Re: Fw: crash on x86_64 - mm related?

[Hugh Dickins <hugh@veritas.com>]

On Thu, 1 Dec 2005, Kai Makisara wrote:
> On Thu, 1 Dec 2005, James Bottomley wrote:
> > 
> > On a side note, I have Kai's patch in the scsi-rc-fixes tree which I'm
> > getting ready to push.  Can we get a consensus on whether it should be
> > removed before I merge upwards?

I too need to decide whether to base my little sg.c,st.c patchset
on top of Kai's (as I see in 2.6.15-rc3-mm1, I presume) or on top
of 2.6.15-rc4.

> I think it should be removed because it is based partly on a wrong 
> assumption: asynchronous writes are _not_ done together with direct i/o. 
> (I have also experimentally verified that this does not happen.)

I'm assuming from this that I'd best base on 2.6.15-rc4;
but by all means overrule me if you've changed your mind.

> The patch includes the patch I sent sent to linux-scsi on Nov 21. Nobody 
> has commented it and I don't know if the user pages have to be explicitly 
> marked dirty after the HBA has read data there. If they have to, then this 
> earlier patch is valid.

What I see in 2.6.15-rc3-mm1 looks like three patches.

One to do with resetting sg_segs to 0 at various points:
I've no appreciation of that patch at all.  If it would help you for
me to add that into my little set, please send me a comment for it.

One to add an "is_read" argument to release_buffering.  Yes, that's
a part of my set too, though in my case called "dirtied" (and I believe
that the call at the end of st_read can say 0, because that's just for
an error path: when it's really dirtied user memory, it'll be read_tape
that does the release_buffering).  sg.c was always saying dirtied, even
when writing from memory; st.c was always saying not dirtied, even when
reading into memory.  Usually the latter is okay, get_user_pages has
said dirty in advance; but under pressure there's a window whereby it's
not good enough.  And SetPageDirty can be counter-productive these days,
so your patch is incomplete in that regard: I'll explain more in mine.

One to move around where release_buffering is called from:
that's the part you've decided was wrong, or at least unnecessary.

> If not, I will send a patch for 2.6.16 to remove the latent code.

I didn't understand that bit, but I probably don't need to.

Hugh
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/