Re: [RFC PATCH v2 1/2] printk-rb: add a new printk ringbuffer implementation

From: Peter Zijlstra
Date: Fri Jun 28 2019 - 11:45:35 EST


On Fri, Jun 28, 2019 at 11:50:33AM +0200, John Ogness wrote:
> On 2019-06-27, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> >>>> +static void add_descr_list(struct prb_reserved_entry *e)
> >>>> +{
> >>>> + struct printk_ringbuffer *rb = e->rb;
> >>>> + struct prb_list *l = &rb->descr_list;
> >>>> + struct prb_descr *d = e->descr;
> >>>> + struct prb_descr *newest_d;
> >>>> + unsigned long newest_id;
> >>>> +
> >>>> + /* set as newest */
> >>>> + do {
> >>>> + /* MB5: synchronize add descr */
> >>>> + newest_id = smp_load_acquire(&l->newest);
> >>>> + newest_d = TO_DESCR(rb, newest_id);
> >>>> +
> >>>> + if (newest_id == EOL)
> >>>> + WRITE_ONCE(d->seq, 1);
> >>>> + else
> >>>> + WRITE_ONCE(d->seq, READ_ONCE(newest_d->seq) + 1);
> >>>> + /*
> >>>> + * MB5: synchronize add descr
> >>>> + *
> >>>> + * In particular: next written before cmpxchg
> >>>> + */
> >>>> + } while (cmpxchg_release(&l->newest, newest_id, e->id) != newest_id);
> >>>
> >>> What does this pair with? I find ->newest usage in:
> >>
> >> It is pairing with the smp_load_acquire() at the beginning of this
> >> loop (also labeled MB5) that is running simultaneously on another
> >> CPU. I am avoiding a possible situation that a new descriptor is
> >> added but the store of "next" from the previous descriptor is not yet
> >> visible and thus the cmpxchg following will fail, which is not
> >> allowed. (Note that "next" is set to EOL shortly before this function
> >> is called.)
> >>
> >> The litmus test for this is:
> >>
> >> P0(int *newest, int *d_next)
> >> {
> >> // set descr->next to EOL (terminates list)
> >> WRITE_ONCE(*d_next, 1);
> >>
> >> // set descr as newest
> >> smp_store_release(*newest, 1);
> >> }
> >>
> >> P1(int *newest, int *d_next)
> >> {
> >> int local_newest;
> >> int local_next;
> >>
> >> // get newest descriptor
> >> local_newest = smp_load_acquire(*newest);
> >>
> >> // a new descriptor is set as the newest
> >> // (not relevant here)
> >>
> >> // read descr->next of previous newest
> >> // (must be EOL!)
> >> local_next = READ_ONCE(*d_next);
> >> }
> >>
> >> exists (1:local_newest=1 /\ 1:local_next=0)
> >
> > I'm having trouble connecting your P1's READ_ONCE() to the actual
> > code. You say that is in the same function, but I cannot find a LOAD
> > there that would care about the ACQUIRE.
>
> P1's READ_ONCE() is the READ part of the cmpxchg a few lines below:
>
> WARN_ON_ONCE(cmpxchg_release(&newest_d->next,
> EOL, e->id) != EOL);
>
> Note that the cmpxchg is a _release because of MB6 (a different memory
> barrier pair). But only the READ part of that cmpxchg synchronizes with
> MB5.
>
> Also note that cmpxchg is used only because of bug checking. If instead
> it becomes a blind store (such as you suggest below), then it changes to
> smp_store_release().
>
> While investigating this (and the lack of a LOAD), I realized that the
> smp_load_acquire() is not needed because @seq is dependent on the load
> of @newest.

That!

> I have implemented and tested these changes. I also added
> setting the list terminator to this function, since all callers would
> have to do it anyway. Also, I spent a lot of time trying to put in
> comments that I think are _understandable_ and _acceptable_.
>
> @Peter: I expect they are way too long for you.
>
> @Andrea: Is this starting to become something that you would like to
> see?
>
> /**
> * add_descr_list() - Add a descriptor to the descriptor list.
> *
> * @e: An entry that has already reserved data.
> *
> * The provided entry contains a pointer to a descriptor that has already
> * been reserved for this entry. However, the reserved descriptor is not
> * yet on the list. Add this descriptor as the newest item.
> *
> * A descriptor is added in two steps. The first step is to make this
> * descriptor the newest. The second step is to update @next of the former
> * newest descriptor to point to this one (or set @oldest to this one if
> * this will be the first descriptor on the list).
> */

I still think it might be useful to explicitly call out the data
structure more. Even if you cannot use a fully abtracted queue.

Also, newest/oldest just looks weird to me; I'm expecting head/tail.

> static void add_descr_list(struct prb_reserved_entry *e)
> {
> struct printk_ringbuffer *rb = e->rb;
> struct prb_list *l = &rb->descr_list;
> struct prb_descr *d = e->descr;
> struct prb_descr *newest_d;
> unsigned long newest_id;
>
> WRITE_ONCE(d->next, EOL);
>
> do {
> newest_id = READ_ONCE(l->newest);
> newest_d = TO_DESC(rb, newest_id);
>
> if (newest_id == EOL) {
> WRITE_ONCE(d->seq, 1);
> } else {
> /*
> * MB5-read: synchronize setting newest descr
> *
> * context-pair: 2 writers adding a descriptor via
> * add_descr_list().
> *
> * @newest will load before @seq due to a data
> * dependency, therefore, the stores of @seq
> * and @next from the pairing MB5-write context
> * will be visible.
> *
> * Although @next is not loaded by this context,
> * this context must overwrite the stored @next
> * value of the pairing MB5-write context.
> */
> WRITE_ONCE(d->seq, READ_ONCE(newest_d->seq) + 1);
> }
>
> /*
> * MB5-write: synchronize setting newest descr
> *
> * context-pair: 2 writers adding a descriptor via
> * add_descr_list().
> *
> * Ensure that @next and @seq are stored before @d is
> * visible via @newest. The pairing MB5-read context
> * must load this @seq value and must overwrite this
> * @next value.
> */
> } while (cmpxchg_release(&l->newest, newest_id, e->id) != newest_id);
>
> if (unlikely(newest_id == EOL)) {
> /*
> * MB0-write: synchronize adding first descr
> *
> * context-pair: 1 writer adding the first descriptor via
> * add_descr_list(), 1 reader getting the beginning of
> * the list via iter_peek_next_id().
> *
> * This context recently assigned new values for @id,
> * @next, @seq. Ensure these are stored before the first
> * store to @oldest so that the new values are visible
> * to the reader in the pairing MB0-read context.
> *
> * Note: Before this store, the value of @oldest is EOL.
> */
> smp_store_release(&l->oldest, e->id);
> } else {
> /*
> * MB6-write: synchronize linking new descr
> *
> * context-pair-1: 1 writer adding a descriptor via
> * add_descr_list(), 1 writer removing a descriptor via
> * remove_oldest_descr().
> *
> * If this is a recycled descriptor, this context
> * recently stored a new @oldest value. Ensure that
> * @oldest is stored before storing @next so that
> * if the pairing MB6-read context sees a non-EOL
> * @next value, it is ensured that it will also see
> * an updated @oldest value.
> *
> * context-pair-2: 1 writer adding a descriptor via
> * add_descr_list(), 1 reader iterating the list via
> * prb_iter_next_valid_entry().
> *
> * This context recently assigned new values for @id,
> * @next, @seq, @data, @data_next. Ensure these are
> * stored before storing @next of the previously
> * newest descriptor so that the new values are
> * visible to the iterating reader in the pairing
> * MB6-read context.
> *
> * Note: Before this store, the value of @next of the
> * previously newest descriptor is EOL.
> */
> smp_store_release(&newest_d->next, e->id);
> }
> }
>
> The smp_rmb() calls in the reader functions are then commented and
> marked with the appropriate MB0-read and MB6-read labels.
>
> > Afaict prb_list is a list head not a list node (calling it just _list
> > is confusing at best).
>
> OK.
>
> > You have a single linked list going from the tail to the head, while
> > adding to the head and removing from the tail. And that sounds like a
> > FIFO queue:
>
> Yes, but with one important feature: the nodes in the FIFO queue are
> labeled with ordered sequence numbers. This is important for printk. I
> talk more about this below.

But nowhere did/do you say what the actual data structure is, with what
modification for which reason.

> > struct lqueue_head {
> > struct lqueue_node *head, *tail;
> > };
> >
> > struct lqueue_node {
> > struct lqueue_node *next;
> > };
> >
> > void lqueue_push(struct lqueue_head *h, struct lqueue_node *n)
> > {
> > struct lqueue_node *prev;
> >
> > n->next = NULL;
>
> Is this safe? Do all compilers understand that @next must be stored
> before the xchg() of @head? I would have chosen WRITE_ONCE().

Yep, xchg() implies an smp_mb() before and after, smp_mb() in turn
implies a compiler barrier. Even if there is compiler induced brain
damage (store-tearing) all that must be done before the actual RmW.

Same with xchg_release(), the RELEASE is sufficient to have all previous
stores complete before the RmW.

> > /*
> > * xchg() implies RELEASE; and thereby ensures @n is
> > * complete before getting published.
> > */
> > prev = xchg(&h->head, n);
>
> Unfortunately it is not that simple because of sequence numbers. A node
> must be assigned a sequence number that is +1 of the previous node. This
> must be done before exchanging the head because immediately after the
> xchg() on the head, another CPU could then add on to us and expects our
> sequence number to already be set.
>
> This is why I need cmpxchg() here.

