Merge branch 'pci-for-jesse' of git://git.kernel.org/pub/scm/linux/kernel/git/tip...
[linux-2.6] / arch / powerpc / platforms / pseries / cmm.c
1 /*
2  * Collaborative memory management interface.
3  *
4  * Copyright (C) 2008 IBM Corporation
5  * Author(s): Brian King (brking@linux.vnet.ibm.com),
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 2 of the License, or
10  * (at your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
20  *
21  */
22
23 #include <linux/ctype.h>
24 #include <linux/delay.h>
25 #include <linux/errno.h>
26 #include <linux/fs.h>
27 #include <linux/init.h>
28 #include <linux/kthread.h>
29 #include <linux/module.h>
30 #include <linux/oom.h>
31 #include <linux/sched.h>
32 #include <linux/stringify.h>
33 #include <linux/swap.h>
34 #include <linux/sysdev.h>
35 #include <asm/firmware.h>
36 #include <asm/hvcall.h>
37 #include <asm/mmu.h>
38 #include <asm/pgalloc.h>
39 #include <asm/uaccess.h>
40
41 #include "plpar_wrappers.h"
42
43 #define CMM_DRIVER_VERSION      "1.0.0"
44 #define CMM_DEFAULT_DELAY       1
45 #define CMM_DEBUG                       0
46 #define CMM_DISABLE             0
47 #define CMM_OOM_KB              1024
48 #define CMM_MIN_MEM_MB          256
49 #define KB2PAGES(_p)            ((_p)>>(PAGE_SHIFT-10))
50 #define PAGES2KB(_p)            ((_p)<<(PAGE_SHIFT-10))
51
52 static unsigned int delay = CMM_DEFAULT_DELAY;
53 static unsigned int oom_kb = CMM_OOM_KB;
54 static unsigned int cmm_debug = CMM_DEBUG;
55 static unsigned int cmm_disabled = CMM_DISABLE;
56 static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
57 static struct sys_device cmm_sysdev;
58
59 MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
60 MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
61 MODULE_LICENSE("GPL");
62 MODULE_VERSION(CMM_DRIVER_VERSION);
63
64 module_param_named(delay, delay, uint, S_IRUGO | S_IWUSR);
65 MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
66                  "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
67 module_param_named(oom_kb, oom_kb, uint, S_IRUGO | S_IWUSR);
68 MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
69                  "[Default=" __stringify(CMM_OOM_KB) "]");
70 module_param_named(min_mem_mb, min_mem_mb, ulong, S_IRUGO | S_IWUSR);
71 MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
72                  "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
73 module_param_named(debug, cmm_debug, uint, S_IRUGO | S_IWUSR);
74 MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
75                  "[Default=" __stringify(CMM_DEBUG) "]");
76
77 #define CMM_NR_PAGES ((PAGE_SIZE - sizeof(void *) - sizeof(unsigned long)) / sizeof(unsigned long))
78
79 #define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
80
81 struct cmm_page_array {
82         struct cmm_page_array *next;
83         unsigned long index;
84         unsigned long page[CMM_NR_PAGES];
85 };
86
87 static unsigned long loaned_pages;
88 static unsigned long loaned_pages_target;
89 static unsigned long oom_freed_pages;
90
91 static struct cmm_page_array *cmm_page_list;
92 static DEFINE_SPINLOCK(cmm_lock);
93
94 static struct task_struct *cmm_thread_ptr;
95
96 /**
97  * cmm_alloc_pages - Allocate pages and mark them as loaned
98  * @nr: number of pages to allocate
99  *
100  * Return value:
101  *      number of pages requested to be allocated which were not
102  **/
103 static long cmm_alloc_pages(long nr)
104 {
105         struct cmm_page_array *pa, *npa;
106         unsigned long addr;
107         long rc;
108
109         cmm_dbg("Begin request for %ld pages\n", nr);
110
111         while (nr) {
112                 addr = __get_free_page(GFP_NOIO | __GFP_NOWARN |
113                                        __GFP_NORETRY | __GFP_NOMEMALLOC);
114                 if (!addr)
115                         break;
116                 spin_lock(&cmm_lock);
117                 pa = cmm_page_list;
