]> git.proxmox.com Git - mirror_ubuntu-focal-kernel.git/blob - mm/vmalloc.c
Merge tag 'v4.6-rc3' into drm-intel-next-queued
[mirror_ubuntu-focal-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.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 <asm/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(pgd_t *pgd, unsigned long addr, unsigned long end)
90 {
91 pud_t *pud;
92 unsigned long next;
93
94 pud = pud_offset(pgd, addr);
95 do {
96 next = pud_addr_end(addr, end);
97 if (pud_clear_huge(pud))
98 continue;
99 if (pud_none_or_clear_bad(pud))
100 continue;
101 vunmap_pmd_range(pud, addr, next);
102 } while (pud++, addr = next, addr != end);
103 }
104
105 static void vunmap_page_range(unsigned long addr, unsigned long end)
106 {
107 pgd_t *pgd;
108 unsigned long next;
109
110 BUG_ON(addr >= end);
111 pgd = pgd_offset_k(addr);
112 do {
113 next = pgd_addr_end(addr, end);
114 if (pgd_none_or_clear_bad(pgd))
115 continue;
116 vunmap_pud_range(pgd, addr, next);
117 } while (pgd++, addr = next, addr != end);
118 }
119
120 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
121 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
122 {
123 pte_t *pte;
124
125 /*
126 * nr is a running index into the array which helps higher level
127 * callers keep track of where we're up to.
128 */
129
130 pte = pte_alloc_kernel(pmd, addr);
131 if (!pte)
132 return -ENOMEM;
133 do {
134 struct page *page = pages[*nr];
135
136 if (WARN_ON(!pte_none(*pte)))
137 return -EBUSY;
138 if (WARN_ON(!page))
139 return -ENOMEM;
140 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
141 (*nr)++;
142 } while (pte++, addr += PAGE_SIZE, addr != end);
143 return 0;
144 }
145
146 static int vmap_pmd_range(pud_t *pud, unsigned long addr,
147 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
148 {
149 pmd_t *pmd;
150 unsigned long next;
151
152 pmd = pmd_alloc(&init_mm, pud, addr);
153 if (!pmd)
154 return -ENOMEM;
155 do {
156 next = pmd_addr_end(addr, end);
157 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
158 return -ENOMEM;
159 } while (pmd++, addr = next, addr != end);
160 return 0;
161 }
162
163 static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
165 {
166 pud_t *pud;
167 unsigned long next;
168
169 pud = pud_alloc(&init_mm, pgd, addr);
170 if (!pud)
171 return -ENOMEM;
172 do {
173 next = pud_addr_end(addr, end);
174 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
175 return -ENOMEM;
176 } while (pud++, addr = next, addr != end);
177 return 0;
178 }
179
180 /*
181 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
182 * will have pfns corresponding to the "pages" array.
183 *
184 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
185 */
186 static int vmap_page_range_noflush(unsigned long start, unsigned long end,
187 pgprot_t prot, struct page **pages)
188 {
189 pgd_t *pgd;
190 unsigned long next;
191 unsigned long addr = start;
192 int err = 0;
193 int nr = 0;
194
195 BUG_ON(addr >= end);
196 pgd = pgd_offset_k(addr);
197 do {
198 next = pgd_addr_end(addr, end);
199 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
200 if (err)
201 return err;
202 } while (pgd++, addr = next, addr != end);
203
204 return nr;
205 }
206
207 static int vmap_page_range(unsigned long start, unsigned long end,
208 pgprot_t prot, struct page **pages)
209 {
210 int ret;
211
212 ret = vmap_page_range_noflush(start, end, prot, pages);
213 flush_cache_vmap(start, end);
214 return ret;
215 }
216
217 int is_vmalloc_or_module_addr(const void *x)
218 {
219 /*
220 * ARM, x86-64 and sparc64 put modules in a special place,
221 * and fall back on vmalloc() if that fails. Others
222 * just put it in the vmalloc space.
223 */
224 #if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
225 unsigned long addr = (unsigned long)x;
226 if (addr >= MODULES_VADDR && addr < MODULES_END)
227 return 1;
228 #endif
229 return is_vmalloc_addr(x);
230 }
231
232 /*
233 * Walk a vmap address to the struct page it maps.
234 */
235 struct page *vmalloc_to_page(const void *vmalloc_addr)
236 {
237 unsigned long addr = (unsigned long) vmalloc_addr;
238 struct page *page = NULL;
239 pgd_t *pgd = pgd_offset_k(addr);
240
241 /*
242 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
243 * architectures that do not vmalloc module space
244 */
245 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
246
247 if (!pgd_none(*pgd)) {
248 pud_t *pud = pud_offset(pgd, addr);
249 if (!pud_none(*pud)) {
250 pmd_t *pmd = pmd_offset(pud, addr);
251 if (!pmd_none(*pmd)) {
252 pte_t *ptep, pte;
253
254 ptep = pte_offset_map(pmd, addr);
255 pte = *ptep;
256 if (pte_present(pte))
257 page = pte_page(pte);
258 pte_unmap(ptep);
259 }
260 }
261 }
262 return page;
263 }
264 EXPORT_SYMBOL(vmalloc_to_page);
265
266 /*
267 * Map a vmalloc()-space virtual address to the physical page frame number.
268 */
269 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
270 {
271 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
272 }
273 EXPORT_SYMBOL(vmalloc_to_pfn);
274
275
276 /*** Global kva allocator ***/
277
278 #define VM_LAZY_FREE 0x01
279 #define VM_LAZY_FREEING 0x02
280 #define VM_VM_AREA 0x04
281
282 static DEFINE_SPINLOCK(vmap_area_lock);
283 /* Export for kexec only */
284 LIST_HEAD(vmap_area_list);
285 static struct rb_root vmap_area_root = RB_ROOT;
286
287 /* The vmap cache globals are protected by vmap_area_lock */
288 static struct rb_node *free_vmap_cache;
289 static unsigned long cached_hole_size;
290 static unsigned long cached_vstart;
291 static unsigned long cached_align;
292
293 static unsigned long vmap_area_pcpu_hole;
294
295 static struct vmap_area *__find_vmap_area(unsigned long addr)
296 {
297 struct rb_node *n = vmap_area_root.rb_node;
298
299 while (n) {
300 struct vmap_area *va;
301
302 va = rb_entry(n, struct vmap_area, rb_node);
303 if (addr < va->va_start)
304 n = n->rb_left;
305 else if (addr >= va->va_end)
306 n = n->rb_right;
307 else
308 return va;
309 }
310
311 return NULL;
312 }
313
314 static void __insert_vmap_area(struct vmap_area *va)
315 {
316 struct rb_node **p = &vmap_area_root.rb_node;
317 struct rb_node *parent = NULL;
318 struct rb_node *tmp;
319
320 while (*p) {
321 struct vmap_area *tmp_va;
322
323 parent = *p;
324 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
325 if (va->va_start < tmp_va->va_end)
326 p = &(*p)->rb_left;
327 else if (va->va_end > tmp_va->va_start)
328 p = &(*p)->rb_right;
329 else
330 BUG();
331 }
332
333 rb_link_node(&va->rb_node, parent, p);
334 rb_insert_color(&va->rb_node, &vmap_area_root);
335
336 /* address-sort this list */
337 tmp = rb_prev(&va->rb_node);
338 if (tmp) {
339 struct vmap_area *prev;
340 prev = rb_entry(tmp, struct vmap_area, rb_node);
341 list_add_rcu(&va->list, &prev->list);
342 } else
343 list_add_rcu(&va->list, &vmap_area_list);
344 }
345
346 static void purge_vmap_area_lazy(void);
347
348 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
349
350 /*
351 * Allocate a region of KVA of the specified size and alignment, within the
352 * vstart and vend.
353 */
354 static struct vmap_area *alloc_vmap_area(unsigned long size,
355 unsigned long align,
356 unsigned long vstart, unsigned long vend,
357 int node, gfp_t gfp_mask)
358 {
359 struct vmap_area *va;
360 struct rb_node *n;
361 unsigned long addr;
362 int purged = 0;
363 struct vmap_area *first;
364
365 BUG_ON(!size);
366 BUG_ON(offset_in_page(size));
367 BUG_ON(!is_power_of_2(align));
368
369 might_sleep_if(gfpflags_allow_blocking(gfp_mask));
370
371 va = kmalloc_node(sizeof(struct vmap_area),
372 gfp_mask & GFP_RECLAIM_MASK, node);
373 if (unlikely(!va))
374 return ERR_PTR(-ENOMEM);
375
376 /*
377 * Only scan the relevant parts containing pointers to other objects
378 * to avoid false negatives.
379 */
380 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
381
382 retry:
383 spin_lock(&vmap_area_lock);
384 /*
385 * Invalidate cache if we have more permissive parameters.
386 * cached_hole_size notes the largest hole noticed _below_
387 * the vmap_area cached in free_vmap_cache: if size fits
388 * into that hole, we want to scan from vstart to reuse
389 * the hole instead of allocating above free_vmap_cache.
390 * Note that __free_vmap_area may update free_vmap_cache
391 * without updating cached_hole_size or cached_align.
