]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
eefa864b JK |
2 | #include <linux/mm.h> |
3 | #include <linux/mmzone.h> | |
57c8a661 | 4 | #include <linux/memblock.h> |
eefa864b JK |
5 | #include <linux/page_ext.h> |
6 | #include <linux/memory.h> | |
7 | #include <linux/vmalloc.h> | |
8 | #include <linux/kmemleak.h> | |
48c96a36 | 9 | #include <linux/page_owner.h> |
33c3fc71 | 10 | #include <linux/page_idle.h> |
df4e817b | 11 | #include <linux/page_table_check.h> |
b1d5488a | 12 | #include <linux/rcupdate.h> |
dcfe378c | 13 | #include <linux/pgalloc_tag.h> |
eefa864b JK |
14 | |
15 | /* | |
16 | * struct page extension | |
17 | * | |
18 | * This is the feature to manage memory for extended data per page. | |
19 | * | |
20 | * Until now, we must modify struct page itself to store extra data per page. | |
21 | * This requires rebuilding the kernel and it is really time consuming process. | |
22 | * And, sometimes, rebuild is impossible due to third party module dependency. | |
23 | * At last, enlarging struct page could cause un-wanted system behaviour change. | |
24 | * | |
25 | * This feature is intended to overcome above mentioned problems. This feature | |
26 | * allocates memory for extended data per page in certain place rather than | |
27 | * the struct page itself. This memory can be accessed by the accessor | |
28 | * functions provided by this code. During the boot process, it checks whether | |
29 | * allocation of huge chunk of memory is needed or not. If not, it avoids | |
30 | * allocating memory at all. With this advantage, we can include this feature | |
31 | * into the kernel in default and can avoid rebuild and solve related problems. | |
32 | * | |
33 | * To help these things to work well, there are two callbacks for clients. One | |
34 | * is the need callback which is mandatory if user wants to avoid useless | |
35 | * memory allocation at boot-time. The other is optional, init callback, which | |
36 | * is used to do proper initialization after memory is allocated. | |
37 | * | |
38 | * The need callback is used to decide whether extended memory allocation is | |
39 | * needed or not. Sometimes users want to deactivate some features in this | |
8958b249 | 40 | * boot and extra memory would be unnecessary. In this case, to avoid |
eefa864b JK |
41 | * allocating huge chunk of memory, each clients represent their need of |
42 | * extra memory through the need callback. If one of the need callbacks | |
43 | * returns true, it means that someone needs extra memory so that | |
44 | * page extension core should allocates memory for page extension. If | |
45 | * none of need callbacks return true, memory isn't needed at all in this boot | |
46 | * and page extension core can skip to allocate memory. As result, | |
47 | * none of memory is wasted. | |
48 | * | |
980ac167 JK |
49 | * When need callback returns true, page_ext checks if there is a request for |
50 | * extra memory through size in struct page_ext_operations. If it is non-zero, | |
51 | * extra space is allocated for each page_ext entry and offset is returned to | |
52 | * user through offset in struct page_ext_operations. | |
53 | * | |
eefa864b JK |
54 | * The init callback is used to do proper initialization after page extension |
55 | * is completely initialized. In sparse memory system, extra memory is | |
