1 /*
2  * Virtio balloon implementation, inspired by Dor Laor and Marcelo
3  * Tosatti's implementations.
4  *
5  *  Copyright 2008 Rusty Russell IBM Corporation
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., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
20  */
21 
22 #include <linux/virtio.h>
23 #include <linux/virtio_balloon.h>
24 #include <linux/swap.h>
25 #include <linux/kthread.h>
26 #include <linux/freezer.h>
27 #include <linux/delay.h>
28 #include <linux/slab.h>
29 #include <linux/module.h>
30 #include <linux/balloon_compaction.h>
31 #include <linux/oom.h>
32 #include <linux/wait.h>
33 
34 /*
35  * Balloon device works in 4K page units.  So each page is pointed to by
36  * multiple balloon pages.  All memory counters in this driver are in balloon
37  * page units.
38  */
39 #define VIRTIO_BALLOON_PAGES_PER_PAGE (unsigned)(PAGE_SIZE >> VIRTIO_BALLOON_PFN_SHIFT)
40 #define VIRTIO_BALLOON_ARRAY_PFNS_MAX 256
41 #define OOM_VBALLOON_DEFAULT_PAGES 256
42 #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80
43 
44 static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES;
45 module_param(oom_pages, int, S_IRUSR | S_IWUSR);
46 MODULE_PARM_DESC(oom_pages, "pages to free on OOM");
47 
48 struct virtio_balloon {
49 	struct virtio_device *vdev;
50 	struct virtqueue *inflate_vq, *deflate_vq, *stats_vq;
51 
52 	/* Where the ballooning thread waits for config to change. */
53 	wait_queue_head_t config_change;
54 
55 	/* The thread servicing the balloon. */
56 	struct task_struct *thread;
57 
58 	/* Waiting for host to ack the pages we released. */
59 	wait_queue_head_t acked;
60 
61 	/* Number of balloon pages we've told the Host we're not using. */
62 	unsigned int num_pages;
63 	/*
64 	 * The pages we've told the Host we're not using are enqueued
65 	 * at vb_dev_info->pages list.
66 	 * Each page on this list adds VIRTIO_BALLOON_PAGES_PER_PAGE
67 	 * to num_pages above.
68 	 */
69 	struct balloon_dev_info vb_dev_info;
70 
71 	/* Synchronize access/update to this struct virtio_balloon elements */
72 	struct mutex balloon_lock;
73 
74 	/* The array of pfns we tell the Host about. */
75 	unsigned int num_pfns;
76 	u32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX];
77 
78 	/* Memory statistics */
79 	int need_stats_update;
80 	struct virtio_balloon_stat stats[VIRTIO_BALLOON_S_NR];
81 
82 	/* To register callback in oom notifier call chain */
83 	struct notifier_block nb;
84 };
85 
86 static struct virtio_device_id id_table[] = {
87 	{ VIRTIO_ID_BALLOON, VIRTIO_DEV_ANY_ID },
88 	{ 0 },
89 };
90 
91 static u32 page_to_balloon_pfn(struct page *page)
92 {
93 	unsigned long pfn = page_to_pfn(page);
94 
95 	BUILD_BUG_ON(PAGE_SHIFT < VIRTIO_BALLOON_PFN_SHIFT);
96 	/* Convert pfn from Linux page size to balloon page size. */
97 	return pfn * VIRTIO_BALLOON_PAGES_PER_PAGE;
98 }
99 
100 static struct page *balloon_pfn_to_page(u32 pfn)
101 {
102 	BUG_ON(pfn % VIRTIO_BALLOON_PAGES_PER_PAGE);
103 	return pfn_to_page(pfn / VIRTIO_BALLOON_PAGES_PER_PAGE);
104 }
105 
106 static void balloon_ack(struct virtqueue *vq)
107 {
108 	struct virtio_balloon *vb = vq->vdev->priv;
109 
110 	wake_up(&vb->acked);
111 }
112 
113 static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq)
114 {
115 	struct scatterlist sg;
116 	unsigned int len;
117 
118 	sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns);
119 
120 	/* We should always be able to add one buffer to an empty queue. */
121 	virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
122 	virtqueue_kick(vq);
123 
124 	/* When host has read buffer, this completes via balloon_ack */
125 	wait_event(vb->acked, virtqueue_get_buf(vq, &len));
126 }
127 
128 static void set_page_pfns(u32 pfns[], struct page *page)
129 {
130 	unsigned int i;
131 
132 	/* Set balloon pfns pointing at this page.
