Skip to contentSkip to conversation

Robust Futex patches available

DISCUSSION David Singleton 1 participant 2 messages Nov 22, 2005 – Nov 24, 2005 1 message not in the archive

lore
Share & tools
lore
#0 lore raw

There are two new patches for Robust Futex support available at

http://source.mvista.com/~dsingleton

patch-2.6.14-rt13-rf3 fixes two locking bugs which caused hangs and
deadlocks.

patch-2.6.14-rt13-rf4 adds support for pthread_mutexes 'malloc'ed on
the heap.

I'd also like some advice on the direction POSIX is heading with
respect to
robust pthread_mutexes and priority inheritance.

It appears there are some not used openposix tests that use
different flags for
defining robustness. Here is a snip from the openposix robust
test's README:

Robust Mutex Tests
------------------------
The tests are under <rtnptl-tests>/robust_test directory.

rt-nptl supports 'robust' behavior, there will be two robust modes, one is PTHREAD_MUTEX_ROBUST_NP mode, the other is PTHREAD_MUTEX_ROBUST_SUN_NP mode. When the owner of a mutex dies in the first mode, the waiter will set the mutex to ENOTRECOVERABLE state, while in the second mode, the waiter needs to call pthread_mutex_setconsistency_np to change the state manually.

Currently the PTHREAD_MUTEX_ROBUST_NP is providing the fucntionality described by the PTHREAD_MUTEX_ROBUST_SUN_NP.

Any advice on which way we should go? I feel we should follow
POSIX and provide both methods and the new pthread_mutex_setconsistency_np function which provides the mutex recovery mechanism.

David

A message that is not in the archive (sent off-list or not yet received): [email protected]

#2 lore raw
Quoted 30 lines · Ulrich Drepper
Quoted 3 lines · David Singleton

I'd also like some advice on the direction POSIX is heading with
respect to
robust pthread_mutexes and priority inheritance.

Robust mutexes are not in POSIX nor have they been proposed for inclusion in the next revision. If a sane semantics can be determined I might submit it for inclusion.

As for priority inheritance, there is nothing to add, the feature is fully described.

Quoted 9 lines

rt-nptl supports 'robust' behavior, there will be two robust modes, one is PTHREAD_MUTEX_ROBUST_NP mode, the other is PTHREAD_MUTEX_ROBUST_SUN_NP mode. When the owner of a mutex dies in the first mode, the waiter will set the mutex to ENOTRECOVERABLE state, while in the second mode, the waiter needs to call pthread_mutex_setconsistency_np to change the state manually.

Currently the PTHREAD_MUTEX_ROBUST_NP is providing the fucntionality described by the PTHREAD_MUTEX_ROBUST_SUN_NP.

This description makes no sense. ENOTRECOVERABLE is the error code used if, surprise, the data/mutex cannot be recovered. I.e., it is *not* consistent. It is identical to calling pthread_mutex_unlock on a mutex which was locked when pthread_mutex_lock returned EDEADOWNER. All waiters are woken and subsequent locking attempts will fail in this case with ENOTRECOVERABLE.

Quoted 5 lines

So, if PTHREAD_MUTEX_ROBUST_NP really causes ENOTRECOVERABLE to be returned then this form is useless as it is trivially achieved with the PTHREAD_MUTEX_ROBUST_SUN_NP form. Plus: not defining it means less confusion since the semantics doesn't differ from Solaris (and there would be a bigger change to get the change added to POSIX).

Great, thanks for the help in understanding this. The ENOTRECOVERABLE made no sense to me either.

I'll keep things the way they are. The code returns -EOWNERDIED to the waiter/ new lock owner and the owner can recover the mutex, if they so choose.

Quoted 8 lines

I cannot really comment much on the kernel side. For my taste the robust futex functionality is impacting the far more commen code path too much. I would like to see much of the added code moved completely out of the fast path. Robust mutexes are slow, this won't make it worse.

Yes, even in the kernel I'd like to see the performance penalty for robustness be an option that the user can select if they want the trade off of robustness for a slower exit path.

Quoted 8 lines

The userlevel part is completely unacceptable. It's broken and what is there in code has the same problem as the kernel code: it punishes code which does not require this new feature. I don't worry about this part, though, since I would write this part myself in any case once there is an agreed upon kernel part.

Quoted 4 lines

Having this said, I certainly would like to get the robust futex part in. But not the priority handling part. The two are completely independent in principal. I don't agree at all with the current implementation of the priority inheritance.

I agree, the PI and Robustness are two logically separate functions and could easily be separated out. I'll start on that now.

Quoted 4 lines

So, if you split out the robust futex still, eliminate all the support added for priority inheritance, you get my full support for adding the code to mainline.

Excellent! I hope to have some patches for you in a while.

David