Re: [RFC PATCH 5/9] PCI/AER: Apply function level reset to RCiEP on fatal error

From: Sean V Kelley
Date: Tue Jul 28 2020 - 13:42:39 EST


On 28 Jul 2020, at 10:02, Jonathan Cameron wrote:

On Tue, 28 Jul 2020 09:14:11 -0700
Sean V Kelley <sean.v.kelley@xxxxxxxxx> wrote:

On 28 Jul 2020, at 6:27, Zhuo, Qiuxu wrote:

From: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
Sent: Monday, July 27, 2020 7:17 PM
To: Kelley, Sean V <sean.v.kelley@xxxxxxxxx>
Cc: bhelgaas@xxxxxxxxxx; rjw@xxxxxxxxxxxxx; ashok.raj@xxxxxxxxxx;
Luck,
Tony <tony.luck@xxxxxxxxx>;
sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx;
linux-pci@xxxxxxxxxxxxxxx;
linux-kernel@xxxxxxxxxxxxxxx; Zhuo, Qiuxu <qiuxu.zhuo@xxxxxxxxx>
Subject: Re: [RFC PATCH 5/9] PCI/AER: Apply function level reset to
RCiEP
on fatal error

On Fri, 24 Jul 2020 10:22:19 -0700
Sean V Kelley <sean.v.kelley@xxxxxxxxx> wrote:

From: Qiuxu Zhuo <qiuxu.zhuo@xxxxxxxxx>

Attempt to do function level reset for an RCiEP associated with an
RCEC device on fatal error.

I'd like to understand more on your reasoning for flr here.
Is it simply that it is all we can do, or is there some basis in a
spec
somewhere?


Yes. Though there isn't the link reset for the RCiEP here, I think we
should still be able to reset the RCiEP via FLR on fatal error, if the
RCiEP supports FLR.

-Qiuxu


Also see PCIe 5.0-1, Sec. 6.6.2 Function Level Reset (FLR)

Implementation of FLR is optional (not required), but is strongly
recommended. For an example use case consider CXL. Function 0 DVSEC
instances control for the CXL functionality of the entire CXL device.
FLR may succeed in recovering from CXL.io domain errors.

That feels a little bit of a weak argument in favour. PCI spec lists examples
of use only for FLR and I can't see this matching any of them, but then they
are only examples, so we could argue it doesn't exclude this use. It's not
allowed to affect the link state, but I guess it 'might' recover from some
other type of error?

I'd have read the statement in the CXL spec you are referring to as matching
with the first example in the PCIe spec which is about recovering from
software errors. For example, unexpected VM tear down.

From my perspective, it can add value as the point is to address device functions and their associated software states. As the section in the spec goes on to state:

âThe FLR mechanism enables software to quiesce and reset Endpoint hardware with Function-level granularity. Three example usage models illustrate the benefits of this feature:ââ

Later changes in CXL 2.0 Section 9.8 (as of 0.9 draft) further look to extend FLR with an eFLR or now referred to as CXL Reset.

âAll Functions in a CXL 2.0 (Single Logical Device) SLD that participate in CXL.cache or CXL.mem are required to support either FLR or CXL Reset. MLDs (Multiple Logical Devices), on the other hand, are required to support CXL Reset.â

In my mind the question is whether this change is too limited in scope with this patch series (RCiEP) and whether FLR should be considered in a broader, i.e., EP, as a âhammerâ so to speak.

Thanks,

Sean


@Bjorn / All. What's your view on using FLR as a reset to do when you don't
have any other hammers to use?

Personally I don't have a particular problem with this, it just doesn't fit
with my mental model of what FLR is for (which may well need adjusting :)

Jonathan



Thanks,

Sean


Signed-off-by: Qiuxu Zhuo <qiuxu.zhuo@xxxxxxxxx>
---
drivers/pci/pcie/err.c | 31 ++++++++++++++++++++++---------
1 file changed, 22 insertions(+), 9 deletions(-)

diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c index
044df004f20b..9b3ec94bdf1d 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -170,6 +170,17 @@ static void pci_walk_dev_affected(struct
pci_dev *dev, int (*cb)(struct pci_dev
}
}

+static enum pci_channel_state flr_on_rciep(struct pci_dev *dev) {
+if (!pcie_has_flr(dev))
+return PCI_ERS_RESULT_NONE;
+
+if (pcie_flr(dev))
+return PCI_ERS_RESULT_DISCONNECT;
+
+return PCI_ERS_RESULT_RECOVERED;
+}
+
pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
enum pci_channel_state state,
pci_ers_result_t (*reset_link)(struct pci_dev *pdev))
@@ -191,15
+202,17 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
if (state == pci_channel_io_frozen) {
pci_walk_dev_affected(dev, report_frozen_detected,
&status);
if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END) {
-pci_warn(dev, "link reset not possible for RCiEP\n");
-status = PCI_ERS_RESULT_NONE;
-goto failed;
-}
-
-status = reset_link(dev);
-if (status != PCI_ERS_RESULT_RECOVERED) {
-pci_warn(dev, "link reset failed\n");
-goto failed;
+status = flr_on_rciep(dev);
+if (status != PCI_ERS_RESULT_RECOVERED) {
+pci_warn(dev, "function level reset failed\n");
+goto failed;
+}
+} else {
+status = reset_link(dev);
+if (status != PCI_ERS_RESULT_RECOVERED) {
+pci_warn(dev, "link reset failed\n");
+goto failed;
+}
}
} else {
pci_walk_dev_affected(dev, report_normal_detected,
&status);