Merge git://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf
[platform/kernel/linux-rpi.git] / drivers / virtio / virtio_balloon.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Virtio balloon implementation, inspired by Dor Laor and Marcelo
4  * Tosatti's implementations.
5  *
6  *  Copyright 2008 Rusty Russell IBM Corporation
7  */
8
9 #include <linux/virtio.h>
10 #include <linux/virtio_balloon.h>
11 #include <linux/swap.h>
12 #include <linux/workqueue.h>
13 #include <linux/delay.h>
14 #include <linux/slab.h>
15 #include <linux/module.h>
16 #include <linux/balloon_compaction.h>
17 #include <linux/oom.h>
18 #include <linux/wait.h>
19 #include <linux/mm.h>
20 #include <linux/mount.h>
21 #include <linux/magic.h>
22 #include <linux/pseudo_fs.h>
23 #include <linux/page_reporting.h>
24
25 /*
26  * Balloon device works in 4K page units.  So each page is pointed to by
27  * multiple balloon pages.  All memory counters in this driver are in balloon
28  * page units.
29  */
30 #define VIRTIO_BALLOON_PAGES_PER_PAGE (unsigned)(PAGE_SIZE >> VIRTIO_BALLOON_PFN_SHIFT)
31 #define VIRTIO_BALLOON_ARRAY_PFNS_MAX 256
32 /* Maximum number of (4k) pages to deflate on OOM notifications. */
33 #define VIRTIO_BALLOON_OOM_NR_PAGES 256
34 #define VIRTIO_BALLOON_OOM_NOTIFY_PRIORITY 80
35
36 #define VIRTIO_BALLOON_FREE_PAGE_ALLOC_FLAG (__GFP_NORETRY | __GFP_NOWARN | \
37                                              __GFP_NOMEMALLOC)
38 /* The order of free page blocks to report to host */
39 #define VIRTIO_BALLOON_HINT_BLOCK_ORDER (MAX_ORDER - 1)
40 /* The size of a free page block in bytes */
41 #define VIRTIO_BALLOON_HINT_BLOCK_BYTES \
42         (1 << (VIRTIO_BALLOON_HINT_BLOCK_ORDER + PAGE_SHIFT))
43 #define VIRTIO_BALLOON_HINT_BLOCK_PAGES (1 << VIRTIO_BALLOON_HINT_BLOCK_ORDER)
44
45 #ifdef CONFIG_BALLOON_COMPACTION
46 static struct vfsmount *balloon_mnt;
47 #endif
48
49 enum virtio_balloon_vq {
50         VIRTIO_BALLOON_VQ_INFLATE,
51         VIRTIO_BALLOON_VQ_DEFLATE,
52         VIRTIO_BALLOON_VQ_STATS,
53         VIRTIO_BALLOON_VQ_FREE_PAGE,
54         VIRTIO_BALLOON_VQ_REPORTING,
55         VIRTIO_BALLOON_VQ_MAX
56 };
57
58 enum virtio_balloon_config_read {
59         VIRTIO_BALLOON_CONFIG_READ_CMD_ID = 0,
60 };
61
62 struct virtio_balloon {
63         struct virtio_device *vdev;
64         struct virtqueue *inflate_vq, *deflate_vq, *stats_vq, *free_page_vq;
65
66         /* Balloon's own wq for cpu-intensive work items */
67         struct workqueue_struct *balloon_wq;
68         /* The free page reporting work item submitted to the balloon wq */
69         struct work_struct report_free_page_work;
70
71         /* The balloon servicing is delegated to a freezable workqueue. */
72         struct work_struct update_balloon_stats_work;
73         struct work_struct update_balloon_size_work;
74
75         /* Prevent updating balloon when it is being canceled. */
76         spinlock_t stop_update_lock;
77         bool stop_update;
78         /* Bitmap to indicate if reading the related config fields are needed */
79         unsigned long config_read_bitmap;
80
81         /* The list of allocated free pages, waiting to be given back to mm */
82         struct list_head free_page_list;
83         spinlock_t free_page_list_lock;
84         /* The number of free page blocks on the above list */
85         unsigned long num_free_page_blocks;
86         /*
87          * The cmd id received from host.
88          * Read it via virtio_balloon_cmd_id_received to get the latest value
89          * sent from host.
90          */
91         u32 cmd_id_received_cache;
92         /* The cmd id that is actively in use */
93         __virtio32 cmd_id_active;
94         /* Buffer to store the stop sign */
95         __virtio32 cmd_id_stop;
96
97         /* Waiting for host to ack the pages we released. */
98         wait_queue_head_t acked;
99
100         /* Number of balloon pages we've told the Host we're not using. */
101         unsigned int num_pages;
102         /*
103          * The pages we've told the Host we're not using are enqueued
104          * at vb_dev_info->pages list.
105          * Each page on this list adds VIRTIO_BALLOON_PAGES_PER_PAGE
106          * to num_pages above.
