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