lkml.org 
[lkml]   [2010]   [Oct]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: "do_IRQ: 0.89 No irq handler for vector (irq -1)"
On Tue, 12 Oct 2010 21:01:17 +0200
"Rafael J. Wysocki" <rjw@sisk.pl> wrote:

> On Tuesday, October 12, 2010, Jesse Barnes wrote:
> > On Mon, 11 Oct 2010 15:48:26 -0700
> > Jesse Barnes <jbarnes@virtuousgeek.org> wrote:
> >
> > > On Fri, 8 Oct 2010 21:46:50 +1000
> > > Dave Airlie <airlied@gmail.com> wrote:
> > > > Not sure how best to fix, I can workaround by calling
> > > > pci_set_power_state(PCI_D0) in the drm drivers, but I sorta thing the
> > > > PCI layer should take care of this.
> > >
> > > So I think we *should* be able to call pci_disable_device at remove
> > > time. But as you say, some platforms may not correctly re-route VGA
> > > space to an existing device or disable it properly when we do that.
> > > AFAICT x86 will be ok here though (seems to work ok locally too).
> >
> > Just tested this some more, and I think it's the right thing to do in
> > the KMS case at least. When we load a KMS driver it takes over the gfx
> > device and nothing can assume anything about VGA state unless using the
> > VGA arbiter. So calling pci_disable_device() in the shutdown path of a
> > KMS driver shouldn't make things any worse, and will work around this
> > issue.
> >
> > Doing so in the non-KMS case violates some PC assumptions though, in
> > that things like vgacon and the BIOS will assume VGA memory is still
> > around, which on some platforms pci_disable_device() may affect (I only
> > checked the x86 implementation).
> >
> > > That said, it seems like we should update the current device state at
> > > load time as well, once we've matched the driver it seems like there
> > > should be no harm.
> > >
> > > Rafael, what do you think? Would having the correct power state at
> > > load time cause any trouble with other PM code? I know we've had
> > > issues with setting it explicitly in the past...
> >
> > So we should probably make pci_enable_device pick up the current state
> > as well, instead of assuming it's unknown just because the enable count
> > was non-zero (which as Dave points out, can be affected by sysfs writes
> > too).
> >
> > The only downside I can think of there is that if the device is already
> > enabled, we generally have to assume another driver owns it, and who
> > knows if the device is actually alive enough to read the current state
> > from. But I think we handle those errors ok too, so pulling it out
> > should be safe.
>
> I remember trying to do something like this and it didn't play well with the
> initialization. Still, I didn't do that in pci_enable_device(), so I can't say
> for sure at the moment. I _think_ it will be fine, though.

Here's what I had in mind. I think it's safer than setting the power
state at enable time, and it works around the enable_cnt leak in the
DRM drivers.

--
Jesse Barnes, Intel Open Source Technology Center

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 7fa3cbd..37facc1 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -994,6 +994,18 @@ static int __pci_enable_device_flags(struct pci_dev *dev,
int err;
int i, bars = 0;

+ /*
+ * Power state could be unknown at this point, either due to a fresh
+ * boot or a device removal call. So get the current power state
+ * so that things like MSI message writing will behave as expected
+ * (e.g. if the device really is in D0 at enable time).
+ */
+ if (dev->pm_cap) {
+ u16 pmcsr;
+ pci_read_config_word(dev, dev->pm_cap + PCI_PM_CTRL, &pmcsr);
+ dev->current_state = (pmcsr & PCI_PM_CTRL_STATE_MASK);
+ }
+
if (atomic_add_return(1, &dev->enable_cnt) > 1)
return 0; /* already enabled */



\
 
 \ /
  Last update: 2010-10-13 23:19    [W:0.090 / U:0.152 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site