drivers/xen/xen-selfballoon.c

   1 /******************************************************************************
   2  * Xen selfballoon driver (and optional frontswap self-shrinking driver)
   3  *
   4  * Copyright (c) 2009-2011, Dan Magenheimer, Oracle Corp.
   5  *
   6  * This code complements the cleancache and frontswap patchsets to optimize
   7  * support for Xen Transcendent Memory ("tmem").  The policy it implements
   8  * is rudimentary and will likely improve over time, but it does work well
   9  * enough today.
  10  *
  11  * Two functionalities are implemented here which both use "control theory"
  12  * (feedback) to optimize memory utilization. In a virtualized environment
  13  * such as Xen, RAM is often a scarce resource and we would like to ensure
  14  * that each of a possibly large number of virtual machines is using RAM
  15  * efficiently, i.e. using as little as possible when under light load
  16  * and obtaining as much as possible when memory demands are high.
  17  * Since RAM needs vary highly dynamically and sometimes dramatically,
  18  * "hysteresis" is used, that is, memory target is determined not just
  19  * on current data but also on past data stored in the system.
  20  *
  21  * "Selfballooning" creates memory pressure by managing the Xen balloon
  22  * driver to decrease and increase available kernel memory, driven
  23  * largely by the target value of "Committed_AS" (see /proc/meminfo).
  24  * Since Committed_AS does not account for clean mapped pages (i.e. pages
  25  * in RAM that are identical to pages on disk), selfballooning has the
  26  * affect of pushing less frequently used clean pagecache pages out of
  27  * kernel RAM and, presumably using cleancache, into Xen tmem where
  28  * Xen can more efficiently optimize RAM utilization for such pages.
  29  *
  30  * When kernel memory demand unexpectedly increases faster than Xen, via
  31  * the selfballoon driver, is able to (or chooses to) provide usable RAM,
  32  * the kernel may invoke swapping.  In most cases, frontswap is able
  33  * to absorb this swapping into Xen tmem.  However, due to the fact
  34  * that the kernel swap subsystem assumes swapping occurs to a disk,
  35  * swapped pages may sit on the disk for a very long time; even if
  36  * the kernel knows the page will never be used again.  This is because
  37  * the disk space costs very little and can be overwritten when
  38  * necessary.  When such stale pages are in frontswap, however, they
  39  * are taking up valuable real estate.  "Frontswap selfshrinking" works
  40  * to resolve this:  When frontswap activity is otherwise stable
  41  * and the guest kernel is not under memory pressure, the "frontswap
  42  * selfshrinking" accounts for this by providing pressure to remove some
  43  * pages from frontswap and return them to kernel memory.
  44  *
  45  * For both "selfballooning" and "frontswap-selfshrinking", a worker
  46  * thread is used and sysfs tunables are provided to adjust the frequency
  47  * and rate of adjustments to achieve the goal, as well as to disable one
  48  * or both functions independently.
  49  *
  50  * While some argue that this functionality can and should be implemented
  51  * in userspace, it has been observed that bad things happen (e.g. OOMs).
  52  *
  53  * System configuration note: Selfballooning should not be enabled on
  54  * systems without a sufficiently large swap device configured; for best
  55  * results, it is recommended that total swap be increased by the size
  56  * of the guest memory. Note, that selfballooning should be disabled by default
  57  * if frontswap is not configured.  Similarly selfballooning should be enabled
  58  * by default if frontswap is configured and can be disabled with the
  59  * "tmem.selfballooning=0" kernel boot option.  Finally, when frontswap is
  60  * configured, frontswap-selfshrinking can be disabled  with the
  61  * "tmem.selfshrink=0" kernel boot option.
  62  *
  63  * Selfballooning is disallowed in domain0 and force-disabled.
  64  *
  65  */
  66
  67 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  68
  69 #include <linux/kernel.h>
  70 #include <linux/bootmem.h>
  71 #include <linux/swap.h>
  72 #include <linux/mm.h>
  73 #include <linux/mman.h>
  74 #include <linux/workqueue.h>
  75 #include <linux/device.h>
  76 #include <xen/balloon.h>
  77 #include <xen/tmem.h>
  78 #include <xen/xen.h>
  79
  80 /* Enable/disable with sysfs. */
  81 static int xen_selfballooning_enabled __read_mostly;
  82
  83 /*
  84  * Controls rate at which memory target (this iteration) approaches
  85  * ultimate goal when memory need is increasing (up-hysteresis) or
  86  * decreasing (down-hysteresis). Higher values of hysteresis cause
  87  * slower increases/decreases. The default values for the various
  88  * parameters were deemed reasonable by experimentation, may be
  89  * workload-dependent, and can all be adjusted via sysfs.