392 */
393 if (!free_vmap_cache ||
394 size < cached_hole_size ||
395 vstart < cached_vstart ||
396 align < cached_align) {
397 nocache:
398 cached_hole_size = 0;
399 free_vmap_cache = NULL;
400 }
401 /* record if we encounter less permissive parameters */
402 cached_vstart = vstart;
403 cached_align = align;
404
405 /* find starting point for our search */
406 if (free_vmap_cache) {
407 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
408 addr = ALIGN(first->va_end, align);
409 if (addr < vstart)
410 goto nocache;
411 if (addr + size < addr)
412 goto overflow;
413
414 } else {
415 addr = ALIGN(vstart, align);
416 if (addr + size < addr)
417 goto overflow;
418
419 n = vmap_area_root.rb_node;
420 first = NULL;
421
422 while (n) {
423 struct vmap_area *tmp;
424 tmp = rb_entry(n, struct vmap_area, rb_node);
425 if (tmp->va_end >= addr) {
426 first = tmp;
427 if (tmp->va_start <= addr)
428 break;
429 n = n->rb_left;
430 } else
431 n = n->rb_right;
432 }
433
434 if (!first)
435 goto found;
436 }
437
438 /* from the starting point, walk areas until a suitable hole is found */
439 while (addr + size > first->va_start && addr + size <= vend) {
440 if (addr + cached_hole_size < first->va_start)
441 cached_hole_size = first->va_start - addr;
442 addr = ALIGN(first->va_end, align);
443 if (addr + size < addr)
444 goto overflow;
445
446 if (list_is_last(&first->list, &vmap_area_list))
447 goto found;
448
449 first = list_next_entry(first, list);
450 }
451
452 found:
453 if (addr + size > vend)
454 goto overflow;
455
456 va->va_start = addr;
457 va->va_end = addr + size;
458 va->flags = 0;
459 __insert_vmap_area(va);
460 free_vmap_cache = &va->rb_node;
461 spin_unlock(&vmap_area_lock);
462
463 BUG_ON(!IS_ALIGNED(va->va_start, align));
464 BUG_ON(va->va_start < vstart);
465 BUG_ON(va->va_end > vend);
466
467 return va;
468
469 overflow:
470 spin_unlock(&vmap_area_lock);
471 if (!purged) {
472 purge_vmap_area_lazy();
473 purged = 1;
474 goto retry;
475 }
476
477 if (gfpflags_allow_blocking(gfp_mask)) {
478 unsigned long freed = 0;
479 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
480 if (freed > 0) {
481 purged = 0;
482 goto retry;
483 }
484 }
485
486 if (printk_ratelimit())
487 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
488 size);
489 kfree(va);
490 return ERR_PTR(-EBUSY);
491 }
492
493 int register_vmap_purge_notifier(struct notifier_block *nb)
494 {
495 return blocking_notifier_chain_register(&vmap_notify_list, nb);
496 }
497 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
498
499 int unregister_vmap_purge_notifier(struct notifier_block *nb)
500 {
501 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
502 }
503 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
504
505 static void __free_vmap_area(struct vmap_area *va)
506 {
507 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
508
509 if (free_vmap_cache) {
510 if (va->va_end < cached_vstart) {
511 free_vmap_cache = NULL;
512 } else {
513 struct vmap_area *cache;
514 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
515 if (va->va_start <= cache->va_start) {
516 free_vmap_cache = rb_prev(&va->rb_node);
517 /*
518 * We don't try to update cached_hole_size or
519 * cached_align, but it won't go very wrong.
520 */
521 }
522 }
523 }
524 rb_erase(&va->rb_node, &vmap_area_root);
525 RB_CLEAR_NODE(&va->rb_node);
526 list_del_rcu(&va->list);
527
528 /*
529 * Track the highest possible candidate for pcpu area
530 * allocation. Areas outside of vmalloc area can be returned
531 * here too, consider only end addresses which fall inside
532 * vmalloc area proper.
533 */
534 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
535 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
536
537 kfree_rcu(va, rcu_head);
538 }
539
540 /*
541 * Free a region of KVA allocated by alloc_vmap_area
542 */
543 static void free_vmap_area(struct vmap_area *va)
544 {
545 spin_lock(&vmap_area_lock);
546 __free_vmap_area(va);
547 spin_unlock(&vmap_area_lock);
548 }
549
550 /*
551 * Clear the pagetable entries of a given vmap_area
552 */
553 static void unmap_vmap_area(struct vmap_area *va)
554 {
555 vunmap_page_range(va->va_start, va->va_end);
556 }
557
558 static void vmap_debug_free_range(unsigned long start, unsigned long end)
559 {
560 /*
561 * Unmap page tables and force a TLB flush immediately if pagealloc
562 * debugging is enabled. This catches use after free bugs similarly to
563 * those in linear kernel virtual address space after a page has been
564 * freed.
565 *
566 * All the lazy freeing logic is still retained, in order to minimise
567 * intrusiveness of this debugging feature.
568 *
569 * This is going to be *slow* (linear kernel virtual address debugging
570 * doesn't do a broadcast TLB flush so it is a lot faster).
571 */
572 if (debug_pagealloc_enabled()) {
573 vunmap_page_range(start, end);
574 flush_tlb_kernel_range(start, end);
575 }
576 }
577
578 /*
579 * lazy_max_pages is the maximum amount of virtual address space we gather up
580 * before attempting to purge with a TLB flush.
581 *
582 * There is a tradeoff here: a larger number will cover more kernel page tables
583 * and take slightly longer to purge, but it will linearly reduce the number of
584 * global TLB flushes that must be performed. It would seem natural to scale
585 * this number up linearly with the number of CPUs (because vmapping activity
586 * could also scale linearly with the number of CPUs), however it is likely
587 * that in practice, workloads might be constrained in other ways that mean
588 * vmap activity will not scale linearly with CPUs. Also, I want to be
589 * conservative and not introduce a big latency on huge systems, so go with
590 * a less aggressive log scale. It will still be an improvement over the old
591 * code, and it will be simple to change the scale factor if we find that it
592 * becomes a problem on bigger systems.
593 */
594 static unsigned long lazy_max_pages(void)
595 {
596 unsigned int log;
597
598 log = fls(num_online_cpus());
599
600 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
601 }
602
603 static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
604
605 /* for per-CPU blocks */
606 static void purge_fragmented_blocks_allcpus(void);
607
608 /*
609 * called before a call to iounmap() if the caller wants vm_area_struct's
610 * immediately freed.
611 */
612 void set_iounmap_nonlazy(void)
613 {
614 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
615 }
616
617 /*
618 * Purges all lazily-freed vmap areas.
619 *
620 * If sync is 0 then don't purge if there is already a purge in progress.
621 * If force_flush is 1, then flush kernel TLBs between *start and *end even
622 * if we found no lazy vmap areas to unmap (callers can use this to optimise
623 * their own TLB flushing).
624 * Returns with *start = min(*start, lowest purged address)
625 * *end = max(*end, highest purged address)
626 */
627 static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
628 int sync, int force_flush)
629 {
630 static DEFINE_SPINLOCK(purge_lock);
631 LIST_HEAD(valist);
632 struct vmap_area *va;
633 struct vmap_area *n_va;
634 int nr = 0;
635
636 /*
637 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
638 * should not expect such behaviour. This just simplifies locking for
639 * the case that isn't actually used at the moment anyway.
640 */
641 if (!sync && !force_flush) {
642 if (!spin_trylock(&purge_lock))
643 return;
644 } else
645 spin_lock(&purge_lock);
646
647 if (sync)
648 purge_fragmented_blocks_allcpus();
649
650 rcu_read_lock();
651 list_for_each_entry_rcu(va, &vmap_area_list, list) {
652 if (va->flags & VM_LAZY_FREE) {
653 if (va->va_start < *start)
654 *start = va->va_start;
655 if (va->va_end > *end)
656 *end = va->va_end;
657 nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
658 list_add_tail(&va->purge_list, &valist);
659 va->flags |= VM_LAZY_FREEING;
660 va->flags &= ~VM_LAZY_FREE;
661 }
662 }
663 rcu_read_unlock();
664
665 if (nr)
666 atomic_sub(nr, &vmap_lazy_nr);
667
668 if (nr || force_flush)
669 flush_tlb_kernel_range(*start, *end);
670
671 if (nr) {
672 spin_lock(&vmap_area_lock);
673 list_for_each_entry_safe(va, n_va, &valist, purge_list)
674 __free_vmap_area(va);
675 spin_unlock(&vmap_area_lock);
676 }
677 spin_unlock(&purge_lock);
678 }
679
680 /*
681 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
682 * is already purging.
683 */
684 static void try_purge_vmap_area_lazy(void)
685 {
686 unsigned long start = ULONG_MAX, end = 0;
687
688 __purge_vmap_area_lazy(&start, &end, 0, 0);
689 }
690
691 /*
692 * Kick off a purge of the outstanding lazy areas.
693 */
694 static void purge_vmap_area_lazy(void)
695 {
696 unsigned long start = ULONG_MAX, end = 0;
697
698 __purge_vmap_area_lazy(&start, &end, 1, 0);
699 }
700
701 /*
702 * Free a vmap area, caller ensuring that the area has been unmapped
703 * and flush_cache_vunmap had been called for the correct range
704 * previously.
705 */
706 static void free_vmap_area_noflush(struct vmap_area *va)
707 {
708 va->flags |= VM_LAZY_FREE;
709 atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
710 if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
711 try_purge_vmap_area_lazy();
712 }
713
714 /*
715 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
716 * called for the correct range previously.
717 */
718 static void free_unmap_vmap_area_noflush(struct vmap_area *va)
719 {
720 unmap_vmap_area(va);
721 free_vmap_area_noflush(va);
722 }
723
724 /*
725 * Free and unmap a vmap area
726 */
727 static void free_unmap_vmap_area(struct vmap_area *va)
728 {
729 flush_cache_vunmap(va->va_start, va->va_end);
730 free_unmap_vmap_area_noflush(va);
731 }
732
733 static struct vmap_area *find_vmap_area(unsigned long addr)
734 {
735 struct vmap_area *va;
736
737 spin_lock(&vmap_area_lock);
738 va = __find_vmap_area(addr);
739 spin_unlock(&vmap_area_lock);
740
741 return va;
742 }
743
744 static void free_unmap_vmap_area_addr(unsigned long addr)
745 {
746 struct vmap_area *va;
747
748 va = find_vmap_area(addr);
749 BUG_ON(!va);
750 free_unmap_vmap_area(va);
751 }
752
753
754 /*** Per cpu kva allocator ***/
755
756 /*
757 * vmap space is limited especially on 32 bit architectures. Ensure there is
758 * room for at least 16 percpu vmap blocks per CPU.
