| Message ID | PAWPR08MB8982FE280E0036D1F981733E832D2@PAWPR08MB8982.eurprd08.prod.outlook.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v2] AArch64: Check kernel version for SVE ifuncs | expand |
The 03/18/2024 14:14, Wilco Dijkstra wrote: > > v2: Add __LINUX_KERNEL_VERSION #ifdefs, improve parser, use 8:8:8 format. > > Old Linux kernels disable SVE after every system call. Calling the > SVE-optimized memcpy afterwards will then cause a trap to reenable SVE. > As a result, applications with a high use of syscalls may run slower with > the SVE memcpy. This is true for kernels between 4.15.0 and before 6.2.0, > except for 5.14.0 which was patched. Avoid this by checking the kernel > version and selecting the SVE ifunc on modern kernels. > > Parse the kernel version reported by uname() into a 24-bit kernel.major.minor > value without calling any library functions. If uname() is not supported or > if the version format is not recognized, assume the kernel is modern. > > Passes regress, OK for commit? OK to commit. (clearly a hack but what can we do..) Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com> > > --- > > diff --git a/sysdeps/aarch64/cpu-features.h b/sysdeps/aarch64/cpu-features.h > index 77a782422af1b6e4b2af32bfebfda37874111510..5f2da91ebbd0adafb0d84ec503b0f902f566da5a 100644 > --- a/sysdeps/aarch64/cpu-features.h > +++ b/sysdeps/aarch64/cpu-features.h > @@ -71,6 +71,7 @@ struct cpu_features > /* Currently, the GLIBC memory tagging tunable only defines 8 bits. */ > uint8_t mte_state; > bool sve; > + bool prefer_sve_ifuncs; > bool mops; > }; > > diff --git a/sysdeps/aarch64/multiarch/init-arch.h b/sysdeps/aarch64/multiarch/init-arch.h > index c52860efb22d70eb4bdf356781f51c7de8ec67dc..61dc40088f4d9e5e06b57bdc7d54bde1e2a686a4 100644 > --- a/sysdeps/aarch64/multiarch/init-arch.h > +++ b/sysdeps/aarch64/multiarch/init-arch.h > @@ -36,5 +36,7 @@ > MTE_ENABLED (); \ > bool __attribute__((unused)) sve = \ > GLRO(dl_aarch64_cpu_features).sve; \ > + bool __attribute__((unused)) prefer_sve_ifuncs = \ > + GLRO(dl_aarch64_cpu_features).prefer_sve_ifuncs; \ > bool __attribute__((unused)) mops = \ > GLRO(dl_aarch64_cpu_features).mops; > diff --git a/sysdeps/aarch64/multiarch/memcpy.c b/sysdeps/aarch64/multiarch/memcpy.c > index d12eccfca51f4bcfef6ccf5aa286edb301e361ac..ce53567dab33c2f00b89b4069235abd4651811a6 100644 > --- a/sysdeps/aarch64/multiarch/memcpy.c > +++ b/sysdeps/aarch64/multiarch/memcpy.c > @@ -47,7 +47,7 @@ select_memcpy_ifunc (void) > { > if (IS_A64FX (midr)) > return __memcpy_a64fx; > - return __memcpy_sve; > + return prefer_sve_ifuncs ? __memcpy_sve : __memcpy_generic; > } > > if (IS_THUNDERX (midr)) > diff --git a/sysdeps/aarch64/multiarch/memmove.c b/sysdeps/aarch64/multiarch/memmove.c > index 2081eeb4d40e0240e67a7b26b64576f44eaf18e3..fe95037be391896c7670ef606bf4d3ba7dfb6a00 100644 > --- a/sysdeps/aarch64/multiarch/memmove.c > +++ b/sysdeps/aarch64/multiarch/memmove.c > @@ -47,7 +47,7 @@ select_memmove_ifunc (void) > { > if (IS_A64FX (midr)) > return __memmove_a64fx; > - return __memmove_sve; > + return prefer_sve_ifuncs ? __memmove_sve : __memmove_generic; > } > > if (IS_THUNDERX (midr)) > diff --git a/sysdeps/unix/sysv/linux/aarch64/cpu-features.c b/sysdeps/unix/sysv/linux/aarch64/cpu-features.c > index