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
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
39 * boot and extra memory would be unnecessary. In this case, to avoid
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 *
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 *
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
63#ifdef CONFIG_SPARSEMEM
64#define PAGE_EXT_INVALID (0x1)
65#endif
66
67#if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
68static bool need_page_idle(void)
69{
70 return true;
71}
72static struct page_ext_operations page_idle_ops __initdata = {
73 .need = need_page_idle,
74 .need_shared_flags = true,
75};
76#endif
77
78static struct page_ext_operations *page_ext_ops[] __initdata = {
79#ifdef CONFIG_PAGE_OWNER
80 &page_owner_ops,
81#endif
82#if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
83 &page_idle_ops,
84#endif
85#ifdef CONFIG_PAGE_TABLE_CHECK
86 &page_table_check_ops,
87#endif
88};
89
90unsigned long page_ext_size;
91
92static unsigned long total_usage;
93static struct page_ext *lookup_page_ext(const struct page *page);
94
95bool early_page_ext __meminitdata;
96static int __init setup_early_page_ext(char *str)
97{
98 early_page_ext = true;
99 return 0;
100}
101early_param("early_page_ext", setup_early_page_ext);
102
103static bool __init invoke_need_callbacks(void)
104{
105 int i;
106 int entries = ARRAY_SIZE(page_ext_ops);
107 bool need = false;
108
109 for (i = 0; i < entries; i++) {
110 if (page_ext_ops[i]->need()) {
111 if (page_ext_ops[i]->need_shared_flags) {
112 page_ext_size = sizeof(struct page_ext);
113 break;
114 }
115 }
116 }
117
118 for (i = 0; i < entries; i++) {
119 if (page_ext_ops[i]->need()) {
120 page_ext_ops[i]->offset = page_ext_size;
121 page_ext_size += page_ext_ops[i]->size;
122 need = true;
123 }
124 }
125
126 return need;
127}
128
129static void __init invoke_init_callbacks(void)
130{
131 int i;
132 int entries = ARRAY_SIZE(page_ext_ops);
133
134 for (i = 0; i < entries; i++) {
135 if (page_ext_ops[i]->init)
136 page_ext_ops[i]->init();
137 }
138}
139
140#ifndef CONFIG_SPARSEMEM
141void __init page_ext_init_flatmem_late(void)
142{
143 invoke_init_callbacks();
144}
145#endif
146
147static inline struct page_ext *get_entry(void *base, unsigned long index)
148{
149 return base + page_ext_size * index;
150}
151
152/**
153 * page_ext_get() - Get the extended information for a page.
154 * @page: The page we're interested in.
155 *
156 * Ensures that the page_ext will remain valid until page_ext_put()
157 * is called.
158 *
159 * Return: NULL if no page_ext exists for this page.
160 * Context: Any context. Caller may not sleep until they have called
161 * page_ext_put().
162 */
163struct page_ext *page_ext_get(struct page *page)
164{
165 struct page_ext *page_ext;
166
167 rcu_read_lock();
168 page_ext = lookup_page_ext(page);
169 if (!page_ext) {
170 rcu_read_unlock();
171 return NULL;
172 }
173
174 return page_ext;
175}
176
177/**
178 * page_ext_put() - Working with page extended information is done.
179 * @page_ext: Page extended information received from page_ext_get().
180 *
181 * The page extended information of the page may not be valid after this
182 * function is called.
183 *
184 * Return: None.
185 * Context: Any context with corresponding page_ext_get() is called.
186 */
187void page_ext_put(struct page_ext *page_ext)
188{
189 if (unlikely(!page_ext))
190 return;
191
192 rcu_read_unlock();
193}
194#ifndef CONFIG_SPARSEMEM
195
196
197void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
198{
199 pgdat->node_page_ext = NULL;
200}
201
202static struct page_ext *lookup_page_ext(const struct page *page)
203{
204 unsigned long pfn = page_to_pfn(page);
205 unsigned long index;
206 struct page_ext *base;
207
208 WARN_ON_ONCE(!rcu_read_lock_held());
209 base = NODE_DATA(nid: page_to_nid(page))->node_page_ext;
210 /*
211 * The sanity checks the page allocator does upon freeing a
212 * page can reach here before the page_ext arrays are
213 * allocated when feeding a range of pages to the allocator
214 * for the first time during bootup or memory hotplug.
