This source file includes following definitions.
- plpar_page_set_loaned
- plpar_page_set_active
- cmm_alloc_pages
- cmm_free_pages
- cmm_oom_notify
- cmm_get_mpp
- cmm_thread
- show_oom_pages
- store_oom_pages
- cmm_release_device
- cmm_sysfs_register
- cmm_unregister_sysfs
- cmm_reboot_notifier
- cmm_count_pages
- cmm_memory_isolate_cb
- cmm_mem_going_offline
- cmm_memory_cb
- cmm_init
- cmm_exit
- cmm_set_disable
1
2
3
4
5
6
7
8
9 #include <linux/ctype.h>
10 #include <linux/delay.h>
11 #include <linux/errno.h>
12 #include <linux/fs.h>
13 #include <linux/gfp.h>
14 #include <linux/kthread.h>
15 #include <linux/module.h>
16 #include <linux/oom.h>
17 #include <linux/reboot.h>
18 #include <linux/sched.h>
19 #include <linux/stringify.h>
20 #include <linux/swap.h>
21 #include <linux/device.h>
22 #include <asm/firmware.h>
23 #include <asm/hvcall.h>
24 #include <asm/mmu.h>
25 #include <asm/pgalloc.h>
26 #include <linux/uaccess.h>
27 #include <linux/memory.h>
28 #include <asm/plpar_wrappers.h>
29
30 #include "pseries.h"
31
32 #define CMM_DRIVER_VERSION "1.0.0"
33 #define CMM_DEFAULT_DELAY 1
34 #define CMM_HOTPLUG_DELAY 5
35 #define CMM_DEBUG 0
36 #define CMM_DISABLE 0
37 #define CMM_OOM_KB 1024
38 #define CMM_MIN_MEM_MB 256
39 #define KB2PAGES(_p) ((_p)>>(PAGE_SHIFT-10))
40 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
41
42
43
44
45 #define CMM_MEM_HOTPLUG_PRI 1
46 #define CMM_MEM_ISOLATE_PRI 15
47
48 static unsigned int delay = CMM_DEFAULT_DELAY;
49 static unsigned int hotplug_delay = CMM_HOTPLUG_DELAY;
50 static unsigned int oom_kb = CMM_OOM_KB;
51 static unsigned int cmm_debug = CMM_DEBUG;
52 static unsigned int cmm_disabled = CMM_DISABLE;
53 static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
54 static struct device cmm_dev;
55
56 MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
57 MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
58 MODULE_LICENSE("GPL");
59 MODULE_VERSION(CMM_DRIVER_VERSION);
60
61 module_param_named(delay, delay, uint, 0644);
62 MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
63 "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
64 module_param_named(hotplug_delay, hotplug_delay, uint, 0644);
65 MODULE_PARM_DESC(hotplug_delay, "Delay (in seconds) after memory hotplug remove "
66 "before loaning resumes. "
67 "[Default=" __stringify(CMM_HOTPLUG_DELAY) "]");
68 module_param_named(oom_kb, oom_kb, uint, 0644);
69 MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
70 "[Default=" __stringify(CMM_OOM_KB) "]");
71 module_param_named(min_mem_mb, min_mem_mb, ulong, 0644);
72 MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
73 "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
74 module_param_named(debug, cmm_debug, uint, 0644);
75 MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
76 "[Default=" __stringify(CMM_DEBUG) "]");
77
78 #define CMM_NR_PAGES ((PAGE_SIZE - sizeof(void *) - sizeof(unsigned long)) / sizeof(unsigned long))
79
80 #define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
81
82 struct cmm_page_array {
83 struct cmm_page_array *next;
84 unsigned long index;
85 unsigned long page[CMM_NR_PAGES];
86 };
87
88 static unsigned long loaned_pages;
89 static unsigned long loaned_pages_target;
90 static unsigned long oom_freed_pages;
91
92 static struct cmm_page_array *cmm_page_list;
93 static DEFINE_SPINLOCK(cmm_lock);
94
95 static DEFINE_MUTEX(hotplug_mutex);
96 static int hotplug_occurred;
97
98 static struct task_struct *cmm_thread_ptr;
99
100 static long plpar_page_set_loaned(unsigned long vpa)
101 {
102 unsigned long cmo_page_sz = cmo_get_page_size();
103 long rc = 0;
104 int i;
105
106 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
