Re: [PATCH 1/2] sched/membarrier: Introduce MEMBARRIER_CMD_GET_REGISTRATIONS

From: Michał Cłapiński
Date: Wed Dec 07 2022 - 13:04:18 EST


On Wed, Dec 7, 2022 at 6:07 PM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> On 2022-12-07 11:43, Michal Clapinski wrote:
> > Provide a method to query previously issued registrations.
> >
> > Signed-off-by: Michal Clapinski <mclapinski@xxxxxxxxxx>
> > ---
> > include/uapi/linux/membarrier.h | 4 ++++
> > kernel/sched/membarrier.c | 39 ++++++++++++++++++++++++++++++++-
> > 2 files changed, 42 insertions(+), 1 deletion(-)
> >
> > diff --git a/include/uapi/linux/membarrier.h b/include/uapi/linux/membarrier.h
> > index 737605897f36..5f3ad6d5be6f 100644
> > --- a/include/uapi/linux/membarrier.h
> > +++ b/include/uapi/linux/membarrier.h
> > @@ -137,6 +137,9 @@
> > * @MEMBARRIER_CMD_SHARED:
> > * Alias to MEMBARRIER_CMD_GLOBAL. Provided for
> > * header backward compatibility.
> > + * @MEMBARRIER_CMD_GET_REGISTRATIONS:
> > + * Returns a bitmask of previously issued
> > + * registration commands.
> > *
> > * Command to be passed to the membarrier system call. The commands need to
> > * be a single bit each, except for MEMBARRIER_CMD_QUERY which is assigned to
> > @@ -153,6 +156,7 @@ enum membarrier_cmd {
> > MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE = (1 << 6),
> > MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ = (1 << 7),
> > MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ = (1 << 8),
> > + MEMBARRIER_CMD_GET_REGISTRATIONS = (1 << 9),

Btw. I could do this as a flag to MEMBARRIER_CMD_QUERY instead of a
separate command. Would that be preferable?


> >
> > /* Alias for header backward compatibility. */
> > MEMBARRIER_CMD_SHARED = MEMBARRIER_CMD_GLOBAL,
> > diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c
> > index 0c5be7ebb1dc..2ad881d07752 100644
> > --- a/kernel/sched/membarrier.c
> > +++ b/kernel/sched/membarrier.c
> > @@ -159,7 +159,8 @@
> > | MEMBARRIER_CMD_PRIVATE_EXPEDITED \
> > | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED \
> > | MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK \
> > - | MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK)
> > + | MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK \
> > + | MEMBARRIER_CMD_GET_REGISTRATIONS)
> >
> > static void ipi_mb(void *info)
> > {
> > @@ -540,6 +541,40 @@ static int membarrier_register_private_expedited(int flags)
> > return 0;
> > }
> >
> > +static int membarrier_get_registrations(void)
> > +{
> > + struct task_struct *p = current;
> > + struct mm_struct *mm = p->mm;
> > + int registrations_mask = 0, membarrier_state, i;
> > + static const int states[] = {
> > + MEMBARRIER_STATE_GLOBAL_EXPEDITED |
> > + MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY,
>
> What is the purpose of checking for the _READY state flag as well here ?

Answered below.


>
>
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED |
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY,
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE |
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY,
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ |
> > + MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY
> > + };
> > + static const int registration_cmds[] = {
> > + MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED,
> > + MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED,
> > + MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE,
> > + MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ
> > + };
> > + BUILD_BUG_ON(ARRAY_SIZE(states) != ARRAY_SIZE(registration_cmds));
> > +
> > + membarrier_state = atomic_read(&mm->membarrier_state);
> > + for (i = 0; i < ARRAY_SIZE(states); ++i) {
> > + if (membarrier_state & states[i]) {
>
> The mask will match if either of the flags to match are set. Is that
> your intent ?

Kind of, it was just the easiest to write. As explained in the cover
letter, I don't really care much about the result of this while the
process is running. And when the process is frozen, either state and
state_ready are set or none of them.


>
>
> > + registrations_mask |= registration_cmds[i];
> > + membarrier_state &= ~states[i];
>
> So I understand that those _READY flags are there purely for making sure
> we clear the membarrier_state for validation that they have all been
> checked with the following WARN_ON_ONCE(). Am I on the right track ?

Yes, exactly. It wastes time but I'm worried about people adding new
states and not updating this function. A suggestion on how to do this
better (especially at compile time) would be greatly appreciated.


>
> > + }
> > + }
> > + WARN_ON_ONCE(membarrier_state != 0);
>
> Thanks,
>
> Mathieu
>
> > + return registrations_mask;
> > +}
> > +
> > /**
> > * sys_membarrier - issue memory barriers on a set of threads
> > * @cmd: Takes command values defined in enum membarrier_cmd.
> > @@ -623,6 +658,8 @@ SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id)
> > return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id);
> > case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ:
> > return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ);
> > + case MEMBARRIER_CMD_GET_REGISTRATIONS:
> > + return membarrier_get_registrations();
> > default:
> > return -EINVAL;
> > }
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com
>