[PATCH] powerpc: mm: fixup preempt undefflow with huge pages

hugepd_free() used __get_cpu_var() once. Nothing ensured that the code
accessing the variable did not migrate from one CPU to another and soon
this was noticed by Tiejun Chen in 94b09d755462 ("powerpc/hugetlb:
Replace __get_cpu_var with get_cpu_var"). So we had it fixed.

Christoph Lameter was doing his __get_cpu_var() replaces and forgot
PowerPC. Then he noticed this and sent his fixed up batch again which
got applied as 69111bac42f5 ("powerpc: Replace __get_cpu_var uses").

The careful reader will noticed one little detail: get_cpu_var() got
replaced with this_cpu_ptr(). So now we have a put_cpu_var() which does
a preempt_enable() and nothing that does preempt_disable() so we
underflow the preempt counter.

Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Christoph Lameter <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Tiejun Chen <redacted>
Cc: <redacted>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 arch/powerpc/mm/hugetlbpage.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/powerpc/mm/hugetlbpage.c b/arch/powerpc/mm/hugetlbpage.c
index 744e24bcb85c..9e8919b39640 100644
--- a/arch/powerpc/mm/hugetlbpage.c
+++ b/arch/powerpc/mm/hugetlbpage.c

@@ -414,7 +414,7 @@ static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
 {
 	struct hugepd_freelist **batchp;
 
-	batchp = this_cpu_ptr(&hugepd_freelist_cur);
+	batchp = get_cpu_ptr(&hugepd_freelist_cur);
 
 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
 	    cpumask_equal(mm_cpumask(tlb->mm),

-- 
2.7.0

Sebastian Andrzej Siewior [off-list ref] writes:

IMHO it would better if we do

batchp = get_cpu_var(hugepd_freelist_cpu);

so that it match the existing

put_cpu_var(hugepd_freelist_cur);

While you are there, can you also fix the wrong indentation on line 423
?

-aneesh

On Mon, 2016-03-07 at 21:04 +0530, Aneesh Kumar K.V wrote:

Sebastian Andrzej Siewior [off-list ref] writes:

While you are there, can you also fix the wrong indentation on line
423
?

 .../...

Also this looks like stable material no ?

Cheers,
Ben.

On 03/08/2016 12:41 AM, Benjamin Herrenschmidt wrote:

On Mon, 2016-03-07 at 21:04 +0530, Aneesh Kumar K.V wrote:

quoted

Sebastian Andrzej Siewior [off-list ref] writes:
 
While you are there, can you also fix the wrong indentation on line
423
?

 .../...

Also this looks like stable material no ?

Yes, I had a stable tag attached. I am going to resend it with the
additional changes Aneesh asked for.

Cheers,
Ben.

Sebastian

hugepd_free() used __get_cpu_var() once. Nothing ensured that the code
accessing the variable did not migrate from one CPU to another and soon
this was noticed by Tiejun Chen in 94b09d755462 ("powerpc/hugetlb:
Replace __get_cpu_var with get_cpu_var"). So we had it fixed.

Christoph Lameter was doing his __get_cpu_var() replaces and forgot
PowerPC. Then he noticed this and sent his fixed up batch again which
got applied as 69111bac42f5 ("powerpc: Replace __get_cpu_var uses").

The careful reader will noticed one little detail: get_cpu_var() got
replaced with this_cpu_ptr(). So now we have a put_cpu_var() which does
a preempt_enable() and nothing that does preempt_disable() so we
underflow the preempt counter.

Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Christoph Lameter <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: <redacted>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
v1=E2=80=A6v2: - use get_cpu_var() instead of get_cpu_ptr()
       - correct indentation of put_cpu_var()

 arch/powerpc/mm/hugetlbpage.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/mm/hugetlbpage.c b/arch/powerpc/mm/hugetlbpage.c
index 744e24bcb85c..4a811ca7ac9d 100644
--- a/arch/powerpc/mm/hugetlbpage.c
+++ b/arch/powerpc/mm/hugetlbpage.c

@@ -414,13 +414,13 @@ static void hugepd_free(struct mmu_gather *tlb, void =

*hugepte)
 {
 	struct hugepd_freelist **batchp;
=20
-	batchp =3D this_cpu_ptr(&hugepd_freelist_cur);
+	batchp =3D &get_cpu_var(hugepd_freelist_cur);
=20
 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
 	    cpumask_equal(mm_cpumask(tlb->mm),
 			  cpumask_of(smp_processor_id()))) {
 		kmem_cache_free(hugepte_cache, hugepte);
-        put_cpu_var(hugepd_freelist_cur);
+		put_cpu_var(hugepd_freelist_cur);
 		return;
 	}
=20
--=20
2.7.0

Sebastian Andrzej Siewior [off-list ref] writes:

[ text/plain ]
hugepd_free() used __get_cpu_var() once. Nothing ensured that the code
accessing the variable did not migrate from one CPU to another and soon
this was noticed by Tiejun Chen in 94b09d755462 ("powerpc/hugetlb:
Replace __get_cpu_var with get_cpu_var"). So we had it fixed.

