]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - mm/vmalloc.c
Merge tag 'usb-4.10-rc7' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb
[mirror_ubuntu-bionic-kernel.git] / mm / vmalloc.c
CommitLineData
1da177e4
LT
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
930fc45a 8 * Numa awareness, Christoph Lameter, SGI, June 2005
1da177e4
LT
9 */
10
db64fe02 11#include <linux/vmalloc.h>
1da177e4
LT
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
d43c36dc 15#include <linux/sched.h>
1da177e4
LT
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
5f6a6a9c 19#include <linux/proc_fs.h>
a10aa579 20#include <linux/seq_file.h>
3ac7fe5a 21#include <linux/debugobjects.h>
23016969 22#include <linux/kallsyms.h>
db64fe02 23#include <linux/list.h>
4da56b99 24#include <linux/notifier.h>
db64fe02
NP
25#include <linux/rbtree.h>
26#include <linux/radix-tree.h>
27#include <linux/rcupdate.h>
f0aa6617 28#include <linux/pfn.h>
89219d37 29#include <linux/kmemleak.h>
60063497 30#include <linux/atomic.h>
3b32123d 31#include <linux/compiler.h>
32fcfd40 32#include <linux/llist.h>
0f616be1 33#include <linux/bitops.h>
3b32123d 34
7c0f6ba6 35#include <linux/uaccess.h>
1da177e4 36#include <asm/tlbflush.h>
2dca6999 37#include <asm/shmparam.h>
1da177e4 38
dd56b046
MG
39#include "internal.h"
40
32fcfd40
AV
41struct vfree_deferred {
42 struct llist_head list;
43 struct work_struct wq;
44};
45static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
46
47static void __vunmap(const void *, int);
48
49static void free_work(struct work_struct *w)
50{
51 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
52 struct llist_node *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
db64fe02 60/*** Page table manipulation functions ***/
b221385b 61
1da177e4
LT
62static 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
db64fe02 73static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
1da177e4
LT
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);
b9820d8f
TK
81 if (pmd_clear_huge(pmd))
82 continue;
1da177e4
LT
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
db64fe02 89static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
1da177e4
LT
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);
b9820d8f
TK
97 if (pud_clear_huge(pud))
98 continue;
1da177e4
LT
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
db64fe02 105static void vunmap_page_range(unsigned long addr, unsigned long end)
1da177e4
LT
106{
107 pgd_t *pgd;
108 unsigned long next;
1da177e4
LT
109
110 BUG_ON(addr >= end);
111 pgd = pgd_offset_k(addr);
1da177e4
LT
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);
1da177e4
LT
118}
119
120static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
db64fe02 121 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
1da177e4
LT
122{
123 pte_t *pte;
124
db64fe02
NP
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
872fec16 130 pte = pte_alloc_kernel(pmd, addr);
1da177e4
LT
131 if (!pte)
132 return -ENOMEM;
133 do {
db64fe02
NP
134 struct page *page = pages[*nr];
135
136 if (WARN_ON(!pte_none(*pte)))
137 return -EBUSY;
138 if (WARN_ON(!page))
1da177e4
LT
139 return -ENOMEM;
140 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
db64fe02 141 (*nr)++;
1da177e4
LT
142 } while (pte++, addr += PAGE_SIZE, addr != end);
143 return 0;
144}
145
db64fe02
NP
146static int vmap_pmd_range(pud_t *pud, unsigned long addr,
147 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
1da177e4
LT
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);
db64fe02 157 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
1da177e4
LT
158 return -ENOMEM;
159 } while (pmd++, addr = next, addr != end);
160 return 0;
161}
162
db64fe02
NP
163static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
1da177e4
LT
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);
db64fe02 174 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
1da177e4
LT
175 return -ENOMEM;
176 } while (pud++, addr = next, addr != end);
177 return 0;
178}
179
db64fe02
NP
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 */
8fc48985
TH
186static int vmap_page_range_noflush(unsigned long start, unsigned long end,
187 pgprot_t prot, struct page **pages)
1da177e4
LT
188{
189 pgd_t *pgd;
190 unsigned long next;
2e4e27c7 191 unsigned long addr = start;
db64fe02
NP
192 int err = 0;
193 int nr = 0;
1da177e4
LT
194
195 BUG_ON(addr >= end);
196 pgd = pgd_offset_k(addr);
1da177e4
LT
197 do {
198 next = pgd_addr_end(addr, end);
db64fe02 199 err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
1da177e4 200 if (err)
bf88c8c8 201 return err;
1da177e4 202 } while (pgd++, addr = next, addr != end);
db64fe02 203
db64fe02 204 return nr;
1da177e4
LT
205}
206
8fc48985
TH
207static 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
81ac3ad9 217int is_vmalloc_or_module_addr(const void *x)
73bdf0a6
LT
218{
219 /*
ab4f2ee1 220 * ARM, x86-64 and sparc64 put modules in a special place,
73bdf0a6
LT
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
48667e7a 232/*
add688fb 233 * Walk a vmap address to the struct page it maps.
48667e7a 234 */
add688fb 235struct page *vmalloc_to_page(const void *vmalloc_addr)
48667e7a
CL
236{
237 unsigned long addr = (unsigned long) vmalloc_addr;
add688fb 238 struct page *page = NULL;
48667e7a 239 pgd_t *pgd = pgd_offset_k(addr);
48667e7a 240
7aa413de
IM
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 */
73bdf0a6 245 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
59ea7463 246
48667e7a 247 if (!pgd_none(*pgd)) {
db64fe02 248 pud_t *pud = pud_offset(pgd, addr);
48667e7a 249 if (!pud_none(*pud)) {
db64fe02 250 pmd_t *pmd = pmd_offset(pud, addr);
48667e7a 251 if (!pmd_none(*pmd)) {
db64fe02
NP
252 pte_t *ptep, pte;
253
48667e7a
CL
254 ptep = pte_offset_map(pmd, addr);
255 pte = *ptep;
256 if (pte_present(pte))
add688fb 257 page = pte_page(pte);
48667e7a
CL
258 pte_unmap(ptep);
259 }
260 }
261 }
add688fb 262 return page;
48667e7a 263}
add688fb 264EXPORT_SYMBOL(vmalloc_to_page);
48667e7a
CL
265
266/*
add688fb 267 * Map a vmalloc()-space virtual address to the physical page frame number.
48667e7a 268 */
add688fb 269unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
48667e7a 270{
add688fb 271 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
48667e7a 272}
add688fb 273EXPORT_SYMBOL(vmalloc_to_pfn);
48667e7a 274
db64fe02
NP
275
276/*** Global kva allocator ***/
277
db64fe02
NP
278#define VM_VM_AREA 0x04
279
db64fe02 280static DEFINE_SPINLOCK(vmap_area_lock);
f1c4069e
JK
281/* Export for kexec only */
282LIST_HEAD(vmap_area_list);
80c4bd7a 283static LLIST_HEAD(vmap_purge_list);
89699605
NP
284static struct rb_root vmap_area_root = RB_ROOT;
285
286/* The vmap cache globals are protected by vmap_area_lock */
287static struct rb_node *free_vmap_cache;
288static unsigned long cached_hole_size;
289static unsigned long cached_vstart;
290static unsigned long cached_align;
291
ca23e405 292static unsigned long vmap_area_pcpu_hole;
db64fe02
NP
293
294static struct vmap_area *__find_vmap_area(unsigned long addr)
1da177e4 295{
db64fe02
NP
296 struct rb_node *n = vmap_area_root.rb_node;
297
298 while (n) {
299 struct vmap_area *va;
300
301 va = rb_entry(n, struct vmap_area, rb_node);
302 if (addr < va->va_start)
303 n = n->rb_left;
cef2ac3f 304 else if (addr >= va->va_end)
db64fe02
NP
305 n = n->rb_right;
306 else
307 return va;
308 }
309
310 return NULL;
311}
312
313static void __insert_vmap_area(struct vmap_area *va)
314{
315 struct rb_node **p = &vmap_area_root.rb_node;
316 struct rb_node *parent = NULL;
317 struct rb_node *tmp;
318
319 while (*p) {
170168d0 320 struct vmap_area *tmp_va;
db64fe02
NP
321
322 parent = *p;
170168d0
NK
323 tmp_va = rb_entry(parent, struct vmap_area, rb_node);
324 if (va->va_start < tmp_va->va_end)
db64fe02 325 p = &(*p)->rb_left;
170168d0 326 else if (va->va_end > tmp_va->va_start)
db64fe02
NP
327 p = &(*p)->rb_right;
328 else
329 BUG();
330 }
331
332 rb_link_node(&va->rb_node, parent, p);
333 rb_insert_color(&va->rb_node, &vmap_area_root);
334
4341fa45 335 /* address-sort this list */
db64fe02
NP
336 tmp = rb_prev(&va->rb_node);
337 if (tmp) {
338 struct vmap_area *prev;
339 prev = rb_entry(tmp, struct vmap_area, rb_node);
340 list_add_rcu(&va->list, &prev->list);
341 } else
342 list_add_rcu(&va->list, &vmap_area_list);
343}
344
345static void purge_vmap_area_lazy(void);
346
4da56b99
CW
347static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
348
db64fe02
NP
349/*
350 * Allocate a region of KVA of the specified size and alignment, within the
351 * vstart and vend.
352 */
353static struct vmap_area *alloc_vmap_area(unsigned long size,
354 unsigned long align,
355 unsigned long vstart, unsigned long vend,
356 int node, gfp_t gfp_mask)
357{
358 struct vmap_area *va;
359 struct rb_node *n;
1da177e4 360 unsigned long addr;
db64fe02 361 int purged = 0;
89699605 362 struct vmap_area *first;
db64fe02 363
7766970c 364 BUG_ON(!size);
891c49ab 365 BUG_ON(offset_in_page(size));
89699605 366 BUG_ON(!is_power_of_2(align));
db64fe02 367
5803ed29 368 might_sleep();
4da56b99 369
db64fe02
NP
370 va = kmalloc_node(sizeof(struct vmap_area),
371 gfp_mask & GFP_RECLAIM_MASK, node);
372 if (unlikely(!va))
373 return ERR_PTR(-ENOMEM);
374
7f88f88f
CM
375 /*
376 * Only scan the relevant parts containing pointers to other objects
377 * to avoid false negatives.
378 */
379 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
380
db64fe02
NP
381retry:
382 spin_lock(&vmap_area_lock);
89699605
NP
383 /*
384 * Invalidate cache if we have more permissive parameters.
385 * cached_hole_size notes the largest hole noticed _below_
386 * the vmap_area cached in free_vmap_cache: if size fits
387 * into that hole, we want to scan from vstart to reuse
388 * the hole instead of allocating above free_vmap_cache.
389 * Note that __free_vmap_area may update free_vmap_cache
390 * without updating cached_hole_size or cached_align.
