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 /*
502 * Check also calculated address against the vstart,
503 * because it can be 0 because of big align request.
504 */
505 if (addr + size > vend || addr < vstart)
506 goto overflow;
507
508 va->va_start = addr;
509 va->va_end = addr + size;
510 va->flags = 0;
511 __insert_vmap_area(va);
512 free_vmap_cache = &va->rb_node;
513 spin_unlock(&vmap_area_lock);
514
515 BUG_ON(!IS_ALIGNED(va->va_start, align));
516 BUG_ON(va->va_start < vstart);
517 BUG_ON(va->va_end > vend);
518
519 return va;
520
521overflow:
522 spin_unlock(&vmap_area_lock);
523 if (!purged) {
524 purge_vmap_area_lazy();
525 purged = 1;
526 goto retry;
527 }
528
529 if (gfpflags_allow_blocking(gfp_mask)) {
530 unsigned long freed = 0;
531 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
532 if (freed > 0) {
533 purged = 0;
534 goto retry;
535 }
536 }
537
538 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
539 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
540 size);
541 kfree(va);
542 return ERR_PTR(-EBUSY);
543}
544
545int register_vmap_purge_notifier(struct notifier_block *nb)
546{
547 return blocking_notifier_chain_register(&vmap_notify_list, nb);
548}
549EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
550
551int unregister_vmap_purge_notifier(struct notifier_block *nb)
552{
553 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
554}
555EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
556
557static void __free_vmap_area(struct vmap_area *va)
558{
559 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
560
561 if (free_vmap_cache) {
562 if (va->va_end < cached_vstart) {
563 free_vmap_cache = NULL;
564 } else {
565 struct vmap_area *cache;
566 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
567 if (va->va_start <= cache->va_start) {
568 free_vmap_cache = rb_prev(&va->rb_node);
569 /*
570 * We don't try to update cached_hole_size or
571 * cached_align, but it won't go very wrong.
572 */
573 }
574 }
575 }
576 rb_erase(&va->rb_node, &vmap_area_root);
577 RB_CLEAR_NODE(&va->rb_node);
578 list_del_rcu(&va->list);
579
580 /*
581 * Track the highest possible candidate for pcpu area
582 * allocation. Areas outside of vmalloc area can be returned
583 * here too, consider only end addresses which fall inside
584 * vmalloc area proper.
585 */
586 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
587 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
588
589 kfree_rcu(va, rcu_head);
590}
591
592/*
593 * Free a region of KVA allocated by alloc_vmap_area
594 */
595static void free_vmap_area(struct vmap_area *va)
596{
597 spin_lock(&vmap_area_lock);
598 __free_vmap_area(va);
599 spin_unlock(&vmap_area_lock);
600}
601
602/*
603 * Clear the pagetable entries of a given vmap_area
604 */
605static void unmap_vmap_area(struct vmap_area *va)
606{
607 vunmap_page_range(va->va_start, va->va_end);
608}
609
610/*
611 * lazy_max_pages is the maximum amount of virtual address space we gather up
612 * before attempting to purge with a TLB flush.
613 *
614 * There is a tradeoff here: a larger number will cover more kernel page tables
615 * and take slightly longer to purge, but it will linearly reduce the number of
616 * global TLB flushes that must be performed. It would seem natural to scale
617 * this number up linearly with the number of CPUs (because vmapping activity
618 * could also scale linearly with the number of CPUs), however it is likely
619 * that in practice, workloads might be constrained in other ways that mean
620 * vmap activity will not scale linearly with CPUs. Also, I want to be
621 * conservative and not introduce a big latency on huge systems, so go with
622 * a less aggressive log scale. It will still be an improvement over the old
623 * code, and it will be simple to change the scale factor if we find that it
624 * becomes a problem on bigger systems.
625 */
626static unsigned long lazy_max_pages(void)
627{
628 unsigned int log;
629
630 log = fls(num_online_cpus());
631
632 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
633}
634
635static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
636
637/*
638 * Serialize vmap purging. There is no actual criticial section protected
639 * by this look, but we want to avoid concurrent calls for performance
640 * reasons and to make the pcpu_get_vm_areas more deterministic.
641 */
642static DEFINE_MUTEX(vmap_purge_lock);
643
644/* for per-CPU blocks */
645static void purge_fragmented_blocks_allcpus(void);
646
647/*
648 * called before a call to iounmap() if the caller wants vm_area_struct's
649 * immediately freed.
650 */
651void set_iounmap_nonlazy(void)
652{
653 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
654}
655
656/*
657 * Purges all lazily-freed vmap areas.
658 */
659static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
660{
661 struct llist_node *valist;
662 struct vmap_area *va;
663 struct vmap_area *n_va;
664 bool do_free = false;
665
666 lockdep_assert_held(&vmap_purge_lock);
667
668 valist = llist_del_all(&vmap_purge_list);
669 llist_for_each_entry(va, valist, purge_list) {
670 if (va->va_start < start)
671 start = va->va_start;
672 if (va->va_end > end)
673 end = va->va_end;
674 do_free = true;
675 }
676
677 if (!do_free)
678 return false;
679
680 flush_tlb_kernel_range(start, end);
681
682 spin_lock(&vmap_area_lock);
683 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
684 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
685
686 __free_vmap_area(va);
687 atomic_sub(nr, &vmap_lazy_nr);
688 cond_resched_lock(&vmap_area_lock);
689 }
690 spin_unlock(&vmap_area_lock);
691 return true;
692}
693
694/*
695 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
696 * is already purging.
697 */
698static void try_purge_vmap_area_lazy(void)
699{
700 if (mutex_trylock(&vmap_purge_lock)) {
701 __purge_vmap_area_lazy(ULONG_MAX, 0);
702 mutex_unlock(&vmap_purge_lock);
703 }
704}
705
706/*
707 * Kick off a purge of the outstanding lazy areas.
708 */
709static void purge_vmap_area_lazy(void)
710{
711 mutex_lock(&vmap_purge_lock);
712 purge_fragmented_blocks_allcpus();
713 __purge_vmap_area_lazy(ULONG_MAX, 0);
714 mutex_unlock(&vmap_purge_lock);
715}
716
717/*
718 * Free a vmap area, caller ensuring that the area has been unmapped
719 * and flush_cache_vunmap had been called for the correct range
720 * previously.
721 */
722static void free_vmap_area_noflush(struct vmap_area *va)
723{
724 int nr_lazy;
725
726 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
727 &vmap_lazy_nr);
728
729 /* After this point, we may free va at any time */
730 llist_add(&va->purge_list, &vmap_purge_list);
731
732 if (unlikely(nr_lazy > lazy_max_pages()))
733 try_purge_vmap_area_lazy();
734}
735
736/*
737 * Free and unmap a vmap area
738 */
739static void free_unmap_vmap_area(struct vmap_area *va)
740{
741 flush_cache_vunmap(va->va_start, va->va_end);
742 unmap_vmap_area(va);
743 if (debug_pagealloc_enabled())
744 flush_tlb_kernel_range(va->va_start, va->va_end);
745
746 free_vmap_area_noflush(va);
747}
748
749static struct vmap_area *find_vmap_area(unsigned long addr)
750{
751 struct vmap_area *va;
752
753 spin_lock(&vmap_area_lock);
754 va = __find_vmap_area(addr);
755 spin_unlock(&vmap_area_lock);
756
757 return va;
758}
759
760/*** Per cpu kva allocator ***/
761
762/*
763 * vmap space is limited especially on 32 bit architectures. Ensure there is
764 * room for at least 16 percpu vmap blocks per CPU.
765 */
766/*
767 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
768 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
769 * instead (we just need a rough idea)
770 */
771#if BITS_PER_LONG == 32
772#define VMALLOC_SPACE (128UL*1024*1024)
773#else
774#define VMALLOC_SPACE (128UL*1024*1024*1024)
775#endif
776
777#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
778#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
779#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
780#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
781#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
782#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
783#define VMAP_BBMAP_BITS \
784 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
785 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
786 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
787
788#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
789
790static bool vmap_initialized __read_mostly = false;
791
792struct vmap_block_queue {
793 spinlock_t lock;
794 struct list_head free;
795};
796
797struct vmap_block {
798 spinlock_t lock;
799 struct vmap_area *va;
800 unsigned long free, dirty;
801 unsigned long dirty_min, dirty_max; /*< dirty range */
802 struct list_head free_list;
803 struct rcu_head rcu_head;
804 struct list_head purge;
805};
806
807/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
808static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
809
810/*
811 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
812 * in the free path. Could get rid of this if we change the API to return a
813 * "cookie" from alloc, to be passed to free. But no big deal yet.
814 */
815static DEFINE_SPINLOCK(vmap_block_tree_lock);
816static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
817
818/*
819 * We should probably have a fallback mechanism to allocate virtual memory
820 * out of partially filled vmap blocks. However vmap block sizing should be
821 * fairly reasonable according to the vmalloc size, so it shouldn't be a
822 * big problem.
823 */
824
825static unsigned long addr_to_vb_idx(unsigned long addr)
826{
827 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
828 addr /= VMAP_BLOCK_SIZE;
829 return addr;
830}
831
832static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
833{
834 unsigned long addr;
835
836 addr = va_start + (pages_off << PAGE_SHIFT);
837 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
838 return (void *)addr;
839}
840
841/**
842 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
843 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
844 * @order: how many 2^order pages should be occupied in newly allocated block
845 * @gfp_mask: flags for the page level allocator
846 *
847 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
848 */
849static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
850{
851 struct vmap_block_queue *vbq;
852 struct vmap_block *vb;
853 struct vmap_area *va;
854 unsigned long vb_idx;
855 int node, err;
856 void *vaddr;
857
858 node = numa_node_id();
859
860 vb = kmalloc_node(sizeof(struct vmap_block),
861 gfp_mask & GFP_RECLAIM_MASK, node);
862 if (unlikely(!vb))
863 return ERR_PTR(-ENOMEM);
864
865 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
866 VMALLOC_START, VMALLOC_END,
867 node, gfp_mask);
868 if (IS_ERR(va)) {
869 kfree(vb);
870 return ERR_CAST(va);
871 }
872
873 err = radix_tree_preload(gfp_mask);
874 if (unlikely(err)) {
875 kfree(vb);
876 free_vmap_area(va);
877 return ERR_PTR(err);
878 }
879
880 vaddr = vmap_block_vaddr(va->va_start, 0);
881 spin_lock_init(&vb->lock);
882 vb->va = va;
883 /* At least something should be left free */
884 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
885 vb->free = VMAP_BBMAP_BITS - (1UL << order);
886 vb->dirty = 0;
887 vb->dirty_min = VMAP_BBMAP_BITS;
888 vb->dirty_max = 0;
889 INIT_LIST_HEAD(&vb->free_list);
890
891 vb_idx = addr_to_vb_idx(va->va_start);
892 spin_lock(&vmap_block_tree_lock);
893 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
894 spin_unlock(&vmap_block_tree_lock);
895 BUG_ON(err);
896 radix_tree_preload_end();
897
898 vbq = &get_cpu_var(vmap_block_queue);
899 spin_lock(&vbq->lock);
900 list_add_tail_rcu(&vb->free_list, &vbq->free);
901 spin_unlock(&vbq->lock);
902 put_cpu_var(vmap_block_queue);
903
904 return vaddr;
905}
906
907static void free_vmap_block(struct vmap_block *vb)
908{
909 struct vmap_block *tmp;
910 unsigned long vb_idx;
911
912 vb_idx = addr_to_vb_idx(vb->va->va_start);
913 spin_lock(&vmap_block_tree_lock);
914 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
915 spin_unlock(&vmap_block_tree_lock);
916 BUG_ON(tmp != vb);
917
918 free_vmap_area_noflush(vb->va);
919 kfree_rcu(vb, rcu_head);
920}
921
922static void purge_fragmented_blocks(int cpu)
923{
924 LIST_HEAD(purge);
925 struct vmap_block *vb;
926 struct vmap_block *n_vb;
927 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
928
929 rcu_read_lock();
930 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
931
932 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
933 continue;
934
935 spin_lock(&vb->lock);
936 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
937 vb->free = 0; /* prevent further allocs after releasing lock */
938 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
939 vb->dirty_min = 0;
940 vb->dirty_max = VMAP_BBMAP_BITS;
941 spin_lock(&vbq->lock);
942 list_del_rcu(&vb->free_list);
943 spin_unlock(&vbq->lock);
944 spin_unlock(&vb->lock);
945 list_add_tail(&vb->purge, &purge);
946 } else
947 spin_unlock(&vb->lock);
948 }
949 rcu_read_unlock();
950
951 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
952 list_del(&vb->purge);
953 free_vmap_block(vb);
954 }
955}
956
957static void purge_fragmented_blocks_allcpus(void)
958{
959 int cpu;
960
961 for_each_possible_cpu(cpu)
962 purge_fragmented_blocks(cpu);
963}
964
965static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
966{
967 struct vmap_block_queue *vbq;
968 struct vmap_block *vb;
969 void *vaddr = NULL;
970 unsigned int order;
971
972 BUG_ON(offset_in_page(size));
973 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
974 if (WARN_ON(size == 0)) {
975 /*
976 * Allocating 0 bytes isn't what caller wants since
977 * get_order(0) returns funny result. Just warn and terminate
978 * early.
979 */
980 return NULL;
981 }
982 order = get_order(size);
983
984 rcu_read_lock();
985 vbq = &get_cpu_var(vmap_block_queue);
986 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
987 unsigned long pages_off;
988
989 spin_lock(&vb->lock);
990 if (vb->free < (1UL << order)) {
991 spin_unlock(&vb->lock);
992 continue;
993 }
994
995 pages_off = VMAP_BBMAP_BITS - vb->free;
996 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
997 vb->free -= 1UL << order;
998 if (vb->free == 0) {
999 spin_lock(&vbq->lock);
1000 list_del_rcu(&vb->free_list);
1001 spin_unlock(&vbq->lock);
1002 }
1003
1004 spin_unlock(&vb->lock);
1005 break;
1006 }
1007
1008 put_cpu_var(vmap_block_queue);
1009 rcu_read_unlock();
1010
1011 /* Allocate new block if nothing was found */
1012 if (!vaddr)
1013 vaddr = new_vmap_block(order, gfp_mask);
1014
1015 return vaddr;
1016}
1017
1018static void vb_free(const void *addr, unsigned long size)
1019{
1020 unsigned long offset;
1021 unsigned long vb_idx;
1022 unsigned int order;
1023 struct vmap_block *vb;
1024
1025 BUG_ON(offset_in_page(size));
1026 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1027
1028 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1029
1030 order = get_order(size);
1031
1032 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1033 offset >>= PAGE_SHIFT;
1034
1035 vb_idx = addr_to_vb_idx((unsigned long)addr);
1036 rcu_read_lock();
1037 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1038 rcu_read_unlock();
1039 BUG_ON(!vb);
1040
1041 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1042
1043 if (debug_pagealloc_enabled())
1044 flush_tlb_kernel_range((unsigned long)addr,
1045 (unsigned long)addr + size);
1046
1047 spin_lock(&vb->lock);
1048
1049 /* Expand dirty range */
1050 vb->dirty_min = min(vb->dirty_min, offset);
1051 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1052
1053 vb->dirty += 1UL << order;
1054 if (vb->dirty == VMAP_BBMAP_BITS) {
1055 BUG_ON(vb->free);
1056 spin_unlock(&vb->lock);
1057 free_vmap_block(vb);
1058 } else
1059 spin_unlock(&vb->lock);
1060}
1061
1062/**
1063 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1064 *
1065 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1066 * to amortize TLB flushing overheads. What this means is that any page you
1067 * have now, may, in a former life, have been mapped into kernel virtual
1068 * address by the vmap layer and so there might be some CPUs with TLB entries
1069 * still referencing that page (additional to the regular 1:1 kernel mapping).
