Re: [PATCH v2] PCI: dwc: Wait for link up only if link is started

[Ajay Agarwal]

On Wed, Apr 05, 2023 at 12:23:47PM -0700, William McVicker wrote:
> On 04/05/2023, William McVicker wrote:
> > On 04/05/2023, Bjorn Helgaas wrote:
> > > On Wed, Apr 05, 2023 at 03:24:36PM +0200, Lorenzo Pieralisi wrote:
> > > > On Thu, Mar 16, 2023 at 06:05:02PM -0500, Sajid Dalvi wrote:
> > > > > On Tue, Feb 28, 2023 at 10:36 PM Sajid Dalvi <sdalvi@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > Thanks for your review Jingoo.
> > > > > > Sajid
> > > > > >
> > > > > > On Tue, Feb 28, 2023 at 4:04 PM Han Jingoo <jingoohan1@xxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Mon, Feb 27, 2023, Sajid Dalvi <sdalvi@xxxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > > In dw_pcie_host_init() regardless of whether the link has been started
> > > > > > > > or not, the code waits for the link to come up. Even in cases where
> > > > > > > > start_link() is not defined the code ends up spinning in a loop for 1
> > > > > > > > second. Since in some systems dw_pcie_host_init() gets called during
> > > > > > > > probe, this one second loop for each pcie interface instance ends up
> > > > > > > > extending the boot time.
> > > > > > > >
> > > > > > > > Call trace when start_link() is not defined:
> > > > > > > > dw_pcie_wait_for_link << spins in a loop for 1 second
> > > > > > > > dw_pcie_host_init
> > > > > > > >
> > > > > > > > Signed-off-by: Sajid Dalvi <sdalvi@xxxxxxxxxx>
> > > > > > >
> > > > > > > (CC'ed Krzysztof Kozlowski)
> > > > > > >
> > > > > > > Acked-by: Jingoo Han <jingoohan1@xxxxxxxxx>
> > > > > > >
> > > > > > > It looks good to me. I also checked the previous thread.
> > > > > > > I agree with Krzysztof's opinion that we should include
> > > > > > > only hardware-related features into DT.
> > > > > > > Thank you.
> > > > > > >
> > > > > > > Best regards,
> > > > > > > Jingoo Han
> > > > > > >
> > > > > > > > ---
> > > > > > > >  drivers/pci/controller/dwc/pcie-designware-host.c | 6 +++---
> > > > > > > >  1 file changed, 3 insertions(+), 3 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/drivers/pci/controller/dwc/pcie-designware-host.c b/drivers/pci/controller/dwc/pcie-designware-host.c
> > > > > > > > index 9952057c8819..9709f69f173e 100644
> > > > > > > > --- a/drivers/pci/controller/dwc/pcie-designware-host.c
> > > > > > > > +++ b/drivers/pci/controller/dwc/pcie-designware-host.c
> > > > > > > > @@ -489,10 +489,10 @@ int dw_pcie_host_init(struct dw_pcie_rp *pp)
> > > > > > > >                 ret = dw_pcie_start_link(pci);
> > > > > > > >                 if (ret)
> > > > > > > >                         goto err_remove_edma;
> > > > > > > > -       }
> > > > > > > >
> > > > > > > > -       /* Ignore errors, the link may come up later */
> > > > > > > > -       dw_pcie_wait_for_link(pci);
> > > > > > > > +               /* Ignore errors, the link may come up later */
> > > > > > > > +               dw_pcie_wait_for_link(pci);
> > > > > > > > +       }
> > > > > > > >
> > > > > > > >         bridge->sysdata = pp;
> > > > > > > >
> > > > > > > > --
> > > > > > > > 2.39.2.722.g9855ee24e9-goog
> > > > > > > >
> > > > > 
> > > > > @bhelgaas Can this be picked up in your tree:
> > > > >  https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/
> > > > 
> > > > This patch seems fine to me. The question I have though is why the
> > > > *current* code is written the way it is. Perhaps it is just the way
> > > > it is, I wonder whether this change can trigger a regression though.