107          */
108         struct balloon_dev_info vb_dev_info;
109
110         /* Synchronize access/update to this struct virtio_balloon elements */
111         struct mutex balloon_lock;
112
113         /* The array of pfns we tell the Host about. */
114         unsigned int num_pfns;
115         __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX];
116
117         /* Memory statistics */
118         struct virtio_balloon_stat stats[VIRTIO_BALLOON_S_NR];
119
120         /* Shrinker to return free pages - VIRTIO_BALLOON_F_FREE_PAGE_HINT */
121         struct shrinker shrinker;
122
123         /* OOM notifier to deflate on OOM - VIRTIO_BALLOON_F_DEFLATE_ON_OOM */
124         struct notifier_block oom_nb;
125
126         /* Free page reporting device */
127         struct virtqueue *reporting_vq;
128         struct page_reporting_dev_info pr_dev_info;
129 };
130
131 static const struct virtio_device_id id_table[] = {
132         { VIRTIO_ID_BALLOON, VIRTIO_DEV_ANY_ID },
133         { 0 },
134 };
135
136 static u32 page_to_balloon_pfn(struct page *page)
137 {
138         unsigned long pfn = page_to_pfn(page);
139
140         BUILD_BUG_ON(PAGE_SHIFT < VIRTIO_BALLOON_PFN_SHIFT);
141         /* Convert pfn from Linux page size to balloon page size. */
142         return pfn * VIRTIO_BALLOON_PAGES_PER_PAGE;
143 }
144
145 static void balloon_ack(struct virtqueue *vq)
146 {
147         struct virtio_balloon *vb = vq->vdev->priv;
148
149         wake_up(&vb->acked);
150 }
151
152 static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq)
153 {
154         struct scatterlist sg;
155         unsigned int len;
156
157         sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns);
158
159         /* We should always be able to add one buffer to an empty queue. */
160         virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
161         virtqueue_kick(vq);
162
163         /* When host has read buffer, this completes via balloon_ack */
164         wait_event(vb->acked, virtqueue_get_buf(vq, &len));
165
166 }
167
168 static int virtballoon_free_page_report(struct page_reporting_dev_info *pr_dev_info,
169                                    struct scatterlist *sg, unsigned int nents)
170 {
171         struct virtio_balloon *vb =
172                 container_of(pr_dev_info, struct virtio_balloon, pr_dev_info);
173         struct virtqueue *vq = vb->reporting_vq;
174         unsigned int unused, err;
175
176         /* We should always be able to add these buffers to an empty queue. */
177         err = virtqueue_add_inbuf(vq, sg, nents, vb, GFP_NOWAIT | __GFP_NOWARN);
178
179         /*
180          * In the extremely unlikely case that something has occurred and we
181          * are able to trigger an error we will simply display a warning
182          * and exit without actually processing the pages.
183          */
184         if (WARN_ON_ONCE(err))
185                 return err;
186
187         virtqueue_kick(vq);
188
189         /* When host has read buffer, this completes via balloon_ack */
190         wait_event(vb->acked, virtqueue_get_buf(vq, &unused));
191
192         return 0;
193 }
194
195 static void set_page_pfns(struct virtio_balloon *vb,
196                           __virtio32 pfns[], struct page *page)
197 {
198         unsigned int i;
199
200         BUILD_BUG_ON(VIRTIO_BALLOON_PAGES_PER_PAGE > VIRTIO_BALLOON_ARRAY_PFNS_MAX);
201
202         /*
203          * Set balloon pfns pointing at this page.
204          * Note that the first pfn points at start of the page.
205          */
206         for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++)
207                 pfns[i] = cpu_to_virtio32(vb->vdev,
208                                           page_to_balloon_pfn(page) + i);
209 }
210
211 static unsigned fill_balloon(struct virtio_balloon *vb, size_t num)
212 {
213         unsigned num_allocated_pages;
214         unsigned num_pfns;
215         struct page *page;
216         LIST_HEAD(pages);
217
218         /* We can only do one array worth at a time. */
219         num = min(num, ARRAY_SIZE(vb->pfns));
220
221         for (num_pfns = 0; num_pfns < num;
222              num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
223                 struct page *page = balloon_page_alloc();
224
225                 if (!page) {
226                         dev_info_ratelimited(&vb->vdev->dev,
227                                              "Out of puff! Can't get %u pages\n",
228                                              VIRTIO_BALLOON_PAGES_PER_PAGE);
229                         /* Sleep for at least 1/5 of a second before retry. */
230                         msleep(200);
231                         break;
232                 }
233
234                 balloon_page_push(&pages, page);
235         }
236
237         mutex_lock(&vb->balloon_lock);
238
239         vb->num_pfns = 0;
240
241         while ((page = balloon_page_pop(&pages))) {
242                 balloon_page_enqueue(&vb->vb_dev_info, page);
243
244                 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
245                 vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE;
246                 if (!virtio_has_feature(vb->vdev,
247                                         VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
248                         adjust_managed_page_count(page, -1);
249                 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE;
250         }
251
252         num_allocated_pages = vb->num_pfns;
253         /* Did we get any? */
254         if (vb->num_pfns != 0)
255                 tell_host(vb, vb->inflate_vq);
256         mutex_unlock(&vb->balloon_lock);
257
258         return num_allocated_pages;
259 }
260
261 static void release_pages_balloon(struct virtio_balloon *vb,
262                                  struct list_head *pages)
263 {
264         struct page *page, *next;
265
266         list_for_each_entry_safe(page, next, pages, lru) {
267                 if (!virtio_has_feature(vb->vdev,
268                                         VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
269                         adjust_managed_page_count(page, 1);
270                 list_del(&page->lru);
271                 put_page(page); /* balloon reference */
272         }
273 }
274
275 static unsigned leak_balloon(struct virtio_balloon *vb, size_t num)
276 {
277         unsigned num_freed_pages;
278         struct page *page;
279         struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
280         LIST_HEAD(pages);
281
282         /* We can only do one array worth at a time. */
283         num = min(num, ARRAY_SIZE(vb->pfns));
284
285         mutex_lock(&vb->balloon_lock);
286         /* We can't release more pages than taken */
287         num = min(num, (size_t)vb->num_pages);
288         for (vb->num_pfns = 0; vb->num_pfns < num;
289              vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
290                 page = balloon_page_dequeue(vb_dev_info);
291                 if (!page)
292                         break;
293                 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
294                 list_add(&page->lru, &pages);
295                 vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE;
296         }
297
298         num_freed_pages = vb->num_pfns;
299         /*
300          * Note that if
301          * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST);
302          * is true, we *have* to do it in this order
303          */
304         if (vb->num_pfns != 0)
305                 tell_host(vb, vb->deflate_vq);
306         release_pages_balloon(vb, &pages);
307         mutex_unlock(&vb->balloon_lock);
308         return num_freed_pages;
309 }
310
311 static inline void update_stat(struct virtio_balloon *vb, int idx,
312                                u16 tag, u64 val)
313 {
314         BUG_ON(idx >= VIRTIO_BALLOON_S_NR);
315         vb->stats[idx].tag = cpu_to_virtio16(vb->vdev, tag);
316         vb->stats[idx].val = cpu_to_virtio64(vb->vdev, val);
317 }
318
319 #define pages_to_bytes(x) ((u64)(x) << PAGE_SHIFT)
320
321 static unsigned int update_balloon_stats(struct virtio_balloon *vb)
322 {
323         unsigned long events[NR_VM_EVENT_ITEMS];
324         struct sysinfo i;
325         unsigned int idx = 0;
326         long available;
327         unsigned long caches;
328
329         all_vm_events(events);
330         si_meminfo(&i);
331
332         available = si_mem_available();
333         caches = global_node_page_state(NR_FILE_PAGES);
334
335 #ifdef CONFIG_VM_EVENT_COUNTERS
336         update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_IN,
337                                 pages_to_bytes(events[PSWPIN]));
338         update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_OUT,
339                                 pages_to_bytes(events[PSWPOUT]));
340         update_stat(vb, idx++, VIRTIO_BALLOON_S_MAJFLT, events[PGMAJFAULT]);
341         update_stat(vb, idx++, VIRTIO_BALLOON_S_MINFLT, events[PGFAULT]);
342 #ifdef CONFIG_HUGETLB_PAGE
343         update_stat(vb, idx++, VIRTIO_BALLOON_S_HTLB_PGALLOC,
344                     events[HTLB_BUDDY_PGALLOC]);
345         update_stat(vb, idx++, VIRTIO_BALLOON_S_HTLB_PGFAIL,
346                     events[HTLB_BUDDY_PGALLOC_FAIL]);
347 #endif
348 #endif
349         update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMFREE,
350                                 pages_to_bytes(i.freeram));
351         update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMTOT,
352                                 pages_to_bytes(i.totalram));
353         update_stat(vb, idx++, VIRTIO_BALLOON_S_AVAIL,
354                                 pages_to_bytes(available));
355         update_stat(vb, idx++, VIRTIO_BALLOON_S_CACHES,
356                                 pages_to_bytes(caches));
357
358         return idx;
359 }
360
361 /*
362  * While most virtqueues communicate guest-initiated requests to the hypervisor,
363  * the stats queue operates in reverse.  The driver initializes the virtqueue
364  * with a single buffer.  From that point forward, all conversations consist of
365  * a hypervisor request (a call to this function) which directs us to refill
366  * the virtqueue with a fresh stats buffer.  Since stats collection can sleep,
367  * we delegate the job to a freezable workqueue that will do the actual work via
368  * stats_handle_request().
369  */
370 static void stats_request(struct virtqueue *vq)
371 {
372         struct virtio_balloon *vb = vq->vdev->priv;
373
374         spin_lock(&vb->stop_update_lock);
375         if (!vb->stop_update)
376                 queue_work(system_freezable_wq, &vb->update_balloon_stats_work);
377         spin_unlock(&vb->stop_update_lock);
378 }
379
380 static void stats_handle_request(struct virtio_balloon *vb)
381 {
382         struct virtqueue *vq;
383         struct scatterlist sg;
384         unsigned int len, num_stats;
385
386         num_stats = update_balloon_stats(vb);
387
388         vq = vb->stats_vq;
389         if (!virtqueue_get_buf(vq, &len))
390                 return;
391         sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
392         virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
393         virtqueue_kick(vq);
394 }
395
396 static inline s64 towards_target(struct virtio_balloon *vb)
397 {
398         s64 target;
399         u32 num_pages;
400
401         /* Legacy balloon config space is LE, unlike all other devices. */
402         virtio_cread_le(vb->vdev, struct virtio_balloon_config, num_pages,
403                         &num_pages);
404
405         target = num_pages;
406         return target - vb->num_pages;
407 }
408
409 /* Gives back @num_to_return blocks of free pages to mm. */
410 static unsigned long return_free_pages_to_mm(struct virtio_balloon *vb,
411                                              unsigned long num_to_return)
412 {
413         struct page *page;
414         unsigned long num_returned;
415
416         spin_lock_irq(&vb->free_page_list_lock);
417         for (num_returned = 0; num_returned < num_to_return; num_returned++) {
418                 page = balloon_page_pop(&vb->free_page_list);
419                 if (!page)
420                         break;
421                 free_pages((unsigned long)page_address(page),
422                            VIRTIO_BALLOON_HINT_BLOCK_ORDER);
423         }
424         vb->num_free_page_blocks -= num_returned;
425         spin_unlock_irq(&vb->free_page_list_lock);
426
427         return num_returned;
428 }
429
430 static void virtio_balloon_queue_free_page_work(struct virtio_balloon *vb)
431 {
432         if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
433                 return;
434
435         /* No need to queue the work if the bit was already set. */
436         if (test_and_set_bit(VIRTIO_BALLOON_CONFIG_READ_CMD_ID,
437                              &vb->config_read_bitmap))
438                 return;
439
440         queue_work(vb->balloon_wq, &vb->report_free_page_work);
441 }
442
443 static void virtballoon_changed(struct virtio_device *vdev)
444 {
445         struct virtio_balloon *vb = vdev->priv;
446         unsigned long flags;
447
448         spin_lock_irqsave(&vb->stop_update_lock, flags);
449         if (!vb->stop_update) {
450                 queue_work(system_freezable_wq,
451                            &vb->update_balloon_size_work);
452                 virtio_balloon_queue_free_page_work(vb);
453         }
454         spin_unlock_irqrestore(&vb->stop_update_lock, flags);
455 }
456
457 static void update_balloon_size(struct virtio_balloon *vb)
458 {
459         u32 actual = vb->num_pages;
460
461         /* Legacy balloon config space is LE, unlike all other devices. */
462         virtio_cwrite_le(vb->vdev, struct virtio_balloon_config, actual,
463                          &actual);
464 }
465
466 static void update_balloon_stats_func(struct work_struct *work)
467 {
468         struct virtio_balloon *vb;
469
470         vb = container_of(work, struct virtio_balloon,
471                           update_balloon_stats_work);
472         stats_handle_request(vb);
473 }
474
475 static void update_balloon_size_func(struct work_struct *work)
476 {
477         struct virtio_balloon *vb;
478         s64 diff;
479
480         vb = container_of(work, struct virtio_balloon,
481                           update_balloon_size_work);
482         diff = towards_target(vb);
483
484         if (!diff)
485                 return;
486
487         if (diff > 0)
488                 diff -= fill_balloon(vb, diff);
489         else
490                 diff += leak_balloon(vb, -diff);
491         update_balloon_size(vb);
492
493         if (diff)
494                 queue_work(system_freezable_wq, work);
495 }
496
497 static int init_vqs(struct virtio_balloon *vb)
498 {
499         struct virtqueue *vqs[VIRTIO_BALLOON_VQ_MAX];
500         vq_callback_t *callbacks[VIRTIO_BALLOON_VQ_MAX];
501         const char *names[VIRTIO_BALLOON_VQ_MAX];
502         int err;
503
504         /*
505          * Inflateq and deflateq are used unconditionally. The names[]
506          * will be NULL if the related feature is not enabled, which will
507          * cause no allocation for the corresponding virtqueue in find_vqs.
508          */
509         callbacks[VIRTIO_BALLOON_VQ_INFLATE] = balloon_ack;
510         names[VIRTIO_BALLOON_VQ_INFLATE] = "inflate";
511         callbacks[VIRTIO_BALLOON_VQ_DEFLATE] = balloon_ack;
512         names[VIRTIO_BALLOON_VQ_DEFLATE] = "deflate";
513         callbacks[VIRTIO_BALLOON_VQ_STATS] = NULL;
514         names[VIRTIO_BALLOON_VQ_STATS] = NULL;
515         callbacks[VIRTIO_BALLOON_VQ_FREE_PAGE] = NULL;
516         names[VIRTIO_BALLOON_VQ_FREE_PAGE] = NULL;
517         names[VIRTIO_BALLOON_VQ_REPORTING] = NULL;
518
519         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
520                 names[VIRTIO_BALLOON_VQ_STATS] = "stats";
521                 callbacks[VIRTIO_BALLOON_VQ_STATS] = stats_request;
522         }
523
524         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
525                 names[VIRTIO_BALLOON_VQ_FREE_PAGE] = "free_page_vq";
526                 callbacks[VIRTIO_BALLOON_VQ_FREE_PAGE] = NULL;
527         }
528
529         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING)) {
530                 names[VIRTIO_BALLOON_VQ_REPORTING] = "reporting_vq";
531                 callbacks[VIRTIO_BALLOON_VQ_REPORTING] = balloon_ack;
532         }
533
534         err = virtio_find_vqs(vb->vdev, VIRTIO_BALLOON_VQ_MAX, vqs,
535                               callbacks, names, NULL);
536         if (err)
537                 return err;
538
539         vb->inflate_vq = vqs[VIRTIO_BALLOON_VQ_INFLATE];
540         vb->deflate_vq = vqs[VIRTIO_BALLOON_VQ_DEFLATE];
541         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
542                 struct scatterlist sg;
543                 unsigned int num_stats;
544                 vb->stats_vq = vqs[VIRTIO_BALLOON_VQ_STATS];
545
546                 /*
547                  * Prime this virtqueue with one buffer so the hypervisor can
548                  * use it to signal us later (it can't be broken yet!).
