Re: [linux-usb-devel] Re: serious 2.6 bug in USB subsystem?

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Wed, 10 Mar 2004 08:22:26 -0800, David Brownell <david-b@pacbell.net> said:

  David.B> It won't add new BUG_ON calls (WARN at worst)

I put them there mostly as assertions.  What I'd really want there is
a DEBUG_BUG_ON, which is more like assert() in user-land (i.e.,
production code would drop the checks).  WARN_ON() would be fine, too.

  >> The current OHCI relies on the internals of the dma_pool()
  >> implementation.  If the implementation changed to, say, modify
  >> the memory that is free or, heaven forbid, return the memory to
  >> the kernel, you'd get _extremely_ difficult to track down race
  >> conditions.

  David.B> It'd be good if you said _how_ you think it relies on such
  David.B> internals.

I thought I did.  Suppose somebody changed the dma_pool code such that
it would overwrite freed memory with an 0xf00000000000000 pattern.  If
the HC can still hold a reference to a freed ED (it can without my
patch), the HC could see this kind of ED:

 hw=(info\0000000 tailpð000000 headp\0000000 nextEDð000000)

If so, the HC would go ahead and try to interpret the memory at
address 0 as a transfer descriptor.  Depending on the memory contents,
this could cause silent data corruption at an arbitrary address.

  >> - thus you might get a case where hwTailP is 0 but hwHeadP is
  >> non-zero, which would cause the HC to happily start dereferencing
  >> the descriptor

  David.B> If you assume a bug where the ED is freed but still in use,
  David.B> that's hardly the only thing that'd go wrong!!  You can't
  David.B> use such a potential bug to prove something else is broken.

You lost me here.  All I'm saying is that the current code has a
dangerous race that can corrupt memory, crash machines, or have all
sorts of other wild side-effects.  I never claimed this bug had
anything todo with the BTC keyboard problem.

	--david