2 * Virtio balloon implementation, inspired by Dor Laor and Marcelo
3 * Tosatti's implementations.
5 * Copyright 2008 Rusty Russell IBM Corporation
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.
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.
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
22 #include <linux/virtio.h>
23 #include <linux/virtio_balloon.h>
24 #include <linux/swap.h>
25 #include <linux/workqueue.h>
26 #include <linux/delay.h>
27 #include <linux/slab.h>
28 #include <linux/module.h>
29 #include <linux/balloon_compaction.h>
30 #include <linux/oom.h>
31 #include <linux/wait.h>
33 #include <linux/mount.h>
36 * Balloon device works in 4K page units. So each page is pointed to by
37 * multiple balloon pages. All memory counters in this driver are in balloon
40 #define VIRTIO_BALLOON_PAGES_PER_PAGE (unsigned)(PAGE_SIZE >> VIRTIO_BALLOON_PFN_SHIFT)
41 #define VIRTIO_BALLOON_ARRAY_PFNS_MAX 256
42 #define OOM_VBALLOON_DEFAULT_PAGES 256
43 #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80
45 static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES;
46 module_param(oom_pages, int, S_IRUSR | S_IWUSR);
47 MODULE_PARM_DESC(oom_pages, "pages to free on OOM");
49 #ifdef CONFIG_BALLOON_COMPACTION
50 static struct vfsmount *balloon_mnt;
53 struct virtio_balloon {
54 struct virtio_device *vdev;
55 struct virtqueue *inflate_vq, *deflate_vq, *stats_vq;
57 /* The balloon servicing is delegated to a freezable workqueue. */
58 struct work_struct update_balloon_stats_work;
59 struct work_struct update_balloon_size_work;
61 /* Prevent updating balloon when it is being canceled. */
62 spinlock_t stop_update_lock;
65 /* Waiting for host to ack the pages we released. */
66 wait_queue_head_t acked;
68 /* Number of balloon pages we've told the Host we're not using. */
69 unsigned int num_pages;
71 * The pages we've told the Host we're not using are enqueued
72 * at vb_dev_info->pages list.
73 * Each page on this list adds VIRTIO_BALLOON_PAGES_PER_PAGE
76 struct balloon_dev_info vb_dev_info;
78 /* Synchronize access/update to this struct virtio_balloon elements */
79 struct mutex balloon_lock;
81 /* The array of pfns we tell the Host about. */
82 unsigned int num_pfns;
83 __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX];
85 /* Memory statistics */
86 struct virtio_balloon_stat stats[VIRTIO_BALLOON_S_NR];
88 /* To register callback in oom notifier call chain */
89 struct notifier_block nb;
92 static struct virtio_device_id id_table[] = {
93 { VIRTIO_ID_BALLOON, VIRTIO_DEV_ANY_ID },
97 static u32 page_to_balloon_pfn(struct page *page)
99 unsigned long pfn = page_to_pfn(page);
101 BUILD_BUG_ON(PAGE_SHIFT < VIRTIO_BALLOON_PFN_SHIFT);
102 /* Convert pfn from Linux page size to balloon page size. */
103 return pfn * VIRTIO_BALLOON_PAGES_PER_PAGE;
106 static struct page *balloon_pfn_to_page(u32 pfn)
108 BUG_ON(pfn % VIRTIO_BALLOON_PAGES_PER_PAGE);
109 return pfn_to_page(pfn / VIRTIO_BALLOON_PAGES_PER_PAGE);
112 static void balloon_ack(struct virtqueue *vq)
114 struct virtio_balloon *vb = vq->vdev->priv;
119 static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq)
121 struct scatterlist sg;
124 sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns);
126 /* We should always be able to add one buffer to an empty queue. */
127 virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
130 /* When host has read buffer, this completes via balloon_ack */
131 wait_event(vb->acked, virtqueue_get_buf(vq, &len));
135 static void set_page_pfns(struct virtio_balloon *vb,
136 __virtio32 pfns[], struct page *page)
140 /* Set balloon pfns pointing at this page.
