powerpc/pseries: Add collaborative memory manager
authorBrian King <brking@linux.vnet.ibm.com>
Wed, 23 Jul 2008 18:30:29 +0000 (04:30 +1000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Fri, 25 Jul 2008 05:44:42 +0000 (15:44 +1000)
Adds a collaborative memory manager, which acts as a simple balloon driver
for System p machines that support cooperative memory overcommitment
(CMO).

Adds a platform configuration option for CMO called PPC_SMLPAR.

Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Robert Jennings <rcj@linux.vnet.ibm.com>
Acked-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/platforms/pseries/Kconfig
arch/powerpc/platforms/pseries/Makefile
arch/powerpc/platforms/pseries/cmm.c [new file with mode: 0644]

index 757c0296e0b83e7456d28f5402f5db3c5918d4f4..97619fd51e393870d1c0c1da8a6bdb1926a52a57 100644 (file)
@@ -40,3 +40,26 @@ config PPC_PSERIES_DEBUG
        depends on PPC_PSERIES && PPC_EARLY_DEBUG
        bool "Enable extra debug logging in platforms/pseries"
        default y
+
+config PPC_SMLPAR
+       bool "Support for shared-memory logical partitions"
+       depends on PPC_PSERIES
+       select LPARCFG
+       default n
+       help
+         Select this option to enable shared memory partition support.
+         With this option a system running in an LPAR can be given more
+         memory than physically available and will allow firmware to
+         balance memory across many LPARs.
+
+config CMM
+       tristate "Collaborative memory management"
+       depends on PPC_SMLPAR
+       default y
+       help
+         Select this option, if you want to enable the kernel interface
+         to reduce the memory size of the system. This is accomplished
+         by allocating pages of memory and put them "on hold". This only
+         makes sense for a system running in an LPAR where the unused pages
+         will be reused for other LPARs. The interface allows firmware to
+         balance memory across many LPARs.
index 554c6e42ef2a9ff1f78340a0d5a9c49f5d46b611..dfe574af2dc090bd5ce3aaac503fef02e8bddb91 100644 (file)
@@ -24,3 +24,4 @@ obj-$(CONFIG_HVC_CONSOLE)     += hvconsole.o
 obj-$(CONFIG_HVCS)             += hvcserver.o
 obj-$(CONFIG_HCALL_STATS)      += hvCall_inst.o
 obj-$(CONFIG_PHYP_DUMP)        += phyp_dump.o
+obj-$(CONFIG_CMM)              += cmm.o
diff --git a/arch/powerpc/platforms/pseries/cmm.c b/arch/powerpc/platforms/pseries/cmm.c
new file mode 100644 (file)
index 0000000..c6b3be0
--- /dev/null
@@ -0,0 +1,468 @@
+/*
+ * Collaborative memory management interface.
+ *
+ * Copyright (C) 2008 IBM Corporation
+ * Author(s): Brian King (brking@linux.vnet.ibm.com),
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+#include <linux/ctype.h>
+#include <linux/delay.h>
+#include <linux/errno.h>
+#include <linux/fs.h>
+#include <linux/init.h>
+#include <linux/kthread.h>
+#include <linux/module.h>
+#include <linux/oom.h>
+#include <linux/sched.h>
+#include <linux/stringify.h>
+#include <linux/swap.h>
+#include <linux/sysdev.h>
+#include <asm/firmware.h>
+#include <asm/hvcall.h>
+#include <asm/mmu.h>
+#include <asm/pgalloc.h>
+#include <asm/uaccess.h>
+
+#include "plpar_wrappers.h"
+
+#define CMM_DRIVER_VERSION     "1.0.0"
+#define CMM_DEFAULT_DELAY      1
+#define CMM_DEBUG                      0
+#define CMM_DISABLE            0
+#define CMM_OOM_KB             1024
+#define CMM_MIN_MEM_MB         256
+#define KB2PAGES(_p)           ((_p)>>(PAGE_SHIFT-10))
+#define PAGES2KB(_p)           ((_p)<<(PAGE_SHIFT-10))
+
+static unsigned int delay = CMM_DEFAULT_DELAY;
+static unsigned int oom_kb = CMM_OOM_KB;
+static unsigned int cmm_debug = CMM_DEBUG;
+static unsigned int cmm_disabled = CMM_DISABLE;
+static unsigned long min_mem_mb = CMM_MIN_MEM_MB;
+static struct sys_device cmm_sysdev;
+
+MODULE_AUTHOR("Brian King <brking@linux.vnet.ibm.com>");
+MODULE_DESCRIPTION("IBM System p Collaborative Memory Manager");
+MODULE_LICENSE("GPL");
+MODULE_VERSION(CMM_DRIVER_VERSION);
+
+module_param_named(delay, delay, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(delay, "Delay (in seconds) between polls to query hypervisor paging requests. "
+                "[Default=" __stringify(CMM_DEFAULT_DELAY) "]");
+module_param_named(oom_kb, oom_kb, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(oom_kb, "Amount of memory in kb to free on OOM. "
+                "[Default=" __stringify(CMM_OOM_KB) "]");
+module_param_named(min_mem_mb, min_mem_mb, ulong, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(min_mem_mb, "Minimum amount of memory (in MB) to not balloon. "
+                "[Default=" __stringify(CMM_MIN_MEM_MB) "]");
+module_param_named(debug, cmm_debug, uint, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(debug, "Enable module debugging logging. Set to 1 to enable. "
