Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
eefa864b JK |
2 | #include <linux/mm.h> |
3 | #include <linux/mmzone.h> | |
4 | #include <linux/bootmem.h> | |
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[] = { | |
e30825f1 | 62 | &debug_guardpage_ops, |
48c96a36 JK |
63 | #ifdef CONFIG_PAGE_OWNER |
64 | &page_owner_ops, | |
65 | #endif | |
33c3fc71 VD |
66 | #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT) |
67 | &page_idle_ops, | |
68 | #endif | |
eefa864b JK |
69 | }; |
70 | ||
71 | static unsigned long total_usage; | |
980ac167 | 72 | static unsigned long extra_mem; |
eefa864b JK |
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 JK |
81 | if (page_ext_ops[i]->need && page_ext_ops[i]->need()) { |
82 | page_ext_ops[i]->offset = sizeof(struct page_ext) + | |
83 | extra_mem; | |
84 | extra_mem += page_ext_ops[i]->size; | |
85 | need = true; | |
86 | } | |
eefa864b JK |
87 | } |
88 | ||
980ac167 | 89 | return need; |
eefa864b JK |
90 | } |
91 | ||
92 | static void __init invoke_init_callbacks(void) | |
93 | { | |
94 | int i; | |
95 | int entries = ARRAY_SIZE(page_ext_ops); | |
96 | ||
97 | for (i = 0; i < entries; i++) { | |
98 | if (page_ext_ops[i]->init) | |
99 | page_ext_ops[i]->init(); | |
100 | } | |
101 | } | |
102 | ||
980ac167 JK |
103 | static unsigned long get_entry_size(void) |
104 | { | |
105 | return sizeof(struct page_ext) + extra_mem; | |
106 | } | |
107 | ||
108 | static inline struct page_ext *get_entry(void *base, unsigned long index) | |
109 | { | |
110 | return base + get_entry_size() * index; | |
111 | } | |
112 | ||
eefa864b JK |
113 | #if !defined(CONFIG_SPARSEMEM) |
114 | ||
115 | ||
116 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
117 | { | |
118 | pgdat->node_page_ext = NULL; | |
119 | } | |
120 | ||
121 | struct page_ext *lookup_page_ext(struct page *page) | |
122 | { | |
123 | unsigned long pfn = page_to_pfn(page); | |
0b06bb3f | 124 | unsigned long index; |
eefa864b JK |
125 | struct page_ext *base; |
126 | ||
127 | base = NODE_DATA(page_to_nid(page))->node_page_ext; | |
bd33ef36 | 128 | #if defined(CONFIG_DEBUG_VM) |
eefa864b JK |
129 | /* |
130 | * The sanity checks the page allocator does upon freeing a | |
131 | * page can reach here before the page_ext arrays are | |
132 | * allocated when feeding a range of pages to the allocator | |
133 | * for the first time during bootup or memory hotplug. | |
134 | */ | |
135 | if (unlikely(!base)) | |
136 | return NULL; | |
137 | #endif | |
0b06bb3f | 138 | index = pfn - round_down(node_start_pfn(page_to_nid(page)), |
eefa864b | 139 | MAX_ORDER_NR_PAGES); |
980ac167 | 140 | return get_entry(base, index); |
eefa864b JK |
141 | } |
142 | ||
143 | static int __init alloc_node_page_ext(int nid) | |
144 | { | |
145 | struct page_ext *base; | |
146 | unsigned long table_size; | |
147 | unsigned long nr_pages; | |
148 | ||
149 | nr_pages = NODE_DATA(nid)->node_spanned_pages; | |
150 | if (!nr_pages) | |
151 | return 0; | |
152 | ||
153 | /* | |
154 | * Need extra space if node range is not aligned with | |
155 | * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm | |
156 | * checks buddy's status, range could be out of exact node range. | |
157 | */ | |
158 | if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) || | |
159 | !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES)) | |
160 | nr_pages += MAX_ORDER_NR_PAGES; | |
161 | ||
980ac167 | 162 | table_size = get_entry_size() * nr_pages; |
eefa864b JK |
163 | |
164 | base = memblock_virt_alloc_try_nid_nopanic( | |
165 | table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS), | |
166 | BOOTMEM_ALLOC_ACCESSIBLE, nid); | |
167 | if (!base) | |
168 | return -ENOMEM; | |
169 | NODE_DATA(nid)->node_page_ext = base; | |
170 | total_usage += table_size; | |
171 | return 0; | |
172 | } | |
173 | ||
174 | void __init page_ext_init_flatmem(void) | |
175 | { | |
176 | ||
177 | int nid, fail; | |