133 	 * Note that the first pfn points at start of the page. */
134 	for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++)
135 		pfns[i] = page_to_balloon_pfn(page) + i;
136 }
137 
138 static void fill_balloon(struct virtio_balloon *vb, size_t num)
139 {
140 	struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
141 
142 	/* We can only do one array worth at a time. */
143 	num = min(num, ARRAY_SIZE(vb->pfns));
144 
145 	mutex_lock(&vb->balloon_lock);
146 	for (vb->num_pfns = 0; vb->num_pfns < num;
147 	     vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
148 		struct page *page = balloon_page_enqueue(vb_dev_info);
149 
150 		if (!page) {
151 			dev_info_ratelimited(&vb->vdev->dev,
152 					     "Out of puff! Can't get %u pages\n",
153 					     VIRTIO_BALLOON_PAGES_PER_PAGE);
154 			/* Sleep for at least 1/5 of a second before retry. */
155 			msleep(200);
156 			break;
157 		}
158 		set_page_pfns(vb->pfns + vb->num_pfns, page);
159 		vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE;
160 		adjust_managed_page_count(page, -1);
161 	}
162 
163 	/* Did we get any? */
164 	if (vb->num_pfns != 0)
165 		tell_host(vb, vb->inflate_vq);
166 	mutex_unlock(&vb->balloon_lock);
167 }
168 
169 static void release_pages_by_pfn(const u32 pfns[], unsigned int num)
170 {
171 	unsigned int i;
172 
173 	/* Find pfns pointing at start of each page, get pages and free them. */
174 	for (i = 0; i < num; i += VIRTIO_BALLOON_PAGES_PER_PAGE) {
175 		struct page *page = balloon_pfn_to_page(pfns[i]);
176 		adjust_managed_page_count(page, 1);
177 		put_page(page); /* balloon reference */
178 	}
179 }
180 
181 static unsigned leak_balloon(struct virtio_balloon *vb, size_t num)
182 {
183 	unsigned num_freed_pages;
184 	struct page *page;
185 	struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
186 
187 	/* We can only do one array worth at a time. */
188 	num = min(num, ARRAY_SIZE(vb->pfns));
189 
190 	mutex_lock(&vb->balloon_lock);
191 	for (vb->num_pfns = 0; vb->num_pfns < num;
192 	     vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
193 		page = balloon_page_dequeue(vb_dev_info);
194 		if (!page)
195 			break;
196 		set_page_pfns(vb->pfns + vb->num_pfns, page);
197 		vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE;
198 	}
199 
200 	num_freed_pages = vb->num_pfns;
201 	/*
202 	 * Note that if
203 	 * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST);
204 	 * is true, we *have* to do it in this order
205 	 */
206 	if (vb->num_pfns != 0)
207 		tell_host(vb, vb->deflate_vq);
208 	mutex_unlock(&vb->balloon_lock);
209 	release_pages_by_pfn(vb->pfns, vb->num_pfns);
210 	return num_freed_pages;
211 }
212 
213 static inline void update_stat(struct virtio_balloon *vb, int idx,
214 			       u16 tag, u64 val)
215 {
216 	BUG_ON(idx >= VIRTIO_BALLOON_S_NR);
217 	vb->stats[idx].tag = tag;
218 	vb->stats[idx].val = val;
219 }
220 
221 #define pages_to_bytes(x) ((u64)(x) << PAGE_SHIFT)
222 
223 static void update_balloon_stats(struct virtio_balloon *vb)
224 {
225 	unsigned long events[NR_VM_EVENT_ITEMS];
226 	struct sysinfo i;
227 	int idx = 0;
228 
229 	all_vm_events(events);
230 	si_meminfo(&i);
231 
232 	update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_IN,
233 				pages_to_bytes(events[PSWPIN]));
234 	update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_OUT,
235 				pages_to_bytes(events[PSWPOUT]));
236 	update_stat(vb, idx++, VIRTIO_BALLOON_S_MAJFLT, events[PGMAJFAULT]);
237 	update_stat(vb, idx++, VIRTIO_BALLOON_S_MINFLT, events[PGFAULT]);
238 	update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMFREE,
239 				pages_to_bytes(i.freeram));
240 	update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMTOT,
241 				pages_to_bytes(i.totalram));
242 }
243 
244 /*
245  * While most virtqueues communicate guest-initiated requests to the hypervisor,
246  * the stats queue operates in reverse.  The driver initializes the virtqueue
247  * with a single buffer.  From that point forward, all conversations consist of
248  * a hypervisor request (a call to this function) which directs us to refill
249  * the virtqueue with a fresh stats buffer.  Since stats collection can sleep,
250  * we notify our kthread which does the actual work via stats_handle_request().
