blob: a7be1c7a79f66d4627b8429d82314afaa4b24b9c [file] [log] [blame]
Joonsoo Kimeefa8642014-12-12 16:55:46 -08001#include <linux/mm.h>
2#include <linux/mmzone.h>
3#include <linux/bootmem.h>
4#include <linux/page_ext.h>
5#include <linux/memory.h>
6#include <linux/vmalloc.h>
7#include <linux/kmemleak.h>
Joonsoo Kim48c96a32014-12-12 16:56:01 -08008#include <linux/page_owner.h>
Vladimir Davydov33c3fc72015-09-09 15:35:45 -07009#include <linux/page_idle.h>
Joonsoo Kimeefa8642014-12-12 16:55:46 -080010
11/*
12 * struct page extension
13 *
14 * This is the feature to manage memory for extended data per page.
15 *
16 * Until now, we must modify struct page itself to store extra data per page.
17 * This requires rebuilding the kernel and it is really time consuming process.
18 * And, sometimes, rebuild is impossible due to third party module dependency.
19 * At last, enlarging struct page could cause un-wanted system behaviour change.
20 *
21 * This feature is intended to overcome above mentioned problems. This feature
22 * allocates memory for extended data per page in certain place rather than
23 * the struct page itself. This memory can be accessed by the accessor
24 * functions provided by this code. During the boot process, it checks whether
25 * allocation of huge chunk of memory is needed or not. If not, it avoids
26 * allocating memory at all. With this advantage, we can include this feature
27 * into the kernel in default and can avoid rebuild and solve related problems.
28 *
29 * To help these things to work well, there are two callbacks for clients. One
30 * is the need callback which is mandatory if user wants to avoid useless
31 * memory allocation at boot-time. The other is optional, init callback, which
32 * is used to do proper initialization after memory is allocated.
33 *
34 * The need callback is used to decide whether extended memory allocation is
35 * needed or not. Sometimes users want to deactivate some features in this
36 * boot and extra memory would be unneccessary. In this case, to avoid
37 * allocating huge chunk of memory, each clients represent their need of
38 * extra memory through the need callback. If one of the need callbacks
39 * returns true, it means that someone needs extra memory so that
40 * page extension core should allocates memory for page extension. If
41 * none of need callbacks return true, memory isn't needed at all in this boot
42 * and page extension core can skip to allocate memory. As result,
43 * none of memory is wasted.
44 *
Joonsoo Kim980ac162016-10-07 16:58:27 -070045 * When need callback returns true, page_ext checks if there is a request for
46 * extra memory through size in struct page_ext_operations. If it is non-zero,
47 * extra space is allocated for each page_ext entry and offset is returned to
48 * user through offset in struct page_ext_operations.
49 *
Joonsoo Kimeefa8642014-12-12 16:55:46 -080050 * The init callback is used to do proper initialization after page extension
51 * is completely initialized. In sparse memory system, extra memory is
52 * allocated some time later than memmap is allocated. In other words, lifetime
53 * of memory for page extension isn't same with memmap for struct page.
54 * Therefore, clients can't store extra data until page extension is
55 * initialized, even if pages are allocated and used freely. This could
56 * cause inadequate state of extra data per page, so, to prevent it, client
57 * can utilize this callback to initialize the state of it correctly.