759 */
760 /*
761 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
762 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
763 * instead (we just need a rough idea)
764 */
765 #if BITS_PER_LONG == 32
766 #define VMALLOC_SPACE (128UL*1024*1024)
767 #else
768 #define VMALLOC_SPACE (128UL*1024*1024*1024)
769 #endif
770
771 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
772 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
773 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
774 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
775 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
776 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
777 #define VMAP_BBMAP_BITS \
778 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
779 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
780 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
781
782 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
783
784 static bool vmap_initialized __read_mostly = false;
785
786 struct vmap_block_queue {
787 spinlock_t lock;
788 struct list_head free;
789 };
790
791 struct vmap_block {
792 spinlock_t lock;
793 struct vmap_area *va;
794 unsigned long free, dirty;
795 unsigned long dirty_min, dirty_max; /*< dirty range */
796 struct list_head free_list;
797 struct rcu_head rcu_head;
798 struct list_head purge;
799 };
800
801 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
802 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
803
804 /*
805 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
806 * in the free path. Could get rid of this if we change the API to return a
807 * "cookie" from alloc, to be passed to free. But no big deal yet.
808 */
809 static DEFINE_SPINLOCK(vmap_block_tree_lock);
810 static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
811
812 /*
813 * We should probably have a fallback mechanism to allocate virtual memory
814 * out of partially filled vmap blocks. However vmap block sizing should be
815 * fairly reasonable according to the vmalloc size, so it shouldn't be a
816 * big problem.
817 */
818
819 static unsigned long addr_to_vb_idx(unsigned long addr)
820 {
821 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
822 addr /= VMAP_BLOCK_SIZE;
823 return addr;
824 }
825
826 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
827 {
828 unsigned long addr;
829
830 addr = va_start + (pages_off << PAGE_SHIFT);
831 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
832 return (void *)addr;
833 }
834
835 /**
836 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
837 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
838 * @order: how many 2^order pages should be occupied in newly allocated block
839 * @gfp_mask: flags for the page level allocator
840 *
841 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
842 */
843 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
844 {
845 struct vmap_block_queue *vbq;
846 struct vmap_block *vb;
847 struct vmap_area *va;
848 unsigned long vb_idx;
849 int node, err;
850 void *vaddr;
851
852 node = numa_node_id();
853
854 vb = kmalloc_node(sizeof(struct vmap_block),
855 gfp_mask & GFP_RECLAIM_MASK, node);
856 if (unlikely(!vb))
857 return ERR_PTR(-ENOMEM);
858
859 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
860 VMALLOC_START, VMALLOC_END,
861 node, gfp_mask);
862 if (IS_ERR(va)) {
863 kfree(vb);
864 return ERR_CAST(va);
865 }
866
867 err = radix_tree_preload(gfp_mask);
868 if (unlikely(err)) {
869 kfree(vb);
870 free_vmap_area(va);
871 return ERR_PTR(err);
872 }
873
874 vaddr = vmap_block_vaddr(va->va_start, 0);
875 spin_lock_init(&vb->lock);
876 vb->va = va;
877 /* At least something should be left free */
878 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
879 vb->free = VMAP_BBMAP_BITS - (1UL << order);
880 vb->dirty = 0;
881 vb->dirty_min = VMAP_BBMAP_BITS;
882 vb->dirty_max = 0;
883 INIT_LIST_HEAD(&vb->free_list);
884
885 vb_idx = addr_to_vb_idx(va->va_start);
886 spin_lock(&vmap_block_tree_lock);
887 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
888 spin_unlock(&vmap_block_tree_lock);
889 BUG_ON(err);
890 radix_tree_preload_end();
891
892 vbq = &get_cpu_var(vmap_block_queue);
893 spin_lock(&vbq->lock);
894 list_add_tail_rcu(&vb->free_list, &vbq->free);
895 spin_unlock(&vbq->lock);
896 put_cpu_var(vmap_block_queue);
897
898 return vaddr;
899 }
900
901 static void free_vmap_block(struct vmap_block *vb)
902 {
903 struct vmap_block *tmp;
904 unsigned long vb_idx;
905
906 vb_idx = addr_to_vb_idx(vb->va->va_start);
907 spin_lock(&vmap_block_tree_lock);
908 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
909 spin_unlock(&vmap_block_tree_lock);
910 BUG_ON(tmp != vb);
911
912 free_vmap_area_noflush(vb->va);
913 kfree_rcu(vb, rcu_head);
914 }
915
916 static void purge_fragmented_blocks(int cpu)
917 {
918 LIST_HEAD(purge);
919 struct vmap_block *vb;
920 struct vmap_block *n_vb;
921 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
922
923 rcu_read_lock();
924 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
925
926 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
927 continue;
928
929 spin_lock(&vb->lock);
930 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
931 vb->free = 0; /* prevent further allocs after releasing lock */
932 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
933 vb->dirty_min = 0;
934 vb->dirty_max = VMAP_BBMAP_BITS;
935 spin_lock(&vbq->lock);
936 list_del_rcu(&vb->free_list);
937 spin_unlock(&vbq->lock);
938 spin_unlock(&vb->lock);
939 list_add_tail(&vb->purge, &purge);
940 } else
941 spin_unlock(&vb->lock);
942 }
943 rcu_read_unlock();
944
945 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
946 list_del(&vb->purge);
947 free_vmap_block(vb);
948 }
949 }
950
951 static void purge_fragmented_blocks_allcpus(void)
952 {
953 int cpu;
954
955 for_each_possible_cpu(cpu)
956 purge_fragmented_blocks(cpu);
957 }
958
959 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
960 {
961 struct vmap_block_queue *vbq;
962 struct vmap_block *vb;
963 void *vaddr = NULL;
964 unsigned int order;
965
966 BUG_ON(offset_in_page(size));
967 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
968 if (WARN_ON(size == 0)) {
969 /*
970 * Allocating 0 bytes isn't what caller wants since
971 * get_order(0) returns funny result. Just warn and terminate
972 * early.
973 */
974 return NULL;
975 }
976 order = get_order(size);
977
978 rcu_read_lock();
979 vbq = &get_cpu_var(vmap_block_queue);
980 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
981 unsigned long pages_off;
982
983 spin_lock(&vb->lock);
984 if (vb->free < (1UL << order)) {
985 spin_unlock(&vb->lock);
986 continue;
987 }
988
989 pages_off = VMAP_BBMAP_BITS - vb->free;
990 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
991 vb->free -= 1UL << order;
992 if (vb->free == 0) {
993 spin_lock(&vbq->lock);
994 list_del_rcu(&vb->free_list);
995 spin_unlock(&vbq->lock);
996 }
997
998 spin_unlock(&vb->lock);
999 break;
1000 }
1001
1002 put_cpu_var(vmap_block_queue);
1003 rcu_read_unlock();
1004
1005 /* Allocate new block if nothing was found */
1006 if (!vaddr)
1007 vaddr = new_vmap_block(order, gfp_mask);
1008
1009 return vaddr;
1010 }
1011
1012 static void vb_free(const void *addr, unsigned long size)
1013 {
1014 unsigned long offset;
1015 unsigned long vb_idx;
1016 unsigned int order;
1017 struct vmap_block *vb;
1018
1019 BUG_ON(offset_in_page(size));
1020 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1021
1022 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1023
1024 order = get_order(size);
1025
1026 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1027 offset >>= PAGE_SHIFT;
1028
1029 vb_idx = addr_to_vb_idx((unsigned long)addr);
1030 rcu_read_lock();
1031 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1032 rcu_read_unlock();
1033 BUG_ON(!vb);
1034
1035 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1036
1037 spin_lock(&vb->lock);
1038
1039 /* Expand dirty range */
1040 vb->dirty_min = min(vb->dirty_min, offset);
1041 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1042
1043 vb->dirty += 1UL << order;
1044 if (vb->dirty == VMAP_BBMAP_BITS) {
1045 BUG_ON(vb->free);
1046 spin_unlock(&vb->lock);
1047 free_vmap_block(vb);
1048 } else
1049 spin_unlock(&vb->lock);
1050 }
1051
1052 /**
1053 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1054 *
1055 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1056 * to amortize TLB flushing overheads. What this means is that any page you
1057 * have now, may, in a former life, have been mapped into kernel virtual
1058 * address by the vmap layer and so there might be some CPUs with TLB entries
1059 * still referencing that page (additional to the regular 1:1 kernel mapping).
1060 *
1061 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1062 * be sure that none of the pages we have control over will have any aliases
1063 * from the vmap layer.
