| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit adds support for AF_XDP zero-copy RX and TX.
We create a dedicated XSK RQ inside the channel, it means that two
RQs are running simultaneously: one for non-XSK traffic and the other
for XSK traffic. The regular and XSK RQs use a single ID namespace split
into two halves: the lower half is regular RQs, and the upper half is
XSK RQs. When any zero-copy AF_XDP socket is active, changing the number
of channels is not allowed, because it would break to mapping between
XSK RQ IDs and channels.
XSK requires different page allocation and release routines. Such
functions as mlx5e_{alloc,free}_rx_mpwqe and mlx5e_{get,put}_rx_frag are
generic enough to be used for both regular and XSK RQs, and they use the
mlx5e_page_{alloc,release} wrappers around the real allocation
functions. Function pointers are not used to avoid losing the
performance with retpolines. Wherever it's certain that the regular
(non-XSK) page release function should be used, it's called directly.
Only the stats that could be meaningful for XSK are exposed to the
userspace. Those that don't take part in the XSK flow are not
considered.
Note that we don't wait for WQEs on the XSK RQ (unlike the regular RQ),
because the newer xdpsock sample doesn't provide any Fill Ring entries
at the setup stage.
We create a dedicated XSK SQ in the channel. This separation has its
advantages:
1. When the UMEM is closed, the XSK SQ can also be closed and stop
receiving completions. If an existing SQ was used for XSK, it would
continue receiving completions for the packets of the closed socket. If
a new UMEM was opened at that point, it would start getting completions
that don't belong to it.
2. Calculating statistics separately.
When the userspace kicks the TX, the driver triggers a hardware
interrupt by posting a NOP to a dedicated XSK ICO (internal control
operations) SQ, in order to trigger NAPI on the right CPU core. This XSK
ICO SQ is protected by a spinlock, as the userspace application may kick
the TX from any core.
Store the pointers to the UMEMs in the net device private context,
independently from the kernel. This way the driver can distinguish
between the zero-copy and non-zero-copy UMEMs. The kernel function
xdp_get_umem_from_qid does not care about this difference, but the
driver is only interested in zero-copy UMEMs, particularly, on the
cleanup it determines whether to close the XSK RQ and SQ or not by
looking at the presence of the UMEM. Use state_lock to protect the
access to this area of UMEM pointers.
LRO isn't compatible with XDP, but there may be active UMEMs while
XDP is off. If this is the case, don't allow LRO to ensure XDP can
be reenabled at any time.
The validation of XSK parameters typically happens when XSK queues
open. However, when the interface is down or the XDP program isn't
set, it's still possible to have active AF_XDP sockets and even to
open new, but the XSK queues will be closed. To cover these cases,
perform the validation also in these flows:
1. A new UMEM is registered, but the XSK queues aren't going to be
created due to missing XDP program or interface being down.
2. MTU changes while there are UMEMs registered.
Having this early check prevents mlx5e_open_channels from failing
at a later stage, where recovery is impossible and the application
has no chance to handle the error, because it got the successful
return value for an MTU change or XSK open operation.
The performance testing was performed on a machine with the following
configuration:
- 24 cores of Intel Xeon E5-2620 v3 @ 2.40 GHz
- Mellanox ConnectX-5 Ex with 100 Gbit/s link
The results with retpoline disabled, single stream:
txonly: 33.3 Mpps (21.5 Mpps with queue and app pinned to the same CPU)
rxdrop: 12.2 Mpps
l2fwd: 9.4 Mpps
The results with retpoline enabled, single stream:
txonly: 21.3 Mpps (14.1 Mpps with queue and app pinned to the same CPU)
rxdrop: 9.9 Mpps
l2fwd: 6.8 Mpps
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
|
|
| |
Use the existing mlx5e_get_linear_rq_headroom function to calculate the
headroom for mlx5e_xdp_max_mtu. This function takes the XSK headroom
into consideration, which will be used in the following patches.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When an XDP program returns XDP_TX, and the RQ is XSK-enabled, it
requires careful handling, because convert_to_xdp_frame creates a new
page and copies the data there, while our driver expects the xdp_frame
to point to the same memory as the xdp_buff. Handle this case
separately: map the page, and in the end unmap it and call
xdp_return_frame.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Put the XDP SQ that is used for XDP_TX into the channel. It used to be a
part of the RQ, but with introduction of AF_XDP there will be one more
RQ that could share the same XDP SQ. This patch is a preparation for
that change.
Separate XDP_TX statistics per RQ were implemented in one of the previous
patches.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, struct mlx5e_xdp_info has some issues that have to be cleaned
up before the upcoming AF_XDP support makes things too complicated and
messy. This structure is used both when sending the packet and on
completion. Moreover, the cleanup procedure on completion depends on the
origin of the packet (XDP_REDIRECT, XDP_TX). Adding AF_XDP support will
add new flows that use this structure even differently. To avoid
overcomplicating the code, this commit refactors the usage of this
structure in the following ways:
1. struct mlx5e_xdp_info is split into two different structures. One is
struct mlx5e_xdp_xmit_data, a transient structure that doesn't need to
be stored and is only used while sending the packet. The other is still
struct mlx5e_xdp_info that is stored in a FIFO and contains the fields
needed on completion.