251  */
252 static void stats_request(struct virtqueue *vq)
253 {
254 	struct virtio_balloon *vb = vq->vdev->priv;
255 
256 	vb->need_stats_update = 1;
257 	wake_up(&vb->config_change);
258 }
259 
260 static void stats_handle_request(struct virtio_balloon *vb)
261 {
262 	struct virtqueue *vq;
263 	struct scatterlist sg;
264 	unsigned int len;
265 
266 	vb->need_stats_update = 0;
267 	update_balloon_stats(vb);
268 
269 	vq = vb->stats_vq;
270 	if (!virtqueue_get_buf(vq, &len))
271 		return;
272 	sg_init_one(&sg, vb->stats, sizeof(vb->stats));
273 	virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
274 	virtqueue_kick(vq);
275 }
276 
277 static void virtballoon_changed(struct virtio_device *vdev)
278 {
279 	struct virtio_balloon *vb = vdev->priv;
280 
281 	wake_up(&vb->config_change);
282 }
283 
284 static inline s64 towards_target(struct virtio_balloon *vb)
285 {
286 	__le32 v;
287 	s64 target;
288 
289 	virtio_cread(vb->vdev, struct virtio_balloon_config, num_pages, &v);
290 
291 	target = le32_to_cpu(v);
292 	return target - vb->num_pages;
293 }
294 
295 static void update_balloon_size(struct virtio_balloon *vb)
296 {
297 	__le32 actual = cpu_to_le32(vb->num_pages);
298 
299 	virtio_cwrite(vb->vdev, struct virtio_balloon_config, actual,
300 		      &actual);
301 }
302 
303 /*
304  * virtballoon_oom_notify - release pages when system is under severe
305  *			    memory pressure (called from out_of_memory())
306  * @self : notifier block struct
307  * @dummy: not used
308  * @parm : returned - number of freed pages
309  *
310  * The balancing of memory by use of the virtio balloon should not cause
311  * the termination of processes while there are pages in the balloon.
312  * If virtio balloon manages to release some memory, it will make the
313  * system return and retry the allocation that forced the OOM killer
314  * to run.
315  */
316 static int virtballoon_oom_notify(struct notifier_block *self,
317 				  unsigned long dummy, void *parm)
318 {
319 	struct virtio_balloon *vb;
320 	unsigned long *freed;
321 	unsigned num_freed_pages;
322 
323 	vb = container_of(self, struct virtio_balloon, nb);
324 	if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
325 		return NOTIFY_OK;
326 
327 	freed = parm;
328 	num_freed_pages = leak_balloon(vb, oom_pages);
329 	update_balloon_size(vb);
330 	*freed += num_freed_pages;
331 
332 	return NOTIFY_OK;
333 }
334 
335 static int balloon(void *_vballoon)
336 {
337 	struct virtio_balloon *vb = _vballoon;
338 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
339 
340 	set_freezable();
341 	while (!kthread_should_stop()) {
342 		s64 diff;
343 
344 		try_to_freeze();
345 
346 		add_wait_queue(&vb->config_change, &wait);
347 		for (;;) {
348 			if ((diff = towards_target(vb)) != 0 ||
349 			    vb->need_stats_update ||
350 			    kthread_should_stop() ||
351 			    freezing(current))
352 				break;
353 			wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
354 		}
355 		remove_wait_queue(&vb->config_change, &wait);
356 
357 		if (vb->need_stats_update)
358 			stats_handle_request(vb);
359 		if (diff > 0)
360 			fill_balloon(vb, diff);
361 		else if (diff < 0)
362 			leak_balloon(vb, -diff);
363 		update_balloon_size(vb);
364 
365 		/*
366 		 * For large balloon changes, we could spend a lot of time
367 		 * and always have work to do.  Be nice if preempt disabled.
