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/uio.h>
37 #include <linux/bitops.h>
38 #include <linux/rbtree_augmented.h>
39 #include <linux/overflow.h>
40 #include <linux/pgtable.h>
41 #include <linux/hugetlb.h>
42 #include <linux/sched/mm.h>
43 #include <asm/tlbflush.h>
44 #include <asm/shmparam.h>
45 #include <linux/page_owner.h>
46
47 #define CREATE_TRACE_POINTS
48 #include <trace/events/vmalloc.h>
49
50 #include "internal.h"
51 #include "pgalloc-track.h"
52
53 #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
54 static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
55
56 static int __init set_nohugeiomap(char *str)
57 {
58 ioremap_max_page_shift = PAGE_SHIFT;
59 return 0;
60 }
61 early_param("nohugeiomap", set_nohugeiomap);
62 #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
63 static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
64 #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65
66 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
67 static bool __ro_after_init vmap_allow_huge = true;
68
69 static int __init set_nohugevmalloc(char *str)
70 {
71 vmap_allow_huge = false;
72 return 0;
73 }
74 early_param("nohugevmalloc", set_nohugevmalloc);
75 #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
76 static const bool vmap_allow_huge = false;
77 #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78
79 bool is_vmalloc_addr(const void *x)
80 {
81 unsigned long addr = (unsigned long)kasan_reset_tag(x);
82
83 return addr >= VMALLOC_START && addr < VMALLOC_END;
84 }
85 EXPORT_SYMBOL(is_vmalloc_addr);
86
87 struct vfree_deferred {
88 struct llist_head list;
89 struct work_struct wq;
90 };
91 static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
92
93 /*** Page table manipulation functions ***/
94 static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
95 phys_addr_t phys_addr, pgprot_t prot,
96 unsigned int max_page_shift, pgtbl_mod_mask *mask)
97 {
98 pte_t *pte;
99 u64 pfn;
100 struct page *page;
101 unsigned long size = PAGE_SIZE;
102
103 pfn = phys_addr >> PAGE_SHIFT;
104 pte = pte_alloc_kernel_track(pmd, addr, mask);
105 if (!pte)
106 return -ENOMEM;
107 do {
108 if (!pte_none(ptep_get(pte))) {
109 if (pfn_valid(pfn)) {
110 page = pfn_to_page(pfn);
111 dump_page(page, "remapping already mapped page");
112 }
113 BUG();
114 }
115
116 #ifdef CONFIG_HUGETLB_PAGE
117 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
118 if (size != PAGE_SIZE) {
119 pte_t entry = pfn_pte(pfn, prot);
120
121 entry = arch_make_huge_pte(entry, ilog2(size), 0);
122 set_huge_pte_at(&init_mm, addr, pte, entry, size);
123 pfn += PFN_DOWN(size);
124 continue;
125 }
126 #endif
127 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
128 pfn++;
129 } while (pte += PFN_DOWN(size), addr += size, addr != end);
130 *mask |= PGTBL_PTE_MODIFIED;
131 return 0;
132 }
133
134 static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
135 phys_addr_t phys_addr, pgprot_t prot,
136 unsigned int max_page_shift)
137 {
138 if (max_page_shift < PMD_SHIFT)
139 return 0;
140
141 if (!arch_vmap_pmd_supported(prot))
142 return 0;
143
144 if ((end - addr) != PMD_SIZE)
145 return 0;
146
147 if (!IS_ALIGNED(addr, PMD_SIZE))
148 return 0;
149
150 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
151 return 0;
152
153 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
154 return 0;
155
156 return pmd_set_huge(pmd, phys_addr, prot);
157 }
158
159 static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
160 phys_addr_t phys_addr, pgprot_t prot,
161 unsigned int max_page_shift, pgtbl_mod_mask *mask)
162 {
163 pmd_t *pmd;
164 unsigned long next;
165
166 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
167 if (!pmd)
168 return -ENOMEM;
169 do {
170 next = pmd_addr_end(addr, end);
171
172 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
173 max_page_shift)) {
174 *mask |= PGTBL_PMD_MODIFIED;
175 continue;
176 }
177
178 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
179 return -ENOMEM;
180 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
181 return 0;
182 }
183
184 static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
185 phys_addr_t phys_addr, pgprot_t prot,
186 unsigned int max_page_shift)
187 {
188 if (max_page_shift < PUD_SHIFT)
189 return 0;
190
191 if (!arch_vmap_pud_supported(prot))
192 return 0;
193
194 if ((end - addr) != PUD_SIZE)
195 return 0;
196
197 if (!IS_ALIGNED(addr, PUD_SIZE))
198 return 0;
199
200 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
201 return 0;
202
203 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
204 return 0;
205
206 return pud_set_huge(pud, phys_addr, prot);
207 }
208
209 static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
210 phys_addr_t phys_addr, pgprot_t prot,
211 unsigned int max_page_shift, pgtbl_mod_mask *mask)
212 {
213 pud_t *pud;
214 unsigned long next;
215
216 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
217 if (!pud)
218 return -ENOMEM;
219 do {
220 next = pud_addr_end(addr, end);
221
222 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
223 max_page_shift)) {
224 *mask |= PGTBL_PUD_MODIFIED;
225 continue;
226 }
227
228 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
229 max_page_shift, mask))
230 return -ENOMEM;
231 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
232 return 0;
233 }
234
235 static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
236 phys_addr_t phys_addr, pgprot_t prot,
237 unsigned int max_page_shift)
238 {
239 if (max_page_shift < P4D_SHIFT)
240 return 0;
241
242 if (!arch_vmap_p4d_supported(prot))
243 return 0;
244
245 if ((end - addr) != P4D_SIZE)
246 return 0;
247
248 if (!IS_ALIGNED(addr, P4D_SIZE))
249 return 0;
250
251 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
252 return 0;
253
254 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
255 return 0;
256
257 return p4d_set_huge(p4d, phys_addr, prot);
258 }
259
260 static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
261 phys_addr_t phys_addr, pgprot_t prot,
262 unsigned int max_page_shift, pgtbl_mod_mask *mask)
263 {
264 p4d_t *p4d;
265 unsigned long next;
266
267 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
268 if (!p4d)
269 return -ENOMEM;
270 do {
271 next = p4d_addr_end(addr, end);
272
273 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
274 max_page_shift)) {
275 *mask |= PGTBL_P4D_MODIFIED;
276 continue;
277 }
278
279 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
280 max_page_shift, mask))
281 return -ENOMEM;
282 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
283 return 0;
284 }
285
286 static int vmap_range_noflush(unsigned long addr, unsigned long end,
287 phys_addr_t phys_addr, pgprot_t prot,
288 unsigned int max_page_shift)
289 {
290 pgd_t *pgd;
291 unsigned long start;
292 unsigned long next;
293 int err;
294 pgtbl_mod_mask mask = 0;
295
296 might_sleep();
297 BUG_ON(addr >= end);
298
299 start = addr;
300 pgd = pgd_offset_k(addr);
301 do {
302 next = pgd_addr_end(addr, end);
303 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
304 max_page_shift, &mask);
305 if (err)
306 break;
307 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
308
309 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
310 arch_sync_kernel_mappings(start, end);
311
312 return err;
313 }
314
315 int vmap_page_range(unsigned long addr, unsigned long end,
316 phys_addr_t phys_addr, pgprot_t prot)
317 {
318 int err;
319
320 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
321 ioremap_max_page_shift);
322 flush_cache_vmap(addr, end);
323 if (!err)
324 err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
325 ioremap_max_page_shift);
326 return err;
327 }
328
329 int ioremap_page_range(unsigned long addr, unsigned long end,
330 phys_addr_t phys_addr, pgprot_t prot)
331 {
332 struct vm_struct *area;
333
334 area = find_vm_area((void *)addr);
335 if (!area || !(area->flags & VM_IOREMAP)) {
336 WARN_ONCE(1, "vm_area at addr %lx is not marked as VM_IOREMAP\n", addr);
337 return -EINVAL;
338 }
339 if (addr != (unsigned long)area->addr ||
340 (void *)end != area->addr + get_vm_area_size(area)) {
341 WARN_ONCE(1, "ioremap request [%lx,%lx) doesn't match vm_area [%lx, %lx)\n",
342 addr, end, (long)area->addr,
343 (long)area->addr + get_vm_area_size(area));
344 return -ERANGE;
345 }
346 return vmap_page_range(addr, end, phys_addr, prot);
347 }
348
349 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
350 pgtbl_mod_mask *mask)
351 {
352 pte_t *pte;
353
354 pte = pte_offset_kernel(pmd, addr);
355 do {
356 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
357 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
358 } while (pte++, addr += PAGE_SIZE, addr != end);
359 *mask |= PGTBL_PTE_MODIFIED;
360 }
361
362 static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
363 pgtbl_mod_mask *mask)
364 {
365 pmd_t *pmd;
366 unsigned long next;
367 int cleared;
368
369 pmd = pmd_offset(pud, addr);
370 do {
371 next = pmd_addr_end(addr, end);
372
373 cleared = pmd_clear_huge(pmd);
374 if (cleared || pmd_bad(*pmd))
375 *mask |= PGTBL_PMD_MODIFIED;
376
377 if (cleared)
378 continue;
379 if (pmd_none_or_clear_bad(pmd))
380 continue;
381 vunmap_pte_range(pmd, addr, next, mask);
382
383 cond_resched();
384 } while (pmd++, addr = next, addr != end);
385 }
386
387 static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
388 pgtbl_mod_mask *mask)
389 {
390 pud_t *pud;
391 unsigned long next;
392 int cleared;
393
394 pud = pud_offset(p4d, addr);
395 do {
396 next = pud_addr_end(addr, end);
397
398 cleared = pud_clear_huge(pud);
399 if (cleared || pud_bad(*pud))
400 *mask |= PGTBL_PUD_MODIFIED;
401
402 if (cleared)
403 continue;
404 if (pud_none_or_clear_bad(pud))
405 continue;
406 vunmap_pmd_range(pud, addr, next, mask);
407 } while (pud++, addr = next, addr != end);
408 }
409
410 static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
411 pgtbl_mod_mask *mask)
412 {
413 p4d_t *p4d;
414 unsigned long next;
415
416 p4d = p4d_offset(pgd, addr);
417 do {
418 next = p4d_addr_end(addr, end);
419
420 p4d_clear_huge(p4d);
421 if (p4d_bad(*p4d))
422 *mask |= PGTBL_P4D_MODIFIED;
423
424 if (p4d_none_or_clear_bad(p4d))
425 continue;
426 vunmap_pud_range(p4d, addr, next, mask);
427 } while (p4d++, addr = next, addr != end);
428 }
429
430 /*
431 * vunmap_range_noflush is similar to vunmap_range, but does not
432 * flush caches or TLBs.
433 *
434 * The caller is responsible for calling flush_cache_vmap() before calling
435 * this function, and flush_tlb_kernel_range after it has returned
436 * successfully (and before the addresses are expected to cause a page fault
437 * or be re-mapped for something else, if TLB flushes are being delayed or
438 * coalesced).
439 *
440 * This is an internal function only. Do not use outside mm/.
441 */
442 void __vunmap_range_noflush(unsigned long start, unsigned long end)
443 {
444 unsigned long next;
445 pgd_t *pgd;
446 unsigned long addr = start;
447 pgtbl_mod_mask mask = 0;
448
449 BUG_ON(addr >= end);
450 pgd = pgd_offset_k(addr);
451 do {
452 next = pgd_addr_end(addr, end);
453 if (pgd_bad(*pgd))
454 mask |= PGTBL_PGD_MODIFIED;
455 if (pgd_none_or_clear_bad(pgd))
456 continue;
457 vunmap_p4d_range(pgd, addr, next, &mask);
458 } while (pgd++, addr = next, addr != end);
459
460 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
461 arch_sync_kernel_mappings(start, end);
462 }
463
464 void vunmap_range_noflush(unsigned long start, unsigned long end)
465 {
466 kmsan_vunmap_range_noflush(start, end);
467 __vunmap_range_noflush(start, end);
468 }
469
470 /**
471 * vunmap_range - unmap kernel virtual addresses
472 * @addr: start of the VM area to unmap
473 * @end: end of the VM area to unmap (non-inclusive)
474 *
475 * Clears any present PTEs in the virtual address range, flushes TLBs and
476 * caches. Any subsequent access to the address before it has been re-mapped
477 * is a kernel bug.
478 */
479 void vunmap_range(unsigned long addr, unsigned long end)
480 {
481 flush_cache_vunmap(addr, end);
482 vunmap_range_noflush(addr, end);
483 flush_tlb_kernel_range(addr, end);
484 }
485
486 static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
487 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
488 pgtbl_mod_mask *mask)
489 {
490 pte_t *pte;
491
492 /*
493 * nr is a running index into the array which helps higher level
494 * callers keep track of where we're up to.
495 */
496
497 pte = pte_alloc_kernel_track(pmd, addr, mask);
498 if (!pte)
499 return -ENOMEM;
500 do {
501 struct page *page = pages[*nr];
502
503 if (WARN_ON(!pte_none(ptep_get(pte))))
504 return -EBUSY;
505 if (WARN_ON(!page))
506 return -ENOMEM;
507 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
508 return -EINVAL;
509
510 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
511 (*nr)++;
512 } while (pte++, addr += PAGE_SIZE, addr != end);
513 *mask |= PGTBL_PTE_MODIFIED;
514 return 0;
515 }
516
517 static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
518 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
519 pgtbl_mod_mask *mask)
520 {
521 pmd_t *pmd;
522 unsigned long next;
523
524 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
525 if (!pmd)
526 return -ENOMEM;
527 do {
528 next = pmd_addr_end(addr, end);
529 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
530 return -ENOMEM;
531 } while (pmd++, addr = next, addr != end);
532 return 0;
533 }
534
535 static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
536 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
537 pgtbl_mod_mask *mask)
538 {
539 pud_t *pud;
540 unsigned long next;
541
542 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
543 if (!pud)
544 return -ENOMEM;
545 do {
546 next = pud_addr_end(addr, end);
547 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
548 return -ENOMEM;
549 } while (pud++, addr = next, addr != end);
550 return 0;
551 }
552
553 static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
554 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
555 pgtbl_mod_mask *mask)
556 {
557 p4d_t *p4d;
558 unsigned long next;
559
560 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
561 if (!p4d)
562 return -ENOMEM;
563 do {
564 next = p4d_addr_end(addr, end);
565 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
566 return -ENOMEM;
567 } while (p4d++, addr = next, addr != end);
568 return 0;
569 }
570
571 static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
572 pgprot_t prot, struct page **pages)
573 {
574 unsigned long start = addr;
575 pgd_t *pgd;
576 unsigned long next;
577 int err = 0;
578 int nr = 0;
579 pgtbl_mod_mask mask = 0;
580
581 BUG_ON(addr >= end);
582 pgd = pgd_offset_k(addr);
583 do {
584 next = pgd_addr_end(addr, end);
585 if (pgd_bad(*pgd))
586 mask |= PGTBL_PGD_MODIFIED;
587 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
588 if (err)
589 return err;
590 } while (pgd++, addr = next, addr != end);
591
592 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
593 arch_sync_kernel_mappings(start, end);
594
595 return 0;
596 }
597
598 /*
599 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
600 * flush caches.
601 *
602 * The caller is responsible for calling flush_cache_vmap() after this
603 * function returns successfully and before the addresses are accessed.
604 *
605 * This is an internal function only. Do not use outside mm/.
606 */
607 int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
608 pgprot_t prot, struct page **pages, unsigned int page_shift)
609 {
610 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
611
612 WARN_ON(page_shift < PAGE_SHIFT);
613
614 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
615 page_shift == PAGE_SHIFT)
616 return vmap_small_pages_range_noflush(addr, end, prot, pages);
617
618 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
619 int err;
620
621 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
622 page_to_phys(pages[i]), prot,
623 page_shift);
624 if (err)
625 return err;
626
627 addr += 1UL << page_shift;
628 }
629
630 return 0;
631 }
632
633 int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
634 pgprot_t prot, struct page **pages, unsigned int page_shift)
635 {
636 int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
637 page_shift);
638
639 if (ret)
640 return ret;
641 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
642 }
643
644 /**
645 * vmap_pages_range - map pages to a kernel virtual address
646 * @addr: start of the VM area to map
647 * @end: end of the VM area to map (non-inclusive)
648 * @prot: page protection flags to use
649 * @pages: pages to map (always PAGE_SIZE pages)
650 * @page_shift: maximum shift that the pages may be mapped with, @pages must
651 * be aligned and contiguous up to at least this shift.
652 *
653 * RETURNS:
654 * 0 on success, -errno on failure.
655 */
656 static int vmap_pages_range(unsigned long addr, unsigned long end,
657 pgprot_t prot, struct page **pages, unsigned int page_shift)
658 {
659 int err;
660
661 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
662 flush_cache_vmap(addr, end);
663 return err;
664 }
665
666 static int check_sparse_vm_area(struct vm_struct *area, unsigned long start,
667 unsigned long end)
668 {
669 might_sleep();
670 if (WARN_ON_ONCE(area->flags & VM_FLUSH_RESET_PERMS))
671 return -EINVAL;
672 if (WARN_ON_ONCE(area->flags & VM_NO_GUARD))
673 return -EINVAL;
674 if (WARN_ON_ONCE(!(area->flags & VM_SPARSE)))
675 return -EINVAL;
676 if ((end - start) >> PAGE_SHIFT > totalram_pages())
677 return -E2BIG;
678 if (start < (unsigned long)area->addr ||
679 (void *)end > area->addr + get_vm_area_size(area))
680 return -ERANGE;
681 return 0;
682 }
683
684 /**
685 * vm_area_map_pages - map pages inside given sparse vm_area
686 * @area: vm_area
687 * @start: start address inside vm_area
688 * @end: end address inside vm_area
689 * @pages: pages to map (always PAGE_SIZE pages)
690 */
691 int vm_area_map_pages(struct vm_struct *area, unsigned long start,
692 unsigned long end, struct page **pages)
693 {
694 int err;
695
696 err = check_sparse_vm_area(area, start, end);
697 if (err)
698 return err;
699
700 return vmap_pages_range(start, end, PAGE_KERNEL, pages, PAGE_SHIFT);
701 }
702
703 /**
704 * vm_area_unmap_pages - unmap pages inside given sparse vm_area
705 * @area: vm_area
706 * @start: start address inside vm_area
707 * @end: end address inside vm_area
708 */
709 void vm_area_unmap_pages(struct vm_struct *area, unsigned long start,
710 unsigned long end)
711 {
712 if (check_sparse_vm_area(area, start, end))
713 return;
714
715 vunmap_range(start, end);
716 }
717
718 int is_vmalloc_or_module_addr(const void *x)
719 {
720 /*
721 * ARM, x86-64 and sparc64 put modules in a special place,
722 * and fall back on vmalloc() if that fails. Others
723 * just put it in the vmalloc space.
724 */
725 #if defined(CONFIG_EXECMEM) && defined(MODULES_VADDR)
726 unsigned long addr = (unsigned long)kasan_reset_tag(x);
727 if (addr >= MODULES_VADDR && addr < MODULES_END)
728 return 1;
729 #endif
730 return is_vmalloc_addr(x);
731 }
732 EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
733
734 /*
735 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
736 * return the tail page that corresponds to the base page address, which
737 * matches small vmap mappings.
