Loading...
Note: File does not exist in v3.1.
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
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 unnecessary. 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 *
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 *
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
61static struct page_ext_operations *page_ext_ops[] = {
62#ifdef CONFIG_PAGE_OWNER
63 &page_owner_ops,
64#endif
65#if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
66 &page_idle_ops,
67#endif
68};
69
70unsigned long page_ext_size = sizeof(struct page_ext);
71
72static unsigned long total_usage;
73
74static bool __init invoke_need_callbacks(void)
75{
76 int i;
77 int entries = ARRAY_SIZE(page_ext_ops);
78 bool need = false;
79
80 for (i = 0; i < entries; i++) {
81 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
82 page_ext_ops[i]->offset = page_ext_size;
83 page_ext_size += page_ext_ops[i]->size;
84 need = true;
85 }
86 }
87
88 return need;
89}
90
91static 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
102#ifndef CONFIG_SPARSEMEM
103void __init page_ext_init_flatmem_late(void)
104{
105 invoke_init_callbacks();
106}
107#endif
108
109static inline struct page_ext *get_entry(void *base, unsigned long index)
110{
111 return base + page_ext_size * index;
112}
113
114#ifndef CONFIG_SPARSEMEM
115
116
117void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
118{
119 pgdat->node_page_ext = NULL;
120}
121
122struct page_ext *lookup_page_ext(const struct page *page)
123{
124 unsigned long pfn = page_to_pfn(page);
125 unsigned long index;
126 struct page_ext *base;
127
128 base = NODE_DATA(page_to_nid(page))->node_page_ext;
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 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
138 MAX_ORDER_NR_PAGES);
139 return get_entry(base, index);
140}
141
142static int __init alloc_node_page_ext(int nid)
143{
144 struct page_ext *base;
145 unsigned long table_size;
146 unsigned long nr_pages;
147
148 nr_pages = NODE_DATA(nid)->node_spanned_pages;
149 if (!nr_pages)
150 return 0;
151
152 /*
153 * Need extra space if node range is not aligned with
154 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
155 * checks buddy's status, range could be out of exact node range.
156 */
157 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
158 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
159 nr_pages += MAX_ORDER_NR_PAGES;
160
161 table_size = page_ext_size * nr_pages;
162
163 base = memblock_alloc_try_nid(
164 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
165 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
166 if (!base)
167 return -ENOMEM;
168 NODE_DATA(nid)->node_page_ext = base;
169 total_usage += table_size;
170 return 0;
171}
172
173void __init page_ext_init_flatmem(void)
174{
175
176 int nid, fail;
177
178 if (!invoke_need_callbacks())
179 return;
180
181 for_each_online_node(nid) {
182 fail = alloc_node_page_ext(nid);
183 if (fail)
184 goto fail;
185 }
186 pr_info("allocated %ld bytes of page_ext\n", total_usage);
187 return;
188
189fail:
190 pr_crit("allocation of page_ext failed.\n");
191 panic("Out of memory");
192}
193
194#else /* CONFIG_FLATMEM */
195
196struct page_ext *lookup_page_ext(const struct page *page)
197{
198 unsigned long pfn = page_to_pfn(page);
199 struct mem_section *section = __pfn_to_section(pfn);
200 /*
201 * The sanity checks the page allocator does upon freeing a
202 * page can reach here before the page_ext arrays are
203 * allocated when feeding a range of pages to the allocator
204 * for the first time during bootup or memory hotplug.
205 */
206 if (!section->page_ext)
207 return NULL;
208 return get_entry(section->page_ext, pfn);
209}
210
211static void *__meminit alloc_page_ext(size_t size, int nid)
212{
213 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
214 void *addr = NULL;
215
216 addr = alloc_pages_exact_nid(nid, size, flags);
217 if (addr) {
218 kmemleak_alloc(addr, size, 1, flags);
219 return addr;
220 }
221
222 addr = vzalloc_node(size, nid);
223
224 return addr;
225}
226
227static int __meminit init_section_page_ext(unsigned long pfn, int nid)
228{
229 struct mem_section *section;
230 struct page_ext *base;
231 unsigned long table_size;
232
233 section = __pfn_to_section(pfn);
234
235 if (section->page_ext)
236 return 0;
237
238 table_size = page_ext_size * PAGES_PER_SECTION;
239 base = alloc_page_ext(table_size, nid);
240
241 /*
242 * The value stored in section->page_ext is (base - pfn)
243 * and it does not point to the memory block allocated above,
244 * causing kmemleak false positives.
245 */
246 kmemleak_not_leak(base);
247
248 if (!base) {
249 pr_err("page ext allocation failure\n");
250 return -ENOMEM;
251 }
252
253 /*
254 * The passed "pfn" may not be aligned to SECTION. For the calculation
255 * we need to apply a mask.
