Commit | Line | Data |
---|---|---|
f41f2ed4 MS |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
dff03381 | 3 | * HugeTLB Vmemmap Optimization (HVO) |
f41f2ed4 | 4 | * |
dff03381 | 5 | * Copyright (c) 2020, ByteDance. All rights reserved. |
f41f2ed4 MS |
6 | * |
7 | * Author: Muchun Song <songmuchun@bytedance.com> | |
8 | * | |
ee65728e | 9 | * See Documentation/mm/vmemmap_dedup.rst |
f41f2ed4 | 10 | */ |
e9fdff87 MS |
11 | #define pr_fmt(fmt) "HugeTLB: " fmt |
12 | ||
998a2997 MS |
13 | #include <linux/pgtable.h> |
14 | #include <linux/bootmem_info.h> | |
15 | #include <asm/pgalloc.h> | |
16 | #include <asm/tlbflush.h> | |
f41f2ed4 MS |
17 | #include "hugetlb_vmemmap.h" |
18 | ||
998a2997 MS |
19 | /** |
20 | * struct vmemmap_remap_walk - walk vmemmap page table | |
21 | * | |
22 | * @remap_pte: called for each lowest-level entry (PTE). | |
23 | * @nr_walked: the number of walked pte. | |
24 | * @reuse_page: the page which is reused for the tail vmemmap pages. | |
25 | * @reuse_addr: the virtual address of the @reuse_page page. | |
26 | * @vmemmap_pages: the list head of the vmemmap pages that can be freed | |
27 | * or is mapped from. | |
28 | */ | |
29 | struct vmemmap_remap_walk { | |
30 | void (*remap_pte)(pte_t *pte, unsigned long addr, | |
31 | struct vmemmap_remap_walk *walk); | |
32 | unsigned long nr_walked; | |
33 | struct page *reuse_page; | |
34 | unsigned long reuse_addr; | |
35 | struct list_head *vmemmap_pages; | |
36 | }; | |
37 | ||
998a2997 MS |
38 | static int __split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start) |
39 | { | |
40 | pmd_t __pmd; | |
41 | int i; | |
42 | unsigned long addr = start; | |
43 | struct page *page = pmd_page(*pmd); | |
44 | pte_t *pgtable = pte_alloc_one_kernel(&init_mm); | |
45 | ||
46 | if (!pgtable) | |
47 | return -ENOMEM; | |
48 | ||
49 | pmd_populate_kernel(&init_mm, &__pmd, pgtable); | |
50 | ||
e38f055d | 51 | for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) { |
998a2997 MS |
52 | pte_t entry, *pte; |
53 | pgprot_t pgprot = PAGE_KERNEL; | |
54 | ||
55 | entry = mk_pte(page + i, pgprot); | |
56 | pte = pte_offset_kernel(&__pmd, addr); | |
57 | set_pte_at(&init_mm, addr, pte, entry); | |
58 | } | |
59 | ||
60 | spin_lock(&init_mm.page_table_lock); | |
61 | if (likely(pmd_leaf(*pmd))) { | |
62 | /* | |
63 | * Higher order allocations from buddy allocator must be able to | |
64 | * be treated as indepdenent small pages (as they can be freed | |
65 | * individually). | |
66 | */ | |
67 | if (!PageReserved(page)) | |
68 | split_page(page, get_order(PMD_SIZE)); | |
69 | ||
70 | /* Make pte visible before pmd. See comment in pmd_install(). */ | |
71 | smp_wmb(); | |
72 | pmd_populate_kernel(&init_mm, pmd, pgtable); | |
73 | flush_tlb_kernel_range(start, start + PMD_SIZE); | |
74 | } else { | |
75 | pte_free_kernel(&init_mm, pgtable); | |
76 | } | |
77 | spin_unlock(&init_mm.page_table_lock); | |
78 | ||
79 | return 0; | |
80 | } | |
81 | ||
82 | static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start) | |
83 | { | |
84 | int leaf; | |
85 | ||
86 | spin_lock(&init_mm.page_table_lock); | |
87 | leaf = pmd_leaf(*pmd); | |
88 | spin_unlock(&init_mm.page_table_lock); | |
89 | ||
90 | if (!leaf) | |
91 | return 0; | |
92 | ||
93 | return __split_vmemmap_huge_pmd(pmd, start); | |
94 | } | |
95 | ||
96 | static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr, | |
97 | unsigned long end, | |
98 | struct vmemmap_remap_walk *walk) | |
99 | { | |
100 | pte_t *pte = pte_offset_kernel(pmd, addr); | |
101 | ||
102 | /* | |
