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#include <linux/page_owner.h>
46
47#define CREATE_TRACE_POINTS
48#include <trace/events/vmalloc.h>
49
50#include "internal.h"
51#include "pgalloc-track.h"
52
53#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
54static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
55
56static int __init set_nohugeiomap(char *str)
57{
58 ioremap_max_page_shift = PAGE_SHIFT;
59 return 0;
60}
61early_param("nohugeiomap", set_nohugeiomap);
62#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
63static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
64#endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */
65
66#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
67static bool __ro_after_init vmap_allow_huge = true;
68
69static int __init set_nohugevmalloc(char *str)
70{
71 vmap_allow_huge = false;
72 return 0;
73}
74early_param("nohugevmalloc", set_nohugevmalloc);
75#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
76static const bool vmap_allow_huge = false;
77#endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
78
79bool is_vmalloc_addr(const void *x)
80{
81 unsigned long addr = (unsigned long)kasan_reset_tag(x);
82
83 return addr >= VMALLOC_START && addr < VMALLOC_END;
84}
85EXPORT_SYMBOL(is_vmalloc_addr);
86
87struct vfree_deferred {
88 struct llist_head list;
89 struct work_struct wq;
90};
91static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
92
93/*** Page table manipulation functions ***/
94static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
95 phys_addr_t phys_addr, pgprot_t prot,
96 unsigned int max_page_shift, pgtbl_mod_mask *mask)
97{
98 pte_t *pte;
99 u64 pfn;
100 struct page *page;
101 unsigned long size = PAGE_SIZE;
102
103 pfn = phys_addr >> PAGE_SHIFT;
104 pte = pte_alloc_kernel_track(pmd, addr, mask);
105 if (!pte)
106 return -ENOMEM;
107 do {
108 if (unlikely(!pte_none(ptep_get(pte)))) {
109 if (pfn_valid(pfn)) {
110 page = pfn_to_page(pfn);
111 dump_page(page, "remapping already mapped page");
112 }
113 BUG();
114 }
115
116#ifdef CONFIG_HUGETLB_PAGE
117 size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
118 if (size != PAGE_SIZE) {
119 pte_t entry = pfn_pte(pfn, prot);
120
121 entry = arch_make_huge_pte(entry, ilog2(size), 0);
122 set_huge_pte_at(&init_mm, addr, pte, entry, size);
123 pfn += PFN_DOWN(size);
124 continue;
125 }
126#endif
127 set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
128 pfn++;
129 } while (pte += PFN_DOWN(size), addr += size, addr != end);
130 *mask |= PGTBL_PTE_MODIFIED;
131 return 0;
132}
133
134static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
135 phys_addr_t phys_addr, pgprot_t prot,
136 unsigned int max_page_shift)
137{
138 if (max_page_shift < PMD_SHIFT)
139 return 0;
140
141 if (!arch_vmap_pmd_supported(prot))
142 return 0;
143
144 if ((end - addr) != PMD_SIZE)
145 return 0;
146
147 if (!IS_ALIGNED(addr, PMD_SIZE))
148 return 0;
149
150 if (!IS_ALIGNED(phys_addr, PMD_SIZE))
151 return 0;
152
153 if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
154 return 0;
155
156 return pmd_set_huge(pmd, phys_addr, prot);
157}
158
159static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
160 phys_addr_t phys_addr, pgprot_t prot,
161 unsigned int max_page_shift, pgtbl_mod_mask *mask)
162{
163 pmd_t *pmd;
164 unsigned long next;
165
166 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
167 if (!pmd)
168 return -ENOMEM;
169 do {
170 next = pmd_addr_end(addr, end);
171
172 if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
173 max_page_shift)) {
174 *mask |= PGTBL_PMD_MODIFIED;
175 continue;
176 }
177
178 if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
179 return -ENOMEM;
180 } while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
181 return 0;
182}
183
184static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
185 phys_addr_t phys_addr, pgprot_t prot,
186 unsigned int max_page_shift)
187{
188 if (max_page_shift < PUD_SHIFT)
189 return 0;
190
191 if (!arch_vmap_pud_supported(prot))
192 return 0;
193
194 if ((end - addr) != PUD_SIZE)
195 return 0;
196
197 if (!IS_ALIGNED(addr, PUD_SIZE))
198 return 0;
199
200 if (!IS_ALIGNED(phys_addr, PUD_SIZE))
201 return 0;
202
203 if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
204 return 0;
205
206 return pud_set_huge(pud, phys_addr, prot);
207}
208
209static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
210 phys_addr_t phys_addr, pgprot_t prot,
211 unsigned int max_page_shift, pgtbl_mod_mask *mask)
212{
213 pud_t *pud;
214 unsigned long next;
215
216 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
217 if (!pud)
218 return -ENOMEM;
219 do {
220 next = pud_addr_end(addr, end);
221
222 if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
223 max_page_shift)) {
224 *mask |= PGTBL_PUD_MODIFIED;
225 continue;
226 }
227
228 if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
229 max_page_shift, mask))
230 return -ENOMEM;
231 } while (pud++, phys_addr += (next - addr), addr = next, addr != end);
232 return 0;
233}
234
235static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
236 phys_addr_t phys_addr, pgprot_t prot,
237 unsigned int max_page_shift)
238{
239 if (max_page_shift < P4D_SHIFT)
240 return 0;
241
242 if (!arch_vmap_p4d_supported(prot))
243 return 0;
244
245 if ((end - addr) != P4D_SIZE)
246 return 0;
247
248 if (!IS_ALIGNED(addr, P4D_SIZE))
249 return 0;
250
251 if (!IS_ALIGNED(phys_addr, P4D_SIZE))
252 return 0;
253
254 if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
255 return 0;
256
257 return p4d_set_huge(p4d, phys_addr, prot);
258}
259
260static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
261 phys_addr_t phys_addr, pgprot_t prot,
262 unsigned int max_page_shift, pgtbl_mod_mask *mask)
263{
264 p4d_t *p4d;
265 unsigned long next;
266
267 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
268 if (!p4d)
269 return -ENOMEM;
270 do {
271 next = p4d_addr_end(addr, end);
272
273 if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
274 max_page_shift)) {
275 *mask |= PGTBL_P4D_MODIFIED;
276 continue;
277 }
278
279 if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
280 max_page_shift, mask))
281 return -ENOMEM;
282 } while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
283 return 0;
284}
285
286static int vmap_range_noflush(unsigned long addr, unsigned long end,
287 phys_addr_t phys_addr, pgprot_t prot,
288 unsigned int max_page_shift)
289{
290 pgd_t *pgd;
291 unsigned long start;
292 unsigned long next;
293 int err;
294 pgtbl_mod_mask mask = 0;
295
296 might_sleep();
297 BUG_ON(addr >= end);
298
299 start = addr;
300 pgd = pgd_offset_k(addr);
301 do {
302 next = pgd_addr_end(addr, end);
303 err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
304 max_page_shift, &mask);
305 if (err)
306 break;
307 } while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
308
309 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
310 arch_sync_kernel_mappings(start, end);
311
312 return err;
313}
314
315int vmap_page_range(unsigned long addr, unsigned long end,
316 phys_addr_t phys_addr, pgprot_t prot)
317{
318 int err;
319
320 err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
321 ioremap_max_page_shift);
322 flush_cache_vmap(addr, end);
323 if (!err)
324 err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
325 ioremap_max_page_shift);
326 return err;
327}
328
329int ioremap_page_range(unsigned long addr, unsigned long end,
330 phys_addr_t phys_addr, pgprot_t prot)
331{
332 struct vm_struct *area;
333
334 area = find_vm_area((void *)addr);
335 if (!area || !(area->flags & VM_IOREMAP)) {
336 WARN_ONCE(1, "vm_area at addr %lx is not marked as VM_IOREMAP\n", addr);
337 return -EINVAL;
338 }
339 if (addr != (unsigned long)area->addr ||
340 (void *)end != area->addr + get_vm_area_size(area)) {
341 WARN_ONCE(1, "ioremap request [%lx,%lx) doesn't match vm_area [%lx, %lx)\n",
342 addr, end, (long)area->addr,
343 (long)area->addr + get_vm_area_size(area));
344 return -ERANGE;
345 }
346 return vmap_page_range(addr, end, phys_addr, prot);
347}
348
349static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
350 pgtbl_mod_mask *mask)
351{
352 pte_t *pte;
353
354 pte = pte_offset_kernel(pmd, addr);
355 do {
356 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
357 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
358 } while (pte++, addr += PAGE_SIZE, addr != end);
359 *mask |= PGTBL_PTE_MODIFIED;
360}
361
362static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
363 pgtbl_mod_mask *mask)
364{
365 pmd_t *pmd;
366 unsigned long next;
367 int cleared;
368
369 pmd = pmd_offset(pud, addr);
370 do {
371 next = pmd_addr_end(addr, end);
372
373 cleared = pmd_clear_huge(pmd);
374 if (cleared || pmd_bad(*pmd))
375 *mask |= PGTBL_PMD_MODIFIED;
376
377 if (cleared)
378 continue;
379 if (pmd_none_or_clear_bad(pmd))
380 continue;
381 vunmap_pte_range(pmd, addr, next, mask);
382
383 cond_resched();
384 } while (pmd++, addr = next, addr != end);
385}
386
387static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
388 pgtbl_mod_mask *mask)
389{
390 pud_t *pud;
391 unsigned long next;
392 int cleared;
393
394 pud = pud_offset(p4d, addr);
395 do {
396 next = pud_addr_end(addr, end);
397
398 cleared = pud_clear_huge(pud);
399 if (cleared || pud_bad(*pud))
400 *mask |= PGTBL_PUD_MODIFIED;
401
402 if (cleared)
403 continue;
404 if (pud_none_or_clear_bad(pud))
405 continue;
406 vunmap_pmd_range(pud, addr, next, mask);
407 } while (pud++, addr = next, addr != end);
408}
409
410static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
411 pgtbl_mod_mask *mask)
412{
413 p4d_t *p4d;
414 unsigned long next;
415
416 p4d = p4d_offset(pgd, addr);
417 do {
418 next = p4d_addr_end(addr, end);
419
420 p4d_clear_huge(p4d);
421 if (p4d_bad(*p4d))
422 *mask |= PGTBL_P4D_MODIFIED;
423
424 if (p4d_none_or_clear_bad(p4d))
425 continue;
426 vunmap_pud_range(p4d, addr, next, mask);
427 } while (p4d++, addr = next, addr != end);
428}
429
430/*
431 * vunmap_range_noflush is similar to vunmap_range, but does not
432 * flush caches or TLBs.
433 *
434 * The caller is responsible for calling flush_cache_vmap() before calling
435 * this function, and flush_tlb_kernel_range after it has returned
436 * successfully (and before the addresses are expected to cause a page fault
437 * or be re-mapped for something else, if TLB flushes are being delayed or
438 * coalesced).
439 *
440 * This is an internal function only. Do not use outside mm/.
441 */
442void __vunmap_range_noflush(unsigned long start, unsigned long end)
443{
444 unsigned long next;
445 pgd_t *pgd;
446 unsigned long addr = start;
447 pgtbl_mod_mask mask = 0;
448
449 BUG_ON(addr >= end);
450 pgd = pgd_offset_k(addr);
451 do {
452 next = pgd_addr_end(addr, end);
453 if (pgd_bad(*pgd))
454 mask |= PGTBL_PGD_MODIFIED;
455 if (pgd_none_or_clear_bad(pgd))
456 continue;
457 vunmap_p4d_range(pgd, addr, next, &mask);
458 } while (pgd++, addr = next, addr != end);
459
460 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
461 arch_sync_kernel_mappings(start, end);
462}
463
464void vunmap_range_noflush(unsigned long start, unsigned long end)
465{
466 kmsan_vunmap_range_noflush(start, end);
467 __vunmap_range_noflush(start, end);
468}
469
470/**
471 * vunmap_range - unmap kernel virtual addresses
472 * @addr: start of the VM area to unmap
473 * @end: end of the VM area to unmap (non-inclusive)
474 *
475 * Clears any present PTEs in the virtual address range, flushes TLBs and
476 * caches. Any subsequent access to the address before it has been re-mapped
477 * is a kernel bug.
478 */
479void vunmap_range(unsigned long addr, unsigned long end)
480{
481 flush_cache_vunmap(addr, end);
482 vunmap_range_noflush(addr, end);
483 flush_tlb_kernel_range(addr, end);
484}
485
486static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
487 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
488 pgtbl_mod_mask *mask)
489{
490 pte_t *pte;
491
492 /*
493 * nr is a running index into the array which helps higher level
494 * callers keep track of where we're up to.
495 */
496
497 pte = pte_alloc_kernel_track(pmd, addr, mask);
498 if (!pte)
499 return -ENOMEM;
500 do {
501 struct page *page = pages[*nr];
502
503 if (WARN_ON(!pte_none(ptep_get(pte))))
504 return -EBUSY;
505 if (WARN_ON(!page))
506 return -ENOMEM;
507 if (WARN_ON(!pfn_valid(page_to_pfn(page))))
508 return -EINVAL;
509
510 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
511 (*nr)++;
512 } while (pte++, addr += PAGE_SIZE, addr != end);
513 *mask |= PGTBL_PTE_MODIFIED;
514 return 0;
515}
516
517static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
518 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
519 pgtbl_mod_mask *mask)
520{
521 pmd_t *pmd;
522 unsigned long next;
523
524 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
525 if (!pmd)
526 return -ENOMEM;
527 do {
528 next = pmd_addr_end(addr, end);
529 if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
530 return -ENOMEM;
531 } while (pmd++, addr = next, addr != end);
532 return 0;
533}
534
535static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
536 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
537 pgtbl_mod_mask *mask)
538{
539 pud_t *pud;
540 unsigned long next;
541
542 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
543 if (!pud)
544 return -ENOMEM;
545 do {
546 next = pud_addr_end(addr, end);
547 if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
548 return -ENOMEM;
549 } while (pud++, addr = next, addr != end);
550 return 0;
551}
552
553static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
554 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
555 pgtbl_mod_mask *mask)
556{
557 p4d_t *p4d;
558 unsigned long next;
559
560 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
561 if (!p4d)
562 return -ENOMEM;
563 do {
564 next = p4d_addr_end(addr, end);
565 if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
566 return -ENOMEM;
567 } while (p4d++, addr = next, addr != end);
568 return 0;
569}
570
571static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
572 pgprot_t prot, struct page **pages)
573{
574 unsigned long start = addr;
575 pgd_t *pgd;
576 unsigned long next;
577 int err = 0;
578 int nr = 0;
579 pgtbl_mod_mask mask = 0;
580
581 BUG_ON(addr >= end);
582 pgd = pgd_offset_k(addr);
583 do {
584 next = pgd_addr_end(addr, end);
585 if (pgd_bad(*pgd))
586 mask |= PGTBL_PGD_MODIFIED;
587 err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
588 if (err)
589 break;
590 } while (pgd++, addr = next, addr != end);
591
592 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
593 arch_sync_kernel_mappings(start, end);
594
595 return err;
596}
597
598/*
599 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
600 * flush caches.
601 *
602 * The caller is responsible for calling flush_cache_vmap() after this
603 * function returns successfully and before the addresses are accessed.
604 *
605 * This is an internal function only. Do not use outside mm/.
606 */
607int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
608 pgprot_t prot, struct page **pages, unsigned int page_shift)
609{
610 unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
611
612 WARN_ON(page_shift < PAGE_SHIFT);
613
614 if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
615 page_shift == PAGE_SHIFT)
616 return vmap_small_pages_range_noflush(addr, end, prot, pages);
617
618 for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
619 int err;
620
621 err = vmap_range_noflush(addr, addr + (1UL << page_shift),
622 page_to_phys(pages[i]), prot,
623 page_shift);
624 if (err)
625 return err;
626
627 addr += 1UL << page_shift;
628 }
629
630 return 0;
631}
632
633int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
634 pgprot_t prot, struct page **pages, unsigned int page_shift)
635{
636 int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
637 page_shift);
638
639 if (ret)
640 return ret;
641 return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
642}
643
644/**
645 * vmap_pages_range - map pages to a kernel virtual address
646 * @addr: start of the VM area to map
647 * @end: end of the VM area to map (non-inclusive)
648 * @prot: page protection flags to use
649 * @pages: pages to map (always PAGE_SIZE pages)
650 * @page_shift: maximum shift that the pages may be mapped with, @pages must
651 * be aligned and contiguous up to at least this shift.
652 *
653 * RETURNS:
654 * 0 on success, -errno on failure.
655 */
656int vmap_pages_range(unsigned long addr, unsigned long end,
657 pgprot_t prot, struct page **pages, unsigned int page_shift)
658{
659 int err;
660
661 err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
662 flush_cache_vmap(addr, end);
663 return err;
664}
665
666static int check_sparse_vm_area(struct vm_struct *area, unsigned long start,
667 unsigned long end)
668{
669 might_sleep();
670 if (WARN_ON_ONCE(area->flags & VM_FLUSH_RESET_PERMS))
671 return -EINVAL;
672 if (WARN_ON_ONCE(area->flags & VM_NO_GUARD))
673 return -EINVAL;
674 if (WARN_ON_ONCE(!(area->flags & VM_SPARSE)))
675 return -EINVAL;
676 if ((end - start) >> PAGE_SHIFT > totalram_pages())
677 return -E2BIG;
678 if (start < (unsigned long)area->addr ||
679 (void *)end > area->addr + get_vm_area_size(area))
680 return -ERANGE;
681 return 0;
682}
683
684/**
685 * vm_area_map_pages - map pages inside given sparse vm_area
686 * @area: vm_area
687 * @start: start address inside vm_area
688 * @end: end address inside vm_area
689 * @pages: pages to map (always PAGE_SIZE pages)
690 */
691int vm_area_map_pages(struct vm_struct *area, unsigned long start,
692 unsigned long end, struct page **pages)
693{
694 int err;
695
696 err = check_sparse_vm_area(area, start, end);
697 if (err)
698 return err;
699
700 return vmap_pages_range(start, end, PAGE_KERNEL, pages, PAGE_SHIFT);
701}
702
703/**
704 * vm_area_unmap_pages - unmap pages inside given sparse vm_area
705 * @area: vm_area
706 * @start: start address inside vm_area
707 * @end: end address inside vm_area
708 */
709void vm_area_unmap_pages(struct vm_struct *area, unsigned long start,
710 unsigned long end)
711{
712 if (check_sparse_vm_area(area, start, end))
713 return;
714
715 vunmap_range(start, end);
716}
717
718int is_vmalloc_or_module_addr(const void *x)
719{
720 /*
721 * ARM, x86-64 and sparc64 put modules in a special place,
722 * and fall back on vmalloc() if that fails. Others
723 * just put it in the vmalloc space.
724 */
725#if defined(CONFIG_EXECMEM) && defined(MODULES_VADDR)
726 unsigned long addr = (unsigned long)kasan_reset_tag(x);
727 if (addr >= MODULES_VADDR && addr < MODULES_END)
728 return 1;
729#endif
730 return is_vmalloc_addr(x);
731}
732EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
733
734/*
735 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
736 * return the tail page that corresponds to the base page address, which
737 * matches small vmap mappings.