1064 */
1065 void vm_unmap_aliases(void)
1066 {
1067 unsigned long start = ULONG_MAX, end = 0;
1068 int cpu;
1069 int flush = 0;
1070
1071 if (unlikely(!vmap_initialized))
1072 return;
1073
1074 for_each_possible_cpu(cpu) {
1075 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1076 struct vmap_block *vb;
1077
1078 rcu_read_lock();
1079 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1080 spin_lock(&vb->lock);
1081 if (vb->dirty) {
1082 unsigned long va_start = vb->va->va_start;
1083 unsigned long s, e;
1084
1085 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1086 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1087
1088 start = min(s, start);
1089 end = max(e, end);
1090
1091 flush = 1;
1092 }
1093 spin_unlock(&vb->lock);
1094 }
1095 rcu_read_unlock();
1096 }
1097
1098 __purge_vmap_area_lazy(&start, &end, 1, flush);
1099 }
1100 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1101
1102 /**
1103 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1104 * @mem: the pointer returned by vm_map_ram
1105 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1106 */
1107 void vm_unmap_ram(const void *mem, unsigned int count)
1108 {
1109 unsigned long size = count << PAGE_SHIFT;
1110 unsigned long addr = (unsigned long)mem;
1111
1112 BUG_ON(!addr);
1113 BUG_ON(addr < VMALLOC_START);
1114 BUG_ON(addr > VMALLOC_END);
1115 BUG_ON(!PAGE_ALIGNED(addr));
1116
1117 debug_check_no_locks_freed(mem, size);
1118 vmap_debug_free_range(addr, addr+size);
1119
1120 if (likely(count <= VMAP_MAX_ALLOC))
1121 vb_free(mem, size);
1122 else
1123 free_unmap_vmap_area_addr(addr);
1124 }
1125 EXPORT_SYMBOL(vm_unmap_ram);
1126
1127 /**
1128 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1129 * @pages: an array of pointers to the pages to be mapped
1130 * @count: number of pages
1131 * @node: prefer to allocate data structures on this node
1132 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1133 *
1134 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1135 * faster than vmap so it's good. But if you mix long-life and short-life
1136 * objects with vm_map_ram(), it could consume lots of address space through
1137 * fragmentation (especially on a 32bit machine). You could see failures in
1138 * the end. Please use this function for short-lived objects.
1139 *
1140 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1141 */
1142 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1143 {
1144 unsigned long size = count << PAGE_SHIFT;
1145 unsigned long addr;
1146 void *mem;
1147
1148 if (likely(count <= VMAP_MAX_ALLOC)) {
1149 mem = vb_alloc(size, GFP_KERNEL);
1150 if (IS_ERR(mem))
1151 return NULL;
1152 addr = (unsigned long)mem;
1153 } else {
1154 struct vmap_area *va;
1155 va = alloc_vmap_area(size, PAGE_SIZE,
1156 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1157 if (IS_ERR(va))
1158 return NULL;
1159
1160 addr = va->va_start;
1161 mem = (void *)addr;
1162 }
1163 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1164 vm_unmap_ram(mem, count);
1165 return NULL;
1166 }
1167 return mem;
1168 }
1169 EXPORT_SYMBOL(vm_map_ram);
1170
1171 static struct vm_struct *vmlist __initdata;
1172 /**
1173 * vm_area_add_early - add vmap area early during boot
1174 * @vm: vm_struct to add
1175 *
1176 * This function is used to add fixed kernel vm area to vmlist before
1177 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1178 * should contain proper values and the other fields should be zero.
1179 *
1180 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1181 */
1182 void __init vm_area_add_early(struct vm_struct *vm)
1183 {
1184 struct vm_struct *tmp, **p;
1185
1186 BUG_ON(vmap_initialized);
1187 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1188 if (tmp->addr >= vm->addr) {
1189 BUG_ON(tmp->addr < vm->addr + vm->size);
1190 break;
1191 } else
1192 BUG_ON(tmp->addr + tmp->size > vm->addr);
1193 }
1194 vm->next = *p;
1195 *p = vm;
1196 }
1197
1198 /**
1199 * vm_area_register_early - register vmap area early during boot
1200 * @vm: vm_struct to register
1201 * @align: requested alignment
1202 *
1203 * This function is used to register kernel vm area before
1204 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1205 * proper values on entry and other fields should be zero. On return,
1206 * vm->addr contains the allocated address.
1207 *
1208 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1209 */
1210 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1211 {
1212 static size_t vm_init_off __initdata;
1213 unsigned long addr;
1214
1215 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1216 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1217
1218 vm->addr = (void *)addr;
1219
1220 vm_area_add_early(vm);
1221 }
1222
1223 void __init vmalloc_init(void)
1224 {
1225 struct vmap_area *va;
1226 struct vm_struct *tmp;
1227 int i;
1228
1229 for_each_possible_cpu(i) {
1230 struct vmap_block_queue *vbq;
1231 struct vfree_deferred *p;
1232
1233 vbq = &per_cpu(vmap_block_queue, i);
1234 spin_lock_init(&vbq->lock);
1235 INIT_LIST_HEAD(&vbq->free);
1236 p = &per_cpu(vfree_deferred, i);
1237 init_llist_head(&p->list);
1238 INIT_WORK(&p->wq, free_work);
1239 }
1240
1241 /* Import existing vmlist entries. */
1242 for (tmp = vmlist; tmp; tmp = tmp->next) {
1243 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1244 va->flags = VM_VM_AREA;
1245 va->va_start = (unsigned long)tmp->addr;
1246 va->va_end = va->va_start + tmp->size;
1247 va->vm = tmp;
1248 __insert_vmap_area(va);
1249 }
1250
1251 vmap_area_pcpu_hole = VMALLOC_END;
1252
1253 vmap_initialized = true;
1254 }
1255
1256 /**
1257 * map_kernel_range_noflush - map kernel VM area with the specified pages
1258 * @addr: start of the VM area to map
1259 * @size: size of the VM area to map
1260 * @prot: page protection flags to use
1261 * @pages: pages to map
1262 *
1263 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1264 * specify should have been allocated using get_vm_area() and its
1265 * friends.
1266 *
1267 * NOTE:
1268 * This function does NOT do any cache flushing. The caller is
1269 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1270 * before calling this function.
1271 *
1272 * RETURNS:
1273 * The number of pages mapped on success, -errno on failure.
1274 */
1275 int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1276 pgprot_t prot, struct page **pages)
1277 {
1278 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1279 }
1280
1281 /**
1282 * unmap_kernel_range_noflush - unmap kernel VM area
1283 * @addr: start of the VM area to unmap
1284 * @size: size of the VM area to unmap
1285 *
1286 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1287 * specify should have been allocated using get_vm_area() and its
1288 * friends.
1289 *
1290 * NOTE:
1291 * This function does NOT do any cache flushing. The caller is
1292 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1293 * before calling this function and flush_tlb_kernel_range() after.
1294 */
1295 void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1296 {
1297 vunmap_page_range(addr, addr + size);
1298 }
1299 EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1300
1301 /**
1302 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1303 * @addr: start of the VM area to unmap
1304 * @size: size of the VM area to unmap
1305 *
1306 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1307 * the unmapping and tlb after.
1308 */
1309 void unmap_kernel_range(unsigned long addr, unsigned long size)
1310 {
1311 unsigned long end = addr + size;
1312
1313 flush_cache_vunmap(addr, end);
1314 vunmap_page_range(addr, end);
1315 flush_tlb_kernel_range(addr, end);
1316 }
1317 EXPORT_SYMBOL_GPL(unmap_kernel_range);
1318
1319 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1320 {
1321 unsigned long addr = (unsigned long)area->addr;
1322 unsigned long end = addr + get_vm_area_size(area);
1323 int err;
1324
1325 err = vmap_page_range(addr, end, prot, pages);
1326
1327 return err > 0 ? 0 : err;
1328 }
1329 EXPORT_SYMBOL_GPL(map_vm_area);
1330
1331 static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1332 unsigned long flags, const void *caller)
1333 {
1334 spin_lock(&vmap_area_lock);
1335 vm->flags = flags;
1336 vm->addr = (void *)va->va_start;
1337 vm->size = va->va_end - va->va_start;
1338 vm->caller = caller;
1339 va->vm = vm;
1340 va->flags |= VM_VM_AREA;
1341 spin_unlock(&vmap_area_lock);
1342 }
1343
1344 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1345 {
1346 /*
1347 * Before removing VM_UNINITIALIZED,
1348 * we should make sure that vm has proper values.
1349 * Pair with smp_rmb() in show_numa_info().
1350 */
1351 smp_wmb();
1352 vm->flags &= ~VM_UNINITIALIZED;
1353 }
1354
1355 static struct vm_struct *__get_vm_area_node(unsigned long size,
1356 unsigned long align, unsigned long flags, unsigned long start,
1357 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1358 {
1359 struct vmap_area *va;
1360 struct vm_struct *area;
1361
1362 BUG_ON(in_interrupt());
1363 if (flags & VM_IOREMAP)
1364 align = 1ul << clamp_t(int, fls_long(size),
1365 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1366
1367 size = PAGE_ALIGN(size);
1368 if (unlikely(!size))
1369 return NULL;
1370
1371 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1372 if (unlikely(!area))
1373 return NULL;
1374
1375 if (!(flags & VM_NO_GUARD))
1376 size += PAGE_SIZE;
1377
1378 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1379 if (IS_ERR(va)) {
1380 kfree(area);
1381 return NULL;
1382 }
1383
1384 setup_vmalloc_vm(area, va, flags, caller);
1385
1386 return area;
1387 }
1388
1389 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1390 unsigned long start, unsigned long end)
1391 {
1392 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1393 GFP_KERNEL, __builtin_return_address(0));
1394 }
1395 EXPORT_SYMBOL_GPL(__get_vm_area);
1396
1397 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1398 unsigned long start, unsigned long end,
1399 const void *caller)
1400 {
1401 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1402 GFP_KERNEL, caller);
1403 }
1404
1405 /**
1406 * get_vm_area - reserve a contiguous kernel virtual area
1407 * @size: size of the area
1408 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1409 *
1410 * Search an area of @size in the kernel virtual mapping area,
1411 * and reserved it for out purposes. Returns the area descriptor
1412 * on success or %NULL on failure.
1413 */
1414 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1415 {
1416 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1417 NUMA_NO_NODE, GFP_KERNEL,
1418 __builtin_return_address(0));
1419 }
1420
1421 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1422 const void *caller)
1423 {
1424 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1425 NUMA_NO_NODE, GFP_KERNEL, caller);
1426 }
1427
1428 /**
1429 * find_vm_area - find a continuous kernel virtual area
1430 * @addr: base address
1431 *
1432 * Search for the kernel VM area starting at @addr, and return it.
