Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
1da177e4 LT |
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 | |
930fc45a | 7 | * Numa awareness, Christoph Lameter, SGI, June 2005 |
d758ffe6 | 8 | * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019 |
1da177e4 LT |
9 | */ |
10 | ||
db64fe02 | 11 | #include <linux/vmalloc.h> |
1da177e4 LT |
12 | #include <linux/mm.h> |
13 | #include <linux/module.h> | |
14 | #include <linux/highmem.h> | |
c3edc401 | 15 | #include <linux/sched/signal.h> |
1da177e4 LT |
16 | #include <linux/slab.h> |
17 | #include <linux/spinlock.h> | |
18 | #include <linux/interrupt.h> | |
5f6a6a9c | 19 | #include <linux/proc_fs.h> |
a10aa579 | 20 | #include <linux/seq_file.h> |
868b104d | 21 | #include <linux/set_memory.h> |
3ac7fe5a | 22 | #include <linux/debugobjects.h> |
23016969 | 23 | #include <linux/kallsyms.h> |
db64fe02 | 24 | #include <linux/list.h> |
4da56b99 | 25 | #include <linux/notifier.h> |
db64fe02 | 26 | #include <linux/rbtree.h> |
0f14599c | 27 | #include <linux/xarray.h> |
5da96bdd | 28 | #include <linux/io.h> |
db64fe02 | 29 | #include <linux/rcupdate.h> |
f0aa6617 | 30 | #include <linux/pfn.h> |
89219d37 | 31 | #include <linux/kmemleak.h> |
60063497 | 32 | #include <linux/atomic.h> |
3b32123d | 33 | #include <linux/compiler.h> |
4e5aa1f4 | 34 | #include <linux/memcontrol.h> |
32fcfd40 | 35 | #include <linux/llist.h> |
4c91c07c | 36 | #include <linux/uio.h> |
0f616be1 | 37 | #include <linux/bitops.h> |
68ad4a33 | 38 | #include <linux/rbtree_augmented.h> |
bdebd6a2 | 39 | #include <linux/overflow.h> |
c0eb315a | 40 | #include <linux/pgtable.h> |
f7ee1f13 | 41 | #include <linux/hugetlb.h> |
451769eb | 42 | #include <linux/sched/mm.h> |
1da177e4 | 43 | #include <asm/tlbflush.h> |
2dca6999 | 44 | #include <asm/shmparam.h> |
21e516b9 | 45 | #include <linux/page_owner.h> |
1da177e4 | 46 | |
cf243da6 URS |
47 | #define CREATE_TRACE_POINTS |
48 | #include <trace/events/vmalloc.h> | |
49 | ||
dd56b046 | 50 | #include "internal.h" |
2a681cfa | 51 | #include "pgalloc-track.h" |
dd56b046 | 52 | |
82a70ce0 CH |
53 | #ifdef CONFIG_HAVE_ARCH_HUGE_VMAP |
54 | static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1; | |
55 | ||
56 | static int __init set_nohugeiomap(char *str) | |
57 | { | |
58 | ioremap_max_page_shift = PAGE_SHIFT; | |
59 | return 0; | |
60 | } | |
61 | early_param("nohugeiomap", set_nohugeiomap); | |
62 | #else /* CONFIG_HAVE_ARCH_HUGE_VMAP */ | |
63 | static const unsigned int ioremap_max_page_shift = PAGE_SHIFT; | |
64 | #endif /* CONFIG_HAVE_ARCH_HUGE_VMAP */ | |
65 | ||
121e6f32 NP |
66 | #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC |
67 | static bool __ro_after_init vmap_allow_huge = true; | |
68 | ||
69 | static int __init set_nohugevmalloc(char *str) | |
70 | { | |
71 | vmap_allow_huge = false; | |
72 | return 0; | |
73 | } | |
74 | early_param("nohugevmalloc", set_nohugevmalloc); | |
75 | #else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */ | |
76 | static const bool vmap_allow_huge = false; | |
77 | #endif /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */ | |
78 | ||
186525bd IM |
79 | bool is_vmalloc_addr(const void *x) |
80 | { | |
4aff1dc4 | 81 | unsigned long addr = (unsigned long)kasan_reset_tag(x); |
186525bd IM |
82 | |
83 | return addr >= VMALLOC_START && addr < VMALLOC_END; | |
84 | } | |
85 | EXPORT_SYMBOL(is_vmalloc_addr); | |
86 | ||
32fcfd40 AV |
87 | struct vfree_deferred { |
88 | struct llist_head list; | |
89 | struct work_struct wq; | |
90 | }; | |
91 | static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred); | |
92 | ||
db64fe02 | 93 | /*** Page table manipulation functions ***/ |
5e9e3d77 NP |
94 | static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end, |
95 | phys_addr_t phys_addr, pgprot_t prot, | |
f7ee1f13 | 96 | unsigned int max_page_shift, pgtbl_mod_mask *mask) |
5e9e3d77 NP |
97 | { |
98 | pte_t *pte; | |
99 | u64 pfn; | |
21e516b9 | 100 | struct page *page; |
f7ee1f13 | 101 | unsigned long size = PAGE_SIZE; |
5e9e3d77 NP |
102 | |
103 | pfn = phys_addr >> PAGE_SHIFT; | |
104 | pte = pte_alloc_kernel_track(pmd, addr, mask); | |
105 | if (!pte) | |
106 | return -ENOMEM; | |
107 | do { | |
21e516b9 HP |
108 | if (!pte_none(ptep_get(pte))) { |
109 | if (pfn_valid(pfn)) { | |
110 | page = pfn_to_page(pfn); | |
111 | dump_page(page, "remapping already mapped page"); | |
112 | } | |
113 | BUG(); | |
114 | } | |
f7ee1f13 CL |
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 | ||
f7ee1f13 | 121 | entry = arch_make_huge_pte(entry, ilog2(size), 0); |
935d4f0c | 122 | set_huge_pte_at(&init_mm, addr, pte, entry, size); |
f7ee1f13 CL |
123 | pfn += PFN_DOWN(size); |
124 | continue; | |
125 | } | |
126 | #endif | |
5e9e3d77 NP |
127 | set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot)); |
128 | pfn++; | |
f7ee1f13 | 129 | } while (pte += PFN_DOWN(size), addr += size, addr != end); |
5e9e3d77 NP |
130 | *mask |= PGTBL_PTE_MODIFIED; |
131 | return 0; | |
132 | } | |
133 | ||
134 | static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end, | |
135 | phys_addr_t phys_addr, pgprot_t prot, | |
136 | unsigned int max_page_shift) | |
137 | { | |
138 | if (max_page_shift < PMD_SHIFT) | |
139 | return 0; | |
140 | ||
141 | if (!arch_vmap_pmd_supported(prot)) | |
142 | return 0; | |
143 | ||
144 | if ((end - addr) != PMD_SIZE) | |
145 | return 0; | |
146 | ||
147 | if (!IS_ALIGNED(addr, PMD_SIZE)) | |
148 | return 0; | |
149 | ||
150 | if (!IS_ALIGNED(phys_addr, PMD_SIZE)) | |
151 | return 0; | |
152 | ||
153 | if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr)) | |
154 | return 0; | |
155 | ||
156 | return pmd_set_huge(pmd, phys_addr, prot); | |
157 | } | |
158 | ||
159 | static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end, | |
160 | phys_addr_t phys_addr, pgprot_t prot, | |
161 | unsigned int max_page_shift, pgtbl_mod_mask *mask) | |
162 | { | |
163 | pmd_t *pmd; | |
164 | unsigned long next; | |
165 | ||
166 | pmd = pmd_alloc_track(&init_mm, pud, addr, mask); | |
167 | if (!pmd) | |
168 | return -ENOMEM; | |
169 | do { | |
170 | next = pmd_addr_end(addr, end); | |
171 | ||
172 | if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot, | |
173 | max_page_shift)) { | |
174 | *mask |= PGTBL_PMD_MODIFIED; | |
175 | continue; | |
176 | } | |
177 | ||
f7ee1f13 | 178 | if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask)) |
5e9e3d77 NP |
179 | return -ENOMEM; |
180 | } while (pmd++, phys_addr += (next - addr), addr = next, addr != end); | |
181 | return 0; | |
182 | } | |
183 | ||
184 | static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end, | |
185 | phys_addr_t phys_addr, pgprot_t prot, | |
186 | unsigned int max_page_shift) | |
187 | { | |
188 | if (max_page_shift < PUD_SHIFT) | |
189 | return 0; | |
190 | ||
191 | if (!arch_vmap_pud_supported(prot)) | |
192 | return 0; | |
193 | ||
194 | if ((end - addr) != PUD_SIZE) | |
195 | return 0; | |
196 | ||
197 | if (!IS_ALIGNED(addr, PUD_SIZE)) | |
198 | return 0; | |
199 | ||
200 | if (!IS_ALIGNED(phys_addr, PUD_SIZE)) | |
201 | return 0; | |
202 | ||
203 | if (pud_present(*pud) && !pud_free_pmd_page(pud, addr)) | |
204 | return 0; | |
205 | ||
206 | return pud_set_huge(pud, phys_addr, prot); | |
207 | } | |
208 | ||
209 | static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end, | |
210 | phys_addr_t phys_addr, pgprot_t prot, | |
211 | unsigned int max_page_shift, pgtbl_mod_mask *mask) | |
212 | { | |
213 | pud_t *pud; | |
214 | unsigned long next; | |
215 | ||
216 | pud = pud_alloc_track(&init_mm, p4d, addr, mask); | |
217 | if (!pud) | |
218 | return -ENOMEM; | |
219 | do { | |
220 | next = pud_addr_end(addr, end); | |
221 | ||
222 | if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot, | |
223 | max_page_shift)) { | |
224 | *mask |= PGTBL_PUD_MODIFIED; | |
225 | continue; | |
226 | } | |
227 | ||
228 | if (vmap_pmd_range(pud, addr, next, phys_addr, prot, | |
229 | max_page_shift, mask)) | |
230 | return -ENOMEM; | |
231 | } while (pud++, phys_addr += (next - addr), addr = next, addr != end); | |
232 | return 0; | |
233 | } | |
234 | ||
235 | static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end, | |
236 | phys_addr_t phys_addr, pgprot_t prot, | |
237 | unsigned int max_page_shift) | |
238 | { | |
239 | if (max_page_shift < P4D_SHIFT) | |
240 | return 0; | |
241 | ||
242 | if (!arch_vmap_p4d_supported(prot)) | |
243 | return 0; | |
244 | ||
245 | if ((end - addr) != P4D_SIZE) | |
246 | return 0; | |
247 | ||
248 | if (!IS_ALIGNED(addr, P4D_SIZE)) | |
249 | return 0; | |
250 | ||
251 | if (!IS_ALIGNED(phys_addr, P4D_SIZE)) | |
252 | return 0; | |
253 | ||
254 | if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr)) | |
255 | return 0; | |
256 | ||
257 | return p4d_set_huge(p4d, phys_addr, prot); | |
258 | } | |
259 | ||
260 | static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end, | |
261 | phys_addr_t phys_addr, pgprot_t prot, | |
262 | unsigned int max_page_shift, pgtbl_mod_mask *mask) | |
263 | { | |
264 | p4d_t *p4d; | |
265 | unsigned long next; | |
266 | ||
267 | p4d = p4d_alloc_track(&init_mm, pgd, addr, mask); | |
268 | if (!p4d) | |
269 | return -ENOMEM; | |
270 | do { | |
271 | next = p4d_addr_end(addr, end); | |
272 | ||
273 | if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot, | |
274 | max_page_shift)) { | |
275 | *mask |= PGTBL_P4D_MODIFIED; | |
276 | continue; | |
277 | } | |
278 | ||
279 | if (vmap_pud_range(p4d, addr, next, phys_addr, prot, | |
280 | max_page_shift, mask)) | |
281 | return -ENOMEM; | |
282 | } while (p4d++, phys_addr += (next - addr), addr = next, addr != end); | |
283 | return 0; | |
284 | } | |
285 | ||
5d87510d | 286 | static int vmap_range_noflush(unsigned long addr, unsigned long end, |
5e9e3d77 NP |
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 | ||
5e9e3d77 NP |
309 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) |
310 | arch_sync_kernel_mappings(start, end); | |
311 | ||
312 | return err; | |
313 | } | |
b221385b | 314 | |
d7bca919 AS |
315 | int vmap_page_range(unsigned long addr, unsigned long end, |
316 | phys_addr_t phys_addr, pgprot_t prot) | |
317 | { | |
318 | int err; | |
319 | ||
320 | err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot), | |
321 | ioremap_max_page_shift); | |
322 | flush_cache_vmap(addr, end); | |
323 | if (!err) | |
324 | err = kmsan_ioremap_page_range(addr, end, phys_addr, prot, | |
325 | ioremap_max_page_shift); | |
326 | return err; | |
327 | } | |
328 | ||
82a70ce0 CH |
329 | int ioremap_page_range(unsigned long addr, unsigned long end, |
330 | phys_addr_t phys_addr, pgprot_t prot) | |
5d87510d | 331 | { |
3e49a866 | 332 | struct vm_struct *area; |
5d87510d | 333 | |
3e49a866 AS |
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 | } | |
d7bca919 | 346 | return vmap_page_range(addr, end, phys_addr, prot); |
5d87510d NP |
347 | } |
348 | ||
2ba3e694 JR |
349 | static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end, |
350 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
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); | |
2ba3e694 | 359 | *mask |= PGTBL_PTE_MODIFIED; |
1da177e4 LT |
360 | } |
361 | ||
2ba3e694 JR |
362 | static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end, |
363 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
364 | { |
365 | pmd_t *pmd; | |
366 | unsigned long next; | |
2ba3e694 | 367 | int cleared; |
1da177e4 LT |
368 | |
369 | pmd = pmd_offset(pud, addr); | |
370 | do { | |
371 | next = pmd_addr_end(addr, end); | |
2ba3e694 JR |
372 | |
373 | cleared = pmd_clear_huge(pmd); | |
374 | if (cleared || pmd_bad(*pmd)) | |
375 | *mask |= PGTBL_PMD_MODIFIED; | |
376 | ||
377 | if (cleared) | |
b9820d8f | 378 | continue; |
1da177e4 LT |
379 | if (pmd_none_or_clear_bad(pmd)) |
380 | continue; | |
2ba3e694 | 381 | vunmap_pte_range(pmd, addr, next, mask); |
e47110e9 AK |
382 | |
383 | cond_resched(); | |
1da177e4 LT |
384 | } while (pmd++, addr = next, addr != end); |
385 | } | |
386 | ||
2ba3e694 JR |
387 | static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end, |
388 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
389 | { |
390 | pud_t *pud; | |
391 | unsigned long next; | |
2ba3e694 | 392 | int cleared; |
1da177e4 | 393 | |
c2febafc | 394 | pud = pud_offset(p4d, addr); |
1da177e4 LT |
395 | do { |
396 | next = pud_addr_end(addr, end); | |
2ba3e694 JR |
397 | |
398 | cleared = pud_clear_huge(pud); | |
399 | if (cleared || pud_bad(*pud)) | |
400 | *mask |= PGTBL_PUD_MODIFIED; | |
401 | ||
402 | if (cleared) | |
b9820d8f | 403 | continue; |
1da177e4 LT |
404 | if (pud_none_or_clear_bad(pud)) |
405 | continue; | |
2ba3e694 | 406 | vunmap_pmd_range(pud, addr, next, mask); |
1da177e4 LT |
407 | } while (pud++, addr = next, addr != end); |
408 | } | |
409 | ||
2ba3e694 JR |
410 | static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end, |
411 | pgtbl_mod_mask *mask) | |
c2febafc KS |
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); | |
2ba3e694 | 419 | |
c8db8c26 L |
420 | p4d_clear_huge(p4d); |
421 | if (p4d_bad(*p4d)) | |
2ba3e694 JR |
422 | *mask |= PGTBL_P4D_MODIFIED; |
423 | ||
c2febafc KS |
424 | if (p4d_none_or_clear_bad(p4d)) |
425 | continue; | |
2ba3e694 | 426 | vunmap_pud_range(p4d, addr, next, mask); |
c2febafc KS |
427 | } while (p4d++, addr = next, addr != end); |
428 | } | |
429 | ||
4ad0ae8c NP |
430 | /* |
431 | * vunmap_range_noflush is similar to vunmap_range, but does not | |
432 | * flush caches or TLBs. | |
b521c43f | 433 | * |
4ad0ae8c NP |
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). | |
b521c43f | 439 | * |
4ad0ae8c | 440 | * This is an internal function only. Do not use outside mm/. |
b521c43f | 441 | */ |
b073d7f8 | 442 | void __vunmap_range_noflush(unsigned long start, unsigned long end) |
1da177e4 | 443 | { |
1da177e4 | 444 | unsigned long next; |
b521c43f | 445 | pgd_t *pgd; |
2ba3e694 JR |
446 | unsigned long addr = start; |
447 | pgtbl_mod_mask mask = 0; | |
1da177e4 LT |
448 | |
449 | BUG_ON(addr >= end); | |
450 | pgd = pgd_offset_k(addr); | |
1da177e4 LT |
451 | do { |
452 | next = pgd_addr_end(addr, end); | |
2ba3e694 JR |
453 | if (pgd_bad(*pgd)) |
454 | mask |= PGTBL_PGD_MODIFIED; | |
1da177e4 LT |
455 | if (pgd_none_or_clear_bad(pgd)) |
456 | continue; | |
2ba3e694 | 457 | vunmap_p4d_range(pgd, addr, next, &mask); |
1da177e4 | 458 | } while (pgd++, addr = next, addr != end); |
2ba3e694 JR |
459 | |
460 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) | |
461 | arch_sync_kernel_mappings(start, end); | |
1da177e4 LT |
462 | } |
463 | ||
b073d7f8 AP |
464 | void vunmap_range_noflush(unsigned long start, unsigned long end) |
465 | { | |
466 | kmsan_vunmap_range_noflush(start, end); | |
467 | __vunmap_range_noflush(start, end); | |
468 | } | |
469 | ||
4ad0ae8c NP |
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 | */ | |
479 | void vunmap_range(unsigned long addr, unsigned long end) | |
480 | { | |
481 | flush_cache_vunmap(addr, end); | |
482 | vunmap_range_noflush(addr, end); | |
483 | flush_tlb_kernel_range(addr, end); | |
484 | } | |
485 | ||
0a264884 | 486 | static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr, |
2ba3e694 JR |
487 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, |
488 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
489 | { |
490 | pte_t *pte; | |
491 | ||
db64fe02 NP |
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 | ||
2ba3e694 | 497 | pte = pte_alloc_kernel_track(pmd, addr, mask); |
1da177e4 LT |
498 | if (!pte) |
499 | return -ENOMEM; | |
500 | do { | |
db64fe02 NP |
501 | struct page *page = pages[*nr]; |
502 | ||
c33c7948 | 503 | if (WARN_ON(!pte_none(ptep_get(pte)))) |
db64fe02 NP |
504 | return -EBUSY; |
505 | if (WARN_ON(!page)) | |
1da177e4 | 506 | return -ENOMEM; |
4fcdcc12 YN |
507 | if (WARN_ON(!pfn_valid(page_to_pfn(page)))) |
508 | return -EINVAL; | |
509 | ||
1da177e4 | 510 | set_pte_at(&init_mm, addr, pte, mk_pte(page, prot)); |
db64fe02 | 511 | (*nr)++; |
1da177e4 | 512 | } while (pte++, addr += PAGE_SIZE, addr != end); |
2ba3e694 | 513 | *mask |= PGTBL_PTE_MODIFIED; |
1da177e4 LT |
514 | return 0; |
515 | } | |
516 | ||
0a264884 | 517 | static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr, |
2ba3e694 JR |
518 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, |
519 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
520 | { |
521 | pmd_t *pmd; | |
522 | unsigned long next; | |
523 | ||
2ba3e694 | 524 | pmd = pmd_alloc_track(&init_mm, pud, addr, mask); |
1da177e4 LT |
525 | if (!pmd) |
526 | return -ENOMEM; | |
527 | do { | |
528 | next = pmd_addr_end(addr, end); | |
0a264884 | 529 | if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask)) |
1da177e4 LT |
530 | return -ENOMEM; |
531 | } while (pmd++, addr = next, addr != end); | |
532 | return 0; | |
533 | } | |
534 | ||
0a264884 | 535 | static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr, |
2ba3e694 JR |
536 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, |
537 | pgtbl_mod_mask *mask) | |
1da177e4 LT |
538 | { |
539 | pud_t *pud; | |
540 | unsigned long next; | |
541 | ||
2ba3e694 | 542 | pud = pud_alloc_track(&init_mm, p4d, addr, mask); |
1da177e4 LT |
543 | if (!pud) |
544 | return -ENOMEM; | |
545 | do { | |
546 | next = pud_addr_end(addr, end); | |
0a264884 | 547 | if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask)) |
1da177e4 LT |
548 | return -ENOMEM; |
549 | } while (pud++, addr = next, addr != end); | |
550 | return 0; | |
551 | } | |
552 | ||
0a264884 | 553 | static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr, |
2ba3e694 JR |
554 | unsigned long end, pgprot_t prot, struct page **pages, int *nr, |
555 | pgtbl_mod_mask *mask) | |
c2febafc KS |
556 | { |
557 | p4d_t *p4d; | |
558 | unsigned long next; | |
559 | ||
2ba3e694 | 560 | p4d = p4d_alloc_track(&init_mm, pgd, addr, mask); |
c2febafc KS |
561 | if (!p4d) |
562 | return -ENOMEM; | |
563 | do { | |
564 | next = p4d_addr_end(addr, end); | |
0a264884 | 565 | if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask)) |
c2febafc KS |
566 | return -ENOMEM; |
567 | } while (p4d++, addr = next, addr != end); | |
568 | return 0; | |
569 | } | |
570 | ||
121e6f32 NP |
571 | static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end, |
572 | pgprot_t prot, struct page **pages) | |
1da177e4 | 573 | { |
2ba3e694 | 574 | unsigned long start = addr; |
b521c43f | 575 | pgd_t *pgd; |
121e6f32 | 576 | unsigned long next; |
db64fe02 NP |
577 | int err = 0; |
578 | int nr = 0; | |
2ba3e694 | 579 | pgtbl_mod_mask mask = 0; |
1da177e4 LT |
580 | |
581 | BUG_ON(addr >= end); | |
582 | pgd = pgd_offset_k(addr); | |
1da177e4 LT |
583 | do { |
584 | next = pgd_addr_end(addr, end); | |
2ba3e694 JR |
585 | if (pgd_bad(*pgd)) |
586 | mask |= PGTBL_PGD_MODIFIED; | |
0a264884 | 587 | err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask); |
1da177e4 | 588 | if (err) |
bf88c8c8 | 589 | return err; |
1da177e4 | 590 | } while (pgd++, addr = next, addr != end); |
db64fe02 | 591 | |
2ba3e694 JR |
592 | if (mask & ARCH_PAGE_TABLE_SYNC_MASK) |
593 | arch_sync_kernel_mappings(start, end); | |
594 | ||
60bb4465 | 595 | return 0; |
1da177e4 LT |
596 | } |
597 | ||
b67177ec NP |
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 | */ | |
b073d7f8 | 607 | int __vmap_pages_range_noflush(unsigned long addr, unsigned long end, |
121e6f32 NP |
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), | |
08262ac5 | 622 | page_to_phys(pages[i]), prot, |
121e6f32 NP |
623 | page_shift); |
624 | if (err) | |
625 | return err; | |
626 | ||
627 | addr += 1UL << page_shift; | |
628 | } | |
629 | ||
630 | return 0; | |
631 | } | |
b073d7f8 AP |
632 | |
633 | int vmap_pages_range_noflush(unsigned long addr, unsigned long end, | |
634 | pgprot_t prot, struct page **pages, unsigned int page_shift) | |
635 | { | |
47ebd031 AP |
636 | int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages, |
637 | page_shift); | |
638 | ||
639 | if (ret) | |
640 | return ret; | |
b073d7f8 AP |
641 | return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift); |
642 | } | |
121e6f32 | 643 | |
121e6f32 | 644 | /** |
b67177ec | 645 | * vmap_pages_range - map pages to a kernel virtual address |
121e6f32 | 646 | * @addr: start of the VM area to map |
b67177ec | 647 | * @end: end of the VM area to map (non-inclusive) |
121e6f32 | 648 | * @prot: page protection flags to use |
b67177ec NP |
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. | |
121e6f32 NP |
652 | * |
653 | * RETURNS: | |
654 | * 0 on success, -errno on failure. | |
655 | */ | |
b67177ec NP |
656 | static int vmap_pages_range(unsigned long addr, unsigned long end, |
657 | pgprot_t prot, struct page **pages, unsigned int page_shift) | |
8fc48985 | 658 | { |
b67177ec | 659 | int err; |
8fc48985 | 660 | |
b67177ec NP |
661 | err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift); |
662 | flush_cache_vmap(addr, end); | |
663 | return err; | |
8fc48985 TH |
664 | } |
665 | ||
e6f79822 AS |
666 | static int check_sparse_vm_area(struct vm_struct *area, unsigned long start, |
667 | unsigned long end) | |
668 | { | |
669 | might_sleep(); | |
670 | if (WARN_ON_ONCE(area->flags & VM_FLUSH_RESET_PERMS)) | |
671 | return -EINVAL; | |
672 | if (WARN_ON_ONCE(area->flags & VM_NO_GUARD)) | |
673 | return -EINVAL; | |
674 | if (WARN_ON_ONCE(!(area->flags & VM_SPARSE))) | |
675 | return -EINVAL; | |
676 | if ((end - start) >> PAGE_SHIFT > totalram_pages()) | |
677 | return -E2BIG; | |
678 | if (start < (unsigned long)area->addr || | |
679 | (void *)end > area->addr + get_vm_area_size(area)) | |
680 | return -ERANGE; | |
681 | return 0; | |
682 | } | |
683 | ||
684 | /** | |
685 | * vm_area_map_pages - map pages inside given sparse vm_area | |
686 | * @area: vm_area | |
687 | * @start: start address inside vm_area | |
688 | * @end: end address inside vm_area | |
689 | * @pages: pages to map (always PAGE_SIZE pages) | |
690 | */ | |
691 | int vm_area_map_pages(struct vm_struct *area, unsigned long start, | |
692 | unsigned long end, struct page **pages) | |
693 | { | |
694 | int err; | |
695 | ||
696 | err = check_sparse_vm_area(area, start, end); | |
697 | if (err) | |
698 | return err; | |
699 | ||
700 | return vmap_pages_range(start, end, PAGE_KERNEL, pages, PAGE_SHIFT); | |
701 | } | |
702 | ||
703 | /** | |
704 | * vm_area_unmap_pages - unmap pages inside given sparse vm_area | |
705 | * @area: vm_area | |
706 | * @start: start address inside vm_area | |
707 | * @end: end address inside vm_area | |
708 | */ | |
709 | void vm_area_unmap_pages(struct vm_struct *area, unsigned long start, | |
710 | unsigned long end) | |
711 | { | |
712 | if (check_sparse_vm_area(area, start, end)) | |
713 | return; | |
714 | ||
