lkml.org 
[lkml]   [2017]   [Apr]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
Subject[PATCH] mm, numa: Fix bad pmd by atomically check for pmd_trans_huge when marking page tables prot_numa
A user reported a bug against a distribution kernel while running
a proprietary workload described as "memory intensive that is not
swapping" that is expected to apply to mainline kernels. The workload
is read/write/modifying ranges of memory and checking the contents. They
reported that within a few hours that a bad PMD would be reported followed
by a memory corruption where expected data was all zeros. A partial report
of the bad PMD looked like

[ 5195.338482] ../mm/pgtable-generic.c:33: bad pmd ffff8888157ba008(000002e0396009e2)
[ 5195.341184] ------------[ cut here ]------------
[ 5195.356880] kernel BUG at ../mm/pgtable-generic.c:35!
....
[ 5195.410033] Call Trace:
[ 5195.410471] [<ffffffff811bc75d>] change_protection_range+0x7dd/0x930
[ 5195.410716] [<ffffffff811d4be8>] change_prot_numa+0x18/0x30
[ 5195.410918] [<ffffffff810adefe>] task_numa_work+0x1fe/0x310
[ 5195.411200] [<ffffffff81098322>] task_work_run+0x72/0x90
[ 5195.411246] [<ffffffff81077139>] exit_to_usermode_loop+0x91/0xc2
[ 5195.411494] [<ffffffff81003a51>] prepare_exit_to_usermode+0x31/0x40
[ 5195.411739] [<ffffffff815e56af>] retint_user+0x8/0x10

Decoding revealed that the PMD was a valid prot_numa PMD and the bad PMD
was a false detection. The bug does not trigger if automatic NUMA balancing
or transparent huge pages is disabled.

The bug is due a race in change_pmd_range between a pmd_trans_huge and
pmd_nond_or_clear_bad check without any locks held. During the pmd_trans_huge
check, a parallel protection update under lock can have cleared the PMD
and filled it with a prot_numa entry between the transhuge check and the
pmd_none_or_clear_bad check.

While this could be fixed with heavy locking, it's only necessary to
make a copy of the PMD on the stack during change_pmd_range and avoid
races. A new helper is created for this as the check if quite subtle and the
existing similar helpful is not suitable. This passed 154 hours of testing
(usually triggers between 20 minutes and 24 hours) without detecting bad
PMDs or corruption. A basic test of an autonuma-intensive workload showed
no significant change in behaviour.

Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Cc: stable@vger.kernel.org
---
include/asm-generic/pgtable.h | 25 +++++++++++++++++++++++++
mm/mprotect.c | 12 ++++++++++--
2 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/include/asm-generic/pgtable.h b/include/asm-generic/pgtable.h
index 1fad160f35de..597fa482cd4a 100644
--- a/include/asm-generic/pgtable.h
+++ b/include/asm-generic/pgtable.h
@@ -819,6 +819,31 @@ static inline int pmd_none_or_trans_huge_or_clear_bad(pmd_t *pmd)
}

/*
+ * Used when setting automatic NUMA hinting protection where it is
+ * critical that a numa hinting PMD is not confused with a bad PMD.
+ */
+static inline int pmd_none_or_clear_bad_unless_trans_huge(pmd_t *pmd)
+{
+ pmd_t pmdval = pmd_read_atomic(pmd);
+
+ /* See pmd_none_or_trans_huge_or_clear_bad for info on barrier */
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+ barrier();
+#endif
+
+ if (pmd_none(pmdval))
+ return 1;
+ if (pmd_trans_huge(pmdval))
+ return 0;
+ if (unlikely(pmd_bad(pmdval))) {
+ pmd_clear_bad(pmd);
+ return 1;
+ }
+ return 0;
+}
+
+
+/*
* This is a noop if Transparent Hugepage Support is not built into
* the kernel. Otherwise it is equivalent to
* pmd_none_or_trans_huge_or_clear_bad(), and shall only be called in
diff --git a/mm/mprotect.c b/mm/mprotect.c
index 8edd0d576254..821ff2904cdb 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -150,8 +150,16 @@ static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
unsigned long this_pages;

next = pmd_addr_end(addr, end);
- if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
- && pmd_none_or_clear_bad(pmd))
+
+ /*
+ * Automatic NUMA balancing walks the tables with mmap_sem
+ * held for read. It's possible a parallel update
+ * to occur between pmd_trans_huge and a pmd_none_or_clear_bad
+ * check leading to a false positive and clearing. Hence, it's
+ * necessary to atomically read the PMD value for all the
+ * checks.
+ */
+ if (!pmd_devmap(*pmd) && pmd_none_or_clear_bad_unless_trans_huge(pmd))
continue;

/* invoke the mmu notifier if the pmd is populated */
\
 
 \ /
  Last update: 2017-04-10 11:48    [W:0.168 / U:4.740 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site