1433 * It is up to the caller to do all required locking to keep the returned
1434 * pointer valid.
1435 */
1436 struct vm_struct *find_vm_area(const void *addr)
1437 {
1438 struct vmap_area *va;
1439
1440 va = find_vmap_area((unsigned long)addr);
1441 if (va && va->flags & VM_VM_AREA)
1442 return va->vm;
1443
1444 return NULL;
1445 }
1446
1447 /**
1448 * remove_vm_area - find and remove a continuous kernel virtual area
1449 * @addr: base address
1450 *
1451 * Search for the kernel VM area starting at @addr, and remove it.
1452 * This function returns the found VM area, but using it is NOT safe
1453 * on SMP machines, except for its size or flags.
1454 */
1455 struct vm_struct *remove_vm_area(const void *addr)
1456 {
1457 struct vmap_area *va;
1458
1459 va = find_vmap_area((unsigned long)addr);
1460 if (va && va->flags & VM_VM_AREA) {
1461 struct vm_struct *vm = va->vm;
1462
1463 spin_lock(&vmap_area_lock);
1464 va->vm = NULL;
1465 va->flags &= ~VM_VM_AREA;
1466 spin_unlock(&vmap_area_lock);
1467
1468 vmap_debug_free_range(va->va_start, va->va_end);
1469 kasan_free_shadow(vm);
1470 free_unmap_vmap_area(va);
1471
1472 return vm;
1473 }
1474 return NULL;
1475 }
1476
1477 static void __vunmap(const void *addr, int deallocate_pages)
1478 {
1479 struct vm_struct *area;
1480
1481 if (!addr)
1482 return;
1483
1484 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1485 addr))
1486 return;
1487
1488 area = remove_vm_area(addr);
1489 if (unlikely(!area)) {
1490 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1491 addr);
1492 return;
1493 }
1494
1495 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1496 debug_check_no_obj_freed(addr, get_vm_area_size(area));
1497
1498 if (deallocate_pages) {
1499 int i;
1500
1501 for (i = 0; i < area->nr_pages; i++) {
1502 struct page *page = area->pages[i];
1503
1504 BUG_ON(!page);
1505 __free_kmem_pages(page, 0);
1506 }
1507
1508 kvfree(area->pages);
1509 }
1510
1511 kfree(area);
1512 return;
1513 }
1514
1515 /**
1516 * vfree - release memory allocated by vmalloc()
1517 * @addr: memory base address
1518 *
1519 * Free the virtually continuous memory area starting at @addr, as
1520 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1521 * NULL, no operation is performed.
1522 *
1523 * Must not be called in NMI context (strictly speaking, only if we don't
1524 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1525 * conventions for vfree() arch-depenedent would be a really bad idea)
1526 *
1527 * NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1528 */
1529 void vfree(const void *addr)
1530 {
1531 BUG_ON(in_nmi());
1532
1533 kmemleak_free(addr);
1534
1535 if (!addr)
1536 return;
1537 if (unlikely(in_interrupt())) {
1538 struct vfree_deferred *p = this_cpu_ptr(&vfree_deferred);
1539 if (llist_add((struct llist_node *)addr, &p->list))
1540 schedule_work(&p->wq);
1541 } else
1542 __vunmap(addr, 1);
1543 }
1544 EXPORT_SYMBOL(vfree);
1545
1546 /**
1547 * vunmap - release virtual mapping obtained by vmap()
1548 * @addr: memory base address
1549 *
1550 * Free the virtually contiguous memory area starting at @addr,
1551 * which was created from the page array passed to vmap().
1552 *
1553 * Must not be called in interrupt context.
1554 */
1555 void vunmap(const void *addr)
1556 {
1557 BUG_ON(in_interrupt());
1558 might_sleep();
1559 if (addr)
1560 __vunmap(addr, 0);
1561 }
1562 EXPORT_SYMBOL(vunmap);
1563
1564 /**
1565 * vmap - map an array of pages into virtually contiguous space
1566 * @pages: array of page pointers
1567 * @count: number of pages to map
1568 * @flags: vm_area->flags
1569 * @prot: page protection for the mapping
1570 *
1571 * Maps @count pages from @pages into contiguous kernel virtual
1572 * space.
1573 */
1574 void *vmap(struct page **pages, unsigned int count,
1575 unsigned long flags, pgprot_t prot)
1576 {
1577 struct vm_struct *area;
1578
1579 might_sleep();
1580
1581 if (count > totalram_pages)
1582 return NULL;
1583
1584 area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1585 __builtin_return_address(0));
1586 if (!area)
1587 return NULL;
1588
1589 if (map_vm_area(area, prot, pages)) {
1590 vunmap(area->addr);
1591 return NULL;
1592 }
1593
1594 return area->addr;
1595 }
1596 EXPORT_SYMBOL(vmap);
1597
1598 static void *__vmalloc_node(unsigned long size, unsigned long align,
1599 gfp_t gfp_mask, pgprot_t prot,
1600 int node, const void *caller);
1601 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1602 pgprot_t prot, int node)
1603 {
1604 const int order = 0;
1605 struct page **pages;
1606 unsigned int nr_pages, array_size, i;
1607 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1608 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1609
1610 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1611 array_size = (nr_pages * sizeof(struct page *));
1612
1613 area->nr_pages = nr_pages;
1614 /* Please note that the recursion is strictly bounded. */
1615 if (array_size > PAGE_SIZE) {
1616 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1617 PAGE_KERNEL, node, area->caller);
1618 } else {
1619 pages = kmalloc_node(array_size, nested_gfp, node);
1620 }
1621 area->pages = pages;
1622 if (!area->pages) {
1623 remove_vm_area(area->addr);
1624 kfree(area);
1625 return NULL;
1626 }
1627
1628 for (i = 0; i < area->nr_pages; i++) {
1629 struct page *page;
1630
1631 if (node == NUMA_NO_NODE)
1632 page = alloc_kmem_pages(alloc_mask, order);
1633 else
1634 page = alloc_kmem_pages_node(node, alloc_mask, order);
1635
1636 if (unlikely(!page)) {
1637 /* Successfully allocated i pages, free them in __vunmap() */
1638 area->nr_pages = i;
1639 goto fail;
1640 }
1641 area->pages[i] = page;
1642 if (gfpflags_allow_blocking(gfp_mask))
1643 cond_resched();
1644 }
1645
1646 if (map_vm_area(area, prot, pages))
1647 goto fail;
1648 return area->addr;
1649
1650 fail:
1651 warn_alloc_failed(gfp_mask, order,
1652 "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1653 (area->nr_pages*PAGE_SIZE), area->size);
1654 vfree(area->addr);
1655 return NULL;
1656 }
1657
1658 /**
1659 * __vmalloc_node_range - allocate virtually contiguous memory
1660 * @size: allocation size
1661 * @align: desired alignment
1662 * @start: vm area range start
1663 * @end: vm area range end
1664 * @gfp_mask: flags for the page level allocator
1665 * @prot: protection mask for the allocated pages
1666 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
1667 * @node: node to use for allocation or NUMA_NO_NODE
1668 * @caller: caller's return address
1669 *
1670 * Allocate enough pages to cover @size from the page level
1671 * allocator with @gfp_mask flags. Map them into contiguous
1672 * kernel virtual space, using a pagetable protection of @prot.
1673 */
1674 void *__vmalloc_node_range(unsigned long size, unsigned long align,
1675 unsigned long start, unsigned long end, gfp_t gfp_mask,
1676 pgprot_t prot, unsigned long vm_flags, int node,
1677 const void *caller)
1678 {
1679 struct vm_struct *area;
1680 void *addr;
1681 unsigned long real_size = size;
1682
1683 size = PAGE_ALIGN(size);
1684 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1685 goto fail;
1686
1687 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1688 vm_flags, start, end, node, gfp_mask, caller);
1689 if (!area)
1690 goto fail;
1691
1692 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1693 if (!addr)
1694 return NULL;
1695
1696 /*
1697 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1698 * flag. It means that vm_struct is not fully initialized.
1699 * Now, it is fully initialized, so remove this flag here.
1700 */
1701 clear_vm_uninitialized_flag(area);
1702
1703 /*
1704 * A ref_count = 2 is needed because vm_struct allocated in
1705 * __get_vm_area_node() contains a reference to the virtual address of
1706 * the vmalloc'ed block.
1707 */
1708 kmemleak_alloc(addr, real_size, 2, gfp_mask);
1709
1710 return addr;
1711
1712 fail:
1713 warn_alloc_failed(gfp_mask, 0,
1714 "vmalloc: allocation failure: %lu bytes\n",
1715 real_size);
1716 return NULL;
1717 }
1718
1719 /**
1720 * __vmalloc_node - allocate virtually contiguous memory
1721 * @size: allocation size
1722 * @align: desired alignment
1723 * @gfp_mask: flags for the page level allocator
1724 * @prot: protection mask for the allocated pages
1725 * @node: node to use for allocation or NUMA_NO_NODE
1726 * @caller: caller's return address
1727 *
1728 * Allocate enough pages to cover @size from the page level
1729 * allocator with @gfp_mask flags. Map them into contiguous
1730 * kernel virtual space, using a pagetable protection of @prot.
1731 */
1732 static void *__vmalloc_node(unsigned long size, unsigned long align,
1733 gfp_t gfp_mask, pgprot_t prot,
1734 int node, const void *caller)
1735 {
1736 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1737 gfp_mask, prot, 0, node, caller);
1738 }
1739
1740 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1741 {
1742 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1743 __builtin_return_address(0));
1744 }
1745 EXPORT_SYMBOL(__vmalloc);
1746
1747 static inline void *__vmalloc_node_flags(unsigned long size,
1748 int node, gfp_t flags)
1749 {
1750 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1751 node, __builtin_return_address(0));
1752 }
1753
1754 /**
1755 * vmalloc - allocate virtually contiguous memory
1756 * @size: allocation size
1757 * Allocate enough pages to cover @size from the page level
1758 * allocator and map them into contiguous kernel virtual space.
