Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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/mount.h>
23#include <linux/pseudo_fs.h>
24#include <linux/magic.h>
25#include <linux/balloon_compaction.h>
26#include <asm/firmware.h>
27#include <asm/hvcall.h>
28#include <asm/mmu.h>
29#include <asm/pgalloc.h>
30#include <linux/uaccess.h>
31#include <linux/memory.h>
32#include <asm/plpar_wrappers.h>
33
34#include "pseries.h"
35
36#define CMM_DRIVER_VERSION "1.0.0"
37#define CMM_DEFAULT_DELAY 1
38#define CMM_HOTPLUG_DELAY 5
39#define CMM_DEBUG 0
40#define CMM_DISABLE 0
41#define CMM_OOM_KB 1024
42#define CMM_MIN_MEM_MB 256
43#define KB2PAGES(_p) ((_p)>>(PAGE_SHIFT-10))
44#define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
45
46#define CMM_MEM_HOTPLUG_PRI 1
47
48static unsigned int delay = CMM_DEFAULT_DELAY;
49static unsigned int hotplug_delay = CMM_HOTPLUG_DELAY;
50static unsigned int oom_kb = CMM_OOM_KB;
51static unsigned int cmm_debug = CMM_DEBUG;
52static unsigned int cmm_disabled = CMM_DISABLE;
53static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
54static bool __read_mostly simulate;
55static unsigned long simulate_loan_target_kb;
56static struct device cmm_dev;
57
58MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
59MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
60MODULE_LICENSE("GPL");
61MODULE_VERSION(CMM_DRIVER_VERSION);
62
63module_param_named(delay, delay, uint, 0644);
64MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
65 "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
66module_param_named(hotplug_delay, hotplug_delay, uint, 0644);
67MODULE_PARM_DESC(hotplug_delay, "Delay (in seconds) after memory hotplug remove "
68 "before loaning resumes. "
69 "[Default=" __stringify(CMM_HOTPLUG_DELAY) "]");
70module_param_named(oom_kb, oom_kb, uint, 0644);
71MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
72 "[Default=" __stringify(CMM_OOM_KB) "]");
73module_param_named(min_mem_mb, min_mem_mb, ulong, 0644);
74MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
75 "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
76module_param_named(debug, cmm_debug, uint, 0644);
77MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
78 "[Default=" __stringify(CMM_DEBUG) "]");
79module_param_named(simulate, simulate, bool, 0444);
80MODULE_PARM_DESC(simulate, "Enable simulation mode (no communication with hw).");
81
82#define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
83
84static atomic_long_t loaned_pages;
85static unsigned long loaned_pages_target;
86static unsigned long oom_freed_pages;
87
88static DEFINE_MUTEX(hotplug_mutex);
89static int hotplug_occurred; /* protected by the hotplug mutex */
90
91static struct task_struct *cmm_thread_ptr;
92static struct balloon_dev_info b_dev_info;
93
94static long plpar_page_set_loaned(struct page *page)
95{
96 const unsigned long vpa = page_to_phys(page);
97 unsigned long cmo_page_sz = cmo_get_page_size();
98 long rc = 0;
99 int i;
100
101 if (unlikely(simulate))
102 return 0;
103
104 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
105 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED, vpa + i, 0);
106
107 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
108 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE,
109 vpa + i - cmo_page_sz, 0);
110
111 return rc;
112}
113
114static long plpar_page_set_active(struct page *page)
115{
116 const unsigned long vpa = page_to_phys(page);
117 unsigned long cmo_page_sz = cmo_get_page_size();
118 long rc = 0;
119 int i;
120
121 if (unlikely(simulate))
122 return 0;
123
124 for (i = 0; !rc && i < PAGE_SIZE; i += cmo_page_sz)
125 rc = plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_ACTIVE, vpa + i, 0);
