On Wed, Jun 03, 2020 at 11:12:48PM +0530, Prabhakar Kushwaha wrote:
> On Sat, May 30, 2020 at 1:03 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > On Fri, May 29, 2020 at 07:48:10PM +0530, Prabhakar Kushwaha wrote:
> > > On Thu, May 28, 2020 at 1:48 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, May 27, 2020 at 05:14:39PM +0530, Prabhakar Kushwaha wrote:
> > > > > On Fri, May 22, 2020 at 4:19 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > > > > On Thu, May 21, 2020 at 09:28:20AM +0530, Prabhakar Kushwaha wrote:
> > > > > > > On Wed, May 20, 2020 at 4:52 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > > > > > > On Thu, May 14, 2020 at 12:47:02PM +0530, Prabhakar Kushwaha wrote:
> > > > > > > > > On Wed, May 13, 2020 at 3:33 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > > > > > > > > On Mon, May 11, 2020 at 07:46:06PM -0700, Prabhakar Kushwaha wrote:
> > > > > > > > > > > An SMMU Stream table is created by the primary kernel. This table is
> > > > > > > > > > > used by the SMMU to perform address translations for device-originated
> > > > > > > > > > > transactions. Any crash (if happened) launches the kdump kernel which
> > > > > > > > > > > re-creates the SMMU Stream table. New transactions will be translated
> > > > > > > > > > > via this new table..
> > > > > > > > > > >
> > > > > > > > > > > There are scenarios, where devices are still having old pending
> > > > > > > > > > > transactions (configured in the primary kernel). These transactions
> > > > > > > > > > > come in-between Stream table creation and device-driver probe.
> > > > > > > > > > > As new stream table does not have entry for older transactions,
> > > > > > > > > > > it will be aborted by SMMU.
> > > > > > > > > > >
> > > > > > > > > > > Similar observations were found with PCIe-Intel 82576 Gigabit
> > > > > > > > > > > Network card. It sends old Memory Read transaction in kdump kernel.
> > > > > > > > > > > Transactions configured for older Stream table entries, that do not
> > > > > > > > > > > exist any longer in the new table, will cause a PCIe Completion Abort.
> > > > > > > > > >
> > > > > > > > > > That sounds like exactly what we want, doesn't it?
> > > > > > > > > >
> > > > > > > > > > Or do you *want* DMA from the previous kernel to complete? That will
> > > > > > > > > > read or scribble on something, but maybe that's not terrible as long
> > > > > > > > > > as it's not memory used by the kdump kernel.
> > > > > > > > >
> > > > > > > > > Yes, Abort should happen. But it should happen in context of driver.
> > > > > > > > > But current abort is happening because of SMMU and no driver/pcie
> > > > > > > > > setup present at this moment.
> > > > > > > >
> > > > > > > > I don't understand what you mean by "in context of driver." The whole
> > > > > > > > problem is that we can't control *when* the abort happens, so it may
> > > > > > > > happen in *any* context. It may happen when a NIC receives a packet
> > > > > > > > or at some other unpredictable time.
> > > > > > > >
> > > > > > > > > Solution of this issue should be at 2 place
> > > > > > > > > a) SMMU level: I still believe, this patch has potential to overcome
> > > > > > > > > issue till finally driver's probe takeover.
> > > > > > > > > b) Device level: Even if something goes wrong. Driver/device should
> > > > > > > > > able to recover.
> > > > > > > > >
> > > > > > > > > > > Returned PCIe completion abort further leads to AER Errors from APEI
> > > > > > > > > > > Generic Hardware Error Source (GHES) with completion timeout.
> > > > > > > > > > > A network device hang is observed even after continuous
> > > > > > > > > > > reset/recovery from driver, Hence device is no more usable.
> > > > > > > > > >
> > > > > > > > > > The fact that the device is no longer usable is definitely a problem.
> > > > > > > > > > But in principle we *should* be able to recover from these errors. If
> > > > > > > > > > we could recover and reliably use the device after the error, that
> > > > > > > > > > seems like it would be a more robust solution that having to add
> > > > > > > > > > special cases in every IOMMU driver.
> > > > > > > > > >
> > > > > > > > > > If you have details about this sort of error, I'd like to try to fix
> > > > > > > > > > it because we want to recover from that sort of error in normal
> > > > > > > > > > (non-crash) situations as well.
