xref: /openbmc/linux/drivers/xen/balloon.c (revision b627b4ed)
1 /******************************************************************************
2  * balloon.c
3  *
4  * Xen balloon driver - enables returning/claiming memory to/from Xen.
5  *
6  * Copyright (c) 2003, B Dragovic
7  * Copyright (c) 2003-2004, M Williamson, K Fraser
8  * Copyright (c) 2005 Dan M. Smith, IBM Corporation
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public License version 2
12  * as published by the Free Software Foundation; or, when distributed
13  * separately from the Linux kernel or incorporated into other
14  * software packages, subject to the following license:
15  *
16  * Permission is hereby granted, free of charge, to any person obtaining a copy
17  * of this source file (the "Software"), to deal in the Software without
18  * restriction, including without limitation the rights to use, copy, modify,
19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20  * and to permit persons to whom the Software is furnished to do so, subject to
21  * the following conditions:
22  *
23  * The above copyright notice and this permission notice shall be included in
24  * all copies or substantial portions of the Software.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32  * IN THE SOFTWARE.
33  */
34 
35 #include <linux/kernel.h>
36 #include <linux/module.h>
37 #include <linux/sched.h>
38 #include <linux/errno.h>
39 #include <linux/mm.h>
40 #include <linux/bootmem.h>
41 #include <linux/pagemap.h>
42 #include <linux/highmem.h>
43 #include <linux/mutex.h>
44 #include <linux/list.h>
45 #include <linux/sysdev.h>
46 
47 #include <asm/page.h>
48 #include <asm/pgalloc.h>
49 #include <asm/pgtable.h>
50 #include <asm/uaccess.h>
51 #include <asm/tlb.h>
52 
53 #include <asm/xen/hypervisor.h>
54 #include <asm/xen/hypercall.h>
55 #include <xen/interface/xen.h>
56 #include <xen/interface/memory.h>
57 #include <xen/xenbus.h>
58 #include <xen/features.h>
59 #include <xen/page.h>
60 
61 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
62 
63 #define BALLOON_CLASS_NAME "xen_memory"
64 
65 struct balloon_stats {
66 	/* We aim for 'current allocation' == 'target allocation'. */
67 	unsigned long current_pages;
68 	unsigned long target_pages;
69 	/* We may hit the hard limit in Xen. If we do then we remember it. */
70 	unsigned long hard_limit;
71 	/*
72 	 * Drivers may alter the memory reservation independently, but they
73 	 * must inform the balloon driver so we avoid hitting the hard limit.
74 	 */
75 	unsigned long driver_pages;
76 	/* Number of pages in high- and low-memory balloons. */
77 	unsigned long balloon_low;
78 	unsigned long balloon_high;
79 };
80 
81 static DEFINE_MUTEX(balloon_mutex);
82 
83 static struct sys_device balloon_sysdev;
84 
85 static int register_balloon(struct sys_device *sysdev);
86 
87 /*
88  * Protects atomic reservation decrease/increase against concurrent increases.
89  * Also protects non-atomic updates of current_pages and driver_pages, and
90  * balloon lists.
