projects/linux·blob v5.0 · refs

2754 lines · 69.7 KiB

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 <[email protected]>, 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
606/*
607 * lazy_max_pages is the maximum amount of virtual address space we gather up
608 * before attempting to purge with a TLB flush.
609 *
610 * There is a tradeoff here: a larger number will cover more kernel page tables
611 * and take slightly longer to purge, but it will linearly reduce the number of
612 * global TLB flushes that must be performed. It would seem natural to scale
613 * this number up linearly with the number of CPUs (because vmapping activity
614 * could also scale linearly with the number of CPUs), however it is likely
615 * that in practice, workloads might be constrained in other ways that mean
616 * vmap activity will not scale linearly with CPUs. Also, I want to be
617 * conservative and not introduce a big latency on huge systems, so go with
618 * a less aggressive log scale. It will still be an improvement over the old
619 * code, and it will be simple to change the scale factor if we find that it
620 * becomes a problem on bigger systems.
621 */
622static unsigned long lazy_max_pages(void)
623{
624 unsigned int log;
625
626 log = fls(num_online_cpus());
627
628 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
629}
630
631static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
632
633/*
634 * Serialize vmap purging. There is no actual criticial section protected
635 * by this look, but we want to avoid concurrent calls for performance
636 * reasons and to make the pcpu_get_vm_areas more deterministic.
637 */
638static DEFINE_MUTEX(vmap_purge_lock);
639
640/* for per-CPU blocks */
641static void purge_fragmented_blocks_allcpus(void);
642
643/*
644 * called before a call to iounmap() if the caller wants vm_area_struct's
645 * immediately freed.
646 */
647void set_iounmap_nonlazy(void)
648{
649 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
650}
651
652/*
653 * Purges all lazily-freed vmap areas.
654 */
655static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
656{
657 struct llist_node *valist;
658 struct vmap_area *va;
659 struct vmap_area *n_va;
660 bool do_free = false;
661
662 lockdep_assert_held(&vmap_purge_lock);
663
664 valist = llist_del_all(&vmap_purge_list);
665 llist_for_each_entry(va, valist, purge_list) {
666 if (va->va_start < start)
667 start = va->va_start;
668 if (va->va_end > end)
669 end = va->va_end;
670 do_free = true;
671 }
672
673 if (!do_free)
674 return false;
675
676 flush_tlb_kernel_range(start, end);
677
678 spin_lock(&vmap_area_lock);
679 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
680 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
681
682 __free_vmap_area(va);
683 atomic_sub(nr, &vmap_lazy_nr);
684 cond_resched_lock(&vmap_area_lock);
685 }
686 spin_unlock(&vmap_area_lock);
687 return true;
688}
689
690/*
691 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
692 * is already purging.
693 */
694static void try_purge_vmap_area_lazy(void)
695{
696 if (mutex_trylock(&vmap_purge_lock)) {
697 __purge_vmap_area_lazy(ULONG_MAX, 0);
698 mutex_unlock(&vmap_purge_lock);
699 }
700}
701
702/*
703 * Kick off a purge of the outstanding lazy areas.
704 */
705static void purge_vmap_area_lazy(void)
706{
707 mutex_lock(&vmap_purge_lock);
708 purge_fragmented_blocks_allcpus();
709 __purge_vmap_area_lazy(ULONG_MAX, 0);
710 mutex_unlock(&vmap_purge_lock);
711}
712
713/*
714 * Free a vmap area, caller ensuring that the area has been unmapped
715 * and flush_cache_vunmap had been called for the correct range
716 * previously.
717 */
718static void free_vmap_area_noflush(struct vmap_area *va)
719{
720 int nr_lazy;
721
722 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
723 &vmap_lazy_nr);
724
725 /* After this point, we may free va at any time */
726 llist_add(&va->purge_list, &vmap_purge_list);
727
728 if (unlikely(nr_lazy > lazy_max_pages()))
729 try_purge_vmap_area_lazy();
730}
731
732/*
733 * Free and unmap a vmap area
734 */
735static void free_unmap_vmap_area(struct vmap_area *va)
736{
737 flush_cache_vunmap(va->va_start, va->va_end);
738 unmap_vmap_area(va);
739 if (debug_pagealloc_enabled())
740 flush_tlb_kernel_range(va->va_start, va->va_end);
741
742 free_vmap_area_noflush(va);
743}
744
745static struct vmap_area *find_vmap_area(unsigned long addr)
746{
747 struct vmap_area *va;
748
749 spin_lock(&vmap_area_lock);
750 va = __find_vmap_area(addr);
751 spin_unlock(&vmap_area_lock);
752
753 return va;
754}
755
756/*** Per cpu kva allocator ***/
757
758/*
759 * vmap space is limited especially on 32 bit architectures. Ensure there is
760 * room for at least 16 percpu vmap blocks per CPU.
761 */
762/*
763 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
764 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
765 * instead (we just need a rough idea)
766 */
767#if BITS_PER_LONG == 32
768#define VMALLOC_SPACE (128UL*1024*1024)
769#else
770#define VMALLOC_SPACE (128UL*1024*1024*1024)
771#endif
772
773#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
774#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
775#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
776#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
777#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
778#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
779#define VMAP_BBMAP_BITS \
780 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
781 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
782 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
783
784#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
785
786static bool vmap_initialized __read_mostly = false;
787
788struct vmap_block_queue {
789 spinlock_t lock;
790 struct list_head free;
791};
792
793struct vmap_block {
794 spinlock_t lock;
795 struct vmap_area *va;
796 unsigned long free, dirty;
797 unsigned long dirty_min, dirty_max; /*< dirty range */
798 struct list_head free_list;
799 struct rcu_head rcu_head;
800 struct list_head purge;
801};
802
803/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
804static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
805
806/*
807 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
808 * in the free path. Could get rid of this if we change the API to return a
809 * "cookie" from alloc, to be passed to free. But no big deal yet.
810 */
811static DEFINE_SPINLOCK(vmap_block_tree_lock);
812static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
813
814/*
815 * We should probably have a fallback mechanism to allocate virtual memory
816 * out of partially filled vmap blocks. However vmap block sizing should be
817 * fairly reasonable according to the vmalloc size, so it shouldn't be a
818 * big problem.
