LXR linux/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/* Enable/disable with sysfs. */
  83static int xen_selfballooning_enabled __read_mostly;
  84
  85/*
  86 * Controls rate at which memory target (this iteration) approaches
  87 * ultimate goal when memory need is increasing (up-hysteresis) or
  88 * decreasing (down-hysteresis). Higher values of hysteresis cause
  89 * slower increases/decreases. The default values for the various
  90 * parameters were deemed reasonable by experimentation, may be
  91 * workload-dependent, and can all be adjusted via sysfs.
  92 */
  93static unsigned int selfballoon_downhysteresis __read_mostly = 8;
  94static unsigned int selfballoon_uphysteresis __read_mostly = 1;
  95
  96/* In HZ, controls frequency of worker invocation. */
  97static unsigned int selfballoon_interval __read_mostly = 5;
  98
  99/*
 100 * Minimum usable RAM in MB for selfballooning target for balloon.
 101 * If non-zero, it is added to totalreserve_pages and self-ballooning
 102 * will not balloon below the sum.  If zero, a piecewise linear function
 103 * is calculated as a minimum and added to totalreserve_pages.  Note that
 104 * setting this value indiscriminately may cause OOMs and crashes.
 105 */
 106static unsigned int selfballoon_min_usable_mb;
 107
 108/*
 109 * Amount of RAM in MB to add to the target number of pages.
 110 * Can be used to reserve some more room for caches and the like.
 111 */
 112static unsigned int selfballoon_reserved_mb;
 113
 114static void selfballoon_process(struct work_struct *work);
 115static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process);
 116
 117#ifdef CONFIG_FRONTSWAP
 118#include <linux/frontswap.h>
 119
 120/* Enable/disable with sysfs. */
 121static bool frontswap_selfshrinking __read_mostly;
 122
 123/* Enable/disable with kernel boot option. */
 124static bool use_frontswap_selfshrink __initdata = true;
 125
 126/*
 127 * The default values for the following parameters were deemed reasonable
 128 * by experimentation, may be workload-dependent, and can all be
 129 * adjusted via sysfs.
 130 */
 131
 132/* Control rate for frontswap shrinking. Higher hysteresis is slower. */
 133static unsigned int frontswap_hysteresis __read_mostly = 20;
 134
 135/*
 136 * Number of selfballoon worker invocations to wait before observing that
 137 * frontswap selfshrinking should commence. Note that selfshrinking does
 138 * not use a separate worker thread.
 139 */
 140static unsigned int frontswap_inertia __read_mostly = 3;
 141
 142/* Countdown to next invocation of frontswap_shrink() */
 143static unsigned long frontswap_inertia_counter;
 144
 145/*
 146 * Invoked by the selfballoon worker thread, uses current number of pages
 147 * in frontswap (frontswap_curr_pages()), previous status, and control
 148 * values (hysteresis and inertia) to determine if frontswap should be
 149 * shrunk and what the new frontswap size should be.  Note that
 150 * frontswap_shrink is essentially a partial swapoff that immediately
 151 * transfers pages from the "swap device" (frontswap) back into kernel
 152 * RAM; despite the name, frontswap "shrinking" is very different from
 153 * the "shrinker" interface used by the kernel MM subsystem to reclaim
 154 * memory.
 155 */
 156static void frontswap_selfshrink(void)
 157{
 158        static unsigned long cur_frontswap_pages;
 159        static unsigned long last_frontswap_pages;
 160        static unsigned long tgt_frontswap_pages;
 161
 162        last_frontswap_pages = cur_frontswap_pages;
 163        cur_frontswap_pages = frontswap_curr_pages();
 164        if (!cur_frontswap_pages ||
 165                        (cur_frontswap_pages > last_frontswap_pages)) {
 166                frontswap_inertia_counter = frontswap_inertia;
 167                return;
 168        }
 169        if (frontswap_inertia_counter && --frontswap_inertia_counter)
 170                return;
 171        if (cur_frontswap_pages <= frontswap_hysteresis)
 172                tgt_frontswap_pages = 0;
 173        else
 174                tgt_frontswap_pages = cur_frontswap_pages -
 175                        (cur_frontswap_pages / frontswap_hysteresis);
 176        frontswap_shrink(tgt_frontswap_pages);
 177}
 178
 179static int __init xen_nofrontswap_selfshrink_setup(char *s)
 180{
 181        use_frontswap_selfshrink = false;
 182        return 1;
 183}
 184
 185__setup("noselfshrink", xen_nofrontswap_selfshrink_setup);
 186
 187/* Disable with kernel boot option. */
 188static bool use_selfballooning __initdata = true;
 189
 190static int __init xen_noselfballooning_setup(char *s)
 191{
 192        use_selfballooning = false;
 193        return 1;
 194}
 195
 196__setup("noselfballooning", xen_noselfballooning_setup);
 197#else /* !CONFIG_FRONTSWAP */
 198/* Enable with kernel boot option. */
 199static bool use_selfballooning __initdata = false;
 200
 201static int __init xen_selfballooning_setup(char *s)
 202{
 203        use_selfballooning = true;
 204        return 1;
 205}
 206
 207__setup("selfballooning", xen_selfballooning_setup);
 208#endif /* CONFIG_FRONTSWAP */
 209
 210#define MB2PAGES(mb)    ((mb) << (20 - PAGE_SHIFT))
 211
 212/*
 213 * Use current balloon size, the goal (vm_committed_as), and hysteresis
 214 * parameters to set a new target balloon size
 215 */
 216static void selfballoon_process(struct work_struct *work)
 217{
 218        unsigned long cur_pages, goal_pages, tgt_pages, floor_pages;
 219        unsigned long useful_pages;
 220        bool reset_timer = false;
 221
 222        if (xen_selfballooning_enabled) {
 223                cur_pages = totalram_pages;
 224                tgt_pages = cur_pages; /* default is no change */
 225                goal_pages = percpu_counter_read_positive(&vm_committed_as) +
 226                                totalreserve_pages +
 227                                MB2PAGES(selfballoon_reserved_mb);
 228#ifdef CONFIG_FRONTSWAP
 229                /* allow space for frontswap pages to be repatriated */
 230                if (frontswap_selfshrinking && frontswap_enabled)
 231                        goal_pages += frontswap_curr_pages();
 232#endif
 233                if (cur_pages > goal_pages)
 234                        tgt_pages = cur_pages -
 235                                ((cur_pages - goal_pages) /
 236                                  selfballoon_downhysteresis);
 237                else if (cur_pages < goal_pages)
 238                        tgt_pages = cur_pages +
 239                                ((goal_pages - cur_pages) /
 240                                  selfballoon_uphysteresis);
 241                /* else if cur_pages == goal_pages, no change */
 242                useful_pages = max_pfn - totalreserve_pages;
 243                if (selfballoon_min_usable_mb != 0)
 244                        floor_pages = totalreserve_pages +
 245                                        MB2PAGES(selfballoon_min_usable_mb);
 246                /* piecewise linear function ending in ~3% slope */
 247                else if (useful_pages < MB2PAGES(16))
 248                        floor_pages = max_pfn; /* not worth ballooning */
 249                else if (useful_pages < MB2PAGES(64))
 250                        floor_pages = totalreserve_pages + MB2PAGES(16) +
 251                                        ((useful_pages - MB2PAGES(16)) >> 1);
 252                else if (useful_pages < MB2PAGES(512))
 253                        floor_pages = totalreserve_pages + MB2PAGES(40) +
 254                                        ((useful_pages - MB2PAGES(40)) >> 3);
 255                else /* useful_pages >= MB2PAGES(512) */
 256                        floor_pages = totalreserve_pages + MB2PAGES(99) +
 257                                        ((useful_pages - MB2PAGES(99)) >> 5);
 258                if (tgt_pages < floor_pages)
 259                        tgt_pages = floor_pages;
 260                balloon_set_new_target(tgt_pages +
 261                        balloon_stats.current_pages - totalram_pages);
 262                reset_timer = true;
 263        }
 264#ifdef CONFIG_FRONTSWAP
 265        if (frontswap_selfshrinking && frontswap_enabled) {
 266                frontswap_selfshrink();
 267                reset_timer = true;
 268        }
 269#endif
 270        if (reset_timer)
 271                schedule_delayed_work(&selfballoon_worker,
 272                        selfballoon_interval * HZ);
 273}
 274
 275#ifdef CONFIG_SYSFS
 276
 277#include <linux/capability.h>
 278
 279#define SELFBALLOON_SHOW(name, format, args...)                         \
 280        static ssize_t show_##name(struct device *dev,  \
