From mboxrd@z Thu Jan  1 00:00:00 1970
From: arnd@arndb.de (Arnd Bergmann)
Date: Tue, 10 Nov 2015 10:45:04 +0100
Subject: [PATCH] PCI: imx6:don't sleep in atomic context
In-Reply-To: <1447148110.3142.6.camel@pengutronix.de>
References: <1447066080-5859-1-git-send-email-sanjeev_sharma@mentor.com>
 <4800389.dqcmlJY3HC@wuerfel> <1447148110.3142.6.camel@pengutronix.de>
Message-ID: <6415545.xzR4A6SaUL@wuerfel>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Tuesday 10 November 2015 10:35:10 Lucas Stach wrote:
> Am Dienstag, den 10.11.2015, 10:28 +0100 schrieb Arnd Bergmann:
> > On Tuesday 10 November 2015 09:41:18 Lucas Stach wrote:
> > > > diff --git a/drivers/pci/host/pci-imx6.c b/drivers/pci/host/pci-imx6.c
> > > > index 233a196..9769b13 100644
> > > > --- a/drivers/pci/host/pci-imx6.c
> > > > +++ b/drivers/pci/host/pci-imx6.c
> > > > @@ -499,7 +499,7 @@ static int imx6_pcie_link_up(struct pcie_port *pp)
> > > >                * Wait a little bit, then re-check if the link finished
> > > >                * the training.
> > > >                */
> > > > -             usleep_range(1000, 2000);
> > > > +             mdelay(1000);
> > > 
> > > A mdelay(1000) is a whole different timescale than a usleep(1000). If
> > > this patch works for you with mdelay(1) or maybe mdelay(2) I would be
> > > fine with it.
> > 
> > mdelay(1) is still a really long time to block the CPU for, on potentially
> > every config space access.
> > 
> > Everybody else just returns the link status here, which seems to be
> > the better alternative. If you need to delay the startup, better have
> > a msleep(1) loop in the initial probe function where you are allowed to
> > sleep.
> > 
> Yes, it's somewhere on my TODO list to rework the link-up handling here,
> but as there are quite a few timing and ordering implications in that
> code, this needs a good thought and a good deal of testing. So I'm
> inclined to ACK the current patch to get rid of the obvious bug and sort
> things out properly in a follow on patchset.

Maybe use that patch with some modifications then:

* add a comment to explain that this is currently called from possibly
  atomic context through pci_config_{read,write} and that the link
  state handling never belonged here.

* instead of looping five times for up to 2ms each, loop 100 times
  around a udelay(20) to hopefully be done earlier. I was going to
  suggest using time_before(timeout, jiffies) as the condition to
  wait for, but that doesn't work if called with interrupts disabled.

	Arnd