685e7c8e1fd631123f2c7093b943dbb38c81c31a
[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 }
20 static unsigned long total_usage;
21
22 #if !defined(CONFIG_SPARSEMEM)
23
24
25 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
26 {
27         pgdat->node_page_cgroup = NULL;
28 }
29
30 struct page_cgroup *lookup_page_cgroup(struct page *page)
31 {
32         unsigned long pfn = page_to_pfn(page);
33         unsigned long offset;
34         struct page_cgroup *base;
35
36         base = NODE_DATA(page_to_nid(page))->node_page_cgroup;
37         if (unlikely(!base))
38                 return NULL;
39
40         offset = pfn - NODE_DATA(page_to_nid(page))->node_start_pfn;
41         return base + offset;
42 }
43
44 static int __init alloc_node_page_cgroup(int nid)
45 {
46         struct page_cgroup *base, *pc;
47         unsigned long table_size;
48         unsigned long start_pfn, nr_pages, index;
49
50         start_pfn = NODE_DATA(nid)->node_start_pfn;
51         nr_pages = NODE_DATA(nid)->node_spanned_pages;
52
53         if (!nr_pages)
54                 return 0;
55
56         table_size = sizeof(struct page_cgroup) * nr_pages;
57
58         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
59                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
60         if (!base)
61                 return -ENOMEM;
62         for (index = 0; index < nr_pages; index++) {
63                 pc = base + index;
64                 __init_page_cgroup(pc, start_pfn + index);
65         }
66         NODE_DATA(nid)->node_page_cgroup = base;
67         total_usage += table_size;
68         return 0;
69 }
70
71 void __init page_cgroup_init(void)
72 {
73
74         int nid, fail;
75
76         if (mem_cgroup_subsys.disabled)
77                 return;
78
79         for_each_online_node(nid)  {
80                 fail = alloc_node_page_cgroup(nid);
81                 if (fail)
82                         goto fail;
83         }
84         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
85         printk(KERN_INFO "please try cgroup_disable=memory option if you"
86         " don't want\n");
87         return;
88 fail:
89         printk(KERN_CRIT "allocation of page_cgroup was failed.\n");
90         printk(KERN_CRIT "please try cgroup_disable=memory boot option\n");
91         panic("Out of memory");
92 }
93
94 #else /* CONFIG_FLAT_NODE_MEM_MAP */
95
96 struct page_cgroup *lookup_page_cgroup(struct page *page)
97 {
98         unsigned long pfn = page_to_pfn(page);
99         struct mem_section *section = __pfn_to_section(pfn);
100
101         return section->page_cgroup + pfn;
102 }
103
104 /* __alloc_bootmem...() is protected by !slab_available() */
105 static int __init_refok init_section_page_cgroup(unsigned long pfn)
106 {
107         struct mem_section *section = __pfn_to_section(pfn);
108         struct page_cgroup *base, *pc;
109         unsigned long table_size;
110         int nid, index;
111
112         if (!section->page_cgroup) {
113                 nid = page_to_nid(pfn_to_page(pfn));
114                 table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
115                 if (slab_is_available()) {
116                         base = kmalloc_node(table_size, GFP_KERNEL, nid);
117                         if (!base)
118                                 base = vmalloc_node(table_size, nid);
119                 } else {
120                         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
121                                 table_size,
122                                 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
123                 }
124         } else {
125                 /*
126                  * We don't have to allocate page_cgroup again, but
127                  * address of memmap may be changed. So, we have to initialize
128                  * again.
