Re: [PATCH for-rc v2 3/3] IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type

[Alex Williamson <alex.williamson@redhat.com>]

On Tue, 11 Sep 2018 20:59:27 -0500
Bjorn Helgaas <helgaas@kernel.org> wrote:

> Thanks for the quick response (and sorry for my late one).
> 
> On Tue, Sep 11, 2018 at 09:23:31PM -0400, Sinan Kaya wrote:
> > On 9/11/2018 9:09 PM, Bjorn Helgaas wrote:  
> > > On Wed, Sep 05, 2018 at 04:08:05PM +0000, Sinan Kaya wrote:  
> > > > pci_reset_bus() is being called from the probe context and causes
> > > > a deadlock since pci_reset_bus() also tries to obtain kernel lock.  
> > > 
> > > This doesn't tell me what the deadlock is.  
> > 
> > Sorry, let me try to set the context.
> > 
> > Driver probe routine is being called with pci_dev_lock() held.
> > 
> > pci_reset_bus() tries to obtain the pci_dev_lock() via this path one
> > more time while holding the lock.
> > 
> > pci_dev_lock()
> > 	enter driver probe()
> > 	pci_reset_bus()
> > 		 __pci_reset_bus()
> > 			pci_bus_save_and_disable()
> > 				pci_dev_lock()
> > 
> > 	exit driver probe()
> > pci_dev_unlock()
> > 
> >   
> > >   
> > > > Use __pci_reset_function_locked() that provides locking guarantees.  
> > > 
> > > And the previous patch that adds the "reset_type" parameter doesn't
> > > tell me anything about what locking guarantees it provides.  
> > 
> > This is sort of implied.  
> 
> Uhh, maybe, but the implication is too subtle for me.  It adds a
> "reset_type" that can be FLR, PM, SLOT, BUS, etc.  Those are different
> types of resets, but those types say nothing about locking.
> 
> > As the name of the function is __pci_reset_function_locked(), it is
> > assumed that the caller is responsible for obtaining the necessary
> > locks before calling this function.
> >   
> > > 
> > > I want to merge minimal changes for v4.19 to fix the known problems.
> > > 
> > > It's not clear to me what actually broke hfi1.  You mention a deadlock
> > > above, so I assume some locking change broke it, but 811c5cb37df4 isn't
> > > obviously related to locking.  
> > 
> > Previous code was calling secondary bus reset function directly. With the  
> 
> Ah, so it looks like 409888e0966e ("IB/hfi1: Use pci_try_reset_bus()
> for initiating PCI Secondary Bus Reset") is an essential part of the
> picture here.  That should probably be part of one of these
> changelogs.
> 
> 409888e0966e is where the locking change happened.  trigger_sbr()
> previously called pci_reset_bridge_secondary_bus(), but after
> 409888e0966e, it calls pci_try_reset_bus(), which at the time called
> pci_bus_trylock() *before* calling pci_reset_bridge_secondary_bus().
> 
> Can we just move the pci_bridge_secondary_bus_reset() decl back to
> include/linux/pci.h temporarily and have trigger_sbr() call that
> again?  I'd rather do that than try to squeeze some API rework into an
> -rc.
> 
> > generalization of the code, we are going through device save and restore
> > as follows:
> > 
> > 1. obtain device lock
> > 2. save device state
> > 3. release device lock
> > 4. perform secondary bus reset
> > 5. obtain device lock
> > 6. restore device state
> > 7. release device lock
> > 
> > We have introduced the locks in step #1, #3, #5 and #7 into the existing
> > code path.
> >   
> > > 
> > > It's obvious that 811c5cb37df4 tested the return value of
> > > pci_probe_reset_slot() incorrectly, so there's no problem with patch 1/1.
> > >   
> > 
> > yup.
> >   
> > > But patches 2 & 3:
> > > 
> > >    PCI: Request reset type as part of function reset
> > >    IB/hfi1: Prefer new __pci_reset_function_locked() API with reset type
> > > 
> > > do not connect the dots for me.  
> > 
> > I hope the above explanation helps. The consensus after conversation with Alex
> > was to reuse the locked API but add masks so that requester can selectively
> > request a slot reset without any locks.  
> 
> I'm not buying this yet.  There's nothing in that API that says to me
> "requesting certain types of resets doesn't acquire locks".

The type of reset doesn't determine the locking, it's that the
_locked() version of pci_reset_function() requires the caller to
provide the locking and the "__" prefix version requires the caller to
handle the device save and restore.  So the hfi1 requirement was nearly
satisfied by the existing exported __pci_reset_function_locked(), but
they want a link level reset, not an FLR, but the hotplug or sbr resets
are the last options of all the pci_reset_function() choices
currently.  Therefore adding a reset type as a parameter fixes both the
locking issue and the device save/restore, which the hfi1 driver
duplicates, and still makes the hfi1 driver work better with hotplug
slots, where their previous sbr might trigger surprise hotplugs.

Additionally, the original series unified pci_reset_bus() such that the
caller no longer has control of whether a sbr or hotplug reset is used,
which is undesirable for vfio-pci (811c5cb37df4 PCI: Unify try slot and
bus reset API) which really wants to specify since we do a fair bit or
work making sure the user owns the right set of devices for the
available type of reset.  Abstracting this allows PCI core and vfio to
get out of sync.  The idea of allowing the caller to specify the type
of reset is applicable here as well, unify the function, but still allow
the caller to specify.

For rc, I'd certainly agree that a simple revert might be a better
option and this API can play out for v4.20, but I definitely want to
get back an API that allows vfio to specify the bus reset type before
implementations diverge.  Thanks,

Alex

> > > > Fixes: 811c5cb37df4 ("PCI: Unify try slot and bus reset API")
> > > > Link: https://bugzilla.kernel.org/show_bug.cgi?id=200985
> > > > Signed-off-by: Sinan Kaya <okaya@kernel.org>
> > > > ---
> > > >   drivers/infiniband/hw/hfi1/pcie.c | 2 +-
> > > >   1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/infiniband/hw/hfi1/pcie.c b/drivers/infiniband/hw/hfi1/pcie.c
> > > > index eec83757d55f..13162289b748 100644
> > > > --- a/drivers/infiniband/hw/hfi1/pcie.c
> > > > +++ b/drivers/infiniband/hw/hfi1/pcie.c
> > > > @@ -900,7 +900,7 @@ static int trigger_sbr(struct hfi1_devdata *dd)
> > > >   	 * delay after a reset is required.  Per spec requirements,
> > > >   	 * the link is either working or not after that point.
> > > >   	 */
> > > > -	return pci_reset_bus(dev);
> > > > +	return __pci_reset_function_locked(dev, PCI_RESET_LINK);
> > > >   }
> > > >   /*
> > > > -- 
> > > > 2.18.0
> > > >   
> > >   
> >