738 */
739struct page *vmalloc_to_page(const void *vmalloc_addr)
740{
741 unsigned long addr = (unsigned long) vmalloc_addr;
742 struct page *page = NULL;
743 pgd_t *pgd = pgd_offset_k(addr);
744 p4d_t *p4d;
745 pud_t *pud;
746 pmd_t *pmd;
747 pte_t *ptep, pte;
748
749 /*
750 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
751 * architectures that do not vmalloc module space
752 */
753 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
754
755 if (pgd_none(*pgd))
756 return NULL;
757 if (WARN_ON_ONCE(pgd_leaf(*pgd)))
758 return NULL; /* XXX: no allowance for huge pgd */
759 if (WARN_ON_ONCE(pgd_bad(*pgd)))
760 return NULL;
761
762 p4d = p4d_offset(pgd, addr);
763 if (p4d_none(*p4d))
764 return NULL;
765 if (p4d_leaf(*p4d))
766 return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
767 if (WARN_ON_ONCE(p4d_bad(*p4d)))
768 return NULL;
769
770 pud = pud_offset(p4d, addr);
771 if (pud_none(*pud))
772 return NULL;
773 if (pud_leaf(*pud))
774 return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
775 if (WARN_ON_ONCE(pud_bad(*pud)))
776 return NULL;
777
778 pmd = pmd_offset(pud, addr);
779 if (pmd_none(*pmd))
780 return NULL;
781 if (pmd_leaf(*pmd))
782 return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
783 if (WARN_ON_ONCE(pmd_bad(*pmd)))
784 return NULL;
785
786 ptep = pte_offset_kernel(pmd, addr);
787 pte = ptep_get(ptep);
788 if (pte_present(pte))
789 page = pte_page(pte);
790
791 return page;
792}
793EXPORT_SYMBOL(vmalloc_to_page);
794
795/*
796 * Map a vmalloc()-space virtual address to the physical page frame number.
797 */
798unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
799{
800 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
801}
802EXPORT_SYMBOL(vmalloc_to_pfn);
803
804
805/*** Global kva allocator ***/
806
807#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
808#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
809
810
811static DEFINE_SPINLOCK(free_vmap_area_lock);
812static bool vmap_initialized __read_mostly;
813
814/*
815 * This kmem_cache is used for vmap_area objects. Instead of
816 * allocating from slab we reuse an object from this cache to
817 * make things faster. Especially in "no edge" splitting of
818 * free block.
819 */
820static struct kmem_cache *vmap_area_cachep;
821
822/*
823 * This linked list is used in pair with free_vmap_area_root.
824 * It gives O(1) access to prev/next to perform fast coalescing.
825 */
826static LIST_HEAD(free_vmap_area_list);
827
828/*
829 * This augment red-black tree represents the free vmap space.
830 * All vmap_area objects in this tree are sorted by va->va_start
831 * address. It is used for allocation and merging when a vmap
832 * object is released.
833 *
834 * Each vmap_area node contains a maximum available free block
835 * of its sub-tree, right or left. Therefore it is possible to
836 * find a lowest match of free area.
837 */
838static struct rb_root free_vmap_area_root = RB_ROOT;
839
840/*
841 * Preload a CPU with one object for "no edge" split case. The
842 * aim is to get rid of allocations from the atomic context, thus
843 * to use more permissive allocation masks.
844 */
845static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
846
847/*
848 * This structure defines a single, solid model where a list and
849 * rb-tree are part of one entity protected by the lock. Nodes are
850 * sorted in ascending order, thus for O(1) access to left/right
851 * neighbors a list is used as well as for sequential traversal.
852 */
853struct rb_list {
854 struct rb_root root;
855 struct list_head head;
856 spinlock_t lock;
857};
858
859/*
860 * A fast size storage contains VAs up to 1M size. A pool consists
861 * of linked between each other ready to go VAs of certain sizes.
862 * An index in the pool-array corresponds to number of pages + 1.
863 */
864#define MAX_VA_SIZE_PAGES 256
865
866struct vmap_pool {
867 struct list_head head;
868 unsigned long len;
869};
870
871/*
872 * An effective vmap-node logic. Users make use of nodes instead
873 * of a global heap. It allows to balance an access and mitigate
874 * contention.
875 */
876static struct vmap_node {
877 /* Simple size segregated storage. */
878 struct vmap_pool pool[MAX_VA_SIZE_PAGES];
879 spinlock_t pool_lock;
880 bool skip_populate;
881
882 /* Bookkeeping data of this node. */
883 struct rb_list busy;
884 struct rb_list lazy;
885
886 /*
887 * Ready-to-free areas.
888 */
889 struct list_head purge_list;
890 struct work_struct purge_work;
891 unsigned long nr_purged;
892} single;
893
894/*
895 * Initial setup consists of one single node, i.e. a balancing
896 * is fully disabled. Later on, after vmap is initialized these
897 * parameters are updated based on a system capacity.
898 */
899static struct vmap_node *vmap_nodes = &single;
900static __read_mostly unsigned int nr_vmap_nodes = 1;
901static __read_mostly unsigned int vmap_zone_size = 1;
902
903static inline unsigned int
904addr_to_node_id(unsigned long addr)
905{
906 return (addr / vmap_zone_size) % nr_vmap_nodes;
907}
908
909static inline struct vmap_node *
910addr_to_node(unsigned long addr)
911{
912 return &vmap_nodes[addr_to_node_id(addr)];
913}
914
915static inline struct vmap_node *
916id_to_node(unsigned int id)
917{
918 return &vmap_nodes[id % nr_vmap_nodes];
919}
920
921/*
922 * We use the value 0 to represent "no node", that is why
923 * an encoded value will be the node-id incremented by 1.
924 * It is always greater then 0. A valid node_id which can
925 * be encoded is [0:nr_vmap_nodes - 1]. If a passed node_id
926 * is not valid 0 is returned.
927 */
928static unsigned int
929encode_vn_id(unsigned int node_id)
930{
931 /* Can store U8_MAX [0:254] nodes. */
932 if (node_id < nr_vmap_nodes)
933 return (node_id + 1) << BITS_PER_BYTE;
934
935 /* Warn and no node encoded. */
936 WARN_ONCE(1, "Encode wrong node id (%u)\n", node_id);
937 return 0;
938}
939
940/*
941 * Returns an encoded node-id, the valid range is within
942 * [0:nr_vmap_nodes-1] values. Otherwise nr_vmap_nodes is
943 * returned if extracted data is wrong.
944 */
945static unsigned int
946decode_vn_id(unsigned int val)
947{
948 unsigned int node_id = (val >> BITS_PER_BYTE) - 1;
949
950 /* Can store U8_MAX [0:254] nodes. */
951 if (node_id < nr_vmap_nodes)
952 return node_id;
953
954 /* If it was _not_ zero, warn. */
955 WARN_ONCE(node_id != UINT_MAX,
956 "Decode wrong node id (%d)\n", node_id);
957
958 return nr_vmap_nodes;
959}
960
961static bool
962is_vn_id_valid(unsigned int node_id)
963{
964 if (node_id < nr_vmap_nodes)
965 return true;
966
967 return false;
968}
969
970static __always_inline unsigned long
971va_size(struct vmap_area *va)
972{
973 return (va->va_end - va->va_start);
974}
975
976static __always_inline unsigned long
977get_subtree_max_size(struct rb_node *node)
978{
979 struct vmap_area *va;
980
981 va = rb_entry_safe(node, struct vmap_area, rb_node);
982 return va ? va->subtree_max_size : 0;
983}
984
985RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
986 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
987
988static void reclaim_and_purge_vmap_areas(void);
989static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
990static void drain_vmap_area_work(struct work_struct *work);
991static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
992
993static atomic_long_t nr_vmalloc_pages;
994
995unsigned long vmalloc_nr_pages(void)
996{
997 return atomic_long_read(&nr_vmalloc_pages);
998}
999
1000static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
1001{
1002 struct rb_node *n = root->rb_node;
1003
1004 addr = (unsigned long)kasan_reset_tag((void *)addr);
1005
1006 while (n) {
1007 struct vmap_area *va;
1008
1009 va = rb_entry(n, struct vmap_area, rb_node);
1010 if (addr < va->va_start)
1011 n = n->rb_left;
1012 else if (addr >= va->va_end)
1013 n = n->rb_right;
1014 else
1015 return va;
1016 }
1017
1018 return NULL;
1019}
1020
1021/* Look up the first VA which satisfies addr < va_end, NULL if none. */
1022static struct vmap_area *
1023__find_vmap_area_exceed_addr(unsigned long addr, struct rb_root *root)
1024{
1025 struct vmap_area *va = NULL;
1026 struct rb_node *n = root->rb_node;
1027
1028 addr = (unsigned long)kasan_reset_tag((void *)addr);
1029
1030 while (n) {
1031 struct vmap_area *tmp;
1032
1033 tmp = rb_entry(n, struct vmap_area, rb_node);
1034 if (tmp->va_end > addr) {
1035 va = tmp;
1036 if (tmp->va_start <= addr)
1037 break;
1038
1039 n = n->rb_left;
1040 } else
1041 n = n->rb_right;
1042 }
1043
1044 return va;
1045}
1046
1047/*
1048 * Returns a node where a first VA, that satisfies addr < va_end, resides.
1049 * If success, a node is locked. A user is responsible to unlock it when a
1050 * VA is no longer needed to be accessed.
1051 *
1052 * Returns NULL if nothing found.
1053 */
1054static struct vmap_node *
1055find_vmap_area_exceed_addr_lock(unsigned long addr, struct vmap_area **va)
1056{
1057 unsigned long va_start_lowest;
1058 struct vmap_node *vn;
1059 int i;
1060
1061repeat:
1062 for (i = 0, va_start_lowest = 0; i < nr_vmap_nodes; i++) {
1063 vn = &vmap_nodes[i];
1064
1065 spin_lock(&vn->busy.lock);
1066 *va = __find_vmap_area_exceed_addr(addr, &vn->busy.root);
1067
1068 if (*va)
1069 if (!va_start_lowest || (*va)->va_start < va_start_lowest)
1070 va_start_lowest = (*va)->va_start;
1071 spin_unlock(&vn->busy.lock);
1072 }
1073
1074 /*
1075 * Check if found VA exists, it might have gone away. In this case we
1076 * repeat the search because a VA has been removed concurrently and we
1077 * need to proceed to the next one, which is a rare case.
1078 */
1079 if (va_start_lowest) {
1080 vn = addr_to_node(va_start_lowest);
1081
1082 spin_lock(&vn->busy.lock);
1083 *va = __find_vmap_area(va_start_lowest, &vn->busy.root);
1084
1085 if (*va)
1086 return vn;
1087
1088 spin_unlock(&vn->busy.lock);
1089 goto repeat;
1090 }
1091
1092 return NULL;
1093}
1094
1095/*
1096 * This function returns back addresses of parent node
1097 * and its left or right link for further processing.
1098 *
1099 * Otherwise NULL is returned. In that case all further
1100 * steps regarding inserting of conflicting overlap range
1101 * have to be declined and actually considered as a bug.
1102 */
1103static __always_inline struct rb_node **
1104find_va_links(struct vmap_area *va,
1105 struct rb_root *root, struct rb_node *from,
1106 struct rb_node **parent)
1107{
1108 struct vmap_area *tmp_va;
1109 struct rb_node **link;
1110
1111 if (root) {
1112 link = &root->rb_node;
1113 if (unlikely(!*link)) {
1114 *parent = NULL;
1115 return link;
1116 }
1117 } else {
1118 link = &from;
1119 }
1120
1121 /*
1122 * Go to the bottom of the tree. When we hit the last point
1123 * we end up with parent rb_node and correct direction, i name
1124 * it link, where the new va->rb_node will be attached to.
1125 */
1126 do {
1127 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
1128
1129 /*
1130 * During the traversal we also do some sanity check.
1131 * Trigger the BUG() if there are sides(left/right)
1132 * or full overlaps.
1133 */
1134 if (va->va_end <= tmp_va->va_start)
1135 link = &(*link)->rb_left;
1136 else if (va->va_start >= tmp_va->va_end)
1137 link = &(*link)->rb_right;
1138 else {
1139 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
1140 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
1141
1142 return NULL;
1143 }
1144 } while (*link);
1145
1146 *parent = &tmp_va->rb_node;
1147 return link;
1148}
1149
1150static __always_inline struct list_head *
1151get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
1152{
1153 struct list_head *list;
1154
1155 if (unlikely(!parent))
1156 /*
1157 * The red-black tree where we try to find VA neighbors
1158 * before merging or inserting is empty, i.e. it means
1159 * there is no free vmap space. Normally it does not
1160 * happen but we handle this case anyway.
1161 */
1162 return NULL;
1163
1164 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
1165 return (&parent->rb_right == link ? list->next : list);
1166}
1167
1168static __always_inline void
1169__link_va(struct vmap_area *va, struct rb_root *root,
1170 struct rb_node *parent, struct rb_node **link,
1171 struct list_head *head, bool augment)
1172{
1173 /*
1174 * VA is still not in the list, but we can
1175 * identify its future previous list_head node.
1176 */
1177 if (likely(parent)) {
1178 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
1179 if (&parent->rb_right != link)
1180 head = head->prev;
1181 }
1182
1183 /* Insert to the rb-tree */
1184 rb_link_node(&va->rb_node, parent, link);
1185 if (augment) {
1186 /*
1187 * Some explanation here. Just perform simple insertion
1188 * to the tree. We do not set va->subtree_max_size to
1189 * its current size before calling rb_insert_augmented().
1190 * It is because we populate the tree from the bottom
1191 * to parent levels when the node _is_ in the tree.
1192 *
1193 * Therefore we set subtree_max_size to zero after insertion,
1194 * to let __augment_tree_propagate_from() puts everything to
1195 * the correct order later on.
1196 */
1197 rb_insert_augmented(&va->rb_node,
1198 root, &free_vmap_area_rb_augment_cb);
1199 va->subtree_max_size = 0;
1200 } else {
1201 rb_insert_color(&va->rb_node, root);
1202 }
1203
1204 /* Address-sort this list */
1205 list_add(&va->list, head);
1206}
1207
1208static __always_inline void
1209link_va(struct vmap_area *va, struct rb_root *root,
1210 struct rb_node *parent, struct rb_node **link,
1211 struct list_head *head)
1212{
1213 __link_va(va, root, parent, link, head, false);
1214}
1215
1216static __always_inline void
1217link_va_augment(struct vmap_area *va, struct rb_root *root,
1218 struct rb_node *parent, struct rb_node **link,
1219 struct list_head *head)
1220{
1221 __link_va(va, root, parent, link, head, true);
1222}
1223
1224static __always_inline void
1225__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
1226{
1227 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
1228 return;
1229
1230 if (augment)
1231 rb_erase_augmented(&va->rb_node,
1232 root, &free_vmap_area_rb_augment_cb);
1233 else
1234 rb_erase(&va->rb_node, root);
1235
1236 list_del_init(&va->list);
1237 RB_CLEAR_NODE(&va->rb_node);
1238}
1239
1240static __always_inline void
1241unlink_va(struct vmap_area *va, struct rb_root *root)
1242{
1243 __unlink_va(va, root, false);
1244}
1245
1246static __always_inline void
1247unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1248{
1249 __unlink_va(va, root, true);
1250}
1251
1252#if DEBUG_AUGMENT_PROPAGATE_CHECK
1253/*
1254 * Gets called when remove the node and rotate.
1255 */
1256static __always_inline unsigned long
1257compute_subtree_max_size(struct vmap_area *va)
1258{
1259 return max3(va_size(va),
1260 get_subtree_max_size(va->rb_node.rb_left),
1261 get_subtree_max_size(va->rb_node.rb_right));
1262}
1263
1264static void
1265augment_tree_propagate_check(void)
1266{
1267 struct vmap_area *va;
1268 unsigned long computed_size;
1269
1270 list_for_each_entry(va, &free_vmap_area_list, list) {
1271 computed_size = compute_subtree_max_size(va);
1272 if (computed_size != va->subtree_max_size)
1273 pr_emerg("tree is corrupted: %lu, %lu\n",
1274 va_size(va), va->subtree_max_size);
1275 }
1276}
1277#endif
1278
1279/*
1280 * This function populates subtree_max_size from bottom to upper
1281 * levels starting from VA point. The propagation must be done
1282 * when VA size is modified by changing its va_start/va_end. Or
1283 * in case of newly inserting of VA to the tree.
1284 *
1285 * It means that __augment_tree_propagate_from() must be called:
1286 * - After VA has been inserted to the tree(free path);
1287 * - After VA has been shrunk(allocation path);
1288 * - After VA has been increased(merging path).
1289 *
1290 * Please note that, it does not mean that upper parent nodes
1291 * and their subtree_max_size are recalculated all the time up
1292 * to the root node.
1293 *
1294 * 4--8
1295 * /\
1296 * / \
1297 * / \
1298 * 2--2 8--8
1299 *
1300 * For example if we modify the node 4, shrinking it to 2, then
1301 * no any modification is required. If we shrink the node 2 to 1
1302 * its subtree_max_size is updated only, and set to 1. If we shrink
1303 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1304 * node becomes 4--6.
1305 */
1306static __always_inline void
1307augment_tree_propagate_from(struct vmap_area *va)
1308{
1309 /*
1310 * Populate the tree from bottom towards the root until
1311 * the calculated maximum available size of checked node
1312 * is equal to its current one.
1313 */
1314 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1315
1316#if DEBUG_AUGMENT_PROPAGATE_CHECK
1317 augment_tree_propagate_check();
1318#endif
1319}
1320
1321static void
1322insert_vmap_area(struct vmap_area *va,
1323 struct rb_root *root, struct list_head *head)
1324{
1325 struct rb_node **link;
1326 struct rb_node *parent;
1327
1328 link = find_va_links(va, root, NULL, &parent);
1329 if (link)
1330 link_va(va, root, parent, link, head);
1331}
1332
1333static void
1334insert_vmap_area_augment(struct vmap_area *va,
1335 struct rb_node *from, struct rb_root *root,
1336 struct list_head *head)
1337{
1338 struct rb_node **link;
1339 struct rb_node *parent;
1340
1341 if (from)
1342 link = find_va_links(va, NULL, from, &parent);
1343 else
1344 link = find_va_links(va, root, NULL, &parent);
1345
1346 if (link) {
1347 link_va_augment(va, root, parent, link, head);
1348 augment_tree_propagate_from(va);
1349 }
1350}
1351
1352/*
1353 * Merge de-allocated chunk of VA memory with previous
1354 * and next free blocks. If coalesce is not done a new
1355 * free area is inserted. If VA has been merged, it is
1356 * freed.
1357 *
1358 * Please note, it can return NULL in case of overlap
1359 * ranges, followed by WARN() report. Despite it is a
1360 * buggy behaviour, a system can be alive and keep
1361 * ongoing.
1362 */
1363static __always_inline struct vmap_area *
1364__merge_or_add_vmap_area(struct vmap_area *va,
1365 struct rb_root *root, struct list_head *head, bool augment)
1366{
1367 struct vmap_area *sibling;
1368 struct list_head *next;
1369 struct rb_node **link;
1370 struct rb_node *parent;
1371 bool merged = false;
1372
1373 /*
1374 * Find a place in the tree where VA potentially will be
1375 * inserted, unless it is merged with its sibling/siblings.
1376 */
1377 link = find_va_links(va, root, NULL, &parent);
1378 if (!link)
1379 return NULL;
1380
1381 /*
1382 * Get next node of VA to check if merging can be done.
1383 */
1384 next = get_va_next_sibling(parent, link);
1385 if (unlikely(next == NULL))
1386 goto insert;
1387
1388 /*
1389 * start end
1390 * | |
1391 * |<------VA------>|<-----Next----->|
1392 * | |
1393 * start end
1394 */
1395 if (next != head) {
1396 sibling = list_entry(next, struct vmap_area, list);
1397 if (sibling->va_start == va->va_end) {
1398 sibling->va_start = va->va_start;
1399
1400 /* Free vmap_area object. */
1401 kmem_cache_free(vmap_area_cachep, va);
1402
1403 /* Point to the new merged area. */
1404 va = sibling;
1405 merged = true;
1406 }
1407 }
1408
1409 /*
1410 * start end
1411 * | |
1412 * |<-----Prev----->|<------VA------>|
1413 * | |
1414 * start end
1415 */
1416 if (next->prev != head) {
1417 sibling = list_entry(next->prev, struct vmap_area, list);
1418 if (sibling->va_end == va->va_start) {
1419 /*
1420 * If both neighbors are coalesced, it is important
1421 * to unlink the "next" node first, followed by merging
1422 * with "previous" one. Otherwise the tree might not be
1423 * fully populated if a sibling's augmented value is
1424 * "normalized" because of rotation operations.
