mirror of
https://github.com/AuxXxilium/linux_dsm_epyc7002.git
synced 2024-12-15 15:46:52 +07:00
ad9001f2f4
Currently Linux does not follow PCIe spec regarding the required delays
after reset. A concrete example is a Thunderbolt add-in-card that consists
of a PCIe switch and two PCIe endpoints:
+-1b.0-[01-6b]----00.0-[02-6b]--+-00.0-[03]----00.0 TBT controller
+-01.0-[04-36]-- DS hotplug port
+-02.0-[37]----00.0 xHCI controller
\-04.0-[38-6b]-- DS hotplug port
The root port (1b.0) and the PCIe switch downstream ports are all PCIe Gen3
so they support 8GT/s link speeds.
We wait for the PCIe hierarchy to enter D3cold (runtime):
pcieport 0000:00:1b.0: power state changed by ACPI to D3cold
When it wakes up from D3cold, according to the PCIe 5.0 section 5.8 the
PCIe switch is put to reset and its power is re-applied. This means that we
must follow the rules in PCIe 5.0 section 6.6.1.
For the PCIe Gen3 ports we are dealing with here, the following applies:
With a Downstream Port that supports Link speeds greater than 5.0 GT/s,
software must wait a minimum of 100 ms after Link training completes
before sending a Configuration Request to the device immediately below
that Port. Software can determine when Link training completes by polling
the Data Link Layer Link Active bit or by setting up an associated
interrupt (see Section 6.7.3.3).
Translating this into the above topology we would need to do this (DLLLA
stands for Data Link Layer Link Active):
0000:00:1b.0: wait for 100 ms after DLLLA is set before access to 0000:01:00.0
0000:02:00.0: wait for 100 ms after DLLLA is set before access to 0000:03:00.0
0000:02:02.0: wait for 100 ms after DLLLA is set before access to 0000:37:00.0
I've instrumented the kernel with some additional logging so we can see the
actual delays performed:
pcieport 0000:00:1b.0: power state changed by ACPI to D0
pcieport 0000:00:1b.0: waiting for D3cold delay of 100 ms
pcieport 0000:00:1b.0: waiting for D3hot delay of 10 ms
pcieport 0000:02:01.0: waiting for D3hot delay of 10 ms
pcieport 0000:02:04.0: waiting for D3hot delay of 10 ms
For the switch upstream port (01:00.0 reachable through 00:1b.0 root port)
we wait for 100 ms but not taking into account the DLLLA requirement. We
then wait 10 ms for D3hot -> D0 transition of the root port and the two
downstream hotplug ports. This means that we deviate from what the spec
requires.
Performing the same check for system sleep (s2idle) transitions it turns
out to be even worse. None of the mandatory delays are performed. If this
would be S3 instead of s2idle then according to PCI FW spec 3.2 section
4.6.8. there is a specific _DSM that allows the OS to skip the delays but
this platform does not provide the _DSM and does not go to S3 anyway so no
firmware is involved that could already handle these delays.
On this particular platform these delays are not actually needed because
there is an additional delay as part of the ACPI power resource that is
used to turn on power to the hierarchy but since that additional delay is
not required by any of standards (PCIe, ACPI) it is not present in the
Intel Ice Lake, for example where missing the mandatory delays causes
pciehp to start tearing down the stack too early (links are not yet
trained). Below is an example how it looks like when this happens:
pcieport 0000:83:04.0: pciehp: Slot(4): Card not present
pcieport 0000:87:04.0: PME# disabled
pcieport 0000:83:04.0: pciehp: pciehp_unconfigure_device: domain🚌dev = 0000:86:00
pcieport 0000:86:00.0: Refused to change power state, currently in D3
pcieport 0000:86:00.0: restoring config space at offset 0x3c (was 0xffffffff, writing 0x201ff)
pcieport 0000:86:00.0: restoring config space at offset 0x38 (was 0xffffffff, writing 0x0)
...
There is also one reported case (see the bugzilla link below) where the
missing delay causes xHCI on a Titan Ridge controller fail to runtime
resume when USB-C dock is plugged. This does not involve pciehp but instead
the PCI core fails to runtime resume the xHCI device:
pcieport 0000:04:02.0: restoring config space at offset 0xc (was 0x10000, writing 0x10020)
pcieport 0000:04:02.0: restoring config space at offset 0x4 (was 0x100000, writing 0x100406)
xhci_hcd 0000:39:00.0: Refused to change power state, currently in D3
xhci_hcd 0000:39:00.0: restoring config space at offset 0x3c (was 0xffffffff, writing 0x1ff)
xhci_hcd 0000:39:00.0: restoring config space at offset 0x38 (was 0xffffffff, writing 0x0)
...
Add a new function pci_bridge_wait_for_secondary_bus() that is called on
PCI core resume and runtime resume paths accordingly if the bridge entered
D3cold (and thus went through reset).
This is second attempt to add the missing delays. The previous solution in
|
||
|---|---|---|
| .. | ||
| controller | ||
| endpoint | ||
| hotplug | ||
| pcie | ||
| switch | ||
| access.c | ||
| ats.c | ||
| bus.c | ||
| ecam.c | ||
| host-bridge.c | ||
| iov.c | ||
| irq.c | ||
| Kconfig | ||
| Makefile | ||
| mmap.c | ||
| msi.c | ||
| of.c | ||
| p2pdma.c | ||
| pci-acpi.c | ||
| pci-bridge-emul.c | ||
| pci-bridge-emul.h | ||
| pci-driver.c | ||
| pci-label.c | ||
| pci-mid.c | ||
| pci-pf-stub.c | ||
| pci-stub.c | ||
| pci-sysfs.c | ||
| pci.c | ||
| pci.h | ||
| probe.c | ||
| proc.c | ||
| quirks.c | ||
| remove.c | ||
| rom.c | ||
| search.c | ||
| setup-bus.c | ||
| setup-irq.c | ||
| setup-res.c | ||
| slot.c | ||
| syscall.c | ||
| vc.c | ||
| vpd.c | ||
| xen-pcifront.c | ||