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. Note, that selfballooning should be disabled by default
  57  * if frontswap is not configured.  Similarly selfballooning should be enabled
  58  * by default if frontswap is configured and can be disabled with the
  59  * "tmem.selfballooning=0" kernel boot option.  Finally, when frontswap is
  60  * configured, frontswap-selfshrinking can be disabled  with the
  61  * "tmem.selfshrink=0" kernel boot option.
  62  *
  63  * Selfballooning is disallowed in domain0 and force-disabled.
  64  *
  65  */
  66
  67 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  68
  69 #include <linux/kernel.h>
  70 #include <linux/bootmem.h>
  71 #include <linux/swap.h>
  72 #include <linux/mm.h>
  73 #include <linux/mman.h>
  74 #include <linux/module.h>
  75 #include <linux/workqueue.h>
  76 #include <linux/device.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 /*
 108  * Amount of RAM in MB to add to the target number of pages.
 109  * Can be used to reserve some more room for caches and the like.
 110  */
 111 static unsigned int selfballoon_reserved_mb;
 112
 113 static void selfballoon_process(struct work_struct *work);
 114 static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
 115
 116 #ifdef CONFIG_FRONTSWAP
 117 #include <linux/frontswap.h>
 118
 119 /* Enable/disable with sysfs. */
 120 static bool frontswap_selfshrinking __read_mostly;
 121
 122 /*
 123  * The default values for the following parameters were deemed reasonable
 124  * by experimentation, may be workload-dependent, and can all be
 125  * adjusted via sysfs.
 126  */
 127
 128 /* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 129 static unsigned int frontswap_hysteresis __read_mostly = 20;
 130
 131 /*
 132  * Number of selfballoon worker invocations to wait before observing that
 133  * frontswap selfshrinking should commence. Note that selfshrinking does
 134  * not use a separate worker thread.
 135  */
 136 static unsigned int frontswap_inertia __read_mostly = 3;
 137
 138 /* Countdown to next invocation of frontswap_shrink() */
 139 static unsigned long frontswap_inertia_counter;
 140
 141 /*
 142  * Invoked by the selfballoon worker thread, uses current number of pages
 143  * in frontswap (frontswap_curr_pages()), previous status, and control
 144  * values (hysteresis and inertia) to determine if frontswap should be
 145  * shrunk and what the new frontswap size should be.  Note that
 146  * frontswap_shrink is essentially a partial swapoff that immediately
 147  * transfers pages from the "swap device" (frontswap) back into kernel
 148  * RAM; despite the name, frontswap "shrinking" is very different from
 149  * the "shrinker" interface used by the kernel MM subsystem to reclaim
 150  * memory.
 151  */
 152 static void frontswap_selfshrink(void)
 153 {
 154         static unsigned long cur_frontswap_pages;
 155         static unsigned long last_frontswap_pages;
 156         static unsigned long tgt_frontswap_pages;
 157
 158         last_frontswap_pages = cur_frontswap_pages;
 159         cur_frontswap_pages = frontswap_curr_pages();
 160         if (!cur_frontswap_pages ||
 161                         (cur_frontswap_pages > last_frontswap_pages)) {
 162                 frontswap_inertia_counter = frontswap_inertia;
 163                 return;
 164         }
 165         if (frontswap_inertia_counter && --frontswap_inertia_counter)
 166                 return;
 167         if (cur_frontswap_pages <= frontswap_hysteresis)
 168                 tgt_frontswap_pages = 0;
 169         else
 170                 tgt_frontswap_pages = cur_frontswap_pages -
 171                         (cur_frontswap_pages / frontswap_hysteresis);
 172         frontswap_shrink(tgt_frontswap_pages);
 173 }
 174
 175 #endif /* CONFIG_FRONTSWAP */
 176
 177 #define MB2PAGES(mb)    ((mb) << (20 - PAGE_SHIFT))
 178
 179 /*
 180  * Use current balloon size, the goal (vm_committed_as), and hysteresis
 181  * parameters to set a new target balloon size
 182  */
 183 static void selfballoon_process(struct work_struct *work)
 184 {
 185         unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
 186         unsigned long useful_pages;
 187         bool reset_timer = false;
 188
 189         if (xen_selfballooning_enabled) {
 190                 cur_pages = totalram_pages;
 191                 tgt_pages = cur_pages; /* default is no change */
 192                 goal_pages = vm_memory_committed() +
 193                                 totalreserve_pages +