256 */
257 pfn &= PAGE_SECTION_MASK;
258 section->page_ext = (void *)base - page_ext_size * pfn;
259 total_usage += table_size;
260 return 0;
261}
262#ifdef CONFIG_MEMORY_HOTPLUG
263static void free_page_ext(void *addr)
264{
265 if (is_vmalloc_addr(addr)) {
266 vfree(addr);
267 } else {
268 struct page *page = virt_to_page(addr);
269 size_t table_size;
270
271 table_size = page_ext_size * PAGES_PER_SECTION;
272
273 BUG_ON(PageReserved(page));
274 kmemleak_free(addr);
275 free_pages_exact(addr, table_size);
276 }
277}
278
279static void __free_page_ext(unsigned long pfn)
280{
281 struct mem_section *ms;
282 struct page_ext *base;
283
284 ms = __pfn_to_section(pfn);
285 if (!ms || !ms->page_ext)
286 return;
287 base = get_entry(ms->page_ext, pfn);
288 free_page_ext(base);
289 ms->page_ext = NULL;
290}
291
292static int __meminit online_page_ext(unsigned long start_pfn,
293 unsigned long nr_pages,
294 int nid)
295{
296 unsigned long start, end, pfn;
297 int fail = 0;
298
299 start = SECTION_ALIGN_DOWN(start_pfn);
300 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
301
302 if (nid == NUMA_NO_NODE) {
303 /*
304 * In this case, "nid" already exists and contains valid memory.
305 * "start_pfn" passed to us is a pfn which is an arg for
306 * online__pages(), and start_pfn should exist.
307 */
308 nid = pfn_to_nid(start_pfn);
309 VM_BUG_ON(!node_state(nid, N_ONLINE));
310 }
311
312 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION)
313 fail = init_section_page_ext(pfn, nid);
314 if (!fail)
315 return 0;
316
317 /* rollback */
318 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
319 __free_page_ext(pfn);
320
321 return -ENOMEM;
322}
323
324static int __meminit offline_page_ext(unsigned long start_pfn,
325 unsigned long nr_pages, int nid)
326{
327 unsigned long start, end, pfn;
328
329 start = SECTION_ALIGN_DOWN(start_pfn);
330 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
331
332 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
333 __free_page_ext(pfn);
334 return 0;
335
336}
337
338static int __meminit page_ext_callback(struct notifier_block *self,
339 unsigned long action, void *arg)
340{
341 struct memory_notify *mn = arg;
342 int ret = 0;
343
344 switch (action) {
345 case MEM_GOING_ONLINE:
346 ret = online_page_ext(mn->start_pfn,
347 mn->nr_pages, mn->status_change_nid);
348 break;
349 case MEM_OFFLINE:
350 offline_page_ext(mn->start_pfn,
351 mn->nr_pages, mn->status_change_nid);
352 break;
353 case MEM_CANCEL_ONLINE:
354 offline_page_ext(mn->start_pfn,
355 mn->nr_pages, mn->status_change_nid);
356 break;
357 case MEM_GOING_OFFLINE:
358 break;
359 case MEM_ONLINE:
360 case MEM_CANCEL_OFFLINE:
361 break;
362 }
363
364 return notifier_from_errno(ret);
365}
366
367#endif
368
369void __init page_ext_init(void)
370{
371 unsigned long pfn;
372 int nid;
373
374 if (!invoke_need_callbacks())
375 return;
376
377 for_each_node_state(nid, N_MEMORY) {
378 unsigned long start_pfn, end_pfn;
379
380 start_pfn = node_start_pfn(nid);
381 end_pfn = node_end_pfn(nid);
382 /*
383 * start_pfn and end_pfn may not be aligned to SECTION and the
384 * page->flags of out of node pages are not initialized. So we
385 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
386 */
387 for (pfn = start_pfn; pfn < end_pfn;
388 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
389
390 if (!pfn_valid(pfn))
391 continue;
392 /*
393 * Nodes's pfns can be overlapping.
394 * We know some arch can have a nodes layout such as
395 * -------------pfn-------------->
396 * N0 | N1 | N2 | N0 | N1 | N2|....
397 */
398 if (pfn_to_nid(pfn) != nid)
399 continue;
400 if (init_section_page_ext(pfn, nid))
401 goto oom;
402 cond_resched();
403 }
404 }
405 hotplug_memory_notifier(page_ext_callback, 0);
406 pr_info("allocated %ld bytes of page_ext\n", total_usage);
407 invoke_init_callbacks();
408 return;
409
410oom:
411 panic("Out of memory");
412}
413
414void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
415{
416}
417
418#endif