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> |
eefa864b JK |
13 | |
14 | /* | |
15 | * struct page extension | |
16 | * | |
17 | * This is the feature to manage memory for extended data per page. | |
18 | * | |
19 | * Until now, we must modify struct page itself to store extra data per page. | |
20 | * This requires rebuilding the kernel and it is really time consuming process. | |
21 | * And, sometimes, rebuild is impossible due to third party module dependency. | |
22 | * At last, enlarging struct page could cause un-wanted system behaviour change. | |
23 | * | |
24 | * This feature is intended to overcome above mentioned problems. This feature | |
25 | * allocates memory for extended data per page in certain place rather than | |
26 | * the struct page itself. This memory can be accessed by the accessor | |
27 | * functions provided by this code. During the boot process, it checks whether | |
28 | * allocation of huge chunk of memory is needed or not. If not, it avoids | |
29 | * allocating memory at all. With this advantage, we can include this feature | |
30 | * into the kernel in default and can avoid rebuild and solve related problems. | |
31 | * | |
32 | * To help these things to work well, there are two callbacks for clients. One | |
33 | * is the need callback which is mandatory if user wants to avoid useless | |
34 | * memory allocation at boot-time. The other is optional, init callback, which | |
35 | * is used to do proper initialization after memory is allocated. | |
36 | * | |
37 | * The need callback is used to decide whether extended memory allocation is | |
38 | * needed or not. Sometimes users want to deactivate some features in this | |
8958b249 | 39 | * boot and extra memory would be unnecessary. In this case, to avoid |
eefa864b JK |
40 | * allocating huge chunk of memory, each clients represent their need of |
41 | * extra memory through the need callback. If one of the need callbacks | |
42 | * returns true, it means that someone needs extra memory so that | |
43 | * page extension core should allocates memory for page extension. If | |
44 | * none of need callbacks return true, memory isn't needed at all in this boot | |
45 | * and page extension core can skip to allocate memory. As result, | |
46 | * none of memory is wasted. | |
47 | * | |
980ac167 JK |
48 | * When need callback returns true, page_ext checks if there is a request for |
49 | * extra memory through size in struct page_ext_operations. If it is non-zero, | |
50 | * extra space is allocated for each page_ext entry and offset is returned to | |
51 | * user through offset in struct page_ext_operations. | |
52 | * | |
eefa864b JK |
53 | * The init callback is used to do proper initialization after page extension |
54 | * is completely initialized. In sparse memory system, extra memory is | |
55 | * allocated some time later than memmap is allocated. In other words, lifetime | |
56 | * of memory for page extension isn't same with memmap for struct page. | |
57 | * Therefore, clients can't store extra data until page extension is | |
58 | * initialized, even if pages are allocated and used freely. This could | |
59 | * cause inadequate state of extra data per page, so, to prevent it, client | |
60 | * can utilize this callback to initialize the state of it correctly. | |
61 | */ | |
62 | ||
b1d5488a CTK |
63 | #ifdef CONFIG_SPARSEMEM |
64 | #define PAGE_EXT_INVALID (0x1) | |
65 | #endif | |
66 | ||
1c676e0d SP |
67 | #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT) |
68 | static bool need_page_idle(void) | |
69 | { | |
70 | return true; | |
71 | } | |
cab0a7c1 | 72 | static struct page_ext_operations page_idle_ops __initdata = { |
1c676e0d SP |
73 | .need = need_page_idle, |
74 | }; | |
75 | #endif | |
76 | ||
