[PATCH RFC 0/2] PCI: rockchip-host: Retry training and hot-reset a JMB585 on RK3399

[Shawn Lin]

在 2026/11/06 星期五 13:09, Igor Velkov 写道:
> Hi Shawn,
> 
> I traced the LTSSM on the Helios64 (RK3399 + JMB585, x2, 2.5 GT/s) by
> polling PCIE_CLIENT_DEBUG_OUT_0 about every 2 us during link training
> and the waits after it. This changes what the RFC said, so here are the
> data and a question.
> 
> A failing link bring-up looks like this:
> 
> - the LTSSM reaches L0 as fast as in a good one;
> - then it leaves L0 for Recovery (RcvrLock/RcvrCfg/Idle, about 5 us)
>    every 199..204 us, for as long as I watch: 488 times per 100 ms;
> - link_st in BASIC_STATUS1 becomes 3 only 15..47 ms after training
>    starts, or not within 700 ms; the latter is the "link training gen1
>    timeout" with LTSSM 0x10;
> - the RC shows nothing: AER and the device status are clear, and a dump
>    of the client, core management and root port config registers is
>    identical to a good bring-up.
> 
> A good bring-up has no Recovery entry after L0. A config read on a bad
> link gets no completion; that is the abort I reported. 200 us is the
> optional Flow Control Update timeout, so I suspect that one side does
> not see the DLLPs of the other.
> 
> The state survives Recovery but not Detect: a port re-init or a hot
> reset is just one more bring-up, with the same odds. So the hot reset
> of patch 2/2 does not cure anything, and the time the link takes to
> return is not a reliable sign either.
> 
> Unbind/bind cycles with a bad bring-up (a cycle is one training plus
> one hot reset):
> 
>    Helios64, JMB585, x2                     6 of 104
>    same, PHY refclk 100 MHz from NPLL       4 of 33
>    same, num-lanes = <1>                    0 of 104
>    ROCK Pi 4A, WD SN770 NVMe, x4            0 of 100
>    same, num-lanes = <2>                    0 of 100
> 
> So it takes both lanes and the JMB585.
> 
> Is anything like this known for the RK3399 with a two-lane endpoint,
> for example lane-to-lane skew of the transmitter, or a PHY setting that

I didn't know an issue like this. But yes, there is an option to disable
the automatic deskew adjusting algorithm by:
phy_wr_cfg(rk_phy, 0x6, BIT(1)); // dis auto deskew
phy_wr_cfg(rk_phy, 0x4, deskew trim value); // deskew value, 0x0 <-> 0xf
phy_wr_cfg(rk_phy, 0xb, deskew magnitude vale); //Default 00 is 2x, 01
is 1x, 11 is 3x and 10 turns it off


> affects it? And is there a bit in DEBUG_OUT_0/1 that tells which side
> directs the Recovery?
> 

These debug section lies on  Appendix section of PCIe chanpter, RK3399
TRM. You could try to find some clue by dumping DEBUG_DATA_OUT info.
Most kinds of these problems should be obvious to the PCIe analyzer
but might not be easy to check from the register IMO.

> Both boards ran the same build: v7.3.1 with the Armbian patches and the
> tracing. The Helios64 DT has no ep-gpios there, the ROCK Pi 4A DT has.
> 
> Thanks,
> Igor
>