rtld: mask signals for default read locks

[kib]

Summary

Rtld locks from libthr defer signals delivery, which means that binding
is not possible while a signal handler is executed.

Binding might upgrade read-locked rtld_bind_lock to write-lock, if
symbol resolution requires loading filters.  If a signal would be delivered
while rtld is in read-locked section, and signal handler needs binding
which upgrades the lock, for non-threaded image that uses default rtld
locks, we get the rtld data structures modified under the top-level
active rtld frame.

To correct the problem, Mask signals for read-locking of default locks
in addition to the write-locking.  It is very cheap now with
sigfastblock(2).

Note that the global state is used to track pre-locked state of either
sigfastblock(2) or signal mask (if sigfastblock(2) is administratively
disabled).  It is fine for non-threaded images since there are no other
threads.  But I believe that it is fine for threaded images using libc_r
as well, since masking signals disables preemption (I did not tested
it).

NetBSD PR:      https://gnats.netbsd.org/56979
Reported by:    tmunro

[markj]

libexec/rtld-elf/rtld_lock.c

148

Is _rel really needed here?

[kib]

libexec/rtld-elf/rtld_lock.c

148

No. And _add_acq can be further optimized by removing _acq and only doing fence_acq when WAFLAG is not set, but I decided to not bother.

[kib]

Remove one _rel