58 */
59
60static struct page_ext_operations *page_ext_ops[] = {
Joonsoo Kime30825f2014-12-12 16:55:49 -080061 &debug_guardpage_ops,
62#ifdef CONFIG_PAGE_POISONING
63 &page_poisoning_ops,
64#endif
Joonsoo Kim48c96a32014-12-12 16:56:01 -080065#ifdef CONFIG_PAGE_OWNER
66 &page_owner_ops,
67#endif
Vladimir Davydov33c3fc72015-09-09 15:35:45 -070068#if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
69 &page_idle_ops,
70#endif
Joonsoo Kimeefa8642014-12-12 16:55:46 -080071};
72
73static unsigned long total_usage;
Joonsoo Kim980ac162016-10-07 16:58:27 -070074static unsigned long extra_mem;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080075
76static bool __init invoke_need_callbacks(void)
77{
78 int i;
79 int entries = ARRAY_SIZE(page_ext_ops);
Joonsoo Kim980ac162016-10-07 16:58:27 -070080 bool need = false;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080081
82 for (i = 0; i < entries; i++) {
Joonsoo Kim980ac162016-10-07 16:58:27 -070083 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
84 page_ext_ops[i]->offset = sizeof(struct page_ext) +
85 extra_mem;
86 extra_mem += page_ext_ops[i]->size;
87 need = true;
88 }
Joonsoo Kimeefa8642014-12-12 16:55:46 -080089 }
90
Joonsoo Kim980ac162016-10-07 16:58:27 -070091 return need;
Joonsoo Kimeefa8642014-12-12 16:55:46 -080092}
93
94static void __init invoke_init_callbacks(void)
95{
96 int i;
97 int entries = ARRAY_SIZE(page_ext_ops);
98
99 for (i = 0; i < entries; i++) {
100 if (page_ext_ops[i]->init)
101 page_ext_ops[i]->init();
102 }
103}
104
Joonsoo Kim980ac162016-10-07 16:58:27 -0700105static unsigned long get_entry_size(void)
106{
107 return sizeof(struct page_ext) + extra_mem;
108}
109
110static inline struct page_ext *get_entry(void *base, unsigned long index)
111{
112 return base + get_entry_size() * index;
113}
114
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800115#if !defined(CONFIG_SPARSEMEM)
116
117
118void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
119{
120 pgdat->node_page_ext = NULL;
121}
122
123struct page_ext *lookup_page_ext(struct page *page)
124{
125 unsigned long pfn = page_to_pfn(page);
Joonsoo Kim0b06bb32016-10-07 16:58:24 -0700126 unsigned long index;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800127 struct page_ext *base;
128
129 base = NODE_DATA(page_to_nid(page))->node_page_ext;
Laura Abbott1414c7f2016-03-15 14:56:30 -0700130#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800131 /*
132 * The sanity checks the page allocator does upon freeing a
133 * page can reach here before the page_ext arrays are
134 * allocated when feeding a range of pages to the allocator
135 * for the first time during bootup or memory hotplug.
Laura Abbott1414c7f2016-03-15 14:56:30 -0700136 *
137 * This check is also necessary for ensuring page poisoning
138 * works as expected when enabled
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800139 */
140 if (unlikely(!base))
141 return NULL;
142#endif
Joonsoo Kim0b06bb32016-10-07 16:58:24 -0700143 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800144 MAX_ORDER_NR_PAGES);
Joonsoo Kim980ac162016-10-07 16:58:27 -0700145 return get_entry(base, index);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800146}
147
148static int __init alloc_node_page_ext(int nid)
149{
150 struct page_ext *base;
151 unsigned long table_size;
152 unsigned long nr_pages;
153
154 nr_pages = NODE_DATA(nid)->node_spanned_pages;
155 if (!nr_pages)
156 return 0;
157
158 /*
159 * Need extra space if node range is not aligned with
160 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
161 * checks buddy's status, range could be out of exact node range.