215 */
216 if (unlikely(!base))
217 return NULL;
218 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
219 MAX_ORDER_NR_PAGES);
220 return get_entry(base, index);
221}
222
223static int __init alloc_node_page_ext(int nid)
224{
225 struct page_ext *base;
226 unsigned long table_size;
227 unsigned long nr_pages;
228
229 nr_pages = NODE_DATA(nid)->node_spanned_pages;
230 if (!nr_pages)
231 return 0;
232
233 /*
234 * Need extra space if node range is not aligned with
235 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
236 * checks buddy's status, range could be out of exact node range.
237 */
238 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
239 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
240 nr_pages += MAX_ORDER_NR_PAGES;
241
242 table_size = page_ext_size * nr_pages;
243
244 base = memblock_alloc_try_nid(
245 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
246 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
247 if (!base)
248 return -ENOMEM;
249 NODE_DATA(nid)->node_page_ext = base;
250 total_usage += table_size;
251 return 0;
252}
253
254void __init page_ext_init_flatmem(void)
255{
256
257 int nid, fail;
258
259 if (!invoke_need_callbacks())
260 return;
261
262 for_each_online_node(nid) {
263 fail = alloc_node_page_ext(nid);
264 if (fail)
265 goto fail;
266 }
267 pr_info("allocated %ld bytes of page_ext\n", total_usage);
268 return;
269
270fail:
271 pr_crit("allocation of page_ext failed.\n");
272 panic(fmt: "Out of memory");
273}
274
275#else /* CONFIG_SPARSEMEM */
276static bool page_ext_invalid(struct page_ext *page_ext)
277{
278 return !page_ext || (((unsigned long)page_ext & PAGE_EXT_INVALID) == PAGE_EXT_INVALID);
279}
280
281static struct page_ext *lookup_page_ext(const struct page *page)
282{
283 unsigned long pfn = page_to_pfn(page);
284 struct mem_section *section = __pfn_to_section(pfn);
285 struct page_ext *page_ext = READ_ONCE(section->page_ext);
286
287 WARN_ON_ONCE(!rcu_read_lock_held());
288 /*
289 * The sanity checks the page allocator does upon freeing a
290 * page can reach here before the page_ext arrays are
291 * allocated when feeding a range of pages to the allocator
292 * for the first time during bootup or memory hotplug.
293 */
294 if (page_ext_invalid(page_ext))
295 return NULL;
296 return get_entry(page_ext, pfn);
297}
298
299static void *__meminit alloc_page_ext(size_t size, int nid)
300{
301 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
302 void *addr = NULL;
303
304 addr = alloc_pages_exact_nid(nid, size, flags);
305 if (addr) {
306 kmemleak_alloc(addr, size, 1, flags);
307 return addr;
308 }
309
310 addr = vzalloc_node(size, nid);
311
312 return addr;
313}
314
315static int __meminit init_section_page_ext(unsigned long pfn, int nid)
316{
317 struct mem_section *section;
318 struct page_ext *base;
319 unsigned long table_size;
320
321 section = __pfn_to_section(pfn);
322
323 if (section->page_ext)
324 return 0;
325
326 table_size = page_ext_size * PAGES_PER_SECTION;
327 base = alloc_page_ext(table_size, nid);
328
329 /*
330 * The value stored in section->page_ext is (base - pfn)
331 * and it does not point to the memory block allocated above,
332 * causing kmemleak false positives.
333 */
334 kmemleak_not_leak(base);
335
336 if (!base) {
337 pr_err("page ext allocation failure\n");
338 return -ENOMEM;
339 }
340
341 /*
342 * The passed "pfn" may not be aligned to SECTION. For the calculation
343 * we need to apply a mask.
344 */
345 pfn &= PAGE_SECTION_MASK;
346 section->page_ext = (void *)base - page_ext_size * pfn;
347 total_usage += table_size;
348 return 0;
349}
350
351static void free_page_ext(void *addr)
352{
353 if (is_vmalloc_addr(addr)) {
354 vfree(addr);
355 } else {
356 struct page *page = virt_to_page(addr);
357 size_t table_size;
358
359 table_size = page_ext_size * PAGES_PER_SECTION;
360
361 BUG_ON(PageReserved(page));
362 kmemleak_free(addr);
363 free_pages_exact(addr, table_size);
364 }
365}
366
367static void __free_page_ext(unsigned long pfn)
368{
369 struct mem_section *ms;
370 struct page_ext *base;
371
372 ms = __pfn_to_section(pfn);
373 if (!ms || !ms->page_ext)
374 return;
375
376 base = READ_ONCE(ms->page_ext);
377 /*
378 * page_ext here can be valid while doing the roll back
379 * operation in online_page_ext().
