[PATCH 00/17] PCI: Improve LNKCTL & LNKCTL2 concurrency control

From: Ilpo Järvinen
Date: Thu May 11 2023 - 09:15:14 EST


LNKCTL register is written by a number of things in the kernel as RMW
operations without any concurrency control. This could in the unlucky
case lead to losing one of the updates. One of the most obvious path
which can race with most of the other LNKCTL RMW operations seems to be
ASPM policy sysfs write which triggers LNKCTL update.

Add pcie_capability_clear_and_set_word_locked() that uses a per device
spinlock to protect the RMW operations. Introduce helpers for updating
LNKCTL and LNKCTL2 that internally do
pcie_capability_clear_and_set_word_locked(). Convert RMW to use the new
helpers.

These could be mostly marked with Fixes tags but I've not spent the
effort to find those out for each and every patch until this series has
seen some discussion. I certainly will try to find the Fixes tags if
asked to.

There could be a few LNKCTL RMW that are so early into the init that
they would be safe but I was not able to convince myself so I've
included them (namely, some ASPM init paths and hp link enable). Even
if that is the case, it seems safer to use an access pattern with these
registers that is safe even if there would be a few cases where locking
would not be stricly necessary.

As for LNKCTL2, I think all current users are safe but these came up as
a part of PCIe bandwidth control work (for thermal reasons) that will
be adding a writer for LNKCTL2. I'll send PCI BW ctrl patches
separately later as there's plenty of patches in this series already.
If most of this series is deemed worthy of Fixes tags, I could separate
those few LNKCTL2 changes into own patches.

The series is based on top of the "PCI/ASPM: Handle link retraining
race" patch I sent earlier but is not yet applied.

Ilpo Järvinen (17):
PCI: Add concurrency safe clear_and_set variants for LNKCTL{,2}
PCI: pciehp: Protect LNKCTL changes
PCI/ASPM: Use pcie_lnkctl_clear_and_set()
drm/amdgpu: Use pcie_lnkctl{,2}_clear_and_set() for changing
LNKCTL{,2}
drm/radeon: Use pcie_lnkctl{,2}_clear_and_set() for changing
LNKCTL{,2}
IB/hfi1: Use pcie_lnkctl{,2}_clear_and_set() for changing LNKCTL{,2}
e1000e: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
net/mlx5: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
wifi: ath9k: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
mt76: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
Bluetooth: hci_bcm4377: Use pcie_lnkctl_clear_and_set() for changing
LNKCTL
misc: rtsx: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
net/tg3: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
r8169: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
wifi: ath11k: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
wifi: ath12k: Use pcie_lnkctl_clear_and_set() for changing LNKCTL
wifi: ath10k: Use pcie_lnkctl_clear_and_set() for changing LNKCTL

drivers/bluetooth/hci_bcm4377.c | 3 +-
drivers/gpu/drm/amd/amdgpu/cik.c | 72 +++++-------------
drivers/gpu/drm/amd/amdgpu/si.c | 74 ++++++-------------
drivers/gpu/drm/radeon/cik.c | 71 +++++-------------
drivers/gpu/drm/radeon/si.c | 72 +++++-------------
drivers/infiniband/hw/hfi1/aspm.c | 16 ++--
drivers/infiniband/hw/hfi1/pcie.c | 28 ++-----
drivers/misc/cardreader/rts5228.c | 6 +-
drivers/misc/cardreader/rts5261.c | 6 +-
drivers/misc/cardreader/rtsx_pcr.c | 8 +-
drivers/net/ethernet/broadcom/tg3.c | 14 ++--
drivers/net/ethernet/intel/e1000e/netdev.c | 6 +-
.../ethernet/mellanox/mlx5/core/fw_reset.c | 9 +--
drivers/net/ethernet/realtek/r8169_main.c | 6 +-
drivers/net/wireless/ath/ath10k/pci.c | 8 +-
drivers/net/wireless/ath/ath11k/pci.c | 8 +-
drivers/net/wireless/ath/ath12k/pci.c | 8 +-
drivers/net/wireless/ath/ath9k/pci.c | 9 ++-
drivers/net/wireless/mediatek/mt76/pci.c | 5 +-
drivers/pci/access.c | 14 ++++
drivers/pci/hotplug/pciehp_hpc.c | 11 +--
drivers/pci/pcie/aspm.c | 48 ++++--------
drivers/pci/probe.c | 1 +
include/linux/pci.h | 17 +++++
24 files changed, 183 insertions(+), 337 deletions(-)

--
2.30.2