Re: [PATCH 1/1] scsi: sd: use max_xfer_blocks for set rw_max if max_xfer_blocks is available

["Martin K. Petersen" <martin.petersen@oracle.com>]

Hello Changheun!

> I want to discuss using max_xfer_blocks instead of opt_xfer_blocks as
> a optional.  For example, device reports opt_xfer_blocks is 512KB and
> 1MB as a max_xfer_blocks too. Currently rw_max is set with 512KB only.

Because that's what the device asks for. If a device explicitly requests
us to use 512 KB I/Os we should not be sending it 1 MB requests.

The spec is very clear. It says that if you send a command *larger* than
opt_xfer_blocks, you should expect *slower* performance. That makes
max_xfer_blocks a particularly poor choice for setting the default I/O
size.

In addition to being slower, max_xfer_blocks could potentially also be
much, much larger than opt_xfer_blocks. I understand your 512 KB vs. 1
MB example. But if the max_xfer_blocks limit is reported as 1 GB, is
that then the right value to use instead of 512 KB? Probably not.

If a device does not report an opt_xfer_blocks value that suits your
workload, just override the resulting max_sectors_kb in sysfs. This is
intentionally a soft limit so it can be adjusted by the user without
having to change the kernel.

-- 
Martin K. Petersen	Oracle Linux Engineering