfp_mask)
406{
407 struct vmap_area *va;
408 struct rb_node *n;
409 unsigned long addr;
410 int purged = 0;
411 struct vmap_area *first;
412
413 BUG_ON(!size);
414 BUG_ON(offset_in_page(size));
415 BUG_ON(!is_power_of_2(align));
416
417 might_sleep();
418
419 va = kmalloc_node(sizeof(struct vmap_area),
420 gfp_mask & GFP_RECLAIM_MASK, node);
421 if (unlikely(!va))
422 return ERR_PTR(-ENOMEM);
423
424 /*
425 * Only scan the relevant parts containing pointers to other objects
426 * to avoid false negatives.
427 */
428 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
429
430retry:
431 spin_lock(&vmap_area_lock);
432 /*
433 * Invalidate cache if we have more permissive parameters.
434 * cached_hole_size notes the largest hole noticed _below_
435 * the vmap_area cached in free_vmap_cache: if size fits
436 * into that hole, we want to scan from vstart to reuse
437 * the hole instead of allocating above free_vmap_cache.
438 * Note that __free_vmap_area may update free_vmap_cache
439 * without updating cached_hole_size or cached_align.
440 */
441 if (!free_vmap_cache ||
442 size < cached_hole_size ||
443 vstart < cached_vstart ||
444 align < cached_align) {
445nocache:
446 cached_hole_size = 0;
447 free_vmap_cache = NULL;
448 }
449 /* record if we encounter less permissive parameters */
450 cached_vstart = vstart;
451 cached_align = align;
452
453 /* find starting point for our search */
454 if (free_vmap_cache) {
455 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
456 addr = ALIGN(first->va_end, align);
457 if (addr < vstart)
458 goto nocache;
459 if (addr + size < addr)
460 goto overflow;
461
462 } else {
463 addr = ALIGN(vstart, align);
464 if (addr + size < addr)
465 goto overflow;
466
467 n = vmap_area_root.rb_node;
468 first = NULL;
469
470 while (n) {
471 struct vmap_area *tmp;
472 tmp = rb_entry(n, struct vmap_area, rb_node);
473 if (tmp->va_end >= addr) {
474 first = tmp;
475 if (tmp->va_start <= addr)
476 break;
477 n = n->rb_left;
478 } else
479 n = n->rb_right;
480 }
481
482 if (!first)
483 goto found;
484 }
485
486 /* from the starting point, walk areas until a suitable hole is found */
487 while (addr + size > first->va_start && addr + size <= vend) {
488 if (addr + cached_hole_size < first->va_start)
489 cached_hole_size = first->va_start - addr;
490 addr = ALIGN(first->va_end, align);
491 if (addr + size < addr)
492 goto overflow;
493
494 if (list_is_last(&first->list, &vmap_area_list))
495 goto found;
496
497 first = list_next_entry(first, list);
498 }
499
500found:
501 if (addr + size > vend)
502 goto overflow;
503
504 va->va_start = addr;
505 va->va_end = addr + size;
506 va->flags = 0;
507 __insert_vmap_area(va);
508 free_vmap_cache = &va->rb_node;
509 spin_unlock(&vmap_area_lock);
510
511 BUG_ON(!IS_ALIGNED(va->va_start, align));
512 BUG_ON(va->va_start < vstart);
513 BUG_ON(va->va_end > vend);
514
515 return va;
516
517overflow:
518 spin_unlock(&vmap_area_lock);
519 if (!purged) {
520 purge_vmap_area_lazy();
521 purged = 1;
522 goto retry;
523 }
524
525 if (gfpflags_allow_blocking(gfp_mask)) {
526 unsigned long freed = 0;
527 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
528 if (freed > 0) {
529 purged = 0;
530 goto retry;
531 }
532 }
533
534 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
535 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
536 size);
537 kfree(va);
538 return ERR_PTR(-EBUSY);
539}
540
541int register_vmap_purge_notifier(struct notifier_block *nb)
542{
543 return blocking_notifier_chain_register(&vmap_notify_list, nb);
544}
545EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
546
547int unregister_vmap_purge_notifier(struct notifier_block *nb)
548{
549 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
550}
551EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
552
553static void __free_vmap_area(struct vmap_area *va)
554{
555 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
556
557 if (free_vmap_cache) {
558 if (va->va_end < cached_vstart) {
559 free_vmap_cache = NULL;
560 } else {
561 struct vmap_area *cache;
562 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
563 if (va->va_start <= cache->va_start) {
564 free_vmap_cache = rb_prev(&va->rb_node);
565 /*
566 * We don't try to update cached_hole_size or
567 * cached_align, but it won't go very wrong.
568 */
569 }
570 }
571 }
572 rb_erase(&va->rb_node, &vmap_area_root);
573 RB_CLEAR_NODE(&va->rb_node);
574 list_del_rcu(&va->list);
575
576 /*
577 * Track the highest possible candidate for pcpu area
578 * allocation. Areas outside of vmalloc area can be returned
579 * here too, consider only end addresses which fall inside
580 * vmalloc area proper.
581 */
582 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
583 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
584
585 kfree_rcu(va, rcu_head);
586}
587
588/*
589 * Free a region of KVA allocated by alloc_vmap_area
590 */
591static void free_vmap_area(struct vmap_area *va)
592{
593 spin_lock(&vmap_area_lock);
594 __free_vmap_area(va);
595 spin_unlock(&vmap_area_lock);
596}
597
598/*
599 * Clear the pagetable entries of a given vmap_area
600 */
601static void unmap_vmap_area(struct vmap_area *va)
602{
603 vunmap_page_range(va->va_start, va->va_end);
604}
605
606static void vmap_debug_free_range(unsigned long start, unsigned long end)
607{
608 /*
609 * Unmap page tables and force a TLB flush immediately if pagealloc
610 * debugging is enabled. This catches use after free bugs similarly to
611 * those in linear kernel virtual address space after a page has been
612 * freed.
613 *
614 * All the lazy freeing logic is still retained, in order to minimise
615 * intrusiveness of this debugging feature.
616 *
617 * This is going to be *slow* (linear kernel virtual address debugging
618 * doesn't do a broadcast TLB flush so it is a lot faster).
619 */
620 if (debug_pagealloc_enabled()) {
621 vunmap_page_range(start, end);
622 flush_tlb_kernel_range(start, end);
623 }
624}
625
626/*
627 * lazy_max_pages is the maximum amount of virtual address space we gather up
628 * before attempting to purge with a TLB flush.
629 *
630 * There is a tradeoff here: a larger number will cover more kernel page tables
631 * and take slightly longer to purge, but it will linearly reduce the number of
632 * global TLB flushes that must be performed. It would seem natural to scale
633 * this number up linearly with the number of CPUs (because vmapping activity
634 * could also scale linearly with the number of CPUs), however it is likely
635 * that in practice, workloads might be constrained in other ways that mean
636 * vmap activity will not scale linearly with CPUs. Also, I want to be
637 * conservative and not introduce a big latency on huge systems, so go with
638 * a less aggressive log scale. It will still be an improvement over the old
639 * code, and it will be simple to change the scale factor if we find that it
640 * becomes a problem on bigger systems.
641 */
642static unsigned long lazy_max_pages(void)
643{
644 unsigned int log;
645
646 log = fls(num_online_cpus());
647
648 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
649}
650
651static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
652
653/*
654 * Serialize vmap purging. There is no actual criticial section protected
655 * by this look, but we want to avoid concurrent calls for performance
656 * reasons and to make the pcpu_get_vm_areas more deterministic.
657 */
658static DEFINE_MUTEX(vmap_purge_lock);
659
660/* for per-CPU blocks */
661static void purge_fragmented_blocks_allcpus(void);
662
663/*
664 * called before a call to iounmap() if the caller wants vm_area_struct's
665 * immediately freed.
666 */
667void set_iounmap_nonlazy(void)
668{
669 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
670}
671
672/*
673 * Purges all lazily-freed vmap areas.
674 */
675static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
676{
677 struct llist_node *valist;
678 struct vmap_area *va;
679 struct vmap_area *n_va;
680 bool do_free = false;
681
682 lockdep_assert_held(&vmap_purge_lock);
683
684 valist = llist_del_all(&vmap_purge_list);
685 llist_for_each_entry(va, valist, purge_list) {
686 if (va->va_start < start)
687 start = va->va_start;
688 if (va->va_end > end)
689 end = va->va_end;
690 do_free = true;
691 }
692
693 if (!do_free)
694 return false;
695
696 flush_tlb_kernel_range(start, end);
697
698 spin_lock(&vmap_area_lock);
699 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
700 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
701
702 __free_vmap_area(va);
703 atomic_sub(nr, &vmap_lazy_nr);
704 cond_resched_lock(&vmap_area_lock);
705 }
706 spin_unlock(&vmap_area_lock);
707 return true;
708}
709
710/*
711 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
712 * is already purging.
713 */
714static void try_purge_vmap_area_lazy(void)
715{
716 if (mutex_trylock(&vmap_purge_lock)) {
717 __purge_vmap_area_lazy(ULONG_MAX, 0);
718 mutex_unlock(&vmap_purge_lock);
719 }
720}
721
722/*
723 * Kick off a purge of the outstanding lazy areas.
724 */
725static void purge_vmap_area_lazy(void)
726{
727 mutex_lock(&vmap_purge_lock);
728 purge_fragmented_blocks_allcpus();
729 __purge_vmap_area_lazy(ULONG_MAX, 0);
730 mutex_unlock(&vmap_purge_lock);
731}
732
733/*
734 * Free a vmap area, caller ensuring that the area has been unmapped
735 * and flush_cache_vunmap had been called for the correct range
736 * previously.
737 */
738static void free_vmap_area_noflush(struct vmap_area *va)
739{
740 int nr_lazy;
741
742 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
743 &vmap_lazy_nr);
744
745 /* After this point, we may free va at any time */
746 llist_add(&va->purge_list, &vmap_purge_list);
747
748 if (unlikely(nr_lazy > lazy_max_pages()))
749 try_purge_vmap_area_lazy();
750}
751
752/*
753 * Free and unmap a vmap area
754 */
755static void free_unmap_vmap_area(struct vmap_area *va)
756{
757 flush_cache_vunmap(va->va_start, va->va_end);
758 unmap_vmap_area(va);
759 free_vmap_area_noflush(va);
760}
761
762static struct vmap_area *find_vmap_area(unsigned long addr)
763{
764 struct vmap_area *va;
765
766 spin_lock(&vmap_area_lock);
767 va = __find_vmap_area(addr);
768 spin_unlock(&vmap_area_lock);
769
770 return va;
771}
772
773/*** Per cpu kva allocator ***/
774
775/*
776 * vmap space is limited especially on 32 bit architectures. Ensure there is
777 * room for at least 16 percpu vmap blocks per CPU.
778 */
779/*
780 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
781 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
782 * instead (we just need a rough idea)
783 */
784#if BITS_PER_LONG == 32
785#define VMALLOC_SPACE (128UL*1024*1024)
786#else
787#define VMALLOC_SPACE (128UL*1024*1024*1024)
788#endif
789
790#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
791#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
792#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
793#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
794#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
795#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
796#define VMAP_BBMAP_BITS \
797 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
798 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
799 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
800
801#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
802
803static bool vmap_initialized __read_mostly = false;
804
805struct vmap_block_queue {
806 spinlock_t lock;
807 struct list_head free;
808};
809
810struct vmap_block {
811 spinlock_t lock;
812 struct vmap_area *va;
813 unsigned long free, dirty;
814 unsigned long dirty_min, dirty_max; /*< dirty range */
815 struct list_head free_list;
816 struct rcu_head rcu_head;
817 struct list_head purge;
818};
819
820/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
821static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
822
823/*
824 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
825 * in the free path. Could get rid of this if we change the API to return a
826 * "cookie" from alloc, to be passed to free. But no big deal yet.
827 */
828static DEFINE_SPINLOCK(vmap_block_tree_lock);
829static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
830
831/*
832 * We should probably have a fallback mechanism to allocate virtual memory
833 * out of partially filled vmap blocks. However vmap block sizing should be
834 * fairly reasonable according to the vmalloc size, so it shouldn't be a
835 * big problem.
836 */
837
838static unsigned long addr_to_vb_idx(unsigned long addr)
839{
840 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
841 addr /= VMAP_BLOCK_SIZE;
842 return addr;
843}
844
845static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
846{
847 unsigned long addr;
848
849 addr = va_start + (pages_off << PAGE_SHIFT);
850 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
851 return (void *)addr;
852}
853
854/**
855 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
856 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
857 * @order: how many 2^order pages should be occupied in newly allocated block
858 * @gfp_mask: flags for the page level allocator
859 *
860 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
861 */
862static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
863{
864 struct vmap_block_queue *vbq;
865 struct vmap_block *vb;
866 struct vmap_area *va;
867 unsigned long vb_idx;
868 int node, err;
869 void *vaddr;
870
871 node = numa_node_id();
872
873 vb = kmalloc_node(sizeof(struct vmap_block),
874 gfp_mask & GFP_RECLAIM_MASK, node);
875 if (unlikely(!vb))
876 return ERR_PTR(-ENOMEM);
877
878 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
879 VMALLOC_START, VMALLOC_END,
880 node, gfp_mask);
881 if (IS_ERR(va)) {
882 kfree(vb);
883 return ERR_CAST(va);
884 }
885
886 err = radix_tree_preload(gfp_mask);
887 if (unlikely(err)) {
888 kfree(vb);
889 free_vmap_area(va);
890 return ERR_PTR(err);
891 }
892
893 vaddr = vmap_block_vaddr(va->va_start, 0);
894 spin_lock_init(&vb->lock);
895 vb->va = va;
896 /* At least something should be left free */
897 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
898 vb->free = VMAP_BBMAP_BITS - (1UL << order);
899 vb->dirty = 0;
900 vb->dirty_min = VMAP_BBMAP_BITS;
901 vb->dirty_max = 0;
902 INIT_LIST_HEAD(&vb->free_list);
903
904 vb_idx = addr_to_vb_idx(va->va_start);
905 spin_lock(&vmap_block_tree_lock);
906 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
907 spin_unlock(&vmap_block_tree_lock);
908 BUG_ON(err);
909 radix_tree_preload_end();
910
911 vbq = &get_cpu_var(vmap_block_queue);
912 spin_lock(&vbq->lock);
913 list_add_tail_rcu(&vb->free_list, &vbq->free);
914 spin_unlock(&vbq->lock);
915 put_cpu_var(vmap_block_queue);
916
917 return vaddr;
918}
919
920static void free_vmap_block(struct vmap_block *vb)
921{
922 struct vmap_block *tmp;
923 unsigned long vb_idx;
924
925 vb_idx = addr_to_vb_idx(vb->va->va_start);
926 spin_lock(&vmap_block_tree_lock);
927 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
928 spin_unlock(&vmap_block_tree_lock);
929 BUG_ON(tmp != vb);
930
931 free_vmap_area_noflush(vb->va);
932 kfree_rcu(vb, rcu_head);
933}
934
935static void purge_fragmented_blocks(int cpu)
936{
937 LIST_HEAD(purge);
938 struct vmap_block *vb;
939 struct vmap_block *n_vb;
940 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
941
942 rcu_read_lock();
943 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
944
945 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
946 continue;
947
948 spin_lock(&vb->lock);
949 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
950 vb->free = 0; /* prevent further allocs after releasing lock */
951 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
952 vb->dirty_min = 0;
953 vb->dirty_max = VMAP_BBMAP_BITS;
954 spin_lock(&vbq->lock);
955 list_del_rcu(&vb->free_list);
956 spin_unlock(&vbq->lock);
957 spin_unlock(&vb->lock);
958 list_add_tail(&vb->purge, &purge);
959 } else
960 spin_unlock(&vb->lock);
961 }
962 rcu_read_unlock();
963
964 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
965 list_del(&vb->purge);
966 free_vmap_block(vb);
967 }
968}
969
970static void purge_fragmented_blocks_allcpus(void)
971{
972 int cpu;
973
974 for_each_possible_cpu(cpu)
975 purge_fragmented_blocks(cpu);
976}
977
978static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
979{
980 struct vmap_block_queue *vbq;
981 struct vmap_block *vb;
982 void *vaddr = NULL;
983 unsigned int order;
984
985 BUG_ON(offset_in_page(size));
986 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
987 if (WARN_ON(size == 0)) {
988 /*
989 * Allocating 0 bytes isn't what caller wants since
990 * get_order(0) returns funny result. Just warn and terminate
991 * early.
992 */
993 return NULL;
994 }
995 order = get_order(size);
996
997 rcu_read_lock();
998 vbq = &get_cpu_var(vmap_block_queue);
999 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1000 unsigned long pages_off;
1001
1002 spin_lock(&vb->lock);
1003 if (vb->free < (1UL << order)) {
1004 spin_unlock(&vb->lock);
1005 continue;
1006 }
1007
1008 pages_off = VMAP_BBMAP_BITS - vb->free;
1009 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1010 vb->free -= 1UL << order;
1011 if (vb->free == 0) {
1012 spin_lock(&vbq->lock);
1013 list_del_rcu(&vb->free_list);
1014 spin_unlock(&vbq->lock);
1015 }
1016
1017 spin_unlock(&vb->lock);
1018 break;
1019 }
1020
1021 put_cpu_var(vmap_block_queue);
1022 rcu_read_unlock();
1023
1024 /* Allocate new block if nothing was found */
1025 if (!vaddr)
1026 vaddr = new_vmap_block(order, gfp_mask);
1027
1028 return vaddr;
1029}
1030
1031static void vb_free(const void *addr, unsigned long size)
1032{
1033 unsigned long offset;
1034 unsigned long vb_idx;
1035 unsigned int order;
1036 struct vmap_block *vb;
1037
1038 BUG_ON(offset_in_page(size));
1039 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1040
1041 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1042
1043 order = get_order(size);
1044
1045 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1046 offset >>= PAGE_SHIFT;
1047
1048 vb_idx = addr_to_vb_idx((unsigned long)addr);
1049 rcu_read_lock();
1050 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1051 rcu_read_unlock();
1052 BUG_ON(!vb);
1053
1054 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1055
1056 spin_lock(&vb->lock);
1057
1058 /* Expand dirty range */
1059 vb->dirty_min = min(vb->dirty_min, offset);
1060 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1061
1062 vb->dirty += 1UL << order;
1063 if (vb->dirty == VMAP_BBMAP_BITS) {
1064 BUG_ON(vb->free);
1065 spin_unlock(&vb->lock);
1066 free_vmap_block(vb);
1067 } else
1068 spin_unlock(&vb->lock);
1069}
1070
1071/**
1072 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1073 *
1074 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1075 * to amortize TLB flushing overheads. What this means is that any page you
1076 * have now, may, in a former life, have been mapped into kernel virtual
1077 * address by the vmap layer and so there might be some CPUs with TLB entries
1078 * still referencing that page (additional to the regular 1:1 kernel mapping).
