Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 1993 Linus Torvalds
4 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
5 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
6 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
7 * Numa awareness, Christoph Lameter, SGI, June 2005
8 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
9 */
10
11#include <linux/vmalloc.h>
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/highmem.h>
15#include <linux/sched/signal.h>
16#include <linux/slab.h>
17#include <linux/spinlock.h>
18#include <linux/interrupt.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/set_memory.h>
22#include <linux/debugobjects.h>
23#include <linux/kallsyms.h>
24#include <linux/list.h>
25#include <linux/notifier.h>
26#include <linux/rbtree.h>
27#include <linux/xarray.h>
28#include <linux/io.h>
29#include <linux/rcupdate.h>
30#include <linux/pfn.h>
31#include <linux/kmemleak.h>
32#include <linux/atomic.h>
33#include <linux/compiler.h>
34#include <linux/memcontrol.h>
35#include <linux/llist.h>
36#include <linux/uio.h>
37#include <linux/bitops.h>
38#include <linux/rbtree_augmented.h>
39#include <linux/overflow.h>
40#include <linux/pgtable.h>
41#include <linux/hugetlb.h>
42#include <linux/sched/mm.h>
43#include <asm/tlbflush.h>
44#include <asm/shmparam.h>
45
46#define CREATE_TRACE_POINTS
47#include <trace/events/vmalloc.h>
48
49#include "internal.h"
50#include "pgalloc-track.h"
51
52#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
53static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
54
55static int __init set_nohugeiomap(char *str)
56{
57 ioremap_max_page_shift = PAGE_SHIFT;
58 return 0;
59}
60early_param("nohugeiomap", set_nohugeiomap);
61#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
62static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
63#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
64
65#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
66static bool __ro_after_init vmap_allow_huge = true;
67
68static int __init set_nohugevmalloc(char *str)
69{
70 vmap_allow_huge = false;
71 return 0;
72}
73early_param("nohugevmalloc", set_nohugevmalloc);
74#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
75static const bool vmap_allow_huge = false;
76#endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
77
78bool is_vmalloc_addr(const void *x)
79{
80 unsigned long addr = (unsigned long)kasan_reset_tag(x);
81
82 return addr >= VMALLOC_START && addr < VMALLOC_END;
83}
84EXPORT_SYMBOL(is_vmalloc_addr);
85
86struct vfree_deferred {
87 struct llist_head list;
88 struct work_struct wq;
89};
90static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
91
92/*** Page table manipulation functions ***/
93static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
94 phys_addr_t phys_addr, pgprot_t prot,
95 unsigned int max_page_shift, pgtbl_mod_mask *mask)
96{
97 pte_t *pte;
98 u64 pfn;
99 unsigned long size = PAGE_SIZE;
100
101 pfn = phys_addr >> PAGE_SHIFT;
102 pte = pte_alloc_kernel_track(pmd, addr, mask);
103 if (!pte)
104 return -ENOMEM;
105 do {
106 BUG_ON(!pte_none(ptep_get(pte)));
107
108#ifdef CONFIG_HUGETLB_PAGE
109 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
110 if (size != PAGE_SIZE) {
111 pte_t entry = pfn_pte(pfn, prot);
112
113 entry = arch_make_huge_pte(entry, ilog2(size), 0);
114 set_huge_pte_at(&init_mm, addr, pte, entry, size);
115 pfn += PFN_DOWN(size);
116 continue;
117 }
118#endif
119 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
120 pfn++;
121 } while (pte += PFN_DOWN(size), addr += size, addr != end);
122 *mask |= PGTBL_PTE_MODIFIED;
123 return 0;
124}
125
126static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
127 phys_addr_t phys_addr, pgprot_t prot,
128 unsigned int max_page_shift)
129{
130 if (max_page_shift < PMD_SHIFT)
131 return 0;
132
133 if (!arch_vmap_pmd_supported(prot))
134 return 0;
135
136 if ((end - addr) != PMD_SIZE)
137 return 0;
138
139 if (!IS_ALIGNED(addr, PMD_SIZE))
140 return 0;
141
142 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
143 return 0;
144
145 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
146 return 0;
147
148 return pmd_set_huge(pmd, phys_addr, prot);
149}
150
151static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
152 phys_addr_t phys_addr, pgprot_t prot,
153 unsigned int max_page_shift, pgtbl_mod_mask *mask)
154{
155 pmd_t *pmd;
156 unsigned long next;
157
158 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
159 if (!pmd)
160 return -ENOMEM;
161 do {
162 next = pmd_addr_end(addr, end);
163
164 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
165 max_page_shift)) {
166 *mask |= PGTBL_PMD_MODIFIED;
167 continue;
168 }
169
170 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
171 return -ENOMEM;
172 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
173 return 0;
174}
175
176static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
177 phys_addr_t phys_addr, pgprot_t prot,
178 unsigned int max_page_shift)
179{
180 if (max_page_shift < PUD_SHIFT)
181 return 0;
182
183 if (!arch_vmap_pud_supported(prot))
184 return 0;
185
186 if ((end - addr) != PUD_SIZE)
187 return 0;
188
189 if (!IS_ALIGNED(addr, PUD_SIZE))
190 return 0;
191
192 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
193 return 0;
194
195 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
196 return 0;
197
198 return pud_set_huge(pud, phys_addr, prot);
199}
200
201static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
202 phys_addr_t phys_addr, pgprot_t prot,
203 unsigned int max_page_shift, pgtbl_mod_mask *mask)
204{
205 pud_t *pud;
206 unsigned long next;
207
208 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
209 if (!pud)
210 return -ENOMEM;
211 do {
212 next = pud_addr_end(addr, end);
213
214 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
215 max_page_shift)) {
216 *mask |= PGTBL_PUD_MODIFIED;
217 continue;
218 }
219
220 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
221 max_page_shift, mask))
222 return -ENOMEM;
223 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
224 return 0;
225}
226
227static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
228 phys_addr_t phys_addr, pgprot_t prot,
229 unsigned int max_page_shift)
230{
231 if (max_page_shift < P4D_SHIFT)
232 return 0;
233
234 if (!arch_vmap_p4d_supported(prot))
235 return 0;
236
237 if ((end - addr) != P4D_SIZE)
238 return 0;
239
240 if (!IS_ALIGNED(addr, P4D_SIZE))
241 return 0;
242
243 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
244 return 0;
245
246 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
247 return 0;
248
249 return p4d_set_huge(p4d, phys_addr, prot);
250}
251
252static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
253 phys_addr_t phys_addr, pgprot_t prot,
254 unsigned int max_page_shift, pgtbl_mod_mask *mask)
255{
256 p4d_t *p4d;
257 unsigned long next;
258
259 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
260 if (!p4d)
261 return -ENOMEM;
262 do {
263 next = p4d_addr_end(addr, end);
264
265 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
266 max_page_shift)) {
267 *mask |= PGTBL_P4D_MODIFIED;
268 continue;
269 }
270
271 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
272 max_page_shift, mask))
273 return -ENOMEM;
274 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
275 return 0;
276}
277
278static int vmap_range_noflush(unsigned long addr, unsigned long end,
279 phys_addr_t phys_addr, pgprot_t prot,
280 unsigned int max_page_shift)
281{
282 pgd_t *pgd;
283 unsigned long start;
284 unsigned long next;
285 int err;
286 pgtbl_mod_mask mask = 0;
287
288 might_sleep();
289 BUG_ON(addr >= end);
290
291 start = addr;
292 pgd = pgd_offset_k(addr);
293 do {
294 next = pgd_addr_end(addr, end);
295 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
296 max_page_shift, &mask);
297 if (err)
298 break;
299 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
300
301 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
302 arch_sync_kernel_mappings(start, end);
303
304 return err;
305}
306
307int ioremap_page_range(unsigned long addr, unsigned long end,
308 phys_addr_t phys_addr, pgprot_t prot)
309{
310 int err;
311
312 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
313 ioremap_max_page_shift);
314 flush_cache_vmap(addr, end);
315 if (!err)
316 err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
317 ioremap_max_page_shift);
318 return err;
319}
320
321static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
322 pgtbl_mod_mask *mask)
323{
324 pte_t *pte;
325
326 pte = pte_offset_kernel(pmd, addr);
327 do {
328 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
329 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
330 } while (pte++, addr += PAGE_SIZE, addr != end);
331 *mask |= PGTBL_PTE_MODIFIED;
332}
333
334static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
335 pgtbl_mod_mask *mask)
336{
337 pmd_t *pmd;
338 unsigned long next;
339 int cleared;
340
341 pmd = pmd_offset(pud, addr);
342 do {
343 next = pmd_addr_end(addr, end);
344
345 cleared = pmd_clear_huge(pmd);
346 if (cleared || pmd_bad(*pmd))
347 *mask |= PGTBL_PMD_MODIFIED;
348
349 if (cleared)
350 continue;
351 if (pmd_none_or_clear_bad(pmd))
352 continue;
353 vunmap_pte_range(pmd, addr, next, mask);
354
355 cond_resched();
356 } while (pmd++, addr = next, addr != end);
357}
358
359static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
360 pgtbl_mod_mask *mask)
361{
362 pud_t *pud;
363 unsigned long next;
364 int cleared;
365
366 pud = pud_offset(p4d, addr);
367 do {
368 next = pud_addr_end(addr, end);
369
370 cleared = pud_clear_huge(pud);
371 if (cleared || pud_bad(*pud))
372 *mask |= PGTBL_PUD_MODIFIED;
373
374 if (cleared)
375 continue;
376 if (pud_none_or_clear_bad(pud))
377 continue;
378 vunmap_pmd_range(pud, addr, next, mask);
379 } while (pud++, addr = next, addr != end);
380}
381
382static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
383 pgtbl_mod_mask *mask)
384{
385 p4d_t *p4d;
386 unsigned long next;
387
388 p4d = p4d_offset(pgd, addr);
389 do {
390 next = p4d_addr_end(addr, end);
391
392 p4d_clear_huge(p4d);
393 if (p4d_bad(*p4d))
394 *mask |= PGTBL_P4D_MODIFIED;
395
396 if (p4d_none_or_clear_bad(p4d))
397 continue;
398 vunmap_pud_range(p4d, addr, next, mask);
399 } while (p4d++, addr = next, addr != end);
400}
401
402/*
403 * vunmap_range_noflush is similar to vunmap_range, but does not
404 * flush caches or TLBs.
405 *
406 * The caller is responsible for calling flush_cache_vmap() before calling
407 * this function, and flush_tlb_kernel_range after it has returned
408 * successfully (and before the addresses are expected to cause a page fault
409 * or be re-mapped for something else, if TLB flushes are being delayed or
410 * coalesced).
411 *
412 * This is an internal function only. Do not use outside mm/.
413 */
414void __vunmap_range_noflush(unsigned long start, unsigned long end)
415{
416 unsigned long next;
417 pgd_t *pgd;
418 unsigned long addr = start;
419 pgtbl_mod_mask mask = 0;
420
421 BUG_ON(addr >= end);
422 pgd = pgd_offset_k(addr);
423 do {
424 next = pgd_addr_end(addr, end);
425 if (pgd_bad(*pgd))
426 mask |= PGTBL_PGD_MODIFIED;
427 if (pgd_none_or_clear_bad(pgd))
428 continue;
429 vunmap_p4d_range(pgd, addr, next, &mask);
430 } while (pgd++, addr = next, addr != end);
431
432 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
433 arch_sync_kernel_mappings(start, end);
434}
435
436void vunmap_range_noflush(unsigned long start, unsigned long end)
437{
438 kmsan_vunmap_range_noflush(start, end);
439 __vunmap_range_noflush(start, end);
440}
441
442/**
443 * vunmap_range - unmap kernel virtual addresses
444 * @addr: start of the VM area to unmap
445 * @end: end of the VM area to unmap (non-inclusive)
446 *
447 * Clears any present PTEs in the virtual address range, flushes TLBs and
448 * caches. Any subsequent access to the address before it has been re-mapped
449 * is a kernel bug.
450 */
451void vunmap_range(unsigned long addr, unsigned long end)
452{
453 flush_cache_vunmap(addr, end);
454 vunmap_range_noflush(addr, end);
455 flush_tlb_kernel_range(addr, end);
456}
457
458static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
459 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
460 pgtbl_mod_mask *mask)
461{
462 pte_t *pte;
463
464 /*
465 * nr is a running index into the array which helps higher level
466 * callers keep track of where we're up to.
467 */
468
469 pte = pte_alloc_kernel_track(pmd, addr, mask);
470 if (!pte)
471 return -ENOMEM;
472 do {
473 struct page *page = pages[*nr];
474
475 if (WARN_ON(!pte_none(ptep_get(pte))))
476 return -EBUSY;
477 if (WARN_ON(!page))
478 return -ENOMEM;
479 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
480 return -EINVAL;
481
482 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
483 (*nr)++;
484 } while (pte++, addr += PAGE_SIZE, addr != end);
485 *mask |= PGTBL_PTE_MODIFIED;
486 return 0;
487}
488
489static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
490 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
491 pgtbl_mod_mask *mask)
492{
493 pmd_t *pmd;
494 unsigned long next;
495
496 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
497 if (!pmd)
498 return -ENOMEM;
499 do {
500 next = pmd_addr_end(addr, end);
501 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
502 return -ENOMEM;
503 } while (pmd++, addr = next, addr != end);
504 return 0;
505}
506
507static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
508 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
509 pgtbl_mod_mask *mask)
510{
511 pud_t *pud;
512 unsigned long next;
513
514 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
515 if (!pud)
516 return -ENOMEM;
517 do {
518 next = pud_addr_end(addr, end);
519 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
520 return -ENOMEM;
521 } while (pud++, addr = next, addr != end);
522 return 0;
523}
524
525static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
526 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
527 pgtbl_mod_mask *mask)
528{
529 p4d_t *p4d;
530 unsigned long next;
531
532 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
533 if (!p4d)
534 return -ENOMEM;
535 do {
536 next = p4d_addr_end(addr, end);
537 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
538 return -ENOMEM;
539 } while (p4d++, addr = next, addr != end);
540 return 0;
541}
542
543static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
544 pgprot_t prot, struct page **pages)
545{
546 unsigned long start = addr;
547 pgd_t *pgd;
548 unsigned long next;
549 int err = 0;
550 int nr = 0;
551 pgtbl_mod_mask mask = 0;
552
553 BUG_ON(addr >= end);
554 pgd = pgd_offset_k(addr);
555 do {
556 next = pgd_addr_end(addr, end);
557 if (pgd_bad(*pgd))
558 mask |= PGTBL_PGD_MODIFIED;
559 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
560 if (err)
561 return err;
562 } while (pgd++, addr = next, addr != end);
563
564 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
565 arch_sync_kernel_mappings(start, end);
566
567 return 0;
568}
569
570/*
571 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
572 * flush caches.
573 *
574 * The caller is responsible for calling flush_cache_vmap() after this
575 * function returns successfully and before the addresses are accessed.
576 *
577 * This is an internal function only. Do not use outside mm/.
578 */
579int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
580 pgprot_t prot, struct page **pages, unsigned int page_shift)
581{
582 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
583
584 WARN_ON(page_shift < PAGE_SHIFT);
585
586 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
587 page_shift == PAGE_SHIFT)
588 return vmap_small_pages_range_noflush(addr, end, prot, pages);
589
590 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
591 int err;
592
593 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
594 page_to_phys(pages[i]), prot,
595 page_shift);
596 if (err)
597 return err;
598
599 addr += 1UL << page_shift;
600 }
601
602 return 0;
603}
604
605int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
606 pgprot_t prot, struct page **pages, unsigned int page_shift)
607{
608 int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
609 page_shift);
610
611 if (ret)
612 return ret;
613 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
614}
615
616/**
617 * vmap_pages_range - map pages to a kernel virtual address
618 * @addr: start of the VM area to map
619 * @end: end of the VM area to map (non-inclusive)
620 * @prot: page protection flags to use
621 * @pages: pages to map (always PAGE_SIZE pages)
622 * @page_shift: maximum shift that the pages may be mapped with, @pages must
623 * be aligned and contiguous up to at least this shift.
624 *
625 * RETURNS:
626 * 0 on success, -errno on failure.
627 */
628static int vmap_pages_range(unsigned long addr, unsigned long end,
629 pgprot_t prot, struct page **pages, unsigned int page_shift)
630{
631 int err;
632
633 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
634 flush_cache_vmap(addr, end);
635 return err;
636}
637
638int is_vmalloc_or_module_addr(const void *x)
639{
640 /*
641 * ARM, x86-64 and sparc64 put modules in a special place,
642 * and fall back on vmalloc() if that fails. Others
643 * just put it in the vmalloc space.
644 */
645#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
646 unsigned long addr = (unsigned long)kasan_reset_tag(x);
647 if (addr >= MODULES_VADDR && addr < MODULES_END)
648 return 1;
649#endif
650 return is_vmalloc_addr(x);
651}
652EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
653
654/*
655 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
656 * return the tail page that corresponds to the base page address, which
657 * matches small vmap mappings.
658 */
659struct page *vmalloc_to_page(const void *vmalloc_addr)
660{
661 unsigned long addr = (unsigned long) vmalloc_addr;
662 struct page *page = NULL;
663 pgd_t *pgd = pgd_offset_k(addr);
664 p4d_t *p4d;
665 pud_t *pud;
666 pmd_t *pmd;
667 pte_t *ptep, pte;
668
669 /*
670 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
671 * architectures that do not vmalloc module space
672 */
673 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
674
675 if (pgd_none(*pgd))
676 return NULL;
677 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
678 return NULL; /* XXX: no allowance for huge pgd */
679 if (WARN_ON_ONCE(pgd_bad(*pgd)))
680 return NULL;
681
682 p4d = p4d_offset(pgd, addr);
683 if (p4d_none(*p4d))
684 return NULL;
685 if (p4d_leaf(*p4d))
686 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
687 if (WARN_ON_ONCE(p4d_bad(*p4d)))
688 return NULL;
689
690 pud = pud_offset(p4d, addr);
691 if (pud_none(*pud))
692 return NULL;
693 if (pud_leaf(*pud))
694 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
695 if (WARN_ON_ONCE(pud_bad(*pud)))
696 return NULL;
697
698 pmd = pmd_offset(pud, addr);
699 if (pmd_none(*pmd))
700 return NULL;
701 if (pmd_leaf(*pmd))
702 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
703 if (WARN_ON_ONCE(pmd_bad(*pmd)))
704 return NULL;
705
706 ptep = pte_offset_kernel(pmd, addr);
707 pte = ptep_get(ptep);
708 if (pte_present(pte))
709 page = pte_page(pte);
710
711 return page;
712}
713EXPORT_SYMBOL(vmalloc_to_page);
714
715/*
716 * Map a vmalloc()-space virtual address to the physical page frame number.
717 */
718unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
719{
720 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
721}
722EXPORT_SYMBOL(vmalloc_to_pfn);
723
724
725/*** Global kva allocator ***/
726
727#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
728#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
729
730
731static DEFINE_SPINLOCK(vmap_area_lock);
732static DEFINE_SPINLOCK(free_vmap_area_lock);
733/* Export for kexec only */
734LIST_HEAD(vmap_area_list);
735static struct rb_root vmap_area_root = RB_ROOT;
736static bool vmap_initialized __read_mostly;
737
738static struct rb_root purge_vmap_area_root = RB_ROOT;
739static LIST_HEAD(purge_vmap_area_list);
740static DEFINE_SPINLOCK(purge_vmap_area_lock);
741
742/*
743 * This kmem_cache is used for vmap_area objects. Instead of
744 * allocating from slab we reuse an object from this cache to
745 * make things faster. Especially in "no edge" splitting of
746 * free block.
747 */
748static struct kmem_cache *vmap_area_cachep;
749
750/*
751 * This linked list is used in pair with free_vmap_area_root.
752 * It gives O(1) access to prev/next to perform fast coalescing.
753 */
754static LIST_HEAD(free_vmap_area_list);
755
756/*
757 * This augment red-black tree represents the free vmap space.
758 * All vmap_area objects in this tree are sorted by va->va_start
759 * address. It is used for allocation and merging when a vmap
760 * object is released.
761 *
762 * Each vmap_area node contains a maximum available free block
763 * of its sub-tree, right or left. Therefore it is possible to
764 * find a lowest match of free area.
765 */
766static struct rb_root free_vmap_area_root = RB_ROOT;
767
768/*
769 * Preload a CPU with one object for "no edge" split case. The
770 * aim is to get rid of allocations from the atomic context, thus
771 * to use more permissive allocation masks.
772 */
773static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
774
775static __always_inline unsigned long
776va_size(struct vmap_area *va)
777{
778 return (va->va_end - va->va_start);
779}
780
781static __always_inline unsigned long
782get_subtree_max_size(struct rb_node *node)
783{
784 struct vmap_area *va;
785
786 va = rb_entry_safe(node, struct vmap_area, rb_node);
787 return va ? va->subtree_max_size : 0;
788}
789
790RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
791 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
792
793static void reclaim_and_purge_vmap_areas(void);
794static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
795static void drain_vmap_area_work(struct work_struct *work);
796static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
797
798static atomic_long_t nr_vmalloc_pages;
799
800unsigned long vmalloc_nr_pages(void)
801{
802 return atomic_long_read(&nr_vmalloc_pages);
803}
804
805/* Look up the first VA which satisfies addr < va_end, NULL if none. */
806static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
807{
808 struct vmap_area *va = NULL;
809 struct rb_node *n = vmap_area_root.rb_node;
810
811 addr = (unsigned long)kasan_reset_tag((void *)addr);
812
813 while (n) {
814 struct vmap_area *tmp;
815
816 tmp = rb_entry(n, struct vmap_area, rb_node);
817 if (tmp->va_end > addr) {
818 va = tmp;
819 if (tmp->va_start <= addr)
820 break;
821
822 n = n->rb_left;
823 } else
824 n = n->rb_right;
825 }
826
827 return va;
828}
829
830static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
831{
832 struct rb_node *n = root->rb_node;
833
834 addr = (unsigned long)kasan_reset_tag((void *)addr);
835
836 while (n) {
837 struct vmap_area *va;
838
839 va = rb_entry(n, struct vmap_area, rb_node);
840 if (addr < va->va_start)
841 n = n->rb_left;
842 else if (addr >= va->va_end)
843 n = n->rb_right;
844 else
845 return va;
846 }
847
848 return NULL;
849}
850
851/*
852 * This function returns back addresses of parent node
853 * and its left or right link for further processing.
854 *
855 * Otherwise NULL is returned. In that case all further
856 * steps regarding inserting of conflicting overlap range
857 * have to be declined and actually considered as a bug.
858 */
859static __always_inline struct rb_node **
860find_va_links(struct vmap_area *va,
861 struct rb_root *root, struct rb_node *from,
862 struct rb_node **parent)
863{
864 struct vmap_area *tmp_va;
865 struct rb_node **link;
866
867 if (root) {
868 link = &root->rb_node;
869 if (unlikely(!*link)) {
870 *parent = NULL;
871 return link;
872 }
873 } else {
874 link = &from;
875 }
876
877 /*
878 * Go to the bottom of the tree. When we hit the last point
879 * we end up with parent rb_node and correct direction, i name
880 * it link, where the new va->rb_node will be attached to.
881 */
882 do {
883 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
884
885 /*
886 * During the traversal we also do some sanity check.
887 * Trigger the BUG() if there are sides(left/right)
888 * or full overlaps.
889 */
890 if (va->va_end <= tmp_va->va_start)
891 link = &(*link)->rb_left;
892 else if (va->va_start >= tmp_va->va_end)
893 link = &(*link)->rb_right;
894 else {
895 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
896 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
897
898 return NULL;
899 }
900 } while (*link);
901
902 *parent = &tmp_va->rb_node;
903 return link;
904}
905
906static __always_inline struct list_head *
907get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
908{
909 struct list_head *list;
910
911 if (unlikely(!parent))
912 /*
913 * The red-black tree where we try to find VA neighbors
914 * before merging or inserting is empty, i.e. it means
915 * there is no free vmap space. Normally it does not
916 * happen but we handle this case anyway.
917 */
918 return NULL;
919
920 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
921 return (&parent->rb_right == link ? list->next : list);
922}
923
924static __always_inline void
925__link_va(struct vmap_area *va, struct rb_root *root,
926 struct rb_node *parent, struct rb_node **link,
927 struct list_head *head, bool augment)
928{
929 /*
930 * VA is still not in the list, but we can
931 * identify its future previous list_head node.
932 */
933 if (likely(parent)) {
934 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
935 if (&parent->rb_right != link)
936 head = head->prev;
937 }
938
939 /* Insert to the rb-tree */
940 rb_link_node(&va->rb_node, parent, link);
941 if (augment) {
942 /*
943 * Some explanation here. Just perform simple insertion
944 * to the tree. We do not set va->subtree_max_size to
945 * its current size before calling rb_insert_augmented().
946 * It is because we populate the tree from the bottom
947 * to parent levels when the node _is_ in the tree.
948 *
949 * Therefore we set subtree_max_size to zero after insertion,
950 * to let __augment_tree_propagate_from() puts everything to
951 * the correct order later on.
952 */
953 rb_insert_augmented(&va->rb_node,
954 root, &free_vmap_area_rb_augment_cb);
955 va->subtree_max_size = 0;
956 } else {
957 rb_insert_color(&va->rb_node, root);
958 }
959
960 /* Address-sort this list */
961 list_add(&va->list, head);
962}
963
964static __always_inline void
965link_va(struct vmap_area *va, struct rb_root *root,
966 struct rb_node *parent, struct rb_node **link,
967 struct list_head *head)
968{
969 __link_va(va, root, parent, link, head, false);
970}
971
972static __always_inline void
973link_va_augment(struct vmap_area *va, struct rb_root *root,
974 struct rb_node *parent, struct rb_node **link,
975 struct list_head *head)
976{
977 __link_va(va, root, parent, link, head, true);
978}
979
980static __always_inline void
981__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
982{
983 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
984 return;
985
986 if (augment)
987 rb_erase_augmented(&va->rb_node,
988 root, &free_vmap_area_rb_augment_cb);
989 else
990 rb_erase(&va->rb_node, root);
991
992 list_del_init(&va->list);
993 RB_CLEAR_NODE(&va->rb_node);
994}
995
996static __always_inline void
997unlink_va(struct vmap_area *va, struct rb_root *root)
998{
999 __unlink_va(va, root, false);
1000}
1001
1002static __always_inline void
1003unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1004{
1005 __unlink_va(va, root, true);
1006}
1007
1008#if DEBUG_AUGMENT_PROPAGATE_CHECK
1009/*
1010 * Gets called when remove the node and rotate.
1011 */
1012static __always_inline unsigned long
1013compute_subtree_max_size(struct vmap_area *va)
1014{
1015 return max3(va_size(va),
1016 get_subtree_max_size(va->rb_node.rb_left),
1017 get_subtree_max_size(va->rb_node.rb_right));
1018}
1019
1020static void
1021augment_tree_propagate_check(void)
1022{
1023 struct vmap_area *va;
1024 unsigned long computed_size;
1025
1026 list_for_each_entry(va, &free_vmap_area_list, list) {
1027 computed_size = compute_subtree_max_size(va);
1028 if (computed_size != va->subtree_max_size)
1029 pr_emerg("tree is corrupted: %lu, %lu\n",
1030 va_size(va), va->subtree_max_size);
1031 }
1032}
1033#endif
1034
1035/*
1036 * This function populates subtree_max_size from bottom to upper
1037 * levels starting from VA point. The propagation must be done
1038 * when VA size is modified by changing its va_start/va_end. Or
1039 * in case of newly inserting of VA to the tree.
1040 *
1041 * It means that __augment_tree_propagate_from() must be called:
1042 * - After VA has been inserted to the tree(free path);
1043 * - After VA has been shrunk(allocation path);
1044 * - After VA has been increased(merging path).
1045 *
1046 * Please note that, it does not mean that upper parent nodes
1047 * and their subtree_max_size are recalculated all the time up
1048 * to the root node.
1049 *
1050 * 4--8
1051 * /\
1052 * / \
1053 * / \
1054 * 2--2 8--8
1055 *
1056 * For example if we modify the node 4, shrinking it to 2, then
1057 * no any modification is required. If we shrink the node 2 to 1
1058 * its subtree_max_size is updated only, and set to 1. If we shrink
1059 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1060 * node becomes 4--6.
1061 */
1062static __always_inline void
1063augment_tree_propagate_from(struct vmap_area *va)
1064{
1065 /*
1066 * Populate the tree from bottom towards the root until
1067 * the calculated maximum available size of checked node
1068 * is equal to its current one.
1069 */
1070 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1071
1072#if DEBUG_AUGMENT_PROPAGATE_CHECK
1073 augment_tree_propagate_check();
1074#endif
1075}
1076
1077static void
1078insert_vmap_area(struct vmap_area *va,
1079 struct rb_root *root, struct list_head *head)
1080{
1081 struct rb_node **link;
1082 struct rb_node *parent;
1083
1084 link = find_va_links(va, root, NULL, &parent);
1085 if (link)
1086 link_va(va, root, parent, link, head);
1087}
1088
1089static void
1090insert_vmap_area_augment(struct vmap_area *va,
1091 struct rb_node *from, struct rb_root *root,
1092 struct list_head *head)
1093{
1094 struct rb_node **link;
1095 struct rb_node *parent;
1096
1097 if (from)
1098 link = find_va_links(va, NULL, from, &parent);
1099 else
1100 link = find_va_links(va, root, NULL, &parent);
1101
1102 if (link) {
1103 link_va_augment(va, root, parent, link, head);
1104 augment_tree_propagate_from(va);
1105 }
1106}
1107
1108/*
1109 * Merge de-allocated chunk of VA memory with previous
1110 * and next free blocks. If coalesce is not done a new
1111 * free area is inserted. If VA has been merged, it is
1112 * freed.
1113 *
1114 * Please note, it can return NULL in case of overlap
1115 * ranges, followed by WARN() report. Despite it is a
1116 * buggy behaviour, a system can be alive and keep
1117 * ongoing.
1118 */
1119static __always_inline struct vmap_area *
1120__merge_or_add_vmap_area(struct vmap_area *va,
1121 struct rb_root *root, struct list_head *head, bool augment)
1122{
1123 struct vmap_area *sibling;
1124 struct list_head *next;
1125 struct rb_node **link;
1126 struct rb_node *parent;
1127 bool merged = false;
1128
1129 /*
1130 * Find a place in the tree where VA potentially will be
1131 * inserted, unless it is merged with its sibling/siblings.
1132 */
1133 link = find_va_links(va, root, NULL, &parent);
1134 if (!link)
1135 return NULL;
1136
1137 /*
1138 * Get next node of VA to check if merging can be done.
1139 */
1140 next = get_va_next_sibling(parent, link);
1141 if (unlikely(next == NULL))
1142 goto insert;
1143
1144 /*
1145 * start end
1146 * | |
1147 * |<------VA------>|<-----Next----->|
1148 * | |
1149 * start end
1150 */
1151 if (next != head) {
1152 sibling = list_entry(next, struct vmap_area, list);
1153 if (sibling->va_start == va->va_end) {
1154 sibling->va_start = va->va_start;
1155
1156 /* Free vmap_area object. */
1157 kmem_cache_free(vmap_area_cachep, va);
1158
1159 /* Point to the new merged area. */
1160 va = sibling;
1161 merged = true;
1162 }
1163 }
1164
1165 /*
1166 * start end
1167 * | |
1168 * |<-----Prev----->|<------VA------>|
1169 * | |
1170 * start end
1171 */
1172 if (next->prev != head) {
1173 sibling = list_entry(next->prev, struct vmap_area, list);
1174 if (sibling->va_end == va->va_start) {
1175 /*
1176 * If both neighbors are coalesced, it is important
1177 * to unlink the "next" node first, followed by merging
1178 * with "previous" one. Otherwise the tree might not be
1179 * fully populated if a sibling's augmented value is
1180 * "normalized" because of rotation operations.
1181 */
1182 if (merged)
1183 __unlink_va(va, root, augment);
1184
1185 sibling->va_end = va->va_end;
1186
1187 /* Free vmap_area object. */
1188 kmem_cache_free(vmap_area_cachep, va);
1189
1190 /* Point to the new merged area. */
1191 va = sibling;
1192 merged = true;
1193 }
1194 }
1195
1196insert:
1197 if (!merged)
1198 __link_va(va, root, parent, link, head, augment);
1199
1200 return va;
1201}
1202
1203static __always_inline struct vmap_area *
1204merge_or_add_vmap_area(struct vmap_area *va,
1205 struct rb_root *root, struct list_head *head)
1206{
1207 return __merge_or_add_vmap_area(va, root, head, false);
1208}
1209
1210static __always_inline struct vmap_area *
1211merge_or_add_vmap_area_augment(struct vmap_area *va,
1212 struct rb_root *root, struct list_head *head)
1213{
1214 va = __merge_or_add_vmap_area(va, root, head, true);
1215 if (va)
1216 augment_tree_propagate_from(va);
1217
1218 return va;
1219}
1220
1221static __always_inline bool
1222is_within_this_va(struct vmap_area *va, unsigned long size,
1223 unsigned long align, unsigned long vstart)
1224{
1225 unsigned long nva_start_addr;
1226
1227 if (va->va_start > vstart)
1228 nva_start_addr = ALIGN(va->va_start, align);
1229 else
1230 nva_start_addr = ALIGN(vstart, align);
1231
1232 /* Can be overflowed due to big size or alignment. */
1233 if (nva_start_addr + size < nva_start_addr ||
1234 nva_start_addr < vstart)
1235 return false;
1236
1237 return (nva_start_addr + size <= va->va_end);
1238}
1239
1240/*
1241 * Find the first free block(lowest start address) in the tree,
1242 * that will accomplish the request corresponding to passing
1243 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1244 * a search length is adjusted to account for worst case alignment
1245 * overhead.
1246 */
1247static __always_inline struct vmap_area *
1248find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1249 unsigned long align, unsigned long vstart, bool adjust_search_size)
1250{
1251 struct vmap_area *va;
1252 struct rb_node *node;
1253 unsigned long length;
1254
1255 /* Start from the root. */
1256 node = root->rb_node;
1257
1258 /* Adjust the search size for alignment overhead. */
1259 length = adjust_search_size ? size + align - 1 : size;
1260
1261 while (node) {
1262 va = rb_entry(node, struct vmap_area, rb_node);
1263
1264 if (get_subtree_max_size(node->rb_left) >= length &&
1265 vstart < va->va_start) {
1266 node = node->rb_left;
1267 } else {
1268 if (is_within_this_va(va, size, align, vstart))
1269 return va;
1270
1271 /*
1272 * Does not make sense to go deeper towards the right
1273 * sub-tree if it does not have a free block that is
1274 * equal or bigger to the requested search length.
1275 */
1276 if (get_subtree_max_size(node->rb_right) >= length) {
1277 node = node->rb_right;
1278 continue;
1279 }
1280
1281 /*
1282 * OK. We roll back and find the first right sub-tree,
1283 * that will satisfy the search criteria. It can happen
1284 * due to "vstart" restriction or an alignment overhead
1285 * that is bigger then PAGE_SIZE.
1286 */
1287 while ((node = rb_parent(node))) {
1288 va = rb_entry(node, struct vmap_area, rb_node);
1289 if (is_within_this_va(va, size, align, vstart))
1290 return va;
1291
1292 if (get_subtree_max_size(node->rb_right) >= length &&
1293 vstart <= va->va_start) {
1294 /*
1295 * Shift the vstart forward. Please note, we update it with
1296 * parent's start address adding "1" because we do not want
1297 * to enter same sub-tree after it has already been checked
1298 * and no suitable free block found there.
1299 */
1300 vstart = va->va_start + 1;
1301 node = node->rb_right;
1302 break;
1303 }
1304 }
1305 }
1306 }
1307
1308 return NULL;
1309}
1310
1311#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1312#include <linux/random.h>
1313
1314static struct vmap_area *
1315find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1316 unsigned long align, unsigned long vstart)
1317{
1318 struct vmap_area *va;
1319
1320 list_for_each_entry(va, head, list) {
1321 if (!is_within_this_va(va, size, align, vstart))
1322 continue;
1323
1324 return va;
1325 }
1326
1327 return NULL;
1328}
1329
1330static void
1331find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1332 unsigned long size, unsigned long align)
1333{
1334 struct vmap_area *va_1, *va_2;
1335 unsigned long vstart;
1336 unsigned int rnd;
1337
1338 get_random_bytes(&rnd, sizeof(rnd));
1339 vstart = VMALLOC_START + rnd;
1340
1341 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1342 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1343
1344 if (va_1 != va_2)
1345 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1346 va_1, va_2, vstart);
1347}
1348#endif
1349
1350enum fit_type {
1351 NOTHING_FIT = 0,
1352 FL_FIT_TYPE = 1, /* full fit */
1353 LE_FIT_TYPE = 2, /* left edge fit */
1354 RE_FIT_TYPE = 3, /* right edge fit */
1355 NE_FIT_TYPE = 4 /* no edge fit */
1356};
1357
1358static __always_inline enum fit_type
1359classify_va_fit_type(struct vmap_area *va,
1360 unsigned long nva_start_addr, unsigned long size)
1361{
1362 enum fit_type type;
1363
1364 /* Check if it is within VA. */
1365 if (nva_start_addr < va->va_start ||
1366 nva_start_addr + size > va->va_end)
1367 return NOTHING_FIT;
1368
1369 /* Now classify. */
1370 if (va->va_start == nva_start_addr) {
1371 if (va->va_end == nva_start_addr + size)
1372 type = FL_FIT_TYPE;
1373 else
1374 type = LE_FIT_TYPE;
1375 } else if (va->va_end == nva_start_addr + size) {
1376 type = RE_FIT_TYPE;
1377 } else {
1378 type = NE_FIT_TYPE;
1379 }
1380
1381 return type;
1382}
1383
1384static __always_inline int
1385adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1386 struct vmap_area *va, unsigned long nva_start_addr,
1387 unsigned long size)
1388{
1389 struct vmap_area *lva = NULL;
1390 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1391
1392 if (type == FL_FIT_TYPE) {
1393 /*
1394 * No need to split VA, it fully fits.
1395 *
1396 * | |
1397 * V NVA V
1398 * |---------------|
1399 */
1400 unlink_va_augment(va, root);
1401 kmem_cache_free(vmap_area_cachep, va);
1402 } else if (type == LE_FIT_TYPE) {
1403 /*
1404 * Split left edge of fit VA.
1405 *
1406 * | |
1407 * V NVA V R
1408 * |-------|-------|
1409 */
1410 va->va_start += size;
1411 } else if (type == RE_FIT_TYPE) {
1412 /*
1413 * Split right edge of fit VA.
1414 *
1415 * | |
1416 * L V NVA V
1417 * |-------|-------|
1418 */
1419 va->va_end = nva_start_addr;
1420 } else if (type == NE_FIT_TYPE) {
1421 /*
1422 * Split no edge of fit VA.
1423 *
1424 * | |
1425 * L V NVA V R
1426 * |---|-------|---|
1427 */
1428 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1429 if (unlikely(!lva)) {
1430 /*
1431 * For percpu allocator we do not do any pre-allocation
1432 * and leave it as it is. The reason is it most likely
1433 * never ends up with NE_FIT_TYPE splitting. In case of
1434 * percpu allocations offsets and sizes are aligned to
1435 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1436 * are its main fitting cases.
1437 *
1438 * There are a few exceptions though, as an example it is
1439 * a first allocation (early boot up) when we have "one"
1440 * big free space that has to be split.
1441 *
1442 * Also we can hit this path in case of regular "vmap"
1443 * allocations, if "this" current CPU was not preloaded.
1444 * See the comment in alloc_vmap_area() why. If so, then
1445 * GFP_NOWAIT is used instead to get an extra object for
1446 * split purpose. That is rare and most time does not
1447 * occur.
1448 *
1449 * What happens if an allocation gets failed. Basically,
1450 * an "overflow" path is triggered to purge lazily freed
1451 * areas to free some memory, then, the "retry" path is
1452 * triggered to repeat one more time. See more details
1453 * in alloc_vmap_area() function.
1454 */
1455 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1456 if (!lva)
1457 return -1;
1458 }
1459
1460 /*
1461 * Build the remainder.
1462 */
1463 lva->va_start = va->va_start;
1464 lva->va_end = nva_start_addr;
1465
1466 /*
1467 * Shrink this VA to remaining size.
1468 */
1469 va->va_start = nva_start_addr + size;
1470 } else {
1471 return -1;
1472 }
1473
1474 if (type != FL_FIT_TYPE) {
1475 augment_tree_propagate_from(va);
1476
1477 if (lva) /* type == NE_FIT_TYPE */
1478 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1479 }
1480
1481 return 0;
1482}
1483
1484/*
1485 * Returns a start address of the newly allocated area, if success.
1486 * Otherwise a vend is returned that indicates failure.
1487 */
1488static __always_inline unsigned long
1489__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1490 unsigned long size, unsigned long align,
1491 unsigned long vstart, unsigned long vend)
1492{
1493 bool adjust_search_size = true;
1494 unsigned long nva_start_addr;
1495 struct vmap_area *va;
1496 int ret;
1497
1498 /*
1499 * Do not adjust when:
1500 * a) align <= PAGE_SIZE, because it does not make any sense.
1501 * All blocks(their start addresses) are at least PAGE_SIZE
1502 * aligned anyway;
1503 * b) a short range where a requested size corresponds to exactly
1504 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1505 * With adjusted search length an allocation would not succeed.
1506 */
1507 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1508 adjust_search_size = false;
1509
1510 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1511 if (unlikely(!va))
1512 return vend;
1513
1514 if (va->va_start > vstart)
1515 nva_start_addr = ALIGN(va->va_start, align);
1516 else
1517 nva_start_addr = ALIGN(vstart, align);
1518
1519 /* Check the "vend" restriction. */
1520 if (nva_start_addr + size > vend)
1521 return vend;
1522
1523 /* Update the free vmap_area. */
1524 ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1525 if (WARN_ON_ONCE(ret))
1526 return vend;
1527
1528#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1529 find_vmap_lowest_match_check(root, head, size, align);
1530#endif
1531
1532 return nva_start_addr;
1533}
1534
1535/*
1536 * Free a region of KVA allocated by alloc_vmap_area
1537 */
1538static void free_vmap_area(struct vmap_area *va)
1539{
1540 /*
1541 * Remove from the busy tree/list.
1542 */
1543 spin_lock(&vmap_area_lock);
1544 unlink_va(va, &vmap_area_root);
1545 spin_unlock(&vmap_area_lock);
1546
1547 /*
1548 * Insert/Merge it back to the free tree/list.
1549 */
1550 spin_lock(&free_vmap_area_lock);
1551 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1552 spin_unlock(&free_vmap_area_lock);
1553}
1554
1555static inline void
1556preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1557{
1558 struct vmap_area *va = NULL;
1559
1560 /*
1561 * Preload this CPU with one extra vmap_area object. It is used
1562 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1563 * a CPU that does an allocation is preloaded.
1564 *
1565 * We do it in non-atomic context, thus it allows us to use more
1566 * permissive allocation masks to be more stable under low memory
1567 * condition and high memory pressure.
1568 */
1569 if (!this_cpu_read(ne_fit_preload_node))
1570 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1571
1572 spin_lock(lock);
1573
1574 if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1575 kmem_cache_free(vmap_area_cachep, va);
1576}
1577
1578/*
1579 * Allocate a region of KVA of the specified size and alignment, within the
1580 * vstart and vend.
1581 */
1582static struct vmap_area *alloc_vmap_area(unsigned long size,
1583 unsigned long align,
1584 unsigned long vstart, unsigned long vend,
1585 int node, gfp_t gfp_mask,
1586 unsigned long va_flags)
1587{
1588 struct vmap_area *va;
1589 unsigned long freed;
1590 unsigned long addr;
1591 int purged = 0;
1592 int ret;
1593
1594 if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1595 return ERR_PTR(-EINVAL);
1596
1597 if (unlikely(!vmap_initialized))
1598 return ERR_PTR(-EBUSY);
1599
1600 might_sleep();
1601 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1602
1603 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1604 if (unlikely(!va))
1605 return ERR_PTR(-ENOMEM);
1606
1607 /*
1608 * Only scan the relevant parts containing pointers to other objects
1609 * to avoid false negatives.
1610 */
1611 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1612
1613retry:
1614 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1615 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1616 size, align, vstart, vend);
1617 spin_unlock(&free_vmap_area_lock);
1618
1619 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1620
1621 /*
1622 * If an allocation fails, the "vend" address is
1623 * returned. Therefore trigger the overflow path.
1624 */
1625 if (unlikely(addr == vend))
1626 goto overflow;
1627
1628 va->va_start = addr;
1629 va->va_end = addr + size;
1630 va->vm = NULL;
1631 va->flags = va_flags;
1632
1633 spin_lock(&vmap_area_lock);
1634 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1635 spin_unlock(&vmap_area_lock);
1636
1637 BUG_ON(!IS_ALIGNED(va->va_start, align));
1638 BUG_ON(va->va_start < vstart);
1639 BUG_ON(va->va_end > vend);
1640
1641 ret = kasan_populate_vmalloc(addr, size);
1642 if (ret) {
1643 free_vmap_area(va);
1644 return ERR_PTR(ret);
1645 }
1646
1647 return va;
1648
1649overflow:
1650 if (!purged) {
1651 reclaim_and_purge_vmap_areas();
1652 purged = 1;
1653 goto retry;
1654 }
1655
1656 freed = 0;
1657 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1658
1659 if (freed > 0) {
1660 purged = 0;
1661 goto retry;
1662 }
1663
1664 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1665 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1666 size);
1667
1668 kmem_cache_free(vmap_area_cachep, va);
1669 return ERR_PTR(-EBUSY);
1670}
1671
1672int register_vmap_purge_notifier(struct notifier_block *nb)
1673{
1674 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1675}
1676EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1677
1678int unregister_vmap_purge_notifier(struct notifier_block *nb)
1679{
1680 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1681}
1682EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1683
1684/*
1685 * lazy_max_pages is the maximum amount of virtual address space we gather up
1686 * before attempting to purge with a TLB flush.
1687 *
1688 * There is a tradeoff here: a larger number will cover more kernel page tables
1689 * and take slightly longer to purge, but it will linearly reduce the number of
1690 * global TLB flushes that must be performed. It would seem natural to scale
1691 * this number up linearly with the number of CPUs (because vmapping activity
1692 * could also scale linearly with the number of CPUs), however it is likely
1693 * that in practice, workloads might be constrained in other ways that mean
1694 * vmap activity will not scale linearly with CPUs. Also, I want to be
1695 * conservative and not introduce a big latency on huge systems, so go with
1696 * a less aggressive log scale. It will still be an improvement over the old
1697 * code, and it will be simple to change the scale factor if we find that it
1698 * becomes a problem on bigger systems.
1699 */
1700static unsigned long lazy_max_pages(void)
1701{
1702 unsigned int log;
1703
1704 log = fls(num_online_cpus());
1705
1706 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1707}
1708
1709static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1710
1711/*
1712 * Serialize vmap purging. There is no actual critical section protected
1713 * by this lock, but we want to avoid concurrent calls for performance
1714 * reasons and to make the pcpu_get_vm_areas more deterministic.
1715 */
1716static DEFINE_MUTEX(vmap_purge_lock);
1717
1718/* for per-CPU blocks */
1719static void purge_fragmented_blocks_allcpus(void);
1720
1721/*
1722 * Purges all lazily-freed vmap areas.
1723 */
1724static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1725{
1726 unsigned long resched_threshold;
1727 unsigned int num_purged_areas = 0;
1728 struct list_head local_purge_list;
1729 struct vmap_area *va, *n_va;
1730
1731 lockdep_assert_held(&vmap_purge_lock);
1732
1733 spin_lock(&purge_vmap_area_lock);
1734 purge_vmap_area_root = RB_ROOT;
1735 list_replace_init(&purge_vmap_area_list, &local_purge_list);
1736 spin_unlock(&purge_vmap_area_lock);
1737
1738 if (unlikely(list_empty(&local_purge_list)))
1739 goto out;
1740
1741 start = min(start,
1742 list_first_entry(&local_purge_list,
1743 struct vmap_area, list)->va_start);
1744
1745 end = max(end,
1746 list_last_entry(&local_purge_list,
1747 struct vmap_area, list)->va_end);
1748
1749 flush_tlb_kernel_range(start, end);
1750 resched_threshold = lazy_max_pages() << 1;
1751
1752 spin_lock(&free_vmap_area_lock);
1753 list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1754 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1755 unsigned long orig_start = va->va_start;
1756 unsigned long orig_end = va->va_end;
1757
1758 /*
1759 * Finally insert or merge lazily-freed area. It is
1760 * detached and there is no need to "unlink" it from
1761 * anything.
1762 */
1763 va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1764 &free_vmap_area_list);
1765
1766 if (!va)
1767 continue;
1768
1769 if (is_vmalloc_or_module_addr((void *)orig_start))
1770 kasan_release_vmalloc(orig_start, orig_end,
1771 va->va_start, va->va_end);
1772
1773 atomic_long_sub(nr, &vmap_lazy_nr);
1774 num_purged_areas++;
1775
1776 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1777 cond_resched_lock(&free_vmap_area_lock);
1778 }
1779 spin_unlock(&free_vmap_area_lock);
1780
1781out:
1782 trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1783 return num_purged_areas > 0;
1784}
1785
1786/*
1787 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
1788 */
1789static void reclaim_and_purge_vmap_areas(void)
1790
1791{
1792 mutex_lock(&vmap_purge_lock);
1793 purge_fragmented_blocks_allcpus();
1794 __purge_vmap_area_lazy(ULONG_MAX, 0);
1795 mutex_unlock(&vmap_purge_lock);
1796}
1797
1798static void drain_vmap_area_work(struct work_struct *work)
1799{
1800 unsigned long nr_lazy;
1801
1802 do {
1803 mutex_lock(&vmap_purge_lock);
1804 __purge_vmap_area_lazy(ULONG_MAX, 0);
1805 mutex_unlock(&vmap_purge_lock);
1806
1807 /* Recheck if further work is required. */
1808 nr_lazy = atomic_long_read(&vmap_lazy_nr);
1809 } while (nr_lazy > lazy_max_pages());
1810}
1811
1812/*
1813 * Free a vmap area, caller ensuring that the area has been unmapped,
1814 * unlinked and flush_cache_vunmap had been called for the correct
1815 * range previously.
1816 */
1817static void free_vmap_area_noflush(struct vmap_area *va)
1818{
1819 unsigned long nr_lazy_max = lazy_max_pages();
1820 unsigned long va_start = va->va_start;
1821 unsigned long nr_lazy;
1822
1823 if (WARN_ON_ONCE(!list_empty(&va->list)))
1824 return;
1825
1826 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1827 PAGE_SHIFT, &vmap_lazy_nr);
1828
1829 /*
1830 * Merge or place it to the purge tree/list.
1831 */
1832 spin_lock(&purge_vmap_area_lock);
1833 merge_or_add_vmap_area(va,
1834 &purge_vmap_area_root, &purge_vmap_area_list);
1835 spin_unlock(&purge_vmap_area_lock);
1836
1837 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1838
1839 /* After this point, we may free va at any time */
1840 if (unlikely(nr_lazy > nr_lazy_max))
1841 schedule_work(&drain_vmap_work);
1842}
1843
1844/*
1845 * Free and unmap a vmap area
1846 */
1847static void free_unmap_vmap_area(struct vmap_area *va)
1848{
1849 flush_cache_vunmap(va->va_start, va->va_end);
1850 vunmap_range_noflush(va->va_start, va->va_end);
1851 if (debug_pagealloc_enabled_static())
1852 flush_tlb_kernel_range(va->va_start, va->va_end);
1853
1854 free_vmap_area_noflush(va);
1855}
1856
1857struct vmap_area *find_vmap_area(unsigned long addr)
1858{
1859 struct vmap_area *va;
1860
1861 spin_lock(&vmap_area_lock);
1862 va = __find_vmap_area(addr, &vmap_area_root);
1863 spin_unlock(&vmap_area_lock);
1864
1865 return va;
1866}
1867
1868static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
1869{
1870 struct vmap_area *va;
1871
1872 spin_lock(&vmap_area_lock);
1873 va = __find_vmap_area(addr, &vmap_area_root);
1874 if (va)
1875 unlink_va(va, &vmap_area_root);
1876 spin_unlock(&vmap_area_lock);
1877
1878 return va;
1879}
1880
1881/*** Per cpu kva allocator ***/
1882
1883/*
1884 * vmap space is limited especially on 32 bit architectures. Ensure there is
1885 * room for at least 16 percpu vmap blocks per CPU.
1886 */
1887/*
1888 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1889 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1890 * instead (we just need a rough idea)
1891 */
1892#if BITS_PER_LONG == 32
1893#define VMALLOC_SPACE (128UL*1024*1024)
1894#else
1895#define VMALLOC_SPACE (128UL*1024*1024*1024)
1896#endif
1897
1898#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1899#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1900#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1901#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1902#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1903#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1904#define VMAP_BBMAP_BITS \
1905 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1906 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1907 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1908
1909#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1910
1911/*
1912 * Purge threshold to prevent overeager purging of fragmented blocks for
1913 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
1914 */
1915#define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
1916
1917#define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
1918#define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
1919#define VMAP_FLAGS_MASK 0x3
1920
1921struct vmap_block_queue {
1922 spinlock_t lock;
1923 struct list_head free;
1924
1925 /*
1926 * An xarray requires an extra memory dynamically to
1927 * be allocated. If it is an issue, we can use rb-tree
1928 * instead.
1929 */
1930 struct xarray vmap_blocks;
1931};
1932
1933struct vmap_block {
1934 spinlock_t lock;
1935 struct vmap_area *va;
1936 unsigned long free, dirty;
1937 DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
1938 unsigned long dirty_min, dirty_max; /*< dirty range */
1939 struct list_head free_list;
1940 struct rcu_head rcu_head;
1941 struct list_head purge;
1942};
1943
1944/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1945static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1946
1947/*
1948 * In order to fast access to any "vmap_block" associated with a
1949 * specific address, we use a hash.
1950 *
1951 * A per-cpu vmap_block_queue is used in both ways, to serialize
1952 * an access to free block chains among CPUs(alloc path) and it
1953 * also acts as a vmap_block hash(alloc/free paths). It means we
1954 * overload it, since we already have the per-cpu array which is
1955 * used as a hash table. When used as a hash a 'cpu' passed to
1956 * per_cpu() is not actually a CPU but rather a hash index.
1957 *
1958 * A hash function is addr_to_vb_xa() which hashes any address
1959 * to a specific index(in a hash) it belongs to. This then uses a
1960 * per_cpu() macro to access an array with generated index.
1961 *
1962 * An example:
1963 *
1964 * CPU_1 CPU_2 CPU_0
1965 * | | |
1966 * V V V
1967 * 0 10 20 30 40 50 60
1968 * |------|------|------|------|------|------|...<vmap address space>
1969 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
1970 *
1971 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
1972 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
1973 *
1974 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
1975 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
1976 *
1977 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
1978 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
1979 *
1980 * This technique almost always avoids lock contention on insert/remove,
1981 * however xarray spinlocks protect against any contention that remains.
1982 */
1983static struct xarray *
1984addr_to_vb_xa(unsigned long addr)
1985{
1986 int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
1987
1988 return &per_cpu(vmap_block_queue, index).vmap_blocks;
1989}
1990
1991/*
1992 * We should probably have a fallback mechanism to allocate virtual memory
1993 * out of partially filled vmap blocks. However vmap block sizing should be
1994 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1995 * big problem.
1996 */
1997
1998static unsigned long addr_to_vb_idx(unsigned long addr)
1999{
2000 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2001 addr /= VMAP_BLOCK_SIZE;
2002 return addr;
2003}
2004
2005static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2006{
2007 unsigned long addr;
2008
2009 addr = va_start + (pages_off << PAGE_SHIFT);
2010 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2011 return (void *)addr;
2012}
2013
2014/**
2015 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2016 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2017 * @order: how many 2^order pages should be occupied in newly allocated block
2018 * @gfp_mask: flags for the page level allocator
2019 *
2020 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2021 */
2022static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2023{
2024 struct vmap_block_queue *vbq;
2025 struct vmap_block *vb;
2026 struct vmap_area *va;
2027 struct xarray *xa;
2028 unsigned long vb_idx;
2029 int node, err;
2030 void *vaddr;
2031
2032 node = numa_node_id();
2033
2034 vb = kmalloc_node(sizeof(struct vmap_block),
2035 gfp_mask & GFP_RECLAIM_MASK, node);
2036 if (unlikely(!vb))
2037 return ERR_PTR(-ENOMEM);
2038
2039 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2040 VMALLOC_START, VMALLOC_END,
2041 node, gfp_mask,
2042 VMAP_RAM|VMAP_BLOCK);
2043 if (IS_ERR(va)) {
2044 kfree(vb);
2045 return ERR_CAST(va);
2046 }
2047
2048 vaddr = vmap_block_vaddr(va->va_start, 0);
2049 spin_lock_init(&vb->lock);
2050 vb->va = va;
2051 /* At least something should be left free */
2052 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2053 bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2054 vb->free = VMAP_BBMAP_BITS - (1UL << order);
2055 vb->dirty = 0;
2056 vb->dirty_min = VMAP_BBMAP_BITS;
2057 vb->dirty_max = 0;
2058 bitmap_set(vb->used_map, 0, (1UL << order));
2059 INIT_LIST_HEAD(&vb->free_list);
2060
2061 xa = addr_to_vb_xa(va->va_start);
2062 vb_idx = addr_to_vb_idx(va->va_start);
2063 err = xa_insert(xa, vb_idx, vb, gfp_mask);
2064 if (err) {
2065 kfree(vb);
2066 free_vmap_area(va);
2067 return ERR_PTR(err);
2068 }
2069
2070 vbq = raw_cpu_ptr(&vmap_block_queue);
2071 spin_lock(&vbq->lock);
2072 list_add_tail_rcu(&vb->free_list, &vbq->free);
2073 spin_unlock(&vbq->lock);
2074
2075 return vaddr;
2076}
2077
2078static void free_vmap_block(struct vmap_block *vb)
2079{
2080 struct vmap_block *tmp;
2081 struct xarray *xa;
2082
2083 xa = addr_to_vb_xa(vb->va->va_start);
2084 tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2085 BUG_ON(tmp != vb);
2086
2087 spin_lock(&vmap_area_lock);
2088 unlink_va(vb->va, &vmap_area_root);
2089 spin_unlock(&vmap_area_lock);
2090
2091 free_vmap_area_noflush(vb->va);
2092 kfree_rcu(vb, rcu_head);
2093}
2094
2095static bool purge_fragmented_block(struct vmap_block *vb,
2096 struct vmap_block_queue *vbq, struct list_head *purge_list,
2097 bool force_purge)
2098{
2099 if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2100 vb->dirty == VMAP_BBMAP_BITS)
2101 return false;
2102
2103 /* Don't overeagerly purge usable blocks unless requested */
2104 if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2105 return false;
2106
2107 /* prevent further allocs after releasing lock */
2108 WRITE_ONCE(vb->free, 0);
2109 /* prevent purging it again */
2110 WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2111 vb->dirty_min = 0;
2112 vb->dirty_max = VMAP_BBMAP_BITS;
2113 spin_lock(&vbq->lock);
2114 list_del_rcu(&vb->free_list);
2115 spin_unlock(&vbq->lock);
2116 list_add_tail(&vb->purge, purge_list);
2117 return true;
2118}
2119
2120static void free_purged_blocks(struct list_head *purge_list)
2121{
2122 struct vmap_block *vb, *n_vb;
2123
2124 list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2125 list_del(&vb->purge);
2126 free_vmap_block(vb);
2127 }
2128}
2129
2130static void purge_fragmented_blocks(int cpu)
2131{
2132 LIST_HEAD(purge);
2133 struct vmap_block *vb;
2134 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2135
2136 rcu_read_lock();
2137 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2138 unsigned long free = READ_ONCE(vb->free);
2139 unsigned long dirty = READ_ONCE(vb->dirty);
2140
2141 if (free + dirty != VMAP_BBMAP_BITS ||
2142 dirty == VMAP_BBMAP_BITS)
2143 continue;
2144
2145 spin_lock(&vb->lock);
2146 purge_fragmented_block(vb, vbq, &purge, true);
2147 spin_unlock(&vb->lock);
2148 }
2149 rcu_read_unlock();
2150 free_purged_blocks(&purge);
2151}
2152
2153static void purge_fragmented_blocks_allcpus(void)
2154{
2155 int cpu;
2156
2157 for_each_possible_cpu(cpu)
2158 purge_fragmented_blocks(cpu);
2159}
2160
2161static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2162{
2163 struct vmap_block_queue *vbq;
2164 struct vmap_block *vb;
2165 void *vaddr = NULL;
2166 unsigned int order;
2167
2168 BUG_ON(offset_in_page(size));
2169 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2170 if (WARN_ON(size == 0)) {
2171 /*
2172 * Allocating 0 bytes isn't what caller wants since
2173 * get_order(0) returns funny result. Just warn and terminate
2174 * early.
