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.  Also, while technically not required to be
  57  * configured, it is highly recommended that frontswap also be configured
  58  * and enabled when selfballooning is running.  So, selfballooning
  59  * is disabled by default if frontswap is not configured and can only
  60  * be enabled with the "selfballooning" kernel boot option; similarly
  61  * selfballooning is enabled by default if frontswap is configured and
  62  * can be disabled with the "noselfballooning" kernel boot option.  Finally,
  63  * when frontswap is configured,frontswap-selfshrinking can be disabled
  64  * with the "tmem.selfshrink=0" kernel boot option.
  65  *
  66  * Selfballooning is disallowed in domain0 and force-disabled.
  67  *
  68  */
  69
  70 #include <linux/kernel.h>
  71 #include <linux/bootmem.h>
  72 #include <linux/swap.h>
  73 #include <linux/mm.h>
  74 #include <linux/mman.h>
  75 #include <linux/module.h>
  76 #include <linux/workqueue.h>
  77 #include <linux/device.h>
  78 #include <xen/balloon.h>
  79 #include <xen/tmem.h>
  80 #include <xen/xen.h>
  81
  82 /* Enable/disable with sysfs. */
  83 static int xen_selfballooning_enabled __read_mostly;
  84
  85 /*
  86  * Controls rate at which memory target (this iteration) approaches
  87  * ultimate goal when memory need is increasing (up-hysteresis) or
  88  * decreasing (down-hysteresis). Higher values of hysteresis cause
  89  * slower increases/decreases. The default values for the various
  90  * parameters were deemed reasonable by experimentation, may be
  91  * workload-dependent, and can all be adjusted via sysfs.
  92  */
  93 static unsigned int selfballoon_downhysteresis __read_mostly = 8;
  94 static unsigned int selfballoon_uphysteresis __read_mostly = 1;
  95
  96 /* In HZ, controls frequency of worker invocation. */
  97 static unsigned int selfballoon_interval __read_mostly = 5;
  98
  99 /*
 100  * Minimum usable RAM in MB for selfballooning target for balloon.
 101  * If non-zero, it is added to totalreserve_pages and self-ballooning
 102  * will not balloon below the sum.  If zero, a piecewise linear function
 103  * is calculated as a minimum and added to totalreserve_pages.  Note that
 104  * setting this value indiscriminately may cause OOMs and crashes.
 105  */
 106 static unsigned int selfballoon_min_usable_mb;
 107
 108 /*
 109  * Amount of RAM in MB to add to the target number of pages.
 110  * Can be used to reserve some more room for caches and the like.
 111  */
 112 static unsigned int selfballoon_reserved_mb;
 113
 114 static void selfballoon_process(struct work_struct *work);
 115 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
 116
 117 #ifdef CONFIG_FRONTSWAP
 118 #include <linux/frontswap.h>
 119
 120 /* Enable/disable with sysfs. */
 121 static bool frontswap_selfshrinking __read_mostly;
 122
 123 /*
 124  * The default values for the following parameters were deemed reasonable
 125  * by experimentation, may be workload-dependent, and can all be
 126  * adjusted via sysfs.
 127  */
 128
 129 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 130 static unsigned int frontswap_hysteresis __read_mostly = 20;
 131
 132 /*
 133  * Number of selfballoon worker invocations to wait before observing that
 134  * frontswap selfshrinking should commence. Note that selfshrinking does
 135  * not use a separate worker thread.
 136  */
 137 static unsigned int frontswap_inertia __read_mostly = 3;
 138
 139 /* Countdown to next invocation of frontswap_shrink() */
 140 static unsigned long frontswap_inertia_counter;
 141
 142 /*
 143  * Invoked by the selfballoon worker thread, uses current number of pages
 144  * in frontswap (frontswap_curr_pages()), previous status, and control
 145  * values (hysteresis and inertia) to determine if frontswap should be
 146  * shrunk and what the new frontswap size should be.  Note that
 147  * frontswap_shrink is essentially a partial swapoff that immediately
 148  * transfers pages from the "swap device" (frontswap) back into kernel
 149  * RAM; despite the name, frontswap "shrinking" is very different from
 150  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 151  * memory.
 152  */
 153 static void frontswap_selfshrink(void)
 154 {
 155         static unsigned long cur_frontswap_pages;
 156         static unsigned long last_frontswap_pages;
 157         static unsigned long tgt_frontswap_pages;
 158
 159         last_frontswap_pages = cur_frontswap_pages;
 160         cur_frontswap_pages = frontswap_curr_pages();
 161         if (!cur_frontswap_pages ||
 162                         (cur_frontswap_pages > last_frontswap_pages)) {
 163                 frontswap_inertia_counter = frontswap_inertia;
 164                 return;
 165         }
 166         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 167                 return;
 168         if (cur_frontswap_pages <= frontswap_hysteresis)
 169                 tgt_frontswap_pages = 0;
 170         else
 171                 tgt_frontswap_pages = cur_frontswap_pages -
 172                         (cur_frontswap_pages / frontswap_hysteresis);
 173         frontswap_shrink(tgt_frontswap_pages);
 174 }
 175
 176 /* Disable with kernel boot option. */
 177 static bool use_selfballooning = true;
 178
 179 static int __init xen_noselfballooning_setup(char *s)
 180 {
 181         use_selfballooning = false;
 182         return 1;
 183 }
 184
 185 __setup("noselfballooning", xen_noselfballooning_setup);
 186 #else /* !CONFIG_FRONTSWAP */
 187 /* Enable with kernel boot option. */
 188 static bool use_selfballooning;
 189
 190 static int __init xen_selfballooning_setup(char *s)
 191 {
 192         use_selfballooning = true;
 193         return 1;
 194 }
 195
 196 __setup("selfballooning", xen_selfballooning_setup);
 197 #endif /* CONFIG_FRONTSWAP */
 198
 199 #define MB2PAGES(mb)    ((mb) << (20 - PAGE_SHIFT))
 200
 201 /*
 202  * Use current balloon size, the goal (vm_committed_as), and hysteresis
 203  * parameters to set a new target balloon size
 204  */
 205 static void selfballoon_process(struct work_struct *work)
 206 {
 207         unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
 208         unsigned long useful_pages;
 209         bool reset_timer = false;
 210
 211         if (xen_selfballooning_enabled) {
 212                 cur_pages = totalram_pages;
 213                 tgt_pages = cur_pages; /* default is no change */
 214                 goal_pages = vm_memory_committed() +
 215                                 totalreserve_pages +
 216                                 MB2PAGES(selfballoon_reserved_mb);
 217 #ifdef CONFIG_FRONTSWAP
 218                 /* allow space for frontswap pages to be repatriated */
 219                 if (frontswap_selfshrinking && frontswap_enabled)
 220                         goal_pages += frontswap_curr_pages();
 221 #endif
 222                 if (cur_pages > goal_pages)
 223                         tgt_pages = cur_pages -
 224                                 ((cur_pages - goal_pages) /
 225                                   selfballoon_downhysteresis);
 226                 else if (cur_pages < goal_pages)
 227                         tgt_pages = cur_pages +
 228                                 ((goal_pages - cur_pages) /
 229                                   selfballoon_uphysteresis);
 230                 /* else if cur_pages == goal_pages, no change */
 231                 useful_pages = max_pfn - totalreserve_pages;
 232                 if (selfballoon_min_usable_mb != 0)
 233                         floor_pages = totalreserve_pages +
 234                                         MB2PAGES(selfballoon_min_usable_mb);
 235                 /* piecewise linear function ending in ~3% slope */
 236                 else if (useful_pages < MB2PAGES(16))
 237                         floor_pages = max_pfn; /* not worth ballooning */
 238                 else if (useful_pages < MB2PAGES(64))
 239                         floor_pages = totalreserve_pages + MB2PAGES(16) +
 240                                         ((useful_pages - MB2PAGES(16)) >> 1);
 241                 else if (useful_pages < MB2PAGES(512))
 242                         floor_pages = totalreserve_pages + MB2PAGES(40) +
 243                                         ((useful_pages - MB2PAGES(40)) >> 3);
 244                 else /* useful_pages >= MB2PAGES(512) */
 245                         floor_pages = totalreserve_pages + MB2PAGES(99) +
 246                                         ((useful_pages - MB2PAGES(99)) >> 5);
 247                 if (tgt_pages < floor_pages)
 248                         tgt_pages = floor_pages;
 249                 balloon_set_new_target(tgt_pages +
 250                         balloon_stats.current_pages - totalram_pages);
 251                 reset_timer = true;
 252         }
 253 #ifdef CONFIG_FRONTSWAP
 254         if (frontswap_selfshrinking && frontswap_enabled) {
 255                 frontswap_selfshrink();
 256                 reset_timer = true;
 257         }
 258 #endif
 259         if (reset_timer)
 260                 schedule_delayed_work(&selfballoon_worker,
 261                         selfballoon_interval * HZ);
 262 }
 263
 264 #ifdef CONFIG_SYSFS
 265
 266 #include <linux/capability.h>
 267
 268 #define SELFBALLOON_SHOW(name, format, args...)                         \
 269         static ssize_t show_##name(struct device *dev,  \
 270                                           struct device_attribute *attr, \
