1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Collaborative memory management interface.
4 *
5 * Copyright (C) 2008 IBM Corporation
6 * Author(s): Brian King (brking@linux.vnet.ibm.com),
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 <linux/balloon_compaction.h>
23 #include <asm/firmware.h>
24 #include <asm/hvcall.h>
25 #include <asm/mmu.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 #define CMM_MEM_HOTPLUG_PRI 1
43
44 static unsigned int delay = CMM_DEFAULT_DELAY;
45 static unsigned int hotplug_delay = CMM_HOTPLUG_DELAY;
46 static unsigned int oom_kb = CMM_OOM_KB;
47 static unsigned int cmm_debug = CMM_DEBUG;
48 static unsigned int cmm_disabled = CMM_DISABLE;
49 static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
50 static bool __read_mostly simulate;
51 static unsigned long simulate_loan_target_kb;
52 static struct device cmm_dev;
53
54 MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
55 MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
56 MODULE_LICENSE("GPL");
57 MODULE_VERSION(CMM_DRIVER_VERSION);
58
59 module_param_named(delay, delay, uint, 0644);
60 MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
61 "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
62 module_param_named(hotplug_delay, hotplug_delay, uint, 0644);
63 MODULE_PARM_DESC(hotplug_delay, "Delay (in seconds) after memory hotplug remove "
64 "before loaning resumes. "
65 "[Default=" __stringify(CMM_HOTPLUG_DELAY) "]");
66 module_param_named(oom_kb, oom_kb, uint, 0644);
67 MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
68 "[Default=" __stringify(CMM_OOM_KB) "]");
69 module_param_named(min_mem_mb, min_mem_mb, ulong, 0644);
70 MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
71 "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
72 module_param_named(debug, cmm_debug, uint, 0644);
73 MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
74 "[Default=" __stringify(CMM_DEBUG) "]");
75 module_param_named(simulate, simulate, bool, 0444);
76 MODULE_PARM_DESC(simulate, "Enable simulation mode (no communication with hw).");
77
78 #define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
79
80 static atomic_long_t loaned_pages;
81 static unsigned long loaned_pages_target;
82 static unsigned long oom_freed_pages;
83
84 static DEFINE_MUTEX(hotplug_mutex);
85 static int hotplug_occurred; /* protected by the hotplug mutex */
86
87 static struct task_struct *cmm_thread_ptr;
88 static struct balloon_dev_info b_dev_info;
89
plpar_page_set_loaned(struct page * page)90 static long plpar_page_set_loaned(struct page *page)
91 {
92 const unsigned long vpa = page_to_phys(page);
93 unsigned long cmo_page_sz = cmo_get_page_size();
94 long rc = 0;
95 int i;
96
97 if (unlikely(simulate))
98 return 0;
99
100 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
101 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED, vpa + i, 0);
102
103 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
104 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE,
105 vpa + i - cmo_page_sz, 0);
106
107 return rc;
108 }
109
plpar_page_set_active(struct page * page)110 static long plpar_page_set_active(struct page *page)
111 {
112 const unsigned long vpa = page_to_phys(page);
113 unsigned long cmo_page_sz = cmo_get_page_size();
114 long rc = 0;
115 int i;
116
117 if (unlikely(simulate))
118 return 0;
119
120 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
121 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE, vpa + i, 0);
122
123 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
124 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED,
125 vpa + i - cmo_page_sz, 0);
126
127 return rc;
128 }
129
130 /**
131 * cmm_alloc_pages - Allocate pages and mark them as loaned
132 * @nr: number of pages to allocate
133 *
134 * Return value:
135 * number of pages requested to be allocated which were not
136 **/
cmm_alloc_pages(long nr)137 static long cmm_alloc_pages(long nr)
138 {
139 struct page *page;
140 long rc;
141
142 cmm_dbg("Begin request for %ld pages\n", nr);
143
144 while (nr) {
145 /* Exit if a hotplug operation is in progress or occurred */
146 if (mutex_trylock(&hotplug_mutex)) {
147 if (hotplug_occurred) {
148 mutex_unlock(&hotplug_mutex);
149 break;
150 }
151 mutex_unlock(&hotplug_mutex);
152 } else {
153 break;
154 }
155
156 page = balloon_page_alloc();
157 if (!page)
158 break;
159 rc = plpar_page_set_loaned(page);
160 if (rc) {
161 pr_err("%s: Can not set page to loaned. rc=%ld\n", __func__, rc);
162 __free_page(page);
163 break;
164 }
165
166 balloon_page_enqueue(&b_dev_info, page);
167 atomic_long_inc(&loaned_pages);
168 adjust_managed_page_count(page, -1);
169 nr--;
170 }
171
172 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
173 return nr;
174 }
175
176 /**
177 * cmm_free_pages - Free pages and mark them as active
178 * @nr: number of pages to free
179 *
180 * Return value:
181 * number of pages requested to be freed which were not
182 **/
cmm_free_pages(long nr)183 static long cmm_free_pages(long nr)
184 {
185 struct page *page;
186
187 cmm_dbg("Begin free of %ld pages.\n", nr);
188 while (nr) {
189 page = balloon_page_dequeue(&b_dev_info);
190 if (!page)
191 break;
192 plpar_page_set_active(page);
193 adjust_managed_page_count(page, 1);
194 __free_page(page);
195 atomic_long_dec(&loaned_pages);
196 nr--;
197 }
198 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
199 return nr;
200 }
201
202 /**
203 * cmm_oom_notify - OOM notifier
204 * @self: notifier block struct
205 * @dummy: not used
206 * @parm: returned - number of pages freed
207 *
208 * Return value:
209 * NOTIFY_OK
210 **/
cmm_oom_notify(struct notifier_block * self,unsigned long dummy,void * parm)211 static int cmm_oom_notify(struct notifier_block *self,
212 unsigned long dummy, void *parm)
213 {
214 unsigned long *freed = parm;
215 long nr = KB2PAGES(oom_kb);
216
217 cmm_dbg("OOM processing started\n");
218 nr = cmm_free_pages(nr);
219 loaned_pages_target = atomic_long_read(&loaned_pages);
220 *freed += KB2PAGES(oom_kb) - nr;
221 oom_freed_pages += KB2PAGES(oom_kb) - nr;
222 cmm_dbg("OOM processing complete\n");
223 return NOTIFY_OK;
224 }
225
226 /**
227 * cmm_get_mpp - Read memory performance parameters
228 *
229 * Makes hcall to query the current page loan request from the hypervisor.
230 *
231 * Return value:
232 * nothing
233 **/
cmm_get_mpp(void)234 static void cmm_get_mpp(void)
235 {
236 const long __loaned_pages = atomic_long_read(&loaned_pages);
237 const long total_pages = totalram_pages() + __loaned_pages;
238 int rc;
239 struct hvcall_mpp_data mpp_data;
240 signed long active_pages_target, page_loan_request, target;
241 signed long min_mem_pages = (min_mem_mb * 1024 * 1024) / PAGE_SIZE;
242
243 if (likely(!simulate)) {
244 rc = h_get_mpp(&mpp_data);
245 if (rc != H_SUCCESS)
246 return;
247 page_loan_request = div_s64((s64)mpp_data.loan_request,
248 PAGE_SIZE);
249 target = page_loan_request + __loaned_pages;
250 } else {
251 target = KB2PAGES(simulate_loan_target_kb);
252 page_loan_request = target - __loaned_pages;
253 }
254
255 if (target < 0 || total_pages < min_mem_pages)
256 target = 0;
257
258 if (target > oom_freed_pages)
259 target -= oom_freed_pages;
260 else
261 target = 0;
262
263 active_pages_target = total_pages - target;
264
265 if (min_mem_pages > active_pages_target)
266 target = total_pages - min_mem_pages;
267
268 if (target < 0)
269 target = 0;
270
271 loaned_pages_target = target;
272
273 cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
