Re: [PATCH v1 8/9] PCI: PLDA: starfive: Add JH7110 PCIe controller

[Pali Rohár]

On Monday 31 July 2023 18:12:23 Bjorn Helgaas wrote:
> [+cc Pali, Marek because I used f76b36d40bee ("PCI: aardvark: Fix link
> training") as an example]
> 
> On Mon, Jul 31, 2023 at 01:52:35PM +0800, Kevin Xie wrote:
> > On 2023/7/28 5:40, Bjorn Helgaas wrote:
> > > On Tue, Jul 25, 2023 at 03:46:35PM -0500, Bjorn Helgaas wrote:
> > >> On Mon, Jul 24, 2023 at 06:48:47PM +0800, Kevin Xie wrote:
> > >> > On 2023/7/21 0:15, Bjorn Helgaas wrote:
> > >> > > On Thu, Jul 20, 2023 at 06:11:59PM +0800, Kevin Xie wrote:
> > >> > >> On 2023/7/20 0:48, Bjorn Helgaas wrote:
> > >> > >> > On Wed, Jul 19, 2023 at 06:20:56PM +0800, Minda Chen wrote:
> > >> > >> >> Add StarFive JH7110 SoC PCIe controller platform
> > >> > >> >> driver codes.
> > >> 
> > >> > >> However, in the compatibility testing with several NVMe SSD, we
> > >> > >> found that Lenovo Thinklife ST8000 NVMe can not get ready in 100ms,
> > >> > >> and it actually needs almost 200ms.  Thus, we increased the T_PVPERL
> > >> > >> value to 300ms for the better device compatibility.
> > >> > > ...
> > >> > > 
> > >> > > Thanks for this valuable information!  This NVMe issue potentially
> > >> > > affects many similar drivers, and we may need a more generic fix so
> > >> > > this device works well with all of them.
> > >> > > 
> > >> > > T_PVPERL is defined to start when power is stable.  Do you have a way
> > >> > > to accurately determine that point?  I'm guessing this:
> > >> > > 
> > >> > >   gpiod_set_value_cansleep(pcie->power_gpio, 1)
> > >> > > 
> > >> > > turns the power on?  But of course that doesn't mean it is instantly
> > >> > > stable.  Maybe your testing is telling you that your driver should
> > >> > > have a hardware-specific 200ms delay to wait for power to become
> > >> > > stable, followed by the standard 100ms for T_PVPERL?
> > >> > 
> > >> > You are right, we did not take the power stable cost into account.
> > >> > T_PVPERL is enough for Lenovo Thinklife ST8000 NVMe SSD to get ready,
> > >> > and the extra cost is from the power circuit of a PCIe to M.2 connector,
> > >> > which is used to verify M.2 SSD with our EVB at early stage.
> > >> 
> > >> Hmm.  That sounds potentially interesting.  I assume you're talking
> > >> about something like this: https://www.amazon.com/dp/B07JKH5VTL
> > >> 
> > >> I'm not familiar with the timing requirements for something like this.
> > >> There is a PCIe M.2 spec with some timing requirements, but I don't
> > >> know whether or how software is supposed to manage this.  There is a
> > >> T_PVPGL (power valid to PERST# inactive) parameter, but it's
> > >> implementation specific, so I don't know what the point of that is.
> > >> And I don't see a way for software to even detect the presence of such
> > >> an adapter.
> > > 
> > > I intended to ask about this on the PCI-SIG forum, but after reading
> > > this thread [1], I don't think we would learn anything.  The question
> > > was:
> > > 
> > >   The M.2 device has 5 voltage rails generated from the 3.3V input
> > >   supply voltage
> > >   -------------------------------------------
> > >   This is re. Table 17 in PCI Express M.2 Specification Revision 1.1
> > >   Power Valid* to PERST# input inactive : Implementation specific;
> > >   recommended 50 ms
> > > 
> > >   What exactly does this mean ?
> > > 
> > >   The Note says
> > > 
> > >     *Power Valid when all the voltage supply rails have reached their
> > >     respective Vmin.
> > > 
> > >   Does this mean that the 50ms to PERSTn is counted from the instant
> > >   when all *5 voltage rails* on the M.2 device have become "good" ?
> > > 
> > > and the answer was:
> > > 
> > >   You wrote;
> > >   Does this mean that the 50ms to PERSTn is counted from the instant
> > >   when all 5 voltage rails on the M.2 device have become "good" ?
> > > 
> > >   Reply:
> > >   This means that counting the recommended 50 ms begins from the time
> > >   when the power rails coming to the device/module, from the host, are