738 */
739 struct page *vmalloc_to_page(const void *vmalloc_addr)
740 {
741 unsigned long addr = (unsigned long) vmalloc_addr;
742 struct page *page = NULL;
743 pgd_t *pgd = pgd_offset_k(addr);
744 p4d_t *p4d;
745 pud_t *pud;
746 pmd_t *pmd;
747 pte_t *ptep, pte;
748
749 /*
750 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
751 * architectures that do not vmalloc module space
752 */
753 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
754
755 if (pgd_none(*pgd))
756 return NULL;
757 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
758 return NULL; /* XXX: no allowance for huge pgd */
759 if (WARN_ON_ONCE(pgd_bad(*pgd)))
760 return NULL;
761
762 p4d = p4d_offset(pgd, addr);
763 if (p4d_none(*p4d))
764 return NULL;
765 if (p4d_leaf(*p4d))
766 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
767 if (WARN_ON_ONCE(p4d_bad(*p4d)))
768 return NULL;
769
770 pud = pud_offset(p4d, addr);
771 if (pud_none(*pud))
772 return NULL;
773 if (pud_leaf(*pud))
774 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
775 if (WARN_ON_ONCE(pud_bad(*pud)))
776 return NULL;
777
778 pmd = pmd_offset(pud, addr);
779 if (pmd_none(*pmd))
780 return NULL;
781 if (pmd_leaf(*pmd))
782 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
783 if (WARN_ON_ONCE(pmd_bad(*pmd)))
784 return NULL;
785
786 ptep = pte_offset_kernel(pmd, addr);
787 pte = ptep_get(ptep);
788 if (pte_present(pte))
789 page = pte_page(pte);
790
791 return page;
792 }
793 EXPORT_SYMBOL(vmalloc_to_page);
794
795 /*
796 * Map a vmalloc()-space virtual address to the physical page frame number.
797 */
798 unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
799 {
800 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
801 }
802 EXPORT_SYMBOL(vmalloc_to_pfn);
803
804
805 /*** Global kva allocator ***/
806
807 #define DEBUG_AUGMENT_PROPAGATE_CHECK 0
808 #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
809
810
811 static DEFINE_SPINLOCK(free_vmap_area_lock);
812 static bool vmap_initialized __read_mostly;
813
814 /*
815 * This kmem_cache is used for vmap_area objects. Instead of
816 * allocating from slab we reuse an object from this cache to
817 * make things faster. Especially in "no edge" splitting of
818 * free block.
819 */
820 static struct kmem_cache *vmap_area_cachep;
821
822 /*
823 * This linked list is used in pair with free_vmap_area_root.
824 * It gives O(1) access to prev/next to perform fast coalescing.
825 */
826 static LIST_HEAD(free_vmap_area_list);
827
828 /*
829 * This augment red-black tree represents the free vmap space.
830 * All vmap_area objects in this tree are sorted by va->va_start
831 * address. It is used for allocation and merging when a vmap
832 * object is released.
833 *
834 * Each vmap_area node contains a maximum available free block
835 * of its sub-tree, right or left. Therefore it is possible to
836 * find a lowest match of free area.
837 */
838 static struct rb_root free_vmap_area_root = RB_ROOT;
839
840 /*
841 * Preload a CPU with one object for "no edge" split case. The
842 * aim is to get rid of allocations from the atomic context, thus
843 * to use more permissive allocation masks.
844 */
845 static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
846
847 /*
848 * This structure defines a single, solid model where a list and
849 * rb-tree are part of one entity protected by the lock. Nodes are
850 * sorted in ascending order, thus for O(1) access to left/right
851 * neighbors a list is used as well as for sequential traversal.
852 */
853 struct rb_list {
854 struct rb_root root;
855 struct list_head head;
856 spinlock_t lock;
857 };
858
859 /*
860 * A fast size storage contains VAs up to 1M size. A pool consists
861 * of linked between each other ready to go VAs of certain sizes.
862 * An index in the pool-array corresponds to number of pages + 1.
863 */
864 #define MAX_VA_SIZE_PAGES 256
865
866 struct vmap_pool {
867 struct list_head head;
868 unsigned long len;
869 };
870
871 /*
872 * An effective vmap-node logic. Users make use of nodes instead
873 * of a global heap. It allows to balance an access and mitigate
874 * contention.
875 */
876 static struct vmap_node {
877 /* Simple size segregated storage. */
878 struct vmap_pool pool[MAX_VA_SIZE_PAGES];
879 spinlock_t pool_lock;
880 bool skip_populate;
881
882 /* Bookkeeping data of this node. */
883 struct rb_list busy;
884 struct rb_list lazy;
885
886 /*
887 * Ready-to-free areas.
888 */
889 struct list_head purge_list;
890 struct work_struct purge_work;
891 unsigned long nr_purged;
892 } single;
893
894 /*
895 * Initial setup consists of one single node, i.e. a balancing
896 * is fully disabled. Later on, after vmap is initialized these
897 * parameters are updated based on a system capacity.
898 */
899 static struct vmap_node *vmap_nodes = &single;
900 static __read_mostly unsigned int nr_vmap_nodes = 1;
901 static __read_mostly unsigned int vmap_zone_size = 1;
902
903 static inline unsigned int
904 addr_to_node_id(unsigned long addr)
905 {
906 return (addr / vmap_zone_size) % nr_vmap_nodes;
907 }
908
909 static inline struct vmap_node *
910 addr_to_node(unsigned long addr)
911 {
912 return &vmap_nodes[addr_to_node_id(addr)];
913 }
914
915 static inline struct vmap_node *
916 id_to_node(unsigned int id)
917 {
918 return &vmap_nodes[id % nr_vmap_nodes];
919 }
920
921 /*
922 * We use the value 0 to represent "no node", that is why
923 * an encoded value will be the node-id incremented by 1.
924 * It is always greater then 0. A valid node_id which can
925 * be encoded is [0:nr_vmap_nodes - 1]. If a passed node_id
926 * is not valid 0 is returned.
927 */
928 static unsigned int
929 encode_vn_id(unsigned int node_id)
930 {
931 /* Can store U8_MAX [0:254] nodes. */
932 if (node_id < nr_vmap_nodes)
933 return (node_id + 1) << BITS_PER_BYTE;
934
935 /* Warn and no node encoded. */
936 WARN_ONCE(1, "Encode wrong node id (%u)\n", node_id);
937 return 0;
938 }
939
940 /*
941 * Returns an encoded node-id, the valid range is within
942 * [0:nr_vmap_nodes-1] values. Otherwise nr_vmap_nodes is
943 * returned if extracted data is wrong.
944 */
945 static unsigned int
946 decode_vn_id(unsigned int val)
947 {
948 unsigned int node_id = (val >> BITS_PER_BYTE) - 1;
949
950 /* Can store U8_MAX [0:254] nodes. */
951 if (node_id < nr_vmap_nodes)
952 return node_id;
953
954 /* If it was _not_ zero, warn. */
955 WARN_ONCE(node_id != UINT_MAX,
956 "Decode wrong node id (%d)\n", node_id);
957
958 return nr_vmap_nodes;
959 }
960
961 static bool
962 is_vn_id_valid(unsigned int node_id)
963 {
964 if (node_id < nr_vmap_nodes)
965 return true;
966
967 return false;
968 }
969
970 static __always_inline unsigned long
971 va_size(struct vmap_area *va)
972 {
973 return (va->va_end - va->va_start);
974 }
975
976 static __always_inline unsigned long
977 get_subtree_max_size(struct rb_node *node)
978 {
979 struct vmap_area *va;
980
981 va = rb_entry_safe(node, struct vmap_area, rb_node);
982 return va ? va->subtree_max_size : 0;
983 }
984
985 RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
986 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
987
988 static void reclaim_and_purge_vmap_areas(void);
989 static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
990 static void drain_vmap_area_work(struct work_struct *work);
991 static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
992
993 static atomic_long_t nr_vmalloc_pages;
994
995 unsigned long vmalloc_nr_pages(void)
996 {
997 return atomic_long_read(&nr_vmalloc_pages);
998 }
999
1000 static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
1001 {
1002 struct rb_node *n = root->rb_node;
1003
1004 addr = (unsigned long)kasan_reset_tag((void *)addr);
1005
1006 while (n) {
1007 struct vmap_area *va;
1008
1009 va = rb_entry(n, struct vmap_area, rb_node);
1010 if (addr < va->va_start)
1011 n = n->rb_left;
1012 else if (addr >= va->va_end)
1013 n = n->rb_right;
1014 else
1015 return va;
1016 }
1017
1018 return NULL;
1019 }
1020
1021 /* Look up the first VA which satisfies addr < va_end, NULL if none. */
1022 static struct vmap_area *
1023 __find_vmap_area_exceed_addr(unsigned long addr, struct rb_root *root)
1024 {
1025 struct vmap_area *va = NULL;
1026 struct rb_node *n = root->rb_node;
1027
1028 addr = (unsigned long)kasan_reset_tag((void *)addr);
1029
1030 while (n) {
1031 struct vmap_area *tmp;
1032
1033 tmp = rb_entry(n, struct vmap_area, rb_node);
1034 if (tmp->va_end > addr) {
1035 va = tmp;
1036 if (tmp->va_start <= addr)
1037 break;
1038
1039 n = n->rb_left;
1040 } else
1041 n = n->rb_right;
1042 }
1043
1044 return va;
1045 }
1046
1047 /*
1048 * Returns a node where a first VA, that satisfies addr < va_end, resides.
1049 * If success, a node is locked. A user is responsible to unlock it when a
1050 * VA is no longer needed to be accessed.
1051 *
1052 * Returns NULL if nothing found.
1053 */
1054 static struct vmap_node *
1055 find_vmap_area_exceed_addr_lock(unsigned long addr, struct vmap_area **va)
1056 {
1057 unsigned long va_start_lowest;
1058 struct vmap_node *vn;
1059 int i;
1060
1061 repeat:
1062 for (i = 0, va_start_lowest = 0; i < nr_vmap_nodes; i++) {
1063 vn = &vmap_nodes[i];
1064
1065 spin_lock(&vn->busy.lock);
1066 *va = __find_vmap_area_exceed_addr(addr, &vn->busy.root);
1067
1068 if (*va)
1069 if (!va_start_lowest || (*va)->va_start < va_start_lowest)
1070 va_start_lowest = (*va)->va_start;
1071 spin_unlock(&vn->busy.lock);
1072 }
1073
1074 /*
1075 * Check if found VA exists, it might have gone away. In this case we
1076 * repeat the search because a VA has been removed concurrently and we
1077 * need to proceed to the next one, which is a rare case.
1078 */
1079 if (va_start_lowest) {
1080 vn = addr_to_node(va_start_lowest);
1081
1082 spin_lock(&vn->busy.lock);
1083 *va = __find_vmap_area(va_start_lowest, &vn->busy.root);
1084
1085 if (*va)
1086 return vn;
1087
1088 spin_unlock(&vn->busy.lock);
1089 goto repeat;
1090 }
1091
1092 return NULL;
1093 }
1094
1095 /*
1096 * This function returns back addresses of parent node
1097 * and its left or right link for further processing.
1098 *
1099 * Otherwise NULL is returned. In that case all further
1100 * steps regarding inserting of conflicting overlap range
1101 * have to be declined and actually considered as a bug.
1102 */
1103 static __always_inline struct rb_node **
1104 find_va_links(struct vmap_area *va,
1105 struct rb_root *root, struct rb_node *from,
1106 struct rb_node **parent)
1107 {
1108 struct vmap_area *tmp_va;
1109 struct rb_node **link;
1110
1111 if (root) {
1112 link = &root->rb_node;
1113 if (unlikely(!*link)) {
1114 *parent = NULL;
1115 return link;
1116 }
1117 } else {
1118 link = &from;
1119 }
1120
1121 /*
1122 * Go to the bottom of the tree. When we hit the last point
1123 * we end up with parent rb_node and correct direction, i name
1124 * it link, where the new va->rb_node will be attached to.
1125 */
1126 do {
1127 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
1128
1129 /*
1130 * During the traversal we also do some sanity check.
1131 * Trigger the BUG() if there are sides(left/right)
1132 * or full overlaps.
1133 */
1134 if (va->va_end <= tmp_va->va_start)
1135 link = &(*link)->rb_left;
1136 else if (va->va_start >= tmp_va->va_end)
1137 link = &(*link)->rb_right;
1138 else {
1139 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
1140 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
1141
1142 return NULL;
1143 }
1144 } while (*link);
1145
1146 *parent = &tmp_va->rb_node;
1147 return link;
1148 }
1149
1150 static __always_inline struct list_head *
1151 get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
1152 {
1153 struct list_head *list;
1154
1155 if (unlikely(!parent))
1156 /*
1157 * The red-black tree where we try to find VA neighbors
1158 * before merging or inserting is empty, i.e. it means
1159 * there is no free vmap space. Normally it does not
1160 * happen but we handle this case anyway.
1161 */
1162 return NULL;
1163
1164 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
1165 return (&parent->rb_right == link ? list->next : list);
1166 }
1167
1168 static __always_inline void
1169 __link_va(struct vmap_area *va, struct rb_root *root,
1170 struct rb_node *parent, struct rb_node **link,
1171 struct list_head *head, bool augment)
1172 {
1173 /*
1174 * VA is still not in the list, but we can
1175 * identify its future previous list_head node.
1176 */
1177 if (likely(parent)) {
1178 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
1179 if (&parent->rb_right != link)
1180 head = head->prev;
1181 }
1182
1183 /* Insert to the rb-tree */
1184 rb_link_node(&va->rb_node, parent, link);
1185 if (augment) {
1186 /*
1187 * Some explanation here. Just perform simple insertion
1188 * to the tree. We do not set va->subtree_max_size to
1189 * its current size before calling rb_insert_augmented().
1190 * It is because we populate the tree from the bottom
1191 * to parent levels when the node _is_ in the tree.
1192 *
1193 * Therefore we set subtree_max_size to zero after insertion,
1194 * to let __augment_tree_propagate_from() puts everything to
1195 * the correct order later on.
1196 */
1197 rb_insert_augmented(&va->rb_node,
1198 root, &free_vmap_area_rb_augment_cb);
1199 va->subtree_max_size = 0;
1200 } else {
1201 rb_insert_color(&va->rb_node, root);
1202 }
1203
1204 /* Address-sort this list */
1205 list_add(&va->list, head);
1206 }
1207
1208 static __always_inline void
1209 link_va(struct vmap_area *va, struct rb_root *root,
1210 struct rb_node *parent, struct rb_node **link,
1211 struct list_head *head)
1212 {
1213 __link_va(va, root, parent, link, head, false);
1214 }
1215
1216 static __always_inline void
1217 link_va_augment(struct vmap_area *va, struct rb_root *root,
1218 struct rb_node *parent, struct rb_node **link,
1219 struct list_head *head)
1220 {
1221 __link_va(va, root, parent, link, head, true);
1222 }
1223
1224 static __always_inline void
1225 __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
1226 {
1227 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
1228 return;
1229
1230 if (augment)
1231 rb_erase_augmented(&va->rb_node,
1232 root, &free_vmap_area_rb_augment_cb);
1233 else
1234 rb_erase(&va->rb_node, root);
1235
1236 list_del_init(&va->list);
1237 RB_CLEAR_NODE(&va->rb_node);
1238 }
1239
1240 static __always_inline void
1241 unlink_va(struct vmap_area *va, struct rb_root *root)
1242 {
1243 __unlink_va(va, root, false);
1244 }
1245
1246 static __always_inline void
1247 unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1248 {
1249 __unlink_va(va, root, true);
1250 }
1251
1252 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1253 /*
1254 * Gets called when remove the node and rotate.
1255 */
1256 static __always_inline unsigned long
1257 compute_subtree_max_size(struct vmap_area *va)
1258 {
1259 return max3(va_size(va),
1260 get_subtree_max_size(va->rb_node.rb_left),
1261 get_subtree_max_size(va->rb_node.rb_right));
1262 }
1263
1264 static void
1265 augment_tree_propagate_check(void)
1266 {
1267 struct vmap_area *va;
1268 unsigned long computed_size;
1269
1270 list_for_each_entry(va, &free_vmap_area_list, list) {
1271 computed_size = compute_subtree_max_size(va);
1272 if (computed_size != va->subtree_max_size)
1273 pr_emerg("tree is corrupted: %lu, %lu\n",
1274 va_size(va), va->subtree_max_size);
1275 }
1276 }
1277 #endif
1278
1279 /*
1280 * This function populates subtree_max_size from bottom to upper
1281 * levels starting from VA point. The propagation must be done
1282 * when VA size is modified by changing its va_start/va_end. Or
1283 * in case of newly inserting of VA to the tree.
1284 *
1285 * It means that __augment_tree_propagate_from() must be called:
1286 * - After VA has been inserted to the tree(free path);
1287 * - After VA has been shrunk(allocation path);
1288 * - After VA has been increased(merging path).
1289 *
1290 * Please note that, it does not mean that upper parent nodes
1291 * and their subtree_max_size are recalculated all the time up
1292 * to the root node.
1293 *
1294 * 4--8
1295 * /\
1296 * / \
1297 * / \
1298 * 2--2 8--8
1299 *
1300 * For example if we modify the node 4, shrinking it to 2, then
1301 * no any modification is required. If we shrink the node 2 to 1
1302 * its subtree_max_size is updated only, and set to 1. If we shrink
1303 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1304 * node becomes 4--6.
1305 */
1306 static __always_inline void
1307 augment_tree_propagate_from(struct vmap_area *va)
1308 {
1309 /*
1310 * Populate the tree from bottom towards the root until
1311 * the calculated maximum available size of checked node
1312 * is equal to its current one.
1313 */
1314 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1315
1316 #if DEBUG_AUGMENT_PROPAGATE_CHECK
1317 augment_tree_propagate_check();
1318 #endif
1319 }
1320
1321 static void
1322 insert_vmap_area(struct vmap_area *va,
1323 struct rb_root *root, struct list_head *head)
1324 {
1325 struct rb_node **link;
1326 struct rb_node *parent;
1327
1328 link = find_va_links(va, root, NULL, &parent);
1329 if (link)
1330 link_va(va, root, parent, link, head);
1331 }
1332
1333 static void
1334 insert_vmap_area_augment(struct vmap_area *va,
1335 struct rb_node *from, struct rb_root *root,
1336 struct list_head *head)
1337 {
1338 struct rb_node **link;
1339 struct rb_node *parent;
1340
1341 if (from)
1342 link = find_va_links(va, NULL, from, &parent);
1343 else
1344 link = find_va_links(va, root, NULL, &parent);
1345
1346 if (link) {
1347 link_va_augment(va, root, parent, link, head);
1348 augment_tree_propagate_from(va);
1349 }
1350 }
1351
1352 /*
1353 * Merge de-allocated chunk of VA memory with previous
1354 * and next free blocks. If coalesce is not done a new
1355 * free area is inserted. If VA has been merged, it is
1356 * freed.
1357 *
1358 * Please note, it can return NULL in case of overlap
1359 * ranges, followed by WARN() report. Despite it is a
1360 * buggy behaviour, a system can be alive and keep
1361 * ongoing.
1362 */
1363 static __always_inline struct vmap_area *
1364 __merge_or_add_vmap_area(struct vmap_area *va,
1365 struct rb_root *root, struct list_head *head, bool augment)
1366 {
1367 struct vmap_area *sibling;
1368 struct list_head *next;
1369 struct rb_node **link;
1370 struct rb_node *parent;
1371 bool merged = false;
1372
1373 /*
1374 * Find a place in the tree where VA potentially will be
1375 * inserted, unless it is merged with its sibling/siblings.
1376 */
1377 link = find_va_links(va, root, NULL, &parent);
1378 if (!link)
1379 return NULL;
1380
1381 /*
1382 * Get next node of VA to check if merging can be done.
