Re: efficient copy_to_user and copy_from_user routines in Linux Kernel

[Mala Anand]

Andrew Morton wrote..

>>This came up about a year back when zerocopy networking was merged.
>>Intel boxes started running more slowly purely because of the 8+8
>>alignment thing.

>>I changed tcp to use a different copy if either source or dest were
>>not eight-byte aligned, and found that the resulting improvement
>>across a mixed networking load was only 1%.  Your numbers are higher,
>>so perhaps there are different alignments in the mix...

>I will test on other workloads when I return back to work after OLS
>and vacation.  However we tested an earlier version of this patch on
>Netbench using sendfile and gained around 3% improvement. The baseline
>profiling showed that Netbench was spending 10% in generic_copy_to_user.
>The tcp options are aligned on an 4-byte boundary, so depending on the
>options used the address to the data (source address to the
>generic_copy_to_user) should fall on an 4 or 8 byte boundary. I agree
>with you more test is needed.

One correction to the above statement...
Due to the tcp options alignment on an 4-byte boundary the source
address to the generic_copy_to_user should fall on an 4, 8, 12
and 16 etc., byte boundary. However,I have seen that 4 and 12 byte
alignment using unrolled loop performed better than the string copy.

>>One question:  have you tested on other CPU types?  This problem is
>>very specific to Intel hardware.  On AMD, the eight-byte alignement
>>artifact does not exist at all.  It could be that your patch is not
>>desirable on such CPUs?

>I tested only on Pentium II and III. I will test it on Pentium IV.
>When I said 8-byte alignment, it is 8 and greater.  I will
>try to check out AMD also.

Same correction here, 8-byte alignment means 8, 16 or greater.

Regards,
    Mala

   Mala Anand
   E-mail:[e-mail]
   Linux Technology Center - Performance
   Phone:[phone]; Tie-line:678-8088