56 | * allocated some time later than memmap is allocated. In other words, lifetime | |
57 | * of memory for page extension isn't same with memmap for struct page. | |
58 | * Therefore, clients can't store extra data until page extension is | |
59 | * initialized, even if pages are allocated and used freely. This could | |
60 | * cause inadequate state of extra data per page, so, to prevent it, client | |
61 | * can utilize this callback to initialize the state of it correctly. | |
62 | */ | |
63 | ||
b1d5488a CTK |
64 | #ifdef CONFIG_SPARSEMEM |
65 | #define PAGE_EXT_INVALID (0x1) | |
66 | #endif | |
67 | ||
1c676e0d SP |
68 | #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT) |
69 | static bool need_page_idle(void) | |
70 | { | |
71 | return true; | |
72 | } | |
cab0a7c1 | 73 | static struct page_ext_operations page_idle_ops __initdata = { |
1c676e0d | 74 | .need = need_page_idle, |
6189eb82 | 75 | .need_shared_flags = true, |
1c676e0d SP |
76 | }; |
77 | #endif | |
78 | ||
cab0a7c1 | 79 | static struct page_ext_operations *page_ext_ops[] __initdata = { |
48c96a36 JK |
80 | #ifdef CONFIG_PAGE_OWNER |
81 | &page_owner_ops, | |
82 | #endif | |
1c676e0d | 83 | #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT) |
33c3fc71 VD |
84 | &page_idle_ops, |
85 | #endif | |
dcfe378c SB |
86 | #ifdef CONFIG_MEM_ALLOC_PROFILING |
87 | &page_alloc_tagging_ops, | |
88 | #endif | |
df4e817b PT |
89 | #ifdef CONFIG_PAGE_TABLE_CHECK |
90 | &page_table_check_ops, | |
91 | #endif | |
eefa864b JK |
92 | }; |
93 | ||
6189eb82 | 94 | unsigned long page_ext_size; |
5556cfe8 | 95 | |
eefa864b JK |
96 | static unsigned long total_usage; |
97 | ||
26865a1b SB |
98 | #ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG |
99 | /* | |
100 | * To ensure correct allocation tagging for pages, page_ext should be available | |
101 | * before the first page allocation. Otherwise early task stacks will be | |
102 | * allocated before page_ext initialization and missing tags will be flagged. | |
103 | */ | |
104 | bool early_page_ext __meminitdata = true; | |
105 | #else | |
7ec7096b | 106 | bool early_page_ext __meminitdata; |
26865a1b | 107 | #endif |
c4f20f14 LZ |
108 | static int __init setup_early_page_ext(char *str) |
109 | { | |
110 | early_page_ext = true; | |
111 | return 0; | |
112 | } | |
113 | early_param("early_page_ext", setup_early_page_ext); | |
114 | ||
eefa864b JK |
115 | static bool __init invoke_need_callbacks(void) |
116 | { | |
117 | int i; | |
118 | int entries = ARRAY_SIZE(page_ext_ops); | |
980ac167 | 119 | bool need = false; |
eefa864b JK |
120 | |
121 | for (i = 0; i < entries; i++) { | |
6189eb82 PT |
122 | if (page_ext_ops[i]->need()) { |
123 | if (page_ext_ops[i]->need_shared_flags) { | |
124 | page_ext_size = sizeof(struct page_ext); | |
125 | break; | |
126 | } | |
127 | } | |
128 | } | |
129 | ||
130 | for (i = 0; i < entries; i++) { | |
131 | if (page_ext_ops[i]->need()) { | |
5556cfe8 VB |
132 | page_ext_ops[i]->offset = page_ext_size; |
133 | page_ext_size += page_ext_ops[i]->size; | |
980ac167 JK |
134 | need = true; |
135 | } | |
eefa864b JK |
136 | } |
137 | ||
980ac167 | 138 | return need; |
eefa864b JK |
139 | } |
140 | ||
141 | static void __init invoke_init_callbacks(void) | |
142 | { | |
143 | int i; | |
144 | int entries = ARRAY_SIZE(page_ext_ops); | |
145 | ||
146 | for (i = 0; i < entries; i++) { | |
147 | if (page_ext_ops[i]->init) | |