1759 *
1760 * For tight control over page level allocator and protection flags
1761 * use __vmalloc() instead.
1762 */
1763 void *vmalloc(unsigned long size)
1764 {
1765 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1766 GFP_KERNEL | __GFP_HIGHMEM);
1767 }
1768 EXPORT_SYMBOL(vmalloc);
1769
1770 /**
1771 * vzalloc - allocate virtually contiguous memory with zero fill
1772 * @size: allocation size
1773 * Allocate enough pages to cover @size from the page level
1774 * allocator and map them into contiguous kernel virtual space.
1775 * The memory allocated is set to zero.
1776 *
1777 * For tight control over page level allocator and protection flags
1778 * use __vmalloc() instead.
1779 */
1780 void *vzalloc(unsigned long size)
1781 {
1782 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1783 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1784 }
1785 EXPORT_SYMBOL(vzalloc);
1786
1787 /**
1788 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1789 * @size: allocation size
1790 *
1791 * The resulting memory area is zeroed so it can be mapped to userspace
1792 * without leaking data.
1793 */
1794 void *vmalloc_user(unsigned long size)
1795 {
1796 struct vm_struct *area;
1797 void *ret;
1798
1799 ret = __vmalloc_node(size, SHMLBA,
1800 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1801 PAGE_KERNEL, NUMA_NO_NODE,
1802 __builtin_return_address(0));
1803 if (ret) {
1804 area = find_vm_area(ret);
1805 area->flags |= VM_USERMAP;
1806 }
1807 return ret;
1808 }
1809 EXPORT_SYMBOL(vmalloc_user);
1810
1811 /**
1812 * vmalloc_node - allocate memory on a specific node
1813 * @size: allocation size
1814 * @node: numa node
1815 *
1816 * Allocate enough pages to cover @size from the page level
1817 * allocator and map them into contiguous kernel virtual space.
1818 *
1819 * For tight control over page level allocator and protection flags
1820 * use __vmalloc() instead.
1821 */
1822 void *vmalloc_node(unsigned long size, int node)
1823 {
1824 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1825 node, __builtin_return_address(0));
1826 }
1827 EXPORT_SYMBOL(vmalloc_node);
1828
1829 /**
1830 * vzalloc_node - allocate memory on a specific node with zero fill
1831 * @size: allocation size
1832 * @node: numa node
1833 *
1834 * Allocate enough pages to cover @size from the page level
1835 * allocator and map them into contiguous kernel virtual space.
1836 * The memory allocated is set to zero.
1837 *
1838 * For tight control over page level allocator and protection flags
1839 * use __vmalloc_node() instead.
1840 */
1841 void *vzalloc_node(unsigned long size, int node)
1842 {
1843 return __vmalloc_node_flags(size, node,
1844 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1845 }
1846 EXPORT_SYMBOL(vzalloc_node);
1847
1848 #ifndef PAGE_KERNEL_EXEC
1849 # define PAGE_KERNEL_EXEC PAGE_KERNEL
1850 #endif
1851
1852 /**
1853 * vmalloc_exec - allocate virtually contiguous, executable memory
1854 * @size: allocation size
1855 *
1856 * Kernel-internal function to allocate enough pages to cover @size
1857 * the page level allocator and map them into contiguous and
1858 * executable kernel virtual space.
1859 *
1860 * For tight control over page level allocator and protection flags
1861 * use __vmalloc() instead.
1862 */
1863
1864 void *vmalloc_exec(unsigned long size)
1865 {
1866 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1867 NUMA_NO_NODE, __builtin_return_address(0));
1868 }
1869
1870 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1871 #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1872 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1873 #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1874 #else
1875 #define GFP_VMALLOC32 GFP_KERNEL
1876 #endif
1877
1878 /**
1879 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1880 * @size: allocation size
1881 *
1882 * Allocate enough 32bit PA addressable pages to cover @size from the
1883 * page level allocator and map them into contiguous kernel virtual space.
1884 */
1885 void *vmalloc_32(unsigned long size)
1886 {
1887 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1888 NUMA_NO_NODE, __builtin_return_address(0));
1889 }
1890 EXPORT_SYMBOL(vmalloc_32);
1891
1892 /**
1893 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1894 * @size: allocation size
1895 *
1896 * The resulting memory area is 32bit addressable and zeroed so it can be
1897 * mapped to userspace without leaking data.
1898 */
1899 void *vmalloc_32_user(unsigned long size)
1900 {
1901 struct vm_struct *area;
1902 void *ret;
1903
1904 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1905 NUMA_NO_NODE, __builtin_return_address(0));
1906 if (ret) {
1907 area = find_vm_area(ret);
1908 area->flags |= VM_USERMAP;
1909 }
1910 return ret;
1911 }
1912 EXPORT_SYMBOL(vmalloc_32_user);
1913
1914 /*
1915 * small helper routine , copy contents to buf from addr.
1916 * If the page is not present, fill zero.
1917 */
1918
1919 static int aligned_vread(char *buf, char *addr, unsigned long count)
1920 {
1921 struct page *p;
1922 int copied = 0;
1923
1924 while (count) {
1925 unsigned long offset, length;
1926
1927 offset = offset_in_page(addr);
1928 length = PAGE_SIZE - offset;
1929 if (length > count)
1930 length = count;
1931 p = vmalloc_to_page(addr);
1932 /*
1933 * To do safe access to this _mapped_ area, we need
1934 * lock. But adding lock here means that we need to add
1935 * overhead of vmalloc()/vfree() calles for this _debug_
1936 * interface, rarely used. Instead of that, we'll use
1937 * kmap() and get small overhead in this access function.
1938 */
1939 if (p) {
1940 /*
1941 * we can expect USER0 is not used (see vread/vwrite's
1942 * function description)
1943 */
1944 void *map = kmap_atomic(p);
1945 memcpy(buf, map + offset, length);
1946 kunmap_atomic(map);
1947 } else
1948 memset(buf, 0, length);
1949
1950 addr += length;
1951 buf += length;
1952 copied += length;
1953 count -= length;
1954 }
1955 return copied;
1956 }
1957
1958 static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1959 {
1960 struct page *p;
1961 int copied = 0;
1962
1963 while (count) {
1964 unsigned long offset, length;
1965
1966 offset = offset_in_page(addr);
1967 length = PAGE_SIZE - offset;
1968 if (length > count)
1969 length = count;
1970 p = vmalloc_to_page(addr);
1971 /*
1972 * To do safe access to this _mapped_ area, we need
1973 * lock. But adding lock here means that we need to add
1974 * overhead of vmalloc()/vfree() calles for this _debug_
1975 * interface, rarely used. Instead of that, we'll use
1976 * kmap() and get small overhead in this access function.
1977 */
1978 if (p) {
1979 /*
1980 * we can expect USER0 is not used (see vread/vwrite's
1981 * function description)
1982 */
1983 void *map = kmap_atomic(p);
1984 memcpy(map + offset, buf, length);
1985 kunmap_atomic(map);
1986 }
1987 addr += length;
1988 buf += length;
1989 copied += length;
1990 count -= length;
1991 }
1992 return copied;
1993 }
1994
1995 /**
1996 * vread() - read vmalloc area in a safe way.
1997 * @buf: buffer for reading data
1998 * @addr: vm address.
1999 * @count: number of bytes to be read.
2000 *
2001 * Returns # of bytes which addr and buf should be increased.
2002 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2003 * includes any intersect with alive vmalloc area.
2004 *
2005 * This function checks that addr is a valid vmalloc'ed area, and
2006 * copy data from that area to a given buffer. If the given memory range
2007 * of [addr...addr+count) includes some valid address, data is copied to
2008 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2009 * IOREMAP area is treated as memory hole and no copy is done.
2010 *
2011 * If [addr...addr+count) doesn't includes any intersects with alive
2012 * vm_struct area, returns 0. @buf should be kernel's buffer.
2013 *
2014 * Note: In usual ops, vread() is never necessary because the caller
2015 * should know vmalloc() area is valid and can use memcpy().
2016 * This is for routines which have to access vmalloc area without
2017 * any informaion, as /dev/kmem.
2018 *
2019 */
2020
2021 long vread(char *buf, char *addr, unsigned long count)
2022 {
2023 struct vmap_area *va;
2024 struct vm_struct *vm;
2025 char *vaddr, *buf_start = buf;
2026 unsigned long buflen = count;
2027 unsigned long n;
2028
2029 /* Don't allow overflow */
2030 if ((unsigned long) addr + count < count)
2031 count = -(unsigned long) addr;
2032
2033 spin_lock(&vmap_area_lock);
2034 list_for_each_entry(va, &vmap_area_list, list) {
2035 if (!count)
2036 break;
2037
2038 if (!(va->flags & VM_VM_AREA))
2039 continue;
2040
2041 vm = va->vm;
2042 vaddr = (char *) vm->addr;
2043 if (addr >= vaddr + get_vm_area_size(vm))
2044 continue;
2045 while (addr < vaddr) {
2046 if (count == 0)
2047 goto finished;
2048 *buf = '\0';
2049 buf++;
2050 addr++;
2051 count--;
2052 }
2053 n = vaddr + get_vm_area_size(vm) - addr;
2054 if (n > count)
2055 n = count;
2056 if (!(vm->flags & VM_IOREMAP))
2057 aligned_vread(buf, addr, n);
2058 else /* IOREMAP area is treated as memory hole */
2059 memset(buf, 0, n);
2060 buf += n;
2061 addr += n;
2062 count -= n;
2063 }
2064 finished:
2065 spin_unlock(&vmap_area_lock);
2066
2067 if (buf == buf_start)
2068 return 0;
2069 /* zero-fill memory holes */
2070 if (buf != buf_start + buflen)
2071 memset(buf, 0, buflen - (buf - buf_start));
2072
2073 return buflen;
2074 }
2075
2076 /**
2077 * vwrite() - write vmalloc area in a safe way.
