Skip to content

RFC net: huge page backed page_pool

Jakub Kicinski netdev 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 RFC Latest12 patches 20 repliesIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/12net: hack together some page sharinginclude/net/dcalloc.h net/core/Makefile net/core/dcalloc.c +1 more+502 -10 replies
2/12net: create a 1G-huge-page-backed allocatorarch/x86/kernel/setup.c include/net/dcalloc.h net/core/dcalloc.c +1 more+243 -10 replies
3/12net: page_pool: hide page_pool_release_page()Documentation/networking/page_pool.rst drivers/net/ethernet/engleder/tsnep_main.c drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 more+9 -200 replies
4/12net: page_pool: merge page_pool_release_page() with page_pool_return_page()net/core/page_pool.c+2 -101 reply
5/12net: page_pool: factor out releasing DMA from releasing the pagenet/core/page_pool.c+16 -90 replies
6/12net: page_pool: create hooks for custom page providersinclude/net/page_pool.h net/core/page_pool.c+97 -32 replies
7/12net: page_pool: add huge page backed memory providersinclude/net/page_pool.h net/core/page_pool.c+517 -00 replies
8/12eth: bnxt: let the page pool manage the DMA mappingdrivers/net/ethernet/broadcom/bnxt/bnxt.c+8 -162 replies
9/12eth: bnxt: use the page pool for data pagesdrivers/net/ethernet/broadcom/bnxt/bnxt.c+22 -212 replies
10/12eth: bnxt: make sure we make for recycle skbs before freeing themdrivers/net/ethernet/broadcom/bnxt/bnxt.c+4 -00 replies
11/12eth: bnxt: wrap coherent allocations into helpersdrivers/net/ethernet/broadcom/bnxt/bnxt.c+32 -260 replies
12/12eth: bnxt: hack in the use of MEPdrivers/net/ethernet/broadcom/bnxt/bnxt.c drivers/net/ethernet/broadcom/bnxt/bnxt.h+31 -30 replies

Reviews trailers per patch

1 of 12 patches has a Reviewed-by or Acked-by.

PatchJesper Dangaard Brouer
1/12
2/12
3/12
4/12A-b
5/12
6/12
7/12
8/12
9/12
10/12
11/12
12/12

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes.

Read the v1 thread

Cover letter [RFC 00/12] net: huge page backed page_pool

Hi!

This is an "early PoC" at best. It seems to work for a basic traffic test but there's no uAPI and a lot more general polish is needed.

The problem we're seeing is that performance of some older NICs degrades quite a bit when IOMMU is used (in non-passthru mode). There is a long tail of old NICs deployed, especially in PoPs/ /on edge. From a conversation I had with Eric a few months ago it sounded like others may have similar issues. So I thought I'd take a swing at getting page pool to feed drivers huge pages. 1G pages require hooking into early init via CMA but it works just fine.

I haven't tested this with a real workload, because I'm still waiting to get my hands on the right machine. But the experiment with bnxt shows a ~90% reduction in IOTLB misses (670k -> 70k).

In terms of the missing parts - uAPI is definitely needed. The rough plan would be to add memory config via the netdev genl family. Should fit nicely there. Have the config stored in struct netdevice. When page pool is created get to the netdev and automatically select the provider without the driver even knowing. Two problems with that are - 1) if the driver follows the recommended flow of allocating new queues before freeing old ones we will have page pools created before the old ones are gone, which means we'd need to reserve 2x the number of 1G pages; 2) there's no callback to the driver to say "I did something behind your back, don't worry about it, but recreate your queues, please" so the change will not take effect until some unrelated change like installing XDP. Which may be fine in practice but is a bit odd.

Then we get into hand-wavy stuff like - if we can link page pools to netdevs, we should also be able to export the page pool stats via the netdev family instead doing it the ethtool -S.. ekhm.. "way". And if we start storing configs behind driver's back why don't we also store other params, like ring size and queue count... A lot of potential improvements as we iron out a new API...

Live tree: https://github.com/kuba-moo/linux/tree/pp-providers

Jakub Kicinski (12):
net: hack together some page sharing
net: create a 1G-huge-page-backed allocator
net: page_pool: hide page_pool_release_page()
net: page_pool: merge page_pool_release_page() with
page_pool_return_page()
net: page_pool: factor out releasing DMA from releasing the page
net: page_pool: create hooks for custom page providers
net: page_pool: add huge page backed memory providers
eth: bnxt: let the page pool manage the DMA mapping
eth: bnxt: use the page pool for data pages
eth: bnxt: make sure we make for recycle skbs before freeing them
eth: bnxt: wrap coherent allocations into helpers
eth: bnxt: hack in the use of MEP

 Documentation/networking/page_pool.rst        |  10 +-
 arch/x86/kernel/setup.c                       |   6 +-
 drivers/net/ethernet/broadcom/bnxt/bnxt.c     | 154 +++--
 drivers/net/ethernet/broadcom/bnxt/bnxt.h     |   5 +
 drivers/net/ethernet/engleder/tsnep_main.c    |   2 +-
 .../net/ethernet/stmicro/stmmac/stmmac_main.c |   4 +-
 include/net/dcalloc.h                         |  28 +
 include/net/page_pool.h                       |  36 +-
 net/core/Makefile                             |   2 +-
 net/core/dcalloc.c                            | 615 +++++++++++++++++
 net/core/dcalloc.h                            |  96 +++
 net/core/page_pool.c                          | 625 +++++++++++++++++-
 12 files changed, 1478 insertions(+), 105 deletions(-)
 create mode 100644 include/net/dcalloc.h
 create mode 100644 net/core/dcalloc.c
 create mode 100644 net/core/dcalloc.h
Signature · 1 line
2.41.0