On Tue, Oct 08, 2019 at 07:57:05PM +0000, Jason Gunthorpe wrote:
> On Sun, Oct 06, 2019 at 06:51:36PM +0300, Leon Romanovsky wrote:
> > From: Erez Alfasi <ereza@xxxxxxxxxxxx>
> >
> > Introduce ODP diagnostic counters and count the following
> > per MR within IB/mlx5 driver:
> > 1) Page faults:
> > Total number of faulted pages.
> > 2) Page invalidations:
> > Total number of pages invalidated by the OS during all
> > invalidation events. The translations can be no longer
> > valid due to either non-present pages or mapping changes.
> > 3) Prefetched pages:
> > When prefetching a page, page fault is generated
> > in order to bring the page to the main memory.
> > The prefetched pages counter will be updated
> > during a page fault flow only if it was derived
> > from prefetching operation.
> >
> > Signed-off-by: Erez Alfasi <ereza@xxxxxxxxxxxx>
> > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > drivers/infiniband/hw/mlx5/mlx5_ib.h | 4 ++++
> > drivers/infiniband/hw/mlx5/odp.c | 18 ++++++++++++++++++
> > include/rdma/ib_verbs.h | 6 ++++++
> > 3 files changed, 28 insertions(+)
> >
> > diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> > index bf30d53d94dc..5aae05ebf64b 100644
> > +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> > @@ -585,6 +585,9 @@ struct mlx5_ib_dm {
> > IB_ACCESS_REMOTE_READ |\
> > IB_ZERO_BASED)
> >
> > +#define mlx5_update_odp_stats(mr, counter_name, value) \
> > + atomic64_add(value, &((mr)->odp_stats.counter_name))
> > +
> > struct mlx5_ib_mr {
> > struct ib_mr ibmr;
> > void *descs;
> > @@ -622,6 +625,7 @@ struct mlx5_ib_mr {
> > wait_queue_head_t q_leaf_free;
> > struct mlx5_async_work cb_work;
> > atomic_t num_pending_prefetch;
> > + struct ib_odp_counters odp_stats;
> > };
> >
> > static inline bool is_odp_mr(struct mlx5_ib_mr *mr)
> > diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
> > index 95cf0249b015..966783bfb557 100644
> > +++ b/drivers/infiniband/hw/mlx5/odp.c
> > @@ -261,6 +261,10 @@ void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start,
> > blk_start_idx = idx;
> > in_block = 1;
> > }
> > +
> > + /* Count page invalidations */
> > + mlx5_update_odp_stats(mr, invalidations,
> > + (idx - blk_start_idx + 1));
>
> I feel like these should be batched and the atomic done once at the
> end of the routine..
We can, but does it worth it?
>
> > } else {
> > u64 umr_offset = idx & umr_block_mask;
> >
> > @@ -287,6 +291,7 @@ void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start,
> >
> > ib_umem_odp_unmap_dma_pages(umem_odp, start, end);
> >
> > +
> > if (unlikely(!umem_odp->npages && mr->parent &&
> > !umem_odp->dying)) {
> > WRITE_ONCE(umem_odp->dying, 1);
> > @@ -801,6 +806,19 @@ static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
> > if (ret < 0)
> > goto srcu_unlock;
> >
> > + /*
> > + * When prefetching a page, page fault is generated
> > + * in order to bring the page to the main memory.
> > + * In the current flow, page faults are being counted.
> > + * Prefetched pages counter will be updated as well
> > + * only if the current page fault flow was derived
> > + * from prefetching flow.
> > + */
> > + mlx5_update_odp_stats(mr, faults, ret);
> > +
> > + if (prefetch)
> > + mlx5_update_odp_stats(mr, prefetched, ret);
>
> Hm, I'm about to post a series that eliminates 'prefetch' here..
Jason,
For various reasons we are delaying this series for months already.
Let's drop "prefetch" counter for now and merge everything without
it.
>
> This is also not quite right for prefetch as we are doing a form of
> prefetching in the mlx5_ib_mr_rdma_pfault_handler() too, although it
> is less clear how to count those. Maybe this should be split to SQ/RQ
> faults?
mlx5_ib_mr_rdma_pfault_handler() calls to pagefault_single_data_segment()
without MLX5_PF_FLAGS_PREFETCH, so I'm unsure that this counter should
count mlx5_ib_mr_rdma_pfault_handler() pagefaults.
However the idea to separate SQ/RQ for everything sounds appealing.
Thanks
>
> Jason