91  */
92 static DEFINE_SPINLOCK(balloon_lock);
93 
94 static struct balloon_stats balloon_stats;
95 
96 /* We increase/decrease in batches which fit in a page */
97 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
98 
99 /* VM /proc information for memory */
100 extern unsigned long totalram_pages;
101 
102 #ifdef CONFIG_HIGHMEM
103 extern unsigned long totalhigh_pages;
104 #define inc_totalhigh_pages() (totalhigh_pages++)
105 #define dec_totalhigh_pages() (totalhigh_pages--)
106 #else
107 #define inc_totalhigh_pages() do {} while(0)
108 #define dec_totalhigh_pages() do {} while(0)
109 #endif
110 
111 /* List of ballooned pages, threaded through the mem_map array. */
112 static LIST_HEAD(ballooned_pages);
113 
114 /* Main work function, always executed in process context. */
115 static void balloon_process(struct work_struct *work);
116 static DECLARE_WORK(balloon_worker, balloon_process);
117 static struct timer_list balloon_timer;
118 
119 /* When ballooning out (allocating memory to return to Xen) we don't really
120    want the kernel to try too hard since that can trigger the oom killer. */
121 #define GFP_BALLOON \
122 	(GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
123 
124 static void scrub_page(struct page *page)
125 {
126 #ifdef CONFIG_XEN_SCRUB_PAGES
127 	clear_highpage(page);
128 #endif
129 }
130 
131 /* balloon_append: add the given page to the balloon. */
132 static void balloon_append(struct page *page)
133 {
134 	/* Lowmem is re-populated first, so highmem pages go at list tail. */
135 	if (PageHighMem(page)) {
136 		list_add_tail(&page->lru, &ballooned_pages);
137 		balloon_stats.balloon_high++;
138 		dec_totalhigh_pages();
139 	} else {
140 		list_add(&page->lru, &ballooned_pages);
141 		balloon_stats.balloon_low++;
142 	}
143 }
144 
145 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
146 static struct page *balloon_retrieve(void)
147 {
148 	struct page *page;
149 
150 	if (list_empty(&ballooned_pages))
151 		return NULL;
152 
153 	page = list_entry(ballooned_pages.next, struct page, lru);
154 	list_del(&page->lru);
155 
156 	if (PageHighMem(page)) {
157 		balloon_stats.balloon_high--;
158 		inc_totalhigh_pages();
159 	}
160 	else
161 		balloon_stats.balloon_low--;
162 
163 	return page;
164 }
165 
166 static struct page *balloon_first_page(void)
167 {
168 	if (list_empty(&ballooned_pages))
169 		return NULL;
170 	return list_entry(ballooned_pages.next, struct page, lru);
171 }
172 
173 static struct page *balloon_next_page(struct page *page)
174 {
175 	struct list_head *next = page->lru.next;
176 	if (next == &ballooned_pages)
177 		return NULL;
178 	return list_entry(next, struct page, lru);
179 }
180 
181 static void balloon_alarm(unsigned long unused)
182 {
183 	schedule_work(&balloon_worker);
184 }
185 
186 static unsigned long current_target(void)
187 {
188 	unsigned long target = min(balloon_stats.target_pages, balloon_stats.hard_limit);
189 
190 	target = min(target,
191 		     balloon_stats.current_pages +
192 		     balloon_stats.balloon_low +
193 		     balloon_stats.balloon_high);
194 
195 	return target;
196 }
197 
198 static int increase_reservation(unsigned long nr_pages)
199 {
200 	unsigned long  pfn, i, flags;
201 	struct page   *page;
202 	long           rc;
203 	struct xen_memory_reservation reservation = {
204 		.address_bits = 0,
205 		.extent_order = 0,
206 		.domid        = DOMID_SELF
207 	};
208 
209 	if (nr_pages > ARRAY_SIZE(frame_list))
210 		nr_pages = ARRAY_SIZE(frame_list);
211 
212 	spin_lock_irqsave(&balloon_lock, flags);
213 
214 	page = balloon_first_page();
215 	for (i = 0; i < nr_pages; i++) {