368 		 */
369 		cond_resched();
370 	}
371 	return 0;
372 }
373 
374 static int init_vqs(struct virtio_balloon *vb)
375 {
376 	struct virtqueue *vqs[3];
377 	vq_callback_t *callbacks[] = { balloon_ack, balloon_ack, stats_request };
378 	const char *names[] = { "inflate", "deflate", "stats" };
379 	int err, nvqs;
380 
381 	/*
382 	 * We expect two virtqueues: inflate and deflate, and
383 	 * optionally stat.
384 	 */
385 	nvqs = virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ) ? 3 : 2;
386 	err = vb->vdev->config->find_vqs(vb->vdev, nvqs, vqs, callbacks, names);
387 	if (err)
388 		return err;
389 
390 	vb->inflate_vq = vqs[0];
391 	vb->deflate_vq = vqs[1];
392 	if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
393 		struct scatterlist sg;
394 		vb->stats_vq = vqs[2];
395 
396 		/*
397 		 * Prime this virtqueue with one buffer so the hypervisor can
398 		 * use it to signal us later (it can't be broken yet!).
399 		 */
400 		sg_init_one(&sg, vb->stats, sizeof vb->stats);
401 		if (virtqueue_add_outbuf(vb->stats_vq, &sg, 1, vb, GFP_KERNEL)
402 		    < 0)
403 			BUG();
404 		virtqueue_kick(vb->stats_vq);
405 	}
406 	return 0;
407 }
408 
409 #ifdef CONFIG_BALLOON_COMPACTION
410 /*
411  * virtballoon_migratepage - perform the balloon page migration on behalf of
412  *			     a compation thread.     (called under page lock)
413  * @vb_dev_info: the balloon device
414  * @newpage: page that will replace the isolated page after migration finishes.
415  * @page   : the isolated (old) page that is about to be migrated to newpage.
416  * @mode   : compaction mode -- not used for balloon page migration.
417  *
418  * After a ballooned page gets isolated by compaction procedures, this is the
419  * function that performs the page migration on behalf of a compaction thread
420  * The page migration for virtio balloon is done in a simple swap fashion which
421  * follows these two macro steps:
422  *  1) insert newpage into vb->pages list and update the host about it;
423  *  2) update the host about the old page removed from vb->pages list;
424  *
425  * This function preforms the balloon page migration task.
426  * Called through balloon_mapping->a_ops->migratepage
427  */
428 static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info,
429 		struct page *newpage, struct page *page, enum migrate_mode mode)
430 {
431 	struct virtio_balloon *vb = container_of(vb_dev_info,
432 			struct virtio_balloon, vb_dev_info);
433 	unsigned long flags;
434 
435 	/*
436 	 * In order to avoid lock contention while migrating pages concurrently
437 	 * to leak_balloon() or fill_balloon() we just give up the balloon_lock
438 	 * this turn, as it is easier to retry the page migration later.
439 	 * This also prevents fill_balloon() getting stuck into a mutex
440 	 * recursion in the case it ends up triggering memory compaction
441 	 * while it is attempting to inflate the ballon.
442 	 */
443 	if (!mutex_trylock(&vb->balloon_lock))
444 		return -EAGAIN;
445 
446 	get_page(newpage); /* balloon reference */
447 
448 	/* balloon's page migration 1st step  -- inflate "newpage" */
449 	spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
450 	balloon_page_insert(vb_dev_info, newpage);
451 	vb_dev_info->isolated_pages--;
452 	__count_vm_event(BALLOON_MIGRATE);
453 	spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
454 	vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
455 	set_page_pfns(vb->pfns, newpage);
456 	tell_host(vb, vb->inflate_vq);
457 
458 	/* balloon's page migration 2nd step -- deflate "page" */
459 	balloon_page_delete(page);
460 	vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
461 	set_page_pfns(vb->pfns, page);
