@@ -495,15 +495,25 @@ void file_kill(struct file *file)
int fops_read_lock(struct file *file)
{
int revoked = (file->f_mode & FMODE_REVOKED);
- if (likely(!revoked))
- atomic_long_inc(&file->f_use);
+ if (likely(!revoked)) {
+ if (likely(atomic_long_read(&file->f_count) == 1))
+ atomic_long_set(&file->f_use,
+ atomic_long_read(&file->f_use) + 1);
+ else
+ atomic_long_inc(&file->f_use);
+ }
return revoked;
}
void fops_read_unlock(struct file *file, int revoked)
{
- if (likely(!revoked))
- atomic_long_dec(&file->f_use);
+ if (likely(!revoked)) {
+ if (likely(atomic_long_read(&file->f_count) == 1))
+ atomic_long_set(&file->f_use,
+ atomic_long_read(&file->f_use) - 1);
+ else
+ atomic_long_dec(&file->f_use);
+ }
}
int fs_may_remount_ro(struct super_block *sb)
After seeing fget_light and the justification for it in the commit log I did not want to introduce something into the common read/write path of file descriptors that could have a significant measurable impact on I/O speed. commit f6435db01709533f270b2dce1e5914770dbc65de Author: akpm <akpm> Date: Thu May 8 05:19:50 2003 +0000 [PATCH] reduced overheads in fget/fput From: Dipankar Sarma <dipankar@in.ibm.com> fget() shows up on profiles, especially on SMP. Dipankar's patch special-cases the situation wherein there are no sharers of current->files. In this situation we know that no other process can close this file, so it is not necessary to increment the file's refcount. It's ugly as sin, but makes a substantial difference. The test is dd if=/dev/zero of=foo bs=1 count=1M On 4CPU P3 xeon with 1MB L2 cache and 512MB ram: kernel sys time std-dev ------------ -------- ------- UP - vanilla 2.104 0.028 UP - file 1.867 0.019 SMP - vanilla 2.976 0.023 SMP - file 2.719 0.026 BKrev: 3eb9e8f6Db0nMWoSx5IdHx6SBal8aw My test case was: dd if=/dev/zero of=/dev/null bs=1 count=1M. As writing to a real file turned out to cover the cost. Without the optimization I am seeing 2.4 - 2.5 MB/s on my idle core2 single socket quad core. With this optimization I am seeing 2.9 - 3.0 MB/s on the same machine. Maybe 2% slower than before I introduced fops_read_lock. The common case is that there is only one thread and so the fget_light optimization applies and f_count remains at 1. Which implies that there is only a single process performing operations through the file descriptor. In that case because there is no possible contention it is possible safely skip the atomic operations, gaining all of the benefits of rcu without requiring a per cpu variable. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> --- fs/file_table.c | 18 ++++++++++++++---- 1 files changed, 14 insertions(+), 4 deletions(-)