net/sched/cls_api.c has overlapping changes to a call to
nlmsg_parse(), one (from 'net') added rtm_tca_policy instead of NULL
to the 5th argument, and another (from 'net-next') added cb->extack
instead of NULL to the 6th argument.
net/ipv4/ipmr_base.c is a case of a bug fix in 'net' being done to
code which moved (to mr_table_dump)) in 'net-next'. Thanks to David
Ahern for the heads up.
Signed-off-by: David S. Miller <davem@davemloft.net>
mlx5e netdevice used to calculate fragment edges by a call to
mlx5_wq_cyc_get_frag_size(). This calculation did not give the correct
indication for queues smaller than a PAGE_SIZE, (broken by default on
PowerPC, where PAGE_SIZE == 64KB). Here it is replaced by the correct new
calls/API.
Since (TX/RX) Work Queues buffers are fragmented, here we introduce
changes to the API in core driver, so that it gets a stride index and
returns the index of last stride on same fragment, and an additional
wrapping function that returns the number of physically contiguous
strides that can be written contiguously to the work queue.
This obsoletes the following API functions, and their buggy
usage in EN driver:
* mlx5_wq_cyc_get_frag_size()
* mlx5_wq_cyc_ctr2fragix()
The new API improves modularity and hides the details of such
calculation for mlx5e netdevice and mlx5_ib rdma drivers.
New calculation is also more efficient, and improves performance
as follows:
Packet rate test: pktgen, UDP / IPv4, 64byte, single ring, 8K ring size.
Before: 16,477,619 pps
After: 17,085,793 pps
3.7% improvement
Fixes: 3a2f703312 ("net/mlx5: Use order-0 allocations for all WQ types")
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Currently we practically never report checksum unnecessary, because
for all IP packets we take the checksum complete path.
Enable non-default runs with reprorting checksum unnecessary, using
an ethtool private flag. This can be useful for performance evals
and other explorations.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
We can report checksum unnecessary also when the L3 checksum
flag on the cqe is set and there's no L4 header.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
CHECKSUM_COMPLETE is not applicable to SCTP protocol.
Setting it for SCTP packets leads to CRC32c validation failure.
Fixes: bbceefce9a ("net/mlx5e: Support RX CHECKSUM_COMPLETE")
Signed-off-by: Alaa Hleihel <alaa@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
In multi-host (MH) NIC scheme, a single HW port serves multiple hosts
or sockets on the same host.
The HW uses a mechanism in the PCIe buffer which monitors
the amount of consumed PCIe buffers per host.
On a certain configuration, under congestion,
the HW emulates a switch doing ECN marking on packets using ECN
indication on the completion descriptor (CQE).
The driver needs to set the ECN bits on the packet SKB,
such that the network stack can react on that, this commit does that.
Signed-off-by: Natali Shechtman <natali@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Since the RQs are shared between all pkey interfaces, the stats
should be taken from where the per-ring stats are stored instead
of the parent RQ.
Fixes: 4c6c615e3f ("net/mlx5e: IPoIB, Add PKEY child interface nic profile")
Signed-off-by: Feras Daoud <ferasda@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A loaded XDP program might write to the xdp_frame data area,
prefetchw() it to avoid a potential cache miss.
Performance tests:
ConnectX-5, XDP_TX packet rate, single ring.
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Before: 13,172,976 pps
After: 13,456,248 pps
2% gain.
Fixes: 22f4539881 ("net/mlx5e: Support XDP over Striding RQ")
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
In the downstream patch that adds support to XDP_REDIRECT-out,
the XDP xmit frame function doesn't share the same run context as
the NAPI that polls the XDP-SQ completion queue.
Hence, need to re-order the XDP-SQ fields to avoid cacheline
false-sharing.
Take redirect_flush and doorbell out of DB, into separated
cachelines.
Add a cacheline breaker within the stats struct.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Take XDP code out of the general EN header and RX file into
new XDP files.
Currently, XDP-SQ resides only within an RQ and used from a
single flow (XDP_TX) triggered upon RX completions.
In a downstream patch, additional type of XDP-SQ instances will be
presented and used for the XDP_REDIRECT flow, totally unrelated to
the RX context.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
No need to expose the MPWQE free function to control path.