> > > 
> > > The new code will look basically like this:
> > > 
> > >   if (!dw_pcie_link_up(pci)) {
> > >     dw_pcie_start_link(pci);
> > >     dw_pcie_wait_for_link(pci);
> > >   }
> > > 
> > > If the link is already up by the time we get here, this change means
> > > we won't get this message emitted by dw_pcie_wait_for_link():
> > > 
> > >   dev_info(pci->dev, "PCIe Gen.%u x%u link up\n", ...)
> > > 
> > > I don't know how important that is, but I bet somebody cares about it.
> > > 
> > > From the commit log, I expected the patch to do something based on
> > > whether ->start_link() was defined, but there really isn't a direct
> > > connection, so maybe the log could be refined.
> > > 
> > > Bjorn
> > > 
> > > -- 
> > > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@xxxxxxxxxxx.
> > > 
> > 
> > After taking a deeper dive into this patch, I found that [1] changes the
> > original intent which was to skip the call to dw_pcie_wait_for_link()
> > when pci->ops->start_link is NULL. I talked to Sajid offline and he
> > agreed we should put back the start_link NULL check. The updated patch
> > should look like this:
> > 
> >   if (!dw_pcie_link_up(pci) && pci->ops && pci->ops->start_link) {
> >     ret = dw_pcie_start_link(pci);
> >     if (ret)
> >       goto err_free_msi;
> >     dw_pcie_wait_for_link(pci);
> >   }
> > 
> > 
> > ...which will ensure that we don't call dw_pcie_wait_for_link() when
> > pci->ops->start_link is NULL.
> > 
> > With regards to the log, I think there are 2 ways to solve this:
> > 
> > 1) We could also call dw_pcie_wait_for_link() in a new else if
> >    dw_pcie_link_up() returns 1.
> > 2) We could add this to the top of dw_pcie_wait_for_link() and leave the
> >    code as is:
> > 
> >    if (!pci->ops || !pci->ops->start_link)
> >      return 0;
> > 
> > I kind of like (2) since that solves both Sajid's original issue and
> > will keep the original log.
> > 
> > [1] https://lore.kernel.org/all/20220624143428.8334-14-Sergey.Semin@xxxxxxxxxxxxxxxxxxxx/
> > 
> > Regards,
> > Will
> 
> Below is what I'm thinking will do the job. I verified on a Pixel 6
> (which doesn't have start_link() defined) that we don't have the 1
> second wait from dw_pcie_wait_for_link() during probe.
> 
> diff --git a/drivers/pci/controller/dwc/pcie-designware.c b/drivers/pci/controller/dwc/pcie-designware.c
> index 8e33e6e59e68..1bf04324ad2d 100644
> --- a/drivers/pci/controller/dwc/pcie-designware.c
> +++ b/drivers/pci/controller/dwc/pcie-designware.c
> @@ -648,13 +648,16 @@ int dw_pcie_wait_for_link(struct dw_pcie *pci)
>  {
>  	u32 offset, val;
>  	int retries;
> +	int link_up = dw_pcie_link_up(pci);
>  
> -	/* Check if the link is up or not */
> -	for (retries = 0; retries < LINK_WAIT_MAX_RETRIES; retries++) {
> -		if (dw_pcie_link_up(pci))
> -			break;
> +	if (!link_up && !(pci->ops && pci->ops->start_link))
> +		return 0;
There is a problem with this approach. A platform driver could enable
link training internally, i.e., it does not have the start_link() pointer
defined. Then it could call `dw_pcie_wait_for_link` to wait for the link
to come up. (See pcie-intel-gw.c for an example of such a platform).
Your logic will end up regressing this driver by exiting early.
>  
> +	/* Check if the link is up or not */
> +	for (retries = 0; !link_up && retries < LINK_WAIT_MAX_RETRIES; retries++) {
>  		usleep_range(LINK_WAIT_USLEEP_MIN, LINK_WAIT_USLEEP_MAX);
> +
> +		link_up = dw_pcie_link_up(pci);
>  	}
>  
>  	if (retries >= LINK_WAIT_MAX_RETRIES) {
> 
The problem of the log is still not solved for a platform which could
have the link up by default, i.e., it does not need to explicitly enable
link training.