107 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED, vpa + i, 0);
108
109 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
110 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE,
111 vpa + i - cmo_page_sz, 0);
112
113 return rc;
114 }
115
116 static long plpar_page_set_active(unsigned long vpa)
117 {
118 unsigned long cmo_page_sz = cmo_get_page_size();
119 long rc = 0;
120 int i;
121
122 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
123 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE, vpa + i, 0);
124
125 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
126 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED,
127 vpa + i - cmo_page_sz, 0);
128
129 return rc;
130 }
131
132
133
134
135
136
137
138
139 static long cmm_alloc_pages(long nr)
140 {
141 struct cmm_page_array *pa, *npa;
142 unsigned long addr;
143 long rc;
144
145 cmm_dbg("Begin request for %ld pages\n", nr);
146
147 while (nr) {
148
149 if (mutex_trylock(&hotplug_mutex)) {
150 if (hotplug_occurred) {
151 mutex_unlock(&hotplug_mutex);
152 break;
153 }
154 mutex_unlock(&hotplug_mutex);
155 } else {
156 break;
157 }
158
159 addr = __get_free_page(GFP_NOIO | __GFP_NOWARN |
160 __GFP_NORETRY | __GFP_NOMEMALLOC);
161 if (!addr)
162 break;
163 spin_lock(&cmm_lock);
164 pa = cmm_page_list;
165 if (!pa || pa->index >= CMM_NR_PAGES) {
166
167 spin_unlock(&cmm_lock);
168 npa = (struct cmm_page_array *)__get_free_page(
169 GFP_NOIO | __GFP_NOWARN |
170 __GFP_NORETRY | __GFP_NOMEMALLOC);
171 if (!npa) {
172 pr_info("%s: Can not allocate new page list\n", __func__);
173 free_page(addr);
174 break;
175 }
176 spin_lock(&cmm_lock);
177 pa = cmm_page_list;
178
179 if (!pa || pa->index >= CMM_NR_PAGES) {
180 npa->next = pa;
181 npa->index = 0;
182 pa = npa;
183 cmm_page_list = pa;
184 } else
185 free_page((unsigned long) npa);
186 }
187
188 if ((rc = plpar_page_set_loaned(__pa(addr)))) {
189 pr_err("%s: Can not set page to loaned. rc=%ld\n", __func__, rc);
190 spin_unlock(&cmm_lock);
191 free_page(addr);
192 break;
193 }
194
195 pa->page[pa->index++] = addr;
196 loaned_pages++;
197 totalram_pages_dec();
198 spin_unlock(&cmm_lock);
199 nr--;
200 }
201
202 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
203 return nr;
204 }
205
206
207
208
209
210
211
212
213 static long cmm_free_pages(long nr)
214 {
215 struct cmm_page_array *pa;
216 unsigned long addr;
217
218 cmm_dbg("Begin free of %ld pages.\n", nr);
219 spin_lock(&cmm_lock);
220 pa = cmm_page_list;
221 while (nr) {
222 if (!pa || pa->index <= 0)
223 break;
224 addr = pa->page[--pa->index];
225
226 if (pa->index == 0) {
227 pa = pa->next;
228 free_page((unsigned long) cmm_page_list);
229 cmm_page_list = pa;
230 }
231
232 plpar_page_set_active(__pa(addr));
233 free_page(addr);
234 loaned_pages--;
235 nr--;
236 totalram_pages_inc();
237 }
238 spin_unlock(&cmm_lock);
239 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
240 return nr;
241 }
242
243
244
245
246
247
248
249
250
251
252 static int cmm_oom_notify(struct notifier_block *self,
253 unsigned long dummy, void *parm)
254 {
255 unsigned long *freed = parm;
256 long nr = KB2PAGES(oom_kb);
257
258 cmm_dbg("OOM processing started\n");
259 nr = cmm_free_pages(nr);
260 loaned_pages_target = loaned_pages;
261 *freed += KB2PAGES(oom_kb) - nr;
262 oom_freed_pages += KB2PAGES(oom_kb) - nr;
263 cmm_dbg("OOM processing complete\n");
264 return NOTIFY_OK;
265 }
266
267
268
269
270
271
272
273
274
275 static void cmm_get_mpp(void)
276 {
277 int rc;
278 struct hvcall_mpp_data mpp_data;
279 signed long active_pages_target, page_loan_request, target;
280 signed long total_pages = totalram_pages() + loaned_pages;
