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
set_nohugeiomap(char * str)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
set_nohugevmalloc(char * str)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
is_vmalloc_addr(const void * x)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 ***/
vmap_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift,pgtbl_mod_mask * mask)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
vmap_try_huge_pmd(pmd_t * pmd,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift)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
vmap_pmd_range(pud_t * pud,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift,pgtbl_mod_mask * mask)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
vmap_try_huge_pud(pud_t * pud,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift)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
vmap_pud_range(p4d_t * p4d,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift,pgtbl_mod_mask * mask)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
vmap_try_huge_p4d(p4d_t * p4d,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift)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
vmap_p4d_range(pgd_t * pgd,unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift,pgtbl_mod_mask * mask)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
vmap_range_noflush(unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot,unsigned int max_page_shift)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
vmap_page_range(unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot)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
ioremap_page_range(unsigned long addr,unsigned long end,phys_addr_t phys_addr,pgprot_t prot)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
vunmap_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,pgtbl_mod_mask * mask)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
vunmap_pmd_range(pud_t * pud,unsigned long addr,unsigned long end,pgtbl_mod_mask * mask)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
vunmap_pud_range(p4d_t * p4d,unsigned long addr,unsigned long end,pgtbl_mod_mask * mask)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
vunmap_p4d_range(pgd_t * pgd,unsigned long addr,unsigned long end,pgtbl_mod_mask * mask)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 */
__vunmap_range_noflush(unsigned long start,unsigned long end)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
vunmap_range_noflush(unsigned long start,unsigned long end)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 */
vunmap_range(unsigned long addr,unsigned long end)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
vmap_pages_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,int * nr,pgtbl_mod_mask * mask)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
vmap_pages_pmd_range(pud_t * pud,unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,int * nr,pgtbl_mod_mask * mask)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
vmap_pages_pud_range(p4d_t * p4d,unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,int * nr,pgtbl_mod_mask * mask)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
vmap_pages_p4d_range(pgd_t * pgd,unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,int * nr,pgtbl_mod_mask * mask)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
vmap_small_pages_range_noflush(unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages)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 */
__vmap_pages_range_noflush(unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,unsigned int page_shift)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
vmap_pages_range_noflush(unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,unsigned int page_shift)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 */
vmap_pages_range(unsigned long addr,unsigned long end,pgprot_t prot,struct page ** pages,unsigned int page_shift)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
check_sparse_vm_area(struct vm_struct * area,unsigned long start,unsigned long end)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 */
vm_area_map_pages(struct vm_struct * area,unsigned long start,unsigned long end,struct page ** pages)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 */
vm_area_unmap_pages(struct vm_struct * area,unsigned long start,unsigned long end)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
is_vmalloc_or_module_addr(const void * x)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 */
vmalloc_to_page(const void * vmalloc_addr)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 */
vmalloc_to_pfn(const void * vmalloc_addr)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
addr_to_node_id(unsigned long addr)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 *
addr_to_node(unsigned long addr)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 *
id_to_node(unsigned int id)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
encode_vn_id(unsigned int node_id)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
decode_vn_id(unsigned int val)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
is_vn_id_valid(unsigned int node_id)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
va_size(struct vmap_area * va)971 va_size(struct vmap_area *va)
972 {
973 return (va->va_end - va->va_start);
974 }
975
976 static __always_inline unsigned long
get_subtree_max_size(struct rb_node * node)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
vmalloc_nr_pages(void)995 unsigned long vmalloc_nr_pages(void)
996 {
997 return atomic_long_read(&nr_vmalloc_pages);
998 }
999
__find_vmap_area(unsigned long addr,struct rb_root * root)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 *
__find_vmap_area_exceed_addr(unsigned long addr,struct rb_root * root)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 *
find_vmap_area_exceed_addr_lock(unsigned long addr,struct vmap_area ** va)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 **
find_va_links(struct vmap_area * va,struct rb_root * root,struct rb_node * from,struct rb_node ** parent)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 *
get_va_next_sibling(struct rb_node * parent,struct rb_node ** link)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
__link_va(struct vmap_area * va,struct rb_root * root,struct rb_node * parent,struct rb_node ** link,struct list_head * head,bool augment)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
link_va(struct vmap_area * va,struct rb_root * root,struct rb_node * parent,struct rb_node ** link,struct list_head * head)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
link_va_augment(struct vmap_area * va,struct rb_root * root,struct rb_node * parent,struct rb_node ** link,struct list_head * head)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
__unlink_va(struct vmap_area * va,struct rb_root * root,bool augment)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
unlink_va(struct vmap_area * va,struct rb_root * root)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
unlink_va_augment(struct vmap_area * va,struct rb_root * root)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
compute_subtree_max_size(struct vmap_area * va)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
augment_tree_propagate_check(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
augment_tree_propagate_from(struct vmap_area * va)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
insert_vmap_area(struct vmap_area * va,struct rb_root * root,struct list_head * head)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
insert_vmap_area_augment(struct vmap_area * va,struct rb_node * from,struct rb_root * root,struct list_head * head)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 *
__merge_or_add_vmap_area(struct vmap_area * va,struct rb_root * root,struct list_head * head,bool augment)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 *
merge_or_add_vmap_area(struct vmap_area * va,struct rb_root * root,struct list_head * head)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 *
merge_or_add_vmap_area_augment(struct vmap_area * va,struct rb_root * root,struct list_head * head)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
is_within_this_va(struct vmap_area * va,unsigned long size,unsigned long align,unsigned long vstart)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 *
find_vmap_lowest_match(struct rb_root * root,unsigned long size,unsigned long align,unsigned long vstart,bool adjust_search_size)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 *
find_vmap_lowest_linear_match(struct list_head * head,unsigned long size,unsigned long align,unsigned long vstart)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
find_vmap_lowest_match_check(struct rb_root * root,struct list_head * head,unsigned long size,unsigned long align)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
classify_va_fit_type(struct vmap_area * va,unsigned long nva_start_addr,unsigned long size)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
va_clip(struct rb_root * root,struct list_head * head,struct vmap_area * va,unsigned long nva_start_addr,unsigned long size)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
va_alloc(struct vmap_area * va,struct rb_root * root,struct list_head * head,unsigned long size,unsigned long align,unsigned long vstart,unsigned long vend)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
__alloc_vmap_area(struct rb_root * root,struct list_head * head,unsigned long size,unsigned long align,unsigned long vstart,unsigned long vend)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 */
free_vmap_area(struct vmap_area * va)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
preload_this_cpu_lock(spinlock_t * lock,gfp_t gfp_mask,int node)1817 preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1818 {
1819 struct vmap_area *va = NULL;
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 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1836 kmem_cache_free(vmap_area_cachep, va);
1837 }
1838
1839 static struct vmap_pool *
size_to_va_pool(struct vmap_node * vn,unsigned long size)1840 size_to_va_pool(struct vmap_node *vn, unsigned long size)
1841 {
1842 unsigned int idx = (size - 1) / PAGE_SIZE;
1843
1844 if (idx < MAX_VA_SIZE_PAGES)
1845 return &vn->pool[idx];
1846
1847 return NULL;
1848 }
1849
1850 static bool
node_pool_add_va(struct vmap_node * n,struct vmap_area * va)1851 node_pool_add_va(struct vmap_node *n, struct vmap_area *va)
1852 {
1853 struct vmap_pool *vp;
1854
1855 vp = size_to_va_pool(n, va_size(va));
1856 if (!vp)
1857 return false;
1858
1859 spin_lock(&n->pool_lock);
1860 list_add(&va->list, &vp->head);
1861 WRITE_ONCE(vp->len, vp->len + 1);
1862 spin_unlock(&n->pool_lock);
1863
1864 return true;
1865 }
1866
1867 static struct vmap_area *
node_pool_del_va(struct vmap_node * vn,unsigned long size,unsigned long align,unsigned long vstart,unsigned long vend)1868 node_pool_del_va(struct vmap_node *vn, unsigned long size,
1869 unsigned long align, unsigned long vstart,
1870 unsigned long vend)
1871 {
1872 struct vmap_area *va = NULL;
1873 struct vmap_pool *vp;
1874 int err = 0;
1875
1876 vp = size_to_va_pool(vn, size);
1877 if (!vp || list_empty(&vp->head))
1878 return NULL;
1879
1880 spin_lock(&vn->pool_lock);
1881 if (!list_empty(&vp->head)) {
1882 va = list_first_entry(&vp->head, struct vmap_area, list);
1883
1884 if (IS_ALIGNED(va->va_start, align)) {
1885 /*
1886 * Do some sanity check and emit a warning
1887 * if one of below checks detects an error.
1888 */
1889 err |= (va_size(va) != size);
1890 err |= (va->va_start < vstart);
1891 err |= (va->va_end > vend);
1892
1893 if (!WARN_ON_ONCE(err)) {
1894 list_del_init(&va->list);
1895 WRITE_ONCE(vp->len, vp->len - 1);
1896 } else {
1897 va = NULL;
1898 }
1899 } else {
1900 list_move_tail(&va->list, &vp->head);
1901 va = NULL;
1902 }
1903 }
1904 spin_unlock(&vn->pool_lock);
1905
1906 return va;
1907 }
1908
1909 static struct vmap_area *
node_alloc(unsigned long size,unsigned long align,unsigned long vstart,unsigned long vend,unsigned long * addr,unsigned int * vn_id)1910 node_alloc(unsigned long size, unsigned long align,
1911 unsigned long vstart, unsigned long vend,
1912 unsigned long *addr, unsigned int *vn_id)
1913 {
1914 struct vmap_area *va;
1915
1916 *vn_id = 0;
1917 *addr = vend;
1918
1919 /*
1920 * Fallback to a global heap if not vmalloc or there
1921 * is only one node.
1922 */
1923 if (vstart != VMALLOC_START || vend != VMALLOC_END ||
1924 nr_vmap_nodes == 1)
1925 return NULL;
1926
1927 *vn_id = raw_smp_processor_id() % nr_vmap_nodes;
1928 va = node_pool_del_va(id_to_node(*vn_id), size, align, vstart, vend);
1929 *vn_id = encode_vn_id(*vn_id);
1930
1931 if (va)
1932 *addr = va->va_start;
1933
1934 return va;
1935 }
1936
setup_vmalloc_vm(struct vm_struct * vm,struct vmap_area * va,unsigned long flags,const void * caller)1937 static inline void setup_vmalloc_vm(struct vm_struct *vm,
1938 struct vmap_area *va, unsigned long flags, const void *caller)
1939 {
1940 vm->flags = flags;
1941 vm->addr = (void *)va->va_start;
1942 vm->size = va->va_end - va->va_start;
1943 vm->caller = caller;
1944 va->vm = vm;
1945 }
1946
1947 /*
1948 * Allocate a region of KVA of the specified size and alignment, within the
1949 * vstart and vend. If vm is passed in, the two will also be bound.
1950 */
alloc_vmap_area(unsigned long size,unsigned long align,unsigned long vstart,unsigned long vend,int node,gfp_t gfp_mask,unsigned long va_flags,struct vm_struct * vm)1951 static struct vmap_area *alloc_vmap_area(unsigned long size,
1952 unsigned long align,
1953 unsigned long vstart, unsigned long vend,
1954 int node, gfp_t gfp_mask,
1955 unsigned long va_flags, struct vm_struct *vm)
1956 {
1957 struct vmap_node *vn;
1958 struct vmap_area *va;
1959 unsigned long freed;
1960 unsigned long addr;
1961 unsigned int vn_id;
1962 int purged = 0;
1963 int ret;
1964
1965 if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1966 return ERR_PTR(-EINVAL);
1967
1968 if (unlikely(!vmap_initialized))
1969 return ERR_PTR(-EBUSY);
1970
1971 might_sleep();
1972
1973 /*
1974 * If a VA is obtained from a global heap(if it fails here)
1975 * it is anyway marked with this "vn_id" so it is returned
1976 * to this pool's node later. Such way gives a possibility
1977 * to populate pools based on users demand.
1978 *
1979 * On success a ready to go VA is returned.
1980 */
1981 va = node_alloc(size, align, vstart, vend, &addr, &vn_id);
1982 if (!va) {
1983 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1984
1985 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1986 if (unlikely(!va))
1987 return ERR_PTR(-ENOMEM);
1988
1989 /*
1990 * Only scan the relevant parts containing pointers to other objects
1991 * to avoid false negatives.
1992 */
1993 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1994 }
1995
1996 retry:
1997 if (addr == vend) {
1998 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1999 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
2000 size, align, vstart, vend);
2001 spin_unlock(&free_vmap_area_lock);
2002 }
2003
2004 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
2005
2006 /*
2007 * If an allocation fails, the "vend" address is
2008 * returned. Therefore trigger the overflow path.
2009 */
2010 if (unlikely(addr == vend))
2011 goto overflow;
2012
2013 va->va_start = addr;
2014 va->va_end = addr + size;
2015 va->vm = NULL;
2016 va->flags = (va_flags | vn_id);
2017
2018 if (vm) {
2019 vm->addr = (void *)va->va_start;
2020 vm->size = va->va_end - va->va_start;
2021 va->vm = vm;
2022 }
2023
2024 vn = addr_to_node(va->va_start);
2025
2026 spin_lock(&vn->busy.lock);
2027 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
2028 spin_unlock(&vn->busy.lock);
2029
2030 BUG_ON(!IS_ALIGNED(va->va_start, align));
2031 BUG_ON(va->va_start < vstart);
2032 BUG_ON(va->va_end > vend);
2033
2034 ret = kasan_populate_vmalloc(addr, size);
2035 if (ret) {
2036 free_vmap_area(va);
2037 return ERR_PTR(ret);
2038 }
2039
2040 return va;
2041
2042 overflow:
2043 if (!purged) {
2044 reclaim_and_purge_vmap_areas();
2045 purged = 1;
2046 goto retry;
2047 }
2048
2049 freed = 0;
2050 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
2051
2052 if (freed > 0) {
2053 purged = 0;
2054 goto retry;
2055 }
2056
2057 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
2058 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
2059 size);
2060
2061 kmem_cache_free(vmap_area_cachep, va);
2062 return ERR_PTR(-EBUSY);
2063 }
2064
register_vmap_purge_notifier(struct notifier_block * nb)2065 int register_vmap_purge_notifier(struct notifier_block *nb)
2066 {
2067 return blocking_notifier_chain_register(&vmap_notify_list, nb);
2068 }
2069 EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
2070
unregister_vmap_purge_notifier(struct notifier_block * nb)2071 int unregister_vmap_purge_notifier(struct notifier_block *nb)
2072 {
2073 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
2074 }
2075 EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
2076
2077 /*
2078 * lazy_max_pages is the maximum amount of virtual address space we gather up
2079 * before attempting to purge with a TLB flush.