  90  */
  91 static unsigned int selfballoon_downhysteresis __read_mostly = 8;
  92 static unsigned int selfballoon_uphysteresis __read_mostly = 1;
  93
  94 /* In HZ, controls frequency of worker invocation. */
  95 static unsigned int selfballoon_interval __read_mostly = 5;
  96
  97 /*
  98  * Minimum usable RAM in MB for selfballooning target for balloon.
  99  * If non-zero, it is added to totalreserve_pages and self-ballooning
 100  * will not balloon below the sum.  If zero, a piecewise linear function
 101  * is calculated as a minimum and added to totalreserve_pages.  Note that
 102  * setting this value indiscriminately may cause OOMs and crashes.
 103  */
 104 static unsigned int selfballoon_min_usable_mb;
 105
 106 /*
 107  * Amount of RAM in MB to add to the target number of pages.
 108  * Can be used to reserve some more room for caches and the like.
 109  */
 110 static unsigned int selfballoon_reserved_mb;
 111
 112 static void selfballoon_process(struct work_struct *work);
 113 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
 114
 115 #ifdef CONFIG_FRONTSWAP
 116 #include <linux/frontswap.h>
 117
 118 /* Enable/disable with sysfs. */
 119 static bool frontswap_selfshrinking __read_mostly;
 120
 121 /*
 122  * The default values for the following parameters were deemed reasonable
 123  * by experimentation, may be workload-dependent, and can all be
 124  * adjusted via sysfs.
 125  */
 126
 127 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 128 static unsigned int frontswap_hysteresis __read_mostly = 20;
 129
 130 /*
 131  * Number of selfballoon worker invocations to wait before observing that
 132  * frontswap selfshrinking should commence. Note that selfshrinking does
 133  * not use a separate worker thread.
 134  */
 135 static unsigned int frontswap_inertia __read_mostly = 3;
 136
 137 /* Countdown to next invocation of frontswap_shrink() */
 138 static unsigned long frontswap_inertia_counter;
 139
 140 /*
 141  * Invoked by the selfballoon worker thread, uses current number of pages
 142  * in frontswap (frontswap_curr_pages()), previous status, and control
 143  * values (hysteresis and inertia) to determine if frontswap should be
 144  * shrunk and what the new frontswap size should be.  Note that
 145  * frontswap_shrink is essentially a partial swapoff that immediately
 146  * transfers pages from the "swap device" (frontswap) back into kernel
 147  * RAM; despite the name, frontswap "shrinking" is very different from
 148  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 149  * memory.
 150  */
 151 static void frontswap_selfshrink(void)
 152 {
 153         static unsigned long cur_frontswap_pages;
 154         unsigned long last_frontswap_pages;
 155         unsigned long tgt_frontswap_pages;
 156
 157         last_frontswap_pages = cur_frontswap_pages;
 158         cur_frontswap_pages = frontswap_curr_pages();
 159         if (!cur_frontswap_pages ||
 160                         (cur_frontswap_pages > last_frontswap_pages)) {
 161                 frontswap_inertia_counter = frontswap_inertia;
 162                 return;
 163         }
 164         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 165                 return;
 166         if (cur_frontswap_pages <= frontswap_hysteresis)
 167                 tgt_frontswap_pages = 0;
 168         else
 169                 tgt_frontswap_pages = cur_frontswap_pages -
 170                         (cur_frontswap_pages / frontswap_hysteresis);
 171         frontswap_shrink(tgt_frontswap_pages);
 172         frontswap_inertia_counter = frontswap_inertia;
 173 }
 174
 175 #endif /* CONFIG_FRONTSWAP */
 176
 177 #define MB2PAGES(mb)    ((mb) << (20 - PAGE_SHIFT))
 178 #define PAGES2MB(pages) ((pages) >> (20 - PAGE_SHIFT))
 179
 180 /*
 181  * Use current balloon size, the goal (vm_committed_as), and hysteresis
 182  * parameters to set a new target balloon size
 183  */
 184 static void selfballoon_process(struct work_struct *work)
 185 {
 186         unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
 187         unsigned long useful_pages;
 188         bool reset_timer = false;
 189
 190         if (xen_selfballooning_enabled) {
 191                 cur_pages = totalram_pages;
 192                 tgt_pages = cur_pages; /* default is no change */
 193                 goal_pages = vm_memory_committed() +
 194                                 totalreserve_pages +
 195                                 MB2PAGES(selfballoon_reserved_mb);
 196 #ifdef CONFIG_FRONTSWAP
 197                 /* allow space for frontswap pages to be repatriated */
 198                 if (frontswap_selfshrinking)
 199                         goal_pages += frontswap_curr_pages();
 200 #endif
 201                 if (cur_pages > goal_pages)
 202                         tgt_pages = cur_pages -