281 signed long min_mem_pages = (min_mem_mb * 1024 * 1024) / PAGE_SIZE;
282
283 rc = h_get_mpp(&mpp_data);
284
285 if (rc != H_SUCCESS)
286 return;
287
288 page_loan_request = div_s64((s64)mpp_data.loan_request, PAGE_SIZE);
289 target = page_loan_request + (signed long)loaned_pages;
290
291 if (target < 0 || total_pages < min_mem_pages)
292 target = 0;
293
294 if (target > oom_freed_pages)
295 target -= oom_freed_pages;
296 else
297 target = 0;
298
299 active_pages_target = total_pages - target;
300
301 if (min_mem_pages > active_pages_target)
302 target = total_pages - min_mem_pages;
303
304 if (target < 0)
305 target = 0;
306
307 loaned_pages_target = target;
308
309 cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
310 page_loan_request, loaned_pages, loaned_pages_target,
311 oom_freed_pages, totalram_pages());
312 }
313
314 static struct notifier_block cmm_oom_nb = {
315 .notifier_call = cmm_oom_notify
316 };
317
318
319
320
321
322
323
324
325 static int cmm_thread(void *dummy)
326 {
327 unsigned long timeleft;
328
329 while (1) {
330 timeleft = msleep_interruptible(delay * 1000);
331
332 if (kthread_should_stop() || timeleft)
333 break;
334
335 if (mutex_trylock(&hotplug_mutex)) {
336 if (hotplug_occurred) {
337 hotplug_occurred = 0;
338 mutex_unlock(&hotplug_mutex);
339 cmm_dbg("Hotplug operation has occurred, "
340 "loaning activity suspended "
341 "for %d seconds.\n",
342 hotplug_delay);
343 timeleft = msleep_interruptible(hotplug_delay *
344 1000);
345 if (kthread_should_stop() || timeleft)
346 break;
347 continue;
348 }
349 mutex_unlock(&hotplug_mutex);
350 } else {
351 cmm_dbg("Hotplug operation in progress, activity "
352 "suspended\n");
353 continue;
354 }
355
356 cmm_get_mpp();
357
358 if (loaned_pages_target > loaned_pages) {
359 if (cmm_alloc_pages(loaned_pages_target - loaned_pages))
360 loaned_pages_target = loaned_pages;
361 } else if (loaned_pages_target < loaned_pages)
362 cmm_free_pages(loaned_pages - loaned_pages_target);
363 }
364 return 0;
365 }
366
367 #define CMM_SHOW(name, format, args...) \
368 static ssize_t show_##name(struct device *dev, \
369 struct device_attribute *attr, \
370 char *buf) \
371 { \
372 return sprintf(buf, format, ##args); \
373 } \
374 static DEVICE_ATTR(name, 0444, show_##name, NULL)
375
376 CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(loaned_pages));
377 CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
378
379 static ssize_t show_oom_pages(struct device *dev,
380 struct device_attribute *attr, char *buf)
381 {
382 return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
383 }
384
385 static ssize_t store_oom_pages(struct device *dev,
386 struct device_attribute *attr,
387 const char *buf, size_t count)
388 {
389 unsigned long val = simple_strtoul (buf, NULL, 10);
390
391 if (!capable(CAP_SYS_ADMIN))
392 return -EPERM;
393 if (val != 0)
394 return -EBADMSG;
395
396 oom_freed_pages = 0;
397 return count;
398 }
399
400 static DEVICE_ATTR(oom_freed_kb, 0644,
401 show_oom_pages, store_oom_pages);
402
403 static struct device_attribute *cmm_attrs[] = {
404 &dev_attr_loaned_kb,
405 &dev_attr_loaned_target_kb,
406 &dev_attr_oom_freed_kb,
407 };
408
409 static struct bus_type cmm_subsys = {
410 .name = "cmm",
411 .dev_name = "cmm",
412 };
413
414 static void cmm_release_device(struct device *dev)
415 {
416 }
417
418
419
420
421
422
423
424 static int cmm_sysfs_register(struct device *dev)
425 {
426 int i, rc;
427
428 if ((rc = subsys_system_register(&cmm_subsys, NULL)))
429 return rc;
430
431 dev->id = 0;
432 dev->bus = &cmm_subsys;
433 dev->release = cmm_release_device;
434
435 if ((rc = device_register(dev)))
436 goto subsys_unregister;
437
438 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
439 if ((rc = device_create_file(dev, cmm_attrs[i])))
440 goto fail;
441 }
442
443 return 0;
444
445 fail:
446 while (--i >= 0)
447 device_remove_file(dev, cmm_attrs[i]);
448 device_unregister(dev);
449 subsys_unregister:
450 bus_unregister(&cmm_subsys);
451 return rc;
452 }
453
454
455
456
457
458 static void cmm_unregister_sysfs(struct device *dev)
459 {
460 int i;
461
462 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
463 device_remove_file(dev, cmm_attrs[i]);
464 device_unregister(dev);
465 bus_unregister(&cmm_subsys);
466 }
467
468
469
470
471
472 static int cmm_reboot_notifier(struct notifier_block *nb,
473 unsigned long action, void *unused)
474 {
475 if (action == SYS_RESTART) {
476 if (cmm_thread_ptr)
477 kthread_stop(cmm_thread_ptr);
478 cmm_thread_ptr = NULL;
479 cmm_free_pages(loaned_pages);
480 }
481 return NOTIFY_DONE;
482 }
483
484 static struct notifier_block cmm_reboot_nb = {
485 .notifier_call = cmm_reboot_notifier,
486 };
487
488
489
490
491
492
493
494
495
496 static unsigned long cmm_count_pages(void *arg)
497 {
498 struct memory_isolate_notify *marg = arg;
499 struct cmm_page_array *pa;
500 unsigned long start = (unsigned long)pfn_to_kaddr(marg->start_pfn);
501 unsigned long end = start + (marg->nr_pages << PAGE_SHIFT);
502 unsigned long idx;
503
504 spin_lock(&cmm_lock);
505 pa = cmm_page_list;
506 while (pa) {
507 if ((unsigned long)pa >= start && (unsigned long)pa < end)
508 marg->pages_found++;
509 for (idx = 0; idx < pa->index; idx++)
510 if (pa->page[idx] >= start && pa->page[idx] < end)
511 marg->pages_found++;
512 pa = pa->next;
513 }
514 spin_unlock(&cmm_lock);
515 return 0;
516 }
517
518
519
520
521
522
523
524
525
526
527 static int cmm_memory_isolate_cb(struct notifier_block *self,
528 unsigned long action, void *arg)
529 {
530 int ret = 0;
531
532 if (action == MEM_ISOLATE_COUNT)
533 ret = cmm_count_pages(arg);
534
535 return notifier_from_errno(ret);
536 }
537
538 static struct notifier_block cmm_mem_isolate_nb = {
539 .notifier_call = cmm_memory_isolate_cb,
540 .priority = CMM_MEM_ISOLATE_PRI
541 };
542
543
544
545
546
547
548
549
550 static int cmm_mem_going_offline(void *arg)
551 {
552 struct memory_notify *marg = arg;
553 unsigned long start_page = (unsigned long)pfn_to_kaddr(marg->start_pfn);
554 unsigned long end_page = start_page + (marg->nr_pages << PAGE_SHIFT);
555 struct cmm_page_array *pa_curr, *pa_last, *npa;
556 unsigned long idx;
557 unsigned long freed = 0;
558
559 cmm_dbg("Memory going offline, searching 0x%lx (%ld pages).\n",
560 start_page, marg->nr_pages);
561 spin_lock(&cmm_lock);
562
563
564 pa_last = pa_curr = cmm_page_list;
565 while (pa_curr) {
566 for (idx = (pa_curr->index - 1); (idx + 1) > 0; idx--) {
567 if ((pa_curr->page[idx] < start_page) ||
568 (pa_curr->page[idx] >= end_page))
569 continue;
570
571 plpar_page_set_active(__pa(pa_curr->page[idx]));
572 free_page(pa_curr->page[idx]);
573 freed++;
574 loaned_pages--;
575 totalram_pages_inc();
576 pa_curr->page[idx] = pa_last->page[--pa_last->index];
577 if (pa_last->index == 0) {
578 if (pa_curr == pa_last)
579 pa_curr = pa_last->next;
580 pa_last = pa_last->next;
581 free_page((unsigned long)cmm_page_list);
582 cmm_page_list = pa_last;
583 }
584 }
585 pa_curr = pa_curr->next;
586 }
587
588
589 pa_last = NULL;
590 pa_curr = cmm_page_list;
591 while (pa_curr) {
592 if (((unsigned long)pa_curr >= start_page) &&
593 ((unsigned long)pa_curr < end_page)) {
594 npa = (struct cmm_page_array *)__get_free_page(
595 GFP_NOIO | __GFP_NOWARN |
596 __GFP_NORETRY | __GFP_NOMEMALLOC);
597 if (!npa) {