2175 */
2176 return NULL;
2177 }
2178 order = get_order(size);
2179
2180 rcu_read_lock();
2181 vbq = raw_cpu_ptr(&vmap_block_queue);
2182 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2183 unsigned long pages_off;
2184
2185 if (READ_ONCE(vb->free) < (1UL << order))
2186 continue;
2187
2188 spin_lock(&vb->lock);
2189 if (vb->free < (1UL << order)) {
2190 spin_unlock(&vb->lock);
2191 continue;
2192 }
2193
2194 pages_off = VMAP_BBMAP_BITS - vb->free;
2195 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2196 WRITE_ONCE(vb->free, vb->free - (1UL << order));
2197 bitmap_set(vb->used_map, pages_off, (1UL << order));
2198 if (vb->free == 0) {
2199 spin_lock(&vbq->lock);
2200 list_del_rcu(&vb->free_list);
2201 spin_unlock(&vbq->lock);
2202 }
2203
2204 spin_unlock(&vb->lock);
2205 break;
2206 }
2207
2208 rcu_read_unlock();
2209
2210 /* Allocate new block if nothing was found */
2211 if (!vaddr)
2212 vaddr = new_vmap_block(order, gfp_mask);
2213
2214 return vaddr;
2215}
2216
2217static void vb_free(unsigned long addr, unsigned long size)
2218{
2219 unsigned long offset;
2220 unsigned int order;
2221 struct vmap_block *vb;
2222 struct xarray *xa;
2223
2224 BUG_ON(offset_in_page(size));
2225 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2226
2227 flush_cache_vunmap(addr, addr + size);
2228
2229 order = get_order(size);
2230 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2231
2232 xa = addr_to_vb_xa(addr);
2233 vb = xa_load(xa, addr_to_vb_idx(addr));
2234
2235 spin_lock(&vb->lock);
2236 bitmap_clear(vb->used_map, offset, (1UL << order));
2237 spin_unlock(&vb->lock);
2238
2239 vunmap_range_noflush(addr, addr + size);
2240
2241 if (debug_pagealloc_enabled_static())
2242 flush_tlb_kernel_range(addr, addr + size);
2243
2244 spin_lock(&vb->lock);
2245
2246 /* Expand the not yet TLB flushed dirty range */
2247 vb->dirty_min = min(vb->dirty_min, offset);
2248 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2249
2250 WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2251 if (vb->dirty == VMAP_BBMAP_BITS) {
2252 BUG_ON(vb->free);
2253 spin_unlock(&vb->lock);
2254 free_vmap_block(vb);
2255 } else
2256 spin_unlock(&vb->lock);
2257}
2258
2259static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2260{
2261 LIST_HEAD(purge_list);
2262 int cpu;
2263
2264 if (unlikely(!vmap_initialized))
2265 return;
2266
2267 mutex_lock(&vmap_purge_lock);
2268
2269 for_each_possible_cpu(cpu) {
2270 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2271 struct vmap_block *vb;
2272 unsigned long idx;
2273
2274 rcu_read_lock();
2275 xa_for_each(&vbq->vmap_blocks, idx, vb) {
2276 spin_lock(&vb->lock);
2277
2278 /*
2279 * Try to purge a fragmented block first. If it's
2280 * not purgeable, check whether there is dirty
2281 * space to be flushed.
2282 */
2283 if (!purge_fragmented_block(vb, vbq, &purge_list, false) &&
2284 vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2285 unsigned long va_start = vb->va->va_start;
2286 unsigned long s, e;
2287
2288 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2289 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2290
2291 start = min(s, start);
2292 end = max(e, end);
2293
2294 /* Prevent that this is flushed again */
2295 vb->dirty_min = VMAP_BBMAP_BITS;
2296 vb->dirty_max = 0;
2297
2298 flush = 1;
2299 }
2300 spin_unlock(&vb->lock);
2301 }
2302 rcu_read_unlock();
2303 }
2304 free_purged_blocks(&purge_list);
2305
2306 if (!__purge_vmap_area_lazy(start, end) && flush)
2307 flush_tlb_kernel_range(start, end);
2308 mutex_unlock(&vmap_purge_lock);
2309}
2310
2311/**
2312 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2313 *
2314 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2315 * to amortize TLB flushing overheads. What this means is that any page you
2316 * have now, may, in a former life, have been mapped into kernel virtual
2317 * address by the vmap layer and so there might be some CPUs with TLB entries
2318 * still referencing that page (additional to the regular 1:1 kernel mapping).
2319 *
2320 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2321 * be sure that none of the pages we have control over will have any aliases
2322 * from the vmap layer.
2323 */
2324void vm_unmap_aliases(void)
2325{
2326 unsigned long start = ULONG_MAX, end = 0;
2327 int flush = 0;
2328
2329 _vm_unmap_aliases(start, end, flush);
2330}
2331EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2332
2333/**
2334 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2335 * @mem: the pointer returned by vm_map_ram
2336 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2337 */
2338void vm_unmap_ram(const void *mem, unsigned int count)
2339{
2340 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2341 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2342 struct vmap_area *va;
2343
2344 might_sleep();
2345 BUG_ON(!addr);
2346 BUG_ON(addr < VMALLOC_START);
2347 BUG_ON(addr > VMALLOC_END);
2348 BUG_ON(!PAGE_ALIGNED(addr));
2349
2350 kasan_poison_vmalloc(mem, size);
2351
2352 if (likely(count <= VMAP_MAX_ALLOC)) {
2353 debug_check_no_locks_freed(mem, size);
2354 vb_free(addr, size);
2355 return;
2356 }
2357
2358 va = find_unlink_vmap_area(addr);
2359 if (WARN_ON_ONCE(!va))
2360 return;
2361
2362 debug_check_no_locks_freed((void *)va->va_start,
2363 (va->va_end - va->va_start));
2364 free_unmap_vmap_area(va);
2365}
2366EXPORT_SYMBOL(vm_unmap_ram);
2367
2368/**
2369 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2370 * @pages: an array of pointers to the pages to be mapped
2371 * @count: number of pages
2372 * @node: prefer to allocate data structures on this node
2373 *
2374 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2375 * faster than vmap so it's good. But if you mix long-life and short-life
2376 * objects with vm_map_ram(), it could consume lots of address space through
2377 * fragmentation (especially on a 32bit machine). You could see failures in
2378 * the end. Please use this function for short-lived objects.
2379 *
2380 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2381 */
2382void *vm_map_ram(struct page **pages, unsigned int count, int node)
2383{
2384 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2385 unsigned long addr;
2386 void *mem;
2387
2388 if (likely(count <= VMAP_MAX_ALLOC)) {
2389 mem = vb_alloc(size, GFP_KERNEL);
2390 if (IS_ERR(mem))
2391 return NULL;
2392 addr = (unsigned long)mem;
2393 } else {
2394 struct vmap_area *va;
2395 va = alloc_vmap_area(size, PAGE_SIZE,
2396 VMALLOC_START, VMALLOC_END,
2397 node, GFP_KERNEL, VMAP_RAM);
2398 if (IS_ERR(va))
2399 return NULL;
2400
2401 addr = va->va_start;
2402 mem = (void *)addr;
2403 }
2404
2405 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2406 pages, PAGE_SHIFT) < 0) {
2407 vm_unmap_ram(mem, count);
2408 return NULL;
2409 }
2410
2411 /*
2412 * Mark the pages as accessible, now that they are mapped.
2413 * With hardware tag-based KASAN, marking is skipped for
2414 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2415 */
2416 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2417
2418 return mem;
2419}
2420EXPORT_SYMBOL(vm_map_ram);
2421
2422static struct vm_struct *vmlist __initdata;
2423
2424static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2425{
2426#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2427 return vm->page_order;
2428#else
2429 return 0;
2430#endif
2431}
2432
2433static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2434{
2435#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2436 vm->page_order = order;
2437#else
2438 BUG_ON(order != 0);
2439#endif
2440}
2441
2442/**
2443 * vm_area_add_early - add vmap area early during boot
2444 * @vm: vm_struct to add
2445 *
2446 * This function is used to add fixed kernel vm area to vmlist before
2447 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
2448 * should contain proper values and the other fields should be zero.
2449 *
2450 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2451 */
2452void __init vm_area_add_early(struct vm_struct *vm)
2453{
2454 struct vm_struct *tmp, **p;
2455
2456 BUG_ON(vmap_initialized);
2457 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2458 if (tmp->addr >= vm->addr) {
2459 BUG_ON(tmp->addr < vm->addr + vm->size);
2460 break;
2461 } else
2462 BUG_ON(tmp->addr + tmp->size > vm->addr);
2463 }
2464 vm->next = *p;
2465 *p = vm;
2466}
2467
2468/**
2469 * vm_area_register_early - register vmap area early during boot
2470 * @vm: vm_struct to register
2471 * @align: requested alignment
2472 *
2473 * This function is used to register kernel vm area before
2474 * vmalloc_init() is called. @vm->size and @vm->flags should contain
2475 * proper values on entry and other fields should be zero. On return,
2476 * vm->addr contains the allocated address.
2477 *
2478 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2479 */
2480void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2481{
2482 unsigned long addr = ALIGN(VMALLOC_START, align);
2483 struct vm_struct *cur, **p;
2484
2485 BUG_ON(vmap_initialized);
2486
2487 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2488 if ((unsigned long)cur->addr - addr >= vm->size)
2489 break;
2490 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2491 }
2492
2493 BUG_ON(addr > VMALLOC_END - vm->size);
2494 vm->addr = (void *)addr;
2495 vm->next = *p;
2496 *p = vm;
2497 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2498}
2499
2500static void vmap_init_free_space(void)
2501{
2502 unsigned long vmap_start = 1;
2503 const unsigned long vmap_end = ULONG_MAX;
2504 struct vmap_area *busy, *free;
2505
2506 /*
2507 * B F B B B F
2508 * -|-----|.....|-----|-----|-----|.....|-
2509 * | The KVA space |
2510 * |<--------------------------------->|
2511 */
2512 list_for_each_entry(busy, &vmap_area_list, list) {
2513 if (busy->va_start - vmap_start > 0) {
2514 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2515 if (!WARN_ON_ONCE(!free)) {
2516 free->va_start = vmap_start;
2517 free->va_end = busy->va_start;
2518
2519 insert_vmap_area_augment(free, NULL,
2520 &free_vmap_area_root,
2521 &free_vmap_area_list);
2522 }
2523 }
2524
2525 vmap_start = busy->va_end;
2526 }
2527
2528 if (vmap_end - vmap_start > 0) {
2529 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2530 if (!WARN_ON_ONCE(!free)) {
2531 free->va_start = vmap_start;
2532 free->va_end = vmap_end;
2533
2534 insert_vmap_area_augment(free, NULL,
2535 &free_vmap_area_root,
2536 &free_vmap_area_list);
2537 }
2538 }
2539}
2540
2541static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2542 struct vmap_area *va, unsigned long flags, const void *caller)
2543{
2544 vm->flags = flags;
2545 vm->addr = (void *)va->va_start;
2546 vm->size = va->va_end - va->va_start;
2547 vm->caller = caller;
2548 va->vm = vm;
2549}
2550
2551static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2552 unsigned long flags, const void *caller)
2553{
2554 spin_lock(&vmap_area_lock);
2555 setup_vmalloc_vm_locked(vm, va, flags, caller);
2556 spin_unlock(&vmap_area_lock);
2557}
2558
2559static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2560{
2561 /*
2562 * Before removing VM_UNINITIALIZED,
2563 * we should make sure that vm has proper values.
2564 * Pair with smp_rmb() in show_numa_info().
2565 */
2566 smp_wmb();
2567 vm->flags &= ~VM_UNINITIALIZED;
2568}
2569
2570static struct vm_struct *__get_vm_area_node(unsigned long size,
2571 unsigned long align, unsigned long shift, unsigned long flags,
2572 unsigned long start, unsigned long end, int node,
2573 gfp_t gfp_mask, const void *caller)
2574{
2575 struct vmap_area *va;
2576 struct vm_struct *area;
2577 unsigned long requested_size = size;
2578
2579 BUG_ON(in_interrupt());
2580 size = ALIGN(size, 1ul << shift);
2581 if (unlikely(!size))
2582 return NULL;
2583
2584 if (flags & VM_IOREMAP)
2585 align = 1ul << clamp_t(int, get_count_order_long(size),
2586 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2587
2588 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2589 if (unlikely(!area))
2590 return NULL;
2591
2592 if (!(flags & VM_NO_GUARD))
2593 size += PAGE_SIZE;
2594
2595 va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0);
2596 if (IS_ERR(va)) {
2597 kfree(area);
2598 return NULL;
2599 }
2600
2601 setup_vmalloc_vm(area, va, flags, caller);
2602
2603 /*
2604 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2605 * best-effort approach, as they can be mapped outside of vmalloc code.
2606 * For VM_ALLOC mappings, the pages are marked as accessible after
2607 * getting mapped in __vmalloc_node_range().
2608 * With hardware tag-based KASAN, marking is skipped for
2609 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2610 */
2611 if (!(flags & VM_ALLOC))
2612 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2613 KASAN_VMALLOC_PROT_NORMAL);
2614
2615 return area;
2616}
2617
2618struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2619 unsigned long start, unsigned long end,
2620 const void *caller)
2621{
2622 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2623 NUMA_NO_NODE, GFP_KERNEL, caller);
2624}
2625
2626/**
2627 * get_vm_area - reserve a contiguous kernel virtual area
2628 * @size: size of the area
2629 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2630 *
2631 * Search an area of @size in the kernel virtual mapping area,
2632 * and reserved it for out purposes. Returns the area descriptor
2633 * on success or %NULL on failure.
2634 *
2635 * Return: the area descriptor on success or %NULL on failure.
2636 */
2637struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2638{
2639 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2640 VMALLOC_START, VMALLOC_END,
2641 NUMA_NO_NODE, GFP_KERNEL,
2642 __builtin_return_address(0));
2643}
2644
2645struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2646 const void *caller)
2647{
2648 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2649 VMALLOC_START, VMALLOC_END,
2650 NUMA_NO_NODE, GFP_KERNEL, caller);
2651}
2652
2653/**
2654 * find_vm_area - find a continuous kernel virtual area
2655 * @addr: base address
2656 *
2657 * Search for the kernel VM area starting at @addr, and return it.
2658 * It is up to the caller to do all required locking to keep the returned
2659 * pointer valid.
2660 *
2661 * Return: the area descriptor on success or %NULL on failure.
2662 */
2663struct vm_struct *find_vm_area(const void *addr)
2664{
2665 struct vmap_area *va;
2666
2667 va = find_vmap_area((unsigned long)addr);
2668 if (!va)
2669 return NULL;
2670
2671 return va->vm;
2672}
2673
2674/**
2675 * remove_vm_area - find and remove a continuous kernel virtual area
2676 * @addr: base address
2677 *
2678 * Search for the kernel VM area starting at @addr, and remove it.
2679 * This function returns the found VM area, but using it is NOT safe
2680 * on SMP machines, except for its size or flags.
2681 *
2682 * Return: the area descriptor on success or %NULL on failure.
2683 */
2684struct vm_struct *remove_vm_area(const void *addr)
2685{
2686 struct vmap_area *va;
2687 struct vm_struct *vm;
2688
2689 might_sleep();
2690
2691 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2692 addr))
2693 return NULL;
2694
2695 va = find_unlink_vmap_area((unsigned long)addr);
2696 if (!va || !va->vm)
2697 return NULL;
2698 vm = va->vm;
2699
2700 debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
2701 debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
2702 kasan_free_module_shadow(vm);
2703 kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
2704
2705 free_unmap_vmap_area(va);
2706 return vm;
2707}
2708
2709static inline void set_area_direct_map(const struct vm_struct *area,
2710 int (*set_direct_map)(struct page *page))
2711{
2712 int i;
2713
2714 /* HUGE_VMALLOC passes small pages to set_direct_map */
2715 for (i = 0; i < area->nr_pages; i++)
2716 if (page_address(area->pages[i]))
2717 set_direct_map(area->pages[i]);
2718}
2719
2720/*
2721 * Flush the vm mapping and reset the direct map.
2722 */
2723static void vm_reset_perms(struct vm_struct *area)
2724{
2725 unsigned long start = ULONG_MAX, end = 0;
2726 unsigned int page_order = vm_area_page_order(area);
2727 int flush_dmap = 0;
2728 int i;
2729
2730 /*
2731 * Find the start and end range of the direct mappings to make sure that
2732 * the vm_unmap_aliases() flush includes the direct map.
2733 */
2734 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2735 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2736
2737 if (addr) {
2738 unsigned long page_size;
2739
2740 page_size = PAGE_SIZE << page_order;
2741 start = min(addr, start);
2742 end = max(addr + page_size, end);
2743 flush_dmap = 1;
2744 }
2745 }
2746
2747 /*
2748 * Set direct map to something invalid so that it won't be cached if
2749 * there are any accesses after the TLB flush, then flush the TLB and
2750 * reset the direct map permissions to the default.
2751 */
2752 set_area_direct_map(area, set_direct_map_invalid_noflush);
2753 _vm_unmap_aliases(start, end, flush_dmap);
2754 set_area_direct_map(area, set_direct_map_default_noflush);
2755}
2756
2757static void delayed_vfree_work(struct work_struct *w)
2758{
2759 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
2760 struct llist_node *t, *llnode;
2761
2762 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
2763 vfree(llnode);
2764}
2765
2766/**
2767 * vfree_atomic - release memory allocated by vmalloc()
2768 * @addr: memory base address
2769 *
2770 * This one is just like vfree() but can be called in any atomic context
2771 * except NMIs.
2772 */
2773void vfree_atomic(const void *addr)
2774{
2775 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2776
2777 BUG_ON(in_nmi());
2778 kmemleak_free(addr);
2779
2780 /*
2781 * Use raw_cpu_ptr() because this can be called from preemptible
2782 * context. Preemption is absolutely fine here, because the llist_add()
2783 * implementation is lockless, so it works even if we are adding to
2784 * another cpu's list. schedule_work() should be fine with this too.
2785 */
2786 if (addr && llist_add((struct llist_node *)addr, &p->list))
2787 schedule_work(&p->wq);
2788}
2789
2790/**
2791 * vfree - Release memory allocated by vmalloc()
2792 * @addr: Memory base address
2793 *
2794 * Free the virtually continuous memory area starting at @addr, as obtained
2795 * from one of the vmalloc() family of APIs. This will usually also free the
2796 * physical memory underlying the virtual allocation, but that memory is
2797 * reference counted, so it will not be freed until the last user goes away.
2798 *
2799 * If @addr is NULL, no operation is performed.
2800 *
2801 * Context:
2802 * May sleep if called *not* from interrupt context.
2803 * Must not be called in NMI context (strictly speaking, it could be
2804 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2805 * conventions for vfree() arch-dependent would be a really bad idea).
2806 */
2807void vfree(const void *addr)
2808{
2809 struct vm_struct *vm;
2810 int i;
2811
2812 if (unlikely(in_interrupt())) {
2813 vfree_atomic(addr);
2814 return;
2815 }
2816
2817 BUG_ON(in_nmi());
2818 kmemleak_free(addr);
2819 might_sleep();
2820
2821 if (!addr)
2822 return;
2823
2824 vm = remove_vm_area(addr);
2825 if (unlikely(!vm)) {
2826 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2827 addr);
2828 return;
2829 }
2830
2831 if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
2832 vm_reset_perms(vm);
2833 for (i = 0; i < vm->nr_pages; i++) {
2834 struct page *page = vm->pages[i];
2835
2836 BUG_ON(!page);
2837 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2838 /*
2839 * High-order allocs for huge vmallocs are split, so
2840 * can be freed as an array of order-0 allocations
2841 */
2842 __free_page(page);
2843 cond_resched();
2844 }
2845 atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
2846 kvfree(vm->pages);
2847 kfree(vm);
2848}
2849EXPORT_SYMBOL(vfree);
2850
2851/**
2852 * vunmap - release virtual mapping obtained by vmap()
2853 * @addr: memory base address
2854 *
2855 * Free the virtually contiguous memory area starting at @addr,
2856 * which was created from the page array passed to vmap().
2857 *
2858 * Must not be called in interrupt context.
2859 */
2860void vunmap(const void *addr)
2861{
2862 struct vm_struct *vm;
2863
2864 BUG_ON(in_interrupt());
2865 might_sleep();
2866
2867 if (!addr)
2868 return;
2869 vm = remove_vm_area(addr);
2870 if (unlikely(!vm)) {
2871 WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
2872 addr);
2873 return;
2874 }
2875 kfree(vm);
2876}
2877EXPORT_SYMBOL(vunmap);
2878
2879/**
2880 * vmap - map an array of pages into virtually contiguous space
2881 * @pages: array of page pointers
2882 * @count: number of pages to map
2883 * @flags: vm_area->flags
2884 * @prot: page protection for the mapping
2885 *
2886 * Maps @count pages from @pages into contiguous kernel virtual space.
2887 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2888 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2889 * are transferred from the caller to vmap(), and will be freed / dropped when
2890 * vfree() is called on the return value.
2891 *
2892 * Return: the address of the area or %NULL on failure
2893 */
2894void *vmap(struct page **pages, unsigned int count,
2895 unsigned long flags, pgprot_t prot)
2896{
2897 struct vm_struct *area;
2898 unsigned long addr;
2899 unsigned long size; /* In bytes */
2900
2901 might_sleep();
2902
2903 if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
2904 return NULL;
2905
2906 /*
2907 * Your top guard is someone else's bottom guard. Not having a top
2908 * guard compromises someone else's mappings too.
2909 */
2910 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2911 flags &= ~VM_NO_GUARD;
2912
2913 if (count > totalram_pages())
2914 return NULL;
2915
2916 size = (unsigned long)count << PAGE_SHIFT;
2917 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2918 if (!area)
2919 return NULL;
2920
2921 addr = (unsigned long)area->addr;
2922 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2923 pages, PAGE_SHIFT) < 0) {
2924 vunmap(area->addr);
2925 return NULL;
2926 }
2927
2928 if (flags & VM_MAP_PUT_PAGES) {
2929 area->pages = pages;
2930 area->nr_pages = count;
2931 }
2932 return area->addr;
2933}
2934EXPORT_SYMBOL(vmap);
2935
2936#ifdef CONFIG_VMAP_PFN
2937struct vmap_pfn_data {
2938 unsigned long *pfns;
2939 pgprot_t prot;
2940 unsigned int idx;
2941};
2942
2943static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2944{
2945 struct vmap_pfn_data *data = private;
2946 unsigned long pfn = data->pfns[data->idx];
2947 pte_t ptent;
2948
2949 if (WARN_ON_ONCE(pfn_valid(pfn)))
2950 return -EINVAL;
2951
2952 ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
2953 set_pte_at(&init_mm, addr, pte, ptent);
2954
2955 data->idx++;
2956 return 0;
2957}
2958
2959/**
2960 * vmap_pfn - map an array of PFNs into virtually contiguous space
2961 * @pfns: array of PFNs
2962 * @count: number of pages to map
2963 * @prot: page protection for the mapping
2964 *
2965 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2966 * the start address of the mapping.
2967 */
2968void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2969{
2970 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2971 struct vm_struct *area;
2972
2973 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2974 __builtin_return_address(0));
2975 if (!area)
2976 return NULL;
2977 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2978 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2979 free_vm_area(area);
2980 return NULL;
2981 }
2982
2983 flush_cache_vmap((unsigned long)area->addr,
2984 (unsigned long)area->addr + count * PAGE_SIZE);
2985
2986 return area->addr;
2987}
2988EXPORT_SYMBOL_GPL(vmap_pfn);
2989#endif /* CONFIG_VMAP_PFN */
2990
2991static inline unsigned int
2992vm_area_alloc_pages(gfp_t gfp, int nid,
2993 unsigned int order, unsigned int nr_pages, struct page **pages)
2994{
2995 unsigned int nr_allocated = 0;
2996 gfp_t alloc_gfp = gfp;
2997 bool nofail = false;
2998 struct page *page;
2999 int i;
3000
3001 /*
3002 * For order-0 pages we make use of bulk allocator, if
3003 * the page array is partly or not at all populated due
3004 * to fails, fallback to a single page allocator that is
3005 * more permissive.
3006 */
3007 if (!order) {
3008 /* bulk allocator doesn't support nofail req. officially */
3009 gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
3010
3011 while (nr_allocated < nr_pages) {
3012 unsigned int nr, nr_pages_request;
3013
3014 /*
3015 * A maximum allowed request is hard-coded and is 100
3016 * pages per call. That is done in order to prevent a
3017 * long preemption off scenario in the bulk-allocator
3018 * so the range is [1:100].
3019 */
3020 nr_pages_request = min(100U, nr_pages - nr_allocated);
3021
3022 /* memory allocation should consider mempolicy, we can't
3023 * wrongly use nearest node when nid == NUMA_NO_NODE,
3024 * otherwise memory may be allocated in only one node,
3025 * but mempolicy wants to alloc memory by interleaving.
3026 */
3027 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3028 nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
3029 nr_pages_request,
3030 pages + nr_allocated);
3031
3032 else
3033 nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
3034 nr_pages_request,
3035 pages + nr_allocated);
3036
3037 nr_allocated += nr;
3038 cond_resched();
3039
3040 /*
3041 * If zero or pages were obtained partly,
3042 * fallback to a single page allocator.
3043 */
3044 if (nr != nr_pages_request)
3045 break;
3046 }
3047 } else if (gfp & __GFP_NOFAIL) {
3048 /*
3049 * Higher order nofail allocations are really expensive and
3050 * potentially dangerous (pre-mature OOM, disruptive reclaim
3051 * and compaction etc.
3052 */
3053 alloc_gfp &= ~__GFP_NOFAIL;
3054 nofail = true;
3055 }
3056
3057 /* High-order pages or fallback path if "bulk" fails. */
3058 while (nr_allocated < nr_pages) {
3059 if (fatal_signal_pending(current))
3060 break;
3061
3062 if (nid == NUMA_NO_NODE)
3063 page = alloc_pages(alloc_gfp, order);
3064 else
3065 page = alloc_pages_node(nid, alloc_gfp, order);
3066 if (unlikely(!page)) {
3067 if (!nofail)
3068 break;
3069
3070 /* fall back to the zero order allocations */
3071 alloc_gfp |= __GFP_NOFAIL;
3072 order = 0;
3073 continue;
3074 }
3075
3076 /*
3077 * Higher order allocations must be able to be treated as
3078 * indepdenent small pages by callers (as they can with
3079 * small-page vmallocs). Some drivers do their own refcounting
3080 * on vmalloc_to_page() pages, some use page->mapping,
3081 * page->lru, etc.