+                "[Default=" __stringify(CMM_DEBUG) "]");
+
+#define CMM_NR_PAGES ((PAGE_SIZE - sizeof(void *) - sizeof(unsigned long)) / sizeof(unsigned long))
+
+#define cmm_dbg(...) if (cmm_debug) { printk(KERN_INFO "cmm: "__VA_ARGS__); }
+
+struct cmm_page_array {
+       struct cmm_page_array *next;
+       unsigned long index;
+       unsigned long page[CMM_NR_PAGES];
+};
+
+static unsigned long loaned_pages;
+static unsigned long loaned_pages_target;
+static unsigned long oom_freed_pages;
+
+static struct cmm_page_array *cmm_page_list;
+static DEFINE_SPINLOCK(cmm_lock);
+
+static struct task_struct *cmm_thread_ptr;
+
+/**
+ * cmm_alloc_pages - Allocate pages and mark them as loaned
+ * @nr:        number of pages to allocate
+ *
+ * Return value:
+ *     number of pages requested to be allocated which were not
+ **/
+static long cmm_alloc_pages(long nr)
+{
+       struct cmm_page_array *pa, *npa;
+       unsigned long addr;
+       long rc;
+
+       cmm_dbg("Begin request for %ld pages\n", nr);
+
+       while (nr) {
+               addr = __get_free_page(GFP_NOIO | __GFP_NOWARN |
+                                      __GFP_NORETRY | __GFP_NOMEMALLOC);
+               if (!addr)
+                       break;
+               spin_lock(&cmm_lock);
+               pa = cmm_page_list;
+               if (!pa || pa->index >= CMM_NR_PAGES) {
+                       /* Need a new page for the page list. */
+                       spin_unlock(&cmm_lock);
+                       npa = (struct cmm_page_array *)__get_free_page(GFP_NOIO | __GFP_NOWARN |
+                                                                      __GFP_NORETRY | __GFP_NOMEMALLOC);
+                       if (!npa) {
+                               pr_info("%s: Can not allocate new page list\n", __FUNCTION__);
+                               free_page(addr);
+                               break;
+                       }
+                       spin_lock(&cmm_lock);
+                       pa = cmm_page_list;
+
+                       if (!pa || pa->index >= CMM_NR_PAGES) {
+                               npa->next = pa;
+                               npa->index = 0;
+                               pa = npa;
+                               cmm_page_list = pa;
+                       } else
+                               free_page((unsigned long) npa);
+               }
+
+               if ((rc = plpar_page_set_loaned(__pa(addr)))) {
+                       pr_err("%s: Can not set page to loaned. rc=%ld\n", __FUNCTION__, rc);
+                       spin_unlock(&cmm_lock);
+                       free_page(addr);
+                       break;
+               }
+
+               pa->page[pa->index++] = addr;
+               loaned_pages++;
+               totalram_pages--;
+               spin_unlock(&cmm_lock);
+               nr--;
+       }
+
+       cmm_dbg("End request with %ld pages unfulfilled\n", nr);
+       return nr;
+}
+
+/**
+ * cmm_free_pages - Free pages and mark them as active
+ * @nr:        number of pages to free
+ *
+ * Return value:
+ *     number of pages requested to be freed which were not
+ **/
+static long cmm_free_pages(long nr)
+{
+       struct cmm_page_array *pa;
+       unsigned long addr;
+
+       cmm_dbg("Begin free of %ld pages.\n", nr);
+       spin_lock(&cmm_lock);
+       pa = cmm_page_list;
+       while (nr) {
+               if (!pa || pa->index <= 0)
+                       break;
+               addr = pa->page[--pa->index];
+
+               if (pa->index == 0) {
+                       pa = pa->next;
+                       free_page((unsigned long) cmm_page_list);
+                       cmm_page_list = pa;
+               }
+
+               plpar_page_set_active(__pa(addr));
+               free_page(addr);
+               loaned_pages--;
+               nr--;
+               totalram_pages++;
+       }
+       spin_unlock(&cmm_lock);
+       cmm_dbg("End request with %ld pages unfulfilled\n", nr);
+       return nr;
+}
+
+/**
+ * cmm_oom_notify - OOM notifier
+ * @self:      notifier block struct
+ * @dummy:     not used
+ * @parm:      returned - number of pages freed
+ *
+ * Return value:
+ *     NOTIFY_OK
+ **/
+static int cmm_oom_notify(struct notifier_block *self,
+                         unsigned long dummy, void *parm)
+{
+       unsigned long *freed = parm;
+       long nr = KB2PAGES(oom_kb);
+
+       cmm_dbg("OOM processing started\n");
+       nr = cmm_free_pages(nr);
+       loaned_pages_target = loaned_pages;
+       *freed += KB2PAGES(oom_kb) - nr;
+       oom_freed_pages += KB2PAGES(oom_kb) - nr;
+       cmm_dbg("OOM processing complete\n");
+       return NOTIFY_OK;
+}
+
+/**
+ * cmm_get_mpp - Read memory performance parameters
+ *
+ * Makes hcall to query the current page loan request from the hypervisor.
