On October 10, 2026 9:24:16 AM PDT, Magnus Lindholm <[email protected]> wrote:
>Hi all,
>
>On Sat, Oct 10, 2026 at 3:42 PM Magnus Lindholm <[email protected]> wrote:
>>
>> Enable the seccomp BPF selftests on Alpha with a fallback seccomp syscall
>> number and architecture-specific syscall register accessors.
>>
>> Use the NT_PRSTATUS register layout rather than struct pt_regs. The
>> register set contains r0-r30 followed by pc and unique. Read and update
>> the syscall number in r1, and set both r0 and the r19/a3 error flag when
>> synthesizing syscall results.
>>
>> At syscall entry, r19 still contains the fourth syscall argument, so
>> callers setting a synthetic return value must also set the syscall
>> number to -1 to skip execution.
>>
>> Signed-off-by: Magnus Lindholm <[email protected]>
>> Reviewed-by: Matt Turner <[email protected]>
>> ---
>> Changes in v3:
>> - Drop the /proc/self/syscall availability check and its commit-message
>> paragraph. Alpha selects HAVE_ARCH_TRACEHOOK.
>> - Document that setting r19/a3 at syscall entry requires callers to also
>> set the syscall number to -1.
>> - Add Matt Turner's Reviewed-by after addressing his review comments.
>>
>> Changes in v2:
>> - Use the NT_PRSTATUS layout for PTRACE_GETREGSET/PTRACE_SETREGSET.
>> - Set the Alpha return value and error flag together.
>> - Check for /proc/self/syscall at runtime (removed in v3).
>>
>> Validation on AlphaServer ES40, GCC 15.2.1, running
>> 7.3.0-rc1-es40-tsunami-v5a+:
>> - Standard v3 and the original baseline binary both report 98 pass,
>> 1 fail, 12 skip. The common failure is the 30-second timeout in
>> user_notification_wait_killable_after_reply.
>> - A separate diagnostic build extending only that test's timeout to
>> 180 seconds completes it in 34.938 seconds. The full diagnostic run
>> reports 99 pass, 0 fail, 12 skip. No timeout change is in this patch.
>>
>> tools/testing/selftests/seccomp/seccomp_bpf.c | 26 +++++++++++++++++++
>> 1 file changed, 26 insertions(+)
>>
>> diff --git a/tools/testing/selftests/seccomp/seccomp_bpf.c
>> b/tools/testing/selftests/seccomp/seccomp_bpf.c
>> index 0622bc2..79e8eb3 100644
>> --- a/tools/testing/selftests/seccomp/seccomp_bpf.c
>> +++ b/tools/testing/selftests/seccomp/seccomp_bpf.c
>> @@ -136,6 +136,8 @@ struct seccomp_data {
>> # define __NR_seccomp 354
>> # elif defined(__x86_64__)
>> # define __NR_seccomp 317
>> +# elif defined(__alpha__)
>> +# define __NR_seccomp 514
>> # elif defined(__arm__)
>> # define __NR_seccomp 383
>> # elif defined(__aarch64__)
>> @@ -1754,6 +1756,30 @@ TEST_F(TRACE_poke, getpid_runs_normally)
>> # define ARCH_REGS struct user_regs_struct
>> # define SYSCALL_NUM(_regs) (_regs).orig_eax
>> # define SYSCALL_RET(_regs) (_regs).eax
>> +#elif defined(__alpha__)
>> +/* NT_PRSTATUS contains r0-r30, pc and unique, not struct pt_regs. */
>> +struct alpha_regs {
>> + unsigned long regs[33];
>> +};
>> +
>> +# define ARCH_REGS struct alpha_regs
>> +/* The kernel keeps the mutable syscall number in r1. */
>> +# define SYSCALL_NUM(_regs) ((_regs).regs[1])
>> +# define SYSCALL_RET(_regs) ((_regs).regs[0])
>> +/*
>> + * Alpha returns positive errno in r0 with the r19/a3 error flag set.
>> + * At a syscall-entry stop, r19 is still the fourth syscall argument.
>> + * Callers must also set the syscall number to -1 to skip the syscall
>> + * before replacing r19 with the error flag.
>> + */
>> +# define SYSCALL_RET_SET(_regs, _val) \
>> + do { \
>> + struct alpha_regs *__regs = &(_regs); \
>> + long __ret = (_val); \
>> + \
>> + __regs->regs[0] = __ret < 0 ? -__ret : __ret; \
>> + __regs->regs[19] = __ret < 0; \
>> + } while (0)
>> #elif defined(__arm__)
>> # define ARCH_REGS struct pt_regs
>> # define SYSCALL_NUM(_regs) (_regs).ARM_r7
>> --
>> 2.43.0
>>
>
>Kees,
>Would you be OK with me taking this patch through my Alpha tree?
>The Alpha kernel support is already upstream.
>
>v3 carries Matt's Reviewed-by. On my ES40, it matches the baseline:
>98 passes, 12 skips and one timeout. Extending that timeout gives
>99 passes and 12 skips; that adjustment is not part of the patch.
>
>Link:
>https://lore.kernel.org/linux-alpha/[email protected]/
>
>Thanks,
>Magnus
Yup, totally fine. Thank you!
-Kees
--
Kees Cook
