[ltt-dev] [RFC git tree] Userspace RCU (urcu) for Linux (repost)
Mathieu Desnoyers
compudj at krystal.dyndns.org
Sun Feb 8 19:40:27 EST 2009
* Paul E. McKenney (paulmck at linux.vnet.ibm.com) wrote:
> On Sun, Feb 08, 2009 at 04:46:10PM -0500, Mathieu Desnoyers wrote:
> > * Paul E. McKenney (paulmck at linux.vnet.ibm.com) wrote:
> > > On Sat, Feb 07, 2009 at 06:38:27PM -0500, Mathieu Desnoyers wrote:
> > > > * Paul E. McKenney (paulmck at linux.vnet.ibm.com) wrote:
> > > > > On Fri, Feb 06, 2009 at 08:34:32AM -0800, Paul E. McKenney wrote:
> > > > > > On Fri, Feb 06, 2009 at 05:06:40AM -0800, Paul E. McKenney wrote:
> > > > > > > On Thu, Feb 05, 2009 at 11:58:41PM -0500, Mathieu Desnoyers wrote:
> > > > > > > > (sorry for repost, I got the ltt-dev email wrong in the previous one)
> > > > > > > >
> > > > > > > > Hi Paul,
> > > > > > > >
> > > > > > > > I figured out I needed some userspace RCU for the userspace tracing part
> > > > > > > > of LTTng (for quick read access to the control variables) to trace
> > > > > > > > userspace pthread applications. So I've done a quick-and-dirty userspace
> > > > > > > > RCU implementation.
> > > > > > > >
> > > > > > > > It works so far, but I have not gone through any formal verification
> > > > > > > > phase. It seems to work on paper, and the tests are also OK (so far),
> > > > > > > > but I offer no guarantee for this 300-lines-ish 1-day hack. :-) If you
> > > > > > > > want to comment on it, it would be welcome. It's a userland-only
> > > > > > > > library. It's also currently x86-only, but only a few basic definitions
> > > > > > > > must be adapted in urcu.h to port it.
> > > > > > > >
> > > > > > > > Here is the link to my git tree :
> > > > > > > >
> > > > > > > > git://lttng.org/userspace-rcu.git
> > > > > > > >
> > > > > > > > http://lttng.org/cgi-bin/gitweb.cgi?p=userspace-rcu.git;a=summary
> > > > > > >
> > > > > > > Very cool!!! I will take a look!
> > > > > > >
> > > > > > > I will also point you at a few that I have put together:
> > > > > > >
> > > > > > > git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/perfbook.git
> > > > > > >
> > > > > > > (In the CodeSamples/defer directory.)
> > > > > >
> > > > > > Interesting approach, using the signal to force memory-barrier execution!
> > > > > >
> > > > > > o One possible optimization would be to avoid sending a signal to
> > > > > > a blocked thread, as the context switch leading to blocking
> > > > > > will have implied a memory barrier -- otherwise it would not
> > > > > > be safe to resume the thread on some other CPU. That said,
> > > > > > not sure whether checking to see whether a thread is blocked is
> > > > > > any faster than sending it a signal and forcing it to wake up.
> > > > > >
> > > > > > Of course, this approach does require that the enclosing
> > > > > > application be willing to give up a signal. I suspect that most
> > > > > > applications would be OK with this, though some might not.
> > > > > >
> > > > > > Of course, I cannot resist pointing to an old LKML thread:
> > > > > >
> > > > > > http://lkml.org/lkml/2001/10/8/189
> > > > > >
> > > > > > But I think that the time is now right. ;-)
> > > > > >
> > > > > > o I don't understand the purpose of rcu_write_lock() and
> > > > > > rcu_write_unlock(). I am concerned that it will lead people
> > > > > > to decide that a single global lock must protect RCU updates,
> > > > > > which is of course absolutely not the case. I strongly
> > > > > > suggest making these internal to the urcu.c file. Yes,
> > > > > > uses of urcu_publish_content() would then hit two locks (the
> > > > > > internal-to-urcu.c one and whatever they are using to protect
> > > > > > their data structure), but let's face it, if you are sending a
> > > > > > signal to each and every thread, the additional overhead of the
> > > > > > extra lock is the least of your worries.
> > > > > >
> > > > > > If you really want to heavily optimize this, I would suggest
> > > > > > setting up a state machine that permits multiple concurrent
> > > > > > calls to urcu_publish_content() to share the same set of signal
> > > > > > invocations. That way, if the caller has partitioned the
> > > > > > data structure, global locking might be avoided completely
> > > > > > (or at least greatly restricted in scope).
