Re: [PATCH v2] PCI: aardvark: Use LTSSM state to build link training flag

[Remi Pommarel]

Hi Lorenzo,

Sorry for duplicates I forgot to include everyone.

On Thu, Apr 25, 2019 at 04:06:40PM +0100, Lorenzo Pieralisi wrote:
> On Thu, Apr 25, 2019 at 04:23:53PM +0200, Remi Pommarel wrote:
> > Hi Lorenzo,
> > 
> > On Thu, Apr 25, 2019 at 12:08:30PM +0100, Lorenzo Pieralisi wrote:
> > > On Sat, Mar 16, 2019 at 05:12:43PM +0100, Remi Pommarel wrote:
> > > > The PCI_EXP_LNKSTA_LT flag in the emulated root device's PCI_EXP_LNKSTA
> > > > config register does not reflect the actual link training state and is
> > > > always cleared. The Link Training and Status State Machine (LTSSM) flag
> > > > in 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 setting the PCI_EXP_LINCTL_RL flag does not
> > > > instantly imply a LTSSM state change (e.g. L0s to recovery state
> > > > transition takes some time), LTSSM can be in L0 but link training has
> > > > not finished yet. Thus a lower L0 LTSSM state followed by a L0 or upper
> > > > state sequence has to be seen to be sure that link training has been
> > > > done.
> > > 
> > > Hi Remi,
> > > 
> > > I am a bit confused, so you are saying that the LTSSM flag in the
> > > LMI config register can't be used to detect when training is completed ?
> > 
> > Not exactly, I am saying that PCI_EXP_LNKSTA_LT from PCI_EXP_LNKSTA
> > register can't be used with this hardware, but can be emulated with
> > LTSSM flag.
> > 
> > > 
> > > Certainly it can't be used by ASPM core that relies on:
> > > 
> > > PCI_EXP_LNKSTA_LT flag
> > > 
> > > in the PCI_EXP_LNKSTA register, and that's what you are setting through
> > > this timeout mechanism IIUC.
> > > 
> > > Please elaborate on that.
> > 
> > The problem here is that the hardware does not change PCI_EXP_LNKSTA_LT
> > at all. So in order to support link re-training feature we need to
> > emulate this flag. To do so LTSSM flag can be used.
> 
> Understood.
> 
> > Indeed we can set the emulated PCI_EXP_LNKSTA_LT as soon as re-training
> > is asked and wait for LTSSM flag to be back to a configured state
> > (e.g. L0, L0s) before clearing it.
> 
> The check for the LTSSM is carried out through advk_pcie_link_up()
> (ie register CFG_REG), correct ?
> 

Yes that is correct.

> > The problem with that is that LTSSM flag does not change instantly after
> > link re-training has been asked, and will stay in configured state for a
> > small amount of time. So the idea is to poll the LTSSM flag and wait for
> > it to enter a recovery state then waiting for it to be back in
> > configured state.
> 
> When you say "poll" you mean checking advk_pcie_link_up() ?
> 

I mean checking advk_pcie_link_up() in a loop. This loop is done by the
user (e.g. ASPM core). ASPM core waits for PCI_EXP_LNKSTA_LT to be
cleared in pcie_aspm_configure_common_clock() just after it has set
PCI_EXP_LNKCTL_RL.

So the idea was to check advk_pcie_link_up() each time ASPM core checks
the PCI_EXP_LNKSTA_LT flag. Please see below patch for an alternative
to that.

> More below on the code.
> 
> > The timeout is only here as a fallback in the unlikely event that we
> > missed the LTSSM flag entering recovery state.
> > 
> > > 
> > > I am picking Bjorn's brain on this patch since what you are doing
> > > seems quite arbitrary and honestly it is a bit of a hack.
> > 
> > Yes, sorry, it is a bit of a hack because I try to workaround a
> > hardware issue.
> 
> No problems, it is not your fault.
> > 
> > Please note that vendor has been contacted about this in the meantime
> > and answered the following:
> > 
> > "FW can poll LTSSM state equals any of the following values: 0xB or 0xD
> > or 0xC or 0xE. After that, polls for LTSSM equals 0x10. For your
> > information, LTSSM will transit from 0x10 -> 0xB -> 0xD -> 0xC or 0xE
> > ........... -> 0x10".
> > 
> > It is basically what this patch does, I've just added a timeout fallback
> > to not poll LTSSM state forever if its transition to 0xB, 0xD, 0xC or
> > 0xE has been missed.
> 
> When you say "missed" you mean advk_pcie_link_up() returning true, right ?
> 

Not exactly, I mean that LTSSM had the time to go down and back up
between advk_pcie_link_up() because, for example, ASPM core loop took
too much time between two PCI_EXP_LNKSTA_LT flag checks.

