]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/vmalloc.c
Merge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[mirror_ubuntu-artful-kernel.git] / mm / vmalloc.c
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
41 struct vfree_deferred {
42 struct llist_head list;
43 struct work_struct wq;
44 };
45 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
46
47 static void __vunmap(const void *, int);
48
49 static void free_work(struct work_struct *w)
50 {
51 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
52 struct llist_node *llnode = llist_del_all(&p->list);
53 while (llnode) {
54 void *p = llnode;
55 llnode = llist_next(llnode);
56 __vunmap(p, 1);
57 }
58 }
59
60 /*** Page table manipulation functions ***/
61
62 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
63 {
64 pte_t *pte;
65
66 pte = pte_offset_kernel(pmd, addr);
67 do {
68 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
69 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
70 } while (pte++, addr += PAGE_SIZE, addr != end);
71 }
72
73 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
74 {
75 pmd_t *pmd;
76 unsigned long next;
77
78 pmd = pmd_offset(pud, addr);
79 do {
80 next = pmd_addr_end(addr, end);
81 if (pmd_clear_huge(pmd))
82 continue;
83 if (pmd_none_or_clear_bad(pmd))
84 continue;
85 vunmap_pte_range(pmd, addr, next);
86 } while (pmd++, addr = next, addr != end);
87 }
88
89 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
90 {
91 pud_t *pud;
92 unsigned long next;
93
94 pud = pud_offset(p4d, addr);
95 do {
96 next = pud_addr_end(addr, end);
97 if (pud_clear_huge(pud))
98 continue;
99 if (pud_none_or_clear_bad(pud))
100 continue;
101 vunmap_pmd_range(pud, addr, next);
102 } while (pud++, addr = next, addr != end);
103 }
104
105 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
106 {
107 p4d_t *p4d;
108 unsigned long next;
109
110 p4d = p4d_offset(pgd, addr);
111 do {
112 next = p4d_addr_end(addr, end);
113 if (p4d_clear_huge(p4d))
114 continue;
115 if (p4d_none_or_clear_bad(p4d))
116 continue;
117 vunmap_pud_range(p4d, addr, next);
118 } while (p4d++, addr = next, addr != end);
119 }
120
121 static void vunmap_page_range(unsigned long addr, unsigned long end)
122 {
123 pgd_t *pgd;
124 unsigned long next;
125
126 BUG_ON(addr >= end);
127 pgd = pgd_offset_k(addr);
128 do {
129 next = pgd_addr_end(addr, end);
130 if (pgd_none_or_clear_bad(pgd))
131 continue;
132 vunmap_p4d_range(pgd, addr, next);
133 } while (pgd++, addr = next, addr != end);
134 }
135
136 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
137 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
138 {
139 pte_t *pte;
140
141 /*
142 * nr is a running index into the array which helps higher level
143 * callers keep track of where we're up to.
144 */
145
146 pte = pte_alloc_kernel(pmd, addr);
147 if (!pte)
148 return -ENOMEM;
149 do {
150 struct page *page = pages[*nr];
151
152 if (WARN_ON(!pte_none(*pte)))
153 return -EBUSY;
154 if (WARN_ON(!page))
155 return -ENOMEM;
156 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
157 (*nr)++;
158 } while (pte++, addr += PAGE_SIZE, addr != end);
159 return 0;
160 }
161
162 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
163 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
164 {
165 pmd_t *pmd;
166 unsigned long next;
167
168 pmd = pmd_alloc(&init_mm, pud, addr);
169 if (!pmd)
170 return -ENOMEM;
171 do {
172 next = pmd_addr_end(addr, end);
173 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
174 return -ENOMEM;
175 } while (pmd++, addr = next, addr != end);
176 return 0;
177 }
178
179 static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
180 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
181 {
182 pud_t *pud;
183 unsigned long next;
184
185 pud = pud_alloc(&init_mm, p4d, addr);
186 if (!pud)
187 return -ENOMEM;
188 do {
189 next = pud_addr_end(addr, end);
190 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
191 return -ENOMEM;
192 } while (pud++, addr = next, addr != end);
193 return 0;
194 }
195
196 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
197 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
198 {
199 p4d_t *p4d;
200 unsigned long next;
201
202 p4d = p4d_alloc(&init_mm, pgd, addr);
203 if (!p4d)
204 return -ENOMEM;
205 do {
206 next = p4d_addr_end(addr, end);
207 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
208 return -ENOMEM;
209 } while (p4d++, addr = next, addr != end);
210 return 0;
211 }
212
213 /*
214 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
215 * will have pfns corresponding to the "pages" array.
216 *
217 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
218 */
219 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
220 pgprot_t prot, struct page **pages)
221 {
222 pgd_t *pgd;
223 unsigned long next;
224 unsigned long addr = start;
225 int err = 0;
226 int nr = 0;
227
228 BUG_ON(addr >= end);
229 pgd = pgd_offset_k(addr);
230 do {
231 next = pgd_addr_end(addr, end);
232 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
233 if (err)
234 return err;
235 } while (pgd++, addr = next, addr != end);
236
237 return nr;
238 }
239
240 static int vmap_page_range(unsigned long start, unsigned long end,
241 pgprot_t prot, struct page **pages)
242 {
243 int ret;
244
245 ret = vmap_page_range_noflush(start, end, prot, pages);
246 flush_cache_vmap(start, end);
247 return ret;
248 }
249
250 int is_vmalloc_or_module_addr(const void *x)
251 {
252 /*
253 * ARM, x86-64 and sparc64 put modules in a special place,
254 * and fall back on vmalloc() if that fails. Others
255 * just put it in the vmalloc space.
256 */
257 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
258 unsigned long addr = (unsigned long)x;
259 if (addr >= MODULES_VADDR && addr < MODULES_END)
260 return 1;
261 #endif
262 return is_vmalloc_addr(x);
263 }
264
265 /*
266 * Walk a vmap address to the struct page it maps.
267 */
268 struct page *vmalloc_to_page(const void *vmalloc_addr)
269 {
270 unsigned long addr = (unsigned long) vmalloc_addr;
271 struct page *page = NULL;
272 pgd_t *pgd = pgd_offset_k(addr);
273 p4d_t *p4d;
274 pud_t *pud;
275 pmd_t *pmd;
276 pte_t *ptep, pte;
277
278 /*
279 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
280 * architectures that do not vmalloc module space
281 */
282 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
283
284 if (pgd_none(*pgd))
285 return NULL;
286 p4d = p4d_offset(pgd, addr);
287 if (p4d_none(*p4d))
288 return NULL;
289 pud = pud_offset(p4d, addr);
290
291 /*
292 * Don't dereference bad PUD or PMD (below) entries. This will also
293 * identify huge mappings, which we may encounter on architectures
294 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
295 * identified as vmalloc addresses by is_vmalloc_addr(), but are
296 * not [unambiguously] associated with a struct page, so there is
297 * no correct value to return for them.
298 */
299 WARN_ON_ONCE(pud_bad(*pud));
300 if (pud_none(*pud) || pud_bad(*pud))
301 return NULL;
302 pmd = pmd_offset(pud, addr);
303 WARN_ON_ONCE(pmd_bad(*pmd));
304 if (pmd_none(*pmd) || pmd_bad(*pmd))
305 return NULL;
306
307 ptep = pte_offset_map(pmd, addr);
308 pte = *ptep;
309 if (pte_present(pte))
310 page = pte_page(pte);
311 pte_unmap(ptep);
312 return page;
313 }
314 EXPORT_SYMBOL(vmalloc_to_page);
315
316 /*
317 * Map a vmalloc()-space virtual address to the physical page frame number.
318 */
319 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
320 {
321 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
322 }
323 EXPORT_SYMBOL(vmalloc_to_pfn);
324
325
326 /*** Global kva allocator ***/
327
328 #define VM_LAZY_FREE 0x02
329 #define VM_VM_AREA 0x04
330
331 static DEFINE_SPINLOCK(vmap_area_lock);
332 /* Export for kexec only */
333 LIST_HEAD(vmap_area_list);
334 static LLIST_HEAD(vmap_purge_list);
335 static struct rb_root vmap_area_root = RB_ROOT;
336
337 /* The vmap cache globals are protected by vmap_area_lock */
338 static struct rb_node *free_vmap_cache;
339 static unsigned long cached_hole_size;
340 static unsigned long cached_vstart;
341 static unsigned long cached_align;
342
343 static unsigned long vmap_area_pcpu_hole;
344
345 static struct vmap_area *__find_vmap_area(unsigned long addr)
346 {
347 struct rb_node *n = vmap_area_root.rb_node;
348
349 while (n) {
350 struct vmap_area *va;
351
352 va = rb_entry(n, struct vmap_area, rb_node);
353 if (addr < va->va_start)
354 n = n->rb_left;
355 else if (addr >= va->va_end)
356 n = n->rb_right;
357 else
358 return va;
359 }
360
361 return NULL;
362 }
363
364 static void __insert_vmap_area(struct vmap_area *va)
365 {
366 struct rb_node **p = &vmap_area_root.rb_node;
367 struct rb_node *parent = NULL;
368 struct rb_node *tmp;
369
370 while (*p) {
371 struct vmap_area *tmp_va;
372
373 parent = *p;
374 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
375 if (va->va_start < tmp_va->va_end)
376 p = &(*p)->rb_left;
377 else if (va->va_end > tmp_va->va_start)
378 p = &(*p)->rb_right;
379 else
380 BUG();
381 }
382
383 rb_link_node(&va->rb_node, parent, p);
384 rb_insert_color(&va->rb_node, &vmap_area_root);
385
386 /* address-sort this list */
387 tmp = rb_prev(&va->rb_node);
388 if (tmp) {
389 struct vmap_area *prev;
390 prev = rb_entry(tmp, struct vmap_area, rb_node);
391 list_add_rcu(&va->list, &prev->list);
392 } else
393 list_add_rcu(&va->list, &vmap_area_list);
394 }
395
396 static void purge_vmap_area_lazy(void);
397
398 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
399
400 /*
401 * Allocate a region of KVA of the specified size and alignment, within the
402 * vstart and vend.
