Re: [PATCH 1/2] nvme: set io-scheduler requirement for ZNS

[Damien Le Moal]

On 2020/09/07 20:24, Kanchan Joshi wrote:
> On Mon, Sep 7, 2020 at 1:52 PM Damien Le Moal <Damien.LeMoal@xxxxxxx> wrote:
>>
>> On 2020/09/07 16:01, Kanchan Joshi wrote:
>>>> Even for SMR, the user is free to set the elevator to none, which disables zone
>>>> write locking. Issuing writes correctly then becomes the responsibility of the
>>>> application. This can be useful for settings that for instance use NCQ I/O
>>>> priorities, which give better results when "none" is used.
>>>
>>> Was it not a problem that even if the application is sending writes
>>> correctly, scheduler may not preserve the order.
>>> And even when none is being used, re-queue can happen which may lead
>>> to different ordering.
>>
>> "Issuing writes correctly" means doing small writes, one per zone at most. In
>> that case, it does not matter if the block layer reorders writes. Per zone, they
>> will still be sequential.
>>
>>>> As far as I know, zoned drives are always used in tightly controlled
>>>> environments. Problems like "does not know what other applications would be
>>>> doing" are non-existent. Setting up the drive correctly for the use case at hand
>>>> is a sysadmin/server setup problem, based on *the* application (singular)
>>>> requirements.
>>>
>>> Fine.
>>> But what about the null-block-zone which sets MQ-deadline but does not
>>> actually use write-lock to avoid race among multiple appends on a
>>> zone.
>>> Does that deserve a fix?
>>
>> In nullblk, commands are executed under a spinlock. So there is no concurrency
>> problem. The spinlock serializes the execution of all commands. null_blk zone
>> append emulation thus does not need to take the scheduler level zone write lock
>> like scsi does.
> 
> I do not see spinlock for that. There is one "nullb->lock", but its
> scope is limited to memory-backed handling.
> For concurrent zone-appends on a zone, multiple threads may set the
> "same" write-pointer into incoming request(s).
> Are you referring to any other spinlock that can avoid "same wp being
> returned to multiple threads".

Checking again, it looks like you are correct. nullb->lock is indeed only used
for processing read/write with memory backing turned on.
We either need to extend that spinlock use, or add one to protect the zone array
when doing zoned commands and checks of read/write against a zone wp.
Care to send a patch ? I can send one too.


-- 
Damien Le Moal
Western Digital Research