1 /* 2 * Virtio balloon implementation, inspired by Dor Laor and Marcelo 3 * Tosatti's implementations. 4 * 5 * Copyright 2008 Rusty Russell IBM Corporation 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License as published by 9 * the Free Software Foundation; either version 2 of the License, or 10 * (at your option) any later version. 11 * 12 * This program is distributed in the hope that it will be useful, 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 15 * GNU General Public License for more details. 16 * 17 * You should have received a copy of the GNU General Public License 18 * along with this program; if not, write to the Free Software 19 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 20 */ 21 22 #include <linux/virtio.h> 23 #include <linux/virtio_balloon.h> 24 #include <linux/swap.h> 25 #include <linux/kthread.h> 26 #include <linux/freezer.h> 27 #include <linux/delay.h> 28 #include <linux/slab.h> 29 #include <linux/module.h> 30 #include <linux/balloon_compaction.h> 31 #include <linux/oom.h> 32 #include <linux/wait.h> 33 34 /* 35 * Balloon device works in 4K page units. So each page is pointed to by 36 * multiple balloon pages. All memory counters in this driver are in balloon 37 * page units. 38 */ 39 #define VIRTIO_BALLOON_PAGES_PER_PAGE (unsigned)(PAGE_SIZE >> VIRTIO_BALLOON_PFN_SHIFT) 40 #define VIRTIO_BALLOON_ARRAY_PFNS_MAX 256 41 #define OOM_VBALLOON_DEFAULT_PAGES 256 42 #define VIRTBALLOON_OOM_NOTIFY_PRIORITY 80 43 44 static int oom_pages = OOM_VBALLOON_DEFAULT_PAGES; 45 module_param(oom_pages, int, S_IRUSR | S_IWUSR); 46 MODULE_PARM_DESC(oom_pages, "pages to free on OOM"); 47 48 struct virtio_balloon { 49 struct virtio_device *vdev; 50 struct virtqueue *inflate_vq, *deflate_vq, *stats_vq; 51 52 /* Where the ballooning thread waits for config to change. */ 53 wait_queue_head_t config_change; 54 55 /* The thread servicing the balloon. */ 56 struct task_struct *thread; 57 58 /* Waiting for host to ack the pages we released. */ 59 wait_queue_head_t acked; 60 61 /* Number of balloon pages we've told the Host we're not using. */ 62 unsigned int num_pages; 63 /* 64 * The pages we've told the Host we're not using are enqueued 65 * at vb_dev_info->pages list. 66 * Each page on this list adds VIRTIO_BALLOON_PAGES_PER_PAGE 67 * to num_pages above. 68 */ 69 struct balloon_dev_info vb_dev_info; 70 71 /* Synchronize access/update to this struct virtio_balloon elements */ 72 struct mutex balloon_lock; 73 74 /* The array of pfns we tell the Host about. */ 75 unsigned int num_pfns; 76 u32 pfns[VIRTIO_BALLOON_ARRAY_PFNS_MAX]; 77 78 /* Memory statistics */ 79 int need_stats_update; 80 struct virtio_balloon_stat stats[VIRTIO_BALLOON_S_NR]; 81 82 /* To register callback in oom notifier call chain */ 83 struct notifier_block nb; 84 }; 85 86 static struct virtio_device_id id_table[] = { 87 { VIRTIO_ID_BALLOON, VIRTIO_DEV_ANY_ID }, 88 { 0 }, 89 }; 90 91 static u32 page_to_balloon_pfn(struct page *page) 92 { 93 unsigned long pfn = page_to_pfn(page); 94 95 BUILD_BUG_ON(PAGE_SHIFT < VIRTIO_BALLOON_PFN_SHIFT); 96 /* Convert pfn from Linux page size to balloon page size. */ 97 return pfn * VIRTIO_BALLOON_PAGES_PER_PAGE; 98 } 99 100 static struct page *balloon_pfn_to_page(u32 pfn) 101 { 102 BUG_ON(pfn % VIRTIO_BALLOON_PAGES_PER_PAGE); 103 return pfn_to_page(pfn / VIRTIO_BALLOON_PAGES_PER_PAGE); 104 } 105 106 static void balloon_ack(struct virtqueue *vq) 107 { 108 struct virtio_balloon *vb = vq->vdev->priv; 109 110 wake_up(&vb->acked); 111 } 112 113 static void tell_host(struct virtio_balloon *vb, struct virtqueue *vq) 114 { 115 struct scatterlist sg; 116 unsigned int len; 117 118 sg_init_one(&sg, vb->pfns, sizeof(vb->pfns[0]) * vb->num_pfns); 119 120 /* We should always be able to add one buffer to an empty queue. */ 121 virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); 