391 */
392 if (!free_vmap_cache ||
393 size < cached_hole_size ||
394 vstart < cached_vstart ||
395 align < cached_align) {
396nocache:
397 cached_hole_size = 0;
398 free_vmap_cache = NULL;
399 }
400 /* record if we encounter less permissive parameters */
401 cached_vstart = vstart;
402 cached_align = align;
403
404 /* find starting point for our search */
405 if (free_vmap_cache) {
406 first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
248ac0e1 407 addr = ALIGN(first->va_end, align);
89699605
NP
408 if (addr < vstart)
409 goto nocache;
bcb615a8 410 if (addr + size < addr)
89699605
NP
411 goto overflow;
412
413 } else {
414 addr = ALIGN(vstart, align);
bcb615a8 415 if (addr + size < addr)
89699605
NP
416 goto overflow;
417
418 n = vmap_area_root.rb_node;
419 first = NULL;
420
421 while (n) {
db64fe02
NP
422 struct vmap_area *tmp;
423 tmp = rb_entry(n, struct vmap_area, rb_node);
424 if (tmp->va_end >= addr) {
db64fe02 425 first = tmp;
89699605
NP
426 if (tmp->va_start <= addr)
427 break;
428 n = n->rb_left;
429 } else
db64fe02 430 n = n->rb_right;
89699605 431 }
db64fe02
NP
432
433 if (!first)
434 goto found;
db64fe02 435 }
89699605
NP
436
437 /* from the starting point, walk areas until a suitable hole is found */
248ac0e1 438 while (addr + size > first->va_start && addr + size <= vend) {
89699605
NP
439 if (addr + cached_hole_size < first->va_start)
440 cached_hole_size = first->va_start - addr;
248ac0e1 441 addr = ALIGN(first->va_end, align);
bcb615a8 442 if (addr + size < addr)
89699605
NP
443 goto overflow;
444
92ca922f 445 if (list_is_last(&first->list, &vmap_area_list))
89699605 446 goto found;
92ca922f 447
6219c2a2 448 first = list_next_entry(first, list);
db64fe02
NP
449 }
450
89699605
NP
451found:
452 if (addr + size > vend)
453 goto overflow;
db64fe02
NP
454
455 va->va_start = addr;
456 va->va_end = addr + size;
457 va->flags = 0;
458 __insert_vmap_area(va);
89699605 459 free_vmap_cache = &va->rb_node;
db64fe02
NP
460 spin_unlock(&vmap_area_lock);
461
61e16557 462 BUG_ON(!IS_ALIGNED(va->va_start, align));
89699605
NP
463 BUG_ON(va->va_start < vstart);
464 BUG_ON(va->va_end > vend);
465
db64fe02 466 return va;
89699605
NP
467
468overflow:
469 spin_unlock(&vmap_area_lock);
470 if (!purged) {
471 purge_vmap_area_lazy();
472 purged = 1;
473 goto retry;
474 }
4da56b99
CW
475
476 if (gfpflags_allow_blocking(gfp_mask)) {
477 unsigned long freed = 0;
478 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
479 if (freed > 0) {
480 purged = 0;
481 goto retry;
482 }
483 }
484
89699605 485 if (printk_ratelimit())
756a025f
JP
486 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
487 size);
89699605
NP
488 kfree(va);
489 return ERR_PTR(-EBUSY);
db64fe02
NP
490}
491
4da56b99
CW
492int register_vmap_purge_notifier(struct notifier_block *nb)
493{
494 return blocking_notifier_chain_register(&vmap_notify_list, nb);
495}
496EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
497
498int unregister_vmap_purge_notifier(struct notifier_block *nb)
499{
500 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
501}
502EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
503
db64fe02
NP
504static void __free_vmap_area(struct vmap_area *va)
505{
506 BUG_ON(RB_EMPTY_NODE(&va->rb_node));
89699605
NP
507
508 if (free_vmap_cache) {
509 if (va->va_end < cached_vstart) {
510 free_vmap_cache = NULL;
511 } else {
512 struct vmap_area *cache;
513 cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
514 if (va->va_start <= cache->va_start) {
515 free_vmap_cache = rb_prev(&va->rb_node);
516 /*
517 * We don't try to update cached_hole_size or
518 * cached_align, but it won't go very wrong.
519 */
520 }
521 }
522 }
db64fe02
NP
523 rb_erase(&va->rb_node, &vmap_area_root);
524 RB_CLEAR_NODE(&va->rb_node);
525 list_del_rcu(&va->list);
526
ca23e405
TH
527 /*
528 * Track the highest possible candidate for pcpu area
529 * allocation. Areas outside of vmalloc area can be returned
530 * here too, consider only end addresses which fall inside
531 * vmalloc area proper.
532 */
533 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
534 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
535
14769de9 536 kfree_rcu(va, rcu_head);
db64fe02
NP
537}
538
539/*
540 * Free a region of KVA allocated by alloc_vmap_area
541 */
542static void free_vmap_area(struct vmap_area *va)
543{
544 spin_lock(&vmap_area_lock);
545 __free_vmap_area(va);
546 spin_unlock(&vmap_area_lock);
547}
548
549/*
550 * Clear the pagetable entries of a given vmap_area
551 */
552static void unmap_vmap_area(struct vmap_area *va)
553{
554 vunmap_page_range(va->va_start, va->va_end);
555}
556
cd52858c
NP
557static void vmap_debug_free_range(unsigned long start, unsigned long end)
558{
559 /*
f48d97f3
JK
560 * Unmap page tables and force a TLB flush immediately if pagealloc
561 * debugging is enabled. This catches use after free bugs similarly to
562 * those in linear kernel virtual address space after a page has been
563 * freed.
cd52858c 564 *
f48d97f3
JK
565 * All the lazy freeing logic is still retained, in order to minimise
566 * intrusiveness of this debugging feature.
cd52858c 567 *
f48d97f3
JK
568 * This is going to be *slow* (linear kernel virtual address debugging
569 * doesn't do a broadcast TLB flush so it is a lot faster).
cd52858c 570 */
f48d97f3
JK
571 if (debug_pagealloc_enabled()) {
572 vunmap_page_range(start, end);
573 flush_tlb_kernel_range(start, end);
574 }
cd52858c
NP
575}
576
db64fe02
NP
577/*
578 * lazy_max_pages is the maximum amount of virtual address space we gather up
579 * before attempting to purge with a TLB flush.
580 *
581 * There is a tradeoff here: a larger number will cover more kernel page tables
582 * and take slightly longer to purge, but it will linearly reduce the number of
583 * global TLB flushes that must be performed. It would seem natural to scale
584 * this number up linearly with the number of CPUs (because vmapping activity
585 * could also scale linearly with the number of CPUs), however it is likely
586 * that in practice, workloads might be constrained in other ways that mean
587 * vmap activity will not scale linearly with CPUs. Also, I want to be
588 * conservative and not introduce a big latency on huge systems, so go with
589 * a less aggressive log scale. It will still be an improvement over the old
590 * code, and it will be simple to change the scale factor if we find that it
591 * becomes a problem on bigger systems.
592 */
593static unsigned long lazy_max_pages(void)
594{
595 unsigned int log;
596
597 log = fls(num_online_cpus());
598
599 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
600}
601
602static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
603
0574ecd1
CH
604/*
605 * Serialize vmap purging. There is no actual criticial section protected
606 * by this look, but we want to avoid concurrent calls for performance
607 * reasons and to make the pcpu_get_vm_areas more deterministic.
608 */
f9e09977 609static DEFINE_MUTEX(vmap_purge_lock);
0574ecd1 610
02b709df
NP
611/* for per-CPU blocks */
612static void purge_fragmented_blocks_allcpus(void);
613
3ee48b6a
CW
614/*
615 * called before a call to iounmap() if the caller wants vm_area_struct's
616 * immediately freed.
617 */
618void set_iounmap_nonlazy(void)
619{
620 atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
621}
622
db64fe02
NP
623/*
624 * Purges all lazily-freed vmap areas.
db64fe02 625 */
0574ecd1 626static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
db64fe02 627{
80c4bd7a 628 struct llist_node *valist;
db64fe02 629 struct vmap_area *va;
cbb76676 630 struct vmap_area *n_va;
763b218d 631 bool do_free = false;
db64fe02 632
0574ecd1 633 lockdep_assert_held(&vmap_purge_lock);
02b709df 634
80c4bd7a
CW
635 valist = llist_del_all(&vmap_purge_list);
636 llist_for_each_entry(va, valist, purge_list) {
0574ecd1
CH
637 if (va->va_start < start)
638 start = va->va_start;
639 if (va->va_end > end)
640 end = va->va_end;
763b218d 641 do_free = true;
db64fe02 642 }
db64fe02 643
763b218d 644 if (!do_free)
0574ecd1 645 return false;
db64fe02 646
0574ecd1 647 flush_tlb_kernel_range(start, end);
db64fe02 648
0574ecd1 649 spin_lock(&vmap_area_lock);
763b218d
JF
650 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
651 int nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
652
0574ecd1 653 __free_vmap_area(va);
763b218d
JF
654 atomic_sub(nr, &vmap_lazy_nr);
655 cond_resched_lock(&vmap_area_lock);
656 }
0574ecd1
CH
657 spin_unlock(&vmap_area_lock);
658 return true;
db64fe02
NP
659}
660
496850e5
NP
661/*
662 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
663 * is already purging.
664 */
665static void try_purge_vmap_area_lazy(void)
666{
f9e09977 667 if (mutex_trylock(&vmap_purge_lock)) {
0574ecd1 668 __purge_vmap_area_lazy(ULONG_MAX, 0);
f9e09977 669 mutex_unlock(&vmap_purge_lock);
0574ecd1 670 }
496850e5
NP
671}
672
db64fe02
NP
673/*
674 * Kick off a purge of the outstanding lazy areas.
675 */
676static void purge_vmap_area_lazy(void)
677{
f9e09977 678 mutex_lock(&vmap_purge_lock);
0574ecd1
CH
679 purge_fragmented_blocks_allcpus();
680 __purge_vmap_area_lazy(ULONG_MAX, 0);
f9e09977 681 mutex_unlock(&vmap_purge_lock);
db64fe02
NP
682}
683
684/*
64141da5
JF
685 * Free a vmap area, caller ensuring that the area has been unmapped
686 * and flush_cache_vunmap had been called for the correct range
687 * previously.
db64fe02 688 */
64141da5 689static void free_vmap_area_noflush(struct vmap_area *va)
db64fe02 690{
80c4bd7a
CW
691 int nr_lazy;
692
693 nr_lazy = atomic_add_return((va->va_end - va->va_start) >> PAGE_SHIFT,
694 &vmap_lazy_nr);
695
696 /* After this point, we may free va at any time */
697 llist_add(&va->purge_list, &vmap_purge_list);
698
699 if (unlikely(nr_lazy > lazy_max_pages()))
496850e5 700 try_purge_vmap_area_lazy();
db64fe02
NP
701}
702
b29acbdc
NP
703/*
704 * Free and unmap a vmap area
705 */
706static void free_unmap_vmap_area(struct vmap_area *va)
707{
708 flush_cache_vunmap(va->va_start, va->va_end);
c8eef01e
CH
709 unmap_vmap_area(va);
710 free_vmap_area_noflush(va);
b29acbdc
NP
711}
712
db64fe02
NP
713static struct vmap_area *find_vmap_area(unsigned long addr)
714{
715 struct vmap_area *va;
716
717 spin_lock(&vmap_area_lock);
718 va = __find_vmap_area(addr);
719 spin_unlock(&vmap_area_lock);
720
721 return va;
722}
723
db64fe02
NP
724/*** Per cpu kva allocator ***/
725
726/*
727 * vmap space is limited especially on 32 bit architectures. Ensure there is
728 * room for at least 16 percpu vmap blocks per CPU.
729 */
730/*
731 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
732 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
733 * instead (we just need a rough idea)
734 */
735#if BITS_PER_LONG == 32
736#define VMALLOC_SPACE (128UL*1024*1024)
737#else
738#define VMALLOC_SPACE (128UL*1024*1024*1024)
739#endif
740
741#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
742#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
743#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
744#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
745#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
746#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
f982f915
CL
747#define VMAP_BBMAP_BITS \
748 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
749 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
750 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
db64fe02
NP
751
752#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
753
9b463334
JF
754static bool vmap_initialized __read_mostly = false;
755
db64fe02
NP
756struct vmap_block_queue {
757 spinlock_t lock;
758 struct list_head free;
db64fe02
NP
759};
760
761struct vmap_block {
762 spinlock_t lock;
763 struct vmap_area *va;
db64fe02 764 unsigned long free, dirty;
7d61bfe8 765 unsigned long dirty_min, dirty_max; /*< dirty range */
de560423
NP
766 struct list_head free_list;
767 struct rcu_head rcu_head;
02b709df 768 struct list_head purge;
db64fe02
NP
769};
770
771/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
772static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
773
774/*
775 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
776 * in the free path. Could get rid of this if we change the API to return a
777 * "cookie" from alloc, to be passed to free. But no big deal yet.
