nvme: fix the placement of set_current_state to TASK_INTERRUPTIBLE
PATCH 2/2 nvme: fix the placement of set_current_state to TASK_INTERRUPTIBLE
In reply to a message not in the archiveQuoted 11 lines · Arjan van de Ven
the nvme driver has a kthread that processes certain events periodically. However, the kthread also gets woken for certain urgent actions.
The current code does not use the
current_statelogic correctly; it callsset_current_state(TASK_INTERRUPTIBLE); right before doing aschedule_timeout(), with the result that there is a race condition where a wakeup can get lost, and thus delayed by one second.This patch moves the
set_current_state(TASK_INTERRUPTIBLE); to before the place where the queue of outstanding work is checked, to close the race condition
The thing is, that's not a queue of outstanding work, that's the list of devices that exist in the system. Your patch makes the kthread do all of its work in the TASK_INTERRUPTIBLE state, which I don't think is right either (is it?)
We could add a flag to indicate that there's urgent work to be done (checked after setting TASK_INTERRUPTIBLE), or just live with the occasional race.
Alternatively, we could get rid of the kthread in favour of timers for cancelling commands and a workqueue for dealing with urgent work. I have this on my long-term todo list, but in the absence of having hardware to test with on large-scale systems, I didn't want to make such a large change.
PATCH 2/2 nvme: fix the placement of set_current_state to TASK_INTERRUPTIBLE
Quoted 9 lines · Matthew Wilcox
Quoted 8 lines · Arjan van de Ven
the nvme driver has a kthread that processes certain events periodically. However, the kthread also gets woken for certain urgent actions.
The current code does not use the
current_statelogic correctly; it callsset_current_state(TASK_INTERRUPTIBLE); right before doing aschedule_timeout(), with the result that there is a race condition where a wakeup can get lost, and thus delayed by one second.
Quoted 8 lines
Quoted 3 lines
This patch moves the
set_current_state(TASK_INTERRUPTIBLE); to before the place where the queue of outstanding work is checked, to close the race conditionThe thing is, that's not a queue of outstanding work, that's the list of devices that exist in the system. Your patch makes the kthread do all of its work in the
TASK_INTERRUPTIBLEstate, which I don't think is right either (is it?)
eh why not?
it counts to load average, but frankly, you're using CPU... you count anyway