1383 */
1384 next = get_va_next_sibling(parent, link);
1385 if (unlikely(next == NULL))
1386 goto insert;
1387
1388 /*
1389 * start end
1390 * | |
1391 * |<------VA------>|<-----Next----->|
1392 * | |
1393 * start end
1394 */
1395 if (next != head) {
1396 sibling = list_entry(next, struct vmap_area, list);
1397 if (sibling->va_start == va->va_end) {
1398 sibling->va_start = va->va_start;
1399
1400 /* Free vmap_area object. */
1401 kmem_cache_free(vmap_area_cachep, va);
1402
1403 /* Point to the new merged area. */
1404 va = sibling;
1405 merged = true;
1406 }
1407 }
1408
1409 /*
1410 * start end
1411 * | |
1412 * |<-----Prev----->|<------VA------>|
1413 * | |
1414 * start end
1415 */
1416 if (next->prev != head) {
1417 sibling = list_entry(next->prev, struct vmap_area, list);
1418 if (sibling->va_end == va->va_start) {
1419 /*
1420 * If both neighbors are coalesced, it is important
1421 * to unlink the "next" node first, followed by merging
1422 * with "previous" one. Otherwise the tree might not be
1423 * fully populated if a sibling's augmented value is
1424 * "normalized" because of rotation operations.
1425 */
1426 if (merged)
1427 __unlink_va(va, root, augment);
1428
1429 sibling->va_end = va->va_end;
1430
1431 /* Free vmap_area object. */
1432 kmem_cache_free(vmap_area_cachep, va);
1433
1434 /* Point to the new merged area. */
1435 va = sibling;
1436 merged = true;
1437 }
1438 }
1439
1440 insert:
1441 if (!merged)
1442 __link_va(va, root, parent, link, head, augment);
1443
1444 return va;
1445 }
1446
1447 static __always_inline struct vmap_area *
1448 merge_or_add_vmap_area(struct vmap_area *va,
1449 struct rb_root *root, struct list_head *head)
1450 {
1451 return __merge_or_add_vmap_area(va, root, head, false);
1452 }
1453
1454 static __always_inline struct vmap_area *
1455 merge_or_add_vmap_area_augment(struct vmap_area *va,
1456 struct rb_root *root, struct list_head *head)
1457 {
1458 va = __merge_or_add_vmap_area(va, root, head, true);
1459 if (va)
1460 augment_tree_propagate_from(va);
1461
1462 return va;
1463 }
1464
1465 static __always_inline bool
1466 is_within_this_va(struct vmap_area *va, unsigned long size,
1467 unsigned long align, unsigned long vstart)
1468 {
1469 unsigned long nva_start_addr;
1470
1471 if (va->va_start > vstart)
1472 nva_start_addr = ALIGN(va->va_start, align);
1473 else
1474 nva_start_addr = ALIGN(vstart, align);
1475
1476 /* Can be overflowed due to big size or alignment. */
1477 if (nva_start_addr + size < nva_start_addr ||
1478 nva_start_addr < vstart)
1479 return false;
1480
1481 return (nva_start_addr + size <= va->va_end);
1482 }
1483
1484 /*
1485 * Find the first free block(lowest start address) in the tree,
1486 * that will accomplish the request corresponding to passing
1487 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1488 * a search length is adjusted to account for worst case alignment
1489 * overhead.
1490 */
1491 static __always_inline struct vmap_area *
1492 find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1493 unsigned long align, unsigned long vstart, bool adjust_search_size)
1494 {
1495 struct vmap_area *va;
1496 struct rb_node *node;
1497 unsigned long length;
1498
1499 /* Start from the root. */
1500 node = root->rb_node;
1501
1502 /* Adjust the search size for alignment overhead. */
1503 length = adjust_search_size ? size + align - 1 : size;
1504
1505 while (node) {
1506 va = rb_entry(node, struct vmap_area, rb_node);
1507
1508 if (get_subtree_max_size(node->rb_left) >= length &&
1509 vstart < va->va_start) {
1510 node = node->rb_left;
1511 } else {
1512 if (is_within_this_va(va, size, align, vstart))
1513 return va;
1514
1515 /*
1516 * Does not make sense to go deeper towards the right
1517 * sub-tree if it does not have a free block that is
1518 * equal or bigger to the requested search length.
1519 */
1520 if (get_subtree_max_size(node->rb_right) >= length) {
1521 node = node->rb_right;
1522 continue;
1523 }
1524
1525 /*
1526 * OK. We roll back and find the first right sub-tree,
1527 * that will satisfy the search criteria. It can happen
1528 * due to "vstart" restriction or an alignment overhead
1529 * that is bigger then PAGE_SIZE.
1530 */
1531 while ((node = rb_parent(node))) {
1532 va = rb_entry(node, struct vmap_area, rb_node);
1533 if (is_within_this_va(va, size, align, vstart))
1534 return va;
1535
1536 if (get_subtree_max_size(node->rb_right) >= length &&
1537 vstart <= va->va_start) {
1538 /*
1539 * Shift the vstart forward. Please note, we update it with
1540 * parent's start address adding "1" because we do not want
1541 * to enter same sub-tree after it has already been checked
1542 * and no suitable free block found there.
1543 */
1544 vstart = va->va_start + 1;
1545 node = node->rb_right;
1546 break;
1547 }
1548 }
1549 }
1550 }
1551
1552 return NULL;
1553 }
1554
1555 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1556 #include <linux/random.h>
1557
1558 static struct vmap_area *
1559 find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1560 unsigned long align, unsigned long vstart)
1561 {
1562 struct vmap_area *va;
1563
1564 list_for_each_entry(va, head, list) {
1565 if (!is_within_this_va(va, size, align, vstart))
1566 continue;
1567
1568 return va;
1569 }
1570
1571 return NULL;
1572 }
1573
1574 static void
1575 find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1576 unsigned long size, unsigned long align)
1577 {
1578 struct vmap_area *va_1, *va_2;
1579 unsigned long vstart;
1580 unsigned int rnd;
1581
1582 get_random_bytes(&rnd, sizeof(rnd));
1583 vstart = VMALLOC_START + rnd;
1584
1585 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1586 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1587
1588 if (va_1 != va_2)
1589 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1590 va_1, va_2, vstart);
1591 }
1592 #endif
1593
1594 enum fit_type {
1595 NOTHING_FIT = 0,
1596 FL_FIT_TYPE = 1, /* full fit */
1597 LE_FIT_TYPE = 2, /* left edge fit */
1598 RE_FIT_TYPE = 3, /* right edge fit */
1599 NE_FIT_TYPE = 4 /* no edge fit */
1600 };
1601
1602 static __always_inline enum fit_type
1603 classify_va_fit_type(struct vmap_area *va,
1604 unsigned long nva_start_addr, unsigned long size)
1605 {
1606 enum fit_type type;
1607
1608 /* Check if it is within VA. */
1609 if (nva_start_addr < va->va_start ||
1610 nva_start_addr + size > va->va_end)
1611 return NOTHING_FIT;
1612
1613 /* Now classify. */
1614 if (va->va_start == nva_start_addr) {
1615 if (va->va_end == nva_start_addr + size)
1616 type = FL_FIT_TYPE;
1617 else
1618 type = LE_FIT_TYPE;
1619 } else if (va->va_end == nva_start_addr + size) {
1620 type = RE_FIT_TYPE;
1621 } else {
1622 type = NE_FIT_TYPE;
1623 }
1624
1625 return type;
1626 }
1627
1628 static __always_inline int
1629 va_clip(struct rb_root *root, struct list_head *head,
1630 struct vmap_area *va, unsigned long nva_start_addr,
1631 unsigned long size)
1632 {
1633 struct vmap_area *lva = NULL;
1634 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1635
1636 if (type == FL_FIT_TYPE) {
1637 /*
1638 * No need to split VA, it fully fits.
1639 *
1640 * | |
1641 * V NVA V
1642 * |---------------|
1643 */
1644 unlink_va_augment(va, root);
1645 kmem_cache_free(vmap_area_cachep, va);
1646 } else if (type == LE_FIT_TYPE) {
1647 /*
1648 * Split left edge of fit VA.
1649 *
1650 * | |
1651 * V NVA V R
1652 * |-------|-------|
1653 */
1654 va->va_start += size;
1655 } else if (type == RE_FIT_TYPE) {
1656 /*
1657 * Split right edge of fit VA.
1658 *
1659 * | |
1660 * L V NVA V
1661 * |-------|-------|
1662 */
1663 va->va_end = nva_start_addr;
1664 } else if (type == NE_FIT_TYPE) {
1665 /*
1666 * Split no edge of fit VA.
1667 *
1668 * | |
1669 * L V NVA V R
1670 * |---|-------|---|
1671 */
1672 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1673 if (unlikely(!lva)) {
1674 /*
1675 * For percpu allocator we do not do any pre-allocation
1676 * and leave it as it is. The reason is it most likely
1677 * never ends up with NE_FIT_TYPE splitting. In case of
1678 * percpu allocations offsets and sizes are aligned to
1679 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1680 * are its main fitting cases.
1681 *
1682 * There are a few exceptions though, as an example it is
1683 * a first allocation (early boot up) when we have "one"
1684 * big free space that has to be split.
1685 *
1686 * Also we can hit this path in case of regular "vmap"
1687 * allocations, if "this" current CPU was not preloaded.
1688 * See the comment in alloc_vmap_area() why. If so, then
1689 * GFP_NOWAIT is used instead to get an extra object for
1690 * split purpose. That is rare and most time does not
1691 * occur.
1692 *
1693 * What happens if an allocation gets failed. Basically,
1694 * an "overflow" path is triggered to purge lazily freed
1695 * areas to free some memory, then, the "retry" path is
1696 * triggered to repeat one more time. See more details
1697 * in alloc_vmap_area() function.
1698 */
1699 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1700 if (!lva)
1701 return -1;
1702 }
1703
1704 /*
1705 * Build the remainder.
1706 */
1707 lva->va_start = va->va_start;
1708 lva->va_end = nva_start_addr;
1709
1710 /*
1711 * Shrink this VA to remaining size.
1712 */
1713 va->va_start = nva_start_addr + size;
1714 } else {
1715 return -1;
1716 }
1717
1718 if (type != FL_FIT_TYPE) {
1719 augment_tree_propagate_from(va);
1720
1721 if (lva) /* type == NE_FIT_TYPE */
1722 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1723 }
1724
1725 return 0;
1726 }
1727
1728 static unsigned long
1729 va_alloc(struct vmap_area *va,
1730 struct rb_root *root, struct list_head *head,
1731 unsigned long size, unsigned long align,
1732 unsigned long vstart, unsigned long vend)
1733 {
1734 unsigned long nva_start_addr;
1735 int ret;
1736
1737 if (va->va_start > vstart)
1738 nva_start_addr = ALIGN(va->va_start, align);
1739 else
1740 nva_start_addr = ALIGN(vstart, align);
1741
1742 /* Check the "vend" restriction. */
1743 if (nva_start_addr + size > vend)
1744 return vend;
1745
1746 /* Update the free vmap_area. */
1747 ret = va_clip(root, head, va, nva_start_addr, size);
1748 if (WARN_ON_ONCE(ret))
1749 return vend;
1750
1751 return nva_start_addr;
1752 }
1753
1754 /*
1755 * Returns a start address of the newly allocated area, if success.
1756 * Otherwise a vend is returned that indicates failure.
1757 */
1758 static __always_inline unsigned long
1759 __alloc_vmap_area(struct rb_root *root, struct list_head *head,
1760 unsigned long size, unsigned long align,
1761 unsigned long vstart, unsigned long vend)
1762 {
1763 bool adjust_search_size = true;
1764 unsigned long nva_start_addr;
1765 struct vmap_area *va;
1766
1767 /*
1768 * Do not adjust when:
1769 * a) align <= PAGE_SIZE, because it does not make any sense.
1770 * All blocks(their start addresses) are at least PAGE_SIZE
1771 * aligned anyway;
1772 * b) a short range where a requested size corresponds to exactly
1773 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1774 * With adjusted search length an allocation would not succeed.
1775 */
1776 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1777 adjust_search_size = false;
1778
1779 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1780 if (unlikely(!va))
1781 return vend;
1782
1783 nva_start_addr = va_alloc(va, root, head, size, align, vstart, vend);
1784 if (nva_start_addr == vend)
1785 return vend;
1786
1787 #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1788 find_vmap_lowest_match_check(root, head, size, align);
1789 #endif
1790
1791 return nva_start_addr;
1792 }
1793
1794 /*
1795 * Free a region of KVA allocated by alloc_vmap_area
1796 */
1797 static void free_vmap_area(struct vmap_area *va)
1798 {
1799 struct vmap_node *vn = addr_to_node(va->va_start);
1800
1801 /*
1802 * Remove from the busy tree/list.
1803 */
1804 spin_lock(&vn->busy.lock);
1805 unlink_va(va, &vn->busy.root);
1806 spin_unlock(&vn->busy.lock);
1807
1808 /*
1809 * Insert/Merge it back to the free tree/list.
1810 */
1811 spin_lock(&free_vmap_area_lock);
1812 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1813 spin_unlock(&free_vmap_area_lock);
1814 }
1815
1816 static inline void
1817 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1818 {
1819 struct vmap_area *va = NULL, *tmp;
1820
1821 /*
1822 * Preload this CPU with one extra vmap_area object. It is used
1823 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1824 * a CPU that does an allocation is preloaded.
1825 *
1826 * We do it in non-atomic context, thus it allows us to use more
1827 * permissive allocation masks to be more stable under low memory
1828 * condition and high memory pressure.
1829 */
1830 if (!this_cpu_read(ne_fit_preload_node))
1831 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1832
1833 spin_lock(lock);
1834
1835 tmp = NULL;
1836 if (va && !__this_cpu_try_cmpxchg(ne_fit_preload_node, &tmp, va))
1837 kmem_cache_free(vmap_area_cachep, va);
1838 }
1839
1840 static struct vmap_pool *
1841 size_to_va_pool(struct vmap_node *vn, unsigned long size)
1842 {
1843 unsigned int idx = (size - 1) / PAGE_SIZE;
1844
1845 if (idx < MAX_VA_SIZE_PAGES)
1846 return &vn->pool[idx];
1847
1848 return NULL;
1849 }
1850
1851 static bool
1852 node_pool_add_va(struct vmap_node *n, struct vmap_area *va)
1853 {
1854 struct vmap_pool *vp;
1855
1856 vp = size_to_va_pool(n, va_size(va));
1857 if (!vp)
1858 return false;
1859
1860 spin_lock(&n->pool_lock);
1861 list_add(&va->list, &vp->head);
1862 WRITE_ONCE(vp->len, vp->len + 1);
1863 spin_unlock(&n->pool_lock);
1864
1865 return true;
1866 }
1867
1868 static struct vmap_area *
1869 node_pool_del_va(struct vmap_node *vn, unsigned long size,
1870 unsigned long align, unsigned long vstart,
1871 unsigned long vend)
1872 {
1873 struct vmap_area *va = NULL;
1874 struct vmap_pool *vp;
1875 int err = 0;
1876
1877 vp = size_to_va_pool(vn, size);
1878 if (!vp || list_empty(&vp->head))
1879 return NULL;
1880
1881 spin_lock(&vn->pool_lock);
1882 if (!list_empty(&vp->head)) {
1883 va = list_first_entry(&vp->head, struct vmap_area, list);
1884
1885 if (IS_ALIGNED(va->va_start, align)) {
1886 /*
1887 * Do some sanity check and emit a warning
1888 * if one of below checks detects an error.
1889 */
1890 err |= (va_size(va) != size);
1891 err |= (va->va_start < vstart);
1892 err |= (va->va_end > vend);
1893
1894 if (!WARN_ON_ONCE(err)) {
1895 list_del_init(&va->list);
1896 WRITE_ONCE(vp->len, vp->len - 1);
1897 } else {
1898 va = NULL;
1899 }
1900 } else {
1901 list_move_tail(&va->list, &vp->head);
1902 va = NULL;
1903 }
1904 }
1905 spin_unlock(&vn->pool_lock);
1906
1907 return va;
1908 }
1909
1910 static struct vmap_area *
1911 node_alloc(unsigned long size, unsigned long align,
1912 unsigned long vstart, unsigned long vend,
1913 unsigned long *addr, unsigned int *vn_id)
1914 {
1915 struct vmap_area *va;
1916
1917 *vn_id = 0;
1918 *addr = vend;
1919
1920 /*
1921 * Fallback to a global heap if not vmalloc or there
1922 * is only one node.
1923 */
1924 if (vstart != VMALLOC_START || vend != VMALLOC_END ||
1925 nr_vmap_nodes == 1)
1926 return NULL;
1927
1928 *vn_id = raw_smp_processor_id() % nr_vmap_nodes;
1929 va = node_pool_del_va(id_to_node(*vn_id), size, align, vstart, vend);
1930 *vn_id = encode_vn_id(*vn_id);
1931
1932 if (va)
1933 *addr = va->va_start;
1934
1935 return va;
1936 }
1937
1938 static inline void setup_vmalloc_vm(struct vm_struct *vm,
1939 struct vmap_area *va, unsigned long flags, const void *caller)
1940 {
1941 vm->flags = flags;
1942 vm->addr = (void *)va->va_start;
1943 vm->size = va->va_end - va->va_start;
1944 vm->caller = caller;
1945 va->vm = vm;
1946 }
1947
1948 /*
1949 * Allocate a region of KVA of the specified size and alignment, within the
1950 * vstart and vend. If vm is passed in, the two will also be bound.
1951 */
1952 static struct vmap_area *alloc_vmap_area(unsigned long size,
1953 unsigned long align,
1954 unsigned long vstart, unsigned long vend,
1955 int node, gfp_t gfp_mask,
1956 unsigned long va_flags, struct vm_struct *vm)
1957 {
1958 struct vmap_node *vn;
1959 struct vmap_area *va;
1960 unsigned long freed;
1961 unsigned long addr;
1962 unsigned int vn_id;
1963 int purged = 0;
1964 int ret;
1965
1966 if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1967 return ERR_PTR(-EINVAL);
1968
1969 if (unlikely(!vmap_initialized))
1970 return ERR_PTR(-EBUSY);
1971
1972 might_sleep();
1973
1974 /*
1975 * If a VA is obtained from a global heap(if it fails here)
1976 * it is anyway marked with this "vn_id" so it is returned
1977 * to this pool's node later. Such way gives a possibility
1978 * to populate pools based on users demand.
1979 *
1980 * On success a ready to go VA is returned.
1981 */
1982 va = node_alloc(size, align, vstart, vend, &addr, &vn_id);
1983 if (!va) {
1984 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1985
1986 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1987 if (unlikely(!va))
1988 return ERR_PTR(-ENOMEM);
1989
1990 /*
1991 * Only scan the relevant parts containing pointers to other objects
1992 * to avoid false negatives.
1993 */
1994 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1995 }
1996
1997 retry:
1998 if (addr == vend) {
1999 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
2000 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
2001 size, align, vstart, vend);
2002 spin_unlock(&free_vmap_area_lock);
2003 }
2004
2005 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
2006
2007 /*
2008 * If an allocation fails, the "vend" address is
2009 * returned. Therefore trigger the overflow path.
