runtime: simplify freedefer logic
Currently, freedefer's API forces a subtle and fragile situation. It
requires that the caller unlink the _defer from the G list, but
freedefer itself is responsible for zeroing some _defer fields. In the
window between these two steps, we have to prevent stack growth
because stack growth walks the defer list (which no longer contains
the unlinked defer) to adjust pointers, and would thus leave an
unadjusted and potentially invalid pointer behind in the _defer before
freedefer zeroes it.
This setup puts part of this subtle responsibility on the caller and
also means freedefer must be nosplit, which forces other shenanigans
to avoid nosplit overflows.
We can simplify all of this by replacing freedefer with a new popDefer
function that's responsible for both unlinking and zeroing the _defer,
in addition to freeing it.
Some history: prior to regabi, defer records contained their argument
frame, which deferreturn copied to the stack before freeing the defer
record (and subsequently running the defer). Since that argument frame
didn't have a valid stack map until we ran the deferred function, the
non-preemptible window was much larger and more difficult to isolate.
Now we use normal closure calls to capture defer state and call the
defer, so the non-preemptible window is narrowed to just the unlinking
step.
Change-Id: I7cf95ba18e1e2e7d73f616b9ed9fb38f5e725d72
Reviewed-on: https://go-review.googlesource.com/c/go/+/553696
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Auto-Submit: Austin Clements <austin@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>