2080 *
2081 * There is a tradeoff here: a larger number will cover more kernel page tables
2082 * and take slightly longer to purge, but it will linearly reduce the number of
2083 * global TLB flushes that must be performed. It would seem natural to scale
2084 * this number up linearly with the number of CPUs (because vmapping activity
2085 * could also scale linearly with the number of CPUs), however it is likely
2086 * that in practice, workloads might be constrained in other ways that mean
2087 * vmap activity will not scale linearly with CPUs. Also, I want to be
2088 * conservative and not introduce a big latency on huge systems, so go with
2089 * a less aggressive log scale. It will still be an improvement over the old
2090 * code, and it will be simple to change the scale factor if we find that it
2091 * becomes a problem on bigger systems.
2092 */
lazy_max_pages(void)2093 static unsigned long lazy_max_pages(void)
2094 {
2095 unsigned int log;
2096
2097 log = fls(num_online_cpus());
2098
2099 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
2100 }
2101
2102 static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
2103
2104 /*
2105 * Serialize vmap purging. There is no actual critical section protected
2106 * by this lock, but we want to avoid concurrent calls for performance
2107 * reasons and to make the pcpu_get_vm_areas more deterministic.
2108 */
2109 static DEFINE_MUTEX(vmap_purge_lock);
2110
2111 /* for per-CPU blocks */
2112 static void purge_fragmented_blocks_allcpus(void);
2113 static cpumask_t purge_nodes;
2114
2115 static void
reclaim_list_global(struct list_head * head)2116 reclaim_list_global(struct list_head *head)
2117 {
2118 struct vmap_area *va, *n;
2119
2120 if (list_empty(head))
2121 return;
2122
2123 spin_lock(&free_vmap_area_lock);
2124 list_for_each_entry_safe(va, n, head, list)
2125 merge_or_add_vmap_area_augment(va,
2126 &free_vmap_area_root, &free_vmap_area_list);
2127 spin_unlock(&free_vmap_area_lock);
2128 }
2129
2130 static void
decay_va_pool_node(struct vmap_node * vn,bool full_decay)2131 decay_va_pool_node(struct vmap_node *vn, bool full_decay)
2132 {
2133 struct vmap_area *va, *nva;
2134 struct list_head decay_list;
2135 struct rb_root decay_root;
2136 unsigned long n_decay;
2137 int i;
2138
2139 decay_root = RB_ROOT;
2140 INIT_LIST_HEAD(&decay_list);
2141
2142 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
2143 struct list_head tmp_list;
2144
2145 if (list_empty(&vn->pool[i].head))
2146 continue;
2147
2148 INIT_LIST_HEAD(&tmp_list);
2149
2150 /* Detach the pool, so no-one can access it. */
2151 spin_lock(&vn->pool_lock);
2152 list_replace_init(&vn->pool[i].head, &tmp_list);
2153 spin_unlock(&vn->pool_lock);
2154
2155 if (full_decay)
2156 WRITE_ONCE(vn->pool[i].len, 0);
2157
2158 /* Decay a pool by ~25% out of left objects. */
2159 n_decay = vn->pool[i].len >> 2;
2160
2161 list_for_each_entry_safe(va, nva, &tmp_list, list) {
2162 list_del_init(&va->list);
2163 merge_or_add_vmap_area(va, &decay_root, &decay_list);
2164
2165 if (!full_decay) {
2166 WRITE_ONCE(vn->pool[i].len, vn->pool[i].len - 1);
2167
2168 if (!--n_decay)
2169 break;
2170 }
2171 }
2172
2173 /*
2174 * Attach the pool back if it has been partly decayed.
2175 * Please note, it is supposed that nobody(other contexts)
2176 * can populate the pool therefore a simple list replace
2177 * operation takes place here.
2178 */
2179 if (!full_decay && !list_empty(&tmp_list)) {
2180 spin_lock(&vn->pool_lock);
2181 list_replace_init(&tmp_list, &vn->pool[i].head);
2182 spin_unlock(&vn->pool_lock);
2183 }
2184 }
2185
2186 reclaim_list_global(&decay_list);
2187 }
2188
purge_vmap_node(struct work_struct * work)2189 static void purge_vmap_node(struct work_struct *work)
2190 {
2191 struct vmap_node *vn = container_of(work,
2192 struct vmap_node, purge_work);
2193 struct vmap_area *va, *n_va;
2194 LIST_HEAD(local_list);
2195
2196 vn->nr_purged = 0;
2197
2198 list_for_each_entry_safe(va, n_va, &vn->purge_list, list) {
2199 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
2200 unsigned long orig_start = va->va_start;
2201 unsigned long orig_end = va->va_end;
2202 unsigned int vn_id = decode_vn_id(va->flags);
2203
2204 list_del_init(&va->list);
2205
2206 if (is_vmalloc_or_module_addr((void *)orig_start))
2207 kasan_release_vmalloc(orig_start, orig_end,
2208 va->va_start, va->va_end);
2209
2210 atomic_long_sub(nr, &vmap_lazy_nr);
2211 vn->nr_purged++;
2212
2213 if (is_vn_id_valid(vn_id) && !vn->skip_populate)
2214 if (node_pool_add_va(vn, va))
2215 continue;
2216
2217 /* Go back to global. */
2218 list_add(&va->list, &local_list);
2219 }
2220
2221 reclaim_list_global(&local_list);
2222 }
2223
2224 /*
2225 * Purges all lazily-freed vmap areas.
2226 */
__purge_vmap_area_lazy(unsigned long start,unsigned long end,bool full_pool_decay)2227 static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
2228 bool full_pool_decay)
2229 {
2230 unsigned long nr_purged_areas = 0;
2231 unsigned int nr_purge_helpers;
2232 unsigned int nr_purge_nodes;
2233 struct vmap_node *vn;
2234 int i;
2235
2236 lockdep_assert_held(&vmap_purge_lock);
2237
2238 /*
2239 * Use cpumask to mark which node has to be processed.
2240 */
2241 purge_nodes = CPU_MASK_NONE;
2242
2243 for (i = 0; i < nr_vmap_nodes; i++) {
2244 vn = &vmap_nodes[i];
2245
2246 INIT_LIST_HEAD(&vn->purge_list);
2247 vn->skip_populate = full_pool_decay;
2248 decay_va_pool_node(vn, full_pool_decay);
2249
2250 if (RB_EMPTY_ROOT(&vn->lazy.root))
2251 continue;
2252
2253 spin_lock(&vn->lazy.lock);
2254 WRITE_ONCE(vn->lazy.root.rb_node, NULL);
2255 list_replace_init(&vn->lazy.head, &vn->purge_list);
2256 spin_unlock(&vn->lazy.lock);
2257
2258 start = min(start, list_first_entry(&vn->purge_list,
2259 struct vmap_area, list)->va_start);
2260
2261 end = max(end, list_last_entry(&vn->purge_list,
2262 struct vmap_area, list)->va_end);
2263
2264 cpumask_set_cpu(i, &purge_nodes);
2265 }
2266
2267 nr_purge_nodes = cpumask_weight(&purge_nodes);
2268 if (nr_purge_nodes > 0) {
2269 flush_tlb_kernel_range(start, end);
2270
2271 /* One extra worker is per a lazy_max_pages() full set minus one. */
2272 nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
2273 nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
2274
2275 for_each_cpu(i, &purge_nodes) {
2276 vn = &vmap_nodes[i];
2277
2278 if (nr_purge_helpers > 0) {
2279 INIT_WORK(&vn->purge_work, purge_vmap_node);
2280
2281 if (cpumask_test_cpu(i, cpu_online_mask))
2282 schedule_work_on(i, &vn->purge_work);
2283 else
2284 schedule_work(&vn->purge_work);
2285
2286 nr_purge_helpers--;
2287 } else {
2288 vn->purge_work.func = NULL;
2289 purge_vmap_node(&vn->purge_work);
2290 nr_purged_areas += vn->nr_purged;
2291 }
2292 }
2293
2294 for_each_cpu(i, &purge_nodes) {
2295 vn = &vmap_nodes[i];
2296
2297 if (vn->purge_work.func) {
2298 flush_work(&vn->purge_work);
2299 nr_purged_areas += vn->nr_purged;
2300 }
2301 }
2302 }
2303
2304 trace_purge_vmap_area_lazy(start, end, nr_purged_areas);
2305 return nr_purged_areas > 0;
2306 }
2307
2308 /*
2309 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
2310 */
reclaim_and_purge_vmap_areas(void)2311 static void reclaim_and_purge_vmap_areas(void)
2312
2313 {
2314 mutex_lock(&vmap_purge_lock);
2315 purge_fragmented_blocks_allcpus();
2316 __purge_vmap_area_lazy(ULONG_MAX, 0, true);
2317 mutex_unlock(&vmap_purge_lock);
2318 }
2319
drain_vmap_area_work(struct work_struct * work)2320 static void drain_vmap_area_work(struct work_struct *work)
2321 {
2322 mutex_lock(&vmap_purge_lock);
2323 __purge_vmap_area_lazy(ULONG_MAX, 0, false);
2324 mutex_unlock(&vmap_purge_lock);
2325 }
2326
2327 /*
2328 * Free a vmap area, caller ensuring that the area has been unmapped,
2329 * unlinked and flush_cache_vunmap had been called for the correct
2330 * range previously.
2331 */
free_vmap_area_noflush(struct vmap_area * va)2332 static void free_vmap_area_noflush(struct vmap_area *va)
2333 {
2334 unsigned long nr_lazy_max = lazy_max_pages();
2335 unsigned long va_start = va->va_start;
2336 unsigned int vn_id = decode_vn_id(va->flags);
2337 struct vmap_node *vn;
2338 unsigned long nr_lazy;
2339
2340 if (WARN_ON_ONCE(!list_empty(&va->list)))
2341 return;
2342
2343 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
2344 PAGE_SHIFT, &vmap_lazy_nr);
2345
2346 /*
2347 * If it was request by a certain node we would like to
2348 * return it to that node, i.e. its pool for later reuse.
2349 */
2350 vn = is_vn_id_valid(vn_id) ?
2351 id_to_node(vn_id):addr_to_node(va->va_start);
2352
2353 spin_lock(&vn->lazy.lock);
2354 insert_vmap_area(va, &vn->lazy.root, &vn->lazy.head);
2355 spin_unlock(&vn->lazy.lock);
2356
2357 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
2358
2359 /* After this point, we may free va at any time */
2360 if (unlikely(nr_lazy > nr_lazy_max))
2361 schedule_work(&drain_vmap_work);
2362 }
2363
2364 /*
2365 * Free and unmap a vmap area
2366 */
free_unmap_vmap_area(struct vmap_area * va)2367 static void free_unmap_vmap_area(struct vmap_area *va)
2368 {
2369 flush_cache_vunmap(va->va_start, va->va_end);
2370 vunmap_range_noflush(va->va_start, va->va_end);
2371 if (debug_pagealloc_enabled_static())
2372 flush_tlb_kernel_range(va->va_start, va->va_end);
2373
2374 free_vmap_area_noflush(va);
2375 }
2376
find_vmap_area(unsigned long addr)2377 struct vmap_area *find_vmap_area(unsigned long addr)
2378 {
2379 struct vmap_node *vn;
2380 struct vmap_area *va;
2381 int i, j;
2382
2383 if (unlikely(!vmap_initialized))
2384 return NULL;
2385
2386 /*
2387 * An addr_to_node_id(addr) converts an address to a node index
2388 * where a VA is located. If VA spans several zones and passed
2389 * addr is not the same as va->va_start, what is not common, we
2390 * may need to scan extra nodes. See an example:
2391 *
2392 * <----va---->
2393 * -|-----|-----|-----|-----|-
2394 * 1 2 0 1
2395 *
2396 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
2397 * addr is within 2 or 0 nodes we should do extra work.
2398 */
2399 i = j = addr_to_node_id(addr);
2400 do {
2401 vn = &vmap_nodes[i];
2402
2403 spin_lock(&vn->busy.lock);
2404 va = __find_vmap_area(addr, &vn->busy.root);
2405 spin_unlock(&vn->busy.lock);
2406
2407 if (va)
2408 return va;
2409 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2410
2411 return NULL;
2412 }
2413
find_unlink_vmap_area(unsigned long addr)2414 static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
2415 {
2416 struct vmap_node *vn;
2417 struct vmap_area *va;
2418 int i, j;
2419
2420 /*
2421 * Check the comment in the find_vmap_area() about the loop.
2422 */
2423 i = j = addr_to_node_id(addr);
2424 do {
2425 vn = &vmap_nodes[i];
2426
2427 spin_lock(&vn->busy.lock);
2428 va = __find_vmap_area(addr, &vn->busy.root);
2429 if (va)
2430 unlink_va(va, &vn->busy.root);
2431 spin_unlock(&vn->busy.lock);
2432
2433 if (va)
2434 return va;
2435 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2436
2437 return NULL;
2438 }
2439
2440 /*** Per cpu kva allocator ***/
2441
2442 /*
2443 * vmap space is limited especially on 32 bit architectures. Ensure there is
2444 * room for at least 16 percpu vmap blocks per CPU.
2445 */
2446 /*
2447 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
2448 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
2449 * instead (we just need a rough idea)
2450 */
2451 #if BITS_PER_LONG == 32
2452 #define VMALLOC_SPACE (128UL*1024*1024)
2453 #else
2454 #define VMALLOC_SPACE (128UL*1024*1024*1024)
2455 #endif
2456
2457 #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
2458 #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
2459 #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
2460 #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
2461 #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
2462 #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
2463 #define VMAP_BBMAP_BITS \
2464 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
2465 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
2466 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
2467
2468 #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
2469
2470 /*
2471 * Purge threshold to prevent overeager purging of fragmented blocks for
2472 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
2473 */
2474 #define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
2475
2476 #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
2477 #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
2478 #define VMAP_FLAGS_MASK 0x3
2479
2480 struct vmap_block_queue {
2481 spinlock_t lock;
2482 struct list_head free;
2483
2484 /*
2485 * An xarray requires an extra memory dynamically to
2486 * be allocated. If it is an issue, we can use rb-tree
2487 * instead.
