On 6/25/25 5:04 PM, Damien Le Moal wrote:
On 6/17/25 07:33, Bart Van Assche wrote:
A later patch will preserve the order of pipelined zoned writes by
submitting all zoned writes per zone to the same software queue as
previously submitted zoned writes. Hence support allocating a request
from a specific software queue.
Why is this needed ? All you need to do is schedule the zone write plug BIO work
on a specific CPU. Then the request used will naturally come from the software
queue of that CPU, no ? Am I missing something ?
Yes. That approach may be acceptable for rotating magnetic storage but
is not acceptable for UFS devices. Inserting a context switch in the hot
path would slow down UFS writes significantly.
Thanks,
Bart.
