2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

Hi,

With the 2.6.16-rc5-mm3 kernel w/ the patch

revert-x86_64-mm-i386-early-alignment.patch

applied I'm able to hang my box (Asus L5D, 1 CPU, x86-64 kernel) solid
by running OpenOffice.org from under KDE (100% of the time but on one
user account only).  Before it hangs I get something like this on the serial console:

BUG: spinlock bad magic on CPU#0, soffice.bin/5293
 lock: ffff81005e174e28, .magic: 000001ff, .owner: .5).@4).06)./0, .owner_cpu: -2141827648
BUG: spinlock lockup on CPU#0, soffice.bin/5293, ffff81005e174e28

Greetings,
Rafael

Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andrew Morton]

"Rafael J. Wysocki" <rjw@sisk.pl> wrote:
>
> Hi,
> 
> With the 2.6.16-rc5-mm3 kernel w/ the patch
> 
> revert-x86_64-mm-i386-early-alignment.patch
> 
> applied I'm able to hang my box (Asus L5D, 1 CPU, x86-64 kernel) solid
> by running OpenOffice.org from under KDE (100% of the time but on one
> user account only).  Before it hangs I get something like this on the serial console:
> 
> BUG: spinlock bad magic on CPU#0, soffice.bin/5293
>  lock: ffff81005e174e28, .magic: 000001ff, .owner: .5).@4).06)./0, .owner_cpu: -2141827648
> BUG: spinlock lockup on CPU#0, soffice.bin/5293, ffff81005e174e28
> 

Is it a !CONFIG_SMP kernel?

There's no stack trace?

Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

On Sunday 12 March 2006 00:36, Andrew Morton wrote:
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > With the 2.6.16-rc5-mm3 kernel w/ the patch
> > 
> > revert-x86_64-mm-i386-early-alignment.patch
> > 
> > applied I'm able to hang my box (Asus L5D, 1 CPU, x86-64 kernel) solid
> > by running OpenOffice.org from under KDE (100% of the time but on one
> > user account only).  Before it hangs I get something like this on the serial console:
> > 
> > BUG: spinlock bad magic on CPU#0, soffice.bin/5293
> >  lock: ffff81005e174e28, .magic: 000001ff, .owner: .5).@4).06)./0, .owner_cpu: -2141827648
> > BUG: spinlock lockup on CPU#0, soffice.bin/5293, ffff81005e174e28
> > 
> 
> Is it a !CONFIG_SMP kernel?

Yes.

> There's no stack trace?

Well, probably there was one but it didn't appear on the serial console because
the console loglevel was too low.  I'll try to increase console_loglevel in
spin_bug() and see what happens.

[BTW, looking at the relevant code, in my case .owner_cpu appears to have only
the most significant bit set (??).]

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

[Dropped discuss@x86-64.org from the Cc list, it's probably OT there]

On Sunday 12 March 2006 11:27, Rafael J. Wysocki wrote:
> On Sunday 12 March 2006 00:36, Andrew Morton wrote:
> > "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > > With the 2.6.16-rc5-mm3 kernel w/ the patch
> > > 
> > > revert-x86_64-mm-i386-early-alignment.patch
> > > 
> > > applied I'm able to hang my box (Asus L5D, 1 CPU, x86-64 kernel) solid
> > > by running OpenOffice.org from under KDE (100% of the time but on one
> > > user account only).  Before it hangs I get something like this on the serial console:
> > > 
> > > BUG: spinlock bad magic on CPU#0, soffice.bin/5293
> > >  lock: ffff81005e174e28, .magic: 000001ff, .owner: .5).@4).06)./0, .owner_cpu: -2141827648
> > > BUG: spinlock lockup on CPU#0, soffice.bin/5293, ffff81005e174e28
> > > 
> > 
> > Is it a !CONFIG_SMP kernel?
> 
> Yes.
> 
> > There's no stack trace?
> 
> Well, probably there was one but it didn't appear on the serial console because
> the console loglevel was too low.  I'll try to increase console_loglevel in
> spin_bug() and see what happens.

