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