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/stat.h>
  72 #include <linux/module.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 static void selfballoon_process(struct work_struct *work);
  98 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
  99
 100 #ifdef CONFIG_FRONTSWAP
 101 #include <linux/frontswap.h>
 102
 103 /* Enable/disable with sysfs. */
 104 static bool frontswap_selfshrinking __read_mostly;
 105
 106 /* Enable/disable with kernel boot option. */
 107 static bool use_frontswap_selfshrink __initdata = true;
 108
 109 /*
 110  * The default values for the following parameters were deemed reasonable
 111  * by experimentation, may be workload-dependent, and can all be
 112  * adjusted via sysfs.
 113  */
 114
 115 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 116 static unsigned int frontswap_hysteresis __read_mostly = 20;
 117
 118 /*
 119  * Number of selfballoon worker invocations to wait before observing that
 120  * frontswap selfshrinking should commence. Note that selfshrinking does
 121  * not use a separate worker thread.
 122  */
 123 static unsigned int frontswap_inertia __read_mostly = 3;
 124
 125 /* Countdown to next invocation of frontswap_shrink() */
 126 static unsigned long frontswap_inertia_counter;
 127
 128 /*
 129  * Invoked by the selfballoon worker thread, uses current number of pages
 130  * in frontswap (frontswap_curr_pages()), previous status, and control
 131  * values (hysteresis and inertia) to determine if frontswap should be
 132  * shrunk and what the new frontswap size should be.  Note that
 133  * frontswap_shrink is essentially a partial swapoff that immediately
 134  * transfers pages from the "swap device" (frontswap) back into kernel
 135  * RAM; despite the name, frontswap "shrinking" is very different from
 136  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 137  * memory.
 138  */
 139 static void frontswap_selfshrink(void)
 140 {
 141         static unsigned long cur_frontswap_pages;
 142         static unsigned long last_frontswap_pages;
 143         static unsigned long tgt_frontswap_pages;
 144
 145         last_frontswap_pages = cur_frontswap_pages;
 146         cur_frontswap_pages = frontswap_curr_pages();
 147         if (!cur_frontswap_pages ||
 148                         (cur_frontswap_pages > last_frontswap_pages)) {
 149                 frontswap_inertia_counter = frontswap_inertia;
 150                 return;
 151         }
 152         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 153                 return;
 154         if (cur_frontswap_pages <= frontswap_hysteresis)
 155                 tgt_frontswap_pages = 0;
 156         else
 157                 tgt_frontswap_pages = cur_frontswap_pages -
 158                         (cur_frontswap_pages / frontswap_hysteresis);
 159         frontswap_shrink(tgt_frontswap_pages);
 160 }
 161
 162 static int __init xen_nofrontswap_selfshrink_setup(char *s)
 163 {
 164         use_frontswap_selfshrink = false;
 165         return 1;
 166 }
 167
 168 __setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
 169
 170 /* Disable with kernel boot option. */
 171 static bool use_selfballooning __initdata = true;
 172
 173 static int __init xen_noselfballooning_setup(char *s)
 174 {
 175         use_selfballooning = false;
 176         return 1;
 177 }
 178
 179 __setup("noselfballooning", xen_noselfballooning_setup);
 180 #else /* !CONFIG_FRONTSWAP */
 181 /* Enable with kernel boot option. */
 182 static bool use_selfballooning __initdata = false;
 183
 184 static int __init xen_selfballooning_setup(char *s)
 185 {
 186         use_selfballooning = true;
 187         return 1;
 188 }
 189
 190 __setup("selfballooning", xen_selfballooning_setup);
 191 #endif /* CONFIG_FRONTSWAP */
 192
 193 /*
 194  * Use current balloon size, the goal (vm_committed_as), and hysteresis
 195  * parameters to set a new target balloon size
 196  */
 197 static void selfballoon_process(struct work_struct *work)
 198 {
 199         unsigned long cur_pages, goal_pages, tgt_pages;
 200         bool reset_timer = false;
 201
 202         if (xen_selfballooning_enabled) {
 203                 cur_pages = balloon_stats.current_pages;
 204                 tgt_pages = cur_pages; /* default is no change */
 205                 goal_pages = percpu_counter_read_positive(&vm_committed_as) +
 206                         balloon_stats.current_pages - totalram_pages;
 207 #ifdef CONFIG_FRONTSWAP
 208                 /* allow space for frontswap pages to be repatriated */
 209                 if (frontswap_selfshrinking && frontswap_enabled)
 210                         goal_pages += frontswap_curr_pages();
 211 #endif
 212                 if (cur_pages > goal_pages)
 213                         tgt_pages = cur_pages -
 214                                 ((cur_pages - goal_pages) /
 215                                   selfballoon_downhysteresis);
 216                 else if (cur_pages < goal_pages)
