memcg: synchronized LRU
[linux-2.6.git] / mm / page_cgroup.c
1 #include <linux/mm.h>
2 #include <linux/mmzone.h>
3 #include <linux/bootmem.h>
4 #include <linux/bit_spinlock.h>
5 #include <linux/page_cgroup.h>
6 #include <linux/hash.h>
7 #include <linux/slab.h>
8 #include <linux/memory.h>
9 #include <linux/vmalloc.h>
10 #include <linux/cgroup.h>
11 #include <linux/swapops.h>
12
13 static void __meminit
14 __init_page_cgroup(struct page_cgroup *pc, unsigned long pfn)
15 {
16         pc->flags = 0;
17         pc->mem_cgroup = NULL;
18         pc->page = pfn_to_page(pfn);
19         INIT_LIST_HEAD(&pc->lru);
20 }
21 static unsigned long total_usage;
22
23 #if !defined(CONFIG_SPARSEMEM)
24
25
26 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
27 {
28         pgdat->node_page_cgroup = NULL;
29 }
30
31 struct page_cgroup *lookup_page_cgroup(struct page *page)
32 {
33         unsigned long pfn = page_to_pfn(page);
34         unsigned long offset;
35         struct page_cgroup *base;
36
37         base = NODE_DATA(page_to_nid(page))->node_page_cgroup;
38         if (unlikely(!base))
39                 return NULL;
40
41         offset = pfn - NODE_DATA(page_to_nid(page))->node_start_pfn;
42         return base + offset;
43 }
44
45 static int __init alloc_node_page_cgroup(int nid)
46 {
47         struct page_cgroup *base, *pc;
48         unsigned long table_size;
49         unsigned long start_pfn, nr_pages, index;
50
51         start_pfn = NODE_DATA(nid)->node_start_pfn;
52         nr_pages = NODE_DATA(nid)->node_spanned_pages;
53
54         if (!nr_pages)
55                 return 0;
56
57         table_size = sizeof(struct page_cgroup) * nr_pages;
58
59         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
60                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
61         if (!base)
62                 return -ENOMEM;
63         for (index = 0; index < nr_pages; index++) {
64                 pc = base + index;
65                 __init_page_cgroup(pc, start_pfn + index);
66         }
67         NODE_DATA(nid)->node_page_cgroup = base;
68         total_usage += table_size;
69         return 0;
70 }
71
72 void __init page_cgroup_init(void)
73 {
74
75         int nid, fail;
76
77         if (mem_cgroup_subsys.disabled)
78                 return;
79
80         for_each_online_node(nid)  {
81                 fail = alloc_node_page_cgroup(nid);
82                 if (fail)
83                         goto fail;
84         }
85         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
86         printk(KERN_INFO "please try cgroup_disable=memory option if you"
87         " don't want\n");
88         return;
89 fail:
90         printk(KERN_CRIT "allocation of page_cgroup was failed.\n");
91         printk(KERN_CRIT "please try cgroup_disable=memory boot option\n");
92         panic("Out of memory");
93 }
94
95 #else /* CONFIG_FLAT_NODE_MEM_MAP */
96
97 struct page_cgroup *lookup_page_cgroup(struct page *page)
98 {
99         unsigned long pfn = page_to_pfn(page);
100         struct mem_section *section = __pfn_to_section(pfn);
101
102         return section->page_cgroup + pfn;
103 }
104
105 /* __alloc_bootmem...() is protected by !slab_available() */
106 static int __init_refok init_section_page_cgroup(unsigned long pfn)
107 {
108         struct mem_section *section = __pfn_to_section(pfn);
109         struct page_cgroup *base, *pc;
110         unsigned long table_size;
111         int nid, index;
112
113         if (!section->page_cgroup) {
114                 nid = page_to_nid(pfn_to_page(pfn));
115                 table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
116                 if (slab_is_available()) {
117                         base = kmalloc_node(table_size, GFP_KERNEL, nid);
118                         if (!base)
119                                 base = vmalloc_node(table_size, nid);
120                 } else {
121                         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
122                                 table_size,
123                                 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
124                 }
125         } else {
126                 /*
127                  * We don't have to allocate page_cgroup again, but
128                  * address of memmap may be changed. So, we have to initialize
129                  * again.