1070 *
1071 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1072 * be sure that none of the pages we have control over will have any aliases
1073 * from the vmap layer.
1074 */
1075void vm_unmap_aliases(void)
1076{
1077 unsigned long start = ULONG_MAX, end = 0;
1078 int cpu;
1079 int flush = 0;
1080
1081 if (unlikely(!vmap_initialized))
1082 return;
1083
1084 might_sleep();
1085
1086 for_each_possible_cpu(cpu) {
1087 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1088 struct vmap_block *vb;
1089
1090 rcu_read_lock();
1091 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1092 spin_lock(&vb->lock);
1093 if (vb->dirty) {
1094 unsigned long va_start = vb->va->va_start;
1095 unsigned long s, e;
1096
1097 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1098 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1099
1100 start = min(s, start);
1101 end = max(e, end);
1102
1103 flush = 1;
1104 }
1105 spin_unlock(&vb->lock);
1106 }
1107 rcu_read_unlock();
1108 }
1109
1110 mutex_lock(&vmap_purge_lock);
1111 purge_fragmented_blocks_allcpus();
1112 if (!__purge_vmap_area_lazy(start, end) && flush)
1113 flush_tlb_kernel_range(start, end);
1114 mutex_unlock(&vmap_purge_lock);
1115}
1116EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1117
1118/**
1119 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1120 * @mem: the pointer returned by vm_map_ram
1121 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1122 */
1123void vm_unmap_ram(const void *mem, unsigned int count)
1124{
1125 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1126 unsigned long addr = (unsigned long)mem;
1127 struct vmap_area *va;
1128
1129 might_sleep();
1130 BUG_ON(!addr);
1131 BUG_ON(addr < VMALLOC_START);
1132 BUG_ON(addr > VMALLOC_END);
1133 BUG_ON(!PAGE_ALIGNED(addr));
1134
1135 if (likely(count <= VMAP_MAX_ALLOC)) {
1136 debug_check_no_locks_freed(mem, size);
1137 vb_free(mem, size);
1138 return;
1139 }
1140
1141 va = find_vmap_area(addr);
1142 BUG_ON(!va);
1143 debug_check_no_locks_freed((void *)va->va_start,
1144 (va->va_end - va->va_start));
1145 free_unmap_vmap_area(va);
1146}
1147EXPORT_SYMBOL(vm_unmap_ram);
1148
1149/**
1150 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1151 * @pages: an array of pointers to the pages to be mapped
1152 * @count: number of pages
1153 * @node: prefer to allocate data structures on this node
1154 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1155 *
1156 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1157 * faster than vmap so it's good. But if you mix long-life and short-life
1158 * objects with vm_map_ram(), it could consume lots of address space through
1159 * fragmentation (especially on a 32bit machine). You could see failures in
1160 * the end. Please use this function for short-lived objects.
1161 *
1162 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1163 */
1164void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1165{
1166 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1167 unsigned long addr;
1168 void *mem;
1169
1170 if (likely(count <= VMAP_MAX_ALLOC)) {
1171 mem = vb_alloc(size, GFP_KERNEL);
1172 if (IS_ERR(mem))
1173 return NULL;
1174 addr = (unsigned long)mem;
1175 } else {
1176 struct vmap_area *va;
1177 va = alloc_vmap_area(size, PAGE_SIZE,
1178 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1179 if (IS_ERR(va))
1180 return NULL;
1181
1182 addr = va->va_start;
1183 mem = (void *)addr;
1184 }
1185 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1186 vm_unmap_ram(mem, count);
1187 return NULL;
1188 }
1189 return mem;
1190}
1191EXPORT_SYMBOL(vm_map_ram);
1192
1193static struct vm_struct *vmlist __initdata;
1194
1195/**
1196 * vm_area_add_early - add vmap area early during boot
1197 * @vm: vm_struct to add
1198 *
1199 * This function is used to add fixed kernel vm area to vmlist before
1200 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1201 * should contain proper values and the other fields should be zero.
1202 *
1203 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1204 */
1205void __init vm_area_add_early(struct vm_struct *vm)
1206{
1207 struct vm_struct *tmp, **p;
1208
1209 BUG_ON(vmap_initialized);
1210 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1211 if (tmp->addr >= vm->addr) {
1212 BUG_ON(tmp->addr < vm->addr + vm->size);
1213 break;
1214 } else
1215 BUG_ON(tmp->addr + tmp->size > vm->addr);
1216 }
1217 vm->next = *p;
1218 *p = vm;
1219}
1220
1221/**
1222 * vm_area_register_early - register vmap area early during boot
1223 * @vm: vm_struct to register
1224 * @align: requested alignment
1225 *
1226 * This function is used to register kernel vm area before
1227 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1228 * proper values on entry and other fields should be zero. On return,
1229 * vm->addr contains the allocated address.
1230 *
1231 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1232 */
1233void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1234{
1235 static size_t vm_init_off __initdata;
1236 unsigned long addr;
1237
1238 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1239 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1240
1241 vm->addr = (void *)addr;
1242
1243 vm_area_add_early(vm);
1244}
1245
1246void __init vmalloc_init(void)
1247{
1248 struct vmap_area *va;
1249 struct vm_struct *tmp;
1250 int i;
1251
1252 for_each_possible_cpu(i) {
1253 struct vmap_block_queue *vbq;
1254 struct vfree_deferred *p;
1255
1256 vbq = &per_cpu(vmap_block_queue, i);
1257 spin_lock_init(&vbq->lock);
1258 INIT_LIST_HEAD(&vbq->free);
1259 p = &per_cpu(vfree_deferred, i);
1260 init_llist_head(&p->list);
1261 INIT_WORK(&p->wq, free_work);
1262 }
1263
1264 /* Import existing vmlist entries. */
1265 for (tmp = vmlist; tmp; tmp = tmp->next) {
1266 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1267 va->flags = VM_VM_AREA;
1268 va->va_start = (unsigned long)tmp->addr;
1269 va->va_end = va->va_start + tmp->size;
1270 va->vm = tmp;
1271 __insert_vmap_area(va);
1272 }
1273
1274 vmap_area_pcpu_hole = VMALLOC_END;
1275
1276 vmap_initialized = true;
1277}
1278
1279/**
1280 * map_kernel_range_noflush - map kernel VM area with the specified pages
1281 * @addr: start of the VM area to map
1282 * @size: size of the VM area to map
1283 * @prot: page protection flags to use
1284 * @pages: pages to map
1285 *
1286 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1287 * specify should have been allocated using get_vm_area() and its
1288 * friends.
1289 *
1290 * NOTE:
1291 * This function does NOT do any cache flushing. The caller is
1292 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1293 * before calling this function.
1294 *
1295 * RETURNS:
1296 * The number of pages mapped on success, -errno on failure.
1297 */
1298int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1299 pgprot_t prot, struct page **pages)
1300{
1301 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1302}
1303
1304/**
1305 * unmap_kernel_range_noflush - unmap kernel VM area
1306 * @addr: start of the VM area to unmap
1307 * @size: size of the VM area to unmap
1308 *
1309 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1310 * specify should have been allocated using get_vm_area() and its
1311 * friends.
1312 *
1313 * NOTE:
1314 * This function does NOT do any cache flushing. The caller is
1315 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1316 * before calling this function and flush_tlb_kernel_range() after.
1317 */
1318void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1319{
1320 vunmap_page_range(addr, addr + size);
1321}
1322EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1323
1324/**
1325 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1326 * @addr: start of the VM area to unmap
1327 * @size: size of the VM area to unmap
1328 *
1329 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1330 * the unmapping and tlb after.
1331 */
1332void unmap_kernel_range(unsigned long addr, unsigned long size)
1333{
1334 unsigned long end = addr + size;
1335
1336 flush_cache_vunmap(addr, end);
1337 vunmap_page_range(addr, end);
1338 flush_tlb_kernel_range(addr, end);
1339}
1340EXPORT_SYMBOL_GPL(unmap_kernel_range);
1341
1342int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1343{
1344 unsigned long addr = (unsigned long)area->addr;
1345 unsigned long end = addr + get_vm_area_size(area);
1346 int err;
1347
1348 err = vmap_page_range(addr, end, prot, pages);
1349
1350 return err > 0 ? 0 : err;
1351}
1352EXPORT_SYMBOL_GPL(map_vm_area);
1353
1354static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1355 unsigned long flags, const void *caller)
1356{
1357 spin_lock(&vmap_area_lock);
1358 vm->flags = flags;
1359 vm->addr = (void *)va->va_start;
1360 vm->size = va->va_end - va->va_start;
1361 vm->caller = caller;
1362 va->vm = vm;
1363 va->flags |= VM_VM_AREA;
1364 spin_unlock(&vmap_area_lock);
1365}
1366
1367static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1368{
1369 /*
1370 * Before removing VM_UNINITIALIZED,
1371 * we should make sure that vm has proper values.
1372 * Pair with smp_rmb() in show_numa_info().
1373 */
1374 smp_wmb();
1375 vm->flags &= ~VM_UNINITIALIZED;
1376}
1377
1378static struct vm_struct *__get_vm_area_node(unsigned long size,
1379 unsigned long align, unsigned long flags, unsigned long start,
1380 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1381{
1382 struct vmap_area *va;
1383 struct vm_struct *area;
1384
1385 BUG_ON(in_interrupt());
1386 size = PAGE_ALIGN(size);
1387 if (unlikely(!size))
1388 return NULL;
1389
1390 if (flags & VM_IOREMAP)
1391 align = 1ul << clamp_t(int, get_count_order_long(size),
1392 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1393
1394 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1395 if (unlikely(!area))
1396 return NULL;
1397
1398 if (!(flags & VM_NO_GUARD))
1399 size += PAGE_SIZE;
1400
1401 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1402 if (IS_ERR(va)) {
1403 kfree(area);
1404 return NULL;
1405 }
1406
1407 setup_vmalloc_vm(area, va, flags, caller);
1408
1409 return area;
1410}
1411
1412struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1413 unsigned long start, unsigned long end)
1414{
1415 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1416 GFP_KERNEL, __builtin_return_address(0));
1417}
1418EXPORT_SYMBOL_GPL(__get_vm_area);
1419
1420struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1421 unsigned long start, unsigned long end,
1422 const void *caller)
1423{
1424 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1425 GFP_KERNEL, caller);
1426}
1427
1428/**
1429 * get_vm_area - reserve a contiguous kernel virtual area
1430 * @size: size of the area
1431 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1432 *
1433 * Search an area of @size in the kernel virtual mapping area,
1434 * and reserved it for out purposes. Returns the area descriptor
1435 * on success or %NULL on failure.
1436 *
1437 * Return: the area descriptor on success or %NULL on failure.
1438 */
1439struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1440{
1441 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1442 NUMA_NO_NODE, GFP_KERNEL,
1443 __builtin_return_address(0));
1444}
1445
1446struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1447 const void *caller)
1448{
1449 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1450 NUMA_NO_NODE, GFP_KERNEL, caller);
1451}
1452
1453/**
1454 * find_vm_area - find a continuous kernel virtual area
1455 * @addr: base address
1456 *
1457 * Search for the kernel VM area starting at @addr, and return it.
1458 * It is up to the caller to do all required locking to keep the returned
1459 * pointer valid.
1460 *
1461 * Return: pointer to the found area or %NULL on faulure
1462 */
1463struct vm_struct *find_vm_area(const void *addr)
1464{
1465 struct vmap_area *va;
1466
1467 va = find_vmap_area((unsigned long)addr);
1468 if (va && va->flags & VM_VM_AREA)
1469 return va->vm;
1470
1471 return NULL;
1472}
1473
1474/**
1475 * remove_vm_area - find and remove a continuous kernel virtual area
1476 * @addr: base address
1477 *
1478 * Search for the kernel VM area starting at @addr, and remove it.
1479 * This function returns the found VM area, but using it is NOT safe
1480 * on SMP machines, except for its size or flags.
1481 *
1482 * Return: pointer to the found area or %NULL on faulure
1483 */
1484struct vm_struct *remove_vm_area(const void *addr)
1485{
1486 struct vmap_area *va;
1487
1488 might_sleep();
1489
1490 va = find_vmap_area((unsigned long)addr);
1491 if (va && va->flags & VM_VM_AREA) {
1492 struct vm_struct *vm = va->vm;
1493
1494 spin_lock(&vmap_area_lock);
1495 va->vm = NULL;
1496 va->flags &= ~VM_VM_AREA;
1497 va->flags |= VM_LAZY_FREE;
1498 spin_unlock(&vmap_area_lock);
1499
1500 kasan_free_shadow(vm);
1501 free_unmap_vmap_area(va);
1502
1503 return vm;
1504 }
1505 return NULL;
1506}
1507
1508static void __vunmap(const void *addr, int deallocate_pages)
1509{
1510 struct vm_struct *area;
1511
1512 if (!addr)
1513 return;
1514
1515 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1516 addr))
1517 return;
1518
1519 area = find_vm_area(addr);
1520 if (unlikely(!area)) {
1521 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1522 addr);
1523 return;
1524 }
1525
1526 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
1527 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
1528
1529 remove_vm_area(addr);
1530 if (deallocate_pages) {
1531 int i;
1532
1533 for (i = 0; i < area->nr_pages; i++) {
1534 struct page *page = area->pages[i];
1535
1536 BUG_ON(!page);
1537 __free_pages(page, 0);
1538 }
1539
1540 kvfree(area->pages);
1541 }
1542
1543 kfree(area);
1544 return;
1545}
1546
1547static inline void __vfree_deferred(const void *addr)
1548{
1549 /*
1550 * Use raw_cpu_ptr() because this can be called from preemptible
1551 * context. Preemption is absolutely fine here, because the llist_add()
1552 * implementation is lockless, so it works even if we are adding to
1553 * nother cpu's list. schedule_work() should be fine with this too.
1554 */
1555 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1556
1557 if (llist_add((struct llist_node *)addr, &p->list))
1558 schedule_work(&p->wq);
1559}
1560
1561/**
1562 * vfree_atomic - release memory allocated by vmalloc()
1563 * @addr: memory base address
1564 *
1565 * This one is just like vfree() but can be called in any atomic context
1566 * except NMIs.
1567 */
1568void vfree_atomic(const void *addr)
1569{
1570 BUG_ON(in_nmi());
1571
1572 kmemleak_free(addr);
1573
1574 if (!addr)
1575 return;
1576 __vfree_deferred(addr);
1577}
1578
1579static void __vfree(const void *addr)
1580{
1581 if (unlikely(in_interrupt()))
1582 __vfree_deferred(addr);
1583 else
1584 __vunmap(addr, 1);
1585}
1586
1587/**
1588 * vfree - release memory allocated by vmalloc()
1589 * @addr: memory base address
1590 *
1591 * Free the virtually continuous memory area starting at @addr, as
1592 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1593 * NULL, no operation is performed.
1594 *
1595 * Must not be called in NMI context (strictly speaking, only if we don't
1596 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1597 * conventions for vfree() arch-depenedent would be a really bad idea)
1598 *
1599 * May sleep if called *not* from interrupt context.
1600 *
1601 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1602 */
1603void vfree(const void *addr)
1604{
1605 BUG_ON(in_nmi());
1606
1607 kmemleak_free(addr);
1608
1609 might_sleep_if(!in_interrupt());
1610
1611 if (!addr)
1612 return;
1613
1614 __vfree(addr);
1615}
1616EXPORT_SYMBOL(vfree);
1617
1618/**
1619 * vunmap - release virtual mapping obtained by vmap()
1620 * @addr: memory base address
1621 *
1622 * Free the virtually contiguous memory area starting at @addr,
1623 * which was created from the page array passed to vmap().
1624 *
1625 * Must not be called in interrupt context.
1626 */
1627void vunmap(const void *addr)
1628{
1629 BUG_ON(in_interrupt());
1630 might_sleep();
1631 if (addr)
1632 __vunmap(addr, 0);
1633}
1634EXPORT_SYMBOL(vunmap);
1635
1636/**
1637 * vmap - map an array of pages into virtually contiguous space
1638 * @pages: array of page pointers
1639 * @count: number of pages to map
1640 * @flags: vm_area->flags
1641 * @prot: page protection for the mapping
1642 *
1643 * Maps @count pages from @pages into contiguous kernel virtual
1644 * space.
1645 *
1646 * Return: the address of the area or %NULL on failure
1647 */
1648void *vmap(struct page **pages, unsigned int count,
1649 unsigned long flags, pgprot_t prot)
1650{
1651 struct vm_struct *area;
1652 unsigned long size; /* In bytes */
1653
1654 might_sleep();
1655
1656 if (count > totalram_pages())
1657 return NULL;
1658
1659 size = (unsigned long)count << PAGE_SHIFT;
1660 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1661 if (!area)
1662 return NULL;
1663
1664 if (map_vm_area(area, prot, pages)) {
1665 vunmap(area->addr);
1666 return NULL;
1667 }
1668
1669 return area->addr;
1670}
1671EXPORT_SYMBOL(vmap);
1672
1673static void *__vmalloc_node(unsigned long size, unsigned long align,
1674 gfp_t gfp_mask, pgprot_t prot,
1675 int node, const void *caller);
1676static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1677 pgprot_t prot, int node)
1678{
1679 struct page **pages;
1680 unsigned int nr_pages, array_size, i;
1681 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1682 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1683 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
1684 0 :
1685 __GFP_HIGHMEM;
1686
1687 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1688 array_size = (nr_pages * sizeof(struct page *));
1689
1690 area->nr_pages = nr_pages;
1691 /* Please note that the recursion is strictly bounded. */
1692 if (array_size > PAGE_SIZE) {
1693 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
1694 PAGE_KERNEL, node, area->caller);
1695 } else {
1696 pages = kmalloc_node(array_size, nested_gfp, node);
1697 }
1698 area->pages = pages;
1699 if (!area->pages) {
1700 remove_vm_area(area->addr);
1701 kfree(area);
1702 return NULL;
1703 }
1704
1705 for (i = 0; i < area->nr_pages; i++) {
1706 struct page *page;
1707
1708 if (node == NUMA_NO_NODE)
1709 page = alloc_page(alloc_mask|highmem_mask);
1710 else
1711 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
1712
1713 if (unlikely(!page)) {
1714 /* Successfully allocated i pages, free them in __vunmap() */
1715 area->nr_pages = i;
1716 goto fail;
1717 }
1718 area->pages[i] = page;
1719 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
1720 cond_resched();
1721 }
1722
1723 if (map_vm_area(area, prot, pages))
1724 goto fail;
1725 return area->addr;
1726
1727fail:
1728 warn_alloc(gfp_mask, NULL,
1729 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1730 (area->nr_pages*PAGE_SIZE), area->size);
1731 __vfree(area->addr);
1732 return NULL;
1733}
1734
1735/**
1736 * __vmalloc_node_range - allocate virtually contiguous memory
1737 * @size: allocation size
1738 * @align: desired alignment
1739 * @start: vm area range start
1740 * @end: vm area range end
1741 * @gfp_mask: flags for the page level allocator
1742 * @prot: protection mask for the allocated pages
1743 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1744 * @node: node to use for allocation or NUMA_NO_NODE
1745 * @caller: caller's return address
1746 *
1747 * Allocate enough pages to cover @size from the page level
1748 * allocator with @gfp_mask flags. Map them into contiguous
1749 * kernel virtual space, using a pagetable protection of @prot.
1750 *
1751 * Return: the address of the area or %NULL on failure
1752 */
1753void *__vmalloc_node_range(unsigned long size, unsigned long align,
1754 unsigned long start, unsigned long end, gfp_t gfp_mask,
1755 pgprot_t prot, unsigned long vm_flags, int node,
1756 const void *caller)
1757{
1758 struct vm_struct *area;
1759 void *addr;
1760 unsigned long real_size = size;
1761
1762 size = PAGE_ALIGN(size);
1763 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
1764 goto fail;
1765
1766 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1767 vm_flags, start, end, node, gfp_mask, caller);
1768 if (!area)
1769 goto fail;
1770
1771 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1772 if (!addr)
1773 return NULL;
1774
1775 /*
1776 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1777 * flag. It means that vm_struct is not fully initialized.
1778 * Now, it is fully initialized, so remove this flag here.
1779 */
1780 clear_vm_uninitialized_flag(area);
1781
1782 kmemleak_vmalloc(area, size, gfp_mask);
1783
1784 return addr;
1785
1786fail:
1787 warn_alloc(gfp_mask, NULL,
1788 "vmalloc: allocation failure: %lu bytes", real_size);
1789 return NULL;
1790}
1791
1792/*
1793 * This is only for performance analysis of vmalloc and stress purpose.
1794 * It is required by vmalloc test module, therefore do not use it other
1795 * than that.
1796 */
1797#ifdef CONFIG_TEST_VMALLOC_MODULE
1798EXPORT_SYMBOL_GPL(__vmalloc_node_range);
1799#endif
1800
1801/**
1802 * __vmalloc_node - allocate virtually contiguous memory
1803 * @size: allocation size
1804 * @align: desired alignment
1805 * @gfp_mask: flags for the page level allocator
1806 * @prot: protection mask for the allocated pages
1807 * @node: node to use for allocation or NUMA_NO_NODE
1808 * @caller: caller's return address
1809 *
1810 * Allocate enough pages to cover @size from the page level
1811 * allocator with @gfp_mask flags. Map them into contiguous
1812 * kernel virtual space, using a pagetable protection of @prot.
