Re: [PATCH v3] PCI: aardvark: Use LTSSM state to build link training flag

[Remi Pommarel]

Hi Lorenzo, Marc,

On Mon, Oct 14, 2019 at 11:01:29AM +0100, Lorenzo Pieralisi wrote:
> On Sun, Oct 13, 2019 at 11:34:15AM +0100, Marc Zyngier wrote:
> > On Tue, 1 Oct 2019 09:05:46 +0100
> > Andrew Murray <andrew.murray@xxxxxxx> wrote:
> > 
> > Hi Lorenzo,
> > 
> > > On Mon, Sep 30, 2019 at 06:52:30PM +0200, Remi Pommarel wrote:
> > > > On Mon, Sep 30, 2019 at 04:40:18PM +0100, Andrew Murray wrote:  
> > > > > On Wed, May 22, 2019 at 11:33:51PM +0200, Remi Pommarel wrote:  
> > > > > > Aardvark's PCI_EXP_LNKSTA_LT flag in its link status register is not
> > > > > > implemented and does not reflect the actual link training state (the
> > > > > > flag is always set to 0). In order to support link re-training feature
> > > > > > this flag has to be emulated. The Link Training and Status State
> > > > > > Machine (LTSSM) flag in Aardvark LMI config register could be used as
> > > > > > a link training indicator. Indeed if the LTSSM is in L0 or upper state
> > > > > > then link training has completed (see [1]).
> > > > > > 
> > > > > > Unfortunately because after asking a link retraining it takes a while
> > > > > > for the LTSSM state to become less than 0x10 (due to L0s to recovery
> > > > > > state transition delays), LTSSM can still be in L0 while link training
> > > > > > has not finished yet. So this waits for link to be in recovery or lesser
> > > > > > state before returning after asking for a link retrain.
> > > > > > 
> > > > > > [1] "PCI Express Base Specification", REV. 4.0
> > > > > >     PCI Express, February 19 2014, Table 4-14
> > > > > > 
> > > > > > Signed-off-by: Remi Pommarel <repk@xxxxxxxxxxxx>
> > > > > > ---
> > > > > > Changes since v1:
> > > > > >   - Rename retraining flag field
> > > > > >   - Fix DEVCTL register writing
> > > > > > 
> > > > > > Changes since v2:
> > > > > >   - Rewrite patch logic so it is more legible
> > > > > > 
> > > > > > Please note that I will unlikely be able to answer any comments from May
> > > > > > 24th to June 10th.
> > > > > > ---
> > > > > >  drivers/pci/controller/pci-aardvark.c | 29 ++++++++++++++++++++++++++-
> > > > > >  1 file changed, 28 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/drivers/pci/controller/pci-aardvark.c b/drivers/pci/controller/pci-aardvark.c
> > > > > > index 134e0306ff00..8803083b2174 100644
> > > > > > --- a/drivers/pci/controller/pci-aardvark.c
> > > > > > +++ b/drivers/pci/controller/pci-aardvark.c
> > > > > > @@ -180,6 +180,8 @@
> > > > > >  #define LINK_WAIT_MAX_RETRIES		10
> > > > > >  #define LINK_WAIT_USLEEP_MIN		90000
> > > > > >  #define LINK_WAIT_USLEEP_MAX		100000
> > > > > > +#define RETRAIN_WAIT_MAX_RETRIES	10
> > > > > > +#define RETRAIN_WAIT_USLEEP_US		2000
> > > > > >  
> > > > > >  #define MSI_IRQ_NUM			32
> > > > > >  
> > > > > > @@ -239,6 +241,17 @@ static int advk_pcie_wait_for_link(struct advk_pcie *pcie)
> > > > > >  	return -ETIMEDOUT;
> > > > > >  }
> > > > > >  
> > > > > > +static void advk_pcie_wait_for_retrain(struct advk_pcie *pcie)
> > > > > > +{
> > > > > > +	size_t retries;
> > > > > > +
> > > > > > +	for (retries = 0; retries < RETRAIN_WAIT_MAX_RETRIES; ++retries) {
> > > > > > +		if (!advk_pcie_link_up(pcie))
> > > > > > +			break;
> > > > > > +		udelay(RETRAIN_WAIT_USLEEP_US);
> > > > > > +	}
> > > > > > +}
> > > > > > +
> > > > > >  static void advk_pcie_setup_hw(struct advk_pcie *pcie)
> > > > > >  {
> > > > > >  	u32 reg;
> > > > > > @@ -426,11 +439,20 @@ advk_pci_bridge_emul_pcie_conf_read(struct pci_bridge_emul *bridge,
> > > > > >  		return PCI_BRIDGE_EMUL_HANDLED;
> > > > > >  	}
> > > > > >  
> > > > > > +	case PCI_EXP_LNKCTL: {