403 */
404 static struct vmap_area *alloc_vmap_area(unsigned long size,
405 unsigned long align,
406 unsigned long vstart, unsigned long vend,
407 int node, gfp_t gfp_mask)
408 {
409 struct vmap_area *va;
410 struct rb_node *n;
411 unsigned long addr;
412 int purged = 0;
413 struct vmap_area *first;
414
415 BUG_ON(!size);
416 BUG_ON(offset_in_page(size));
417 BUG_ON(!is_power_of_2(align));
418
419 might_sleep();
420
421 va = kmalloc_node(sizeof(struct vmap_area),
422 gfp_mask & GFP_RECLAIM_MASK, node);
423 if (unlikely(!va))
424 return ERR_PTR(-ENOMEM);
425
426 /*
427 * Only scan the relevant parts containing pointers to other objects
428 * to avoid false negatives.
429 */
430 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
431
432 retry:
433 spin_lock(&vmap_area_lock);
434 /*
435 * Invalidate cache if we have more permissive parameters.
436 * cached_hole_size notes the largest hole noticed _below_
437 * the vmap_area cached in free_vmap_cache: if size fits
438 * into that hole, we want to scan from vstart to reuse
439 * the hole instead of allocating above free_vmap_cache.
440 * Note that __free_vmap_area may update free_vmap_cache
441 * without updating cached_hole_size or cached_align.
442 */
443 if (!free_vmap_cache ||
444 size < cached_hole_size ||
445 vstart < cached_vstart ||
446 align < cached_align) {
447 nocache:
448 cached_hole_size = 0;
449 free_vmap_cache = NULL;
450 }
451 /* record if we encounter less permissive parameters */
452 cached_vstart = vstart;
453 cached_align = align;
454
455 /* find starting point for our search */
456 if (free_vmap_cache) {
457 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
458 addr = ALIGN(first->va_end, align);
459 if (addr < vstart)
460 goto nocache;
461 if (addr + size < addr)
462 goto overflow;
463
464 } else {
465 addr = ALIGN(vstart, align);
466 if (addr + size < addr)
467 goto overflow;
468
469 n = vmap_area_root.rb_node;
470 first = NULL;
471
472 while (n) {
473 struct vmap_area *tmp;
474 tmp = rb_entry(n, struct vmap_area, rb_node);
475 if (tmp->va_end >= addr) {
476 first = tmp;
477 if (tmp->va_start <= addr)
478 break;
479 n = n->rb_left;
480 } else
481 n = n->rb_right;
482 }
483
484 if (!first)
485 goto found;
486 }
487
488 /* from the starting point, walk areas until a suitable hole is found */
489 while (addr + size > first->va_start && addr + size <= vend) {
490 if (addr + cached_hole_size < first->va_start)
491 cached_hole_size = first->va_start - addr;
492 addr = ALIGN(first->va_end, align);
493 if (addr + size < addr)
494 goto overflow;
495
496 if (list_is_last(&first->list, &vmap_area_list))
497 goto found;
498
499 first = list_next_entry(first, list);
500 }
501
502 found:
503 if (addr + size > vend)
504 goto overflow;
505
506 va->va_start = addr;
507 va->va_end = addr + size;
508 va->flags = 0;
509 __insert_vmap_area(va);
510 free_vmap_cache = &va->rb_node;
511 spin_unlock(&vmap_area_lock);
512
513 BUG_ON(!IS_ALIGNED(va->va_start, align));
514 BUG_ON(va->va_start < vstart);
515 BUG_ON(va->va_end > vend);
516
517 return va;
518
519 overflow:
520 spin_unlock(&vmap_area_lock);
521 if (!purged) {
522 purge_vmap_area_lazy();
523 purged = 1;
524 goto retry;
525 }
526
527 if (gfpflags_allow_blocking(gfp_mask)) {
528 unsigned long freed = 0;
529 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
530 if (freed > 0) {
531 purged = 0;
532 goto retry;
533 }
534 }
535
536 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
537 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
538 size);
539 kfree(va);
540 return ERR_PTR(-EBUSY);
541 }
542
543 int register_vmap_purge_notifier(struct notifier_block *nb)
544 {
545 return blocking_notifier_chain_register(&vmap_notify_list, nb);
546 }
547 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
548
549 int unregister_vmap_purge_notifier(struct notifier_block *nb)
550 {
551 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
552 }
553 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
554
555 static void __free_vmap_area(struct vmap_area *va)
556 {
557 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
558
559 if (free_vmap_cache) {
560 if (va->va_end < cached_vstart) {
561 free_vmap_cache = NULL;
562 } else {
563 struct vmap_area *cache;
564 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
565 if (va->va_start <= cache->va_start) {
566 free_vmap_cache = rb_prev(&va->rb_node);
567 /*
568 * We don't try to update cached_hole_size or
569 * cached_align, but it won't go very wrong.
570 */
571 }
572 }
573 }
574 rb_erase(&va->rb_node, &vmap_area_root);
575 RB_CLEAR_NODE(&va->rb_node);
576 list_del_rcu(&va->list);
577
578 /*
579 * Track the highest possible candidate for pcpu area
580 * allocation. Areas outside of vmalloc area can be returned
581 * here too, consider only end addresses which fall inside
582 * vmalloc area proper.
583 */
584 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
585 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
586
587 kfree_rcu(va, rcu_head);
588 }
589
590 /*
591 * Free a region of KVA allocated by alloc_vmap_area
592 */
593 static void free_vmap_area(struct vmap_area *va)
594 {
595 spin_lock(&vmap_area_lock);
596 __free_vmap_area(va);
597 spin_unlock(&vmap_area_lock);
598 }
599
600 /*
601 * Clear the pagetable entries of a given vmap_area
602 */
603 static void unmap_vmap_area(struct vmap_area *va)
604 {
605 vunmap_page_range(va->va_start, va->va_end);
606 }
607
608 static void vmap_debug_free_range(unsigned long start, unsigned long end)
609 {
610 /*
611 * Unmap page tables and force a TLB flush immediately if pagealloc
612 * debugging is enabled. This catches use after free bugs similarly to
613 * those in linear kernel virtual address space after a page has been
614 * freed.
615 *
616 * All the lazy freeing logic is still retained, in order to minimise
617 * intrusiveness of this debugging feature.
618 *
619 * This is going to be *slow* (linear kernel virtual address debugging
620 * doesn't do a broadcast TLB flush so it is a lot faster).
621 */
622 if (debug_pagealloc_enabled()) {
623 vunmap_page_range(start, end);
624 flush_tlb_kernel_range(start, end);
625 }
626 }
627
628 /*
629 * lazy_max_pages is the maximum amount of virtual address space we gather up
630 * before attempting to purge with a TLB flush.
631 *
632 * There is a tradeoff here: a larger number will cover more kernel page tables
633 * and take slightly longer to purge, but it will linearly reduce the number of
634 * global TLB flushes that must be performed. It would seem natural to scale
635 * this number up linearly with the number of CPUs (because vmapping activity
636 * could also scale linearly with the number of CPUs), however it is likely
637 * that in practice, workloads might be constrained in other ways that mean
638 * vmap activity will not scale linearly with CPUs. Also, I want to be
639 * conservative and not introduce a big latency on huge systems, so go with
640 * a less aggressive log scale. It will still be an improvement over the old
641 * code, and it will be simple to change the scale factor if we find that it
642 * becomes a problem on bigger systems.
643 */
644 static unsigned long lazy_max_pages(void)
645 {
646 unsigned int log;
647
648 log = fls(num_online_cpus());
649
650 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
651 }
652
653 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
654
655 /*
656 * Serialize vmap purging. There is no actual criticial section protected
657 * by this look, but we want to avoid concurrent calls for performance
658 * reasons and to make the pcpu_get_vm_areas more deterministic.
659 */
660 static DEFINE_MUTEX(vmap_purge_lock);
661
662 /* for per-CPU blocks */
663 static void purge_fragmented_blocks_allcpus(void);
664
665 /*
666 * called before a call to iounmap() if the caller wants vm_area_struct's
667 * immediately freed.
668 */
669 void set_iounmap_nonlazy(void)
670 {
671 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
672 }
673
674 /*
675 * Purges all lazily-freed vmap areas.
676 */
677 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
678 {
679 struct llist_node *valist;
680 struct vmap_area *va;
681 struct vmap_area *n_va;
682 bool do_free = false;
683
684 lockdep_assert_held(&vmap_purge_lock);
685
686 valist = llist_del_all(&vmap_purge_list);
687 llist_for_each_entry(va, valist, purge_list) {
688 if (va->va_start < start)
689 start = va->va_start;
690 if (va->va_end > end)
691 end = va->va_end;
692 do_free = true;
693 }
694
695 if (!do_free)
696 return false;
697
698 flush_tlb_kernel_range(start, end);
699
700 spin_lock(&vmap_area_lock);
701 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
702 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
703
704 __free_vmap_area(va);
705 atomic_sub(nr, &vmap_lazy_nr);
706 cond_resched_lock(&vmap_area_lock);
707 }
708 spin_unlock(&vmap_area_lock);
709 return true;
710 }
711
712 /*
713 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
714 * is already purging.
715 */
716 static void try_purge_vmap_area_lazy(void)
717 {
718 if (mutex_trylock(&vmap_purge_lock)) {
719 __purge_vmap_area_lazy(ULONG_MAX, 0);
720 mutex_unlock(&vmap_purge_lock);
721 }
722 }
723
724 /*
725 * Kick off a purge of the outstanding lazy areas.
726 */
727 static void purge_vmap_area_lazy(void)
728 {
729 mutex_lock(&vmap_purge_lock);
730 purge_fragmented_blocks_allcpus();
731 __purge_vmap_area_lazy(ULONG_MAX, 0);
732 mutex_unlock(&vmap_purge_lock);
733 }
734
735 /*
736 * Free a vmap area, caller ensuring that the area has been unmapped
737 * and flush_cache_vunmap had been called for the correct range
738 * previously.