b1a3f673f067280bdacfddd92723a81e418023e5..c0b047bc0dbeae428c89e12688b7d802e4cb3a43 100644 > --- a/sysdeps/unix/sysv/linux/aarch64/cpu-features.c > +++ b/sysdeps/unix/sysv/linux/aarch64/cpu-features.c > @@ -21,6 +21,7 @@ > #include <sys/auxv.h> > #include <elf/dl-hwcaps.h> > #include <sys/prctl.h> > +#include <sys/utsname.h> > #include <dl-tunables-parse.h> > > #define DCZID_DZP_MASK (1 << 4) > @@ -62,6 +63,46 @@ get_midr_from_mcpu (const struct tunable_str_t *mcpu) > return UINT64_MAX; > } > > +#if __LINUX_KERNEL_VERSION < 0x060200 > + > +/* Return true if we prefer using SVE in string ifuncs. Old kernels disable > + SVE after every system call which results in unnecessary traps if memcpy > + uses SVE. This is true for kernels between 4.15.0 and before 6.2.0, except > + for 5.14.0 which was patched. For these versions return false to avoid using > + SVE ifuncs. > + Parse the kernel version into a 24-bit kernel.major.minor value without > + calling any library functions. If uname() is not supported or if the version > + format is not recognized, assume the kernel is modern and return true. */ > + > +static inline bool > +prefer_sve_ifuncs (void) > +{ > + struct utsname buf; > + const char *p = &buf.release[0]; > + int kernel = 0; > + int val; > + > + if (__uname (&buf) < 0) > + return true; > + > + for (int shift = 16; shift >= 0; shift -= 8) > + { > + for (val = 0; *p >= '0' && *p <= '9'; p++) > + val = val * 10 + *p - '0'; > + kernel |= (val & 255) << shift; > + if (*p++ != '.') > + break; > + } > + > + if (kernel >= 0x060200 || kernel == 0x050e00) > + return true; > + if (kernel >= 0x040f00) > + return false; > + return true; > +} > + > +#endif > + > static inline void > init_cpu_features (struct cpu_features *cpu_features) > { > @@ -126,6 +167,13 @@ init_cpu_features (struct cpu_features *cpu_features) > /* Check if SVE is supported. */ > cpu_features->sve = GLRO (dl_hwcap) & HWCAP_SVE; > > + cpu_features->prefer_sve_ifuncs = cpu_features->sve; > + > +#if __LINUX_KERNEL_VERSION < 0x060200 > + if (cpu_features->sve) > + cpu_features->prefer_sve_ifuncs = prefer_sve_ifuncs (); > +#endif > + > /* Check if MOPS is supported. */ > cpu_features->mops = GLRO (dl_hwcap2) & HWCAP2_MOPS; > } >
* Szabolcs Nagy: > The 03/18/2024 14:14, Wilco Dijkstra wrote: >> >> v2: Add __LINUX_KERNEL_VERSION #ifdefs, improve parser, use 8:8:8 format. >> >> Old Linux kernels disable SVE after every system call. Calling the >> SVE-optimized memcpy afterwards will then cause a trap to reenable SVE. >> As a result, applications with a high use of syscalls may run slower with >> the SVE memcpy. This is true for kernels between 4.15.0 and before 6.2.0, >> except for 5.14.0 which was patched. Avoid this by checking the kernel >> version and selecting the SVE ifunc on modern kernels. >> >> Parse the kernel version reported by uname() into a 24-bit kernel.major.minor >> value without calling any library functions. If uname() is not supported or >> if the version format is not recognized, assume the kernel is modern. >> >> Passes regress, OK for commit? > > OK to commit. (clearly a hack but what can we do..) > > Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com> I have not had a chance to test this yet with the el9 kernel. I will try to do this tomorrow. Do we need to include other distribution LTS kernels in the version check? Thanks, Florian
