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> |
eefa864b JK |
11 | |
12 | /* | |
13 | * struct page extension | |
14 | * | |
15 | * This is the feature to manage memory for extended data per page. | |
16 | * | |
17 | * Until now, we must modify struct page itself to store extra data per page. | |
18 | * This requires rebuilding the kernel and it is really time consuming process. | |
19 | * And, sometimes, rebuild is impossible due to third party module dependency. | |
20 | * At last, enlarging struct page could cause un-wanted system behaviour change. | |
21 | * | |
22 | * This feature is intended to overcome above mentioned problems. This feature | |
23 | * allocates memory for extended data per page in certain place rather than | |
24 | * the struct page itself. This memory can be accessed by the accessor | |
25 | * functions provided by this code. During the boot process, it checks whether | |
26 | * allocation of huge chunk of memory is needed or not. If not, it avoids | |
27 | * allocating memory at all. With this advantage, we can include this feature | |
28 | * into the kernel in default and can avoid rebuild and solve related problems. | |
29 | * | |
30 | * To help these things to work well, there are two callbacks for clients. One | |
31 | * is the need callback which is mandatory if user wants to avoid useless | |
32 | * memory allocation at boot-time. The other is optional, init callback, which | |
33 | * is used to do proper initialization after memory is allocated. | |
34 | * | |
35 | * The need callback is used to decide whether extended memory allocation is | |
36 | * needed or not. Sometimes users want to deactivate some features in this | |
37 | * boot and extra memory would be unneccessary. In this case, to avoid | |
38 | * allocating huge chunk of memory, each clients represent their need of | |
39 | * extra memory through the need callback. If one of the need callbacks | |
40 | * returns true, it means that someone needs extra memory so that | |
41 | * page extension core should allocates memory for page extension. If | |
42 | * none of need callbacks return true, memory isn't needed at all in this boot | |
43 | * and page extension core can skip to allocate memory. As result, | |
44 | * none of memory is wasted. | |
45 | * | |
980ac167 JK |
46 | * When need callback returns true, page_ext checks if there is a request for |
47 | * extra memory through size in struct page_ext_operations. If it is non-zero, | |
48 | * extra space is allocated for each page_ext entry and offset is returned to | |
49 | * user through offset in struct page_ext_operations. | |
50 | * | |
eefa864b JK |
51 | * The init callback is used to do proper initialization after page extension |
52 | * is completely initialized. In sparse memory system, extra memory is | |
53 | * allocated some time later than memmap is allocated. In other words, lifetime | |
54 | * of memory for page extension isn't same with memmap for struct page. | |
55 | * Therefore, clients can't store extra data until page extension is | |
56 | * initialized, even if pages are allocated and used freely. This could | |
57 | * cause inadequate state of extra data per page, so, to prevent it, client | |
58 | * can utilize this callback to initialize the state of it correctly. | |
59 | */ | |
60 | ||
61 | static struct page_ext_operations *page_ext_ops[] = { | |
48c96a36 JK |
62 | #ifdef CONFIG_PAGE_OWNER |
63 | &page_owner_ops, | |
64 | #endif | |
33c3fc71 VD |
65 | #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT) |
66 | &page_idle_ops, | |
67 | #endif | |
eefa864b JK |
68 | }; |
69 | ||
5556cfe8 VB |
70 | unsigned long page_ext_size = sizeof(struct page_ext); |
71 | ||
eefa864b JK |
