drivers/virtio/virtio_balloon.c

   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/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/wait.h>
  31 #include <linux/mm.h>
  32 #include <linux/mount.h>
  33 #include <linux/magic.h>
  34
  35 /*
  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
  38  * page units.
  39  */
  40 #define VIRTIO_BALLOON_PAGES_PER_PAGE (unsigned)(PAGE_SIZE >> VIRTIO_BALLOON_PFN_SHIFT)
  41 #define VIRTIO_BALLOON_ARRAY_PFNS_MAX 256
  42 #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80
  43
  44 #define VIRTIO_BALLOON_FREE_PAGE_ALLOC_FLAG (__GFP_NORETRY | __GFP_NOWARN | \
  45                                              __GFP_NOMEMALLOC)
  46 /* The order of free page blocks to report to host */
  47 #define VIRTIO_BALLOON_FREE_PAGE_ORDER (MAX_ORDER - 1)
  48 /* The size of a free page block in bytes */
  49 #define VIRTIO_BALLOON_FREE_PAGE_SIZE \
  50         (1 << (VIRTIO_BALLOON_FREE_PAGE_ORDER + PAGE_SHIFT))
  51
  52 #ifdef CONFIG_BALLOON_COMPACTION
  53 static struct vfsmount *balloon_mnt;
  54 #endif
  55
  56 enum virtio_balloon_vq {
  57         VIRTIO_BALLOON_VQ_INFLATE,
  58         VIRTIO_BALLOON_VQ_DEFLATE,
  59         VIRTIO_BALLOON_VQ_STATS,
  60         VIRTIO_BALLOON_VQ_FREE_PAGE,
  61         VIRTIO_BALLOON_VQ_MAX
  62 };
  63
  64 enum virtio_balloon_config_read {
  65         VIRTIO_BALLOON_CONFIG_READ_CMD_ID = 0,
  66 };
  67
  68 struct virtio_balloon {
  69         struct virtio_device *vdev;
  70         struct virtqueue *inflate_vq, *deflate_vq, *stats_vq, *free_page_vq;
  71
  72         /* Balloon's own wq for cpu-intensive work items */
  73         struct workqueue_struct *balloon_wq;
  74         /* The free page reporting work item submitted to the balloon wq */
  75         struct work_struct report_free_page_work;
  76
  77         /* The balloon servicing is delegated to a freezable workqueue. */
  78         struct work_struct update_balloon_stats_work;
  79         struct work_struct update_balloon_size_work;
  80
  81         /* Prevent updating balloon when it is being canceled. */
  82         spinlock_t stop_update_lock;
  83         bool stop_update;
  84         /* Bitmap to indicate if reading the related config fields are needed */
  85         unsigned long config_read_bitmap;
  86
  87         /* The list of allocated free pages, waiting to be given back to mm */
  88         struct list_head free_page_list;
  89         spinlock_t free_page_list_lock;
  90         /* The number of free page blocks on the above list */
  91         unsigned long num_free_page_blocks;
  92         /*
  93          * The cmd id received from host.
  94          * Read it via virtio_balloon_cmd_id_received to get the latest value
  95          * sent from host.
  96          */
  97         u32 cmd_id_received_cache;
  98         /* The cmd id that is actively in use */
  99         __virtio32 cmd_id_active;
 100         /* Buffer to store the stop sign */
 101         __virtio32 cmd_id_stop;
 102
 103         /* Waiting for host to ack the pages we released. */
 104         wait_queue_head_t acked;
 105
 106         /* Number of balloon pages we've told the Host we're not using. */
 107         unsigned int num_pages;
 108         /*
 109          * The pages we've told the Host we're not using are enqueued
 110          * at vb_dev_info->pages list.
 111          * Each page on this list adds VIRTIO_BALLOON_PAGES_PER_PAGE
 112          * to num_pages above.
 113          */
 114         struct balloon_dev_info vb_dev_info;
 115
 116         /* Synchronize access/update to this struct virtio_balloon elements */
 117         struct mutex balloon_lock;
 118
 119         /* The array of pfns we tell the Host about. */
 120         unsigned int num_pfns;
 121         __virtio32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX];
 122
 123         /* Memory statistics */
 124         struct virtio_balloon_stat stats[VIRTIO_BALLOON_S_NR];
 125
 126         /* To register a shrinker to shrink memory upon memory pressure */
 127         struct shrinker shrinker;
 128 };
 129
 130 static struct virtio_device_id id_table[] = {
 131         { VIRTIO_ID_BALLOON, VIRTIO_DEV_ANY_ID },
 132         { 0 },
 133 };
 134
 135 static u32 page_to_balloon_pfn(struct page *page)
 136 {
 137         unsigned long pfn = page_to_pfn(page);
 138
 139         BUILD_BUG_ON(PAGE_SHIFT < VIRTIO_BALLOON_PFN_SHIFT);
 140         /* Convert pfn from Linux page size to balloon page size. */
 141         return pfn * VIRTIO_BALLOON_PAGES_PER_PAGE;
 142 }
 143
 144 static void balloon_ack(struct virtqueue *vq)
 145 {
 146         struct virtio_balloon *vb = vq->vdev->priv;
 147
 148         wake_up(&vb->acked);
 149 }
 150
 151 static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq)
 152 {
 153         struct scatterlist sg;
 154         unsigned int len;
 155
 156         sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns);
 157
 158         /* We should always be able to add one buffer to an empty queue. */
 159         virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
 160         virtqueue_kick(vq);
 161
 162         /* When host has read buffer, this completes via balloon_ack */
 163         wait_event(vb->acked, virtqueue_get_buf(vq, &len));
 164
 165 }
 166
 167 static void set_page_pfns(struct virtio_balloon *vb,
 168                           __virtio32 pfns[], struct page *page)
 169 {
 170         unsigned int i;
 171
 172         /*
 173          * Set balloon pfns pointing at this page.
