Skip to contentSkip to conversation

Linux Preemptive patch success 2.4.10-pre4 + lots of other patches

DISCUSSION Christoph Lameter 5 participants 8 messages Sep 7, 2001 1 message not in the archive

lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive

Given the minimal nature of the patch I would suggest that it become part of 2.4.10 or 11

Quoted 5 lines · Robert Love

As for integrating into the kernel, there has been talk of preemption becoming standard for 2.5. This is one reason why I keep the patches up to date. When 2.5 starts, I will have a working patch that has seen some use and hopefully generated some benchmarks and discussions, and I can send a diff off to Linus. We shall see.

#2 lore raw
Quoted 2 lines · Christoph Lameter

Given the minimal nature of the patch I would suggest that it become part of 2.4.10 or 11

Are you kidding? We will be lucky to see this in during 2.5.

Its a pretty big change. It makes the Linux kernel preemptible. This
is a fairly big move, one I don't think any of the major Unices have done. The only reason the patch is not _huge_ is because the Linux kernel is already setup for concurrency of this nature -- it does SMP.

I suggest you read
http://www.linuxdevices.com/articles/AT4185744181.html
http://www.linuxdevices.com/articles/AT5152980814.html
http://kpreempt.sourceforge.net

and my previous threads on this issue, for more informaiton.

Signature · 3 lines
Robert M. Love
rml at ufl.edu
rml at tech9.net
#3 lore raw
In reply to Christoph Lameter
Quoted 6 lines · Robert Love
Quoted 2 lines · Christoph Lameter

Given the minimal nature of the patch I would suggest that it become part of 2.4.10 or 11

Are you kidding? We will be lucky to see this in during 2.5. Its a pretty big change. It makes the Linux kernel preemptible.

CONFIG_PREEMPT

Quoted 2 lines

This is a fairly big move, one I don't think any of the major Unices have done.

The other Unices are at least evenly split, or mostly preemptible. Typically, a more complex strategy is used where spinlocks can sleep after a few spins. This patch is very conservative in that regard, it basically just uses the structure we already have, SMP spinlocks.

Quoted 9 lines

The only reason the patch is not _huge_ is because the Linux kernel is already setup for concurrency of this nature -- it does SMP.

I suggest you read
http://www.linuxdevices.com/articles/AT4185744181.html
http://www.linuxdevices.com/articles/AT5152980814.html
http://kpreempt.sourceforge.net

and my previous threads on this issue, for more informaiton.

Hmm, how did you read those and come to such a different conclusion?

--
Daniel

#4 lore raw
Quoted 9 lines · Daniel Phillips
Quoted 6 lines · Robert Love
Quoted 2 lines · Christoph Lameter

Given the minimal nature of the patch I would suggest that it become part of 2.4.10 or 11

Are you kidding? We will be lucky to see this in during 2.5. Its a pretty big change. It makes the Linux kernel preemptible.

CONFIG_PREEMPT

and... ?

Quoted 7 lines
Quoted 2 lines

This is a fairly big move, one I don't think any of the major Unices have done.

The other Unices are at least evenly split, or mostly preemptible. Typically, a more complex strategy is used where spinlocks can sleep after a few spins. This patch is very conservative in that regard, it basically just uses the structure we already have, SMP spinlocks.

I did not know other Unices were (in general) preemptible. Solaris is? The only one I thought was preemptible was Irix.

Anyhow, you are right about the simplistic approach we take. There are a few alternatives: mixing mutexes and shorter locks, priority-bearing semaphores, changing the way the preemption count works, etc.

Quoted 11 lines
Quoted 9 lines

The only reason the patch is not _huge_ is because the Linux kernel is already setup for concurrency of this nature -- it does SMP.

I suggest you read
http://www.linuxdevices.com/articles/AT4185744181.html
http://www.linuxdevices.com/articles/AT5152980814.html
http://kpreempt.sourceforge.net

and my previous threads on this issue, for more informaiton.

Hmm, how did you read those and come to such a different conclusion?

What different conclusion? What are you even arguing with me about?

Do you think I am against a preemptible kernel? I _posted_ the damn patch, of course I am not.

I probably agree with whatever you are thinking.

Signature · 3 lines
Robert M. Love
rml at ufl.edu
rml at tech9.net
#6 lore raw
In reply to Christoph Lameter
Quoted 6 lines · Robert Love
Quoted 5 lines · Daniel Phillips
Quoted 4 lines · Robert Love
Quoted 2 lines · Christoph Lameter

Given the minimal nature of the patch I would suggest that it become part of 2.4.10 or 11

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

Are you kidding? We will be lucky to see this in during 2.5. Its a pretty big change. It makes the Linux kernel preemptible.

CONFIG_PREEMPT

and... ?

Sorry, I thought that was self-evident. It's not a change, it's an option.

Quoted 6 lines

What different conclusion? What are you even arguing with me about?

Do you think I am against a preemptible kernel? I _posted_ the damn patch, of course I am not.

I probably agree with whatever you are thinking.

Yes, violent agreement. OK, I disagree with your assessment that it's a huge change. Big in effect yes, but not in structural impact.

And you're right, I did think you were arguing against your own patch.

--
Daniel

#7 lore raw
Quoted 1 line · Daniel Phillips

Sorry, I thought that was self-evident. It's not a change, it's an option.

Oh, yes. It is completely an option -- although I would like to see it be the norm :)

Quoted 2 lines

Yes, violent agreement. OK, I disagree with your assessment that it's a huge change. Big in effect yes, but not in structural impact.

Right, actually I agree with you. That was my point about using SMP lock points. Since it uses the existing structure for SMP concurrency, the patch is small -- but its effects are pretty large. Its a fairly big deal.

Quoted 1 line

And you're right, I did think you were arguing against your own patch.

No, but I had to take that approach as it was suggested the patch be merged for 2.4.10!

I am glad you take a pro side to the preemption issue. Hopefully I can get some continued support and see some work towards inclusion in 2.5. Any help is appreciated.

Sorry for the confusion.

Signature · 3 lines
Robert M. Love
rml at ufl.edu
rml at tech9.net
#8 lore raw
In reply to Christoph Lameter

I'm using the 2.4.9-ac9 preempt patch to test things out. Everything is fine and stable except it seems that sudden and rapid network accesses cause tiny "pops" in the soundcard that didn't occur before the patch. This is not increased system load or anything so the audio isn't skipping due to that. It has only happened so far when grabbing many tiny files in rapid succession (note. only at 70K/s). Weird.

#5 lore raw
In reply to Robert Love
Quoted 8 lines · Robert Love
Quoted 4 lines · Daniel Phillips

The other Unices are at least evenly split, or mostly preemptible. Typically, a more complex strategy is used where spinlocks can sleep after a few spins. This patch is very conservative in that regard, it basically just uses the structure we already have, SMP spinlocks.

I did not know other Unices were (in general) preemptible. Solaris is? The only one I thought was preemptible was Irix.

Solaris is, and has been since good ol' Solaris 2.0. So's AIX and even more obscure things like DG/UX. I'd think all SysVR4 derivatives have to be, by virtue of the SysV schedular (threads have priorities from 0 to 159; system threads run from 60 to 99, so the schedular must be able to preempt system threads for the real-time threads from 100 to 159).

later,
chris

Signature · 2 lines
Chris Ricker                                               [email protected]
                                                          [email protected]