 271                                           char *buf) \
 272         { \
 273                 return sprintf(buf, format, ##args); \
 274         }
 275
 276 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 277
 278 static ssize_t store_selfballooning(struct device *dev,
 279                             struct device_attribute *attr,
 280                             const char *buf,
 281                             size_t count)
 282 {
 283         bool was_enabled = xen_selfballooning_enabled;
 284         unsigned long tmp;
 285         int err;
 286
 287         if (!capable(CAP_SYS_ADMIN))
 288                 return -EPERM;
 289
 290         err = strict_strtoul(buf, 10, &tmp);
 291         if (err || ((tmp != 0) && (tmp != 1)))
 292                 return -EINVAL;
 293
 294         xen_selfballooning_enabled = !!tmp;
 295         if (!was_enabled && xen_selfballooning_enabled)
 296                 schedule_delayed_work(&selfballoon_worker,
 297                         selfballoon_interval * HZ);
 298
 299         return count;
 300 }
 301
 302 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 303                    show_selfballooning, store_selfballooning);
 304
 305 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 306
 307 static ssize_t store_selfballoon_interval(struct device *dev,
 308                                           struct device_attribute *attr,
 309                                           const char *buf,
 310                                           size_t count)
 311 {
 312         unsigned long val;
 313         int err;
 314
 315         if (!capable(CAP_SYS_ADMIN))
 316                 return -EPERM;
 317         err = strict_strtoul(buf, 10, &val);
 318         if (err || val == 0)
 319                 return -EINVAL;
 320         selfballoon_interval = val;
 321         return count;
 322 }
 323
 324 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 325                    show_selfballoon_interval, store_selfballoon_interval);
 326
 327 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 328
 329 static ssize_t store_selfballoon_downhys(struct device *dev,
 330                                          struct device_attribute *attr,
 331                                          const char *buf,
 332                                          size_t count)
 333 {
 334         unsigned long val;
 335         int err;
 336
 337         if (!capable(CAP_SYS_ADMIN))
 338                 return -EPERM;
 339         err = strict_strtoul(buf, 10, &val);
 340         if (err || val == 0)
 341                 return -EINVAL;
 342         selfballoon_downhysteresis = val;
 343         return count;
 344 }
 345
 346 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 347                    show_selfballoon_downhys, store_selfballoon_downhys);
 348
 349
 350 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 351
 352 static ssize_t store_selfballoon_uphys(struct device *dev,
 353                                        struct device_attribute *attr,
 354                                        const char *buf,
 355                                        size_t count)
 356 {
 357         unsigned long val;
 358         int err;
 359
 360         if (!capable(CAP_SYS_ADMIN))
 361                 return -EPERM;
 362         err = strict_strtoul(buf, 10, &val);
 363         if (err || val == 0)
 364                 return -EINVAL;
 365         selfballoon_uphysteresis = val;
 366         return count;
 367 }
 368
 369 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 370                    show_selfballoon_uphys, store_selfballoon_uphys);
 371
 372 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 373                                 selfballoon_min_usable_mb);
 374
 375 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
 376                                                struct device_attribute *attr,
 377                                                const char *buf,
 378                                                size_t count)
 379 {
 380         unsigned long val;
 381         int err;
 382
 383         if (!capable(CAP_SYS_ADMIN))
 384                 return -EPERM;
 385         err = strict_strtoul(buf, 10, &val);
 386         if (err || val == 0)
 387                 return -EINVAL;
 388         selfballoon_min_usable_mb = val;
 389         return count;
 390 }
 391
 392 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 393                    show_selfballoon_min_usable_mb,
 394                    store_selfballoon_min_usable_mb);
 395
 396 SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
 397                                 selfballoon_reserved_mb);
 398
 399 static ssize_t store_selfballoon_reserved_mb(struct device *dev,
 400                                              struct device_attribute *attr,
 401                                              const char *buf,
 402                                              size_t count)
 403 {
 404         unsigned long val;
 405         int err;
 406
 407         if (!capable(CAP_SYS_ADMIN))
 408                 return -EPERM;
 409         err = strict_strtoul(buf, 10, &val);
 410         if (err || val == 0)
 411                 return -EINVAL;
 412         selfballoon_reserved_mb = val;
 413         return count;
 414 }
 415