162 */
163 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
164 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
165 nr_pages += MAX_ORDER_NR_PAGES;
166
Joonsoo Kim980ac162016-10-07 16:58:27 -0700167 table_size = get_entry_size() * nr_pages;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800168
169 base = memblock_virt_alloc_try_nid_nopanic(
170 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
171 BOOTMEM_ALLOC_ACCESSIBLE, nid);
172 if (!base)
173 return -ENOMEM;
174 NODE_DATA(nid)->node_page_ext = base;
175 total_usage += table_size;
176 return 0;
177}
178
179void __init page_ext_init_flatmem(void)
180{
181
182 int nid, fail;
183
184 if (!invoke_need_callbacks())
185 return;
186
187 for_each_online_node(nid) {
188 fail = alloc_node_page_ext(nid);
189 if (fail)
190 goto fail;
191 }
192 pr_info("allocated %ld bytes of page_ext\n", total_usage);
193 invoke_init_callbacks();
194 return;
195
196fail:
197 pr_crit("allocation of page_ext failed.\n");
198 panic("Out of memory");
199}
200
201#else /* CONFIG_FLAT_NODE_MEM_MAP */
202
203struct page_ext *lookup_page_ext(struct page *page)
204{
205 unsigned long pfn = page_to_pfn(page);
206 struct mem_section *section = __pfn_to_section(pfn);
Laura Abbott1414c7f2016-03-15 14:56:30 -0700207#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800208 /*
209 * The sanity checks the page allocator does upon freeing a
210 * page can reach here before the page_ext arrays are
211 * allocated when feeding a range of pages to the allocator
212 * for the first time during bootup or memory hotplug.
Laura Abbott1414c7f2016-03-15 14:56:30 -0700213 *
214 * This check is also necessary for ensuring page poisoning
215 * works as expected when enabled
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800216 */
217 if (!section->page_ext)
218 return NULL;
219#endif
Joonsoo Kim980ac162016-10-07 16:58:27 -0700220 return get_entry(section->page_ext, pfn);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800221}
222
223static void *__meminit alloc_page_ext(size_t size, int nid)
224{
225 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
226 void *addr = NULL;
227
228 addr = alloc_pages_exact_nid(nid, size, flags);
229 if (addr) {
230 kmemleak_alloc(addr, size, 1, flags);
231 return addr;
232 }
233
234 if (node_state(nid, N_HIGH_MEMORY))
235 addr = vzalloc_node(size, nid);
236 else
237 addr = vzalloc(size);
238
239 return addr;
240}
241
242static int __meminit init_section_page_ext(unsigned long pfn, int nid)
243{
244 struct mem_section *section;
245 struct page_ext *base;
246 unsigned long table_size;
247
248 section = __pfn_to_section(pfn);
249
250 if (section->page_ext)
251 return 0;
252
Joonsoo Kim980ac162016-10-07 16:58:27 -0700253 table_size = get_entry_size() * PAGES_PER_SECTION;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800254 base = alloc_page_ext(table_size, nid);
255
256 /*
257 * The value stored in section->page_ext is (base - pfn)
258 * and it does not point to the memory block allocated above,
259 * causing kmemleak false positives.
260 */
261 kmemleak_not_leak(base);
262
263 if (!base) {
264 pr_err("page ext allocation failure\n");
265 return -ENOMEM;
266 }
267
268 /*
269 * The passed "pfn" may not be aligned to SECTION. For the calculation
270 * we need to apply a mask.
271 */
272 pfn &= PAGE_SECTION_MASK;
Joonsoo Kim980ac162016-10-07 16:58:27 -0700273 section->page_ext = (void *)base - get_entry_size() * pfn;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800274 total_usage += table_size;
275 return 0;
276}
277#ifdef CONFIG_MEMORY_HOTPLUG
278static void free_page_ext(void *addr)
279{
280 if (is_vmalloc_addr(addr)) {
281 vfree(addr);
282 } else {
283 struct page *page = virt_to_page(addr);
284 size_t table_size;
285
Joonsoo Kim980ac162016-10-07 16:58:27 -0700286 table_size = get_entry_size() * PAGES_PER_SECTION;
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800287
288 BUG_ON(PageReserved(page));
Qian Cai7880fc292019-03-05 15:49:46 -0800289 kmemleak_free(addr);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800290 free_pages_exact(addr, table_size);
291 }
292}
293
294static void __free_page_ext(unsigned long pfn)
295{
296 struct mem_section *ms;
297 struct page_ext *base;
298
299 ms = __pfn_to_section(pfn);
300 if (!ms || !ms->page_ext)
301 return;
Joonsoo Kim980ac162016-10-07 16:58:27 -0700302 base = get_entry(ms->page_ext, pfn);
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800303 free_page_ext(base);
304 ms->page_ext = NULL;
305}
306
307static int __meminit online_page_ext(unsigned long start_pfn,
308 unsigned long nr_pages,
309 int nid)
310{
311 unsigned long start, end, pfn;
312 int fail = 0;
313
314 start = SECTION_ALIGN_DOWN(start_pfn);
315 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
316
317 if (nid == -1) {
318 /*
319 * In this case, "nid" already exists and contains valid memory.
320 * "start_pfn" passed to us is a pfn which is an arg for
321 * online__pages(), and start_pfn should exist.
322 */
323 nid = pfn_to_nid(start_pfn);
324 VM_BUG_ON(!node_state(nid, N_ONLINE));
325 }
326
327 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
328 if (!pfn_present(pfn))
329 continue;
330 fail = init_section_page_ext(pfn, nid);
331 }
332 if (!fail)
333 return 0;
334
335 /* rollback */
336 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
337 __free_page_ext(pfn);
338
339 return -ENOMEM;
340}
341
342static int __meminit offline_page_ext(unsigned long start_pfn,
343 unsigned long nr_pages, int nid)
344{
345 unsigned long start, end, pfn;
346
347 start = SECTION_ALIGN_DOWN(start_pfn);
348 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
349
350 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
351 __free_page_ext(pfn);
352 return 0;
353
354}
355
356static int __meminit page_ext_callback(struct notifier_block *self,
357 unsigned long action, void *arg)
358{
359 struct memory_notify *mn = arg;
360 int ret = 0;
361
362 switch (action) {
363 case MEM_GOING_ONLINE:
364 ret = online_page_ext(mn->start_pfn,
365 mn->nr_pages, mn->status_change_nid);
366 break;
367 case MEM_OFFLINE:
368 offline_page_ext(mn->start_pfn,
369 mn->nr_pages, mn->status_change_nid);
370 break;
371 case MEM_CANCEL_ONLINE:
372 offline_page_ext(mn->start_pfn,
373 mn->nr_pages, mn->status_change_nid);
374 break;
375 case MEM_GOING_OFFLINE:
376 break;
377 case MEM_ONLINE:
378 case MEM_CANCEL_OFFLINE:
379 break;
380 }
381
382 return notifier_from_errno(ret);
383}
384
385#endif
386
387void __init page_ext_init(void)
388{
389 unsigned long pfn;
390 int nid;
391
392 if (!invoke_need_callbacks())
393 return;
394
395 for_each_node_state(nid, N_MEMORY) {
396 unsigned long start_pfn, end_pfn;
397
398 start_pfn = node_start_pfn(nid);
399 end_pfn = node_end_pfn(nid);
400 /*
401 * start_pfn and end_pfn may not be aligned to SECTION and the
402 * page->flags of out of node pages are not initialized. So we
403 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
404 */
405 for (pfn = start_pfn; pfn < end_pfn;
406 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
407
408 if (!pfn_valid(pfn))
409 continue;
410 /*
411 * Nodes's pfns can be overlapping.
412 * We know some arch can have a nodes layout such as
413 * -------------pfn-------------->
414 * N0 | N1 | N2 | N0 | N1 | N2|....
Yang Shife53ca52016-05-27 14:27:30 -0700415 *
416 * Take into account DEFERRED_STRUCT_PAGE_INIT.
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800417 */
Yang Shife53ca52016-05-27 14:27:30 -0700418 if (early_pfn_to_nid(pfn) != nid)
Joonsoo Kimeefa8642014-12-12 16:55:46 -0800419 continue;
420 if (init_section_page_ext(pfn, nid))
421 goto oom;
422 }
423 }
424 hotplug_memory_notifier(page_ext_callback, 0);
425 pr_info("allocated %ld bytes of page_ext\n", total_usage);
426 invoke_init_callbacks();
427 return;
428
429oom:
430 panic("Out of memory");
431}
432
433void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
434{
435}
436
437#endif