On Wed, Oct 07, 2026 at 02:01:13PM +0100, Bruce Richardson wrote:
> On Tue, Oct 06, 2026 at 08:20:56AM +0000, Dragos Tarcatu wrote:
> > rte_cpu_get_flag_enabled() reads only the CPUID feature bits. CPUID
> > tells what the CPU implements, not what the OS enables. If the OS does
> > not enable the related state components in XCR0, the first AVX or AVX512
> > instruction raises #UD.
> >
> > CPUID can report AVX512 while XCR0 has no AVX512 state. Old kernels
> > without AVX512 state support leave XCR0 at 0x7. A kernel booted with
> > clearcpuid=avx512f removes the AVX512 state from XCR0, but CPUID still
> > reports AVX512F. Similarly, gather_data_sampling=force removes the AVX
> > state.
> >
> > On such a system, the rte_net_crc constructor selects the AVX512 code
> > from CPUID alone. The dynamic loader runs this constructor before main(),
> > so the application gets SIGILL before EAL can parse any argument.
> >
> > For the AVX-class and AVX512-class flags, also read XCR0 with XGETBV.
> > Report the flag only when the OS has enabled the related state. Use
> > XGETBV only when CPUID reports OSXSAVE, because the instruction faults
> > otherwise.
> >
> > This change also affects rte_cpu_is_supported(). If DPDK is built for
> > AVX512 and the OS does not enable AVX512 state, EAL can now stop with an
> > error that names the missing flag. This is possible only if no AVX512
> > instruction executes before rte_eal_init(), for example in a constructor.
> >
> > Fixes: 3934bb741354 ("eal/x86: identify AVX512 CPU flag") Cc:
> > [email protected]
> >
> > Signed-off-by: Dragos Tarcatu <[email protected]> ---
> > lib/eal/include/generic/rte_cpuflags.h | 3 + lib/eal/x86/rte_cpuflags.c
> > | 102 ++++++++++++++++++++----- 2 files changed, 86 insertions(+), 19
> > deletions(-)
> >
> Tested this using two "x86-generic" builds from test-meson-builds.sh, one
> with this patch and one without. The test app from the one without the
> patch gives illegal instruction on an initializer for avx512 crc, while the
> one with runs successfully.
Just for clarification, the test was run on a system where I'd passed
clearcpuid=avx512f to the kernel (as described above in the commit log message).
>
> Code looks correct to me also.
>
> Tested-by: Bruce Richardson <[email protected]>
> Acked-by: Bruce Richardson <[email protected]>
>