 217                         tgt_pages = cur_pages +
 218                                 ((goal_pages - cur_pages) /
 219                                   selfballoon_uphysteresis);
 220                 /* else if cur_pages == goal_pages, no change */
 221                 balloon_set_new_target(tgt_pages);
 222                 reset_timer = true;
 223         }
 224 #ifdef CONFIG_FRONTSWAP
 225         if (frontswap_selfshrinking && frontswap_enabled) {
 226                 frontswap_selfshrink();
 227                 reset_timer = true;
 228         }
 229 #endif
 230         if (reset_timer)
 231                 schedule_delayed_work(&selfballoon_worker,
 232                         selfballoon_interval * HZ);
 233 }
 234
 235 #ifdef CONFIG_SYSFS
 236
 237 #include <linux/sysdev.h>
 238 #include <linux/capability.h>
 239
 240 #define SELFBALLOON_SHOW(name, format, args...)                         \
 241         static ssize_t show_##name(struct sys_device *dev,      \
 242                                            struct sysdev_attribute *attr, \
 243                                            char *buf) \
 244         { \
 245                 return sprintf(buf, format, ##args); \
 246         }
 247
 248 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 249
 250 static ssize_t store_selfballooning(struct sys_device *dev,
 251                             struct sysdev_attribute *attr,
 252                             const char *buf,
 253                             size_t count)
 254 {
 255         bool was_enabled = xen_selfballooning_enabled;
 256         unsigned long tmp;
 257         int err;
 258
 259         if (!capable(CAP_SYS_ADMIN))
 260                 return -EPERM;
 261
 262         err = strict_strtoul(buf, 10, &tmp);
 263         if (err || ((tmp != 0) && (tmp != 1)))
 264                 return -EINVAL;
 265
 266         xen_selfballooning_enabled = !!tmp;
 267         if (!was_enabled && xen_selfballooning_enabled)
 268                 schedule_delayed_work(&selfballoon_worker,
 269                         selfballoon_interval * HZ);
 270
 271         return count;
 272 }
 273
 274 static SYSDEV_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 275                    show_selfballooning, store_selfballooning);
 276
 277 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 278
 279 static ssize_t store_selfballoon_interval(struct sys_device *dev,
 280                                           struct sysdev_attribute *attr,
 281                                           const char *buf,
 282                                           size_t count)
 283 {
 284         unsigned long val;
 285         int err;
 286
 287         if (!capable(CAP_SYS_ADMIN))
 288                 return -EPERM;
 289         err = strict_strtoul(buf, 10, &val);
 290         if (err || val == 0)
 291                 return -EINVAL;
 292         selfballoon_interval = val;
 293         return count;
 294 }
 295
 296 static SYSDEV_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 297                    show_selfballoon_interval, store_selfballoon_interval);
 298
 299 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 300
 301 static ssize_t store_selfballoon_downhys(struct sys_device *dev,
 302                                          struct sysdev_attribute *attr,
 303                                          const char *buf,
 304                                          size_t count)
 305 {
 306         unsigned long val;
 307         int err;
 308
 309         if (!capable(CAP_SYS_ADMIN))
 310                 return -EPERM;
 311         err = strict_strtoul(buf, 10, &val);
 312         if (err || val == 0)
 313                 return -EINVAL;
 314         selfballoon_downhysteresis = val;
 315         return count;
 316 }
 317
 318 static SYSDEV_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 319                    show_selfballoon_downhys, store_selfballoon_downhys);
 320
 321
 322 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 323
 324 static ssize_t store_selfballoon_uphys(struct sys_device *dev,
 325                                        struct sysdev_attribute *attr,
 326                                        const char *buf,
 327                                        size_t count)
 328 {
 329         unsigned long val;
 330         int err;
 331
 332         if (!capable(CAP_SYS_ADMIN))
 333                 return -EPERM;
 334         err = strict_strtoul(buf, 10, &val);
 335         if (err || val == 0)
 336                 return -EINVAL;
 337         selfballoon_uphysteresis = val;
 338         return count;
 339 }
 340
 341 static SYSDEV_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 342                    show_selfballoon_uphys, store_selfballoon_uphys);
 343
 344 #ifdef CONFIG_FRONTSWAP
 345 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 346
 347 static ssize_t store_frontswap_selfshrinking(struct sys_device *dev,