 281                                          struct device_attribute *attr, \
 282                                          char *buf) \
 283        { \
 284                return sprintf(buf, format, ##args); \
 285        }
 286
 287SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled);
 288
 289static ssize_t store_selfballooning(struct device *dev,
 290                            struct device_attribute *attr,
 291                            const char *buf,
 292                            size_t count)
 293{
 294        bool was_enabled = xen_selfballooning_enabled;
 295        unsigned long tmp;
 296        int err;
 297
 298        if (!capable(CAP_SYS_ADMIN))
 299                return -EPERM;
 300
 301        err = strict_strtoul(buf, 10, &tmp);
 302        if (err || ((tmp != 0) && (tmp != 1)))
 303                return -EINVAL;
 304
 305        xen_selfballooning_enabled = !!tmp;
 306        if (!was_enabled && xen_selfballooning_enabled)
 307                schedule_delayed_work(&selfballoon_worker,
 308                        selfballoon_interval * HZ);
 309
 310        return count;
 311}
 312
 313static DEVICE_ATTR(selfballooning, S_IRUGO | S_IWUSR,
 314                   show_selfballooning, store_selfballooning);
 315
 316SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval);
 317
 318static ssize_t store_selfballoon_interval(struct device *dev,
 319                                          struct device_attribute *attr,
 320                                          const char *buf,
 321                                          size_t count)
 322{
 323        unsigned long val;
 324        int err;
 325
 326        if (!capable(CAP_SYS_ADMIN))
 327                return -EPERM;
 328        err = strict_strtoul(buf, 10, &val);
 329        if (err || val == 0)
 330                return -EINVAL;
 331        selfballoon_interval = val;
 332        return count;
 333}
 334
 335static DEVICE_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR,
 336                   show_selfballoon_interval, store_selfballoon_interval);
 337
 338SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis);
 339
 340static ssize_t store_selfballoon_downhys(struct device *dev,
 341                                         struct device_attribute *attr,
 342                                         const char *buf,
 343                                         size_t count)
 344{
 345        unsigned long val;
 346        int err;
 347
 348        if (!capable(CAP_SYS_ADMIN))
 349                return -EPERM;
 350        err = strict_strtoul(buf, 10, &val);
 351        if (err || val == 0)
 352                return -EINVAL;
 353        selfballoon_downhysteresis = val;
 354        return count;
 355}
 356
 357static DEVICE_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR,
 358                   show_selfballoon_downhys, store_selfballoon_downhys);
 359
 360
 361SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis);
 362
 363static ssize_t store_selfballoon_uphys(struct device *dev,
 364                                       struct device_attribute *attr,
 365                                       const char *buf,
 366                                       size_t count)
 367{
 368        unsigned long val;
 369        int err;
 370
 371        if (!capable(CAP_SYS_ADMIN))
 372                return -EPERM;
 373        err = strict_strtoul(buf, 10, &val);
 374        if (err || val == 0)
 375                return -EINVAL;
 376        selfballoon_uphysteresis = val;
 377        return count;
 378}
 379
 380static DEVICE_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR,
 381                   show_selfballoon_uphys, store_selfballoon_uphys);
 382
 383SELFBALLOON_SHOW(selfballoon_min_usable_mb, "%d\n",
 384                                selfballoon_min_usable_mb);
 385
 386static ssize_t store_selfballoon_min_usable_mb(struct device *dev,
 387                                               struct device_attribute *attr,
 388                                               const char *buf,
 389                                               size_t count)
 390{
 391        unsigned long val;
 392        int err;
 393
 394        if (!capable(CAP_SYS_ADMIN))
 395                return -EPERM;
 396        err = strict_strtoul(buf, 10, &val);
 397        if (err || val == 0)
 398                return -EINVAL;
 399        selfballoon_min_usable_mb = val;
 400        return count;
 401}
 402
 403static DEVICE_ATTR(selfballoon_min_usable_mb, S_IRUGO | S_IWUSR,
 404                   show_selfballoon_min_usable_mb,
 405                   store_selfballoon_min_usable_mb);
 406
 407SELFBALLOON_SHOW(selfballoon_reserved_mb, "%d\n",
 408                                selfballoon_reserved_mb);
 409
 410static ssize_t store_selfballoon_reserved_mb(struct device *dev,
 411                                             struct device_attribute *attr,
 412                                             const char *buf,
 413                                             size_t count)
 414{
 415        unsigned long val;
 416        int err;
 417
 418        if (!capable(CAP_SYS_ADMIN))
 419                return -EPERM;
 420        err = strict_strtoul(buf, 10, &val);
 421        if (err || val == 0)
 422                return -EINVAL;
 423        selfballoon_reserved_mb = val;
 424        return count;