819 */
820
821static unsigned long addr_to_vb_idx(unsigned long addr)
822{
823 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
824 addr /= VMAP_BLOCK_SIZE;
825 return addr;
826}
827
828static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
829{
830 unsigned long addr;
831
832 addr = va_start + (pages_off << PAGE_SHIFT);
833 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
834 return (void *)addr;
835}
836
837/**
838 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
839 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
840 * @order: how many 2^order pages should be occupied in newly allocated block
841 * @gfp_mask: flags for the page level allocator
842 *
843 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
844 */
845static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
846{
847 struct vmap_block_queue *vbq;
848 struct vmap_block *vb;
849 struct vmap_area *va;
850 unsigned long vb_idx;
851 int node, err;
852 void *vaddr;
853
854 node = numa_node_id();
855
856 vb = kmalloc_node(sizeof(struct vmap_block),
857 gfp_mask & GFP_RECLAIM_MASK, node);
858 if (unlikely(!vb))
859 return ERR_PTR(-ENOMEM);
860
861 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
862 VMALLOC_START, VMALLOC_END,
863 node, gfp_mask);
864 if (IS_ERR(va)) {
865 kfree(vb);
866 return ERR_CAST(va);
867 }
868
869 err = radix_tree_preload(gfp_mask);
870 if (unlikely(err)) {
871 kfree(vb);
872 free_vmap_area(va);
873 return ERR_PTR(err);
874 }
875
876 vaddr = vmap_block_vaddr(va->va_start, 0);
877 spin_lock_init(&vb->lock);
878 vb->va = va;
879 /* At least something should be left free */
880 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
881 vb->free = VMAP_BBMAP_BITS - (1UL << order);
882 vb->dirty = 0;
883 vb->dirty_min = VMAP_BBMAP_BITS;
884 vb->dirty_max = 0;
885 INIT_LIST_HEAD(&vb->free_list);
886
887 vb_idx = addr_to_vb_idx(va->va_start);
888 spin_lock(&vmap_block_tree_lock);
889 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
890 spin_unlock(&vmap_block_tree_lock);
891 BUG_ON(err);
892 radix_tree_preload_end();
893
894 vbq = &get_cpu_var(vmap_block_queue);
895 spin_lock(&vbq->lock);
896 list_add_tail_rcu(&vb->free_list, &vbq->free);
897 spin_unlock(&vbq->lock);
898 put_cpu_var(vmap_block_queue);
899
900 return vaddr;
901}
902
903static void free_vmap_block(struct vmap_block *vb)
904{
905 struct vmap_block *tmp;
906 unsigned long vb_idx;
907
908 vb_idx = addr_to_vb_idx(vb->va->va_start);
909 spin_lock(&vmap_block_tree_lock);
910 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
911 spin_unlock(&vmap_block_tree_lock);
912 BUG_ON(tmp != vb);
913
914 free_vmap_area_noflush(vb->va);
915 kfree_rcu(vb, rcu_head);
916}
917
918static void purge_fragmented_blocks(int cpu)
919{
920 LIST_HEAD(purge);
921 struct vmap_block *vb;
922 struct vmap_block *n_vb;
923 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
924
925 rcu_read_lock();
926 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
927
928 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
929 continue;
930
931 spin_lock(&vb->lock);
932 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
933 vb->free = 0; /* prevent further allocs after releasing lock */
934 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
935 vb->dirty_min = 0;
936 vb->dirty_max = VMAP_BBMAP_BITS;
937 spin_lock(&vbq->lock);
938 list_del_rcu(&vb->free_list);
939 spin_unlock(&vbq->lock);
940 spin_unlock(&vb->lock);
941 list_add_tail(&vb->purge, &purge);
942 } else
943 spin_unlock(&vb->lock);
944 }
945 rcu_read_unlock();
946
947 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
948 list_del(&vb->purge);
949 free_vmap_block(vb);
950 }
951}
952
953static void purge_fragmented_blocks_allcpus(void)
954{
955 int cpu;
956
957 for_each_possible_cpu(cpu)
958 purge_fragmented_blocks(cpu);
959}
960
961static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
962{
963 struct vmap_block_queue *vbq;
964 struct vmap_block *vb;
965 void *vaddr = NULL;
966 unsigned int order;
967
968 BUG_ON(offset_in_page(size));
969 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
970 if (WARN_ON(size == 0)) {
971 /*
972 * Allocating 0 bytes isn't what caller wants since
973 * get_order(0) returns funny result. Just warn and terminate
974 * early.
975 */
976 return NULL;
977 }
978 order = get_order(size);
979
980 rcu_read_lock();
981 vbq = &get_cpu_var(vmap_block_queue);
982 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
983 unsigned long pages_off;
984
985 spin_lock(&vb->lock);
986 if (vb->free < (1UL << order)) {
987 spin_unlock(&vb->lock);
988 continue;
989 }
990
991 pages_off = VMAP_BBMAP_BITS - vb->free;
992 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
993 vb->free -= 1UL << order;
994 if (vb->free == 0) {
995 spin_lock(&vbq->lock);
996 list_del_rcu(&vb->free_list);
997 spin_unlock(&vbq->lock);
998 }
999
1000 spin_unlock(&vb->lock);
1001 break;
1002 }
1003
1004 put_cpu_var(vmap_block_queue);
1005 rcu_read_unlock();
1006
1007 /* Allocate new block if nothing was found */
1008 if (!vaddr)
1009 vaddr = new_vmap_block(order, gfp_mask);
1010
1011 return vaddr;
1012}
1013
1014static void vb_free(const void *addr, unsigned long size)
1015{
1016 unsigned long offset;
1017 unsigned long vb_idx;
1018 unsigned int order;
1019 struct vmap_block *vb;
1020
1021 BUG_ON(offset_in_page(size));
1022 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1023
1024 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1025
1026 order = get_order(size);
1027
1028 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1029 offset >>= PAGE_SHIFT;
1030
1031 vb_idx = addr_to_vb_idx((unsigned long)addr);
1032 rcu_read_lock();
1033 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1034 rcu_read_unlock();
1035 BUG_ON(!vb);
1036
1037 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1038
1039 if (debug_pagealloc_enabled())
1040 flush_tlb_kernel_range((unsigned long)addr,
1041 (unsigned long)addr + size);
1042
1043 spin_lock(&vb->lock);
1044
1045 /* Expand dirty range */
1046 vb->dirty_min = min(vb->dirty_min, offset);
1047 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1048
1049 vb->dirty += 1UL << order;
1050 if (vb->dirty == VMAP_BBMAP_BITS) {
1051 BUG_ON(vb->free);
1052 spin_unlock(&vb->lock);
1053 free_vmap_block(vb);
1054 } else
1055 spin_unlock(&vb->lock);
1056}
1057
1058/**
1059 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1060 *
1061 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1062 * to amortize TLB flushing overheads. What this means is that any page you
1063 * have now, may, in a former life, have been mapped into kernel virtual
1064 * address by the vmap layer and so there might be some CPUs with TLB entries
1065 * still referencing that page (additional to the regular 1:1 kernel mapping).
1066 *
1067 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1068 * be sure that none of the pages we have control over will have any aliases
1069 * from the vmap layer.
1070 */
1071void vm_unmap_aliases(void)
1072{
1073 unsigned long start = ULONG_MAX, end = 0;
1074 int cpu;
1075 int flush = 0;
1076
1077 if (unlikely(!vmap_initialized))
1078 return;
1079
1080 might_sleep();
1081
1082 for_each_possible_cpu(cpu) {
1083 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1084 struct vmap_block *vb;
1085
1086 rcu_read_lock();
1087 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1088 spin_lock(&vb->lock);
1089 if (vb->dirty) {
1090 unsigned long va_start = vb->va->va_start;
1091 unsigned long s, e;
1092
1093 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1094 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1095
1096 start = min(s, start);
1097 end = max(e, end);
1098
1099 flush = 1;
1100 }
1101 spin_unlock(&vb->lock);
1102 }
1103 rcu_read_unlock();
1104 }
1105
1106 mutex_lock(&vmap_purge_lock);
1107 purge_fragmented_blocks_allcpus();
1108 if (!__purge_vmap_area_lazy(start, end) && flush)
1109 flush_tlb_kernel_range(start, end);
1110 mutex_unlock(&vmap_purge_lock);
1111}
1112EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1113
1114/**
1115 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1116 * @mem: the pointer returned by vm_map_ram
1117 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1118 */
1119void vm_unmap_ram(const void *mem, unsigned int count)
1120{
1121 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1122 unsigned long addr = (unsigned long)mem;
1123 struct vmap_area *va;
1124
1125 might_sleep();
1126 BUG_ON(!addr);
1127 BUG_ON(addr < VMALLOC_START);
1128 BUG_ON(addr > VMALLOC_END);
1129 BUG_ON(!PAGE_ALIGNED(addr));
1130
1131 if (likely(count <= VMAP_MAX_ALLOC)) {
1132 debug_check_no_locks_freed(mem, size);
1133 vb_free(mem, size);
1134 return;
1135 }
1136
1137 va = find_vmap_area(addr);
1138 BUG_ON(!va);
1139 debug_check_no_locks_freed((void *)va->va_start,
1140 (va->va_end - va->va_start));
1141 free_unmap_vmap_area(va);
1142}
1143EXPORT_SYMBOL(vm_unmap_ram);
1144
1145/**
1146 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1147 * @pages: an array of pointers to the pages to be mapped
1148 * @count: number of pages
1149 * @node: prefer to allocate data structures on this node
1150 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1151 *
1152 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1153 * faster than vmap so it's good. But if you mix long-life and short-life
1154 * objects with vm_map_ram(), it could consume lots of address space through
1155 * fragmentation (especially on a 32bit machine). You could see failures in
1156 * the end. Please use this function for short-lived objects.