380 */
381 if (page_ext_invalid(base))
382 base = (void *)base - PAGE_EXT_INVALID;
383 WRITE_ONCE(ms->page_ext, NULL);
384
385 base = get_entry(base, pfn);
386 free_page_ext(base);
387}
388
389static void __invalidate_page_ext(unsigned long pfn)
390{
391 struct mem_section *ms;
392 void *val;
393
394 ms = __pfn_to_section(pfn);
395 if (!ms || !ms->page_ext)
396 return;
397 val = (void *)ms->page_ext + PAGE_EXT_INVALID;
398 WRITE_ONCE(ms->page_ext, val);
399}
400
401static int __meminit online_page_ext(unsigned long start_pfn,
402 unsigned long nr_pages,
403 int nid)
404{
405 unsigned long start, end, pfn;
406 int fail = 0;
407
408 start = SECTION_ALIGN_DOWN(start_pfn);
409 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
410
411 if (nid == NUMA_NO_NODE) {
412 /*
413 * In this case, "nid" already exists and contains valid memory.
414 * "start_pfn" passed to us is a pfn which is an arg for
415 * online__pages(), and start_pfn should exist.
416 */
417 nid = pfn_to_nid(start_pfn);
418 VM_BUG_ON(!node_online(nid));
419 }
420
421 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION)
422 fail = init_section_page_ext(pfn, nid);
423 if (!fail)
424 return 0;
425
426 /* rollback */
427 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
428 __free_page_ext(pfn);
429
430 return -ENOMEM;
431}
432
433static int __meminit offline_page_ext(unsigned long start_pfn,
434 unsigned long nr_pages)
435{
436 unsigned long start, end, pfn;
437
438 start = SECTION_ALIGN_DOWN(start_pfn);
439 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
440
441 /*
442 * Freeing of page_ext is done in 3 steps to avoid
443 * use-after-free of it:
444 * 1) Traverse all the sections and mark their page_ext
445 * as invalid.
446 * 2) Wait for all the existing users of page_ext who
447 * started before invalidation to finish.
448 * 3) Free the page_ext.
449 */
450 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
451 __invalidate_page_ext(pfn);
452
453 synchronize_rcu();
454
455 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
456 __free_page_ext(pfn);
457 return 0;
458
459}
460
461static int __meminit page_ext_callback(struct notifier_block *self,
462 unsigned long action, void *arg)
463{
464 struct memory_notify *mn = arg;
465 int ret = 0;
466
467 switch (action) {
468 case MEM_GOING_ONLINE:
469 ret = online_page_ext(mn->start_pfn,
470 mn->nr_pages, mn->status_change_nid);
471 break;
472 case MEM_OFFLINE:
473 offline_page_ext(mn->start_pfn,
474 mn->nr_pages);
475 break;
476 case MEM_CANCEL_ONLINE:
477 offline_page_ext(mn->start_pfn,
478 mn->nr_pages);
479 break;
480 case MEM_GOING_OFFLINE:
481 break;
482 case MEM_ONLINE:
483 case MEM_CANCEL_OFFLINE:
484 break;
485 }
486
487 return notifier_from_errno(ret);
488}
489
490void __init page_ext_init(void)
491{
492 unsigned long pfn;
493 int nid;
494
495 if (!invoke_need_callbacks())
496 return;
497
498 for_each_node_state(nid, N_MEMORY) {
499 unsigned long start_pfn, end_pfn;
500
501 start_pfn = node_start_pfn(nid);
502 end_pfn = node_end_pfn(nid);
503 /*
504 * start_pfn and end_pfn may not be aligned to SECTION and the
505 * page->flags of out of node pages are not initialized. So we
506 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
507 */
508 for (pfn = start_pfn; pfn < end_pfn;
509 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
510
511 if (!pfn_valid(pfn))
512 continue;
513 /*
514 * Nodes's pfns can be overlapping.
515 * We know some arch can have a nodes layout such as
516 * -------------pfn-------------->
517 * N0 | N1 | N2 | N0 | N1 | N2|....
518 */
519 if (pfn_to_nid(pfn) != nid)
520 continue;
521 if (init_section_page_ext(pfn, nid))
522 goto oom;
523 cond_resched();
524 }
525 }
526 hotplug_memory_notifier(page_ext_callback, DEFAULT_CALLBACK_PRI);
527 pr_info("allocated %ld bytes of page_ext\n", total_usage);
528 invoke_init_callbacks();
529 return;
530
531oom:
532 panic("Out of memory");
533}
534
535void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
536{
537}
538
539#endif
540

source code of linux/mm/page_ext.c