148 | page_ext_ops[i]->init(); | |
149 | } | |
150 | } | |
151 | ||
980ac167 JK |
152 | static inline struct page_ext *get_entry(void *base, unsigned long index) |
153 | { | |
5556cfe8 | 154 | return base + page_ext_size * index; |
980ac167 JK |
155 | } |
156 | ||
eb0da7f6 KS |
157 | #ifndef CONFIG_SPARSEMEM |
158 | void __init page_ext_init_flatmem_late(void) | |
b1d5488a | 159 | { |
eb0da7f6 | 160 | invoke_init_callbacks(); |
b1d5488a | 161 | } |
eefa864b JK |
162 | |
163 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
164 | { | |
165 | pgdat->node_page_ext = NULL; | |
166 | } | |
167 | ||
b1d5488a | 168 | static struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
169 | { |
170 | unsigned long pfn = page_to_pfn(page); | |
0b06bb3f | 171 | unsigned long index; |
eefa864b JK |
172 | struct page_ext *base; |
173 | ||
b1d5488a | 174 | WARN_ON_ONCE(!rcu_read_lock_held()); |
eefa864b | 175 | base = NODE_DATA(page_to_nid(page))->node_page_ext; |
eefa864b JK |
176 | /* |
177 | * The sanity checks the page allocator does upon freeing a | |
178 | * page can reach here before the page_ext arrays are | |
179 | * allocated when feeding a range of pages to the allocator | |
180 | * for the first time during bootup or memory hotplug. | |
181 | */ | |
182 | if (unlikely(!base)) | |
183 | return NULL; | |
0b06bb3f | 184 | index = pfn - round_down(node_start_pfn(page_to_nid(page)), |
eefa864b | 185 | MAX_ORDER_NR_PAGES); |
980ac167 | 186 | return get_entry(base, index); |
eefa864b JK |
187 | } |
188 | ||
189 | static int __init alloc_node_page_ext(int nid) | |
190 | { | |
191 | struct page_ext *base; | |
192 | unsigned long table_size; | |
193 | unsigned long nr_pages; | |
194 | ||
195 | nr_pages = NODE_DATA(nid)->node_spanned_pages; | |
196 | if (!nr_pages) | |
197 | return 0; | |
198 | ||
199 | /* | |
200 | * Need extra space if node range is not aligned with | |
201 | * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm | |
202 | * checks buddy's status, range could be out of exact node range. | |
203 | */ | |
204 | if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) || | |
205 | !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES)) | |
206 | nr_pages += MAX_ORDER_NR_PAGES; | |
207 | ||
5556cfe8 | 208 | table_size = page_ext_size * nr_pages; |
eefa864b | 209 | |
26fb3dae | 210 | base = memblock_alloc_try_nid( |
eefa864b | 211 | table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS), |
97ad1087 | 212 | MEMBLOCK_ALLOC_ACCESSIBLE, nid); |
eefa864b JK |
213 | if (!base) |
214 | return -ENOMEM; | |
215 | NODE_DATA(nid)->node_page_ext = base; | |
216 | total_usage += table_size; | |
217 | return 0; | |
218 | } | |
219 | ||
220 | void __init page_ext_init_flatmem(void) | |
221 | { | |
222 | ||
223 | int nid, fail; | |
224 | ||
225 | if (!invoke_need_callbacks()) | |
226 | return; | |
227 | ||
228 | for_each_online_node(nid) { | |
229 | fail = alloc_node_page_ext(nid); | |
230 | if (fail) | |
231 | goto fail; | |
232 | } | |
233 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
eefa864b JK |
234 | return; |
235 | ||
236 | fail: | |
237 | pr_crit("allocation of page_ext failed.\n"); | |
238 | panic("Out of memory"); | |
239 | } | |
240 | ||
d1fea155 | 241 | #else /* CONFIG_SPARSEMEM */ |
b1d5488a CTK |
242 | static bool page_ext_invalid(struct page_ext *page_ext) |
243 | { | |
