2018-07-30 23:03:50

by Mike Travis

[permalink] [raw]
Subject: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large Reference Count warnings

With the Intel BIOS support for 8 processor sockets with a full complement
of NVDIMMS potentially installable, and there are empty sockets without
NVDIMMS, there is an extremely large amount of the following warnings:

ACPI Warning: Large Reference Count (0x1001) in object ffff99453fc71750,
Type=0x0A

On a 4 socket system with 4 NVDIMMs there were over 6000 of these warning
messages and it has been seem on systems from 4 to 32 sockets.

Through some guidance from the BIOS developers and testing, it appears
that simply bumping up the threshold for warnings from 0x1000 to 0x2000
eliminates these messages. Changing them to be ACPI DEBUG messages, or
even removing them are other options, but it would then defeat the purpose
of the warnings as the problem would effectively be hidden.

Reviewed-by: Dimitri Sivanich <[email protected]>
Tested-by: Russ Anderson <[email protected]>
Signed-off-by: Mike Travis <[email protected]>
---
include/acpi/acconfig.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

--- linux-4.12.orig/include/acpi/acconfig.h
+++ linux-4.12/include/acpi/acconfig.h
@@ -123,7 +123,7 @@

/* Maximum object reference count (detects object deletion issues) */

-#define ACPI_MAX_REFERENCE_COUNT 0x1000
+#define ACPI_MAX_REFERENCE_COUNT 0x2000

/* Default page size for use in mapping memory for operation regions */


--



2018-08-01 19:50:25

by Moore, Robert

[permalink] [raw]
Subject: RE: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large Reference Count warnings

Are there any actual memory leaks, or this a consequence of the number of sockets?


> -----Original Message-----
> From: Mike Travis [mailto:[email protected]]
> Sent: Monday, July 30, 2018 12:56 PM
> To: Moore, Robert <[email protected]>; Schmauss, Erik
> <[email protected]>; Wysocki, Rafael J
> <[email protected]>; Len Brown <[email protected]>
> Cc: Russ Anderson <[email protected]>; Dimitri Sivanich
> <[email protected]>; Ross Zwisler <[email protected]>;
> Williams, Dan J <[email protected]>; Verma, Vishal L
> <[email protected]>; Jiang, Dave <[email protected]>; linux-
> [email protected]; [email protected]; [email protected];
> [email protected]
> Subject: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large
> Reference Count warnings
>
> With the Intel BIOS support for 8 processor sockets with a full
> complement of NVDIMMS potentially installable, and there are empty
> sockets without NVDIMMS, there is an extremely large amount of the
> following warnings:
>
> ACPI Warning: Large Reference Count (0x1001) in object ffff99453fc71750,
> Type=0x0A
>
> On a 4 socket system with 4 NVDIMMs there were over 6000 of these
> warning messages and it has been seem on systems from 4 to 32 sockets.
>
> Through some guidance from the BIOS developers and testing, it appears
> that simply bumping up the threshold for warnings from 0x1000 to 0x2000
> eliminates these messages. Changing them to be ACPI DEBUG messages, or
> even removing them are other options, but it would then defeat the
> purpose of the warnings as the problem would effectively be hidden.
>
> Reviewed-by: Dimitri Sivanich <[email protected]>
> Tested-by: Russ Anderson <[email protected]>
> Signed-off-by: Mike Travis <[email protected]>
> ---
> include/acpi/acconfig.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> --- linux-4.12.orig/include/acpi/acconfig.h
> +++ linux-4.12/include/acpi/acconfig.h
> @@ -123,7 +123,7 @@
>
> /* Maximum object reference count (detects object deletion issues) */
>
> -#define ACPI_MAX_REFERENCE_COUNT 0x1000
> +#define ACPI_MAX_REFERENCE_COUNT 0x2000
>
> /* Default page size for use in mapping memory for operation regions */
>
>
> --


2018-08-01 20:39:36

by Russ Anderson

[permalink] [raw]
Subject: Re: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large Reference Count warnings

On Wed, Aug 01, 2018 at 07:49:14PM +0000, Moore, Robert wrote:
> Are there any actual memory leaks, or this a consequence of the number of sockets?

It is a consequence of the number of sockets.
It is not an actual memory leak.

Thanks.

> > -----Original Message-----
> > From: Mike Travis [mailto:[email protected]]
> > Sent: Monday, July 30, 2018 12:56 PM
> > To: Moore, Robert <[email protected]>; Schmauss, Erik
> > <[email protected]>; Wysocki, Rafael J
> > <[email protected]>; Len Brown <[email protected]>
> > Cc: Russ Anderson <[email protected]>; Dimitri Sivanich
> > <[email protected]>; Ross Zwisler <[email protected]>;
> > Williams, Dan J <[email protected]>; Verma, Vishal L
> > <[email protected]>; Jiang, Dave <[email protected]>; linux-
> > [email protected]; [email protected]; [email protected];
> > [email protected]
> > Subject: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large
> > Reference Count warnings
> >
> > With the Intel BIOS support for 8 processor sockets with a full
> > complement of NVDIMMS potentially installable, and there are empty
> > sockets without NVDIMMS, there is an extremely large amount of the
> > following warnings:
> >
> > ACPI Warning: Large Reference Count (0x1001) in object ffff99453fc71750,
> > Type=0x0A
> >
> > On a 4 socket system with 4 NVDIMMs there were over 6000 of these
> > warning messages and it has been seem on systems from 4 to 32 sockets.
> >
> > Through some guidance from the BIOS developers and testing, it appears
> > that simply bumping up the threshold for warnings from 0x1000 to 0x2000
> > eliminates these messages. Changing them to be ACPI DEBUG messages, or
> > even removing them are other options, but it would then defeat the
> > purpose of the warnings as the problem would effectively be hidden.
> >
> > Reviewed-by: Dimitri Sivanich <[email protected]>
> > Tested-by: Russ Anderson <[email protected]>
> > Signed-off-by: Mike Travis <[email protected]>
> > ---
> > include/acpi/acconfig.h | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > --- linux-4.12.orig/include/acpi/acconfig.h
> > +++ linux-4.12/include/acpi/acconfig.h
> > @@ -123,7 +123,7 @@
> >
> > /* Maximum object reference count (detects object deletion issues) */
> >
> > -#define ACPI_MAX_REFERENCE_COUNT 0x1000
> > +#define ACPI_MAX_REFERENCE_COUNT 0x2000
> >
> > /* Default page size for use in mapping memory for operation regions */
> >
> >
> > --
>

