memcg: add interface to move charge at task migration
[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_flatmem(void)
73 {
74
75         int nid, fail;
76
77         if (mem_cgroup_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 memory cgroups\n");
88         return;
89 fail:
90         printk(KERN_CRIT "allocation of page_cgroup 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         if (!section->page_cgroup)
103                 return NULL;
104         return section->page_cgroup + pfn;
105 }
106
107 /* __alloc_bootmem...() is protected by !slab_available() */
108 static int __init_refok init_section_page_cgroup(unsigned long pfn)
109 {
110         struct mem_section *section = __pfn_to_section(pfn);
111         struct page_cgroup *base, *pc;
112         unsigned long table_size;
113         int nid, index;
114
115         if (!section->page_cgroup) {
116                 nid = page_to_nid(pfn_to_page(pfn));
117                 table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
118                 VM_BUG_ON(!slab_is_available());
119                 if (node_state(nid, N_HIGH_MEMORY)) {
120                         base = kmalloc_node(table_size,
121                                 GFP_KERNEL | __GFP_NOWARN, nid);
122                         if (!base)
123                                 base = vmalloc_node(table_size, nid);
124                 } else {
125                         base = kmalloc(table_size, GFP_KERNEL | __GFP_NOWARN);
126                         if (!base)
127                                 base = vmalloc(table_size);
128                 }
129         } else {
130                 /*
131                  * We don't have to allocate page_cgroup again, but
132                  * address of memmap may be changed. So, we have to initialize
133                  * again.
134                  */
135                 base = section->page_cgroup + pfn;
136                 table_size = 0;
137                 /* check address of memmap is changed or not. */
138                 if (base->page == pfn_to_page(pfn))
139                         return 0;
140         }
141
142         if (!base) {
143                 printk(KERN_ERR "page cgroup allocation failure\n");
144                 return -ENOMEM;
145         }
146
147         for (index = 0; index < PAGES_PER_SECTION; index++) {
148                 pc = base + index;
149                 __init_page_cgroup(pc, pfn + index);
150         }
151
152         section->page_cgroup = base - pfn;
153         total_usage += table_size;
154         return 0;
155 }
156 #ifdef CONFIG_MEMORY_HOTPLUG
157 void __free_page_cgroup(unsigned long pfn)
158 {
159         struct mem_section *ms;
160         struct page_cgroup *base;
161
162         ms = __pfn_to_section(pfn);
163         if (!ms || !ms->page_cgroup)
164                 return;
165         base = ms->page_cgroup + pfn;
166         if (is_vmalloc_addr(base)) {
167                 vfree(base);
168                 ms->page_cgroup = NULL;
169         } else {
170                 struct page *page = virt_to_page(base);
171                 if (!PageReserved(page)) { /* Is bootmem ? */
172                         kfree(base);
173                         ms->page_cgroup = NULL;
174                 }
175         }
176 }
177
178 int __meminit online_page_cgroup(unsigned long start_pfn,
179                         unsigned long nr_pages,
180                         int nid)
181 {
182         unsigned long start, end, pfn;
183         int fail = 0;
184
185         start = start_pfn & ~(PAGES_PER_SECTION - 1);
186         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
187
188         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
189                 if (!pfn_present(pfn))
190                         continue;
191                 fail = init_section_page_cgroup(pfn);
192         }
193         if (!fail)
194                 return 0;
195
196         /* rollback */
197         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
198                 __free_page_cgroup(pfn);
199
200         return -ENOMEM;
201 }
202
203 int __meminit offline_page_cgroup(unsigned long start_pfn,
204                 unsigned long nr_pages, int nid)
205 {
206         unsigned long start, end, pfn;
207
208         start = start_pfn & ~(PAGES_PER_SECTION - 1);
209         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
210
211         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
212                 __free_page_cgroup(pfn);
213         return 0;
214
215 }
216
217 static int __meminit page_cgroup_callback(struct notifier_block *self,
218                                unsigned long action, void *arg)
219 {
220         struct memory_notify *mn = arg;
221         int ret = 0;
222         switch (action) {
223         case MEM_GOING_ONLINE:
224                 ret = online_page_cgroup(mn->start_pfn,
225                                    mn->nr_pages, mn->status_change_nid);
226                 break;
227         case MEM_OFFLINE:
228                 offline_page_cgroup(mn->start_pfn,
229                                 mn->nr_pages, mn->status_change_nid);
230                 break;
231         case MEM_CANCEL_ONLINE:
232         case MEM_GOING_OFFLINE:
233                 break;
234         case MEM_ONLINE:
235         case MEM_CANCEL_OFFLINE:
236                 break;
237         }
238
239         if (ret)
240                 ret = notifier_from_errno(ret);
241         else
242                 ret = NOTIFY_OK;
243
244         return ret;
245 }
246
247 #endif
248
249 void __init page_cgroup_init(void)
250 {
251         unsigned long pfn;
252         int fail = 0;
253
254         if (mem_cgroup_disabled())
255                 return;
256
257         for (pfn = 0; !fail && pfn < max_pfn; pfn += PAGES_PER_SECTION) {
258                 if (!pfn_present(pfn))
259                         continue;
260                 fail = init_section_page_cgroup(pfn);
261         }
262         if (fail) {
263                 printk(KERN_CRIT "try 'cgroup_disable=memory' boot option\n");
264                 panic("Out of memory");
265         } else {
266                 hotplug_memory_notifier(page_cgroup_callback, 0);
267         }
268         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
269         printk(KERN_INFO "please try 'cgroup_disable=memory' option if you don't"
270         " want memory cgroups\n");
271 }
272
273 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
274 {
275         return;
276 }
277
278 #endif
279
280
281 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
282
283 static DEFINE_MUTEX(swap_cgroup_mutex);
284 struct swap_cgroup_ctrl {
285         struct page **map;
286         unsigned long length;
287 };
288
289 struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
290
291 struct swap_cgroup {
292         unsigned short          id;
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         ctrl = &swap_cgroup_ctrl[type];
321
322         for (idx = 0; idx < ctrl->length; idx++) {
323                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
324                 if (!page)
325                         goto not_enough_page;
326                 ctrl->map[idx] = page;
327         }
328         return 0;
329 not_enough_page:
330         max = idx;
331         for (idx = 0; idx < max; idx++)
332                 __free_page(ctrl->map[idx]);
333
334         return -ENOMEM;
335 }
336
337 /**
338  * swap_cgroup_record - record mem_cgroup for this swp_entry.