778 */
779static DEFINE_SPINLOCK(vmap_block_tree_lock);
780static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
781
782/*
783 * We should probably have a fallback mechanism to allocate virtual memory
784 * out of partially filled vmap blocks. However vmap block sizing should be
785 * fairly reasonable according to the vmalloc size, so it shouldn't be a
786 * big problem.
787 */
788
789static unsigned long addr_to_vb_idx(unsigned long addr)
790{
791 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
792 addr /= VMAP_BLOCK_SIZE;
793 return addr;
794}
795
cf725ce2
RP
796static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
797{
798 unsigned long addr;
799
800 addr = va_start + (pages_off << PAGE_SHIFT);
801 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
802 return (void *)addr;
803}
804
805/**
806 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
807 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
808 * @order: how many 2^order pages should be occupied in newly allocated block
809 * @gfp_mask: flags for the page level allocator
810 *
811 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
812 */
813static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
db64fe02
NP
814{
815 struct vmap_block_queue *vbq;
816 struct vmap_block *vb;
817 struct vmap_area *va;
818 unsigned long vb_idx;
819 int node, err;
cf725ce2 820 void *vaddr;
db64fe02
NP
821
822 node = numa_node_id();
823
824 vb = kmalloc_node(sizeof(struct vmap_block),
825 gfp_mask & GFP_RECLAIM_MASK, node);
826 if (unlikely(!vb))
827 return ERR_PTR(-ENOMEM);
828
829 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
830 VMALLOC_START, VMALLOC_END,
831 node, gfp_mask);
ddf9c6d4 832 if (IS_ERR(va)) {
db64fe02 833 kfree(vb);
e7d86340 834 return ERR_CAST(va);
db64fe02
NP
835 }
836
837 err = radix_tree_preload(gfp_mask);
838 if (unlikely(err)) {
839 kfree(vb);
840 free_vmap_area(va);
841 return ERR_PTR(err);
842 }
843
cf725ce2 844 vaddr = vmap_block_vaddr(va->va_start, 0);
db64fe02
NP
845 spin_lock_init(&vb->lock);
846 vb->va = va;
cf725ce2
RP
847 /* At least something should be left free */
848 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
849 vb->free = VMAP_BBMAP_BITS - (1UL << order);
db64fe02 850 vb->dirty = 0;
7d61bfe8
RP
851 vb->dirty_min = VMAP_BBMAP_BITS;
852 vb->dirty_max = 0;
db64fe02 853 INIT_LIST_HEAD(&vb->free_list);
db64fe02
NP
854
855 vb_idx = addr_to_vb_idx(va->va_start);
856 spin_lock(&vmap_block_tree_lock);
857 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
858 spin_unlock(&vmap_block_tree_lock);
859 BUG_ON(err);
860 radix_tree_preload_end();
861
862 vbq = &get_cpu_var(vmap_block_queue);
db64fe02 863 spin_lock(&vbq->lock);
68ac546f 864 list_add_tail_rcu(&vb->free_list, &vbq->free);
db64fe02 865 spin_unlock(&vbq->lock);
3f04ba85 866 put_cpu_var(vmap_block_queue);
db64fe02 867
cf725ce2 868 return vaddr;
db64fe02
NP
869}
870
db64fe02
NP
871static void free_vmap_block(struct vmap_block *vb)
872{
873 struct vmap_block *tmp;
874 unsigned long vb_idx;
875
db64fe02
NP
876 vb_idx = addr_to_vb_idx(vb->va->va_start);
877 spin_lock(&vmap_block_tree_lock);
878 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
879 spin_unlock(&vmap_block_tree_lock);
880 BUG_ON(tmp != vb);
881
64141da5 882 free_vmap_area_noflush(vb->va);
22a3c7d1 883 kfree_rcu(vb, rcu_head);
db64fe02
NP
884}
885
02b709df
NP
886static void purge_fragmented_blocks(int cpu)
887{
888 LIST_HEAD(purge);
889 struct vmap_block *vb;
890 struct vmap_block *n_vb;
891 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
892
893 rcu_read_lock();
894 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
895
896 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
897 continue;
898
899 spin_lock(&vb->lock);
900 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
901 vb->free = 0; /* prevent further allocs after releasing lock */
902 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
7d61bfe8
RP
903 vb->dirty_min = 0;
904 vb->dirty_max = VMAP_BBMAP_BITS;
02b709df
NP
905 spin_lock(&vbq->lock);
906 list_del_rcu(&vb->free_list);
907 spin_unlock(&vbq->lock);
908 spin_unlock(&vb->lock);
909 list_add_tail(&vb->purge, &purge);
910 } else
911 spin_unlock(&vb->lock);
912 }
913 rcu_read_unlock();
914
915 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
916 list_del(&vb->purge);
917 free_vmap_block(vb);
918 }
919}
920
02b709df
NP
921static void purge_fragmented_blocks_allcpus(void)
922{
923 int cpu;
924
925 for_each_possible_cpu(cpu)
926 purge_fragmented_blocks(cpu);
927}
928
db64fe02
NP
929static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
930{
931 struct vmap_block_queue *vbq;
932 struct vmap_block *vb;
cf725ce2 933 void *vaddr = NULL;
db64fe02
NP
934 unsigned int order;
935
891c49ab 936 BUG_ON(offset_in_page(size));
db64fe02 937 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
aa91c4d8
JK
938 if (WARN_ON(size == 0)) {
939 /*
940 * Allocating 0 bytes isn't what caller wants since
941 * get_order(0) returns funny result. Just warn and terminate
942 * early.
943 */
944 return NULL;
945 }
db64fe02
NP
946 order = get_order(size);
947
db64fe02
NP
948 rcu_read_lock();
949 vbq = &get_cpu_var(vmap_block_queue);
950 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
cf725ce2 951 unsigned long pages_off;
db64fe02
NP
952
953 spin_lock(&vb->lock);
cf725ce2
RP
954 if (vb->free < (1UL << order)) {
955 spin_unlock(&vb->lock);
956 continue;
957 }
02b709df 958
cf725ce2
RP
959 pages_off = VMAP_BBMAP_BITS - vb->free;
960 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
02b709df
NP
961 vb->free -= 1UL << order;
962 if (vb->free == 0) {
963 spin_lock(&vbq->lock);
964 list_del_rcu(&vb->free_list);
965 spin_unlock(&vbq->lock);
966 }
cf725ce2 967
02b709df
NP
968 spin_unlock(&vb->lock);
969 break;
db64fe02 970 }
02b709df 971
3f04ba85 972 put_cpu_var(vmap_block_queue);
db64fe02
NP
973 rcu_read_unlock();
974
cf725ce2
RP
975 /* Allocate new block if nothing was found */
976 if (!vaddr)
977 vaddr = new_vmap_block(order, gfp_mask);
db64fe02 978
cf725ce2 979 return vaddr;
db64fe02
NP
980}
981
982static void vb_free(const void *addr, unsigned long size)
983{
984 unsigned long offset;
985 unsigned long vb_idx;
986 unsigned int order;
987 struct vmap_block *vb;
988
891c49ab 989 BUG_ON(offset_in_page(size));
db64fe02 990 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
b29acbdc
NP
991
992 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
993
db64fe02
NP
994 order = get_order(size);
995
996 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
7d61bfe8 997 offset >>= PAGE_SHIFT;
db64fe02
NP
998
999 vb_idx = addr_to_vb_idx((unsigned long)addr);
1000 rcu_read_lock();
1001 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1002 rcu_read_unlock();
1003 BUG_ON(!vb);
1004
64141da5
JF
1005 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1006
db64fe02 1007 spin_lock(&vb->lock);
7d61bfe8
RP
1008
1009 /* Expand dirty range */
1010 vb->dirty_min = min(vb->dirty_min, offset);
1011 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
d086817d 1012
db64fe02
NP
1013 vb->dirty += 1UL << order;
1014 if (vb->dirty == VMAP_BBMAP_BITS) {
de560423 1015 BUG_ON(vb->free);
db64fe02
NP
1016 spin_unlock(&vb->lock);
1017 free_vmap_block(vb);
1018 } else
1019 spin_unlock(&vb->lock);
1020}
1021
1022/**
1023 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1024 *
1025 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1026 * to amortize TLB flushing overheads. What this means is that any page you
1027 * have now, may, in a former life, have been mapped into kernel virtual
1028 * address by the vmap layer and so there might be some CPUs with TLB entries
1029 * still referencing that page (additional to the regular 1:1 kernel mapping).
1030 *
1031 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1032 * be sure that none of the pages we have control over will have any aliases
1033 * from the vmap layer.
1034 */
1035void vm_unmap_aliases(void)
1036{
1037 unsigned long start = ULONG_MAX, end = 0;
1038 int cpu;
1039 int flush = 0;
1040
9b463334
JF
1041 if (unlikely(!vmap_initialized))
1042 return;
1043
5803ed29
CH
1044 might_sleep();
1045
db64fe02
NP
1046 for_each_possible_cpu(cpu) {
1047 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1048 struct vmap_block *vb;
1049
1050 rcu_read_lock();
1051 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
db64fe02 1052 spin_lock(&vb->lock);
7d61bfe8
RP
1053 if (vb->dirty) {
1054 unsigned long va_start = vb->va->va_start;
db64fe02 1055 unsigned long s, e;
b136be5e 1056
7d61bfe8
RP
1057 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1058 e = va_start + (vb->dirty_max << PAGE_SHIFT);
db64fe02 1059
7d61bfe8
RP
1060 start = min(s, start);
1061 end = max(e, end);
db64fe02 1062
7d61bfe8 1063 flush = 1;
db64fe02
NP
1064 }
1065 spin_unlock(&vb->lock);
1066 }
1067 rcu_read_unlock();
1068 }
1069
f9e09977 1070 mutex_lock(&vmap_purge_lock);
0574ecd1
CH
1071 purge_fragmented_blocks_allcpus();
1072 if (!__purge_vmap_area_lazy(start, end) && flush)
1073 flush_tlb_kernel_range(start, end);
f9e09977 1074 mutex_unlock(&vmap_purge_lock);
db64fe02
NP
1075}
1076EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1077
1078/**
1079 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1080 * @mem: the pointer returned by vm_map_ram
1081 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1082 */
1083void vm_unmap_ram(const void *mem, unsigned int count)
1084{
65ee03c4 1085 unsigned long size = (unsigned long)count << PAGE_SHIFT;
db64fe02 1086 unsigned long addr = (unsigned long)mem;
9c3acf60 1087 struct vmap_area *va;
db64fe02 1088
5803ed29 1089 might_sleep();
db64fe02
NP
1090 BUG_ON(!addr);
1091 BUG_ON(addr < VMALLOC_START);
1092 BUG_ON(addr > VMALLOC_END);
a1c0b1a0 1093 BUG_ON(!PAGE_ALIGNED(addr));
db64fe02
NP
1094
1095 debug_check_no_locks_freed(mem, size);
cd52858c 1096 vmap_debug_free_range(addr, addr+size);
db64fe02 1097
9c3acf60 1098 if (likely(count <= VMAP_MAX_ALLOC)) {
db64fe02 1099 vb_free(mem, size);
9c3acf60
CH
1100 return;
1101 }
1102
1103 va = find_vmap_area(addr);
1104 BUG_ON(!va);
1105 free_unmap_vmap_area(va);
db64fe02
NP
1106}
1107EXPORT_SYMBOL(vm_unmap_ram);
1108
1109/**
1110 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1111 * @pages: an array of pointers to the pages to be mapped
1112 * @count: number of pages
1113 * @node: prefer to allocate data structures on this node
1114 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
e99c97ad 1115 *
36437638
GK
1116 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1117 * faster than vmap so it's good. But if you mix long-life and short-life
1118 * objects with vm_map_ram(), it could consume lots of address space through
1119 * fragmentation (especially on a 32bit machine). You could see failures in
1120 * the end. Please use this function for short-lived objects.
1121 *
e99c97ad 1122 * Returns: a pointer to the address that has been mapped, or %NULL on failure
db64fe02
NP
1123 */
1124void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1125{
65ee03c4 1126 unsigned long size = (unsigned long)count << PAGE_SHIFT;
db64fe02
NP
1127 unsigned long addr;
1128 void *mem;
1129
1130 if (likely(count <= VMAP_MAX_ALLOC)) {
1131 mem = vb_alloc(size, GFP_KERNEL);
1132 if (IS_ERR(mem))
1133 return NULL;
1134 addr = (unsigned long)mem;
1135 } else {
1136 struct vmap_area *va;
1137 va = alloc_vmap_area(size, PAGE_SIZE,
1138 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1139 if (IS_ERR(va))
1140 return NULL;
1141
1142 addr = va->va_start;
1143 mem = (void *)addr;
1144 }
1145 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1146 vm_unmap_ram(mem, count);
1147 return NULL;
1148 }
1149 return mem;
1150}
1151EXPORT_SYMBOL(vm_map_ram);
1152
4341fa45 1153static struct vm_struct *vmlist __initdata;
be9b7335
NP
1154/**
1155 * vm_area_add_early - add vmap area early during boot
1156 * @vm: vm_struct to add
1157 *
1158 * This function is used to add fixed kernel vm area to vmlist before
1159 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1160 * should contain proper values and the other fields should be zero.
1161 *
1162 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1163 */
1164void __init vm_area_add_early(struct vm_struct *vm)
1165{
1166 struct vm_struct *tmp, **p;
1167
1168 BUG_ON(vmap_initialized);
1169 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1170 if (tmp->addr >= vm->addr) {
1171 BUG_ON(tmp->addr < vm->addr + vm->size);
1172 break;
1173 } else
1174 BUG_ON(tmp->addr + tmp->size > vm->addr);
1175 }
1176 vm->next = *p;
1177 *p = vm;
1178}
1179
f0aa6617
TH
1180/**
1181 * vm_area_register_early - register vmap area early during boot
1182 * @vm: vm_struct to register
c0c0a293 1183 * @align: requested alignment
f0aa6617
TH
1184 *
1185 * This function is used to register kernel vm area before
1186 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1187 * proper values on entry and other fields should be zero. On return,
1188 * vm->addr contains the allocated address.
1189 *
1190 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1191 */
c0c0a293 1192void __init vm_area_register_early(struct vm_struct *vm, size_t align)
f0aa6617
TH
1193{
1194 static size_t vm_init_off __initdata;
c0c0a293
TH
1195 unsigned long addr;
1196
1197 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1198 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
f0aa6617 1199
c0c0a293 1200 vm->addr = (void *)addr;
f0aa6617 1201
be9b7335 1202 vm_area_add_early(vm);
f0aa6617
TH
1203}
1204
db64fe02
NP
1205void __init vmalloc_init(void)
1206{
822c18f2
IK
1207 struct vmap_area *va;
1208 struct vm_struct *tmp;
db64fe02
NP
1209 int i;
1210
1211 for_each_possible_cpu(i) {
1212 struct vmap_block_queue *vbq;
32fcfd40 1213 struct vfree_deferred *p;
db64fe02
NP
1214
1215 vbq = &per_cpu(vmap_block_queue, i);
1216 spin_lock_init(&vbq->lock);
1217 INIT_LIST_HEAD(&vbq->free);
32fcfd40
AV
1218 p = &per_cpu(vfree_deferred, i);
1219 init_llist_head(&p->list);
1220 INIT_WORK(&p->wq, free_work);
db64fe02 1221 }
9b463334 1222
822c18f2
IK
1223 /* Import existing vmlist entries. */
1224 for (tmp = vmlist; tmp; tmp = tmp->next) {
43ebdac4 1225 va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
dbda591d 1226 va->flags = VM_VM_AREA;
822c18f2
IK
1227 va->va_start = (unsigned long)tmp->addr;
1228 va->va_end = va->va_start + tmp->size;
dbda591d 1229 va->vm = tmp;
822c18f2
IK
1230 __insert_vmap_area(va);
1231 }
ca23e405
TH
1232
1233 vmap_area_pcpu_hole = VMALLOC_END;
1234
9b463334 1235 vmap_initialized = true;
db64fe02
NP
1236}
1237
8fc48985
TH
1238/**
1239 * map_kernel_range_noflush - map kernel VM area with the specified pages
1240 * @addr: start of the VM area to map
1241 * @size: size of the VM area to map
1242 * @prot: page protection flags to use
1243 * @pages: pages to map
1244 *
1245 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1246 * specify should have been allocated using get_vm_area() and its
1247 * friends.
1248 *
1249 * NOTE:
1250 * This function does NOT do any cache flushing. The caller is
1251 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1252 * before calling this function.
1253 *
1254 * RETURNS:
1255 * The number of pages mapped on success, -errno on failure.
1256 */
1257int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1258 pgprot_t prot, struct page **pages)
1259{
1260 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1261}
1262
1263/**
1264 * unmap_kernel_range_noflush - unmap kernel VM area
1265 * @addr: start of the VM area to unmap
1266 * @size: size of the VM area to unmap
1267 *
1268 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1269 * specify should have been allocated using get_vm_area() and its
1270 * friends.
1271 *
1272 * NOTE:
1273 * This function does NOT do any cache flushing. The caller is
1274 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1275 * before calling this function and flush_tlb_kernel_range() after.
1276 */
1277void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1278{
1279 vunmap_page_range(addr, addr + size);
1280}
81e88fdc 1281EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
8fc48985
TH
1282
1283/**
1284 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1285 * @addr: start of the VM area to unmap
1286 * @size: size of the VM area to unmap
1287 *
1288 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1289 * the unmapping and tlb after.
1290 */
db64fe02
NP
1291void unmap_kernel_range(unsigned long addr, unsigned long size)
1292{
1293 unsigned long end = addr + size;
f6fcba70
TH
1294
1295 flush_cache_vunmap(addr, end);
db64fe02
NP
1296 vunmap_page_range(addr, end);
1297 flush_tlb_kernel_range(addr, end);
1298}
93ef6d6c 1299EXPORT_SYMBOL_GPL(unmap_kernel_range);
db64fe02 1300
f6f8ed47 1301int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
db64fe02
NP
1302{
1303 unsigned long addr = (unsigned long)area->addr;
762216ab 1304 unsigned long end = addr + get_vm_area_size(area);
db64fe02
NP
1305 int err;
1306
f6f8ed47 1307 err = vmap_page_range(addr, end, prot, pages);
db64fe02 1308
f6f8ed47 1309 return err > 0 ? 0 : err;
db64fe02
NP
1310}
1311EXPORT_SYMBOL_GPL(map_vm_area);
1312
f5252e00 1313static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
5e6cafc8 1314 unsigned long flags, const void *caller)
cf88c790 1315{
c69480ad 1316 spin_lock(&vmap_area_lock);
cf88c790
TH
1317 vm->flags = flags;
1318 vm->addr = (void *)va->va_start;
1319 vm->size = va->va_end - va->va_start;
1320 vm->caller = caller;
db1aecaf 1321 va->vm = vm;
cf88c790 1322 va->flags |= VM_VM_AREA;
c69480ad 1323 spin_unlock(&vmap_area_lock);
f5252e00 1324}
cf88c790 1325
20fc02b4 1326static void clear_vm_uninitialized_flag(struct vm_struct *vm)
f5252e00 1327{
d4033afd 1328 /*
20fc02b4 1329 * Before removing VM_UNINITIALIZED,
d4033afd
JK
1330 * we should make sure that vm has proper values.
1331 * Pair with smp_rmb() in show_numa_info().
1332 */
1333 smp_wmb();
20fc02b4 1334 vm->flags &= ~VM_UNINITIALIZED;
cf88c790
TH
1335}
1336
db64fe02 1337static struct vm_struct *__get_vm_area_node(unsigned long size,
2dca6999 1338 unsigned long align, unsigned long flags, unsigned long start,
5e6cafc8 1339 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
db64fe02 1340{
0006526d 1341 struct vmap_area *va;
db64fe02 1342 struct vm_struct *area;
1da177e4 1343
52fd24ca 1344 BUG_ON(in_interrupt());
1da177e4 1345 size = PAGE_ALIGN(size);
31be8309
OH
1346 if (unlikely(!size))
1347 return NULL;
1da177e4 1348
252e5c6e 1349 if (flags & VM_IOREMAP)
1350 align = 1ul << clamp_t(int, get_count_order_long(size),
1351 PAGE_SHIFT, IOREMAP_MAX_ORDER);
1352
cf88c790 1353 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1da177e4
LT
1354 if (unlikely(!area))
1355 return NULL;
1356
71394fe5
AR
1357 if (!(flags & VM_NO_GUARD))
1358 size += PAGE_SIZE;
1da177e4 1359
db64fe02
NP
1360 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1361 if (IS_ERR(va)) {
1362 kfree(area);
1363 return NULL;
1da177e4 1364 }
1da177e4 1365
d82b1d85 1366 setup_vmalloc_vm(area, va, flags, caller);
f5252e00 1367
1da177e4 1368 return area;
1da177e4
LT
1369}
1370
930fc45a
CL
1371struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1372 unsigned long start, unsigned long end)
1373{
00ef2d2f
DR
1374 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1375 GFP_KERNEL, __builtin_return_address(0));
930fc45a 1376}
5992b6da 1377EXPORT_SYMBOL_GPL(__get_vm_area);
930fc45a 1378
c2968612
BH
1379struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1380 unsigned long start, unsigned long end,
5e6cafc8 1381 const void *caller)
c2968612 1382{
00ef2d2f
DR
1383 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1384 GFP_KERNEL, caller);
c2968612
BH
1385}
1386
1da177e4 1387/**
183ff22b 1388 * get_vm_area - reserve a contiguous kernel virtual area
1da177e4
LT
1389 * @size: size of the area
1390 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
1391 *
1392 * Search an area of @size in the kernel virtual mapping area,
1393 * and reserved it for out purposes. Returns the area descriptor
1394 * on success or %NULL on failure.
1395 */
1396struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1397{
2dca6999 1398 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
00ef2d2f
DR
1399 NUMA_NO_NODE, GFP_KERNEL,
1400 __builtin_return_address(0));
23016969
CL
1401}
1402
1403struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
5e6cafc8 1404 const void *caller)
23016969 1405{
2dca6999 1406 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
00ef2d2f 1407 NUMA_NO_NODE, GFP_KERNEL, caller);
1da177e4
LT
1408}
1409
e9da6e99
MS
1410/**
1411 * find_vm_area - find a continuous kernel virtual area
1412 * @addr: base address
1413 *
1414 * Search for the kernel VM area starting at @addr, and return it.
1415 * It is up to the caller to do all required locking to keep the returned
1416 * pointer valid.
1417 */
1418struct vm_struct *find_vm_area(const void *addr)
83342314 1419{
db64fe02 1420 struct vmap_area *va;
83342314 1421
db64fe02
NP
1422 va = find_vmap_area((unsigned long)addr);
1423 if (va && va->flags & VM_VM_AREA)
db1aecaf 1424 return va->vm;
1da177e4 1425
1da177e4 1426 return NULL;
1da177e4
LT
1427}
1428
7856dfeb 1429/**
183ff22b 1430 * remove_vm_area - find and remove a continuous kernel virtual area
7856dfeb
AK
1431 * @addr: base address
1432 *
1433 * Search for the kernel VM area starting at @addr, and remove it.
1434 * This function returns the found VM area, but using it is NOT safe
1435 * on SMP machines, except for its size or flags.
1436 */
b3bdda02 1437struct vm_struct *remove_vm_area(const void *addr)
7856dfeb 1438{
db64fe02
NP
1439 struct vmap_area *va;
1440
5803ed29
CH
1441 might_sleep();
1442
db64fe02
NP
1443 va = find_vmap_area((unsigned long)addr);
1444 if (va && va->flags & VM_VM_AREA) {
db1aecaf 1445 struct vm_struct *vm = va->vm;
f5252e00 1446
c69480ad
JK
1447 spin_lock(&vmap_area_lock);
1448 va->vm = NULL;
1449 va->flags &= ~VM_VM_AREA;
1450 spin_unlock(&vmap_area_lock);
1451
dd32c279 1452 vmap_debug_free_range(va->va_start, va->va_end);
a5af5aa8 1453 kasan_free_shadow(vm);
dd32c279 1454 free_unmap_vmap_area(va);
dd32c279 1455
db64fe02
NP
1456 return vm;
1457 }
1458 return NULL;
7856dfeb
AK
1459}
1460
b3bdda02 1461static void __vunmap(const void *addr, int deallocate_pages)
1da177e4
LT
1462{
1463 struct vm_struct *area;
1464
1465 if (!addr)
1466 return;
1467
e69e9d4a 1468 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
ab15d9b4 1469 addr))
1da177e4 1470 return;
1da177e4
LT
1471
1472 area = remove_vm_area(addr);
1473 if (unlikely(!area)) {
4c8573e2 1474 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1da177e4 1475 addr);
1da177e4
LT
1476 return;
1477 }
1478
7511c3ed
JM
1479 debug_check_no_locks_freed(addr, get_vm_area_size(area));
1480 debug_check_no_obj_freed(addr, get_vm_area_size(area));
9a11b49a 1481
1da177e4
LT
1482 if (deallocate_pages) {
1483 int i;
1484
1485 for (i = 0; i < area->nr_pages; i++) {
bf53d6f8
CL
1486 struct page *page = area->pages[i];
1487
1488 BUG_ON(!page);
4949148a 1489 __free_pages(page, 0);
1da177e4
LT
1490 }
1491
244d63ee 1492 kvfree(area->pages);
1da177e4
LT
1493 }
1494
1495 kfree(area);
1496 return;
1497}
bf22e37a
AR
1498
1499static inline void __vfree_deferred(const void *addr)
1500{
1501 /*
1502 * Use raw_cpu_ptr() because this can be called from preemptible
1503 * context. Preemption is absolutely fine here, because the llist_add()
1504 * implementation is lockless, so it works even if we are adding to
1505 * nother cpu's list. schedule_work() should be fine with this too.
1506 */
1507 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
1508
1509 if (llist_add((struct llist_node *)addr, &p->list))
1510 schedule_work(&p->wq);
1511}
1512
1513/**
1514 * vfree_atomic - release memory allocated by vmalloc()
1515 * @addr: memory base address
1516 *
1517 * This one is just like vfree() but can be called in any atomic context
1518 * except NMIs.
1519 */
1520void vfree_atomic(const void *addr)
1521{
1522 BUG_ON(in_nmi());
1523
1524 kmemleak_free(addr);
1525
1526 if (!addr)
1527 return;
1528 __vfree_deferred(addr);
1529}
1530
1da177e4
LT
1531/**
1532 * vfree - release memory allocated by vmalloc()
1da177e4
LT
1533 * @addr: memory base address
1534 *
183ff22b 1535 * Free the virtually continuous memory area starting at @addr, as
80e93eff
PE
1536 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1537 * NULL, no operation is performed.
1da177e4 1538 *
32fcfd40
AV
1539 * Must not be called in NMI context (strictly speaking, only if we don't
1540 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1541 * conventions for vfree() arch-depenedent would be a really bad idea)
c9fcee51
AM
1542 *
1543 * NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1da177e4 1544 */
b3bdda02 1545void vfree(const void *addr)
1da177e4 1546{
32fcfd40 1547 BUG_ON(in_nmi());
89219d37
CM
1548
1549 kmemleak_free(addr);
1550
32fcfd40
AV
1551 if (!addr)
1552 return;
bf22e37a
AR
1553 if (unlikely(in_interrupt()))
1554 __vfree_deferred(addr);
1555 else
32fcfd40 1556 __vunmap(addr, 1);
1da177e4 1557}
1da177e4
LT
1558EXPORT_SYMBOL(vfree);
1559
1560/**
1561 * vunmap - release virtual mapping obtained by vmap()
1da177e4
LT
1562 * @addr: memory base address
1563 *
1564 * Free the virtually contiguous memory area starting at @addr,
1565 * which was created from the page array passed to vmap().
1566 *
80e93eff 1567 * Must not be called in interrupt context.
1da177e4 1568 */
b3bdda02 1569void vunmap(const void *addr)
1da177e4
LT
1570{
1571 BUG_ON(in_interrupt());
34754b69 1572 might_sleep();
32fcfd40
AV
1573 if (addr)
1574 __vunmap(addr, 0);
1da177e4 1575}
1da177e4
LT
1576EXPORT_SYMBOL(vunmap);
1577
1578/**
1579 * vmap - map an array of pages into virtually contiguous space
1da177e4
LT
1580 * @pages: array of page pointers
1581 * @count: number of pages to map
1582 * @flags: vm_area->flags
1583 * @prot: page protection for the mapping
1584 *
1585 * Maps @count pages from @pages into contiguous kernel virtual
1586 * space.
1587 */
1588void *vmap(struct page **pages, unsigned int count,
1589 unsigned long flags, pgprot_t prot)
1590{
1591 struct vm_struct *area;
65ee03c4 1592 unsigned long size; /* In bytes */
1da177e4 1593
34754b69
PZ
1594 might_sleep();
1595
4481374c 1596 if (count > totalram_pages)
1da177e4
LT
1597 return NULL;
1598
65ee03c4
GJM
1599 size = (unsigned long)count << PAGE_SHIFT;
1600 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
1da177e4
LT
1601 if (!area)
1602 return NULL;
23016969 1603
f6f8ed47 1604 if (map_vm_area(area, prot, pages)) {
1da177e4
LT
1605 vunmap(area->addr);
1606 return NULL;
1607 }
1608
1609 return area->addr;
1610}
1da177e4
LT
1611EXPORT_SYMBOL(vmap);
1612
2dca6999
DM
1613static void *__vmalloc_node(unsigned long size, unsigned long align,
1614 gfp_t gfp_mask, pgprot_t prot,
5e6cafc8 1615 int node, const void *caller);
e31d9eb5 1616static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3722e13c 1617 pgprot_t prot, int node)
1da177e4
LT
1618{
1619 struct page **pages;
1620 unsigned int nr_pages, array_size, i;
930f036b
DR
1621 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1622 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1da177e4 1623
762216ab 1624 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1da177e4
LT
1625 array_size = (nr_pages * sizeof(struct page *));
1626
1627 area->nr_pages = nr_pages;
1628 /* Please note that the recursion is strictly bounded. */
8757d5fa 1629 if (array_size > PAGE_SIZE) {
976d6dfb 1630 pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
3722e13c 1631 PAGE_KERNEL, node, area->caller);
286e1ea3 1632 } else {
976d6dfb 1633 pages = kmalloc_node(array_size, nested_gfp, node);
286e1ea3 1634 }
1da177e4
LT
1635 area->pages = pages;
1636 if (!area->pages) {
1637 remove_vm_area(area->addr);
1638 kfree(area);
1639 return NULL;
1640 }
1da177e4
LT
1641
1642 for (i = 0; i < area->nr_pages; i++) {
bf53d6f8
CL
1643 struct page *page;
1644
4b90951c 1645 if (node == NUMA_NO_NODE)
7877cdcc 1646 page = alloc_page(alloc_mask);
930fc45a 1647 else
7877cdcc 1648 page = alloc_pages_node(node, alloc_mask, 0);
bf53d6f8
CL
1649
1650 if (unlikely(!page)) {
1da177e4
LT
1651 /* Successfully allocated i pages, free them in __vunmap() */
1652 area->nr_pages = i;
1653 goto fail;
1654 }
bf53d6f8 1655 area->pages[i] = page;
d0164adc 1656 if (gfpflags_allow_blocking(gfp_mask))
660654f9 1657 cond_resched();
1da177e4
LT
1658 }
1659
f6f8ed47 1660 if (map_vm_area(area, prot, pages))
1da177e4
LT
1661 goto fail;
1662 return area->addr;
1663
1664fail:
7877cdcc
MH
1665 warn_alloc(gfp_mask,
1666 "vmalloc: allocation failure, allocated %ld of %ld bytes",
22943ab1 1667 (area->nr_pages*PAGE_SIZE), area->size);
1da177e4
LT
1668 vfree(area->addr);
1669 return NULL;
1670}
1671
1672/**
d0a21265 1673 * __vmalloc_node_range - allocate virtually contiguous memory
1da177e4 1674 * @size: allocation size
2dca6999 1675 * @align: desired alignment
d0a21265
DR
1676 * @start: vm area range start
1677 * @end: vm area range end
1da177e4
LT
1678 * @gfp_mask: flags for the page level allocator
1679 * @prot: protection mask for the allocated pages
cb9e3c29 1680 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
00ef2d2f 1681 * @node: node to use for allocation or NUMA_NO_NODE
c85d194b 1682 * @caller: caller's return address
1da177e4
LT
1683 *
1684 * Allocate enough pages to cover @size from the page level
1685 * allocator with @gfp_mask flags. Map them into contiguous
1686 * kernel virtual space, using a pagetable protection of @prot.
1687 */
d0a21265
DR
1688void *__vmalloc_node_range(unsigned long size, unsigned long align,
1689 unsigned long start, unsigned long end, gfp_t gfp_mask,
cb9e3c29
AR
1690 pgprot_t prot, unsigned long vm_flags, int node,
1691 const void *caller)
1da177e4
LT
1692{
1693 struct vm_struct *area;
89219d37
CM
1694 void *addr;
1695 unsigned long real_size = size;
1da177e4
LT
1696
1697 size = PAGE_ALIGN(size);
4481374c 1698 if (!size || (size >> PAGE_SHIFT) > totalram_pages)
de7d2b56 1699 goto fail;
1da177e4 1700
cb9e3c29
AR
1701 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1702 vm_flags, start, end, node, gfp_mask, caller);
1da177e4 1703 if (!area)
de7d2b56 1704 goto fail;
1da177e4 1705
3722e13c 1706 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1368edf0 1707 if (!addr)
b82225f3 1708 return NULL;
89219d37 1709
f5252e00 1710 /*
20fc02b4
ZY
1711 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1712 * flag. It means that vm_struct is not fully initialized.
4341fa45 1713 * Now, it is fully initialized, so remove this flag here.
f5252e00 1714 */
20fc02b4 1715 clear_vm_uninitialized_flag(area);
f5252e00 1716
89219d37 1717 /*
7f88f88f
CM
1718 * A ref_count = 2 is needed because vm_struct allocated in
1719 * __get_vm_area_node() contains a reference to the virtual address of
1720 * the vmalloc'ed block.
89219d37 1721 */
7f88f88f 1722 kmemleak_alloc(addr, real_size, 2, gfp_mask);
89219d37
CM
1723
1724 return addr;
de7d2b56
JP
1725
1726fail:
7877cdcc
MH
1727 warn_alloc(gfp_mask,
1728 "vmalloc: allocation failure: %lu bytes", real_size);
de7d2b56 1729 return NULL;
1da177e4
LT
1730}
1731
d0a21265
DR
1732/**
1733 * __vmalloc_node - allocate virtually contiguous memory
1734 * @size: allocation size
1735 * @align: desired alignment
1736 * @gfp_mask: flags for the page level allocator
1737 * @prot: protection mask for the allocated pages
00ef2d2f 1738 * @node: node to use for allocation or NUMA_NO_NODE
d0a21265
DR
1739 * @caller: caller's return address
1740 *
1741 * Allocate enough pages to cover @size from the page level
1742 * allocator with @gfp_mask flags. Map them into contiguous
1743 * kernel virtual space, using a pagetable protection of @prot.
1744 */
1745static void *__vmalloc_node(unsigned long size, unsigned long align,
1746 gfp_t gfp_mask, pgprot_t prot,
5e6cafc8 1747 int node, const void *caller)
d0a21265
DR
1748{
1749 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
cb9e3c29 1750 gfp_mask, prot, 0, node, caller);
d0a21265
DR
1751}
1752
930fc45a
CL
1753void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1754{
00ef2d2f 1755 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
23016969 1756 __builtin_return_address(0));
930fc45a 1757}
1da177e4
LT
1758EXPORT_SYMBOL(__vmalloc);
1759
e1ca7788
DY
1760static inline void *__vmalloc_node_flags(unsigned long size,
1761 int node, gfp_t flags)
1762{
1763 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1764 node, __builtin_return_address(0));
1765}
1766
1da177e4
LT
1767/**
1768 * vmalloc - allocate virtually contiguous memory
1da177e4 1769 * @size: allocation size
1da177e4
LT
1770 * Allocate enough pages to cover @size from the page level
1771 * allocator and map them into contiguous kernel virtual space.
1772 *
c1c8897f 1773 * For tight control over page level allocator and protection flags
1da177e4
LT
1774 * use __vmalloc() instead.
1775 */
1776void *vmalloc(unsigned long size)
1777{
00ef2d2f
DR
1778 return __vmalloc_node_flags(size, NUMA_NO_NODE,
1779 GFP_KERNEL | __GFP_HIGHMEM);
1da177e4 1780}
1da177e4
LT
1781EXPORT_SYMBOL(vmalloc);
1782
e1ca7788
DY
1783/**
1784 * vzalloc - allocate virtually contiguous memory with zero fill
1785 * @size: allocation size
1786 * Allocate enough pages to cover @size from the page level
1787 * allocator and map them into contiguous kernel virtual space.
1788 * The memory allocated is set to zero.
1789 *
1790 * For tight control over page level allocator and protection flags
1791 * use __vmalloc() instead.
1792 */
1793void *vzalloc(unsigned long size)
1794{
00ef2d2f 1795 return __vmalloc_node_flags(size, NUMA_NO_NODE,
e1ca7788
DY
1796 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1797}
1798EXPORT_SYMBOL(vzalloc);
1799
83342314 1800/**
ead04089
REB
1801 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1802 * @size: allocation size
83342314 1803 *
ead04089
REB
1804 * The resulting memory area is zeroed so it can be mapped to userspace
1805 * without leaking data.
83342314
NP
1806 */
1807void *vmalloc_user(unsigned long size)
1808{
1809 struct vm_struct *area;
1810 void *ret;
1811
2dca6999
DM
1812 ret = __vmalloc_node(size, SHMLBA,
1813 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
00ef2d2f
DR
1814 PAGE_KERNEL, NUMA_NO_NODE,
1815 __builtin_return_address(0));
2b4ac44e 1816 if (ret) {
db64fe02 1817 area = find_vm_area(ret);
2b4ac44e 1818 area->flags |= VM_USERMAP;
2b4ac44e 1819 }
83342314
NP
1820 return ret;
1821}
1822EXPORT_SYMBOL(vmalloc_user);
1823
930fc45a
CL
1824/**
1825 * vmalloc_node - allocate memory on a specific node
930fc45a 1826 * @size: allocation size
d44e0780 1827 * @node: numa node
930fc45a
CL
1828 *
1829 * Allocate enough pages to cover @size from the page level
1830 * allocator and map them into contiguous kernel virtual space.
1831 *
c1c8897f 1832 * For tight control over page level allocator and protection flags
930fc45a
CL
1833 * use __vmalloc() instead.
1834 */
1835void *vmalloc_node(unsigned long size, int node)
1836{
2dca6999 1837 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
23016969 1838 node, __builtin_return_address(0));
930fc45a
CL
1839}
1840EXPORT_SYMBOL(vmalloc_node);
1841
e1ca7788
DY
1842/**
1843 * vzalloc_node - allocate memory on a specific node with zero fill
1844 * @size: allocation size
1845 * @node: numa node
1846 *
1847 * Allocate enough pages to cover @size from the page level
1848 * allocator and map them into contiguous kernel virtual space.
1849 * The memory allocated is set to zero.
1850 *
1851 * For tight control over page level allocator and protection flags
1852 * use __vmalloc_node() instead.
1853 */
1854void *vzalloc_node(unsigned long size, int node)
1855{
1856 return __vmalloc_node_flags(size, node,
1857 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1858}
1859EXPORT_SYMBOL(vzalloc_node);
1860
4dc3b16b
PP
1861#ifndef PAGE_KERNEL_EXEC
1862# define PAGE_KERNEL_EXEC PAGE_KERNEL
1863#endif
1864
1da177e4
LT
1865/**
1866 * vmalloc_exec - allocate virtually contiguous, executable memory
1da177e4
LT
1867 * @size: allocation size
1868 *
1869 * Kernel-internal function to allocate enough pages to cover @size
1870 * the page level allocator and map them into contiguous and
1871 * executable kernel virtual space.
1872 *
c1c8897f 1873 * For tight control over page level allocator and protection flags
1da177e4
LT
1874 * use __vmalloc() instead.
1875 */
1876
1da177e4
LT
1877void *vmalloc_exec(unsigned long size)
1878{
2dca6999 1879 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
00ef2d2f 1880 NUMA_NO_NODE, __builtin_return_address(0));
1da177e4
LT
1881}
1882
0d08e0d3 1883#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
7ac674f5 1884#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
0d08e0d3 1885#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
7ac674f5 1886#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
0d08e0d3
AK
1887#else
1888#define GFP_VMALLOC32 GFP_KERNEL
1889#endif
1890
1da177e4
LT
1891/**
1892 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
1da177e4
LT
1893 * @size: allocation size
1894 *
1895 * Allocate enough 32bit PA addressable pages to cover @size from the
1896 * page level allocator and map them into contiguous kernel virtual space.
1897 */
1898void *vmalloc_32(unsigned long size)
1899{
2dca6999 1900 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
00ef2d2f 1901 NUMA_NO_NODE, __builtin_return_address(0));
1da177e4 1902}
1da177e4
LT
1903EXPORT_SYMBOL(vmalloc_32);
1904
83342314 1905/**
ead04089 1906 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
83342314 1907 * @size: allocation size
ead04089
REB
1908 *
1909 * The resulting memory area is 32bit addressable and zeroed so it can be
1910 * mapped to userspace without leaking data.
83342314
NP
1911 */
1912void *vmalloc_32_user(unsigned long size)
1913{
1914 struct vm_struct *area;
1915 void *ret;
1916
2dca6999 1917 ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
00ef2d2f 1918 NUMA_NO_NODE, __builtin_return_address(0));
2b4ac44e 1919 if (ret) {
db64fe02 1920 area = find_vm_area(ret);
2b4ac44e 1921 area->flags |= VM_USERMAP;
2b4ac44e 1922 }
83342314
NP
1923 return ret;
1924}
1925EXPORT_SYMBOL(vmalloc_32_user);
1926
d0107eb0
KH
1927/*
1928 * small helper routine , copy contents to buf from addr.
1929 * If the page is not present, fill zero.
1930 */
1931
1932static int aligned_vread(char *buf, char *addr, unsigned long count)
1933{
1934 struct page *p;
1935 int copied = 0;
1936
1937 while (count) {
1938 unsigned long offset, length;
1939
891c49ab 1940 offset = offset_in_page(addr);
d0107eb0
KH
1941 length = PAGE_SIZE - offset;
1942 if (length > count)
1943 length = count;
1944 p = vmalloc_to_page(addr);
1945 /*
1946 * To do safe access to this _mapped_ area, we need
1947 * lock. But adding lock here means that we need to add
1948 * overhead of vmalloc()/vfree() calles for this _debug_
1949 * interface, rarely used. Instead of that, we'll use
1950 * kmap() and get small overhead in this access function.
1951 */
1952 if (p) {
1953 /*
1954 * we can expect USER0 is not used (see vread/vwrite's
1955 * function description)
1956 */
9b04c5fe 1957 void *map = kmap_atomic(p);
d0107eb0 1958 memcpy(buf, map + offset, length);
9b04c5fe 1959 kunmap_atomic(map);
d0107eb0
KH
1960 } else
1961 memset(buf, 0, length);
1962
1963 addr += length;
1964 buf += length;
1965 copied += length;
1966 count -= length;
1967 }
1968 return copied;
1969}
1970
1971static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1972{
1973 struct page *p;
1974 int copied = 0;
1975
1976 while (count) {
1977 unsigned long offset, length;
1978
891c49ab 1979 offset = offset_in_page(addr);
d0107eb0
KH
1980 length = PAGE_SIZE - offset;
1981 if (length > count)
1982 length = count;
1983 p = vmalloc_to_page(addr);
1984 /*
1985 * To do safe access to this _mapped_ area, we need
1986 * lock. But adding lock here means that we need to add
1987 * overhead of vmalloc()/vfree() calles for this _debug_
1988 * interface, rarely used. Instead of that, we'll use
1989 * kmap() and get small overhead in this access function.
1990 */
1991 if (p) {
1992 /*
1993 * we can expect USER0 is not used (see vread/vwrite's
1994 * function description)
1995 */
9b04c5fe 1996 void *map = kmap_atomic(p);
d0107eb0 1997 memcpy(map + offset, buf, length);
9b04c5fe 1998 kunmap_atomic(map);
d0107eb0
KH
1999 }
2000 addr += length;
2001 buf += length;
2002 copied += length;
2003 count -= length;
2004 }
2005 return copied;
2006}
2007
2008/**
2009 * vread() - read vmalloc area in a safe way.
2010 * @buf: buffer for reading data
2011 * @addr: vm address.
2012 * @count: number of bytes to be read.
2013 *
2014 * Returns # of bytes which addr and buf should be increased.
2015 * (same number to @count). Returns 0 if [addr...addr+count) doesn't
2016 * includes any intersect with alive vmalloc area.
2017 *
2018 * This function checks that addr is a valid vmalloc'ed area, and
2019 * copy data from that area to a given buffer. If the given memory range
2020 * of [addr...addr+count) includes some valid address, data is copied to
2021 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2022 * IOREMAP area is treated as memory hole and no copy is done.
2023 *
2024 * If [addr...addr+count) doesn't includes any intersects with alive
a8e5202d 2025 * vm_struct area, returns 0. @buf should be kernel's buffer.
d0107eb0
KH
2026 *
2027 * Note: In usual ops, vread() is never necessary because the caller
2028 * should know vmalloc() area is valid and can use memcpy().
2029 * This is for routines which have to access vmalloc area without
2030 * any informaion, as /dev/kmem.
