On Tue, Mar 3, 2020 at 5:04 PM Bart Van Assche <bvanassche@xxxxxxx> wrote:
>
> On 3/2/20 5:20 AM, Danil Kipnis wrote:
> > On Sun, Mar 1, 2020 at 2:33 AM Bart Van Assche <bvanassche@xxxxxxx> wrote:
> >> On 2020-02-21 02:47, Jack Wang wrote:
> >>> +static struct rtrs_permit *
> >>> +__rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
> >>> +{
> >>> + size_t max_depth = clt->queue_depth;
> >>> + struct rtrs_permit *permit;
> >>> + int cpu, bit;
> >>> +
> >>> + /* Combined with cq_vector, we pin the IO to the the cpu it comes */
> >>
> >> This comment is confusing. Please clarify this comment. All I see below
> >> is that preemption is disabled. I don't see pinning of I/O to the CPU of
> >> the caller.
> > The comment is addressing a use-case of the driver: The user can
> > assign (under /proc/irq/) the irqs of the HCA cq_vectors "one-to-one"
> > to each cpu. This will "force" the driver to process io response on
> > the same cpu the io has been submitted on.
> > In the code below only preemption is disabled. This can lead to the
> > situation that callers from different cpus will grab the same bit,
> > since find_first_zero_bit is not atomic. But then the
> > test_and_set_bit_lock will fail for all the callers but one, so that
> > they will loop again. This way an explicit spinlock is not required.
> > Will extend the comment.
>
> If the purpose of get_cpu() and put_cpu() calls is to serialize code
> against other threads, please use locking instead of disabling
> preemption. This will help tools that verify locking like lockdep and
> the kernel thread sanitizer (https://github.com/google/ktsan/wiki).
We can look into it, but I'm afraid converting to spinlock might have
a performance impact.
>
> >>> +static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
> >>> + struct rtrs_clt_io_req *req,
> >>> + struct rtrs_rbuf *rbuf, u32 off,
> >>> + u32 imm, struct ib_send_wr *wr)
> >>> +{
> >>> + struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
> >>> + enum ib_send_flags flags;
> >>> + struct ib_sge sge;
> >>> +
> >>> + if (unlikely(!req->sg_size)) {
> >>> + rtrs_wrn(con->c.sess,
> >>> + "Doing RDMA Write failed, no data supplied\n");
> >>> + return -EINVAL;
> >>> + }
> >>> +
> >>> + /* user data and user message in the first list element */
> >>> + sge.addr = req->iu->dma_addr;
> >>> + sge.length = req->sg_size;
> >>> + sge.lkey = sess->s.dev->ib_pd->local_dma_lkey;
> >>> +
> >>> + /*
> >>> + * From time to time we have to post signalled sends,
> >>> + * or send queue will fill up and only QP reset can help.
> >>> + */
> >>> + flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
> >>> + 0 : IB_SEND_SIGNALED;
> >>> +
> >>> + ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
> >>> + req->sg_size, DMA_TO_DEVICE);
> >>> +
> >>> + return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
> >>> + rbuf->rkey, rbuf->addr + off,
> >>> + imm, flags, wr);
> >>> +}
> >>
> >> I don't think that posting a signalled send from time to time is
> >> sufficient to prevent send queue overflow. Please address Jason's
> >> comment from January 7th: "Not quite. If the SQ depth is 16 and you post
> >> 16 things and then signal the last one, you *cannot* post new work until
> >> you see the completion. More SQ space *ONLY* becomes available upon
> >> receipt of a completion. This is why you can't have an unsignaled SQ."
> >
> >> See also https://lore.kernel.org/linux-rdma/20200107182528.GB26174@xxxxxxxx/
> > In our case we set the send queue of each QP belonging to one
> > "session" to the one supported by the hardware (max_qp_wr) which is
> > around 5K on our hardware. The queue depth of our "session" is 512.
> > Those 512 are "shared" by all the QPs (number of CPUs on client side)
> > belonging to that session. So we have at most 512 and 512/num_cpus on
> > average inflights on each QP. We never experienced send queue full
> > event in any of our performance tests or production usage. The
> > alternative would be to count submitted requests and completed
> > requests, check the difference before submission and wait if the
> > difference multiplied by the queue depth of "session" exceeds the max
> > supported by the hardware. The check will require quite some code and
> > will most probably affect performance. I do not think it is worth it
> > to introduce a code path which is triggered only on a condition which
> > is known to never become true.
> > Jason, do you think it's necessary to implement such tracking?
>
> Please either make sure that send queues do not overflow by providing
> enough space for 512 in-flight requests fit or implement tracking for
> the number of in-flight requests.
We do have enough space for send queue.
>
> Thanks,
>
> Bart.
>
Thanks Bart!
