AF_XDP Tx descriptor cleaning in ice driver currently works in a "lazy"
way - descriptors are not cleaned immediately after send. We rather hold
on with cleaning until we see that free space in ring drops below
particular threshold. This was supposed to reduce the amount of
unnecessary work related to cleaning and instead of keeping the ring
empty, ring was rather saturated.
In AF_XDP realm cleaning Tx descriptors implies producing them to CQ.
This is a way of letting know user space that particular descriptor has
been sent, as John points out in [0].
We tried to implement serial descriptor cleaning which would be used in
conjunction with batched cleaning but it made code base more convoluted
and probably harder to maintain in future. Therefore we step away from
batched cleaning in a current form in favor of an approach where we set
RS bit on every last descriptor from a batch and clean always at the
beginning of ice_xmit_zc().
This means that we give up a bit of Tx performance, but this doesn't
hurt l2fwd scenario which is way more meaningful than txonly as this can
be treaten as AF_XDP based packet generator. l2fwd is not hurt due to
the fact that Tx side is much faster than Rx and Rx is the one that has
to catch Tx up.
FWIW Tx descriptors are still produced in a batched way.
[0]: https://lore.kernel.org/bpf/62b0a20232920_3573208ab@john.notmuch/
Fixes: 126cdfe100 ("ice: xsk: Improve AF_XDP ZC Tx and use batching API")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: George Kuruvinakunnel <george.kuruvinakunnel@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Fix leak, when user changes ring parameters.
During reallocation of RX buffers, new DMA mappings are created for
those buffers. New buffers with different RX ring count should
substitute older ones, but those buffers were freed in ice_vsi_cfg_rxq
and reallocated again with ice_alloc_rx_buf. kfree on rx_buf caused
leak of already mapped DMA.
Reallocate ZC with xdp_buf struct, when BPF program loads. Reallocate
back to rx_buf, when BPF program unloads.
If BPF program is loaded/unloaded and XSK pools are created, reallocate
RX queues accordingly in XDP_SETUP_XSK_POOL handler.
Steps for reproduction:
while :
do
for ((i=0; i<=8160; i=i+32))
do
ethtool -G enp130s0f0 rx $i tx $i
sleep 0.5
ethtool -g enp130s0f0
done
done
Fixes: 617f3e1b58 ("ice: xsk: allocate separate memory for XDP SW ring")
Signed-off-by: Przemyslaw Patynowski <przemyslawx.patynowski@intel.com>
Signed-off-by: Mateusz Palczewski <mateusz.palczewski@intel.com>
Tested-by: Chandan <chandanx.rout@intel.com> (A Contingent Worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Several headers in the ice driver include ice.h even though they are
themselves included by that header. The most notable of these is
ice_common.h, but several other headers also do this.
Such a recursive inclusion is problematic as it forces headers to be
included in a strict order, otherwise compilation errors can result. The
circular inclusions do not trigger an endless loop due to standard
header inclusion guards, however other errors can occur.
For example, ice_flow.h defines ice_rss_hash_cfg, which is used by
ice_sriov.h as part of the definition of ice_vf_hash_ip_ctx.
ice_flow.h includes ice_acl.h, which includes ice_common.h, and which
finally includes ice.h. Since ice.h itself includes ice_sriov.h, this
creates a circular dependency.
The definition in ice_sriov.h requires things from ice_flow.h, but
ice_flow.h itself will lead to trying to load ice_sriov.h as part of its
process for expanding ice.h. The current code avoids this issue by
having an implicit dependency without the include of ice_flow.h.
If we were to fix that so that ice_sriov.h explicitly depends on
ice_flow.h the following pattern would occur:
ice_flow.h -> ice_acl.h -> ice_common.h -> ice.h -> ice_sriov.h
At this point, during the expansion of, the header guard for ice_flow.h
is already set, so when ice_sriov.h attempts to load the ice_flow.h
header it is skipped. Then, we go on to begin including the rest of
ice_sriov.h, including structure definitions which depend on
ice_rss_hash_cfg. This produces a compiler warning because
ice_rss_hash_cfg hasn't yet been included. Remember, we're just at the
start of ice_flow.h!
If the order of headers is incorrect (ice_flow.h is not implicitly
loaded first in all files which include ice_sriov.h) then we get the
same failure.
Removing this recursive inclusion requires fixing a few cases where some
headers depended on the header inclusions from ice.h. In addition, a few
other changes are also required.
Most notably, ice_hw_to_dev is implemented as a macro in ice_osdep.h,
which is the likely reason that ice_common.h includes ice.h at all. This
macro implementation requires the full definition of ice_pf in order to
properly compile.
