Re: [PATCH v5 6/7] module: Improve support for asynchronous module exit code

[Luis Chamberlain]

On Fri, Sep 30, 2022 at 12:39:27PM -0700, Luis Chamberlain wrote:
> On Wed, Sep 28, 2022 at 11:17:02AM -0700, Bart Van Assche wrote:
> > On 9/27/22 17:02, Luis Chamberlain wrote:
> > > On Tue, Sep 20, 2022 at 10:13:40AM -0700, Bart Van Assche wrote:
> > > > On 9/14/22 15:56, Bart Van Assche wrote:
> > > > > Some kernel modules call device_del() from their module exit code and
> > > > > schedule asynchronous work from inside the .release callback without waiting
> > > > > until that callback has finished. As an example, many SCSI LLD drivers call
> > > > > scsi_remove_host() from their module exit code. scsi_remove_host() may
> > > > > invoke scsi_device_dev_release_usercontext() asynchronously.
> > > > > scsi_device_dev_release_usercontext() uses the host template pointer and
> > > > > that pointer usually exists in static storage in the SCSI LLD. Support
> > > > > using the module reference count to keep the module around until
> > > > > asynchronous module exiting has completed by waiting in the delete_module()
> > > > > system call until the module reference count drops to zero.
> > > > 
> > > > Hi Luis,
> > > > 
> > > > I'd like to know your opinion about this patch since you are the maintainer
> > > > of the kernel module system.
> > > 
> > > See this patch which extends the documentation of try_module_get():
> > > 
> > > https://lkml.kernel.org/r/20211029184500.2821444-7-mcgrof@xxxxxxxxxx
> > > 
> > > You can ignore discussion around the thread as sadly it is just
> > > irrelevant stuff not about that patch. But the logic it spells out
> > > is still true.
> > > 
> > > So, in short, using try_module_get() on exit is actually the wrong
> > > thing to do and it is no surprise it would fail. I haven't gotten
> > > yet around to reviewing Mauro's driver API which let's you unbind
> > > drivers, but it sounds related so I CC'd you on that.
> > > 
> > > So I'd like to ask instead if an alternative to using try_module_get()
> > > on exit would be better here and for the future.
> > 
> > Hi Luis,
> > 
> > The extended documentation of try_module_get() is very helpful. But please
> > note that this patch is not related to try_module_get() at all. See also
> > patch 7/7 in this series (https://lore.kernel.org/linux-scsi/20220914225621.415631-8-bvanassche@xxxxxxx/).
> 
> I cannot see how this patch set is no way related to try_module_get()
> given the 7/7 patch you posted replaces try_module_get() with __module_get().
> My point, and hint, is that the original construct that added try_module_get()
> on removal was flawed and I'm not sure trying to expand on that idea would
> or even *should* be fruitful given the issues / tribal knowledge I tried
> extending documentation for.
> 
> It would beg the question if instead re-evaluating the goal could be
> done in such a way that the new documentation I suggested on try_module_get()
> would be seriously taken into account.

Yeah I've gone ahead and re-read your original patch again the issue
with that is it waits *once* for the refcnt to go to 0, but that does
not forbit it from going back up, at which point you have a race which
can still create the situation. Every subsystem is different, but for
instance simply running a loop opening a device block file should
suffice to bump the refcnt of a respective block driver module. So
at least the patch itself won't ultimately address this issue I'm
afraid.

  Luis