 416 static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
 417                    show_selfballoon_reserved_mb,
 418                    store_selfballoon_reserved_mb);
 419
 420
 421 #ifdef CONFIG_FRONTSWAP
 422 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 423
 424 static ssize_t store_frontswap_selfshrinking(struct device *dev,
 425                                              struct device_attribute *attr,
 426                                              const char *buf,
 427                                              size_t count)
 428 {
 429         bool was_enabled = frontswap_selfshrinking;
 430         unsigned long tmp;
 431         int err;
 432
 433         if (!capable(CAP_SYS_ADMIN))
 434                 return -EPERM;
 435         err = strict_strtoul(buf, 10, &tmp);
 436         if (err || ((tmp != 0) && (tmp != 1)))
 437                 return -EINVAL;
 438         frontswap_selfshrinking = !!tmp;
 439         if (!was_enabled && !xen_selfballooning_enabled &&
 440              frontswap_selfshrinking)
 441                 schedule_delayed_work(&selfballoon_worker,
 442                         selfballoon_interval * HZ);
 443
 444         return count;
 445 }
 446
 447 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 448                    show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 449
 450 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 451
 452 static ssize_t store_frontswap_inertia(struct device *dev,
 453                                        struct device_attribute *attr,
 454                                        const char *buf,
 455                                        size_t count)
 456 {
 457         unsigned long val;
 458         int err;
 459
 460         if (!capable(CAP_SYS_ADMIN))
 461                 return -EPERM;
 462         err = strict_strtoul(buf, 10, &val);
 463         if (err || val == 0)
 464                 return -EINVAL;
 465         frontswap_inertia = val;
 466         frontswap_inertia_counter = val;
 467         return count;
 468 }
 469
 470 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 471                    show_frontswap_inertia, store_frontswap_inertia);
 472
 473 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 474
 475 static ssize_t store_frontswap_hysteresis(struct device *dev,
 476                                           struct device_attribute *attr,
 477                                           const char *buf,
 478                                           size_t count)
 479 {
 480         unsigned long val;
 481         int err;
 482
 483         if (!capable(CAP_SYS_ADMIN))
 484                 return -EPERM;
 485         err = strict_strtoul(buf, 10, &val);
 486         if (err || val == 0)
 487                 return -EINVAL;
 488         frontswap_hysteresis = val;
 489         return count;
 490 }
 491
 492 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 493                    show_frontswap_hysteresis, store_frontswap_hysteresis);
 494
 495 #endif /* CONFIG_FRONTSWAP */
 496
 497 static struct attribute *selfballoon_attrs[] = {
 498         &dev_attr_selfballooning.attr,
 499         &dev_attr_selfballoon_interval.attr,
 500         &dev_attr_selfballoon_downhysteresis.attr,
 501         &dev_attr_selfballoon_uphysteresis.attr,
 502         &dev_attr_selfballoon_min_usable_mb.attr,
 503         &dev_attr_selfballoon_reserved_mb.attr,
 504 #ifdef CONFIG_FRONTSWAP
 505         &dev_attr_frontswap_selfshrinking.attr,
 506         &dev_attr_frontswap_hysteresis.attr,
 507         &dev_attr_frontswap_inertia.attr,
 508 #endif
 509         NULL
 510 };
 511
 512 static const struct attribute_group selfballoon_group = {
 513         .name = "selfballoon",
 514         .attrs = selfballoon_attrs
 515 };
 516 #endif
 517
 518 int register_xen_selfballooning(struct device *dev)
 519 {
 520         int error = -1;
 521
 522 #ifdef CONFIG_SYSFS
 523         error = sysfs_create_group(&dev->kobj, &selfballoon_group);
 524 #endif
 525         return error;
 526 }
 527 EXPORT_SYMBOL(register_xen_selfballooning);
 528
 529 int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
 530 {
 531         bool enable = false;
 532
 533         if (!xen_domain())
 534                 return -ENODEV;
 535
 536         if (xen_initial_domain()) {
 537                 pr_info("xen/balloon: Xen selfballooning driver "
 538                                 "disabled for domain0.\n");
 539                 return -ENODEV;
 540         }
 541
 542         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 543         if (xen_selfballooning_enabled) {
 544                 pr_info("xen/balloon: Initializing Xen "
 545                                         "selfballooning driver.\n");
 546                 enable = true;
 547         }
 548 #ifdef CONFIG_FRONTSWAP
 549         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 550         if (frontswap_selfshrinking) {
 551                 pr_info("xen/balloon: Initializing frontswap "
 552                                         "selfshrinking driver.\n");
 553                 enable = true;
 554         }
 555 #endif
 556         if (!enable)
 557                 return -ENODEV;
 558
 559         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 560
 561         return 0;
 562 }
 563 EXPORT_SYMBOL(xen_selfballoon_init);