 425}
 426
 427static DEVICE_ATTR(selfballoon_reserved_mb, S_IRUGO | S_IWUSR,
 428                   show_selfballoon_reserved_mb,
 429                   store_selfballoon_reserved_mb);
 430
 431
 432#ifdef CONFIG_FRONTSWAP
 433SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking);
 434
 435static ssize_t store_frontswap_selfshrinking(struct device *dev,
 436                                             struct device_attribute *attr,
 437                                             const char *buf,
 438                                             size_t count)
 439{
 440        bool was_enabled = frontswap_selfshrinking;
 441        unsigned long tmp;
 442        int err;
 443
 444        if (!capable(CAP_SYS_ADMIN))
 445                return -EPERM;
 446        err = strict_strtoul(buf, 10, &tmp);
 447        if (err || ((tmp != 0) && (tmp != 1)))
 448                return -EINVAL;
 449        frontswap_selfshrinking = !!tmp;
 450        if (!was_enabled && !xen_selfballooning_enabled &&
 451             frontswap_selfshrinking)
 452                schedule_delayed_work(&selfballoon_worker,
 453                        selfballoon_interval * HZ);
 454
 455        return count;
 456}
 457
 458static DEVICE_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR,
 459                   show_frontswap_selfshrinking, store_frontswap_selfshrinking);
 460
 461SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia);
 462
 463static ssize_t store_frontswap_inertia(struct device *dev,
 464                                       struct device_attribute *attr,
 465                                       const char *buf,
 466                                       size_t count)
 467{
 468        unsigned long val;
 469        int err;
 470
 471        if (!capable(CAP_SYS_ADMIN))
 472                return -EPERM;
 473        err = strict_strtoul(buf, 10, &val);
 474        if (err || val == 0)
 475                return -EINVAL;
 476        frontswap_inertia = val;
 477        frontswap_inertia_counter = val;
 478        return count;
 479}
 480
 481static DEVICE_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR,
 482                   show_frontswap_inertia, store_frontswap_inertia);
 483
 484SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis);
 485
 486static ssize_t store_frontswap_hysteresis(struct device *dev,
 487                                          struct device_attribute *attr,
 488                                          const char *buf,
 489                                          size_t count)
 490{
 491        unsigned long val;
 492        int err;
 493
 494        if (!capable(CAP_SYS_ADMIN))
 495                return -EPERM;
 496        err = strict_strtoul(buf, 10, &val);
 497        if (err || val == 0)
 498                return -EINVAL;
 499        frontswap_hysteresis = val;
 500        return count;
 501}
 502
 503static DEVICE_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR,
 504                   show_frontswap_hysteresis, store_frontswap_hysteresis);
 505
 506#endif /* CONFIG_FRONTSWAP */
 507
 508static struct attribute *selfballoon_attrs[] = {
 509        &dev_attr_selfballooning.attr,
 510        &dev_attr_selfballoon_interval.attr,
 511        &dev_attr_selfballoon_downhysteresis.attr,
 512        &dev_attr_selfballoon_uphysteresis.attr,
 513        &dev_attr_selfballoon_min_usable_mb.attr,
 514        &dev_attr_selfballoon_reserved_mb.attr,
 515#ifdef CONFIG_FRONTSWAP
 516        &dev_attr_frontswap_selfshrinking.attr,
 517        &dev_attr_frontswap_hysteresis.attr,
 518        &dev_attr_frontswap_inertia.attr,
 519#endif
 520        NULL
 521};
 522
 523static const struct attribute_group selfballoon_group = {
 524        .name = "selfballoon",
 525        .attrs = selfballoon_attrs
 526};
 527#endif
 528
 529int register_xen_selfballooning(struct device *dev)
 530{
 531        int error = -1;
 532
 533#ifdef CONFIG_SYSFS
 534        error = sysfs_create_group(&dev->kobj, &selfballoon_group);
 535#endif
 536        return error;
 537}
 538EXPORT_SYMBOL(register_xen_selfballooning);
 539
 540static int __init xen_selfballoon_init(void)
 541{
 542        bool enable = false;
 543
 544        if (!xen_domain())
 545                return -ENODEV;
 546
 547        if (xen_initial_domain()) {
 548                pr_info("xen/balloon: Xen selfballooning driver "
 549                                "disabled for domain0.\n");
 550                return -ENODEV;
 551        }
 552
 553        xen_selfballooning_enabled = tmem_enabled && use_selfballooning;
 554        if (xen_selfballooning_enabled) {
 555                pr_info("xen/balloon: Initializing Xen "
 556                                        "selfballooning driver.\n");
 557                enable = true;
 558        }
 559#ifdef CONFIG_FRONTSWAP
 560        frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink;
 561        if (frontswap_selfshrinking) {
 562                pr_info("xen/balloon: Initializing frontswap "
 563                                        "selfshrinking driver.\n");
 564                enable = true;
 565        }
 566#endif
 567        if (!enable)
 568                return -ENODEV;
 569
 570        schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ);
 571
 572        return 0;
 573}
 574
 575subsys_initcall(xen_selfballoon_init);
 576
 577MODULE_LICENSE("GPL");
 578