> > > > > >
> > > > > > Of course, if updates are rare, the optimization would not
> > > > > > help, but in that case, acquiring two locks would be even less
> > > > > > of a problem.
> > > > > >
> > > > > > o Is urcu_qparity relying on initialization to zero? Or on the
> > > > > > fact that, for all x, 1-x!=x mod 2^32? Ah, given that this is
> > > > > > used to index urcu_active_readers[], you must be relying on
> > > > > > initialization to zero.
> > > > > >
> > > > > > o In rcu_read_lock(), why is a non-atomic increment of the
> > > > > > urcu_active_readers[urcu_parity] element safe? Are you
> > > > > > relying on the compiler generating an x86 add-to-memory
> > > > > > instruction?
> > > > > >
> > > > > > Ditto for rcu_read_unlock().
> > > > > >
> > > > > > Ah, never mind!!! I now see the __thread specification,
> > > > > > and the keeping of references to it in the reader_data list.
> > > > > >
> > > > > > o Combining the equivalent of rcu_assign_pointer() and
> > > > > > synchronize_rcu() into urcu_publish_content() is an interesting
> > > > > > approach. Not yet sure whether or not it is a good idea. I
> > > > > > guess trying it out on several applications would be the way
> > > > > > to find out. ;-)
> > > > > >
> > > > > > That said, I suspect that it would be very convenient in a
> > > > > > number of situations.
> > > > > >
> > > > > > o It would be good to avoid having to pass the return value
> > > > > > of rcu_read_lock() into rcu_read_unlock(). It should be
> > > > > > possible to avoid this via counter value tricks, though this
> > > > > > would add a bit more code in rcu_read_lock() on 32-bit machines.
> > > > > > (64-bit machines don't have to worry about counter overflow.)
> > > > > >
> > > > > > See the recently updated version of CodeSamples/defer/rcu_nest.[ch]
> > > > > > in the aforementioned git archive for a way to do this.
> > > > > > (And perhaps I should apply this change to SRCU...)
> > > > > >
> > > > > > o Your test looks a bit strange, not sure why you test all the
> > > > > > different variables. It would be nice to take a test duration
> > > > > > as an argument and run the test for that time.
> > > > > >
> > > > > > I killed the test after better part of an hour on my laptop,
> > > > > > will retry on a larger machine (after noting the 18 threads
> > > > > > created!). (And yes, I first tried Power, which objected
> > > > > > strenously to the "mfence" and "lock; incl" instructions,
> > > > > > so getting an x86 machine to try on.)
> > > > > >
> > > > > > Again, looks interesting! Looks plausible, although I have not 100%
> > > > > > convinced myself that it is perfectly bug-free. But I do maintain
> > > > > > a healthy skepticism of purported RCU algorithms, especially ones that
> > > > > > I have written. ;-)
> > > > >
> > > > > OK, here is one sequence of concern...
> > > > >
> > > >
> > > > Let's see..
> > > >
> > > > > o Thread 0 starts rcu_read_lock(), picking up the current
> > > > > get_urcu_qparity() into the local variable urcu_parity().
> > > > > Assume that the value returned is zero.
> > > > >
> > > > > o Thread 0 is now preempted.
> > > > >
> > > > > o Thread 1 invokes urcu_publish_content():
> > > > >
> > > > > o It substitutes the pointer.
> > > > >
> > > > > o It forces all threads to execute a memory barrier
> > > > > (thread 0 runs just long enough to process its signal
> > > > > and then is immediately preempted again).
> > > > >
> > > > > o It switches the parity, which is now one.
> > > > >
> > > > > o It waits for all readers on parity zero, and there are
> > > > > none, because thread 0 has not yet registered itself.
> > > > >
> > > > > o It therefore returns the old pointer. So far, so good.
> > > > >
> > > > > o Thread 0 now resumes:
> > > > >
> > > > > o It increments its urcu_active_readers[0].
> > > > >
> > > > > o It forces a compiler barrier.
> > > > >
> > > > > o It returns zero (why not store this in thread-local
> > > > > storage rather than returning?).
> > > > >
> > > >
> > > > To support nested rcu_read_locks. (that's the only reason)
> > >
> > > A patch below to allow nested rcu_read_lock() while keeping to the Linux
> > > kernel API, just FYI. One can argue that the overhead of accessing the
> > > extra per-thread variables is offset by the fact that there no longer
> > > needs to be a return value from rcu_read_lock() nor an argument to
> > > rcu_read_unlock(), but hard to say.