2488 */
2489 struct xarray vmap_blocks;
2490 };
2491
2492 struct vmap_block {
2493 spinlock_t lock;
2494 struct vmap_area *va;
2495 unsigned long free, dirty;
2496 DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
2497 unsigned long dirty_min, dirty_max; /*< dirty range */
2498 struct list_head free_list;
2499 struct rcu_head rcu_head;
2500 struct list_head purge;
2501 };
2502
2503 /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
2504 static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
2505
2506 /*
2507 * In order to fast access to any "vmap_block" associated with a
2508 * specific address, we use a hash.
2509 *
2510 * A per-cpu vmap_block_queue is used in both ways, to serialize
2511 * an access to free block chains among CPUs(alloc path) and it
2512 * also acts as a vmap_block hash(alloc/free paths). It means we
2513 * overload it, since we already have the per-cpu array which is
2514 * used as a hash table. When used as a hash a 'cpu' passed to
2515 * per_cpu() is not actually a CPU but rather a hash index.
2516 *
2517 * A hash function is addr_to_vb_xa() which hashes any address
2518 * to a specific index(in a hash) it belongs to. This then uses a
2519 * per_cpu() macro to access an array with generated index.
2520 *
2521 * An example:
2522 *
2523 * CPU_1 CPU_2 CPU_0
2524 * | | |
2525 * V V V
2526 * 0 10 20 30 40 50 60
2527 * |------|------|------|------|------|------|...<vmap address space>
2528 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
2529 *
2530 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
2531 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
2532 *
2533 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
2534 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
2535 *
2536 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
2537 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
2538 *
2539 * This technique almost always avoids lock contention on insert/remove,
2540 * however xarray spinlocks protect against any contention that remains.
2541 */
2542 static struct xarray *
addr_to_vb_xa(unsigned long addr)2543 addr_to_vb_xa(unsigned long addr)
2544 {
2545 int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
2546
2547 return &per_cpu(vmap_block_queue, index).vmap_blocks;
2548 }
2549
2550 /*
2551 * We should probably have a fallback mechanism to allocate virtual memory
2552 * out of partially filled vmap blocks. However vmap block sizing should be
2553 * fairly reasonable according to the vmalloc size, so it shouldn't be a
2554 * big problem.
2555 */
2556
addr_to_vb_idx(unsigned long addr)2557 static unsigned long addr_to_vb_idx(unsigned long addr)
2558 {
2559 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2560 addr /= VMAP_BLOCK_SIZE;
2561 return addr;
2562 }
2563
vmap_block_vaddr(unsigned long va_start,unsigned long pages_off)2564 static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2565 {
2566 unsigned long addr;
2567
2568 addr = va_start + (pages_off << PAGE_SHIFT);
2569 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2570 return (void *)addr;
2571 }
2572
2573 /**
2574 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2575 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2576 * @order: how many 2^order pages should be occupied in newly allocated block
2577 * @gfp_mask: flags for the page level allocator
2578 *
2579 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2580 */
new_vmap_block(unsigned int order,gfp_t gfp_mask)2581 static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2582 {
2583 struct vmap_block_queue *vbq;
2584 struct vmap_block *vb;
2585 struct vmap_area *va;
2586 struct xarray *xa;
2587 unsigned long vb_idx;
2588 int node, err;
2589 void *vaddr;
2590
2591 node = numa_node_id();
2592
2593 vb = kmalloc_node(sizeof(struct vmap_block),
2594 gfp_mask & GFP_RECLAIM_MASK, node);
2595 if (unlikely(!vb))
2596 return ERR_PTR(-ENOMEM);
2597
2598 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2599 VMALLOC_START, VMALLOC_END,
2600 node, gfp_mask,
2601 VMAP_RAM|VMAP_BLOCK, NULL);
2602 if (IS_ERR(va)) {
2603 kfree(vb);
2604 return ERR_CAST(va);
2605 }
2606
2607 vaddr = vmap_block_vaddr(va->va_start, 0);
2608 spin_lock_init(&vb->lock);
2609 vb->va = va;
2610 /* At least something should be left free */
2611 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2612 bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2613 vb->free = VMAP_BBMAP_BITS - (1UL << order);
2614 vb->dirty = 0;
2615 vb->dirty_min = VMAP_BBMAP_BITS;
2616 vb->dirty_max = 0;
2617 bitmap_set(vb->used_map, 0, (1UL << order));
2618 INIT_LIST_HEAD(&vb->free_list);
2619
2620 xa = addr_to_vb_xa(va->va_start);
2621 vb_idx = addr_to_vb_idx(va->va_start);
2622 err = xa_insert(xa, vb_idx, vb, gfp_mask);
2623 if (err) {
2624 kfree(vb);
2625 free_vmap_area(va);
2626 return ERR_PTR(err);
2627 }
2628
2629 vbq = raw_cpu_ptr(&vmap_block_queue);
2630 spin_lock(&vbq->lock);
2631 list_add_tail_rcu(&vb->free_list, &vbq->free);
2632 spin_unlock(&vbq->lock);
2633
2634 return vaddr;
2635 }
2636
free_vmap_block(struct vmap_block * vb)2637 static void free_vmap_block(struct vmap_block *vb)
2638 {
2639 struct vmap_node *vn;
2640 struct vmap_block *tmp;
2641 struct xarray *xa;
2642
2643 xa = addr_to_vb_xa(vb->va->va_start);
2644 tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2645 BUG_ON(tmp != vb);
2646
2647 vn = addr_to_node(vb->va->va_start);
2648 spin_lock(&vn->busy.lock);
2649 unlink_va(vb->va, &vn->busy.root);
2650 spin_unlock(&vn->busy.lock);
2651
2652 free_vmap_area_noflush(vb->va);
2653 kfree_rcu(vb, rcu_head);
2654 }
2655
purge_fragmented_block(struct vmap_block * vb,struct vmap_block_queue * vbq,struct list_head * purge_list,bool force_purge)2656 static bool purge_fragmented_block(struct vmap_block *vb,
2657 struct vmap_block_queue *vbq, struct list_head *purge_list,
2658 bool force_purge)
2659 {
2660 if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2661 vb->dirty == VMAP_BBMAP_BITS)
2662 return false;
2663
2664 /* Don't overeagerly purge usable blocks unless requested */
2665 if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2666 return false;
2667
2668 /* prevent further allocs after releasing lock */
2669 WRITE_ONCE(vb->free, 0);
2670 /* prevent purging it again */
2671 WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2672 vb->dirty_min = 0;
2673 vb->dirty_max = VMAP_BBMAP_BITS;
2674 spin_lock(&vbq->lock);
2675 list_del_rcu(&vb->free_list);
2676 spin_unlock(&vbq->lock);
2677 list_add_tail(&vb->purge, purge_list);
2678 return true;
2679 }
2680
free_purged_blocks(struct list_head * purge_list)2681 static void free_purged_blocks(struct list_head *purge_list)
2682 {
2683 struct vmap_block *vb, *n_vb;
2684
2685 list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2686 list_del(&vb->purge);
2687 free_vmap_block(vb);
2688 }
2689 }
2690
purge_fragmented_blocks(int cpu)2691 static void purge_fragmented_blocks(int cpu)
2692 {
2693 LIST_HEAD(purge);
2694 struct vmap_block *vb;
2695 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2696
2697 rcu_read_lock();
2698 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2699 unsigned long free = READ_ONCE(vb->free);
2700 unsigned long dirty = READ_ONCE(vb->dirty);
2701
2702 if (free + dirty != VMAP_BBMAP_BITS ||
2703 dirty == VMAP_BBMAP_BITS)
2704 continue;
2705
2706 spin_lock(&vb->lock);
2707 purge_fragmented_block(vb, vbq, &purge, true);
2708 spin_unlock(&vb->lock);
2709 }
2710 rcu_read_unlock();
2711 free_purged_blocks(&purge);
2712 }
2713
purge_fragmented_blocks_allcpus(void)2714 static void purge_fragmented_blocks_allcpus(void)
2715 {
2716 int cpu;
2717
2718 for_each_possible_cpu(cpu)
2719 purge_fragmented_blocks(cpu);
2720 }
2721
vb_alloc(unsigned long size,gfp_t gfp_mask)2722 static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2723 {
2724 struct vmap_block_queue *vbq;
2725 struct vmap_block *vb;
2726 void *vaddr = NULL;
2727 unsigned int order;
2728
2729 BUG_ON(offset_in_page(size));
2730 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2731 if (WARN_ON(size == 0)) {
2732 /*
2733 * Allocating 0 bytes isn't what caller wants since
2734 * get_order(0) returns funny result. Just warn and terminate
2735 * early.
2736 */
2737 return ERR_PTR(-EINVAL);
2738 }
2739 order = get_order(size);
2740
2741 rcu_read_lock();
2742 vbq = raw_cpu_ptr(&vmap_block_queue);
2743 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2744 unsigned long pages_off;
2745
2746 if (READ_ONCE(vb->free) < (1UL << order))
2747 continue;
2748
2749 spin_lock(&vb->lock);
2750 if (vb->free < (1UL << order)) {
2751 spin_unlock(&vb->lock);
2752 continue;
2753 }
2754
2755 pages_off = VMAP_BBMAP_BITS - vb->free;
2756 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2757 WRITE_ONCE(vb->free, vb->free - (1UL << order));
2758 bitmap_set(vb->used_map, pages_off, (1UL << order));
2759 if (vb->free == 0) {
2760 spin_lock(&vbq->lock);
2761 list_del_rcu(&vb->free_list);
2762 spin_unlock(&vbq->lock);
2763 }
2764
2765 spin_unlock(&vb->lock);
2766 break;
2767 }
2768
2769 rcu_read_unlock();
2770
2771 /* Allocate new block if nothing was found */
2772 if (!vaddr)
2773 vaddr = new_vmap_block(order, gfp_mask);
2774
2775 return vaddr;
2776 }
2777
vb_free(unsigned long addr,unsigned long size)2778 static void vb_free(unsigned long addr, unsigned long size)
2779 {
2780 unsigned long offset;
2781 unsigned int order;
2782 struct vmap_block *vb;
2783 struct xarray *xa;
2784
2785 BUG_ON(offset_in_page(size));
2786 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2787
2788 flush_cache_vunmap(addr, addr + size);
2789
2790 order = get_order(size);
2791 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2792
2793 xa = addr_to_vb_xa(addr);
2794 vb = xa_load(xa, addr_to_vb_idx(addr));
2795
2796 spin_lock(&vb->lock);
2797 bitmap_clear(vb->used_map, offset, (1UL << order));
2798 spin_unlock(&vb->lock);
2799
2800 vunmap_range_noflush(addr, addr + size);
2801
2802 if (debug_pagealloc_enabled_static())
2803 flush_tlb_kernel_range(addr, addr + size);
2804
2805 spin_lock(&vb->lock);
2806
2807 /* Expand the not yet TLB flushed dirty range */
2808 vb->dirty_min = min(vb->dirty_min, offset);
2809 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2810
2811 WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2812 if (vb->dirty == VMAP_BBMAP_BITS) {
2813 BUG_ON(vb->free);
2814 spin_unlock(&vb->lock);
2815 free_vmap_block(vb);
2816 } else
2817 spin_unlock(&vb->lock);
2818 }
2819
_vm_unmap_aliases(unsigned long start,unsigned long end,int flush)2820 static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2821 {
2822 LIST_HEAD(purge_list);
2823 int cpu;
2824
2825 if (unlikely(!vmap_initialized))
2826 return;
2827
2828 mutex_lock(&vmap_purge_lock);
2829
2830 for_each_possible_cpu(cpu) {
2831 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2832 struct vmap_block *vb;
2833 unsigned long idx;
2834
2835 rcu_read_lock();
2836 xa_for_each(&vbq->vmap_blocks, idx, vb) {
2837 spin_lock(&vb->lock);
2838
2839 /*
2840 * Try to purge a fragmented block first. If it's
2841 * not purgeable, check whether there is dirty
2842 * space to be flushed.
2843 */
2844 if (!purge_fragmented_block(vb, vbq, &purge_list, false) &&
2845 vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2846 unsigned long va_start = vb->va->va_start;
2847 unsigned long s, e;
2848
2849 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2850 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2851
2852 start = min(s, start);
2853 end = max(e, end);
2854
2855 /* Prevent that this is flushed again */
2856 vb->dirty_min = VMAP_BBMAP_BITS;
2857 vb->dirty_max = 0;
2858
2859 flush = 1;
2860 }
2861 spin_unlock(&vb->lock);
2862 }
2863 rcu_read_unlock();
2864 }
2865 free_purged_blocks(&purge_list);
2866
2867 if (!__purge_vmap_area_lazy(start, end, false) && flush)
2868 flush_tlb_kernel_range(start, end);
2869 mutex_unlock(&vmap_purge_lock);
2870 }
2871
2872 /**
2873 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2874 *
2875 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2876 * to amortize TLB flushing overheads. What this means is that any page you
2877 * have now, may, in a former life, have been mapped into kernel virtual
2878 * address by the vmap layer and so there might be some CPUs with TLB entries
2879 * still referencing that page (additional to the regular 1:1 kernel mapping).
2880 *
2881 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2882 * be sure that none of the pages we have control over will have any aliases
2883 * from the vmap layer.
2884 */
vm_unmap_aliases(void)2885 void vm_unmap_aliases(void)
2886 {
2887 unsigned long start = ULONG_MAX, end = 0;
2888 int flush = 0;
2889
2890 _vm_unmap_aliases(start, end, flush);
2891 }
2892 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2893
2894 /**
2895 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2896 * @mem: the pointer returned by vm_map_ram
2897 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2898 */
vm_unmap_ram(const void * mem,unsigned int count)2899 void vm_unmap_ram(const void *mem, unsigned int count)
2900 {
2901 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2902 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2903 struct vmap_area *va;
2904
2905 might_sleep();
2906 BUG_ON(!addr);
2907 BUG_ON(addr < VMALLOC_START);
2908 BUG_ON(addr > VMALLOC_END);
2909 BUG_ON(!PAGE_ALIGNED(addr));
2910
2911 kasan_poison_vmalloc(mem, size);
2912
2913 if (likely(count <= VMAP_MAX_ALLOC)) {
2914 debug_check_no_locks_freed(mem, size);
2915 vb_free(addr, size);
2916 return;
2917 }
2918
2919 va = find_unlink_vmap_area(addr);
2920 if (WARN_ON_ONCE(!va))
2921 return;
2922
2923 debug_check_no_locks_freed((void *)va->va_start,
2924 (va->va_end - va->va_start));
2925 free_unmap_vmap_area(va);
2926 }
2927 EXPORT_SYMBOL(vm_unmap_ram);
2928
2929 /**
2930 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2931 * @pages: an array of pointers to the pages to be mapped
2932 * @count: number of pages
2933 * @node: prefer to allocate data structures on this node
2934 *
2935 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2936 * faster than vmap so it's good. But if you mix long-life and short-life
2937 * objects with vm_map_ram(), it could consume lots of address space through
2938 * fragmentation (especially on a 32bit machine). You could see failures in
2939 * the end. Please use this function for short-lived objects.