178 | ||
179 | if (!invoke_need_callbacks()) | |
180 | return; | |
181 | ||
182 | for_each_online_node(nid) { | |
183 | fail = alloc_node_page_ext(nid); | |
184 | if (fail) | |
185 | goto fail; | |
186 | } | |
187 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
188 | invoke_init_callbacks(); | |
189 | return; | |
190 | ||
191 | fail: | |
192 | pr_crit("allocation of page_ext failed.\n"); | |
193 | panic("Out of memory"); | |
194 | } | |
195 | ||
196 | #else /* CONFIG_FLAT_NODE_MEM_MAP */ | |
197 | ||
198 | struct page_ext *lookup_page_ext(struct page *page) | |
199 | { | |
200 | unsigned long pfn = page_to_pfn(page); | |
201 | struct mem_section *section = __pfn_to_section(pfn); | |
bd33ef36 | 202 | #if defined(CONFIG_DEBUG_VM) |
eefa864b JK |
203 | /* |
204 | * The sanity checks the page allocator does upon freeing a | |
205 | * page can reach here before the page_ext arrays are | |
206 | * allocated when feeding a range of pages to the allocator | |
207 | * for the first time during bootup or memory hotplug. | |
208 | */ | |
209 | if (!section->page_ext) | |
210 | return NULL; | |
211 | #endif | |
980ac167 | 212 | return get_entry(section->page_ext, pfn); |
eefa864b JK |
213 | } |
214 | ||
215 | static void *__meminit alloc_page_ext(size_t size, int nid) | |
216 | { | |
217 | gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN; | |
218 | void *addr = NULL; | |
219 | ||
220 | addr = alloc_pages_exact_nid(nid, size, flags); | |
221 | if (addr) { | |
222 | kmemleak_alloc(addr, size, 1, flags); | |
223 | return addr; | |
224 | } | |
225 | ||
b95046b0 | 226 | addr = vzalloc_node(size, nid); |
eefa864b JK |
227 | |
228 | return addr; | |
229 | } | |
230 | ||
231 | static int __meminit init_section_page_ext(unsigned long pfn, int nid) | |
232 | { | |
233 | struct mem_section *section; | |
234 | struct page_ext *base; | |
235 | unsigned long table_size; | |
236 | ||
237 | section = __pfn_to_section(pfn); | |
238 | ||
239 | if (section->page_ext) | |
240 | return 0; | |
241 | ||
980ac167 | 242 | table_size = get_entry_size() * PAGES_PER_SECTION; |
eefa864b JK |
243 | base = alloc_page_ext(table_size, nid); |
244 | ||
245 | /* | |
246 | * The value stored in section->page_ext is (base - pfn) | |
247 | * and it does not point to the memory block allocated above, | |
248 | * causing kmemleak false positives. | |
249 | */ | |
250 | kmemleak_not_leak(base); | |
251 | ||
252 | if (!base) { | |
253 | pr_err("page ext allocation failure\n"); | |
254 | return -ENOMEM; | |
255 | } | |
256 | ||
257 | /* | |
258 | * The passed "pfn" may not be aligned to SECTION. For the calculation | |
259 | * we need to apply a mask. | |
260 | */ | |
261 | pfn &= PAGE_SECTION_MASK; | |
980ac167 | 262 | section->page_ext = (void *)base - get_entry_size() * pfn; |
eefa864b JK |
263 | total_usage += table_size; |
264 | return 0; | |
265 | } | |
266 | #ifdef CONFIG_MEMORY_HOTPLUG | |
267 | static void free_page_ext(void *addr) | |
268 | { | |
269 | if (is_vmalloc_addr(addr)) { | |
270 | vfree(addr); | |
271 | } else { | |
272 | struct page *page = virt_to_page(addr); | |
273 | size_t table_size; | |
274 | ||
980ac167 | 275 | table_size = get_entry_size() * PAGES_PER_SECTION; |
eefa864b JK |
276 | |
277 | BUG_ON(PageReserved(page)); | |
278 | free_pages_exact(addr, table_size); | |
279 | } | |
280 | } | |
281 | ||
282 | static void __free_page_ext(unsigned long pfn) | |
283 | { | |
284 | struct mem_section *ms; | |
285 | struct page_ext *base; | |
286 | ||
287 | ms = __pfn_to_section(pfn); | |
288 | if (!ms || !ms->page_ext) | |
289 | return; | |
980ac167 | 290 | base = get_entry(ms->page_ext, pfn); |
eefa864b JK |
291 | free_page_ext(base); |
292 | ms->page_ext = NULL; | |
293 | } | |
294 | ||
295 | static int __meminit online_page_ext(unsigned long start_pfn, | |
296 | unsigned long nr_pages, | |
297 | int nid) | |
298 | { | |
299 | unsigned long start, end, pfn; | |