> > >
> >
> > I ran your modified version within my benchmarks :
> >
> > with return value : 14.164 cycles per read
> > without return value : 16.4017 cycles per read
> >
> > So we have a 14% performance decrease due to this. We also pollute the
> > branch prediction buffer and we add a cache access due to the added
> > variables in the TLS. Returning the value has the clear advantage of
> > letting the compiler keep it around in registers or on the stack, which
> > clearly costs less.
> >
> > So I think the speed factor outweights the visual considerations. Maybe
> > we could switch to something like :
> >
> > unsigned int qparity;
> >
> > urcu_read_lock(&qparity);
> > ...
> > urcu_read_unlock(&qparity);
> >
> > That would be a bit like local_irq_save() in the kernel, except that we
> > could do it in a static inline because we pass the address. I
> > personnally dislike the local_irq_save() way of hiding the fact that it
> > writes to the variable in a "clever" macro. I'd really prefer to leave
> > the " & ".
> >
> > What is your opinion ?
>
> My current opinion is that I can avoid the overflow problem and the
> need to recheck, which might get rid of the need for both arguments
> and return values while still maintaining good performance. The trick
> is to use only the topmost bit for the grace-period counter, and all
> the rest of the bits for nesting. That way, no matter what value of
> global counter one picks up, it will be waited for (since there are but
> two values that the global counter takes on).
>
> But just now coding it, so will see if it actually works.
>
I look forward to see and test it.
> > > > > o It enters its critical section, obtaining a reference
> > > > > to the new pointer that thread 1 just published.
> > > > >
> > > > > o Thread 1 now again invokes urcu_publish_content():
> > > > >
> > > > > o It substitutes the pointer.
> > > > >
> > > > > o It forces all threads to execute a memory barrier,
> > > > > including thread 0.
> > > > >
> > > > > o It switches the parity, which is now zero.
> > > > >
> > > > > o It waits for all readers on parity one, and there are
> > > > > none, because thread 0 has registered itself on parity
> > > > > zero!!!
> > > > >
> > > > > o Thread 1 therefore returns the old pointer.
> > > > >
> > > > > o Thread 1 frees the old pointer, which thread 0 is still
> > > > > using!!!
> > > > >
> > > >
> > > > Ah, yes, you are right.
> > > >
> > > > > So, how to fix? Here are some approaches:
> > > > >
> > > > > o Make urcu_publish_content() do two parity flips rather than one.
> > > > > I use this approach in my rcu_rcpg, rcu_rcpl, and rcu_rcpls
> > > > > algorithms in CodeSamples/defer.
> > > >
> > > > This approach seems very interesting.
> > >
> > > Patch in earlier email. ;-)
> > >
> > > > > o Use a single free-running counter, in a manner similar to rcu_nest,
> > > > > as suggested earlier. This one is interesting, as I rely on a
> > > > > read-side memory barrier to handle the long-preemption case.
> > > > > However, if you believe that any thread that waits several minutes
> > > > > between executing adjacent instructions must have been preempted
> > > > > (which the memory barriers that are required to do a context
> > > > > switch), then a compiler barrier suffices. ;-)
> > > >
> > > > Hrm, I'm trying to figure out what kind of memory backend you need to
> > > > put your counters for each quiescent state period. Is this free-running
> > > > counter indexing a very large array ? I doubt it does. Then how does it
> > > > make sure we don't roll back to the old array entries ?
> > >
> > > There is no array, just a global counter that is incremented by a modest
> > > power of two for each grace period. Then the outermost rcu_read_lock()
> > > records the one greater than current value of the global counter in its
> > > per-thread variable.
> > >
> > > Now, rcu_read_lock() can tell that it is outermost by examining the
> > > low-order bits of its per-thread variable -- if these bits are zero,
> > > then this is the outermost rcu_read_lock(). So if rcu_read_lock() sees
> > > that it is nested, it simply increments its per-thread counter.
> > >
> > > Then rcu_read_unlock() simply decrements its per-thread variable.
> > >
> > > If the counter is only 32 bits, it is subject to overflow. In that case,
> > > it is necessary to check for the counter having been incremented a huge
> > > number of times between the time the outermost rcu_read_lock() fetched
> > > the counter value and the time that it stored into its per-thread
> > > variable.
> > >
> > > An admittedly crude implementation of this approach may be found in
> > > CodeSamples/defer/rcu_nest.[hc] in:
> > >
> > > git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/perfbook.git
> > >
> > > Of course, if the counter is 64 bits, overflow can safely be ignored.
