On Wed, Mar 11, 2020 at 03:11:06PM -0700, Kuppuswamy Sathyanarayanan wrote: > On 3/11/20 2:53 PM, Austin.Bolen@xxxxxxxx wrote: > > On 3/11/2020 4:27 PM, Kuppuswamy Sathyanarayanan wrote: > > > On 3/11/20 1:33 PM, Bjorn Helgaas wrote: > > > > On Wed, Mar 11, 2020 at 05:27:35PM +0000, Austin.Bolen@xxxxxxxx wrote: > > > > > On 3/11/2020 12:12 PM, Bjorn Helgaas wrote: > > > > > <SNIP> > > > > > > I'm probably missing your intent, but that sounds like "the OS can > > > > > > read/write AER bits whenever it wants, regardless of ownership." > > > > > > > > > > > > That doesn't sound practical to me, and I don't think it's really > > > > > > similar to DPC, where it's pretty clear that the OS can touch DPC bits > > > > > > it doesn't own but only *during the EDR processing window*. > > > > > Yes, by treating AER bits like DPC bits I meant I'd define the specific > > > > > time windows when OS can touch the AER status bits similar to how it's > > > > > done for DPC in the current ECN. > > > > Makes sense, thanks. > > > > > > > > > > > > > For the normative text describing when OS clears the AER bits > > > > > > > > > following the informative flow chart, it could say that OS clears > > > > > > > > > AER as soon as possible after OST returns and before OS processes > > > > > > > > > _HPX and loading drivers. Open to other suggestions as well. > > > > > > > > I'm not sure what to do with "as soon as possible" either. That > > > > > > > > doesn't seem like something firmware and the OS can agree on. > > > > > > > I can just state that it's done after OST returns but before _HPX or > > > > > > > driver is loaded. Any time in that range is fine. I can't get super > > > > > > > specific here because different OSes do different things. Even for > > > > > > > a given OS they change over time. And I need something generic > > > > > > > enough to support a wide variety of OS implementations. > > > > > > Yeah. I don't know how to solve this. > > > > > > > > > > > > Linux doesn't actually unload and reload drivers for the child devices > > > > > > (Sathy, correct me if I'm wrong here) even though DPC containment > > > > > > takes the link down and effectively unplugs and replugs the device. I > > > > > > would *like* to handle it like hotplug, but some higher-level software > > > > > > doesn't deal well with things like storage devices disappearing and > > > > > > reappearing. > > > > > > > > > > > > Since Linux doesn't actually re-enumerate the child devices, it > > > > > > wouldn't evaluate _HPX again. It would probably be cleaner if it did, > > > > > > but it's all tied up with the whole unplug/replug problem. > > > > > DPC resets everything below it and so to get it back up and running it > > > > > would mean that all buses and resources need to be assigned, _HPX > > > > > evaluated, and drivers reloaded. If those things don't happen then the > > > > > whole hierarchy below the port that triggered DPC will be inaccessible. > > > > Hmm, I think I might be confusing this with another situation. Sathy, > > > > can you help me understand this? I don't have a way to actually > > > > exercise this EDR path. Is there some way the pciehp hotplug driver > > > > gets involved here? > > If the port has hot-plug enabled then DPC trigger will cause the link to > > go down (disabled state) and will generate a DLLSC hot-plug interrupt. > > When DPC is released, the link will become active and generate another > > DLLSC hot-plug interrupt. > Yes, device/driver enumeration and removal will triggered by DLLSC > state change interrupt in pciehp driver. > > > > > > Here's how this seems to work as far as I can tell: > > > > > > > > - Linux does not have DPC or AER control > > > > > > > > - Linux installs EDR notify handler > > > > > > > > - Linux evaluates DPC Enable _DSM > > > > > > > > - DPC containment event occurs > > > > > > > > - Firmware fields DPC interrupt > > > > > > > > - DPC event is not a surprise remove > > > > > > > > - Firmware sends EDR notification > > > > > > > > - Linux EDR notify handler evaluates Locate _DSM > > > > > > > > - Linux reads and logs DPC and AER error information for port in > > > > containment mode. [If it was an RP PIO error, Linux clears RP PIO > > > > error status, which is an asymmetry with the non-RP PIO path.] > > > > > > > > - Linux clears AER error status (pci_aer_raw_clear_status()) > > > > > > > > - Linux calls driver .error_detected() methods for all child devices > > > > of the port in containment mode (pcie_do_recovery()). These > > > > devices are inaccessible because the link is down. > > > > > > > > - Linux clears DPC Trigger Status (dpc_reset_link() from > > > > pcie_do_recovery()). > > > > > > > > - Linux calls driver .mmio_enabled() methods for all child devices. > > > > > > > > This is where I get lost. These child devices are now accessible, but > > > > they've been reset, so I don't know how their config space got > > > > restored. Did pciehp enumerate them? Did we do something like > > > > pci_restore_state()? I don't see where either of these happens. > > > AFAIK, AER error status registers are sticky (RW1CS) and hence > > > will be preserved during reset. > > In our testing, the device directly connected to the port that was > > contained does get reprogrammed and the driver is reloaded. These are > > hot-plug slots and so might be due to DLLSC hot-plug interrupt when > > containment is released and link goes back to active state. > > > > However, if a switch is connected to the port where DPC was triggered > > then we do not see the whole switch hierarchy being re-enumerated. > Now that I have a hardware to verify this scenario, I will look into > it. I suspect there is a transient state in link status which causes > this disconnect issue. But I think this issue is not related to > EDR support and hence should be reproducible in native handling > as well. > > > > Also, DPC could be enabled on non-hot-plug slots so can't always rely on > > hot-plug to re-init devices in the recovery path. > If hotplug is not supported then there is support to enumerate > devices via polling or ACPI events. But a point to note > here is, enumeration path is independent of error handler path, and > hence there is no explicit trigger or event from error handler path > to enumeration path to kick start the enumeration. Is any synchronization needed here between the EDR path and the hotplug/enumeration path?
