[PATCH net-next 0/2] net: atlantic: convert RX path to page_pool

[Yangyu Chen <cyy@cyyself.name>]

On systems where the NIC sits behind an IOMMU, the atlantic RX path
cannot reach line rate: every RX buffer is allocated with
dev_alloc_pages() and mapped with dma_map_page(), then unmapped and
freed once the stack has consumed the packet. Every map/unmap is an
IOTLB/pagetable operation, and at 10G rates this dominates the RX
path. On an AMD Strix Halo system with a Thunderbolt-attached QNAP
QNA-T310G1S (MTU 1500, TCP over IPv6, iperf3 -R), RX tops out at
about 2.2 Gbit/s.

An earlier patch [1] worked around this by making the RX page order
tunable via a module parameter, amortizing one map/unmap over eight
pages worth of frames. The review feedback was to convert the driver
to the page_pool API instead of adding a knob. This series does that
conversion.

Patch 1 fixes a long-standing teardown leak first: TX buffers beyond
the budgeted single aq_ring_tx_clean() pass, or still posted past the
frozen hw_head, are stranded when the interface goes down. Today that
is a silent leak; after the conversion a stranded XDP_TX frame would
hold a page_pool fragment reference and turn every such ifdown into a
permanently stalled pool shutdown, so it needs fixing before the
conversion lands.

Patch 2 converts the RX path to page_pool with the fragment API.
Pages are DMA-mapped once when they enter the per-ring pool
(PP_FLAG_DMA_MAP) and stay mapped while they recycle between the
driver and the stack, so steady-state RX performs no IOMMU work.
page_pool_dev_alloc_frag() takes over the sub-page splitting the
driver's hand-rolled "page flip" scheme did based on page_ref_count(),
buffer ownership becomes transfer-based, and the XDP memory model
switches to MEM_TYPE_PAGE_POOL.

Performance, QNA-T310G1S (AQC100) behind Thunderbolt/IOMMU, MTU 1500,
TCP over IPv6, iperf3 -R:

  before:  2.24 Gbit/s
  after:   9.14 Gbit/s

matching what previously required the rxpageorder=3 workaround, but
with order-0 pages and no tunable.

Tested on the same setup: line rate with plain RX and with XDP_PASS;
XDP_TX with a MAC-swap reflector under sustained traffic; repeated
ifdown/ifup and module unload cycles under both plain RX and XDP_TX
load complete without "stalled pool shutdown" warnings or other splats.

[1] https://lore.kernel.org/lkml/tencent_E71C2F71D9631843941A5DF87204D1B5B509@qq.com/

Yangyu Chen (2):
  net: atlantic: free stranded TX buffers on ring deinit
  net: atlantic: convert RX path to page_pool

 drivers/net/ethernet/aquantia/Kconfig         |   1 +
 .../ethernet/aquantia/atlantic/aq_ethtool.c   |   3 -
 .../net/ethernet/aquantia/atlantic/aq_ptp.c   |  20 +-
 .../net/ethernet/aquantia/atlantic/aq_ring.c  | 268 +++++++++---------
 .../net/ethernet/aquantia/atlantic/aq_ring.h  |   7 +-
 .../net/ethernet/aquantia/atlantic/aq_vec.c   |  17 +-
 6 files changed, 172 insertions(+), 144 deletions(-)

-- 
2.47.3