2010 */
2011 if (unlikely(addr == vend))
2012 goto overflow;
2013
2014 va->va_start = addr;
2015 va->va_end = addr + size;
2016 va->vm = NULL;
2017 va->flags = (va_flags | vn_id);
2018
2019 if (vm) {
2020 vm->addr = (void *)va->va_start;
2021 vm->size = va->va_end - va->va_start;
2022 va->vm = vm;
2023 }
2024
2025 vn = addr_to_node(va->va_start);
2026
2027 spin_lock(&vn->busy.lock);
2028 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
2029 spin_unlock(&vn->busy.lock);
2030
2031 BUG_ON(!IS_ALIGNED(va->va_start, align));
2032 BUG_ON(va->va_start < vstart);
2033 BUG_ON(va->va_end > vend);
2034
2035 ret = kasan_populate_vmalloc(addr, size);
2036 if (ret) {
2037 free_vmap_area(va);
2038 return ERR_PTR(ret);
2039 }
2040
2041 return va;
2042
2043 overflow:
2044 if (!purged) {
2045 reclaim_and_purge_vmap_areas();
2046 purged = 1;
2047 goto retry;
2048 }
2049
2050 freed = 0;
2051 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
2052
2053 if (freed > 0) {
2054 purged = 0;
2055 goto retry;
2056 }
2057
2058 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
2059 pr_warn("vmalloc_node_range for size %lu failed: Address range restricted to %#lx - %#lx\n",
2060 size, vstart, vend);
2061
2062 kmem_cache_free(vmap_area_cachep, va);
2063 return ERR_PTR(-EBUSY);
2064 }
2065
2066 int register_vmap_purge_notifier(struct notifier_block *nb)
2067 {
2068 return blocking_notifier_chain_register(&vmap_notify_list, nb);
2069 }
2070 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
2071
2072 int unregister_vmap_purge_notifier(struct notifier_block *nb)
2073 {
2074 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
2075 }
2076 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
2077
2078 /*
2079 * lazy_max_pages is the maximum amount of virtual address space we gather up
2080 * before attempting to purge with a TLB flush.
2081 *
2082 * There is a tradeoff here: a larger number will cover more kernel page tables
2083 * and take slightly longer to purge, but it will linearly reduce the number of
2084 * global TLB flushes that must be performed. It would seem natural to scale
2085 * this number up linearly with the number of CPUs (because vmapping activity
2086 * could also scale linearly with the number of CPUs), however it is likely
2087 * that in practice, workloads might be constrained in other ways that mean
2088 * vmap activity will not scale linearly with CPUs. Also, I want to be
2089 * conservative and not introduce a big latency on huge systems, so go with
2090 * a less aggressive log scale. It will still be an improvement over the old
2091 * code, and it will be simple to change the scale factor if we find that it
2092 * becomes a problem on bigger systems.
2093 */
2094 static unsigned long lazy_max_pages(void)
2095 {
2096 unsigned int log;
2097
2098 log = fls(num_online_cpus());
2099
2100 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
2101 }
2102
2103 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
2104
2105 /*
2106 * Serialize vmap purging. There is no actual critical section protected
2107 * by this lock, but we want to avoid concurrent calls for performance
2108 * reasons and to make the pcpu_get_vm_areas more deterministic.
2109 */
2110 static DEFINE_MUTEX(vmap_purge_lock);
2111
2112 /* for per-CPU blocks */
2113 static void purge_fragmented_blocks_allcpus(void);
2114 static cpumask_t purge_nodes;
2115
2116 static void
2117 reclaim_list_global(struct list_head *head)
2118 {
2119 struct vmap_area *va, *n;
2120
2121 if (list_empty(head))
2122 return;
2123
2124 spin_lock(&free_vmap_area_lock);
2125 list_for_each_entry_safe(va, n, head, list)
2126 merge_or_add_vmap_area_augment(va,
2127 &free_vmap_area_root, &free_vmap_area_list);
2128 spin_unlock(&free_vmap_area_lock);
2129 }
2130
2131 static void
2132 decay_va_pool_node(struct vmap_node *vn, bool full_decay)
2133 {
2134 struct vmap_area *va, *nva;
2135 struct list_head decay_list;
2136 struct rb_root decay_root;
2137 unsigned long n_decay;
2138 int i;
2139
2140 decay_root = RB_ROOT;
2141 INIT_LIST_HEAD(&decay_list);
2142
2143 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
2144 struct list_head tmp_list;
2145
2146 if (list_empty(&vn->pool[i].head))
2147 continue;
2148
2149 INIT_LIST_HEAD(&tmp_list);
2150
2151 /* Detach the pool, so no-one can access it. */
2152 spin_lock(&vn->pool_lock);
2153 list_replace_init(&vn->pool[i].head, &tmp_list);
2154 spin_unlock(&vn->pool_lock);
2155
2156 if (full_decay)
2157 WRITE_ONCE(vn->pool[i].len, 0);
2158
2159 /* Decay a pool by ~25% out of left objects. */
2160 n_decay = vn->pool[i].len >> 2;
2161
2162 list_for_each_entry_safe(va, nva, &tmp_list, list) {
2163 list_del_init(&va->list);
2164 merge_or_add_vmap_area(va, &decay_root, &decay_list);
2165
2166 if (!full_decay) {
2167 WRITE_ONCE(vn->pool[i].len, vn->pool[i].len - 1);
2168
2169 if (!--n_decay)
2170 break;
2171 }
2172 }
2173
2174 /*
2175 * Attach the pool back if it has been partly decayed.
2176 * Please note, it is supposed that nobody(other contexts)
2177 * can populate the pool therefore a simple list replace
2178 * operation takes place here.
2179 */
2180 if (!full_decay && !list_empty(&tmp_list)) {
2181 spin_lock(&vn->pool_lock);
2182 list_replace_init(&tmp_list, &vn->pool[i].head);
2183 spin_unlock(&vn->pool_lock);
2184 }
2185 }
2186
2187 reclaim_list_global(&decay_list);
2188 }
2189
2190 static void purge_vmap_node(struct work_struct *work)
2191 {
2192 struct vmap_node *vn = container_of(work,
2193 struct vmap_node, purge_work);
2194 unsigned long nr_purged_pages = 0;
2195 struct vmap_area *va, *n_va;
2196 LIST_HEAD(local_list);
2197
2198 vn->nr_purged = 0;
2199
2200 list_for_each_entry_safe(va, n_va, &vn->purge_list, list) {
2201 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
2202 unsigned long orig_start = va->va_start;
2203 unsigned long orig_end = va->va_end;
2204 unsigned int vn_id = decode_vn_id(va->flags);
2205
2206 list_del_init(&va->list);
2207
2208 if (is_vmalloc_or_module_addr((void *)orig_start))
2209 kasan_release_vmalloc(orig_start, orig_end,
2210 va->va_start, va->va_end);
2211
2212 nr_purged_pages += nr;
2213 vn->nr_purged++;
2214
2215 if (is_vn_id_valid(vn_id) && !vn->skip_populate)
2216 if (node_pool_add_va(vn, va))
2217 continue;
2218
2219 /* Go back to global. */
2220 list_add(&va->list, &local_list);
2221 }
2222
2223 atomic_long_sub(nr_purged_pages, &vmap_lazy_nr);
2224
2225 reclaim_list_global(&local_list);
2226 }
2227
2228 /*
2229 * Purges all lazily-freed vmap areas.
2230 */
2231 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
2232 bool full_pool_decay)
2233 {
2234 unsigned long nr_purged_areas = 0;
2235 unsigned int nr_purge_helpers;
2236 unsigned int nr_purge_nodes;
2237 struct vmap_node *vn;
2238 int i;
2239
2240 lockdep_assert_held(&vmap_purge_lock);
2241
2242 /*
2243 * Use cpumask to mark which node has to be processed.
2244 */
2245 purge_nodes = CPU_MASK_NONE;
2246
2247 for (i = 0; i < nr_vmap_nodes; i++) {
2248 vn = &vmap_nodes[i];
2249
2250 INIT_LIST_HEAD(&vn->purge_list);
2251 vn->skip_populate = full_pool_decay;
2252 decay_va_pool_node(vn, full_pool_decay);
2253
2254 if (RB_EMPTY_ROOT(&vn->lazy.root))
2255 continue;
2256
2257 spin_lock(&vn->lazy.lock);
2258 WRITE_ONCE(vn->lazy.root.rb_node, NULL);
2259 list_replace_init(&vn->lazy.head, &vn->purge_list);
2260 spin_unlock(&vn->lazy.lock);
2261
2262 start = min(start, list_first_entry(&vn->purge_list,
2263 struct vmap_area, list)->va_start);
2264
2265 end = max(end, list_last_entry(&vn->purge_list,
2266 struct vmap_area, list)->va_end);
2267
2268 cpumask_set_cpu(i, &purge_nodes);
2269 }
2270
2271 nr_purge_nodes = cpumask_weight(&purge_nodes);
2272 if (nr_purge_nodes > 0) {
2273 flush_tlb_kernel_range(start, end);
2274
2275 /* One extra worker is per a lazy_max_pages() full set minus one. */
2276 nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
2277 nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
2278
2279 for_each_cpu(i, &purge_nodes) {
2280 vn = &vmap_nodes[i];
2281
2282 if (nr_purge_helpers > 0) {
2283 INIT_WORK(&vn->purge_work, purge_vmap_node);
2284
2285 if (cpumask_test_cpu(i, cpu_online_mask))
2286 schedule_work_on(i, &vn->purge_work);
2287 else
2288 schedule_work(&vn->purge_work);
2289
2290 nr_purge_helpers--;
2291 } else {
2292 vn->purge_work.func = NULL;
2293 purge_vmap_node(&vn->purge_work);
2294 nr_purged_areas += vn->nr_purged;
2295 }
2296 }
2297
2298 for_each_cpu(i, &purge_nodes) {
2299 vn = &vmap_nodes[i];
2300
2301 if (vn->purge_work.func) {
2302 flush_work(&vn->purge_work);
2303 nr_purged_areas += vn->nr_purged;
2304 }
2305 }
2306 }
2307
2308 trace_purge_vmap_area_lazy(start, end, nr_purged_areas);
2309 return nr_purged_areas > 0;
2310 }
2311
2312 /*
2313 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
2314 */
2315 static void reclaim_and_purge_vmap_areas(void)
2316
2317 {
2318 mutex_lock(&vmap_purge_lock);
2319 purge_fragmented_blocks_allcpus();
2320 __purge_vmap_area_lazy(ULONG_MAX, 0, true);
2321 mutex_unlock(&vmap_purge_lock);
2322 }
2323
2324 static void drain_vmap_area_work(struct work_struct *work)
2325 {
2326 mutex_lock(&vmap_purge_lock);
2327 __purge_vmap_area_lazy(ULONG_MAX, 0, false);
2328 mutex_unlock(&vmap_purge_lock);
2329 }
2330
2331 /*
2332 * Free a vmap area, caller ensuring that the area has been unmapped,
2333 * unlinked and flush_cache_vunmap had been called for the correct
2334 * range previously.
2335 */
2336 static void free_vmap_area_noflush(struct vmap_area *va)
2337 {
2338 unsigned long nr_lazy_max = lazy_max_pages();
2339 unsigned long va_start = va->va_start;
2340 unsigned int vn_id = decode_vn_id(va->flags);
2341 struct vmap_node *vn;
2342 unsigned long nr_lazy;
2343
2344 if (WARN_ON_ONCE(!list_empty(&va->list)))
2345 return;
2346
2347 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
2348 PAGE_SHIFT, &vmap_lazy_nr);
2349
2350 /*
2351 * If it was request by a certain node we would like to
2352 * return it to that node, i.e. its pool for later reuse.
2353 */
2354 vn = is_vn_id_valid(vn_id) ?
2355 id_to_node(vn_id):addr_to_node(va->va_start);
2356
2357 spin_lock(&vn->lazy.lock);
2358 insert_vmap_area(va, &vn->lazy.root, &vn->lazy.head);
2359 spin_unlock(&vn->lazy.lock);
2360
2361 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
2362
2363 /* After this point, we may free va at any time */
2364 if (unlikely(nr_lazy > nr_lazy_max))
2365 schedule_work(&drain_vmap_work);
2366 }
2367
2368 /*
2369 * Free and unmap a vmap area
2370 */
2371 static void free_unmap_vmap_area(struct vmap_area *va)
2372 {
2373 flush_cache_vunmap(va->va_start, va->va_end);
2374 vunmap_range_noflush(va->va_start, va->va_end);
2375 if (debug_pagealloc_enabled_static())
2376 flush_tlb_kernel_range(va->va_start, va->va_end);
2377
2378 free_vmap_area_noflush(va);
2379 }
2380
2381 struct vmap_area *find_vmap_area(unsigned long addr)
2382 {
2383 struct vmap_node *vn;
2384 struct vmap_area *va;
2385 int i, j;
2386
2387 if (unlikely(!vmap_initialized))
2388 return NULL;
2389
2390 /*
2391 * An addr_to_node_id(addr) converts an address to a node index
2392 * where a VA is located. If VA spans several zones and passed
2393 * addr is not the same as va->va_start, what is not common, we
2394 * may need to scan extra nodes. See an example:
2395 *
2396 * <----va---->
2397 * -|-----|-----|-----|-----|-
2398 * 1 2 0 1
2399 *
2400 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
2401 * addr is within 2 or 0 nodes we should do extra work.
2402 */
2403 i = j = addr_to_node_id(addr);
2404 do {
2405 vn = &vmap_nodes[i];
2406
2407 spin_lock(&vn->busy.lock);
2408 va = __find_vmap_area(addr, &vn->busy.root);
2409 spin_unlock(&vn->busy.lock);
2410
2411 if (va)
2412 return va;
2413 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2414
2415 return NULL;
2416 }
2417
2418 static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
2419 {
2420 struct vmap_node *vn;
2421 struct vmap_area *va;
2422 int i, j;
2423
2424 /*
2425 * Check the comment in the find_vmap_area() about the loop.
2426 */
2427 i = j = addr_to_node_id(addr);
2428 do {
2429 vn = &vmap_nodes[i];
2430
2431 spin_lock(&vn->busy.lock);
2432 va = __find_vmap_area(addr, &vn->busy.root);
2433 if (va)
2434 unlink_va(va, &vn->busy.root);
2435 spin_unlock(&vn->busy.lock);
2436
2437 if (va)
2438 return va;
2439 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2440
2441 return NULL;
2442 }
2443
2444 /*** Per cpu kva allocator ***/
2445
2446 /*
2447 * vmap space is limited especially on 32 bit architectures. Ensure there is
2448 * room for at least 16 percpu vmap blocks per CPU.
2449 */
2450 /*
2451 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
2452 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
2453 * instead (we just need a rough idea)
2454 */
2455 #if BITS_PER_LONG == 32
2456 #define VMALLOC_SPACE (128UL*1024*1024)
2457 #else
2458 #define VMALLOC_SPACE (128UL*1024*1024*1024)
2459 #endif
2460
2461 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
2462 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
2463 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
2464 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
2465 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
2466 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
2467 #define VMAP_BBMAP_BITS \
2468 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
2469 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
2470 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
2471
2472 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
2473
2474 /*
2475 * Purge threshold to prevent overeager purging of fragmented blocks for
2476 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
2477 */
2478 #define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
2479
2480 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
2481 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
2482 #define VMAP_FLAGS_MASK 0x3
2483
2484 struct vmap_block_queue {
2485 spinlock_t lock;
2486 struct list_head free;
2487
2488 /*
2489 * An xarray requires an extra memory dynamically to
2490 * be allocated. If it is an issue, we can use rb-tree
2491 * instead.
2492 */
2493 struct xarray vmap_blocks;
2494 };
2495
2496 struct vmap_block {
2497 spinlock_t lock;
2498 struct vmap_area *va;
2499 unsigned long free, dirty;
2500 DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
2501 unsigned long dirty_min, dirty_max; /*< dirty range */
2502 struct list_head free_list;
2503 struct rcu_head rcu_head;
2504 struct list_head purge;
2505 unsigned int cpu;
2506 };
2507
2508 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
2509 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
2510
2511 /*
2512 * In order to fast access to any "vmap_block" associated with a
2513 * specific address, we use a hash.
2514 *
2515 * A per-cpu vmap_block_queue is used in both ways, to serialize
2516 * an access to free block chains among CPUs(alloc path) and it
2517 * also acts as a vmap_block hash(alloc/free paths). It means we
2518 * overload it, since we already have the per-cpu array which is
2519 * used as a hash table. When used as a hash a 'cpu' passed to
2520 * per_cpu() is not actually a CPU but rather a hash index.
2521 *
2522 * A hash function is addr_to_vb_xa() which hashes any address
2523 * to a specific index(in a hash) it belongs to. This then uses a
2524 * per_cpu() macro to access an array with generated index.
2525 *
2526 * An example:
2527 *
2528 * CPU_1 CPU_2 CPU_0
2529 * | | |
2530 * V V V
2531 * 0 10 20 30 40 50 60
2532 * |------|------|------|------|------|------|...<vmap address space>
2533 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
2534 *
2535 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
2536 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
2537 *
2538 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
2539 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
2540 *
2541 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
2542 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
2543 *
2544 * This technique almost always avoids lock contention on insert/remove,
2545 * however xarray spinlocks protect against any contention that remains.
2546 */
2547 static struct xarray *
2548 addr_to_vb_xa(unsigned long addr)
2549 {
2550 int index = (addr / VMAP_BLOCK_SIZE) % nr_cpu_ids;
2551
2552 /*
2553 * Please note, nr_cpu_ids points on a highest set
2554 * possible bit, i.e. we never invoke cpumask_next()
2555 * if an index points on it which is nr_cpu_ids - 1.
2556 */
2557 if (!cpu_possible(index))
2558 index = cpumask_next(index, cpu_possible_mask);
2559
2560 return &per_cpu(vmap_block_queue, index).vmap_blocks;
2561 }
2562
2563 /*
2564 * We should probably have a fallback mechanism to allocate virtual memory
2565 * out of partially filled vmap blocks. However vmap block sizing should be
2566 * fairly reasonable according to the vmalloc size, so it shouldn't be a
2567 * big problem.
2568 */
2569
2570 static unsigned long addr_to_vb_idx(unsigned long addr)
2571 {
2572 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2573 addr /= VMAP_BLOCK_SIZE;
2574 return addr;
2575 }
2576
2577 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2578 {
2579 unsigned long addr;
2580
2581 addr = va_start + (pages_off << PAGE_SHIFT);
2582 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2583 return (void *)addr;
2584 }
2585
2586 /**
2587 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2588 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2589 * @order: how many 2^order pages should be occupied in newly allocated block
2590 * @gfp_mask: flags for the page level allocator
2591 *
2592 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2593 */
2594 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2595 {
2596 struct vmap_block_queue *vbq;
2597 struct vmap_block *vb;
2598 struct vmap_area *va;
2599 struct xarray *xa;
2600 unsigned long vb_idx;
2601 int node, err;
2602 void *vaddr;
2603
2604 node = numa_node_id();
2605
2606 vb = kmalloc_node(sizeof(struct vmap_block),
2607 gfp_mask & GFP_RECLAIM_MASK, node);
2608 if (unlikely(!vb))
2609 return ERR_PTR(-ENOMEM);
2610
2611 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2612 VMALLOC_START, VMALLOC_END,
2613 node, gfp_mask,
2614 VMAP_RAM|VMAP_BLOCK, NULL);
2615 if (IS_ERR(va)) {
2616 kfree(vb);
2617 return ERR_CAST(va);
2618 }
2619
2620 vaddr = vmap_block_vaddr(va->va_start, 0);
2621 spin_lock_init(&vb->lock);
2622 vb->va = va;
2623 /* At least something should be left free */
2624 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2625 bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2626 vb->free = VMAP_BBMAP_BITS - (1UL << order);
2627 vb->dirty = 0;
2628 vb->dirty_min = VMAP_BBMAP_BITS;
2629 vb->dirty_max = 0;
2630 bitmap_set(vb->used_map, 0, (1UL << order));
2631 INIT_LIST_HEAD(&vb->free_list);
2632 vb->cpu = raw_smp_processor_id();
2633
2634 xa = addr_to_vb_xa(va->va_start);
2635 vb_idx = addr_to_vb_idx(va->va_start);
2636 err = xa_insert(xa, vb_idx, vb, gfp_mask);
2637 if (err) {
2638 kfree(vb);
2639 free_vmap_area(va);
2640 return ERR_PTR(err);
2641 }
2642 /*
2643 * list_add_tail_rcu could happened in another core
2644 * rather than vb->cpu due to task migration, which
2645 * is safe as list_add_tail_rcu will ensure the list's
2646 * integrity together with list_for_each_rcu from read
2647 * side.