118                 if (!pa || pa->index >= CMM_NR_PAGES) {
119                         /* Need a new page for the page list. */
120                         spin_unlock(&cmm_lock);
121                         npa = (struct cmm_page_array *)__get_free_page(GFP_NOIO | __GFP_NOWARN |
122                                                                        __GFP_NORETRY | __GFP_NOMEMALLOC);
123                         if (!npa) {
124                                 pr_info("%s: Can not allocate new page list\n", __FUNCTION__);
125                                 free_page(addr);
126                                 break;
127                         }
128                         spin_lock(&cmm_lock);
129                         pa = cmm_page_list;
130
131                         if (!pa || pa->index >= CMM_NR_PAGES) {
132                                 npa->next = pa;
133                                 npa->index = 0;
134                                 pa = npa;
135                                 cmm_page_list = pa;
136                         } else
137                                 free_page((unsigned long) npa);
138                 }
139
140                 if ((rc = plpar_page_set_loaned(__pa(addr)))) {
141                         pr_err("%s: Can not set page to loaned. rc=%ld\n", __FUNCTION__, rc);
142                         spin_unlock(&cmm_lock);
143                         free_page(addr);
144                         break;
145                 }
146
147                 pa->page[pa->index++] = addr;
148                 loaned_pages++;
149                 totalram_pages--;
150                 spin_unlock(&cmm_lock);
151                 nr--;
152         }
153
154         cmm_dbg("End request with %ld pages unfulfilled\n", nr);
155         return nr;
156 }
157
158 /**
159  * cmm_free_pages - Free pages and mark them as active
160  * @nr: number of pages to free
161  *
162  * Return value:
163  *      number of pages requested to be freed which were not
164  **/
165 static long cmm_free_pages(long nr)
166 {
167         struct cmm_page_array *pa;
168         unsigned long addr;
169
170         cmm_dbg("Begin free of %ld pages.\n", nr);
171         spin_lock(&cmm_lock);
172         pa = cmm_page_list;
173         while (nr) {
174                 if (!pa || pa->index <= 0)
175                         break;
176                 addr = pa->page[--pa->index];
177
178                 if (pa->index == 0) {
179                         pa = pa->next;
180                         free_page((unsigned long) cmm_page_list);
181                         cmm_page_list = pa;
182                 }
183
184                 plpar_page_set_active(__pa(addr));
185                 free_page(addr);
186                 loaned_pages--;
187                 nr--;
188                 totalram_pages++;
189         }
190         spin_unlock(&cmm_lock);
191         cmm_dbg("End request with %ld pages unfulfilled\n", nr);
192         return nr;
193 }
194
195 /**
196  * cmm_oom_notify - OOM notifier
197  * @self:       notifier block struct
198  * @dummy:      not used
199  * @parm:       returned - number of pages freed
200  *
201  * Return value:
202  *      NOTIFY_OK
203  **/
204 static int cmm_oom_notify(struct notifier_block *self,
205                           unsigned long dummy, void *parm)
206 {
207         unsigned long *freed = parm;
208         long nr = KB2PAGES(oom_kb);
209
210         cmm_dbg("OOM processing started\n");
211         nr = cmm_free_pages(nr);
212         loaned_pages_target = loaned_pages;
213         *freed += KB2PAGES(oom_kb) - nr;
214         oom_freed_pages += KB2PAGES(oom_kb) - nr;
215         cmm_dbg("OOM processing complete\n");
216         return NOTIFY_OK;
217 }
218
219 /**
220  * cmm_get_mpp - Read memory performance parameters
221  *
222  * Makes hcall to query the current page loan request from the hypervisor.
223  *
224  * Return value:
225  *      nothing
226  **/
227 static void cmm_get_mpp(void)
228 {
229         int rc;
230         struct hvcall_mpp_data mpp_data;
231         unsigned long active_pages_target;
232         signed long page_loan_request;
233
234         rc = h_get_mpp(&mpp_data);
235
236         if (rc != H_SUCCESS)
237                 return;
238
239         page_loan_request = div_s64((s64)mpp_data.loan_request, PAGE_SIZE);
240         loaned_pages_target = page_loan_request + loaned_pages;
241         if (loaned_pages_target > oom_freed_pages)
242                 loaned_pages_target -= oom_freed_pages;
243         else
244                 loaned_pages_target = 0;
245
246         active_pages_target = totalram_pages + loaned_pages - loaned_pages_target;
247
248         if ((min_mem_mb * 1024 * 1024) > (active_pages_target * PAGE_SIZE))
249                 loaned_pages_target = totalram_pages + loaned_pages -
250                         ((min_mem_mb * 1024 * 1024) / PAGE_SIZE);
251
252         cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
253                 page_loan_request, loaned_pages, loaned_pages_target,
254                 oom_freed_pages, totalram_pages);
255 }
256
257 static struct notifier_block cmm_oom_nb = {
258         .notifier_call = cmm_oom_notify
259 };
260
261 /**
262  * cmm_thread - CMM task thread
263  * @dummy:      not used
264  *
265  * Return value:
266  *      0
267  **/
268 static int cmm_thread(void *dummy)
269 {
270         unsigned long timeleft;
271
272         while (1) {
273                 timeleft = msleep_interruptible(delay * 1000);
274
275                 if (kthread_should_stop() || timeleft) {
276                         loaned_pages_target = loaned_pages;
277                         break;
278                 }
279
280                 cmm_get_mpp();
281
282                 if (loaned_pages_target > loaned_pages) {
283                         if (cmm_alloc_pages(loaned_pages_target - loaned_pages))
284                                 loaned_pages_target = loaned_pages;
285                 } else if (loaned_pages_target < loaned_pages)
286                         cmm_free_pages(loaned_pages - loaned_pages_target);
287         }
288         return 0;
289 }
290
291 #define CMM_SHOW(name, format, args...)                 \
292         static ssize_t show_##name(struct sys_device *dev,      \
293                                    struct sysdev_attribute *attr,       \
294                                    char *buf)                   \
295         {                                                       \
296                 return sprintf(buf, format, ##args);            \
297         }                                                       \
298         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
299
300 CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(loaned_pages));
301 CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
302
303 static ssize_t show_oom_pages(struct sys_device *dev,
304                               struct sysdev_attribute *attr, char *buf)
305 {
306         return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
307 }
308
309 static ssize_t store_oom_pages(struct sys_device *dev,
310                                struct sysdev_attribute *attr,
311                                const char *buf, size_t count)
312 {
313         unsigned long val = simple_strtoul (buf, NULL, 10);
314
315         if (!capable(CAP_SYS_ADMIN))
316                 return -EPERM;
317         if (val != 0)
318                 return -EBADMSG;
319
320         oom_freed_pages = 0;
321         return count;
322 }
323
324 static SYSDEV_ATTR(oom_freed_kb, S_IWUSR| S_IRUGO,
325                    show_oom_pages, store_oom_pages);
326
327 static struct sysdev_attribute *cmm_attrs[] = {
328         &attr_loaned_kb,
329         &attr_loaned_target_kb,
330         &attr_oom_freed_kb,
331 };
332
333 static struct sysdev_class cmm_sysdev_class = {
