Loading...
1/*
2 * linux/mm/vmalloc.c
3 *
4 * Copyright (C) 1993 Linus Torvalds
5 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
6 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
7 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
8 * Numa awareness, Christoph Lameter, SGI, June 2005
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched/signal.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/debugobjects.h>
22#include <linux/kallsyms.h>
23#include <linux/list.h>
24#include <linux/notifier.h>
25#include <linux/rbtree.h>
26#include <linux/radix-tree.h>
27#include <linux/rcupdate.h>
28#include <linux/pfn.h>
29#include <linux/kmemleak.h>
30#include <linux/atomic.h>
31#include <linux/compiler.h>
32#include <linux/llist.h>
33#include <linux/bitops.h>
34
35#include <linux/uaccess.h>
36#include <asm/tlbflush.h>
37#include <asm/shmparam.h>
38
39#include "internal.h"
40
41struct vfree_deferred {
42 struct llist_head list;
43 struct work_struct wq;
44};
45static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
46
47static void __vunmap(const void *, int);
48
49static void free_work(struct work_struct *w)
50{
51 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
52 struct llist_node *t, *llnode;
53
54 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
55 __vunmap((void *)llnode, 1);
56}
57
58/*** Page table manipulation functions ***/
59
60static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
61{
62 pte_t *pte;
63
64 pte = pte_offset_kernel(pmd, addr);
65 do {
66 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
67 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
68 } while (pte++, addr += PAGE_SIZE, addr != end);
69}
70
71static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
72{
73 pmd_t *pmd;
74 unsigned long next;
75
76 pmd = pmd_offset(pud, addr);
77 do {
78 next = pmd_addr_end(addr, end);
79 if (pmd_clear_huge(pmd))
80 continue;
81 if (pmd_none_or_clear_bad(pmd))
82 continue;
83 vunmap_pte_range(pmd, addr, next);
84 } while (pmd++, addr = next, addr != end);
85}
86
87static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
88{
89 pud_t *pud;
90 unsigned long next;
91
92 pud = pud_offset(p4d, addr);
93 do {
94 next = pud_addr_end(addr, end);
95 if (pud_clear_huge(pud))
96 continue;
97 if (pud_none_or_clear_bad(pud))
98 continue;
99 vunmap_pmd_range(pud, addr, next);
100 } while (pud++, addr = next, addr != end);
101}
102
103static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
104{
105 p4d_t *p4d;
106 unsigned long next;
107
108 p4d = p4d_offset(pgd, addr);
109 do {
110 next = p4d_addr_end(addr, end);
111 if (p4d_clear_huge(p4d))
112 continue;
113 if (p4d_none_or_clear_bad(p4d))
114 continue;
115 vunmap_pud_range(p4d, addr, next);
116 } while (p4d++, addr = next, addr != end);
117}
118
119static void vunmap_page_range(unsigned long addr, unsigned long end)
120{
121 pgd_t *pgd;
122 unsigned long next;
123
124 BUG_ON(addr >= end);
125 pgd = pgd_offset_k(addr);
126 do {
127 next = pgd_addr_end(addr, end);
128 if (pgd_none_or_clear_bad(pgd))
129 continue;
130 vunmap_p4d_range(pgd, addr, next);
131 } while (pgd++, addr = next, addr != end);
132}
133
134static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
135 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
136{
137 pte_t *pte;
138
139 /*
140 * nr is a running index into the array which helps higher level
141 * callers keep track of where we're up to.
142 */
143
144 pte = pte_alloc_kernel(pmd, addr);
145 if (!pte)
146 return -ENOMEM;
147 do {
148 struct page *page = pages[*nr];
149
150 if (WARN_ON(!pte_none(*pte)))
151 return -EBUSY;
152 if (WARN_ON(!page))
153 return -ENOMEM;
154 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
155 (*nr)++;
156 } while (pte++, addr += PAGE_SIZE, addr != end);
157 return 0;
158}
159
160static int vmap_pmd_range(pud_t *pud, unsigned long addr,
161 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
162{
163 pmd_t *pmd;
164 unsigned long next;
165
166 pmd = pmd_alloc(&init_mm, pud, addr);
167 if (!pmd)
168 return -ENOMEM;
169 do {
170 next = pmd_addr_end(addr, end);
171 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
172 return -ENOMEM;
173 } while (pmd++, addr = next, addr != end);
174 return 0;
175}
176
177static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
178 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
179{
180 pud_t *pud;
181 unsigned long next;
182
183 pud = pud_alloc(&init_mm, p4d, addr);
184 if (!pud)
185 return -ENOMEM;
186 do {
187 next = pud_addr_end(addr, end);
188 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
189 return -ENOMEM;
190 } while (pud++, addr = next, addr != end);
191 return 0;
192}
193
194static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
195 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
196{
197 p4d_t *p4d;
198 unsigned long next;
199
200 p4d = p4d_alloc(&init_mm, pgd, addr);
201 if (!p4d)
202 return -ENOMEM;
203 do {
204 next = p4d_addr_end(addr, end);
205 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
206 return -ENOMEM;
207 } while (p4d++, addr = next, addr != end);
208 return 0;
209}
210
211/*
212 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
213 * will have pfns corresponding to the "pages" array.
214 *
215 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
216 */
217static int vmap_page_range_noflush(unsigned long start, unsigned long end,
218 pgprot_t prot, struct page **pages)
219{
220 pgd_t *pgd;
221 unsigned long next;
222 unsigned long addr = start;
223 int err = 0;
224 int nr = 0;
225
226 BUG_ON(addr >= end);
227 pgd = pgd_offset_k(addr);
228 do {
229 next = pgd_addr_end(addr, end);
230 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
231 if (err)
232 return err;
233 } while (pgd++, addr = next, addr != end);
234
235 return nr;
236}
237
238static int vmap_page_range(unsigned long start, unsigned long end,
239 pgprot_t prot, struct page **pages)
240{
241 int ret;
242
243 ret = vmap_page_range_noflush(start, end, prot, pages);
244 flush_cache_vmap(start, end);
245 return ret;
246}
247
248int is_vmalloc_or_module_addr(const void *x)
249{
250 /*
251 * ARM, x86-64 and sparc64 put modules in a special place,
252 * and fall back on vmalloc() if that fails. Others
253 * just put it in the vmalloc space.
254 */
255#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
256 unsigned long addr = (unsigned long)x;
257 if (addr >= MODULES_VADDR && addr < MODULES_END)
258 return 1;
259#endif
260 return is_vmalloc_addr(x);
261}
262
263/*
264 * Walk a vmap address to the struct page it maps.
265 */
266struct page *vmalloc_to_page(const void *vmalloc_addr)
267{
268 unsigned long addr = (unsigned long) vmalloc_addr;
269 struct page *page = NULL;
270 pgd_t *pgd = pgd_offset_k(addr);
271 p4d_t *p4d;
272 pud_t *pud;
273 pmd_t *pmd;
274 pte_t *ptep, pte;
275
276 /*
277 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
278 * architectures that do not vmalloc module space
279 */
280 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
281
282 if (pgd_none(*pgd))
283 return NULL;
284 p4d = p4d_offset(pgd, addr);
285 if (p4d_none(*p4d))
286 return NULL;
287 pud = pud_offset(p4d, addr);
288
289 /*
290 * Don't dereference bad PUD or PMD (below) entries. This will also
291 * identify huge mappings, which we may encounter on architectures
292 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
293 * identified as vmalloc addresses by is_vmalloc_addr(), but are
294 * not [unambiguously] associated with a struct page, so there is
295 * no correct value to return for them.
296 */
297 WARN_ON_ONCE(pud_bad(*pud));
298 if (pud_none(*pud) || pud_bad(*pud))
299 return NULL;
300 pmd = pmd_offset(pud, addr);
301 WARN_ON_ONCE(pmd_bad(*pmd));
302 if (pmd_none(*pmd) || pmd_bad(*pmd))
303 return NULL;
304
305 ptep = pte_offset_map(pmd, addr);
306 pte = *ptep;
307 if (pte_present(pte))
308 page = pte_page(pte);
309 pte_unmap(ptep);
310 return page;
311}
312EXPORT_SYMBOL(vmalloc_to_page);
313
314/*
315 * Map a vmalloc()-space virtual address to the physical page frame number.
316 */
317unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
318{
319 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
320}
321EXPORT_SYMBOL(vmalloc_to_pfn);
322
323
324/*** Global kva allocator ***/
325
326#define VM_LAZY_FREE 0x02
327#define VM_VM_AREA 0x04
328
329static DEFINE_SPINLOCK(vmap_area_lock);
330/* Export for kexec only */
331LIST_HEAD(vmap_area_list);
332static LLIST_HEAD(vmap_purge_list);
333static struct rb_root vmap_area_root = RB_ROOT;
334
335/* The vmap cache globals are protected by vmap_area_lock */
336static struct rb_node *free_vmap_cache;
337static unsigned long cached_hole_size;
338static unsigned long cached_vstart;
339static unsigned long cached_align;
340
341static unsigned long vmap_area_pcpu_hole;
342
343static struct vmap_area *__find_vmap_area(unsigned long addr)
344{
345 struct rb_node *n = vmap_area_root.rb_node;
346
347 while (n) {
348 struct vmap_area *va;
349
350 va = rb_entry(n, struct vmap_area, rb_node);
351 if (addr < va->va_start)
352 n = n->rb_left;
353 else if (addr >= va->va_end)
354 n = n->rb_right;
355 else
356 return va;
357 }
358
359 return NULL;
360}
361
362static void __insert_vmap_area(struct vmap_area *va)
363{
364 struct rb_node **p = &vmap_area_root.rb_node;
365 struct rb_node *parent = NULL;
366 struct rb_node *tmp;
367
368 while (*p) {
369 struct vmap_area *tmp_va;
370
371 parent = *p;
372 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
373 if (va->va_start < tmp_va->va_end)
374 p = &(*p)->rb_left;
375 else if (va->va_end > tmp_va->va_start)
376 p = &(*p)->rb_right;
377 else
378 BUG();
379 }
380
381 rb_link_node(&va->rb_node, parent, p);
382 rb_insert_color(&va->rb_node, &vmap_area_root);
383
384 /* address-sort this list */
385 tmp = rb_prev(&va->rb_node);
386 if (tmp) {
387 struct vmap_area *prev;
388 prev = rb_entry(tmp, struct vmap_area, rb_node);
389 list_add_rcu(&va->list, &prev->list);
390 } else
391 list_add_rcu(&va->list, &vmap_area_list);
392}
393
394static void purge_vmap_area_lazy(void);
395
396static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
397
398/*
399 * Allocate a region of KVA of the specified size and alignment, within the
400 * vstart and vend.
401 */
402static struct vmap_area *alloc_vmap_area(unsigned long size,
403 unsigned long align,
404 unsigned long vstart, unsigned long vend,
405 int node, gfp_t gfp_mask)
406{
407 struct vmap_area *va;
408 struct rb_node *n;
409 unsigned long addr;
410 int purged = 0;
411 struct vmap_area *first;
412
413 BUG_ON(!size);
414 BUG_ON(offset_in_page(size));
415 BUG_ON(!is_power_of_2(align));
416
417 might_sleep();
418
419 va = kmalloc_node(sizeof(struct vmap_area),
420 gfp_mask & GFP_RECLAIM_MASK, node);
421 if (unlikely(!va))
422 return ERR_PTR(-ENOMEM);
423
424 /*
425 * Only scan the relevant parts containing pointers to other objects
426 * to avoid false negatives.
427 */
428 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
429
430retry:
431 spin_lock(&vmap_area_lock);
432 /*
433 * Invalidate cache if we have more permissive parameters.
434 * cached_hole_size notes the largest hole noticed _below_
435 * the vmap_area cached in free_vmap_cache: if size fits
436 * into that hole, we want to scan from vstart to reuse
437 * the hole instead of allocating above free_vmap_cache.
438 * Note that __free_vmap_area may update free_vmap_cache
439 * without updating cached_hole_size or cached_align.
440 */
441 if (!free_vmap_cache ||
442 size < cached_hole_size ||
443 vstart < cached_vstart ||
444 align < cached_align) {
445nocache:
446 cached_hole_size = 0;
447 free_vmap_cache = NULL;
448 }
449 /* record if we encounter less permissive parameters */
450 cached_vstart = vstart;
451 cached_align = align;
452
453 /* find starting point for our search */
454 if (free_vmap_cache) {
455 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
456 addr = ALIGN(first->va_end, align);
457 if (addr < vstart)
458 goto nocache;
459 if (addr + size < addr)
460 goto overflow;
461
462 } else {
463 addr = ALIGN(vstart, align);
464 if (addr + size < addr)
465 goto overflow;
466
467 n = vmap_area_root.rb_node;
468 first = NULL;
469
470 while (n) {
471 struct vmap_area *tmp;
472 tmp = rb_entry(n, struct vmap_area, rb_node);
473 if (tmp->va_end >= addr) {
474 first = tmp;
475 if (tmp->va_start <= addr)
476 break;
477 n = n->rb_left;
478 } else
479 n = n->rb_right;
480 }
481
482 if (!first)
483 goto found;
484 }
485
486 /* from the starting point, walk areas until a suitable hole is found */
487 while (addr + size > first->va_start && addr + size <= vend) {
488 if (addr + cached_hole_size < first->va_start)
489 cached_hole_size = first->va_start - addr;
490 addr = ALIGN(first->va_end, align);
491 if (addr + size < addr)
492 goto overflow;
493
494 if (list_is_last(&first->list, &vmap_area_list))
495 goto found;
496
497 first = list_next_entry(first, list);
498 }
499
500found:
501 if (addr + size > vend)
502 goto overflow;
503
504 va->va_start = addr;
505 va->va_end = addr + size;
506 va->flags = 0;
507 __insert_vmap_area(va);
508 free_vmap_cache = &va->rb_node;
509 spin_unlock(&vmap_area_lock);
510
511 BUG_ON(!IS_ALIGNED(va->va_start, align));
512 BUG_ON(va->va_start < vstart);
513 BUG_ON(va->va_end > vend);
514
515 return va;
516
517overflow:
518 spin_unlock(&vmap_area_lock);
519 if (!purged) {
520 purge_vmap_area_lazy();
521 purged = 1;
522 goto retry;
523 }
524
525 if (gfpflags_allow_blocking(gfp_mask)) {
526 unsigned long freed = 0;
527 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
528 if (freed > 0) {
529 purged = 0;
530 goto retry;
531 }
532 }
533
534 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
535 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
536 size);
537 kfree(va);
538 return ERR_PTR(-EBUSY);
539}
540
541int register_vmap_purge_notifier(struct notifier_block *nb)
542{
543 return blocking_notifier_chain_register(&vmap_notify_list, nb);
544}
545EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
546
547int unregister_vmap_purge_notifier(struct notifier_block *nb)
548{
549 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
550}
551EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
552
553static void __free_vmap_area(struct vmap_area *va)
554{
555 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
556
557 if (free_vmap_cache) {
558 if (va->va_end < cached_vstart) {
559 free_vmap_cache = NULL;
560 } else {
561 struct vmap_area *cache;
562 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
563 if (va->va_start <= cache->va_start) {
564 free_vmap_cache = rb_prev(&va->rb_node);
565 /*
566 * We don't try to update cached_hole_size or
567 * cached_align, but it won't go very wrong.
568 */
569 }
570 }
571 }
572 rb_erase(&va->rb_node, &vmap_area_root);
573 RB_CLEAR_NODE(&va->rb_node);
574 list_del_rcu(&va->list);
575
576 /*
577 * Track the highest possible candidate for pcpu area
578 * allocation. Areas outside of vmalloc area can be returned
579 * here too, consider only end addresses which fall inside
580 * vmalloc area proper.
581 */
582 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
583 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
584
585 kfree_rcu(va, rcu_head);
586}
587
588/*
589 * Free a region of KVA allocated by alloc_vmap_area
590 */
591static void free_vmap_area(struct vmap_area *va)
592{
593 spin_lock(&vmap_area_lock);
594 __free_vmap_area(va);
595 spin_unlock(&vmap_area_lock);
596}
597
598/*
599 * Clear the pagetable entries of a given vmap_area
600 */
601static void unmap_vmap_area(struct vmap_area *va)
602{
603 vunmap_page_range(va->va_start, va->va_end);
604}
605
606static void vmap_debug_free_range(unsigned long start, unsigned long end)
607{
608 /*
609 * Unmap page tables and force a TLB flush immediately if pagealloc
610 * debugging is enabled. This catches use after free bugs similarly to
611 * those in linear kernel virtual address space after a page has been
612 * freed.
613 *
614 * All the lazy freeing logic is still retained, in order to minimise
615 * intrusiveness of this debugging feature.
616 *
617 * This is going to be *slow* (linear kernel virtual address debugging
618 * doesn't do a broadcast TLB flush so it is a lot faster).
619 */
620 if (debug_pagealloc_enabled()) {
621 vunmap_page_range(start, end);
622 flush_tlb_kernel_range(start, end);
623 }
624}
625
626/*
627 * lazy_max_pages is the maximum amount of virtual address space we gather up
628 * before attempting to purge with a TLB flush.
629 *
630 * There is a tradeoff here: a larger number will cover more kernel page tables
631 * and take slightly longer to purge, but it will linearly reduce the number of
632 * global TLB flushes that must be performed. It would seem natural to scale
633 * this number up linearly with the number of CPUs (because vmapping activity
634 * could also scale linearly with the number of CPUs), however it is likely
635 * that in practice, workloads might be constrained in other ways that mean
636 * vmap activity will not scale linearly with CPUs. Also, I want to be
637 * conservative and not introduce a big latency on huge systems, so go with
638 * a less aggressive log scale. It will still be an improvement over the old
639 * code, and it will be simple to change the scale factor if we find that it
640 * becomes a problem on bigger systems.
641 */
642static unsigned long lazy_max_pages(void)
643{
644 unsigned int log;
645
646 log = fls(num_online_cpus());
647
648 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
649}
650
651static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
652
653/*
654 * Serialize vmap purging. There is no actual criticial section protected
655 * by this look, but we want to avoid concurrent calls for performance
656 * reasons and to make the pcpu_get_vm_areas more deterministic.
657 */
658static DEFINE_MUTEX(vmap_purge_lock);
659
660/* for per-CPU blocks */
661static void purge_fragmented_blocks_allcpus(void);
662
663/*
664 * called before a call to iounmap() if the caller wants vm_area_struct's
665 * immediately freed.
666 */
667void set_iounmap_nonlazy(void)
668{
669 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
670}
671
672/*
673 * Purges all lazily-freed vmap areas.
674 */
675static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
676{
677 struct llist_node *valist;
678 struct vmap_area *va;
679 struct vmap_area *n_va;
680 bool do_free = false;
681
682 lockdep_assert_held(&vmap_purge_lock);
683
684 valist = llist_del_all(&vmap_purge_list);
685 llist_for_each_entry(va, valist, purge_list) {
686 if (va->va_start < start)
687 start = va->va_start;
688 if (va->va_end > end)
689 end = va->va_end;
690 do_free = true;
691 }
692
693 if (!do_free)
694 return false;
695
696 flush_tlb_kernel_range(start, end);
697
698 spin_lock(&vmap_area_lock);
699 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
700 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
701
702 __free_vmap_area(va);
703 atomic_sub(nr, &vmap_lazy_nr);
704 cond_resched_lock(&vmap_area_lock);
705 }
706 spin_unlock(&vmap_area_lock);
707 return true;
708}
709
710/*
711 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
712 * is already purging.
713 */
714static void try_purge_vmap_area_lazy(void)
715{
716 if (mutex_trylock(&vmap_purge_lock)) {
717 __purge_vmap_area_lazy(ULONG_MAX, 0);
718 mutex_unlock(&vmap_purge_lock);
719 }
720}
721
722/*
723 * Kick off a purge of the outstanding lazy areas.
724 */
725static void purge_vmap_area_lazy(void)
726{
727 mutex_lock(&vmap_purge_lock);
728 purge_fragmented_blocks_allcpus();
729 __purge_vmap_area_lazy(ULONG_MAX, 0);
730 mutex_unlock(&vmap_purge_lock);
731}
732
733/*
734 * Free a vmap area, caller ensuring that the area has been unmapped
735 * and flush_cache_vunmap had been called for the correct range
736 * previously.
737 */
738static void free_vmap_area_noflush(struct vmap_area *va)
739{
740 int nr_lazy;
741
742 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
743 &vmap_lazy_nr);
744
745 /* After this point, we may free va at any time */
746 llist_add(&va->purge_list, &vmap_purge_list);
747
748 if (unlikely(nr_lazy > lazy_max_pages()))
749 try_purge_vmap_area_lazy();
750}
751
752/*
753 * Free and unmap a vmap area
754 */
755static void free_unmap_vmap_area(struct vmap_area *va)
756{
757 flush_cache_vunmap(va->va_start, va->va_end);
758 unmap_vmap_area(va);
759 free_vmap_area_noflush(va);
760}
761
762static struct vmap_area *find_vmap_area(unsigned long addr)
763{
764 struct vmap_area *va;
765
766 spin_lock(&vmap_area_lock);
767 va = __find_vmap_area(addr);
768 spin_unlock(&vmap_area_lock);
769
770 return va;
771}
772
773/*** Per cpu kva allocator ***/
774
775/*
776 * vmap space is limited especially on 32 bit architectures. Ensure there is
777 * room for at least 16 percpu vmap blocks per CPU.
778 */
779/*
780 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
781 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
782 * instead (we just need a rough idea)
783 */
784#if BITS_PER_LONG == 32
785#define VMALLOC_SPACE (128UL*1024*1024)
786#else
787#define VMALLOC_SPACE (128UL*1024*1024*1024)
788#endif
789
790#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
791#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
792#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
793#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
794#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
795#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
796#define VMAP_BBMAP_BITS \
797 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
798 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
799 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
800
801#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
802
803static bool vmap_initialized __read_mostly = false;
804
805struct vmap_block_queue {
806 spinlock_t lock;
807 struct list_head free;
808};
809
810struct vmap_block {
811 spinlock_t lock;
812 struct vmap_area *va;
813 unsigned long free, dirty;
814 unsigned long dirty_min, dirty_max; /*< dirty range */
815 struct list_head free_list;
816 struct rcu_head rcu_head;
817 struct list_head purge;
818};
819
820/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
821static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
822
823/*
824 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
825 * in the free path. Could get rid of this if we change the API to return a
826 * "cookie" from alloc, to be passed to free. But no big deal yet.
827 */
828static DEFINE_SPINLOCK(vmap_block_tree_lock);
829static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
830
831/*
832 * We should probably have a fallback mechanism to allocate virtual memory
833 * out of partially filled vmap blocks. However vmap block sizing should be
834 * fairly reasonable according to the vmalloc size, so it shouldn't be a
835 * big problem.
