From: Dan Magenheimer Date: Fri, 8 Jul 2011 18:26:21 +0000 (-0600) Subject: xen: tmem: self-ballooning and frontswap-selfshrinking X-Git-Url: http://git.cdn.openwrt.org/?a=commitdiff_plain;h=a50777c791031d7345ce95785ea6220f67339d90;p=openwrt%2Fstaging%2Fblogic.git xen: tmem: self-ballooning and frontswap-selfshrinking This patch introduces two in-kernel drivers for Xen transcendent memory ("tmem") functionality that complement cleancache and frontswap. Both use control theory to dynamically adjust and optimize memory utilization. Selfballooning controls the in-kernel Xen balloon driver, targeting a goal value (vm_committed_as), thus pushing less frequently used clean page cache pages (through the cleancache code) into Xen tmem where Xen can balance needs across all VMs residing on the physical machine. Frontswap-selfshrinking controls the number of pages in frontswap, driving it towards zero (effectively doing a partial swapoff) when in-kernel memory pressure subsides, freeing up RAM for other VMs. More detail is provided in the header comment of xen-selfballooning.c. Signed-off-by: Dan Magenheimer [v8: konrad.wilk@oracle.com: set default enablement depending on frontswap] [v7: konrad.wilk@oracle.com: fix capitalization and punctuation in comments] [v6: fix frontswap-selfshrinking initialization] [v6: konrad.wilk@oracle.com: fix init pr_infos; add comments about swap] [v5: konrad.wilk@oracle.com: add NULL to attr list; move inits up to decls] [v4: dkiper@net-space.pl: use strict_strtoul plus a few syntactic nits] [v3: konrad.wilk@oracle.com: fix potential divides-by-zero] [v3: konrad.wilk@oracle.com: add many more comments, fix nits] [v2: rebased to linux-3.0-rc1] [v2: Ian.Campbell@citrix.com: reorganize as new file (xen-selfballoon.c)] [v2: dkiper@net-space.pl: proper access to vm_committed_as] [v2: dkiper@net-space.pl: accounting fixes] Cc: Jan Beulich Cc: Jeremy Fitzhardinge Cc: --- diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig index a59638b37c1a..fd3266696a4d 100644 --- a/drivers/xen/Kconfig +++ b/drivers/xen/Kconfig @@ -9,6 +9,23 @@ config XEN_BALLOON the system to expand the domain's memory allocation, or alternatively return unneeded memory to the system. +config XEN_SELFBALLOONING + bool "Dynamically self-balloon kernel memory to target" + depends on XEN && XEN_BALLOON && CLEANCACHE && SWAP + default n + help + Self-ballooning dynamically balloons available kernel memory driven + by the current usage of anonymous memory ("committed AS") and + controlled by various sysfs-settable parameters. Configuring + FRONTSWAP is highly recommended; if it is not configured, self- + ballooning is disabled by default but can be enabled with the + 'selfballooning' kernel boot parameter. If FRONTSWAP is configured, + frontswap-selfshrinking is enabled by default but can be disabled + with the 'noselfshrink' kernel boot parameter; and self-ballooning + is enabled by default but can be disabled with the 'noselfballooning' + kernel boot parameter. Note that systems without a sufficiently + large swap device should not enable self-ballooning. + config XEN_SCRUB_PAGES bool "Scrub pages before returning them to system" depends on XEN_BALLOON diff --git a/drivers/xen/Makefile b/drivers/xen/Makefile index bbc18258ecc5..458057464dd1 100644 --- a/drivers/xen/Makefile +++ b/drivers/xen/Makefile @@ -9,6 +9,7 @@ obj-$(CONFIG_BLOCK) += biomerge.o obj-$(CONFIG_HOTPLUG_CPU) += cpu_hotplug.o obj-$(CONFIG_XEN_XENCOMM) += xencomm.o obj-$(CONFIG_XEN_BALLOON) += xen-balloon.o +obj-$(CONFIG_XEN_SELFBALLOONING) += xen-selfballoon.o obj-$(CONFIG_XEN_DEV_EVTCHN) += xen-evtchn.o obj-$(CONFIG_XEN_GNTDEV) += xen-gntdev.o obj-$(CONFIG_XEN_GRANT_DEV_ALLOC) += xen-gntalloc.o diff --git a/drivers/xen/xen-balloon.c b/drivers/xen/xen-balloon.c index a4ff225ee868..5c9dc43c1e94 100644 --- a/drivers/xen/xen-balloon.c +++ b/drivers/xen/xen-balloon.c @@ -98,6 +98,8 @@ static int __init balloon_init(void) register_balloon(&balloon_sysdev); + register_xen_selfballooning(&balloon_sysdev); + target_watch.callback = watch_target; xenstore_notifier.notifier_call = balloon_init_watcher; diff --git a/drivers/xen/xen-selfballoon.c b/drivers/xen/xen-selfballoon.c new file mode 100644 index 000000000000..010937b5a7c9 --- /dev/null +++ b/drivers/xen/xen-selfballoon.c @@ -0,0 +1,485 @@ +/****************************************************************************** + * Xen selfballoon driver (and optional frontswap self-shrinking driver) + * + * Copyright (c) 2009-2011, Dan Magenheimer, Oracle Corp. + * + * This code complements the cleancache and frontswap patchsets to optimize + * support for Xen Transcendent Memory ("tmem"). The policy it implements + * is rudimentary and will likely improve over time, but it does work well + * enough today. + * + * Two functionalities are implemented here which both use "control theory" + * (feedback) to optimize memory utilization. In a virtualized environment + * such as Xen, RAM is often a scarce resource and we would like to ensure + * that each of a possibly large number of virtual machines is using RAM + * efficiently, i.e. using as little as possible when under light load + * and obtaining as much as possible when memory demands are high. + * Since RAM needs vary highly dynamically and sometimes dramatically, + * "hysteresis" is used, that is, memory target is determined not just + * on current data but also on past data stored in the system. + * + * "Selfballooning" creates memory pressure by managing the Xen balloon + * driver to decrease and increase available kernel memory, driven + * largely by the target value of "Committed_AS" (see /proc/meminfo). + * Since Committed_AS does not account for clean mapped pages (i.e. pages + * in RAM that are identical to pages on disk), selfballooning has the + * affect of pushing less frequently used clean pagecache pages out of + * kernel RAM and, presumably using cleancache, into Xen tmem where + * Xen can more efficiently optimize RAM utilization for such pages. + * + * When kernel memory demand unexpectedly increases faster than Xen, via + * the selfballoon driver, is able to (or chooses to) provide usable RAM, + * the kernel may invoke swapping. In most cases, frontswap is able + * to absorb this swapping into Xen tmem. However, due to the fact + * that the kernel swap subsystem assumes swapping occurs to a disk, + * swapped pages may sit on the disk for a very long time; even if + * the kernel knows the page will never be used again. This is because + * the disk space costs very little and can be overwritten when + * necessary. When such stale pages are in frontswap, however, they + * are taking up valuable real estate. "Frontswap selfshrinking" works + * to resolve this: When frontswap activity is otherwise stable + * and the guest kernel is not under memory pressure, the "frontswap + * selfshrinking" accounts for this by providing pressure to remove some + * pages from frontswap and return them to kernel memory. + * + * For both "selfballooning" and "frontswap-selfshrinking", a worker + * thread is used and sysfs tunables are provided to adjust the frequency + * and rate of adjustments to achieve the goal, as well as to disable one + * or both functions independently. + * + * While some argue that this functionality can and should be implemented + * in userspace, it has been observed that bad things happen (e.g. OOMs). + * + * System configuration note: Selfballooning should not be enabled on + * systems without a sufficiently large swap device configured; for best + * results, it is recommended that total swap be increased by the size + * of the guest memory. Also, while technically not required to be + * configured, it is highly recommended that frontswap also be configured + * and enabled when selfballooning is running. So, selfballooning + * is disabled by default if frontswap is not configured and can only + * be enabled with the "selfballooning" kernel boot option; similarly + * selfballooning is enabled by default if frontswap is configured and + * can be disabled with the "noselfballooning" kernel boot option. Finally, + * when frontswap is configured, frontswap-selfshrinking can be disabled + * with the "noselfshrink" kernel boot option. + * + * Selfballooning is disallowed in domain0 and force-disabled. + * + */ + +#include +#include +#include + +#include + +#include + +/* Enable/disable with sysfs. */ +static int xen_selfballooning_enabled __read_mostly; + +/* + * Controls rate at which memory target (this iteration) approaches + * ultimate goal when memory need is increasing (up-hysteresis) or + * decreasing (down-hysteresis). Higher values of hysteresis cause + * slower increases/decreases. The default values for the various + * parameters were deemed reasonable by experimentation, may be + * workload-dependent, and can all be adjusted via sysfs. + */ +static unsigned int selfballoon_downhysteresis __read_mostly = 8; +static unsigned int selfballoon_uphysteresis __read_mostly = 1; + +/* In HZ, controls frequency of worker invocation. */ +static unsigned int selfballoon_interval __read_mostly = 5; + +static void selfballoon_process(struct work_struct *work); +static DECLARE_DELAYED_WORK(selfballoon_worker, selfballoon_process); + +#ifdef CONFIG_FRONTSWAP +#include + +/* Enable/disable with sysfs. */ +static bool frontswap_selfshrinking __read_mostly; + +/* Enable/disable with kernel boot option. */ +static bool use_frontswap_selfshrink __initdata = true; + +/* + * The default values for the following parameters were deemed reasonable + * by experimentation, may be workload-dependent, and can all be + * adjusted via sysfs. + */ + +/* Control rate for frontswap shrinking. Higher hysteresis is slower. */ +static unsigned int frontswap_hysteresis __read_mostly = 20; + +/* + * Number of selfballoon worker invocations to wait before observing that + * frontswap selfshrinking should commence. Note that selfshrinking does + * not use a separate worker thread. + */ +static unsigned int frontswap_inertia __read_mostly = 3; + +/* Countdown to next invocation of frontswap_shrink() */ +static unsigned long frontswap_inertia_counter; + +/* + * Invoked by the selfballoon worker thread, uses current number of pages + * in frontswap (frontswap_curr_pages()), previous status, and control + * values (hysteresis and inertia) to determine if frontswap should be + * shrunk and what the new frontswap size should be. Note that + * frontswap_shrink is essentially a partial swapoff that immediately + * transfers pages from the "swap device" (frontswap) back into kernel + * RAM; despite the name, frontswap "shrinking" is very different from + * the "shrinker" interface used by the kernel MM subsystem to reclaim + * memory. + */ +static void frontswap_selfshrink(void) +{ + static unsigned long cur_frontswap_pages; + static unsigned long last_frontswap_pages; + static unsigned long tgt_frontswap_pages; + + last_frontswap_pages = cur_frontswap_pages; + cur_frontswap_pages = frontswap_curr_pages(); + if (!cur_frontswap_pages || + (cur_frontswap_pages > last_frontswap_pages)) { + frontswap_inertia_counter = frontswap_inertia; + return; + } + if (frontswap_inertia_counter && --frontswap_inertia_counter) + return; + if (cur_frontswap_pages <= frontswap_hysteresis) + tgt_frontswap_pages = 0; + else + tgt_frontswap_pages = cur_frontswap_pages - + (cur_frontswap_pages / frontswap_hysteresis); + frontswap_shrink(tgt_frontswap_pages); +} + +static int __init xen_nofrontswap_selfshrink_setup(char *s) +{ + use_frontswap_selfshrink = false; + return 1; +} + +__setup("noselfshrink", xen_nofrontswap_selfshrink_setup); + +/* Disable with kernel boot option. */ +static bool use_selfballooning __initdata = true; + +static int __init xen_noselfballooning_setup(char *s) +{ + use_selfballooning = false; + return 1; +} + +__setup("noselfballooning", xen_noselfballooning_setup); +#else /* !CONFIG_FRONTSWAP */ +/* Enable with kernel boot option. */ +static bool use_selfballooning __initdata = false; + +static int __init