Re: [PATCH v3] proc,fcntl: introduce F_SET_DESCRIPTION

From: Pascal Bouchareine <kalou@tfz.net>
To: Alexey Dobriyan <adobriyan@gmail.com>
Cc: linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
	linux-api@vger.kernel.org, Al Viro <viro@zeniv.linux.org.uk>,
	Jeff Layton <jlayton@poochiereds.net>,
	"J. Bruce Fields" <bfields@fieldses.org>
Subject: Re: [PATCH v3] proc,fcntl: introduce F_SET_DESCRIPTION
Date: Mon, 27 Jul 2020 18:39:38 -0700	[thread overview]
Message-ID: <CAGbU3_nkBfrJK55-ZiCc6ve1F+CZBeuSmm6Dike0kkKja-RmVg@mail.gmail.com> (raw)
In-Reply-To: <20200727142140.GA116567@localhost.localdomain>

Thanks for reviewing, I added some questions inline below

On Mon, Jul 27, 2020 at 7:21 AM Alexey Dobriyan <adobriyan@gmail.com> wrote:
> > +     d = strndup_user(desc, MAX_FILE_DESC_SIZE);
>
> This should be kmem accounted because allocation is persistent.
> To make things more entertaining, strndup_user() doesn't have gfp_t argument.

I had to look it up so I might be very far from it, but is that
__GFP_ACCOUNT and would it be correct to assume memdup_user() should
use it unconditionally?

Otherwise my simple option would be to implement the logic in the
set_description, but the benefit would be very local.

Please let me know what you think is best, happy to read more doc if
there's a more productive way to address that

>
> > +     if (IS_ERR(d))
> > +             return PTR_ERR(d);
> > +
> > +     spin_lock(&file->f_lock);
> > +     kfree(file->f_description);
> > +     file->f_description = d;
> > +     spin_unlock(&file->f_lock);
>
> Generally kfree under spinlock is not good idea.
> You can replace the pointer and free without spinlock.

Sorry I also need some pointers here - do you suggest I move the kfree
out of the locked section or that there is a safe way other than
spinlock?

> struct file is nicely aligned to 256 bytes on distro configs.
> Will this break everything?
>
>         $ cat /sys/kernel/slab/filp/object_size

Indeed on the config I'm using here this jumped to 264 bytes

Would it be better to move this to the inode struct? I don't know the
implications of this - any other option?

Thanks!