[PATCH 1/2] arm64: Improve error reporting on set_pte_at() checks

[PATCH 1/2] arm64: Improve error reporting on set_pte_at() checks

[James Morse]

Hi Andrew,

On 16/12/15 03:01, Andrew Pinski wrote:
> On Tue, Dec 9, 2015 at 17:26:56, Catalin Marinas
> <catalin.marinas@arm.com> wrote:
>>
>> Currently the BUG_ON() checks do not give enough information about the PTEs being set. This patch changes BUG_ON to WARN_ONCE and dumps the values of the old and new PTEs.
> 
> 
> This change broke building the mantis driver:
> 
> In file included from ./arch/arm64/include/asm/io.h:30:0,
>                  from drivers/media/pci/mantis/mantis_i2c.c:21:
> ./arch/arm64/include/asm/pgtable.h: In function ?set_pte_at?:
> ./arch/arm64/include/asm/pgtable.h:281:3: error: implicit declaration
> of function ?BUILD_BUG_ON_INVALID?
> [-Werror=implicit-function-declaration]
>    VM_WARN_ONCE(!pte_young(pte),
>    ^

This is due to a missing include in include/linux/mmdebug.h, which
Julien Grall saw with CONFIG_XEN & !CONFIG_DEBUG_VM. There was some
discussion at [0], and a patch at [1].

The quick-and-dirty workaround is to toggle CONFIG_DEBUG_VM. It should
be fixed in the next rc.


Thanks,

James

[0] https://lkml.org/lkml/2015/12/14/489
[1] http://www.spinics.net/lists/linux-mm/msg98868.html

[PATCH 0/2] arm64: Non-racy PTE setting in the presence of HW AF/DBM

[Catalin Marinas]

This series addresses a potentially racy default implementation of
ptep_set_access_flags() when hardware update of the access or dirty
states is enabled. The first patch is some clean-up in set_pte_at() to
improve the information reporting and replace BUG with WARN. The second
patch contains the arm64-specific ptep_set_access_flags()
implementation.

Possible racy scenarios are described in patch 2. I think this series
could be simplified on the following assumptions:

a) if the CPUs do not support HW AF/DBM or it is disabled, no other
   agent in the system will perform such updates

b) if one CPU supports HW AF/DBM, all of them must do (don't mix such
   features)

Point (a) means that the current code works fine and BUG_ON() is not
necessary.

Point (b) however requires a ptep_set_access_flags() similar to the x86
one, i.e. only do the setting if (changed && dirty), otherwise let the
hardware handle the updates.

Anyway, while patch 2 is still debatable, I'd like to merge the first
patch in 4.4 to avoid an unnecessary BUG_ON on hardware that doesn't
even do DBM.

Catalin


Catalin Marinas (2):
  arm64: Improve error reporting on set_pte_at() checks
  arm64: Implement ptep_set_access_flags() for hardware AF/DBM

 arch/arm64/include/asm/pgtable.h | 16 ++++++++---
 arch/arm64/mm/fault.c            | 57 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 69 insertions(+), 4 deletions(-)

[PATCH 1/2] arm64: Improve error reporting on set_pte_at() checks

[Catalin Marinas]

Currently the BUG_ON() checks do not give enough information about the
PTEs being set. This patch changes BUG_ON to WARN_ONCE and dumps the
values of the old and new PTEs.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
---
 arch/arm64/include/asm/pgtable.h | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 7e074f93f383..002dc61a4dff 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -276,10 +276,13 @@ static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
 	 * hardware updates of the pte (ptep_set_access_flags safely changes
 	 * valid ptes without going through an invalid entry).
 	 */
-	if (IS_ENABLED(CONFIG_DEBUG_VM) && IS_ENABLED(CONFIG_ARM64_HW_AFDBM) &&
-	    pte_valid(*ptep)) {
-		BUG_ON(!pte_young(pte));
-		BUG_ON(pte_write(*ptep) && !pte_dirty(pte));
+	if (IS_ENABLED(CONFIG_ARM64_HW_AFDBM) && pte_valid(*ptep)) {
+		VM_WARN_ONCE(!pte_young(pte),
+			     "%s: racy access flag clearing: %016llx -> %016llx",
+			     __func__, pte_val(*ptep), pte_val(pte));
+		VM_WARN_ONCE(pte_write(*ptep) && !pte_dirty(pte),
+			     "%s: racy dirty state clearing: %016llx -> %016llx",
+			     __func__, pte_val(*ptep), pte_val(pte));
 	}
 
 	set_pte(ptep, pte);