1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/mm.h>
3 #include <linux/mmzone.h>
4 #include <linux/memblock.h>
5 #include <linux/page_ext.h>
6 #include <linux/memory.h>
7 #include <linux/vmalloc.h>
8 #include <linux/kmemleak.h>
9 #include <linux/page_owner.h>
10 #include <linux/page_idle.h>
11 #include <linux/page_table_check.h>
12 #include <linux/rcupdate.h>
13 #include <linux/pgalloc_tag.h>
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
40 * boot and extra memory would be unnecessary. In this case, to avoid
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 *
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 *
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
64 #ifdef CONFIG_SPARSEMEM
65 #define PAGE_EXT_INVALID (0x1)
66 #endif
67
68 #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
69 static bool need_page_idle(void)
70 {
71 return true;
72 }
73 static struct page_ext_operations page_idle_ops __initdata = {
74 .need = need_page_idle,
75 .need_shared_flags = true,
76 };
77 #endif
78
79 static struct page_ext_operations *page_ext_ops[] __initdata = {
80 #ifdef CONFIG_PAGE_OWNER
81 &page_owner_ops,
82 #endif
83 #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
84 &page_idle_ops,
85 #endif
86 #ifdef CONFIG_MEM_ALLOC_PROFILING
87 &page_alloc_tagging_ops,
88 #endif
89 #ifdef CONFIG_PAGE_TABLE_CHECK
90 &page_table_check_ops,
91 #endif
92 };
93
94 unsigned long page_ext_size;
95
96 static unsigned long total_usage;
97
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
106 bool early_page_ext __meminitdata;
107 #endif
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
115 static bool __init invoke_need_callbacks(void)
116 {
117 int i;
118 int entries = ARRAY_SIZE(page_ext_ops);
119 bool need = false;
120
121 for (i = 0; i < entries; i++) {
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()) {
132 page_ext_ops[i]->offset = page_ext_size;
133 page_ext_size += page_ext_ops[i]->size;
134 need = true;
135 }
136 }
137
138 return need;
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
152 static inline struct page_ext *get_entry(void *base, unsigned long index)
153 {
154 return base + page_ext_size * index;
155 }
156
157 #ifndef CONFIG_SPARSEMEM
158 void __init page_ext_init_flatmem_late(void)
159 {
160 invoke_init_callbacks();
161 }
162
163 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
164 {
165 pgdat->node_page_ext = NULL;
166 }
167
168 static struct page_ext *lookup_page_ext(const struct page *page)
169 {
170 unsigned long pfn = page_to_pfn(page);
171 unsigned long index;
172 struct page_ext *base;
173
174 WARN_ON_ONCE(!rcu_read_lock_held());
175 base = NODE_DATA(page_to_nid(page))->node_page_ext;
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;
184 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
185 MAX_ORDER_NR_PAGES);
186 return get_entry(base, index);
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
208 table_size = page_ext_size * nr_pages;
209
210 base = memblock_alloc_try_nid(
211 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
212 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
213 if (!base)
214 return -ENOMEM;
215 NODE_DATA(nid)->node_page_ext = base;
216 total_usage += table_size;
217 mod_node_page_state(NODE_DATA(nid), NR_MEMMAP_BOOT,
218 DIV_ROUND_UP(table_size, PAGE_SIZE));
219 return 0;
220 }
221
222 void __init page_ext_init_flatmem(void)
223 {
224
225 int nid, fail;
226
227 if (!invoke_need_callbacks())
228 return;
229
230 for_each_online_node(nid) {
231 fail = alloc_node_page_ext(nid);
232 if (fail)
233 goto fail;
234 }
235 pr_info("allocated %ld bytes of page_ext\n", total_usage);
236 return;
237
238 fail:
239 pr_crit("allocation of page_ext failed.\n");
240 panic("Out of memory");
241 }
242
243 #else /* CONFIG_SPARSEMEM */
244 static bool page_ext_invalid(struct page_ext *page_ext)
245 {
246 return !page_ext || (((unsigned long)page_ext & PAGE_EXT_INVALID) == PAGE_EXT_INVALID);
247 }
248
249 static struct page_ext *lookup_page_ext(const struct page *page)
250 {
251 unsigned long pfn = page_to_pfn(page);
252 struct mem_section *section = __pfn_to_section(pfn);
253 struct page_ext *page_ext = READ_ONCE(section->page_ext);
254
255 WARN_ON_ONCE(!rcu_read_lock_held());
256 /*
257 * The sanity checks the page allocator does upon freeing a
258 * page can reach here before the page_ext arrays are
259 * allocated when feeding a range of pages to the allocator
260 * for the first time during bootup or memory hotplug.
261 */
262 if (page_ext_invalid(page_ext))
263 return NULL;
264 return get_entry(page_ext, pfn);
265 }
266
267 static void *__meminit alloc_page_ext(size_t size, int nid)
268 {
269 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
270 void *addr = NULL;
271
272 addr = alloc_pages_exact_nid(nid, size, flags);
273 if (addr)
274 kmemleak_alloc(addr, size, 1, flags);
275 else
276 addr = vzalloc_node(size, nid);
277
278 if (addr) {
279 mod_node_page_state(NODE_DATA(nid), NR_MEMMAP,
280 DIV_ROUND_UP(size, PAGE_SIZE));
281 }
282
283 return addr;
284 }
285
286 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
287 {
288 struct mem_section *section;
289 struct page_ext *base;
290 unsigned long table_size;
291
292 section = __pfn_to_section(pfn);
293
294 if (section->page_ext)
295 return 0;
296
297 table_size = page_ext_size * PAGES_PER_SECTION;
298 base = alloc_page_ext(table_size, nid);
299
300 /*
301 * The value stored in section->page_ext is (base - pfn)
302 * and it does not point to the memory block allocated above,
303 * causing kmemleak false positives.
304 */
305 kmemleak_not_leak(base);
306
307 if (!base) {
308 pr_err("page ext allocation failure\n");
309 return -ENOMEM;
310 }
311
312 /*
313 * The passed "pfn" may not be aligned to SECTION. For the calculation
314 * we need to apply a mask.
315 */
316 pfn &= PAGE_SECTION_MASK;
317 section->page_ext = (void *)base - page_ext_size * pfn;
318 total_usage += table_size;
319 return 0;
320 }
321
322 static void free_page_ext(void *addr)
323 {
324 size_t table_size;
325 struct page *page;
326 struct pglist_data *pgdat;
327
328 table_size = page_ext_size * PAGES_PER_SECTION;
329
330 if (is_vmalloc_addr(addr)) {
331 page = vmalloc_to_page(addr);
332 pgdat = page_pgdat(page);
333 vfree(addr);
334 } else {
335 page = virt_to_page(addr);
336 pgdat = page_pgdat(page);
337 BUG_ON(PageReserved(page));
338 kmemleak_free(addr);
339 free_pages_exact(addr, table_size);
340 }
341
342 mod_node_page_state(pgdat, NR_MEMMAP,
343 -1L * (DIV_ROUND_UP(table_size, PAGE_SIZE)));
344
345 }
346
347 static void __free_page_ext(unsigned long pfn)
348 {
349 struct mem_section *ms;
350 struct page_ext *base;
351
352 ms = __pfn_to_section(pfn);
353 if (!ms || !ms->page_ext)
354 return;
355
356 base = READ_ONCE(ms->page_ext);
357 /*
358 * page_ext here can be valid while doing the roll back
359 * operation in online_page_ext().
360 */
361 if (page_ext_invalid(base))
362 base = (void *)base - PAGE_EXT_INVALID;
363 WRITE_ONCE(ms->page_ext, NULL);
364
365 base = get_entry(base, pfn);
366 free_page_ext(base);
367 }
368
369 static void __invalidate_page_ext(unsigned long pfn)
370 {
371 struct mem_section *ms;
372 void *val;
373
374 ms = __pfn_to_section(pfn);
375 if (!ms || !ms->page_ext)
376 return;
377 val = (void *)ms->page_ext + PAGE_EXT_INVALID;
378 WRITE_ONCE(ms->page_ext, val);
379 }
380
381 static int __meminit online_page_ext(unsigned long start_pfn,
382 unsigned long nr_pages,
383 int nid)
384 {
385 unsigned long start, end, pfn;
386 int fail = 0;
387
388 start = SECTION_ALIGN_DOWN(start_pfn);
389 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
390
391 if (nid == NUMA_NO_NODE) {
392 /*
393 * In this case, "nid" already exists and contains valid memory.
394 * "start_pfn" passed to us is a pfn which is an arg for
395 * online__pages(), and start_pfn should exist.
