PCIe EP should not be filtered by --skip-dp-nic-ms in Linux SBSA PCI_MSI_2 test.

[anncheng-arm]

PCI_MSI_2 is currently skipped in SystemReady-band Linux SBSA runs because the run uses --skip-dp-nic-ms, which filters out Network endpoints.

For this rule, that looks too aggressive.

PCI_MSI_2 checks whether allocated MSI/MSI-X vectors are unique across endpoint devices. It does not appear to depend on BAR memory access or the MMIO-style access patterns that typically motivate --skip-dp-nic-ms.

As a result, the current policy can reduce the candidate endpoint set so much that the test cannot form a comparison pair and ends up as SKIPPED, even when the platform has multiple PCIe endpoints with MSI/MSI-X capability.

@sunnywang-arm for vis.

[Sujana-M]

Hi @anncheng-arm,

PCI_MSI_2 collects MSI/MSI-X vector information through pal_get_msi_vectors(). For kernel version greater than 5.16, the code uses msi_for_each_desc() to iterate over the MSI descriptors, and that requires the descriptor list to be protected with msi_lock_descs()/__msi_lock_descs().
This locks protects the MSI descriptor storage of a device before operating on it. For the devices skipped (GP/NIC/MS), taking this lock causes an exception. Hence the devices are currently skipped.

Thanks,
Sujana
ACS Team

[anncheng-arm]

Hi @Sujana-M, thanks for the updates. I will close this one but create a new one that we may need a clear document/checklist for prebuild image/arm-systemready repo what could be default skip in automation flow.

@sunnywang-arm for vis.

[sunnywang-arm]

Hi @Sujana-M

Can we consider reporting WARNING instead of SKIPPED when running the test with --skip-dp-nic-ms? Also, update the message to “DP/NIC/MAS/RES device are skipped. Please individually run the test without --skip-dp-nic-ms to check the compliance”. if you all agree with this, I will raise issue for it.

Can we also review all tests that are affected by --skip-dp-nic-ms and see if we need to do the same for them?

• PCI_IN_13
• PCI_IN_19
• PCI_SM_02
• PCI_MM_01
• RE_REG_1, IE_REG_1, IE_REG_3
• RI_RST_1
• PCI_MM_03
• PCI_MSI_2

Cc @anncheng-arm @chetan-rathore

[SapthagiriP]

Hi @sunnywang-arm ,

below is my analysis of usage of -skip_dp_nic_ms in below rules
PCI_IN_13 : Both tests p005,004 write into the downstream BAR window and temporarily shrink the Root Port’s Type 1 memory window before doing anout‑of‑range access. Display/NIC endpoints can be touchy: a stray MMIO write may flip device state, and temporarily chopping the window can break ongoing traffic/DMA or even kill the console/network.
PCI_IN_19:P030.c decodes error by injecting the error condition by clearing Memory Space Enable (MSE) bit this can disrupt normal functions of devices like display port and nic

PCI_SM_02: does a full reset of the EP this could disrupt the DP and NIC

PCI_MM_01 : No skip done

RE_REG_1, IE_REG_1, IE_REG_3 : decodes error by injecting the error condition by clearing Memory Space Enable (MSE) bit this can disrupt normal functions of devices like display port and nic

RI_RST_1: Does function level reset. it needs skip because nic / dp can transfer packets continuously having this tested can disrupt the functionality

PCI_MM_03 : No skip

PCI_MSI_2: Skip Network Controller devices as locking access to MSI descriptors is causing an exception and for the device with base class codes greater than 13h as they are reserved.

and coming to modifying skip to warn for these rules when -skip_dp_nic_ms , we will implement the same.

[sunnywang-arm]

Thanks, @SapthagiriP
Your analysis for PCI_MM_01 and PCI_MM_03 are "no skip. Do you mean you will remove the --skip-dp-nic-ms for these two tests?
PCI_MM_01 https://github.com/ARM-software/sysarch-acs/blob/main/test_pool/pcie/p045.c#L113
PCI_MM_03 https://github.com/ARM-software/sysarch-acs/blob/main/test_pool/pcie/p094.c#L114

[SapthagiriP]

Hi @sunnywang-arm ,
I missed to take a look at these to tests somehow and my analysis for PCI_MM_01 and 03 are for equivalent exerciser tests.
Now coming to the tests mentioned.

Skip for dp, nic and ms is reqiured since the tests usess ioremap to remap the BAR with DEVICE_nGnRnE and NORMAL_NC then reads and writes an hardcoded data. which might produce an unknown result. So better to skip for DP,NIC and MS