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 "noselfshrink" 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/workqueue.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 static void selfballoon_process(struct work_struct *work);
 107 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
 108
 109 #ifdef CONFIG_FRONTSWAP
 110 #include <linux/frontswap.h>
 111
 112 /* Enable/disable with sysfs. */
 113 static bool frontswap_selfshrinking __read_mostly;
 114
 115 /* Enable/disable with kernel boot option. */
 116 static bool use_frontswap_selfshrink __initdata = true;
 117
 118 /*
 119  * The default values for the following parameters were deemed reasonable
 120  * by experimentation, may be workload-dependent, and can all be
 121  * adjusted via sysfs.
 122  */
 123
 124 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 125 static unsigned int frontswap_hysteresis __read_mostly = 20;
 126
 127 /*
 128  * Number of selfballoon worker invocations to wait before observing that
 129  * frontswap selfshrinking should commence. Note that selfshrinking does
 130  * not use a separate worker thread.
 131  */
 132 static unsigned int frontswap_inertia __read_mostly = 3;
 133
 134 /* Countdown to next invocation of frontswap_shrink() */
 135 static unsigned long frontswap_inertia_counter;
 136
 137 /*
 138  * Invoked by the selfballoon worker thread, uses current number of pages
 139  * in frontswap (frontswap_curr_pages()), previous status, and control
 140  * values (hysteresis and inertia) to determine if frontswap should be
 141  * shrunk and what the new frontswap size should be.  Note that
 142  * frontswap_shrink is essentially a partial swapoff that immediately
 143  * transfers pages from the "swap device" (frontswap) back into kernel
 144  * RAM; despite the name, frontswap "shrinking" is very different from
 145  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 146  * memory.
 147  */
 148 static void frontswap_selfshrink(void)
 149 {
 150         static unsigned long cur_frontswap_pages;
 151         static unsigned long last_frontswap_pages;
 152         static unsigned long tgt_frontswap_pages;
 153
 154         last_frontswap_pages = cur_frontswap_pages;
 155         cur_frontswap_pages = frontswap_curr_pages();
 156         if (!cur_frontswap_pages ||
 157                         (cur_frontswap_pages > last_frontswap_pages)) {
 158                 frontswap_inertia_counter = frontswap_inertia;
 159                 return;
 160         }
 161         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 162                 return;
 163         if (cur_frontswap_pages <= frontswap_hysteresis)
 164                 tgt_frontswap_pages = 0;
 165         else
 166                 tgt_frontswap_pages = cur_frontswap_pages -
 167                         (cur_frontswap_pages / frontswap_hysteresis);
 168         frontswap_shrink(tgt_frontswap_pages);
 169 }
 170
 171 static int __init xen_nofrontswap_selfshrink_setup(char *s)
 172 {
 173         use_frontswap_selfshrink = false;
 174         return 1;
 175 }
 176
 177 __setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
 178
 179 /* Disable with kernel boot option. */
 180 static bool use_selfballooning __initdata = true;
 181
 182 static int __init xen_noselfballooning_setup(char *s)
 183 {
 184         use_selfballooning = false;
 185         return 1;
 186 }
 187
 188 __setup("noselfballooning", xen_noselfballooning_setup);
 189 #else /* !CONFIG_FRONTSWAP */
 190 /* Enable with kernel boot option. */
 191 static bool use_selfballooning __initdata = false;
 192
 193 static int __init xen_selfballooning_setup(char *s)
 194 {
 195         use_selfballooning = true;
 196         return 1;
 197 }
 198
 199 __setup("selfballooning", xen_selfballooning_setup);
 200 #endif /* CONFIG_FRONTSWAP */
 201
 202 #define MB2PAGES(mb)    ((mb) << (20 - PAGE_SHIFT))
 203
 204 /*
 205  * Use current balloon size, the goal (vm_committed_as), and hysteresis
 206  * parameters to set a new target balloon size
 207  */
 208 static void selfballoon_process(struct work_struct *work)
 209 {
 210         unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
 211         unsigned long useful_pages;
 212         bool reset_timer = false;
 213
 214         if (xen_selfballooning_enabled) {
 215                 cur_pages = totalram_pages;
 216                 tgt_pages = cur_pages; /* default is no change */
 217                 goal_pages = percpu_counter_read_positive(&vm_committed_as) +
 218                                 totalreserve_pages;
 219 #ifdef CONFIG_FRONTSWAP
 220                 /* allow space for frontswap pages to be repatriated */
 221                 if (frontswap_selfshrinking && frontswap_enabled)
 222                         goal_pages += frontswap_curr_pages();
 223 #endif
 224                 if (cur_pages > goal_pages)