1079 *
1080 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1081 * be sure that none of the pages we have control over will have any aliases
1082 * from the vmap layer.
1083 */
1084void vm_unmap_aliases(void)
1085{
1086 unsigned long start = ULONG_MAX, end = 0;
1087 int cpu;
1088 int flush = 0;
1089
1090 if (unlikely(!vmap_initialized))
1091 return;
1092
1093 might_sleep();
1094
1095 for_each_possible_cpu(cpu) {
1096 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1097 struct vmap_block *vb;
1098
1099 rcu_read_lock();
1100 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1101 spin_lock(&vb->lock);
1102 if (vb->dirty) {
1103 unsigned long va_start = vb->va->va_start;
1104 unsigned long s, e;
1105
1106 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1107 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1108
1109 start = min(s, start);
1110 end = max(e, end);
1111
1112 flush = 1;
1113 }
1114 spin_unlock(&vb->lock);
1115 }
1116 rcu_read_unlock();
1117 }
1118
1119 mutex_lock(&vmap_purge_lock);
1120 purge_fragmented_blocks_allcpus();
1121 if (!__purge_vmap_area_lazy(start, end) && flush)
1122 flush_tlb_kernel_range(start, end);
1123 mutex_unlock(&vmap_purge_lock);
1124}
1125EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1126
1127/**
1128 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1129 * @mem: the pointer returned by vm_map_ram
1130 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1131 */
1132void vm_unmap_ram(const void *mem, unsigned int count)
1133{
1134 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1135 unsigned long addr = (unsigned long)mem;
1136 struct vmap_area *va;
1137
1138 might_sleep();
1139 BUG_ON(!addr);
1140 BUG_ON(addr < VMALLOC_START);
1141 BUG_ON(addr > VMALLOC_END);
1142 BUG_ON(!PAGE_ALIGNED(addr));
1143
1144 debug_check_no_locks_freed(mem, size);
1145 vmap_debug_free_range(addr, addr+size);
1146
1147 if (likely(count <= VMAP_MAX_ALLOC)) {
1148 vb_free(mem, size);
1149 return;
1150 }
1151
1152 va = find_vmap_area(addr);
1153 BUG_ON(!va);
1154 free_unmap_vmap_area(va);
1155}
1156EXPORT_SYMBOL(vm_unmap_ram);
1157
1158/**
1159 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1160 * @pages: an array of pointers to the pages to be mapped
1161 * @count: number of pages
1162 * @node: prefer to allocate data structures on this node
1163 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1164 *
1165 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1166 * faster than vmap so it's good. But if you mix long-life and short-life
1167 * objects with vm_map_ram(), it could consume lots of address space through
1168 * fragmentation (especially on a 32bit machine). You could see failures in
1169 * the end. Please use this function for short-lived objects.
1170 *
1171 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1172 */
1173void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1174{
1175 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1176 unsigned long addr;
1177 void *mem;
1178
1179 if (likely(count <= VMAP_MAX_ALLOC)) {
1180 mem = vb_alloc(size, GFP_KERNEL);
1181 if (IS_ERR(mem))
1182 return NULL;
1183 addr = (unsigned long)mem;
1184 } else {
1185 struct vmap_area *va;
1186 va = alloc_vmap_area(size, PAGE_SIZE,
1187 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1188 if (IS_ERR(va))
1189 return NULL;
1190
1191 addr = va->va_start;
1192 mem = (void *)addr;
1193 }
1194 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1195 vm_unmap_ram(mem, count);
1196 return NULL;
1197 }
1198 return mem;
1199}
1200EXPORT_SYMBOL(vm_map_ram);
1201
1202static struct vm_struct *vmlist __initdata;
1203/**
1204 * vm_area_add_early - add vmap area early during boot
1205 * @vm: vm_struct to add
1206 *
1207 * This function is used to add fixed kernel vm area to vmlist before
1208 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1209 * should contain proper values and the other fields should be zero.
1210 *
1211 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1212 */
1213void __init vm_area_add_early(struct vm_struct *vm)
1214{
1215 struct vm_struct *tmp, **p;
1216
1217 BUG_ON(vmap_initialized);
1218 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1219 if (tmp->addr >= vm->addr) {
1220 BUG_ON(tmp->addr < vm->addr + vm->size);
1221 break;
1222 } else
1223 BUG_ON(tmp->addr + tmp->size > vm->addr);
1224 }
1225 vm->next = *p;
1226 *p = vm;
1227}
1228
1229/**
1230 * vm_area_register_early - register vmap area early during boot
1231 * @vm: vm_struct to register
1232 * @align: requested alignment
1233 *
1234 * This function is used to register kernel vm area before
1235 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1236 * proper values on entry and other fields should be zero. On return,
1237 * vm->addr contains the allocated address.
1238 *
1239 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1240 */
1241void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1242{
1243 static size_t vm_init_off __initdata;
1244 unsigned long addr;
1245
1246 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1247 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1248
1249 vm->addr = (void *)addr;
1250
1251 vm_area_add_early(vm);
1252}
1253
1254void __init vmalloc_init(void)
1255{
1256 struct vmap_area *va;
1257 struct vm_struct *tmp;
1258 int i;
1259
1260 for_each_possible_cpu(i) {
1261 struct vmap_block_queue *vbq;
1262 struct vfree_deferred *p;
1263
1264 vbq = &per_cpu(vmap_block_queue, i);
1265 spin_lock_init(&vbq->lock);
1266 INIT_LIST_HEAD(&vbq->free);
1267 p = &per_cpu(vfree_deferred, i);
1268 init_llist_head(&p->list);
1269 INIT_WORK(&p->wq, free_work);
1270 }
1271
1272 /* Import existing vmlist entries. */
1273 for (tmp = vmlist; tmp; tmp = tmp->next) {
1274 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1275 va->flags = VM_VM_AREA;
1276 va->va_start = (unsigned long)tmp->addr;
1277 va->va_end = va->va_start + tmp->size;
1278 va->vm = tmp;
1279 __insert_vmap_area(va);
1280 }
1281
1282 vmap_area_pcpu_hole = VMALLOC_END;
1283
1284 vmap_initialized = true;
1285}
1286
1287/**
1288 * map_kernel_range_noflush - map kernel VM area with the specified pages
1289 * @addr: start of the VM area to map
1290 * @size: size of the VM area to map
1291 * @prot: page protection flags to use
1292 * @pages: pages to map
1293 *
1294 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1295 * specify should have been allocated using get_vm_area() and its
1296 * friends.
1297 *
1298 * NOTE:
1299 * This function does NOT do any cache flushing. The caller is
1300 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1301 * before calling this function.
1302 *
1303 * RETURNS:
1304 * The number of pages mapped on success, -errno on failure.
1305 */
1306int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1307 pgprot_t prot, struct page **pages)
1308{
1309 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1310}
1311
1312/**
1313 * unmap_kernel_range_noflush - unmap kernel VM area
1314 * @addr: start of the VM area to unmap
1315 * @size: size of the VM area to unmap
1316 *
1317 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1318 * specify should have been allocated using get_vm_area() and its
1319 * friends.
1320 *
1321 * NOTE:
1322 * This function does NOT do any cache flushing. The caller is
1323 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1324 * before calling this function and flush_tlb_kernel_range() after.
1325 */
1326void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1327{
1328 vunmap_page_range(addr, addr + size);
1329}
1330EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1331
1332/**
1333 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1334 * @addr: start of the VM area to unmap
1335 * @size: size of the VM area to unmap
1336 *
1337 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1338 * the unmapping and tlb after.
1339 */
1340void unmap_kernel_range(unsigned long addr, unsigned long size)
1341{
1342 unsigned long end = addr + size;
1343
1344 flush_cache_vunmap(addr, end);
1345 vunmap_page_range(addr, end);
1346 flush_tlb_kernel_range(addr, end);
1347}
1348EXPORT_SYMBOL_GPL(unmap_kernel_range);
1349
1350int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1351{
1352 unsigned long addr = (unsigned long)area->addr;
1353 unsigned long end = addr + get_vm_area_size(area);
1354 int err;
1355
1356 err = vmap_page_range(addr, end, prot, pages);
1357
1358 return err > 0 ? 0 : err;
1359}
1360EXPORT_SYMBOL_GPL(map_vm_area);
1361
1362static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1363 unsigned long flags, const void *caller)
1364{
1365 spin_lock(&vmap_area_lock);
1366 vm->flags = flags;
1367 vm->addr = (void *)va->va_start;
1368 vm->size = va->va_end - va->va_start;
1369 vm->caller = caller;
1370 va->vm = vm;
1371 va->flags |= VM_VM_AREA;
1372 spin_unlock(&vmap_area_lock);
1373}
1374
1375static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1376{
1377 /*
1378 * Before removing VM_UNINITIALIZED,
1379 * we should make sure that vm has proper values.
1380 * Pair with smp_rmb() in show_numa_info().
1381 */
1382 smp_wmb();
1383 vm->flags &= ~VM_UNINITIALIZED;
1384}
1385
1386static struct vm_struct *__get_vm_area_node(unsigned long size,
1387 unsigned long align, unsigned long flags, unsigned long start,
1388 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1389{
1390 struct vmap_area *va;
1391 struct vm_struct *area;
1392
1393 BUG_ON(in_interrupt());
1394 size = PAGE_ALIGN(size);
1395 if (unlikely(!size))
1396 return NULL;
1397
1398 if (flags & VM_IOREMAP)
1399 align = 1ul << clamp_t(int, get_count_order_long(size),
1400 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1401
1402 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1403 if (unlikely(!area))
1404 return NULL;
1405
1406 if (!(flags & VM_NO_GUARD))
1407 size += PAGE_SIZE;
1408
1409 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1410 if (IS_ERR(va)) {
1411 kfree(area);
1412 return NULL;
1413 }
1414
1415 setup_vmalloc_vm(area, va, flags, caller);
1416
1417 return area;
1418}
1419
1420struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1421 unsigned long start, unsigned long end)
1422{
1423 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1424 GFP_KERNEL, __builtin_return_address(0));
1425}
1426EXPORT_SYMBOL_GPL(__get_vm_area);
1427
1428struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1429 unsigned long start, unsigned long end,
1430 const void *caller)
1431{
1432 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1433 GFP_KERNEL, caller);
1434}
1435
1436/**
1437 * get_vm_area - reserve a contiguous kernel virtual area
1438 * @size: size of the area
1439 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1440 *
1441 * Search an area of @size in the kernel virtual mapping area,
1442 * and reserved it for out purposes. Returns the area descriptor
1443 * on success or %NULL on failure.
1444 */
1445struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1446{
1447 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1448 NUMA_NO_NODE, GFP_KERNEL,
1449 __builtin_return_address(0));
1450}
1451
1452struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1453 const void *caller)
1454{
1455 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1456 NUMA_NO_NODE, GFP_KERNEL, caller);
1457}
1458
1459/**
1460 * find_vm_area - find a continuous kernel virtual area
1461 * @addr: base address
1462 *
1463 * Search for the kernel VM area starting at @addr, and return it.
1464 * It is up to the caller to do all required locking to keep the returned
1465 * pointer valid.
1466 */
1467struct vm_struct *find_vm_area(const void *addr)
1468{
1469 struct vmap_area *va;
1470
1471 va = find_vmap_area((unsigned long)addr);
1472 if (va && va->flags & VM_VM_AREA)
1473 return va->vm;
1474
1475 return NULL;
1476}
1477
1478/**
1479 * remove_vm_area - find and remove a continuous kernel virtual area
1480 * @addr: base address
1481 *
1482 * Search for the kernel VM area starting at @addr, and remove it.
1483 * This function returns the found VM area, but using it is NOT safe
1484 * on SMP machines, except for its size or flags.
1485 */
1486struct vm_struct *remove_vm_area(const void *addr)
1487{
1488 struct vmap_area *va;
1489
1490 might_sleep();
1491
1492 va = find_vmap_area((unsigned long)addr);
1493 if (va && va->flags & VM_VM_AREA) {
1494 struct vm_struct *vm = va->vm;
1495
1496 spin_lock(&vmap_area_lock);
1497 va->vm = NULL;
1498 va->flags &= ~VM_VM_AREA;
1499 va->flags |= VM_LAZY_FREE;
1500 spin_unlock(&vmap_area_lock);
1501
1502 vmap_debug_free_range(va->va_start, va->va_end);
1503 kasan_free_shadow(vm);
1504 free_unmap_vmap_area(va);
1505
1506 return vm;
1507 }
1508 return NULL;
1509}
1510
1511static void __vunmap(const void *addr, int deallocate_pages)
1512{
1513 struct vm_struct *area;
1514
1515 if (!addr)
1516 return;
1517
1518 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1519 addr))
1520 return;
1521
1522 area = remove_vm_area(addr);
1523 if (unlikely(!area)) {
1524 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1525 addr);
1526 return;
1527 }
1528
1529 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1530 debug_check_no_obj_freed(addr, get_vm_area_size(area));
1531
1532 if (deallocate_pages) {
1533 int i;
1534
1535 for (i = 0; i < area->nr_pages; i++) {
1536 struct page *page = area->pages[i];
1537
1538 BUG_ON(!page);
1539 __free_pages(page, 0);
1540 }
1541
1542 kvfree(area->pages);
1543 }
1544
1545 kfree(area);
1546 return;
1547}
1548
1549static inline void __vfree_deferred(const void *addr)
1550{
1551 /*
1552 * Use raw_cpu_ptr() because this can be called from preemptible
1553 * context. Preemption is absolutely fine here, because the llist_add()
1554 * implementation is lockless, so it works even if we are adding to
1555 * nother cpu's list. schedule_work() should be fine with this too.
1556 */
1557 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1558
1559 if (llist_add((struct llist_node *)addr, &p->list))
1560 schedule_work(&p->wq);
1561}
1562
1563/**
1564 * vfree_atomic - release memory allocated by vmalloc()
1565 * @addr: memory base address
1566 *
1567 * This one is just like vfree() but can be called in any atomic context
1568 * except NMIs.
1569 */
1570void vfree_atomic(const void *addr)
1571{
1572 BUG_ON(in_nmi());
1573
1574 kmemleak_free(addr);
1575
1576 if (!addr)
1577 return;
1578 __vfree_deferred(addr);
1579}
1580
1581/**
1582 * vfree - release memory allocated by vmalloc()
1583 * @addr: memory base address
1584 *
1585 * Free the virtually continuous memory area starting at @addr, as
1586 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1587 * NULL, no operation is performed.
1588 *
1589 * Must not be called in NMI context (strictly speaking, only if we don't
1590 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1591 * conventions for vfree() arch-depenedent would be a really bad idea)
1592 *
1593 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1594 */
1595void vfree(const void *addr)
1596{
1597 BUG_ON(in_nmi());
1598
1599 kmemleak_free(addr);
1600
1601 if (!addr)
1602 return;
1603 if (unlikely(in_interrupt()))
1604 __vfree_deferred(addr);
1605 else
1606 __vunmap(addr, 1);
1607}
1608EXPORT_SYMBOL(vfree);
1609
1610/**
1611 * vunmap - release virtual mapping obtained by vmap()
1612 * @addr: memory base address
1613 *
1614 * Free the virtually contiguous memory area starting at @addr,
1615 * which was created from the page array passed to vmap().
1616 *
1617 * Must not be called in interrupt context.
1618 */
1619void vunmap(const void *addr)
1620{
1621 BUG_ON(in_interrupt());
1622 might_sleep();
1623 if (addr)
1624 __vunmap(addr, 0);
1625}
1626EXPORT_SYMBOL(vunmap);
1627
1628/**
1629 * vmap - map an array of pages into virtually contiguous space
1630 * @pages: array of page pointers
1631 * @count: number of pages to map
1632 * @flags: vm_area->flags
1633 * @prot: page protection for the mapping
1634 *
1635 * Maps @count pages from @pages into contiguous kernel virtual
1636 * space.
1637 */
1638void *vmap(struct page **pages, unsigned int count,
1639 unsigned long flags, pgprot_t prot)
1640{
1641 struct vm_struct *area;
1642 unsigned long size; /* In bytes */
1643
1644 might_sleep();
1645
1646 if (count > totalram_pages)
1647 return NULL;
1648
1649 size = (unsigned long)count << PAGE_SHIFT;
1650 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1651 if (!area)
1652 return NULL;
1653
1654 if (map_vm_area(area, prot, pages)) {
1655 vunmap(area->addr);
1656 return NULL;
1657 }
1658
1659 return area->addr;
1660}
1661EXPORT_SYMBOL(vmap);
1662
1663static void *__vmalloc_node(unsigned long size, unsigned long align,
1664 gfp_t gfp_mask, pgprot_t prot,
1665 int node, const void *caller);
1666static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1667 pgprot_t prot, int node)
1668{
1669 struct page **pages;
1670 unsigned int nr_pages, array_size, i;
1671 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1672 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1673 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
1674 0 :
1675 __GFP_HIGHMEM;
1676
1677 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1678 array_size = (nr_pages * sizeof(struct page *));
1679
1680 area->nr_pages = nr_pages;
1681 /* Please note that the recursion is strictly bounded. */
1682 if (array_size > PAGE_SIZE) {
1683 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
1684 PAGE_KERNEL, node, area->caller);
1685 } else {
1686 pages = kmalloc_node(array_size, nested_gfp, node);
1687 }
1688 area->pages = pages;
1689 if (!area->pages) {
1690 remove_vm_area(area->addr);
1691 kfree(area);
1692 return NULL;
1693 }
1694
1695 for (i = 0; i < area->nr_pages; i++) {
1696 struct page *page;
1697
1698 if (node == NUMA_NO_NODE)
1699 page = alloc_page(alloc_mask|highmem_mask);
1700 else
1701 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
1702
1703 if (unlikely(!page)) {
1704 /* Successfully allocated i pages, free them in __vunmap() */
1705 area->nr_pages = i;
1706 goto fail;
1707 }
1708 area->pages[i] = page;
1709 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
1710 cond_resched();
1711 }
1712
1713 if (map_vm_area(area, prot, pages))
1714 goto fail;
1715 return area->addr;
1716
1717fail:
1718 warn_alloc(gfp_mask, NULL,
1719 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1720 (area->nr_pages*PAGE_SIZE), area->size);
1721 vfree(area->addr);
1722 return NULL;
1723}
1724
1725/**
1726 * __vmalloc_node_range - allocate virtually contiguous memory
1727 * @size: allocation size
1728 * @align: desired alignment
1729 * @start: vm area range start
1730 * @end: vm area range end
1731 * @gfp_mask: flags for the page level allocator
1732 * @prot: protection mask for the allocated pages
1733 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1734 * @node: node to use for allocation or NUMA_NO_NODE
1735 * @caller: caller's return address
1736 *
1737 * Allocate enough pages to cover @size from the page level
1738 * allocator with @gfp_mask flags. Map them into contiguous
1739 * kernel virtual space, using a pagetable protection of @prot.