122 virtqueue_kick(vq); 123 124 /* When host has read buffer, this completes via balloon_ack */ 125 wait_event(vb->acked, virtqueue_get_buf(vq, &len)); 126 } 127 128 static void set_page_pfns(u32 pfns[], struct page *page) 129 { 130 unsigned int i; 131 132 /* Set balloon pfns pointing at this page. 133 * Note that the first pfn points at start of the page. */ 134 for (i = 0; i < VIRTIO_BALLOON_PAGES_PER_PAGE; i++) 135 pfns[i] = page_to_balloon_pfn(page) + i; 136 } 137 138 static void fill_balloon(struct virtio_balloon *vb, size_t num) 139 { 140 struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; 141 142 /* We can only do one array worth at a time. */ 143 num = min(num, ARRAY_SIZE(vb->pfns)); 144 145 mutex_lock(&vb->balloon_lock); 146 for (vb->num_pfns = 0; vb->num_pfns < num; 147 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) { 148 struct page *page = balloon_page_enqueue(vb_dev_info); 149 150 if (!page) { 151 dev_info_ratelimited(&vb->vdev->dev, 152 "Out of puff! Can't get %u pages\n", 153 VIRTIO_BALLOON_PAGES_PER_PAGE); 154 /* Sleep for at least 1/5 of a second before retry. */ 155 msleep(200); 156 break; 157 } 158 set_page_pfns(vb->pfns + vb->num_pfns, page); 159 vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE; 160 if (!virtio_has_feature(vb->vdev, 161 VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) 162 adjust_managed_page_count(page, -1); 163 } 164 165 /* Did we get any? */ 166 if (vb->num_pfns != 0) 167 tell_host(vb, vb->inflate_vq); 168 mutex_unlock(&vb->balloon_lock); 169 } 170 171 static void release_pages_balloon(struct virtio_balloon *vb) 172 { 173 unsigned int i; 174 175 /* Find pfns pointing at start of each page, get pages and free them. */ 176 for (i = 0; i < vb->num_pfns; i += VIRTIO_BALLOON_PAGES_PER_PAGE) { 177 struct page *page = balloon_pfn_to_page(vb->pfns[i]); 178 if (!virtio_has_feature(vb->vdev, 179 VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) 180 adjust_managed_page_count(page, 1); 181 put_page(page); /* balloon reference */ 182 } 183 } 184 185 static unsigned leak_balloon(struct virtio_balloon *vb, size_t num) 186 { 187 unsigned num_freed_pages; 188 struct page *page; 189 struct balloon_dev_info *vb_dev_info = &vb->vb_dev_info; 190 191 /* We can only do one array worth at a time. */ 192 num = min(num, ARRAY_SIZE(vb->pfns)); 193 194 mutex_lock(&vb->balloon_lock); 195 for (vb->num_pfns = 0; vb->num_pfns < num; 196 vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE) { 197 page = balloon_page_dequeue(vb_dev_info); 198 if (!page) 199 break; 200 set_page_pfns(vb->pfns + vb->num_pfns, page); 201 vb->num_pages -= VIRTIO_BALLOON_PAGES_PER_PAGE; 202 } 203 204 num_freed_pages = vb->num_pfns; 205 /* 206 * Note that if 207 * virtio_has_feature(vdev, VIRTIO_BALLOON_F_MUST_TELL_HOST); 208 * is true, we *have* to do it in this order 209 */ 210 if (vb->num_pfns != 0) 211 tell_host(vb, vb->deflate_vq); 212 mutex_unlock(&vb->balloon_lock); 213 release_pages_balloon(vb); 214 return num_freed_pages; 215 } 216 217 static inline void update_stat(struct virtio_balloon *vb, int idx, 218 u16 tag, u64 val) 219 { 220 BUG_ON(idx >= VIRTIO_BALLOON_S_NR); 221 vb->stats[idx].tag = cpu_to_virtio16(vb->vdev, tag); 222 vb->stats[idx].val = cpu_to_virtio64(vb->vdev, val); 223 } 224 225 #define pages_to_bytes(x) ((u64)(x) << PAGE_SHIFT) 226 227 static void update_balloon_stats(struct virtio_balloon *vb) 228 { 229 unsigned long events[NR_VM_EVENT_ITEMS]; 230 struct sysinfo i; 231 int idx = 0; 232 233 all_vm_events(events); 234 si_meminfo(&i); 235 236 update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_IN, 237 pages_to_bytes(events[PSWPIN])); 238 update_stat(vb, idx++, VIRTIO_BALLOON_S_SWAP_OUT, 239 pages_to_bytes(events[PSWPOUT])); 240 update_stat(vb, idx++, VIRTIO_BALLOON_S_MAJFLT, events[PGMAJFAULT]); 