On 15.11.22 19:03, Peter Xu wrote:
On Tue, Nov 15, 2022 at 06:22:03PM +0100, David Hildenbrand wrote:
That's precisely what I had in mind recently, and I am happy to hear that you have similar idea:
https://lkml.kernel.org/r/20221108174652.198904-6-david@redhat.com
" Note that we don't optimize for the actual migration case: (1) When migration succeeds the new PTE will not be writable because the source PTE was not writable (protnone); in the future we might just optimize that case similarly by reusing can_change_pte_writable()/can_change_pmd_writable() when removing migration PTEs. "
I see, sorry I haven't yet read it, but sounds doable indeed.
Currently, "readable_migration_entry" is even wrong: it might be PROT_NONE and not even readable.
Do you mean mprotect(PROT_NONE)?
If we read the "read migration entry" as "migration entry with no write bit", it seems still fine, and code-wise after pte recovered it should still be PROT_NONE iiuc because mk_pte() will just make a pte without e.g. _PRESENT bit set on x86 while it'll have the _PROT_NONE bit.
Exactly that's the unintuitive interpretation of "readable_migration_entry". By "wrong" I meant: the naming is wrong.
May not keep true for numa balancing though: when migration happens after a numa hint applied to a pte, it seems to me it's prone to lose the hint after migration completes (assuming this migration is not the numa balancing operation itself caused by a page access). Doesn't sound like a severe issue though even if I didn't miss something, since if the page got moved around the original hint may need to reconsider anyway.
Yes, I think any migration will lose fake PROT_NONE. "Fake" as in "not VMA permissions" but "additional permissions imposed by NUMA hinting faults."