Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched/signal.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/set_memory.h>
22#include <linux/debugobjects.h>
23#include <linux/kallsyms.h>
24#include <linux/list.h>
25#include <linux/notifier.h>
26#include <linux/rbtree.h>
27#include <linux/xarray.h>
28#include <linux/io.h>
29#include <linux/rcupdate.h>
30#include <linux/pfn.h>
31#include <linux/kmemleak.h>
32#include <linux/atomic.h>
33#include <linux/compiler.h>
34#include <linux/memcontrol.h>
35#include <linux/llist.h>
36#include <linux/bitops.h>
37#include <linux/rbtree_augmented.h>
38#include <linux/overflow.h>
39#include <linux/pgtable.h>
40#include <linux/uaccess.h>
41#include <linux/hugetlb.h>
42#include <linux/sched/mm.h>
43#include <asm/tlbflush.h>
44#include <asm/shmparam.h>
45
46#define CREATE_TRACE_POINTS
47#include <trace/events/vmalloc.h>
48
49#include "internal.h"
50#include "pgalloc-track.h"
51
52#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
54
55static int __init set_nohugeiomap(char *str)
56{
57 ioremap_max_page_shift = PAGE_SHIFT;
58 return 0;
59}
60early_param("nohugeiomap", set_nohugeiomap);
61#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
63#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
64
65#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66static bool __ro_after_init vmap_allow_huge = true;
67
68static int __init set_nohugevmalloc(char *str)
69{
70 vmap_allow_huge = false;
71 return 0;
72}
73early_param("nohugevmalloc", set_nohugevmalloc);
74#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75static const bool vmap_allow_huge = false;
76#endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
77
78bool is_vmalloc_addr(const void *x)
79{
80 unsigned long addr = (unsigned long)kasan_reset_tag(x);
81
82 return addr >= VMALLOC_START && addr < VMALLOC_END;
83}
84EXPORT_SYMBOL(is_vmalloc_addr);
85
86struct vfree_deferred {
87 struct llist_head list;
88 struct work_struct wq;
89};
90static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
91
92static void __vunmap(const void *, int);
93
94static void free_work(struct work_struct *w)
95{
96 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
97 struct llist_node *t, *llnode;
98
99 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
100 __vunmap((void *)llnode, 1);
101}
102
103/*** Page table manipulation functions ***/
104static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
105 phys_addr_t phys_addr, pgprot_t prot,
106 unsigned int max_page_shift, pgtbl_mod_mask *mask)
107{
108 pte_t *pte;
109 u64 pfn;
110 unsigned long size = PAGE_SIZE;
111
112 pfn = phys_addr >> PAGE_SHIFT;
113 pte = pte_alloc_kernel_track(pmd, addr, mask);
114 if (!pte)
115 return -ENOMEM;
116 do {
117 BUG_ON(!pte_none(*pte));
118
119#ifdef CONFIG_HUGETLB_PAGE
120 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
121 if (size != PAGE_SIZE) {
122 pte_t entry = pfn_pte(pfn, prot);
123
124 entry = arch_make_huge_pte(entry, ilog2(size), 0);
125 set_huge_pte_at(&init_mm, addr, pte, entry);
126 pfn += PFN_DOWN(size);
127 continue;
128 }
129#endif
130 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
131 pfn++;
132 } while (pte += PFN_DOWN(size), addr += size, addr != end);
133 *mask |= PGTBL_PTE_MODIFIED;
134 return 0;
135}
136
137static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
138 phys_addr_t phys_addr, pgprot_t prot,
139 unsigned int max_page_shift)
140{
141 if (max_page_shift < PMD_SHIFT)
142 return 0;
143
144 if (!arch_vmap_pmd_supported(prot))
145 return 0;
146
147 if ((end - addr) != PMD_SIZE)
148 return 0;
149
150 if (!IS_ALIGNED(addr, PMD_SIZE))
151 return 0;
152
153 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
154 return 0;
155
156 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
157 return 0;
158
159 return pmd_set_huge(pmd, phys_addr, prot);
160}
161
162static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
163 phys_addr_t phys_addr, pgprot_t prot,
164 unsigned int max_page_shift, pgtbl_mod_mask *mask)
165{
166 pmd_t *pmd;
167 unsigned long next;
168
169 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
170 if (!pmd)
171 return -ENOMEM;
172 do {
173 next = pmd_addr_end(addr, end);
174
175 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
176 max_page_shift)) {
177 *mask |= PGTBL_PMD_MODIFIED;
178 continue;
179 }
180
181 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
182 return -ENOMEM;
183 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
184 return 0;
185}
186
187static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
188 phys_addr_t phys_addr, pgprot_t prot,
189 unsigned int max_page_shift)
190{
191 if (max_page_shift < PUD_SHIFT)
192 return 0;
193
194 if (!arch_vmap_pud_supported(prot))
195 return 0;
196
197 if ((end - addr) != PUD_SIZE)
198 return 0;
199
200 if (!IS_ALIGNED(addr, PUD_SIZE))
201 return 0;
202
203 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
204 return 0;
205
206 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
207 return 0;
208
209 return pud_set_huge(pud, phys_addr, prot);
210}
211
212static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
213 phys_addr_t phys_addr, pgprot_t prot,
214 unsigned int max_page_shift, pgtbl_mod_mask *mask)
215{
216 pud_t *pud;
217 unsigned long next;
218
219 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
220 if (!pud)
221 return -ENOMEM;
222 do {
223 next = pud_addr_end(addr, end);
224
225 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
226 max_page_shift)) {
227 *mask |= PGTBL_PUD_MODIFIED;
228 continue;
229 }
230
231 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
232 max_page_shift, mask))
233 return -ENOMEM;
234 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
235 return 0;
236}
237
238static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
239 phys_addr_t phys_addr, pgprot_t prot,
240 unsigned int max_page_shift)
241{
242 if (max_page_shift < P4D_SHIFT)
243 return 0;
244
245 if (!arch_vmap_p4d_supported(prot))
246 return 0;
247
248 if ((end - addr) != P4D_SIZE)
249 return 0;
250
251 if (!IS_ALIGNED(addr, P4D_SIZE))
252 return 0;
253
254 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
255 return 0;
256
257 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
258 return 0;
259
260 return p4d_set_huge(p4d, phys_addr, prot);
261}
262
263static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
264 phys_addr_t phys_addr, pgprot_t prot,
265 unsigned int max_page_shift, pgtbl_mod_mask *mask)
266{
267 p4d_t *p4d;
268 unsigned long next;
269
270 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
271 if (!p4d)
272 return -ENOMEM;
273 do {
274 next = p4d_addr_end(addr, end);
275
276 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
277 max_page_shift)) {
278 *mask |= PGTBL_P4D_MODIFIED;
279 continue;
280 }
281
282 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
283 max_page_shift, mask))
284 return -ENOMEM;
285 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
286 return 0;
287}
288
289static int vmap_range_noflush(unsigned long addr, unsigned long end,
290 phys_addr_t phys_addr, pgprot_t prot,
291 unsigned int max_page_shift)
292{
293 pgd_t *pgd;
294 unsigned long start;
295 unsigned long next;
296 int err;
297 pgtbl_mod_mask mask = 0;
298
299 might_sleep();
300 BUG_ON(addr >= end);
301
302 start = addr;
303 pgd = pgd_offset_k(addr);
304 do {
305 next = pgd_addr_end(addr, end);
306 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
307 max_page_shift, &mask);
308 if (err)
309 break;
310 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
311
312 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
313 arch_sync_kernel_mappings(start, end);
314
315 return err;
316}
317
318int ioremap_page_range(unsigned long addr, unsigned long end,
319 phys_addr_t phys_addr, pgprot_t prot)
320{
321 int err;
322
323 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
324 ioremap_max_page_shift);
325 flush_cache_vmap(addr, end);
326 if (!err)
327 kmsan_ioremap_page_range(addr, end, phys_addr, prot,
328 ioremap_max_page_shift);
329 return err;
330}
331
332static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
333 pgtbl_mod_mask *mask)
334{
335 pte_t *pte;
336
337 pte = pte_offset_kernel(pmd, addr);
338 do {
339 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
340 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
341 } while (pte++, addr += PAGE_SIZE, addr != end);
342 *mask |= PGTBL_PTE_MODIFIED;
343}
344
345static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
346 pgtbl_mod_mask *mask)
347{
348 pmd_t *pmd;
349 unsigned long next;
350 int cleared;
351
352 pmd = pmd_offset(pud, addr);
353 do {
354 next = pmd_addr_end(addr, end);
355
356 cleared = pmd_clear_huge(pmd);
357 if (cleared || pmd_bad(*pmd))
358 *mask |= PGTBL_PMD_MODIFIED;
359
360 if (cleared)
361 continue;
362 if (pmd_none_or_clear_bad(pmd))
363 continue;
364 vunmap_pte_range(pmd, addr, next, mask);
365
366 cond_resched();
367 } while (pmd++, addr = next, addr != end);
368}
369
370static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
371 pgtbl_mod_mask *mask)
372{
373 pud_t *pud;
374 unsigned long next;
375 int cleared;
376
377 pud = pud_offset(p4d, addr);
378 do {
379 next = pud_addr_end(addr, end);
380
381 cleared = pud_clear_huge(pud);
382 if (cleared || pud_bad(*pud))
383 *mask |= PGTBL_PUD_MODIFIED;
384
385 if (cleared)
386 continue;
387 if (pud_none_or_clear_bad(pud))
388 continue;
389 vunmap_pmd_range(pud, addr, next, mask);
390 } while (pud++, addr = next, addr != end);
391}
392
393static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
394 pgtbl_mod_mask *mask)
395{
396 p4d_t *p4d;
397 unsigned long next;
398
399 p4d = p4d_offset(pgd, addr);
400 do {
401 next = p4d_addr_end(addr, end);
402
403 p4d_clear_huge(p4d);
404 if (p4d_bad(*p4d))
405 *mask |= PGTBL_P4D_MODIFIED;
406
407 if (p4d_none_or_clear_bad(p4d))
408 continue;
409 vunmap_pud_range(p4d, addr, next, mask);
410 } while (p4d++, addr = next, addr != end);
411}
412
413/*
414 * vunmap_range_noflush is similar to vunmap_range, but does not
415 * flush caches or TLBs.
416 *
417 * The caller is responsible for calling flush_cache_vmap() before calling
418 * this function, and flush_tlb_kernel_range after it has returned
419 * successfully (and before the addresses are expected to cause a page fault
420 * or be re-mapped for something else, if TLB flushes are being delayed or
421 * coalesced).
422 *
423 * This is an internal function only. Do not use outside mm/.
424 */
425void __vunmap_range_noflush(unsigned long start, unsigned long end)
426{
427 unsigned long next;
428 pgd_t *pgd;
429 unsigned long addr = start;
430 pgtbl_mod_mask mask = 0;
431
432 BUG_ON(addr >= end);
433 pgd = pgd_offset_k(addr);
434 do {
435 next = pgd_addr_end(addr, end);
436 if (pgd_bad(*pgd))
437 mask |= PGTBL_PGD_MODIFIED;
438 if (pgd_none_or_clear_bad(pgd))
439 continue;
440 vunmap_p4d_range(pgd, addr, next, &mask);
441 } while (pgd++, addr = next, addr != end);
442
443 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
444 arch_sync_kernel_mappings(start, end);
445}
446
447void vunmap_range_noflush(unsigned long start, unsigned long end)
448{
449 kmsan_vunmap_range_noflush(start, end);
450 __vunmap_range_noflush(start, end);
451}
452
453/**
454 * vunmap_range - unmap kernel virtual addresses
455 * @addr: start of the VM area to unmap
456 * @end: end of the VM area to unmap (non-inclusive)
457 *
458 * Clears any present PTEs in the virtual address range, flushes TLBs and
459 * caches. Any subsequent access to the address before it has been re-mapped
460 * is a kernel bug.
461 */
462void vunmap_range(unsigned long addr, unsigned long end)
463{
464 flush_cache_vunmap(addr, end);
465 vunmap_range_noflush(addr, end);
466 flush_tlb_kernel_range(addr, end);
467}
468
469static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
470 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
471 pgtbl_mod_mask *mask)
472{
473 pte_t *pte;
474
475 /*
476 * nr is a running index into the array which helps higher level
477 * callers keep track of where we're up to.
478 */
479
480 pte = pte_alloc_kernel_track(pmd, addr, mask);
481 if (!pte)
482 return -ENOMEM;
483 do {
484 struct page *page = pages[*nr];
485
486 if (WARN_ON(!pte_none(*pte)))
487 return -EBUSY;
488 if (WARN_ON(!page))
489 return -ENOMEM;
490 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
491 return -EINVAL;
492
493 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
494 (*nr)++;
495 } while (pte++, addr += PAGE_SIZE, addr != end);
496 *mask |= PGTBL_PTE_MODIFIED;
497 return 0;
498}
499
500static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
501 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
502 pgtbl_mod_mask *mask)
503{
504 pmd_t *pmd;
505 unsigned long next;
506
507 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
508 if (!pmd)
509 return -ENOMEM;
510 do {
511 next = pmd_addr_end(addr, end);
512 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
513 return -ENOMEM;
514 } while (pmd++, addr = next, addr != end);
515 return 0;
516}
517
518static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
519 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
520 pgtbl_mod_mask *mask)
521{
522 pud_t *pud;
523 unsigned long next;
524
525 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
526 if (!pud)
527 return -ENOMEM;
528 do {
529 next = pud_addr_end(addr, end);
530 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
531 return -ENOMEM;
532 } while (pud++, addr = next, addr != end);
533 return 0;
534}
535
536static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
537 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
538 pgtbl_mod_mask *mask)
539{
540 p4d_t *p4d;
541 unsigned long next;
542
543 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
544 if (!p4d)
545 return -ENOMEM;
546 do {
547 next = p4d_addr_end(addr, end);
548 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
549 return -ENOMEM;
550 } while (p4d++, addr = next, addr != end);
551 return 0;
552}
553
554static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
555 pgprot_t prot, struct page **pages)
556{
557 unsigned long start = addr;
558 pgd_t *pgd;
559 unsigned long next;
560 int err = 0;
561 int nr = 0;
562 pgtbl_mod_mask mask = 0;
563
564 BUG_ON(addr >= end);
565 pgd = pgd_offset_k(addr);
566 do {
567 next = pgd_addr_end(addr, end);
568 if (pgd_bad(*pgd))
569 mask |= PGTBL_PGD_MODIFIED;
570 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
571 if (err)
572 return err;
573 } while (pgd++, addr = next, addr != end);
574
575 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
576 arch_sync_kernel_mappings(start, end);
577
578 return 0;
579}
580
581/*
582 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
583 * flush caches.
584 *
585 * The caller is responsible for calling flush_cache_vmap() after this
586 * function returns successfully and before the addresses are accessed.
587 *
588 * This is an internal function only. Do not use outside mm/.
589 */
590int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
591 pgprot_t prot, struct page **pages, unsigned int page_shift)
592{
593 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
594
595 WARN_ON(page_shift < PAGE_SHIFT);
596
597 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
598 page_shift == PAGE_SHIFT)
599 return vmap_small_pages_range_noflush(addr, end, prot, pages);
600
601 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
602 int err;
603
604 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
605 page_to_phys(pages[i]), prot,
606 page_shift);
607 if (err)
608 return err;
609
610 addr += 1UL << page_shift;
611 }
612
613 return 0;
614}
615
616int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
617 pgprot_t prot, struct page **pages, unsigned int page_shift)
618{
619 kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
620 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
621}
622
623/**
624 * vmap_pages_range - map pages to a kernel virtual address
625 * @addr: start of the VM area to map
626 * @end: end of the VM area to map (non-inclusive)
627 * @prot: page protection flags to use
628 * @pages: pages to map (always PAGE_SIZE pages)
629 * @page_shift: maximum shift that the pages may be mapped with, @pages must
630 * be aligned and contiguous up to at least this shift.
631 *
632 * RETURNS:
633 * 0 on success, -errno on failure.
634 */
635static int vmap_pages_range(unsigned long addr, unsigned long end,
636 pgprot_t prot, struct page **pages, unsigned int page_shift)
637{
638 int err;
639
640 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
641 flush_cache_vmap(addr, end);
642 return err;
643}
644
645int is_vmalloc_or_module_addr(const void *x)
646{
647 /*
648 * ARM, x86-64 and sparc64 put modules in a special place,
649 * and fall back on vmalloc() if that fails. Others
650 * just put it in the vmalloc space.
651 */
652#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
653 unsigned long addr = (unsigned long)kasan_reset_tag(x);
654 if (addr >= MODULES_VADDR && addr < MODULES_END)
655 return 1;
656#endif
657 return is_vmalloc_addr(x);
658}
659
660/*
661 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
662 * return the tail page that corresponds to the base page address, which
663 * matches small vmap mappings.
664 */
665struct page *vmalloc_to_page(const void *vmalloc_addr)
666{
667 unsigned long addr = (unsigned long) vmalloc_addr;
668 struct page *page = NULL;
669 pgd_t *pgd = pgd_offset_k(addr);
670 p4d_t *p4d;
671 pud_t *pud;
672 pmd_t *pmd;
673 pte_t *ptep, pte;
674
675 /*
676 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
677 * architectures that do not vmalloc module space
678 */
679 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
680
681 if (pgd_none(*pgd))
682 return NULL;
683 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
684 return NULL; /* XXX: no allowance for huge pgd */
685 if (WARN_ON_ONCE(pgd_bad(*pgd)))
686 return NULL;
687
688 p4d = p4d_offset(pgd, addr);
689 if (p4d_none(*p4d))
690 return NULL;
691 if (p4d_leaf(*p4d))
692 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
693 if (WARN_ON_ONCE(p4d_bad(*p4d)))
694 return NULL;
695
696 pud = pud_offset(p4d, addr);
697 if (pud_none(*pud))
698 return NULL;
699 if (pud_leaf(*pud))
700 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
701 if (WARN_ON_ONCE(pud_bad(*pud)))
702 return NULL;
703
704 pmd = pmd_offset(pud, addr);
705 if (pmd_none(*pmd))
706 return NULL;
707 if (pmd_leaf(*pmd))
708 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
709 if (WARN_ON_ONCE(pmd_bad(*pmd)))
710 return NULL;
711
712 ptep = pte_offset_map(pmd, addr);
713 pte = *ptep;
714 if (pte_present(pte))
715 page = pte_page(pte);
716 pte_unmap(ptep);
717
718 return page;
719}
720EXPORT_SYMBOL(vmalloc_to_page);
721
722/*
723 * Map a vmalloc()-space virtual address to the physical page frame number.
724 */
725unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
726{
727 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
728}
729EXPORT_SYMBOL(vmalloc_to_pfn);
730
731
732/*** Global kva allocator ***/
733
734#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
735#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
736
737
738static DEFINE_SPINLOCK(vmap_area_lock);
739static DEFINE_SPINLOCK(free_vmap_area_lock);
740/* Export for kexec only */
741LIST_HEAD(vmap_area_list);
742static struct rb_root vmap_area_root = RB_ROOT;
743static bool vmap_initialized __read_mostly;
744
745static struct rb_root purge_vmap_area_root = RB_ROOT;
746static LIST_HEAD(purge_vmap_area_list);
747static DEFINE_SPINLOCK(purge_vmap_area_lock);
748
749/*
750 * This kmem_cache is used for vmap_area objects. Instead of
751 * allocating from slab we reuse an object from this cache to
752 * make things faster. Especially in "no edge" splitting of
753 * free block.
754 */
755static struct kmem_cache *vmap_area_cachep;
756
757/*
758 * This linked list is used in pair with free_vmap_area_root.
759 * It gives O(1) access to prev/next to perform fast coalescing.
760 */
761static LIST_HEAD(free_vmap_area_list);
762
763/*
764 * This augment red-black tree represents the free vmap space.
765 * All vmap_area objects in this tree are sorted by va->va_start
766 * address. It is used for allocation and merging when a vmap
767 * object is released.
768 *
769 * Each vmap_area node contains a maximum available free block
770 * of its sub-tree, right or left. Therefore it is possible to
771 * find a lowest match of free area.
772 */
773static struct rb_root free_vmap_area_root = RB_ROOT;
774
775/*
776 * Preload a CPU with one object for "no edge" split case. The
777 * aim is to get rid of allocations from the atomic context, thus
778 * to use more permissive allocation masks.
779 */
780static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
781
782static __always_inline unsigned long
783va_size(struct vmap_area *va)
784{
785 return (va->va_end - va->va_start);
786}
787
788static __always_inline unsigned long
789get_subtree_max_size(struct rb_node *node)
790{
791 struct vmap_area *va;
792
793 va = rb_entry_safe(node, struct vmap_area, rb_node);
794 return va ? va->subtree_max_size : 0;
795}
796
797RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
798 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
799
800static void purge_vmap_area_lazy(void);
801static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
802static void drain_vmap_area_work(struct work_struct *work);
803static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
804
805static atomic_long_t nr_vmalloc_pages;
806
807unsigned long vmalloc_nr_pages(void)
808{
809 return atomic_long_read(&nr_vmalloc_pages);
810}
811
812/* Look up the first VA which satisfies addr < va_end, NULL if none. */
813static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
814{
815 struct vmap_area *va = NULL;
816 struct rb_node *n = vmap_area_root.rb_node;
817
818 addr = (unsigned long)kasan_reset_tag((void *)addr);
819
820 while (n) {
821 struct vmap_area *tmp;
822
823 tmp = rb_entry(n, struct vmap_area, rb_node);
824 if (tmp->va_end > addr) {
825 va = tmp;
826 if (tmp->va_start <= addr)
827 break;
828
829 n = n->rb_left;
830 } else
831 n = n->rb_right;
832 }
833
834 return va;
835}
836
837static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
838{
839 struct rb_node *n = root->rb_node;
840
841 addr = (unsigned long)kasan_reset_tag((void *)addr);
842
843 while (n) {
844 struct vmap_area *va;
845
846 va = rb_entry(n, struct vmap_area, rb_node);
847 if (addr < va->va_start)
848 n = n->rb_left;
849 else if (addr >= va->va_end)
850 n = n->rb_right;
851 else
852 return va;
853 }
854
855 return NULL;
856}
857
858/*
859 * This function returns back addresses of parent node
860 * and its left or right link for further processing.
861 *
862 * Otherwise NULL is returned. In that case all further
863 * steps regarding inserting of conflicting overlap range
864 * have to be declined and actually considered as a bug.
865 */
866static __always_inline struct rb_node **
867find_va_links(struct vmap_area *va,
868 struct rb_root *root, struct rb_node *from,
869 struct rb_node **parent)
870{
871 struct vmap_area *tmp_va;
872 struct rb_node **link;
873
874 if (root) {
875 link = &root->rb_node;
876 if (unlikely(!*link)) {
877 *parent = NULL;
878 return link;
879 }
880 } else {
881 link = &from;
882 }
883
884 /*
885 * Go to the bottom of the tree. When we hit the last point
886 * we end up with parent rb_node and correct direction, i name
887 * it link, where the new va->rb_node will be attached to.
888 */
889 do {
890 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
891
892 /*
893 * During the traversal we also do some sanity check.
894 * Trigger the BUG() if there are sides(left/right)
895 * or full overlaps.
896 */
897 if (va->va_end <= tmp_va->va_start)
898 link = &(*link)->rb_left;
899 else if (va->va_start >= tmp_va->va_end)
900 link = &(*link)->rb_right;
901 else {
902 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
903 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
904
905 return NULL;
906 }
907 } while (*link);
908
909 *parent = &tmp_va->rb_node;
910 return link;
911}
912
913static __always_inline struct list_head *
914get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
915{
916 struct list_head *list;
917
918 if (unlikely(!parent))
919 /*
920 * The red-black tree where we try to find VA neighbors
921 * before merging or inserting is empty, i.e. it means
922 * there is no free vmap space. Normally it does not
923 * happen but we handle this case anyway.
924 */
925 return NULL;
926
927 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
928 return (&parent->rb_right == link ? list->next : list);
929}
930
931static __always_inline void
932__link_va(struct vmap_area *va, struct rb_root *root,
933 struct rb_node *parent, struct rb_node **link,
934 struct list_head *head, bool augment)
935{
936 /*
937 * VA is still not in the list, but we can
938 * identify its future previous list_head node.
939 */
940 if (likely(parent)) {
941 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
942 if (&parent->rb_right != link)
943 head = head->prev;
944 }
945
946 /* Insert to the rb-tree */
947 rb_link_node(&va->rb_node, parent, link);
948 if (augment) {
949 /*
950 * Some explanation here. Just perform simple insertion
951 * to the tree. We do not set va->subtree_max_size to
952 * its current size before calling rb_insert_augmented().
953 * It is because we populate the tree from the bottom
954 * to parent levels when the node _is_ in the tree.
955 *
956 * Therefore we set subtree_max_size to zero after insertion,
957 * to let __augment_tree_propagate_from() puts everything to
958 * the correct order later on.
959 */
960 rb_insert_augmented(&va->rb_node,
961 root, &free_vmap_area_rb_augment_cb);
962 va->subtree_max_size = 0;
963 } else {
964 rb_insert_color(&va->rb_node, root);
965 }
966
967 /* Address-sort this list */
968 list_add(&va->list, head);
969}
970
971static __always_inline void
972link_va(struct vmap_area *va, struct rb_root *root,
973 struct rb_node *parent, struct rb_node **link,
974 struct list_head *head)
975{
976 __link_va(va, root, parent, link, head, false);
977}
978
979static __always_inline void
980link_va_augment(struct vmap_area *va, struct rb_root *root,
981 struct rb_node *parent, struct rb_node **link,
982 struct list_head *head)
983{
984 __link_va(va, root, parent, link, head, true);
985}
986
987static __always_inline void
988__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
989{
990 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
991 return;
992
993 if (augment)
994 rb_erase_augmented(&va->rb_node,
995 root, &free_vmap_area_rb_augment_cb);
996 else
997 rb_erase(&va->rb_node, root);
998
999 list_del_init(&va->list);
1000 RB_CLEAR_NODE(&va->rb_node);
1001}
1002
1003static __always_inline void
1004unlink_va(struct vmap_area *va, struct rb_root *root)
1005{
1006 __unlink_va(va, root, false);
1007}
1008
1009static __always_inline void
1010unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1011{
1012 __unlink_va(va, root, true);
1013}
1014
1015#if DEBUG_AUGMENT_PROPAGATE_CHECK
1016/*
1017 * Gets called when remove the node and rotate.
1018 */
1019static __always_inline unsigned long
1020compute_subtree_max_size(struct vmap_area *va)
1021{
1022 return max3(va_size(va),
1023 get_subtree_max_size(va->rb_node.rb_left),
1024 get_subtree_max_size(va->rb_node.rb_right));
1025}
1026
1027static void
1028augment_tree_propagate_check(void)
1029{
1030 struct vmap_area *va;
1031 unsigned long computed_size;
1032
1033 list_for_each_entry(va, &free_vmap_area_list, list) {
1034 computed_size = compute_subtree_max_size(va);
1035 if (computed_size != va->subtree_max_size)
1036 pr_emerg("tree is corrupted: %lu, %lu\n",
1037 va_size(va), va->subtree_max_size);
1038 }
1039}
1040#endif
1041
1042/*
1043 * This function populates subtree_max_size from bottom to upper
1044 * levels starting from VA point. The propagation must be done
1045 * when VA size is modified by changing its va_start/va_end. Or
1046 * in case of newly inserting of VA to the tree.
1047 *
1048 * It means that __augment_tree_propagate_from() must be called:
1049 * - After VA has been inserted to the tree(free path);
1050 * - After VA has been shrunk(allocation path);
1051 * - After VA has been increased(merging path).
1052 *
1053 * Please note that, it does not mean that upper parent nodes
1054 * and their subtree_max_size are recalculated all the time up
1055 * to the root node.
1056 *
1057 * 4--8
1058 * /\
1059 * / \
1060 * / \
1061 * 2--2 8--8
1062 *
1063 * For example if we modify the node 4, shrinking it to 2, then
1064 * no any modification is required. If we shrink the node 2 to 1
1065 * its subtree_max_size is updated only, and set to 1. If we shrink
1066 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1067 * node becomes 4--6.
1068 */
1069static __always_inline void
1070augment_tree_propagate_from(struct vmap_area *va)
1071{
1072 /*
1073 * Populate the tree from bottom towards the root until
1074 * the calculated maximum available size of checked node
1075 * is equal to its current one.
1076 */
1077 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1078
1079#if DEBUG_AUGMENT_PROPAGATE_CHECK
1080 augment_tree_propagate_check();
1081#endif
1082}
1083
1084static void
1085insert_vmap_area(struct vmap_area *va,
1086 struct rb_root *root, struct list_head *head)
1087{
1088 struct rb_node **link;
1089 struct rb_node *parent;
1090
1091 link = find_va_links(va, root, NULL, &parent);
1092 if (link)
1093 link_va(va, root, parent, link, head);
1094}
1095
1096static void
1097insert_vmap_area_augment(struct vmap_area *va,
1098 struct rb_node *from, struct rb_root *root,
1099 struct list_head *head)
1100{
1101 struct rb_node **link;
1102 struct rb_node *parent;
1103
1104 if (from)
1105 link = find_va_links(va, NULL, from, &parent);
1106 else
1107 link = find_va_links(va, root, NULL, &parent);
1108
1109 if (link) {
1110 link_va_augment(va, root, parent, link, head);
1111 augment_tree_propagate_from(va);
1112 }
1113}
1114
1115/*
1116 * Merge de-allocated chunk of VA memory with previous
1117 * and next free blocks. If coalesce is not done a new
1118 * free area is inserted. If VA has been merged, it is
1119 * freed.