241 update_stat(vb, idx++, VIRTIO_BALLOON_S_MINFLT, events[PGFAULT]); 242 update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMFREE, 243 pages_to_bytes(i.freeram)); 244 update_stat(vb, idx++, VIRTIO_BALLOON_S_MEMTOT, 245 pages_to_bytes(i.totalram)); 246 } 247 248 /* 249 * While most virtqueues communicate guest-initiated requests to the hypervisor, 250 * the stats queue operates in reverse. The driver initializes the virtqueue 251 * with a single buffer. From that point forward, all conversations consist of 252 * a hypervisor request (a call to this function) which directs us to refill 253 * the virtqueue with a fresh stats buffer. Since stats collection can sleep, 254 * we notify our kthread which does the actual work via stats_handle_request(). 255 */ 256 static void stats_request(struct virtqueue *vq) 257 { 258 struct virtio_balloon *vb = vq->vdev->priv; 259 260 vb->need_stats_update = 1; 261 wake_up(&vb->config_change); 262 } 263 264 static void stats_handle_request(struct virtio_balloon *vb) 265 { 266 struct virtqueue *vq; 267 struct scatterlist sg; 268 unsigned int len; 269 270 vb->need_stats_update = 0; 271 update_balloon_stats(vb); 272 273 vq = vb->stats_vq; 274 if (!virtqueue_get_buf(vq, &len)) 275 return; 276 sg_init_one(&sg, vb->stats, sizeof(vb->stats)); 277 virtqueue_add_outbuf(vq, &sg, 1, vb, GFP_KERNEL); 278 virtqueue_kick(vq); 279 } 280 281 static void virtballoon_changed(struct virtio_device *vdev) 282 { 283 struct virtio_balloon *vb = vdev->priv; 284 285 wake_up(&vb->config_change); 286 } 287 288 static inline s64 towards_target(struct virtio_balloon *vb) 289 { 290 s64 target; 291 u32 num_pages; 292 293 virtio_cread(vb->vdev, struct virtio_balloon_config, num_pages, 294 &num_pages); 295 296 /* Legacy balloon config space is LE, unlike all other devices. */ 297 if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1)) 298 num_pages = le32_to_cpu((__force __le32)num_pages); 299 300 target = num_pages; 301 return target - vb->num_pages; 302 } 303 304 static void update_balloon_size(struct virtio_balloon *vb) 305 { 306 u32 actual = vb->num_pages; 307 308 /* Legacy balloon config space is LE, unlike all other devices. */ 309 if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1)) 310 actual = (__force u32)cpu_to_le32(actual); 311 312 virtio_cwrite(vb->vdev, struct virtio_balloon_config, actual, 313 &actual); 314 } 315 316 /* 317 * virtballoon_oom_notify - release pages when system is under severe 318 * memory pressure (called from out_of_memory()) 319 * @self : notifier block struct 320 * @dummy: not used 321 * @parm : returned - number of freed pages 322 * 323 * The balancing of memory by use of the virtio balloon should not cause 324 * the termination of processes while there are pages in the balloon. 325 * If virtio balloon manages to release some memory, it will make the 326 * system return and retry the allocation that forced the OOM killer 327 * to run. 328 */ 329 static int virtballoon_oom_notify(struct notifier_block *self, 330 unsigned long dummy, void *parm) 331 { 332 struct virtio_balloon *vb; 333 unsigned long *freed; 334 unsigned num_freed_pages; 335 336 vb = container_of(self, struct virtio_balloon, nb); 337 if (!virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM)) 338 return NOTIFY_OK; 339 340 freed = parm; 341 num_freed_pages = leak_balloon(vb, oom_pages); 342 update_balloon_size(vb); 343 *freed += num_freed_pages; 344 345 return NOTIFY_OK; 346 } 347 348 static int balloon(void *_vballoon) 349 { 350 struct virtio_balloon *vb = _vballoon; 351 DEFINE_WAIT_FUNC(wait, woken_wake_function); 352 353 set_freezable(); 354 while (!kthread_should_stop()) { 355 s64 diff; 356 357 try_to_freeze(); 358 359 add_wait_queue(&vb->config_change, &wait); 360 for (;;) { 361 if ((diff = towards_target(vb)) != 0 || 362 vb->need_stats_update || 363 kthread_should_stop() || 364 freezing(current)) 365 break; 366 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 367 } 368 remove_wait_queue(&vb->config_change, &wait); 369 370 if (vb->need_stats_update) 371 stats_handle_request(vb); 372 if (diff > 0) 373 fill_balloon(vb, diff); 374 else if (diff < 0) 375 leak_balloon(vb, -diff); 376 update_balloon_size(vb); 377 378 /* 379 * For large balloon changes, we could spend a lot of time 380 * and always have work to do. Be nice if preempt disabled. 