John Paul reported a warning about bogus NUMA distance values spurred by
commit:
620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
In this case, the afflicted machine comes up with a reported 256 possible
nodes, all of which are 0 distance away from one another. This was
previously silently ignored, but is now caught by the aforementioned
commit.
The culprit is ia64's node_possible_map which remains unchanged from its
initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
have any SRAT nor SLIT table, but AIUI the possible map remains untouched
regardless of what ACPI tables end up being parsed. Thus, !online &&
possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
"reserved and have no meaning" as per the ACPI spec).
Follow x86 / drivers/base/arch_numa's example and set the possible map to
the parsed map, which in this case seems to be the online map.
Link: http://lore.kernel.org/r/[email protected]
Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
Reported-by: John Paul Adrian Glaubitz <[email protected]>
Signed-off-by: Valentin Schneider <[email protected]>
---
This might need an earlier Fixes: tag, but all of this is quite old and
dusty (the git blame rabbit hole leads me to ~2008/2007)
Alternatively, can we deprecate ia64 already?
---
arch/ia64/kernel/acpi.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/arch/ia64/kernel/acpi.c b/arch/ia64/kernel/acpi.c
index a5636524af76..e2af6b172200 100644
--- a/arch/ia64/kernel/acpi.c
+++ b/arch/ia64/kernel/acpi.c
@@ -446,7 +446,8 @@ void __init acpi_numa_fixup(void)
if (srat_num_cpus == 0) {
node_set_online(0);
node_cpuid[0].phys_id = hard_smp_processor_id();
- return;
+ slit_distance(0, 0) = LOCAL_DISTANCE;
+ goto out;
}
/*
@@ -489,7 +490,7 @@ void __init acpi_numa_fixup(void)
for (j = 0; j < MAX_NUMNODES; j++)
slit_distance(i, j) = i == j ?
LOCAL_DISTANCE : REMOTE_DISTANCE;
- return;
+ goto out;
}
memset(numa_slit, -1, sizeof(numa_slit));
@@ -514,6 +515,8 @@ void __init acpi_numa_fixup(void)
printk("\n");
}
#endif
+out:
+ node_possible_map = node_online_map;
}
#endif /* CONFIG_ACPI_NUMA */
--
2.25.1
Hi Valentin!
On 3/18/21 2:06 PM, Valentin Schneider wrote:
> John Paul reported a warning about bogus NUMA distance values spurred by
> commit:
>
> 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>
> In this case, the afflicted machine comes up with a reported 256 possible
> nodes, all of which are 0 distance away from one another. This was
> previously silently ignored, but is now caught by the aforementioned
> commit.
>
> The culprit is ia64's node_possible_map which remains unchanged from its
> initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
> have any SRAT nor SLIT table, but AIUI the possible map remains untouched
> regardless of what ACPI tables end up being parsed. Thus, !online &&
> possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
> "reserved and have no meaning" as per the ACPI spec).
>
> Follow x86 / drivers/base/arch_numa's example and set the possible map to
> the parsed map, which in this case seems to be the online map.
>
> Link: http://lore.kernel.org/r/[email protected]
> Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
> Reported-by: John Paul Adrian Glaubitz <[email protected]>
> Signed-off-by: Valentin Schneider <[email protected]>
> ---
> This might need an earlier Fixes: tag, but all of this is quite old and
> dusty (the git blame rabbit hole leads me to ~2008/2007)
>
> Alternatively, can we deprecate ia64 already?
> ---
> arch/ia64/kernel/acpi.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/arch/ia64/kernel/acpi.c b/arch/ia64/kernel/acpi.c
> index a5636524af76..e2af6b172200 100644
> --- a/arch/ia64/kernel/acpi.c
> +++ b/arch/ia64/kernel/acpi.c
> @@ -446,7 +446,8 @@ void __init acpi_numa_fixup(void)
> if (srat_num_cpus == 0) {
> node_set_online(0);
> node_cpuid[0].phys_id = hard_smp_processor_id();
> - return;
> + slit_distance(0, 0) = LOCAL_DISTANCE;
> + goto out;
> }
>
> /*
> @@ -489,7 +490,7 @@ void __init acpi_numa_fixup(void)
> for (j = 0; j < MAX_NUMNODES; j++)
> slit_distance(i, j) = i == j ?