836 */
837
838static unsigned long addr_to_vb_idx(unsigned long addr)
839{
840 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
841 addr /= VMAP_BLOCK_SIZE;
842 return addr;
843}
844
845static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
846{
847 unsigned long addr;
848
849 addr = va_start + (pages_off << PAGE_SHIFT);
850 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
851 return (void *)addr;
852}
853
854/**
855 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
856 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
857 * @order: how many 2^order pages should be occupied in newly allocated block
858 * @gfp_mask: flags for the page level allocator
859 *
860 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
861 */
862static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
863{
864 struct vmap_block_queue *vbq;
865 struct vmap_block *vb;
866 struct vmap_area *va;
867 unsigned long vb_idx;
868 int node, err;
869 void *vaddr;
870
871 node = numa_node_id();
872
873 vb = kmalloc_node(sizeof(struct vmap_block),
874 gfp_mask & GFP_RECLAIM_MASK, node);
875 if (unlikely(!vb))
876 return ERR_PTR(-ENOMEM);
877
878 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
879 VMALLOC_START, VMALLOC_END,
880 node, gfp_mask);
881 if (IS_ERR(va)) {
882 kfree(vb);
883 return ERR_CAST(va);
884 }
885
886 err = radix_tree_preload(gfp_mask);
887 if (unlikely(err)) {
888 kfree(vb);
889 free_vmap_area(va);
890 return ERR_PTR(err);
891 }
892
893 vaddr = vmap_block_vaddr(va->va_start, 0);
894 spin_lock_init(&vb->lock);
895 vb->va = va;
896 /* At least something should be left free */
897 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
898 vb->free = VMAP_BBMAP_BITS - (1UL << order);
899 vb->dirty = 0;
900 vb->dirty_min = VMAP_BBMAP_BITS;
901 vb->dirty_max = 0;
902 INIT_LIST_HEAD(&vb->free_list);
903
904 vb_idx = addr_to_vb_idx(va->va_start);
905 spin_lock(&vmap_block_tree_lock);
906 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
907 spin_unlock(&vmap_block_tree_lock);
908 BUG_ON(err);
909 radix_tree_preload_end();
910
911 vbq = &get_cpu_var(vmap_block_queue);
912 spin_lock(&vbq->lock);
913 list_add_tail_rcu(&vb->free_list, &vbq->free);
914 spin_unlock(&vbq->lock);
915 put_cpu_var(vmap_block_queue);
916
917 return vaddr;
918}
919
920static void free_vmap_block(struct vmap_block *vb)
921{
922 struct vmap_block *tmp;
923 unsigned long vb_idx;
924
925 vb_idx = addr_to_vb_idx(vb->va->va_start);
926 spin_lock(&vmap_block_tree_lock);
927 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
928 spin_unlock(&vmap_block_tree_lock);
929 BUG_ON(tmp != vb);
930
931 free_vmap_area_noflush(vb->va);
932 kfree_rcu(vb, rcu_head);
933}
934
935static void purge_fragmented_blocks(int cpu)
936{
937 LIST_HEAD(purge);
938 struct vmap_block *vb;
939 struct vmap_block *n_vb;
940 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
941
942 rcu_read_lock();
943 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
944
945 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
946 continue;
947
948 spin_lock(&vb->lock);
949 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
950 vb->free = 0; /* prevent further allocs after releasing lock */
951 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
952 vb->dirty_min = 0;
953 vb->dirty_max = VMAP_BBMAP_BITS;
954 spin_lock(&vbq->lock);
955 list_del_rcu(&vb->free_list);
956 spin_unlock(&vbq->lock);
957 spin_unlock(&vb->lock);
958 list_add_tail(&vb->purge, &purge);
959 } else
960 spin_unlock(&vb->lock);
961 }
962 rcu_read_unlock();
963
964 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
965 list_del(&vb->purge);
966 free_vmap_block(vb);
967 }
968}
969
970static void purge_fragmented_blocks_allcpus(void)
971{
972 int cpu;
973
974 for_each_possible_cpu(cpu)
975 purge_fragmented_blocks(cpu);
976}
977
978static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
979{
980 struct vmap_block_queue *vbq;
981 struct vmap_block *vb;
982 void *vaddr = NULL;
983 unsigned int order;
984
985 BUG_ON(offset_in_page(size));
986 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
987 if (WARN_ON(size == 0)) {
988 /*
989 * Allocating 0 bytes isn't what caller wants since
990 * get_order(0) returns funny result. Just warn and terminate
991 * early.
992 */
993 return NULL;
994 }
995 order = get_order(size);
996
997 rcu_read_lock();
998 vbq = &get_cpu_var(vmap_block_queue);
999 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1000 unsigned long pages_off;
1001
1002 spin_lock(&vb->lock);
1003 if (vb->free < (1UL << order)) {
1004 spin_unlock(&vb->lock);
1005 continue;
1006 }
1007
1008 pages_off = VMAP_BBMAP_BITS - vb->free;
1009 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1010 vb->free -= 1UL << order;
1011 if (vb->free == 0) {
1012 spin_lock(&vbq->lock);
1013 list_del_rcu(&vb->free_list);
1014 spin_unlock(&vbq->lock);
1015 }
1016
1017 spin_unlock(&vb->lock);
1018 break;
1019 }
1020
1021 put_cpu_var(vmap_block_queue);
1022 rcu_read_unlock();
1023
1024 /* Allocate new block if nothing was found */
1025 if (!vaddr)
1026 vaddr = new_vmap_block(order, gfp_mask);
1027
1028 return vaddr;
1029}
1030
1031static void vb_free(const void *addr, unsigned long size)
1032{
1033 unsigned long offset;
1034 unsigned long vb_idx;
1035 unsigned int order;
1036 struct vmap_block *vb;
1037
1038 BUG_ON(offset_in_page(size));
1039 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1040
1041 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1042
1043 order = get_order(size);
1044
1045 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1046 offset >>= PAGE_SHIFT;
1047
1048 vb_idx = addr_to_vb_idx((unsigned long)addr);
1049 rcu_read_lock();
1050 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1051 rcu_read_unlock();
1052 BUG_ON(!vb);
1053
1054 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1055
1056 spin_lock(&vb->lock);
1057
1058 /* Expand dirty range */
1059 vb->dirty_min = min(vb->dirty_min, offset);
1060 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1061
1062 vb->dirty += 1UL << order;
1063 if (vb->dirty == VMAP_BBMAP_BITS) {
1064 BUG_ON(vb->free);
1065 spin_unlock(&vb->lock);
1066 free_vmap_block(vb);
1067 } else
1068 spin_unlock(&vb->lock);
1069}
1070
1071/**
1072 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1073 *
1074 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1075 * to amortize TLB flushing overheads. What this means is that any page you
1076 * have now, may, in a former life, have been mapped into kernel virtual
1077 * address by the vmap layer and so there might be some CPUs with TLB entries
1078 * still referencing that page (additional to the regular 1:1 kernel mapping).
1079 *
1080 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1081 * be sure that none of the pages we have control over will have any aliases
1082 * from the vmap layer.
1083 */
1084void vm_unmap_aliases(void)
1085{
1086 unsigned long start = ULONG_MAX, end = 0;
1087 int cpu;
1088 int flush = 0;
1089
1090 if (unlikely(!vmap_initialized))
1091 return;
1092
1093 might_sleep();
1094
1095 for_each_possible_cpu(cpu) {
1096 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1097 struct vmap_block *vb;
1098
1099 rcu_read_lock();
1100 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1101 spin_lock(&vb->lock);
1102 if (vb->dirty) {
1103 unsigned long va_start = vb->va->va_start;
1104 unsigned long s, e;
1105
1106 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1107 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1108
1109 start = min(s, start);
1110 end = max(e, end);
1111
1112 flush = 1;
1113 }
1114 spin_unlock(&vb->lock);
1115 }
1116 rcu_read_unlock();
1117 }
1118
1119 mutex_lock(&vmap_purge_lock);
1120 purge_fragmented_blocks_allcpus();
1121 if (!__purge_vmap_area_lazy(start, end) && flush)
1122 flush_tlb_kernel_range(start, end);
1123 mutex_unlock(&vmap_purge_lock);
1124}
1125EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1126
1127/**
1128 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1129 * @mem: the pointer returned by vm_map_ram
1130 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1131 */
1132void vm_unmap_ram(const void *mem, unsigned int count)
1133{
1134 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1135 unsigned long addr = (unsigned long)mem;
1136 struct vmap_area *va;
1137
1138 might_sleep();
1139 BUG_ON(!addr);
1140 BUG_ON(addr < VMALLOC_START);
1141 BUG_ON(addr > VMALLOC_END);
1142 BUG_ON(!PAGE_ALIGNED(addr));
1143
1144 debug_check_no_locks_freed(mem, size);
1145 vmap_debug_free_range(addr, addr+size);
1146
1147 if (likely(count <= VMAP_MAX_ALLOC)) {
1148 vb_free(mem, size);
1149 return;
1150 }
1151
1152 va = find_vmap_area(addr);
1153 BUG_ON(!va);
1154 free_unmap_vmap_area(va);
1155}
1156EXPORT_SYMBOL(vm_unmap_ram);
1157
1158/**
1159 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1160 * @pages: an array of pointers to the pages to be mapped
1161 * @count: number of pages
1162 * @node: prefer to allocate data structures on this node
1163 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1164 *
1165 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1166 * faster than vmap so it's good. But if you mix long-life and short-life
1167 * objects with vm_map_ram(), it could consume lots of address space through
1168 * fragmentation (especially on a 32bit machine). You could see failures in
1169 * the end. Please use this function for short-lived objects.
1170 *
1171 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1172 */
1173void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1174{
1175 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1176 unsigned long addr;
1177 void *mem;
1178
1179 if (likely(count <= VMAP_MAX_ALLOC)) {
1180 mem = vb_alloc(size, GFP_KERNEL);
1181 if (IS_ERR(mem))
1182 return NULL;
1183 addr = (unsigned long)mem;
1184 } else {
1185 struct vmap_area *va;
1186 va = alloc_vmap_area(size, PAGE_SIZE,
1187 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1188 if (IS_ERR(va))
1189 return NULL;
1190
1191 addr = va->va_start;
1192 mem = (void *)addr;
1193 }
1194 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1195 vm_unmap_ram(mem, count);
1196 return NULL;
1197 }
1198 return mem;
1199}
1200EXPORT_SYMBOL(vm_map_ram);
1201
1202static struct vm_struct *vmlist __initdata;
1203/**
1204 * vm_area_add_early - add vmap area early during boot
1205 * @vm: vm_struct to add
1206 *
1207 * This function is used to add fixed kernel vm area to vmlist before
1208 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1209 * should contain proper values and the other fields should be zero.
1210 *
1211 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1212 */
1213void __init vm_area_add_early(struct vm_struct *vm)
1214{
1215 struct vm_struct *tmp, **p;
1216
1217 BUG_ON(vmap_initialized);
1218 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1219 if (tmp->addr >= vm->addr) {
1220 BUG_ON(tmp->addr < vm->addr + vm->size);
1221 break;
1222 } else
1223 BUG_ON(tmp->addr + tmp->size > vm->addr);
1224 }
1225 vm->next = *p;
1226 *p = vm;
1227}
1228
1229/**
1230 * vm_area_register_early - register vmap area early during boot
1231 * @vm: vm_struct to register
1232 * @align: requested alignment
1233 *
1234 * This function is used to register kernel vm area before
1235 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1236 * proper values on entry and other fields should be zero. On return,
1237 * vm->addr contains the allocated address.
1238 *
1239 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1240 */
1241void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1242{
1243 static size_t vm_init_off __initdata;
1244 unsigned long addr;
1245
1246 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1247 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1248
1249 vm->addr = (void *)addr;
1250
1251 vm_area_add_early(vm);
1252}
1253
1254void __init vmalloc_init(void)
1255{
1256 struct vmap_area *va;
1257 struct vm_struct *tmp;
1258 int i;
1259
1260 for_each_possible_cpu(i) {
1261 struct vmap_block_queue *vbq;
1262 struct vfree_deferred *p;
1263
1264 vbq = &per_cpu(vmap_block_queue, i);
1265 spin_lock_init(&vbq->lock);
1266 INIT_LIST_HEAD(&vbq->free);
1267 p = &per_cpu(vfree_deferred, i);
1268 init_llist_head(&p->list);
1269 INIT_WORK(&p->wq, free_work);
1270 }
1271
1272 /* Import existing vmlist entries. */
1273 for (tmp = vmlist; tmp; tmp = tmp->next) {
1274 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1275 va->flags = VM_VM_AREA;
1276 va->va_start = (unsigned long)tmp->addr;
1277 va->va_end = va->va_start + tmp->size;
1278 va->vm = tmp;
1279 __insert_vmap_area(va);
1280 }
1281
1282 vmap_area_pcpu_hole = VMALLOC_END;
1283
1284 vmap_initialized = true;
1285}
1286
1287/**
1288 * map_kernel_range_noflush - map kernel VM area with the specified pages
1289 * @addr: start of the VM area to map
1290 * @size: size of the VM area to map
1291 * @prot: page protection flags to use
1292 * @pages: pages to map
1293 *
1294 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1295 * specify should have been allocated using get_vm_area() and its
1296 * friends.
1297 *
1298 * NOTE:
1299 * This function does NOT do any cache flushing. The caller is
1300 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1301 * before calling this function.
1302 *
1303 * RETURNS:
1304 * The number of pages mapped on success, -errno on failure.
1305 */
1306int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1307 pgprot_t prot, struct page **pages)
1308{
1309 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1310}
1311
1312/**
1313 * unmap_kernel_range_noflush - unmap kernel VM area
1314 * @addr: start of the VM area to unmap
1315 * @size: size of the VM area to unmap
1316 *
1317 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1318 * specify should have been allocated using get_vm_area() and its
1319 * friends.
1320 *
1321 * NOTE:
1322 * This function does NOT do any cache flushing. The caller is
1323 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1324 * before calling this function and flush_tlb_kernel_range() after.
1325 */
1326void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1327{
1328 vunmap_page_range(addr, addr + size);
1329}
1330EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1331
1332/**
1333 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1334 * @addr: start of the VM area to unmap
1335 * @size: size of the VM area to unmap
1336 *
1337 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1338 * the unmapping and tlb after.
1339 */
1340void unmap_kernel_range(unsigned long addr, unsigned long size)
1341{
1342 unsigned long end = addr + size;
1343
1344 flush_cache_vunmap(addr, end);
1345 vunmap_page_range(addr, end);
1346 flush_tlb_kernel_range(addr, end);
1347}
1348EXPORT_SYMBOL_GPL(unmap_kernel_range);
1349
1350int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1351{
1352 unsigned long addr = (unsigned long)area->addr;
1353 unsigned long end = addr + get_vm_area_size(area);
1354 int err;
1355
1356 err = vmap_page_range(addr, end, prot, pages);
1357
1358 return err > 0 ? 0 : err;
1359}
1360EXPORT_SYMBOL_GPL(map_vm_area);
1361
1362static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1363 unsigned long flags, const void *caller)
1364{
1365 spin_lock(&vmap_area_lock);
1366 vm->flags = flags;
1367 vm->addr = (void *)va->va_start;
1368 vm->size = va->va_end - va->va_start;
1369 vm->caller = caller;
1370 va->vm = vm;
1371 va->flags |= VM_VM_AREA;
1372 spin_unlock(&vmap_area_lock);
1373}
1374
1375static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1376{
1377 /*
1378 * Before removing VM_UNINITIALIZED,
1379 * we should make sure that vm has proper values.
1380 * Pair with smp_rmb() in show_numa_info().
1381 */
1382 smp_wmb();
1383 vm->flags &= ~VM_UNINITIALIZED;
1384}
1385
1386static struct vm_struct *__get_vm_area_node(unsigned long size,
1387 unsigned long align, unsigned long flags, unsigned long start,
1388 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1389{
1390 struct vmap_area *va;
1391 struct vm_struct *area;
1392
1393 BUG_ON(in_interrupt());
1394 size = PAGE_ALIGN(size);
1395 if (unlikely(!size))
1396 return NULL;
1397
1398 if (flags & VM_IOREMAP)
1399 align = 1ul << clamp_t(int, get_count_order_long(size),
1400 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1401
1402 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1403 if (unlikely(!area))
1404 return NULL;
1405
1406 if (!(flags & VM_NO_GUARD))
1407 size += PAGE_SIZE;
1408
1409 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1410 if (IS_ERR(va)) {
1411 kfree(area);
1412 return NULL;
1413 }
1414
1415 setup_vmalloc_vm(area, va, flags, caller);
1416
1417 return area;
1418}
1419
1420struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1421 unsigned long start, unsigned long end)
1422{
1423 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1424 GFP_KERNEL, __builtin_return_address(0));
1425}
1426EXPORT_SYMBOL_GPL(__get_vm_area);
1427
1428struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1429 unsigned long start, unsigned long end,
1430 const void *caller)
1431{
1432 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1433 GFP_KERNEL, caller);
1434}
1435
1436/**
1437 * get_vm_area - reserve a contiguous kernel virtual area
1438 * @size: size of the area
1439 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1440 *
1441 * Search an area of @size in the kernel virtual mapping area,
1442 * and reserved it for out purposes. Returns the area descriptor
1443 * on success or %NULL on failure.
1444 */
1445struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1446{
1447 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1448 NUMA_NO_NODE, GFP_KERNEL,
1449 __builtin_return_address(0));
1450}
1451
1452struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1453 const void *caller)
1454{
1455 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1456 NUMA_NO_NODE, GFP_KERNEL, caller);
1457}
1458
1459/**
1460 * find_vm_area - find a continuous kernel virtual area
1461 * @addr: base address
1462 *
1463 * Search for the kernel VM area starting at @addr, and return it.
1464 * It is up to the caller to do all required locking to keep the returned
1465 * pointer valid.
1466 */
1467struct vm_struct *find_vm_area(const void *addr)
1468{
1469 struct vmap_area *va;
1470
1471 va = find_vmap_area((unsigned long)addr);
1472 if (va && va->flags & VM_VM_AREA)
1473 return va->vm;
1474
1475 return NULL;
1476}
1477
1478/**
1479 * remove_vm_area - find and remove a continuous kernel virtual area
1480 * @addr: base address
1481 *
1482 * Search for the kernel VM area starting at @addr, and remove it.
1483 * This function returns the found VM area, but using it is NOT safe
1484 * on SMP machines, except for its size or flags.
1485 */
1486struct vm_struct *remove_vm_area(const void *addr)
1487{
1488 struct vmap_area *va;
1489
1490 might_sleep();
1491
1492 va = find_vmap_area((unsigned long)addr);
1493 if (va && va->flags & VM_VM_AREA) {
1494 struct vm_struct *vm = va->vm;
1495
1496 spin_lock(&vmap_area_lock);
1497 va->vm = NULL;
1498 va->flags &= ~VM_VM_AREA;
1499 va->flags |= VM_LAZY_FREE;
1500 spin_unlock(&vmap_area_lock);
1501
1502 vmap_debug_free_range(va->va_start, va->va_end);
1503 kasan_free_shadow(vm);
1504 free_unmap_vmap_area(va);
1505
1506 return vm;
1507 }
1508 return NULL;
1509}
1510
1511static void __vunmap(const void *addr, int deallocate_pages)
1512{
1513 struct vm_struct *area;
1514
1515 if (!addr)
1516 return;
1517
1518 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1519 addr))
1520 return;
1521
1522 area = remove_vm_area(addr);
1523 if (unlikely(!area)) {
1524 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1525 addr);
1526 return;
1527 }
1528
1529 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1530 debug_check_no_obj_freed(addr, get_vm_area_size(area));
1531
1532 if (deallocate_pages) {
1533 int i;
1534
1535 for (i = 0; i < area->nr_pages; i++) {
1536 struct page *page = area->pages[i];
1537
1538 BUG_ON(!page);
1539 __free_pages(page, 0);
1540 }
1541
1542 kvfree(area->pages);
1543 }
1544
1545 kfree(area);
1546 return;
1547}
1548
1549static inline void __vfree_deferred(const void *addr)
1550{
1551 /*
1552 * Use raw_cpu_ptr() because this can be called from preemptible
1553 * context. Preemption is absolutely fine here, because the llist_add()
1554 * implementation is lockless, so it works even if we are adding to
1555 * nother cpu's list. schedule_work() should be fine with this too.
1556 */
1557 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1558
1559 if (llist_add((struct llist_node *)addr, &p->list))
1560 schedule_work(&p->wq);
1561}
1562
1563/**
1564 * vfree_atomic - release memory allocated by vmalloc()
1565 * @addr: memory base address
1566 *
1567 * This one is just like vfree() but can be called in any atomic context
1568 * except NMIs.
1569 */
1570void vfree_atomic(const void *addr)
1571{
1572 BUG_ON(in_nmi());
1573
1574 kmemleak_free(addr);
1575
1576 if (!addr)
1577 return;
1578 __vfree_deferred(addr);
1579}
1580
1581/**
1582 * vfree - release memory allocated by vmalloc()
1583 * @addr: memory base address
1584 *
1585 * Free the virtually continuous memory area starting at @addr, as
1586 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1587 * NULL, no operation is performed.
1588 *
1589 * Must not be called in NMI context (strictly speaking, only if we don't
1590 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1591 * conventions for vfree() arch-depenedent would be a really bad idea)
1592 *
1593 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1594 */
1595void vfree(const void *addr)
1596{
1597 BUG_ON(in_nmi());
1598
1599 kmemleak_free(addr);
1600
1601 if (!addr)
1602 return;
1603 if (unlikely(in_interrupt()))
1604 __vfree_deferred(addr);
1605 else
1606 __vunmap(addr, 1);
1607}
1608EXPORT_SYMBOL(vfree);
1609
1610/**
1611 * vunmap - release virtual mapping obtained by vmap()
1612 * @addr: memory base address
1613 *
1614 * Free the virtually contiguous memory area starting at @addr,
1615 * which was created from the page array passed to vmap().
1616 *
1617 * Must not be called in interrupt context.
1618 */
1619void vunmap(const void *addr)
1620{
1621 BUG_ON(in_interrupt());
1622 might_sleep();
1623 if (addr)
1624 __vunmap(addr, 0);
1625}
1626EXPORT_SYMBOL(vunmap);
1627
1628/**
1629 * vmap - map an array of pages into virtually contiguous space
1630 * @pages: array of page pointers
1631 * @count: number of pages to map
1632 * @flags: vm_area->flags
1633 * @prot: page protection for the mapping
1634 *
1635 * Maps @count pages from @pages into contiguous kernel virtual
1636 * space.