130                  */
131                 base = section->page_cgroup + pfn;
132                 table_size = 0;
133                 /* check address of memmap is changed or not. */
134                 if (base->page == pfn_to_page(pfn))
135                         return 0;
136         }
137
138         if (!base) {
139                 printk(KERN_ERR "page cgroup allocation failure\n");
140                 return -ENOMEM;
141         }
142
143         for (index = 0; index < PAGES_PER_SECTION; index++) {
144                 pc = base + index;
145                 __init_page_cgroup(pc, pfn + index);
146         }
147
148         section->page_cgroup = base - pfn;
149         total_usage += table_size;
150         return 0;
151 }
152 #ifdef CONFIG_MEMORY_HOTPLUG
153 void __free_page_cgroup(unsigned long pfn)
154 {
155         struct mem_section *ms;
156         struct page_cgroup *base;
157
158         ms = __pfn_to_section(pfn);
159         if (!ms || !ms->page_cgroup)
160                 return;
161         base = ms->page_cgroup + pfn;
162         if (is_vmalloc_addr(base)) {
163                 vfree(base);
164                 ms->page_cgroup = NULL;
165         } else {
166                 struct page *page = virt_to_page(base);
167                 if (!PageReserved(page)) { /* Is bootmem ? */
168                         kfree(base);
169                         ms->page_cgroup = NULL;
170                 }
171         }
172 }
173
174 int __meminit online_page_cgroup(unsigned long start_pfn,
175                         unsigned long nr_pages,
176                         int nid)
177 {
178         unsigned long start, end, pfn;
179         int fail = 0;
180
181         start = start_pfn & ~(PAGES_PER_SECTION - 1);
182         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
183
184         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
185                 if (!pfn_present(pfn))
186                         continue;
187                 fail = init_section_page_cgroup(pfn);
188         }
189         if (!fail)
190                 return 0;
191
192         /* rollback */
193         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
194                 __free_page_cgroup(pfn);
195
196         return -ENOMEM;
197 }
198
199 int __meminit offline_page_cgroup(unsigned long start_pfn,
200                 unsigned long nr_pages, int nid)
201 {
202         unsigned long start, end, pfn;
203
204         start = start_pfn & ~(PAGES_PER_SECTION - 1);
205         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
206
207         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
208                 __free_page_cgroup(pfn);
209         return 0;
210
211 }
212
213 static int __meminit page_cgroup_callback(struct notifier_block *self,
214                                unsigned long action, void *arg)
215 {
216         struct memory_notify *mn = arg;
217         int ret = 0;
218         switch (action) {
219         case MEM_GOING_ONLINE:
220                 ret = online_page_cgroup(mn->start_pfn,
221                                    mn->nr_pages, mn->status_change_nid);
222                 break;
223         case MEM_OFFLINE:
224                 offline_page_cgroup(mn->start_pfn,
225                                 mn->nr_pages, mn->status_change_nid);
226                 break;
227         case MEM_CANCEL_ONLINE:
228         case MEM_GOING_OFFLINE:
229                 break;
230         case MEM_ONLINE:
231         case MEM_CANCEL_OFFLINE:
232                 break;
233         }
234
235         if (ret)
236                 ret = notifier_from_errno(ret);
237         else
238                 ret = NOTIFY_OK;
239
240         return ret;
241 }
242
243 #endif
244
245 void __init page_cgroup_init(void)
246 {
247         unsigned long pfn;
248         int fail = 0;
249
250         if (mem_cgroup_subsys.disabled)
251                 return;
252
253         for (pfn = 0; !fail && pfn < max_pfn; pfn += PAGES_PER_SECTION) {
254                 if (!pfn_present(pfn))
255                         continue;
256                 fail = init_section_page_cgroup(pfn);
257         }
258         if (fail) {
259                 printk(KERN_CRIT "try cgroup_disable=memory boot option\n");
260                 panic("Out of memory");
261         } else {
262                 hotplug_memory_notifier(page_cgroup_callback, 0);
263         }
264         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
265         printk(KERN_INFO "please try cgroup_disable=memory option if you don't"
266         " want\n");
267 }
268
269 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
270 {
271         return;
272 }
273
274 #endif
275
276
277 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
278
279 static DEFINE_MUTEX(swap_cgroup_mutex);
280 struct swap_cgroup_ctrl {
281         struct page **map;
282         unsigned long length;
283 };
284
285 struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
286
287 /*
288  * This 8bytes seems big..maybe we can reduce this when we can use "id" for
289  * cgroup rather than pointer.
290  */
291 struct swap_cgroup {
292         struct mem_cgroup       *val;
293 };
294 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
295 #define SC_POS_MASK     (SC_PER_PAGE - 1)
296
297 /*
298  * SwapCgroup implements "lookup" and "exchange" operations.
299  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
300  * against SwapCache. At swap_free(), this is accessed directly from swap.
301  *
302  * This means,
303  *  - we have no race in "exchange" when we're accessed via SwapCache because
304  *    SwapCache(and its swp_entry) is under lock.
305  *  - When called via swap_free(), there is no user of this entry and no race.
306  * Then, we don't need lock around "exchange".
307  *
308  * TODO: we can push these buffers out to HIGHMEM.
309  */
310
311 /*
312  * allocate buffer for swap_cgroup.