2. The fields of struct mlx5e_xdp_info that are used in different flows
are put into a union. A special enum indicates the cleanup mode and
helps choose the right union member. This approach is clear and
explicit. Although it could be possible to "guess" the mode by looking
at the values of the fields and at the XDP SQ type, it wouldn't be that
clear and extendable and would require looking through the whole chain
to understand what's going on.
For the reference, there are the fields of struct mlx5e_xdp_info that
are used in different flows (including AF_XDP ones):
Packet origin | Fields used on completion | Cleanup steps
-----------------------+---------------------------+------------------
XDP_REDIRECT, | xdpf, dma_addr | DMA unmap and
XDP_TX from XSK RQ | | xdp_return_frame.
-----------------------+---------------------------+------------------
XDP_TX from regular RQ | di | Recycle page.
-----------------------+---------------------------+------------------
AF_XDP TX | (none) | Increment the
| | producer index in
| | Completion Ring.
On send, the same set of mlx5e_xdp_xmit_data fields is used in all
flows: DMA and virtual addresses and length.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
|
| |
The PCI API for DMA is deprecated, and PCI_DMA_TODEVICE is just defined
to DMA_TO_DEVICE for backward compatibility. Just use DMA_TO_DEVICE.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
|
|
|
|
|
|
|
|
|
| |
The same code that returns XDP frames and releases pages is used both in
mlx5e_poll_xdpsq_cq and mlx5e_free_xdpsq_descs. Create a function that
cleans up an MPWQE.
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|\
| |
| |
| |
| |
| | |
Two easy cases of overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
MLX5E_XDP_MAX_MTU was calculated incorrectly. It didn't account for
NET_IP_ALIGN and MLX5E_HW2SW_MTU, and it also misused MLX5_SKB_FRAG_SZ.
This commit fixes the calculations and adds a brief explanation for the
formula used.
Fixes: a26a5bdf3ee2d ("net/mlx5e: Restrict the combination of large MTU and XDP")
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
xdp_return_frame releases the frame. It leads to releasing the page, so
it's not allowed to access xdpi.xdpf->len after that, because xdpi.xdpf
is at xdp->data_hard_start after convert_to_xdp_frame. This patch moves
the memory access to precede the return of the frame.
Fixes: 58b99ee3e3ebe ("net/mlx5e: Add support for XDP_REDIRECT in device-out side")
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Upon high packet rate with multiple CPUs TX workloads, much of the HCA's
resources are spent on prefetching TX descriptors, thus affecting
transmission rates.
This patch comes to mitigate this problem by moving some workload to the
CPU and reducing the HW data prefetch overhead for small packets (<= 256B).
When forwarding packets with XDP, a packet that is smaller
than a certain size (set to ~256 bytes) would be sent inline within
its WQE TX descrptor (mem-copied), when the hardware tx queue is congested
beyond a pre-defined water-mark.
This is added to better utilize the HW resources (which now makes
one less packet data prefetch) and allow better scalability, on the
account of CPU usage (which now 'memcpy's the packet into the WQE).
To load balance between HW and CPU and get max packet rate, we use
watermarks to detect how much the HW is congested and move the work
loads back and forth between HW and CPU.
Performance:
Tested packet rate for UDP 64Byte multi-stream
over two dual port ConnectX-5 100Gbps NICs.
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
* Tested with hyper-threading disabled
XDP_TX:
| | before | after | |
| 24 rings | 51Mpps | 116Mpps | +126% |
| 1 ring | 12Mpps | 12Mpps | same |
XDP_REDIRECT:
** Below is the transmit rate, not the redirection rate
which might be larger, and is not affected by this patch.
| | before | after | |
| 32 rings | 64Mpps | 92Mpps | +43% |
| 1 ring | 6.4Mpps | 6.4Mpps | same |
As we can see, feature significantly improves scaling, without
hurting single ring performance.
Signed-off-by: Shay Agroskin <shayag@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
| |
| |
| |
| |
| |
| |
| |
| | |
This counter tracks how many TX MPWQE sessions are started in XDP SQ
in XDP TX/REDIRECT flow. It counts per-channel and global stats.
Signed-off-by: Shay Agroskin <shayag@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|/
|
|
|
|
|
|
|
|
|
| |
The XDP redirect flush indication belongs to the receive queue,
not to its XDP send queue.
For this, use a new bit on rq->flags.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Shay Agroskin <shayag@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently mlx5 driver creates xdp redirect hw queues unconditionally on
netdevice open, This is great until someone starts redirecting XDP traffic
via ndo_xdp_xmit on mlx5 device and changes the device configuration at
the same time, this might cause crashes, since the other device's napi
is not aware of the mlx5 state change (resources un-availability).
To fix this we must synchronize with other devices napi's on the system.
