On Tue, Oct 29, 2019 at 01:59:00PM +0300, Andrew Rybchenko wrote:
> On 10/28/19 5:01 PM, Olivier Matz wrote:
> > When populating a mempool, ensure that objects are not located across
> > several pages, except if user did not request iova contiguous objects.
>
> I think it breaks distribution across memory channels which could
> affect performance significantly.
With 2M hugepages, there are ~900 mbufs per page, and all of them will
be distributed across memory channels. For larger objects, I don't think
the distribution is that important.
With small pages, that may be true. I think the problem was already
there except in IOVA=VA mode.
This should be fixable, but I'm not sure a use-case where we can see a
regression really exists.
> > Signed-off-by: Vamsi Krishna Attunuru <vattun...@marvell.com>
> > Signed-off-by: Olivier Matz <olivier.m...@6wind.com>
> > ---
> > lib/librte_mempool/rte_mempool.c | 23 +++++-----------
> > lib/librte_mempool/rte_mempool_ops_default.c | 29 ++++++++++++++++++--
> > 2 files changed, 33 insertions(+), 19 deletions(-)
> >
> > diff --git a/lib/librte_mempool/rte_mempool.c
> > b/lib/librte_mempool/rte_mempool.c
> > index 7664764e5..b23fd1b06 100644
> > --- a/lib/librte_mempool/rte_mempool.c
> > +++ b/lib/librte_mempool/rte_mempool.c
> > @@ -428,8 +428,6 @@ rte_mempool_get_page_size(struct rte_mempool *mp,
> > size_t *pg_sz)
> > if (!need_iova_contig_obj)
> > *pg_sz = 0;
> > - else if (!alloc_in_ext_mem && rte_eal_iova_mode() == RTE_IOVA_VA)
> > - *pg_sz = 0;
> > else if (rte_eal_has_hugepages() || alloc_in_ext_mem)
> > *pg_sz = get_min_page_size(mp->socket_id);
> > else
> > @@ -478,17 +476,15 @@ rte_mempool_populate_default(struct rte_mempool *mp)
> > * then just set page shift and page size to 0, because the user has
> > * indicated that there's no need to care about anything.
> > *
> > - * if we do need contiguous objects, there is also an option to reserve
> > - * the entire mempool memory as one contiguous block of memory, in
> > - * which case the page shift and alignment wouldn't matter as well.
> > + * if we do need contiguous objects (if a mempool driver has its
> > + * own calc_size() method returning min_chunk_size = mem_size),
> > + * there is also an option to reserve the entire mempool memory
> > + * as one contiguous block of memory.
> > *
> > * if we require contiguous objects, but not necessarily the entire
> > - * mempool reserved space to be contiguous, then there are two options.
> > - *
> > - * if our IO addresses are virtual, not actual physical (IOVA as VA
> > - * case), then no page shift needed - our memory allocation will give us
> > - * contiguous IO memory as far as the hardware is concerned, so
> > - * act as if we're getting contiguous memory.
> > + * mempool reserved space to be contiguous, pg_sz will be != 0,
> > + * and the default ops->populate() will take care of not placing
> > + * objects across pages.
> > *
> > * if our IO addresses are physical, we may get memory from bigger
> > * pages, or we might get memory from smaller pages, and how much of it
> > @@ -501,11 +497,6 @@ rte_mempool_populate_default(struct rte_mempool *mp)
> > *
> > * If we fail to get enough contiguous memory, then we'll go and
> > * reserve space in smaller chunks.
> > - *
> > - * We also have to take into account the fact that memory that we're
> > - * going to allocate from can belong to an externally allocated memory
> > - * area, in which case the assumption of IOVA as VA mode being
> > - * synonymous with IOVA contiguousness will not hold.
> > */
> > need_iova_contig_obj = !(mp->flags & MEMPOOL_F_NO_IOVA_CONTIG);
> > diff --git a/lib/librte_mempool/rte_mempool_ops_default.c
> > b/lib/librte_mempool/rte_mempool_ops_default.c
> > index f6aea7662..dd09a0a32 100644
> > --- a/lib/librte_mempool/rte_mempool_ops_default.c
> > +++ b/lib/librte_mempool/rte_mempool_ops_default.c
> > @@ -61,21 +61,44 @@ rte_mempool_op_calc_mem_size_default(const struct
> > rte_mempool *mp,
> > return mem_size;
> > }
> > +/* Returns -1 if object crosses a page boundary, else returns 0 */
> > +static int
> > +check_obj_bounds(char *obj, size_t pg_sz, size_t elt_sz)
> > +{
> > + if (pg_sz == 0)
> > + return 0;
> > + if (elt_sz > pg_sz)
> > + return 0;
> > + if (RTE_PTR_ALIGN(obj, pg_sz) != RTE_PTR_ALIGN(obj + elt_sz - 1, pg_sz))
> > + return -1;
> > + return 0;
> > +}
> > +
> > int
> > rte_mempool_op_populate_default(struct rte_mempool *mp, unsigned int
> > max_objs,
> > void *vaddr, rte_iova_t iova, size_t len,
> > rte_mempool_populate_obj_cb_t *obj_cb, void *obj_cb_arg)
> > {
> > - size_t total_elt_sz;
> > + char *va = vaddr;
> > + size_t total_elt_sz, pg_sz;
> > size_t off;
> > unsigned int i;
> > void *obj;
> > + rte_mempool_get_page_size(mp, &pg_sz);
> > +
>
> The function may return an error which should be taken into account here.
That would be better, indeed.
> > total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
> > - for (off = 0, i = 0; off + total_elt_sz <= len && i < max_objs; i++) {
> > + for (off = 0, i = 0; i < max_objs; i++) {
> > + /* align offset to next page start if required */
> > + if (check_obj_bounds(va + off, pg_sz, total_elt_sz) < 0)
> > + off += RTE_PTR_ALIGN_CEIL(va + off, pg_sz) - (va + off);
> > +
> > + if (off + total_elt_sz > len)
> > + break;
> > +
> > off += mp->header_size;
> > - obj = (char *)vaddr + off;
> > + obj = va + off;
> > obj_cb(mp, obj_cb_arg, obj,
> > (iova == RTE_BAD_IOVA) ? RTE_BAD_IOVA : (iova + off));
> > rte_mempool_ops_enqueue_bulk(mp, &obj, 1);
>
Thanks for all the comments!
I will send a v2 tomorrow.
Olivier
