Sunil Mushran
2010-Dec-29 07:26 UTC
[Ocfs2-devel] [PATCH] ocfs2: Use hrtimer to track ocfs2 fs lock stats
Patch makes use of the hrtimer to track times in ocfs2 lock stats.
The patch is a bit involved to ensure no additional impact on the memory
footprint. The size of ocfs2_inode_cache remains 1280 bytes on 32-bit systems.
A related change was to modify the unit of the max wait time from nanosec to
microsec allowing us to track max time larger than 4 secs. This change
necessitated the bumping of the output version in the debugfs file,
locking_state, from 2 to 3.
Signed-off-by: Sunil Mushran <sunil.mushran at oracle.com>
---
fs/ocfs2/dlmglue.c | 97 +++++++++++++++++++++++++++-------------------------
fs/ocfs2/ocfs2.h | 23 +++++++-----
2 files changed, 64 insertions(+), 56 deletions(-)
diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index e8d94d7..7484bb9 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -64,7 +64,7 @@ struct ocfs2_mask_waiter {
unsigned long mw_mask;
unsigned long mw_goal;
#ifdef CONFIG_OCFS2_FS_STATS
- unsigned long long mw_lock_start;
+ ktime_t mw_lock_start;
#endif
};
@@ -435,44 +435,41 @@ static void ocfs2_remove_lockres_tracking(struct
ocfs2_lock_res *res)
#ifdef CONFIG_OCFS2_FS_STATS
static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
{
- res->l_lock_num_prmode = 0;
- res->l_lock_num_prmode_failed = 0;
- res->l_lock_total_prmode = 0;
- res->l_lock_max_prmode = 0;
- res->l_lock_num_exmode = 0;
- res->l_lock_num_exmode_failed = 0;
- res->l_lock_total_exmode = 0;
- res->l_lock_max_exmode = 0;
res->l_lock_refresh = 0;
+ memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
+ memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
}
static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
struct ocfs2_mask_waiter *mw, int ret)
{
- unsigned long long *num, *sum;
- unsigned int *max, *failed;
- struct timespec ts = current_kernel_time();
- unsigned long long time = timespec_to_ns(&ts) - mw->mw_lock_start;
-
- if (level == LKM_PRMODE) {
- num = &res->l_lock_num_prmode;
- sum = &res->l_lock_total_prmode;
- max = &res->l_lock_max_prmode;
- failed = &res->l_lock_num_prmode_failed;
- } else if (level == LKM_EXMODE) {
- num = &res->l_lock_num_exmode;
- sum = &res->l_lock_total_exmode;
- max = &res->l_lock_max_exmode;
- failed = &res->l_lock_num_exmode_failed;
- } else
+ u32 usec;
+ ktime_t kt;
+ struct ocfs2_lock_stats *stats;
+
+ if (level == LKM_PRMODE)
+ stats = &res->l_lock_prmode;
+ else if (level == LKM_EXMODE)
+ stats = &res->l_lock_exmode;
+ else
return;
- (*num)++;
- (*sum) += time;
- if (time > *max)
- *max = time;
+ kt = ktime_sub(ktime_get(), mw->mw_lock_start);
+ usec = ktime_to_us(kt);
+
+ stats->ls_gets++;
+ stats->ls_total += ktime_to_ns(kt);
+ /* overflow */
+ if (unlikely(stats->ls_gets) == 0) {
+ stats->ls_gets++;
+ stats->ls_total = ktime_to_ns(kt);
+ }
+
+ if (stats->ls_max < usec)
+ stats->ls_max = usec;
+
if (ret)
- (*failed)++;
+ stats->ls_fail++;
}
static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
@@ -482,8 +479,7 @@ static inline void ocfs2_track_lock_refresh(struct
ocfs2_lock_res *lockres)
static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
{
- struct timespec ts = current_kernel_time();
- mw->mw_lock_start = timespec_to_ns(&ts);
+ mw->mw_lock_start = ktime_get();
}
#else
static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
@@ -2869,8 +2865,15 @@ static void *ocfs2_dlm_seq_next(struct seq_file *m, void
*v, loff_t *pos)
return iter;
}
-/* So that debugfs.ocfs2 can determine which format is being used */
-#define OCFS2_DLM_DEBUG_STR_VERSION 2
+/*
+ * Version is used by debugfs.ocfs2 to determine the format being used
+ *
+ * New in version 2
+ * - Lock stats printed
+ * New in version 3
+ * - Max time in lock stats is in usecs (instead of nsecs)
+ */
+#define OCFS2_DLM_DEBUG_STR_VERSION 3
static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
{
int i;
@@ -2912,18 +2915,18 @@ static int ocfs2_dlm_seq_show(struct seq_file *m, void
*v)
seq_printf(m, "0x%x\t", lvb[i]);
#ifdef CONFIG_OCFS2_FS_STATS
