[PATCH v8 02/12] accel/rocket: wait for a running IRQ handler before resetting a core

[Jiaxing Hu]

Hi Igor,

I checked all of it against the tree and it holds.

The lockdep point goes in the commit message. That the wait is on a
waitqueue rather than a lock, so nothing would have reported the deadlock,
is a better argument for the placement than mine.

Masking before the sync, yes. INTERRUPT_MASK is armed in hw_submit and
cleared only in the hardirq, and rocket_reset never touches it, so on an
ordinary timeout it is live. Your line numbers are next-20260814 and mine
have the series on top, so here it is rocket_job.c:165 and :499. I agree
it is not a hole and that reset.pending already closes the resubmit
branch. I want the sentence the patch adds to be true on its own, not true
because something else prevents the case.

The runtime PM facts are right. rocket_job_is_idle is atomic_read on
credit_count, runtime_suspend returns -EBUSY only on that and then drops
the clocks, and drm_sched_stop zeroes the counter until drm_sched_start at
the end of the reset. The driver does claim idle for the whole body while
holding nothing, and the two puts differ the way you describe.

I am not folding that into 2/12. It changes behaviour in the shared path
instead of adding a fence, and on this SoC it meets a power domain that
cycles a bus reset on power-on, so it wants its own patch and a board run.
It is next in the reset path either way.

Your question. MMU_DTE_ADDR predates the rail by about a hundred rounds.
The entry after a timed out job has come back at one constant with that
reset error beside it since round 138, recorded in the paper and the
log-book, and the rail moved to domain-supply in round 241.

One near counterexample, since you would find it. The round before the one
in the cover letter timed a job out and logged no MMU_DTE_ADDR. It had no
job after the timeout, because the shape that times out runs last on
purpose, so the attach never happened. Your first bullet from the other
side.

I ran your test on the archive. The run with the two MMU_DTE_ADDR lines
has zero "stall request timed out" and zero "paging request timed out", so
the MMU is not responding at all rather than sitting in a wrong state.

The third bullet is what I will build. If put_noidle leaves the device
active with no idle request pending, the domain never drops between the
failed job and the next one, and the bus reset that 9/12 cycles on
power-on never gets cycled. That fits what I have, including the block
being fine after a reboot and not otherwise. The next image swaps
put_noidle for put_autosuspend, and separately forces a suspend and resume
before the next job, so the two do not confound. You get the result either
way.

Your aside is right and it is not mine. rocket_reset_work is defined,
INIT_WORK'd and never queued, and it is that way in the base this series
sits on, untouched by the twelve.

No tag wanted, for the reason you give.

Jiaxing