lock against themselves, causing infinite spinning. Brian Feldman
found this problem when testing with Mozilla and supplied the fix,
which I have revised slightly.
Here is the failure scenario. A thread calls dlopen() and acquires
the writer lock. While the thread still holds the lock, a signal
is delivered and caught. The signal handler tries to call a function
which hasn't been bound yet. It thus enters the dynamic linker
and tries to acquire the reader lock. Since the writer lock is
already held, it will spin forever in the signal handler. The
thread holding the lock won't be able to progress and release the
lock.
The solution is to block almost all signals while holding the
exclusive lock.
A similar problem could conceivably occur in the opposite order.
Namely, a thread is holding the reader lock and then a signal
handler calls dlopen() or dlclose() and spins waiting for the writer
lock. We deal with this administratively by proclaiming that signal
handlers aren't allowed to call dlopen() or dlclose(). Actually
we don't have to proclaim a thing, since signal handlers aren't
allowed to call any system functions except those which are explicitly
permitted.
Submitted by: Brian Fundakowski Feldman <green>
and for all (I hope). Packages such as wine, JDK, and linuxthreads
should no longer have any problems with re-entering the dynamic
linker.
This commit replaces the locking used in the dynamic linker with a
new spinlock-based reader/writer lock implementation. Brian
Fundakowski Feldman <green> argued for this from the very beginning,
but it took me a long time to come around to his point of view.
Spinlocks are the only kinds of locks that work with all thread
packages. But on uniprocessor systems they can be inefficient,
because while a contender for the lock is spinning the holder of the
lock cannot make any progress toward releasing it. To alleviate
this disadvantage I have borrowed a trick from Sleepycat's Berkeley
DB implementation. When spinning for a lock, the requester does a
nanosleep() call for 1 usec. each time around the loop. This will
generally yield the CPU to other threads, allowing the lock holder
to finish its business and release the lock. I chose 1 usec. as the
minimum sleep which would with reasonable certainty not be rounded
down to 0.
The formerly machine-independent file "lockdflt.c" has been moved
into the architecture-specific subdirectories by repository copy.
It now contains the machine-dependent spinlocking code. For the
spinlocks I used the very nifty "simple, non-scalable reader-preference
lock" which I found at
<http://www.cs.rochester.edu/u/scott/synchronization/pseudocode/rw.html>
on all CPUs except the 80386 (the specific CPU model, not the
architecture). The 80386 CPU doesn't support the necessary "cmpxchg"
instruction, so on that CPU a simple exclusive test-and-set lock
is used instead. 80386 CPUs are detected at initialization time by
trying to execute "cmpxchg" and catching the resulting SIGILL
signal.
To reduce contention for the locks, I have revamped a couple of
key data structures, permitting all common operations to be done
under non-exclusive (reader) locking. The only operations that
require exclusive locking now are the rare intrusive operations
such as dlopen() and dlclose().
The dllockinit() interface is now deprecated. It still exists,
but only as a do-nothing stub. I plan to remove it as soon as is
reasonably possible. (From the very beginning it was clearly
labeled as experimental and subject to change.) As far as I know,
only the linuxthreads port uses dllockinit(). This interface turned
out to have several problems. As one example, when the dynamic
linker called a client-supplied locking function, that function
sometimes needed lazy binding, causing re-entry into the dynamic
linker and a big looping mess. And in any case, it turned out to be
too burdensome to require threads packages to register themselves
with the dynamic linker.
just a few of them. This looks like it solves the recent
ld-elf.so.1: assert failed: /usr/src/libexec/rtld-elf/lockdflt.c:55
failures seen by some applications such as JDK.
init and fini functions. Now the code is very careful to hold no
locks when calling these functions. Thus the dynamic linker cannot
be re-entered with a lock already held.
Remove the tolerance for recursive locking that I added in revision
1.2 of dllockinit.c. Recursive locking shouldn't happen any more.
Mozilla and JDK users: I'd appreciate confirmation that things still
work right (or at least the same) with these changes.
locking functions. If an application loads a shared object with
dlopen() and the shared object has an init function which requires
lazy binding, then _rtld_bind is called when the thread is already
inside the dynamic linker. This leads to a recursive acquisition
of the lock, which I was not expecting -- hence the assert failure.
This work-around makes the default locking functions handle recursive
locking. It is NOT the correct fix -- that should be implemented
at the generic locking level rather than in the default locking
functions. I will implement the correct fix in a future commit.
Since the dllockinit() interface will likely need to change, warn
about that in both the man page and the header file.
functions to be used by the dynamic linker. This can be called by
threads packages at start-up time. I will add the call to libc_r
soon.
Also add a default locking method that is used up until dllockinit()
is called. The default method works by blocking SIGVTALRM, SIGPROF,
and SIGALRM in critical sections. It is based on the observation
that most user-space threads packages implement thread preemption
with one of these signals (usually SIGVTALRM).
The dynamic linker has never been reentrant, but it became less
reentrant in revision 1.34 of "src/libexec/rtld-elf/rtld.c".
Starting with that revision, multiple threads each doing lazy
binding could interfere with each other. The usual symptom was
that a symbol was falsely reported as undefined at start-up time.
It was rare but not unseen. This commit fixes it.