1120 *
1121 * Please note, it can return NULL in case of overlap
1122 * ranges, followed by WARN() report. Despite it is a
1123 * buggy behaviour, a system can be alive and keep
1124 * ongoing.
1125 */
1126static __always_inline struct vmap_area *
1127__merge_or_add_vmap_area(struct vmap_area *va,
1128 struct rb_root *root, struct list_head *head, bool augment)
1129{
1130 struct vmap_area *sibling;
1131 struct list_head *next;
1132 struct rb_node **link;
1133 struct rb_node *parent;
1134 bool merged = false;
1135
1136 /*
1137 * Find a place in the tree where VA potentially will be
1138 * inserted, unless it is merged with its sibling/siblings.
1139 */
1140 link = find_va_links(va, root, NULL, &parent);
1141 if (!link)
1142 return NULL;
1143
1144 /*
1145 * Get next node of VA to check if merging can be done.
1146 */
1147 next = get_va_next_sibling(parent, link);
1148 if (unlikely(next == NULL))
1149 goto insert;
1150
1151 /*
1152 * start end
1153 * | |
1154 * |<------VA------>|<-----Next----->|
1155 * | |
1156 * start end
1157 */
1158 if (next != head) {
1159 sibling = list_entry(next, struct vmap_area, list);
1160 if (sibling->va_start == va->va_end) {
1161 sibling->va_start = va->va_start;
1162
1163 /* Free vmap_area object. */
1164 kmem_cache_free(vmap_area_cachep, va);
1165
1166 /* Point to the new merged area. */
1167 va = sibling;
1168 merged = true;
1169 }
1170 }
1171
1172 /*
1173 * start end
1174 * | |
1175 * |<-----Prev----->|<------VA------>|
1176 * | |
1177 * start end
1178 */
1179 if (next->prev != head) {
1180 sibling = list_entry(next->prev, struct vmap_area, list);
1181 if (sibling->va_end == va->va_start) {
1182 /*
1183 * If both neighbors are coalesced, it is important
1184 * to unlink the "next" node first, followed by merging
1185 * with "previous" one. Otherwise the tree might not be
1186 * fully populated if a sibling's augmented value is
1187 * "normalized" because of rotation operations.
1188 */
1189 if (merged)
1190 __unlink_va(va, root, augment);
1191
1192 sibling->va_end = va->va_end;
1193
1194 /* Free vmap_area object. */
1195 kmem_cache_free(vmap_area_cachep, va);
1196
1197 /* Point to the new merged area. */
1198 va = sibling;
1199 merged = true;
1200 }
1201 }
1202
1203insert:
1204 if (!merged)
1205 __link_va(va, root, parent, link, head, augment);
1206
1207 return va;
1208}
1209
1210static __always_inline struct vmap_area *
1211merge_or_add_vmap_area(struct vmap_area *va,
1212 struct rb_root *root, struct list_head *head)
1213{
1214 return __merge_or_add_vmap_area(va, root, head, false);
1215}
1216
1217static __always_inline struct vmap_area *
1218merge_or_add_vmap_area_augment(struct vmap_area *va,
1219 struct rb_root *root, struct list_head *head)
1220{
1221 va = __merge_or_add_vmap_area(va, root, head, true);
1222 if (va)
1223 augment_tree_propagate_from(va);
1224
1225 return va;
1226}
1227
1228static __always_inline bool
1229is_within_this_va(struct vmap_area *va, unsigned long size,
1230 unsigned long align, unsigned long vstart)
1231{
1232 unsigned long nva_start_addr;
1233
1234 if (va->va_start > vstart)
1235 nva_start_addr = ALIGN(va->va_start, align);
1236 else
1237 nva_start_addr = ALIGN(vstart, align);
1238
1239 /* Can be overflowed due to big size or alignment. */
1240 if (nva_start_addr + size < nva_start_addr ||
1241 nva_start_addr < vstart)
1242 return false;
1243
1244 return (nva_start_addr + size <= va->va_end);
1245}
1246
1247/*
1248 * Find the first free block(lowest start address) in the tree,
1249 * that will accomplish the request corresponding to passing
1250 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1251 * a search length is adjusted to account for worst case alignment
1252 * overhead.
1253 */
1254static __always_inline struct vmap_area *
1255find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1256 unsigned long align, unsigned long vstart, bool adjust_search_size)
1257{
1258 struct vmap_area *va;
1259 struct rb_node *node;
1260 unsigned long length;
1261
1262 /* Start from the root. */
1263 node = root->rb_node;
1264
1265 /* Adjust the search size for alignment overhead. */
1266 length = adjust_search_size ? size + align - 1 : size;
1267
1268 while (node) {
1269 va = rb_entry(node, struct vmap_area, rb_node);
1270
1271 if (get_subtree_max_size(node->rb_left) >= length &&
1272 vstart < va->va_start) {
1273 node = node->rb_left;
1274 } else {
1275 if (is_within_this_va(va, size, align, vstart))
1276 return va;
1277
1278 /*
1279 * Does not make sense to go deeper towards the right
1280 * sub-tree if it does not have a free block that is
1281 * equal or bigger to the requested search length.
1282 */
1283 if (get_subtree_max_size(node->rb_right) >= length) {
1284 node = node->rb_right;
1285 continue;
1286 }
1287
1288 /*
1289 * OK. We roll back and find the first right sub-tree,
1290 * that will satisfy the search criteria. It can happen
1291 * due to "vstart" restriction or an alignment overhead
1292 * that is bigger then PAGE_SIZE.
1293 */
1294 while ((node = rb_parent(node))) {
1295 va = rb_entry(node, struct vmap_area, rb_node);
1296 if (is_within_this_va(va, size, align, vstart))
1297 return va;
1298
1299 if (get_subtree_max_size(node->rb_right) >= length &&
1300 vstart <= va->va_start) {
1301 /*
1302 * Shift the vstart forward. Please note, we update it with
1303 * parent's start address adding "1" because we do not want
1304 * to enter same sub-tree after it has already been checked
1305 * and no suitable free block found there.
1306 */
1307 vstart = va->va_start + 1;
1308 node = node->rb_right;
1309 break;
1310 }
1311 }
1312 }
1313 }
1314
1315 return NULL;
1316}
1317
1318#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1319#include <linux/random.h>
1320
1321static struct vmap_area *
1322find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1323 unsigned long align, unsigned long vstart)
1324{
1325 struct vmap_area *va;
1326
1327 list_for_each_entry(va, head, list) {
1328 if (!is_within_this_va(va, size, align, vstart))
1329 continue;
1330
1331 return va;
1332 }
1333
1334 return NULL;
1335}
1336
1337static void
1338find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1339 unsigned long size, unsigned long align)
1340{
1341 struct vmap_area *va_1, *va_2;
1342 unsigned long vstart;
1343 unsigned int rnd;
1344
1345 get_random_bytes(&rnd, sizeof(rnd));
1346 vstart = VMALLOC_START + rnd;
1347
1348 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1349 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1350
1351 if (va_1 != va_2)
1352 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1353 va_1, va_2, vstart);
1354}
1355#endif
1356
1357enum fit_type {
1358 NOTHING_FIT = 0,
1359 FL_FIT_TYPE = 1, /* full fit */
1360 LE_FIT_TYPE = 2, /* left edge fit */
1361 RE_FIT_TYPE = 3, /* right edge fit */
1362 NE_FIT_TYPE = 4 /* no edge fit */
1363};
1364
1365static __always_inline enum fit_type
1366classify_va_fit_type(struct vmap_area *va,
1367 unsigned long nva_start_addr, unsigned long size)
1368{
1369 enum fit_type type;
1370
1371 /* Check if it is within VA. */
1372 if (nva_start_addr < va->va_start ||
1373 nva_start_addr + size > va->va_end)
1374 return NOTHING_FIT;
1375
1376 /* Now classify. */
1377 if (va->va_start == nva_start_addr) {
1378 if (va->va_end == nva_start_addr + size)
1379 type = FL_FIT_TYPE;
1380 else
1381 type = LE_FIT_TYPE;
1382 } else if (va->va_end == nva_start_addr + size) {
1383 type = RE_FIT_TYPE;
1384 } else {
1385 type = NE_FIT_TYPE;
1386 }
1387
1388 return type;
1389}
1390
1391static __always_inline int
1392adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1393 struct vmap_area *va, unsigned long nva_start_addr,
1394 unsigned long size)
1395{
1396 struct vmap_area *lva = NULL;
1397 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1398
1399 if (type == FL_FIT_TYPE) {
1400 /*
1401 * No need to split VA, it fully fits.
1402 *
1403 * | |
1404 * V NVA V
1405 * |---------------|
1406 */
1407 unlink_va_augment(va, root);
1408 kmem_cache_free(vmap_area_cachep, va);
1409 } else if (type == LE_FIT_TYPE) {
1410 /*
1411 * Split left edge of fit VA.
1412 *
1413 * | |
1414 * V NVA V R
1415 * |-------|-------|
1416 */
1417 va->va_start += size;
1418 } else if (type == RE_FIT_TYPE) {
1419 /*
1420 * Split right edge of fit VA.
1421 *
1422 * | |
1423 * L V NVA V
1424 * |-------|-------|
1425 */
1426 va->va_end = nva_start_addr;
1427 } else if (type == NE_FIT_TYPE) {
1428 /*
1429 * Split no edge of fit VA.
1430 *
1431 * | |
1432 * L V NVA V R
1433 * |---|-------|---|
1434 */
1435 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1436 if (unlikely(!lva)) {
1437 /*
1438 * For percpu allocator we do not do any pre-allocation
1439 * and leave it as it is. The reason is it most likely
1440 * never ends up with NE_FIT_TYPE splitting. In case of
1441 * percpu allocations offsets and sizes are aligned to
1442 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1443 * are its main fitting cases.
1444 *
1445 * There are a few exceptions though, as an example it is
1446 * a first allocation (early boot up) when we have "one"
1447 * big free space that has to be split.
1448 *
1449 * Also we can hit this path in case of regular "vmap"
1450 * allocations, if "this" current CPU was not preloaded.
1451 * See the comment in alloc_vmap_area() why. If so, then
1452 * GFP_NOWAIT is used instead to get an extra object for
1453 * split purpose. That is rare and most time does not
1454 * occur.
1455 *
1456 * What happens if an allocation gets failed. Basically,
1457 * an "overflow" path is triggered to purge lazily freed
1458 * areas to free some memory, then, the "retry" path is
1459 * triggered to repeat one more time. See more details
1460 * in alloc_vmap_area() function.
1461 */
1462 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1463 if (!lva)
1464 return -1;
1465 }
1466
1467 /*
1468 * Build the remainder.
1469 */
1470 lva->va_start = va->va_start;
1471 lva->va_end = nva_start_addr;
1472
1473 /*
1474 * Shrink this VA to remaining size.
1475 */
1476 va->va_start = nva_start_addr + size;
1477 } else {
1478 return -1;
1479 }
1480
1481 if (type != FL_FIT_TYPE) {
1482 augment_tree_propagate_from(va);
1483
1484 if (lva) /* type == NE_FIT_TYPE */
1485 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1486 }
1487
1488 return 0;
1489}
1490
1491/*
1492 * Returns a start address of the newly allocated area, if success.
1493 * Otherwise a vend is returned that indicates failure.
1494 */
1495static __always_inline unsigned long
1496__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1497 unsigned long size, unsigned long align,
1498 unsigned long vstart, unsigned long vend)
1499{
1500 bool adjust_search_size = true;
1501 unsigned long nva_start_addr;
1502 struct vmap_area *va;
1503 int ret;
1504
1505 /*
1506 * Do not adjust when:
1507 * a) align <= PAGE_SIZE, because it does not make any sense.
1508 * All blocks(their start addresses) are at least PAGE_SIZE
1509 * aligned anyway;
1510 * b) a short range where a requested size corresponds to exactly
1511 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1512 * With adjusted search length an allocation would not succeed.
1513 */
1514 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1515 adjust_search_size = false;
1516
1517 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1518 if (unlikely(!va))
1519 return vend;
1520
1521 if (va->va_start > vstart)
1522 nva_start_addr = ALIGN(va->va_start, align);
1523 else
1524 nva_start_addr = ALIGN(vstart, align);
1525
1526 /* Check the "vend" restriction. */
1527 if (nva_start_addr + size > vend)
1528 return vend;
1529
1530 /* Update the free vmap_area. */
1531 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1532 if (WARN_ON_ONCE(ret))
1533 return vend;
1534
1535#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1536 find_vmap_lowest_match_check(root, head, size, align);
1537#endif
1538
1539 return nva_start_addr;
1540}
1541
1542/*
1543 * Free a region of KVA allocated by alloc_vmap_area
1544 */
1545static void free_vmap_area(struct vmap_area *va)
1546{
1547 /*
1548 * Remove from the busy tree/list.
1549 */
1550 spin_lock(&vmap_area_lock);
1551 unlink_va(va, &vmap_area_root);
1552 spin_unlock(&vmap_area_lock);
1553
1554 /*
1555 * Insert/Merge it back to the free tree/list.
1556 */
1557 spin_lock(&free_vmap_area_lock);
1558 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1559 spin_unlock(&free_vmap_area_lock);
1560}
1561
1562static inline void
1563preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1564{
1565 struct vmap_area *va = NULL;
1566
1567 /*
1568 * Preload this CPU with one extra vmap_area object. It is used
1569 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1570 * a CPU that does an allocation is preloaded.
1571 *
1572 * We do it in non-atomic context, thus it allows us to use more
1573 * permissive allocation masks to be more stable under low memory
1574 * condition and high memory pressure.
1575 */
1576 if (!this_cpu_read(ne_fit_preload_node))
1577 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1578
1579 spin_lock(lock);
1580
1581 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1582 kmem_cache_free(vmap_area_cachep, va);
1583}
1584
1585/*
1586 * Allocate a region of KVA of the specified size and alignment, within the
1587 * vstart and vend.
1588 */
1589static struct vmap_area *alloc_vmap_area(unsigned long size,
1590 unsigned long align,
1591 unsigned long vstart, unsigned long vend,
1592 int node, gfp_t gfp_mask)
1593{
1594 struct vmap_area *va;
1595 unsigned long freed;
1596 unsigned long addr;
1597 int purged = 0;
1598 int ret;
1599
1600 BUG_ON(!size);
1601 BUG_ON(offset_in_page(size));
1602 BUG_ON(!is_power_of_2(align));
1603
1604 if (unlikely(!vmap_initialized))
1605 return ERR_PTR(-EBUSY);
1606
1607 might_sleep();
1608 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1609
1610 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1611 if (unlikely(!va))
1612 return ERR_PTR(-ENOMEM);
1613
1614 /*
1615 * Only scan the relevant parts containing pointers to other objects
1616 * to avoid false negatives.
1617 */
1618 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1619
1620retry:
1621 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1622 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1623 size, align, vstart, vend);
1624 spin_unlock(&free_vmap_area_lock);
1625
1626 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1627
1628 /*
1629 * If an allocation fails, the "vend" address is
1630 * returned. Therefore trigger the overflow path.
1631 */
1632 if (unlikely(addr == vend))
1633 goto overflow;
1634
1635 va->va_start = addr;
1636 va->va_end = addr + size;
1637 va->vm = NULL;
1638
1639 spin_lock(&vmap_area_lock);
1640 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1641 spin_unlock(&vmap_area_lock);
1642
1643 BUG_ON(!IS_ALIGNED(va->va_start, align));
1644 BUG_ON(va->va_start < vstart);
1645 BUG_ON(va->va_end > vend);
1646
1647 ret = kasan_populate_vmalloc(addr, size);
1648 if (ret) {
1649 free_vmap_area(va);
1650 return ERR_PTR(ret);
1651 }
1652
1653 return va;
1654
1655overflow:
1656 if (!purged) {
1657 purge_vmap_area_lazy();
1658 purged = 1;
1659 goto retry;
1660 }
1661
1662 freed = 0;
1663 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1664
1665 if (freed > 0) {
1666 purged = 0;
1667 goto retry;
1668 }
1669
1670 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1671 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1672 size);
1673
1674 kmem_cache_free(vmap_area_cachep, va);
1675 return ERR_PTR(-EBUSY);
1676}
1677
1678int register_vmap_purge_notifier(struct notifier_block *nb)
1679{
1680 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1681}
1682EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1683
1684int unregister_vmap_purge_notifier(struct notifier_block *nb)
1685{
1686 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1687}
1688EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1689
1690/*
1691 * lazy_max_pages is the maximum amount of virtual address space we gather up
1692 * before attempting to purge with a TLB flush.
1693 *
1694 * There is a tradeoff here: a larger number will cover more kernel page tables
1695 * and take slightly longer to purge, but it will linearly reduce the number of
1696 * global TLB flushes that must be performed. It would seem natural to scale
1697 * this number up linearly with the number of CPUs (because vmapping activity
1698 * could also scale linearly with the number of CPUs), however it is likely
1699 * that in practice, workloads might be constrained in other ways that mean
1700 * vmap activity will not scale linearly with CPUs. Also, I want to be
1701 * conservative and not introduce a big latency on huge systems, so go with
1702 * a less aggressive log scale. It will still be an improvement over the old
1703 * code, and it will be simple to change the scale factor if we find that it
1704 * becomes a problem on bigger systems.
1705 */
1706static unsigned long lazy_max_pages(void)
1707{
1708 unsigned int log;
1709
1710 log = fls(num_online_cpus());
1711
1712 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1713}
1714
1715static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1716
1717/*
1718 * Serialize vmap purging. There is no actual critical section protected
1719 * by this lock, but we want to avoid concurrent calls for performance
1720 * reasons and to make the pcpu_get_vm_areas more deterministic.
1721 */
1722static DEFINE_MUTEX(vmap_purge_lock);
1723
1724/* for per-CPU blocks */
1725static void purge_fragmented_blocks_allcpus(void);
1726
1727/*
1728 * Purges all lazily-freed vmap areas.
1729 */
1730static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1731{
1732 unsigned long resched_threshold;
1733 unsigned int num_purged_areas = 0;
1734 struct list_head local_purge_list;
1735 struct vmap_area *va, *n_va;
1736
1737 lockdep_assert_held(&vmap_purge_lock);
1738
1739 spin_lock(&purge_vmap_area_lock);
1740 purge_vmap_area_root = RB_ROOT;
1741 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1742 spin_unlock(&purge_vmap_area_lock);
1743
1744 if (unlikely(list_empty(&local_purge_list)))
1745 goto out;
1746
1747 start = min(start,
1748 list_first_entry(&local_purge_list,
1749 struct vmap_area, list)->va_start);
1750
1751 end = max(end,
1752 list_last_entry(&local_purge_list,
1753 struct vmap_area, list)->va_end);
1754
1755 flush_tlb_kernel_range(start, end);
1756 resched_threshold = lazy_max_pages() << 1;
1757
1758 spin_lock(&free_vmap_area_lock);
1759 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1760 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1761 unsigned long orig_start = va->va_start;
1762 unsigned long orig_end = va->va_end;
1763
1764 /*
1765 * Finally insert or merge lazily-freed area. It is
1766 * detached and there is no need to "unlink" it from
1767 * anything.
1768 */
1769 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1770 &free_vmap_area_list);
1771
1772 if (!va)
1773 continue;
1774
1775 if (is_vmalloc_or_module_addr((void *)orig_start))
1776 kasan_release_vmalloc(orig_start, orig_end,
1777 va->va_start, va->va_end);
1778
1779 atomic_long_sub(nr, &vmap_lazy_nr);
1780 num_purged_areas++;
1781
1782 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1783 cond_resched_lock(&free_vmap_area_lock);
1784 }
1785 spin_unlock(&free_vmap_area_lock);
1786
1787out:
1788 trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1789 return num_purged_areas > 0;
1790}
1791
1792/*
1793 * Kick off a purge of the outstanding lazy areas.
1794 */
1795static void purge_vmap_area_lazy(void)
1796{
1797 mutex_lock(&vmap_purge_lock);
1798 purge_fragmented_blocks_allcpus();
1799 __purge_vmap_area_lazy(ULONG_MAX, 0);
1800 mutex_unlock(&vmap_purge_lock);
1801}
1802
1803static void drain_vmap_area_work(struct work_struct *work)
1804{
1805 unsigned long nr_lazy;
1806
1807 do {
1808 mutex_lock(&vmap_purge_lock);
1809 __purge_vmap_area_lazy(ULONG_MAX, 0);
1810 mutex_unlock(&vmap_purge_lock);
1811
1812 /* Recheck if further work is required. */
1813 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1814 } while (nr_lazy > lazy_max_pages());
1815}
1816
1817/*
1818 * Free a vmap area, caller ensuring that the area has been unmapped
1819 * and flush_cache_vunmap had been called for the correct range
1820 * previously.
1821 */
1822static void free_vmap_area_noflush(struct vmap_area *va)
1823{
1824 unsigned long nr_lazy_max = lazy_max_pages();
1825 unsigned long va_start = va->va_start;
1826 unsigned long nr_lazy;
1827
1828 spin_lock(&vmap_area_lock);
1829 unlink_va(va, &vmap_area_root);
1830 spin_unlock(&vmap_area_lock);
1831
1832 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1833 PAGE_SHIFT, &vmap_lazy_nr);
1834
1835 /*
1836 * Merge or place it to the purge tree/list.
1837 */
1838 spin_lock(&purge_vmap_area_lock);
1839 merge_or_add_vmap_area(va,
1840 &purge_vmap_area_root, &purge_vmap_area_list);
1841 spin_unlock(&purge_vmap_area_lock);
1842
1843 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1844
1845 /* After this point, we may free va at any time */
1846 if (unlikely(nr_lazy > nr_lazy_max))
1847 schedule_work(&drain_vmap_work);
1848}
1849
1850/*
1851 * Free and unmap a vmap area
1852 */
1853static void free_unmap_vmap_area(struct vmap_area *va)
1854{
1855 flush_cache_vunmap(va->va_start, va->va_end);
1856 vunmap_range_noflush(va->va_start, va->va_end);
1857 if (debug_pagealloc_enabled_static())
1858 flush_tlb_kernel_range(va->va_start, va->va_end);
1859
1860 free_vmap_area_noflush(va);
1861}
1862
1863struct vmap_area *find_vmap_area(unsigned long addr)
1864{
1865 struct vmap_area *va;
1866
1867 spin_lock(&vmap_area_lock);
1868 va = __find_vmap_area(addr, &vmap_area_root);
1869 spin_unlock(&vmap_area_lock);
1870
1871 return va;
1872}
1873
1874/*** Per cpu kva allocator ***/
1875
1876/*
1877 * vmap space is limited especially on 32 bit architectures. Ensure there is
1878 * room for at least 16 percpu vmap blocks per CPU.
1879 */
1880/*
1881 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1882 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1883 * instead (we just need a rough idea)
1884 */
1885#if BITS_PER_LONG == 32
1886#define VMALLOC_SPACE (128UL*1024*1024)
1887#else
1888#define VMALLOC_SPACE (128UL*1024*1024*1024)
1889#endif
1890
1891#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1892#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1893#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1894#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1895#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1896#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1897#define VMAP_BBMAP_BITS \
1898 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1899 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1900 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1901
1902#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1903
1904struct vmap_block_queue {
1905 spinlock_t lock;
1906 struct list_head free;
1907};
1908
1909struct vmap_block {
1910 spinlock_t lock;
1911 struct vmap_area *va;
1912 unsigned long free, dirty;
1913 unsigned long dirty_min, dirty_max; /*< dirty range */
1914 struct list_head free_list;
1915 struct rcu_head rcu_head;
1916 struct list_head purge;
1917};
1918
1919/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1920static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1921
1922/*
1923 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1924 * in the free path. Could get rid of this if we change the API to return a
1925 * "cookie" from alloc, to be passed to free. But no big deal yet.
1926 */
1927static DEFINE_XARRAY(vmap_blocks);
1928
1929/*
1930 * We should probably have a fallback mechanism to allocate virtual memory
1931 * out of partially filled vmap blocks. However vmap block sizing should be
1932 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1933 * big problem.
1934 */
1935
1936static unsigned long addr_to_vb_idx(unsigned long addr)
1937{
1938 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1939 addr /= VMAP_BLOCK_SIZE;
1940 return addr;
1941}
1942
1943static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1944{
1945 unsigned long addr;
1946
1947 addr = va_start + (pages_off << PAGE_SHIFT);
1948 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1949 return (void *)addr;
1950}
1951
1952/**
1953 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1954 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1955 * @order: how many 2^order pages should be occupied in newly allocated block
1956 * @gfp_mask: flags for the page level allocator
1957 *
1958 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1959 */
1960static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1961{
1962 struct vmap_block_queue *vbq;
1963 struct vmap_block *vb;
1964 struct vmap_area *va;
1965 unsigned long vb_idx;
1966 int node, err;
1967 void *vaddr;
1968
1969 node = numa_node_id();
1970
1971 vb = kmalloc_node(sizeof(struct vmap_block),
1972 gfp_mask & GFP_RECLAIM_MASK, node);
1973 if (unlikely(!vb))
1974 return ERR_PTR(-ENOMEM);
1975
1976 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1977 VMALLOC_START, VMALLOC_END,
1978 node, gfp_mask);
1979 if (IS_ERR(va)) {
1980 kfree(vb);
1981 return ERR_CAST(va);
1982 }
1983
1984 vaddr = vmap_block_vaddr(va->va_start, 0);
1985 spin_lock_init(&vb->lock);
1986 vb->va = va;
1987 /* At least something should be left free */
1988 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1989 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1990 vb->dirty = 0;
1991 vb->dirty_min = VMAP_BBMAP_BITS;
1992 vb->dirty_max = 0;
1993 INIT_LIST_HEAD(&vb->free_list);
1994
1995 vb_idx = addr_to_vb_idx(va->va_start);
1996 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1997 if (err) {
1998 kfree(vb);
1999 free_vmap_area(va);
2000 return ERR_PTR(err);
2001 }
2002
2003 vbq = raw_cpu_ptr(&vmap_block_queue);
2004 spin_lock(&vbq->lock);
2005 list_add_tail_rcu(&vb->free_list, &vbq->free);
2006 spin_unlock(&vbq->lock);
2007
2008 return vaddr;
2009}
2010
2011static void free_vmap_block(struct vmap_block *vb)
2012{
2013 struct vmap_block *tmp;
2014
2015 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2016 BUG_ON(tmp != vb);
2017
2018 free_vmap_area_noflush(vb->va);
2019 kfree_rcu(vb, rcu_head);
2020}
2021
2022static void purge_fragmented_blocks(int cpu)
2023{
2024 LIST_HEAD(purge);
2025 struct vmap_block *vb;
2026 struct vmap_block *n_vb;
2027 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2028
2029 rcu_read_lock();
2030 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2031
2032 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2033 continue;
2034
2035 spin_lock(&vb->lock);
2036 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2037 vb->free = 0; /* prevent further allocs after releasing lock */
2038 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2039 vb->dirty_min = 0;
2040 vb->dirty_max = VMAP_BBMAP_BITS;
2041 spin_lock(&vbq->lock);
2042 list_del_rcu(&vb->free_list);
2043 spin_unlock(&vbq->lock);
2044 spin_unlock(&vb->lock);
2045 list_add_tail(&vb->purge, &purge);
2046 } else
2047 spin_unlock(&vb->lock);
2048 }
2049 rcu_read_unlock();
2050
2051 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2052 list_del(&vb->purge);
2053 free_vmap_block(vb);
2054 }
2055}
2056
2057static void purge_fragmented_blocks_allcpus(void)
2058{
2059 int cpu;
2060
2061 for_each_possible_cpu(cpu)
2062 purge_fragmented_blocks(cpu);
2063}
2064
2065static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2066{
2067 struct vmap_block_queue *vbq;
2068 struct vmap_block *vb;
2069 void *vaddr = NULL;
2070 unsigned int order;
2071
2072 BUG_ON(offset_in_page(size));
2073 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2074 if (WARN_ON(size == 0)) {
2075 /*
2076 * Allocating 0 bytes isn't what caller wants since
2077 * get_order(0) returns funny result. Just warn and terminate
2078 * early.