1157 *
1158 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1159 */
1160void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1161{
1162 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1163 unsigned long addr;
1164 void *mem;
1165
1166 if (likely(count <= VMAP_MAX_ALLOC)) {
1167 mem = vb_alloc(size, GFP_KERNEL);
1168 if (IS_ERR(mem))
1169 return NULL;
1170 addr = (unsigned long)mem;
1171 } else {
1172 struct vmap_area *va;
1173 va = alloc_vmap_area(size, PAGE_SIZE,
1174 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1175 if (IS_ERR(va))
1176 return NULL;
1177
1178 addr = va->va_start;
1179 mem = (void *)addr;
1180 }
1181 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1182 vm_unmap_ram(mem, count);
1183 return NULL;
1184 }
1185 return mem;
1186}
1187EXPORT_SYMBOL(vm_map_ram);
1188
1189static struct vm_struct *vmlist __initdata;
1190/**
1191 * vm_area_add_early - add vmap area early during boot
1192 * @vm: vm_struct to add
1193 *
1194 * This function is used to add fixed kernel vm area to vmlist before
1195 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1196 * should contain proper values and the other fields should be zero.
1197 *
1198 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1199 */
1200void __init vm_area_add_early(struct vm_struct *vm)
1201{
1202 struct vm_struct *tmp, **p;
1203
1204 BUG_ON(vmap_initialized);
1205 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1206 if (tmp->addr >= vm->addr) {
1207 BUG_ON(tmp->addr < vm->addr + vm->size);
1208 break;
1209 } else
1210 BUG_ON(tmp->addr + tmp->size > vm->addr);
1211 }
1212 vm->next = *p;
1213 *p = vm;
1214}
1215
1216/**
1217 * vm_area_register_early - register vmap area early during boot
1218 * @vm: vm_struct to register
1219 * @align: requested alignment
1220 *
1221 * This function is used to register kernel vm area before
1222 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1223 * proper values on entry and other fields should be zero. On return,
1224 * vm->addr contains the allocated address.
1225 *
1226 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1227 */
1228void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1229{
1230 static size_t vm_init_off __initdata;
1231 unsigned long addr;
1232
1233 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1234 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1235
1236 vm->addr = (void *)addr;
1237
1238 vm_area_add_early(vm);
1239}
1240
1241void __init vmalloc_init(void)
1242{
1243 struct vmap_area *va;
1244 struct vm_struct *tmp;
1245 int i;
1246
1247 for_each_possible_cpu(i) {
1248 struct vmap_block_queue *vbq;
1249 struct vfree_deferred *p;
1250
1251 vbq = &per_cpu(vmap_block_queue, i);
1252 spin_lock_init(&vbq->lock);
1253 INIT_LIST_HEAD(&vbq->free);
1254 p = &per_cpu(vfree_deferred, i);
1255 init_llist_head(&p->list);
1256 INIT_WORK(&p->wq, free_work);
1257 }
1258
1259 /* Import existing vmlist entries. */
1260 for (tmp = vmlist; tmp; tmp = tmp->next) {
1261 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1262 va->flags = VM_VM_AREA;
1263 va->va_start = (unsigned long)tmp->addr;
1264 va->va_end = va->va_start + tmp->size;
1265 va->vm = tmp;
1266 __insert_vmap_area(va);
1267 }
1268
1269 vmap_area_pcpu_hole = VMALLOC_END;
1270
1271 vmap_initialized = true;
1272}
1273
1274/**
1275 * map_kernel_range_noflush - map kernel VM area with the specified pages
1276 * @addr: start of the VM area to map
1277 * @size: size of the VM area to map
1278 * @prot: page protection flags to use
1279 * @pages: pages to map
1280 *
1281 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1282 * specify should have been allocated using get_vm_area() and its
1283 * friends.
1284 *
1285 * NOTE:
1286 * This function does NOT do any cache flushing. The caller is
1287 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1288 * before calling this function.
1289 *
1290 * RETURNS:
1291 * The number of pages mapped on success, -errno on failure.
1292 */
1293int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1294 pgprot_t prot, struct page **pages)
1295{
1296 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1297}
1298
1299/**
1300 * unmap_kernel_range_noflush - unmap kernel VM area
1301 * @addr: start of the VM area to unmap
1302 * @size: size of the VM area to unmap
1303 *
1304 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1305 * specify should have been allocated using get_vm_area() and its
1306 * friends.
1307 *
1308 * NOTE:
1309 * This function does NOT do any cache flushing. The caller is
1310 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1311 * before calling this function and flush_tlb_kernel_range() after.
1312 */
1313void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1314{
1315 vunmap_page_range(addr, addr + size);
1316}
1317EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1318
1319/**
1320 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1321 * @addr: start of the VM area to unmap
1322 * @size: size of the VM area to unmap
1323 *
1324 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1325 * the unmapping and tlb after.
1326 */
1327void unmap_kernel_range(unsigned long addr, unsigned long size)
1328{
1329 unsigned long end = addr + size;
1330
1331 flush_cache_vunmap(addr, end);
1332 vunmap_page_range(addr, end);
1333 flush_tlb_kernel_range(addr, end);
1334}
1335EXPORT_SYMBOL_GPL(unmap_kernel_range);
1336
1337int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1338{
1339 unsigned long addr = (unsigned long)area->addr;
1340 unsigned long end = addr + get_vm_area_size(area);
1341 int err;
1342
1343 err = vmap_page_range(addr, end, prot, pages);
1344
1345 return err > 0 ? 0 : err;
1346}
1347EXPORT_SYMBOL_GPL(map_vm_area);
1348
1349static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1350 unsigned long flags, const void *caller)
1351{
1352 spin_lock(&vmap_area_lock);
1353 vm->flags = flags;
1354 vm->addr = (void *)va->va_start;
1355 vm->size = va->va_end - va->va_start;
1356 vm->caller = caller;
1357 va->vm = vm;
1358 va->flags |= VM_VM_AREA;
1359 spin_unlock(&vmap_area_lock);
1360}
1361
1362static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1363{
1364 /*
1365 * Before removing VM_UNINITIALIZED,
1366 * we should make sure that vm has proper values.
1367 * Pair with smp_rmb() in show_numa_info().
1368 */
1369 smp_wmb();
1370 vm->flags &= ~VM_UNINITIALIZED;
1371}
1372
1373static struct vm_struct *__get_vm_area_node(unsigned long size,
1374 unsigned long align, unsigned long flags, unsigned long start,
1375 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1376{
1377 struct vmap_area *va;
1378 struct vm_struct *area;
1379
1380 BUG_ON(in_interrupt());
1381 size = PAGE_ALIGN(size);
1382 if (unlikely(!size))
1383 return NULL;
1384
1385 if (flags & VM_IOREMAP)
1386 align = 1ul << clamp_t(int, get_count_order_long(size),
1387 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1388
1389 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1390 if (unlikely(!area))
1391 return NULL;
1392
1393 if (!(flags & VM_NO_GUARD))
1394 size += PAGE_SIZE;
1395
1396 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1397 if (IS_ERR(va)) {
1398 kfree(area);
1399 return NULL;
1400 }
1401
1402 setup_vmalloc_vm(area, va, flags, caller);
1403
1404 return area;
1405}
1406
1407struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1408 unsigned long start, unsigned long end)
1409{
1410 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1411 GFP_KERNEL, __builtin_return_address(0));
1412}
1413EXPORT_SYMBOL_GPL(__get_vm_area);
1414
1415struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1416 unsigned long start, unsigned long end,
1417 const void *caller)
1418{
1419 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1420 GFP_KERNEL, caller);
1421}
1422
1423/**
1424 * get_vm_area - reserve a contiguous kernel virtual area
1425 * @size: size of the area
1426 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1427 *
1428 * Search an area of @size in the kernel virtual mapping area,
1429 * and reserved it for out purposes. Returns the area descriptor
1430 * on success or %NULL on failure.