Christoph Lameter was doing his __get_cpu_var() replaces and forgot
PowerPC. Then he noticed this and sent his fixed up batch again which
got applied as 69111bac42f5 ("powerpc: Replace __get_cpu_var uses").

The careful reader will noticed one little detail: get_cpu_var() got
replaced with this_cpu_ptr(). So now we have a put_cpu_var() which does
a preempt_enable() and nothing that does preempt_disable() so we
underflow the preempt counter.

Reviewed-by: Aneesh Kumar K.V <redacted>

d *hugepte)

 {
 	struct hugepd_freelist **batchp;
=20=20
-	batchp =3D this_cpu_ptr(&hugepd_freelist_cur);
+	batchp =3D &get_cpu_var(hugepd_freelist_cur);
=20=20
 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
 	    cpumask_equal(mm_cpumask(tlb->mm),
 			  cpumask_of(smp_processor_id()))) {
 		kmem_cache_free(hugepte_cache, hugepte);
-        put_cpu_var(hugepd_freelist_cur);
+		put_cpu_var(hugepd_freelist_cur);
 		return;
 	}
=20=20
--=20
2.7.0

On 2016-03-10 01:04:24 [+0530], Aneesh Kumar K.V wrote:

Sebastian Andrzej Siewior [off-list ref] writes:

*ping*
http://patchwork.ozlabs.org/patch/593943/

quoted

[ text/plain ]
hugepd_free() used __get_cpu_var() once. Nothing ensured that the code
accessing the variable did not migrate from one CPU to another and soon
this was noticed by Tiejun Chen in 94b09d755462 ("powerpc/hugetlb:
Replace __get_cpu_var with get_cpu_var"). So we had it fixed.

Christoph Lameter was doing his __get_cpu_var() replaces and forgot
PowerPC. Then he noticed this and sent his fixed up batch again which
got applied as 69111bac42f5 ("powerpc: Replace __get_cpu_var uses").

The careful reader will noticed one little detail: get_cpu_var() got
replaced with this_cpu_ptr(). So now we have a put_cpu_var() which does
a preempt_enable() and nothing that does preempt_disable() so we
underflow the preempt counter.

=20
Reviewed-by: Aneesh Kumar K.V <redacted>
=20

quoted

Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Christoph Lameter <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: <redacted>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
v1=E2=80=A6v2: - use get_cpu_var() instead of get_cpu_ptr()
       - correct indentation of put_cpu_var()

 arch/powerpc/mm/hugetlbpage.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/mm/hugetlbpage.c b/arch/powerpc/mm/hugetlbpag=

e.c

quoted

index 744e24bcb85c..4a811ca7ac9d 100644

--- a/arch/powerpc/mm/hugetlbpage.c
+++ b/arch/powerpc/mm/hugetlbpage.c

@@ -414,13 +414,13 @@ static void hugepd_free(struct mmu_gather *tlb, v=

oid *hugepte)

quoted

 {
 	struct hugepd_freelist **batchp;
=20
-	batchp =3D this_cpu_ptr(&hugepd_freelist_cur);
+	batchp =3D &get_cpu_var(hugepd_freelist_cur);
=20
 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
 	    cpumask_equal(mm_cpumask(tlb->mm),
 			  cpumask_of(smp_processor_id()))) {
 		kmem_cache_free(hugepte_cache, hugepte);
-        put_cpu_var(hugepd_freelist_cur);
+		put_cpu_var(hugepd_freelist_cur);
 		return;
 	}
=20

Sebastian

On Tue, 2016-03-29 at 15:40 +0200, Sebastian Andrzej Siewior wrote:

On 2016-03-10 01:04:24 [+0530], Aneesh Kumar K.V wrote:

quoted

Sebastian Andrzej Siewior [off-list ref] writes:

*ping*
http://patchwork.ozlabs.org/patch/593943/

*pong*

The merge window just closed, I'm still recovering.

I've got it in my fixes branch locally, I'll probably push that today to
linux-next.

cheers

On 03/30/2016 02:41 AM, Michael Ellerman wrote:

The merge window just closed, I'm still recovering.

I've got it in my fixes branch locally, I'll probably push that today to
linux-next.

Thank you.

cheers

Sebastian

On Tue, 2016-08-03 at 09:03:56 UTC, Sebastian Siewior wrote:

hugepd_free() used __get_cpu_var() once. Nothing ensured that the code
accessing the variable did not migrate from one CPU to another and soon
this was noticed by Tiejun Chen in 94b09d755462 ("powerpc/hugetlb:
Replace __get_cpu_var with get_cpu_var"). So we had it fixed.

Christoph Lameter was doing his __get_cpu_var() replaces and forgot
PowerPC. Then he noticed this and sent his fixed up batch again which
got applied as 69111bac42f5 ("powerpc: Replace __get_cpu_var uses").

The careful reader will noticed one little detail: get_cpu_var() got
replaced with this_cpu_ptr(). So now we have a put_cpu_var() which does
a preempt_enable() and nothing that does preempt_disable() so we
underflow the preempt counter.

Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Christoph Lameter <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: <redacted>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Aneesh Kumar K.V <redacted>

Applied to powerpc fixes, thanks.

https://git.kernel.org/powerpc/c/08a5bb2921e490939f78f38fd0

cheers