On Thu, 26 Sep 2024, Ilpo Järvinen wrote: > On Thu, 26 Sep 2024, Jian-Hong Pan wrote: > > > David E. Box <david.e.box@xxxxxxxxxxxxxxx> 於 2024年9月26日 週四 上午10:51寫道： > > > > > > Hi Jian-Hong, > > > > > > On Tue, 2024-09-24 at 15:29 +0800, Jian-Hong Pan wrote: > > > > PCI devices' parameters on the VMD bus have been programmed properly > > > > originally. But, cleared after pci_reset_bus() and have not been restored > > > > correctly. This leads the link's L1.2 between PCIe Root Port and child > > > > device gets wrong configs. > > > > > > > > Here is a failed example on ASUS B1400CEAE with enabled VMD. Both PCIe > > > > bridge and NVMe device should have the same LTR1.2_Threshold value. > > > > However, they are configured as different values in this case: > > > > > > > > 10000:e0:06.0 PCI bridge [0604]: Intel Corporation 11th Gen Core Processor > > > > PCIe Controller [8086:9a09] (rev 01) (prog-if 00 [Normal decode]) > > > > ... > > > > Capabilities: [200 v1] L1 PM Substates > > > > L1SubCap: PCI-PM_L1.2+ PCI-PM_L1.1+ ASPM_L1.2+ ASPM_L1.1+ L1_PM_Substates+ > > > > PortCommonModeRestoreTime=45us PortTPowerOnTime=50us > > > > L1SubCtl1: PCI-PM_L1.2- PCI-PM_L1.1- ASPM_L1.2+ ASPM_L1.1- > > > > T_CommonMode=0us LTR1.2_Threshold=0ns > > > > L1SubCtl2: T_PwrOn=0us > > > > > > > > 10000:e1:00.0 Non-Volatile memory controller [0108]: Sandisk Corp WD Blue > > > > SN550 NVMe SSD [15b7:5009] (rev 01) (prog-if 02 [NVM Express]) > > > > ... > > > > Capabilities: [900 v1] L1 PM Substates > > > > L1SubCap: PCI-PM_L1.2+ PCI-PM_L1.1- ASPM_L1.2+ ASPM_L1.1- L1_PM_Substates+ > > > > PortCommonModeRestoreTime=32us PortTPowerOnTime=10us > > > > L1SubCtl1: PCI-PM_L1.2- PCI-PM_L1.1- ASPM_L1.2+ ASPM_L1.1- > > > > T_CommonMode=0us LTR1.2_Threshold=101376ns > > > > L1SubCtl2: T_PwrOn=50us > > > > > > > > Here is VMD mapped PCI device tree: > > > > > > > > -+-[0000:00]-+-00.0 Intel Corporation Device 9a04 > > > > | ... > > > > \-[10000:e0]-+-06.0-[e1]----00.0 Sandisk Corp WD Blue SN550 NVMe SSD > > > > \-17.0 Intel Corporation Tiger Lake-LP SATA Controller > > > > > > > > When pci_reset_bus() resets the bus [e1] of the NVMe, it only saves and > > > > restores NVMe's state before and after reset. Because bus [e1] has only one > > > > device: 10000:e1:00.0 NVMe. The PCIe bridge is missed. However, when it > > > > restores the NVMe's state, it also restores the ASPM L1SS between the PCIe > > > > bridge and the NVMe by pci_restore_aspm_l1ss_state(). The NVMe's L1SS is > > > > restored correctly. But, the PCIe bridge's L1SS is restored with the wrong > > > > value 0x0. Becuase, the PCIe bridge's state is not saved before reset. > > > > That is why pci_restore_aspm_l1ss_state() gets the parent device (PCIe > > > > bridge)'s saved state capability data and restores L1SS with value 0. > > > > > > > > So, save the PCIe bridge's state before pci_reset_bus(). Then, restore the > > > > state after pci_reset_bus(). The restoring state also consumes the saving > > > > state. > > > > > > > > Link: https://www.spinics.net/lists/linux-pci/msg159267.html > > > > Signed-off-by: Jian-Hong Pan <jhp@xxxxxxxxxxxxx> > > > > --- > > > > v9: > > > > - Drop the v8 fix about drivers/pci/pcie/aspm.c. Use this in VMD instead. > > > > > > > > drivers/pci/controller/vmd.c | 7 +++++++ > > > > 1 file changed, 7 insertions(+) > > > > > > > > diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c > > > > index 11870d1fc818..2820005165b4 100644 > > > > --- a/drivers/pci/controller/vmd.c > > > > +++ b/drivers/pci/controller/vmd.c > > > > @@ -938,9 +938,16 @@ static int vmd_enable_domain(struct vmd_dev *vmd, > > > > unsigned long features) > > > > if (!list_empty(&child->devices)) { > > > > dev = list_first_entry(&child->devices, > > > > struct pci_dev, bus_list); > > > > > > Newline here > > > > + /* To avoid pci_reset_bus restore wrong ASPM L1SS > > > > + * configuration due to missed saving parent device's > > > > + * states, save & restore the parent device's states > > > > + * as well. > > > > + */ > > > > > > No text on first line of comment. > > > > Oops! Thanks > > > > > /* > > > * To avoid pci_reset_bus restore wrong ASPM L1SS > > > * ... > > > */ > > > > > > > + pci_save_state(dev->bus->self); > > > > ret = pci_reset_bus(dev); > > > > if (ret) > > > > pci_warn(dev, "can't reset device: %d\n", > > > > ret); > > > > + pci_restore_state(dev->bus->self); > > > > > > I think Ilpo's point was that pci_save_aspm_l1ss_state() and > > > pci_restore_aspm_l1ss_state() should be more symmetric. If > > > pci_save_aspm_l1ss_state() is changed to also handle the state for the device > > > and its parent in the same call, then no change is needed here. But that would > > > only handle L1SS settings and not anything else that might need to be restored > > > after the bus reset. So I'm okay with it either way. > > Why would something else need to be restored? The spec explicitly says the > bus reset should not cause config changes on the parent other than > to status bits. > > Based on what is seen so far, the problem here is not the reset itself > breaking parent's config but ASPM code restoring values from state beyond > what it had saved and thus it programs pseudogarbage into the L1SS > settings. > > > Yes, that made me think the whole day before I sent out this patch. I > > do not know what is the best reset bus procedure, especially how other > > drivers reset the bus. > > > > If trace the code path: > > pci_reset_bus() > > __pci_reset_bus() > > pci_bus_reset() > > pci_bus_save_and_disable_locked() > > > > pci_bus_save_and_disable_locked()'s comment shows "Save and disable > > devices from the top of the tree down while holding the @dev mutex > > lock for the entire tree". I think that means the PCI(e) bridge should > > save state first, then the following bus' devices. However, VMD resets > > the child device (10000:e1:00.0 NVMe)'s bus first and only saves the > > child device (10000:e1:00.0 NVMe)'s state. It does not visit the tree > > from the top to down. So, it misses the PCIe bridge. Therefore, I > > make it save & restore its parent (10000:e0:06.0 PCIe bridge)'s state > > as compensation. > > The problem with your fix is it only takes care of part of the cases (i.e. > VMD). But there are other callers of pci_reset_bus() which would also > restore incorrect L1SS settings, no? > > I'd suggest making the L1SS code symmetric on this, even if it means > saving the register value twice when walking the tree downwards (it seems > harmless to write the same value twice). Perhaps parent->state_saved == true could even be used to avoid doing it twice/unnecessarily if the parent is already saved. -- i.