1813 *
1814 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1815 * and __GFP_NOFAIL are not supported
1816 *
1817 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1818 * with mm people.
1819 *
1820 * Return: pointer to the allocated memory or %NULL on error
1821 */
1822static void *__vmalloc_node(unsigned long size, unsigned long align,
1823 gfp_t gfp_mask, pgprot_t prot,
1824 int node, const void *caller)
1825{
1826 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1827 gfp_mask, prot, 0, node, caller);
1828}
1829
1830void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1831{
1832 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1833 __builtin_return_address(0));
1834}
1835EXPORT_SYMBOL(__vmalloc);
1836
1837static inline void *__vmalloc_node_flags(unsigned long size,
1838 int node, gfp_t flags)
1839{
1840 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1841 node, __builtin_return_address(0));
1842}
1843
1844
1845void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
1846 void *caller)
1847{
1848 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
1849}
1850
1851/**
1852 * vmalloc - allocate virtually contiguous memory
1853 * @size: allocation size
1854 *
1855 * Allocate enough pages to cover @size from the page level
1856 * allocator and map them into contiguous kernel virtual space.
1857 *
1858 * For tight control over page level allocator and protection flags
1859 * use __vmalloc() instead.
1860 *
1861 * Return: pointer to the allocated memory or %NULL on error
1862 */
1863void *vmalloc(unsigned long size)
1864{
1865 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1866 GFP_KERNEL);
1867}
1868EXPORT_SYMBOL(vmalloc);
1869
1870/**
1871 * vzalloc - allocate virtually contiguous memory with zero fill
1872 * @size: allocation size
1873 *
1874 * Allocate enough pages to cover @size from the page level
1875 * allocator and map them into contiguous kernel virtual space.
1876 * The memory allocated is set to zero.
1877 *
1878 * For tight control over page level allocator and protection flags
1879 * use __vmalloc() instead.
1880 *
1881 * Return: pointer to the allocated memory or %NULL on error
1882 */
1883void *vzalloc(unsigned long size)
1884{
1885 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1886 GFP_KERNEL | __GFP_ZERO);
1887}
1888EXPORT_SYMBOL(vzalloc);
1889
1890/**
1891 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1892 * @size: allocation size
1893 *
1894 * The resulting memory area is zeroed so it can be mapped to userspace
1895 * without leaking data.
1896 *
1897 * Return: pointer to the allocated memory or %NULL on error
1898 */
1899void *vmalloc_user(unsigned long size)
1900{
1901 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
1902 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
1903 VM_USERMAP, NUMA_NO_NODE,
1904 __builtin_return_address(0));
1905}
1906EXPORT_SYMBOL(vmalloc_user);
1907
1908/**
1909 * vmalloc_node - allocate memory on a specific node
1910 * @size: allocation size
1911 * @node: numa node
1912 *
1913 * Allocate enough pages to cover @size from the page level
1914 * allocator and map them into contiguous kernel virtual space.
1915 *
1916 * For tight control over page level allocator and protection flags
1917 * use __vmalloc() instead.
1918 *
1919 * Return: pointer to the allocated memory or %NULL on error
1920 */
1921void *vmalloc_node(unsigned long size, int node)
1922{
1923 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
1924 node, __builtin_return_address(0));
1925}
1926EXPORT_SYMBOL(vmalloc_node);
1927
1928/**
1929 * vzalloc_node - allocate memory on a specific node with zero fill
1930 * @size: allocation size
1931 * @node: numa node
1932 *
1933 * Allocate enough pages to cover @size from the page level
1934 * allocator and map them into contiguous kernel virtual space.
1935 * The memory allocated is set to zero.
1936 *
1937 * For tight control over page level allocator and protection flags
1938 * use __vmalloc_node() instead.
1939 *
1940 * Return: pointer to the allocated memory or %NULL on error
1941 */
1942void *vzalloc_node(unsigned long size, int node)
1943{
1944 return __vmalloc_node_flags(size, node,
1945 GFP_KERNEL | __GFP_ZERO);
1946}
1947EXPORT_SYMBOL(vzalloc_node);
1948
1949/**
1950 * vmalloc_exec - allocate virtually contiguous, executable memory
1951 * @size: allocation size
1952 *
1953 * Kernel-internal function to allocate enough pages to cover @size
1954 * the page level allocator and map them into contiguous and
1955 * executable kernel virtual space.
1956 *
1957 * For tight control over page level allocator and protection flags
1958 * use __vmalloc() instead.
1959 *
1960 * Return: pointer to the allocated memory or %NULL on error
1961 */
1962void *vmalloc_exec(unsigned long size)
1963{
1964 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL_EXEC,
1965 NUMA_NO_NODE, __builtin_return_address(0));
1966}
1967
1968#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1969#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
1970#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1971#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
1972#else
1973/*
1974 * 64b systems should always have either DMA or DMA32 zones. For others
1975 * GFP_DMA32 should do the right thing and use the normal zone.
1976 */
1977#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1978#endif
1979
1980/**
1981 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1982 * @size: allocation size
1983 *
1984 * Allocate enough 32bit PA addressable pages to cover @size from the
1985 * page level allocator and map them into contiguous kernel virtual space.
1986 *
1987 * Return: pointer to the allocated memory or %NULL on error
1988 */
1989void *vmalloc_32(unsigned long size)
1990{
1991 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1992 NUMA_NO_NODE, __builtin_return_address(0));
1993}
1994EXPORT_SYMBOL(vmalloc_32);
1995
1996/**
1997 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1998 * @size: allocation size
1999 *
2000 * The resulting memory area is 32bit addressable and zeroed so it can be
2001 * mapped to userspace without leaking data.
2002 *
2003 * Return: pointer to the allocated memory or %NULL on error
2004 */
2005void *vmalloc_32_user(unsigned long size)
2006{
2007 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2008 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2009 VM_USERMAP, NUMA_NO_NODE,
2010 __builtin_return_address(0));
2011}
2012EXPORT_SYMBOL(vmalloc_32_user);
2013
2014/*
2015 * small helper routine , copy contents to buf from addr.
2016 * If the page is not present, fill zero.
2017 */
2018
2019static int aligned_vread(char *buf, char *addr, unsigned long count)
2020{
2021 struct page *p;
2022 int copied = 0;
2023
2024 while (count) {
2025 unsigned long offset, length;
2026
2027 offset = offset_in_page(addr);
2028 length = PAGE_SIZE - offset;
2029 if (length > count)
2030 length = count;
2031 p = vmalloc_to_page(addr);
2032 /*
2033 * To do safe access to this _mapped_ area, we need
2034 * lock. But adding lock here means that we need to add
2035 * overhead of vmalloc()/vfree() calles for this _debug_
2036 * interface, rarely used. Instead of that, we'll use
2037 * kmap() and get small overhead in this access function.
2038 */
2039 if (p) {
2040 /*
2041 * we can expect USER0 is not used (see vread/vwrite's
2042 * function description)
2043 */
2044 void *map = kmap_atomic(p);
2045 memcpy(buf, map + offset, length);
2046 kunmap_atomic(map);
2047 } else
2048 memset(buf, 0, length);
2049
2050 addr += length;
2051 buf += length;
2052 copied += length;
2053 count -= length;
2054 }
2055 return copied;
2056}
2057
2058static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2059{
2060 struct page *p;
2061 int copied = 0;
2062
2063 while (count) {
2064 unsigned long offset, length;
2065
2066 offset = offset_in_page(addr);
2067 length = PAGE_SIZE - offset;
2068 if (length > count)
2069 length = count;
2070 p = vmalloc_to_page(addr);
2071 /*
2072 * To do safe access to this _mapped_ area, we need
2073 * lock. But adding lock here means that we need to add
2074 * overhead of vmalloc()/vfree() calles for this _debug_
2075 * interface, rarely used. Instead of that, we'll use
2076 * kmap() and get small overhead in this access function.