So far that doens't make sense yet, +1 is implicit in the list order
surely. But yes, if you need the seq like that, then cmpxchg it is.

> > /*
> > * xchg() implies ACQUIRE; and thereby ensures @tail is
> > * written after @head, see lqueue_pop()'s smp_rmb().
> > */
> > if (prev)
> > WRITE_ONCE(prev->next, n);
>
> This needs to be a store_release() so that a reader cannot read @n but
> the store to @next is not yet visible. The memory barriers of the above
> xchg() do not apply here because readers never read @head.

Confused, all stores to @n are before the xchg() so the barrier from
xchg() also order those stores and this store.

> > else
> > WRITE_ONCE(h->tail, n);
>
> Ditto, but for the tail node in particular.
>
> > }
> >
> > struct lqueue_node *lqueue_pop(struct lqueue_head *h)
> > {
> > struct lqueue_node *head, *tail, *next;
> >
> > do {
> > tail = READ_ONCE(h->tail);
> > /* If the list is empty, nothing to remove. */
> > if (!tail)
> > return NULL;
> >
> > /*
> > * If we see @tail, we must then also see @head.
> > * Pairs with the xchg() in lqueue_push(),
> > * ensure no false positive on the singleton
> > * test below.
> > */
> > smp_rmb();
> > head = READ_ONCE(h->head);
> >
> > /* If there is but one item; fail to remove. */
> > if (head == tail)
> > return NULL;
> >
> > next = smp_cond_load_relaxed(&tail->next, VAL);
>
> What if a writer is adding a 2nd node to the queue and is interrupted by
> an NMI directly after the xchg() in lqueue_push()? Then we have:
>
> * head != tail
> * tail->next == NULL
>
> If that interrupting NMI calls lqueue_pop(), the NMI will spin
> forever. The following cmpxchg() is not allowed to happen as long as
> tail->next is NULL.

Indeed. I forgot that you actually use pop on the producer side.

(Note that the qspinlock has a queue not unlike this, but that again
doesn't have to bother with NMIs)

> This is why I synchronize on @next instead, using (tail && !tail->next)
> for the singleton test.

OK.

> > } while (cmpxchg(h->tail, tail, next) != tail);
> >
> > return tail;
> > }
> >
> > Now, you appear to be using desc_ids instead of pointers, but since
> > you're not using the actual wrap value; I don't see the benefit of
> > using those IDs over straight pointers.
>
> The documentation mentions that descriptor ids are used to identify
> pointers to invalid descriptors. This is used by the readers, see
> iter_peek_next_id() and prb_iter_next_valid_entry().
>
> IDs are used for:
>
> - @next of descriptors on the list
> - @id, @id_next in the reader iterator
> - @id in the data blocks
>
> If changed to pointers, iterators would need to additionally store @seq
> values to be able to identifiy if the entry they are pointing to is the
> entry they expect.
>
> The only advantage I see with pointers is that the ringbuffer could be
> more useful generally, independent of whether the data is separate or
> within the nodes or if the nodes are statically or dynamically
> allocated. That is something worth having, even if it is not printk
> related.
>
> Are you implicitly requesting me to split the prb_ringbuffer and instead
> base it on a new "lockless multi-writer multi-reader sequenced FIFO
> queue" data structure?

Not specifically; I was just trying to untangle the code and found a
queue. I still (sorry!) haven't gotten through the lot of it to see how
all the parts fit together.

> > That is, unless I've overlooked some subtle ABA issue, but then, your
> > code doesn't seem to mention that, and I think we're good because if
> > we re-use an entry, it can never get back in the same location, since
> > we never allow an empty list
>
> I do not understand what you mean here. If a reader has a pointer to an
> entry, the entry behind that pointer can certainly change. But that
> isn't a problem. The reader will recognize that.

ABA is where a cmpxchg has a false positive due to values matching but
not the structure.

For example, in the above pop, if h->tail would (again) be @tail, but
@next would not be the correct value. Something like that could happen
if before the cmpxchg a concurrent pop takes the element off and then
sticks it back on, but with a different ->next.

Then our cmpxchg wil succeed and corrupt.

> > (might also be fixable, haven't tought too hard on this).
>
> :-)
>
> > That said, the above has cmpxchg() vs WRITE_ONCE() and is therefore
> > not safe on a number of our architectures. We can either not care
> > about performance and use xchg() for the ->tail store, or use
> > atomic_long_t and suffer ugly casting.
>
> cmpxchg_release() vs WRITE_ONCE() is not safe?! Can you point me to
> documentation about this?

Documentation/atomic_t.txt has this, see the SEMANTICS section on
atomic-set.