1637 */
1638void *vmap(struct page **pages, unsigned int count,
1639 unsigned long flags, pgprot_t prot)
1640{
1641 struct vm_struct *area;
1642 unsigned long size; /* In bytes */
1643
1644 might_sleep();
1645
1646 if (count > totalram_pages)
1647 return NULL;
1648
1649 size = (unsigned long)count << PAGE_SHIFT;
1650 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1651 if (!area)
1652 return NULL;
1653
1654 if (map_vm_area(area, prot, pages)) {
1655 vunmap(area->addr);
1656 return NULL;
1657 }
1658
1659 return area->addr;
1660}
1661EXPORT_SYMBOL(vmap);
1662
1663static void *__vmalloc_node(unsigned long size, unsigned long align,
1664 gfp_t gfp_mask, pgprot_t prot,
1665 int node, const void *caller);
1666static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1667 pgprot_t prot, int node)
1668{
1669 struct page **pages;
1670 unsigned int nr_pages, array_size, i;
1671 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1672 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1673 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
1674 0 :
1675 __GFP_HIGHMEM;
1676
1677 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1678 array_size = (nr_pages * sizeof(struct page *));
1679
1680 area->nr_pages = nr_pages;
1681 /* Please note that the recursion is strictly bounded. */
1682 if (array_size > PAGE_SIZE) {
1683 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
1684 PAGE_KERNEL, node, area->caller);
1685 } else {
1686 pages = kmalloc_node(array_size, nested_gfp, node);
1687 }
1688 area->pages = pages;
1689 if (!area->pages) {
1690 remove_vm_area(area->addr);
1691 kfree(area);
1692 return NULL;
1693 }
1694
1695 for (i = 0; i < area->nr_pages; i++) {
1696 struct page *page;
1697
1698 if (node == NUMA_NO_NODE)
1699 page = alloc_page(alloc_mask|highmem_mask);
1700 else
1701 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
1702
1703 if (unlikely(!page)) {
1704 /* Successfully allocated i pages, free them in __vunmap() */
1705 area->nr_pages = i;
1706 goto fail;
1707 }
1708 area->pages[i] = page;
1709 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
1710 cond_resched();
1711 }
1712
1713 if (map_vm_area(area, prot, pages))
1714 goto fail;
1715 return area->addr;
1716
1717fail:
1718 warn_alloc(gfp_mask, NULL,
1719 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1720 (area->nr_pages*PAGE_SIZE), area->size);
1721 vfree(area->addr);
1722 return NULL;
1723}
1724
1725/**
1726 * __vmalloc_node_range - allocate virtually contiguous memory
1727 * @size: allocation size
1728 * @align: desired alignment
1729 * @start: vm area range start
1730 * @end: vm area range end
1731 * @gfp_mask: flags for the page level allocator
1732 * @prot: protection mask for the allocated pages
1733 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1734 * @node: node to use for allocation or NUMA_NO_NODE
1735 * @caller: caller's return address
1736 *
1737 * Allocate enough pages to cover @size from the page level
1738 * allocator with @gfp_mask flags. Map them into contiguous
1739 * kernel virtual space, using a pagetable protection of @prot.
1740 */
1741void *__vmalloc_node_range(unsigned long size, unsigned long align,
1742 unsigned long start, unsigned long end, gfp_t gfp_mask,
1743 pgprot_t prot, unsigned long vm_flags, int node,
1744 const void *caller)
1745{
1746 struct vm_struct *area;
1747 void *addr;
1748 unsigned long real_size = size;
1749
1750 size = PAGE_ALIGN(size);
1751 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1752 goto fail;
1753
1754 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1755 vm_flags, start, end, node, gfp_mask, caller);
1756 if (!area)
1757 goto fail;
1758
1759 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1760 if (!addr)
1761 return NULL;
1762
1763 /*
1764 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1765 * flag. It means that vm_struct is not fully initialized.
1766 * Now, it is fully initialized, so remove this flag here.
1767 */
1768 clear_vm_uninitialized_flag(area);
1769
1770 kmemleak_vmalloc(area, size, gfp_mask);
1771
1772 return addr;
1773
1774fail:
1775 warn_alloc(gfp_mask, NULL,
1776 "vmalloc: allocation failure: %lu bytes", real_size);
1777 return NULL;
1778}
1779
1780/**
1781 * __vmalloc_node - allocate virtually contiguous memory
1782 * @size: allocation size
1783 * @align: desired alignment
1784 * @gfp_mask: flags for the page level allocator
1785 * @prot: protection mask for the allocated pages
1786 * @node: node to use for allocation or NUMA_NO_NODE
1787 * @caller: caller's return address
1788 *
1789 * Allocate enough pages to cover @size from the page level
1790 * allocator with @gfp_mask flags. Map them into contiguous
1791 * kernel virtual space, using a pagetable protection of @prot.
1792 *
1793 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1794 * and __GFP_NOFAIL are not supported
1795 *
1796 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1797 * with mm people.
1798 *
1799 */
1800static void *__vmalloc_node(unsigned long size, unsigned long align,
1801 gfp_t gfp_mask, pgprot_t prot,
1802 int node, const void *caller)
1803{
1804 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1805 gfp_mask, prot, 0, node, caller);
1806}
1807
1808void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1809{
1810 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1811 __builtin_return_address(0));
1812}
1813EXPORT_SYMBOL(__vmalloc);
1814
1815static inline void *__vmalloc_node_flags(unsigned long size,
1816 int node, gfp_t flags)
1817{
1818 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1819 node, __builtin_return_address(0));
1820}
1821
1822
1823void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
1824 void *caller)
1825{
1826 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
1827}
1828
1829/**
1830 * vmalloc - allocate virtually contiguous memory
1831 * @size: allocation size
1832 * Allocate enough pages to cover @size from the page level
1833 * allocator and map them into contiguous kernel virtual space.
1834 *
1835 * For tight control over page level allocator and protection flags
1836 * use __vmalloc() instead.
1837 */
1838void *vmalloc(unsigned long size)
1839{
1840 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1841 GFP_KERNEL);
1842}
1843EXPORT_SYMBOL(vmalloc);
1844
1845/**
1846 * vzalloc - allocate virtually contiguous memory with zero fill
1847 * @size: allocation size
1848 * Allocate enough pages to cover @size from the page level
1849 * allocator and map them into contiguous kernel virtual space.
1850 * The memory allocated is set to zero.
1851 *
1852 * For tight control over page level allocator and protection flags
1853 * use __vmalloc() instead.
1854 */
1855void *vzalloc(unsigned long size)
1856{
1857 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1858 GFP_KERNEL | __GFP_ZERO);
1859}
1860EXPORT_SYMBOL(vzalloc);
1861
1862/**
1863 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1864 * @size: allocation size
1865 *
1866 * The resulting memory area is zeroed so it can be mapped to userspace
1867 * without leaking data.
1868 */
1869void *vmalloc_user(unsigned long size)
1870{
1871 struct vm_struct *area;
1872 void *ret;
1873
1874 ret = __vmalloc_node(size, SHMLBA,
1875 GFP_KERNEL | __GFP_ZERO,
1876 PAGE_KERNEL, NUMA_NO_NODE,
1877 __builtin_return_address(0));
1878 if (ret) {
1879 area = find_vm_area(ret);
1880 area->flags |= VM_USERMAP;
1881 }
1882 return ret;
1883}
1884EXPORT_SYMBOL(vmalloc_user);
1885
1886/**
1887 * vmalloc_node - allocate memory on a specific node
1888 * @size: allocation size
1889 * @node: numa node
1890 *
1891 * Allocate enough pages to cover @size from the page level
1892 * allocator and map them into contiguous kernel virtual space.
1893 *
1894 * For tight control over page level allocator and protection flags
1895 * use __vmalloc() instead.
1896 */
1897void *vmalloc_node(unsigned long size, int node)
1898{
1899 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
1900 node, __builtin_return_address(0));
1901}
1902EXPORT_SYMBOL(vmalloc_node);
1903
1904/**
1905 * vzalloc_node - allocate memory on a specific node with zero fill
1906 * @size: allocation size
1907 * @node: numa node
1908 *
1909 * Allocate enough pages to cover @size from the page level
1910 * allocator and map them into contiguous kernel virtual space.
1911 * The memory allocated is set to zero.
1912 *
1913 * For tight control over page level allocator and protection flags
1914 * use __vmalloc_node() instead.
1915 */
1916void *vzalloc_node(unsigned long size, int node)
1917{
1918 return __vmalloc_node_flags(size, node,
1919 GFP_KERNEL | __GFP_ZERO);
1920}
1921EXPORT_SYMBOL(vzalloc_node);
1922
1923#ifndef PAGE_KERNEL_EXEC
1924# define PAGE_KERNEL_EXEC PAGE_KERNEL
1925#endif
1926
1927/**
1928 * vmalloc_exec - allocate virtually contiguous, executable memory
1929 * @size: allocation size
1930 *
1931 * Kernel-internal function to allocate enough pages to cover @size
1932 * the page level allocator and map them into contiguous and
1933 * executable kernel virtual space.
1934 *
1935 * For tight control over page level allocator and protection flags
1936 * use __vmalloc() instead.
1937 */
1938
1939void *vmalloc_exec(unsigned long size)
1940{
1941 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL_EXEC,
1942 NUMA_NO_NODE, __builtin_return_address(0));
1943}
1944
1945#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1946#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
1947#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1948#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
1949#else
1950/*
1951 * 64b systems should always have either DMA or DMA32 zones. For others
1952 * GFP_DMA32 should do the right thing and use the normal zone.
1953 */
1954#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1955#endif
1956
1957/**
1958 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1959 * @size: allocation size
1960 *
1961 * Allocate enough 32bit PA addressable pages to cover @size from the
1962 * page level allocator and map them into contiguous kernel virtual space.
1963 */
1964void *vmalloc_32(unsigned long size)
1965{
1966 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1967 NUMA_NO_NODE, __builtin_return_address(0));
1968}
1969EXPORT_SYMBOL(vmalloc_32);
1970
1971/**
1972 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1973 * @size: allocation size
1974 *
1975 * The resulting memory area is 32bit addressable and zeroed so it can be
1976 * mapped to userspace without leaking data.
1977 */
1978void *vmalloc_32_user(unsigned long size)
1979{
1980 struct vm_struct *area;
1981 void *ret;
1982
1983 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1984 NUMA_NO_NODE, __builtin_return_address(0));
1985 if (ret) {
1986 area = find_vm_area(ret);
1987 area->flags |= VM_USERMAP;
1988 }
1989 return ret;
1990}
1991EXPORT_SYMBOL(vmalloc_32_user);
1992
1993/*
1994 * small helper routine , copy contents to buf from addr.
1995 * If the page is not present, fill zero.
1996 */
1997
1998static int aligned_vread(char *buf, char *addr, unsigned long count)
1999{
2000 struct page *p;
2001 int copied = 0;
2002
2003 while (count) {
2004 unsigned long offset, length;
2005
2006 offset = offset_in_page(addr);
2007 length = PAGE_SIZE - offset;
2008 if (length > count)
2009 length = count;
2010 p = vmalloc_to_page(addr);
2011 /*
2012 * To do safe access to this _mapped_ area, we need
2013 * lock. But adding lock here means that we need to add
2014 * overhead of vmalloc()/vfree() calles for this _debug_
2015 * interface, rarely used. Instead of that, we'll use
2016 * kmap() and get small overhead in this access function.
2017 */
2018 if (p) {
2019 /*
2020 * we can expect USER0 is not used (see vread/vwrite's
2021 * function description)
2022 */
2023 void *map = kmap_atomic(p);
2024 memcpy(buf, map + offset, length);
2025 kunmap_atomic(map);
2026 } else
2027 memset(buf, 0, length);
2028
2029 addr += length;
2030 buf += length;
2031 copied += length;
2032 count -= length;
2033 }
2034 return copied;
2035}
2036
2037static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2038{
2039 struct page *p;
2040 int copied = 0;
2041
2042 while (count) {
2043 unsigned long offset, length;
2044
2045 offset = offset_in_page(addr);
2046 length = PAGE_SIZE - offset;
2047 if (length > count)
2048 length = count;
2049 p = vmalloc_to_page(addr);
2050 /*
2051 * To do safe access to this _mapped_ area, we need
2052 * lock. But adding lock here means that we need to add
2053 * overhead of vmalloc()/vfree() calles for this _debug_
2054 * interface, rarely used. Instead of that, we'll use
2055 * kmap() and get small overhead in this access function.
2056 */
2057 if (p) {
2058 /*
2059 * we can expect USER0 is not used (see vread/vwrite's
2060 * function description)
2061 */
2062 void *map = kmap_atomic(p);
2063 memcpy(map + offset, buf, length);
2064 kunmap_atomic(map);
2065 }
2066 addr += length;
2067 buf += length;
2068 copied += length;
2069 count -= length;
2070 }
2071 return copied;
2072}
2073
2074/**
2075 * vread() - read vmalloc area in a safe way.
2076 * @buf: buffer for reading data
2077 * @addr: vm address.
2078 * @count: number of bytes to be read.
2079 *
2080 * Returns # of bytes which addr and buf should be increased.
2081 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2082 * includes any intersect with alive vmalloc area.
2083 *
2084 * This function checks that addr is a valid vmalloc'ed area, and
2085 * copy data from that area to a given buffer. If the given memory range
2086 * of [addr...addr+count) includes some valid address, data is copied to
2087 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2088 * IOREMAP area is treated as memory hole and no copy is done.
2089 *
2090 * If [addr...addr+count) doesn't includes any intersects with alive
2091 * vm_struct area, returns 0. @buf should be kernel's buffer.
2092 *
2093 * Note: In usual ops, vread() is never necessary because the caller
2094 * should know vmalloc() area is valid and can use memcpy().
2095 * This is for routines which have to access vmalloc area without
2096 * any informaion, as /dev/kmem.
2097 *
2098 */
2099
2100long vread(char *buf, char *addr, unsigned long count)
2101{
2102 struct vmap_area *va;
2103 struct vm_struct *vm;
2104 char *vaddr, *buf_start = buf;
2105 unsigned long buflen = count;
2106 unsigned long n;
2107
2108 /* Don't allow overflow */
2109 if ((unsigned long) addr + count < count)
2110 count = -(unsigned long) addr;
2111
2112 spin_lock(&vmap_area_lock);
2113 list_for_each_entry(va, &vmap_area_list, list) {
2114 if (!count)
2115 break;
2116
2117 if (!(va->flags & VM_VM_AREA))
2118 continue;
2119
2120 vm = va->vm;
2121 vaddr = (char *) vm->addr;
2122 if (addr >= vaddr + get_vm_area_size(vm))
2123 continue;
2124 while (addr < vaddr) {
2125 if (count == 0)
2126 goto finished;
2127 *buf = '\0';
2128 buf++;
2129 addr++;
2130 count--;
2131 }
2132 n = vaddr + get_vm_area_size(vm) - addr;
2133 if (n > count)
2134 n = count;
2135 if (!(vm->flags & VM_IOREMAP))
2136 aligned_vread(buf, addr, n);
2137 else /* IOREMAP area is treated as memory hole */
2138 memset(buf, 0, n);
2139 buf += n;
2140 addr += n;
2141 count -= n;
2142 }
2143finished:
2144 spin_unlock(&vmap_area_lock);
2145
2146 if (buf == buf_start)
2147 return 0;
2148 /* zero-fill memory holes */
2149 if (buf != buf_start + buflen)
2150 memset(buf, 0, buflen - (buf - buf_start));
2151
2152 return buflen;
2153}
2154
2155/**
2156 * vwrite() - write vmalloc area in a safe way.
2157 * @buf: buffer for source data
2158 * @addr: vm address.
2159 * @count: number of bytes to be read.
2160 *
2161 * Returns # of bytes which addr and buf should be incresed.
2162 * (same number to @count).
2163 * If [addr...addr+count) doesn't includes any intersect with valid
2164 * vmalloc area, returns 0.
2165 *
2166 * This function checks that addr is a valid vmalloc'ed area, and
2167 * copy data from a buffer to the given addr. If specified range of
2168 * [addr...addr+count) includes some valid address, data is copied from
2169 * proper area of @buf. If there are memory holes, no copy to hole.
2170 * IOREMAP area is treated as memory hole and no copy is done.
2171 *
2172 * If [addr...addr+count) doesn't includes any intersects with alive
2173 * vm_struct area, returns 0. @buf should be kernel's buffer.
2174 *
2175 * Note: In usual ops, vwrite() is never necessary because the caller
2176 * should know vmalloc() area is valid and can use memcpy().
2177 * This is for routines which have to access vmalloc area without
2178 * any informaion, as /dev/kmem.
2179 */
2180
2181long vwrite(char *buf, char *addr, unsigned long count)
2182{
2183 struct vmap_area *va;
2184 struct vm_struct *vm;
2185 char *vaddr;
2186 unsigned long n, buflen;
2187 int copied = 0;
2188
2189 /* Don't allow overflow */
2190 if ((unsigned long) addr + count < count)
2191 count = -(unsigned long) addr;
2192 buflen = count;
2193
2194 spin_lock(&vmap_area_lock);
2195 list_for_each_entry(va, &vmap_area_list, list) {
2196 if (!count)
2197 break;
2198
2199 if (!(va->flags & VM_VM_AREA))
2200 continue;
2201
2202 vm = va->vm;
2203 vaddr = (char *) vm->addr;
2204 if (addr >= vaddr + get_vm_area_size(vm))
2205 continue;
2206 while (addr < vaddr) {
2207 if (count == 0)
2208 goto finished;
2209 buf++;
2210 addr++;
2211 count--;
2212 }
2213 n = vaddr + get_vm_area_size(vm) - addr;
2214 if (n > count)
2215 n = count;
2216 if (!(vm->flags & VM_IOREMAP)) {
2217 aligned_vwrite(buf, addr, n);
2218 copied++;
2219 }
2220 buf += n;
2221 addr += n;
2222 count -= n;
2223 }
2224finished:
2225 spin_unlock(&vmap_area_lock);
2226 if (!copied)
2227 return 0;
2228 return buflen;
2229}
2230
2231/**
2232 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2233 * @vma: vma to cover
2234 * @uaddr: target user address to start at
2235 * @kaddr: virtual address of vmalloc kernel memory
2236 * @size: size of map area
2237 *
2238 * Returns: 0 for success, -Exxx on failure
2239 *
2240 * This function checks that @kaddr is a valid vmalloc'ed area,
2241 * and that it is big enough to cover the range starting at
2242 * @uaddr in @vma. Will return failure if that criteria isn't
2243 * met.
2244 *
2245 * Similar to remap_pfn_range() (see mm/memory.c)
2246 */
2247int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2248 void *kaddr, unsigned long size)
2249{
2250 struct vm_struct *area;
2251
2252 size = PAGE_ALIGN(size);
2253
2254 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2255 return -EINVAL;
2256
2257 area = find_vm_area(kaddr);
2258 if (!area)
2259 return -EINVAL;
2260
2261 if (!(area->flags & VM_USERMAP))
2262 return -EINVAL;
2263
2264 if (kaddr + size > area->addr + area->size)
2265 return -EINVAL;
2266
2267 do {
2268 struct page *page = vmalloc_to_page(kaddr);
2269 int ret;
2270
2271 ret = vm_insert_page(vma, uaddr, page);
2272 if (ret)
2273 return ret;
2274
2275 uaddr += PAGE_SIZE;
2276 kaddr += PAGE_SIZE;
2277 size -= PAGE_SIZE;
2278 } while (size > 0);
2279
2280 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2281
2282 return 0;
2283}
2284EXPORT_SYMBOL(remap_vmalloc_range_partial);
2285
2286/**
2287 * remap_vmalloc_range - map vmalloc pages to userspace
2288 * @vma: vma to cover (map full range of vma)
2289 * @addr: vmalloc memory
2290 * @pgoff: number of pages into addr before first page to map
2291 *
2292 * Returns: 0 for success, -Exxx on failure
2293 *
2294 * This function checks that addr is a valid vmalloc'ed area, and
2295 * that it is big enough to cover the vma. Will return failure if
2296 * that criteria isn't met.
2297 *
2298 * Similar to remap_pfn_range() (see mm/memory.c)
2299 */
2300int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2301 unsigned long pgoff)
2302{
2303 return remap_vmalloc_range_partial(vma, vma->vm_start,
2304 addr + (pgoff << PAGE_SHIFT),
2305 vma->vm_end - vma->vm_start);
2306}
2307EXPORT_SYMBOL(remap_vmalloc_range);
2308
2309/*
2310 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2311 * have one.
2312 */
2313void __weak vmalloc_sync_all(void)
2314{
2315}
2316
2317
2318static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2319{
2320 pte_t ***p = data;
2321
2322 if (p) {
2323 *(*p) = pte;
2324 (*p)++;
2325 }
2326 return 0;
2327}
2328
2329/**
2330 * alloc_vm_area - allocate a range of kernel address space
2331 * @size: size of the area
2332 * @ptes: returns the PTEs for the address space
2333 *
2334 * Returns: NULL on failure, vm_struct on success
2335 *
2336 * This function reserves a range of kernel address space, and
2337 * allocates pagetables to map that range. No actual mappings
2338 * are created.
2339 *
2340 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2341 * allocated for the VM area are returned.
2342 */
2343struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2344{
2345 struct vm_struct *area;
2346
2347 area = get_vm_area_caller(size, VM_IOREMAP,
2348 __builtin_return_address(0));
2349 if (area == NULL)
2350 return NULL;
2351
2352 /*
2353 * This ensures that page tables are constructed for this region
2354 * of kernel virtual address space and mapped into init_mm.
2355 */
2356 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2357 size, f, ptes ? &ptes : NULL)) {
2358 free_vm_area(area);
2359 return NULL;
2360 }
2361
2362 return area;
2363}
2364EXPORT_SYMBOL_GPL(alloc_vm_area);
2365
2366void free_vm_area(struct vm_struct *area)
2367{
2368 struct vm_struct *ret;
2369 ret = remove_vm_area(area->addr);
2370 BUG_ON(ret != area);
2371 kfree(area);
2372}
2373EXPORT_SYMBOL_GPL(free_vm_area);
2374
2375#ifdef CONFIG_SMP
2376static struct vmap_area *node_to_va(struct rb_node *n)
2377{
2378 return rb_entry_safe(n, struct vmap_area, rb_node);
2379}
2380
2381/**
2382 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2383 * @end: target address
2384 * @pnext: out arg for the next vmap_area
2385 * @pprev: out arg for the previous vmap_area
2386 *
2387 * Returns: %true if either or both of next and prev are found,
2388 * %false if no vmap_area exists
2389 *
2390 * Find vmap_areas end addresses of which enclose @end. ie. if not
2391 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2392 */
2393static bool pvm_find_next_prev(unsigned long end,
2394 struct vmap_area **pnext,
2395 struct vmap_area **pprev)
2396{
2397 struct rb_node *n = vmap_area_root.rb_node;
2398 struct vmap_area *va = NULL;
2399
2400 while (n) {
2401 va = rb_entry(n, struct vmap_area, rb_node);
2402 if (end < va->va_end)
2403 n = n->rb_left;
2404 else if (end > va->va_end)
2405 n = n->rb_right;
2406 else
2407 break;
2408 }
2409
2410 if (!va)
2411 return false;
2412
2413 if (va->va_end > end) {
2414 *pnext = va;
2415 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2416 } else {
2417 *pprev = va;
2418 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2419 }
2420 return true;
2421}
2422
2423/**
2424 * pvm_determine_end - find the highest aligned address between two vmap_areas
2425 * @pnext: in/out arg for the next vmap_area
2426 * @pprev: in/out arg for the previous vmap_area
2427 * @align: alignment
2428 *
2429 * Returns: determined end address
2430 *
2431 * Find the highest aligned address between *@pnext and *@pprev below
2432 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2433 * down address is between the end addresses of the two vmap_areas.
2434 *
2435 * Please note that the address returned by this function may fall
2436 * inside *@pnext vmap_area. The caller is responsible for checking
2437 * that.
2438 */
2439static unsigned long pvm_determine_end(struct vmap_area **pnext,
2440 struct vmap_area **pprev,
2441 unsigned long align)
2442{
2443 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2444 unsigned long addr;
2445
2446 if (*pnext)
2447 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2448 else
2449 addr = vmalloc_end;
2450
2451 while (*pprev && (*pprev)->va_end > addr) {
2452 *pnext = *pprev;
2453 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2454 }
2455
2456 return addr;
2457}
2458
2459/**
2460 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2461 * @offsets: array containing offset of each area
2462 * @sizes: array containing size of each area
2463 * @nr_vms: the number of areas to allocate
2464 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2465 *
2466 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2467 * vm_structs on success, %NULL on failure
2468 *
2469 * Percpu allocator wants to use congruent vm areas so that it can
2470 * maintain the offsets among percpu areas. This function allocates
2471 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2472 * be scattered pretty far, distance between two areas easily going up
2473 * to gigabytes. To avoid interacting with regular vmallocs, these
2474 * areas are allocated from top.
2475 *
2476 * Despite its complicated look, this allocator is rather simple. It
2477 * does everything top-down and scans areas from the end looking for
2478 * matching slot. While scanning, if any of the areas overlaps with
2479 * existing vmap_area, the base address is pulled down to fit the
2480 * area. Scanning is repeated till all the areas fit and then all
2481 * necessary data structures are inserted and the result is returned.
2482 */
2483struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2484 const size_t *sizes, int nr_vms,
2485 size_t align)
2486{
2487 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2488 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2489 struct vmap_area **vas, *prev, *next;
2490 struct vm_struct **vms;
2491 int area, area2, last_area, term_area;
2492 unsigned long base, start, end, last_end;
2493 bool purged = false;
2494
2495 /* verify parameters and allocate data structures */
2496 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2497 for (last_area = 0, area = 0; area < nr_vms; area++) {
2498 start = offsets[area];
2499 end = start + sizes[area];
2500
2501 /* is everything aligned properly? */
2502 BUG_ON(!IS_ALIGNED(offsets[area], align));
2503 BUG_ON(!IS_ALIGNED(sizes[area], align));
2504
2505 /* detect the area with the highest address */
2506 if (start > offsets[last_area])
2507 last_area = area;
2508
2509 for (area2 = area + 1; area2 < nr_vms; area2++) {
2510 unsigned long start2 = offsets[area2];
2511 unsigned long end2 = start2 + sizes[area2];
2512
2513 BUG_ON(start2 < end && start < end2);
2514 }
2515 }
2516 last_end = offsets[last_area] + sizes[last_area];
2517
2518 if (vmalloc_end - vmalloc_start < last_end) {
2519 WARN_ON(true);
2520 return NULL;
2521 }
2522
2523 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2524 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2525 if (!vas || !vms)
2526 goto err_free2;
2527
2528 for (area = 0; area < nr_vms; area++) {
2529 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2530 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2531 if (!vas[area] || !vms[area])
2532 goto err_free;
2533 }
2534retry:
2535 spin_lock(&vmap_area_lock);
2536
2537 /* start scanning - we scan from the top, begin with the last area */
2538 area = term_area = last_area;
2539 start = offsets[area];
2540 end = start + sizes[area];
2541
2542 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2543 base = vmalloc_end - last_end;
2544 goto found;
2545 }
2546 base = pvm_determine_end(&next, &prev, align) - end;
2547
2548 while (true) {
2549 BUG_ON(next && next->va_end <= base + end);
2550 BUG_ON(prev && prev->va_end > base + end);
2551
2552 /*
2553 * base might have underflowed, add last_end before
2554 * comparing.
2555 */
2556 if (base + last_end < vmalloc_start + last_end) {
2557 spin_unlock(&vmap_area_lock);
2558 if (!purged) {
2559 purge_vmap_area_lazy();
2560 purged = true;
2561 goto retry;
2562 }
2563 goto err_free;
2564 }
2565
2566 /*
2567 * If next overlaps, move base downwards so that it's
2568 * right below next and then recheck.
2569 */
2570 if (next && next->va_start < base + end) {
2571 base = pvm_determine_end(&next, &prev, align) - end;
2572 term_area = area;
2573 continue;
2574 }
2575
2576 /*
2577 * If prev overlaps, shift down next and prev and move
2578 * base so that it's right below new next and then
2579 * recheck.
2580 */
2581 if (prev && prev->va_end > base + start) {
2582 next = prev;
2583 prev = node_to_va(rb_prev(&next->rb_node));
2584 base = pvm_determine_end(&next, &prev, align) - end;
2585 term_area = area;
2586 continue;
2587 }
2588
2589 /*
2590 * This area fits, move on to the previous one. If
2591 * the previous one is the terminal one, we're done.
2592 */
2593 area = (area + nr_vms - 1) % nr_vms;
2594 if (area == term_area)
2595 break;
2596 start = offsets[area];
2597 end = start + sizes[area];
2598 pvm_find_next_prev(base + end, &next, &prev);
2599 }
2600found:
2601 /* we've found a fitting base, insert all va's */
2602 for (area = 0; area < nr_vms; area++) {
2603 struct vmap_area *va = vas[area];
2604
2605 va->va_start = base + offsets[area];
2606 va->va_end = va->va_start + sizes[area];
2607 __insert_vmap_area(va);
2608 }
2609
2610 vmap_area_pcpu_hole = base + offsets[last_area];
2611
2612 spin_unlock(&vmap_area_lock);
2613
2614 /* insert all vm's */
2615 for (area = 0; area < nr_vms; area++)
2616 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2617 pcpu_get_vm_areas);
2618
2619 kfree(vas);
2620 return vms;
2621
2622err_free:
2623 for (area = 0; area < nr_vms; area++) {
2624 kfree(vas[area]);
2625 kfree(vms[area]);
2626 }
2627err_free2:
2628 kfree(vas);
2629 kfree(vms);
2630 return NULL;
2631}
2632
2633/**
2634 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2635 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2636 * @nr_vms: the number of allocated areas
2637 *
2638 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2639 */
2640void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2641{
2642 int i;
2643
2644 for (i = 0; i < nr_vms; i++)
2645 free_vm_area(vms[i]);
2646 kfree(vms);
2647}
2648#endif /* CONFIG_SMP */
2649
2650#ifdef CONFIG_PROC_FS
2651static void *s_start(struct seq_file *m, loff_t *pos)
2652 __acquires(&vmap_area_lock)
2653{
2654 spin_lock(&vmap_area_lock);
2655 return seq_list_start(&vmap_area_list, *pos);
2656}
2657
2658static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2659{
2660 return seq_list_next(p, &vmap_area_list, pos);
2661}
2662
2663static void s_stop(struct seq_file *m, void *p)
2664 __releases(&vmap_area_lock)
2665{
2666 spin_unlock(&vmap_area_lock);
2667}
2668
2669static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2670{
2671 if (IS_ENABLED(CONFIG_NUMA)) {
2672 unsigned int nr, *counters = m->private;
2673
2674 if (!counters)
2675 return;
2676
2677 if (v->flags & VM_UNINITIALIZED)
2678 return;
2679 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2680 smp_rmb();
2681
2682 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2683
2684 for (nr = 0; nr < v->nr_pages; nr++)
2685 counters[page_to_nid(v->pages[nr])]++;
2686
2687 for_each_node_state(nr, N_HIGH_MEMORY)
2688 if (counters[nr])
2689 seq_printf(m, " N%u=%u", nr, counters[nr]);
2690 }
2691}
2692
2693static int s_show(struct seq_file *m, void *p)
2694{
2695 struct vmap_area *va;
2696 struct vm_struct *v;
2697
2698 va = list_entry(p, struct vmap_area, list);
2699
2700 /*
2701 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2702 * behalf of vmap area is being tear down or vm_map_ram allocation.
2703 */
2704 if (!(va->flags & VM_VM_AREA)) {
2705 seq_printf(m, "0x%pK-0x%pK %7ld %s\n",
2706 (void *)va->va_start, (void *)va->va_end,
2707 va->va_end - va->va_start,
2708 va->flags & VM_LAZY_FREE ? "unpurged vm_area" : "vm_map_ram");
2709
2710 return 0;
2711 }
2712
2713 v = va->vm;
2714
2715 seq_printf(m, "0x%pK-0x%pK %7ld",
2716 v->addr, v->addr + v->size, v->size);
2717
2718 if (v->caller)
2719 seq_printf(m, " %pS", v->caller);
2720
2721 if (v->nr_pages)
2722 seq_printf(m, " pages=%d", v->nr_pages);
2723
2724 if (v->phys_addr)
2725 seq_printf(m, " phys=%pa", &v->phys_addr);
2726
2727 if (v->flags & VM_IOREMAP)
2728 seq_puts(m, " ioremap");
2729
2730 if (v->flags & VM_ALLOC)
2731 seq_puts(m, " vmalloc");
2732
2733 if (v->flags & VM_MAP)
2734 seq_puts(m, " vmap");
2735
2736 if (v->flags & VM_USERMAP)
2737 seq_puts(m, " user");
2738
2739 if (is_vmalloc_addr(v->pages))
2740 seq_puts(m, " vpages");
2741
2742 show_numa_info(m, v);
2743 seq_putc(m, '\n');
2744 return 0;
2745}
2746
2747static const struct seq_operations vmalloc_op = {
2748 .start = s_start,
2749 .next = s_next,
2750 .stop = s_stop,
2751 .show = s_show,
2752};
2753
2754static int vmalloc_open(struct inode *inode, struct file *file)
2755{
2756 if (IS_ENABLED(CONFIG_NUMA))
2757 return seq_open_private(file, &vmalloc_op,
2758 nr_node_ids * sizeof(unsigned int));
2759 else
2760 return seq_open(file, &vmalloc_op);
2761}
2762
2763static const struct file_operations proc_vmalloc_operations = {
2764 .open = vmalloc_open,
2765 .read = seq_read,
2766 .llseek = seq_lseek,
2767 .release = seq_release_private,
2768};
2769
2770static int __init proc_vmalloc_init(void)
2771{
2772 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2773 return 0;
2774}
2775module_init(proc_vmalloc_init);
2776
2777#endif
2778
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmalloc.c
4 *
5 * Copyright (C) 1993 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
8 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
9 * Numa awareness, Christoph Lameter, SGI, June 2005
10 */
11
12#include <linux/vmalloc.h>
13#include <linux/mm.h>
14#include <linux/module.h>
15#include <linux/highmem.h>
16#include <linux/sched/signal.h>
17#include <linux/slab.h>
18#include <linux/spinlock.h>
19#include <linux/interrupt.h>
20#include <linux/proc_fs.h>
21#include <linux/seq_file.h>
22#include <linux/set_memory.h>
23#include <linux/debugobjects.h>
24#include <linux/kallsyms.h>
25#include <linux/list.h>
26#include <linux/notifier.h>
27#include <linux/rbtree.h>
28#include <linux/radix-tree.h>
29#include <linux/rcupdate.h>
30#include <linux/pfn.h>
31#include <linux/kmemleak.h>
32#include <linux/atomic.h>
33#include <linux/compiler.h>
34#include <linux/llist.h>
35#include <linux/bitops.h>
36#include <linux/rbtree_augmented.h>
37
38#include <linux/uaccess.h>
39#include <asm/tlbflush.h>
40#include <asm/shmparam.h>
41
42#include "internal.h"
43
44struct vfree_deferred {
45 struct llist_head list;
46 struct work_struct wq;
47};
48static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
49
50static void __vunmap(const void *, int);
51
52static void free_work(struct work_struct *w)
53{
54 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
55 struct llist_node *t, *llnode;
56
57 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
58 __vunmap((void *)llnode, 1);
59}
60
61/*** Page table manipulation functions ***/
62
63static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
64{
65 pte_t *pte;
66
67 pte = pte_offset_kernel(pmd, addr);
68 do {
69 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
70 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
71 } while (pte++, addr += PAGE_SIZE, addr != end);
72}
73
74static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
75{
76 pmd_t *pmd;
77 unsigned long next;
78
79 pmd = pmd_offset(pud, addr);
80 do {
81 next = pmd_addr_end(addr, end);
82 if (pmd_clear_huge(pmd))
83 continue;
84 if (pmd_none_or_clear_bad(pmd))
85 continue;
86 vunmap_pte_range(pmd, addr, next);
87 } while (pmd++, addr = next, addr != end);
88}
89
90static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
91{
92 pud_t *pud;
93 unsigned long next;
94
95 pud = pud_offset(p4d, addr);
96 do {
97 next = pud_addr_end(addr, end);
98 if (pud_clear_huge(pud))
99 continue;
100 if (pud_none_or_clear_bad(pud))
101 continue;
102 vunmap_pmd_range(pud, addr, next);
103 } while (pud++, addr = next, addr != end);
104}
105
106static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
107{
108 p4d_t *p4d;
109 unsigned long next;
110
111 p4d = p4d_offset(pgd, addr);
112 do {
113 next = p4d_addr_end(addr, end);
114 if (p4d_clear_huge(p4d))
115 continue;
116 if (p4d_none_or_clear_bad(p4d))
117 continue;
118 vunmap_pud_range(p4d, addr, next);
119 } while (p4d++, addr = next, addr != end);
120}
121
122static void vunmap_page_range(unsigned long addr, unsigned long end)
123{
124 pgd_t *pgd;
125 unsigned long next;
126
127 BUG_ON(addr >= end);
128 pgd = pgd_offset_k(addr);
129 do {
130 next = pgd_addr_end(addr, end);
131 if (pgd_none_or_clear_bad(pgd))
132 continue;
133 vunmap_p4d_range(pgd, addr, next);
134 } while (pgd++, addr = next, addr != end);
135}
136
137static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
138 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
139{
140 pte_t *pte;
141
142 /*
143 * nr is a running index into the array which helps higher level
144 * callers keep track of where we're up to.
145 */
146
147 pte = pte_alloc_kernel(pmd, addr);
148 if (!pte)
149 return -ENOMEM;
150 do {
151 struct page *page = pages[*nr];
152
153 if (WARN_ON(!pte_none(*pte)))
154 return -EBUSY;
155 if (WARN_ON(!page))
156 return -ENOMEM;
157 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
158 (*nr)++;
159 } while (pte++, addr += PAGE_SIZE, addr != end);
160 return 0;
161}
162
163static int vmap_pmd_range(pud_t *pud, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
165{
166 pmd_t *pmd;
167 unsigned long next;
168
169 pmd = pmd_alloc(&init_mm, pud, addr);
170 if (!pmd)
171 return -ENOMEM;
172 do {
173 next = pmd_addr_end(addr, end);
174 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
175 return -ENOMEM;
176 } while (pmd++, addr = next, addr != end);
177 return 0;
178}
179
180static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
181 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
182{
183 pud_t *pud;
184 unsigned long next;
185
186 pud = pud_alloc(&init_mm, p4d, addr);
187 if (!pud)
188 return -ENOMEM;
189 do {
190 next = pud_addr_end(addr, end);
191 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
192 return -ENOMEM;
193 } while (pud++, addr = next, addr != end);
194 return 0;
195}
196
197static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
198 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
199{
200 p4d_t *p4d;
201 unsigned long next;
202
203 p4d = p4d_alloc(&init_mm, pgd, addr);
204 if (!p4d)
205 return -ENOMEM;
206 do {
207 next = p4d_addr_end(addr, end);
208 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
209 return -ENOMEM;
210 } while (p4d++, addr = next, addr != end);
211 return 0;
212}
213
214/*
215 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
216 * will have pfns corresponding to the "pages" array.
217 *
218 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
219 */
220static int vmap_page_range_noflush(unsigned long start, unsigned long end,
221 pgprot_t prot, struct page **pages)
222{
223 pgd_t *pgd;
224 unsigned long next;
225 unsigned long addr = start;
226 int err = 0;
227 int nr = 0;
228
229 BUG_ON(addr >= end);
230 pgd = pgd_offset_k(addr);
231 do {
232 next = pgd_addr_end(addr, end);
233 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
234 if (err)
235 return err;
236 } while (pgd++, addr = next, addr != end);
237
238 return nr;
239}
240
241static int vmap_page_range(unsigned long start, unsigned long end,
242 pgprot_t prot, struct page **pages)
243{
244 int ret;
245
246 ret = vmap_page_range_noflush(start, end, prot, pages);
247 flush_cache_vmap(start, end);
248 return ret;
249}
250
251int is_vmalloc_or_module_addr(const void *x)
252{
253 /*
254 * ARM, x86-64 and sparc64 put modules in a special place,
255 * and fall back on vmalloc() if that fails. Others
256 * just put it in the vmalloc space.
257 */
258#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
259 unsigned long addr = (unsigned long)x;
260 if (addr >= MODULES_VADDR && addr < MODULES_END)
261 return 1;
262#endif
263 return is_vmalloc_addr(x);
264}
265
266/*
267 * Walk a vmap address to the struct page it maps.
268 */
269struct page *vmalloc_to_page(const void *vmalloc_addr)
270{
271 unsigned long addr = (unsigned long) vmalloc_addr;
272 struct page *page = NULL;
273 pgd_t *pgd = pgd_offset_k(addr);
274 p4d_t *p4d;
275 pud_t *pud;
276 pmd_t *pmd;
277 pte_t *ptep, pte;
278
279 /*
280 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
281 * architectures that do not vmalloc module space
282 */
283 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
284
285 if (pgd_none(*pgd))
286 return NULL;
287 p4d = p4d_offset(pgd, addr);
288 if (p4d_none(*p4d))
289 return NULL;
290 pud = pud_offset(p4d, addr);
291
292 /*
293 * Don't dereference bad PUD or PMD (below) entries. This will also
294 * identify huge mappings, which we may encounter on architectures
295 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
296 * identified as vmalloc addresses by is_vmalloc_addr(), but are
297 * not [unambiguously] associated with a struct page, so there is
298 * no correct value to return for them.
299 */
300 WARN_ON_ONCE(pud_bad(*pud));
301 if (pud_none(*pud) || pud_bad(*pud))
302 return NULL;
303 pmd = pmd_offset(pud, addr);
304 WARN_ON_ONCE(pmd_bad(*pmd));
305 if (pmd_none(*pmd) || pmd_bad(*pmd))
306 return NULL;
307
308 ptep = pte_offset_map(pmd, addr);
309 pte = *ptep;
310 if (pte_present(pte))
311 page = pte_page(pte);
312 pte_unmap(ptep);
313 return page;
314}
315EXPORT_SYMBOL(vmalloc_to_page);
316
317/*
318 * Map a vmalloc()-space virtual address to the physical page frame number.
319 */
320unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
321{
322 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
323}
324EXPORT_SYMBOL(vmalloc_to_pfn);
325
326
327/*** Global kva allocator ***/
328
329#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
330#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
331
332
333static DEFINE_SPINLOCK(vmap_area_lock);
334/* Export for kexec only */
335LIST_HEAD(vmap_area_list);
336static LLIST_HEAD(vmap_purge_list);
337static struct rb_root vmap_area_root = RB_ROOT;
338static bool vmap_initialized __read_mostly;
339
340/*
341 * This kmem_cache is used for vmap_area objects. Instead of
342 * allocating from slab we reuse an object from this cache to
343 * make things faster. Especially in "no edge" splitting of
344 * free block.
345 */
346static struct kmem_cache *vmap_area_cachep;
347
348/*
349 * This linked list is used in pair with free_vmap_area_root.
350 * It gives O(1) access to prev/next to perform fast coalescing.
351 */
352static LIST_HEAD(free_vmap_area_list);
353
354/*
355 * This augment red-black tree represents the free vmap space.
356 * All vmap_area objects in this tree are sorted by va->va_start
357 * address. It is used for allocation and merging when a vmap
358 * object is released.
359 *
360 * Each vmap_area node contains a maximum available free block
361 * of its sub-tree, right or left. Therefore it is possible to
362 * find a lowest match of free area.
363 */
364static struct rb_root free_vmap_area_root = RB_ROOT;
365
366/*
367 * Preload a CPU with one object for "no edge" split case. The
368 * aim is to get rid of allocations from the atomic context, thus
369 * to use more permissive allocation masks.
370 */
371static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
372
373static __always_inline unsigned long
374va_size(struct vmap_area *va)
375{
376 return (va->va_end - va->va_start);
377}
378
379static __always_inline unsigned long
380get_subtree_max_size(struct rb_node *node)
381{
382 struct vmap_area *va;
383
384 va = rb_entry_safe(node, struct vmap_area, rb_node);
385 return va ? va->subtree_max_size : 0;
386}
387
388/*
389 * Gets called when remove the node and rotate.
390 */
391static __always_inline unsigned long
392compute_subtree_max_size(struct vmap_area *va)
393{
394 return max3(va_size(va),
395 get_subtree_max_size(va->rb_node.rb_left),
396 get_subtree_max_size(va->rb_node.rb_right));
397}
398
399RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
400 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
401
402static void purge_vmap_area_lazy(void);
403static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
404static unsigned long lazy_max_pages(void);
405
406static atomic_long_t nr_vmalloc_pages;
407
408unsigned long vmalloc_nr_pages(void)
409{
410 return atomic_long_read(&nr_vmalloc_pages);
411}
412
413static struct vmap_area *__find_vmap_area(unsigned long addr)
414{
415 struct rb_node *n = vmap_area_root.rb_node;
416
417 while (n) {
418 struct vmap_area *va;
419
420 va = rb_entry(n, struct vmap_area, rb_node);
421 if (addr < va->va_start)
422 n = n->rb_left;
423 else if (addr >= va->va_end)
424 n = n->rb_right;
425 else
426 return va;
427 }
428
429 return NULL;
430}
431
432/*
433 * This function returns back addresses of parent node
434 * and its left or right link for further processing.
435 */
436static __always_inline struct rb_node **
437find_va_links(struct vmap_area *va,
438 struct rb_root *root, struct rb_node *from,
439 struct rb_node **parent)
440{
441 struct vmap_area *tmp_va;
442 struct rb_node **link;
443
444 if (root) {
445 link = &root->rb_node;
446 if (unlikely(!*link)) {
447 *parent = NULL;
448 return link;
449 }
450 } else {
451 link = &from;
452 }
453
454 /*
455 * Go to the bottom of the tree. When we hit the last point
456 * we end up with parent rb_node and correct direction, i name
457 * it link, where the new va->rb_node will be attached to.
458 */
459 do {
460 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
461
462 /*
463 * During the traversal we also do some sanity check.
464 * Trigger the BUG() if there are sides(left/right)
465 * or full overlaps.
466 */
467 if (va->va_start < tmp_va->va_end &&
468 va->va_end <= tmp_va->va_start)
469 link = &(*link)->rb_left;
470 else if (va->va_end > tmp_va->va_start &&
471 va->va_start >= tmp_va->va_end)
472 link = &(*link)->rb_right;
473 else
474 BUG();
475 } while (*link);
476
477 *parent = &tmp_va->rb_node;
478 return link;
479}
480
481static __always_inline struct list_head *
482get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
483{
484 struct list_head *list;
485
486 if (unlikely(!parent))
487 /*
488 * The red-black tree where we try to find VA neighbors
489 * before merging or inserting is empty, i.e. it means
490 * there is no free vmap space. Normally it does not
491 * happen but we handle this case anyway.
492 */
493 return NULL;
494
495 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
496 return (&parent->rb_right == link ? list->next : list);
497}
498
499static __always_inline void
500link_va(struct vmap_area *va, struct rb_root *root,
501 struct rb_node *parent, struct rb_node **link, struct list_head *head)
502{
503 /*
504 * VA is still not in the list, but we can
505 * identify its future previous list_head node.
506 */
507 if (likely(parent)) {
508 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
509 if (&parent->rb_right != link)
510 head = head->prev;
511 }
512
513 /* Insert to the rb-tree */
514 rb_link_node(&va->rb_node, parent, link);
515 if (root == &free_vmap_area_root) {
516 /*
517 * Some explanation here. Just perform simple insertion
518 * to the tree. We do not set va->subtree_max_size to
519 * its current size before calling rb_insert_augmented().
520 * It is because of we populate the tree from the bottom
521 * to parent levels when the node _is_ in the tree.
522 *
523 * Therefore we set subtree_max_size to zero after insertion,
524 * to let __augment_tree_propagate_from() puts everything to
525 * the correct order later on.
526 */
527 rb_insert_augmented(&va->rb_node,
528 root, &free_vmap_area_rb_augment_cb);
529 va->subtree_max_size = 0;
530 } else {
531 rb_insert_color(&va->rb_node, root);
532 }
533
534 /* Address-sort this list */
535 list_add(&va->list, head);
536}
537
538static __always_inline void
539unlink_va(struct vmap_area *va, struct rb_root *root)
540{
541 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
542 return;
543
544 if (root == &free_vmap_area_root)
545 rb_erase_augmented(&va->rb_node,
546 root, &free_vmap_area_rb_augment_cb);
547 else
548 rb_erase(&va->rb_node, root);
549
550 list_del(&va->list);
551 RB_CLEAR_NODE(&va->rb_node);
552}
553
554#if DEBUG_AUGMENT_PROPAGATE_CHECK
555static void
556augment_tree_propagate_check(struct rb_node *n)
557{
558 struct vmap_area *va;
559 struct rb_node *node;
560 unsigned long size;
561 bool found = false;
562
563 if (n == NULL)
564 return;
565
566 va = rb_entry(n, struct vmap_area, rb_node);
567 size = va->subtree_max_size;
568 node = n;
569
570 while (node) {
571 va = rb_entry(node, struct vmap_area, rb_node);
572
573 if (get_subtree_max_size(node->rb_left) == size) {
574 node = node->rb_left;
575 } else {
576 if (va_size(va) == size) {
577 found = true;
578 break;
579 }
580
581 node = node->rb_right;
582 }
583 }
584
585 if (!found) {
586 va = rb_entry(n, struct vmap_area, rb_node);
587 pr_emerg("tree is corrupted: %lu, %lu\n",
588 va_size(va), va->subtree_max_size);
589 }
590
591 augment_tree_propagate_check(n->rb_left);
592 augment_tree_propagate_check(n->rb_right);
593}
594#endif
595
596/*
597 * This function populates subtree_max_size from bottom to upper
598 * levels starting from VA point. The propagation must be done
599 * when VA size is modified by changing its va_start/va_end. Or
600 * in case of newly inserting of VA to the tree.
601 *
602 * It means that __augment_tree_propagate_from() must be called:
603 * - After VA has been inserted to the tree(free path);
604 * - After VA has been shrunk(allocation path);
605 * - After VA has been increased(merging path).
606 *
607 * Please note that, it does not mean that upper parent nodes
608 * and their subtree_max_size are recalculated all the time up
609 * to the root node.
610 *
611 * 4--8
612 * /\
613 * / \
614 * / \
615 * 2--2 8--8
616 *
617 * For example if we modify the node 4, shrinking it to 2, then
618 * no any modification is required. If we shrink the node 2 to 1
619 * its subtree_max_size is updated only, and set to 1. If we shrink
620 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
621 * node becomes 4--6.
622 */
623static __always_inline void
624augment_tree_propagate_from(struct vmap_area *va)
625{
626 struct rb_node *node = &va->rb_node;
627 unsigned long new_va_sub_max_size;
628
629 while (node) {
630 va = rb_entry(node, struct vmap_area, rb_node);
631 new_va_sub_max_size = compute_subtree_max_size(va);
632
633 /*
634 * If the newly calculated maximum available size of the
635 * subtree is equal to the current one, then it means that
636 * the tree is propagated correctly. So we have to stop at
637 * this point to save cycles.
638 */
639 if (va->subtree_max_size == new_va_sub_max_size)
640 break;
641
642 va->subtree_max_size = new_va_sub_max_size;
643 node = rb_parent(&va->rb_node);
644 }
645
646#if DEBUG_AUGMENT_PROPAGATE_CHECK
647 augment_tree_propagate_check(free_vmap_area_root.rb_node);
648#endif
649}
650
651static void
652insert_vmap_area(struct vmap_area *va,
653 struct rb_root *root, struct list_head *head)
654{
655 struct rb_node **link;
656 struct rb_node *parent;
657
658 link = find_va_links(va, root, NULL, &parent);
659 link_va(va, root, parent, link, head);
660}
661
662static void
663insert_vmap_area_augment(struct vmap_area *va,
664 struct rb_node *from, struct rb_root *root,
665 struct list_head *head)
666{
667 struct rb_node **link;
668 struct rb_node *parent;
669
670 if (from)
671 link = find_va_links(va, NULL, from, &parent);
672 else
673 link = find_va_links(va, root, NULL, &parent);
674
675 link_va(va, root, parent, link, head);
676 augment_tree_propagate_from(va);
677}
678
679/*
680 * Merge de-allocated chunk of VA memory with previous
681 * and next free blocks. If coalesce is not done a new
682 * free area is inserted. If VA has been merged, it is
683 * freed.
684 */
685static __always_inline void
686merge_or_add_vmap_area(struct vmap_area *va,
687 struct rb_root *root, struct list_head *head)
688{
689 struct vmap_area *sibling;
690 struct list_head *next;
691 struct rb_node **link;
692 struct rb_node *parent;
693 bool merged = false;
694
695 /*
696 * Find a place in the tree where VA potentially will be
697 * inserted, unless it is merged with its sibling/siblings.
698 */
699 link = find_va_links(va, root, NULL, &parent);
700
701 /*
702 * Get next node of VA to check if merging can be done.
703 */
704 next = get_va_next_sibling(parent, link);
705 if (unlikely(next == NULL))
706 goto insert;
707
708 /*
709 * start end
710 * | |
711 * |<------VA------>|<-----Next----->|
712 * | |
713 * start end
714 */
715 if (next != head) {
716 sibling = list_entry(next, struct vmap_area, list);
717 if (sibling->va_start == va->va_end) {
718 sibling->va_start = va->va_start;
719
720 /* Check and update the tree if needed. */
721 augment_tree_propagate_from(sibling);
722
723 /* Free vmap_area object. */
724 kmem_cache_free(vmap_area_cachep, va);
725
726 /* Point to the new merged area. */
727 va = sibling;
728 merged = true;
729 }
730 }
731
732 /*
733 * start end
734 * | |
735 * |<-----Prev----->|<------VA------>|
736 * | |
737 * start end
738 */
739 if (next->prev != head) {
740 sibling = list_entry(next->prev, struct vmap_area, list);
741 if (sibling->va_end == va->va_start) {
742 sibling->va_end = va->va_end;
743
744 /* Check and update the tree if needed. */
745 augment_tree_propagate_from(sibling);
746
747 if (merged)
748 unlink_va(va, root);
749
750 /* Free vmap_area object. */
751 kmem_cache_free(vmap_area_cachep, va);
752 return;
753 }
754 }
755
756insert:
757 if (!merged) {
758 link_va(va, root, parent, link, head);
759 augment_tree_propagate_from(va);
760 }
761}
762
763static __always_inline bool
764is_within_this_va(struct vmap_area *va, unsigned long size,
765 unsigned long align, unsigned long vstart)
766{
767 unsigned long nva_start_addr;
768
769 if (va->va_start > vstart)
770 nva_start_addr = ALIGN(va->va_start, align);
771 else
772 nva_start_addr = ALIGN(vstart, align);
773
774 /* Can be overflowed due to big size or alignment. */
775 if (nva_start_addr + size < nva_start_addr ||
776 nva_start_addr < vstart)
777 return false;
778
779 return (nva_start_addr + size <= va->va_end);
780}
781
782/*
783 * Find the first free block(lowest start address) in the tree,
784 * that will accomplish the request corresponding to passing
785 * parameters.
786 */
787static __always_inline struct vmap_area *
788find_vmap_lowest_match(unsigned long size,
789 unsigned long align, unsigned long vstart)
790{
791 struct vmap_area *va;
792 struct rb_node *node;
793 unsigned long length;
794
795 /* Start from the root. */
796 node = free_vmap_area_root.rb_node;
797
798 /* Adjust the search size for alignment overhead. */
799 length = size + align - 1;
800
801 while (node) {
802 va = rb_entry(node, struct vmap_area, rb_node);
803
804 if (get_subtree_max_size(node->rb_left) >= length &&
805 vstart < va->va_start) {
806 node = node->rb_left;
807 } else {
808 if (is_within_this_va(va, size, align, vstart))
809 return va;
810
811 /*
812 * Does not make sense to go deeper towards the right
813 * sub-tree if it does not have a free block that is
814 * equal or bigger to the requested search length.
815 */
816 if (get_subtree_max_size(node->rb_right) >= length) {
817 node = node->rb_right;
818 continue;
819 }
820
821 /*
822 * OK. We roll back and find the first right sub-tree,
823 * that will satisfy the search criteria. It can happen
824 * only once due to "vstart" restriction.
825 */
826 while ((node = rb_parent(node))) {
827 va = rb_entry(node, struct vmap_area, rb_node);
828 if (is_within_this_va(va, size, align, vstart))
829 return va;
830
831 if (get_subtree_max_size(node->rb_right) >= length &&
832 vstart <= va->va_start) {
833 node = node->rb_right;
834 break;
835 }
836 }
837 }
838 }
839
840 return NULL;
841}
842
843#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
844#include <linux/random.h>
845
846static struct vmap_area *
847find_vmap_lowest_linear_match(unsigned long size,
848 unsigned long align, unsigned long vstart)
849{
850 struct vmap_area *va;
851
852 list_for_each_entry(va, &free_vmap_area_list, list) {
853 if (!is_within_this_va(va, size, align, vstart))
854 continue;
855
856 return va;
857 }
858
859 return NULL;
860}
861
862static void
863find_vmap_lowest_match_check(unsigned long size)
864{
865 struct vmap_area *va_1, *va_2;
866 unsigned long vstart;
867 unsigned int rnd;
868
869 get_random_bytes(&rnd, sizeof(rnd));
870 vstart = VMALLOC_START + rnd;
871
872 va_1 = find_vmap_lowest_match(size, 1, vstart);
873 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
874
875 if (va_1 != va_2)
876 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
877 va_1, va_2, vstart);
878}
879#endif
880
881enum fit_type {
882 NOTHING_FIT = 0,
883 FL_FIT_TYPE = 1, /* full fit */
884 LE_FIT_TYPE = 2, /* left edge fit */
885 RE_FIT_TYPE = 3, /* right edge fit */
886 NE_FIT_TYPE = 4 /* no edge fit */
887};
888
889static __always_inline enum fit_type
890classify_va_fit_type(struct vmap_area *va,
891 unsigned long nva_start_addr, unsigned long size)
892{
893 enum fit_type type;
894
895 /* Check if it is within VA. */
896 if (nva_start_addr < va->va_start ||
897 nva_start_addr + size > va->va_end)
898 return NOTHING_FIT;
899
900 /* Now classify. */
901 if (va->va_start == nva_start_addr) {
902 if (va->va_end == nva_start_addr + size)
903 type = FL_FIT_TYPE;
904 else
905 type = LE_FIT_TYPE;
906 } else if (va->va_end == nva_start_addr + size) {
907 type = RE_FIT_TYPE;
908 } else {
909 type = NE_FIT_TYPE;
910 }
911
912 return type;
913}
914
915static __always_inline int
916adjust_va_to_fit_type(struct vmap_area *va,
917 unsigned long nva_start_addr, unsigned long size,
918 enum fit_type type)
919{
920 struct vmap_area *lva = NULL;
921
922 if (type == FL_FIT_TYPE) {
923 /*
924 * No need to split VA, it fully fits.
925 *
926 * | |
927 * V NVA V
928 * |---------------|
929 */
930 unlink_va(va, &free_vmap_area_root);
931 kmem_cache_free(vmap_area_cachep, va);
932 } else if (type == LE_FIT_TYPE) {
933 /*
934 * Split left edge of fit VA.
935 *
936 * | |
937 * V NVA V R
938 * |-------|-------|
939 */
940 va->va_start += size;
941 } else if (type == RE_FIT_TYPE) {
942 /*
943 * Split right edge of fit VA.
944 *
945 * | |
946 * L V NVA V
947 * |-------|-------|
948 */
949 va->va_end = nva_start_addr;
950 } else if (type == NE_FIT_TYPE) {
951 /*
952 * Split no edge of fit VA.
953 *
954 * | |
955 * L V NVA V R
956 * |---|-------|---|
957 */
958 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
959 if (unlikely(!lva)) {
960 /*
961 * For percpu allocator we do not do any pre-allocation
962 * and leave it as it is. The reason is it most likely
963 * never ends up with NE_FIT_TYPE splitting. In case of
964 * percpu allocations offsets and sizes are aligned to
965 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
966 * are its main fitting cases.
967 *
968 * There are a few exceptions though, as an example it is
969 * a first allocation (early boot up) when we have "one"
970 * big free space that has to be split.
971 */
972 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
973 if (!lva)
974 return -1;
975 }
976
977 /*
978 * Build the remainder.
979 */
980 lva->va_start = va->va_start;
981 lva->va_end = nva_start_addr;
982
983 /*
984 * Shrink this VA to remaining size.
985 */
986 va->va_start = nva_start_addr + size;
987 } else {
988 return -1;
989 }
990
991 if (type != FL_FIT_TYPE) {
992 augment_tree_propagate_from(va);
993
994 if (lva) /* type == NE_FIT_TYPE */
995 insert_vmap_area_augment(lva, &va->rb_node,
996 &free_vmap_area_root, &free_vmap_area_list);
997 }
998
999 return 0;
1000}
1001
1002/*
1003 * Returns a start address of the newly allocated area, if success.
1004 * Otherwise a vend is returned that indicates failure.
1005 */
1006static __always_inline unsigned long
1007__alloc_vmap_area(unsigned long size, unsigned long align,
1008 unsigned long vstart, unsigned long vend)
1009{
1010 unsigned long nva_start_addr;
1011 struct vmap_area *va;
1012 enum fit_type type;
1013 int ret;
1014
1015 va = find_vmap_lowest_match(size, align, vstart);
1016 if (unlikely(!va))
1017 return vend;
1018
1019 if (va->va_start > vstart)
1020 nva_start_addr = ALIGN(va->va_start, align);
1021 else
1022 nva_start_addr = ALIGN(vstart, align);
1023
1024 /* Check the "vend" restriction. */
1025 if (nva_start_addr + size > vend)
1026 return vend;
1027
1028 /* Classify what we have found. */
1029 type = classify_va_fit_type(va, nva_start_addr, size);
1030 if (WARN_ON_ONCE(type == NOTHING_FIT))
1031 return vend;
1032
1033 /* Update the free vmap_area. */
1034 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1035 if (ret)
1036 return vend;
1037
1038#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1039 find_vmap_lowest_match_check(size);
1040#endif
1041
1042 return nva_start_addr;
1043}
1044
1045/*
1046 * Allocate a region of KVA of the specified size and alignment, within the
1047 * vstart and vend.
1048 */
1049static struct vmap_area *alloc_vmap_area(unsigned long size,
1050 unsigned long align,
1051 unsigned long vstart, unsigned long vend,
1052 int node, gfp_t gfp_mask)
1053{
1054 struct vmap_area *va, *pva;
1055 unsigned long addr;
1056 int purged = 0;
1057
1058 BUG_ON(!size);
1059 BUG_ON(offset_in_page(size));
1060 BUG_ON(!is_power_of_2(align));
1061
1062 if (unlikely(!vmap_initialized))
1063 return ERR_PTR(-EBUSY);
1064
1065 might_sleep();
1066
1067 va = kmem_cache_alloc_node(vmap_area_cachep,
1068 gfp_mask & GFP_RECLAIM_MASK, node);
1069 if (unlikely(!va))
1070 return ERR_PTR(-ENOMEM);
1071
1072 /*
1073 * Only scan the relevant parts containing pointers to other objects
1074 * to avoid false negatives.
1075 */
1076 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
1077
1078retry:
1079 /*
1080 * Preload this CPU with one extra vmap_area object to ensure
1081 * that we have it available when fit type of free area is
1082 * NE_FIT_TYPE.
1083 *
1084 * The preload is done in non-atomic context, thus it allows us
1085 * to use more permissive allocation masks to be more stable under
1086 * low memory condition and high memory pressure.
1087 *
1088 * Even if it fails we do not really care about that. Just proceed
1089 * as it is. "overflow" path will refill the cache we allocate from.
1090 */
1091 preempt_disable();
1092 if (!__this_cpu_read(ne_fit_preload_node)) {
1093 preempt_enable();
1094 pva = kmem_cache_alloc_node(vmap_area_cachep, GFP_KERNEL, node);
1095 preempt_disable();
1096
1097 if (__this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva)) {
1098 if (pva)
1099 kmem_cache_free(vmap_area_cachep, pva);
1100 }
1101 }
1102
1103 spin_lock(&vmap_area_lock);
1104 preempt_enable();
1105
1106 /*
1107 * If an allocation fails, the "vend" address is
1108 * returned. Therefore trigger the overflow path.
1109 */
1110 addr = __alloc_vmap_area(size, align, vstart, vend);
1111 if (unlikely(addr == vend))
1112 goto overflow;
1113
1114 va->va_start = addr;
1115 va->va_end = addr + size;
1116 va->vm = NULL;
1117 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1118
1119 spin_unlock(&vmap_area_lock);
1120
1121 BUG_ON(!IS_ALIGNED(va->va_start, align));
1122 BUG_ON(va->va_start < vstart);
1123 BUG_ON(va->va_end > vend);
1124
1125 return va;
1126
1127overflow:
1128 spin_unlock(&vmap_area_lock);
1129 if (!purged) {
1130 purge_vmap_area_lazy();
1131 purged = 1;
1132 goto retry;
1133 }
1134
1135 if (gfpflags_allow_blocking(gfp_mask)) {
1136 unsigned long freed = 0;
1137 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1138 if (freed > 0) {
1139 purged = 0;
1140 goto retry;
1141 }
1142 }
1143
1144 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1145 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1146 size);
1147
1148 kmem_cache_free(vmap_area_cachep, va);
1149 return ERR_PTR(-EBUSY);
1150}
1151
1152int register_vmap_purge_notifier(struct notifier_block *nb)
1153{
1154 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1155}
1156EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1157
1158int unregister_vmap_purge_notifier(struct notifier_block *nb)
1159{
1160 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1161}
1162EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1163
1164static void __free_vmap_area(struct vmap_area *va)
1165{
1166 /*
1167 * Remove from the busy tree/list.
1168 */
1169 unlink_va(va, &vmap_area_root);
1170
1171 /*
1172 * Merge VA with its neighbors, otherwise just add it.
1173 */
1174 merge_or_add_vmap_area(va,
1175 &free_vmap_area_root, &free_vmap_area_list);
1176}
1177
1178/*
1179 * Free a region of KVA allocated by alloc_vmap_area
1180 */
1181static void free_vmap_area(struct vmap_area *va)
1182{
1183 spin_lock(&vmap_area_lock);
1184 __free_vmap_area(va);
1185 spin_unlock(&vmap_area_lock);
1186}
1187
1188/*
1189 * Clear the pagetable entries of a given vmap_area
1190 */
1191static void unmap_vmap_area(struct vmap_area *va)
1192{
1193 vunmap_page_range(va->va_start, va->va_end);
1194}
1195
1196/*
1197 * lazy_max_pages is the maximum amount of virtual address space we gather up
1198 * before attempting to purge with a TLB flush.
1199 *
1200 * There is a tradeoff here: a larger number will cover more kernel page tables
1201 * and take slightly longer to purge, but it will linearly reduce the number of
1202 * global TLB flushes that must be performed. It would seem natural to scale
1203 * this number up linearly with the number of CPUs (because vmapping activity
1204 * could also scale linearly with the number of CPUs), however it is likely
1205 * that in practice, workloads might be constrained in other ways that mean
1206 * vmap activity will not scale linearly with CPUs. Also, I want to be
1207 * conservative and not introduce a big latency on huge systems, so go with
1208 * a less aggressive log scale. It will still be an improvement over the old
1209 * code, and it will be simple to change the scale factor if we find that it
1210 * becomes a problem on bigger systems.
1211 */
1212static unsigned long lazy_max_pages(void)
1213{
1214 unsigned int log;
1215
1216 log = fls(num_online_cpus());
1217
1218 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1219}
1220
1221static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1222
1223/*
1224 * Serialize vmap purging. There is no actual criticial section protected
1225 * by this look, but we want to avoid concurrent calls for performance
1226 * reasons and to make the pcpu_get_vm_areas more deterministic.
