[BUG] PCI: spacemit-k1: port C probe hard-hangs a CPU with one endpoint on Milk-V Jupiter

[Bruno Banelli <bbanelli@gmail.com>]

Hi,

On a Milk-V Jupiter, probing the PCIe controller at ca800000 (port C, the
card slot) permanently wedges the CPU that runs the probe when one
particular add-in card is installed.  The CPU stops responding to NMI, and
because the probe is asynchronous, kernel_init() then blocks forever in
async_synchronize_full() and the machine never finishes booting.

The same card, in the same slot, on the same board, does *not* hang the
vendor 6.6 kernel -- it reports "Phy link never came up" and boots normally.
Six other cards do not hang mainline either.  So whatever the electrical
cause, this looks like a robustness problem in pcie-spacemit-k1: an endpoint
should not be able to hang a host-side DBI register access.


HARDWARE
--------
  Milk-V Jupiter v1.1, SpacemiT M1 (socinfo: CPU[M1-8571] REV[C] DRO[130]),
  16 GiB LPDDR4X.
  Firmware: stock vendor U-Boot 2022.10 (k1-bl-v2.2.9), unmodified.
  Port B (ca400000, M.2) has a Samsung PM9B1 NVMe and works throughout.
  Port C (ca800000) is the card slot -- an x8-length connector, silkscreened
  PCIE_X2, wired x2.


REPRODUCED ON
-------------
  v7.1 and v7.2, riscv defconfig (plus PHY_SPACEMIT_K1_USB2, USB_DWC3,
  SPACEMIT_K1_TSENSOR, IGB, IGC, NVMe/ext4 built in).
  gcc 13.3.0 (cross) and gcc 16.2.0 (native, Debian sid).
  Identical failure in all combinations.  Not a regression -- port C has
  never worked with this card on mainline.

Command line:
  console=ttyS0,115200 earlycon root=/dev/nvme0n1p2 rootwait rw
  swiotlb=65536 clk_ignore_unused pd_ignore_unused


SYMPTOM
-------
Port C prints its address ranges and then never speaks again.  (Log below is
from a run with port B disabled in DT, so nothing is interleaved.)

  [ 1.290074] spacemit-k1-pcie ca800000.pcie: host bridge /soc/pcie-bus/pcie@ca800000 ranges:
  [ 1.297283] spacemit-k1-pcie ca800000.pcie:       IO 0x00b7002000..0x00b7101fff -> 0x0000000000
  [ 1.312783] spacemit-k1-pcie ca800000.pcie:      MEM 0x00a0000000..0x00afffffff -> 0x00a0000000
  [ 1.326753] spacemit-k1-pcie ca800000.pcie:      MEM 0x00b0000000..0x00b6ffffff -> 0x00b0000000
  [22.348490] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
  [22.351764] rcu:     4-...0: (12 GPs behind) idle=051c/1/0x4000000000000000 softirq=43/43 fqs=1908
  [22.367040] Sending NMI from CPU 2 to CPUs 4:
  [32.367049] After 10 seconds, these CPUS still haven't responded to the NMI: 4

The CPU ignoring an NMI for ten seconds is why I read this as an MMIO access
that never receives a completion rather than a spin or a deadlock.


LOCALISATION
------------
I added a dev_info() before each step of k1_pcie_init() (patch at the end of
this mail).  The last marker port C prints is the one immediately before the
first DBI access:

  [1.347050] spacemit-k1-pcie ca800000.pcie: K1DBG 1 toggle_soft_reset
  [1.362635] spacemit-k1-pcie ca800000.pcie: K1DBG 2 enable_resources
  [1.370918] spacemit-k1-pcie ca800000.pcie: K1DBG 3 first DBI write (vendor/device ID)
  <nothing further>

i.e. it dies in

        dw_pcie_dbi_ro_wr_en(pci);
        dw_pcie_writew_dbi(pci, PCI_VENDOR_ID, PCI_VENDOR_ID_SPACEMIT);

which is the first register access to the controller after
k1_pcie_enable_resources() has enabled the clocks and deasserted the resets.