3082 */
3083 if (order)
3084 split_page(page, order);
3085
3086 /*
3087 * Careful, we allocate and map page-order pages, but
3088 * tracking is done per PAGE_SIZE page so as to keep the
3089 * vm_struct APIs independent of the physical/mapped size.
3090 */
3091 for (i = 0; i < (1U << order); i++)
3092 pages[nr_allocated + i] = page + i;
3093
3094 cond_resched();
3095 nr_allocated += 1U << order;
3096 }
3097
3098 return nr_allocated;
3099}
3100
3101static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3102 pgprot_t prot, unsigned int page_shift,
3103 int node)
3104{
3105 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3106 bool nofail = gfp_mask & __GFP_NOFAIL;
3107 unsigned long addr = (unsigned long)area->addr;
3108 unsigned long size = get_vm_area_size(area);
3109 unsigned long array_size;
3110 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3111 unsigned int page_order;
3112 unsigned int flags;
3113 int ret;
3114
3115 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3116
3117 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3118 gfp_mask |= __GFP_HIGHMEM;
3119
3120 /* Please note that the recursion is strictly bounded. */
3121 if (array_size > PAGE_SIZE) {
3122 area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3123 area->caller);
3124 } else {
3125 area->pages = kmalloc_node(array_size, nested_gfp, node);
3126 }
3127
3128 if (!area->pages) {
3129 warn_alloc(gfp_mask, NULL,
3130 "vmalloc error: size %lu, failed to allocated page array size %lu",
3131 nr_small_pages * PAGE_SIZE, array_size);
3132 free_vm_area(area);
3133 return NULL;
3134 }
3135
3136 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3137 page_order = vm_area_page_order(area);
3138
3139 area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3140 node, page_order, nr_small_pages, area->pages);
3141
3142 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3143 if (gfp_mask & __GFP_ACCOUNT) {
3144 int i;
3145
3146 for (i = 0; i < area->nr_pages; i++)
3147 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3148 }
3149
3150 /*
3151 * If not enough pages were obtained to accomplish an
3152 * allocation request, free them via vfree() if any.
3153 */
3154 if (area->nr_pages != nr_small_pages) {
3155 /*
3156 * vm_area_alloc_pages() can fail due to insufficient memory but
3157 * also:-
3158 *
3159 * - a pending fatal signal
3160 * - insufficient huge page-order pages
3161 *
3162 * Since we always retry allocations at order-0 in the huge page
3163 * case a warning for either is spurious.
3164 */
3165 if (!fatal_signal_pending(current) && page_order == 0)
3166 warn_alloc(gfp_mask, NULL,
3167 "vmalloc error: size %lu, failed to allocate pages",
3168 area->nr_pages * PAGE_SIZE);
3169 goto fail;
3170 }
3171
3172 /*
3173 * page tables allocations ignore external gfp mask, enforce it
3174 * by the scope API
3175 */
3176 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3177 flags = memalloc_nofs_save();
3178 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3179 flags = memalloc_noio_save();
3180
3181 do {
3182 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3183 page_shift);
3184 if (nofail && (ret < 0))
3185 schedule_timeout_uninterruptible(1);
3186 } while (nofail && (ret < 0));
3187
3188 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3189 memalloc_nofs_restore(flags);
3190 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3191 memalloc_noio_restore(flags);
3192
3193 if (ret < 0) {
3194 warn_alloc(gfp_mask, NULL,
3195 "vmalloc error: size %lu, failed to map pages",
3196 area->nr_pages * PAGE_SIZE);
3197 goto fail;
3198 }
3199
3200 return area->addr;
3201
3202fail:
3203 vfree(area->addr);
3204 return NULL;
3205}
3206
3207/**
3208 * __vmalloc_node_range - allocate virtually contiguous memory
3209 * @size: allocation size
3210 * @align: desired alignment
3211 * @start: vm area range start
3212 * @end: vm area range end
3213 * @gfp_mask: flags for the page level allocator
3214 * @prot: protection mask for the allocated pages
3215 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3216 * @node: node to use for allocation or NUMA_NO_NODE
3217 * @caller: caller's return address
3218 *
3219 * Allocate enough pages to cover @size from the page level
3220 * allocator with @gfp_mask flags. Please note that the full set of gfp
3221 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3222 * supported.
3223 * Zone modifiers are not supported. From the reclaim modifiers
3224 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3225 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3226 * __GFP_RETRY_MAYFAIL are not supported).
3227 *
3228 * __GFP_NOWARN can be used to suppress failures messages.
3229 *
3230 * Map them into contiguous kernel virtual space, using a pagetable
3231 * protection of @prot.
3232 *
3233 * Return: the address of the area or %NULL on failure
3234 */
3235void *__vmalloc_node_range(unsigned long size, unsigned long align,
3236 unsigned long start, unsigned long end, gfp_t gfp_mask,
3237 pgprot_t prot, unsigned long vm_flags, int node,
3238 const void *caller)
3239{
3240 struct vm_struct *area;
3241 void *ret;
3242 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3243 unsigned long real_size = size;
3244 unsigned long real_align = align;
3245 unsigned int shift = PAGE_SHIFT;
3246
3247 if (WARN_ON_ONCE(!size))
3248 return NULL;
3249
3250 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3251 warn_alloc(gfp_mask, NULL,
3252 "vmalloc error: size %lu, exceeds total pages",
3253 real_size);
3254 return NULL;
3255 }
3256
3257 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3258 unsigned long size_per_node;
3259
3260 /*
3261 * Try huge pages. Only try for PAGE_KERNEL allocations,
3262 * others like modules don't yet expect huge pages in
3263 * their allocations due to apply_to_page_range not
3264 * supporting them.
3265 */
3266
3267 size_per_node = size;
3268 if (node == NUMA_NO_NODE)
3269 size_per_node /= num_online_nodes();
3270 if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3271 shift = PMD_SHIFT;
3272 else
3273 shift = arch_vmap_pte_supported_shift(size_per_node);
3274
3275 align = max(real_align, 1UL << shift);
3276 size = ALIGN(real_size, 1UL << shift);
3277 }
3278
3279again:
3280 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3281 VM_UNINITIALIZED | vm_flags, start, end, node,
3282 gfp_mask, caller);
3283 if (!area) {
3284 bool nofail = gfp_mask & __GFP_NOFAIL;
3285 warn_alloc(gfp_mask, NULL,
3286 "vmalloc error: size %lu, vm_struct allocation failed%s",
3287 real_size, (nofail) ? ". Retrying." : "");
3288 if (nofail) {
3289 schedule_timeout_uninterruptible(1);
3290 goto again;
3291 }
3292 goto fail;
3293 }
3294
3295 /*
3296 * Prepare arguments for __vmalloc_area_node() and
3297 * kasan_unpoison_vmalloc().
3298 */
3299 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3300 if (kasan_hw_tags_enabled()) {
3301 /*
3302 * Modify protection bits to allow tagging.
3303 * This must be done before mapping.
3304 */
3305 prot = arch_vmap_pgprot_tagged(prot);
3306
3307 /*
3308 * Skip page_alloc poisoning and zeroing for physical
3309 * pages backing VM_ALLOC mapping. Memory is instead
3310 * poisoned and zeroed by kasan_unpoison_vmalloc().
3311 */
3312 gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3313 }
3314
3315 /* Take note that the mapping is PAGE_KERNEL. */
3316 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3317 }
3318
3319 /* Allocate physical pages and map them into vmalloc space. */
3320 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3321 if (!ret)
3322 goto fail;
3323
3324 /*
3325 * Mark the pages as accessible, now that they are mapped.
3326 * The condition for setting KASAN_VMALLOC_INIT should complement the
3327 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3328 * to make sure that memory is initialized under the same conditions.
3329 * Tag-based KASAN modes only assign tags to normal non-executable
3330 * allocations, see __kasan_unpoison_vmalloc().
3331 */
3332 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3333 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3334 (gfp_mask & __GFP_SKIP_ZERO))
3335 kasan_flags |= KASAN_VMALLOC_INIT;
3336 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3337 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3338
3339 /*
3340 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3341 * flag. It means that vm_struct is not fully initialized.
3342 * Now, it is fully initialized, so remove this flag here.
3343 */
3344 clear_vm_uninitialized_flag(area);
3345
3346 size = PAGE_ALIGN(size);
3347 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3348 kmemleak_vmalloc(area, size, gfp_mask);
3349
3350 return area->addr;
3351
3352fail:
3353 if (shift > PAGE_SHIFT) {
3354 shift = PAGE_SHIFT;
3355 align = real_align;
3356 size = real_size;
3357 goto again;
3358 }
3359
3360 return NULL;
3361}
3362
3363/**
3364 * __vmalloc_node - allocate virtually contiguous memory
3365 * @size: allocation size
3366 * @align: desired alignment
3367 * @gfp_mask: flags for the page level allocator
3368 * @node: node to use for allocation or NUMA_NO_NODE
3369 * @caller: caller's return address
3370 *
3371 * Allocate enough pages to cover @size from the page level allocator with
3372 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3373 *
3374 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3375 * and __GFP_NOFAIL are not supported
3376 *
3377 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3378 * with mm people.
3379 *
3380 * Return: pointer to the allocated memory or %NULL on error
3381 */
3382void *__vmalloc_node(unsigned long size, unsigned long align,
3383 gfp_t gfp_mask, int node, const void *caller)
3384{
3385 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3386 gfp_mask, PAGE_KERNEL, 0, node, caller);
3387}
3388/*
3389 * This is only for performance analysis of vmalloc and stress purpose.
3390 * It is required by vmalloc test module, therefore do not use it other
3391 * than that.
3392 */
3393#ifdef CONFIG_TEST_VMALLOC_MODULE
3394EXPORT_SYMBOL_GPL(__vmalloc_node);
3395#endif
3396
3397void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3398{
3399 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3400 __builtin_return_address(0));
3401}
3402EXPORT_SYMBOL(__vmalloc);
3403
3404/**
3405 * vmalloc - allocate virtually contiguous memory
3406 * @size: allocation size
3407 *
3408 * Allocate enough pages to cover @size from the page level
3409 * allocator and map them into contiguous kernel virtual space.
3410 *
3411 * For tight control over page level allocator and protection flags
3412 * use __vmalloc() instead.
3413 *
3414 * Return: pointer to the allocated memory or %NULL on error
3415 */
3416void *vmalloc(unsigned long size)
3417{
3418 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3419 __builtin_return_address(0));
3420}
3421EXPORT_SYMBOL(vmalloc);
3422
3423/**
3424 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3425 * @size: allocation size
3426 * @gfp_mask: flags for the page level allocator
3427 *
3428 * Allocate enough pages to cover @size from the page level
3429 * allocator and map them into contiguous kernel virtual space.
3430 * If @size is greater than or equal to PMD_SIZE, allow using
3431 * huge pages for the memory
3432 *
3433 * Return: pointer to the allocated memory or %NULL on error
3434 */
3435void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3436{
3437 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3438 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3439 NUMA_NO_NODE, __builtin_return_address(0));
3440}
3441EXPORT_SYMBOL_GPL(vmalloc_huge);
3442
3443/**
3444 * vzalloc - allocate virtually contiguous memory with zero fill
3445 * @size: allocation size
3446 *
3447 * Allocate enough pages to cover @size from the page level
3448 * allocator and map them into contiguous kernel virtual space.
3449 * The memory allocated is set to zero.
3450 *
3451 * For tight control over page level allocator and protection flags
3452 * use __vmalloc() instead.
3453 *
3454 * Return: pointer to the allocated memory or %NULL on error
3455 */
3456void *vzalloc(unsigned long size)
3457{
3458 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3459 __builtin_return_address(0));
3460}
3461EXPORT_SYMBOL(vzalloc);
3462
3463/**
3464 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3465 * @size: allocation size
3466 *
3467 * The resulting memory area is zeroed so it can be mapped to userspace
3468 * without leaking data.
3469 *
3470 * Return: pointer to the allocated memory or %NULL on error
3471 */
3472void *vmalloc_user(unsigned long size)
3473{
3474 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3475 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3476 VM_USERMAP, NUMA_NO_NODE,
3477 __builtin_return_address(0));
3478}
3479EXPORT_SYMBOL(vmalloc_user);
3480
3481/**
3482 * vmalloc_node - allocate memory on a specific node
3483 * @size: allocation size
3484 * @node: numa node
3485 *
3486 * Allocate enough pages to cover @size from the page level
3487 * allocator and map them into contiguous kernel virtual space.
3488 *
3489 * For tight control over page level allocator and protection flags
3490 * use __vmalloc() instead.
3491 *
3492 * Return: pointer to the allocated memory or %NULL on error
3493 */
3494void *vmalloc_node(unsigned long size, int node)
3495{
3496 return __vmalloc_node(size, 1, GFP_KERNEL, node,
3497 __builtin_return_address(0));
3498}
3499EXPORT_SYMBOL(vmalloc_node);
3500
3501/**
3502 * vzalloc_node - allocate memory on a specific node with zero fill
3503 * @size: allocation size
3504 * @node: numa node
3505 *
3506 * Allocate enough pages to cover @size from the page level
3507 * allocator and map them into contiguous kernel virtual space.
3508 * The memory allocated is set to zero.
3509 *
3510 * Return: pointer to the allocated memory or %NULL on error
3511 */
3512void *vzalloc_node(unsigned long size, int node)
3513{
3514 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3515 __builtin_return_address(0));
3516}
3517EXPORT_SYMBOL(vzalloc_node);
3518
3519#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3520#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3521#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3522#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3523#else
3524/*
3525 * 64b systems should always have either DMA or DMA32 zones. For others
3526 * GFP_DMA32 should do the right thing and use the normal zone.
3527 */
3528#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3529#endif
3530
3531/**
3532 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3533 * @size: allocation size
3534 *
3535 * Allocate enough 32bit PA addressable pages to cover @size from the
3536 * page level allocator and map them into contiguous kernel virtual space.
3537 *
3538 * Return: pointer to the allocated memory or %NULL on error
3539 */
3540void *vmalloc_32(unsigned long size)
3541{
3542 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3543 __builtin_return_address(0));
3544}
3545EXPORT_SYMBOL(vmalloc_32);
3546
3547/**
3548 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3549 * @size: allocation size
3550 *
3551 * The resulting memory area is 32bit addressable and zeroed so it can be
3552 * mapped to userspace without leaking data.
3553 *
3554 * Return: pointer to the allocated memory or %NULL on error
3555 */
3556void *vmalloc_32_user(unsigned long size)
3557{
3558 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
3559 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3560 VM_USERMAP, NUMA_NO_NODE,
3561 __builtin_return_address(0));
3562}
3563EXPORT_SYMBOL(vmalloc_32_user);
3564
3565/*
3566 * Atomically zero bytes in the iterator.
3567 *
3568 * Returns the number of zeroed bytes.
3569 */
3570static size_t zero_iter(struct iov_iter *iter, size_t count)
3571{
3572 size_t remains = count;
3573
3574 while (remains > 0) {
3575 size_t num, copied;
3576
3577 num = min_t(size_t, remains, PAGE_SIZE);
3578 copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
3579 remains -= copied;
3580
3581 if (copied < num)
3582 break;
3583 }
3584
3585 return count - remains;
3586}
3587
3588/*
3589 * small helper routine, copy contents to iter from addr.
3590 * If the page is not present, fill zero.
3591 *
3592 * Returns the number of copied bytes.
3593 */
3594static size_t aligned_vread_iter(struct iov_iter *iter,
3595 const char *addr, size_t count)
3596{
3597 size_t remains = count;
3598 struct page *page;
3599
3600 while (remains > 0) {
3601 unsigned long offset, length;
3602 size_t copied = 0;
3603
3604 offset = offset_in_page(addr);
3605 length = PAGE_SIZE - offset;
3606 if (length > remains)
3607 length = remains;
3608 page = vmalloc_to_page(addr);
3609 /*
3610 * To do safe access to this _mapped_ area, we need lock. But
3611 * adding lock here means that we need to add overhead of
3612 * vmalloc()/vfree() calls for this _debug_ interface, rarely
3613 * used. Instead of that, we'll use an local mapping via
3614 * copy_page_to_iter_nofault() and accept a small overhead in
3615 * this access function.
3616 */
3617 if (page)
3618 copied = copy_page_to_iter_nofault(page, offset,
3619 length, iter);
3620 else
3621 copied = zero_iter(iter, length);
3622
3623 addr += copied;
3624 remains -= copied;
3625
3626 if (copied != length)
3627 break;
3628 }
3629
3630 return count - remains;
3631}
3632
3633/*
3634 * Read from a vm_map_ram region of memory.
3635 *
3636 * Returns the number of copied bytes.
3637 */
3638static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
3639 size_t count, unsigned long flags)
3640{
3641 char *start;
3642 struct vmap_block *vb;
3643 struct xarray *xa;
3644 unsigned long offset;
3645 unsigned int rs, re;
3646 size_t remains, n;
3647
3648 /*
3649 * If it's area created by vm_map_ram() interface directly, but
3650 * not further subdividing and delegating management to vmap_block,
3651 * handle it here.
3652 */
3653 if (!(flags & VMAP_BLOCK))
3654 return aligned_vread_iter(iter, addr, count);
3655
3656 remains = count;
3657
3658 /*
3659 * Area is split into regions and tracked with vmap_block, read out
3660 * each region and zero fill the hole between regions.
3661 */
3662 xa = addr_to_vb_xa((unsigned long) addr);
3663 vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
3664 if (!vb)
3665 goto finished_zero;
3666
3667 spin_lock(&vb->lock);
3668 if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
3669 spin_unlock(&vb->lock);
3670 goto finished_zero;
3671 }
3672
3673 for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
3674 size_t copied;
3675
3676 if (remains == 0)
3677 goto finished;
3678
3679 start = vmap_block_vaddr(vb->va->va_start, rs);
3680
3681 if (addr < start) {
3682 size_t to_zero = min_t(size_t, start - addr, remains);
3683 size_t zeroed = zero_iter(iter, to_zero);
3684
3685 addr += zeroed;
3686 remains -= zeroed;
3687
3688 if (remains == 0 || zeroed != to_zero)
3689 goto finished;
3690 }
3691
3692 /*it could start reading from the middle of used region*/
3693 offset = offset_in_page(addr);
3694 n = ((re - rs + 1) << PAGE_SHIFT) - offset;
3695 if (n > remains)
3696 n = remains;
3697
3698 copied = aligned_vread_iter(iter, start + offset, n);
3699
3700 addr += copied;
3701 remains -= copied;
3702
3703 if (copied != n)
3704 goto finished;
3705 }
3706
3707 spin_unlock(&vb->lock);
3708
3709finished_zero:
3710 /* zero-fill the left dirty or free regions */
3711 return count - remains + zero_iter(iter, remains);
3712finished:
3713 /* We couldn't copy/zero everything */
3714 spin_unlock(&vb->lock);
3715 return count - remains;
3716}
3717
3718/**
3719 * vread_iter() - read vmalloc area in a safe way to an iterator.
3720 * @iter: the iterator to which data should be written.
3721 * @addr: vm address.
3722 * @count: number of bytes to be read.
3723 *
3724 * This function checks that addr is a valid vmalloc'ed area, and
3725 * copy data from that area to a given buffer. If the given memory range
3726 * of [addr...addr+count) includes some valid address, data is copied to
3727 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3728 * IOREMAP area is treated as memory hole and no copy is done.
3729 *
3730 * If [addr...addr+count) doesn't includes any intersects with alive
3731 * vm_struct area, returns 0. @buf should be kernel's buffer.
3732 *
3733 * Note: In usual ops, vread() is never necessary because the caller
3734 * should know vmalloc() area is valid and can use memcpy().
3735 * This is for routines which have to access vmalloc area without
3736 * any information, as /proc/kcore.
3737 *
3738 * Return: number of bytes for which addr and buf should be increased
3739 * (same number as @count) or %0 if [addr...addr+count) doesn't
3740 * include any intersection with valid vmalloc area
3741 */
3742long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
3743{
3744 struct vmap_area *va;
3745 struct vm_struct *vm;
3746 char *vaddr;
3747 size_t n, size, flags, remains;
3748
3749 addr = kasan_reset_tag(addr);
3750
3751 /* Don't allow overflow */
3752 if ((unsigned long) addr + count < count)
3753 count = -(unsigned long) addr;
3754
3755 remains = count;
3756
3757 spin_lock(&vmap_area_lock);
3758 va = find_vmap_area_exceed_addr((unsigned long)addr);
3759 if (!va)
3760 goto finished_zero;
3761
3762 /* no intersects with alive vmap_area */
3763 if ((unsigned long)addr + remains <= va->va_start)
3764 goto finished_zero;
3765
3766 list_for_each_entry_from(va, &vmap_area_list, list) {
3767 size_t copied;
3768
3769 if (remains == 0)
3770 goto finished;
3771
3772 vm = va->vm;
3773 flags = va->flags & VMAP_FLAGS_MASK;
3774 /*
3775 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
3776 * be set together with VMAP_RAM.
3777 */
3778 WARN_ON(flags == VMAP_BLOCK);
3779
3780 if (!vm && !flags)
3781 continue;
3782
3783 if (vm && (vm->flags & VM_UNINITIALIZED))
3784 continue;
3785
3786 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3787 smp_rmb();
3788
3789 vaddr = (char *) va->va_start;
3790 size = vm ? get_vm_area_size(vm) : va_size(va);
3791
3792 if (addr >= vaddr + size)
3793 continue;
3794
3795 if (addr < vaddr) {
3796 size_t to_zero = min_t(size_t, vaddr - addr, remains);
3797 size_t zeroed = zero_iter(iter, to_zero);
3798
3799 addr += zeroed;
3800 remains -= zeroed;
3801
3802 if (remains == 0 || zeroed != to_zero)
3803 goto finished;
3804 }
3805
3806 n = vaddr + size - addr;
3807 if (n > remains)
3808 n = remains;
3809
3810 if (flags & VMAP_RAM)
3811 copied = vmap_ram_vread_iter(iter, addr, n, flags);
3812 else if (!(vm && (vm->flags & VM_IOREMAP)))
3813 copied = aligned_vread_iter(iter, addr, n);
3814 else /* IOREMAP area is treated as memory hole */
3815 copied = zero_iter(iter, n);
3816
3817 addr += copied;
3818 remains -= copied;
3819
3820 if (copied != n)
3821 goto finished;
3822 }
3823
3824finished_zero:
3825 spin_unlock(&vmap_area_lock);
3826 /* zero-fill memory holes */
3827 return count - remains + zero_iter(iter, remains);
3828finished:
3829 /* Nothing remains, or We couldn't copy/zero everything. */
3830 spin_unlock(&vmap_area_lock);
3831
3832 return count - remains;
3833}
3834
3835/**
3836 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3837 * @vma: vma to cover
3838 * @uaddr: target user address to start at
3839 * @kaddr: virtual address of vmalloc kernel memory
3840 * @pgoff: offset from @kaddr to start at
3841 * @size: size of map area
3842 *
3843 * Returns: 0 for success, -Exxx on failure
3844 *
3845 * This function checks that @kaddr is a valid vmalloc'ed area,
3846 * and that it is big enough to cover the range starting at
3847 * @uaddr in @vma. Will return failure if that criteria isn't
3848 * met.
3849 *
3850 * Similar to remap_pfn_range() (see mm/memory.c)
3851 */
3852int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3853 void *kaddr, unsigned long pgoff,
3854 unsigned long size)
3855{
3856 struct vm_struct *area;
3857 unsigned long off;
3858 unsigned long end_index;
3859
3860 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3861 return -EINVAL;
3862
3863 size = PAGE_ALIGN(size);
3864
3865 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3866 return -EINVAL;
3867
3868 area = find_vm_area(kaddr);
3869 if (!area)
3870 return -EINVAL;
3871
3872 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3873 return -EINVAL;
3874
3875 if (check_add_overflow(size, off, &end_index) ||
3876 end_index > get_vm_area_size(area))
3877 return -EINVAL;
3878 kaddr += off;
3879
3880 do {
3881 struct page *page = vmalloc_to_page(kaddr);
3882 int ret;
3883
3884 ret = vm_insert_page(vma, uaddr, page);
3885 if (ret)
3886 return ret;
3887
3888 uaddr += PAGE_SIZE;
3889 kaddr += PAGE_SIZE;
3890 size -= PAGE_SIZE;
3891 } while (size > 0);
3892
3893 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
3894
3895 return 0;
3896}
3897
3898/**
3899 * remap_vmalloc_range - map vmalloc pages to userspace
3900 * @vma: vma to cover (map full range of vma)
3901 * @addr: vmalloc memory
3902 * @pgoff: number of pages into addr before first page to map
3903 *
3904 * Returns: 0 for success, -Exxx on failure
3905 *
3906 * This function checks that addr is a valid vmalloc'ed area, and
3907 * that it is big enough to cover the vma. Will return failure if
3908 * that criteria isn't met.
3909 *
3910 * Similar to remap_pfn_range() (see mm/memory.c)
3911 */
3912int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3913 unsigned long pgoff)
3914{
3915 return remap_vmalloc_range_partial(vma, vma->vm_start,
3916 addr, pgoff,
3917 vma->vm_end - vma->vm_start);
3918}
3919EXPORT_SYMBOL(remap_vmalloc_range);
3920
3921void free_vm_area(struct vm_struct *area)
3922{
3923 struct vm_struct *ret;
3924 ret = remove_vm_area(area->addr);
3925 BUG_ON(ret != area);
3926 kfree(area);
3927}
3928EXPORT_SYMBOL_GPL(free_vm_area);
3929
3930#ifdef CONFIG_SMP
3931static struct vmap_area *node_to_va(struct rb_node *n)
3932{
3933 return rb_entry_safe(n, struct vmap_area, rb_node);
3934}
3935
3936/**
3937 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3938 * @addr: target address
3939 *
3940 * Returns: vmap_area if it is found. If there is no such area
3941 * the first highest(reverse order) vmap_area is returned
3942 * i.e. va->va_start < addr && va->va_end < addr or NULL
3943 * if there are no any areas before @addr.
3944 */
3945static struct vmap_area *
3946pvm_find_va_enclose_addr(unsigned long addr)
3947{
3948 struct vmap_area *va, *tmp;
3949 struct rb_node *n;
3950
3951 n = free_vmap_area_root.rb_node;
3952 va = NULL;
3953
3954 while (n) {
3955 tmp = rb_entry(n, struct vmap_area, rb_node);
3956 if (tmp->va_start <= addr) {
3957 va = tmp;
3958 if (tmp->va_end >= addr)
3959 break;
3960
3961 n = n->rb_right;
3962 } else {
3963 n = n->rb_left;
3964 }
3965 }
3966
3967 return va;
3968}
3969
3970/**
3971 * pvm_determine_end_from_reverse - find the highest aligned address
3972 * of free block below VMALLOC_END
3973 * @va:
3974 * in - the VA we start the search(reverse order);
3975 * out - the VA with the highest aligned end address.
3976 * @align: alignment for required highest address
3977 *
3978 * Returns: determined end address within vmap_area
3979 */
3980static unsigned long
3981pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3982{
3983 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3984 unsigned long addr;
3985
3986 if (likely(*va)) {
3987 list_for_each_entry_from_reverse((*va),
3988 &free_vmap_area_list, list) {
3989 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3990 if ((*va)->va_start < addr)
3991 return addr;
3992 }
3993 }
3994
3995 return 0;
3996}
3997
3998/**
3999 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4000 * @offsets: array containing offset of each area
4001 * @sizes: array containing size of each area
4002 * @nr_vms: the number of areas to allocate
4003 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4004 *
4005 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4006 * vm_structs on success, %NULL on failure
4007 *
4008 * Percpu allocator wants to use congruent vm areas so that it can
4009 * maintain the offsets among percpu areas. This function allocates
4010 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4011 * be scattered pretty far, distance between two areas easily going up
4012 * to gigabytes. To avoid interacting with regular vmallocs, these
4013 * areas are allocated from top.
4014 *
4015 * Despite its complicated look, this allocator is rather simple. It
4016 * does everything top-down and scans free blocks from the end looking
4017 * for matching base. While scanning, if any of the areas do not fit the
4018 * base address is pulled down to fit the area. Scanning is repeated till
4019 * all the areas fit and then all necessary data structures are inserted
4020 * and the result is returned.