274 page_loan_request, __loaned_pages, loaned_pages_target,
275 oom_freed_pages, totalram_pages());
276 }
277
278 static struct notifier_block cmm_oom_nb = {
279 .notifier_call = cmm_oom_notify
280 };
281
282 /**
283 * cmm_thread - CMM task thread
284 * @dummy: not used
285 *
286 * Return value:
287 * 0
288 **/
cmm_thread(void * dummy)289 static int cmm_thread(void *dummy)
290 {
291 unsigned long timeleft;
292 long __loaned_pages;
293
294 while (1) {
295 timeleft = msleep_interruptible(delay * 1000);
296
297 if (kthread_should_stop() || timeleft)
298 break;
299
300 if (mutex_trylock(&hotplug_mutex)) {
301 if (hotplug_occurred) {
302 hotplug_occurred = 0;
303 mutex_unlock(&hotplug_mutex);
304 cmm_dbg("Hotplug operation has occurred, "
305 "loaning activity suspended "
306 "for %d seconds.\n",
307 hotplug_delay);
308 timeleft = msleep_interruptible(hotplug_delay *
309 1000);
310 if (kthread_should_stop() || timeleft)
311 break;
312 continue;
313 }
314 mutex_unlock(&hotplug_mutex);
315 } else {
316 cmm_dbg("Hotplug operation in progress, activity "
317 "suspended\n");
318 continue;
319 }
320
321 cmm_get_mpp();
322
323 __loaned_pages = atomic_long_read(&loaned_pages);
324 if (loaned_pages_target > __loaned_pages) {
325 if (cmm_alloc_pages(loaned_pages_target - __loaned_pages))
326 loaned_pages_target = __loaned_pages;
327 } else if (loaned_pages_target < __loaned_pages)
328 cmm_free_pages(__loaned_pages - loaned_pages_target);
329 }
330 return 0;
331 }
332
333 #define CMM_SHOW(name, format, args...) \
334 static ssize_t show_##name(struct device *dev, \
335 struct device_attribute *attr, \
336 char *buf) \
337 { \
338 return sprintf(buf, format, ##args); \
339 } \
340 static DEVICE_ATTR(name, 0444, show_##name, NULL)
341
342 CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(atomic_long_read(&loaned_pages)));
343 CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
344
show_oom_pages(struct device * dev,struct device_attribute * attr,char * buf)345 static ssize_t show_oom_pages(struct device *dev,
346 struct device_attribute *attr, char *buf)
347 {
348 return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
349 }
350
store_oom_pages(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)351 static ssize_t store_oom_pages(struct device *dev,
352 struct device_attribute *attr,
353 const char *buf, size_t count)
354 {
355 unsigned long val = simple_strtoul (buf, NULL, 10);
356
357 if (!capable(CAP_SYS_ADMIN))
358 return -EPERM;
359 if (val != 0)
360 return -EBADMSG;
361
362 oom_freed_pages = 0;
363 return count;
364 }
365
366 static DEVICE_ATTR(oom_freed_kb, 0644,
367 show_oom_pages, store_oom_pages);
368
369 static struct device_attribute *cmm_attrs[] = {
370 &dev_attr_loaned_kb,
371 &dev_attr_loaned_target_kb,
372 &dev_attr_oom_freed_kb,
373 };
374
375 static DEVICE_ULONG_ATTR(simulate_loan_target_kb, 0644,
376 simulate_loan_target_kb);
377
378 static struct bus_type cmm_subsys = {
379 .name = "cmm",
380 .dev_name = "cmm",
381 };
382
cmm_release_device(struct device * dev)383 static void cmm_release_device(struct device *dev)
384 {
385 }
386
387 /**
388 * cmm_sysfs_register - Register with sysfs
389 *
390 * Return value:
391 * 0 on success / other on failure
392 **/
cmm_sysfs_register(struct device * dev)393 static int cmm_sysfs_register(struct device *dev)
394 {
395 int i, rc;
396
397 if ((rc = subsys_system_register(&cmm_subsys, NULL)))
398 return rc;
399
400 dev->id = 0;
401 dev->bus = &cmm_subsys;
402 dev->release = cmm_release_device;
403
404 if ((rc = device_register(dev)))
405 goto subsys_unregister;
406
407 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
408 if ((rc = device_create_file(dev, cmm_attrs[i])))
409 goto fail;
410 }
411
412 if (!simulate)
413 return 0;
414 rc = device_create_file(dev, &dev_attr_simulate_loan_target_kb.attr);