715 | vunmap_range(start, end); | |
716 | } | |
717 | ||
81ac3ad9 | 718 | int is_vmalloc_or_module_addr(const void *x) |
73bdf0a6 LT |
719 | { |
720 | /* | |
ab4f2ee1 | 721 | * ARM, x86-64 and sparc64 put modules in a special place, |
73bdf0a6 LT |
722 | * and fall back on vmalloc() if that fails. Others |
723 | * just put it in the vmalloc space. | |
724 | */ | |
0105eaab | 725 | #if defined(CONFIG_EXECMEM) && defined(MODULES_VADDR) |
4aff1dc4 | 726 | unsigned long addr = (unsigned long)kasan_reset_tag(x); |
73bdf0a6 LT |
727 | if (addr >= MODULES_VADDR && addr < MODULES_END) |
728 | return 1; | |
729 | #endif | |
730 | return is_vmalloc_addr(x); | |
731 | } | |
01858469 | 732 | EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr); |
73bdf0a6 | 733 | |
48667e7a | 734 | /* |
c0eb315a NP |
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. | |
48667e7a | 738 | */ |
add688fb | 739 | struct page *vmalloc_to_page(const void *vmalloc_addr) |
48667e7a CL |
740 | { |
741 | unsigned long addr = (unsigned long) vmalloc_addr; | |
add688fb | 742 | struct page *page = NULL; |
48667e7a | 743 | pgd_t *pgd = pgd_offset_k(addr); |
c2febafc KS |
744 | p4d_t *p4d; |
745 | pud_t *pud; | |
746 | pmd_t *pmd; | |
747 | pte_t *ptep, pte; | |
48667e7a | 748 | |
7aa413de IM |
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 | */ | |
73bdf0a6 | 753 | VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr)); |
59ea7463 | 754 | |
c2febafc KS |
755 | if (pgd_none(*pgd)) |
756 | return NULL; | |
c0eb315a NP |
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 | ||
c2febafc KS |
762 | p4d = p4d_offset(pgd, addr); |
763 | if (p4d_none(*p4d)) | |
764 | return NULL; | |
c0eb315a NP |
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; | |
029c54b0 | 769 | |
c0eb315a NP |
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))) | |
c2febafc | 776 | return NULL; |
c0eb315a | 777 | |
c2febafc | 778 | pmd = pmd_offset(pud, addr); |
c0eb315a NP |
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))) | |
c2febafc KS |
784 | return NULL; |
785 | ||
0d1c81ed | 786 | ptep = pte_offset_kernel(pmd, addr); |
c33c7948 | 787 | pte = ptep_get(ptep); |
c2febafc KS |
788 | if (pte_present(pte)) |
789 | page = pte_page(pte); | |
c0eb315a | 790 | |
add688fb | 791 | return page; |
48667e7a | 792 | } |
add688fb | 793 | EXPORT_SYMBOL(vmalloc_to_page); |
48667e7a CL |
794 | |
795 | /* | |
add688fb | 796 | * Map a vmalloc()-space virtual address to the physical page frame number. |
48667e7a | 797 | */ |
add688fb | 798 | unsigned long vmalloc_to_pfn(const void *vmalloc_addr) |
48667e7a | 799 | { |
add688fb | 800 | return page_to_pfn(vmalloc_to_page(vmalloc_addr)); |
48667e7a | 801 | } |
add688fb | 802 | EXPORT_SYMBOL(vmalloc_to_pfn); |
48667e7a | 803 | |
db64fe02 NP |
804 | |
805 | /*** Global kva allocator ***/ | |
806 | ||
bb850f4d | 807 | #define DEBUG_AUGMENT_PROPAGATE_CHECK 0 |
a6cf4e0f | 808 | #define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0 |
bb850f4d | 809 | |
db64fe02 | 810 | |
e36176be | 811 | static DEFINE_SPINLOCK(free_vmap_area_lock); |
68ad4a33 | 812 | static bool vmap_initialized __read_mostly; |
89699605 | 813 | |
68ad4a33 URS |
814 | /* |
815 | * This kmem_cache is used for vmap_area objects. Instead of | |
816 | * allocating from slab we reuse an object from this cache to | |
817 | * make things faster. Especially in "no edge" splitting of | |
818 | * free block. | |
819 | */ | |
820 | static struct kmem_cache *vmap_area_cachep; | |
821 | ||
822 | /* | |
823 | * This linked list is used in pair with free_vmap_area_root. | |
824 | * It gives O(1) access to prev/next to perform fast coalescing. | |
825 | */ | |
826 | static LIST_HEAD(free_vmap_area_list); | |
827 | ||
828 | /* | |
829 | * This augment red-black tree represents the free vmap space. | |
830 | * All vmap_area objects in this tree are sorted by va->va_start | |
831 | * address. It is used for allocation and merging when a vmap | |
832 | * object is released. | |
833 | * | |
834 | * Each vmap_area node contains a maximum available free block | |
835 | * of its sub-tree, right or left. Therefore it is possible to | |
836 | * find a lowest match of free area. | |
837 | */ | |
838 | static struct rb_root free_vmap_area_root = RB_ROOT; | |
839 | ||
82dd23e8 URS |
840 | /* |
841 | * Preload a CPU with one object for "no edge" split case. The | |
842 | * aim is to get rid of allocations from the atomic context, thus | |
843 | * to use more permissive allocation masks. | |
844 | */ | |
845 | static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node); | |
846 | ||
d0936029 | 847 | /* |
15e02a39 URS |
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. | |
d0936029 URS |
852 | */ |
853 | struct rb_list { | |
854 | struct rb_root root; | |
855 | struct list_head head; | |
856 | spinlock_t lock; | |
857 | }; | |
858 | ||
15e02a39 URS |
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 | ||
72210662 URS |
866 | struct vmap_pool { |
867 | struct list_head head; | |
868 | unsigned long len; | |
869 | }; | |
870 | ||
871 | /* | |
15e02a39 URS |
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. | |
72210662 | 875 | */ |
d0936029 | 876 | static struct vmap_node { |
72210662 URS |
877 | /* Simple size segregated storage. */ |
878 | struct vmap_pool pool[MAX_VA_SIZE_PAGES]; | |
879 | spinlock_t pool_lock; | |
880 | bool skip_populate; | |
881 | ||
d0936029 URS |
882 | /* Bookkeeping data of this node. */ |
883 | struct rb_list busy; | |
282631cb URS |
884 | struct rb_list lazy; |
885 | ||
886 | /* | |
887 | * Ready-to-free areas. | |
888 | */ | |
889 | struct list_head purge_list; | |
72210662 URS |
890 | struct work_struct purge_work; |
891 | unsigned long nr_purged; | |
d0936029 URS |
892 | } single; |
893 | ||
15e02a39 URS |
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 | */ | |
d0936029 URS |
899 | static struct vmap_node *vmap_nodes = &single; |
900 | static __read_mostly unsigned int nr_vmap_nodes = 1; | |
901 | static __read_mostly unsigned int vmap_zone_size = 1; | |
902 | ||
903 | static inline unsigned int | |
904 | addr_to_node_id(unsigned long addr) | |
905 | { | |
906 | return (addr / vmap_zone_size) % nr_vmap_nodes; | |
907 | } | |
908 | ||
909 | static inline struct vmap_node * | |
910 | addr_to_node(unsigned long addr) | |
911 | { | |
912 | return &vmap_nodes[addr_to_node_id(addr)]; | |
913 | } | |
914 | ||
72210662 URS |
915 | static inline struct vmap_node * |
916 | id_to_node(unsigned int id) | |
917 | { | |
918 | return &vmap_nodes[id % nr_vmap_nodes]; | |
919 | } | |
920 | ||
921 | /* | |
922 | * We use the value 0 to represent "no node", that is why | |
923 | * an encoded value will be the node-id incremented by 1. | |
924 | * It is always greater then 0. A valid node_id which can | |
925 | * be encoded is [0:nr_vmap_nodes - 1]. If a passed node_id | |
926 | * is not valid 0 is returned. | |
927 | */ | |
928 | static unsigned int | |
929 | encode_vn_id(unsigned int node_id) | |
930 | { | |
931 | /* Can store U8_MAX [0:254] nodes. */ | |
932 | if (node_id < nr_vmap_nodes) | |
933 | return (node_id + 1) << BITS_PER_BYTE; | |
934 | ||
935 | /* Warn and no node encoded. */ | |
936 | WARN_ONCE(1, "Encode wrong node id (%u)\n", node_id); | |
937 | return 0; | |
938 | } | |
939 | ||
940 | /* | |
941 | * Returns an encoded node-id, the valid range is within | |
942 | * [0:nr_vmap_nodes-1] values. Otherwise nr_vmap_nodes is | |
943 | * returned if extracted data is wrong. | |
944 | */ | |
945 | static unsigned int | |
946 | decode_vn_id(unsigned int val) | |
947 | { | |
948 | unsigned int node_id = (val >> BITS_PER_BYTE) - 1; | |
949 | ||
950 | /* Can store U8_MAX [0:254] nodes. */ | |
951 | if (node_id < nr_vmap_nodes) | |
952 | return node_id; | |
953 | ||
954 | /* If it was _not_ zero, warn. */ | |
955 | WARN_ONCE(node_id != UINT_MAX, | |
956 | "Decode wrong node id (%d)\n", node_id); | |
957 | ||
958 | return nr_vmap_nodes; | |
959 | } | |
960 | ||
961 | static bool | |
962 | is_vn_id_valid(unsigned int node_id) | |
963 | { | |
964 | if (node_id < nr_vmap_nodes) | |
965 | return true; | |
966 | ||
967 | return false; | |
968 | } | |
969 | ||
68ad4a33 URS |
970 | static __always_inline unsigned long |
971 | va_size(struct vmap_area *va) | |
972 | { | |
973 | return (va->va_end - va->va_start); | |
974 | } | |
975 | ||
976 | static __always_inline unsigned long | |
977 | get_subtree_max_size(struct rb_node *node) | |
978 | { | |
979 | struct vmap_area *va; | |
980 | ||
981 | va = rb_entry_safe(node, struct vmap_area, rb_node); | |
982 | return va ? va->subtree_max_size : 0; | |
983 | } | |
89699605 | 984 | |
315cc066 ML |
985 | RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb, |
986 | struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size) | |
68ad4a33 | 987 | |
77e50af0 | 988 | static void reclaim_and_purge_vmap_areas(void); |
68ad4a33 | 989 | static BLOCKING_NOTIFIER_HEAD(vmap_notify_list); |
690467c8 URS |
990 | static void drain_vmap_area_work(struct work_struct *work); |
991 | static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work); | |
db64fe02 | 992 | |
97105f0a RG |
993 | static atomic_long_t nr_vmalloc_pages; |
994 | ||
995 | unsigned long vmalloc_nr_pages(void) | |
996 | { | |
997 | return atomic_long_read(&nr_vmalloc_pages); | |
998 | } | |
999 | ||
fc2c2269 URS |
1000 | static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root) |
1001 | { | |
1002 | struct rb_node *n = root->rb_node; | |
1003 | ||
1004 | addr = (unsigned long)kasan_reset_tag((void *)addr); | |
1005 | ||
1006 | while (n) { | |
1007 | struct vmap_area *va; | |
1008 | ||
1009 | va = rb_entry(n, struct vmap_area, rb_node); | |
1010 | if (addr < va->va_start) | |
1011 | n = n->rb_left; | |
1012 | else if (addr >= va->va_end) | |
1013 | n = n->rb_right; | |
1014 | else | |
1015 | return va; | |
1016 | } | |
1017 | ||
1018 | return NULL; | |
1019 | } | |
1020 | ||
153090f2 | 1021 | /* Look up the first VA which satisfies addr < va_end, NULL if none. */ |
d0936029 | 1022 | static struct vmap_area * |
53becf32 | 1023 | __find_vmap_area_exceed_addr(unsigned long addr, struct rb_root *root) |
f181234a CW |
1024 | { |
1025 | struct vmap_area *va = NULL; | |
d0936029 | 1026 | struct rb_node *n = root->rb_node; |
f181234a | 1027 | |
4aff1dc4 AK |
1028 | addr = (unsigned long)kasan_reset_tag((void *)addr); |
1029 | ||
f181234a CW |
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 | ||
53becf32 URS |
1047 | /* |
1048 | * Returns a node where a first VA, that satisfies addr < va_end, resides. | |
1049 | * If success, a node is locked. A user is responsible to unlock it when a | |
1050 | * VA is no longer needed to be accessed. | |
1051 | * | |
1052 | * Returns NULL if nothing found. | |
1053 | */ | |
1054 | static struct vmap_node * | |
1055 | find_vmap_area_exceed_addr_lock(unsigned long addr, struct vmap_area **va) | |
1056 | { | |
fc2c2269 URS |
1057 | unsigned long va_start_lowest; |
1058 | struct vmap_node *vn; | |
53becf32 URS |
1059 | int i; |
1060 | ||
fc2c2269 URS |
1061 | repeat: |
1062 | for (i = 0, va_start_lowest = 0; i < nr_vmap_nodes; i++) { | |
53becf32 URS |
1063 | vn = &vmap_nodes[i]; |
1064 | ||
1065 | spin_lock(&vn->busy.lock); | |
fc2c2269 URS |
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; | |
53becf32 URS |
1071 | spin_unlock(&vn->busy.lock); |
1072 | } | |
1073 | ||
fc2c2269 URS |
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); | |
53becf32 | 1081 | |
fc2c2269 URS |
1082 | spin_lock(&vn->busy.lock); |
1083 | *va = __find_vmap_area(va_start_lowest, &vn->busy.root); | |
db64fe02 | 1084 | |
fc2c2269 URS |
1085 | if (*va) |
1086 | return vn; | |
4aff1dc4 | 1087 | |
fc2c2269 URS |
1088 | spin_unlock(&vn->busy.lock); |
1089 | goto repeat; | |
db64fe02 NP |
1090 | } |
1091 | ||
1092 | return NULL; | |
1093 | } | |
1094 | ||
68ad4a33 URS |
1095 | /* |
1096 | * This function returns back addresses of parent node | |
1097 | * and its left or right link for further processing. | |
9c801f61 URS |
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. | |
68ad4a33 URS |
1102 | */ |
1103 | static __always_inline struct rb_node ** | |
1104 | find_va_links(struct vmap_area *va, | |
1105 | struct rb_root *root, struct rb_node *from, | |
1106 | struct rb_node **parent) | |
1107 | { | |
1108 | struct vmap_area *tmp_va; | |
1109 | struct rb_node **link; | |
1110 | ||
1111 | if (root) { | |
1112 | link = &root->rb_node; | |
1113 | if (unlikely(!*link)) { | |
1114 | *parent = NULL; | |
1115 | return link; | |
1116 | } | |
1117 | } else { | |
1118 | link = &from; | |
1119 | } | |
db64fe02 | 1120 | |
68ad4a33 URS |
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); | |
db64fe02 | 1128 | |
68ad4a33 URS |
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 | */ | |
753df96b | 1134 | if (va->va_end <= tmp_va->va_start) |
68ad4a33 | 1135 | link = &(*link)->rb_left; |
753df96b | 1136 | else if (va->va_start >= tmp_va->va_end) |
68ad4a33 | 1137 | link = &(*link)->rb_right; |
9c801f61 URS |
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 | } | |
68ad4a33 URS |
1144 | } while (*link); |
1145 | ||
1146 | *parent = &tmp_va->rb_node; | |
1147 | return link; | |
1148 | } | |
1149 | ||
1150 | static __always_inline struct list_head * | |
1151 | get_va_next_sibling(struct rb_node *parent, struct rb_node **link) | |
1152 | { | |
1153 | struct list_head *list; | |
1154 | ||
1155 | if (unlikely(!parent)) | |
1156 | /* | |
1157 | * The red-black tree where we try to find VA neighbors | |
1158 | * before merging or inserting is empty, i.e. it means | |
1159 | * there is no free vmap space. Normally it does not | |
1160 | * happen but we handle this case anyway. | |
1161 | */ | |
1162 | return NULL; | |
1163 | ||
1164 | list = &rb_entry(parent, struct vmap_area, rb_node)->list; | |
1165 | return (&parent->rb_right == link ? list->next : list); | |
1166 | } | |
1167 | ||
1168 | static __always_inline void | |
8eb510db URS |
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) | |
68ad4a33 URS |
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; | |
db64fe02 NP |
1181 | } |
1182 | ||
68ad4a33 URS |
1183 | /* Insert to the rb-tree */ |
1184 | rb_link_node(&va->rb_node, parent, link); | |
8eb510db | 1185 | if (augment) { |
68ad4a33 URS |
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(). | |
153090f2 | 1190 | * It is because we populate the tree from the bottom |
68ad4a33 URS |
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 | } | |
db64fe02 | 1203 | |
68ad4a33 URS |
1204 | /* Address-sort this list */ |
1205 | list_add(&va->list, head); | |
db64fe02 NP |
1206 | } |
1207 | ||
68ad4a33 | 1208 | static __always_inline void |
8eb510db URS |
1209 | link_va(struct vmap_area *va, struct rb_root *root, |
1210 | struct rb_node *parent, struct rb_node **link, | |
1211 | struct list_head *head) | |
1212 | { | |
1213 | __link_va(va, root, parent, link, head, false); | |
1214 | } | |
1215 | ||
1216 | static __always_inline void | |
1217 | link_va_augment(struct vmap_area *va, struct rb_root *root, | |
1218 | struct rb_node *parent, struct rb_node **link, | |
1219 | struct list_head *head) | |
1220 | { | |
1221 | __link_va(va, root, parent, link, head, true); | |
1222 | } | |
1223 | ||
1224 | static __always_inline void | |
1225 | __unlink_va(struct vmap_area *va, struct rb_root *root, bool augment) | |
68ad4a33 | 1226 | { |
460e42d1 URS |
1227 | if (WARN_ON(RB_EMPTY_NODE(&va->rb_node))) |
1228 | return; | |
db64fe02 | 1229 | |
8eb510db | 1230 | if (augment) |
460e42d1 URS |
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 | ||
5d7a7c54 | 1236 | list_del_init(&va->list); |
460e42d1 | 1237 | RB_CLEAR_NODE(&va->rb_node); |
68ad4a33 URS |
1238 | } |
1239 | ||
8eb510db URS |
1240 | static __always_inline void |
1241 | unlink_va(struct vmap_area *va, struct rb_root *root) | |
1242 | { | |
1243 | __unlink_va(va, root, false); | |
1244 | } | |
1245 | ||
1246 | static __always_inline void | |
1247 | unlink_va_augment(struct vmap_area *va, struct rb_root *root) | |
1248 | { | |
1249 | __unlink_va(va, root, true); | |
1250 | } | |
1251 | ||
bb850f4d | 1252 | #if DEBUG_AUGMENT_PROPAGATE_CHECK |
c3385e84 JC |
1253 | /* |
1254 | * Gets called when remove the node and rotate. | |
1255 | */ | |
1256 | static __always_inline unsigned long | |
1257 | compute_subtree_max_size(struct vmap_area *va) | |
1258 | { | |
1259 | return max3(va_size(va), | |
1260 | get_subtree_max_size(va->rb_node.rb_left), | |
1261 | get_subtree_max_size(va->rb_node.rb_right)); | |
1262 | } | |
1263 | ||
bb850f4d | 1264 | static void |
da27c9ed | 1265 | augment_tree_propagate_check(void) |
bb850f4d URS |
1266 | { |
1267 | struct vmap_area *va; | |
da27c9ed | 1268 | unsigned long computed_size; |
bb850f4d | 1269 | |
da27c9ed URS |
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); | |
bb850f4d | 1275 | } |
bb850f4d URS |
1276 | } |
1277 | #endif | |
1278 | ||
68ad4a33 URS |
1279 | /* |
1280 | * This function populates subtree_max_size from bottom to upper | |
1281 | * levels starting from VA point. The propagation must be done | |
1282 | * when VA size is modified by changing its va_start/va_end. Or | |
1283 | * in case of newly inserting of VA to the tree. | |
1284 | * | |
1285 | * It means that __augment_tree_propagate_from() must be called: | |
1286 | * - After VA has been inserted to the tree(free path); | |
1287 | * - After VA has been shrunk(allocation path); | |
1288 | * - After VA has been increased(merging path). | |
1289 | * | |
1290 | * Please note that, it does not mean that upper parent nodes | |
1291 | * and their subtree_max_size are recalculated all the time up | |
1292 | * to the root node. | |
1293 | * | |
1294 | * 4--8 | |
1295 | * /\ | |
1296 | * / \ | |
1297 | * / \ | |
1298 | * 2--2 8--8 | |
1299 | * | |
1300 | * For example if we modify the node 4, shrinking it to 2, then | |
1301 | * no any modification is required. If we shrink the node 2 to 1 | |
1302 | * its subtree_max_size is updated only, and set to 1. If we shrink | |
1303 | * the node 8 to 6, then its subtree_max_size is set to 6 and parent | |
1304 | * node becomes 4--6. | |
1305 | */ | |
1306 | static __always_inline void | |
1307 | augment_tree_propagate_from(struct vmap_area *va) | |
1308 | { | |
15ae144f URS |
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); | |
bb850f4d URS |
1315 | |
1316 | #if DEBUG_AUGMENT_PROPAGATE_CHECK | |
da27c9ed | 1317 | augment_tree_propagate_check(); |
bb850f4d | 1318 | #endif |
68ad4a33 URS |
1319 | } |
1320 | ||
1321 | static void | |
1322 | insert_vmap_area(struct vmap_area *va, | |
1323 | struct rb_root *root, struct list_head *head) | |
1324 | { | |
1325 | struct rb_node **link; | |
1326 | struct rb_node *parent; | |
1327 | ||
1328 | link = find_va_links(va, root, NULL, &parent); | |
9c801f61 URS |
1329 | if (link) |
1330 | link_va(va, root, parent, link, head); | |
68ad4a33 URS |
1331 | } |
1332 | ||
1333 | static void | |
1334 | insert_vmap_area_augment(struct vmap_area *va, | |
1335 | struct rb_node *from, struct rb_root *root, | |
1336 | struct list_head *head) | |
1337 | { | |
1338 | struct rb_node **link; | |
1339 | struct rb_node *parent; | |
1340 | ||
1341 | if (from) | |
1342 | link = find_va_links(va, NULL, from, &parent); | |
1343 | else | |
1344 | link = find_va_links(va, root, NULL, &parent); | |
1345 | ||
9c801f61 | 1346 | if (link) { |
8eb510db | 1347 | link_va_augment(va, root, parent, link, head); |
9c801f61 URS |
1348 | augment_tree_propagate_from(va); |
1349 | } | |
68ad4a33 URS |
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. | |
9c801f61 URS |
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. | |
68ad4a33 | 1362 | */ |
3c5c3cfb | 1363 | static __always_inline struct vmap_area * |
8eb510db URS |
1364 | __merge_or_add_vmap_area(struct vmap_area *va, |
1365 | struct rb_root *root, struct list_head *head, bool augment) | |
68ad4a33 URS |
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); | |
9c801f61 URS |
1378 | if (!link) |
1379 | return NULL; | |
68ad4a33 URS |
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 | ||
68ad4a33 URS |
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) { | |
5dd78640 URS |
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 | */ | |
54f63d9d | 1426 | if (merged) |
8eb510db | 1427 | __unlink_va(va, root, augment); |
68ad4a33 | 1428 | |
5dd78640 URS |
1429 | sibling->va_end = va->va_end; |
1430 | ||
68ad4a33 URS |
1431 | /* Free vmap_area object. */ |
1432 | kmem_cache_free(vmap_area_cachep, va); | |
3c5c3cfb DA |
1433 | |
1434 | /* Point to the new merged area. */ | |
1435 | va = sibling; | |
1436 | merged = true; | |
68ad4a33 URS |
1437 | } |
1438 | } | |
1439 | ||
1440 | insert: | |
5dd78640 | 1441 | if (!merged) |
8eb510db | 1442 | __link_va(va, root, parent, link, head, augment); |
3c5c3cfb | 1443 | |
96e2db45 URS |
1444 | return va; |
1445 | } | |
1446 | ||
8eb510db URS |
1447 | static __always_inline struct vmap_area * |
1448 | merge_or_add_vmap_area(struct vmap_area *va, | |
1449 | struct rb_root *root, struct list_head *head) | |
1450 | { | |
1451 | return __merge_or_add_vmap_area(va, root, head, false); | |
1452 | } | |
1453 | ||
96e2db45 URS |
1454 | static __always_inline struct vmap_area * |
1455 | merge_or_add_vmap_area_augment(struct vmap_area *va, | |
1456 | struct rb_root *root, struct list_head *head) | |
1457 | { | |
8eb510db | 1458 | va = __merge_or_add_vmap_area(va, root, head, true); |
96e2db45 URS |
1459 | if (va) |
1460 | augment_tree_propagate_from(va); | |
1461 | ||
3c5c3cfb | 1462 | return va; |
68ad4a33 URS |
1463 | } |
1464 | ||
1465 | static __always_inline bool | |
1466 | is_within_this_va(struct vmap_area *va, unsigned long size, | |
1467 | unsigned long align, unsigned long vstart) | |
1468 | { | |
1469 | unsigned long nva_start_addr; | |
1470 | ||
1471 | if (va->va_start > vstart) | |
1472 | nva_start_addr = ALIGN(va->va_start, align); | |
1473 | else | |
1474 | nva_start_addr = ALIGN(vstart, align); | |
1475 | ||
1476 | /* Can be overflowed due to big size or alignment. */ | |
1477 | if (nva_start_addr + size < nva_start_addr || | |
1478 | nva_start_addr < vstart) | |
1479 | return false; | |
1480 | ||
1481 | return (nva_start_addr + size <= va->va_end); | |
1482 | } | |
1483 | ||
1484 | /* | |
1485 | * Find the first free block(lowest start address) in the tree, | |
1486 | * that will accomplish the request corresponding to passing | |
9333fe98 UR |
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. | |
68ad4a33 URS |
1490 | */ |
1491 | static __always_inline struct vmap_area * | |
f9863be4 URS |
1492 | find_vmap_lowest_match(struct rb_root *root, unsigned long size, |
1493 | unsigned long align, unsigned long vstart, bool adjust_search_size) | |
