| Message ID | 1409463591-661-1-git-send-email-leoy@marvell.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 31 Aug 2014, at 06:39, Leo Yan <leoy@marvell.com> wrote: > Now arm64 defers reloading FPSIMD state, but this optimization also > introduces the bug after cpu resume back from low power mode. You are right, I can see a bug here. > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c > index ad8aebb..4caef93 100644 > --- a/arch/arm64/kernel/fpsimd.c > +++ b/arch/arm64/kernel/fpsimd.c > @@ -268,13 +268,9 @@ static int fpsimd_cpu_pm_notifier(struct notifier_block *self, > { > switch (cmd) { > case CPU_PM_ENTER: > - if (current->mm && !test_thread_flag(TIF_FOREIGN_FPSTATE)) > - fpsimd_save_state(¤t->thread.fpsimd_state); That’s needed if we enter a low power state directly from a task with a valid mm (user). I’m not sure that’s possible, but just in case. > + this_cpu_write(fpsimd_last_state, NULL); That’s correct. In most cases, we enter low power state from the idle thread which does not have an mm, so the CPU_PM_EXIT case wouldn’t set a TIF_FOREIGN_FPSTATE, so thread switching would not detect the change. > break; > case CPU_PM_EXIT: > - if (current->mm) > - set_thread_flag(TIF_FOREIGN_FPSTATE); See above for why this may still be needed (it case it returns to user directly without fpsimd_thread_switch() to detect fpsimd_last_state). Catalin
On 09/01/2014 06:11 AM, Catalin Marinas wrote: > On 31 Aug 2014, at 06:39, Leo Yan <leoy@marvell.com> wrote: >> Now arm64 defers reloading FPSIMD state, but this optimization also >> introduces the bug after cpu resume back from low power mode. > > You are right, I can see a bug here. > >> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c >> index ad8aebb..4caef93 100644 >> --- a/arch/arm64/kernel/fpsimd.c >> +++ b/arch/arm64/kernel/fpsimd.c >> @@ -268,13 +268,9 @@ static int fpsimd_cpu_pm_notifier(struct notifier_block *self, >> { >> switch (cmd) { >> case CPU_PM_ENTER: >> - if (current->mm && !test_thread_flag(TIF_FOREIGN_FPSTATE)) >> - fpsimd_save_state(¤t->thread.fpsimd_state); > > That’s needed if we enter a low power state directly from a task with > a valid mm (user). I’m not sure that’s possible, but just in case. u are right, in some special cases (such like IKS, etc) a normal thread also may call cpu_suspend() function so that it will run here; i will commit another patch to reserve this part. >> + this_cpu_write(fpsimd_last_state, NULL); > > That’s correct. In most cases, we enter low power state from the idle > thread which does not have an mm, so the CPU_PM_EXIT case wouldn’t set > a TIF_FOREIGN_FPSTATE, so thread switching would not detect the change. > >> break; >> case CPU_PM_EXIT: >> - if (current->mm) >> - set_thread_flag(TIF_FOREIGN_FPSTATE); > > See above for why this may still be needed (it case it returns to user > directly without fpsimd_thread_switch() to detect fpsimd_last_state). Will reserve this part as well. Thanks, Leo Yan
diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c index ad8aebb..4caef93 100644 --- a/arch/arm64/kernel/fpsimd.c +++ b/arch/arm64/kernel/fpsimd.c @@ -268,13 +268,9 @@ static int fpsimd_cpu_pm_notifier(struct notifier_block *self, { switch (cmd) { case CPU_PM_ENTER: - if (current->mm && !test_thread_flag(TIF_FOREIGN_FPSTATE)) - fpsimd_save_state(¤t->thread.fpsimd_state); + this_cpu_write(fpsimd_last_state, NULL); break; case CPU_PM_EXIT: - if (current->mm) - set_thread_flag(TIF_FOREIGN_FPSTATE); - break; case CPU_PM_ENTER_FAILED: default: return NOTIFY_DONE;
Now arm64 defers reloading FPSIMD state, but this optimization also introduces the bug after cpu resume back from low power mode. The reason is after the cpu has been powered off, s/w need set the cpu's fpsimd_last_state to NULL so that it will force to reload FPSIMD state for the thread, otherwise there has the chance to meet the condition for both the task's fpsimd_state.cpu field contains the id of the current cpu, and the cpu's fpsimd_last_state per-cpu variable points to the task's fpsimd_state, so finally kernel will skip to reload the context during it return back to userland. Furthermore, this patch also removes the redundant operation for saving the state for idle thread; this operation is needless due kernel's idle thread will not use any VFP/NEON instruction. so that this also can benefit a bit for the latency during cpu's power off. Signed-off-by: Leo Yan <leoy@marvell.com> --- arch/arm64/kernel/fpsimd.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)