1740 */
1741void *__vmalloc_node_range(unsigned long size, unsigned long align,
1742 unsigned long start, unsigned long end, gfp_t gfp_mask,
1743 pgprot_t prot, unsigned long vm_flags, int node,
1744 const void *caller)
1745{
1746 struct vm_struct *area;
1747 void *addr;
1748 unsigned long real_size = size;
1749
1750 size = PAGE_ALIGN(size);
1751 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1752 goto fail;
1753
1754 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1755 vm_flags, start, end, node, gfp_mask, caller);
1756 if (!area)
1757 goto fail;
1758
1759 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1760 if (!addr)
1761 return NULL;
1762
1763 /*
1764 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1765 * flag. It means that vm_struct is not fully initialized.
1766 * Now, it is fully initialized, so remove this flag here.
1767 */
1768 clear_vm_uninitialized_flag(area);
1769
1770 kmemleak_vmalloc(area, size, gfp_mask);
1771
1772 return addr;
1773
1774fail:
1775 warn_alloc(gfp_mask, NULL,
1776 "vmalloc: allocation failure: %lu bytes", real_size);
1777 return NULL;
1778}
1779
1780/**
1781 * __vmalloc_node - allocate virtually contiguous memory
1782 * @size: allocation size
1783 * @align: desired alignment
1784 * @gfp_mask: flags for the page level allocator
1785 * @prot: protection mask for the allocated pages
1786 * @node: node to use for allocation or NUMA_NO_NODE
1787 * @caller: caller's return address
1788 *
1789 * Allocate enough pages to cover @size from the page level
1790 * allocator with @gfp_mask flags. Map them into contiguous
1791 * kernel virtual space, using a pagetable protection of @prot.
1792 *
1793 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1794 * and __GFP_NOFAIL are not supported
1795 *
1796 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1797 * with mm people.
1798 *
1799 */
1800static void *__vmalloc_node(unsigned long size, unsigned long align,
1801 gfp_t gfp_mask, pgprot_t prot,
1802 int node, const void *caller)
1803{
1804 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1805 gfp_mask, prot, 0, node, caller);
1806}
1807
1808void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1809{
1810 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1811 __builtin_return_address(0));
1812}
1813EXPORT_SYMBOL(__vmalloc);
1814
1815static inline void *__vmalloc_node_flags(unsigned long size,
1816 int node, gfp_t flags)
1817{
1818 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1819 node, __builtin_return_address(0));
1820}
1821
1822
1823void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
1824 void *caller)
1825{
1826 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
1827}
1828
1829/**
1830 * vmalloc - allocate virtually contiguous memory
1831 * @size: allocation size
1832 * Allocate enough pages to cover @size from the page level
1833 * allocator and map them into contiguous kernel virtual space.
1834 *
1835 * For tight control over page level allocator and protection flags
1836 * use __vmalloc() instead.
1837 */
1838void *vmalloc(unsigned long size)
1839{
1840 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1841 GFP_KERNEL);
1842}
1843EXPORT_SYMBOL(vmalloc);
1844
1845/**
1846 * vzalloc - allocate virtually contiguous memory with zero fill
1847 * @size: allocation size
1848 * Allocate enough pages to cover @size from the page level
1849 * allocator and map them into contiguous kernel virtual space.
1850 * The memory allocated is set to zero.
1851 *
1852 * For tight control over page level allocator and protection flags
1853 * use __vmalloc() instead.
1854 */
1855void *vzalloc(unsigned long size)
1856{
1857 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1858 GFP_KERNEL | __GFP_ZERO);
1859}
1860EXPORT_SYMBOL(vzalloc);
1861
1862/**
1863 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1864 * @size: allocation size
1865 *
1866 * The resulting memory area is zeroed so it can be mapped to userspace
1867 * without leaking data.
1868 */
1869void *vmalloc_user(unsigned long size)
1870{
1871 struct vm_struct *area;
1872 void *ret;
1873
1874 ret = __vmalloc_node(size, SHMLBA,
1875 GFP_KERNEL | __GFP_ZERO,
1876 PAGE_KERNEL, NUMA_NO_NODE,
1877 __builtin_return_address(0));
1878 if (ret) {
1879 area = find_vm_area(ret);
1880 area->flags |= VM_USERMAP;
1881 }
1882 return ret;
1883}
1884EXPORT_SYMBOL(vmalloc_user);
1885
1886/**
1887 * vmalloc_node - allocate memory on a specific node
1888 * @size: allocation size
1889 * @node: numa node
1890 *
1891 * Allocate enough pages to cover @size from the page level
1892 * allocator and map them into contiguous kernel virtual space.
1893 *
1894 * For tight control over page level allocator and protection flags
1895 * use __vmalloc() instead.
1896 */
1897void *vmalloc_node(unsigned long size, int node)
1898{
1899 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
1900 node, __builtin_return_address(0));
1901}
1902EXPORT_SYMBOL(vmalloc_node);
1903
1904/**
1905 * vzalloc_node - allocate memory on a specific node with zero fill
1906 * @size: allocation size
1907 * @node: numa node
1908 *
1909 * Allocate enough pages to cover @size from the page level
1910 * allocator and map them into contiguous kernel virtual space.
1911 * The memory allocated is set to zero.
1912 *
1913 * For tight control over page level allocator and protection flags
1914 * use __vmalloc_node() instead.
1915 */
1916void *vzalloc_node(unsigned long size, int node)
1917{
1918 return __vmalloc_node_flags(size, node,
1919 GFP_KERNEL | __GFP_ZERO);
1920}
1921EXPORT_SYMBOL(vzalloc_node);
1922
1923#ifndef PAGE_KERNEL_EXEC
1924# define PAGE_KERNEL_EXEC PAGE_KERNEL
1925#endif
1926
1927/**
1928 * vmalloc_exec - allocate virtually contiguous, executable memory
1929 * @size: allocation size
1930 *
1931 * Kernel-internal function to allocate enough pages to cover @size
1932 * the page level allocator and map them into contiguous and
1933 * executable kernel virtual space.
1934 *
1935 * For tight control over page level allocator and protection flags
1936 * use __vmalloc() instead.
1937 */
1938
1939void *vmalloc_exec(unsigned long size)
1940{
1941 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL_EXEC,
1942 NUMA_NO_NODE, __builtin_return_address(0));
1943}
1944
1945#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1946#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
1947#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1948#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
1949#else
1950/*
1951 * 64b systems should always have either DMA or DMA32 zones. For others
1952 * GFP_DMA32 should do the right thing and use the normal zone.
1953 */
1954#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1955#endif
1956
1957/**
1958 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1959 * @size: allocation size
1960 *
1961 * Allocate enough 32bit PA addressable pages to cover @size from the
1962 * page level allocator and map them into contiguous kernel virtual space.
1963 */
1964void *vmalloc_32(unsigned long size)
1965{
1966 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1967 NUMA_NO_NODE, __builtin_return_address(0));
1968}
1969EXPORT_SYMBOL(vmalloc_32);
1970
1971/**
1972 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1973 * @size: allocation size
1974 *
1975 * The resulting memory area is 32bit addressable and zeroed so it can be
1976 * mapped to userspace without leaking data.
1977 */
1978void *vmalloc_32_user(unsigned long size)
1979{
1980 struct vm_struct *area;
1981 void *ret;
1982
1983 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1984 NUMA_NO_NODE, __builtin_return_address(0));
1985 if (ret) {
1986 area = find_vm_area(ret);
1987 area->flags |= VM_USERMAP;
1988 }
1989 return ret;
1990}
1991EXPORT_SYMBOL(vmalloc_32_user);
1992
1993/*
1994 * small helper routine , copy contents to buf from addr.
1995 * If the page is not present, fill zero.
1996 */
1997
1998static int aligned_vread(char *buf, char *addr, unsigned long count)
1999{
2000 struct page *p;
2001 int copied = 0;
2002
2003 while (count) {
2004 unsigned long offset, length;
2005
2006 offset = offset_in_page(addr);
2007 length = PAGE_SIZE - offset;
2008 if (length > count)
2009 length = count;
2010 p = vmalloc_to_page(addr);
2011 /*
2012 * To do safe access to this _mapped_ area, we need
2013 * lock. But adding lock here means that we need to add
2014 * overhead of vmalloc()/vfree() calles for this _debug_
2015 * interface, rarely used. Instead of that, we'll use
2016 * kmap() and get small overhead in this access function.
2017 */
2018 if (p) {
2019 /*
2020 * we can expect USER0 is not used (see vread/vwrite's
2021 * function description)
2022 */
2023 void *map = kmap_atomic(p);
2024 memcpy(buf, map + offset, length);
2025 kunmap_atomic(map);
2026 } else
2027 memset(buf, 0, length);
2028
2029 addr += length;
2030 buf += length;
2031 copied += length;
2032 count -= length;
2033 }
2034 return copied;
2035}
2036
2037static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2038{
2039 struct page *p;
2040 int copied = 0;
2041
2042 while (count) {
2043 unsigned long offset, length;
2044
2045 offset = offset_in_page(addr);
2046 length = PAGE_SIZE - offset;
2047 if (length > count)
2048 length = count;
2049 p = vmalloc_to_page(addr);
2050 /*
2051 * To do safe access to this _mapped_ area, we need
2052 * lock. But adding lock here means that we need to add
2053 * overhead of vmalloc()/vfree() calles for this _debug_
2054 * interface, rarely used. Instead of that, we'll use
2055 * kmap() and get small overhead in this access function.
2056 */
2057 if (p) {
2058 /*
2059 * we can expect USER0 is not used (see vread/vwrite's
2060 * function description)
2061 */
2062 void *map = kmap_atomic(p);
2063 memcpy(map + offset, buf, length);
2064 kunmap_atomic(map);
2065 }
2066 addr += length;
2067 buf += length;
2068 copied += length;
2069 count -= length;
2070 }
2071 return copied;
2072}
2073
2074/**
2075 * vread() - read vmalloc area in a safe way.
2076 * @buf: buffer for reading data
2077 * @addr: vm address.
2078 * @count: number of bytes to be read.
2079 *
2080 * Returns # of bytes which addr and buf should be increased.
2081 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2082 * includes any intersect with alive vmalloc area.
2083 *
2084 * This function checks that addr is a valid vmalloc'ed area, and
2085 * copy data from that area to a given buffer. If the given memory range
2086 * of [addr...addr+count) includes some valid address, data is copied to
2087 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2088 * IOREMAP area is treated as memory hole and no copy is done.
2089 *
2090 * If [addr...addr+count) doesn't includes any intersects with alive
2091 * vm_struct area, returns 0. @buf should be kernel's buffer.
2092 *
2093 * Note: In usual ops, vread() is never necessary because the caller
2094 * should know vmalloc() area is valid and can use memcpy().
2095 * This is for routines which have to access vmalloc area without
2096 * any informaion, as /dev/kmem.
2097 *
2098 */
2099
2100long vread(char *buf, char *addr, unsigned long count)
2101{
2102 struct vmap_area *va;
2103 struct vm_struct *vm;
2104 char *vaddr, *buf_start = buf;
2105 unsigned long buflen = count;
2106 unsigned long n;
2107
2108 /* Don't allow overflow */
2109 if ((unsigned long) addr + count < count)
2110 count = -(unsigned long) addr;
2111
2112 spin_lock(&vmap_area_lock);
2113 list_for_each_entry(va, &vmap_area_list, list) {
2114 if (!count)
2115 break;
2116
2117 if (!(va->flags & VM_VM_AREA))
2118 continue;
2119
2120 vm = va->vm;
2121 vaddr = (char *) vm->addr;
2122 if (addr >= vaddr + get_vm_area_size(vm))
2123 continue;
2124 while (addr < vaddr) {
2125 if (count == 0)
2126 goto finished;
2127 *buf = '\0';
2128 buf++;
2129 addr++;
2130 count--;
2131 }
2132 n = vaddr + get_vm_area_size(vm) - addr;
2133 if (n > count)
2134 n = count;
2135 if (!(vm->flags & VM_IOREMAP))
2136 aligned_vread(buf, addr, n);
2137 else /* IOREMAP area is treated as memory hole */
2138 memset(buf, 0, n);
2139 buf += n;
2140 addr += n;
2141 count -= n;
2142 }
2143finished:
2144 spin_unlock(&vmap_area_lock);
2145
2146 if (buf == buf_start)
2147 return 0;
2148 /* zero-fill memory holes */
2149 if (buf != buf_start + buflen)
2150 memset(buf, 0, buflen - (buf - buf_start));
2151
2152 return buflen;
2153}
2154
2155/**
2156 * vwrite() - write vmalloc area in a safe way.
2157 * @buf: buffer for source data
2158 * @addr: vm address.
2159 * @count: number of bytes to be read.
2160 *
2161 * Returns # of bytes which addr and buf should be incresed.
2162 * (same number to @count).
2163 * If [addr...addr+count) doesn't includes any intersect with valid
2164 * vmalloc area, returns 0.
2165 *
2166 * This function checks that addr is a valid vmalloc'ed area, and
2167 * copy data from a buffer to the given addr. If specified range of
2168 * [addr...addr+count) includes some valid address, data is copied from
2169 * proper area of @buf. If there are memory holes, no copy to hole.
2170 * IOREMAP area is treated as memory hole and no copy is done.
2171 *
2172 * If [addr...addr+count) doesn't includes any intersects with alive
2173 * vm_struct area, returns 0. @buf should be kernel's buffer.
2174 *
2175 * Note: In usual ops, vwrite() is never necessary because the caller
2176 * should know vmalloc() area is valid and can use memcpy().
2177 * This is for routines which have to access vmalloc area without
2178 * any informaion, as /dev/kmem.
2179 */
2180
2181long vwrite(char *buf, char *addr, unsigned long count)
2182{
2183 struct vmap_area *va;
2184 struct vm_struct *vm;
2185 char *vaddr;
2186 unsigned long n, buflen;
2187 int copied = 0;
2188
2189 /* Don't allow overflow */
2190 if ((unsigned long) addr + count < count)
2191 count = -(unsigned long) addr;
2192 buflen = count;
2193
2194 spin_lock(&vmap_area_lock);
2195 list_for_each_entry(va, &vmap_area_list, list) {
2196 if (!count)
2197 break;
2198
2199 if (!(va->flags & VM_VM_AREA))
2200 continue;
2201
2202 vm = va->vm;
2203 vaddr = (char *) vm->addr;
2204 if (addr >= vaddr + get_vm_area_size(vm))
2205 continue;
2206 while (addr < vaddr) {
2207 if (count == 0)
2208 goto finished;
2209 buf++;
2210 addr++;
2211 count--;
2212 }
2213 n = vaddr + get_vm_area_size(vm) - addr;
2214 if (n > count)
2215 n = count;
2216 if (!(vm->flags & VM_IOREMAP)) {
2217 aligned_vwrite(buf, addr, n);
2218 copied++;
2219 }
2220 buf += n;
2221 addr += n;
2222 count -= n;
2223 }
2224finished:
2225 spin_unlock(&vmap_area_lock);
2226 if (!copied)
2227 return 0;
2228 return buflen;
2229}
2230
2231/**
2232 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2233 * @vma: vma to cover
2234 * @uaddr: target user address to start at
2235 * @kaddr: virtual address of vmalloc kernel memory
2236 * @size: size of map area
2237 *
2238 * Returns: 0 for success, -Exxx on failure
2239 *
2240 * This function checks that @kaddr is a valid vmalloc'ed area,
2241 * and that it is big enough to cover the range starting at
2242 * @uaddr in @vma. Will return failure if that criteria isn't
2243 * met.
2244 *
2245 * Similar to remap_pfn_range() (see mm/memory.c)
2246 */
2247int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2248 void *kaddr, unsigned long size)
2249{
2250 struct vm_struct *area;
2251
2252 size = PAGE_ALIGN(size);
2253
2254 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2255 return -EINVAL;
2256
2257 area = find_vm_area(kaddr);
2258 if (!area)
2259 return -EINVAL;
2260
2261 if (!(area->flags & VM_USERMAP))
2262 return -EINVAL;
2263
2264 if (kaddr + size > area->addr + area->size)
2265 return -EINVAL;
2266
2267 do {
2268 struct page *page = vmalloc_to_page(kaddr);
2269 int ret;
2270
2271 ret = vm_insert_page(vma, uaddr, page);
2272 if (ret)
2273 return ret;
2274
2275 uaddr += PAGE_SIZE;
2276 kaddr += PAGE_SIZE;
2277 size -= PAGE_SIZE;
2278 } while (size > 0);
2279
2280 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2281
2282 return 0;
2283}
2284EXPORT_SYMBOL(remap_vmalloc_range_partial);
2285
2286/**
2287 * remap_vmalloc_range - map vmalloc pages to userspace
2288 * @vma: vma to cover (map full range of vma)
2289 * @addr: vmalloc memory
2290 * @pgoff: number of pages into addr before first page to map
2291 *
2292 * Returns: 0 for success, -Exxx on failure
2293 *
2294 * This function checks that addr is a valid vmalloc'ed area, and
2295 * that it is big enough to cover the vma. Will return failure if
2296 * that criteria isn't met.
