1. Patchew
  2. linux
  3. View series

[PATCH] vdso: change PAGE_MASK to signed on all 32-bit architectures

Arnd Bergmann posted 1 patch 1 month ago 
include/vdso/page.h | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
[PATCH] vdso: change PAGE_MASK to signed on all 32-bit architectures
Posted by Arnd Bergmann 1 month ago 
From: Arnd Bergmann <arnd@arndb.de>

With the introduction of an architecture-independent defintion of
PAGE_MASK, we had to make a choice between defining it as 'unsigned
long' as on 64-bit architectures, or as signed 'long' as required for
architectures with a 64-bit phys_addr_t.

To reduce the risk for regressions and minimize the changes in behavior,
the result was using the signed value only when CONFIG_PHYS_ADDR_T_64BIT
is set, but that ended up causing a regression after all in the
early_init_dt_add_memory_arch() function that uses 64-bit integers for
address calculation.

Presumably the same regression also affects mips32 and powerpc32 when
dealing with large amounts of memory on DT platforms: like arm32, they
were using the signed version unconditionally.

The two most sensible options that I see for addressing the regiression
are either to go back to an architecture specific definition, using a
signed constant on arm/powerpc/mips and unsigned on the others, or to
use the same definition everywhere.

Use the simpler of those two and change them all to the signed version,
in the hope that this does not cause a different type of bug. Most
of the other 32-bit architectures have no large physical address
support and are rarely used, so it seems more likely that using the
same definition helps than hurts here.

In particular, x86-32 does have physical addressing extensions, so it
already changed to the signed version after the previous patch,
so it makes sense to use the same version on non-PAE as well.

Fixes: efe8419ae78d ("vdso: Introduce vdso/page.h")
Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
Tested-by: Anders Roxell <anders.roxell@linaro.org>
Cc: Alex Bennee <alex.bennee@linaro.org>,
Link: https://lore.kernel.org/lkml/CA+G9fYt86bUAu_v5dXPWnDUwQNVipj+Wq3Djir1KUSKdr9QLNg@mail.gmail.com/
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 include/vdso/page.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/vdso/page.h b/include/vdso/page.h
index 4ada1ba6bd1f..710ae2414e68 100644
--- a/include/vdso/page.h
+++ b/include/vdso/page.h
@@ -14,13 +14,14 @@
 
 #define PAGE_SIZE	(_AC(1,UL) << CONFIG_PAGE_SHIFT)
 
-#if defined(CONFIG_PHYS_ADDR_T_64BIT) && !defined(CONFIG_64BIT)
+#if !defined(CONFIG_64BIT)
 /*
- * Applies only to 32-bit architectures with a 64-bit phys_addr_t.
+ * Applies only to 32-bit architectures.
  *
  * Subtle: (1 << CONFIG_PAGE_SHIFT) is an int, not an unsigned long.
  * So if we assign PAGE_MASK to a larger type it gets extended the
- * way we want (i.e. with 1s in the high bits)
+ * way we want (i.e. with 1s in the high bits) while masking a
+ * 64-bit value such as phys_addr_t.
  */
 #define PAGE_MASK	(~((1 << CONFIG_PAGE_SHIFT) - 1))
 #else
-- 
2.39.5
Re: [PATCH] vdso: change PAGE_MASK to signed on all 32-bit architectures
Posted by Vincenzo Frascino 4 weeks ago 

On 24/10/2024 14:34, Arnd Bergmann wrote:
> From: Arnd Bergmann <arnd@arndb.de>
> 
> With the introduction of an architecture-independent defintion of
> PAGE_MASK, we had to make a choice between defining it as 'unsigned
> long' as on 64-bit architectures, or as signed 'long' as required for
> architectures with a 64-bit phys_addr_t.
> 
> To reduce the risk for regressions and minimize the changes in behavior,
> the result was using the signed value only when CONFIG_PHYS_ADDR_T_64BIT
> is set, but that ended up causing a regression after all in the
> early_init_dt_add_memory_arch() function that uses 64-bit integers for
> address calculation.
> 
> Presumably the same regression also affects mips32 and powerpc32 when
> dealing with large amounts of memory on DT platforms: like arm32, they
> were using the signed version unconditionally.
> 
> The two most sensible options that I see for addressing the regiression
> are either to go back to an architecture specific definition, using a
> signed constant on arm/powerpc/mips and unsigned on the others, or to
> use the same definition everywhere.
> 
> Use the simpler of those two and change them all to the signed version,
> in the hope that this does not cause a different type of bug. Most
> of the other 32-bit architectures have no large physical address
> support and are rarely used, so it seems more likely that using the
> same definition helps than hurts here.
> 
> In particular, x86-32 does have physical addressing extensions, so it
> already changed to the signed version after the previous patch,
> so it makes sense to use the same version on non-PAE as well.
> 
> Fixes: efe8419ae78d ("vdso: Introduce vdso/page.h")
> Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
> Tested-by: Anders Roxell <anders.roxell@linaro.org>
> Cc: Alex Bennee <alex.bennee@linaro.org>,
> Link: https://lore.kernel.org/lkml/CA+G9fYt86bUAu_v5dXPWnDUwQNVipj+Wq3Djir1KUSKdr9QLNg@mail.gmail.com/

Thanks Arnd, I tested it this morning and seems working correctly with qemu 9.1.1.