Done, and now it looks like this:

BUG: spinlock bad magic on CPU#0, soffice.bin/5192
 lock: ffff81005f79ae28, .magic: 000001ff, .owner: 1..1..|1. |1.|1..|1.�1..|1..1. 1./-1,
.owner_cpu: -2141838208

Call Trace: <ffffffff80210383>{__alloc_pages+99} <ffffffff802156c3>{spin_bug+195}
       <ffffffff802077dc>{_raw_spin_lock+44} <ffffffff80270a4e>{_spin_lock+30}
       <ffffffff8033712d>{journal_extend+77} <ffffffff80327255>{ext3_get_block+165}
       <ffffffff8022c2f9>{do_mpage_readpage+425} <ffffffff80270cc4>{_write_unlock_irq+20}
       <ffffffff8020cce2>{add_to_page_cache+162} <ffffffff8023fdee>{mpage_readpages+254}
       <ffffffff803271b0>{ext3_get_block+0} <ffffffff803271b0>{ext3_get_block+0}
       <ffffffff803146df>{get_cnode+95} <ffffffff8020a3bb>{get_page_from_freelist+619}
       <ffffffff80210383>{__alloc_pages+99} <ffffffff80323c1a>{ext3_readpages+26}
       <ffffffff80214030>{__do_page_cache_readahead+416} <ffffffff80213c12>{poison_obj+66}
       <ffffffff80232058>{wake_up_bit+40} <ffffffff80243152>{unlock_buffer+18}
       <ffffffff80315bb8>{reiserfs_prepare_for_journal+104}
       <ffffffff802b6ab4>{do_page_cache_readahead+100} <ffffffff80215942>{filemap_nopage+322}
       <ffffffff80208b2c>{__handle_mm_fault+1004} <ffffffff80270e7d>{_spin_unlock_irqrestore+29}
       <ffffffff8020ae99>{do_page_fault+1257} <ffffffff8026af8d>{error_exit+0}
       <ffffffff802ffb40>{reiserfs_copy_from_user_to_file_region+80}
       <ffffffff80302446>{reiserfs_file_write+6102} <ffffffff802f8f4e>{reiserfs_add_entry+1054}
       <ffffffff8033c1ff>{journal_cancel_revoke+351} <ffffffff80213c12>{poison_obj+66}
       <ffffffff80236d27>{cache_free_debugcheck+711} <ffffffff80335734>{journal_stop+772}
       <ffffffff80270f30>{_spin_unlock+16} <ffffffff802193a2>{vfs_write+226}
       <ffffffff80219c80>{sys_write+80} <ffffffff8026d234>{cstar_do_call+27}
BUG: spinlock lockup on CPU#0, soffice.bin/5192, ffff81005f79ae28

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andi Kleen]

On Sunday 12 March 2006 23:26, Andrew Morton wrote:

> 
> It's a pretty vile backtrace.  I supposed you have CONFIG_FRAME_POINTER=n.

Won't make a difference because the oops backtracer doesn't
use them.

> Still.  It seems that what's happened is that we took a pagefault while
> reiserfs had a transaction open.  The fault is against a mmapped ext3 file
> and we ended up in the recently-reworked ext3_get_block() which tests
> journal_current_handle() to work out whether we're in a write or a read. 
> oops.  The presence of reiserfs journal_info makes it decide it's a write,
> not a read so it starts treating a reiserfs journal_info as an ext3 one.
> 
> The code used to work OK because it was only for direct-IO, which doesn't
> get recurred into like this.  But it got used for regular I/O in -mm.

Oops. Can this happen in more situations?

-Andi

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andrew Morton]

"Rafael J. Wysocki" <rjw@sisk.pl> wrote:
>
>  Done, and now it looks like this:

Helps a lot, thanks.