1425 */
1426 if (merged)
1427 __unlink_va(va, root, augment);
1428
1429 sibling->va_end = va->va_end;
1430
1431 /* Free vmap_area object. */
1432 kmem_cache_free(vmap_area_cachep, va);
1433
1434 /* Point to the new merged area. */
1435 va = sibling;
1436 merged = true;
1437 }
1438 }
1439
1440insert:
1441 if (!merged)
1442 __link_va(va, root, parent, link, head, augment);
1443
1444 return va;
1445}
1446
1447static __always_inline struct vmap_area *
1448merge_or_add_vmap_area(struct vmap_area *va,
1449 struct rb_root *root, struct list_head *head)
1450{
1451 return __merge_or_add_vmap_area(va, root, head, false);
1452}
1453
1454static __always_inline struct vmap_area *
1455merge_or_add_vmap_area_augment(struct vmap_area *va,
1456 struct rb_root *root, struct list_head *head)
1457{
1458 va = __merge_or_add_vmap_area(va, root, head, true);
1459 if (va)
1460 augment_tree_propagate_from(va);
1461
1462 return va;
1463}
1464
1465static __always_inline bool
1466is_within_this_va(struct vmap_area *va, unsigned long size,
1467 unsigned long align, unsigned long vstart)
1468{
1469 unsigned long nva_start_addr;
1470
1471 if (va->va_start > vstart)
1472 nva_start_addr = ALIGN(va->va_start, align);
1473 else
1474 nva_start_addr = ALIGN(vstart, align);
1475
1476 /* Can be overflowed due to big size or alignment. */
1477 if (nva_start_addr + size < nva_start_addr ||
1478 nva_start_addr < vstart)
1479 return false;
1480
1481 return (nva_start_addr + size <= va->va_end);
1482}
1483
1484/*
1485 * Find the first free block(lowest start address) in the tree,
1486 * that will accomplish the request corresponding to passing
1487 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1488 * a search length is adjusted to account for worst case alignment
1489 * overhead.
1490 */
1491static __always_inline struct vmap_area *
1492find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1493 unsigned long align, unsigned long vstart, bool adjust_search_size)
1494{
1495 struct vmap_area *va;
1496 struct rb_node *node;
1497 unsigned long length;
1498
1499 /* Start from the root. */
1500 node = root->rb_node;
1501
1502 /* Adjust the search size for alignment overhead. */
1503 length = adjust_search_size ? size + align - 1 : size;
1504
1505 while (node) {
1506 va = rb_entry(node, struct vmap_area, rb_node);
1507
1508 if (get_subtree_max_size(node->rb_left) >= length &&
1509 vstart < va->va_start) {
1510 node = node->rb_left;
1511 } else {
1512 if (is_within_this_va(va, size, align, vstart))
1513 return va;
1514
1515 /*
1516 * Does not make sense to go deeper towards the right
1517 * sub-tree if it does not have a free block that is
1518 * equal or bigger to the requested search length.
1519 */
1520 if (get_subtree_max_size(node->rb_right) >= length) {
1521 node = node->rb_right;
1522 continue;
1523 }
1524
1525 /*
1526 * OK. We roll back and find the first right sub-tree,
1527 * that will satisfy the search criteria. It can happen
1528 * due to "vstart" restriction or an alignment overhead
1529 * that is bigger then PAGE_SIZE.
1530 */
1531 while ((node = rb_parent(node))) {
1532 va = rb_entry(node, struct vmap_area, rb_node);
1533 if (is_within_this_va(va, size, align, vstart))
1534 return va;
1535
1536 if (get_subtree_max_size(node->rb_right) >= length &&
1537 vstart <= va->va_start) {
1538 /*
1539 * Shift the vstart forward. Please note, we update it with
1540 * parent's start address adding "1" because we do not want
1541 * to enter same sub-tree after it has already been checked
1542 * and no suitable free block found there.
1543 */
1544 vstart = va->va_start + 1;
1545 node = node->rb_right;
1546 break;
1547 }
1548 }
1549 }
1550 }
1551
1552 return NULL;
1553}
1554
1555#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1556#include <linux/random.h>
1557
1558static struct vmap_area *
1559find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1560 unsigned long align, unsigned long vstart)
1561{
1562 struct vmap_area *va;
1563
1564 list_for_each_entry(va, head, list) {
1565 if (!is_within_this_va(va, size, align, vstart))
1566 continue;
1567
1568 return va;
1569 }
1570
1571 return NULL;
1572}
1573
1574static void
1575find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1576 unsigned long size, unsigned long align)
1577{
1578 struct vmap_area *va_1, *va_2;
1579 unsigned long vstart;
1580 unsigned int rnd;
1581
1582 get_random_bytes(&rnd, sizeof(rnd));
1583 vstart = VMALLOC_START + rnd;
1584
1585 va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1586 va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1587
1588 if (va_1 != va_2)
1589 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1590 va_1, va_2, vstart);
1591}
1592#endif
1593
1594enum fit_type {
1595 NOTHING_FIT = 0,
1596 FL_FIT_TYPE = 1, /* full fit */
1597 LE_FIT_TYPE = 2, /* left edge fit */
1598 RE_FIT_TYPE = 3, /* right edge fit */
1599 NE_FIT_TYPE = 4 /* no edge fit */
1600};
1601
1602static __always_inline enum fit_type
1603classify_va_fit_type(struct vmap_area *va,
1604 unsigned long nva_start_addr, unsigned long size)
1605{
1606 enum fit_type type;
1607
1608 /* Check if it is within VA. */
1609 if (nva_start_addr < va->va_start ||
1610 nva_start_addr + size > va->va_end)
1611 return NOTHING_FIT;
1612
1613 /* Now classify. */
1614 if (va->va_start == nva_start_addr) {
1615 if (va->va_end == nva_start_addr + size)
1616 type = FL_FIT_TYPE;
1617 else
1618 type = LE_FIT_TYPE;
1619 } else if (va->va_end == nva_start_addr + size) {
1620 type = RE_FIT_TYPE;
1621 } else {
1622 type = NE_FIT_TYPE;
1623 }
1624
1625 return type;
1626}
1627
1628static __always_inline int
1629va_clip(struct rb_root *root, struct list_head *head,
1630 struct vmap_area *va, unsigned long nva_start_addr,
1631 unsigned long size)
1632{
1633 struct vmap_area *lva = NULL;
1634 enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1635
1636 if (type == FL_FIT_TYPE) {
1637 /*
1638 * No need to split VA, it fully fits.
1639 *
1640 * | |
1641 * V NVA V
1642 * |---------------|
1643 */
1644 unlink_va_augment(va, root);
1645 kmem_cache_free(vmap_area_cachep, va);
1646 } else if (type == LE_FIT_TYPE) {
1647 /*
1648 * Split left edge of fit VA.
1649 *
1650 * | |
1651 * V NVA V R
1652 * |-------|-------|
1653 */
1654 va->va_start += size;
1655 } else if (type == RE_FIT_TYPE) {
1656 /*
1657 * Split right edge of fit VA.
1658 *
1659 * | |
1660 * L V NVA V
1661 * |-------|-------|
1662 */
1663 va->va_end = nva_start_addr;
1664 } else if (type == NE_FIT_TYPE) {
1665 /*
1666 * Split no edge of fit VA.
1667 *
1668 * | |
1669 * L V NVA V R
1670 * |---|-------|---|
1671 */
1672 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1673 if (unlikely(!lva)) {
1674 /*
1675 * For percpu allocator we do not do any pre-allocation
1676 * and leave it as it is. The reason is it most likely
1677 * never ends up with NE_FIT_TYPE splitting. In case of
1678 * percpu allocations offsets and sizes are aligned to
1679 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1680 * are its main fitting cases.
1681 *
1682 * There are a few exceptions though, as an example it is
1683 * a first allocation (early boot up) when we have "one"
1684 * big free space that has to be split.
1685 *
1686 * Also we can hit this path in case of regular "vmap"
1687 * allocations, if "this" current CPU was not preloaded.
1688 * See the comment in alloc_vmap_area() why. If so, then
1689 * GFP_NOWAIT is used instead to get an extra object for
1690 * split purpose. That is rare and most time does not
1691 * occur.
1692 *
1693 * What happens if an allocation gets failed. Basically,
1694 * an "overflow" path is triggered to purge lazily freed
1695 * areas to free some memory, then, the "retry" path is
1696 * triggered to repeat one more time. See more details
1697 * in alloc_vmap_area() function.
1698 */
1699 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1700 if (!lva)
1701 return -1;
1702 }
1703
1704 /*
1705 * Build the remainder.
1706 */
1707 lva->va_start = va->va_start;
1708 lva->va_end = nva_start_addr;
1709
1710 /*
1711 * Shrink this VA to remaining size.
1712 */
1713 va->va_start = nva_start_addr + size;
1714 } else {
1715 return -1;
1716 }
1717
1718 if (type != FL_FIT_TYPE) {
1719 augment_tree_propagate_from(va);
1720
1721 if (lva) /* type == NE_FIT_TYPE */
1722 insert_vmap_area_augment(lva, &va->rb_node, root, head);
1723 }
1724
1725 return 0;
1726}
1727
1728static unsigned long
1729va_alloc(struct vmap_area *va,
1730 struct rb_root *root, struct list_head *head,
1731 unsigned long size, unsigned long align,
1732 unsigned long vstart, unsigned long vend)
1733{
1734 unsigned long nva_start_addr;
1735 int ret;
1736
1737 if (va->va_start > vstart)
1738 nva_start_addr = ALIGN(va->va_start, align);
1739 else
1740 nva_start_addr = ALIGN(vstart, align);
1741
1742 /* Check the "vend" restriction. */
1743 if (nva_start_addr + size > vend)
1744 return vend;
1745
1746 /* Update the free vmap_area. */
1747 ret = va_clip(root, head, va, nva_start_addr, size);
1748 if (WARN_ON_ONCE(ret))
1749 return vend;
1750
1751 return nva_start_addr;
1752}
1753
1754/*
1755 * Returns a start address of the newly allocated area, if success.
1756 * Otherwise a vend is returned that indicates failure.
1757 */
1758static __always_inline unsigned long
1759__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1760 unsigned long size, unsigned long align,
1761 unsigned long vstart, unsigned long vend)
1762{
1763 bool adjust_search_size = true;
1764 unsigned long nva_start_addr;
1765 struct vmap_area *va;
1766
1767 /*
1768 * Do not adjust when:
1769 * a) align <= PAGE_SIZE, because it does not make any sense.
1770 * All blocks(their start addresses) are at least PAGE_SIZE
1771 * aligned anyway;
1772 * b) a short range where a requested size corresponds to exactly
1773 * specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1774 * With adjusted search length an allocation would not succeed.
1775 */
1776 if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1777 adjust_search_size = false;
1778
1779 va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1780 if (unlikely(!va))
1781 return vend;
1782
1783 nva_start_addr = va_alloc(va, root, head, size, align, vstart, vend);
1784 if (nva_start_addr == vend)
1785 return vend;
1786
1787#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1788 find_vmap_lowest_match_check(root, head, size, align);
1789#endif
1790
1791 return nva_start_addr;
1792}
1793
1794/*
1795 * Free a region of KVA allocated by alloc_vmap_area
1796 */
1797static void free_vmap_area(struct vmap_area *va)
1798{
1799 struct vmap_node *vn = addr_to_node(va->va_start);
1800
1801 /*
1802 * Remove from the busy tree/list.
1803 */
1804 spin_lock(&vn->busy.lock);
1805 unlink_va(va, &vn->busy.root);
1806 spin_unlock(&vn->busy.lock);
1807
1808 /*
1809 * Insert/Merge it back to the free tree/list.
1810 */
1811 spin_lock(&free_vmap_area_lock);
1812 merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1813 spin_unlock(&free_vmap_area_lock);
1814}
1815
1816static inline void
1817preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1818{
1819 struct vmap_area *va = NULL, *tmp;
1820
1821 /*
1822 * Preload this CPU with one extra vmap_area object. It is used
1823 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1824 * a CPU that does an allocation is preloaded.
1825 *
1826 * We do it in non-atomic context, thus it allows us to use more
1827 * permissive allocation masks to be more stable under low memory
1828 * condition and high memory pressure.
1829 */
1830 if (!this_cpu_read(ne_fit_preload_node))
1831 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1832
1833 spin_lock(lock);
1834
1835 tmp = NULL;
1836 if (va && !__this_cpu_try_cmpxchg(ne_fit_preload_node, &tmp, va))
1837 kmem_cache_free(vmap_area_cachep, va);
1838}
1839
1840static struct vmap_pool *
1841size_to_va_pool(struct vmap_node *vn, unsigned long size)
1842{
1843 unsigned int idx = (size - 1) / PAGE_SIZE;
1844
1845 if (idx < MAX_VA_SIZE_PAGES)
1846 return &vn->pool[idx];
1847
1848 return NULL;
1849}
1850
1851static bool
1852node_pool_add_va(struct vmap_node *n, struct vmap_area *va)
1853{
1854 struct vmap_pool *vp;
1855
1856 vp = size_to_va_pool(n, va_size(va));
1857 if (!vp)
1858 return false;
1859
1860 spin_lock(&n->pool_lock);
1861 list_add(&va->list, &vp->head);
1862 WRITE_ONCE(vp->len, vp->len + 1);
1863 spin_unlock(&n->pool_lock);
1864
1865 return true;
1866}
1867
1868static struct vmap_area *
1869node_pool_del_va(struct vmap_node *vn, unsigned long size,
1870 unsigned long align, unsigned long vstart,
1871 unsigned long vend)
1872{
1873 struct vmap_area *va = NULL;
1874 struct vmap_pool *vp;
1875 int err = 0;
1876
1877 vp = size_to_va_pool(vn, size);
1878 if (!vp || list_empty(&vp->head))
1879 return NULL;
1880
1881 spin_lock(&vn->pool_lock);
1882 if (!list_empty(&vp->head)) {
1883 va = list_first_entry(&vp->head, struct vmap_area, list);
1884
1885 if (IS_ALIGNED(va->va_start, align)) {
1886 /*
1887 * Do some sanity check and emit a warning
1888 * if one of below checks detects an error.
1889 */
1890 err |= (va_size(va) != size);
1891 err |= (va->va_start < vstart);
1892 err |= (va->va_end > vend);
1893
1894 if (!WARN_ON_ONCE(err)) {
1895 list_del_init(&va->list);
1896 WRITE_ONCE(vp->len, vp->len - 1);
1897 } else {
1898 va = NULL;
1899 }
1900 } else {
1901 list_move_tail(&va->list, &vp->head);
1902 va = NULL;
1903 }
1904 }
1905 spin_unlock(&vn->pool_lock);
1906
1907 return va;
1908}
1909
1910static struct vmap_area *
1911node_alloc(unsigned long size, unsigned long align,
1912 unsigned long vstart, unsigned long vend,
1913 unsigned long *addr, unsigned int *vn_id)
1914{
1915 struct vmap_area *va;
1916
1917 *vn_id = 0;
1918 *addr = vend;
1919
1920 /*
1921 * Fallback to a global heap if not vmalloc or there
1922 * is only one node.
1923 */
1924 if (vstart != VMALLOC_START || vend != VMALLOC_END ||
1925 nr_vmap_nodes == 1)
1926 return NULL;
1927
1928 *vn_id = raw_smp_processor_id() % nr_vmap_nodes;
1929 va = node_pool_del_va(id_to_node(*vn_id), size, align, vstart, vend);
1930 *vn_id = encode_vn_id(*vn_id);
1931
1932 if (va)
1933 *addr = va->va_start;
1934
1935 return va;
1936}
1937
1938static inline void setup_vmalloc_vm(struct vm_struct *vm,
1939 struct vmap_area *va, unsigned long flags, const void *caller)
1940{
1941 vm->flags = flags;
1942 vm->addr = (void *)va->va_start;
1943 vm->size = va_size(va);
1944 vm->caller = caller;
1945 va->vm = vm;
1946}
1947
1948/*
1949 * Allocate a region of KVA of the specified size and alignment, within the
1950 * vstart and vend. If vm is passed in, the two will also be bound.
1951 */
1952static struct vmap_area *alloc_vmap_area(unsigned long size,
1953 unsigned long align,
1954 unsigned long vstart, unsigned long vend,
1955 int node, gfp_t gfp_mask,
1956 unsigned long va_flags, struct vm_struct *vm)
1957{
1958 struct vmap_node *vn;
1959 struct vmap_area *va;
1960 unsigned long freed;
1961 unsigned long addr;
1962 unsigned int vn_id;
1963 int purged = 0;
1964 int ret;
1965
1966 if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1967 return ERR_PTR(-EINVAL);
1968
1969 if (unlikely(!vmap_initialized))
1970 return ERR_PTR(-EBUSY);
1971
1972 might_sleep();
1973
1974 /*
1975 * If a VA is obtained from a global heap(if it fails here)
1976 * it is anyway marked with this "vn_id" so it is returned
1977 * to this pool's node later. Such way gives a possibility
1978 * to populate pools based on users demand.
1979 *
1980 * On success a ready to go VA is returned.
1981 */
1982 va = node_alloc(size, align, vstart, vend, &addr, &vn_id);
1983 if (!va) {
1984 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1985
1986 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1987 if (unlikely(!va))
1988 return ERR_PTR(-ENOMEM);
1989
1990 /*
1991 * Only scan the relevant parts containing pointers to other objects
1992 * to avoid false negatives.
1993 */
1994 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1995 }
1996
1997retry:
1998 if (addr == vend) {
1999 preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
2000 addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
2001 size, align, vstart, vend);
2002 spin_unlock(&free_vmap_area_lock);
2003 }
2004
2005 trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
2006
2007 /*
2008 * If an allocation fails, the "vend" address is
2009 * returned. Therefore trigger the overflow path.
2010 */
2011 if (unlikely(addr == vend))
2012 goto overflow;
2013
2014 va->va_start = addr;
2015 va->va_end = addr + size;
2016 va->vm = NULL;
2017 va->flags = (va_flags | vn_id);
2018
2019 if (vm) {
2020 vm->addr = (void *)va->va_start;
2021 vm->size = va_size(va);
2022 va->vm = vm;
2023 }
2024
2025 vn = addr_to_node(va->va_start);
2026
2027 spin_lock(&vn->busy.lock);
2028 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
2029 spin_unlock(&vn->busy.lock);
2030
2031 BUG_ON(!IS_ALIGNED(va->va_start, align));
2032 BUG_ON(va->va_start < vstart);
2033 BUG_ON(va->va_end > vend);
2034
2035 ret = kasan_populate_vmalloc(addr, size);
2036 if (ret) {
2037 free_vmap_area(va);
2038 return ERR_PTR(ret);
2039 }
2040
2041 return va;
2042
2043overflow:
2044 if (!purged) {
2045 reclaim_and_purge_vmap_areas();
2046 purged = 1;
2047 goto retry;
2048 }
2049
2050 freed = 0;
2051 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
2052
2053 if (freed > 0) {
2054 purged = 0;
2055 goto retry;
2056 }
2057
2058 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
2059 pr_warn("vmalloc_node_range for size %lu failed: Address range restricted to %#lx - %#lx\n",
2060 size, vstart, vend);
2061
2062 kmem_cache_free(vmap_area_cachep, va);
2063 return ERR_PTR(-EBUSY);
2064}
2065
2066int register_vmap_purge_notifier(struct notifier_block *nb)
2067{
2068 return blocking_notifier_chain_register(&vmap_notify_list, nb);
2069}
2070EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
2071
2072int unregister_vmap_purge_notifier(struct notifier_block *nb)
2073{
2074 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
2075}
2076EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
2077
2078/*
2079 * lazy_max_pages is the maximum amount of virtual address space we gather up
2080 * before attempting to purge with a TLB flush.
