Re: [PATCH] sg: disable interrupts inside sg_copy_buffer

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Fri, 2008-09-12 at 23:12 +0900, FUJITA Tomonori wrote:
> On Fri, 12 Sep 2008 08:48:39 -0500
> James Bottomley <James.Bottomley@HansenPartnership.com> wrote:
> 
> > On Fri, 2008-09-12 at 15:26 +0900, FUJITA Tomonori wrote:
> > > On Fri, 12 Sep 2008 08:05:46 +0200
> > > Jens Axboe <jens.axboe@oracle.com> wrote:
> > > 
> > > > On Fri, Sep 12 2008, FUJITA Tomonori wrote:
> > > > > On Thu, 11 Sep 2008 18:33:03 +0200
> > > > > Jens Axboe <jens.axboe@oracle.com> wrote:
> > > > > 
> > > > > > On Thu, Sep 11 2008, FUJITA Tomonori wrote:
> > > > > > > The callers of sg_copy_buffer must disable interrupts before calling
> > > > > > > it (since it uses kmap_atomic). Some callers use it on
> > > > > > > interrupt-disabled code but some need to take the trouble to disable
> > > > > > > interrupts just for this. No wonder they forget about it and we hit a
> > > > > > > bug like:
> > > > > > > 
> > > > > > > http://bugzilla.kernel.org/show_bug.cgi?id=11529
> > > > > > > 
> > > > > > > James said that it might be better to disable interrupts inside the
> > > > > > > function rather than risk the callers getting it wrong.
> > > > > > > 
> > > > > > > Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> > > > > > > ---
> > > > > > >  lib/scatterlist.c |    5 +++++
> > > > > > >  1 files changed, 5 insertions(+), 0 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/lib/scatterlist.c b/lib/scatterlist.c
> > > > > > > index 876ba6d..dd52cd5 100644
> > > > > > > --- a/lib/scatterlist.c
> > > > > > > +++ b/lib/scatterlist.c
> > > > > > > @@ -422,6 +422,9 @@ static size_t sg_copy_buffer(struct scatterlist *sgl, unsigned int nents,
> > > > > > >  {
> > > > > > >  	unsigned int offset = 0;
> > > > > > >  	struct sg_mapping_iter miter;
> > > > > > > +	unsigned long flags;
> > > > > > > +
> > > > > > > +	local_irq_save(flags);
> > > > > > >  
> > > > > > >  	sg_miter_start(&miter, sgl, nents, SG_MITER_ATOMIC);
> > > > > > >  
> > > > > > > @@ -442,6 +445,8 @@ static size_t sg_copy_buffer(struct scatterlist *sgl, unsigned int nents,
> > > > > > >  
> > > > > > >  	sg_miter_stop(&miter);
> > > > > > >  
> > > > > > > +	local_irq_restore(flags);
> > > > > > > +
> > > > > > >  	return offset;
> > > > > > >  }
> > > > > > 
> > > > > > Agreed and applied.
> > > > > 
> > > > > Jens, this won't go to 2.6.27 (goes to 2.6.28)?
> > > > > 
> > > > > If so, I need to ask James to apply this workaround to 2.6.27:
> > > > > 
> > > > > http://marc.info/?l=linux-scsi&m=122106973807028&w=2
> > > > 
> > > > It will go into 2.6.27, it's already upstream since ~10 hours or so.
> > > 
> > > Thanks, I've just found it.
> > 
> > You forgot to add the cc: stable@kernel.org tag.  You'll have to send
> > this to stable manually quoting the upsream commit id.
> 
> You forgot that Tejun rewrote the sg copy code after 2.6.26. This
> patch can't be applied to 2.6.26.X.
> 
> I need to push a patch that is not in mainline to stable trees. Can I
> do that ?

Yes, if we're absolutely sure it's necessary.  You'll have to quote the
upstream commit ID, explain why the backport has to be different and add
the patch.

James