4021 */
4022struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4023 const size_t *sizes, int nr_vms,
4024 size_t align)
4025{
4026 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4027 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4028 struct vmap_area **vas, *va;
4029 struct vm_struct **vms;
4030 int area, area2, last_area, term_area;
4031 unsigned long base, start, size, end, last_end, orig_start, orig_end;
4032 bool purged = false;
4033
4034 /* verify parameters and allocate data structures */
4035 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4036 for (last_area = 0, area = 0; area < nr_vms; area++) {
4037 start = offsets[area];
4038 end = start + sizes[area];
4039
4040 /* is everything aligned properly? */
4041 BUG_ON(!IS_ALIGNED(offsets[area], align));
4042 BUG_ON(!IS_ALIGNED(sizes[area], align));
4043
4044 /* detect the area with the highest address */
4045 if (start > offsets[last_area])
4046 last_area = area;
4047
4048 for (area2 = area + 1; area2 < nr_vms; area2++) {
4049 unsigned long start2 = offsets[area2];
4050 unsigned long end2 = start2 + sizes[area2];
4051
4052 BUG_ON(start2 < end && start < end2);
4053 }
4054 }
4055 last_end = offsets[last_area] + sizes[last_area];
4056
4057 if (vmalloc_end - vmalloc_start < last_end) {
4058 WARN_ON(true);
4059 return NULL;
4060 }
4061
4062 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4063 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4064 if (!vas || !vms)
4065 goto err_free2;
4066
4067 for (area = 0; area < nr_vms; area++) {
4068 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4069 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4070 if (!vas[area] || !vms[area])
4071 goto err_free;
4072 }
4073retry:
4074 spin_lock(&free_vmap_area_lock);
4075
4076 /* start scanning - we scan from the top, begin with the last area */
4077 area = term_area = last_area;
4078 start = offsets[area];
4079 end = start + sizes[area];
4080
4081 va = pvm_find_va_enclose_addr(vmalloc_end);
4082 base = pvm_determine_end_from_reverse(&va, align) - end;
4083
4084 while (true) {
4085 /*
4086 * base might have underflowed, add last_end before
4087 * comparing.
4088 */
4089 if (base + last_end < vmalloc_start + last_end)
4090 goto overflow;
4091
4092 /*
4093 * Fitting base has not been found.
4094 */
4095 if (va == NULL)
4096 goto overflow;
4097
4098 /*
4099 * If required width exceeds current VA block, move
4100 * base downwards and then recheck.
4101 */
4102 if (base + end > va->va_end) {
4103 base = pvm_determine_end_from_reverse(&va, align) - end;
4104 term_area = area;
4105 continue;
4106 }
4107
4108 /*
4109 * If this VA does not fit, move base downwards and recheck.
4110 */
4111 if (base + start < va->va_start) {
4112 va = node_to_va(rb_prev(&va->rb_node));
4113 base = pvm_determine_end_from_reverse(&va, align) - end;
4114 term_area = area;
4115 continue;
4116 }
4117
4118 /*
4119 * This area fits, move on to the previous one. If
4120 * the previous one is the terminal one, we're done.
4121 */
4122 area = (area + nr_vms - 1) % nr_vms;
4123 if (area == term_area)
4124 break;
4125
4126 start = offsets[area];
4127 end = start + sizes[area];
4128 va = pvm_find_va_enclose_addr(base + end);
4129 }
4130
4131 /* we've found a fitting base, insert all va's */
4132 for (area = 0; area < nr_vms; area++) {
4133 int ret;
4134
4135 start = base + offsets[area];
4136 size = sizes[area];
4137
4138 va = pvm_find_va_enclose_addr(start);
4139 if (WARN_ON_ONCE(va == NULL))
4140 /* It is a BUG(), but trigger recovery instead. */
4141 goto recovery;
4142
4143 ret = adjust_va_to_fit_type(&free_vmap_area_root,
4144 &free_vmap_area_list,
4145 va, start, size);
4146 if (WARN_ON_ONCE(unlikely(ret)))
4147 /* It is a BUG(), but trigger recovery instead. */
4148 goto recovery;
4149
4150 /* Allocated area. */
4151 va = vas[area];
4152 va->va_start = start;
4153 va->va_end = start + size;
4154 }
4155
4156 spin_unlock(&free_vmap_area_lock);
4157
4158 /* populate the kasan shadow space */
4159 for (area = 0; area < nr_vms; area++) {
4160 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4161 goto err_free_shadow;
4162 }
4163
4164 /* insert all vm's */
4165 spin_lock(&vmap_area_lock);
4166 for (area = 0; area < nr_vms; area++) {
4167 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
4168
4169 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
4170 pcpu_get_vm_areas);
4171 }
4172 spin_unlock(&vmap_area_lock);
4173
4174 /*
4175 * Mark allocated areas as accessible. Do it now as a best-effort
4176 * approach, as they can be mapped outside of vmalloc code.
4177 * With hardware tag-based KASAN, marking is skipped for
4178 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4179 */
4180 for (area = 0; area < nr_vms; area++)
4181 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4182 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4183
4184 kfree(vas);
4185 return vms;
4186
4187recovery:
4188 /*
4189 * Remove previously allocated areas. There is no
4190 * need in removing these areas from the busy tree,
4191 * because they are inserted only on the final step
4192 * and when pcpu_get_vm_areas() is success.
4193 */
4194 while (area--) {
4195 orig_start = vas[area]->va_start;
4196 orig_end = vas[area]->va_end;
4197 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4198 &free_vmap_area_list);
4199 if (va)
4200 kasan_release_vmalloc(orig_start, orig_end,
4201 va->va_start, va->va_end);
4202 vas[area] = NULL;
4203 }
4204
4205overflow:
4206 spin_unlock(&free_vmap_area_lock);
4207 if (!purged) {
4208 reclaim_and_purge_vmap_areas();
4209 purged = true;
4210
4211 /* Before "retry", check if we recover. */
4212 for (area = 0; area < nr_vms; area++) {
4213 if (vas[area])
4214 continue;
4215
4216 vas[area] = kmem_cache_zalloc(
4217 vmap_area_cachep, GFP_KERNEL);
4218 if (!vas[area])
4219 goto err_free;
4220 }
4221
4222 goto retry;
4223 }
4224
4225err_free:
4226 for (area = 0; area < nr_vms; area++) {
4227 if (vas[area])
4228 kmem_cache_free(vmap_area_cachep, vas[area]);
4229
4230 kfree(vms[area]);
4231 }
4232err_free2:
4233 kfree(vas);
4234 kfree(vms);
4235 return NULL;
4236
4237err_free_shadow:
4238 spin_lock(&free_vmap_area_lock);
4239 /*
4240 * We release all the vmalloc shadows, even the ones for regions that
4241 * hadn't been successfully added. This relies on kasan_release_vmalloc
4242 * being able to tolerate this case.
4243 */
4244 for (area = 0; area < nr_vms; area++) {
4245 orig_start = vas[area]->va_start;
4246 orig_end = vas[area]->va_end;
4247 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4248 &free_vmap_area_list);
4249 if (va)
4250 kasan_release_vmalloc(orig_start, orig_end,
4251 va->va_start, va->va_end);
4252 vas[area] = NULL;
4253 kfree(vms[area]);
4254 }
4255 spin_unlock(&free_vmap_area_lock);
4256 kfree(vas);
4257 kfree(vms);
4258 return NULL;
4259}
4260
4261/**
4262 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4263 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4264 * @nr_vms: the number of allocated areas
4265 *
4266 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4267 */
4268void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4269{
4270 int i;
4271
4272 for (i = 0; i < nr_vms; i++)
4273 free_vm_area(vms[i]);
4274 kfree(vms);
4275}
4276#endif /* CONFIG_SMP */
4277
4278#ifdef CONFIG_PRINTK
4279bool vmalloc_dump_obj(void *object)
4280{
4281 void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4282 const void *caller;
4283 struct vm_struct *vm;
4284 struct vmap_area *va;
4285 unsigned long addr;
4286 unsigned int nr_pages;
4287
4288 if (!spin_trylock(&vmap_area_lock))
4289 return false;
4290 va = __find_vmap_area((unsigned long)objp, &vmap_area_root);
4291 if (!va) {
4292 spin_unlock(&vmap_area_lock);
4293 return false;
4294 }
4295
4296 vm = va->vm;
4297 if (!vm) {
4298 spin_unlock(&vmap_area_lock);
4299 return false;
4300 }
4301 addr = (unsigned long)vm->addr;
4302 caller = vm->caller;
4303 nr_pages = vm->nr_pages;
4304 spin_unlock(&vmap_area_lock);
4305 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4306 nr_pages, addr, caller);
4307 return true;
4308}
4309#endif
4310
4311#ifdef CONFIG_PROC_FS
4312static void *s_start(struct seq_file *m, loff_t *pos)
4313 __acquires(&vmap_purge_lock)
4314 __acquires(&vmap_area_lock)
4315{
4316 mutex_lock(&vmap_purge_lock);
4317 spin_lock(&vmap_area_lock);
4318
4319 return seq_list_start(&vmap_area_list, *pos);
4320}
4321
4322static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4323{
4324 return seq_list_next(p, &vmap_area_list, pos);
4325}
4326
4327static void s_stop(struct seq_file *m, void *p)
4328 __releases(&vmap_area_lock)
4329 __releases(&vmap_purge_lock)
4330{
4331 spin_unlock(&vmap_area_lock);
4332 mutex_unlock(&vmap_purge_lock);
4333}
4334
4335static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4336{
4337 if (IS_ENABLED(CONFIG_NUMA)) {
4338 unsigned int nr, *counters = m->private;
4339 unsigned int step = 1U << vm_area_page_order(v);
4340
4341 if (!counters)
4342 return;
4343
4344 if (v->flags & VM_UNINITIALIZED)
4345 return;
4346 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4347 smp_rmb();
4348
4349 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4350
4351 for (nr = 0; nr < v->nr_pages; nr += step)
4352 counters[page_to_nid(v->pages[nr])] += step;
4353 for_each_node_state(nr, N_HIGH_MEMORY)
4354 if (counters[nr])
4355 seq_printf(m, " N%u=%u", nr, counters[nr]);
4356 }
4357}
4358
4359static void show_purge_info(struct seq_file *m)
4360{
4361 struct vmap_area *va;
4362
4363 spin_lock(&purge_vmap_area_lock);
4364 list_for_each_entry(va, &purge_vmap_area_list, list) {
4365 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4366 (void *)va->va_start, (void *)va->va_end,
4367 va->va_end - va->va_start);
4368 }
4369 spin_unlock(&purge_vmap_area_lock);
4370}
4371
4372static int s_show(struct seq_file *m, void *p)
4373{
4374 struct vmap_area *va;
4375 struct vm_struct *v;
4376
4377 va = list_entry(p, struct vmap_area, list);
4378
4379 if (!va->vm) {
4380 if (va->flags & VMAP_RAM)
4381 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4382 (void *)va->va_start, (void *)va->va_end,
4383 va->va_end - va->va_start);
4384
4385 goto final;
4386 }
4387
4388 v = va->vm;
4389
4390 seq_printf(m, "0x%pK-0x%pK %7ld",
4391 v->addr, v->addr + v->size, v->size);
4392
4393 if (v->caller)
4394 seq_printf(m, " %pS", v->caller);
4395
4396 if (v->nr_pages)
4397 seq_printf(m, " pages=%d", v->nr_pages);
4398
4399 if (v->phys_addr)
4400 seq_printf(m, " phys=%pa", &v->phys_addr);
4401
4402 if (v->flags & VM_IOREMAP)
4403 seq_puts(m, " ioremap");
4404
4405 if (v->flags & VM_ALLOC)
4406 seq_puts(m, " vmalloc");
4407
4408 if (v->flags & VM_MAP)
4409 seq_puts(m, " vmap");
4410
4411 if (v->flags & VM_USERMAP)
4412 seq_puts(m, " user");
4413
4414 if (v->flags & VM_DMA_COHERENT)
4415 seq_puts(m, " dma-coherent");
4416
4417 if (is_vmalloc_addr(v->pages))
4418 seq_puts(m, " vpages");
4419
4420 show_numa_info(m, v);
4421 seq_putc(m, '\n');
4422
4423 /*
4424 * As a final step, dump "unpurged" areas.
4425 */
4426final:
4427 if (list_is_last(&va->list, &vmap_area_list))
4428 show_purge_info(m);
4429
4430 return 0;
4431}
4432
4433static const struct seq_operations vmalloc_op = {
4434 .start = s_start,
4435 .next = s_next,
4436 .stop = s_stop,
4437 .show = s_show,
4438};
4439
4440static int __init proc_vmalloc_init(void)
4441{
4442 if (IS_ENABLED(CONFIG_NUMA))
4443 proc_create_seq_private("vmallocinfo", 0400, NULL,
4444 &vmalloc_op,
4445 nr_node_ids * sizeof(unsigned int), NULL);
4446 else
4447 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4448 return 0;
4449}
4450module_init(proc_vmalloc_init);
4451
4452#endif
4453
4454void __init vmalloc_init(void)
4455{
4456 struct vmap_area *va;
4457 struct vm_struct *tmp;
4458 int i;
4459
4460 /*
4461 * Create the cache for vmap_area objects.
4462 */
4463 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
4464
4465 for_each_possible_cpu(i) {
4466 struct vmap_block_queue *vbq;
4467 struct vfree_deferred *p;
4468
4469 vbq = &per_cpu(vmap_block_queue, i);
4470 spin_lock_init(&vbq->lock);
4471 INIT_LIST_HEAD(&vbq->free);
4472 p = &per_cpu(vfree_deferred, i);
4473 init_llist_head(&p->list);
4474 INIT_WORK(&p->wq, delayed_vfree_work);
4475 xa_init(&vbq->vmap_blocks);
4476 }
4477
4478 /* Import existing vmlist entries. */
4479 for (tmp = vmlist; tmp; tmp = tmp->next) {
4480 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4481 if (WARN_ON_ONCE(!va))
4482 continue;
4483
4484 va->va_start = (unsigned long)tmp->addr;
4485 va->va_end = va->va_start + tmp->size;
4486 va->vm = tmp;
4487 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
4488 }
4489
4490 /*
4491 * Now we can initialize a free vmap space.
4492 */
4493 vmap_init_free_space();
4494 vmap_initialized = true;
4495}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmalloc.c
4 *
5 * Copyright (C) 1993 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
8 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
9 * Numa awareness, Christoph Lameter, SGI, June 2005
10 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 */
12
13#include <linux/vmalloc.h>
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/highmem.h>
17#include <linux/sched/signal.h>
18#include <linux/slab.h>
19#include <linux/spinlock.h>
20#include <linux/interrupt.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/set_memory.h>
24#include <linux/debugobjects.h>
25#include <linux/kallsyms.h>
26#include <linux/list.h>
27#include <linux/notifier.h>
28#include <linux/rbtree.h>
29#include <linux/xarray.h>
30#include <linux/rcupdate.h>
31#include <linux/pfn.h>
32#include <linux/kmemleak.h>
33#include <linux/atomic.h>
34#include <linux/compiler.h>
35#include <linux/llist.h>
36#include <linux/bitops.h>
37#include <linux/rbtree_augmented.h>
38#include <linux/overflow.h>
39
40#include <linux/uaccess.h>
41#include <asm/tlbflush.h>
42#include <asm/shmparam.h>
43
44#include "internal.h"
45#include "pgalloc-track.h"
46
47bool is_vmalloc_addr(const void *x)
48{
49 unsigned long addr = (unsigned long)x;
50
51 return addr >= VMALLOC_START && addr < VMALLOC_END;
52}
53EXPORT_SYMBOL(is_vmalloc_addr);
54
55struct vfree_deferred {
56 struct llist_head list;
57 struct work_struct wq;
58};
59static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
60
61static void __vunmap(const void *, int);
62
63static void free_work(struct work_struct *w)
64{
65 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
66 struct llist_node *t, *llnode;
67
68 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
69 __vunmap((void *)llnode, 1);
70}
71
72/*** Page table manipulation functions ***/
73
74static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
75 pgtbl_mod_mask *mask)
76{
77 pte_t *pte;
78
79 pte = pte_offset_kernel(pmd, addr);
80 do {
81 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
82 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
83 } while (pte++, addr += PAGE_SIZE, addr != end);
84 *mask |= PGTBL_PTE_MODIFIED;
85}
86
87static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
88 pgtbl_mod_mask *mask)
89{
90 pmd_t *pmd;
91 unsigned long next;
92 int cleared;
93
94 pmd = pmd_offset(pud, addr);
95 do {
96 next = pmd_addr_end(addr, end);
97
98 cleared = pmd_clear_huge(pmd);
99 if (cleared || pmd_bad(*pmd))
100 *mask |= PGTBL_PMD_MODIFIED;
101
102 if (cleared)
103 continue;
104 if (pmd_none_or_clear_bad(pmd))
105 continue;
106 vunmap_pte_range(pmd, addr, next, mask);
107
108 cond_resched();
109 } while (pmd++, addr = next, addr != end);
110}
111
112static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
113 pgtbl_mod_mask *mask)
114{
115 pud_t *pud;
116 unsigned long next;
117 int cleared;
118
119 pud = pud_offset(p4d, addr);
120 do {
121 next = pud_addr_end(addr, end);
122
123 cleared = pud_clear_huge(pud);
124 if (cleared || pud_bad(*pud))
125 *mask |= PGTBL_PUD_MODIFIED;
126
127 if (cleared)
128 continue;
129 if (pud_none_or_clear_bad(pud))
130 continue;
131 vunmap_pmd_range(pud, addr, next, mask);
132 } while (pud++, addr = next, addr != end);
133}
134
135static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
136 pgtbl_mod_mask *mask)
137{
138 p4d_t *p4d;
139 unsigned long next;
140 int cleared;
141
142 p4d = p4d_offset(pgd, addr);
143 do {
144 next = p4d_addr_end(addr, end);
145
146 cleared = p4d_clear_huge(p4d);
147 if (cleared || p4d_bad(*p4d))
148 *mask |= PGTBL_P4D_MODIFIED;
149
150 if (cleared)
151 continue;
152 if (p4d_none_or_clear_bad(p4d))
153 continue;
154 vunmap_pud_range(p4d, addr, next, mask);
155 } while (p4d++, addr = next, addr != end);
156}
157
158/**
159 * unmap_kernel_range_noflush - unmap kernel VM area
160 * @start: start of the VM area to unmap
161 * @size: size of the VM area to unmap
162 *
163 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
164 * should have been allocated using get_vm_area() and its friends.
165 *
166 * NOTE:
167 * This function does NOT do any cache flushing. The caller is responsible
168 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
169 * function and flush_tlb_kernel_range() after.
170 */
171void unmap_kernel_range_noflush(unsigned long start, unsigned long size)
172{
173 unsigned long end = start + size;
174 unsigned long next;
175 pgd_t *pgd;
176 unsigned long addr = start;
177 pgtbl_mod_mask mask = 0;
178
179 BUG_ON(addr >= end);
180 pgd = pgd_offset_k(addr);
181 do {
182 next = pgd_addr_end(addr, end);
183 if (pgd_bad(*pgd))
184 mask |= PGTBL_PGD_MODIFIED;
185 if (pgd_none_or_clear_bad(pgd))
186 continue;
187 vunmap_p4d_range(pgd, addr, next, &mask);
188 } while (pgd++, addr = next, addr != end);
189
190 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
191 arch_sync_kernel_mappings(start, end);
192}
193
194static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
195 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
196 pgtbl_mod_mask *mask)
197{
198 pte_t *pte;
199
200 /*
201 * nr is a running index into the array which helps higher level
202 * callers keep track of where we're up to.
203 */
204
205 pte = pte_alloc_kernel_track(pmd, addr, mask);
206 if (!pte)
207 return -ENOMEM;
208 do {
209 struct page *page = pages[*nr];
210
211 if (WARN_ON(!pte_none(*pte)))
212 return -EBUSY;
213 if (WARN_ON(!page))
214 return -ENOMEM;
215 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
216 (*nr)++;
217 } while (pte++, addr += PAGE_SIZE, addr != end);
218 *mask |= PGTBL_PTE_MODIFIED;
219 return 0;
220}
221
222static int vmap_pmd_range(pud_t *pud, unsigned long addr,
223 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
224 pgtbl_mod_mask *mask)
225{
226 pmd_t *pmd;
227 unsigned long next;
228
229 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
230 if (!pmd)
231 return -ENOMEM;
232 do {
233 next = pmd_addr_end(addr, end);
234 if (vmap_pte_range(pmd, addr, next, prot, pages, nr, mask))
235 return -ENOMEM;
236 } while (pmd++, addr = next, addr != end);
237 return 0;
238}
239
240static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
241 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
242 pgtbl_mod_mask *mask)
243{
244 pud_t *pud;
245 unsigned long next;
246
247 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
248 if (!pud)
249 return -ENOMEM;
250 do {
251 next = pud_addr_end(addr, end);
252 if (vmap_pmd_range(pud, addr, next, prot, pages, nr, mask))
253 return -ENOMEM;
254 } while (pud++, addr = next, addr != end);
255 return 0;
256}
257
258static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
259 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
260 pgtbl_mod_mask *mask)
261{
262 p4d_t *p4d;
263 unsigned long next;
264
265 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
266 if (!p4d)
267 return -ENOMEM;
268 do {
269 next = p4d_addr_end(addr, end);
270 if (vmap_pud_range(p4d, addr, next, prot, pages, nr, mask))
271 return -ENOMEM;
272 } while (p4d++, addr = next, addr != end);
273 return 0;
274}
275
276/**
277 * map_kernel_range_noflush - map kernel VM area with the specified pages
278 * @addr: start of the VM area to map
279 * @size: size of the VM area to map
280 * @prot: page protection flags to use
281 * @pages: pages to map
282 *
283 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
284 * have been allocated using get_vm_area() and its friends.
285 *
286 * NOTE:
287 * This function does NOT do any cache flushing. The caller is responsible for
288 * calling flush_cache_vmap() on to-be-mapped areas before calling this
289 * function.
290 *
291 * RETURNS:
292 * 0 on success, -errno on failure.
293 */
294int map_kernel_range_noflush(unsigned long addr, unsigned long size,
295 pgprot_t prot, struct page **pages)
296{
297 unsigned long start = addr;
298 unsigned long end = addr + size;
299 unsigned long next;
300 pgd_t *pgd;
301 int err = 0;
302 int nr = 0;
303 pgtbl_mod_mask mask = 0;
304
305 BUG_ON(addr >= end);
306 pgd = pgd_offset_k(addr);
307 do {
308 next = pgd_addr_end(addr, end);
309 if (pgd_bad(*pgd))
310 mask |= PGTBL_PGD_MODIFIED;
311 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
312 if (err)
313 return err;
314 } while (pgd++, addr = next, addr != end);
315
316 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
317 arch_sync_kernel_mappings(start, end);
318
319 return 0;
320}
321
322int map_kernel_range(unsigned long start, unsigned long size, pgprot_t prot,
323 struct page **pages)
324{
325 int ret;
326
327 ret = map_kernel_range_noflush(start, size, prot, pages);
328 flush_cache_vmap(start, start + size);
329 return ret;
330}
331
332int is_vmalloc_or_module_addr(const void *x)
333{
334 /*
335 * ARM, x86-64 and sparc64 put modules in a special place,
336 * and fall back on vmalloc() if that fails. Others
337 * just put it in the vmalloc space.
338 */
339#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
340 unsigned long addr = (unsigned long)x;
341 if (addr >= MODULES_VADDR && addr < MODULES_END)
342 return 1;
343#endif
344 return is_vmalloc_addr(x);
345}
346
347/*
348 * Walk a vmap address to the struct page it maps.
349 */
350struct page *vmalloc_to_page(const void *vmalloc_addr)
351{
352 unsigned long addr = (unsigned long) vmalloc_addr;
353 struct page *page = NULL;
354 pgd_t *pgd = pgd_offset_k(addr);
355 p4d_t *p4d;
356 pud_t *pud;
357 pmd_t *pmd;
358 pte_t *ptep, pte;
359
360 /*
361 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
362 * architectures that do not vmalloc module space
363 */
364 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
365
366 if (pgd_none(*pgd))
367 return NULL;
368 p4d = p4d_offset(pgd, addr);
369 if (p4d_none(*p4d))
370 return NULL;
371 pud = pud_offset(p4d, addr);
372
373 /*
374 * Don't dereference bad PUD or PMD (below) entries. This will also
375 * identify huge mappings, which we may encounter on architectures
376 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
377 * identified as vmalloc addresses by is_vmalloc_addr(), but are
378 * not [unambiguously] associated with a struct page, so there is
379 * no correct value to return for them.
380 */
381 WARN_ON_ONCE(pud_bad(*pud));
382 if (pud_none(*pud) || pud_bad(*pud))
383 return NULL;
384 pmd = pmd_offset(pud, addr);
385 WARN_ON_ONCE(pmd_bad(*pmd));
386 if (pmd_none(*pmd) || pmd_bad(*pmd))
387 return NULL;
388
389 ptep = pte_offset_map(pmd, addr);
390 pte = *ptep;
391 if (pte_present(pte))
392 page = pte_page(pte);
393 pte_unmap(ptep);
394 return page;
395}
396EXPORT_SYMBOL(vmalloc_to_page);
397
398/*
399 * Map a vmalloc()-space virtual address to the physical page frame number.
400 */
401unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
402{
403 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
404}
405EXPORT_SYMBOL(vmalloc_to_pfn);
406
407
408/*** Global kva allocator ***/
409
410#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
411#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
412
413
414static DEFINE_SPINLOCK(vmap_area_lock);
415static DEFINE_SPINLOCK(free_vmap_area_lock);
416/* Export for kexec only */
417LIST_HEAD(vmap_area_list);
418static LLIST_HEAD(vmap_purge_list);
419static struct rb_root vmap_area_root = RB_ROOT;
420static bool vmap_initialized __read_mostly;
421
422/*
423 * This kmem_cache is used for vmap_area objects. Instead of
424 * allocating from slab we reuse an object from this cache to
425 * make things faster. Especially in "no edge" splitting of
426 * free block.
427 */
428static struct kmem_cache *vmap_area_cachep;
429
430/*
431 * This linked list is used in pair with free_vmap_area_root.
432 * It gives O(1) access to prev/next to perform fast coalescing.
433 */
434static LIST_HEAD(free_vmap_area_list);
435
436/*
437 * This augment red-black tree represents the free vmap space.
438 * All vmap_area objects in this tree are sorted by va->va_start
439 * address. It is used for allocation and merging when a vmap
440 * object is released.
441 *
442 * Each vmap_area node contains a maximum available free block
443 * of its sub-tree, right or left. Therefore it is possible to
444 * find a lowest match of free area.
445 */
446static struct rb_root free_vmap_area_root = RB_ROOT;
447
448/*
449 * Preload a CPU with one object for "no edge" split case. The
450 * aim is to get rid of allocations from the atomic context, thus
451 * to use more permissive allocation masks.
452 */
453static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
454
455static __always_inline unsigned long
456va_size(struct vmap_area *va)
457{
458 return (va->va_end - va->va_start);
459}
460
461static __always_inline unsigned long
462get_subtree_max_size(struct rb_node *node)
463{
464 struct vmap_area *va;
465
466 va = rb_entry_safe(node, struct vmap_area, rb_node);
467 return va ? va->subtree_max_size : 0;
468}
469
470/*
471 * Gets called when remove the node and rotate.
472 */
473static __always_inline unsigned long
474compute_subtree_max_size(struct vmap_area *va)
475{
476 return max3(va_size(va),
477 get_subtree_max_size(va->rb_node.rb_left),
478 get_subtree_max_size(va->rb_node.rb_right));
479}
480
481RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
482 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
483
484static void purge_vmap_area_lazy(void);
485static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
486static unsigned long lazy_max_pages(void);
487
488static atomic_long_t nr_vmalloc_pages;
489
490unsigned long vmalloc_nr_pages(void)
491{
492 return atomic_long_read(&nr_vmalloc_pages);
493}
494
495static struct vmap_area *__find_vmap_area(unsigned long addr)
496{
497 struct rb_node *n = vmap_area_root.rb_node;
498
499 while (n) {
500 struct vmap_area *va;
501
502 va = rb_entry(n, struct vmap_area, rb_node);
503 if (addr < va->va_start)
504 n = n->rb_left;
505 else if (addr >= va->va_end)
506 n = n->rb_right;
507 else
508 return va;
509 }
510
511 return NULL;
512}
513
514/*
515 * This function returns back addresses of parent node
516 * and its left or right link for further processing.
517 *
518 * Otherwise NULL is returned. In that case all further
519 * steps regarding inserting of conflicting overlap range
520 * have to be declined and actually considered as a bug.
521 */
522static __always_inline struct rb_node **
523find_va_links(struct vmap_area *va,
524 struct rb_root *root, struct rb_node *from,
525 struct rb_node **parent)
526{
527 struct vmap_area *tmp_va;
528 struct rb_node **link;
529
530 if (root) {
531 link = &root->rb_node;
532 if (unlikely(!*link)) {
533 *parent = NULL;
534 return link;
535 }
536 } else {
537 link = &from;
538 }
539
540 /*
541 * Go to the bottom of the tree. When we hit the last point
542 * we end up with parent rb_node and correct direction, i name
543 * it link, where the new va->rb_node will be attached to.
544 */
545 do {
546 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
547
548 /*
549 * During the traversal we also do some sanity check.
550 * Trigger the BUG() if there are sides(left/right)
551 * or full overlaps.
552 */
553 if (va->va_start < tmp_va->va_end &&
554 va->va_end <= tmp_va->va_start)
555 link = &(*link)->rb_left;
556 else if (va->va_end > tmp_va->va_start &&
557 va->va_start >= tmp_va->va_end)
558 link = &(*link)->rb_right;
559 else {
560 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
561 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
562
563 return NULL;
564 }
565 } while (*link);
566
567 *parent = &tmp_va->rb_node;
568 return link;
569}
570
571static __always_inline struct list_head *
572get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
573{
574 struct list_head *list;
575
576 if (unlikely(!parent))
577 /*
578 * The red-black tree where we try to find VA neighbors
579 * before merging or inserting is empty, i.e. it means
580 * there is no free vmap space. Normally it does not
581 * happen but we handle this case anyway.
