xfs_buf and buffercache/pagecache connection

xfs_buf and buffercache/pagecache connection

[Yannis Klonatos]

Hello,

         I was looking to add a kernel hook to my system in order to 
monitor buffer-cache hit and misses. Initially I was
planning to add my modifications to the __getblk(). However, i noticed 
that XFS does not directly use the buffer-cache
for its pages but it seems to implement its own buffer.
         What I am now looking for is 1) the place where XFS checks 
whether a page exists in its buffer or not and 2)
what are the possible interactions between xfs_buf and the Linux kernel 
buffer-cache.
         I would appreciate any information regarding the above issues.

Thanks in advance,
Yannis Klonatos





_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: xfs_buf and buffercache/pagecache connection

[Andi Kleen]

Yannis Klonatos <klonatos@ics.forth.gr> writes:

>         I was looking to add a kernel hook to my system in order to
> monitor buffer-cache hit and misses. Initially I was
> planning to add my modifications to the __getblk(). However, i noticed
> that XFS does not directly use the buffer-cache
> for its pages but it seems to implement its own buffer.
>         What I am now looking for is 1) the place where XFS checks
> whether a page exists in its buffer or not and 2)
> what are the possible interactions between xfs_buf and the Linux
> kernel buffer-cache.
>         I would appreciate any information regarding the above issues.

The kernel does not track all accesses, e.g. through mmap.
So you can only get misses (which is essentially IO rate and already
accounted), but not hits.

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: xfs_buf and buffercache/pagecache connection

[Yannis Klonatos]

[-- Attachment #1.1: Type: text/plain, Size: 2259 bytes --]

???? 5/31/2010 9:04 PM, O/H Andi Kleen ??????:
> Yannis Klonatos<klonatos@ics.forth.gr>  writes:
>
>    
>>          I was looking to add a kernel hook to my system in order to
>> monitor buffer-cache hit and misses. Initially I was
>> planning to add my modifications to the __getblk(). However, i noticed
>> that XFS does not directly use the buffer-cache
>> for its pages but it seems to implement its own buffer.
>>          What I am now looking for is 1) the place where XFS checks
>> whether a page exists in its buffer or not and 2)
>> what are the possible interactions between xfs_buf and the Linux
>> kernel buffer-cache.
>>          I would appreciate any information regarding the above issues.
>>      
> The kernel does not track all accesses, e.g. through mmap.
> So you can only get misses (which is essentially IO rate and already
> accounted), but not hits.
>
> -Andi
>    

First of all thanks for your quick reply.

So, if i understand correctly, what you are saying is that it is 
basically impossible to modify xfs code
so that i get that specific information? This sounds a bit strange since 
if XFS was indeed using the
buffercache as ext3 or other fs does, the following modification would 
suffice (file fs/buffer.c):

structbuffer_head  <+code=buffer_head>  *
find_get_block  <+code=__find_get_block>(structblock_device  <+code=block_device>  *bdev  <+code=bdev>,sector_t  <+code=sector_t>  block  <+code=block>, intsize  <+code=size>) {

      	structbuffer_head  <+code=buffer_head>  *bh  <+code=bh>  =lookup_bh_lru  <+code=lookup_bh_lru>(bdev  <+code=bdev>,block  <+code=block>,size  <+code=size>);
	/* BEGIN MODIFICATION */
	if (bh) buffercache_hits++;
	else buffercache_misses++;
	/* END MODIFICATION */
         if (bh  <+code=bh>  ==NULL  <+code=NULL>) {
                bh  <+code=bh>  =__find_get_block_slow  <+code=__find_get_block_slow>(bdev  <+code=bdev>,block  <+code=block>);
                if (bh  <+code=bh>)
                        bh_lru_install  <+code=bh_lru_install>(bh  <+code=bh>);
        }
        if (bh  <+code=bh>) {
                touch_buffer  <+code=touch_buffer>(bh  <+code=bh>);
        returnbh  <+code=bh>;

}

XPORT_SYMBOL  <+code=EXPORT_SYMBOL>(__find_get_block  <+code=__find_get_block>);



[-- Attachment #1.2: Type: text/html, Size: 4086 bytes --]

[-- Attachment #2: Type: text/plain, Size: 121 bytes --]

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: xfs_buf and buffercache/pagecache connection

[Dave Chinner]

On Mon, May 31, 2010 at 07:31:39PM +0300, Yannis Klonatos wrote:
> Hello,
> 
>         I was looking to add a kernel hook to my system in order to
> monitor buffer-cache hit and misses. Initially I was
> planning to add my modifications to the __getblk(). However, i
> noticed that XFS does not directly use the buffer-cache
> for its pages but it seems to implement its own buffer.
>         What I am now looking for is 1) the place where XFS checks
> whether a page exists in its buffer or not and 2)
> what are the possible interactions between xfs_buf and the Linux
> kernel buffer-cache.
>         I would appreciate any information regarding the above issues.

There are already stats in place to tell you about this.

http://xfs.org/index.php/Runtime_Stats#buf_-_Buf_Statistics

They aren't documented there, but if you look at
fs/xfs/linux-2.6/xfs_buf.c an search for XFS_STATS_INC you'll find
what they all mean from the code. It isn't as simple as hit or miss
counters - there's different counters for different types of hits
and misses....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs