Re: [PATCH 08/32] genirq: Synchronize only with single thread on free_irq()

[Lukas Wunner <lukas@wunner.de>]

On Mon, Jul 16, 2018 at 07:37:19AM -0500, Bjorn Helgaas wrote:
> On Fri, Jul 13, 2018 at 09:21:09AM +0200, Lukas Wunner wrote:
> > On Thu, Jul 12, 2018 at 05:21:09PM -0500, Bjorn Helgaas wrote:
> > > On Sat, Jun 16, 2018 at 09:25:00PM +0200, Lukas Wunner wrote:
> > > > When pciehp is converted to threaded IRQ handling, removal of unplugged
> > > > devices below a PCIe hotplug port happens synchronously in the IRQ
> > > > thread.  Removal of devices typically entails a call to free_irq() by
> > > > their drivers.
> > > > 
> > > > If those devices share their IRQ with the hotplug port, free_irq()
> > > > deadlocks because it calls synchronize_irq() to wait for all hard IRQ
> > > > handlers as well as all threads sharing the IRQ to finish.
> > > > 
> > > > Actually it's sufficient to wait only for the IRQ thread of the removed
> > > > device, so call synchronize_hardirq() to wait for all hard IRQ handlers
> > > > to finish, but no longer for any threads.  Compensate by rearranging the
> > > > control flow in irq_wait_for_interrupt() such that the device's thread
> > > > is allowed to run one last time after kthread_stop() has been called.
> > > 
> > > I assume this would need to be merged along with the rest of the
> > > series, which should probably go through the PCI tree, but I'm
> > > definitely not qualified to review this IRQ patch.  And it would need
> > > an ack from Thomas in any case.
> > 
> > A v2 of this patch has already been merged through the tip tree on June 24,
> > it's in linux-next as commit 519cc8652b3a, and ISTR that I marked this patch
> > either as "Obsoleted" or "Not Applicable" in pci-patchwork.  There was no
> > build-dependency of the succeeding patches in the series on this patch,
> > hence merging through a different tree was possible.
> 
> Great!  Do I need to make sure the tip tree is merged before the PCI
> tree during the merge window?

In theory the tip tree would have to be merged beforehand, but in practice
the odds are extremely low that someone with an affected machine (primarily
a Mac introduced 2011/2012) happens to unplug a PCI device during a bisect
while running a kernel which has the pci tree merged but not the tip tree.
So I wouldn't worry about it.

Thanks,

Lukas