cab0a7c1 | 77 | static struct page_ext_operations *page_ext_ops[] __initdata = { |
48c96a36 JK |
78 | #ifdef CONFIG_PAGE_OWNER |
79 | &page_owner_ops, | |
80 | #endif | |
1c676e0d | 81 | #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT) |
33c3fc71 VD |
82 | &page_idle_ops, |
83 | #endif | |
df4e817b PT |
84 | #ifdef CONFIG_PAGE_TABLE_CHECK |
85 | &page_table_check_ops, | |
86 | #endif | |
eefa864b JK |
87 | }; |
88 | ||
5556cfe8 VB |
89 | unsigned long page_ext_size = sizeof(struct page_ext); |
90 | ||
eefa864b | 91 | static unsigned long total_usage; |
b1d5488a | 92 | static struct page_ext *lookup_page_ext(const struct page *page); |
eefa864b | 93 | |
c4f20f14 LZ |
94 | bool early_page_ext; |
95 | static int __init setup_early_page_ext(char *str) | |
96 | { | |
97 | early_page_ext = true; | |
98 | return 0; | |
99 | } | |
100 | early_param("early_page_ext", setup_early_page_ext); | |
101 | ||
eefa864b JK |
102 | static bool __init invoke_need_callbacks(void) |
103 | { | |
104 | int i; | |
105 | int entries = ARRAY_SIZE(page_ext_ops); | |
980ac167 | 106 | bool need = false; |
eefa864b JK |
107 | |
108 | for (i = 0; i < entries; i++) { | |
980ac167 | 109 | if (page_ext_ops[i]->need && page_ext_ops[i]->need()) { |
5556cfe8 VB |
110 | page_ext_ops[i]->offset = page_ext_size; |
111 | page_ext_size += page_ext_ops[i]->size; | |
980ac167 JK |
112 | need = true; |
113 | } | |
eefa864b JK |
114 | } |
115 | ||
980ac167 | 116 | return need; |
eefa864b JK |
117 | } |
118 | ||
119 | static void __init invoke_init_callbacks(void) | |
120 | { | |
121 | int i; | |
122 | int entries = ARRAY_SIZE(page_ext_ops); | |
123 | ||
124 | for (i = 0; i < entries; i++) { | |
125 | if (page_ext_ops[i]->init) | |
126 | page_ext_ops[i]->init(); | |
127 | } | |
128 | } | |
129 | ||
7fb7ab6d ZH |
130 | #ifndef CONFIG_SPARSEMEM |
131 | void __init page_ext_init_flatmem_late(void) | |
132 | { | |
133 | invoke_init_callbacks(); | |
134 | } | |
135 | #endif | |
136 | ||
980ac167 JK |
137 | static inline struct page_ext *get_entry(void *base, unsigned long index) |
138 | { | |
5556cfe8 | 139 | return base + page_ext_size * index; |
980ac167 JK |
140 | } |
141 | ||
b1d5488a CTK |
142 | /** |
143 | * page_ext_get() - Get the extended information for a page. | |
144 | * @page: The page we're interested in. | |
145 | * | |
146 | * Ensures that the page_ext will remain valid until page_ext_put() | |
147 | * is called. | |
148 | * | |
149 | * Return: NULL if no page_ext exists for this page. | |
150 | * Context: Any context. Caller may not sleep until they have called | |
151 | * page_ext_put(). | |
152 | */ | |
153 | struct page_ext *page_ext_get(struct page *page) | |
154 | { | |
155 | struct page_ext *page_ext; | |
156 | ||
157 | rcu_read_lock(); | |
158 | page_ext = lookup_page_ext(page); | |
159 | if (!page_ext) { | |
160 | rcu_read_unlock(); | |
161 | return NULL; | |
162 | } | |
163 | ||
164 | return page_ext; | |
165 | } | |
166 | ||
167 | /** | |
168 | * page_ext_put() - Working with page extended information is done. | |
169 | * @page_ext - Page extended information received from page_ext_get(). | |
170 | * | |
171 | * The page extended information of the page may not be valid after this | |
172 | * function is called. | |
173 | * | |
174 | * Return: None. | |
175 | * Context: Any context with corresponding page_ext_get() is called. | |
176 | */ | |
177 | void page_ext_put(struct page_ext *page_ext) | |
178 | { | |
179 | if (unlikely(!page_ext)) | |
180 | return; | |
181 | ||
182 | rcu_read_unlock(); | |
183 | } | |
7fb7ab6d | 184 | #ifndef CONFIG_SPARSEMEM |
eefa864b JK |
185 | |
186 | ||
187 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
188 | { | |
189 | pgdat->node_page_ext = NULL; | |
190 | } | |
191 | ||