415 if (rc)
416 goto fail;
417 return 0;
418
419 fail:
420 while (--i >= 0)
421 device_remove_file(dev, cmm_attrs[i]);
422 device_unregister(dev);
423 subsys_unregister:
424 bus_unregister(&cmm_subsys);
425 return rc;
426 }
427
428 /**
429 * cmm_unregister_sysfs - Unregister from sysfs
430 *
431 **/
cmm_unregister_sysfs(struct device * dev)432 static void cmm_unregister_sysfs(struct device *dev)
433 {
434 int i;
435
436 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
437 device_remove_file(dev, cmm_attrs[i]);
438 device_unregister(dev);
439 bus_unregister(&cmm_subsys);
440 }
441
442 /**
443 * cmm_reboot_notifier - Make sure pages are not still marked as "loaned"
444 *
445 **/
cmm_reboot_notifier(struct notifier_block * nb,unsigned long action,void * unused)446 static int cmm_reboot_notifier(struct notifier_block *nb,
447 unsigned long action, void *unused)
448 {
449 if (action == SYS_RESTART) {
450 if (cmm_thread_ptr)
451 kthread_stop(cmm_thread_ptr);
452 cmm_thread_ptr = NULL;
453 cmm_free_pages(atomic_long_read(&loaned_pages));
454 }
455 return NOTIFY_DONE;
456 }
457
458 static struct notifier_block cmm_reboot_nb = {
459 .notifier_call = cmm_reboot_notifier,
460 };
461
462 /**
463 * cmm_memory_cb - Handle memory hotplug notifier calls
464 * @self: notifier block struct
465 * @action: action to take
466 * @arg: struct memory_notify data for handler
467 *
468 * Return value:
469 * NOTIFY_OK or notifier error based on subfunction return value
470 *
471 **/
cmm_memory_cb(struct notifier_block * self,unsigned long action,void * arg)472 static int cmm_memory_cb(struct notifier_block *self,
473 unsigned long action, void *arg)
474 {
475 switch (action) {
476 case MEM_GOING_OFFLINE:
477 mutex_lock(&hotplug_mutex);
478 hotplug_occurred = 1;
479 break;
480 case MEM_OFFLINE:
481 case MEM_CANCEL_OFFLINE:
482 mutex_unlock(&hotplug_mutex);
483 cmm_dbg("Memory offline operation complete.\n");
484 break;
485 case MEM_GOING_ONLINE:
486 case MEM_ONLINE:
487 case MEM_CANCEL_ONLINE:
488 break;
489 }
490
491 return NOTIFY_OK;
492 }
493
494 static struct notifier_block cmm_mem_nb = {
495 .notifier_call = cmm_memory_cb,
496 .priority = CMM_MEM_HOTPLUG_PRI
497 };
498
499 #ifdef CONFIG_BALLOON_COMPACTION
cmm_migratepage(struct balloon_dev_info * b_dev_info,struct page * newpage,struct page * page,enum migrate_mode mode)500 static int cmm_migratepage(struct balloon_dev_info *b_dev_info,
501 struct page *newpage, struct page *page,
502 enum migrate_mode mode)
503 {
504 unsigned long flags;
505
506 /*
507 * loan/"inflate" the newpage first.
508 *
509 * We might race against the cmm_thread who might discover after our
510 * loan request that another page is to be unloaned. However, once
511 * the cmm_thread runs again later, this error will automatically
512 * be corrected.
513 */
514 if (plpar_page_set_loaned(newpage)) {
515 /* Unlikely, but possible. Tell the caller not to retry now. */
516 pr_err_ratelimited("%s: Cannot set page to loaned.", __func__);
517 return -EBUSY;
518 }
519
520 /* balloon page list reference */
521 get_page(newpage);
522
523 /*
524 * When we migrate a page to a different zone, we have to fixup the
525 * count of both involved zones as we adjusted the managed page count
526 * when inflating.
527 */
528 if (page_zone(page) != page_zone(newpage)) {
529 adjust_managed_page_count(page, 1);
530 adjust_managed_page_count(newpage, -1);
531 }
532
533 spin_lock_irqsave(&b_dev_info->pages_lock, flags);
534 balloon_page_insert(b_dev_info, newpage);
535 balloon_page_delete(page);
536 b_dev_info->isolated_pages--;
537 spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
538
539 /*
540 * activate/"deflate" the old page. We ignore any errors just like the
541 * other callers.