141 * Note that the first pfn points at start of the page. */
142 for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++)
143 pfns[i] = cpu_to_virtio32(vb->vdev,
144 page_to_balloon_pfn(page) + i);
147 static unsigned fill_balloon(struct virtio_balloon *vb, size_t num)
149 struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
150 unsigned num_allocated_pages;
152 /* We can only do one array worth at a time. */
153 num = min(num, ARRAY_SIZE(vb->pfns));
155 mutex_lock(&vb->balloon_lock);
156 for (vb->num_pfns = 0; vb->num_pfns < num;
157 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
158 struct page *page = balloon_page_enqueue(vb_dev_info);
161 dev_info_ratelimited(&vb->vdev->dev,
162 "Out of puff! Can't get %u pages\n",
163 VIRTIO_BALLOON_PAGES_PER_PAGE);
164 /* Sleep for at least 1/5 of a second before retry. */
168 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
169 vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE;
170 if (!virtio_has_feature(vb->vdev,
171 VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
172 adjust_managed_page_count(page, -1);
175 num_allocated_pages = vb->num_pfns;
176 /* Did we get any? */
177 if (vb->num_pfns != 0)
178 tell_host(vb, vb->inflate_vq);
179 mutex_unlock(&vb->balloon_lock);
181 return num_allocated_pages;
184 static void release_pages_balloon(struct virtio_balloon *vb)
189 /* Find pfns pointing at start of each page, get pages and free them. */
190 for (i = 0; i < vb->num_pfns; i += VIRTIO_BALLOON_PAGES_PER_PAGE) {
191 page = balloon_pfn_to_page(virtio32_to_cpu(vb->vdev,
193 if (!virtio_has_feature(vb->vdev,
194 VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
195 adjust_managed_page_count(page, 1);
196 put_page(page); /* balloon reference */
200 static unsigned leak_balloon(struct virtio_balloon *vb, size_t num)
202 unsigned num_freed_pages;
204 struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
206 /* We can only do one array worth at a time. */
207 num = min(num, ARRAY_SIZE(vb->pfns));
209 mutex_lock(&vb->balloon_lock);
210 /* We can't release more pages than taken */
211 num = min(num, (size_t)vb->num_pages);
212 for (vb->num_pfns = 0; vb->num_pfns < num;
213 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
214 page = balloon_page_dequeue(vb_dev_info);
217 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
218 vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE;
221 num_freed_pages = vb->num_pfns;
224 * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST);
225 * is true, we *have* to do it in this order
227 if (vb->num_pfns != 0)
228 tell_host(vb, vb->deflate_vq);
229 release_pages_balloon(vb);
230 mutex_unlock(&vb->balloon_lock);
231 return num_freed_pages;
234 static inline void update_stat(struct virtio_balloon *vb, int idx,
237 BUG_ON(idx >= VIRTIO_BALLOON_S_NR);
238 vb->stats[idx].tag = cpu_to_virtio16(vb->vdev, tag);
239 vb->stats[idx].val = cpu_to_virtio64(vb->vdev, val);
242 #define pages_to_bytes(x) ((u64)(x) << PAGE_SHIFT)
244 static unsigned int update_balloon_stats(struct virtio_balloon *vb)
246 unsigned long events[NR_VM_EVENT_ITEMS];
248 unsigned int idx = 0;
251 all_vm_events(events);
254 available = si_mem_available();
256 #ifdef CONFIG_VM_EVENT_COUNTERS
257 update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_IN,
258 pages_to_bytes(events[PSWPIN]));
259 update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_OUT,
260 pages_to_bytes(events[PSWPOUT]));
261 update_stat(vb, idx++, VIRTIO_BALLOON_S_MAJFLT, events[PGMAJFAULT]);
262 update_stat(vb, idx++, VIRTIO_BALLOON_S_MINFLT, events[PGFAULT]);
264 update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMFREE,
265 pages_to_bytes(i.freeram));
266 update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMTOT,
267 pages_to_bytes(i.totalram));
268 update_stat(vb, idx++, VIRTIO_BALLOON_S_AVAIL,
269 pages_to_bytes(available));
275 * While most virtqueues communicate guest-initiated requests to the hypervisor,
276 * the stats queue operates in reverse. The driver initializes the virtqueue
277 * with a single buffer. From that point forward, all conversations consist of
278 * a hypervisor request (a call to this function) which directs us to refill
279 * the virtqueue with a fresh stats buffer. Since stats collection can sleep,
280 * we delegate the job to a freezable workqueue that will do the actual work via
281 * stats_handle_request().