103 | * The reuse_page is found 'first' in table walk before we start | |
104 | * remapping (which is calling @walk->remap_pte). | |
105 | */ | |
106 | if (!walk->reuse_page) { | |
107 | walk->reuse_page = pte_page(*pte); | |
108 | /* | |
109 | * Because the reuse address is part of the range that we are | |
110 | * walking, skip the reuse address range. | |
111 | */ | |
112 | addr += PAGE_SIZE; | |
113 | pte++; | |
114 | walk->nr_walked++; | |
115 | } | |
116 | ||
117 | for (; addr != end; addr += PAGE_SIZE, pte++) { | |
118 | walk->remap_pte(pte, addr, walk); | |
119 | walk->nr_walked++; | |
120 | } | |
121 | } | |
122 | ||
123 | static int vmemmap_pmd_range(pud_t *pud, unsigned long addr, | |
124 | unsigned long end, | |
125 | struct vmemmap_remap_walk *walk) | |
126 | { | |
127 | pmd_t *pmd; | |
128 | unsigned long next; | |
129 | ||
130 | pmd = pmd_offset(pud, addr); | |
131 | do { | |
132 | int ret; | |
133 | ||
134 | ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK); | |
135 | if (ret) | |
136 | return ret; | |
137 | ||
138 | next = pmd_addr_end(addr, end); | |
139 | vmemmap_pte_range(pmd, addr, next, walk); | |
140 | } while (pmd++, addr = next, addr != end); | |
141 | ||
142 | return 0; | |
143 | } | |
144 | ||
145 | static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr, | |
146 | unsigned long end, | |
147 | struct vmemmap_remap_walk *walk) | |
148 | { | |
149 | pud_t *pud; | |
150 | unsigned long next; | |
151 | ||
152 | pud = pud_offset(p4d, addr); | |
153 | do { | |
154 | int ret; | |
155 | ||
156 | next = pud_addr_end(addr, end); | |
157 | ret = vmemmap_pmd_range(pud, addr, next, walk); | |
158 | if (ret) | |
159 | return ret; | |
160 | } while (pud++, addr = next, addr != end); | |
161 | ||
162 | return 0; | |
163 | } | |
164 | ||
165 | static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr, | |
166 | unsigned long end, | |
167 | struct vmemmap_remap_walk *walk) | |
168 | { | |
169 | p4d_t *p4d; | |
170 | unsigned long next; | |
171 | ||
172 | p4d = p4d_offset(pgd, addr); | |
173 | do { | |
174 | int ret; | |
175 | ||
176 | next = p4d_addr_end(addr, end); | |
177 | ret = vmemmap_pud_range(p4d, addr, next, walk); | |
178 | if (ret) | |
179 | return ret; | |
180 | } while (p4d++, addr = next, addr != end); | |
181 | ||
182 | return 0; | |
183 | } | |
184 | ||
185 | static int vmemmap_remap_range(unsigned long start, unsigned long end, | |
186 | struct vmemmap_remap_walk *walk) | |
187 | { | |
188 | unsigned long addr = start; | |
189 | unsigned long next; | |
190 | pgd_t *pgd; | |
191 | ||
192 | VM_BUG_ON(!PAGE_ALIGNED(start)); | |
193 | VM_BUG_ON(!PAGE_ALIGNED(end)); | |
194 | ||
195 | pgd = pgd_offset_k(addr); | |
196 | do { | |
197 | int ret; | |
198 | ||
199 | next = pgd_addr_end(addr, end); | |
200 | ret = vmemmap_p4d_range(pgd, addr, next, walk); | |
201 | if (ret) | |
202 | return ret; | |
203 | } while (pgd++, addr = next, addr != end); | |
204 | ||
205 | /* | |
206 | * We only change the mapping of the vmemmap virtual address range | |
207 | * [@start + PAGE_SIZE, end), so we only need to flush the TLB which | |
208 | * belongs to the range. | |
209 | */ | |
210 | flush_tlb_kernel_range(start + PAGE_SIZE, end); | |
211 | ||
212 | return 0; | |
213 | } | |
214 | ||
215 | /* | |
216 | * Free a vmemmap page. A vmemmap page can be allocated from the memblock | |
217 | * allocator or buddy allocator. If the PG_reserved flag is set, it means | |
218 | * that it allocated from the memblock allocator, just free it via the | |
219 | * free_bootmem_page(). Otherwise, use __free_page(). | |
