On 08/07/2012 12:10 PM, Jiang Liu wrote:
From: Jiang Liu<liuj97@xxxxxxxxx>
This is the second take to resolve race conditions when hot-plugging PCI
devices/host bridges. Instead of using a globla lock to serialize all hotplug
operations as in previous version, now we introduce a state machine and bit
lock mechanism for PCI buses to serialize hotplug operations. For discussions
related to previous version, please refer to:
http://comments.gmane.org/gmane.linux.kernel.pci/15007
This patch-set is still in early stages, so sending it out just requesting
for comments. Any comments are welcomed, especially about whether it's the
right/suitable way to solve these race condition issues.
patch 1-5:
Preparing for coming PCI bus lock
patch 6-7:
Core of the new PCI bus lock mechanism.
patch 8-13:
Enhance PCI core to support PCI bus lock mechanism.
patch 14-18:
Enhance several PCI hotplug drivers to use PCI bus lock to serialize
hotplug operations.
patch 19-20:
Enable PCI bus lock mechanism for x86 and IA64, still need to enable
PCI bus lock for other archs.
patch 21-22:
Cleanups for unsed code.
There are multiple methods to trigger PCI hotplug requests/operations
concurrently, such as:
1. Sysfs interfaces exported by the PCI core subsystem
/sys/devices/pcissss:bb/ssss:bb:dd.f/.../remove
/sys/devices/pcissss:bb/ssss:bb:dd.f/.../rescan
/sys/devices/pcissss:bb/ssss:bb:dd.f/.../pci_bus/ssss:bb/rescan
/sys/bus/pci/rescan
2. Sysfs interfaces exported by the PCI hotplug subsystem
/sys/bus/pci/slots/xx/power
3. PCI hotplug events triggered by PCI Hotplug Controllers
4. ACPI hotplug events for PCI host bridges
5. Driver binding/unbinding events
binding/unbinding pci drivers with SR-IOV support
6. PCI reset
--> a PCIe device-level reset is done by KVM when it assigns a device
to a guest. a PCI config-save before reset, and PCI config-restore after reset
is done in this case.
--> VF devices are interesting, since they are reset, then bound to
pci-stub driver. when more than 1 VF is enabled in a PF,
and several device-assignments are done simultaneously, you
get a storm of reset (save/restore pci cfg space), and pci-stub binding
(pci cfg read for resource allocation/deallocation), and depending on
the hw design: an AER caused by the FLR reset -- not suppose to, but
hw has bugs too! ;-)
PCI locking is 'challenged' in the above scenario.
So, I ask: have you tried your patch set doing something like:
a) modprobe an SRIOV device with > 1 vf enabled
you may also have to do:
b) while assigning another SRIOV device's VF to another KVM guest
Unfortunately, the PCI cfg-space locking, esp. on x86 (ok, I'll say it:
damn, mutually exclusive, IO-port-based cfg registers), doesn't lend itself
to this multi-task, dynamic PCI scenario.
Much less complicated on linearly-mapped, PCI-mmconf-only accesses.
- Don
With current implementation, the PCI core subsystem doesn't support
concurrent hotplug operations yet. The existing pci_bus_sem lock only
protects several lists in struct pci_bus, such as children list,
devices list, but it doesn't protect the pci_bus or pci_dev structure
themselves.
Let's take pci_remove_bus_device() as an example, which are used by
PCI hotplug drivers to hot-remove PCI devices. Currently all these
are free running without any protection, so it can't support reentrance.
pci_remove_bus_device()
->pci_stop_bus_device()
->pci_stop_bus_device()
->pci_stop_bus_devices()
->pci_stop_dev()
Jiang Liu (22):
PCI: use pci_get_domain_bus_and_slot() to avoid race conditions
PCI: trivial cleanups for drivers/pci/remove.c
PCI: change PCI device management code to better follow device model
PCI: split PCI bus device registration into two stages
PCI: introduce pci_bus_{get|put}() to manage PCI bus reference count
PCI: use a global lock to serialize PCI root bridge hotplug
operations
PCI: introduce PCI bus lock to serialize PCI hotplug operations
PCI: introduce hotplug safe search interfaces for PCI bus/device
PCI: enhance PCI probe logic to support PCI bus lock mechanism
PCI: enhance PCI bus specific logic to support PCI bus lock mechanism
PCI: enhance PCI resource assignment logic to support PCI bus lock
mechanism
PCI: enhance PCI remove logic to support PCI bus lock mechanism
PCI: make each PCI device hold a reference to its parent PCI bus
PCI/sysfs: use PCI bus lock to avoid race conditions
PCI/eeepc: use PCI bus lock to avoid race conditions
PCI/asus-wmi: use PCI bus lock to avoid race conditions
PCI/pciehp: use PCI bus lock to avoid race conditions
PCI/acpiphp: use PCI bus lock to avoid race conditions
PCI/x86: enable PCI bus lock mechanism for x86 platforms
PCI/IA64: enable PCI bus lock mechanism for IA64 platforms
PCI: cleanups for PCI bus lock implementation
PCI: unexport pci_root_buses
arch/ia64/pci/pci.c | 2 +
arch/ia64/sn/kernel/io_common.c | 4 +-
arch/ia64/sn/kernel/io_init.c | 1 +
arch/ia64/sn/pci/tioca_provider.c | 4 +-
arch/x86/pci/acpi.c | 6 +-
arch/x86/pci/common.c | 12 +++
drivers/acpi/pci_root.c | 8 +-
drivers/edac/i7core_edac.c | 16 ++-
drivers/gpu/drm/drm_fops.c | 6 +-
drivers/gpu/vga/vgaarb.c | 15 +--
drivers/pci/bus.c | 188 +++++++++++++++++++++++++++++-----
drivers/pci/host-bridge.c | 19 ++++
drivers/pci/hotplug/acpiphp_glue.c | 13 ++-
drivers/pci/hotplug/cpcihp_generic.c | 8 +-
drivers/pci/hotplug/pciehp_pci.c | 15 +++
drivers/pci/hotplug/sgi_hotplug.c | 2 +
drivers/pci/iov.c | 11 +-
drivers/pci/pci-sysfs.c | 37 ++++---
drivers/pci/probe.c | 83 +++++++++++----
drivers/pci/remove.c | 176 +++++++++++++++++--------------
drivers/pci/search.c | 53 ++++++++--
drivers/pci/setup-bus.c | 65 +++++++++---
drivers/pci/xen-pcifront.c | 10 +-
drivers/platform/x86/asus-wmi.c | 23 ++++-
drivers/platform/x86/eeepc-laptop.c | 20 ++--
include/linux/pci.h | 56 +++++++++-
26 files changed, 629 insertions(+), 224 deletions(-)
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