126
127 for (i -= cmo_page_sz; rc && i != 0; i -= cmo_page_sz)
128 plpar_hcall_norets(H_PAGE_INIT, H_PAGE_SET_LOANED,
129 vpa + i - cmo_page_sz, 0);
130
131 return rc;
132}
133
134/**
135 * cmm_alloc_pages - Allocate pages and mark them as loaned
136 * @nr: number of pages to allocate
137 *
138 * Return value:
139 * number of pages requested to be allocated which were not
140 **/
141static long cmm_alloc_pages(long nr)
142{
143 struct page *page;
144 long rc;
145
146 cmm_dbg("Begin request for %ld pages\n", nr);
147
148 while (nr) {
149 /* Exit if a hotplug operation is in progress or occurred */
150 if (mutex_trylock(&hotplug_mutex)) {
151 if (hotplug_occurred) {
152 mutex_unlock(&hotplug_mutex);
153 break;
154 }
155 mutex_unlock(&hotplug_mutex);
156 } else {
157 break;
158 }
159
160 page = balloon_page_alloc();
161 if (!page)
162 break;
163 rc = plpar_page_set_loaned(page);
164 if (rc) {
165 pr_err("%s: Can not set page to loaned. rc=%ld\n", __func__, rc);
166 __free_page(page);
167 break;
168 }
169
170 balloon_page_enqueue(&b_dev_info, page);
171 atomic_long_inc(&loaned_pages);
172 adjust_managed_page_count(page, -1);
173 nr--;
174 }
175
176 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
177 return nr;
178}
179
180/**
181 * cmm_free_pages - Free pages and mark them as active
182 * @nr: number of pages to free
183 *
184 * Return value:
185 * number of pages requested to be freed which were not
186 **/
187static long cmm_free_pages(long nr)
188{
189 struct page *page;
190
191 cmm_dbg("Begin free of %ld pages.\n", nr);
192 while (nr) {
193 page = balloon_page_dequeue(&b_dev_info);
194 if (!page)
195 break;
196 plpar_page_set_active(page);
197 adjust_managed_page_count(page, 1);
198 __free_page(page);
199 atomic_long_dec(&loaned_pages);
200 nr--;
201 }
202 cmm_dbg("End request with %ld pages unfulfilled\n", nr);
203 return nr;
204}
205
206/**
207 * cmm_oom_notify - OOM notifier
208 * @self: notifier block struct
209 * @dummy: not used
210 * @parm: returned - number of pages freed
211 *
212 * Return value:
213 * NOTIFY_OK
214 **/
215static int cmm_oom_notify(struct notifier_block *self,
216 unsigned long dummy, void *parm)
217{
218 unsigned long *freed = parm;
219 long nr = KB2PAGES(oom_kb);
220
221 cmm_dbg("OOM processing started\n");
222 nr = cmm_free_pages(nr);
223 loaned_pages_target = atomic_long_read(&loaned_pages);
224 *freed += KB2PAGES(oom_kb) - nr;
225 oom_freed_pages += KB2PAGES(oom_kb) - nr;
226 cmm_dbg("OOM processing complete\n");
227 return NOTIFY_OK;
228}
229
230/**
231 * cmm_get_mpp - Read memory performance parameters
232 *
233 * Makes hcall to query the current page loan request from the hypervisor.
234 *
235 * Return value:
236 * nothing
237 **/
238static void cmm_get_mpp(void)
239{
240 const long __loaned_pages = atomic_long_read(&loaned_pages);
241 const long total_pages = totalram_pages() + __loaned_pages;
242 int rc;
243 struct hvcall_mpp_data mpp_data;
244 signed long active_pages_target, page_loan_request, target;
245 signed long min_mem_pages = (min_mem_mb * 1024 * 1024) / PAGE_SIZE;
246
247 if (likely(!simulate)) {
248 rc = h_get_mpp(&mpp_data);
249 if (rc != H_SUCCESS)
250 return;
251 page_loan_request = div_s64((s64)mpp_data.loan_request,
252 PAGE_SIZE);
253 target = page_loan_request + __loaned_pages;
254 } else {
255 target = KB2PAGES(simulate_loan_target_kb);
256 page_loan_request = target - __loaned_pages;
257 }
258
259 if (target < 0 || total_pages < min_mem_pages)
260 target = 0;
261
262 if (target > oom_freed_pages)
263 target -= oom_freed_pages;
264 else
265 target = 0;
266
267 active_pages_target = total_pages - target;
268
269 if (min_mem_pages > active_pages_target)
270 target = total_pages - min_mem_pages;
271
