xref: /linux/mm/page_ext.c (revision 8bc7c5e525584903ea83332e18a2118ed3b1985e)
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 	memmap_boot_pages_add(DIV_ROUND_UP(table_size, PAGE_SIZE));
218 	return 0;
219 }
220 
221 void __init page_ext_init_flatmem(void)
222 {
223 
224 	int nid, fail;
225 
226 	if (!invoke_need_callbacks())
227 		return;
228 
229 	for_each_online_node(nid)  {
230 		fail = alloc_node_page_ext(nid);
231 		if (fail)
232 			goto fail;
233 	}
234 	pr_info("allocated %ld bytes of page_ext\n", total_usage);
235 	return;
236 
237 fail:
238 	pr_crit("allocation of page_ext failed.\n");
239 	panic("Out of memory");
240 }
241 
242 #else /* CONFIG_SPARSEMEM */
243 static bool page_ext_invalid(struct page_ext *page_ext)
244 {
245 	return !page_ext || (((unsigned long)page_ext & PAGE_EXT_INVALID) == PAGE_EXT_INVALID);
246 }
247 
248 static struct page_ext *lookup_page_ext(const struct page *page)
249 {
250 	unsigned long pfn = page_to_pfn(page);
251 	struct mem_section *section = __pfn_to_section(pfn);
252 	struct page_ext *page_ext = READ_ONCE(section->page_ext);
253 
254 	WARN_ON_ONCE(!rcu_read_lock_held());
255 	/*
256 	 * The sanity checks the page allocator does upon freeing a
257 	 * page can reach here before the page_ext arrays are
258 	 * allocated when feeding a range of pages to the allocator
259 	 * for the first time during bootup or memory hotplug.
260 	 */
261 	if (page_ext_invalid(page_ext))
262 		return NULL;
263 	return get_entry(page_ext, pfn);
264 }
265 
266 static void *__meminit alloc_page_ext(size_t size, int nid)
267 {
268 	gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
269 	void *addr = NULL;
270 
271 	addr = alloc_pages_exact_nid(nid, size, flags);
272 	if (addr)
273 		kmemleak_alloc(addr, size, 1, flags);
274 	else
275 		addr = vzalloc_node(size, nid);
276 
277 	if (addr)
278 		memmap_pages_add(DIV_ROUND_UP(size, PAGE_SIZE));
279 
280 	return addr;
281 }
282 
283 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
284 {
285 	struct mem_section *section;
286 	struct page_ext *base;
287 	unsigned long table_size;
288 
289 	section = __pfn_to_section(pfn);
290 
291 	if (section->page_ext)
292 		return 0;
293 
294 	table_size = page_ext_size * PAGES_PER_SECTION;
295 	base = alloc_page_ext(table_size, nid);
296 
297 	/*
298 	 * The value stored in section->page_ext is (base - pfn)
299 	 * and it does not point to the memory block allocated above,
300 	 * causing kmemleak false positives.
301 	 */
302 	kmemleak_not_leak(base);
303 
304 	if (!base) {
305 		pr_err("page ext allocation failure\n");
306 		return -ENOMEM;
307 	}
308 
309 	/*
310 	 * The passed "pfn" may not be aligned to SECTION.  For the calculation
311 	 * we need to apply a mask.
312 	 */
313 	pfn &= PAGE_SECTION_MASK;
314 	section->page_ext = (void *)base - page_ext_size * pfn;
315 	total_usage += table_size;
316 	return 0;
317 }
318 
319 static void free_page_ext(void *addr)
320 {
321 	size_t table_size;
322 	struct page *page;
323 
324 	table_size = page_ext_size * PAGES_PER_SECTION;
325 	memmap_pages_add(-1L * (DIV_ROUND_UP(table_size, PAGE_SIZE)));
326 
327 	if (is_vmalloc_addr(addr)) {
328 		vfree(addr);
329 	} else {
330 		page = virt_to_page(addr);
331 		BUG_ON(PageReserved(page));
332 		kmemleak_free(addr);
333 		free_pages_exact(addr, table_size);
334 	}
335 }
336 
337 static void __free_page_ext(unsigned long pfn)
338 {
339 	struct mem_section *ms;
340 	struct page_ext *base;
341 
342 	ms = __pfn_to_section(pfn);
343 	if (!ms || !ms->page_ext)
344 		return;
345 
346 	base = READ_ONCE(ms->page_ext);
347 	/*
348 	 * page_ext here can be valid while doing the roll back
349 	 * operation in online_page_ext().
350 	 */
351 	if (page_ext_invalid(base))
352 		base = (void *)base - PAGE_EXT_INVALID;
353 	WRITE_ONCE(ms->page_ext, NULL);
354 
355 	base = get_entry(base, pfn);
356 	free_page_ext(base);
357 }
358 
359 static void __invalidate_page_ext(unsigned long pfn)
360 {
361 	struct mem_section *ms;
362 	void *val;
363 
364 	ms = __pfn_to_section(pfn);
365 	if (!ms || !ms->page_ext)
366 		return;
367 	val = (void *)ms->page_ext + PAGE_EXT_INVALID;
368 	WRITE_ONCE(ms->page_ext, val);
369 }
370 
371 static int __meminit online_page_ext(unsigned long start_pfn,
372 				unsigned long nr_pages,
373 				int nid)
374 {
375 	unsigned long start, end, pfn;
376 	int fail = 0;
377 
378 	start = SECTION_ALIGN_DOWN(start_pfn);
379 	end = SECTION_ALIGN_UP(start_pfn + nr_pages);
380 
381 	if (nid == NUMA_NO_NODE) {
382 		/*
383 		 * In this case, "nid" already exists and contains valid memory.
384 		 * "start_pfn" passed to us is a pfn which is an arg for
385 		 * online__pages(), and start_pfn should exist.
386 		 */
387 		nid = pfn_to_nid(start_pfn);
388 		VM_BUG_ON(!node_online(nid));
389 	}
390 
391 	for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION)
392 		fail = init_section_page_ext(pfn, nid);
393 	if (!fail)
394 		return 0;
395 
396 	/* rollback */
397 	end = pfn - PAGES_PER_SECTION;
398 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
399 		__free_page_ext(pfn);
400 
401 	return -ENOMEM;
402 }
403 
404 static void __meminit offline_page_ext(unsigned long start_pfn,
405 				unsigned long nr_pages)
406 {
407 	unsigned long start, end, pfn;
408 
409 	start = SECTION_ALIGN_DOWN(start_pfn);
410 	end = SECTION_ALIGN_UP(start_pfn + nr_pages);
411 
412 	/*
413 	 * Freeing of page_ext is done in 3 steps to avoid
414 	 * use-after-free of it:
415 	 * 1) Traverse all the sections and mark their page_ext
416 	 *    as invalid.
417 	 * 2) Wait for all the existing users of page_ext who
418 	 *    started before invalidation to finish.
419 	 * 3) Free the page_ext.
420 	 */
421 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
422 		__invalidate_page_ext(pfn);
423 
424 	synchronize_rcu();
425 
426 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
427 		__free_page_ext(pfn);
428 }
429 
430 static int __meminit page_ext_callback(struct notifier_block *self,
431 			       unsigned long action, void *arg)
432 {
433 	struct memory_notify *mn = arg;
434 	int ret = 0;
435 
436 	switch (action) {
437 	case MEM_GOING_ONLINE:
438 		ret = online_page_ext(mn->start_pfn,
439 				   mn->nr_pages, mn->status_change_nid);
440 		break;
441 	case MEM_OFFLINE:
442 		offline_page_ext(mn->start_pfn,
443 				mn->nr_pages);
444 		break;
445 	case MEM_CANCEL_ONLINE:
446 		offline_page_ext(mn->start_pfn,
447 				mn->nr_pages);
448 		break;
449 	case MEM_GOING_OFFLINE:
450 		break;
451 	case MEM_ONLINE:
452 	case MEM_CANCEL_OFFLINE:
453 		break;
454 	}
455 
456 	return notifier_from_errno(ret);
457 }
458 
459 void __init page_ext_init(void)
460 {
461 	unsigned long pfn;
462 	int nid;
463 
464 	if (!invoke_need_callbacks())
465 		return;
466 
467 	for_each_node_state(nid, N_MEMORY) {
468 		unsigned long start_pfn, end_pfn;
469 
470 		start_pfn = node_start_pfn(nid);
471 		end_pfn = node_end_pfn(nid);
472 		/*
473 		 * start_pfn and end_pfn may not be aligned to SECTION and the
474 		 * page->flags of out of node pages are not initialized.  So we
475 		 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
476 		 */
477 		for (pfn = start_pfn; pfn < end_pfn;
478 			pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
479 
480 			if (!pfn_valid(pfn))
481 				continue;
482 			/*
483 			 * Nodes's pfns can be overlapping.
484 			 * We know some arch can have a nodes layout such as
485 			 * -------------pfn-------------->
486 			 * N0 | N1 | N2 | N0 | N1 | N2|....
487 			 */
488 			if (pfn_to_nid(pfn) != nid)
489 				continue;
490 			if (init_section_page_ext(pfn, nid))
491 				goto oom;
492 			cond_resched();
493 		}
494 	}
495 	hotplug_memory_notifier(page_ext_callback, DEFAULT_CALLBACK_PRI);
496 	pr_info("allocated %ld bytes of page_ext\n", total_usage);
497 	invoke_init_callbacks();
498 	return;
499 
500 oom:
501 	panic("Out of memory");
502 }
503 
504 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
505 {
506 }
507 
508 #endif
509 
510 /**
511  * page_ext_get() - Get the extended information for a page.
512  * @page: The page we're interested in.
513  *
514  * Ensures that the page_ext will remain valid until page_ext_put()
515  * is called.
516  *
517  * Return: NULL if no page_ext exists for this page.
518  * Context: Any context.  Caller may not sleep until they have called
519  * page_ext_put().
520  */
521 struct page_ext *page_ext_get(const struct page *page)
522 {
523 	struct page_ext *page_ext;
524 
525 	rcu_read_lock();
526 	page_ext = lookup_page_ext(page);
527 	if (!page_ext) {
528 		rcu_read_unlock();
529 		return NULL;
530 	}
531 
532 	return page_ext;
533 }
534 
535 /**
536  * page_ext_put() - Working with page extended information is done.
537  * @page_ext: Page extended information received from page_ext_get().
538  *
539  * The page extended information of the page may not be valid after this
540  * function is called.
541  *
542  * Return: None.
543  * Context: Any context with corresponding page_ext_get() is called.
544  */
545 void page_ext_put(struct page_ext *page_ext)
546 {
547 	if (unlikely(!page_ext))
548 		return;
549 
550 	rcu_read_unlock();
551 }
552