On Thu, Aug 31, 2017 at 06:00:43PM +0100, Dave P Martin wrote:
> +/*
> + * Handle SVE state across fork():
> + *
> + * dst and src must not end up with aliases of the same sve_state.
> + * Because a task cannot fork except in a syscall, we can discard SVE
> + * state for dst here, so long as we take care to retain the FPSIMD
> + * subset of the state if SVE is in use. Reallocation of the SVE state
> + * will be deferred until dst tries to use SVE.
> + */
> +void fpsimd_dup_sve(struct task_struct *dst, struct task_struct const *src)
> +{
> + if (test_and_clear_tsk_thread_flag(dst, TIF_SVE)) {
> + WARN_ON(dst->mm && !in_syscall(task_pt_regs(dst)));
> + sve_to_fpsimd(dst);
> + }
> +
> + dst->thread.sve_state = NULL;
> +}
I first thought the thread flags are not visible in dst yet since
dup_task_struct() calls arch_dup_task_struct() before
setup_thread_stack(). However, at the end of the last year we enabled
CONFIG_THREAD_INFO_IN_TASK_STRUCT. But I don't particularly like relying
on this.
Anyway, IIUC we don't need sve_to_fpsimd() here. The
arch_dup_task_struct() already called fpsimd_preserve_current_state()
for src, so the FPSIMD state (which we care about) is transferred during
the *dst = *src assignment. So you'd only need the last statement,
possibly with a different function name like fpsimd_erase_sve (and maybe
make the function static inline in the header).
[...]
> int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
> @@ -246,6 +247,9 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
> if (current->mm)
> fpsimd_preserve_current_state();
> *dst = *src;
> +
> + fpsimd_dup_sve(dst, src);
> +
> return 0;
> }
--
Catalin
