Skip to contentSkip to conversation

do_get_mempolicy()

bug ? John Blackwood 3 participants 5 messages Jul 3, 2008 – Jul 8, 2008

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hi Lee,

I'm having unexpected results with get_mempolicy(2) in 2.6.26, and I am hoping that you can either agree with me, or maybe comment on my misconceptions.

When I have a task with no special task mempolicy (the default mempolicy), when I call get_mempolicy(2), it returns a policy value of 2 (MPOL_BIND) with a NULL nodemask.

I believe that this is because of the code in do_get_mempolicy() that does:

*policy |= pol->flags;

in the else case when flags do not contain MPOL_F_NODE.

I guess I don't understand why we are ORing in the pol->flags into the *policy value. For example, when this is for the default_policy, the MPOL_F_LOCAL flag (which has a value of 2) gets stuffed into the *policy location, and a get_mempolicy(2) caller sees this as the MPOL_BIND mempolicy.

Maybe the "*policy |= pol->flags;" line should be removed ?

That is, maybe it was valid at some point, but subsequent changes make this line of code no longer valid ?

Sorry if I'm out-to-lunch here...

Thanks very much for you time and considerations on this issue.

#1 lore raw
Quoted 16 lines · John Blackwood

Hi Lee,

I'm having unexpected results with get_mempolicy(2) in 2.6.26, and I am hoping that you can either agree with me, or maybe comment on my misconceptions.

When I have a task with no special task mempolicy (the default mempolicy), when I call get_mempolicy(2), it returns a policy value of 2 (MPOL_BIND) with a NULL nodemask.

I believe that this is because of the code in do_get_mempolicy() that does:

*policy |= pol->flags;

in the else case when flags do not contain MPOL_F_NODE.

Quoted 8 lines

I guess I don't understand why we are ORing in the pol->flags into the *policy value. For example, when this is for the default_policy, the MPOL_F_LOCAL flag (which has a value of 2) gets stuffed into the *policy location, and a get_mempolicy(2) caller sees this as the MPOL_BIND mempolicy.

Maybe the "*policy |= pol->flags;" line should be removed ?

You're right, the flags member of struct mempolicy has subsequently changed to carry "internal" flags that are not supposed to be exposed to userspace via the get_mempolicy() API.

The following patch probably fixes it.

Lee?

---
 mm/mempolicy.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
mm/mempolicy.c +1 -1
@@ -729,7 +729,7 @@ static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 	} else {
 		*policy = pol == &default_policy ? MPOL_DEFAULT :
 						pol->mode;
-		*policy |= pol->flags;
+		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 	}
 
 	if (vma) {
#2 lore raw
Quoted 4 lines · David Rientjes

You're right, the flags member of struct mempolicy has subsequently changed to carry "internal" flags that are not supposed to be exposed to userspace via the get_mempolicy() API.

John, please give 2.6.26-rc9 a try and let me know if there's any outstanding issues.

Thanks.

#4 lore raw
In reply to John Blackwood
Quoted 1 line · John Blackwood

Hi Lee,

John: I'm just getting back from a long weekend and I'm wading through a big e-mail back log. I'll take a look at this when I get to the end of my inbox. I/m probably responsible for that line. I recall thinking that get_mempolicy() should return the policy that one would pass back in to achieve the same effect. But, I blew it.

Quoted 9 lines

I'm having unexpected results with get_mempolicy(2) in 2.6.26, and I am hoping that you can either agree with me, or maybe comment on my misconceptions.

When I have a task with no special task mempolicy (the default mempolicy), when I call get_mempolicy(2), it returns a policy value of 2 (MPOL_BIND) with a NULL nodemask.

Quoted 5 lines

I believe that this is because of the code in do_get_mempolicy() that does:

  *policy |= pol->flags;

in the else case when flags do not contain MPOL_F_NODE.

I think that need to mask off the internal flags, and shift the remaining ones up to the correct location. I'll send a patch, if no one beats me to it.

Quoted 7 lines

I guess I don't understand why we are ORing in the pol->flags into the *policy value. For example, when this is for the default_policy, the MPOL_F_LOCAL flag (which has a value of 2) gets stuffed into the *policy location, and a get_mempolicy(2) caller sees this as the MPOL_BIND mempolicy.

Quoted 6 lines

Maybe the "*policy |= pol->flags;" line should be removed ?

That is, maybe it was valid at some point, but subsequent changes make this line of code no longer valid ?

Sorry if I'm out-to-lunch here...

No, doesn't appear that way..

Quoted 3 lines

Thanks very much for you time and considerations on this issue.

thanks for reporting it.

Lee

#3 lore raw
In reply to David Rientjes
Quoted 27 lines · David Rientjes
Quoted 16 lines · John Blackwood

Hi Lee,

I'm having unexpected results with get_mempolicy(2) in 2.6.26, and I am hoping that you can either agree with me, or maybe comment on my misconceptions.

When I have a task with no special task mempolicy (the default mempolicy), when I call get_mempolicy(2), it returns a policy value of 2 (MPOL_BIND) with a NULL nodemask.

I believe that this is because of the code in do_get_mempolicy() that does:

*policy |= pol->flags;

in the else case when flags do not contain MPOL_F_NODE.

Quoted 8 lines

I guess I don't understand why we are ORing in the pol->flags into the *policy value. For example, when this is for the default_policy, the MPOL_F_LOCAL flag (which has a value of 2) gets stuffed into the *policy location, and a get_mempolicy(2) caller sees this as the MPOL_BIND mempolicy.

Maybe the "*policy |= pol->flags;" line should be removed ?

Quoted 7 lines

You're right, the flags member of struct mempolicy has subsequently changed to carry "internal" flags that are not supposed to be exposed to userspace via the get_mempolicy() API.

The following patch probably fixes it.

Lee?

David: Just getting back from long weekend. This looks good. I was thinking that, in addition to masking off the internal flags, we need to shift the flags into the upper half word. However, external flags are already shifted to the correct position, so just the mask is needed.

Quoted 2 lines
Quoted 16 lines
---
 mm/mempolicy.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
mm/mempolicy.c +1 -1
@@ -729,7 +729,7 @@ static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 	} else {
 		*policy = pol == &default_policy ? MPOL_DEFAULT :
 						pol->mode;
-		*policy |= pol->flags;
+		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 	}
 
 	if (vma) {