> > > If you have a grace period every microsecond and allow RCU read-side
> > > critical sections to be nested 255 deep, it would take more than 2,000
> > > years to overflow. ;-)
> > >
> >
> > Looking at the code, my first thought is : if we find out that the
> > array-based solution and the counter-based solution have the same
> > performance, I would definitely prefer the array-based version because
> > there are far less overflow considerations. It's therefore more solid
> > algorithmically and can be proven formally.
> >
> > Also, I'm not sure I fully understand where your overflow test is going.
> > So let's pretend we are a reader, nested inside other rcu read locks,
> > and we arrive much later after the outermost reader has read the
> > rcu_gp_ctr. After 255 increments actually :
> >
> > static void rcu_read_lock(void)
> > {
> > long tmp;
> > long *rrgp;
> >
> > /*
> > * If this is the outermost RCU read-side critical section,
> > * copy the global grace-period counter. In either case,
> > * increment the nesting count held in the low-order bits.
> > */
> >
> > rrgp = &__get_thread_var(rcu_reader_gp);
> > retry:
> > tmp = *rrgp;
> > # we read the local rrgp
> > if ((tmp & RCU_GP_CTR_NEST_MASK) == 0)
> > tmp = rcu_gp_ctr;
> > # not executed, innermost and nested.
> > tmp++;
> > *rrgp = tmp;
> > # increment the local count and write it to the local rrgp
> > smp_mb();
> > if (((tmp & RCU_GP_CTR_NEST_MASK) == 1) &&
> > ((rcu_gp_ctr - tmp) > (RCU_GP_CTR_NEST_MASK << 8)) != 0) {
> > (*rrgp)--;
> > goto retry;
> > # If we are more than 255 increments away from rcu_gp_ctr, decrement
> > # rrgp and loop
> > }
> > }
> >
> > The problem is : rcu_gp_ctr is advancing. So if we have tmp stucked at a
> > given value, and we are nested over the outermost read lock (therefore
> > we are making it impossible to go end its execution), then when the
> > rcu_gp_crt will advance (which is the only way things can eventually go
> > forward, because the local rrgp is set back to its original value), we
> > are just going to be _farther_ away from it (not closer). So we'll have
> > to wait for a complete type overflow (will take a while on 32-bits, and
> > a very long while on 64-bits) to have the test returning false and then
> > going forward.
> >
> > Or there might be something I misunderstood ?
>
> The first clause of the "if" statement should prevent this -- if we are
> not the outermost rcu_read_lock(), then we never retry. (If I understand
> your scenario.)
>
Ah, yes. The if (((tmp & RCU_GP_CTR_NEST_MASK) == 1) && tests if this
is the outermost read lock due to the tmp++. My mistake.
> > > > This latter solution could break jump-based probing of programs
> > > > soon-to-be available in gcc. The probes are meant to be of short
> > > > duration, but the fact is that this design lets the debugger inject code
> > > > without resorting to a breakpoint, which might therefore break your
> > > > "short time between instructions" assumption. It's very unlikely, but
> > > > possible.
> > >
> > > But would the debugger's code injection take more than a minute without
> > > doing a context switch? Ah -- you are thinking of a probe that spins
> > > for several minutes. Yes, this would be strange, but not impossible.
> > >
> > > OK, so for this usage, solution 1 it is!
> >
> > Yes, it's unlikely, but possible.. and I like to design things assuming
> > the worse case scenario, even if it's almost impossible.
>
> That is indeed the only way to get even semi-reliable software!
>
Yes. By the way, I just committed the "duration" modification to
rcu_test.c. I also added some debugging which calls sched_yield() either
for the reader, the writer, or both. I also integrated some randomness
to leave some going quickly and others slowly.
Mathieu
> Thanx, Paul
>
> > Mathieu
> >
> > > > > Of course, the probability of seeing this failure during test is quite
> > > > > low, since it is unlikely that thread 0 would run just long enough to
> > > > > execute its signal handler. However, it could happen. And if you were
> > > > > to adapt this algorithm for use in a real-time application, then priority
> > > > > boosting could cause this to happen naturally.
> > > >
> > > > Yes. It's not because we are not able to create the faulty condition
> > > > that it will _never_ happen. It must therefore be taken care of.
> > >
> > > No argument here!!! ;-) See the earlier patch for one way to fix.
> > >
> > > The following patch makes rcu_read_lock() back into a void function
> > > while still permitting nesting, for whatever it is worth.