72 | static unsigned long total_usage; |
73 | ||
74 | static bool __init invoke_need_callbacks(void) | |
75 | { | |
76 | int i; | |
77 | int entries = ARRAY_SIZE(page_ext_ops); | |
980ac167 | 78 | bool need = false; |
eefa864b JK |
79 | |
80 | for (i = 0; i < entries; i++) { | |
980ac167 | 81 | if (page_ext_ops[i]->need && page_ext_ops[i]->need()) { |
5556cfe8 VB |
82 | page_ext_ops[i]->offset = page_ext_size; |
83 | page_ext_size += page_ext_ops[i]->size; | |
980ac167 JK |
84 | need = true; |
85 | } | |
eefa864b JK |
86 | } |
87 | ||
980ac167 | 88 | return need; |
eefa864b JK |
89 | } |
90 | ||
91 | static void __init invoke_init_callbacks(void) | |
92 | { | |
93 | int i; | |
94 | int entries = ARRAY_SIZE(page_ext_ops); | |
95 | ||
96 | for (i = 0; i < entries; i++) { | |
97 | if (page_ext_ops[i]->init) | |
98 | page_ext_ops[i]->init(); | |
99 | } | |
100 | } | |
101 | ||
980ac167 JK |
102 | static inline struct page_ext *get_entry(void *base, unsigned long index) |
103 | { | |
5556cfe8 | 104 | return base + page_ext_size * index; |
980ac167 JK |
105 | } |
106 | ||
eefa864b JK |
107 | #if !defined(CONFIG_SPARSEMEM) |
108 | ||
109 | ||
110 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
111 | { | |
112 | pgdat->node_page_ext = NULL; | |
113 | } | |
114 | ||
10ed6341 | 115 | struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
116 | { |
117 | unsigned long pfn = page_to_pfn(page); | |
0b06bb3f | 118 | unsigned long index; |
eefa864b JK |
119 | struct page_ext *base; |
120 | ||
121 | base = NODE_DATA(page_to_nid(page))->node_page_ext; | |
eefa864b JK |
122 | /* |
123 | * The sanity checks the page allocator does upon freeing a | |
124 | * page can reach here before the page_ext arrays are | |
125 | * allocated when feeding a range of pages to the allocator | |
126 | * for the first time during bootup or memory hotplug. | |
127 | */ | |
128 | if (unlikely(!base)) | |
129 | return NULL; | |
0b06bb3f | 130 | index = pfn - round_down(node_start_pfn(page_to_nid(page)), |
eefa864b | 131 | MAX_ORDER_NR_PAGES); |
980ac167 | 132 | return get_entry(base, index); |
eefa864b JK |
133 | } |
134 | ||
135 | static int __init alloc_node_page_ext(int nid) | |
136 | { | |
137 | struct page_ext *base; | |
138 | unsigned long table_size; | |
139 | unsigned long nr_pages; | |
140 | ||
141 | nr_pages = NODE_DATA(nid)->node_spanned_pages; | |
142 | if (!nr_pages) | |
143 | return 0; | |
144 | ||
145 | /* | |
146 | * Need extra space if node range is not aligned with | |
147 | * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm | |
148 | * checks buddy's status, range could be out of exact node range. | |
149 | */ | |
150 | if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) || | |
151 | !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES)) | |
152 | nr_pages += MAX_ORDER_NR_PAGES; | |
153 | ||
5556cfe8 | 154 | table_size = page_ext_size * nr_pages; |
eefa864b | 155 | |
26fb3dae | 156 | base = memblock_alloc_try_nid( |
eefa864b | 157 | table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS), |
97ad1087 | 158 | MEMBLOCK_ALLOC_ACCESSIBLE, nid); |
eefa864b JK |
159 | if (!base) |
160 | return -ENOMEM; | |
161 | NODE_DATA(nid)->node_page_ext = base; | |
162 | total_usage += table_size; | |
163 | return 0; | |
164 | } | |
165 | ||
166 | void __init page_ext_init_flatmem(void) | |
167 | { | |
168 | ||
169 | int nid, fail; | |
170 | ||
171 | if (!invoke_need_callbacks()) | |
172 | return; | |
173 | ||
174 | for_each_online_node(nid) { | |
175 | fail = alloc_node_page_ext(nid); | |
176 | if (fail) | |
177 | goto fail; | |
178 | } | |
179 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
180 | invoke_init_callbacks(); | |
181 | return; | |
182 | ||