The dealloc function already exposed, use it.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
This patch fixes the byte count indication in CQE for processed IPsec
packets that contain a metadata header.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Implement the TLS rx offload data path according to the
requirements of the TLS generic NIC offload infrastructure.
Special metadata ethertype is used to pass information to
the hardware.
When hardware loses synchronization a special resync request
metadata message is used to request resync.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add ethtool counter to indicate the number of strides consumed
by filler CQEs.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Add per-ring and global ethtool counters for congested UMR requests.
These events indicate congestion in UMR handlers in HW.
Such event is concluded when there's an outstanding UMR post,
yet the SW consumed at least two additional MPWQEs in the meanwhile.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Add per-ring and global ethtool counters for XDP_TX completions.
This helps us monitor and analyze XDP_TX flow performance.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Local variable 'wq' already points to &sq->wq, use it.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Enhance the memory scheme of the legacy RQ, such that
only order-0 pages are used.
Whenever possible, prefer using a linear SKB, and build it
wrapping the WQE buffer.
Otherwise (for example, jumbo frames on x86), use non-linear SKB,
with as many frags as needed. In this case, multiple WQE
scatter entries are used, up to a maximum of 4 frags and 10KB of MTU.
This implied to remove support of HW LRO in legacy RQ, as it would
require large number of page allocations and scatter entries per WQE
on archs with PAGE_SIZE = 4KB, yielding bad performance.
In earlier patches, we guaranteed that all completions are in-order,
and that we use a cyclic WQ.
This creates an oppurtunity for a performance optimization:
The mapping between a "struct mlx5e_dma_info", and the
WQEs (struct mlx5e_wqe_frag_info) pointing to it, is constant
across different cycles of a WQ. This allows initializing
the mapping in the time of RQ creation, and not handle it
in datapath.
A struct mlx5e_dma_info that is shared between different WQEs
is allocated by the first WQE, and freed by the last one.
This implies an important requirement: WQEs that share the same
struct mlx5e_dma_info must be posted within the same NAPI.
Otherwise, upon completion, struct mlx5e_wqe_frag_info would mistakenly
point to the new struct mlx5e_dma_info, not the one that was posted
(and the HW wrote to).
This bulking requirement is actually good also for performance reasons,
hence we extend the bulk beyong the minimal requirement above.
With this memory scheme, the RQs memory footprint is reduce by a
factor of 2 on x86, and by a factor of 32 on PowerPC.
Same factors apply for the number of pages in a GRO session.
Performance tests:
ConnectX-4, single core, single RX ring, default MTU.
x86:
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Packet rate (early drop in TC): no degradation
TCP streams: ~5% improvement
PowerPC:
CPU: POWER8 (raw), altivec supported
Packet rate (early drop in TC): 20% gain
TCP streams: 25% gain
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Now that LRO is not supported for Legacy RQ, there is no source of
out-of-order completions in the WQ, and we can use a cyclic one.
This has multiple advantages:
- reduces the WQE size (smaller PCI transactions).
- lower overhead in datapath (no handling of 'next' pointers).
- no reserved WQE for the WQ head (was need in linked-list).
- allows using a constant map between frag and dma_info struct, in downstream patch.
Performance tests:
ConnectX-4, single core, single RX ring.
Major gain in packet rate of single ring XDP drop.
Bottleneck is shifted form HW (at 16Mpps) to SW (at 20Mpps).
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Replace the common RQ WQ object with two separate ones for the
different RQ types.
This is in preparation for switching to using a cyclic WQ type
in Legacy RQ.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Get the logic of copying the packet header into the SKB linear part
into a generic function. Function does copy length alignment
and dma buffer sync.
It is currently called only within the MPWQE flow.
In a downstream patch, it will be called within the legacy RQ flow
as well.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Rename it and pass truesize as an extra argument, as it will be used also
in Legacy RQ in a downstream patch.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Make name more generic by dropping MPWRQ from it, as it will be
used also in Legacy RQ in a downstream patch.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
This series includes updates for mlx5e netdev driver.
1) Allowr flow based VF vport mirroring under sriov switchdev scheme,
added support for offloading the TC mirred mirror sub-action, from
Chris Mi.
