A bug in the gcc-6.0 prerelease version caused at least one
driver (lpfc) to have excessive stack usage when dealing with
wwn data, on the ARM architecture.
lpfc_scsi.c: In function 'lpfc_find_next_oas_lun':
lpfc_scsi.c:117:1: warning: the frame size of 1152 bytes is larger than 1024 bytes [-Wframe-larger-than=]
I have reported this as a gcc regression in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70232
However, using a better implementation of wwn_to_u64() not only
helps with the particular gcc problem but also leads to better
object code for any version or architecture.
The kernel already provides get_unaligned_be64() and
put_unaligned_be64() helper functions that provide an
optimized implementation with the desired semantics.
The lpfc_find_next_oas_lun() function in the example that
grew from 1146 bytes to 5144 bytes when moving from gcc-5.3
to gcc-6.0 is now 804 bytes, as the optimized
get_unaligned_be64() load can be done in three instructions.
The stack usage is now down to 28 bytes from 128 bytes with
gcc-5.3 before.
Signed-off-by: Arnd Bergmann <[email protected]>
---
include/scsi/scsi_transport_fc.h | 15 +++------------
1 file changed, 3 insertions(+), 12 deletions(-)
diff --git a/include/scsi/scsi_transport_fc.h b/include/scsi/scsi_transport_fc.h
index 784bc2c0929f..bf66ea6bed2b 100644
--- a/include/scsi/scsi_transport_fc.h
+++ b/include/scsi/scsi_transport_fc.h
@@ -28,6 +28,7 @@
#define SCSI_TRANSPORT_FC_H
#include <linux/sched.h>
+#include <asm/unaligned.h>
#include <scsi/scsi.h>
#include <scsi/scsi_netlink.h>
@@ -797,22 +798,12 @@ fc_remote_port_chkready(struct fc_rport *rport)
static inline u64 wwn_to_u64(u8 *wwn)
{
- return (u64)wwn[0] << 56 | (u64)wwn[1] << 48 |
- (u64)wwn[2] << 40 | (u64)wwn[3] << 32 |
- (u64)wwn[4] << 24 | (u64)wwn[5] << 16 |
- (u64)wwn[6] << 8 | (u64)wwn[7];
+ return get_unaligned_be64(wwn);
}
static inline void u64_to_wwn(u64 inm, u8 *wwn)
{
- wwn[0] = (inm >> 56) & 0xff;
- wwn[1] = (inm >> 48) & 0xff;
- wwn[2] = (inm >> 40) & 0xff;
- wwn[3] = (inm >> 32) & 0xff;
- wwn[4] = (inm >> 24) & 0xff;
- wwn[5] = (inm >> 16) & 0xff;
- wwn[6] = (inm >> 8) & 0xff;
- wwn[7] = inm & 0xff;
+ put_unaligned_be64(inm, wwn);
}
/**
--
2.7.0
On Wed, 2016-03-16 at 17:39 +0100, Arnd Bergmann wrote:
> A bug in the gcc-6.0 prerelease version caused at least one
> driver (lpfc) to have excessive stack usage when dealing with
> wwn data, on the ARM architecture.
>
> lpfc_scsi.c: In function 'lpfc_find_next_oas_lun':
> lpfc_scsi.c:117:1: warning: the frame size of 1152 bytes is larger than 1024 bytes [-Wframe-larger-than=]
>
> I have reported this as a gcc regression in
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70232
>
> However, using a better implementation of wwn_to_u64() not only
> helps with the particular gcc problem but also leads to better
> object code for any version or architecture.
>
> The kernel already provides get_unaligned_be64() and
> put_unaligned_be64() helper functions that provide an
> optimized implementation with the desired semantics.
>
> The lpfc_find_next_oas_lun() function in the example that
> grew from 1146 bytes to 5144 bytes when moving from gcc-5.3
> to gcc-6.0 is now 804 bytes, as the optimized
> get_unaligned_be64() load can be done in three instructions.
> The stack usage is now down to 28 bytes from 128 bytes with
> gcc-5.3 before.