b1d5488a | 192 | static struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
193 | { |
194 | unsigned long pfn = page_to_pfn(page); | |
0b06bb3f | 195 | unsigned long index; |
eefa864b JK |
196 | struct page_ext *base; |
197 | ||
b1d5488a | 198 | WARN_ON_ONCE(!rcu_read_lock_held()); |
eefa864b | 199 | base = NODE_DATA(page_to_nid(page))->node_page_ext; |
eefa864b JK |
200 | /* |
201 | * The sanity checks the page allocator does upon freeing a | |
202 | * page can reach here before the page_ext arrays are | |
203 | * allocated when feeding a range of pages to the allocator | |
204 | * for the first time during bootup or memory hotplug. | |
205 | */ | |
206 | if (unlikely(!base)) | |
207 | return NULL; | |
0b06bb3f | 208 | index = pfn - round_down(node_start_pfn(page_to_nid(page)), |
eefa864b | 209 | MAX_ORDER_NR_PAGES); |
980ac167 | 210 | return get_entry(base, index); |
eefa864b JK |
211 | } |
212 | ||
213 | static int __init alloc_node_page_ext(int nid) | |
214 | { | |
215 | struct page_ext *base; | |
216 | unsigned long table_size; | |
217 | unsigned long nr_pages; | |
218 | ||
219 | nr_pages = NODE_DATA(nid)->node_spanned_pages; | |
220 | if (!nr_pages) | |
221 | return 0; | |
222 | ||
223 | /* | |
224 | * Need extra space if node range is not aligned with | |
225 | * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm | |
226 | * checks buddy's status, range could be out of exact node range. | |
227 | */ | |
228 | if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) || | |
229 | !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES)) | |
230 | nr_pages += MAX_ORDER_NR_PAGES; | |
231 | ||
5556cfe8 | 232 | table_size = page_ext_size * nr_pages; |
eefa864b | 233 | |
26fb3dae | 234 | base = memblock_alloc_try_nid( |
eefa864b | 235 | table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS), |
97ad1087 | 236 | MEMBLOCK_ALLOC_ACCESSIBLE, nid); |
eefa864b JK |
237 | if (!base) |
238 | return -ENOMEM; | |
239 | NODE_DATA(nid)->node_page_ext = base; | |
240 | total_usage += table_size; | |
241 | return 0; | |
242 | } | |
243 | ||
244 | void __init page_ext_init_flatmem(void) | |
245 | { | |
246 | ||
247 | int nid, fail; | |
248 | ||
249 | if (!invoke_need_callbacks()) | |
250 | return; | |
251 | ||
252 | for_each_online_node(nid) { | |
253 | fail = alloc_node_page_ext(nid); | |
254 | if (fail) | |
255 | goto fail; | |
256 | } | |
257 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
eefa864b JK |
258 | return; |
259 | ||
260 | fail: | |
261 | pr_crit("allocation of page_ext failed.\n"); | |
262 | panic("Out of memory"); | |
263 | } | |
264 | ||
d1fea155 | 265 | #else /* CONFIG_SPARSEMEM */ |
b1d5488a CTK |
266 | static bool page_ext_invalid(struct page_ext *page_ext) |
267 | { | |
268 | return !page_ext || (((unsigned long)page_ext & PAGE_EXT_INVALID) == PAGE_EXT_INVALID); | |
269 | } | |
eefa864b | 270 | |
b1d5488a | 271 | static struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
272 | { |
273 | unsigned long pfn = page_to_pfn(page); | |
274 | struct mem_section *section = __pfn_to_section(pfn); | |
b1d5488a CTK |
275 | struct page_ext *page_ext = READ_ONCE(section->page_ext); |
276 | ||
277 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
eefa864b JK |
278 | /* |
279 | * The sanity checks the page allocator does upon freeing a | |
280 | * page can reach here before the page_ext arrays are | |
281 | * allocated when feeding a range of pages to the allocator | |
282 | * for the first time during bootup or memory hotplug. | |
283 | */ | |
b1d5488a | 284 | if (page_ext_invalid(page_ext)) |
eefa864b | 285 | return NULL; |
b1d5488a | 286 | return get_entry(page_ext, pfn); |
eefa864b JK |
287 | } |
288 | ||
289 | static void *__meminit alloc_page_ext(size_t size, int nid) | |
290 | { | |
291 | gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN; | |
