memtest failed to find bad memory when compiled with clang. So use
{WRITE,READ}_ONCE to access memory to avoid compiler over optimization.
Cc: <[email protected]>
Signed-off-by: Qiang Zhang <[email protected]>
---
mm/memtest.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/memtest.c b/mm/memtest.c
index 32f3e9dda837..c2c609c39119 100644
--- a/mm/memtest.c
+++ b/mm/memtest.c
@@ -51,10 +51,10 @@ static void __init memtest(u64 pattern, phys_addr_t start_phys, phys_addr_t size
last_bad = 0;
for (p = start; p < end; p++)
- *p = pattern;
+ WRITE_ONCE(*p, pattern);
for (p = start; p < end; p++, start_phys_aligned += incr) {
- if (*p == pattern)
+ if (READ_ONCE(*p) == pattern)
continue;
if (start_phys_aligned == last_bad + incr) {
last_bad += incr;
--
2.39.2
On Tue, Mar 12, 2024 at 04:04:23PM +0800, Qiang Zhang wrote:
> memtest failed to find bad memory when compiled with clang. So use
> {WRITE,READ}_ONCE to access memory to avoid compiler over optimization.
This commit message is severely lacking in details in my opinion,
especially for a patch marked for stable. Did a kernel or LLVM change
cause this (i.e., has this always been an issue or is it a recent
regression)? What is the transformation that LLVM does to break the test
and why is using READ_ONCE() or WRITE_ONCE() sufficient to resolve it?
> Cc: <[email protected]>
> Signed-off-by: Qiang Zhang <[email protected]>
> ---
> mm/memtest.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/mm/memtest.c b/mm/memtest.c
> index 32f3e9dda837..c2c609c39119 100644
> --- a/mm/memtest.c
> +++ b/mm/memtest.c
> @@ -51,10 +51,10 @@ static void __init memtest(u64 pattern, phys_addr_t start_phys, phys_addr_t size
> last_bad = 0;
>
> for (p = start; p < end; p++)
> - *p = pattern;
> + WRITE_ONCE(*p, pattern);
>
> for (p = start; p < end; p++, start_phys_aligned += incr) {
> - if (*p == pattern)
> + if (READ_ONCE(*p) == pattern)
> continue;
> if (start_phys_aligned == last_bad + incr) {
> last_bad += incr;
> --
> 2.39.2
>
Hi, Nathan
Sorry for the incomplete commit message.
I have tried to compile with gcc and clang-{11,13,14} on Debian 12. On my test environment, hypervisor emulates a range of bad memory where writes are ignored and reads always returns all ones.
Memtest compiled with all clang-{11,13,14} can't find the bad memory without this patch. But gcc works fine. So it seems not a regression in clang.
I don't have expertise in compilers. But I think {READ,WRITE}_ONCE can force the compiler to treat the iterating pointer as volatile.
Welcome more comments !
BR
Qiang
-----Original Message-----
From: Nathan Chancellor <[email protected]>
Sent: Thursday, March 14, 2024 1:22 AM
To: Zhang, Qiang4 <[email protected]>
Cc: Andrew Morton <[email protected]>; Nick Desaulniers <[email protected]>; Bill Wendling <[email protected]>; Justin Stitt <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] memtest: use {READ,WRITE}_ONCE in memory scanning
On Tue, Mar 12, 2024 at 04:04:23PM +0800, Qiang Zhang wrote:
> memtest failed to find bad memory when compiled with clang. So use
> {WRITE,READ}_ONCE to access memory to avoid compiler over optimization.
This commit message is severely lacking in details in my opinion, especially for a patch marked for stable. Did a kernel or LLVM change cause this (i.e., has this always been an issue or is it a recent regression)? What is the transformation that LLVM does to break the test and why is using READ_ONCE() or WRITE_ONCE() sufficient to resolve it?
> Cc: <[email protected]>
> Signed-off-by: Qiang Zhang <[email protected]>
> ---
> mm/memtest.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/mm/memtest.c b/mm/memtest.c index
> 32f3e9dda837..c2c609c39119 100644
> --- a/mm/memtest.c
> +++ b/mm/memtest.c
> @@ -51,10 +51,10 @@ static void __init memtest(u64 pattern, phys_addr_t start_phys, phys_addr_t size
> last_bad = 0;
>
> for (p = start; p < end; p++)
> - *p = pattern;
> + WRITE_ONCE(*p, pattern);
>
> for (p = start; p < end; p++, start_phys_aligned += incr) {
> - if (*p == pattern)
> + if (READ_ONCE(*p) == pattern)
> continue;
> if (start_phys_aligned == last_bad + incr) {
> last_bad += incr;
> --
> 2.39.2
>