2940 *
2941 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2942 */
vm_map_ram(struct page ** pages,unsigned int count,int node)2943 void *vm_map_ram(struct page **pages, unsigned int count, int node)
2944 {
2945 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2946 unsigned long addr;
2947 void *mem;
2948
2949 if (likely(count <= VMAP_MAX_ALLOC)) {
2950 mem = vb_alloc(size, GFP_KERNEL);
2951 if (IS_ERR(mem))
2952 return NULL;
2953 addr = (unsigned long)mem;
2954 } else {
2955 struct vmap_area *va;
2956 va = alloc_vmap_area(size, PAGE_SIZE,
2957 VMALLOC_START, VMALLOC_END,
2958 node, GFP_KERNEL, VMAP_RAM,
2959 NULL);
2960 if (IS_ERR(va))
2961 return NULL;
2962
2963 addr = va->va_start;
2964 mem = (void *)addr;
2965 }
2966
2967 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2968 pages, PAGE_SHIFT) < 0) {
2969 vm_unmap_ram(mem, count);
2970 return NULL;
2971 }
2972
2973 /*
2974 * Mark the pages as accessible, now that they are mapped.
2975 * With hardware tag-based KASAN, marking is skipped for
2976 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2977 */
2978 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2979
2980 return mem;
2981 }
2982 EXPORT_SYMBOL(vm_map_ram);
2983
2984 static struct vm_struct *vmlist __initdata;
2985
vm_area_page_order(struct vm_struct * vm)2986 static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2987 {
2988 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2989 return vm->page_order;
2990 #else
2991 return 0;
2992 #endif
2993 }
2994
set_vm_area_page_order(struct vm_struct * vm,unsigned int order)2995 static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2996 {
2997 #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2998 vm->page_order = order;
2999 #else
3000 BUG_ON(order != 0);
3001 #endif
3002 }
3003
3004 /**
3005 * vm_area_add_early - add vmap area early during boot
3006 * @vm: vm_struct to add
3007 *
3008 * This function is used to add fixed kernel vm area to vmlist before
3009 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
3010 * should contain proper values and the other fields should be zero.
3011 *
3012 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3013 */
vm_area_add_early(struct vm_struct * vm)3014 void __init vm_area_add_early(struct vm_struct *vm)
3015 {
3016 struct vm_struct *tmp, **p;
3017
3018 BUG_ON(vmap_initialized);
3019 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
3020 if (tmp->addr >= vm->addr) {
3021 BUG_ON(tmp->addr < vm->addr + vm->size);
3022 break;
3023 } else
3024 BUG_ON(tmp->addr + tmp->size > vm->addr);
3025 }
3026 vm->next = *p;
3027 *p = vm;
3028 }
3029
3030 /**
3031 * vm_area_register_early - register vmap area early during boot
3032 * @vm: vm_struct to register
3033 * @align: requested alignment
3034 *
3035 * This function is used to register kernel vm area before
3036 * vmalloc_init() is called. @vm->size and @vm->flags should contain
3037 * proper values on entry and other fields should be zero. On return,
3038 * vm->addr contains the allocated address.
3039 *
3040 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3041 */
vm_area_register_early(struct vm_struct * vm,size_t align)3042 void __init vm_area_register_early(struct vm_struct *vm, size_t align)
3043 {
3044 unsigned long addr = ALIGN(VMALLOC_START, align);
3045 struct vm_struct *cur, **p;
3046
3047 BUG_ON(vmap_initialized);
3048
3049 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
3050 if ((unsigned long)cur->addr - addr >= vm->size)
3051 break;
3052 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
3053 }
3054
3055 BUG_ON(addr > VMALLOC_END - vm->size);
3056 vm->addr = (void *)addr;
3057 vm->next = *p;
3058 *p = vm;
3059 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
3060 }
3061
clear_vm_uninitialized_flag(struct vm_struct * vm)3062 static void clear_vm_uninitialized_flag(struct vm_struct *vm)
3063 {
3064 /*
3065 * Before removing VM_UNINITIALIZED,
3066 * we should make sure that vm has proper values.
3067 * Pair with smp_rmb() in show_numa_info().
3068 */
3069 smp_wmb();
3070 vm->flags &= ~VM_UNINITIALIZED;
3071 }
3072
__get_vm_area_node(unsigned long size,unsigned long align,unsigned long shift,unsigned long flags,unsigned long start,unsigned long end,int node,gfp_t gfp_mask,const void * caller)3073 static struct vm_struct *__get_vm_area_node(unsigned long size,
3074 unsigned long align, unsigned long shift, unsigned long flags,
3075 unsigned long start, unsigned long end, int node,
3076 gfp_t gfp_mask, const void *caller)
3077 {
3078 struct vmap_area *va;
3079 struct vm_struct *area;
3080 unsigned long requested_size = size;
3081
3082 BUG_ON(in_interrupt());
3083 size = ALIGN(size, 1ul << shift);
3084 if (unlikely(!size))
3085 return NULL;
3086
3087 if (flags & VM_IOREMAP)
3088 align = 1ul << clamp_t(int, get_count_order_long(size),
3089 PAGE_SHIFT, IOREMAP_MAX_ORDER);
3090
3091 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
3092 if (unlikely(!area))
3093 return NULL;
3094
3095 if (!(flags & VM_NO_GUARD))
3096 size += PAGE_SIZE;
3097
3098 area->flags = flags;
3099 area->caller = caller;
3100
3101 va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0, area);
3102 if (IS_ERR(va)) {
3103 kfree(area);
3104 return NULL;
3105 }
3106
3107 /*
3108 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
3109 * best-effort approach, as they can be mapped outside of vmalloc code.
3110 * For VM_ALLOC mappings, the pages are marked as accessible after
3111 * getting mapped in __vmalloc_node_range().
3112 * With hardware tag-based KASAN, marking is skipped for
3113 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3114 */
3115 if (!(flags & VM_ALLOC))
3116 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
3117 KASAN_VMALLOC_PROT_NORMAL);
3118
3119 return area;
3120 }
3121
__get_vm_area_caller(unsigned long size,unsigned long flags,unsigned long start,unsigned long end,const void * caller)3122 struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
3123 unsigned long start, unsigned long end,
3124 const void *caller)
3125 {
3126 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
3127 NUMA_NO_NODE, GFP_KERNEL, caller);
3128 }
3129
3130 /**
3131 * get_vm_area - reserve a contiguous kernel virtual area
3132 * @size: size of the area
3133 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
3134 *
3135 * Search an area of @size in the kernel virtual mapping area,
3136 * and reserved it for out purposes. Returns the area descriptor
3137 * on success or %NULL on failure.
3138 *
3139 * Return: the area descriptor on success or %NULL on failure.
3140 */
get_vm_area(unsigned long size,unsigned long flags)3141 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
3142 {
3143 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3144 VMALLOC_START, VMALLOC_END,
3145 NUMA_NO_NODE, GFP_KERNEL,
3146 __builtin_return_address(0));
3147 }
3148
get_vm_area_caller(unsigned long size,unsigned long flags,const void * caller)3149 struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
3150 const void *caller)
3151 {
3152 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3153 VMALLOC_START, VMALLOC_END,
3154 NUMA_NO_NODE, GFP_KERNEL, caller);
3155 }
3156
3157 /**
3158 * find_vm_area - find a continuous kernel virtual area
3159 * @addr: base address
3160 *
3161 * Search for the kernel VM area starting at @addr, and return it.
3162 * It is up to the caller to do all required locking to keep the returned
3163 * pointer valid.
3164 *
3165 * Return: the area descriptor on success or %NULL on failure.
3166 */
find_vm_area(const void * addr)3167 struct vm_struct *find_vm_area(const void *addr)
3168 {
3169 struct vmap_area *va;
3170
3171 va = find_vmap_area((unsigned long)addr);
3172 if (!va)
3173 return NULL;
3174
3175 return va->vm;
3176 }
3177
3178 /**
3179 * remove_vm_area - find and remove a continuous kernel virtual area
3180 * @addr: base address
3181 *
3182 * Search for the kernel VM area starting at @addr, and remove it.
3183 * This function returns the found VM area, but using it is NOT safe
3184 * on SMP machines, except for its size or flags.
3185 *
3186 * Return: the area descriptor on success or %NULL on failure.
3187 */
remove_vm_area(const void * addr)3188 struct vm_struct *remove_vm_area(const void *addr)
3189 {
3190 struct vmap_area *va;
3191 struct vm_struct *vm;
3192
3193 might_sleep();
3194
3195 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
3196 addr))
3197 return NULL;
3198
3199 va = find_unlink_vmap_area((unsigned long)addr);
3200 if (!va || !va->vm)
3201 return NULL;
3202 vm = va->vm;
3203
3204 debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
3205 debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
3206 kasan_free_module_shadow(vm);
3207 kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
3208
3209 free_unmap_vmap_area(va);
3210 return vm;
3211 }
3212
set_area_direct_map(const struct vm_struct * area,int (* set_direct_map)(struct page * page))3213 static inline void set_area_direct_map(const struct vm_struct *area,
3214 int (*set_direct_map)(struct page *page))
3215 {
3216 int i;
3217
3218 /* HUGE_VMALLOC passes small pages to set_direct_map */
3219 for (i = 0; i < area->nr_pages; i++)
3220 if (page_address(area->pages[i]))
3221 set_direct_map(area->pages[i]);
3222 }
3223
3224 /*
3225 * Flush the vm mapping and reset the direct map.
3226 */
vm_reset_perms(struct vm_struct * area)3227 static void vm_reset_perms(struct vm_struct *area)
3228 {
3229 unsigned long start = ULONG_MAX, end = 0;
3230 unsigned int page_order = vm_area_page_order(area);
3231 int flush_dmap = 0;
3232 int i;
3233
3234 /*
3235 * Find the start and end range of the direct mappings to make sure that
3236 * the vm_unmap_aliases() flush includes the direct map.
3237 */
3238 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
3239 unsigned long addr = (unsigned long)page_address(area->pages[i]);
3240
3241 if (addr) {
3242 unsigned long page_size;
3243
3244 page_size = PAGE_SIZE << page_order;
3245 start = min(addr, start);
3246 end = max(addr + page_size, end);
3247 flush_dmap = 1;
3248 }
3249 }
3250
3251 /*
3252 * Set direct map to something invalid so that it won't be cached if
3253 * there are any accesses after the TLB flush, then flush the TLB and
3254 * reset the direct map permissions to the default.
3255 */
3256 set_area_direct_map(area, set_direct_map_invalid_noflush);
3257 _vm_unmap_aliases(start, end, flush_dmap);
3258 set_area_direct_map(area, set_direct_map_default_noflush);
3259 }
3260
delayed_vfree_work(struct work_struct * w)3261 static void delayed_vfree_work(struct work_struct *w)
3262 {
3263 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
3264 struct llist_node *t, *llnode;
3265
3266 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
3267 vfree(llnode);
3268 }
3269
3270 /**
3271 * vfree_atomic - release memory allocated by vmalloc()
3272 * @addr: memory base address
3273 *
3274 * This one is just like vfree() but can be called in any atomic context
3275 * except NMIs.
3276 */
vfree_atomic(const void * addr)3277 void vfree_atomic(const void *addr)
3278 {
3279 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
3280
3281 BUG_ON(in_nmi());
3282 kmemleak_free(addr);
3283
3284 /*
3285 * Use raw_cpu_ptr() because this can be called from preemptible
3286 * context. Preemption is absolutely fine here, because the llist_add()
3287 * implementation is lockless, so it works even if we are adding to
3288 * another cpu's list. schedule_work() should be fine with this too.
3289 */
3290 if (addr && llist_add((struct llist_node *)addr, &p->list))
3291 schedule_work(&p->wq);
3292 }
3293
3294 /**
3295 * vfree - Release memory allocated by vmalloc()
3296 * @addr: Memory base address
3297 *
3298 * Free the virtually continuous memory area starting at @addr, as obtained
3299 * from one of the vmalloc() family of APIs. This will usually also free the
3300 * physical memory underlying the virtual allocation, but that memory is
3301 * reference counted, so it will not be freed until the last user goes away.
3302 *
3303 * If @addr is NULL, no operation is performed.
3304 *
3305 * Context:
3306 * May sleep if called *not* from interrupt context.
3307 * Must not be called in NMI context (strictly speaking, it could be
3308 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
3309 * conventions for vfree() arch-dependent would be a really bad idea).
3310 */
vfree(const void * addr)3311 void vfree(const void *addr)
3312 {
3313 struct vm_struct *vm;
3314 int i;
3315
3316 if (unlikely(in_interrupt())) {
3317 vfree_atomic(addr);
3318 return;
3319 }
3320
3321 BUG_ON(in_nmi());
3322 kmemleak_free(addr);
3323 might_sleep();
3324
3325 if (!addr)
3326 return;
3327
3328 vm = remove_vm_area(addr);
3329 if (unlikely(!vm)) {
3330 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
3331 addr);
3332 return;
3333 }
3334
3335 if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
3336 vm_reset_perms(vm);
3337 for (i = 0; i < vm->nr_pages; i++) {
3338 struct page *page = vm->pages[i];
3339
3340 BUG_ON(!page);
3341 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
3342 /*
3343 * High-order allocs for huge vmallocs are split, so
3344 * can be freed as an array of order-0 allocations
3345 */
3346 __free_page(page);
3347 cond_resched();
3348 }
3349 atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
3350 kvfree(vm->pages);
3351 kfree(vm);
3352 }
3353 EXPORT_SYMBOL(vfree);
3354
3355 /**
3356 * vunmap - release virtual mapping obtained by vmap()
3357 * @addr: memory base address
3358 *
3359 * Free the virtually contiguous memory area starting at @addr,
3360 * which was created from the page array passed to vmap().
3361 *
3362 * Must not be called in interrupt context.
