Re: [PATCH] dma-debug: Add dma map/unmap error tracking support

[Shuah Khan <shuah.khan@hp.com>]

On Wed, 2012-09-05 at 07:57 -0400, Konrad Rzeszutek Wilk wrote:
> On Tue, Sep 04, 2012 at 04:57:56PM -0600, Shuah Khan wrote:
> > On Tue, 2012-09-04 at 17:05 -0400, Konrad Rzeszutek Wilk wrote:
> > > On Sun, Sep 02, 2012 at 08:14:17AM -0600, Shuah Khan wrote:
> > > > A recent dma mapping error analysis effort showed that a large precentage
> > > > of dma_map_single() and dma_map_page() returns are not checked for mapping
> > > > errors. Reference: https://lkml.org/lkml/2012/8/10/326
> > > > 
> > > 
> > > So were you able to catch some naughty drivers with this?
> > 
> > I did compile a complete list of drivers that don't check dma mapping
> > errors from my analysis. Are you interested in seeing the full analysis?
> 
> Yes, plus the authors of the drivers are probably interested in it as
> well.
> ..snip..

I compiled very detailed information on the nature of problems with dma
mapping error checking, ranging from not checking mapping errors to not
unmapping etc. I would like to share this information in such a way that
others can leverage the research I already did. One good way to
communicate this information would be on one of the kernel wikis. Ideas,
suggestions?

> > > I was initially thinking that this patch would contain a state for the driver
> > > of whether after map it has called dma_mapping_error. So this function would
> > > increment some internal state, and if dma_mapping_error on that specific dma_addr
> > > it would decrement it. If it never occured, then we would print on the unmap
> > > that the device never had called dma_mapping_error on said dma_addr?
> > 
> > That is a good idea. Let me see if I understand what you are saying
> > correctly. Add a new field to dma_debug_entry structure and keep state
> > and clear it if dma_mapping_error() is called. This will require adding
> > a debug interface for dma_mapping_error() which is not hard to do. Is
> > this close to what you are thinking?
> 
> Right. It is more complex than this patch but it should provide a nicer
> "trap" mechanism to alert driver writers that they are not checking DMA
> addresses properly.

I will go back and work on improving this patch. I will have to think
about how to track overflow buffer triggers as well without impacting
the production kernel. That is a separate effort.

-- Shuah