+ *
+ * Return value:
+ *     nothing
+ **/
+static void cmm_get_mpp(void)
+{
+       int rc;
+       struct hvcall_mpp_data mpp_data;
+       unsigned long active_pages_target;
+       signed long page_loan_request;
+
+       rc = h_get_mpp(&mpp_data);
+
+       if (rc != H_SUCCESS)
+               return;
+
+       page_loan_request = div_s64((s64)mpp_data.loan_request, PAGE_SIZE);
+       loaned_pages_target = page_loan_request + loaned_pages;
+       if (loaned_pages_target > oom_freed_pages)
+               loaned_pages_target -= oom_freed_pages;
+       else
+               loaned_pages_target = 0;
+
+       active_pages_target = totalram_pages + loaned_pages - loaned_pages_target;
+
+       if ((min_mem_mb * 1024 * 1024) > (active_pages_target * PAGE_SIZE))
+               loaned_pages_target = totalram_pages + loaned_pages -
+                       ((min_mem_mb * 1024 * 1024) / PAGE_SIZE);
+
+       cmm_dbg("delta = %ld, loaned = %lu, target = %lu, oom = %lu, totalram = %lu\n",
+               page_loan_request, loaned_pages, loaned_pages_target,
+               oom_freed_pages, totalram_pages);
+}
+
+static struct notifier_block cmm_oom_nb = {
+       .notifier_call = cmm_oom_notify
+};
+
+/**
+ * cmm_thread - CMM task thread
+ * @dummy:     not used
+ *
+ * Return value:
+ *     0
+ **/
+static int cmm_thread(void *dummy)
+{
+       unsigned long timeleft;
+
+       while (1) {
+               timeleft = msleep_interruptible(delay * 1000);
+
+               if (kthread_should_stop() || timeleft) {
+                       loaned_pages_target = loaned_pages;
+                       break;
+               }
+
+               cmm_get_mpp();
+
+               if (loaned_pages_target > loaned_pages) {
+                       if (cmm_alloc_pages(loaned_pages_target - loaned_pages))
+                               loaned_pages_target = loaned_pages;
+               } else if (loaned_pages_target < loaned_pages)
+                       cmm_free_pages(loaned_pages - loaned_pages_target);
+       }
+       return 0;
+}
+
+#define CMM_SHOW(name, format, args...)                        \
+       static ssize_t show_##name(struct sys_device *dev, char *buf)   \
+       {                                                       \
+               return sprintf(buf, format, ##args);            \
+       }                                                       \
+       static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
+
+CMM_SHOW(loaned_kb, "%lu\n", PAGES2KB(loaned_pages));
+CMM_SHOW(loaned_target_kb, "%lu\n", PAGES2KB(loaned_pages_target));
+
+static ssize_t show_oom_pages(struct sys_device *dev, char *buf)
+{
+       return sprintf(buf, "%lu\n", PAGES2KB(oom_freed_pages));
+}
+
+static ssize_t store_oom_pages(struct sys_device *dev,
+                              const char *buf, size_t count)
+{
+       unsigned long val = simple_strtoul (buf, NULL, 10);
+
+       if (!capable(CAP_SYS_ADMIN))
+               return -EPERM;
+       if (val != 0)
+               return -EBADMSG;
+
+       oom_freed_pages = 0;
+       return count;
+}
+
+static SYSDEV_ATTR(oom_freed_kb, S_IWUSR| S_IRUGO,
+                  show_oom_pages, store_oom_pages);
+
+static struct sysdev_attribute *cmm_attrs[] = {
+       &attr_loaned_kb,
+       &attr_loaned_target_kb,
+       &attr_oom_freed_kb,
+};
+
+static struct sysdev_class cmm_sysdev_class = {
+       .name = "cmm",
+};
+
+/**
+ * cmm_sysfs_register - Register with sysfs
+ *