1227 */
1228static DEFINE_MUTEX(vmap_purge_lock);
1229
1230/* for per-CPU blocks */
1231static void purge_fragmented_blocks_allcpus(void);
1232
1233/*
1234 * called before a call to iounmap() if the caller wants vm_area_struct's
1235 * immediately freed.
1236 */
1237void set_iounmap_nonlazy(void)
1238{
1239 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1240}
1241
1242/*
1243 * Purges all lazily-freed vmap areas.
1244 */
1245static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1246{
1247 unsigned long resched_threshold;
1248 struct llist_node *valist;
1249 struct vmap_area *va;
1250 struct vmap_area *n_va;
1251
1252 lockdep_assert_held(&vmap_purge_lock);
1253
1254 valist = llist_del_all(&vmap_purge_list);
1255 if (unlikely(valist == NULL))
1256 return false;
1257
1258 /*
1259 * First make sure the mappings are removed from all page-tables
1260 * before they are freed.
1261 */
1262 vmalloc_sync_all();
1263
1264 /*
1265 * TODO: to calculate a flush range without looping.
1266 * The list can be up to lazy_max_pages() elements.
1267 */
1268 llist_for_each_entry(va, valist, purge_list) {
1269 if (va->va_start < start)
1270 start = va->va_start;
1271 if (va->va_end > end)
1272 end = va->va_end;
1273 }
1274
1275 flush_tlb_kernel_range(start, end);
1276 resched_threshold = lazy_max_pages() << 1;
1277
1278 spin_lock(&vmap_area_lock);
1279 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1280 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1281
1282 /*
1283 * Finally insert or merge lazily-freed area. It is
1284 * detached and there is no need to "unlink" it from
1285 * anything.
1286 */
1287 merge_or_add_vmap_area(va,
1288 &free_vmap_area_root, &free_vmap_area_list);
1289
1290 atomic_long_sub(nr, &vmap_lazy_nr);
1291
1292 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1293 cond_resched_lock(&vmap_area_lock);
1294 }
1295 spin_unlock(&vmap_area_lock);
1296 return true;
1297}
1298
1299/*
1300 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1301 * is already purging.
1302 */
1303static void try_purge_vmap_area_lazy(void)
1304{
1305 if (mutex_trylock(&vmap_purge_lock)) {
1306 __purge_vmap_area_lazy(ULONG_MAX, 0);
1307 mutex_unlock(&vmap_purge_lock);
1308 }
1309}
1310
1311/*
1312 * Kick off a purge of the outstanding lazy areas.
1313 */
1314static void purge_vmap_area_lazy(void)
1315{
1316 mutex_lock(&vmap_purge_lock);
1317 purge_fragmented_blocks_allcpus();
1318 __purge_vmap_area_lazy(ULONG_MAX, 0);
1319 mutex_unlock(&vmap_purge_lock);
1320}
1321
1322/*
1323 * Free a vmap area, caller ensuring that the area has been unmapped
1324 * and flush_cache_vunmap had been called for the correct range
1325 * previously.
1326 */
1327static void free_vmap_area_noflush(struct vmap_area *va)
1328{
1329 unsigned long nr_lazy;
1330
1331 spin_lock(&vmap_area_lock);
1332 unlink_va(va, &vmap_area_root);
1333 spin_unlock(&vmap_area_lock);
1334
1335 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1336 PAGE_SHIFT, &vmap_lazy_nr);
1337
1338 /* After this point, we may free va at any time */
1339 llist_add(&va->purge_list, &vmap_purge_list);
1340
1341 if (unlikely(nr_lazy > lazy_max_pages()))
1342 try_purge_vmap_area_lazy();
1343}
1344
1345/*
1346 * Free and unmap a vmap area
1347 */
1348static void free_unmap_vmap_area(struct vmap_area *va)
1349{
1350 flush_cache_vunmap(va->va_start, va->va_end);
1351 unmap_vmap_area(va);
1352 if (debug_pagealloc_enabled())
1353 flush_tlb_kernel_range(va->va_start, va->va_end);
1354
1355 free_vmap_area_noflush(va);
1356}
1357
1358static struct vmap_area *find_vmap_area(unsigned long addr)
1359{
1360 struct vmap_area *va;
1361
1362 spin_lock(&vmap_area_lock);
1363 va = __find_vmap_area(addr);
1364 spin_unlock(&vmap_area_lock);
1365
1366 return va;
1367}
1368
1369/*** Per cpu kva allocator ***/
1370
1371/*
1372 * vmap space is limited especially on 32 bit architectures. Ensure there is
1373 * room for at least 16 percpu vmap blocks per CPU.
1374 */
1375/*
1376 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1377 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1378 * instead (we just need a rough idea)
1379 */
1380#if BITS_PER_LONG == 32
1381#define VMALLOC_SPACE (128UL*1024*1024)
1382#else
1383#define VMALLOC_SPACE (128UL*1024*1024*1024)
1384#endif
1385
1386#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1387#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1388#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1389#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1390#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1391#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1392#define VMAP_BBMAP_BITS \
1393 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1394 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1395 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1396
1397#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1398
1399struct vmap_block_queue {
1400 spinlock_t lock;
1401 struct list_head free;
1402};
1403
1404struct vmap_block {
1405 spinlock_t lock;
1406 struct vmap_area *va;
1407 unsigned long free, dirty;
1408 unsigned long dirty_min, dirty_max; /*< dirty range */
1409 struct list_head free_list;
1410 struct rcu_head rcu_head;
1411 struct list_head purge;
1412};
1413
1414/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1415static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1416
1417/*
1418 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1419 * in the free path. Could get rid of this if we change the API to return a
1420 * "cookie" from alloc, to be passed to free. But no big deal yet.
1421 */
1422static DEFINE_SPINLOCK(vmap_block_tree_lock);
1423static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
1424
1425/*
1426 * We should probably have a fallback mechanism to allocate virtual memory
1427 * out of partially filled vmap blocks. However vmap block sizing should be
1428 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1429 * big problem.
1430 */
1431
1432static unsigned long addr_to_vb_idx(unsigned long addr)
1433{
1434 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1435 addr /= VMAP_BLOCK_SIZE;
1436 return addr;
1437}
1438
1439static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1440{
1441 unsigned long addr;
1442
1443 addr = va_start + (pages_off << PAGE_SHIFT);
1444 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1445 return (void *)addr;
1446}
1447
1448/**
1449 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1450 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1451 * @order: how many 2^order pages should be occupied in newly allocated block
1452 * @gfp_mask: flags for the page level allocator
1453 *
1454 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1455 */
1456static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1457{
1458 struct vmap_block_queue *vbq;
1459 struct vmap_block *vb;
1460 struct vmap_area *va;
1461 unsigned long vb_idx;
1462 int node, err;
1463 void *vaddr;
1464
1465 node = numa_node_id();
1466
1467 vb = kmalloc_node(sizeof(struct vmap_block),
1468 gfp_mask & GFP_RECLAIM_MASK, node);
1469 if (unlikely(!vb))
1470 return ERR_PTR(-ENOMEM);
1471
1472 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1473 VMALLOC_START, VMALLOC_END,
1474 node, gfp_mask);
1475 if (IS_ERR(va)) {
1476 kfree(vb);
1477 return ERR_CAST(va);
1478 }
1479
1480 err = radix_tree_preload(gfp_mask);
1481 if (unlikely(err)) {
1482 kfree(vb);
1483 free_vmap_area(va);
1484 return ERR_PTR(err);
1485 }
1486
1487 vaddr = vmap_block_vaddr(va->va_start, 0);
1488 spin_lock_init(&vb->lock);
1489 vb->va = va;
1490 /* At least something should be left free */
1491 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1492 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1493 vb->dirty = 0;
1494 vb->dirty_min = VMAP_BBMAP_BITS;
1495 vb->dirty_max = 0;
1496 INIT_LIST_HEAD(&vb->free_list);
1497
1498 vb_idx = addr_to_vb_idx(va->va_start);
1499 spin_lock(&vmap_block_tree_lock);
1500 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
1501 spin_unlock(&vmap_block_tree_lock);
1502 BUG_ON(err);
1503 radix_tree_preload_end();
1504
1505 vbq = &get_cpu_var(vmap_block_queue);
1506 spin_lock(&vbq->lock);
1507 list_add_tail_rcu(&vb->free_list, &vbq->free);
1508 spin_unlock(&vbq->lock);
1509 put_cpu_var(vmap_block_queue);
1510
1511 return vaddr;
1512}
1513
1514static void free_vmap_block(struct vmap_block *vb)
1515{
1516 struct vmap_block *tmp;
1517 unsigned long vb_idx;
1518
1519 vb_idx = addr_to_vb_idx(vb->va->va_start);
1520 spin_lock(&vmap_block_tree_lock);
1521 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
1522 spin_unlock(&vmap_block_tree_lock);
1523 BUG_ON(tmp != vb);
1524
1525 free_vmap_area_noflush(vb->va);
1526 kfree_rcu(vb, rcu_head);
1527}
1528
1529static void purge_fragmented_blocks(int cpu)
1530{
1531 LIST_HEAD(purge);
1532 struct vmap_block *vb;
1533 struct vmap_block *n_vb;
1534 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1535
1536 rcu_read_lock();
1537 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1538
1539 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1540 continue;
1541
1542 spin_lock(&vb->lock);
1543 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1544 vb->free = 0; /* prevent further allocs after releasing lock */
1545 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1546 vb->dirty_min = 0;
1547 vb->dirty_max = VMAP_BBMAP_BITS;
1548 spin_lock(&vbq->lock);
1549 list_del_rcu(&vb->free_list);
1550 spin_unlock(&vbq->lock);
1551 spin_unlock(&vb->lock);
1552 list_add_tail(&vb->purge, &purge);
1553 } else
1554 spin_unlock(&vb->lock);
1555 }
1556 rcu_read_unlock();
1557
1558 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1559 list_del(&vb->purge);
1560 free_vmap_block(vb);
1561 }
1562}
1563
1564static void purge_fragmented_blocks_allcpus(void)
1565{
1566 int cpu;
1567
1568 for_each_possible_cpu(cpu)
1569 purge_fragmented_blocks(cpu);
1570}
1571
1572static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1573{
1574 struct vmap_block_queue *vbq;
1575 struct vmap_block *vb;
1576 void *vaddr = NULL;
1577 unsigned int order;
1578
1579 BUG_ON(offset_in_page(size));
1580 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1581 if (WARN_ON(size == 0)) {
1582 /*
1583 * Allocating 0 bytes isn't what caller wants since
1584 * get_order(0) returns funny result. Just warn and terminate
1585 * early.
1586 */
1587 return NULL;
1588 }
1589 order = get_order(size);
1590
1591 rcu_read_lock();
1592 vbq = &get_cpu_var(vmap_block_queue);
1593 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1594 unsigned long pages_off;
1595
1596 spin_lock(&vb->lock);
1597 if (vb->free < (1UL << order)) {
1598 spin_unlock(&vb->lock);
1599 continue;
1600 }
1601
1602 pages_off = VMAP_BBMAP_BITS - vb->free;
1603 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1604 vb->free -= 1UL << order;
1605 if (vb->free == 0) {
1606 spin_lock(&vbq->lock);
1607 list_del_rcu(&vb->free_list);
1608 spin_unlock(&vbq->lock);
1609 }
1610
1611 spin_unlock(&vb->lock);
1612 break;
1613 }
1614
1615 put_cpu_var(vmap_block_queue);
1616 rcu_read_unlock();
1617
1618 /* Allocate new block if nothing was found */
1619 if (!vaddr)
1620 vaddr = new_vmap_block(order, gfp_mask);
1621
1622 return vaddr;
1623}
1624
1625static void vb_free(const void *addr, unsigned long size)
1626{
1627 unsigned long offset;
1628 unsigned long vb_idx;
1629 unsigned int order;
1630 struct vmap_block *vb;
1631
1632 BUG_ON(offset_in_page(size));
1633 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1634
1635 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1636
1637 order = get_order(size);
1638
1639 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1640 offset >>= PAGE_SHIFT;
1641
1642 vb_idx = addr_to_vb_idx((unsigned long)addr);
1643 rcu_read_lock();
1644 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1645 rcu_read_unlock();
1646 BUG_ON(!vb);
1647
1648 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1649
1650 if (debug_pagealloc_enabled())
1651 flush_tlb_kernel_range((unsigned long)addr,
1652 (unsigned long)addr + size);
1653
1654 spin_lock(&vb->lock);
1655
1656 /* Expand dirty range */
1657 vb->dirty_min = min(vb->dirty_min, offset);
1658 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1659
1660 vb->dirty += 1UL << order;
1661 if (vb->dirty == VMAP_BBMAP_BITS) {
1662 BUG_ON(vb->free);
1663 spin_unlock(&vb->lock);
1664 free_vmap_block(vb);
1665 } else
1666 spin_unlock(&vb->lock);
1667}
1668
1669static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1670{
1671 int cpu;
1672
1673 if (unlikely(!vmap_initialized))
1674 return;
1675
1676 might_sleep();
1677
1678 for_each_possible_cpu(cpu) {
1679 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1680 struct vmap_block *vb;
1681
1682 rcu_read_lock();
1683 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1684 spin_lock(&vb->lock);
1685 if (vb->dirty) {
1686 unsigned long va_start = vb->va->va_start;
1687 unsigned long s, e;
1688
1689 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1690 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1691
1692 start = min(s, start);
1693 end = max(e, end);
1694
1695 flush = 1;
1696 }
1697 spin_unlock(&vb->lock);
1698 }
1699 rcu_read_unlock();
1700 }
1701
1702 mutex_lock(&vmap_purge_lock);
1703 purge_fragmented_blocks_allcpus();
1704 if (!__purge_vmap_area_lazy(start, end) && flush)
1705 flush_tlb_kernel_range(start, end);
1706 mutex_unlock(&vmap_purge_lock);
1707}
1708
1709/**
1710 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1711 *
1712 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1713 * to amortize TLB flushing overheads. What this means is that any page you
1714 * have now, may, in a former life, have been mapped into kernel virtual
1715 * address by the vmap layer and so there might be some CPUs with TLB entries
1716 * still referencing that page (additional to the regular 1:1 kernel mapping).
1717 *
1718 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1719 * be sure that none of the pages we have control over will have any aliases
1720 * from the vmap layer.
1721 */
1722void vm_unmap_aliases(void)
1723{
1724 unsigned long start = ULONG_MAX, end = 0;
1725 int flush = 0;
1726
1727 _vm_unmap_aliases(start, end, flush);
1728}
1729EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1730
1731/**
1732 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1733 * @mem: the pointer returned by vm_map_ram
1734 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1735 */
1736void vm_unmap_ram(const void *mem, unsigned int count)
1737{
1738 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1739 unsigned long addr = (unsigned long)mem;
1740 struct vmap_area *va;
1741
1742 might_sleep();
1743 BUG_ON(!addr);
1744 BUG_ON(addr < VMALLOC_START);
1745 BUG_ON(addr > VMALLOC_END);
1746 BUG_ON(!PAGE_ALIGNED(addr));
1747
1748 if (likely(count <= VMAP_MAX_ALLOC)) {
1749 debug_check_no_locks_freed(mem, size);
1750 vb_free(mem, size);
1751 return;
1752 }
1753
1754 va = find_vmap_area(addr);
1755 BUG_ON(!va);
1756 debug_check_no_locks_freed((void *)va->va_start,
1757 (va->va_end - va->va_start));
1758 free_unmap_vmap_area(va);
1759}
1760EXPORT_SYMBOL(vm_unmap_ram);
1761
1762/**
1763 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1764 * @pages: an array of pointers to the pages to be mapped
1765 * @count: number of pages
1766 * @node: prefer to allocate data structures on this node
1767 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1768 *
1769 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1770 * faster than vmap so it's good. But if you mix long-life and short-life
1771 * objects with vm_map_ram(), it could consume lots of address space through
1772 * fragmentation (especially on a 32bit machine). You could see failures in
1773 * the end. Please use this function for short-lived objects.
1774 *
1775 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1776 */
1777void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1778{
1779 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1780 unsigned long addr;
1781 void *mem;
1782
1783 if (likely(count <= VMAP_MAX_ALLOC)) {
1784 mem = vb_alloc(size, GFP_KERNEL);
1785 if (IS_ERR(mem))
1786 return NULL;
1787 addr = (unsigned long)mem;
1788 } else {
1789 struct vmap_area *va;
1790 va = alloc_vmap_area(size, PAGE_SIZE,
1791 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1792 if (IS_ERR(va))
1793 return NULL;
1794
1795 addr = va->va_start;
1796 mem = (void *)addr;
1797 }
1798 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1799 vm_unmap_ram(mem, count);
1800 return NULL;
1801 }
1802 return mem;
1803}
1804EXPORT_SYMBOL(vm_map_ram);
1805
1806static struct vm_struct *vmlist __initdata;
1807
1808/**
1809 * vm_area_add_early - add vmap area early during boot
1810 * @vm: vm_struct to add
1811 *
1812 * This function is used to add fixed kernel vm area to vmlist before
1813 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1814 * should contain proper values and the other fields should be zero.
1815 *
1816 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1817 */
1818void __init vm_area_add_early(struct vm_struct *vm)
1819{
1820 struct vm_struct *tmp, **p;
1821
1822 BUG_ON(vmap_initialized);
1823 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1824 if (tmp->addr >= vm->addr) {
1825 BUG_ON(tmp->addr < vm->addr + vm->size);
1826 break;
1827 } else
1828 BUG_ON(tmp->addr + tmp->size > vm->addr);
1829 }
1830 vm->next = *p;
1831 *p = vm;
1832}
1833
1834/**
1835 * vm_area_register_early - register vmap area early during boot
1836 * @vm: vm_struct to register
1837 * @align: requested alignment
1838 *
1839 * This function is used to register kernel vm area before
1840 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1841 * proper values on entry and other fields should be zero. On return,
1842 * vm->addr contains the allocated address.
1843 *
1844 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1845 */
1846void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1847{
1848 static size_t vm_init_off __initdata;
1849 unsigned long addr;
1850
1851 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1852 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1853
1854 vm->addr = (void *)addr;
1855
1856 vm_area_add_early(vm);
1857}
1858
1859static void vmap_init_free_space(void)
1860{
1861 unsigned long vmap_start = 1;
1862 const unsigned long vmap_end = ULONG_MAX;
1863 struct vmap_area *busy, *free;
1864
1865 /*
1866 * B F B B B F
1867 * -|-----|.....|-----|-----|-----|.....|-
1868 * | The KVA space |
1869 * |<--------------------------------->|
1870 */
1871 list_for_each_entry(busy, &vmap_area_list, list) {
1872 if (busy->va_start - vmap_start > 0) {
1873 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1874 if (!WARN_ON_ONCE(!free)) {
1875 free->va_start = vmap_start;
1876 free->va_end = busy->va_start;
1877
1878 insert_vmap_area_augment(free, NULL,
1879 &free_vmap_area_root,
1880 &free_vmap_area_list);
1881 }
1882 }
1883
1884 vmap_start = busy->va_end;
1885 }
1886
1887 if (vmap_end - vmap_start > 0) {
1888 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1889 if (!WARN_ON_ONCE(!free)) {
1890 free->va_start = vmap_start;
1891 free->va_end = vmap_end;
1892
1893 insert_vmap_area_augment(free, NULL,
1894 &free_vmap_area_root,
1895 &free_vmap_area_list);
1896 }
1897 }
1898}
1899
1900void __init vmalloc_init(void)
1901{
1902 struct vmap_area *va;
1903 struct vm_struct *tmp;
1904 int i;
1905
1906 /*
1907 * Create the cache for vmap_area objects.
1908 */
1909 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1910
1911 for_each_possible_cpu(i) {
1912 struct vmap_block_queue *vbq;
1913 struct vfree_deferred *p;
1914
1915 vbq = &per_cpu(vmap_block_queue, i);
1916 spin_lock_init(&vbq->lock);
1917 INIT_LIST_HEAD(&vbq->free);
1918 p = &per_cpu(vfree_deferred, i);
1919 init_llist_head(&p->list);
1920 INIT_WORK(&p->wq, free_work);
1921 }
1922
1923 /* Import existing vmlist entries. */
1924 for (tmp = vmlist; tmp; tmp = tmp->next) {
1925 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1926 if (WARN_ON_ONCE(!va))
1927 continue;
1928
1929 va->va_start = (unsigned long)tmp->addr;
1930 va->va_end = va->va_start + tmp->size;
1931 va->vm = tmp;
1932 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1933 }
1934
1935 /*
1936 * Now we can initialize a free vmap space.
1937 */
1938 vmap_init_free_space();
1939 vmap_initialized = true;
1940}
1941
1942/**
1943 * map_kernel_range_noflush - map kernel VM area with the specified pages
1944 * @addr: start of the VM area to map
1945 * @size: size of the VM area to map
1946 * @prot: page protection flags to use
1947 * @pages: pages to map
1948 *
1949 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1950 * specify should have been allocated using get_vm_area() and its
1951 * friends.
1952 *
1953 * NOTE:
1954 * This function does NOT do any cache flushing. The caller is
1955 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1956 * before calling this function.
1957 *
1958 * RETURNS:
1959 * The number of pages mapped on success, -errno on failure.
1960 */
1961int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1962 pgprot_t prot, struct page **pages)
1963{
1964 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1965}
1966
1967/**
1968 * unmap_kernel_range_noflush - unmap kernel VM area
1969 * @addr: start of the VM area to unmap
1970 * @size: size of the VM area to unmap
1971 *
1972 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1973 * specify should have been allocated using get_vm_area() and its
1974 * friends.
1975 *
1976 * NOTE:
1977 * This function does NOT do any cache flushing. The caller is
1978 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1979 * before calling this function and flush_tlb_kernel_range() after.
1980 */
1981void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1982{
1983 vunmap_page_range(addr, addr + size);
1984}
1985EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1986
1987/**
1988 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1989 * @addr: start of the VM area to unmap
1990 * @size: size of the VM area to unmap
1991 *
1992 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1993 * the unmapping and tlb after.