739 */
740 static void free_vmap_area_noflush(struct vmap_area *va)
741 {
742 int nr_lazy;
743
744 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
745 &vmap_lazy_nr);
746
747 /* After this point, we may free va at any time */
748 llist_add(&va->purge_list, &vmap_purge_list);
749
750 if (unlikely(nr_lazy > lazy_max_pages()))
751 try_purge_vmap_area_lazy();
752 }
753
754 /*
755 * Free and unmap a vmap area
756 */
757 static void free_unmap_vmap_area(struct vmap_area *va)
758 {
759 flush_cache_vunmap(va->va_start, va->va_end);
760 unmap_vmap_area(va);
761 free_vmap_area_noflush(va);
762 }
763
764 static struct vmap_area *find_vmap_area(unsigned long addr)
765 {
766 struct vmap_area *va;
767
768 spin_lock(&vmap_area_lock);
769 va = __find_vmap_area(addr);
770 spin_unlock(&vmap_area_lock);
771
772 return va;
773 }
774
775 /*** Per cpu kva allocator ***/
776
777 /*
778 * vmap space is limited especially on 32 bit architectures. Ensure there is
779 * room for at least 16 percpu vmap blocks per CPU.
780 */
781 /*
782 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
783 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
784 * instead (we just need a rough idea)
785 */
786 #if BITS_PER_LONG == 32
787 #define VMALLOC_SPACE (128UL*1024*1024)
788 #else
789 #define VMALLOC_SPACE (128UL*1024*1024*1024)
790 #endif
791
792 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
793 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
794 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
795 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
796 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
797 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
798 #define VMAP_BBMAP_BITS \
799 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
800 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
801 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
802
803 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
804
805 static bool vmap_initialized __read_mostly = false;
806
807 struct vmap_block_queue {
808 spinlock_t lock;
809 struct list_head free;
810 };
811
812 struct vmap_block {
813 spinlock_t lock;
814 struct vmap_area *va;
815 unsigned long free, dirty;
816 unsigned long dirty_min, dirty_max; /*< dirty range */
817 struct list_head free_list;
818 struct rcu_head rcu_head;
819 struct list_head purge;
820 };
821
822 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
823 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
824
825 /*
826 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
827 * in the free path. Could get rid of this if we change the API to return a
828 * "cookie" from alloc, to be passed to free. But no big deal yet.
829 */
830 static DEFINE_SPINLOCK(vmap_block_tree_lock);
831 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
832
833 /*
834 * We should probably have a fallback mechanism to allocate virtual memory
835 * out of partially filled vmap blocks. However vmap block sizing should be
836 * fairly reasonable according to the vmalloc size, so it shouldn't be a
837 * big problem.
838 */
839
840 static unsigned long addr_to_vb_idx(unsigned long addr)
841 {
842 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
843 addr /= VMAP_BLOCK_SIZE;
844 return addr;
845 }
846
847 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
848 {
849 unsigned long addr;
850
851 addr = va_start + (pages_off << PAGE_SHIFT);
852 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
853 return (void *)addr;
854 }
855
856 /**
857 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
858 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
859 * @order: how many 2^order pages should be occupied in newly allocated block
860 * @gfp_mask: flags for the page level allocator
861 *
862 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
863 */
864 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
865 {
866 struct vmap_block_queue *vbq;
867 struct vmap_block *vb;
868 struct vmap_area *va;
869 unsigned long vb_idx;
870 int node, err;
871 void *vaddr;
872
873 node = numa_node_id();
874
875 vb = kmalloc_node(sizeof(struct vmap_block),
876 gfp_mask & GFP_RECLAIM_MASK, node);
877 if (unlikely(!vb))
878 return ERR_PTR(-ENOMEM);
879
880 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
881 VMALLOC_START, VMALLOC_END,
882 node, gfp_mask);
883 if (IS_ERR(va)) {
884 kfree(vb);
885 return ERR_CAST(va);
886 }
887
888 err = radix_tree_preload(gfp_mask);
889 if (unlikely(err)) {
890 kfree(vb);
891 free_vmap_area(va);
892 return ERR_PTR(err);
893 }
894
895 vaddr = vmap_block_vaddr(va->va_start, 0);
896 spin_lock_init(&vb->lock);
897 vb->va = va;
898 /* At least something should be left free */
899 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
900 vb->free = VMAP_BBMAP_BITS - (1UL << order);
901 vb->dirty = 0;
902 vb->dirty_min = VMAP_BBMAP_BITS;
903 vb->dirty_max = 0;
904 INIT_LIST_HEAD(&vb->free_list);
905
906 vb_idx = addr_to_vb_idx(va->va_start);
907 spin_lock(&vmap_block_tree_lock);
908 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
909 spin_unlock(&vmap_block_tree_lock);
910 BUG_ON(err);
911 radix_tree_preload_end();
912
913 vbq = &get_cpu_var(vmap_block_queue);
914 spin_lock(&vbq->lock);
915 list_add_tail_rcu(&vb->free_list, &vbq->free);
916 spin_unlock(&vbq->lock);
917 put_cpu_var(vmap_block_queue);
918
919 return vaddr;
920 }
921
922 static void free_vmap_block(struct vmap_block *vb)
923 {
924 struct vmap_block *tmp;
925 unsigned long vb_idx;
926
927 vb_idx = addr_to_vb_idx(vb->va->va_start);
928 spin_lock(&vmap_block_tree_lock);
929 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
930 spin_unlock(&vmap_block_tree_lock);
931 BUG_ON(tmp != vb);
932
933 free_vmap_area_noflush(vb->va);
934 kfree_rcu(vb, rcu_head);
935 }
936
937 static void purge_fragmented_blocks(int cpu)
938 {
939 LIST_HEAD(purge);
940 struct vmap_block *vb;
941 struct vmap_block *n_vb;
942 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
943
944 rcu_read_lock();
945 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
946
947 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
948 continue;
949
950 spin_lock(&vb->lock);
951 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
952 vb->free = 0; /* prevent further allocs after releasing lock */
953 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
954 vb->dirty_min = 0;
955 vb->dirty_max = VMAP_BBMAP_BITS;
956 spin_lock(&vbq->lock);
957 list_del_rcu(&vb->free_list);
958 spin_unlock(&vbq->lock);
959 spin_unlock(&vb->lock);
960 list_add_tail(&vb->purge, &purge);
961 } else
962 spin_unlock(&vb->lock);
963 }
964 rcu_read_unlock();
965
966 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
967 list_del(&vb->purge);
968 free_vmap_block(vb);
969 }
970 }
971
972 static void purge_fragmented_blocks_allcpus(void)
973 {
974 int cpu;
975
976 for_each_possible_cpu(cpu)
977 purge_fragmented_blocks(cpu);
978 }
979
980 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
981 {
982 struct vmap_block_queue *vbq;
983 struct vmap_block *vb;
984 void *vaddr = NULL;
985 unsigned int order;
986
987 BUG_ON(offset_in_page(size));
988 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
989 if (WARN_ON(size == 0)) {
990 /*
991 * Allocating 0 bytes isn't what caller wants since
992 * get_order(0) returns funny result. Just warn and terminate
993 * early.
994 */
995 return NULL;
996 }
997 order = get_order(size);
998
999 rcu_read_lock();
1000 vbq = &get_cpu_var(vmap_block_queue);
1001 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1002 unsigned long pages_off;
1003
1004 spin_lock(&vb->lock);
1005 if (vb->free < (1UL << order)) {
1006 spin_unlock(&vb->lock);
1007 continue;
1008 }
1009
1010 pages_off = VMAP_BBMAP_BITS - vb->free;
1011 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1012 vb->free -= 1UL << order;
1013 if (vb->free == 0) {
1014 spin_lock(&vbq->lock);
1015 list_del_rcu(&vb->free_list);
1016 spin_unlock(&vbq->lock);
1017 }
1018
1019 spin_unlock(&vb->lock);
1020 break;
1021 }
1022
1023 put_cpu_var(vmap_block_queue);
1024 rcu_read_unlock();
1025
1026 /* Allocate new block if nothing was found */
1027 if (!vaddr)
1028 vaddr = new_vmap_block(order, gfp_mask);
1029
1030 return vaddr;
1031 }
1032
1033 static void vb_free(const void *addr, unsigned long size)
1034 {
1035 unsigned long offset;
1036 unsigned long vb_idx;
1037 unsigned int order;
1038 struct vmap_block *vb;
1039
1040 BUG_ON(offset_in_page(size));
1041 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1042
1043 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1044
1045 order = get_order(size);
1046
1047 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1048 offset >>= PAGE_SHIFT;
1049
1050 vb_idx = addr_to_vb_idx((unsigned long)addr);
1051 rcu_read_lock();
1052 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1053 rcu_read_unlock();
1054 BUG_ON(!vb);
1055
1056 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1057
1058 spin_lock(&vb->lock);
1059
1060 /* Expand dirty range */
1061 vb->dirty_min = min(vb->dirty_min, offset);
1062 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1063
1064 vb->dirty += 1UL << order;
1065 if (vb->dirty == VMAP_BBMAP_BITS) {
1066 BUG_ON(vb->free);
1067 spin_unlock(&vb->lock);
1068 free_vmap_block(vb);
1069 } else
1070 spin_unlock(&vb->lock);
1071 }
1072
1073 /**
1074 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1075 *
1076 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1077 * to amortize TLB flushing overheads. What this means is that any page you
1078 * have now, may, in a former life, have been mapped into kernel virtual
1079 * address by the vmap layer and so there might be some CPUs with TLB entries
1080 * still referencing that page (additional to the regular 1:1 kernel mapping).
1081 *
1082 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1083 * be sure that none of the pages we have control over will have any aliases
1084 * from the vmap layer.