283 static void stats_request(struct virtqueue *vq)
285 struct virtio_balloon *vb = vq->vdev->priv;
287 spin_lock(&vb->stop_update_lock);
288 if (!vb->stop_update)
289 queue_work(system_freezable_wq, &vb->update_balloon_stats_work);
290 spin_unlock(&vb->stop_update_lock);
293 static void stats_handle_request(struct virtio_balloon *vb)
295 struct virtqueue *vq;
296 struct scatterlist sg;
297 unsigned int len, num_stats;
299 num_stats = update_balloon_stats(vb);
302 if (!virtqueue_get_buf(vq, &len))
304 sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
305 virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
309 static void virtballoon_changed(struct virtio_device *vdev)
311 struct virtio_balloon *vb = vdev->priv;
314 spin_lock_irqsave(&vb->stop_update_lock, flags);
315 if (!vb->stop_update)
316 queue_work(system_freezable_wq, &vb->update_balloon_size_work);
317 spin_unlock_irqrestore(&vb->stop_update_lock, flags);
320 static inline s64 towards_target(struct virtio_balloon *vb)
325 virtio_cread(vb->vdev, struct virtio_balloon_config, num_pages,
328 /* Legacy balloon config space is LE, unlike all other devices. */
329 if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1))
330 num_pages = le32_to_cpu((__force __le32)num_pages);
333 return target - vb->num_pages;
336 static void update_balloon_size(struct virtio_balloon *vb)
338 u32 actual = vb->num_pages;
340 /* Legacy balloon config space is LE, unlike all other devices. */
341 if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1))
342 actual = (__force u32)cpu_to_le32(actual);
344 virtio_cwrite(vb->vdev, struct virtio_balloon_config, actual,
349 * virtballoon_oom_notify - release pages when system is under severe
350 * memory pressure (called from out_of_memory())
351 * @self : notifier block struct
353 * @parm : returned - number of freed pages
355 * The balancing of memory by use of the virtio balloon should not cause
356 * the termination of processes while there are pages in the balloon.
357 * If virtio balloon manages to release some memory, it will make the
358 * system return and retry the allocation that forced the OOM killer
361 static int virtballoon_oom_notify(struct notifier_block *self,
362 unsigned long dummy, void *parm)
364 struct virtio_balloon *vb;
365 unsigned long *freed;
366 unsigned num_freed_pages;
368 vb = container_of(self, struct virtio_balloon, nb);
369 if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
373 num_freed_pages = leak_balloon(vb, oom_pages);
374 update_balloon_size(vb);
375 *freed += num_freed_pages;
380 static void update_balloon_stats_func(struct work_struct *work)
382 struct virtio_balloon *vb;
384 vb = container_of(work, struct virtio_balloon,
385 update_balloon_stats_work);
386 stats_handle_request(vb);
389 static void update_balloon_size_func(struct work_struct *work)
391 struct virtio_balloon *vb;
394 vb = container_of(work, struct virtio_balloon,
395 update_balloon_size_work);
396 diff = towards_target(vb);
399 diff -= fill_balloon(vb, diff);
401 diff += leak_balloon(vb, -diff);
402 update_balloon_size(vb);
405 queue_work(system_freezable_wq, work);
408 static int init_vqs(struct virtio_balloon *vb)
410 struct virtqueue *vqs[3];
411 vq_callback_t *callbacks[] = { balloon_ack, balloon_ack, stats_request };
412 static const char * const names[] = { "inflate", "deflate", "stats" };
416 * We expect two virtqueues: inflate and deflate, and
419 nvqs = virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ) ? 3 : 2;
420 err = vb->vdev->config->find_vqs(vb->vdev, nvqs, vqs, callbacks, names);
424 vb->inflate_vq = vqs[0];
425 vb->deflate_vq = vqs[1];
426 if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
427 struct scatterlist sg;
428 unsigned int num_stats;
429 vb->stats_vq = vqs[2];
432 * Prime this virtqueue with one buffer so the hypervisor can
433 * use it to signal us later (it can't be broken yet!).
435 num_stats = update_balloon_stats(vb);
437 sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
438 if (virtqueue_add_outbuf(vb->stats_vq, &sg, 1, vb, GFP_KERNEL)
441 virtqueue_kick(vb->stats_vq);
446 #ifdef CONFIG_BALLOON_COMPACTION
448 * virtballoon_migratepage - perform the balloon page migration on behalf of
449 * a compation thread. (called under page lock)
450 * @vb_dev_info: the balloon device
451 * @newpage: page that will replace the isolated page after migration finishes.
452 * @page : the isolated (old) page that is about to be migrated to newpage.
453 * @mode : compaction mode -- not used for balloon page migration.
455 * After a ballooned page gets isolated by compaction procedures, this is the
456 * function that performs the page migration on behalf of a compaction thread
457 * The page migration for virtio balloon is done in a simple swap fashion which
458 * follows these two macro steps:
459 * 1) insert newpage into vb->pages list and update the host about it;
460 * 2) update the host about the old page removed from vb->pages list;
462 * This function preforms the balloon page migration task.
463 * Called through balloon_mapping->a_ops->migratepage
465 static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info,
466 struct page *newpage, struct page *page, enum migrate_mode mode)
468 struct virtio_balloon *vb = container_of(vb_dev_info,
469 struct virtio_balloon, vb_dev_info);
473 * In order to avoid lock contention while migrating pages concurrently
474 * to leak_balloon() or fill_balloon() we just give up the balloon_lock
475 * this turn, as it is easier to retry the page migration later.
476 * This also prevents fill_balloon() getting stuck into a mutex
477 * recursion in the case it ends up triggering memory compaction
478 * while it is attempting to inflate the ballon.
480 if (!mutex_trylock(&vb->balloon_lock))
483 get_page(newpage); /* balloon reference */
486 * When we migrate a page to a different zone and adjusted the
487 * managed page count when inflating, we have to fixup the count of
488 * both involved zones.
490 if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM) &&
491 page_zone(page) != page_zone(newpage)) {
492 adjust_managed_page_count(page, 1);
493 adjust_managed_page_count(newpage, -1);
496 /* balloon's page migration 1st step -- inflate "newpage" */
497 spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
498 balloon_page_insert(vb_dev_info, newpage);
499 vb_dev_info->isolated_pages--;
500 __count_vm_event(BALLOON_MIGRATE);
501 spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
502 vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
503 set_page_pfns(vb, vb->pfns, newpage);
504 tell_host(vb, vb->inflate_vq);
506 /* balloon's page migration 2nd step -- deflate "page" */
507 spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
508 balloon_page_delete(page);
509 spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
510 vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
511 set_page_pfns(vb, vb->pfns, page);
512 tell_host(vb, vb->deflate_vq);
514 mutex_unlock(&vb->balloon_lock);
516 put_page(page); /* balloon reference */
518 return MIGRATEPAGE_SUCCESS;
521 static struct dentry *balloon_mount(struct file_system_type *fs_type,
522 int flags, const char *dev_name, void *data)
524 static const struct dentry_operations ops = {
525 .d_dname = simple_dname,
528 return mount_pseudo(fs_type, "balloon-kvm:", NULL, &ops,
532 static struct file_system_type balloon_fs = {
533 .name = "balloon-kvm",
534 .mount = balloon_mount,