549                  */
550                 num_stats = update_balloon_stats(vb);
551
552                 sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
553                 err = virtqueue_add_outbuf(vb->stats_vq, &sg, 1, vb,
554                                            GFP_KERNEL);
555                 if (err) {
556                         dev_warn(&vb->vdev->dev, "%s: add stat_vq failed\n",
557                                  __func__);
558                         return err;
559                 }
560                 virtqueue_kick(vb->stats_vq);
561         }
562
563         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
564                 vb->free_page_vq = vqs[VIRTIO_BALLOON_VQ_FREE_PAGE];
565
566         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING))
567                 vb->reporting_vq = vqs[VIRTIO_BALLOON_VQ_REPORTING];
568
569         return 0;
570 }
571
572 static u32 virtio_balloon_cmd_id_received(struct virtio_balloon *vb)
573 {
574         if (test_and_clear_bit(VIRTIO_BALLOON_CONFIG_READ_CMD_ID,
575                                &vb->config_read_bitmap)) {
576                 /* Legacy balloon config space is LE, unlike all other devices. */
577                 virtio_cread_le(vb->vdev, struct virtio_balloon_config,
578                                 free_page_hint_cmd_id,
579                                 &vb->cmd_id_received_cache);
580         }
581
582         return vb->cmd_id_received_cache;
583 }
584
585 static int send_cmd_id_start(struct virtio_balloon *vb)
586 {
587         struct scatterlist sg;
588         struct virtqueue *vq = vb->free_page_vq;
589         int err, unused;
590
591         /* Detach all the used buffers from the vq */
592         while (virtqueue_get_buf(vq, &unused))
593                 ;
594
595         vb->cmd_id_active = cpu_to_virtio32(vb->vdev,
596                                         virtio_balloon_cmd_id_received(vb));
597         sg_init_one(&sg, &vb->cmd_id_active, sizeof(vb->cmd_id_active));
598         err = virtqueue_add_outbuf(vq, &sg, 1, &vb->cmd_id_active, GFP_KERNEL);
599         if (!err)
600                 virtqueue_kick(vq);
601         return err;
602 }
603
604 static int send_cmd_id_stop(struct virtio_balloon *vb)
605 {
606         struct scatterlist sg;
607         struct virtqueue *vq = vb->free_page_vq;
608         int err, unused;
609
610         /* Detach all the used buffers from the vq */
611         while (virtqueue_get_buf(vq, &unused))
612                 ;
613
614         sg_init_one(&sg, &vb->cmd_id_stop, sizeof(vb->cmd_id_stop));
615         err = virtqueue_add_outbuf(vq, &sg, 1, &vb->cmd_id_stop, GFP_KERNEL);
616         if (!err)
617                 virtqueue_kick(vq);
618         return err;
619 }
620
621 static int get_free_page_and_send(struct virtio_balloon *vb)
622 {
623         struct virtqueue *vq = vb->free_page_vq;
624         struct page *page;
625         struct scatterlist sg;
626         int err, unused;
627         void *p;
628
629         /* Detach all the used buffers from the vq */
630         while (virtqueue_get_buf(vq, &unused))
631                 ;
632
633         page = alloc_pages(VIRTIO_BALLOON_FREE_PAGE_ALLOC_FLAG,
634                            VIRTIO_BALLOON_HINT_BLOCK_ORDER);
635         /*
636          * When the allocation returns NULL, it indicates that we have got all
637          * the possible free pages, so return -EINTR to stop.
638          */
639         if (!page)
640                 return -EINTR;
641
642         p = page_address(page);
643         sg_init_one(&sg, p, VIRTIO_BALLOON_HINT_BLOCK_BYTES);
644         /* There is always 1 entry reserved for the cmd id to use. */
645         if (vq->num_free > 1) {
646                 err = virtqueue_add_inbuf(vq, &sg, 1, p, GFP_KERNEL);
647                 if (unlikely(err)) {
648                         free_pages((unsigned long)p,
649                                    VIRTIO_BALLOON_HINT_BLOCK_ORDER);
650                         return err;
651                 }
652                 virtqueue_kick(vq);
653                 spin_lock_irq(&vb->free_page_list_lock);
654                 balloon_page_push(&vb->free_page_list, page);
655                 vb->num_free_page_blocks++;
656                 spin_unlock_irq(&vb->free_page_list_lock);
657         } else {
658                 /*
659                  * The vq has no available entry to add this page block, so
660                  * just free it.
661                  */
662                 free_pages((unsigned long)p, VIRTIO_BALLOON_HINT_BLOCK_ORDER);
663         }
664
665         return 0;
666 }
667
668 static int send_free_pages(struct virtio_balloon *vb)
669 {
670         int err;
671         u32 cmd_id_active;
672
673         while (1) {
674                 /*
675                  * If a stop id or a new cmd id was just received from host,
676                  * stop the reporting.
677                  */
678                 cmd_id_active = virtio32_to_cpu(vb->vdev, vb->cmd_id_active);
679                 if (unlikely(cmd_id_active !=
680                              virtio_balloon_cmd_id_received(vb)))
681                         break;
682
683                 /*
684                  * The free page blocks are allocated and sent to host one by
685                  * one.
686                  */
687                 err = get_free_page_and_send(vb);
688                 if (err == -EINTR)
689                         break;
690                 else if (unlikely(err))
691                         return err;
692         }
693
694         return 0;
695 }
696
697 static void virtio_balloon_report_free_page(struct virtio_balloon *vb)
698 {
699         int err;
700         struct device *dev = &vb->vdev->dev;
701
702         /* Start by sending the received cmd id to host with an outbuf. */
703         err = send_cmd_id_start(vb);
704         if (unlikely(err))
705                 dev_err(dev, "Failed to send a start id, err = %d\n", err);
706
707         err = send_free_pages(vb);
708         if (unlikely(err))
709                 dev_err(dev, "Failed to send a free page, err = %d\n", err);
710
711         /* End by sending a stop id to host with an outbuf. */
712         err = send_cmd_id_stop(vb);
713         if (unlikely(err))
714                 dev_err(dev, "Failed to send a stop id, err = %d\n", err);
715 }
716
717 static void report_free_page_func(struct work_struct *work)
718 {
719         struct virtio_balloon *vb = container_of(work, struct virtio_balloon,
720                                                  report_free_page_work);
721         u32 cmd_id_received;
722
723         cmd_id_received = virtio_balloon_cmd_id_received(vb);
724         if (cmd_id_received == VIRTIO_BALLOON_CMD_ID_DONE) {
725                 /* Pass ULONG_MAX to give back all the free pages */
726                 return_free_pages_to_mm(vb, ULONG_MAX);
727         } else if (cmd_id_received != VIRTIO_BALLOON_CMD_ID_STOP &&
728                    cmd_id_received !=
729                    virtio32_to_cpu(vb->vdev, vb->cmd_id_active)) {
730                 virtio_balloon_report_free_page(vb);
731         }
732 }
733
734 #ifdef CONFIG_BALLOON_COMPACTION
735 /*
736  * virtballoon_migratepage - perform the balloon page migration on behalf of
737  *                           a compaction thread.     (called under page lock)
738  * @vb_dev_info: the balloon device
739  * @newpage: page that will replace the isolated page after migration finishes.