1085 */
1086 void vm_unmap_aliases(void)
1087 {
1088 unsigned long start = ULONG_MAX, end = 0;
1089 int cpu;
1090 int flush = 0;
1091
1092 if (unlikely(!vmap_initialized))
1093 return;
1094
1095 might_sleep();
1096
1097 for_each_possible_cpu(cpu) {
1098 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1099 struct vmap_block *vb;
1100
1101 rcu_read_lock();
1102 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1103 spin_lock(&vb->lock);
1104 if (vb->dirty) {
1105 unsigned long va_start = vb->va->va_start;
1106 unsigned long s, e;
1107
1108 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1109 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1110
1111 start = min(s, start);
1112 end = max(e, end);
1113
1114 flush = 1;
1115 }
1116 spin_unlock(&vb->lock);
1117 }
1118 rcu_read_unlock();
1119 }
1120
1121 mutex_lock(&vmap_purge_lock);
1122 purge_fragmented_blocks_allcpus();
1123 if (!__purge_vmap_area_lazy(start, end) && flush)
1124 flush_tlb_kernel_range(start, end);
1125 mutex_unlock(&vmap_purge_lock);
1126 }
1127 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1128
1129 /**
1130 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1131 * @mem: the pointer returned by vm_map_ram
1132 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1133 */
1134 void vm_unmap_ram(const void *mem, unsigned int count)
1135 {
1136 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1137 unsigned long addr = (unsigned long)mem;
1138 struct vmap_area *va;
1139
1140 might_sleep();
1141 BUG_ON(!addr);
1142 BUG_ON(addr < VMALLOC_START);
1143 BUG_ON(addr > VMALLOC_END);
1144 BUG_ON(!PAGE_ALIGNED(addr));
1145
1146 debug_check_no_locks_freed(mem, size);
1147 vmap_debug_free_range(addr, addr+size);
1148
1149 if (likely(count <= VMAP_MAX_ALLOC)) {
1150 vb_free(mem, size);
1151 return;
1152 }
1153
1154 va = find_vmap_area(addr);
1155 BUG_ON(!va);
1156 free_unmap_vmap_area(va);
1157 }
1158 EXPORT_SYMBOL(vm_unmap_ram);
1159
1160 /**
1161 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1162 * @pages: an array of pointers to the pages to be mapped
1163 * @count: number of pages
1164 * @node: prefer to allocate data structures on this node
1165 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1166 *
1167 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1168 * faster than vmap so it's good. But if you mix long-life and short-life
1169 * objects with vm_map_ram(), it could consume lots of address space through
1170 * fragmentation (especially on a 32bit machine). You could see failures in
1171 * the end. Please use this function for short-lived objects.
1172 *
1173 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1174 */
1175 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1176 {
1177 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1178 unsigned long addr;
1179 void *mem;
1180
1181 if (likely(count <= VMAP_MAX_ALLOC)) {
1182 mem = vb_alloc(size, GFP_KERNEL);
1183 if (IS_ERR(mem))
1184 return NULL;
1185 addr = (unsigned long)mem;
1186 } else {
1187 struct vmap_area *va;
1188 va = alloc_vmap_area(size, PAGE_SIZE,
1189 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1190 if (IS_ERR(va))
1191 return NULL;
1192
1193 addr = va->va_start;
1194 mem = (void *)addr;
1195 }
1196 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1197 vm_unmap_ram(mem, count);
1198 return NULL;
1199 }
1200 return mem;
1201 }
1202 EXPORT_SYMBOL(vm_map_ram);
1203
1204 static struct vm_struct *vmlist __initdata;
1205 /**
1206 * vm_area_add_early - add vmap area early during boot
1207 * @vm: vm_struct to add
1208 *
1209 * This function is used to add fixed kernel vm area to vmlist before
1210 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1211 * should contain proper values and the other fields should be zero.
1212 *
1213 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1214 */
1215 void __init vm_area_add_early(struct vm_struct *vm)
1216 {
1217 struct vm_struct *tmp, **p;
1218
1219 BUG_ON(vmap_initialized);
1220 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1221 if (tmp->addr >= vm->addr) {
1222 BUG_ON(tmp->addr < vm->addr + vm->size);
1223 break;
1224 } else
1225 BUG_ON(tmp->addr + tmp->size > vm->addr);
1226 }
1227 vm->next = *p;
1228 *p = vm;
1229 }
1230
1231 /**
1232 * vm_area_register_early - register vmap area early during boot
1233 * @vm: vm_struct to register
1234 * @align: requested alignment
1235 *
1236 * This function is used to register kernel vm area before
1237 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1238 * proper values on entry and other fields should be zero. On return,
1239 * vm->addr contains the allocated address.
1240 *
1241 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1242 */
1243 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1244 {
1245 static size_t vm_init_off __initdata;
1246 unsigned long addr;
1247
1248 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1249 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1250
1251 vm->addr = (void *)addr;
1252
1253 vm_area_add_early(vm);
1254 }
1255
1256 void __init vmalloc_init(void)
1257 {
1258 struct vmap_area *va;
1259 struct vm_struct *tmp;
1260 int i;
1261
1262 for_each_possible_cpu(i) {
1263 struct vmap_block_queue *vbq;
1264 struct vfree_deferred *p;
1265
1266 vbq = &per_cpu(vmap_block_queue, i);
1267 spin_lock_init(&vbq->lock);
1268 INIT_LIST_HEAD(&vbq->free);
1269 p = &per_cpu(vfree_deferred, i);
1270 init_llist_head(&p->list);
1271 INIT_WORK(&p->wq, free_work);
1272 }
1273
1274 /* Import existing vmlist entries. */
1275 for (tmp = vmlist; tmp; tmp = tmp->next) {
1276 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1277 va->flags = VM_VM_AREA;
1278 va->va_start = (unsigned long)tmp->addr;
1279 va->va_end = va->va_start + tmp->size;
1280 va->vm = tmp;
1281 __insert_vmap_area(va);
1282 }
1283
1284 vmap_area_pcpu_hole = VMALLOC_END;
1285
1286 vmap_initialized = true;
1287 }
1288
1289 /**
1290 * map_kernel_range_noflush - map kernel VM area with the specified pages
1291 * @addr: start of the VM area to map
1292 * @size: size of the VM area to map
1293 * @prot: page protection flags to use
1294 * @pages: pages to map
1295 *
1296 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1297 * specify should have been allocated using get_vm_area() and its
1298 * friends.
1299 *
1300 * NOTE:
1301 * This function does NOT do any cache flushing. The caller is
1302 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1303 * before calling this function.
1304 *
1305 * RETURNS:
1306 * The number of pages mapped on success, -errno on failure.
1307 */
1308 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1309 pgprot_t prot, struct page **pages)
1310 {
1311 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1312 }
1313
1314 /**
1315 * unmap_kernel_range_noflush - unmap kernel VM area
1316 * @addr: start of the VM area to unmap
1317 * @size: size of the VM area to unmap
1318 *
1319 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1320 * specify should have been allocated using get_vm_area() and its
1321 * friends.
1322 *
1323 * NOTE:
1324 * This function does NOT do any cache flushing. The caller is
1325 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1326 * before calling this function and flush_tlb_kernel_range() after.
1327 */
1328 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1329 {
1330 vunmap_page_range(addr, addr + size);
1331 }
1332 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1333
1334 /**
1335 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1336 * @addr: start of the VM area to unmap
1337 * @size: size of the VM area to unmap
1338 *
1339 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1340 * the unmapping and tlb after.
1341 */
1342 void unmap_kernel_range(unsigned long addr, unsigned long size)
1343 {
1344 unsigned long end = addr + size;
1345
1346 flush_cache_vunmap(addr, end);
1347 vunmap_page_range(addr, end);
1348 flush_tlb_kernel_range(addr, end);
1349 }
1350 EXPORT_SYMBOL_GPL(unmap_kernel_range);
1351
1352 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1353 {
1354 unsigned long addr = (unsigned long)area->addr;
1355 unsigned long end = addr + get_vm_area_size(area);
1356 int err;
1357
1358 err = vmap_page_range(addr, end, prot, pages);
1359
1360 return err > 0 ? 0 : err;
1361 }
1362 EXPORT_SYMBOL_GPL(map_vm_area);
1363
1364 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1365 unsigned long flags, const void *caller)
1366 {
1367 spin_lock(&vmap_area_lock);
1368 vm->flags = flags;
1369 vm->addr = (void *)va->va_start;
1370 vm->size = va->va_end - va->va_start;
1371 vm->caller = caller;
1372 va->vm = vm;
1373 va->flags |= VM_VM_AREA;
1374 spin_unlock(&vmap_area_lock);
1375 }
1376
1377 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1378 {
1379 /*
1380 * Before removing VM_UNINITIALIZED,
1381 * we should make sure that vm has proper values.
1382 * Pair with smp_rmb() in show_numa_info().
1383 */
1384 smp_wmb();
1385 vm->flags &= ~VM_UNINITIALIZED;
1386 }
1387
1388 static struct vm_struct *__get_vm_area_node(unsigned long size,
1389 unsigned long align, unsigned long flags, unsigned long start,
1390 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1391 {
1392 struct vmap_area *va;
1393 struct vm_struct *area;
1394
1395 BUG_ON(in_interrupt());
1396 size = PAGE_ALIGN(size);
1397 if (unlikely(!size))
1398 return NULL;
1399
1400 if (flags & VM_IOREMAP)
1401 align = 1ul << clamp_t(int, get_count_order_long(size),
1402 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1403
1404 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1405 if (unlikely(!area))
1406 return NULL;
1407
1408 if (!(flags & VM_NO_GUARD))
1409 size += PAGE_SIZE;
1410
1411 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1412 if (IS_ERR(va)) {
1413 kfree(area);
1414 return NULL;
1415 }
1416
1417 setup_vmalloc_vm(area, va, flags, caller);
1418
1419 return area;
1420 }
1421
1422 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1423 unsigned long start, unsigned long end)
1424 {
1425 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1426 GFP_KERNEL, __builtin_return_address(0));
1427 }
1428 EXPORT_SYMBOL_GPL(__get_vm_area);
1429
1430 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1431 unsigned long start, unsigned long end,
1432 const void *caller)
1433 {
1434 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1435 GFP_KERNEL, caller);
1436 }
1437
1438 /**
1439 * get_vm_area - reserve a contiguous kernel virtual area
1440 * @size: size of the area
1441 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1442 *
1443 * Search an area of @size in the kernel virtual mapping area,
1444 * and reserved it for out purposes. Returns the area descriptor
1445 * on success or %NULL on failure.
