Oddities in brd queue limits

Oddities in brd queue limits

[Hannes Reinecke]

Hi Christoph,

brd ends up with the following queue limits:

optimal_io_size: 0
minimum_io_size: 4096
hw_sector_size: 512
physical_block_size: 4096

which I find particularly odd; how can the minimum I/O size be _larger_ 
than the hw_sector_size? Wouldn't that imply that we can only send I/O
in units of physical block size, rendering the hw_sector_size pretty 
much pointless?

Or what is the idea here?

Btw, I would have expected brd to set 'optimal_io_size' to 4k, and
minimum_io_size to 512 bytes. Which would've been an alternative fix.

Cheers,

Hannes

Re: Oddities in brd queue limits

[Christoph Hellwig]

On Tue, Apr 02, 2024 at 03:17:26PM +0200, Hannes Reinecke wrote:
> Hi Christoph,
>
> brd ends up with the following queue limits:
>
> optimal_io_size: 0
> minimum_io_size: 4096
> hw_sector_size: 512
> physical_block_size: 4096
>
> which I find particularly odd; how can the minimum I/O size be _larger_ 
> than the hw_sector_size? Wouldn't that imply that we can only send I/O
> in units of physical block size, rendering the hw_sector_size pretty much 
> pointless?

The minimum_io_size is always larger or equal to hw sector size.
It really is the minimal efficient I/O size.

Re: Oddities in brd queue limits

[Hannes Reinecke]

On 4/2/24 15:18, Christoph Hellwig wrote:
> On Tue, Apr 02, 2024 at 03:17:26PM +0200, Hannes Reinecke wrote:
>> Hi Christoph,
>>
>> brd ends up with the following queue limits:
>>
>> optimal_io_size: 0
>> minimum_io_size: 4096
>> hw_sector_size: 512
>> physical_block_size: 4096
>>
>> which I find particularly odd; how can the minimum I/O size be _larger_
>> than the hw_sector_size? Wouldn't that imply that we can only send I/O
>> in units of physical block size, rendering the hw_sector_size pretty much
>> pointless?
> 
> The minimum_io_size is always larger or equal to hw sector size.
> It really is the minimal efficient I/O size.
> 

So is it a hard limit (as in: we cannot send I/O smaller than that)
or a soft limit (as in: we should not send I/O smaller than that)?

Cheers,

Hannes

Re: Oddities in brd queue limits

[Christoph Hellwig]

On Tue, Apr 02, 2024 at 03:21:00PM +0200, Hannes Reinecke wrote:
> So is it a hard limit (as in: we cannot send I/O smaller than that)
> or a soft limit (as in: we should not send I/O smaller than that)?

It is a completely soft hint that isn't used by anything in the
kernel I/O path.

(as a quick grep would tell..)