=================
From: Or Gerlitz <ogerlitz@mellanox.com>
The user will typically set the actions order such that the mirror
port (mirror VF) sees packets as the original port (VF under
mirroring) sent them or as it will receive them. In the general case,
it means that packets are potentially sent to the mirror port before
or after some actions were applied on them.
To properly do that, we follow on the exact action order as set for
the flow and make sure this will also be the case when we program the
HW offload.
If all the actions should apply before forwarding to the mirror and dest port,
mirroring is just multicasting to the two vports. Otherwise, we split
the TC flow to two HW rules, where the 1st applies only the actions
needed up to the mirror (if there are such) and the 2nd the rest of
the actions plus the forwarding to the dest vport.
=================
2) Move to order-0 only allocations (using fragmented work queues) for all
work queues used by the driver, RX and TX descriptor rings
(RQs, SQs and Completion Queues (CQs)), from Tariq Toukan.
3) Avoid resetting netdevice statistics on netdevice
state changes, from Eran Ben Elisha.
-----BEGIN PGP SIGNATURE-----
iQEcBAABAgAGBQJbCKCcAAoJEEg/ir3gV/o++7sH/1FPmwvpf4qNEsusr714lNnl
PxzVjnwxKZNYyovIjr6QGSxMM1qDiPejyYnZIAzH00B+XWCq3zn8H3sfJLFbmxN3
clayd6dGV27HzLZwV2aD9vXfVb7snNhQtTp5zoajsnZY4xO335n3FA3kF5TxLqKO
bxsnY2xSaSpbrBH2z2UvHc+ib9KnvY1Q+gr2WqdRnN5Tm51Zq+bgaUw0nROefGJK
XR/aVBub3PsjA8W/0/b3DGfiP1bgPeU6QmLhdjn3IEufFtbEJH+K/4u53l3A4RLR
fXmJrWKlZn8j7LUBFOD0/G43RU/YzqRgwF/iUwCxbUU2wK/J0cPv23drLDVYzaY=
=n7Zz
-----END PGP SIGNATURE-----
Merge tag 'mlx5e-updates-2018-05-25' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux
Saeed Mahameed says:
====================
mlx5e-updates-2018-05-25
This series includes updates for mlx5e netdev driver.
1) Allowr flow based VF vport mirroring under sriov switchdev scheme,
added support for offloading the TC mirred mirror sub-action, from
Chris Mi.
=================
From: Or Gerlitz <ogerlitz@mellanox.com>
The user will typically set the actions order such that the mirror
port (mirror VF) sees packets as the original port (VF under
mirroring) sent them or as it will receive them. In the general case,
it means that packets are potentially sent to the mirror port before
or after some actions were applied on them.
To properly do that, we follow on the exact action order as set for
the flow and make sure this will also be the case when we program the
HW offload.
If all the actions should apply before forwarding to the mirror and dest port,
mirroring is just multicasting to the two vports. Otherwise, we split
the TC flow to two HW rules, where the 1st applies only the actions
needed up to the mirror (if there are such) and the 2nd the rest of
the actions plus the forwarding to the dest vport.
=================
2) Move to order-0 only allocations (using fragmented work queues) for all
work queues used by the driver, RX and TX descriptor rings
(RQs, SQs and Completion Queues (CQs)), from Tariq Toukan.
3) Avoid resetting netdevice statistics on netdevice
state changes, from Eran Ben Elisha.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Move all RQ, SQ and channel counters from the channel objects into the
priv structure. With this change, counters will not be reset upon
channel configuration changes.
Channel's statistics for SQs which are associated with TCs higher than
zero will be presented in ethtool -S, only for SQs which were opened at
least once since the module was loaded (regardless of their open/close
current status). This is done in order to decrease the total amount of
statistics presented and calculated for the common out of box use (no
QoS).
mlx5e_channel_stats is a compound of CH,RQ,SQs stats in order to
create locality for the NAPI when handling TX and RX of the same
channel.
Align the new statistics struct per ring to avoid several channels
update to the same cache line at the same time.
Packet rate was tested, no degradation sensed.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
CC: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Complete the transition of all WQ types to use fragmented
order-0 coherent memory instead of high-order allocations.
