| Message ID | 20170628145814.24763-1-luc.vanoostenryck@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, Jun 28, 2017 at 04:58:14PM +0200, Luc Van Oostenryck wrote: > ARM64 implementation of ip_fast_csum() do most of the work > in 128 or 64 bit and call csum_fold() to finalize. csum_fold() > itself take a __wsum argument, to insure that this value is > always a 32bit native-order value. > > Fix this by using an helper __csum_fold() taking the native > 32 bit value and doing the needed folding to 16 bit (and reuse > this helper for csum_fold() itself). > > Note: a simpler patch would to use something like: > return csum_fold((__wsum __force)(sum >> 32)); > but using an helper __csum_fold() allow to avoid > a forced cast. But you've added a __force cast in csum_fold, and we still have the one in the return statement of __csum_fold, so I think I prefer the simpler patch. Will > > Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com> > --- > arch/arm64/include/asm/checksum.h | 10 +++++++--- > 1 file changed, 7 insertions(+), 3 deletions(-) > > diff --git a/arch/arm64/include/asm/checksum.h b/arch/arm64/include/asm/checksum.h > index 09f65339d..dcc655137 100644 > --- a/arch/arm64/include/asm/checksum.h > +++ b/arch/arm64/include/asm/checksum.h > @@ -18,12 +18,16 @@ > > #include <linux/types.h> > > -static inline __sum16 csum_fold(__wsum csum) > +static inline __sum16 __csum_fold(u32 sum) > { > - u32 sum = (__force u32)csum; > sum += (sum >> 16) | (sum << 16); > return ~(__force __sum16)(sum >> 16); > } > + > +static inline __sum16 csum_fold(__wsum csum) > +{ > + return __csum_fold((__force u32)csum); > +} > #define csum_fold csum_fold > > static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > @@ -42,7 +46,7 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) > } while (--ihl); > > sum += ((sum >> 32) | (sum << 32)); > - return csum_fold(sum >> 32); > + return __csum_fold(sum >> 32); > } > #define ip_fast_csum ip_fast_csum > > -- > 2.13.0 >
diff --git a/arch/arm64/include/asm/checksum.h b/arch/arm64/include/asm/checksum.h index 09f65339d..dcc655137 100644 --- a/arch/arm64/include/asm/checksum.h +++ b/arch/arm64/include/asm/checksum.h @@ -18,12 +18,16 @@ #include <linux/types.h> -static inline __sum16 csum_fold(__wsum csum) +static inline __sum16 __csum_fold(u32 sum) { - u32 sum = (__force u32)csum; sum += (sum >> 16) | (sum << 16); return ~(__force __sum16)(sum >> 16); } + +static inline __sum16 csum_fold(__wsum csum) +{ + return __csum_fold((__force u32)csum); +} #define csum_fold csum_fold static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) @@ -42,7 +46,7 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl) } while (--ihl); sum += ((sum >> 32) | (sum << 32)); - return csum_fold(sum >> 32); + return __csum_fold(sum >> 32); } #define ip_fast_csum ip_fast_csum
ARM64 implementation of ip_fast_csum() do most of the work in 128 or 64 bit and call csum_fold() to finalize. csum_fold() itself take a __wsum argument, to insure that this value is always a 32bit native-order value. Fix this by using an helper __csum_fold() taking the native 32 bit value and doing the needed folding to 16 bit (and reuse this helper for csum_fold() itself). Note: a simpler patch would to use something like: return csum_fold((__wsum __force)(sum >> 32)); but using an helper __csum_fold() allow to avoid a forced cast. Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com> --- arch/arm64/include/asm/checksum.h | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)