272 if (target < 0)
273 target = 0;
274
275 loaned_pages_target = target;
276
277 cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
278 page_loan_request, __loaned_pages, loaned_pages_target,
279 oom_freed_pages, totalram_pages());
280}
281
282static struct notifier_block cmm_oom_nb = {
283 .notifier_call = cmm_oom_notify
284};
285
286/**
287 * cmm_thread - CMM task thread
288 * @dummy: not used
289 *
290 * Return value:
291 * 0
292 **/
293static int cmm_thread(void *dummy)
294{
295 unsigned long timeleft;
296 long __loaned_pages;
297
298 while (1) {
299 timeleft = msleep_interruptible(delay * 1000);
300
301 if (kthread_should_stop() || timeleft)
302 break;
303
304 if (mutex_trylock(&hotplug_mutex)) {
305 if (hotplug_occurred) {
306 hotplug_occurred = 0;
307 mutex_unlock(&hotplug_mutex);
308 cmm_dbg("Hotplug operation has occurred, "
309 "loaning activity suspended "
310 "for %d seconds.\n",
311 hotplug_delay);
312 timeleft = msleep_interruptible(hotplug_delay *
313 1000);
314 if (kthread_should_stop() || timeleft)
315 break;
316 continue;
317 }
318 mutex_unlock(&hotplug_mutex);
319 } else {
320 cmm_dbg("Hotplug operation in progress, activity "
321 "suspended\n");
322 continue;
323 }
324
325 cmm_get_mpp();
326
327 __loaned_pages = atomic_long_read(&loaned_pages);
328 if (loaned_pages_target > __loaned_pages) {
329 if (cmm_alloc_pages(loaned_pages_target - __loaned_pages))
330 loaned_pages_target = __loaned_pages;
331 } else if (loaned_pages_target < __loaned_pages)
332 cmm_free_pages(__loaned_pages - loaned_pages_target);
333 }
334 return 0;
335}
336
337#define CMM_SHOW(name, format, args...) \
338 static ssize_t show_##name(struct device *dev, \
339 struct device_attribute *attr, \
340 char *buf) \
341 { \
342 return sprintf(buf, format, ##args); \
343 } \
344 static DEVICE_ATTR(name, 0444, show_##name, NULL)
345
346CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(atomic_long_read(&loaned_pages)));
347CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
348
349static ssize_t show_oom_pages(struct device *dev,
350 struct device_attribute *attr, char *buf)
351{
352 return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
353}
354
355static ssize_t store_oom_pages(struct device *dev,
356 struct device_attribute *attr,
357 const char *buf, size_t count)
358{
359 unsigned long val = simple_strtoul (buf, NULL, 10);
360
361 if (!capable(CAP_SYS_ADMIN))
362 return -EPERM;
363 if (val != 0)
364 return -EBADMSG;
365
366 oom_freed_pages = 0;
367 return count;
368}
369
370static DEVICE_ATTR(oom_freed_kb, 0644,
371 show_oom_pages, store_oom_pages);
372
373static struct device_attribute *cmm_attrs[] = {
374 &dev_attr_loaned_kb,
375 &dev_attr_loaned_target_kb,
376 &dev_attr_oom_freed_kb,
377};
378
379static DEVICE_ULONG_ATTR(simulate_loan_target_kb, 0644,
380 simulate_loan_target_kb);
381
382static struct bus_type cmm_subsys = {
383 .name = "cmm",
384 .dev_name = "cmm",
385};
386
387static void cmm_release_device(struct device *dev)
388{
389}
390
391/**
392 * cmm_sysfs_register - Register with sysfs
393 *
394 * Return value:
395 * 0 on success / other on failure
396 **/
397static int cmm_sysfs_register(struct device *dev)
398{
399 int i, rc;
400
401 if ((rc = subsys_system_register(&cmm_subsys, NULL)))
402 return rc;
403
404 dev->id = 0;
405 dev->bus = &cmm_subsys;
406 dev->release = cmm_release_device;
407
408 if ((rc = device_register(dev)))
409 goto subsys_unregister;
410
411 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
412 if ((rc = device_create_file(dev, cmm_attrs[i])))
413 goto fail;
414 }
415
416 if (!simulate)
417 return 0;
418 rc = device_create_file(dev, &dev_attr_simulate_loan_target_kb.attr);
419 if (rc)
420 goto fail;
421 return 0;