68ad4a33 URS |
1494 | { |
1495 | struct vmap_area *va; | |
1496 | struct rb_node *node; | |
9333fe98 | 1497 | unsigned long length; |
68ad4a33 URS |
1498 | |
1499 | /* Start from the root. */ | |
f9863be4 | 1500 | node = root->rb_node; |
68ad4a33 | 1501 | |
9333fe98 UR |
1502 | /* Adjust the search size for alignment overhead. */ |
1503 | length = adjust_search_size ? size + align - 1 : size; | |
1504 | ||
68ad4a33 URS |
1505 | while (node) { |
1506 | va = rb_entry(node, struct vmap_area, rb_node); | |
1507 | ||
9333fe98 | 1508 | if (get_subtree_max_size(node->rb_left) >= length && |
68ad4a33 URS |
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 | |
9333fe98 | 1518 | * equal or bigger to the requested search length. |
68ad4a33 | 1519 | */ |
9333fe98 | 1520 | if (get_subtree_max_size(node->rb_right) >= length) { |
68ad4a33 URS |
1521 | node = node->rb_right; |
1522 | continue; | |
1523 | } | |
1524 | ||
1525 | /* | |
3806b041 | 1526 | * OK. We roll back and find the first right sub-tree, |
68ad4a33 | 1527 | * that will satisfy the search criteria. It can happen |
9f531973 URS |
1528 | * due to "vstart" restriction or an alignment overhead |
1529 | * that is bigger then PAGE_SIZE. | |
68ad4a33 URS |
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 | ||
9333fe98 | 1536 | if (get_subtree_max_size(node->rb_right) >= length && |
68ad4a33 | 1537 | vstart <= va->va_start) { |
9f531973 URS |
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; | |
68ad4a33 URS |
1545 | node = node->rb_right; |
1546 | break; | |
1547 | } | |
1548 | } | |
1549 | } | |
1550 | } | |
1551 | ||
1552 | return NULL; | |
1553 | } | |
1554 | ||
a6cf4e0f URS |
1555 | #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK |
1556 | #include <linux/random.h> | |
1557 | ||
1558 | static struct vmap_area * | |
bd1264c3 | 1559 | find_vmap_lowest_linear_match(struct list_head *head, unsigned long size, |
a6cf4e0f URS |
1560 | unsigned long align, unsigned long vstart) |
1561 | { | |
1562 | struct vmap_area *va; | |
1563 | ||
bd1264c3 | 1564 | list_for_each_entry(va, head, list) { |
a6cf4e0f URS |
1565 | if (!is_within_this_va(va, size, align, vstart)) |
1566 | continue; | |
1567 | ||
1568 | return va; | |
1569 | } | |
1570 | ||
1571 | return NULL; | |
1572 | } | |
1573 | ||
1574 | static void | |
bd1264c3 SL |
1575 | find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head, |
1576 | unsigned long size, unsigned long align) | |
a6cf4e0f URS |
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 | ||
bd1264c3 SL |
1585 | va_1 = find_vmap_lowest_match(root, size, align, vstart, false); |
1586 | va_2 = find_vmap_lowest_linear_match(head, size, align, vstart); | |
a6cf4e0f URS |
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 | ||
68ad4a33 URS |
1594 | enum fit_type { |
1595 | NOTHING_FIT = 0, | |
1596 | FL_FIT_TYPE = 1, /* full fit */ | |
1597 | LE_FIT_TYPE = 2, /* left edge fit */ | |
1598 | RE_FIT_TYPE = 3, /* right edge fit */ | |
1599 | NE_FIT_TYPE = 4 /* no edge fit */ | |
1600 | }; | |
1601 | ||
1602 | static __always_inline enum fit_type | |
1603 | classify_va_fit_type(struct vmap_area *va, | |
1604 | unsigned long nva_start_addr, unsigned long size) | |
1605 | { | |
1606 | enum fit_type type; | |
1607 | ||
1608 | /* Check if it is within VA. */ | |
1609 | if (nva_start_addr < va->va_start || | |
1610 | nva_start_addr + size > va->va_end) | |
1611 | return NOTHING_FIT; | |
1612 | ||
1613 | /* Now classify. */ | |
1614 | if (va->va_start == nva_start_addr) { | |
1615 | if (va->va_end == nva_start_addr + size) | |
1616 | type = FL_FIT_TYPE; | |
1617 | else | |
1618 | type = LE_FIT_TYPE; | |
1619 | } else if (va->va_end == nva_start_addr + size) { | |
1620 | type = RE_FIT_TYPE; | |
1621 | } else { | |
1622 | type = NE_FIT_TYPE; | |
1623 | } | |
1624 | ||
1625 | return type; | |
1626 | } | |
1627 | ||
1628 | static __always_inline int | |
5b75b8e1 URS |
1629 | va_clip(struct rb_root *root, struct list_head *head, |
1630 | struct vmap_area *va, unsigned long nva_start_addr, | |
1631 | unsigned long size) | |
68ad4a33 | 1632 | { |
2c929233 | 1633 | struct vmap_area *lva = NULL; |
1b23ff80 | 1634 | enum fit_type type = classify_va_fit_type(va, nva_start_addr, size); |
68ad4a33 URS |
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 | */ | |
f9863be4 | 1644 | unlink_va_augment(va, root); |
68ad4a33 URS |
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 | */ | |
82dd23e8 URS |
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. | |
060650a2 URS |
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. | |
82dd23e8 URS |
1698 | */ |
1699 | lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT); | |
1700 | if (!lva) | |
1701 | return -1; | |
1702 | } | |
68ad4a33 URS |
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 | ||
2c929233 | 1721 | if (lva) /* type == NE_FIT_TYPE */ |
f9863be4 | 1722 | insert_vmap_area_augment(lva, &va->rb_node, root, head); |
68ad4a33 URS |
1723 | } |
1724 | ||
1725 | return 0; | |
1726 | } | |
1727 | ||
38f6b9af URS |
1728 | static unsigned long |
1729 | va_alloc(struct vmap_area *va, | |
1730 | struct rb_root *root, struct list_head *head, | |
1731 | unsigned long size, unsigned long align, | |
1732 | unsigned long vstart, unsigned long vend) | |
1733 | { | |
1734 | unsigned long nva_start_addr; | |
1735 | int ret; | |
1736 | ||
1737 | if (va->va_start > vstart) | |
1738 | nva_start_addr = ALIGN(va->va_start, align); | |
1739 | else | |
1740 | nva_start_addr = ALIGN(vstart, align); | |
1741 | ||
1742 | /* Check the "vend" restriction. */ | |
1743 | if (nva_start_addr + size > vend) | |
1744 | return vend; | |
1745 | ||
1746 | /* Update the free vmap_area. */ | |
5b75b8e1 | 1747 | ret = va_clip(root, head, va, nva_start_addr, size); |
38f6b9af URS |
1748 | if (WARN_ON_ONCE(ret)) |
1749 | return vend; | |
1750 | ||
1751 | return nva_start_addr; | |
1752 | } | |
1753 | ||
68ad4a33 URS |
1754 | /* |
1755 | * Returns a start address of the newly allocated area, if success. | |
1756 | * Otherwise a vend is returned that indicates failure. | |
1757 | */ | |
1758 | static __always_inline unsigned long | |
f9863be4 URS |
1759 | __alloc_vmap_area(struct rb_root *root, struct list_head *head, |
1760 | unsigned long size, unsigned long align, | |
cacca6ba | 1761 | unsigned long vstart, unsigned long vend) |
68ad4a33 | 1762 | { |
9333fe98 | 1763 | bool adjust_search_size = true; |
68ad4a33 URS |
1764 | unsigned long nva_start_addr; |
1765 | struct vmap_area *va; | |
68ad4a33 | 1766 | |
9333fe98 UR |
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 | ||
f9863be4 | 1779 | va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size); |
68ad4a33 URS |
1780 | if (unlikely(!va)) |
1781 | return vend; | |
1782 | ||
38f6b9af URS |
1783 | nva_start_addr = va_alloc(va, root, head, size, align, vstart, vend); |
1784 | if (nva_start_addr == vend) | |
68ad4a33 URS |
1785 | return vend; |
1786 | ||
a6cf4e0f | 1787 | #if DEBUG_AUGMENT_LOWEST_MATCH_CHECK |
bd1264c3 | 1788 | find_vmap_lowest_match_check(root, head, size, align); |
a6cf4e0f URS |
1789 | #endif |
1790 | ||
68ad4a33 URS |
1791 | return nva_start_addr; |
1792 | } | |
4da56b99 | 1793 | |
d98c9e83 AR |
1794 | /* |
1795 | * Free a region of KVA allocated by alloc_vmap_area | |
1796 | */ | |
1797 | static void free_vmap_area(struct vmap_area *va) | |
1798 | { | |
d0936029 URS |
1799 | struct vmap_node *vn = addr_to_node(va->va_start); |
1800 | ||
d98c9e83 AR |
1801 | /* |
1802 | * Remove from the busy tree/list. | |
1803 | */ | |
d0936029 URS |
1804 | spin_lock(&vn->busy.lock); |
1805 | unlink_va(va, &vn->busy.root); | |
1806 | spin_unlock(&vn->busy.lock); | |
d98c9e83 AR |
1807 | |
1808 | /* | |
1809 | * Insert/Merge it back to the free tree/list. | |
1810 | */ | |
1811 | spin_lock(&free_vmap_area_lock); | |
96e2db45 | 1812 | merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list); |
d98c9e83 AR |
1813 | spin_unlock(&free_vmap_area_lock); |
1814 | } | |
1815 | ||
187f8cc4 URS |
1816 | static inline void |
1817 | preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node) | |
1818 | { | |
f56810c9 | 1819 | struct vmap_area *va = NULL, *tmp; |
187f8cc4 URS |
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 | ||
f56810c9 UB |
1835 | tmp = NULL; |
1836 | if (va && !__this_cpu_try_cmpxchg(ne_fit_preload_node, &tmp, va)) | |
187f8cc4 URS |
1837 | kmem_cache_free(vmap_area_cachep, va); |
1838 | } | |
1839 | ||
72210662 URS |
1840 | static struct vmap_pool * |
1841 | size_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 | ||
1851 | static bool | |
1852 | node_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 | ||
1868 | static struct vmap_area * | |
1869 | node_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 | ||
1910 | static struct vmap_area * | |
1911 | node_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 | ||
aaab830a | 1938 | static 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->va_end - va->va_start; | |
1944 | vm->caller = caller; | |
1945 | va->vm = vm; | |
1946 | } | |
1947 | ||
db64fe02 NP |
1948 | /* |
1949 | * Allocate a region of KVA of the specified size and alignment, within the | |
aaab830a | 1950 | * vstart and vend. If vm is passed in, the two will also be bound. |
db64fe02 NP |
1951 | */ |
1952 | static struct vmap_area *alloc_vmap_area(unsigned long size, | |
1953 | unsigned long align, | |
1954 | unsigned long vstart, unsigned long vend, | |
869176a0 | 1955 | int node, gfp_t gfp_mask, |
4b68a773 | 1956 | unsigned long va_flags, struct vm_struct *vm) |
db64fe02 | 1957 | { |
d0936029 | 1958 | struct vmap_node *vn; |
187f8cc4 | 1959 | struct vmap_area *va; |
12e376a6 | 1960 | unsigned long freed; |
1da177e4 | 1961 | unsigned long addr; |
72210662 | 1962 | unsigned int vn_id; |
db64fe02 | 1963 | int purged = 0; |
d98c9e83 | 1964 | int ret; |
db64fe02 | 1965 | |
7e4a32c0 HL |
1966 | if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align))) |
1967 | return ERR_PTR(-EINVAL); | |
db64fe02 | 1968 | |
68ad4a33 URS |
1969 | if (unlikely(!vmap_initialized)) |
1970 | return ERR_PTR(-EBUSY); | |
1971 | ||
5803ed29 | 1972 | might_sleep(); |
db64fe02 | 1973 | |
7f88f88f | 1974 | /* |
72210662 URS |
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. | |
7f88f88f | 1981 | */ |
72210662 URS |
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); | |
db64fe02 | 1989 | |
96aa8437 URS |
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 | } | |
7f88f88f | 1996 | |
db64fe02 | 1997 | retry: |
72210662 URS |
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 | } | |
89699605 | 2004 | |
cf243da6 URS |
2005 | trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend); |
2006 | ||
afd07389 | 2007 | /* |
68ad4a33 URS |
2008 | * If an allocation fails, the "vend" address is |
2009 | * returned. Therefore trigger the overflow path. | |
afd07389 | 2010 | */ |
68ad4a33 | 2011 | if (unlikely(addr == vend)) |
89699605 | 2012 | goto overflow; |
db64fe02 NP |
2013 | |
2014 | va->va_start = addr; | |
2015 | va->va_end = addr + size; | |
688fcbfc | 2016 | va->vm = NULL; |
72210662 | 2017 | va->flags = (va_flags | vn_id); |
68ad4a33 | 2018 | |
4b68a773 BH |
2019 | if (vm) { |
2020 | vm->addr = (void *)va->va_start; | |
2021 | vm->size = va->va_end - va->va_start; | |
2022 | va->vm = vm; | |
2023 | } | |
aaab830a | 2024 | |
d0936029 URS |
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); | |
db64fe02 | 2030 | |
61e16557 | 2031 | BUG_ON(!IS_ALIGNED(va->va_start, align)); |
89699605 NP |
2032 | BUG_ON(va->va_start < vstart); |
2033 | BUG_ON(va->va_end > vend); | |
2034 | ||
d98c9e83 AR |
2035 | ret = kasan_populate_vmalloc(addr, size); |
2036 | if (ret) { | |
2037 | free_vmap_area(va); | |
2038 | return ERR_PTR(ret); | |
2039 | } | |
2040 | ||
db64fe02 | 2041 | return va; |
89699605 NP |
2042 | |
2043 | overflow: | |
89699605 | 2044 | if (!purged) { |
77e50af0 | 2045 | reclaim_and_purge_vmap_areas(); |
89699605 NP |
2046 | purged = 1; |
2047 | goto retry; | |
2048 | } | |
4da56b99 | 2049 | |
12e376a6 URS |
2050 | freed = 0; |
2051 | blocking_notifier_call_chain(&vmap_notify_list, 0, &freed); | |
2052 | ||
2053 | if (freed > 0) { | |
2054 | purged = 0; | |
2055 | goto retry; | |
4da56b99 CW |
2056 | } |
2057 | ||
03497d76 | 2058 | if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) |
55ccad6f SKO |
2059 | pr_warn("vmalloc_node_range for size %lu failed: Address range restricted to %#lx - %#lx\n", |
2060 | size, vstart, vend); | |
68ad4a33 URS |
2061 | |
2062 | kmem_cache_free(vmap_area_cachep, va); | |
89699605 | 2063 | return ERR_PTR(-EBUSY); |
db64fe02 NP |
2064 | } |
2065 | ||
4da56b99 CW |
2066 | int register_vmap_purge_notifier(struct notifier_block *nb) |
2067 | { | |
2068 | return blocking_notifier_chain_register(&vmap_notify_list, nb); | |
2069 | } | |
2070 | EXPORT_SYMBOL_GPL(register_vmap_purge_notifier); | |
2071 | ||
2072 | int unregister_vmap_purge_notifier(struct notifier_block *nb) | |
2073 | { | |
2074 | return blocking_notifier_chain_unregister(&vmap_notify_list, nb); | |
2075 | } | |
2076 | EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier); | |
2077 | ||
db64fe02 NP |
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 | */ | |
2094 | static 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 | ||
4d36e6f8 | 2103 | static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0); |
db64fe02 | 2104 | |
0574ecd1 | 2105 | /* |
f0953a1b | 2106 | * Serialize vmap purging. There is no actual critical section protected |
153090f2 | 2107 | * by this lock, but we want to avoid concurrent calls for performance |
0574ecd1 CH |
2108 | * reasons and to make the pcpu_get_vm_areas more deterministic. |
2109 | */ | |
f9e09977 | 2110 | static DEFINE_MUTEX(vmap_purge_lock); |
0574ecd1 | 2111 | |
02b709df NP |
2112 | /* for per-CPU blocks */ |
2113 | static void purge_fragmented_blocks_allcpus(void); | |
282631cb | 2114 | static cpumask_t purge_nodes; |
02b709df | 2115 | |
72210662 URS |
2116 | static void |
2117 | reclaim_list_global(struct list_head *head) | |
db64fe02 | 2118 | { |
72210662 | 2119 | struct vmap_area *va, *n; |
db64fe02 | 2120 | |
72210662 URS |
2121 | if (list_empty(head)) |
2122 | return; | |
02b709df | 2123 | |
e36176be | 2124 | spin_lock(&free_vmap_area_lock); |
72210662 URS |
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 | } | |
96e2db45 | 2130 | |
72210662 URS |
2131 | static void |
2132 | decay_va_pool_node(struct vmap_node *vn, bool full_decay) | |
2133 | { | |
2134 | struct vmap_area *va, *nva; | |
2135 | struct list_head decay_list; | |
2136 | struct rb_root decay_root; | |
2137 | unsigned long n_decay; | |
2138 | int i; | |
68571be9 | 2139 | |
72210662 URS |
2140 | decay_root = RB_ROOT; |
2141 | INIT_LIST_HEAD(&decay_list); | |
96e2db45 | 2142 | |
72210662 URS |
2143 | for (i = 0; i < MAX_VA_SIZE_PAGES; i++) { |
2144 | struct list_head tmp_list; | |
db64fe02 | 2145 | |
72210662 URS |
2146 | if (list_empty(&vn->pool[i].head)) |
2147 | continue; | |
db64fe02 | 2148 | |
72210662 URS |
2149 | INIT_LIST_HEAD(&tmp_list); |
2150 | ||
2151 | /* Detach the pool, so no-one can access it. */ | |
2152 | spin_lock(&vn->pool_lock); | |
2153 | list_replace_init(&vn->pool[i].head, &tmp_list); | |
2154 | spin_unlock(&vn->pool_lock); | |
2155 | ||
2156 | if (full_decay) | |
2157 | WRITE_ONCE(vn->pool[i].len, 0); | |
2158 | ||
2159 | /* Decay a pool by ~25% out of left objects. */ | |
2160 | n_decay = vn->pool[i].len >> 2; | |
2161 | ||
2162 | list_for_each_entry_safe(va, nva, &tmp_list, list) { | |
2163 | list_del_init(&va->list); | |
2164 | merge_or_add_vmap_area(va, &decay_root, &decay_list); | |
2165 | ||
2166 | if (!full_decay) { | |
2167 | WRITE_ONCE(vn->pool[i].len, vn->pool[i].len - 1); | |
2168 | ||
2169 | if (!--n_decay) | |
2170 | break; | |
2171 | } | |
2172 | } | |
763b218d | 2173 | |
dd3b8353 | 2174 | /* |
15e02a39 URS |
2175 | * Attach the pool back if it has been partly decayed. |
2176 | * Please note, it is supposed that nobody(other contexts) | |
2177 | * can populate the pool therefore a simple list replace | |
2178 | * operation takes place here. | |
dd3b8353 | 2179 | */ |
72210662 URS |
2180 | if (!full_decay && !list_empty(&tmp_list)) { |
2181 | spin_lock(&vn->pool_lock); | |
2182 | list_replace_init(&tmp_list, &vn->pool[i].head); | |
2183 | spin_unlock(&vn->pool_lock); | |
2184 | } | |
2185 | } | |
3c5c3cfb | 2186 | |
72210662 URS |
2187 | reclaim_list_global(&decay_list); |
2188 | } | |
2189 | ||
2190 | static void purge_vmap_node(struct work_struct *work) | |
2191 | { | |
2192 | struct vmap_node *vn = container_of(work, | |
2193 | struct vmap_node, purge_work); | |
409faf8c | 2194 | unsigned long nr_purged_pages = 0; |
72210662 URS |
2195 | struct vmap_area *va, *n_va; |
2196 | LIST_HEAD(local_list); | |
2197 | ||
2198 | vn->nr_purged = 0; | |
2199 | ||
282631cb | 2200 | list_for_each_entry_safe(va, n_va, &vn->purge_list, list) { |
4d36e6f8 | 2201 | unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT; |
3c5c3cfb DA |
2202 | unsigned long orig_start = va->va_start; |
2203 | unsigned long orig_end = va->va_end; | |
72210662 | 2204 | unsigned int vn_id = decode_vn_id(va->flags); |
763b218d | 2205 | |
72210662 | 2206 | list_del_init(&va->list); |
9c801f61 | 2207 | |
3c5c3cfb DA |
2208 | if (is_vmalloc_or_module_addr((void *)orig_start)) |
2209 | kasan_release_vmalloc(orig_start, orig_end, | |
2210 | va->va_start, va->va_end); | |
dd3b8353 | 2211 | |
409faf8c | 2212 | nr_purged_pages += nr; |
72210662 | 2213 | vn->nr_purged++; |
68571be9 | 2214 | |
72210662 URS |
2215 | if (is_vn_id_valid(vn_id) && !vn->skip_populate) |
2216 | if (node_pool_add_va(vn, va)) | |
2217 | continue; | |
2218 | ||
2219 | /* Go back to global. */ | |
2220 | list_add(&va->list, &local_list); | |
763b218d | 2221 | } |
6030fd5f | 2222 | |
409faf8c AH |
2223 | atomic_long_sub(nr_purged_pages, &vmap_lazy_nr); |
2224 | ||
72210662 | 2225 | reclaim_list_global(&local_list); |
282631cb URS |
2226 | } |
2227 | ||
2228 | /* | |
2229 | * Purges all lazily-freed vmap areas. | |
2230 | */ | |
72210662 URS |
2231 | static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end, |
2232 | bool full_pool_decay) | |
282631cb | 2233 | { |
72210662 URS |
2234 | unsigned long nr_purged_areas = 0; |
2235 | unsigned int nr_purge_helpers; | |
2236 | unsigned int nr_purge_nodes; | |
282631cb URS |
2237 | struct vmap_node *vn; |
2238 | int i; | |
2239 | ||
2240 | lockdep_assert_held(&vmap_purge_lock); | |
72210662 URS |
2241 | |
2242 | /* | |
2243 | * Use cpumask to mark which node has to be processed. | |
2244 | */ | |
282631cb URS |
2245 | purge_nodes = CPU_MASK_NONE; |
2246 | ||
2247 | for (i = 0; i < nr_vmap_nodes; i++) { | |
2248 | vn = &vmap_nodes[i]; | |
2249 | ||
2250 | INIT_LIST_HEAD(&vn->purge_list); | |
72210662 URS |
2251 | vn->skip_populate = full_pool_decay; |
2252 | decay_va_pool_node(vn, full_pool_decay); | |
282631cb URS |
2253 | |
2254 | if (RB_EMPTY_ROOT(&vn->lazy.root)) | |
2255 | continue; | |
2256 | ||
2257 | spin_lock(&vn->lazy.lock); | |
2258 | WRITE_ONCE(vn->lazy.root.rb_node, NULL); | |
2259 | list_replace_init(&vn->lazy.head, &vn->purge_list); | |
2260 | spin_unlock(&vn->lazy.lock); | |
2261 | ||
2262 | start = min(start, list_first_entry(&vn->purge_list, | |
2263 | struct vmap_area, list)->va_start); | |
2264 | ||
2265 | end = max(end, list_last_entry(&vn->purge_list, | |
2266 | struct vmap_area, list)->va_end); | |
2267 | ||
2268 | cpumask_set_cpu(i, &purge_nodes); | |
2269 | } | |
2270 | ||
72210662 URS |
2271 | nr_purge_nodes = cpumask_weight(&purge_nodes); |
2272 | if (nr_purge_nodes > 0) { | |
282631cb URS |
2273 | flush_tlb_kernel_range(start, end); |
2274 | ||
72210662 URS |
2275 | /* One extra worker is per a lazy_max_pages() full set minus one. */ |
2276 | nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages(); | |
2277 | nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1; | |
2278 | ||
282631cb | 2279 | for_each_cpu(i, &purge_nodes) { |
72210662 URS |
2280 | vn = &vmap_nodes[i]; |
2281 | ||
2282 | if (nr_purge_helpers > 0) { | |
2283 | INIT_WORK(&vn->purge_work, purge_vmap_node); | |
2284 | ||
2285 | if (cpumask_test_cpu(i, cpu_online_mask)) | |
2286 | schedule_work_on(i, &vn->purge_work); | |
2287 | else | |
2288 | schedule_work(&vn->purge_work); | |
2289 | ||
2290 | nr_purge_helpers--; | |
2291 | } else { | |
2292 | vn->purge_work.func = NULL; | |
2293 | purge_vmap_node(&vn->purge_work); | |
2294 | nr_purged_areas += vn->nr_purged; | |
2295 | } | |
2296 | } | |
2297 | ||
2298 | for_each_cpu(i, &purge_nodes) { | |
2299 | vn = &vmap_nodes[i]; | |
2300 | ||
2301 | if (vn->purge_work.func) { | |
2302 | flush_work(&vn->purge_work); | |
2303 | nr_purged_areas += vn->nr_purged; | |
2304 | } | |
282631cb URS |
2305 | } |
2306 | } | |
2307 | ||
72210662 URS |
2308 | trace_purge_vmap_area_lazy(start, end, nr_purged_areas); |
2309 | return nr_purged_areas > 0; | |
db64fe02 NP |
2310 | } |
2311 | ||
2312 | /* | |
77e50af0 | 2313 | * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list. |
db64fe02 | 2314 | */ |
77e50af0 TG |
2315 | static void reclaim_and_purge_vmap_areas(void) |
2316 | ||
db64fe02 | 2317 | { |
f9e09977 | 2318 | mutex_lock(&vmap_purge_lock); |
0574ecd1 | 2319 | purge_fragmented_blocks_allcpus(); |
72210662 | 2320 | __purge_vmap_area_lazy(ULONG_MAX, 0, true); |
f9e09977 | 2321 | mutex_unlock(&vmap_purge_lock); |
db64fe02 NP |
2322 | } |
2323 | ||
690467c8 URS |
2324 | static void drain_vmap_area_work(struct work_struct *work) |
2325 | { | |
282631cb | 2326 | mutex_lock(&vmap_purge_lock); |
72210662 | 2327 | __purge_vmap_area_lazy(ULONG_MAX, 0, false); |
282631cb | 2328 | mutex_unlock(&vmap_purge_lock); |
690467c8 URS |
2329 | } |
2330 | ||
db64fe02 | 2331 | /* |
edd89818 URS |
2332 | * Free a vmap area, caller ensuring that the area has been unmapped, |
2333 | * unlinked and flush_cache_vunmap had been called for the correct | |
2334 | * range previously. | |
db64fe02 | 2335 | */ |
64141da5 | 2336 | static void free_vmap_area_noflush(struct vmap_area *va) |
db64fe02 | 2337 | { |
8c4196fe URS |
2338 | unsigned long nr_lazy_max = lazy_max_pages(); |
2339 | unsigned long va_start = va->va_start; | |
72210662 URS |
2340 | unsigned int vn_id = decode_vn_id(va->flags); |
2341 | struct vmap_node *vn; | |
4d36e6f8 | 2342 | unsigned long nr_lazy; |
80c4bd7a | 2343 | |
edd89818 URS |
2344 | if (WARN_ON_ONCE(!list_empty(&va->list))) |
2345 | return; | |
dd3b8353 | 2346 | |
4d36e6f8 URS |
2347 | nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> |
2348 | PAGE_SHIFT, &vmap_lazy_nr); | |
80c4bd7a | 2349 | |
96e2db45 | 2350 | /* |
72210662 URS |
2351 | * If it was request by a certain node we would like to |
2352 | * return it to that node, i.e. its pool for later reuse. | |