1431 */
1432struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1433{
1434 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1435 NUMA_NO_NODE, GFP_KERNEL,
1436 __builtin_return_address(0));
1437}
1438
1439struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1440 const void *caller)
1441{
1442 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1443 NUMA_NO_NODE, GFP_KERNEL, caller);
1444}
1445
1446/**
1447 * find_vm_area - find a continuous kernel virtual area
1448 * @addr: base address
1449 *
1450 * Search for the kernel VM area starting at @addr, and return it.
1451 * It is up to the caller to do all required locking to keep the returned
1452 * pointer valid.
1453 */
1454struct vm_struct *find_vm_area(const void *addr)
1455{
1456 struct vmap_area *va;
1457
1458 va = find_vmap_area((unsigned long)addr);
1459 if (va && va->flags & VM_VM_AREA)
1460 return va->vm;
1461
1462 return NULL;
1463}
1464
1465/**
1466 * remove_vm_area - find and remove a continuous kernel virtual area
1467 * @addr: base address
1468 *
1469 * Search for the kernel VM area starting at @addr, and remove it.
1470 * This function returns the found VM area, but using it is NOT safe
1471 * on SMP machines, except for its size or flags.
1472 */
1473struct vm_struct *remove_vm_area(const void *addr)
1474{
1475 struct vmap_area *va;
1476
1477 might_sleep();
1478
1479 va = find_vmap_area((unsigned long)addr);
1480 if (va && va->flags & VM_VM_AREA) {
1481 struct vm_struct *vm = va->vm;
1482
1483 spin_lock(&vmap_area_lock);
1484 va->vm = NULL;
1485 va->flags &= ~VM_VM_AREA;
1486 va->flags |= VM_LAZY_FREE;
1487 spin_unlock(&vmap_area_lock);
1488
1489 kasan_free_shadow(vm);
1490 free_unmap_vmap_area(va);
1491
1492 return vm;
1493 }
1494 return NULL;
1495}
1496
1497static void __vunmap(const void *addr, int deallocate_pages)
1498{
1499 struct vm_struct *area;
1500
1501 if (!addr)
1502 return;
1503
1504 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1505 addr))
1506 return;
1507
1508 area = find_vmap_area((unsigned long)addr)->vm;
1509 if (unlikely(!area)) {
1510 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1511 addr);
1512 return;
1513 }
1514
1515 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
1516 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
1517
1518 remove_vm_area(addr);
1519 if (deallocate_pages) {
1520 int i;
1521
1522 for (i = 0; i < area->nr_pages; i++) {
1523 struct page *page = area->pages[i];
1524
1525 BUG_ON(!page);
1526 __free_pages(page, 0);
1527 }
1528
1529 kvfree(area->pages);
1530 }
1531
1532 kfree(area);
1533 return;
1534}
1535
1536static inline void __vfree_deferred(const void *addr)
1537{
1538 /*
1539 * Use raw_cpu_ptr() because this can be called from preemptible
1540 * context. Preemption is absolutely fine here, because the llist_add()
1541 * implementation is lockless, so it works even if we are adding to
1542 * nother cpu's list. schedule_work() should be fine with this too.
1543 */
1544 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1545
1546 if (llist_add((struct llist_node *)addr, &p->list))
1547 schedule_work(&p->wq);
1548}
1549
1550/**
1551 * vfree_atomic - release memory allocated by vmalloc()
1552 * @addr: memory base address
1553 *
1554 * This one is just like vfree() but can be called in any atomic context
1555 * except NMIs.
1556 */
1557void vfree_atomic(const void *addr)
1558{
1559 BUG_ON(in_nmi());
1560
1561 kmemleak_free(addr);
1562
1563 if (!addr)
1564 return;
1565 __vfree_deferred(addr);
1566}
1567
1568/**
1569 * vfree - release memory allocated by vmalloc()
1570 * @addr: memory base address
1571 *
1572 * Free the virtually continuous memory area starting at @addr, as
1573 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1574 * NULL, no operation is performed.
1575 *
1576 * Must not be called in NMI context (strictly speaking, only if we don't
1577 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1578 * conventions for vfree() arch-depenedent would be a really bad idea)
1579 *
1580 * May sleep if called *not* from interrupt context.
1581 *
1582 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1583 */
1584void vfree(const void *addr)
1585{
1586 BUG_ON(in_nmi());
1587
1588 kmemleak_free(addr);
1589
1590 might_sleep_if(!in_interrupt());
1591
1592 if (!addr)
1593 return;
1594 if (unlikely(in_interrupt()))
1595 __vfree_deferred(addr);
1596 else
1597 __vunmap(addr, 1);
1598}
1599EXPORT_SYMBOL(vfree);
1600
1601/**
1602 * vunmap - release virtual mapping obtained by vmap()
1603 * @addr: memory base address
1604 *
1605 * Free the virtually contiguous memory area starting at @addr,
1606 * which was created from the page array passed to vmap().
1607 *
1608 * Must not be called in interrupt context.
1609 */
1610void vunmap(const void *addr)
1611{
1612 BUG_ON(in_interrupt());
1613 might_sleep();
1614 if (addr)
1615 __vunmap(addr, 0);
1616}
1617EXPORT_SYMBOL(vunmap);
1618
1619/**
1620 * vmap - map an array of pages into virtually contiguous space
1621 * @pages: array of page pointers
1622 * @count: number of pages to map
1623 * @flags: vm_area->flags
1624 * @prot: page protection for the mapping
1625 *
1626 * Maps @count pages from @pages into contiguous kernel virtual
1627 * space.
1628 */
1629void *vmap(struct page **pages, unsigned int count,
1630 unsigned long flags, pgprot_t prot)
1631{
1632 struct vm_struct *area;
1633 unsigned long size; /* In bytes */
1634
1635 might_sleep();
1636
1637 if (count > totalram_pages())
1638 return NULL;
1639
1640 size = (unsigned long)count << PAGE_SHIFT;
1641 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1642 if (!area)
1643 return NULL;
1644
1645 if (map_vm_area(area, prot, pages)) {
1646 vunmap(area->addr);
1647 return NULL;
1648 }
1649
1650 return area->addr;
1651}
1652EXPORT_SYMBOL(vmap);
1653
1654static void *__vmalloc_node(unsigned long size, unsigned long align,
1655 gfp_t gfp_mask, pgprot_t prot,
1656 int node, const void *caller);
1657static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1658 pgprot_t prot, int node)
1659{
1660 struct page **pages;
1661 unsigned int nr_pages, array_size, i;
1662 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1663 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1664 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
1665 0 :
1666 __GFP_HIGHMEM;
1667
1668 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1669 array_size = (nr_pages * sizeof(struct page *));
1670
1671 area->nr_pages = nr_pages;
1672 /* Please note that the recursion is strictly bounded. */
1673 if (array_size > PAGE_SIZE) {
1674 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
1675 PAGE_KERNEL, node, area->caller);
1676 } else {
1677 pages = kmalloc_node(array_size, nested_gfp, node);
1678 }
1679 area->pages = pages;
1680 if (!area->pages) {
1681 remove_vm_area(area->addr);
1682 kfree(area);
1683 return NULL;
1684 }
1685
1686 for (i = 0; i < area->nr_pages; i++) {
1687 struct page *page;
1688
1689 if (node == NUMA_NO_NODE)
1690 page = alloc_page(alloc_mask|highmem_mask);
1691 else
1692 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
1693
1694 if (unlikely(!page)) {
1695 /* Successfully allocated i pages, free them in __vunmap() */
1696 area->nr_pages = i;
1697 goto fail;
1698 }
1699 area->pages[i] = page;
1700 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
1701 cond_resched();
1702 }
1703
1704 if (map_vm_area(area, prot, pages))
1705 goto fail;
1706 return area->addr;
1707
1708fail:
1709 warn_alloc(gfp_mask, NULL,
1710 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1711 (area->nr_pages*PAGE_SIZE), area->size);
1712 vfree(area->addr);
1713 return NULL;
1714}
1715
1716/**
1717 * __vmalloc_node_range - allocate virtually contiguous memory
1718 * @size: allocation size
1719 * @align: desired alignment
1720 * @start: vm area range start
1721 * @end: vm area range end
1722 * @gfp_mask: flags for the page level allocator
1723 * @prot: protection mask for the allocated pages
1724 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1725 * @node: node to use for allocation or NUMA_NO_NODE
1726 * @caller: caller's return address
1727 *
1728 * Allocate enough pages to cover @size from the page level
1729 * allocator with @gfp_mask flags. Map them into contiguous
1730 * kernel virtual space, using a pagetable protection of @prot.