216 		BUG_ON(page == NULL);
217 		frame_list[i] = page_to_pfn(page);;
218 		page = balloon_next_page(page);
219 	}
220 
221 	set_xen_guest_handle(reservation.extent_start, frame_list);
222 	reservation.nr_extents = nr_pages;
223 	rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
224 	if (rc < nr_pages) {
225 		if (rc > 0) {
226 			int ret;
227 
228 			/* We hit the Xen hard limit: reprobe. */
229 			reservation.nr_extents = rc;
230 			ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
231 						   &reservation);
232 			BUG_ON(ret != rc);
233 		}
234 		if (rc >= 0)
235 			balloon_stats.hard_limit = (balloon_stats.current_pages + rc -
236 						    balloon_stats.driver_pages);
237 		goto out;
238 	}
239 
240 	for (i = 0; i < nr_pages; i++) {
241 		page = balloon_retrieve();
242 		BUG_ON(page == NULL);
243 
244 		pfn = page_to_pfn(page);
245 		BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
246 		       phys_to_machine_mapping_valid(pfn));
247 
248 		set_phys_to_machine(pfn, frame_list[i]);
249 
250 		/* Link back into the page tables if not highmem. */
251 		if (pfn < max_low_pfn) {
252 			int ret;
253 			ret = HYPERVISOR_update_va_mapping(
254 				(unsigned long)__va(pfn << PAGE_SHIFT),
255 				mfn_pte(frame_list[i], PAGE_KERNEL),
256 				0);
257 			BUG_ON(ret);
258 		}
259 
260 		/* Relinquish the page back to the allocator. */
261 		ClearPageReserved(page);
262 		init_page_count(page);
263 		__free_page(page);
264 	}
265 
266 	balloon_stats.current_pages += nr_pages;
267 	totalram_pages = balloon_stats.current_pages;
268 
269  out:
270 	spin_unlock_irqrestore(&balloon_lock, flags);
271 
272 	return 0;
273 }
274 
275 static int decrease_reservation(unsigned long nr_pages)
276 {
277 	unsigned long  pfn, i, flags;
278 	struct page   *page;
279 	int            need_sleep = 0;
280 	int ret;
281 	struct xen_memory_reservation reservation = {
282 		.address_bits = 0,
283 		.extent_order = 0,
284 		.domid        = DOMID_SELF
285 	};
286 
287 	if (nr_pages > ARRAY_SIZE(frame_list))
288 		nr_pages = ARRAY_SIZE(frame_list);
289 
290 	for (i = 0; i < nr_pages; i++) {
291 		if ((page = alloc_page(GFP_BALLOON)) == NULL) {
292 			nr_pages = i;
293 			need_sleep = 1;
294 			break;
295 		}
296 
297 		pfn = page_to_pfn(page);
298 		frame_list[i] = pfn_to_mfn(pfn);
299 
300 		scrub_page(page);
301 
302 		if (!PageHighMem(page)) {
303 			ret = HYPERVISOR_update_va_mapping(
304 				(unsigned long)__va(pfn << PAGE_SHIFT),
305 				__pte_ma(0), 0);
306 			BUG_ON(ret);
307                 }
308 
309 	}
310 
311 	/* Ensure that ballooned highmem pages don't have kmaps. */
312 	kmap_flush_unused();
313 	flush_tlb_all();
314 
315 	spin_lock_irqsave(&balloon_lock, flags);
316 
317 	/* No more mappings: invalidate P2M and add to balloon. */
318 	for (i = 0; i < nr_pages; i++) {
319 		pfn = mfn_to_pfn(frame_list[i]);
320 		set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
321 		balloon_append(pfn_to_page(pfn));
322 	}
323 
324 	set_xen_guest_handle(reservation.extent_start, frame_list);
325 	reservation.nr_extents   = nr_pages;
326 	ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
327 	BUG_ON(ret != nr_pages);
328 
329 	balloon_stats.current_pages -= nr_pages;
330 	totalram_pages = balloon_stats.current_pages;
331 
332 	spin_unlock_irqrestore(&balloon_lock, flags);
333 
334 	return need_sleep;
335 }
336 
337 /*
338  * We avoid multiple worker processes conflicting via the balloon mutex.
339  * We may of course race updates of the target counts (which are protected
340  * by the balloon lock), or with changes to the Xen hard limit, but we will
341  * recover from these in time.