334         .name = "cmm",
335 };
336
337 /**
338  * cmm_sysfs_register - Register with sysfs
339  *
340  * Return value:
341  *      0 on success / other on failure
342  **/
343 static int cmm_sysfs_register(struct sys_device *sysdev)
344 {
345         int i, rc;
346
347         if ((rc = sysdev_class_register(&cmm_sysdev_class)))
348                 return rc;
349
350         sysdev->id = 0;
351         sysdev->cls = &cmm_sysdev_class;
352
353         if ((rc = sysdev_register(sysdev)))
354                 goto class_unregister;
355
356         for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
357                 if ((rc = sysdev_create_file(sysdev, cmm_attrs[i])))
358                         goto fail;
359         }
360
361         return 0;
362
363 fail:
364         while (--i >= 0)
365                 sysdev_remove_file(sysdev, cmm_attrs[i]);
366         sysdev_unregister(sysdev);
367 class_unregister:
368         sysdev_class_unregister(&cmm_sysdev_class);
369         return rc;
370 }
371
372 /**
373  * cmm_unregister_sysfs - Unregister from sysfs
374  *
375  **/
376 static void cmm_unregister_sysfs(struct sys_device *sysdev)
377 {
378         int i;
379
380         for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
381                 sysdev_remove_file(sysdev, cmm_attrs[i]);
382         sysdev_unregister(sysdev);
383         sysdev_class_unregister(&cmm_sysdev_class);
384 }
385
386 /**
387  * cmm_init - Module initialization
388  *
389  * Return value:
390  *      0 on success / other on failure
391  **/
392 static int cmm_init(void)
393 {
394         int rc = -ENOMEM;
395
396         if (!firmware_has_feature(FW_FEATURE_CMO))
397                 return -EOPNOTSUPP;
398
399         if ((rc = register_oom_notifier(&cmm_oom_nb)) < 0)
400                 return rc;
401
402         if ((rc = cmm_sysfs_register(&cmm_sysdev)))
403                 goto out_oom_notifier;
404
405         if (cmm_disabled)
406                 return rc;
407
408         cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
409         if (IS_ERR(cmm_thread_ptr)) {
410                 rc = PTR_ERR(cmm_thread_ptr);
411                 goto out_unregister_sysfs;
412         }
413
414         return rc;
415
416 out_unregister_sysfs:
417         cmm_unregister_sysfs(&cmm_sysdev);
418 out_oom_notifier:
419         unregister_oom_notifier(&cmm_oom_nb);
420         return rc;
421 }
422
423 /**
424  * cmm_exit - Module exit
425  *
426  * Return value:
427  *      nothing
428  **/
429 static void cmm_exit(void)
430 {
431         if (cmm_thread_ptr)
432                 kthread_stop(cmm_thread_ptr);
433         unregister_oom_notifier(&cmm_oom_nb);
434         cmm_free_pages(loaned_pages);
435         cmm_unregister_sysfs(&cmm_sysdev);
436 }
437
438 /**
439  * cmm_set_disable - Disable/Enable CMM
440  *
441  * Return value:
442  *      0 on success / other on failure
443  **/
444 static int cmm_set_disable(const char *val, struct kernel_param *kp)
445 {
446         int disable = simple_strtoul(val, NULL, 10);
447
448         if (disable != 0 && disable != 1)
449                 return -EINVAL;
450
451         if (disable && !cmm_disabled) {
452                 if (cmm_thread_ptr)
453                         kthread_stop(cmm_thread_ptr);
454                 cmm_thread_ptr = NULL;
455                 cmm_free_pages(loaned_pages);
456         } else if (!disable && cmm_disabled) {
457                 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
458                 if (IS_ERR(cmm_thread_ptr))
459                         return PTR_ERR(cmm_thread_ptr);
460         }
461
462         cmm_disabled = disable;
463         return 0;
464 }
465
466 module_param_call(disable, cmm_set_disable, param_get_uint,
467                   &cmm_disabled, S_IRUGO | S_IWUSR);
468 MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
469                  "[Default=" __stringify(CMM_DISABLE) "]");
470
471 module_init(cmm_init);
472 module_exit(cmm_exit);