Messages in this thread | | | Date | Mon, 11 May 2020 00:25:41 +0200 | From | Joerg Roedel <> | Subject | Re: [git pull] IOMMU Fixes for Linux v5.7-rc4 |
| |
On Sun, May 10, 2020 at 11:34:49AM -0700, Linus Torvalds wrote: > On Sun, May 10, 2020 at 5:26 AM Joerg Roedel <joro@8bytes.org> wrote: > > > > The first race condition was around > > the non-atomic update of the domain page-table root pointer > > and the variable containing the page-table depth (called > > mode). This is fixed now be merging page-table root and mode > > into one 64-bit field which is read/written atomically. > > This seems a bit odd. > > The pointer part is always page-aligned, and the "mode" is just three bits. > > Why isn't it just encoded as one pointer with the low three bits being the mode? > > The thing is, the 64-bit atomic reads/writes are very expensive on > 32-bit x86. If it was just a native pointer, it would be much cheaper > than an "atomic64_t".
Yeah, when I think about it again, you are right. I think I used atomic64_t just to be on the safe side with memory odering and all. But in this case it doesn't really matter when a reader observes the update, it is only important that the reader does not observe one field updated while the other is not. And that should already be fullfilled with 64-bit writes on x86-64, like a native pointer write.
I'll send a patch to Qian to test this, just to be sure I am not missing anything.
Thanks,
Joerg
| |