Whamcloud - gitweb
LU-2139 osc: Track and limit "unstable" pages
[fs/lustre-release.git] / lustre / obdclass / class_obd.c
index dd981e9..cb0ea79 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, Whamcloud, Inc.
+ * Copyright (c) 2011, 2012, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 #include <lprocfs_status.h>
 #include <lustre/lustre_build_version.h>
 #include <libcfs/list.h>
+#include <cl_object.h>
+#ifdef HAVE_SERVER_SUPPORT
+# include <dt_object.h>
+# include <md_object.h>
+#endif /* HAVE_SERVER_SUPPORT */
 #include "llog_internal.h"
 
 #ifndef __KERNEL__
@@ -60,14 +65,17 @@ EXPORT_SYMBOL(obd_devs);
 cfs_list_t obd_types;
 DEFINE_RWLOCK(obd_dev_lock);
 
-#ifndef __KERNEL__
 __u64 obd_max_pages = 0;
 __u64 obd_max_alloc = 0;
+#ifndef __KERNEL__
 __u64 obd_alloc;
 __u64 obd_pages;
 #endif
+DEFINE_SPINLOCK(obd_updatemax_lock);
 
 /* The following are visible and mutable through /proc/sys/lustre/. */
+unsigned int obd_alloc_fail_rate = 0;
+EXPORT_SYMBOL(obd_alloc_fail_rate);
 unsigned int obd_debug_peer_on_timeout;
 EXPORT_SYMBOL(obd_debug_peer_on_timeout);
 unsigned int obd_dump_on_timeout;
@@ -76,6 +84,8 @@ unsigned int obd_dump_on_eviction;
 EXPORT_SYMBOL(obd_dump_on_eviction);
 unsigned int obd_max_dirty_pages = 256;
 EXPORT_SYMBOL(obd_max_dirty_pages);
+cfs_atomic_t obd_unstable_pages;
+EXPORT_SYMBOL(obd_unstable_pages);
 cfs_atomic_t obd_dirty_pages;
 EXPORT_SYMBOL(obd_dirty_pages);
 unsigned int obd_timeout = OBD_TIMEOUT_DEFAULT;   /* seconds */
@@ -104,6 +114,11 @@ EXPORT_SYMBOL(obd_dirty_transit_pages);
 char obd_jobid_var[JOBSTATS_JOBID_VAR_MAX_LEN + 1] = JOBSTATS_DISABLE;
 EXPORT_SYMBOL(obd_jobid_var);
 
+#ifdef LPROCFS
+struct lprocfs_stats *obd_memory = NULL;
+EXPORT_SYMBOL(obd_memory);
+#endif
+
 /* Get jobid of current process by reading the environment variable
  * stored in between the "env_start" & "env_end" of task struct.
  *
@@ -129,8 +144,8 @@ int lustre_get_jobid(char *jobid)
 
        /* Use process name + fsuid as jobid */
        if (strcmp(obd_jobid_var, JOBSTATS_PROCNAME_UID) == 0) {
-               snprintf(jobid, JOBSTATS_JOBID_SIZE, "%s_%u",
-                        cfs_curproc_comm(), cfs_curproc_fsuid());
+               snprintf(jobid, JOBSTATS_JOBID_SIZE, "%s.%u",
+                        current_comm(), current_fsuid());
                RETURN(0);
        }
 
@@ -162,6 +177,27 @@ int lustre_get_jobid(char *jobid)
 }
 EXPORT_SYMBOL(lustre_get_jobid);
 