183 | fail: | |
184 | pr_crit("allocation of page_ext failed.\n"); | |
185 | panic("Out of memory"); | |
186 | } | |
187 | ||
188 | #else /* CONFIG_FLAT_NODE_MEM_MAP */ | |
189 | ||
10ed6341 | 190 | struct page_ext *lookup_page_ext(const struct page *page) |
eefa864b JK |
191 | { |
192 | unsigned long pfn = page_to_pfn(page); | |
193 | struct mem_section *section = __pfn_to_section(pfn); | |
eefa864b JK |
194 | /* |
195 | * The sanity checks the page allocator does upon freeing a | |
196 | * page can reach here before the page_ext arrays are | |
197 | * allocated when feeding a range of pages to the allocator | |
198 | * for the first time during bootup or memory hotplug. | |
199 | */ | |
200 | if (!section->page_ext) | |
201 | return NULL; | |
980ac167 | 202 | return get_entry(section->page_ext, pfn); |
eefa864b JK |
203 | } |
204 | ||
205 | static void *__meminit alloc_page_ext(size_t size, int nid) | |
206 | { | |
207 | gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN; | |
208 | void *addr = NULL; | |
209 | ||
210 | addr = alloc_pages_exact_nid(nid, size, flags); | |
211 | if (addr) { | |
212 | kmemleak_alloc(addr, size, 1, flags); | |
213 | return addr; | |
214 | } | |
215 | ||
b95046b0 | 216 | addr = vzalloc_node(size, nid); |
eefa864b JK |
217 | |
218 | return addr; | |
219 | } | |
220 | ||
221 | static int __meminit init_section_page_ext(unsigned long pfn, int nid) | |
222 | { | |
223 | struct mem_section *section; | |
224 | struct page_ext *base; | |
225 | unsigned long table_size; | |
226 | ||
227 | section = __pfn_to_section(pfn); | |
228 | ||
229 | if (section->page_ext) | |
230 | return 0; | |
231 | ||
5556cfe8 | 232 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
233 | base = alloc_page_ext(table_size, nid); |
234 | ||
235 | /* | |
236 | * The value stored in section->page_ext is (base - pfn) | |
237 | * and it does not point to the memory block allocated above, | |
238 | * causing kmemleak false positives. | |
239 | */ | |
240 | kmemleak_not_leak(base); | |
241 | ||
242 | if (!base) { | |
243 | pr_err("page ext allocation failure\n"); | |
244 | return -ENOMEM; | |
245 | } | |
246 | ||
247 | /* | |
248 | * The passed "pfn" may not be aligned to SECTION. For the calculation | |
249 | * we need to apply a mask. | |
250 | */ | |
251 | pfn &= PAGE_SECTION_MASK; | |
5556cfe8 | 252 | section->page_ext = (void *)base - page_ext_size * pfn; |
eefa864b JK |
253 | total_usage += table_size; |
254 | return 0; | |
255 | } | |
256 | #ifdef CONFIG_MEMORY_HOTPLUG | |
257 | static void free_page_ext(void *addr) | |
258 | { | |
259 | if (is_vmalloc_addr(addr)) { | |
260 | vfree(addr); | |
261 | } else { | |
262 | struct page *page = virt_to_page(addr); | |
263 | size_t table_size; | |
264 | ||
5556cfe8 | 265 | table_size = page_ext_size * PAGES_PER_SECTION; |
eefa864b JK |
266 | |
267 | BUG_ON(PageReserved(page)); | |
0c815854 | 268 | kmemleak_free(addr); |
eefa864b JK |
269 | free_pages_exact(addr, table_size); |
270 | } | |
271 | } | |
272 | ||
273 | static void __free_page_ext(unsigned long pfn) | |
274 | { | |
275 | struct mem_section *ms; | |
276 | struct page_ext *base; | |
277 | ||
278 | ms = __pfn_to_section(pfn); | |
279 | if (!ms || !ms->page_ext) | |
280 | return; | |
980ac167 | 281 | base = get_entry(ms->page_ext, pfn); |
eefa864b JK |
282 | free_page_ext(base); |
283 | ms->page_ext = NULL; | |
284 | } | |
285 | ||
286 | static int __meminit online_page_ext(unsigned long start_pfn, | |
287 | unsigned long nr_pages, | |
288 | int nid) | |
289 | { | |
290 | unsigned long start, end, pfn; | |
291 | int fail = 0; | |
292 | ||
293 | start = SECTION_ALIGN_DOWN(start_pfn); | |
294 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
295 | ||