> LOCAL_DISTANCE : REMOTE_DISTANCE;
> - return;
> + goto out;
> }
>
> memset(numa_slit, -1, sizeof(numa_slit));
> @@ -514,6 +515,8 @@ void __init acpi_numa_fixup(void)
> printk("\n");
> }
> #endif
> +out:
> + node_possible_map = node_online_map;
> }
> #endif /* CONFIG_ACPI_NUMA */
>
>
Tested-by: John Paul Adrian Glaubitz <[email protected]>
Could you send this patch through Andrew Morton's tree? The ia64 port currently
has no maintainer, so we have to use an alternative tree.
@Sergei: Could you test/ack this patch as well?
Thanks,
Adrian
--
.''`. John Paul Adrian Glaubitz
: :' : Debian Developer - [email protected]
`. `' Freie Universitaet Berlin - [email protected]
`- GPG: 62FF 8A75 84E0 2956 9546 0006 7426 3B37 F5B5 F913
On Fri, 19 Mar 2021 15:47:09 +0100
John Paul Adrian Glaubitz <[email protected]> wrote:
> Hi Valentin!
>
> On 3/18/21 2:06 PM, Valentin Schneider wrote:
> > John Paul reported a warning about bogus NUMA distance values spurred by
> > commit:
> >
> > 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
> >
> > In this case, the afflicted machine comes up with a reported 256 possible
> > nodes, all of which are 0 distance away from one another. This was
> > previously silently ignored, but is now caught by the aforementioned
> > commit.
> >
> > The culprit is ia64's node_possible_map which remains unchanged from its
> > initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
> > have any SRAT nor SLIT table, but AIUI the possible map remains untouched
> > regardless of what ACPI tables end up being parsed. Thus, !online &&
> > possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
> > "reserved and have no meaning" as per the ACPI spec).
> >
> > Follow x86 / drivers/base/arch_numa's example and set the possible map to
> > the parsed map, which in this case seems to be the online map.
> >
> > Link: http://lore.kernel.org/r/[email protected]
> > Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
> > Reported-by: John Paul Adrian Glaubitz <[email protected]>
> > Signed-off-by: Valentin Schneider <[email protected]>
> > ---
> > This might need an earlier Fixes: tag, but all of this is quite old and
> > dusty (the git blame rabbit hole leads me to ~2008/2007)
> >
> > Alternatively, can we deprecate ia64 already?
> > ---
> > arch/ia64/kernel/acpi.c | 7 +++++--
> > 1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/ia64/kernel/acpi.c b/arch/ia64/kernel/acpi.c
> > index a5636524af76..e2af6b172200 100644
> > --- a/arch/ia64/kernel/acpi.c
> > +++ b/arch/ia64/kernel/acpi.c
> > @@ -446,7 +446,8 @@ void __init acpi_numa_fixup(void)
> > if (srat_num_cpus == 0) {
> > node_set_online(0);
> > node_cpuid[0].phys_id = hard_smp_processor_id();
> > - return;
> > + slit_distance(0, 0) = LOCAL_DISTANCE;
> > + goto out;
> > }
> >
> > /*
> > @@ -489,7 +490,7 @@ void __init acpi_numa_fixup(void)
> > for (j = 0; j < MAX_NUMNODES; j++)
> > slit_distance(i, j) = i == j ?