96e2db45 | 2353 | */ |
72210662 URS |
2354 | vn = is_vn_id_valid(vn_id) ? |
2355 | id_to_node(vn_id):addr_to_node(va->va_start); | |
2356 | ||
282631cb | 2357 | spin_lock(&vn->lazy.lock); |
72210662 | 2358 | insert_vmap_area(va, &vn->lazy.root, &vn->lazy.head); |
282631cb | 2359 | spin_unlock(&vn->lazy.lock); |
80c4bd7a | 2360 | |
8c4196fe URS |
2361 | trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max); |
2362 | ||
96e2db45 | 2363 | /* After this point, we may free va at any time */ |
8c4196fe | 2364 | if (unlikely(nr_lazy > nr_lazy_max)) |
690467c8 | 2365 | schedule_work(&drain_vmap_work); |
db64fe02 NP |
2366 | } |
2367 | ||
b29acbdc NP |
2368 | /* |
2369 | * Free and unmap a vmap area | |
2370 | */ | |
2371 | static void free_unmap_vmap_area(struct vmap_area *va) | |
2372 | { | |
2373 | flush_cache_vunmap(va->va_start, va->va_end); | |
4ad0ae8c | 2374 | vunmap_range_noflush(va->va_start, va->va_end); |
8e57f8ac | 2375 | if (debug_pagealloc_enabled_static()) |
82a2e924 CP |
2376 | flush_tlb_kernel_range(va->va_start, va->va_end); |
2377 | ||
c8eef01e | 2378 | free_vmap_area_noflush(va); |
b29acbdc NP |
2379 | } |
2380 | ||
993d0b28 | 2381 | struct vmap_area *find_vmap_area(unsigned long addr) |
db64fe02 | 2382 | { |
d0936029 | 2383 | struct vmap_node *vn; |
db64fe02 | 2384 | struct vmap_area *va; |
d0936029 | 2385 | int i, j; |
db64fe02 | 2386 | |
4ed91fa9 URS |
2387 | if (unlikely(!vmap_initialized)) |
2388 | return NULL; | |
2389 | ||
d0936029 URS |
2390 | /* |
2391 | * An addr_to_node_id(addr) converts an address to a node index | |
2392 | * where a VA is located. If VA spans several zones and passed | |
2393 | * addr is not the same as va->va_start, what is not common, we | |
15e02a39 | 2394 | * may need to scan extra nodes. See an example: |
d0936029 | 2395 | * |
15e02a39 | 2396 | * <----va----> |
d0936029 URS |
2397 | * -|-----|-----|-----|-----|- |
2398 | * 1 2 0 1 | |
2399 | * | |
15e02a39 URS |
2400 | * VA resides in node 1 whereas it spans 1, 2 an 0. If passed |
2401 | * addr is within 2 or 0 nodes we should do extra work. | |
d0936029 URS |
2402 | */ |
2403 | i = j = addr_to_node_id(addr); | |
2404 | do { | |
2405 | vn = &vmap_nodes[i]; | |
db64fe02 | 2406 | |
d0936029 URS |
2407 | spin_lock(&vn->busy.lock); |
2408 | va = __find_vmap_area(addr, &vn->busy.root); | |
2409 | spin_unlock(&vn->busy.lock); | |
2410 | ||
2411 | if (va) | |
2412 | return va; | |
2413 | } while ((i = (i + 1) % nr_vmap_nodes) != j); | |
2414 | ||
2415 | return NULL; | |
db64fe02 NP |
2416 | } |
2417 | ||
edd89818 URS |
2418 | static struct vmap_area *find_unlink_vmap_area(unsigned long addr) |
2419 | { | |
d0936029 | 2420 | struct vmap_node *vn; |
edd89818 | 2421 | struct vmap_area *va; |
d0936029 | 2422 | int i, j; |
edd89818 | 2423 | |
15e02a39 URS |
2424 | /* |
2425 | * Check the comment in the find_vmap_area() about the loop. | |
2426 | */ | |
d0936029 URS |
2427 | i = j = addr_to_node_id(addr); |
2428 | do { | |
2429 | vn = &vmap_nodes[i]; | |
edd89818 | 2430 | |
d0936029 URS |
2431 | spin_lock(&vn->busy.lock); |
2432 | va = __find_vmap_area(addr, &vn->busy.root); | |
2433 | if (va) | |
2434 | unlink_va(va, &vn->busy.root); | |
2435 | spin_unlock(&vn->busy.lock); | |
2436 | ||
2437 | if (va) | |
2438 | return va; | |
2439 | } while ((i = (i + 1) % nr_vmap_nodes) != j); | |
2440 | ||
2441 | return NULL; | |
edd89818 URS |
2442 | } |
2443 | ||
db64fe02 NP |
2444 | /*** Per cpu kva allocator ***/ |
2445 | ||
2446 | /* | |
2447 | * vmap space is limited especially on 32 bit architectures. Ensure there is | |
2448 | * room for at least 16 percpu vmap blocks per CPU. | |
2449 | */ | |
2450 | /* | |
2451 | * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able | |
2452 | * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess | |
2453 | * instead (we just need a rough idea) | |
2454 | */ | |
2455 | #if BITS_PER_LONG == 32 | |
2456 | #define VMALLOC_SPACE (128UL*1024*1024) | |
2457 | #else | |
2458 | #define VMALLOC_SPACE (128UL*1024*1024*1024) | |
2459 | #endif | |
2460 | ||
2461 | #define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE) | |
2462 | #define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */ | |
2463 | #define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */ | |
2464 | #define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2) | |
2465 | #define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */ | |
2466 | #define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */ | |
f982f915 CL |
2467 | #define VMAP_BBMAP_BITS \ |
2468 | VMAP_MIN(VMAP_BBMAP_BITS_MAX, \ | |
2469 | VMAP_MAX(VMAP_BBMAP_BITS_MIN, \ | |
2470 | VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16)) | |
db64fe02 NP |
2471 | |
2472 | #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE) | |
2473 | ||
77e50af0 TG |
2474 | /* |
2475 | * Purge threshold to prevent overeager purging of fragmented blocks for | |
2476 | * regular operations: Purge if vb->free is less than 1/4 of the capacity. | |
2477 | */ | |
2478 | #define VMAP_PURGE_THRESHOLD (VMAP_BBMAP_BITS / 4) | |
2479 | ||
869176a0 BH |
2480 | #define VMAP_RAM 0x1 /* indicates vm_map_ram area*/ |
2481 | #define VMAP_BLOCK 0x2 /* mark out the vmap_block sub-type*/ | |
2482 | #define VMAP_FLAGS_MASK 0x3 | |
2483 | ||
db64fe02 NP |
2484 | struct vmap_block_queue { |
2485 | spinlock_t lock; | |
2486 | struct list_head free; | |
062eacf5 URS |
2487 | |
2488 | /* | |
2489 | * An xarray requires an extra memory dynamically to | |
2490 | * be allocated. If it is an issue, we can use rb-tree | |
2491 | * instead. | |
2492 | */ | |
2493 | struct xarray vmap_blocks; | |
db64fe02 NP |
2494 | }; |
2495 | ||
2496 | struct vmap_block { | |
2497 | spinlock_t lock; | |
2498 | struct vmap_area *va; | |
db64fe02 | 2499 | unsigned long free, dirty; |
d76f9954 | 2500 | DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS); |
7d61bfe8 | 2501 | unsigned long dirty_min, dirty_max; /*< dirty range */ |
de560423 NP |
2502 | struct list_head free_list; |
2503 | struct rcu_head rcu_head; | |
02b709df | 2504 | struct list_head purge; |
8c61291f | 2505 | unsigned int cpu; |
db64fe02 NP |
2506 | }; |
2507 | ||
2508 | /* Queue of free and dirty vmap blocks, for allocation and flushing purposes */ | |
2509 | static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue); | |
2510 | ||
2511 | /* | |
062eacf5 URS |
2512 | * In order to fast access to any "vmap_block" associated with a |
2513 | * specific address, we use a hash. | |
2514 | * | |
2515 | * A per-cpu vmap_block_queue is used in both ways, to serialize | |
2516 | * an access to free block chains among CPUs(alloc path) and it | |
2517 | * also acts as a vmap_block hash(alloc/free paths). It means we | |
2518 | * overload it, since we already have the per-cpu array which is | |
2519 | * used as a hash table. When used as a hash a 'cpu' passed to | |
2520 | * per_cpu() is not actually a CPU but rather a hash index. | |
2521 | * | |
fa1c77c1 | 2522 | * A hash function is addr_to_vb_xa() which hashes any address |
062eacf5 URS |
2523 | * to a specific index(in a hash) it belongs to. This then uses a |
2524 | * per_cpu() macro to access an array with generated index. | |
2525 | * | |
2526 | * An example: | |
2527 | * | |
2528 | * CPU_1 CPU_2 CPU_0 | |
2529 | * | | | | |
2530 | * V V V | |
2531 | * 0 10 20 30 40 50 60 | |
2532 | * |------|------|------|------|------|------|...<vmap address space> | |
2533 | * CPU0 CPU1 CPU2 CPU0 CPU1 CPU2 | |
2534 | * | |
2535 | * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus | |
2536 | * it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock; | |
2537 | * | |
2538 | * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus | |
2539 | * it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock; | |
2540 | * | |
2541 | * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus | |
2542 | * it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock. | |
2543 | * | |
2544 | * This technique almost always avoids lock contention on insert/remove, | |
2545 | * however xarray spinlocks protect against any contention that remains. | |
db64fe02 | 2546 | */ |
062eacf5 | 2547 | static struct xarray * |
fa1c77c1 | 2548 | addr_to_vb_xa(unsigned long addr) |
062eacf5 | 2549 | { |
a34acf30 URS |
2550 | int index = (addr / VMAP_BLOCK_SIZE) % nr_cpu_ids; |
2551 | ||
2552 | /* | |
2553 | * Please note, nr_cpu_ids points on a highest set | |
2554 | * possible bit, i.e. we never invoke cpumask_next() | |
2555 | * if an index points on it which is nr_cpu_ids - 1. | |
2556 | */ | |
2557 | if (!cpu_possible(index)) | |
2558 | index = cpumask_next(index, cpu_possible_mask); | |
062eacf5 URS |
2559 | |
2560 | return &per_cpu(vmap_block_queue, index).vmap_blocks; | |
2561 | } | |
db64fe02 NP |
2562 | |
2563 | /* | |
2564 | * We should probably have a fallback mechanism to allocate virtual memory | |
2565 | * out of partially filled vmap blocks. However vmap block sizing should be | |
2566 | * fairly reasonable according to the vmalloc size, so it shouldn't be a | |
2567 | * big problem. | |
2568 | */ | |
2569 | ||
2570 | static unsigned long addr_to_vb_idx(unsigned long addr) | |
2571 | { | |
2572 | addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1); | |
2573 | addr /= VMAP_BLOCK_SIZE; | |
2574 | return addr; | |
2575 | } | |
2576 | ||
cf725ce2 RP |
2577 | static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off) |
2578 | { | |
2579 | unsigned long addr; | |
2580 | ||
2581 | addr = va_start + (pages_off << PAGE_SHIFT); | |
2582 | BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start)); | |
2583 | return (void *)addr; | |
2584 | } | |
2585 | ||
2586 | /** | |
2587 | * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this | |
2588 | * block. Of course pages number can't exceed VMAP_BBMAP_BITS | |
2589 | * @order: how many 2^order pages should be occupied in newly allocated block | |
2590 | * @gfp_mask: flags for the page level allocator | |
2591 | * | |
a862f68a | 2592 | * Return: virtual address in a newly allocated block or ERR_PTR(-errno) |
cf725ce2 RP |
2593 | */ |
2594 | static void *new_vmap_block(unsigned int order, gfp_t gfp_mask) | |
db64fe02 NP |
2595 | { |
2596 | struct vmap_block_queue *vbq; | |
2597 | struct vmap_block *vb; | |
2598 | struct vmap_area *va; | |
062eacf5 | 2599 | struct xarray *xa; |
db64fe02 NP |
2600 | unsigned long vb_idx; |
2601 | int node, err; | |
cf725ce2 | 2602 | void *vaddr; |
db64fe02 NP |
2603 | |
2604 | node = numa_node_id(); | |
2605 | ||
2606 | vb = kmalloc_node(sizeof(struct vmap_block), | |
2607 | gfp_mask & GFP_RECLAIM_MASK, node); | |
2608 | if (unlikely(!vb)) | |
2609 | return ERR_PTR(-ENOMEM); | |
2610 | ||
2611 | va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE, | |
2612 | VMALLOC_START, VMALLOC_END, | |
869176a0 | 2613 | node, gfp_mask, |
4b68a773 | 2614 | VMAP_RAM|VMAP_BLOCK, NULL); |
ddf9c6d4 | 2615 | if (IS_ERR(va)) { |
db64fe02 | 2616 | kfree(vb); |
e7d86340 | 2617 | return ERR_CAST(va); |
db64fe02 NP |
2618 | } |
2619 | ||
cf725ce2 | 2620 | vaddr = vmap_block_vaddr(va->va_start, 0); |
db64fe02 NP |
2621 | spin_lock_init(&vb->lock); |
2622 | vb->va = va; | |
cf725ce2 RP |
2623 | /* At least something should be left free */ |
2624 | BUG_ON(VMAP_BBMAP_BITS <= (1UL << order)); | |
d76f9954 | 2625 | bitmap_zero(vb->used_map, VMAP_BBMAP_BITS); |
cf725ce2 | 2626 | vb->free = VMAP_BBMAP_BITS - (1UL << order); |
db64fe02 | 2627 | vb->dirty = 0; |
7d61bfe8 RP |
2628 | vb->dirty_min = VMAP_BBMAP_BITS; |
2629 | vb->dirty_max = 0; | |
d76f9954 | 2630 | bitmap_set(vb->used_map, 0, (1UL << order)); |
db64fe02 | 2631 | INIT_LIST_HEAD(&vb->free_list); |
3e3de794 | 2632 | vb->cpu = raw_smp_processor_id(); |
db64fe02 | 2633 | |
fa1c77c1 | 2634 | xa = addr_to_vb_xa(va->va_start); |
db64fe02 | 2635 | vb_idx = addr_to_vb_idx(va->va_start); |
062eacf5 | 2636 | err = xa_insert(xa, vb_idx, vb, gfp_mask); |
0f14599c MWO |
2637 | if (err) { |
2638 | kfree(vb); | |
2639 | free_vmap_area(va); | |
2640 | return ERR_PTR(err); | |
2641 | } | |
8c61291f ZH |
2642 | /* |
2643 | * list_add_tail_rcu could happened in another core | |
2644 | * rather than vb->cpu due to task migration, which | |
2645 | * is safe as list_add_tail_rcu will ensure the list's | |
2646 | * integrity together with list_for_each_rcu from read | |
2647 | * side. | |
2648 | */ | |
8c61291f | 2649 | vbq = per_cpu_ptr(&vmap_block_queue, vb->cpu); |
db64fe02 | 2650 | spin_lock(&vbq->lock); |
68ac546f | 2651 | list_add_tail_rcu(&vb->free_list, &vbq->free); |
db64fe02 | 2652 | spin_unlock(&vbq->lock); |
db64fe02 | 2653 | |
cf725ce2 | 2654 | return vaddr; |
db64fe02 NP |
2655 | } |
2656 | ||
db64fe02 NP |
2657 | static void free_vmap_block(struct vmap_block *vb) |
2658 | { | |
d0936029 | 2659 | struct vmap_node *vn; |
db64fe02 | 2660 | struct vmap_block *tmp; |
062eacf5 | 2661 | struct xarray *xa; |
db64fe02 | 2662 | |
fa1c77c1 | 2663 | xa = addr_to_vb_xa(vb->va->va_start); |
062eacf5 | 2664 | tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start)); |
db64fe02 NP |
2665 | BUG_ON(tmp != vb); |
2666 | ||
d0936029 URS |
2667 | vn = addr_to_node(vb->va->va_start); |
2668 | spin_lock(&vn->busy.lock); | |
2669 | unlink_va(vb->va, &vn->busy.root); | |
2670 | spin_unlock(&vn->busy.lock); | |
edd89818 | 2671 | |
64141da5 | 2672 | free_vmap_area_noflush(vb->va); |
22a3c7d1 | 2673 | kfree_rcu(vb, rcu_head); |
db64fe02 NP |
2674 | } |
2675 | ||
ca5e46c3 | 2676 | static bool purge_fragmented_block(struct vmap_block *vb, |
8c61291f | 2677 | struct list_head *purge_list, bool force_purge) |
ca5e46c3 | 2678 | { |
8c61291f ZH |
2679 | struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, vb->cpu); |
2680 | ||
ca5e46c3 TG |
2681 | if (vb->free + vb->dirty != VMAP_BBMAP_BITS || |
2682 | vb->dirty == VMAP_BBMAP_BITS) | |
2683 | return false; | |
2684 | ||
77e50af0 TG |
2685 | /* Don't overeagerly purge usable blocks unless requested */ |
2686 | if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD)) | |
2687 | return false; | |
2688 | ||
ca5e46c3 | 2689 | /* prevent further allocs after releasing lock */ |
7f48121e | 2690 | WRITE_ONCE(vb->free, 0); |
ca5e46c3 | 2691 | /* prevent purging it again */ |
7f48121e | 2692 | WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS); |
ca5e46c3 TG |
2693 | vb->dirty_min = 0; |
2694 | vb->dirty_max = VMAP_BBMAP_BITS; | |
2695 | spin_lock(&vbq->lock); | |
2696 | list_del_rcu(&vb->free_list); | |
2697 | spin_unlock(&vbq->lock); | |
2698 | list_add_tail(&vb->purge, purge_list); | |
2699 | return true; | |
2700 | } | |
2701 | ||
2702 | static void free_purged_blocks(struct list_head *purge_list) | |
2703 | { | |
2704 | struct vmap_block *vb, *n_vb; | |
2705 | ||
2706 | list_for_each_entry_safe(vb, n_vb, purge_list, purge) { | |
2707 | list_del(&vb->purge); | |
2708 | free_vmap_block(vb); | |
2709 | } | |
2710 | } | |
2711 | ||
02b709df NP |
2712 | static void purge_fragmented_blocks(int cpu) |
2713 | { | |
2714 | LIST_HEAD(purge); | |
2715 | struct vmap_block *vb; | |
02b709df NP |
2716 | struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu); |
2717 | ||
2718 | rcu_read_lock(); | |
2719 | list_for_each_entry_rcu(vb, &vbq->free, free_list) { | |
7f48121e TG |
2720 | unsigned long free = READ_ONCE(vb->free); |
2721 | unsigned long dirty = READ_ONCE(vb->dirty); | |
2722 | ||
2723 | if (free + dirty != VMAP_BBMAP_BITS || | |
2724 | dirty == VMAP_BBMAP_BITS) | |
02b709df NP |
2725 | continue; |
2726 | ||
2727 | spin_lock(&vb->lock); | |
8c61291f | 2728 | purge_fragmented_block(vb, &purge, true); |
ca5e46c3 | 2729 | spin_unlock(&vb->lock); |
02b709df NP |
2730 | } |
2731 | rcu_read_unlock(); | |
ca5e46c3 | 2732 | free_purged_blocks(&purge); |
02b709df NP |
2733 | } |
2734 | ||
02b709df NP |
2735 | static void purge_fragmented_blocks_allcpus(void) |
2736 | { | |
2737 | int cpu; | |
2738 | ||
2739 | for_each_possible_cpu(cpu) | |
2740 | purge_fragmented_blocks(cpu); | |
2741 | } | |
2742 | ||
db64fe02 NP |
2743 | static void *vb_alloc(unsigned long size, gfp_t gfp_mask) |
2744 | { | |
2745 | struct vmap_block_queue *vbq; | |
2746 | struct vmap_block *vb; | |
cf725ce2 | 2747 | void *vaddr = NULL; |
db64fe02 NP |
2748 | unsigned int order; |
2749 | ||
891c49ab | 2750 | BUG_ON(offset_in_page(size)); |
db64fe02 | 2751 | BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC); |
aa91c4d8 JK |
2752 | if (WARN_ON(size == 0)) { |
2753 | /* | |
2754 | * Allocating 0 bytes isn't what caller wants since | |
2755 | * get_order(0) returns funny result. Just warn and terminate | |
2756 | * early. | |
2757 | */ | |
ac0476e8 | 2758 | return ERR_PTR(-EINVAL); |
aa91c4d8 | 2759 | } |
db64fe02 NP |
2760 | order = get_order(size); |
2761 | ||
db64fe02 | 2762 | rcu_read_lock(); |
3f804920 | 2763 | vbq = raw_cpu_ptr(&vmap_block_queue); |
db64fe02 | 2764 | list_for_each_entry_rcu(vb, &vbq->free, free_list) { |
cf725ce2 | 2765 | unsigned long pages_off; |
db64fe02 | 2766 | |
43d76502 TG |
2767 | if (READ_ONCE(vb->free) < (1UL << order)) |
2768 | continue; | |
2769 | ||
db64fe02 | 2770 | spin_lock(&vb->lock); |
cf725ce2 RP |
2771 | if (vb->free < (1UL << order)) { |
2772 | spin_unlock(&vb->lock); | |
2773 | continue; | |
2774 | } | |
02b709df | 2775 | |
cf725ce2 RP |
2776 | pages_off = VMAP_BBMAP_BITS - vb->free; |
2777 | vaddr = vmap_block_vaddr(vb->va->va_start, pages_off); | |
43d76502 | 2778 | WRITE_ONCE(vb->free, vb->free - (1UL << order)); |
d76f9954 | 2779 | bitmap_set(vb->used_map, pages_off, (1UL << order)); |
02b709df NP |
2780 | if (vb->free == 0) { |
2781 | spin_lock(&vbq->lock); | |
2782 | list_del_rcu(&vb->free_list); | |
2783 | spin_unlock(&vbq->lock); | |
2784 | } | |
cf725ce2 | 2785 | |
02b709df NP |
2786 | spin_unlock(&vb->lock); |
2787 | break; | |
db64fe02 | 2788 | } |
02b709df | 2789 | |
db64fe02 NP |
2790 | rcu_read_unlock(); |
2791 | ||
cf725ce2 RP |
2792 | /* Allocate new block if nothing was found */ |
2793 | if (!vaddr) | |
2794 | vaddr = new_vmap_block(order, gfp_mask); | |
db64fe02 | 2795 | |
cf725ce2 | 2796 | return vaddr; |
db64fe02 NP |
2797 | } |
2798 | ||
78a0e8c4 | 2799 | static void vb_free(unsigned long addr, unsigned long size) |
db64fe02 NP |
2800 | { |
2801 | unsigned long offset; | |
db64fe02 NP |
2802 | unsigned int order; |
2803 | struct vmap_block *vb; | |
062eacf5 | 2804 | struct xarray *xa; |
db64fe02 | 2805 | |
891c49ab | 2806 | BUG_ON(offset_in_page(size)); |
db64fe02 | 2807 | BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC); |
b29acbdc | 2808 | |
78a0e8c4 | 2809 | flush_cache_vunmap(addr, addr + size); |
b29acbdc | 2810 | |
db64fe02 | 2811 | order = get_order(size); |
78a0e8c4 | 2812 | offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT; |
062eacf5 | 2813 | |
fa1c77c1 | 2814 | xa = addr_to_vb_xa(addr); |
062eacf5 URS |
2815 | vb = xa_load(xa, addr_to_vb_idx(addr)); |
2816 | ||
d76f9954 BH |
2817 | spin_lock(&vb->lock); |
2818 | bitmap_clear(vb->used_map, offset, (1UL << order)); | |
2819 | spin_unlock(&vb->lock); | |
db64fe02 | 2820 | |
4ad0ae8c | 2821 | vunmap_range_noflush(addr, addr + size); |
64141da5 | 2822 | |
8e57f8ac | 2823 | if (debug_pagealloc_enabled_static()) |
78a0e8c4 | 2824 | flush_tlb_kernel_range(addr, addr + size); |
82a2e924 | 2825 | |
db64fe02 | 2826 | spin_lock(&vb->lock); |
7d61bfe8 | 2827 | |
a09fad96 | 2828 | /* Expand the not yet TLB flushed dirty range */ |
7d61bfe8 RP |
2829 | vb->dirty_min = min(vb->dirty_min, offset); |
2830 | vb->dirty_max = max(vb->dirty_max, offset + (1UL << order)); | |
d086817d | 2831 | |
7f48121e | 2832 | WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order)); |
db64fe02 | 2833 | if (vb->dirty == VMAP_BBMAP_BITS) { |
de560423 | 2834 | BUG_ON(vb->free); |
db64fe02 NP |
2835 | spin_unlock(&vb->lock); |
2836 | free_vmap_block(vb); | |
2837 | } else | |
2838 | spin_unlock(&vb->lock); | |
2839 | } | |
2840 | ||
868b104d | 2841 | static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush) |
db64fe02 | 2842 | { |
ca5e46c3 | 2843 | LIST_HEAD(purge_list); |
db64fe02 | 2844 | int cpu; |
db64fe02 | 2845 | |
9b463334 JF |
2846 | if (unlikely(!vmap_initialized)) |
2847 | return; | |
2848 | ||
ca5e46c3 | 2849 | mutex_lock(&vmap_purge_lock); |
5803ed29 | 2850 | |
db64fe02 NP |
2851 | for_each_possible_cpu(cpu) { |
2852 | struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu); | |
2853 | struct vmap_block *vb; | |
fc1e0d98 | 2854 | unsigned long idx; |
db64fe02 NP |
2855 | |
2856 | rcu_read_lock(); | |
fc1e0d98 | 2857 | xa_for_each(&vbq->vmap_blocks, idx, vb) { |
db64fe02 | 2858 | spin_lock(&vb->lock); |
ca5e46c3 TG |
2859 | |
2860 | /* | |
2861 | * Try to purge a fragmented block first. If it's | |
2862 | * not purgeable, check whether there is dirty | |
2863 | * space to be flushed. | |
2864 | */ | |
8c61291f | 2865 | if (!purge_fragmented_block(vb, &purge_list, false) && |
a09fad96 | 2866 | vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) { |
7d61bfe8 | 2867 | unsigned long va_start = vb->va->va_start; |
db64fe02 | 2868 | unsigned long s, e; |
b136be5e | 2869 | |
7d61bfe8 RP |
2870 | s = va_start + (vb->dirty_min << PAGE_SHIFT); |
2871 | e = va_start + (vb->dirty_max << PAGE_SHIFT); | |
db64fe02 | 2872 | |
7d61bfe8 RP |
2873 | start = min(s, start); |
2874 | end = max(e, end); | |
db64fe02 | 2875 | |
a09fad96 TG |
2876 | /* Prevent that this is flushed again */ |
2877 | vb->dirty_min = VMAP_BBMAP_BITS; | |
2878 | vb->dirty_max = 0; | |
2879 | ||
7d61bfe8 | 2880 | flush = 1; |
db64fe02 NP |
2881 | } |
2882 | spin_unlock(&vb->lock); | |
2883 | } | |
2884 | rcu_read_unlock(); | |
2885 | } | |
ca5e46c3 | 2886 | free_purged_blocks(&purge_list); |
db64fe02 | 2887 | |
72210662 | 2888 | if (!__purge_vmap_area_lazy(start, end, false) && flush) |
0574ecd1 | 2889 | flush_tlb_kernel_range(start, end); |
f9e09977 | 2890 | mutex_unlock(&vmap_purge_lock); |
db64fe02 | 2891 | } |
868b104d RE |
2892 | |
2893 | /** | |
2894 | * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer | |
2895 | * | |
2896 | * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily | |
2897 | * to amortize TLB flushing overheads. What this means is that any page you | |
2898 | * have now, may, in a former life, have been mapped into kernel virtual | |
2899 | * address by the vmap layer and so there might be some CPUs with TLB entries | |
2900 | * still referencing that page (additional to the regular 1:1 kernel mapping). | |
2901 | * | |
2902 | * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can | |
2903 | * be sure that none of the pages we have control over will have any aliases | |
2904 | * from the vmap layer. | |
2905 | */ | |
2906 | void vm_unmap_aliases(void) | |
2907 | { | |
2908 | unsigned long start = ULONG_MAX, end = 0; | |
2909 | int flush = 0; | |
2910 | ||
2911 | _vm_unmap_aliases(start, end, flush); | |
