Re: [PATCH]: Proposed fix to the I2O problems with > 4GB space

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 25 Jul 2008 13:16:37 +0100 Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:

> From: Alan Cox <alan@redhat.com>
> 
> The I2O ioctls assume 32bits. In itself that is fine as they are old cards
> and nobody uses 64bit. However on LKML it was noted this assumption is also
> made for some ioctl allocated memory and is unsafe on 64bit systems.
> 
> Fixing this is a mess. It turns out there is tons of crap buried in a header
> file that does racy 32/64bit filtering on the masks.
> 
> So we:
> - Verify all callers of the racy code can sleep (i2o_dma_[re]alloc)
> - Move the code into a new i2o/memory.c file
> - Remove the gfp_mask argument so nobody can try and misuse the function
> - Wrap a mutex around the problem area (a single mutex is easy to do and
>   none of this is performance relevant)
> - Switch the remaining problem kmalloc holdout to use i2o_dma_alloc
> 
> There is a remaining race but that causes 32bit allocations to occur very
> rarely during block I/O when 64bit could be used. I don't think that is worth
> the cost of fixing as the impact is just a negligible performance loss.
> 
> Signed-off-by: Alan Cox <alan@redhat.com>
> 

I think you sent the wrong version of this patch - i2o_dma_alloc() is
all mucked up.  Sometimes takes three args, sometimes four.  When it's
four the fourth arg is sometimes a gfp_t, sometimes a PCI_DMA_foo.

It fell afoul of the dma_mapping_error() interface change too.

> (but could do with more testing, especially on 64bit > 4GB boxes with 64bit
> capable I2O cards)

A bit of git-trolling indicates that Vasily might be able to help out
with a bit of testing?