-# define lock_num_prmode(_l) (_l)->l_lock_num_prmode
-# define lock_num_exmode(_l) (_l)->l_lock_num_exmode
-# define lock_num_prmode_failed(_l) (_l)->l_lock_num_prmode_failed
-# define lock_num_exmode_failed(_l) (_l)->l_lock_num_exmode_failed
-# define lock_total_prmode(_l) (_l)->l_lock_total_prmode
-# define lock_total_exmode(_l) (_l)->l_lock_total_exmode
-# define lock_max_prmode(_l) (_l)->l_lock_max_prmode
-# define lock_max_exmode(_l) (_l)->l_lock_max_exmode
-# define lock_refresh(_l) (_l)->l_lock_refresh
+# define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets)
+# define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets)
+# define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail)
+# define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail)
+# define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total)
+# define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total)
+# define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max)
+# define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max)
+# define lock_refresh(_l) ((_l)->l_lock_refresh)
#else
-# define lock_num_prmode(_l) (0ULL)
-# define lock_num_exmode(_l) (0ULL)
+# define lock_num_prmode(_l) (0)
+# define lock_num_exmode(_l) (0)
# define lock_num_prmode_failed(_l) (0)
# define lock_num_exmode_failed(_l) (0)
# define lock_total_prmode(_l) (0ULL)
@@ -2933,8 +2936,8 @@ static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
# define lock_refresh(_l) (0)
#endif
/* The following seq_print was added in version 2 of this output */
- seq_printf(m, "%llu\t"
- "%llu\t"
+ seq_printf(m, "%u\t"
+ "%u\t"
"%u\t"
"%u\t"
"%llu\t"
diff --git a/fs/ocfs2/ocfs2.h b/fs/ocfs2/ocfs2.h
index 70dd3b1..0b04cfd 100644
--- a/fs/ocfs2/ocfs2.h
+++ b/fs/ocfs2/ocfs2.h
@@ -147,6 +147,17 @@ struct ocfs2_lock_res_ops;
typedef void (*ocfs2_lock_callback)(int status, unsigned long data);
+#ifdef CONFIG_OCFS2_FS_STATS
+struct ocfs2_lock_stats {
+ u64 ls_total; /* Total wait in NSEC */
+ u32 ls_gets; /* Num acquires */
+ u32 ls_fail; /* Num failed acquires */
+
+ /* Storing max wait in usecs saves 24 bytes per inode */
+ u32 ls_max; /* Max wait in USEC */
+};
+#endif
+
struct ocfs2_lock_res {
void *l_priv;
struct ocfs2_lock_res_ops *l_ops;
@@ -182,15 +193,9 @@ struct ocfs2_lock_res {
struct list_head l_debug_list;
#ifdef CONFIG_OCFS2_FS_STATS
- unsigned long long l_lock_num_prmode; /* PR acquires */
- unsigned long long l_lock_num_exmode; /* EX acquires */
- unsigned int l_lock_num_prmode_failed; /* Failed PR gets */
- unsigned int l_lock_num_exmode_failed; /* Failed EX gets */
- unsigned long long l_lock_total_prmode; /* Tot wait for PR */
- unsigned long long l_lock_total_exmode; /* Tot wait for EX */
- unsigned int l_lock_max_prmode; /* Max wait for PR */
- unsigned int l_lock_max_exmode; /* Max wait for EX */
- unsigned int l_lock_refresh; /* Disk refreshes */
+ struct ocfs2_lock_stats l_lock_prmode; /* PR mode stats */
+ u32 l_lock_refresh; /* Disk refreshes */
+ struct ocfs2_lock_stats l_lock_exmode; /* EX mode stats */
#endif
#ifdef CONFIG_DEBUG_LOCK_ALLOC
struct lockdep_map l_lockdep_map;
--
1.7.1
Joel Becker
2011-Feb-20 12:07 UTC
[Ocfs2-devel] [PATCH] ocfs2: Use hrtimer to track ocfs2 fs lock stats
On Tue, Dec 28, 2010 at 11:26:03PM -0800, Sunil Mushran wrote:> Patch makes use of the hrtimer to track times in ocfs2 lock stats. > > The patch is a bit involved to ensure no additional impact on the memory > footprint. The size of ocfs2_inode_cache remains 1280 bytes on 32-bit systems. > > A related change was to modify the unit of the max wait time from nanosec to > microsec allowing us to track max time larger than 4 secs. This change > necessitated the bumping of the output version in the debugfs file, > locking_state, from 2 to 3. > > Signed-off-by: Sunil Mushran <sunil.mushran at oracle.com>This patch is now in the merge-window branch of ocfs2.git. Joel -- Life's Little Instruction Book #198 "Feed a stranger's expired parking meter." http://www.jlbec.org/ jlbec at evilplan.org