 203                                 ((cur_pages - goal_pages) /
 204                                   selfballoon_downhysteresis);
 205                 else if (cur_pages < goal_pages)
 206                         tgt_pages = cur_pages +
 207                                 ((goal_pages - cur_pages) /
 208                                   selfballoon_uphysteresis);
 209                 /* else if cur_pages == goal_pages, no change */
 210                 useful_pages = max_pfn - totalreserve_pages;
 211                 if (selfballoon_min_usable_mb != 0)
 212                         floor_pages = totalreserve_pages +
 213                                         MB2PAGES(selfballoon_min_usable_mb);
 214                 /* piecewise linear function ending in ~3% slope */
 215                 else if (useful_pages < MB2PAGES(16))
 216                         floor_pages = max_pfn; /* not worth ballooning */
 217                 else if (useful_pages < MB2PAGES(64))
 218                         floor_pages = totalreserve_pages + MB2PAGES(16) +
 219                                         ((useful_pages - MB2PAGES(16)) >> 1);
 220                 else if (useful_pages < MB2PAGES(512))
 221                         floor_pages = totalreserve_pages + MB2PAGES(40) +
 222                                         ((useful_pages - MB2PAGES(40)) >> 3);
 223                 else /* useful_pages >= MB2PAGES(512) */
 224                         floor_pages = totalreserve_pages + MB2PAGES(99) +
 225                                         ((useful_pages - MB2PAGES(99)) >> 5);
 226                 if (tgt_pages < floor_pages)
 227                         tgt_pages = floor_pages;
 228                 balloon_set_new_target(tgt_pages +
 229                         balloon_stats.current_pages - totalram_pages);
 230                 reset_timer = true;
 231         }
 232 #ifdef CONFIG_FRONTSWAP
 233         if (frontswap_selfshrinking) {
 234                 frontswap_selfshrink();
 235                 reset_timer = true;
 236         }
 237 #endif
 238         if (reset_timer)
 239                 schedule_delayed_work(&selfballoon_worker,
 240                         selfballoon_interval * HZ);
 241 }
 242
 243 #ifdef CONFIG_SYSFS
 244
 245 #include <linux/capability.h>
 246
 247 #define SELFBALLOON_SHOW(name, format, args...)                         \
 248         static ssize_t show_##name(struct device *dev,  \
 249                                           struct device_attribute *attr, \
 250                                           char *buf) \
 251         { \
 252                 return sprintf(buf, format, ##args); \
 253         }
 254
 255 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 256
 257 static ssize_t store_selfballooning(struct device *dev,
 258                             struct device_attribute *attr,
 259                             const char *buf,
 260                             size_t count)
 261 {
 262         bool was_enabled = xen_selfballooning_enabled;
 263         unsigned long tmp;
 264         int err;
 265
 266         if (!capable(CAP_SYS_ADMIN))
 267                 return -EPERM;
 268
 269         err = kstrtoul(buf, 10, &tmp);
 270         if (err)
 271                 return err;
 272         if ((tmp != 0) && (tmp != 1))
 273                 return -EINVAL;
 274
 275         xen_selfballooning_enabled = !!tmp;
 276         if (!was_enabled && xen_selfballooning_enabled)
 277                 schedule_delayed_work(&selfballoon_worker,
 278                         selfballoon_interval * HZ);
 279
 280         return count;
 281 }
 282
 283 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 284                    show_selfballooning, store_selfballooning);
 285
 286 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 287
 288 static ssize_t store_selfballoon_interval(struct device *dev,
 289                                           struct device_attribute *attr,
 290                                           const char *buf,
 291                                           size_t count)
 292 {
 293         unsigned long val;
 294         int err;
 295
 296         if (!capable(CAP_SYS_ADMIN))
 297                 return -EPERM;
 298         err = kstrtoul(buf, 10, &val);
 299         if (err)
 300                 return err;
 301         if (val == 0)
 302                 return -EINVAL;
 303         selfballoon_interval = val;
 304         return count;
 305 }
 306
 307 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 308                    show_selfballoon_interval, store_selfballoon_interval);
 309
 310 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 311
 312 static ssize_t store_selfballoon_downhys(struct device *dev,
 313                                          struct device_attribute *attr,