 225                         tgt_pages = cur_pages -
 226                                 ((cur_pages - goal_pages) /
 227                                   selfballoon_downhysteresis);
 228                 else if (cur_pages < goal_pages)
 229                         tgt_pages = cur_pages +
 230                                 ((goal_pages - cur_pages) /
 231                                   selfballoon_uphysteresis);
 232                 /* else if cur_pages == goal_pages, no change */
 233                 useful_pages = max_pfn - totalreserve_pages;
 234                 if (selfballoon_min_usable_mb != 0)
 235                         floor_pages = totalreserve_pages +
 236                                         MB2PAGES(selfballoon_min_usable_mb);
 237                 /* piecewise linear function ending in ~3% slope */
 238                 else if (useful_pages < MB2PAGES(16))
 239                         floor_pages = max_pfn; /* not worth ballooning */
 240                 else if (useful_pages < MB2PAGES(64))
 241                         floor_pages = totalreserve_pages + MB2PAGES(16) +
 242                                         ((useful_pages - MB2PAGES(16)) >> 1);
 243                 else if (useful_pages < MB2PAGES(512))
 244                         floor_pages = totalreserve_pages + MB2PAGES(40) +
 245                                         ((useful_pages - MB2PAGES(40)) >> 3);
 246                 else /* useful_pages >= MB2PAGES(512) */
 247                         floor_pages = totalreserve_pages + MB2PAGES(99) +
 248                                         ((useful_pages - MB2PAGES(99)) >> 5);
 249                 if (tgt_pages < floor_pages)
 250                         tgt_pages = floor_pages;
 251                 balloon_set_new_target(tgt_pages +
 252                         balloon_stats.current_pages - totalram_pages);
 253                 reset_timer = true;
 254         }
 255 #ifdef CONFIG_FRONTSWAP
 256         if (frontswap_selfshrinking && frontswap_enabled) {
 257                 frontswap_selfshrink();
 258                 reset_timer = true;
 259         }
 260 #endif
 261         if (reset_timer)
 262                 schedule_delayed_work(&selfballoon_worker,
 263                         selfballoon_interval * HZ);
 264 }
 265
 266 #ifdef CONFIG_SYSFS
 267
 268 #include <linux/sysdev.h>
 269 #include <linux/capability.h>
 270
 271 #define SELFBALLOON_SHOW(name, format, args...)                         \
 272         static ssize_t show_##name(struct sys_device *dev,      \
 273                                            struct sysdev_attribute *attr, \
 274                                            char *buf) \
 275         { \
 276                 return sprintf(buf, format, ##args); \
 277         }
 278
 279 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 280
 281 static ssize_t store_selfballooning(struct sys_device *dev,
 282                             struct sysdev_attribute *attr,
 283                             const char *buf,
 284                             size_t count)
 285 {
 286         bool was_enabled = xen_selfballooning_enabled;
 287         unsigned long tmp;
 288         int err;
 289
 290         if (!capable(CAP_SYS_ADMIN))
 291                 return -EPERM;
 292
 293         err = strict_strtoul(buf, 10, &tmp);
 294         if (err || ((tmp != 0) && (tmp != 1)))
 295                 return -EINVAL;
 296
 297         xen_selfballooning_enabled = !!tmp;
 298         if (!was_enabled && xen_selfballooning_enabled)
 299                 schedule_delayed_work(&selfballoon_worker,
 300                         selfballoon_interval * HZ);
 301
 302         return count;
 303 }
 304
 305 static SYSDEV_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 306                    show_selfballooning, store_selfballooning);
 307
 308 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 309
 310 static ssize_t store_selfballoon_interval(struct sys_device *dev,
 311                                           struct sysdev_attribute *attr,
 312                                           const char *buf,
 313                                           size_t count)
 314 {
 315         unsigned long val;
 316         int err;
 317
 318         if (!capable(CAP_SYS_ADMIN))
 319                 return -EPERM;
 320         err = strict_strtoul(buf, 10, &val);
 321         if (err || val == 0)
 322                 return -EINVAL;
 323         selfballoon_interval = val;
 324         return count;
 325 }
 326
 327 static SYSDEV_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 328                    show_selfballoon_interval, store_selfballoon_interval);
 329
 330 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 331
 332 static ssize_t store_selfballoon_downhys(struct sys_device *dev,
 333                                          struct sysdev_attribute *attr,
