Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma
[linux-2.6-block.git] / mm / page_ext.c
CommitLineData
eefa864b
JK
1#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>
48c96a36 8#include <linux/page_owner.h>
33c3fc71 9#include <linux/page_idle.h>
eefa864b
JK
10
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 *
45 * The init callback is used to do proper initialization after page extension
46 * is completely initialized. In sparse memory system, extra memory is
47 * allocated some time later than memmap is allocated. In other words, lifetime
48 * of memory for page extension isn't same with memmap for struct page.
49 * Therefore, clients can't store extra data until page extension is
50 * initialized, even if pages are allocated and used freely. This could
51 * cause inadequate state of extra data per page, so, to prevent it, client
52 * can utilize this callback to initialize the state of it correctly.
53 */
54
55static struct page_ext_operations *page_ext_ops[] = {
e30825f1
JK
56 &debug_guardpage_ops,
57#ifdef CONFIG_PAGE_POISONING
58 &page_poisoning_ops,
59#endif
48c96a36
JK
60#ifdef CONFIG_PAGE_OWNER
61 &page_owner_ops,
62#endif
33c3fc71
VD
63#if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
64 &page_idle_ops,
65#endif
eefa864b
JK
66};
67
68static unsigned long total_usage;
69
70static bool __init invoke_need_callbacks(void)
71{
72 int i;
73 int entries = ARRAY_SIZE(page_ext_ops);
74
75 for (i = 0; i < entries; i++) {
76 if (page_ext_ops[i]->need && page_ext_ops[i]->need())
77 return true;
78 }
79
80 return false;
81}
82
83static void __init invoke_init_callbacks(void)
84{
85 int i;
86 int entries = ARRAY_SIZE(page_ext_ops);
87
88 for (i = 0; i < entries; i++) {
89 if (page_ext_ops[i]->init)
90 page_ext_ops[i]->init();
91 }
92}
93
94#if !defined(CONFIG_SPARSEMEM)
95
96
97void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
98{
99 pgdat->node_page_ext = NULL;
100}
101
102struct page_ext *lookup_page_ext(struct page *page)
103{
104 unsigned long pfn = page_to_pfn(page);
105 unsigned long offset;
106 struct page_ext *base;
107
108 base = NODE_DATA(page_to_nid(page))->node_page_ext;
1414c7f4 109#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
eefa864b
JK
110 /*
111 * The sanity checks the page allocator does upon freeing a
112 * page can reach here before the page_ext arrays are
113 * allocated when feeding a range of pages to the allocator
114 * for the first time during bootup or memory hotplug.
1414c7f4
LA
115 *
116 * This check is also necessary for ensuring page poisoning
117 * works as expected when enabled
eefa864b
JK
118 */
119 if (unlikely(!base))
120 return NULL;
121#endif
122 offset = pfn - round_down(node_start_pfn(page_to_nid(page)),
123 MAX_ORDER_NR_PAGES);
124 return base + offset;
125}
126
127static int __init alloc_node_page_ext(int nid)
128{
129 struct page_ext *base;
130 unsigned long table_size;
131 unsigned long nr_pages;
132
133 nr_pages = NODE_DATA(nid)->node_spanned_pages;
134 if (!nr_pages)
135 return 0;
136
137 /*
138 * Need extra space if node range is not aligned with
139 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
140 * checks buddy's status, range could be out of exact node range.
141 */
142 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
143 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
144 nr_pages += MAX_ORDER_NR_PAGES;
145
146 table_size = sizeof(struct page_ext) * nr_pages;
147
148 base = memblock_virt_alloc_try_nid_nopanic(
149 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
150 BOOTMEM_ALLOC_ACCESSIBLE, nid);
151 if (!base)
152 return -ENOMEM;
153 NODE_DATA(nid)->node_page_ext = base;
154 total_usage += table_size;
155 return 0;
156}
157
158void __init page_ext_init_flatmem(void)
159{
160
161 int nid, fail;
162
163 if (!invoke_need_callbacks())
164 return;
165
166 for_each_online_node(nid) {
167 fail = alloc_node_page_ext(nid);
168 if (fail)
169 goto fail;
170 }
171 pr_info("allocated %ld bytes of page_ext\n", total_usage);
172 invoke_init_callbacks();
173 return;
174
175fail:
176 pr_crit("allocation of page_ext failed.\n");
177 panic("Out of memory");
178}
179
180#else /* CONFIG_FLAT_NODE_MEM_MAP */
181
182struct page_ext *lookup_page_ext(struct page *page)
183{
184 unsigned long pfn = page_to_pfn(page);
185 struct mem_section *section = __pfn_to_section(pfn);
1414c7f4 186#if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
eefa864b
JK
187 /*
188 * The sanity checks the page allocator does upon freeing a
189 * page can reach here before the page_ext arrays are
190 * allocated when feeding a range of pages to the allocator
191 * for the first time during bootup or memory hotplug.