> > > > > > > > > >
> > > > > > > > > Completion abort case should be gracefully handled. And device should
> > > > > > > > > always remain usable.
> > > > > > > > >
> > > > > > > > > There are 2 scenario which I am testing with Ethernet card PCIe-Intel
> > > > > > > > > 82576 Gigabit Network card.
> > > > > > > > >
> > > > > > > > > I) Crash testing using kdump root file system: De-facto scenario
> > > > > > > > > - kdump file system does not have Ethernet driver
> > > > > > > > > - A lot of AER prints [1], making it impossible to work on shell
> > > > > > > > > of kdump root file system.
> > > > > > > >
> > > > > > > > In this case, I think report_error_detected() is deciding that because
> > > > > > > > the device has no driver, we can't do anything. The flow is like
> > > > > > > > this:
> > > > > > > >
> > > > > > > > aer_recover_work_func # aer_recover_work
> > > > > > > > kfifo_get(aer_recover_ring, entry)
> > > > > > > > dev = pci_get_domain_bus_and_slot
> > > > > > > > cper_print_aer(dev, ...)
> > > > > > > > pci_err("AER: aer_status:")
> > > > > > > > pci_err("AER: [14] CmpltTO")
> > > > > > > > pci_err("AER: aer_layer=")
> > > > > > > > if (AER_NONFATAL)
> > > > > > > > pcie_do_recovery(dev, pci_channel_io_normal)
> > > > > > > > status = CAN_RECOVER
> > > > > > > > pci_walk_bus(report_normal_detected)
> > > > > > > > report_error_detected
> > > > > > > > if (!dev->driver)
> > > > > > > > vote = NO_AER_DRIVER
> > > > > > > > pci_info("can't recover (no error_detected callback)")
> > > > > > > > *result = merge_result(*, NO_AER_DRIVER)
> > > > > > > > # always NO_AER_DRIVER
> > > > > > > > status is now NO_AER_DRIVER
> > > > > > > >
> > > > > > > > So pcie_do_recovery() does not call .report_mmio_enabled() or .slot_reset(),
> > > > > > > > and status is not RECOVERED, so it skips .resume().
> > > > > > > >
> > > > > > > > I don't remember the history there, but if a device has no driver and
> > > > > > > > the device generates errors, it seems like we ought to be able to
> > > > > > > > reset it.
> > > > > > >
> > > > > > > But how to reset the device considering there is no driver.
> > > > > > > Hypothetically, this case should be taken care by PCIe subsystem to
> > > > > > > perform reset at PCIe level.
> > > > > >
> > > > > > I don't understand your question. The PCI core (not the device
> > > > > > driver) already does the reset. When pcie_do_recovery() calls
> > > > > > reset_link(), all devices on the other side of the link are reset.
> > > > > >
> > > > > > > > We should be able to field one (or a few) AER errors, reset the
> > > > > > > > device, and you should be able to use the shell in the kdump kernel.
> > > > > > > >
> > > > > > > here kdump shell is usable only problem is a "lot of AER Errors". One
> > > > > > > cannot see what they are typing.
> > > > > >
> > > > > > Right, that's what I expect. If the PCI core resets the device, you
> > > > > > should get just a few AER errors, and they should stop after the
> > > > > > device is reset.
> > > > > >
> > > > > > > > > - Note kdump shell allows to use makedumpfile, vmcore-dmesg applications.
> > > > > > > > >
> > > > > > > > > II) Crash testing using default root file system: Specific case to
> > > > > > > > > test Ethernet driver in second kernel
> > > > > > > > > - Default root file system have Ethernet driver
> > > > > > > > > - AER error comes even before the driver probe starts.
> > > > > > > > > - Driver does reset Ethernet card as part of probe but no success.
> > > > > > > > > - AER also tries to recover. but no success. [2]
> > > > > > > > > - I also tries to remove AER errors by using "pci=noaer" bootargs
> > > > > > > > > and commenting ghes_handle_aer() from GHES driver..
> > > > > > > > > than different set of errors come which also never able to recover [3]
> > > > > > > > >
> > > > > > >
> > > > > > > Please suggest your view on this case. Here driver is preset.
> > > > > > > (driver/net/ethernet/intel/igb/igb_main.c)
> > > > > > > In this case AER errors starts even before driver probe starts.