--
Russ Anderson, SuperDome Flex Linux Kernel Group Manager
HPE - Hewlett Packard Enterprise (formerly SGI) [email protected]

2018-08-01 21:25:27

by Moore, Robert

[permalink] [raw]
Subject: RE: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large Reference Count warnings

> -----Original Message-----
> From: Russ Anderson [mailto:[email protected]]
> Sent: Wednesday, August 1, 2018 1:38 PM
> To: Moore, Robert <[email protected]>
> Cc: Mike Travis <[email protected]>; Schmauss, Erik
> <[email protected]>; Wysocki, Rafael J
> <[email protected]>; Len Brown <[email protected]>; Russ Anderson
> <[email protected]>; Dimitri Sivanich <[email protected]>;
> Ross Zwisler <[email protected]>; Williams, Dan J
> <[email protected]>; Verma, Vishal L <[email protected]>;
> Jiang, Dave <[email protected]>; [email protected];
> [email protected]; [email protected]; linux-
> [email protected]
> Subject: Re: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large
> Reference Count warnings
>
> On Wed, Aug 01, 2018 at 07:49:14PM +0000, Moore, Robert wrote:
> > Are there any actual memory leaks, or this a consequence of the number
> of sockets?
>
> It is a consequence of the number of sockets.
> It is not an actual memory leak.
>
> Thanks.
>
[Moore, Robert]

It does bother me a bit that there are over 4096 references to a single object, however. Could you send us the ACPI dump for the machine?

This issue sounds vaguely familiar though. Perhaps Dan remembers, it was something about large reference counts and the NFIT/NVDIMM stuff. If so, this has been fixed for some time now.

Bob


> > > -----Original Message-----
> > > From: Mike Travis [mailto:[email protected]]
> > > Sent: Monday, July 30, 2018 12:56 PM
> > > To: Moore, Robert <[email protected]>; Schmauss, Erik
> > > <[email protected]>; Wysocki, Rafael J
> > > <[email protected]>; Len Brown <[email protected]>
> > > Cc: Russ Anderson <[email protected]>; Dimitri Sivanich
> > > <[email protected]>; Ross Zwisler
> > > <[email protected]>; Williams, Dan J
> > > <[email protected]>; Verma, Vishal L
> > > <[email protected]>; Jiang, Dave <[email protected]>;
> > > linux- [email protected]; [email protected];
> > > [email protected]; [email protected]
> > > Subject: [PATCH 1/1] x86, pmem, acpi: Remove excessive ACPI Large
> > > Reference Count warnings
> > >
> > > With the Intel BIOS support for 8 processor sockets with a full
> > > complement of NVDIMMS potentially installable, and there are empty
> > > sockets without NVDIMMS, there is an extremely large amount of the
> > > following warnings:
> > >
> > > ACPI Warning: Large Reference Count (0x1001) in object
> ffff99453fc71750,
> > > Type=0x0A
> > >
> > > On a 4 socket system with 4 NVDIMMs there were over 6000 of these
> > > warning messages and it has been seem on systems from 4 to 32
> sockets.
> > >
> > > Through some guidance from the BIOS developers and testing, it
> > > appears that simply bumping up the threshold for warnings from
> > > 0x1000 to 0x2000 eliminates these messages. Changing them to be
> > > ACPI DEBUG messages, or even removing them are other options, but it
> > > would then defeat the purpose of the warnings as the problem would
> effectively be hidden.
> > >
> > > Reviewed-by: Dimitri Sivanich <[email protected]>
> > > Tested-by: Russ Anderson <[email protected]>
> > > Signed-off-by: Mike Travis <[email protected]>
> > > ---
> > > include/acpi/acconfig.h | 2 +-
> > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > --- linux-4.12.orig/include/acpi/acconfig.h
> > > +++ linux-4.12/include/acpi/acconfig.h
> > > @@ -123,7 +123,7 @@
> > >
> > > /* Maximum object reference count (detects object deletion issues)
> > > */
> > >
> > > -#define ACPI_MAX_REFERENCE_COUNT 0x1000
> > > +#define ACPI_MAX_REFERENCE_COUNT 0x2000
> > >
> > > /* Default page size for use in mapping memory for operation
> > > regions */
> > >
> > >
> > > --
> >
>
> --
> Russ Anderson, SuperDome Flex Linux Kernel Group Manager HPE - Hewlett
> Packard Enterprise (formerly SGI) [email protected]