396 */
397 nid = pfn_to_nid(start_pfn);
398 VM_BUG_ON(!node_online(nid));
399 }
400
401 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION)
402 fail = init_section_page_ext(pfn, nid);
403 if (!fail)
404 return 0;
405
406 /* rollback */
407 end = pfn - PAGES_PER_SECTION;
408 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
409 __free_page_ext(pfn);
410
411 return -ENOMEM;
412 }
413
414 static void __meminit offline_page_ext(unsigned long start_pfn,
415 unsigned long nr_pages)
416 {
417 unsigned long start, end, pfn;
418
419 start = SECTION_ALIGN_DOWN(start_pfn);
420 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
421
422 /*
423 * Freeing of page_ext is done in 3 steps to avoid
424 * use-after-free of it:
425 * 1) Traverse all the sections and mark their page_ext
426 * as invalid.
427 * 2) Wait for all the existing users of page_ext who
428 * started before invalidation to finish.
429 * 3) Free the page_ext.
430 */
431 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
432 __invalidate_page_ext(pfn);
433
434 synchronize_rcu();
435
436 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
437 __free_page_ext(pfn);
438 }
439
440 static int __meminit page_ext_callback(struct notifier_block *self,
441 unsigned long action, void *arg)
442 {
443 struct memory_notify *mn = arg;
444 int ret = 0;
445
446 switch (action) {
447 case MEM_GOING_ONLINE:
448 ret = online_page_ext(mn->start_pfn,
449 mn->nr_pages, mn->status_change_nid);
450 break;
451 case MEM_OFFLINE:
452 offline_page_ext(mn->start_pfn,
453 mn->nr_pages);
454 break;
455 case MEM_CANCEL_ONLINE:
456 offline_page_ext(mn->start_pfn,
457 mn->nr_pages);
458 break;
459 case MEM_GOING_OFFLINE:
460 break;
461 case MEM_ONLINE:
462 case MEM_CANCEL_OFFLINE:
463 break;
464 }
465
466 return notifier_from_errno(ret);
467 }
468
469 void __init page_ext_init(void)
470 {
471 unsigned long pfn;
472 int nid;
473
474 if (!invoke_need_callbacks())
475 return;
476
477 for_each_node_state(nid, N_MEMORY) {
478 unsigned long start_pfn, end_pfn;
479
480 start_pfn = node_start_pfn(nid);
481 end_pfn = node_end_pfn(nid);
482 /*
483 * start_pfn and end_pfn may not be aligned to SECTION and the
484 * page->flags of out of node pages are not initialized. So we
485 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
486 */
487 for (pfn = start_pfn; pfn < end_pfn;
488 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
489
490 if (!pfn_valid(pfn))
491 continue;
492 /*
493 * Nodes's pfns can be overlapping.
494 * We know some arch can have a nodes layout such as
495 * -------------pfn-------------->
496 * N0 | N1 | N2 | N0 | N1 | N2|....
497 */
498 if (pfn_to_nid(pfn) != nid)
499 continue;
500 if (init_section_page_ext(pfn, nid))
501 goto oom;
502 cond_resched();
503 }
504 }
505 hotplug_memory_notifier(page_ext_callback, DEFAULT_CALLBACK_PRI);
506 pr_info("allocated %ld bytes of page_ext\n", total_usage);
507 invoke_init_callbacks();
508 return;
509
510 oom:
511 panic("Out of memory");
512 }
513
514 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
515 {
516 }
517
518 #endif
519
520 /**
521 * page_ext_get() - Get the extended information for a page.
522 * @page: The page we're interested in.
523 *
524 * Ensures that the page_ext will remain valid until page_ext_put()
525 * is called.
526 *
527 * Return: NULL if no page_ext exists for this page.
528 * Context: Any context. Caller may not sleep until they have called
529 * page_ext_put().
530 */
531 struct page_ext *page_ext_get(const struct page *page)
532 {
533 struct page_ext *page_ext;
534
535 rcu_read_lock();
536 page_ext = lookup_page_ext(page);
537 if (!page_ext) {
538 rcu_read_unlock();
539 return NULL;
540 }
541
542 return page_ext;
543 }
544
545 /**
546 * page_ext_put() - Working with page extended information is done.
547 * @page_ext: Page extended information received from page_ext_get().
548 *
549 * The page extended information of the page may not be valid after this
550 * function is called.
551 *
552 * Return: None.
553 * Context: Any context with corresponding page_ext_get() is called.
554 */
555 void page_ext_put(struct page_ext *page_ext)
556 {
557 if (unlikely(!page_ext))
558 return;
559
560 rcu_read_unlock();
561 }
562
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.