[v3,0/7] mempool: avoid objects allocations across pages

Message ID 20191104151254.6354-1-olivier.matz@6wind.com (mailing list archive)


Olivier Matz Nov. 4, 2019, 3:12 p.m. UTC
  KNI supposes that mbufs are contiguous in kernel virtual memory. This
may not be true when using the IOVA=VA mode. To fix this, a possibility
is to ensure that objects do not cross page boundaries in mempool. This
patchset implements this in the last patch (5/5).

The previous patches prepare the job:
- allow to populate with an unaligned virtual area (1/5).
- reduce spaced wasted in the mempool size calculation when not using
  the iova-contiguous allocation (2/5).
- remove the iova-contiguous allocation when populating mempool (3/5):
  a va-contiguous alloc does the job as well if we want to populate
  without crossing page boundaries, so simplify the mempool populate
- export a function to get the minimum page used in a mempool (4/5)

Memory consumption impact when using hugepages:
- worst case: + ~0.1% for a mbuf pool (objsize ~= 2368)
- best case: -50% for if pool size is just above page size

The memory consumption impact with 4K pages in IOVA=VA mode could
however consume up to 75% more memory for mbuf pool, because there will
be only 1 mbuf per page. Not sure how common this usecase is.

Caveat: this changes the behavior of the mempool (calc_mem_size and
populate), and there is a small risk to break things, especially with
alternate mempool drivers.


* introduce new helpers to calculate required memory size and to
  populate mempool, use them in drivers: the alignment constraint
  of octeontx/octeontx2 is managed in this common code.
* fix octeontx mempool driver by taking alignment constraint in account
  like in octeontx2
* fix bucket mempool driver with 4K pages: limit bucket size in this
  case to ensure that objects do not cross page boundaries. With larger
  pages, it was already ok, because bucket size (64K) is smaller than
  a page.
* fix some api comments in mempool header file


* update octeontx2 driver to keep alignment constraint (issue seen by
* add a new patch to use RTE_MEMPOOL_ALIGN (Andrew)
* fix initialization of for loop in rte_mempool_populate_virt() (Andrew)
* use rte_mempool_populate_iova() if mz_flags has
* check rte_mempool_get_page_size() return value (Andrew)
* some other minor style improvements

rfc -> v1

* remove first cleanup patch, it was pushed separately
  a2b5a8722f20 ("mempool: clarify default populate function")
* add missing change in rte_mempool_op_calc_mem_size_default()
* allow unaligned addr/len in populate virt
* better split patches
* try to better explain the change
* use DPDK align macros when relevant

Olivier Matz (7):
  mempool: allow unaligned addr/len in populate virt
  mempool: reduce wasted space on mempool populate
  mempool: remove optimistic IOVA-contiguous allocation
  mempool: introduce function to get mempool page size
  mempool: introduce helpers for populate and calc mem size
  mempool: prevent objects from being across pages
  mempool: use the specific macro for object alignment

 drivers/mempool/bucket/Makefile               |   2 +
 drivers/mempool/bucket/meson.build            |   3 +
 drivers/mempool/bucket/rte_mempool_bucket.c   |  10 +-
 drivers/mempool/dpaa/dpaa_mempool.c           |   4 +-
 drivers/mempool/dpaa2/dpaa2_hw_mempool.c      |   4 +-
 drivers/mempool/octeontx/Makefile             |   3 +
 drivers/mempool/octeontx/meson.build          |   3 +
 .../mempool/octeontx/rte_mempool_octeontx.c   |  21 +--
 drivers/mempool/octeontx2/Makefile            |   6 +
 drivers/mempool/octeontx2/meson.build         |   6 +
 drivers/mempool/octeontx2/otx2_mempool_ops.c  |  21 ++-
 lib/librte_mempool/rte_mempool.c              | 147 +++++++-----------
 lib/librte_mempool/rte_mempool.h              |  64 ++++++--
 lib/librte_mempool/rte_mempool_ops.c          |   4 +-
 lib/librte_mempool/rte_mempool_ops_default.c  | 113 +++++++++++---
 lib/librte_mempool/rte_mempool_version.map    |   3 +
 16 files changed, 272 insertions(+), 142 deletions(-)