381 */ 382 cond_resched(); 383 } 384 return 0; 385 } 386 387 static int init_vqs(struct virtio_balloon *vb) 388 { 389 struct virtqueue *vqs[3]; 390 vq_callback_t *callbacks[] = { balloon_ack, balloon_ack, stats_request }; 391 const char *names[] = { "inflate", "deflate", "stats" }; 392 int err, nvqs; 393 394 /* 395 * We expect two virtqueues: inflate and deflate, and 396 * optionally stat. 397 */ 398 nvqs = virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ) ? 3 : 2; 399 err = vb->vdev->config->find_vqs(vb->vdev, nvqs, vqs, callbacks, names); 400 if (err) 401 return err; 402 403 vb->inflate_vq = vqs[0]; 404 vb->deflate_vq = vqs[1]; 405 if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_STATS_VQ)) { 406 struct scatterlist sg; 407 vb->stats_vq = vqs[2]; 408 409 /* 410 * Prime this virtqueue with one buffer so the hypervisor can 411 * use it to signal us later (it can't be broken yet!). 412 */ 413 sg_init_one(&sg, vb->stats, sizeof vb->stats); 414 if (virtqueue_add_outbuf(vb->stats_vq, &sg, 1, vb, GFP_KERNEL) 415 < 0) 416 BUG(); 417 virtqueue_kick(vb->stats_vq); 418 } 419 return 0; 420 } 421 422 #ifdef CONFIG_BALLOON_COMPACTION 423 /* 424 * virtballoon_migratepage - perform the balloon page migration on behalf of 425 * a compation thread. (called under page lock) 426 * @vb_dev_info: the balloon device 427 * @newpage: page that will replace the isolated page after migration finishes. 428 * @page : the isolated (old) page that is about to be migrated to newpage. 429 * @mode : compaction mode -- not used for balloon page migration. 430 * 431 * After a ballooned page gets isolated by compaction procedures, this is the 432 * function that performs the page migration on behalf of a compaction thread 433 * The page migration for virtio balloon is done in a simple swap fashion which 434 * follows these two macro steps: 435 * 1) insert newpage into vb->pages list and update the host about it; 436 * 2) update the host about the old page removed from vb->pages list; 437 * 438 * This function preforms the balloon page migration task. 439 * Called through balloon_mapping->a_ops->migratepage 440 */ 441 static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info, 442 struct page *newpage, struct page *page, enum migrate_mode mode) 443 { 444 struct virtio_balloon *vb = container_of(vb_dev_info, 445 struct virtio_balloon, vb_dev_info); 446 unsigned long flags; 447 448 /* 449 * In order to avoid lock contention while migrating pages concurrently 450 * to leak_balloon() or fill_balloon() we just give up the balloon_lock 451 * this turn, as it is easier to retry the page migration later. 452 * This also prevents fill_balloon() getting stuck into a mutex 453 * recursion in the case it ends up triggering memory compaction 454 * while it is attempting to inflate the ballon. 