CQ-WQ already uses order-0.
Here we do the same for cyclic and linked-list WQs.
This allows the driver to load cleanly on systems with a highly
fragmented coherent memory.
Performance tests:
ConnectX-5 100Gbps, CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Packet rate of 64B packets, single transmit ring, size 8K.
No degradation is sensed.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
We fill SQ edge with NOPs to avoid WQEs wrap.
Here, instead of doing that in advance for the maximum possible
WQE size, we do it on-demand using the actual WQE size.
We re-order some parts in mlx5e_sq_xmit to finish the calculation
of WQE size (ds_cnt) before doing any writes to the WQE buffer.
When SQ work queue is fragmented (introduced in an downstream patch),
dealing with WQE wraps becomes more frequent. This change would drastically
reduce the overhead in this case.
Performance tests:
ConnectX-5 100Gbps, CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Packet rate of 64B packets, single transmit ring, size 8K.
Before: 14.9 Mpps
After: 15.8 Mpps
Improvement of 6%.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Use the WQ API to get the WQ size, and to map a counter
into a WQ entry index.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When RXFCS feature is enabled, the HW do not strip the FCS data,
however it is not present in the checksum calculated by the HW.
Fix that by manually calculating the FCS checksum and adding it to the SKB
checksum field.
Add helper function to find the FCS data for all SKB forms (linear,
one fragment or more).
Fixes: 102722fc68 ("net/mlx5e: Add support for RXFCS feature flag")
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
MLX5E_TEST_BIT macro is the same as the already existent test_bit,
remove it and replace all usages.
Signed-off-by: Gal Pressman <galp@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Range of LRO number of segments fits in u8.
Also, bring initialization and declaration together to
save code.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Changing API xdp_return_frame() to take struct xdp_frame as argument,
seems like a natural choice. But there are some subtle performance
details here that needs extra care, which is a deliberate choice.
When de-referencing xdp_frame on a remote CPU during DMA-TX
completion, result in the cache-line is change to "Shared"
state. Later when the page is reused for RX, then this xdp_frame
cache-line is written, which change the state to "Modified".
This situation already happens (naturally) for, virtio_net, tun and
cpumap as the xdp_frame pointer is the queued object. In tun and
cpumap, the ptr_ring is used for efficiently transferring cache-lines
(with pointers) between CPUs. Thus, the only option is to
de-referencing xdp_frame.
It is only the ixgbe driver that had an optimization, in which it can
avoid doing the de-reference of xdp_frame. The driver already have
TX-ring queue, which (in case of remote DMA-TX completion) have to be
transferred between CPUs anyhow. In this data area, we stored a
struct xdp_mem_info and a data pointer, which allowed us to avoid
de-referencing xdp_frame.
To compensate for this, a prefetchw is used for telling the cache
coherency protocol about our access pattern. My benchmarks show that
this prefetchw is enough to compensate the ixgbe driver.
V7: Adjust for commit d9314c474d ("i40e: add support for XDP_REDIRECT")
V8: Adjust for commit bd658dda42 ("net/mlx5e: Separate dma base address
and offset in dma_sync call")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch shows how it is possible to have both the driver local page
cache, which uses elevated refcnt for "catching"/avoiding SKB
put_page returns the page through the page allocator. And at the
same time, have pages getting returned to the page_pool from
ndp_xdp_xmit DMA completion.
The performance improvement for XDP_REDIRECT in this patch is really
good. Especially considering that (currently) the xdp_return_frame
API and page_pool_put_page() does per frame operations of both
rhashtable ID-lookup and locked return into (page_pool) ptr_ring.
(It is the plan to remove these per frame operation in a followup
patchset).
The benchmark performed was RX on mlx5 and XDP_REDIRECT out ixgbe,
with xdp_redirect_map (using devmap) . And the target/maximum
capability of ixgbe is 13Mpps (on this HW setup).
Before this patch for mlx5, XDP redirected frames were returned via
the page allocator. The single flow performance was 6Mpps, and if I
started two flows the collective performance drop to 4Mpps, because we
hit the page allocator lock (further negative scaling occurs).
Two test scenarios need to be covered, for xdp_return_frame API, which
is DMA-TX completion running on same-CPU or cross-CPU free/return.