 314                                          const char *buf,
 315                                          size_t count)
 316 {
 317         unsigned long val;
 318         int err;
 319
 320         if (!capable(CAP_SYS_ADMIN))
 321                 return -EPERM;
 322         err = kstrtoul(buf, 10, &val);
 323         if (err)
 324                 return err;
 325         if (val == 0)
 326                 return -EINVAL;
 327         selfballoon_downhysteresis = val;
 328         return count;
 329 }
 330
 331 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 332                    show_selfballoon_downhys, store_selfballoon_downhys);
 333
 334
 335 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 336
 337 static ssize_t store_selfballoon_uphys(struct device *dev,
 338                                        struct device_attribute *attr,
 339                                        const char *buf,
 340                                        size_t count)
 341 {
 342         unsigned long val;
 343         int err;
 344
 345         if (!capable(CAP_SYS_ADMIN))
 346                 return -EPERM;
 347         err = kstrtoul(buf, 10, &val);
 348         if (err)
 349                 return err;
 350         if (val == 0)
 351                 return -EINVAL;
 352         selfballoon_uphysteresis = val;
 353         return count;
 354 }
 355
 356 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 357                    show_selfballoon_uphys, store_selfballoon_uphys);
 358
 359 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 360                                 selfballoon_min_usable_mb);
 361
 362 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
 363                                                struct device_attribute *attr,
 364                                                const char *buf,
 365                                                size_t count)
 366 {
 367         unsigned long val;
 368         int err;
 369
 370         if (!capable(CAP_SYS_ADMIN))
 371                 return -EPERM;
 372         err = kstrtoul(buf, 10, &val);
 373         if (err)
 374                 return err;
 375         if (val == 0)
 376                 return -EINVAL;
 377         selfballoon_min_usable_mb = val;
 378         return count;
 379 }
 380
 381 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 382                    show_selfballoon_min_usable_mb,
 383                    store_selfballoon_min_usable_mb);
 384
 385 SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
 386                                 selfballoon_reserved_mb);
 387
 388 static ssize_t store_selfballoon_reserved_mb(struct device *dev,
 389                                              struct device_attribute *attr,
 390                                              const char *buf,
 391                                              size_t count)
 392 {
 393         unsigned long val;
 394         int err;
 395
 396         if (!capable(CAP_SYS_ADMIN))
 397                 return -EPERM;
 398         err = kstrtoul(buf, 10, &val);
 399         if (err)
 400                 return err;
 401         if (val == 0)
 402                 return -EINVAL;
 403         selfballoon_reserved_mb = val;
 404         return count;
 405 }
 406
 407 static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
 408                    show_selfballoon_reserved_mb,
 409                    store_selfballoon_reserved_mb);
 410
 411
 412 #ifdef CONFIG_FRONTSWAP
 413 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 414
 415 static ssize_t store_frontswap_selfshrinking(struct device *dev,
 416                                              struct device_attribute *attr,
 417                                              const char *buf,
 418                                              size_t count)
 419 {
 420         bool was_enabled = frontswap_selfshrinking;
 421         unsigned long tmp;
 422         int err;
 423
 424         if (!capable(CAP_SYS_ADMIN))
 425                 return -EPERM;
 426         err = kstrtoul(buf, 10, &tmp);
 427         if (err)
 428                 return err;
 429         if ((tmp != 0) && (tmp != 1))
 430                 return -EINVAL;
 431         frontswap_selfshrinking = !!tmp;
 432         if (!was_enabled && !xen_selfballooning_enabled &&
 433              frontswap_selfshrinking)
 434                 schedule_delayed_work(&selfballoon_worker,
 435                         selfballoon_interval * HZ);
 436
 437         return count;
 438 }
 439
 440 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 441                    show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 442
 443 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 444
 445 static ssize_t store_frontswap_inertia(struct device *dev,
 446                                        struct device_attribute *attr,