Fix this by moving it to a function declared in ice_main.c, so that we
do not require all files to depend on the layout of the ice_pf
structure.
Note that this change only fixes circular dependencies, but it does not
fully resolve all implicit dependencies where one header may depend on
the inclusion of another. I tried to fix as many of the implicit
dependencies as I noticed, but fixing them all requires a somewhat
tedious analysis of each header and attempting to compile it separately.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Gurucharan G <gurucharanx.g@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Pragma unroll was introduced around GCC 8, whereas current xsk code in
ice that prepares loop_unrolled_for macro that is based on mentioned
pragma, compares GCC version against 4, which is wrong and Stephen
found this out by compiling kernel with GCC 5.4 [0].
Fix this mistake and check if GCC version is >= 8.
[0]: https://lore.kernel.org/netdev/20220307213659.47658125@canb.auug.org.au/
Fixes: 126cdfe100 ("ice: xsk: Improve AF_XDP ZC Tx and use batching API")
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/r/20220307231353.56638-1-maciej.fijalkowski@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Apply the logic that was done for regular XDP from commit 9610bd988d
("ice: optimize XDP_TX workloads") to the ZC side of the driver. On top
of that, introduce batching to Tx that is inspired by i40e's
implementation with adjustments to the cleaning logic - take into the
account NAPI budget in ice_clean_xdp_irq_zc().
Separating the stats structs onto separate cache lines seemed to improve
the performance.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Link: https://lore.kernel.org/bpf/20220125160446.78976-8-maciej.fijalkowski@intel.com
While it was convenient to have a generic ring structure that served
both Tx and Rx sides, next commits are going to introduce several
Tx-specific fields, so in order to avoid hurting the Rx side, let's
pull out the Tx ring onto new ice_tx_ring and ice_rx_ring structs.
Rx ring could be handled by the old ice_ring which would reduce the code
churn within this patch, but this would make things asymmetric.
Make the union out of the ring container within ice_q_vector so that it
is possible to iterate over newly introduced ice_tx_ring.
Remove the @size as it's only accessed from control path and it can be
calculated pretty easily.
Change definitions of ice_update_ring_stats and
ice_fetch_u64_stats_per_ring so that they are ring agnostic and can be
used for both Rx and Tx rings.
Sizes of Rx and Tx ring structs are 256 and 192 bytes, respectively. In
Rx ring xdp_rxq_info occupies its own cacheline, so it's the major
difference now.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Gurucharan G <gurucharanx.g@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Trivial:
The driver had previously attempted to use #define
macros to make functions that have no use in certain
configs disappear. Using static inlines instead allows
for certain static checkers to process the code better,
and results in no functional change.
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Replace the explicit umem reference passed to the driver in AF_XDP
zero-copy mode with the buffer pool instead. This in preparation for
extending the functionality of the zero-copy mode so that umems can be
shared between queues on the same netdev and also between netdevs. In
this commit, only an umem reference has been added to the buffer pool
struct. But later commits will add other entities to it. These are
going to be entities that are different between different queue ids
and netdevs even though the umem is shared between them.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1598603189-32145-2-git-send-email-magnus.karlsson@intel.com
Remove MEM_TYPE_ZERO_COPY in favor of the new MEM_TYPE_XSK_BUFF_POOL
APIs.
v4->v5: Fixed "warning: Excess function parameter 'alloc' description
in 'ice_alloc_rx_bufs_zc'" and "warning: Excess function
parameter 'xdp' description in
'ice_construct_skb_zc'". (Jakub)
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Cc: intel-wired-lan@lists.osuosl.org
Link: https://lore.kernel.org/bpf/20200520192103.355233-10-bjorn.topel@gmail.com
Using ENOTSUPP almost always results in some bizarre error message to
be printed in userspace. This is likely because ENOTSUPP was defined for
the NFS protocol (as per a comment in include/linux/errno.h). Use
EOPNOTSUPP instead.
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Add zero copy AF_XDP support. This patch adds zero copy support for
Tx and Rx; code for zero copy is added to ice_xsk.h and ice_xsk.c.
For Tx, implement ndo_xsk_wakeup. As with other drivers, reuse
existing XDP Tx queues for this task, since XDP_REDIRECT guarantees
mutual exclusion between different NAPI contexts based on CPU ID. In
turn, a netdev can XDP_REDIRECT to another netdev with a different
NAPI context, since the operation is bound to a specific core and each
core has its own hardware ring.
For Rx, allocate frames as MEM_TYPE_ZERO_COPY on queues that AF_XDP is
enabled.
Signed-off-by: Krzysztof Kazimierczak <krzysztof.kazimierczak@intel.com>
Co-developed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