 194                                 MB2PAGES(selfballoon_reserved_mb);
 195 #ifdef CONFIG_FRONTSWAP
 196                 /* allow space for frontswap pages to be repatriated */
 197                 if (frontswap_selfshrinking && frontswap_enabled)
 198                         goal_pages += frontswap_curr_pages();
 199 #endif
 200                 if (cur_pages > goal_pages)
 201                         tgt_pages = cur_pages -
 202                                 ((cur_pages - goal_pages) /
 203                                   selfballoon_downhysteresis);
 204                 else if (cur_pages < goal_pages)
 205                         tgt_pages = cur_pages +
 206                                 ((goal_pages - cur_pages) /
 207                                   selfballoon_uphysteresis);
 208                 /* else if cur_pages == goal_pages, no change */
 209                 useful_pages = max_pfn - totalreserve_pages;
 210                 if (selfballoon_min_usable_mb != 0)
 211                         floor_pages = totalreserve_pages +
 212                                         MB2PAGES(selfballoon_min_usable_mb);
 213                 /* piecewise linear function ending in ~3% slope */
 214                 else if (useful_pages < MB2PAGES(16))
 215                         floor_pages = max_pfn; /* not worth ballooning */
 216                 else if (useful_pages < MB2PAGES(64))
 217                         floor_pages = totalreserve_pages + MB2PAGES(16) +
 218                                         ((useful_pages - MB2PAGES(16)) >> 1);
 219                 else if (useful_pages < MB2PAGES(512))
 220                         floor_pages = totalreserve_pages + MB2PAGES(40) +
 221                                         ((useful_pages - MB2PAGES(40)) >> 3);
 222                 else /* useful_pages >= MB2PAGES(512) */
 223                         floor_pages = totalreserve_pages + MB2PAGES(99) +
 224                                         ((useful_pages - MB2PAGES(99)) >> 5);
 225                 if (tgt_pages < floor_pages)
 226                         tgt_pages = floor_pages;
 227                 balloon_set_new_target(tgt_pages +
 228                         balloon_stats.current_pages - totalram_pages);
 229                 reset_timer = true;
 230         }
 231 #ifdef CONFIG_FRONTSWAP
 232         if (frontswap_selfshrinking && frontswap_enabled) {
 233                 frontswap_selfshrink();
 234                 reset_timer = true;
 235         }
 236 #endif
 237         if (reset_timer)
 238                 schedule_delayed_work(&selfballoon_worker,
 239                         selfballoon_interval * HZ);
 240 }
 241
 242 #ifdef CONFIG_SYSFS
 243
 244 #include <linux/capability.h>
 245
 246 #define SELFBALLOON_SHOW(name, format, args...)                         \
 247         static ssize_t show_##name(struct device *dev,  \
 248                                           struct device_attribute *attr, \
 249                                           char *buf) \
 250         { \
 251                 return sprintf(buf, format, ##args); \
 252         }
 253
 254 SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 255
 256 static ssize_t store_selfballooning(struct device *dev,
 257                             struct device_attribute *attr,
 258                             const char *buf,
 259                             size_t count)
 260 {
 261         bool was_enabled = xen_selfballooning_enabled;
 262         unsigned long tmp;
 263         int err;
 264
 265         if (!capable(CAP_SYS_ADMIN))
 266                 return -EPERM;
 267
 268         err = kstrtoul(buf, 10, &tmp);
 269         if (err)
 270                 return err;
 271         if ((tmp != 0) && (tmp != 1))
 272                 return -EINVAL;
 273
 274         xen_selfballooning_enabled = !!tmp;
 275         if (!was_enabled && xen_selfballooning_enabled)
 276                 schedule_delayed_work(&selfballoon_worker,
 277                         selfballoon_interval * HZ);
 278
 279         return count;
 280 }
 281
 282 static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 283                    show_selfballooning, store_selfballooning);
 284
 285 SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 286
 287 static ssize_t store_selfballoon_interval(struct device *dev,
 288                                           struct device_attribute *attr,
 289                                           const char *buf,
 290                                           size_t count)
 291 {
 292         unsigned long val;