1414c7f4
LA
192 *
193 * This check is also necessary for ensuring page poisoning
194 * works as expected when enabled
eefa864b
JK
195 */
196 if (!section->page_ext)
197 return NULL;
198#endif
199 return section->page_ext + pfn;
200}
201
202static void *__meminit alloc_page_ext(size_t size, int nid)
203{
204 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
205 void *addr = NULL;
206
207 addr = alloc_pages_exact_nid(nid, size, flags);
208 if (addr) {
209 kmemleak_alloc(addr, size, 1, flags);
210 return addr;
211 }
212
213 if (node_state(nid, N_HIGH_MEMORY))
214 addr = vzalloc_node(size, nid);
215 else
216 addr = vzalloc(size);
217
218 return addr;
219}
220
221static 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
232 table_size = sizeof(struct page_ext) * PAGES_PER_SECTION;
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;
252 section->page_ext = base - pfn;
253 total_usage += table_size;
254 return 0;
255}
256#ifdef CONFIG_MEMORY_HOTPLUG
257static 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
265 table_size = sizeof(struct page_ext) * PAGES_PER_SECTION;
266
267 BUG_ON(PageReserved(page));
268 free_pages_exact(addr, table_size);
269 }
270}
271
272static void __free_page_ext(unsigned long pfn)
273{
274 struct mem_section *ms;
275 struct page_ext *base;
276
277 ms = __pfn_to_section(pfn);
278 if (!ms || !ms->page_ext)
279 return;
280 base = ms->page_ext + pfn;
281 free_page_ext(base);
282 ms->page_ext = NULL;
283}
284
285static int __meminit online_page_ext(unsigned long start_pfn,
286 unsigned long nr_pages,
287 int nid)
288{
289 unsigned long start, end, pfn;
290 int fail = 0;
291
292 start = SECTION_ALIGN_DOWN(start_pfn);
293 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
294
295 if (nid == -1) {
296 /*
297 * In this case, "nid" already exists and contains valid memory.
298 * "start_pfn" passed to us is a pfn which is an arg for
299 * online__pages(), and start_pfn should exist.
300 */
301 nid = pfn_to_nid(start_pfn);
302 VM_BUG_ON(!node_state(nid, N_ONLINE));
303 }
304
305 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
306 if (!pfn_present(pfn))
307 continue;
308 fail = init_section_page_ext(pfn, nid);
309 }
310 if (!fail)
311 return 0;
312
313 /* rollback */
314 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
315 __free_page_ext(pfn);
316
317 return -ENOMEM;
318}
319
320static int __meminit offline_page_ext(unsigned long start_pfn,
321 unsigned long nr_pages, int nid)
322{
323 unsigned long start, end, pfn;
324
325 start = SECTION_ALIGN_DOWN(start_pfn);
326 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
327
328 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
329 __free_page_ext(pfn);
330 return 0;
331
332}
333
334static int __meminit page_ext_callback(struct notifier_block *self,
335 unsigned long action, void *arg)
336{
337 struct memory_notify *mn = arg;
338 int ret = 0;
339
340 switch (action) {
341 case MEM_GOING_ONLINE:
342 ret = online_page_ext(mn->start_pfn,
343 mn->nr_pages, mn->status_change_nid);
344 break;
345 case MEM_OFFLINE:
346 offline_page_ext(mn->start_pfn,
347 mn->nr_pages, mn->status_change_nid);
348 break;
349 case MEM_CANCEL_ONLINE:
350 offline_page_ext(mn->start_pfn,
351 mn->nr_pages, mn->status_change_nid);
352 break;
353 case MEM_GOING_OFFLINE:
354 break;
355 case MEM_ONLINE:
356 case MEM_CANCEL_OFFLINE:
357 break;
358 }
359
360 return notifier_from_errno(ret);
361}
362
363#endif
364
365void __init page_ext_init(void)
366{
367 unsigned long pfn;
368 int nid;
369
370 if (!invoke_need_callbacks())
371 return;
372
373 for_each_node_state(nid, N_MEMORY) {
374 unsigned long start_pfn, end_pfn;
375
376 start_pfn = node_start_pfn(nid);
377 end_pfn = node_end_pfn(nid);
378 /*
379 * start_pfn and end_pfn may not be aligned to SECTION and the
380 * page->flags of out of node pages are not initialized. So we
381 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
382 */
383 for (pfn = start_pfn; pfn < end_pfn;
384 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
385
386 if (!pfn_valid(pfn))
387 continue;
388 /*
389 * Nodes's pfns can be overlapping.
390 * We know some arch can have a nodes layout such as
391 * -------------pfn-------------->
392 * N0 | N1 | N2 | N0 | N1 | N2|....
fe53ca54
YS
393 *
394 * Take into account DEFERRED_STRUCT_PAGE_INIT.
eefa864b 395 */
fe53ca54 396 if (early_pfn_to_nid(pfn) != nid)
eefa864b
JK
397 continue;
398 if (init_section_page_ext(pfn, nid))
399 goto oom;
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
407oom:
408 panic("Out of memory");
409}
410
411void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
412{
413}
414
415#endif