129                  */
130                 base = section->page_cgroup + pfn;
131                 table_size = 0;
132                 /* check address of memmap is changed or not. */
133                 if (base->page == pfn_to_page(pfn))
134                         return 0;
135         }
136
137         if (!base) {
138                 printk(KERN_ERR "page cgroup allocation failure\n");
139                 return -ENOMEM;
140         }
141
142         for (index = 0; index < PAGES_PER_SECTION; index++) {
143                 pc = base + index;
144                 __init_page_cgroup(pc, pfn + index);
145         }
146
147         section->page_cgroup = base - pfn;
148         total_usage += table_size;
149         return 0;
150 }
151 #ifdef CONFIG_MEMORY_HOTPLUG
152 void __free_page_cgroup(unsigned long pfn)
153 {
154         struct mem_section *ms;
155         struct page_cgroup *base;
156
157         ms = __pfn_to_section(pfn);
158         if (!ms || !ms->page_cgroup)
159                 return;
160         base = ms->page_cgroup + pfn;
161         if (is_vmalloc_addr(base)) {
162                 vfree(base);
163                 ms->page_cgroup = NULL;
164         } else {
165                 struct page *page = virt_to_page(base);
166                 if (!PageReserved(page)) { /* Is bootmem ? */
167                         kfree(base);
168                         ms->page_cgroup = NULL;
169                 }
170         }
171 }
172
173 int __meminit online_page_cgroup(unsigned long start_pfn,
174                         unsigned long nr_pages,
175                         int nid)
176 {
177         unsigned long start, end, pfn;
178         int fail = 0;
179
180         start = start_pfn & ~(PAGES_PER_SECTION - 1);
181         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
182
183         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
184                 if (!pfn_present(pfn))
185                         continue;
186                 fail = init_section_page_cgroup(pfn);
187         }
188         if (!fail)
189                 return 0;
190
191         /* rollback */
192         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
193                 __free_page_cgroup(pfn);
194
195         return -ENOMEM;
196 }
197
198 int __meminit offline_page_cgroup(unsigned long start_pfn,
199                 unsigned long nr_pages, int nid)
200 {
201         unsigned long start, end, pfn;
202
203         start = start_pfn & ~(PAGES_PER_SECTION - 1);
204         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
205
206         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
207                 __free_page_cgroup(pfn);
208         return 0;
209
210 }
211
212 static int __meminit page_cgroup_callback(struct notifier_block *self,
213                                unsigned long action, void *arg)
214 {
215         struct memory_notify *mn = arg;
216         int ret = 0;
217         switch (action) {
218         case MEM_GOING_ONLINE:
219                 ret = online_page_cgroup(mn->start_pfn,
220                                    mn->nr_pages, mn->status_change_nid);
221                 break;
222         case MEM_OFFLINE:
223                 offline_page_cgroup(mn->start_pfn,
224                                 mn->nr_pages, mn->status_change_nid);
225                 break;
226         case MEM_CANCEL_ONLINE:
227         case MEM_GOING_OFFLINE:
228                 break;
229         case MEM_ONLINE:
230         case MEM_CANCEL_OFFLINE:
231                 break;
232         }
233
234         if (ret)
235                 ret = notifier_from_errno(ret);
236         else
237                 ret = NOTIFY_OK;
238
239         return ret;
240 }
241
242 #endif
243
244 void __init page_cgroup_init(void)
245 {
246         unsigned long pfn;
247         int fail = 0;
248
249         if (mem_cgroup_subsys.disabled)
250                 return;
251
252         for (pfn = 0; !fail && pfn < max_pfn; pfn += PAGES_PER_SECTION) {
253                 if (!pfn_present(pfn))
254                         continue;
255                 fail = init_section_page_cgroup(pfn);
256         }
257         if (fail) {
258                 printk(KERN_CRIT "try cgroup_disable=memory boot option\n");
259                 panic("Out of memory");
260         } else {
261                 hotplug_memory_notifier(page_cgroup_callback, 0);
262         }
263         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
264         printk(KERN_INFO "please try cgroup_disable=memory option if you don't"
265         " want\n");
266 }
267
268 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
269 {
270         return;
271 }
272
273 #endif
274
275
276 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
277
278 static DEFINE_MUTEX(swap_cgroup_mutex);
279 struct swap_cgroup_ctrl {
280         struct page **map;
281         unsigned long length;
282 };
283
284 struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
285
286 /*
287  * This 8bytes seems big..maybe we can reduce this when we can use "id" for
288  * cgroup rather than pointer.
289  */
290 struct swap_cgroup {
291         struct mem_cgroup       *val;
292 };
293 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
294 #define SC_POS_MASK     (SC_PER_PAGE - 1)
295
296 /*
297  * SwapCgroup implements "lookup" and "exchange" operations.
298  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
299  * against SwapCache. At swap_free(), this is accessed directly from swap.
300  *
301  * This means,
302  *  - we have no race in "exchange" when we're accessed via SwapCache because
303  *    SwapCache(and its swp_entry) is under lock.
304  *  - When called via swap_free(), there is no user of this entry and no race.
305  * Then, we don't need lock around "exchange".
306  *
307  * TODO: we can push these buffers out to HIGHMEM.
308  */
309
310 /*
311  * allocate buffer for swap_cgroup.
312  */
313 static int swap_cgroup_prepare(int type)
314 {
315         struct page *page;
316         struct swap_cgroup_ctrl *ctrl;
317         unsigned long idx, max;
318
319         if (!do_swap_account)
320                 return 0;
321         ctrl = &swap_cgroup_ctrl[type];
322
323         for (idx = 0; idx < ctrl->length; idx++) {
324                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
325                 if (!page)
326                         goto not_enough_page;
327                 ctrl->map[idx] = page;
328         }
329         return 0;
330 not_enough_page:
331         max = idx;
332         for (idx = 0; idx < max; idx++)
333                 __free_page(ctrl->map[idx]);
334
335         return -ENOMEM;
336 }
337
338 /**
339  * swap_cgroup_record - record mem_cgroup for this swp_entry.