 174          * Note that the first pfn points at start of the page.
 175          */
 176         for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++)
 177                 pfns[i] = cpu_to_virtio32(vb->vdev,
 178                                           page_to_balloon_pfn(page) + i);
 179 }
 180
 181 static unsigned fill_balloon(struct virtio_balloon *vb, size_t num)
 182 {
 183         unsigned num_allocated_pages;
 184         unsigned num_pfns;
 185         struct page *page;
 186         LIST_HEAD(pages);
 187
 188         /* We can only do one array worth at a time. */
 189         num = min(num, ARRAY_SIZE(vb->pfns));
 190
 191         for (num_pfns = 0; num_pfns < num;
 192              num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
 193                 struct page *page = balloon_page_alloc();
 194
 195                 if (!page) {
 196                         dev_info_ratelimited(&vb->vdev->dev,
 197                                              "Out of puff! Can't get %u pages\n",
 198                                              VIRTIO_BALLOON_PAGES_PER_PAGE);
 199                         /* Sleep for at least 1/5 of a second before retry. */
 200                         msleep(200);
 201                         break;
 202                 }
 203
 204                 balloon_page_push(&pages, page);
 205         }
 206
 207         mutex_lock(&vb->balloon_lock);
 208
 209         vb->num_pfns = 0;
 210
 211         while ((page = balloon_page_pop(&pages))) {
 212                 balloon_page_enqueue(&vb->vb_dev_info, page);
 213
 214                 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
 215                 vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE;
 216                 if (!virtio_has_feature(vb->vdev,
 217                                         VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
 218                         adjust_managed_page_count(page, -1);
 219                 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE;
 220         }
 221
 222         num_allocated_pages = vb->num_pfns;
 223         /* Did we get any? */
 224         if (vb->num_pfns != 0)
 225                 tell_host(vb, vb->inflate_vq);
 226         mutex_unlock(&vb->balloon_lock);
 227
 228         return num_allocated_pages;
 229 }
 230
 231 static void release_pages_balloon(struct virtio_balloon *vb,
 232                                  struct list_head *pages)
 233 {
 234         struct page *page, *next;
 235
 236         list_for_each_entry_safe(page, next, pages, lru) {
 237                 if (!virtio_has_feature(vb->vdev,
 238                                         VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
 239                         adjust_managed_page_count(page, 1);
 240                 list_del(&page->lru);
 241                 put_page(page); /* balloon reference */
 242         }
 243 }
 244
 245 static unsigned leak_balloon(struct virtio_balloon *vb, size_t num)
 246 {
 247         unsigned num_freed_pages;
 248         struct page *page;
 249         struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info;
 250         LIST_HEAD(pages);
 251
 252         /* We can only do one array worth at a time. */
 253         num = min(num, ARRAY_SIZE(vb->pfns));
 254
 255         mutex_lock(&vb->balloon_lock);
 256         /* We can't release more pages than taken */
 257         num = min(num, (size_t)vb->num_pages);
 258         for (vb->num_pfns = 0; vb->num_pfns < num;
 259              vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) {
 260                 page = balloon_page_dequeue(vb_dev_info);
 261                 if (!page)
 262                         break;
 263                 set_page_pfns(vb, vb->pfns + vb->num_pfns, page);
 264                 list_add(&page->lru, &pages);
 265                 vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE;
 266         }
 267
 268         num_freed_pages = vb->num_pfns;
 269         /*
 270          * Note that if
 271          * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST);
 272          * is true, we *have* to do it in this order
 273          */
 274         if (vb->num_pfns != 0)
 275                 tell_host(vb, vb->deflate_vq);
 276         release_pages_balloon(vb, &pages);
 277         mutex_unlock(&vb->balloon_lock);
 278         return num_freed_pages;
 279 }
 280
 281 static inline void update_stat(struct virtio_balloon *vb, int idx,
 282                                u16 tag, u64 val)
 283 {
 284         BUG_ON(idx >= VIRTIO_BALLOON_S_NR);
 285         vb->stats[idx].tag = cpu_to_virtio16(vb->vdev, tag);
