blockfile access patterns logging

blockfile access patterns logging

[bert hubert]

Jens,

For some time I've wanted to log exactly what linux is reading and writing
from my harddisk - for a variety of reasons. The current reason is that my
very idle laptop writes to disk every once in a while (or reads, I don't
know).

Now, conceptually this should not be very hard, but I'd like to ask your
thoughts on where I might insert some crude logging? There are lots of
places that might be better or worse for some reason.

I'd love to be as close to the physical block device as possible, short of
rewriting actual IDE drivers.

Any tips? Or is this idea crazy?

Thanks!

-- 
http://www.PowerDNS.com      Open source, database driven DNS Software 
http://lartc.org           Linux Advanced Routing & Traffic Control HOWTO

Re: blockfile access patterns logging

[Jens Axboe]

On Thu, Jan 08 2004, bert hubert wrote:
> Jens,
> 
> For some time I've wanted to log exactly what linux is reading and writing
> from my harddisk - for a variety of reasons. The current reason is that my
> very idle laptop writes to disk every once in a while (or reads, I don't
> know).
> 
> Now, conceptually this should not be very hard, but I'd like to ask your
> thoughts on where I might insert some crude logging? There are lots of
> places that might be better or worse for some reason.
> 
> I'd love to be as close to the physical block device as possible, short of
> rewriting actual IDE drivers.
> 
> Any tips? Or is this idea crazy?

If you have the laptop mode patch (it's in 2.4 current, and in 2.6-mm as
well), then you can enable block dump by echoing 1 to
/proc/sys/vm/block_dump - this will dump all reads/writes to any device
in the system. Sounds like this is what you want.

-- 
Jens Axboe

Re: blockfile access patterns logging

[Valdis.Kletnieks]

[-- Attachment #1: Type: text/plain, Size: 1115 bytes --]

On Thu, 08 Jan 2004 13:00:08 +0100, bert hubert said:

> For some time I've wanted to log exactly what linux is reading and writing
> from my harddisk - for a variety of reasons. The current reason is that my
> very idle laptop writes to disk every once in a while (or reads, I don't
> know).
> 
> Now, conceptually this should not be very hard, but I'd like to ask your
> thoughts on where I might insert some crude logging? There are lots of
> places that might be better or worse for some reason.
> 
> I'd love to be as close to the physical block device as possible, short of
> rewriting actual IDE drivers.

You probably want to do logging at a higher level.  It's totally useless to find out
that LBA 2234324567 got re-written.  Mapping it to a partition on the disk so
you know it was something on /dev/hda7 is a bit better.  And being able to tell
that somebody updated the atime on /var/log/messages is most informative of all.

The other problem is that unless your laptop is *VERY* idle, you will have a scrolling
problem and buffering issues - so you end up writing to disk to log the buffers and... ;)

[-- Attachment #2: Type: application/pgp-signature, Size: 226 bytes --]

Re: blockfile access patterns logging

[Jens Axboe]

On Thu, Jan 08 2004, Valdis.Kletnieks@vt.edu wrote:
> On Thu, 08 Jan 2004 13:00:08 +0100, bert hubert said:
> 
> > For some time I've wanted to log exactly what linux is reading and writing
> > from my harddisk - for a variety of reasons. The current reason is that my
> > very idle laptop writes to disk every once in a while (or reads, I don't
> > know).
> > 
> > Now, conceptually this should not be very hard, but I'd like to ask your
> > thoughts on where I might insert some crude logging? There are lots of
> > places that might be better or worse for some reason.
> > 
> > I'd love to be as close to the physical block device as possible, short of
> > rewriting actual IDE drivers.
> 
> You probably want to do logging at a higher level.  It's totally
> useless to find out that LBA 2234324567 got re-written.  Mapping it to
> a partition on the disk so you know it was something on /dev/hda7 is a
> bit better.  And being able to tell that somebody updated the atime on
> /var/log/messages is most informative of all.

For laptops, it's often most interesting to find out _what_ process
dirtied what data (which in turn caused bdflush to sync it), or what
process keeps doing small reads. And block_dump does exactly that (it
was invented for exactly that purpose :)


> The other problem is that unless your laptop is *VERY* idle, you will
> have a scrolling problem and buffering issues - so you end up writing
> to disk to log the buffers and... ;)

I don't think you understand what Bert is looking for. He explicitly
mentions that the machine is very idle, so he's probably looking for
culprits that spin up the drive occasionally.

It doesn't provide tcpdump like logs of course, that's far more
invasive.

-- 
Jens Axboe

Re: blockfile access patterns logging

[Valdis.Kletnieks]

[-- Attachment #1: Type: text/plain, Size: 728 bytes --]

On Thu, 08 Jan 2004 15:39:08 +0100, Jens Axboe said:

> For laptops, it's often most interesting to find out _what_ process
> dirtied what data (which in turn caused bdflush to sync it), or what
> process keeps doing small reads. And block_dump does exactly that (it
> was invented for exactly that purpose :)

I submit that "what process ID did it" is even more remote from the disk
than "what inode" ;)

> I don't think you understand what Bert is looking for. He explicitly
> mentions that the machine is very idle, so he's probably looking for
> culprits that spin up the drive occasionally.

Oh, I understand very well.. Been there myself.  I'd just forgotten that
the laptop patch provided the process-ID information. ;)


[-- Attachment #2: Type: application/pgp-signature, Size: 226 bytes --]

Re: blockfile access patterns logging

[Jens Axboe]

On Thu, Jan 08 2004, Valdis.Kletnieks@vt.edu wrote:
> On Thu, 08 Jan 2004 15:39:08 +0100, Jens Axboe said:
> 
> > For laptops, it's often most interesting to find out _what_ process
> > dirtied what data (which in turn caused bdflush to sync it), or what
> > process keeps doing small reads. And block_dump does exactly that (it
> > was invented for exactly that purpose :)
> 
> I submit that "what process ID did it" is even more remote from the disk
> than "what inode" ;)

There are no inodes involved at that point, it's left the fs and is
purely in area between block layer and disk driver. Besides, what would
do with the inode for this problem? If you really want to find eg the
file in question (if you cannot _guess_ when you know the process), then
you could go back from the block number. I've never ever had a need to
go that far, though.

-- 
Jens Axboe

blockfile access patterns logging

[Samium Gromoff]

> Jens,                                                                                    
>                                                                                          
> For some time I've wanted to log exactly what linux is reading and writing               
> from my harddisk - for a variety of reasons. The current reason is that my               
> very idle laptop writes to disk every once in a while (or reads, I don't                 
> know).                                                                                   

Have you mounted the fs with -o noatime?

> Now, conceptually this should not be very hard, but I'd like to ask your                 
> thoughts on where I might insert some crude logging? There are lots of                   
> places that might be better or worse for some reason.                                    
>                                                                                          
> I'd love to be as close to the physical block device as possible, short of               
> rewriting actual IDE drivers.                                                            
>                                                                                          
> Any tips? Or is this idea crazy?                                                         
>                                                                                          
> Thanks!                                                                                  
>                                                                                          
> --                                                                                       
> http://www.PowerDNS.com      Open source, database driven DNS Software                   
> http://lartc.org           Linux Advanced Routing & Traffic Control HOWTO                

regards, Samium Gromoff