>  BUG: spinlock bad magic on CPU#0, soffice.bin/5192
>   lock: ffff81005f79ae28, .magic: 000001ff, .owner: 1..1..|1. |1.|1..|1.___1..|1..1. 1./-1,
>  .owner_cpu: -2141838208
> 
>  Call Trace: <ffffffff80210383>{__alloc_pages+99} <ffffffff802156c3>{spin_bug+195}
>         <ffffffff802077dc>{_raw_spin_lock+44} <ffffffff80270a4e>{_spin_lock+30}
>         <ffffffff8033712d>{journal_extend+77} <ffffffff80327255>{ext3_get_block+165}
>         <ffffffff8022c2f9>{do_mpage_readpage+425} <ffffffff80270cc4>{_write_unlock_irq+20}
>         <ffffffff8020cce2>{add_to_page_cache+162} <ffffffff8023fdee>{mpage_readpages+254}
>         <ffffffff803271b0>{ext3_get_block+0} <ffffffff803271b0>{ext3_get_block+0}
>         <ffffffff803146df>{get_cnode+95} <ffffffff8020a3bb>{get_page_from_freelist+619}
>         <ffffffff80210383>{__alloc_pages+99} <ffffffff80323c1a>{ext3_readpages+26}
>         <ffffffff80214030>{__do_page_cache_readahead+416} <ffffffff80213c12>{poison_obj+66}
>         <ffffffff80232058>{wake_up_bit+40} <ffffffff80243152>{unlock_buffer+18}
>         <ffffffff80315bb8>{reiserfs_prepare_for_journal+104}
>         <ffffffff802b6ab4>{do_page_cache_readahead+100} <ffffffff80215942>{filemap_nopage+322}
>         <ffffffff80208b2c>{__handle_mm_fault+1004} <ffffffff80270e7d>{_spin_unlock_irqrestore+29}
>         <ffffffff8020ae99>{do_page_fault+1257} <ffffffff8026af8d>{error_exit+0}
>         <ffffffff802ffb40>{reiserfs_copy_from_user_to_file_region+80}
>         <ffffffff80302446>{reiserfs_file_write+6102} <ffffffff802f8f4e>{reiserfs_add_entry+1054}
>         <ffffffff8033c1ff>{journal_cancel_revoke+351} <ffffffff80213c12>{poison_obj+66}
>         <ffffffff80236d27>{cache_free_debugcheck+711} <ffffffff80335734>{journal_stop+772}
>         <ffffffff80270f30>{_spin_unlock+16} <ffffffff802193a2>{vfs_write+226}
>         <ffffffff80219c80>{sys_write+80} <ffffffff8026d234>{cstar_do_call+27}
>  BUG: spinlock lockup on CPU#0, soffice.bin/5192, ffff81005f79ae28

It's a pretty vile backtrace.  I supposed you have CONFIG_FRAME_POINTER=n.

Still.  It seems that what's happened is that we took a pagefault while
reiserfs had a transaction open.  The fault is against a mmapped ext3 file
and we ended up in the recently-reworked ext3_get_block() which tests
journal_current_handle() to work out whether we're in a write or a read. 
oops.  The presence of reiserfs journal_info makes it decide it's a write,
not a read so it starts treating a reiserfs journal_info as an ext3 one.

The code used to work OK because it was only for direct-IO, which doesn't
get recurred into like this.  But it got used for regular I/O in -mm.