+int obd_alloc_fail(const void *ptr, const char *name, const char *type,
+                  size_t size, const char *file, int line)
+{
+       if (ptr == NULL ||
+           (cfs_rand() & OBD_ALLOC_FAIL_MASK) < obd_alloc_fail_rate) {
+               CERROR("%s%salloc of %s ("LPU64" bytes) failed at %s:%d\n",
+                      ptr ? "force " :"", type, name, (__u64)size, file,
+                      line);
+               CERROR(LPU64" total bytes and "LPU64" total pages "
+                      "("LPU64" bytes) allocated by Lustre, "
+                      "%d total bytes by LNET\n",
+                      obd_memory_sum(),
+                      obd_pages_sum() << PAGE_CACHE_SHIFT,
+                      obd_pages_sum(),
+                       cfs_atomic_read(&libcfs_kmemory));
+               return 1;
+       }
+       return 0;
+}
+EXPORT_SYMBOL(obd_alloc_fail);
+
 static inline void obd_data2conn(struct lustre_handle *conn,
                                  struct obd_ioctl_data *data)
 {
@@ -239,7 +275,7 @@ int class_handle_ioctl(unsigned int cmd, unsigned long arg)
                 OBD_ALLOC(lcfg, data->ioc_plen1);
                 if (lcfg == NULL)
                         GOTO(out, err = -ENOMEM);
-                err = cfs_copy_from_user(lcfg, data->ioc_pbuf1,
+               err = copy_from_user(lcfg, data->ioc_pbuf1,
                                          data->ioc_plen1);
                 if (!err)
                         err = lustre_cfg_sanity_check(lcfg, data->ioc_plen1);
@@ -419,9 +455,9 @@ int class_handle_ioctl(unsigned int cmd, unsigned long arg)
 } /* class_handle_ioctl */
 
 #ifdef __KERNEL__
-extern cfs_psdev_t obd_psdev;
+extern struct miscdevice obd_psdev;
 #else
-void *obd_psdev = NULL;
+struct miscdevice obd_psdev;
 #endif
 
 #define OBD_INIT_CHECK
@@ -485,9 +521,9 @@ int obd_init_checks(void)
                 CWARN("LPD64 wrong length! strlen(%s)=%d != 2\n", buf, len);
                 ret = -EINVAL;
         }
-        if ((u64val & ~CFS_PAGE_MASK) >= CFS_PAGE_SIZE) {
+       if ((u64val & ~CFS_PAGE_MASK) >= PAGE_CACHE_SIZE) {
                 CWARN("mask failed: u64val "LPU64" >= "LPU64"\n", u64val,
-                      (__u64)CFS_PAGE_SIZE);
+                     (__u64)PAGE_CACHE_SIZE);
                 ret = -EINVAL;
         }
 
@@ -497,7 +533,7 @@ int obd_init_checks(void)
 #define obd_init_checks() do {} while(0)
 #endif
 
-extern cfs_spinlock_t obd_types_lock;
+extern spinlock_t obd_types_lock;
 extern int class_procfs_init(void);
 extern int class_procfs_clean(void);
 
@@ -517,7 +553,7 @@ int init_obdclass(void)
 
         LCONSOLE_INFO("Lustre: Build Version: "BUILD_VERSION"\n");
 
-        cfs_spin_lock_init(&obd_types_lock);
+       spin_lock_init(&obd_types_lock);
         obd_zombie_impexp_init();
 #ifdef LPROCFS
         obd_memory = lprocfs_alloc_stats(OBD_STATS_NUM,
@@ -546,11 +582,11 @@ int init_obdclass(void)
 
         CFS_INIT_LIST_HEAD(&obd_types);
 
-        err = cfs_psdev_register(&obd_psdev);
-        if (err) {
-                CERROR("cannot register %d err %d\n", OBD_DEV_MINOR, err);
-                return err;
-        }
+       err = misc_register(&obd_psdev);
+       if (err) {
+               CERROR("cannot register %d err %d\n", OBD_DEV_MINOR, err);
+               return err;
+       }
 