1446 */
1447 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1448 {
1449 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1450 NUMA_NO_NODE, GFP_KERNEL,
1451 __builtin_return_address(0));
1452 }
1453
1454 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1455 const void *caller)
1456 {
1457 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1458 NUMA_NO_NODE, GFP_KERNEL, caller);
1459 }
1460
1461 /**
1462 * find_vm_area - find a continuous kernel virtual area
1463 * @addr: base address
1464 *
1465 * Search for the kernel VM area starting at @addr, and return it.
1466 * It is up to the caller to do all required locking to keep the returned
1467 * pointer valid.
1468 */
1469 struct vm_struct *find_vm_area(const void *addr)
1470 {
1471 struct vmap_area *va;
1472
1473 va = find_vmap_area((unsigned long)addr);
1474 if (va && va->flags & VM_VM_AREA)
1475 return va->vm;
1476
1477 return NULL;
1478 }
1479
1480 /**
1481 * remove_vm_area - find and remove a continuous kernel virtual area
1482 * @addr: base address
1483 *
1484 * Search for the kernel VM area starting at @addr, and remove it.
1485 * This function returns the found VM area, but using it is NOT safe
1486 * on SMP machines, except for its size or flags.
1487 */
1488 struct vm_struct *remove_vm_area(const void *addr)
1489 {
1490 struct vmap_area *va;
1491
1492 might_sleep();
1493
1494 va = find_vmap_area((unsigned long)addr);
1495 if (va && va->flags & VM_VM_AREA) {
1496 struct vm_struct *vm = va->vm;
1497
1498 spin_lock(&vmap_area_lock);
1499 va->vm = NULL;
1500 va->flags &= ~VM_VM_AREA;
1501 va->flags |= VM_LAZY_FREE;
1502 spin_unlock(&vmap_area_lock);
1503
1504 vmap_debug_free_range(va->va_start, va->va_end);
1505 kasan_free_shadow(vm);
1506 free_unmap_vmap_area(va);
1507
1508 return vm;
1509 }
1510 return NULL;
1511 }
1512
1513 static void __vunmap(const void *addr, int deallocate_pages)
1514 {
1515 struct vm_struct *area;
1516
1517 if (!addr)
1518 return;
1519
1520 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1521 addr))
1522 return;
1523
1524 area = remove_vm_area(addr);
1525 if (unlikely(!area)) {
1526 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1527 addr);
1528 return;
1529 }
1530
1531 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1532 debug_check_no_obj_freed(addr, get_vm_area_size(area));
1533
1534 if (deallocate_pages) {
1535 int i;
1536
1537 for (i = 0; i < area->nr_pages; i++) {
1538 struct page *page = area->pages[i];
1539
1540 BUG_ON(!page);
1541 __free_pages(page, 0);
1542 }
1543
1544 kvfree(area->pages);
1545 }
1546
1547 kfree(area);
1548 return;
1549 }
1550
1551 static inline void __vfree_deferred(const void *addr)
1552 {
1553 /*
1554 * Use raw_cpu_ptr() because this can be called from preemptible
1555 * context. Preemption is absolutely fine here, because the llist_add()
1556 * implementation is lockless, so it works even if we are adding to
1557 * nother cpu's list. schedule_work() should be fine with this too.
1558 */
1559 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1560
1561 if (llist_add((struct llist_node *)addr, &p->list))
1562 schedule_work(&p->wq);
1563 }
1564
1565 /**
1566 * vfree_atomic - release memory allocated by vmalloc()
1567 * @addr: memory base address
1568 *
1569 * This one is just like vfree() but can be called in any atomic context
1570 * except NMIs.
1571 */
1572 void vfree_atomic(const void *addr)
1573 {
1574 BUG_ON(in_nmi());
1575
1576 kmemleak_free(addr);
1577
1578 if (!addr)
1579 return;
1580 __vfree_deferred(addr);
1581 }
1582
1583 /**
1584 * vfree - release memory allocated by vmalloc()
1585 * @addr: memory base address
1586 *
1587 * Free the virtually continuous memory area starting at @addr, as
1588 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1589 * NULL, no operation is performed.
1590 *
1591 * Must not be called in NMI context (strictly speaking, only if we don't
1592 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1593 * conventions for vfree() arch-depenedent would be a really bad idea)
1594 *
1595 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
1596 */
1597 void vfree(const void *addr)
1598 {
1599 BUG_ON(in_nmi());
1600
1601 kmemleak_free(addr);
1602
1603 if (!addr)
1604 return;
1605 if (unlikely(in_interrupt()))
1606 __vfree_deferred(addr);
1607 else
1608 __vunmap(addr, 1);
1609 }
1610 EXPORT_SYMBOL(vfree);
1611
1612 /**
1613 * vunmap - release virtual mapping obtained by vmap()
1614 * @addr: memory base address
1615 *
1616 * Free the virtually contiguous memory area starting at @addr,
1617 * which was created from the page array passed to vmap().
1618 *
1619 * Must not be called in interrupt context.
1620 */
1621 void vunmap(const void *addr)
1622 {
1623 BUG_ON(in_interrupt());
1624 might_sleep();
1625 if (addr)
1626 __vunmap(addr, 0);
1627 }
1628 EXPORT_SYMBOL(vunmap);
1629
1630 /**
1631 * vmap - map an array of pages into virtually contiguous space
1632 * @pages: array of page pointers
1633 * @count: number of pages to map
1634 * @flags: vm_area->flags
1635 * @prot: page protection for the mapping
1636 *
1637 * Maps @count pages from @pages into contiguous kernel virtual
1638 * space.
1639 */
1640 void *vmap(struct page **pages, unsigned int count,
1641 unsigned long flags, pgprot_t prot)
1642 {
1643 struct vm_struct *area;
1644 unsigned long size; /* In bytes */
1645
1646 might_sleep();
1647
1648 if (count > totalram_pages)
1649 return NULL;
1650
1651 size = (unsigned long)count << PAGE_SHIFT;
1652 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1653 if (!area)
1654 return NULL;
1655
1656 if (map_vm_area(area, prot, pages)) {
1657 vunmap(area->addr);
1658 return NULL;
1659 }
1660
1661 return area->addr;
1662 }
1663 EXPORT_SYMBOL(vmap);
1664
1665 static void *__vmalloc_node(unsigned long size, unsigned long align,
1666 gfp_t gfp_mask, pgprot_t prot,
1667 int node, const void *caller);
1668 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1669 pgprot_t prot, int node)
1670 {
1671 struct page **pages;
1672 unsigned int nr_pages, array_size, i;
1673 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1674 const gfp_t alloc_mask = gfp_mask | __GFP_HIGHMEM | __GFP_NOWARN;
1675
1676 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1677 array_size = (nr_pages * sizeof(struct page *));
1678
1679 area->nr_pages = nr_pages;
1680 /* Please note that the recursion is strictly bounded. */
1681 if (array_size > PAGE_SIZE) {
1682 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1683 PAGE_KERNEL, node, area->caller);
1684 } else {
1685 pages = kmalloc_node(array_size, nested_gfp, node);
1686 }
1687 area->pages = pages;
1688 if (!area->pages) {
1689 remove_vm_area(area->addr);
1690 kfree(area);
1691 return NULL;
1692 }
1693
1694 for (i = 0; i < area->nr_pages; i++) {
1695 struct page *page;
1696
1697 if (fatal_signal_pending(current)) {
1698 area->nr_pages = i;
1699 goto fail_no_warn;
1700 }
1701
1702 if (node == NUMA_NO_NODE)
1703 page = alloc_page(alloc_mask);
1704 else
1705 page = alloc_pages_node(node, alloc_mask, 0);
1706
1707 if (unlikely(!page)) {
1708 /* Successfully allocated i pages, free them in __vunmap() */
1709 area->nr_pages = i;
1710 goto fail;
1711 }
1712 area->pages[i] = page;
1713 if (gfpflags_allow_blocking(gfp_mask))
1714 cond_resched();
1715 }
1716
1717 if (map_vm_area(area, prot, pages))
1718 goto fail;
1719 return area->addr;
1720
1721 fail:
1722 warn_alloc(gfp_mask, NULL,
1723 "vmalloc: allocation failure, allocated %ld of %ld bytes",
1724 (area->nr_pages*PAGE_SIZE), area->size);
1725 fail_no_warn:
1726 vfree(area->addr);
1727 return NULL;
1728 }
1729
1730 /**
1731 * __vmalloc_node_range - allocate virtually contiguous memory
1732 * @size: allocation size
1733 * @align: desired alignment
1734 * @start: vm area range start
1735 * @end: vm area range end
1736 * @gfp_mask: flags for the page level allocator
1737 * @prot: protection mask for the allocated pages
1738 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1739 * @node: node to use for allocation or NUMA_NO_NODE
1740 * @caller: caller's return address
1741 *
1742 * Allocate enough pages to cover @size from the page level
1743 * allocator with @gfp_mask flags. Map them into contiguous
1744 * kernel virtual space, using a pagetable protection of @prot.
1745 */
1746 void *__vmalloc_node_range(unsigned long size, unsigned long align,
1747 unsigned long start, unsigned long end, gfp_t gfp_mask,
1748 pgprot_t prot, unsigned long vm_flags, int node,
1749 const void *caller)
1750 {
1751 struct vm_struct *area;
1752 void *addr;
1753 unsigned long real_size = size;
1754
1755 size = PAGE_ALIGN(size);
1756 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1757 goto fail;
1758
1759 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1760 vm_flags, start, end, node, gfp_mask, caller);
1761 if (!area)
1762 goto fail;
1763
1764 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1765 if (!addr)
1766 return NULL;
1767
1768 /*
1769 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1770 * flag. It means that vm_struct is not fully initialized.
1771 * Now, it is fully initialized, so remove this flag here.
1772 */
1773 clear_vm_uninitialized_flag(area);
1774
1775 kmemleak_vmalloc(area, size, gfp_mask);
1776
1777 return addr;
1778
1779 fail:
1780 warn_alloc(gfp_mask, NULL,
1781 "vmalloc: allocation failure: %lu bytes", real_size);
1782 return NULL;
1783 }
1784
1785 /**
1786 * __vmalloc_node - allocate virtually contiguous memory
1787 * @size: allocation size
1788 * @align: desired alignment
1789 * @gfp_mask: flags for the page level allocator
1790 * @prot: protection mask for the allocated pages
1791 * @node: node to use for allocation or NUMA_NO_NODE
1792 * @caller: caller's return address
1793 *
1794 * Allocate enough pages to cover @size from the page level
1795 * allocator with @gfp_mask flags. Map them into contiguous
1796 * kernel virtual space, using a pagetable protection of @prot.