2081 *
2082 * There is a tradeoff here: a larger number will cover more kernel page tables
2083 * and take slightly longer to purge, but it will linearly reduce the number of
2084 * global TLB flushes that must be performed. It would seem natural to scale
2085 * this number up linearly with the number of CPUs (because vmapping activity
2086 * could also scale linearly with the number of CPUs), however it is likely
2087 * that in practice, workloads might be constrained in other ways that mean
2088 * vmap activity will not scale linearly with CPUs. Also, I want to be
2089 * conservative and not introduce a big latency on huge systems, so go with
2090 * a less aggressive log scale. It will still be an improvement over the old
2091 * code, and it will be simple to change the scale factor if we find that it
2092 * becomes a problem on bigger systems.
2093 */
2094static unsigned long lazy_max_pages(void)
2095{
2096 unsigned int log;
2097
2098 log = fls(num_online_cpus());
2099
2100 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
2101}
2102
2103static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
2104
2105/*
2106 * Serialize vmap purging. There is no actual critical section protected
2107 * by this lock, but we want to avoid concurrent calls for performance
2108 * reasons and to make the pcpu_get_vm_areas more deterministic.
2109 */
2110static DEFINE_MUTEX(vmap_purge_lock);
2111
2112/* for per-CPU blocks */
2113static void purge_fragmented_blocks_allcpus(void);
2114static cpumask_t purge_nodes;
2115
2116static void
2117reclaim_list_global(struct list_head *head)
2118{
2119 struct vmap_area *va, *n;
2120
2121 if (list_empty(head))
2122 return;
2123
2124 spin_lock(&free_vmap_area_lock);
2125 list_for_each_entry_safe(va, n, head, list)
2126 merge_or_add_vmap_area_augment(va,
2127 &free_vmap_area_root, &free_vmap_area_list);
2128 spin_unlock(&free_vmap_area_lock);
2129}
2130
2131static void
2132decay_va_pool_node(struct vmap_node *vn, bool full_decay)
2133{
2134 LIST_HEAD(decay_list);
2135 struct rb_root decay_root = RB_ROOT;
2136 struct vmap_area *va, *nva;
2137 unsigned long n_decay;
2138 int i;
2139
2140 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
2141 LIST_HEAD(tmp_list);
2142
2143 if (list_empty(&vn->pool[i].head))
2144 continue;
2145
2146 /* Detach the pool, so no-one can access it. */
2147 spin_lock(&vn->pool_lock);
2148 list_replace_init(&vn->pool[i].head, &tmp_list);
2149 spin_unlock(&vn->pool_lock);
2150
2151 if (full_decay)
2152 WRITE_ONCE(vn->pool[i].len, 0);
2153
2154 /* Decay a pool by ~25% out of left objects. */
2155 n_decay = vn->pool[i].len >> 2;
2156
2157 list_for_each_entry_safe(va, nva, &tmp_list, list) {
2158 list_del_init(&va->list);
2159 merge_or_add_vmap_area(va, &decay_root, &decay_list);
2160
2161 if (!full_decay) {
2162 WRITE_ONCE(vn->pool[i].len, vn->pool[i].len - 1);
2163
2164 if (!--n_decay)
2165 break;
2166 }
2167 }
2168
2169 /*
2170 * Attach the pool back if it has been partly decayed.
2171 * Please note, it is supposed that nobody(other contexts)
2172 * can populate the pool therefore a simple list replace
2173 * operation takes place here.
2174 */
2175 if (!full_decay && !list_empty(&tmp_list)) {
2176 spin_lock(&vn->pool_lock);
2177 list_replace_init(&tmp_list, &vn->pool[i].head);
2178 spin_unlock(&vn->pool_lock);
2179 }
2180 }
2181
2182 reclaim_list_global(&decay_list);
2183}
2184
2185static void
2186kasan_release_vmalloc_node(struct vmap_node *vn)
2187{
2188 struct vmap_area *va;
2189 unsigned long start, end;
2190
2191 start = list_first_entry(&vn->purge_list, struct vmap_area, list)->va_start;
2192 end = list_last_entry(&vn->purge_list, struct vmap_area, list)->va_end;
2193
2194 list_for_each_entry(va, &vn->purge_list, list) {
2195 if (is_vmalloc_or_module_addr((void *) va->va_start))
2196 kasan_release_vmalloc(va->va_start, va->va_end,
2197 va->va_start, va->va_end,
2198 KASAN_VMALLOC_PAGE_RANGE);
2199 }
2200
2201 kasan_release_vmalloc(start, end, start, end, KASAN_VMALLOC_TLB_FLUSH);
2202}
2203
2204static void purge_vmap_node(struct work_struct *work)
2205{
2206 struct vmap_node *vn = container_of(work,
2207 struct vmap_node, purge_work);
2208 unsigned long nr_purged_pages = 0;
2209 struct vmap_area *va, *n_va;
2210 LIST_HEAD(local_list);
2211
2212 if (IS_ENABLED(CONFIG_KASAN_VMALLOC))
2213 kasan_release_vmalloc_node(vn);
2214
2215 vn->nr_purged = 0;
2216
2217 list_for_each_entry_safe(va, n_va, &vn->purge_list, list) {
2218 unsigned long nr = va_size(va) >> PAGE_SHIFT;
2219 unsigned int vn_id = decode_vn_id(va->flags);
2220
2221 list_del_init(&va->list);
2222
2223 nr_purged_pages += nr;
2224 vn->nr_purged++;
2225
2226 if (is_vn_id_valid(vn_id) && !vn->skip_populate)
2227 if (node_pool_add_va(vn, va))
2228 continue;
2229
2230 /* Go back to global. */
2231 list_add(&va->list, &local_list);
2232 }
2233
2234 atomic_long_sub(nr_purged_pages, &vmap_lazy_nr);
2235
2236 reclaim_list_global(&local_list);
2237}
2238
2239/*
2240 * Purges all lazily-freed vmap areas.
2241 */
2242static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
2243 bool full_pool_decay)
2244{
2245 unsigned long nr_purged_areas = 0;
2246 unsigned int nr_purge_helpers;
2247 unsigned int nr_purge_nodes;
2248 struct vmap_node *vn;
2249 int i;
2250
2251 lockdep_assert_held(&vmap_purge_lock);
2252
2253 /*
2254 * Use cpumask to mark which node has to be processed.
2255 */
2256 purge_nodes = CPU_MASK_NONE;
2257
2258 for (i = 0; i < nr_vmap_nodes; i++) {
2259 vn = &vmap_nodes[i];
2260
2261 INIT_LIST_HEAD(&vn->purge_list);
2262 vn->skip_populate = full_pool_decay;
2263 decay_va_pool_node(vn, full_pool_decay);
2264
2265 if (RB_EMPTY_ROOT(&vn->lazy.root))
2266 continue;
2267
2268 spin_lock(&vn->lazy.lock);
2269 WRITE_ONCE(vn->lazy.root.rb_node, NULL);
2270 list_replace_init(&vn->lazy.head, &vn->purge_list);
2271 spin_unlock(&vn->lazy.lock);
2272
2273 start = min(start, list_first_entry(&vn->purge_list,
2274 struct vmap_area, list)->va_start);
2275
2276 end = max(end, list_last_entry(&vn->purge_list,
2277 struct vmap_area, list)->va_end);
2278
2279 cpumask_set_cpu(i, &purge_nodes);
2280 }
2281
2282 nr_purge_nodes = cpumask_weight(&purge_nodes);
2283 if (nr_purge_nodes > 0) {
2284 flush_tlb_kernel_range(start, end);
2285
2286 /* One extra worker is per a lazy_max_pages() full set minus one. */
2287 nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
2288 nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
2289
2290 for_each_cpu(i, &purge_nodes) {
2291 vn = &vmap_nodes[i];
2292
2293 if (nr_purge_helpers > 0) {
2294 INIT_WORK(&vn->purge_work, purge_vmap_node);
2295
2296 if (cpumask_test_cpu(i, cpu_online_mask))
2297 schedule_work_on(i, &vn->purge_work);
2298 else
2299 schedule_work(&vn->purge_work);
2300
2301 nr_purge_helpers--;
2302 } else {
2303 vn->purge_work.func = NULL;
2304 purge_vmap_node(&vn->purge_work);
2305 nr_purged_areas += vn->nr_purged;
2306 }
2307 }
2308
2309 for_each_cpu(i, &purge_nodes) {
2310 vn = &vmap_nodes[i];
2311
2312 if (vn->purge_work.func) {
2313 flush_work(&vn->purge_work);
2314 nr_purged_areas += vn->nr_purged;
2315 }
2316 }
2317 }
2318
2319 trace_purge_vmap_area_lazy(start, end, nr_purged_areas);
2320 return nr_purged_areas > 0;
2321}
2322
2323/*
2324 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
2325 */
2326static void reclaim_and_purge_vmap_areas(void)
2327
2328{
2329 mutex_lock(&vmap_purge_lock);
2330 purge_fragmented_blocks_allcpus();
2331 __purge_vmap_area_lazy(ULONG_MAX, 0, true);
2332 mutex_unlock(&vmap_purge_lock);
2333}
2334
2335static void drain_vmap_area_work(struct work_struct *work)
2336{
2337 mutex_lock(&vmap_purge_lock);
2338 __purge_vmap_area_lazy(ULONG_MAX, 0, false);
2339 mutex_unlock(&vmap_purge_lock);
2340}
2341
2342/*
2343 * Free a vmap area, caller ensuring that the area has been unmapped,
2344 * unlinked and flush_cache_vunmap had been called for the correct
2345 * range previously.
2346 */
2347static void free_vmap_area_noflush(struct vmap_area *va)
2348{
2349 unsigned long nr_lazy_max = lazy_max_pages();
2350 unsigned long va_start = va->va_start;
2351 unsigned int vn_id = decode_vn_id(va->flags);
2352 struct vmap_node *vn;
2353 unsigned long nr_lazy;
2354
2355 if (WARN_ON_ONCE(!list_empty(&va->list)))
2356 return;
2357
2358 nr_lazy = atomic_long_add_return(va_size(va) >> PAGE_SHIFT,
2359 &vmap_lazy_nr);
2360
2361 /*
2362 * If it was request by a certain node we would like to
2363 * return it to that node, i.e. its pool for later reuse.
2364 */
2365 vn = is_vn_id_valid(vn_id) ?
2366 id_to_node(vn_id):addr_to_node(va->va_start);
2367
2368 spin_lock(&vn->lazy.lock);
2369 insert_vmap_area(va, &vn->lazy.root, &vn->lazy.head);
2370 spin_unlock(&vn->lazy.lock);
2371
2372 trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
2373
2374 /* After this point, we may free va at any time */
2375 if (unlikely(nr_lazy > nr_lazy_max))
2376 schedule_work(&drain_vmap_work);
2377}
2378
2379/*
2380 * Free and unmap a vmap area
2381 */
2382static void free_unmap_vmap_area(struct vmap_area *va)
2383{
2384 flush_cache_vunmap(va->va_start, va->va_end);
2385 vunmap_range_noflush(va->va_start, va->va_end);
2386 if (debug_pagealloc_enabled_static())
2387 flush_tlb_kernel_range(va->va_start, va->va_end);
2388
2389 free_vmap_area_noflush(va);
2390}
2391
2392struct vmap_area *find_vmap_area(unsigned long addr)
2393{
2394 struct vmap_node *vn;
2395 struct vmap_area *va;
2396 int i, j;
2397
2398 if (unlikely(!vmap_initialized))
2399 return NULL;
2400
2401 /*
2402 * An addr_to_node_id(addr) converts an address to a node index
2403 * where a VA is located. If VA spans several zones and passed
2404 * addr is not the same as va->va_start, what is not common, we
2405 * may need to scan extra nodes. See an example:
2406 *
2407 * <----va---->
2408 * -|-----|-----|-----|-----|-
2409 * 1 2 0 1
2410 *
2411 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
2412 * addr is within 2 or 0 nodes we should do extra work.
2413 */
2414 i = j = addr_to_node_id(addr);
2415 do {
2416 vn = &vmap_nodes[i];
2417
2418 spin_lock(&vn->busy.lock);
2419 va = __find_vmap_area(addr, &vn->busy.root);
2420 spin_unlock(&vn->busy.lock);
2421
2422 if (va)
2423 return va;
2424 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2425
2426 return NULL;
2427}
2428
2429static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
2430{
2431 struct vmap_node *vn;
2432 struct vmap_area *va;
2433 int i, j;
2434
2435 /*
2436 * Check the comment in the find_vmap_area() about the loop.
2437 */
2438 i = j = addr_to_node_id(addr);
2439 do {
2440 vn = &vmap_nodes[i];
2441
2442 spin_lock(&vn->busy.lock);
2443 va = __find_vmap_area(addr, &vn->busy.root);
2444 if (va)
2445 unlink_va(va, &vn->busy.root);
2446 spin_unlock(&vn->busy.lock);
2447
2448 if (va)
2449 return va;
2450 } while ((i = (i + 1) % nr_vmap_nodes) != j);
2451
2452 return NULL;
2453}
2454
2455/*** Per cpu kva allocator ***/
2456
2457/*
2458 * vmap space is limited especially on 32 bit architectures. Ensure there is
2459 * room for at least 16 percpu vmap blocks per CPU.
2460 */
2461/*
2462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
2463 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
2464 * instead (we just need a rough idea)
2465 */
2466#if BITS_PER_LONG == 32
2467#define VMALLOC_SPACE (128UL*1024*1024)
2468#else
2469#define VMALLOC_SPACE (128UL*1024*1024*1024)
2470#endif
2471
2472#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
2473#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
2474#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
2475#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
2476#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
2477#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
2478#define VMAP_BBMAP_BITS \
2479 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
2480 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
2481 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
2482
2483#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
2484
2485/*
2486 * Purge threshold to prevent overeager purging of fragmented blocks for
2487 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
2488 */
2489#define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4)
2490
2491#define VMAP_RAM 0x1 /* indicates vm_map_ram area*/
2492#define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/
2493#define VMAP_FLAGS_MASK 0x3
2494
2495struct vmap_block_queue {
2496 spinlock_t lock;
2497 struct list_head free;
2498
2499 /*
2500 * An xarray requires an extra memory dynamically to
2501 * be allocated. If it is an issue, we can use rb-tree
2502 * instead.
2503 */
2504 struct xarray vmap_blocks;
2505};
2506
2507struct vmap_block {
2508 spinlock_t lock;
2509 struct vmap_area *va;
2510 unsigned long free, dirty;
2511 DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
2512 unsigned long dirty_min, dirty_max; /*< dirty range */
2513 struct list_head free_list;
2514 struct rcu_head rcu_head;
2515 struct list_head purge;
2516 unsigned int cpu;
2517};
2518
2519/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
2520static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
2521
2522/*
2523 * In order to fast access to any "vmap_block" associated with a
2524 * specific address, we use a hash.
2525 *
2526 * A per-cpu vmap_block_queue is used in both ways, to serialize
2527 * an access to free block chains among CPUs(alloc path) and it
2528 * also acts as a vmap_block hash(alloc/free paths). It means we
2529 * overload it, since we already have the per-cpu array which is
2530 * used as a hash table. When used as a hash a 'cpu' passed to
2531 * per_cpu() is not actually a CPU but rather a hash index.
2532 *
2533 * A hash function is addr_to_vb_xa() which hashes any address
2534 * to a specific index(in a hash) it belongs to. This then uses a
2535 * per_cpu() macro to access an array with generated index.
2536 *
2537 * An example:
2538 *
2539 * CPU_1 CPU_2 CPU_0
2540 * | | |
2541 * V V V
2542 * 0 10 20 30 40 50 60
2543 * |------|------|------|------|------|------|...<vmap address space>
2544 * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2
2545 *
2546 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
2547 * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
2548 *
2549 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
2550 * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
2551 *
2552 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
2553 * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
2554 *
2555 * This technique almost always avoids lock contention on insert/remove,
2556 * however xarray spinlocks protect against any contention that remains.
2557 */
2558static struct xarray *
2559addr_to_vb_xa(unsigned long addr)
2560{
2561 int index = (addr / VMAP_BLOCK_SIZE) % nr_cpu_ids;
2562
2563 /*
2564 * Please note, nr_cpu_ids points on a highest set
2565 * possible bit, i.e. we never invoke cpumask_next()
2566 * if an index points on it which is nr_cpu_ids - 1.
2567 */
2568 if (!cpu_possible(index))
2569 index = cpumask_next(index, cpu_possible_mask);
2570
2571 return &per_cpu(vmap_block_queue, index).vmap_blocks;
2572}
2573
2574/*
2575 * We should probably have a fallback mechanism to allocate virtual memory
2576 * out of partially filled vmap blocks. However vmap block sizing should be
2577 * fairly reasonable according to the vmalloc size, so it shouldn't be a
2578 * big problem.
2579 */
2580
2581static unsigned long addr_to_vb_idx(unsigned long addr)
2582{
2583 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2584 addr /= VMAP_BLOCK_SIZE;
2585 return addr;
2586}
2587
2588static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2589{
2590 unsigned long addr;
2591
2592 addr = va_start + (pages_off << PAGE_SHIFT);
2593 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2594 return (void *)addr;
2595}
2596
2597/**
2598 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2599 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
2600 * @order: how many 2^order pages should be occupied in newly allocated block
2601 * @gfp_mask: flags for the page level allocator
2602 *
2603 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2604 */
2605static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2606{
2607 struct vmap_block_queue *vbq;
2608 struct vmap_block *vb;
2609 struct vmap_area *va;
2610 struct xarray *xa;
2611 unsigned long vb_idx;
2612 int node, err;
2613 void *vaddr;
2614
2615 node = numa_node_id();
2616
2617 vb = kmalloc_node(sizeof(struct vmap_block),
2618 gfp_mask & GFP_RECLAIM_MASK, node);
2619 if (unlikely(!vb))
2620 return ERR_PTR(-ENOMEM);
2621
2622 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2623 VMALLOC_START, VMALLOC_END,
2624 node, gfp_mask,
2625 VMAP_RAM|VMAP_BLOCK, NULL);
2626 if (IS_ERR(va)) {
2627 kfree(vb);
2628 return ERR_CAST(va);
2629 }
2630
2631 vaddr = vmap_block_vaddr(va->va_start, 0);
2632 spin_lock_init(&vb->lock);
2633 vb->va = va;
2634 /* At least something should be left free */
2635 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2636 bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2637 vb->free = VMAP_BBMAP_BITS - (1UL << order);
2638 vb->dirty = 0;
2639 vb->dirty_min = VMAP_BBMAP_BITS;
2640 vb->dirty_max = 0;
2641 bitmap_set(vb->used_map, 0, (1UL << order));
2642 INIT_LIST_HEAD(&vb->free_list);
2643 vb->cpu = raw_smp_processor_id();
2644
2645 xa = addr_to_vb_xa(va->va_start);
2646 vb_idx = addr_to_vb_idx(va->va_start);
2647 err = xa_insert(xa, vb_idx, vb, gfp_mask);
2648 if (err) {
2649 kfree(vb);
2650 free_vmap_area(va);
2651 return ERR_PTR(err);
2652 }
2653 /*
2654 * list_add_tail_rcu could happened in another core
2655 * rather than vb->cpu due to task migration, which
2656 * is safe as list_add_tail_rcu will ensure the list's
2657 * integrity together with list_for_each_rcu from read
2658 * side.