1731 */
1732void *__vmalloc_node_range(unsigned long size, unsigned long align,
1733 unsigned long start, unsigned long end, gfp_t gfp_mask,
1734 pgprot_t prot, unsigned long vm_flags, int node,
1735 const void *caller)
1736{
1737 struct vm_struct *area;
1738 void *addr;
1739 unsigned long real_size = size;
1740
1741 size = PAGE_ALIGN(size);
1742 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
1743 goto fail;
1744
1745 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1746 vm_flags, start, end, node, gfp_mask, caller);
1747 if (!area)
1748 goto fail;
1749
1750 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1751 if (!addr)
1752 return NULL;
1753
1754 /*
1755 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1756 * flag. It means that vm_struct is not fully initialized.
1757 * Now, it is fully initialized, so remove this flag here.
1758 */
1759 clear_vm_uninitialized_flag(area);
1760
1761 kmemleak_vmalloc(area, size, gfp_mask);
1762
1763 return addr;
1764
1765fail:
1766 warn_alloc(gfp_mask, NULL,
1767 "vmalloc: allocation failure: %lu bytes", real_size);
1768 return NULL;
1769}
1770
1771/**
1772 * __vmalloc_node - allocate virtually contiguous memory
1773 * @size: allocation size
1774 * @align: desired alignment
1775 * @gfp_mask: flags for the page level allocator
1776 * @prot: protection mask for the allocated pages
1777 * @node: node to use for allocation or NUMA_NO_NODE
1778 * @caller: caller's return address
1779 *
1780 * Allocate enough pages to cover @size from the page level
1781 * allocator with @gfp_mask flags. Map them into contiguous
1782 * kernel virtual space, using a pagetable protection of @prot.
1783 *
1784 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1785 * and __GFP_NOFAIL are not supported
1786 *
1787 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1788 * with mm people.
1789 *
1790 */
1791static void *__vmalloc_node(unsigned long size, unsigned long align,
1792 gfp_t gfp_mask, pgprot_t prot,
1793 int node, const void *caller)
1794{
1795 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1796 gfp_mask, prot, 0, node, caller);
1797}
1798
1799void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1800{
1801 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1802 __builtin_return_address(0));
1803}
1804EXPORT_SYMBOL(__vmalloc);
1805
1806static inline void *__vmalloc_node_flags(unsigned long size,
1807 int node, gfp_t flags)
1808{
1809 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1810 node, __builtin_return_address(0));
1811}
1812
1813
1814void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
1815 void *caller)
1816{
1817 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
1818}
1819
1820/**
1821 * vmalloc - allocate virtually contiguous memory
1822 * @size: allocation size
1823 * Allocate enough pages to cover @size from the page level
1824 * allocator and map them into contiguous kernel virtual space.
1825 *
1826 * For tight control over page level allocator and protection flags
1827 * use __vmalloc() instead.
1828 */
1829void *vmalloc(unsigned long size)
1830{
1831 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1832 GFP_KERNEL);
1833}
1834EXPORT_SYMBOL(vmalloc);
1835
1836/**
1837 * vzalloc - allocate virtually contiguous memory with zero fill
1838 * @size: allocation size
1839 * Allocate enough pages to cover @size from the page level
1840 * allocator and map them into contiguous kernel virtual space.
1841 * The memory allocated is set to zero.
1842 *
1843 * For tight control over page level allocator and protection flags
1844 * use __vmalloc() instead.
1845 */
1846void *vzalloc(unsigned long size)
1847{
1848 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1849 GFP_KERNEL | __GFP_ZERO);
1850}
1851EXPORT_SYMBOL(vzalloc);
1852
1853/**
1854 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1855 * @size: allocation size
1856 *
1857 * The resulting memory area is zeroed so it can be mapped to userspace
1858 * without leaking data.
1859 */
1860void *vmalloc_user(unsigned long size)
1861{
1862 struct vm_struct *area;
1863 void *ret;
1864
1865 ret = __vmalloc_node(size, SHMLBA,
1866 GFP_KERNEL | __GFP_ZERO,
1867 PAGE_KERNEL, NUMA_NO_NODE,
1868 __builtin_return_address(0));
1869 if (ret) {
1870 area = find_vm_area(ret);
1871 area->flags |= VM_USERMAP;
1872 }
1873 return ret;
1874}
1875EXPORT_SYMBOL(vmalloc_user);
1876
1877/**
1878 * vmalloc_node - allocate memory on a specific node
1879 * @size: allocation size
1880 * @node: numa node
1881 *
1882 * Allocate enough pages to cover @size from the page level
1883 * allocator and map them into contiguous kernel virtual space.
1884 *
1885 * For tight control over page level allocator and protection flags
1886 * use __vmalloc() instead.
1887 */
1888void *vmalloc_node(unsigned long size, int node)
1889{
1890 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
1891 node, __builtin_return_address(0));
1892}
1893EXPORT_SYMBOL(vmalloc_node);
1894
1895/**
1896 * vzalloc_node - allocate memory on a specific node with zero fill
1897 * @size: allocation size
1898 * @node: numa node
1899 *
1900 * Allocate enough pages to cover @size from the page level
1901 * allocator and map them into contiguous kernel virtual space.
1902 * The memory allocated is set to zero.
1903 *
1904 * For tight control over page level allocator and protection flags
1905 * use __vmalloc_node() instead.
1906 */
1907void *vzalloc_node(unsigned long size, int node)
1908{
1909 return __vmalloc_node_flags(size, node,
1910 GFP_KERNEL | __GFP_ZERO);
1911}
1912EXPORT_SYMBOL(vzalloc_node);
1913
1914/**
1915 * vmalloc_exec - allocate virtually contiguous, executable memory
1916 * @size: allocation size
1917 *
1918 * Kernel-internal function to allocate enough pages to cover @size
1919 * the page level allocator and map them into contiguous and
1920 * executable kernel virtual space.
1921 *
1922 * For tight control over page level allocator and protection flags
1923 * use __vmalloc() instead.
1924 */
1925
1926void *vmalloc_exec(unsigned long size)
1927{
1928 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL_EXEC,
1929 NUMA_NO_NODE, __builtin_return_address(0));
1930}
1931
1932#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1933#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
1934#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1935#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
1936#else
1937/*
1938 * 64b systems should always have either DMA or DMA32 zones. For others
1939 * GFP_DMA32 should do the right thing and use the normal zone.
1940 */
1941#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1942#endif
1943
1944/**
1945 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1946 * @size: allocation size
1947 *
1948 * Allocate enough 32bit PA addressable pages to cover @size from the
1949 * page level allocator and map them into contiguous kernel virtual space.
1950 */
1951void *vmalloc_32(unsigned long size)
1952{
1953 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1954 NUMA_NO_NODE, __builtin_return_address(0));
1955}
1956EXPORT_SYMBOL(vmalloc_32);
1957
1958/**
1959 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1960 * @size: allocation size
1961 *
1962 * The resulting memory area is 32bit addressable and zeroed so it can be
1963 * mapped to userspace without leaking data.
1964 */
1965void *vmalloc_32_user(unsigned long size)
1966{
1967 struct vm_struct *area;
1968 void *ret;
1969
1970 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1971 NUMA_NO_NODE, __builtin_return_address(0));
1972 if (ret) {
1973 area = find_vm_area(ret);
1974 area->flags |= VM_USERMAP;
1975 }
1976 return ret;
1977}
1978EXPORT_SYMBOL(vmalloc_32_user);
1979
1980/*
1981 * small helper routine , copy contents to buf from addr.