1797 *
1798 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
1799 * and __GFP_NOFAIL are not supported
1800 *
1801 * Any use of gfp flags outside of GFP_KERNEL should be consulted
1802 * with mm people.
1803 *
1804 */
1805 static void *__vmalloc_node(unsigned long size, unsigned long align,
1806 gfp_t gfp_mask, pgprot_t prot,
1807 int node, const void *caller)
1808 {
1809 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1810 gfp_mask, prot, 0, node, caller);
1811 }
1812
1813 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1814 {
1815 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1816 __builtin_return_address(0));
1817 }
1818 EXPORT_SYMBOL(__vmalloc);
1819
1820 static inline void *__vmalloc_node_flags(unsigned long size,
1821 int node, gfp_t flags)
1822 {
1823 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1824 node, __builtin_return_address(0));
1825 }
1826
1827
1828 void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
1829 void *caller)
1830 {
1831 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
1832 }
1833
1834 /**
1835 * vmalloc - allocate virtually contiguous memory
1836 * @size: allocation size
1837 * Allocate enough pages to cover @size from the page level
1838 * allocator and map them into contiguous kernel virtual space.
1839 *
1840 * For tight control over page level allocator and protection flags
1841 * use __vmalloc() instead.
1842 */
1843 void *vmalloc(unsigned long size)
1844 {
1845 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1846 GFP_KERNEL);
1847 }
1848 EXPORT_SYMBOL(vmalloc);
1849
1850 /**
1851 * vzalloc - allocate virtually contiguous memory with zero fill
1852 * @size: allocation size
1853 * Allocate enough pages to cover @size from the page level
1854 * allocator and map them into contiguous kernel virtual space.
1855 * The memory allocated is set to zero.
1856 *
1857 * For tight control over page level allocator and protection flags
1858 * use __vmalloc() instead.
1859 */
1860 void *vzalloc(unsigned long size)
1861 {
1862 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1863 GFP_KERNEL | __GFP_ZERO);
1864 }
1865 EXPORT_SYMBOL(vzalloc);
1866
1867 /**
1868 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1869 * @size: allocation size
1870 *
1871 * The resulting memory area is zeroed so it can be mapped to userspace
1872 * without leaking data.
1873 */
1874 void *vmalloc_user(unsigned long size)
1875 {
1876 struct vm_struct *area;
1877 void *ret;
1878
1879 ret = __vmalloc_node(size, SHMLBA,
1880 GFP_KERNEL | __GFP_ZERO,
1881 PAGE_KERNEL, NUMA_NO_NODE,
1882 __builtin_return_address(0));
1883 if (ret) {
1884 area = find_vm_area(ret);
1885 area->flags |= VM_USERMAP;
1886 }
1887 return ret;
1888 }
1889 EXPORT_SYMBOL(vmalloc_user);
1890
1891 /**
1892 * vmalloc_node - allocate memory on a specific node
1893 * @size: allocation size
1894 * @node: numa node
1895 *
1896 * Allocate enough pages to cover @size from the page level
1897 * allocator and map them into contiguous kernel virtual space.
1898 *
1899 * For tight control over page level allocator and protection flags
1900 * use __vmalloc() instead.
1901 */
1902 void *vmalloc_node(unsigned long size, int node)
1903 {
1904 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
1905 node, __builtin_return_address(0));
1906 }
1907 EXPORT_SYMBOL(vmalloc_node);
1908
1909 /**
1910 * vzalloc_node - allocate memory on a specific node with zero fill
1911 * @size: allocation size
1912 * @node: numa node
1913 *
1914 * Allocate enough pages to cover @size from the page level
1915 * allocator and map them into contiguous kernel virtual space.
1916 * The memory allocated is set to zero.
1917 *
1918 * For tight control over page level allocator and protection flags
1919 * use __vmalloc_node() instead.
1920 */
1921 void *vzalloc_node(unsigned long size, int node)
1922 {
1923 return __vmalloc_node_flags(size, node,
1924 GFP_KERNEL | __GFP_ZERO);
1925 }
1926 EXPORT_SYMBOL(vzalloc_node);
1927
1928 #ifndef PAGE_KERNEL_EXEC
1929 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1930 #endif
1931
1932 /**
1933 * vmalloc_exec - allocate virtually contiguous, executable memory
1934 * @size: allocation size
1935 *
1936 * Kernel-internal function to allocate enough pages to cover @size
1937 * the page level allocator and map them into contiguous and
1938 * executable kernel virtual space.
1939 *
1940 * For tight control over page level allocator and protection flags
1941 * use __vmalloc() instead.
1942 */
1943
1944 void *vmalloc_exec(unsigned long size)
1945 {
1946 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL_EXEC,
1947 NUMA_NO_NODE, __builtin_return_address(0));
1948 }
1949
1950 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1951 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1952 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1953 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1954 #else
1955 #define GFP_VMALLOC32 GFP_KERNEL
1956 #endif
1957
1958 /**
1959 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1960 * @size: allocation size
1961 *
1962 * Allocate enough 32bit PA addressable pages to cover @size from the
1963 * page level allocator and map them into contiguous kernel virtual space.
1964 */
1965 void *vmalloc_32(unsigned long size)
1966 {
1967 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1968 NUMA_NO_NODE, __builtin_return_address(0));
1969 }
1970 EXPORT_SYMBOL(vmalloc_32);
1971
1972 /**
1973 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1974 * @size: allocation size
1975 *
1976 * The resulting memory area is 32bit addressable and zeroed so it can be
1977 * mapped to userspace without leaking data.
1978 */
1979 void *vmalloc_32_user(unsigned long size)
1980 {
1981 struct vm_struct *area;
1982 void *ret;
1983
1984 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1985 NUMA_NO_NODE, __builtin_return_address(0));
1986 if (ret) {
1987 area = find_vm_area(ret);
1988 area->flags |= VM_USERMAP;
1989 }
1990 return ret;
1991 }
1992 EXPORT_SYMBOL(vmalloc_32_user);
1993
1994 /*
1995 * small helper routine , copy contents to buf from addr.
1996 * If the page is not present, fill zero.
1997 */
1998
1999 static int aligned_vread(char *buf, char *addr, unsigned long count)
2000 {
2001 struct page *p;
2002 int copied = 0;
2003
2004 while (count) {
2005 unsigned long offset, length;
2006
2007 offset = offset_in_page(addr);
2008 length = PAGE_SIZE - offset;
2009 if (length > count)
2010 length = count;
2011 p = vmalloc_to_page(addr);
2012 /*
2013 * To do safe access to this _mapped_ area, we need
2014 * lock. But adding lock here means that we need to add
2015 * overhead of vmalloc()/vfree() calles for this _debug_
2016 * interface, rarely used. Instead of that, we'll use
2017 * kmap() and get small overhead in this access function.
2018 */
2019 if (p) {
2020 /*
2021 * we can expect USER0 is not used (see vread/vwrite's
2022 * function description)
2023 */
2024 void *map = kmap_atomic(p);
2025 memcpy(buf, map + offset, length);
2026 kunmap_atomic(map);
2027 } else
2028 memset(buf, 0, length);
2029
2030 addr += length;
2031 buf += length;
2032 copied += length;
2033 count -= length;
2034 }
2035 return copied;
2036 }
2037
2038 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2039 {
2040 struct page *p;
2041 int copied = 0;
2042
2043 while (count) {
2044 unsigned long offset, length;
2045
2046 offset = offset_in_page(addr);
2047 length = PAGE_SIZE - offset;
2048 if (length > count)
2049 length = count;
2050 p = vmalloc_to_page(addr);
2051 /*
2052 * To do safe access to this _mapped_ area, we need
2053 * lock. But adding lock here means that we need to add
2054 * overhead of vmalloc()/vfree() calles for this _debug_
2055 * interface, rarely used. Instead of that, we'll use
2056 * kmap() and get small overhead in this access function.
2057 */
2058 if (p) {
2059 /*
2060 * we can expect USER0 is not used (see vread/vwrite's
2061 * function description)
2062 */
2063 void *map = kmap_atomic(p);
2064 memcpy(map + offset, buf, length);
2065 kunmap_atomic(map);
2066 }
2067 addr += length;
2068 buf += length;
2069 copied += length;
2070 count -= length;
2071 }
2072 return copied;
2073 }
2074
2075 /**
2076 * vread() - read vmalloc area in a safe way.
2077 * @buf: buffer for reading data
2078 * @addr: vm address.
2079 * @count: number of bytes to be read.
2080 *
2081 * Returns # of bytes which addr and buf should be increased.
2082 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2083 * includes any intersect with alive vmalloc area.
2084 *
2085 * This function checks that addr is a valid vmalloc'ed area, and
2086 * copy data from that area to a given buffer. If the given memory range
2087 * of [addr...addr+count) includes some valid address, data is copied to
2088 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2089 * IOREMAP area is treated as memory hole and no copy is done.
2090 *
2091 * If [addr...addr+count) doesn't includes any intersects with alive
2092 * vm_struct area, returns 0. @buf should be kernel's buffer.
2093 *
2094 * Note: In usual ops, vread() is never necessary because the caller
2095 * should know vmalloc() area is valid and can use memcpy().
2096 * This is for routines which have to access vmalloc area without
2097 * any informaion, as /dev/kmem.
