Date:   Tue,  3 Mar 2020 17:55:35 +0200
From:   Madalin Bucur <madalin.bucur@....nxp.com>
To:     davem@...emloft.net
Cc:     netdev@...r.kernel.org, robh+dt@...nel.org, mark.rutland@....com,
        shawnguo@...nel.org, leoyang.li@....com,
        devicetree@...r.kernel.org, linux-kernel@...r.kernel.org,
        Madalin Bucur <madalin.bucur@....nxp.com>
Subject: [PATCH net 0/4] QorIQ DPAA FMan erratum A050385 workaround

The patch set implements the workaround for FMan erratum A050385:

FMAN DMA read or writes under heavy traffic load may cause FMAN
internal resource leak; thus stopping further packet processing.

The FMAN internal queue can overflow when FMAN splits single
read or write transactions into multiple smaller transactions
such that more than 17 AXI transactions are in flight from FMAN
to interconnect. When the FMAN internal queue overflows, it can
stall further packet processing. The issue can occur with any one
of the following three conditions:

  1. FMAN AXI transaction crosses 4K address boundary (Errata
	 A010022)
  2. FMAN DMA address for an AXI transaction is not 16 byte
	 aligned, i.e. the last 4 bits of an address are non-zero
  3. Scatter Gather (SG) frames have more than one SG buffer in
	 the SG list and any one of the buffers, except the last
	 buffer in the SG list has data size that is not a multiple
	 of 16 bytes, i.e., other than 16, 32, 48, 64, etc.

With any one of the above three conditions present, there is
likelihood of stalled FMAN packet processing, especially under
stress with multiple ports injecting line-rate traffic.

To avoid situations that stall FMAN packet processing, all of the
above three conditions must be avoided; therefore, configure the
system with the following rules:

  1. Frame buffers must not span a 4KB address boundary, unless
	 the frame start address is 256 byte aligned
  2. All FMAN DMA start addresses (for example, BMAN buffer
	 address, FD[address] + FD[offset]) are 16B aligned
  3. SG table and buffer addresses are 16B aligned and the size
	 of SG buffers are multiple of 16 bytes, except for the last
	 SG buffer that can be of any size.

Additional workaround notes:
- Address alignment of 64 bytes is recommended for maximally
efficient system bus transactions (although 16 byte alignment is
sufficient to avoid the stall condition)
- To support frame sizes that are larger than 4K bytes, there are
two options:
  1. Large single buffer frames that span a 4KB page boundary can
	 be converted into SG frames to avoid transaction splits at
	 the 4KB boundary,
  2. Align the large single buffer to 256B address boundaries,
	 ensure that the frame address plus offset is 256B aligned.
- If software generated SG frames have buffers that are unaligned
and with random non-multiple of 16 byte lengths, before
transmitting such frames via FMAN, frames will need to be copied
into a new single buffer or multiple buffer SG frame that is
compliant with the three rules listed above.

Madalin Bucur (4):
  dt-bindings: net: FMan erratum A050385
  arm64: dts: ls1043a: FMan erratum A050385
  fsl/fman: detect FMan erratum A050385
  dpaa_eth: FMan erratum A050385 workaround

 Documentation/devicetree/bindings/net/fsl-fman.txt |   7 ++
 arch/arm64/boot/dts/freescale/fsl-ls1043-post.dtsi |   2 +
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c     | 111 ++++++++++++++++++++-
 drivers/net/ethernet/freescale/fman/fman.c         |  18 ++++
 drivers/net/ethernet/freescale/fman/fman.h         |  10 ++
 5 files changed, 145 insertions(+), 3 deletions(-)

-- 
2.1.0

Powered by blists - more mailing lists