> > LOCAL_DISTANCE : REMOTE_DISTANCE;
> > - return;
> > + goto out;
> > }
> >
> > memset(numa_slit, -1, sizeof(numa_slit));
> > @@ -514,6 +515,8 @@ void __init acpi_numa_fixup(void)
> > printk("\n");
> > }
> > #endif
> > +out:
> > + node_possible_map = node_online_map;
> > }
> > #endif /* CONFIG_ACPI_NUMA */
> >
> >
>
> Tested-by: John Paul Adrian Glaubitz <[email protected]>
>
> Could you send this patch through Andrew Morton's tree? The ia64 port currently
> has no maintainer, so we have to use an alternative tree.
>
> @Sergei: Could you test/ack this patch as well?
Booted successfully without problems on rx3600.
Tested-by: Sergei Trofimovich <[email protected]>
--
Sergei
On 3/19/21 8:10 PM, Sergei Trofimovich wrote:
> On Fri, 19 Mar 2021 15:47:09 +0100
> John Paul Adrian Glaubitz <[email protected]> wrote:
>
>> Hi Valentin!
>>
>> On 3/18/21 2:06 PM, Valentin Schneider wrote:
>>> John Paul reported a warning about bogus NUMA distance values spurred by
>>> commit:
>>>
>>> 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>>>
>>> In this case, the afflicted machine comes up with a reported 256 possible
>>> nodes, all of which are 0 distance away from one another. This was
>>> previously silently ignored, but is now caught by the aforementioned
>>> commit.
>>>
>>> The culprit is ia64's node_possible_map which remains unchanged from its
>>> initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
>>> have any SRAT nor SLIT table, but AIUI the possible map remains untouched
>>> regardless of what ACPI tables end up being parsed. Thus, !online &&
>>> possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
>>> "reserved and have no meaning" as per the ACPI spec).
>>>
>>> Follow x86 / drivers/base/arch_numa's example and set the possible map to
>>> the parsed map, which in this case seems to be the online map.
>>>
>>> Link: http://lore.kernel.org/r/[email protected]
>>> Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>>> Reported-by: John Paul Adrian Glaubitz <[email protected]>
>>> Signed-off-by: Valentin Schneider <[email protected]>
>>> ---
>>> This might need an earlier Fixes: tag, but all of this is quite old and
>>> dusty (the git blame rabbit hole leads me to ~2008/2007)
>>>
>>> Alternatively, can we deprecate ia64 already?
>>> ---
>>> arch/ia64/kernel/acpi.c | 7 +++++--
>>> 1 file changed, 5 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/arch/ia64/kernel/acpi.c b/arch/ia64/kernel/acpi.c
>>> index a5636524af76..e2af6b172200 100644
>>> --- a/arch/ia64/kernel/acpi.c
>>> +++ b/arch/ia64/kernel/acpi.c
>>> @@ -446,7 +446,8 @@ void __init acpi_numa_fixup(void)
>>> if (srat_num_cpus == 0) {
>>> node_set_online(0);
>>> node_cpuid[0].phys_id = hard_smp_processor_id();
>>> - return;
>>> + slit_distance(0, 0) = LOCAL_DISTANCE;
>>> + goto out;
>>> }
>>>
>>> /*
>>> @@ -489,7 +490,7 @@ void __init acpi_numa_fixup(void)
>>> for (j = 0; j < MAX_NUMNODES; j++)
>>> slit_distance(i, j) = i == j ?
>>> LOCAL_DISTANCE : REMOTE_DISTANCE;
>>> - return;
>>> + goto out;
>>> }
>>>
>>> memset(numa_slit, -1, sizeof(numa_slit));
>>> @@ -514,6 +515,8 @@ void __init acpi_numa_fixup(void)
>>> printk("\n");
>>> }
>>> #endif
>>> +out:
>>> + node_possible_map = node_online_map;
>>> }
>>> #endif /* CONFIG_ACPI_NUMA */
>>>
>>>
>>
>> Tested-by: John Paul Adrian Glaubitz <[email protected]>
>>
>> Could you send this patch through Andrew Morton's tree? The ia64 port currently
>> has no maintainer, so we have to use an alternative tree.
