Re: [PATCH] fs: fix a data race in i_size_write/i_size_read

From: Al Viro
Date: Wed Feb 19 2020 - 00:23:40 EST

On Wed, Feb 19, 2020 at 12:08:40AM -0500, Qian Cai wrote:
> > On Feb 18, 2020, at 11:52 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > If aligned 64bit stores on 64bit host (note the BITS_PER_LONG ifdefs) end up
> > being split, the kernel is FUBAR anyway. Details, please - how could that
> > end up happening?
> My understanding is the compiler might decide to split the load into saying two 4-byte loads. Then, we might have,
> Load1
> Store
> Load2
> where the load value could be a garbage. Also, Marco (the KCSAN maintainer) who knew more of compiler than me mentioned that there is no guarantee that the store will not be split either. Thus, the WRITE_ONCE().

I would suggest
* if some compiler does that, ask the persons responsible for that
"optimization" which flags should be used to disable it.
* if they fail to provide such, educate them regarding the
usefulness of their idea
* if that does not help, don't use the bloody piece of garbage.

Again, is that pure theory (because I can't come up with
any reason why splitting a 32bit load would be any less legitimate
than doing the same to a 64bit one on a 64bit architecture),
or is there anything that really would pull that off?