 334                                          const char *buf,
 335                                          size_t count)
 336 {
 337         unsigned long val;
 338         int err;
 339
 340         if (!capable(CAP_SYS_ADMIN))
 341                 return -EPERM;
 342         err = strict_strtoul(buf, 10, &val);
 343         if (err || val == 0)
 344                 return -EINVAL;
 345         selfballoon_downhysteresis = val;
 346         return count;
 347 }
 348
 349 static SYSDEV_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 350                    show_selfballoon_downhys, store_selfballoon_downhys);
 351
 352
 353 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 354
 355 static ssize_t store_selfballoon_uphys(struct sys_device *dev,
 356                                        struct sysdev_attribute *attr,
 357                                        const char *buf,
 358                                        size_t count)
 359 {
 360         unsigned long val;
 361         int err;
 362
 363         if (!capable(CAP_SYS_ADMIN))
 364                 return -EPERM;
 365         err = strict_strtoul(buf, 10, &val);
 366         if (err || val == 0)
 367                 return -EINVAL;
 368         selfballoon_uphysteresis = val;
 369         return count;
 370 }
 371
 372 static SYSDEV_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 373                    show_selfballoon_uphys, store_selfballoon_uphys);
 374
 375 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 376                                 selfballoon_min_usable_mb);
 377
 378 static ssize_t store_selfballoon_min_usable_mb(struct sys_device *dev,
 379                                                struct sysdev_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_min_usable_mb = val;
 392         return count;
 393 }
 394
 395 static SYSDEV_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 396                    show_selfballoon_min_usable_mb,
 397                    store_selfballoon_min_usable_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 sys_device *dev,
 404                                              struct sysdev_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 SYSDEV_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 sys_device *dev,
 432                                        struct sysdev_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 SYSDEV_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 sys_device *dev,
 455                                           struct sysdev_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 SYSDEV_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         &attr_selfballooning.attr,
 478         &attr_selfballoon_interval.attr,
 479         &attr_selfballoon_downhysteresis.attr,
 480         &attr_selfballoon_uphysteresis.attr,
 481         &attr_selfballoon_min_usable_mb.attr,
 482 #ifdef CONFIG_FRONTSWAP
 483         &attr_frontswap_selfshrinking.attr,
 484         &attr_frontswap_hysteresis.attr,
 485         &attr_frontswap_inertia.attr,
 486 #endif
 487         NULL
 488 };
 489
 490 static struct attribute_group selfballoon_group = {
 491         .name = "selfballoon",
 492         .attrs = selfballoon_attrs
 493 };
 494 #endif
 495
 496 int register_xen_selfballooning(struct sys_device *sysdev)
 497 {
 498         int error = -1;
 499
 500 #ifdef CONFIG_SYSFS
 501         error = sysfs_create_group(&sysdev->kobj, &selfballoon_group);
 502 #endif
 503         return error;
 504 }
 505 EXPORT_SYMBOL(register_xen_selfballooning);
 506
 507 static int __init xen_selfballoon_init(void)
 508 {
 509         bool enable = false;
 510
 511         if (!xen_domain())
 512                 return -ENODEV;
 513
 514         if (xen_initial_domain()) {
 515                 pr_info("xen/balloon: Xen selfballooning driver "
 516                                 "disabled for domain0.\n");
 517                 return -ENODEV;
 518         }
 519
 520         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 521         if (xen_selfballooning_enabled) {
 522                 pr_info("xen/balloon: Initializing Xen "
 523                                         "selfballooning driver.\n");
 524                 enable = true;
 525         }
 526 #ifdef CONFIG_FRONTSWAP
 527         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 528         if (frontswap_selfshrinking) {
 529                 pr_info("xen/balloon: Initializing frontswap "
 530                                         "selfshrinking driver.\n");
 531                 enable = true;
 532         }
 533 #endif
 534         if (!enable)
 535                 return -ENODEV;
 536
 537         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 538
 539         return 0;
 540 }
 541
 542 subsys_initcall(xen_selfballoon_init);
 543
 544 MODULE_LICENSE("GPL");