2297 *
2298 * Similar to remap_pfn_range() (see mm/memory.c)
2299 */
2300int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2301 unsigned long pgoff)
2302{
2303 return remap_vmalloc_range_partial(vma, vma->vm_start,
2304 addr + (pgoff << PAGE_SHIFT),
2305 vma->vm_end - vma->vm_start);
2306}
2307EXPORT_SYMBOL(remap_vmalloc_range);
2308
2309/*
2310 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2311 * have one.
2312 */
2313void __weak vmalloc_sync_all(void)
2314{
2315}
2316
2317
2318static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2319{
2320 pte_t ***p = data;
2321
2322 if (p) {
2323 *(*p) = pte;
2324 (*p)++;
2325 }
2326 return 0;
2327}
2328
2329/**
2330 * alloc_vm_area - allocate a range of kernel address space
2331 * @size: size of the area
2332 * @ptes: returns the PTEs for the address space
2333 *
2334 * Returns: NULL on failure, vm_struct on success
2335 *
2336 * This function reserves a range of kernel address space, and
2337 * allocates pagetables to map that range. No actual mappings
2338 * are created.
2339 *
2340 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2341 * allocated for the VM area are returned.
2342 */
2343struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2344{
2345 struct vm_struct *area;
2346
2347 area = get_vm_area_caller(size, VM_IOREMAP,
2348 __builtin_return_address(0));
2349 if (area == NULL)
2350 return NULL;
2351
2352 /*
2353 * This ensures that page tables are constructed for this region
2354 * of kernel virtual address space and mapped into init_mm.
2355 */
2356 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2357 size, f, ptes ? &ptes : NULL)) {
2358 free_vm_area(area);
2359 return NULL;
2360 }
2361
2362 return area;
2363}
2364EXPORT_SYMBOL_GPL(alloc_vm_area);
2365
2366void free_vm_area(struct vm_struct *area)
2367{
2368 struct vm_struct *ret;
2369 ret = remove_vm_area(area->addr);
2370 BUG_ON(ret != area);
2371 kfree(area);
2372}
2373EXPORT_SYMBOL_GPL(free_vm_area);
2374
2375#ifdef CONFIG_SMP
2376static struct vmap_area *node_to_va(struct rb_node *n)
2377{
2378 return rb_entry_safe(n, struct vmap_area, rb_node);
2379}
2380
2381/**
2382 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2383 * @end: target address
2384 * @pnext: out arg for the next vmap_area
2385 * @pprev: out arg for the previous vmap_area
2386 *
2387 * Returns: %true if either or both of next and prev are found,
2388 * %false if no vmap_area exists
2389 *
2390 * Find vmap_areas end addresses of which enclose @end. ie. if not
2391 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2392 */
2393static bool pvm_find_next_prev(unsigned long end,
2394 struct vmap_area **pnext,
2395 struct vmap_area **pprev)
2396{
2397 struct rb_node *n = vmap_area_root.rb_node;
2398 struct vmap_area *va = NULL;
2399
2400 while (n) {
2401 va = rb_entry(n, struct vmap_area, rb_node);
2402 if (end < va->va_end)
2403 n = n->rb_left;
2404 else if (end > va->va_end)
2405 n = n->rb_right;
2406 else
2407 break;
2408 }
2409
2410 if (!va)
2411 return false;
2412
2413 if (va->va_end > end) {
2414 *pnext = va;
2415 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2416 } else {
2417 *pprev = va;
2418 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2419 }
2420 return true;
2421}
2422
2423/**
2424 * pvm_determine_end - find the highest aligned address between two vmap_areas
2425 * @pnext: in/out arg for the next vmap_area
2426 * @pprev: in/out arg for the previous vmap_area
2427 * @align: alignment
2428 *
2429 * Returns: determined end address
2430 *
2431 * Find the highest aligned address between *@pnext and *@pprev below
2432 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2433 * down address is between the end addresses of the two vmap_areas.
2434 *
2435 * Please note that the address returned by this function may fall
2436 * inside *@pnext vmap_area. The caller is responsible for checking
2437 * that.
2438 */
2439static unsigned long pvm_determine_end(struct vmap_area **pnext,
2440 struct vmap_area **pprev,
2441 unsigned long align)
2442{
2443 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2444 unsigned long addr;
2445
2446 if (*pnext)
2447 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2448 else
2449 addr = vmalloc_end;
2450
2451 while (*pprev && (*pprev)->va_end > addr) {
2452 *pnext = *pprev;
2453 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2454 }
2455
2456 return addr;
2457}
2458
2459/**
2460 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2461 * @offsets: array containing offset of each area
2462 * @sizes: array containing size of each area
2463 * @nr_vms: the number of areas to allocate
2464 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2465 *
2466 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2467 * vm_structs on success, %NULL on failure
2468 *
2469 * Percpu allocator wants to use congruent vm areas so that it can
2470 * maintain the offsets among percpu areas. This function allocates
2471 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2472 * be scattered pretty far, distance between two areas easily going up
2473 * to gigabytes. To avoid interacting with regular vmallocs, these
2474 * areas are allocated from top.
2475 *
2476 * Despite its complicated look, this allocator is rather simple. It
2477 * does everything top-down and scans areas from the end looking for
2478 * matching slot. While scanning, if any of the areas overlaps with
2479 * existing vmap_area, the base address is pulled down to fit the
2480 * area. Scanning is repeated till all the areas fit and then all
2481 * necessary data structures are inserted and the result is returned.
2482 */
2483struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2484 const size_t *sizes, int nr_vms,
2485 size_t align)
2486{
2487 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2488 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2489 struct vmap_area **vas, *prev, *next;
2490 struct vm_struct **vms;
2491 int area, area2, last_area, term_area;
2492 unsigned long base, start, end, last_end;
2493 bool purged = false;
2494
2495 /* verify parameters and allocate data structures */
2496 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2497 for (last_area = 0, area = 0; area < nr_vms; area++) {
2498 start = offsets[area];
2499 end = start + sizes[area];
2500
2501 /* is everything aligned properly? */
2502 BUG_ON(!IS_ALIGNED(offsets[area], align));
2503 BUG_ON(!IS_ALIGNED(sizes[area], align));
2504
2505 /* detect the area with the highest address */
2506 if (start > offsets[last_area])
2507 last_area = area;
2508
2509 for (area2 = area + 1; area2 < nr_vms; area2++) {
2510 unsigned long start2 = offsets[area2];
2511 unsigned long end2 = start2 + sizes[area2];
2512
2513 BUG_ON(start2 < end && start < end2);
2514 }
2515 }
2516 last_end = offsets[last_area] + sizes[last_area];
2517
2518 if (vmalloc_end - vmalloc_start < last_end) {
2519 WARN_ON(true);
2520 return NULL;
2521 }
2522
2523 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2524 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2525 if (!vas || !vms)
2526 goto err_free2;
2527
2528 for (area = 0; area < nr_vms; area++) {
2529 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2530 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2531 if (!vas[area] || !vms[area])
2532 goto err_free;
2533 }
2534retry:
2535 spin_lock(&vmap_area_lock);
2536
2537 /* start scanning - we scan from the top, begin with the last area */
2538 area = term_area = last_area;
2539 start = offsets[area];
2540 end = start + sizes[area];
2541
2542 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2543 base = vmalloc_end - last_end;
2544 goto found;
2545 }
2546 base = pvm_determine_end(&next, &prev, align) - end;
2547
2548 while (true) {
2549 BUG_ON(next && next->va_end <= base + end);
2550 BUG_ON(prev && prev->va_end > base + end);
2551
2552 /*
2553 * base might have underflowed, add last_end before
2554 * comparing.
2555 */
2556 if (base + last_end < vmalloc_start + last_end) {
2557 spin_unlock(&vmap_area_lock);
2558 if (!purged) {
2559 purge_vmap_area_lazy();
2560 purged = true;
2561 goto retry;
2562 }
2563 goto err_free;
2564 }
2565
2566 /*
2567 * If next overlaps, move base downwards so that it's
2568 * right below next and then recheck.
2569 */
2570 if (next && next->va_start < base + end) {
2571 base = pvm_determine_end(&next, &prev, align) - end;
2572 term_area = area;
2573 continue;
2574 }
2575
2576 /*
2577 * If prev overlaps, shift down next and prev and move
2578 * base so that it's right below new next and then
2579 * recheck.
2580 */
2581 if (prev && prev->va_end > base + start) {
2582 next = prev;
2583 prev = node_to_va(rb_prev(&next->rb_node));
2584 base = pvm_determine_end(&next, &prev, align) - end;
2585 term_area = area;
2586 continue;
2587 }
2588
2589 /*
2590 * This area fits, move on to the previous one. If
2591 * the previous one is the terminal one, we're done.
2592 */
2593 area = (area + nr_vms - 1) % nr_vms;
2594 if (area == term_area)
2595 break;
2596 start = offsets[area];
2597 end = start + sizes[area];
2598 pvm_find_next_prev(base + end, &next, &prev);
2599 }
2600found:
2601 /* we've found a fitting base, insert all va's */
2602 for (area = 0; area < nr_vms; area++) {
2603 struct vmap_area *va = vas[area];
2604
2605 va->va_start = base + offsets[area];
2606 va->va_end = va->va_start + sizes[area];
2607 __insert_vmap_area(va);
2608 }
2609
2610 vmap_area_pcpu_hole = base + offsets[last_area];
2611
2612 spin_unlock(&vmap_area_lock);
2613
2614 /* insert all vm's */
2615 for (area = 0; area < nr_vms; area++)
2616 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2617 pcpu_get_vm_areas);
2618
2619 kfree(vas);
2620 return vms;
2621
2622err_free:
2623 for (area = 0; area < nr_vms; area++) {
2624 kfree(vas[area]);
2625 kfree(vms[area]);
2626 }
2627err_free2:
2628 kfree(vas);
2629 kfree(vms);
2630 return NULL;
2631}
2632
2633/**
2634 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2635 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2636 * @nr_vms: the number of allocated areas
2637 *
2638 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2639 */
2640void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2641{
2642 int i;
2643
2644 for (i = 0; i < nr_vms; i++)
2645 free_vm_area(vms[i]);
2646 kfree(vms);
2647}
2648#endif /* CONFIG_SMP */
2649
2650#ifdef CONFIG_PROC_FS
2651static void *s_start(struct seq_file *m, loff_t *pos)
2652 __acquires(&vmap_area_lock)
2653{
2654 spin_lock(&vmap_area_lock);
2655 return seq_list_start(&vmap_area_list, *pos);
2656}
2657
2658static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2659{
2660 return seq_list_next(p, &vmap_area_list, pos);
2661}
2662
2663static void s_stop(struct seq_file *m, void *p)
2664 __releases(&vmap_area_lock)
2665{
2666 spin_unlock(&vmap_area_lock);
2667}
2668
2669static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2670{
2671 if (IS_ENABLED(CONFIG_NUMA)) {
2672 unsigned int nr, *counters = m->private;
2673
2674 if (!counters)
2675 return;
2676
2677 if (v->flags & VM_UNINITIALIZED)
2678 return;
2679 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2680 smp_rmb();
2681
2682 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2683
2684 for (nr = 0; nr < v->nr_pages; nr++)
2685 counters[page_to_nid(v->pages[nr])]++;
2686
2687 for_each_node_state(nr, N_HIGH_MEMORY)
2688 if (counters[nr])
2689 seq_printf(m, " N%u=%u", nr, counters[nr]);
2690 }
2691}
2692
2693static int s_show(struct seq_file *m, void *p)
2694{
2695 struct vmap_area *va;
2696 struct vm_struct *v;
2697
2698 va = list_entry(p, struct vmap_area, list);
2699
2700 /*
2701 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2702 * behalf of vmap area is being tear down or vm_map_ram allocation.
2703 */
2704 if (!(va->flags & VM_VM_AREA)) {
2705 seq_printf(m, "0x%pK-0x%pK %7ld %s\n",
2706 (void *)va->va_start, (void *)va->va_end,
2707 va->va_end - va->va_start,
2708 va->flags & VM_LAZY_FREE ? "unpurged vm_area" : "vm_map_ram");
2709
2710 return 0;
2711 }
2712
2713 v = va->vm;
2714
2715 seq_printf(m, "0x%pK-0x%pK %7ld",
2716 v->addr, v->addr + v->size, v->size);
2717
2718 if (v->caller)
2719 seq_printf(m, " %pS", v->caller);
2720
2721 if (v->nr_pages)
2722 seq_printf(m, " pages=%d", v->nr_pages);
2723
2724 if (v->phys_addr)
2725 seq_printf(m, " phys=%pa", &v->phys_addr);
2726
2727 if (v->flags & VM_IOREMAP)
2728 seq_puts(m, " ioremap");
2729
2730 if (v->flags & VM_ALLOC)
2731 seq_puts(m, " vmalloc");
2732
2733 if (v->flags & VM_MAP)
2734 seq_puts(m, " vmap");
2735
2736 if (v->flags & VM_USERMAP)
2737 seq_puts(m, " user");
2738
2739 if (is_vmalloc_addr(v->pages))
2740 seq_puts(m, " vpages");
2741
2742 show_numa_info(m, v);
2743 seq_putc(m, '\n');
2744 return 0;
2745}
2746
2747static const struct seq_operations vmalloc_op = {
2748 .start = s_start,
2749 .next = s_next,
2750 .stop = s_stop,
2751 .show = s_show,
2752};
2753
2754static int vmalloc_open(struct inode *inode, struct file *file)
2755{
2756 if (IS_ENABLED(CONFIG_NUMA))
2757 return seq_open_private(file, &vmalloc_op,
2758 nr_node_ids * sizeof(unsigned int));
2759 else
2760 return seq_open(file, &vmalloc_op);
2761}
2762
2763static const struct file_operations proc_vmalloc_operations = {
2764 .open = vmalloc_open,
2765 .read = seq_read,
2766 .llseek = seq_lseek,
2767 .release = seq_release_private,
2768};
2769
2770static int __init proc_vmalloc_init(void)
2771{
2772 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2773 return 0;
2774}
2775module_init(proc_vmalloc_init);
2776
2777#endif
2778
1/*
2 * linux/mm/vmalloc.c
3 *
4 * Copyright (C) 1993 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8 * Numa awareness, Christoph Lameter, SGI, June 2005
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/debugobjects.h>
22#include <linux/kallsyms.h>
23#include <linux/list.h>
24#include <linux/rbtree.h>
25#include <linux/radix-tree.h>
26#include <linux/rcupdate.h>
27#include <linux/pfn.h>
28#include <linux/kmemleak.h>
29#include <linux/atomic.h>
30#include <linux/compiler.h>
31#include <linux/llist.h>
32
33#include <asm/uaccess.h>
34#include <asm/tlbflush.h>
35#include <asm/shmparam.h>
36
37struct vfree_deferred {
38 struct llist_head list;
39 struct work_struct wq;
40};
41static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
42
43static void __vunmap(const void *, int);
44
45static void free_work(struct work_struct *w)
46{
47 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
48 struct llist_node *llnode = llist_del_all(&p->list);
49 while (llnode) {
50 void *p = llnode;
51 llnode = llist_next(llnode);
52 __vunmap(p, 1);
53 }
54}
55
56/*** Page table manipulation functions ***/
57
58static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
59{
60 pte_t *pte;
61
62 pte = pte_offset_kernel(pmd, addr);
63 do {
64 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
65 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
66 } while (pte++, addr += PAGE_SIZE, addr != end);
67}
68
69static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
70{
71 pmd_t *pmd;
72 unsigned long next;
73
74 pmd = pmd_offset(pud, addr);
75 do {
76 next = pmd_addr_end(addr, end);
77 if (pmd_none_or_clear_bad(pmd))
78 continue;
79 vunmap_pte_range(pmd, addr, next);
80 } while (pmd++, addr = next, addr != end);
81}
82
83static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
84{
85 pud_t *pud;
86 unsigned long next;
87
88 pud = pud_offset(pgd, addr);
89 do {
90 next = pud_addr_end(addr, end);
91 if (pud_none_or_clear_bad(pud))
92 continue;
93 vunmap_pmd_range(pud, addr, next);
94 } while (pud++, addr = next, addr != end);
95}
96
97static void vunmap_page_range(unsigned long addr, unsigned long end)
98{
99 pgd_t *pgd;
100 unsigned long next;
101
102 BUG_ON(addr >= end);
103 pgd = pgd_offset_k(addr);
104 do {
105 next = pgd_addr_end(addr, end);
106 if (pgd_none_or_clear_bad(pgd))
107 continue;
108 vunmap_pud_range(pgd, addr, next);
109 } while (pgd++, addr = next, addr != end);
110}
111
112static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
113 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
114{
115 pte_t *pte;
116
117 /*
118 * nr is a running index into the array which helps higher level
119 * callers keep track of where we're up to.
120 */
121
122 pte = pte_alloc_kernel(pmd, addr);
123 if (!pte)
124 return -ENOMEM;
125 do {
126 struct page *page = pages[*nr];
127
128 if (WARN_ON(!pte_none(*pte)))
129 return -EBUSY;
130 if (WARN_ON(!page))
131 return -ENOMEM;
132 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
133 (*nr)++;
134 } while (pte++, addr += PAGE_SIZE, addr != end);
135 return 0;
136}
137
138static int vmap_pmd_range(pud_t *pud, unsigned long addr,
139 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
140{
141 pmd_t *pmd;
142 unsigned long next;
143
144 pmd = pmd_alloc(&init_mm, pud, addr);
145 if (!pmd)
146 return -ENOMEM;
147 do {
148 next = pmd_addr_end(addr, end);
149 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
150 return -ENOMEM;
151 } while (pmd++, addr = next, addr != end);
152 return 0;
153}
154
155static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
156 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
157{
158 pud_t *pud;
159 unsigned long next;
160
161 pud = pud_alloc(&init_mm, pgd, addr);
162 if (!pud)
163 return -ENOMEM;
164 do {
165 next = pud_addr_end(addr, end);
166 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
167 return -ENOMEM;
168 } while (pud++, addr = next, addr != end);
169 return 0;
170}
171
172/*
173 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
174 * will have pfns corresponding to the "pages" array.