2648 */
2649 vbq = per_cpu_ptr(&vmap_block_queue, vb->cpu);
2650 spin_lock(&vbq->lock);
2651 list_add_tail_rcu(&vb->free_list, &vbq->free);
2652 spin_unlock(&vbq->lock);
2653
2654 return vaddr;
2655 }
2656
2657 static void free_vmap_block(struct vmap_block *vb)
2658 {
2659 struct vmap_node *vn;
2660 struct vmap_block *tmp;
2661 struct xarray *xa;
2662
2663 xa = addr_to_vb_xa(vb->va->va_start);
2664 tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2665 BUG_ON(tmp != vb);
2666
2667 vn = addr_to_node(vb->va->va_start);
2668 spin_lock(&vn->busy.lock);
2669 unlink_va(vb->va, &vn->busy.root);
2670 spin_unlock(&vn->busy.lock);
2671
2672 free_vmap_area_noflush(vb->va);
2673 kfree_rcu(vb, rcu_head);
2674 }
2675
2676 static bool purge_fragmented_block(struct vmap_block *vb,
2677 struct list_head *purge_list, bool force_purge)
2678 {
2679 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, vb->cpu);
2680
2681 if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2682 vb->dirty == VMAP_BBMAP_BITS)
2683 return false;
2684
2685 /* Don't overeagerly purge usable blocks unless requested */
2686 if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2687 return false;
2688
2689 /* prevent further allocs after releasing lock */
2690 WRITE_ONCE(vb->free, 0);
2691 /* prevent purging it again */
2692 WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2693 vb->dirty_min = 0;
2694 vb->dirty_max = VMAP_BBMAP_BITS;
2695 spin_lock(&vbq->lock);
2696 list_del_rcu(&vb->free_list);
2697 spin_unlock(&vbq->lock);
2698 list_add_tail(&vb->purge, purge_list);
2699 return true;
2700 }
2701
2702 static void free_purged_blocks(struct list_head *purge_list)
2703 {
2704 struct vmap_block *vb, *n_vb;
2705
2706 list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2707 list_del(&vb->purge);
2708 free_vmap_block(vb);
2709 }
2710 }
2711
2712 static void purge_fragmented_blocks(int cpu)
2713 {
2714 LIST_HEAD(purge);
2715 struct vmap_block *vb;
2716 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2717
2718 rcu_read_lock();
2719 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2720 unsigned long free = READ_ONCE(vb->free);
2721 unsigned long dirty = READ_ONCE(vb->dirty);
2722
2723 if (free + dirty != VMAP_BBMAP_BITS ||
2724 dirty == VMAP_BBMAP_BITS)
2725 continue;
2726
2727 spin_lock(&vb->lock);
2728 purge_fragmented_block(vb, &purge, true);
2729 spin_unlock(&vb->lock);
2730 }
2731 rcu_read_unlock();
2732 free_purged_blocks(&purge);
2733 }
2734
2735 static void purge_fragmented_blocks_allcpus(void)
2736 {
2737 int cpu;
2738
2739 for_each_possible_cpu(cpu)
2740 purge_fragmented_blocks(cpu);
2741 }
2742
2743 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2744 {
2745 struct vmap_block_queue *vbq;
2746 struct vmap_block *vb;
2747 void *vaddr = NULL;
2748 unsigned int order;
2749
2750 BUG_ON(offset_in_page(size));
2751 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2752 if (WARN_ON(size == 0)) {
2753 /*
2754 * Allocating 0 bytes isn't what caller wants since
2755 * get_order(0) returns funny result. Just warn and terminate
2756 * early.
2757 */
2758 return ERR_PTR(-EINVAL);
2759 }
2760 order = get_order(size);
2761
2762 rcu_read_lock();
2763 vbq = raw_cpu_ptr(&vmap_block_queue);
2764 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2765 unsigned long pages_off;
2766
2767 if (READ_ONCE(vb->free) < (1UL << order))
2768 continue;
2769
2770 spin_lock(&vb->lock);
2771 if (vb->free < (1UL << order)) {
2772 spin_unlock(&vb->lock);
2773 continue;
2774 }
2775
2776 pages_off = VMAP_BBMAP_BITS - vb->free;
2777 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2778 WRITE_ONCE(vb->free, vb->free - (1UL << order));
2779 bitmap_set(vb->used_map, pages_off, (1UL << order));
2780 if (vb->free == 0) {
2781 spin_lock(&vbq->lock);
2782 list_del_rcu(&vb->free_list);
2783 spin_unlock(&vbq->lock);
2784 }
2785
2786 spin_unlock(&vb->lock);
2787 break;
2788 }
2789
2790 rcu_read_unlock();
2791
2792 /* Allocate new block if nothing was found */
2793 if (!vaddr)
2794 vaddr = new_vmap_block(order, gfp_mask);
2795
2796 return vaddr;
2797 }
2798
2799 static void vb_free(unsigned long addr, unsigned long size)
2800 {
2801 unsigned long offset;
2802 unsigned int order;
2803 struct vmap_block *vb;
2804 struct xarray *xa;
2805
2806 BUG_ON(offset_in_page(size));
2807 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2808
2809 flush_cache_vunmap(addr, addr + size);
2810
2811 order = get_order(size);
2812 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2813
2814 xa = addr_to_vb_xa(addr);
2815 vb = xa_load(xa, addr_to_vb_idx(addr));
2816
2817 spin_lock(&vb->lock);
2818 bitmap_clear(vb->used_map, offset, (1UL << order));
2819 spin_unlock(&vb->lock);
2820
2821 vunmap_range_noflush(addr, addr + size);
2822
2823 if (debug_pagealloc_enabled_static())
2824 flush_tlb_kernel_range(addr, addr + size);
2825
2826 spin_lock(&vb->lock);
2827
2828 /* Expand the not yet TLB flushed dirty range */
2829 vb->dirty_min = min(vb->dirty_min, offset);
2830 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2831
2832 WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2833 if (vb->dirty == VMAP_BBMAP_BITS) {
2834 BUG_ON(vb->free);
2835 spin_unlock(&vb->lock);
2836 free_vmap_block(vb);
2837 } else
2838 spin_unlock(&vb->lock);
2839 }
2840
2841 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2842 {
2843 LIST_HEAD(purge_list);
2844 int cpu;
2845
2846 if (unlikely(!vmap_initialized))
2847 return;
2848
2849 mutex_lock(&vmap_purge_lock);
2850
2851 for_each_possible_cpu(cpu) {
2852 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2853 struct vmap_block *vb;
2854 unsigned long idx;
2855
2856 rcu_read_lock();
2857 xa_for_each(&vbq->vmap_blocks, idx, vb) {
2858 spin_lock(&vb->lock);
2859
2860 /*
2861 * Try to purge a fragmented block first. If it's
2862 * not purgeable, check whether there is dirty
2863 * space to be flushed.
2864 */
2865 if (!purge_fragmented_block(vb, &purge_list, false) &&
2866 vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2867 unsigned long va_start = vb->va->va_start;
2868 unsigned long s, e;
2869
2870 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2871 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2872
2873 start = min(s, start);
2874 end = max(e, end);
2875
2876 /* Prevent that this is flushed again */
2877 vb->dirty_min = VMAP_BBMAP_BITS;
2878 vb->dirty_max = 0;
2879
2880 flush = 1;
2881 }
2882 spin_unlock(&vb->lock);
2883 }
2884 rcu_read_unlock();
2885 }
2886 free_purged_blocks(&purge_list);
2887
2888 if (!__purge_vmap_area_lazy(start, end, false) && flush)
2889 flush_tlb_kernel_range(start, end);
2890 mutex_unlock(&vmap_purge_lock);
2891 }
2892
2893 /**
2894 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2895 *
2896 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2897 * to amortize TLB flushing overheads. What this means is that any page you
2898 * have now, may, in a former life, have been mapped into kernel virtual
2899 * address by the vmap layer and so there might be some CPUs with TLB entries
2900 * still referencing that page (additional to the regular 1:1 kernel mapping).
2901 *
2902 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2903 * be sure that none of the pages we have control over will have any aliases
2904 * from the vmap layer.
2905 */
2906 void vm_unmap_aliases(void)
2907 {
2908 unsigned long start = ULONG_MAX, end = 0;
2909 int flush = 0;
2910
2911 _vm_unmap_aliases(start, end, flush);
2912 }
2913 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2914
2915 /**
2916 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2917 * @mem: the pointer returned by vm_map_ram
2918 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2919 */
2920 void vm_unmap_ram(const void *mem, unsigned int count)
2921 {
2922 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2923 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2924 struct vmap_area *va;
2925
2926 might_sleep();
2927 BUG_ON(!addr);
2928 BUG_ON(addr < VMALLOC_START);
2929 BUG_ON(addr > VMALLOC_END);
2930 BUG_ON(!PAGE_ALIGNED(addr));
2931
2932 kasan_poison_vmalloc(mem, size);
2933
2934 if (likely(count <= VMAP_MAX_ALLOC)) {
2935 debug_check_no_locks_freed(mem, size);
2936 vb_free(addr, size);
2937 return;
2938 }
2939
2940 va = find_unlink_vmap_area(addr);
2941 if (WARN_ON_ONCE(!va))
2942 return;
2943
2944 debug_check_no_locks_freed((void *)va->va_start,
2945 (va->va_end - va->va_start));
2946 free_unmap_vmap_area(va);
2947 }
2948 EXPORT_SYMBOL(vm_unmap_ram);
2949
2950 /**
2951 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2952 * @pages: an array of pointers to the pages to be mapped
2953 * @count: number of pages
2954 * @node: prefer to allocate data structures on this node
2955 *
2956 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2957 * faster than vmap so it's good. But if you mix long-life and short-life
2958 * objects with vm_map_ram(), it could consume lots of address space through
2959 * fragmentation (especially on a 32bit machine). You could see failures in
2960 * the end. Please use this function for short-lived objects.
2961 *
2962 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2963 */
2964 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2965 {
2966 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2967 unsigned long addr;
2968 void *mem;
2969
2970 if (likely(count <= VMAP_MAX_ALLOC)) {
2971 mem = vb_alloc(size, GFP_KERNEL);
2972 if (IS_ERR(mem))
2973 return NULL;
2974 addr = (unsigned long)mem;
2975 } else {
2976 struct vmap_area *va;
2977 va = alloc_vmap_area(size, PAGE_SIZE,
2978 VMALLOC_START, VMALLOC_END,
2979 node, GFP_KERNEL, VMAP_RAM,
2980 NULL);
2981 if (IS_ERR(va))
2982 return NULL;
2983
2984 addr = va->va_start;
2985 mem = (void *)addr;
2986 }
2987
2988 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2989 pages, PAGE_SHIFT) < 0) {
2990 vm_unmap_ram(mem, count);
2991 return NULL;
2992 }
2993
2994 /*
2995 * Mark the pages as accessible, now that they are mapped.
2996 * With hardware tag-based KASAN, marking is skipped for
2997 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2998 */
2999 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
3000
3001 return mem;
3002 }
3003 EXPORT_SYMBOL(vm_map_ram);
3004
3005 static struct vm_struct *vmlist __initdata;
3006
3007 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
3008 {
3009 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3010 return vm->page_order;
3011 #else
3012 return 0;
3013 #endif
3014 }
3015
3016 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
3017 {
3018 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3019 vm->page_order = order;
3020 #else
3021 BUG_ON(order != 0);
3022 #endif
3023 }
3024
3025 /**
3026 * vm_area_add_early - add vmap area early during boot
3027 * @vm: vm_struct to add
3028 *
3029 * This function is used to add fixed kernel vm area to vmlist before
3030 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
3031 * should contain proper values and the other fields should be zero.
3032 *
3033 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3034 */
3035 void __init vm_area_add_early(struct vm_struct *vm)
3036 {
3037 struct vm_struct *tmp, **p;
3038
3039 BUG_ON(vmap_initialized);
3040 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
3041 if (tmp->addr >= vm->addr) {
3042 BUG_ON(tmp->addr < vm->addr + vm->size);
3043 break;
3044 } else
3045 BUG_ON(tmp->addr + tmp->size > vm->addr);
3046 }
3047 vm->next = *p;
3048 *p = vm;
3049 }
3050
3051 /**
3052 * vm_area_register_early - register vmap area early during boot
3053 * @vm: vm_struct to register
3054 * @align: requested alignment
3055 *
3056 * This function is used to register kernel vm area before
3057 * vmalloc_init() is called. @vm->size and @vm->flags should contain
3058 * proper values on entry and other fields should be zero. On return,
3059 * vm->addr contains the allocated address.
3060 *
3061 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3062 */
3063 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
3064 {
3065 unsigned long addr = ALIGN(VMALLOC_START, align);
3066 struct vm_struct *cur, **p;
3067
3068 BUG_ON(vmap_initialized);
3069
3070 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
3071 if ((unsigned long)cur->addr - addr >= vm->size)
3072 break;
3073 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
3074 }
3075
3076 BUG_ON(addr > VMALLOC_END - vm->size);
3077 vm->addr = (void *)addr;
3078 vm->next = *p;
3079 *p = vm;
3080 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
3081 }
3082
3083 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
3084 {
3085 /*
3086 * Before removing VM_UNINITIALIZED,
3087 * we should make sure that vm has proper values.
3088 * Pair with smp_rmb() in show_numa_info().
3089 */
3090 smp_wmb();
3091 vm->flags &= ~VM_UNINITIALIZED;
3092 }
3093
3094 static struct vm_struct *__get_vm_area_node(unsigned long size,
3095 unsigned long align, unsigned long shift, unsigned long flags,
3096 unsigned long start, unsigned long end, int node,
3097 gfp_t gfp_mask, const void *caller)
3098 {
3099 struct vmap_area *va;
3100 struct vm_struct *area;
3101 unsigned long requested_size = size;
3102
3103 BUG_ON(in_interrupt());
3104 size = ALIGN(size, 1ul << shift);
3105 if (unlikely(!size))
3106 return NULL;
3107
3108 if (flags & VM_IOREMAP)
3109 align = 1ul << clamp_t(int, get_count_order_long(size),
3110 PAGE_SHIFT, IOREMAP_MAX_ORDER);
3111
3112 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
3113 if (unlikely(!area))
3114 return NULL;
3115
3116 if (!(flags & VM_NO_GUARD))
3117 size += PAGE_SIZE;
3118
3119 area->flags = flags;
3120 area->caller = caller;
3121
3122 va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0, area);
3123 if (IS_ERR(va)) {
3124 kfree(area);
3125 return NULL;
3126 }
3127
3128 /*
3129 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
3130 * best-effort approach, as they can be mapped outside of vmalloc code.
3131 * For VM_ALLOC mappings, the pages are marked as accessible after
3132 * getting mapped in __vmalloc_node_range().
3133 * With hardware tag-based KASAN, marking is skipped for
3134 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3135 */
3136 if (!(flags & VM_ALLOC))
3137 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
3138 KASAN_VMALLOC_PROT_NORMAL);
3139
3140 return area;
3141 }
3142
3143 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
3144 unsigned long start, unsigned long end,
3145 const void *caller)
3146 {
3147 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
3148 NUMA_NO_NODE, GFP_KERNEL, caller);
3149 }
3150
3151 /**
3152 * get_vm_area - reserve a contiguous kernel virtual area
3153 * @size: size of the area
3154 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
3155 *
3156 * Search an area of @size in the kernel virtual mapping area,
3157 * and reserved it for out purposes. Returns the area descriptor
3158 * on success or %NULL on failure.
3159 *
3160 * Return: the area descriptor on success or %NULL on failure.
3161 */
3162 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
3163 {
3164 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3165 VMALLOC_START, VMALLOC_END,
3166 NUMA_NO_NODE, GFP_KERNEL,
3167 __builtin_return_address(0));
3168 }
3169
3170 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
3171 const void *caller)
3172 {
3173 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3174 VMALLOC_START, VMALLOC_END,
3175 NUMA_NO_NODE, GFP_KERNEL, caller);
3176 }
3177
3178 /**
3179 * find_vm_area - find a continuous kernel virtual area
3180 * @addr: base address
3181 *
3182 * Search for the kernel VM area starting at @addr, and return it.
3183 * It is up to the caller to do all required locking to keep the returned
3184 * pointer valid.
3185 *
3186 * Return: the area descriptor on success or %NULL on failure.
3187 */
3188 struct vm_struct *find_vm_area(const void *addr)
3189 {
3190 struct vmap_area *va;
3191
3192 va = find_vmap_area((unsigned long)addr);
3193 if (!va)
3194 return NULL;
3195
3196 return va->vm;
3197 }
3198
3199 /**
3200 * remove_vm_area - find and remove a continuous kernel virtual area
3201 * @addr: base address
3202 *
3203 * Search for the kernel VM area starting at @addr, and remove it.
3204 * This function returns the found VM area, but using it is NOT safe
3205 * on SMP machines, except for its size or flags.
3206 *
3207 * Return: the area descriptor on success or %NULL on failure.
3208 */
3209 struct vm_struct *remove_vm_area(const void *addr)
3210 {
3211 struct vmap_area *va;
3212 struct vm_struct *vm;
3213
3214 might_sleep();
3215
3216 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
3217 addr))
3218 return NULL;
3219
3220 va = find_unlink_vmap_area((unsigned long)addr);
3221 if (!va || !va->vm)
3222 return NULL;
3223 vm = va->vm;
3224
3225 debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
3226 debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
3227 kasan_free_module_shadow(vm);
3228 kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
3229
3230 free_unmap_vmap_area(va);
3231 return vm;
3232 }
3233
3234 static inline void set_area_direct_map(const struct vm_struct *area,
3235 int (*set_direct_map)(struct page *page))
3236 {
3237 int i;
3238
3239 /* HUGE_VMALLOC passes small pages to set_direct_map */
3240 for (i = 0; i < area->nr_pages; i++)
3241 if (page_address(area->pages[i]))
3242 set_direct_map(area->pages[i]);
3243 }
3244
3245 /*
3246 * Flush the vm mapping and reset the direct map.
3247 */
3248 static void vm_reset_perms(struct vm_struct *area)
3249 {
3250 unsigned long start = ULONG_MAX, end = 0;
3251 unsigned int page_order = vm_area_page_order(area);
3252 int flush_dmap = 0;
3253 int i;
3254
3255 /*
3256 * Find the start and end range of the direct mappings to make sure that
3257 * the vm_unmap_aliases() flush includes the direct map.
3258 */
3259 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
3260 unsigned long addr = (unsigned long)page_address(area->pages[i]);
3261
3262 if (addr) {
3263 unsigned long page_size;
3264
3265 page_size = PAGE_SIZE << page_order;
3266 start = min(addr, start);
3267 end = max(addr + page_size, end);
3268 flush_dmap = 1;
3269 }
3270 }
3271
3272 /*
3273 * Set direct map to something invalid so that it won't be cached if
3274 * there are any accesses after the TLB flush, then flush the TLB and
3275 * reset the direct map permissions to the default.
