Re: WARNING: at mm/mremap.c:211 move_page_tables in i386

[Matthew Wilcox]

On Sat, Jul 11, 2020 at 11:12:58AM -0700, Linus Torvalds wrote:
> Yeah, that's just the commit that enables the code, not the commit
> that introduces the fundamental problem.
> 
> That said, this is a prime example of why I absolutely detest patch
> series that do this kind of thing, and are several patches that create
> new functionality, followed by one patch to enable it.
> 
> If you can't get things working incrementally, maybe you shouldn't do
> them at all. Doing a big series of "hidden work" and then enabling it
> later is wrong.

I'm struggling with exactly this for my current THP-in-pagecache patches.
There are about fifty patches, each fixing something that won't work if
the page is a THP.  And then there's the one patch which actually starts
creating THPs, and that's the one patch any bisect will point to.

But I don't see any other way to do it.  It's not like I can put THPs
in the page cache before fixing the things that won't work.

This probably wasn't the kind of thing you had in mind when you wrote
the above, but if you had some advice for my situation, I'd welcome it.