[lkml]   [2020]   [Mar]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
SubjectRe: [PATCH v18 03/11] PCI/DPC: Fix DPC recovery issue in non hotplug case
On Tue, Mar 24, 2020 at 06:17:44PM -0700, Kuppuswamy, Sathyanarayanan wrote:
> On 3/24/20 4:49 PM, Bjorn Helgaas wrote:
> > I don't understand why hotplug is relevant here. This path
> > (dpc_reset_link()) is only used for downstream ports that support DPC.
> > DPC has already disabled the link, which resets everything below the
> > port, regardless of whether the port supports hotplug.
> >
> > I do see that PCI_ERS_RESULT_NEED_RESET seems to promise a lot more
> > than it actually *does*. The doc (pci-error-recovery.rst) says
> > .error_detected() can return PCI_ERS_RESULT_NEED_RESET to *request* a
> > slot reset. But if that happens, pcie_do_recovery() doesn't do a
> > reset at all. It calls the driver's .slot_reset() method, which tells
> > the driver "we've reset your device; please re-initialize the
> > hardware."
> >
> > I guess this abuses PCI_ERS_RESULT_NEED_RESET by taking advantage of
> > that implementation deficiency in pcie_do_recovery(): we know the
> > downstream devices have already been reset via DPC, and returning
> > PCI_ERS_RESULT_NEED_RESET means we'll call .slot_reset() to tell the
> > driver about that reset.
> >
> > I can see how this achieves the desired result, but if/when we fix
> > pcie_do_recovery() to actually *do* the reset promised by
> > PCI_ERS_RESULT_NEED_RESET, we will be doing *two* resets: the first
> > via DPC and a second via whatever slot reset mechanism
> > pcie_do_recovery() would use.
> When we fix this issue, if we make sure the reset logic is
> implemented before we call .reset_link callback we should be
> able to avoid resetting the device twice. Before we call DPC
> .reset_link callback, the device link will not up and hence we
> should not able to reset it.
> > So I guess the real issue (as you allude to in the commit log) is that
> > we rely on hotplug to unbind/rebind the driver, and without hotplug we
> > need to at least tell the driver the device was reset.
> Agree
> > I'll try to expand the comment here so it reminds me what's going on
> > when we have to look at this again:) Let me know if I'm on the right
> > track.
> Yes, your understanding is correct.

OK, thanks. I'm still uncomfortable with this issue, so I think I'm
going to apply this series but omit this patch. Here's why:

1) The fact that resets may cause hotplug events isn't specific to
DPC, so I don't think dpc_reset_link() is the right place. For
instance, aer_root_reset() ultimately does a secondary bus reset. The
pci_slot_reset() -> pciehp_reset_slot() path goes to some trouble to
ignore the resulting hotplug event, but the pci_bus_reset() path does

2) I'm not convinced that "hotplug_is_native()" is the correct test.
Even if we're using ACPI hotplug (acpiphp), that will detach the
drivers and remove the devices, won't it?

I considered something like the patch below, which partly addresses my
first concern, but not the second. Even the first one is awfully
messy because of the different ways the aer_root_reset() path can

PCI/ERR: Skip driver callbacks if reset causes hotplug remove/add

diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
index 1ac57e9e1e71..000551a06013 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -208,6 +208,18 @@ void pcie_do_recovery(struct pci_dev *dev, enum pci_channel_state state,
status = reset_link(dev, service);
goto failed;
+ /*
+ * If pdev supports hotplug, a link reset causes a hotplug
+ * remove event. If we have a hotplug driver, it will
+ * detach all drivers of downstream devices and remove the
+ * devices, so we can't call any driver error recovery
+ * callbacks. Bringing the link back up causes a hotplug
+ * add event, and the devices should be re-enumerated and
+ * the drivers re-attached.
+ */
+ if (hotplug_is_native(pdev))
+ goto succeeded;
} else {
pci_walk_bus(bus, report_normal_detected, &status);
@@ -224,7 +236,11 @@ void pcie_do_recovery(struct pci_dev *dev, enum pci_channel_state state,
* functions to reset slot before calling
* drivers' slot_reset callbacks?
+ pci_warn(pdev, "driver requested reset, but that's not implemented\n");
+ }
+ if (status == PCI_ERS_RESULT_RECOVERED) {
pci_dbg(dev, "broadcast slot_reset message\n");
pci_walk_bus(bus, report_slot_reset, &status);
@@ -235,6 +251,7 @@ void pcie_do_recovery(struct pci_dev *dev, enum pci_channel_state state,
pci_dbg(dev, "broadcast resume message\n");
pci_walk_bus(bus, report_resume, &status);

pci_info(dev, "device recovery successful\n");
 \ /
  Last update: 2020-03-28 18:11    [W:0.082 / U:0.160 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site