mm/pagewalk: fix bootstopping regression from extra pte_unmap()

Commit Message

Hugh Dickins Sept. 2, 2023, 3:29 p.m. UTC

Mikhail reports early-6.6-based Fedora Rawhide not booting: "rcu_preempt
detected expedited stalls", minutes wait, and then hung_task splat while
kworker trying to synchronize_rcu_expedited().  Nothing logged to disk.

He bisected to my 6.6 a349d72fd9ef ("mm/pgtable: add rcu_read_lock() and
rcu_read_unlock()s"): but the one to blame is my 6.5 commit to fix the
espfix "bad pmd" warnings when booting x86_64 with CONFIG_EFI_PGT_DUMP=y.

Gaah, that added an "addr >= TASK_SIZE" check to avoid pte_offset_map(),
but failed to add the equivalent check when choosing to pte_unmap().

It's not a problem on 6.5 (for different reasons, it's harmless on both
64-bit and 32-bit), but becomes a bootstopper on 6.6 with the unbalanced
rcu_read_unlock() - RCU has a WARN_ON_ONCE for that, but it would have
scrolled off Mikhail's console too quickly.

Reported-by: Mikhail Gavrilov <[e-mail]>
Closes: https://lore.kernel.org/linux-mm/[e-mail]/
Fixes: 8b1cb4a2e819 ("mm/pagewalk: fix EFI_PGT_DUMP of espfix area")
Fixes: a349d72fd9ef ("mm/pgtable: add rcu_read_lock() and rcu_read_unlock()s")
Signed-off-by: Hugh Dickins <[e-mail]>
Tested-by: Mikhail Gavrilov <[e-mail]>
---
 mm/pagewalk.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Patch

diff --git a/mm/pagewalk.c b/mm/pagewalk.c
index 9b2d23fbf4d3..b7d7e4fcfad7 100644
--- a/mm/pagewalk.c
+++ b/mm/pagewalk.c
@@ -58,7 +58,7 @@  static int walk_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 			pte = pte_offset_map(pmd, addr);
 		if (pte) {
 			err = walk_pte_range_inner(pte, addr, end, walk);
-			if (walk->mm != &init_mm)
+			if (walk->mm != &init_mm && addr < TASK_SIZE)
 				pte_unmap(pte);
 		}
 	} else {