3363 */
vunmap(const void * addr)3364 void vunmap(const void *addr)
3365 {
3366 struct vm_struct *vm;
3367
3368 BUG_ON(in_interrupt());
3369 might_sleep();
3370
3371 if (!addr)
3372 return;
3373 vm = remove_vm_area(addr);
3374 if (unlikely(!vm)) {
3375 WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
3376 addr);
3377 return;
3378 }
3379 kfree(vm);
3380 }
3381 EXPORT_SYMBOL(vunmap);
3382
3383 /**
3384 * vmap - map an array of pages into virtually contiguous space
3385 * @pages: array of page pointers
3386 * @count: number of pages to map
3387 * @flags: vm_area->flags
3388 * @prot: page protection for the mapping
3389 *
3390 * Maps @count pages from @pages into contiguous kernel virtual space.
3391 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
3392 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
3393 * are transferred from the caller to vmap(), and will be freed / dropped when
3394 * vfree() is called on the return value.
3395 *
3396 * Return: the address of the area or %NULL on failure
3397 */
vmap(struct page ** pages,unsigned int count,unsigned long flags,pgprot_t prot)3398 void *vmap(struct page **pages, unsigned int count,
3399 unsigned long flags, pgprot_t prot)
3400 {
3401 struct vm_struct *area;
3402 unsigned long addr;
3403 unsigned long size; /* In bytes */
3404
3405 might_sleep();
3406
3407 if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
3408 return NULL;
3409
3410 /*
3411 * Your top guard is someone else's bottom guard. Not having a top
3412 * guard compromises someone else's mappings too.
3413 */
3414 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
3415 flags &= ~VM_NO_GUARD;
3416
3417 if (count > totalram_pages())
3418 return NULL;
3419
3420 size = (unsigned long)count << PAGE_SHIFT;
3421 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
3422 if (!area)
3423 return NULL;
3424
3425 addr = (unsigned long)area->addr;
3426 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
3427 pages, PAGE_SHIFT) < 0) {
3428 vunmap(area->addr);
3429 return NULL;
3430 }
3431
3432 if (flags & VM_MAP_PUT_PAGES) {
3433 area->pages = pages;
3434 area->nr_pages = count;
3435 }
3436 return area->addr;
3437 }
3438 EXPORT_SYMBOL(vmap);
3439
3440 #ifdef CONFIG_VMAP_PFN
3441 struct vmap_pfn_data {
3442 unsigned long *pfns;
3443 pgprot_t prot;
3444 unsigned int idx;
3445 };
3446
vmap_pfn_apply(pte_t * pte,unsigned long addr,void * private)3447 static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
3448 {
3449 struct vmap_pfn_data *data = private;
3450 unsigned long pfn = data->pfns[data->idx];
3451 pte_t ptent;
3452
3453 if (WARN_ON_ONCE(pfn_valid(pfn)))
3454 return -EINVAL;
3455
3456 ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
3457 set_pte_at(&init_mm, addr, pte, ptent);
3458
3459 data->idx++;
3460 return 0;
3461 }
3462
3463 /**
3464 * vmap_pfn - map an array of PFNs into virtually contiguous space
3465 * @pfns: array of PFNs
3466 * @count: number of pages to map
3467 * @prot: page protection for the mapping
3468 *
3469 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
3470 * the start address of the mapping.
3471 */
vmap_pfn(unsigned long * pfns,unsigned int count,pgprot_t prot)3472 void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
3473 {
3474 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
3475 struct vm_struct *area;
3476
3477 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
3478 __builtin_return_address(0));
3479 if (!area)
3480 return NULL;
3481 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3482 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
3483 free_vm_area(area);
3484 return NULL;
3485 }
3486
3487 flush_cache_vmap((unsigned long)area->addr,
3488 (unsigned long)area->addr + count * PAGE_SIZE);
3489
3490 return area->addr;
3491 }
3492 EXPORT_SYMBOL_GPL(vmap_pfn);
3493 #endif /* CONFIG_VMAP_PFN */
3494
3495 static inline unsigned int
vm_area_alloc_pages(gfp_t gfp,int nid,unsigned int order,unsigned int nr_pages,struct page ** pages)3496 vm_area_alloc_pages(gfp_t gfp, int nid,
3497 unsigned int order, unsigned int nr_pages, struct page **pages)
3498 {
3499 unsigned int nr_allocated = 0;
3500 gfp_t alloc_gfp = gfp;
3501 bool nofail = gfp & __GFP_NOFAIL;
3502 struct page *page;
3503 int i;
3504
3505 /*
3506 * For order-0 pages we make use of bulk allocator, if
3507 * the page array is partly or not at all populated due
3508 * to fails, fallback to a single page allocator that is
3509 * more permissive.
3510 */
3511 if (!order) {
3512 /* bulk allocator doesn't support nofail req. officially */
3513 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
3514
3515 while (nr_allocated < nr_pages) {
3516 unsigned int nr, nr_pages_request;
3517
3518 /*
3519 * A maximum allowed request is hard-coded and is 100
3520 * pages per call. That is done in order to prevent a
3521 * long preemption off scenario in the bulk-allocator
3522 * so the range is [1:100].
3523 */
3524 nr_pages_request = min(100U, nr_pages - nr_allocated);
3525
3526 /* memory allocation should consider mempolicy, we can't
3527 * wrongly use nearest node when nid == NUMA_NO_NODE,
3528 * otherwise memory may be allocated in only one node,
3529 * but mempolicy wants to alloc memory by interleaving.
3530 */
3531 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3532 nr = alloc_pages_bulk_array_mempolicy_noprof(bulk_gfp,
3533 nr_pages_request,
3534 pages + nr_allocated);
3535
3536 else
3537 nr = alloc_pages_bulk_array_node_noprof(bulk_gfp, nid,
3538 nr_pages_request,
3539 pages + nr_allocated);
3540
3541 nr_allocated += nr;
3542 cond_resched();
3543
3544 /*
3545 * If zero or pages were obtained partly,
3546 * fallback to a single page allocator.
3547 */
3548 if (nr != nr_pages_request)
3549 break;
3550 }
3551 } else if (gfp & __GFP_NOFAIL) {
3552 /*
3553 * Higher order nofail allocations are really expensive and
3554 * potentially dangerous (pre-mature OOM, disruptive reclaim
3555 * and compaction etc.
3556 */
3557 alloc_gfp &= ~__GFP_NOFAIL;
3558 }
3559
3560 /* High-order pages or fallback path if "bulk" fails. */
3561 while (nr_allocated < nr_pages) {
3562 if (!nofail && fatal_signal_pending(current))
3563 break;
3564
3565 if (nid == NUMA_NO_NODE)
3566 page = alloc_pages_noprof(alloc_gfp, order);
3567 else
3568 page = alloc_pages_node_noprof(nid, alloc_gfp, order);
3569 if (unlikely(!page)) {
3570 if (!nofail)
3571 break;
3572
3573 /* fall back to the zero order allocations */
3574 alloc_gfp |= __GFP_NOFAIL;
3575 order = 0;
3576 continue;
3577 }
3578
3579 /*
3580 * Higher order allocations must be able to be treated as
3581 * indepdenent small pages by callers (as they can with
3582 * small-page vmallocs). Some drivers do their own refcounting
3583 * on vmalloc_to_page() pages, some use page->mapping,
3584 * page->lru, etc.
3585 */
3586 if (order)
3587 split_page(page, order);
3588
3589 /*
3590 * Careful, we allocate and map page-order pages, but
3591 * tracking is done per PAGE_SIZE page so as to keep the
3592 * vm_struct APIs independent of the physical/mapped size.
3593 */
3594 for (i = 0; i < (1U << order); i++)
3595 pages[nr_allocated + i] = page + i;
3596
3597 cond_resched();
3598 nr_allocated += 1U << order;
3599 }
3600
3601 return nr_allocated;
3602 }
3603
__vmalloc_area_node(struct vm_struct * area,gfp_t gfp_mask,pgprot_t prot,unsigned int page_shift,int node)3604 static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3605 pgprot_t prot, unsigned int page_shift,
3606 int node)
3607 {
3608 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3609 bool nofail = gfp_mask & __GFP_NOFAIL;
3610 unsigned long addr = (unsigned long)area->addr;
3611 unsigned long size = get_vm_area_size(area);
3612 unsigned long array_size;
3613 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3614 unsigned int page_order;
3615 unsigned int flags;
3616 int ret;
3617
3618 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3619
3620 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3621 gfp_mask |= __GFP_HIGHMEM;
3622
3623 /* Please note that the recursion is strictly bounded. */
3624 if (array_size > PAGE_SIZE) {
3625 area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node,
3626 area->caller);
3627 } else {
3628 area->pages = kmalloc_node_noprof(array_size, nested_gfp, node);
3629 }
3630
3631 if (!area->pages) {
3632 warn_alloc(gfp_mask, NULL,
3633 "vmalloc error: size %lu, failed to allocated page array size %lu",
3634 nr_small_pages * PAGE_SIZE, array_size);
3635 free_vm_area(area);
3636 return NULL;
3637 }
3638
3639 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3640 page_order = vm_area_page_order(area);
3641
3642 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3643 node, page_order, nr_small_pages, area->pages);
3644
3645 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3646 if (gfp_mask & __GFP_ACCOUNT) {
3647 int i;
3648
3649 for (i = 0; i < area->nr_pages; i++)
3650 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3651 }
3652
3653 /*
3654 * If not enough pages were obtained to accomplish an
3655 * allocation request, free them via vfree() if any.
3656 */
3657 if (area->nr_pages != nr_small_pages) {
3658 /*
3659 * vm_area_alloc_pages() can fail due to insufficient memory but
3660 * also:-
3661 *
3662 * - a pending fatal signal
3663 * - insufficient huge page-order pages
3664 *
3665 * Since we always retry allocations at order-0 in the huge page
3666 * case a warning for either is spurious.
3667 */
3668 if (!fatal_signal_pending(current) && page_order == 0)
3669 warn_alloc(gfp_mask, NULL,
3670 "vmalloc error: size %lu, failed to allocate pages",
3671 area->nr_pages * PAGE_SIZE);
3672 goto fail;
3673 }
3674
3675 /*
3676 * page tables allocations ignore external gfp mask, enforce it
3677 * by the scope API
3678 */
3679 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3680 flags = memalloc_nofs_save();
3681 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3682 flags = memalloc_noio_save();
3683
3684 do {
3685 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3686 page_shift);
3687 if (nofail && (ret < 0))
3688 schedule_timeout_uninterruptible(1);
3689 } while (nofail && (ret < 0));
3690
3691 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3692 memalloc_nofs_restore(flags);
3693 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3694 memalloc_noio_restore(flags);
3695
3696 if (ret < 0) {
3697 warn_alloc(gfp_mask, NULL,
3698 "vmalloc error: size %lu, failed to map pages",
3699 area->nr_pages * PAGE_SIZE);
3700 goto fail;
3701 }
3702
3703 return area->addr;
3704
3705 fail:
3706 vfree(area->addr);
3707 return NULL;
3708 }
3709
3710 /**
3711 * __vmalloc_node_range - allocate virtually contiguous memory
3712 * @size: allocation size
3713 * @align: desired alignment
3714 * @start: vm area range start
3715 * @end: vm area range end
3716 * @gfp_mask: flags for the page level allocator
3717 * @prot: protection mask for the allocated pages
3718 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3719 * @node: node to use for allocation or NUMA_NO_NODE
3720 * @caller: caller's return address
3721 *
3722 * Allocate enough pages to cover @size from the page level
3723 * allocator with @gfp_mask flags. Please note that the full set of gfp
3724 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3725 * supported.
3726 * Zone modifiers are not supported. From the reclaim modifiers
3727 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3728 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3729 * __GFP_RETRY_MAYFAIL are not supported).
3730 *
3731 * __GFP_NOWARN can be used to suppress failures messages.
3732 *
3733 * Map them into contiguous kernel virtual space, using a pagetable
3734 * protection of @prot.
3735 *
3736 * Return: the address of the area or %NULL on failure
3737 */
__vmalloc_node_range_noprof(unsigned long size,unsigned long align,unsigned long start,unsigned long end,gfp_t gfp_mask,pgprot_t prot,unsigned long vm_flags,int node,const void * caller)3738 void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align,
3739 unsigned long start, unsigned long end, gfp_t gfp_mask,
3740 pgprot_t prot, unsigned long vm_flags, int node,
3741 const void *caller)
3742 {
3743 struct vm_struct *area;
3744 void *ret;
3745 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3746 unsigned long real_size = size;
3747 unsigned long real_align = align;
3748 unsigned int shift = PAGE_SHIFT;
3749
3750 if (WARN_ON_ONCE(!size))
3751 return NULL;
3752
3753 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3754 warn_alloc(gfp_mask, NULL,
3755 "vmalloc error: size %lu, exceeds total pages",
3756 real_size);
3757 return NULL;
3758 }
3759
3760 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3761 unsigned long size_per_node;
3762
3763 /*
3764 * Try huge pages. Only try for PAGE_KERNEL allocations,
3765 * others like modules don't yet expect huge pages in
3766 * their allocations due to apply_to_page_range not
3767 * supporting them.
3768 */
3769
3770 size_per_node = size;
3771 if (node == NUMA_NO_NODE)
3772 size_per_node /= num_online_nodes();
3773 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3774 shift = PMD_SHIFT;
3775 else
3776 shift = arch_vmap_pte_supported_shift(size_per_node);
3777
3778 align = max(real_align, 1UL << shift);
3779 size = ALIGN(real_size, 1UL << shift);
3780 }
3781
3782 again:
3783 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3784 VM_UNINITIALIZED | vm_flags, start, end, node,
3785 gfp_mask, caller);
3786 if (!area) {
3787 bool nofail = gfp_mask & __GFP_NOFAIL;
3788 warn_alloc(gfp_mask, NULL,
3789 "vmalloc error: size %lu, vm_struct allocation failed%s",
3790 real_size, (nofail) ? ". Retrying." : "");
3791 if (nofail) {
3792 schedule_timeout_uninterruptible(1);
3793 goto again;
3794 }
3795 goto fail;
3796 }
3797
3798 /*
3799 * Prepare arguments for __vmalloc_area_node() and
3800 * kasan_unpoison_vmalloc().
3801 */
3802 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3803 if (kasan_hw_tags_enabled()) {
3804 /*
3805 * Modify protection bits to allow tagging.
3806 * This must be done before mapping.
3807 */
3808 prot = arch_vmap_pgprot_tagged(prot);
3809
3810 /*
3811 * Skip page_alloc poisoning and zeroing for physical
3812 * pages backing VM_ALLOC mapping. Memory is instead
3813 * poisoned and zeroed by kasan_unpoison_vmalloc().
3814 */
3815 gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3816 }
3817
3818 /* Take note that the mapping is PAGE_KERNEL. */
3819 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3820 }
3821
3822 /* Allocate physical pages and map them into vmalloc space. */
3823 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3824 if (!ret)
3825 goto fail;
3826
3827 /*
3828 * Mark the pages as accessible, now that they are mapped.
