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