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