> > > > > > +		/* u32 contains both PCI_EXP_LNKCTL and PCI_EXP_LNKSTA */
> > > > > > +		u32 val = advk_readl(pcie, PCIE_CORE_PCIEXP_CAP + reg) &
> > > > > > +			~(PCI_EXP_LNKSTA_LT << 16);  
> > > > > 
> > > > > The commit message says "the flag is always set to 0" - therefore I guess
> > > > > you don't *need* to mask out the LT bit here? I assume this is just
> > > > > belt-and-braces but thought I'd check incase I've misunderstood or if your
> > > > > commit message is inaccurate.
> > > > > 
> > > > > In any case masking out the bit (or adding a comment) makes this code more
> > > > > readable as the reader doesn't need to know what the hardware does with this
> > > > > bit.  
> > > > 
> > > > Actually vendor eventually responded that the bit was reserved, but
> > > > during my tests it remains to 0.
> > > > 
> > > > So yes I am masking this out mainly for belt-and-braces and legibility.  
> > > 
> > > Thanks for the clarification.
> > > 
> > > >   
> > > > > > +		if (!advk_pcie_link_up(pcie))
> > > > > > +			val |= (PCI_EXP_LNKSTA_LT << 16);
> > > > > > +		*value = val;
> > > > > > +		return PCI_BRIDGE_EMUL_HANDLED;
> > > > > > +	}
> > > > > > +
> > > > > >  	case PCI_CAP_LIST_ID:
> > > > > >  	case PCI_EXP_DEVCAP:
> > > > > >  	case PCI_EXP_DEVCTL:
> > > > > >  	case PCI_EXP_LNKCAP:
> > > > > > -	case PCI_EXP_LNKCTL:
> > > > > >  		*value = advk_readl(pcie, PCIE_CORE_PCIEXP_CAP + reg);
> > > > > >  		return PCI_BRIDGE_EMUL_HANDLED;
> > > > > >  	default:
> > > > > > @@ -447,8 +469,13 @@ advk_pci_bridge_emul_pcie_conf_write(struct pci_bridge_emul *bridge,
> > > > > >  
> > > > > >  	switch (reg) {
> > > > > >  	case PCI_EXP_DEVCTL:
> > > > > > +		advk_writel(pcie, new, PCIE_CORE_PCIEXP_CAP + reg);
> > > > > > +		break;  
> > > > > 
> > > > > Why is this here?
> > > > >   
> > > > 
> > > > Before PCI_EXP_DEVCTL and PCI_EXP_LNKCTL were doing the same thing, but
> > > > as now PCI_EXP_LNKCTL does extra things (i.e. wait for link to
> > > > successfully retrain), they do have different behaviours.
> > > > 
> > > > So this is here so PCI_EXP_DEVCTL keeps its old behaviour and do not
> > > > wait for link retrain in case an unrelated (PCI_EXP_LNKCTL_RL) bit is
> > > > set.  
> > > 
> > > Oh yes, of course!
> > > 
> > > Thanks and:
> > > 
> > > Reviewed-by: Andrew Murray <andrew.murray@xxxxxxx>
> > 
> > Is there any hope for this workaround to make it into 5.4?
> > 
> > My EspressoBin (which is blessed with this joke of a PCIe controller)
> > is pretty much a doorstop without it and dies with a SError early at
> > boot.
> > 
> > FWIW:
> > 
> > Tested-by: Marc Zyngier <maz@xxxxxxxxxx>
> 
> Hi Marc,
> 
> First thing I will have to mark it as a Fixes: (if Remi can provide
> me with a tag that'd be great), usually we send fixes at -rc* for
> patches that fix code that went in the current (eg 5.4) material,
> I will ask Bjorn to see if we can send this in one of the upcoming
> -rc* even if it fixes older code.

Sure, I think this could be considered a fix for the following commit :
Fixes: 8a3ebd8de328 ("PCI: aardvark: Implement emulated root PCI bridge config space")

Moreover, Marc, I am also a bit supprised that you did not have to use
[1] to even be able to boot.

Also if you want to be completely immune to this kind of SError (that
could theoretically happen if the link goes down for other reasons than
being retrained) you would have to use mainline ATF along with [2]. But
the chances to hit that are low (could only happen in case of link
errors).

[1] [v3] PCI: aardvark: Use LTSSM state to build link training flag
    https://patchwork.ozlabs.org/patch/1115864/
[2] [v3] PCI: aardvark: Don't rely on jiffies while holding spinlock
    https://patchwork.ozlabs.org/patch/1168349/

-- 
Remi