535 .kill_sb = kill_anon_super,
538 #endif /* CONFIG_BALLOON_COMPACTION */
540 static int virtballoon_probe(struct virtio_device *vdev)
542 struct virtio_balloon *vb;
545 if (!vdev->config->get) {
546 dev_err(&vdev->dev, "%s failure: config access disabled\n",
551 vdev->priv = vb = kmalloc(sizeof(*vb), GFP_KERNEL);
557 INIT_WORK(&vb->update_balloon_stats_work, update_balloon_stats_func);
558 INIT_WORK(&vb->update_balloon_size_work, update_balloon_size_func);
559 spin_lock_init(&vb->stop_update_lock);
560 vb->stop_update = false;
562 mutex_init(&vb->balloon_lock);
563 init_waitqueue_head(&vb->acked);
566 balloon_devinfo_init(&vb->vb_dev_info);
572 vb->nb.notifier_call = virtballoon_oom_notify;
573 vb->nb.priority = VIRTBALLOON_OOM_NOTIFY_PRIORITY;
574 err = register_oom_notifier(&vb->nb);
578 #ifdef CONFIG_BALLOON_COMPACTION
579 balloon_mnt = kern_mount(&balloon_fs);
580 if (IS_ERR(balloon_mnt)) {
581 err = PTR_ERR(balloon_mnt);
582 unregister_oom_notifier(&vb->nb);
586 vb->vb_dev_info.migratepage = virtballoon_migratepage;
587 vb->vb_dev_info.inode = alloc_anon_inode(balloon_mnt->mnt_sb);
588 if (IS_ERR(vb->vb_dev_info.inode)) {
589 err = PTR_ERR(vb->vb_dev_info.inode);
590 kern_unmount(balloon_mnt);
591 unregister_oom_notifier(&vb->nb);
592 vb->vb_dev_info.inode = NULL;
595 vb->vb_dev_info.inode->i_mapping->a_ops = &balloon_aops;
598 virtio_device_ready(vdev);
600 if (towards_target(vb))
601 virtballoon_changed(vdev);
605 vdev->config->del_vqs(vdev);
612 static void remove_common(struct virtio_balloon *vb)
614 /* There might be pages left in the balloon: free them. */
615 while (vb->num_pages)
616 leak_balloon(vb, vb->num_pages);
617 update_balloon_size(vb);
619 /* Now we reset the device so we can clean up the queues. */
620 vb->vdev->config->reset(vb->vdev);
622 vb->vdev->config->del_vqs(vb->vdev);
625 static void virtballoon_remove(struct virtio_device *vdev)
627 struct virtio_balloon *vb = vdev->priv;
629 unregister_oom_notifier(&vb->nb);
631 spin_lock_irq(&vb->stop_update_lock);
632 vb->stop_update = true;
633 spin_unlock_irq(&vb->stop_update_lock);
634 cancel_work_sync(&vb->update_balloon_size_work);
635 cancel_work_sync(&vb->update_balloon_stats_work);
638 #ifdef CONFIG_BALLOON_COMPACTION
639 if (vb->vb_dev_info.inode)
640 iput(vb->vb_dev_info.inode);
642 kern_unmount(balloon_mnt);
647 #ifdef CONFIG_PM_SLEEP
648 static int virtballoon_freeze(struct virtio_device *vdev)
650 struct virtio_balloon *vb = vdev->priv;
653 * The workqueue is already frozen by the PM core before this
654 * function is called.
660 static int virtballoon_restore(struct virtio_device *vdev)
662 struct virtio_balloon *vb = vdev->priv;
665 ret = init_vqs(vdev->priv);
669 virtio_device_ready(vdev);
671 if (towards_target(vb))
672 virtballoon_changed(vdev);
673 update_balloon_size(vb);
678 static unsigned int features[] = {
679 VIRTIO_BALLOON_F_MUST_TELL_HOST,
680 VIRTIO_BALLOON_F_STATS_VQ,
681 VIRTIO_BALLOON_F_DEFLATE_ON_OOM,
684 static struct virtio_driver virtio_balloon_driver = {
685 .feature_table = features,
686 .feature_table_size = ARRAY_SIZE(features),
687 .driver.name = KBUILD_MODNAME,
688 .driver.owner = THIS_MODULE,
689 .id_table = id_table,
690 .probe = virtballoon_probe,
691 .remove = virtballoon_remove,
692 .config_changed = virtballoon_changed,
693 #ifdef CONFIG_PM_SLEEP
694 .freeze = virtballoon_freeze,
695 .restore = virtballoon_restore,
699 module_virtio_driver(virtio_balloon_driver);
700 MODULE_DEVICE_TABLE(virtio, id_table);
701 MODULE_DESCRIPTION("Virtio balloon driver");
702 MODULE_LICENSE("GPL");