 286         vb->stats[idx].val = cpu_to_virtio64(vb->vdev, val);
 287 }
 288
 289 #define pages_to_bytes(x) ((u64)(x) << PAGE_SHIFT)
 290
 291 static unsigned int update_balloon_stats(struct virtio_balloon *vb)
 292 {
 293         unsigned long events[NR_VM_EVENT_ITEMS];
 294         struct sysinfo i;
 295         unsigned int idx = 0;
 296         long available;
 297         unsigned long caches;
 298
 299         all_vm_events(events);
 300         si_meminfo(&i);
 301
 302         available = si_mem_available();
 303         caches = global_node_page_state(NR_FILE_PAGES);
 304
 305 #ifdef CONFIG_VM_EVENT_COUNTERS
 306         update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_IN,
 307                                 pages_to_bytes(events[PSWPIN]));
 308         update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_OUT,
 309                                 pages_to_bytes(events[PSWPOUT]));
 310         update_stat(vb, idx++, VIRTIO_BALLOON_S_MAJFLT, events[PGMAJFAULT]);
 311         update_stat(vb, idx++, VIRTIO_BALLOON_S_MINFLT, events[PGFAULT]);
 312 #ifdef CONFIG_HUGETLB_PAGE
 313         update_stat(vb, idx++, VIRTIO_BALLOON_S_HTLB_PGALLOC,
 314                     events[HTLB_BUDDY_PGALLOC]);
 315         update_stat(vb, idx++, VIRTIO_BALLOON_S_HTLB_PGFAIL,
 316                     events[HTLB_BUDDY_PGALLOC_FAIL]);
 317 #endif
 318 #endif
 319         update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMFREE,
 320                                 pages_to_bytes(i.freeram));
 321         update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMTOT,
 322                                 pages_to_bytes(i.totalram));
 323         update_stat(vb, idx++, VIRTIO_BALLOON_S_AVAIL,
 324                                 pages_to_bytes(available));
 325         update_stat(vb, idx++, VIRTIO_BALLOON_S_CACHES,
 326                                 pages_to_bytes(caches));
 327
 328         return idx;
 329 }
 330
 331 /*
 332  * While most virtqueues communicate guest-initiated requests to the hypervisor,
 333  * the stats queue operates in reverse.  The driver initializes the virtqueue
 334  * with a single buffer.  From that point forward, all conversations consist of
 335  * a hypervisor request (a call to this function) which directs us to refill
 336  * the virtqueue with a fresh stats buffer.  Since stats collection can sleep,
 337  * we delegate the job to a freezable workqueue that will do the actual work via
 338  * stats_handle_request().
 339  */
 340 static void stats_request(struct virtqueue *vq)
 341 {
 342         struct virtio_balloon *vb = vq->vdev->priv;
 343
 344         spin_lock(&vb->stop_update_lock);
 345         if (!vb->stop_update)
 346                 queue_work(system_freezable_wq, &vb->update_balloon_stats_work);
 347         spin_unlock(&vb->stop_update_lock);
 348 }
 349
 350 static void stats_handle_request(struct virtio_balloon *vb)
 351 {
 352         struct virtqueue *vq;
 353         struct scatterlist sg;
 354         unsigned int len, num_stats;
 355
 356         num_stats = update_balloon_stats(vb);
 357
 358         vq = vb->stats_vq;
 359         if (!virtqueue_get_buf(vq, &len))
 360                 return;
 361         sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
 362         virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL);
 363         virtqueue_kick(vq);
 364 }
 365
 366 static inline s64 towards_target(struct virtio_balloon *vb)
 367 {
 368         s64 target;
 369         u32 num_pages;
 370
 371         virtio_cread(vb->vdev, struct virtio_balloon_config, num_pages,
 372                      &num_pages);
 373
 374         /* Legacy balloon config space is LE, unlike all other devices. */
 375         if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1))
 376                 num_pages = le32_to_cpu((__force __le32)num_pages);
 377
 378         target = num_pages;
 379         return target - vb->num_pages;
 380 }
 381
 382 /* Gives back @num_to_return blocks of free pages to mm. */
 383 static unsigned long return_free_pages_to_mm(struct virtio_balloon *vb,
 384                                              unsigned long num_to_return)
 385 {
 386         struct page *page;
 387         unsigned long num_returned;
 388
 389         spin_lock_irq(&vb->free_page_list_lock);
 390         for (num_returned = 0; num_returned < num_to_return; num_returned++) {
 391                 page = balloon_page_pop(&vb->free_page_list);
 392                 if (!page)
 393                         break;
 394                 free_pages((unsigned long)page_address(page),
 395                            VIRTIO_BALLOON_FREE_PAGE_ORDER);
 396         }
