Re: [PATCH v2] PCI: Re-enable downstream port LTR if it was previously enabled

[Mingchuang Qiao]

On Fri, 2021-01-22 at 07:20 -0600, Bjorn Helgaas wrote:
> On Fri, Jan 22, 2021 at 03:03:11PM +0800, Mingchuang Qiao wrote:
> > On Thu, 2021-01-21 at 16:31 -0600, Bjorn Helgaas wrote:
> > > [+cc Alex and Mingchuang et al from
> > > https://lore.kernel.org/r/20210112072739.31624-1-mingchuang.qiao@xxxxxxxxxxxx]
> > > 
> > > On Tue, Jan 19, 2021 at 04:14:10PM +0300, Mika Westerberg wrote:
> > > > PCIe r5.0, sec 7.5.3.16 says that the downstream ports must reset the
> > > > LTR enable bit if the link goes down (port goes DL_Down status). Now, if
> > > > we had LTR previously enabled and the PCIe endpoint gets hot-removed and
> > > > then hot-added back the ->ltr_path of the downstream port is still set
> > > > but the port now does not have the LTR enable bit set anymore.
> > > > 
> > > > For this reason check if the bridge upstream had LTR enabled previously
> > > > and re-enable it before enabling LTR for the endpoint.
> > > > 
> > > > Reported-by: Utkarsh H Patel <utkarsh.h.patel@xxxxxxxxx>
> > > > Signed-off-by: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
> > > 
> > > I think this and Mingchuang's patch, which is essentially identical,
> > > are right and solves the problem for hot-remove/hot-add.  In that
> > > scenario we call pci_configure_ltr() on the hot-added device, and
> > > with this patch, we'll re-enable LTR on the bridge leading to the new
> > > device before enabling LTR on the new device itself.
> > > 
> > > But don't we have a similar problem if we simply do a Fundamental
> > > Reset on a device?  I think the reset path will restore the device's
> > > state, including PCI_EXP_DEVCTL2, but it doesn't do anything with the
> > > upstream bridge, does it?
> > 
> > Yes. I think the same problem exists under such scenario, and that’s the
> > issue my patch intends to resolve.
> > I also prepared a v2 patch for review(update the patch description).
> > Shall I submit the v2 patch for review?
> 
> How does your patch solve this for the reset path?  I don't think we
> call pci_configure_ltr() when we reset a device.
> 

Sorry, I misunderstand the reset path. When we do a Fundamental Reset on
a device, we can trigger device removal and rescan flow to restore the
device. In device rescan flow, pci_configure_ltr() will be invoked. I
regard the "remove and rescan flow" as a part of reset path for this
case :)
If we restore device just with pci_restore_state() in device driver
after device resets, the LTR problem also exists due to
pci_restore_state() does nothing with upstream bridge. In next patch, I
would like to re-enable LTR for upstream bridge before restoring
device's PCI_EXP_DEVCTL2 if it is needed.

> > > So if a bridge and a device below it both have LTR enabled, can't we
> > > have the following:
> > > 
> > >   - bridge LTR enabled
> > >   - device LTR enabled
> > >   - reset device, e.g., via Secondary Bus Reset
> > >   - link goes down, bridge disables LTR
> > >   - link comes back up, LTR disabled in both bridge and device
> > >   - restore device state, including LTR enable
> > >   - device sends LTR message
> > >   - bridge reports Unsupported Request
> > > 
> > > > ---
> > > > Previous version can be found here:
> > > > 
> > > >   https://lore.kernel.org/linux-pci/20210114134724.79511-1-mika.westerberg@xxxxxxxxxxxxxxx/
> > > > 
> > > > Changes from the previous version:
> > > > 
> > > >   * Corrected typos in the commit message
> > > >   * No need to call pcie_downstream_port()
> > > > 
> > > >  drivers/pci/probe.c | 17 ++++++++++++++++-
> > > >  1 file changed, 16 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> > > > index 0eb68b47354f..a4a8c0305fb9 100644
> > > > --- a/drivers/pci/probe.c
> > > > +++ b/drivers/pci/probe.c
> > > > @@ -2153,7 +2153,7 @@ static void pci_configure_ltr(struct pci_dev *dev)
> > > >  {
> > > >  #ifdef CONFIG_PCIEASPM
> > > >  	struct pci_host_bridge *host = pci_find_host_bridge(dev->bus);
> > > > -	struct pci_dev *bridge;
> > > > +	struct pci_dev *bridge = NULL;
> > > >  	u32 cap, ctl;
> > > >  
> > > >  	if (!pci_is_pcie(dev))
> > > > @@ -2191,6 +2191,21 @@ static void pci_configure_ltr(struct pci_dev *dev)
> > > >  	if (pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT ||
> > > >  	    ((bridge = pci_upstream_bridge(dev)) &&
> > > >  	      bridge->ltr_path)) {
> > > > +		/*
> > > > +		 * Downstream ports reset the LTR enable bit when the
> > > > +		 * link goes down (e.g on hot-remove) so re-enable the
> > > > +		 * bit here if not set anymore.
> > > > +		 * PCIe r5.0, sec 7.5.3.16.
> > > > +		 */
> > > > +		if (bridge) {
> > > > +			pcie_capability_read_dword(bridge, PCI_EXP_DEVCTL2, &ctl);
> > > > +			if (!(ctl & PCI_EXP_DEVCTL2_LTR_EN)) {
> > > > +				pci_dbg(bridge, "re-enabling LTR\n");
> > > > +				pcie_capability_set_word(bridge, PCI_EXP_DEVCTL2,
> > > > +							 PCI_EXP_DEVCTL2_LTR_EN);
> > > > +			}
> > > > +		}
> > > > +		pci_dbg(dev, "enabling LTR\n");
> > > >  		pcie_capability_set_word(dev, PCI_EXP_DEVCTL2,
> > > >  					 PCI_EXP_DEVCTL2_LTR_EN);
> > > >  		dev->ltr_path = 1;
> > > > -- 
> > > > 2.29.2
> > > > 
> >