The series of patches bases on linux-poerpc-next initially and intends to
resolve
the following problems:
- On pSeries platform, the EEH doesn't work after PHB hotplug
with "drmgr". The root cause is that the EEH resources (
EEH devices, EEH caches) aren't released correctly. For the
problem, we add one hook (pcibios_stop_dev), which is called
on pci_stop_and_remove_device(). In pcibios_stop_dev(), we
release the EEH resources.
- Another issue is that we need put the domain (PE or PHB) into
quite state while doing reset on that domain. However, some
deivces in the domain might not have EEH sensitive drivers, or
even don't have driver. Those deivces can't be put into quite
state and possibly keep issuing PCI-CFG or MMIO request during
resetting the domain. That possibly causes the failure of reset
and eventually failure of EEH recovery. For the issue, we introduces
so-called "partial hotplug". That means, those devices without driver
or
without EEH sensitive driver are removed before doing reset, and
plugged (probed) into the system after reset.
- We need traverse EEH devices of one specific PE with safe variant
of list tranverse function. The EEH device might be removed while
doing iteration.
- When doing plug for PCI bus, we need check if we need reassign the
resources for subordinate devices (PCI_REASSIGN_ALL_RSRC) and do that
accordingly.
The patchset is verified on pSeires and PowerNV platforms:
pSeries Platform:
drmgr -c phb -r -s "PHB 513"
drmgr -c phb -a -s "PHB 513"
errinjct eeh -f 1 -s net/eth2
PowerNV Platform:
cd
/sys/devices/pci0005:00/0005:00:00.0/0005:01:00.0/0005:02:08.0/0005:80:00.0/0005:90:01.0
while true; do od -x config > /dev/null; sleep 1; done
echo 1 > /sys/kernel/debug/powerpc/PCI0005/err_injct
---
v3 -> v4:
* Add some comments to explain why we needn't check the return
value of pci_scan_slot() in pcibios_add_pci_devices().
* Check PCI_PROBE_ONLY while assigning those unassigned resources
in pcibios_finish_adding_to_bus().
v2 -> v3:
* Make pcibios_add_pci_devices() to support "partial" hotplug
according to Ben's comments. arch/powerpc/kernel/pci_of_scan.c
has been adjusted for that.
* Use pcibios_add_pci_devices() to do "partial" hotplug inside
eeh_reset_device().
* Introduce flag EEH_DEV_SYSFS to trace the state of sysfs entries
of the EEH device (then PCI device) to avoid race condition during
"partial" hotplug.
v1 -> v2:
* Rebase to 3.11.rc1 in order to use pcibios_release_device().
* Use pcibios_release_device() to release EEH cache and detach
EEH device from PCI device.
* Remove reference to PCI device in EEH cache since we're relying
on pcibios_release_device().
* PCI device instance (struct pci_dev) isn't available during BAR
restore and avoid use the instance that time.
* Fix unbalanced enable for IRQ in eeh_driver.c
* Retest the series of patches on Firebird-L/VPL3/VPL4
---
arch/powerpc/include/asm/eeh.h | 30 ++++++++--
arch/powerpc/include/asm/pci-bridge.h | 1 -
arch/powerpc/kernel/eeh.c | 70 +++++++++++------------
arch/powerpc/kernel/eeh_cache.c | 18 ++----
arch/powerpc/kernel/eeh_driver.c | 77 +++++++++++++++++++++++++-
arch/powerpc/kernel/eeh_pe.c | 58 ++++++++-----------
arch/powerpc/kernel/eeh_sysfs.c | 21 +++++++
arch/powerpc/kernel/pci-common.c | 2 +
arch/powerpc/kernel/pci-hotplug.c | 49 ++++++++--------
arch/powerpc/kernel/pci_of_scan.c | 56 +++++++++++++-----
arch/powerpc/platforms/powernv/eeh-powernv.c | 17 +++++-
arch/powerpc/platforms/pseries/eeh_pseries.c | 67 +++++++++++++++++++++-
drivers/pci/hotplug/rpadlpar_core.c | 1 -
13 files changed, 327 insertions(+), 140 deletions(-)
Thanks,
Gavin
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