1994 */
1995void unmap_kernel_range(unsigned long addr, unsigned long size)
1996{
1997 unsigned long end = addr + size;
1998
1999 flush_cache_vunmap(addr, end);
2000 vunmap_page_range(addr, end);
2001 flush_tlb_kernel_range(addr, end);
2002}
2003EXPORT_SYMBOL_GPL(unmap_kernel_range);
2004
2005int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
2006{
2007 unsigned long addr = (unsigned long)area->addr;
2008 unsigned long end = addr + get_vm_area_size(area);
2009 int err;
2010
2011 err = vmap_page_range(addr, end, prot, pages);
2012
2013 return err > 0 ? 0 : err;
2014}
2015EXPORT_SYMBOL_GPL(map_vm_area);
2016
2017static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2018 unsigned long flags, const void *caller)
2019{
2020 spin_lock(&vmap_area_lock);
2021 vm->flags = flags;
2022 vm->addr = (void *)va->va_start;
2023 vm->size = va->va_end - va->va_start;
2024 vm->caller = caller;
2025 va->vm = vm;
2026 spin_unlock(&vmap_area_lock);
2027}
2028
2029static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2030{
2031 /*
2032 * Before removing VM_UNINITIALIZED,
2033 * we should make sure that vm has proper values.
2034 * Pair with smp_rmb() in show_numa_info().
2035 */
2036 smp_wmb();
2037 vm->flags &= ~VM_UNINITIALIZED;
2038}
2039
2040static struct vm_struct *__get_vm_area_node(unsigned long size,
2041 unsigned long align, unsigned long flags, unsigned long start,
2042 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2043{
2044 struct vmap_area *va;
2045 struct vm_struct *area;
2046
2047 BUG_ON(in_interrupt());
2048 size = PAGE_ALIGN(size);
2049 if (unlikely(!size))
2050 return NULL;
2051
2052 if (flags & VM_IOREMAP)
2053 align = 1ul << clamp_t(int, get_count_order_long(size),
2054 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2055
2056 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2057 if (unlikely(!area))
2058 return NULL;
2059
2060 if (!(flags & VM_NO_GUARD))
2061 size += PAGE_SIZE;
2062
2063 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2064 if (IS_ERR(va)) {
2065 kfree(area);
2066 return NULL;
2067 }
2068
2069 setup_vmalloc_vm(area, va, flags, caller);
2070
2071 return area;
2072}
2073
2074struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
2075 unsigned long start, unsigned long end)
2076{
2077 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2078 GFP_KERNEL, __builtin_return_address(0));
2079}
2080EXPORT_SYMBOL_GPL(__get_vm_area);
2081
2082struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2083 unsigned long start, unsigned long end,
2084 const void *caller)
2085{
2086 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2087 GFP_KERNEL, caller);
2088}
2089
2090/**
2091 * get_vm_area - reserve a contiguous kernel virtual area
2092 * @size: size of the area
2093 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2094 *
2095 * Search an area of @size in the kernel virtual mapping area,
2096 * and reserved it for out purposes. Returns the area descriptor
2097 * on success or %NULL on failure.
2098 *
2099 * Return: the area descriptor on success or %NULL on failure.
2100 */
2101struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2102{
2103 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2104 NUMA_NO_NODE, GFP_KERNEL,
2105 __builtin_return_address(0));
2106}
2107
2108struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2109 const void *caller)
2110{
2111 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2112 NUMA_NO_NODE, GFP_KERNEL, caller);
2113}
2114
2115/**
2116 * find_vm_area - find a continuous kernel virtual area
2117 * @addr: base address
2118 *
2119 * Search for the kernel VM area starting at @addr, and return it.
2120 * It is up to the caller to do all required locking to keep the returned
2121 * pointer valid.
2122 *
2123 * Return: pointer to the found area or %NULL on faulure
2124 */
2125struct vm_struct *find_vm_area(const void *addr)
2126{
2127 struct vmap_area *va;
2128
2129 va = find_vmap_area((unsigned long)addr);
2130 if (!va)
2131 return NULL;
2132
2133 return va->vm;
2134}
2135
2136/**
2137 * remove_vm_area - find and remove a continuous kernel virtual area
2138 * @addr: base address
2139 *
2140 * Search for the kernel VM area starting at @addr, and remove it.
2141 * This function returns the found VM area, but using it is NOT safe
2142 * on SMP machines, except for its size or flags.
2143 *
2144 * Return: pointer to the found area or %NULL on faulure
2145 */
2146struct vm_struct *remove_vm_area(const void *addr)
2147{
2148 struct vmap_area *va;
2149
2150 might_sleep();
2151
2152 spin_lock(&vmap_area_lock);
2153 va = __find_vmap_area((unsigned long)addr);
2154 if (va && va->vm) {
2155 struct vm_struct *vm = va->vm;
2156
2157 va->vm = NULL;
2158 spin_unlock(&vmap_area_lock);
2159
2160 kasan_free_shadow(vm);
2161 free_unmap_vmap_area(va);
2162
2163 return vm;
2164 }
2165
2166 spin_unlock(&vmap_area_lock);
2167 return NULL;
2168}
2169
2170static inline void set_area_direct_map(const struct vm_struct *area,
2171 int (*set_direct_map)(struct page *page))
2172{
2173 int i;
2174
2175 for (i = 0; i < area->nr_pages; i++)
2176 if (page_address(area->pages[i]))
2177 set_direct_map(area->pages[i]);
2178}
2179
2180/* Handle removing and resetting vm mappings related to the vm_struct. */
2181static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2182{
2183 unsigned long start = ULONG_MAX, end = 0;
2184 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2185 int flush_dmap = 0;
2186 int i;
2187
2188 remove_vm_area(area->addr);
2189
2190 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2191 if (!flush_reset)
2192 return;
2193
2194 /*
2195 * If not deallocating pages, just do the flush of the VM area and
2196 * return.
2197 */
2198 if (!deallocate_pages) {
2199 vm_unmap_aliases();
2200 return;
2201 }
2202
2203 /*
2204 * If execution gets here, flush the vm mapping and reset the direct
2205 * map. Find the start and end range of the direct mappings to make sure
2206 * the vm_unmap_aliases() flush includes the direct map.
2207 */
2208 for (i = 0; i < area->nr_pages; i++) {
2209 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2210 if (addr) {
2211 start = min(addr, start);
2212 end = max(addr + PAGE_SIZE, end);
2213 flush_dmap = 1;
2214 }
2215 }
2216
2217 /*
2218 * Set direct map to something invalid so that it won't be cached if
2219 * there are any accesses after the TLB flush, then flush the TLB and
2220 * reset the direct map permissions to the default.
2221 */
2222 set_area_direct_map(area, set_direct_map_invalid_noflush);
2223 _vm_unmap_aliases(start, end, flush_dmap);
2224 set_area_direct_map(area, set_direct_map_default_noflush);
2225}
2226
2227static void __vunmap(const void *addr, int deallocate_pages)
2228{
2229 struct vm_struct *area;
2230
2231 if (!addr)
2232 return;
2233
2234 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2235 addr))
2236 return;
2237
2238 area = find_vm_area(addr);
2239 if (unlikely(!area)) {
2240 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2241 addr);
2242 return;
2243 }
2244
2245 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2246 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2247
2248 vm_remove_mappings(area, deallocate_pages);
2249
2250 if (deallocate_pages) {
2251 int i;
2252
2253 for (i = 0; i < area->nr_pages; i++) {
2254 struct page *page = area->pages[i];
2255
2256 BUG_ON(!page);
2257 __free_pages(page, 0);
2258 }
2259 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2260
2261 kvfree(area->pages);
2262 }
2263
2264 kfree(area);
2265 return;
2266}
2267
2268static inline void __vfree_deferred(const void *addr)
2269{
2270 /*
2271 * Use raw_cpu_ptr() because this can be called from preemptible
2272 * context. Preemption is absolutely fine here, because the llist_add()
2273 * implementation is lockless, so it works even if we are adding to
2274 * nother cpu's list. schedule_work() should be fine with this too.
2275 */
2276 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2277
2278 if (llist_add((struct llist_node *)addr, &p->list))
2279 schedule_work(&p->wq);
2280}
2281
2282/**
2283 * vfree_atomic - release memory allocated by vmalloc()
2284 * @addr: memory base address
2285 *
2286 * This one is just like vfree() but can be called in any atomic context
2287 * except NMIs.
2288 */
2289void vfree_atomic(const void *addr)
2290{
2291 BUG_ON(in_nmi());
2292
2293 kmemleak_free(addr);
2294
2295 if (!addr)
2296 return;
2297 __vfree_deferred(addr);
2298}
2299
2300static void __vfree(const void *addr)
2301{
2302 if (unlikely(in_interrupt()))
2303 __vfree_deferred(addr);
2304 else
2305 __vunmap(addr, 1);
2306}
2307
2308/**
2309 * vfree - release memory allocated by vmalloc()
2310 * @addr: memory base address
2311 *
2312 * Free the virtually continuous memory area starting at @addr, as
2313 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2314 * NULL, no operation is performed.
2315 *
2316 * Must not be called in NMI context (strictly speaking, only if we don't
2317 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2318 * conventions for vfree() arch-depenedent would be a really bad idea)
2319 *
2320 * May sleep if called *not* from interrupt context.
2321 *
2322 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2323 */
2324void vfree(const void *addr)
2325{
2326 BUG_ON(in_nmi());
2327
2328 kmemleak_free(addr);
2329
2330 might_sleep_if(!in_interrupt());
2331
2332 if (!addr)
2333 return;
2334
2335 __vfree(addr);
2336}
2337EXPORT_SYMBOL(vfree);
2338
2339/**
2340 * vunmap - release virtual mapping obtained by vmap()
2341 * @addr: memory base address
2342 *
2343 * Free the virtually contiguous memory area starting at @addr,
2344 * which was created from the page array passed to vmap().
2345 *
2346 * Must not be called in interrupt context.
2347 */
2348void vunmap(const void *addr)
2349{
2350 BUG_ON(in_interrupt());
2351 might_sleep();
2352 if (addr)
2353 __vunmap(addr, 0);
2354}
2355EXPORT_SYMBOL(vunmap);
2356
2357/**
2358 * vmap - map an array of pages into virtually contiguous space
2359 * @pages: array of page pointers
2360 * @count: number of pages to map
2361 * @flags: vm_area->flags
2362 * @prot: page protection for the mapping
2363 *
2364 * Maps @count pages from @pages into contiguous kernel virtual
2365 * space.
2366 *
2367 * Return: the address of the area or %NULL on failure
2368 */
2369void *vmap(struct page **pages, unsigned int count,
2370 unsigned long flags, pgprot_t prot)
2371{
2372 struct vm_struct *area;
2373 unsigned long size; /* In bytes */
2374
2375 might_sleep();
2376
2377 if (count > totalram_pages())
2378 return NULL;
2379
2380 size = (unsigned long)count << PAGE_SHIFT;
2381 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2382 if (!area)
2383 return NULL;
2384
2385 if (map_vm_area(area, prot, pages)) {
2386 vunmap(area->addr);
2387 return NULL;
2388 }
2389
2390 return area->addr;
2391}
2392EXPORT_SYMBOL(vmap);
2393
2394static void *__vmalloc_node(unsigned long size, unsigned long align,
2395 gfp_t gfp_mask, pgprot_t prot,
2396 int node, const void *caller);
2397static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2398 pgprot_t prot, int node)
2399{
2400 struct page **pages;
2401 unsigned int nr_pages, array_size, i;
2402 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2403 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2404 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2405 0 :
2406 __GFP_HIGHMEM;
2407
2408 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2409 array_size = (nr_pages * sizeof(struct page *));
2410
2411 /* Please note that the recursion is strictly bounded. */
2412 if (array_size > PAGE_SIZE) {
2413 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2414 PAGE_KERNEL, node, area->caller);
2415 } else {
2416 pages = kmalloc_node(array_size, nested_gfp, node);
2417 }
2418
2419 if (!pages) {
2420 remove_vm_area(area->addr);
2421 kfree(area);
2422 return NULL;
2423 }
2424
2425 area->pages = pages;
2426 area->nr_pages = nr_pages;
2427
2428 for (i = 0; i < area->nr_pages; i++) {
2429 struct page *page;
2430
2431 if (node == NUMA_NO_NODE)
2432 page = alloc_page(alloc_mask|highmem_mask);
2433 else
2434 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2435
2436 if (unlikely(!page)) {
2437 /* Successfully allocated i pages, free them in __vunmap() */
2438 area->nr_pages = i;
2439 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2440 goto fail;
2441 }
2442 area->pages[i] = page;
2443 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
2444 cond_resched();
2445 }
2446 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2447
2448 if (map_vm_area(area, prot, pages))
2449 goto fail;
2450 return area->addr;
2451
2452fail:
2453 warn_alloc(gfp_mask, NULL,
2454 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2455 (area->nr_pages*PAGE_SIZE), area->size);
2456 __vfree(area->addr);
2457 return NULL;
2458}
2459
2460/**
2461 * __vmalloc_node_range - allocate virtually contiguous memory
2462 * @size: allocation size
2463 * @align: desired alignment
2464 * @start: vm area range start
2465 * @end: vm area range end
2466 * @gfp_mask: flags for the page level allocator
2467 * @prot: protection mask for the allocated pages
2468 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2469 * @node: node to use for allocation or NUMA_NO_NODE
2470 * @caller: caller's return address
2471 *
2472 * Allocate enough pages to cover @size from the page level
2473 * allocator with @gfp_mask flags. Map them into contiguous
2474 * kernel virtual space, using a pagetable protection of @prot.
2475 *
2476 * Return: the address of the area or %NULL on failure
2477 */
2478void *__vmalloc_node_range(unsigned long size, unsigned long align,
2479 unsigned long start, unsigned long end, gfp_t gfp_mask,
2480 pgprot_t prot, unsigned long vm_flags, int node,
2481 const void *caller)
2482{
2483 struct vm_struct *area;
2484 void *addr;
2485 unsigned long real_size = size;
2486
2487 size = PAGE_ALIGN(size);
2488 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2489 goto fail;
2490
2491 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
2492 vm_flags, start, end, node, gfp_mask, caller);
2493 if (!area)
2494 goto fail;
2495
2496 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2497 if (!addr)
2498 return NULL;
2499
2500 /*
2501 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2502 * flag. It means that vm_struct is not fully initialized.
2503 * Now, it is fully initialized, so remove this flag here.
2504 */
2505 clear_vm_uninitialized_flag(area);
2506
2507 kmemleak_vmalloc(area, size, gfp_mask);
2508
2509 return addr;
2510
2511fail:
2512 warn_alloc(gfp_mask, NULL,
2513 "vmalloc: allocation failure: %lu bytes", real_size);
2514 return NULL;
2515}
2516
2517/*
2518 * This is only for performance analysis of vmalloc and stress purpose.
2519 * It is required by vmalloc test module, therefore do not use it other
2520 * than that.
2521 */
2522#ifdef CONFIG_TEST_VMALLOC_MODULE
2523EXPORT_SYMBOL_GPL(__vmalloc_node_range);
2524#endif
2525
2526/**
2527 * __vmalloc_node - allocate virtually contiguous memory
2528 * @size: allocation size
2529 * @align: desired alignment
2530 * @gfp_mask: flags for the page level allocator
2531 * @prot: protection mask for the allocated pages
2532 * @node: node to use for allocation or NUMA_NO_NODE
2533 * @caller: caller's return address
2534 *
2535 * Allocate enough pages to cover @size from the page level
2536 * allocator with @gfp_mask flags. Map them into contiguous
2537 * kernel virtual space, using a pagetable protection of @prot.
2538 *
2539 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2540 * and __GFP_NOFAIL are not supported
2541 *
2542 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2543 * with mm people.
2544 *
2545 * Return: pointer to the allocated memory or %NULL on error
2546 */
2547static void *__vmalloc_node(unsigned long size, unsigned long align,
2548 gfp_t gfp_mask, pgprot_t prot,
2549 int node, const void *caller)
2550{
2551 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2552 gfp_mask, prot, 0, node, caller);
2553}
2554
2555void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
2556{
2557 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
2558 __builtin_return_address(0));
2559}
2560EXPORT_SYMBOL(__vmalloc);
2561
2562static inline void *__vmalloc_node_flags(unsigned long size,
2563 int node, gfp_t flags)
2564{
2565 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
2566 node, __builtin_return_address(0));
2567}
2568
2569
2570void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
2571 void *caller)
2572{
2573 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
2574}
2575
2576/**
2577 * vmalloc - allocate virtually contiguous memory
2578 * @size: allocation size
2579 *
2580 * Allocate enough pages to cover @size from the page level
2581 * allocator and map them into contiguous kernel virtual space.
2582 *
2583 * For tight control over page level allocator and protection flags
2584 * use __vmalloc() instead.
2585 *
2586 * Return: pointer to the allocated memory or %NULL on error
2587 */
2588void *vmalloc(unsigned long size)
2589{
2590 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2591 GFP_KERNEL);
2592}
2593EXPORT_SYMBOL(vmalloc);
2594
2595/**
2596 * vzalloc - allocate virtually contiguous memory with zero fill
2597 * @size: allocation size
2598 *
2599 * Allocate enough pages to cover @size from the page level
2600 * allocator and map them into contiguous kernel virtual space.
2601 * The memory allocated is set to zero.
2602 *
2603 * For tight control over page level allocator and protection flags
2604 * use __vmalloc() instead.
2605 *
2606 * Return: pointer to the allocated memory or %NULL on error
2607 */
2608void *vzalloc(unsigned long size)
2609{
2610 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2611 GFP_KERNEL | __GFP_ZERO);
2612}
2613EXPORT_SYMBOL(vzalloc);
2614
2615/**
2616 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2617 * @size: allocation size
2618 *
2619 * The resulting memory area is zeroed so it can be mapped to userspace
2620 * without leaking data.
2621 *
2622 * Return: pointer to the allocated memory or %NULL on error
2623 */
2624void *vmalloc_user(unsigned long size)
2625{
2626 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2627 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2628 VM_USERMAP, NUMA_NO_NODE,
2629 __builtin_return_address(0));
2630}
2631EXPORT_SYMBOL(vmalloc_user);
2632
2633/**
2634 * vmalloc_node - allocate memory on a specific node
2635 * @size: allocation size
2636 * @node: numa node
2637 *
2638 * Allocate enough pages to cover @size from the page level
2639 * allocator and map them into contiguous kernel virtual space.
2640 *
2641 * For tight control over page level allocator and protection flags
2642 * use __vmalloc() instead.
2643 *
2644 * Return: pointer to the allocated memory or %NULL on error
2645 */
2646void *vmalloc_node(unsigned long size, int node)
2647{
2648 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
2649 node, __builtin_return_address(0));
2650}
2651EXPORT_SYMBOL(vmalloc_node);
2652
2653/**
2654 * vzalloc_node - allocate memory on a specific node with zero fill
2655 * @size: allocation size
2656 * @node: numa node
2657 *
2658 * Allocate enough pages to cover @size from the page level
2659 * allocator and map them into contiguous kernel virtual space.
2660 * The memory allocated is set to zero.
2661 *
2662 * For tight control over page level allocator and protection flags
2663 * use __vmalloc_node() instead.
2664 *
2665 * Return: pointer to the allocated memory or %NULL on error
2666 */
2667void *vzalloc_node(unsigned long size, int node)
2668{
2669 return __vmalloc_node_flags(size, node,
2670 GFP_KERNEL | __GFP_ZERO);
2671}
2672EXPORT_SYMBOL(vzalloc_node);
2673
2674/**
2675 * vmalloc_exec - allocate virtually contiguous, executable memory
2676 * @size: allocation size
2677 *
2678 * Kernel-internal function to allocate enough pages to cover @size
2679 * the page level allocator and map them into contiguous and
2680 * executable kernel virtual space.
2681 *
2682 * For tight control over page level allocator and protection flags
2683 * use __vmalloc() instead.
2684 *
2685 * Return: pointer to the allocated memory or %NULL on error
2686 */
2687void *vmalloc_exec(unsigned long size)
2688{
2689 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
2690 GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
2691 NUMA_NO_NODE, __builtin_return_address(0));
2692}
2693
2694#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2695#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2696#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2697#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2698#else
2699/*
2700 * 64b systems should always have either DMA or DMA32 zones. For others
2701 * GFP_DMA32 should do the right thing and use the normal zone.
2702 */
2703#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2704#endif
2705
2706/**
2707 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2708 * @size: allocation size
2709 *
2710 * Allocate enough 32bit PA addressable pages to cover @size from the
2711 * page level allocator and map them into contiguous kernel virtual space.
2712 *
2713 * Return: pointer to the allocated memory or %NULL on error
2714 */
2715void *vmalloc_32(unsigned long size)
2716{
2717 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
2718 NUMA_NO_NODE, __builtin_return_address(0));
2719}
2720EXPORT_SYMBOL(vmalloc_32);
2721
2722/**
2723 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2724 * @size: allocation size
2725 *
2726 * The resulting memory area is 32bit addressable and zeroed so it can be
2727 * mapped to userspace without leaking data.
2728 *
2729 * Return: pointer to the allocated memory or %NULL on error
2730 */
2731void *vmalloc_32_user(unsigned long size)
2732{
2733 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2734 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2735 VM_USERMAP, NUMA_NO_NODE,
2736 __builtin_return_address(0));
2737}
2738EXPORT_SYMBOL(vmalloc_32_user);
2739
2740/*
2741 * small helper routine , copy contents to buf from addr.
2742 * If the page is not present, fill zero.
2743 */
2744
2745static int aligned_vread(char *buf, char *addr, unsigned long count)
2746{
2747 struct page *p;
2748 int copied = 0;
2749
2750 while (count) {
2751 unsigned long offset, length;
2752
2753 offset = offset_in_page(addr);
2754 length = PAGE_SIZE - offset;
2755 if (length > count)
2756 length = count;
2757 p = vmalloc_to_page(addr);
2758 /*
2759 * To do safe access to this _mapped_ area, we need
2760 * lock. But adding lock here means that we need to add
2761 * overhead of vmalloc()/vfree() calles for this _debug_
2762 * interface, rarely used. Instead of that, we'll use
2763 * kmap() and get small overhead in this access function.
2764 */
2765 if (p) {
2766 /*
2767 * we can expect USER0 is not used (see vread/vwrite's
2768 * function description)
2769 */
2770 void *map = kmap_atomic(p);
2771 memcpy(buf, map + offset, length);
2772 kunmap_atomic(map);
2773 } else
2774 memset(buf, 0, length);
2775
2776 addr += length;
2777 buf += length;
2778 copied += length;
2779 count -= length;
2780 }
2781 return copied;
2782}
2783
2784static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2785{
2786 struct page *p;
2787 int copied = 0;
2788
2789 while (count) {
2790 unsigned long offset, length;
2791
2792 offset = offset_in_page(addr);
2793 length = PAGE_SIZE - offset;
2794 if (length > count)
2795 length = count;
2796 p = vmalloc_to_page(addr);
2797 /*
2798 * To do safe access to this _mapped_ area, we need
2799 * lock. But adding lock here means that we need to add
2800 * overhead of vmalloc()/vfree() calles for this _debug_
2801 * interface, rarely used. Instead of that, we'll use
2802 * kmap() and get small overhead in this access function.