422
423fail:
424 while (--i >= 0)
425 device_remove_file(dev, cmm_attrs[i]);
426 device_unregister(dev);
427subsys_unregister:
428 bus_unregister(&cmm_subsys);
429 return rc;
430}
431
432/**
433 * cmm_unregister_sysfs - Unregister from sysfs
434 *
435 **/
436static void cmm_unregister_sysfs(struct device *dev)
437{
438 int i;
439
440 for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
441 device_remove_file(dev, cmm_attrs[i]);
442 device_unregister(dev);
443 bus_unregister(&cmm_subsys);
444}
445
446/**
447 * cmm_reboot_notifier - Make sure pages are not still marked as "loaned"
448 *
449 **/
450static int cmm_reboot_notifier(struct notifier_block *nb,
451 unsigned long action, void *unused)
452{
453 if (action == SYS_RESTART) {
454 if (cmm_thread_ptr)
455 kthread_stop(cmm_thread_ptr);
456 cmm_thread_ptr = NULL;
457 cmm_free_pages(atomic_long_read(&loaned_pages));
458 }
459 return NOTIFY_DONE;
460}
461
462static struct notifier_block cmm_reboot_nb = {
463 .notifier_call = cmm_reboot_notifier,
464};
465
466/**
467 * cmm_memory_cb - Handle memory hotplug notifier calls
468 * @self: notifier block struct
469 * @action: action to take
470 * @arg: struct memory_notify data for handler
471 *
472 * Return value:
473 * NOTIFY_OK or notifier error based on subfunction return value
474 *
475 **/
476static int cmm_memory_cb(struct notifier_block *self,
477 unsigned long action, void *arg)
478{
479 int ret = 0;
480
481 switch (action) {
482 case MEM_GOING_OFFLINE:
483 mutex_lock(&hotplug_mutex);
484 hotplug_occurred = 1;
485 break;
486 case MEM_OFFLINE:
487 case MEM_CANCEL_OFFLINE:
488 mutex_unlock(&hotplug_mutex);
489 cmm_dbg("Memory offline operation complete.\n");
490 break;
491 case MEM_GOING_ONLINE:
492 case MEM_ONLINE:
493 case MEM_CANCEL_ONLINE:
494 break;
495 }
496
497 return notifier_from_errno(ret);
498}
499
500static struct notifier_block cmm_mem_nb = {
501 .notifier_call = cmm_memory_cb,
502 .priority = CMM_MEM_HOTPLUG_PRI
503};
504
505#ifdef CONFIG_BALLOON_COMPACTION
506static struct vfsmount *balloon_mnt;
507
508static int cmm_init_fs_context(struct fs_context *fc)
509{
510 return init_pseudo(fc, PPC_CMM_MAGIC) ? 0 : -ENOMEM;
511}
512
513static struct file_system_type balloon_fs = {
514 .name = "ppc-cmm",
515 .init_fs_context = cmm_init_fs_context,
516 .kill_sb = kill_anon_super,
517};
518
519static int cmm_migratepage(struct balloon_dev_info *b_dev_info,
520 struct page *newpage, struct page *page,
521 enum migrate_mode mode)
522{
523 unsigned long flags;
524
525 /*
526 * loan/"inflate" the newpage first.
527 *
528 * We might race against the cmm_thread who might discover after our
529 * loan request that another page is to be unloaned. However, once
530 * the cmm_thread runs again later, this error will automatically
531 * be corrected.
532 */
533 if (plpar_page_set_loaned(newpage)) {
534 /* Unlikely, but possible. Tell the caller not to retry now. */
535 pr_err_ratelimited("%s: Cannot set page to loaned.", __func__);
536 return -EBUSY;
537 }
538
539 /* balloon page list reference */
540 get_page(newpage);
541
542 /*
543 * When we migrate a page to a different zone, we have to fixup the
544 * count of both involved zones as we adjusted the managed page count
545 * when inflating.
546 */
547 if (page_zone(page) != page_zone(newpage)) {
548 adjust_managed_page_count(page, 1);
549 adjust_managed_page_count(newpage, -1);
550 }
551
552 spin_lock_irqsave(&b_dev_info->pages_lock, flags);
553 balloon_page_insert(b_dev_info, newpage);
554 balloon_page_delete(page);
555 b_dev_info->isolated_pages--;
556 spin_unlock_irqrestore(&b_dev_info->pages_lock, flags);
557
558 /*
559 * activate/"deflate" the old page. We ignore any errors just like the
560 * other callers.