 397         vb->num_free_page_blocks -= num_returned;
 398         spin_unlock_irq(&vb->free_page_list_lock);
 399
 400         return num_returned;
 401 }
 402
 403 static void virtio_balloon_queue_free_page_work(struct virtio_balloon *vb)
 404 {
 405         if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
 406                 return;
 407
 408         /* No need to queue the work if the bit was already set. */
 409         if (test_and_set_bit(VIRTIO_BALLOON_CONFIG_READ_CMD_ID,
 410                              &vb->config_read_bitmap))
 411                 return;
 412
 413         queue_work(vb->balloon_wq, &vb->report_free_page_work);
 414 }
 415
 416 static void virtballoon_changed(struct virtio_device *vdev)
 417 {
 418         struct virtio_balloon *vb = vdev->priv;
 419         unsigned long flags;
 420
 421         spin_lock_irqsave(&vb->stop_update_lock, flags);
 422         if (!vb->stop_update) {
 423                 queue_work(system_freezable_wq,
 424                            &vb->update_balloon_size_work);
 425                 virtio_balloon_queue_free_page_work(vb);
 426         }
 427         spin_unlock_irqrestore(&vb->stop_update_lock, flags);
 428 }
 429
 430 static void update_balloon_size(struct virtio_balloon *vb)
 431 {
 432         u32 actual = vb->num_pages;
 433
 434         /* Legacy balloon config space is LE, unlike all other devices. */
 435         if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1))
 436                 actual = (__force u32)cpu_to_le32(actual);
 437
 438         virtio_cwrite(vb->vdev, struct virtio_balloon_config, actual,
 439                       &actual);
 440 }
 441
 442 static void update_balloon_stats_func(struct work_struct *work)
 443 {
 444         struct virtio_balloon *vb;
 445
 446         vb = container_of(work, struct virtio_balloon,
 447                           update_balloon_stats_work);
 448         stats_handle_request(vb);
 449 }
 450
 451 static void update_balloon_size_func(struct work_struct *work)
 452 {
 453         struct virtio_balloon *vb;
 454         s64 diff;
 455
 456         vb = container_of(work, struct virtio_balloon,
 457                           update_balloon_size_work);
 458         diff = towards_target(vb);
 459
 460         if (!diff)
 461                 return;
 462
 463         if (diff > 0)
 464                 diff -= fill_balloon(vb, diff);
 465         else
 466                 diff += leak_balloon(vb, -diff);
 467         update_balloon_size(vb);
 468
 469         if (diff)
 470                 queue_work(system_freezable_wq, work);
 471 }
 472
 473 static int init_vqs(struct virtio_balloon *vb)
 474 {
 475         struct virtqueue *vqs[VIRTIO_BALLOON_VQ_MAX];
 476         vq_callback_t *callbacks[VIRTIO_BALLOON_VQ_MAX];
 477         const char *names[VIRTIO_BALLOON_VQ_MAX];
 478         int err;
 479
 480         /*
 481          * Inflateq and deflateq are used unconditionally. The names[]
 482          * will be NULL if the related feature is not enabled, which will
 483          * cause no allocation for the corresponding virtqueue in find_vqs.
 484          */
 485         callbacks[VIRTIO_BALLOON_VQ_INFLATE] = balloon_ack;
 486         names[VIRTIO_BALLOON_VQ_INFLATE] = "inflate";
 487         callbacks[VIRTIO_BALLOON_VQ_DEFLATE] = balloon_ack;
 488         names[VIRTIO_BALLOON_VQ_DEFLATE] = "deflate";
 489         names[VIRTIO_BALLOON_VQ_STATS] = NULL;
 490         names[VIRTIO_BALLOON_VQ_FREE_PAGE] = NULL;
 491
 492         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
 493                 names[VIRTIO_BALLOON_VQ_STATS] = "stats";
 494                 callbacks[VIRTIO_BALLOON_VQ_STATS] = stats_request;
 495         }
 496
 497         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
 498                 names[VIRTIO_BALLOON_VQ_FREE_PAGE] = "free_page_vq";
 499                 callbacks[VIRTIO_BALLOON_VQ_FREE_PAGE] = NULL;
 500         }
 501
 502         err = vb->vdev->config->find_vqs(vb->vdev, VIRTIO_BALLOON_VQ_MAX,
 503                                          vqs, callbacks, names, NULL, NULL);
 504         if (err)
 505                 return err;
 506
 507         vb->inflate_vq = vqs[VIRTIO_BALLOON_VQ_INFLATE];
 508         vb->deflate_vq = vqs[VIRTIO_BALLOON_VQ_DEFLATE];
 509         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) {
 510                 struct scatterlist sg;
 511                 unsigned int num_stats;
 512                 vb->stats_vq = vqs[VIRTIO_BALLOON_VQ_STATS];
 513
 514                 /*
 515                  * Prime this virtqueue with one buffer so the hypervisor can
 516                  * use it to signal us later (it can't be broken yet!).