I also tried moving phy_init() ahead of that block, in case the DBI domain
needed the PHY running (port B is masked here, because vendor U-Boot
initialises port B's PHY and never touches port C's).  phy_init() returned
success and the DBI access still hung:

  K1DBG 1 toggle_soft_reset
  K1DBG 2 enable_resources
  K1DBG 4 assert PERST# + 100ms
  K1DBG 5 RC mode + AUX_PWR_DET
  K1DBG 6 phy_init
  K1DBG 3 first DBI write (vendor/device ID)     <- still the last line


CARD MATRIX (port C, mainline 7.2, otherwise identical boots)
-------------------------------------------------------------
  empty slot                                          boots
  Intel I210        [8086:1533]  Gen1 x1   link up, enumerates
  HP NC360T         [8086:105e]  Gen1 x2   link up, enumerates
  NVIDIA Quadro P400[10de:1cb3]  Gen1 x2   link up, enumerates
  NVIDIA Quadro T400[10de:1fb2]  Gen1 x1   link up, enumerates
  AMD Radeon RX 550 [1002:699f]  Gen1 x2   link up, enumerates
  Sun ATLS1QGE                             "Device found, but not active"
  HP NC375T                                "Device found, but not active"
  Intel I225-V rev 01                      *** CPU HANG ***

Note the two cards that do not train fail *politely* -- the DWC core logs
"Device found, but not active", the probe completes, an empty bus 0002:00 is
created and the machine boots.  So port C is perfectly capable of handling a
link that never comes up.  The I225-V is different, and it fails long before
link training is reached.

The I225-V card itself is good: it works on a MACCHIATObin (Armada 8040) in
the same office, and it works on this same Jupiter under the vendor kernel.


THE VENDOR DRIVER SURVIVES THE SAME CARD
----------------------------------------
Bianbu 2.3.5 / Linux 6.6.63, vendor k1x-dwc-pcie driver, same board, same
slot, same I225-V:

  [ 1.657482] k1x-dwc-pcie ca400000.pcie: PCIe Gen.2 x2 link up
  [ 2.949762] k1x-dwc-pcie ca800000.pcie: Phy link never came up
  [ 2.952808] k1x-dwc-pcie ca800000.pcie: PCI host bridge to bus 0002:00
  ... boots to a login prompt

The vendor driver does not do the "set the PCI vendor and device ID" DBI
write at that point in its sequence.


HYPOTHESES ELIMINATED BY EXPERIMENT
-----------------------------------
  nvme driver          initcall_blacklist=nvme_init  -- still hangs
  power domains        pd_ignore_unused              -- still hangs
  link training        dies before dw_pcie_iatu_detect()
  kernel version       identical on v7.1 and v7.2
  compiler             identical with gcc 13.3 and gcc 16.2
  PHY init ordering    phy_init() first -- returns 0, still hangs
  CLKREQ# pinmux       CLKREQ# removed from pcie2_4_cfg -- still hangs
  port B interference  port B disabled in DT -- still hangs
  power supply         12 V / 12.5 A bench supply, not USB-C PD


A POSSIBLY RELATED OBSERVATION
------------------------------
k1_pcie_init() writes PCI_VENDOR_ID_SPACEMIT / PCI_DEVICE_ID_SPACEMIT_K1 to
both ports.  On this board the write takes effect on port B but not on port C
-- including on the boots where port C works fine:

  pci 0001:00:00.0: [201f:0001] type 01 class 0x060400 PCIe Root Port
  pci 0002:00:00.0: [1e5d:3003] type 01 class 0x060400 PCIe Root Port

1e5d:3003 is the hardware default (ASR Microelectronics).  So that same DBI
read-only write is being silently dropped on port C even when it does not
hang.  I do not know whether this is the same underlying issue, but it is in
the same function and on the same port, so it seemed worth mentioning.


WORKAROUND
----------
        &pcie2 { status = "disabled"; };

or, from U-Boot, before booti:

        fdt set /soc/pcie-bus/pcie@ca800000 status disabled

With that, mainline 7.2 boots Debian happily on this board with root on the
M.2 NVMe.  Nobody's board is stuck; the slot is just unusable with this card.

Note this is not a problem with the DTS enablement of &pcie2 (added in 7.1) --
an empty slot, and six of seven cards, work fine.


WHAT I AM ASKING
----------------
I do not have the K1 documentation, so I have gone as far as I can from
outside.  I would appreciate a pointer to what could gate that first DBI
access on port C -- CLK_PCIE2_DBI/MASTER/SLAVE, RESET_PCIE2_*, or something
in the PMU/APMU block -- and I am happy to run any test you like on this
board.

Separately, and regardless of the cause: a hard CPU hang on an unanswered
DBI read is an unpleasant failure mode, since there is no completion timeout
and no machine check to abort it on RISC-V.  If there is a sane way to bound
it, that seems worth having.

Full logs for every boot referenced above are available on request.

Thanks,
Bruno Banelli


--------------------------------------------------------------------------------
Instrumentation used for the localisation above (not for merging):

diff --git a/drivers/pci/controller/dwc/pcie-spacemit-k1.c b/drivers/pci/controller/dwc/pcie-spacemit-k1.c
index 04241df8f..d4c08e021 100644
--- a/drivers/pci/controller/dwc/pcie-spacemit-k1.c
+++ b/drivers/pci/controller/dwc/pcie-spacemit-k1.c
@@ -130,17 +130,21 @@ static int k1_pcie_init(struct dw_pcie_rp *pp)
 {
 	struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
 	struct k1_pcie *k1 = to_k1_pcie(pci);
+	struct device *dev = pci->dev;
 	u32 reset_ctrl;
 	u32 val;
 	int ret;
 
+	dev_info(dev, "K1DBG 1 toggle_soft_reset\n");
 	k1_pcie_toggle_soft_reset(k1);
 
+	dev_info(dev, "K1DBG 2 enable_resources\n");
 	ret = k1_pcie_enable_resources(k1);
 	if (ret)
 		return ret;
 
 	/* Set the PCI vendor and device ID */
+	dev_info(dev, "K1DBG 3 first DBI write (vendor/device ID)\n");
 	dw_pcie_dbi_ro_wr_en(pci);
 	dw_pcie_writew_dbi(pci, PCI_VENDOR_ID, PCI_VENDOR_ID_SPACEMIT);
 	dw_pcie_writew_dbi(pci, PCI_DEVICE_ID, PCI_DEVICE_ID_SPACEMIT_K1);
@@ -153,6 +157,7 @@ static int k1_pcie_init(struct dw_pcie_rp *pp)
 	 * delay first.  Write, then read it back to guarantee the write
 	 * reaches the device before we start the delay.
 	 */
+	dev_info(dev, "K1DBG 4 assert PERST# + 100ms\n");
 	reset_ctrl = k1->pmu_off + PCIE_CLK_RESET_CONTROL;
 	regmap_set_bits(k1->pmu, reset_ctrl, PCIE_RC_PERST);
 	regmap_read(k1->pmu, reset_ctrl, &val);
@@ -162,8 +167,10 @@ static int k1_pcie_init(struct dw_pcie_rp *pp)
 	 * Put the controller in root complex mode, and indicate that
 	 * Vaux (3.3v) is present.
 	 */
+	dev_info(dev, "K1DBG 5 RC mode + AUX_PWR_DET\n");
 	regmap_set_bits(k1->pmu, reset_ctrl, DEVICE_TYPE_RC | PCIE_AUX_PWR_DET);
 
+	dev_info(dev, "K1DBG 6 phy_init\n");
 	ret = phy_init(k1->phy);
 	if (ret) {
 		k1_pcie_disable_resources(k1);
@@ -172,11 +179,14 @@ static int k1_pcie_init(struct dw_pcie_rp *pp)
 	}
 
 	/* Deassert fundamental reset (drive PERST# high) */
+	dev_info(dev, "K1DBG 7 deassert PERST#\n");
 	regmap_clear_bits(k1->pmu, reset_ctrl, PCIE_RC_PERST);
 
 	/* Finally, as a workaround, disable ASPM L1 */
+	dev_info(dev, "K1DBG 8 disable_aspm_l1\n");
 	k1_pcie_disable_aspm_l1(k1);
 
+	dev_info(dev, "K1DBG 9 init complete\n");
 	return 0;
 }