The 03/20/2024 16:39, Florian Weimer wrote: > * Szabolcs Nagy: > > > The 03/18/2024 14:14, Wilco Dijkstra wrote: > >> > >> v2: Add __LINUX_KERNEL_VERSION #ifdefs, improve parser, use 8:8:8 format. > >> > >> Old Linux kernels disable SVE after every system call. Calling the > >> SVE-optimized memcpy afterwards will then cause a trap to reenable SVE. > >> As a result, applications with a high use of syscalls may run slower with > >> the SVE memcpy. This is true for kernels between 4.15.0 and before 6.2.0, > >> except for 5.14.0 which was patched. Avoid this by checking the kernel > >> version and selecting the SVE ifunc on modern kernels. > >> > >> Parse the kernel version reported by uname() into a 24-bit kernel.major.minor > >> value without calling any library functions. If uname() is not supported or > >> if the version format is not recognized, assume the kernel is modern. > >> > >> Passes regress, OK for commit? > > > > OK to commit. (clearly a hack but what can we do..) > > > > Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com> > > I have not had a chance to test this yet with the el9 kernel. > > I will try to do this tomorrow. > > Do we need to include other distribution LTS kernels in the version > check? i checked some (suse,ubuntu,debian) and they didn't have the SVE backport (maybe we should request those backports, but today the version check looks ok) > > Thanks, > Florian >
* Szabolcs Nagy: > The 03/20/2024 16:39, Florian Weimer wrote: >> * Szabolcs Nagy: >> >> > The 03/18/2024 14:14, Wilco Dijkstra wrote: >> >> >> >> v2: Add __LINUX_KERNEL_VERSION #ifdefs, improve parser, use 8:8:8 format. >> >> >> >> Old Linux kernels disable SVE after every system call. Calling the >> >> SVE-optimized memcpy afterwards will then cause a trap to reenable SVE. >> >> As a result, applications with a high use of syscalls may run slower with >> >> the SVE memcpy. This is true for kernels between 4.15.0 and before 6.2.0, >> >> except for 5.14.0 which was patched. Avoid this by checking the kernel >> >> version and selecting the SVE ifunc on modern kernels. >> >> >> >> Parse the kernel version reported by uname() into a 24-bit kernel.major.minor >> >> value without calling any library functions. If uname() is not supported or >> >> if the version format is not recognized, assume the kernel is modern. >> >> >> >> Passes regress, OK for commit? >> > >> > OK to commit. (clearly a hack but what can we do..) >> > >> > Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com> >> >> I have not had a chance to test this yet with the el9 kernel. >> >> I will try to do this tomorrow. >> >> Do we need to include other distribution LTS kernels in the version >> check? > > i checked some (suse,ubuntu,debian) and they didn't have the > SVE backport (maybe we should request those backports, but > today the version check looks ok) I think we should make this change in one commit if at all possible, otherwise we risk creating too much divergence. But if you consider further kernel backports unlikely, the v2 patch is okay. Anyway, I verified this against 4.18 and 5.14 kernels, and got the expected results (__memcpy_generic and __memcpy_sve). Tested-by: Florian Weimer <fweimer@redhat.com> Thanks, Florian