455 */ 456 if (!mutex_trylock(&vb->balloon_lock)) 457 return -EAGAIN; 458 459 get_page(newpage); /* balloon reference */ 460 461 /* balloon's page migration 1st step -- inflate "newpage" */ 462 spin_lock_irqsave(&vb_dev_info->pages_lock, flags); 463 balloon_page_insert(vb_dev_info, newpage); 464 vb_dev_info->isolated_pages--; 465 __count_vm_event(BALLOON_MIGRATE); 466 spin_unlock_irqrestore(&vb_dev_info->pages_lock, flags); 467 vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; 468 set_page_pfns(vb->pfns, newpage); 469 tell_host(vb, vb->inflate_vq); 470 471 /* balloon's page migration 2nd step -- deflate "page" */ 472 balloon_page_delete(page); 473 vb->num_pfns = VIRTIO_BALLOON_PAGES_PER_PAGE; 474 set_page_pfns(vb->pfns, page); 475 tell_host(vb, vb->deflate_vq); 476 477 mutex_unlock(&vb->balloon_lock); 478 479 put_page(page); /* balloon reference */ 480 481 return MIGRATEPAGE_SUCCESS; 482 } 483 #endif /* CONFIG_BALLOON_COMPACTION */ 484 485 static int virtballoon_probe(struct virtio_device *vdev) 486 { 487 struct virtio_balloon *vb; 488 int err; 489 490 if (!vdev->config->get) { 491 dev_err(&vdev->dev, "%s failure: config access disabled\n", 492 __func__); 493 return -EINVAL; 494 } 495 496 vdev->priv = vb = kmalloc(sizeof(*vb), GFP_KERNEL); 497 if (!vb) { 498 err = -ENOMEM; 499 goto out; 500 } 501 502 vb->num_pages = 0; 503 mutex_init(&vb->balloon_lock); 504 init_waitqueue_head(&vb->config_change); 505 init_waitqueue_head(&vb->acked); 506 vb->vdev = vdev; 507 vb->need_stats_update = 0; 508 509 balloon_devinfo_init(&vb->vb_dev_info); 510 #ifdef CONFIG_BALLOON_COMPACTION 511 vb->vb_dev_info.migratepage = virtballoon_migratepage; 512 #endif 513 514 err = init_vqs(vb); 515 if (err) 516 goto out_free_vb; 517 518 vb->nb.notifier_call = virtballoon_oom_notify; 519 vb->nb.priority = VIRTBALLOON_OOM_NOTIFY_PRIORITY; 520 err = register_oom_notifier(&vb->nb); 521 if (err < 0) 522 goto out_oom_notify; 523 524 virtio_device_ready(vdev); 525 526 vb->thread = kthread_run(balloon, vb, "vballoon"); 527 if (IS_ERR(vb->thread)) { 528 err = PTR_ERR(vb->thread); 529 goto out_del_vqs; 530 } 531 532 return 0; 533 534 out_del_vqs: 535 unregister_oom_notifier(&vb->nb); 536 out_oom_notify: 537 vdev->config->del_vqs(vdev); 538 out_free_vb: 539 kfree(vb); 540 out: 541 return err; 542 } 543 544 static void remove_common(struct virtio_balloon *vb) 545 { 546 /* There might be pages left in the balloon: free them. */ 547 while (vb->num_pages) 548 leak_balloon(vb, vb->num_pages); 549 update_balloon_size(vb); 550 551 /* Now we reset the device so we can clean up the queues. */ 552 vb->vdev->config->reset(vb->vdev); 553 554 vb->vdev->config->del_vqs(vb->vdev); 555 } 556 557 static void virtballoon_remove(struct virtio_device *vdev) 558 { 559 struct virtio_balloon *vb = vdev->priv; 560 561 unregister_oom_notifier(&vb->nb); 562 kthread_stop(vb->thread); 563 remove_common(vb); 564 kfree(vb); 565 } 566 567 #ifdef CONFIG_PM_SLEEP 568 static int virtballoon_freeze(struct virtio_device *vdev) 569 { 570 struct virtio_balloon *vb = vdev->priv; 571 572 /* 573 * The kthread is already frozen by the PM core before this 574 * function is called. 575 */ 576 577 remove_common(vb); 578 return 0; 579 } 580 581 static int virtballoon_restore(struct virtio_device *vdev) 582 { 583 struct virtio_balloon *vb = vdev->priv; 584 int ret; 585 586 ret = init_vqs(vdev->priv); 587 if (ret) 588 return ret; 589 590 virtio_device_ready(vdev); 591 592 fill_balloon(vb, towards_target(vb)); 593 update_balloon_size(vb); 594 return 0; 595 } 596 #endif 597 598 static unsigned int features[] = { 599 VIRTIO_BALLOON_F_MUST_TELL_HOST, 600 VIRTIO_BALLOON_F_STATS_VQ, 601 VIRTIO_BALLOON_F_DEFLATE_ON_OOM, 602 }; 603 604 static struct virtio_driver virtio_balloon_driver = { 605 .feature_table = features, 606 .feature_table_size = ARRAY_SIZE(features), 607 .driver.name = KBUILD_MODNAME, 608 .driver.owner = THIS_MODULE, 609 .id_table = id_table, 610 .probe = virtballoon_probe, 611 .remove = virtballoon_remove, 612 .config_changed = virtballoon_changed, 613 #ifdef CONFIG_PM_SLEEP 614 .freeze = virtballoon_freeze, 615 .restore = virtballoon_restore, 616 #endif 617 }; 618 619 module_virtio_driver(virtio_balloon_driver); 620 MODULE_DEVICE_TABLE(virtio, id_table); 621 MODULE_DESCRIPTION("Virtio balloon driver"); 622 MODULE_LICENSE("GPL"); 623