On Thu, 2018-09-06 at 13:56 -0700, Linus Torvalds wrote:
> 
> A lot of that was that the actual *fixes* were marked for stable, but
> quite often they were preceded by cleanups and other updates that
> didn't actually fix things directly, and that weren't in themselves
> explicitly marked for stable and didn't have a Fixes: tag, because
> they were prep-work.
> 
> So we had _several_ nasty regressions in stable that never showed up
> in mainline, because there was some non-obvious dependency that didn't
> cause a merge conflict, but did cause a "this commit needed that other
> commit to work right".

This time round, for L1TF I think we did relatively OK on the stable
backports. The dependencies we were missing were stuff that was just
randomly cleaned up in the intervening kernels.

For example, the 'Simplify p[g4um]d_page() macros' commit (fd7e315988)
start out as *purely* a cosmetic improvement to reduce duplication by
introducing a pgd_pfn() macro.

Some time later, completely unpredicted at the time of that commit, we
came along and made pgd_pfn() do something different. It's *those*
"dependencies" which tend to get missed, and they can't be predicted at
the time of the original commit.

I don't think we did so badly with the commits which were done as
cleanups in preparation for the event itself.