244 | return !page_ext || (((unsigned long)page_ext & PAGE_EXT_INVALID) == PAGE_EXT_INVALID); | |
245 | } | |
eefa864b | 246 | |
b1d5488a | 247 | static struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
248 | { |
249 | unsigned long pfn = page_to_pfn(page); | |
250 | struct mem_section *section = __pfn_to_section(pfn); | |
b1d5488a CTK |
251 | struct page_ext *page_ext = READ_ONCE(section->page_ext); |
252 | ||
253 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
eefa864b JK |
254 | /* |
255 | * The sanity checks the page allocator does upon freeing a | |
256 | * page can reach here before the page_ext arrays are | |
257 | * allocated when feeding a range of pages to the allocator | |
258 | * for the first time during bootup or memory hotplug. | |
259 | */ | |
b1d5488a | 260 | if (page_ext_invalid(page_ext)) |
eefa864b | 261 | return NULL; |
b1d5488a | 262 | return get_entry(page_ext, pfn); |
eefa864b JK |
263 | } |
264 | ||
265 | static void *__meminit alloc_page_ext(size_t size, int nid) | |
266 | { | |
267 | gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN; | |
268 | void *addr = NULL; | |
269 | ||
270 | addr = alloc_pages_exact_nid(nid, size, flags); | |
271 | if (addr) { | |
272 | kmemleak_alloc(addr, size, 1, flags); | |
273 | return addr; | |
274 | } | |
275 | ||
b95046b0 | 276 | addr = vzalloc_node(size, nid); |
eefa864b JK |
277 | |
278 | return addr; | |
279 | } | |
280 | ||
281 | static int __meminit init_section_page_ext(unsigned long pfn, int nid) | |
282 | { | |
283 | struct mem_section *section; | |
284 | struct page_ext *base; | |
285 | unsigned long table_size; | |
286 | ||
287 | section = __pfn_to_section(pfn); | |
288 | ||
289 | if (section->page_ext) | |
290 | return 0; | |
291 | ||
5556cfe8 | 292 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
293 | base = alloc_page_ext(table_size, nid); |
294 | ||
295 | /* | |
296 | * The value stored in section->page_ext is (base - pfn) | |
297 | * and it does not point to the memory block allocated above, | |
298 | * causing kmemleak false positives. | |
299 | */ | |
300 | kmemleak_not_leak(base); | |
301 | ||
302 | if (!base) { | |
303 | pr_err("page ext allocation failure\n"); | |
304 | return -ENOMEM; | |
305 | } | |
306 | ||
307 | /* | |
308 | * The passed "pfn" may not be aligned to SECTION. For the calculation | |
309 | * we need to apply a mask. | |
310 | */ | |
311 | pfn &= PAGE_SECTION_MASK; | |
5556cfe8 | 312 | section->page_ext = (void *)base - page_ext_size * pfn; |
eefa864b JK |
313 | total_usage += table_size; |
314 | return 0; | |
315 | } | |
76af6a05 | 316 | |
eefa864b JK |
317 | static void free_page_ext(void *addr) |
318 | { | |
319 | if (is_vmalloc_addr(addr)) { | |
320 | vfree(addr); | |
321 | } else { | |
322 | struct page *page = virt_to_page(addr); | |
323 | size_t table_size; | |
324 | ||
5556cfe8 | 325 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
326 | |
327 | BUG_ON(PageReserved(page)); | |
0c815854 | 328 | kmemleak_free(addr); |
eefa864b JK |
329 | free_pages_exact(addr, table_size); |
330 | } | |
331 | } | |
332 | ||
333 | static void __free_page_ext(unsigned long pfn) | |
334 | { | |
335 | struct mem_section *ms; | |
336 | struct page_ext *base; | |
337 | ||
338 | ms = __pfn_to_section(pfn); | |
339 | if (!ms || !ms->page_ext) | |
340 | return; | |
b1d5488a CTK |
341 | |
342 | base = READ_ONCE(ms->page_ext); | |
343 | /* | |