 293         int err;
 294
 295         if (!capable(CAP_SYS_ADMIN))
 296                 return -EPERM;
 297         err = kstrtoul(buf, 10, &val);
 298         if (err)
 299                 return err;
 300         if (val == 0)
 301                 return -EINVAL;
 302         selfballoon_interval = val;
 303         return count;
 304 }
 305
 306 static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 307                    show_selfballoon_interval, store_selfballoon_interval);
 308
 309 SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 310
 311 static ssize_t store_selfballoon_downhys(struct device *dev,
 312                                          struct device_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 = kstrtoul(buf, 10, &val);
 322         if (err)
 323                 return err;
 324         if (val == 0)
 325                 return -EINVAL;
 326         selfballoon_downhysteresis = val;
 327         return count;
 328 }
 329
 330 static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 331                    show_selfballoon_downhys, store_selfballoon_downhys);
 332
 333
 334 SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 335
 336 static ssize_t store_selfballoon_uphys(struct device *dev,
 337                                        struct device_attribute *attr,
 338                                        const char *buf,
 339                                        size_t count)
 340 {
 341         unsigned long val;
 342         int err;
 343
 344         if (!capable(CAP_SYS_ADMIN))
 345                 return -EPERM;
 346         err = kstrtoul(buf, 10, &val);
 347         if (err)
 348                 return err;
 349         if (val == 0)
 350                 return -EINVAL;
 351         selfballoon_uphysteresis = val;
 352         return count;
 353 }
 354
 355 static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 356                    show_selfballoon_uphys, store_selfballoon_uphys);
 357
 358 SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 359                                 selfballoon_min_usable_mb);
 360
 361 static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
 362                                                struct device_attribute *attr,
 363                                                const char *buf,
 364                                                size_t count)
 365 {
 366         unsigned long val;
 367         int err;
 368
 369         if (!capable(CAP_SYS_ADMIN))
 370                 return -EPERM;
 371         err = kstrtoul(buf, 10, &val);
 372         if (err)
 373                 return err;
 374         if (val == 0)
 375                 return -EINVAL;
 376         selfballoon_min_usable_mb = val;
 377         return count;
 378 }
 379
 380 static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 381                    show_selfballoon_min_usable_mb,
 382                    store_selfballoon_min_usable_mb);
 383
 384 SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
 385                                 selfballoon_reserved_mb);
 386
 387 static ssize_t store_selfballoon_reserved_mb(struct device *dev,
 388                                              struct device_attribute *attr,
 389                                              const char *buf,
 390                                              size_t count)
 391 {
 392         unsigned long val;
 393         int err;
 394
 395         if (!capable(CAP_SYS_ADMIN))
 396                 return -EPERM;
 397         err = kstrtoul(buf, 10, &val);
 398         if (err)
 399                 return err;
 400         if (val == 0)
 401                 return -EINVAL;
 402         selfballoon_reserved_mb = val;
 403         return count;
 404 }
 405
 406 static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
 407                    show_selfballoon_reserved_mb,
 408                    store_selfballoon_reserved_mb);
 409
 410
 411 #ifdef CONFIG_FRONTSWAP
 412 SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 413
 414 static ssize_t store_frontswap_selfshrinking(struct device *dev,
 415                                              struct device_attribute *attr,
 416                                              const char *buf,
 417                                              size_t count)
 418 {
 419         bool was_enabled = frontswap_selfshrinking;
 420         unsigned long tmp;
 421         int err;
 422
 423         if (!capable(CAP_SYS_ADMIN))
 424                 return -EPERM;
 425         err = kstrtoul(buf, 10, &tmp);