3829 * The condition for setting KASAN_VMALLOC_INIT should complement the
3830 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3831 * to make sure that memory is initialized under the same conditions.
3832 * Tag-based KASAN modes only assign tags to normal non-executable
3833 * allocations, see __kasan_unpoison_vmalloc().
3834 */
3835 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3836 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3837 (gfp_mask & __GFP_SKIP_ZERO))
3838 kasan_flags |= KASAN_VMALLOC_INIT;
3839 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3840 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3841
3842 /*
3843 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3844 * flag. It means that vm_struct is not fully initialized.
3845 * Now, it is fully initialized, so remove this flag here.
3846 */
3847 clear_vm_uninitialized_flag(area);
3848
3849 size = PAGE_ALIGN(size);
3850 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3851 kmemleak_vmalloc(area, size, gfp_mask);
3852
3853 return area->addr;
3854
3855 fail:
3856 if (shift > PAGE_SHIFT) {
3857 shift = PAGE_SHIFT;
3858 align = real_align;
3859 size = real_size;
3860 goto again;
3861 }
3862
3863 return NULL;
3864 }
3865
3866 /**
3867 * __vmalloc_node - allocate virtually contiguous memory
3868 * @size: allocation size
3869 * @align: desired alignment
3870 * @gfp_mask: flags for the page level allocator
3871 * @node: node to use for allocation or NUMA_NO_NODE
3872 * @caller: caller's return address
3873 *
3874 * Allocate enough pages to cover @size from the page level allocator with
3875 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3876 *
3877 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3878 * and __GFP_NOFAIL are not supported
3879 *
3880 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3881 * with mm people.
3882 *
3883 * Return: pointer to the allocated memory or %NULL on error
3884 */
__vmalloc_node_noprof(unsigned long size,unsigned long align,gfp_t gfp_mask,int node,const void * caller)3885 void *__vmalloc_node_noprof(unsigned long size, unsigned long align,
3886 gfp_t gfp_mask, int node, const void *caller)
3887 {
3888 return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END,
3889 gfp_mask, PAGE_KERNEL, 0, node, caller);
3890 }
3891 /*
3892 * This is only for performance analysis of vmalloc and stress purpose.
3893 * It is required by vmalloc test module, therefore do not use it other
3894 * than that.
3895 */
3896 #ifdef CONFIG_TEST_VMALLOC_MODULE
3897 EXPORT_SYMBOL_GPL(__vmalloc_node_noprof);
3898 #endif
3899
__vmalloc_noprof(unsigned long size,gfp_t gfp_mask)3900 void *__vmalloc_noprof(unsigned long size, gfp_t gfp_mask)
3901 {
3902 return __vmalloc_node_noprof(size, 1, gfp_mask, NUMA_NO_NODE,
3903 __builtin_return_address(0));
3904 }
3905 EXPORT_SYMBOL(__vmalloc_noprof);
3906
3907 /**
3908 * vmalloc - allocate virtually contiguous memory
3909 * @size: allocation size
3910 *
3911 * Allocate enough pages to cover @size from the page level
3912 * allocator and map them into contiguous kernel virtual space.
3913 *
3914 * For tight control over page level allocator and protection flags
3915 * use __vmalloc() instead.
3916 *
3917 * Return: pointer to the allocated memory or %NULL on error
3918 */
vmalloc_noprof(unsigned long size)3919 void *vmalloc_noprof(unsigned long size)
3920 {
3921 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3922 __builtin_return_address(0));
3923 }
3924 EXPORT_SYMBOL(vmalloc_noprof);
3925
3926 /**
3927 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3928 * @size: allocation size
3929 * @gfp_mask: flags for the page level allocator
3930 *
3931 * Allocate enough pages to cover @size from the page level
3932 * allocator and map them into contiguous kernel virtual space.
3933 * If @size is greater than or equal to PMD_SIZE, allow using
3934 * huge pages for the memory
3935 *
3936 * Return: pointer to the allocated memory or %NULL on error
3937 */
vmalloc_huge_noprof(unsigned long size,gfp_t gfp_mask)3938 void *vmalloc_huge_noprof(unsigned long size, gfp_t gfp_mask)
3939 {
3940 return __vmalloc_node_range_noprof(size, 1, VMALLOC_START, VMALLOC_END,
3941 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3942 NUMA_NO_NODE, __builtin_return_address(0));
3943 }
3944 EXPORT_SYMBOL_GPL(vmalloc_huge_noprof);
3945
3946 /**
3947 * vzalloc - allocate virtually contiguous memory with zero fill
3948 * @size: allocation size
3949 *
3950 * Allocate enough pages to cover @size from the page level
3951 * allocator and map them into contiguous kernel virtual space.
3952 * The memory allocated is set to zero.
3953 *
3954 * For tight control over page level allocator and protection flags
3955 * use __vmalloc() instead.
3956 *
3957 * Return: pointer to the allocated memory or %NULL on error
3958 */
vzalloc_noprof(unsigned long size)3959 void *vzalloc_noprof(unsigned long size)
3960 {
3961 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3962 __builtin_return_address(0));
3963 }
3964 EXPORT_SYMBOL(vzalloc_noprof);
3965
3966 /**
3967 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3968 * @size: allocation size
3969 *
3970 * The resulting memory area is zeroed so it can be mapped to userspace
3971 * without leaking data.
3972 *
3973 * Return: pointer to the allocated memory or %NULL on error
3974 */
vmalloc_user_noprof(unsigned long size)3975 void *vmalloc_user_noprof(unsigned long size)
3976 {
3977 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3978 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3979 VM_USERMAP, NUMA_NO_NODE,
3980 __builtin_return_address(0));
3981 }
3982 EXPORT_SYMBOL(vmalloc_user_noprof);
3983
3984 /**
3985 * vmalloc_node - allocate memory on a specific node
3986 * @size: allocation size
3987 * @node: numa node
3988 *
3989 * Allocate enough pages to cover @size from the page level
3990 * allocator and map them into contiguous kernel virtual space.
3991 *
3992 * For tight control over page level allocator and protection flags
3993 * use __vmalloc() instead.
3994 *
3995 * Return: pointer to the allocated memory or %NULL on error
3996 */
vmalloc_node_noprof(unsigned long size,int node)3997 void *vmalloc_node_noprof(unsigned long size, int node)
3998 {
3999 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, node,
4000 __builtin_return_address(0));
4001 }
4002 EXPORT_SYMBOL(vmalloc_node_noprof);
4003
4004 /**
4005 * vzalloc_node - allocate memory on a specific node with zero fill
4006 * @size: allocation size
4007 * @node: numa node
4008 *
4009 * Allocate enough pages to cover @size from the page level
4010 * allocator and map them into contiguous kernel virtual space.
4011 * The memory allocated is set to zero.
4012 *
4013 * Return: pointer to the allocated memory or %NULL on error
4014 */
vzalloc_node_noprof(unsigned long size,int node)4015 void *vzalloc_node_noprof(unsigned long size, int node)
4016 {
4017 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, node,
4018 __builtin_return_address(0));
4019 }
4020 EXPORT_SYMBOL(vzalloc_node_noprof);
4021
4022 #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
4023 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4024 #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
4025 #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
4026 #else
4027 /*
4028 * 64b systems should always have either DMA or DMA32 zones. For others
4029 * GFP_DMA32 should do the right thing and use the normal zone.
4030 */
4031 #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4032 #endif
4033
4034 /**
4035 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
4036 * @size: allocation size
4037 *
4038 * Allocate enough 32bit PA addressable pages to cover @size from the
4039 * page level allocator and map them into contiguous kernel virtual space.
4040 *
4041 * Return: pointer to the allocated memory or %NULL on error
4042 */
vmalloc_32_noprof(unsigned long size)4043 void *vmalloc_32_noprof(unsigned long size)
4044 {
4045 return __vmalloc_node_noprof(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
4046 __builtin_return_address(0));
4047 }
4048 EXPORT_SYMBOL(vmalloc_32_noprof);
4049
4050 /**
4051 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
4052 * @size: allocation size
4053 *
4054 * The resulting memory area is 32bit addressable and zeroed so it can be
4055 * mapped to userspace without leaking data.
4056 *
4057 * Return: pointer to the allocated memory or %NULL on error
4058 */
vmalloc_32_user_noprof(unsigned long size)4059 void *vmalloc_32_user_noprof(unsigned long size)
4060 {
4061 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
4062 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
4063 VM_USERMAP, NUMA_NO_NODE,
4064 __builtin_return_address(0));
4065 }
4066 EXPORT_SYMBOL(vmalloc_32_user_noprof);
4067
4068 /*
4069 * Atomically zero bytes in the iterator.
4070 *
4071 * Returns the number of zeroed bytes.
4072 */
zero_iter(struct iov_iter * iter,size_t count)4073 static size_t zero_iter(struct iov_iter *iter, size_t count)
4074 {
4075 size_t remains = count;
4076
4077 while (remains > 0) {
4078 size_t num, copied;
4079
4080 num = min_t(size_t, remains, PAGE_SIZE);
4081 copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
4082 remains -= copied;
4083
4084 if (copied < num)
4085 break;
4086 }
4087
4088 return count - remains;
4089 }
4090
4091 /*
4092 * small helper routine, copy contents to iter from addr.
4093 * If the page is not present, fill zero.
4094 *
4095 * Returns the number of copied bytes.
4096 */
aligned_vread_iter(struct iov_iter * iter,const char * addr,size_t count)4097 static size_t aligned_vread_iter(struct iov_iter *iter,
4098 const char *addr, size_t count)
4099 {
4100 size_t remains = count;
4101 struct page *page;
4102
4103 while (remains > 0) {
4104 unsigned long offset, length;
4105 size_t copied = 0;
4106
4107 offset = offset_in_page(addr);
4108 length = PAGE_SIZE - offset;
4109 if (length > remains)
4110 length = remains;
4111 page = vmalloc_to_page(addr);
4112 /*
4113 * To do safe access to this _mapped_ area, we need lock. But
4114 * adding lock here means that we need to add overhead of
4115 * vmalloc()/vfree() calls for this _debug_ interface, rarely
4116 * used. Instead of that, we'll use an local mapping via
4117 * copy_page_to_iter_nofault() and accept a small overhead in
4118 * this access function.
4119 */
4120 if (page)
4121 copied = copy_page_to_iter_nofault(page, offset,
4122 length, iter);
4123 else
4124 copied = zero_iter(iter, length);
4125
4126 addr += copied;
4127 remains -= copied;
4128
4129 if (copied != length)
4130 break;
4131 }
4132
4133 return count - remains;
4134 }
4135
4136 /*
4137 * Read from a vm_map_ram region of memory.
4138 *
4139 * Returns the number of copied bytes.
4140 */
vmap_ram_vread_iter(struct iov_iter * iter,const char * addr,size_t count,unsigned long flags)4141 static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
4142 size_t count, unsigned long flags)
4143 {
4144 char *start;
4145 struct vmap_block *vb;
4146 struct xarray *xa;
4147 unsigned long offset;
4148 unsigned int rs, re;
4149 size_t remains, n;
4150
4151 /*
4152 * If it's area created by vm_map_ram() interface directly, but
4153 * not further subdividing and delegating management to vmap_block,
4154 * handle it here.
4155 */
4156 if (!(flags & VMAP_BLOCK))
4157 return aligned_vread_iter(iter, addr, count);
4158
4159 remains = count;
4160
4161 /*
4162 * Area is split into regions and tracked with vmap_block, read out
4163 * each region and zero fill the hole between regions.
4164 */
4165 xa = addr_to_vb_xa((unsigned long) addr);
4166 vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
4167 if (!vb)
4168 goto finished_zero;
4169
4170 spin_lock(&vb->lock);
4171 if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
4172 spin_unlock(&vb->lock);
4173 goto finished_zero;
4174 }
4175
4176 for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
4177 size_t copied;
4178
4179 if (remains == 0)
4180 goto finished;
4181
4182 start = vmap_block_vaddr(vb->va->va_start, rs);
4183
4184 if (addr < start) {
4185 size_t to_zero = min_t(size_t, start - addr, remains);
4186 size_t zeroed = zero_iter(iter, to_zero);
4187
4188 addr += zeroed;
4189 remains -= zeroed;
4190
4191 if (remains == 0 || zeroed != to_zero)
4192 goto finished;
4193 }
4194
4195 /*it could start reading from the middle of used region*/
4196 offset = offset_in_page(addr);
4197 n = ((re - rs + 1) << PAGE_SHIFT) - offset;
4198 if (n > remains)
4199 n = remains;
4200
4201 copied = aligned_vread_iter(iter, start + offset, n);
4202
4203 addr += copied;
4204 remains -= copied;
4205
4206 if (copied != n)
4207 goto finished;
4208 }
4209
4210 spin_unlock(&vb->lock);
4211
4212 finished_zero:
4213 /* zero-fill the left dirty or free regions */
4214 return count - remains + zero_iter(iter, remains);
4215 finished:
4216 /* We couldn't copy/zero everything */
4217 spin_unlock(&vb->lock);
4218 return count - remains;
4219 }
4220
4221 /**
4222 * vread_iter() - read vmalloc area in a safe way to an iterator.
4223 * @iter: the iterator to which data should be written.
4224 * @addr: vm address.
4225 * @count: number of bytes to be read.
4226 *
4227 * This function checks that addr is a valid vmalloc'ed area, and
4228 * copy data from that area to a given buffer. If the given memory range
4229 * of [addr...addr+count) includes some valid address, data is copied to
4230 * proper area of @buf. If there are memory holes, they'll be zero-filled.
4231 * IOREMAP area is treated as memory hole and no copy is done.
4232 *
4233 * If [addr...addr+count) doesn't includes any intersects with alive
4234 * vm_struct area, returns 0. @buf should be kernel's buffer.
4235 *
4236 * Note: In usual ops, vread() is never necessary because the caller
4237 * should know vmalloc() area is valid and can use memcpy().
4238 * This is for routines which have to access vmalloc area without
4239 * any information, as /proc/kcore.
