Skip to contentSkip to conversation

Do any Alpha systems include InfiniBand?

DISCUSSION Paul E. McKenney 4 participants 7 messages Oct 20, 2017 – Oct 23, 2017 Alpha port

lore
Share & tools
lore
#0 lore raw

Hello, Michael,

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no". If I am wrong, please let me know, as I will otherwise interpret silence as assent. ;-)

Thanx, Paul

maintainer
#1 lore raw
Quoted 6 lines · Paul E. McKenney

Hello, Michael,

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no". If I am wrong, please let me know, as I will otherwise interpret silence as assent. ;-)

I'm not aware of any shipping with Infiniband, and I haven't heard of anyone putting Infiniband in an Alpha.

Why do you ask? :)

#2 lore raw
Quoted 9 lines · Matt Turner
Quoted 6 lines · Paul E. McKenney

Hello, Michael,

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no". If I am wrong, please let me know, as I will otherwise interpret silence as assent. ;-)

Quoted 4 lines

I'm not aware of any shipping with Infiniband, and I haven't heard of anyone putting Infiniband in an Alpha.

Why do you ask? :)

Because the Linux-kernel InfiniBand driver contains a bunch of instances of smp_read_barrier_depends(), which matter only on Alpha. Plus the InfiniBand maintainers have not responded to queries about what these smp_read_barrier_depends() instances are supposed to do. Now it might be that they are just busy, or it might be that they don't know themselves.

If Alpha doesn't use InfiniBand, they can simply be removed. On the other hand, if Alpha -did- use InfiniBand, it would be necessary to inspect the code to determine whether if READ_ONCE() needs to be added somewhere to make up for the removal of smp_read_barrier_depends().

The reason this came up to begin with is that Will Deacon is adding an smp_read_barrier_depends() to READ_ONCE() and Mark Rutland is converting all ACCESS_ONCE() instances to either READ_ONCE() or WRITE_ONCE(). With these changes, core code need not know about Alpha, and almost all instances of smp_read_barrier_depends() can be removed.

Hey, you asked! ;-)

Thanx, Paul

#3 lore raw
In reply to Paul E. McKenney
Quoted 2 lines · Paul E. McKenney

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no".

I am not aware of any Alpha systems with InfiniBand.

Cheers
Michael.

#4 lore raw
Quoted 5 lines · Michael Cree
Quoted 2 lines · Paul E. McKenney

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no".

I am not aware of any Alpha systems with InfiniBand.

Very good, then we need not worry about any Alpha-related consequences of removing smp_read_barrier_depends() invocations from the InfiniBand driver. ;-)

Thanx, Paul

#5 lore raw
Quoted 10 lines · Paul E. McKenney
Quoted 5 lines · Michael Cree
Quoted 2 lines · Paul E. McKenney

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no".

I am not aware of any Alpha systems with InfiniBand.

Very good, then we need not worry about any Alpha-related consequences of removing smp_read_barrier_depends() invocations from the InfiniBand driver. ;-)

But maybe make that driver depend on !CONFIG_ALPHA, so as to avoid planting any subtle ordering bugs?

Will

#6 lore raw
Quoted 6 lines · Will Deacon
Quoted 5 lines · Paul E. McKenney
Quoted 4 lines · Michael Cree
Quoted 2 lines · Paul E. McKenney

Do any of the DEC Alpha systems that run recent kernels have InfiniBand? Given my understanding of the history, I believe the answer to be "no".

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

I am not aware of any Alpha systems with InfiniBand.

Very good, then we need not worry about any Alpha-related consequences of removing smp_read_barrier_depends() invocations from the InfiniBand driver. ;-)

But maybe make that driver depend on !CONFIG_ALPHA, so as to avoid planting any subtle ordering bugs?

Good point, that does sound like a very sensible precaution!

Thanx, Paul