> > >   stable *at the device connector*.
> > > 
> > >   As for the time it takes voltages derived inside the device from any
> > >   of the host power rails (e.g., 3.3V rail) to become stable, that is
> > >   part of the 50ms the host should wait before de-asserting PERST#, in
> > >   order ensure that most devices will be ready by then.
> > > 
> > >   Strictly speaking, nothing disastrous happens if a host violates the
> > >   50ms. If it de-asserts too soon, the device may not be ready, but
> > >   most hosts will try again. If the host de-asserts too late, the
> > >   device has even more time to stabilize. This is why the WG felt that
> > >   an exact minimum number for >>Tpvpgl, was not valid in practice, and
> > >   we made it a recommendation.
> > > 
> > > Since T_PVPGL is implementation-specific, we can't really base
> > > anything in software on the 50ms recommendation.  It sounds to me like
> > > they are counting on software to retry config reads when enumerating.
> > > 
> > > I guess the delays we *can* observe are:
> > > 
> > >   100ms T_PVPERL "Power stable to PERST# inactive" (CEM 2.9.2)
> > >   100ms software delay between reset and config request (Base 6.6.1)
> > 
> > Refer to Figure2-10 in CEM Spec V2.0, I guess this two delays are T2 & T4?
> > In the PATCH v2[4/4], T2 is the msleep(100) for T_PVPERL,
> > and T4 is done by starfive_pcie_host_wait_for_link().
> 
> Yes, I think "T2" is T_PVPERL.  The CEM r2.0 Figure 2-10 note is
> "2. Minimum time from power rails within specified tolerance to
> PERST# inactive (T_PVPERL)."
> 
> As far as T4 ("Minimum PERST# inactive to PCI Express link out of
> electrical idle"), I don't see a name or a value for that parameter,
> and I don't think it is the delay required by PCIe r6.0, sec 6.6.1.
> 
> The delay required by sec 6.6.1 is a minimum of 100ms following exit
> from reset or, for fast links, 100ms after link training completes.
> 
> The comment at the call of advk_pcie_wait_for_link() [2] says it is
> the delay required by sec 6.6.1, but that doesn't seem right to me.
> 
> For one thing, I don't think 6.6.1 says anything about "link up" being
> the end of a delay.  So if we want to do the delay required by 6.6.1,
> "wait_for_link()" doesn't seem like quite the right name.
> 
> For another, all the *_wait_for_link() functions can return success
> after 0ms, 90ms, 180ms, etc.  They're unlikely to return after 0ms,
> but 90ms is quite possible.  If we avoided the 0ms return and
> LINK_WAIT_USLEEP_MIN were 100ms instead of 90ms, that should be enough
> for slow links, where we need 100ms following "exit from reset."
> 
> But it's still not enough for fast links where we need 100ms "after
> link training completes" because we don't know when training
> completed.  If training completed 89ms into *_wait_for_link(), we only
> delay 1ms after that.

Please look into discussion "How long should be PCIe card in Warm Reset
state?" including external references where are more interesting details:
https://lore.kernel.org/linux-pci/20210310110535.zh4pnn4vpmvzwl5q@pali/

About wait for the link, this should be done asynchronously...

> > > The PCI core doesn't know how to assert PERST#, so the T_PVPERL delay
> > > definitely has to be in the host controller driver.
> > > 
> > > The PCI core observes the second 100ms delay after a reset in
> > > pci_bridge_wait_for_secondary_bus().  But this 100ms delay does not
> > > happen during initial enumeration.  I think the assumption of the PCI
> > > core is that when the host controller driver calls pci_host_probe(),
> > > we can issue config requests immediately.
> > > 
> > > So I think that to be safe, we probably need to do both of those 100ms
> > > delays in the host controller driver.  Maybe there's some hope of
> > > supporting the latter one in the PCI core someday, but that's not
> > > today.
> > > 
> > > Bjorn
> > > 
> > > [1] https://forum.pcisig.com/viewtopic.php?f=74&t=1037
> 
> [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/controller/pci-aardvark.c?id=v6.4#n433