Results were same-CPU=10Mpps, and cross-CPU=12Mpps. This is very
close to our 13Mpps max target.
The reason max target isn't reached in cross-CPU test, is likely due
to RX-ring DMA unmap/map overhead (which doesn't occur in ixgbe to
ixgbe testing). It is also planned to remove this unnecessary DMA
unmap in a later patchset
V2: Adjustments requested by Tariq
- Changed page_pool_create return codes not return NULL, only
ERR_PTR, as this simplifies err handling in drivers.
- Save a branch in mlx5e_page_release
- Correct page_pool size calc for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
V5: Updated patch desc
V8: Adjust for b0cedc844c ("net/mlx5e: Remove rq_headroom field from params")
V9:
- Adjust for 121e892754 ("net/mlx5e: Refactor RQ XDP_TX indication")
- Adjust for 73281b78a3 ("net/mlx5e: Derive Striding RQ size from MTU")
- Correct handling if page_pool_create fail for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
V10: Req from Tariq
- Change pool_size calc for MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This implements basic XDP redirect support in mlx5 driver.
Notice that the ndo_xdp_xmit() is NOT implemented, because that API
need some changes that this patchset is working towards.
The main purpose of this patch is have different drivers doing
XDP_REDIRECT to show how different memory models behave in a cross
driver world.
Update(pre-RFCv2 Tariq): Need to DMA unmap page before xdp_do_redirect,
as the return API does not exist yet to to keep this mapped.
Update(pre-RFCv3 Saeed): Don't mix XDP_TX and XDP_REDIRECT flushing,
introduce xdpsq.db.redirect_flush boolian.
V9: Adjust for commit 121e892754 ("net/mlx5e: Refactor RQ XDP_TX indication")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Upon a new UMR post, check if the WQE buffer contains
a previous UMR WQE. If so, modify the dynamic fields
instead of a whole WQE overwrite. This saves a memcpy.
In current setting, after 2 WQ cycles (12 UMR posts),
this will always be the case.
No degradation sensed.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
All UMR WQEs of an RQ share many common fields. We use
pre-initialized structures to save calculations in datapath.
One field (xlt_offset) was the only reason we saved a pre-initialized
copy per WQE index.
Here we remove its initialization (move its calculation to datapath),
and reduce the number of copies to one-per-RQ.
A very small datapath calculation is added, it occurs once per a MPWQE
(i.e. once every 256KB), but reduces memory consumption and gives
better cache utilization.
Performance testing:
Tested packet rate, no degradation sensed.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When many packets reside on the same page, the bulking of
page_ref modifications reduces the total number of atomic
operations executed.
Besides the necessary 2 operations on page alloc/free, we
have the following extra ops per page:
- one on WQE allocation (bump refcnt to maximum possible),
- zero ops for SKBs,
- one on WQE free,
a constant of two operations in total, no matter how many
packets/SKBs actually populate the page.
Without this bulking, we have:
- no ops on WQE allocation or free,
- one op per SKB,
Comparing the two methods when PAGE_SIZE is 4K:
- As mentioned above, bulking method always executes 2 operations,
not more, but not less.
- In the default MTU configuration (1500, stride size is 2K),
the non-bulking method execute 2 ops as well.
- For larger MTUs with stride size of 4K, non-bulking method
executes only a single op.
- For XDP (stride size of 4K, no SKBs), non-bulking method
executes no ops at all!
Hence, to optimize the flows with linear SKB and XDP over Striding RQ,
we here remove the page_ref bulking method.
Performance testing:
ConnectX-5, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz.
Single core packet rate (64 bytes).
Early drop in TC: no degradation.
XDP_DROP:
before: 14,270,188 pps
after: 20,503,603 pps, 43% improvement.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Add XDP support over Striding RQ.
Now that linear SKB is supported over Striding RQ,
we can support XDP by setting stride size to PAGE_SIZE
and headroom to XDP_PACKET_HEADROOM.
Upon a MPWQE free, do not release pages that are being
XDP xmit, they will be released upon completions.
Striding RQ is capable of a higher packet-rate than
conventional RQ.
A performance gain is expected for all cases that had
a HW packet-rate bottleneck. This is the case whenever
using many flows that distribute to many cores.
