mm: critical shortage of bounce buffers

mm: critical shortage of bounce buffers

[Sergey Kubushin]

Can anybody explain what does "mm: critical shortage of bounce buffers"
mean?

I have a 2xP-III/850 system with 2Gbyte of RAM. I'm trying to run
ImageMagick on this system with quite big files (convert consumes 1+ Gbyte
of RAM). The system crushes immediately with that message in log file and a
whole screen of constantly scrolling allocation failure messages.

Should I change some kernel define to be able to use those 2 Gbytes?

It does crush even when given "mem=960M" boot option. Both 4 and 64 Gbytes
RAM configurations do crush. Work like a charm with high memory disabled.

---
Sergey Kubushin				Sr. Unix Administrator
CyberBills, Inc.			Phone:	702-567-8857
874 American Pacific Dr,		Fax:	702-567-8808
Henderson, NV 89014

mm: critical shortage of bounce buffers

[Steffen Persvold]

Hi list members,

I've recently encountered the following message on a machine running RedHat's
2.4.3-12 kernel :

"mm: critical shortage of bounce buffers"

I've searched through the kernel sources, but my 'find' just can't locate this
string anywhere.

Why does this message appear (apparently during high network load with the intel
eepro100 driver or e1000 driver). Is bounce buffers really in use on a x86
machine with 2GB of RAM (normal smp RedHat kernel, not enterprise)??

I appreciate any feedback.


Thanks,
-- 
  Steffen Persvold   | Scalable Linux Systems |   Try out the world's best   
 mailto:sp@scali.no  |  http://www.scali.com  | performing MPI implementation:
Tel: (+47) 2262 8950 |   Olaf Helsets vei 6   |      - ScaMPI 1.12.2 -         
Fax: (+47) 2262 8951 |   N0621 Oslo, NORWAY   | >300MBytes/s and <4uS latency

Re: mm: critical shortage of bounce buffers

[Alan Cox]

> I've recently encountered the following message on a machine running RedHat's
> 2.4.3-12 kernel :
> 
> "mm: critical shortage of bounce buffers"
> 
> I've searched through the kernel sources, but my 'find' just can't locate this
> string anywhere.

Its in the high mem handling routines. It means the machine stalled for
a moment doing I/O because it had no memory below 1Gb to use.

> Why does this message appear (apparently during high network load with the intel
> eepro100 driver or e1000 driver). Is bounce buffers really in use on a x86
> machine with 2GB of RAM (normal smp RedHat kernel, not enterprise)??

The answer is yes. You can actually build yourself a custom none bounce
buffer 1.8GB kernel with about 2Gb of user virtual space per app. For some
applications it will perform better.

Re: mm: critical shortage of bounce buffers

[Steffen Persvold]

Alan Cox wrote:
> 
> > I've recently encountered the following message on a machine running RedHat's
> > 2.4.3-12 kernel :
> >
> > "mm: critical shortage of bounce buffers"
> >
> > I've searched through the kernel sources, but my 'find' just can't locate this
> > string anywhere.
> 
> Its in the high mem handling routines. It means the machine stalled for
> a moment doing I/O because it had no memory below 1Gb to use.

But why does it need to have memory below 1Gb ?? Normally, 32bit PCI DMA
controllers (such as network cards and disk controllers) can access up to 4GB of
physical memory within the machine, so unless you are using the CONFIG_HIGHMEM4G
option it shouldn't need bounce buffers. Or am I missing something here
(something like the second physical GB is actually in the address range
4GB->)???

> 
> > Why does this message appear (apparently during high network load with the intel
> > eepro100 driver or e1000 driver). Is bounce buffers really in use on a x86
> > machine with 2GB of RAM (normal smp RedHat kernel, not enterprise)??
> 
> The answer is yes. You can actually build yourself a custom none bounce
> buffer 1.8GB kernel with about 2Gb of user virtual space per app. For some
> applications it will perform better.

There's no CONFIG option for that right (like the old CONFIG_2GB option on 2.2
kernels) ? I'll have to manually go in and edit the header files in asm-i386.
Would the 2.2 (e.g 2.2.19) kernel have the same problems with bounce buffers or
is this not implemented on 2.2 ??

Regards,
-- 
  Steffen Persvold   | Scalable Linux Systems |   Try out the world's best   
 mailto:sp@scali.no  |  http://www.scali.com  | performing MPI implementation:
Tel: (+47) 2262 8950 |   Olaf Helsets vei 6   |      - ScaMPI 1.12.2 -         
Fax: (+47) 2262 8951 |   N0621 Oslo, NORWAY   | >300MBytes/s and <4uS latency

Re: mm: critical shortage of bounce buffers

[Steffen Persvold]

Steffen Persvold wrote:
> 
> Alan Cox wrote:
> >
> > > I've recently encountered the following message on a machine running RedHat's
> > > 2.4.3-12 kernel :
> > >
> > > "mm: critical shortage of bounce buffers"
> > >
> > > I've searched through the kernel sources, but my 'find' just can't locate this
> > > string anywhere.
> >
> > Its in the high mem handling routines. It means the machine stalled for
> > a moment doing I/O because it had no memory below 1Gb to use.
> 
> But why does it need to have memory below 1Gb ?? Normally, 32bit PCI DMA
> controllers (such as network cards and disk controllers) can access up to 4GB of
> physical memory within the machine, so unless you are using the CONFIG_HIGHMEM4G
> option it shouldn't need bounce buffers.

Sorry  that should have been the CONFIG_HIGHMEM64G option. I think the
CONFIG_HIGHMEM4G option could manage without bounce buffers.

Regards,
-- 
  Steffen Persvold   | Scalable Linux Systems |   Try out the world's best   
 mailto:sp@scali.no  |  http://www.scali.com  | performing MPI implementation:
Tel: (+47) 2262 8950 |   Olaf Helsets vei 6   |      - ScaMPI 1.12.2 -         
Fax: (+47) 2262 8951 |   N0621 Oslo, NORWAY   | >300MBytes/s and <4uS latency

Re: mm: critical shortage of bounce buffers

[Alan Cox]

> > Its in the high mem handling routines. It means the machine stalled for
> > a moment doing I/O because it had no memory below 1Gb to use.
> 
> But why does it need to have memory below 1Gb ?? Normally, 32bit PCI DMA
> controllers (such as network cards and disk controllers) can access up to 4GB of
> physical memory within the machine, so unless you are using the CONFIG_HIGHMEM4G
> option it shouldn't need bounce buffers. Or am I missing something here
> (something like the second physical GB is actually in the address range
> 4GB->)???

The core code doesn't know if the device is PIO or DMA, and it cannot 
pass virtually unmapped objects to the block layer. This is what Jens has
been fixing.