2912 | } | |
db64fe02 NP |
2913 | EXPORT_SYMBOL_GPL(vm_unmap_aliases); |
2914 | ||
2915 | /** | |
2916 | * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram | |
2917 | * @mem: the pointer returned by vm_map_ram | |
2918 | * @count: the count passed to that vm_map_ram call (cannot unmap partial) | |
2919 | */ | |
2920 | void vm_unmap_ram(const void *mem, unsigned int count) | |
2921 | { | |
65ee03c4 | 2922 | unsigned long size = (unsigned long)count << PAGE_SHIFT; |
4aff1dc4 | 2923 | unsigned long addr = (unsigned long)kasan_reset_tag(mem); |
9c3acf60 | 2924 | struct vmap_area *va; |
db64fe02 | 2925 | |
5803ed29 | 2926 | might_sleep(); |
db64fe02 NP |
2927 | BUG_ON(!addr); |
2928 | BUG_ON(addr < VMALLOC_START); | |
2929 | BUG_ON(addr > VMALLOC_END); | |
a1c0b1a0 | 2930 | BUG_ON(!PAGE_ALIGNED(addr)); |
db64fe02 | 2931 | |
d98c9e83 AR |
2932 | kasan_poison_vmalloc(mem, size); |
2933 | ||
9c3acf60 | 2934 | if (likely(count <= VMAP_MAX_ALLOC)) { |
05e3ff95 | 2935 | debug_check_no_locks_freed(mem, size); |
78a0e8c4 | 2936 | vb_free(addr, size); |
9c3acf60 CH |
2937 | return; |
2938 | } | |
2939 | ||
edd89818 | 2940 | va = find_unlink_vmap_area(addr); |
14687619 URS |
2941 | if (WARN_ON_ONCE(!va)) |
2942 | return; | |
2943 | ||
05e3ff95 CP |
2944 | debug_check_no_locks_freed((void *)va->va_start, |
2945 | (va->va_end - va->va_start)); | |
9c3acf60 | 2946 | free_unmap_vmap_area(va); |
db64fe02 NP |
2947 | } |
2948 | EXPORT_SYMBOL(vm_unmap_ram); | |
2949 | ||
2950 | /** | |
2951 | * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space) | |
2952 | * @pages: an array of pointers to the pages to be mapped | |
2953 | * @count: number of pages | |
2954 | * @node: prefer to allocate data structures on this node | |
e99c97ad | 2955 | * |
36437638 GK |
2956 | * If you use this function for less than VMAP_MAX_ALLOC pages, it could be |
2957 | * faster than vmap so it's good. But if you mix long-life and short-life | |
2958 | * objects with vm_map_ram(), it could consume lots of address space through | |
2959 | * fragmentation (especially on a 32bit machine). You could see failures in | |
2960 | * the end. Please use this function for short-lived objects. | |
2961 | * | |
e99c97ad | 2962 | * Returns: a pointer to the address that has been mapped, or %NULL on failure |
db64fe02 | 2963 | */ |
d4efd79a | 2964 | void *vm_map_ram(struct page **pages, unsigned int count, int node) |
db64fe02 | 2965 | { |
65ee03c4 | 2966 | unsigned long size = (unsigned long)count << PAGE_SHIFT; |
db64fe02 NP |
2967 | unsigned long addr; |
2968 | void *mem; | |
2969 | ||
2970 | if (likely(count <= VMAP_MAX_ALLOC)) { | |
2971 | mem = vb_alloc(size, GFP_KERNEL); | |
2972 | if (IS_ERR(mem)) | |
2973 | return NULL; | |
2974 | addr = (unsigned long)mem; | |
2975 | } else { | |
2976 | struct vmap_area *va; | |
2977 | va = alloc_vmap_area(size, PAGE_SIZE, | |
869176a0 | 2978 | VMALLOC_START, VMALLOC_END, |
aaab830a | 2979 | node, GFP_KERNEL, VMAP_RAM, |
4b68a773 | 2980 | NULL); |
db64fe02 NP |
2981 | if (IS_ERR(va)) |
2982 | return NULL; | |
2983 | ||
2984 | addr = va->va_start; | |
2985 | mem = (void *)addr; | |
2986 | } | |
d98c9e83 | 2987 | |
b67177ec NP |
2988 | if (vmap_pages_range(addr, addr + size, PAGE_KERNEL, |
2989 | pages, PAGE_SHIFT) < 0) { | |
db64fe02 NP |
2990 | vm_unmap_ram(mem, count); |
2991 | return NULL; | |
2992 | } | |
b67177ec | 2993 | |
23689e91 AK |
2994 | /* |
2995 | * Mark the pages as accessible, now that they are mapped. | |
2996 | * With hardware tag-based KASAN, marking is skipped for | |
2997 | * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc(). | |
2998 | */ | |
f6e39794 | 2999 | mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL); |
19f1c3ac | 3000 | |
db64fe02 NP |
3001 | return mem; |
3002 | } | |
3003 | EXPORT_SYMBOL(vm_map_ram); | |
3004 | ||
4341fa45 | 3005 | static struct vm_struct *vmlist __initdata; |
92eac168 | 3006 | |
121e6f32 NP |
3007 | static inline unsigned int vm_area_page_order(struct vm_struct *vm) |
3008 | { | |
3009 | #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC | |
3010 | return vm->page_order; | |
3011 | #else | |
3012 | return 0; | |
3013 | #endif | |
3014 | } | |
3015 | ||
3016 | static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order) | |
3017 | { | |
3018 | #ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC | |
3019 | vm->page_order = order; | |
3020 | #else | |
3021 | BUG_ON(order != 0); | |
3022 | #endif | |
3023 | } | |
3024 | ||
be9b7335 NP |
3025 | /** |
3026 | * vm_area_add_early - add vmap area early during boot | |
3027 | * @vm: vm_struct to add | |
3028 | * | |
3029 | * This function is used to add fixed kernel vm area to vmlist before | |
3030 | * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags | |
3031 | * should contain proper values and the other fields should be zero. | |
3032 | * | |
3033 | * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING. | |
3034 | */ | |
3035 | void __init vm_area_add_early(struct vm_struct *vm) | |
3036 | { | |
3037 | struct vm_struct *tmp, **p; | |
3038 | ||
3039 | BUG_ON(vmap_initialized); | |
3040 | for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) { | |
3041 | if (tmp->addr >= vm->addr) { | |
3042 | BUG_ON(tmp->addr < vm->addr + vm->size); | |
3043 | break; | |
3044 | } else | |
3045 | BUG_ON(tmp->addr + tmp->size > vm->addr); | |
3046 | } | |
3047 | vm->next = *p; | |
3048 | *p = vm; | |
3049 | } | |
3050 | ||
f0aa6617 TH |
3051 | /** |
3052 | * vm_area_register_early - register vmap area early during boot | |
3053 | * @vm: vm_struct to register | |
c0c0a293 | 3054 | * @align: requested alignment |
f0aa6617 TH |
3055 | * |
3056 | * This function is used to register kernel vm area before | |
3057 | * vmalloc_init() is called. @vm->size and @vm->flags should contain | |
3058 | * proper values on entry and other fields should be zero. On return, | |
3059 | * vm->addr contains the allocated address. | |
3060 | * | |
3061 | * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING. | |
3062 | */ | |
c0c0a293 | 3063 | void __init vm_area_register_early(struct vm_struct *vm, size_t align) |
f0aa6617 | 3064 | { |
0eb68437 KW |
3065 | unsigned long addr = ALIGN(VMALLOC_START, align); |
3066 | struct vm_struct *cur, **p; | |
c0c0a293 | 3067 | |
0eb68437 | 3068 | BUG_ON(vmap_initialized); |
f0aa6617 | 3069 | |
0eb68437 KW |
3070 | for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) { |
3071 | if ((unsigned long)cur->addr - addr >= vm->size) | |
3072 | break; | |
3073 | addr = ALIGN((unsigned long)cur->addr + cur->size, align); | |
3074 | } | |
f0aa6617 | 3075 | |
0eb68437 KW |
3076 | BUG_ON(addr > VMALLOC_END - vm->size); |
3077 | vm->addr = (void *)addr; | |
3078 | vm->next = *p; | |
3079 | *p = vm; | |
3252b1d8 | 3080 | kasan_populate_early_vm_area_shadow(vm->addr, vm->size); |
f0aa6617 TH |
3081 | } |
3082 | ||
20fc02b4 | 3083 | static void clear_vm_uninitialized_flag(struct vm_struct *vm) |
f5252e00 | 3084 | { |
d4033afd | 3085 | /* |
20fc02b4 | 3086 | * Before removing VM_UNINITIALIZED, |
d4033afd JK |
3087 | * we should make sure that vm has proper values. |
3088 | * Pair with smp_rmb() in show_numa_info(). | |
3089 | */ | |
3090 | smp_wmb(); | |
20fc02b4 | 3091 | vm->flags &= ~VM_UNINITIALIZED; |
cf88c790 TH |
3092 | } |
3093 | ||
db64fe02 | 3094 | static struct vm_struct *__get_vm_area_node(unsigned long size, |
7ca3027b DA |
3095 | unsigned long align, unsigned long shift, unsigned long flags, |
3096 | unsigned long start, unsigned long end, int node, | |
3097 | gfp_t gfp_mask, const void *caller) | |
db64fe02 | 3098 | { |
0006526d | 3099 | struct vmap_area *va; |
db64fe02 | 3100 | struct vm_struct *area; |
d98c9e83 | 3101 | unsigned long requested_size = size; |
1da177e4 | 3102 | |
52fd24ca | 3103 | BUG_ON(in_interrupt()); |
7ca3027b | 3104 | size = ALIGN(size, 1ul << shift); |
31be8309 OH |
3105 | if (unlikely(!size)) |
3106 | return NULL; | |
1da177e4 | 3107 | |
252e5c6e | 3108 | if (flags & VM_IOREMAP) |
3109 | align = 1ul << clamp_t(int, get_count_order_long(size), | |
3110 | PAGE_SHIFT, IOREMAP_MAX_ORDER); | |
3111 | ||
cf88c790 | 3112 | area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node); |
1da177e4 LT |
3113 | if (unlikely(!area)) |
3114 | return NULL; | |
3115 | ||
71394fe5 AR |
3116 | if (!(flags & VM_NO_GUARD)) |
3117 | size += PAGE_SIZE; | |
1da177e4 | 3118 | |
4b68a773 BH |
3119 | area->flags = flags; |
3120 | area->caller = caller; | |
3121 | ||
3122 | va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0, area); | |
db64fe02 NP |
3123 | if (IS_ERR(va)) { |
3124 | kfree(area); | |
3125 | return NULL; | |
1da177e4 | 3126 | } |
1da177e4 | 3127 | |
19f1c3ac AK |
3128 | /* |
3129 | * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a | |
3130 | * best-effort approach, as they can be mapped outside of vmalloc code. | |
3131 | * For VM_ALLOC mappings, the pages are marked as accessible after | |
3132 | * getting mapped in __vmalloc_node_range(). | |
23689e91 AK |
3133 | * With hardware tag-based KASAN, marking is skipped for |
3134 | * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc(). | |
19f1c3ac AK |
3135 | */ |
3136 | if (!(flags & VM_ALLOC)) | |
23689e91 | 3137 | area->addr = kasan_unpoison_vmalloc(area->addr, requested_size, |
f6e39794 | 3138 | KASAN_VMALLOC_PROT_NORMAL); |
1d96320f | 3139 | |
1da177e4 | 3140 | return area; |
1da177e4 LT |
3141 | } |
3142 | ||
c2968612 BH |
3143 | struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags, |
3144 | unsigned long start, unsigned long end, | |
5e6cafc8 | 3145 | const void *caller) |
c2968612 | 3146 | { |
7ca3027b DA |
3147 | return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end, |
3148 | NUMA_NO_NODE, GFP_KERNEL, caller); | |
c2968612 BH |
3149 | } |
3150 | ||
1da177e4 | 3151 | /** |
92eac168 MR |
3152 | * get_vm_area - reserve a contiguous kernel virtual area |
3153 | * @size: size of the area | |
3154 | * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC | |
1da177e4 | 3155 | * |
92eac168 MR |
3156 | * Search an area of @size in the kernel virtual mapping area, |
3157 | * and reserved it for out purposes. Returns the area descriptor | |
3158 | * on success or %NULL on failure. | |
a862f68a MR |
3159 | * |
3160 | * Return: the area descriptor on success or %NULL on failure. | |
1da177e4 LT |
3161 | */ |
3162 | struct vm_struct *get_vm_area(unsigned long size, unsigned long flags) | |
3163 | { | |
7ca3027b DA |
3164 | return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, |
3165 | VMALLOC_START, VMALLOC_END, | |
00ef2d2f DR |
3166 | NUMA_NO_NODE, GFP_KERNEL, |
3167 | __builtin_return_address(0)); | |
23016969 CL |
3168 | } |
3169 | ||
3170 | struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags, | |
5e6cafc8 | 3171 | const void *caller) |
23016969 | 3172 | { |
7ca3027b DA |
3173 | return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, |
3174 | VMALLOC_START, VMALLOC_END, | |
00ef2d2f | 3175 | NUMA_NO_NODE, GFP_KERNEL, caller); |
1da177e4 LT |
3176 | } |
3177 | ||
e9da6e99 | 3178 | /** |
92eac168 MR |
3179 | * find_vm_area - find a continuous kernel virtual area |
3180 | * @addr: base address | |
e9da6e99 | 3181 | * |
92eac168 MR |
3182 | * Search for the kernel VM area starting at @addr, and return it. |
3183 | * It is up to the caller to do all required locking to keep the returned | |
3184 | * pointer valid. | |
a862f68a | 3185 | * |
74640617 | 3186 | * Return: the area descriptor on success or %NULL on failure. |
e9da6e99 MS |
3187 | */ |
3188 | struct vm_struct *find_vm_area(const void *addr) | |
83342314 | 3189 | { |
db64fe02 | 3190 | struct vmap_area *va; |
83342314 | 3191 | |
db64fe02 | 3192 | va = find_vmap_area((unsigned long)addr); |
688fcbfc PL |
3193 | if (!va) |
3194 | return NULL; | |
1da177e4 | 3195 | |
688fcbfc | 3196 | return va->vm; |
1da177e4 LT |
3197 | } |
3198 | ||
7856dfeb | 3199 | /** |
92eac168 MR |
3200 | * remove_vm_area - find and remove a continuous kernel virtual area |
3201 | * @addr: base address | |
7856dfeb | 3202 | * |
92eac168 MR |
3203 | * Search for the kernel VM area starting at @addr, and remove it. |
3204 | * This function returns the found VM area, but using it is NOT safe | |
3205 | * on SMP machines, except for its size or flags. | |
a862f68a | 3206 | * |
74640617 | 3207 | * Return: the area descriptor on success or %NULL on failure. |
7856dfeb | 3208 | */ |
b3bdda02 | 3209 | struct vm_struct *remove_vm_area(const void *addr) |
7856dfeb | 3210 | { |
db64fe02 | 3211 | struct vmap_area *va; |
75c59ce7 | 3212 | struct vm_struct *vm; |
db64fe02 | 3213 | |
5803ed29 CH |
3214 | might_sleep(); |
3215 | ||
17d3ef43 CH |
3216 | if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n", |
3217 | addr)) | |
3218 | return NULL; | |
c69480ad | 3219 | |
75c59ce7 CH |
3220 | va = find_unlink_vmap_area((unsigned long)addr); |
3221 | if (!va || !va->vm) | |
3222 | return NULL; | |
3223 | vm = va->vm; | |
dd32c279 | 3224 | |
17d3ef43 CH |
3225 | debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm)); |
3226 | debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm)); | |
75c59ce7 | 3227 | kasan_free_module_shadow(vm); |
17d3ef43 | 3228 | kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm)); |
dd3b8353 | 3229 | |
75c59ce7 CH |
3230 | free_unmap_vmap_area(va); |
3231 | return vm; | |
7856dfeb AK |
3232 | } |
3233 | ||
868b104d RE |
3234 | static inline void set_area_direct_map(const struct vm_struct *area, |
3235 | int (*set_direct_map)(struct page *page)) | |
3236 | { | |
3237 | int i; | |
3238 | ||
121e6f32 | 3239 | /* HUGE_VMALLOC passes small pages to set_direct_map */ |
868b104d RE |
3240 | for (i = 0; i < area->nr_pages; i++) |
3241 | if (page_address(area->pages[i])) | |
3242 | set_direct_map(area->pages[i]); | |
3243 | } | |
3244 | ||
9e5fa0ae CH |
3245 | /* |
3246 | * Flush the vm mapping and reset the direct map. | |
3247 | */ | |
3248 | static void vm_reset_perms(struct vm_struct *area) | |
868b104d | 3249 | { |
868b104d | 3250 | unsigned long start = ULONG_MAX, end = 0; |
121e6f32 | 3251 | unsigned int page_order = vm_area_page_order(area); |
31e67340 | 3252 | int flush_dmap = 0; |
868b104d RE |
3253 | int i; |
3254 | ||
868b104d | 3255 | /* |
9e5fa0ae | 3256 | * Find the start and end range of the direct mappings to make sure that |
868b104d RE |
3257 | * the vm_unmap_aliases() flush includes the direct map. |
3258 | */ | |
121e6f32 | 3259 | for (i = 0; i < area->nr_pages; i += 1U << page_order) { |
8e41f872 | 3260 | unsigned long addr = (unsigned long)page_address(area->pages[i]); |
9e5fa0ae | 3261 | |
8e41f872 | 3262 | if (addr) { |
121e6f32 NP |
3263 | unsigned long page_size; |
3264 | ||
3265 | page_size = PAGE_SIZE << page_order; | |
868b104d | 3266 | start = min(addr, start); |
121e6f32 | 3267 | end = max(addr + page_size, end); |
31e67340 | 3268 | flush_dmap = 1; |
868b104d RE |
3269 | } |
3270 | } | |
3271 | ||
3272 | /* | |
3273 | * Set direct map to something invalid so that it won't be cached if | |
3274 | * there are any accesses after the TLB flush, then flush the TLB and | |
3275 | * reset the direct map permissions to the default. | |
3276 | */ | |
3277 | set_area_direct_map(area, set_direct_map_invalid_noflush); | |
31e67340 | 3278 | _vm_unmap_aliases(start, end, flush_dmap); |
868b104d RE |
3279 | set_area_direct_map(area, set_direct_map_default_noflush); |
3280 | } | |
3281 | ||
208162f4 | 3282 | static void delayed_vfree_work(struct work_struct *w) |
1da177e4 | 3283 | { |
208162f4 CH |
3284 | struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq); |
3285 | struct llist_node *t, *llnode; | |
bf22e37a | 3286 | |
208162f4 | 3287 | llist_for_each_safe(llnode, t, llist_del_all(&p->list)) |
5d3d31d6 | 3288 | vfree(llnode); |
bf22e37a AR |
3289 | } |
3290 | ||
3291 | /** | |
92eac168 MR |
3292 | * vfree_atomic - release memory allocated by vmalloc() |
3293 | * @addr: memory base address | |
bf22e37a | 3294 | * |
92eac168 MR |
3295 | * This one is just like vfree() but can be called in any atomic context |
3296 | * except NMIs. | |
bf22e37a AR |
3297 | */ |
3298 | void vfree_atomic(const void *addr) | |
3299 | { | |
01e2e839 | 3300 | struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred); |
bf22e37a | 3301 | |
01e2e839 | 3302 | BUG_ON(in_nmi()); |
bf22e37a AR |
3303 | kmemleak_free(addr); |
3304 | ||
01e2e839 CH |
3305 | /* |
3306 | * Use raw_cpu_ptr() because this can be called from preemptible | |
3307 | * context. Preemption is absolutely fine here, because the llist_add() | |
3308 | * implementation is lockless, so it works even if we are adding to | |
3309 | * another cpu's list. schedule_work() should be fine with this too. | |
3310 | */ | |
3311 | if (addr && llist_add((struct llist_node *)addr, &p->list)) | |
3312 | schedule_work(&p->wq); | |
c67dc624 RP |
3313 | } |
3314 | ||
1da177e4 | 3315 | /** |
fa307474 MWO |
3316 | * vfree - Release memory allocated by vmalloc() |
3317 | * @addr: Memory base address | |
1da177e4 | 3318 | * |
fa307474 MWO |
3319 | * Free the virtually continuous memory area starting at @addr, as obtained |
3320 | * from one of the vmalloc() family of APIs. This will usually also free the | |
3321 | * physical memory underlying the virtual allocation, but that memory is | |
3322 | * reference counted, so it will not be freed until the last user goes away. | |
1da177e4 | 3323 | * |
fa307474 | 3324 | * If @addr is NULL, no operation is performed. |
c9fcee51 | 3325 | * |
fa307474 | 3326 | * Context: |
92eac168 | 3327 | * May sleep if called *not* from interrupt context. |
fa307474 MWO |
3328 | * Must not be called in NMI context (strictly speaking, it could be |
3329 | * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling | |
f0953a1b | 3330 | * conventions for vfree() arch-dependent would be a really bad idea). |
1da177e4 | 3331 | */ |
b3bdda02 | 3332 | void vfree(const void *addr) |
1da177e4 | 3333 | { |
79311c1f CH |
3334 | struct vm_struct *vm; |
3335 | int i; | |
89219d37 | 3336 | |
01e2e839 CH |
3337 | if (unlikely(in_interrupt())) { |
3338 | vfree_atomic(addr); | |
3339 | return; | |
3340 | } | |
89219d37 | 3341 | |
01e2e839 | 3342 | BUG_ON(in_nmi()); |
89219d37 | 3343 | kmemleak_free(addr); |
01e2e839 | 3344 | might_sleep(); |
a8dda165 | 3345 | |
32fcfd40 AV |
3346 | if (!addr) |
3347 | return; | |
c67dc624 | 3348 | |
79311c1f CH |
3349 | vm = remove_vm_area(addr); |
3350 | if (unlikely(!vm)) { | |
3351 | WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", | |
3352 | addr); | |
3353 | return; | |
3354 | } | |
3355 | ||
9e5fa0ae CH |
3356 | if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS)) |
3357 | vm_reset_perms(vm); | |
79311c1f CH |
3358 | for (i = 0; i < vm->nr_pages; i++) { |
3359 | struct page *page = vm->pages[i]; | |
3360 | ||
3361 | BUG_ON(!page); | |
3362 | mod_memcg_page_state(page, MEMCG_VMALLOC, -1); | |
3363 | /* | |
3364 | * High-order allocs for huge vmallocs are split, so | |
3365 | * can be freed as an array of order-0 allocations | |
3366 | */ | |
dcc1be11 | 3367 | __free_page(page); |
79311c1f CH |
3368 | cond_resched(); |
3369 | } | |
3370 | atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages); | |
3371 | kvfree(vm->pages); | |
3372 | kfree(vm); | |
1da177e4 | 3373 | } |
1da177e4 LT |
3374 | EXPORT_SYMBOL(vfree); |
3375 | ||
3376 | /** | |
92eac168 MR |
3377 | * vunmap - release virtual mapping obtained by vmap() |
3378 | * @addr: memory base address | |
1da177e4 | 3379 | * |
92eac168 MR |
3380 | * Free the virtually contiguous memory area starting at @addr, |
3381 | * which was created from the page array passed to vmap(). | |
1da177e4 | 3382 | * |
92eac168 | 3383 | * Must not be called in interrupt context. |
1da177e4 | 3384 | */ |
b3bdda02 | 3385 | void vunmap(const void *addr) |
1da177e4 | 3386 | { |
79311c1f CH |
3387 | struct vm_struct *vm; |
3388 | ||
1da177e4 | 3389 | BUG_ON(in_interrupt()); |
34754b69 | 3390 | might_sleep(); |
79311c1f CH |
3391 | |
3392 | if (!addr) | |
3393 | return; | |
3394 | vm = remove_vm_area(addr); | |
3395 | if (unlikely(!vm)) { | |
3396 | WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n", | |
3397 | addr); | |
3398 | return; | |
3399 | } | |
3400 | kfree(vm); | |
1da177e4 | 3401 | } |
1da177e4 LT |
3402 | EXPORT_SYMBOL(vunmap); |
3403 | ||
3404 | /** | |
92eac168 MR |
3405 | * vmap - map an array of pages into virtually contiguous space |
3406 | * @pages: array of page pointers | |
3407 | * @count: number of pages to map | |
3408 | * @flags: vm_area->flags | |
3409 | * @prot: page protection for the mapping | |
3410 | * | |
b944afc9 CH |
3411 | * Maps @count pages from @pages into contiguous kernel virtual space. |
3412 | * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself | |
3413 | * (which must be kmalloc or vmalloc memory) and one reference per pages in it | |
3414 | * are transferred from the caller to vmap(), and will be freed / dropped when | |
3415 | * vfree() is called on the return value. | |
a862f68a MR |
3416 | * |
3417 | * Return: the address of the area or %NULL on failure | |
1da177e4 LT |
3418 | */ |
3419 | void *vmap(struct page **pages, unsigned int count, | |
92eac168 | 3420 | unsigned long flags, pgprot_t prot) |
1da177e4 LT |
3421 | { |
3422 | struct vm_struct *area; | |
b67177ec | 3423 | unsigned long addr; |
65ee03c4 | 3424 | unsigned long size; /* In bytes */ |
1da177e4 | 3425 | |
34754b69 PZ |
3426 | might_sleep(); |
3427 | ||
37f3605e CH |
3428 | if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS)) |
3429 | return NULL; | |
3430 | ||
bd1a8fb2 PZ |
3431 | /* |
3432 | * Your top guard is someone else's bottom guard. Not having a top | |
3433 | * guard compromises someone else's mappings too. | |
3434 | */ | |
3435 | if (WARN_ON_ONCE(flags & VM_NO_GUARD)) | |
3436 | flags &= ~VM_NO_GUARD; | |
3437 | ||
ca79b0c2 | 3438 | if (count > totalram_pages()) |
1da177e4 LT |
3439 | return NULL; |
3440 | ||
65ee03c4 GJM |
3441 | size = (unsigned long)count << PAGE_SHIFT; |
3442 | area = get_vm_area_caller(size, flags, __builtin_return_address(0)); | |
1da177e4 LT |
3443 | if (!area) |
3444 | return NULL; | |
23016969 | 3445 | |
b67177ec NP |
3446 | addr = (unsigned long)area->addr; |
3447 | if (vmap_pages_range(addr, addr + size, pgprot_nx(prot), | |
3448 | pages, PAGE_SHIFT) < 0) { | |
1da177e4 LT |
3449 | vunmap(area->addr); |
3450 | return NULL; | |
3451 | } | |
3452 | ||
c22ee528 | 3453 | if (flags & VM_MAP_PUT_PAGES) { |
b944afc9 | 3454 | area->pages = pages; |
c22ee528 ML |
3455 | area->nr_pages = count; |
3456 | } | |
1da177e4 LT |
3457 | return area->addr; |
3458 | } | |
1da177e4 LT |
3459 | EXPORT_SYMBOL(vmap); |
3460 | ||
3e9a9e25 CH |
3461 | #ifdef CONFIG_VMAP_PFN |
3462 | struct vmap_pfn_data { | |
3463 | unsigned long *pfns; | |