This should fix:

--- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
+++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
@@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
 	handle_t *handle = journal_current_handle();
 	int ret = 0;
 
-	if (!handle)
+	if (!create)
 		goto get_block;		/* A read */
 
 	if (max_blocks == 1)
_

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andrew Morton]

Andi Kleen <ak@suse.de> wrote:
>
> > Still.  It seems that what's happened is that we took a pagefault while
>  > reiserfs had a transaction open.  The fault is against a mmapped ext3 file
>  > and we ended up in the recently-reworked ext3_get_block() which tests
>  > journal_current_handle() to work out whether we're in a write or a read. 
>  > oops.  The presence of reiserfs journal_info makes it decide it's a write,
>  > not a read so it starts treating a reiserfs journal_info as an ext3 one.
>  > 
>  > The code used to work OK because it was only for direct-IO, which doesn't
>  > get recurred into like this.  But it got used for regular I/O in -mm.
> 
>  Oops. Can this happen in more situations?

I don't _think_ so, but it's pretty scary.  The code's been this way for a
while.

Typical scenario:

	reiserfs_journal_start		- sets current->journal_info
	  copy_from_user
	    pagefault
	      ext3_readpages()

As long as ext3_readpage[s]() doesn't try to start a transaction we're OK. 
And it shouldn't, if create==0.

Fortunately filemap_nopage() doesn't do atime updates.  If it did, things
would get messy.

We do have deadlock possibilities in there - filemap_nopage() does
lock_page() inside journal_start(), whereas generic_file_write() does
journal_start() inside lock_page().  Chris Mason and I have stared
unhappily at that a few times.  Hard to fix.

But I don't _think_ we have any more journal_start() recursions like this -
ext3 tends to get pretty noisy if it detects that in unexpected places.

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

On Sunday 12 March 2006 23:26, Andrew Morton wrote:
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> >
> >  Done, and now it looks like this:
> 
> Helps a lot, thanks.
> 
> >  BUG: spinlock bad magic on CPU#0, soffice.bin/5192
> >   lock: ffff81005f79ae28, .magic: 000001ff, .owner: 1..1..|1. |1.|1..|1.___1..|1..1. 1./-1,
> >  .owner_cpu: -2141838208
> > 
> >  Call Trace: <ffffffff80210383>{__alloc_pages+99} <ffffffff802156c3>{spin_bug+195}
> >         <ffffffff802077dc>{_raw_spin_lock+44} <ffffffff80270a4e>{_spin_lock+30}
> >         <ffffffff8033712d>{journal_extend+77} <ffffffff80327255>{ext3_get_block+165}
> >         <ffffffff8022c2f9>{do_mpage_readpage+425} <ffffffff80270cc4>{_write_unlock_irq+20}
> >         <ffffffff8020cce2>{add_to_page_cache+162} <ffffffff8023fdee>{mpage_readpages+254}
> >         <ffffffff803271b0>{ext3_get_block+0} <ffffffff803271b0>{ext3_get_block+0}
> >         <ffffffff803146df>{get_cnode+95} <ffffffff8020a3bb>{get_page_from_freelist+619}
> >         <ffffffff80210383>{__alloc_pages+99} <ffffffff80323c1a>{ext3_readpages+26}
> >         <ffffffff80214030>{__do_page_cache_readahead+416} <ffffffff80213c12>{poison_obj+66}
> >         <ffffffff80232058>{wake_up_bit+40} <ffffffff80243152>{unlock_buffer+18}
> >         <ffffffff80315bb8>{reiserfs_prepare_for_journal+104}
> >         <ffffffff802b6ab4>{do_page_cache_readahead+100} <ffffffff80215942>{filemap_nopage+322}
> >         <ffffffff80208b2c>{__handle_mm_fault+1004} <ffffffff80270e7d>{_spin_unlock_irqrestore+29}
> >         <ffffffff8020ae99>{do_page_fault+1257} <ffffffff8026af8d>{error_exit+0}
> >         <ffffffff802ffb40>{reiserfs_copy_from_user_to_file_region+80}
> >         <ffffffff80302446>{reiserfs_file_write+6102} <ffffffff802f8f4e>{reiserfs_add_entry+1054}
> >         <ffffffff8033c1ff>{journal_cancel_revoke+351} <ffffffff80213c12>{poison_obj+66}
> >         <ffffffff80236d27>{cache_free_debugcheck+711} <ffffffff80335734>{journal_stop+772}
> >         <ffffffff80270f30>{_spin_unlock+16} <ffffffff802193a2>{vfs_write+226}
> >         <ffffffff80219c80>{sys_write+80} <ffffffff8026d234>{cstar_do_call+27}
> >  BUG: spinlock lockup on CPU#0, soffice.bin/5192, ffff81005f79ae28
> 
> It's a pretty vile backtrace.  I supposed you have CONFIG_FRAME_POINTER=n.
> 
> Still.  It seems that what's happened is that we took a pagefault while
> reiserfs had a transaction open.  The fault is against a mmapped ext3 file
> and we ended up in the recently-reworked ext3_get_block() which tests
> journal_current_handle() to work out whether we're in a write or a read. 
> oops.  The presence of reiserfs journal_info makes it decide it's a write,
> not a read so it starts treating a reiserfs journal_info as an ext3 one.
> 
> The code used to work OK because it was only for direct-IO, which doesn't
> get recurred into like this.  But it got used for regular I/O in -mm.
> 
> This should fix:
> 
> --- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
> +++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
> @@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
>  	handle_t *handle = journal_current_handle();
>  	int ret = 0;
>  
> -	if (!handle)
> +	if (!create)
>  		goto get_block;		/* A read */
>  
>  	if (max_blocks == 1)