2659 */
2660 vbq = per_cpu_ptr(&vmap_block_queue, vb->cpu);
2661 spin_lock(&vbq->lock);
2662 list_add_tail_rcu(&vb->free_list, &vbq->free);
2663 spin_unlock(&vbq->lock);
2664
2665 return vaddr;
2666}
2667
2668static void free_vmap_block(struct vmap_block *vb)
2669{
2670 struct vmap_node *vn;
2671 struct vmap_block *tmp;
2672 struct xarray *xa;
2673
2674 xa = addr_to_vb_xa(vb->va->va_start);
2675 tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2676 BUG_ON(tmp != vb);
2677
2678 vn = addr_to_node(vb->va->va_start);
2679 spin_lock(&vn->busy.lock);
2680 unlink_va(vb->va, &vn->busy.root);
2681 spin_unlock(&vn->busy.lock);
2682
2683 free_vmap_area_noflush(vb->va);
2684 kfree_rcu(vb, rcu_head);
2685}
2686
2687static bool purge_fragmented_block(struct vmap_block *vb,
2688 struct list_head *purge_list, bool force_purge)
2689{
2690 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, vb->cpu);
2691
2692 if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2693 vb->dirty == VMAP_BBMAP_BITS)
2694 return false;
2695
2696 /* Don't overeagerly purge usable blocks unless requested */
2697 if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2698 return false;
2699
2700 /* prevent further allocs after releasing lock */
2701 WRITE_ONCE(vb->free, 0);
2702 /* prevent purging it again */
2703 WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2704 vb->dirty_min = 0;
2705 vb->dirty_max = VMAP_BBMAP_BITS;
2706 spin_lock(&vbq->lock);
2707 list_del_rcu(&vb->free_list);
2708 spin_unlock(&vbq->lock);
2709 list_add_tail(&vb->purge, purge_list);
2710 return true;
2711}
2712
2713static void free_purged_blocks(struct list_head *purge_list)
2714{
2715 struct vmap_block *vb, *n_vb;
2716
2717 list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2718 list_del(&vb->purge);
2719 free_vmap_block(vb);
2720 }
2721}
2722
2723static void purge_fragmented_blocks(int cpu)
2724{
2725 LIST_HEAD(purge);
2726 struct vmap_block *vb;
2727 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2728
2729 rcu_read_lock();
2730 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2731 unsigned long free = READ_ONCE(vb->free);
2732 unsigned long dirty = READ_ONCE(vb->dirty);
2733
2734 if (free + dirty != VMAP_BBMAP_BITS ||
2735 dirty == VMAP_BBMAP_BITS)
2736 continue;
2737
2738 spin_lock(&vb->lock);
2739 purge_fragmented_block(vb, &purge, true);
2740 spin_unlock(&vb->lock);
2741 }
2742 rcu_read_unlock();
2743 free_purged_blocks(&purge);
2744}
2745
2746static void purge_fragmented_blocks_allcpus(void)
2747{
2748 int cpu;
2749
2750 for_each_possible_cpu(cpu)
2751 purge_fragmented_blocks(cpu);
2752}
2753
2754static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2755{
2756 struct vmap_block_queue *vbq;
2757 struct vmap_block *vb;
2758 void *vaddr = NULL;
2759 unsigned int order;
2760
2761 BUG_ON(offset_in_page(size));
2762 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2763 if (WARN_ON(size == 0)) {
2764 /*
2765 * Allocating 0 bytes isn't what caller wants since
2766 * get_order(0) returns funny result. Just warn and terminate
2767 * early.
2768 */
2769 return ERR_PTR(-EINVAL);
2770 }
2771 order = get_order(size);
2772
2773 rcu_read_lock();
2774 vbq = raw_cpu_ptr(&vmap_block_queue);
2775 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2776 unsigned long pages_off;
2777
2778 if (READ_ONCE(vb->free) < (1UL << order))
2779 continue;
2780
2781 spin_lock(&vb->lock);
2782 if (vb->free < (1UL << order)) {
2783 spin_unlock(&vb->lock);
2784 continue;
2785 }
2786
2787 pages_off = VMAP_BBMAP_BITS - vb->free;
2788 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2789 WRITE_ONCE(vb->free, vb->free - (1UL << order));
2790 bitmap_set(vb->used_map, pages_off, (1UL << order));
2791 if (vb->free == 0) {
2792 spin_lock(&vbq->lock);
2793 list_del_rcu(&vb->free_list);
2794 spin_unlock(&vbq->lock);
2795 }
2796
2797 spin_unlock(&vb->lock);
2798 break;
2799 }
2800
2801 rcu_read_unlock();
2802
2803 /* Allocate new block if nothing was found */
2804 if (!vaddr)
2805 vaddr = new_vmap_block(order, gfp_mask);
2806
2807 return vaddr;
2808}
2809
2810static void vb_free(unsigned long addr, unsigned long size)
2811{
2812 unsigned long offset;
2813 unsigned int order;
2814 struct vmap_block *vb;
2815 struct xarray *xa;
2816
2817 BUG_ON(offset_in_page(size));
2818 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2819
2820 flush_cache_vunmap(addr, addr + size);
2821
2822 order = get_order(size);
2823 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2824
2825 xa = addr_to_vb_xa(addr);
2826 vb = xa_load(xa, addr_to_vb_idx(addr));
2827
2828 spin_lock(&vb->lock);
2829 bitmap_clear(vb->used_map, offset, (1UL << order));
2830 spin_unlock(&vb->lock);
2831
2832 vunmap_range_noflush(addr, addr + size);
2833
2834 if (debug_pagealloc_enabled_static())
2835 flush_tlb_kernel_range(addr, addr + size);
2836
2837 spin_lock(&vb->lock);
2838
2839 /* Expand the not yet TLB flushed dirty range */
2840 vb->dirty_min = min(vb->dirty_min, offset);
2841 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2842
2843 WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2844 if (vb->dirty == VMAP_BBMAP_BITS) {
2845 BUG_ON(vb->free);
2846 spin_unlock(&vb->lock);
2847 free_vmap_block(vb);
2848 } else
2849 spin_unlock(&vb->lock);
2850}
2851
2852static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2853{
2854 LIST_HEAD(purge_list);
2855 int cpu;
2856
2857 if (unlikely(!vmap_initialized))
2858 return;
2859
2860 mutex_lock(&vmap_purge_lock);
2861
2862 for_each_possible_cpu(cpu) {
2863 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2864 struct vmap_block *vb;
2865 unsigned long idx;
2866
2867 rcu_read_lock();
2868 xa_for_each(&vbq->vmap_blocks, idx, vb) {
2869 spin_lock(&vb->lock);
2870
2871 /*
2872 * Try to purge a fragmented block first. If it's
2873 * not purgeable, check whether there is dirty
2874 * space to be flushed.
2875 */
2876 if (!purge_fragmented_block(vb, &purge_list, false) &&
2877 vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2878 unsigned long va_start = vb->va->va_start;
2879 unsigned long s, e;
2880
2881 s = va_start + (vb->dirty_min << PAGE_SHIFT);
2882 e = va_start + (vb->dirty_max << PAGE_SHIFT);
2883
2884 start = min(s, start);
2885 end = max(e, end);
2886
2887 /* Prevent that this is flushed again */
2888 vb->dirty_min = VMAP_BBMAP_BITS;
2889 vb->dirty_max = 0;
2890
2891 flush = 1;
2892 }
2893 spin_unlock(&vb->lock);
2894 }
2895 rcu_read_unlock();
2896 }
2897 free_purged_blocks(&purge_list);
2898
2899 if (!__purge_vmap_area_lazy(start, end, false) && flush)
2900 flush_tlb_kernel_range(start, end);
2901 mutex_unlock(&vmap_purge_lock);
2902}
2903
2904/**
2905 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2906 *
2907 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2908 * to amortize TLB flushing overheads. What this means is that any page you
2909 * have now, may, in a former life, have been mapped into kernel virtual
2910 * address by the vmap layer and so there might be some CPUs with TLB entries
2911 * still referencing that page (additional to the regular 1:1 kernel mapping).
2912 *
2913 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2914 * be sure that none of the pages we have control over will have any aliases
2915 * from the vmap layer.
2916 */
2917void vm_unmap_aliases(void)
2918{
2919 unsigned long start = ULONG_MAX, end = 0;
2920 int flush = 0;
2921
2922 _vm_unmap_aliases(start, end, flush);
2923}
2924EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2925
2926/**
2927 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2928 * @mem: the pointer returned by vm_map_ram
2929 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2930 */
2931void vm_unmap_ram(const void *mem, unsigned int count)
2932{
2933 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2934 unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2935 struct vmap_area *va;
2936
2937 might_sleep();
2938 BUG_ON(!addr);
2939 BUG_ON(addr < VMALLOC_START);
2940 BUG_ON(addr > VMALLOC_END);
2941 BUG_ON(!PAGE_ALIGNED(addr));
2942
2943 kasan_poison_vmalloc(mem, size);
2944
2945 if (likely(count <= VMAP_MAX_ALLOC)) {
2946 debug_check_no_locks_freed(mem, size);
2947 vb_free(addr, size);
2948 return;
2949 }
2950
2951 va = find_unlink_vmap_area(addr);
2952 if (WARN_ON_ONCE(!va))
2953 return;
2954
2955 debug_check_no_locks_freed((void *)va->va_start, va_size(va));
2956 free_unmap_vmap_area(va);
2957}
2958EXPORT_SYMBOL(vm_unmap_ram);
2959
2960/**
2961 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2962 * @pages: an array of pointers to the pages to be mapped
2963 * @count: number of pages
2964 * @node: prefer to allocate data structures on this node
2965 *
2966 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2967 * faster than vmap so it's good. But if you mix long-life and short-life
2968 * objects with vm_map_ram(), it could consume lots of address space through
2969 * fragmentation (especially on a 32bit machine). You could see failures in
2970 * the end. Please use this function for short-lived objects.
2971 *
2972 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2973 */
2974void *vm_map_ram(struct page **pages, unsigned int count, int node)
2975{
2976 unsigned long size = (unsigned long)count << PAGE_SHIFT;
2977 unsigned long addr;
2978 void *mem;
2979
2980 if (likely(count <= VMAP_MAX_ALLOC)) {
2981 mem = vb_alloc(size, GFP_KERNEL);
2982 if (IS_ERR(mem))
2983 return NULL;
2984 addr = (unsigned long)mem;
2985 } else {
2986 struct vmap_area *va;
2987 va = alloc_vmap_area(size, PAGE_SIZE,
2988 VMALLOC_START, VMALLOC_END,
2989 node, GFP_KERNEL, VMAP_RAM,
2990 NULL);
2991 if (IS_ERR(va))
2992 return NULL;
2993
2994 addr = va->va_start;
2995 mem = (void *)addr;
2996 }
2997
2998 if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2999 pages, PAGE_SHIFT) < 0) {
3000 vm_unmap_ram(mem, count);
3001 return NULL;
3002 }
3003
3004 /*
3005 * Mark the pages as accessible, now that they are mapped.
3006 * With hardware tag-based KASAN, marking is skipped for
3007 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3008 */
3009 mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
3010
3011 return mem;
3012}
3013EXPORT_SYMBOL(vm_map_ram);
3014
3015static struct vm_struct *vmlist __initdata;
3016
3017static inline unsigned int vm_area_page_order(struct vm_struct *vm)
3018{
3019#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3020 return vm->page_order;
3021#else
3022 return 0;
3023#endif
3024}
3025
3026unsigned int get_vm_area_page_order(struct vm_struct *vm)
3027{
3028 return vm_area_page_order(vm);
3029}
3030
3031static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
3032{
3033#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3034 vm->page_order = order;
3035#else
3036 BUG_ON(order != 0);
3037#endif
3038}
3039
3040/**
3041 * vm_area_add_early - add vmap area early during boot
3042 * @vm: vm_struct to add
3043 *
3044 * This function is used to add fixed kernel vm area to vmlist before
3045 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
3046 * should contain proper values and the other fields should be zero.
3047 *
3048 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3049 */
3050void __init vm_area_add_early(struct vm_struct *vm)
3051{
3052 struct vm_struct *tmp, **p;
3053
3054 BUG_ON(vmap_initialized);
3055 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
3056 if (tmp->addr >= vm->addr) {
3057 BUG_ON(tmp->addr < vm->addr + vm->size);
3058 break;
3059 } else
3060 BUG_ON(tmp->addr + tmp->size > vm->addr);
3061 }
3062 vm->next = *p;
3063 *p = vm;
3064}
3065
3066/**
3067 * vm_area_register_early - register vmap area early during boot
3068 * @vm: vm_struct to register
3069 * @align: requested alignment
3070 *
3071 * This function is used to register kernel vm area before
3072 * vmalloc_init() is called. @vm->size and @vm->flags should contain
3073 * proper values on entry and other fields should be zero. On return,
3074 * vm->addr contains the allocated address.
3075 *
3076 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3077 */
3078void __init vm_area_register_early(struct vm_struct *vm, size_t align)
3079{
3080 unsigned long addr = ALIGN(VMALLOC_START, align);
3081 struct vm_struct *cur, **p;
3082
3083 BUG_ON(vmap_initialized);
3084
3085 for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
3086 if ((unsigned long)cur->addr - addr >= vm->size)
3087 break;
3088 addr = ALIGN((unsigned long)cur->addr + cur->size, align);
3089 }
3090
3091 BUG_ON(addr > VMALLOC_END - vm->size);
3092 vm->addr = (void *)addr;
3093 vm->next = *p;
3094 *p = vm;
3095 kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
3096}
3097
3098static void clear_vm_uninitialized_flag(struct vm_struct *vm)
3099{
3100 /*
3101 * Before removing VM_UNINITIALIZED,
3102 * we should make sure that vm has proper values.
3103 * Pair with smp_rmb() in show_numa_info().
3104 */
3105 smp_wmb();
3106 vm->flags &= ~VM_UNINITIALIZED;
3107}
3108
3109struct vm_struct *__get_vm_area_node(unsigned long size,
3110 unsigned long align, unsigned long shift, unsigned long flags,
3111 unsigned long start, unsigned long end, int node,
3112 gfp_t gfp_mask, const void *caller)
3113{
3114 struct vmap_area *va;
3115 struct vm_struct *area;
3116 unsigned long requested_size = size;
3117
3118 BUG_ON(in_interrupt());
3119 size = ALIGN(size, 1ul << shift);
3120 if (unlikely(!size))
3121 return NULL;
3122
3123 if (flags & VM_IOREMAP)
3124 align = 1ul << clamp_t(int, get_count_order_long(size),
3125 PAGE_SHIFT, IOREMAP_MAX_ORDER);
3126
3127 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
3128 if (unlikely(!area))
3129 return NULL;
3130
3131 if (!(flags & VM_NO_GUARD))
3132 size += PAGE_SIZE;
3133
3134 area->flags = flags;
3135 area->caller = caller;
3136
3137 va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0, area);
3138 if (IS_ERR(va)) {
3139 kfree(area);
3140 return NULL;
3141 }
3142
3143 /*
3144 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
3145 * best-effort approach, as they can be mapped outside of vmalloc code.
3146 * For VM_ALLOC mappings, the pages are marked as accessible after
3147 * getting mapped in __vmalloc_node_range().
3148 * With hardware tag-based KASAN, marking is skipped for
3149 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3150 */
3151 if (!(flags & VM_ALLOC))
3152 area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
3153 KASAN_VMALLOC_PROT_NORMAL);
3154
3155 return area;
3156}
3157
3158struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
3159 unsigned long start, unsigned long end,
3160 const void *caller)
3161{
3162 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
3163 NUMA_NO_NODE, GFP_KERNEL, caller);
3164}
3165
3166/**
3167 * get_vm_area - reserve a contiguous kernel virtual area
3168 * @size: size of the area
3169 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
3170 *
3171 * Search an area of @size in the kernel virtual mapping area,
3172 * and reserved it for out purposes. Returns the area descriptor
3173 * on success or %NULL on failure.
3174 *
3175 * Return: the area descriptor on success or %NULL on failure.
3176 */
3177struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
3178{
3179 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3180 VMALLOC_START, VMALLOC_END,
3181 NUMA_NO_NODE, GFP_KERNEL,
3182 __builtin_return_address(0));
3183}
3184
3185struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
3186 const void *caller)
3187{
3188 return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3189 VMALLOC_START, VMALLOC_END,
3190 NUMA_NO_NODE, GFP_KERNEL, caller);
3191}
3192
3193/**
3194 * find_vm_area - find a continuous kernel virtual area
3195 * @addr: base address
3196 *
3197 * Search for the kernel VM area starting at @addr, and return it.
3198 * It is up to the caller to do all required locking to keep the returned
3199 * pointer valid.
3200 *
3201 * Return: the area descriptor on success or %NULL on failure.
3202 */
3203struct vm_struct *find_vm_area(const void *addr)
3204{
3205 struct vmap_area *va;
3206
3207 va = find_vmap_area((unsigned long)addr);
3208 if (!va)
3209 return NULL;
3210
3211 return va->vm;
3212}
3213
3214/**
3215 * remove_vm_area - find and remove a continuous kernel virtual area
3216 * @addr: base address
3217 *
3218 * Search for the kernel VM area starting at @addr, and remove it.
3219 * This function returns the found VM area, but using it is NOT safe
3220 * on SMP machines, except for its size or flags.
3221 *
3222 * Return: the area descriptor on success or %NULL on failure.
3223 */
3224struct vm_struct *remove_vm_area(const void *addr)
3225{
3226 struct vmap_area *va;
3227 struct vm_struct *vm;
3228
3229 might_sleep();
3230
3231 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
3232 addr))
3233 return NULL;
3234
3235 va = find_unlink_vmap_area((unsigned long)addr);
3236 if (!va || !va->vm)
3237 return NULL;
3238 vm = va->vm;
3239
3240 debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
3241 debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
3242 kasan_free_module_shadow(vm);
3243 kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
3244
3245 free_unmap_vmap_area(va);
3246 return vm;
3247}
3248
3249static inline void set_area_direct_map(const struct vm_struct *area,
3250 int (*set_direct_map)(struct page *page))
3251{
3252 int i;
3253
3254 /* HUGE_VMALLOC passes small pages to set_direct_map */
3255 for (i = 0; i < area->nr_pages; i++)
3256 if (page_address(area->pages[i]))
3257 set_direct_map(area->pages[i]);
3258}
3259
3260/*
3261 * Flush the vm mapping and reset the direct map.
3262 */
3263static void vm_reset_perms(struct vm_struct *area)
3264{
3265 unsigned long start = ULONG_MAX, end = 0;
3266 unsigned int page_order = vm_area_page_order(area);
3267 int flush_dmap = 0;
3268 int i;
3269
3270 /*
3271 * Find the start and end range of the direct mappings to make sure that
3272 * the vm_unmap_aliases() flush includes the direct map.
3273 */
3274 for (i = 0; i < area->nr_pages; i += 1U << page_order) {
3275 unsigned long addr = (unsigned long)page_address(area->pages[i]);
3276
3277 if (addr) {
3278 unsigned long page_size;
3279
3280 page_size = PAGE_SIZE << page_order;
3281 start = min(addr, start);
3282 end = max(addr + page_size, end);
3283 flush_dmap = 1;
3284 }
3285 }
3286
3287 /*
3288 * Set direct map to something invalid so that it won't be cached if
3289 * there are any accesses after the TLB flush, then flush the TLB and
3290 * reset the direct map permissions to the default.
3291 */
3292 set_area_direct_map(area, set_direct_map_invalid_noflush);
3293 _vm_unmap_aliases(start, end, flush_dmap);
3294 set_area_direct_map(area, set_direct_map_default_noflush);
3295}
3296
3297static void delayed_vfree_work(struct work_struct *w)
3298{
3299 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
3300 struct llist_node *t, *llnode;
3301
3302 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
3303 vfree(llnode);
3304}
3305
3306/**
3307 * vfree_atomic - release memory allocated by vmalloc()
3308 * @addr: memory base address
3309 *
3310 * This one is just like vfree() but can be called in any atomic context
3311 * except NMIs.
3312 */
3313void vfree_atomic(const void *addr)
3314{
3315 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
3316
3317 BUG_ON(in_nmi());
3318 kmemleak_free(addr);
3319
3320 /*
3321 * Use raw_cpu_ptr() because this can be called from preemptible
3322 * context. Preemption is absolutely fine here, because the llist_add()
3323 * implementation is lockless, so it works even if we are adding to
3324 * another cpu's list. schedule_work() should be fine with this too.
