On 9/4/2018 6:52 PM, Alex Williamson wrote:
On Fri, 31 Aug 2018 16:03:38 -0700
Sinan Kaya <okaya@xxxxxxxxxx> wrote:
Hi Bjorn, Alex;
Since my recent series to relocate secondary bus reset code into the PCI core,
we hit a deadlock while trying to obtain a device lock during probe since
device lock is already held.
https://bugzilla.kernel.org/show_bug.cgi?id=200985
I posted a patch into the bugzilla so that we skip locks if we are probing
to follow the same strategy found in other lock routines in pci.c.
https://bugzilla.kernel.org/attachment.cgi?id=278221
I wanted to hear some opinions since this is a regression and need to find
a solution and also waiting for test feedback
Responded on the bugzilla, the patch seems to pass basic tests. I will
queue up some more but so far so good.
TBH, I'm having a hard time seeing the value in the original series.
Commit 811c5cb37df4 is fiddling with two drivers, hfi1 and vfio-pci.
In the case of hfi1 the comment before the call is specifically talking
about performing a secondary bus reset, so all they're looking for is
bouncing the link, it seems they don't even necessarily want a slot
reset, which might entail a full power cycle.
Correct. Keep it simple as possible is usually the best approach.
I won't deny that we're getting a little sloppy with some of the reset
interfaces and I'd love to solve the device lock issue more generally,
but trying to abstract slot vs bus and hide the lower level interfaces
from drivers doesn't seem like it's really working here.
The issue I have with all of this is we used to have a function that did
a very simple thing. Do an SBR, get out. Changing the API, I'm fine
with, but these changes did more than just change the API. They changed
the underlying behavior and what is happening now is different.
It sounds like all that is being done is put a band-aid on things and
there is some redesign work needed. Perhaps the best thing to do is back
this out for 4.19 and revisit with a more general solution for 4.20.
That's a call for you folks to make. My concern is getting our HW
functioning again with 4.19 and Sinan's patch along with my 1/1 seems to
allow that at least.
-Denny