598 spin_unlock(&cmm_lock);
599 cmm_dbg("Failed to allocate memory for list "
600 "management. Memory hotplug "
601 "failed.\n");
602 return -ENOMEM;
603 }
604 memcpy(npa, pa_curr, PAGE_SIZE);
605 if (pa_curr == cmm_page_list)
606 cmm_page_list = npa;
607 if (pa_last)
608 pa_last->next = npa;
609 free_page((unsigned long) pa_curr);
610 freed++;
611 pa_curr = npa;
612 }
613
614 pa_last = pa_curr;
615 pa_curr = pa_curr->next;
616 }
617
618 spin_unlock(&cmm_lock);
619 cmm_dbg("Released %ld pages in the search range.\n", freed);
620
621 return 0;
622 }
623
624
625
626
627
628
629
630
631
632
633
634 static int cmm_memory_cb(struct notifier_block *self,
635 unsigned long action, void *arg)
636 {
637 int ret = 0;
638
639 switch (action) {
640 case MEM_GOING_OFFLINE:
641 mutex_lock(&hotplug_mutex);
642 hotplug_occurred = 1;
643 ret = cmm_mem_going_offline(arg);
644 break;
645 case MEM_OFFLINE:
646 case MEM_CANCEL_OFFLINE:
647 mutex_unlock(&hotplug_mutex);
648 cmm_dbg("Memory offline operation complete.\n");
649 break;
650 case MEM_GOING_ONLINE:
651 case MEM_ONLINE:
652 case MEM_CANCEL_ONLINE:
653 break;
654 }
655
656 return notifier_from_errno(ret);
657 }
658
659 static struct notifier_block cmm_mem_nb = {
660 .notifier_call = cmm_memory_cb,
661 .priority = CMM_MEM_HOTPLUG_PRI
662 };
663
664
665
666
667
668
669
670 static int cmm_init(void)
671 {
672 int rc = -ENOMEM;
673
674 if (!firmware_has_feature(FW_FEATURE_CMO))
675 return -EOPNOTSUPP;
676
677 if ((rc = register_oom_notifier(&cmm_oom_nb)) < 0)
678 return rc;
679
680 if ((rc = register_reboot_notifier(&cmm_reboot_nb)))
681 goto out_oom_notifier;
682
683 if ((rc = cmm_sysfs_register(&cmm_dev)))
684 goto out_reboot_notifier;
685
686 if (register_memory_notifier(&cmm_mem_nb) ||
687 register_memory_isolate_notifier(&cmm_mem_isolate_nb))
688 goto out_unregister_notifier;
689
690 if (cmm_disabled)
691 return rc;
692
693 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
694 if (IS_ERR(cmm_thread_ptr)) {
695 rc = PTR_ERR(cmm_thread_ptr);
696 goto out_unregister_notifier;
697 }
698
699 return rc;
700
701 out_unregister_notifier:
702 unregister_memory_notifier(&cmm_mem_nb);
703 unregister_memory_isolate_notifier(&cmm_mem_isolate_nb);
704 cmm_unregister_sysfs(&cmm_dev);
705 out_reboot_notifier:
706 unregister_reboot_notifier(&cmm_reboot_nb);
707 out_oom_notifier:
708 unregister_oom_notifier(&cmm_oom_nb);
709 return rc;
710 }
711
712
713
714
715
716
717
718 static void cmm_exit(void)
719 {
720 if (cmm_thread_ptr)
721 kthread_stop(cmm_thread_ptr);
722 unregister_oom_notifier(&cmm_oom_nb);
723 unregister_reboot_notifier(&cmm_reboot_nb);
724 unregister_memory_notifier(&cmm_mem_nb);
725 unregister_memory_isolate_notifier(&cmm_mem_isolate_nb);
726 cmm_free_pages(loaned_pages);
727 cmm_unregister_sysfs(&cmm_dev);
728 }
729
730
731
732
733
734
735
736 static int cmm_set_disable(const char *val, const struct kernel_param *kp)
737 {
738 int disable = simple_strtoul(val, NULL, 10);
739
740 if (disable != 0 && disable != 1)
741 return -EINVAL;
742
743 if (disable && !cmm_disabled) {
744 if (cmm_thread_ptr)
745 kthread_stop(cmm_thread_ptr);
746 cmm_thread_ptr = NULL;
747 cmm_free_pages(loaned_pages);
748 } else if (!disable && cmm_disabled) {
749 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
750 if (IS_ERR(cmm_thread_ptr))
751 return PTR_ERR(cmm_thread_ptr);
752 }
753
754 cmm_disabled = disable;
755 return 0;
756 }
757
758 module_param_call(disable, cmm_set_disable, param_get_uint,
759 &cmm_disabled, 0644);
760 MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
761 "[Default=" __stringify(CMM_DISABLE) "]");
762
763 module_init(cmm_init);
764 module_exit(cmm_exit);