Er, it doesn't apply to either 2.6.16-rc5-mm3 or 2.6.16-rc6-mm1.

There's no function ext3_direct_io_get_blocks() there in fs/ext3/inode.c, AFAICT.

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Andrew Morton]

"Rafael J. Wysocki" <rjw@sisk.pl> wrote:
>
> > This should fix:
>  > 
>  > --- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
>  > +++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
>  > @@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
>  >  	handle_t *handle = journal_current_handle();
>  >  	int ret = 0;
>  >  
>  > -	if (!handle)
>  > +	if (!create)
>  >  		goto get_block;		/* A read */
>  >  
>  >  	if (max_blocks == 1)
> 
>  Er, it doesn't apply to either 2.6.16-rc5-mm3 or 2.6.16-rc6-mm1.

Nope, it applies OK to rc6-mm1.

But whatever - it's only a single line?

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

On Monday 13 March 2006 12:45, Andrew Morton wrote:
> "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> >
> > > This should fix:
> >  > 
> >  > --- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
> >  > +++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
> >  > @@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
> >  >  	handle_t *handle = journal_current_handle();
> >  >  	int ret = 0;
> >  >  
> >  > -	if (!handle)
> >  > +	if (!create)
> >  >  		goto get_block;		/* A read */
> >  >  
> >  >  	if (max_blocks == 1)
> > 
> >  Er, it doesn't apply to either 2.6.16-rc5-mm3 or 2.6.16-rc6-mm1.
> 
> Nope, it applies OK to rc6-mm1.

Well, this means my rc6-mm1 is different to what you have. :-)

Anyway in "my" version there's a function ext3_get_block() which reads like this:

static int ext3_get_block(struct inode *inode, sector_t iblock,
                        struct buffer_head *bh_result, int create)
{
        handle_t *handle = journal_current_handle();
        int ret = 0;
        unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;

        if (!handle)
                goto get_block;         /* A read */

        if (max_blocks == 1)
                goto get_block;         /* A single block get */

etc.

I guess I should replace the "if (!handle)" with "if (!create)"?

