On 8/17/20 7:32 AM, Peter Lieven wrote:
Hi,
I am currently debugging a performance issue in qemu-img convert. I think I
have found the cause and will send a patch later.
But is there any reason why BDRV_REQUEST_MAX_SECTORS is not at least aligned
down to 8 (4k sectors)?
Any operation that is not able to determinate an optimal or maximum request
size from a driver will likely use this value and
might generate unaligned requests (as qemu-img convert does) as soon as the
first iteration of an operation is done.
Vladimir's work to make the block layer 64-bit clean should help resolve this,
as we will be relying more heavily on values aligned to the driver's limits
rather than arbitrary 32-bit cutoffs.