2098 *
2099 */
2100
2101 long vread(char *buf, char *addr, unsigned long count)
2102 {
2103 struct vmap_area *va;
2104 struct vm_struct *vm;
2105 char *vaddr, *buf_start = buf;
2106 unsigned long buflen = count;
2107 unsigned long n;
2108
2109 /* Don't allow overflow */
2110 if ((unsigned long) addr + count < count)
2111 count = -(unsigned long) addr;
2112
2113 spin_lock(&vmap_area_lock);
2114 list_for_each_entry(va, &vmap_area_list, list) {
2115 if (!count)
2116 break;
2117
2118 if (!(va->flags & VM_VM_AREA))
2119 continue;
2120
2121 vm = va->vm;
2122 vaddr = (char *) vm->addr;
2123 if (addr >= vaddr + get_vm_area_size(vm))
2124 continue;
2125 while (addr < vaddr) {
2126 if (count == 0)
2127 goto finished;
2128 *buf = '\0';
2129 buf++;
2130 addr++;
2131 count--;
2132 }
2133 n = vaddr + get_vm_area_size(vm) - addr;
2134 if (n > count)
2135 n = count;
2136 if (!(vm->flags & VM_IOREMAP))
2137 aligned_vread(buf, addr, n);
2138 else /* IOREMAP area is treated as memory hole */
2139 memset(buf, 0, n);
2140 buf += n;
2141 addr += n;
2142 count -= n;
2143 }
2144 finished:
2145 spin_unlock(&vmap_area_lock);
2146
2147 if (buf == buf_start)
2148 return 0;
2149 /* zero-fill memory holes */
2150 if (buf != buf_start + buflen)
2151 memset(buf, 0, buflen - (buf - buf_start));
2152
2153 return buflen;
2154 }
2155
2156 /**
2157 * vwrite() - write vmalloc area in a safe way.
2158 * @buf: buffer for source data
2159 * @addr: vm address.
2160 * @count: number of bytes to be read.
2161 *
2162 * Returns # of bytes which addr and buf should be incresed.
2163 * (same number to @count).
2164 * If [addr...addr+count) doesn't includes any intersect with valid
2165 * vmalloc area, returns 0.
2166 *
2167 * This function checks that addr is a valid vmalloc'ed area, and
2168 * copy data from a buffer to the given addr. If specified range of
2169 * [addr...addr+count) includes some valid address, data is copied from
2170 * proper area of @buf. If there are memory holes, no copy to hole.
2171 * IOREMAP area is treated as memory hole and no copy is done.
2172 *
2173 * If [addr...addr+count) doesn't includes any intersects with alive
2174 * vm_struct area, returns 0. @buf should be kernel's buffer.
2175 *
2176 * Note: In usual ops, vwrite() is never necessary because the caller
2177 * should know vmalloc() area is valid and can use memcpy().
2178 * This is for routines which have to access vmalloc area without
2179 * any informaion, as /dev/kmem.
2180 */
2181
2182 long vwrite(char *buf, char *addr, unsigned long count)
2183 {
2184 struct vmap_area *va;
2185 struct vm_struct *vm;
2186 char *vaddr;
2187 unsigned long n, buflen;
2188 int copied = 0;
2189
2190 /* Don't allow overflow */
2191 if ((unsigned long) addr + count < count)
2192 count = -(unsigned long) addr;
2193 buflen = count;
2194
2195 spin_lock(&vmap_area_lock);
2196 list_for_each_entry(va, &vmap_area_list, list) {
2197 if (!count)
2198 break;
2199
2200 if (!(va->flags & VM_VM_AREA))
2201 continue;
2202
2203 vm = va->vm;
2204 vaddr = (char *) vm->addr;
2205 if (addr >= vaddr + get_vm_area_size(vm))
2206 continue;
2207 while (addr < vaddr) {
2208 if (count == 0)
2209 goto finished;
2210 buf++;
2211 addr++;
2212 count--;
2213 }
2214 n = vaddr + get_vm_area_size(vm) - addr;
2215 if (n > count)
2216 n = count;
2217 if (!(vm->flags & VM_IOREMAP)) {
2218 aligned_vwrite(buf, addr, n);
2219 copied++;
2220 }
2221 buf += n;
2222 addr += n;
2223 count -= n;
2224 }
2225 finished:
2226 spin_unlock(&vmap_area_lock);
2227 if (!copied)
2228 return 0;
2229 return buflen;
2230 }
2231
2232 /**
2233 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2234 * @vma: vma to cover
2235 * @uaddr: target user address to start at
2236 * @kaddr: virtual address of vmalloc kernel memory
2237 * @size: size of map area
2238 *
2239 * Returns: 0 for success, -Exxx on failure
2240 *
2241 * This function checks that @kaddr is a valid vmalloc'ed area,
2242 * and that it is big enough to cover the range starting at
2243 * @uaddr in @vma. Will return failure if that criteria isn't
2244 * met.
2245 *
2246 * Similar to remap_pfn_range() (see mm/memory.c)
2247 */
2248 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2249 void *kaddr, unsigned long size)
2250 {
2251 struct vm_struct *area;
2252
2253 size = PAGE_ALIGN(size);
2254
2255 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2256 return -EINVAL;
2257
2258 area = find_vm_area(kaddr);
2259 if (!area)
2260 return -EINVAL;
2261
2262 if (!(area->flags & VM_USERMAP))
2263 return -EINVAL;
2264
2265 if (kaddr + size > area->addr + area->size)
2266 return -EINVAL;
2267
2268 do {
2269 struct page *page = vmalloc_to_page(kaddr);
2270 int ret;
2271
2272 ret = vm_insert_page(vma, uaddr, page);
2273 if (ret)
2274 return ret;
2275
2276 uaddr += PAGE_SIZE;
2277 kaddr += PAGE_SIZE;
2278 size -= PAGE_SIZE;
2279 } while (size > 0);
2280
2281 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2282
2283 return 0;
2284 }
2285 EXPORT_SYMBOL(remap_vmalloc_range_partial);
2286
2287 /**
2288 * remap_vmalloc_range - map vmalloc pages to userspace
2289 * @vma: vma to cover (map full range of vma)
2290 * @addr: vmalloc memory
2291 * @pgoff: number of pages into addr before first page to map
2292 *
2293 * Returns: 0 for success, -Exxx on failure
2294 *
2295 * This function checks that addr is a valid vmalloc'ed area, and
2296 * that it is big enough to cover the vma. Will return failure if
2297 * that criteria isn't met.
2298 *
2299 * Similar to remap_pfn_range() (see mm/memory.c)
2300 */
2301 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2302 unsigned long pgoff)
2303 {
2304 return remap_vmalloc_range_partial(vma, vma->vm_start,
2305 addr + (pgoff << PAGE_SHIFT),
2306 vma->vm_end - vma->vm_start);
2307 }
2308 EXPORT_SYMBOL(remap_vmalloc_range);
2309
2310 /*
2311 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2312 * have one.
2313 */
2314 void __weak vmalloc_sync_all(void)
2315 {
2316 }
2317
2318
2319 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2320 {
2321 pte_t ***p = data;
2322
2323 if (p) {
2324 *(*p) = pte;
2325 (*p)++;
2326 }
2327 return 0;
2328 }
2329
2330 /**
2331 * alloc_vm_area - allocate a range of kernel address space
2332 * @size: size of the area
2333 * @ptes: returns the PTEs for the address space
2334 *
2335 * Returns: NULL on failure, vm_struct on success
2336 *
2337 * This function reserves a range of kernel address space, and
2338 * allocates pagetables to map that range. No actual mappings
2339 * are created.
2340 *
2341 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2342 * allocated for the VM area are returned.
2343 */
2344 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2345 {
2346 struct vm_struct *area;
2347
2348 area = get_vm_area_caller(size, VM_IOREMAP,
2349 __builtin_return_address(0));
2350 if (area == NULL)
2351 return NULL;
2352
2353 /*
2354 * This ensures that page tables are constructed for this region
2355 * of kernel virtual address space and mapped into init_mm.
2356 */
2357 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2358 size, f, ptes ? &ptes : NULL)) {
2359 free_vm_area(area);
2360 return NULL;
2361 }
2362
2363 return area;
2364 }
2365 EXPORT_SYMBOL_GPL(alloc_vm_area);
2366
2367 void free_vm_area(struct vm_struct *area)
2368 {
2369 struct vm_struct *ret;
2370 ret = remove_vm_area(area->addr);
2371 BUG_ON(ret != area);
2372 kfree(area);
2373 }
2374 EXPORT_SYMBOL_GPL(free_vm_area);
2375
2376 #ifdef CONFIG_SMP
2377 static struct vmap_area *node_to_va(struct rb_node *n)
2378 {
2379 return rb_entry_safe(n, struct vmap_area, rb_node);
2380 }
2381
2382 /**
2383 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2384 * @end: target address
2385 * @pnext: out arg for the next vmap_area
2386 * @pprev: out arg for the previous vmap_area
2387 *
2388 * Returns: %true if either or both of next and prev are found,
2389 * %false if no vmap_area exists
2390 *
2391 * Find vmap_areas end addresses of which enclose @end. ie. if not
2392 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2393 */
2394 static bool pvm_find_next_prev(unsigned long end,
2395 struct vmap_area **pnext,
2396 struct vmap_area **pprev)
2397 {
2398 struct rb_node *n = vmap_area_root.rb_node;
2399 struct vmap_area *va = NULL;
2400
2401 while (n) {
2402 va = rb_entry(n, struct vmap_area, rb_node);
2403 if (end < va->va_end)
2404 n = n->rb_left;
2405 else if (end > va->va_end)
2406 n = n->rb_right;
2407 else
2408 break;
2409 }
2410
2411 if (!va)
2412 return false;
2413
2414 if (va->va_end > end) {
2415 *pnext = va;
2416 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2417 } else {
2418 *pprev = va;
2419 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2420 }
2421 return true;
2422 }
2423
2424 /**
2425 * pvm_determine_end - find the highest aligned address between two vmap_areas
2426 * @pnext: in/out arg for the next vmap_area
2427 * @pprev: in/out arg for the previous vmap_area
2428 * @align: alignment
2429 *
2430 * Returns: determined end address
2431 *
2432 * Find the highest aligned address between *@pnext and *@pprev below
2433 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2434 * down address is between the end addresses of the two vmap_areas.
2435 *
2436 * Please note that the address returned by this function may fall
2437 * inside *@pnext vmap_area. The caller is responsible for checking
2438 * that.