 517                  */
 518                 num_stats = update_balloon_stats(vb);
 519
 520                 sg_init_one(&sg, vb->stats, sizeof(vb->stats[0]) * num_stats);
 521                 err = virtqueue_add_outbuf(vb->stats_vq, &sg, 1, vb,
 522                                            GFP_KERNEL);
 523                 if (err) {
 524                         dev_warn(&vb->vdev->dev, "%s: add stat_vq failed\n",
 525                                  __func__);
 526                         return err;
 527                 }
 528                 virtqueue_kick(vb->stats_vq);
 529         }
 530
 531         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
 532                 vb->free_page_vq = vqs[VIRTIO_BALLOON_VQ_FREE_PAGE];
 533
 534         return 0;
 535 }
 536
 537 static u32 virtio_balloon_cmd_id_received(struct virtio_balloon *vb)
 538 {
 539         if (test_and_clear_bit(VIRTIO_BALLOON_CONFIG_READ_CMD_ID,
 540                                &vb->config_read_bitmap))
 541                 virtio_cread(vb->vdev, struct virtio_balloon_config,
 542                              free_page_report_cmd_id,
 543                              &vb->cmd_id_received_cache);
 544
 545         return vb->cmd_id_received_cache;
 546 }
 547
 548 static int send_cmd_id_start(struct virtio_balloon *vb)
 549 {
 550         struct scatterlist sg;
 551         struct virtqueue *vq = vb->free_page_vq;
 552         int err, unused;
 553
 554         /* Detach all the used buffers from the vq */
 555         while (virtqueue_get_buf(vq, &unused))
 556                 ;
 557
 558         vb->cmd_id_active = virtio32_to_cpu(vb->vdev,
 559                                         virtio_balloon_cmd_id_received(vb));
 560         sg_init_one(&sg, &vb->cmd_id_active, sizeof(vb->cmd_id_active));
 561         err = virtqueue_add_outbuf(vq, &sg, 1, &vb->cmd_id_active, GFP_KERNEL);
 562         if (!err)
 563                 virtqueue_kick(vq);
 564         return err;
 565 }
 566
 567 static int send_cmd_id_stop(struct virtio_balloon *vb)
 568 {
 569         struct scatterlist sg;
 570         struct virtqueue *vq = vb->free_page_vq;
 571         int err, unused;
 572
 573         /* Detach all the used buffers from the vq */
 574         while (virtqueue_get_buf(vq, &unused))
 575                 ;
 576
 577         sg_init_one(&sg, &vb->cmd_id_stop, sizeof(vb->cmd_id_stop));
 578         err = virtqueue_add_outbuf(vq, &sg, 1, &vb->cmd_id_stop, GFP_KERNEL);
 579         if (!err)
 580                 virtqueue_kick(vq);
 581         return err;
 582 }
 583
 584 static int get_free_page_and_send(struct virtio_balloon *vb)
 585 {
 586         struct virtqueue *vq = vb->free_page_vq;
 587         struct page *page;
 588         struct scatterlist sg;
 589         int err, unused;
 590         void *p;
 591
 592         /* Detach all the used buffers from the vq */
 593         while (virtqueue_get_buf(vq, &unused))
 594                 ;
 595
 596         page = alloc_pages(VIRTIO_BALLOON_FREE_PAGE_ALLOC_FLAG,
 597                            VIRTIO_BALLOON_FREE_PAGE_ORDER);
 598         /*
 599          * When the allocation returns NULL, it indicates that we have got all
 600          * the possible free pages, so return -EINTR to stop.
 601          */
 602         if (!page)
 603                 return -EINTR;
 604
 605         p = page_address(page);
 606         sg_init_one(&sg, p, VIRTIO_BALLOON_FREE_PAGE_SIZE);
 607         /* There is always 1 entry reserved for the cmd id to use. */
 608         if (vq->num_free > 1) {
 609                 err = virtqueue_add_inbuf(vq, &sg, 1, p, GFP_KERNEL);
 610                 if (unlikely(err)) {
 611                         free_pages((unsigned long)p,
 612                                    VIRTIO_BALLOON_FREE_PAGE_ORDER);
 613                         return err;
 614                 }
 615                 virtqueue_kick(vq);
 616                 spin_lock_irq(&vb->free_page_list_lock);
 617                 balloon_page_push(&vb->free_page_list, page);
 618                 vb->num_free_page_blocks++;
 619                 spin_unlock_irq(&vb->free_page_list_lock);
 620         } else {
 621                 /*
 622                  * The vq has no available entry to add this page block, so
 623                  * just free it.
 624                  */
 625                 free_pages((unsigned long)p, VIRTIO_BALLOON_FREE_PAGE_ORDER);
 626         }
 627
 628         return 0;
 629 }
 630
 631 static int send_free_pages(struct virtio_balloon *vb)
 632 {
 633         int err;
 634         u32 cmd_id_active;
 635
 636         while (1) {
 637                 /*
 638                  * If a stop id or a new cmd id was just received from host,
 639                  * stop the reporting.
 640                  */
 641                 cmd_id_active = virtio32_to_cpu(vb->vdev, vb->cmd_id_active);
 642                 if (unlikely(cmd_id_active !=
 643                              virtio_balloon_cmd_id_received(vb)))
 644                         break;
 645
 646                 /*
 647                  * The free page blocks are allocated and sent to host one by
 648                  * one.
 649                  */
 650                 err = get_free_page_and_send(vb);
 651                 if (err == -EINTR)
 652                         break;
 653                 else if (unlikely(err))
 654                         return err;
 655         }
 656
 657         return 0;
 658 }
 659
 660 static void virtio_balloon_report_free_page(struct virtio_balloon *vb)
 661 {
 662         int err;
 663         struct device *dev = &vb->vdev->dev;
 664
 665         /* Start by sending the received cmd id to host with an outbuf. */
 666         err = send_cmd_id_start(vb);
 667         if (unlikely(err))
 668                 dev_err(dev, "Failed to send a start id, err = %d\n", err);
 669
 670         err = send_free_pages(vb);
 671         if (unlikely(err))
 672                 dev_err(dev, "Failed to send a free page, err = %d\n", err);
 673
 674         /* End by sending a stop id to host with an outbuf. */
 675         err = send_cmd_id_stop(vb);
 676         if (unlikely(err))
 677                 dev_err(dev, "Failed to send a stop id, err = %d\n", err);
 678 }
 679
 680 static void report_free_page_func(struct work_struct *work)
 681 {
 682         struct virtio_balloon *vb = container_of(work, struct virtio_balloon,
 683                                                  report_free_page_work);
 684         u32 cmd_id_received;
 685
 686         cmd_id_received = virtio_balloon_cmd_id_received(vb);
 687         if (cmd_id_received == VIRTIO_BALLOON_CMD_ID_DONE) {
 688                 /* Pass ULONG_MAX to give back all the free pages */
 689                 return_free_pages_to_mm(vb, ULONG_MAX);
 690         } else if (cmd_id_received != VIRTIO_BALLOON_CMD_ID_STOP &&
 691                    cmd_id_received !=
 692                    virtio32_to_cpu(vb->vdev, vb->cmd_id_active)) {
 693                 virtio_balloon_report_free_page(vb);
 694         }
 695 }
 696
 697 #ifdef CONFIG_BALLOON_COMPACTION
 698 /*
 699  * virtballoon_migratepage - perform the balloon page migration on behalf of
 700  *                           a compation thread.     (called under page lock)
 701  * @vb_dev_info: the balloon device
 702  * @newpage: page that will replace the isolated page after migration finishes.
