Re: [PATCH RFC] gen_init_cpio: Do fsync() only on regular files

[David Disseldorp]

On Mon, 6 Oct 2025 21:40:03 -0700, Christoph Hellwig wrote:

> On Tue, Oct 07, 2025 at 12:55:03AM +0100, Dmitry Safonov via B4 Relay wrote:
> > From: Dmitry Safonov <dima@xxxxxxxxxx>
> > 
> > Here at Arista gen_init_cpio is used in testing in order to create
> > an initramfs for specific tests. Most notably, there is a test that does
> > essentially a fork-bomb in kdump/panic kernel, replacing build-time
> > generated init script: instead of doing makedumpfile, it does call
> > shell tests.  
> 
> Why is is using fsync at all?  Seems like this was added in
> 
> commit ae18b94099b04264b32e33b057114024bc72c993
> Author: David Disseldorp <ddiss@xxxxxxx>
> Date:   Tue Aug 19 13:05:45 2025 +1000
> 
>     gen_init_cpio: support -o <output_file> parameter
> 
> without any good explanation.  In general doing a per-file fsync
> is going to horrible wreck performance, and given that no one is
> interested in partial initramfs archives also rather pointless.

I should have explained why in the commit, sorry. The intention was to
catch any FS I/O errors during output archive writeback. fsync() is
called only once as the final I/O.