Re: [PATCH] Simplify /proc/<pid|self>/exe symlink code

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 12 Jul 2007 19:00:12 -0700 Matt Helsley <matthltc@us.ibm.com> wrote:

> This patch avoids holding the mmap semaphore while walking VMAs in response to
> programs which read or follow the /proc/<pid|self>/exe symlink. This also allows
> us to merge mmu and nommu proc_exe_link() functions. The costs are holding the
> task lock, a separate reference to the executable file stored in the task
> struct, and increased code in fork, exec, and exit paths.
> 
> Signed-off-by: Matt Helsley <matthltc@us.ibm.com>
> ---
> 
> Changelog:
> 
> Hold task_lock() while using task->exe_file. With this change I haven't
> 	been able to reproduce Chris Wright's Oops report:
> 		http://lkml.org/lkml/2007/5/31/34 
> 	I used a 4-way, x86 system running kernbench. I also tried a 4-way x86_64
> 	system running pidof. I used oprofile during all runs but I could not
> 	reproduce Chris' Oops with the new patch.
> 
> Compiled and passed simple tests for regressions when patched against a 2.6.20
> and a 2.6.22 kernel. Regression tests included a variety of file operations on
> /proc/<pid|self>/exe such as stat, lstat, open, close, readlink, and unlink. All
> produced the expected, baseline output results.
> 
> Andrew, please consider this patch for inclusion in -mm.

I wish we had a description of the bug which this fixes.  That email of
Chris's is referencing code which diddles with task_struct.exe_file, but
your patch _adds_ task_struct.exe_file, so I am all confused.

Your patch does lots of fput()s under task_lock(), but fput() can sleep.

Plus what Al said.