Re: [PATCH 14/32] aio: Make aio_read_evt() more efficient, convertto hrtimers

From: Kent Overstreet
Date: Mon Jan 07 2013 - 20:28:05 EST

On Mon, Jan 07, 2013 at 05:00:55PM -0800, Andrew Morton wrote:
> aio_read_events_ring() is called via the
> wait_event_interruptible_hrtimeout() macro's call to `condition' - to
> work out whether aio_read_events_ring() should terminate.
> A problem we should think about is "under what circumstances will
> aio_read_events_ring() set us into TASK_RUNNING?". We don't want
> aio_read_events_ring() to do this too often because it will cause
> schedule() to fall through and we end up in a busy loop, chewing CPU.
> afacit, aio_read_events_ring() will usually return non-zero if it
> flipped us into TASK_RUNNING state. An exception is where the
> mutex_trylock() failed, in which case the thread slept in mutex_lock(),
> whcih will help with the CPU-chewing. But aio_read_events_ring() can
> then end up returning 0 but in state TASK_RUNNING which will cause a
> small cpu-chew in wait_event_interruptible_hrtimeout().

Yeah, that was my reasoning too.

> I think :( It is unfortunately complex and it would be nice to make
> this dynamic behaviour more clear and solid. Or at least documented!
> Explain how this code avoid getting stuck in a cpu-burning loop. To
> help prevent people from causing a cpu-burning loop when they later
> change the code.


> > However - I was told that calling mutex_lock() in TASK_INTERRUPTIBLE
> > state was bad, but thinking about it more I'm not seeing how that's the
> > case. Either mutex_lock() finds the lock uncontended and doesn't touch
> > the task state, or it does and leaves it in TASK_RUNNING when it
> > returns.
> >
> > IOW, I don't see how it'd behave any differently from what I'd doing.
> >
> > Any light you could shed would be most appreciated.
> Well, the problem with running mutex_lock() in TASK_[UN]INTERRUPTIBLE
> is just that: it may or may not flip you into TASK_RUNNING, so what the
> heck is the caller thinking of? It's strange to set the task state a
> particular way, then call a function which will randomly go and undo
> that.
> The cause of all this is the wish to use a wait_event `condition'
> predicate which must take a mutex. hrm.

I've run into this problem before, and I've yet to come up with a
satisfactory solution. What we kind of want is just pthreads style
condition variables. Or something. I'm surprised this doesn't come up
more often.

But, this code has been through like 5 iterations (with Zach Brown
picking most of them apart) and I think this is the best we've come up
with. Trying to get the task state stuff exactly right led to it being
_much_ more contorted, I think.

Does the patch below help?

> > > IOW, I don't have the foggiest clue what you're trying to do here and
> > > you owe us all a code comment. At least.
> >
> > Yeah, will do.
> Excited!
> > This look better for the types?
> yup.
> Also, it's unclear why kioctx.shadow_tail exists. Some overviewy
> explanation at its definitions site is needed, IMO.

Ah, that's mostly just to reduce cacheline bouncing - in practice the
tail pointer that aio_complete() uses tends to be a lot more contended
than the head pointer, since events get delivered one at a time and then
pulled off all at once. So aio_complete() keeps it up to date and then
aio_read_events() doesn't have to compete for the tail cacheline.

commit ab92ba18a0a891821edd967c46dc988326ef6bb0
Author: Kent Overstreet <koverstreet@xxxxxxxxxx>
Date: Mon Jan 7 17:27:19 2013 -0800

aio: Document, clarify aio_read_events() and shadow_tail

Signed-off-by: Kent Overstreet <koverstreet@xxxxxxxxxx>

diff --git a/fs/aio.c b/fs/aio.c
index 21b2c27..932170a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -102,6 +102,19 @@ struct kioctx {
struct {
struct mutex ring_lock;
wait_queue_head_t wait;
+ /*
+ * Copy of the real tail, that aio_complete uses - to reduce
+ * cacheline bouncing. The real tail will tend to be much more
+ * contended - since typically events are delivered one at a
+ * time, and then aio_read_events() slurps them up a bunch at a
+ * time - so it's helpful if aio_read_events() isn't also
+ * contending for the tail. So, aio_complete() updates
+ * shadow_tail whenever it updates tail.
+ *
+ * Also needed because tail is used as a hacky lock and isn't
+ * always the real tail.
+ */
unsigned shadow_tail;
} ____cacheline_aligned_in_smp;

@@ -845,10 +858,7 @@ static long aio_read_events_ring(struct kioctx *ctx,
long ret = 0;
int copy_ret;

- if (!mutex_trylock(&ctx->ring_lock)) {
- __set_current_state(TASK_RUNNING);
- mutex_lock(&ctx->ring_lock);
- }
+ mutex_lock(&ctx->ring_lock);

ring = kmap_atomic(ctx->ring_pages[0]);
head = ring->head;
@@ -859,8 +869,6 @@ static long aio_read_events_ring(struct kioctx *ctx,
if (head == ctx->shadow_tail)
goto out;

- __set_current_state(TASK_RUNNING);
while (ret < nr) {
long avail = (head < ctx->shadow_tail
? ctx->shadow_tail : ctx->nr) - head;
@@ -939,6 +947,20 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
until = timespec_to_ktime(ts);

+ /*
+ * Note that aio_read_events() is being called as the conditional - i.e.
+ * we're calling it after prepare_to_wait() has set task state to
+ *
+ * But aio_read_events() can block, and if it blocks it's going to flip
+ * the task state back to TASK_RUNNING.
+ *
+ * This should be ok, provided it doesn't flip the state back to
+ * TASK_RUNNING and return 0 too much - that causes us to spin. That
+ * will only happen if the mutex_lock() call blocks, and we then find
+ * the ringbuffer empty. So in practice we should be ok, but it's
+ * something to be aware of when touching this code.
+ */
aio_read_events(ctx, min_nr, nr, event, &ret), until);

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at