 703  * @page   : the isolated (old) page that is about to be migrated to newpage.
 704  * @mode   : compaction mode -- not used for balloon page migration.
 705  *
 706  * After a ballooned page gets isolated by compaction procedures, this is the
 707  * function that performs the page migration on behalf of a compaction thread
 708  * The page migration for virtio balloon is done in a simple swap fashion which
 709  * follows these two macro steps:
 710  *  1) insert newpage into vb->pages list and update the host about it;
 711  *  2) update the host about the old page removed from vb->pages list;
 712  *
 713  * This function preforms the balloon page migration task.
 714  * Called through balloon_mapping->a_ops->migratepage
 715  */
 716 static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info,
 717                 struct page *newpage, struct page *page, enum migrate_mode mode)
 718 {
 719         struct virtio_balloon *vb = container_of(vb_dev_info,
 720                         struct virtio_balloon, vb_dev_info);
 721         unsigned long flags;
 722
 723         /*
 724          * In order to avoid lock contention while migrating pages concurrently
 725          * to leak_balloon() or fill_balloon() we just give up the balloon_lock
 726          * this turn, as it is easier to retry the page migration later.
 727          * This also prevents fill_balloon() getting stuck into a mutex
 728          * recursion in the case it ends up triggering memory compaction
 729          * while it is attempting to inflate the ballon.
 730          */
 731         if (!mutex_trylock(&vb->balloon_lock))
 732                 return -EAGAIN;
 733
 734         get_page(newpage); /* balloon reference */
 735
 736         /* balloon's page migration 1st step  -- inflate "newpage" */
 737         spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
 738         balloon_page_insert(vb_dev_info, newpage);
 739         vb_dev_info->isolated_pages--;
 740         __count_vm_event(BALLOON_MIGRATE);
 741         spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
 742         vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
 743         set_page_pfns(vb, vb->pfns, newpage);
 744         tell_host(vb, vb->inflate_vq);
 745
 746         /* balloon's page migration 2nd step -- deflate "page" */
 747         spin_lock_irqsave(&vb_dev_info->pages_lock, flags);
 748         balloon_page_delete(page);
 749         spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags);
 750         vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE;
 751         set_page_pfns(vb, vb->pfns, page);
 752         tell_host(vb, vb->deflate_vq);
 753
 754         mutex_unlock(&vb->balloon_lock);
 755
 756         put_page(page); /* balloon reference */
 757
 758         return MIGRATEPAGE_SUCCESS;
 759 }
 760
 761 static struct dentry *balloon_mount(struct file_system_type *fs_type,
 762                 int flags, const char *dev_name, void *data)
 763 {
 764         static const struct dentry_operations ops = {
 765                 .d_dname = simple_dname,
 766         };
 767
 768         return mount_pseudo(fs_type, "balloon-kvm:", NULL, &ops,
 769                                 BALLOON_KVM_MAGIC);
 770 }
 771
 772 static struct file_system_type balloon_fs = {
 773         .name           = "balloon-kvm",
 774         .mount          = balloon_mount,
 775         .kill_sb        = kill_anon_super,
 776 };
 777
 778 #endif /* CONFIG_BALLOON_COMPACTION */
 779
 780 static unsigned long shrink_free_pages(struct virtio_balloon *vb,
 781                                        unsigned long pages_to_free)
 782 {
 783         unsigned long blocks_to_free, blocks_freed;
 784
 785         pages_to_free = round_up(pages_to_free,
 786                                  1 << VIRTIO_BALLOON_FREE_PAGE_ORDER);
 787         blocks_to_free = pages_to_free >> VIRTIO_BALLOON_FREE_PAGE_ORDER;
 788         blocks_freed = return_free_pages_to_mm(vb, blocks_to_free);
 789
 790         return blocks_freed << VIRTIO_BALLOON_FREE_PAGE_ORDER;
 791 }
 792
 793 static unsigned long shrink_balloon_pages(struct virtio_balloon *vb,
 794                                           unsigned long pages_to_free)
 795 {
 796         unsigned long pages_freed = 0;
 797
 798         /*
 799          * One invocation of leak_balloon can deflate at most
 800          * VIRTIO_BALLOON_ARRAY_PFNS_MAX balloon pages, so we call it
 801          * multiple times to deflate pages till reaching pages_to_free.