 447                                        const char *buf,
 448                                        size_t count)
 449 {
 450         unsigned long val;
 451         int err;
 452
 453         if (!capable(CAP_SYS_ADMIN))
 454                 return -EPERM;
 455         err = kstrtoul(buf, 10, &val);
 456         if (err)
 457                 return err;
 458         if (val == 0)
 459                 return -EINVAL;
 460         frontswap_inertia = val;
 461         frontswap_inertia_counter = val;
 462         return count;
 463 }
 464
 465 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 466                    show_frontswap_inertia, store_frontswap_inertia);
 467
 468 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 469
 470 static ssize_t store_frontswap_hysteresis(struct device *dev,
 471                                           struct device_attribute *attr,
 472                                           const char *buf,
 473                                           size_t count)
 474 {
 475         unsigned long val;
 476         int err;
 477
 478         if (!capable(CAP_SYS_ADMIN))
 479                 return -EPERM;
 480         err = kstrtoul(buf, 10, &val);
 481         if (err)
 482                 return err;
 483         if (val == 0)
 484                 return -EINVAL;
 485         frontswap_hysteresis = val;
 486         return count;
 487 }
 488
 489 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 490                    show_frontswap_hysteresis, store_frontswap_hysteresis);
 491
 492 #endif /* CONFIG_FRONTSWAP */
 493
 494 static struct attribute *selfballoon_attrs[] = {
 495         &dev_attr_selfballooning.attr,
 496         &dev_attr_selfballoon_interval.attr,
 497         &dev_attr_selfballoon_downhysteresis.attr,
 498         &dev_attr_selfballoon_uphysteresis.attr,
 499         &dev_attr_selfballoon_min_usable_mb.attr,
 500         &dev_attr_selfballoon_reserved_mb.attr,
 501 #ifdef CONFIG_FRONTSWAP
 502         &dev_attr_frontswap_selfshrinking.attr,
 503         &dev_attr_frontswap_hysteresis.attr,
 504         &dev_attr_frontswap_inertia.attr,
 505 #endif
 506         NULL
 507 };
 508
 509 static const struct attribute_group selfballoon_group = {
 510         .name = "selfballoon",
 511         .attrs = selfballoon_attrs
 512 };
 513 #endif
 514
 515 int register_xen_selfballooning(struct device *dev)
 516 {
 517         int error = -1;
 518
 519 #ifdef CONFIG_SYSFS
 520         error = sysfs_create_group(&dev->kobj, &selfballoon_group);
 521 #endif
 522         return error;
 523 }
 524 EXPORT_SYMBOL(register_xen_selfballooning);
 525
 526 int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
 527 {
 528         bool enable = false;
 529         unsigned long reserve_pages;
 530
 531         if (!xen_domain())
 532                 return -ENODEV;
 533
 534         if (xen_initial_domain()) {
 535                 pr_info("Xen selfballooning driver disabled for domain0\n");
 536                 return -ENODEV;
 537         }
 538
 539         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 540         if (xen_selfballooning_enabled) {
 541                 pr_info("Initializing Xen selfballooning driver\n");
 542                 enable = true;
 543         }
 544 #ifdef CONFIG_FRONTSWAP
 545         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 546         if (frontswap_selfshrinking) {
 547                 pr_info("Initializing frontswap selfshrinking driver\n");
 548                 enable = true;
 549         }
 550 #endif
 551         if (!enable)
 552                 return -ENODEV;
 553
 554         /*
 555          * Give selfballoon_reserved_mb a default value(10% of total ram pages)
 556          * to make selfballoon not so aggressive.
 557          *
 558          * There are mainly two reasons:
 559          * 1) The original goal_page didn't consider some pages used by kernel
 560          *    space, like slab pages and memory used by device drivers.
 561          *
 562          * 2) The balloon driver may not give back memory to guest OS fast
 563          *    enough when the workload suddenly aquries a lot of physical memory.
 564          *
 565          * In both cases, the guest OS will suffer from memory pressure and
 566          * OOM killer may be triggered.
 567          * By reserving extra 10% of total ram pages, we can keep the system
 568          * much more reliably and response faster in some cases.
 569          */
 570         if (!selfballoon_reserved_mb) {
 571                 reserve_pages = totalram_pages / 10;
 572                 selfballoon_reserved_mb = PAGES2MB(reserve_pages);
 573         }
 574         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 575
 576         return 0;
 577 }
 578 EXPORT_SYMBOL(xen_selfballoon_init);