Re: [PATCH 1/1] NFSD: Fix SCSI reservation conflict causing pNFS client to revert I/O to MDS

[Dai Ngo <dai.ngo@oracle.com>]

On 10/9/25 12:46 PM, Benjamin Coddington wrote:
> On 9 Oct 2025, at 13:43, Dai Ngo wrote:
>
>> On 10/7/25 1:37 PM, Benjamin Coddington wrote:
>>> On 7 Oct 2025, at 11:59, Dai Ngo wrote:
>>>
>>>> On 10/1/25 10:36 AM, Dai Ngo wrote:
>>>>> On 10/1/25 3:54 AM, Benjamin Coddington wrote:
>>>>>> On 30 Sep 2025, at 17:41, Dai Ngo wrote:
>>>>>>
>>>>>>> Hi Ben,
>>>>>>>
>>>>>>> On 9/30/25 12:15 PM, Benjamin Coddington wrote:
>>>>>>>> Hi Dai,
>>>>>>>>
>>>>>>>> On 30 Sep 2025, at 12:28, Dai Ngo wrote:
>>>>>>>>
>>>>>>>>> When servicing the GETDEVICEINFO call from an NFS client, the NFS server
>>>>>>>>> creates a SCSI persistent reservation on the target device using the
>>>>>>>>> reservation type PR_EXCLUSIVE_ACCESS_REG_ONLY. This setting restricts
>>>>>>>>> device access so that only hosts registered with a reservation key can
>>>>>>>>> perform read or write operations. Any unregistered initiator is completely
>>>>>>>>> blocked, including standard SCSI commands such as READCAPACITY.
>>>>>>>> SBC-4, table 13 shows that READ CAPACITY should be allowed from any I_T
>>>>>>>> nexus, no matter the state of the reservation on the LU.
>>>>>>>>
>>>>>>>> Is it possible that your SCSI implementation might be out of the spec?  Also
>>>>>>>> possible that SBC-4 has been updated, I haven't been following the SCSI
>>>>>>>> specification updates..
>>>>>>>>
>>>>>>>> Ben
>>>>>>> I don't have access to SBC-4 spec, t10.org does not allow guest access
>>>>>>> to their docs. Can you please share the content of table 13 here?
>>>>>> The document's licensing prohibits me from doing this, I'm sorry to report.
>>>>>> I have a single-user copy that prohibits me from copying or transmitting any
>>>>>> part or whole.  Looks like you can get SBC-5 from the ANSI webstore for $60:
>>>>>>
>>>>>> https://urldefense.com/v3/__https://webstore.ansi.org/standards/incits/incits5712025__;!!ACWV5N9M2RV99hQ!N4FtetrpMVBPf88WPTlz6EuwsK0kPhNqw04MXvtXGUwMzzAf0NPkCYhL5HYx32ZZVogW2MKS0Jr8P8M$
>>>>>>
>>>>>> The reason your patch caught my eye was because we'd previously fixed the
>>>>>> same problem in the SCSI LIO target.
>>>>> Thank you Ben, I'll get the spec from the ANSI webstore.
>>>> You're right Ben! The SBC-4 spec says read capacity is allowed in this
>>>> case.
>>>>
>>>> The problem was caused by the DS was running an older version of the
>>>> kernel that did not have your fix:
>>>>
>>>> 28c58f8a0947f scsi: target: Enable READ CAPACITY for PR EARO
>>>>
>>>> This fix did not include the SERVICE_ACTION_IN_16 with Service Action
>>>> READ_CAPACITY. However, the Linux client tries SERVICE_ACTION_IN_16
>>>> three times then switches to READ CAPACITY (0x25).
>>>>
>>>> Thank you for pointing this out.
>>> Would you be willing to test this patch for SERVICE_ACTION_IN_16?
>>>
>>>   From d7fa5d5f593dcfe39b7456dd6f23eb042fb2636f Mon Sep 17 00:00:00 2001
>>> Message-ID: <d7fa5d5f593dcfe39b7456dd6f23eb042fb2636f.1759869410.git.bcodding@redhat.com>
>>> From: Benjamin Coddington <bcodding@redhat.com>
>>> Date: Tue, 7 Oct 2025 16:34:37 -0400
>>> Subject: [PATCH] scsi: target: Fixup two more cases for PR EARO
>>>
>>> Allow READ_CAPACITY_16 and REPORT_REFERALS for SERVICE_ACTION_IN_16
>>> in the SCSI target driver.
>>>
>>> Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
>>> ---
>>>    drivers/target/target_core_pr.c | 7 +++++++
>>>    1 file changed, 7 insertions(+)
>>>
>>> diff --git a/drivers/target/target_core_pr.c b/drivers/target/target_core_pr.c
>>> index 83e172c92238..0b6803754422 100644
>>> --- a/drivers/target/target_core_pr.c
>>> +++ b/drivers/target/target_core_pr.c
>>> @@ -465,6 +465,13 @@ static int core_scsi3_pr_seq_non_holder(struct se_cmd *cmd, u32 pr_reg_type,
>>>                           return -EINVAL;
>>>                   }
>>>                   break;
>>> +       case SERVICE_ACTION_IN_16:
>>> +               switch (cdb[1] & 0x1f) {
>>> +               case SAI_READ_CAPACITY_16:
>>> +               case SAI_REPORT_REFERRALS:
>>> +                       ret = 0;
>>> +               }
>>> +               break;
>>>           case ACCESS_CONTROL_IN:
>>>           case ACCESS_CONTROL_OUT:
>>>           case INQUIRY:
>> The patch worked fine for READ_CAPACITY_16.
>>
>> The REPORT_REFERRALS got a Check Condition with sense code 0x2000
>> (Invalid command Operation code), whether there is a PR or not,
>> because the SCSI target does not support Referrals as reported by
>> 'sg_vpd' command.
>>
>> Thanks for doing this, Ben.
>>
>> -Dai
> Thanks for the test.  I'll send this to the target list, and will add your
> Tested-by unless you'd rather it withheld.

Please feel free to add me as a Tested-by tag.

Thanks,
-Dai