xc_lowpri and xc_thread are racy and xc_wait may return during/before
executing all xcall callbacks, resulting in a kernel panic at worst.
xc_lowpri serializes multiple jobs by a mutex and a cv. If all xcall
callbacks are done, xc_wait returns and also xc_lowpri accepts a next job.
The problem is that a counter that counts the number of finished xcall
callbacks is incremented *before* actually executing a xcall callback
(see xc_tailp++ in xc_thread). So xc_lowpri accepts a next job before
all xcall callbacks complete and a next job begins to run its xcall callbacks.
Even worse the counter is global and shared between jobs, so if a xcall
callback of the next job completes, the shared counter is incremented,
which confuses wc_wait of the previous job as all xcall callbacks of the
previous job are done and wc_wait of the previous job returns during/before
executing its xcall callbacks.
How to fix: there are actually two counters that count the number of finished
xcall callbacks for low priority xcall for historical reasons (I guess):
xc_tailp and xc_low_pri.xc_donep. xc_low_pri.xc_donep is incremented correctly
while xc_tailp is incremented wrongly, i.e., before executing a xcall callback.
We can fix the issue by dropping xc_tailp and using only xc_low_pri.xc_donep.
PR kern/51632
Adjust pcu(9) to this xcall(9) change. This may fix the problems after
x86 FPU was converted to use PCU, since it avoids heavy contention at the
lower levels (particularly, IPL_SOFTNET). This is a good illustration why
software interrupts should generally avoid any blocking on locks.
tech-kern:
- Invert priority space so that zero is the lowest priority. Rearrange
number and type of priority levels into bands. Add new bands like
'kernel real time'.
- Ignore the priority level passed to tsleep. Compute priority for
sleep dynamically.
- For SCHED_4BSD, make priority adjustment per-LWP, not per-process.
