Re: [PATCH v3] media: uvc: don't do DMA on stack

['Alan Stern' <stern@rowland.harvard.edu>]

On Tue, Jun 22, 2021 at 02:21:27PM +0000, David Laight wrote:
> From: Alan Stern
> > Sent: 22 June 2021 14:29
> ...
> > > Thought...
> > >
> > > Is kmalloc(1, GFP_KERNEL) guaranteed to return a pointer into
> > > a cache line that will not be accessed by any other code?
> > > (This is slightly weaker than requiring a cache-line aligned
> > > pointer - but very similar.)
> > 
> > As I understand it, on architectures that do not have cache-coherent
> > I/O, kmalloc is guaranteed to return a buffer that is
> > cacheline-aligned and whose length is a multiple of the cacheline
> > size.
> > 
> > Now, whether that buffer ends up being accessed by any other code
> > depends on what your driver does with the pointer it gets from
> > kmalloc.  :-)
> 
> Thanks for the clarification.
> 
> Most of the small allocates in the usb stack are for transmits
> where it is only necessary to ensure a cache write-back.
> 
> I know there has been some confusion because one of the
> allocators can add a small header to every allocation.
> This can lead to unexpectedly inadequately aligned pointers.
> If it is updated when the preceding block is freed (as some
> user-space mallocs do) then it would need to be in a
> completely separate cache line.

If you really want to find out what the true story is, you should ask 
on the linux-mm mailing list.  The rest of us are not experts on this 
stuff.

Alan Stern