740  * @page   : the isolated (old) page that is about to be migrated to newpage.
741  * @mode   : compaction mode -- not used for balloon page migration.
742  *
743  * After a ballooned page gets isolated by compaction procedures, this is the
744  * function that performs the page migration on behalf of a compaction thread
745  * The page migration for virtio balloon is done in a simple swap fashion which
746  * follows these two macro steps:
747  *  1) insert newpage into vb->pages list and update the host about it;
748  *  2) update the host about the old page removed from vb->pages list;
749  *
750  * This function preforms the balloon page migration task.
751  * Called through balloon_mapping->a_ops->migratepage
752  */
753 static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info,
754                 struct page *newpage, struct page *page, enum migrate_mode mode)
755 {
756         struct virtio_balloon *vb = container_of(vb_dev_info,
757                         struct virtio_balloon, vb_dev_info);
758         unsigned long flags;
759
760         /*
761          * In order to avoid lock contention while migrating pages concurrently
762          * to leak_balloon() or fill_balloon() we just give up the balloon_lock
763          * this turn, as it is easier to retry the page migration later.
764          * This also prevents fill_balloon() getting stuck into a mutex
765          * recursion in the case it ends up triggering memory compaction
766          * while it is attempting to inflate the ballon.
767          */
768         if (!mutex_trylock(&vb->balloon_lock))
769                 return -EAGAIN;
770
771         get_page(newpage); /* balloon reference */
772
773         /*
774           * When we migrate a page to a different zone and adjusted the
775           * managed page count when inflating, we have to fixup the count of
776           * both involved zones.
777           */
778         if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM) &&
779             page_zone(page) != page_zone(newpage)) {
780                 adjust_managed_page_count(page, 1);
781                 adjust_managed_page_count(newpage, -1);
782         }
783
784         /* balloon's page migration 1st step  -- inflate "newpage" */
785         spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
786         balloon_page_insert(vb_dev_info, newpage);
787         vb_dev_info->isolated_pages--;
788         __count_vm_event(BALLOON_MIGRATE);
789         spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
790         vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
791         set_page_pfns(vb, vb->pfns, newpage);
792         tell_host(vb, vb->inflate_vq);
793
794         /* balloon's page migration 2nd step -- deflate "page" */
795         spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
796         balloon_page_delete(page);
797         spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
798         vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
799         set_page_pfns(vb, vb->pfns, page);
800         tell_host(vb, vb->deflate_vq);
801
802         mutex_unlock(&vb->balloon_lock);
803
804         put_page(page); /* balloon reference */
805
806         return MIGRATEPAGE_SUCCESS;
807 }
808
809 static int balloon_init_fs_context(struct fs_context *fc)
810 {
811         return init_pseudo(fc, BALLOON_KVM_MAGIC) ? 0 : -ENOMEM;
812 }
813
814 static struct file_system_type balloon_fs = {
815         .name           = "balloon-kvm",
816         .init_fs_context = balloon_init_fs_context,
817         .kill_sb        = kill_anon_super,
818 };
819
820 #endif /* CONFIG_BALLOON_COMPACTION */
821
822 static unsigned long shrink_free_pages(struct virtio_balloon *vb,
823                                        unsigned long pages_to_free)
824 {
825         unsigned long blocks_to_free, blocks_freed;
826
827         pages_to_free = round_up(pages_to_free,
828                                  VIRTIO_BALLOON_HINT_BLOCK_PAGES);
829         blocks_to_free = pages_to_free / VIRTIO_BALLOON_HINT_BLOCK_PAGES;
830         blocks_freed = return_free_pages_to_mm(vb, blocks_to_free);
831
832         return blocks_freed * VIRTIO_BALLOON_HINT_BLOCK_PAGES;
833 }
834
835 static unsigned long virtio_balloon_shrinker_scan(struct shrinker *shrinker,
836                                                   struct shrink_control *sc)
837 {
838         struct virtio_balloon *vb = container_of(shrinker,
839                                         struct virtio_balloon, shrinker);
840
841         return shrink_free_pages(vb, sc->nr_to_scan);
842 }
843
844 static unsigned long virtio_balloon_shrinker_count(struct shrinker *shrinker,
845                                                    struct shrink_control *sc)
846 {
847         struct virtio_balloon *vb = container_of(shrinker,
848                                         struct virtio_balloon, shrinker);
849
850         return vb->num_free_page_blocks * VIRTIO_BALLOON_HINT_BLOCK_PAGES;
851 }
852
853 static int virtio_balloon_oom_notify(struct notifier_block *nb,
854                                      unsigned long dummy, void *parm)
855 {
856         struct virtio_balloon *vb = container_of(nb,
857                                                  struct virtio_balloon, oom_nb);
858         unsigned long *freed = parm;
859
860         *freed += leak_balloon(vb, VIRTIO_BALLOON_OOM_NR_PAGES) /
861                   VIRTIO_BALLOON_PAGES_PER_PAGE;
862         update_balloon_size(vb);
863
864         return NOTIFY_OK;
865 }
866
867 static void virtio_balloon_unregister_shrinker(struct virtio_balloon *vb)
868 {
869         unregister_shrinker(&vb->shrinker);
870 }
871
872 static int virtio_balloon_register_shrinker(struct virtio_balloon *vb)
873 {
874         vb->shrinker.scan_objects = virtio_balloon_shrinker_scan;
875         vb->shrinker.count_objects = virtio_balloon_shrinker_count;
876         vb->shrinker.seeks = DEFAULT_SEEKS;
877