 802          */
 803         while (vb->num_pages && pages_to_free) {
 804                 pages_freed += leak_balloon(vb, pages_to_free) /
 805                                         VIRTIO_BALLOON_PAGES_PER_PAGE;
 806                 pages_to_free -= pages_freed;
 807         }
 808         update_balloon_size(vb);
 809
 810         return pages_freed;
 811 }
 812
 813 static unsigned long virtio_balloon_shrinker_scan(struct shrinker *shrinker,
 814                                                   struct shrink_control *sc)
 815 {
 816         unsigned long pages_to_free, pages_freed = 0;
 817         struct virtio_balloon *vb = container_of(shrinker,
 818                                         struct virtio_balloon, shrinker);
 819
 820         pages_to_free = sc->nr_to_scan * VIRTIO_BALLOON_PAGES_PER_PAGE;
 821
 822         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
 823                 pages_freed = shrink_free_pages(vb, pages_to_free);
 824
 825         if (pages_freed >= pages_to_free)
 826                 return pages_freed;
 827
 828         pages_freed += shrink_balloon_pages(vb, pages_to_free - pages_freed);
 829
 830         return pages_freed;
 831 }
 832
 833 static unsigned long virtio_balloon_shrinker_count(struct shrinker *shrinker,
 834                                                    struct shrink_control *sc)
 835 {
 836         struct virtio_balloon *vb = container_of(shrinker,
 837                                         struct virtio_balloon, shrinker);
 838         unsigned long count;
 839
 840         count = vb->num_pages / VIRTIO_BALLOON_PAGES_PER_PAGE;
 841         count += vb->num_free_page_blocks >> VIRTIO_BALLOON_FREE_PAGE_ORDER;
 842
 843         return count;
 844 }
 845
 846 static void virtio_balloon_unregister_shrinker(struct virtio_balloon *vb)
 847 {
 848         unregister_shrinker(&vb->shrinker);
 849 }
 850
 851 static int virtio_balloon_register_shrinker(struct virtio_balloon *vb)
 852 {
 853         vb->shrinker.scan_objects = virtio_balloon_shrinker_scan;
 854         vb->shrinker.count_objects = virtio_balloon_shrinker_count;
 855         vb->shrinker.seeks = DEFAULT_SEEKS;
 856
 857         return register_shrinker(&vb->shrinker);
 858 }
 859
 860 static int virtballoon_probe(struct virtio_device *vdev)
 861 {
 862         struct virtio_balloon *vb;
 863         __u32 poison_val;
 864         int err;
 865
 866         if (!vdev->config->get) {
 867                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
 868                         __func__);
 869                 return -EINVAL;
 870         }
 871
 872         vdev->priv = vb = kzalloc(sizeof(*vb), GFP_KERNEL);
 873         if (!vb) {
 874                 err = -ENOMEM;
 875                 goto out;
 876         }
 877
 878         INIT_WORK(&vb->update_balloon_stats_work, update_balloon_stats_func);
 879         INIT_WORK(&vb->update_balloon_size_work, update_balloon_size_func);
 880         spin_lock_init(&vb->stop_update_lock);
 881         mutex_init(&vb->balloon_lock);
 882         init_waitqueue_head(&vb->acked);
 883         vb->vdev = vdev;
 884
 885         balloon_devinfo_init(&vb->vb_dev_info);
 886
 887         err = init_vqs(vb);
 888         if (err)
 889                 goto out_free_vb;
 890
 891 #ifdef CONFIG_BALLOON_COMPACTION
 892         balloon_mnt = kern_mount(&balloon_fs);
 893         if (IS_ERR(balloon_mnt)) {
 894                 err = PTR_ERR(balloon_mnt);
 895                 goto out_del_vqs;
 896         }
 897
 898         vb->vb_dev_info.migratepage = virtballoon_migratepage;
 899         vb->vb_dev_info.inode = alloc_anon_inode(balloon_mnt->mnt_sb);
 900         if (IS_ERR(vb->vb_dev_info.inode)) {
 901                 err = PTR_ERR(vb->vb_dev_info.inode);
 902                 kern_unmount(balloon_mnt);
 903                 goto out_del_vqs;
 904         }
 905         vb->vb_dev_info.inode->i_mapping->a_ops = &balloon_aops;
 906 #endif
 907         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
 908                 /*
 909                  * There is always one entry reserved for cmd id, so the ring
 910                  * size needs to be at least two to report free page hints.
 911                  */
 912                 if (virtqueue_get_vring_size(vb->free_page_vq) < 2) {
 913                         err = -ENOSPC;
 914                         goto out_del_vqs;
 915                 }
 916                 vb->balloon_wq = alloc_workqueue("balloon-wq",
 917                                         WQ_FREEZABLE | WQ_CPU_INTENSIVE, 0);
 918                 if (!vb->balloon_wq) {
 919                         err = -ENOMEM;
 920                         goto out_del_vqs;
 921                 }
 922                 INIT_WORK(&vb->report_free_page_work, report_free_page_func);
 923                 vb->cmd_id_received_cache = VIRTIO_BALLOON_CMD_ID_STOP;
 924                 vb->cmd_id_active = cpu_to_virtio32(vb->vdev,
 925                                                   VIRTIO_BALLOON_CMD_ID_STOP);
 926                 vb->cmd_id_stop = cpu_to_virtio32(vb->vdev,
 927                                                   VIRTIO_BALLOON_CMD_ID_STOP);
 928                 spin_lock_init(&vb->free_page_list_lock);
 929                 INIT_LIST_HEAD(&vb->free_page_list);
 930                 if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_PAGE_POISON)) {
 931                         memset(&poison_val, PAGE_POISON, sizeof(poison_val));
 932                         virtio_cwrite(vb->vdev, struct virtio_balloon_config,
 933                                       poison_val, &poison_val);
 934                 }
 935         }
 936         /*
 937          * We continue to use VIRTIO_BALLOON_F_DEFLATE_ON_OOM to decide if a
 938          * shrinker needs to be registered to relieve memory pressure.