220 | */ | |
221 | static inline void free_vmemmap_page(struct page *page) | |
222 | { | |
223 | if (PageReserved(page)) | |
224 | free_bootmem_page(page); | |
225 | else | |
226 | __free_page(page); | |
227 | } | |
228 | ||
229 | /* Free a list of the vmemmap pages */ | |
230 | static void free_vmemmap_page_list(struct list_head *list) | |
231 | { | |
232 | struct page *page, *next; | |
233 | ||
234 | list_for_each_entry_safe(page, next, list, lru) { | |
235 | list_del(&page->lru); | |
236 | free_vmemmap_page(page); | |
237 | } | |
238 | } | |
239 | ||
240 | static void vmemmap_remap_pte(pte_t *pte, unsigned long addr, | |
241 | struct vmemmap_remap_walk *walk) | |
242 | { | |
243 | /* | |
244 | * Remap the tail pages as read-only to catch illegal write operation | |
245 | * to the tail pages. | |
246 | */ | |
247 | pgprot_t pgprot = PAGE_KERNEL_RO; | |
248 | pte_t entry = mk_pte(walk->reuse_page, pgprot); | |
249 | struct page *page = pte_page(*pte); | |
250 | ||
251 | list_add_tail(&page->lru, walk->vmemmap_pages); | |
252 | set_pte_at(&init_mm, addr, pte, entry); | |
253 | } | |
254 | ||
255 | /* | |
256 | * How many struct page structs need to be reset. When we reuse the head | |
257 | * struct page, the special metadata (e.g. page->flags or page->mapping) | |
258 | * cannot copy to the tail struct page structs. The invalid value will be | |
259 | * checked in the free_tail_pages_check(). In order to avoid the message | |
260 | * of "corrupted mapping in tail page". We need to reset at least 3 (one | |
261 | * head struct page struct and two tail struct page structs) struct page | |
262 | * structs. | |
263 | */ | |
264 | #define NR_RESET_STRUCT_PAGE 3 | |
265 | ||
266 | static inline void reset_struct_pages(struct page *start) | |
267 | { | |
998a2997 MS |
268 | struct page *from = start + NR_RESET_STRUCT_PAGE; |
269 | ||
33febb51 MS |
270 | BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page)); |
271 | memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE); | |
998a2997 MS |
272 | } |
273 | ||
274 | static void vmemmap_restore_pte(pte_t *pte, unsigned long addr, | |
275 | struct vmemmap_remap_walk *walk) | |
276 | { | |
277 | pgprot_t pgprot = PAGE_KERNEL; | |
278 | struct page *page; | |
279 | void *to; | |
280 | ||
281 | BUG_ON(pte_page(*pte) != walk->reuse_page); | |
282 | ||
283 | page = list_first_entry(walk->vmemmap_pages, struct page, lru); | |
284 | list_del(&page->lru); | |
285 | to = page_to_virt(page); | |
286 | copy_page(to, (void *)walk->reuse_addr); | |
287 | reset_struct_pages(to); | |
288 | ||
939de63d ML |
289 | /* |
290 | * Makes sure that preceding stores to the page contents become visible | |
291 | * before the set_pte_at() write. | |
292 | */ | |
293 | smp_wmb(); | |
998a2997 MS |
294 | set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot)); |
295 | } | |
296 | ||
297 | /** | |
298 | * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end) | |
299 | * to the page which @reuse is mapped to, then free vmemmap | |
300 | * which the range are mapped to. | |
301 | * @start: start address of the vmemmap virtual address range that we want | |
302 | * to remap. | |
303 | * @end: end address of the vmemmap virtual address range that we want to | |
304 | * remap. | |
305 | * @reuse: reuse address. | |
306 | * | |
307 | * Return: %0 on success, negative error code otherwise. | |
308 | */ | |
309 | static int vmemmap_remap_free(unsigned long start, unsigned long end, | |
310 | unsigned long reuse) | |
311 | { | |
312 | int ret; | |
313 | LIST_HEAD(vmemmap_pages); | |