3464 | pgprot_t prot; | |
3465 | unsigned int idx; | |
3466 | }; | |
3467 | ||
3468 | static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private) | |
3469 | { | |
3470 | struct vmap_pfn_data *data = private; | |
b3f78e74 RR |
3471 | unsigned long pfn = data->pfns[data->idx]; |
3472 | pte_t ptent; | |
3e9a9e25 | 3473 | |
b3f78e74 | 3474 | if (WARN_ON_ONCE(pfn_valid(pfn))) |
3e9a9e25 | 3475 | return -EINVAL; |
b3f78e74 RR |
3476 | |
3477 | ptent = pte_mkspecial(pfn_pte(pfn, data->prot)); | |
3478 | set_pte_at(&init_mm, addr, pte, ptent); | |
3479 | ||
3480 | data->idx++; | |
3e9a9e25 CH |
3481 | return 0; |
3482 | } | |
3483 | ||
3484 | /** | |
3485 | * vmap_pfn - map an array of PFNs into virtually contiguous space | |
3486 | * @pfns: array of PFNs | |
3487 | * @count: number of pages to map | |
3488 | * @prot: page protection for the mapping | |
3489 | * | |
3490 | * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns | |
3491 | * the start address of the mapping. | |
3492 | */ | |
3493 | void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot) | |
3494 | { | |
3495 | struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) }; | |
3496 | struct vm_struct *area; | |
3497 | ||
3498 | area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP, | |
3499 | __builtin_return_address(0)); | |
3500 | if (!area) | |
3501 | return NULL; | |
3502 | if (apply_to_page_range(&init_mm, (unsigned long)area->addr, | |
3503 | count * PAGE_SIZE, vmap_pfn_apply, &data)) { | |
3504 | free_vm_area(area); | |
3505 | return NULL; | |
3506 | } | |
a50420c7 AG |
3507 | |
3508 | flush_cache_vmap((unsigned long)area->addr, | |
3509 | (unsigned long)area->addr + count * PAGE_SIZE); | |
3510 | ||
3e9a9e25 CH |
3511 | return area->addr; |
3512 | } | |
3513 | EXPORT_SYMBOL_GPL(vmap_pfn); | |
3514 | #endif /* CONFIG_VMAP_PFN */ | |
3515 | ||
12b9f873 UR |
3516 | static inline unsigned int |
3517 | vm_area_alloc_pages(gfp_t gfp, int nid, | |
343ab817 | 3518 | unsigned int order, unsigned int nr_pages, struct page **pages) |
12b9f873 UR |
3519 | { |
3520 | unsigned int nr_allocated = 0; | |
e9c3cda4 | 3521 | gfp_t alloc_gfp = gfp; |
8e0545c8 | 3522 | bool nofail = gfp & __GFP_NOFAIL; |
ffb29b1c CW |
3523 | struct page *page; |
3524 | int i; | |
12b9f873 UR |
3525 | |
3526 | /* | |
3527 | * For order-0 pages we make use of bulk allocator, if | |
3528 | * the page array is partly or not at all populated due | |
3529 | * to fails, fallback to a single page allocator that is | |
3530 | * more permissive. | |
3531 | */ | |
c00b6b96 | 3532 | if (!order) { |
e9c3cda4 | 3533 | /* bulk allocator doesn't support nofail req. officially */ |
9376130c MH |
3534 | gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL; |
3535 | ||
343ab817 URS |
3536 | while (nr_allocated < nr_pages) { |
3537 | unsigned int nr, nr_pages_request; | |
3538 | ||
3539 | /* | |
3540 | * A maximum allowed request is hard-coded and is 100 | |
3541 | * pages per call. That is done in order to prevent a | |
3542 | * long preemption off scenario in the bulk-allocator | |
3543 | * so the range is [1:100]. | |
3544 | */ | |
3545 | nr_pages_request = min(100U, nr_pages - nr_allocated); | |
3546 | ||
c00b6b96 CW |
3547 | /* memory allocation should consider mempolicy, we can't |
3548 | * wrongly use nearest node when nid == NUMA_NO_NODE, | |
3549 | * otherwise memory may be allocated in only one node, | |
98af39d5 | 3550 | * but mempolicy wants to alloc memory by interleaving. |
c00b6b96 CW |
3551 | */ |
3552 | if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE) | |
88ae5fb7 | 3553 | nr = alloc_pages_bulk_array_mempolicy_noprof(bulk_gfp, |
c00b6b96 CW |
3554 | nr_pages_request, |
3555 | pages + nr_allocated); | |
3556 | ||
3557 | else | |
88ae5fb7 | 3558 | nr = alloc_pages_bulk_array_node_noprof(bulk_gfp, nid, |
c00b6b96 CW |
3559 | nr_pages_request, |
3560 | pages + nr_allocated); | |
343ab817 URS |
3561 | |
3562 | nr_allocated += nr; | |
3563 | cond_resched(); | |
3564 | ||
3565 | /* | |
3566 | * If zero or pages were obtained partly, | |
3567 | * fallback to a single page allocator. | |
3568 | */ | |
3569 | if (nr != nr_pages_request) | |
3570 | break; | |
3571 | } | |
e9c3cda4 MH |
3572 | } else if (gfp & __GFP_NOFAIL) { |
3573 | /* | |
3574 | * Higher order nofail allocations are really expensive and | |
3575 | * potentially dangerous (pre-mature OOM, disruptive reclaim | |
3576 | * and compaction etc. | |
3577 | */ | |
3578 | alloc_gfp &= ~__GFP_NOFAIL; | |
3b8000ae | 3579 | } |
12b9f873 UR |
3580 | |
3581 | /* High-order pages or fallback path if "bulk" fails. */ | |
ffb29b1c | 3582 | while (nr_allocated < nr_pages) { |
8e0545c8 | 3583 | if (!nofail && fatal_signal_pending(current)) |
dd544141 VA |
3584 | break; |
3585 | ||
ffb29b1c | 3586 | if (nid == NUMA_NO_NODE) |
88ae5fb7 | 3587 | page = alloc_pages_noprof(alloc_gfp, order); |
ffb29b1c | 3588 | else |
88ae5fb7 | 3589 | page = alloc_pages_node_noprof(nid, alloc_gfp, order); |
61ebe5a7 HL |
3590 | if (unlikely(!page)) |
3591 | break; | |
e9c3cda4 | 3592 | |
3b8000ae NP |
3593 | /* |
3594 | * Higher order allocations must be able to be treated as | |
3595 | * indepdenent small pages by callers (as they can with | |
3596 | * small-page vmallocs). Some drivers do their own refcounting | |
3597 | * on vmalloc_to_page() pages, some use page->mapping, | |
3598 | * page->lru, etc. | |
3599 | */ | |
3600 | if (order) | |
3601 | split_page(page, order); | |
12b9f873 UR |
3602 | |
3603 | /* | |
3604 | * Careful, we allocate and map page-order pages, but | |
3605 | * tracking is done per PAGE_SIZE page so as to keep the | |
3606 | * vm_struct APIs independent of the physical/mapped size. | |
3607 | */ | |
3608 | for (i = 0; i < (1U << order); i++) | |
3609 | pages[nr_allocated + i] = page + i; | |
3610 | ||
12e376a6 | 3611 | cond_resched(); |
12b9f873 UR |
3612 | nr_allocated += 1U << order; |
3613 | } | |
3614 | ||
3615 | return nr_allocated; | |
3616 | } | |
3617 | ||
e31d9eb5 | 3618 | static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask, |
121e6f32 NP |
3619 | pgprot_t prot, unsigned int page_shift, |
3620 | int node) | |
1da177e4 | 3621 | { |
930f036b | 3622 | const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO; |
9376130c | 3623 | bool nofail = gfp_mask & __GFP_NOFAIL; |
121e6f32 NP |
3624 | unsigned long addr = (unsigned long)area->addr; |
3625 | unsigned long size = get_vm_area_size(area); | |
34fe6537 | 3626 | unsigned long array_size; |
121e6f32 NP |
3627 | unsigned int nr_small_pages = size >> PAGE_SHIFT; |
3628 | unsigned int page_order; | |
451769eb MH |
3629 | unsigned int flags; |
3630 | int ret; | |
1da177e4 | 3631 | |
121e6f32 | 3632 | array_size = (unsigned long)nr_small_pages * sizeof(struct page *); |
80b1d8fd | 3633 | |
f255935b CH |
3634 | if (!(gfp_mask & (GFP_DMA | GFP_DMA32))) |
3635 | gfp_mask |= __GFP_HIGHMEM; | |
1da177e4 | 3636 | |
1da177e4 | 3637 | /* Please note that the recursion is strictly bounded. */ |
8757d5fa | 3638 | if (array_size > PAGE_SIZE) { |
88ae5fb7 | 3639 | area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node, |
f255935b | 3640 | area->caller); |
286e1ea3 | 3641 | } else { |
88ae5fb7 | 3642 | area->pages = kmalloc_node_noprof(array_size, nested_gfp, node); |
286e1ea3 | 3643 | } |
7ea36242 | 3644 | |
5c1f4e69 | 3645 | if (!area->pages) { |
c3d77172 | 3646 | warn_alloc(gfp_mask, NULL, |
f4bdfeaf URS |
3647 | "vmalloc error: size %lu, failed to allocated page array size %lu", |
3648 | nr_small_pages * PAGE_SIZE, array_size); | |
cd61413b | 3649 | free_vm_area(area); |
1da177e4 LT |
3650 | return NULL; |
3651 | } | |
1da177e4 | 3652 | |
121e6f32 | 3653 | set_vm_area_page_order(area, page_shift - PAGE_SHIFT); |
121e6f32 | 3654 | page_order = vm_area_page_order(area); |
bf53d6f8 | 3655 | |
c3d77172 URS |
3656 | area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN, |
3657 | node, page_order, nr_small_pages, area->pages); | |
5c1f4e69 | 3658 | |
97105f0a | 3659 | atomic_long_add(area->nr_pages, &nr_vmalloc_pages); |
4e5aa1f4 | 3660 | if (gfp_mask & __GFP_ACCOUNT) { |
3b8000ae | 3661 | int i; |
4e5aa1f4 | 3662 | |
3b8000ae NP |
3663 | for (i = 0; i < area->nr_pages; i++) |
3664 | mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1); | |
4e5aa1f4 | 3665 | } |
1da177e4 | 3666 | |
5c1f4e69 URS |
3667 | /* |
3668 | * If not enough pages were obtained to accomplish an | |
f41f036b | 3669 | * allocation request, free them via vfree() if any. |
5c1f4e69 URS |
3670 | */ |
3671 | if (area->nr_pages != nr_small_pages) { | |
95a301ee LS |
3672 | /* |
3673 | * vm_area_alloc_pages() can fail due to insufficient memory but | |
3674 | * also:- | |
3675 | * | |
3676 | * - a pending fatal signal | |
3677 | * - insufficient huge page-order pages | |
3678 | * | |
3679 | * Since we always retry allocations at order-0 in the huge page | |
3680 | * case a warning for either is spurious. | |
3681 | */ | |
3682 | if (!fatal_signal_pending(current) && page_order == 0) | |
f349b15e | 3683 | warn_alloc(gfp_mask, NULL, |
95a301ee LS |
3684 | "vmalloc error: size %lu, failed to allocate pages", |
3685 | area->nr_pages * PAGE_SIZE); | |
5c1f4e69 URS |
3686 | goto fail; |
3687 | } | |
3688 | ||
451769eb MH |
3689 | /* |
3690 | * page tables allocations ignore external gfp mask, enforce it | |
3691 | * by the scope API | |
3692 | */ | |
3693 | if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO) | |
3694 | flags = memalloc_nofs_save(); | |
3695 | else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0) | |
3696 | flags = memalloc_noio_save(); | |
3697 | ||
9376130c MH |
3698 | do { |
3699 | ret = vmap_pages_range(addr, addr + size, prot, area->pages, | |
451769eb | 3700 | page_shift); |
9376130c MH |
3701 | if (nofail && (ret < 0)) |
3702 | schedule_timeout_uninterruptible(1); | |
3703 | } while (nofail && (ret < 0)); | |
451769eb MH |
3704 | |
3705 | if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO) | |
3706 | memalloc_nofs_restore(flags); | |
3707 | else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0) | |
3708 | memalloc_noio_restore(flags); | |
3709 | ||
3710 | if (ret < 0) { | |
c3d77172 | 3711 | warn_alloc(gfp_mask, NULL, |
f4bdfeaf URS |
3712 | "vmalloc error: size %lu, failed to map pages", |
3713 | area->nr_pages * PAGE_SIZE); | |
1da177e4 | 3714 | goto fail; |
d70bec8c | 3715 | } |
ed1f324c | 3716 | |
1da177e4 LT |
3717 | return area->addr; |
3718 | ||
3719 | fail: | |
f41f036b | 3720 | vfree(area->addr); |
1da177e4 LT |
3721 | return NULL; |
3722 | } | |
3723 | ||
3724 | /** | |
92eac168 MR |
3725 | * __vmalloc_node_range - allocate virtually contiguous memory |
3726 | * @size: allocation size | |
3727 | * @align: desired alignment | |
3728 | * @start: vm area range start | |
3729 | * @end: vm area range end | |
3730 | * @gfp_mask: flags for the page level allocator | |
3731 | * @prot: protection mask for the allocated pages | |
3732 | * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD) | |
3733 | * @node: node to use for allocation or NUMA_NO_NODE | |
3734 | * @caller: caller's return address | |
3735 | * | |
3736 | * Allocate enough pages to cover @size from the page level | |
b7d90e7a | 3737 | * allocator with @gfp_mask flags. Please note that the full set of gfp |
30d3f011 MH |
3738 | * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all |
3739 | * supported. | |
3740 | * Zone modifiers are not supported. From the reclaim modifiers | |
3741 | * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported) | |
3742 | * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and | |
3743 | * __GFP_RETRY_MAYFAIL are not supported). | |
3744 | * | |
3745 | * __GFP_NOWARN can be used to suppress failures messages. | |
b7d90e7a MH |
3746 | * |
3747 | * Map them into contiguous kernel virtual space, using a pagetable | |
3748 | * protection of @prot. | |
a862f68a MR |
3749 | * |
3750 | * Return: the address of the area or %NULL on failure | |
1da177e4 | 3751 | */ |
88ae5fb7 | 3752 | void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align, |
d0a21265 | 3753 | unsigned long start, unsigned long end, gfp_t gfp_mask, |
cb9e3c29 AR |
3754 | pgprot_t prot, unsigned long vm_flags, int node, |
3755 | const void *caller) | |
1da177e4 LT |
3756 | { |
3757 | struct vm_struct *area; | |
19f1c3ac | 3758 | void *ret; |
f6e39794 | 3759 | kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE; |
89219d37 | 3760 | unsigned long real_size = size; |
121e6f32 NP |
3761 | unsigned long real_align = align; |
3762 | unsigned int shift = PAGE_SHIFT; | |
1da177e4 | 3763 | |
d70bec8c NP |
3764 | if (WARN_ON_ONCE(!size)) |
3765 | return NULL; | |
3766 | ||
3767 | if ((size >> PAGE_SHIFT) > totalram_pages()) { | |
3768 | warn_alloc(gfp_mask, NULL, | |
f4bdfeaf URS |
3769 | "vmalloc error: size %lu, exceeds total pages", |
3770 | real_size); | |
d70bec8c | 3771 | return NULL; |
121e6f32 NP |
3772 | } |
3773 | ||
559089e0 | 3774 | if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) { |
121e6f32 | 3775 | unsigned long size_per_node; |
1da177e4 | 3776 | |
121e6f32 NP |
3777 | /* |
3778 | * Try huge pages. Only try for PAGE_KERNEL allocations, | |
3779 | * others like modules don't yet expect huge pages in | |
3780 | * their allocations due to apply_to_page_range not | |
3781 | * supporting them. | |
3782 | */ | |
3783 | ||
3784 | size_per_node = size; | |
3785 | if (node == NUMA_NO_NODE) | |
3786 | size_per_node /= num_online_nodes(); | |
3382bbee | 3787 | if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE) |
121e6f32 | 3788 | shift = PMD_SHIFT; |
3382bbee CL |
3789 | else |
3790 | shift = arch_vmap_pte_supported_shift(size_per_node); | |
3791 | ||
3792 | align = max(real_align, 1UL << shift); | |
3793 | size = ALIGN(real_size, 1UL << shift); | |
121e6f32 NP |
3794 | } |
3795 | ||
3796 | again: | |
7ca3027b DA |
3797 | area = __get_vm_area_node(real_size, align, shift, VM_ALLOC | |
3798 | VM_UNINITIALIZED | vm_flags, start, end, node, | |
3799 | gfp_mask, caller); | |
d70bec8c | 3800 | if (!area) { |
9376130c | 3801 | bool nofail = gfp_mask & __GFP_NOFAIL; |
d70bec8c | 3802 | warn_alloc(gfp_mask, NULL, |
9376130c MH |
3803 | "vmalloc error: size %lu, vm_struct allocation failed%s", |
3804 | real_size, (nofail) ? ". Retrying." : ""); | |
3805 | if (nofail) { | |
3806 | schedule_timeout_uninterruptible(1); | |
3807 | goto again; | |
3808 | } | |
de7d2b56 | 3809 | goto fail; |
d70bec8c | 3810 | } |
1da177e4 | 3811 | |
f6e39794 AK |
3812 | /* |
3813 | * Prepare arguments for __vmalloc_area_node() and | |
3814 | * kasan_unpoison_vmalloc(). | |
3815 | */ | |
3816 | if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) { | |
3817 | if (kasan_hw_tags_enabled()) { | |
3818 | /* | |
3819 | * Modify protection bits to allow tagging. | |
3820 | * This must be done before mapping. | |
3821 | */ | |
3822 | prot = arch_vmap_pgprot_tagged(prot); | |
01d92c7f | 3823 | |
f6e39794 AK |
3824 | /* |
3825 | * Skip page_alloc poisoning and zeroing for physical | |
3826 | * pages backing VM_ALLOC mapping. Memory is instead | |
3827 | * poisoned and zeroed by kasan_unpoison_vmalloc(). | |
3828 | */ | |
0a54864f | 3829 | gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO; |
f6e39794 AK |
3830 | } |
3831 | ||
3832 | /* Take note that the mapping is PAGE_KERNEL. */ | |
3833 | kasan_flags |= KASAN_VMALLOC_PROT_NORMAL; | |
23689e91 AK |
3834 | } |
3835 | ||
01d92c7f | 3836 | /* Allocate physical pages and map them into vmalloc space. */ |
19f1c3ac AK |
3837 | ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node); |
3838 | if (!ret) | |
121e6f32 | 3839 | goto fail; |
89219d37 | 3840 | |
23689e91 AK |
3841 | /* |
3842 | * Mark the pages as accessible, now that they are mapped. | |
6c2f761d AK |
3843 | * The condition for setting KASAN_VMALLOC_INIT should complement the |
3844 | * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check | |
3845 | * to make sure that memory is initialized under the same conditions. | |
f6e39794 AK |
3846 | * Tag-based KASAN modes only assign tags to normal non-executable |
3847 | * allocations, see __kasan_unpoison_vmalloc(). | |
23689e91 | 3848 | */ |
f6e39794 | 3849 | kasan_flags |= KASAN_VMALLOC_VM_ALLOC; |
6c2f761d AK |
3850 | if (!want_init_on_free() && want_init_on_alloc(gfp_mask) && |
3851 | (gfp_mask & __GFP_SKIP_ZERO)) | |
23689e91 | 3852 | kasan_flags |= KASAN_VMALLOC_INIT; |
f6e39794 | 3853 | /* KASAN_VMALLOC_PROT_NORMAL already set if required. */ |
23689e91 | 3854 | area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags); |
19f1c3ac | 3855 | |
f5252e00 | 3856 | /* |
20fc02b4 ZY |
3857 | * In this function, newly allocated vm_struct has VM_UNINITIALIZED |
3858 | * flag. It means that vm_struct is not fully initialized. | |
4341fa45 | 3859 | * Now, it is fully initialized, so remove this flag here. |
f5252e00 | 3860 | */ |
20fc02b4 | 3861 | clear_vm_uninitialized_flag(area); |
f5252e00 | 3862 | |
7ca3027b | 3863 | size = PAGE_ALIGN(size); |
60115fa5 KW |
3864 | if (!(vm_flags & VM_DEFER_KMEMLEAK)) |
3865 | kmemleak_vmalloc(area, size, gfp_mask); | |
89219d37 | 3866 | |
19f1c3ac | 3867 | return area->addr; |
de7d2b56 JP |
3868 | |
3869 | fail: | |
121e6f32 NP |
3870 | if (shift > PAGE_SHIFT) { |
3871 | shift = PAGE_SHIFT; | |
3872 | align = real_align; | |
3873 | size = real_size; | |
3874 | goto again; | |
3875 | } | |
3876 | ||
de7d2b56 | 3877 | return NULL; |
1da177e4 LT |
3878 | } |
3879 | ||
d0a21265 | 3880 | /** |
92eac168 MR |
3881 | * __vmalloc_node - allocate virtually contiguous memory |
3882 | * @size: allocation size | |
3883 | * @align: desired alignment | |
3884 | * @gfp_mask: flags for the page level allocator | |
92eac168 MR |
3885 | * @node: node to use for allocation or NUMA_NO_NODE |
3886 | * @caller: caller's return address | |
a7c3e901 | 3887 | * |
f38fcb9c CH |
3888 | * Allocate enough pages to cover @size from the page level allocator with |
3889 | * @gfp_mask flags. Map them into contiguous kernel virtual space. | |
a7c3e901 | 3890 | * |
92eac168 MR |
3891 | * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL |
3892 | * and __GFP_NOFAIL are not supported | |
a7c3e901 | 3893 | * |
92eac168 MR |
3894 | * Any use of gfp flags outside of GFP_KERNEL should be consulted |
3895 | * with mm people. | |
a862f68a MR |
3896 | * |
3897 | * Return: pointer to the allocated memory or %NULL on error | |
d0a21265 | 3898 | */ |
88ae5fb7 | 3899 | void *__vmalloc_node_noprof(unsigned long size, unsigned long align, |
f38fcb9c | 3900 | gfp_t gfp_mask, int node, const void *caller) |
d0a21265 | 3901 | { |
88ae5fb7 | 3902 | return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END, |
f38fcb9c | 3903 | gfp_mask, PAGE_KERNEL, 0, node, caller); |
d0a21265 | 3904 | } |
c3f896dc CH |
3905 | /* |
3906 | * This is only for performance analysis of vmalloc and stress purpose. | |
3907 | * It is required by vmalloc test module, therefore do not use it other | |
3908 | * than that. | |
3909 | */ | |
3910 | #ifdef CONFIG_TEST_VMALLOC_MODULE | |
88ae5fb7 | 3911 | EXPORT_SYMBOL_GPL(__vmalloc_node_noprof); |
c3f896dc | 3912 | #endif |
d0a21265 | 3913 | |
88ae5fb7 | 3914 | void *__vmalloc_noprof(unsigned long size, gfp_t gfp_mask) |
930fc45a | 3915 | { |
88ae5fb7 | 3916 | return __vmalloc_node_noprof(size, 1, gfp_mask, NUMA_NO_NODE, |
23016969 | 3917 | __builtin_return_address(0)); |
930fc45a | 3918 | } |
88ae5fb7 | 3919 | EXPORT_SYMBOL(__vmalloc_noprof); |
1da177e4 LT |
3920 | |
3921 | /** | |
92eac168 MR |
3922 | * vmalloc - allocate virtually contiguous memory |
3923 | * @size: allocation size | |
3924 | * | |
3925 | * Allocate enough pages to cover @size from the page level | |
3926 | * allocator and map them into contiguous kernel virtual space. | |
1da177e4 | 3927 | * |
92eac168 MR |
3928 | * For tight control over page level allocator and protection flags |
3929 | * use __vmalloc() instead. | |
a862f68a MR |
3930 | * |
3931 | * Return: pointer to the allocated memory or %NULL on error | |
1da177e4 | 3932 | */ |
88ae5fb7 | 3933 | void *vmalloc_noprof(unsigned long size) |
1da177e4 | 3934 | { |
88ae5fb7 | 3935 | return __vmalloc_node_noprof(size, 1, GFP_KERNEL, NUMA_NO_NODE, |
4d39d728 | 3936 | __builtin_return_address(0)); |
1da177e4 | 3937 | } |
88ae5fb7 | 3938 | EXPORT_SYMBOL(vmalloc_noprof); |
1da177e4 | 3939 | |
15a64f5a | 3940 | /** |
559089e0 SL |
3941 | * vmalloc_huge - allocate virtually contiguous memory, allow huge pages |
3942 | * @size: allocation size | |
3943 | * @gfp_mask: flags for the page level allocator | |
15a64f5a | 3944 | * |
559089e0 | 3945 | * Allocate enough pages to cover @size from the page level |
15a64f5a | 3946 | * allocator and map them into contiguous kernel virtual space. |
559089e0 SL |
3947 | * If @size is greater than or equal to PMD_SIZE, allow using |
3948 | * huge pages for the memory | |
15a64f5a CI |
3949 | * |
3950 | * Return: pointer to the allocated memory or %NULL on error | |
3951 | */ | |
88ae5fb7 | 3952 | void *vmalloc_huge_noprof(unsigned long size, gfp_t gfp_mask) |
15a64f5a | 3953 | { |
88ae5fb7 | 3954 | return __vmalloc_node_range_noprof(size, 1, VMALLOC_START, VMALLOC_END, |
559089e0 | 3955 | gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP, |
15a64f5a CI |
3956 | NUMA_NO_NODE, __builtin_return_address(0)); |
3957 | } | |
88ae5fb7 | 3958 | EXPORT_SYMBOL_GPL(vmalloc_huge_noprof); |
15a64f5a | 3959 | |
e1ca7788 | 3960 | /** |
92eac168 MR |
3961 | * vzalloc - allocate virtually contiguous memory with zero fill |
3962 | * @size: allocation size | |
3963 | * | |
3964 | * Allocate enough pages to cover @size from the page level | |
3965 | * allocator and map them into contiguous kernel virtual space. | |
3966 | * The memory allocated is set to zero. | |
3967 | * | |
3968 | * For tight control over page level allocator and protection flags | |
3969 | * use __vmalloc() instead. | |
a862f68a MR |
3970 | * |
3971 | * Return: pointer to the allocated memory or %NULL on error | |
e1ca7788 | 3972 | */ |
88ae5fb7 | 3973 | void *vzalloc_noprof(unsigned long size) |
e1ca7788 | 3974 | { |
88ae5fb7 | 3975 | return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE, |
4d39d728 | 3976 | __builtin_return_address(0)); |
e1ca7788 | 3977 | } |
88ae5fb7 | 3978 | EXPORT_SYMBOL(vzalloc_noprof); |
e1ca7788 | 3979 | |
83342314 | 3980 | /** |
ead04089 REB |
3981 | * vmalloc_user - allocate zeroed virtually contiguous memory for userspace |
3982 | * @size: allocation size | |
83342314 | 3983 | * |
ead04089 REB |
3984 | * The resulting memory area is zeroed so it can be mapped to userspace |
3985 | * without leaking data. | |
a862f68a MR |
3986 | * |
3987 | * Return: pointer to the allocated memory or %NULL on error | |
83342314 | 3988 | */ |
88ae5fb7 | 3989 | void *vmalloc_user_noprof(unsigned long size) |