582 */
583 return NULL;
584
585 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
586 return (&parent->rb_right == link ? list->next : list);
587}
588
589static __always_inline void
590link_va(struct vmap_area *va, struct rb_root *root,
591 struct rb_node *parent, struct rb_node **link, struct list_head *head)
592{
593 /*
594 * VA is still not in the list, but we can
595 * identify its future previous list_head node.
596 */
597 if (likely(parent)) {
598 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
599 if (&parent->rb_right != link)
600 head = head->prev;
601 }
602
603 /* Insert to the rb-tree */
604 rb_link_node(&va->rb_node, parent, link);
605 if (root == &free_vmap_area_root) {
606 /*
607 * Some explanation here. Just perform simple insertion
608 * to the tree. We do not set va->subtree_max_size to
609 * its current size before calling rb_insert_augmented().
610 * It is because of we populate the tree from the bottom
611 * to parent levels when the node _is_ in the tree.
612 *
613 * Therefore we set subtree_max_size to zero after insertion,
614 * to let __augment_tree_propagate_from() puts everything to
615 * the correct order later on.
616 */
617 rb_insert_augmented(&va->rb_node,
618 root, &free_vmap_area_rb_augment_cb);
619 va->subtree_max_size = 0;
620 } else {
621 rb_insert_color(&va->rb_node, root);
622 }
623
624 /* Address-sort this list */
625 list_add(&va->list, head);
626}
627
628static __always_inline void
629unlink_va(struct vmap_area *va, struct rb_root *root)
630{
631 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
632 return;
633
634 if (root == &free_vmap_area_root)
635 rb_erase_augmented(&va->rb_node,
636 root, &free_vmap_area_rb_augment_cb);
637 else
638 rb_erase(&va->rb_node, root);
639
640 list_del(&va->list);
641 RB_CLEAR_NODE(&va->rb_node);
642}
643
644#if DEBUG_AUGMENT_PROPAGATE_CHECK
645static void
646augment_tree_propagate_check(void)
647{
648 struct vmap_area *va;
649 unsigned long computed_size;
650
651 list_for_each_entry(va, &free_vmap_area_list, list) {
652 computed_size = compute_subtree_max_size(va);
653 if (computed_size != va->subtree_max_size)
654 pr_emerg("tree is corrupted: %lu, %lu\n",
655 va_size(va), va->subtree_max_size);
656 }
657}
658#endif
659
660/*
661 * This function populates subtree_max_size from bottom to upper
662 * levels starting from VA point. The propagation must be done
663 * when VA size is modified by changing its va_start/va_end. Or
664 * in case of newly inserting of VA to the tree.
665 *
666 * It means that __augment_tree_propagate_from() must be called:
667 * - After VA has been inserted to the tree(free path);
668 * - After VA has been shrunk(allocation path);
669 * - After VA has been increased(merging path).
670 *
671 * Please note that, it does not mean that upper parent nodes
672 * and their subtree_max_size are recalculated all the time up
673 * to the root node.
674 *
675 * 4--8
676 * /\
677 * / \
678 * / \
679 * 2--2 8--8
680 *
681 * For example if we modify the node 4, shrinking it to 2, then
682 * no any modification is required. If we shrink the node 2 to 1
683 * its subtree_max_size is updated only, and set to 1. If we shrink
684 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
685 * node becomes 4--6.
686 */
687static __always_inline void
688augment_tree_propagate_from(struct vmap_area *va)
689{
690 /*
691 * Populate the tree from bottom towards the root until
692 * the calculated maximum available size of checked node
693 * is equal to its current one.
694 */
695 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
696
697#if DEBUG_AUGMENT_PROPAGATE_CHECK
698 augment_tree_propagate_check();
699#endif
700}
701
702static void
703insert_vmap_area(struct vmap_area *va,
704 struct rb_root *root, struct list_head *head)
705{
706 struct rb_node **link;
707 struct rb_node *parent;
708
709 link = find_va_links(va, root, NULL, &parent);
710 if (link)
711 link_va(va, root, parent, link, head);
712}
713
714static void
715insert_vmap_area_augment(struct vmap_area *va,
716 struct rb_node *from, struct rb_root *root,
717 struct list_head *head)
718{
719 struct rb_node **link;
720 struct rb_node *parent;
721
722 if (from)
723 link = find_va_links(va, NULL, from, &parent);
724 else
725 link = find_va_links(va, root, NULL, &parent);
726
727 if (link) {
728 link_va(va, root, parent, link, head);
729 augment_tree_propagate_from(va);
730 }
731}
732
733/*
734 * Merge de-allocated chunk of VA memory with previous
735 * and next free blocks. If coalesce is not done a new
736 * free area is inserted. If VA has been merged, it is
737 * freed.
738 *
739 * Please note, it can return NULL in case of overlap
740 * ranges, followed by WARN() report. Despite it is a
741 * buggy behaviour, a system can be alive and keep
742 * ongoing.
743 */
744static __always_inline struct vmap_area *
745merge_or_add_vmap_area(struct vmap_area *va,
746 struct rb_root *root, struct list_head *head)
747{
748 struct vmap_area *sibling;
749 struct list_head *next;
750 struct rb_node **link;
751 struct rb_node *parent;
752 bool merged = false;
753
754 /*
755 * Find a place in the tree where VA potentially will be
756 * inserted, unless it is merged with its sibling/siblings.
757 */
758 link = find_va_links(va, root, NULL, &parent);
759 if (!link)
760 return NULL;
761
762 /*
763 * Get next node of VA to check if merging can be done.
764 */
765 next = get_va_next_sibling(parent, link);
766 if (unlikely(next == NULL))
767 goto insert;
768
769 /*
770 * start end
771 * | |
772 * |<------VA------>|<-----Next----->|
773 * | |
774 * start end
775 */
776 if (next != head) {
777 sibling = list_entry(next, struct vmap_area, list);
778 if (sibling->va_start == va->va_end) {
779 sibling->va_start = va->va_start;
780
781 /* Free vmap_area object. */
782 kmem_cache_free(vmap_area_cachep, va);
783
784 /* Point to the new merged area. */
785 va = sibling;
786 merged = true;
787 }
788 }
789
790 /*
791 * start end
792 * | |
793 * |<-----Prev----->|<------VA------>|
794 * | |
795 * start end
796 */
797 if (next->prev != head) {
798 sibling = list_entry(next->prev, struct vmap_area, list);
799 if (sibling->va_end == va->va_start) {
800 /*
801 * If both neighbors are coalesced, it is important
802 * to unlink the "next" node first, followed by merging
803 * with "previous" one. Otherwise the tree might not be
804 * fully populated if a sibling's augmented value is
805 * "normalized" because of rotation operations.
806 */
807 if (merged)
808 unlink_va(va, root);
809
810 sibling->va_end = va->va_end;
811
812 /* Free vmap_area object. */
813 kmem_cache_free(vmap_area_cachep, va);
814
815 /* Point to the new merged area. */
816 va = sibling;
817 merged = true;
818 }
819 }
820
821insert:
822 if (!merged)
823 link_va(va, root, parent, link, head);
824
825 /*
826 * Last step is to check and update the tree.
827 */
828 augment_tree_propagate_from(va);
829 return va;
830}
831
832static __always_inline bool
833is_within_this_va(struct vmap_area *va, unsigned long size,
834 unsigned long align, unsigned long vstart)
835{
836 unsigned long nva_start_addr;
837
838 if (va->va_start > vstart)
839 nva_start_addr = ALIGN(va->va_start, align);
840 else
841 nva_start_addr = ALIGN(vstart, align);
842
843 /* Can be overflowed due to big size or alignment. */
844 if (nva_start_addr + size < nva_start_addr ||
845 nva_start_addr < vstart)
846 return false;
847
848 return (nva_start_addr + size <= va->va_end);
849}
850
851/*
852 * Find the first free block(lowest start address) in the tree,
853 * that will accomplish the request corresponding to passing
854 * parameters.
855 */
856static __always_inline struct vmap_area *
857find_vmap_lowest_match(unsigned long size,
858 unsigned long align, unsigned long vstart)
859{
860 struct vmap_area *va;
861 struct rb_node *node;
862 unsigned long length;
863
864 /* Start from the root. */
865 node = free_vmap_area_root.rb_node;
866
867 /* Adjust the search size for alignment overhead. */
868 length = size + align - 1;
869
870 while (node) {
871 va = rb_entry(node, struct vmap_area, rb_node);
872
873 if (get_subtree_max_size(node->rb_left) >= length &&
874 vstart < va->va_start) {
875 node = node->rb_left;
876 } else {
877 if (is_within_this_va(va, size, align, vstart))
878 return va;
879
880 /*
881 * Does not make sense to go deeper towards the right
882 * sub-tree if it does not have a free block that is
883 * equal or bigger to the requested search length.
884 */
885 if (get_subtree_max_size(node->rb_right) >= length) {
886 node = node->rb_right;
887 continue;
888 }
889
890 /*
891 * OK. We roll back and find the first right sub-tree,
892 * that will satisfy the search criteria. It can happen
893 * only once due to "vstart" restriction.
894 */
895 while ((node = rb_parent(node))) {
896 va = rb_entry(node, struct vmap_area, rb_node);
897 if (is_within_this_va(va, size, align, vstart))
898 return va;
899
900 if (get_subtree_max_size(node->rb_right) >= length &&
901 vstart <= va->va_start) {
902 node = node->rb_right;
903 break;
904 }
905 }
906 }
907 }
908
909 return NULL;
910}
911
912#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
913#include <linux/random.h>
914
915static struct vmap_area *
916find_vmap_lowest_linear_match(unsigned long size,
917 unsigned long align, unsigned long vstart)
918{
919 struct vmap_area *va;
920
921 list_for_each_entry(va, &free_vmap_area_list, list) {
922 if (!is_within_this_va(va, size, align, vstart))
923 continue;
924
925 return va;
926 }
927
928 return NULL;
929}
930
931static void
932find_vmap_lowest_match_check(unsigned long size)
933{
934 struct vmap_area *va_1, *va_2;
935 unsigned long vstart;
936 unsigned int rnd;
937
938 get_random_bytes(&rnd, sizeof(rnd));
939 vstart = VMALLOC_START + rnd;
940
941 va_1 = find_vmap_lowest_match(size, 1, vstart);
942 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
943
944 if (va_1 != va_2)
945 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
946 va_1, va_2, vstart);
947}
948#endif
949
950enum fit_type {
951 NOTHING_FIT = 0,
952 FL_FIT_TYPE = 1, /* full fit */
953 LE_FIT_TYPE = 2, /* left edge fit */
954 RE_FIT_TYPE = 3, /* right edge fit */
955 NE_FIT_TYPE = 4 /* no edge fit */
956};
957
958static __always_inline enum fit_type
959classify_va_fit_type(struct vmap_area *va,
960 unsigned long nva_start_addr, unsigned long size)
961{
962 enum fit_type type;
963
964 /* Check if it is within VA. */
965 if (nva_start_addr < va->va_start ||
966 nva_start_addr + size > va->va_end)
967 return NOTHING_FIT;
968
969 /* Now classify. */
970 if (va->va_start == nva_start_addr) {
971 if (va->va_end == nva_start_addr + size)
972 type = FL_FIT_TYPE;
973 else
974 type = LE_FIT_TYPE;
975 } else if (va->va_end == nva_start_addr + size) {
976 type = RE_FIT_TYPE;
977 } else {
978 type = NE_FIT_TYPE;
979 }
980
981 return type;
982}
983
984static __always_inline int
985adjust_va_to_fit_type(struct vmap_area *va,
986 unsigned long nva_start_addr, unsigned long size,
987 enum fit_type type)
988{
989 struct vmap_area *lva = NULL;
990
991 if (type == FL_FIT_TYPE) {
992 /*
993 * No need to split VA, it fully fits.
994 *
995 * | |
996 * V NVA V
997 * |---------------|
998 */
999 unlink_va(va, &free_vmap_area_root);
1000 kmem_cache_free(vmap_area_cachep, va);
1001 } else if (type == LE_FIT_TYPE) {
1002 /*
1003 * Split left edge of fit VA.
1004 *
1005 * | |
1006 * V NVA V R
1007 * |-------|-------|
1008 */
1009 va->va_start += size;
1010 } else if (type == RE_FIT_TYPE) {
1011 /*
1012 * Split right edge of fit VA.
1013 *
1014 * | |
1015 * L V NVA V
1016 * |-------|-------|
1017 */
1018 va->va_end = nva_start_addr;
1019 } else if (type == NE_FIT_TYPE) {
1020 /*
1021 * Split no edge of fit VA.
1022 *
1023 * | |
1024 * L V NVA V R
1025 * |---|-------|---|
1026 */
1027 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1028 if (unlikely(!lva)) {
1029 /*
1030 * For percpu allocator we do not do any pre-allocation
1031 * and leave it as it is. The reason is it most likely
1032 * never ends up with NE_FIT_TYPE splitting. In case of
1033 * percpu allocations offsets and sizes are aligned to
1034 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1035 * are its main fitting cases.
1036 *
1037 * There are a few exceptions though, as an example it is
1038 * a first allocation (early boot up) when we have "one"
1039 * big free space that has to be split.
1040 *
1041 * Also we can hit this path in case of regular "vmap"
1042 * allocations, if "this" current CPU was not preloaded.
1043 * See the comment in alloc_vmap_area() why. If so, then
1044 * GFP_NOWAIT is used instead to get an extra object for
1045 * split purpose. That is rare and most time does not
1046 * occur.
1047 *
1048 * What happens if an allocation gets failed. Basically,
1049 * an "overflow" path is triggered to purge lazily freed
1050 * areas to free some memory, then, the "retry" path is
1051 * triggered to repeat one more time. See more details
1052 * in alloc_vmap_area() function.
1053 */
1054 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1055 if (!lva)
1056 return -1;
1057 }
1058
1059 /*
1060 * Build the remainder.
1061 */
1062 lva->va_start = va->va_start;
1063 lva->va_end = nva_start_addr;
1064
1065 /*
1066 * Shrink this VA to remaining size.
1067 */
1068 va->va_start = nva_start_addr + size;
1069 } else {
1070 return -1;
1071 }
1072
1073 if (type != FL_FIT_TYPE) {
1074 augment_tree_propagate_from(va);
1075
1076 if (lva) /* type == NE_FIT_TYPE */
1077 insert_vmap_area_augment(lva, &va->rb_node,
1078 &free_vmap_area_root, &free_vmap_area_list);
1079 }
1080
1081 return 0;
1082}
1083
1084/*
1085 * Returns a start address of the newly allocated area, if success.
1086 * Otherwise a vend is returned that indicates failure.
1087 */
1088static __always_inline unsigned long
1089__alloc_vmap_area(unsigned long size, unsigned long align,
1090 unsigned long vstart, unsigned long vend)
1091{
1092 unsigned long nva_start_addr;
1093 struct vmap_area *va;
1094 enum fit_type type;
1095 int ret;
1096
1097 va = find_vmap_lowest_match(size, align, vstart);
1098 if (unlikely(!va))
1099 return vend;
1100
1101 if (va->va_start > vstart)
1102 nva_start_addr = ALIGN(va->va_start, align);
1103 else
1104 nva_start_addr = ALIGN(vstart, align);
1105
1106 /* Check the "vend" restriction. */
1107 if (nva_start_addr + size > vend)
1108 return vend;
1109
1110 /* Classify what we have found. */
1111 type = classify_va_fit_type(va, nva_start_addr, size);
1112 if (WARN_ON_ONCE(type == NOTHING_FIT))
1113 return vend;
1114
1115 /* Update the free vmap_area. */
1116 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1117 if (ret)
1118 return vend;
1119
1120#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1121 find_vmap_lowest_match_check(size);
1122#endif
1123
1124 return nva_start_addr;
1125}
1126
1127/*
1128 * Free a region of KVA allocated by alloc_vmap_area
1129 */
1130static void free_vmap_area(struct vmap_area *va)
1131{
1132 /*
1133 * Remove from the busy tree/list.
1134 */
1135 spin_lock(&vmap_area_lock);
1136 unlink_va(va, &vmap_area_root);
1137 spin_unlock(&vmap_area_lock);
1138
1139 /*
1140 * Insert/Merge it back to the free tree/list.
1141 */
1142 spin_lock(&free_vmap_area_lock);
1143 merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list);
1144 spin_unlock(&free_vmap_area_lock);
1145}
1146
1147/*
1148 * Allocate a region of KVA of the specified size and alignment, within the
1149 * vstart and vend.
1150 */
1151static struct vmap_area *alloc_vmap_area(unsigned long size,
1152 unsigned long align,
1153 unsigned long vstart, unsigned long vend,
1154 int node, gfp_t gfp_mask)
1155{
1156 struct vmap_area *va, *pva;
1157 unsigned long addr;
1158 int purged = 0;
1159 int ret;
1160
1161 BUG_ON(!size);
1162 BUG_ON(offset_in_page(size));
1163 BUG_ON(!is_power_of_2(align));
1164
1165 if (unlikely(!vmap_initialized))
1166 return ERR_PTR(-EBUSY);
1167
1168 might_sleep();
1169 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1170
1171 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1172 if (unlikely(!va))
1173 return ERR_PTR(-ENOMEM);
1174
1175 /*
1176 * Only scan the relevant parts containing pointers to other objects
1177 * to avoid false negatives.
1178 */
1179 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1180
1181retry:
1182 /*
1183 * Preload this CPU with one extra vmap_area object. It is used
1184 * when fit type of free area is NE_FIT_TYPE. Please note, it
1185 * does not guarantee that an allocation occurs on a CPU that
1186 * is preloaded, instead we minimize the case when it is not.
1187 * It can happen because of cpu migration, because there is a
1188 * race until the below spinlock is taken.
1189 *
1190 * The preload is done in non-atomic context, thus it allows us
1191 * to use more permissive allocation masks to be more stable under
1192 * low memory condition and high memory pressure. In rare case,
1193 * if not preloaded, GFP_NOWAIT is used.
1194 *
1195 * Set "pva" to NULL here, because of "retry" path.
1196 */
1197 pva = NULL;
1198
1199 if (!this_cpu_read(ne_fit_preload_node))
1200 /*
1201 * Even if it fails we do not really care about that.
1202 * Just proceed as it is. If needed "overflow" path
1203 * will refill the cache we allocate from.
1204 */
1205 pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1206
1207 spin_lock(&free_vmap_area_lock);
1208
1209 if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva))
1210 kmem_cache_free(vmap_area_cachep, pva);
1211
1212 /*
1213 * If an allocation fails, the "vend" address is
1214 * returned. Therefore trigger the overflow path.
1215 */
1216 addr = __alloc_vmap_area(size, align, vstart, vend);
1217 spin_unlock(&free_vmap_area_lock);
1218
1219 if (unlikely(addr == vend))
1220 goto overflow;
1221
1222 va->va_start = addr;
1223 va->va_end = addr + size;
1224 va->vm = NULL;
1225
1226
1227 spin_lock(&vmap_area_lock);
1228 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1229 spin_unlock(&vmap_area_lock);
1230
1231 BUG_ON(!IS_ALIGNED(va->va_start, align));
1232 BUG_ON(va->va_start < vstart);
1233 BUG_ON(va->va_end > vend);
1234
1235 ret = kasan_populate_vmalloc(addr, size);
1236 if (ret) {
1237 free_vmap_area(va);
1238 return ERR_PTR(ret);
1239 }
1240
1241 return va;
1242
1243overflow:
1244 if (!purged) {
1245 purge_vmap_area_lazy();
1246 purged = 1;
1247 goto retry;
1248 }
1249
1250 if (gfpflags_allow_blocking(gfp_mask)) {
1251 unsigned long freed = 0;
1252 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1253 if (freed > 0) {
1254 purged = 0;
1255 goto retry;
1256 }
1257 }
1258
1259 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1260 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1261 size);
1262
1263 kmem_cache_free(vmap_area_cachep, va);
1264 return ERR_PTR(-EBUSY);
1265}
1266
1267int register_vmap_purge_notifier(struct notifier_block *nb)
1268{
1269 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1270}
1271EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1272
1273int unregister_vmap_purge_notifier(struct notifier_block *nb)
1274{
1275 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1276}
1277EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1278
1279/*
1280 * lazy_max_pages is the maximum amount of virtual address space we gather up
1281 * before attempting to purge with a TLB flush.
1282 *
1283 * There is a tradeoff here: a larger number will cover more kernel page tables
1284 * and take slightly longer to purge, but it will linearly reduce the number of
1285 * global TLB flushes that must be performed. It would seem natural to scale
1286 * this number up linearly with the number of CPUs (because vmapping activity
1287 * could also scale linearly with the number of CPUs), however it is likely
1288 * that in practice, workloads might be constrained in other ways that mean
1289 * vmap activity will not scale linearly with CPUs. Also, I want to be
1290 * conservative and not introduce a big latency on huge systems, so go with
1291 * a less aggressive log scale. It will still be an improvement over the old
1292 * code, and it will be simple to change the scale factor if we find that it
1293 * becomes a problem on bigger systems.
1294 */
1295static unsigned long lazy_max_pages(void)
1296{
1297 unsigned int log;
1298
1299 log = fls(num_online_cpus());
1300
1301 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1302}
1303
1304static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1305
1306/*
1307 * Serialize vmap purging. There is no actual criticial section protected
1308 * by this look, but we want to avoid concurrent calls for performance
1309 * reasons and to make the pcpu_get_vm_areas more deterministic.
1310 */
1311static DEFINE_MUTEX(vmap_purge_lock);
1312
1313/* for per-CPU blocks */
1314static void purge_fragmented_blocks_allcpus(void);
1315
1316/*
1317 * called before a call to iounmap() if the caller wants vm_area_struct's
1318 * immediately freed.
1319 */
1320void set_iounmap_nonlazy(void)
1321{
1322 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1323}
1324
1325/*
1326 * Purges all lazily-freed vmap areas.
1327 */
1328static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1329{
1330 unsigned long resched_threshold;
1331 struct llist_node *valist;
1332 struct vmap_area *va;
1333 struct vmap_area *n_va;
1334
1335 lockdep_assert_held(&vmap_purge_lock);
1336
1337 valist = llist_del_all(&vmap_purge_list);
1338 if (unlikely(valist == NULL))
1339 return false;
1340
1341 /*
1342 * TODO: to calculate a flush range without looping.
1343 * The list can be up to lazy_max_pages() elements.
1344 */
1345 llist_for_each_entry(va, valist, purge_list) {
1346 if (va->va_start < start)
1347 start = va->va_start;
1348 if (va->va_end > end)
1349 end = va->va_end;
1350 }
1351
1352 flush_tlb_kernel_range(start, end);
1353 resched_threshold = lazy_max_pages() << 1;
1354
1355 spin_lock(&free_vmap_area_lock);
1356 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1357 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1358 unsigned long orig_start = va->va_start;
1359 unsigned long orig_end = va->va_end;
1360
1361 /*
1362 * Finally insert or merge lazily-freed area. It is
1363 * detached and there is no need to "unlink" it from
1364 * anything.
1365 */
1366 va = merge_or_add_vmap_area(va, &free_vmap_area_root,
1367 &free_vmap_area_list);
1368
1369 if (!va)
1370 continue;
1371
1372 if (is_vmalloc_or_module_addr((void *)orig_start))
1373 kasan_release_vmalloc(orig_start, orig_end,
1374 va->va_start, va->va_end);
1375
1376 atomic_long_sub(nr, &vmap_lazy_nr);
1377
1378 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1379 cond_resched_lock(&free_vmap_area_lock);
1380 }
1381 spin_unlock(&free_vmap_area_lock);
1382 return true;
1383}
1384
1385/*
1386 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1387 * is already purging.
1388 */
1389static void try_purge_vmap_area_lazy(void)
1390{
1391 if (mutex_trylock(&vmap_purge_lock)) {
1392 __purge_vmap_area_lazy(ULONG_MAX, 0);
1393 mutex_unlock(&vmap_purge_lock);
1394 }
1395}
1396
1397/*
1398 * Kick off a purge of the outstanding lazy areas.
1399 */
1400static void purge_vmap_area_lazy(void)
1401{
1402 mutex_lock(&vmap_purge_lock);
1403 purge_fragmented_blocks_allcpus();
1404 __purge_vmap_area_lazy(ULONG_MAX, 0);
1405 mutex_unlock(&vmap_purge_lock);
1406}
1407
1408/*
1409 * Free a vmap area, caller ensuring that the area has been unmapped
1410 * and flush_cache_vunmap had been called for the correct range
1411 * previously.
1412 */
1413static void free_vmap_area_noflush(struct vmap_area *va)
1414{
1415 unsigned long nr_lazy;
1416
1417 spin_lock(&vmap_area_lock);
1418 unlink_va(va, &vmap_area_root);
1419 spin_unlock(&vmap_area_lock);
1420
1421 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1422 PAGE_SHIFT, &vmap_lazy_nr);
1423
1424 /* After this point, we may free va at any time */
1425 llist_add(&va->purge_list, &vmap_purge_list);
1426
1427 if (unlikely(nr_lazy > lazy_max_pages()))
1428 try_purge_vmap_area_lazy();
1429}
1430
1431/*
1432 * Free and unmap a vmap area
1433 */
1434static void free_unmap_vmap_area(struct vmap_area *va)
1435{
1436 flush_cache_vunmap(va->va_start, va->va_end);
1437 unmap_kernel_range_noflush(va->va_start, va->va_end - va->va_start);
1438 if (debug_pagealloc_enabled_static())
1439 flush_tlb_kernel_range(va->va_start, va->va_end);
1440
1441 free_vmap_area_noflush(va);
1442}
1443
1444static struct vmap_area *find_vmap_area(unsigned long addr)
1445{
1446 struct vmap_area *va;
1447
1448 spin_lock(&vmap_area_lock);
1449 va = __find_vmap_area(addr);
1450 spin_unlock(&vmap_area_lock);
1451
1452 return va;
1453}
1454
1455/*** Per cpu kva allocator ***/
1456
1457/*
1458 * vmap space is limited especially on 32 bit architectures. Ensure there is
1459 * room for at least 16 percpu vmap blocks per CPU.
1460 */
1461/*
1462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1463 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1464 * instead (we just need a rough idea)
1465 */
1466#if BITS_PER_LONG == 32
1467#define VMALLOC_SPACE (128UL*1024*1024)
1468#else
1469#define VMALLOC_SPACE (128UL*1024*1024*1024)
1470#endif
1471
1472#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1473#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1474#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1475#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1476#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1477#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1478#define VMAP_BBMAP_BITS \
1479 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1480 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1481 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1482
1483#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1484
1485struct vmap_block_queue {
1486 spinlock_t lock;
1487 struct list_head free;
1488};
1489
1490struct vmap_block {
1491 spinlock_t lock;
1492 struct vmap_area *va;
1493 unsigned long free, dirty;
1494 unsigned long dirty_min, dirty_max; /*< dirty range */
1495 struct list_head free_list;
1496 struct rcu_head rcu_head;
1497 struct list_head purge;
1498};
1499
1500/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1501static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1502
1503/*
1504 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1505 * in the free path. Could get rid of this if we change the API to return a
1506 * "cookie" from alloc, to be passed to free. But no big deal yet.
1507 */
1508static DEFINE_XARRAY(vmap_blocks);
1509
1510/*
1511 * We should probably have a fallback mechanism to allocate virtual memory
1512 * out of partially filled vmap blocks. However vmap block sizing should be
1513 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1514 * big problem.
1515 */
1516
1517static unsigned long addr_to_vb_idx(unsigned long addr)
1518{
1519 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1520 addr /= VMAP_BLOCK_SIZE;
1521 return addr;
1522}
1523
1524static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1525{
1526 unsigned long addr;
1527
1528 addr = va_start + (pages_off << PAGE_SHIFT);
1529 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1530 return (void *)addr;
1531}
1532
1533/**
1534 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1535 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1536 * @order: how many 2^order pages should be occupied in newly allocated block
1537 * @gfp_mask: flags for the page level allocator
1538 *
1539 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1540 */
1541static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1542{
1543 struct vmap_block_queue *vbq;
1544 struct vmap_block *vb;
1545 struct vmap_area *va;
1546 unsigned long vb_idx;
1547 int node, err;
1548 void *vaddr;
1549
1550 node = numa_node_id();
1551
1552 vb = kmalloc_node(sizeof(struct vmap_block),
1553 gfp_mask & GFP_RECLAIM_MASK, node);
1554 if (unlikely(!vb))
1555 return ERR_PTR(-ENOMEM);
1556
1557 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1558 VMALLOC_START, VMALLOC_END,
1559 node, gfp_mask);
1560 if (IS_ERR(va)) {
1561 kfree(vb);
1562 return ERR_CAST(va);
1563 }
1564
1565 vaddr = vmap_block_vaddr(va->va_start, 0);
1566 spin_lock_init(&vb->lock);
1567 vb->va = va;
1568 /* At least something should be left free */
1569 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1570 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1571 vb->dirty = 0;
1572 vb->dirty_min = VMAP_BBMAP_BITS;
1573 vb->dirty_max = 0;
1574 INIT_LIST_HEAD(&vb->free_list);
1575
1576 vb_idx = addr_to_vb_idx(va->va_start);
1577 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1578 if (err) {
1579 kfree(vb);
1580 free_vmap_area(va);
1581 return ERR_PTR(err);
1582 }
1583
1584 vbq = &get_cpu_var(vmap_block_queue);
1585 spin_lock(&vbq->lock);
1586 list_add_tail_rcu(&vb->free_list, &vbq->free);
1587 spin_unlock(&vbq->lock);
1588 put_cpu_var(vmap_block_queue);
1589
1590 return vaddr;
1591}
1592
1593static void free_vmap_block(struct vmap_block *vb)
1594{
1595 struct vmap_block *tmp;
1596
1597 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1598 BUG_ON(tmp != vb);
1599
1600 free_vmap_area_noflush(vb->va);
1601 kfree_rcu(vb, rcu_head);
1602}
1603
1604static void purge_fragmented_blocks(int cpu)
1605{
1606 LIST_HEAD(purge);
1607 struct vmap_block *vb;
1608 struct vmap_block *n_vb;
1609 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1610
1611 rcu_read_lock();
1612 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1613
1614 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1615 continue;
1616
1617 spin_lock(&vb->lock);
1618 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1619 vb->free = 0; /* prevent further allocs after releasing lock */
1620 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1621 vb->dirty_min = 0;
1622 vb->dirty_max = VMAP_BBMAP_BITS;
1623 spin_lock(&vbq->lock);
1624 list_del_rcu(&vb->free_list);
1625 spin_unlock(&vbq->lock);
1626 spin_unlock(&vb->lock);
1627 list_add_tail(&vb->purge, &purge);
1628 } else
1629 spin_unlock(&vb->lock);
1630 }
1631 rcu_read_unlock();
1632
1633 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1634 list_del(&vb->purge);
1635 free_vmap_block(vb);
1636 }
1637}
1638
1639static void purge_fragmented_blocks_allcpus(void)
1640{
1641 int cpu;
1642
1643 for_each_possible_cpu(cpu)
1644 purge_fragmented_blocks(cpu);
1645}
1646
1647static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1648{
1649 struct vmap_block_queue *vbq;
1650 struct vmap_block *vb;
1651 void *vaddr = NULL;
1652 unsigned int order;
1653
1654 BUG_ON(offset_in_page(size));
1655 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1656 if (WARN_ON(size == 0)) {
1657 /*
1658 * Allocating 0 bytes isn't what caller wants since
1659 * get_order(0) returns funny result. Just warn and terminate
1660 * early.
1661 */
1662 return NULL;
1663 }
1664 order = get_order(size);
1665
1666 rcu_read_lock();
1667 vbq = &get_cpu_var(vmap_block_queue);
1668 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1669 unsigned long pages_off;
1670
1671 spin_lock(&vb->lock);
1672 if (vb->free < (1UL << order)) {
1673 spin_unlock(&vb->lock);
1674 continue;
1675 }
1676
1677 pages_off = VMAP_BBMAP_BITS - vb->free;
1678 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1679 vb->free -= 1UL << order;
1680 if (vb->free == 0) {
1681 spin_lock(&vbq->lock);
1682 list_del_rcu(&vb->free_list);
1683 spin_unlock(&vbq->lock);
1684 }
1685
1686 spin_unlock(&vb->lock);
1687 break;
1688 }
1689
1690 put_cpu_var(vmap_block_queue);
1691 rcu_read_unlock();
1692
1693 /* Allocate new block if nothing was found */
1694 if (!vaddr)
1695 vaddr = new_vmap_block(order, gfp_mask);
1696
1697 return vaddr;
1698}
1699
1700static void vb_free(unsigned long addr, unsigned long size)
1701{
1702 unsigned long offset;
1703 unsigned int order;
1704 struct vmap_block *vb;
1705
1706 BUG_ON(offset_in_page(size));
1707 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1708
1709 flush_cache_vunmap(addr, addr + size);
1710
1711 order = get_order(size);
1712 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
1713 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
1714
1715 unmap_kernel_range_noflush(addr, size);
1716
1717 if (debug_pagealloc_enabled_static())
1718 flush_tlb_kernel_range(addr, addr + size);
1719
1720 spin_lock(&vb->lock);
1721
1722 /* Expand dirty range */
1723 vb->dirty_min = min(vb->dirty_min, offset);
1724 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1725
1726 vb->dirty += 1UL << order;
1727 if (vb->dirty == VMAP_BBMAP_BITS) {
1728 BUG_ON(vb->free);
1729 spin_unlock(&vb->lock);
1730 free_vmap_block(vb);
1731 } else
1732 spin_unlock(&vb->lock);
1733}
1734
1735static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1736{
1737 int cpu;
1738
1739 if (unlikely(!vmap_initialized))
1740 return;
1741
1742 might_sleep();
1743
1744 for_each_possible_cpu(cpu) {
1745 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1746 struct vmap_block *vb;
1747
1748 rcu_read_lock();
1749 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1750 spin_lock(&vb->lock);
1751 if (vb->dirty) {
1752 unsigned long va_start = vb->va->va_start;
1753 unsigned long s, e;
1754
1755 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1756 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1757
1758 start = min(s, start);
1759 end = max(e, end);
1760
1761 flush = 1;
1762 }
1763 spin_unlock(&vb->lock);
1764 }
1765 rcu_read_unlock();
1766 }
1767
1768 mutex_lock(&vmap_purge_lock);
1769 purge_fragmented_blocks_allcpus();
1770 if (!__purge_vmap_area_lazy(start, end) && flush)
1771 flush_tlb_kernel_range(start, end);
1772 mutex_unlock(&vmap_purge_lock);
1773}
1774
1775/**
1776 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1777 *
1778 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1779 * to amortize TLB flushing overheads. What this means is that any page you
1780 * have now, may, in a former life, have been mapped into kernel virtual
1781 * address by the vmap layer and so there might be some CPUs with TLB entries
1782 * still referencing that page (additional to the regular 1:1 kernel mapping).
1783 *
1784 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1785 * be sure that none of the pages we have control over will have any aliases
1786 * from the vmap layer.
1787 */
1788void vm_unmap_aliases(void)
1789{
1790 unsigned long start = ULONG_MAX, end = 0;
1791 int flush = 0;
1792
1793 _vm_unmap_aliases(start, end, flush);
1794}
1795EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1796
1797/**
1798 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1799 * @mem: the pointer returned by vm_map_ram
1800 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1801 */
1802void vm_unmap_ram(const void *mem, unsigned int count)
1803{
1804 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1805 unsigned long addr = (unsigned long)mem;
1806 struct vmap_area *va;
1807
1808 might_sleep();
1809 BUG_ON(!addr);
1810 BUG_ON(addr < VMALLOC_START);
1811 BUG_ON(addr > VMALLOC_END);
1812 BUG_ON(!PAGE_ALIGNED(addr));
1813
1814 kasan_poison_vmalloc(mem, size);
1815
1816 if (likely(count <= VMAP_MAX_ALLOC)) {
1817 debug_check_no_locks_freed(mem, size);
1818 vb_free(addr, size);
1819 return;
1820 }
1821
1822 va = find_vmap_area(addr);
1823 BUG_ON(!va);
1824 debug_check_no_locks_freed((void *)va->va_start,
1825 (va->va_end - va->va_start));
1826 free_unmap_vmap_area(va);
1827}
1828EXPORT_SYMBOL(vm_unmap_ram);
1829
1830/**
1831 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1832 * @pages: an array of pointers to the pages to be mapped
1833 * @count: number of pages
1834 * @node: prefer to allocate data structures on this node
1835 *
1836 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1837 * faster than vmap so it's good. But if you mix long-life and short-life
1838 * objects with vm_map_ram(), it could consume lots of address space through
1839 * fragmentation (especially on a 32bit machine). You could see failures in
1840 * the end. Please use this function for short-lived objects.
1841 *
1842 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1843 */
1844void *vm_map_ram(struct page **pages, unsigned int count, int node)
1845{
1846 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1847 unsigned long addr;
1848 void *mem;
1849
1850 if (likely(count <= VMAP_MAX_ALLOC)) {
1851 mem = vb_alloc(size, GFP_KERNEL);
1852 if (IS_ERR(mem))
1853 return NULL;
1854 addr = (unsigned long)mem;
1855 } else {
1856 struct vmap_area *va;
1857 va = alloc_vmap_area(size, PAGE_SIZE,
1858 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1859 if (IS_ERR(va))
1860 return NULL;
1861
1862 addr = va->va_start;
1863 mem = (void *)addr;
1864 }
1865
1866 kasan_unpoison_vmalloc(mem, size);
1867
1868 if (map_kernel_range(addr, size, PAGE_KERNEL, pages) < 0) {
1869 vm_unmap_ram(mem, count);
1870 return NULL;
1871 }
1872 return mem;
1873}
1874EXPORT_SYMBOL(vm_map_ram);
1875
1876static struct vm_struct *vmlist __initdata;
1877
1878/**
1879 * vm_area_add_early - add vmap area early during boot
1880 * @vm: vm_struct to add
1881 *
1882 * This function is used to add fixed kernel vm area to vmlist before
1883 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1884 * should contain proper values and the other fields should be zero.
1885 *
1886 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1887 */
1888void __init vm_area_add_early(struct vm_struct *vm)
1889{
1890 struct vm_struct *tmp, **p;
1891
1892 BUG_ON(vmap_initialized);
1893 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1894 if (tmp->addr >= vm->addr) {
1895 BUG_ON(tmp->addr < vm->addr + vm->size);
1896 break;
1897 } else
1898 BUG_ON(tmp->addr + tmp->size > vm->addr);
1899 }
1900 vm->next = *p;
1901 *p = vm;
1902}
1903
1904/**
1905 * vm_area_register_early - register vmap area early during boot
1906 * @vm: vm_struct to register
1907 * @align: requested alignment
1908 *
1909 * This function is used to register kernel vm area before
1910 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1911 * proper values on entry and other fields should be zero. On return,
1912 * vm->addr contains the allocated address.
1913 *
1914 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1915 */
1916void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1917{
1918 static size_t vm_init_off __initdata;
1919 unsigned long addr;
1920
1921 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1922 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1923
1924 vm->addr = (void *)addr;
1925
1926 vm_area_add_early(vm);
1927}
1928
1929static void vmap_init_free_space(void)
1930{
1931 unsigned long vmap_start = 1;
1932 const unsigned long vmap_end = ULONG_MAX;
1933 struct vmap_area *busy, *free;
1934
1935 /*
1936 * B F B B B F
1937 * -|-----|.....|-----|-----|-----|.....|-
1938 * | The KVA space |
1939 * |<--------------------------------->|
1940 */
1941 list_for_each_entry(busy, &vmap_area_list, list) {
1942 if (busy->va_start - vmap_start > 0) {
1943 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1944 if (!WARN_ON_ONCE(!free)) {
1945 free->va_start = vmap_start;
1946 free->va_end = busy->va_start;
1947
1948 insert_vmap_area_augment(free, NULL,
1949 &free_vmap_area_root,
1950 &free_vmap_area_list);
1951 }
1952 }
1953
1954 vmap_start = busy->va_end;
1955 }
1956
1957 if (vmap_end - vmap_start > 0) {
1958 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1959 if (!WARN_ON_ONCE(!free)) {
1960 free->va_start = vmap_start;
1961 free->va_end = vmap_end;
1962
1963 insert_vmap_area_augment(free, NULL,
1964 &free_vmap_area_root,
1965 &free_vmap_area_list);
1966 }
1967 }
1968}
1969
1970void __init vmalloc_init(void)
1971{
1972 struct vmap_area *va;
1973 struct vm_struct *tmp;
1974 int i;
1975
1976 /*
1977 * Create the cache for vmap_area objects.
1978 */
1979 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1980
1981 for_each_possible_cpu(i) {
1982 struct vmap_block_queue *vbq;
1983 struct vfree_deferred *p;
1984
1985 vbq = &per_cpu(vmap_block_queue, i);
1986 spin_lock_init(&vbq->lock);
1987 INIT_LIST_HEAD(&vbq->free);
1988 p = &per_cpu(vfree_deferred, i);
1989 init_llist_head(&p->list);
1990 INIT_WORK(&p->wq, free_work);
1991 }
1992
1993 /* Import existing vmlist entries. */
1994 for (tmp = vmlist; tmp; tmp = tmp->next) {
1995 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1996 if (WARN_ON_ONCE(!va))
1997 continue;
1998
1999 va->va_start = (unsigned long)tmp->addr;
2000 va->va_end = va->va_start + tmp->size;
2001 va->vm = tmp;
2002 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2003 }
2004
2005 /*
2006 * Now we can initialize a free vmap space.
2007 */
2008 vmap_init_free_space();
2009 vmap_initialized = true;
2010}
2011
2012/**
2013 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2014 * @addr: start of the VM area to unmap
2015 * @size: size of the VM area to unmap
2016 *
2017 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2018 * the unmapping and tlb after.
2019 */
2020void unmap_kernel_range(unsigned long addr, unsigned long size)
2021{
2022 unsigned long end = addr + size;
2023
2024 flush_cache_vunmap(addr, end);
2025 unmap_kernel_range_noflush(addr, size);
2026 flush_tlb_kernel_range(addr, end);
2027}
2028
2029static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2030 struct vmap_area *va, unsigned long flags, const void *caller)
2031{
2032 vm->flags = flags;
2033 vm->addr = (void *)va->va_start;
2034 vm->size = va->va_end - va->va_start;
2035 vm->caller = caller;
2036 va->vm = vm;
2037}
2038
2039static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2040 unsigned long flags, const void *caller)
2041{
2042 spin_lock(&vmap_area_lock);
2043 setup_vmalloc_vm_locked(vm, va, flags, caller);
2044 spin_unlock(&vmap_area_lock);
2045}
2046
2047static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2048{
2049 /*
2050 * Before removing VM_UNINITIALIZED,
2051 * we should make sure that vm has proper values.
2052 * Pair with smp_rmb() in show_numa_info().
2053 */
2054 smp_wmb();
2055 vm->flags &= ~VM_UNINITIALIZED;
2056}
2057
2058static struct vm_struct *__get_vm_area_node(unsigned long size,
2059 unsigned long align, unsigned long flags, unsigned long start,
2060 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2061{
2062 struct vmap_area *va;
2063 struct vm_struct *area;
2064 unsigned long requested_size = size;
2065
2066 BUG_ON(in_interrupt());
2067 size = PAGE_ALIGN(size);
2068 if (unlikely(!size))
2069 return NULL;
2070
2071 if (flags & VM_IOREMAP)
2072 align = 1ul << clamp_t(int, get_count_order_long(size),
2073 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2074
2075 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2076 if (unlikely(!area))
2077 return NULL;
2078
2079 if (!(flags & VM_NO_GUARD))
2080 size += PAGE_SIZE;
2081
2082 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2083 if (IS_ERR(va)) {
2084 kfree(area);
2085 return NULL;
2086 }
2087
2088 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2089
2090 setup_vmalloc_vm(area, va, flags, caller);
2091
2092 return area;
2093}
2094
2095struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2096 unsigned long start, unsigned long end,
2097 const void *caller)
2098{
2099 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2100 GFP_KERNEL, caller);
2101}
2102
2103/**
2104 * get_vm_area - reserve a contiguous kernel virtual area
2105 * @size: size of the area
2106 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2107 *
2108 * Search an area of @size in the kernel virtual mapping area,
2109 * and reserved it for out purposes. Returns the area descriptor
2110 * on success or %NULL on failure.
2111 *
2112 * Return: the area descriptor on success or %NULL on failure.
2113 */
2114struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2115{
2116 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2117 NUMA_NO_NODE, GFP_KERNEL,
2118 __builtin_return_address(0));
2119}
2120
2121struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2122 const void *caller)
2123{
2124 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2125 NUMA_NO_NODE, GFP_KERNEL, caller);
2126}
2127
2128/**
2129 * find_vm_area - find a continuous kernel virtual area
2130 * @addr: base address
2131 *
2132 * Search for the kernel VM area starting at @addr, and return it.
2133 * It is up to the caller to do all required locking to keep the returned
2134 * pointer valid.
2135 *
2136 * Return: pointer to the found area or %NULL on faulure
2137 */
2138struct vm_struct *find_vm_area(const void *addr)
2139{
2140 struct vmap_area *va;
2141
2142 va = find_vmap_area((unsigned long)addr);
2143 if (!va)
2144 return NULL;
2145
2146 return va->vm;
2147}
2148
2149/**
2150 * remove_vm_area - find and remove a continuous kernel virtual area
2151 * @addr: base address
2152 *
2153 * Search for the kernel VM area starting at @addr, and remove it.
2154 * This function returns the found VM area, but using it is NOT safe
2155 * on SMP machines, except for its size or flags.
2156 *
2157 * Return: pointer to the found area or %NULL on faulure
2158 */
2159struct vm_struct *remove_vm_area(const void *addr)
2160{
2161 struct vmap_area *va;
2162
2163 might_sleep();
2164
2165 spin_lock(&vmap_area_lock);
2166 va = __find_vmap_area((unsigned long)addr);
2167 if (va && va->vm) {
2168 struct vm_struct *vm = va->vm;
2169
2170 va->vm = NULL;
2171 spin_unlock(&vmap_area_lock);
2172
2173 kasan_free_shadow(vm);
2174 free_unmap_vmap_area(va);
2175
2176 return vm;
2177 }
2178
2179 spin_unlock(&vmap_area_lock);
2180 return NULL;
2181}
2182
2183static inline void set_area_direct_map(const struct vm_struct *area,
2184 int (*set_direct_map)(struct page *page))
2185{
2186 int i;
2187
2188 for (i = 0; i < area->nr_pages; i++)
2189 if (page_address(area->pages[i]))
2190 set_direct_map(area->pages[i]);
2191}
2192
2193/* Handle removing and resetting vm mappings related to the vm_struct. */
2194static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2195{
2196 unsigned long start = ULONG_MAX, end = 0;
2197 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2198 int flush_dmap = 0;
2199 int i;
2200
2201 remove_vm_area(area->addr);
2202
2203 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2204 if (!flush_reset)
2205 return;
2206
2207 /*
2208 * If not deallocating pages, just do the flush of the VM area and
2209 * return.
2210 */
2211 if (!deallocate_pages) {
2212 vm_unmap_aliases();
2213 return;
2214 }
2215
2216 /*
2217 * If execution gets here, flush the vm mapping and reset the direct
2218 * map. Find the start and end range of the direct mappings to make sure
2219 * the vm_unmap_aliases() flush includes the direct map.
2220 */
2221 for (i = 0; i < area->nr_pages; i++) {
2222 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2223 if (addr) {
2224 start = min(addr, start);
2225 end = max(addr + PAGE_SIZE, end);
2226 flush_dmap = 1;
2227 }
2228 }
2229
2230 /*
2231 * Set direct map to something invalid so that it won't be cached if
2232 * there are any accesses after the TLB flush, then flush the TLB and
2233 * reset the direct map permissions to the default.
2234 */
2235 set_area_direct_map(area, set_direct_map_invalid_noflush);
2236 _vm_unmap_aliases(start, end, flush_dmap);
2237 set_area_direct_map(area, set_direct_map_default_noflush);
2238}
2239
2240static void __vunmap(const void *addr, int deallocate_pages)
2241{
2242 struct vm_struct *area;
2243
2244 if (!addr)
2245 return;
2246
2247 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2248 addr))
2249 return;
2250
2251 area = find_vm_area(addr);
2252 if (unlikely(!area)) {
2253 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2254 addr);
2255 return;
2256 }
2257
2258 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2259 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2260
2261 kasan_poison_vmalloc(area->addr, area->size);
2262
2263 vm_remove_mappings(area, deallocate_pages);
2264
2265 if (deallocate_pages) {
2266 int i;
2267
2268 for (i = 0; i < area->nr_pages; i++) {
2269 struct page *page = area->pages[i];
2270
2271 BUG_ON(!page);
2272 __free_pages(page, 0);
2273 }
2274 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2275
2276 kvfree(area->pages);
2277 }
2278
2279 kfree(area);
2280 return;
2281}
2282
2283static inline void __vfree_deferred(const void *addr)
2284{
2285 /*
2286 * Use raw_cpu_ptr() because this can be called from preemptible
2287 * context. Preemption is absolutely fine here, because the llist_add()
2288 * implementation is lockless, so it works even if we are adding to
2289 * another cpu's list. schedule_work() should be fine with this too.
2290 */
2291 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2292
2293 if (llist_add((struct llist_node *)addr, &p->list))
2294 schedule_work(&p->wq);
2295}
2296
2297/**
2298 * vfree_atomic - release memory allocated by vmalloc()
2299 * @addr: memory base address
2300 *
2301 * This one is just like vfree() but can be called in any atomic context
2302 * except NMIs.
2303 */
2304void vfree_atomic(const void *addr)
2305{
2306 BUG_ON(in_nmi());
2307
2308 kmemleak_free(addr);
2309
2310 if (!addr)
2311 return;
2312 __vfree_deferred(addr);
2313}
2314
2315static void __vfree(const void *addr)
2316{
2317 if (unlikely(in_interrupt()))
2318 __vfree_deferred(addr);
2319 else
2320 __vunmap(addr, 1);
2321}
2322
2323/**
2324 * vfree - release memory allocated by vmalloc()
2325 * @addr: memory base address
2326 *
2327 * Free the virtually continuous memory area starting at @addr, as
2328 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2329 * NULL, no operation is performed.
2330 *
2331 * Must not be called in NMI context (strictly speaking, only if we don't
2332 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2333 * conventions for vfree() arch-depenedent would be a really bad idea)
2334 *
2335 * May sleep if called *not* from interrupt context.
2336 *
2337 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2338 */
2339void vfree(const void *addr)
2340{
2341 BUG_ON(in_nmi());
2342
2343 kmemleak_free(addr);
2344
2345 might_sleep_if(!in_interrupt());
2346
2347 if (!addr)
2348 return;
2349
2350 __vfree(addr);
2351}
2352EXPORT_SYMBOL(vfree);
2353
2354/**
2355 * vunmap - release virtual mapping obtained by vmap()
2356 * @addr: memory base address
2357 *
2358 * Free the virtually contiguous memory area starting at @addr,
2359 * which was created from the page array passed to vmap().
2360 *
2361 * Must not be called in interrupt context.
2362 */
2363void vunmap(const void *addr)
2364{
2365 BUG_ON(in_interrupt());
2366 might_sleep();
2367 if (addr)
2368 __vunmap(addr, 0);
2369}
2370EXPORT_SYMBOL(vunmap);
2371
2372/**
2373 * vmap - map an array of pages into virtually contiguous space
2374 * @pages: array of page pointers
2375 * @count: number of pages to map
2376 * @flags: vm_area->flags
2377 * @prot: page protection for the mapping
2378 *
2379 * Maps @count pages from @pages into contiguous kernel virtual
2380 * space.
2381 *
2382 * Return: the address of the area or %NULL on failure
2383 */
2384void *vmap(struct page **pages, unsigned int count,
2385 unsigned long flags, pgprot_t prot)
2386{
2387 struct vm_struct *area;
2388 unsigned long size; /* In bytes */
2389
2390 might_sleep();
2391
2392 if (count > totalram_pages())
2393 return NULL;
2394
2395 size = (unsigned long)count << PAGE_SHIFT;
2396 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2397 if (!area)
2398 return NULL;
2399
2400 if (map_kernel_range((unsigned long)area->addr, size, pgprot_nx(prot),
2401 pages) < 0) {
2402 vunmap(area->addr);
2403 return NULL;
2404 }
2405
2406 return area->addr;
2407}
2408EXPORT_SYMBOL(vmap);
2409
2410static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2411 pgprot_t prot, int node)
2412{
2413 struct page **pages;
2414 unsigned int nr_pages, array_size, i;
2415 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2416 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2417 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2418 0 :
2419 __GFP_HIGHMEM;
2420
2421 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2422 array_size = (nr_pages * sizeof(struct page *));
2423
2424 /* Please note that the recursion is strictly bounded. */
2425 if (array_size > PAGE_SIZE) {
2426 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2427 node, area->caller);
2428 } else {
2429 pages = kmalloc_node(array_size, nested_gfp, node);
2430 }
2431
2432 if (!pages) {
2433 remove_vm_area(area->addr);
2434 kfree(area);
2435 return NULL;
2436 }
2437
2438 area->pages = pages;
2439 area->nr_pages = nr_pages;
2440
2441 for (i = 0; i < area->nr_pages; i++) {
2442 struct page *page;
2443
2444 if (node == NUMA_NO_NODE)
2445 page = alloc_page(alloc_mask|highmem_mask);
2446 else
2447 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2448
2449 if (unlikely(!page)) {
2450 /* Successfully allocated i pages, free them in __vunmap() */
2451 area->nr_pages = i;
2452 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2453 goto fail;
2454 }
2455 area->pages[i] = page;
2456 if (gfpflags_allow_blocking(gfp_mask))
2457 cond_resched();
2458 }
2459 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2460
2461 if (map_kernel_range((unsigned long)area->addr, get_vm_area_size(area),
2462 prot, pages) < 0)
2463 goto fail;
2464
2465 return area->addr;
2466
2467fail:
2468 warn_alloc(gfp_mask, NULL,
2469 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2470 (area->nr_pages*PAGE_SIZE), area->size);
2471 __vfree(area->addr);
2472 return NULL;
2473}
2474
2475/**
2476 * __vmalloc_node_range - allocate virtually contiguous memory
2477 * @size: allocation size
2478 * @align: desired alignment
2479 * @start: vm area range start
2480 * @end: vm area range end
2481 * @gfp_mask: flags for the page level allocator
2482 * @prot: protection mask for the allocated pages
2483 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2484 * @node: node to use for allocation or NUMA_NO_NODE
2485 * @caller: caller's return address
2486 *
2487 * Allocate enough pages to cover @size from the page level
2488 * allocator with @gfp_mask flags. Map them into contiguous
2489 * kernel virtual space, using a pagetable protection of @prot.
2490 *
2491 * Return: the address of the area or %NULL on failure
2492 */
2493void *__vmalloc_node_range(unsigned long size, unsigned long align,
2494 unsigned long start, unsigned long end, gfp_t gfp_mask,
2495 pgprot_t prot, unsigned long vm_flags, int node,
2496 const void *caller)
2497{
2498 struct vm_struct *area;
2499 void *addr;
2500 unsigned long real_size = size;
2501
2502 size = PAGE_ALIGN(size);
2503 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2504 goto fail;
2505
2506 area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED |
2507 vm_flags, start, end, node, gfp_mask, caller);
2508 if (!area)
2509 goto fail;
2510
2511 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2512 if (!addr)
2513 return NULL;
2514
2515 /*
2516 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2517 * flag. It means that vm_struct is not fully initialized.
2518 * Now, it is fully initialized, so remove this flag here.
2519 */
2520 clear_vm_uninitialized_flag(area);
2521
2522 kmemleak_vmalloc(area, size, gfp_mask);
2523
2524 return addr;
2525
2526fail:
2527 warn_alloc(gfp_mask, NULL,
2528 "vmalloc: allocation failure: %lu bytes", real_size);
2529 return NULL;
2530}
2531
2532/**
2533 * __vmalloc_node - allocate virtually contiguous memory
2534 * @size: allocation size
2535 * @align: desired alignment
2536 * @gfp_mask: flags for the page level allocator
2537 * @node: node to use for allocation or NUMA_NO_NODE
2538 * @caller: caller's return address
2539 *
2540 * Allocate enough pages to cover @size from the page level allocator with
2541 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2542 *
2543 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2544 * and __GFP_NOFAIL are not supported
2545 *
2546 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2547 * with mm people.
2548 *
2549 * Return: pointer to the allocated memory or %NULL on error
2550 */
2551void *__vmalloc_node(unsigned long size, unsigned long align,
2552 gfp_t gfp_mask, int node, const void *caller)
2553{
2554 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2555 gfp_mask, PAGE_KERNEL, 0, node, caller);
2556}
2557/*
2558 * This is only for performance analysis of vmalloc and stress purpose.
2559 * It is required by vmalloc test module, therefore do not use it other
2560 * than that.
2561 */
2562#ifdef CONFIG_TEST_VMALLOC_MODULE
2563EXPORT_SYMBOL_GPL(__vmalloc_node);
2564#endif
2565
2566void *__vmalloc(unsigned long size, gfp_t gfp_mask)
2567{
2568 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
2569 __builtin_return_address(0));
2570}
2571EXPORT_SYMBOL(__vmalloc);
2572
2573/**
2574 * vmalloc - allocate virtually contiguous memory
2575 * @size: allocation size
2576 *
2577 * Allocate enough pages to cover @size from the page level
2578 * allocator and map them into contiguous kernel virtual space.
2579 *
2580 * For tight control over page level allocator and protection flags
2581 * use __vmalloc() instead.
2582 *
2583 * Return: pointer to the allocated memory or %NULL on error
2584 */
2585void *vmalloc(unsigned long size)
2586{
2587 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
2588 __builtin_return_address(0));
2589}
2590EXPORT_SYMBOL(vmalloc);
2591
2592/**
2593 * vzalloc - allocate virtually contiguous memory with zero fill
2594 * @size: allocation size
2595 *
2596 * Allocate enough pages to cover @size from the page level
2597 * allocator and map them into contiguous kernel virtual space.
2598 * The memory allocated is set to zero.
2599 *
2600 * For tight control over page level allocator and protection flags
2601 * use __vmalloc() instead.