462 	tell_host(vb, vb->deflate_vq);
463 
464 	mutex_unlock(&vb->balloon_lock);
465 
466 	put_page(page); /* balloon reference */
467 
468 	return MIGRATEPAGE_SUCCESS;
469 }
470 #endif /* CONFIG_BALLOON_COMPACTION */
471 
472 static int virtballoon_probe(struct virtio_device *vdev)
473 {
474 	struct virtio_balloon *vb;
475 	int err;
476 
477 	if (!vdev->config->get) {
478 		dev_err(&vdev->dev, "%s failure: config access disabled\n",
479 			__func__);
480 		return -EINVAL;
481 	}
482 
483 	vdev->priv = vb = kmalloc(sizeof(*vb), GFP_KERNEL);
484 	if (!vb) {
485 		err = -ENOMEM;
486 		goto out;
487 	}
488 
489 	vb->num_pages = 0;
490 	mutex_init(&vb->balloon_lock);
491 	init_waitqueue_head(&vb->config_change);
492 	init_waitqueue_head(&vb->acked);
493 	vb->vdev = vdev;
494 	vb->need_stats_update = 0;
495 
496 	balloon_devinfo_init(&vb->vb_dev_info);
497 #ifdef CONFIG_BALLOON_COMPACTION
498 	vb->vb_dev_info.migratepage = virtballoon_migratepage;
499 #endif
500 
501 	err = init_vqs(vb);
502 	if (err)
503 		goto out_free_vb;
504 
505 	vb->nb.notifier_call = virtballoon_oom_notify;
506 	vb->nb.priority = VIRTBALLOON_OOM_NOTIFY_PRIORITY;
507 	err = register_oom_notifier(&vb->nb);
508 	if (err < 0)
509 		goto out_oom_notify;
510 
511 	virtio_device_ready(vdev);
512 
513 	vb->thread = kthread_run(balloon, vb, "vballoon");
514 	if (IS_ERR(vb->thread)) {
515 		err = PTR_ERR(vb->thread);
516 		goto out_del_vqs;
517 	}
518 
519 	return 0;
520 
521 out_del_vqs:
522 	unregister_oom_notifier(&vb->nb);
523 out_oom_notify:
524 	vdev->config->del_vqs(vdev);
525 out_free_vb:
526 	kfree(vb);
527 out:
528 	return err;
529 }
530 
531 static void remove_common(struct virtio_balloon *vb)
532 {
533 	/* There might be pages left in the balloon: free them. */
534 	while (vb->num_pages)
535 		leak_balloon(vb, vb->num_pages);
536 	update_balloon_size(vb);
537 
538 	/* Now we reset the device so we can clean up the queues. */
539 	vb->vdev->config->reset(vb->vdev);
540 
541 	vb->vdev->config->del_vqs(vb->vdev);
542 }
543 
544 static void virtballoon_remove(struct virtio_device *vdev)
545 {
546 	struct virtio_balloon *vb = vdev->priv;
547 
548 	unregister_oom_notifier(&vb->nb);
549 	kthread_stop(vb->thread);
550 	remove_common(vb);
551 	kfree(vb);
552 }
553 
554 #ifdef CONFIG_PM_SLEEP
555 static int virtballoon_freeze(struct virtio_device *vdev)
556 {
557 	struct virtio_balloon *vb = vdev->priv;
558 
559 	/*
560 	 * The kthread is already frozen by the PM core before this
561 	 * function is called.
562 	 */
563 
564 	remove_common(vb);
565 	return 0;
566 }
567 
568 static int virtballoon_restore(struct virtio_device *vdev)
569 {
570 	struct virtio_balloon *vb = vdev->priv;
571 	int ret;
572 
573 	ret = init_vqs(vdev->priv);
574 	if (ret)
575 		return ret;
576 
577 	virtio_device_ready(vdev);
578 
579 	fill_balloon(vb, towards_target(vb));
580 	update_balloon_size(vb);
581 	return 0;
582 }
583 #endif
584 
585 static unsigned int features[] = {
586 	VIRTIO_BALLOON_F_MUST_TELL_HOST,
587 	VIRTIO_BALLOON_F_STATS_VQ,
588 	VIRTIO_BALLOON_F_DEFLATE_ON_OOM,
589 };
590 
591 static struct virtio_driver virtio_balloon_driver = {
592 	.feature_table = features,
593 	.feature_table_size = ARRAY_SIZE(features),
594 	.driver.name =	KBUILD_MODNAME,
595 	.driver.owner =	THIS_MODULE,
596 	.id_table =	id_table,
597 	.probe =	virtballoon_probe,
598 	.remove =	virtballoon_remove,
599 	.config_changed = virtballoon_changed,
600 #ifdef CONFIG_PM_SLEEP
601 	.freeze	=	virtballoon_freeze,
602 	.restore =	virtballoon_restore,
603 #endif
604 };
605 
606 module_virtio_driver(virtio_balloon_driver);
607 MODULE_DEVICE_TABLE(virtio, id_table);
608 MODULE_DESCRIPTION("Virtio balloon driver");
609 MODULE_LICENSE("GPL");
610