mbox series

[RFC,0/8] pci: rescan/remove locking rework

Message ID 20240722151936.1452299-1-kbusch@meta.com (mailing list archive)
Headers show
Series pci: rescan/remove locking rework | expand

Message

Keith Busch July 22, 2024, 3:19 p.m. UTC
From: Keith Busch <kbusch@kernel.org>

This patch set targets a subset of pci bus scanning and removals that
were shown to be problematic with deep pci topologies that support
native hotplug. I've tried to capture the common pci components, but
there are definitely many subsystems accessing the topology in their own
way, many of which I can't possibly test, and I have not tried to
convert every user to this new locking scheme. However, if I did this
correctly, they should be no worse off than today!

The earlier patches are just cleanups and/or making it a little easier
to change the locking schemes. The real stuff happens from patches 7 and
8.

I've run this with lockdep enabled, tested concurrent hotplug events on
various x86 platforms with layers of pci switches. That said, as
mentioned earlier, there are many paths to here that I haven't been able
to test, so the final patch might be considered experimental.

Manual concurrent concurrent device removals is still broken. I've sent
a different patch for that here, but it does not sound like it is
acceptable:

  https://lore.kernel.org/linux-pci/20240719185513.3376321-1-kbusch@meta.com/T/#u

Keith Busch (8):
  pci: make pci_stop_dev concurrent safe
  pci: make pci_destroy_dev concurrent safe
  pci: move the walk bus lock to where its needed
  pci: walk bus recursively
  pci: unexport pci_walk_bus_locked
  pci: add helpers for stop and remove bus
  pci: reference count subordinate
  pci: use finer grain locking for bus protection

 drivers/pci/bus.c                | 70 ++++++++++++-------------
 drivers/pci/hotplug/pciehp_pci.c | 21 +++++---
 drivers/pci/pci-sysfs.c          |  2 +
 drivers/pci/pci.c                | 31 ++++++++---
 drivers/pci/pci.h                | 16 +++++-
 drivers/pci/pcie/aspm.c          |  7 ++-
 drivers/pci/probe.c              | 16 +++++-
 drivers/pci/remove.c             | 90 +++++++++++++++++++-------------
 include/linux/pci.h              | 11 ++++
 9 files changed, 175 insertions(+), 89 deletions(-)

Comments

Keith Busch Aug. 7, 2024, 3:40 p.m. UTC | #1
On Mon, Jul 22, 2024 at 08:19:28AM -0700, Keith Busch wrote:
> From: Keith Busch <kbusch@kernel.org>
> 
> This patch set targets a subset of pci bus scanning and removals that
> were shown to be problematic with deep pci topologies that support
> native hotplug. I've tried to capture the common pci components, but
> there are definitely many subsystems accessing the topology in their own
> way, many of which I can't possibly test, and I have not tried to
> convert every user to this new locking scheme. However, if I did this
> correctly, they should be no worse off than today!
> 
> The earlier patches are just cleanups and/or making it a little easier
> to change the locking schemes. The real stuff happens from patches 7 and
> 8.
> 
> I've run this with lockdep enabled, tested concurrent hotplug events on
> various x86 platforms with layers of pci switches. That said, as
> mentioned earlier, there are many paths to here that I haven't been able
> to test, so the final patch might be considered experimental.

Any thoughts on this new scheme? I know I sent this during the merge
window, so maybe bad timing on my part. Changing the locks like this is
kind of scary too, so if that's the case, could we reconsider the
band-aid patch from before?

  https://lore.kernel.org/linux-pci/20240612181625.3604512-2-kbusch@meta.com/