i.MX8M Plus PCIe link regression

[Hongxing Zhu]

> -----Original Message-----
> From: Manivannan Sadhasivam <mani at kernel.org>
> Sent: 2026年3月4日 0:42
> To: Hongxing Zhu <hongxing.zhu at nxp.com>; Alexander Stein
> <alexander.stein at ew.tq-group.com>
> Cc: linux-pci at vger.kernel.org; linux-arm-kernel at lists.infradead.org;
> imx at lists.linux.dev
> Subject: Re: i.MX8M Plus PCIe link regression
> 
> On Tue, Mar 03, 2026 at 01:38:13PM +0100, Alexander Stein wrote:
> > Hi,
> >
> > these days I noticed that there is a PCIe link regression on my
> > i.MX8MP platform (TQMa8MPxL) running next-20260227.
> > I could bisect it back to commit 9c03e30e3ade3 ("PCI: imx6: Skip link
> > up workaround for newer platforms"). I always get the following errors:
> >
> > imx6q-pcie 33800000.pcie: Link failed to come up. LTSSM:
> > CFG_LINKWD_START imx6q-pcie 33800000.pcie: probe with driver
> > imx6q-pcie failed with error -110
> >
> > Connected is a Gen1 PCIe -> Ethernet adapter. Interestingly a Gen2
> > device is detected without issues.
> >
> > Reverting 3c96a61dd2e098dda8dcac3dce3d38a3c87afbfc and
> > b9a8d28ebbf118bb3eac953f4a37abbd341257ab "fixes" my platform, both
> Gen
> > 1 and Gen 2 devices are detect. Commit
> > 3c96a61dd2e098dda8dcac3dce3d38a3c87afbfc is only required for conflict
> free revert.
> >
> > Here is a summary with outputs:
> > Gen 1 device
> > > 00:00.0 PCI bridge: Synopsys, Inc. DWC_usb3 / PCIe bridge (rev 01)
> > > 01:00.0 Ethernet controller: Realtek Semiconductor Co., Ltd.
> > > RTL8111/8168/8211/8411 PCI Express Gigabit Ethernet Controller (rev
> > > 01)
> >
> > Gen 2 device
> > > 00:00.0 PCI bridge: Synopsys, Inc. DWC_usb3 / PCIe bridge (rev 01)
> > > 01:00.0 SATA controller: Marvell Technology Group Ltd. 88SE9128 PCIe
> > > SATA 6 Gb/s RAID controller (rev 20)
> >
> > output of "dmesg | grep imx6q-pcie". Common part for all tests:
> > > imx6q-pcie 33800000.pcie: host bridge /soc at 0/pcie at 33800000 ranges:
> > > imx6q-pcie 33800000.pcie:       IO 0x001ff80000..0x001ff8ffff ->
> > > 0x0000000000
> > > imx6q-pcie 33800000.pcie:      MEM 0x0018000000..0x001fefffff ->
> > > 0x0018000000
> > > imx6q-pcie 33800000.pcie: config reg[1] 0x1ff00000 == cpu 0x1ff00000
> > > imx6q-pcie 33800000.pcie: iATU: unroll T, 4 ob, 4 ib, align 64K,
> > > limit 16G
> >
> > next-20260227
> > Gen 1
> > > imx6q-pcie 33800000.pcie: Link failed to come up. LTSSM:
> > > CFG_LINKWD_START imx6q-pcie 33800000.pcie: probe with driver
> > > imx6q-pcie failed with error -110
> >
> > Gen 2
> > > imx6q-pcie 33800000.pcie: PCIe Gen.2 x1 link up imx6q-pcie
> > > 33800000.pcie: PCI host bridge to bus 0000:00
> >
> > next-20260227 + reverts
> > Gen 1
> > > imx6q-pcie 33800000.pcie: PCIe Gen.1 x1 link up imx6q-pcie
> > > 33800000.pcie: PCIe Gen.1 x1 link up imx6q-pcie 33800000.pcie: PCI
> > > host bridge to bus 0000:00
> >
> > Gen 2
> > > imx6q-pcie 33800000.pcie: PCIe Gen.1 x1 link up imx6q-pcie
> > > 33800000.pcie: PCIe Gen.2 x1 link up imx6q-pcie 33800000.pcie: PCI
> > > host bridge to bus 0000:00
> >
> > What can we do here? I'm wondering why Gen 2 trains correctly, while
> > Gen 1 doesn't.
> >
> 
> Thanks for the report and sorry for the breakage. Commit 9c03e30e3ade3,
> mentions that the workaround is only needed for the i.MX6 platforms. So
> I'm not sure why the patch breaks i.MX8M and that too only for Gen 1
> speed.
> 
> Also, before 9c03e30e3ade3, LTSSM was always started in Gen 1, but as per
> your finding, only Gen 1 devices fail to work. So the code till
> imx_pcie_wait_for_speed_change() shouldn't have an impact. Maybe the
> issue is due to skipping imx_pcie_wait_for_speed_change()?
> 
> Richard, thoughts?
Hi Maini:
Prior to calling imx_pcie_wait_for_speed_change(), the PCIe link must
 already be established. Before commit 9c03e30e3ade3, the speed change
 procedure was only initiated after the initial Gen1 link was successfully
 brought up.

Hi Alexander:
Sorry for the breakage.
Before commit 9c03e30e3ade3, i.MX8MP PCIe forced Gen1 link training
 initially, then switched to Gen3 after link-up by triggering a speed
 change. After commit 9c03e30e3ade3, link training starts directly at
 the Gen3 capability level.

The two link training methods differ only in the i.MX8MP Root Complex (RC)
 link capability configuration prior to link training: one method forces
 Gen1 speed, while the other retains the default Gen3 setting in your tests.

Fortunately, I also have a RTL8168E Gigabit network card on hand.
Based on v7.0-rc1, here is the tests log on my i.MX8MP EVK board and one M.2
Key E to PCIe standard slot exchange adaptor.

Logs:
root at imx8mpevk:~# dmesg | grep pci
[    2.172420] imx6q-pcie 33800000.pcie: host bridge /soc at 0/pcie at 33800000 ranges:
[    2.172516] imx6q-pcie 33800000.pcie:       IO 0x001ff80000..0x001ff8ffff -> 0x0000000000
[    2.172545] imx6q-pcie 33800000.pcie:      MEM 0x0018000000..0x001fefffff -> 0x0018000000
[    2.172631] imx6q-pcie 33800000.pcie: config reg[1] 0x1ff00000 == cpu 0x1ff00000
[    2.380065] imx6q-pcie 33800000.pcie: iATU: unroll T, 4 ob, 4 ib, align 64K, limit 16G
[    2.579604] imx6q-pcie 33800000.pcie: PCIe Gen.1 x1 link up
[    2.580020] imx6q-pcie 33800000.pcie: PCI host bridge to bus 0000:00
<snipped>
[    2.583653] pci 0000:01:00.0: [10ec:8168] type 00 class 0x020000 PCIe Endpoint
[    2.583858] pci 0000:01:00.0: BAR 0 [io  0x0000-0x00ff]
[    2.583885] pci 0000:01:00.0: BAR 2 [mem 0x00000000-0x00000fff 64bit]
[    2.583902] pci 0000:01:00.0: BAR 4 [mem 0x00000000-0x00003fff 64bit pref]
[    2.583914] pci 0000:01:00.0: ROM [mem 0x00000000-0x0001ffff pref]
[    2.584280] pci 0000:01:00.0: supports D1 D2
[    2.584320] pci 0000:01:00.0: PME# supported from D0 D1 D2 D3hot D3cold
[    2.595615] pci 0000:01:00.0: ASPM: default states L0s L1
root at imx8mpevk:~#
root at imx8mpevk:~# lspci
00:00.0 PCI bridge: Synopsys, Inc. DWC_usb3 / PCIe bridge (rev 01)
01:00.0 Ethernet controller: Realtek Semiconductor Co., Ltd. RTL8111/8168/8211/8411 PCI Express Gigabit Ethernet Controller (rev 06)

Best Regards
Richard
> 
> - Mani
> 
> --
> மணிவண்ணன் சதாசிவம்