On Fri, Dec 8, 2023 at 4:47 AM Bjorn Helgaas helgaas@kernel.org wrote:
[+cc Kai-Heng]
On Tue, Nov 28, 2023 at 09:15:07AM +0100, Johan Hovold wrote:
Add a helper for enabling link states that can be used in contexts where a pci_bus_sem read lock is already held (e.g. from pci_walk_bus()).
This helper will be used to fix a couple of potential deadlocks where the current helper is called with the lock already held, hence the CC stable tag.
Fixes: f492edb40b54 ("PCI: vmd: Add quirk to configure PCIe ASPM and LTR") Cc: stable@vger.kernel.org # 6.3 Cc: Michael Bottini michael.a.bottini@linux.intel.com Cc: David E. Box david.e.box@linux.intel.com Reviewed-by: Manivannan Sadhasivam manivannan.sadhasivam@linaro.org Signed-off-by: Johan Hovold johan+linaro@kernel.org
drivers/pci/pcie/aspm.c | 53 +++++++++++++++++++++++++++++++---------- include/linux/pci.h | 3 +++ 2 files changed, 43 insertions(+), 13 deletions(-)
diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c index 50b04ae5c394..5eb462772354 100644 --- a/drivers/pci/pcie/aspm.c +++ b/drivers/pci/pcie/aspm.c @@ -1109,17 +1109,7 @@ int pci_disable_link_state(struct pci_dev *pdev, int state) } EXPORT_SYMBOL(pci_disable_link_state);
-/**
- pci_enable_link_state - Clear and set the default device link state so that
- the link may be allowed to enter the specified states. Note that if the
- BIOS didn't grant ASPM control to the OS, this does nothing because we can't
- touch the LNKCTL register. Also note that this does not enable states
- disabled by pci_disable_link_state(). Return 0 or a negative errno.
- @pdev: PCI device
- @state: Mask of ASPM link states to enable
- */
-int pci_enable_link_state(struct pci_dev *pdev, int state) +static int __pci_enable_link_state(struct pci_dev *pdev, int state, bool locked) { struct pcie_link_state *link = pcie_aspm_get_link(pdev);
@@ -1136,7 +1126,8 @@ int pci_enable_link_state(struct pci_dev *pdev, int state) return -EPERM; }
down_read(&pci_bus_sem);
if (!locked)down_read(&pci_bus_sem); mutex_lock(&aspm_lock); link->aspm_default = 0; if (state & PCIE_LINK_STATE_L0S)@@ -1157,12 +1148,48 @@ int pci_enable_link_state(struct pci_dev *pdev, int state) link->clkpm_default = (state & PCIE_LINK_STATE_CLKPM) ? 1 : 0; pcie_set_clkpm(link, policy_to_clkpm_state(link)); mutex_unlock(&aspm_lock);
up_read(&pci_bus_sem);
if (!locked)up_read(&pci_bus_sem); return 0;}
+/**
- pci_enable_link_state - Clear and set the default device link state so that
- the link may be allowed to enter the specified states. Note that if the
- BIOS didn't grant ASPM control to the OS, this does nothing because we can't
- touch the LNKCTL register. Also note that this does not enable states
- disabled by pci_disable_link_state(). Return 0 or a negative errno.
- @pdev: PCI device
- @state: Mask of ASPM link states to enable
- */
+int pci_enable_link_state(struct pci_dev *pdev, int state) +{
return __pci_enable_link_state(pdev, state, false);+} EXPORT_SYMBOL(pci_enable_link_state);
As far as I can see, we end up with pci_enable_link_state() defined but never called and pci_enable_link_state_locked() being called only by pcie-qcom.c and vmd.c.
Can we just rename pci_enable_link_state() to pci_enable_link_state_locked() and assert that pci_bus_sem is held, so we don't end up with a function that's never used?
I hope we can obsolete this whole idea someday. Using pci_walk_bus() in qcom and vmd to enable ASPM is an ugly hack to work around this weird idea that "the OS isn't allowed to enable more ASPM states than the BIOS did because the BIOS might have left ASPM disabled because it knows about hardware issues." More history at https://lore.kernel.org/linux-pci/20230615070421.1704133-1-kai.heng.feng@can...
I think we need to get to a point where Linux enables all supported ASPM features by default. If we really think x86 BIOS assumes an implicit contract that the OS will never enable ASPM more aggressively, we might need some kind of arch quirk for that.
The reality is that PC ODM toggles ASPM to workaround hardware defects, assuming that OS will honor what's set by the BIOS. If ASPM gets enabled for all devices, many devices will break.
Kai-Heng
If we can get there, the qcom use of pci_enable_link_state() could go away, and the vmd use could be replaced by some kind of "if device is below VMD, get rid of the legacy x86 ASPM assumption" quirk.
Bjorn