Sorry, I missed out Dave's email, so now I'm taking my time to page (hah!) all of this.
On 25/07/23 15:21, Peter Zijlstra wrote:
On Mon, Jul 24, 2023 at 10:40:04AM -0700, Dave Hansen wrote:
TLB flushes for freed page tables are another game entirely. The CPU is free to cache any part of the paging hierarchy it wants at any time. It's also free to set accessed and dirty bits at any time, even for instructions that may never execute architecturally.
That basically means that if you have *ANY* freed page table page *ANYWHERE* in the page table hierarchy of any CPU at any time ... you're screwed.
There's no reasoning about accesses or ordering. As soon as the CPU does *anything*, it's out to get you.
OK, I feel like I need to go back do some more reading now, but I think I get the difference. Thanks for spelling it out.
You're going to need to do something a lot more radical to deal with free page table pages.
Ha! IIRC the only thing we can reasonably do there is to have strict per-cpu page-tables such that NOHZ_FULL CPUs can be isolated. That is, as long we the per-cpu tables do not contain -- and have never contained -- a particular table page, we can avoid flushing it. Because if it never was there, it also couldn't have speculatively loaded it.
Now, x86 doesn't really do per-cpu page tables easily (otherwise we'd have done them ages ago) and doing them is going to be *major* surgery and pain.
Other than that, we must take the TLBI-IPI when freeing page-table-pages.
But yeah, I think Nadav is right, vmalloc.c never frees page-tables (or at least, I couldn't find it in a hurry either), but if we're going to be doing this, then that file must include a very prominent comment explaining it must never actually do so either.
I also couldn't find any freeing of the page-table-pages, I'll do another pass and sharpen my quill for a big fat comment.
Not being able to free page-tables might be a 'problem' if we're going to be doing more of HUGE_VMALLOC, because that means it becomes rather hard to swizzle from small to large pages.