2439 */
2440 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2441 struct vmap_area **pprev,
2442 unsigned long align)
2443 {
2444 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2445 unsigned long addr;
2446
2447 if (*pnext)
2448 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2449 else
2450 addr = vmalloc_end;
2451
2452 while (*pprev && (*pprev)->va_end > addr) {
2453 *pnext = *pprev;
2454 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2455 }
2456
2457 return addr;
2458 }
2459
2460 /**
2461 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2462 * @offsets: array containing offset of each area
2463 * @sizes: array containing size of each area
2464 * @nr_vms: the number of areas to allocate
2465 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2466 *
2467 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2468 * vm_structs on success, %NULL on failure
2469 *
2470 * Percpu allocator wants to use congruent vm areas so that it can
2471 * maintain the offsets among percpu areas. This function allocates
2472 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2473 * be scattered pretty far, distance between two areas easily going up
2474 * to gigabytes. To avoid interacting with regular vmallocs, these
2475 * areas are allocated from top.
2476 *
2477 * Despite its complicated look, this allocator is rather simple. It
2478 * does everything top-down and scans areas from the end looking for
2479 * matching slot. While scanning, if any of the areas overlaps with
2480 * existing vmap_area, the base address is pulled down to fit the
2481 * area. Scanning is repeated till all the areas fit and then all
2482 * necessary data structres are inserted and the result is returned.
2483 */
2484 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2485 const size_t *sizes, int nr_vms,
2486 size_t align)
2487 {
2488 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2489 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2490 struct vmap_area **vas, *prev, *next;
2491 struct vm_struct **vms;
2492 int area, area2, last_area, term_area;
2493 unsigned long base, start, end, last_end;
2494 bool purged = false;
2495
2496 /* verify parameters and allocate data structures */
2497 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2498 for (last_area = 0, area = 0; area < nr_vms; area++) {
2499 start = offsets[area];
2500 end = start + sizes[area];
2501
2502 /* is everything aligned properly? */
2503 BUG_ON(!IS_ALIGNED(offsets[area], align));
2504 BUG_ON(!IS_ALIGNED(sizes[area], align));
2505
2506 /* detect the area with the highest address */
2507 if (start > offsets[last_area])
2508 last_area = area;
2509
2510 for (area2 = 0; area2 < nr_vms; area2++) {
2511 unsigned long start2 = offsets[area2];
2512 unsigned long end2 = start2 + sizes[area2];
2513
2514 if (area2 == area)
2515 continue;
2516
2517 BUG_ON(start2 >= start && start2 < end);
2518 BUG_ON(end2 <= end && end2 > start);
2519 }
2520 }
2521 last_end = offsets[last_area] + sizes[last_area];
2522
2523 if (vmalloc_end - vmalloc_start < last_end) {
2524 WARN_ON(true);
2525 return NULL;
2526 }
2527
2528 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2529 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2530 if (!vas || !vms)
2531 goto err_free2;
2532
2533 for (area = 0; area < nr_vms; area++) {
2534 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2535 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2536 if (!vas[area] || !vms[area])
2537 goto err_free;
2538 }
2539 retry:
2540 spin_lock(&vmap_area_lock);
2541
2542 /* start scanning - we scan from the top, begin with the last area */
2543 area = term_area = last_area;
2544 start = offsets[area];
2545 end = start + sizes[area];
2546
2547 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2548 base = vmalloc_end - last_end;
2549 goto found;
2550 }
2551 base = pvm_determine_end(&next, &prev, align) - end;
2552
2553 while (true) {
2554 BUG_ON(next && next->va_end <= base + end);
2555 BUG_ON(prev && prev->va_end > base + end);
2556
2557 /*
2558 * base might have underflowed, add last_end before
2559 * comparing.
2560 */
2561 if (base + last_end < vmalloc_start + last_end) {
2562 spin_unlock(&vmap_area_lock);
2563 if (!purged) {
2564 purge_vmap_area_lazy();
2565 purged = true;
2566 goto retry;
2567 }
2568 goto err_free;
2569 }
2570
2571 /*
2572 * If next overlaps, move base downwards so that it's
2573 * right below next and then recheck.
2574 */
2575 if (next && next->va_start < base + end) {
2576 base = pvm_determine_end(&next, &prev, align) - end;
2577 term_area = area;
2578 continue;
2579 }
2580
2581 /*
2582 * If prev overlaps, shift down next and prev and move
2583 * base so that it's right below new next and then
2584 * recheck.
2585 */
2586 if (prev && prev->va_end > base + start) {
2587 next = prev;
2588 prev = node_to_va(rb_prev(&next->rb_node));
2589 base = pvm_determine_end(&next, &prev, align) - end;
2590 term_area = area;
2591 continue;
2592 }
2593
2594 /*
2595 * This area fits, move on to the previous one. If
2596 * the previous one is the terminal one, we're done.
2597 */
2598 area = (area + nr_vms - 1) % nr_vms;
2599 if (area == term_area)
2600 break;
2601 start = offsets[area];
2602 end = start + sizes[area];
2603 pvm_find_next_prev(base + end, &next, &prev);
2604 }
2605 found:
2606 /* we've found a fitting base, insert all va's */
2607 for (area = 0; area < nr_vms; area++) {
2608 struct vmap_area *va = vas[area];
2609
2610 va->va_start = base + offsets[area];
2611 va->va_end = va->va_start + sizes[area];
2612 __insert_vmap_area(va);
2613 }
2614
2615 vmap_area_pcpu_hole = base + offsets[last_area];
2616
2617 spin_unlock(&vmap_area_lock);
2618
2619 /* insert all vm's */
2620 for (area = 0; area < nr_vms; area++)
2621 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2622 pcpu_get_vm_areas);
2623
2624 kfree(vas);
2625 return vms;
2626
2627 err_free:
2628 for (area = 0; area < nr_vms; area++) {
2629 kfree(vas[area]);
2630 kfree(vms[area]);
2631 }
2632 err_free2:
2633 kfree(vas);
2634 kfree(vms);
2635 return NULL;
2636 }
2637
2638 /**
2639 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2640 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2641 * @nr_vms: the number of allocated areas
2642 *
2643 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2644 */
2645 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2646 {
2647 int i;
2648
2649 for (i = 0; i < nr_vms; i++)
2650 free_vm_area(vms[i]);
2651 kfree(vms);
2652 }
2653 #endif /* CONFIG_SMP */
2654
2655 #ifdef CONFIG_PROC_FS
2656 static void *s_start(struct seq_file *m, loff_t *pos)
2657 __acquires(&vmap_area_lock)
2658 {
2659 spin_lock(&vmap_area_lock);
2660 return seq_list_start(&vmap_area_list, *pos);
2661 }
2662
2663 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2664 {
2665 return seq_list_next(p, &vmap_area_list, pos);
2666 }
2667
2668 static void s_stop(struct seq_file *m, void *p)
2669 __releases(&vmap_area_lock)
2670 {
2671 spin_unlock(&vmap_area_lock);
2672 }
2673
2674 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2675 {
2676 if (IS_ENABLED(CONFIG_NUMA)) {
2677 unsigned int nr, *counters = m->private;
2678
2679 if (!counters)
2680 return;
2681
2682 if (v->flags & VM_UNINITIALIZED)
2683 return;
2684 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2685 smp_rmb();
2686
2687 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2688
2689 for (nr = 0; nr < v->nr_pages; nr++)
2690 counters[page_to_nid(v->pages[nr])]++;
2691
2692 for_each_node_state(nr, N_HIGH_MEMORY)
2693 if (counters[nr])
2694 seq_printf(m, " N%u=%u", nr, counters[nr]);
2695 }
2696 }
2697
2698 static int s_show(struct seq_file *m, void *p)
2699 {
2700 struct vmap_area *va;
2701 struct vm_struct *v;
2702
2703 va = list_entry(p, struct vmap_area, list);
2704
2705 /*
2706 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2707 * behalf of vmap area is being tear down or vm_map_ram allocation.
2708 */
2709 if (!(va->flags & VM_VM_AREA)) {
2710 seq_printf(m, "0x%pK-0x%pK %7ld %s\n",
2711 (void *)va->va_start, (void *)va->va_end,
2712 va->va_end - va->va_start,
2713 va->flags & VM_LAZY_FREE ? "unpurged vm_area" : "vm_map_ram");
2714
2715 return 0;
2716 }
2717
2718 v = va->vm;
2719
2720 seq_printf(m, "0x%pK-0x%pK %7ld",
2721 v->addr, v->addr + v->size, v->size);
2722
2723 if (v->caller)
2724 seq_printf(m, " %pS", v->caller);
2725
2726 if (v->nr_pages)
2727 seq_printf(m, " pages=%d", v->nr_pages);
2728
2729 if (v->phys_addr)
2730 seq_printf(m, " phys=%pa", &v->phys_addr);
2731
2732 if (v->flags & VM_IOREMAP)
2733 seq_puts(m, " ioremap");
2734
2735 if (v->flags & VM_ALLOC)
2736 seq_puts(m, " vmalloc");
2737
2738 if (v->flags & VM_MAP)
2739 seq_puts(m, " vmap");
2740
2741 if (v->flags & VM_USERMAP)
2742 seq_puts(m, " user");
2743
2744 if (is_vmalloc_addr(v->pages))
2745 seq_puts(m, " vpages");
2746
2747 show_numa_info(m, v);
2748 seq_putc(m, '\n');
2749 return 0;
2750 }
2751
2752 static const struct seq_operations vmalloc_op = {
2753 .start = s_start,
2754 .next = s_next,
2755 .stop = s_stop,
2756 .show = s_show,
2757 };
2758
2759 static int vmalloc_open(struct inode *inode, struct file *file)
2760 {
2761 if (IS_ENABLED(CONFIG_NUMA))
2762 return seq_open_private(file, &vmalloc_op,
2763 nr_node_ids * sizeof(unsigned int));
2764 else
2765 return seq_open(file, &vmalloc_op);
2766 }
2767
2768 static const struct file_operations proc_vmalloc_operations = {
2769 .open = vmalloc_open,
2770 .read = seq_read,
2771 .llseek = seq_lseek,
2772 .release = seq_release_private,
2773 };
2774
2775 static int __init proc_vmalloc_init(void)
2776 {
2777 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2778 return 0;
2779 }
2780 module_init(proc_vmalloc_init);
2781
2782 #endif
2783