2079 */
2080 return NULL;
2081 }
2082 order = get_order(size);
2083
2084 rcu_read_lock();
2085 vbq = raw_cpu_ptr(&vmap_block_queue);
2086 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2087 unsigned long pages_off;
2088
2089 spin_lock(&vb->lock);
2090 if (vb->free < (1UL << order)) {
2091 spin_unlock(&vb->lock);
2092 continue;
2093 }
2094
2095 pages_off = VMAP_BBMAP_BITS - vb->free;
2096 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2097 vb->free -= 1UL << order;
2098 if (vb->free == 0) {
2099 spin_lock(&vbq->lock);
2100 list_del_rcu(&vb->free_list);
2101 spin_unlock(&vbq->lock);
2102 }
2103
2104 spin_unlock(&vb->lock);
2105 break;
2106 }
2107
2108 rcu_read_unlock();
2109
2110 /* Allocate new block if nothing was found */
2111 if (!vaddr)
2112 vaddr = new_vmap_block(order, gfp_mask);
2113
2114 return vaddr;
2115}
2116
2117static void vb_free(unsigned long addr, unsigned long size)
2118{
2119 unsigned long offset;
2120 unsigned int order;
2121 struct vmap_block *vb;
2122
2123 BUG_ON(offset_in_page(size));
2124 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2125
2126 flush_cache_vunmap(addr, addr + size);
2127
2128 order = get_order(size);
2129 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2130 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2131
2132 vunmap_range_noflush(addr, addr + size);
2133
2134 if (debug_pagealloc_enabled_static())
2135 flush_tlb_kernel_range(addr, addr + size);
2136
2137 spin_lock(&vb->lock);
2138
2139 /* Expand dirty range */
2140 vb->dirty_min = min(vb->dirty_min, offset);
2141 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2142
2143 vb->dirty += 1UL << order;
2144 if (vb->dirty == VMAP_BBMAP_BITS) {
2145 BUG_ON(vb->free);
2146 spin_unlock(&vb->lock);
2147 free_vmap_block(vb);
2148 } else
2149 spin_unlock(&vb->lock);
2150}
2151
2152static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2153{
2154 int cpu;
2155
2156 if (unlikely(!vmap_initialized))
2157 return;
2158
2159 might_sleep();
2160
2161 for_each_possible_cpu(cpu) {
2162 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2163 struct vmap_block *vb;
2164
2165 rcu_read_lock();
2166 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2167 spin_lock(&vb->lock);
2168 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2169 unsigned long va_start = vb->va->va_start;
2170 unsigned long s, e;
2171
2172 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2173 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2174
2175 start = min(s, start);
2176 end = max(e, end);
2177
2178 flush = 1;
2179 }
2180 spin_unlock(&vb->lock);
2181 }
2182 rcu_read_unlock();
2183 }
2184
2185 mutex_lock(&vmap_purge_lock);
2186 purge_fragmented_blocks_allcpus();
2187 if (!__purge_vmap_area_lazy(start, end) && flush)
2188 flush_tlb_kernel_range(start, end);
2189 mutex_unlock(&vmap_purge_lock);
2190}
2191
2192/**
2193 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2194 *
2195 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2196 * to amortize TLB flushing overheads. What this means is that any page you
2197 * have now, may, in a former life, have been mapped into kernel virtual
2198 * address by the vmap layer and so there might be some CPUs with TLB entries
2199 * still referencing that page (additional to the regular 1:1 kernel mapping).
2200 *
2201 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2202 * be sure that none of the pages we have control over will have any aliases
2203 * from the vmap layer.
2204 */
2205void vm_unmap_aliases(void)
2206{
2207 unsigned long start = ULONG_MAX, end = 0;
2208 int flush = 0;
2209
2210 _vm_unmap_aliases(start, end, flush);
2211}
2212EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2213
2214/**
2215 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2216 * @mem: the pointer returned by vm_map_ram
2217 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2218 */
2219void vm_unmap_ram(const void *mem, unsigned int count)
2220{
2221 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2222 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2223 struct vmap_area *va;
2224
2225 might_sleep();
2226 BUG_ON(!addr);
2227 BUG_ON(addr < VMALLOC_START);
2228 BUG_ON(addr > VMALLOC_END);
2229 BUG_ON(!PAGE_ALIGNED(addr));
2230
2231 kasan_poison_vmalloc(mem, size);
2232
2233 if (likely(count <= VMAP_MAX_ALLOC)) {
2234 debug_check_no_locks_freed(mem, size);
2235 vb_free(addr, size);
2236 return;
2237 }
2238
2239 va = find_vmap_area(addr);
2240 BUG_ON(!va);
2241 debug_check_no_locks_freed((void *)va->va_start,
2242 (va->va_end - va->va_start));
2243 free_unmap_vmap_area(va);
2244}
2245EXPORT_SYMBOL(vm_unmap_ram);
2246
2247/**
2248 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2249 * @pages: an array of pointers to the pages to be mapped
2250 * @count: number of pages
2251 * @node: prefer to allocate data structures on this node
2252 *
2253 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2254 * faster than vmap so it's good. But if you mix long-life and short-life
2255 * objects with vm_map_ram(), it could consume lots of address space through
2256 * fragmentation (especially on a 32bit machine). You could see failures in
2257 * the end. Please use this function for short-lived objects.
2258 *
2259 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2260 */
2261void *vm_map_ram(struct page **pages, unsigned int count, int node)
2262{
2263 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2264 unsigned long addr;
2265 void *mem;
2266
2267 if (likely(count <= VMAP_MAX_ALLOC)) {
2268 mem = vb_alloc(size, GFP_KERNEL);
2269 if (IS_ERR(mem))
2270 return NULL;
2271 addr = (unsigned long)mem;
2272 } else {
2273 struct vmap_area *va;
2274 va = alloc_vmap_area(size, PAGE_SIZE,
2275 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2276 if (IS_ERR(va))
2277 return NULL;
2278
2279 addr = va->va_start;
2280 mem = (void *)addr;
2281 }
2282
2283 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2284 pages, PAGE_SHIFT) < 0) {
2285 vm_unmap_ram(mem, count);
2286 return NULL;
2287 }
2288
2289 /*
2290 * Mark the pages as accessible, now that they are mapped.
2291 * With hardware tag-based KASAN, marking is skipped for
2292 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2293 */
2294 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2295
2296 return mem;
2297}
2298EXPORT_SYMBOL(vm_map_ram);
2299
2300static struct vm_struct *vmlist __initdata;
2301
2302static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2303{
2304#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2305 return vm->page_order;
2306#else
2307 return 0;
2308#endif
2309}
2310
2311static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2312{
2313#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2314 vm->page_order = order;
2315#else
2316 BUG_ON(order != 0);
2317#endif
2318}
2319
2320/**
2321 * vm_area_add_early - add vmap area early during boot
2322 * @vm: vm_struct to add
2323 *
2324 * This function is used to add fixed kernel vm area to vmlist before
2325 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2326 * should contain proper values and the other fields should be zero.
2327 *
2328 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2329 */
2330void __init vm_area_add_early(struct vm_struct *vm)
2331{
2332 struct vm_struct *tmp, **p;
2333
2334 BUG_ON(vmap_initialized);
2335 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2336 if (tmp->addr >= vm->addr) {
2337 BUG_ON(tmp->addr < vm->addr + vm->size);
2338 break;
2339 } else
2340 BUG_ON(tmp->addr + tmp->size > vm->addr);
2341 }
2342 vm->next = *p;
2343 *p = vm;
2344}
2345
2346/**
2347 * vm_area_register_early - register vmap area early during boot
2348 * @vm: vm_struct to register
2349 * @align: requested alignment
2350 *
2351 * This function is used to register kernel vm area before
2352 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2353 * proper values on entry and other fields should be zero. On return,
2354 * vm->addr contains the allocated address.
2355 *
2356 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2357 */
2358void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2359{
2360 unsigned long addr = ALIGN(VMALLOC_START, align);
2361 struct vm_struct *cur, **p;
2362
2363 BUG_ON(vmap_initialized);
2364
2365 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2366 if ((unsigned long)cur->addr - addr >= vm->size)
2367 break;
2368 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2369 }
2370
2371 BUG_ON(addr > VMALLOC_END - vm->size);
2372 vm->addr = (void *)addr;
2373 vm->next = *p;
2374 *p = vm;
2375 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2376}
2377
2378static void vmap_init_free_space(void)
2379{
2380 unsigned long vmap_start = 1;
2381 const unsigned long vmap_end = ULONG_MAX;
2382 struct vmap_area *busy, *free;
2383
2384 /*
2385 * B F B B B F
2386 * -|-----|.....|-----|-----|-----|.....|-
2387 * | The KVA space |
2388 * |<--------------------------------->|
2389 */
2390 list_for_each_entry(busy, &vmap_area_list, list) {
2391 if (busy->va_start - vmap_start > 0) {
2392 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2393 if (!WARN_ON_ONCE(!free)) {
2394 free->va_start = vmap_start;
2395 free->va_end = busy->va_start;
2396
2397 insert_vmap_area_augment(free, NULL,
2398 &free_vmap_area_root,
2399 &free_vmap_area_list);
2400 }
2401 }
2402
2403 vmap_start = busy->va_end;
2404 }
2405
2406 if (vmap_end - vmap_start > 0) {
2407 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2408 if (!WARN_ON_ONCE(!free)) {
2409 free->va_start = vmap_start;
2410 free->va_end = vmap_end;
2411
2412 insert_vmap_area_augment(free, NULL,
2413 &free_vmap_area_root,
2414 &free_vmap_area_list);
2415 }
2416 }
2417}
2418
2419void __init vmalloc_init(void)
2420{
2421 struct vmap_area *va;
2422 struct vm_struct *tmp;
2423 int i;
2424
2425 /*
2426 * Create the cache for vmap_area objects.
2427 */
2428 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2429
2430 for_each_possible_cpu(i) {
2431 struct vmap_block_queue *vbq;
2432 struct vfree_deferred *p;
2433
2434 vbq = &per_cpu(vmap_block_queue, i);
2435 spin_lock_init(&vbq->lock);
2436 INIT_LIST_HEAD(&vbq->free);
2437 p = &per_cpu(vfree_deferred, i);
2438 init_llist_head(&p->list);
2439 INIT_WORK(&p->wq, free_work);
2440 }
2441
2442 /* Import existing vmlist entries. */
2443 for (tmp = vmlist; tmp; tmp = tmp->next) {
2444 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2445 if (WARN_ON_ONCE(!va))
2446 continue;
2447
2448 va->va_start = (unsigned long)tmp->addr;
2449 va->va_end = va->va_start + tmp->size;
2450 va->vm = tmp;
2451 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2452 }
2453
2454 /*
2455 * Now we can initialize a free vmap space.
2456 */
2457 vmap_init_free_space();
2458 vmap_initialized = true;
2459}
2460
2461static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2462 struct vmap_area *va, unsigned long flags, const void *caller)
2463{
2464 vm->flags = flags;
2465 vm->addr = (void *)va->va_start;
2466 vm->size = va->va_end - va->va_start;
2467 vm->caller = caller;
2468 va->vm = vm;
2469}
2470
2471static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2472 unsigned long flags, const void *caller)
2473{
2474 spin_lock(&vmap_area_lock);
2475 setup_vmalloc_vm_locked(vm, va, flags, caller);
2476 spin_unlock(&vmap_area_lock);
2477}
2478
2479static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2480{
2481 /*
2482 * Before removing VM_UNINITIALIZED,
2483 * we should make sure that vm has proper values.
2484 * Pair with smp_rmb() in show_numa_info().
2485 */
2486 smp_wmb();
2487 vm->flags &= ~VM_UNINITIALIZED;
2488}
2489
2490static struct vm_struct *__get_vm_area_node(unsigned long size,
2491 unsigned long align, unsigned long shift, unsigned long flags,
2492 unsigned long start, unsigned long end, int node,
2493 gfp_t gfp_mask, const void *caller)
2494{
2495 struct vmap_area *va;
2496 struct vm_struct *area;
2497 unsigned long requested_size = size;
2498
2499 BUG_ON(in_interrupt());
2500 size = ALIGN(size, 1ul << shift);
2501 if (unlikely(!size))
2502 return NULL;
2503
2504 if (flags & VM_IOREMAP)
2505 align = 1ul << clamp_t(int, get_count_order_long(size),
2506 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2507
2508 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2509 if (unlikely(!area))
2510 return NULL;
2511
2512 if (!(flags & VM_NO_GUARD))
2513 size += PAGE_SIZE;
2514
2515 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2516 if (IS_ERR(va)) {
2517 kfree(area);
2518 return NULL;
2519 }
2520
2521 setup_vmalloc_vm(area, va, flags, caller);
2522
2523 /*
2524 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2525 * best-effort approach, as they can be mapped outside of vmalloc code.
2526 * For VM_ALLOC mappings, the pages are marked as accessible after
2527 * getting mapped in __vmalloc_node_range().
2528 * With hardware tag-based KASAN, marking is skipped for
2529 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2530 */
2531 if (!(flags & VM_ALLOC))
2532 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2533 KASAN_VMALLOC_PROT_NORMAL);
2534
2535 return area;
2536}
2537
2538struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2539 unsigned long start, unsigned long end,
2540 const void *caller)
2541{
2542 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2543 NUMA_NO_NODE, GFP_KERNEL, caller);
2544}
2545
2546/**
2547 * get_vm_area - reserve a contiguous kernel virtual area
2548 * @size: size of the area
2549 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2550 *
2551 * Search an area of @size in the kernel virtual mapping area,
2552 * and reserved it for out purposes. Returns the area descriptor
2553 * on success or %NULL on failure.
2554 *
2555 * Return: the area descriptor on success or %NULL on failure.
2556 */
2557struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2558{
2559 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2560 VMALLOC_START, VMALLOC_END,
2561 NUMA_NO_NODE, GFP_KERNEL,
2562 __builtin_return_address(0));
2563}
2564
2565struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2566 const void *caller)
2567{
2568 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2569 VMALLOC_START, VMALLOC_END,
2570 NUMA_NO_NODE, GFP_KERNEL, caller);
2571}
2572
2573/**
2574 * find_vm_area - find a continuous kernel virtual area
2575 * @addr: base address
2576 *
2577 * Search for the kernel VM area starting at @addr, and return it.
2578 * It is up to the caller to do all required locking to keep the returned
2579 * pointer valid.
2580 *
2581 * Return: the area descriptor on success or %NULL on failure.
2582 */
2583struct vm_struct *find_vm_area(const void *addr)
2584{
2585 struct vmap_area *va;
2586
2587 va = find_vmap_area((unsigned long)addr);
2588 if (!va)
2589 return NULL;
2590
2591 return va->vm;
2592}
2593
2594/**
2595 * remove_vm_area - find and remove a continuous kernel virtual area
2596 * @addr: base address
2597 *
2598 * Search for the kernel VM area starting at @addr, and remove it.
2599 * This function returns the found VM area, but using it is NOT safe
2600 * on SMP machines, except for its size or flags.
2601 *
2602 * Return: the area descriptor on success or %NULL on failure.
2603 */
2604struct vm_struct *remove_vm_area(const void *addr)
2605{
2606 struct vmap_area *va;
2607
2608 might_sleep();
2609
2610 spin_lock(&vmap_area_lock);
2611 va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2612 if (va && va->vm) {
2613 struct vm_struct *vm = va->vm;
2614
2615 va->vm = NULL;
2616 spin_unlock(&vmap_area_lock);
2617
2618 kasan_free_module_shadow(vm);
2619 free_unmap_vmap_area(va);
2620
2621 return vm;
2622 }
2623
2624 spin_unlock(&vmap_area_lock);
2625 return NULL;
2626}
2627
2628static inline void set_area_direct_map(const struct vm_struct *area,
2629 int (*set_direct_map)(struct page *page))
2630{
2631 int i;
2632
2633 /* HUGE_VMALLOC passes small pages to set_direct_map */
2634 for (i = 0; i < area->nr_pages; i++)
2635 if (page_address(area->pages[i]))
2636 set_direct_map(area->pages[i]);
2637}
2638
2639/* Handle removing and resetting vm mappings related to the vm_struct. */
2640static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2641{
2642 unsigned long start = ULONG_MAX, end = 0;
2643 unsigned int page_order = vm_area_page_order(area);
2644 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2645 int flush_dmap = 0;
2646 int i;
2647
2648 remove_vm_area(area->addr);
2649
2650 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2651 if (!flush_reset)
2652 return;
2653
2654 /*
2655 * If not deallocating pages, just do the flush of the VM area and
2656 * return.
2657 */
2658 if (!deallocate_pages) {
2659 vm_unmap_aliases();
2660 return;
2661 }
2662
2663 /*
2664 * If execution gets here, flush the vm mapping and reset the direct
2665 * map. Find the start and end range of the direct mappings to make sure
2666 * the vm_unmap_aliases() flush includes the direct map.
2667 */
2668 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2669 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2670 if (addr) {
2671 unsigned long page_size;
2672
2673 page_size = PAGE_SIZE << page_order;
2674 start = min(addr, start);
2675 end = max(addr + page_size, end);
2676 flush_dmap = 1;
2677 }
2678 }
2679
2680 /*
2681 * Set direct map to something invalid so that it won't be cached if
2682 * there are any accesses after the TLB flush, then flush the TLB and
2683 * reset the direct map permissions to the default.
2684 */
2685 set_area_direct_map(area, set_direct_map_invalid_noflush);
2686 _vm_unmap_aliases(start, end, flush_dmap);
2687 set_area_direct_map(area, set_direct_map_default_noflush);
2688}
2689
2690static void __vunmap(const void *addr, int deallocate_pages)
2691{
2692 struct vm_struct *area;
2693
2694 if (!addr)
2695 return;
2696
2697 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2698 addr))
2699 return;
2700
2701 area = find_vm_area(addr);
2702 if (unlikely(!area)) {
2703 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2704 addr);
2705 return;
2706 }
2707
2708 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2709 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2710
2711 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2712
2713 vm_remove_mappings(area, deallocate_pages);
2714
2715 if (deallocate_pages) {
2716 int i;
2717
2718 for (i = 0; i < area->nr_pages; i++) {
2719 struct page *page = area->pages[i];
2720
2721 BUG_ON(!page);
2722 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2723 /*
2724 * High-order allocs for huge vmallocs are split, so
2725 * can be freed as an array of order-0 allocations
2726 */
2727 __free_pages(page, 0);
2728 cond_resched();
2729 }
2730 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2731
2732 kvfree(area->pages);
2733 }
2734
2735 kfree(area);
2736}
2737
2738static inline void __vfree_deferred(const void *addr)
2739{
2740 /*
2741 * Use raw_cpu_ptr() because this can be called from preemptible
2742 * context. Preemption is absolutely fine here, because the llist_add()
2743 * implementation is lockless, so it works even if we are adding to
2744 * another cpu's list. schedule_work() should be fine with this too.
2745 */
2746 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2747
2748 if (llist_add((struct llist_node *)addr, &p->list))
2749 schedule_work(&p->wq);
2750}
2751
2752/**
2753 * vfree_atomic - release memory allocated by vmalloc()
2754 * @addr: memory base address
2755 *
2756 * This one is just like vfree() but can be called in any atomic context
2757 * except NMIs.
2758 */
2759void vfree_atomic(const void *addr)
2760{
2761 BUG_ON(in_nmi());
2762
2763 kmemleak_free(addr);
2764
2765 if (!addr)
2766 return;
2767 __vfree_deferred(addr);
2768}
2769
2770static void __vfree(const void *addr)
2771{
2772 if (unlikely(in_interrupt()))
2773 __vfree_deferred(addr);
2774 else
2775 __vunmap(addr, 1);
2776}
2777
2778/**
2779 * vfree - Release memory allocated by vmalloc()
2780 * @addr: Memory base address
2781 *
2782 * Free the virtually continuous memory area starting at @addr, as obtained
2783 * from one of the vmalloc() family of APIs. This will usually also free the
2784 * physical memory underlying the virtual allocation, but that memory is
2785 * reference counted, so it will not be freed until the last user goes away.
2786 *
2787 * If @addr is NULL, no operation is performed.
2788 *
2789 * Context:
2790 * May sleep if called *not* from interrupt context.
2791 * Must not be called in NMI context (strictly speaking, it could be
2792 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2793 * conventions for vfree() arch-dependent would be a really bad idea).
2794 */
2795void vfree(const void *addr)
2796{
2797 BUG_ON(in_nmi());
2798
2799 kmemleak_free(addr);
2800
2801 might_sleep_if(!in_interrupt());
2802
2803 if (!addr)
2804 return;
2805
2806 __vfree(addr);
2807}
2808EXPORT_SYMBOL(vfree);
2809
2810/**
2811 * vunmap - release virtual mapping obtained by vmap()
2812 * @addr: memory base address
2813 *
2814 * Free the virtually contiguous memory area starting at @addr,
2815 * which was created from the page array passed to vmap().
2816 *
2817 * Must not be called in interrupt context.
2818 */
2819void vunmap(const void *addr)
2820{
2821 BUG_ON(in_interrupt());
2822 might_sleep();
2823 if (addr)
2824 __vunmap(addr, 0);
2825}
2826EXPORT_SYMBOL(vunmap);
2827
2828/**
2829 * vmap - map an array of pages into virtually contiguous space
2830 * @pages: array of page pointers
2831 * @count: number of pages to map
2832 * @flags: vm_area->flags
2833 * @prot: page protection for the mapping
2834 *
2835 * Maps @count pages from @pages into contiguous kernel virtual space.
2836 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2837 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2838 * are transferred from the caller to vmap(), and will be freed / dropped when
2839 * vfree() is called on the return value.
2840 *
2841 * Return: the address of the area or %NULL on failure
2842 */
2843void *vmap(struct page **pages, unsigned int count,
2844 unsigned long flags, pgprot_t prot)
2845{
2846 struct vm_struct *area;
2847 unsigned long addr;
2848 unsigned long size; /* In bytes */
2849
2850 might_sleep();
2851
2852 /*
2853 * Your top guard is someone else's bottom guard. Not having a top
2854 * guard compromises someone else's mappings too.
2855 */
2856 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2857 flags &= ~VM_NO_GUARD;
2858
2859 if (count > totalram_pages())
2860 return NULL;
2861
2862 size = (unsigned long)count << PAGE_SHIFT;
2863 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2864 if (!area)
2865 return NULL;
2866
2867 addr = (unsigned long)area->addr;
2868 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2869 pages, PAGE_SHIFT) < 0) {
2870 vunmap(area->addr);
2871 return NULL;
2872 }
2873
2874 if (flags & VM_MAP_PUT_PAGES) {
2875 area->pages = pages;
2876 area->nr_pages = count;
2877 }
2878 return area->addr;
2879}
2880EXPORT_SYMBOL(vmap);
2881
2882#ifdef CONFIG_VMAP_PFN
2883struct vmap_pfn_data {
2884 unsigned long *pfns;
2885 pgprot_t prot;
2886 unsigned int idx;
2887};
2888
2889static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2890{
2891 struct vmap_pfn_data *data = private;
2892
2893 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2894 return -EINVAL;
2895 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2896 return 0;
2897}
2898
2899/**
2900 * vmap_pfn - map an array of PFNs into virtually contiguous space
2901 * @pfns: array of PFNs
2902 * @count: number of pages to map
2903 * @prot: page protection for the mapping
2904 *
2905 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2906 * the start address of the mapping.
2907 */
2908void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2909{
2910 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2911 struct vm_struct *area;
2912
2913 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2914 __builtin_return_address(0));
2915 if (!area)
2916 return NULL;
2917 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2918 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2919 free_vm_area(area);
2920 return NULL;
2921 }
2922 return area->addr;
2923}
2924EXPORT_SYMBOL_GPL(vmap_pfn);
2925#endif /* CONFIG_VMAP_PFN */
2926
2927static inline unsigned int
2928vm_area_alloc_pages(gfp_t gfp, int nid,
2929 unsigned int order, unsigned int nr_pages, struct page **pages)
2930{
2931 unsigned int nr_allocated = 0;
2932 struct page *page;
2933 int i;
2934
2935 /*
2936 * For order-0 pages we make use of bulk allocator, if
2937 * the page array is partly or not at all populated due
2938 * to fails, fallback to a single page allocator that is
2939 * more permissive.
2940 */
2941 if (!order) {
2942 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2943
2944 while (nr_allocated < nr_pages) {
2945 unsigned int nr, nr_pages_request;
2946
2947 /*
2948 * A maximum allowed request is hard-coded and is 100
2949 * pages per call. That is done in order to prevent a
2950 * long preemption off scenario in the bulk-allocator
2951 * so the range is [1:100].
2952 */
2953 nr_pages_request = min(100U, nr_pages - nr_allocated);
2954
2955 /* memory allocation should consider mempolicy, we can't
2956 * wrongly use nearest node when nid == NUMA_NO_NODE,
2957 * otherwise memory may be allocated in only one node,
2958 * but mempolicy wants to alloc memory by interleaving.
2959 */
2960 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2961 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2962 nr_pages_request,
2963 pages + nr_allocated);
2964
2965 else
2966 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2967 nr_pages_request,
2968 pages + nr_allocated);
2969
2970 nr_allocated += nr;
2971 cond_resched();
2972
2973 /*
2974 * If zero or pages were obtained partly,
2975 * fallback to a single page allocator.
2976 */
2977 if (nr != nr_pages_request)
2978 break;
2979 }
2980 }
2981
2982 /* High-order pages or fallback path if "bulk" fails. */
2983
2984 while (nr_allocated < nr_pages) {
2985 if (fatal_signal_pending(current))
2986 break;
2987
2988 if (nid == NUMA_NO_NODE)
2989 page = alloc_pages(gfp, order);
2990 else
2991 page = alloc_pages_node(nid, gfp, order);
2992 if (unlikely(!page))
2993 break;
2994 /*
2995 * Higher order allocations must be able to be treated as
2996 * indepdenent small pages by callers (as they can with
2997 * small-page vmallocs). Some drivers do their own refcounting
2998 * on vmalloc_to_page() pages, some use page->mapping,
2999 * page->lru, etc.
3000 */
3001 if (order)
3002 split_page(page, order);
3003
3004 /*
3005 * Careful, we allocate and map page-order pages, but
3006 * tracking is done per PAGE_SIZE page so as to keep the
3007 * vm_struct APIs independent of the physical/mapped size.
3008 */
3009 for (i = 0; i < (1U << order); i++)
3010 pages[nr_allocated + i] = page + i;
3011
3012 cond_resched();
3013 nr_allocated += 1U << order;
3014 }
3015
3016 return nr_allocated;
3017}
3018
3019static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3020 pgprot_t prot, unsigned int page_shift,
3021 int node)
3022{
3023 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3024 bool nofail = gfp_mask & __GFP_NOFAIL;
3025 unsigned long addr = (unsigned long)area->addr;
3026 unsigned long size = get_vm_area_size(area);
3027 unsigned long array_size;
3028 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3029 unsigned int page_order;
3030 unsigned int flags;
3031 int ret;
3032
3033 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3034 gfp_mask |= __GFP_NOWARN;
3035 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3036 gfp_mask |= __GFP_HIGHMEM;
3037
3038 /* Please note that the recursion is strictly bounded. */
3039 if (array_size > PAGE_SIZE) {
3040 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3041 area->caller);
3042 } else {
3043 area->pages = kmalloc_node(array_size, nested_gfp, node);
3044 }
3045
3046 if (!area->pages) {
3047 warn_alloc(gfp_mask, NULL,
3048 "vmalloc error: size %lu, failed to allocated page array size %lu",
3049 nr_small_pages * PAGE_SIZE, array_size);
3050 free_vm_area(area);
3051 return NULL;
3052 }
3053
3054 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3055 page_order = vm_area_page_order(area);
3056
3057 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3058 node, page_order, nr_small_pages, area->pages);
3059
3060 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3061 if (gfp_mask & __GFP_ACCOUNT) {
3062 int i;
3063
3064 for (i = 0; i < area->nr_pages; i++)
3065 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3066 }
3067
3068 /*
3069 * If not enough pages were obtained to accomplish an
3070 * allocation request, free them via __vfree() if any.
3071 */
3072 if (area->nr_pages != nr_small_pages) {
3073 warn_alloc(gfp_mask, NULL,
3074 "vmalloc error: size %lu, page order %u, failed to allocate pages",
3075 area->nr_pages * PAGE_SIZE, page_order);
3076 goto fail;
3077 }
3078
3079 /*
3080 * page tables allocations ignore external gfp mask, enforce it
3081 * by the scope API
3082 */
3083 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3084 flags = memalloc_nofs_save();
3085 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3086 flags = memalloc_noio_save();
3087
3088 do {
3089 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3090 page_shift);
3091 if (nofail && (ret < 0))
3092 schedule_timeout_uninterruptible(1);
3093 } while (nofail && (ret < 0));
3094
3095 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3096 memalloc_nofs_restore(flags);
3097 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3098 memalloc_noio_restore(flags);
3099
3100 if (ret < 0) {
3101 warn_alloc(gfp_mask, NULL,
3102 "vmalloc error: size %lu, failed to map pages",
3103 area->nr_pages * PAGE_SIZE);
3104 goto fail;
3105 }
3106
3107 return area->addr;
3108
3109fail:
3110 __vfree(area->addr);
3111 return NULL;
3112}
3113
3114/**
3115 * __vmalloc_node_range - allocate virtually contiguous memory
3116 * @size: allocation size
3117 * @align: desired alignment
3118 * @start: vm area range start
3119 * @end: vm area range end
3120 * @gfp_mask: flags for the page level allocator
3121 * @prot: protection mask for the allocated pages
3122 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3123 * @node: node to use for allocation or NUMA_NO_NODE
3124 * @caller: caller's return address
3125 *
3126 * Allocate enough pages to cover @size from the page level
3127 * allocator with @gfp_mask flags. Please note that the full set of gfp
3128 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3129 * supported.