83342314 | 3990 | { |
88ae5fb7 | 3991 | return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END, |
bc84c535 RP |
3992 | GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL, |
3993 | VM_USERMAP, NUMA_NO_NODE, | |
3994 | __builtin_return_address(0)); | |
83342314 | 3995 | } |
88ae5fb7 | 3996 | EXPORT_SYMBOL(vmalloc_user_noprof); |
83342314 | 3997 | |
930fc45a | 3998 | /** |
92eac168 MR |
3999 | * vmalloc_node - allocate memory on a specific node |
4000 | * @size: allocation size | |
4001 | * @node: numa node | |
930fc45a | 4002 | * |
92eac168 MR |
4003 | * Allocate enough pages to cover @size from the page level |
4004 | * allocator and map them into contiguous kernel virtual space. | |
930fc45a | 4005 | * |
92eac168 MR |
4006 | * For tight control over page level allocator and protection flags |
4007 | * use __vmalloc() instead. | |
a862f68a MR |
4008 | * |
4009 | * Return: pointer to the allocated memory or %NULL on error | |
930fc45a | 4010 | */ |
88ae5fb7 | 4011 | void *vmalloc_node_noprof(unsigned long size, int node) |
930fc45a | 4012 | { |
88ae5fb7 | 4013 | return __vmalloc_node_noprof(size, 1, GFP_KERNEL, node, |
f38fcb9c | 4014 | __builtin_return_address(0)); |
930fc45a | 4015 | } |
88ae5fb7 | 4016 | EXPORT_SYMBOL(vmalloc_node_noprof); |
930fc45a | 4017 | |
e1ca7788 DY |
4018 | /** |
4019 | * vzalloc_node - allocate memory on a specific node with zero fill | |
4020 | * @size: allocation size | |
4021 | * @node: numa node | |
4022 | * | |
4023 | * Allocate enough pages to cover @size from the page level | |
4024 | * allocator and map them into contiguous kernel virtual space. | |
4025 | * The memory allocated is set to zero. | |
4026 | * | |
a862f68a | 4027 | * Return: pointer to the allocated memory or %NULL on error |
e1ca7788 | 4028 | */ |
88ae5fb7 | 4029 | void *vzalloc_node_noprof(unsigned long size, int node) |
e1ca7788 | 4030 | { |
88ae5fb7 | 4031 | return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, node, |
4d39d728 | 4032 | __builtin_return_address(0)); |
e1ca7788 | 4033 | } |
88ae5fb7 | 4034 | EXPORT_SYMBOL(vzalloc_node_noprof); |
e1ca7788 | 4035 | |
0d08e0d3 | 4036 | #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32) |
698d0831 | 4037 | #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL) |
0d08e0d3 | 4038 | #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA) |
698d0831 | 4039 | #define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL) |
0d08e0d3 | 4040 | #else |
698d0831 MH |
4041 | /* |
4042 | * 64b systems should always have either DMA or DMA32 zones. For others | |
4043 | * GFP_DMA32 should do the right thing and use the normal zone. | |
4044 | */ | |
68d68ff6 | 4045 | #define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL) |
0d08e0d3 AK |
4046 | #endif |
4047 | ||
1da177e4 | 4048 | /** |
92eac168 MR |
4049 | * vmalloc_32 - allocate virtually contiguous memory (32bit addressable) |
4050 | * @size: allocation size | |
1da177e4 | 4051 | * |
92eac168 MR |
4052 | * Allocate enough 32bit PA addressable pages to cover @size from the |
4053 | * page level allocator and map them into contiguous kernel virtual space. | |
a862f68a MR |
4054 | * |
4055 | * Return: pointer to the allocated memory or %NULL on error | |
1da177e4 | 4056 | */ |
88ae5fb7 | 4057 | void *vmalloc_32_noprof(unsigned long size) |
1da177e4 | 4058 | { |
88ae5fb7 | 4059 | return __vmalloc_node_noprof(size, 1, GFP_VMALLOC32, NUMA_NO_NODE, |
f38fcb9c | 4060 | __builtin_return_address(0)); |
1da177e4 | 4061 | } |
88ae5fb7 | 4062 | EXPORT_SYMBOL(vmalloc_32_noprof); |
1da177e4 | 4063 | |
83342314 | 4064 | /** |
ead04089 | 4065 | * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory |
92eac168 | 4066 | * @size: allocation size |
ead04089 REB |
4067 | * |
4068 | * The resulting memory area is 32bit addressable and zeroed so it can be | |
4069 | * mapped to userspace without leaking data. | |
a862f68a MR |
4070 | * |
4071 | * Return: pointer to the allocated memory or %NULL on error | |
83342314 | 4072 | */ |
88ae5fb7 | 4073 | void *vmalloc_32_user_noprof(unsigned long size) |
83342314 | 4074 | { |
88ae5fb7 | 4075 | return __vmalloc_node_range_noprof(size, SHMLBA, VMALLOC_START, VMALLOC_END, |
bc84c535 RP |
4076 | GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL, |
4077 | VM_USERMAP, NUMA_NO_NODE, | |
4078 | __builtin_return_address(0)); | |
83342314 | 4079 | } |
88ae5fb7 | 4080 | EXPORT_SYMBOL(vmalloc_32_user_noprof); |
83342314 | 4081 | |
d0107eb0 | 4082 | /* |
4c91c07c LS |
4083 | * Atomically zero bytes in the iterator. |
4084 | * | |
4085 | * Returns the number of zeroed bytes. | |
d0107eb0 | 4086 | */ |
4c91c07c LS |
4087 | static size_t zero_iter(struct iov_iter *iter, size_t count) |
4088 | { | |
4089 | size_t remains = count; | |
4090 | ||
4091 | while (remains > 0) { | |
4092 | size_t num, copied; | |
4093 | ||
0e4bc271 | 4094 | num = min_t(size_t, remains, PAGE_SIZE); |
4c91c07c LS |
4095 | copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter); |
4096 | remains -= copied; | |
4097 | ||
4098 | if (copied < num) | |
4099 | break; | |
4100 | } | |
d0107eb0 | 4101 | |
4c91c07c LS |
4102 | return count - remains; |
4103 | } | |
4104 | ||
4105 | /* | |
4106 | * small helper routine, copy contents to iter from addr. | |
4107 | * If the page is not present, fill zero. | |
4108 | * | |
4109 | * Returns the number of copied bytes. | |
4110 | */ | |
4111 | static size_t aligned_vread_iter(struct iov_iter *iter, | |
4112 | const char *addr, size_t count) | |
d0107eb0 | 4113 | { |
4c91c07c LS |
4114 | size_t remains = count; |
4115 | struct page *page; | |
d0107eb0 | 4116 | |
4c91c07c | 4117 | while (remains > 0) { |
d0107eb0 | 4118 | unsigned long offset, length; |
4c91c07c | 4119 | size_t copied = 0; |
d0107eb0 | 4120 | |
891c49ab | 4121 | offset = offset_in_page(addr); |
d0107eb0 | 4122 | length = PAGE_SIZE - offset; |
4c91c07c LS |
4123 | if (length > remains) |
4124 | length = remains; | |
4125 | page = vmalloc_to_page(addr); | |
d0107eb0 | 4126 | /* |
4c91c07c LS |
4127 | * To do safe access to this _mapped_ area, we need lock. But |
4128 | * adding lock here means that we need to add overhead of | |
4129 | * vmalloc()/vfree() calls for this _debug_ interface, rarely | |
4130 | * used. Instead of that, we'll use an local mapping via | |
4131 | * copy_page_to_iter_nofault() and accept a small overhead in | |
4132 | * this access function. | |
d0107eb0 | 4133 | */ |
4c91c07c LS |
4134 | if (page) |
4135 | copied = copy_page_to_iter_nofault(page, offset, | |
4136 | length, iter); | |
4137 | else | |
4138 | copied = zero_iter(iter, length); | |
d0107eb0 | 4139 | |
4c91c07c LS |
4140 | addr += copied; |
4141 | remains -= copied; | |
4142 | ||
4143 | if (copied != length) | |
4144 | break; | |
d0107eb0 | 4145 | } |
4c91c07c LS |
4146 | |
4147 | return count - remains; | |
d0107eb0 KH |
4148 | } |
4149 | ||
4c91c07c LS |
4150 | /* |
4151 | * Read from a vm_map_ram region of memory. | |
4152 | * | |
4153 | * Returns the number of copied bytes. | |
4154 | */ | |
4155 | static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr, | |
4156 | size_t count, unsigned long flags) | |
06c89946 BH |
4157 | { |
4158 | char *start; | |
4159 | struct vmap_block *vb; | |
062eacf5 | 4160 | struct xarray *xa; |
06c89946 | 4161 | unsigned long offset; |
4c91c07c LS |
4162 | unsigned int rs, re; |
4163 | size_t remains, n; | |
06c89946 BH |
4164 | |
4165 | /* | |
4166 | * If it's area created by vm_map_ram() interface directly, but | |
4167 | * not further subdividing and delegating management to vmap_block, | |
4168 | * handle it here. | |
4169 | */ | |
4c91c07c LS |
4170 | if (!(flags & VMAP_BLOCK)) |
4171 | return aligned_vread_iter(iter, addr, count); | |
4172 | ||
4173 | remains = count; | |
06c89946 BH |
4174 | |
4175 | /* | |
4176 | * Area is split into regions and tracked with vmap_block, read out | |
4177 | * each region and zero fill the hole between regions. | |
4178 | */ | |
fa1c77c1 | 4179 | xa = addr_to_vb_xa((unsigned long) addr); |
062eacf5 | 4180 | vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr)); |
06c89946 | 4181 | if (!vb) |
4c91c07c | 4182 | goto finished_zero; |
06c89946 BH |
4183 | |
4184 | spin_lock(&vb->lock); | |
4185 | if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) { | |
4186 | spin_unlock(&vb->lock); | |
4c91c07c | 4187 | goto finished_zero; |
06c89946 | 4188 | } |
4c91c07c | 4189 | |
06c89946 | 4190 | for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) { |
4c91c07c LS |
4191 | size_t copied; |
4192 | ||
4193 | if (remains == 0) | |
4194 | goto finished; | |
4195 | ||
06c89946 | 4196 | start = vmap_block_vaddr(vb->va->va_start, rs); |
4c91c07c LS |
4197 | |
4198 | if (addr < start) { | |
4199 | size_t to_zero = min_t(size_t, start - addr, remains); | |
4200 | size_t zeroed = zero_iter(iter, to_zero); | |
4201 | ||
4202 | addr += zeroed; | |
4203 | remains -= zeroed; | |
4204 | ||
4205 | if (remains == 0 || zeroed != to_zero) | |
4206 | goto finished; | |
06c89946 | 4207 | } |
4c91c07c | 4208 | |
06c89946 BH |
4209 | /*it could start reading from the middle of used region*/ |
4210 | offset = offset_in_page(addr); | |
4211 | n = ((re - rs + 1) << PAGE_SHIFT) - offset; | |
4c91c07c LS |
4212 | if (n > remains) |
4213 | n = remains; | |
4214 | ||
4215 | copied = aligned_vread_iter(iter, start + offset, n); | |
06c89946 | 4216 | |
4c91c07c LS |
4217 | addr += copied; |
4218 | remains -= copied; | |
4219 | ||
4220 | if (copied != n) | |
4221 | goto finished; | |
06c89946 | 4222 | } |
4c91c07c | 4223 | |
06c89946 BH |
4224 | spin_unlock(&vb->lock); |
4225 | ||
4c91c07c | 4226 | finished_zero: |
06c89946 | 4227 | /* zero-fill the left dirty or free regions */ |
4c91c07c LS |
4228 | return count - remains + zero_iter(iter, remains); |
4229 | finished: | |
4230 | /* We couldn't copy/zero everything */ | |
4231 | spin_unlock(&vb->lock); | |
4232 | return count - remains; | |
06c89946 BH |
4233 | } |
4234 | ||
d0107eb0 | 4235 | /** |
4c91c07c LS |
4236 | * vread_iter() - read vmalloc area in a safe way to an iterator. |
4237 | * @iter: the iterator to which data should be written. | |
4238 | * @addr: vm address. | |
4239 | * @count: number of bytes to be read. | |
92eac168 | 4240 | * |
92eac168 MR |
4241 | * This function checks that addr is a valid vmalloc'ed area, and |
4242 | * copy data from that area to a given buffer. If the given memory range | |
4243 | * of [addr...addr+count) includes some valid address, data is copied to | |
4244 | * proper area of @buf. If there are memory holes, they'll be zero-filled. | |
4245 | * IOREMAP area is treated as memory hole and no copy is done. | |
4246 | * | |
4247 | * If [addr...addr+count) doesn't includes any intersects with alive | |
4248 | * vm_struct area, returns 0. @buf should be kernel's buffer. | |
4249 | * | |
4250 | * Note: In usual ops, vread() is never necessary because the caller | |
4251 | * should know vmalloc() area is valid and can use memcpy(). | |
4252 | * This is for routines which have to access vmalloc area without | |
bbcd53c9 | 4253 | * any information, as /proc/kcore. |
a862f68a MR |
4254 | * |
4255 | * Return: number of bytes for which addr and buf should be increased | |
4256 | * (same number as @count) or %0 if [addr...addr+count) doesn't | |
4257 | * include any intersection with valid vmalloc area | |
d0107eb0 | 4258 | */ |
4c91c07c | 4259 | long vread_iter(struct iov_iter *iter, const char *addr, size_t count) |
1da177e4 | 4260 | { |
d0936029 | 4261 | struct vmap_node *vn; |
e81ce85f JK |
4262 | struct vmap_area *va; |
4263 | struct vm_struct *vm; | |
4c91c07c LS |
4264 | char *vaddr; |
4265 | size_t n, size, flags, remains; | |
53becf32 | 4266 | unsigned long next; |
1da177e4 | 4267 | |
4aff1dc4 AK |
4268 | addr = kasan_reset_tag(addr); |
4269 | ||
1da177e4 LT |
4270 | /* Don't allow overflow */ |
4271 | if ((unsigned long) addr + count < count) | |
4272 | count = -(unsigned long) addr; | |
4273 | ||
4c91c07c LS |
4274 | remains = count; |
4275 | ||
53becf32 URS |
4276 | vn = find_vmap_area_exceed_addr_lock((unsigned long) addr, &va); |
4277 | if (!vn) | |
4c91c07c | 4278 | goto finished_zero; |
f181234a CW |
4279 | |
4280 | /* no intersects with alive vmap_area */ | |
4c91c07c LS |
4281 | if ((unsigned long)addr + remains <= va->va_start) |
4282 | goto finished_zero; | |
f181234a | 4283 | |
53becf32 | 4284 | do { |
4c91c07c LS |
4285 | size_t copied; |
4286 | ||
4287 | if (remains == 0) | |
4288 | goto finished; | |
e81ce85f | 4289 | |
06c89946 BH |
4290 | vm = va->vm; |
4291 | flags = va->flags & VMAP_FLAGS_MASK; | |
4292 | /* | |
4293 | * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need | |
4294 | * be set together with VMAP_RAM. | |
4295 | */ | |
4296 | WARN_ON(flags == VMAP_BLOCK); | |
4297 | ||
4298 | if (!vm && !flags) | |
53becf32 | 4299 | goto next_va; |
e81ce85f | 4300 | |
30a7a9b1 | 4301 | if (vm && (vm->flags & VM_UNINITIALIZED)) |
53becf32 | 4302 | goto next_va; |
4c91c07c | 4303 | |
30a7a9b1 BH |
4304 | /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */ |
4305 | smp_rmb(); | |
4306 | ||
06c89946 BH |
4307 | vaddr = (char *) va->va_start; |
4308 | size = vm ? get_vm_area_size(vm) : va_size(va); | |
4309 | ||
4310 | if (addr >= vaddr + size) | |
53becf32 | 4311 | goto next_va; |
4c91c07c LS |
4312 | |
4313 | if (addr < vaddr) { | |
4314 | size_t to_zero = min_t(size_t, vaddr - addr, remains); | |
4315 | size_t zeroed = zero_iter(iter, to_zero); | |
4316 | ||
4317 | addr += zeroed; | |
4318 | remains -= zeroed; | |
4319 | ||
4320 | if (remains == 0 || zeroed != to_zero) | |
1da177e4 | 4321 | goto finished; |
1da177e4 | 4322 | } |
4c91c07c | 4323 | |
06c89946 | 4324 | n = vaddr + size - addr; |
4c91c07c LS |
4325 | if (n > remains) |
4326 | n = remains; | |
06c89946 BH |
4327 | |
4328 | if (flags & VMAP_RAM) | |
4c91c07c | 4329 | copied = vmap_ram_vread_iter(iter, addr, n, flags); |
e6f79822 | 4330 | else if (!(vm && (vm->flags & (VM_IOREMAP | VM_SPARSE)))) |
4c91c07c | 4331 | copied = aligned_vread_iter(iter, addr, n); |
e6f79822 | 4332 | else /* IOREMAP | SPARSE area is treated as memory hole */ |
4c91c07c LS |
4333 | copied = zero_iter(iter, n); |
4334 | ||
4335 | addr += copied; | |
4336 | remains -= copied; | |
4337 | ||
4338 | if (copied != n) | |
4339 | goto finished; | |
53becf32 URS |
4340 | |
4341 | next_va: | |
4342 | next = va->va_end; | |
4343 | spin_unlock(&vn->busy.lock); | |
4344 | } while ((vn = find_vmap_area_exceed_addr_lock(next, &va))); | |
d0107eb0 | 4345 | |
4c91c07c | 4346 | finished_zero: |
53becf32 URS |
4347 | if (vn) |
4348 | spin_unlock(&vn->busy.lock); | |
4349 | ||
d0107eb0 | 4350 | /* zero-fill memory holes */ |
4c91c07c LS |
4351 | return count - remains + zero_iter(iter, remains); |
4352 | finished: | |
4353 | /* Nothing remains, or We couldn't copy/zero everything. */ | |
53becf32 URS |
4354 | if (vn) |
4355 | spin_unlock(&vn->busy.lock); | |
d0107eb0 | 4356 | |
4c91c07c | 4357 | return count - remains; |
1da177e4 LT |
4358 | } |
4359 | ||
83342314 | 4360 | /** |
92eac168 MR |
4361 | * remap_vmalloc_range_partial - map vmalloc pages to userspace |
4362 | * @vma: vma to cover | |
4363 | * @uaddr: target user address to start at | |
4364 | * @kaddr: virtual address of vmalloc kernel memory | |
bdebd6a2 | 4365 | * @pgoff: offset from @kaddr to start at |
92eac168 | 4366 | * @size: size of map area |
7682486b | 4367 | * |
92eac168 | 4368 | * Returns: 0 for success, -Exxx on failure |
83342314 | 4369 | * |
92eac168 MR |
4370 | * This function checks that @kaddr is a valid vmalloc'ed area, |
4371 | * and that it is big enough to cover the range starting at | |
4372 | * @uaddr in @vma. Will return failure if that criteria isn't | |
4373 | * met. | |
83342314 | 4374 | * |
92eac168 | 4375 | * Similar to remap_pfn_range() (see mm/memory.c) |
83342314 | 4376 | */ |
e69e9d4a | 4377 | int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr, |
bdebd6a2 JH |
4378 | void *kaddr, unsigned long pgoff, |
4379 | unsigned long size) | |
83342314 NP |
4380 | { |
4381 | struct vm_struct *area; | |
bdebd6a2 JH |
4382 | unsigned long off; |
4383 | unsigned long end_index; | |
4384 | ||
4385 | if (check_shl_overflow(pgoff, PAGE_SHIFT, &off)) | |
4386 | return -EINVAL; | |
83342314 | 4387 | |
e69e9d4a HD |
4388 | size = PAGE_ALIGN(size); |
4389 | ||
4390 | if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr)) | |
83342314 NP |
4391 | return -EINVAL; |
4392 | ||
e69e9d4a | 4393 | area = find_vm_area(kaddr); |
83342314 | 4394 | if (!area) |
db64fe02 | 4395 | return -EINVAL; |
83342314 | 4396 | |
fe9041c2 | 4397 | if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT))) |
db64fe02 | 4398 | return -EINVAL; |
83342314 | 4399 | |
bdebd6a2 JH |
4400 | if (check_add_overflow(size, off, &end_index) || |
4401 | end_index > get_vm_area_size(area)) | |
db64fe02 | 4402 | return -EINVAL; |
bdebd6a2 | 4403 | kaddr += off; |
83342314 | 4404 | |
83342314 | 4405 | do { |
e69e9d4a | 4406 | struct page *page = vmalloc_to_page(kaddr); |
db64fe02 NP |
4407 | int ret; |
4408 | ||
83342314 NP |
4409 | ret = vm_insert_page(vma, uaddr, page); |
4410 | if (ret) | |
4411 | return ret; | |
4412 | ||
4413 | uaddr += PAGE_SIZE; | |
e69e9d4a HD |
4414 | kaddr += PAGE_SIZE; |
4415 | size -= PAGE_SIZE; | |
4416 | } while (size > 0); | |
83342314 | 4417 | |
1c71222e | 4418 | vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP); |
83342314 | 4419 | |
db64fe02 | 4420 | return 0; |
83342314 | 4421 | } |
e69e9d4a HD |
4422 | |
4423 | /** | |
92eac168 MR |
4424 | * remap_vmalloc_range - map vmalloc pages to userspace |
4425 | * @vma: vma to cover (map full range of vma) | |
4426 | * @addr: vmalloc memory | |
4427 | * @pgoff: number of pages into addr before first page to map | |
e69e9d4a | 4428 | * |
92eac168 | 4429 | * Returns: 0 for success, -Exxx on failure |
e69e9d4a | 4430 | * |
92eac168 MR |
4431 | * This function checks that addr is a valid vmalloc'ed area, and |
4432 | * that it is big enough to cover the vma. Will return failure if | |
4433 | * that criteria isn't met. | |
e69e9d4a | 4434 | * |
92eac168 | 4435 | * Similar to remap_pfn_range() (see mm/memory.c) |
e69e9d4a HD |
4436 | */ |
4437 | int remap_vmalloc_range(struct vm_area_struct *vma, void *addr, | |
4438 | unsigned long pgoff) | |
4439 | { | |
4440 | return remap_vmalloc_range_partial(vma, vma->vm_start, | |
bdebd6a2 | 4441 | addr, pgoff, |
e69e9d4a HD |
4442 | vma->vm_end - vma->vm_start); |
4443 | } | |
83342314 NP |
4444 | EXPORT_SYMBOL(remap_vmalloc_range); |
4445 | ||
5f4352fb JF |
4446 | void free_vm_area(struct vm_struct *area) |
4447 | { | |
4448 | struct vm_struct *ret; | |
4449 | ret = remove_vm_area(area->addr); | |
4450 | BUG_ON(ret != area); | |
4451 | kfree(area); | |
4452 | } | |
4453 | EXPORT_SYMBOL_GPL(free_vm_area); | |
a10aa579 | 4454 | |
4f8b02b4 | 4455 | #ifdef CONFIG_SMP |
ca23e405 TH |
4456 | static struct vmap_area *node_to_va(struct rb_node *n) |
4457 | { | |
4583e773 | 4458 | return rb_entry_safe(n, struct vmap_area, rb_node); |
ca23e405 TH |
4459 | } |
4460 | ||
4461 | /** | |
68ad4a33 URS |
4462 | * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to |
4463 | * @addr: target address | |
ca23e405 | 4464 | * |
68ad4a33 URS |
4465 | * Returns: vmap_area if it is found. If there is no such area |
4466 | * the first highest(reverse order) vmap_area is returned | |
4467 | * i.e. va->va_start < addr && va->va_end < addr or NULL | |
4468 | * if there are no any areas before @addr. | |
ca23e405 | 4469 | */ |
68ad4a33 URS |
4470 | static struct vmap_area * |
4471 | pvm_find_va_enclose_addr(unsigned long addr) | |
ca23e405 | 4472 | { |
68ad4a33 URS |
4473 | struct vmap_area *va, *tmp; |
4474 | struct rb_node *n; | |
4475 | ||
4476 | n = free_vmap_area_root.rb_node; | |
4477 | va = NULL; | |
ca23e405 TH |
4478 | |
4479 | while (n) { | |
68ad4a33 URS |
4480 | tmp = rb_entry(n, struct vmap_area, rb_node); |
4481 | if (tmp->va_start <= addr) { | |
4482 | va = tmp; | |
4483 | if (tmp->va_end >= addr) | |
4484 | break; | |
4485 | ||
ca23e405 | 4486 | n = n->rb_right; |
68ad4a33 URS |
4487 | } else { |
4488 | n = n->rb_left; | |
4489 | } | |
ca23e405 TH |
4490 | } |
4491 | ||
68ad4a33 | 4492 | return va; |
ca23e405 TH |
4493 | } |
4494 | ||
4495 | /** | |
68ad4a33 URS |
4496 | * pvm_determine_end_from_reverse - find the highest aligned address |
4497 | * of free block below VMALLOC_END | |
4498 | * @va: | |
4499 | * in - the VA we start the search(reverse order); | |
4500 | * out - the VA with the highest aligned end address. | |
799fa85d | 4501 | * @align: alignment for required highest address |
ca23e405 | 4502 | * |
68ad4a33 | 4503 | * Returns: determined end address within vmap_area |
ca23e405 | 4504 | */ |
68ad4a33 URS |
4505 | static unsigned long |
4506 | pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align) | |
ca23e405 | 4507 | { |
68ad4a33 | 4508 | unsigned long vmalloc_end = VMALLOC_END & ~(align - 1); |
ca23e405 TH |
4509 | unsigned long addr; |
4510 | ||
68ad4a33 URS |
4511 | if (likely(*va)) { |
4512 | list_for_each_entry_from_reverse((*va), | |
4513 | &free_vmap_area_list, list) { | |
4514 | addr = min((*va)->va_end & ~(align - 1), vmalloc_end); | |
4515 | if ((*va)->va_start < addr) | |
4516 | return addr; | |
4517 | } | |
ca23e405 TH |
4518 | } |
4519 | ||
68ad4a33 | 4520 | return 0; |
ca23e405 TH |
4521 | } |
4522 | ||
4523 | /** | |
4524 | * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator | |
4525 | * @offsets: array containing offset of each area | |
4526 | * @sizes: array containing size of each area | |
4527 | * @nr_vms: the number of areas to allocate | |
4528 | * @align: alignment, all entries in @offsets and @sizes must be aligned to this | |
ca23e405 TH |
4529 | * |
4530 | * Returns: kmalloc'd vm_struct pointer array pointing to allocated | |
4531 | * vm_structs on success, %NULL on failure | |
4532 | * | |
4533 | * Percpu allocator wants to use congruent vm areas so that it can | |
4534 | * maintain the offsets among percpu areas. This function allocates | |
ec3f64fc DR |
4535 | * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to |
4536 | * be scattered pretty far, distance between two areas easily going up | |
4537 | * to gigabytes. To avoid interacting with regular vmallocs, these | |
4538 | * areas are allocated from top. | |
ca23e405 | 4539 | * |
68ad4a33 URS |
4540 | * Despite its complicated look, this allocator is rather simple. It |
4541 | * does everything top-down and scans free blocks from the end looking | |
4542 | * for matching base. While scanning, if any of the areas do not fit the | |
4543 | * base address is pulled down to fit the area. Scanning is repeated till | |
4544 | * all the areas fit and then all necessary data structures are inserted | |