>
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---
> include/scsi/scsi_transport_fc.h | 15 +++------------
> 1 file changed, 3 insertions(+), 12 deletions(-)
>
> diff --git a/include/scsi/scsi_transport_fc.h b/include/scsi/scsi_transport_fc.h
> index 784bc2c0929f..bf66ea6bed2b 100644
> --- a/include/scsi/scsi_transport_fc.h
> +++ b/include/scsi/scsi_transport_fc.h
> @@ -28,6 +28,7 @@
> #define SCSI_TRANSPORT_FC_H
>
> #include <linux/sched.h>
> +#include <asm/unaligned.h>
> #include <scsi/scsi.h>
> #include <scsi/scsi_netlink.h>
>
> @@ -797,22 +798,12 @@ fc_remote_port_chkready(struct fc_rport *rport)
>
> static inline u64 wwn_to_u64(u8 *wwn)
> {
> - return (u64)wwn[0] << 56 | (u64)wwn[1] << 48 |
> - (u64)wwn[2] << 40 | (u64)wwn[3] << 32 |
> - (u64)wwn[4] << 24 | (u64)wwn[5] << 16 |
> - (u64)wwn[6] << 8 | (u64)wwn[7];
> + return get_unaligned_be64(wwn);
> }
>
> static inline void u64_to_wwn(u64 inm, u8 *wwn)
> {
> - wwn[0] = (inm >> 56) & 0xff;
> - wwn[1] = (inm >> 48) & 0xff;
> - wwn[2] = (inm >> 40) & 0xff;
> - wwn[3] = (inm >> 32) & 0xff;
> - wwn[4] = (inm >> 24) & 0xff;
> - wwn[5] = (inm >> 16) & 0xff;
> - wwn[6] = (inm >> 8) & 0xff;
> - wwn[7] = inm & 0xff;
> + put_unaligned_be64(inm, wwn);
> }
>
> /**
It would be nice to get rid of these functions completely and just
change the callers to use get/put_unaligned_be64() directly, like libfc
does, but that involves changing 7 drivers and scsi_transport_fc.
Reviewed-by: Ewan D. Milne <[email protected]>
On 03/16/2016 05:39 PM, Arnd Bergmann wrote:
> A bug in the gcc-6.0 prerelease version caused at least one
> driver (lpfc) to have excessive stack usage when dealing with
> wwn data, on the ARM architecture.
>
> lpfc_scsi.c: In function 'lpfc_find_next_oas_lun':
> lpfc_scsi.c:117:1: warning: the frame size of 1152 bytes is larger than 1024 bytes [-Wframe-larger-than=]
>
> I have reported this as a gcc regression in
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70232
>
> However, using a better implementation of wwn_to_u64() not only
> helps with the particular gcc problem but also leads to better
> object code for any version or architecture.
>
> The kernel already provides get_unaligned_be64() and
> put_unaligned_be64() helper functions that provide an
> optimized implementation with the desired semantics.
>
> The lpfc_find_next_oas_lun() function in the example that
> grew from 1146 bytes to 5144 bytes when moving from gcc-5.3
> to gcc-6.0 is now 804 bytes, as the optimized
> get_unaligned_be64() load can be done in three instructions.
> The stack usage is now down to 28 bytes from 128 bytes with
> gcc-5.3 before.
>
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---
> include/scsi/scsi_transport_fc.h | 15 +++------------
> 1 file changed, 3 insertions(+), 12 deletions(-)
>
Reviewed-by: Hannes Reinecke <[email protected]>
Cheers,
Hannes
--
Dr. Hannes Reinecke Teamlead Storage & Networking
[email protected] +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 N?rnberg
GF: F. Imend?rffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG N?rnberg)
>>>>> "Arnd" == Arnd Bergmann <[email protected]> writes:
Arnd> A bug in the gcc-6.0 prerelease version caused at least one driver
Arnd> (lpfc) to have excessive stack usage when dealing with wwn data,
Arnd> on the ARM architecture.
Applied to 4.6/scsi-fixes.
--
Martin K. Petersen Oracle Linux Engineering