3276 */
3277 set_area_direct_map(area, set_direct_map_invalid_noflush);
3278 _vm_unmap_aliases(start, end, flush_dmap);
3279 set_area_direct_map(area, set_direct_map_default_noflush);
3280 }
3281
3282 static void delayed_vfree_work(struct work_struct *w)
3283 {
3284 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
3285 struct llist_node *t, *llnode;
3286
3287 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
3288 vfree(llnode);
3289 }
3290
3291 /**
3292 * vfree_atomic - release memory allocated by vmalloc()
3293 * @addr: memory base address
3294 *
3295 * This one is just like vfree() but can be called in any atomic context
3296 * except NMIs.
3297 */
3298 void vfree_atomic(const void *addr)
3299 {
3300 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
3301
3302 BUG_ON(in_nmi());
3303 kmemleak_free(addr);
3304
3305 /*
3306 * Use raw_cpu_ptr() because this can be called from preemptible
3307 * context. Preemption is absolutely fine here, because the llist_add()
3308 * implementation is lockless, so it works even if we are adding to
3309 * another cpu's list. schedule_work() should be fine with this too.
3310 */
3311 if (addr && llist_add((struct llist_node *)addr, &p->list))
3312 schedule_work(&p->wq);
3313 }
3314
3315 /**
3316 * vfree - Release memory allocated by vmalloc()
3317 * @addr: Memory base address
3318 *
3319 * Free the virtually continuous memory area starting at @addr, as obtained
3320 * from one of the vmalloc() family of APIs. This will usually also free the
3321 * physical memory underlying the virtual allocation, but that memory is
3322 * reference counted, so it will not be freed until the last user goes away.
3323 *
3324 * If @addr is NULL, no operation is performed.
3325 *
3326 * Context:
3327 * May sleep if called *not* from interrupt context.
3328 * Must not be called in NMI context (strictly speaking, it could be
3329 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
3330 * conventions for vfree() arch-dependent would be a really bad idea).
3331 */
3332 void vfree(const void *addr)
3333 {
3334 struct vm_struct *vm;
3335 int i;
3336
3337 if (unlikely(in_interrupt())) {
3338 vfree_atomic(addr);
3339 return;
3340 }
3341
3342 BUG_ON(in_nmi());
3343 kmemleak_free(addr);
3344 might_sleep();
3345
3346 if (!addr)
3347 return;
3348
3349 vm = remove_vm_area(addr);
3350 if (unlikely(!vm)) {
3351 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
3352 addr);
3353 return;
3354 }
3355
3356 if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
3357 vm_reset_perms(vm);
3358 for (i = 0; i < vm->nr_pages; i++) {
3359 struct page *page = vm->pages[i];
3360
3361 BUG_ON(!page);
3362 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
3363 /*
3364 * High-order allocs for huge vmallocs are split, so
3365 * can be freed as an array of order-0 allocations
3366 */
3367 __free_page(page);
3368 cond_resched();
3369 }
3370 atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
3371 kvfree(vm->pages);
3372 kfree(vm);
3373 }
3374 EXPORT_SYMBOL(vfree);
3375
3376 /**
3377 * vunmap - release virtual mapping obtained by vmap()
3378 * @addr: memory base address
3379 *
3380 * Free the virtually contiguous memory area starting at @addr,
3381 * which was created from the page array passed to vmap().
3382 *
3383 * Must not be called in interrupt context.
3384 */
3385 void vunmap(const void *addr)
3386 {
3387 struct vm_struct *vm;
3388
3389 BUG_ON(in_interrupt());
3390 might_sleep();
3391
3392 if (!addr)
3393 return;
3394 vm = remove_vm_area(addr);
3395 if (unlikely(!vm)) {
3396 WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
3397 addr);
3398 return;
3399 }
3400 kfree(vm);
3401 }
3402 EXPORT_SYMBOL(vunmap);
3403
3404 /**
3405 * vmap - map an array of pages into virtually contiguous space
3406 * @pages: array of page pointers
3407 * @count: number of pages to map
3408 * @flags: vm_area->flags
3409 * @prot: page protection for the mapping
3410 *
3411 * Maps @count pages from @pages into contiguous kernel virtual space.
3412 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
3413 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
3414 * are transferred from the caller to vmap(), and will be freed / dropped when
3415 * vfree() is called on the return value.
3416 *
3417 * Return: the address of the area or %NULL on failure
3418 */
3419 void *vmap(struct page **pages, unsigned int count,
3420 unsigned long flags, pgprot_t prot)
3421 {
3422 struct vm_struct *area;
3423 unsigned long addr;
3424 unsigned long size; /* In bytes */
3425
3426 might_sleep();
3427
3428 if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
3429 return NULL;
3430
3431 /*
3432 * Your top guard is someone else's bottom guard. Not having a top
3433 * guard compromises someone else's mappings too.
3434 */
3435 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
3436 flags &= ~VM_NO_GUARD;
3437
3438 if (count > totalram_pages())
3439 return NULL;
3440
3441 size = (unsigned long)count << PAGE_SHIFT;
3442 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
3443 if (!area)
3444 return NULL;
3445
3446 addr = (unsigned long)area->addr;
3447 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
3448 pages, PAGE_SHIFT) < 0) {
3449 vunmap(area->addr);
3450 return NULL;
3451 }
3452
3453 if (flags & VM_MAP_PUT_PAGES) {
3454 area->pages = pages;
3455 area->nr_pages = count;
3456 }
3457 return area->addr;
3458 }
3459 EXPORT_SYMBOL(vmap);
3460
3461 #ifdef CONFIG_VMAP_PFN
3462 struct vmap_pfn_data {
3463 unsigned long *pfns;
3464 pgprot_t prot;
3465 unsigned int idx;
3466 };
3467
3468 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
3469 {
3470 struct vmap_pfn_data *data = private;
3471 unsigned long pfn = data->pfns[data->idx];
3472 pte_t ptent;
3473
3474 if (WARN_ON_ONCE(pfn_valid(pfn)))
3475 return -EINVAL;
3476
3477 ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
3478 set_pte_at(&init_mm, addr, pte, ptent);
3479
3480 data->idx++;
3481 return 0;
3482 }
3483
3484 /**
3485 * vmap_pfn - map an array of PFNs into virtually contiguous space
3486 * @pfns: array of PFNs
3487 * @count: number of pages to map
3488 * @prot: page protection for the mapping
3489 *
3490 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
3491 * the start address of the mapping.
3492 */
3493 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
3494 {
3495 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
3496 struct vm_struct *area;
3497
3498 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
3499 __builtin_return_address(0));
3500 if (!area)
3501 return NULL;
3502 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3503 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
3504 free_vm_area(area);
3505 return NULL;
3506 }
3507
3508 flush_cache_vmap((unsigned long)area->addr,
3509 (unsigned long)area->addr + count * PAGE_SIZE);
3510
3511 return area->addr;
3512 }
3513 EXPORT_SYMBOL_GPL(vmap_pfn);
3514 #endif /* CONFIG_VMAP_PFN */
3515
3516 static inline unsigned int
3517 vm_area_alloc_pages(gfp_t gfp, int nid,
3518 unsigned int order, unsigned int nr_pages, struct page **pages)
3519 {
3520 unsigned int nr_allocated = 0;
3521 gfp_t alloc_gfp = gfp;
3522 bool nofail = gfp & __GFP_NOFAIL;
3523 struct page *page;
3524 int i;
3525
3526 /*
3527 * For order-0 pages we make use of bulk allocator, if
3528 * the page array is partly or not at all populated due
3529 * to fails, fallback to a single page allocator that is
3530 * more permissive.
3531 */
3532 if (!order) {
3533 /* bulk allocator doesn't support nofail req. officially */
3534 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
3535
3536 while (nr_allocated < nr_pages) {
3537 unsigned int nr, nr_pages_request;
3538
3539 /*
3540 * A maximum allowed request is hard-coded and is 100
3541 * pages per call. That is done in order to prevent a
3542 * long preemption off scenario in the bulk-allocator
3543 * so the range is [1:100].
3544 */
3545 nr_pages_request = min(100U, nr_pages - nr_allocated);
3546
3547 /* memory allocation should consider mempolicy, we can't
3548 * wrongly use nearest node when nid == NUMA_NO_NODE,
3549 * otherwise memory may be allocated in only one node,
3550 * but mempolicy wants to alloc memory by interleaving.
3551 */
3552 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3553 nr = alloc_pages_bulk_array_mempolicy_noprof(bulk_gfp,
3554 nr_pages_request,
3555 pages + nr_allocated);
3556
3557 else
3558 nr = alloc_pages_bulk_array_node_noprof(bulk_gfp, nid,
3559 nr_pages_request,
3560 pages + nr_allocated);
3561
3562 nr_allocated += nr;
3563 cond_resched();
3564
3565 /*
3566 * If zero or pages were obtained partly,
3567 * fallback to a single page allocator.
3568 */
3569 if (nr != nr_pages_request)
3570 break;
3571 }
3572 } else if (gfp & __GFP_NOFAIL) {
3573 /*
3574 * Higher order nofail allocations are really expensive and
3575 * potentially dangerous (pre-mature OOM, disruptive reclaim
3576 * and compaction etc.
3577 */
3578 alloc_gfp &= ~__GFP_NOFAIL;
3579 }
3580
3581 /* High-order pages or fallback path if "bulk" fails. */
3582 while (nr_allocated < nr_pages) {
3583 if (!nofail && fatal_signal_pending(current))
3584 break;
3585
3586 if (nid == NUMA_NO_NODE)
3587 page = alloc_pages_noprof(alloc_gfp, order);
3588 else
3589 page = alloc_pages_node_noprof(nid, alloc_gfp, order);
3590 if (unlikely(!page))
3591 break;
3592
3593 /*
3594 * Higher order allocations must be able to be treated as
3595 * indepdenent small pages by callers (as they can with
3596 * small-page vmallocs). Some drivers do their own refcounting
3597 * on vmalloc_to_page() pages, some use page->mapping,
3598 * page->lru, etc.
3599 */
3600 if (order)
3601 split_page(page, order);
3602
3603 /*
3604 * Careful, we allocate and map page-order pages, but
3605 * tracking is done per PAGE_SIZE page so as to keep the
3606 * vm_struct APIs independent of the physical/mapped size.
3607 */
3608 for (i = 0; i < (1U << order); i++)
3609 pages[nr_allocated + i] = page + i;
3610
3611 cond_resched();
3612 nr_allocated += 1U << order;
3613 }
3614
3615 return nr_allocated;
3616 }
3617
3618 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3619 pgprot_t prot, unsigned int page_shift,
3620 int node)
3621 {
3622 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3623 bool nofail = gfp_mask & __GFP_NOFAIL;
3624 unsigned long addr = (unsigned long)area->addr;
3625 unsigned long size = get_vm_area_size(area);
3626 unsigned long array_size;
3627 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3628 unsigned int page_order;
3629 unsigned int flags;
3630 int ret;
3631
3632 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3633
3634 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3635 gfp_mask |= __GFP_HIGHMEM;
3636
3637 /* Please note that the recursion is strictly bounded. */
3638 if (array_size > PAGE_SIZE) {
3639 area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node,
3640 area->caller);
3641 } else {
3642 area->pages = kmalloc_node_noprof(array_size, nested_gfp, node);
3643 }
3644
3645 if (!area->pages) {
3646 warn_alloc(gfp_mask, NULL,
3647 "vmalloc error: size %lu, failed to allocated page array size %lu",
3648 nr_small_pages * PAGE_SIZE, array_size);
3649 free_vm_area(area);
3650 return NULL;
3651 }
3652
3653 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3654 page_order = vm_area_page_order(area);
3655
3656 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3657 node, page_order, nr_small_pages, area->pages);
3658
3659 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3660 if (gfp_mask & __GFP_ACCOUNT) {
3661 int i;
3662
3663 for (i = 0; i < area->nr_pages; i++)
3664 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3665 }
3666
3667 /*
3668 * If not enough pages were obtained to accomplish an
3669 * allocation request, free them via vfree() if any.
3670 */
3671 if (area->nr_pages != nr_small_pages) {
3672 /*
3673 * vm_area_alloc_pages() can fail due to insufficient memory but
3674 * also:-
3675 *
3676 * - a pending fatal signal
3677 * - insufficient huge page-order pages
3678 *
3679 * Since we always retry allocations at order-0 in the huge page
3680 * case a warning for either is spurious.
3681 */
3682 if (!fatal_signal_pending(current) && page_order == 0)
3683 warn_alloc(gfp_mask, NULL,
3684 "vmalloc error: size %lu, failed to allocate pages",
3685 area->nr_pages * PAGE_SIZE);
3686 goto fail;
3687 }
3688
3689 /*
3690 * page tables allocations ignore external gfp mask, enforce it
3691 * by the scope API
3692 */
3693 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3694 flags = memalloc_nofs_save();
3695 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3696 flags = memalloc_noio_save();
3697
3698 do {
3699 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3700 page_shift);
3701 if (nofail && (ret < 0))
3702 schedule_timeout_uninterruptible(1);
3703 } while (nofail && (ret < 0));
3704
3705 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3706 memalloc_nofs_restore(flags);
3707 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3708 memalloc_noio_restore(flags);
3709
3710 if (ret < 0) {
3711 warn_alloc(gfp_mask, NULL,
3712 "vmalloc error: size %lu, failed to map pages",
3713 area->nr_pages * PAGE_SIZE);
3714 goto fail;
3715 }
3716
3717 return area->addr;
3718
3719 fail:
3720 vfree(area->addr);
3721 return NULL;
3722 }
3723
3724 /**
3725 * __vmalloc_node_range - allocate virtually contiguous memory
3726 * @size: allocation size
3727 * @align: desired alignment
3728 * @start: vm area range start
3729 * @end: vm area range end
3730 * @gfp_mask: flags for the page level allocator
3731 * @prot: protection mask for the allocated pages
3732 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3733 * @node: node to use for allocation or NUMA_NO_NODE
3734 * @caller: caller's return address
3735 *
3736 * Allocate enough pages to cover @size from the page level
3737 * allocator with @gfp_mask flags. Please note that the full set of gfp
3738 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3739 * supported.
3740 * Zone modifiers are not supported. From the reclaim modifiers
3741 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3742 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3743 * __GFP_RETRY_MAYFAIL are not supported).
3744 *
3745 * __GFP_NOWARN can be used to suppress failures messages.
3746 *
3747 * Map them into contiguous kernel virtual space, using a pagetable
3748 * protection of @prot.
3749 *
3750 * Return: the address of the area or %NULL on failure
3751 */
3752 void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align,
3753 unsigned long start, unsigned long end, gfp_t gfp_mask,
3754 pgprot_t prot, unsigned long vm_flags, int node,
3755 const void *caller)
3756 {
3757 struct vm_struct *area;
3758 void *ret;
3759 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3760 unsigned long real_size = size;
3761 unsigned long real_align = align;
3762 unsigned int shift = PAGE_SHIFT;
3763
3764 if (WARN_ON_ONCE(!size))
3765 return NULL;
3766
3767 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3768 warn_alloc(gfp_mask, NULL,
3769 "vmalloc error: size %lu, exceeds total pages",
3770 real_size);
3771 return NULL;
3772 }
3773
3774 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3775 unsigned long size_per_node;
3776
3777 /*
3778 * Try huge pages. Only try for PAGE_KERNEL allocations,
3779 * others like modules don't yet expect huge pages in
3780 * their allocations due to apply_to_page_range not
3781 * supporting them.
3782 */
3783
3784 size_per_node = size;
3785 if (node == NUMA_NO_NODE)
3786 size_per_node /= num_online_nodes();
3787 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3788 shift = PMD_SHIFT;
3789 else
3790 shift = arch_vmap_pte_supported_shift(size_per_node);
3791
3792 align = max(real_align, 1UL << shift);
3793 size = ALIGN(real_size, 1UL << shift);
3794 }
3795
3796 again:
3797 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3798 VM_UNINITIALIZED | vm_flags, start, end, node,
3799 gfp_mask, caller);
3800 if (!area) {
3801 bool nofail = gfp_mask & __GFP_NOFAIL;
3802 warn_alloc(gfp_mask, NULL,
3803 "vmalloc error: size %lu, vm_struct allocation failed%s",
3804 real_size, (nofail) ? ". Retrying." : "");
3805 if (nofail) {
3806 schedule_timeout_uninterruptible(1);
3807 goto again;
3808 }
3809 goto fail;
3810 }
3811
3812 /*
3813 * Prepare arguments for __vmalloc_area_node() and
3814 * kasan_unpoison_vmalloc().
3815 */
3816 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3817 if (kasan_hw_tags_enabled()) {
3818 /*
3819 * Modify protection bits to allow tagging.
3820 * This must be done before mapping.
3821 */
3822 prot = arch_vmap_pgprot_tagged(prot);
3823
3824 /*
3825 * Skip page_alloc poisoning and zeroing for physical
3826 * pages backing VM_ALLOC mapping. Memory is instead
3827 * poisoned and zeroed by kasan_unpoison_vmalloc().
3828 */
3829 gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3830 }
3831
3832 /* Take note that the mapping is PAGE_KERNEL. */
3833 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3834 }
3835
3836 /* Allocate physical pages and map them into vmalloc space. */
3837 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3838 if (!ret)
3839 goto fail;
3840
3841 /*
3842 * Mark the pages as accessible, now that they are mapped.
3843 * The condition for setting KASAN_VMALLOC_INIT should complement the
3844 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3845 * to make sure that memory is initialized under the same conditions.
3846 * Tag-based KASAN modes only assign tags to normal non-executable
3847 * allocations, see __kasan_unpoison_vmalloc().
3848 */
3849 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3850 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3851 (gfp_mask & __GFP_SKIP_ZERO))
3852 kasan_flags |= KASAN_VMALLOC_INIT;
3853 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3854 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3855
3856 /*
3857 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3858 * flag. It means that vm_struct is not fully initialized.
3859 * Now, it is fully initialized, so remove this flag here.
3860 */
3861 clear_vm_uninitialized_flag(area);
3862
3863 size = PAGE_ALIGN(size);
3864 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3865 kmemleak_vmalloc(area, size, gfp_mask);
3866
3867 return area->addr;
3868
3869 fail:
3870 if (shift > PAGE_SHIFT) {
3871 shift = PAGE_SHIFT;
3872 align = real_align;
3873 size = real_size;
3874 goto again;
3875 }
3876
3877 return NULL;
3878 }
3879
3880 /**
3881 * __vmalloc_node - allocate virtually contiguous memory
3882 * @size: allocation size
3883 * @align: desired alignment
3884 * @gfp_mask: flags for the page level allocator
3885 * @node: node to use for allocation or NUMA_NO_NODE
3886 * @caller: caller's return address
3887 *
3888 * Allocate enough pages to cover @size from the page level allocator with
3889 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3890 *
3891 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3892 * and __GFP_NOFAIL are not supported
3893 *
3894 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3895 * with mm people.
3896 *
3897 * Return: pointer to the allocated memory or %NULL on error
3898 */
3899 void *__vmalloc_node_noprof(unsigned long size, unsigned long align,
3900 gfp_t gfp_mask, int node, const void *caller)
3901 {
3902 return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END,
3903 gfp_mask, PAGE_KERNEL, 0, node, caller);
3904 }
3905 /*
3906 * This is only for performance analysis of vmalloc and stress purpose.
3907 * It is required by vmalloc test module, therefore do not use it other
3908 * than that.