340  * @ent: swap entry to be recorded into
341  * @mem: mem_cgroup to be recorded
342  *
343  * Returns old value at success, NULL at failure.
344  * (Of course, old value can be NULL.)
345  */
346 struct mem_cgroup *swap_cgroup_record(swp_entry_t ent, struct mem_cgroup *mem)
347 {
348         int type = swp_type(ent);
349         unsigned long offset = swp_offset(ent);
350         unsigned long idx = offset / SC_PER_PAGE;
351         unsigned long pos = offset & SC_POS_MASK;
352         struct swap_cgroup_ctrl *ctrl;
353         struct page *mappage;
354         struct swap_cgroup *sc;
355         struct mem_cgroup *old;
356
357         if (!do_swap_account)
358                 return NULL;
359
360         ctrl = &swap_cgroup_ctrl[type];
361
362         mappage = ctrl->map[idx];
363         sc = page_address(mappage);
364         sc += pos;
365         old = sc->val;
366         sc->val = mem;
367
368         return old;
369 }
370
371 /**
372  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
373  * @ent: swap entry to be looked up.
374  *
375  * Returns pointer to mem_cgroup at success. NULL at failure.
376  */
377 struct mem_cgroup *lookup_swap_cgroup(swp_entry_t ent)
378 {
379         int type = swp_type(ent);
380         unsigned long offset = swp_offset(ent);
381         unsigned long idx = offset / SC_PER_PAGE;
382         unsigned long pos = offset & SC_POS_MASK;
383         struct swap_cgroup_ctrl *ctrl;
384         struct page *mappage;
385         struct swap_cgroup *sc;
386         struct mem_cgroup *ret;
387
388         if (!do_swap_account)
389                 return NULL;
390
391         ctrl = &swap_cgroup_ctrl[type];
392         mappage = ctrl->map[idx];
393         sc = page_address(mappage);
394         sc += pos;
395         ret = sc->val;
396         return ret;
397 }
398
399 int swap_cgroup_swapon(int type, unsigned long max_pages)
400 {
401         void *array;
402         unsigned long array_size;
403         unsigned long length;
404         struct swap_cgroup_ctrl *ctrl;
405
406         if (!do_swap_account)
407                 return 0;
408
409         length = ((max_pages/SC_PER_PAGE) + 1);
410         array_size = length * sizeof(void *);
411
412         array = vmalloc(array_size);
413         if (!array)
414                 goto nomem;
415
416         memset(array, 0, array_size);
417         ctrl = &swap_cgroup_ctrl[type];
418         mutex_lock(&swap_cgroup_mutex);
419         ctrl->length = length;
420         ctrl->map = array;
421         if (swap_cgroup_prepare(type)) {
422                 /* memory shortage */
423                 ctrl->map = NULL;
424                 ctrl->length = 0;
425                 vfree(array);
426                 mutex_unlock(&swap_cgroup_mutex);
427                 goto nomem;
428         }
429         mutex_unlock(&swap_cgroup_mutex);
430
431         printk(KERN_INFO
432                 "swap_cgroup: uses %ld bytes of vmalloc for pointer array space"
433                 " and %ld bytes to hold mem_cgroup pointers on swap\n",
434                 array_size, length * PAGE_SIZE);
435         printk(KERN_INFO
436         "swap_cgroup can be disabled by noswapaccount boot option.\n");
437
438         return 0;
439 nomem:
440         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
441         printk(KERN_INFO
442                 "swap_cgroup can be disabled by noswapaccount boot option\n");
443         return -ENOMEM;
444 }
445
446 void swap_cgroup_swapoff(int type)
447 {
448         int i;
449         struct swap_cgroup_ctrl *ctrl;
450
451         if (!do_swap_account)
452                 return;
453
454         mutex_lock(&swap_cgroup_mutex);
455         ctrl = &swap_cgroup_ctrl[type];
456         if (ctrl->map) {
457                 for (i = 0; i < ctrl->length; i++) {
458                         struct page *page = ctrl->map[i];
459                         if (page)
460                                 __free_page(page);
461                 }
462                 vfree(ctrl->map);
463                 ctrl->map = NULL;
464                 ctrl->length = 0;
465         }
466         mutex_unlock(&swap_cgroup_mutex);
467 }
468
469 #endif