[UPDATE] Zerocopy, last one today I promise :-)

[UPDATE] Zerocopy, last one today I promise :-)

[David S. Miller]

Ok, it's pretty fast, but due to a critical bug fix, there's a new
one going up:

ftp://ftp.kernel.org/pub/linux/kernel/people/davem/zerocopy-2.4.1p10-3.diff.gz

o If sock_writepage is called on path via device without SG support,
  the cooked up sock_sendmsg() call needs to switch to KERNEL_DS.
  Discovered and fixed by Ingo Molnar.

This does show that not too many people are testing this all that
thoroughly :-) Basically, any sys_sendfile() over TCP using a network
card other than loopback/3c59x/sunhme/acenic would fail with -EFAULT
or even worse a kernel crash depending upon architecture.

Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[Tobias Ringstrom]

On Thu, 25 Jan 2001, David S. Miller wrote:
> This does show that not too many people are testing this all that
> thoroughly :-) Basically, any sys_sendfile() over TCP using a network
> card other than loopback/3c59x/sunhme/acenic would fail with -EFAULT
> or even worse a kernel crash depending upon architecture.

You may have said it before, but since you're the one who wants it tested,
I'm sure you're happy to repeat it, right? :-)

I understand from your comment that you want people to run it on all kinds
of hardware, both with and without hw checksumming, but how do you want us
to test it?  Is "my computer works as usual with this patch included" what
you are looking for, or do you want us to run specific tests or
benchmarks?

Should I get a speed increase in a normal TCP session, or do I have to use
sendfile to see any change?  Of course, stability is the most important
factor right now, but it would be nice if I will get a performance boost
from my old tired P90.  Right now it peaks at about 40 Mb/s for TCP.

/Tobias, a soon to be zerocopy patch tester

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[David S. Miller]

Tobias Ringstrom writes:
 > I understand from your comment that you want people to run it on all kinds
 > of hardware, both with and without hw checksumming, but how do you want us
 > to test it?  Is "my computer works as usual with this patch included" what
 > you are looking for, or do you want us to run specific tests or
 > benchmarks?

Basically, make use of the following:

1) TCP applications using normal write/sendmsg to send data
2) TCP applications using sys_sendfile to send data
   (f.e. pftpd or some other server which makes use of Linux's
    sendfile())
3) NFS client side activity

on both cards supporting sg+csum and those which do not.

Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[Tigran Aivazian]

On Thu, 25 Jan 2001, David S. Miller wrote:
> 3) NFS client side activity

this bit interesting (to me). Do you mean Linux NFS client as present in
the Linux kernel source (nfs filesystem) or just the type of traffic
generated by arbitrary NFS client (e.g. the builtin one used by SPEC SFS
benchmark suite)?

Regards,
Tigran

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[David S. Miller]

Tigran Aivazian writes:
 > On Thu, 25 Jan 2001, David S. Miller wrote:
 > > 3) NFS client side activity
 > 
 > this bit interesting (to me). Do you mean Linux NFS client as present in
 > the Linux kernel source (nfs filesystem) or just the type of traffic
 > generated by arbitrary NFS client (e.g. the builtin one used by SPEC SFS
 > benchmark suite)?

The normal NFS client in the kernel.... Performance should improve due
to the delayed defragmentation we do in the zerocopy patches.

Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[James H. Cloos Jr.]

What exaclty were the issues with the intel cards and sg+csum?

Any idea how much work it'd require to surmount them?

-JimC
-- 
James H. Cloos, Jr.  <http://jhcloos.com/public_key>     1024D/ED7DAEA6 
<cloos@jhcloos.com>  E9E9 F828 61A4 6EA9 0F2B  63E7 997A 9F17 ED7D AEA6
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[David S. Miller]

James H. Cloos Jr. writes:
 > What exaclty were the issues with the intel cards and sg+csum?
 > 
 > Any idea how much work it'd require to surmount them?

Getting Intel to release full specs on how to make use of
TX hardware checksum assist with the eepro100.

Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[Andi Kleen]

On Thu, Jan 25, 2001 at 08:40:16AM -0800, David S. Miller wrote:
> 
> James H. Cloos Jr. writes:
>  > What exaclty were the issues with the intel cards and sg+csum?
>  > 
>  > Any idea how much work it'd require to surmount them?
> 
> Getting Intel to release full specs on how to make use of
> TX hardware checksum assist with the eepro100.

It's halfway documented in their e100 driver source, but no real specs 
unfortunately.


-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[kuznet]

Hello!

>  > What exaclty were the issues with the intel cards and sg+csum?
>  > 
>  > Any idea how much work it'd require to surmount them?
> 
> Getting Intel to release full specs on how to make use of
> TX hardware checksum assist with the eepro100.

It simply does not exist for 82559* in all the steppings.
eepro100 is pretty poor device.

Probably, it exists for card identified as Gamla (D102) (82559 is D101).

Alexey
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[Chris Wedgwood]

On Thu, Jan 25, 2001 at 10:40:40PM +0300, kuznet@ms2.inr.ac.ru wrote:

    It simply does not exist for 82559* in all the steppings.
    eepro100 is pretty poor device.
    
    Probably, it exists for card identified as Gamla (D102) (82559 is
    D101).

I have code from Audrey to enable HWCK for received only and code
from you for SG -- but no way to enable HWCK for TX. The Intel
provided drivers themselves don't seem to have any way of doing this
so i wonder if it is either undocumented so Intel can sell the same
hardware at a premium (as they do for NT, the only different being
drivers) or simply because the HW is poor.



  --cw
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: [UPDATE] Zerocopy, last one today I promise :-)

[Anton Blanchard]

 
> o If sock_writepage is called on path via device without SG support,
>   the cooked up sock_sendmsg() call needs to switch to KERNEL_DS.
>   Discovered and fixed by Ingo Molnar.

Good catch.

> This does show that not too many people are testing this all that
> thoroughly :-) Basically, any sys_sendfile() over TCP using a network
> card other than loopback/3c59x/sunhme/acenic would fail with -EFAULT
> or even worse a kernel crash depending upon architecture.

Hey now, I was seeing it but I hadn't got around to chasing the bug down. :)

Anton
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/