175 *
176 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
177 */
178static int vmap_page_range_noflush(unsigned long start, unsigned long end,
179 pgprot_t prot, struct page **pages)
180{
181 pgd_t *pgd;
182 unsigned long next;
183 unsigned long addr = start;
184 int err = 0;
185 int nr = 0;
186
187 BUG_ON(addr >= end);
188 pgd = pgd_offset_k(addr);
189 do {
190 next = pgd_addr_end(addr, end);
191 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
192 if (err)
193 return err;
194 } while (pgd++, addr = next, addr != end);
195
196 return nr;
197}
198
199static int vmap_page_range(unsigned long start, unsigned long end,
200 pgprot_t prot, struct page **pages)
201{
202 int ret;
203
204 ret = vmap_page_range_noflush(start, end, prot, pages);
205 flush_cache_vmap(start, end);
206 return ret;
207}
208
209int is_vmalloc_or_module_addr(const void *x)
210{
211 /*
212 * ARM, x86-64 and sparc64 put modules in a special place,
213 * and fall back on vmalloc() if that fails. Others
214 * just put it in the vmalloc space.
215 */
216#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
217 unsigned long addr = (unsigned long)x;
218 if (addr >= MODULES_VADDR && addr < MODULES_END)
219 return 1;
220#endif
221 return is_vmalloc_addr(x);
222}
223
224/*
225 * Walk a vmap address to the struct page it maps.
226 */
227struct page *vmalloc_to_page(const void *vmalloc_addr)
228{
229 unsigned long addr = (unsigned long) vmalloc_addr;
230 struct page *page = NULL;
231 pgd_t *pgd = pgd_offset_k(addr);
232
233 /*
234 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
235 * architectures that do not vmalloc module space
236 */
237 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
238
239 if (!pgd_none(*pgd)) {
240 pud_t *pud = pud_offset(pgd, addr);
241 if (!pud_none(*pud)) {
242 pmd_t *pmd = pmd_offset(pud, addr);
243 if (!pmd_none(*pmd)) {
244 pte_t *ptep, pte;
245
246 ptep = pte_offset_map(pmd, addr);
247 pte = *ptep;
248 if (pte_present(pte))
249 page = pte_page(pte);
250 pte_unmap(ptep);
251 }
252 }
253 }
254 return page;
255}
256EXPORT_SYMBOL(vmalloc_to_page);
257
258/*
259 * Map a vmalloc()-space virtual address to the physical page frame number.
260 */
261unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
262{
263 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
264}
265EXPORT_SYMBOL(vmalloc_to_pfn);
266
267
268/*** Global kva allocator ***/
269
270#define VM_LAZY_FREE 0x01
271#define VM_LAZY_FREEING 0x02
272#define VM_VM_AREA 0x04
273
274static DEFINE_SPINLOCK(vmap_area_lock);
275/* Export for kexec only */
276LIST_HEAD(vmap_area_list);
277static struct rb_root vmap_area_root = RB_ROOT;
278
279/* The vmap cache globals are protected by vmap_area_lock */
280static struct rb_node *free_vmap_cache;
281static unsigned long cached_hole_size;
282static unsigned long cached_vstart;
283static unsigned long cached_align;
284
285static unsigned long vmap_area_pcpu_hole;
286
287static struct vmap_area *__find_vmap_area(unsigned long addr)
288{
289 struct rb_node *n = vmap_area_root.rb_node;
290
291 while (n) {
292 struct vmap_area *va;
293
294 va = rb_entry(n, struct vmap_area, rb_node);
295 if (addr < va->va_start)
296 n = n->rb_left;
297 else if (addr >= va->va_end)
298 n = n->rb_right;
299 else
300 return va;
301 }
302
303 return NULL;
304}
305
306static void __insert_vmap_area(struct vmap_area *va)
307{
308 struct rb_node **p = &vmap_area_root.rb_node;
309 struct rb_node *parent = NULL;
310 struct rb_node *tmp;
311
312 while (*p) {
313 struct vmap_area *tmp_va;
314
315 parent = *p;
316 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
317 if (va->va_start < tmp_va->va_end)
318 p = &(*p)->rb_left;
319 else if (va->va_end > tmp_va->va_start)
320 p = &(*p)->rb_right;
321 else
322 BUG();
323 }
324
325 rb_link_node(&va->rb_node, parent, p);
326 rb_insert_color(&va->rb_node, &vmap_area_root);
327
328 /* address-sort this list */
329 tmp = rb_prev(&va->rb_node);
330 if (tmp) {
331 struct vmap_area *prev;
332 prev = rb_entry(tmp, struct vmap_area, rb_node);
333 list_add_rcu(&va->list, &prev->list);
334 } else
335 list_add_rcu(&va->list, &vmap_area_list);
336}
337
338static void purge_vmap_area_lazy(void);
339
340/*
341 * Allocate a region of KVA of the specified size and alignment, within the
342 * vstart and vend.
343 */
344static struct vmap_area *alloc_vmap_area(unsigned long size,
345 unsigned long align,
346 unsigned long vstart, unsigned long vend,
347 int node, gfp_t gfp_mask)
348{
349 struct vmap_area *va;
350 struct rb_node *n;
351 unsigned long addr;
352 int purged = 0;
353 struct vmap_area *first;
354
355 BUG_ON(!size);
356 BUG_ON(size & ~PAGE_MASK);
357 BUG_ON(!is_power_of_2(align));
358
359 va = kmalloc_node(sizeof(struct vmap_area),
360 gfp_mask & GFP_RECLAIM_MASK, node);
361 if (unlikely(!va))
362 return ERR_PTR(-ENOMEM);
363
364 /*
365 * Only scan the relevant parts containing pointers to other objects
366 * to avoid false negatives.
367 */
368 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
369
370retry:
371 spin_lock(&vmap_area_lock);
372 /*
373 * Invalidate cache if we have more permissive parameters.
374 * cached_hole_size notes the largest hole noticed _below_
375 * the vmap_area cached in free_vmap_cache: if size fits
376 * into that hole, we want to scan from vstart to reuse
377 * the hole instead of allocating above free_vmap_cache.
378 * Note that __free_vmap_area may update free_vmap_cache
379 * without updating cached_hole_size or cached_align.
380 */
381 if (!free_vmap_cache ||
382 size < cached_hole_size ||
383 vstart < cached_vstart ||
384 align < cached_align) {
385nocache:
386 cached_hole_size = 0;
387 free_vmap_cache = NULL;
388 }
389 /* record if we encounter less permissive parameters */
390 cached_vstart = vstart;
391 cached_align = align;
392
393 /* find starting point for our search */
394 if (free_vmap_cache) {
395 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
396 addr = ALIGN(first->va_end, align);
397 if (addr < vstart)
398 goto nocache;
399 if (addr + size < addr)
400 goto overflow;
401
402 } else {
403 addr = ALIGN(vstart, align);
404 if (addr + size < addr)
405 goto overflow;
406
407 n = vmap_area_root.rb_node;
408 first = NULL;
409
410 while (n) {
411 struct vmap_area *tmp;
412 tmp = rb_entry(n, struct vmap_area, rb_node);
413 if (tmp->va_end >= addr) {
414 first = tmp;
415 if (tmp->va_start <= addr)
416 break;
417 n = n->rb_left;
418 } else
419 n = n->rb_right;
420 }
421
422 if (!first)
423 goto found;
424 }
425
426 /* from the starting point, walk areas until a suitable hole is found */
427 while (addr + size > first->va_start && addr + size <= vend) {
428 if (addr + cached_hole_size < first->va_start)
429 cached_hole_size = first->va_start - addr;
430 addr = ALIGN(first->va_end, align);
431 if (addr + size < addr)
432 goto overflow;
433
434 if (list_is_last(&first->list, &vmap_area_list))
435 goto found;
436
437 first = list_entry(first->list.next,
438 struct vmap_area, list);
439 }
440
441found:
442 if (addr + size > vend)
443 goto overflow;
444
445 va->va_start = addr;
446 va->va_end = addr + size;
447 va->flags = 0;
448 __insert_vmap_area(va);
449 free_vmap_cache = &va->rb_node;
450 spin_unlock(&vmap_area_lock);
451
452 BUG_ON(va->va_start & (align-1));
453 BUG_ON(va->va_start < vstart);
454 BUG_ON(va->va_end > vend);
455
456 return va;
457
458overflow:
459 spin_unlock(&vmap_area_lock);
460 if (!purged) {
461 purge_vmap_area_lazy();
462 purged = 1;
463 goto retry;
464 }
465 if (printk_ratelimit())
466 printk(KERN_WARNING
467 "vmap allocation for size %lu failed: "
468 "use vmalloc=<size> to increase size.\n", size);
469 kfree(va);
470 return ERR_PTR(-EBUSY);
471}
472
473static void __free_vmap_area(struct vmap_area *va)
474{
475 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
476
477 if (free_vmap_cache) {
478 if (va->va_end < cached_vstart) {
479 free_vmap_cache = NULL;
480 } else {
481 struct vmap_area *cache;
482 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
483 if (va->va_start <= cache->va_start) {
484 free_vmap_cache = rb_prev(&va->rb_node);
485 /*
486 * We don't try to update cached_hole_size or
487 * cached_align, but it won't go very wrong.
488 */
489 }
490 }
491 }
492 rb_erase(&va->rb_node, &vmap_area_root);
493 RB_CLEAR_NODE(&va->rb_node);
494 list_del_rcu(&va->list);
495
496 /*
497 * Track the highest possible candidate for pcpu area
498 * allocation. Areas outside of vmalloc area can be returned
499 * here too, consider only end addresses which fall inside
500 * vmalloc area proper.
501 */
502 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
503 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
504
505 kfree_rcu(va, rcu_head);
506}
507
508/*
509 * Free a region of KVA allocated by alloc_vmap_area
510 */
511static void free_vmap_area(struct vmap_area *va)
512{
513 spin_lock(&vmap_area_lock);
514 __free_vmap_area(va);
515 spin_unlock(&vmap_area_lock);
516}
517
518/*
519 * Clear the pagetable entries of a given vmap_area
520 */
521static void unmap_vmap_area(struct vmap_area *va)
522{
523 vunmap_page_range(va->va_start, va->va_end);
524}
525
526static void vmap_debug_free_range(unsigned long start, unsigned long end)
527{
528 /*
529 * Unmap page tables and force a TLB flush immediately if
530 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
531 * bugs similarly to those in linear kernel virtual address
532 * space after a page has been freed.
533 *
534 * All the lazy freeing logic is still retained, in order to
535 * minimise intrusiveness of this debugging feature.
536 *
537 * This is going to be *slow* (linear kernel virtual address
538 * debugging doesn't do a broadcast TLB flush so it is a lot
539 * faster).
540 */
541#ifdef CONFIG_DEBUG_PAGEALLOC
542 vunmap_page_range(start, end);
543 flush_tlb_kernel_range(start, end);
544#endif
545}
546
547/*
548 * lazy_max_pages is the maximum amount of virtual address space we gather up
549 * before attempting to purge with a TLB flush.
550 *
551 * There is a tradeoff here: a larger number will cover more kernel page tables
552 * and take slightly longer to purge, but it will linearly reduce the number of
553 * global TLB flushes that must be performed. It would seem natural to scale
554 * this number up linearly with the number of CPUs (because vmapping activity
555 * could also scale linearly with the number of CPUs), however it is likely
556 * that in practice, workloads might be constrained in other ways that mean
557 * vmap activity will not scale linearly with CPUs. Also, I want to be
558 * conservative and not introduce a big latency on huge systems, so go with
559 * a less aggressive log scale. It will still be an improvement over the old
560 * code, and it will be simple to change the scale factor if we find that it
561 * becomes a problem on bigger systems.
562 */
563static unsigned long lazy_max_pages(void)
564{
565 unsigned int log;
566
567 log = fls(num_online_cpus());
568
569 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
570}
571
572static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
573
574/* for per-CPU blocks */
575static void purge_fragmented_blocks_allcpus(void);
576
577/*
578 * called before a call to iounmap() if the caller wants vm_area_struct's
579 * immediately freed.
580 */
581void set_iounmap_nonlazy(void)
582{
583 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
584}
585
586/*
587 * Purges all lazily-freed vmap areas.
588 *
589 * If sync is 0 then don't purge if there is already a purge in progress.
590 * If force_flush is 1, then flush kernel TLBs between *start and *end even
591 * if we found no lazy vmap areas to unmap (callers can use this to optimise
592 * their own TLB flushing).
593 * Returns with *start = min(*start, lowest purged address)
594 * *end = max(*end, highest purged address)
595 */
596static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
597 int sync, int force_flush)
598{
599 static DEFINE_SPINLOCK(purge_lock);
600 LIST_HEAD(valist);
601 struct vmap_area *va;
602 struct vmap_area *n_va;
603 int nr = 0;
604
605 /*
606 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
607 * should not expect such behaviour. This just simplifies locking for
608 * the case that isn't actually used at the moment anyway.
609 */
610 if (!sync && !force_flush) {
611 if (!spin_trylock(&purge_lock))
612 return;
613 } else
614 spin_lock(&purge_lock);
615
616 if (sync)
617 purge_fragmented_blocks_allcpus();
618
619 rcu_read_lock();
620 list_for_each_entry_rcu(va, &vmap_area_list, list) {
621 if (va->flags & VM_LAZY_FREE) {
622 if (va->va_start < *start)
623 *start = va->va_start;
624 if (va->va_end > *end)
625 *end = va->va_end;
626 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
627 list_add_tail(&va->purge_list, &valist);
628 va->flags |= VM_LAZY_FREEING;
629 va->flags &= ~VM_LAZY_FREE;
630 }
631 }
632 rcu_read_unlock();
633
634 if (nr)
635 atomic_sub(nr, &vmap_lazy_nr);
636
637 if (nr || force_flush)
638 flush_tlb_kernel_range(*start, *end);
639
640 if (nr) {
641 spin_lock(&vmap_area_lock);
642 list_for_each_entry_safe(va, n_va, &valist, purge_list)
643 __free_vmap_area(va);
644 spin_unlock(&vmap_area_lock);
645 }
646 spin_unlock(&purge_lock);
647}
648
649/*
650 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
651 * is already purging.
652 */
653static void try_purge_vmap_area_lazy(void)
654{
655 unsigned long start = ULONG_MAX, end = 0;
656
657 __purge_vmap_area_lazy(&start, &end, 0, 0);
658}
659
660/*
661 * Kick off a purge of the outstanding lazy areas.
662 */
663static void purge_vmap_area_lazy(void)
664{
665 unsigned long start = ULONG_MAX, end = 0;
666
667 __purge_vmap_area_lazy(&start, &end, 1, 0);
668}
669
670/*
671 * Free a vmap area, caller ensuring that the area has been unmapped
672 * and flush_cache_vunmap had been called for the correct range
673 * previously.
674 */
675static void free_vmap_area_noflush(struct vmap_area *va)
676{
677 va->flags |= VM_LAZY_FREE;
678 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
679 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
680 try_purge_vmap_area_lazy();
681}
682
683/*
684 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
685 * called for the correct range previously.
686 */
687static void free_unmap_vmap_area_noflush(struct vmap_area *va)
688{
689 unmap_vmap_area(va);
690 free_vmap_area_noflush(va);
691}
692
693/*
694 * Free and unmap a vmap area
695 */
696static void free_unmap_vmap_area(struct vmap_area *va)
697{
698 flush_cache_vunmap(va->va_start, va->va_end);
699 free_unmap_vmap_area_noflush(va);
700}
701
702static struct vmap_area *find_vmap_area(unsigned long addr)
703{
704 struct vmap_area *va;
705
706 spin_lock(&vmap_area_lock);
707 va = __find_vmap_area(addr);
708 spin_unlock(&vmap_area_lock);
709
710 return va;
711}
712
713static void free_unmap_vmap_area_addr(unsigned long addr)
714{
715 struct vmap_area *va;
716
717 va = find_vmap_area(addr);
718 BUG_ON(!va);
719 free_unmap_vmap_area(va);
720}
721
722
723/*** Per cpu kva allocator ***/
724
725/*
726 * vmap space is limited especially on 32 bit architectures. Ensure there is
727 * room for at least 16 percpu vmap blocks per CPU.
728 */
729/*
730 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
731 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
732 * instead (we just need a rough idea)
733 */
734#if BITS_PER_LONG == 32
735#define VMALLOC_SPACE (128UL*1024*1024)
736#else
737#define VMALLOC_SPACE (128UL*1024*1024*1024)
738#endif
739
740#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
741#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
742#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
743#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
744#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
745#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
746#define VMAP_BBMAP_BITS \
747 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
748 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
749 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
750
751#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
752
753static bool vmap_initialized __read_mostly = false;
754
755struct vmap_block_queue {
756 spinlock_t lock;
757 struct list_head free;
758};
759
760struct vmap_block {
761 spinlock_t lock;
762 struct vmap_area *va;
763 unsigned long free, dirty;
764 DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
765 struct list_head free_list;
766 struct rcu_head rcu_head;
767 struct list_head purge;
768};
769
770/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
771static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
772
773/*
774 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
775 * in the free path. Could get rid of this if we change the API to return a
776 * "cookie" from alloc, to be passed to free. But no big deal yet.
777 */
778static DEFINE_SPINLOCK(vmap_block_tree_lock);
779static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
780
781/*
782 * We should probably have a fallback mechanism to allocate virtual memory
783 * out of partially filled vmap blocks. However vmap block sizing should be
784 * fairly reasonable according to the vmalloc size, so it shouldn't be a
785 * big problem.