292 | void *addr = NULL; | |
293 | ||
294 | addr = alloc_pages_exact_nid(nid, size, flags); | |
295 | if (addr) { | |
296 | kmemleak_alloc(addr, size, 1, flags); | |
297 | return addr; | |
298 | } | |
299 | ||
b95046b0 | 300 | addr = vzalloc_node(size, nid); |
eefa864b JK |
301 | |
302 | return addr; | |
303 | } | |
304 | ||
305 | static int __meminit init_section_page_ext(unsigned long pfn, int nid) | |
306 | { | |
307 | struct mem_section *section; | |
308 | struct page_ext *base; | |
309 | unsigned long table_size; | |
310 | ||
311 | section = __pfn_to_section(pfn); | |
312 | ||
313 | if (section->page_ext) | |
314 | return 0; | |
315 | ||
5556cfe8 | 316 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
317 | base = alloc_page_ext(table_size, nid); |
318 | ||
319 | /* | |
320 | * The value stored in section->page_ext is (base - pfn) | |
321 | * and it does not point to the memory block allocated above, | |
322 | * causing kmemleak false positives. | |
323 | */ | |
324 | kmemleak_not_leak(base); | |
325 | ||
326 | if (!base) { | |
327 | pr_err("page ext allocation failure\n"); | |
328 | return -ENOMEM; | |
329 | } | |
330 | ||
331 | /* | |
332 | * The passed "pfn" may not be aligned to SECTION. For the calculation | |
333 | * we need to apply a mask. | |
334 | */ | |
335 | pfn &= PAGE_SECTION_MASK; | |
5556cfe8 | 336 | section->page_ext = (void *)base - page_ext_size * pfn; |
eefa864b JK |
337 | total_usage += table_size; |
338 | return 0; | |
339 | } | |
76af6a05 | 340 | |
eefa864b JK |
341 | static void free_page_ext(void *addr) |
342 | { | |
343 | if (is_vmalloc_addr(addr)) { | |
344 | vfree(addr); | |
345 | } else { | |
346 | struct page *page = virt_to_page(addr); | |
347 | size_t table_size; | |
348 | ||
5556cfe8 | 349 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
350 | |
351 | BUG_ON(PageReserved(page)); | |
0c815854 | 352 | kmemleak_free(addr); |
eefa864b JK |
353 | free_pages_exact(addr, table_size); |
354 | } | |
355 | } | |
356 | ||
357 | static void __free_page_ext(unsigned long pfn) | |
358 | { | |
359 | struct mem_section *ms; | |
360 | struct page_ext *base; | |
361 | ||
362 | ms = __pfn_to_section(pfn); | |
363 | if (!ms || !ms->page_ext) | |
364 | return; | |
b1d5488a CTK |
365 | |
366 | base = READ_ONCE(ms->page_ext); | |
367 | /* | |
368 | * page_ext here can be valid while doing the roll back | |
369 | * operation in online_page_ext(). | |
370 | */ | |
371 | if (page_ext_invalid(base)) | |
372 | base = (void *)base - PAGE_EXT_INVALID; | |
373 | WRITE_ONCE(ms->page_ext, NULL); | |
374 | ||
375 | base = get_entry(base, pfn); | |
eefa864b | 376 | free_page_ext(base); |
b1d5488a CTK |
377 | } |
378 | ||
379 | static void __invalidate_page_ext(unsigned long pfn) | |
380 | { | |
381 | struct mem_section *ms; | |
382 | void *val; | |
383 | ||
384 | ms = __pfn_to_section(pfn); | |
385 | if (!ms || !ms->page_ext) | |
386 | return; | |
387 | val = (void *)ms->page_ext + PAGE_EXT_INVALID; | |
388 | WRITE_ONCE(ms->page_ext, val); | |
eefa864b JK |
389 | } |
390 | ||
391 | static int __meminit online_page_ext(unsigned long start_pfn, | |
392 | unsigned long nr_pages, | |
393 | int nid) | |
394 | { | |
395 | unsigned long start, end, pfn; | |
396 | int fail = 0; | |
397 | ||
398 | start = SECTION_ALIGN_DOWN(start_pfn); | |
399 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
400 | ||
98fa15f3 | 401 | if (nid == NUMA_NO_NODE) { |
eefa864b JK |
402 | /* |
403 | * In this case, "nid" already exists and contains valid memory. | |
404 | * "start_pfn" passed to us is a pfn which is an arg for | |
405 | * online__pages(), and start_pfn should exist. | |
406 | */ | |
407 | nid = pfn_to_nid(start_pfn); | |
30a51400 | 408 | VM_BUG_ON(!node_online(nid)); |
eefa864b JK |
409 | } |
410 | ||