>>
>> @Sergei: Could you test/ack this patch as well?
>
> Booted successfully without problems on rx3600.
>
> Tested-by: Sergei Trofimovich <[email protected]>
Great, thanks!
@Andrew: Could you pick up this patch through your tree?
Adrian
--
.''`. John Paul Adrian Glaubitz
: :' : Debian Developer - [email protected]
`. `' Freie Universitaet Berlin - [email protected]
`- GPG: 62FF 8A75 84E0 2956 9546 0006 7426 3B37 F5B5 F913
Hi!
On 3/24/21 7:54 PM, Andrew Morton wrote:
> On Thu, 18 Mar 2021 13:06:17 +0000 Valentin Schneider <[email protected]> wrote:
>
>> John Paul reported a warning about bogus NUMA distance values spurred by
>> commit:
>>
>> 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>>
>> In this case, the afflicted machine comes up with a reported 256 possible
>> nodes, all of which are 0 distance away from one another. This was
>> previously silently ignored, but is now caught by the aforementioned
>> commit.
>>
>> The culprit is ia64's node_possible_map which remains unchanged from its
>> initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
>> have any SRAT nor SLIT table, but AIUI the possible map remains untouched
>> regardless of what ACPI tables end up being parsed. Thus, !online &&
>> possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
>> "reserved and have no meaning" as per the ACPI spec).
>>
>> Follow x86 / drivers/base/arch_numa's example and set the possible map to
>> the parsed map, which in this case seems to be the online map.
>>
>> Link: http://lore.kernel.org/r/[email protected]
>> Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>> Reported-by: John Paul Adrian Glaubitz <[email protected]>
>> Signed-off-by: Valentin Schneider <[email protected]>
>> ---
>> This might need an earlier Fixes: tag, but all of this is quite old and
>> dusty (the git blame rabbit hole leads me to ~2008/2007)
>>
>
> Thanks. Is this worth a cc:stable tag?
Looks like the regression was introduced 5.12-rc1, so no need for backporting.
Adrian
--
.''`. John Paul Adrian Glaubitz
: :' : Debian Developer - [email protected]
`. `' Freie Universitaet Berlin - [email protected]
`- GPG: 62FF 8A75 84E0 2956 9546 0006 7426 3B37 F5B5 F913
On Thu, 18 Mar 2021 13:06:17 +0000 Valentin Schneider <[email protected]> wrote:
> John Paul reported a warning about bogus NUMA distance values spurred by
> commit:
>
> 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
>
> In this case, the afflicted machine comes up with a reported 256 possible
> nodes, all of which are 0 distance away from one another. This was
> previously silently ignored, but is now caught by the aforementioned
> commit.
>
> The culprit is ia64's node_possible_map which remains unchanged from its
> initialization value of NODE_MASK_ALL. In John's case, the machine doesn't
> have any SRAT nor SLIT table, but AIUI the possible map remains untouched
> regardless of what ACPI tables end up being parsed. Thus, !online &&
> possible nodes remain with a bogus distance of 0 (distances \in [0, 9] are
> "reserved and have no meaning" as per the ACPI spec).
>
> Follow x86 / drivers/base/arch_numa's example and set the possible map to
> the parsed map, which in this case seems to be the online map.
>
> Link: http://lore.kernel.org/r/[email protected]
> Fixes: 620a6dc40754 ("sched/topology: Make sched_init_numa() use a set for the deduplicating sort")
> Reported-by: John Paul Adrian Glaubitz <[email protected]>
> Signed-off-by: Valentin Schneider <[email protected]>
> ---
> This might need an earlier Fixes: tag, but all of this is quite old and
> dusty (the git blame rabbit hole leads me to ~2008/2007)
>
Thanks. Is this worth a cc:stable tag?