561 */
562 plpar_page_set_active(page);
563
564 /* balloon page list reference */
565 put_page(page);
566
567 return MIGRATEPAGE_SUCCESS;
568}
569
570static int cmm_balloon_compaction_init(void)
571{
572 int rc;
573
574 balloon_devinfo_init(&b_dev_info);
575 b_dev_info.migratepage = cmm_migratepage;
576
577 balloon_mnt = kern_mount(&balloon_fs);
578 if (IS_ERR(balloon_mnt)) {
579 rc = PTR_ERR(balloon_mnt);
580 balloon_mnt = NULL;
581 return rc;
582 }
583
584 b_dev_info.inode = alloc_anon_inode(balloon_mnt->mnt_sb);
585 if (IS_ERR(b_dev_info.inode)) {
586 rc = PTR_ERR(b_dev_info.inode);
587 b_dev_info.inode = NULL;
588 kern_unmount(balloon_mnt);
589 balloon_mnt = NULL;
590 return rc;
591 }
592
593 b_dev_info.inode->i_mapping->a_ops = &balloon_aops;
594 return 0;
595}
596static void cmm_balloon_compaction_deinit(void)
597{
598 if (b_dev_info.inode)
599 iput(b_dev_info.inode);
600 b_dev_info.inode = NULL;
601 kern_unmount(balloon_mnt);
602 balloon_mnt = NULL;
603}
604#else /* CONFIG_BALLOON_COMPACTION */
605static int cmm_balloon_compaction_init(void)
606{
607 return 0;
608}
609
610static void cmm_balloon_compaction_deinit(void)
611{
612}
613#endif /* CONFIG_BALLOON_COMPACTION */
614
615/**
616 * cmm_init - Module initialization
617 *
618 * Return value:
619 * 0 on success / other on failure
620 **/
621static int cmm_init(void)
622{
623 int rc;
624
625 if (!firmware_has_feature(FW_FEATURE_CMO) && !simulate)
626 return -EOPNOTSUPP;
627
628 rc = cmm_balloon_compaction_init();
629 if (rc)
630 return rc;
631
632 rc = register_oom_notifier(&cmm_oom_nb);
633 if (rc < 0)
634 goto out_balloon_compaction;
635
636 if ((rc = register_reboot_notifier(&cmm_reboot_nb)))
637 goto out_oom_notifier;
638
639 if ((rc = cmm_sysfs_register(&cmm_dev)))
640 goto out_reboot_notifier;
641
642 rc = register_memory_notifier(&cmm_mem_nb);
643 if (rc)
644 goto out_unregister_notifier;
645
646 if (cmm_disabled)
647 return 0;
648
649 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
650 if (IS_ERR(cmm_thread_ptr)) {
651 rc = PTR_ERR(cmm_thread_ptr);
652 goto out_unregister_notifier;
653 }
654
655 return 0;
656out_unregister_notifier:
657 unregister_memory_notifier(&cmm_mem_nb);
658 cmm_unregister_sysfs(&cmm_dev);
659out_reboot_notifier:
660 unregister_reboot_notifier(&cmm_reboot_nb);
661out_oom_notifier:
662 unregister_oom_notifier(&cmm_oom_nb);
663out_balloon_compaction:
664 cmm_balloon_compaction_deinit();
665 return rc;
666}
667
668/**
669 * cmm_exit - Module exit
670 *
671 * Return value:
672 * nothing
673 **/
674static void cmm_exit(void)
675{
676 if (cmm_thread_ptr)
677 kthread_stop(cmm_thread_ptr);
678 unregister_oom_notifier(&cmm_oom_nb);
679 unregister_reboot_notifier(&cmm_reboot_nb);
680 unregister_memory_notifier(&cmm_mem_nb);
681 cmm_free_pages(atomic_long_read(&loaned_pages));
682 cmm_unregister_sysfs(&cmm_dev);
683 cmm_balloon_compaction_deinit();
684}
685
686/**
687 * cmm_set_disable - Disable/Enable CMM
688 *
689 * Return value:
690 * 0 on success / other on failure
691 **/
692static int cmm_set_disable(const char *val, const struct kernel_param *kp)
693{
694 int disable = simple_strtoul(val, NULL, 10);
695
696 if (disable != 0 && disable != 1)
697 return -EINVAL;
698
699 if (disable && !cmm_disabled) {
700 if (cmm_thread_ptr)
701 kthread_stop(cmm_thread_ptr);
702 cmm_thread_ptr = NULL;
703 cmm_free_pages(atomic_long_read(&loaned_pages));
704 } else if (!disable && cmm_disabled) {
705 cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
706 if (IS_ERR(cmm_thread_ptr))
707 return PTR_ERR(cmm_thread_ptr);
708 }
709
710 cmm_disabled = disable;
711 return 0;
712}
713
714module_param_call(disable, cmm_set_disable, param_get_uint,
715 &cmm_disabled, 0644);
716MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
717 "[Default=" __stringify(CMM_DISABLE) "]");
718
719module_init(cmm_init);
720module_exit(cmm_exit);