Performance testing:
ConnectX-5, 24 rings, default MTU.
CQE compression ON (to reduce completions BW in PCI).
XDP_DROP packet rate:
--------------------------------------------------
| pkt size | XDP rate | 100GbE linerate | pct% |
--------------------------------------------------
| 64byte | 126.2 Mpps | 148.0 Mpps | 85% |
| 128byte | 80.0 Mpps | 84.8 Mpps | 94% |
| 256byte | 42.7 Mpps | 42.7 Mpps | 100% |
| 512byte | 23.4 Mpps | 23.4 Mpps | 100% |
--------------------------------------------------
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Make the xdp_xmit indication available for Striding RQ
by taking it out of the type-specific union.
This refactor is a preparation for a downstream patch that
adds XDP support over Striding RQ.
In addition, use a bitmap instead of a boolean for possible
future flags.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Current Striding RQ HW feature utilizes the RX buffers so that
there is no wasted room between the strides. This maximises
the memory utilization.
This prevents the use of build_skb() (which requires headroom
and tailroom), and demands to memcpy the packets headers into
the skb linear part.
In this patch, whenever a set of conditions holds, we apply
an RQ configuration that allows combining the use of linear SKB
on top of a Striding RQ.
To use build_skb() with Striding RQ, the following must hold:
1. packet does not cross a page boundary.
2. there is enough headroom and tailroom surrounding the packet.
We can satisfy 1 and 2 by configuring:
stride size = MTU + headroom + tailoom.
This is possible only when:
a. (MTU - headroom - tailoom) does not exceed PAGE_SIZE.
b. HW LRO is turned off.
Using linear SKB has many advantages:
- Saves a memcpy of the headers.
- No page-boundary checks in datapath.
- No filler CQEs.
- Significantly smaller CQ.
- SKB data continuously resides in linear part, and not split to
small amount (linear part) and large amount (fragment).
This saves datapath cycles in driver and improves utilization
of SKB fragments in GRO.
- The fragments of a resulting GRO SKB follow the IP forwarding
assumption of equal-size fragments.
Some implementation details:
HW writes the packets to the beginning of a stride,
i.e. does not keep headroom. To overcome this we make sure we can
extend backwards and use the last bytes of stride i-1.
Extra care is needed for stride 0 as it has no preceding stride.
We make sure headroom bytes are available by shifting the buffer
pointer passed to HW by headroom bytes.
This configuration now becomes default, whenever capable.
Of course, this implies turning LRO off.
Performance testing:
ConnectX-5, single core, single RX ring, default MTU.
UDP packet rate, early drop in TC layer:
--------------------------------------------
| pkt size | before | after | ratio |
--------------------------------------------
| 1500byte | 4.65 Mpps | 5.96 Mpps | 1.28x |
| 500byte | 5.23 Mpps | 5.97 Mpps | 1.14x |
| 64byte | 5.94 Mpps | 5.96 Mpps | 1.00x |
--------------------------------------------
TCP streams: ~20% gain
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When modifying the page mapping of a HW memory region
(via a UMR post), post the new values inlined in WQE,
instead of using a data pointer.
This is a micro-optimization, inline UMR WQEs of different
rings scale better in HW.
In addition, this obsoletes a few control flows and helps
delete ~50 LOC.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Do not busy-wait a pending UMR completion. Under high HW load,
busy-waiting a delayed completion would fully utilize the CPU core
and mistakenly indicate a SW bottleneck.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Gets the process of a UMR WQE post in one function,
in preparation for a downstream patch that inlines
the WQE data.
No functional change here.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Knowing the MTU is required for RQ creation flow.
By our design, channels creation flow is totally isolated
from priv/netdev, and can be completed with access to
channels params and mdev.
Adding the MTU to the channels params helps preserving that.
In addition, we save it in RQ to make its access faster in
datapath checks.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
In copying skb header to skb->data, replace the call to
skb_copy_to_linear_data_offset() with a zero offset with
the call to the no-offset function skb_copy_to_linear_data().
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Pass the base dma address and offset to dma_sync_single_range_for_cpu(),
instead of doing the pre-calculation.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
David S. Miller