339  * @ent: swap entry to be recorded into
340  * @mem: mem_cgroup to be recorded
341  *
342  * Returns old value at success, 0 at failure.
343  * (Of course, old value can be 0.)
344  */
345 unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
346 {
347         int type = swp_type(ent);
348         unsigned long offset = swp_offset(ent);
349         unsigned long idx = offset / SC_PER_PAGE;
350         unsigned long pos = offset & SC_POS_MASK;
351         struct swap_cgroup_ctrl *ctrl;
352         struct page *mappage;
353         struct swap_cgroup *sc;
354         unsigned short old;
355
356         ctrl = &swap_cgroup_ctrl[type];
357
358         mappage = ctrl->map[idx];
359         sc = page_address(mappage);
360         sc += pos;
361         old = sc->id;
362         sc->id = id;
363
364         return old;
365 }
366
367 /**
368  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
369  * @ent: swap entry to be looked up.
370  *
371  * Returns CSS ID of mem_cgroup at success. 0 at failure. (0 is invalid ID)
372  */
373 unsigned short lookup_swap_cgroup(swp_entry_t ent)
374 {
375         int type = swp_type(ent);
376         unsigned long offset = swp_offset(ent);
377         unsigned long idx = offset / SC_PER_PAGE;
378         unsigned long pos = offset & SC_POS_MASK;
379         struct swap_cgroup_ctrl *ctrl;
380         struct page *mappage;
381         struct swap_cgroup *sc;
382         unsigned short ret;
383
384         ctrl = &swap_cgroup_ctrl[type];
385         mappage = ctrl->map[idx];
386         sc = page_address(mappage);
387         sc += pos;
388         ret = sc->id;
389         return ret;
390 }
391
392 int swap_cgroup_swapon(int type, unsigned long max_pages)
393 {
394         void *array;
395         unsigned long array_size;
396         unsigned long length;
397         struct swap_cgroup_ctrl *ctrl;
398
399         if (!do_swap_account)
400                 return 0;
401
402         length = ((max_pages/SC_PER_PAGE) + 1);
403         array_size = length * sizeof(void *);
404
405         array = vmalloc(array_size);
406         if (!array)
407                 goto nomem;
408
409         memset(array, 0, array_size);
410         ctrl = &swap_cgroup_ctrl[type];
411         mutex_lock(&swap_cgroup_mutex);
412         ctrl->length = length;
413         ctrl->map = array;
414         if (swap_cgroup_prepare(type)) {
415                 /* memory shortage */
416                 ctrl->map = NULL;
417                 ctrl->length = 0;
418                 vfree(array);
419                 mutex_unlock(&swap_cgroup_mutex);
420                 goto nomem;
421         }
422         mutex_unlock(&swap_cgroup_mutex);
423
424         return 0;
425 nomem:
426         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
427         printk(KERN_INFO
428                 "swap_cgroup can be disabled by noswapaccount boot option\n");
429         return -ENOMEM;
430 }
431
432 void swap_cgroup_swapoff(int type)
433 {
434         int i;
435         struct swap_cgroup_ctrl *ctrl;
436
437         if (!do_swap_account)
438                 return;
439
440         mutex_lock(&swap_cgroup_mutex);
441         ctrl = &swap_cgroup_ctrl[type];
442         if (ctrl->map) {
443                 for (i = 0; i < ctrl->length; i++) {
444                         struct page *page = ctrl->map[i];
445                         if (page)
446                                 __free_page(page);
447                 }
448                 vfree(ctrl->map);
449                 ctrl->map = NULL;
450                 ctrl->length = 0;
451         }
452         mutex_unlock(&swap_cgroup_mutex);
453 }
454
455 #endif