[PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Andi Kleen]

For some reason it triggers always with NFS root and spams the kernel
logs of my nfs root boxes a lot.

Cc: trond.myklebust@fys.uio.no
Signed-off-by: Andi Kleen <ak@suse.de>

---
 mm/filemap.c |    2 --
 1 files changed, 2 deletions(-)

Index: linux-2.6.18-rc2-git7/mm/filemap.c
===================================================================
--- linux-2.6.18-rc2-git7.orig/mm/filemap.c
+++ linux-2.6.18-rc2-git7/mm/filemap.c
@@ -849,8 +849,6 @@ static void shrink_readahead_size_eio(st
 		return;
 
 	ra->ra_pages /= 4;
-	printk(KERN_WARNING "Reducing readahead size to %luK\n",
-			ra->ra_pages << (PAGE_CACHE_SHIFT - 10));
 }
 
 /**

Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Trond Myklebust]

On Sat, 2006-07-29 at 21:42 +0200, Andi Kleen wrote:
> For some reason it triggers always with NFS root and spams the kernel
> logs of my nfs root boxes a lot.
> 
> Cc: trond.myklebust@fys.uio.no
> Signed-off-by: Andi Kleen <ak@suse.de>

Big ACK! I never really understood why we needed this printk, and yes,
it does spam the syslog heavily on all NFS clients...

Cheers,
  Trond

> ---
>  mm/filemap.c |    2 --
>  1 files changed, 2 deletions(-)
> 
> Index: linux-2.6.18-rc2-git7/mm/filemap.c
> ===================================================================
> --- linux-2.6.18-rc2-git7.orig/mm/filemap.c
> +++ linux-2.6.18-rc2-git7/mm/filemap.c
> @@ -849,8 +849,6 @@ static void shrink_readahead_size_eio(st
>  		return;
>  
>  	ra->ra_pages /= 4;
> -	printk(KERN_WARNING "Reducing readahead size to %luK\n",
> -			ra->ra_pages << (PAGE_CACHE_SHIFT - 10));
>  }
>  
>  /**

Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Nate Diller]

On 7/29/06, Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> On Sat, 2006-07-29 at 21:42 +0200, Andi Kleen wrote:
> > For some reason it triggers always with NFS root and spams the kernel
> > logs of my nfs root boxes a lot.
> >
> > Cc: trond.myklebust@fys.uio.no
> > Signed-off-by: Andi Kleen <ak@suse.de>
>
> Big ACK! I never really understood why we needed this printk, and yes,
> it does spam the syslog heavily on all NFS clients...

not an NFS user, but it seems like this is indicating an actual
performance problem.  the printk was introduced by an attempt to
reduce retries after -EIO on scratched CDs/DVDs, and is intended to
curb pathological behavior when one disk sector consitently cannot be
read.

for some reason (maybe just busy networks, dunno), you're getting I/O
errors on NFS, and the readahead window is being reduced.  IMO, the
proper behavior in this case would be to keep it large, or even
increase it, since that would enable the network to more efficiently
transfer data.

of course, it could be that some quirk of the NFS client VFS interface
causes "spurious" -EIO returns.  either way, i'd rather see it fixed
rather than the printk removed, since it is useful to point out that
some performance degradation is occuring.

NATE

Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Trond Myklebust]

On Sat, 2006-07-29 at 18:04 -0700, Nate Diller wrote:
> On 7/29/06, Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> > On Sat, 2006-07-29 at 21:42 +0200, Andi Kleen wrote:
> > > For some reason it triggers always with NFS root and spams the kernel
> > > logs of my nfs root boxes a lot.
> > >
> > > Cc: trond.myklebust@fys.uio.no
> > > Signed-off-by: Andi Kleen <ak@suse.de>
> >
> > Big ACK! I never really understood why we needed this printk, and yes,
> > it does spam the syslog heavily on all NFS clients...
> 
> not an NFS user, but it seems like this is indicating an actual
> performance problem.  the printk was introduced by an attempt to
> reduce retries after -EIO on scratched CDs/DVDs, and is intended to
> curb pathological behavior when one disk sector consitently cannot be
> read.
>
> for some reason (maybe just busy networks, dunno), you're getting I/O
> errors on NFS, and the readahead window is being reduced.  IMO, the
> proper behavior in this case would be to keep it large, or even
> increase it, since that would enable the network to more efficiently
> transfer data.
> 
> of course, it could be that some quirk of the NFS client VFS interface
> causes "spurious" -EIO returns.  either way, i'd rather see it fixed
> rather than the printk removed, since it is useful to point out that
> some performance degradation is occuring.

We have no way of telling. That printk doesn't give us any useful
information whatsoever for debugging that sort of problem. It should
either be replaced with something that does, or it should be thrown out.

  Trond

Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Andrew Morton]

On Sun, 30 Jul 2006 00:22:14 -0400
Trond Myklebust <trond.myklebust@fys.uio.no> wrote:

> > of course, it could be that some quirk of the NFS client VFS interface
> > causes "spurious" -EIO returns.  either way, i'd rather see it fixed
> > rather than the printk removed, since it is useful to point out that
> > some performance degradation is occuring.
> 
> We have no way of telling. That printk doesn't give us any useful
> information whatsoever for debugging that sort of problem. It should
> either be replaced with something that does, or it should be thrown out.

err, the printk has found a probable bug in NFS.  That was pretty useful
of it.

Do we know why nfs's readpage isn't bringing the page up to date?

Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Trond Myklebust]

On Sat, 2006-07-29 at 23:26 -0700, Andrew Morton wrote:
> On Sun, 30 Jul 2006 00:22:14 -0400
> Trond Myklebust <trond.myklebust@fys.uio.no> wrote:
> 
> > > of course, it could be that some quirk of the NFS client VFS interface
> > > causes "spurious" -EIO returns.  either way, i'd rather see it fixed
> > > rather than the printk removed, since it is useful to point out that
> > > some performance degradation is occuring.
> > 
> > We have no way of telling. That printk doesn't give us any useful
> > information whatsoever for debugging that sort of problem. It should
> > either be replaced with something that does, or it should be thrown out.
> 
> err, the printk has found a probable bug in NFS.  That was pretty useful
> of it.

Not necessarily. AFAICS, the spam could be triggered by perfectly
legitimate activity. For instance, someone on the server may have
revoked your read permissions to the file, or may have deleted it.

> Do we know why nfs's readpage isn't bringing the page up to date?

It may be that other lurking issues were also triggering the printk. For
instance I know of a couple of corner cases in the krb5 privacy code
that could result in readpage failing. Those issues are being looked
into.

Cheers,
  Trond

Re: [discuss] Re: [PATCH for 2.6.18] [8/8] MM: Remove rogue readahead printk

[Andi Kleen]

> Not necessarily. AFAICS, the spam could be triggered by perfectly
> legitimate activity. For instance, someone on the server may have
> revoked your read permissions to the file, or may have deleted it.

I don't think any of this was the case in the cases I saw it.

It would be probably good to investigate post 2.6.18, but for 2.6.18
itself I think removing the printk was the best short term "fix".

> > Do we know why nfs's readpage isn't bringing the page up to date?
> 
> It may be that other lurking issues were also triggering the printk. For
> instance I know of a couple of corner cases in the krb5 privacy code
> that could result in readpage failing. Those issues are being looked
> into.

I don't use kerberos either.

-Andi