Added a new flag under mlx5e_priv to determine XDP TX resources are
available, set/clear it up when necessary and use synchronize_rcu()
when the flag is turned off, so other napi's are in-sync with it, before
we actually cleanup the hw resources.
The flag is tested prior to committing to transmit on mlx5e_xdp_xmit, and
it is sufficient to determine if it safe to transmit or not. The other
two internal flags (MLX5E_STATE_OPENED and MLX5E_SQ_STATE_ENABLED) become
unnecessary. Thus, they are removed from data path.
Fixes: 58b99ee3e3eb ("net/mlx5e: Add support for XDP_REDIRECT in device-out side")
Reported-by: Toke Høiland-Jørgensen <toke@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add support for the HW feature of multi-packet WQE in XDP
xmit flow.
The conventional TX descriptor (WQE, Work Queue Element) serves
a single packet. Our HW has support for multi-packet WQE (MPWQE)
in which a single descriptor serves multiple TX packets.
This reduces both the PCI overhead and the CPU cycles wasted on
writing them.
In this patch we add support for the HW feature, which is supported
starting from ConnectX-5.
Performance:
Tested packet rate for UDP 64Byte multi-stream over ConnectX-5 NICs.
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
XDP_TX:
We see a huge gain on single port ConnectX-5, and reach the 100 Mpps
milestone.
* Single-port HCA:
Before: 70 Mpps
After: 100 Mpps (+42.8%)
* Dual-port HCA:
Before: 51.7 Mpps
After: 57.3 Mpps (+10.8%)
* In both cases we tested traffic on one port and for now On Dual-port HCAs
we see only small gain, we are working to overcome this bottleneck, but
for the moment only with experimental firmware on dual port HCAs we can
reach the wanted numbers as seen on Single-port HCAs.
XDP_REDIRECT:
Redirect from (A) ConnectX-5 to (B) ConnectX-5.
Due to a setup limitation, (A) and (B) are on different NUMA nodes,
so absolute performance numbers are not optimal.
Note:
Below is the transmit rate of (B), not the redirect rate of (A)
which is in some cases higher.
* (B) is single-port:
Before: 77 Mpps
After: 90 Mpps (+16.8%)
* (B) is dual-port:
Before: 61 Mpps
After: 72 Mpps (+18%)
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
| |
Each xdp_wqe_info instance describes the number of data-segments
and WQEBBs of the WQE.
This is useful for a downstream patch that adds support for
Multi-Packet TX WQE feature.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
| |
This provides infrastructure to have multiple xdp_info instances
for the same consumer index.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
| |
Instead of calculating the control segment to be used upon an
XDP xmit doorbell, save it in SQ structure.
Nullify when no pending doorbell.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
| |
Do not ignore the CQE opcode.
This helps expose issues and debug them.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Do not maintain an SQ state bit to indicate whether an
XDP SQ serves redirect operations.
Instead, rely on the fact that such an XDP SQ doesn't reside
in an RQ instance, while the others do.
This info is not known to the XDP SQ functions themselves,
and they rely on their callers to distinguish between the cases.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
program check
At the end of the RQ polling loop, some XDP-related operations
might be required. Before checking them one by one, check if
an XDP program is even loaded.
Combine all the checks and operations in a single function in xdp files.
This saves unnecessary checks for non-XDP flows.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
| |
In preparation to enabling -Wimplicit-fallthrough, mark switch cases
where we are expecting to fall through.
Addresses-Coverity-ID: 114808 ("Missing break in switch")
Addresses-Coverity-ID: 114802 ("Missing break in switch")
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Commit 58b99ee3e3eb ("net/mlx5e: Add support for XDP_REDIRECT in device-out side")
forgot to return/free the xdp_frame in case the DMA mapping failed, correct this.
Also DMA unmap the frame in case mlx5e_xmit_xdp_frame() fails.
Fixes: 58b99ee3e3eb ("net/mlx5e: Add support for XDP_REDIRECT in device-out side")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add implementation for the ndo_xdp_xmit callback.
Dedicate a new set of XDP-SQ instances to satisfy the XDP_REDIRECT
requests. These instances are totally separated from the existing
XDP-SQ objects that satisfy local XDP_TX actions.
Performance tests:
xdp_redirect_map from ConnectX-5 to ConnectX-5.
CPU: Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
Packet-rate of 64B packets.
Single queue: 7 Mpps.
Multi queue: 55 Mpps.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Eugenia Emantayev <eugenia@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>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Separate the XDP counters into two sets:
(1) One set reside in the RQ stats, and they monitor XDP stats
in the RQ side.
(2) Another set is per XDP-SQ, and they monitor XDP stats that
are related to XDP transmit flow.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Eugenia Emantayev <eugenia@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
|
|
|
| |
Convert the XDP xmit functions to use the generic xdp_frame API
in XDP_TX flow.
Same functions will be used later in this series to transmit
the XDP redirect-out packets as well.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Eugenia Emantayev <eugenia@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
|
|
|
|
|
|
|
|
| |
Add per-ring and total stats for received packets that
goes into XDP redirection.
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>
|