542 */
543 plpar_page_set_active(page);
544
545 /* balloon page list reference */
546 put_page(page);
547
548 return MIGRATEPAGE_SUCCESS;
549 }
550
cmm_balloon_compaction_init(void)551 static void cmm_balloon_compaction_init(void)
552 {
553 balloon_devinfo_init(&b_dev_info);
554 b_dev_info.migratepage = cmm_migratepage;
555 }
556 #else /* CONFIG_BALLOON_COMPACTION */
cmm_balloon_compaction_init(void)557 static void cmm_balloon_compaction_init(void)
558 {
559 }
560 #endif /* CONFIG_BALLOON_COMPACTION */
561
562 /**
563 * cmm_init - Module initialization
564 *
565 * Return value:
566 * 0 on success / other on failure
567 **/
cmm_init(void)568 static int cmm_init(void)
569 {
570 int rc;
571
572 if (!firmware_has_feature(FW_FEATURE_CMO) && !simulate)
573 return -EOPNOTSUPP;
574
575 cmm_balloon_compaction_init();
576
577 rc = register_oom_notifier(&cmm_oom_nb);
578 if (rc < 0)
579 goto out_balloon_compaction;
580
581 if ((rc = register_reboot_notifier(&cmm_reboot_nb)))
582 goto out_oom_notifier;
583
584 if ((rc = cmm_sysfs_register(&cmm_dev)))
585 goto out_reboot_notifier;
586
587 rc = register_memory_notifier(&cmm_mem_nb);
588 if (rc)
589 goto out_unregister_notifier;
590
591 if (cmm_disabled)
592 return 0;
593
594 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
595 if (IS_ERR(cmm_thread_ptr)) {
596 rc = PTR_ERR(cmm_thread_ptr);
597 goto out_unregister_notifier;
598 }
599
600 return 0;
601 out_unregister_notifier:
602 unregister_memory_notifier(&cmm_mem_nb);
603 cmm_unregister_sysfs(&cmm_dev);
604 out_reboot_notifier:
605 unregister_reboot_notifier(&cmm_reboot_nb);
606 out_oom_notifier:
607 unregister_oom_notifier(&cmm_oom_nb);
608 out_balloon_compaction:
609 return rc;
610 }
611
612 /**
613 * cmm_exit - Module exit
614 *
615 * Return value:
616 * nothing
617 **/
cmm_exit(void)618 static void cmm_exit(void)
619 {
620 if (cmm_thread_ptr)
621 kthread_stop(cmm_thread_ptr);
622 unregister_oom_notifier(&cmm_oom_nb);
623 unregister_reboot_notifier(&cmm_reboot_nb);
624 unregister_memory_notifier(&cmm_mem_nb);
625 cmm_free_pages(atomic_long_read(&loaned_pages));
626 cmm_unregister_sysfs(&cmm_dev);
627 }
628
629 /**
630 * cmm_set_disable - Disable/Enable CMM
631 *
632 * Return value:
633 * 0 on success / other on failure
634 **/
cmm_set_disable(const char * val,const struct kernel_param * kp)635 static int cmm_set_disable(const char *val, const struct kernel_param *kp)
636 {
637 int disable = simple_strtoul(val, NULL, 10);
638
639 if (disable != 0 && disable != 1)
640 return -EINVAL;
641
642 if (disable && !cmm_disabled) {
643 if (cmm_thread_ptr)
644 kthread_stop(cmm_thread_ptr);
645 cmm_thread_ptr = NULL;
646 cmm_free_pages(atomic_long_read(&loaned_pages));
647 } else if (!disable && cmm_disabled) {
648 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
649 if (IS_ERR(cmm_thread_ptr))
650 return PTR_ERR(cmm_thread_ptr);
651 }
652
653 cmm_disabled = disable;
654 return 0;
655 }
656
657 module_param_call(disable, cmm_set_disable, param_get_uint,
658 &cmm_disabled, 0644);
659 MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
660 "[Default=" __stringify(CMM_DISABLE) "]");
661
662 module_init(cmm_init);
663 module_exit(cmm_exit);
664