3909 */
3910 #ifdef CONFIG_TEST_VMALLOC_MODULE
3911 EXPORT_SYMBOL_GPL(__vmalloc_node_noprof);
3912 #endif
3913
3914 void *__vmalloc_noprof(unsigned long size, gfp_t gfp_mask)
3915 {
3916 return __vmalloc_node_noprof(size, 1, gfp_mask, NUMA_NO_NODE,
3917 __builtin_return_address(0));
3918 }
3919 EXPORT_SYMBOL(__vmalloc_noprof);
3920
3921 /**
3922 * vmalloc - allocate virtually contiguous memory
3923 * @size: allocation size
3924 *
3925 * Allocate enough pages to cover @size from the page level
3926 * allocator and map them into contiguous kernel virtual space.
3927 *
3928 * For tight control over page level allocator and protection flags
3929 * use __vmalloc() instead.
3930 *
3931 * Return: pointer to the allocated memory or %NULL on error
3932 */
3933 void *vmalloc_noprof(unsigned long size)
3934 {
3935 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3936 __builtin_return_address(0));
3937 }
3938 EXPORT_SYMBOL(vmalloc_noprof);
3939
3940 /**
3941 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3942 * @size: allocation size
3943 * @gfp_mask: flags for the page level allocator
3944 *
3945 * Allocate enough pages to cover @size from the page level
3946 * allocator and map them into contiguous kernel virtual space.
3947 * If @size is greater than or equal to PMD_SIZE, allow using
3948 * huge pages for the memory
3949 *
3950 * Return: pointer to the allocated memory or %NULL on error
3951 */
3952 void *vmalloc_huge_noprof(unsigned long size, gfp_t gfp_mask)
3953 {
3954 return __vmalloc_node_range_noprof(size, 1, VMALLOC_START, VMALLOC_END,
3955 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3956 NUMA_NO_NODE, __builtin_return_address(0));
3957 }
3958 EXPORT_SYMBOL_GPL(vmalloc_huge_noprof);
3959
3960 /**
3961 * vzalloc - allocate virtually contiguous memory with zero fill
3962 * @size: allocation size
3963 *
3964 * Allocate enough pages to cover @size from the page level
3965 * allocator and map them into contiguous kernel virtual space.
3966 * The memory allocated is set to zero.
3967 *
3968 * For tight control over page level allocator and protection flags
3969 * use __vmalloc() instead.
3970 *
3971 * Return: pointer to the allocated memory or %NULL on error
3972 */
3973 void *vzalloc_noprof(unsigned long size)
3974 {
3975 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3976 __builtin_return_address(0));
3977 }
3978 EXPORT_SYMBOL(vzalloc_noprof);
3979
3980 /**
3981 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3982 * @size: allocation size
3983 *
3984 * The resulting memory area is zeroed so it can be mapped to userspace
3985 * without leaking data.
3986 *
3987 * Return: pointer to the allocated memory or %NULL on error
3988 */
3989 void *vmalloc_user_noprof(unsigned long size)
3990 {
3991 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3992 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3993 VM_USERMAP, NUMA_NO_NODE,
3994 __builtin_return_address(0));
3995 }
3996 EXPORT_SYMBOL(vmalloc_user_noprof);
3997
3998 /**
3999 * vmalloc_node - allocate memory on a specific node
4000 * @size: allocation size
4001 * @node: numa node
4002 *
4003 * Allocate enough pages to cover @size from the page level
4004 * allocator and map them into contiguous kernel virtual space.
4005 *
4006 * For tight control over page level allocator and protection flags
4007 * use __vmalloc() instead.
4008 *
4009 * Return: pointer to the allocated memory or %NULL on error
4010 */
4011 void *vmalloc_node_noprof(unsigned long size, int node)
4012 {
4013 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, node,
4014 __builtin_return_address(0));
4015 }
4016 EXPORT_SYMBOL(vmalloc_node_noprof);
4017
4018 /**
4019 * vzalloc_node - allocate memory on a specific node with zero fill
4020 * @size: allocation size
4021 * @node: numa node
4022 *
4023 * Allocate enough pages to cover @size from the page level
4024 * allocator and map them into contiguous kernel virtual space.
4025 * The memory allocated is set to zero.
4026 *
4027 * Return: pointer to the allocated memory or %NULL on error
4028 */
4029 void *vzalloc_node_noprof(unsigned long size, int node)
4030 {
4031 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, node,
4032 __builtin_return_address(0));
4033 }
4034 EXPORT_SYMBOL(vzalloc_node_noprof);
4035
4036 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
4037 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4038 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
4039 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
4040 #else
4041 /*
4042 * 64b systems should always have either DMA or DMA32 zones. For others
4043 * GFP_DMA32 should do the right thing and use the normal zone.
4044 */
4045 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4046 #endif
4047
4048 /**
4049 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
4050 * @size: allocation size
4051 *
4052 * Allocate enough 32bit PA addressable pages to cover @size from the
4053 * page level allocator and map them into contiguous kernel virtual space.
4054 *
4055 * Return: pointer to the allocated memory or %NULL on error
4056 */
4057 void *vmalloc_32_noprof(unsigned long size)
4058 {
4059 return __vmalloc_node_noprof(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
4060 __builtin_return_address(0));
4061 }
4062 EXPORT_SYMBOL(vmalloc_32_noprof);
4063
4064 /**
4065 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
4066 * @size: allocation size
4067 *
4068 * The resulting memory area is 32bit addressable and zeroed so it can be
4069 * mapped to userspace without leaking data.
4070 *
4071 * Return: pointer to the allocated memory or %NULL on error
4072 */
4073 void *vmalloc_32_user_noprof(unsigned long size)
4074 {
4075 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
4076 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
4077 VM_USERMAP, NUMA_NO_NODE,
4078 __builtin_return_address(0));
4079 }
4080 EXPORT_SYMBOL(vmalloc_32_user_noprof);
4081
4082 /*
4083 * Atomically zero bytes in the iterator.
4084 *
4085 * Returns the number of zeroed bytes.
4086 */
4087 static size_t zero_iter(struct iov_iter *iter, size_t count)
4088 {
4089 size_t remains = count;
4090
4091 while (remains > 0) {
4092 size_t num, copied;
4093
4094 num = min_t(size_t, remains, PAGE_SIZE);
4095 copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
4096 remains -= copied;
4097
4098 if (copied < num)
4099 break;
4100 }
4101
4102 return count - remains;
4103 }
4104
4105 /*
4106 * small helper routine, copy contents to iter from addr.
4107 * If the page is not present, fill zero.
4108 *
4109 * Returns the number of copied bytes.
4110 */
4111 static size_t aligned_vread_iter(struct iov_iter *iter,
4112 const char *addr, size_t count)
4113 {
4114 size_t remains = count;
4115 struct page *page;
4116
4117 while (remains > 0) {
4118 unsigned long offset, length;
4119 size_t copied = 0;
4120
4121 offset = offset_in_page(addr);
4122 length = PAGE_SIZE - offset;
4123 if (length > remains)
4124 length = remains;
4125 page = vmalloc_to_page(addr);
4126 /*
4127 * To do safe access to this _mapped_ area, we need lock. But
4128 * adding lock here means that we need to add overhead of
4129 * vmalloc()/vfree() calls for this _debug_ interface, rarely
4130 * used. Instead of that, we'll use an local mapping via
4131 * copy_page_to_iter_nofault() and accept a small overhead in
4132 * this access function.
4133 */
4134 if (page)
4135 copied = copy_page_to_iter_nofault(page, offset,
4136 length, iter);
4137 else
4138 copied = zero_iter(iter, length);
4139
4140 addr += copied;
4141 remains -= copied;
4142
4143 if (copied != length)
4144 break;
4145 }
4146
4147 return count - remains;
4148 }
4149
4150 /*
4151 * Read from a vm_map_ram region of memory.
4152 *
4153 * Returns the number of copied bytes.
4154 */
4155 static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
4156 size_t count, unsigned long flags)
4157 {
4158 char *start;
4159 struct vmap_block *vb;
4160 struct xarray *xa;
4161 unsigned long offset;
4162 unsigned int rs, re;
4163 size_t remains, n;
4164
4165 /*
4166 * If it's area created by vm_map_ram() interface directly, but
4167 * not further subdividing and delegating management to vmap_block,
4168 * handle it here.
4169 */
4170 if (!(flags & VMAP_BLOCK))
4171 return aligned_vread_iter(iter, addr, count);
4172
4173 remains = count;
4174
4175 /*
4176 * Area is split into regions and tracked with vmap_block, read out
4177 * each region and zero fill the hole between regions.
4178 */
4179 xa = addr_to_vb_xa((unsigned long) addr);
4180 vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
4181 if (!vb)
4182 goto finished_zero;
4183
4184 spin_lock(&vb->lock);
4185 if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
4186 spin_unlock(&vb->lock);
4187 goto finished_zero;
4188 }
4189
4190 for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
4191 size_t copied;
4192
4193 if (remains == 0)
4194 goto finished;
4195
4196 start = vmap_block_vaddr(vb->va->va_start, rs);
4197
4198 if (addr < start) {
4199 size_t to_zero = min_t(size_t, start - addr, remains);
4200 size_t zeroed = zero_iter(iter, to_zero);
4201
4202 addr += zeroed;
4203 remains -= zeroed;
4204
4205 if (remains == 0 || zeroed != to_zero)
4206 goto finished;
4207 }
4208
4209 /*it could start reading from the middle of used region*/
4210 offset = offset_in_page(addr);
4211 n = ((re - rs + 1) << PAGE_SHIFT) - offset;
4212 if (n > remains)
4213 n = remains;
4214
4215 copied = aligned_vread_iter(iter, start + offset, n);
4216
4217 addr += copied;
4218 remains -= copied;
4219
4220 if (copied != n)
4221 goto finished;
4222 }
4223
4224 spin_unlock(&vb->lock);
4225
4226 finished_zero:
4227 /* zero-fill the left dirty or free regions */
4228 return count - remains + zero_iter(iter, remains);
4229 finished:
4230 /* We couldn't copy/zero everything */
4231 spin_unlock(&vb->lock);
4232 return count - remains;
4233 }
4234
4235 /**
4236 * vread_iter() - read vmalloc area in a safe way to an iterator.
4237 * @iter: the iterator to which data should be written.
4238 * @addr: vm address.
4239 * @count: number of bytes to be read.
4240 *
4241 * This function checks that addr is a valid vmalloc'ed area, and
4242 * copy data from that area to a given buffer. If the given memory range
4243 * of [addr...addr+count) includes some valid address, data is copied to
4244 * proper area of @buf. If there are memory holes, they'll be zero-filled.
4245 * IOREMAP area is treated as memory hole and no copy is done.
4246 *
4247 * If [addr...addr+count) doesn't includes any intersects with alive
4248 * vm_struct area, returns 0. @buf should be kernel's buffer.
4249 *
4250 * Note: In usual ops, vread() is never necessary because the caller
4251 * should know vmalloc() area is valid and can use memcpy().
4252 * This is for routines which have to access vmalloc area without
4253 * any information, as /proc/kcore.
4254 *
4255 * Return: number of bytes for which addr and buf should be increased
4256 * (same number as @count) or %0 if [addr...addr+count) doesn't
4257 * include any intersection with valid vmalloc area
4258 */
4259 long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
4260 {
4261 struct vmap_node *vn;
4262 struct vmap_area *va;
4263 struct vm_struct *vm;
4264 char *vaddr;
4265 size_t n, size, flags, remains;
4266 unsigned long next;
4267
4268 addr = kasan_reset_tag(addr);
4269
4270 /* Don't allow overflow */
4271 if ((unsigned long) addr + count < count)
4272 count = -(unsigned long) addr;
4273
4274 remains = count;
4275
4276 vn = find_vmap_area_exceed_addr_lock((unsigned long) addr, &va);
4277 if (!vn)
4278 goto finished_zero;
4279
4280 /* no intersects with alive vmap_area */
4281 if ((unsigned long)addr + remains <= va->va_start)
4282 goto finished_zero;
4283
4284 do {
4285 size_t copied;
4286
4287 if (remains == 0)
4288 goto finished;
4289
4290 vm = va->vm;
4291 flags = va->flags & VMAP_FLAGS_MASK;
4292 /*
4293 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
4294 * be set together with VMAP_RAM.
4295 */
4296 WARN_ON(flags == VMAP_BLOCK);
4297
4298 if (!vm && !flags)
4299 goto next_va;
4300
4301 if (vm && (vm->flags & VM_UNINITIALIZED))
4302 goto next_va;
4303
4304 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4305 smp_rmb();
4306
4307 vaddr = (char *) va->va_start;
4308 size = vm ? get_vm_area_size(vm) : va_size(va);
4309
4310 if (addr >= vaddr + size)
4311 goto next_va;
4312
4313 if (addr < vaddr) {
4314 size_t to_zero = min_t(size_t, vaddr - addr, remains);
4315 size_t zeroed = zero_iter(iter, to_zero);
4316
4317 addr += zeroed;
4318 remains -= zeroed;
4319
4320 if (remains == 0 || zeroed != to_zero)
4321 goto finished;
4322 }
4323
4324 n = vaddr + size - addr;
4325 if (n > remains)
4326 n = remains;
4327
4328 if (flags & VMAP_RAM)
4329 copied = vmap_ram_vread_iter(iter, addr, n, flags);
4330 else if (!(vm && (vm->flags & (VM_IOREMAP | VM_SPARSE))))
4331 copied = aligned_vread_iter(iter, addr, n);
4332 else /* IOREMAP | SPARSE area is treated as memory hole */
4333 copied = zero_iter(iter, n);
4334
4335 addr += copied;
4336 remains -= copied;
4337
4338 if (copied != n)
4339 goto finished;
4340
4341 next_va:
4342 next = va->va_end;
4343 spin_unlock(&vn->busy.lock);
4344 } while ((vn = find_vmap_area_exceed_addr_lock(next, &va)));
4345
4346 finished_zero:
4347 if (vn)
4348 spin_unlock(&vn->busy.lock);
4349
4350 /* zero-fill memory holes */
4351 return count - remains + zero_iter(iter, remains);
4352 finished:
4353 /* Nothing remains, or We couldn't copy/zero everything. */
4354 if (vn)
4355 spin_unlock(&vn->busy.lock);
4356
4357 return count - remains;
4358 }
4359
4360 /**
4361 * remap_vmalloc_range_partial - map vmalloc pages to userspace
4362 * @vma: vma to cover
4363 * @uaddr: target user address to start at
4364 * @kaddr: virtual address of vmalloc kernel memory
4365 * @pgoff: offset from @kaddr to start at
4366 * @size: size of map area
4367 *
4368 * Returns: 0 for success, -Exxx on failure
4369 *
4370 * This function checks that @kaddr is a valid vmalloc'ed area,
4371 * and that it is big enough to cover the range starting at
4372 * @uaddr in @vma. Will return failure if that criteria isn't
4373 * met.
4374 *
4375 * Similar to remap_pfn_range() (see mm/memory.c)
4376 */
4377 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
4378 void *kaddr, unsigned long pgoff,
4379 unsigned long size)
4380 {
4381 struct vm_struct *area;
4382 unsigned long off;
4383 unsigned long end_index;
4384
4385 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
4386 return -EINVAL;
4387
4388 size = PAGE_ALIGN(size);
4389
4390 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
4391 return -EINVAL;
4392
4393 area = find_vm_area(kaddr);
4394 if (!area)
4395 return -EINVAL;
4396
4397 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
4398 return -EINVAL;
4399
4400 if (check_add_overflow(size, off, &end_index) ||
4401 end_index > get_vm_area_size(area))
4402 return -EINVAL;
4403 kaddr += off;
4404
4405 do {
4406 struct page *page = vmalloc_to_page(kaddr);
4407 int ret;
4408
4409 ret = vm_insert_page(vma, uaddr, page);
4410 if (ret)
4411 return ret;
4412
4413 uaddr += PAGE_SIZE;
4414 kaddr += PAGE_SIZE;
4415 size -= PAGE_SIZE;
4416 } while (size > 0);
4417
4418 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
4419
4420 return 0;
4421 }
4422
4423 /**
4424 * remap_vmalloc_range - map vmalloc pages to userspace
4425 * @vma: vma to cover (map full range of vma)
4426 * @addr: vmalloc memory
4427 * @pgoff: number of pages into addr before first page to map
4428 *
4429 * Returns: 0 for success, -Exxx on failure
4430 *
4431 * This function checks that addr is a valid vmalloc'ed area, and
4432 * that it is big enough to cover the vma. Will return failure if
4433 * that criteria isn't met.
4434 *
4435 * Similar to remap_pfn_range() (see mm/memory.c)
4436 */
4437 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
4438 unsigned long pgoff)
4439 {
4440 return remap_vmalloc_range_partial(vma, vma->vm_start,
4441 addr, pgoff,
4442 vma->vm_end - vma->vm_start);
4443 }
4444 EXPORT_SYMBOL(remap_vmalloc_range);
4445
4446 void free_vm_area(struct vm_struct *area)
4447 {
4448 struct vm_struct *ret;
4449 ret = remove_vm_area(area->addr);
4450 BUG_ON(ret != area);
4451 kfree(area);
4452 }
4453 EXPORT_SYMBOL_GPL(free_vm_area);
4454
4455 #ifdef CONFIG_SMP
4456 static struct vmap_area *node_to_va(struct rb_node *n)
4457 {
4458 return rb_entry_safe(n, struct vmap_area, rb_node);
4459 }
4460
4461 /**
4462 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
4463 * @addr: target address
4464 *
4465 * Returns: vmap_area if it is found. If there is no such area
4466 * the first highest(reverse order) vmap_area is returned
4467 * i.e. va->va_start < addr && va->va_end < addr or NULL
4468 * if there are no any areas before @addr.
4469 */
4470 static struct vmap_area *
4471 pvm_find_va_enclose_addr(unsigned long addr)
4472 {
4473 struct vmap_area *va, *tmp;
4474 struct rb_node *n;
4475
4476 n = free_vmap_area_root.rb_node;
4477 va = NULL;
4478
4479 while (n) {
4480 tmp = rb_entry(n, struct vmap_area, rb_node);
4481 if (tmp->va_start <= addr) {
4482 va = tmp;
4483 if (tmp->va_end >= addr)
4484 break;
4485
4486 n = n->rb_right;
4487 } else {
4488 n = n->rb_left;
4489 }
4490 }
4491
4492 return va;
4493 }
4494
4495 /**
4496 * pvm_determine_end_from_reverse - find the highest aligned address
4497 * of free block below VMALLOC_END
4498 * @va:
4499 * in - the VA we start the search(reverse order);
4500 * out - the VA with the highest aligned end address.
4501 * @align: alignment for required highest address
4502 *
4503 * Returns: determined end address within vmap_area
4504 */
4505 static unsigned long
4506 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
4507 {
4508 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4509 unsigned long addr;
4510
4511 if (likely(*va)) {
4512 list_for_each_entry_from_reverse((*va),
4513 &free_vmap_area_list, list) {
4514 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
4515 if ((*va)->va_start < addr)
4516 return addr;
4517 }
4518 }
4519
4520 return 0;
4521 }
4522
4523 /**
4524 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4525 * @offsets: array containing offset of each area
4526 * @sizes: array containing size of each area
4527 * @nr_vms: the number of areas to allocate
4528 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4529 *
4530 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4531 * vm_structs on success, %NULL on failure
4532 *
4533 * Percpu allocator wants to use congruent vm areas so that it can
4534 * maintain the offsets among percpu areas. This function allocates
4535 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4536 * be scattered pretty far, distance between two areas easily going up
4537 * to gigabytes. To avoid interacting with regular vmallocs, these
4538 * areas are allocated from top.
4539 *
4540 * Despite its complicated look, this allocator is rather simple. It
4541 * does everything top-down and scans free blocks from the end looking
4542 * for matching base. While scanning, if any of the areas do not fit the
4543 * base address is pulled down to fit the area. Scanning is repeated till
4544 * all the areas fit and then all necessary data structures are inserted
4545 * and the result is returned.
4546 */
4547 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4548 const size_t *sizes, int nr_vms,
4549 size_t align)
4550 {
4551 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4552 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4553 struct vmap_area **vas, *va;
4554 struct vm_struct **vms;
4555 int area, area2, last_area, term_area;
4556 unsigned long base, start, size, end, last_end, orig_start, orig_end;
4557 bool purged = false;
4558
4559 /* verify parameters and allocate data structures */
4560 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4561 for (last_area = 0, area = 0; area < nr_vms; area++) {
4562 start = offsets[area];
4563 end = start + sizes[area];
4564
4565 /* is everything aligned properly? */
4566 BUG_ON(!IS_ALIGNED(offsets[area], align));
4567 BUG_ON(!IS_ALIGNED(sizes[area], align));
4568
4569 /* detect the area with the highest address */
4570 if (start > offsets[last_area])
4571 last_area = area;
4572
4573 for (area2 = area + 1; area2 < nr_vms; area2++) {
4574 unsigned long start2 = offsets[area2];
4575 unsigned long end2 = start2 + sizes[area2];
4576
4577 BUG_ON(start2 < end && start < end2);
4578 }
4579 }
4580 last_end = offsets[last_area] + sizes[last_area];
4581
4582 if (vmalloc_end - vmalloc_start < last_end) {
4583 WARN_ON(true);
4584 return NULL;
4585 }
4586
4587 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4588 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4589 if (!vas || !vms)
4590 goto err_free2;
4591
4592 for (area = 0; area < nr_vms; area++) {
4593 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4594 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4595 if (!vas[area] || !vms[area])
4596 goto err_free;
4597 }
4598 retry:
4599 spin_lock(&free_vmap_area_lock);
4600
4601 /* start scanning - we scan from the top, begin with the last area */
4602 area = term_area = last_area;
4603 start = offsets[area];
4604 end = start + sizes[area];
4605
4606 va = pvm_find_va_enclose_addr(vmalloc_end);
4607 base = pvm_determine_end_from_reverse(&va, align) - end;
4608
4609 while (true) {
4610 /*
4611 * base might have underflowed, add last_end before
4612 * comparing.
4613 */
4614 if (base + last_end < vmalloc_start + last_end)
4615 goto overflow;
4616
4617 /*
4618 * Fitting base has not been found.
4619 */
4620 if (va == NULL)
4621 goto overflow;
4622
4623 /*
4624 * If required width exceeds current VA block, move
4625 * base downwards and then recheck.
4626 */
4627 if (base + end > va->va_end) {
4628 base = pvm_determine_end_from_reverse(&va, align) - end;
4629 term_area = area;
4630 continue;
4631 }
4632
4633 /*
4634 * If this VA does not fit, move base downwards and recheck.
4635 */
4636 if (base + start < va->va_start) {
4637 va = node_to_va(rb_prev(&va->rb_node));
4638 base = pvm_determine_end_from_reverse(&va, align) - end;
4639 term_area = area;
4640 continue;
4641 }
4642
4643 /*
4644 * This area fits, move on to the previous one. If
4645 * the previous one is the terminal one, we're done.
4646 */
4647 area = (area + nr_vms - 1) % nr_vms;
4648 if (area == term_area)
4649 break;
4650
4651 start = offsets[area];
4652 end = start + sizes[area];
4653 va = pvm_find_va_enclose_addr(base + end);
4654 }
4655
4656 /* we've found a fitting base, insert all va's */
4657 for (area = 0; area < nr_vms; area++) {
4658 int ret;
4659
4660 start = base + offsets[area];
4661 size = sizes[area];
4662
4663 va = pvm_find_va_enclose_addr(start);
4664 if (WARN_ON_ONCE(va == NULL))
4665 /* It is a BUG(), but trigger recovery instead. */
4666 goto recovery;
4667
4668 ret = va_clip(&free_vmap_area_root,
4669 &free_vmap_area_list, va, start, size);
4670 if (WARN_ON_ONCE(unlikely(ret)))
4671 /* It is a BUG(), but trigger recovery instead. */
4672 goto recovery;
4673
4674 /* Allocated area. */
4675 va = vas[area];
4676 va->va_start = start;
4677 va->va_end = start + size;
4678 }
4679
4680 spin_unlock(&free_vmap_area_lock);
4681
4682 /* populate the kasan shadow space */
4683 for (area = 0; area < nr_vms; area++) {
4684 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4685 goto err_free_shadow;
4686 }
4687
4688 /* insert all vm's */
4689 for (area = 0; area < nr_vms; area++) {
4690 struct vmap_node *vn = addr_to_node(vas[area]->va_start);
4691
4692 spin_lock(&vn->busy.lock);
4693 insert_vmap_area(vas[area], &vn->busy.root, &vn->busy.head);
4694 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
4695 pcpu_get_vm_areas);
4696 spin_unlock(&vn->busy.lock);
4697 }
4698
4699 /*
4700 * Mark allocated areas as accessible. Do it now as a best-effort
4701 * approach, as they can be mapped outside of vmalloc code.
4702 * With hardware tag-based KASAN, marking is skipped for
4703 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4704 */
4705 for (area = 0; area < nr_vms; area++)
4706 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4707 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4708
4709 kfree(vas);
4710 return vms;
4711
4712 recovery:
4713 /*
4714 * Remove previously allocated areas. There is no
4715 * need in removing these areas from the busy tree,
4716 * because they are inserted only on the final step
4717 * and when pcpu_get_vm_areas() is success.
4718 */
4719 while (area--) {
4720 orig_start = vas[area]->va_start;
4721 orig_end = vas[area]->va_end;
4722 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4723 &free_vmap_area_list);
4724 if (va)
4725 kasan_release_vmalloc(orig_start, orig_end,
4726 va->va_start, va->va_end);
4727 vas[area] = NULL;
4728 }
4729
4730 overflow:
4731 spin_unlock(&free_vmap_area_lock);
4732 if (!purged) {
4733 reclaim_and_purge_vmap_areas();
4734 purged = true;
4735
4736 /* Before "retry", check if we recover. */
4737 for (area = 0; area < nr_vms; area++) {
4738 if (vas[area])
4739 continue;
4740
4741 vas[area] = kmem_cache_zalloc(
4742 vmap_area_cachep, GFP_KERNEL);
4743 if (!vas[area])
4744 goto err_free;
4745 }
4746
4747 goto retry;
4748 }
4749
4750 err_free:
4751 for (area = 0; area < nr_vms; area++) {
4752 if (vas[area])
4753 kmem_cache_free(vmap_area_cachep, vas[area]);
4754
4755 kfree(vms[area]);
4756 }
4757 err_free2:
4758 kfree(vas);
4759 kfree(vms);
4760 return NULL;
4761
4762 err_free_shadow:
4763 spin_lock(&free_vmap_area_lock);
4764 /*
4765 * We release all the vmalloc shadows, even the ones for regions that
4766 * hadn't been successfully added. This relies on kasan_release_vmalloc
4767 * being able to tolerate this case.
4768 */
4769 for (area = 0; area < nr_vms; area++) {
4770 orig_start = vas[area]->va_start;
4771 orig_end = vas[area]->va_end;
4772 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4773 &free_vmap_area_list);
4774 if (va)
4775 kasan_release_vmalloc(orig_start, orig_end,
4776 va->va_start, va->va_end);
4777 vas[area] = NULL;
4778 kfree(vms[area]);
4779 }
4780 spin_unlock(&free_vmap_area_lock);
4781 kfree(vas);
4782 kfree(vms);
4783 return NULL;
4784 }
4785
4786 /**
4787 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4788 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4789 * @nr_vms: the number of allocated areas
4790 *
4791 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4792 */
4793 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4794 {
4795 int i;
4796
4797 for (i = 0; i < nr_vms; i++)
4798 free_vm_area(vms[i]);
4799 kfree(vms);
4800 }
4801 #endif /* CONFIG_SMP */
4802
4803 #ifdef CONFIG_PRINTK
4804 bool vmalloc_dump_obj(void *object)
4805 {
4806 const void *caller;
4807 struct vm_struct *vm;
4808 struct vmap_area *va;
4809 struct vmap_node *vn;
4810 unsigned long addr;
4811 unsigned int nr_pages;
4812
4813 addr = PAGE_ALIGN((unsigned long) object);
4814 vn = addr_to_node(addr);
4815
4816 if (!spin_trylock(&vn->busy.lock))
4817 return false;
4818
4819 va = __find_vmap_area(addr, &vn->busy.root);
4820 if (!va || !va->vm) {
4821 spin_unlock(&vn->busy.lock);
4822 return false;
4823 }
4824
4825 vm = va->vm;
4826 addr = (unsigned long) vm->addr;
4827 caller = vm->caller;
4828 nr_pages = vm->nr_pages;
4829 spin_unlock(&vn->busy.lock);
4830
4831 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4832 nr_pages, addr, caller);
4833
4834 return true;
4835 }
4836 #endif
4837
4838 #ifdef CONFIG_PROC_FS
4839 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4840 {
4841 if (IS_ENABLED(CONFIG_NUMA)) {
4842 unsigned int nr, *counters = m->private;
4843 unsigned int step = 1U << vm_area_page_order(v);
4844
4845 if (!counters)
4846 return;
4847
4848 if (v->flags & VM_UNINITIALIZED)
4849 return;
4850 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4851 smp_rmb();
4852
4853 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4854
4855 for (nr = 0; nr < v->nr_pages; nr += step)
4856 counters[page_to_nid(v->pages[nr])] += step;
4857 for_each_node_state(nr, N_HIGH_MEMORY)
4858 if (counters[nr])
4859 seq_printf(m, " N%u=%u", nr, counters[nr]);
4860 }
4861 }
4862
4863 static void show_purge_info(struct seq_file *m)
4864 {
4865 struct vmap_node *vn;
4866 struct vmap_area *va;
4867 int i;
4868
4869 for (i = 0; i < nr_vmap_nodes; i++) {
4870 vn = &vmap_nodes[i];
4871
4872 spin_lock(&vn->lazy.lock);
4873 list_for_each_entry(va, &vn->lazy.head, list) {
4874 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4875 (void *)va->va_start, (void *)va->va_end,
4876 va->va_end - va->va_start);
4877 }
4878 spin_unlock(&vn->lazy.lock);
4879 }
4880 }
4881
4882 static int vmalloc_info_show(struct seq_file *m, void *p)
4883 {
4884 struct vmap_node *vn;
4885 struct vmap_area *va;
4886 struct vm_struct *v;
4887 int i;
4888
4889 for (i = 0; i < nr_vmap_nodes; i++) {
4890 vn = &vmap_nodes[i];
4891
4892 spin_lock(&vn->busy.lock);
4893 list_for_each_entry(va, &vn->busy.head, list) {
4894 if (!va->vm) {
4895 if (va->flags & VMAP_RAM)
4896 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4897 (void *)va->va_start, (void *)va->va_end,
4898 va->va_end - va->va_start);
4899
4900 continue;
4901 }
4902
4903 v = va->vm;
4904
4905 seq_printf(m, "0x%pK-0x%pK %7ld",
4906 v->addr, v->addr + v->size, v->size);
4907
4908 if (v->caller)
4909 seq_printf(m, " %pS", v->caller);
4910
4911 if (v->nr_pages)
4912 seq_printf(m, " pages=%d", v->nr_pages);
4913
4914 if (v->phys_addr)
4915 seq_printf(m, " phys=%pa", &v->phys_addr);
4916
4917 if (v->flags & VM_IOREMAP)
4918 seq_puts(m, " ioremap");
4919
4920 if (v->flags & VM_SPARSE)
4921 seq_puts(m, " sparse");
4922
4923 if (v->flags & VM_ALLOC)
4924 seq_puts(m, " vmalloc");
4925
4926 if (v->flags & VM_MAP)
4927 seq_puts(m, " vmap");
4928
4929 if (v->flags & VM_USERMAP)
4930 seq_puts(m, " user");
4931
4932 if (v->flags & VM_DMA_COHERENT)
4933 seq_puts(m, " dma-coherent");
4934
4935 if (is_vmalloc_addr(v->pages))
4936 seq_puts(m, " vpages");
4937
4938 show_numa_info(m, v);
4939 seq_putc(m, '\n');
4940 }
4941 spin_unlock(&vn->busy.lock);
4942 }
4943
4944 /*
4945 * As a final step, dump "unpurged" areas.
4946 */
4947 show_purge_info(m);
4948 return 0;
4949 }
4950
4951 static int __init proc_vmalloc_init(void)
4952 {
4953 void *priv_data = NULL;
4954
4955 if (IS_ENABLED(CONFIG_NUMA))
4956 priv_data = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
4957
4958 proc_create_single_data("vmallocinfo",
4959 0400, NULL, vmalloc_info_show, priv_data);
4960
4961 return 0;
4962 }
4963 module_init(proc_vmalloc_init);
4964
4965 #endif
4966
4967 static void __init vmap_init_free_space(void)
4968 {
4969 unsigned long vmap_start = 1;
4970 const unsigned long vmap_end = ULONG_MAX;
4971 struct vmap_area *free;
4972 struct vm_struct *busy;
4973
4974 /*
4975 * B F B B B F
4976 * -|-----|.....|-----|-----|-----|.....|-
4977 * | The KVA space |
4978 * |<--------------------------------->|
4979 */
4980 for (busy = vmlist; busy; busy = busy->next) {
4981 if ((unsigned long) busy->addr - vmap_start > 0) {
4982 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4983 if (!WARN_ON_ONCE(!free)) {
4984 free->va_start = vmap_start;
4985 free->va_end = (unsigned long) busy->addr;
4986
4987 insert_vmap_area_augment(free, NULL,
4988 &free_vmap_area_root,
4989 &free_vmap_area_list);
4990 }
4991 }
4992
4993 vmap_start = (unsigned long) busy->addr + busy->size;
4994 }
4995
4996 if (vmap_end - vmap_start > 0) {
4997 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4998 if (!WARN_ON_ONCE(!free)) {
4999 free->va_start = vmap_start;
5000 free->va_end = vmap_end;
5001
5002 insert_vmap_area_augment(free, NULL,
5003 &free_vmap_area_root,
5004 &free_vmap_area_list);
5005 }
5006 }
5007 }
5008
5009 static void vmap_init_nodes(void)
5010 {
5011 struct vmap_node *vn;
5012 int i, n;
5013
5014 #if BITS_PER_LONG == 64
5015 /*
5016 * A high threshold of max nodes is fixed and bound to 128,
5017 * thus a scale factor is 1 for systems where number of cores
5018 * are less or equal to specified threshold.
5019 *
5020 * As for NUMA-aware notes. For bigger systems, for example
5021 * NUMA with multi-sockets, where we can end-up with thousands
5022 * of cores in total, a "sub-numa-clustering" should be added.
5023 *
5024 * In this case a NUMA domain is considered as a single entity
5025 * with dedicated sub-nodes in it which describe one group or
5026 * set of cores. Therefore a per-domain purging is supposed to
5027 * be added as well as a per-domain balancing.
5028 */
5029 n = clamp_t(unsigned int, num_possible_cpus(), 1, 128);
5030
5031 if (n > 1) {
5032 vn = kmalloc_array(n, sizeof(*vn), GFP_NOWAIT | __GFP_NOWARN);
5033 if (vn) {
5034 /* Node partition is 16 pages. */
5035 vmap_zone_size = (1 << 4) * PAGE_SIZE;
5036 nr_vmap_nodes = n;
5037 vmap_nodes = vn;
5038 } else {
5039 pr_err("Failed to allocate an array. Disable a node layer\n");
5040 }
5041 }
5042 #endif
5043
5044 for (n = 0; n < nr_vmap_nodes; n++) {
5045 vn = &vmap_nodes[n];
5046 vn->busy.root = RB_ROOT;
5047 INIT_LIST_HEAD(&vn->busy.head);
5048 spin_lock_init(&vn->busy.lock);
5049
5050 vn->lazy.root = RB_ROOT;
5051 INIT_LIST_HEAD(&vn->lazy.head);
5052 spin_lock_init(&vn->lazy.lock);
5053
5054 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
5055 INIT_LIST_HEAD(&vn->pool[i].head);
5056 WRITE_ONCE(vn->pool[i].len, 0);
5057 }
5058
5059 spin_lock_init(&vn->pool_lock);
5060 }
5061 }
5062
5063 static unsigned long
5064 vmap_node_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5065 {
5066 unsigned long count;
5067 struct vmap_node *vn;
5068 int i, j;
5069
5070 for (count = 0, i = 0; i < nr_vmap_nodes; i++) {
5071 vn = &vmap_nodes[i];
5072
5073 for (j = 0; j < MAX_VA_SIZE_PAGES; j++)
5074 count += READ_ONCE(vn->pool[j].len);
5075 }
5076
5077 return count ? count : SHRINK_EMPTY;
5078 }
5079
5080 static unsigned long
5081 vmap_node_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5082 {
5083 int i;
5084
5085 for (i = 0; i < nr_vmap_nodes; i++)
5086 decay_va_pool_node(&vmap_nodes[i], true);
5087
5088 return SHRINK_STOP;
5089 }
5090
5091 void __init vmalloc_init(void)
5092 {
5093 struct shrinker *vmap_node_shrinker;
5094 struct vmap_area *va;
5095 struct vmap_node *vn;
5096 struct vm_struct *tmp;
5097 int i;
5098
5099 /*
5100 * Create the cache for vmap_area objects.
5101 */
5102 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
5103
5104 for_each_possible_cpu(i) {
5105 struct vmap_block_queue *vbq;
5106 struct vfree_deferred *p;
5107
5108 vbq = &per_cpu(vmap_block_queue, i);
5109 spin_lock_init(&vbq->lock);
5110 INIT_LIST_HEAD(&vbq->free);
5111 p = &per_cpu(vfree_deferred, i);
5112 init_llist_head(&p->list);
5113 INIT_WORK(&p->wq, delayed_vfree_work);
5114 xa_init(&vbq->vmap_blocks);
5115 }
5116
5117 /*
5118 * Setup nodes before importing vmlist.
5119 */
5120 vmap_init_nodes();
5121
5122 /* Import existing vmlist entries. */
5123 for (tmp = vmlist; tmp; tmp = tmp->next) {
5124 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5125 if (WARN_ON_ONCE(!va))
5126 continue;
5127
5128 va->va_start = (unsigned long)tmp->addr;
5129 va->va_end = va->va_start + tmp->size;
5130 va->vm = tmp;
5131
5132 vn = addr_to_node(va->va_start);
5133 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
5134 }
5135
5136 /*
5137 * Now we can initialize a free vmap space.
5138 */
5139 vmap_init_free_space();
5140 vmap_initialized = true;
5141
5142 vmap_node_shrinker = shrinker_alloc(0, "vmap-node");
5143 if (!vmap_node_shrinker) {
5144 pr_err("Failed to allocate vmap-node shrinker!\n");
5145 return;
5146 }
5147
5148 vmap_node_shrinker->count_objects = vmap_node_shrink_count;
5149 vmap_node_shrinker->scan_objects = vmap_node_shrink_scan;
5150 shrinker_register(vmap_node_shrinker);
5151 }
5152
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.