dccacf8d | 411 | for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) |
eefa864b | 412 | fail = init_section_page_ext(pfn, nid); |
eefa864b JK |
413 | if (!fail) |
414 | return 0; | |
415 | ||
416 | /* rollback */ | |
417 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
418 | __free_page_ext(pfn); | |
419 | ||
420 | return -ENOMEM; | |
421 | } | |
422 | ||
423 | static int __meminit offline_page_ext(unsigned long start_pfn, | |
7b5a0b66 | 424 | unsigned long nr_pages) |
eefa864b JK |
425 | { |
426 | unsigned long start, end, pfn; | |
427 | ||
428 | start = SECTION_ALIGN_DOWN(start_pfn); | |
429 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
430 | ||
b1d5488a CTK |
431 | /* |
432 | * Freeing of page_ext is done in 3 steps to avoid | |
433 | * use-after-free of it: | |
434 | * 1) Traverse all the sections and mark their page_ext | |
435 | * as invalid. | |
436 | * 2) Wait for all the existing users of page_ext who | |
437 | * started before invalidation to finish. | |
438 | * 3) Free the page_ext. | |
439 | */ | |
440 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
441 | __invalidate_page_ext(pfn); | |
442 | ||
443 | synchronize_rcu(); | |
444 | ||
eefa864b JK |
445 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) |
446 | __free_page_ext(pfn); | |
447 | return 0; | |
448 | ||
449 | } | |
450 | ||
451 | static int __meminit page_ext_callback(struct notifier_block *self, | |
452 | unsigned long action, void *arg) | |
453 | { | |
454 | struct memory_notify *mn = arg; | |
455 | int ret = 0; | |
456 | ||
457 | switch (action) { | |
458 | case MEM_GOING_ONLINE: | |
459 | ret = online_page_ext(mn->start_pfn, | |
460 | mn->nr_pages, mn->status_change_nid); | |
461 | break; | |
462 | case MEM_OFFLINE: | |
463 | offline_page_ext(mn->start_pfn, | |
7b5a0b66 | 464 | mn->nr_pages); |
eefa864b JK |
465 | break; |
466 | case MEM_CANCEL_ONLINE: | |
467 | offline_page_ext(mn->start_pfn, | |
7b5a0b66 | 468 | mn->nr_pages); |
eefa864b JK |
469 | break; |
470 | case MEM_GOING_OFFLINE: | |
471 | break; | |
472 | case MEM_ONLINE: | |
473 | case MEM_CANCEL_OFFLINE: | |
474 | break; | |
475 | } | |
476 | ||
477 | return notifier_from_errno(ret); | |
478 | } | |
479 | ||
eefa864b JK |
480 | void __init page_ext_init(void) |
481 | { | |
482 | unsigned long pfn; | |
483 | int nid; | |
484 | ||
485 | if (!invoke_need_callbacks()) | |
486 | return; | |
487 | ||
488 | for_each_node_state(nid, N_MEMORY) { | |
489 | unsigned long start_pfn, end_pfn; | |
490 | ||
491 | start_pfn = node_start_pfn(nid); | |
492 | end_pfn = node_end_pfn(nid); | |
493 | /* | |
494 | * start_pfn and end_pfn may not be aligned to SECTION and the | |
495 | * page->flags of out of node pages are not initialized. So we | |
496 | * scan [start_pfn, the biggest section's pfn < end_pfn) here. | |
497 | */ | |
498 | for (pfn = start_pfn; pfn < end_pfn; | |
499 | pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) { | |
500 | ||
501 | if (!pfn_valid(pfn)) | |
502 | continue; | |
503 | /* | |
504 | * Nodes's pfns can be overlapping. | |
505 | * We know some arch can have a nodes layout such as | |
506 | * -------------pfn--------------> | |
507 | * N0 | N1 | N2 | N0 | N1 | N2|.... | |
508 | */ | |
2f1ee091 | 509 | if (pfn_to_nid(pfn) != nid) |
eefa864b JK |
510 | continue; |
511 | if (init_section_page_ext(pfn, nid)) | |
512 | goto oom; | |
0fc542b7 | 513 | cond_resched(); |
eefa864b JK |
514 | } |
515 | } | |
516 | hotplug_memory_notifier(page_ext_callback, 0); | |
517 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
518 | invoke_init_callbacks(); | |
519 | return; | |
520 | ||
521 | oom: | |
522 | panic("Out of memory"); | |
523 | } | |
524 | ||
525 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
526 | { | |
527 | } | |
528 | ||
529 | #endif |