1982 * If the page is not present, fill zero.
1983 */
1984
1985static int aligned_vread(char *buf, char *addr, unsigned long count)
1986{
1987 struct page *p;
1988 int copied = 0;
1989
1990 while (count) {
1991 unsigned long offset, length;
1992
1993 offset = offset_in_page(addr);
1994 length = PAGE_SIZE - offset;
1995 if (length > count)
1996 length = count;
1997 p = vmalloc_to_page(addr);
1998 /*
1999 * To do safe access to this _mapped_ area, we need
2000 * lock. But adding lock here means that we need to add
2001 * overhead of vmalloc()/vfree() calles for this _debug_
2002 * interface, rarely used. Instead of that, we'll use
2003 * kmap() and get small overhead in this access function.
2004 */
2005 if (p) {
2006 /*
2007 * we can expect USER0 is not used (see vread/vwrite's
2008 * function description)
2009 */
2010 void *map = kmap_atomic(p);
2011 memcpy(buf, map + offset, length);
2012 kunmap_atomic(map);
2013 } else
2014 memset(buf, 0, length);
2015
2016 addr += length;
2017 buf += length;
2018 copied += length;
2019 count -= length;
2020 }
2021 return copied;
2022}
2023
2024static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2025{
2026 struct page *p;
2027 int copied = 0;
2028
2029 while (count) {
2030 unsigned long offset, length;
2031
2032 offset = offset_in_page(addr);
2033 length = PAGE_SIZE - offset;
2034 if (length > count)
2035 length = count;
2036 p = vmalloc_to_page(addr);
2037 /*
2038 * To do safe access to this _mapped_ area, we need
2039 * lock. But adding lock here means that we need to add
2040 * overhead of vmalloc()/vfree() calles for this _debug_
2041 * interface, rarely used. Instead of that, we'll use
2042 * kmap() and get small overhead in this access function.
2043 */
2044 if (p) {
2045 /*
2046 * we can expect USER0 is not used (see vread/vwrite's
2047 * function description)
2048 */
2049 void *map = kmap_atomic(p);
2050 memcpy(map + offset, buf, length);
2051 kunmap_atomic(map);
2052 }
2053 addr += length;
2054 buf += length;
2055 copied += length;
2056 count -= length;
2057 }
2058 return copied;
2059}
2060
2061/**
2062 * vread() - read vmalloc area in a safe way.
2063 * @buf: buffer for reading data
2064 * @addr: vm address.
2065 * @count: number of bytes to be read.
2066 *
2067 * Returns # of bytes which addr and buf should be increased.
2068 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2069 * includes any intersect with alive vmalloc area.
2070 *
2071 * This function checks that addr is a valid vmalloc'ed area, and
2072 * copy data from that area to a given buffer. If the given memory range
2073 * of [addr...addr+count) includes some valid address, data is copied to
2074 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2075 * IOREMAP area is treated as memory hole and no copy is done.
2076 *
2077 * If [addr...addr+count) doesn't includes any intersects with alive
2078 * vm_struct area, returns 0. @buf should be kernel's buffer.
2079 *
2080 * Note: In usual ops, vread() is never necessary because the caller
2081 * should know vmalloc() area is valid and can use memcpy().
2082 * This is for routines which have to access vmalloc area without
2083 * any informaion, as /dev/kmem.
2084 *
2085 */
2086
2087long vread(char *buf, char *addr, unsigned long count)
2088{
2089 struct vmap_area *va;
2090 struct vm_struct *vm;
2091 char *vaddr, *buf_start = buf;
2092 unsigned long buflen = count;
2093 unsigned long n;
2094
2095 /* Don't allow overflow */
2096 if ((unsigned long) addr + count < count)
2097 count = -(unsigned long) addr;
2098
2099 spin_lock(&vmap_area_lock);
2100 list_for_each_entry(va, &vmap_area_list, list) {
2101 if (!count)
2102 break;
2103
2104 if (!(va->flags & VM_VM_AREA))
2105 continue;
2106
2107 vm = va->vm;
2108 vaddr = (char *) vm->addr;
2109 if (addr >= vaddr + get_vm_area_size(vm))
2110 continue;
2111 while (addr < vaddr) {
2112 if (count == 0)
2113 goto finished;
2114 *buf = '\0';
2115 buf++;
2116 addr++;
2117 count--;
2118 }
2119 n = vaddr + get_vm_area_size(vm) - addr;
2120 if (n > count)
2121 n = count;
2122 if (!(vm->flags & VM_IOREMAP))
2123 aligned_vread(buf, addr, n);
2124 else /* IOREMAP area is treated as memory hole */
2125 memset(buf, 0, n);
2126 buf += n;
2127 addr += n;
2128 count -= n;
2129 }
2130finished:
2131 spin_unlock(&vmap_area_lock);
2132
2133 if (buf == buf_start)
2134 return 0;
2135 /* zero-fill memory holes */
2136 if (buf != buf_start + buflen)
2137 memset(buf, 0, buflen - (buf - buf_start));
2138
2139 return buflen;
2140}
2141
2142/**
2143 * vwrite() - write vmalloc area in a safe way.
2144 * @buf: buffer for source data
2145 * @addr: vm address.
2146 * @count: number of bytes to be read.
2147 *
2148 * Returns # of bytes which addr and buf should be incresed.
2149 * (same number to @count).
2150 * If [addr...addr+count) doesn't includes any intersect with valid
2151 * vmalloc area, returns 0.
2152 *
2153 * This function checks that addr is a valid vmalloc'ed area, and
2154 * copy data from a buffer to the given addr. If specified range of
2155 * [addr...addr+count) includes some valid address, data is copied from
2156 * proper area of @buf. If there are memory holes, no copy to hole.
2157 * IOREMAP area is treated as memory hole and no copy is done.
2158 *
2159 * If [addr...addr+count) doesn't includes any intersects with alive
2160 * vm_struct area, returns 0. @buf should be kernel's buffer.
2161 *
2162 * Note: In usual ops, vwrite() is never necessary because the caller
2163 * should know vmalloc() area is valid and can use memcpy().
2164 * This is for routines which have to access vmalloc area without
2165 * any informaion, as /dev/kmem.
2166 */
2167
2168long vwrite(char *buf, char *addr, unsigned long count)
2169{
2170 struct vmap_area *va;
2171 struct vm_struct *vm;
2172 char *vaddr;
2173 unsigned long n, buflen;
2174 int copied = 0;
2175
2176 /* Don't allow overflow */
2177 if ((unsigned long) addr + count < count)
2178 count = -(unsigned long) addr;
2179 buflen = count;
2180
2181 spin_lock(&vmap_area_lock);
2182 list_for_each_entry(va, &vmap_area_list, list) {
2183 if (!count)
2184 break;
2185
2186 if (!(va->flags & VM_VM_AREA))
2187 continue;
2188
2189 vm = va->vm;
2190 vaddr = (char *) vm->addr;
2191 if (addr >= vaddr + get_vm_area_size(vm))
2192 continue;
2193 while (addr < vaddr) {
2194 if (count == 0)
2195 goto finished;
2196 buf++;
2197 addr++;
2198 count--;
2199 }
2200 n = vaddr + get_vm_area_size(vm) - addr;
2201 if (n > count)
2202 n = count;
2203 if (!(vm->flags & VM_IOREMAP)) {
2204 aligned_vwrite(buf, addr, n);
2205 copied++;
2206 }
2207 buf += n;
2208 addr += n;
2209 count -= n;
2210 }
2211finished:
2212 spin_unlock(&vmap_area_lock);
2213 if (!copied)
2214 return 0;
2215 return buflen;
2216}
2217
2218/**
2219 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2220 * @vma: vma to cover
2221 * @uaddr: target user address to start at
2222 * @kaddr: virtual address of vmalloc kernel memory
2223 * @size: size of map area
2224 *
2225 * Returns: 0 for success, -Exxx on failure
2226 *
2227 * This function checks that @kaddr is a valid vmalloc'ed area,
2228 * and that it is big enough to cover the range starting at
2229 * @uaddr in @vma. Will return failure if that criteria isn't
2230 * met.