 939          */
 940         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) {
 941                 err = virtio_balloon_register_shrinker(vb);
 942                 if (err)
 943                         goto out_del_balloon_wq;
 944         }
 945         virtio_device_ready(vdev);
 946
 947         if (towards_target(vb))
 948                 virtballoon_changed(vdev);
 949         return 0;
 950
 951 out_del_balloon_wq:
 952         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT))
 953                 destroy_workqueue(vb->balloon_wq);
 954 out_del_vqs:
 955         vdev->config->del_vqs(vdev);
 956 out_free_vb:
 957         kfree(vb);
 958 out:
 959         return err;
 960 }
 961
 962 static void remove_common(struct virtio_balloon *vb)
 963 {
 964         /* There might be pages left in the balloon: free them. */
 965         while (vb->num_pages)
 966                 leak_balloon(vb, vb->num_pages);
 967         update_balloon_size(vb);
 968
 969         /* Now we reset the device so we can clean up the queues. */
 970         vb->vdev->config->reset(vb->vdev);
 971
 972         vb->vdev->config->del_vqs(vb->vdev);
 973 }
 974
 975 static void virtballoon_remove(struct virtio_device *vdev)
 976 {
 977         struct virtio_balloon *vb = vdev->priv;
 978
 979         if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
 980                 virtio_balloon_unregister_shrinker(vb);
 981         spin_lock_irq(&vb->stop_update_lock);
 982         vb->stop_update = true;
 983         spin_unlock_irq(&vb->stop_update_lock);
 984         cancel_work_sync(&vb->update_balloon_size_work);
 985         cancel_work_sync(&vb->update_balloon_stats_work);
 986
 987         if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT)) {
 988                 cancel_work_sync(&vb->report_free_page_work);
 989                 destroy_workqueue(vb->balloon_wq);
 990         }
 991
 992         remove_common(vb);
 993 #ifdef CONFIG_BALLOON_COMPACTION
 994         if (vb->vb_dev_info.inode)
 995                 iput(vb->vb_dev_info.inode);
 996
 997         kern_unmount(balloon_mnt);
 998 #endif
 999         kfree(vb);
1000 }
1001
1002 #ifdef CONFIG_PM_SLEEP
1003 static int virtballoon_freeze(struct virtio_device *vdev)
1004 {
1005         struct virtio_balloon *vb = vdev->priv;
1006
1007         /*
1008          * The workqueue is already frozen by the PM core before this
1009          * function is called.
1010          */
1011         remove_common(vb);
1012         return 0;
1013 }
1014
1015 static int virtballoon_restore(struct virtio_device *vdev)
1016 {
1017         struct virtio_balloon *vb = vdev->priv;
1018         int ret;
1019
1020         ret = init_vqs(vdev->priv);
1021         if (ret)
1022                 return ret;
1023
1024         virtio_device_ready(vdev);
1025
1026         if (towards_target(vb))
1027                 virtballoon_changed(vdev);
1028         update_balloon_size(vb);
1029         return 0;
1030 }
1031 #endif
1032
1033 static int virtballoon_validate(struct virtio_device *vdev)
1034 {
1035         if (!page_poisoning_enabled())
1036                 __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_PAGE_POISON);
1037
1038         __virtio_clear_bit(vdev, VIRTIO_F_IOMMU_PLATFORM);
1039         return 0;
1040 }
1041
1042 static unsigned int features[] = {
1043         VIRTIO_BALLOON_F_MUST_TELL_HOST,
1044         VIRTIO_BALLOON_F_STATS_VQ,
1045         VIRTIO_BALLOON_F_DEFLATE_ON_OOM,
1046         VIRTIO_BALLOON_F_FREE_PAGE_HINT,
1047         VIRTIO_BALLOON_F_PAGE_POISON,
1048 };
1049
1050 static struct virtio_driver virtio_balloon_driver = {
1051         .feature_table = features,
1052         .feature_table_size = ARRAY_SIZE(features),
1053         .driver.name =  KBUILD_MODNAME,
1054         .driver.owner = THIS_MODULE,
1055         .id_table =     id_table,
1056         .validate =     virtballoon_validate,
1057         .probe =        virtballoon_probe,
1058         .remove =       virtballoon_remove,
1059         .config_changed = virtballoon_changed,
1060 #ifdef CONFIG_PM_SLEEP
1061         .freeze =       virtballoon_freeze,
1062         .restore =      virtballoon_restore,
1063 #endif
1064 };
1065
1066 module_virtio_driver(virtio_balloon_driver);
1067 MODULE_DEVICE_TABLE(virtio, id_table);
1068 MODULE_DESCRIPTION("Virtio balloon driver");
1069 MODULE_LICENSE("GPL");