Re: [PATCH v2 05/13] PCI/ASPM: Add pci_enable_link_state()

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

[-- Attachment #1: Type: text/plain, Size: 4122 bytes --]

On Fri, 13 Oct 2023, Bjorn Helgaas wrote:

> On Thu, Oct 12, 2023 at 03:53:39PM +0300, Ilpo Järvinen wrote:
> > On Wed, 11 Oct 2023, Bjorn Helgaas wrote:
> > > On Mon, Sep 18, 2023 at 04:10:55PM +0300, Ilpo Järvinen wrote:
> > > > pci_disable_link_state() lacks a symmetric pair. Some drivers want to
> > > > disable ASPM during certain phases of their operation but then
> > > > re-enable it later on. If pci_disable_link_state() is made for the
> > > > device, there is currently no way to re-enable the states that were
> > > > disabled.
> > > 
> > > pci_disable_link_state() gives drivers a way to disable specified ASPM
> > > states using a bitmask (PCIE_LINK_STATE_L0S, PCIE_LINK_STATE_L1,
> > > PCIE_LINK_STATE_L1_1, etc), but IIUC the driver can't tell exactly
> > > what changed and can't directly restore the original state, e.g.,
> > > 
> > >   - PCIE_LINK_STATE_L1 enabled initially
> > >   - driver calls pci_disable_link_state(PCIE_LINK_STATE_L0S)
> > >   - driver calls pci_enable_link_state(PCIE_LINK_STATE_L0S)
> > >   - PCIE_LINK_STATE_L0S and PCIE_LINK_STATE_L1 are enabled now
> > > 
> > > Now PCIE_LINK_STATE_L0S is enabled even though it was not initially
> > > enabled.  Maybe that's what we want; I dunno.
> > > 
> > > pci_disable_link_state() currently returns success/failure, but only
> > > r8169 and mt76 even check, and only rtl_init_one() (r8169) has a
> > > non-trivial reason, so it's conceivable that it could return a bitmask
> > > instead.
> > 
> > It's great that you suggested this since it's actually what also I've been 
> > started to think should be done instead of this straightforward approach
> > I used in V2. 
> > 
> > That is, don't have the drivers to get anything directly from LNKCTL
> > but they should get everything through the API provided by the 
> > disable/enable calls which makes it easy for the driver to pass the same
> > value back into the enable call.
> > 
> > > > Add pci_enable_link_state() to remove ASPM states from the state
> > > > disable mask.
> > > > 
> > > > Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@linux.intel.com>
> > > > ---
> > > >  drivers/pci/pcie/aspm.c | 42 +++++++++++++++++++++++++++++++++++++++++
> > > >  include/linux/pci.h     |  2 ++
> > > >  2 files changed, 44 insertions(+)
> > > > 
> > > > diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> > > > index 91dc95aca90f..f45d18d47c20 100644
> > > > --- a/drivers/pci/pcie/aspm.c
> > > > +++ b/drivers/pci/pcie/aspm.c
> > > > @@ -1117,6 +1117,48 @@ int pci_disable_link_state(struct pci_dev *pdev, int state)
> > > >  }
> > > >  EXPORT_SYMBOL(pci_disable_link_state);
> > > >  
> > > > +/**
> > > > + * pci_enable_link_state - Re-enable device's link state
> > > > + * @pdev: PCI device
> > > > + * @state: ASPM link states to re-enable
> > > > + *
> > > > + * Enable device's link state that were previously disable so the link is
> > > 
> > > "state[s] that were previously disable[d]" alludes to the use case you
> > > have in mind, but I don't think it describes how this function
> > > actually works.  This function just makes it possible to enable the
> > > specified states.  The @state parameter may have nothing to do with
> > > any previously disabled states.
> > 
> > Yes, it's what I've been thinking between the lines. But I see your point 
> > that this API didn't make it easy/obvious as is.
> > 
> > Would you want me to enforce it too besides altering the API such that the 
> > states are actually returned from disable call? (I don't personally find
> > that necessary as long as the API pair itself makes it obvious what the 
> > driver is expect to pass there.)
> 
> This was just a comment about the doc not matching the function
> behavior.
> 
> I think we have to support pci_enable_link_state() even if the driver
> hasn't previously called pci_disable_link_state(), so drivers have to
> be able to specify the pci_enable_link_state() @state from scratch.
> 
> Does that answer the enforcement question?

Yes.

-- 
 i.

> I don't think we can
> really enforce anything other than that @state specifies valid ASPM
> states.
> 
> Bjorn
>