344 | * page_ext here can be valid while doing the roll back | |
345 | * operation in online_page_ext(). | |
346 | */ | |
347 | if (page_ext_invalid(base)) | |
348 | base = (void *)base - PAGE_EXT_INVALID; | |
349 | WRITE_ONCE(ms->page_ext, NULL); | |
350 | ||
351 | base = get_entry(base, pfn); | |
eefa864b | 352 | free_page_ext(base); |
b1d5488a CTK |
353 | } |
354 | ||
355 | static void __invalidate_page_ext(unsigned long pfn) | |
356 | { | |
357 | struct mem_section *ms; | |
358 | void *val; | |
359 | ||
360 | ms = __pfn_to_section(pfn); | |
361 | if (!ms || !ms->page_ext) | |
362 | return; | |
363 | val = (void *)ms->page_ext + PAGE_EXT_INVALID; | |
364 | WRITE_ONCE(ms->page_ext, val); | |
eefa864b JK |
365 | } |
366 | ||
367 | static int __meminit online_page_ext(unsigned long start_pfn, | |
368 | unsigned long nr_pages, | |
369 | int nid) | |
370 | { | |
371 | unsigned long start, end, pfn; | |
372 | int fail = 0; | |
373 | ||
374 | start = SECTION_ALIGN_DOWN(start_pfn); | |
375 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
376 | ||
98fa15f3 | 377 | if (nid == NUMA_NO_NODE) { |
eefa864b JK |
378 | /* |
379 | * In this case, "nid" already exists and contains valid memory. | |
380 | * "start_pfn" passed to us is a pfn which is an arg for | |
381 | * online__pages(), and start_pfn should exist. | |
382 | */ | |
383 | nid = pfn_to_nid(start_pfn); | |
30a51400 | 384 | VM_BUG_ON(!node_online(nid)); |
eefa864b JK |
385 | } |
386 | ||
dccacf8d | 387 | for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) |
eefa864b | 388 | fail = init_section_page_ext(pfn, nid); |
eefa864b JK |
389 | if (!fail) |
390 | return 0; | |
391 | ||
392 | /* rollback */ | |
3c09be5a | 393 | end = pfn - PAGES_PER_SECTION; |
eefa864b JK |
394 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) |
395 | __free_page_ext(pfn); | |
396 | ||
397 | return -ENOMEM; | |
398 | } | |
399 | ||
063ff7cd | 400 | static void __meminit offline_page_ext(unsigned long start_pfn, |
7b5a0b66 | 401 | unsigned long nr_pages) |
eefa864b JK |
402 | { |
403 | unsigned long start, end, pfn; | |
404 | ||
405 | start = SECTION_ALIGN_DOWN(start_pfn); | |
406 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
407 | ||
b1d5488a CTK |
408 | /* |
409 | * Freeing of page_ext is done in 3 steps to avoid | |
410 | * use-after-free of it: | |
411 | * 1) Traverse all the sections and mark their page_ext | |
412 | * as invalid. | |
413 | * 2) Wait for all the existing users of page_ext who | |
414 | * started before invalidation to finish. | |
415 | * 3) Free the page_ext. | |
416 | */ | |
417 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
418 | __invalidate_page_ext(pfn); | |
419 | ||
420 | synchronize_rcu(); | |
421 | ||
eefa864b JK |
422 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) |
423 | __free_page_ext(pfn); | |
eefa864b JK |
424 | } |
425 | ||
426 | static int __meminit page_ext_callback(struct notifier_block *self, | |
427 | unsigned long action, void *arg) | |
428 | { | |
429 | struct memory_notify *mn = arg; | |
430 | int ret = 0; | |
431 | ||
432 | switch (action) { | |
433 | case MEM_GOING_ONLINE: | |
434 | ret = online_page_ext(mn->start_pfn, | |
435 | mn->nr_pages, mn->status_change_nid); | |
436 | break; | |
437 | case MEM_OFFLINE: | |
438 | offline_page_ext(mn->start_pfn, | |
7b5a0b66 | 439 | mn->nr_pages); |
eefa864b JK |
440 | break; |