xen_selfballooning_setup(char *s) +{ + use_selfballooning = true; + return 1; +} + +__setup("selfballooning", xen_selfballooning_setup); +#endif /* CONFIG_FRONTSWAP */ + +/* + * Use current balloon size, the goal (vm_committed_as), and hysteresis + * parameters to set a new target balloon size + */ +static void selfballoon_process(struct work_struct *work) +{ + unsigned long cur_pages, goal_pages, tgt_pages; + bool reset_timer = false; + + if (xen_selfballooning_enabled) { + cur_pages = balloon_stats.current_pages; + tgt_pages = cur_pages; /* default is no change */ + goal_pages = percpu_counter_read_positive(&vm_committed_as) + + balloon_stats.current_pages - totalram_pages; +#ifdef CONFIG_FRONTSWAP + /* allow space for frontswap pages to be repatriated */ + if (frontswap_selfshrinking && frontswap_enabled) + goal_pages += frontswap_curr_pages(); +#endif + if (cur_pages > goal_pages) + tgt_pages = cur_pages - + ((cur_pages - goal_pages) / + selfballoon_downhysteresis); + else if (cur_pages < goal_pages) + tgt_pages = cur_pages + + ((goal_pages - cur_pages) / + selfballoon_uphysteresis); + /* else if cur_pages == goal_pages, no change */ + balloon_set_new_target(tgt_pages); + reset_timer = true; + } +#ifdef CONFIG_FRONTSWAP + if (frontswap_selfshrinking && frontswap_enabled) { + frontswap_selfshrink(); + reset_timer = true; + } +#endif + if (reset_timer) + schedule_delayed_work(&selfballoon_worker, + selfballoon_interval * HZ); +} + +#ifdef CONFIG_SYSFS + +#include +#include + +#define SELFBALLOON_SHOW(name, format, args...) \ + static ssize_t show_##name(struct sys_device *dev, \ + struct sysdev_attribute *attr, \ + char *buf) \ + { \ + return sprintf(buf, format, ##args); \ + } + +SELFBALLOON_SHOW(selfballooning, "%d\n", xen_selfballooning_enabled); + +static ssize_t store_selfballooning(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + bool was_enabled = xen_selfballooning_enabled; + unsigned long tmp; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + + err = strict_strtoul(buf, 10, &tmp); + if (err || ((tmp != 0) && (tmp != 1))) + return -EINVAL; + + xen_selfballooning_enabled = !!tmp; + if (!was_enabled && xen_selfballooning_enabled) + schedule_delayed_work(&selfballoon_worker, + selfballoon_interval * HZ); + + return count; +} + +static SYSDEV_ATTR(selfballooning, S_IRUGO | S_IWUSR, + show_selfballooning, store_selfballooning); + +SELFBALLOON_SHOW(selfballoon_interval, "%d\n", selfballoon_interval); + +static ssize_t store_selfballoon_interval(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + unsigned long val; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &val); + if (err || val == 0) + return -EINVAL; + selfballoon_interval = val; + return count; +} + +static SYSDEV_ATTR(selfballoon_interval, S_IRUGO | S_IWUSR, + show_selfballoon_interval, store_selfballoon_interval); + +SELFBALLOON_SHOW(selfballoon_downhys, "%d\n", selfballoon_downhysteresis); + +static ssize_t store_selfballoon_downhys(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + unsigned long val; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &val); + if (err || val == 0) + return -EINVAL; + selfballoon_downhysteresis = val; + return count; +} + +static SYSDEV_ATTR(selfballoon_downhysteresis, S_IRUGO | S_IWUSR, + show_selfballoon_downhys, store_selfballoon_downhys); + + +SELFBALLOON_SHOW(selfballoon_uphys, "%d\n", selfballoon_uphysteresis); + +static ssize_t store_selfballoon_uphys(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + unsigned long val; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &val); + if (err || val == 0) + return -EINVAL; + selfballoon_uphysteresis = val; + return count; +} + +static SYSDEV_ATTR(selfballoon_uphysteresis, S_IRUGO | S_IWUSR, + show_selfballoon_uphys, store_selfballoon_uphys); + +#ifdef CONFIG_FRONTSWAP +SELFBALLOON_SHOW(frontswap_selfshrinking, "%d\n", frontswap_selfshrinking); + +static ssize_t store_frontswap_selfshrinking(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + bool was_enabled = frontswap_selfshrinking; + unsigned long tmp; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &tmp); + if (err || ((tmp != 0) && (tmp != 1))) + return -EINVAL; + frontswap_selfshrinking = !!tmp; + if (!was_enabled && !xen_selfballooning_enabled && + frontswap_selfshrinking) + schedule_delayed_work(&selfballoon_worker, + selfballoon_interval * HZ); + + return count; +} + +static SYSDEV_ATTR(frontswap_selfshrinking, S_IRUGO | S_IWUSR, + show_frontswap_selfshrinking, store_frontswap_selfshrinking); + +SELFBALLOON_SHOW(frontswap_inertia, "%d\n", frontswap_inertia); + +static ssize_t store_frontswap_inertia(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + unsigned long val; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &val); + if (err || val == 0) + return -EINVAL; + frontswap_inertia = val; + frontswap_inertia_counter = val; + return count; +} + +static SYSDEV_ATTR(frontswap_inertia, S_IRUGO | S_IWUSR, + show_frontswap_inertia, store_frontswap_inertia); + +SELFBALLOON_SHOW(frontswap_hysteresis, "%d\n", frontswap_hysteresis); + +static ssize_t store_frontswap_hysteresis(struct sys_device *dev, + struct sysdev_attribute *attr, + const char *buf, + size_t count) +{ + unsigned long val; + int err; + + if (!capable(CAP_SYS_ADMIN)) + return -EPERM; + err = strict_strtoul(buf, 10, &val); + if (err || val == 0) + return -EINVAL; + frontswap_hysteresis = val; + return count; +} + +static SYSDEV_ATTR(frontswap_hysteresis, S_IRUGO | S_IWUSR, + show_frontswap_hysteresis, store_frontswap_hysteresis); + +#endif /* CONFIG_FRONTSWAP */ + +static struct attribute *selfballoon_attrs[] = { + &attr_selfballooning.attr, + &attr_selfballoon_interval.attr, + &attr_selfballoon_downhysteresis.attr, + &attr_selfballoon_uphysteresis.attr, +#ifdef CONFIG_FRONTSWAP + &attr_frontswap_selfshrinking.attr, + &attr_frontswap_hysteresis.attr, + &attr_frontswap_inertia.attr, +#endif + NULL +}; + +static struct attribute_group selfballoon_group = { + .name = "selfballoon", + .attrs = selfballoon_attrs +}; +#endif + +int register_xen_selfballooning(struct sys_device *sysdev) +{ + int error = -1; + +#ifdef CONFIG_SYSFS + error = sysfs_create_group(&sysdev->kobj, &selfballoon_group); +#endif + return error; +} +EXPORT_SYMBOL(register_xen_selfballooning); + +static int __init xen_selfballoon_init(void) +{ + bool enable = false; + + if (!xen_domain()) + return -ENODEV; + + if (xen_initial_domain()) { + pr_info("xen/balloon: Xen selfballooning driver " + "disabled for domain0.\n"); + return -ENODEV; + } + + xen_selfballooning_enabled = tmem_enabled && use_selfballooning; + if (xen_selfballooning_enabled) { + pr_info("xen/balloon: Initializing Xen " + "selfballooning driver.\n"); + enable = true; + } +#ifdef CONFIG_FRONTSWAP + frontswap_selfshrinking = tmem_enabled && use_frontswap_selfshrink; + if (frontswap_selfshrinking) { + pr_info("xen/balloon: Initializing frontswap " + "selfshrinking driver.\n"); + enable = true; + } +#endif + if (!enable) + return -ENODEV; + + schedule_delayed_work(&selfballoon_worker, selfballoon_interval * HZ); + + return 0; +} + +subsys_initcall(xen_selfballoon_init); + +MODULE_LICENSE("GPL"); diff --git a/include/xen/balloon.h b/include/xen/balloon.h index a2b22f01a51d..4076ed72afbd 100644 --- a/include/xen/balloon.h +++ b/include/xen/balloon.h @@ -23,3 +23,13 @@ void balloon_set_new_target(unsigned long target); int alloc_xenballooned_pages(int nr_pages, struct page** pages); void free_xenballooned_pages(int nr_pages, struct page** pages); + +struct sys_device; +#ifdef CONFIG_XEN_SELFBALLOONING +extern int register_xen_selfballooning(struct sys_device *sysdev); +#else +static inline int register_xen_selfballooning(struct sys_device *sysdev) +{ + return -ENOSYS; +} +#endif diff --git a/include/xen/tmem.h b/include/xen/tmem.h new file mode 100644 index 000000000000..82e2c83a32f5 --- /dev/null +++ b/include/xen/tmem.h @@ -0,0 +1,5 @@ +#ifndef _XEN_TMEM_H +#define _XEN_TMEM_H +/* defined in drivers/xen/tmem.c */ +extern int tmem_enabled; +#endif /* _XEN_TMEM_H */