[PATCH v3] PCI: cadence: Fix Gen2 Link Retraining process

[Lorenzo Pieralisi]

On Mon, Jun 12, 2023 at 09:56:27AM +0530, Siddharth Vadapalli wrote:
> 
> 
> On 09/06/23 23:09, Bjorn Helgaas wrote:
> > On Wed, Jun 07, 2023 at 02:44:27PM +0530, Siddharth Vadapalli wrote:
> >> The Link Retraining process is initiated to account for the Gen2 defect in
> >> the Cadence PCIe controller in J721E SoC. The errata corresponding to this
> >> is i2085, documented at:
> >> https://www.ti.com/lit/er/sprz455c/sprz455c.pdf
> >>
> >> The existing workaround implemented for the errata waits for the Data Link
> >> initialization to complete and assumes that the link retraining process
> >> at the Physical Layer has completed. However, it is possible that the
> >> Physical Layer training might be ongoing as indicated by the
> >> PCI_EXP_LNKSTA_LT bit in the PCI_EXP_LNKSTA register.
> >>
> >> Fix the existing workaround, to ensure that the Physical Layer training
> >> has also completed, in addition to the Data Link initialization.
> >>
> >> Fixes: 4740b969aaf5 ("PCI: cadence: Retrain Link to work around Gen2 training defect")
> >> Signed-off-by: Siddharth Vadapalli <s-vadapalli at ti.com>
> >> Reviewed-by: Vignesh Raghavendra <vigneshr at ti.com>
> >> ---
> >>
> >> Hello,
> >>
> >> This patch is based on linux-next tagged next-20230606.
> >>
> >> v2:
> >> https://lore.kernel.org/r/20230315070800.1615527-1-s-vadapalli@ti.com/
> >> Changes since v2:
> >> - Merge the cdns_pcie_host_training_complete() function with the
> >>   cdns_pcie_host_wait_for_link() function, as suggested by Bjorn
> >>   for the v2 patch.
> >> - Add dev_err() to notify when Link Training fails, since this is a
> >>   fatal error and proceeding from this point will almost always crash
> >>   the kernel.
> >>
> >> v1:
> >> https://lore.kernel.org/r/20230102075656.260333-1-s-vadapalli@ti.com/
> >> Changes since v1:
> >> - Collect Reviewed-by tag from Vignesh Raghavendra.
> >> - Rebase on next-20230315.
> >>
> >> Regards,
> >> Siddharth.
> >>
> >>  .../controller/cadence/pcie-cadence-host.c    | 20 +++++++++++++++++++
> >>  1 file changed, 20 insertions(+)
> >>
> >> diff --git a/drivers/pci/controller/cadence/pcie-cadence-host.c b/drivers/pci/controller/cadence/pcie-cadence-host.c
> >> index 940c7dd701d6..70a5f581ff4f 100644
> >> --- a/drivers/pci/controller/cadence/pcie-cadence-host.c
> >> +++ b/drivers/pci/controller/cadence/pcie-cadence-host.c
> >> @@ -12,6 +12,8 @@
> >>  
> >>  #include "pcie-cadence.h"
> >>  
> >> +#define LINK_RETRAIN_TIMEOUT HZ
> >> +
> >>  static u64 bar_max_size[] = {
> >>  	[RP_BAR0] = _ULL(128 * SZ_2G),
> >>  	[RP_BAR1] = SZ_2G,
> >> @@ -80,8 +82,26 @@ static struct pci_ops cdns_pcie_host_ops = {
> >>  static int cdns_pcie_host_wait_for_link(struct cdns_pcie *pcie)
> >>  {
> >>  	struct device *dev = pcie->dev;
> >> +	unsigned long end_jiffies;
> >> +	u16 link_status;
> >>  	int retries;
> >>  
> >> +	/* Wait for link training to complete */
> >> +	end_jiffies = jiffies + LINK_RETRAIN_TIMEOUT;
> >> +	do {
> >> +		link_status = cdns_pcie_rp_readw(pcie, CDNS_PCIE_RP_CAP_OFFSET + PCI_EXP_LNKSTA);
> >> +		if (!(link_status & PCI_EXP_LNKSTA_LT))
> >> +			break;
> >> +		usleep_range(0, 1000);
> >> +	} while (time_before(jiffies, end_jiffies));
> >> +
> >> +	if (!(link_status & PCI_EXP_LNKSTA_LT)) {
> >> +		dev_info(dev, "Link training complete\n");
> >> +	} else {
> >> +		dev_err(dev, "Fatal! Link training incomplete\n");
> >> +		return -ETIMEDOUT;
> >> +	}
> > 
> > Can I have a brown paper bag, please?  I totally blew it here, and I'm
> > sorry.
> > 
> > You took my advice by combining this with the existing
> > cdns_pcie_host_wait_for_link(), but I think my advice was poor because
> > (a) now this additional wait is not clearly connected with the
> > erratum, and (b) it affects devices that don't have the erratum.
> > 
> > IIUC, this is all part of a workaround for the i2085 erratum.  The
> > original workaround, 4740b969aaf5 ("PCI: cadence: Retrain Link to work
> > around Gen2 training defect"), added this:
> > 
> >   if (!ret && rc->quirk_retrain_flag)
> >     ret = cdns_pcie_retrain(pcie);
> > 
> > I think the wait for link train to complete should also be in
> > cdns_pcie_retrain() so it's clearly connected with the quirk, which
> > also means we'd only do the wait for devices with the erratum.
> > 
> > Which is EXACTLY what your first patch did, and I missed it.  I am
> > very sorry.  I guess maybe I thought cdns_pcie_retrain() was a
> > general-purpose thing, but in fact it's only used for this quirk.
> 
> With the current approach implemented in this patch, I could do the following:
> In the cdns_pcie_host_wait_for_link() function, I obtain the reference to the
> struct cdns_pcie_rc *rc, using:
> struct cdns_pcie_rc *rc = container_of(pcie, struct cdns_pcie_rc, pcie);
> followed by checking if the quirk "quirk_retrain_flag" is set, before proceeding
> with the Link Training check added by this patch. With this, only the
> controllers with the quirk will check for the Link Training completion before
> proceeding. However, the difference with this new approach compared to the
> approach in the v2 patch is that in this new approach, even in the Link Training
> Phase, the Link Training check is performed for the controllers with the quirk,
> unlike the v2 patch where the Link Training check was performed only during the
> Link Retraining Phase through the cdns_pcie_retrain() function.
> 
> Also, based on Mani's suggestion, I have measured the latency introduced by the
> Link Training check for both quirky and non-quirky controllers at:
> https://lore.kernel.org/r/a63fc8b0-581b-897f-cac6-cb0a0e82c63e@ti.com/
> If the latency is acceptable, then the current implementation in this v3 patch
> could be fine too.
> 
> Kindly let me know which approach among the following seems to be the best one:
> 1. The approach implemented in v2 patch (I will make minor changes to the patch
> to print out the "Fatal" error, so that users will be informed of the cause of
> the crash, followed by posting a v4 patch with this change).
> 2. The current implementation in the v3 patch with a check added to see if the
> controller has the quirk_retrain_flag set, before proceeding with the Link
> Training check.
> 3. The current implementation in the v3 patch as is, without any modification,
> if the latency introduced is not a concern and the sanity check for Link
> Training completion for non-quirky controllers appears acceptable.

The point is, you stated it yourself that the non-quirky path is broken
too in its *current* form, I don't think there is any other option on
the table other than (3) (unless we want to rely on probe time timing
to hide the issue; that to me it is not even considerable as an option).

Lorenzo