On Sat, 2018-07-14 at 10:37 +-0800, Ming Lei wrote: +AD4- OK, I am thinking another idea for addressing this issue. +AD4- +AD4- We may introduce one logical admin(pm) request queue for each scsi+AF8-device, +AD4- and this queue shares tag with IO queue, with NO+AF8-SCHED set, and always +AD4- use atomic mode of the queue usage refcounter. Then we may send PM +AD4- command to device after the IO queue is frozen. +AD4- +AD4- Also PREEMPT+AF8-ONLY can be removed too in this way. +AD4- +AD4- Even in future, all pass-through commands may be sent to this admin queue. +AD4- +AD4- If no one objects, I will cook patches towards this direction. Wouldn't it be overkill to introduce a separate admin queue for SCSI devices? Will that new queue be visible under /sys/block as a request queue? If so, how many scripts will break due to that sysfs change? If the new queue won't be visible under /sys/block, how many changes will have to be made to the block layer to avoid that it becomes visible? Regarding the device+AF8-busy, device+AF8-blocked, cmd+AF8-list and other attributes in struct scsi+AF8-device: will the SCSI commands submitted to the admin queue affect these attributes? If not, which assumptions in the SCSI core will break? And if SCSI commands submitted to the admin queue will affect these attributes, how will these attributes be shared between the regular and the admin request queues? Since the power management core guarantees that runtime PM operations do not occur during suspend nor during resume operations: has it been considered to drop the RQF+AF8-PM flag and to use the RQF+AF8-PREEMPT flag instead and also to set and reset QUEUE+AF8-FLAG+AF8-PREEMPT+AF8-ONLY as needed? Bart.