3130 * Zone modifiers are not supported. From the reclaim modifiers
3131 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3132 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3133 * __GFP_RETRY_MAYFAIL are not supported).
3134 *
3135 * __GFP_NOWARN can be used to suppress failures messages.
3136 *
3137 * Map them into contiguous kernel virtual space, using a pagetable
3138 * protection of @prot.
3139 *
3140 * Return: the address of the area or %NULL on failure
3141 */
3142void *__vmalloc_node_range(unsigned long size, unsigned long align,
3143 unsigned long start, unsigned long end, gfp_t gfp_mask,
3144 pgprot_t prot, unsigned long vm_flags, int node,
3145 const void *caller)
3146{
3147 struct vm_struct *area;
3148 void *ret;
3149 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3150 unsigned long real_size = size;
3151 unsigned long real_align = align;
3152 unsigned int shift = PAGE_SHIFT;
3153
3154 if (WARN_ON_ONCE(!size))
3155 return NULL;
3156
3157 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3158 warn_alloc(gfp_mask, NULL,
3159 "vmalloc error: size %lu, exceeds total pages",
3160 real_size);
3161 return NULL;
3162 }
3163
3164 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3165 unsigned long size_per_node;
3166
3167 /*
3168 * Try huge pages. Only try for PAGE_KERNEL allocations,
3169 * others like modules don't yet expect huge pages in
3170 * their allocations due to apply_to_page_range not
3171 * supporting them.
3172 */
3173
3174 size_per_node = size;
3175 if (node == NUMA_NO_NODE)
3176 size_per_node /= num_online_nodes();
3177 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3178 shift = PMD_SHIFT;
3179 else
3180 shift = arch_vmap_pte_supported_shift(size_per_node);
3181
3182 align = max(real_align, 1UL << shift);
3183 size = ALIGN(real_size, 1UL << shift);
3184 }
3185
3186again:
3187 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3188 VM_UNINITIALIZED | vm_flags, start, end, node,
3189 gfp_mask, caller);
3190 if (!area) {
3191 bool nofail = gfp_mask & __GFP_NOFAIL;
3192 warn_alloc(gfp_mask, NULL,
3193 "vmalloc error: size %lu, vm_struct allocation failed%s",
3194 real_size, (nofail) ? ". Retrying." : "");
3195 if (nofail) {
3196 schedule_timeout_uninterruptible(1);
3197 goto again;
3198 }
3199 goto fail;
3200 }
3201
3202 /*
3203 * Prepare arguments for __vmalloc_area_node() and
3204 * kasan_unpoison_vmalloc().
3205 */
3206 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3207 if (kasan_hw_tags_enabled()) {
3208 /*
3209 * Modify protection bits to allow tagging.
3210 * This must be done before mapping.
3211 */
3212 prot = arch_vmap_pgprot_tagged(prot);
3213
3214 /*
3215 * Skip page_alloc poisoning and zeroing for physical
3216 * pages backing VM_ALLOC mapping. Memory is instead
3217 * poisoned and zeroed by kasan_unpoison_vmalloc().
3218 */
3219 gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3220 }
3221
3222 /* Take note that the mapping is PAGE_KERNEL. */
3223 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3224 }
3225
3226 /* Allocate physical pages and map them into vmalloc space. */
3227 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3228 if (!ret)
3229 goto fail;
3230
3231 /*
3232 * Mark the pages as accessible, now that they are mapped.
3233 * The condition for setting KASAN_VMALLOC_INIT should complement the
3234 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3235 * to make sure that memory is initialized under the same conditions.
3236 * Tag-based KASAN modes only assign tags to normal non-executable
3237 * allocations, see __kasan_unpoison_vmalloc().
3238 */
3239 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3240 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3241 (gfp_mask & __GFP_SKIP_ZERO))
3242 kasan_flags |= KASAN_VMALLOC_INIT;
3243 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3244 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3245
3246 /*
3247 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3248 * flag. It means that vm_struct is not fully initialized.
3249 * Now, it is fully initialized, so remove this flag here.
3250 */
3251 clear_vm_uninitialized_flag(area);
3252
3253 size = PAGE_ALIGN(size);
3254 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3255 kmemleak_vmalloc(area, size, gfp_mask);
3256
3257 return area->addr;
3258
3259fail:
3260 if (shift > PAGE_SHIFT) {
3261 shift = PAGE_SHIFT;
3262 align = real_align;
3263 size = real_size;
3264 goto again;
3265 }
3266
3267 return NULL;
3268}
3269
3270/**
3271 * __vmalloc_node - allocate virtually contiguous memory
3272 * @size: allocation size
3273 * @align: desired alignment
3274 * @gfp_mask: flags for the page level allocator
3275 * @node: node to use for allocation or NUMA_NO_NODE
3276 * @caller: caller's return address
3277 *
3278 * Allocate enough pages to cover @size from the page level allocator with
3279 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3280 *
3281 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3282 * and __GFP_NOFAIL are not supported
3283 *
3284 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3285 * with mm people.
3286 *
3287 * Return: pointer to the allocated memory or %NULL on error
3288 */
3289void *__vmalloc_node(unsigned long size, unsigned long align,
3290 gfp_t gfp_mask, int node, const void *caller)
3291{
3292 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3293 gfp_mask, PAGE_KERNEL, 0, node, caller);
3294}
3295/*
3296 * This is only for performance analysis of vmalloc and stress purpose.
3297 * It is required by vmalloc test module, therefore do not use it other
3298 * than that.
3299 */
3300#ifdef CONFIG_TEST_VMALLOC_MODULE
3301EXPORT_SYMBOL_GPL(__vmalloc_node);
3302#endif
3303
3304void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3305{
3306 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3307 __builtin_return_address(0));
3308}
3309EXPORT_SYMBOL(__vmalloc);
3310
3311/**
3312 * vmalloc - allocate virtually contiguous memory
3313 * @size: allocation size
3314 *
3315 * Allocate enough pages to cover @size from the page level
3316 * allocator and map them into contiguous kernel virtual space.
3317 *
3318 * For tight control over page level allocator and protection flags
3319 * use __vmalloc() instead.
3320 *
3321 * Return: pointer to the allocated memory or %NULL on error
3322 */
3323void *vmalloc(unsigned long size)
3324{
3325 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3326 __builtin_return_address(0));
3327}
3328EXPORT_SYMBOL(vmalloc);
3329
3330/**
3331 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3332 * @size: allocation size
3333 * @gfp_mask: flags for the page level allocator
3334 *
3335 * Allocate enough pages to cover @size from the page level
3336 * allocator and map them into contiguous kernel virtual space.
3337 * If @size is greater than or equal to PMD_SIZE, allow using
3338 * huge pages for the memory
3339 *
3340 * Return: pointer to the allocated memory or %NULL on error
3341 */
3342void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3343{
3344 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3345 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3346 NUMA_NO_NODE, __builtin_return_address(0));
3347}
3348EXPORT_SYMBOL_GPL(vmalloc_huge);
3349
3350/**
3351 * vzalloc - allocate virtually contiguous memory with zero fill
3352 * @size: allocation size
3353 *
3354 * Allocate enough pages to cover @size from the page level
3355 * allocator and map them into contiguous kernel virtual space.
3356 * The memory allocated is set to zero.
3357 *
3358 * For tight control over page level allocator and protection flags
3359 * use __vmalloc() instead.
3360 *
3361 * Return: pointer to the allocated memory or %NULL on error
3362 */
3363void *vzalloc(unsigned long size)
3364{
3365 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3366 __builtin_return_address(0));
3367}
3368EXPORT_SYMBOL(vzalloc);
3369
3370/**
3371 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3372 * @size: allocation size
3373 *
3374 * The resulting memory area is zeroed so it can be mapped to userspace
3375 * without leaking data.
3376 *
3377 * Return: pointer to the allocated memory or %NULL on error
3378 */
3379void *vmalloc_user(unsigned long size)
3380{
3381 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3382 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3383 VM_USERMAP, NUMA_NO_NODE,
3384 __builtin_return_address(0));
3385}
3386EXPORT_SYMBOL(vmalloc_user);
3387
3388/**
3389 * vmalloc_node - allocate memory on a specific node
3390 * @size: allocation size
3391 * @node: numa node
3392 *
3393 * Allocate enough pages to cover @size from the page level
3394 * allocator and map them into contiguous kernel virtual space.
3395 *
3396 * For tight control over page level allocator and protection flags
3397 * use __vmalloc() instead.
3398 *
3399 * Return: pointer to the allocated memory or %NULL on error
3400 */
3401void *vmalloc_node(unsigned long size, int node)
3402{
3403 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3404 __builtin_return_address(0));
3405}
3406EXPORT_SYMBOL(vmalloc_node);
3407
3408/**
3409 * vzalloc_node - allocate memory on a specific node with zero fill
3410 * @size: allocation size
3411 * @node: numa node
3412 *
3413 * Allocate enough pages to cover @size from the page level
3414 * allocator and map them into contiguous kernel virtual space.
3415 * The memory allocated is set to zero.
3416 *
3417 * Return: pointer to the allocated memory or %NULL on error
3418 */
3419void *vzalloc_node(unsigned long size, int node)
3420{
3421 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3422 __builtin_return_address(0));
3423}
3424EXPORT_SYMBOL(vzalloc_node);
3425
3426#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3427#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3428#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3429#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3430#else
3431/*
3432 * 64b systems should always have either DMA or DMA32 zones. For others
3433 * GFP_DMA32 should do the right thing and use the normal zone.
3434 */
3435#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3436#endif
3437
3438/**
3439 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3440 * @size: allocation size
3441 *
3442 * Allocate enough 32bit PA addressable pages to cover @size from the
3443 * page level allocator and map them into contiguous kernel virtual space.
3444 *
3445 * Return: pointer to the allocated memory or %NULL on error
3446 */
3447void *vmalloc_32(unsigned long size)
3448{
3449 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3450 __builtin_return_address(0));
3451}
3452EXPORT_SYMBOL(vmalloc_32);
3453
3454/**
3455 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3456 * @size: allocation size
3457 *
3458 * The resulting memory area is 32bit addressable and zeroed so it can be
3459 * mapped to userspace without leaking data.
3460 *
3461 * Return: pointer to the allocated memory or %NULL on error
3462 */
3463void *vmalloc_32_user(unsigned long size)
3464{
3465 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3466 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3467 VM_USERMAP, NUMA_NO_NODE,
3468 __builtin_return_address(0));
3469}
3470EXPORT_SYMBOL(vmalloc_32_user);
3471
3472/*
3473 * small helper routine , copy contents to buf from addr.
3474 * If the page is not present, fill zero.
3475 */
3476
3477static int aligned_vread(char *buf, char *addr, unsigned long count)
3478{
3479 struct page *p;
3480 int copied = 0;
3481
3482 while (count) {
3483 unsigned long offset, length;
3484
3485 offset = offset_in_page(addr);
3486 length = PAGE_SIZE - offset;
3487 if (length > count)
3488 length = count;
3489 p = vmalloc_to_page(addr);
3490 /*
3491 * To do safe access to this _mapped_ area, we need
3492 * lock. But adding lock here means that we need to add
3493 * overhead of vmalloc()/vfree() calls for this _debug_
3494 * interface, rarely used. Instead of that, we'll use
3495 * kmap() and get small overhead in this access function.
3496 */
3497 if (p) {
3498 /* We can expect USER0 is not used -- see vread() */
3499 void *map = kmap_atomic(p);
3500 memcpy(buf, map + offset, length);
3501 kunmap_atomic(map);
3502 } else
3503 memset(buf, 0, length);
3504
3505 addr += length;
3506 buf += length;
3507 copied += length;
3508 count -= length;
3509 }
3510 return copied;
3511}
3512
3513/**
3514 * vread() - read vmalloc area in a safe way.
3515 * @buf: buffer for reading data
3516 * @addr: vm address.
3517 * @count: number of bytes to be read.
3518 *
3519 * This function checks that addr is a valid vmalloc'ed area, and
3520 * copy data from that area to a given buffer. If the given memory range
3521 * of [addr...addr+count) includes some valid address, data is copied to
3522 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3523 * IOREMAP area is treated as memory hole and no copy is done.
3524 *
3525 * If [addr...addr+count) doesn't includes any intersects with alive
3526 * vm_struct area, returns 0. @buf should be kernel's buffer.
3527 *
3528 * Note: In usual ops, vread() is never necessary because the caller
3529 * should know vmalloc() area is valid and can use memcpy().
3530 * This is for routines which have to access vmalloc area without
3531 * any information, as /proc/kcore.
3532 *
3533 * Return: number of bytes for which addr and buf should be increased
3534 * (same number as @count) or %0 if [addr...addr+count) doesn't
3535 * include any intersection with valid vmalloc area
3536 */
3537long vread(char *buf, char *addr, unsigned long count)
3538{
3539 struct vmap_area *va;
3540 struct vm_struct *vm;
3541 char *vaddr, *buf_start = buf;
3542 unsigned long buflen = count;
3543 unsigned long n;
3544
3545 addr = kasan_reset_tag(addr);
3546
3547 /* Don't allow overflow */
3548 if ((unsigned long) addr + count < count)
3549 count = -(unsigned long) addr;
3550
3551 spin_lock(&vmap_area_lock);
3552 va = find_vmap_area_exceed_addr((unsigned long)addr);
3553 if (!va)
3554 goto finished;
3555
3556 /* no intersects with alive vmap_area */
3557 if ((unsigned long)addr + count <= va->va_start)
3558 goto finished;
3559
3560 list_for_each_entry_from(va, &vmap_area_list, list) {
3561 if (!count)
3562 break;
3563
3564 if (!va->vm)
3565 continue;
3566
3567 vm = va->vm;
3568 vaddr = (char *) vm->addr;
3569 if (addr >= vaddr + get_vm_area_size(vm))
3570 continue;
3571 while (addr < vaddr) {
3572 if (count == 0)
3573 goto finished;
3574 *buf = '\0';
3575 buf++;
3576 addr++;
3577 count--;
3578 }
3579 n = vaddr + get_vm_area_size(vm) - addr;
3580 if (n > count)
3581 n = count;
3582 if (!(vm->flags & VM_IOREMAP))
3583 aligned_vread(buf, addr, n);
3584 else /* IOREMAP area is treated as memory hole */
3585 memset(buf, 0, n);
3586 buf += n;
3587 addr += n;
3588 count -= n;
3589 }
3590finished:
3591 spin_unlock(&vmap_area_lock);
3592
3593 if (buf == buf_start)
3594 return 0;
3595 /* zero-fill memory holes */
3596 if (buf != buf_start + buflen)
3597 memset(buf, 0, buflen - (buf - buf_start));
3598
3599 return buflen;
3600}
3601
3602/**
3603 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3604 * @vma: vma to cover
3605 * @uaddr: target user address to start at
3606 * @kaddr: virtual address of vmalloc kernel memory
3607 * @pgoff: offset from @kaddr to start at
3608 * @size: size of map area
3609 *
3610 * Returns: 0 for success, -Exxx on failure
3611 *
3612 * This function checks that @kaddr is a valid vmalloc'ed area,
3613 * and that it is big enough to cover the range starting at
3614 * @uaddr in @vma. Will return failure if that criteria isn't
3615 * met.
3616 *
3617 * Similar to remap_pfn_range() (see mm/memory.c)
3618 */
3619int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3620 void *kaddr, unsigned long pgoff,
3621 unsigned long size)
3622{
3623 struct vm_struct *area;
3624 unsigned long off;
3625 unsigned long end_index;
3626
3627 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3628 return -EINVAL;
3629
3630 size = PAGE_ALIGN(size);
3631
3632 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3633 return -EINVAL;
3634
3635 area = find_vm_area(kaddr);
3636 if (!area)
3637 return -EINVAL;
3638
3639 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3640 return -EINVAL;
3641
3642 if (check_add_overflow(size, off, &end_index) ||
3643 end_index > get_vm_area_size(area))
3644 return -EINVAL;
3645 kaddr += off;
3646
3647 do {
3648 struct page *page = vmalloc_to_page(kaddr);
3649 int ret;
3650
3651 ret = vm_insert_page(vma, uaddr, page);
3652 if (ret)
3653 return ret;
3654
3655 uaddr += PAGE_SIZE;
3656 kaddr += PAGE_SIZE;
3657 size -= PAGE_SIZE;
3658 } while (size > 0);
3659
3660 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3661
3662 return 0;
3663}
3664
3665/**
3666 * remap_vmalloc_range - map vmalloc pages to userspace
3667 * @vma: vma to cover (map full range of vma)
3668 * @addr: vmalloc memory
3669 * @pgoff: number of pages into addr before first page to map
3670 *
3671 * Returns: 0 for success, -Exxx on failure
3672 *
3673 * This function checks that addr is a valid vmalloc'ed area, and
3674 * that it is big enough to cover the vma. Will return failure if
3675 * that criteria isn't met.
3676 *
3677 * Similar to remap_pfn_range() (see mm/memory.c)
3678 */
3679int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3680 unsigned long pgoff)
3681{
3682 return remap_vmalloc_range_partial(vma, vma->vm_start,
3683 addr, pgoff,
3684 vma->vm_end - vma->vm_start);
3685}
3686EXPORT_SYMBOL(remap_vmalloc_range);
3687
3688void free_vm_area(struct vm_struct *area)
3689{
3690 struct vm_struct *ret;
3691 ret = remove_vm_area(area->addr);
3692 BUG_ON(ret != area);
3693 kfree(area);
3694}
3695EXPORT_SYMBOL_GPL(free_vm_area);
3696
3697#ifdef CONFIG_SMP
3698static struct vmap_area *node_to_va(struct rb_node *n)
3699{
3700 return rb_entry_safe(n, struct vmap_area, rb_node);
3701}
3702
3703/**
3704 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3705 * @addr: target address
3706 *
3707 * Returns: vmap_area if it is found. If there is no such area
3708 * the first highest(reverse order) vmap_area is returned
3709 * i.e. va->va_start < addr && va->va_end < addr or NULL
3710 * if there are no any areas before @addr.
3711 */
3712static struct vmap_area *
3713pvm_find_va_enclose_addr(unsigned long addr)
3714{
3715 struct vmap_area *va, *tmp;
3716 struct rb_node *n;
3717
3718 n = free_vmap_area_root.rb_node;
3719 va = NULL;
3720
3721 while (n) {
3722 tmp = rb_entry(n, struct vmap_area, rb_node);
3723 if (tmp->va_start <= addr) {
3724 va = tmp;
3725 if (tmp->va_end >= addr)
3726 break;
3727
3728 n = n->rb_right;
3729 } else {
3730 n = n->rb_left;
3731 }
3732 }
3733
3734 return va;
3735}
3736
3737/**
3738 * pvm_determine_end_from_reverse - find the highest aligned address
3739 * of free block below VMALLOC_END
3740 * @va:
3741 * in - the VA we start the search(reverse order);
3742 * out - the VA with the highest aligned end address.
3743 * @align: alignment for required highest address
3744 *
3745 * Returns: determined end address within vmap_area
3746 */
3747static unsigned long
3748pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3749{
3750 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3751 unsigned long addr;
3752
3753 if (likely(*va)) {
3754 list_for_each_entry_from_reverse((*va),
3755 &free_vmap_area_list, list) {
3756 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3757 if ((*va)->va_start < addr)
3758 return addr;
3759 }
3760 }
3761
3762 return 0;
3763}
3764
3765/**
3766 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3767 * @offsets: array containing offset of each area
3768 * @sizes: array containing size of each area
3769 * @nr_vms: the number of areas to allocate
3770 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3771 *
3772 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3773 * vm_structs on success, %NULL on failure
3774 *
3775 * Percpu allocator wants to use congruent vm areas so that it can
3776 * maintain the offsets among percpu areas. This function allocates
3777 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3778 * be scattered pretty far, distance between two areas easily going up
3779 * to gigabytes. To avoid interacting with regular vmallocs, these
3780 * areas are allocated from top.
3781 *
3782 * Despite its complicated look, this allocator is rather simple. It
3783 * does everything top-down and scans free blocks from the end looking
3784 * for matching base. While scanning, if any of the areas do not fit the
3785 * base address is pulled down to fit the area. Scanning is repeated till
3786 * all the areas fit and then all necessary data structures are inserted
3787 * and the result is returned.
3788 */
3789struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3790 const size_t *sizes, int nr_vms,
3791 size_t align)
3792{
3793 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3794 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3795 struct vmap_area **vas, *va;
3796 struct vm_struct **vms;
3797 int area, area2, last_area, term_area;
3798 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3799 bool purged = false;
3800
3801 /* verify parameters and allocate data structures */
3802 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3803 for (last_area = 0, area = 0; area < nr_vms; area++) {
3804 start = offsets[area];
3805 end = start + sizes[area];
3806
3807 /* is everything aligned properly? */
3808 BUG_ON(!IS_ALIGNED(offsets[area], align));
3809 BUG_ON(!IS_ALIGNED(sizes[area], align));
3810
3811 /* detect the area with the highest address */
3812 if (start > offsets[last_area])
3813 last_area = area;
3814
3815 for (area2 = area + 1; area2 < nr_vms; area2++) {
3816 unsigned long start2 = offsets[area2];
3817 unsigned long end2 = start2 + sizes[area2];
3818
3819 BUG_ON(start2 < end && start < end2);
3820 }
3821 }
3822 last_end = offsets[last_area] + sizes[last_area];
3823
3824 if (vmalloc_end - vmalloc_start < last_end) {
3825 WARN_ON(true);
3826 return NULL;
3827 }
3828
3829 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3830 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3831 if (!vas || !vms)
3832 goto err_free2;
3833
3834 for (area = 0; area < nr_vms; area++) {
3835 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3836 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3837 if (!vas[area] || !vms[area])
3838 goto err_free;
3839 }
3840retry:
3841 spin_lock(&free_vmap_area_lock);
3842
3843 /* start scanning - we scan from the top, begin with the last area */
3844 area = term_area = last_area;
3845 start = offsets[area];
3846 end = start + sizes[area];
3847
3848 va = pvm_find_va_enclose_addr(vmalloc_end);
3849 base = pvm_determine_end_from_reverse(&va, align) - end;
3850
3851 while (true) {
3852 /*
3853 * base might have underflowed, add last_end before
3854 * comparing.
3855 */
3856 if (base + last_end < vmalloc_start + last_end)
3857 goto overflow;
3858
3859 /*
3860 * Fitting base has not been found.
3861 */
3862 if (va == NULL)
3863 goto overflow;
3864
3865 /*
3866 * If required width exceeds current VA block, move
3867 * base downwards and then recheck.
3868 */
3869 if (base + end > va->va_end) {
3870 base = pvm_determine_end_from_reverse(&va, align) - end;
3871 term_area = area;
3872 continue;
3873 }
3874
3875 /*
3876 * If this VA does not fit, move base downwards and recheck.
3877 */
3878 if (base + start < va->va_start) {
3879 va = node_to_va(rb_prev(&va->rb_node));
3880 base = pvm_determine_end_from_reverse(&va, align) - end;
3881 term_area = area;
3882 continue;
3883 }
3884
3885 /*
3886 * This area fits, move on to the previous one. If
3887 * the previous one is the terminal one, we're done.
3888 */
3889 area = (area + nr_vms - 1) % nr_vms;
3890 if (area == term_area)
3891 break;
3892
3893 start = offsets[area];
3894 end = start + sizes[area];
3895 va = pvm_find_va_enclose_addr(base + end);
3896 }
3897
3898 /* we've found a fitting base, insert all va's */
3899 for (area = 0; area < nr_vms; area++) {
3900 int ret;
3901
3902 start = base + offsets[area];
3903 size = sizes[area];
3904
3905 va = pvm_find_va_enclose_addr(start);
3906 if (WARN_ON_ONCE(va == NULL))
3907 /* It is a BUG(), but trigger recovery instead. */
3908 goto recovery;
3909
3910 ret = adjust_va_to_fit_type(&free_vmap_area_root,
3911 &free_vmap_area_list,
3912 va, start, size);
3913 if (WARN_ON_ONCE(unlikely(ret)))
3914 /* It is a BUG(), but trigger recovery instead. */
3915 goto recovery;
3916
3917 /* Allocated area. */
3918 va = vas[area];
3919 va->va_start = start;
3920 va->va_end = start + size;
3921 }
3922
3923 spin_unlock(&free_vmap_area_lock);
3924
3925 /* populate the kasan shadow space */
3926 for (area = 0; area < nr_vms; area++) {
3927 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3928 goto err_free_shadow;
3929 }
3930
3931 /* insert all vm's */
3932 spin_lock(&vmap_area_lock);
3933 for (area = 0; area < nr_vms; area++) {
3934 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3935
3936 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3937 pcpu_get_vm_areas);
3938 }
3939 spin_unlock(&vmap_area_lock);
3940
3941 /*
3942 * Mark allocated areas as accessible. Do it now as a best-effort
3943 * approach, as they can be mapped outside of vmalloc code.
3944 * With hardware tag-based KASAN, marking is skipped for
3945 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3946 */
3947 for (area = 0; area < nr_vms; area++)
3948 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3949 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3950
3951 kfree(vas);
3952 return vms;
3953
3954recovery:
3955 /*
3956 * Remove previously allocated areas. There is no
3957 * need in removing these areas from the busy tree,
3958 * because they are inserted only on the final step
3959 * and when pcpu_get_vm_areas() is success.
3960 */
3961 while (area--) {
3962 orig_start = vas[area]->va_start;
3963 orig_end = vas[area]->va_end;
3964 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3965 &free_vmap_area_list);
3966 if (va)
3967 kasan_release_vmalloc(orig_start, orig_end,
3968 va->va_start, va->va_end);
3969 vas[area] = NULL;
3970 }
3971
3972overflow:
3973 spin_unlock(&free_vmap_area_lock);
3974 if (!purged) {
3975 purge_vmap_area_lazy();
3976 purged = true;
3977
3978 /* Before "retry", check if we recover. */
3979 for (area = 0; area < nr_vms; area++) {
3980 if (vas[area])
3981 continue;
3982
3983 vas[area] = kmem_cache_zalloc(
3984 vmap_area_cachep, GFP_KERNEL);
3985 if (!vas[area])
3986 goto err_free;
3987 }
3988
3989 goto retry;
3990 }
3991
3992err_free:
3993 for (area = 0; area < nr_vms; area++) {
3994 if (vas[area])
3995 kmem_cache_free(vmap_area_cachep, vas[area]);
3996
3997 kfree(vms[area]);
3998 }
3999err_free2:
4000 kfree(vas);
4001 kfree(vms);
4002 return NULL;
4003
4004err_free_shadow:
4005 spin_lock(&free_vmap_area_lock);
4006 /*
4007 * We release all the vmalloc shadows, even the ones for regions that
4008 * hadn't been successfully added. This relies on kasan_release_vmalloc
4009 * being able to tolerate this case.
4010 */
4011 for (area = 0; area < nr_vms; area++) {
4012 orig_start = vas[area]->va_start;
4013 orig_end = vas[area]->va_end;
4014 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4015 &free_vmap_area_list);
4016 if (va)
4017 kasan_release_vmalloc(orig_start, orig_end,
4018 va->va_start, va->va_end);
4019 vas[area] = NULL;
4020 kfree(vms[area]);
4021 }
4022 spin_unlock(&free_vmap_area_lock);
4023 kfree(vas);
4024 kfree(vms);
4025 return NULL;
4026}
4027
4028/**
4029 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4030 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4031 * @nr_vms: the number of allocated areas
4032 *
4033 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4034 */
4035void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4036{
4037 int i;
4038
4039 for (i = 0; i < nr_vms; i++)
4040 free_vm_area(vms[i]);
4041 kfree(vms);
4042}
4043#endif /* CONFIG_SMP */
4044
4045#ifdef CONFIG_PRINTK
4046bool vmalloc_dump_obj(void *object)
4047{
4048 struct vm_struct *vm;
4049 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4050
4051 vm = find_vm_area(objp);
4052 if (!vm)
4053 return false;
4054 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4055 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4056 return true;
4057}
4058#endif
4059
4060#ifdef CONFIG_PROC_FS
4061static void *s_start(struct seq_file *m, loff_t *pos)
4062 __acquires(&vmap_purge_lock)
4063 __acquires(&vmap_area_lock)
4064{
4065 mutex_lock(&vmap_purge_lock);
4066 spin_lock(&vmap_area_lock);
4067
4068 return seq_list_start(&vmap_area_list, *pos);
4069}
4070
4071static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4072{
4073 return seq_list_next(p, &vmap_area_list, pos);
4074}
4075
4076static void s_stop(struct seq_file *m, void *p)
4077 __releases(&vmap_area_lock)
4078 __releases(&vmap_purge_lock)
4079{
4080 spin_unlock(&vmap_area_lock);
4081 mutex_unlock(&vmap_purge_lock);
4082}
4083
4084static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4085{
4086 if (IS_ENABLED(CONFIG_NUMA)) {
4087 unsigned int nr, *counters = m->private;
4088 unsigned int step = 1U << vm_area_page_order(v);
4089
4090 if (!counters)
4091 return;
4092
4093 if (v->flags & VM_UNINITIALIZED)
4094 return;
4095 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4096 smp_rmb();
4097
4098 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4099
4100 for (nr = 0; nr < v->nr_pages; nr += step)
4101 counters[page_to_nid(v->pages[nr])] += step;
4102 for_each_node_state(nr, N_HIGH_MEMORY)
4103 if (counters[nr])
4104 seq_printf(m, " N%u=%u", nr, counters[nr]);
4105 }
4106}
4107
4108static void show_purge_info(struct seq_file *m)
4109{
4110 struct vmap_area *va;
4111
4112 spin_lock(&purge_vmap_area_lock);
4113 list_for_each_entry(va, &purge_vmap_area_list, list) {
4114 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4115 (void *)va->va_start, (void *)va->va_end,
4116 va->va_end - va->va_start);
4117 }
4118 spin_unlock(&purge_vmap_area_lock);
4119}
4120
4121static int s_show(struct seq_file *m, void *p)
4122{
4123 struct vmap_area *va;
4124 struct vm_struct *v;
4125
4126 va = list_entry(p, struct vmap_area, list);
4127
4128 /*
4129 * s_show can encounter race with remove_vm_area, !vm on behalf
4130 * of vmap area is being tear down or vm_map_ram allocation.