2077 */
2078 if (p) {
2079 /*
2080 * we can expect USER0 is not used (see vread/vwrite's
2081 * function description)
2082 */
2083 void *map = kmap_atomic(p);
2084 memcpy(map + offset, buf, length);
2085 kunmap_atomic(map);
2086 }
2087 addr += length;
2088 buf += length;
2089 copied += length;
2090 count -= length;
2091 }
2092 return copied;
2093}
2094
2095/**
2096 * vread() - read vmalloc area in a safe way.
2097 * @buf: buffer for reading data
2098 * @addr: vm address.
2099 * @count: number of bytes to be read.
2100 *
2101 * This function checks that addr is a valid vmalloc'ed area, and
2102 * copy data from that area to a given buffer. If the given memory range
2103 * of [addr...addr+count) includes some valid address, data is copied to
2104 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2105 * IOREMAP area is treated as memory hole and no copy is done.
2106 *
2107 * If [addr...addr+count) doesn't includes any intersects with alive
2108 * vm_struct area, returns 0. @buf should be kernel's buffer.
2109 *
2110 * Note: In usual ops, vread() is never necessary because the caller
2111 * should know vmalloc() area is valid and can use memcpy().
2112 * This is for routines which have to access vmalloc area without
2113 * any informaion, as /dev/kmem.
2114 *
2115 * Return: number of bytes for which addr and buf should be increased
2116 * (same number as @count) or %0 if [addr...addr+count) doesn't
2117 * include any intersection with valid vmalloc area
2118 */
2119long vread(char *buf, char *addr, unsigned long count)
2120{
2121 struct vmap_area *va;
2122 struct vm_struct *vm;
2123 char *vaddr, *buf_start = buf;
2124 unsigned long buflen = count;
2125 unsigned long n;
2126
2127 /* Don't allow overflow */
2128 if ((unsigned long) addr + count < count)
2129 count = -(unsigned long) addr;
2130
2131 spin_lock(&vmap_area_lock);
2132 list_for_each_entry(va, &vmap_area_list, list) {
2133 if (!count)
2134 break;
2135
2136 if (!(va->flags & VM_VM_AREA))
2137 continue;
2138
2139 vm = va->vm;
2140 vaddr = (char *) vm->addr;
2141 if (addr >= vaddr + get_vm_area_size(vm))
2142 continue;
2143 while (addr < vaddr) {
2144 if (count == 0)
2145 goto finished;
2146 *buf = '\0';
2147 buf++;
2148 addr++;
2149 count--;
2150 }
2151 n = vaddr + get_vm_area_size(vm) - addr;
2152 if (n > count)
2153 n = count;
2154 if (!(vm->flags & VM_IOREMAP))
2155 aligned_vread(buf, addr, n);
2156 else /* IOREMAP area is treated as memory hole */
2157 memset(buf, 0, n);
2158 buf += n;
2159 addr += n;
2160 count -= n;
2161 }
2162finished:
2163 spin_unlock(&vmap_area_lock);
2164
2165 if (buf == buf_start)
2166 return 0;
2167 /* zero-fill memory holes */
2168 if (buf != buf_start + buflen)
2169 memset(buf, 0, buflen - (buf - buf_start));
2170
2171 return buflen;
2172}
2173
2174/**
2175 * vwrite() - write vmalloc area in a safe way.
2176 * @buf: buffer for source data
2177 * @addr: vm address.
2178 * @count: number of bytes to be read.
2179 *
2180 * This function checks that addr is a valid vmalloc'ed area, and
2181 * copy data from a buffer to the given addr. If specified range of
2182 * [addr...addr+count) includes some valid address, data is copied from
2183 * proper area of @buf. If there are memory holes, no copy to hole.
2184 * IOREMAP area is treated as memory hole and no copy is done.
2185 *
2186 * If [addr...addr+count) doesn't includes any intersects with alive
2187 * vm_struct area, returns 0. @buf should be kernel's buffer.
2188 *
2189 * Note: In usual ops, vwrite() is never necessary because the caller
2190 * should know vmalloc() area is valid and can use memcpy().
2191 * This is for routines which have to access vmalloc area without
2192 * any informaion, as /dev/kmem.
2193 *
2194 * Return: number of bytes for which addr and buf should be
2195 * increased (same number as @count) or %0 if [addr...addr+count)
2196 * doesn't include any intersection with valid vmalloc area
2197 */
2198long vwrite(char *buf, char *addr, unsigned long count)
2199{
2200 struct vmap_area *va;
2201 struct vm_struct *vm;
2202 char *vaddr;
2203 unsigned long n, buflen;
2204 int copied = 0;
2205
2206 /* Don't allow overflow */
2207 if ((unsigned long) addr + count < count)
2208 count = -(unsigned long) addr;
2209 buflen = count;
2210
2211 spin_lock(&vmap_area_lock);
2212 list_for_each_entry(va, &vmap_area_list, list) {
2213 if (!count)
2214 break;
2215
2216 if (!(va->flags & VM_VM_AREA))
2217 continue;
2218
2219 vm = va->vm;
2220 vaddr = (char *) vm->addr;
2221 if (addr >= vaddr + get_vm_area_size(vm))
2222 continue;
2223 while (addr < vaddr) {
2224 if (count == 0)
2225 goto finished;
2226 buf++;
2227 addr++;
2228 count--;
2229 }
2230 n = vaddr + get_vm_area_size(vm) - addr;
2231 if (n > count)
2232 n = count;
2233 if (!(vm->flags & VM_IOREMAP)) {
2234 aligned_vwrite(buf, addr, n);
2235 copied++;
2236 }
2237 buf += n;
2238 addr += n;
2239 count -= n;
2240 }
2241finished:
2242 spin_unlock(&vmap_area_lock);
2243 if (!copied)
2244 return 0;
2245 return buflen;
2246}
2247
2248/**
2249 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2250 * @vma: vma to cover
2251 * @uaddr: target user address to start at
2252 * @kaddr: virtual address of vmalloc kernel memory
2253 * @size: size of map area
2254 *
2255 * Returns: 0 for success, -Exxx on failure
2256 *
2257 * This function checks that @kaddr is a valid vmalloc'ed area,
2258 * and that it is big enough to cover the range starting at
2259 * @uaddr in @vma. Will return failure if that criteria isn't
2260 * met.
2261 *
2262 * Similar to remap_pfn_range() (see mm/memory.c)
2263 */
2264int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2265 void *kaddr, unsigned long size)
2266{
2267 struct vm_struct *area;
2268
2269 size = PAGE_ALIGN(size);
2270
2271 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2272 return -EINVAL;
2273
2274 area = find_vm_area(kaddr);
2275 if (!area)
2276 return -EINVAL;
2277
2278 if (!(area->flags & VM_USERMAP))
2279 return -EINVAL;
2280
2281 if (kaddr + size > area->addr + get_vm_area_size(area))
2282 return -EINVAL;
2283
2284 do {
2285 struct page *page = vmalloc_to_page(kaddr);
2286 int ret;
2287
2288 ret = vm_insert_page(vma, uaddr, page);
2289 if (ret)
2290 return ret;
2291
2292 uaddr += PAGE_SIZE;
2293 kaddr += PAGE_SIZE;
2294 size -= PAGE_SIZE;
2295 } while (size > 0);
2296
2297 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2298
2299 return 0;
2300}
2301EXPORT_SYMBOL(remap_vmalloc_range_partial);
2302
2303/**
2304 * remap_vmalloc_range - map vmalloc pages to userspace
2305 * @vma: vma to cover (map full range of vma)
2306 * @addr: vmalloc memory
2307 * @pgoff: number of pages into addr before first page to map
2308 *
2309 * Returns: 0 for success, -Exxx on failure
2310 *
2311 * This function checks that addr is a valid vmalloc'ed area, and
2312 * that it is big enough to cover the vma. Will return failure if
2313 * that criteria isn't met.