> > > > > > > After probe, driver does the device reset with no success and even AER
> > > > > > > recovery does not work.
> > > > > >
> > > > > > This case should be the same as the one above. If we can change the
> > > > > > PCI core so it can reset the device when there's no driver, that would
> > > > > > apply to case I (where there will never be a driver) and to case II
> > > > > > (where there is no driver now, but a driver will probe the device
> > > > > > later).
> > > > >
> > > > > Does this means change are required in PCI core.
> > > >
> > > > Yes, I am suggesting that the PCI core does not do the right thing
> > > > here.
> > > >
> > > > > I tried following changes in pcie_do_recovery() but it did not help.
> > > > > Same error as before.
> > > > >
> > > > > -- a/drivers/pci/pcie/err.c
> > > > > +++ b/drivers/pci/pcie/err.c
> > > > > pci_info(dev, "broadcast resume message\n");
> > > > > pci_walk_bus(bus, report_resume, &status);
> > > > > @@ -203,7 +207,12 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
> > > > > return status;
> > > > >
> > > > > failed:
> > > > > pci_uevent_ers(dev, PCI_ERS_RESULT_DISCONNECT);
> > > > > + pci_reset_function(dev);
> > > > > + pci_aer_clear_device_status(dev);
> > > > > + pci_aer_clear_nonfatal_status(dev);
> > > >
> > > > Did you confirm that this resets the devices in question (0000:09:00.0
> > > > and 0000:09:00.1, I think), and what reset mechanism this uses (FLR,
> > > > PM, etc)?
> > >
> > > Earlier reset was happening with P2P bridge(0000:00:09.0) this the
> > > reason no effect. After making following changes, both devices are
> > > now getting reset.
> > > Both devices are using FLR.
> > >
> > > diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
> > > index 117c0a2b2ba4..26b908f55aef 100644
> > > --- a/drivers/pci/pcie/err.c
> > > +++ b/drivers/pci/pcie/err.c
> > > @@ -66,6 +66,20 @@ static int report_error_detected(struct pci_dev *dev,
> > > if (dev->hdr_type != PCI_HEADER_TYPE_BRIDGE) {
> > > vote = PCI_ERS_RESULT_NO_AER_DRIVER;
> > > pci_info(dev, "can't recover (no
> > > error_detected callback)\n");
> > > +
> > > + pci_save_state(dev);
> > > + pci_cfg_access_lock(dev);
> > > +
> > > + /* Quiesce the device completely */
> > > + pci_write_config_word(dev, PCI_COMMAND,
> > > + PCI_COMMAND_INTX_DISABLE);
> > > + if (!__pci_reset_function_locked(dev)) {
> > > + vote = PCI_ERS_RESULT_RECOVERED;
> > > + pci_info(dev, "recovered via pci level
> > > reset\n");
> > > + }
> >
> > Why do we need to save the state and quiesce the device? The reset
> > should disable interrupts anyway. In this particular case where
> > there's no driver, I don't think we should have to restore the state.
> > We maybe should *remove* the device and re-enumerate it after the
> > reset, but the state from before the reset should be irrelevant.
>
> I tried pci_reset_function_locked without save/restore then I got the
> synchronous abort during igb_probe (case 2 i.e. with driver). This is
> 100% reproducible.
> looks like pci_reset_function_locked is causing PCI configuration
> space random. Same is mentioned here
> https://www.kernel.org/doc/html/latest/driver-api/pci/pci.html
That documentation is poorly worded. A reset doesn't make the
contents of config space "random," but of course it sets config space
registers to their initialization values, including things like the
device BARs. After a reset, the device BARs are zero, so it won't
respond at the address we expect, and I'm sure that's what's causing
the external abort.
So I guess we *do* need to save the state before the reset and restore
it (either that or enumerate the device from scratch just like we
would if it had been hot-added). I'm not really thrilled with trying
to save the state after the device has already reported an error. I'd
rather do it earlier, maybe during enumeration, like in
pci_init_capabilities(). But I don't understand all the subtleties of
dev->state_saved, so that requires some legwork.
I don't think we should set INTX_DISABLE; the reset will make whatever
we do with it irrelevant anyway.
Remind me why the pci_cfg_access_lock()?
Bjorn
