On 05/01, Peter Zijlstra wrote:
>
> Anyway; I cobbled together the below. Oleg, could you have a look, I'm
> sure I messed it up.
Oh, I will need to read this carefully. but at first glance I do not see
any hole...
> +static void readers_block(struct percpu_rw_semaphore *sem)
> +{
> + wait_event_cmd(sem->writer, !sem->readers_block,
> + __up_read(&sem->rw_sem), __down_read(&sem->rw_sem));
> +}
> +
> +static void block_readers(struct percpu_rw_semaphore *sem)
> +{
> + wait_event_exclusive_cmd(sem->writer, !sem->readers_block,
> + __up_write(&sem->rw_sem),
> + __down_write(&sem->rw_sem));
> + /*
> + * Notify new readers to block; up until now, and thus throughout the
> + * longish rcu_sync_enter() above, new readers could still come in.
> + */
> + WRITE_ONCE(sem->readers_block, 1);
> +}
So iiuc, despite it name block_readers() also serializes the writers, ->rw_sem
can be dropped by down_write_non_owner() so the new writer can take this lock.
And note that the caller of readers_block() does down_read(), the caller of
block_readers() does down_write(). So perhaps it makes sense to shift these
down_read/write into the helpers above and rename them,
void xxx_down_read(struct percpu_rw_semaphore *sem)
{
__down_read(&sem->rw_sem);
wait_event_cmd(sem->writer, !sem->readers_block,
__up_read(&sem->rw_sem), __down_read(&sem->rw_sem));
}
void xxx_down_write(struct percpu_rw_semaphore *sem)
{
down_write(&sem->rw_sem);
wait_event_exclusive_cmd(sem->writer, !sem->readers_block,
__up_write(&sem->rw_sem),
__down_write(&sem->rw_sem));
/*
* Notify new readers to block; up until now, and thus throughout the
* longish rcu_sync_enter() above, new readers could still come in.
*/
WRITE_ONCE(sem->readers_block, 1);
}
to make this logic more clear? Or even
bool ck_read(struct percpu_rw_semaphore *sem)
{
__down_read(&sem->rw_sem);
if (!sem->readers_block)
return true;
__up_read(&sem->rw_sem);
}
bool ck_write(struct percpu_rw_semaphore *sem)
{
down_write(&sem->rw_sem);
if (!sem->readers_block)
return true;
up_write(&sem->rw_sem);
}
Then percpu_down_read/write can simply do wait_event(ck_read(sem)) and
wait_event_exclusive(ck_write(sem)) respectively.
But this all is cosmetic, it seems that we can remove ->rw_sem altogether
but I am not sure...
Oleg.