2314 *
2315 * Similar to remap_pfn_range() (see mm/memory.c)
2316 */
2317int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2318 unsigned long pgoff)
2319{
2320 return remap_vmalloc_range_partial(vma, vma->vm_start,
2321 addr + (pgoff << PAGE_SHIFT),
2322 vma->vm_end - vma->vm_start);
2323}
2324EXPORT_SYMBOL(remap_vmalloc_range);
2325
2326/*
2327 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2328 * have one.
2329 */
2330void __weak vmalloc_sync_all(void)
2331{
2332}
2333
2334
2335static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2336{
2337 pte_t ***p = data;
2338
2339 if (p) {
2340 *(*p) = pte;
2341 (*p)++;
2342 }
2343 return 0;
2344}
2345
2346/**
2347 * alloc_vm_area - allocate a range of kernel address space
2348 * @size: size of the area
2349 * @ptes: returns the PTEs for the address space
2350 *
2351 * Returns: NULL on failure, vm_struct on success
2352 *
2353 * This function reserves a range of kernel address space, and
2354 * allocates pagetables to map that range. No actual mappings
2355 * are created.
2356 *
2357 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2358 * allocated for the VM area are returned.
2359 */
2360struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2361{
2362 struct vm_struct *area;
2363
2364 area = get_vm_area_caller(size, VM_IOREMAP,
2365 __builtin_return_address(0));
2366 if (area == NULL)
2367 return NULL;
2368
2369 /*
2370 * This ensures that page tables are constructed for this region
2371 * of kernel virtual address space and mapped into init_mm.
2372 */
2373 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2374 size, f, ptes ? &ptes : NULL)) {
2375 free_vm_area(area);
2376 return NULL;
2377 }
2378
2379 return area;
2380}
2381EXPORT_SYMBOL_GPL(alloc_vm_area);
2382
2383void free_vm_area(struct vm_struct *area)
2384{
2385 struct vm_struct *ret;
2386 ret = remove_vm_area(area->addr);
2387 BUG_ON(ret != area);
2388 kfree(area);
2389}
2390EXPORT_SYMBOL_GPL(free_vm_area);
2391
2392#ifdef CONFIG_SMP
2393static struct vmap_area *node_to_va(struct rb_node *n)
2394{
2395 return rb_entry_safe(n, struct vmap_area, rb_node);
2396}
2397
2398/**
2399 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2400 * @end: target address
2401 * @pnext: out arg for the next vmap_area
2402 * @pprev: out arg for the previous vmap_area
2403 *
2404 * Returns: %true if either or both of next and prev are found,
2405 * %false if no vmap_area exists
2406 *
2407 * Find vmap_areas end addresses of which enclose @end. ie. if not
2408 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2409 */
2410static bool pvm_find_next_prev(unsigned long end,
2411 struct vmap_area **pnext,
2412 struct vmap_area **pprev)
2413{
2414 struct rb_node *n = vmap_area_root.rb_node;
2415 struct vmap_area *va = NULL;
2416
2417 while (n) {
2418 va = rb_entry(n, struct vmap_area, rb_node);
2419 if (end < va->va_end)
2420 n = n->rb_left;
2421 else if (end > va->va_end)
2422 n = n->rb_right;
2423 else
2424 break;
2425 }
2426
2427 if (!va)
2428 return false;
2429
2430 if (va->va_end > end) {
2431 *pnext = va;
2432 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2433 } else {
2434 *pprev = va;
2435 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2436 }
2437 return true;
2438}
2439
2440/**
2441 * pvm_determine_end - find the highest aligned address between two vmap_areas
2442 * @pnext: in/out arg for the next vmap_area
2443 * @pprev: in/out arg for the previous vmap_area
2444 * @align: alignment
2445 *
2446 * Returns: determined end address
2447 *
2448 * Find the highest aligned address between *@pnext and *@pprev below
2449 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2450 * down address is between the end addresses of the two vmap_areas.
2451 *
2452 * Please note that the address returned by this function may fall
2453 * inside *@pnext vmap_area. The caller is responsible for checking
2454 * that.
2455 */
2456static unsigned long pvm_determine_end(struct vmap_area **pnext,
2457 struct vmap_area **pprev,
2458 unsigned long align)
2459{
2460 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2461 unsigned long addr;
2462
2463 if (*pnext)
2464 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2465 else
2466 addr = vmalloc_end;
2467
2468 while (*pprev && (*pprev)->va_end > addr) {
2469 *pnext = *pprev;
2470 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2471 }
2472
2473 return addr;
2474}
2475
2476/**
2477 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2478 * @offsets: array containing offset of each area
2479 * @sizes: array containing size of each area
2480 * @nr_vms: the number of areas to allocate
2481 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2482 *
2483 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2484 * vm_structs on success, %NULL on failure
2485 *
2486 * Percpu allocator wants to use congruent vm areas so that it can
2487 * maintain the offsets among percpu areas. This function allocates
2488 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2489 * be scattered pretty far, distance between two areas easily going up
2490 * to gigabytes. To avoid interacting with regular vmallocs, these
2491 * areas are allocated from top.
2492 *
2493 * Despite its complicated look, this allocator is rather simple. It
2494 * does everything top-down and scans areas from the end looking for
2495 * matching slot. While scanning, if any of the areas overlaps with
2496 * existing vmap_area, the base address is pulled down to fit the
2497 * area. Scanning is repeated till all the areas fit and then all
2498 * necessary data structures are inserted and the result is returned.
2499 */
2500struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2501 const size_t *sizes, int nr_vms,
2502 size_t align)
2503{
2504 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2505 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2506 struct vmap_area **vas, *prev, *next;
2507 struct vm_struct **vms;
2508 int area, area2, last_area, term_area;
2509 unsigned long base, start, end, last_end;
2510 bool purged = false;
2511
2512 /* verify parameters and allocate data structures */
2513 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2514 for (last_area = 0, area = 0; area < nr_vms; area++) {
2515 start = offsets[area];
2516 end = start + sizes[area];
2517
2518 /* is everything aligned properly? */
2519 BUG_ON(!IS_ALIGNED(offsets[area], align));
2520 BUG_ON(!IS_ALIGNED(sizes[area], align));
2521
2522 /* detect the area with the highest address */
2523 if (start > offsets[last_area])
2524 last_area = area;
2525
2526 for (area2 = area + 1; area2 < nr_vms; area2++) {
2527 unsigned long start2 = offsets[area2];
2528 unsigned long end2 = start2 + sizes[area2];
2529
2530 BUG_ON(start2 < end && start < end2);
2531 }
2532 }
2533 last_end = offsets[last_area] + sizes[last_area];
2534
2535 if (vmalloc_end - vmalloc_start < last_end) {
2536 WARN_ON(true);
2537 return NULL;
2538 }
2539
2540 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2541 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2542 if (!vas || !vms)
2543 goto err_free2;
2544
2545 for (area = 0; area < nr_vms; area++) {
2546 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);