3325 */
3326 if (addr && llist_add((struct llist_node *)addr, &p->list))
3327 schedule_work(&p->wq);
3328}
3329
3330/**
3331 * vfree - Release memory allocated by vmalloc()
3332 * @addr: Memory base address
3333 *
3334 * Free the virtually continuous memory area starting at @addr, as obtained
3335 * from one of the vmalloc() family of APIs. This will usually also free the
3336 * physical memory underlying the virtual allocation, but that memory is
3337 * reference counted, so it will not be freed until the last user goes away.
3338 *
3339 * If @addr is NULL, no operation is performed.
3340 *
3341 * Context:
3342 * May sleep if called *not* from interrupt context.
3343 * Must not be called in NMI context (strictly speaking, it could be
3344 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
3345 * conventions for vfree() arch-dependent would be a really bad idea).
3346 */
3347void vfree(const void *addr)
3348{
3349 struct vm_struct *vm;
3350 int i;
3351
3352 if (unlikely(in_interrupt())) {
3353 vfree_atomic(addr);
3354 return;
3355 }
3356
3357 BUG_ON(in_nmi());
3358 kmemleak_free(addr);
3359 might_sleep();
3360
3361 if (!addr)
3362 return;
3363
3364 vm = remove_vm_area(addr);
3365 if (unlikely(!vm)) {
3366 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
3367 addr);
3368 return;
3369 }
3370
3371 if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
3372 vm_reset_perms(vm);
3373 for (i = 0; i < vm->nr_pages; i++) {
3374 struct page *page = vm->pages[i];
3375
3376 BUG_ON(!page);
3377 if (!(vm->flags & VM_MAP_PUT_PAGES))
3378 mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
3379 /*
3380 * High-order allocs for huge vmallocs are split, so
3381 * can be freed as an array of order-0 allocations
3382 */
3383 __free_page(page);
3384 cond_resched();
3385 }
3386 if (!(vm->flags & VM_MAP_PUT_PAGES))
3387 atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
3388 kvfree(vm->pages);
3389 kfree(vm);
3390}
3391EXPORT_SYMBOL(vfree);
3392
3393/**
3394 * vunmap - release virtual mapping obtained by vmap()
3395 * @addr: memory base address
3396 *
3397 * Free the virtually contiguous memory area starting at @addr,
3398 * which was created from the page array passed to vmap().
3399 *
3400 * Must not be called in interrupt context.
3401 */
3402void vunmap(const void *addr)
3403{
3404 struct vm_struct *vm;
3405
3406 BUG_ON(in_interrupt());
3407 might_sleep();
3408
3409 if (!addr)
3410 return;
3411 vm = remove_vm_area(addr);
3412 if (unlikely(!vm)) {
3413 WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
3414 addr);
3415 return;
3416 }
3417 kfree(vm);
3418}
3419EXPORT_SYMBOL(vunmap);
3420
3421/**
3422 * vmap - map an array of pages into virtually contiguous space
3423 * @pages: array of page pointers
3424 * @count: number of pages to map
3425 * @flags: vm_area->flags
3426 * @prot: page protection for the mapping
3427 *
3428 * Maps @count pages from @pages into contiguous kernel virtual space.
3429 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
3430 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
3431 * are transferred from the caller to vmap(), and will be freed / dropped when
3432 * vfree() is called on the return value.
3433 *
3434 * Return: the address of the area or %NULL on failure
3435 */
3436void *vmap(struct page **pages, unsigned int count,
3437 unsigned long flags, pgprot_t prot)
3438{
3439 struct vm_struct *area;
3440 unsigned long addr;
3441 unsigned long size; /* In bytes */
3442
3443 might_sleep();
3444
3445 if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
3446 return NULL;
3447
3448 /*
3449 * Your top guard is someone else's bottom guard. Not having a top
3450 * guard compromises someone else's mappings too.
3451 */
3452 if (WARN_ON_ONCE(flags & VM_NO_GUARD))
3453 flags &= ~VM_NO_GUARD;
3454
3455 if (count > totalram_pages())
3456 return NULL;
3457
3458 size = (unsigned long)count << PAGE_SHIFT;
3459 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
3460 if (!area)
3461 return NULL;
3462
3463 addr = (unsigned long)area->addr;
3464 if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
3465 pages, PAGE_SHIFT) < 0) {
3466 vunmap(area->addr);
3467 return NULL;
3468 }
3469
3470 if (flags & VM_MAP_PUT_PAGES) {
3471 area->pages = pages;
3472 area->nr_pages = count;
3473 }
3474 return area->addr;
3475}
3476EXPORT_SYMBOL(vmap);
3477
3478#ifdef CONFIG_VMAP_PFN
3479struct vmap_pfn_data {
3480 unsigned long *pfns;
3481 pgprot_t prot;
3482 unsigned int idx;
3483};
3484
3485static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
3486{
3487 struct vmap_pfn_data *data = private;
3488 unsigned long pfn = data->pfns[data->idx];
3489 pte_t ptent;
3490
3491 if (WARN_ON_ONCE(pfn_valid(pfn)))
3492 return -EINVAL;
3493
3494 ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
3495 set_pte_at(&init_mm, addr, pte, ptent);
3496
3497 data->idx++;
3498 return 0;
3499}
3500
3501/**
3502 * vmap_pfn - map an array of PFNs into virtually contiguous space
3503 * @pfns: array of PFNs
3504 * @count: number of pages to map
3505 * @prot: page protection for the mapping
3506 *
3507 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
3508 * the start address of the mapping.
3509 */
3510void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
3511{
3512 struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
3513 struct vm_struct *area;
3514
3515 area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
3516 __builtin_return_address(0));
3517 if (!area)
3518 return NULL;
3519 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3520 count * PAGE_SIZE, vmap_pfn_apply, &data)) {
3521 free_vm_area(area);
3522 return NULL;
3523 }
3524
3525 flush_cache_vmap((unsigned long)area->addr,
3526 (unsigned long)area->addr + count * PAGE_SIZE);
3527
3528 return area->addr;
3529}
3530EXPORT_SYMBOL_GPL(vmap_pfn);
3531#endif /* CONFIG_VMAP_PFN */
3532
3533static inline unsigned int
3534vm_area_alloc_pages(gfp_t gfp, int nid,
3535 unsigned int order, unsigned int nr_pages, struct page **pages)
3536{
3537 unsigned int nr_allocated = 0;
3538 struct page *page;
3539 int i;
3540
3541 /*
3542 * For order-0 pages we make use of bulk allocator, if
3543 * the page array is partly or not at all populated due
3544 * to fails, fallback to a single page allocator that is
3545 * more permissive.
3546 */
3547 if (!order) {
3548 while (nr_allocated < nr_pages) {
3549 unsigned int nr, nr_pages_request;
3550
3551 /*
3552 * A maximum allowed request is hard-coded and is 100
3553 * pages per call. That is done in order to prevent a
3554 * long preemption off scenario in the bulk-allocator
3555 * so the range is [1:100].
3556 */
3557 nr_pages_request = min(100U, nr_pages - nr_allocated);
3558
3559 /* memory allocation should consider mempolicy, we can't
3560 * wrongly use nearest node when nid == NUMA_NO_NODE,
3561 * otherwise memory may be allocated in only one node,
3562 * but mempolicy wants to alloc memory by interleaving.
3563 */
3564 if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3565 nr = alloc_pages_bulk_array_mempolicy_noprof(gfp,
3566 nr_pages_request,
3567 pages + nr_allocated);
3568 else
3569 nr = alloc_pages_bulk_array_node_noprof(gfp, nid,
3570 nr_pages_request,
3571 pages + nr_allocated);
3572
3573 nr_allocated += nr;
3574 cond_resched();
3575
3576 /*
3577 * If zero or pages were obtained partly,
3578 * fallback to a single page allocator.
3579 */
3580 if (nr != nr_pages_request)
3581 break;
3582 }
3583 }
3584
3585 /* High-order pages or fallback path if "bulk" fails. */
3586 while (nr_allocated < nr_pages) {
3587 if (!(gfp & __GFP_NOFAIL) && fatal_signal_pending(current))
3588 break;
3589
3590 if (nid == NUMA_NO_NODE)
3591 page = alloc_pages_noprof(gfp, order);
3592 else
3593 page = alloc_pages_node_noprof(nid, gfp, order);
3594
3595 if (unlikely(!page))
3596 break;
3597
3598 /*
3599 * High-order allocations must be able to be treated as
3600 * independent small pages by callers (as they can with
3601 * small-page vmallocs). Some drivers do their own refcounting
3602 * on vmalloc_to_page() pages, some use page->mapping,
3603 * page->lru, etc.
3604 */
3605 if (order)
3606 split_page(page, order);
3607
3608 /*
3609 * Careful, we allocate and map page-order pages, but
3610 * tracking is done per PAGE_SIZE page so as to keep the
3611 * vm_struct APIs independent of the physical/mapped size.
3612 */
3613 for (i = 0; i < (1U << order); i++)
3614 pages[nr_allocated + i] = page + i;
3615
3616 cond_resched();
3617 nr_allocated += 1U << order;
3618 }
3619
3620 return nr_allocated;
3621}
3622
3623static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3624 pgprot_t prot, unsigned int page_shift,
3625 int node)
3626{
3627 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3628 bool nofail = gfp_mask & __GFP_NOFAIL;
3629 unsigned long addr = (unsigned long)area->addr;
3630 unsigned long size = get_vm_area_size(area);
3631 unsigned long array_size;
3632 unsigned int nr_small_pages = size >> PAGE_SHIFT;
3633 unsigned int page_order;
3634 unsigned int flags;
3635 int ret;
3636
3637 array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3638
3639 if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3640 gfp_mask |= __GFP_HIGHMEM;
3641
3642 /* Please note that the recursion is strictly bounded. */
3643 if (array_size > PAGE_SIZE) {
3644 area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node,
3645 area->caller);
3646 } else {
3647 area->pages = kmalloc_node_noprof(array_size, nested_gfp, node);
3648 }
3649
3650 if (!area->pages) {
3651 warn_alloc(gfp_mask, NULL,
3652 "vmalloc error: size %lu, failed to allocated page array size %lu",
3653 nr_small_pages * PAGE_SIZE, array_size);
3654 free_vm_area(area);
3655 return NULL;
3656 }
3657
3658 set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3659 page_order = vm_area_page_order(area);
3660
3661 /*
3662 * High-order nofail allocations are really expensive and
3663 * potentially dangerous (pre-mature OOM, disruptive reclaim
3664 * and compaction etc.
3665 *
3666 * Please note, the __vmalloc_node_range_noprof() falls-back
3667 * to order-0 pages if high-order attempt is unsuccessful.
3668 */
3669 area->nr_pages = vm_area_alloc_pages((page_order ?
3670 gfp_mask & ~__GFP_NOFAIL : gfp_mask) | __GFP_NOWARN,
3671 node, page_order, nr_small_pages, area->pages);
3672
3673 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3674 if (gfp_mask & __GFP_ACCOUNT) {
3675 int i;
3676
3677 for (i = 0; i < area->nr_pages; i++)
3678 mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3679 }
3680
3681 /*
3682 * If not enough pages were obtained to accomplish an
3683 * allocation request, free them via vfree() if any.
3684 */
3685 if (area->nr_pages != nr_small_pages) {
3686 /*
3687 * vm_area_alloc_pages() can fail due to insufficient memory but
3688 * also:-
3689 *
3690 * - a pending fatal signal
3691 * - insufficient huge page-order pages
3692 *
3693 * Since we always retry allocations at order-0 in the huge page
3694 * case a warning for either is spurious.
3695 */
3696 if (!fatal_signal_pending(current) && page_order == 0)
3697 warn_alloc(gfp_mask, NULL,
3698 "vmalloc error: size %lu, failed to allocate pages",
3699 area->nr_pages * PAGE_SIZE);
3700 goto fail;
3701 }
3702
3703 /*
3704 * page tables allocations ignore external gfp mask, enforce it
3705 * by the scope API
3706 */
3707 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3708 flags = memalloc_nofs_save();
3709 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3710 flags = memalloc_noio_save();
3711
3712 do {
3713 ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3714 page_shift);
3715 if (nofail && (ret < 0))
3716 schedule_timeout_uninterruptible(1);
3717 } while (nofail && (ret < 0));
3718
3719 if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3720 memalloc_nofs_restore(flags);
3721 else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3722 memalloc_noio_restore(flags);
3723
3724 if (ret < 0) {
3725 warn_alloc(gfp_mask, NULL,
3726 "vmalloc error: size %lu, failed to map pages",
3727 area->nr_pages * PAGE_SIZE);
3728 goto fail;
3729 }
3730
3731 return area->addr;
3732
3733fail:
3734 vfree(area->addr);
3735 return NULL;
3736}
3737
3738/**
3739 * __vmalloc_node_range - allocate virtually contiguous memory
3740 * @size: allocation size
3741 * @align: desired alignment
3742 * @start: vm area range start
3743 * @end: vm area range end
3744 * @gfp_mask: flags for the page level allocator
3745 * @prot: protection mask for the allocated pages
3746 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
3747 * @node: node to use for allocation or NUMA_NO_NODE
3748 * @caller: caller's return address
3749 *
3750 * Allocate enough pages to cover @size from the page level
3751 * allocator with @gfp_mask flags. Please note that the full set of gfp
3752 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3753 * supported.
3754 * Zone modifiers are not supported. From the reclaim modifiers
3755 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3756 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3757 * __GFP_RETRY_MAYFAIL are not supported).
3758 *
3759 * __GFP_NOWARN can be used to suppress failures messages.
3760 *
3761 * Map them into contiguous kernel virtual space, using a pagetable
3762 * protection of @prot.
3763 *
3764 * Return: the address of the area or %NULL on failure
3765 */
3766void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align,
3767 unsigned long start, unsigned long end, gfp_t gfp_mask,
3768 pgprot_t prot, unsigned long vm_flags, int node,
3769 const void *caller)
3770{
3771 struct vm_struct *area;
3772 void *ret;
3773 kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3774 unsigned long real_size = size;
3775 unsigned long real_align = align;
3776 unsigned int shift = PAGE_SHIFT;
3777
3778 if (WARN_ON_ONCE(!size))
3779 return NULL;
3780
3781 if ((size >> PAGE_SHIFT) > totalram_pages()) {
3782 warn_alloc(gfp_mask, NULL,
3783 "vmalloc error: size %lu, exceeds total pages",
3784 real_size);
3785 return NULL;
3786 }
3787
3788 if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3789 /*
3790 * Try huge pages. Only try for PAGE_KERNEL allocations,
3791 * others like modules don't yet expect huge pages in
3792 * their allocations due to apply_to_page_range not
3793 * supporting them.
3794 */
3795
3796 if (arch_vmap_pmd_supported(prot) && size >= PMD_SIZE)
3797 shift = PMD_SHIFT;
3798 else
3799 shift = arch_vmap_pte_supported_shift(size);
3800
3801 align = max(real_align, 1UL << shift);
3802 size = ALIGN(real_size, 1UL << shift);
3803 }
3804
3805again:
3806 area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3807 VM_UNINITIALIZED | vm_flags, start, end, node,
3808 gfp_mask, caller);
3809 if (!area) {
3810 bool nofail = gfp_mask & __GFP_NOFAIL;
3811 warn_alloc(gfp_mask, NULL,
3812 "vmalloc error: size %lu, vm_struct allocation failed%s",
3813 real_size, (nofail) ? ". Retrying." : "");
3814 if (nofail) {
3815 schedule_timeout_uninterruptible(1);
3816 goto again;
3817 }
3818 goto fail;
3819 }
3820
3821 /*
3822 * Prepare arguments for __vmalloc_area_node() and
3823 * kasan_unpoison_vmalloc().
3824 */
3825 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3826 if (kasan_hw_tags_enabled()) {
3827 /*
3828 * Modify protection bits to allow tagging.
3829 * This must be done before mapping.
3830 */
3831 prot = arch_vmap_pgprot_tagged(prot);
3832
3833 /*
3834 * Skip page_alloc poisoning and zeroing for physical
3835 * pages backing VM_ALLOC mapping. Memory is instead
3836 * poisoned and zeroed by kasan_unpoison_vmalloc().
3837 */
3838 gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3839 }
3840
3841 /* Take note that the mapping is PAGE_KERNEL. */
3842 kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3843 }
3844
3845 /* Allocate physical pages and map them into vmalloc space. */
3846 ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3847 if (!ret)
3848 goto fail;
3849
3850 /*
3851 * Mark the pages as accessible, now that they are mapped.
3852 * The condition for setting KASAN_VMALLOC_INIT should complement the
3853 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3854 * to make sure that memory is initialized under the same conditions.
3855 * Tag-based KASAN modes only assign tags to normal non-executable
3856 * allocations, see __kasan_unpoison_vmalloc().
3857 */
3858 kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3859 if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3860 (gfp_mask & __GFP_SKIP_ZERO))
3861 kasan_flags |= KASAN_VMALLOC_INIT;
3862 /* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3863 area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3864
3865 /*
3866 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3867 * flag. It means that vm_struct is not fully initialized.
3868 * Now, it is fully initialized, so remove this flag here.
3869 */
3870 clear_vm_uninitialized_flag(area);
3871
3872 size = PAGE_ALIGN(size);
3873 if (!(vm_flags & VM_DEFER_KMEMLEAK))
3874 kmemleak_vmalloc(area, size, gfp_mask);
3875
3876 return area->addr;
3877
3878fail:
3879 if (shift > PAGE_SHIFT) {
3880 shift = PAGE_SHIFT;
3881 align = real_align;
3882 size = real_size;
3883 goto again;
3884 }
3885
3886 return NULL;
3887}
3888
3889/**
3890 * __vmalloc_node - allocate virtually contiguous memory
3891 * @size: allocation size
3892 * @align: desired alignment
3893 * @gfp_mask: flags for the page level allocator
3894 * @node: node to use for allocation or NUMA_NO_NODE
3895 * @caller: caller's return address
3896 *
3897 * Allocate enough pages to cover @size from the page level allocator with
3898 * @gfp_mask flags. Map them into contiguous kernel virtual space.
3899 *
3900 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3901 * and __GFP_NOFAIL are not supported
3902 *
3903 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3904 * with mm people.
3905 *
3906 * Return: pointer to the allocated memory or %NULL on error
3907 */
3908void *__vmalloc_node_noprof(unsigned long size, unsigned long align,
3909 gfp_t gfp_mask, int node, const void *caller)
3910{
3911 return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END,
3912 gfp_mask, PAGE_KERNEL, 0, node, caller);
3913}
3914/*
3915 * This is only for performance analysis of vmalloc and stress purpose.
3916 * It is required by vmalloc test module, therefore do not use it other
3917 * than that.
3918 */
3919#ifdef CONFIG_TEST_VMALLOC_MODULE
3920EXPORT_SYMBOL_GPL(__vmalloc_node_noprof);
3921#endif
3922
3923void *__vmalloc_noprof(unsigned long size, gfp_t gfp_mask)
3924{
3925 return __vmalloc_node_noprof(size, 1, gfp_mask, NUMA_NO_NODE,
3926 __builtin_return_address(0));
3927}
3928EXPORT_SYMBOL(__vmalloc_noprof);
3929
3930/**
3931 * vmalloc - allocate virtually contiguous memory
3932 * @size: allocation size
3933 *
3934 * Allocate enough pages to cover @size from the page level
3935 * allocator and map them into contiguous kernel virtual space.
3936 *
3937 * For tight control over page level allocator and protection flags
3938 * use __vmalloc() instead.
3939 *
3940 * Return: pointer to the allocated memory or %NULL on error
3941 */
3942void *vmalloc_noprof(unsigned long size)
3943{
3944 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3945 __builtin_return_address(0));
3946}
3947EXPORT_SYMBOL(vmalloc_noprof);
3948
3949/**
3950 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3951 * @size: allocation size
3952 * @gfp_mask: flags for the page level allocator
3953 *
3954 * Allocate enough pages to cover @size from the page level
3955 * allocator and map them into contiguous kernel virtual space.