2231 *
2232 * Similar to remap_pfn_range() (see mm/memory.c)
2233 */
2234int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2235 void *kaddr, unsigned long size)
2236{
2237 struct vm_struct *area;
2238
2239 size = PAGE_ALIGN(size);
2240
2241 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2242 return -EINVAL;
2243
2244 area = find_vm_area(kaddr);
2245 if (!area)
2246 return -EINVAL;
2247
2248 if (!(area->flags & VM_USERMAP))
2249 return -EINVAL;
2250
2251 if (kaddr + size > area->addr + area->size)
2252 return -EINVAL;
2253
2254 do {
2255 struct page *page = vmalloc_to_page(kaddr);
2256 int ret;
2257
2258 ret = vm_insert_page(vma, uaddr, page);
2259 if (ret)
2260 return ret;
2261
2262 uaddr += PAGE_SIZE;
2263 kaddr += PAGE_SIZE;
2264 size -= PAGE_SIZE;
2265 } while (size > 0);
2266
2267 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2268
2269 return 0;
2270}
2271EXPORT_SYMBOL(remap_vmalloc_range_partial);
2272
2273/**
2274 * remap_vmalloc_range - map vmalloc pages to userspace
2275 * @vma: vma to cover (map full range of vma)
2276 * @addr: vmalloc memory
2277 * @pgoff: number of pages into addr before first page to map
2278 *
2279 * Returns: 0 for success, -Exxx on failure
2280 *
2281 * This function checks that addr is a valid vmalloc'ed area, and
2282 * that it is big enough to cover the vma. Will return failure if
2283 * that criteria isn't met.
2284 *
2285 * Similar to remap_pfn_range() (see mm/memory.c)
2286 */
2287int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2288 unsigned long pgoff)
2289{
2290 return remap_vmalloc_range_partial(vma, vma->vm_start,
2291 addr + (pgoff << PAGE_SHIFT),
2292 vma->vm_end - vma->vm_start);
2293}
2294EXPORT_SYMBOL(remap_vmalloc_range);
2295
2296/*
2297 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2298 * have one.
2299 */
2300void __weak vmalloc_sync_all(void)
2301{
2302}
2303
2304
2305static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2306{
2307 pte_t ***p = data;
2308
2309 if (p) {
2310 *(*p) = pte;
2311 (*p)++;
2312 }
2313 return 0;
2314}
2315
2316/**
2317 * alloc_vm_area - allocate a range of kernel address space
2318 * @size: size of the area
2319 * @ptes: returns the PTEs for the address space
2320 *
2321 * Returns: NULL on failure, vm_struct on success
2322 *
2323 * This function reserves a range of kernel address space, and
2324 * allocates pagetables to map that range. No actual mappings
2325 * are created.
2326 *
2327 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2328 * allocated for the VM area are returned.
2329 */
2330struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2331{
2332 struct vm_struct *area;
2333
2334 area = get_vm_area_caller(size, VM_IOREMAP,
2335 __builtin_return_address(0));
2336 if (area == NULL)
2337 return NULL;
2338
2339 /*
2340 * This ensures that page tables are constructed for this region
2341 * of kernel virtual address space and mapped into init_mm.
2342 */
2343 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2344 size, f, ptes ? &ptes : NULL)) {
2345 free_vm_area(area);
2346 return NULL;
2347 }
2348
2349 return area;
2350}
2351EXPORT_SYMBOL_GPL(alloc_vm_area);
2352
2353void free_vm_area(struct vm_struct *area)
2354{
2355 struct vm_struct *ret;
2356 ret = remove_vm_area(area->addr);
2357 BUG_ON(ret != area);
2358 kfree(area);
2359}
2360EXPORT_SYMBOL_GPL(free_vm_area);
2361
2362#ifdef CONFIG_SMP
2363static struct vmap_area *node_to_va(struct rb_node *n)
2364{
2365 return rb_entry_safe(n, struct vmap_area, rb_node);
2366}
2367
2368/**
2369 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2370 * @end: target address
2371 * @pnext: out arg for the next vmap_area
2372 * @pprev: out arg for the previous vmap_area
2373 *
2374 * Returns: %true if either or both of next and prev are found,
2375 * %false if no vmap_area exists
2376 *
2377 * Find vmap_areas end addresses of which enclose @end. ie. if not
2378 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2379 */
2380static bool pvm_find_next_prev(unsigned long end,
2381 struct vmap_area **pnext,
2382 struct vmap_area **pprev)
2383{
2384 struct rb_node *n = vmap_area_root.rb_node;
2385 struct vmap_area *va = NULL;
2386
2387 while (n) {
2388 va = rb_entry(n, struct vmap_area, rb_node);
2389 if (end < va->va_end)
2390 n = n->rb_left;
2391 else if (end > va->va_end)
2392 n = n->rb_right;
2393 else
2394 break;
2395 }
2396
2397 if (!va)
2398 return false;
2399
2400 if (va->va_end > end) {
2401 *pnext = va;
2402 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2403 } else {
2404 *pprev = va;
2405 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2406 }
2407 return true;
2408}
2409
2410/**
2411 * pvm_determine_end - find the highest aligned address between two vmap_areas
2412 * @pnext: in/out arg for the next vmap_area
2413 * @pprev: in/out arg for the previous vmap_area
2414 * @align: alignment
2415 *
2416 * Returns: determined end address
2417 *
2418 * Find the highest aligned address between *@pnext and *@pprev below
2419 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2420 * down address is between the end addresses of the two vmap_areas.
2421 *
2422 * Please note that the address returned by this function may fall
2423 * inside *@pnext vmap_area. The caller is responsible for checking
2424 * that.
2425 */
2426static unsigned long pvm_determine_end(struct vmap_area **pnext,
2427 struct vmap_area **pprev,
2428 unsigned long align)
2429{
2430 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2431 unsigned long addr;
2432
2433 if (*pnext)
2434 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2435 else
2436 addr = vmalloc_end;
2437
2438 while (*pprev && (*pprev)->va_end > addr) {
2439 *pnext = *pprev;
2440 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2441 }
2442
2443 return addr;
2444}
2445
2446/**
2447 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2448 * @offsets: array containing offset of each area
2449 * @sizes: array containing size of each area
2450 * @nr_vms: the number of areas to allocate
2451 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2452 *
2453 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2454 * vm_structs on success, %NULL on failure
2455 *
2456 * Percpu allocator wants to use congruent vm areas so that it can
2457 * maintain the offsets among percpu areas. This function allocates
2458 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2459 * be scattered pretty far, distance between two areas easily going up
2460 * to gigabytes. To avoid interacting with regular vmallocs, these
2461 * areas are allocated from top.
2462 *
2463 * Despite its complicated look, this allocator is rather simple. It
2464 * does everything top-down and scans areas from the end looking for
2465 * matching slot. While scanning, if any of the areas overlaps with
2466 * existing vmap_area, the base address is pulled down to fit the
2467 * area. Scanning is repeated till all the areas fit and then all
2468 * necessary data structures are inserted and the result is returned.
