Re: [PATCH v8 4/4] PCI/ASPM: Fix L1.2 parameters when enable link state

["Ilpo Järvinen" <ilpo.jarvinen@linux.intel.com>]

[-- Attachment #1: Type: text/plain, Size: 9617 bytes --]

On Wed, 7 Aug 2024, David E. Box wrote:

> On Wed, 2024-08-07 at 14:18 +0300, Ilpo Järvinen wrote:
> > On Wed, 7 Aug 2024, Jian-Hong Pan wrote:
> > 
> > > David E. Box <david.e.box@linux.intel.com> 於 2024年8月6日 週二 上午4:26寫道：
> > > > 
> > > > Hi Jian-Hong,
> > > > 
> > > > On Fri, 2024-08-02 at 16:24 +0800, Jian-Hong Pan wrote:
> > > > > Jian-Hong Pan <jhp@endlessos.org> 於 2024年7月19日 週五 下午4:04寫道：
> > > > > > 
> > > > > > Currently, when enable link's L1.2 features with
> > > > > > __pci_enable_link_state(),
> > > > > > it configs the link directly without ensuring related L1.2 parameters,
> > > > > > such
> > > > > > as T_POWER_ON, Common_Mode_Restore_Time, and LTR_L1.2_THRESHOLD have
> > > > > > been
> > > > > > programmed.
> > > > > > 
> > > > > > This leads the link's L1.2 between PCIe Root Port and child device
> > > > > > gets
> > > > > > wrong configs when a caller tries to enabled it.
> > > > > > 
> > > > > > Here is a failed example on ASUS B1400CEAE with enabled VMD:
> > > > > > 
> > > > > > 10000:e0:06.0 PCI bridge: Intel Corporation 11th Gen Core Processor
> > > > > > PCIe
> > > > > > Controller (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=45us LTR1.2_Threshold=101376ns
> > > > > >         L1SubCtl2: T_PwrOn=50us
> > > > > > 
> > > > > > 10000:e1:00.0 Non-Volatile memory controller: Sandisk Corp WD Blue
> > > > > > SN550
> > > > > > NVMe SSD (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=0ns
> > > > > >         L1SubCtl2: T_PwrOn=10us
> > > > > > 
> > > > > > According to "PCIe r6.0, sec 5.5.4", before enabling ASPM L1.2 on the
> > > > > > PCIe
> > > > > > Root Port and the child NVMe, they should be programmed with the same
> > > > > > LTR1.2_Threshold value. However, they have different values in this
> > > > > > case.
> > > > > > 
> > > > > > Invoke aspm_calc_l12_info() to program the L1.2 parameters properly
> > > > > > before
> > > > > > enable L1.2 bits of L1 PM Substates Control Register in
> > > > > > __pci_enable_link_state().
> > > > > > 
> > > > > > Link: https://bugzilla.kernel.org/show_bug.cgi?id=218394
> > > > > > Signed-off-by: Jian-Hong Pan <jhp@endlessos.org>
> > > > > > ---
> > > > > > v2:
> > > > > > - Prepare the PCIe LTR parameters before enable L1 Substates
> > > > > > 
> > > > > > v3:
> > > > > > - Only enable supported features for the L1 Substates part
> > > > > > 
> > > > > > v4:
> > > > > > - Focus on fixing L1.2 parameters, instead of re-initializing whole
> > > > > > L1SS
> > > > > > 
> > > > > > v5:
> > > > > > - Fix typo and commit message
> > > > > > - Split introducing aspm_get_l1ss_cap() to "PCI/ASPM: Introduce
> > > > > >   aspm_get_l1ss_cap()"
> > > > > > 
> > > > > > v6:
> > > > > > - Skipped
> > > > > > 
> > > > > > v7:
> > > > > > - Pick back and rebase on the new version kernel
> > > > > > - Drop the link state flag check. And, always config link state's
> > > > > > timing
> > > > > >   parameters
> > > > > > 
> > > > > > v8:
> > > > > > - Because pcie_aspm_get_link() might return the link as NULL, move
> > > > > >   getting the link's parent and child devices after check the link is
> > > > > >   not NULL. This avoids NULL memory access.
> > > > > > 
> > > > > >  drivers/pci/pcie/aspm.c | 15 +++++++++++++++
> > > > > >  1 file changed, 15 insertions(+)
> > > > > > 
> > > > > > diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> > > > > > index 5db1044c9895..55ff1d26fcea 100644
> > > > > > --- a/drivers/pci/pcie/aspm.c
> > > > > > +++ b/drivers/pci/pcie/aspm.c
> > > > > > @@ -1411,9 +1411,15 @@ EXPORT_SYMBOL(pci_disable_link_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);
> > > > > > +       u32 parent_l1ss_cap, child_l1ss_cap;
> > > > > > +       struct pci_dev *parent, *child;
> > > > > > 
> > > > > >         if (!link)
> > > > > >                 return -EINVAL;
> > > > > > +
> > > > > > +       parent = link->pdev;
> > > > > > +       child = link->downstream;
