Re: [PATCH 2/2] PCI: uniphier: Add checking whether PERST# is deasserted

From: Kunihiko Hayashi
Date: Sun Dec 08 2019 - 21:35:49 EST


Hi Kishon,

On Fri, 6 Dec 2019 14:31:17 +0530 <kishon@xxxxxx> wrote:

> Hi,
>
> On 06/12/19 2:28 pm, Kunihiko Hayashi wrote:
> > Hi Kishon,
> > > On Fri, 6 Dec 2019 12:28:29 +0530 <kishon@xxxxxx> wrote:
> > >> Hi,
> >>
> >> On 04/12/19 3:35 pm, Kunihiko Hayashi wrote:
> >>> On Fri, 22 Nov 2019 20:53:16 +0900 <hayashi.kunihiko@xxxxxxxxxxxxx> wrote:
> >>>>> Hello Lorenzo,
> >>>>
> >>>> On Thu, 21 Nov 2019 16:47:05 +0000 <lorenzo.pieralisi@xxxxxxx> wrote:
> >>>>
> >>>>> On Fri, Nov 08, 2019 at 04:30:27PM +0900, Kunihiko Hayashi wrote:
> >>>>>>> However, If I understand correctly, doesn't your solution only work some
> >>>>>>> of the time? What happens if you boot both machines at the same time,
> >>>>>>> and PERST# isn't asserted prior to the kernel booting?
> >>>>>>
> >>>>>> I think it contains an annoying problem.
> >>>>>>
> >>>>>> If PERST# isn't toggled prior to the kernel booting, PERST# remains asserted
> >>>>>> and the RC driver can't access PCI bus.
> >>>>>>
> >>>>>> As a result, this patch works and deasserts PERST# (and EP configuration will
> >>>>>> be lost). So boot sequence needs to include deasserting PERST#.
> >>>>>
> >>>>> I am sorry but I have lost you. Can you explain to us why checking
> >>>>> that PERST# is deasserted guarantees you that:
> >>>>>
> >>>>> - The EP has bootstrapped
> >>>>> - It is safe not to toggle it again (and also skip
> >>>>> uniphier_pcie_ltssm_enable())
> >>>>>
> >>>>> Please provide details of the HW configuration so that we understand
> >>>>> what's actually supposed to happen and why this patch fixes the
> >>>>> issue you are facing.
> >>>>
> >>>> I tried to connect between the following boards, and do pci-epf-test:
> >>>> - "RC board": UniPhier ld20 board that has DWC RC controller
> >>>> - "EP board": UniPhier legacy board that has DWC EP controller
> >>>>
> >>>> This EP has power-on-state configuration, but it's necessary to set
> >>>> class ID, BAR sizes, etc. after starting up.
> >>>>
> >>>> In case of that starting up RC board before EP board, the RC driver
> >>>> can't establish link. So we need to boot EP board first.
> >>>> At that point, I've considered why RC can't establish link,
> >>> and found that the waitng time was too short.
> >>>> - EP/RC: power on both boards
> >>>> - RC: start up the kernel on RC board
> >>>> - RC: wait for link up (long time enough)
> >>>> - EP: start up the kernel on EP board
> >>>> - EP: configurate pci-epf-test
> >>>> When the endpoint configuration is done and the EP driver enables LTSSM,
> >>> the RC driver will quit from waiting for link up.
> >>>> Currently DWC RC driver calls dwc_pcie_wait_for_link(), however,
> >>> the function tries to link up 10 times only, that is defined
> >>> as LINK_WAIT_MAX_RETRIES in pcie-designware.h, it's too short
> >>> to configurate the endpoint.
> >>>> Now the patch to bypass PERST# is not necessary.
> >>>> Instead for DWC RC drivers, I think that the number of retries
> >>> should be changed according to the usage.
> >>> And the same issue remains with other RC drivers.
> >>
> >> If EP is configured using Linux, then PERST# cannot be used as it's difficult to boot linux and initialize EP within the specified time interval. Can't you prevent PERST from being propagated at all?
> > > Surely it might be difficult for RC to decide the time to wait for EP.
> > Since RC almost toggles PERST# in boot time, I'd like to think about
> > how to prevent from first PERST# at least.
>
> It can be prevented in the HW (If that's possible). I modify the cable connecting RC and EP to not propagate PERST#.

I understand. Although it's possible in case of a cable,
in case of an edge connector, EP side needs hardware mechanism not to propagate PERST#.

Thank you,

---
Best Regards,
Kunihiko Hayashi