Re: [PATCH 0/2] nvmet: add namespace-level debugfs for reservation state

在 2026/7/6 20:57, Shin'ichiro Kawasaki 写道:
> On Jul 06, 2026 / 18:28, Guixin Liu wrote:
>>
>> 在 2026/7/6 15:59, Daniel Wagner 写道:
>>> On Mon, Jul 06, 2026 at 11:10:41AM +0800, Guixin Liu wrote:
>>>>>>> Example output with two registered hosts and an active holder:
>>>>>>>
>>>>>>>      $ cat /sys/kernel/debug/nvmet/testnqn/ns1/reservation
>>>>>>>      enable : 1
>>>>>>>      generation : 2
>>>>>>>      notify_mask : 0x0
>>>>>>>      rtype : write_exclusive
>>>>>>>      holder : hostid=11111111-1111-1111-1111-111111111111, rkey=0x1111
>>>>>>>      registrants:
>>>>>>>        hostid=11111111-1111-1111-1111-111111111111, rkey=0x1111
>>>>>>>        hostid=22222222-2222-2222-2222-222222222222, rkey=0x2222
>>>>>> I expect blktests will make use of this interface eventually. The parser
>>>>>> for this wont be dead simple, sure doable but not simple. If you see a
>>>>>> way to make a bit simpler I wouldn't mind.
>>>>> How about this:
>>>>>
>>>>> 1. One holer, multi registrants:
>>>>>     enable=1
>>>>>     generation=2
>>>>>     notify_mask=0x0
>>>>>     rtype=write_exclusive
>>>>>     holder=11111111-1111-1111-1111-111111111111,0x1111
>>>>>     reg=11111111-1111-1111-1111-111111111111,0x1111
>>>>>     reg=22222222-2222-2222-2222-222222222222,0x2222
>>>>>
>>>>> 2. No holder, multi registrants:
>>>>>     enable=1
>>>>>     generation=1
>>>>>     notify_mask=0x0
>>>>>     rtype=none
>>>>>     holder=none
>>>>>     reg=11111111-1111-1111-1111-111111111111,0x1111
>>>>>     reg=22222222-2222-2222-2222-222222222222,0x2222
>>>>>
>>>>> 3. No registrant:
>>>>>     enable=1
>>>>>     generation=0
>>>>>     notify_mask=0x0
>>>>>     rtype=none
>>>>>     holder=none
>>>>>
>>>>> 4. Reservation not enabled:
>>>>>     enable=0
> ...
>> Hi Shin'ichiro:
>>      Does this work for you?
> Hi Guixin, yes, I think it is good enough. If this is sysfs, "one value
> per file" and "array by files in a directory" would be guided, but this is
> debugfs readonly data. So, it sounds reasonable to put some data structure into
> a single debufs attribute file. Uevent like key=value format sounds good. It is
> a bit wacky to use comma to hold both host id and key for a holder and a
> registrant, but I cannot think of better solution.
OK, I will change this in v2, thanks.

Best Regards,
Guixin Liu