> > > > > > +
> > > > > >         /*
> > > > > >          * A driver requested that ASPM be enabled on this device, but
> > > > > >          * if we don't have permission to manage ASPM (e.g., on ACPI
> > > > > > @@ -1428,6 +1434,15 @@ static int __pci_enable_link_state(struct
> > > > > > pci_dev
> > > > > > *pdev, int state, bool locked)
> > > > > >         if (!locked)
> > > > > >                 down_read(&pci_bus_sem);
> > > > > >         mutex_lock(&aspm_lock);
> > > > > > +       /*
> > > > > > +        * Ensure L1.2 parameters: Common_Mode_Restore_Times,
> > > > > > T_POWER_ON and
> > > > > > +        * LTR_L1.2_THRESHOLD are programmed properly before enable
> > > > > > bits for
> > > > > > +        * L1.2, per PCIe r6.0, sec 5.5.4.
> > > > > > +        */
> > > > > > +       parent_l1ss_cap = aspm_get_l1ss_cap(parent);
> > > > > > +       child_l1ss_cap = aspm_get_l1ss_cap(child);
> > > > > > +       aspm_calc_l12_info(link, parent_l1ss_cap, child_l1ss_cap);
> > > > 
> > > > I still don't think this is the place to recalculate the L1.2 parameters
> > > > especially when know the calculation was done but was cleared by
> > > > pci_bus_reset(). Can't we just do a pci_save/restore_state() before/after
> > > > pci_bus_reset() in vmd.c?
> > > 
> > > I have not thought pci_save/restore_state() around pci_bus_reset()
> > > before.  It is an interesting direction.
> > > 
> > > So, I prepare modification below for test.  Include "[PATCH v8 1/4]
> > > PCI: vmd: Set PCI devices to D0 before enable PCI PM's L1 substates",
> > > too.  Then, both the PCIe bridge and the PCIe device have the same
> > > LTR_L1.2_THRESHOLD 101376ns as expected.
> > > 
> > > diff --git a/drivers/pci/controller/vmd.c b/drivers/pci/controller/vmd.c
> > > index bbf4a47e7b31..6b8dd4f30127 100644
> > > --- a/drivers/pci/controller/vmd.c
> > > +++ b/drivers/pci/controller/vmd.c
> > > @@ -727,6 +727,18 @@ static void vmd_copy_host_bridge_flags(struct
> > > pci_host_bridge *root_bridge,
> > >         vmd_bridge->native_dpc = root_bridge->native_dpc;
> > >  }
> > > 
> > > +static int vmd_pci_save_state(struct pci_dev *pdev, void *userdata)
> > > +{
> > > +       pci_save_state(pdev);
> > > +       return 0;
> > > +}
> > > +
> > > +static int vmd_pci_restore_state(struct pci_dev *pdev, void *userdata)
> > > +{
> > > +       pci_restore_state(pdev);
> > > +       return 0;
> > > +}
> > > +
> > >  /*
> > >   * Enable ASPM and LTR settings on devices that aren't configured by BIOS.
> > >   */
> > > @@ -927,6 +939,7 @@ static int vmd_enable_domain(struct vmd_dev *vmd,
> > > unsigned long features)
> > >         pci_scan_child_bus(vmd->bus);
> > >         vmd_domain_reset(vmd);
> > > 
> > > +       pci_walk_bus(vmd->bus, vmd_pci_save_state, NULL);
> > >         /* When Intel VMD is enabled, the OS does not discover the Root
> > > Ports
> > >          * owned by Intel VMD within the MMCFG space. pci_reset_bus()
> > > applies
> > >          * a reset to the parent of the PCI device supplied as argument.
> > > This
> > > @@ -945,6 +958,7 @@ static int vmd_enable_domain(struct vmd_dev *vmd,
> > > unsigned long features)
> > >                         break;
> > >                 }
> > >         }
> > > +       pci_walk_bus(vmd->bus, vmd_pci_restore_state, NULL);
> > 
> > Why not call pci_reset_bus() (or __pci_reset_bus()) then in 
> > vmd_enable_domain() which preserves state unlike pci_reset_bus()?
> > 
> > (Don't tell me naming of these functions is a horrible mess. :-/)
> 
> Hmm. So this *is* calling pci_reset_bus().

Yeah, I managed to get confused by the names myself, I somehow 
ended up thinking it calls pci_bus_reset() which is not correct...

> L1.2 configuration has specific
> ordering requirements for changes to parent & child devices. Could be why it's
> not getting restored properly.

Indeed, it has to be something else since the patch above doesn't even 
restore anything because dev->state_saved should get set to false by the 
first pci_restore_state() called from 
__pci_reset_bus() -> pci_bus_restore_locked() -> pci_dev_restore(), I 
think!?

-- 
 i.