On Thu, 2019-11-21 at 08:54 +0800, Ming Lei wrote: > On Wed, Nov 20, 2019 at 12:00:19PM -0500, Ewan D. Milne wrote: > > On Wed, 2019-11-20 at 11:05 +0100, Hannes Reinecke wrote: > > > > > > Hmm. > > > > > > I must admit I patently don't like this explicit dependency on > > > blk_nonrot(). Having a conditional counter is just an open invitation to > > > getting things wrong... > > > > > > > This concerns me as well, it seems like the SCSI ML should have it's > > own per-device attribute if we actually need to control this per-device > > instead of on a per-host or per-driver basis. And it seems like this > > is something that is specific to high-performance drivers, so changing > > the way this works for all drivers seems a bit much. > > > > Ordinarily I'd prefer a host template attribute as Sumanesh proposed, > > but I dislike wrapping the examination of that and the queue flag in > > a macro that makes it not obvious how the behavior is affected. > > (Plus Hannes just submitted submitted the patches to remove .use_cmd_list, > > which was another piece of ML functionality used by only a few drivers.) > > > > Ming's patch does freeze the queue if NONROT is changed by sysfs, but > > the flag can be changed by other kernel code, e.g. sd_revalidate_disk() > > clears it and then calls sd_read_block_characteristics() which may set > > it again. So it's not clear to me how reliable this is. > > The queue freeze is applied in sd_revalidate_disk() too, isn't it? > Yes, sorry, you are right, your patch does add this. But if anything else changes the NONROT attribute for a queue associated with a SCSI device in the future it would have to freeze the queue. Because the device_busy counter mechanism would rely on it to work right. This isn't an obvious connection, it seems to me. -Ewan