4131 */
4132 if (!va->vm) {
4133 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4134 (void *)va->va_start, (void *)va->va_end,
4135 va->va_end - va->va_start);
4136
4137 goto final;
4138 }
4139
4140 v = va->vm;
4141
4142 seq_printf(m, "0x%pK-0x%pK %7ld",
4143 v->addr, v->addr + v->size, v->size);
4144
4145 if (v->caller)
4146 seq_printf(m, " %pS", v->caller);
4147
4148 if (v->nr_pages)
4149 seq_printf(m, " pages=%d", v->nr_pages);
4150
4151 if (v->phys_addr)
4152 seq_printf(m, " phys=%pa", &v->phys_addr);
4153
4154 if (v->flags & VM_IOREMAP)
4155 seq_puts(m, " ioremap");
4156
4157 if (v->flags & VM_ALLOC)
4158 seq_puts(m, " vmalloc");
4159
4160 if (v->flags & VM_MAP)
4161 seq_puts(m, " vmap");
4162
4163 if (v->flags & VM_USERMAP)
4164 seq_puts(m, " user");
4165
4166 if (v->flags & VM_DMA_COHERENT)
4167 seq_puts(m, " dma-coherent");
4168
4169 if (is_vmalloc_addr(v->pages))
4170 seq_puts(m, " vpages");
4171
4172 show_numa_info(m, v);
4173 seq_putc(m, '\n');
4174
4175 /*
4176 * As a final step, dump "unpurged" areas.
4177 */
4178final:
4179 if (list_is_last(&va->list, &vmap_area_list))
4180 show_purge_info(m);
4181
4182 return 0;
4183}
4184
4185static const struct seq_operations vmalloc_op = {
4186 .start = s_start,
4187 .next = s_next,
4188 .stop = s_stop,
4189 .show = s_show,
4190};
4191
4192static int __init proc_vmalloc_init(void)
4193{
4194 if (IS_ENABLED(CONFIG_NUMA))
4195 proc_create_seq_private("vmallocinfo", 0400, NULL,
4196 &vmalloc_op,
4197 nr_node_ids * sizeof(unsigned int), NULL);
4198 else
4199 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4200 return 0;
4201}
4202module_init(proc_vmalloc_init);
4203
4204#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched/signal.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/set_memory.h>
22#include <linux/debugobjects.h>
23#include <linux/kallsyms.h>
24#include <linux/list.h>
25#include <linux/notifier.h>
26#include <linux/rbtree.h>
27#include <linux/xarray.h>
28#include <linux/io.h>
29#include <linux/rcupdate.h>
30#include <linux/pfn.h>
31#include <linux/kmemleak.h>
32#include <linux/atomic.h>
33#include <linux/compiler.h>
34#include <linux/llist.h>
35#include <linux/bitops.h>
36#include <linux/rbtree_augmented.h>
37#include <linux/overflow.h>
38#include <linux/pgtable.h>
39#include <linux/uaccess.h>
40#include <linux/hugetlb.h>
41#include <asm/tlbflush.h>
42#include <asm/shmparam.h>
43
44#include "internal.h"
45#include "pgalloc-track.h"
46
47#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
48static bool __ro_after_init vmap_allow_huge = true;
49
50static int __init set_nohugevmalloc(char *str)
51{
52 vmap_allow_huge = false;
53 return 0;
54}
55early_param("nohugevmalloc", set_nohugevmalloc);
56#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
57static const bool vmap_allow_huge = false;
58#endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
59
60bool is_vmalloc_addr(const void *x)
61{
62 unsigned long addr = (unsigned long)x;
63
64 return addr >= VMALLOC_START && addr < VMALLOC_END;
65}
66EXPORT_SYMBOL(is_vmalloc_addr);
67
68struct vfree_deferred {
69 struct llist_head list;
70 struct work_struct wq;
71};
72static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
73
74static void __vunmap(const void *, int);
75
76static void free_work(struct work_struct *w)
77{
78 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
79 struct llist_node *t, *llnode;
80
81 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
82 __vunmap((void *)llnode, 1);
83}
84
85/*** Page table manipulation functions ***/
86static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
87 phys_addr_t phys_addr, pgprot_t prot,
88 unsigned int max_page_shift, pgtbl_mod_mask *mask)
89{
90 pte_t *pte;
91 u64 pfn;
92 unsigned long size = PAGE_SIZE;
93
94 pfn = phys_addr >> PAGE_SHIFT;
95 pte = pte_alloc_kernel_track(pmd, addr, mask);
96 if (!pte)
97 return -ENOMEM;
98 do {
99 BUG_ON(!pte_none(*pte));
100
101#ifdef CONFIG_HUGETLB_PAGE
102 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
103 if (size != PAGE_SIZE) {
104 pte_t entry = pfn_pte(pfn, prot);
105
106 entry = pte_mkhuge(entry);
107 entry = arch_make_huge_pte(entry, ilog2(size), 0);
108 set_huge_pte_at(&init_mm, addr, pte, entry);
109 pfn += PFN_DOWN(size);
110 continue;
111 }
112#endif
113 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
114 pfn++;
115 } while (pte += PFN_DOWN(size), addr += size, addr != end);
116 *mask |= PGTBL_PTE_MODIFIED;
117 return 0;
118}
119
120static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
121 phys_addr_t phys_addr, pgprot_t prot,
122 unsigned int max_page_shift)
123{
124 if (max_page_shift < PMD_SHIFT)
125 return 0;
126
127 if (!arch_vmap_pmd_supported(prot))
128 return 0;
129
130 if ((end - addr) != PMD_SIZE)
131 return 0;
132
133 if (!IS_ALIGNED(addr, PMD_SIZE))
134 return 0;
135
136 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
137 return 0;
138
139 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
140 return 0;
141
142 return pmd_set_huge(pmd, phys_addr, prot);
143}
144
145static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
146 phys_addr_t phys_addr, pgprot_t prot,
147 unsigned int max_page_shift, pgtbl_mod_mask *mask)
148{
149 pmd_t *pmd;
150 unsigned long next;
151
152 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
153 if (!pmd)
154 return -ENOMEM;
155 do {
156 next = pmd_addr_end(addr, end);
157
158 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
159 max_page_shift)) {
160 *mask |= PGTBL_PMD_MODIFIED;
161 continue;
162 }
163
164 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
165 return -ENOMEM;
166 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
167 return 0;
168}
169
170static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
171 phys_addr_t phys_addr, pgprot_t prot,
172 unsigned int max_page_shift)
173{
174 if (max_page_shift < PUD_SHIFT)
175 return 0;
176
177 if (!arch_vmap_pud_supported(prot))
178 return 0;
179
180 if ((end - addr) != PUD_SIZE)
181 return 0;
182
183 if (!IS_ALIGNED(addr, PUD_SIZE))
184 return 0;
185
186 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
187 return 0;
188
189 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
190 return 0;
191
192 return pud_set_huge(pud, phys_addr, prot);
193}
194
195static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
196 phys_addr_t phys_addr, pgprot_t prot,
197 unsigned int max_page_shift, pgtbl_mod_mask *mask)
198{
199 pud_t *pud;
200 unsigned long next;
201
202 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
203 if (!pud)
204 return -ENOMEM;
205 do {
206 next = pud_addr_end(addr, end);
207
208 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
209 max_page_shift)) {
210 *mask |= PGTBL_PUD_MODIFIED;
211 continue;
212 }
213
214 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
215 max_page_shift, mask))
216 return -ENOMEM;
217 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
218 return 0;
219}
220
221static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
222 phys_addr_t phys_addr, pgprot_t prot,
223 unsigned int max_page_shift)
224{
225 if (max_page_shift < P4D_SHIFT)
226 return 0;
227
228 if (!arch_vmap_p4d_supported(prot))
229 return 0;
230
231 if ((end - addr) != P4D_SIZE)
232 return 0;
233
234 if (!IS_ALIGNED(addr, P4D_SIZE))
235 return 0;
236
237 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
238 return 0;
239
240 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
241 return 0;
242
243 return p4d_set_huge(p4d, phys_addr, prot);
244}
245
246static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
247 phys_addr_t phys_addr, pgprot_t prot,
248 unsigned int max_page_shift, pgtbl_mod_mask *mask)
249{
250 p4d_t *p4d;
251 unsigned long next;
252
253 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
254 if (!p4d)
255 return -ENOMEM;
256 do {
257 next = p4d_addr_end(addr, end);
258
259 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
260 max_page_shift)) {
261 *mask |= PGTBL_P4D_MODIFIED;
262 continue;
263 }
264
265 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
266 max_page_shift, mask))
267 return -ENOMEM;
268 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
269 return 0;
270}
271
272static int vmap_range_noflush(unsigned long addr, unsigned long end,
273 phys_addr_t phys_addr, pgprot_t prot,
274 unsigned int max_page_shift)
275{
276 pgd_t *pgd;
277 unsigned long start;
278 unsigned long next;
279 int err;
280 pgtbl_mod_mask mask = 0;
281
282 might_sleep();
283 BUG_ON(addr >= end);
284
285 start = addr;
286 pgd = pgd_offset_k(addr);
287 do {
288 next = pgd_addr_end(addr, end);
289 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
290 max_page_shift, &mask);
291 if (err)
292 break;
293 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
294
295 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
296 arch_sync_kernel_mappings(start, end);
297
298 return err;
299}
300
301int vmap_range(unsigned long addr, unsigned long end,
302 phys_addr_t phys_addr, pgprot_t prot,
303 unsigned int max_page_shift)
304{
305 int err;
306
307 err = vmap_range_noflush(addr, end, phys_addr, prot, max_page_shift);
308 flush_cache_vmap(addr, end);
309
310 return err;
311}
312
313static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
314 pgtbl_mod_mask *mask)
315{
316 pte_t *pte;
317
318 pte = pte_offset_kernel(pmd, addr);
319 do {
320 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
321 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
322 } while (pte++, addr += PAGE_SIZE, addr != end);
323 *mask |= PGTBL_PTE_MODIFIED;
324}
325
326static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
327 pgtbl_mod_mask *mask)
328{
329 pmd_t *pmd;
330 unsigned long next;
331 int cleared;
332
333 pmd = pmd_offset(pud, addr);
334 do {
335 next = pmd_addr_end(addr, end);
336
337 cleared = pmd_clear_huge(pmd);
338 if (cleared || pmd_bad(*pmd))
339 *mask |= PGTBL_PMD_MODIFIED;
340
341 if (cleared)
342 continue;
343 if (pmd_none_or_clear_bad(pmd))
344 continue;
345 vunmap_pte_range(pmd, addr, next, mask);
346
347 cond_resched();
348 } while (pmd++, addr = next, addr != end);
349}
350
351static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
352 pgtbl_mod_mask *mask)
353{
354 pud_t *pud;
355 unsigned long next;
356 int cleared;
357
358 pud = pud_offset(p4d, addr);
359 do {
360 next = pud_addr_end(addr, end);
361
362 cleared = pud_clear_huge(pud);
363 if (cleared || pud_bad(*pud))
364 *mask |= PGTBL_PUD_MODIFIED;
365
366 if (cleared)
367 continue;
368 if (pud_none_or_clear_bad(pud))
369 continue;
370 vunmap_pmd_range(pud, addr, next, mask);
371 } while (pud++, addr = next, addr != end);
372}
373
374static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
375 pgtbl_mod_mask *mask)
376{
377 p4d_t *p4d;
378 unsigned long next;
379 int cleared;
380
381 p4d = p4d_offset(pgd, addr);
382 do {
383 next = p4d_addr_end(addr, end);
384
385 cleared = p4d_clear_huge(p4d);
386 if (cleared || p4d_bad(*p4d))
387 *mask |= PGTBL_P4D_MODIFIED;
388
389 if (cleared)
390 continue;
391 if (p4d_none_or_clear_bad(p4d))
392 continue;
393 vunmap_pud_range(p4d, addr, next, mask);
394 } while (p4d++, addr = next, addr != end);
395}
396
397/*
398 * vunmap_range_noflush is similar to vunmap_range, but does not
399 * flush caches or TLBs.
400 *
401 * The caller is responsible for calling flush_cache_vmap() before calling
402 * this function, and flush_tlb_kernel_range after it has returned
403 * successfully (and before the addresses are expected to cause a page fault
404 * or be re-mapped for something else, if TLB flushes are being delayed or
405 * coalesced).
406 *
407 * This is an internal function only. Do not use outside mm/.
408 */
409void vunmap_range_noflush(unsigned long start, unsigned long end)
410{
411 unsigned long next;
412 pgd_t *pgd;
413 unsigned long addr = start;
414 pgtbl_mod_mask mask = 0;
415
416 BUG_ON(addr >= end);
417 pgd = pgd_offset_k(addr);
418 do {
419 next = pgd_addr_end(addr, end);
420 if (pgd_bad(*pgd))
421 mask |= PGTBL_PGD_MODIFIED;
422 if (pgd_none_or_clear_bad(pgd))
423 continue;
424 vunmap_p4d_range(pgd, addr, next, &mask);
425 } while (pgd++, addr = next, addr != end);
426
427 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
428 arch_sync_kernel_mappings(start, end);
429}
430
431/**
432 * vunmap_range - unmap kernel virtual addresses
433 * @addr: start of the VM area to unmap
434 * @end: end of the VM area to unmap (non-inclusive)
435 *
436 * Clears any present PTEs in the virtual address range, flushes TLBs and
437 * caches. Any subsequent access to the address before it has been re-mapped
438 * is a kernel bug.
439 */
440void vunmap_range(unsigned long addr, unsigned long end)
441{
442 flush_cache_vunmap(addr, end);
443 vunmap_range_noflush(addr, end);
444 flush_tlb_kernel_range(addr, end);
445}
446
447static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
448 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
449 pgtbl_mod_mask *mask)
450{
451 pte_t *pte;
452
453 /*
454 * nr is a running index into the array which helps higher level
455 * callers keep track of where we're up to.
456 */
457
458 pte = pte_alloc_kernel_track(pmd, addr, mask);
459 if (!pte)
460 return -ENOMEM;
461 do {
462 struct page *page = pages[*nr];
463
464 if (WARN_ON(!pte_none(*pte)))
465 return -EBUSY;
466 if (WARN_ON(!page))
467 return -ENOMEM;
468 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
469 (*nr)++;
470 } while (pte++, addr += PAGE_SIZE, addr != end);
471 *mask |= PGTBL_PTE_MODIFIED;
472 return 0;
473}
474
475static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
476 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
477 pgtbl_mod_mask *mask)
478{
479 pmd_t *pmd;
480 unsigned long next;
481
482 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
483 if (!pmd)
484 return -ENOMEM;
485 do {
486 next = pmd_addr_end(addr, end);
487 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
488 return -ENOMEM;
489 } while (pmd++, addr = next, addr != end);
490 return 0;
491}
492
493static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
494 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
495 pgtbl_mod_mask *mask)
496{
497 pud_t *pud;
498 unsigned long next;
499
500 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
501 if (!pud)
502 return -ENOMEM;
503 do {
504 next = pud_addr_end(addr, end);
505 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
506 return -ENOMEM;
507 } while (pud++, addr = next, addr != end);
508 return 0;
509}
510
511static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
512 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
513 pgtbl_mod_mask *mask)
514{
515 p4d_t *p4d;
516 unsigned long next;
517
518 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
519 if (!p4d)
520 return -ENOMEM;
521 do {
522 next = p4d_addr_end(addr, end);
523 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
524 return -ENOMEM;
525 } while (p4d++, addr = next, addr != end);
526 return 0;
527}
528
529static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
530 pgprot_t prot, struct page **pages)
531{
532 unsigned long start = addr;
533 pgd_t *pgd;
534 unsigned long next;
535 int err = 0;
536 int nr = 0;
537 pgtbl_mod_mask mask = 0;
538
539 BUG_ON(addr >= end);
540 pgd = pgd_offset_k(addr);
541 do {
542 next = pgd_addr_end(addr, end);
543 if (pgd_bad(*pgd))
544 mask |= PGTBL_PGD_MODIFIED;
545 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
546 if (err)
547 return err;
548 } while (pgd++, addr = next, addr != end);
549
550 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
551 arch_sync_kernel_mappings(start, end);
552
553 return 0;
554}
555
556/*
557 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
558 * flush caches.
559 *
560 * The caller is responsible for calling flush_cache_vmap() after this
561 * function returns successfully and before the addresses are accessed.
562 *
563 * This is an internal function only. Do not use outside mm/.
564 */
565int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
566 pgprot_t prot, struct page **pages, unsigned int page_shift)
567{
568 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
569
570 WARN_ON(page_shift < PAGE_SHIFT);
571
572 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
573 page_shift == PAGE_SHIFT)
574 return vmap_small_pages_range_noflush(addr, end, prot, pages);
575
576 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
577 int err;
578
579 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
580 __pa(page_address(pages[i])), prot,
581 page_shift);
582 if (err)
583 return err;
584
585 addr += 1UL << page_shift;
586 }
587
588 return 0;
589}
590
591/**
592 * vmap_pages_range - map pages to a kernel virtual address
593 * @addr: start of the VM area to map
594 * @end: end of the VM area to map (non-inclusive)
595 * @prot: page protection flags to use
596 * @pages: pages to map (always PAGE_SIZE pages)
597 * @page_shift: maximum shift that the pages may be mapped with, @pages must
598 * be aligned and contiguous up to at least this shift.
599 *
600 * RETURNS:
601 * 0 on success, -errno on failure.
602 */
603static int vmap_pages_range(unsigned long addr, unsigned long end,
604 pgprot_t prot, struct page **pages, unsigned int page_shift)
605{
606 int err;
607
608 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
609 flush_cache_vmap(addr, end);
610 return err;
611}
612
613int is_vmalloc_or_module_addr(const void *x)
614{
615 /*
616 * ARM, x86-64 and sparc64 put modules in a special place,
617 * and fall back on vmalloc() if that fails. Others
618 * just put it in the vmalloc space.
619 */
620#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
621 unsigned long addr = (unsigned long)x;
622 if (addr >= MODULES_VADDR && addr < MODULES_END)
623 return 1;
624#endif
625 return is_vmalloc_addr(x);
626}
627
628/*
629 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
630 * return the tail page that corresponds to the base page address, which
631 * matches small vmap mappings.
632 */
633struct page *vmalloc_to_page(const void *vmalloc_addr)
634{
635 unsigned long addr = (unsigned long) vmalloc_addr;
636 struct page *page = NULL;
637 pgd_t *pgd = pgd_offset_k(addr);
638 p4d_t *p4d;
639 pud_t *pud;
640 pmd_t *pmd;
641 pte_t *ptep, pte;
642
643 /*
644 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
645 * architectures that do not vmalloc module space
646 */
647 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
648
649 if (pgd_none(*pgd))
650 return NULL;
651 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
652 return NULL; /* XXX: no allowance for huge pgd */
653 if (WARN_ON_ONCE(pgd_bad(*pgd)))
654 return NULL;
655
656 p4d = p4d_offset(pgd, addr);
657 if (p4d_none(*p4d))
658 return NULL;
659 if (p4d_leaf(*p4d))
660 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
661 if (WARN_ON_ONCE(p4d_bad(*p4d)))
662 return NULL;
663
664 pud = pud_offset(p4d, addr);
665 if (pud_none(*pud))
666 return NULL;
667 if (pud_leaf(*pud))
668 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
669 if (WARN_ON_ONCE(pud_bad(*pud)))
670 return NULL;
671
672 pmd = pmd_offset(pud, addr);
673 if (pmd_none(*pmd))
674 return NULL;
675 if (pmd_leaf(*pmd))
676 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
677 if (WARN_ON_ONCE(pmd_bad(*pmd)))
678 return NULL;
679
680 ptep = pte_offset_map(pmd, addr);
681 pte = *ptep;
682 if (pte_present(pte))
683 page = pte_page(pte);
684 pte_unmap(ptep);
685
686 return page;
687}
688EXPORT_SYMBOL(vmalloc_to_page);
689
690/*
691 * Map a vmalloc()-space virtual address to the physical page frame number.
692 */
693unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
694{
695 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
696}
697EXPORT_SYMBOL(vmalloc_to_pfn);
698
699
700/*** Global kva allocator ***/
701
702#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
703#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
704
705
706static DEFINE_SPINLOCK(vmap_area_lock);
707static DEFINE_SPINLOCK(free_vmap_area_lock);
708/* Export for kexec only */
709LIST_HEAD(vmap_area_list);
710static struct rb_root vmap_area_root = RB_ROOT;
711static bool vmap_initialized __read_mostly;
712
713static struct rb_root purge_vmap_area_root = RB_ROOT;
714static LIST_HEAD(purge_vmap_area_list);
715static DEFINE_SPINLOCK(purge_vmap_area_lock);
716
717/*
718 * This kmem_cache is used for vmap_area objects. Instead of
719 * allocating from slab we reuse an object from this cache to
720 * make things faster. Especially in "no edge" splitting of
721 * free block.
722 */
723static struct kmem_cache *vmap_area_cachep;
724
725/*
726 * This linked list is used in pair with free_vmap_area_root.
727 * It gives O(1) access to prev/next to perform fast coalescing.
728 */
729static LIST_HEAD(free_vmap_area_list);
730
731/*
732 * This augment red-black tree represents the free vmap space.
733 * All vmap_area objects in this tree are sorted by va->va_start
734 * address. It is used for allocation and merging when a vmap
735 * object is released.
736 *
737 * Each vmap_area node contains a maximum available free block
738 * of its sub-tree, right or left. Therefore it is possible to
739 * find a lowest match of free area.
740 */
741static struct rb_root free_vmap_area_root = RB_ROOT;
742
743/*
744 * Preload a CPU with one object for "no edge" split case. The
745 * aim is to get rid of allocations from the atomic context, thus
746 * to use more permissive allocation masks.
747 */
748static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
749
750static __always_inline unsigned long
751va_size(struct vmap_area *va)
752{
753 return (va->va_end - va->va_start);
754}
755
756static __always_inline unsigned long
757get_subtree_max_size(struct rb_node *node)
758{
759 struct vmap_area *va;
760
761 va = rb_entry_safe(node, struct vmap_area, rb_node);
762 return va ? va->subtree_max_size : 0;
763}
764
765/*
766 * Gets called when remove the node and rotate.
767 */
768static __always_inline unsigned long
769compute_subtree_max_size(struct vmap_area *va)
770{
771 return max3(va_size(va),
772 get_subtree_max_size(va->rb_node.rb_left),
773 get_subtree_max_size(va->rb_node.rb_right));
774}
775
776RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
777 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
778
779static void purge_vmap_area_lazy(void);
780static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
781static unsigned long lazy_max_pages(void);
782
783static atomic_long_t nr_vmalloc_pages;
784
785unsigned long vmalloc_nr_pages(void)
786{
787 return atomic_long_read(&nr_vmalloc_pages);
788}
789
790static struct vmap_area *__find_vmap_area(unsigned long addr)
791{
792 struct rb_node *n = vmap_area_root.rb_node;
793
794 while (n) {
795 struct vmap_area *va;
796
797 va = rb_entry(n, struct vmap_area, rb_node);
798 if (addr < va->va_start)
799 n = n->rb_left;
800 else if (addr >= va->va_end)
801 n = n->rb_right;
802 else
803 return va;
804 }
805
806 return NULL;
807}
808
809/*
810 * This function returns back addresses of parent node
811 * and its left or right link for further processing.
812 *
813 * Otherwise NULL is returned. In that case all further
814 * steps regarding inserting of conflicting overlap range
815 * have to be declined and actually considered as a bug.
816 */
817static __always_inline struct rb_node **
818find_va_links(struct vmap_area *va,
819 struct rb_root *root, struct rb_node *from,
820 struct rb_node **parent)
821{
822 struct vmap_area *tmp_va;
823 struct rb_node **link;
824
825 if (root) {
826 link = &root->rb_node;
827 if (unlikely(!*link)) {
828 *parent = NULL;
829 return link;
830 }
831 } else {
832 link = &from;
833 }
834
835 /*
836 * Go to the bottom of the tree. When we hit the last point
837 * we end up with parent rb_node and correct direction, i name
838 * it link, where the new va->rb_node will be attached to.
839 */
840 do {
841 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
842
843 /*
844 * During the traversal we also do some sanity check.
845 * Trigger the BUG() if there are sides(left/right)
846 * or full overlaps.
847 */
848 if (va->va_start < tmp_va->va_end &&
849 va->va_end <= tmp_va->va_start)
850 link = &(*link)->rb_left;
851 else if (va->va_end > tmp_va->va_start &&
852 va->va_start >= tmp_va->va_end)
853 link = &(*link)->rb_right;
854 else {
855 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
856 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
857
858 return NULL;
859 }
860 } while (*link);
861
862 *parent = &tmp_va->rb_node;
863 return link;
864}
865
866static __always_inline struct list_head *
867get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
868{
869 struct list_head *list;
870
871 if (unlikely(!parent))
872 /*
873 * The red-black tree where we try to find VA neighbors
874 * before merging or inserting is empty, i.e. it means
875 * there is no free vmap space. Normally it does not
876 * happen but we handle this case anyway.
877 */
878 return NULL;
879
880 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
881 return (&parent->rb_right == link ? list->next : list);
882}
883
884static __always_inline void
885link_va(struct vmap_area *va, struct rb_root *root,
886 struct rb_node *parent, struct rb_node **link, struct list_head *head)
887{
888 /*
889 * VA is still not in the list, but we can
890 * identify its future previous list_head node.
891 */
892 if (likely(parent)) {
893 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
894 if (&parent->rb_right != link)
895 head = head->prev;
896 }
897
898 /* Insert to the rb-tree */
899 rb_link_node(&va->rb_node, parent, link);
900 if (root == &free_vmap_area_root) {
901 /*
902 * Some explanation here. Just perform simple insertion
903 * to the tree. We do not set va->subtree_max_size to
904 * its current size before calling rb_insert_augmented().
905 * It is because of we populate the tree from the bottom
906 * to parent levels when the node _is_ in the tree.
907 *
908 * Therefore we set subtree_max_size to zero after insertion,
909 * to let __augment_tree_propagate_from() puts everything to
910 * the correct order later on.
911 */
912 rb_insert_augmented(&va->rb_node,
913 root, &free_vmap_area_rb_augment_cb);
914 va->subtree_max_size = 0;
915 } else {
916 rb_insert_color(&va->rb_node, root);
917 }
918
919 /* Address-sort this list */
920 list_add(&va->list, head);
921}
922
923static __always_inline void
924unlink_va(struct vmap_area *va, struct rb_root *root)
925{
926 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
927 return;
928
929 if (root == &free_vmap_area_root)
930 rb_erase_augmented(&va->rb_node,
931 root, &free_vmap_area_rb_augment_cb);
932 else
933 rb_erase(&va->rb_node, root);
934
935 list_del(&va->list);
936 RB_CLEAR_NODE(&va->rb_node);
937}
938
939#if DEBUG_AUGMENT_PROPAGATE_CHECK
940static void
941augment_tree_propagate_check(void)
942{
943 struct vmap_area *va;
944 unsigned long computed_size;
945
946 list_for_each_entry(va, &free_vmap_area_list, list) {
947 computed_size = compute_subtree_max_size(va);
948 if (computed_size != va->subtree_max_size)
949 pr_emerg("tree is corrupted: %lu, %lu\n",
950 va_size(va), va->subtree_max_size);
951 }
952}
953#endif
954
955/*
956 * This function populates subtree_max_size from bottom to upper
957 * levels starting from VA point. The propagation must be done
958 * when VA size is modified by changing its va_start/va_end. Or
959 * in case of newly inserting of VA to the tree.
