Re: [GIT PULL] percpu fixes for v5.14-rc1

[Linus Torvalds]

On Wed, Jul 7, 2021 at 6:00 AM Dennis Zhou <dennis@xxxxxxxxxx> wrote:
>
> This is just a single change to fix percpu depopulation. The code relied
> on depopulation code written specifically for the free path and relied
> on vmalloc to do the tlb flush lazily. As we're modifying the backing
> pages during the lifetime of a chunk, we need to also flush the tlb
> accordingly.

I pulled this, but I ended up unpulling after looking at the fix.

The fix may be perfectly correct, but I'm looking at that
pcpu_reclaim_populated() function, and I want somebody to explain to
me what it's ok to drop and re-take the 'pcpu_lock' and just continue.

Because whatever it was protecting is now not protected any more.

It *looks* like it's intended to protect the pcpu_chunk_lists[]
content, and some other functions that do this look ok. So for
example, pcpu_balance_free() at least removes the 'chunk' from the
pcpu_chunk_lists[] before it drops the lock and then works on the
chunk contents.

But pcpu_reclaim_populated() seems to *leave* chunk on the
pcpu_chunk_lists[], drop the lock, and then continue to use 'chunk'.

That odd "release lock and continue to use the data it's supposed to
protect" seems to be pre-existing, but

 (a) this is the code that caused problems to begin with

and

 (b) it seems to now happen even more.

So maybe this code is right. But it looks very odd to me, and I'd like
to get more explanations of _why_ it would be ok before I pull this
fix, since there seems to be a deeper underlying problem in the code
that this tries to fix.

                 Linus