878         return register_shrinker(&vb->shrinker);
879 }
880
881 static int virtballoon_probe(struct virtio_device *vdev)
882 {
883         struct virtio_balloon *vb;
884         int err;
885
886         if (!vdev->config->get) {
887                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
888                         __func__);
889                 return -EINVAL;
890         }
891
892         vdev->priv = vb = kzalloc(sizeof(*vb), GFP_KERNEL);
893         if (!vb) {
894                 err = -ENOMEM;
895                 goto out;
896         }
897
898         INIT_WORK(&vb->update_balloon_stats_work, update_balloon_stats_func);
899         INIT_WORK(&vb->update_balloon_size_work, update_balloon_size_func);
900         spin_lock_init(&vb->stop_update_lock);
901         mutex_init(&vb->balloon_lock);
902         init_waitqueue_head(&vb->acked);
903         vb->vdev = vdev;
904
905         balloon_devinfo_init(&vb->vb_dev_info);
906
907         err = init_vqs(vb);
908         if (err)
909                 goto out_free_vb;
910
911 #ifdef CONFIG_BALLOON_COMPACTION
912         balloon_mnt = kern_mount(&balloon_fs);
913         if (IS_ERR(balloon_mnt)) {
914                 err = PTR_ERR(balloon_mnt);
915                 goto out_del_vqs;
916         }
917
918         vb->vb_dev_info.migratepage = virtballoon_migratepage;
919         vb->vb_dev_info.inode = alloc_anon_inode(balloon_mnt->mnt_sb);
920         if (IS_ERR(vb->vb_dev_info.inode)) {
921                 err = PTR_ERR(vb->vb_dev_info.inode);
922                 goto out_kern_unmount;
923         }
924         vb->vb_dev_info.inode->i_mapping->a_ops = &balloon_aops;
925 #endif
926         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
927                 /*
928                  * There is always one entry reserved for cmd id, so the ring
929                  * size needs to be at least two to report free page hints.
930                  */
931                 if (virtqueue_get_vring_size(vb->free_page_vq) < 2) {
932                         err = -ENOSPC;
933                         goto out_iput;
934                 }
935                 vb->balloon_wq = alloc_workqueue("balloon-wq",
936                                         WQ_FREEZABLE | WQ_CPU_INTENSIVE, 0);
937                 if (!vb->balloon_wq) {
938                         err = -ENOMEM;
939                         goto out_iput;
940                 }
941                 INIT_WORK(&vb->report_free_page_work, report_free_page_func);
942                 vb->cmd_id_received_cache = VIRTIO_BALLOON_CMD_ID_STOP;
943                 vb->cmd_id_active = cpu_to_virtio32(vb->vdev,
944                                                   VIRTIO_BALLOON_CMD_ID_STOP);
945                 vb->cmd_id_stop = cpu_to_virtio32(vb->vdev,
946                                                   VIRTIO_BALLOON_CMD_ID_STOP);
947                 spin_lock_init(&vb->free_page_list_lock);
948                 INIT_LIST_HEAD(&vb->free_page_list);
949                 /*
950                  * We're allowed to reuse any free pages, even if they are
951                  * still to be processed by the host.
952                  */
953                 err = virtio_balloon_register_shrinker(vb);
954                 if (err)
955                         goto out_del_balloon_wq;
956         }
957
958         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) {
959                 vb->oom_nb.notifier_call = virtio_balloon_oom_notify;
960                 vb->oom_nb.priority = VIRTIO_BALLOON_OOM_NOTIFY_PRIORITY;
961                 err = register_oom_notifier(&vb->oom_nb);
962                 if (err < 0)
963                         goto out_unregister_shrinker;
964         }
965
966         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_PAGE_POISON)) {
967                 /* Start with poison val of 0 representing general init */
968                 __u32 poison_val = 0;
969
970                 /*
971                  * Let the hypervisor know that we are expecting a
972                  * specific value to be written back in balloon pages.
973                  *
974                  * If the PAGE_POISON value was larger than a byte we would
975                  * need to byte swap poison_val here to guarantee it is
976                  * little-endian. However for now it is a single byte so we
977                  * can pass it as-is.
978                  */
979                 if (!want_init_on_free())
980                         memset(&poison_val, PAGE_POISON, sizeof(poison_val));
981
982                 virtio_cwrite_le(vb->vdev, struct virtio_balloon_config,
983                                  poison_val, &poison_val);
984         }
985
986         vb->pr_dev_info.report = virtballoon_free_page_report;
987         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING)) {
988                 unsigned int capacity;
989
990                 capacity = virtqueue_get_vring_size(vb->reporting_vq);
991                 if (capacity < PAGE_REPORTING_CAPACITY) {
992                         err = -ENOSPC;
993                         goto out_unregister_oom;
994                 }
995
996                 /*
997                  * The default page reporting order is @pageblock_order, which
998                  * corresponds to 512MB in size on ARM64 when 64KB base page
999                  * size is used. The page reporting won't be triggered if the
1000                  * freeing page can't come up with a free area like that huge.
1001                  * So we specify the page reporting order to 5, corresponding
1002                  * to 2MB. It helps to avoid THP splitting if 4KB base page
1003                  * size is used by host.
1004                  *
1005                  * Ideally, the page reporting order is selected based on the
1006                  * host's base page size. However, it needs more work to report
1007                  * that value. The hard-coded order would be fine currently.