 426         if (err)
 427                 return err;
 428         if ((tmp != 0) && (tmp != 1))
 429                 return -EINVAL;
 430         frontswap_selfshrinking = !!tmp;
 431         if (!was_enabled && !xen_selfballooning_enabled &&
 432              frontswap_selfshrinking)
 433                 schedule_delayed_work(&selfballoon_worker,
 434                         selfballoon_interval * HZ);
 435
 436         return count;
 437 }
 438
 439 static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 440                    show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 441
 442 SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 443
 444 static ssize_t store_frontswap_inertia(struct device *dev,
 445                                        struct device_attribute *attr,
 446                                        const char *buf,
 447                                        size_t count)
 448 {
 449         unsigned long val;
 450         int err;
 451
 452         if (!capable(CAP_SYS_ADMIN))
 453                 return -EPERM;
 454         err = kstrtoul(buf, 10, &val);
 455         if (err)
 456                 return err;
 457         if (val == 0)
 458                 return -EINVAL;
 459         frontswap_inertia = val;
 460         frontswap_inertia_counter = val;
 461         return count;
 462 }
 463
 464 static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 465                    show_frontswap_inertia, store_frontswap_inertia);
 466
 467 SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 468
 469 static ssize_t store_frontswap_hysteresis(struct device *dev,
 470                                           struct device_attribute *attr,
 471                                           const char *buf,
 472                                           size_t count)
 473 {
 474         unsigned long val;
 475         int err;
 476
 477         if (!capable(CAP_SYS_ADMIN))
 478                 return -EPERM;
 479         err = kstrtoul(buf, 10, &val);
 480         if (err)
 481                 return err;
 482         if (val == 0)
 483                 return -EINVAL;
 484         frontswap_hysteresis = val;
 485         return count;
 486 }
 487
 488 static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 489                    show_frontswap_hysteresis, store_frontswap_hysteresis);
 490
 491 #endif /* CONFIG_FRONTSWAP */
 492
 493 static struct attribute *selfballoon_attrs[] = {
 494         &dev_attr_selfballooning.attr,
 495         &dev_attr_selfballoon_interval.attr,
 496         &dev_attr_selfballoon_downhysteresis.attr,
 497         &dev_attr_selfballoon_uphysteresis.attr,
 498         &dev_attr_selfballoon_min_usable_mb.attr,
 499         &dev_attr_selfballoon_reserved_mb.attr,
 500 #ifdef CONFIG_FRONTSWAP
 501         &dev_attr_frontswap_selfshrinking.attr,
 502         &dev_attr_frontswap_hysteresis.attr,
 503         &dev_attr_frontswap_inertia.attr,
 504 #endif
 505         NULL
 506 };
 507
 508 static const struct attribute_group selfballoon_group = {
 509         .name = "selfballoon",
 510         .attrs = selfballoon_attrs
 511 };
 512 #endif
 513
 514 int register_xen_selfballooning(struct device *dev)
 515 {
 516         int error = -1;
 517
 518 #ifdef CONFIG_SYSFS
 519         error = sysfs_create_group(&dev->kobj, &selfballoon_group);
 520 #endif
 521         return error;
 522 }
 523 EXPORT_SYMBOL(register_xen_selfballooning);
 524
 525 int xen_selfballoon_init(bool use_selfballooning, bool use_frontswap_selfshrink)
 526 {
 527         bool enable = false;
 528
 529         if (!xen_domain())
 530                 return -ENODEV;
 531
 532         if (xen_initial_domain()) {
 533                 pr_info("Xen selfballooning driver disabled for domain0\n");
 534                 return -ENODEV;
 535         }
 536
 537         xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 538         if (xen_selfballooning_enabled) {
 539                 pr_info("Initializing Xen selfballooning driver\n");
 540                 enable = true;
 541         }
 542 #ifdef CONFIG_FRONTSWAP
 543         frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 544         if (frontswap_selfshrinking) {
 545                 pr_info("Initializing frontswap selfshrinking driver\n");
 546                 enable = true;
 547         }
 548 #endif
 549         if (!enable)
 550                 return -ENODEV;
 551
 552         schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 553
 554         return 0;
 555 }
 556 EXPORT_SYMBOL(xen_selfballoon_init);