Re: [PATCH 1/1] fsopen: fsconfig syscall restart fix

[Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>]

On Wed, 23 Sep 2020 18:03:22 +0100
Al Viro <viro@zeniv.linux.org.uk> wrote:

> On Wed, Sep 23, 2020 at 07:46:36PM +0300, Alexander Mikhalitsyn wrote:
> > During execution of vfs_fsconfig_locked function we can get ERESTARTNOINTR
> > error (or other interrupt error). But we changing fs context fc->phase
> > field to transient states and our entry fc->phase checks in switch cases
> > (see FS_CONTEXT_CREATE_PARAMS, FS_CONTEXT_RECONF_PARAMS) will always fail
> > after syscall restart which will lead to returning -EBUSY to the userspace.
> > 
> > The idea of the fix is to save entry-time fs_context phase field value and
> > recover fc->phase value to the original one before exiting with
> > "interrupt error" (ERESTARTNOINTR or similar).
> 
> If you have e.g. vfs_create_tree() fail in the middle of ->get_tree(),
> the only thing you can do to that thing is to discard it.  The state is
> *NOT* required to be recoverable after a failure exit - quite a bit of
> config might've been consumed and freed by that point.
> 
> CREATE and RECONFIGURE are simply not restartable.

Thank you for quick response!

I got you idea. But as far as I understand fsopen/fsconfig API is in
early-development stage and we can think about convenience here.

Consider the typical code here:
int fsfd;
fsfd = fsopen("somefs", 0);
// a lot of:
fsconfig(fsfd, FSCONFIG_SET_FLAG, ...);
fsconfig(fsfd, FSCONFIG_SET_STRING, ...);
fsconfig(fsfd, FSCONFIG_SET_BINARY, ...);
//...

// now call:
fsconfig(sfd, FSCONFIG_CMD_CREATE, NULL, NULL, 0)
-> get signal here or something else
-> syscall restarted but this doesn't work because
of broken fc->phase state
-> get EBUSY
-> now we need to repeat *all* steps with
fsconfig(fsfd, FSCONFIG_SET_FLAG/FSCONFIG_SET_STRING, ...).
Speaking honestly, this looks weird.

Regards,
Alex.