Thanks, Ivan.
In the process of reviewing this, I starting thinking if the !shmem case was
also susceptible to a similar race, and I *think* it might be. Unfortunately, my
time has ran out, and I haven't been able to validate ; I'm less familiar with
the file-side of things.
The underlying problem is race with truncation/hole-punch under OOM condition.
The nice do-while loop near the top of collapse_file() attempts to avoid this
scenario by making sure enough slots are available. However, when we drop xarray
lock, we open ourselves up to concurrent removal + slot deletion. Those slots
then need to be allocated again -- which under OOM condition is failable.
The syzbot reproducer picks on shmem, but I think this can occur for file as
well. If we find a hole, we unlock the xarray and call
page_cache_sync_readahead(), which if it succeeds, IIUC, will have allocated a
new slot in our mapping pointing to the new page. We *then* locks the page. Only
after the page is locked are we protected from concurrent removal (Note: this is
what provides us protection in many of the xas_store() cases ; we've held the
slot's contained page-lock since verifying the slot exists, protecting us from
removal / reallocation races).
Maybe I'm just low on caffeine at the end of the day, and am missing something,
but if I had more time, I'd be looking into the file-side some more to verify.
Apologies that hasn't occurred to me until now ; I was looking at one of your
comments and double-checked why I *thought* we were safe.
Anyways, irrespective of that looming issues, some more notes to follow:
The 'xas_store' call during page cache scanning can potentially
translate 'xas' into the error state (with the reproducer provided
by the syzkaller the error code is -ENOMEM). However, there are no
further checks after the 'xas_store', and the next call of 'xas_next'
at the start of the scanning cycle doesn't increase the xa_index,
and the issue occurs.
This patch will add the xarray state error checking after the
'xas_store' and the corresponding result error code. It will
also add xarray state error checking via WARN_ON_ONCE macros,
to be sure that ENOMEM or other possible errors don't occur
at the places they shouldn't.
Thanks for the additions here. I think it's worthwhile providing even more
details about the specifics of the race we are fixing and/or guarding against to
help ppl understand how that -ENOMEM comes about if the do-while loop has
"Ensured" we have slots available (additionally, I think that comment can be
augmented).
Tested via syzbot.
Reported-by: syzbot+9578faa5475acb35fa50@xxxxxxxxxxxxxxxxxxxxxxxxx
Link: https://syzkaller.appspot.com/bug?id=7d6bb3760e026ece7524500fe44fb024a0e959fc
Signed-off-by: Ivan Orlov <ivan.orlov0322@xxxxxxxxx>
---
V1 -> V2: Add WARN_ON_ONCE error checking and comments
mm/khugepaged.c | 20 ++++++++++++++++++++
1 file changed, 20 insertions(+)
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 92e6f56a932d..8b6580b13339 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -55,6 +55,7 @@ enum scan_result {
SCAN_CGROUP_CHARGE_FAIL,
SCAN_TRUNCATED,
SCAN_PAGE_HAS_PRIVATE,
+ SCAN_STORE_FAILED,
};
I'm still reluctant to add a new error code for this as this seems like quite a
rare race that requires OOM to trigger. I'd be happier just reusing SCAN_FAIL,
or, something we might get some millage out of later: SCAN_OOM.
Also, a reminder to update include/trace/events/huge_memory.h, if you go that
route.
#define CREATE_TRACE_POINTS
@@ -1840,6 +1841,15 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
goto xa_locked;
}
xas_store(&xas, hpage);
+ if (xas_error(&xas)) {
+ /* revert shmem_charge performed
+ * in the previous condition
+ */
Nit: Here, and following, I think standard convention for multiline comment is
to have an empty first and last line, eg:
+ /*
+ * revert shmem_charge performed
+ * in the previous condition
+ */
Though, checkpatch.pl --strict didn't seem to care.
+ mapping->nrpages--;
+ shmem_uncharge(mapping->host, 1);
+ result = SCAN_STORE_FAILED;
+ goto xa_locked;
+ }
nr_none++;
continue;
}
@@ -1992,6 +2002,11 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
/* Finally, replace with the new page. */
xas_store(&xas, hpage);
+ /* We can't get an ENOMEM here (because the allocation happened before)
+ * but let's check for errors (XArray implementation can be
+ * changed in the future)
+ */
+ WARN_ON_ONCE(xas_error(&xas));
Nit: it's not just that allocation happened before -- need some guarantee we've
been protected from concurrent removal. This is what made me look at the file
side.
continue;
out_unlock:
unlock_page(page);
@@ -2029,6 +2044,11 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
/* Join all the small entries into a single multi-index entry */
xas_set_order(&xas, start, HPAGE_PMD_ORDER);
xas_store(&xas, hpage);
+ /* Here we can't get an ENOMEM (because entries were
+ * previously allocated) But let's check for errors
+ * (XArray implementation can be changed in the future)
+ */
+ WARN_ON_ONCE(xas_error(&xas));
Ditto.
Apologies I won't be around to see this change through -- I'm just out of time,
and will be shutting my computer down tomorrow for 3 months. Sorry for the poor
timing, for raising issues, then disappearing. Hopefully I'm wrong and the
file-side isn't a concern.
Best,
Zach
xa_locked:
xas_unlock_irq(&xas);
xa_unlocked:
--
2.34.1