         /* This struct is already zeroed for us (static global) */
         for (i = 0; i < class_devno_max(); i++)
@@ -559,10 +595,10 @@ int init_obdclass(void)
         /* Default the dirty page cache cap to 1/2 of system memory.
          * For clients with less memory, a larger fraction is needed
          * for other purposes (mostly for BGL). */
-        if (cfs_num_physpages <= 512 << (20 - CFS_PAGE_SHIFT))
-                obd_max_dirty_pages = cfs_num_physpages / 4;
-        else
-                obd_max_dirty_pages = cfs_num_physpages / 2;
+       if (totalram_pages <= 512 << (20 - PAGE_CACHE_SHIFT))
+               obd_max_dirty_pages = totalram_pages / 4;
+       else
+               obd_max_dirty_pages = totalram_pages / 2;
 
         err = obd_init_caches();
         if (err)
@@ -573,9 +609,27 @@ int init_obdclass(void)
                 return err;
 #endif
 
-        err = lu_global_init();
-        if (err)
-                return err;
+       err = lu_global_init();
+       if (err)
+               return err;
+
+       err = lu_capainfo_init();
+       if (err)
+               return err;
+
+       err = cl_global_init();
+       if (err != 0)
+               return err;
+
+#if defined(__KERNEL__) && defined(HAVE_SERVER_SUPPORT)
+       err = dt_global_init();
+       if (err != 0)
+               return err;
+
+       err = lu_ucred_global_init();
+       if (err != 0)
+               return err;
+#endif
 
        err = llog_info_init();
        if (err)
@@ -588,6 +642,48 @@ int init_obdclass(void)
         return err;
 }
 
+void obd_update_maxusage(void)
+{
+       __u64 max1, max2;
+
+       max1 = obd_pages_sum();
+       max2 = obd_memory_sum();
+
+       spin_lock(&obd_updatemax_lock);
+       if (max1 > obd_max_pages)
+               obd_max_pages = max1;
+       if (max2 > obd_max_alloc)
+               obd_max_alloc = max2;
+       spin_unlock(&obd_updatemax_lock);
+}
+EXPORT_SYMBOL(obd_update_maxusage);
+
+#ifdef LPROCFS
+__u64 obd_memory_max(void)
+{
+       __u64 ret;
+
+       spin_lock(&obd_updatemax_lock);
+       ret = obd_max_alloc;
+       spin_unlock(&obd_updatemax_lock);
+
+       return ret;
+}
+EXPORT_SYMBOL(obd_memory_max);
+
+__u64 obd_pages_max(void)
+{
+       __u64 ret;
+
+       spin_lock(&obd_updatemax_lock);
+       ret = obd_max_pages;
+       spin_unlock(&obd_updatemax_lock);
+
+       return ret;
+}
+EXPORT_SYMBOL(obd_pages_max);
+#endif
+
 /* liblustre doesn't call cleanup_obdclass, apparently.  we carry on in this
  * ifdef to the end of the file to cover module and versioning goo.*/
 #ifdef __KERNEL__
@@ -601,18 +697,24 @@ static void cleanup_obdclass(void)
 
         lustre_unregister_fs();
 
-        cfs_psdev_deregister(&obd_psdev);
-        for (i = 0; i < class_devno_max(); i++) {
-                struct obd_device *obd = class_num2obd(i);
-                if (obd && obd->obd_set_up &&
-                    OBT(obd) && OBP(obd, detach)) {
-                        /* XXX should this call generic detach otherwise? */
-                        LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
-                        OBP(obd, detach)(obd);
-                }
-        }
+       misc_deregister(&obd_psdev);
+       for (i = 0; i < class_devno_max(); i++) {
+               struct obd_device *obd = class_num2obd(i);
+               if (obd && obd->obd_set_up &&
+                   OBT(obd) && OBP(obd, detach)) {
+                       /* XXX should this call generic detach otherwise? */
+                       LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
+                       OBP(obd, detach)(obd);
+               }
+       }
        llog_info_fini();
-        lu_global_fini();
+#ifdef HAVE_SERVER_SUPPORT
+       lu_ucred_global_fini();
+       dt_global_fini();
+#endif
+       cl_global_fini();
+       lu_capainfo_fini();
+       lu_global_fini();
 
         obd_cleanup_caches();
         obd_sysctl_clean();