On Fri, Jun 02, 2023 at 10:42:45PM +0500, Muhammad Usama Anjum wrote:
Thank you for reviewing and helping out.
On 6/2/23 9:47 PM, Peter Xu wrote: [...]
static inline void make_uffd_wp_huge_pte(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep, pte_t ptent) { if (is_hugetlb_entry_hwpoisoned(ptent) || is_pte_marker(ptent)) return;
if (is_hugetlb_entry_migration(ptent)) set_huge_pte_at(vma->vm_mm, addr, ptep, pte_swp_mkuffd_wp(ptent)); else if (!huge_pte_none(ptent)) ptep_modify_prot_commit(vma, addr, ptep, ptent, huge_pte_mkuffd_wp(ptent)); else set_huge_pte_at(vma->vm_mm, addr, ptep, make_pte_marker(PTE_MARKER_UFFD_WP)); }
the is_pte_marker() check can be extended to double check pte_marker_uffd_wp() bit, but shouldn't matter a lot since besides the uffd-wp bit currently we only support swapin error which should sigbus when accessed, so no point in tracking anyway.
Yeah, we are good with what we have as even if more bits are supported in pte markers, this function is only reached when UNPOPULATED + ASYNC WP are enabled. So no other bit would be set on the marker.
I think we don't know? swapin error bit can be set there afaiu, if someone swapoff -a and found a swap device broken for some swapped out ptes.
Oops, so I remove returning on pte marker detection. Instead the last else is already setting marker wp-ed.
What I meant is your current code is fine, please don't remove the check.
Removing is_pte_marker() means you could potentially overwrite one swap error entry with a marker only having uffd-wp bit set. It can corrupt the guest because swapin error is not recoverable and higher priority than wp.
But again I think that's fine for now.
As we always set UNPOPULATED, so markers are always set on none ptes initially. Is it possible that a none pte becomes present, then swapped and finally none again? So I'll do the following addition for make_uffd_wp_pte():
--- a/fs/proc/task_mmu.c +++ b/fs/proc/task_mmu.c @@ -1800,6 +1800,9 @@ static inline void make_uffd_wp_pte(struct vm_area_struct *vma, } else if (is_swap_pte(ptent)) { ptent = pte_swp_mkuffd_wp(ptent); set_pte_at(vma->vm_mm, addr, pte, ptent);
- } else {
set_pte_at(vma->vm_mm, addr, pte,make_pte_marker(PTE_MARKER_UFFD_WP));}
Makes sense, you can leverage userfaultfd_wp_use_markers() here, and you should probably keep the protocol (only set the marker when WP_UNPOPULATED for anon).
This function is only reachable when UNPOPULATED + Async WP are set. So we don't need to use userfaultfd_wp_use_markers().
I don't remember where you explicitly checked that to make sure it'll always be the case, but if you do it'll still be nice if you can add a comment right above here explaining.
I was only testing for WP and WP Async in pagemap_scan_test_walk() as WP async can only be enabled if unpopulated is enabled which in turn enables the markers. I'll add userfaultfd_wp_use_markers() to pagemap_scan_test_walk().
OK.
Or, maybe you can still use userfaultfd_wp_use_markers() here, then you can just WARN_ON_ONCE() on it, which looks even better?
[...]
Hmm, is it a bug for pagemap? pagemapread.buffer should be linear to the address range to be scanned to me. If it skips some unstable pmd without filling in anything it seems everything later will be shifted with PMD_SIZE.. I had a feeling that it should set walk->action==ACTION_AGAIN before return.
I don't think this is a bug if this is how it was implemented in the first place. In this task_mmu.c file, we can find several examples of the same pattern that error isn't returned if pmd_trans_unstable() succeeds.
I don't see why multiple same usages mean it's correct.. maybe they're all buggy?
I can post a patch for this to collect opinions to see if I missed something. I'd suggest you figure out what's the right thing to do for the new interface and make it right from the start, no matter how it was implemented elsewhere.
Alright. At first sight, it seems I should return -EAGAIN here. But there maybe a case where there are 3 THPs. I've cleared WP over the first THP and put data in the user buffer. If I return -EAGAIN, the data in the user buffer would be not used by the user correctly as negative value has been returned. So the best way here would be to skip this second VMA and don't abort the operation. Thus we'll not be giving any information about the second THP as it was in transition.
I'll think about it again before sending next version.
Here when reaching unstable==true I think it means a rare race happened, that a thp was _just_ installed. You can try to read the comment over pmd_trans_unstable() and pmd_none_or_trans_huge_or_clear_bad() to understand what it wants to avoid.
Basically afaiu that's the ultimate guard if one wants to walk the pte pgtable (that fact can change again if Hugh's series will land, but that's another story).
As said, a patch will be posted (will have you copied) soon just for that, so people can also comment and may help you / us know which is the right way, probably not in a few hours (need to go very soon..), but shouldn't be too long. I just need to double check more users out of task_mmu.c when we're at it.