 348                                              struct sysdev_attribute *attr,
 349                                              const char *buf,
 350                                              size_t count)
 351 {
 352         bool was_enabled = frontswap_selfshrinking;
 353         unsigned long tmp;
 354         int err;
 355
 356         if (!capable(CAP_SYS_ADMIN))
 357                 return -EPERM;
 358         err = strict_strtoul(buf, 10, &tmp);
 359         if (err || ((tmp != 0) && (tmp != 1)))
 360                 return -EINVAL;
 361         frontswap_selfshrinking = !!tmp;
 362         if (!was_enabled && !xen_selfballooning_enabled &&
 363              frontswap_selfshrinking)
 364                 schedule_delayed_work(&selfballoon_worker,
 365                         selfballoon_interval * HZ);
 366
 367         return count;
 368 }
 369
 370 static SYSDEV_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 371                    show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 372
 373 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 374
 375 static ssize_t store_frontswap_inertia(struct sys_device *dev,
 376                                        struct sysdev_attribute *attr,
 377                                        const char *buf,
 378                                        size_t count)
 379 {
 380         unsigned long val;
 381         int err;
 382
 383         if (!capable(CAP_SYS_ADMIN))
 384                 return -EPERM;
 385         err = strict_strtoul(buf, 10, &val);
 386         if (err || val == 0)
 387                 return -EINVAL;
 388         frontswap_inertia = val;
 389         frontswap_inertia_counter = val;
 390         return count;
 391 }
 392
 393 static SYSDEV_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 394                    show_frontswap_inertia, store_frontswap_inertia);
 395
 396 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 397
 398 static ssize_t store_frontswap_hysteresis(struct sys_device *dev,
 399                                           struct sysdev_attribute *attr,
 400                                           const char *buf,
 401                                           size_t count)
 402 {
 403         unsigned long val;
 404         int err;
 405
 406         if (!capable(CAP_SYS_ADMIN))
 407                 return -EPERM;
 408         err = strict_strtoul(buf, 10, &val);
 409         if (err || val == 0)
 410                 return -EINVAL;
 411         frontswap_hysteresis = val;
 412         return count;
 413 }
 414
 415 static SYSDEV_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 416                    show_frontswap_hysteresis, store_frontswap_hysteresis);
 417
 418 #endif /* CONFIG_FRONTSWAP */
 419
 420 static struct attribute *selfballoon_attrs[] = {
 421         &attr_selfballooning.attr,
 422         &attr_selfballoon_interval.attr,
 423         &attr_selfballoon_downhysteresis.attr,
 424         &attr_selfballoon_uphysteresis.attr,
 425 #ifdef CONFIG_FRONTSWAP
 426         &attr_frontswap_selfshrinking.attr,
 427         &attr_frontswap_hysteresis.attr,
 428         &attr_frontswap_inertia.attr,
 429 #endif
 430         NULL
 431 };
 432
 433 static struct attribute_group selfballoon_group = {
 434         .name = "selfballoon",
 435         .attrs = selfballoon_attrs
 436 };
 437 #endif
 438
 439 int register_xen_selfballooning(struct sys_device *sysdev)
 440 {
 441         int error = -1;
 442
 443 #ifdef CONFIG_SYSFS
 444         error = sysfs_create_group(&sysdev->kobj, &selfballoon_group);
 445 #endif
 446         return error;
 447 }
 448 EXPORT_SYMBOL(register_xen_selfballooning);
 449
 450 static int __init xen_selfballoon_init(void)
 451 {
 452         bool enable = false;
 453
 454         if (!xen_domain())
 455                 return -ENODEV;
 456
 457         if (xen_initial_domain()) {
 458                 pr_info("xen/balloon: Xen selfballooning driver "
 459                                 "disabled for domain0.\n");
 460                 return -ENODEV;
 461         }
 462
 463         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 464         if (xen_selfballooning_enabled) {
 465                 pr_info("xen/balloon: Initializing Xen "
 466                                         "selfballooning driver.\n");
 467                 enable = true;
 468         }
 469 #ifdef CONFIG_FRONTSWAP
 470         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 471         if (frontswap_selfshrinking) {
 472                 pr_info("xen/balloon: Initializing frontswap "
 473                                         "selfshrinking driver.\n");
 474                 enable = true;
 475         }
 476 #endif
 477         if (!enable)
 478                 return -ENODEV;
 479
 480         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 481
 482         return 0;
 483 }
 484
 485 subsys_initcall(xen_selfballoon_init);
 486
 487 MODULE_LICENSE("GPL");