3956 * If @size is greater than or equal to PMD_SIZE, allow using
3957 * huge pages for the memory
3958 *
3959 * Return: pointer to the allocated memory or %NULL on error
3960 */
3961void *vmalloc_huge_noprof(unsigned long size, gfp_t gfp_mask)
3962{
3963 return __vmalloc_node_range_noprof(size, 1, VMALLOC_START, VMALLOC_END,
3964 gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3965 NUMA_NO_NODE, __builtin_return_address(0));
3966}
3967EXPORT_SYMBOL_GPL(vmalloc_huge_noprof);
3968
3969/**
3970 * vzalloc - allocate virtually contiguous memory with zero fill
3971 * @size: allocation size
3972 *
3973 * Allocate enough pages to cover @size from the page level
3974 * allocator and map them into contiguous kernel virtual space.
3975 * The memory allocated is set to zero.
3976 *
3977 * For tight control over page level allocator and protection flags
3978 * use __vmalloc() instead.
3979 *
3980 * Return: pointer to the allocated memory or %NULL on error
3981 */
3982void *vzalloc_noprof(unsigned long size)
3983{
3984 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3985 __builtin_return_address(0));
3986}
3987EXPORT_SYMBOL(vzalloc_noprof);
3988
3989/**
3990 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3991 * @size: allocation size
3992 *
3993 * The resulting memory area is zeroed so it can be mapped to userspace
3994 * without leaking data.
3995 *
3996 * Return: pointer to the allocated memory or %NULL on error
3997 */
3998void *vmalloc_user_noprof(unsigned long size)
3999{
4000 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
4001 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
4002 VM_USERMAP, NUMA_NO_NODE,
4003 __builtin_return_address(0));
4004}
4005EXPORT_SYMBOL(vmalloc_user_noprof);
4006
4007/**
4008 * vmalloc_node - allocate memory on a specific node
4009 * @size: allocation size
4010 * @node: numa node
4011 *
4012 * Allocate enough pages to cover @size from the page level
4013 * allocator and map them into contiguous kernel virtual space.
4014 *
4015 * For tight control over page level allocator and protection flags
4016 * use __vmalloc() instead.
4017 *
4018 * Return: pointer to the allocated memory or %NULL on error
4019 */
4020void *vmalloc_node_noprof(unsigned long size, int node)
4021{
4022 return __vmalloc_node_noprof(size, 1, GFP_KERNEL, node,
4023 __builtin_return_address(0));
4024}
4025EXPORT_SYMBOL(vmalloc_node_noprof);
4026
4027/**
4028 * vzalloc_node - allocate memory on a specific node with zero fill
4029 * @size: allocation size
4030 * @node: numa node
4031 *
4032 * Allocate enough pages to cover @size from the page level
4033 * allocator and map them into contiguous kernel virtual space.
4034 * The memory allocated is set to zero.
4035 *
4036 * Return: pointer to the allocated memory or %NULL on error
4037 */
4038void *vzalloc_node_noprof(unsigned long size, int node)
4039{
4040 return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, node,
4041 __builtin_return_address(0));
4042}
4043EXPORT_SYMBOL(vzalloc_node_noprof);
4044
4045/**
4046 * vrealloc - reallocate virtually contiguous memory; contents remain unchanged
4047 * @p: object to reallocate memory for
4048 * @size: the size to reallocate
4049 * @flags: the flags for the page level allocator
4050 *
4051 * If @p is %NULL, vrealloc() behaves exactly like vmalloc(). If @size is 0 and
4052 * @p is not a %NULL pointer, the object pointed to is freed.
4053 *
4054 * If __GFP_ZERO logic is requested, callers must ensure that, starting with the
4055 * initial memory allocation, every subsequent call to this API for the same
4056 * memory allocation is flagged with __GFP_ZERO. Otherwise, it is possible that
4057 * __GFP_ZERO is not fully honored by this API.
4058 *
4059 * In any case, the contents of the object pointed to are preserved up to the
4060 * lesser of the new and old sizes.
4061 *
4062 * This function must not be called concurrently with itself or vfree() for the
4063 * same memory allocation.
4064 *
4065 * Return: pointer to the allocated memory; %NULL if @size is zero or in case of
4066 * failure
4067 */
4068void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
4069{
4070 size_t old_size = 0;
4071 void *n;
4072
4073 if (!size) {
4074 vfree(p);
4075 return NULL;
4076 }
4077
4078 if (p) {
4079 struct vm_struct *vm;
4080
4081 vm = find_vm_area(p);
4082 if (unlikely(!vm)) {
4083 WARN(1, "Trying to vrealloc() nonexistent vm area (%p)\n", p);
4084 return NULL;
4085 }
4086
4087 old_size = get_vm_area_size(vm);
4088 }
4089
4090 /*
4091 * TODO: Shrink the vm_area, i.e. unmap and free unused pages. What
4092 * would be a good heuristic for when to shrink the vm_area?
4093 */
4094 if (size <= old_size) {
4095 /* Zero out spare memory. */
4096 if (want_init_on_alloc(flags))
4097 memset((void *)p + size, 0, old_size - size);
4098 kasan_poison_vmalloc(p + size, old_size - size);
4099 kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
4100 return (void *)p;
4101 }
4102
4103 /* TODO: Grow the vm_area, i.e. allocate and map additional pages. */
4104 n = __vmalloc_noprof(size, flags);
4105 if (!n)
4106 return NULL;
4107
4108 if (p) {
4109 memcpy(n, p, old_size);
4110 vfree(p);
4111 }
4112
4113 return n;
4114}
4115
4116#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
4117#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4118#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
4119#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
4120#else
4121/*
4122 * 64b systems should always have either DMA or DMA32 zones. For others
4123 * GFP_DMA32 should do the right thing and use the normal zone.
4124 */
4125#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4126#endif
4127
4128/**
4129 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
4130 * @size: allocation size
4131 *
4132 * Allocate enough 32bit PA addressable pages to cover @size from the
4133 * page level allocator and map them into contiguous kernel virtual space.
4134 *
4135 * Return: pointer to the allocated memory or %NULL on error
4136 */
4137void *vmalloc_32_noprof(unsigned long size)
4138{
4139 return __vmalloc_node_noprof(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
4140 __builtin_return_address(0));
4141}
4142EXPORT_SYMBOL(vmalloc_32_noprof);
4143
4144/**
4145 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
4146 * @size: allocation size
4147 *
4148 * The resulting memory area is 32bit addressable and zeroed so it can be
4149 * mapped to userspace without leaking data.
4150 *
4151 * Return: pointer to the allocated memory or %NULL on error
4152 */
4153void *vmalloc_32_user_noprof(unsigned long size)
4154{
4155 return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END,
4156 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
4157 VM_USERMAP, NUMA_NO_NODE,
4158 __builtin_return_address(0));
4159}
4160EXPORT_SYMBOL(vmalloc_32_user_noprof);
4161
4162/*
4163 * Atomically zero bytes in the iterator.
4164 *
4165 * Returns the number of zeroed bytes.
4166 */
4167static size_t zero_iter(struct iov_iter *iter, size_t count)
4168{
4169 size_t remains = count;
4170
4171 while (remains > 0) {
4172 size_t num, copied;
4173
4174 num = min_t(size_t, remains, PAGE_SIZE);
4175 copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
4176 remains -= copied;
4177
4178 if (copied < num)
4179 break;
4180 }
4181
4182 return count - remains;
4183}
4184
4185/*
4186 * small helper routine, copy contents to iter from addr.
4187 * If the page is not present, fill zero.
4188 *
4189 * Returns the number of copied bytes.
4190 */
4191static size_t aligned_vread_iter(struct iov_iter *iter,
4192 const char *addr, size_t count)
4193{
4194 size_t remains = count;
4195 struct page *page;
4196
4197 while (remains > 0) {
4198 unsigned long offset, length;
4199 size_t copied = 0;
4200
4201 offset = offset_in_page(addr);
4202 length = PAGE_SIZE - offset;
4203 if (length > remains)
4204 length = remains;
4205 page = vmalloc_to_page(addr);
4206 /*
4207 * To do safe access to this _mapped_ area, we need lock. But
4208 * adding lock here means that we need to add overhead of
4209 * vmalloc()/vfree() calls for this _debug_ interface, rarely
4210 * used. Instead of that, we'll use an local mapping via
4211 * copy_page_to_iter_nofault() and accept a small overhead in
4212 * this access function.
4213 */
4214 if (page)
4215 copied = copy_page_to_iter_nofault(page, offset,
4216 length, iter);
4217 else
4218 copied = zero_iter(iter, length);
4219
4220 addr += copied;
4221 remains -= copied;
4222
4223 if (copied != length)
4224 break;
4225 }
4226
4227 return count - remains;
4228}
4229
4230/*
4231 * Read from a vm_map_ram region of memory.
4232 *
4233 * Returns the number of copied bytes.
4234 */
4235static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
4236 size_t count, unsigned long flags)
4237{
4238 char *start;
4239 struct vmap_block *vb;
4240 struct xarray *xa;
4241 unsigned long offset;
4242 unsigned int rs, re;
4243 size_t remains, n;
4244
4245 /*
4246 * If it's area created by vm_map_ram() interface directly, but
4247 * not further subdividing and delegating management to vmap_block,
4248 * handle it here.
4249 */
4250 if (!(flags & VMAP_BLOCK))
4251 return aligned_vread_iter(iter, addr, count);
4252
4253 remains = count;
4254
4255 /*
4256 * Area is split into regions and tracked with vmap_block, read out
4257 * each region and zero fill the hole between regions.
4258 */
4259 xa = addr_to_vb_xa((unsigned long) addr);
4260 vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
4261 if (!vb)
4262 goto finished_zero;
4263
4264 spin_lock(&vb->lock);
4265 if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
4266 spin_unlock(&vb->lock);
4267 goto finished_zero;
4268 }
4269
4270 for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
4271 size_t copied;
4272
4273 if (remains == 0)
4274 goto finished;
4275
4276 start = vmap_block_vaddr(vb->va->va_start, rs);
4277
4278 if (addr < start) {
4279 size_t to_zero = min_t(size_t, start - addr, remains);
4280 size_t zeroed = zero_iter(iter, to_zero);
4281
4282 addr += zeroed;
4283 remains -= zeroed;
4284
4285 if (remains == 0 || zeroed != to_zero)
4286 goto finished;
4287 }
4288
4289 /*it could start reading from the middle of used region*/
4290 offset = offset_in_page(addr);
4291 n = ((re - rs + 1) << PAGE_SHIFT) - offset;
4292 if (n > remains)
4293 n = remains;
4294
4295 copied = aligned_vread_iter(iter, start + offset, n);
4296
4297 addr += copied;
4298 remains -= copied;
4299
4300 if (copied != n)
4301 goto finished;
4302 }
4303
4304 spin_unlock(&vb->lock);
4305
4306finished_zero:
4307 /* zero-fill the left dirty or free regions */
4308 return count - remains + zero_iter(iter, remains);
4309finished:
4310 /* We couldn't copy/zero everything */
4311 spin_unlock(&vb->lock);
4312 return count - remains;
4313}
4314
4315/**
4316 * vread_iter() - read vmalloc area in a safe way to an iterator.
4317 * @iter: the iterator to which data should be written.
4318 * @addr: vm address.
4319 * @count: number of bytes to be read.
4320 *
4321 * This function checks that addr is a valid vmalloc'ed area, and
4322 * copy data from that area to a given buffer. If the given memory range
4323 * of [addr...addr+count) includes some valid address, data is copied to
4324 * proper area of @buf. If there are memory holes, they'll be zero-filled.
4325 * IOREMAP area is treated as memory hole and no copy is done.
4326 *
4327 * If [addr...addr+count) doesn't includes any intersects with alive
4328 * vm_struct area, returns 0. @buf should be kernel's buffer.
4329 *
4330 * Note: In usual ops, vread() is never necessary because the caller
4331 * should know vmalloc() area is valid and can use memcpy().
4332 * This is for routines which have to access vmalloc area without
4333 * any information, as /proc/kcore.
4334 *
4335 * Return: number of bytes for which addr and buf should be increased
4336 * (same number as @count) or %0 if [addr...addr+count) doesn't
4337 * include any intersection with valid vmalloc area
4338 */
4339long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
4340{
4341 struct vmap_node *vn;
4342 struct vmap_area *va;
4343 struct vm_struct *vm;
4344 char *vaddr;
4345 size_t n, size, flags, remains;
4346 unsigned long next;
4347
4348 addr = kasan_reset_tag(addr);
4349
4350 /* Don't allow overflow */
4351 if ((unsigned long) addr + count < count)
4352 count = -(unsigned long) addr;
4353
4354 remains = count;
4355
4356 vn = find_vmap_area_exceed_addr_lock((unsigned long) addr, &va);
4357 if (!vn)
4358 goto finished_zero;
4359
4360 /* no intersects with alive vmap_area */
4361 if ((unsigned long)addr + remains <= va->va_start)
4362 goto finished_zero;
4363
4364 do {
4365 size_t copied;
4366
4367 if (remains == 0)
4368 goto finished;
4369
4370 vm = va->vm;
4371 flags = va->flags & VMAP_FLAGS_MASK;
4372 /*
4373 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
4374 * be set together with VMAP_RAM.
4375 */
4376 WARN_ON(flags == VMAP_BLOCK);
4377
4378 if (!vm && !flags)
4379 goto next_va;
4380
4381 if (vm && (vm->flags & VM_UNINITIALIZED))
4382 goto next_va;
4383
4384 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4385 smp_rmb();
4386
4387 vaddr = (char *) va->va_start;
4388 size = vm ? get_vm_area_size(vm) : va_size(va);
4389
4390 if (addr >= vaddr + size)
4391 goto next_va;
4392
4393 if (addr < vaddr) {
4394 size_t to_zero = min_t(size_t, vaddr - addr, remains);
4395 size_t zeroed = zero_iter(iter, to_zero);
4396
4397 addr += zeroed;
4398 remains -= zeroed;
4399
4400 if (remains == 0 || zeroed != to_zero)
4401 goto finished;
4402 }
4403
4404 n = vaddr + size - addr;
4405 if (n > remains)
4406 n = remains;
4407
4408 if (flags & VMAP_RAM)
4409 copied = vmap_ram_vread_iter(iter, addr, n, flags);
4410 else if (!(vm && (vm->flags & (VM_IOREMAP | VM_SPARSE))))
4411 copied = aligned_vread_iter(iter, addr, n);
4412 else /* IOREMAP | SPARSE area is treated as memory hole */
4413 copied = zero_iter(iter, n);
4414
4415 addr += copied;
4416 remains -= copied;
4417
4418 if (copied != n)
4419 goto finished;
4420
4421 next_va:
4422 next = va->va_end;
4423 spin_unlock(&vn->busy.lock);
4424 } while ((vn = find_vmap_area_exceed_addr_lock(next, &va)));
4425
4426finished_zero:
4427 if (vn)
4428 spin_unlock(&vn->busy.lock);
4429
4430 /* zero-fill memory holes */
4431 return count - remains + zero_iter(iter, remains);
4432finished:
4433 /* Nothing remains, or We couldn't copy/zero everything. */
4434 if (vn)
4435 spin_unlock(&vn->busy.lock);
4436
4437 return count - remains;
4438}
4439
4440/**
4441 * remap_vmalloc_range_partial - map vmalloc pages to userspace
4442 * @vma: vma to cover
4443 * @uaddr: target user address to start at
4444 * @kaddr: virtual address of vmalloc kernel memory
4445 * @pgoff: offset from @kaddr to start at
4446 * @size: size of map area
4447 *
4448 * Returns: 0 for success, -Exxx on failure
4449 *
4450 * This function checks that @kaddr is a valid vmalloc'ed area,
4451 * and that it is big enough to cover the range starting at
4452 * @uaddr in @vma. Will return failure if that criteria isn't
4453 * met.
4454 *
4455 * Similar to remap_pfn_range() (see mm/memory.c)
4456 */
4457int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
4458 void *kaddr, unsigned long pgoff,
4459 unsigned long size)
4460{
4461 struct vm_struct *area;
4462 unsigned long off;
4463 unsigned long end_index;
4464
4465 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
4466 return -EINVAL;
4467
4468 size = PAGE_ALIGN(size);
4469
4470 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
4471 return -EINVAL;
4472
4473 area = find_vm_area(kaddr);
4474 if (!area)
4475 return -EINVAL;
4476
4477 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
4478 return -EINVAL;
4479
4480 if (check_add_overflow(size, off, &end_index) ||
4481 end_index > get_vm_area_size(area))
4482 return -EINVAL;
4483 kaddr += off;
4484
4485 do {
4486 struct page *page = vmalloc_to_page(kaddr);
4487 int ret;
4488
4489 ret = vm_insert_page(vma, uaddr, page);
4490 if (ret)
4491 return ret;
4492
4493 uaddr += PAGE_SIZE;
4494 kaddr += PAGE_SIZE;
4495 size -= PAGE_SIZE;
4496 } while (size > 0);
4497
4498 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
4499
4500 return 0;
4501}
4502
4503/**
4504 * remap_vmalloc_range - map vmalloc pages to userspace
4505 * @vma: vma to cover (map full range of vma)
4506 * @addr: vmalloc memory
4507 * @pgoff: number of pages into addr before first page to map
4508 *
4509 * Returns: 0 for success, -Exxx on failure
4510 *
4511 * This function checks that addr is a valid vmalloc'ed area, and
4512 * that it is big enough to cover the vma. Will return failure if
4513 * that criteria isn't met.
4514 *
4515 * Similar to remap_pfn_range() (see mm/memory.c)
4516 */
4517int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
4518 unsigned long pgoff)
4519{
4520 return remap_vmalloc_range_partial(vma, vma->vm_start,
4521 addr, pgoff,
4522 vma->vm_end - vma->vm_start);
4523}
4524EXPORT_SYMBOL(remap_vmalloc_range);
4525
4526void free_vm_area(struct vm_struct *area)
4527{
4528 struct vm_struct *ret;
4529 ret = remove_vm_area(area->addr);
4530 BUG_ON(ret != area);
4531 kfree(area);
4532}
4533EXPORT_SYMBOL_GPL(free_vm_area);
4534
4535#ifdef CONFIG_SMP
4536static struct vmap_area *node_to_va(struct rb_node *n)
4537{
4538 return rb_entry_safe(n, struct vmap_area, rb_node);
4539}
4540
4541/**
4542 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
4543 * @addr: target address
4544 *
4545 * Returns: vmap_area if it is found. If there is no such area
4546 * the first highest(reverse order) vmap_area is returned
4547 * i.e. va->va_start < addr && va->va_end < addr or NULL
4548 * if there are no any areas before @addr.
4549 */
4550static struct vmap_area *
4551pvm_find_va_enclose_addr(unsigned long addr)
4552{
4553 struct vmap_area *va, *tmp;
4554 struct rb_node *n;
4555
4556 n = free_vmap_area_root.rb_node;
4557 va = NULL;
4558
4559 while (n) {
4560 tmp = rb_entry(n, struct vmap_area, rb_node);
4561 if (tmp->va_start <= addr) {
4562 va = tmp;
4563 if (tmp->va_end >= addr)
4564 break;
4565
4566 n = n->rb_right;
4567 } else {
4568 n = n->rb_left;
4569 }
4570 }
4571
4572 return va;
4573}
4574
4575/**
4576 * pvm_determine_end_from_reverse - find the highest aligned address
4577 * of free block below VMALLOC_END
4578 * @va:
4579 * in - the VA we start the search(reverse order);
4580 * out - the VA with the highest aligned end address.
4581 * @align: alignment for required highest address
4582 *
4583 * Returns: determined end address within vmap_area
4584 */
4585static unsigned long
4586pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
4587{
4588 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4589 unsigned long addr;
4590
4591 if (likely(*va)) {
4592 list_for_each_entry_from_reverse((*va),
4593 &free_vmap_area_list, list) {
4594 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
4595 if ((*va)->va_start < addr)
4596 return addr;
4597 }
4598 }
4599
4600 return 0;
4601}
4602
4603/**
4604 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4605 * @offsets: array containing offset of each area
4606 * @sizes: array containing size of each area
4607 * @nr_vms: the number of areas to allocate
4608 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4609 *
4610 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4611 * vm_structs on success, %NULL on failure
4612 *
4613 * Percpu allocator wants to use congruent vm areas so that it can
4614 * maintain the offsets among percpu areas. This function allocates
4615 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
4616 * be scattered pretty far, distance between two areas easily going up
4617 * to gigabytes. To avoid interacting with regular vmallocs, these
4618 * areas are allocated from top.