342  */
343 static void balloon_process(struct work_struct *work)
344 {
345 	int need_sleep = 0;
346 	long credit;
347 
348 	mutex_lock(&balloon_mutex);
349 
350 	do {
351 		credit = current_target() - balloon_stats.current_pages;
352 		if (credit > 0)
353 			need_sleep = (increase_reservation(credit) != 0);
354 		if (credit < 0)
355 			need_sleep = (decrease_reservation(-credit) != 0);
356 
357 #ifndef CONFIG_PREEMPT
358 		if (need_resched())
359 			schedule();
360 #endif
361 	} while ((credit != 0) && !need_sleep);
362 
363 	/* Schedule more work if there is some still to be done. */
364 	if (current_target() != balloon_stats.current_pages)
365 		mod_timer(&balloon_timer, jiffies + HZ);
366 
367 	mutex_unlock(&balloon_mutex);
368 }
369 
370 /* Resets the Xen limit, sets new target, and kicks off processing. */
371 static void balloon_set_new_target(unsigned long target)
372 {
373 	/* No need for lock. Not read-modify-write updates. */
374 	balloon_stats.hard_limit   = ~0UL;
375 	balloon_stats.target_pages = target;
376 	schedule_work(&balloon_worker);
377 }
378 
379 static struct xenbus_watch target_watch =
380 {
381 	.node = "memory/target"
382 };
383 
384 /* React to a change in the target key */
385 static void watch_target(struct xenbus_watch *watch,
386 			 const char **vec, unsigned int len)
387 {
388 	unsigned long long new_target;
389 	int err;
390 
391 	err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
392 	if (err != 1) {
393 		/* This is ok (for domain0 at least) - so just return */
394 		return;
395 	}
396 
397 	/* The given memory/target value is in KiB, so it needs converting to
398 	 * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
399 	 */
400 	balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
401 }
402 
403 static int balloon_init_watcher(struct notifier_block *notifier,
404 				unsigned long event,
405 				void *data)
406 {
407 	int err;
408 
409 	err = register_xenbus_watch(&target_watch);
410 	if (err)
411 		printk(KERN_ERR "Failed to set balloon watcher\n");
412 
413 	return NOTIFY_DONE;
414 }
415 
416 static struct notifier_block xenstore_notifier;
417 
418 static int __init balloon_init(void)
419 {
420 	unsigned long pfn;
421 	struct page *page;
422 
423 	if (!xen_pv_domain())
424 		return -ENODEV;
425 
426 	pr_info("xen_balloon: Initialising balloon driver.\n");
427 
428 	balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
429 	totalram_pages   = balloon_stats.current_pages;
430 	balloon_stats.target_pages  = balloon_stats.current_pages;
431 	balloon_stats.balloon_low   = 0;
432 	balloon_stats.balloon_high  = 0;
433 	balloon_stats.driver_pages  = 0UL;
434 	balloon_stats.hard_limit    = ~0UL;
435 
436 	init_timer(&balloon_timer);
437 	balloon_timer.data = 0;
438 	balloon_timer.function = balloon_alarm;
439 
440 	register_balloon(&balloon_sysdev);
441 
442 	/* Initialise the balloon with excess memory space. */
443 	for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
444 		page = pfn_to_page(pfn);
445 		if (!PageReserved(page))
446 			balloon_append(page);
447 	}
448 
449 	target_watch.callback = watch_target;
450 	xenstore_notifier.notifier_call = balloon_init_watcher;
451 
452 	register_xenstore_notifier(&xenstore_notifier);
453 
454 	return 0;
455 }
456 
457 subsys_initcall(balloon_init);
458 
459 static void balloon_exit(void)
460 {
461     /* XXX - release balloon here */
462     return;
463 }
464 
465 module_exit(balloon_exit);
466 
467 #define BALLOON_SHOW(name, format, args...)				\
468 	static ssize_t show_##name(struct sys_device *dev,		\