4240 *
4241 * Return: number of bytes for which addr and buf should be increased
4242 * (same number as @count) or %0 if [addr...addr+count) doesn't
4243 * include any intersection with valid vmalloc area
4244 */
vread_iter(struct iov_iter * iter,const char * addr,size_t count)4245 long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
4246 {
4247 struct vmap_node *vn;
4248 struct vmap_area *va;
4249 struct vm_struct *vm;
4250 char *vaddr;
4251 size_t n, size, flags, remains;
4252 unsigned long next;
4253
4254 addr = kasan_reset_tag(addr);
4255
4256 /* Don't allow overflow */
4257 if ((unsigned long) addr + count < count)
4258 count = -(unsigned long) addr;
4259
4260 remains = count;
4261
4262 vn = find_vmap_area_exceed_addr_lock((unsigned long) addr, &va);
4263 if (!vn)
4264 goto finished_zero;
4265
4266 /* no intersects with alive vmap_area */
4267 if ((unsigned long)addr + remains <= va->va_start)
4268 goto finished_zero;
4269
4270 do {
4271 size_t copied;
4272
4273 if (remains == 0)
4274 goto finished;
4275
4276 vm = va->vm;
4277 flags = va->flags & VMAP_FLAGS_MASK;
4278 /*
4279 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
4280 * be set together with VMAP_RAM.
4281 */
4282 WARN_ON(flags == VMAP_BLOCK);
4283
4284 if (!vm && !flags)
4285 goto next_va;
4286
4287 if (vm && (vm->flags & VM_UNINITIALIZED))
4288 goto next_va;
4289
4290 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4291 smp_rmb();
4292
4293 vaddr = (char *) va->va_start;
4294 size = vm ? get_vm_area_size(vm) : va_size(va);
4295
4296 if (addr >= vaddr + size)
4297 goto next_va;
4298
4299 if (addr < vaddr) {
4300 size_t to_zero = min_t(size_t, vaddr - addr, remains);
4301 size_t zeroed = zero_iter(iter, to_zero);
4302
4303 addr += zeroed;
4304 remains -= zeroed;
4305
4306 if (remains == 0 || zeroed != to_zero)
4307 goto finished;
4308 }
4309
4310 n = vaddr + size - addr;
4311 if (n > remains)
4312 n = remains;
4313
4314 if (flags & VMAP_RAM)
4315 copied = vmap_ram_vread_iter(iter, addr, n, flags);
4316 else if (!(vm && (vm->flags & (VM_IOREMAP | VM_SPARSE))))
4317 copied = aligned_vread_iter(iter, addr, n);
4318 else /* IOREMAP | SPARSE area is treated as memory hole */
4319 copied = zero_iter(iter, n);
4320
4321 addr += copied;
4322 remains -= copied;
4323
4324 if (copied != n)
4325 goto finished;
4326
4327 next_va:
4328 next = va->va_end;
4329 spin_unlock(&vn->busy.lock);
4330 } while ((vn = find_vmap_area_exceed_addr_lock(next, &va)));
4331
4332 finished_zero:
4333 if (vn)
4334 spin_unlock(&vn->busy.lock);
4335
4336 /* zero-fill memory holes */
4337 return count - remains + zero_iter(iter, remains);
4338 finished:
4339 /* Nothing remains, or We couldn't copy/zero everything. */
4340 if (vn)
4341 spin_unlock(&vn->busy.lock);
4342
4343 return count - remains;
4344 }
4345
4346 /**
4347 * remap_vmalloc_range_partial - map vmalloc pages to userspace
4348 * @vma: vma to cover
4349 * @uaddr: target user address to start at
4350 * @kaddr: virtual address of vmalloc kernel memory
4351 * @pgoff: offset from @kaddr to start at
4352 * @size: size of map area
4353 *
4354 * Returns: 0 for success, -Exxx on failure
4355 *
4356 * This function checks that @kaddr is a valid vmalloc'ed area,
4357 * and that it is big enough to cover the range starting at
4358 * @uaddr in @vma. Will return failure if that criteria isn't
4359 * met.
4360 *
4361 * Similar to remap_pfn_range() (see mm/memory.c)
4362 */
remap_vmalloc_range_partial(struct vm_area_struct * vma,unsigned long uaddr,void * kaddr,unsigned long pgoff,unsigned long size)4363 int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
4364 void *kaddr, unsigned long pgoff,
4365 unsigned long size)
4366 {
4367 struct vm_struct *area;
4368 unsigned long off;
4369 unsigned long end_index;
4370
4371 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
4372 return -EINVAL;
4373
4374 size = PAGE_ALIGN(size);
4375
4376 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
4377 return -EINVAL;
4378
4379 area = find_vm_area(kaddr);
4380 if (!area)
4381 return -EINVAL;
4382
4383 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
4384 return -EINVAL;
4385
4386 if (check_add_overflow(size, off, &end_index) ||
4387 end_index > get_vm_area_size(area))
4388 return -EINVAL;
4389 kaddr += off;
4390
4391 do {
4392 struct page *page = vmalloc_to_page(kaddr);
4393 int ret;
4394
4395 ret = vm_insert_page(vma, uaddr, page);
4396 if (ret)
4397 return ret;
4398
4399 uaddr += PAGE_SIZE;
4400 kaddr += PAGE_SIZE;
4401 size -= PAGE_SIZE;
4402 } while (size > 0);
4403
4404 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
4405
4406 return 0;
4407 }
4408
4409 /**
4410 * remap_vmalloc_range - map vmalloc pages to userspace
4411 * @vma: vma to cover (map full range of vma)
4412 * @addr: vmalloc memory
4413 * @pgoff: number of pages into addr before first page to map
4414 *
4415 * Returns: 0 for success, -Exxx on failure
4416 *
4417 * This function checks that addr is a valid vmalloc'ed area, and
4418 * that it is big enough to cover the vma. Will return failure if
4419 * that criteria isn't met.
4420 *
4421 * Similar to remap_pfn_range() (see mm/memory.c)
4422 */
remap_vmalloc_range(struct vm_area_struct * vma,void * addr,unsigned long pgoff)4423 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
4424 unsigned long pgoff)
4425 {
4426 return remap_vmalloc_range_partial(vma, vma->vm_start,
4427 addr, pgoff,
4428 vma->vm_end - vma->vm_start);
4429 }
4430 EXPORT_SYMBOL(remap_vmalloc_range);
4431
free_vm_area(struct vm_struct * area)4432 void free_vm_area(struct vm_struct *area)
4433 {
4434 struct vm_struct *ret;
4435 ret = remove_vm_area(area->addr);
4436 BUG_ON(ret != area);
4437 kfree(area);
4438 }
4439 EXPORT_SYMBOL_GPL(free_vm_area);
4440
4441 #ifdef CONFIG_SMP
node_to_va(struct rb_node * n)4442 static struct vmap_area *node_to_va(struct rb_node *n)
4443 {
4444 return rb_entry_safe(n, struct vmap_area, rb_node);
4445 }
4446
4447 /**
4448 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
4449 * @addr: target address
4450 *
4451 * Returns: vmap_area if it is found. If there is no such area
4452 * the first highest(reverse order) vmap_area is returned
4453 * i.e. va->va_start < addr && va->va_end < addr or NULL
4454 * if there are no any areas before @addr.
4455 */
4456 static struct vmap_area *
pvm_find_va_enclose_addr(unsigned long addr)4457 pvm_find_va_enclose_addr(unsigned long addr)
4458 {
4459 struct vmap_area *va, *tmp;
4460 struct rb_node *n;
4461
4462 n = free_vmap_area_root.rb_node;
4463 va = NULL;
4464
4465 while (n) {
4466 tmp = rb_entry(n, struct vmap_area, rb_node);
4467 if (tmp->va_start <= addr) {
4468 va = tmp;
4469 if (tmp->va_end >= addr)
4470 break;
4471
4472 n = n->rb_right;
4473 } else {
4474 n = n->rb_left;
4475 }
4476 }
4477
4478 return va;
4479 }
4480
4481 /**
4482 * pvm_determine_end_from_reverse - find the highest aligned address
4483 * of free block below VMALLOC_END
4484 * @va:
4485 * in - the VA we start the search(reverse order);
4486 * out - the VA with the highest aligned end address.
4487 * @align: alignment for required highest address
4488 *
4489 * Returns: determined end address within vmap_area
4490 */
4491 static unsigned long
pvm_determine_end_from_reverse(struct vmap_area ** va,unsigned long align)4492 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
4493 {
4494 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4495 unsigned long addr;
4496
4497 if (likely(*va)) {
4498 list_for_each_entry_from_reverse((*va),
4499 &free_vmap_area_list, list) {
4500 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
4501 if ((*va)->va_start < addr)
4502 return addr;
4503 }
4504 }
4505
4506 return 0;
4507 }
4508
4509 /**
4510 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4511 * @offsets: array containing offset of each area
4512 * @sizes: array containing size of each area
4513 * @nr_vms: the number of areas to allocate
4514 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4515 *
4516 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4517 * vm_structs on success, %NULL on failure
4518 *
4519 * Percpu allocator wants to use congruent vm areas so that it can
4520 * maintain the offsets among percpu areas. This function allocates
4521 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4522 * be scattered pretty far, distance between two areas easily going up
4523 * to gigabytes. To avoid interacting with regular vmallocs, these
4524 * areas are allocated from top.
4525 *
4526 * Despite its complicated look, this allocator is rather simple. It
4527 * does everything top-down and scans free blocks from the end looking
4528 * for matching base. While scanning, if any of the areas do not fit the
4529 * base address is pulled down to fit the area. Scanning is repeated till
4530 * all the areas fit and then all necessary data structures are inserted
4531 * and the result is returned.
4532 */
pcpu_get_vm_areas(const unsigned long * offsets,const size_t * sizes,int nr_vms,size_t align)4533 struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4534 const size_t *sizes, int nr_vms,
4535 size_t align)
4536 {
4537 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4538 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4539 struct vmap_area **vas, *va;
4540 struct vm_struct **vms;
4541 int area, area2, last_area, term_area;
4542 unsigned long base, start, size, end, last_end, orig_start, orig_end;
4543 bool purged = false;
4544
4545 /* verify parameters and allocate data structures */
4546 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4547 for (last_area = 0, area = 0; area < nr_vms; area++) {
4548 start = offsets[area];
4549 end = start + sizes[area];
4550
4551 /* is everything aligned properly? */
4552 BUG_ON(!IS_ALIGNED(offsets[area], align));
4553 BUG_ON(!IS_ALIGNED(sizes[area], align));
4554
4555 /* detect the area with the highest address */
4556 if (start > offsets[last_area])
4557 last_area = area;
4558
4559 for (area2 = area + 1; area2 < nr_vms; area2++) {
4560 unsigned long start2 = offsets[area2];
4561 unsigned long end2 = start2 + sizes[area2];
4562
4563 BUG_ON(start2 < end && start < end2);
4564 }
4565 }
4566 last_end = offsets[last_area] + sizes[last_area];
4567
4568 if (vmalloc_end - vmalloc_start < last_end) {
4569 WARN_ON(true);
4570 return NULL;
4571 }
4572
4573 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4574 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4575 if (!vas || !vms)
4576 goto err_free2;
4577
4578 for (area = 0; area < nr_vms; area++) {
4579 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4580 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4581 if (!vas[area] || !vms[area])
4582 goto err_free;
4583 }
4584 retry:
4585 spin_lock(&free_vmap_area_lock);
4586
4587 /* start scanning - we scan from the top, begin with the last area */
4588 area = term_area = last_area;
4589 start = offsets[area];
4590 end = start + sizes[area];
4591
4592 va = pvm_find_va_enclose_addr(vmalloc_end);
4593 base = pvm_determine_end_from_reverse(&va, align) - end;
4594
4595 while (true) {
4596 /*
4597 * base might have underflowed, add last_end before
4598 * comparing.
4599 */
4600 if (base + last_end < vmalloc_start + last_end)
4601 goto overflow;
4602
4603 /*
4604 * Fitting base has not been found.
4605 */
4606 if (va == NULL)
4607 goto overflow;
4608
4609 /*
4610 * If required width exceeds current VA block, move
4611 * base downwards and then recheck.
4612 */
4613 if (base + end > va->va_end) {
4614 base = pvm_determine_end_from_reverse(&va, align) - end;
4615 term_area = area;
4616 continue;
4617 }
4618
4619 /*
4620 * If this VA does not fit, move base downwards and recheck.
4621 */
4622 if (base + start < va->va_start) {
4623 va = node_to_va(rb_prev(&va->rb_node));
4624 base = pvm_determine_end_from_reverse(&va, align) - end;
4625 term_area = area;
4626 continue;
4627 }
4628
4629 /*
4630 * This area fits, move on to the previous one. If
4631 * the previous one is the terminal one, we're done.
4632 */
4633 area = (area + nr_vms - 1) % nr_vms;
4634 if (area == term_area)
4635 break;
4636
4637 start = offsets[area];
4638 end = start + sizes[area];
4639 va = pvm_find_va_enclose_addr(base + end);
4640 }
4641
4642 /* we've found a fitting base, insert all va's */
4643 for (area = 0; area < nr_vms; area++) {
4644 int ret;
4645
4646 start = base + offsets[area];
4647 size = sizes[area];
4648
4649 va = pvm_find_va_enclose_addr(start);
4650 if (WARN_ON_ONCE(va == NULL))
4651 /* It is a BUG(), but trigger recovery instead. */
4652 goto recovery;
4653
4654 ret = va_clip(&free_vmap_area_root,
4655 &free_vmap_area_list, va, start, size);
4656 if (WARN_ON_ONCE(unlikely(ret)))
4657 /* It is a BUG(), but trigger recovery instead. */
4658 goto recovery;
4659
4660 /* Allocated area. */
4661 va = vas[area];
4662 va->va_start = start;
4663 va->va_end = start + size;
4664 }
4665
4666 spin_unlock(&free_vmap_area_lock);
4667
4668 /* populate the kasan shadow space */
4669 for (area = 0; area < nr_vms; area++) {
4670 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4671 goto err_free_shadow;
4672 }
4673
4674 /* insert all vm's */
4675 for (area = 0; area < nr_vms; area++) {
4676 struct vmap_node *vn = addr_to_node(vas[area]->va_start);
4677
4678 spin_lock(&vn->busy.lock);
4679 insert_vmap_area(vas[area], &vn->busy.root, &vn->busy.head);
4680 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
4681 pcpu_get_vm_areas);
4682 spin_unlock(&vn->busy.lock);
4683 }
4684
4685 /*
4686 * Mark allocated areas as accessible. Do it now as a best-effort
4687 * approach, as they can be mapped outside of vmalloc code.
4688 * With hardware tag-based KASAN, marking is skipped for
4689 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4690 */
4691 for (area = 0; area < nr_vms; area++)
4692 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4693 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4694
4695 kfree(vas);
4696 return vms;
4697
4698 recovery:
4699 /*
4700 * Remove previously allocated areas. There is no
4701 * need in removing these areas from the busy tree,
4702 * because they are inserted only on the final step
4703 * and when pcpu_get_vm_areas() is success.