98fa15f3 | 296 | if (nid == NUMA_NO_NODE) { |
eefa864b JK |
297 | /* |
298 | * In this case, "nid" already exists and contains valid memory. | |
299 | * "start_pfn" passed to us is a pfn which is an arg for | |
300 | * online__pages(), and start_pfn should exist. | |
301 | */ | |
302 | nid = pfn_to_nid(start_pfn); | |
303 | VM_BUG_ON(!node_state(nid, N_ONLINE)); | |
304 | } | |
305 | ||
306 | for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) { | |
e03d1f78 | 307 | if (!pfn_in_present_section(pfn)) |
eefa864b JK |
308 | continue; |
309 | fail = init_section_page_ext(pfn, nid); | |
310 | } | |
311 | if (!fail) | |
312 | return 0; | |
313 | ||
314 | /* rollback */ | |
315 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
316 | __free_page_ext(pfn); | |
317 | ||
318 | return -ENOMEM; | |
319 | } | |
320 | ||
321 | static int __meminit offline_page_ext(unsigned long start_pfn, | |
322 | unsigned long nr_pages, int nid) | |
323 | { | |
324 | unsigned long start, end, pfn; | |
325 | ||
326 | start = SECTION_ALIGN_DOWN(start_pfn); | |
327 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
328 | ||
329 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
330 | __free_page_ext(pfn); | |
331 | return 0; | |
332 | ||
333 | } | |
334 | ||
335 | static int __meminit page_ext_callback(struct notifier_block *self, | |
336 | unsigned long action, void *arg) | |
337 | { | |
338 | struct memory_notify *mn = arg; | |
339 | int ret = 0; | |
340 | ||
341 | switch (action) { | |
342 | case MEM_GOING_ONLINE: | |
343 | ret = online_page_ext(mn->start_pfn, | |
344 | mn->nr_pages, mn->status_change_nid); | |
345 | break; | |
346 | case MEM_OFFLINE: | |
347 | offline_page_ext(mn->start_pfn, | |
348 | mn->nr_pages, mn->status_change_nid); | |
349 | break; | |
350 | case MEM_CANCEL_ONLINE: | |
351 | offline_page_ext(mn->start_pfn, | |
352 | mn->nr_pages, mn->status_change_nid); | |
353 | break; | |
354 | case MEM_GOING_OFFLINE: | |
355 | break; | |
356 | case MEM_ONLINE: | |
357 | case MEM_CANCEL_OFFLINE: | |
358 | break; | |
359 | } | |
360 | ||
361 | return notifier_from_errno(ret); | |
362 | } | |
363 | ||
364 | #endif | |
365 | ||
366 | void __init page_ext_init(void) | |
367 | { | |
368 | unsigned long pfn; | |
369 | int nid; | |
370 | ||
371 | if (!invoke_need_callbacks()) | |
372 | return; | |
373 | ||
374 | for_each_node_state(nid, N_MEMORY) { | |
375 | unsigned long start_pfn, end_pfn; | |
376 | ||
377 | start_pfn = node_start_pfn(nid); | |
378 | end_pfn = node_end_pfn(nid); | |
379 | /* | |
380 | * start_pfn and end_pfn may not be aligned to SECTION and the | |
381 | * page->flags of out of node pages are not initialized. So we | |
382 | * scan [start_pfn, the biggest section's pfn < end_pfn) here. | |
383 | */ | |
384 | for (pfn = start_pfn; pfn < end_pfn; | |
385 | pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) { | |
386 | ||
387 | if (!pfn_valid(pfn)) | |
388 | continue; | |
389 | /* | |
390 | * Nodes's pfns can be overlapping. | |
391 | * We know some arch can have a nodes layout such as | |
392 | * -------------pfn--------------> | |
393 | * N0 | N1 | N2 | N0 | N1 | N2|.... | |
394 | */ | |
2f1ee091 | 395 | if (pfn_to_nid(pfn) != nid) |
eefa864b JK |
396 | continue; |
397 | if (init_section_page_ext(pfn, nid)) | |
398 | goto oom; | |
0fc542b7 | 399 | cond_resched(); |
eefa864b JK |
400 | } |
401 | } | |
402 | hotplug_memory_notifier(page_ext_callback, 0); | |
403 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
404 | invoke_init_callbacks(); | |
405 | return; | |
406 | ||
407 | oom: | |
408 | panic("Out of memory"); | |
409 | } | |
410 | ||
411 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
412 | { | |
413 | } | |
414 | ||
415 | #endif |