960 *
961 * It means that __augment_tree_propagate_from() must be called:
962 * - After VA has been inserted to the tree(free path);
963 * - After VA has been shrunk(allocation path);
964 * - After VA has been increased(merging path).
965 *
966 * Please note that, it does not mean that upper parent nodes
967 * and their subtree_max_size are recalculated all the time up
968 * to the root node.
969 *
970 * 4--8
971 * /\
972 * / \
973 * / \
974 * 2--2 8--8
975 *
976 * For example if we modify the node 4, shrinking it to 2, then
977 * no any modification is required. If we shrink the node 2 to 1
978 * its subtree_max_size is updated only, and set to 1. If we shrink
979 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
980 * node becomes 4--6.
981 */
982static __always_inline void
983augment_tree_propagate_from(struct vmap_area *va)
984{
985 /*
986 * Populate the tree from bottom towards the root until
987 * the calculated maximum available size of checked node
988 * is equal to its current one.
989 */
990 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
991
992#if DEBUG_AUGMENT_PROPAGATE_CHECK
993 augment_tree_propagate_check();
994#endif
995}
996
997static void
998insert_vmap_area(struct vmap_area *va,
999 struct rb_root *root, struct list_head *head)
1000{
1001 struct rb_node **link;
1002 struct rb_node *parent;
1003
1004 link = find_va_links(va, root, NULL, &parent);
1005 if (link)
1006 link_va(va, root, parent, link, head);
1007}
1008
1009static void
1010insert_vmap_area_augment(struct vmap_area *va,
1011 struct rb_node *from, struct rb_root *root,
1012 struct list_head *head)
1013{
1014 struct rb_node **link;
1015 struct rb_node *parent;
1016
1017 if (from)
1018 link = find_va_links(va, NULL, from, &parent);
1019 else
1020 link = find_va_links(va, root, NULL, &parent);
1021
1022 if (link) {
1023 link_va(va, root, parent, link, head);
1024 augment_tree_propagate_from(va);
1025 }
1026}
1027
1028/*
1029 * Merge de-allocated chunk of VA memory with previous
1030 * and next free blocks. If coalesce is not done a new
1031 * free area is inserted. If VA has been merged, it is
1032 * freed.
1033 *
1034 * Please note, it can return NULL in case of overlap
1035 * ranges, followed by WARN() report. Despite it is a
1036 * buggy behaviour, a system can be alive and keep
1037 * ongoing.
1038 */
1039static __always_inline struct vmap_area *
1040merge_or_add_vmap_area(struct vmap_area *va,
1041 struct rb_root *root, struct list_head *head)
1042{
1043 struct vmap_area *sibling;
1044 struct list_head *next;
1045 struct rb_node **link;
1046 struct rb_node *parent;
1047 bool merged = false;
1048
1049 /*
1050 * Find a place in the tree where VA potentially will be
1051 * inserted, unless it is merged with its sibling/siblings.
1052 */
1053 link = find_va_links(va, root, NULL, &parent);
1054 if (!link)
1055 return NULL;
1056
1057 /*
1058 * Get next node of VA to check if merging can be done.
1059 */
1060 next = get_va_next_sibling(parent, link);
1061 if (unlikely(next == NULL))
1062 goto insert;
1063
1064 /*
1065 * start end
1066 * | |
1067 * |<------VA------>|<-----Next----->|
1068 * | |
1069 * start end
1070 */
1071 if (next != head) {
1072 sibling = list_entry(next, struct vmap_area, list);
1073 if (sibling->va_start == va->va_end) {
1074 sibling->va_start = va->va_start;
1075
1076 /* Free vmap_area object. */
1077 kmem_cache_free(vmap_area_cachep, va);
1078
1079 /* Point to the new merged area. */
1080 va = sibling;
1081 merged = true;
1082 }
1083 }
1084
1085 /*
1086 * start end
1087 * | |
1088 * |<-----Prev----->|<------VA------>|
1089 * | |
1090 * start end
1091 */
1092 if (next->prev != head) {
1093 sibling = list_entry(next->prev, struct vmap_area, list);
1094 if (sibling->va_end == va->va_start) {
1095 /*
1096 * If both neighbors are coalesced, it is important
1097 * to unlink the "next" node first, followed by merging
1098 * with "previous" one. Otherwise the tree might not be
1099 * fully populated if a sibling's augmented value is
1100 * "normalized" because of rotation operations.
1101 */
1102 if (merged)
1103 unlink_va(va, root);
1104
1105 sibling->va_end = va->va_end;
1106
1107 /* Free vmap_area object. */
1108 kmem_cache_free(vmap_area_cachep, va);
1109
1110 /* Point to the new merged area. */
1111 va = sibling;
1112 merged = true;
1113 }
1114 }
1115
1116insert:
1117 if (!merged)
1118 link_va(va, root, parent, link, head);
1119
1120 return va;
1121}
1122
1123static __always_inline struct vmap_area *
1124merge_or_add_vmap_area_augment(struct vmap_area *va,
1125 struct rb_root *root, struct list_head *head)
1126{
1127 va = merge_or_add_vmap_area(va, root, head);
1128 if (va)
1129 augment_tree_propagate_from(va);
1130
1131 return va;
1132}
1133
1134static __always_inline bool
1135is_within_this_va(struct vmap_area *va, unsigned long size,
1136 unsigned long align, unsigned long vstart)
1137{
1138 unsigned long nva_start_addr;
1139
1140 if (va->va_start > vstart)
1141 nva_start_addr = ALIGN(va->va_start, align);
1142 else
1143 nva_start_addr = ALIGN(vstart, align);
1144
1145 /* Can be overflowed due to big size or alignment. */
1146 if (nva_start_addr + size < nva_start_addr ||
1147 nva_start_addr < vstart)
1148 return false;
1149
1150 return (nva_start_addr + size <= va->va_end);
1151}
1152
1153/*
1154 * Find the first free block(lowest start address) in the tree,
1155 * that will accomplish the request corresponding to passing
1156 * parameters.
1157 */
1158static __always_inline struct vmap_area *
1159find_vmap_lowest_match(unsigned long size,
1160 unsigned long align, unsigned long vstart)
1161{
1162 struct vmap_area *va;
1163 struct rb_node *node;
1164 unsigned long length;
1165
1166 /* Start from the root. */
1167 node = free_vmap_area_root.rb_node;
1168
1169 /* Adjust the search size for alignment overhead. */
1170 length = size + align - 1;
1171
1172 while (node) {
1173 va = rb_entry(node, struct vmap_area, rb_node);
1174
1175 if (get_subtree_max_size(node->rb_left) >= length &&
1176 vstart < va->va_start) {
1177 node = node->rb_left;
1178 } else {
1179 if (is_within_this_va(va, size, align, vstart))
1180 return va;
1181
1182 /*
1183 * Does not make sense to go deeper towards the right
1184 * sub-tree if it does not have a free block that is
1185 * equal or bigger to the requested search length.
1186 */
1187 if (get_subtree_max_size(node->rb_right) >= length) {
1188 node = node->rb_right;
1189 continue;
1190 }
1191
1192 /*
1193 * OK. We roll back and find the first right sub-tree,
1194 * that will satisfy the search criteria. It can happen
1195 * only once due to "vstart" restriction.
1196 */
1197 while ((node = rb_parent(node))) {
1198 va = rb_entry(node, struct vmap_area, rb_node);
1199 if (is_within_this_va(va, size, align, vstart))
1200 return va;
1201
1202 if (get_subtree_max_size(node->rb_right) >= length &&
1203 vstart <= va->va_start) {
1204 node = node->rb_right;
1205 break;
1206 }
1207 }
1208 }
1209 }
1210
1211 return NULL;
1212}
1213
1214#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1215#include <linux/random.h>
1216
1217static struct vmap_area *
1218find_vmap_lowest_linear_match(unsigned long size,
1219 unsigned long align, unsigned long vstart)
1220{
1221 struct vmap_area *va;
1222
1223 list_for_each_entry(va, &free_vmap_area_list, list) {
1224 if (!is_within_this_va(va, size, align, vstart))
1225 continue;
1226
1227 return va;
1228 }
1229
1230 return NULL;
1231}
1232
1233static void
1234find_vmap_lowest_match_check(unsigned long size)
1235{
1236 struct vmap_area *va_1, *va_2;
1237 unsigned long vstart;
1238 unsigned int rnd;
1239
1240 get_random_bytes(&rnd, sizeof(rnd));
1241 vstart = VMALLOC_START + rnd;
1242
1243 va_1 = find_vmap_lowest_match(size, 1, vstart);
1244 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
1245
1246 if (va_1 != va_2)
1247 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1248 va_1, va_2, vstart);
1249}
1250#endif
1251
1252enum fit_type {
1253 NOTHING_FIT = 0,
1254 FL_FIT_TYPE = 1, /* full fit */
1255 LE_FIT_TYPE = 2, /* left edge fit */
1256 RE_FIT_TYPE = 3, /* right edge fit */
1257 NE_FIT_TYPE = 4 /* no edge fit */
1258};
1259
1260static __always_inline enum fit_type
1261classify_va_fit_type(struct vmap_area *va,
1262 unsigned long nva_start_addr, unsigned long size)
1263{
1264 enum fit_type type;
1265
1266 /* Check if it is within VA. */
1267 if (nva_start_addr < va->va_start ||
1268 nva_start_addr + size > va->va_end)
1269 return NOTHING_FIT;
1270
1271 /* Now classify. */
1272 if (va->va_start == nva_start_addr) {
1273 if (va->va_end == nva_start_addr + size)
1274 type = FL_FIT_TYPE;
1275 else
1276 type = LE_FIT_TYPE;
1277 } else if (va->va_end == nva_start_addr + size) {
1278 type = RE_FIT_TYPE;
1279 } else {
1280 type = NE_FIT_TYPE;
1281 }
1282
1283 return type;
1284}
1285
1286static __always_inline int
1287adjust_va_to_fit_type(struct vmap_area *va,
1288 unsigned long nva_start_addr, unsigned long size,
1289 enum fit_type type)
1290{
1291 struct vmap_area *lva = NULL;
1292
1293 if (type == FL_FIT_TYPE) {
1294 /*
1295 * No need to split VA, it fully fits.
1296 *
1297 * | |
1298 * V NVA V
1299 * |---------------|
1300 */
1301 unlink_va(va, &free_vmap_area_root);
1302 kmem_cache_free(vmap_area_cachep, va);
1303 } else if (type == LE_FIT_TYPE) {
1304 /*
1305 * Split left edge of fit VA.
1306 *
1307 * | |
1308 * V NVA V R
1309 * |-------|-------|
1310 */
1311 va->va_start += size;
1312 } else if (type == RE_FIT_TYPE) {
1313 /*
1314 * Split right edge of fit VA.
1315 *
1316 * | |
1317 * L V NVA V
1318 * |-------|-------|
1319 */
1320 va->va_end = nva_start_addr;
1321 } else if (type == NE_FIT_TYPE) {
1322 /*
1323 * Split no edge of fit VA.
1324 *
1325 * | |
1326 * L V NVA V R
1327 * |---|-------|---|
1328 */
1329 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1330 if (unlikely(!lva)) {
1331 /*
1332 * For percpu allocator we do not do any pre-allocation
1333 * and leave it as it is. The reason is it most likely
1334 * never ends up with NE_FIT_TYPE splitting. In case of
1335 * percpu allocations offsets and sizes are aligned to
1336 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1337 * are its main fitting cases.
1338 *
1339 * There are a few exceptions though, as an example it is
1340 * a first allocation (early boot up) when we have "one"
1341 * big free space that has to be split.
1342 *
1343 * Also we can hit this path in case of regular "vmap"
1344 * allocations, if "this" current CPU was not preloaded.
1345 * See the comment in alloc_vmap_area() why. If so, then
1346 * GFP_NOWAIT is used instead to get an extra object for
1347 * split purpose. That is rare and most time does not
1348 * occur.
1349 *
1350 * What happens if an allocation gets failed. Basically,
1351 * an "overflow" path is triggered to purge lazily freed
1352 * areas to free some memory, then, the "retry" path is
1353 * triggered to repeat one more time. See more details
1354 * in alloc_vmap_area() function.
1355 */
1356 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1357 if (!lva)
1358 return -1;
1359 }
1360
1361 /*
1362 * Build the remainder.
1363 */
1364 lva->va_start = va->va_start;
1365 lva->va_end = nva_start_addr;
1366
1367 /*
1368 * Shrink this VA to remaining size.
1369 */
1370 va->va_start = nva_start_addr + size;
1371 } else {
1372 return -1;
1373 }
1374
1375 if (type != FL_FIT_TYPE) {
1376 augment_tree_propagate_from(va);
1377
1378 if (lva) /* type == NE_FIT_TYPE */
1379 insert_vmap_area_augment(lva, &va->rb_node,
1380 &free_vmap_area_root, &free_vmap_area_list);
1381 }
1382
1383 return 0;
1384}
1385
1386/*
1387 * Returns a start address of the newly allocated area, if success.
1388 * Otherwise a vend is returned that indicates failure.
1389 */
1390static __always_inline unsigned long
1391__alloc_vmap_area(unsigned long size, unsigned long align,
1392 unsigned long vstart, unsigned long vend)
1393{
1394 unsigned long nva_start_addr;
1395 struct vmap_area *va;
1396 enum fit_type type;
1397 int ret;
1398
1399 va = find_vmap_lowest_match(size, align, vstart);
1400 if (unlikely(!va))
1401 return vend;
1402
1403 if (va->va_start > vstart)
1404 nva_start_addr = ALIGN(va->va_start, align);
1405 else
1406 nva_start_addr = ALIGN(vstart, align);
1407
1408 /* Check the "vend" restriction. */
1409 if (nva_start_addr + size > vend)
1410 return vend;
1411
1412 /* Classify what we have found. */
1413 type = classify_va_fit_type(va, nva_start_addr, size);
1414 if (WARN_ON_ONCE(type == NOTHING_FIT))
1415 return vend;
1416
1417 /* Update the free vmap_area. */
1418 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1419 if (ret)
1420 return vend;
1421
1422#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1423 find_vmap_lowest_match_check(size);
1424#endif
1425
1426 return nva_start_addr;
1427}
1428
1429/*
1430 * Free a region of KVA allocated by alloc_vmap_area
1431 */
1432static void free_vmap_area(struct vmap_area *va)
1433{
1434 /*
1435 * Remove from the busy tree/list.
1436 */
1437 spin_lock(&vmap_area_lock);
1438 unlink_va(va, &vmap_area_root);
1439 spin_unlock(&vmap_area_lock);
1440
1441 /*
1442 * Insert/Merge it back to the free tree/list.
1443 */
1444 spin_lock(&free_vmap_area_lock);
1445 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1446 spin_unlock(&free_vmap_area_lock);
1447}
1448
1449static inline void
1450preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1451{
1452 struct vmap_area *va = NULL;
1453
1454 /*
1455 * Preload this CPU with one extra vmap_area object. It is used
1456 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1457 * a CPU that does an allocation is preloaded.
1458 *
1459 * We do it in non-atomic context, thus it allows us to use more
1460 * permissive allocation masks to be more stable under low memory
1461 * condition and high memory pressure.
1462 */
1463 if (!this_cpu_read(ne_fit_preload_node))
1464 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1465
1466 spin_lock(lock);
1467
1468 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1469 kmem_cache_free(vmap_area_cachep, va);
1470}
1471
1472/*
1473 * Allocate a region of KVA of the specified size and alignment, within the
1474 * vstart and vend.
1475 */
1476static struct vmap_area *alloc_vmap_area(unsigned long size,
1477 unsigned long align,
1478 unsigned long vstart, unsigned long vend,
1479 int node, gfp_t gfp_mask)
1480{
1481 struct vmap_area *va;
1482 unsigned long addr;
1483 int purged = 0;
1484 int ret;
1485
1486 BUG_ON(!size);
1487 BUG_ON(offset_in_page(size));
1488 BUG_ON(!is_power_of_2(align));
1489
1490 if (unlikely(!vmap_initialized))
1491 return ERR_PTR(-EBUSY);
1492
1493 might_sleep();
1494 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1495
1496 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1497 if (unlikely(!va))
1498 return ERR_PTR(-ENOMEM);
1499
1500 /*
1501 * Only scan the relevant parts containing pointers to other objects
1502 * to avoid false negatives.
1503 */
1504 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1505
1506retry:
1507 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1508 addr = __alloc_vmap_area(size, align, vstart, vend);
1509 spin_unlock(&free_vmap_area_lock);
1510
1511 /*
1512 * If an allocation fails, the "vend" address is
1513 * returned. Therefore trigger the overflow path.
1514 */
1515 if (unlikely(addr == vend))
1516 goto overflow;
1517
1518 va->va_start = addr;
1519 va->va_end = addr + size;
1520 va->vm = NULL;
1521
1522 spin_lock(&vmap_area_lock);
1523 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1524 spin_unlock(&vmap_area_lock);
1525
1526 BUG_ON(!IS_ALIGNED(va->va_start, align));
1527 BUG_ON(va->va_start < vstart);
1528 BUG_ON(va->va_end > vend);
1529
1530 ret = kasan_populate_vmalloc(addr, size);
1531 if (ret) {
1532 free_vmap_area(va);
1533 return ERR_PTR(ret);
1534 }
1535
1536 return va;
1537
1538overflow:
1539 if (!purged) {
1540 purge_vmap_area_lazy();
1541 purged = 1;
1542 goto retry;
1543 }
1544
1545 if (gfpflags_allow_blocking(gfp_mask)) {
1546 unsigned long freed = 0;
1547 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1548 if (freed > 0) {
1549 purged = 0;
1550 goto retry;
1551 }
1552 }
1553
1554 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1555 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1556 size);
1557
1558 kmem_cache_free(vmap_area_cachep, va);
1559 return ERR_PTR(-EBUSY);
1560}
1561
1562int register_vmap_purge_notifier(struct notifier_block *nb)
1563{
1564 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1565}
1566EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1567
1568int unregister_vmap_purge_notifier(struct notifier_block *nb)
1569{
1570 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1571}
1572EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1573
1574/*
1575 * lazy_max_pages is the maximum amount of virtual address space we gather up
1576 * before attempting to purge with a TLB flush.
1577 *
1578 * There is a tradeoff here: a larger number will cover more kernel page tables
1579 * and take slightly longer to purge, but it will linearly reduce the number of
1580 * global TLB flushes that must be performed. It would seem natural to scale
1581 * this number up linearly with the number of CPUs (because vmapping activity
1582 * could also scale linearly with the number of CPUs), however it is likely
1583 * that in practice, workloads might be constrained in other ways that mean
1584 * vmap activity will not scale linearly with CPUs. Also, I want to be
1585 * conservative and not introduce a big latency on huge systems, so go with
1586 * a less aggressive log scale. It will still be an improvement over the old
1587 * code, and it will be simple to change the scale factor if we find that it
1588 * becomes a problem on bigger systems.
1589 */
1590static unsigned long lazy_max_pages(void)
1591{
1592 unsigned int log;
1593
1594 log = fls(num_online_cpus());
1595
1596 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1597}
1598
1599static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1600
1601/*
1602 * Serialize vmap purging. There is no actual critical section protected
1603 * by this look, but we want to avoid concurrent calls for performance
1604 * reasons and to make the pcpu_get_vm_areas more deterministic.
1605 */
1606static DEFINE_MUTEX(vmap_purge_lock);
1607
1608/* for per-CPU blocks */
1609static void purge_fragmented_blocks_allcpus(void);
1610
1611#ifdef CONFIG_X86_64
1612/*
1613 * called before a call to iounmap() if the caller wants vm_area_struct's
1614 * immediately freed.
1615 */
1616void set_iounmap_nonlazy(void)
1617{
1618 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1619}
1620#endif /* CONFIG_X86_64 */
1621
1622/*
1623 * Purges all lazily-freed vmap areas.
1624 */
1625static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1626{
1627 unsigned long resched_threshold;
1628 struct list_head local_pure_list;
1629 struct vmap_area *va, *n_va;
1630
1631 lockdep_assert_held(&vmap_purge_lock);
1632
1633 spin_lock(&purge_vmap_area_lock);
1634 purge_vmap_area_root = RB_ROOT;
1635 list_replace_init(&purge_vmap_area_list, &local_pure_list);
1636 spin_unlock(&purge_vmap_area_lock);
1637
1638 if (unlikely(list_empty(&local_pure_list)))
1639 return false;
1640
1641 start = min(start,
1642 list_first_entry(&local_pure_list,
1643 struct vmap_area, list)->va_start);
1644
1645 end = max(end,
1646 list_last_entry(&local_pure_list,
1647 struct vmap_area, list)->va_end);
1648
1649 flush_tlb_kernel_range(start, end);
1650 resched_threshold = lazy_max_pages() << 1;
1651
1652 spin_lock(&free_vmap_area_lock);
1653 list_for_each_entry_safe(va, n_va, &local_pure_list, list) {
1654 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1655 unsigned long orig_start = va->va_start;
1656 unsigned long orig_end = va->va_end;
1657
1658 /*
1659 * Finally insert or merge lazily-freed area. It is
1660 * detached and there is no need to "unlink" it from
1661 * anything.
1662 */
1663 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1664 &free_vmap_area_list);
1665
1666 if (!va)
1667 continue;
1668
1669 if (is_vmalloc_or_module_addr((void *)orig_start))
1670 kasan_release_vmalloc(orig_start, orig_end,
1671 va->va_start, va->va_end);
1672
1673 atomic_long_sub(nr, &vmap_lazy_nr);
1674
1675 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1676 cond_resched_lock(&free_vmap_area_lock);
1677 }
1678 spin_unlock(&free_vmap_area_lock);
1679 return true;
1680}
1681
1682/*
1683 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1684 * is already purging.
1685 */
1686static void try_purge_vmap_area_lazy(void)
1687{
1688 if (mutex_trylock(&vmap_purge_lock)) {
1689 __purge_vmap_area_lazy(ULONG_MAX, 0);
1690 mutex_unlock(&vmap_purge_lock);
1691 }
1692}
1693
1694/*
1695 * Kick off a purge of the outstanding lazy areas.
1696 */
1697static void purge_vmap_area_lazy(void)
1698{
1699 mutex_lock(&vmap_purge_lock);
1700 purge_fragmented_blocks_allcpus();
1701 __purge_vmap_area_lazy(ULONG_MAX, 0);
1702 mutex_unlock(&vmap_purge_lock);
1703}
1704
1705/*
1706 * Free a vmap area, caller ensuring that the area has been unmapped
1707 * and flush_cache_vunmap had been called for the correct range
1708 * previously.
1709 */
1710static void free_vmap_area_noflush(struct vmap_area *va)
1711{
1712 unsigned long nr_lazy;
1713
1714 spin_lock(&vmap_area_lock);
1715 unlink_va(va, &vmap_area_root);
1716 spin_unlock(&vmap_area_lock);
1717
1718 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1719 PAGE_SHIFT, &vmap_lazy_nr);
1720
1721 /*
1722 * Merge or place it to the purge tree/list.
1723 */
1724 spin_lock(&purge_vmap_area_lock);
1725 merge_or_add_vmap_area(va,
1726 &purge_vmap_area_root, &purge_vmap_area_list);
1727 spin_unlock(&purge_vmap_area_lock);
1728
1729 /* After this point, we may free va at any time */
1730 if (unlikely(nr_lazy > lazy_max_pages()))
1731 try_purge_vmap_area_lazy();
1732}
1733
1734/*
1735 * Free and unmap a vmap area
1736 */
1737static void free_unmap_vmap_area(struct vmap_area *va)
1738{
1739 flush_cache_vunmap(va->va_start, va->va_end);
1740 vunmap_range_noflush(va->va_start, va->va_end);
1741 if (debug_pagealloc_enabled_static())
1742 flush_tlb_kernel_range(va->va_start, va->va_end);
1743
1744 free_vmap_area_noflush(va);
1745}
1746
1747static struct vmap_area *find_vmap_area(unsigned long addr)
1748{
1749 struct vmap_area *va;
1750
1751 spin_lock(&vmap_area_lock);
1752 va = __find_vmap_area(addr);
1753 spin_unlock(&vmap_area_lock);
1754
1755 return va;
1756}
1757
1758/*** Per cpu kva allocator ***/
1759
1760/*
1761 * vmap space is limited especially on 32 bit architectures. Ensure there is
1762 * room for at least 16 percpu vmap blocks per CPU.
1763 */
1764/*
1765 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1766 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1767 * instead (we just need a rough idea)
1768 */
1769#if BITS_PER_LONG == 32
1770#define VMALLOC_SPACE (128UL*1024*1024)
1771#else
1772#define VMALLOC_SPACE (128UL*1024*1024*1024)
1773#endif
1774
1775#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1776#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1777#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1778#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1779#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1780#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1781#define VMAP_BBMAP_BITS \
1782 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1783 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1784 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1785
1786#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1787
1788struct vmap_block_queue {
1789 spinlock_t lock;
1790 struct list_head free;
1791};
1792
1793struct vmap_block {
1794 spinlock_t lock;
1795 struct vmap_area *va;
1796 unsigned long free, dirty;
1797 unsigned long dirty_min, dirty_max; /*< dirty range */
1798 struct list_head free_list;
1799 struct rcu_head rcu_head;
1800 struct list_head purge;
1801};
1802
1803/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1804static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1805
1806/*
1807 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1808 * in the free path. Could get rid of this if we change the API to return a
1809 * "cookie" from alloc, to be passed to free. But no big deal yet.
1810 */
1811static DEFINE_XARRAY(vmap_blocks);
1812
1813/*
1814 * We should probably have a fallback mechanism to allocate virtual memory
1815 * out of partially filled vmap blocks. However vmap block sizing should be
1816 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1817 * big problem.
1818 */
1819
1820static unsigned long addr_to_vb_idx(unsigned long addr)
1821{
1822 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1823 addr /= VMAP_BLOCK_SIZE;
1824 return addr;
1825}
1826
1827static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1828{
1829 unsigned long addr;
1830
1831 addr = va_start + (pages_off << PAGE_SHIFT);
1832 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1833 return (void *)addr;
1834}
1835
1836/**
1837 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1838 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1839 * @order: how many 2^order pages should be occupied in newly allocated block
1840 * @gfp_mask: flags for the page level allocator
1841 *
1842 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1843 */
1844static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1845{
1846 struct vmap_block_queue *vbq;
1847 struct vmap_block *vb;
1848 struct vmap_area *va;
1849 unsigned long vb_idx;
1850 int node, err;
1851 void *vaddr;
1852
1853 node = numa_node_id();
1854
1855 vb = kmalloc_node(sizeof(struct vmap_block),
1856 gfp_mask & GFP_RECLAIM_MASK, node);
1857 if (unlikely(!vb))
1858 return ERR_PTR(-ENOMEM);
1859
1860 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1861 VMALLOC_START, VMALLOC_END,
1862 node, gfp_mask);
1863 if (IS_ERR(va)) {
1864 kfree(vb);
1865 return ERR_CAST(va);
1866 }
1867
1868 vaddr = vmap_block_vaddr(va->va_start, 0);
1869 spin_lock_init(&vb->lock);
1870 vb->va = va;
1871 /* At least something should be left free */
1872 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1873 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1874 vb->dirty = 0;
1875 vb->dirty_min = VMAP_BBMAP_BITS;
1876 vb->dirty_max = 0;
1877 INIT_LIST_HEAD(&vb->free_list);
1878
1879 vb_idx = addr_to_vb_idx(va->va_start);
1880 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1881 if (err) {
1882 kfree(vb);
1883 free_vmap_area(va);
1884 return ERR_PTR(err);
1885 }
1886
1887 vbq = &get_cpu_var(vmap_block_queue);
1888 spin_lock(&vbq->lock);
1889 list_add_tail_rcu(&vb->free_list, &vbq->free);
1890 spin_unlock(&vbq->lock);
1891 put_cpu_var(vmap_block_queue);
1892
1893 return vaddr;
1894}
1895
1896static void free_vmap_block(struct vmap_block *vb)
1897{
1898 struct vmap_block *tmp;
1899
1900 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1901 BUG_ON(tmp != vb);
1902
1903 free_vmap_area_noflush(vb->va);
1904 kfree_rcu(vb, rcu_head);
1905}
1906
1907static void purge_fragmented_blocks(int cpu)
1908{
1909 LIST_HEAD(purge);
1910 struct vmap_block *vb;
1911 struct vmap_block *n_vb;
1912 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1913
1914 rcu_read_lock();
1915 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1916
1917 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1918 continue;
1919
1920 spin_lock(&vb->lock);
1921 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1922 vb->free = 0; /* prevent further allocs after releasing lock */
1923 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1924 vb->dirty_min = 0;
1925 vb->dirty_max = VMAP_BBMAP_BITS;
1926 spin_lock(&vbq->lock);
1927 list_del_rcu(&vb->free_list);
1928 spin_unlock(&vbq->lock);
1929 spin_unlock(&vb->lock);
1930 list_add_tail(&vb->purge, &purge);
1931 } else
1932 spin_unlock(&vb->lock);
1933 }
1934 rcu_read_unlock();
1935
1936 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1937 list_del(&vb->purge);
1938 free_vmap_block(vb);
1939 }
1940}
1941
1942static void purge_fragmented_blocks_allcpus(void)
1943{
1944 int cpu;
1945
1946 for_each_possible_cpu(cpu)
1947 purge_fragmented_blocks(cpu);
1948}
1949
1950static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1951{
1952 struct vmap_block_queue *vbq;
1953 struct vmap_block *vb;
1954 void *vaddr = NULL;
1955 unsigned int order;
1956
1957 BUG_ON(offset_in_page(size));
1958 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1959 if (WARN_ON(size == 0)) {
1960 /*
1961 * Allocating 0 bytes isn't what caller wants since
1962 * get_order(0) returns funny result. Just warn and terminate
1963 * early.