4704 */
4705 while (area--) {
4706 orig_start = vas[area]->va_start;
4707 orig_end = vas[area]->va_end;
4708 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4709 &free_vmap_area_list);
4710 if (va)
4711 kasan_release_vmalloc(orig_start, orig_end,
4712 va->va_start, va->va_end);
4713 vas[area] = NULL;
4714 }
4715
4716 overflow:
4717 spin_unlock(&free_vmap_area_lock);
4718 if (!purged) {
4719 reclaim_and_purge_vmap_areas();
4720 purged = true;
4721
4722 /* Before "retry", check if we recover. */
4723 for (area = 0; area < nr_vms; area++) {
4724 if (vas[area])
4725 continue;
4726
4727 vas[area] = kmem_cache_zalloc(
4728 vmap_area_cachep, GFP_KERNEL);
4729 if (!vas[area])
4730 goto err_free;
4731 }
4732
4733 goto retry;
4734 }
4735
4736 err_free:
4737 for (area = 0; area < nr_vms; area++) {
4738 if (vas[area])
4739 kmem_cache_free(vmap_area_cachep, vas[area]);
4740
4741 kfree(vms[area]);
4742 }
4743 err_free2:
4744 kfree(vas);
4745 kfree(vms);
4746 return NULL;
4747
4748 err_free_shadow:
4749 spin_lock(&free_vmap_area_lock);
4750 /*
4751 * We release all the vmalloc shadows, even the ones for regions that
4752 * hadn't been successfully added. This relies on kasan_release_vmalloc
4753 * being able to tolerate this case.
4754 */
4755 for (area = 0; area < nr_vms; area++) {
4756 orig_start = vas[area]->va_start;
4757 orig_end = vas[area]->va_end;
4758 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4759 &free_vmap_area_list);
4760 if (va)
4761 kasan_release_vmalloc(orig_start, orig_end,
4762 va->va_start, va->va_end);
4763 vas[area] = NULL;
4764 kfree(vms[area]);
4765 }
4766 spin_unlock(&free_vmap_area_lock);
4767 kfree(vas);
4768 kfree(vms);
4769 return NULL;
4770 }
4771
4772 /**
4773 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4774 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4775 * @nr_vms: the number of allocated areas
4776 *
4777 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4778 */
pcpu_free_vm_areas(struct vm_struct ** vms,int nr_vms)4779 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4780 {
4781 int i;
4782
4783 for (i = 0; i < nr_vms; i++)
4784 free_vm_area(vms[i]);
4785 kfree(vms);
4786 }
4787 #endif /* CONFIG_SMP */
4788
4789 #ifdef CONFIG_PRINTK
vmalloc_dump_obj(void * object)4790 bool vmalloc_dump_obj(void *object)
4791 {
4792 const void *caller;
4793 struct vm_struct *vm;
4794 struct vmap_area *va;
4795 struct vmap_node *vn;
4796 unsigned long addr;
4797 unsigned int nr_pages;
4798
4799 addr = PAGE_ALIGN((unsigned long) object);
4800 vn = addr_to_node(addr);
4801
4802 if (!spin_trylock(&vn->busy.lock))
4803 return false;
4804
4805 va = __find_vmap_area(addr, &vn->busy.root);
4806 if (!va || !va->vm) {
4807 spin_unlock(&vn->busy.lock);
4808 return false;
4809 }
4810
4811 vm = va->vm;
4812 addr = (unsigned long) vm->addr;
4813 caller = vm->caller;
4814 nr_pages = vm->nr_pages;
4815 spin_unlock(&vn->busy.lock);
4816
4817 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4818 nr_pages, addr, caller);
4819
4820 return true;
4821 }
4822 #endif
4823
4824 #ifdef CONFIG_PROC_FS
show_numa_info(struct seq_file * m,struct vm_struct * v)4825 static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4826 {
4827 if (IS_ENABLED(CONFIG_NUMA)) {
4828 unsigned int nr, *counters = m->private;
4829 unsigned int step = 1U << vm_area_page_order(v);
4830
4831 if (!counters)
4832 return;
4833
4834 if (v->flags & VM_UNINITIALIZED)
4835 return;
4836 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4837 smp_rmb();
4838
4839 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4840
4841 for (nr = 0; nr < v->nr_pages; nr += step)
4842 counters[page_to_nid(v->pages[nr])] += step;
4843 for_each_node_state(nr, N_HIGH_MEMORY)
4844 if (counters[nr])
4845 seq_printf(m, " N%u=%u", nr, counters[nr]);
4846 }
4847 }
4848
show_purge_info(struct seq_file * m)4849 static void show_purge_info(struct seq_file *m)
4850 {
4851 struct vmap_node *vn;
4852 struct vmap_area *va;
4853 int i;
4854
4855 for (i = 0; i < nr_vmap_nodes; i++) {
4856 vn = &vmap_nodes[i];
4857
4858 spin_lock(&vn->lazy.lock);
4859 list_for_each_entry(va, &vn->lazy.head, list) {
4860 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4861 (void *)va->va_start, (void *)va->va_end,
4862 va->va_end - va->va_start);
4863 }
4864 spin_unlock(&vn->lazy.lock);
4865 }
4866 }
4867
vmalloc_info_show(struct seq_file * m,void * p)4868 static int vmalloc_info_show(struct seq_file *m, void *p)
4869 {
4870 struct vmap_node *vn;
4871 struct vmap_area *va;
4872 struct vm_struct *v;
4873 int i;
4874
4875 for (i = 0; i < nr_vmap_nodes; i++) {
4876 vn = &vmap_nodes[i];
4877
4878 spin_lock(&vn->busy.lock);
4879 list_for_each_entry(va, &vn->busy.head, list) {
4880 if (!va->vm) {
4881 if (va->flags & VMAP_RAM)
4882 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4883 (void *)va->va_start, (void *)va->va_end,
4884 va->va_end - va->va_start);
4885
4886 continue;
4887 }
4888
4889 v = va->vm;
4890
4891 seq_printf(m, "0x%pK-0x%pK %7ld",
4892 v->addr, v->addr + v->size, v->size);
4893
4894 if (v->caller)
4895 seq_printf(m, " %pS", v->caller);
4896
4897 if (v->nr_pages)
4898 seq_printf(m, " pages=%d", v->nr_pages);
4899
4900 if (v->phys_addr)
4901 seq_printf(m, " phys=%pa", &v->phys_addr);
4902
4903 if (v->flags & VM_IOREMAP)
4904 seq_puts(m, " ioremap");
4905
4906 if (v->flags & VM_SPARSE)
4907 seq_puts(m, " sparse");
4908
4909 if (v->flags & VM_ALLOC)
4910 seq_puts(m, " vmalloc");
4911
4912 if (v->flags & VM_MAP)
4913 seq_puts(m, " vmap");
4914
4915 if (v->flags & VM_USERMAP)
4916 seq_puts(m, " user");
4917
4918 if (v->flags & VM_DMA_COHERENT)
4919 seq_puts(m, " dma-coherent");
4920
4921 if (is_vmalloc_addr(v->pages))
4922 seq_puts(m, " vpages");
4923
4924 show_numa_info(m, v);
4925 seq_putc(m, '\n');
4926 }
4927 spin_unlock(&vn->busy.lock);
4928 }
4929
4930 /*
4931 * As a final step, dump "unpurged" areas.
4932 */
4933 show_purge_info(m);
4934 return 0;
4935 }
4936
proc_vmalloc_init(void)4937 static int __init proc_vmalloc_init(void)
4938 {
4939 void *priv_data = NULL;
4940
4941 if (IS_ENABLED(CONFIG_NUMA))
4942 priv_data = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
4943
4944 proc_create_single_data("vmallocinfo",
4945 0400, NULL, vmalloc_info_show, priv_data);
4946
4947 return 0;
4948 }
4949 module_init(proc_vmalloc_init);
4950
4951 #endif
4952
vmap_init_free_space(void)4953 static void __init vmap_init_free_space(void)
4954 {
4955 unsigned long vmap_start = 1;
4956 const unsigned long vmap_end = ULONG_MAX;
4957 struct vmap_area *free;
4958 struct vm_struct *busy;
4959
4960 /*
4961 * B F B B B F
4962 * -|-----|.....|-----|-----|-----|.....|-
4963 * | The KVA space |
4964 * |<--------------------------------->|
4965 */
4966 for (busy = vmlist; busy; busy = busy->next) {
4967 if ((unsigned long) busy->addr - vmap_start > 0) {
4968 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4969 if (!WARN_ON_ONCE(!free)) {
4970 free->va_start = vmap_start;
4971 free->va_end = (unsigned long) busy->addr;
4972
4973 insert_vmap_area_augment(free, NULL,
4974 &free_vmap_area_root,
4975 &free_vmap_area_list);
4976 }
4977 }
4978
4979 vmap_start = (unsigned long) busy->addr + busy->size;
4980 }
4981
4982 if (vmap_end - vmap_start > 0) {
4983 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4984 if (!WARN_ON_ONCE(!free)) {
4985 free->va_start = vmap_start;
4986 free->va_end = vmap_end;
4987
4988 insert_vmap_area_augment(free, NULL,
4989 &free_vmap_area_root,
4990 &free_vmap_area_list);
4991 }
4992 }
4993 }
4994
vmap_init_nodes(void)4995 static void vmap_init_nodes(void)
4996 {
4997 struct vmap_node *vn;
4998 int i, n;
4999
5000 #if BITS_PER_LONG == 64
5001 /*
5002 * A high threshold of max nodes is fixed and bound to 128,
5003 * thus a scale factor is 1 for systems where number of cores
5004 * are less or equal to specified threshold.
5005 *
5006 * As for NUMA-aware notes. For bigger systems, for example
5007 * NUMA with multi-sockets, where we can end-up with thousands
5008 * of cores in total, a "sub-numa-clustering" should be added.
5009 *
5010 * In this case a NUMA domain is considered as a single entity
5011 * with dedicated sub-nodes in it which describe one group or
5012 * set of cores. Therefore a per-domain purging is supposed to
5013 * be added as well as a per-domain balancing.
5014 */
5015 n = clamp_t(unsigned int, num_possible_cpus(), 1, 128);
5016
5017 if (n > 1) {
5018 vn = kmalloc_array(n, sizeof(*vn), GFP_NOWAIT | __GFP_NOWARN);
5019 if (vn) {
5020 /* Node partition is 16 pages. */
5021 vmap_zone_size = (1 << 4) * PAGE_SIZE;
5022 nr_vmap_nodes = n;
5023 vmap_nodes = vn;
5024 } else {
5025 pr_err("Failed to allocate an array. Disable a node layer\n");
5026 }
5027 }
5028 #endif
5029
5030 for (n = 0; n < nr_vmap_nodes; n++) {
5031 vn = &vmap_nodes[n];
5032 vn->busy.root = RB_ROOT;
5033 INIT_LIST_HEAD(&vn->busy.head);
5034 spin_lock_init(&vn->busy.lock);
5035
5036 vn->lazy.root = RB_ROOT;
5037 INIT_LIST_HEAD(&vn->lazy.head);
5038 spin_lock_init(&vn->lazy.lock);
5039
5040 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
5041 INIT_LIST_HEAD(&vn->pool[i].head);
5042 WRITE_ONCE(vn->pool[i].len, 0);
5043 }
5044
5045 spin_lock_init(&vn->pool_lock);
5046 }
5047 }
5048
5049 static unsigned long
vmap_node_shrink_count(struct shrinker * shrink,struct shrink_control * sc)5050 vmap_node_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5051 {
5052 unsigned long count;
5053 struct vmap_node *vn;
5054 int i, j;
5055
5056 for (count = 0, i = 0; i < nr_vmap_nodes; i++) {
5057 vn = &vmap_nodes[i];
5058
5059 for (j = 0; j < MAX_VA_SIZE_PAGES; j++)
5060 count += READ_ONCE(vn->pool[j].len);
5061 }
5062
5063 return count ? count : SHRINK_EMPTY;
5064 }
5065
5066 static unsigned long
vmap_node_shrink_scan(struct shrinker * shrink,struct shrink_control * sc)5067 vmap_node_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5068 {
5069 int i;
5070
5071 for (i = 0; i < nr_vmap_nodes; i++)
5072 decay_va_pool_node(&vmap_nodes[i], true);
5073
5074 return SHRINK_STOP;
5075 }
5076
vmalloc_init(void)5077 void __init vmalloc_init(void)
5078 {
5079 struct shrinker *vmap_node_shrinker;
5080 struct vmap_area *va;
5081 struct vmap_node *vn;
5082 struct vm_struct *tmp;
5083 int i;
5084
5085 /*
5086 * Create the cache for vmap_area objects.
5087 */
5088 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
5089
5090 for_each_possible_cpu(i) {
5091 struct vmap_block_queue *vbq;
5092 struct vfree_deferred *p;
5093
5094 vbq = &per_cpu(vmap_block_queue, i);
5095 spin_lock_init(&vbq->lock);
5096 INIT_LIST_HEAD(&vbq->free);
5097 p = &per_cpu(vfree_deferred, i);
5098 init_llist_head(&p->list);
5099 INIT_WORK(&p->wq, delayed_vfree_work);
5100 xa_init(&vbq->vmap_blocks);
5101 }
5102
5103 /*
5104 * Setup nodes before importing vmlist.
5105 */
5106 vmap_init_nodes();
5107
5108 /* Import existing vmlist entries. */
5109 for (tmp = vmlist; tmp; tmp = tmp->next) {
5110 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5111 if (WARN_ON_ONCE(!va))
5112 continue;
5113
5114 va->va_start = (unsigned long)tmp->addr;
5115 va->va_end = va->va_start + tmp->size;
5116 va->vm = tmp;
5117
5118 vn = addr_to_node(va->va_start);
5119 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
5120 }
5121
5122 /*
5123 * Now we can initialize a free vmap space.
5124 */
5125 vmap_init_free_space();
5126 vmap_initialized = true;
5127
5128 vmap_node_shrinker = shrinker_alloc(0, "vmap-node");
5129 if (!vmap_node_shrinker) {
5130 pr_err("Failed to allocate vmap-node shrinker!\n");
5131 return;
5132 }
5133
5134 vmap_node_shrinker->count_objects = vmap_node_shrink_count;
5135 vmap_node_shrinker->scan_objects = vmap_node_shrink_scan;
5136 shrinker_register(vmap_node_shrinker);
5137 }
5138