An update below. Also, add Peter.
On 1/20/2025 9:20 PM, jane.chu@xxxxxxxxxx wrote:
Thanks for the quick comment!
On 1/20/2025 9:00 PM, Matthew Wilcox wrote:
On Mon, Jan 20, 2025 at 09:18:49PM -0700, Jane Chu wrote:
When a process consumes a UE in a page, the memory failure handler
attempts to collect information for a potential SIGBUS.
If the page is an anonymous page, page_mapped_in_vma(page, vma) is
invoked in order to
1. retrieve the vaddr from the process' address space,
2. verify that the vaddr is indeed mapped to the poisoned page,
where 'page' is the precise small page with UE.
It's been observed that when injecting poison to a non-head subpage
of an anonymous hugetlb page, no SIGBUS show up; while injecting to
the head page produces a SIGBUS. The casue is that, though
hugetlb_walk()
returns a valid pmd entry (on x86), but check_pte() detects mismatch
between the head page per the pmd and the input subpage. Thus the vaddr
is considered not mapped to the subpage and the process is not
collected
for SIGBUS purpose. This is the calling stack
collect_procs_anon
page_mapped_in_vma
page_vma_mapped_walk
hugetlb_walk
huge_pte_lock
check_pte
It seems that the most obvious place to fix the issue is by making
page_mapped_in_vma() hugetlb walk aware. The precise subpage in the
input is useful in providing PAGE_SIZE granularity vaddr.
I don't like this solution because it adds yet another special case for
hugetlb. If we don't split a PMD-mapped THP, we'd have the same
problem, right?
check_pte() would succeed if we set pvmw->pfn to folio_pfn() and
pvmw->nr_pages to folio_nr_pages(), right? I just don't know what else
might be affected by that.
I like one of these two options:
@@ -206,6 +206,7 @@ bool page_vma_mapped_walk(struct
page_vma_mapped_walk *pvmw)
pvmw->pte = hugetlb_walk(vma, pvmw->address, size);
if (!pvmw->pte)
return false;
+ pvmw->pte += pvmw->address & (size - PAGE_SIZE);
pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
if (!check_pte(pvmw))
(that needs a bit of tidying up; you can't just do that, but I think
you get the basic idea -- correct the pte to point to the precise page
instead of the hugetlb pfn)
That'll work, let me think about how to tidy it up. More below.
It appears that check_pte() is supposed to be able to check range
overlap for leaf pte among other things.
But the currently implementation doesn't do that. Fixing this takes
care of the subject issue.
More below.
The option I really prefer is much more work but matches our preferred
direction of getting rid of hugetlb specific code. Something like this:
@@ -192,27 +192,6 @@ bool page_vma_mapped_walk(struct
page_vma_mapped_walk *pvmw)
if (pvmw->pmd && !pvmw->pte)
return not_found(pvmw);
- if (unlikely(is_vm_hugetlb_page(vma))) {
- struct hstate *hstate = hstate_vma(vma);
- unsigned long size = huge_page_size(hstate);
- /* The only possible mapping was handled on last
iteration */
[...]
- pvmw->ptl = huge_pte_lock(hstate, mm, pvmw->pte);
- if (!check_pte(pvmw))
- return not_found(pvmw);
- return true;
- }
-
end = vma_address_end(pvmw);
if (pvmw->pte)
goto next_pte;
@@ -229,7 +208,19 @@ bool page_vma_mapped_walk(struct
page_vma_mapped_walk *pvmw continue;
}
pud = pud_offset(p4d, pvmw->address);
- if (!pud_present(*pud)) {
+ pude = *pud;
+ if (pud_trans_huge(pude) ||
+ (pud_present(pude) && pud_devmap(pude))) {
+ pvmw->ptl = pud_lock(mm, pvmw->pud);
+ ...
+ if (likely(pud_trans_huge(pude) ||
pud_devmap(pude))) {
+ if (pvmw->flags & PVMW_MIGRATION)
+ return not_found(pvmw);
+ if (!check_pud(pud_pfn(pude), pvmw))
+ return not_found(pvmw);
+ return true;
+ }
+ } else if (!pud_present(pude)) {
step_forward(pvmw, PUD_SIZE);
continue;
}
ie get rid of all the hugetlb-specific code, and add support for the
PUD level to the common code. You'd also need to write check_pud().
Good idea! I'd like to give this more generic approach a try as well.
I ran a simple experiment on page_vma_mapped_walk() checking on a PMD
level hugetlb page but with the hugetlb{} logic commented out.
The experiment includes both test/move_pages and the memory poison
tests, and the latter hit a NULL pointer deref. in try_to_unmap_one()
that I am looking into.
Besides, there are broader things to consider, such as pxd_trans_huge()
are defined under CONFIG_TRANSPARENT_HUGEPAGE that doesn't apply to
hugetlb. Mabybe try pxd_leaf()? but they're not the same thing, does
the difference matter?
Hugetlb doesn't use pxd_lock(), it has its own locking though after
peeling off the wrappers appear to be the similar thing, but not quite
the same thing.
There are also the hugetlb unique PMD sharing and implied locking
requirement.
Maybe there are more as I dig further, it's becoming more like a project
on its own which I will update on a separate thread in future.
Thanks,
-jane
I'll understand if you don't want to do all the extra work. And
thanks for tracking down this bug.
Thanks a lot!
-jane