313  */
314 static int swap_cgroup_prepare(int type)
315 {
316         struct page *page;
317         struct swap_cgroup_ctrl *ctrl;
318         unsigned long idx, max;
319
320         if (!do_swap_account)
321                 return 0;
322         ctrl = &swap_cgroup_ctrl[type];
323
324         for (idx = 0; idx < ctrl->length; idx++) {
325                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
326                 if (!page)
327                         goto not_enough_page;
328                 ctrl->map[idx] = page;
329         }
330         return 0;
331 not_enough_page:
332         max = idx;
333         for (idx = 0; idx < max; idx++)
334                 __free_page(ctrl->map[idx]);
335
336         return -ENOMEM;
337 }
338
339 /**
340  * swap_cgroup_record - record mem_cgroup for this swp_entry.
341  * @ent: swap entry to be recorded into
342  * @mem: mem_cgroup to be recorded
343  *
344  * Returns old value at success, NULL at failure.
345  * (Of course, old value can be NULL.)
346  */
347 struct mem_cgroup *swap_cgroup_record(swp_entry_t ent, struct mem_cgroup *mem)
348 {
349         int type = swp_type(ent);
350         unsigned long offset = swp_offset(ent);
351         unsigned long idx = offset / SC_PER_PAGE;
352         unsigned long pos = offset & SC_POS_MASK;
353         struct swap_cgroup_ctrl *ctrl;
354         struct page *mappage;
355         struct swap_cgroup *sc;
356         struct mem_cgroup *old;
357
358         if (!do_swap_account)
359                 return NULL;
360
361         ctrl = &swap_cgroup_ctrl[type];
362
363         mappage = ctrl->map[idx];
364         sc = page_address(mappage);
365         sc += pos;
366         old = sc->val;
367         sc->val = mem;
368
369         return old;
370 }
371
372 /**
373  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
374  * @ent: swap entry to be looked up.
375  *
376  * Returns pointer to mem_cgroup at success. NULL at failure.
377  */
378 struct mem_cgroup *lookup_swap_cgroup(swp_entry_t ent)
379 {
380         int type = swp_type(ent);
381         unsigned long offset = swp_offset(ent);
382         unsigned long idx = offset / SC_PER_PAGE;
383         unsigned long pos = offset & SC_POS_MASK;
384         struct swap_cgroup_ctrl *ctrl;
385         struct page *mappage;
386         struct swap_cgroup *sc;
387         struct mem_cgroup *ret;
388
389         if (!do_swap_account)
390                 return NULL;
391
392         ctrl = &swap_cgroup_ctrl[type];
393         mappage = ctrl->map[idx];
394         sc = page_address(mappage);
395         sc += pos;
396         ret = sc->val;
397         return ret;
398 }
399
400 int swap_cgroup_swapon(int type, unsigned long max_pages)
401 {
402         void *array;
403         unsigned long array_size;
404         unsigned long length;
405         struct swap_cgroup_ctrl *ctrl;
406
407         if (!do_swap_account)
408                 return 0;
409
410         length = ((max_pages/SC_PER_PAGE) + 1);
411         array_size = length * sizeof(void *);
412
413         array = vmalloc(array_size);
414         if (!array)
415                 goto nomem;
416
417         memset(array, 0, array_size);
418         ctrl = &swap_cgroup_ctrl[type];
419         mutex_lock(&swap_cgroup_mutex);
420         ctrl->length = length;
421         ctrl->map = array;
422         if (swap_cgroup_prepare(type)) {
423                 /* memory shortage */
424                 ctrl->map = NULL;
425                 ctrl->length = 0;
426                 vfree(array);
427                 mutex_unlock(&swap_cgroup_mutex);
428                 goto nomem;
429         }
430         mutex_unlock(&swap_cgroup_mutex);
431
432         printk(KERN_INFO
433                 "swap_cgroup: uses %ld bytes of vmalloc for pointer array space"
434                 " and %ld bytes to hold mem_cgroup pointers on swap\n",
435                 array_size, length * PAGE_SIZE);
436         printk(KERN_INFO
437         "swap_cgroup can be disabled by noswapaccount boot option.\n");
438
439         return 0;
440 nomem:
441         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
442         printk(KERN_INFO
443                 "swap_cgroup can be disabled by noswapaccount boot option\n");
444         return -ENOMEM;
445 }
446
447 void swap_cgroup_swapoff(int type)
448 {
449         int i;
450         struct swap_cgroup_ctrl *ctrl;
451
452         if (!do_swap_account)
453                 return;
454
455         mutex_lock(&swap_cgroup_mutex);
456         ctrl = &swap_cgroup_ctrl[type];
457         if (ctrl->map) {
458                 for (i = 0; i < ctrl->length; i++) {
459                         struct page *page = ctrl->map[i];
460                         if (page)
461                                 __free_page(page);
462                 }
463                 vfree(ctrl->map);
464                 ctrl->map = NULL;
465                 ctrl->length = 0;
466         }
467         mutex_unlock(&swap_cgroup_mutex);
468 }
469
470 #endif