On 21/03/24 08:44, Florian Weimer wrote: > * Szabolcs Nagy: > >> The 03/20/2024 16:39, Florian Weimer wrote: >>> * Szabolcs Nagy: >>> >>>> The 03/18/2024 14:14, Wilco Dijkstra wrote: >>>>> >>>>> v2: Add __LINUX_KERNEL_VERSION #ifdefs, improve parser, use 8:8:8 format. >>>>> >>>>> Old Linux kernels disable SVE after every system call. Calling the >>>>> SVE-optimized memcpy afterwards will then cause a trap to reenable SVE. >>>>> As a result, applications with a high use of syscalls may run slower with >>>>> the SVE memcpy. This is true for kernels between 4.15.0 and before 6.2.0, >>>>> except for 5.14.0 which was patched. Avoid this by checking the kernel >>>>> version and selecting the SVE ifunc on modern kernels. >>>>> >>>>> Parse the kernel version reported by uname() into a 24-bit kernel.major.minor >>>>> value without calling any library functions. If uname() is not supported or >>>>> if the version format is not recognized, assume the kernel is modern. >>>>> >>>>> Passes regress, OK for commit? >>>> >>>> OK to commit. (clearly a hack but what can we do..) >>>> >>>> Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com> >>> >>> I have not had a chance to test this yet with the el9 kernel. >>> >>> I will try to do this tomorrow. >>> >>> Do we need to include other distribution LTS kernels in the version >>> check? >> >> i checked some (suse,ubuntu,debian) and they didn't have the >> SVE backport (maybe we should request those backports, but >> today the version check looks ok) > > I think we should make this change in one commit if at all possible, > otherwise we risk creating too much divergence. But if you consider > further kernel backports unlikely, the v2 patch is okay. > > Anyway, I verified this against 4.18 and 5.14 kernels, and got the > expected results (__memcpy_generic and __memcpy_sve). If you have check through the string tests, you might not get the correct answer since it uses the ifunc-impl-list.c and this patch does not changes the reported functions. > > Tested-by: Florian Weimer <fweimer@redhat.com> > > Thanks, > Florian >
* Adhemerval Zanella Netto: >> Anyway, I verified this against 4.18 and 5.14 kernels, and got the >> expected results (__memcpy_generic and __memcpy_sve). > > If you have check through the string tests, you might not get the > correct answer since it uses the ifunc-impl-list.c and this patch > does not changes the reported functions. No, I used GDB to resolve the addresses returned by dlsym back to a symbol. Thanks, Florian
diff --git a/sysdeps/aarch64/cpu-features.h b/sysdeps/aarch64/cpu-features.h index 77a782422af1b6e4b2af32bfebfda37874111510..5f2da91ebbd0adafb0d84ec503b0f902f566da5a 100644 --- a/sysdeps/aarch64/cpu-features.h +++ b/sysdeps/aarch64/cpu-features.h @@ -71,6 +71,7 @@ struct cpu_features /* Currently, the GLIBC memory tagging tunable only defines 8 bits. */ uint8_t mte_state; bool sve; + bool prefer_sve_ifuncs; bool mops; }; diff --git a/sysdeps/aarch64/multiarch/init-arch.h b/sysdeps/aarch64/multiarch/init-arch.h index c52860efb22d70eb4bdf356781f51c7de8ec67dc..61dc40088f4d9e5e06b57bdc7d54bde1e2a686a4 100644 --- a/sysdeps/aarch64/multiarch/init-arch.h +++ b/sysdeps/aarch64/multiarch/init-arch.h @@ -36,5 +36,7 @@ MTE_ENABLED (); \ bool __attribute__((unused)) sve = \ GLRO(dl_aarch64_cpu_features).sve; \ + bool __attribute__((unused)) prefer_sve_ifuncs = \ + GLRO(dl_aarch64_cpu_features).prefer_sve_ifuncs; \ bool __attribute__((unused)) mops = \ GLRO(dl_aarch64_cpu_features).mops; diff --git a/sysdeps/aarch64/multiarch/memcpy.c b/sysdeps/aarch64/multiarch/memcpy.c index d12eccfca51f4bcfef6ccf5aa286edb301e361ac..ce53567dab33c2f00b89b4069235abd4651811a6 100644 --- a/sysdeps/aarch64/multiarch/memcpy.c +++ b/sysdeps/aarch64/multiarch/memcpy.c @@ -47,7 +47,7 @@ select_memcpy_ifunc (void) { if (IS_A64FX (midr)) return __memcpy_a64fx; - return __memcpy_sve; + return prefer_sve_ifuncs ? __memcpy_sve : __memcpy_generic; } if (IS_THUNDERX (midr)) diff --git a/sysdeps/aarch64/multiarch/memmove.c b/sysdeps/aarch64/multiarch/memmove.c index 2081eeb4d40e0240e67a7b26b64576f44eaf18e3..fe95037be391896c7670ef606bf4d3ba7dfb6a00 100644 --- a/sysdeps/aarch64/multiarch/memmove.c +++ b/sysdeps/aarch64/multiarch/memmove.c @@ -47,7 +47,7 @@ select_memmove_ifunc (void) { if (IS_A64FX (midr)) return __memmove_a64fx; - return __memmove_sve; + return prefer_sve_ifuncs ? __memmove_sve : __memmove_generic; } if (IS_THUNDERX (midr)) diff --git a/sysdeps/unix/sysv/linux/aarch64/cpu-features.c b/sysdeps/unix/sysv/linux/aarch64/cpu-features.c index b1a3f673f067280bdacfddd92723a81e418023e5..c0b047bc0dbeae428c89e12688b7d802e4cb3a43 100644 --- a/sysdeps/unix/sysv/linux/aarch64/cpu-features.c +++ b/sysdeps/unix/sysv/linux/aarch64/cpu-features.c @@ -21,6 +21,7 @@ #include <sys/auxv.h> #include <elf/dl-hwcaps.h> #include <sys/prctl.h> +#include <sys/utsname.h> #include <dl-tunables-parse.h> #define DCZID_DZP_MASK (1 << 4) @@ -62,6 +63,46 @@ get_midr_from_mcpu (const struct tunable_str_t *mcpu) return UINT64_MAX; } +#if __LINUX_KERNEL_VERSION < 0x060200 + +/* Return true if we prefer using SVE in string ifuncs. Old kernels disable + SVE after every system call which results in unnecessary traps if memcpy + uses SVE. This is true for kernels between 4.15.0 and before 6.2.0, except + for 5.14.0 which was patched. For these versions return false to avoid using + SVE ifuncs. + Parse the kernel version into a 24-bit kernel.major.minor value without + calling any library functions. If uname() is not supported or if the version + format is not recognized, assume the kernel is modern and return true. */ + +static inline bool +prefer_sve_ifuncs (void) +{ + struct utsname buf; + const char *p = &buf.release[0]; + int kernel = 0; + int val; + + if (__uname (&buf) < 0) + return true; + + for (int shift = 16; shift >= 0; shift -= 8) + { + for (val = 0; *p >= '0' && *p <= '9'; p++) + val = val * 10 + *p - '0'; + kernel |= (val & 255) << shift; + if (*p++ != '.') + break; + } + + if (kernel >= 0x060200 || kernel == 0x050e00) + return true; + if (kernel >= 0x040f00) + return false; + return true; +} + +#endif + static inline void init_cpu_features (struct cpu_features *cpu_features) { @@ -126,6 +167,13 @@ init_cpu_features (struct cpu_features *cpu_features) /* Check if SVE is supported. */ cpu_features->sve = GLRO (dl_hwcap) & HWCAP_SVE; + cpu_features->prefer_sve_ifuncs = cpu_features->sve; + +#if __LINUX_KERNEL_VERSION < 0x060200 + if (cpu_features->sve) + cpu_features->prefer_sve_ifuncs = prefer_sve_ifuncs (); +#endif + /* Check if MOPS is supported. */ cpu_features->mops = GLRO (dl_hwcap2) & HWCAP2_MOPS; }