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Badari Pulavarty]

On Mon, 2006-03-13 at 13:07 +0100, Rafael J. Wysocki wrote:
> On Monday 13 March 2006 12:45, Andrew Morton wrote:
> > "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > >
> > > > This should fix:
> > >  > 
> > >  > --- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
> > >  > +++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
> > >  > @@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
> > >  >  	handle_t *handle = journal_current_handle();
> > >  >  	int ret = 0;
> > >  >  
> > >  > -	if (!handle)
> > >  > +	if (!create)
> > >  >  		goto get_block;		/* A read */
> > >  >  
> > >  >  	if (max_blocks == 1)
> > > 
> > >  Er, it doesn't apply to either 2.6.16-rc5-mm3 or 2.6.16-rc6-mm1.
> > 
> > Nope, it applies OK to rc6-mm1.
> 
> Well, this means my rc6-mm1 is different to what you have. :-)
> 
> Anyway in "my" version there's a function ext3_get_block() which reads like this:
> 
> static int ext3_get_block(struct inode *inode, sector_t iblock,
>                         struct buffer_head *bh_result, int create)
> {
>         handle_t *handle = journal_current_handle();
>         int ret = 0;
>         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
> 
>         if (!handle)
>                 goto get_block;         /* A read */
> 
>         if (max_blocks == 1)
>                 goto get_block;         /* A single block get */
> 
> etc.
> 
> I guess I should replace the "if (!handle)" with "if (!create)"?

Yes. In "-mm" ext3_get_block() == ext3_direct_io_getblocks() in
mainline. 

I renamed ext3_direct_io_getblocks() to ext3_get_block() (in -mm) 
since both of them do same thing now. (both can deal with mapping
multiple blocks).

Thanks,
Badari

Re: [discuss] Re: 2.6.16-rc5-mm3: spinlock bad magic on CPU#0 on AMD64

[Rafael J. Wysocki]

On Monday 13 March 2006 17:02, Badari Pulavarty wrote:
> On Mon, 2006-03-13 at 13:07 +0100, Rafael J. Wysocki wrote:
> > On Monday 13 March 2006 12:45, Andrew Morton wrote:
> > > "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > > >
> > > > > This should fix:
> > > >  > 
> > > >  > --- devel/fs/ext3/inode.c~ext3-get-blocks-maping-multiple-blocks-at-a-once-journal-reentry-fix	2006-03-12 14:25:04.000000000 -0800
> > > >  > +++ devel-akpm/fs/ext3/inode.c	2006-03-12 14:25:04.000000000 -0800
> > > >  > @@ -830,7 +830,7 @@ ext3_direct_io_get_blocks(struct inode *
> > > >  >  	handle_t *handle = journal_current_handle();
> > > >  >  	int ret = 0;
> > > >  >  
> > > >  > -	if (!handle)
> > > >  > +	if (!create)
> > > >  >  		goto get_block;		/* A read */
> > > >  >  
> > > >  >  	if (max_blocks == 1)
> > > > 
> > > >  Er, it doesn't apply to either 2.6.16-rc5-mm3 or 2.6.16-rc6-mm1.
> > > 
> > > Nope, it applies OK to rc6-mm1.
> > 
> > Well, this means my rc6-mm1 is different to what you have. :-)
> > 
> > Anyway in "my" version there's a function ext3_get_block() which reads like this:
> > 
> > static int ext3_get_block(struct inode *inode, sector_t iblock,
> >                         struct buffer_head *bh_result, int create)
> > {
> >         handle_t *handle = journal_current_handle();
> >         int ret = 0;
> >         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
> > 
> >         if (!handle)
> >                 goto get_block;         /* A read */
> > 
> >         if (max_blocks == 1)
> >                 goto get_block;         /* A single block get */
> > 
> > etc.
> > 
> > I guess I should replace the "if (!handle)" with "if (!create)"?
> 
> Yes. In "-mm" ext3_get_block() == ext3_direct_io_getblocks() in
> mainline. 
> 
> I renamed ext3_direct_io_getblocks() to ext3_get_block() (in -mm) 
> since both of them do same thing now. (both can deal with mapping
> multiple blocks).

Ah, I see, thanks.

Replacing the  "if (!handle)" with "if (!create)" fixed the problem for me.

Greetings,
Rafael