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