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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 2023/7/28 5:40, Bjorn Helgaas wrote:
> [+cc Mika, Maciej since they've worked on similar delays recently]
> 
> 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().

I am sorry for the late feedback to you, because we keep on testing since last week.
Several NVMe SSD are verified with this patch, and they work fine.

It is a pity that we lost the Thinklife NVMe SSD mentioned before,
because it belongs to a departing employee.
We bought two new SSD in the same model for testing,
the issue can not be reproduced, and all of then work fine with V1 & V2 patch.

> 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



[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux