On Wed, Jan 22, 2020 at 3:07 PM Leon Romanovsky <leon@xxxxxxxxxx> wrote:
>
> On Wed, Jan 22, 2020 at 02:12:19PM +0100, Jinpu Wang wrote:
> > On Wed, Jan 22, 2020 at 1:25 PM Leon Romanovsky <leon@xxxxxxxxxx> wrote:
> > >
> > > On Wed, Jan 22, 2020 at 12:22:43PM +0100, Jinpu Wang wrote:
> > > > > > +/**
> > > > > > + * rnbd_get_cpu_qlist() - finds a list with HW queues to be rerun
> > > > > > + * @sess: Session to find a queue for
> > > > > > + * @cpu: Cpu to start the search from
> > > > > > + *
> > > > > > + * Description:
> > > > > > + * Each CPU has a list of HW queues, which needs to be rerun. If a list
> > > > > > + * is not empty - it is marked with a bit. This function finds first
> > > > > > + * set bit in a bitmap and returns corresponding CPU list.
> > > > > > + */
> > > > > > +static struct rnbd_cpu_qlist *
> > > > > > +rnbd_get_cpu_qlist(struct rnbd_clt_session *sess, int cpu)
> > > > > > +{
> > > > > > + int bit;
> > > > > > +
> > > > > > + /* First half */
> > > > > > + bit = find_next_bit(sess->cpu_queues_bm, nr_cpu_ids, cpu);
> > > > >
> > > > > Is it protected by any lock?
> > > > We hold requeue_lock when set/clear bit, and disable preemption via
> > > > get_cpu_ptr when find_next_bit.
> > > > even it fails to get latest bit, it just cause an rerun the queue.
> > >
> > > It is not clear here at all.
> > >
> > > > >
> > > > > > + if (bit < nr_cpu_ids) {
> > > > > > + return per_cpu_ptr(sess->cpu_queues, bit);
> > > > > > + } else if (cpu != 0) {
> > > > > > + /* Second half */
> > > > > > + bit = find_next_bit(sess->cpu_queues_bm, cpu, 0);
> > > > > > + if (bit < cpu)
> > > > > > + return per_cpu_ptr(sess->cpu_queues, bit);
> > > > > > + }
> > > > > > +
> > > > > > + return NULL;
> > > > > > +}
> > > > > > +
> > > > > > +static inline int nxt_cpu(int cpu)
> > > > > > +{
> > > > > > + return (cpu + 1) % nr_cpu_ids;
> > > > > > +}
> > > > > > +
> > > > > > +/**
> > > > > > + * rnbd_rerun_if_needed() - rerun next queue marked as stopped
> > > > > > + * @sess: Session to rerun a queue on
> > > > > > + *
> > > > > > + * Description:
> > > > > > + * Each CPU has it's own list of HW queues, which should be rerun.
> > > > > > + * Function finds such list with HW queues, takes a list lock, picks up
> > > > > > + * the first HW queue out of the list and requeues it.
> > > > > > + *
> > > > > > + * Return:
> > > > > > + * True if the queue was requeued, false otherwise.
> > > > > > + *
> > > > > > + * Context:
> > > > > > + * Does not matter.
> > > > > > + */
> > > > > > +static inline bool rnbd_rerun_if_needed(struct rnbd_clt_session *sess)
> > > > >
> > > > > No inline function in C files.
> > > > First time saw such request, there are so many inline functions in C
> > >
> > > 15) The inline disease
> > > https://elixir.bootlin.com/linux/latest/source/Documentation/process/coding-style.rst#L882
> > ok, will review the inline usage, and drop some.
> > >
> > > > files across the tree
> > > > grep inline drivers/infiniband/core/*.c
> > > > drivers/infiniband/core/addr.c:static inline bool
> > > > ib_nl_is_good_ip_resp(const struct nlmsghdr *nlh)
> > > > drivers/infiniband/core/cma.c:static inline u8 cma_get_ip_ver(const
> > > > struct cma_hdr *hdr)
> > > > drivers/infiniband/core/cma.c:static inline void cma_set_ip_ver(struct
> > > > cma_hdr *hdr, u8 ip_ver)
> > > > drivers/infiniband/core/cma.c:static inline void release_mc(struct kref *kref)
> > > > drivers/infiniband/core/cma.c:static inline struct sockaddr
> > > > *cma_src_addr(struct rdma_id_private *id_priv)
> > > > drivers/infiniband/core/cma.c:static inline struct sockaddr
> > > > *cma_dst_addr(struct rdma_id_private *id_priv)
> > > >
> > > > >
> > > > > > +{
> > > > > > + struct rnbd_queue *q = NULL;
> > > > > > + struct rnbd_cpu_qlist *cpu_q;
> > > > > > + unsigned long flags;
> > > > > > + int *cpup;
> > > > > > +
> > > > > > + /*
> > > > > > + * To keep fairness and not to let other queues starve we always
> > > > > > + * try to wake up someone else in round-robin manner. That of course
> > > > > > + * increases latency but queues always have a chance to be executed.
> > > > > > + */
> > > > > > + cpup = get_cpu_ptr(sess->cpu_rr);
> > > > > > + for (cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(*cpup)); cpu_q;
> > > > > > + cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(cpu_q->cpu))) {
> > > > > > + if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
> > > > > > + continue;
> > > > > > + if (likely(test_bit(cpu_q->cpu, sess->cpu_queues_bm))) {
> > > > >
> > > > > Success oriented approach please.
> > > > sorry, I don't quite get your point.
> > >
> > > The flows are better to be written:
> > > if (err)
> > > return or conitnue
> > > <...>
> > > do_something
> > >
> > > in your case
> > > if (!test_bit(...))
> > > continue;
> > > do_work_here.
> > In our case,
> > if we failed to get requeue_lock, we continue to next cpu_q and do the work
> > I guess you miss read the code.
>
> I don't think so, this is is how it is expected to be.
>
> + if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
> + continue;
> + if (!test_bit(cpu_q->cpu, sess->cpu_queues_bm))
> + goto unlock;
> +
> + q = list_first_entry_or_null(&cpu_q->requeue_list,
> + typeof(*q), requeue_list);
> + if (!q)
> + goto clear_bit;
> + list_del_init(&q->requeue_list);
> + clear_bit_unlock(0, &q->in_list);
> ....
>
>
> >
> > Thanks
Got it, I will prepare a patch and test it with our regression test
Thanks
