Re: [PATCH 03/16] dma-debug: add hash functions fordma_debug_entries

[FUJITA Tomonori]

On Wed, 14 Jan 2009 18:48:04 +0100
Ingo Molnar <mingo@xxxxxxx> wrote:

> 
> * Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Wed, 14 Jan 2009 12:43:47 +0100 Ingo Molnar <mingo@xxxxxxx> wrote:
> > 
> > > 
> > > * Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > 
> > > > On Fri, 9 Jan 2009 17:19:17 +0100 Joerg Roedel <joerg.roedel@xxxxxxx> wrote:
> > > > 
> > > > > +struct hash_bucket {
> > > > > +	struct list_head list;
> > > > > +	spinlock_t lock;
> > > > > +} ____cacheline_aligned;
> > > > 
> > > > __cacheline_aligned_in_smp.
> > > > 
> > > > This all looks like an exotically large amount of code for a debug 
> > > > thingy?
> > > 
> > > this code checks the DMA usage of ~1 million lines of kernel code - all 
> > > the DMA using drivers. I think Joerg's feature is hugely relevant as DMA 
> > > scribbles are one of the hardest to debug kernel bugs: they can end up in 
> > > permanent data corruption or other hard to find bugs. In fact i think his 
> > > patchset is rather simple and even having 10 times as much debug code 
> > > would pay for its existence in the long run.
> > > 
> > 
> > Have we previously found bugs by other means which this facility would 
> > have detected?  I don't recall any...
> 
> btw., during the past decade we have had countless very ugly driver DMA 
> bugs in the past that took vendors months and specialized equipment to 
> track down.
> 
> I cannot give you a number breakdown, only an impression: storage drivers 
> tended to be the hardest hit (due to the severity of the bugs and due to 
> their inherent complexity) - but DMA bugs in networking drivers can be 
> hard to track down too.

We have had ugly DMA bugs in scsi drivers but I think that this new
feature can find very few of them. I can't think of any SCSI driver
bugs that this could find. This feature can't find any popular DMA
bugs in scsi drivers, such as messing up driver's dma descriptor from
a scatter gather list.

But this can find some kinds of DMA bugs and it's is just about 1,000
lines. I don't see any problem about merging this. 1,000 lines it too
large? Maybe cutting just 1,000 lines up into too many pieces is
deceptive.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/