300 | int fail = 0; | |
301 | ||
302 | start = SECTION_ALIGN_DOWN(start_pfn); | |
303 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
304 | ||
305 | if (nid == -1) { | |
306 | /* | |
307 | * In this case, "nid" already exists and contains valid memory. | |
308 | * "start_pfn" passed to us is a pfn which is an arg for | |
309 | * online__pages(), and start_pfn should exist. | |
310 | */ | |
311 | nid = pfn_to_nid(start_pfn); | |
312 | VM_BUG_ON(!node_state(nid, N_ONLINE)); | |
313 | } | |
314 | ||
315 | for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) { | |
316 | if (!pfn_present(pfn)) | |
317 | continue; | |
318 | fail = init_section_page_ext(pfn, nid); | |
319 | } | |
320 | if (!fail) | |
321 | return 0; | |
322 | ||
323 | /* rollback */ | |
324 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
325 | __free_page_ext(pfn); | |
326 | ||
327 | return -ENOMEM; | |
328 | } | |
329 | ||
330 | static int __meminit offline_page_ext(unsigned long start_pfn, | |
331 | unsigned long nr_pages, int nid) | |
332 | { | |
333 | unsigned long start, end, pfn; | |
334 | ||
335 | start = SECTION_ALIGN_DOWN(start_pfn); | |
336 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); | |
337 | ||
338 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) | |
339 | __free_page_ext(pfn); | |
340 | return 0; | |
341 | ||
342 | } | |
343 | ||
344 | static int __meminit page_ext_callback(struct notifier_block *self, | |
345 | unsigned long action, void *arg) | |
346 | { | |
347 | struct memory_notify *mn = arg; | |
348 | int ret = 0; | |
349 | ||
350 | switch (action) { | |
351 | case MEM_GOING_ONLINE: | |
352 | ret = online_page_ext(mn->start_pfn, | |
353 | mn->nr_pages, mn->status_change_nid); | |
354 | break; | |
355 | case MEM_OFFLINE: | |
356 | offline_page_ext(mn->start_pfn, | |
357 | mn->nr_pages, mn->status_change_nid); | |
358 | break; | |
359 | case MEM_CANCEL_ONLINE: | |
360 | offline_page_ext(mn->start_pfn, | |
361 | mn->nr_pages, mn->status_change_nid); | |
362 | break; | |
363 | case MEM_GOING_OFFLINE: | |
364 | break; | |
365 | case MEM_ONLINE: | |
366 | case MEM_CANCEL_OFFLINE: | |
367 | break; | |
368 | } | |
369 | ||
370 | return notifier_from_errno(ret); | |
371 | } | |
372 | ||
373 | #endif | |
374 | ||
375 | void __init page_ext_init(void) | |
376 | { | |
377 | unsigned long pfn; | |
378 | int nid; | |
379 | ||
380 | if (!invoke_need_callbacks()) | |
381 | return; | |
382 | ||
383 | for_each_node_state(nid, N_MEMORY) { | |
384 | unsigned long start_pfn, end_pfn; | |
385 | ||
386 | start_pfn = node_start_pfn(nid); | |
387 | end_pfn = node_end_pfn(nid); | |
388 | /* | |
389 | * start_pfn and end_pfn may not be aligned to SECTION and the | |
390 | * page->flags of out of node pages are not initialized. So we | |
391 | * scan [start_pfn, the biggest section's pfn < end_pfn) here. | |
392 | */ | |
393 | for (pfn = start_pfn; pfn < end_pfn; | |
394 | pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) { | |
395 | ||
396 | if (!pfn_valid(pfn)) | |
397 | continue; | |
398 | /* | |
399 | * Nodes's pfns can be overlapping. | |
400 | * We know some arch can have a nodes layout such as | |
401 | * -------------pfn--------------> | |
402 | * N0 | N1 | N2 | N0 | N1 | N2|.... | |
fe53ca54 YS |
403 | * |
404 | * Take into account DEFERRED_STRUCT_PAGE_INIT. | |
eefa864b | 405 | */ |
fe53ca54 | 406 | if (early_pfn_to_nid(pfn) != nid) |
eefa864b JK |
407 | continue; |
408 | if (init_section_page_ext(pfn, nid)) | |
409 | goto oom; | |
0fc542b7 | 410 | cond_resched(); |
eefa864b JK |
411 | } |
412 | } | |
413 | hotplug_memory_notifier(page_ext_callback, 0); | |
414 | pr_info("allocated %ld bytes of page_ext\n", total_usage); | |
415 | invoke_init_callbacks(); | |
416 | return; | |
417 | ||
418 | oom: | |
419 | panic("Out of memory"); | |
420 | } | |
421 | ||
422 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) | |
423 | { | |
424 | } | |
425 | ||
426 | #endif |