2469 */
2470struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2471 const size_t *sizes, int nr_vms,
2472 size_t align)
2473{
2474 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2475 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2476 struct vmap_area **vas, *prev, *next;
2477 struct vm_struct **vms;
2478 int area, area2, last_area, term_area;
2479 unsigned long base, start, end, last_end;
2480 bool purged = false;
2481
2482 /* verify parameters and allocate data structures */
2483 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2484 for (last_area = 0, area = 0; area < nr_vms; area++) {
2485 start = offsets[area];
2486 end = start + sizes[area];
2487
2488 /* is everything aligned properly? */
2489 BUG_ON(!IS_ALIGNED(offsets[area], align));
2490 BUG_ON(!IS_ALIGNED(sizes[area], align));
2491
2492 /* detect the area with the highest address */
2493 if (start > offsets[last_area])
2494 last_area = area;
2495
2496 for (area2 = area + 1; area2 < nr_vms; area2++) {
2497 unsigned long start2 = offsets[area2];
2498 unsigned long end2 = start2 + sizes[area2];
2499
2500 BUG_ON(start2 < end && start < end2);
2501 }
2502 }
2503 last_end = offsets[last_area] + sizes[last_area];
2504
2505 if (vmalloc_end - vmalloc_start < last_end) {
2506 WARN_ON(true);
2507 return NULL;
2508 }
2509
2510 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2511 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2512 if (!vas || !vms)
2513 goto err_free2;
2514
2515 for (area = 0; area < nr_vms; area++) {
2516 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2517 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2518 if (!vas[area] || !vms[area])
2519 goto err_free;
2520 }
2521retry:
2522 spin_lock(&vmap_area_lock);
2523
2524 /* start scanning - we scan from the top, begin with the last area */
2525 area = term_area = last_area;
2526 start = offsets[area];
2527 end = start + sizes[area];
2528
2529 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2530 base = vmalloc_end - last_end;
2531 goto found;
2532 }
2533 base = pvm_determine_end(&next, &prev, align) - end;
2534
2535 while (true) {
2536 BUG_ON(next && next->va_end <= base + end);
2537 BUG_ON(prev && prev->va_end > base + end);
2538
2539 /*
2540 * base might have underflowed, add last_end before
2541 * comparing.
2542 */
2543 if (base + last_end < vmalloc_start + last_end) {
2544 spin_unlock(&vmap_area_lock);
2545 if (!purged) {
2546 purge_vmap_area_lazy();
2547 purged = true;
2548 goto retry;
2549 }
2550 goto err_free;
2551 }
2552
2553 /*
2554 * If next overlaps, move base downwards so that it's
2555 * right below next and then recheck.
2556 */
2557 if (next && next->va_start < base + end) {
2558 base = pvm_determine_end(&next, &prev, align) - end;
2559 term_area = area;
2560 continue;
2561 }
2562
2563 /*
2564 * If prev overlaps, shift down next and prev and move
2565 * base so that it's right below new next and then
2566 * recheck.
2567 */
2568 if (prev && prev->va_end > base + start) {
2569 next = prev;
2570 prev = node_to_va(rb_prev(&next->rb_node));
2571 base = pvm_determine_end(&next, &prev, align) - end;
2572 term_area = area;
2573 continue;
2574 }
2575
2576 /*
2577 * This area fits, move on to the previous one. If
2578 * the previous one is the terminal one, we're done.
2579 */
2580 area = (area + nr_vms - 1) % nr_vms;
2581 if (area == term_area)
2582 break;
2583 start = offsets[area];
2584 end = start + sizes[area];
2585 pvm_find_next_prev(base + end, &next, &prev);
2586 }
2587found:
2588 /* we've found a fitting base, insert all va's */
2589 for (area = 0; area < nr_vms; area++) {
2590 struct vmap_area *va = vas[area];
2591
2592 va->va_start = base + offsets[area];
2593 va->va_end = va->va_start + sizes[area];
2594 __insert_vmap_area(va);
2595 }
2596
2597 vmap_area_pcpu_hole = base + offsets[last_area];
2598
2599 spin_unlock(&vmap_area_lock);
2600
2601 /* insert all vm's */
2602 for (area = 0; area < nr_vms; area++)
2603 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2604 pcpu_get_vm_areas);
2605
2606 kfree(vas);
2607 return vms;
2608
2609err_free:
2610 for (area = 0; area < nr_vms; area++) {
2611 kfree(vas[area]);
2612 kfree(vms[area]);
2613 }
2614err_free2:
2615 kfree(vas);
2616 kfree(vms);
2617 return NULL;
2618}
2619
2620/**
2621 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2622 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2623 * @nr_vms: the number of allocated areas
2624 *
2625 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2626 */
2627void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2628{
2629 int i;
2630
2631 for (i = 0; i < nr_vms; i++)
2632 free_vm_area(vms[i]);
2633 kfree(vms);
2634}
2635#endif /* CONFIG_SMP */
2636
2637#ifdef CONFIG_PROC_FS
2638static void *s_start(struct seq_file *m, loff_t *pos)
2639 __acquires(&vmap_area_lock)
2640{
2641 spin_lock(&vmap_area_lock);
2642 return seq_list_start(&vmap_area_list, *pos);
2643}
2644
2645static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2646{
2647 return seq_list_next(p, &vmap_area_list, pos);
2648}
2649
2650static void s_stop(struct seq_file *m, void *p)
2651 __releases(&vmap_area_lock)
2652{
2653 spin_unlock(&vmap_area_lock);
2654}
2655
2656static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2657{
2658 if (IS_ENABLED(CONFIG_NUMA)) {
2659 unsigned int nr, *counters = m->private;
2660
2661 if (!counters)
2662 return;
2663
2664 if (v->flags & VM_UNINITIALIZED)
2665 return;
2666 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2667 smp_rmb();
2668
2669 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2670
2671 for (nr = 0; nr < v->nr_pages; nr++)
2672 counters[page_to_nid(v->pages[nr])]++;
2673
2674 for_each_node_state(nr, N_HIGH_MEMORY)
2675 if (counters[nr])
2676 seq_printf(m, " N%u=%u", nr, counters[nr]);
2677 }
2678}
2679
2680static int s_show(struct seq_file *m, void *p)
2681{
2682 struct vmap_area *va;
2683 struct vm_struct *v;
2684
2685 va = list_entry(p, struct vmap_area, list);
2686
2687 /*
2688 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2689 * behalf of vmap area is being tear down or vm_map_ram allocation.
2690 */
2691 if (!(va->flags & VM_VM_AREA)) {
2692 seq_printf(m, "0x%pK-0x%pK %7ld %s\n",
2693 (void *)va->va_start, (void *)va->va_end,
2694 va->va_end - va->va_start,
2695 va->flags & VM_LAZY_FREE ? "unpurged vm_area" : "vm_map_ram");
2696
2697 return 0;
2698 }
2699
2700 v = va->vm;
2701
2702 seq_printf(m, "0x%pK-0x%pK %7ld",
2703 v->addr, v->addr + v->size, v->size);
2704
2705 if (v->caller)
2706 seq_printf(m, " %pS", v->caller);
2707
2708 if (v->nr_pages)
2709 seq_printf(m, " pages=%d", v->nr_pages);
2710
2711 if (v->phys_addr)
2712 seq_printf(m, " phys=%pa", &v->phys_addr);
2713
2714 if (v->flags & VM_IOREMAP)
2715 seq_puts(m, " ioremap");
2716
2717 if (v->flags & VM_ALLOC)
2718 seq_puts(m, " vmalloc");
2719
2720 if (v->flags & VM_MAP)
2721 seq_puts(m, " vmap");
2722
2723 if (v->flags & VM_USERMAP)
2724 seq_puts(m, " user");
2725
2726 if (is_vmalloc_addr(v->pages))
2727 seq_puts(m, " vpages");
2728
2729 show_numa_info(m, v);
2730 seq_putc(m, '\n');
2731 return 0;
2732}
2733
2734static const struct seq_operations vmalloc_op = {
2735 .start = s_start,
2736 .next = s_next,
2737 .stop = s_stop,
2738 .show = s_show,
2739};
2740
2741static int __init proc_vmalloc_init(void)
2742{
2743 if (IS_ENABLED(CONFIG_NUMA))
2744 proc_create_seq_private("vmallocinfo", 0400, NULL,
2745 &vmalloc_op,
2746 nr_node_ids * sizeof(unsigned int), NULL);
2747 else
2748 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
2749 return 0;
2750}
2751module_init(proc_vmalloc_init);
2752
2753#endif
2754