4545 | * and the result is returned. | |
ca23e405 TH |
4546 | */ |
4547 | struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets, | |
4548 | const size_t *sizes, int nr_vms, | |
ec3f64fc | 4549 | size_t align) |
ca23e405 TH |
4550 | { |
4551 | const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align); | |
4552 | const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1); | |
68ad4a33 | 4553 | struct vmap_area **vas, *va; |
ca23e405 TH |
4554 | struct vm_struct **vms; |
4555 | int area, area2, last_area, term_area; | |
253a496d | 4556 | unsigned long base, start, size, end, last_end, orig_start, orig_end; |
ca23e405 TH |
4557 | bool purged = false; |
4558 | ||
ca23e405 | 4559 | /* verify parameters and allocate data structures */ |
891c49ab | 4560 | BUG_ON(offset_in_page(align) || !is_power_of_2(align)); |
ca23e405 TH |
4561 | for (last_area = 0, area = 0; area < nr_vms; area++) { |
4562 | start = offsets[area]; | |
4563 | end = start + sizes[area]; | |
4564 | ||
4565 | /* is everything aligned properly? */ | |
4566 | BUG_ON(!IS_ALIGNED(offsets[area], align)); | |
4567 | BUG_ON(!IS_ALIGNED(sizes[area], align)); | |
4568 | ||
4569 | /* detect the area with the highest address */ | |
4570 | if (start > offsets[last_area]) | |
4571 | last_area = area; | |
4572 | ||
c568da28 | 4573 | for (area2 = area + 1; area2 < nr_vms; area2++) { |
ca23e405 TH |
4574 | unsigned long start2 = offsets[area2]; |
4575 | unsigned long end2 = start2 + sizes[area2]; | |
4576 | ||
c568da28 | 4577 | BUG_ON(start2 < end && start < end2); |
ca23e405 TH |
4578 | } |
4579 | } | |
4580 | last_end = offsets[last_area] + sizes[last_area]; | |
4581 | ||
4582 | if (vmalloc_end - vmalloc_start < last_end) { | |
4583 | WARN_ON(true); | |
4584 | return NULL; | |
4585 | } | |
4586 | ||
4d67d860 TM |
4587 | vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL); |
4588 | vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL); | |
ca23e405 | 4589 | if (!vas || !vms) |
f1db7afd | 4590 | goto err_free2; |
ca23e405 TH |
4591 | |
4592 | for (area = 0; area < nr_vms; area++) { | |
68ad4a33 | 4593 | vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL); |
ec3f64fc | 4594 | vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL); |
ca23e405 TH |
4595 | if (!vas[area] || !vms[area]) |
4596 | goto err_free; | |
4597 | } | |
4598 | retry: | |
e36176be | 4599 | spin_lock(&free_vmap_area_lock); |
ca23e405 TH |
4600 | |
4601 | /* start scanning - we scan from the top, begin with the last area */ | |
4602 | area = term_area = last_area; | |
4603 | start = offsets[area]; | |
4604 | end = start + sizes[area]; | |
4605 | ||
68ad4a33 URS |
4606 | va = pvm_find_va_enclose_addr(vmalloc_end); |
4607 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
ca23e405 TH |
4608 | |
4609 | while (true) { | |
ca23e405 TH |
4610 | /* |
4611 | * base might have underflowed, add last_end before | |
4612 | * comparing. | |
4613 | */ | |
68ad4a33 URS |
4614 | if (base + last_end < vmalloc_start + last_end) |
4615 | goto overflow; | |
ca23e405 TH |
4616 | |
4617 | /* | |
68ad4a33 | 4618 | * Fitting base has not been found. |
ca23e405 | 4619 | */ |
68ad4a33 URS |
4620 | if (va == NULL) |
4621 | goto overflow; | |
ca23e405 | 4622 | |
5336e52c | 4623 | /* |
d8cc323d | 4624 | * If required width exceeds current VA block, move |
5336e52c KS |
4625 | * base downwards and then recheck. |
4626 | */ | |
4627 | if (base + end > va->va_end) { | |
4628 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
4629 | term_area = area; | |
4630 | continue; | |
4631 | } | |
4632 | ||
ca23e405 | 4633 | /* |
68ad4a33 | 4634 | * If this VA does not fit, move base downwards and recheck. |
ca23e405 | 4635 | */ |
5336e52c | 4636 | if (base + start < va->va_start) { |
68ad4a33 URS |
4637 | va = node_to_va(rb_prev(&va->rb_node)); |
4638 | base = pvm_determine_end_from_reverse(&va, align) - end; | |
ca23e405 TH |
4639 | term_area = area; |
4640 | continue; | |
4641 | } | |
4642 | ||
4643 | /* | |
4644 | * This area fits, move on to the previous one. If | |
4645 | * the previous one is the terminal one, we're done. | |
4646 | */ | |
4647 | area = (area + nr_vms - 1) % nr_vms; | |
4648 | if (area == term_area) | |
4649 | break; | |
68ad4a33 | 4650 | |
ca23e405 TH |
4651 | start = offsets[area]; |
4652 | end = start + sizes[area]; | |
68ad4a33 | 4653 | va = pvm_find_va_enclose_addr(base + end); |
ca23e405 | 4654 | } |
68ad4a33 | 4655 | |
ca23e405 TH |
4656 | /* we've found a fitting base, insert all va's */ |
4657 | for (area = 0; area < nr_vms; area++) { | |
68ad4a33 | 4658 | int ret; |
ca23e405 | 4659 | |
68ad4a33 URS |
4660 | start = base + offsets[area]; |
4661 | size = sizes[area]; | |
ca23e405 | 4662 | |
68ad4a33 URS |
4663 | va = pvm_find_va_enclose_addr(start); |
4664 | if (WARN_ON_ONCE(va == NULL)) | |
4665 | /* It is a BUG(), but trigger recovery instead. */ | |
4666 | goto recovery; | |
4667 | ||
5b75b8e1 URS |
4668 | ret = va_clip(&free_vmap_area_root, |
4669 | &free_vmap_area_list, va, start, size); | |
1b23ff80 | 4670 | if (WARN_ON_ONCE(unlikely(ret))) |
68ad4a33 URS |
4671 | /* It is a BUG(), but trigger recovery instead. */ |
4672 | goto recovery; | |
4673 | ||
68ad4a33 URS |
4674 | /* Allocated area. */ |
4675 | va = vas[area]; | |
4676 | va->va_start = start; | |
4677 | va->va_end = start + size; | |
68ad4a33 | 4678 | } |
ca23e405 | 4679 | |
e36176be | 4680 | spin_unlock(&free_vmap_area_lock); |
ca23e405 | 4681 | |
253a496d DA |
4682 | /* populate the kasan shadow space */ |
4683 | for (area = 0; area < nr_vms; area++) { | |
4684 | if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area])) | |
4685 | goto err_free_shadow; | |
253a496d DA |
4686 | } |
4687 | ||
ca23e405 | 4688 | /* insert all vm's */ |
e36176be | 4689 | for (area = 0; area < nr_vms; area++) { |
d0936029 | 4690 | struct vmap_node *vn = addr_to_node(vas[area]->va_start); |
e36176be | 4691 | |
d0936029 URS |
4692 | spin_lock(&vn->busy.lock); |
4693 | insert_vmap_area(vas[area], &vn->busy.root, &vn->busy.head); | |
aaab830a | 4694 | setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC, |
3645cb4a | 4695 | pcpu_get_vm_areas); |
d0936029 | 4696 | spin_unlock(&vn->busy.lock); |
e36176be | 4697 | } |
ca23e405 | 4698 | |
19f1c3ac AK |
4699 | /* |
4700 | * Mark allocated areas as accessible. Do it now as a best-effort | |
4701 | * approach, as they can be mapped outside of vmalloc code. | |
23689e91 AK |
4702 | * With hardware tag-based KASAN, marking is skipped for |
4703 | * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc(). | |
19f1c3ac | 4704 | */ |
1d96320f AK |
4705 | for (area = 0; area < nr_vms; area++) |
4706 | vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr, | |
f6e39794 | 4707 | vms[area]->size, KASAN_VMALLOC_PROT_NORMAL); |
1d96320f | 4708 | |
ca23e405 TH |
4709 | kfree(vas); |
4710 | return vms; | |
4711 | ||
68ad4a33 | 4712 | recovery: |
e36176be URS |
4713 | /* |
4714 | * Remove previously allocated areas. There is no | |
4715 | * need in removing these areas from the busy tree, | |
4716 | * because they are inserted only on the final step | |
4717 | * and when pcpu_get_vm_areas() is success. | |
4718 | */ | |
68ad4a33 | 4719 | while (area--) { |
253a496d DA |
4720 | orig_start = vas[area]->va_start; |
4721 | orig_end = vas[area]->va_end; | |
96e2db45 URS |
4722 | va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root, |
4723 | &free_vmap_area_list); | |
9c801f61 URS |
4724 | if (va) |
4725 | kasan_release_vmalloc(orig_start, orig_end, | |
4726 | va->va_start, va->va_end); | |
68ad4a33 URS |
4727 | vas[area] = NULL; |
4728 | } | |
4729 | ||
4730 | overflow: | |
e36176be | 4731 | spin_unlock(&free_vmap_area_lock); |
68ad4a33 | 4732 | if (!purged) { |
77e50af0 | 4733 | reclaim_and_purge_vmap_areas(); |
68ad4a33 URS |
4734 | purged = true; |
4735 | ||
4736 | /* Before "retry", check if we recover. */ | |
4737 | for (area = 0; area < nr_vms; area++) { | |
4738 | if (vas[area]) | |
4739 | continue; | |
4740 | ||
4741 | vas[area] = kmem_cache_zalloc( | |
4742 | vmap_area_cachep, GFP_KERNEL); | |
4743 | if (!vas[area]) | |
4744 | goto err_free; | |
4745 | } | |
4746 | ||
4747 | goto retry; | |
4748 | } | |
4749 | ||
ca23e405 TH |
4750 | err_free: |
4751 | for (area = 0; area < nr_vms; area++) { | |
68ad4a33 URS |
4752 | if (vas[area]) |
4753 | kmem_cache_free(vmap_area_cachep, vas[area]); | |
4754 | ||
f1db7afd | 4755 | kfree(vms[area]); |
ca23e405 | 4756 | } |
f1db7afd | 4757 | err_free2: |
ca23e405 TH |
4758 | kfree(vas); |
4759 | kfree(vms); | |
4760 | return NULL; | |
253a496d DA |
4761 | |
4762 | err_free_shadow: | |
4763 | spin_lock(&free_vmap_area_lock); | |
4764 | /* | |
4765 | * We release all the vmalloc shadows, even the ones for regions that | |
4766 | * hadn't been successfully added. This relies on kasan_release_vmalloc | |
4767 | * being able to tolerate this case. | |
4768 | */ | |
4769 | for (area = 0; area < nr_vms; area++) { | |
4770 | orig_start = vas[area]->va_start; | |
4771 | orig_end = vas[area]->va_end; | |
96e2db45 URS |
4772 | va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root, |
4773 | &free_vmap_area_list); | |
9c801f61 URS |
4774 | if (va) |
4775 | kasan_release_vmalloc(orig_start, orig_end, | |
4776 | va->va_start, va->va_end); | |
253a496d DA |
4777 | vas[area] = NULL; |
4778 | kfree(vms[area]); | |
4779 | } | |
4780 | spin_unlock(&free_vmap_area_lock); | |
4781 | kfree(vas); | |
4782 | kfree(vms); | |
4783 | return NULL; | |
ca23e405 TH |
4784 | } |
4785 | ||
4786 | /** | |
4787 | * pcpu_free_vm_areas - free vmalloc areas for percpu allocator | |
4788 | * @vms: vm_struct pointer array returned by pcpu_get_vm_areas() | |
4789 | * @nr_vms: the number of allocated areas | |
4790 | * | |
4791 | * Free vm_structs and the array allocated by pcpu_get_vm_areas(). | |
4792 | */ | |
4793 | void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms) | |
4794 | { | |
4795 | int i; | |
4796 | ||
4797 | for (i = 0; i < nr_vms; i++) | |
4798 | free_vm_area(vms[i]); | |
4799 | kfree(vms); | |
4800 | } | |
4f8b02b4 | 4801 | #endif /* CONFIG_SMP */ |
a10aa579 | 4802 | |
5bb1bb35 | 4803 | #ifdef CONFIG_PRINTK |
98f18083 PM |
4804 | bool vmalloc_dump_obj(void *object) |
4805 | { | |
0818e739 JFG |
4806 | const void *caller; |
4807 | struct vm_struct *vm; | |
4808 | struct vmap_area *va; | |
d0936029 | 4809 | struct vmap_node *vn; |
0818e739 JFG |
4810 | unsigned long addr; |
4811 | unsigned int nr_pages; | |
98f18083 | 4812 | |
8be4d46e URS |
4813 | addr = PAGE_ALIGN((unsigned long) object); |
4814 | vn = addr_to_node(addr); | |
d0936029 | 4815 | |
8be4d46e | 4816 | if (!spin_trylock(&vn->busy.lock)) |
0818e739 | 4817 | return false; |
d0936029 | 4818 | |
8be4d46e URS |
4819 | va = __find_vmap_area(addr, &vn->busy.root); |
4820 | if (!va || !va->vm) { | |
d0936029 | 4821 | spin_unlock(&vn->busy.lock); |
98f18083 | 4822 | return false; |
0818e739 JFG |
4823 | } |
4824 | ||
4825 | vm = va->vm; | |
8be4d46e | 4826 | addr = (unsigned long) vm->addr; |
0818e739 JFG |
4827 | caller = vm->caller; |
4828 | nr_pages = vm->nr_pages; | |
8be4d46e | 4829 | spin_unlock(&vn->busy.lock); |
d0936029 | 4830 | |
bd34dcd4 | 4831 | pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n", |
0818e739 | 4832 | nr_pages, addr, caller); |
8be4d46e | 4833 | |
98f18083 PM |
4834 | return true; |
4835 | } | |
5bb1bb35 | 4836 | #endif |
98f18083 | 4837 | |
a10aa579 | 4838 | #ifdef CONFIG_PROC_FS |
a47a126a ED |
4839 | static void show_numa_info(struct seq_file *m, struct vm_struct *v) |
4840 | { | |
e5adfffc | 4841 | if (IS_ENABLED(CONFIG_NUMA)) { |
a47a126a | 4842 | unsigned int nr, *counters = m->private; |
51e50b3a | 4843 | unsigned int step = 1U << vm_area_page_order(v); |
a47a126a ED |
4844 | |
4845 | if (!counters) | |
4846 | return; | |
4847 | ||
af12346c WL |
4848 | if (v->flags & VM_UNINITIALIZED) |
4849 | return; | |
7e5b528b DV |
4850 | /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */ |
4851 | smp_rmb(); | |
af12346c | 4852 | |
a47a126a ED |
4853 | memset(counters, 0, nr_node_ids * sizeof(unsigned int)); |
4854 | ||
51e50b3a ED |
4855 | for (nr = 0; nr < v->nr_pages; nr += step) |
4856 | counters[page_to_nid(v->pages[nr])] += step; | |
a47a126a ED |
4857 | for_each_node_state(nr, N_HIGH_MEMORY) |
4858 | if (counters[nr]) | |
4859 | seq_printf(m, " N%u=%u", nr, counters[nr]); | |
4860 | } | |
4861 | } | |
4862 | ||
dd3b8353 URS |
4863 | static void show_purge_info(struct seq_file *m) |
4864 | { | |
282631cb | 4865 | struct vmap_node *vn; |
dd3b8353 | 4866 | struct vmap_area *va; |
282631cb | 4867 | int i; |
dd3b8353 | 4868 | |
282631cb URS |
4869 | for (i = 0; i < nr_vmap_nodes; i++) { |
4870 | vn = &vmap_nodes[i]; | |
dd3b8353 | 4871 | |
282631cb URS |
4872 | spin_lock(&vn->lazy.lock); |
4873 | list_for_each_entry(va, &vn->lazy.head, list) { | |
4874 | seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n", | |
4875 | (void *)va->va_start, (void *)va->va_end, | |
4876 | va->va_end - va->va_start); | |
4877 | } | |
4878 | spin_unlock(&vn->lazy.lock); | |
dd3b8353 URS |
4879 | } |
4880 | } | |
4881 | ||
8e1d743f | 4882 | static int vmalloc_info_show(struct seq_file *m, void *p) |
a10aa579 | 4883 | { |
d0936029 | 4884 | struct vmap_node *vn; |
3f500069 | 4885 | struct vmap_area *va; |
d4033afd | 4886 | struct vm_struct *v; |
8e1d743f | 4887 | int i; |
d4033afd | 4888 | |
8e1d743f URS |
4889 | for (i = 0; i < nr_vmap_nodes; i++) { |
4890 | vn = &vmap_nodes[i]; | |
3f500069 | 4891 | |
8e1d743f URS |
4892 | spin_lock(&vn->busy.lock); |
4893 | list_for_each_entry(va, &vn->busy.head, list) { | |
4894 | if (!va->vm) { | |
4895 | if (va->flags & VMAP_RAM) | |
4896 | seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n", | |
4897 | (void *)va->va_start, (void *)va->va_end, | |
4898 | va->va_end - va->va_start); | |
78c72746 | 4899 | |
8e1d743f URS |
4900 | continue; |
4901 | } | |
d4033afd | 4902 | |
8e1d743f | 4903 | v = va->vm; |
a10aa579 | 4904 | |
8e1d743f URS |
4905 | seq_printf(m, "0x%pK-0x%pK %7ld", |
4906 | v->addr, v->addr + v->size, v->size); | |
a10aa579 | 4907 | |
8e1d743f URS |
4908 | if (v->caller) |
4909 | seq_printf(m, " %pS", v->caller); | |
23016969 | 4910 | |
8e1d743f URS |
4911 | if (v->nr_pages) |
4912 | seq_printf(m, " pages=%d", v->nr_pages); | |
a10aa579 | 4913 | |
8e1d743f URS |
4914 | if (v->phys_addr) |
4915 | seq_printf(m, " phys=%pa", &v->phys_addr); | |
a10aa579 | 4916 | |
8e1d743f URS |
4917 | if (v->flags & VM_IOREMAP) |
4918 | seq_puts(m, " ioremap"); | |
a10aa579 | 4919 | |
902861e3 LT |
4920 | if (v->flags & VM_SPARSE) |
4921 | seq_puts(m, " sparse"); | |
e6f79822 | 4922 | |
8e1d743f URS |
4923 | if (v->flags & VM_ALLOC) |
4924 | seq_puts(m, " vmalloc"); | |
a10aa579 | 4925 | |
8e1d743f URS |
4926 | if (v->flags & VM_MAP) |
4927 | seq_puts(m, " vmap"); | |
a10aa579 | 4928 | |
8e1d743f URS |
4929 | if (v->flags & VM_USERMAP) |
4930 | seq_puts(m, " user"); | |
a10aa579 | 4931 | |
8e1d743f URS |
4932 | if (v->flags & VM_DMA_COHERENT) |
4933 | seq_puts(m, " dma-coherent"); | |
fe9041c2 | 4934 | |
8e1d743f URS |
4935 | if (is_vmalloc_addr(v->pages)) |
4936 | seq_puts(m, " vpages"); | |
a10aa579 | 4937 | |
8e1d743f URS |
4938 | show_numa_info(m, v); |
4939 | seq_putc(m, '\n'); | |
4940 | } | |
4941 | spin_unlock(&vn->busy.lock); | |
4942 | } | |
dd3b8353 URS |
4943 | |
4944 | /* | |
96e2db45 | 4945 | * As a final step, dump "unpurged" areas. |
dd3b8353 | 4946 | */ |
8e1d743f | 4947 | show_purge_info(m); |
a10aa579 CL |
4948 | return 0; |
4949 | } | |
4950 | ||
5f6a6a9c AD |
4951 | static int __init proc_vmalloc_init(void) |
4952 | { | |
8e1d743f URS |
4953 | void *priv_data = NULL; |
4954 | ||
fddda2b7 | 4955 | if (IS_ENABLED(CONFIG_NUMA)) |
8e1d743f URS |
4956 | priv_data = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL); |
4957 | ||
4958 | proc_create_single_data("vmallocinfo", | |
4959 | 0400, NULL, vmalloc_info_show, priv_data); | |
4960 | ||
5f6a6a9c AD |
4961 | return 0; |
4962 | } | |
4963 | module_init(proc_vmalloc_init); | |
db3808c1 | 4964 | |
a10aa579 | 4965 | #endif |
208162f4 | 4966 | |
d0936029 | 4967 | static void __init vmap_init_free_space(void) |
7fa8cee0 URS |
4968 | { |
4969 | unsigned long vmap_start = 1; | |
4970 | const unsigned long vmap_end = ULONG_MAX; | |
d0936029 URS |
4971 | struct vmap_area *free; |
4972 | struct vm_struct *busy; | |
7fa8cee0 URS |
4973 | |
4974 | /* | |
4975 | * B F B B B F | |
4976 | * -|-----|.....|-----|-----|-----|.....|- | |
4977 | * | The KVA space | | |
4978 | * |<--------------------------------->| | |
4979 | */ | |
d0936029 URS |
4980 | for (busy = vmlist; busy; busy = busy->next) { |
4981 | if ((unsigned long) busy->addr - vmap_start > 0) { | |
7fa8cee0 URS |
4982 | free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); |
4983 | if (!WARN_ON_ONCE(!free)) { | |
4984 | free->va_start = vmap_start; | |
d0936029 | 4985 | free->va_end = (unsigned long) busy->addr; |
7fa8cee0 URS |
4986 | |
4987 | insert_vmap_area_augment(free, NULL, | |
4988 | &free_vmap_area_root, | |
4989 | &free_vmap_area_list); | |
4990 | } | |
4991 | } | |
4992 | ||
d0936029 | 4993 | vmap_start = (unsigned long) busy->addr + busy->size; |
7fa8cee0 URS |
4994 | } |
4995 | ||
4996 | if (vmap_end - vmap_start > 0) { | |
4997 | free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); | |
4998 | if (!WARN_ON_ONCE(!free)) { | |
4999 | free->va_start = vmap_start; | |
5000 | free->va_end = vmap_end; | |
5001 | ||
5002 | insert_vmap_area_augment(free, NULL, | |
5003 | &free_vmap_area_root, | |
5004 | &free_vmap_area_list); | |
5005 | } | |
5006 | } | |
5007 | } | |
5008 | ||
d0936029 URS |
5009 | static void vmap_init_nodes(void) |
5010 | { | |
5011 | struct vmap_node *vn; | |
8f33a2ff URS |
5012 | int i, n; |
5013 | ||
5014 | #if BITS_PER_LONG == 64 | |
15e02a39 URS |
5015 | /* |
5016 | * A high threshold of max nodes is fixed and bound to 128, | |
5017 | * thus a scale factor is 1 for systems where number of cores | |
5018 | * are less or equal to specified threshold. | |
5019 | * | |
5020 | * As for NUMA-aware notes. For bigger systems, for example | |
5021 | * NUMA with multi-sockets, where we can end-up with thousands | |
5022 | * of cores in total, a "sub-numa-clustering" should be added. | |
5023 | * | |
5024 | * In this case a NUMA domain is considered as a single entity | |
5025 | * with dedicated sub-nodes in it which describe one group or | |
5026 | * set of cores. Therefore a per-domain purging is supposed to | |
5027 | * be added as well as a per-domain balancing. | |
5028 | */ | |
8f33a2ff URS |
5029 | n = clamp_t(unsigned int, num_possible_cpus(), 1, 128); |
5030 | ||
5031 | if (n > 1) { | |
5032 | vn = kmalloc_array(n, sizeof(*vn), GFP_NOWAIT | __GFP_NOWARN); | |
5033 | if (vn) { | |
5034 | /* Node partition is 16 pages. */ | |
5035 | vmap_zone_size = (1 << 4) * PAGE_SIZE; | |
5036 | nr_vmap_nodes = n; | |
5037 | vmap_nodes = vn; | |
5038 | } else { | |
5039 | pr_err("Failed to allocate an array. Disable a node layer\n"); | |
5040 | } | |
5041 | } | |
5042 | #endif | |
d0936029 | 5043 | |
8f33a2ff URS |
5044 | for (n = 0; n < nr_vmap_nodes; n++) { |
5045 | vn = &vmap_nodes[n]; | |
d0936029 URS |
5046 | vn->busy.root = RB_ROOT; |
5047 | INIT_LIST_HEAD(&vn->busy.head); | |
5048 | spin_lock_init(&vn->busy.lock); | |
282631cb URS |
5049 | |
5050 | vn->lazy.root = RB_ROOT; | |
5051 | INIT_LIST_HEAD(&vn->lazy.head); | |
5052 | spin_lock_init(&vn->lazy.lock); | |
72210662 | 5053 | |
8f33a2ff URS |
5054 | for (i = 0; i < MAX_VA_SIZE_PAGES; i++) { |
5055 | INIT_LIST_HEAD(&vn->pool[i].head); | |
5056 | WRITE_ONCE(vn->pool[i].len, 0); | |
72210662 URS |
5057 | } |
5058 | ||
5059 | spin_lock_init(&vn->pool_lock); | |
d0936029 URS |
5060 | } |
5061 | } | |
5062 | ||
7679ba6b URS |
5063 | static unsigned long |
5064 | vmap_node_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
5065 | { | |
5066 | unsigned long count; | |
5067 | struct vmap_node *vn; | |
5068 | int i, j; | |
5069 | ||
5070 | for (count = 0, i = 0; i < nr_vmap_nodes; i++) { | |
5071 | vn = &vmap_nodes[i]; | |
5072 | ||
5073 | for (j = 0; j < MAX_VA_SIZE_PAGES; j++) | |
5074 | count += READ_ONCE(vn->pool[j].len); | |
5075 | } | |
5076 | ||
5077 | return count ? count : SHRINK_EMPTY; | |
5078 | } | |
5079 | ||
5080 | static unsigned long | |
5081 | vmap_node_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
5082 | { | |
5083 | int i; | |
5084 | ||
5085 | for (i = 0; i < nr_vmap_nodes; i++) | |
5086 | decay_va_pool_node(&vmap_nodes[i], true); | |
5087 | ||
5088 | return SHRINK_STOP; | |
5089 | } | |
5090 | ||
208162f4 CH |
5091 | void __init vmalloc_init(void) |
5092 | { | |
7679ba6b | 5093 | struct shrinker *vmap_node_shrinker; |
208162f4 | 5094 | struct vmap_area *va; |
d0936029 | 5095 | struct vmap_node *vn; |
208162f4 CH |
5096 | struct vm_struct *tmp; |
5097 | int i; | |
5098 | ||
5099 | /* | |
5100 | * Create the cache for vmap_area objects. | |
5101 | */ | |
5102 | vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC); | |
5103 | ||
5104 | for_each_possible_cpu(i) { | |
5105 | struct vmap_block_queue *vbq; | |
5106 | struct vfree_deferred *p; | |
5107 | ||
5108 | vbq = &per_cpu(vmap_block_queue, i); | |
5109 | spin_lock_init(&vbq->lock); | |
5110 | INIT_LIST_HEAD(&vbq->free); | |
5111 | p = &per_cpu(vfree_deferred, i); | |
5112 | init_llist_head(&p->list); | |
5113 | INIT_WORK(&p->wq, delayed_vfree_work); | |
062eacf5 | 5114 | xa_init(&vbq->vmap_blocks); |
208162f4 CH |
5115 | } |
5116 | ||
d0936029 URS |
5117 | /* |
5118 | * Setup nodes before importing vmlist. | |
5119 | */ | |
5120 | vmap_init_nodes(); | |
5121 | ||
208162f4 CH |
5122 | /* Import existing vmlist entries. */ |
5123 | for (tmp = vmlist; tmp; tmp = tmp->next) { | |
5124 | va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT); | |
5125 | if (WARN_ON_ONCE(!va)) | |
5126 | continue; | |
5127 | ||
5128 | va->va_start = (unsigned long)tmp->addr; | |
5129 | va->va_end = va->va_start + tmp->size; | |
5130 | va->vm = tmp; | |
d0936029 URS |
5131 | |
5132 | vn = addr_to_node(va->va_start); | |
5133 | insert_vmap_area(va, &vn->busy.root, &vn->busy.head); | |
208162f4 CH |
5134 | } |
5135 | ||
5136 | /* | |
5137 | * Now we can initialize a free vmap space. | |
5138 | */ | |
5139 | vmap_init_free_space(); | |
5140 | vmap_initialized = true; | |
7679ba6b URS |
5141 | |
5142 | vmap_node_shrinker = shrinker_alloc(0, "vmap-node"); | |
5143 | if (!vmap_node_shrinker) { | |
5144 | pr_err("Failed to allocate vmap-node shrinker!\n"); | |
5145 | return; | |
5146 | } | |
5147 | ||
5148 | vmap_node_shrinker->count_objects = vmap_node_shrink_count; | |
5149 | vmap_node_shrinker->scan_objects = vmap_node_shrink_scan; | |
5150 | shrinker_register(vmap_node_shrinker); | |
208162f4 | 5151 | } |