lkml.org 
[lkml]   [2006]   [Mar]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
Subject[RFC, PATCH] avoid some atomics in open()/close() for monothreaded processes
Goal : Avoid some locking/unlocking 'struct files_struct'->file_lock for mono 
threaded processes.

We define files_multithreaded() function .

static inline int files_multithreaded(const struct files_struct *files)
{
return sizeof(files->file_lock) > 0 && atomic_read(&files->count) > 1;
}

On plain UP kernel (not preemptable nor spinlock debug), this function is a
const 0, so that gcc can wipe out some code.

On preemptible or SMP, or spinlock debug kernels, the result is true only if
the ref count is greater than 1 (multi threaded process or /proc/{pid}/fd is
under investigation by another task)

This patch increases kernel size but pros are worth the cons, as said Linus
himself, we should increase performance of mono-threaded tasks....

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
--- a/include/linux/file.h 2006-03-22 06:23:02.000000000 +0100
+++ b/include/linux/file.h 2006-03-22 07:11:08.000000000 +0100
@@ -42,6 +42,11 @@
spinlock_t file_lock; /* Protects concurrent writers. Nests inside tsk->alloc_lock */
};

+static inline int files_multithreaded(const struct files_struct *files)
+{
+ return sizeof(files->file_lock) > 0 && atomic_read(&files->count) > 1;
+}
+
#define files_fdtable(files) (rcu_dereference((files)->fdt))

extern void FASTCALL(__fput(struct file *));
--- a/fs/open.c.orig 2006-03-22 06:24:34.000000000 +0100
+++ b/fs/open.c 2006-03-22 06:30:54.000000000 +0100
@@ -1050,11 +1050,17 @@
{
struct files_struct *files = current->files;
struct fdtable *fdt;
- spin_lock(&files->file_lock);
+ int fl_locked = 0;
+
+ if (files_multithreaded(files)) {
+ spin_lock(&files->file_lock);
+ fl_locked = 1;
+ }
fdt = files_fdtable(files);
BUG_ON(fdt->fd[fd] != NULL);
rcu_assign_pointer(fdt->fd[fd], file);
- spin_unlock(&files->file_lock);
+ if (fl_locked)
+ spin_unlock(&files->file_lock);
}

EXPORT_SYMBOL(fd_install);
@@ -1147,8 +1153,12 @@
struct file * filp;
struct files_struct *files = current->files;
struct fdtable *fdt;
+ int fl_locked = 0;

- spin_lock(&files->file_lock);
+ if (files_multithreaded(files)) {
+ spin_lock(&files->file_lock);
+ fl_locked = 1;
+ }
fdt = files_fdtable(files);
if (fd >= fdt->max_fds)
goto out_unlock;
@@ -1158,11 +1168,13 @@
rcu_assign_pointer(fdt->fd[fd], NULL);
FD_CLR(fd, fdt->close_on_exec);
__put_unused_fd(files, fd);
- spin_unlock(&files->file_lock);
+ if (fl_locked)
+ spin_unlock(&files->file_lock);
return filp_close(filp, files);

out_unlock:
- spin_unlock(&files->file_lock);
+ if (fl_locked)
+ spin_unlock(&files->file_lock);
return -EBADF;
}
\
 
 \ /
  Last update: 2006-03-22 07:26    [W:0.068 / U:0.364 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site