314 | struct vmemmap_remap_walk walk = { | |
315 | .remap_pte = vmemmap_remap_pte, | |
316 | .reuse_addr = reuse, | |
317 | .vmemmap_pages = &vmemmap_pages, | |
318 | }; | |
319 | ||
320 | /* | |
321 | * In order to make remapping routine most efficient for the huge pages, | |
322 | * the routine of vmemmap page table walking has the following rules | |
323 | * (see more details from the vmemmap_pte_range()): | |
324 | * | |
325 | * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE) | |
326 | * should be continuous. | |
327 | * - The @reuse address is part of the range [@reuse, @end) that we are | |
328 | * walking which is passed to vmemmap_remap_range(). | |
329 | * - The @reuse address is the first in the complete range. | |
330 | * | |
331 | * So we need to make sure that @start and @reuse meet the above rules. | |
332 | */ | |
333 | BUG_ON(start - reuse != PAGE_SIZE); | |
334 | ||
335 | mmap_read_lock(&init_mm); | |
336 | ret = vmemmap_remap_range(reuse, end, &walk); | |
337 | if (ret && walk.nr_walked) { | |
338 | end = reuse + walk.nr_walked * PAGE_SIZE; | |
339 | /* | |
340 | * vmemmap_pages contains pages from the previous | |
341 | * vmemmap_remap_range call which failed. These | |
342 | * are pages which were removed from the vmemmap. | |
343 | * They will be restored in the following call. | |
344 | */ | |
345 | walk = (struct vmemmap_remap_walk) { | |
346 | .remap_pte = vmemmap_restore_pte, | |
347 | .reuse_addr = reuse, | |
348 | .vmemmap_pages = &vmemmap_pages, | |
349 | }; | |
350 | ||
351 | vmemmap_remap_range(reuse, end, &walk); | |
352 | } | |
353 | mmap_read_unlock(&init_mm); | |
354 | ||
355 | free_vmemmap_page_list(&vmemmap_pages); | |
356 | ||
357 | return ret; | |
358 | } | |
359 | ||
360 | static int alloc_vmemmap_page_list(unsigned long start, unsigned long end, | |
361 | gfp_t gfp_mask, struct list_head *list) | |
362 | { | |
363 | unsigned long nr_pages = (end - start) >> PAGE_SHIFT; | |
364 | int nid = page_to_nid((struct page *)start); | |
365 | struct page *page, *next; | |
366 | ||
367 | while (nr_pages--) { | |
368 | page = alloc_pages_node(nid, gfp_mask, 0); | |
369 | if (!page) | |
370 | goto out; | |
371 | list_add_tail(&page->lru, list); | |
372 | } | |
373 | ||
374 | return 0; | |
375 | out: | |
376 | list_for_each_entry_safe(page, next, list, lru) | |
377 | __free_pages(page, 0); | |
378 | return -ENOMEM; | |
379 | } | |
380 | ||
381 | /** | |
382 | * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end) | |
383 | * to the page which is from the @vmemmap_pages | |
384 | * respectively. | |
385 | * @start: start address of the vmemmap virtual address range that we want | |
386 | * to remap. | |
387 | * @end: end address of the vmemmap virtual address range that we want to | |
388 | * remap. | |
389 | * @reuse: reuse address. | |
390 | * @gfp_mask: GFP flag for allocating vmemmap pages. | |
391 | * | |
392 | * Return: %0 on success, negative error code otherwise. | |
393 | */ | |
394 | static int vmemmap_remap_alloc(unsigned long start, unsigned long end, | |
395 | unsigned long reuse, gfp_t gfp_mask) | |
396 | { | |
397 | LIST_HEAD(vmemmap_pages); | |
398 | struct vmemmap_remap_walk walk = { | |
399 | .remap_pte = vmemmap_restore_pte, | |
400 | .reuse_addr = reuse, | |
401 | .vmemmap_pages = &vmemmap_pages, | |
402 | }; | |
403 | ||
404 | /* See the comment in the vmemmap_remap_free(). */ | |
405 | BUG_ON(start - reuse != PAGE_SIZE); | |
406 | ||
407 | if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages)) | |
408 | return -ENOMEM; | |
