Re: [PATCH 1/3] libertas_spi: Use workqueue in hw_host_to_card

[Dan Williams <dcbw@redhat.com>]

On Fri, 2011-01-21 at 23:24 +0200, Vasily Khoruzhick wrote:
> On Friday 21 January 2011 23:22:37 Marek Vasut wrote:
> > On Friday 21 January 2011 21:44:48 Vasily Khoruzhick wrote:
> > > Use workqueue to perform SPI xfers, it's necessary to fix
> > > nasty "BUG: scheduling while atomic", because
> > > spu_write() calls spi_sync() and spi_sync() may sleep, but
> > > hw_host_to_card() callback can be called from atomic context.
> > > Remove kthread completely, workqueue now does its job.
> > > Restore intermediate buffers which were removed in commit
> > > 86c34fe89e9cad9e1ba4d1a8bbf98259035f4caf that introduced
> > > mentioned bug.
> > 
> > I have two questions:
> > 
> > 1) Why not leave kthread there? ie. why switch to workqueue
> 
> Because it's not easy to ensure that kthread did its job in suspend handler,
> and to make if_spi.c look similar to if_sdio.c.
> 
> > 2) This should be split into two patches I guess -- a) revert the change b)
> > convert to workqueue -- so they can be (N)ACKed separatedly
> 
> Actually just reverting commit does not make driver work (it will fail on 
> rmmod), and it can impact on future bisect (if it'll be necessary). But if 
> it's requirement - ok.

I'll disagree with Marek; I don't think it's a hard requirement as long
as there's a good reason.  And if the driver doesn't work by reverting,
then you're right bisect is broken and that sucks.  I don't have a
problem with the "mega" patch (given that it's not really that large).

Dan