2602 *
2603 * Return: pointer to the allocated memory or %NULL on error
2604 */
2605void *vzalloc(unsigned long size)
2606{
2607 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
2608 __builtin_return_address(0));
2609}
2610EXPORT_SYMBOL(vzalloc);
2611
2612/**
2613 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2614 * @size: allocation size
2615 *
2616 * The resulting memory area is zeroed so it can be mapped to userspace
2617 * without leaking data.
2618 *
2619 * Return: pointer to the allocated memory or %NULL on error
2620 */
2621void *vmalloc_user(unsigned long size)
2622{
2623 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2624 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2625 VM_USERMAP, NUMA_NO_NODE,
2626 __builtin_return_address(0));
2627}
2628EXPORT_SYMBOL(vmalloc_user);
2629
2630/**
2631 * vmalloc_node - allocate memory on a specific node
2632 * @size: allocation size
2633 * @node: numa node
2634 *
2635 * Allocate enough pages to cover @size from the page level
2636 * allocator and map them into contiguous kernel virtual space.
2637 *
2638 * For tight control over page level allocator and protection flags
2639 * use __vmalloc() instead.
2640 *
2641 * Return: pointer to the allocated memory or %NULL on error
2642 */
2643void *vmalloc_node(unsigned long size, int node)
2644{
2645 return __vmalloc_node(size, 1, GFP_KERNEL, node,
2646 __builtin_return_address(0));
2647}
2648EXPORT_SYMBOL(vmalloc_node);
2649
2650/**
2651 * vzalloc_node - allocate memory on a specific node with zero fill
2652 * @size: allocation size
2653 * @node: numa node
2654 *
2655 * Allocate enough pages to cover @size from the page level
2656 * allocator and map them into contiguous kernel virtual space.
2657 * The memory allocated is set to zero.
2658 *
2659 * Return: pointer to the allocated memory or %NULL on error
2660 */
2661void *vzalloc_node(unsigned long size, int node)
2662{
2663 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
2664 __builtin_return_address(0));
2665}
2666EXPORT_SYMBOL(vzalloc_node);
2667
2668#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2669#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2670#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2671#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2672#else
2673/*
2674 * 64b systems should always have either DMA or DMA32 zones. For others
2675 * GFP_DMA32 should do the right thing and use the normal zone.
2676 */
2677#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2678#endif
2679
2680/**
2681 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2682 * @size: allocation size
2683 *
2684 * Allocate enough 32bit PA addressable pages to cover @size from the
2685 * page level allocator and map them into contiguous kernel virtual space.
2686 *
2687 * Return: pointer to the allocated memory or %NULL on error
2688 */
2689void *vmalloc_32(unsigned long size)
2690{
2691 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
2692 __builtin_return_address(0));
2693}
2694EXPORT_SYMBOL(vmalloc_32);
2695
2696/**
2697 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2698 * @size: allocation size
2699 *
2700 * The resulting memory area is 32bit addressable and zeroed so it can be
2701 * mapped to userspace without leaking data.
2702 *
2703 * Return: pointer to the allocated memory or %NULL on error
2704 */
2705void *vmalloc_32_user(unsigned long size)
2706{
2707 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2708 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2709 VM_USERMAP, NUMA_NO_NODE,
2710 __builtin_return_address(0));
2711}
2712EXPORT_SYMBOL(vmalloc_32_user);
2713
2714/*
2715 * small helper routine , copy contents to buf from addr.
2716 * If the page is not present, fill zero.
2717 */
2718
2719static int aligned_vread(char *buf, char *addr, unsigned long count)
2720{
2721 struct page *p;
2722 int copied = 0;
2723
2724 while (count) {
2725 unsigned long offset, length;
2726
2727 offset = offset_in_page(addr);
2728 length = PAGE_SIZE - offset;
2729 if (length > count)
2730 length = count;
2731 p = vmalloc_to_page(addr);
2732 /*
2733 * To do safe access to this _mapped_ area, we need
2734 * lock. But adding lock here means that we need to add
2735 * overhead of vmalloc()/vfree() calles for this _debug_
2736 * interface, rarely used. Instead of that, we'll use
2737 * kmap() and get small overhead in this access function.
2738 */
2739 if (p) {
2740 /*
2741 * we can expect USER0 is not used (see vread/vwrite's
2742 * function description)
2743 */
2744 void *map = kmap_atomic(p);
2745 memcpy(buf, map + offset, length);
2746 kunmap_atomic(map);
2747 } else
2748 memset(buf, 0, length);
2749
2750 addr += length;
2751 buf += length;
2752 copied += length;
2753 count -= length;
2754 }
2755 return copied;
2756}
2757
2758static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2759{
2760 struct page *p;
2761 int copied = 0;
2762
2763 while (count) {
2764 unsigned long offset, length;
2765
2766 offset = offset_in_page(addr);
2767 length = PAGE_SIZE - offset;
2768 if (length > count)
2769 length = count;
2770 p = vmalloc_to_page(addr);
2771 /*
2772 * To do safe access to this _mapped_ area, we need
2773 * lock. But adding lock here means that we need to add
2774 * overhead of vmalloc()/vfree() calles for this _debug_
2775 * interface, rarely used. Instead of that, we'll use
2776 * kmap() and get small overhead in this access function.
2777 */
2778 if (p) {
2779 /*
2780 * we can expect USER0 is not used (see vread/vwrite's
2781 * function description)
2782 */
2783 void *map = kmap_atomic(p);
2784 memcpy(map + offset, buf, length);
2785 kunmap_atomic(map);
2786 }
2787 addr += length;
2788 buf += length;
2789 copied += length;
2790 count -= length;
2791 }
2792 return copied;
2793}
2794
2795/**
2796 * vread() - read vmalloc area in a safe way.
2797 * @buf: buffer for reading data
2798 * @addr: vm address.
2799 * @count: number of bytes to be read.
2800 *
2801 * This function checks that addr is a valid vmalloc'ed area, and
2802 * copy data from that area to a given buffer. If the given memory range
2803 * of [addr...addr+count) includes some valid address, data is copied to
2804 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2805 * IOREMAP area is treated as memory hole and no copy is done.
2806 *
2807 * If [addr...addr+count) doesn't includes any intersects with alive
2808 * vm_struct area, returns 0. @buf should be kernel's buffer.
2809 *
2810 * Note: In usual ops, vread() is never necessary because the caller
2811 * should know vmalloc() area is valid and can use memcpy().
2812 * This is for routines which have to access vmalloc area without
2813 * any information, as /dev/kmem.
2814 *
2815 * Return: number of bytes for which addr and buf should be increased
2816 * (same number as @count) or %0 if [addr...addr+count) doesn't
2817 * include any intersection with valid vmalloc area
2818 */
2819long vread(char *buf, char *addr, unsigned long count)
2820{
2821 struct vmap_area *va;
2822 struct vm_struct *vm;
2823 char *vaddr, *buf_start = buf;
2824 unsigned long buflen = count;
2825 unsigned long n;
2826
2827 /* Don't allow overflow */
2828 if ((unsigned long) addr + count < count)
2829 count = -(unsigned long) addr;
2830
2831 spin_lock(&vmap_area_lock);
2832 list_for_each_entry(va, &vmap_area_list, list) {
2833 if (!count)
2834 break;
2835
2836 if (!va->vm)
2837 continue;
2838
2839 vm = va->vm;
2840 vaddr = (char *) vm->addr;
2841 if (addr >= vaddr + get_vm_area_size(vm))
2842 continue;
2843 while (addr < vaddr) {
2844 if (count == 0)
2845 goto finished;
2846 *buf = '\0';
2847 buf++;
2848 addr++;
2849 count--;
2850 }
2851 n = vaddr + get_vm_area_size(vm) - addr;
2852 if (n > count)
2853 n = count;
2854 if (!(vm->flags & VM_IOREMAP))
2855 aligned_vread(buf, addr, n);
2856 else /* IOREMAP area is treated as memory hole */
2857 memset(buf, 0, n);
2858 buf += n;
2859 addr += n;
2860 count -= n;
2861 }
2862finished:
2863 spin_unlock(&vmap_area_lock);
2864
2865 if (buf == buf_start)
2866 return 0;
2867 /* zero-fill memory holes */
2868 if (buf != buf_start + buflen)
2869 memset(buf, 0, buflen - (buf - buf_start));
2870
2871 return buflen;
2872}
2873
2874/**
2875 * vwrite() - write vmalloc area in a safe way.
2876 * @buf: buffer for source data
2877 * @addr: vm address.
2878 * @count: number of bytes to be read.
2879 *
2880 * This function checks that addr is a valid vmalloc'ed area, and
2881 * copy data from a buffer to the given addr. If specified range of
2882 * [addr...addr+count) includes some valid address, data is copied from
2883 * proper area of @buf. If there are memory holes, no copy to hole.
2884 * IOREMAP area is treated as memory hole and no copy is done.
2885 *
2886 * If [addr...addr+count) doesn't includes any intersects with alive
2887 * vm_struct area, returns 0. @buf should be kernel's buffer.
2888 *
2889 * Note: In usual ops, vwrite() is never necessary because the caller
2890 * should know vmalloc() area is valid and can use memcpy().
2891 * This is for routines which have to access vmalloc area without
2892 * any information, as /dev/kmem.
2893 *
2894 * Return: number of bytes for which addr and buf should be
2895 * increased (same number as @count) or %0 if [addr...addr+count)
2896 * doesn't include any intersection with valid vmalloc area
2897 */
2898long vwrite(char *buf, char *addr, unsigned long count)
2899{
2900 struct vmap_area *va;
2901 struct vm_struct *vm;
2902 char *vaddr;
2903 unsigned long n, buflen;
2904 int copied = 0;
2905
2906 /* Don't allow overflow */
2907 if ((unsigned long) addr + count < count)
2908 count = -(unsigned long) addr;
2909 buflen = count;
2910
2911 spin_lock(&vmap_area_lock);
2912 list_for_each_entry(va, &vmap_area_list, list) {
2913 if (!count)
2914 break;
2915
2916 if (!va->vm)
2917 continue;
2918
2919 vm = va->vm;
2920 vaddr = (char *) vm->addr;
2921 if (addr >= vaddr + get_vm_area_size(vm))
2922 continue;
2923 while (addr < vaddr) {
2924 if (count == 0)
2925 goto finished;
2926 buf++;
2927 addr++;
2928 count--;
2929 }
2930 n = vaddr + get_vm_area_size(vm) - addr;
2931 if (n > count)
2932 n = count;
2933 if (!(vm->flags & VM_IOREMAP)) {
2934 aligned_vwrite(buf, addr, n);
2935 copied++;
2936 }
2937 buf += n;
2938 addr += n;
2939 count -= n;
2940 }
2941finished:
2942 spin_unlock(&vmap_area_lock);
2943 if (!copied)
2944 return 0;
2945 return buflen;
2946}
2947
2948/**
2949 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2950 * @vma: vma to cover
2951 * @uaddr: target user address to start at
2952 * @kaddr: virtual address of vmalloc kernel memory
2953 * @pgoff: offset from @kaddr to start at
2954 * @size: size of map area
2955 *
2956 * Returns: 0 for success, -Exxx on failure
2957 *
2958 * This function checks that @kaddr is a valid vmalloc'ed area,
2959 * and that it is big enough to cover the range starting at
2960 * @uaddr in @vma. Will return failure if that criteria isn't
2961 * met.
2962 *
2963 * Similar to remap_pfn_range() (see mm/memory.c)
2964 */
2965int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2966 void *kaddr, unsigned long pgoff,
2967 unsigned long size)
2968{
2969 struct vm_struct *area;
2970 unsigned long off;
2971 unsigned long end_index;
2972
2973 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
2974 return -EINVAL;
2975
2976 size = PAGE_ALIGN(size);
2977
2978 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2979 return -EINVAL;
2980
2981 area = find_vm_area(kaddr);
2982 if (!area)
2983 return -EINVAL;
2984
2985 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
2986 return -EINVAL;
2987
2988 if (check_add_overflow(size, off, &end_index) ||
2989 end_index > get_vm_area_size(area))
2990 return -EINVAL;
2991 kaddr += off;
2992
2993 do {
2994 struct page *page = vmalloc_to_page(kaddr);
2995 int ret;
2996
2997 ret = vm_insert_page(vma, uaddr, page);
2998 if (ret)
2999 return ret;
3000
3001 uaddr += PAGE_SIZE;
3002 kaddr += PAGE_SIZE;
3003 size -= PAGE_SIZE;
3004 } while (size > 0);
3005
3006 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3007
3008 return 0;
3009}
3010EXPORT_SYMBOL(remap_vmalloc_range_partial);
3011
3012/**
3013 * remap_vmalloc_range - map vmalloc pages to userspace
3014 * @vma: vma to cover (map full range of vma)
3015 * @addr: vmalloc memory
3016 * @pgoff: number of pages into addr before first page to map
3017 *
3018 * Returns: 0 for success, -Exxx on failure
3019 *
3020 * This function checks that addr is a valid vmalloc'ed area, and
3021 * that it is big enough to cover the vma. Will return failure if
3022 * that criteria isn't met.
3023 *
3024 * Similar to remap_pfn_range() (see mm/memory.c)
3025 */
3026int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3027 unsigned long pgoff)
3028{
3029 return remap_vmalloc_range_partial(vma, vma->vm_start,
3030 addr, pgoff,
3031 vma->vm_end - vma->vm_start);
3032}
3033EXPORT_SYMBOL(remap_vmalloc_range);
3034
3035static int f(pte_t *pte, unsigned long addr, void *data)
3036{
3037 pte_t ***p = data;
3038
3039 if (p) {
3040 *(*p) = pte;
3041 (*p)++;
3042 }
3043 return 0;
3044}
3045
3046/**
3047 * alloc_vm_area - allocate a range of kernel address space
3048 * @size: size of the area
3049 * @ptes: returns the PTEs for the address space
3050 *
3051 * Returns: NULL on failure, vm_struct on success
3052 *
3053 * This function reserves a range of kernel address space, and
3054 * allocates pagetables to map that range. No actual mappings
3055 * are created.
3056 *
3057 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3058 * allocated for the VM area are returned.
3059 */
3060struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3061{
3062 struct vm_struct *area;
3063
3064 area = get_vm_area_caller(size, VM_IOREMAP,
3065 __builtin_return_address(0));
3066 if (area == NULL)
3067 return NULL;
3068
3069 /*
3070 * This ensures that page tables are constructed for this region
3071 * of kernel virtual address space and mapped into init_mm.
3072 */
3073 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3074 size, f, ptes ? &ptes : NULL)) {
3075 free_vm_area(area);
3076 return NULL;
3077 }
3078
3079 return area;
3080}
3081EXPORT_SYMBOL_GPL(alloc_vm_area);
3082
3083void free_vm_area(struct vm_struct *area)
3084{
3085 struct vm_struct *ret;
3086 ret = remove_vm_area(area->addr);
3087 BUG_ON(ret != area);
3088 kfree(area);
3089}
3090EXPORT_SYMBOL_GPL(free_vm_area);
3091
3092#ifdef CONFIG_SMP
3093static struct vmap_area *node_to_va(struct rb_node *n)
3094{
3095 return rb_entry_safe(n, struct vmap_area, rb_node);
3096}
3097
3098/**
3099 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3100 * @addr: target address
3101 *
3102 * Returns: vmap_area if it is found. If there is no such area
3103 * the first highest(reverse order) vmap_area is returned
3104 * i.e. va->va_start < addr && va->va_end < addr or NULL
3105 * if there are no any areas before @addr.
3106 */
3107static struct vmap_area *
3108pvm_find_va_enclose_addr(unsigned long addr)
3109{
3110 struct vmap_area *va, *tmp;
3111 struct rb_node *n;
3112
3113 n = free_vmap_area_root.rb_node;
3114 va = NULL;
3115
3116 while (n) {
3117 tmp = rb_entry(n, struct vmap_area, rb_node);
3118 if (tmp->va_start <= addr) {
3119 va = tmp;
3120 if (tmp->va_end >= addr)
3121 break;
3122
3123 n = n->rb_right;
3124 } else {
3125 n = n->rb_left;
3126 }
3127 }
3128
3129 return va;
3130}
3131
3132/**
3133 * pvm_determine_end_from_reverse - find the highest aligned address
3134 * of free block below VMALLOC_END
3135 * @va:
3136 * in - the VA we start the search(reverse order);
3137 * out - the VA with the highest aligned end address.
3138 *
3139 * Returns: determined end address within vmap_area
3140 */
3141static unsigned long
3142pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3143{
3144 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3145 unsigned long addr;
3146
3147 if (likely(*va)) {
3148 list_for_each_entry_from_reverse((*va),
3149 &free_vmap_area_list, list) {
3150 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3151 if ((*va)->va_start < addr)
3152 return addr;
3153 }
3154 }
3155
3156 return 0;
3157}
3158
3159/**
3160 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3161 * @offsets: array containing offset of each area
3162 * @sizes: array containing size of each area
3163 * @nr_vms: the number of areas to allocate
3164 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3165 *
3166 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3167 * vm_structs on success, %NULL on failure
3168 *
3169 * Percpu allocator wants to use congruent vm areas so that it can
3170 * maintain the offsets among percpu areas. This function allocates
3171 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3172 * be scattered pretty far, distance between two areas easily going up
3173 * to gigabytes. To avoid interacting with regular vmallocs, these
3174 * areas are allocated from top.
3175 *
3176 * Despite its complicated look, this allocator is rather simple. It
3177 * does everything top-down and scans free blocks from the end looking
3178 * for matching base. While scanning, if any of the areas do not fit the
3179 * base address is pulled down to fit the area. Scanning is repeated till
3180 * all the areas fit and then all necessary data structures are inserted
3181 * and the result is returned.
3182 */
3183struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3184 const size_t *sizes, int nr_vms,
3185 size_t align)
3186{
3187 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3188 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3189 struct vmap_area **vas, *va;
3190 struct vm_struct **vms;
3191 int area, area2, last_area, term_area;
3192 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3193 bool purged = false;
3194 enum fit_type type;
3195
3196 /* verify parameters and allocate data structures */
3197 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3198 for (last_area = 0, area = 0; area < nr_vms; area++) {
3199 start = offsets[area];
3200 end = start + sizes[area];
3201
3202 /* is everything aligned properly? */
3203 BUG_ON(!IS_ALIGNED(offsets[area], align));
3204 BUG_ON(!IS_ALIGNED(sizes[area], align));
3205
3206 /* detect the area with the highest address */
3207 if (start > offsets[last_area])
3208 last_area = area;
3209
3210 for (area2 = area + 1; area2 < nr_vms; area2++) {
3211 unsigned long start2 = offsets[area2];
3212 unsigned long end2 = start2 + sizes[area2];
3213
3214 BUG_ON(start2 < end && start < end2);
3215 }
3216 }
3217 last_end = offsets[last_area] + sizes[last_area];
3218
3219 if (vmalloc_end - vmalloc_start < last_end) {
3220 WARN_ON(true);
3221 return NULL;
3222 }
3223
3224 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3225 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3226 if (!vas || !vms)
3227 goto err_free2;
3228
3229 for (area = 0; area < nr_vms; area++) {
3230 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3231 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3232 if (!vas[area] || !vms[area])
3233 goto err_free;
3234 }
3235retry:
3236 spin_lock(&free_vmap_area_lock);
3237
3238 /* start scanning - we scan from the top, begin with the last area */
3239 area = term_area = last_area;
3240 start = offsets[area];
3241 end = start + sizes[area];
3242
3243 va = pvm_find_va_enclose_addr(vmalloc_end);
3244 base = pvm_determine_end_from_reverse(&va, align) - end;
3245
3246 while (true) {
3247 /*
3248 * base might have underflowed, add last_end before
3249 * comparing.
3250 */
3251 if (base + last_end < vmalloc_start + last_end)
3252 goto overflow;
3253
3254 /*
3255 * Fitting base has not been found.
3256 */
3257 if (va == NULL)
3258 goto overflow;
3259
3260 /*
3261 * If required width exceeds current VA block, move
3262 * base downwards and then recheck.
3263 */
3264 if (base + end > va->va_end) {
3265 base = pvm_determine_end_from_reverse(&va, align) - end;
3266 term_area = area;
3267 continue;
3268 }
3269
3270 /*
3271 * If this VA does not fit, move base downwards and recheck.
3272 */
3273 if (base + start < va->va_start) {
3274 va = node_to_va(rb_prev(&va->rb_node));
3275 base = pvm_determine_end_from_reverse(&va, align) - end;
3276 term_area = area;
3277 continue;
3278 }
3279
3280 /*
3281 * This area fits, move on to the previous one. If
3282 * the previous one is the terminal one, we're done.
3283 */
3284 area = (area + nr_vms - 1) % nr_vms;
3285 if (area == term_area)
3286 break;
3287
3288 start = offsets[area];
3289 end = start + sizes[area];
3290 va = pvm_find_va_enclose_addr(base + end);
3291 }
3292
3293 /* we've found a fitting base, insert all va's */
3294 for (area = 0; area < nr_vms; area++) {
3295 int ret;
3296
3297 start = base + offsets[area];
3298 size = sizes[area];
3299
3300 va = pvm_find_va_enclose_addr(start);
3301 if (WARN_ON_ONCE(va == NULL))
3302 /* It is a BUG(), but trigger recovery instead. */
3303 goto recovery;
3304
3305 type = classify_va_fit_type(va, start, size);
3306 if (WARN_ON_ONCE(type == NOTHING_FIT))
3307 /* It is a BUG(), but trigger recovery instead. */
3308 goto recovery;
3309
3310 ret = adjust_va_to_fit_type(va, start, size, type);
3311 if (unlikely(ret))
3312 goto recovery;
3313
3314 /* Allocated area. */
3315 va = vas[area];
3316 va->va_start = start;
3317 va->va_end = start + size;
3318 }
3319
3320 spin_unlock(&free_vmap_area_lock);
3321
3322 /* populate the kasan shadow space */
3323 for (area = 0; area < nr_vms; area++) {
3324 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3325 goto err_free_shadow;
3326
3327 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3328 sizes[area]);
3329 }
3330
3331 /* insert all vm's */
3332 spin_lock(&vmap_area_lock);
3333 for (area = 0; area < nr_vms; area++) {
3334 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3335
3336 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3337 pcpu_get_vm_areas);
3338 }
3339 spin_unlock(&vmap_area_lock);
3340
3341 kfree(vas);
3342 return vms;
3343
3344recovery:
3345 /*
3346 * Remove previously allocated areas. There is no
3347 * need in removing these areas from the busy tree,
3348 * because they are inserted only on the final step
3349 * and when pcpu_get_vm_areas() is success.
3350 */
3351 while (area--) {
3352 orig_start = vas[area]->va_start;
3353 orig_end = vas[area]->va_end;
3354 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3355 &free_vmap_area_list);
3356 if (va)
3357 kasan_release_vmalloc(orig_start, orig_end,
3358 va->va_start, va->va_end);
3359 vas[area] = NULL;
3360 }
3361
3362overflow:
3363 spin_unlock(&free_vmap_area_lock);
3364 if (!purged) {
3365 purge_vmap_area_lazy();
3366 purged = true;
3367
3368 /* Before "retry", check if we recover. */
3369 for (area = 0; area < nr_vms; area++) {
3370 if (vas[area])
3371 continue;
3372
3373 vas[area] = kmem_cache_zalloc(
3374 vmap_area_cachep, GFP_KERNEL);
3375 if (!vas[area])
3376 goto err_free;
3377 }
3378
3379 goto retry;
3380 }
3381
3382err_free:
3383 for (area = 0; area < nr_vms; area++) {
3384 if (vas[area])
3385 kmem_cache_free(vmap_area_cachep, vas[area]);
3386
3387 kfree(vms[area]);
3388 }
3389err_free2:
3390 kfree(vas);
3391 kfree(vms);
3392 return NULL;
3393
3394err_free_shadow:
3395 spin_lock(&free_vmap_area_lock);
3396 /*
3397 * We release all the vmalloc shadows, even the ones for regions that
3398 * hadn't been successfully added. This relies on kasan_release_vmalloc
3399 * being able to tolerate this case.
3400 */
3401 for (area = 0; area < nr_vms; area++) {
3402 orig_start = vas[area]->va_start;
3403 orig_end = vas[area]->va_end;
3404 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3405 &free_vmap_area_list);
3406 if (va)
3407 kasan_release_vmalloc(orig_start, orig_end,
3408 va->va_start, va->va_end);
3409 vas[area] = NULL;
3410 kfree(vms[area]);
3411 }
3412 spin_unlock(&free_vmap_area_lock);
3413 kfree(vas);
3414 kfree(vms);
3415 return NULL;
3416}
3417
3418/**
3419 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3420 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3421 * @nr_vms: the number of allocated areas
3422 *
3423 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3424 */
3425void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3426{
3427 int i;
3428
3429 for (i = 0; i < nr_vms; i++)
3430 free_vm_area(vms[i]);
3431 kfree(vms);
3432}
3433#endif /* CONFIG_SMP */
3434
3435#ifdef CONFIG_PROC_FS
3436static void *s_start(struct seq_file *m, loff_t *pos)
3437 __acquires(&vmap_purge_lock)
3438 __acquires(&vmap_area_lock)
3439{
3440 mutex_lock(&vmap_purge_lock);
3441 spin_lock(&vmap_area_lock);
3442
3443 return seq_list_start(&vmap_area_list, *pos);
3444}
3445
3446static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3447{
3448 return seq_list_next(p, &vmap_area_list, pos);
3449}
3450
3451static void s_stop(struct seq_file *m, void *p)
3452 __releases(&vmap_purge_lock)
3453 __releases(&vmap_area_lock)
3454{
3455 mutex_unlock(&vmap_purge_lock);
3456 spin_unlock(&vmap_area_lock);
3457}
3458
3459static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3460{
3461 if (IS_ENABLED(CONFIG_NUMA)) {
3462 unsigned int nr, *counters = m->private;
3463
3464 if (!counters)
3465 return;
3466
3467 if (v->flags & VM_UNINITIALIZED)
3468 return;
3469 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3470 smp_rmb();
3471
3472 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3473
3474 for (nr = 0; nr < v->nr_pages; nr++)
3475 counters[page_to_nid(v->pages[nr])]++;
3476
3477 for_each_node_state(nr, N_HIGH_MEMORY)
3478 if (counters[nr])
3479 seq_printf(m, " N%u=%u", nr, counters[nr]);
3480 }
3481}
3482
3483static void show_purge_info(struct seq_file *m)
3484{
3485 struct llist_node *head;
3486 struct vmap_area *va;
3487
3488 head = READ_ONCE(vmap_purge_list.first);
3489 if (head == NULL)
3490 return;
3491
3492 llist_for_each_entry(va, head, purge_list) {
3493 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3494 (void *)va->va_start, (void *)va->va_end,
3495 va->va_end - va->va_start);
3496 }
3497}
3498
3499static int s_show(struct seq_file *m, void *p)
3500{
3501 struct vmap_area *va;
3502 struct vm_struct *v;
3503
3504 va = list_entry(p, struct vmap_area, list);
3505
3506 /*
3507 * s_show can encounter race with remove_vm_area, !vm on behalf
3508 * of vmap area is being tear down or vm_map_ram allocation.
3509 */
3510 if (!va->vm) {
3511 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3512 (void *)va->va_start, (void *)va->va_end,
3513 va->va_end - va->va_start);
3514
3515 return 0;
3516 }
3517
3518 v = va->vm;
3519
3520 seq_printf(m, "0x%pK-0x%pK %7ld",
3521 v->addr, v->addr + v->size, v->size);
3522
3523 if (v->caller)
3524 seq_printf(m, " %pS", v->caller);
3525
3526 if (v->nr_pages)
3527 seq_printf(m, " pages=%d", v->nr_pages);
3528
3529 if (v->phys_addr)
3530 seq_printf(m, " phys=%pa", &v->phys_addr);
3531
3532 if (v->flags & VM_IOREMAP)
3533 seq_puts(m, " ioremap");
3534
3535 if (v->flags & VM_ALLOC)
3536 seq_puts(m, " vmalloc");
3537
3538 if (v->flags & VM_MAP)
3539 seq_puts(m, " vmap");
3540
3541 if (v->flags & VM_USERMAP)
3542 seq_puts(m, " user");
3543
3544 if (v->flags & VM_DMA_COHERENT)
3545 seq_puts(m, " dma-coherent");
3546
3547 if (is_vmalloc_addr(v->pages))
3548 seq_puts(m, " vpages");
3549
3550 show_numa_info(m, v);
3551 seq_putc(m, '\n');
3552
3553 /*
3554 * As a final step, dump "unpurged" areas. Note,
3555 * that entire "/proc/vmallocinfo" output will not
3556 * be address sorted, because the purge list is not
3557 * sorted.
3558 */
3559 if (list_is_last(&va->list, &vmap_area_list))
3560 show_purge_info(m);
3561
3562 return 0;
3563}
3564
3565static const struct seq_operations vmalloc_op = {
3566 .start = s_start,
3567 .next = s_next,
3568 .stop = s_stop,
3569 .show = s_show,
3570};
3571
3572static int __init proc_vmalloc_init(void)
3573{
3574 if (IS_ENABLED(CONFIG_NUMA))
3575 proc_create_seq_private("vmallocinfo", 0400, NULL,
3576 &vmalloc_op,
3577 nr_node_ids * sizeof(unsigned int), NULL);
3578 else
3579 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3580 return 0;
3581}
3582module_init(proc_vmalloc_init);
3583
3584#endif