2078 * @buf: buffer for source data
2079 * @addr: vm address.
2080 * @count: number of bytes to be read.
2081 *
2082 * Returns # of bytes which addr and buf should be incresed.
2083 * (same number to @count).
2084 * If [addr...addr+count) doesn't includes any intersect with valid
2085 * vmalloc area, returns 0.
2086 *
2087 * This function checks that addr is a valid vmalloc'ed area, and
2088 * copy data from a buffer to the given addr. If specified range of
2089 * [addr...addr+count) includes some valid address, data is copied from
2090 * proper area of @buf. If there are memory holes, no copy to hole.
2091 * IOREMAP area is treated as memory hole and no copy is done.
2092 *
2093 * If [addr...addr+count) doesn't includes any intersects with alive
2094 * vm_struct area, returns 0. @buf should be kernel's buffer.
2095 *
2096 * Note: In usual ops, vwrite() is never necessary because the caller
2097 * should know vmalloc() area is valid and can use memcpy().
2098 * This is for routines which have to access vmalloc area without
2099 * any informaion, as /dev/kmem.
2100 */
2101
2102 long vwrite(char *buf, char *addr, unsigned long count)
2103 {
2104 struct vmap_area *va;
2105 struct vm_struct *vm;
2106 char *vaddr;
2107 unsigned long n, buflen;
2108 int copied = 0;
2109
2110 /* Don't allow overflow */
2111 if ((unsigned long) addr + count < count)
2112 count = -(unsigned long) addr;
2113 buflen = count;
2114
2115 spin_lock(&vmap_area_lock);
2116 list_for_each_entry(va, &vmap_area_list, list) {
2117 if (!count)
2118 break;
2119
2120 if (!(va->flags & VM_VM_AREA))
2121 continue;
2122
2123 vm = va->vm;
2124 vaddr = (char *) vm->addr;
2125 if (addr >= vaddr + get_vm_area_size(vm))
2126 continue;
2127 while (addr < vaddr) {
2128 if (count == 0)
2129 goto finished;
2130 buf++;
2131 addr++;
2132 count--;
2133 }
2134 n = vaddr + get_vm_area_size(vm) - addr;
2135 if (n > count)
2136 n = count;
2137 if (!(vm->flags & VM_IOREMAP)) {
2138 aligned_vwrite(buf, addr, n);
2139 copied++;
2140 }
2141 buf += n;
2142 addr += n;
2143 count -= n;
2144 }
2145 finished:
2146 spin_unlock(&vmap_area_lock);
2147 if (!copied)
2148 return 0;
2149 return buflen;
2150 }
2151
2152 /**
2153 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2154 * @vma: vma to cover
2155 * @uaddr: target user address to start at
2156 * @kaddr: virtual address of vmalloc kernel memory
2157 * @size: size of map area
2158 *
2159 * Returns: 0 for success, -Exxx on failure
2160 *
2161 * This function checks that @kaddr is a valid vmalloc'ed area,
2162 * and that it is big enough to cover the range starting at
2163 * @uaddr in @vma. Will return failure if that criteria isn't
2164 * met.
2165 *
2166 * Similar to remap_pfn_range() (see mm/memory.c)
2167 */
2168 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2169 void *kaddr, unsigned long size)
2170 {
2171 struct vm_struct *area;
2172
2173 size = PAGE_ALIGN(size);
2174
2175 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2176 return -EINVAL;
2177
2178 area = find_vm_area(kaddr);
2179 if (!area)
2180 return -EINVAL;
2181
2182 if (!(area->flags & VM_USERMAP))
2183 return -EINVAL;
2184
2185 if (kaddr + size > area->addr + area->size)
2186 return -EINVAL;
2187
2188 do {
2189 struct page *page = vmalloc_to_page(kaddr);
2190 int ret;
2191
2192 ret = vm_insert_page(vma, uaddr, page);
2193 if (ret)
2194 return ret;
2195
2196 uaddr += PAGE_SIZE;
2197 kaddr += PAGE_SIZE;
2198 size -= PAGE_SIZE;
2199 } while (size > 0);
2200
2201 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2202
2203 return 0;
2204 }
2205 EXPORT_SYMBOL(remap_vmalloc_range_partial);
2206
2207 /**
2208 * remap_vmalloc_range - map vmalloc pages to userspace
2209 * @vma: vma to cover (map full range of vma)
2210 * @addr: vmalloc memory
2211 * @pgoff: number of pages into addr before first page to map
2212 *
2213 * Returns: 0 for success, -Exxx on failure
2214 *
2215 * This function checks that addr is a valid vmalloc'ed area, and
2216 * that it is big enough to cover the vma. Will return failure if
2217 * that criteria isn't met.
2218 *
2219 * Similar to remap_pfn_range() (see mm/memory.c)
2220 */
2221 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2222 unsigned long pgoff)
2223 {
2224 return remap_vmalloc_range_partial(vma, vma->vm_start,
2225 addr + (pgoff << PAGE_SHIFT),
2226 vma->vm_end - vma->vm_start);
2227 }
2228 EXPORT_SYMBOL(remap_vmalloc_range);
2229
2230 /*
2231 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2232 * have one.
2233 */
2234 void __weak vmalloc_sync_all(void)
2235 {
2236 }
2237
2238
2239 static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2240 {
2241 pte_t ***p = data;
2242
2243 if (p) {
2244 *(*p) = pte;
2245 (*p)++;
2246 }
2247 return 0;
2248 }
2249
2250 /**
2251 * alloc_vm_area - allocate a range of kernel address space
2252 * @size: size of the area
2253 * @ptes: returns the PTEs for the address space
2254 *
2255 * Returns: NULL on failure, vm_struct on success
2256 *
2257 * This function reserves a range of kernel address space, and
2258 * allocates pagetables to map that range. No actual mappings
2259 * are created.
2260 *
2261 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2262 * allocated for the VM area are returned.
2263 */
2264 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2265 {
2266 struct vm_struct *area;
2267
2268 area = get_vm_area_caller(size, VM_IOREMAP,
2269 __builtin_return_address(0));
2270 if (area == NULL)
2271 return NULL;
2272
2273 /*
2274 * This ensures that page tables are constructed for this region
2275 * of kernel virtual address space and mapped into init_mm.
2276 */
2277 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2278 size, f, ptes ? &ptes : NULL)) {
2279 free_vm_area(area);
2280 return NULL;
2281 }
2282
2283 return area;
2284 }
2285 EXPORT_SYMBOL_GPL(alloc_vm_area);
2286
2287 void free_vm_area(struct vm_struct *area)
2288 {
2289 struct vm_struct *ret;
2290 ret = remove_vm_area(area->addr);
2291 BUG_ON(ret != area);
2292 kfree(area);
2293 }
2294 EXPORT_SYMBOL_GPL(free_vm_area);
2295
2296 #ifdef CONFIG_SMP
2297 static struct vmap_area *node_to_va(struct rb_node *n)
2298 {
2299 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2300 }
2301
2302 /**
2303 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2304 * @end: target address
2305 * @pnext: out arg for the next vmap_area
2306 * @pprev: out arg for the previous vmap_area
2307 *
2308 * Returns: %true if either or both of next and prev are found,
2309 * %false if no vmap_area exists
2310 *
2311 * Find vmap_areas end addresses of which enclose @end. ie. if not
2312 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2313 */
2314 static bool pvm_find_next_prev(unsigned long end,
2315 struct vmap_area **pnext,
2316 struct vmap_area **pprev)
2317 {
2318 struct rb_node *n = vmap_area_root.rb_node;
2319 struct vmap_area *va = NULL;
2320
2321 while (n) {
2322 va = rb_entry(n, struct vmap_area, rb_node);
2323 if (end < va->va_end)
2324 n = n->rb_left;
2325 else if (end > va->va_end)
2326 n = n->rb_right;
2327 else
2328 break;
2329 }
2330
2331 if (!va)
2332 return false;
2333
2334 if (va->va_end > end) {
2335 *pnext = va;
2336 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2337 } else {
2338 *pprev = va;
2339 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2340 }
2341 return true;
2342 }
2343
2344 /**
2345 * pvm_determine_end - find the highest aligned address between two vmap_areas
2346 * @pnext: in/out arg for the next vmap_area
2347 * @pprev: in/out arg for the previous vmap_area
2348 * @align: alignment
2349 *
2350 * Returns: determined end address
2351 *
2352 * Find the highest aligned address between *@pnext and *@pprev below
2353 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2354 * down address is between the end addresses of the two vmap_areas.
2355 *
2356 * Please note that the address returned by this function may fall
2357 * inside *@pnext vmap_area. The caller is responsible for checking
2358 * that.
2359 */
2360 static unsigned long pvm_determine_end(struct vmap_area **pnext,
2361 struct vmap_area **pprev,
2362 unsigned long align)
2363 {
2364 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2365 unsigned long addr;
2366
2367 if (*pnext)
2368 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2369 else
2370 addr = vmalloc_end;
2371
2372 while (*pprev && (*pprev)->va_end > addr) {
2373 *pnext = *pprev;
2374 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2375 }
2376
2377 return addr;
2378 }
2379
2380 /**
2381 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2382 * @offsets: array containing offset of each area
2383 * @sizes: array containing size of each area
2384 * @nr_vms: the number of areas to allocate
2385 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2386 *
2387 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2388 * vm_structs on success, %NULL on failure
2389 *
2390 * Percpu allocator wants to use congruent vm areas so that it can
2391 * maintain the offsets among percpu areas. This function allocates
2392 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2393 * be scattered pretty far, distance between two areas easily going up
2394 * to gigabytes. To avoid interacting with regular vmallocs, these
2395 * areas are allocated from top.
