If gcAssistAlloc is unable to steal or perform enough scan work, it
calls timeSleep, which allocates. If this allocation requires
obtaining a new span, it will in turn attempt to assist GC. Since
there's likely still no way to satisfy the assist, it will sleep
again, and so on, leading to potentially deep (not infinite, but also
not bounded) recursion.
Fix this by disallowing assists during the timeSleep.
This same problem was fixed on master by
65aa2da. That commit built on
several other changes and hence can't be directly cherry-picked. This
commit implements the same idea.
Fixes #12894.
Change-Id: I152977eb1d0a3005c42ff3985d58778f054a86d4
Reviewed-on: https://go-review.googlesource.com/15720
Reviewed-by: Rick Hudson <rlh@golang.org>
// more, so go around again after performing an
// interruptible sleep for 100 us (the same as the
// getfull barrier) to let other mutators run.
+
+ // timeSleep may allocate, so avoid recursive assist.
+ gcalloc := gp.gcalloc
+ gp.gcalloc = 0
timeSleep(100 * 1000)
+ gp.gcalloc = gcalloc
goto retry
}
}