786 */
787
788static unsigned long addr_to_vb_idx(unsigned long addr)
789{
790 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
791 addr /= VMAP_BLOCK_SIZE;
792 return addr;
793}
794
795static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
796{
797 struct vmap_block_queue *vbq;
798 struct vmap_block *vb;
799 struct vmap_area *va;
800 unsigned long vb_idx;
801 int node, err;
802
803 node = numa_node_id();
804
805 vb = kmalloc_node(sizeof(struct vmap_block),
806 gfp_mask & GFP_RECLAIM_MASK, node);
807 if (unlikely(!vb))
808 return ERR_PTR(-ENOMEM);
809
810 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
811 VMALLOC_START, VMALLOC_END,
812 node, gfp_mask);
813 if (IS_ERR(va)) {
814 kfree(vb);
815 return ERR_CAST(va);
816 }
817
818 err = radix_tree_preload(gfp_mask);
819 if (unlikely(err)) {
820 kfree(vb);
821 free_vmap_area(va);
822 return ERR_PTR(err);
823 }
824
825 spin_lock_init(&vb->lock);
826 vb->va = va;
827 vb->free = VMAP_BBMAP_BITS;
828 vb->dirty = 0;
829 bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
830 INIT_LIST_HEAD(&vb->free_list);
831
832 vb_idx = addr_to_vb_idx(va->va_start);
833 spin_lock(&vmap_block_tree_lock);
834 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
835 spin_unlock(&vmap_block_tree_lock);
836 BUG_ON(err);
837 radix_tree_preload_end();
838
839 vbq = &get_cpu_var(vmap_block_queue);
840 spin_lock(&vbq->lock);
841 list_add_rcu(&vb->free_list, &vbq->free);
842 spin_unlock(&vbq->lock);
843 put_cpu_var(vmap_block_queue);
844
845 return vb;
846}
847
848static void free_vmap_block(struct vmap_block *vb)
849{
850 struct vmap_block *tmp;
851 unsigned long vb_idx;
852
853 vb_idx = addr_to_vb_idx(vb->va->va_start);
854 spin_lock(&vmap_block_tree_lock);
855 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
856 spin_unlock(&vmap_block_tree_lock);
857 BUG_ON(tmp != vb);
858
859 free_vmap_area_noflush(vb->va);
860 kfree_rcu(vb, rcu_head);
861}
862
863static void purge_fragmented_blocks(int cpu)
864{
865 LIST_HEAD(purge);
866 struct vmap_block *vb;
867 struct vmap_block *n_vb;
868 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
869
870 rcu_read_lock();
871 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
872
873 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
874 continue;
875
876 spin_lock(&vb->lock);
877 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
878 vb->free = 0; /* prevent further allocs after releasing lock */
879 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
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_allcpus(void)
898{
899 int cpu;
900
901 for_each_possible_cpu(cpu)
902 purge_fragmented_blocks(cpu);
903}
904
905static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
906{
907 struct vmap_block_queue *vbq;
908 struct vmap_block *vb;
909 unsigned long addr = 0;
910 unsigned int order;
911
912 BUG_ON(size & ~PAGE_MASK);
913 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
914 if (WARN_ON(size == 0)) {
915 /*
916 * Allocating 0 bytes isn't what caller wants since
917 * get_order(0) returns funny result. Just warn and terminate
918 * early.
919 */
920 return NULL;
921 }
922 order = get_order(size);
923
924again:
925 rcu_read_lock();
926 vbq = &get_cpu_var(vmap_block_queue);
927 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
928 int i;
929
930 spin_lock(&vb->lock);
931 if (vb->free < 1UL << order)
932 goto next;
933
934 i = VMAP_BBMAP_BITS - vb->free;
935 addr = vb->va->va_start + (i << PAGE_SHIFT);
936 BUG_ON(addr_to_vb_idx(addr) !=
937 addr_to_vb_idx(vb->va->va_start));
938 vb->free -= 1UL << order;
939 if (vb->free == 0) {
940 spin_lock(&vbq->lock);
941 list_del_rcu(&vb->free_list);
942 spin_unlock(&vbq->lock);
943 }
944 spin_unlock(&vb->lock);
945 break;
946next:
947 spin_unlock(&vb->lock);
948 }
949
950 put_cpu_var(vmap_block_queue);
951 rcu_read_unlock();
952
953 if (!addr) {
954 vb = new_vmap_block(gfp_mask);
955 if (IS_ERR(vb))
956 return vb;
957 goto again;
958 }
959
960 return (void *)addr;
961}
962
963static void vb_free(const void *addr, unsigned long size)
964{
965 unsigned long offset;
966 unsigned long vb_idx;
967 unsigned int order;
968 struct vmap_block *vb;
969
970 BUG_ON(size & ~PAGE_MASK);
971 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
972
973 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
974
975 order = get_order(size);
976
977 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
978
979 vb_idx = addr_to_vb_idx((unsigned long)addr);
980 rcu_read_lock();
981 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
982 rcu_read_unlock();
983 BUG_ON(!vb);
984
985 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
986
987 spin_lock(&vb->lock);
988 BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
989
990 vb->dirty += 1UL << order;
991 if (vb->dirty == VMAP_BBMAP_BITS) {
992 BUG_ON(vb->free);
993 spin_unlock(&vb->lock);
994 free_vmap_block(vb);
995 } else
996 spin_unlock(&vb->lock);
997}
998
999/**
1000 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1001 *
1002 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1003 * to amortize TLB flushing overheads. What this means is that any page you
1004 * have now, may, in a former life, have been mapped into kernel virtual
1005 * address by the vmap layer and so there might be some CPUs with TLB entries
1006 * still referencing that page (additional to the regular 1:1 kernel mapping).
1007 *
1008 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1009 * be sure that none of the pages we have control over will have any aliases
1010 * from the vmap layer.
1011 */
1012void vm_unmap_aliases(void)
1013{
1014 unsigned long start = ULONG_MAX, end = 0;
1015 int cpu;
1016 int flush = 0;
1017
1018 if (unlikely(!vmap_initialized))
1019 return;
1020
1021 for_each_possible_cpu(cpu) {
1022 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1023 struct vmap_block *vb;
1024
1025 rcu_read_lock();
1026 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1027 int i, j;
1028
1029 spin_lock(&vb->lock);
1030 i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1031 if (i < VMAP_BBMAP_BITS) {
1032 unsigned long s, e;
1033
1034 j = find_last_bit(vb->dirty_map,
1035 VMAP_BBMAP_BITS);
1036 j = j + 1; /* need exclusive index */
1037
1038 s = vb->va->va_start + (i << PAGE_SHIFT);
1039 e = vb->va->va_start + (j << PAGE_SHIFT);
1040 flush = 1;
1041
1042 if (s < start)
1043 start = s;
1044 if (e > end)
1045 end = e;
1046 }
1047 spin_unlock(&vb->lock);
1048 }
1049 rcu_read_unlock();
1050 }
1051
1052 __purge_vmap_area_lazy(&start, &end, 1, flush);
1053}
1054EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1055
1056/**
1057 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1058 * @mem: the pointer returned by vm_map_ram
1059 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1060 */
1061void vm_unmap_ram(const void *mem, unsigned int count)
1062{
1063 unsigned long size = count << PAGE_SHIFT;
1064 unsigned long addr = (unsigned long)mem;
1065
1066 BUG_ON(!addr);
1067 BUG_ON(addr < VMALLOC_START);
1068 BUG_ON(addr > VMALLOC_END);
1069 BUG_ON(addr & (PAGE_SIZE-1));
1070
1071 debug_check_no_locks_freed(mem, size);
1072 vmap_debug_free_range(addr, addr+size);
1073
1074 if (likely(count <= VMAP_MAX_ALLOC))
1075 vb_free(mem, size);
1076 else
1077 free_unmap_vmap_area_addr(addr);
1078}
1079EXPORT_SYMBOL(vm_unmap_ram);
1080
1081/**
1082 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1083 * @pages: an array of pointers to the pages to be mapped
1084 * @count: number of pages
1085 * @node: prefer to allocate data structures on this node
1086 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1087 *
1088 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1089 * faster than vmap so it's good. But if you mix long-life and short-life
1090 * objects with vm_map_ram(), it could consume lots of address space through
1091 * fragmentation (especially on a 32bit machine). You could see failures in
1092 * the end. Please use this function for short-lived objects.
1093 *
1094 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1095 */
1096void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1097{
1098 unsigned long size = count << PAGE_SHIFT;
1099 unsigned long addr;
1100 void *mem;
1101
1102 if (likely(count <= VMAP_MAX_ALLOC)) {
1103 mem = vb_alloc(size, GFP_KERNEL);
1104 if (IS_ERR(mem))
1105 return NULL;
1106 addr = (unsigned long)mem;
1107 } else {
1108 struct vmap_area *va;
1109 va = alloc_vmap_area(size, PAGE_SIZE,
1110 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1111 if (IS_ERR(va))
1112 return NULL;
1113
1114 addr = va->va_start;
1115 mem = (void *)addr;
1116 }
1117 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1118 vm_unmap_ram(mem, count);
1119 return NULL;
1120 }
1121 return mem;
1122}
1123EXPORT_SYMBOL(vm_map_ram);
1124
1125static struct vm_struct *vmlist __initdata;
1126/**
1127 * vm_area_add_early - add vmap area early during boot
1128 * @vm: vm_struct to add
1129 *
1130 * This function is used to add fixed kernel vm area to vmlist before
1131 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1132 * should contain proper values and the other fields should be zero.
1133 *
1134 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1135 */
1136void __init vm_area_add_early(struct vm_struct *vm)
1137{
1138 struct vm_struct *tmp, **p;
1139
1140 BUG_ON(vmap_initialized);
1141 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1142 if (tmp->addr >= vm->addr) {
1143 BUG_ON(tmp->addr < vm->addr + vm->size);
1144 break;
1145 } else
1146 BUG_ON(tmp->addr + tmp->size > vm->addr);
1147 }
1148 vm->next = *p;
1149 *p = vm;
1150}
1151
1152/**
1153 * vm_area_register_early - register vmap area early during boot
1154 * @vm: vm_struct to register
1155 * @align: requested alignment
1156 *
1157 * This function is used to register kernel vm area before
1158 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1159 * proper values on entry and other fields should be zero. On return,
1160 * vm->addr contains the allocated address.
1161 *
1162 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1163 */
1164void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1165{
1166 static size_t vm_init_off __initdata;
1167 unsigned long addr;
1168
1169 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1170 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1171
1172 vm->addr = (void *)addr;
1173
1174 vm_area_add_early(vm);
1175}
1176
1177void __init vmalloc_init(void)
1178{
1179 struct vmap_area *va;
1180 struct vm_struct *tmp;
1181 int i;
1182
1183 for_each_possible_cpu(i) {
1184 struct vmap_block_queue *vbq;
1185 struct vfree_deferred *p;
1186
1187 vbq = &per_cpu(vmap_block_queue, i);
1188 spin_lock_init(&vbq->lock);
1189 INIT_LIST_HEAD(&vbq->free);
1190 p = &per_cpu(vfree_deferred, i);
1191 init_llist_head(&p->list);
1192 INIT_WORK(&p->wq, free_work);
1193 }
1194
1195 /* Import existing vmlist entries. */
1196 for (tmp = vmlist; tmp; tmp = tmp->next) {
1197 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1198 va->flags = VM_VM_AREA;
1199 va->va_start = (unsigned long)tmp->addr;
1200 va->va_end = va->va_start + tmp->size;
1201 va->vm = tmp;
1202 __insert_vmap_area(va);
1203 }
1204
1205 vmap_area_pcpu_hole = VMALLOC_END;
1206
1207 vmap_initialized = true;
1208}
1209
1210/**
1211 * map_kernel_range_noflush - map kernel VM area with the specified pages
1212 * @addr: start of the VM area to map
1213 * @size: size of the VM area to map
1214 * @prot: page protection flags to use
1215 * @pages: pages to map
1216 *
1217 * Map 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_vmap() on to-be-mapped areas
1224 * before calling this function.
1225 *
1226 * RETURNS:
1227 * The number of pages mapped on success, -errno on failure.
1228 */
1229int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1230 pgprot_t prot, struct page **pages)
1231{
1232 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1233}
1234
1235/**
1236 * unmap_kernel_range_noflush - unmap kernel VM area
1237 * @addr: start of the VM area to unmap
1238 * @size: size of the VM area to unmap
1239 *
1240 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1241 * specify should have been allocated using get_vm_area() and its
1242 * friends.
1243 *
1244 * NOTE:
1245 * This function does NOT do any cache flushing. The caller is
1246 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1247 * before calling this function and flush_tlb_kernel_range() after.
1248 */
1249void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1250{
1251 vunmap_page_range(addr, addr + size);
1252}
1253EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1254
1255/**
1256 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1257 * @addr: start of the VM area to unmap
1258 * @size: size of the VM area to unmap
1259 *
1260 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1261 * the unmapping and tlb after.
1262 */
1263void unmap_kernel_range(unsigned long addr, unsigned long size)
1264{
1265 unsigned long end = addr + size;
1266
1267 flush_cache_vunmap(addr, end);
1268 vunmap_page_range(addr, end);
1269 flush_tlb_kernel_range(addr, end);
1270}
1271
1272int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1273{
1274 unsigned long addr = (unsigned long)area->addr;
1275 unsigned long end = addr + get_vm_area_size(area);
1276 int err;
1277
1278 err = vmap_page_range(addr, end, prot, *pages);
1279 if (err > 0) {
1280 *pages += err;
1281 err = 0;
1282 }
1283
1284 return err;
1285}
1286EXPORT_SYMBOL_GPL(map_vm_area);
1287
1288static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1289 unsigned long flags, const void *caller)
1290{
1291 spin_lock(&vmap_area_lock);
1292 vm->flags = flags;
1293 vm->addr = (void *)va->va_start;
1294 vm->size = va->va_end - va->va_start;
1295 vm->caller = caller;
1296 va->vm = vm;
1297 va->flags |= VM_VM_AREA;
1298 spin_unlock(&vmap_area_lock);
1299}
1300
1301static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1302{
1303 /*
1304 * Before removing VM_UNINITIALIZED,
1305 * we should make sure that vm has proper values.
1306 * Pair with smp_rmb() in show_numa_info().
1307 */
1308 smp_wmb();
1309 vm->flags &= ~VM_UNINITIALIZED;
1310}
1311
1312static struct vm_struct *__get_vm_area_node(unsigned long size,
1313 unsigned long align, unsigned long flags, unsigned long start,
1314 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1315{
1316 struct vmap_area *va;
1317 struct vm_struct *area;
1318
1319 BUG_ON(in_interrupt());
1320 if (flags & VM_IOREMAP)
1321 align = 1ul << clamp(fls(size), PAGE_SHIFT, IOREMAP_MAX_ORDER);
1322
1323 size = PAGE_ALIGN(size);
1324 if (unlikely(!size))
1325 return NULL;
1326
1327 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1328 if (unlikely(!area))
1329 return NULL;
1330
1331 /*
1332 * We always allocate a guard page.
1333 */
1334 size += PAGE_SIZE;
1335
1336 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1337 if (IS_ERR(va)) {
1338 kfree(area);
1339 return NULL;
1340 }
1341
1342 setup_vmalloc_vm(area, va, flags, caller);
1343
1344 return area;
1345}
1346
1347struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1348 unsigned long start, unsigned long end)
1349{
1350 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1351 GFP_KERNEL, __builtin_return_address(0));
1352}
1353EXPORT_SYMBOL_GPL(__get_vm_area);
1354
1355struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1356 unsigned long start, unsigned long end,
1357 const void *caller)
1358{
1359 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1360 GFP_KERNEL, caller);
1361}
1362
1363/**
1364 * get_vm_area - reserve a contiguous kernel virtual area
1365 * @size: size of the area
1366 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1367 *
1368 * Search an area of @size in the kernel virtual mapping area,
1369 * and reserved it for out purposes. Returns the area descriptor
1370 * on success or %NULL on failure.
1371 */
1372struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1373{
1374 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1375 NUMA_NO_NODE, GFP_KERNEL,
1376 __builtin_return_address(0));
1377}
1378
1379struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1380 const void *caller)
1381{
1382 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1383 NUMA_NO_NODE, GFP_KERNEL, caller);
1384}
1385
1386/**
1387 * find_vm_area - find a continuous kernel virtual area
1388 * @addr: base address
1389 *
1390 * Search for the kernel VM area starting at @addr, and return it.
1391 * It is up to the caller to do all required locking to keep the returned
1392 * pointer valid.
1393 */
1394struct vm_struct *find_vm_area(const void *addr)
1395{
1396 struct vmap_area *va;
1397
1398 va = find_vmap_area((unsigned long)addr);
1399 if (va && va->flags & VM_VM_AREA)
1400 return va->vm;
1401
1402 return NULL;
1403}
1404
1405/**
1406 * remove_vm_area - find and remove a continuous kernel virtual area
1407 * @addr: base address
1408 *
1409 * Search for the kernel VM area starting at @addr, and remove it.
1410 * This function returns the found VM area, but using it is NOT safe
1411 * on SMP machines, except for its size or flags.
1412 */
1413struct vm_struct *remove_vm_area(const void *addr)
1414{
1415 struct vmap_area *va;
1416
1417 va = find_vmap_area((unsigned long)addr);
1418 if (va && va->flags & VM_VM_AREA) {
1419 struct vm_struct *vm = va->vm;
1420
1421 spin_lock(&vmap_area_lock);
1422 va->vm = NULL;
1423 va->flags &= ~VM_VM_AREA;
1424 spin_unlock(&vmap_area_lock);
1425
1426 vmap_debug_free_range(va->va_start, va->va_end);
1427 free_unmap_vmap_area(va);
1428 vm->size -= PAGE_SIZE;
1429
1430 return vm;
1431 }
1432 return NULL;
1433}
1434
1435static void __vunmap(const void *addr, int deallocate_pages)
1436{
1437 struct vm_struct *area;
1438
1439 if (!addr)
1440 return;
1441
1442 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1443 addr))
1444 return;
1445
1446 area = remove_vm_area(addr);
1447 if (unlikely(!area)) {
1448 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1449 addr);
1450 return;
1451 }
1452
1453 debug_check_no_locks_freed(addr, area->size);
1454 debug_check_no_obj_freed(addr, area->size);
1455
1456 if (deallocate_pages) {
1457 int i;
1458
1459 for (i = 0; i < area->nr_pages; i++) {
1460 struct page *page = area->pages[i];
1461
1462 BUG_ON(!page);
1463 __free_page(page);
1464 }
1465
1466 if (area->flags & VM_VPAGES)
1467 vfree(area->pages);
1468 else
1469 kfree(area->pages);
1470 }
1471
1472 kfree(area);
1473 return;
1474}
1475
1476/**
1477 * vfree - release memory allocated by vmalloc()
1478 * @addr: memory base address
1479 *
1480 * Free the virtually continuous memory area starting at @addr, as
1481 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1482 * NULL, no operation is performed.
1483 *
1484 * Must not be called in NMI context (strictly speaking, only if we don't
1485 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1486 * conventions for vfree() arch-depenedent would be a really bad idea)
1487 *
1488 * NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1489 */
1490void vfree(const void *addr)
1491{
1492 BUG_ON(in_nmi());
1493
1494 kmemleak_free(addr);
1495
1496 if (!addr)
1497 return;
1498 if (unlikely(in_interrupt())) {
1499 struct vfree_deferred *p = &__get_cpu_var(vfree_deferred);
1500 if (llist_add((struct llist_node *)addr, &p->list))
1501 schedule_work(&p->wq);
1502 } else
1503 __vunmap(addr, 1);
1504}
1505EXPORT_SYMBOL(vfree);
1506
1507/**
1508 * vunmap - release virtual mapping obtained by vmap()
1509 * @addr: memory base address
1510 *
1511 * Free the virtually contiguous memory area starting at @addr,
1512 * which was created from the page array passed to vmap().
1513 *
1514 * Must not be called in interrupt context.
1515 */
1516void vunmap(const void *addr)
1517{
1518 BUG_ON(in_interrupt());
1519 might_sleep();
1520 if (addr)
1521 __vunmap(addr, 0);
1522}
1523EXPORT_SYMBOL(vunmap);
1524
1525/**
1526 * vmap - map an array of pages into virtually contiguous space
1527 * @pages: array of page pointers
1528 * @count: number of pages to map
1529 * @flags: vm_area->flags
1530 * @prot: page protection for the mapping
1531 *
1532 * Maps @count pages from @pages into contiguous kernel virtual
1533 * space.
1534 */
1535void *vmap(struct page **pages, unsigned int count,
1536 unsigned long flags, pgprot_t prot)
1537{
1538 struct vm_struct *area;
1539
1540 might_sleep();
1541
1542 if (count > totalram_pages)
1543 return NULL;
1544
1545 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1546 __builtin_return_address(0));
1547 if (!area)
1548 return NULL;
1549
1550 if (map_vm_area(area, prot, &pages)) {
1551 vunmap(area->addr);
1552 return NULL;
1553 }
1554
1555 return area->addr;
1556}
1557EXPORT_SYMBOL(vmap);
1558
1559static void *__vmalloc_node(unsigned long size, unsigned long align,
1560 gfp_t gfp_mask, pgprot_t prot,
1561 int node, const void *caller);
1562static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1563 pgprot_t prot, int node)
1564{
1565 const int order = 0;
1566 struct page **pages;
1567 unsigned int nr_pages, array_size, i;
1568 gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1569
1570 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1571 array_size = (nr_pages * sizeof(struct page *));
1572
1573 area->nr_pages = nr_pages;
1574 /* Please note that the recursion is strictly bounded. */
1575 if (array_size > PAGE_SIZE) {
1576 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1577 PAGE_KERNEL, node, area->caller);
1578 area->flags |= VM_VPAGES;
1579 } else {
1580 pages = kmalloc_node(array_size, nested_gfp, node);
1581 }
1582 area->pages = pages;
1583 if (!area->pages) {
1584 remove_vm_area(area->addr);
1585 kfree(area);
1586 return NULL;
1587 }
1588
1589 for (i = 0; i < area->nr_pages; i++) {
1590 struct page *page;
1591 gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1592
1593 if (node == NUMA_NO_NODE)
1594 page = alloc_page(tmp_mask);
1595 else
1596 page = alloc_pages_node(node, tmp_mask, order);
1597
1598 if (unlikely(!page)) {
1599 /* Successfully allocated i pages, free them in __vunmap() */
1600 area->nr_pages = i;
1601 goto fail;
1602 }
1603 area->pages[i] = page;
1604 }
1605
1606 if (map_vm_area(area, prot, &pages))
1607 goto fail;
1608 return area->addr;
1609
1610fail:
1611 warn_alloc_failed(gfp_mask, order,
1612 "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1613 (area->nr_pages*PAGE_SIZE), area->size);
1614 vfree(area->addr);
1615 return NULL;
1616}
1617
1618/**
1619 * __vmalloc_node_range - allocate virtually contiguous memory
1620 * @size: allocation size
1621 * @align: desired alignment
1622 * @start: vm area range start
1623 * @end: vm area range end
1624 * @gfp_mask: flags for the page level allocator
1625 * @prot: protection mask for the allocated pages
1626 * @node: node to use for allocation or NUMA_NO_NODE
1627 * @caller: caller's return address
1628 *
1629 * Allocate enough pages to cover @size from the page level
1630 * allocator with @gfp_mask flags. Map them into contiguous
1631 * kernel virtual space, using a pagetable protection of @prot.
1632 */
1633void *__vmalloc_node_range(unsigned long size, unsigned long align,
1634 unsigned long start, unsigned long end, gfp_t gfp_mask,
1635 pgprot_t prot, int node, const void *caller)
1636{
1637 struct vm_struct *area;
1638 void *addr;
1639 unsigned long real_size = size;
1640
1641 size = PAGE_ALIGN(size);
1642 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1643 goto fail;
1644
1645 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED,
1646 start, end, node, gfp_mask, caller);
1647 if (!area)
1648 goto fail;
1649
1650 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1651 if (!addr)
1652 return NULL;
1653
1654 /*
1655 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1656 * flag. It means that vm_struct is not fully initialized.
1657 * Now, it is fully initialized, so remove this flag here.
1658 */
1659 clear_vm_uninitialized_flag(area);
1660
1661 /*
1662 * A ref_count = 2 is needed because vm_struct allocated in
1663 * __get_vm_area_node() contains a reference to the virtual address of
1664 * the vmalloc'ed block.
1665 */
1666 kmemleak_alloc(addr, real_size, 2, gfp_mask);
1667
1668 return addr;
1669
1670fail:
1671 warn_alloc_failed(gfp_mask, 0,
1672 "vmalloc: allocation failure: %lu bytes\n",
1673 real_size);
1674 return NULL;
1675}
1676
1677/**
1678 * __vmalloc_node - allocate virtually contiguous memory
1679 * @size: allocation size
1680 * @align: desired alignment
1681 * @gfp_mask: flags for the page level allocator
1682 * @prot: protection mask for the allocated pages
1683 * @node: node to use for allocation or NUMA_NO_NODE
1684 * @caller: caller's return address
1685 *
1686 * Allocate enough pages to cover @size from the page level
1687 * allocator with @gfp_mask flags. Map them into contiguous
1688 * kernel virtual space, using a pagetable protection of @prot.
1689 */
1690static void *__vmalloc_node(unsigned long size, unsigned long align,
1691 gfp_t gfp_mask, pgprot_t prot,
1692 int node, const void *caller)
1693{
1694 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1695 gfp_mask, prot, node, caller);
1696}
1697
1698void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1699{
1700 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1701 __builtin_return_address(0));
1702}
1703EXPORT_SYMBOL(__vmalloc);
1704
1705static inline void *__vmalloc_node_flags(unsigned long size,
1706 int node, gfp_t flags)
1707{
1708 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1709 node, __builtin_return_address(0));
1710}
1711
1712/**
1713 * vmalloc - allocate virtually contiguous memory
1714 * @size: allocation size
1715 * Allocate enough pages to cover @size from the page level
1716 * allocator and map them into contiguous kernel virtual space.
1717 *
1718 * For tight control over page level allocator and protection flags
1719 * use __vmalloc() instead.
1720 */
1721void *vmalloc(unsigned long size)
1722{
1723 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1724 GFP_KERNEL | __GFP_HIGHMEM);
1725}
1726EXPORT_SYMBOL(vmalloc);
1727
1728/**
1729 * vzalloc - allocate virtually contiguous memory with zero fill
1730 * @size: allocation size
1731 * Allocate enough pages to cover @size from the page level
1732 * allocator and map them into contiguous kernel virtual space.
1733 * The memory allocated is set to zero.
1734 *
1735 * For tight control over page level allocator and protection flags
1736 * use __vmalloc() instead.
1737 */
1738void *vzalloc(unsigned long size)
1739{
1740 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1741 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1742}
1743EXPORT_SYMBOL(vzalloc);
1744
1745/**
1746 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1747 * @size: allocation size
1748 *
1749 * The resulting memory area is zeroed so it can be mapped to userspace
1750 * without leaking data.
1751 */
1752void *vmalloc_user(unsigned long size)
1753{
1754 struct vm_struct *area;
1755 void *ret;
1756
1757 ret = __vmalloc_node(size, SHMLBA,
1758 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1759 PAGE_KERNEL, NUMA_NO_NODE,
1760 __builtin_return_address(0));
1761 if (ret) {
1762 area = find_vm_area(ret);
1763 area->flags |= VM_USERMAP;
1764 }
1765 return ret;
1766}
1767EXPORT_SYMBOL(vmalloc_user);
1768
1769/**
1770 * vmalloc_node - allocate memory on a specific node
1771 * @size: allocation size
1772 * @node: numa node
1773 *
1774 * Allocate enough pages to cover @size from the page level
1775 * allocator and map them into contiguous kernel virtual space.
1776 *
1777 * For tight control over page level allocator and protection flags
1778 * use __vmalloc() instead.
1779 */
1780void *vmalloc_node(unsigned long size, int node)
1781{
1782 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1783 node, __builtin_return_address(0));
1784}
1785EXPORT_SYMBOL(vmalloc_node);
1786
1787/**
1788 * vzalloc_node - allocate memory on a specific node with zero fill
1789 * @size: allocation size
1790 * @node: numa node
1791 *
1792 * Allocate enough pages to cover @size from the page level
1793 * allocator and map them into contiguous kernel virtual space.
1794 * The memory allocated is set to zero.
1795 *
1796 * For tight control over page level allocator and protection flags
1797 * use __vmalloc_node() instead.
1798 */
1799void *vzalloc_node(unsigned long size, int node)
1800{
1801 return __vmalloc_node_flags(size, node,
1802 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1803}
1804EXPORT_SYMBOL(vzalloc_node);
1805
1806#ifndef PAGE_KERNEL_EXEC
1807# define PAGE_KERNEL_EXEC PAGE_KERNEL
1808#endif
1809
1810/**
1811 * vmalloc_exec - allocate virtually contiguous, executable memory
1812 * @size: allocation size
1813 *
1814 * Kernel-internal function to allocate enough pages to cover @size
1815 * the page level allocator and map them into contiguous and
1816 * executable kernel virtual space.
1817 *
1818 * For tight control over page level allocator and protection flags
1819 * use __vmalloc() instead.
1820 */
1821
1822void *vmalloc_exec(unsigned long size)
1823{
1824 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1825 NUMA_NO_NODE, __builtin_return_address(0));
1826}
1827
1828#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1829#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1830#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1831#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1832#else
1833#define GFP_VMALLOC32 GFP_KERNEL
1834#endif
1835
1836/**
1837 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1838 * @size: allocation size
1839 *
1840 * Allocate enough 32bit PA addressable pages to cover @size from the
1841 * page level allocator and map them into contiguous kernel virtual space.
1842 */
1843void *vmalloc_32(unsigned long size)
1844{
1845 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1846 NUMA_NO_NODE, __builtin_return_address(0));
1847}
1848EXPORT_SYMBOL(vmalloc_32);
1849
1850/**
1851 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1852 * @size: allocation size
1853 *
1854 * The resulting memory area is 32bit addressable and zeroed so it can be
1855 * mapped to userspace without leaking data.
1856 */
1857void *vmalloc_32_user(unsigned long size)
1858{
1859 struct vm_struct *area;
1860 void *ret;
1861
1862 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1863 NUMA_NO_NODE, __builtin_return_address(0));
1864 if (ret) {
1865 area = find_vm_area(ret);
1866 area->flags |= VM_USERMAP;
1867 }
1868 return ret;
1869}
1870EXPORT_SYMBOL(vmalloc_32_user);
1871
1872/*
1873 * small helper routine , copy contents to buf from addr.
1874 * If the page is not present, fill zero.
1875 */
1876
1877static int aligned_vread(char *buf, char *addr, unsigned long count)
1878{
1879 struct page *p;
1880 int copied = 0;
1881
1882 while (count) {
1883 unsigned long offset, length;
1884
1885 offset = (unsigned long)addr & ~PAGE_MASK;
1886 length = PAGE_SIZE - offset;
1887 if (length > count)
1888 length = count;
1889 p = vmalloc_to_page(addr);
1890 /*
1891 * To do safe access to this _mapped_ area, we need
1892 * lock. But adding lock here means that we need to add
1893 * overhead of vmalloc()/vfree() calles for this _debug_
1894 * interface, rarely used. Instead of that, we'll use
1895 * kmap() and get small overhead in this access function.
1896 */
1897 if (p) {
1898 /*
1899 * we can expect USER0 is not used (see vread/vwrite's
1900 * function description)
1901 */
1902 void *map = kmap_atomic(p);
1903 memcpy(buf, map + offset, length);
1904 kunmap_atomic(map);
1905 } else
1906 memset(buf, 0, length);
1907
1908 addr += length;
1909 buf += length;
1910 copied += length;
1911 count -= length;
1912 }
1913 return copied;
1914}
1915
1916static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1917{
1918 struct page *p;
1919 int copied = 0;
1920
1921 while (count) {
1922 unsigned long offset, length;
1923
1924 offset = (unsigned long)addr & ~PAGE_MASK;
1925 length = PAGE_SIZE - offset;
1926 if (length > count)
1927 length = count;
1928 p = vmalloc_to_page(addr);
1929 /*
1930 * To do safe access to this _mapped_ area, we need
1931 * lock. But adding lock here means that we need to add
1932 * overhead of vmalloc()/vfree() calles for this _debug_
1933 * interface, rarely used. Instead of that, we'll use
1934 * kmap() and get small overhead in this access function.
1935 */
1936 if (p) {
1937 /*
1938 * we can expect USER0 is not used (see vread/vwrite's
1939 * function description)
1940 */
1941 void *map = kmap_atomic(p);
1942 memcpy(map + offset, buf, length);
1943 kunmap_atomic(map);
1944 }
1945 addr += length;
1946 buf += length;
1947 copied += length;
1948 count -= length;
1949 }
1950 return copied;
1951}
1952
1953/**
1954 * vread() - read vmalloc area in a safe way.
1955 * @buf: buffer for reading data
1956 * @addr: vm address.
1957 * @count: number of bytes to be read.
1958 *
1959 * Returns # of bytes which addr and buf should be increased.
1960 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
1961 * includes any intersect with alive vmalloc area.
1962 *
1963 * This function checks that addr is a valid vmalloc'ed area, and
1964 * copy data from that area to a given buffer. If the given memory range
1965 * of [addr...addr+count) includes some valid address, data is copied to
1966 * proper area of @buf. If there are memory holes, they'll be zero-filled.
1967 * IOREMAP area is treated as memory hole and no copy is done.
1968 *
1969 * If [addr...addr+count) doesn't includes any intersects with alive
1970 * vm_struct area, returns 0. @buf should be kernel's buffer.
1971 *
1972 * Note: In usual ops, vread() is never necessary because the caller
1973 * should know vmalloc() area is valid and can use memcpy().
1974 * This is for routines which have to access vmalloc area without
1975 * any informaion, as /dev/kmem.
1976 *
1977 */
1978
1979long vread(char *buf, char *addr, unsigned long count)
1980{
1981 struct vmap_area *va;
1982 struct vm_struct *vm;
1983 char *vaddr, *buf_start = buf;
1984 unsigned long buflen = count;
1985 unsigned long n;
1986
1987 /* Don't allow overflow */
1988 if ((unsigned long) addr + count < count)
1989 count = -(unsigned long) addr;
1990
1991 spin_lock(&vmap_area_lock);
1992 list_for_each_entry(va, &vmap_area_list, list) {
1993 if (!count)
1994 break;
1995
1996 if (!(va->flags & VM_VM_AREA))
1997 continue;
1998
1999 vm = va->vm;
2000 vaddr = (char *) vm->addr;
2001 if (addr >= vaddr + get_vm_area_size(vm))
2002 continue;
2003 while (addr < vaddr) {
2004 if (count == 0)
2005 goto finished;
2006 *buf = '\0';
2007 buf++;
2008 addr++;
2009 count--;
2010 }
2011 n = vaddr + get_vm_area_size(vm) - addr;
2012 if (n > count)
2013 n = count;
2014 if (!(vm->flags & VM_IOREMAP))
2015 aligned_vread(buf, addr, n);
2016 else /* IOREMAP area is treated as memory hole */
2017 memset(buf, 0, n);
2018 buf += n;
2019 addr += n;
2020 count -= n;
2021 }
2022finished:
2023 spin_unlock(&vmap_area_lock);
2024
2025 if (buf == buf_start)
2026 return 0;
2027 /* zero-fill memory holes */
2028 if (buf != buf_start + buflen)
2029 memset(buf, 0, buflen - (buf - buf_start));
2030
2031 return buflen;
2032}
2033
2034/**
2035 * vwrite() - write vmalloc area in a safe way.
2036 * @buf: buffer for source data
2037 * @addr: vm address.
2038 * @count: number of bytes to be read.
2039 *
2040 * Returns # of bytes which addr and buf should be incresed.
2041 * (same number to @count).
2042 * If [addr...addr+count) doesn't includes any intersect with valid
2043 * vmalloc area, returns 0.
2044 *
2045 * This function checks that addr is a valid vmalloc'ed area, and
2046 * copy data from a buffer to the given addr. If specified range of
2047 * [addr...addr+count) includes some valid address, data is copied from
2048 * proper area of @buf. If there are memory holes, no copy to hole.
2049 * IOREMAP area is treated as memory hole and no copy is done.
2050 *
2051 * If [addr...addr+count) doesn't includes any intersects with alive
2052 * vm_struct area, returns 0. @buf should be kernel's buffer.
2053 *
2054 * Note: In usual ops, vwrite() is never necessary because the caller
2055 * should know vmalloc() area is valid and can use memcpy().
2056 * This is for routines which have to access vmalloc area without
2057 * any informaion, as /dev/kmem.
2058 */
2059
2060long vwrite(char *buf, char *addr, unsigned long count)
2061{
2062 struct vmap_area *va;
2063 struct vm_struct *vm;
2064 char *vaddr;
2065 unsigned long n, buflen;
2066 int copied = 0;
2067
2068 /* Don't allow overflow */
2069 if ((unsigned long) addr + count < count)
2070 count = -(unsigned long) addr;
2071 buflen = count;
2072
2073 spin_lock(&vmap_area_lock);
2074 list_for_each_entry(va, &vmap_area_list, list) {
2075 if (!count)
2076 break;
2077
2078 if (!(va->flags & VM_VM_AREA))
2079 continue;
2080
2081 vm = va->vm;
2082 vaddr = (char *) vm->addr;
2083 if (addr >= vaddr + get_vm_area_size(vm))
2084 continue;
2085 while (addr < vaddr) {
2086 if (count == 0)
2087 goto finished;
2088 buf++;
2089 addr++;
2090 count--;
2091 }
2092 n = vaddr + get_vm_area_size(vm) - addr;
2093 if (n > count)
2094 n = count;
2095 if (!(vm->flags & VM_IOREMAP)) {
2096 aligned_vwrite(buf, addr, n);
2097 copied++;
2098 }
2099 buf += n;
2100 addr += n;
2101 count -= n;
2102 }
2103finished:
2104 spin_unlock(&vmap_area_lock);
2105 if (!copied)
2106 return 0;
2107 return buflen;
2108}
2109
2110/**
2111 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2112 * @vma: vma to cover
2113 * @uaddr: target user address to start at
2114 * @kaddr: virtual address of vmalloc kernel memory
2115 * @size: size of map area
2116 *
2117 * Returns: 0 for success, -Exxx on failure
2118 *
2119 * This function checks that @kaddr is a valid vmalloc'ed area,
2120 * and that it is big enough to cover the range starting at
2121 * @uaddr in @vma. Will return failure if that criteria isn't
2122 * met.
2123 *
2124 * Similar to remap_pfn_range() (see mm/memory.c)
2125 */
2126int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2127 void *kaddr, unsigned long size)
2128{
2129 struct vm_struct *area;
2130
2131 size = PAGE_ALIGN(size);
2132
2133 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2134 return -EINVAL;
2135
2136 area = find_vm_area(kaddr);
2137 if (!area)
2138 return -EINVAL;
2139
2140 if (!(area->flags & VM_USERMAP))
2141 return -EINVAL;
2142
2143 if (kaddr + size > area->addr + area->size)
2144 return -EINVAL;
2145
2146 do {
2147 struct page *page = vmalloc_to_page(kaddr);
2148 int ret;
2149
2150 ret = vm_insert_page(vma, uaddr, page);
2151 if (ret)
2152 return ret;
2153
2154 uaddr += PAGE_SIZE;
2155 kaddr += PAGE_SIZE;
2156 size -= PAGE_SIZE;
2157 } while (size > 0);
2158
2159 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2160
2161 return 0;
2162}
2163EXPORT_SYMBOL(remap_vmalloc_range_partial);
2164
2165/**
2166 * remap_vmalloc_range - map vmalloc pages to userspace
2167 * @vma: vma to cover (map full range of vma)
2168 * @addr: vmalloc memory
2169 * @pgoff: number of pages into addr before first page to map
2170 *
2171 * Returns: 0 for success, -Exxx on failure
2172 *
2173 * This function checks that addr is a valid vmalloc'ed area, and
2174 * that it is big enough to cover the vma. Will return failure if
2175 * that criteria isn't met.
2176 *
2177 * Similar to remap_pfn_range() (see mm/memory.c)
2178 */
2179int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2180 unsigned long pgoff)
2181{
2182 return remap_vmalloc_range_partial(vma, vma->vm_start,
2183 addr + (pgoff << PAGE_SHIFT),
2184 vma->vm_end - vma->vm_start);
2185}
2186EXPORT_SYMBOL(remap_vmalloc_range);
2187
2188/*
2189 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2190 * have one.
2191 */
2192void __weak vmalloc_sync_all(void)
2193{
2194}
2195
2196
2197static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2198{
2199 pte_t ***p = data;
2200
2201 if (p) {
2202 *(*p) = pte;
2203 (*p)++;
2204 }
2205 return 0;
2206}
2207
2208/**
2209 * alloc_vm_area - allocate a range of kernel address space
2210 * @size: size of the area
2211 * @ptes: returns the PTEs for the address space
2212 *
2213 * Returns: NULL on failure, vm_struct on success
2214 *
2215 * This function reserves a range of kernel address space, and
2216 * allocates pagetables to map that range. No actual mappings
2217 * are created.
2218 *
2219 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2220 * allocated for the VM area are returned.
2221 */
2222struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2223{
2224 struct vm_struct *area;
2225
2226 area = get_vm_area_caller(size, VM_IOREMAP,
2227 __builtin_return_address(0));
2228 if (area == NULL)
2229 return NULL;
2230
2231 /*
2232 * This ensures that page tables are constructed for this region
2233 * of kernel virtual address space and mapped into init_mm.
2234 */
2235 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2236 size, f, ptes ? &ptes : NULL)) {
2237 free_vm_area(area);
2238 return NULL;
2239 }
2240
2241 return area;
2242}
2243EXPORT_SYMBOL_GPL(alloc_vm_area);
2244
2245void free_vm_area(struct vm_struct *area)
2246{
2247 struct vm_struct *ret;
2248 ret = remove_vm_area(area->addr);
2249 BUG_ON(ret != area);
2250 kfree(area);
2251}
2252EXPORT_SYMBOL_GPL(free_vm_area);
2253
2254#ifdef CONFIG_SMP
2255static struct vmap_area *node_to_va(struct rb_node *n)
2256{
2257 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2258}
2259
2260/**
2261 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2262 * @end: target address
2263 * @pnext: out arg for the next vmap_area
2264 * @pprev: out arg for the previous vmap_area
2265 *
2266 * Returns: %true if either or both of next and prev are found,
2267 * %false if no vmap_area exists
2268 *
2269 * Find vmap_areas end addresses of which enclose @end. ie. if not
2270 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2271 */
2272static bool pvm_find_next_prev(unsigned long end,
2273 struct vmap_area **pnext,
2274 struct vmap_area **pprev)
2275{
2276 struct rb_node *n = vmap_area_root.rb_node;
2277 struct vmap_area *va = NULL;
2278
2279 while (n) {
2280 va = rb_entry(n, struct vmap_area, rb_node);
2281 if (end < va->va_end)
2282 n = n->rb_left;
2283 else if (end > va->va_end)
2284 n = n->rb_right;
2285 else
2286 break;
2287 }
2288
2289 if (!va)
2290 return false;
2291
2292 if (va->va_end > end) {
2293 *pnext = va;
2294 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2295 } else {
2296 *pprev = va;
2297 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2298 }
2299 return true;
2300}
2301
2302/**
2303 * pvm_determine_end - find the highest aligned address between two vmap_areas
2304 * @pnext: in/out arg for the next vmap_area
2305 * @pprev: in/out arg for the previous vmap_area
2306 * @align: alignment
2307 *
2308 * Returns: determined end address
2309 *
2310 * Find the highest aligned address between *@pnext and *@pprev below
2311 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2312 * down address is between the end addresses of the two vmap_areas.
2313 *
2314 * Please note that the address returned by this function may fall
2315 * inside *@pnext vmap_area. The caller is responsible for checking
2316 * that.
2317 */
2318static unsigned long pvm_determine_end(struct vmap_area **pnext,
2319 struct vmap_area **pprev,
2320 unsigned long align)
2321{
2322 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2323 unsigned long addr;
2324
2325 if (*pnext)
2326 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2327 else
2328 addr = vmalloc_end;
2329
2330 while (*pprev && (*pprev)->va_end > addr) {
2331 *pnext = *pprev;
2332 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2333 }
2334
2335 return addr;
2336}
2337
2338/**
2339 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2340 * @offsets: array containing offset of each area
2341 * @sizes: array containing size of each area
2342 * @nr_vms: the number of areas to allocate
2343 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2344 *
2345 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2346 * vm_structs on success, %NULL on failure
2347 *
2348 * Percpu allocator wants to use congruent vm areas so that it can
2349 * maintain the offsets among percpu areas. This function allocates
2350 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2351 * be scattered pretty far, distance between two areas easily going up
2352 * to gigabytes. To avoid interacting with regular vmallocs, these
2353 * areas are allocated from top.
2354 *
2355 * Despite its complicated look, this allocator is rather simple. It
2356 * does everything top-down and scans areas from the end looking for
2357 * matching slot. While scanning, if any of the areas overlaps with
2358 * existing vmap_area, the base address is pulled down to fit the
2359 * area. Scanning is repeated till all the areas fit and then all
2360 * necessary data structres are inserted and the result is returned.
2361 */
2362struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2363 const size_t *sizes, int nr_vms,
2364 size_t align)
2365{
2366 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2367 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2368 struct vmap_area **vas, *prev, *next;
2369 struct vm_struct **vms;
2370 int area, area2, last_area, term_area;
2371 unsigned long base, start, end, last_end;
2372 bool purged = false;
2373
2374 /* verify parameters and allocate data structures */
2375 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2376 for (last_area = 0, area = 0; area < nr_vms; area++) {
2377 start = offsets[area];
2378 end = start + sizes[area];
2379
2380 /* is everything aligned properly? */
2381 BUG_ON(!IS_ALIGNED(offsets[area], align));
2382 BUG_ON(!IS_ALIGNED(sizes[area], align));
2383
2384 /* detect the area with the highest address */
2385 if (start > offsets[last_area])
2386 last_area = area;
2387
2388 for (area2 = 0; area2 < nr_vms; area2++) {
2389 unsigned long start2 = offsets[area2];
2390 unsigned long end2 = start2 + sizes[area2];
2391
2392 if (area2 == area)
2393 continue;
2394
2395 BUG_ON(start2 >= start && start2 < end);
2396 BUG_ON(end2 <= end && end2 > start);
2397 }
2398 }
2399 last_end = offsets[last_area] + sizes[last_area];
2400
2401 if (vmalloc_end - vmalloc_start < last_end) {
2402 WARN_ON(true);
2403 return NULL;
2404 }
2405
2406 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2407 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2408 if (!vas || !vms)
2409 goto err_free2;
2410
2411 for (area = 0; area < nr_vms; area++) {
2412 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2413 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2414 if (!vas[area] || !vms[area])
2415 goto err_free;
2416 }
2417retry:
2418 spin_lock(&vmap_area_lock);
2419
2420 /* start scanning - we scan from the top, begin with the last area */
2421 area = term_area = last_area;
2422 start = offsets[area];
2423 end = start + sizes[area];
2424
2425 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2426 base = vmalloc_end - last_end;
2427 goto found;
2428 }
2429 base = pvm_determine_end(&next, &prev, align) - end;
2430
2431 while (true) {
2432 BUG_ON(next && next->va_end <= base + end);
2433 BUG_ON(prev && prev->va_end > base + end);
2434
2435 /*
2436 * base might have underflowed, add last_end before
2437 * comparing.
2438 */
2439 if (base + last_end < vmalloc_start + last_end) {
2440 spin_unlock(&vmap_area_lock);
2441 if (!purged) {
2442 purge_vmap_area_lazy();
2443 purged = true;
2444 goto retry;
2445 }
2446 goto err_free;
2447 }
2448
2449 /*
2450 * If next overlaps, move base downwards so that it's
2451 * right below next and then recheck.
2452 */
2453 if (next && next->va_start < base + end) {
2454 base = pvm_determine_end(&next, &prev, align) - end;
2455 term_area = area;
2456 continue;
2457 }
2458
2459 /*
2460 * If prev overlaps, shift down next and prev and move
2461 * base so that it's right below new next and then
2462 * recheck.
2463 */
2464 if (prev && prev->va_end > base + start) {
2465 next = prev;
2466 prev = node_to_va(rb_prev(&next->rb_node));
2467 base = pvm_determine_end(&next, &prev, align) - end;
2468 term_area = area;
2469 continue;
2470 }
2471
2472 /*
2473 * This area fits, move on to the previous one. If
2474 * the previous one is the terminal one, we're done.
2475 */
2476 area = (area + nr_vms - 1) % nr_vms;
2477 if (area == term_area)
2478 break;
2479 start = offsets[area];
2480 end = start + sizes[area];
2481 pvm_find_next_prev(base + end, &next, &prev);
2482 }
2483found:
2484 /* we've found a fitting base, insert all va's */
2485 for (area = 0; area < nr_vms; area++) {
2486 struct vmap_area *va = vas[area];
2487
2488 va->va_start = base + offsets[area];
2489 va->va_end = va->va_start + sizes[area];
2490 __insert_vmap_area(va);
2491 }
2492
2493 vmap_area_pcpu_hole = base + offsets[last_area];
2494
2495 spin_unlock(&vmap_area_lock);
2496
2497 /* insert all vm's */
2498 for (area = 0; area < nr_vms; area++)
2499 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2500 pcpu_get_vm_areas);
2501
2502 kfree(vas);
2503 return vms;
2504
2505err_free:
2506 for (area = 0; area < nr_vms; area++) {
2507 kfree(vas[area]);
2508 kfree(vms[area]);
2509 }
2510err_free2:
2511 kfree(vas);
2512 kfree(vms);
2513 return NULL;
2514}
2515
2516/**
2517 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2518 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2519 * @nr_vms: the number of allocated areas
2520 *
2521 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2522 */
2523void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2524{
2525 int i;
2526
2527 for (i = 0; i < nr_vms; i++)
2528 free_vm_area(vms[i]);
2529 kfree(vms);
2530}
2531#endif /* CONFIG_SMP */
2532
2533#ifdef CONFIG_PROC_FS
2534static void *s_start(struct seq_file *m, loff_t *pos)
2535 __acquires(&vmap_area_lock)
2536{
2537 loff_t n = *pos;
2538 struct vmap_area *va;
2539
2540 spin_lock(&vmap_area_lock);
2541 va = list_entry((&vmap_area_list)->next, typeof(*va), list);
2542 while (n > 0 && &va->list != &vmap_area_list) {
2543 n--;
2544 va = list_entry(va->list.next, typeof(*va), list);
2545 }
2546 if (!n && &va->list != &vmap_area_list)
2547 return va;
2548
2549 return NULL;
2550
2551}
2552
2553static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2554{
2555 struct vmap_area *va = p, *next;
2556
2557 ++*pos;
2558 next = list_entry(va->list.next, typeof(*va), list);
2559 if (&next->list != &vmap_area_list)
2560 return next;
2561
2562 return NULL;
2563}
2564
2565static void s_stop(struct seq_file *m, void *p)
2566 __releases(&vmap_area_lock)
2567{
2568 spin_unlock(&vmap_area_lock);
2569}
2570
2571static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2572{
2573 if (IS_ENABLED(CONFIG_NUMA)) {
2574 unsigned int nr, *counters = m->private;
2575
2576 if (!counters)
2577 return;
2578
2579 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2580 smp_rmb();
2581 if (v->flags & VM_UNINITIALIZED)
2582 return;
2583
2584 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2585
2586 for (nr = 0; nr < v->nr_pages; nr++)
2587 counters[page_to_nid(v->pages[nr])]++;
2588
2589 for_each_node_state(nr, N_HIGH_MEMORY)
2590 if (counters[nr])
2591 seq_printf(m, " N%u=%u", nr, counters[nr]);
2592 }
2593}
2594
2595static int s_show(struct seq_file *m, void *p)
2596{
2597 struct vmap_area *va = p;
2598 struct vm_struct *v;
2599
2600 /*
2601 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2602 * behalf of vmap area is being tear down or vm_map_ram allocation.
2603 */
2604 if (!(va->flags & VM_VM_AREA))
2605 return 0;
2606
2607 v = va->vm;
2608
2609 seq_printf(m, "0x%pK-0x%pK %7ld",
2610 v->addr, v->addr + v->size, v->size);
2611
2612 if (v->caller)
2613 seq_printf(m, " %pS", v->caller);
2614
2615 if (v->nr_pages)
2616 seq_printf(m, " pages=%d", v->nr_pages);
2617
2618 if (v->phys_addr)
2619 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2620
2621 if (v->flags & VM_IOREMAP)
2622 seq_printf(m, " ioremap");
2623
2624 if (v->flags & VM_ALLOC)
2625 seq_printf(m, " vmalloc");
2626
2627 if (v->flags & VM_MAP)
2628 seq_printf(m, " vmap");
2629
2630 if (v->flags & VM_USERMAP)
2631 seq_printf(m, " user");
2632
2633 if (v->flags & VM_VPAGES)
2634 seq_printf(m, " vpages");
2635
2636 show_numa_info(m, v);
2637 seq_putc(m, '\n');
2638 return 0;
2639}
2640
2641static const struct seq_operations vmalloc_op = {
2642 .start = s_start,
2643 .next = s_next,
2644 .stop = s_stop,
2645 .show = s_show,
2646};
2647
2648static int vmalloc_open(struct inode *inode, struct file *file)
2649{
2650 unsigned int *ptr = NULL;
2651 int ret;
2652
2653 if (IS_ENABLED(CONFIG_NUMA)) {
2654 ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2655 if (ptr == NULL)
2656 return -ENOMEM;
2657 }
2658 ret = seq_open(file, &vmalloc_op);
2659 if (!ret) {
2660 struct seq_file *m = file->private_data;
2661 m->private = ptr;
2662 } else
2663 kfree(ptr);
2664 return ret;
2665}
2666
2667static const struct file_operations proc_vmalloc_operations = {
2668 .open = vmalloc_open,
2669 .read = seq_read,
2670 .llseek = seq_lseek,
2671 .release = seq_release_private,
2672};
2673
2674static int __init proc_vmalloc_init(void)
2675{
2676 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2677 return 0;
2678}
2679module_init(proc_vmalloc_init);
2680
2681void get_vmalloc_info(struct vmalloc_info *vmi)
2682{
2683 struct vmap_area *va;
2684 unsigned long free_area_size;
2685 unsigned long prev_end;
2686
2687 vmi->used = 0;
2688 vmi->largest_chunk = 0;
2689
2690 prev_end = VMALLOC_START;
2691
2692 spin_lock(&vmap_area_lock);
2693
2694 if (list_empty(&vmap_area_list)) {
2695 vmi->largest_chunk = VMALLOC_TOTAL;
2696 goto out;
2697 }
2698
2699 list_for_each_entry(va, &vmap_area_list, list) {
2700 unsigned long addr = va->va_start;
2701
2702 /*
2703 * Some archs keep another range for modules in vmalloc space
2704 */
2705 if (addr < VMALLOC_START)
2706 continue;
2707 if (addr >= VMALLOC_END)
2708 break;
2709
2710 if (va->flags & (VM_LAZY_FREE | VM_LAZY_FREEING))
2711 continue;
2712
2713 vmi->used += (va->va_end - va->va_start);
2714
2715 free_area_size = addr - prev_end;
2716 if (vmi->largest_chunk < free_area_size)
2717 vmi->largest_chunk = free_area_size;
2718
2719 prev_end = va->va_end;
2720 }
2721
2722 if (VMALLOC_END - prev_end > vmi->largest_chunk)
2723 vmi->largest_chunk = VMALLOC_END - prev_end;
2724
2725out:
2726 spin_unlock(&vmap_area_lock);
2727}
2728#endif
2729