With this:

Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com>
Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com>

> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> ---
>  include/vdso/page.h | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/include/vdso/page.h b/include/vdso/page.h
> index 4ada1ba6bd1f..710ae2414e68 100644
> --- a/include/vdso/page.h
> +++ b/include/vdso/page.h
> @@ -14,13 +14,14 @@
>  
>  #define PAGE_SIZE	(_AC(1,UL) << CONFIG_PAGE_SHIFT)
>  
> -#if defined(CONFIG_PHYS_ADDR_T_64BIT) && !defined(CONFIG_64BIT)
> +#if !defined(CONFIG_64BIT)
>  /*
> - * Applies only to 32-bit architectures with a 64-bit phys_addr_t.
> + * Applies only to 32-bit architectures.
>   *
>   * Subtle: (1 << CONFIG_PAGE_SHIFT) is an int, not an unsigned long.
>   * So if we assign PAGE_MASK to a larger type it gets extended the
> - * way we want (i.e. with 1s in the high bits)
> + * way we want (i.e. with 1s in the high bits) while masking a
> + * 64-bit value such as phys_addr_t.
>   */
>  #define PAGE_MASK	(~((1 << CONFIG_PAGE_SHIFT) - 1))
>  #else

-- 
Regards,
Vincenzo
[tip: timers/vdso] vdso: Change PAGE_MASK to signed on all 32-bit architectures
Posted by tip-bot2 for Arnd Bergmann 4 weeks ago 
The following commit has been merged into the timers/vdso branch of tip:

Commit-ID:     d4a65302dd849fade9e2ca712826c35b8d068ecb
Gitweb:        https://git.kernel.org/tip/d4a65302dd849fade9e2ca712826c35b8d068ecb
Author:        Arnd Bergmann <arnd@arndb.de>
AuthorDate:    Thu, 24 Oct 2024 13:34:26 
Committer:     Thomas Gleixner <tglx@linutronix.de>
CommitterDate: Mon, 28 Oct 2024 16:56:15 +01:00

vdso: Change PAGE_MASK to signed on all 32-bit architectures

With the introduction of an architecture-independent defintion of
PAGE_MASK, we had to make a choice between defining it as 'unsigned long'
as on 64-bit architectures, or as signed 'long' as required for
architectures with a 64-bit phys_addr_t.

To reduce the risk for regressions and minimize the changes in behavior,
the result was using the signed value only when CONFIG_PHYS_ADDR_T_64BIT
is set, but that ended up causing a regression after all in the
early_init_dt_add_memory_arch() function that uses 64-bit integers for
address calculation.

Presumably the same regression also affects mips32 and powerpc32 when
dealing with large amounts of memory on DT platforms: like arm32, they were
using the signed version unconditionally.

The two most sensible options for addressing the regression are either to
go back to an architecture specific definition, using a signed constant on
arm/powerpc/mips and unsigned on the others, or to use the same definition
everywhere.

Use the simpler of those two and change them all to the signed version, in
the hope that this does not cause a different type of bug. Most of the
other 32-bit architectures have no large physical address support and are
rarely used, so it seems more likely that using the same definition helps
than hurts here.

In particular, x86-32 does have physical addressing extensions, so it
already changed to the signed version after the previous patch, so it makes
sense to use the same version on non-PAE as well.

Fixes: efe8419ae78d ("vdso: Introduce vdso/page.h")
Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Anders Roxell <anders.roxell@linaro.org>
Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com>
Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com>
Link: https://lore.kernel.org/all/20241024133447.3117273-1-arnd@kernel.org
Link: https://lore.kernel.org/lkml/CA+G9fYt86bUAu_v5dXPWnDUwQNVipj+Wq3Djir1KUSKdr9QLNg@mail.gmail.com/

---
 include/vdso/page.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/vdso/page.h b/include/vdso/page.h
index 4ada1ba..710ae24 100644
--- a/include/vdso/page.h
+++ b/include/vdso/page.h
@@ -14,13 +14,14 @@
 
 #define PAGE_SIZE	(_AC(1,UL) << CONFIG_PAGE_SHIFT)
 
-#if defined(CONFIG_PHYS_ADDR_T_64BIT) && !defined(CONFIG_64BIT)
+#if !defined(CONFIG_64BIT)
 /*
- * Applies only to 32-bit architectures with a 64-bit phys_addr_t.
+ * Applies only to 32-bit architectures.
  *
  * Subtle: (1 << CONFIG_PAGE_SHIFT) is an int, not an unsigned long.
  * So if we assign PAGE_MASK to a larger type it gets extended the
- * way we want (i.e. with 1s in the high bits)
+ * way we want (i.e. with 1s in the high bits) while masking a
+ * 64-bit value such as phys_addr_t.
  */
 #define PAGE_MASK	(~((1 << CONFIG_PAGE_SHIFT) - 1))
 #else

Patchew 2.1-devel-b67e4c2

© 2016 - 2024 Red Hat, Inc.