409 | ||
410 | mmap_read_lock(&init_mm); | |
411 | vmemmap_remap_range(reuse, end, &walk); | |
412 | mmap_read_unlock(&init_mm); | |
413 | ||
414 | return 0; | |
415 | } | |
416 | ||
cf5472e5 | 417 | DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key); |
f10f1442 | 418 | EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key); |
e9fdff87 | 419 | |
30152245 MS |
420 | static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON); |
421 | core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0); | |
f41f2ed4 | 422 | |
6213834c MS |
423 | /** |
424 | * hugetlb_vmemmap_restore - restore previously optimized (by | |
425 | * hugetlb_vmemmap_optimize()) vmemmap pages which | |
426 | * will be reallocated and remapped. | |
427 | * @h: struct hstate. | |
428 | * @head: the head page whose vmemmap pages will be restored. | |
429 | * | |
430 | * Return: %0 if @head's vmemmap pages have been reallocated and remapped, | |
431 | * negative error code otherwise. | |
ad2fa371 | 432 | */ |
6213834c | 433 | int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head) |
ad2fa371 MS |
434 | { |
435 | int ret; | |
6213834c MS |
436 | unsigned long vmemmap_start = (unsigned long)head, vmemmap_end; |
437 | unsigned long vmemmap_reuse; | |
ad2fa371 MS |
438 | |
439 | if (!HPageVmemmapOptimized(head)) | |
440 | return 0; | |
441 | ||
6213834c MS |
442 | vmemmap_end = vmemmap_start + hugetlb_vmemmap_size(h); |
443 | vmemmap_reuse = vmemmap_start; | |
444 | vmemmap_start += HUGETLB_VMEMMAP_RESERVE_SIZE; | |
5981611d | 445 | |
ad2fa371 | 446 | /* |
6213834c | 447 | * The pages which the vmemmap virtual address range [@vmemmap_start, |
ad2fa371 MS |
448 | * @vmemmap_end) are mapped to are freed to the buddy allocator, and |
449 | * the range is mapped to the page which @vmemmap_reuse is mapped to. | |
450 | * When a HugeTLB page is freed to the buddy allocator, previously | |
451 | * discarded vmemmap pages must be allocated and remapping. | |
452 | */ | |
6213834c | 453 | ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse, |
ad2fa371 | 454 | GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE); |
78f39084 | 455 | if (!ret) { |
ad2fa371 | 456 | ClearHPageVmemmapOptimized(head); |
78f39084 MS |
457 | static_branch_dec(&hugetlb_optimize_vmemmap_key); |
458 | } | |
ad2fa371 MS |
459 | |
460 | return ret; | |
461 | } | |
462 | ||
6213834c MS |
463 | /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */ |
464 | static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head) | |
66361095 | 465 | { |
cf5472e5 | 466 | if (!READ_ONCE(vmemmap_optimize_enabled)) |
6213834c MS |
467 | return false; |
468 | ||
469 | if (!hugetlb_vmemmap_optimizable(h)) | |
470 | return false; | |
66361095 MS |
471 | |
472 | if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) { | |
473 | pmd_t *pmdp, pmd; | |
474 | struct page *vmemmap_page; | |
475 | unsigned long vaddr = (unsigned long)head; | |
476 | ||
477 | /* | |
478 | * Only the vmemmap page's vmemmap page can be self-hosted. | |
479 | * Walking the page tables to find the backing page of the | |
480 | * vmemmap page. | |
481 | */ | |
482 | pmdp = pmd_off_k(vaddr); | |
483 | /* | |
484 | * The READ_ONCE() is used to stabilize *pmdp in a register or | |
485 | * on the stack so that it will stop changing under the code. | |
486 | * The only concurrent operation where it can be changed is | |
487 | * split_vmemmap_huge_pmd() (*pmdp will be stable after this | |
488 | * operation). | |
489 | */ | |
490 | pmd = READ_ONCE(*pmdp); | |
491 | if (pmd_leaf(pmd)) | |