2031 *
2032 */
2033
1da177e4
LT
2034long vread(char *buf, char *addr, unsigned long count)
2035{
e81ce85f
JK
2036 struct vmap_area *va;
2037 struct vm_struct *vm;
1da177e4 2038 char *vaddr, *buf_start = buf;
d0107eb0 2039 unsigned long buflen = count;
1da177e4
LT
2040 unsigned long n;
2041
2042 /* Don't allow overflow */
2043 if ((unsigned long) addr + count < count)
2044 count = -(unsigned long) addr;
2045
e81ce85f
JK
2046 spin_lock(&vmap_area_lock);
2047 list_for_each_entry(va, &vmap_area_list, list) {
2048 if (!count)
2049 break;
2050
2051 if (!(va->flags & VM_VM_AREA))
2052 continue;
2053
2054 vm = va->vm;
2055 vaddr = (char *) vm->addr;
762216ab 2056 if (addr >= vaddr + get_vm_area_size(vm))
1da177e4
LT
2057 continue;
2058 while (addr < vaddr) {
2059 if (count == 0)
2060 goto finished;
2061 *buf = '\0';
2062 buf++;
2063 addr++;
2064 count--;
2065 }
762216ab 2066 n = vaddr + get_vm_area_size(vm) - addr;
d0107eb0
KH
2067 if (n > count)
2068 n = count;
e81ce85f 2069 if (!(vm->flags & VM_IOREMAP))
d0107eb0
KH
2070 aligned_vread(buf, addr, n);
2071 else /* IOREMAP area is treated as memory hole */
2072 memset(buf, 0, n);
2073 buf += n;
2074 addr += n;
2075 count -= n;
1da177e4
LT
2076 }
2077finished:
e81ce85f 2078 spin_unlock(&vmap_area_lock);
d0107eb0
KH
2079
2080 if (buf == buf_start)
2081 return 0;
2082 /* zero-fill memory holes */
2083 if (buf != buf_start + buflen)
2084 memset(buf, 0, buflen - (buf - buf_start));
2085
2086 return buflen;
1da177e4
LT
2087}
2088
d0107eb0
KH
2089/**
2090 * vwrite() - write vmalloc area in a safe way.
2091 * @buf: buffer for source data
2092 * @addr: vm address.
2093 * @count: number of bytes to be read.
2094 *
2095 * Returns # of bytes which addr and buf should be incresed.
2096 * (same number to @count).
2097 * If [addr...addr+count) doesn't includes any intersect with valid
2098 * vmalloc area, returns 0.
2099 *
2100 * This function checks that addr is a valid vmalloc'ed area, and
2101 * copy data from a buffer to the given addr. If specified range of
2102 * [addr...addr+count) includes some valid address, data is copied from
2103 * proper area of @buf. If there are memory holes, no copy to hole.
2104 * IOREMAP area is treated as memory hole and no copy is done.
2105 *
2106 * If [addr...addr+count) doesn't includes any intersects with alive
a8e5202d 2107 * vm_struct area, returns 0. @buf should be kernel's buffer.
d0107eb0
KH
2108 *
2109 * Note: In usual ops, vwrite() is never necessary because the caller
2110 * should know vmalloc() area is valid and can use memcpy().
2111 * This is for routines which have to access vmalloc area without
2112 * any informaion, as /dev/kmem.
d0107eb0
KH
2113 */
2114
1da177e4
LT
2115long vwrite(char *buf, char *addr, unsigned long count)
2116{
e81ce85f
JK
2117 struct vmap_area *va;
2118 struct vm_struct *vm;
d0107eb0
KH
2119 char *vaddr;
2120 unsigned long n, buflen;
2121 int copied = 0;
1da177e4
LT
2122
2123 /* Don't allow overflow */
2124 if ((unsigned long) addr + count < count)
2125 count = -(unsigned long) addr;
d0107eb0 2126 buflen = count;
1da177e4 2127
e81ce85f
JK
2128 spin_lock(&vmap_area_lock);
2129 list_for_each_entry(va, &vmap_area_list, list) {
2130 if (!count)
2131 break;
2132
2133 if (!(va->flags & VM_VM_AREA))
2134 continue;
2135
2136 vm = va->vm;
2137 vaddr = (char *) vm->addr;
762216ab 2138 if (addr >= vaddr + get_vm_area_size(vm))
1da177e4
LT
2139 continue;
2140 while (addr < vaddr) {
2141 if (count == 0)
2142 goto finished;
2143 buf++;
2144 addr++;
2145 count--;
2146 }
762216ab 2147 n = vaddr + get_vm_area_size(vm) - addr;
d0107eb0
KH
2148 if (n > count)
2149 n = count;
e81ce85f 2150 if (!(vm->flags & VM_IOREMAP)) {
d0107eb0
KH
2151 aligned_vwrite(buf, addr, n);
2152 copied++;
2153 }
2154 buf += n;
2155 addr += n;
2156 count -= n;
1da177e4
LT
2157 }
2158finished:
e81ce85f 2159 spin_unlock(&vmap_area_lock);
d0107eb0
KH
2160 if (!copied)
2161 return 0;
2162 return buflen;
1da177e4 2163}
83342314
NP
2164
2165/**
e69e9d4a
HD
2166 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2167 * @vma: vma to cover
2168 * @uaddr: target user address to start at
2169 * @kaddr: virtual address of vmalloc kernel memory
2170 * @size: size of map area
7682486b
RD
2171 *
2172 * Returns: 0 for success, -Exxx on failure
83342314 2173 *
e69e9d4a
HD
2174 * This function checks that @kaddr is a valid vmalloc'ed area,
2175 * and that it is big enough to cover the range starting at
2176 * @uaddr in @vma. Will return failure if that criteria isn't
2177 * met.
83342314 2178 *
72fd4a35 2179 * Similar to remap_pfn_range() (see mm/memory.c)
83342314 2180 */
e69e9d4a
HD
2181int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2182 void *kaddr, unsigned long size)
83342314
NP
2183{
2184 struct vm_struct *area;
83342314 2185
e69e9d4a
HD
2186 size = PAGE_ALIGN(size);
2187
2188 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
83342314
NP
2189 return -EINVAL;
2190
e69e9d4a 2191 area = find_vm_area(kaddr);
83342314 2192 if (!area)
db64fe02 2193 return -EINVAL;
83342314
NP
2194
2195 if (!(area->flags & VM_USERMAP))
db64fe02 2196 return -EINVAL;
83342314 2197
e69e9d4a 2198 if (kaddr + size > area->addr + area->size)
db64fe02 2199 return -EINVAL;
83342314 2200
83342314 2201 do {
e69e9d4a 2202 struct page *page = vmalloc_to_page(kaddr);
db64fe02
NP
2203 int ret;
2204
83342314
NP
2205 ret = vm_insert_page(vma, uaddr, page);
2206 if (ret)
2207 return ret;
2208
2209 uaddr += PAGE_SIZE;
e69e9d4a
HD
2210 kaddr += PAGE_SIZE;
2211 size -= PAGE_SIZE;
2212 } while (size > 0);
83342314 2213
314e51b9 2214 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
83342314 2215
db64fe02 2216 return 0;
83342314 2217}
e69e9d4a
HD
2218EXPORT_SYMBOL(remap_vmalloc_range_partial);
2219
2220/**
2221 * remap_vmalloc_range - map vmalloc pages to userspace
2222 * @vma: vma to cover (map full range of vma)
2223 * @addr: vmalloc memory
2224 * @pgoff: number of pages into addr before first page to map
2225 *
2226 * Returns: 0 for success, -Exxx on failure
2227 *
2228 * This function checks that addr is a valid vmalloc'ed area, and
2229 * that it is big enough to cover the vma. Will return failure if
2230 * that criteria isn't met.
2231 *
2232 * Similar to remap_pfn_range() (see mm/memory.c)
2233 */
2234int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2235 unsigned long pgoff)
2236{
2237 return remap_vmalloc_range_partial(vma, vma->vm_start,
2238 addr + (pgoff << PAGE_SHIFT),
2239 vma->vm_end - vma->vm_start);
2240}
83342314
NP
2241EXPORT_SYMBOL(remap_vmalloc_range);
2242
1eeb66a1
CH
2243/*
2244 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2245 * have one.
2246 */
3b32123d 2247void __weak vmalloc_sync_all(void)
1eeb66a1
CH
2248{
2249}
5f4352fb
JF
2250
2251
2f569afd 2252static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
5f4352fb 2253{
cd12909c
DV
2254 pte_t ***p = data;
2255
2256 if (p) {
2257 *(*p) = pte;
2258 (*p)++;
2259 }
5f4352fb
JF
2260 return 0;
2261}
2262
2263/**
2264 * alloc_vm_area - allocate a range of kernel address space
2265 * @size: size of the area
cd12909c 2266 * @ptes: returns the PTEs for the address space
7682486b
RD
2267 *
2268 * Returns: NULL on failure, vm_struct on success
5f4352fb
JF
2269 *
2270 * This function reserves a range of kernel address space, and
2271 * allocates pagetables to map that range. No actual mappings
cd12909c
DV
2272 * are created.
2273 *
2274 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2275 * allocated for the VM area are returned.
5f4352fb 2276 */
cd12909c 2277struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
5f4352fb
JF
2278{
2279 struct vm_struct *area;
2280
23016969
CL
2281 area = get_vm_area_caller(size, VM_IOREMAP,
2282 __builtin_return_address(0));
5f4352fb
JF
2283 if (area == NULL)
2284 return NULL;
2285
2286 /*
2287 * This ensures that page tables are constructed for this region
2288 * of kernel virtual address space and mapped into init_mm.
2289 */
2290 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
cd12909c 2291 size, f, ptes ? &ptes : NULL)) {
5f4352fb
JF
2292 free_vm_area(area);
2293 return NULL;
2294 }
2295
5f4352fb
JF
2296 return area;
2297}
2298EXPORT_SYMBOL_GPL(alloc_vm_area);
2299
2300void free_vm_area(struct vm_struct *area)
2301{
2302 struct vm_struct *ret;
2303 ret = remove_vm_area(area->addr);
2304 BUG_ON(ret != area);
2305 kfree(area);
2306}
2307EXPORT_SYMBOL_GPL(free_vm_area);
a10aa579 2308
4f8b02b4 2309#ifdef CONFIG_SMP
ca23e405
TH
2310static struct vmap_area *node_to_va(struct rb_node *n)
2311{
2312 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2313}
2314
2315/**
2316 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2317 * @end: target address
2318 * @pnext: out arg for the next vmap_area
2319 * @pprev: out arg for the previous vmap_area
2320 *
2321 * Returns: %true if either or both of next and prev are found,
2322 * %false if no vmap_area exists
2323 *
2324 * Find vmap_areas end addresses of which enclose @end. ie. if not
2325 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2326 */
2327static bool pvm_find_next_prev(unsigned long end,
2328 struct vmap_area **pnext,
2329 struct vmap_area **pprev)
2330{
2331 struct rb_node *n = vmap_area_root.rb_node;
2332 struct vmap_area *va = NULL;
2333
2334 while (n) {
2335 va = rb_entry(n, struct vmap_area, rb_node);
2336 if (end < va->va_end)
2337 n = n->rb_left;
2338 else if (end > va->va_end)
2339 n = n->rb_right;
2340 else
2341 break;
2342 }
2343
2344 if (!va)
2345 return false;
2346
2347 if (va->va_end > end) {
2348 *pnext = va;
2349 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2350 } else {
2351 *pprev = va;
2352 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2353 }
2354 return true;
2355}
2356
2357/**
2358 * pvm_determine_end - find the highest aligned address between two vmap_areas
2359 * @pnext: in/out arg for the next vmap_area
2360 * @pprev: in/out arg for the previous vmap_area
2361 * @align: alignment
2362 *
2363 * Returns: determined end address
2364 *
2365 * Find the highest aligned address between *@pnext and *@pprev below
2366 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
2367 * down address is between the end addresses of the two vmap_areas.
