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 "tmem.selfballooning=1" kernel boot option; similarly
  61  * selfballooning is enabled by default if frontswap is configured and
  62  * can be disabled with the "tmem.selfballooning=0" 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 #endif /* CONFIG_FRONTSWAP */
 177
 178 #define MB2PAGES(mb)    ((mb) << (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 && frontswap_enabled)
 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 && frontswap_enabled) {
 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 = strict_strtoul(buf, 10, &tmp);
 270         if (err || ((tmp != 0) && (tmp != 1)))
 271                 return -EINVAL;
 272
 273         xen_selfballooning_enabled = !!tmp;
 274         if (!was_enabled && xen_selfballooning_enabled)
 275                 schedule_delayed_work(&selfballoon_worker,
 276                         selfballoon_interval * HZ);
 277
 278         return count;
 279 }
 280
 281 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 282                    show_selfballooning, store_selfballooning);
 283
 284 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 285
 286 static ssize_t store_selfballoon_interval(struct device *dev,
 287                                           struct device_attribute *attr,
 288                                           const char *buf,
 289                                           size_t count)
 290 {
 291         unsigned long val;
 292         int err;
 293
 294         if (!capable(CAP_SYS_ADMIN))
 295                 return -EPERM;
 296         err = strict_strtoul(buf, 10, &val);
 297         if (err || val == 0)
 298                 return -EINVAL;
 299         selfballoon_interval = val;
 300         return count;
 301 }
 302
 303 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 304                    show_selfballoon_interval, store_selfballoon_interval);
 305
 306 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 307
 308 static ssize_t store_selfballoon_downhys(struct device *dev,
 309                                          struct device_attribute *attr,
 310                                          const char *buf,
 311                                          size_t count)
 312 {
 313         unsigned long val;
 314         int err;
 315
 316         if (!capable(CAP_SYS_ADMIN))
 317                 return -EPERM;
 318         err = strict_strtoul(buf, 10, &val);
 319         if (err || val == 0)
 320                 return -EINVAL;
 321         selfballoon_downhysteresis = val;
 322         return count;
 323 }
 324
 325 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 326                    show_selfballoon_downhys, store_selfballoon_downhys);
 327
 328
 329 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 330
 331 static ssize_t store_selfballoon_uphys(struct device *dev,
 332                                        struct device_attribute *attr,
 333                                        const char *buf,
 334                                        size_t count)
 335 {
 336         unsigned long val;
 337         int err;
 338
 339         if (!capable(CAP_SYS_ADMIN))
 340                 return -EPERM;
 341         err = strict_strtoul(buf, 10, &val);
 342         if (err || val == 0)
 343                 return -EINVAL;
 344         selfballoon_uphysteresis = val;
 345         return count;
 346 }
 347
 348 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 349                    show_selfballoon_uphys, store_selfballoon_uphys);
 350
 351 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 352                                 selfballoon_min_usable_mb);
 353
 354 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
 355                                                struct device_attribute *attr,
 356                                                const char *buf,
 357                                                size_t count)
 358 {
 359         unsigned long val;
 360         int err;
 361
 362         if (!capable(CAP_SYS_ADMIN))
 363                 return -EPERM;
 364         err = strict_strtoul(buf, 10, &val);
 365         if (err || val == 0)
 366                 return -EINVAL;
 367         selfballoon_min_usable_mb = val;
 368         return count;
 369 }
 370
 371 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 372                    show_selfballoon_min_usable_mb,
 373                    store_selfballoon_min_usable_mb);
 374
 375 SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
 376                                 selfballoon_reserved_mb);
 377
 378 static ssize_t store_selfballoon_reserved_mb(struct device *dev,
 379                                              struct device_attribute *attr,
 380                                              const char *buf,
 381                                              size_t count)
 382 {
 383         unsigned long val;
 384         int err;
 385
 386         if (!capable(CAP_SYS_ADMIN))
 387                 return -EPERM;
 388         err = strict_strtoul(buf, 10, &val);
 389         if (err || val == 0)
 390                 return -EINVAL;
 391         selfballoon_reserved_mb = val;
 392         return count;
 393 }
 394
 395 static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