1964 */
1965 return NULL;
1966 }
1967 order = get_order(size);
1968
1969 rcu_read_lock();
1970 vbq = &get_cpu_var(vmap_block_queue);
1971 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1972 unsigned long pages_off;
1973
1974 spin_lock(&vb->lock);
1975 if (vb->free < (1UL << order)) {
1976 spin_unlock(&vb->lock);
1977 continue;
1978 }
1979
1980 pages_off = VMAP_BBMAP_BITS - vb->free;
1981 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1982 vb->free -= 1UL << order;
1983 if (vb->free == 0) {
1984 spin_lock(&vbq->lock);
1985 list_del_rcu(&vb->free_list);
1986 spin_unlock(&vbq->lock);
1987 }
1988
1989 spin_unlock(&vb->lock);
1990 break;
1991 }
1992
1993 put_cpu_var(vmap_block_queue);
1994 rcu_read_unlock();
1995
1996 /* Allocate new block if nothing was found */
1997 if (!vaddr)
1998 vaddr = new_vmap_block(order, gfp_mask);
1999
2000 return vaddr;
2001}
2002
2003static void vb_free(unsigned long addr, unsigned long size)
2004{
2005 unsigned long offset;
2006 unsigned int order;
2007 struct vmap_block *vb;
2008
2009 BUG_ON(offset_in_page(size));
2010 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2011
2012 flush_cache_vunmap(addr, addr + size);
2013
2014 order = get_order(size);
2015 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2016 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2017
2018 vunmap_range_noflush(addr, addr + size);
2019
2020 if (debug_pagealloc_enabled_static())
2021 flush_tlb_kernel_range(addr, addr + size);
2022
2023 spin_lock(&vb->lock);
2024
2025 /* Expand dirty range */
2026 vb->dirty_min = min(vb->dirty_min, offset);
2027 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2028
2029 vb->dirty += 1UL << order;
2030 if (vb->dirty == VMAP_BBMAP_BITS) {
2031 BUG_ON(vb->free);
2032 spin_unlock(&vb->lock);
2033 free_vmap_block(vb);
2034 } else
2035 spin_unlock(&vb->lock);
2036}
2037
2038static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2039{
2040 int cpu;
2041
2042 if (unlikely(!vmap_initialized))
2043 return;
2044
2045 might_sleep();
2046
2047 for_each_possible_cpu(cpu) {
2048 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2049 struct vmap_block *vb;
2050
2051 rcu_read_lock();
2052 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2053 spin_lock(&vb->lock);
2054 if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2055 unsigned long va_start = vb->va->va_start;
2056 unsigned long s, e;
2057
2058 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2059 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2060
2061 start = min(s, start);
2062 end = max(e, end);
2063
2064 flush = 1;
2065 }
2066 spin_unlock(&vb->lock);
2067 }
2068 rcu_read_unlock();
2069 }
2070
2071 mutex_lock(&vmap_purge_lock);
2072 purge_fragmented_blocks_allcpus();
2073 if (!__purge_vmap_area_lazy(start, end) && flush)
2074 flush_tlb_kernel_range(start, end);
2075 mutex_unlock(&vmap_purge_lock);
2076}
2077
2078/**
2079 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2080 *
2081 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2082 * to amortize TLB flushing overheads. What this means is that any page you
2083 * have now, may, in a former life, have been mapped into kernel virtual
2084 * address by the vmap layer and so there might be some CPUs with TLB entries
2085 * still referencing that page (additional to the regular 1:1 kernel mapping).
2086 *
2087 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2088 * be sure that none of the pages we have control over will have any aliases
2089 * from the vmap layer.
2090 */
2091void vm_unmap_aliases(void)
2092{
2093 unsigned long start = ULONG_MAX, end = 0;
2094 int flush = 0;
2095
2096 _vm_unmap_aliases(start, end, flush);
2097}
2098EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2099
2100/**
2101 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2102 * @mem: the pointer returned by vm_map_ram
2103 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2104 */
2105void vm_unmap_ram(const void *mem, unsigned int count)
2106{
2107 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2108 unsigned long addr = (unsigned long)mem;
2109 struct vmap_area *va;
2110
2111 might_sleep();
2112 BUG_ON(!addr);
2113 BUG_ON(addr < VMALLOC_START);
2114 BUG_ON(addr > VMALLOC_END);
2115 BUG_ON(!PAGE_ALIGNED(addr));
2116
2117 kasan_poison_vmalloc(mem, size);
2118
2119 if (likely(count <= VMAP_MAX_ALLOC)) {
2120 debug_check_no_locks_freed(mem, size);
2121 vb_free(addr, size);
2122 return;
2123 }
2124
2125 va = find_vmap_area(addr);
2126 BUG_ON(!va);
2127 debug_check_no_locks_freed((void *)va->va_start,
2128 (va->va_end - va->va_start));
2129 free_unmap_vmap_area(va);
2130}
2131EXPORT_SYMBOL(vm_unmap_ram);
2132
2133/**
2134 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2135 * @pages: an array of pointers to the pages to be mapped
2136 * @count: number of pages
2137 * @node: prefer to allocate data structures on this node
2138 *
2139 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2140 * faster than vmap so it's good. But if you mix long-life and short-life
2141 * objects with vm_map_ram(), it could consume lots of address space through
2142 * fragmentation (especially on a 32bit machine). You could see failures in
2143 * the end. Please use this function for short-lived objects.
2144 *
2145 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2146 */
2147void *vm_map_ram(struct page **pages, unsigned int count, int node)
2148{
2149 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2150 unsigned long addr;
2151 void *mem;
2152
2153 if (likely(count <= VMAP_MAX_ALLOC)) {
2154 mem = vb_alloc(size, GFP_KERNEL);
2155 if (IS_ERR(mem))
2156 return NULL;
2157 addr = (unsigned long)mem;
2158 } else {
2159 struct vmap_area *va;
2160 va = alloc_vmap_area(size, PAGE_SIZE,
2161 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2162 if (IS_ERR(va))
2163 return NULL;
2164
2165 addr = va->va_start;
2166 mem = (void *)addr;
2167 }
2168
2169 kasan_unpoison_vmalloc(mem, size);
2170
2171 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2172 pages, PAGE_SHIFT) < 0) {
2173 vm_unmap_ram(mem, count);
2174 return NULL;
2175 }
2176
2177 return mem;
2178}
2179EXPORT_SYMBOL(vm_map_ram);
2180
2181static struct vm_struct *vmlist __initdata;
2182
2183static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2184{
2185#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2186 return vm->page_order;
2187#else
2188 return 0;
2189#endif
2190}
2191
2192static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2193{
2194#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2195 vm->page_order = order;
2196#else
2197 BUG_ON(order != 0);
2198#endif
2199}
2200
2201/**
2202 * vm_area_add_early - add vmap area early during boot
2203 * @vm: vm_struct to add
2204 *
2205 * This function is used to add fixed kernel vm area to vmlist before
2206 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2207 * should contain proper values and the other fields should be zero.
2208 *
2209 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2210 */
2211void __init vm_area_add_early(struct vm_struct *vm)
2212{
2213 struct vm_struct *tmp, **p;
2214
2215 BUG_ON(vmap_initialized);
2216 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2217 if (tmp->addr >= vm->addr) {
2218 BUG_ON(tmp->addr < vm->addr + vm->size);
2219 break;
2220 } else
2221 BUG_ON(tmp->addr + tmp->size > vm->addr);
2222 }
2223 vm->next = *p;
2224 *p = vm;
2225}
2226
2227/**
2228 * vm_area_register_early - register vmap area early during boot
2229 * @vm: vm_struct to register
2230 * @align: requested alignment
2231 *
2232 * This function is used to register kernel vm area before
2233 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2234 * proper values on entry and other fields should be zero. On return,
2235 * vm->addr contains the allocated address.
2236 *
2237 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2238 */
2239void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2240{
2241 static size_t vm_init_off __initdata;
2242 unsigned long addr;
2243
2244 addr = ALIGN(VMALLOC_START + vm_init_off, align);
2245 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
2246
2247 vm->addr = (void *)addr;
2248
2249 vm_area_add_early(vm);
2250}
2251
2252static void vmap_init_free_space(void)
2253{
2254 unsigned long vmap_start = 1;
2255 const unsigned long vmap_end = ULONG_MAX;
2256 struct vmap_area *busy, *free;
2257
2258 /*
2259 * B F B B B F
2260 * -|-----|.....|-----|-----|-----|.....|-
2261 * | The KVA space |
2262 * |<--------------------------------->|
2263 */
2264 list_for_each_entry(busy, &vmap_area_list, list) {
2265 if (busy->va_start - vmap_start > 0) {
2266 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2267 if (!WARN_ON_ONCE(!free)) {
2268 free->va_start = vmap_start;
2269 free->va_end = busy->va_start;
2270
2271 insert_vmap_area_augment(free, NULL,
2272 &free_vmap_area_root,
2273 &free_vmap_area_list);
2274 }
2275 }
2276
2277 vmap_start = busy->va_end;
2278 }
2279
2280 if (vmap_end - vmap_start > 0) {
2281 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2282 if (!WARN_ON_ONCE(!free)) {
2283 free->va_start = vmap_start;
2284 free->va_end = vmap_end;
2285
2286 insert_vmap_area_augment(free, NULL,
2287 &free_vmap_area_root,
2288 &free_vmap_area_list);
2289 }
2290 }
2291}
2292
2293void __init vmalloc_init(void)
2294{
2295 struct vmap_area *va;
2296 struct vm_struct *tmp;
2297 int i;
2298
2299 /*
2300 * Create the cache for vmap_area objects.
2301 */
2302 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2303
2304 for_each_possible_cpu(i) {
2305 struct vmap_block_queue *vbq;
2306 struct vfree_deferred *p;
2307
2308 vbq = &per_cpu(vmap_block_queue, i);
2309 spin_lock_init(&vbq->lock);
2310 INIT_LIST_HEAD(&vbq->free);
2311 p = &per_cpu(vfree_deferred, i);
2312 init_llist_head(&p->list);
2313 INIT_WORK(&p->wq, free_work);
2314 }
2315
2316 /* Import existing vmlist entries. */
2317 for (tmp = vmlist; tmp; tmp = tmp->next) {
2318 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2319 if (WARN_ON_ONCE(!va))
2320 continue;
2321
2322 va->va_start = (unsigned long)tmp->addr;
2323 va->va_end = va->va_start + tmp->size;
2324 va->vm = tmp;
2325 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2326 }
2327
2328 /*
2329 * Now we can initialize a free vmap space.
2330 */
2331 vmap_init_free_space();
2332 vmap_initialized = true;
2333}
2334
2335static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2336 struct vmap_area *va, unsigned long flags, const void *caller)
2337{
2338 vm->flags = flags;
2339 vm->addr = (void *)va->va_start;
2340 vm->size = va->va_end - va->va_start;
2341 vm->caller = caller;
2342 va->vm = vm;
2343}
2344
2345static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2346 unsigned long flags, const void *caller)
2347{
2348 spin_lock(&vmap_area_lock);
2349 setup_vmalloc_vm_locked(vm, va, flags, caller);
2350 spin_unlock(&vmap_area_lock);
2351}
2352
2353static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2354{
2355 /*
2356 * Before removing VM_UNINITIALIZED,
2357 * we should make sure that vm has proper values.
2358 * Pair with smp_rmb() in show_numa_info().
2359 */
2360 smp_wmb();
2361 vm->flags &= ~VM_UNINITIALIZED;
2362}
2363
2364static struct vm_struct *__get_vm_area_node(unsigned long size,
2365 unsigned long align, unsigned long shift, unsigned long flags,
2366 unsigned long start, unsigned long end, int node,
2367 gfp_t gfp_mask, const void *caller)
2368{
2369 struct vmap_area *va;
2370 struct vm_struct *area;
2371 unsigned long requested_size = size;
2372
2373 BUG_ON(in_interrupt());
2374 size = ALIGN(size, 1ul << shift);
2375 if (unlikely(!size))
2376 return NULL;
2377
2378 if (flags & VM_IOREMAP)
2379 align = 1ul << clamp_t(int, get_count_order_long(size),
2380 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2381
2382 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2383 if (unlikely(!area))
2384 return NULL;
2385
2386 if (!(flags & VM_NO_GUARD))
2387 size += PAGE_SIZE;
2388
2389 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2390 if (IS_ERR(va)) {
2391 kfree(area);
2392 return NULL;
2393 }
2394
2395 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2396
2397 setup_vmalloc_vm(area, va, flags, caller);
2398
2399 return area;
2400}
2401
2402struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2403 unsigned long start, unsigned long end,
2404 const void *caller)
2405{
2406 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2407 NUMA_NO_NODE, GFP_KERNEL, caller);
2408}
2409
2410/**
2411 * get_vm_area - reserve a contiguous kernel virtual area
2412 * @size: size of the area
2413 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2414 *
2415 * Search an area of @size in the kernel virtual mapping area,
2416 * and reserved it for out purposes. Returns the area descriptor
2417 * on success or %NULL on failure.
2418 *
2419 * Return: the area descriptor on success or %NULL on failure.
2420 */
2421struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2422{
2423 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2424 VMALLOC_START, VMALLOC_END,
2425 NUMA_NO_NODE, GFP_KERNEL,
2426 __builtin_return_address(0));
2427}
2428
2429struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2430 const void *caller)
2431{
2432 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2433 VMALLOC_START, VMALLOC_END,
2434 NUMA_NO_NODE, GFP_KERNEL, caller);
2435}
2436
2437/**
2438 * find_vm_area - find a continuous kernel virtual area
2439 * @addr: base address
2440 *
2441 * Search for the kernel VM area starting at @addr, and return it.
2442 * It is up to the caller to do all required locking to keep the returned
2443 * pointer valid.
2444 *
2445 * Return: the area descriptor on success or %NULL on failure.
2446 */
2447struct vm_struct *find_vm_area(const void *addr)
2448{
2449 struct vmap_area *va;
2450
2451 va = find_vmap_area((unsigned long)addr);
2452 if (!va)
2453 return NULL;
2454
2455 return va->vm;
2456}
2457
2458/**
2459 * remove_vm_area - find and remove a continuous kernel virtual area
2460 * @addr: base address
2461 *
2462 * Search for the kernel VM area starting at @addr, and remove it.
2463 * This function returns the found VM area, but using it is NOT safe
2464 * on SMP machines, except for its size or flags.
2465 *
2466 * Return: the area descriptor on success or %NULL on failure.
2467 */
2468struct vm_struct *remove_vm_area(const void *addr)
2469{
2470 struct vmap_area *va;
2471
2472 might_sleep();
2473
2474 spin_lock(&vmap_area_lock);
2475 va = __find_vmap_area((unsigned long)addr);
2476 if (va && va->vm) {
2477 struct vm_struct *vm = va->vm;
2478
2479 va->vm = NULL;
2480 spin_unlock(&vmap_area_lock);
2481
2482 kasan_free_shadow(vm);
2483 free_unmap_vmap_area(va);
2484
2485 return vm;
2486 }
2487
2488 spin_unlock(&vmap_area_lock);
2489 return NULL;
2490}
2491
2492static inline void set_area_direct_map(const struct vm_struct *area,
2493 int (*set_direct_map)(struct page *page))
2494{
2495 int i;
2496
2497 /* HUGE_VMALLOC passes small pages to set_direct_map */
2498 for (i = 0; i < area->nr_pages; i++)
2499 if (page_address(area->pages[i]))
2500 set_direct_map(area->pages[i]);
2501}
2502
2503/* Handle removing and resetting vm mappings related to the vm_struct. */
2504static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2505{
2506 unsigned long start = ULONG_MAX, end = 0;
2507 unsigned int page_order = vm_area_page_order(area);
2508 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2509 int flush_dmap = 0;
2510 int i;
2511
2512 remove_vm_area(area->addr);
2513
2514 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2515 if (!flush_reset)
2516 return;
2517
2518 /*
2519 * If not deallocating pages, just do the flush of the VM area and
2520 * return.
2521 */
2522 if (!deallocate_pages) {
2523 vm_unmap_aliases();
2524 return;
2525 }
2526
2527 /*
2528 * If execution gets here, flush the vm mapping and reset the direct
2529 * map. Find the start and end range of the direct mappings to make sure
2530 * the vm_unmap_aliases() flush includes the direct map.
2531 */
2532 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2533 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2534 if (addr) {
2535 unsigned long page_size;
2536
2537 page_size = PAGE_SIZE << page_order;
2538 start = min(addr, start);
2539 end = max(addr + page_size, end);
2540 flush_dmap = 1;
2541 }
2542 }
2543
2544 /*
2545 * Set direct map to something invalid so that it won't be cached if
2546 * there are any accesses after the TLB flush, then flush the TLB and
2547 * reset the direct map permissions to the default.
2548 */
2549 set_area_direct_map(area, set_direct_map_invalid_noflush);
2550 _vm_unmap_aliases(start, end, flush_dmap);
2551 set_area_direct_map(area, set_direct_map_default_noflush);
2552}
2553
2554static void __vunmap(const void *addr, int deallocate_pages)
2555{
2556 struct vm_struct *area;
2557
2558 if (!addr)
2559 return;
2560
2561 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2562 addr))
2563 return;
2564
2565 area = find_vm_area(addr);
2566 if (unlikely(!area)) {
2567 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2568 addr);
2569 return;
2570 }
2571
2572 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2573 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2574
2575 kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2576
2577 vm_remove_mappings(area, deallocate_pages);
2578
2579 if (deallocate_pages) {
2580 unsigned int page_order = vm_area_page_order(area);
2581 int i;
2582
2583 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2584 struct page *page = area->pages[i];
2585
2586 BUG_ON(!page);
2587 __free_pages(page, page_order);
2588 cond_resched();
2589 }
2590 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2591
2592 kvfree(area->pages);
2593 }
2594
2595 kfree(area);
2596}
2597
2598static inline void __vfree_deferred(const void *addr)
2599{
2600 /*
2601 * Use raw_cpu_ptr() because this can be called from preemptible
2602 * context. Preemption is absolutely fine here, because the llist_add()
2603 * implementation is lockless, so it works even if we are adding to
2604 * another cpu's list. schedule_work() should be fine with this too.
2605 */
2606 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2607
2608 if (llist_add((struct llist_node *)addr, &p->list))
2609 schedule_work(&p->wq);
2610}
2611
2612/**
2613 * vfree_atomic - release memory allocated by vmalloc()
2614 * @addr: memory base address
2615 *
2616 * This one is just like vfree() but can be called in any atomic context
2617 * except NMIs.
2618 */
2619void vfree_atomic(const void *addr)
2620{
2621 BUG_ON(in_nmi());
2622
2623 kmemleak_free(addr);
2624
2625 if (!addr)
2626 return;
2627 __vfree_deferred(addr);
2628}
2629
2630static void __vfree(const void *addr)
2631{
2632 if (unlikely(in_interrupt()))
2633 __vfree_deferred(addr);
2634 else
2635 __vunmap(addr, 1);
2636}
2637
2638/**
2639 * vfree - Release memory allocated by vmalloc()
2640 * @addr: Memory base address
2641 *
2642 * Free the virtually continuous memory area starting at @addr, as obtained
2643 * from one of the vmalloc() family of APIs. This will usually also free the
2644 * physical memory underlying the virtual allocation, but that memory is
2645 * reference counted, so it will not be freed until the last user goes away.
2646 *
2647 * If @addr is NULL, no operation is performed.
2648 *
2649 * Context:
2650 * May sleep if called *not* from interrupt context.
2651 * Must not be called in NMI context (strictly speaking, it could be
2652 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2653 * conventions for vfree() arch-dependent would be a really bad idea).
2654 */
2655void vfree(const void *addr)
2656{
2657 BUG_ON(in_nmi());
2658
2659 kmemleak_free(addr);
2660
2661 might_sleep_if(!in_interrupt());
2662
2663 if (!addr)
2664 return;
2665
2666 __vfree(addr);
2667}
2668EXPORT_SYMBOL(vfree);
2669
2670/**
2671 * vunmap - release virtual mapping obtained by vmap()
2672 * @addr: memory base address
2673 *
2674 * Free the virtually contiguous memory area starting at @addr,
2675 * which was created from the page array passed to vmap().
2676 *
2677 * Must not be called in interrupt context.
2678 */
2679void vunmap(const void *addr)
2680{
2681 BUG_ON(in_interrupt());
2682 might_sleep();
2683 if (addr)
2684 __vunmap(addr, 0);
2685}
2686EXPORT_SYMBOL(vunmap);
2687
2688/**
2689 * vmap - map an array of pages into virtually contiguous space
2690 * @pages: array of page pointers
2691 * @count: number of pages to map
2692 * @flags: vm_area->flags
2693 * @prot: page protection for the mapping
2694 *
2695 * Maps @count pages from @pages into contiguous kernel virtual space.
2696 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2697 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2698 * are transferred from the caller to vmap(), and will be freed / dropped when
2699 * vfree() is called on the return value.
2700 *
2701 * Return: the address of the area or %NULL on failure
2702 */
2703void *vmap(struct page **pages, unsigned int count,
2704 unsigned long flags, pgprot_t prot)
2705{
2706 struct vm_struct *area;
2707 unsigned long addr;
2708 unsigned long size; /* In bytes */
2709
2710 might_sleep();
2711
2712 if (count > totalram_pages())
2713 return NULL;
2714
2715 size = (unsigned long)count << PAGE_SHIFT;
2716 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2717 if (!area)
2718 return NULL;
2719
2720 addr = (unsigned long)area->addr;
2721 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2722 pages, PAGE_SHIFT) < 0) {
2723 vunmap(area->addr);
2724 return NULL;
2725 }
2726
2727 if (flags & VM_MAP_PUT_PAGES) {
2728 area->pages = pages;
2729 area->nr_pages = count;
2730 }
2731 return area->addr;
2732}
2733EXPORT_SYMBOL(vmap);
2734
2735#ifdef CONFIG_VMAP_PFN
2736struct vmap_pfn_data {
2737 unsigned long *pfns;
2738 pgprot_t prot;
2739 unsigned int idx;
2740};
2741
2742static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2743{
2744 struct vmap_pfn_data *data = private;
2745
2746 if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2747 return -EINVAL;
2748 *pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2749 return 0;
2750}
2751
2752/**
2753 * vmap_pfn - map an array of PFNs into virtually contiguous space
2754 * @pfns: array of PFNs
2755 * @count: number of pages to map
2756 * @prot: page protection for the mapping
2757 *
2758 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2759 * the start address of the mapping.
2760 */
2761void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2762{
2763 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2764 struct vm_struct *area;
2765
2766 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2767 __builtin_return_address(0));
2768 if (!area)
2769 return NULL;
2770 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2771 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2772 free_vm_area(area);
2773 return NULL;
2774 }
2775 return area->addr;
2776}
2777EXPORT_SYMBOL_GPL(vmap_pfn);
2778#endif /* CONFIG_VMAP_PFN */
2779
2780static inline unsigned int
2781vm_area_alloc_pages(gfp_t gfp, int nid,
2782 unsigned int order, unsigned long nr_pages, struct page **pages)
2783{
2784 unsigned int nr_allocated = 0;
2785
2786 /*
2787 * For order-0 pages we make use of bulk allocator, if
2788 * the page array is partly or not at all populated due
2789 * to fails, fallback to a single page allocator that is
2790 * more permissive.
2791 */
2792 if (!order)
2793 nr_allocated = alloc_pages_bulk_array_node(
2794 gfp, nid, nr_pages, pages);
2795 else
2796 /*
2797 * Compound pages required for remap_vmalloc_page if
2798 * high-order pages.
2799 */
2800 gfp |= __GFP_COMP;
2801
2802 /* High-order pages or fallback path if "bulk" fails. */
2803 while (nr_allocated < nr_pages) {
2804 struct page *page;
2805 int i;
2806
2807 page = alloc_pages_node(nid, gfp, order);
2808 if (unlikely(!page))
2809 break;
2810
2811 /*
2812 * Careful, we allocate and map page-order pages, but
2813 * tracking is done per PAGE_SIZE page so as to keep the
2814 * vm_struct APIs independent of the physical/mapped size.
2815 */
2816 for (i = 0; i < (1U << order); i++)
2817 pages[nr_allocated + i] = page + i;
2818
2819 if (gfpflags_allow_blocking(gfp))
2820 cond_resched();
2821
2822 nr_allocated += 1U << order;
2823 }
2824
2825 return nr_allocated;
2826}
2827
2828static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2829 pgprot_t prot, unsigned int page_shift,
2830 int node)
2831{
2832 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2833 unsigned long addr = (unsigned long)area->addr;
2834 unsigned long size = get_vm_area_size(area);
2835 unsigned long array_size;
2836 unsigned int nr_small_pages = size >> PAGE_SHIFT;
2837 unsigned int page_order;
2838
2839 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
2840 gfp_mask |= __GFP_NOWARN;
2841 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
2842 gfp_mask |= __GFP_HIGHMEM;
2843
2844 /* Please note that the recursion is strictly bounded. */
2845 if (array_size > PAGE_SIZE) {
2846 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
2847 area->caller);
2848 } else {
2849 area->pages = kmalloc_node(array_size, nested_gfp, node);
2850 }
2851
2852 if (!area->pages) {
2853 warn_alloc(gfp_mask, NULL,
2854 "vmalloc error: size %lu, failed to allocated page array size %lu",
2855 nr_small_pages * PAGE_SIZE, array_size);
2856 free_vm_area(area);
2857 return NULL;
2858 }
2859
2860 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
2861 page_order = vm_area_page_order(area);
2862
2863 area->nr_pages = vm_area_alloc_pages(gfp_mask, node,
2864 page_order, nr_small_pages, area->pages);
2865
2866 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2867
2868 /*
2869 * If not enough pages were obtained to accomplish an
2870 * allocation request, free them via __vfree() if any.
2871 */
2872 if (area->nr_pages != nr_small_pages) {
2873 warn_alloc(gfp_mask, NULL,
2874 "vmalloc error: size %lu, page order %u, failed to allocate pages",
2875 area->nr_pages * PAGE_SIZE, page_order);
2876 goto fail;
2877 }
2878
2879 if (vmap_pages_range(addr, addr + size, prot, area->pages,
2880 page_shift) < 0) {
2881 warn_alloc(gfp_mask, NULL,
2882 "vmalloc error: size %lu, failed to map pages",
2883 area->nr_pages * PAGE_SIZE);
2884 goto fail;
2885 }
2886
2887 return area->addr;
2888
2889fail:
2890 __vfree(area->addr);
2891 return NULL;
2892}
2893
2894/**
2895 * __vmalloc_node_range - allocate virtually contiguous memory
2896 * @size: allocation size
2897 * @align: desired alignment
2898 * @start: vm area range start
2899 * @end: vm area range end
2900 * @gfp_mask: flags for the page level allocator
2901 * @prot: protection mask for the allocated pages
2902 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2903 * @node: node to use for allocation or NUMA_NO_NODE
2904 * @caller: caller's return address
2905 *
2906 * Allocate enough pages to cover @size from the page level
2907 * allocator with @gfp_mask flags. Map them into contiguous
2908 * kernel virtual space, using a pagetable protection of @prot.
2909 *
2910 * Return: the address of the area or %NULL on failure
2911 */
2912void *__vmalloc_node_range(unsigned long size, unsigned long align,
2913 unsigned long start, unsigned long end, gfp_t gfp_mask,
2914 pgprot_t prot, unsigned long vm_flags, int node,
2915 const void *caller)
2916{
2917 struct vm_struct *area;
2918 void *addr;
2919 unsigned long real_size = size;
2920 unsigned long real_align = align;
2921 unsigned int shift = PAGE_SHIFT;
2922
2923 if (WARN_ON_ONCE(!size))
2924 return NULL;
2925
2926 if ((size >> PAGE_SHIFT) > totalram_pages()) {
2927 warn_alloc(gfp_mask, NULL,
2928 "vmalloc error: size %lu, exceeds total pages",
2929 real_size);
2930 return NULL;
2931 }
2932
2933 if (vmap_allow_huge && !(vm_flags & VM_NO_HUGE_VMAP)) {
2934 unsigned long size_per_node;
2935
2936 /*
2937 * Try huge pages. Only try for PAGE_KERNEL allocations,
2938 * others like modules don't yet expect huge pages in
2939 * their allocations due to apply_to_page_range not
2940 * supporting them.