2368 *
2369 * Please note that the address returned by this function may fall
2370 * inside *@pnext vmap_area. The caller is responsible for checking
2371 * that.
2372 */
2373static unsigned long pvm_determine_end(struct vmap_area **pnext,
2374 struct vmap_area **pprev,
2375 unsigned long align)
2376{
2377 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2378 unsigned long addr;
2379
2380 if (*pnext)
2381 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2382 else
2383 addr = vmalloc_end;
2384
2385 while (*pprev && (*pprev)->va_end > addr) {
2386 *pnext = *pprev;
2387 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2388 }
2389
2390 return addr;
2391}
2392
2393/**
2394 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2395 * @offsets: array containing offset of each area
2396 * @sizes: array containing size of each area
2397 * @nr_vms: the number of areas to allocate
2398 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
ca23e405
TH
2399 *
2400 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2401 * vm_structs on success, %NULL on failure
2402 *
2403 * Percpu allocator wants to use congruent vm areas so that it can
2404 * maintain the offsets among percpu areas. This function allocates
ec3f64fc
DR
2405 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
2406 * be scattered pretty far, distance between two areas easily going up
2407 * to gigabytes. To avoid interacting with regular vmallocs, these
2408 * areas are allocated from top.
ca23e405
TH
2409 *
2410 * Despite its complicated look, this allocator is rather simple. It
2411 * does everything top-down and scans areas from the end looking for
2412 * matching slot. While scanning, if any of the areas overlaps with
2413 * existing vmap_area, the base address is pulled down to fit the
2414 * area. Scanning is repeated till all the areas fit and then all
2415 * necessary data structres are inserted and the result is returned.
2416 */
2417struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2418 const size_t *sizes, int nr_vms,
ec3f64fc 2419 size_t align)
ca23e405
TH
2420{
2421 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2422 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2423 struct vmap_area **vas, *prev, *next;
2424 struct vm_struct **vms;
2425 int area, area2, last_area, term_area;
2426 unsigned long base, start, end, last_end;
2427 bool purged = false;
2428
ca23e405 2429 /* verify parameters and allocate data structures */
891c49ab 2430 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
ca23e405
TH
2431 for (last_area = 0, area = 0; area < nr_vms; area++) {
2432 start = offsets[area];
2433 end = start + sizes[area];
2434
2435 /* is everything aligned properly? */
2436 BUG_ON(!IS_ALIGNED(offsets[area], align));
2437 BUG_ON(!IS_ALIGNED(sizes[area], align));
2438
2439 /* detect the area with the highest address */
2440 if (start > offsets[last_area])
2441 last_area = area;
2442
2443 for (area2 = 0; area2 < nr_vms; area2++) {
2444 unsigned long start2 = offsets[area2];
2445 unsigned long end2 = start2 + sizes[area2];
2446
2447 if (area2 == area)
2448 continue;
2449
2450 BUG_ON(start2 >= start && start2 < end);
2451 BUG_ON(end2 <= end && end2 > start);
2452 }
2453 }
2454 last_end = offsets[last_area] + sizes[last_area];
2455
2456 if (vmalloc_end - vmalloc_start < last_end) {
2457 WARN_ON(true);
2458 return NULL;
2459 }
2460
4d67d860
TM
2461 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2462 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
ca23e405 2463 if (!vas || !vms)
f1db7afd 2464 goto err_free2;
ca23e405
TH
2465
2466 for (area = 0; area < nr_vms; area++) {
ec3f64fc
DR
2467 vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2468 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
ca23e405
TH
2469 if (!vas[area] || !vms[area])
2470 goto err_free;
2471 }
2472retry:
2473 spin_lock(&vmap_area_lock);
2474
2475 /* start scanning - we scan from the top, begin with the last area */
2476 area = term_area = last_area;
2477 start = offsets[area];
2478 end = start + sizes[area];
2479
2480 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2481 base = vmalloc_end - last_end;
2482 goto found;
2483 }
2484 base = pvm_determine_end(&next, &prev, align) - end;
2485
2486 while (true) {
2487 BUG_ON(next && next->va_end <= base + end);
2488 BUG_ON(prev && prev->va_end > base + end);
2489
2490 /*
2491 * base might have underflowed, add last_end before
2492 * comparing.
2493 */
2494 if (base + last_end < vmalloc_start + last_end) {
2495 spin_unlock(&vmap_area_lock);
2496 if (!purged) {
2497 purge_vmap_area_lazy();
2498 purged = true;
2499 goto retry;
2500 }
2501 goto err_free;
2502 }
2503
2504 /*
2505 * If next overlaps, move base downwards so that it's
2506 * right below next and then recheck.
2507 */
2508 if (next && next->va_start < base + end) {
2509 base = pvm_determine_end(&next, &prev, align) - end;
2510 term_area = area;
2511 continue;
2512 }
2513
2514 /*
2515 * If prev overlaps, shift down next and prev and move
2516 * base so that it's right below new next and then
2517 * recheck.
2518 */
2519 if (prev && prev->va_end > base + start) {
2520 next = prev;
2521 prev = node_to_va(rb_prev(&next->rb_node));
2522 base = pvm_determine_end(&next, &prev, align) - end;
2523 term_area = area;
2524 continue;
2525 }
2526
2527 /*
2528 * This area fits, move on to the previous one. If
2529 * the previous one is the terminal one, we're done.
2530 */
2531 area = (area + nr_vms - 1) % nr_vms;
2532 if (area == term_area)
2533 break;
2534 start = offsets[area];
2535 end = start + sizes[area];
2536 pvm_find_next_prev(base + end, &next, &prev);
2537 }
2538found:
2539 /* we've found a fitting base, insert all va's */
2540 for (area = 0; area < nr_vms; area++) {
2541 struct vmap_area *va = vas[area];
2542
2543 va->va_start = base + offsets[area];
2544 va->va_end = va->va_start + sizes[area];
2545 __insert_vmap_area(va);
2546 }
2547
2548 vmap_area_pcpu_hole = base + offsets[last_area];
2549
2550 spin_unlock(&vmap_area_lock);
2551
2552 /* insert all vm's */
2553 for (area = 0; area < nr_vms; area++)
3645cb4a
ZY
2554 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2555 pcpu_get_vm_areas);
ca23e405
TH
2556
2557 kfree(vas);
2558 return vms;
2559
2560err_free:
2561 for (area = 0; area < nr_vms; area++) {
f1db7afd
KC
2562 kfree(vas[area]);
2563 kfree(vms[area]);
ca23e405 2564 }
f1db7afd 2565err_free2:
ca23e405
TH
2566 kfree(vas);
2567 kfree(vms);
2568 return NULL;
2569}
2570
2571/**
2572 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2573 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2574 * @nr_vms: the number of allocated areas
2575 *
2576 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2577 */
2578void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2579{
2580 int i;
2581
2582 for (i = 0; i < nr_vms; i++)
2583 free_vm_area(vms[i]);
2584 kfree(vms);
2585}
4f8b02b4 2586#endif /* CONFIG_SMP */
a10aa579
CL
2587
2588#ifdef CONFIG_PROC_FS
2589static void *s_start(struct seq_file *m, loff_t *pos)
d4033afd 2590 __acquires(&vmap_area_lock)
a10aa579 2591{
d4033afd 2592 spin_lock(&vmap_area_lock);
3f500069 2593 return seq_list_start(&vmap_area_list, *pos);
a10aa579
CL
2594}
2595
2596static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2597{
3f500069 2598 return seq_list_next(p, &vmap_area_list, pos);
a10aa579
CL
2599}
2600
2601static void s_stop(struct seq_file *m, void *p)
d4033afd 2602 __releases(&vmap_area_lock)
a10aa579 2603{
d4033afd 2604 spin_unlock(&vmap_area_lock);
a10aa579
CL
2605}
2606
a47a126a
ED
2607static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2608{
e5adfffc 2609 if (IS_ENABLED(CONFIG_NUMA)) {
a47a126a
ED
2610 unsigned int nr, *counters = m->private;
2611
2612 if (!counters)
2613 return;
2614
af12346c
WL
2615 if (v->flags & VM_UNINITIALIZED)
2616 return;
7e5b528b
DV
2617 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2618 smp_rmb();
af12346c 2619
a47a126a
ED
2620 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2621
2622 for (nr = 0; nr < v->nr_pages; nr++)
2623 counters[page_to_nid(v->pages[nr])]++;
2624
2625 for_each_node_state(nr, N_HIGH_MEMORY)
2626 if (counters[nr])
2627 seq_printf(m, " N%u=%u", nr, counters[nr]);
2628 }
2629}
2630
a10aa579
CL
2631static int s_show(struct seq_file *m, void *p)
2632{
3f500069 2633 struct vmap_area *va;
d4033afd
JK
2634 struct vm_struct *v;
2635
3f500069 2636 va = list_entry(p, struct vmap_area, list);
2637
c2ce8c14
WL
2638 /*
2639 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2640 * behalf of vmap area is being tear down or vm_map_ram allocation.
2641 */
2642 if (!(va->flags & VM_VM_AREA))
d4033afd 2643 return 0;
d4033afd
JK
2644
2645 v = va->vm;
a10aa579 2646
45ec1690 2647 seq_printf(m, "0x%pK-0x%pK %7ld",
a10aa579
CL
2648 v->addr, v->addr + v->size, v->size);
2649
62c70bce
JP
2650 if (v->caller)
2651 seq_printf(m, " %pS", v->caller);
23016969 2652
a10aa579
CL
2653 if (v->nr_pages)
2654 seq_printf(m, " pages=%d", v->nr_pages);
2655
2656 if (v->phys_addr)
ffa71f33 2657 seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
a10aa579
CL
2658
2659 if (v->flags & VM_IOREMAP)
f4527c90 2660 seq_puts(m, " ioremap");
a10aa579
CL
2661
2662 if (v->flags & VM_ALLOC)
f4527c90 2663 seq_puts(m, " vmalloc");
a10aa579
CL
2664
2665 if (v->flags & VM_MAP)
f4527c90 2666 seq_puts(m, " vmap");
a10aa579
CL
2667
2668 if (v->flags & VM_USERMAP)
f4527c90 2669 seq_puts(m, " user");
a10aa579 2670
244d63ee 2671 if (is_vmalloc_addr(v->pages))
f4527c90 2672 seq_puts(m, " vpages");
a10aa579 2673
a47a126a 2674 show_numa_info(m, v);
a10aa579
CL
2675 seq_putc(m, '\n');
2676 return 0;
2677}
2678
5f6a6a9c 2679static const struct seq_operations vmalloc_op = {
a10aa579
CL
2680 .start = s_start,
2681 .next = s_next,
2682 .stop = s_stop,
2683 .show = s_show,
2684};
5f6a6a9c
AD
2685
2686static int vmalloc_open(struct inode *inode, struct file *file)
2687{
703394c1
RJ
2688 if (IS_ENABLED(CONFIG_NUMA))
2689 return seq_open_private(file, &vmalloc_op,
2690 nr_node_ids * sizeof(unsigned int));
2691 else
2692 return seq_open(file, &vmalloc_op);
5f6a6a9c
AD
2693}
2694
2695static const struct file_operations proc_vmalloc_operations = {
2696 .open = vmalloc_open,
2697 .read = seq_read,
2698 .llseek = seq_lseek,
2699 .release = seq_release_private,
2700};
2701
2702static int __init proc_vmalloc_init(void)
2703{
2704 proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2705 return 0;
2706}
2707module_init(proc_vmalloc_init);
db3808c1 2708
a10aa579
CL
2709#endif
2710