469 				   struct sysdev_attribute *attr,	\
470 				   char *buf)				\
471 	{								\
472 		return sprintf(buf, format, ##args);			\
473 	}								\
474 	static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
475 
476 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
477 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
478 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
479 BALLOON_SHOW(hard_limit_kb,
480 	     (balloon_stats.hard_limit!=~0UL) ? "%lu\n" : "???\n",
481 	     (balloon_stats.hard_limit!=~0UL) ? PAGES2KB(balloon_stats.hard_limit) : 0);
482 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
483 
484 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
485 			      char *buf)
486 {
487 	return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
488 }
489 
490 static ssize_t store_target_kb(struct sys_device *dev,
491 			       struct sysdev_attribute *attr,
492 			       const char *buf,
493 			       size_t count)
494 {
495 	char *endchar;
496 	unsigned long long target_bytes;
497 
498 	if (!capable(CAP_SYS_ADMIN))
499 		return -EPERM;
500 
501 	target_bytes = simple_strtoull(buf, &endchar, 0) * 1024;
502 
503 	balloon_set_new_target(target_bytes >> PAGE_SHIFT);
504 
505 	return count;
506 }
507 
508 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
509 		   show_target_kb, store_target_kb);
510 
511 
512 static ssize_t show_target(struct sys_device *dev, struct sysdev_attribute *attr,
513 			      char *buf)
514 {
515 	return sprintf(buf, "%llu\n",
516 		       (u64)balloon_stats.target_pages << PAGE_SHIFT);
517 }
518 
519 static ssize_t store_target(struct sys_device *dev,
520 			    struct sysdev_attribute *attr,
521 			    const char *buf,
522 			    size_t count)
523 {
524 	char *endchar;
525 	unsigned long long target_bytes;
526 
527 	if (!capable(CAP_SYS_ADMIN))
528 		return -EPERM;
529 
530 	target_bytes = memparse(buf, &endchar);
531 
532 	balloon_set_new_target(target_bytes >> PAGE_SHIFT);
533 
534 	return count;
535 }
536 
537 static SYSDEV_ATTR(target, S_IRUGO | S_IWUSR,
538 		   show_target, store_target);
539 
540 
541 static struct sysdev_attribute *balloon_attrs[] = {
542 	&attr_target_kb,
543 	&attr_target,
544 };
545 
546 static struct attribute *balloon_info_attrs[] = {
547 	&attr_current_kb.attr,
548 	&attr_low_kb.attr,
549 	&attr_high_kb.attr,
550 	&attr_hard_limit_kb.attr,
551 	&attr_driver_kb.attr,
552 	NULL
553 };
554 
555 static struct attribute_group balloon_info_group = {
556 	.name = "info",
557 	.attrs = balloon_info_attrs,
558 };
559 
560 static struct sysdev_class balloon_sysdev_class = {
561 	.name = BALLOON_CLASS_NAME,
562 };
563 
564 static int register_balloon(struct sys_device *sysdev)
565 {
566 	int i, error;
567 
568 	error = sysdev_class_register(&balloon_sysdev_class);
569 	if (error)
570 		return error;
571 
572 	sysdev->id = 0;
573 	sysdev->cls = &balloon_sysdev_class;
574 
575 	error = sysdev_register(sysdev);
576 	if (error) {
577 		sysdev_class_unregister(&balloon_sysdev_class);
578 		return error;
579 	}
580 
581 	for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
582 		error = sysdev_create_file(sysdev, balloon_attrs[i]);
583 		if (error)
584 			goto fail;
585 	}
586 
587 	error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
588 	if (error)
589 		goto fail;
590 
591 	return 0;
592 
593  fail:
594 	while (--i >= 0)
595 		sysdev_remove_file(sysdev, balloon_attrs[i]);
596 	sysdev_unregister(sysdev);
597 	sysdev_class_unregister(&balloon_sysdev_class);
598 	return error;
599 }
600 
601 MODULE_LICENSE("GPL");
602