2803 */
2804 if (p) {
2805 /*
2806 * we can expect USER0 is not used (see vread/vwrite's
2807 * function description)
2808 */
2809 void *map = kmap_atomic(p);
2810 memcpy(map + offset, buf, length);
2811 kunmap_atomic(map);
2812 }
2813 addr += length;
2814 buf += length;
2815 copied += length;
2816 count -= length;
2817 }
2818 return copied;
2819}
2820
2821/**
2822 * vread() - read vmalloc area in a safe way.
2823 * @buf: buffer for reading data
2824 * @addr: vm address.
2825 * @count: number of bytes to be read.
2826 *
2827 * This function checks that addr is a valid vmalloc'ed area, and
2828 * copy data from that area to a given buffer. If the given memory range
2829 * of [addr...addr+count) includes some valid address, data is copied to
2830 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2831 * IOREMAP area is treated as memory hole and no copy is done.
2832 *
2833 * If [addr...addr+count) doesn't includes any intersects with alive
2834 * vm_struct area, returns 0. @buf should be kernel's buffer.
2835 *
2836 * Note: In usual ops, vread() is never necessary because the caller
2837 * should know vmalloc() area is valid and can use memcpy().
2838 * This is for routines which have to access vmalloc area without
2839 * any information, as /dev/kmem.
2840 *
2841 * Return: number of bytes for which addr and buf should be increased
2842 * (same number as @count) or %0 if [addr...addr+count) doesn't
2843 * include any intersection with valid vmalloc area
2844 */
2845long vread(char *buf, char *addr, unsigned long count)
2846{
2847 struct vmap_area *va;
2848 struct vm_struct *vm;
2849 char *vaddr, *buf_start = buf;
2850 unsigned long buflen = count;
2851 unsigned long n;
2852
2853 /* Don't allow overflow */
2854 if ((unsigned long) addr + count < count)
2855 count = -(unsigned long) addr;
2856
2857 spin_lock(&vmap_area_lock);
2858 list_for_each_entry(va, &vmap_area_list, list) {
2859 if (!count)
2860 break;
2861
2862 if (!va->vm)
2863 continue;
2864
2865 vm = va->vm;
2866 vaddr = (char *) vm->addr;
2867 if (addr >= vaddr + get_vm_area_size(vm))
2868 continue;
2869 while (addr < vaddr) {
2870 if (count == 0)
2871 goto finished;
2872 *buf = '\0';
2873 buf++;
2874 addr++;
2875 count--;
2876 }
2877 n = vaddr + get_vm_area_size(vm) - addr;
2878 if (n > count)
2879 n = count;
2880 if (!(vm->flags & VM_IOREMAP))
2881 aligned_vread(buf, addr, n);
2882 else /* IOREMAP area is treated as memory hole */
2883 memset(buf, 0, n);
2884 buf += n;
2885 addr += n;
2886 count -= n;
2887 }
2888finished:
2889 spin_unlock(&vmap_area_lock);
2890
2891 if (buf == buf_start)
2892 return 0;
2893 /* zero-fill memory holes */
2894 if (buf != buf_start + buflen)
2895 memset(buf, 0, buflen - (buf - buf_start));
2896
2897 return buflen;
2898}
2899
2900/**
2901 * vwrite() - write vmalloc area in a safe way.
2902 * @buf: buffer for source data
2903 * @addr: vm address.
2904 * @count: number of bytes to be read.
2905 *
2906 * This function checks that addr is a valid vmalloc'ed area, and
2907 * copy data from a buffer to the given addr. If specified range of
2908 * [addr...addr+count) includes some valid address, data is copied from
2909 * proper area of @buf. If there are memory holes, no copy to hole.
2910 * IOREMAP area is treated as memory hole and no copy is done.
2911 *
2912 * If [addr...addr+count) doesn't includes any intersects with alive
2913 * vm_struct area, returns 0. @buf should be kernel's buffer.
2914 *
2915 * Note: In usual ops, vwrite() is never necessary because the caller
2916 * should know vmalloc() area is valid and can use memcpy().
2917 * This is for routines which have to access vmalloc area without
2918 * any information, as /dev/kmem.
2919 *
2920 * Return: number of bytes for which addr and buf should be
2921 * increased (same number as @count) or %0 if [addr...addr+count)
2922 * doesn't include any intersection with valid vmalloc area
2923 */
2924long vwrite(char *buf, char *addr, unsigned long count)
2925{
2926 struct vmap_area *va;
2927 struct vm_struct *vm;
2928 char *vaddr;
2929 unsigned long n, buflen;
2930 int copied = 0;
2931
2932 /* Don't allow overflow */
2933 if ((unsigned long) addr + count < count)
2934 count = -(unsigned long) addr;
2935 buflen = count;
2936
2937 spin_lock(&vmap_area_lock);
2938 list_for_each_entry(va, &vmap_area_list, list) {
2939 if (!count)
2940 break;
2941
2942 if (!va->vm)
2943 continue;
2944
2945 vm = va->vm;
2946 vaddr = (char *) vm->addr;
2947 if (addr >= vaddr + get_vm_area_size(vm))
2948 continue;
2949 while (addr < vaddr) {
2950 if (count == 0)
2951 goto finished;
2952 buf++;
2953 addr++;
2954 count--;
2955 }
2956 n = vaddr + get_vm_area_size(vm) - addr;
2957 if (n > count)
2958 n = count;
2959 if (!(vm->flags & VM_IOREMAP)) {
2960 aligned_vwrite(buf, addr, n);
2961 copied++;
2962 }
2963 buf += n;
2964 addr += n;
2965 count -= n;
2966 }
2967finished:
2968 spin_unlock(&vmap_area_lock);
2969 if (!copied)
2970 return 0;
2971 return buflen;
2972}
2973
2974/**
2975 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2976 * @vma: vma to cover
2977 * @uaddr: target user address to start at
2978 * @kaddr: virtual address of vmalloc kernel memory
2979 * @size: size of map area
2980 *
2981 * Returns: 0 for success, -Exxx on failure
2982 *
2983 * This function checks that @kaddr is a valid vmalloc'ed area,
2984 * and that it is big enough to cover the range starting at
2985 * @uaddr in @vma. Will return failure if that criteria isn't
2986 * met.
2987 *
2988 * Similar to remap_pfn_range() (see mm/memory.c)
2989 */
2990int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2991 void *kaddr, unsigned long size)
2992{
2993 struct vm_struct *area;
2994
2995 size = PAGE_ALIGN(size);
2996
2997 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2998 return -EINVAL;
2999
3000 area = find_vm_area(kaddr);
3001 if (!area)
3002 return -EINVAL;
3003
3004 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3005 return -EINVAL;
3006
3007 if (kaddr + size > area->addr + get_vm_area_size(area))
3008 return -EINVAL;
3009
3010 do {
3011 struct page *page = vmalloc_to_page(kaddr);
3012 int ret;
3013
3014 ret = vm_insert_page(vma, uaddr, page);
3015 if (ret)
3016 return ret;
3017
3018 uaddr += PAGE_SIZE;
3019 kaddr += PAGE_SIZE;
3020 size -= PAGE_SIZE;
3021 } while (size > 0);
3022
3023 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3024
3025 return 0;
3026}
3027EXPORT_SYMBOL(remap_vmalloc_range_partial);
3028
3029/**
3030 * remap_vmalloc_range - map vmalloc pages to userspace
3031 * @vma: vma to cover (map full range of vma)
3032 * @addr: vmalloc memory
3033 * @pgoff: number of pages into addr before first page to map
3034 *
3035 * Returns: 0 for success, -Exxx on failure
3036 *
3037 * This function checks that addr is a valid vmalloc'ed area, and
3038 * that it is big enough to cover the vma. Will return failure if
3039 * that criteria isn't met.
3040 *
3041 * Similar to remap_pfn_range() (see mm/memory.c)
3042 */
3043int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3044 unsigned long pgoff)
3045{
3046 return remap_vmalloc_range_partial(vma, vma->vm_start,
3047 addr + (pgoff << PAGE_SHIFT),
3048 vma->vm_end - vma->vm_start);
3049}
3050EXPORT_SYMBOL(remap_vmalloc_range);
3051
3052/*
3053 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3054 * have one.
3055 *
3056 * The purpose of this function is to make sure the vmalloc area
3057 * mappings are identical in all page-tables in the system.
3058 */
3059void __weak vmalloc_sync_all(void)
3060{
3061}
3062
3063
3064static int f(pte_t *pte, unsigned long addr, void *data)
3065{
3066 pte_t ***p = data;
3067
3068 if (p) {
3069 *(*p) = pte;
3070 (*p)++;
3071 }
3072 return 0;
3073}
3074
3075/**
3076 * alloc_vm_area - allocate a range of kernel address space
3077 * @size: size of the area
3078 * @ptes: returns the PTEs for the address space
3079 *
3080 * Returns: NULL on failure, vm_struct on success
3081 *
3082 * This function reserves a range of kernel address space, and
3083 * allocates pagetables to map that range. No actual mappings
3084 * are created.
3085 *
3086 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3087 * allocated for the VM area are returned.
3088 */
3089struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3090{
3091 struct vm_struct *area;
3092
3093 area = get_vm_area_caller(size, VM_IOREMAP,
3094 __builtin_return_address(0));
3095 if (area == NULL)
3096 return NULL;
3097
3098 /*
3099 * This ensures that page tables are constructed for this region
3100 * of kernel virtual address space and mapped into init_mm.
3101 */
3102 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3103 size, f, ptes ? &ptes : NULL)) {
3104 free_vm_area(area);
3105 return NULL;
3106 }
3107
3108 return area;
3109}
3110EXPORT_SYMBOL_GPL(alloc_vm_area);
3111
3112void free_vm_area(struct vm_struct *area)
3113{
3114 struct vm_struct *ret;
3115 ret = remove_vm_area(area->addr);
3116 BUG_ON(ret != area);
3117 kfree(area);
3118}
3119EXPORT_SYMBOL_GPL(free_vm_area);
3120
3121#ifdef CONFIG_SMP
3122static struct vmap_area *node_to_va(struct rb_node *n)
3123{
3124 return rb_entry_safe(n, struct vmap_area, rb_node);
3125}
3126
3127/**
3128 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3129 * @addr: target address
3130 *
3131 * Returns: vmap_area if it is found. If there is no such area
3132 * the first highest(reverse order) vmap_area is returned
3133 * i.e. va->va_start < addr && va->va_end < addr or NULL
3134 * if there are no any areas before @addr.
3135 */
3136static struct vmap_area *
3137pvm_find_va_enclose_addr(unsigned long addr)
3138{
3139 struct vmap_area *va, *tmp;
3140 struct rb_node *n;
3141
3142 n = free_vmap_area_root.rb_node;
3143 va = NULL;
3144
3145 while (n) {
3146 tmp = rb_entry(n, struct vmap_area, rb_node);
3147 if (tmp->va_start <= addr) {
3148 va = tmp;
3149 if (tmp->va_end >= addr)
3150 break;
3151
3152 n = n->rb_right;
3153 } else {
3154 n = n->rb_left;
3155 }
3156 }
3157
3158 return va;
3159}
3160
3161/**
3162 * pvm_determine_end_from_reverse - find the highest aligned address
3163 * of free block below VMALLOC_END
3164 * @va:
3165 * in - the VA we start the search(reverse order);
3166 * out - the VA with the highest aligned end address.
3167 *
3168 * Returns: determined end address within vmap_area
3169 */
3170static unsigned long
3171pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3172{
3173 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3174 unsigned long addr;
3175
3176 if (likely(*va)) {
3177 list_for_each_entry_from_reverse((*va),
3178 &free_vmap_area_list, list) {
3179 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3180 if ((*va)->va_start < addr)
3181 return addr;
3182 }
3183 }
3184
3185 return 0;
3186}
3187
3188/**
3189 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3190 * @offsets: array containing offset of each area
3191 * @sizes: array containing size of each area
3192 * @nr_vms: the number of areas to allocate
3193 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3194 *
3195 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3196 * vm_structs on success, %NULL on failure
3197 *
3198 * Percpu allocator wants to use congruent vm areas so that it can
3199 * maintain the offsets among percpu areas. This function allocates
3200 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3201 * be scattered pretty far, distance between two areas easily going up
3202 * to gigabytes. To avoid interacting with regular vmallocs, these
3203 * areas are allocated from top.
3204 *
3205 * Despite its complicated look, this allocator is rather simple. It
3206 * does everything top-down and scans free blocks from the end looking
3207 * for matching base. While scanning, if any of the areas do not fit the
3208 * base address is pulled down to fit the area. Scanning is repeated till
3209 * all the areas fit and then all necessary data structures are inserted
3210 * and the result is returned.
3211 */
3212struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3213 const size_t *sizes, int nr_vms,
3214 size_t align)
3215{
3216 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3217 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3218 struct vmap_area **vas, *va;
3219 struct vm_struct **vms;
3220 int area, area2, last_area, term_area;
3221 unsigned long base, start, size, end, last_end;
3222 bool purged = false;
3223 enum fit_type type;
3224
3225 /* verify parameters and allocate data structures */
3226 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3227 for (last_area = 0, area = 0; area < nr_vms; area++) {
3228 start = offsets[area];
3229 end = start + sizes[area];
3230
3231 /* is everything aligned properly? */
3232 BUG_ON(!IS_ALIGNED(offsets[area], align));
3233 BUG_ON(!IS_ALIGNED(sizes[area], align));
3234
3235 /* detect the area with the highest address */
3236 if (start > offsets[last_area])
3237 last_area = area;
3238
3239 for (area2 = area + 1; area2 < nr_vms; area2++) {
3240 unsigned long start2 = offsets[area2];
3241 unsigned long end2 = start2 + sizes[area2];
3242
3243 BUG_ON(start2 < end && start < end2);
3244 }
3245 }
3246 last_end = offsets[last_area] + sizes[last_area];
3247
3248 if (vmalloc_end - vmalloc_start < last_end) {
3249 WARN_ON(true);
3250 return NULL;
3251 }
3252
3253 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3254 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3255 if (!vas || !vms)
3256 goto err_free2;
3257
3258 for (area = 0; area < nr_vms; area++) {
3259 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3260 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3261 if (!vas[area] || !vms[area])
3262 goto err_free;
3263 }
3264retry:
3265 spin_lock(&vmap_area_lock);
3266
3267 /* start scanning - we scan from the top, begin with the last area */
3268 area = term_area = last_area;
3269 start = offsets[area];
3270 end = start + sizes[area];
3271
3272 va = pvm_find_va_enclose_addr(vmalloc_end);
3273 base = pvm_determine_end_from_reverse(&va, align) - end;
3274
3275 while (true) {
3276 /*
3277 * base might have underflowed, add last_end before
3278 * comparing.
3279 */
3280 if (base + last_end < vmalloc_start + last_end)
3281 goto overflow;
3282
3283 /*
3284 * Fitting base has not been found.
3285 */
3286 if (va == NULL)
3287 goto overflow;
3288
3289 /*
3290 * If required width exeeds current VA block, move
3291 * base downwards and then recheck.
3292 */
3293 if (base + end > va->va_end) {
3294 base = pvm_determine_end_from_reverse(&va, align) - end;
3295 term_area = area;
3296 continue;
3297 }
3298
3299 /*
3300 * If this VA does not fit, move base downwards and recheck.
3301 */
3302 if (base + start < va->va_start) {
3303 va = node_to_va(rb_prev(&va->rb_node));
3304 base = pvm_determine_end_from_reverse(&va, align) - end;
3305 term_area = area;
3306 continue;
3307 }
3308
3309 /*
3310 * This area fits, move on to the previous one. If
3311 * the previous one is the terminal one, we're done.
3312 */
3313 area = (area + nr_vms - 1) % nr_vms;
3314 if (area == term_area)
3315 break;
3316
3317 start = offsets[area];
3318 end = start + sizes[area];
3319 va = pvm_find_va_enclose_addr(base + end);
3320 }
3321
3322 /* we've found a fitting base, insert all va's */
3323 for (area = 0; area < nr_vms; area++) {
3324 int ret;
3325
3326 start = base + offsets[area];
3327 size = sizes[area];
3328
3329 va = pvm_find_va_enclose_addr(start);
3330 if (WARN_ON_ONCE(va == NULL))
3331 /* It is a BUG(), but trigger recovery instead. */
3332 goto recovery;
3333
3334 type = classify_va_fit_type(va, start, size);
3335 if (WARN_ON_ONCE(type == NOTHING_FIT))
3336 /* It is a BUG(), but trigger recovery instead. */
3337 goto recovery;
3338
3339 ret = adjust_va_to_fit_type(va, start, size, type);
3340 if (unlikely(ret))
3341 goto recovery;
3342
3343 /* Allocated area. */
3344 va = vas[area];
3345 va->va_start = start;
3346 va->va_end = start + size;
3347
3348 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
3349 }
3350
3351 spin_unlock(&vmap_area_lock);
3352
3353 /* insert all vm's */
3354 for (area = 0; area < nr_vms; area++)
3355 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
3356 pcpu_get_vm_areas);
3357
3358 kfree(vas);
3359 return vms;
3360
3361recovery:
3362 /* Remove previously inserted areas. */
3363 while (area--) {
3364 __free_vmap_area(vas[area]);
3365 vas[area] = NULL;
3366 }
3367
3368overflow:
3369 spin_unlock(&vmap_area_lock);
3370 if (!purged) {
3371 purge_vmap_area_lazy();
3372 purged = true;
3373
3374 /* Before "retry", check if we recover. */
3375 for (area = 0; area < nr_vms; area++) {
3376 if (vas[area])
3377 continue;
3378
3379 vas[area] = kmem_cache_zalloc(
3380 vmap_area_cachep, GFP_KERNEL);
3381 if (!vas[area])
3382 goto err_free;
3383 }
3384
3385 goto retry;
3386 }
3387
3388err_free:
3389 for (area = 0; area < nr_vms; area++) {
3390 if (vas[area])
3391 kmem_cache_free(vmap_area_cachep, vas[area]);
3392
3393 kfree(vms[area]);
3394 }
3395err_free2:
3396 kfree(vas);
3397 kfree(vms);
3398 return NULL;
3399}
3400
3401/**
3402 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3403 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3404 * @nr_vms: the number of allocated areas
3405 *
3406 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3407 */
3408void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3409{
3410 int i;
3411
3412 for (i = 0; i < nr_vms; i++)
3413 free_vm_area(vms[i]);
3414 kfree(vms);
3415}
3416#endif /* CONFIG_SMP */
3417
3418#ifdef CONFIG_PROC_FS
3419static void *s_start(struct seq_file *m, loff_t *pos)
3420 __acquires(&vmap_area_lock)
3421{
3422 spin_lock(&vmap_area_lock);
3423 return seq_list_start(&vmap_area_list, *pos);
3424}
3425
3426static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3427{
3428 return seq_list_next(p, &vmap_area_list, pos);
3429}
3430
3431static void s_stop(struct seq_file *m, void *p)
3432 __releases(&vmap_area_lock)
3433{
3434 spin_unlock(&vmap_area_lock);
3435}
3436
3437static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3438{
3439 if (IS_ENABLED(CONFIG_NUMA)) {
3440 unsigned int nr, *counters = m->private;
3441
3442 if (!counters)
3443 return;
3444
3445 if (v->flags & VM_UNINITIALIZED)
3446 return;
3447 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3448 smp_rmb();
3449
3450 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3451
3452 for (nr = 0; nr < v->nr_pages; nr++)
3453 counters[page_to_nid(v->pages[nr])]++;
3454
3455 for_each_node_state(nr, N_HIGH_MEMORY)
3456 if (counters[nr])
3457 seq_printf(m, " N%u=%u", nr, counters[nr]);
3458 }
3459}
3460
3461static void show_purge_info(struct seq_file *m)
3462{
3463 struct llist_node *head;
3464 struct vmap_area *va;
3465
3466 head = READ_ONCE(vmap_purge_list.first);
3467 if (head == NULL)
3468 return;
3469
3470 llist_for_each_entry(va, head, purge_list) {
3471 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3472 (void *)va->va_start, (void *)va->va_end,
3473 va->va_end - va->va_start);
3474 }
3475}
3476
3477static int s_show(struct seq_file *m, void *p)
3478{
3479 struct vmap_area *va;
3480 struct vm_struct *v;
3481
3482 va = list_entry(p, struct vmap_area, list);
3483
3484 /*
3485 * s_show can encounter race with remove_vm_area, !vm on behalf
3486 * of vmap area is being tear down or vm_map_ram allocation.
3487 */
3488 if (!va->vm) {
3489 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3490 (void *)va->va_start, (void *)va->va_end,
3491 va->va_end - va->va_start);
3492
3493 return 0;
3494 }
3495
3496 v = va->vm;
3497
3498 seq_printf(m, "0x%pK-0x%pK %7ld",
3499 v->addr, v->addr + v->size, v->size);
3500
3501 if (v->caller)
3502 seq_printf(m, " %pS", v->caller);
3503
3504 if (v->nr_pages)
3505 seq_printf(m, " pages=%d", v->nr_pages);
3506
3507 if (v->phys_addr)
3508 seq_printf(m, " phys=%pa", &v->phys_addr);
3509
3510 if (v->flags & VM_IOREMAP)
3511 seq_puts(m, " ioremap");
3512
3513 if (v->flags & VM_ALLOC)
3514 seq_puts(m, " vmalloc");
3515
3516 if (v->flags & VM_MAP)
3517 seq_puts(m, " vmap");
3518
3519 if (v->flags & VM_USERMAP)
3520 seq_puts(m, " user");
3521
3522 if (v->flags & VM_DMA_COHERENT)
3523 seq_puts(m, " dma-coherent");
3524
3525 if (is_vmalloc_addr(v->pages))
3526 seq_puts(m, " vpages");
3527
3528 show_numa_info(m, v);
3529 seq_putc(m, '\n');
3530
3531 /*
3532 * As a final step, dump "unpurged" areas. Note,
3533 * that entire "/proc/vmallocinfo" output will not
3534 * be address sorted, because the purge list is not
3535 * sorted.
3536 */
3537 if (list_is_last(&va->list, &vmap_area_list))
3538 show_purge_info(m);
3539
3540 return 0;
3541}
3542
3543static const struct seq_operations vmalloc_op = {
3544 .start = s_start,
3545 .next = s_next,
3546 .stop = s_stop,
3547 .show = s_show,
3548};
3549
3550static int __init proc_vmalloc_init(void)
3551{
3552 if (IS_ENABLED(CONFIG_NUMA))
3553 proc_create_seq_private("vmallocinfo", 0400, NULL,
3554 &vmalloc_op,
3555 nr_node_ids * sizeof(unsigned int), NULL);
3556 else
3557 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3558 return 0;
3559}
3560module_init(proc_vmalloc_init);
3561
3562#endif