> > >
> > > Signed-off-by: Paul E. McKenney <paulmck at linux.vnet.ibm.com>
> > > ---
> > >
> > > test_urcu.c | 6 +++---
> > > urcu.c | 2 ++
> > > urcu.h | 40 ++++++++++++++++++++++++----------------
> > > 3 files changed, 29 insertions(+), 19 deletions(-)
> > >
> > > diff --git a/test_urcu.c b/test_urcu.c
> > > index db0b68c..16b212b 100644
> > > --- a/test_urcu.c
> > > +++ b/test_urcu.c
> > > @@ -33,7 +33,7 @@ static struct test_array *test_rcu_pointer;
> > >
> > > void *thr_reader(void *arg)
> > > {
> > > - int qparity, i, j;
> > > + int i, j;
> > > struct test_array *local_ptr;
> > >
> > > printf("thread %s, thread id : %lu, pid %lu\n",
> > > @@ -44,14 +44,14 @@ void *thr_reader(void *arg)
> > >
> > > for (i = 0; i < 100000; i++) {
> > > for (j = 0; j < 100000000; j++) {
> > > - qparity = rcu_read_lock();
> > > + rcu_read_lock();
> > > local_ptr = rcu_dereference(test_rcu_pointer);
> > > if (local_ptr) {
> > > assert(local_ptr->a == 8);
> > > assert(local_ptr->b == 12);
> > > assert(local_ptr->c[55] == 2);
> > > }
> > > - rcu_read_unlock(qparity);
> > > + rcu_read_unlock();
> > > }
> > > }
> > >
> > > diff --git a/urcu.c b/urcu.c
> > > index 1a276ce..95eea4e 100644
> > > --- a/urcu.c
> > > +++ b/urcu.c
> > > @@ -23,6 +23,8 @@ pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
> > > int urcu_qparity;
> > >
> > > int __thread urcu_active_readers[2];
> > > +int __thread urcu_reader_nesting;
> > > +int __thread urcu_reader_parity;
> > >
> > > /* Thread IDs of registered readers */
> > > #define INIT_NUM_THREADS 4
> > > diff --git a/urcu.h b/urcu.h
> > > index 9431da5..6d28ea2 100644
> > > --- a/urcu.h
> > > +++ b/urcu.h
> > > @@ -70,6 +70,8 @@ static inline void atomic_inc(int *v)
> > > extern int urcu_qparity;
> > >
> > > extern int __thread urcu_active_readers[2];
> > > +extern int __thread urcu_reader_nesting;
> > > +extern int __thread urcu_reader_parity;
> > >
> > > static inline int get_urcu_qparity(void)
> > > {
> > > @@ -79,26 +81,32 @@ static inline int get_urcu_qparity(void)
> > > /*
> > > * returns urcu_parity.
> > > */
> > > -static inline int rcu_read_lock(void)
> > > +static inline void rcu_read_lock(void)
> > > {
> > > - int urcu_parity = get_urcu_qparity();
> > > - urcu_active_readers[urcu_parity]++;
> > > - /*
> > > - * Increment active readers count before accessing the pointer.
> > > - * See force_mb_all_threads().
> > > - */
> > > - barrier();
> > > - return urcu_parity;
> > > + int urcu_parity;
> > > +
> > > + if (urcu_reader_nesting++ == 0) {
> > > + urcu_parity = get_urcu_qparity();
> > > + urcu_active_readers[urcu_parity]++;
> > > + urcu_reader_parity = urcu_parity;
> > > + /*
> > > + * Increment active readers count before accessing the pointer.
> > > + * See force_mb_all_threads().
> > > + */
> > > + barrier();
> > > + }
> > > }
> > >
> > > -static inline void rcu_read_unlock(int urcu_parity)
> > > +static inline void rcu_read_unlock(void)
> > > {
> > > - barrier();
> > > - /*
> > > - * Finish using rcu before decrementing the pointer.
> > > - * See force_mb_all_threads().
> > > - */
> > > - urcu_active_readers[urcu_parity]--;
> > > + if (--urcu_reader_nesting == 0) {
> > > + barrier();
> > > + /*
> > > + * Finish using rcu before decrementing the pointer.
> > > + * See force_mb_all_threads().
> > > + */
> > > + urcu_active_readers[urcu_reader_parity]--;
> > > + }
> > > }
> > >
> > > extern void rcu_write_lock(void);
> > >
> >
> > --
> > Mathieu Desnoyers
> > OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
>
> _______________________________________________
> ltt-dev mailing list
> ltt-dev at lists.casi.polymtl.ca
> http://lists.casi.polymtl.ca/cgi-bin/mailman/listinfo/ltt-dev
>
--
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
More information about the lttng-dev
mailing list