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