2941 */
2942
2943 size_per_node = size;
2944 if (node == NUMA_NO_NODE)
2945 size_per_node /= num_online_nodes();
2946 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
2947 shift = PMD_SHIFT;
2948 else
2949 shift = arch_vmap_pte_supported_shift(size_per_node);
2950
2951 align = max(real_align, 1UL << shift);
2952 size = ALIGN(real_size, 1UL << shift);
2953 }
2954
2955again:
2956 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
2957 VM_UNINITIALIZED | vm_flags, start, end, node,
2958 gfp_mask, caller);
2959 if (!area) {
2960 warn_alloc(gfp_mask, NULL,
2961 "vmalloc error: size %lu, vm_struct allocation failed",
2962 real_size);
2963 goto fail;
2964 }
2965
2966 addr = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
2967 if (!addr)
2968 goto fail;
2969
2970 /*
2971 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2972 * flag. It means that vm_struct is not fully initialized.
2973 * Now, it is fully initialized, so remove this flag here.
2974 */
2975 clear_vm_uninitialized_flag(area);
2976
2977 size = PAGE_ALIGN(size);
2978 kmemleak_vmalloc(area, size, gfp_mask);
2979
2980 return addr;
2981
2982fail:
2983 if (shift > PAGE_SHIFT) {
2984 shift = PAGE_SHIFT;
2985 align = real_align;
2986 size = real_size;
2987 goto again;
2988 }
2989
2990 return NULL;
2991}
2992
2993/**
2994 * __vmalloc_node - allocate virtually contiguous memory
2995 * @size: allocation size
2996 * @align: desired alignment
2997 * @gfp_mask: flags for the page level allocator
2998 * @node: node to use for allocation or NUMA_NO_NODE
2999 * @caller: caller's return address
3000 *
3001 * Allocate enough pages to cover @size from the page level allocator with
3002 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3003 *
3004 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3005 * and __GFP_NOFAIL are not supported
3006 *
3007 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3008 * with mm people.
3009 *
3010 * Return: pointer to the allocated memory or %NULL on error
3011 */
3012void *__vmalloc_node(unsigned long size, unsigned long align,
3013 gfp_t gfp_mask, int node, const void *caller)
3014{
3015 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3016 gfp_mask, PAGE_KERNEL, 0, node, caller);
3017}
3018/*
3019 * This is only for performance analysis of vmalloc and stress purpose.
3020 * It is required by vmalloc test module, therefore do not use it other
3021 * than that.
3022 */
3023#ifdef CONFIG_TEST_VMALLOC_MODULE
3024EXPORT_SYMBOL_GPL(__vmalloc_node);
3025#endif
3026
3027void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3028{
3029 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3030 __builtin_return_address(0));
3031}
3032EXPORT_SYMBOL(__vmalloc);
3033
3034/**
3035 * vmalloc - allocate virtually contiguous memory
3036 * @size: allocation size
3037 *
3038 * Allocate enough pages to cover @size from the page level
3039 * allocator and map them into contiguous kernel virtual space.
3040 *
3041 * For tight control over page level allocator and protection flags
3042 * use __vmalloc() instead.
3043 *
3044 * Return: pointer to the allocated memory or %NULL on error
3045 */
3046void *vmalloc(unsigned long size)
3047{
3048 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3049 __builtin_return_address(0));
3050}
3051EXPORT_SYMBOL(vmalloc);
3052
3053/**
3054 * vmalloc_no_huge - allocate virtually contiguous memory using small pages
3055 * @size: allocation size
3056 *
3057 * Allocate enough non-huge pages to cover @size from the page level
3058 * allocator and map them into contiguous kernel virtual space.
3059 *
3060 * Return: pointer to the allocated memory or %NULL on error
3061 */
3062void *vmalloc_no_huge(unsigned long size)
3063{
3064 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3065 GFP_KERNEL, PAGE_KERNEL, VM_NO_HUGE_VMAP,
3066 NUMA_NO_NODE, __builtin_return_address(0));
3067}
3068EXPORT_SYMBOL(vmalloc_no_huge);
3069
3070/**
3071 * vzalloc - allocate virtually contiguous memory with zero fill
3072 * @size: allocation size
3073 *
3074 * Allocate enough pages to cover @size from the page level
3075 * allocator and map them into contiguous kernel virtual space.
3076 * The memory allocated is set to zero.
3077 *
3078 * For tight control over page level allocator and protection flags
3079 * use __vmalloc() instead.
3080 *
3081 * Return: pointer to the allocated memory or %NULL on error
3082 */
3083void *vzalloc(unsigned long size)
3084{
3085 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3086 __builtin_return_address(0));
3087}
3088EXPORT_SYMBOL(vzalloc);
3089
3090/**
3091 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3092 * @size: allocation size
3093 *
3094 * The resulting memory area is zeroed so it can be mapped to userspace
3095 * without leaking data.
3096 *
3097 * Return: pointer to the allocated memory or %NULL on error
3098 */
3099void *vmalloc_user(unsigned long size)
3100{
3101 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3102 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3103 VM_USERMAP, NUMA_NO_NODE,
3104 __builtin_return_address(0));
3105}
3106EXPORT_SYMBOL(vmalloc_user);
3107
3108/**
3109 * vmalloc_node - allocate memory on a specific node
3110 * @size: allocation size
3111 * @node: numa node
3112 *
3113 * Allocate enough pages to cover @size from the page level
3114 * allocator and map them into contiguous kernel virtual space.
3115 *
3116 * For tight control over page level allocator and protection flags
3117 * use __vmalloc() instead.
3118 *
3119 * Return: pointer to the allocated memory or %NULL on error
3120 */
3121void *vmalloc_node(unsigned long size, int node)
3122{
3123 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3124 __builtin_return_address(0));
3125}
3126EXPORT_SYMBOL(vmalloc_node);
3127
3128/**
3129 * vzalloc_node - allocate memory on a specific node with zero fill
3130 * @size: allocation size
3131 * @node: numa node
3132 *
3133 * Allocate enough pages to cover @size from the page level
3134 * allocator and map them into contiguous kernel virtual space.
3135 * The memory allocated is set to zero.
3136 *
3137 * Return: pointer to the allocated memory or %NULL on error
3138 */
3139void *vzalloc_node(unsigned long size, int node)
3140{
3141 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3142 __builtin_return_address(0));
3143}
3144EXPORT_SYMBOL(vzalloc_node);
3145
3146#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3147#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3148#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3149#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3150#else
3151/*
3152 * 64b systems should always have either DMA or DMA32 zones. For others
3153 * GFP_DMA32 should do the right thing and use the normal zone.
3154 */
3155#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3156#endif
3157
3158/**
3159 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3160 * @size: allocation size
3161 *
3162 * Allocate enough 32bit PA addressable pages to cover @size from the
3163 * page level allocator and map them into contiguous kernel virtual space.
3164 *
3165 * Return: pointer to the allocated memory or %NULL on error
3166 */
3167void *vmalloc_32(unsigned long size)
3168{
3169 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3170 __builtin_return_address(0));
3171}
3172EXPORT_SYMBOL(vmalloc_32);
3173
3174/**
3175 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3176 * @size: allocation size
3177 *
3178 * The resulting memory area is 32bit addressable and zeroed so it can be
3179 * mapped to userspace without leaking data.
3180 *
3181 * Return: pointer to the allocated memory or %NULL on error
3182 */
3183void *vmalloc_32_user(unsigned long size)
3184{
3185 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3186 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3187 VM_USERMAP, NUMA_NO_NODE,
3188 __builtin_return_address(0));
3189}
3190EXPORT_SYMBOL(vmalloc_32_user);
3191
3192/*
3193 * small helper routine , copy contents to buf from addr.
3194 * If the page is not present, fill zero.
3195 */
3196
3197static int aligned_vread(char *buf, char *addr, unsigned long count)
3198{
3199 struct page *p;
3200 int copied = 0;
3201
3202 while (count) {
3203 unsigned long offset, length;
3204
3205 offset = offset_in_page(addr);
3206 length = PAGE_SIZE - offset;
3207 if (length > count)
3208 length = count;
3209 p = vmalloc_to_page(addr);
3210 /*
3211 * To do safe access to this _mapped_ area, we need
3212 * lock. But adding lock here means that we need to add
3213 * overhead of vmalloc()/vfree() calls for this _debug_
3214 * interface, rarely used. Instead of that, we'll use
3215 * kmap() and get small overhead in this access function.
3216 */
3217 if (p) {
3218 /* We can expect USER0 is not used -- see vread() */
3219 void *map = kmap_atomic(p);
3220 memcpy(buf, map + offset, length);
3221 kunmap_atomic(map);
3222 } else
3223 memset(buf, 0, length);
3224
3225 addr += length;
3226 buf += length;
3227 copied += length;
3228 count -= length;
3229 }
3230 return copied;
3231}
3232
3233/**
3234 * vread() - read vmalloc area in a safe way.
3235 * @buf: buffer for reading data
3236 * @addr: vm address.
3237 * @count: number of bytes to be read.
3238 *
3239 * This function checks that addr is a valid vmalloc'ed area, and
3240 * copy data from that area to a given buffer. If the given memory range
3241 * of [addr...addr+count) includes some valid address, data is copied to
3242 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3243 * IOREMAP area is treated as memory hole and no copy is done.
3244 *
3245 * If [addr...addr+count) doesn't includes any intersects with alive
3246 * vm_struct area, returns 0. @buf should be kernel's buffer.
3247 *
3248 * Note: In usual ops, vread() is never necessary because the caller
3249 * should know vmalloc() area is valid and can use memcpy().
3250 * This is for routines which have to access vmalloc area without
3251 * any information, as /proc/kcore.
3252 *
3253 * Return: number of bytes for which addr and buf should be increased
3254 * (same number as @count) or %0 if [addr...addr+count) doesn't
3255 * include any intersection with valid vmalloc area
3256 */
3257long vread(char *buf, char *addr, unsigned long count)
3258{
3259 struct vmap_area *va;
3260 struct vm_struct *vm;
3261 char *vaddr, *buf_start = buf;
3262 unsigned long buflen = count;
3263 unsigned long n;
3264
3265 /* Don't allow overflow */
3266 if ((unsigned long) addr + count < count)
3267 count = -(unsigned long) addr;
3268
3269 spin_lock(&vmap_area_lock);
3270 va = __find_vmap_area((unsigned long)addr);
3271 if (!va)
3272 goto finished;
3273 list_for_each_entry_from(va, &vmap_area_list, list) {
3274 if (!count)
3275 break;
3276
3277 if (!va->vm)
3278 continue;
3279
3280 vm = va->vm;
3281 vaddr = (char *) vm->addr;
3282 if (addr >= vaddr + get_vm_area_size(vm))
3283 continue;
3284 while (addr < vaddr) {
3285 if (count == 0)
3286 goto finished;
3287 *buf = '\0';
3288 buf++;
3289 addr++;
3290 count--;
3291 }
3292 n = vaddr + get_vm_area_size(vm) - addr;
3293 if (n > count)
3294 n = count;
3295 if (!(vm->flags & VM_IOREMAP))
3296 aligned_vread(buf, addr, n);
3297 else /* IOREMAP area is treated as memory hole */
3298 memset(buf, 0, n);
3299 buf += n;
3300 addr += n;
3301 count -= n;
3302 }
3303finished:
3304 spin_unlock(&vmap_area_lock);
3305
3306 if (buf == buf_start)
3307 return 0;
3308 /* zero-fill memory holes */
3309 if (buf != buf_start + buflen)
3310 memset(buf, 0, buflen - (buf - buf_start));
3311
3312 return buflen;
3313}
3314
3315/**
3316 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3317 * @vma: vma to cover
3318 * @uaddr: target user address to start at
3319 * @kaddr: virtual address of vmalloc kernel memory
3320 * @pgoff: offset from @kaddr to start at
3321 * @size: size of map area
3322 *
3323 * Returns: 0 for success, -Exxx on failure
3324 *
3325 * This function checks that @kaddr is a valid vmalloc'ed area,
3326 * and that it is big enough to cover the range starting at
3327 * @uaddr in @vma. Will return failure if that criteria isn't
3328 * met.
3329 *
3330 * Similar to remap_pfn_range() (see mm/memory.c)
3331 */
3332int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3333 void *kaddr, unsigned long pgoff,
3334 unsigned long size)
3335{
3336 struct vm_struct *area;
3337 unsigned long off;
3338 unsigned long end_index;
3339
3340 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3341 return -EINVAL;
3342
3343 size = PAGE_ALIGN(size);
3344
3345 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3346 return -EINVAL;
3347
3348 area = find_vm_area(kaddr);
3349 if (!area)
3350 return -EINVAL;
3351
3352 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3353 return -EINVAL;
3354
3355 if (check_add_overflow(size, off, &end_index) ||
3356 end_index > get_vm_area_size(area))
3357 return -EINVAL;
3358 kaddr += off;
3359
3360 do {
3361 struct page *page = vmalloc_to_page(kaddr);
3362 int ret;
3363
3364 ret = vm_insert_page(vma, uaddr, page);
3365 if (ret)
3366 return ret;
3367
3368 uaddr += PAGE_SIZE;
3369 kaddr += PAGE_SIZE;
3370 size -= PAGE_SIZE;
3371 } while (size > 0);
3372
3373 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3374
3375 return 0;
3376}
3377
3378/**
3379 * remap_vmalloc_range - map vmalloc pages to userspace
3380 * @vma: vma to cover (map full range of vma)
3381 * @addr: vmalloc memory
3382 * @pgoff: number of pages into addr before first page to map
3383 *
3384 * Returns: 0 for success, -Exxx on failure
3385 *
3386 * This function checks that addr is a valid vmalloc'ed area, and
3387 * that it is big enough to cover the vma. Will return failure if
3388 * that criteria isn't met.
3389 *
3390 * Similar to remap_pfn_range() (see mm/memory.c)
3391 */
3392int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3393 unsigned long pgoff)
3394{
3395 return remap_vmalloc_range_partial(vma, vma->vm_start,
3396 addr, pgoff,
3397 vma->vm_end - vma->vm_start);
3398}
3399EXPORT_SYMBOL(remap_vmalloc_range);
3400
3401void free_vm_area(struct vm_struct *area)
3402{
3403 struct vm_struct *ret;
3404 ret = remove_vm_area(area->addr);
3405 BUG_ON(ret != area);
3406 kfree(area);
3407}
3408EXPORT_SYMBOL_GPL(free_vm_area);
3409
3410#ifdef CONFIG_SMP
3411static struct vmap_area *node_to_va(struct rb_node *n)
3412{
3413 return rb_entry_safe(n, struct vmap_area, rb_node);
3414}
3415
3416/**
3417 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3418 * @addr: target address
3419 *
3420 * Returns: vmap_area if it is found. If there is no such area
3421 * the first highest(reverse order) vmap_area is returned
3422 * i.e. va->va_start < addr && va->va_end < addr or NULL
3423 * if there are no any areas before @addr.
3424 */
3425static struct vmap_area *
3426pvm_find_va_enclose_addr(unsigned long addr)
3427{
3428 struct vmap_area *va, *tmp;
3429 struct rb_node *n;
3430
3431 n = free_vmap_area_root.rb_node;
3432 va = NULL;
3433
3434 while (n) {
3435 tmp = rb_entry(n, struct vmap_area, rb_node);
3436 if (tmp->va_start <= addr) {
3437 va = tmp;
3438 if (tmp->va_end >= addr)
3439 break;
3440
3441 n = n->rb_right;
3442 } else {
3443 n = n->rb_left;
3444 }
3445 }
3446
3447 return va;
3448}
3449
3450/**
3451 * pvm_determine_end_from_reverse - find the highest aligned address
3452 * of free block below VMALLOC_END
3453 * @va:
3454 * in - the VA we start the search(reverse order);
3455 * out - the VA with the highest aligned end address.
3456 * @align: alignment for required highest address
3457 *
3458 * Returns: determined end address within vmap_area
3459 */
3460static unsigned long
3461pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3462{
3463 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3464 unsigned long addr;
3465
3466 if (likely(*va)) {
3467 list_for_each_entry_from_reverse((*va),
3468 &free_vmap_area_list, list) {
3469 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3470 if ((*va)->va_start < addr)
3471 return addr;
3472 }
3473 }
3474
3475 return 0;
3476}
3477
3478/**
3479 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3480 * @offsets: array containing offset of each area
3481 * @sizes: array containing size of each area
3482 * @nr_vms: the number of areas to allocate
3483 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3484 *
3485 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3486 * vm_structs on success, %NULL on failure
3487 *
3488 * Percpu allocator wants to use congruent vm areas so that it can
3489 * maintain the offsets among percpu areas. This function allocates
3490 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3491 * be scattered pretty far, distance between two areas easily going up
3492 * to gigabytes. To avoid interacting with regular vmallocs, these
3493 * areas are allocated from top.
3494 *
3495 * Despite its complicated look, this allocator is rather simple. It
3496 * does everything top-down and scans free blocks from the end looking
3497 * for matching base. While scanning, if any of the areas do not fit the
3498 * base address is pulled down to fit the area. Scanning is repeated till
3499 * all the areas fit and then all necessary data structures are inserted
3500 * and the result is returned.
3501 */
3502struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3503 const size_t *sizes, int nr_vms,
3504 size_t align)
3505{
3506 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3507 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3508 struct vmap_area **vas, *va;
3509 struct vm_struct **vms;
3510 int area, area2, last_area, term_area;
3511 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3512 bool purged = false;
3513 enum fit_type type;
3514
3515 /* verify parameters and allocate data structures */
3516 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3517 for (last_area = 0, area = 0; area < nr_vms; area++) {
3518 start = offsets[area];
3519 end = start + sizes[area];
3520
3521 /* is everything aligned properly? */
3522 BUG_ON(!IS_ALIGNED(offsets[area], align));
3523 BUG_ON(!IS_ALIGNED(sizes[area], align));
3524
3525 /* detect the area with the highest address */
3526 if (start > offsets[last_area])
3527 last_area = area;
3528
3529 for (area2 = area + 1; area2 < nr_vms; area2++) {
3530 unsigned long start2 = offsets[area2];
3531 unsigned long end2 = start2 + sizes[area2];
3532
3533 BUG_ON(start2 < end && start < end2);
3534 }
3535 }
3536 last_end = offsets[last_area] + sizes[last_area];
3537
3538 if (vmalloc_end - vmalloc_start < last_end) {
3539 WARN_ON(true);
3540 return NULL;
3541 }
3542
3543 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3544 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3545 if (!vas || !vms)
3546 goto err_free2;
3547
3548 for (area = 0; area < nr_vms; area++) {
3549 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3550 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3551 if (!vas[area] || !vms[area])
3552 goto err_free;
3553 }
3554retry:
3555 spin_lock(&free_vmap_area_lock);
3556
3557 /* start scanning - we scan from the top, begin with the last area */
3558 area = term_area = last_area;
3559 start = offsets[area];
3560 end = start + sizes[area];
3561
3562 va = pvm_find_va_enclose_addr(vmalloc_end);
3563 base = pvm_determine_end_from_reverse(&va, align) - end;
3564
3565 while (true) {
3566 /*
3567 * base might have underflowed, add last_end before
3568 * comparing.
3569 */
3570 if (base + last_end < vmalloc_start + last_end)
3571 goto overflow;
3572
3573 /*
3574 * Fitting base has not been found.
3575 */
3576 if (va == NULL)
3577 goto overflow;
3578
3579 /*
3580 * If required width exceeds current VA block, move
3581 * base downwards and then recheck.
3582 */
3583 if (base + end > va->va_end) {
3584 base = pvm_determine_end_from_reverse(&va, align) - end;
3585 term_area = area;
3586 continue;
3587 }
3588
3589 /*
3590 * If this VA does not fit, move base downwards and recheck.
3591 */
3592 if (base + start < va->va_start) {
3593 va = node_to_va(rb_prev(&va->rb_node));
3594 base = pvm_determine_end_from_reverse(&va, align) - end;
3595 term_area = area;
3596 continue;
3597 }
3598
3599 /*
3600 * This area fits, move on to the previous one. If
3601 * the previous one is the terminal one, we're done.
3602 */
3603 area = (area + nr_vms - 1) % nr_vms;
3604 if (area == term_area)
3605 break;
3606
3607 start = offsets[area];
3608 end = start + sizes[area];
3609 va = pvm_find_va_enclose_addr(base + end);
3610 }
3611
3612 /* we've found a fitting base, insert all va's */
3613 for (area = 0; area < nr_vms; area++) {
3614 int ret;
3615
3616 start = base + offsets[area];
3617 size = sizes[area];
3618
3619 va = pvm_find_va_enclose_addr(start);
3620 if (WARN_ON_ONCE(va == NULL))
3621 /* It is a BUG(), but trigger recovery instead. */
3622 goto recovery;
3623
3624 type = classify_va_fit_type(va, start, size);
3625 if (WARN_ON_ONCE(type == NOTHING_FIT))
3626 /* It is a BUG(), but trigger recovery instead. */
3627 goto recovery;
3628
3629 ret = adjust_va_to_fit_type(va, start, size, type);
3630 if (unlikely(ret))
3631 goto recovery;
3632
3633 /* Allocated area. */
3634 va = vas[area];
3635 va->va_start = start;
3636 va->va_end = start + size;
3637 }
3638
3639 spin_unlock(&free_vmap_area_lock);
3640
3641 /* populate the kasan shadow space */
3642 for (area = 0; area < nr_vms; area++) {
3643 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3644 goto err_free_shadow;
3645
3646 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3647 sizes[area]);
3648 }
3649
3650 /* insert all vm's */
3651 spin_lock(&vmap_area_lock);
3652 for (area = 0; area < nr_vms; area++) {
3653 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3654
3655 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3656 pcpu_get_vm_areas);
3657 }
3658 spin_unlock(&vmap_area_lock);
3659
3660 kfree(vas);
3661 return vms;
3662
3663recovery:
3664 /*
3665 * Remove previously allocated areas. There is no
3666 * need in removing these areas from the busy tree,
3667 * because they are inserted only on the final step
3668 * and when pcpu_get_vm_areas() is success.
3669 */
3670 while (area--) {
3671 orig_start = vas[area]->va_start;
3672 orig_end = vas[area]->va_end;
3673 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3674 &free_vmap_area_list);
3675 if (va)
3676 kasan_release_vmalloc(orig_start, orig_end,
3677 va->va_start, va->va_end);
3678 vas[area] = NULL;
3679 }
3680
3681overflow:
3682 spin_unlock(&free_vmap_area_lock);
3683 if (!purged) {
3684 purge_vmap_area_lazy();
3685 purged = true;
3686
3687 /* Before "retry", check if we recover. */
3688 for (area = 0; area < nr_vms; area++) {
3689 if (vas[area])
3690 continue;
3691
3692 vas[area] = kmem_cache_zalloc(
3693 vmap_area_cachep, GFP_KERNEL);
3694 if (!vas[area])
3695 goto err_free;
3696 }
3697
3698 goto retry;
3699 }
3700
3701err_free:
3702 for (area = 0; area < nr_vms; area++) {
3703 if (vas[area])
3704 kmem_cache_free(vmap_area_cachep, vas[area]);
3705
3706 kfree(vms[area]);
3707 }
3708err_free2:
3709 kfree(vas);
3710 kfree(vms);
3711 return NULL;
3712
3713err_free_shadow:
3714 spin_lock(&free_vmap_area_lock);
3715 /*
3716 * We release all the vmalloc shadows, even the ones for regions that
3717 * hadn't been successfully added. This relies on kasan_release_vmalloc
3718 * being able to tolerate this case.
3719 */
3720 for (area = 0; area < nr_vms; area++) {
3721 orig_start = vas[area]->va_start;
3722 orig_end = vas[area]->va_end;
3723 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3724 &free_vmap_area_list);
3725 if (va)
3726 kasan_release_vmalloc(orig_start, orig_end,
3727 va->va_start, va->va_end);
3728 vas[area] = NULL;
3729 kfree(vms[area]);
3730 }
3731 spin_unlock(&free_vmap_area_lock);
3732 kfree(vas);
3733 kfree(vms);
3734 return NULL;
3735}
3736
3737/**
3738 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3739 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3740 * @nr_vms: the number of allocated areas
3741 *
3742 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3743 */
3744void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3745{
3746 int i;
3747
3748 for (i = 0; i < nr_vms; i++)
3749 free_vm_area(vms[i]);
3750 kfree(vms);
3751}
3752#endif /* CONFIG_SMP */
3753
3754#ifdef CONFIG_PRINTK
3755bool vmalloc_dump_obj(void *object)
3756{
3757 struct vm_struct *vm;
3758 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
3759
3760 vm = find_vm_area(objp);
3761 if (!vm)
3762 return false;
3763 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
3764 vm->nr_pages, (unsigned long)vm->addr, vm->caller);
3765 return true;
3766}
3767#endif
3768
3769#ifdef CONFIG_PROC_FS
3770static void *s_start(struct seq_file *m, loff_t *pos)
3771 __acquires(&vmap_purge_lock)
3772 __acquires(&vmap_area_lock)
3773{
3774 mutex_lock(&vmap_purge_lock);
3775 spin_lock(&vmap_area_lock);
3776
3777 return seq_list_start(&vmap_area_list, *pos);
3778}
3779
3780static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3781{
3782 return seq_list_next(p, &vmap_area_list, pos);
3783}
3784
3785static void s_stop(struct seq_file *m, void *p)
3786 __releases(&vmap_area_lock)
3787 __releases(&vmap_purge_lock)
3788{
3789 spin_unlock(&vmap_area_lock);
3790 mutex_unlock(&vmap_purge_lock);
3791}
3792
3793static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3794{
3795 if (IS_ENABLED(CONFIG_NUMA)) {
3796 unsigned int nr, *counters = m->private;
3797
3798 if (!counters)
3799 return;
3800
3801 if (v->flags & VM_UNINITIALIZED)
3802 return;
3803 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3804 smp_rmb();
3805
3806 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3807
3808 for (nr = 0; nr < v->nr_pages; nr++)
3809 counters[page_to_nid(v->pages[nr])]++;
3810
3811 for_each_node_state(nr, N_HIGH_MEMORY)
3812 if (counters[nr])
3813 seq_printf(m, " N%u=%u", nr, counters[nr]);
3814 }
3815}
3816
3817static void show_purge_info(struct seq_file *m)
3818{
3819 struct vmap_area *va;
3820
3821 spin_lock(&purge_vmap_area_lock);
3822 list_for_each_entry(va, &purge_vmap_area_list, list) {
3823 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3824 (void *)va->va_start, (void *)va->va_end,
3825 va->va_end - va->va_start);
3826 }
3827 spin_unlock(&purge_vmap_area_lock);
3828}
3829
3830static int s_show(struct seq_file *m, void *p)
3831{
3832 struct vmap_area *va;
3833 struct vm_struct *v;
3834
3835 va = list_entry(p, struct vmap_area, list);
3836
3837 /*
3838 * s_show can encounter race with remove_vm_area, !vm on behalf
3839 * of vmap area is being tear down or vm_map_ram allocation.
3840 */
3841 if (!va->vm) {
3842 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3843 (void *)va->va_start, (void *)va->va_end,
3844 va->va_end - va->va_start);
3845
3846 return 0;
3847 }
3848
3849 v = va->vm;
3850
3851 seq_printf(m, "0x%pK-0x%pK %7ld",
3852 v->addr, v->addr + v->size, v->size);
3853
3854 if (v->caller)
3855 seq_printf(m, " %pS", v->caller);
3856
3857 if (v->nr_pages)
3858 seq_printf(m, " pages=%d", v->nr_pages);
3859
3860 if (v->phys_addr)
3861 seq_printf(m, " phys=%pa", &v->phys_addr);
3862
3863 if (v->flags & VM_IOREMAP)
3864 seq_puts(m, " ioremap");
3865
3866 if (v->flags & VM_ALLOC)
3867 seq_puts(m, " vmalloc");
3868
3869 if (v->flags & VM_MAP)
3870 seq_puts(m, " vmap");
3871
3872 if (v->flags & VM_USERMAP)
3873 seq_puts(m, " user");
3874
3875 if (v->flags & VM_DMA_COHERENT)
3876 seq_puts(m, " dma-coherent");
3877
3878 if (is_vmalloc_addr(v->pages))
3879 seq_puts(m, " vpages");
3880
3881 show_numa_info(m, v);
3882 seq_putc(m, '\n');
3883
3884 /*
3885 * As a final step, dump "unpurged" areas.
3886 */
3887 if (list_is_last(&va->list, &vmap_area_list))
3888 show_purge_info(m);
3889
3890 return 0;
3891}
3892
3893static const struct seq_operations vmalloc_op = {
3894 .start = s_start,
3895 .next = s_next,
3896 .stop = s_stop,
3897 .show = s_show,
3898};
3899
3900static int __init proc_vmalloc_init(void)
3901{
3902 if (IS_ENABLED(CONFIG_NUMA))
3903 proc_create_seq_private("vmallocinfo", 0400, NULL,
3904 &vmalloc_op,
3905 nr_node_ids * sizeof(unsigned int), NULL);
3906 else
3907 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3908 return 0;
3909}
3910module_init(proc_vmalloc_init);
3911
3912#endif