 396                    show_selfballoon_reserved_mb,
 397                    store_selfballoon_reserved_mb);
 398
 399
 400 #ifdef CONFIG_FRONTSWAP
 401 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 402
 403 static ssize_t store_frontswap_selfshrinking(struct device *dev,
 404                                              struct device_attribute *attr,
 405                                              const char *buf,
 406                                              size_t count)
 407 {
 408         bool was_enabled = frontswap_selfshrinking;
 409         unsigned long tmp;
 410         int err;
 411
 412         if (!capable(CAP_SYS_ADMIN))
 413                 return -EPERM;
 414         err = strict_strtoul(buf, 10, &tmp);
 415         if (err || ((tmp != 0) && (tmp != 1)))
 416                 return -EINVAL;
 417         frontswap_selfshrinking = !!tmp;
 418         if (!was_enabled && !xen_selfballooning_enabled &&
 419              frontswap_selfshrinking)
 420                 schedule_delayed_work(&selfballoon_worker,
 421                         selfballoon_interval * HZ);
 422
 423         return count;
 424 }
 425
 426 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 427                    show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 428
 429 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 430
 431 static ssize_t store_frontswap_inertia(struct device *dev,
 432                                        struct device_attribute *attr,
 433                                        const char *buf,
 434                                        size_t count)
 435 {
 436         unsigned long val;
 437         int err;
 438
 439         if (!capable(CAP_SYS_ADMIN))
 440                 return -EPERM;
 441         err = strict_strtoul(buf, 10, &val);
 442         if (err || val == 0)
 443                 return -EINVAL;
 444         frontswap_inertia = val;
 445         frontswap_inertia_counter = val;
 446         return count;
 447 }
 448
 449 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 450                    show_frontswap_inertia, store_frontswap_inertia);
 451
 452 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 453
 454 static ssize_t store_frontswap_hysteresis(struct device *dev,
 455                                           struct device_attribute *attr,
 456                                           const char *buf,
 457                                           size_t count)
 458 {
 459         unsigned long val;
 460         int err;
 461
 462         if (!capable(CAP_SYS_ADMIN))
 463                 return -EPERM;
 464         err = strict_strtoul(buf, 10, &val);
 465         if (err || val == 0)
 466                 return -EINVAL;
 467         frontswap_hysteresis = val;
 468         return count;
 469 }
 470
 471 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 472                    show_frontswap_hysteresis, store_frontswap_hysteresis);
 473
 474 #endif /* CONFIG_FRONTSWAP */
 475
 476 static struct attribute *selfballoon_attrs[] = {
 477         &dev_attr_selfballooning.attr,
 478         &dev_attr_selfballoon_interval.attr,
 479         &dev_attr_selfballoon_downhysteresis.attr,
 480         &dev_attr_selfballoon_uphysteresis.attr,
 481         &dev_attr_selfballoon_min_usable_mb.attr,
 482         &dev_attr_selfballoon_reserved_mb.attr,
 483 #ifdef CONFIG_FRONTSWAP
 484         &dev_attr_frontswap_selfshrinking.attr,
 485         &dev_attr_frontswap_hysteresis.attr,
 486         &dev_attr_frontswap_inertia.attr,
 487 #endif
 488         NULL
 489 };
 490
 491 static const struct attribute_group selfballoon_group = {
 492         .name = "selfballoon",
 493         .attrs = selfballoon_attrs
 494 };
 495 #endif
 496
 497 int register_xen_selfballooning(struct device *dev)
 498 {
 499         int error = -1;
 500
 501 #ifdef CONFIG_SYSFS
 502         error = sysfs_create_group(&dev->kobj, &selfballoon_group);
 503 #endif
 504         return error;
 505 }
 506 EXPORT_SYMBOL(register_xen_selfballooning);
 507
 508 int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
 509 {
 510         bool enable = false;
 511
 512         if (!xen_domain())
 513                 return -ENODEV;
 514
 515         if (xen_initial_domain()) {
 516                 pr_info("xen/balloon: Xen selfballooning driver "
 517                                 "disabled for domain0.\n");
 518                 return -ENODEV;
 519         }
 520
 521         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 522         if (xen_selfballooning_enabled) {
 523                 pr_info("xen/balloon: Initializing Xen "
 524                                         "selfballooning driver.\n");
 525                 enable = true;
 526         }
 527 #ifdef CONFIG_FRONTSWAP
 528         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 529         if (frontswap_selfshrinking) {
 530                 pr_info("xen/balloon: Initializing frontswap "
 531                                         "selfshrinking driver.\n");
 532                 enable = true;
 533         }
 534 #endif
 535         if (!enable)
 536                 return -ENODEV;
 537
 538         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 539
 540         return 0;
 541 }
 542 EXPORT_SYMBOL(xen_selfballoon_init);