2396 *
2397 * Despite its complicated look, this allocator is rather simple. It
2398 * does everything top-down and scans areas from the end looking for
2399 * matching slot. While scanning, if any of the areas overlaps with
2400 * existing vmap_area, the base address is pulled down to fit the
2401 * area. Scanning is repeated till all the areas fit and then all
2402 * necessary data structres are inserted and the result is returned.
2403 */
2404 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2405 const size_t *sizes, int nr_vms,
2406 size_t align)
2407 {
2408 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2409 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2410 struct vmap_area **vas, *prev, *next;
2411 struct vm_struct **vms;
2412 int area, area2, last_area, term_area;
2413 unsigned long base, start, end, last_end;
2414 bool purged = false;
2415
2416 /* verify parameters and allocate data structures */
2417 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2418 for (last_area = 0, area = 0; area < nr_vms; area++) {
2419 start = offsets[area];
2420 end = start + sizes[area];
2421
2422 /* is everything aligned properly? */
2423 BUG_ON(!IS_ALIGNED(offsets[area], align));
2424 BUG_ON(!IS_ALIGNED(sizes[area], align));
2425
2426 /* detect the area with the highest address */
2427 if (start > offsets[last_area])
2428 last_area = area;
2429
2430 for (area2 = 0; area2 < nr_vms; area2++) {
2431 unsigned long start2 = offsets[area2];
2432 unsigned long end2 = start2 + sizes[area2];
2433
2434 if (area2 == area)
2435 continue;
2436
2437 BUG_ON(start2 >= start && start2 < end);
2438 BUG_ON(end2 <= end && end2 > start);
2439 }
2440 }
2441 last_end = offsets[last_area] + sizes[last_area];
2442
2443 if (vmalloc_end - vmalloc_start < last_end) {
2444 WARN_ON(true);
2445 return NULL;
2446 }
2447
2448 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2449 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2450 if (!vas || !vms)
2451 goto err_free2;
2452
2453 for (area = 0; area < nr_vms; area++) {
2454 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2455 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2456 if (!vas[area] || !vms[area])
2457 goto err_free;
2458 }
2459 retry:
2460 spin_lock(&vmap_area_lock);
2461
2462 /* start scanning - we scan from the top, begin with the last area */
2463 area = term_area = last_area;
2464 start = offsets[area];
2465 end = start + sizes[area];
2466
2467 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2468 base = vmalloc_end - last_end;
2469 goto found;
2470 }
2471 base = pvm_determine_end(&next, &prev, align) - end;
2472
2473 while (true) {
2474 BUG_ON(next && next->va_end <= base + end);
2475 BUG_ON(prev && prev->va_end > base + end);
2476
2477 /*
2478 * base might have underflowed, add last_end before
2479 * comparing.
2480 */
2481 if (base + last_end < vmalloc_start + last_end) {
2482 spin_unlock(&vmap_area_lock);
2483 if (!purged) {
2484 purge_vmap_area_lazy();
2485 purged = true;
2486 goto retry;
2487 }
2488 goto err_free;
2489 }
2490
2491 /*
2492 * If next overlaps, move base downwards so that it's
2493 * right below next and then recheck.
2494 */
2495 if (next && next->va_start < base + end) {
2496 base = pvm_determine_end(&next, &prev, align) - end;
2497 term_area = area;
2498 continue;
2499 }
2500
2501 /*
2502 * If prev overlaps, shift down next and prev and move
2503 * base so that it's right below new next and then
2504 * recheck.
2505 */
2506 if (prev && prev->va_end > base + start) {
2507 next = prev;
2508 prev = node_to_va(rb_prev(&next->rb_node));
2509 base = pvm_determine_end(&next, &prev, align) - end;
2510 term_area = area;
2511 continue;
2512 }
2513
2514 /*
2515 * This area fits, move on to the previous one. If
2516 * the previous one is the terminal one, we're done.
2517 */
2518 area = (area + nr_vms - 1) % nr_vms;
2519 if (area == term_area)
2520 break;
2521 start = offsets[area];
2522 end = start + sizes[area];
2523 pvm_find_next_prev(base + end, &next, &prev);
2524 }
2525 found:
2526 /* we've found a fitting base, insert all va's */
2527 for (area = 0; area < nr_vms; area++) {
2528 struct vmap_area *va = vas[area];
2529
2530 va->va_start = base + offsets[area];
2531 va->va_end = va->va_start + sizes[area];
2532 __insert_vmap_area(va);
2533 }
2534
2535 vmap_area_pcpu_hole = base + offsets[last_area];
2536
2537 spin_unlock(&vmap_area_lock);
2538
2539 /* insert all vm's */
2540 for (area = 0; area < nr_vms; area++)
2541 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2542 pcpu_get_vm_areas);
2543
2544 kfree(vas);
2545 return vms;
2546
2547 err_free:
2548 for (area = 0; area < nr_vms; area++) {
2549 kfree(vas[area]);
2550 kfree(vms[area]);
2551 }
2552 err_free2:
2553 kfree(vas);
2554 kfree(vms);
2555 return NULL;
2556 }
2557
2558 /**
2559 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2560 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2561 * @nr_vms: the number of allocated areas
2562 *
2563 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2564 */
2565 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2566 {
2567 int i;
2568
2569 for (i = 0; i < nr_vms; i++)
2570 free_vm_area(vms[i]);
2571 kfree(vms);
2572 }
2573 #endif /* CONFIG_SMP */
2574
2575 #ifdef CONFIG_PROC_FS
2576 static void *s_start(struct seq_file *m, loff_t *pos)
2577 __acquires(&vmap_area_lock)
2578 {
2579 loff_t n = *pos;
2580 struct vmap_area *va;
2581
2582 spin_lock(&vmap_area_lock);
2583 va = list_first_entry(&vmap_area_list, typeof(*va), list);
2584 while (n > 0 && &va->list != &vmap_area_list) {
2585 n--;
2586 va = list_next_entry(va, list);
2587 }
2588 if (!n && &va->list != &vmap_area_list)
2589 return va;
2590
2591 return NULL;
2592
2593 }
2594
2595 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2596 {
2597 struct vmap_area *va = p, *next;
2598
2599 ++*pos;
2600 next = list_next_entry(va, list);
2601 if (&next->list != &vmap_area_list)
2602 return next;
2603
2604 return NULL;
2605 }
2606
2607 static void s_stop(struct seq_file *m, void *p)
2608 __releases(&vmap_area_lock)
2609 {
2610 spin_unlock(&vmap_area_lock);
2611 }
2612
2613 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2614 {
2615 if (IS_ENABLED(CONFIG_NUMA)) {
2616 unsigned int nr, *counters = m->private;
2617
2618 if (!counters)
2619 return;
2620
2621 if (v->flags & VM_UNINITIALIZED)
2622 return;
2623 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2624 smp_rmb();
2625
2626 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2627
2628 for (nr = 0; nr < v->nr_pages; nr++)
2629 counters[page_to_nid(v->pages[nr])]++;
2630
2631 for_each_node_state(nr, N_HIGH_MEMORY)
2632 if (counters[nr])
2633 seq_printf(m, " N%u=%u", nr, counters[nr]);
2634 }
2635 }
2636
2637 static int s_show(struct seq_file *m, void *p)
2638 {
2639 struct vmap_area *va = p;
2640 struct vm_struct *v;
2641
2642 /*
2643 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2644 * behalf of vmap area is being tear down or vm_map_ram allocation.
2645 */
2646 if (!(va->flags & VM_VM_AREA))
2647 return 0;
2648
2649 v = va->vm;
2650
2651 seq_printf(m, "0x%pK-0x%pK %7ld",
2652 v->addr, v->addr + v->size, v->size);
2653
2654 if (v->caller)
2655 seq_printf(m, " %pS", v->caller);
2656
2657 if (v->nr_pages)
2658 seq_printf(m, " pages=%d", v->nr_pages);
2659
2660 if (v->phys_addr)
2661 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2662
2663 if (v->flags & VM_IOREMAP)
2664 seq_puts(m, " ioremap");
2665
2666 if (v->flags & VM_ALLOC)
2667 seq_puts(m, " vmalloc");
2668
2669 if (v->flags & VM_MAP)
2670 seq_puts(m, " vmap");
2671
2672 if (v->flags & VM_USERMAP)
2673 seq_puts(m, " user");
2674
2675 if (is_vmalloc_addr(v->pages))
2676 seq_puts(m, " vpages");
2677
2678 show_numa_info(m, v);
2679 seq_putc(m, '\n');
2680 return 0;
2681 }
2682
2683 static const struct seq_operations vmalloc_op = {
2684 .start = s_start,
2685 .next = s_next,
2686 .stop = s_stop,
2687 .show = s_show,
2688 };
2689
2690 static int vmalloc_open(struct inode *inode, struct file *file)
2691 {
2692 if (IS_ENABLED(CONFIG_NUMA))
2693 return seq_open_private(file, &vmalloc_op,
2694 nr_node_ids * sizeof(unsigned int));
2695 else
2696 return seq_open(file, &vmalloc_op);
2697 }
2698
2699 static const struct file_operations proc_vmalloc_operations = {
2700 .open = vmalloc_open,
2701 .read = seq_read,
2702 .llseek = seq_lseek,
2703 .release = seq_release_private,
2704 };
2705
2706 static int __init proc_vmalloc_init(void)
2707 {
2708 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2709 return 0;
2710 }
2711 module_init(proc_vmalloc_init);
2712
2713 #endif
2714