441 | case MEM_CANCEL_ONLINE: | |
442 | offline_page_ext(mn->start_pfn, | |
7b5a0b66 | 443 | mn->nr_pages); |
eefa864b JK |
444 | break; |
445 | case MEM_GOING_OFFLINE: | |
446 | break; | |
447 | case MEM_ONLINE: | |
448 | case MEM_CANCEL_OFFLINE: | |
449 | break; | |
450 | } | |
451 | ||
452 | return notifier_from_errno(ret); | |
453 | } | |
454 | ||
eefa864b JK |
455 | void __init page_ext_init(void) |
456 | { | |
457 | unsigned long pfn; | |
458 | int nid; | |
459 | ||
460 | if (!invoke_need_callbacks()) | |
461 | return; | |
462 | ||
463 | for_each_node_state(nid, N_MEMORY) { | |
464 | unsigned long start_pfn, end_pfn; | |
465 | ||
466 | start_pfn = node_start_pfn(nid); | |
467 | end_pfn = node_end_pfn(nid); | |
468 | /* | |
469 | * start_pfn and end_pfn may not be aligned to SECTION and the | |
470 | * page->flags of out of node pages are not initialized. So we | |
471 | * scan [start_pfn, the biggest section's pfn < end_pfn) here. | |
472 | */ | |
473 | for (pfn = start_pfn; pfn < end_pfn; | |
474 | pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) { | |
475 | ||
476 | if (!pfn_valid(pfn)) | |
477 | continue; | |
478 | /* | |
479 | * Nodes's pfns can be overlapping. | |
480 | * We know some arch can have a nodes layout such as | |
481 | * -------------pfn--------------> | |
482 | * N0 | N1 | N2 | N0 | N1 | N2|.... | |
483 | */ | |
2f1ee091 | 484 | if (pfn_to_nid(pfn) != nid) |
eefa864b JK |
485 | continue; |
486 | if (init_section_page_ext(pfn, nid)) | |
487 | goto oom; | |
0fc542b7 | 488 | cond_resched(); |
eefa864b JK |
489 | } |
490 | } | |
1eeaa4fd | 491 | hotplug_memory_notifier(page_ext_callback, DEFAULT_CALLBACK_PRI); |
eefa864b JK |
492 | pr_info("allocated %ld bytes of page_ext\n", total_usage); |
493 | invoke_init_callbacks(); | |
494 | return; | |
495 | ||
496 | oom: | |
497 | panic("Out of memory"); | |
498 | } | |
499 | ||
500 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
501 | { | |
502 | } | |
503 | ||
504 | #endif | |
eb0da7f6 KS |
505 | |
506 | /** | |
507 | * page_ext_get() - Get the extended information for a page. | |
508 | * @page: The page we're interested in. | |
509 | * | |
510 | * Ensures that the page_ext will remain valid until page_ext_put() | |
511 | * is called. | |
512 | * | |
513 | * Return: NULL if no page_ext exists for this page. | |
514 | * Context: Any context. Caller may not sleep until they have called | |
515 | * page_ext_put(). | |
516 | */ | |
6e65aa55 | 517 | struct page_ext *page_ext_get(const struct page *page) |
eb0da7f6 KS |
518 | { |
519 | struct page_ext *page_ext; | |
520 | ||
521 | rcu_read_lock(); | |
522 | page_ext = lookup_page_ext(page); | |
523 | if (!page_ext) { | |
524 | rcu_read_unlock(); | |
525 | return NULL; | |
526 | } | |
527 | ||
528 | return page_ext; | |
529 | } | |
530 | ||
531 | /** | |
532 | * page_ext_put() - Working with page extended information is done. | |
533 | * @page_ext: Page extended information received from page_ext_get(). | |
534 | * | |
535 | * The page extended information of the page may not be valid after this | |
536 | * function is called. | |
537 | * | |
538 | * Return: None. | |
539 | * Context: Any context with corresponding page_ext_get() is called. | |
540 | */ | |
541 | void page_ext_put(struct page_ext *page_ext) | |
542 | { | |
543 | if (unlikely(!page_ext)) | |
544 | return; | |
545 | ||
546 | rcu_read_unlock(); | |
547 | } |