+ * Return value:
+ *     0 on success / other on failure
+ **/
+static int cmm_sysfs_register(struct sys_device *sysdev)
+{
+       int i, rc;
+
+       if ((rc = sysdev_class_register(&cmm_sysdev_class)))
+               return rc;
+
+       sysdev->id = 0;
+       sysdev->cls = &cmm_sysdev_class;
+
+       if ((rc = sysdev_register(sysdev)))
+               goto class_unregister;
+
+       for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++) {
+               if ((rc = sysdev_create_file(sysdev, cmm_attrs[i])))
+                       goto fail;
+       }
+
+       return 0;
+
+fail:
+       while (--i >= 0)
+               sysdev_remove_file(sysdev, cmm_attrs[i]);
+       sysdev_unregister(sysdev);
+class_unregister:
+       sysdev_class_unregister(&cmm_sysdev_class);
+       return rc;
+}
+
+/**
+ * cmm_unregister_sysfs - Unregister from sysfs
+ *
+ **/
+static void cmm_unregister_sysfs(struct sys_device *sysdev)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(cmm_attrs); i++)
+               sysdev_remove_file(sysdev, cmm_attrs[i]);
+       sysdev_unregister(sysdev);
+       sysdev_class_unregister(&cmm_sysdev_class);
+}
+
+/**
+ * cmm_init - Module initialization
+ *
+ * Return value:
+ *     0 on success / other on failure
+ **/
+static int cmm_init(void)
+{
+       int rc = -ENOMEM;
+
+       if (!firmware_has_feature(FW_FEATURE_CMO))
+               return -EOPNOTSUPP;
+
+       if ((rc = register_oom_notifier(&cmm_oom_nb)) < 0)
+               return rc;
+
+       if ((rc = cmm_sysfs_register(&cmm_sysdev)))
+               goto out_oom_notifier;
+
+       if (cmm_disabled)
+               return rc;
+
+       cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
+       if (IS_ERR(cmm_thread_ptr)) {
+               rc = PTR_ERR(cmm_thread_ptr);
+               goto out_unregister_sysfs;
+       }
+
+       return rc;
+
+out_unregister_sysfs:
+       cmm_unregister_sysfs(&cmm_sysdev);
+out_oom_notifier:
+       unregister_oom_notifier(&cmm_oom_nb);
+       return rc;
+}
+
+/**
+ * cmm_exit - Module exit
+ *
+ * Return value:
+ *     nothing
+ **/
+static void cmm_exit(void)
+{
+       if (cmm_thread_ptr)
+               kthread_stop(cmm_thread_ptr);
+       unregister_oom_notifier(&cmm_oom_nb);
+       cmm_free_pages(loaned_pages);
+       cmm_unregister_sysfs(&cmm_sysdev);
+}
+
+/**
+ * cmm_set_disable - Disable/Enable CMM
+ *
+ * Return value:
+ *     0 on success / other on failure
+ **/
+static int cmm_set_disable(const char *val, struct kernel_param *kp)
+{
+       int disable = simple_strtoul(val, NULL, 10);
+
+       if (disable != 0 && disable != 1)
+               return -EINVAL;
+
+       if (disable && !cmm_disabled) {
+               if (cmm_thread_ptr)
+                       kthread_stop(cmm_thread_ptr);
+               cmm_thread_ptr = NULL;
+               cmm_free_pages(loaned_pages);
+       } else if (!disable && cmm_disabled) {
+               cmm_thread_ptr = kthread_run(cmm_thread, NULL, "cmmthread");
+               if (IS_ERR(cmm_thread_ptr))
+                       return PTR_ERR(cmm_thread_ptr);
+       }
+
+       cmm_disabled = disable;
+       return 0;
+}
+
+module_param_call(disable, cmm_set_disable, param_get_uint,
+                 &cmm_disabled, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(disable, "Disable CMM. Set to 1 to disable. "
+                "[Default=" __stringify(CMM_DISABLE) "]");
+
+module_init(cmm_init);
+module_exit(cmm_exit);