1008                  */
1009 #if defined(CONFIG_ARM64) && defined(CONFIG_ARM64_64K_PAGES)
1010                 vb->pr_dev_info.order = 5;
1011 #endif
1012
1013                 err = page_reporting_register(&vb->pr_dev_info);
1014                 if (err)
1015                         goto out_unregister_oom;
1016         }
1017
1018         virtio_device_ready(vdev);
1019
1020         if (towards_target(vb))
1021                 virtballoon_changed(vdev);
1022         return 0;
1023
1024 out_unregister_oom:
1025         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
1026                 unregister_oom_notifier(&vb->oom_nb);
1027 out_unregister_shrinker:
1028         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
1029                 virtio_balloon_unregister_shrinker(vb);
1030 out_del_balloon_wq:
1031         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
1032                 destroy_workqueue(vb->balloon_wq);
1033 out_iput:
1034 #ifdef CONFIG_BALLOON_COMPACTION
1035         iput(vb->vb_dev_info.inode);
1036 out_kern_unmount:
1037         kern_unmount(balloon_mnt);
1038 out_del_vqs:
1039 #endif
1040         vdev->config->del_vqs(vdev);
1041 out_free_vb:
1042         kfree(vb);
1043 out:
1044         return err;
1045 }
1046
1047 static void remove_common(struct virtio_balloon *vb)
1048 {
1049         /* There might be pages left in the balloon: free them. */
1050         while (vb->num_pages)
1051                 leak_balloon(vb, vb->num_pages);
1052         update_balloon_size(vb);
1053
1054         /* There might be free pages that are being reported: release them. */
1055         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
1056                 return_free_pages_to_mm(vb, ULONG_MAX);
1057
1058         /* Now we reset the device so we can clean up the queues. */
1059         vb->vdev->config->reset(vb->vdev);
1060
1061         vb->vdev->config->del_vqs(vb->vdev);
1062 }
1063
1064 static void virtballoon_remove(struct virtio_device *vdev)
1065 {
1066         struct virtio_balloon *vb = vdev->priv;
1067
1068         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING))
1069                 page_reporting_unregister(&vb->pr_dev_info);
1070         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
1071                 unregister_oom_notifier(&vb->oom_nb);
1072         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
1073                 virtio_balloon_unregister_shrinker(vb);
1074         spin_lock_irq(&vb->stop_update_lock);
1075         vb->stop_update = true;
1076         spin_unlock_irq(&vb->stop_update_lock);
1077         cancel_work_sync(&vb->update_balloon_size_work);
1078         cancel_work_sync(&vb->update_balloon_stats_work);
1079
1080         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
1081                 cancel_work_sync(&vb->report_free_page_work);
1082                 destroy_workqueue(vb->balloon_wq);
1083         }
1084
1085         remove_common(vb);
1086 #ifdef CONFIG_BALLOON_COMPACTION
1087         if (vb->vb_dev_info.inode)
1088                 iput(vb->vb_dev_info.inode);
1089
1090         kern_unmount(balloon_mnt);
1091 #endif
1092         kfree(vb);
1093 }
1094
1095 #ifdef CONFIG_PM_SLEEP
1096 static int virtballoon_freeze(struct virtio_device *vdev)
1097 {
1098         struct virtio_balloon *vb = vdev->priv;
1099
1100         /*
1101          * The workqueue is already frozen by the PM core before this
1102          * function is called.
1103          */
1104         remove_common(vb);
1105         return 0;
1106 }
1107
1108 static int virtballoon_restore(struct virtio_device *vdev)
1109 {
1110         struct virtio_balloon *vb = vdev->priv;
1111         int ret;
1112
1113         ret = init_vqs(vdev->priv);
1114         if (ret)
1115                 return ret;
1116
1117         virtio_device_ready(vdev);
1118
1119         if (towards_target(vb))
1120                 virtballoon_changed(vdev);
1121         update_balloon_size(vb);
1122         return 0;
1123 }
1124 #endif
1125
1126 static int virtballoon_validate(struct virtio_device *vdev)
1127 {
1128         /*
1129          * Inform the hypervisor that our pages are poisoned or
1130          * initialized. If we cannot do that then we should disable
1131          * page reporting as it could potentially change the contents
1132          * of our free pages.
1133          */
1134         if (!want_init_on_free() && !page_poisoning_enabled_static())
1135                 __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_PAGE_POISON);
1136         else if (!virtio_has_feature(vdev, VIRTIO_BALLOON_F_PAGE_POISON))
1137                 __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_REPORTING);
1138
1139         __virtio_clear_bit(vdev, VIRTIO_F_ACCESS_PLATFORM);
1140         return 0;
1141 }
1142
1143 static unsigned int features[] = {
1144         VIRTIO_BALLOON_F_MUST_TELL_HOST,
1145         VIRTIO_BALLOON_F_STATS_VQ,
1146         VIRTIO_BALLOON_F_DEFLATE_ON_OOM,
1147         VIRTIO_BALLOON_F_FREE_PAGE_HINT,
1148         VIRTIO_BALLOON_F_PAGE_POISON,
1149         VIRTIO_BALLOON_F_REPORTING,
1150 };
1151
1152 static struct virtio_driver virtio_balloon_driver = {
1153         .feature_table = features,
1154         .feature_table_size = ARRAY_SIZE(features),
1155         .driver.name =  KBUILD_MODNAME,
1156         .driver.owner = THIS_MODULE,
1157         .id_table =     id_table,
1158         .validate =     virtballoon_validate,
1159         .probe =        virtballoon_probe,
1160         .remove =       virtballoon_remove,
1161         .config_changed = virtballoon_changed,
1162 #ifdef CONFIG_PM_SLEEP
1163         .freeze =       virtballoon_freeze,
1164         .restore =      virtballoon_restore,
1165 #endif
1166 };
1167
1168 module_virtio_driver(virtio_balloon_driver);
1169 MODULE_DEVICE_TABLE(virtio, id_table);
1170 MODULE_DESCRIPTION("Virtio balloon driver");
1171 MODULE_LICENSE("GPL");