492 | vmemmap_page = pmd_page(pmd) + pte_index(vaddr); | |
493 | else | |
494 | vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr)); | |
495 | /* | |
496 | * Due to HugeTLB alignment requirements and the vmemmap pages | |
497 | * being at the start of the hotplugged memory region in | |
498 | * memory_hotplug.memmap_on_memory case. Checking any vmemmap | |
499 | * page's vmemmap page if it is marked as VmemmapSelfHosted is | |
500 | * sufficient. | |
501 | * | |
502 | * [ hotplugged memory ] | |
503 | * [ section ][...][ section ] | |
504 | * [ vmemmap ][ usable memory ] | |
505 | * ^ | | | | |
506 | * +---+ | | | |
507 | * ^ | | | |
508 | * +-------+ | | |
509 | * ^ | | |
510 | * +-------------------------------------------+ | |
511 | */ | |
512 | if (PageVmemmapSelfHosted(vmemmap_page)) | |
6213834c | 513 | return false; |
66361095 MS |
514 | } |
515 | ||
6213834c | 516 | return true; |
66361095 MS |
517 | } |
518 | ||
6213834c MS |
519 | /** |
520 | * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages. | |
521 | * @h: struct hstate. | |
522 | * @head: the head page whose vmemmap pages will be optimized. | |
523 | * | |
524 | * This function only tries to optimize @head's vmemmap pages and does not | |
525 | * guarantee that the optimization will succeed after it returns. The caller | |
526 | * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages | |
527 | * have been optimized. | |
528 | */ | |
529 | void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head) | |
f41f2ed4 | 530 | { |
6213834c MS |
531 | unsigned long vmemmap_start = (unsigned long)head, vmemmap_end; |
532 | unsigned long vmemmap_reuse; | |
f41f2ed4 | 533 | |
6213834c | 534 | if (!vmemmap_should_optimize(h, head)) |
f41f2ed4 MS |
535 | return; |
536 | ||
78f39084 MS |
537 | static_branch_inc(&hugetlb_optimize_vmemmap_key); |
538 | ||
6213834c MS |
539 | vmemmap_end = vmemmap_start + hugetlb_vmemmap_size(h); |
540 | vmemmap_reuse = vmemmap_start; | |
541 | vmemmap_start += HUGETLB_VMEMMAP_RESERVE_SIZE; | |
f41f2ed4 MS |
542 | |
543 | /* | |
6213834c | 544 | * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end) |
f41f2ed4 | 545 | * to the page which @vmemmap_reuse is mapped to, then free the pages |
6213834c | 546 | * which the range [@vmemmap_start, @vmemmap_end] is mapped to. |
f41f2ed4 | 547 | */ |
6213834c | 548 | if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse)) |
78f39084 MS |
549 | static_branch_dec(&hugetlb_optimize_vmemmap_key); |
550 | else | |
3bc2b6a7 | 551 | SetHPageVmemmapOptimized(head); |
f41f2ed4 | 552 | } |
77490587 | 553 | |
78f39084 MS |
554 | static struct ctl_table hugetlb_vmemmap_sysctls[] = { |
555 | { | |
556 | .procname = "hugetlb_optimize_vmemmap", | |
cf5472e5 MS |
557 | .data = &vmemmap_optimize_enabled, |
558 | .maxlen = sizeof(int), | |
78f39084 | 559 | .mode = 0644, |
cf5472e5 | 560 | .proc_handler = proc_dobool, |
78f39084 MS |
561 | }, |
562 | { } | |
563 | }; | |
564 | ||
6213834c | 565 | static int __init hugetlb_vmemmap_init(void) |
78f39084 | 566 | { |
6213834c MS |
567 | /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */ |
568 | BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE); | |
569 | ||
570 | if (IS_ENABLED(CONFIG_PROC_SYSCTL)) { | |
571 | const struct hstate *h; | |
572 | ||
573 | for_each_hstate(h) { | |
574 | if (hugetlb_vmemmap_optimizable(h)) { | |
575 | register_sysctl_init("vm", hugetlb_vmemmap_sysctls); | |
576 | break; | |
577 | } | |
578 | } | |
579 | } | |
78f39084 MS |
580 | return 0; |
581 | } | |
6213834c | 582 | late_initcall(hugetlb_vmemmap_init); |