4619 *
4620 * Despite its complicated look, this allocator is rather simple. It
4621 * does everything top-down and scans free blocks from the end looking
4622 * for matching base. While scanning, if any of the areas do not fit the
4623 * base address is pulled down to fit the area. Scanning is repeated till
4624 * all the areas fit and then all necessary data structures are inserted
4625 * and the result is returned.
4626 */
4627struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4628 const size_t *sizes, int nr_vms,
4629 size_t align)
4630{
4631 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4632 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4633 struct vmap_area **vas, *va;
4634 struct vm_struct **vms;
4635 int area, area2, last_area, term_area;
4636 unsigned long base, start, size, end, last_end, orig_start, orig_end;
4637 bool purged = false;
4638
4639 /* verify parameters and allocate data structures */
4640 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4641 for (last_area = 0, area = 0; area < nr_vms; area++) {
4642 start = offsets[area];
4643 end = start + sizes[area];
4644
4645 /* is everything aligned properly? */
4646 BUG_ON(!IS_ALIGNED(offsets[area], align));
4647 BUG_ON(!IS_ALIGNED(sizes[area], align));
4648
4649 /* detect the area with the highest address */
4650 if (start > offsets[last_area])
4651 last_area = area;
4652
4653 for (area2 = area + 1; area2 < nr_vms; area2++) {
4654 unsigned long start2 = offsets[area2];
4655 unsigned long end2 = start2 + sizes[area2];
4656
4657 BUG_ON(start2 < end && start < end2);
4658 }
4659 }
4660 last_end = offsets[last_area] + sizes[last_area];
4661
4662 if (vmalloc_end - vmalloc_start < last_end) {
4663 WARN_ON(true);
4664 return NULL;
4665 }
4666
4667 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4668 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4669 if (!vas || !vms)
4670 goto err_free2;
4671
4672 for (area = 0; area < nr_vms; area++) {
4673 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4674 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4675 if (!vas[area] || !vms[area])
4676 goto err_free;
4677 }
4678retry:
4679 spin_lock(&free_vmap_area_lock);
4680
4681 /* start scanning - we scan from the top, begin with the last area */
4682 area = term_area = last_area;
4683 start = offsets[area];
4684 end = start + sizes[area];
4685
4686 va = pvm_find_va_enclose_addr(vmalloc_end);
4687 base = pvm_determine_end_from_reverse(&va, align) - end;
4688
4689 while (true) {
4690 /*
4691 * base might have underflowed, add last_end before
4692 * comparing.
4693 */
4694 if (base + last_end < vmalloc_start + last_end)
4695 goto overflow;
4696
4697 /*
4698 * Fitting base has not been found.
4699 */
4700 if (va == NULL)
4701 goto overflow;
4702
4703 /*
4704 * If required width exceeds current VA block, move
4705 * base downwards and then recheck.
4706 */
4707 if (base + end > va->va_end) {
4708 base = pvm_determine_end_from_reverse(&va, align) - end;
4709 term_area = area;
4710 continue;
4711 }
4712
4713 /*
4714 * If this VA does not fit, move base downwards and recheck.
4715 */
4716 if (base + start < va->va_start) {
4717 va = node_to_va(rb_prev(&va->rb_node));
4718 base = pvm_determine_end_from_reverse(&va, align) - end;
4719 term_area = area;
4720 continue;
4721 }
4722
4723 /*
4724 * This area fits, move on to the previous one. If
4725 * the previous one is the terminal one, we're done.
4726 */
4727 area = (area + nr_vms - 1) % nr_vms;
4728 if (area == term_area)
4729 break;
4730
4731 start = offsets[area];
4732 end = start + sizes[area];
4733 va = pvm_find_va_enclose_addr(base + end);
4734 }
4735
4736 /* we've found a fitting base, insert all va's */
4737 for (area = 0; area < nr_vms; area++) {
4738 int ret;
4739
4740 start = base + offsets[area];
4741 size = sizes[area];
4742
4743 va = pvm_find_va_enclose_addr(start);
4744 if (WARN_ON_ONCE(va == NULL))
4745 /* It is a BUG(), but trigger recovery instead. */
4746 goto recovery;
4747
4748 ret = va_clip(&free_vmap_area_root,
4749 &free_vmap_area_list, va, start, size);
4750 if (WARN_ON_ONCE(unlikely(ret)))
4751 /* It is a BUG(), but trigger recovery instead. */
4752 goto recovery;
4753
4754 /* Allocated area. */
4755 va = vas[area];
4756 va->va_start = start;
4757 va->va_end = start + size;
4758 }
4759
4760 spin_unlock(&free_vmap_area_lock);
4761
4762 /* populate the kasan shadow space */
4763 for (area = 0; area < nr_vms; area++) {
4764 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4765 goto err_free_shadow;
4766 }
4767
4768 /* insert all vm's */
4769 for (area = 0; area < nr_vms; area++) {
4770 struct vmap_node *vn = addr_to_node(vas[area]->va_start);
4771
4772 spin_lock(&vn->busy.lock);
4773 insert_vmap_area(vas[area], &vn->busy.root, &vn->busy.head);
4774 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
4775 pcpu_get_vm_areas);
4776 spin_unlock(&vn->busy.lock);
4777 }
4778
4779 /*
4780 * Mark allocated areas as accessible. Do it now as a best-effort
4781 * approach, as they can be mapped outside of vmalloc code.
4782 * With hardware tag-based KASAN, marking is skipped for
4783 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4784 */
4785 for (area = 0; area < nr_vms; area++)
4786 vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4787 vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4788
4789 kfree(vas);
4790 return vms;
4791
4792recovery:
4793 /*
4794 * Remove previously allocated areas. There is no
4795 * need in removing these areas from the busy tree,
4796 * because they are inserted only on the final step
4797 * and when pcpu_get_vm_areas() is success.
4798 */
4799 while (area--) {
4800 orig_start = vas[area]->va_start;
4801 orig_end = vas[area]->va_end;
4802 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4803 &free_vmap_area_list);
4804 if (va)
4805 kasan_release_vmalloc(orig_start, orig_end,
4806 va->va_start, va->va_end,
4807 KASAN_VMALLOC_PAGE_RANGE | KASAN_VMALLOC_TLB_FLUSH);
4808 vas[area] = NULL;
4809 }
4810
4811overflow:
4812 spin_unlock(&free_vmap_area_lock);
4813 if (!purged) {
4814 reclaim_and_purge_vmap_areas();
4815 purged = true;
4816
4817 /* Before "retry", check if we recover. */
4818 for (area = 0; area < nr_vms; area++) {
4819 if (vas[area])
4820 continue;
4821
4822 vas[area] = kmem_cache_zalloc(
4823 vmap_area_cachep, GFP_KERNEL);
4824 if (!vas[area])
4825 goto err_free;
4826 }
4827
4828 goto retry;
4829 }
4830
4831err_free:
4832 for (area = 0; area < nr_vms; area++) {
4833 if (vas[area])
4834 kmem_cache_free(vmap_area_cachep, vas[area]);
4835
4836 kfree(vms[area]);
4837 }
4838err_free2:
4839 kfree(vas);
4840 kfree(vms);
4841 return NULL;
4842
4843err_free_shadow:
4844 spin_lock(&free_vmap_area_lock);
4845 /*
4846 * We release all the vmalloc shadows, even the ones for regions that
4847 * hadn't been successfully added. This relies on kasan_release_vmalloc
4848 * being able to tolerate this case.
4849 */
4850 for (area = 0; area < nr_vms; area++) {
4851 orig_start = vas[area]->va_start;
4852 orig_end = vas[area]->va_end;
4853 va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4854 &free_vmap_area_list);
4855 if (va)
4856 kasan_release_vmalloc(orig_start, orig_end,
4857 va->va_start, va->va_end,
4858 KASAN_VMALLOC_PAGE_RANGE | KASAN_VMALLOC_TLB_FLUSH);
4859 vas[area] = NULL;
4860 kfree(vms[area]);
4861 }
4862 spin_unlock(&free_vmap_area_lock);
4863 kfree(vas);
4864 kfree(vms);
4865 return NULL;
4866}
4867
4868/**
4869 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4870 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4871 * @nr_vms: the number of allocated areas
4872 *
4873 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4874 */
4875void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4876{
4877 int i;
4878
4879 for (i = 0; i < nr_vms; i++)
4880 free_vm_area(vms[i]);
4881 kfree(vms);
4882}
4883#endif /* CONFIG_SMP */
4884
4885#ifdef CONFIG_PRINTK
4886bool vmalloc_dump_obj(void *object)
4887{
4888 const void *caller;
4889 struct vm_struct *vm;
4890 struct vmap_area *va;
4891 struct vmap_node *vn;
4892 unsigned long addr;
4893 unsigned int nr_pages;
4894
4895 addr = PAGE_ALIGN((unsigned long) object);
4896 vn = addr_to_node(addr);
4897
4898 if (!spin_trylock(&vn->busy.lock))
4899 return false;
4900
4901 va = __find_vmap_area(addr, &vn->busy.root);
4902 if (!va || !va->vm) {
4903 spin_unlock(&vn->busy.lock);
4904 return false;
4905 }
4906
4907 vm = va->vm;
4908 addr = (unsigned long) vm->addr;
4909 caller = vm->caller;
4910 nr_pages = vm->nr_pages;
4911 spin_unlock(&vn->busy.lock);
4912
4913 pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4914 nr_pages, addr, caller);
4915
4916 return true;
4917}
4918#endif
4919
4920#ifdef CONFIG_PROC_FS
4921static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4922{
4923 if (IS_ENABLED(CONFIG_NUMA)) {
4924 unsigned int nr, *counters = m->private;
4925 unsigned int step = 1U << vm_area_page_order(v);
4926
4927 if (!counters)
4928 return;
4929
4930 if (v->flags & VM_UNINITIALIZED)
4931 return;
4932 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4933 smp_rmb();
4934
4935 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4936
4937 for (nr = 0; nr < v->nr_pages; nr += step)
4938 counters[page_to_nid(v->pages[nr])] += step;
4939 for_each_node_state(nr, N_HIGH_MEMORY)
4940 if (counters[nr])
4941 seq_printf(m, " N%u=%u", nr, counters[nr]);
4942 }
4943}
4944
4945static void show_purge_info(struct seq_file *m)
4946{
4947 struct vmap_node *vn;
4948 struct vmap_area *va;
4949 int i;
4950
4951 for (i = 0; i < nr_vmap_nodes; i++) {
4952 vn = &vmap_nodes[i];
4953
4954 spin_lock(&vn->lazy.lock);
4955 list_for_each_entry(va, &vn->lazy.head, list) {
4956 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4957 (void *)va->va_start, (void *)va->va_end,
4958 va_size(va));
4959 }
4960 spin_unlock(&vn->lazy.lock);
4961 }
4962}
4963
4964static int vmalloc_info_show(struct seq_file *m, void *p)
4965{
4966 struct vmap_node *vn;
4967 struct vmap_area *va;
4968 struct vm_struct *v;
4969 int i;
4970
4971 for (i = 0; i < nr_vmap_nodes; i++) {
4972 vn = &vmap_nodes[i];
4973
4974 spin_lock(&vn->busy.lock);
4975 list_for_each_entry(va, &vn->busy.head, list) {
4976 if (!va->vm) {
4977 if (va->flags & VMAP_RAM)
4978 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4979 (void *)va->va_start, (void *)va->va_end,
4980 va_size(va));
4981
4982 continue;
4983 }
4984
4985 v = va->vm;
4986
4987 seq_printf(m, "0x%pK-0x%pK %7ld",
4988 v->addr, v->addr + v->size, v->size);
4989
4990 if (v->caller)
4991 seq_printf(m, " %pS", v->caller);
4992
4993 if (v->nr_pages)
4994 seq_printf(m, " pages=%d", v->nr_pages);
4995
4996 if (v->phys_addr)
4997 seq_printf(m, " phys=%pa", &v->phys_addr);
4998
4999 if (v->flags & VM_IOREMAP)
5000 seq_puts(m, " ioremap");
5001
5002 if (v->flags & VM_SPARSE)
5003 seq_puts(m, " sparse");
5004
5005 if (v->flags & VM_ALLOC)
5006 seq_puts(m, " vmalloc");
5007
5008 if (v->flags & VM_MAP)
5009 seq_puts(m, " vmap");
5010
5011 if (v->flags & VM_USERMAP)
5012 seq_puts(m, " user");
5013
5014 if (v->flags & VM_DMA_COHERENT)
5015 seq_puts(m, " dma-coherent");
5016
5017 if (is_vmalloc_addr(v->pages))
5018 seq_puts(m, " vpages");
5019
5020 show_numa_info(m, v);
5021 seq_putc(m, '\n');
5022 }
5023 spin_unlock(&vn->busy.lock);
5024 }
5025
5026 /*
5027 * As a final step, dump "unpurged" areas.
5028 */
5029 show_purge_info(m);
5030 return 0;
5031}
5032
5033static int __init proc_vmalloc_init(void)
5034{
5035 void *priv_data = NULL;
5036
5037 if (IS_ENABLED(CONFIG_NUMA))
5038 priv_data = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
5039
5040 proc_create_single_data("vmallocinfo",
5041 0400, NULL, vmalloc_info_show, priv_data);
5042
5043 return 0;
5044}
5045module_init(proc_vmalloc_init);
5046
5047#endif
5048
5049static void __init vmap_init_free_space(void)
5050{
5051 unsigned long vmap_start = 1;
5052 const unsigned long vmap_end = ULONG_MAX;
5053 struct vmap_area *free;
5054 struct vm_struct *busy;
5055
5056 /*
5057 * B F B B B F
5058 * -|-----|.....|-----|-----|-----|.....|-
5059 * | The KVA space |
5060 * |<--------------------------------->|
5061 */
5062 for (busy = vmlist; busy; busy = busy->next) {
5063 if ((unsigned long) busy->addr - vmap_start > 0) {
5064 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5065 if (!WARN_ON_ONCE(!free)) {
5066 free->va_start = vmap_start;
5067 free->va_end = (unsigned long) busy->addr;
5068
5069 insert_vmap_area_augment(free, NULL,
5070 &free_vmap_area_root,
5071 &free_vmap_area_list);
5072 }
5073 }
5074
5075 vmap_start = (unsigned long) busy->addr + busy->size;
5076 }
5077
5078 if (vmap_end - vmap_start > 0) {
5079 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5080 if (!WARN_ON_ONCE(!free)) {
5081 free->va_start = vmap_start;
5082 free->va_end = vmap_end;
5083
5084 insert_vmap_area_augment(free, NULL,
5085 &free_vmap_area_root,
5086 &free_vmap_area_list);
5087 }
5088 }
5089}
5090
5091static void vmap_init_nodes(void)
5092{
5093 struct vmap_node *vn;
5094 int i, n;
5095
5096#if BITS_PER_LONG == 64
5097 /*
5098 * A high threshold of max nodes is fixed and bound to 128,
5099 * thus a scale factor is 1 for systems where number of cores
5100 * are less or equal to specified threshold.
5101 *
5102 * As for NUMA-aware notes. For bigger systems, for example
5103 * NUMA with multi-sockets, where we can end-up with thousands
5104 * of cores in total, a "sub-numa-clustering" should be added.
5105 *
5106 * In this case a NUMA domain is considered as a single entity
5107 * with dedicated sub-nodes in it which describe one group or
5108 * set of cores. Therefore a per-domain purging is supposed to
5109 * be added as well as a per-domain balancing.
5110 */
5111 n = clamp_t(unsigned int, num_possible_cpus(), 1, 128);
5112
5113 if (n > 1) {
5114 vn = kmalloc_array(n, sizeof(*vn), GFP_NOWAIT | __GFP_NOWARN);
5115 if (vn) {
5116 /* Node partition is 16 pages. */
5117 vmap_zone_size = (1 << 4) * PAGE_SIZE;
5118 nr_vmap_nodes = n;
5119 vmap_nodes = vn;
5120 } else {
5121 pr_err("Failed to allocate an array. Disable a node layer\n");
5122 }
5123 }
5124#endif
5125
5126 for (n = 0; n < nr_vmap_nodes; n++) {
5127 vn = &vmap_nodes[n];
5128 vn->busy.root = RB_ROOT;
5129 INIT_LIST_HEAD(&vn->busy.head);
5130 spin_lock_init(&vn->busy.lock);
5131
5132 vn->lazy.root = RB_ROOT;
5133 INIT_LIST_HEAD(&vn->lazy.head);
5134 spin_lock_init(&vn->lazy.lock);
5135
5136 for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
5137 INIT_LIST_HEAD(&vn->pool[i].head);
5138 WRITE_ONCE(vn->pool[i].len, 0);
5139 }
5140
5141 spin_lock_init(&vn->pool_lock);
5142 }
5143}
5144
5145static unsigned long
5146vmap_node_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5147{
5148 unsigned long count;
5149 struct vmap_node *vn;
5150 int i, j;
5151
5152 for (count = 0, i = 0; i < nr_vmap_nodes; i++) {
5153 vn = &vmap_nodes[i];
5154
5155 for (j = 0; j < MAX_VA_SIZE_PAGES; j++)
5156 count += READ_ONCE(vn->pool[j].len);
5157 }
5158
5159 return count ? count : SHRINK_EMPTY;
5160}
5161
5162static unsigned long
5163vmap_node_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5164{
5165 int i;
5166
5167 for (i = 0; i < nr_vmap_nodes; i++)
5168 decay_va_pool_node(&vmap_nodes[i], true);
5169
5170 return SHRINK_STOP;
5171}
5172
5173void __init vmalloc_init(void)
5174{
5175 struct shrinker *vmap_node_shrinker;
5176 struct vmap_area *va;
5177 struct vmap_node *vn;
5178 struct vm_struct *tmp;
5179 int i;
5180
5181 /*
5182 * Create the cache for vmap_area objects.
5183 */
5184 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
5185
5186 for_each_possible_cpu(i) {
5187 struct vmap_block_queue *vbq;
5188 struct vfree_deferred *p;
5189
5190 vbq = &per_cpu(vmap_block_queue, i);
5191 spin_lock_init(&vbq->lock);
5192 INIT_LIST_HEAD(&vbq->free);
5193 p = &per_cpu(vfree_deferred, i);
5194 init_llist_head(&p->list);
5195 INIT_WORK(&p->wq, delayed_vfree_work);
5196 xa_init(&vbq->vmap_blocks);
5197 }
5198
5199 /*
5200 * Setup nodes before importing vmlist.
5201 */
5202 vmap_init_nodes();
5203
5204 /* Import existing vmlist entries. */
5205 for (tmp = vmlist; tmp; tmp = tmp->next) {
5206 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5207 if (WARN_ON_ONCE(!va))
5208 continue;
5209
5210 va->va_start = (unsigned long)tmp->addr;
5211 va->va_end = va->va_start + tmp->size;
5212 va->vm = tmp;
5213
5214 vn = addr_to_node(va->va_start);
5215 insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
5216 }
5217
5218 /*
5219 * Now we can initialize a free vmap space.
5220 */
5221 vmap_init_free_space();
5222 vmap_initialized = true;
5223
5224 vmap_node_shrinker = shrinker_alloc(0, "vmap-node");
5225 if (!vmap_node_shrinker) {
5226 pr_err("Failed to allocate vmap-node shrinker!\n");
5227 return;
5228 }
5229
5230 vmap_node_shrinker->count_objects = vmap_node_shrink_count;
5231 vmap_node_shrinker->scan_objects = vmap_node_shrink_scan;
5232 shrinker_register(vmap_node_shrinker);
5233}