> [...]
> 
> > > > +static int advk_pcie_link_retraining(struct advk_pcie *pcie)
> > > > +{
> > > > +	if (!advk_pcie_link_up(pcie)) {
> 
> That's the bit I find confusing. Is this check here to detect if the
> link went through the sequence below ? Should not it be carried
> out only if (pcie->rl_asked == 1) ?
> 
> "... LTSSM will transit from 0x10 -> 0xB -> 0xD -> 0xC or 0xE
>  ........... -> 0x10".

Yes it is the check to detect the sequence. advk_pcie_link_up() returns
false if LTSSM <= 0x10.

This cannot be done only if (pcie->rl_asked == 1) because I still
want this function to return 1 if link is still down.

> 
> > > > +		pcie->rl_asked = 0;
> 
> Why ?
> 

rl_asked is not a good name, I could have called it
pcie->wait_for_link_down instead. So if advk_pcie_link_up() returns
false that means that we don't need to wait for link being down any more
and just wait for (LTSSM >= 0x10). In this case the delay is not needed.

> > > > +		return 1;
> > > > +	}
> > > > +
> > > > +	if (pcie->rl_asked && time_before(jiffies, pcie->rl_deadline))
> > > > +		return 1;
> 
> This ensures that if the LTSSM >= 0x10 we still wait for a delay before
> considering the link up (because I suppose, after asking a retraining
> it takes a while for the LTSSM state to become < 0x10), correct ?

Yes it takes a while to become < 0x10 after retraining hence the delay.
But here we don't need to always wait for a delay. Indeed if we've
already seen the link being < 0x10 (i.e if "pcie->rl_asked == 0") and
if after that link is >= 0x10 then we know that retraining process has
finished.

Anyway I did it this way because I wanted to keep
advk_pci_bridge_emul_pcie_conf_write() from polling. But this is
obviously a bad reason as it makes the code way too complex and relies
on user (ASPM core) to do the poll instead.

So if you find the following better I'll send a v3 with that:

---
diff --git a/drivers/pci/controller/pci-aardvark.c b/drivers/pci/controller/pci-aardvark.c
index eb58dfdaba1b..67e8ae4e313e 100644
--- a/drivers/pci/controller/pci-aardvark.c
+++ b/drivers/pci/controller/pci-aardvark.c
@@ -180,6 +180,9 @@
 #define LINK_WAIT_MAX_RETRIES		10
 #define LINK_WAIT_USLEEP_MIN		90000
 #define LINK_WAIT_USLEEP_MAX		100000
+#define RETRAIN_WAIT_MAX_RETRIES	20
+#define RETRAIN_WAIT_USLEEP_MIN		2000
+#define RETRAIN_WAIT_USLEEP_MAX		5000
 
 #define MSI_IRQ_NUM			32
 
@@ -239,6 +242,17 @@ 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;
+		usleep_range(RETRAIN_WAIT_USLEEP_MIN, RETRAIN_WAIT_USLEEP_MAX);
+	}
+}
+
 static void advk_pcie_setup_hw(struct advk_pcie *pcie)
 {
 	u32 reg;
@@ -426,11 +440,19 @@ advk_pci_bridge_emul_pcie_conf_read(struct pci_bridge_emul *bridge,
 		return PCI_BRIDGE_EMUL_HANDLED;
 	}
 
+	case PCI_EXP_LNKCTL: {
+		u32 val = advk_readl(pcie, PCIE_CORE_PCIEXP_CAP + reg) &
+			~(PCI_EXP_LNKSTA_LT << 16);
+		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;
+
 	case PCI_EXP_LNKCTL:
 		advk_writel(pcie, new, PCIE_CORE_PCIEXP_CAP + reg);
+		if (new & PCI_EXP_LNKCTL_RL)
+			advk_pcie_wait_for_retrain(pcie);
 		break;
 
 	case PCI_EXP_RTCTL: