Kuniyuki Iwashima says:
====================
af_unix: Fix lockless access of sk->sk_state and others fields.
The patch 1 fixes a bug where SOCK_DGRAM's sk->sk_state is changed
to TCP_CLOSE even if the socket is connect()ed to another socket.
The rest of this series annotates lockless accesses to the following
fields.
* sk->sk_state
* sk->sk_sndbuf
* net->unx.sysctl_max_dgram_qlen
* sk->sk_receive_queue.qlen
* sk->sk_shutdown
Note that with this series there is skb_queue_empty() left in
unix_dgram_disconnected() that needs to be changed to lockless
version, and unix_peer(other) access there should be protected
by unix_state_lock().
This will require some refactoring, so another series will follow.
Changes:
v2:
* Patch 1: Fix wrong double lock
v1: https://lore.kernel.org/netdev/20240603143231.62085-1-kuniyu@amazon.com/
====================
Link: https://lore.kernel.org/r/20240604165241.44758-1-kuniyu@amazon.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
While dumping sockets via UNIX_DIAG, we do not hold unix_state_lock().
Let's use READ_ONCE() to read sk->sk_shutdown.
Fixes: e4e541a848 ("sock-diag: Report shutdown for inet and unix sockets (v2)")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We can dump the socket queue length via UNIX_DIAG by specifying
UDIAG_SHOW_RQLEN.
If sk->sk_state is TCP_LISTEN, we return the recv queue length,
but here we do not hold recvq lock.
Let's use skb_queue_len_lockless() in sk_diag_show_rqlen().
Fixes: c9da99e647 ("unix_diag: Fixup RQLEN extension report")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
If the socket type is SOCK_STREAM or SOCK_SEQPACKET, unix_release_sock()
checks the length of the peer socket's recvq under unix_state_lock().
However, unix_stream_read_generic() calls skb_unlink() after releasing
the lock. Also, for SOCK_SEQPACKET, __skb_try_recv_datagram() unlinks
skb without unix_state_lock().
Thues, unix_state_lock() does not protect qlen.
Let's use skb_queue_empty_lockless() in unix_release_sock().
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Once sk->sk_state is changed to TCP_LISTEN, it never changes.
unix_accept() takes advantage of this characteristics; it does not
hold the listener's unix_state_lock() and only acquires recvq lock
to pop one skb.
It means unix_state_lock() does not prevent the queue length from
changing in unix_stream_connect().
Thus, we need to use unix_recvq_full_lockless() to avoid data-race.
Now we remove unix_recvq_full() as no one uses it.
Note that we can remove READ_ONCE() for sk->sk_max_ack_backlog in
unix_recvq_full_lockless() because of the following reasons:
(1) For SOCK_DGRAM, it is a written-once field in unix_create1()
(2) For SOCK_STREAM and SOCK_SEQPACKET, it is changed under the
listener's unix_state_lock() in unix_listen(), and we hold
the lock in unix_stream_connect()
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
net->unx.sysctl_max_dgram_qlen is exposed as a sysctl knob and can be
changed concurrently.
Let's use READ_ONCE() in unix_create1().
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
sk_setsockopt() changes sk->sk_sndbuf under lock_sock(), but it's
not used in af_unix.c.
Let's use READ_ONCE() to read sk->sk_sndbuf in unix_writable(),
unix_dgram_sendmsg(), and unix_stream_sendmsg().
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
While dumping AF_UNIX sockets via UNIX_DIAG, sk->sk_state is read
locklessly.
Let's use READ_ONCE() there.
Note that the result could be inconsistent if the socket is dumped
during the state change. This is common for other SOCK_DIAG and
similar interfaces.
Fixes: c9da99e647 ("unix_diag: Fixup RQLEN extension report")
Fixes: 2aac7a2cb0 ("unix_diag: Pending connections IDs NLA")
Fixes: 45a96b9be6 ("unix_diag: Dumping all sockets core")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
unix_stream_read_skb() is called from sk->sk_data_ready() context
where unix_state_lock() is not held.
Let's use READ_ONCE() there.
Fixes: 77462de14a ("af_unix: Add read_sock for stream socket types")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The following functions read sk->sk_state locklessly and proceed only if
the state is TCP_ESTABLISHED.
* unix_stream_sendmsg
* unix_stream_read_generic
* unix_seqpacket_sendmsg
* unix_seqpacket_recvmsg
Let's use READ_ONCE() there.
Fixes: a05d2ad1c1 ("af_unix: Only allow recv on connected seqpacket sockets.")
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Once sk->sk_state is changed to TCP_LISTEN, it never changes.
unix_accept() takes the advantage and reads sk->sk_state without
holding unix_state_lock().
Let's use READ_ONCE() there.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
As small optimisation, unix_stream_connect() prefetches the client's
sk->sk_state without unix_state_lock() and checks if it's TCP_CLOSE.
Later, sk->sk_state is checked again under unix_state_lock().
Let's use READ_ONCE() for the first check and TCP_CLOSE directly for
the second check.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
unix_poll() and unix_dgram_poll() read sk->sk_state locklessly and
calls unix_writable() which also reads sk->sk_state without holding
unix_state_lock().
Let's use READ_ONCE() in unix_poll() and unix_dgram_poll() and pass
it to unix_writable().
While at it, we remove TCP_SYN_SENT check in unix_dgram_poll() as
that state does not exist for AF_UNIX socket since the code was added.
Fixes: 1586a5877d ("af_unix: do not report POLLOUT on listeners")
Fixes: 3c73419c09 ("af_unix: fix 'poll for write'/ connected DGRAM sockets")
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
ioctl(SIOCINQ) calls unix_inq_len() that checks sk->sk_state first
and returns -EINVAL if it's TCP_LISTEN.
Then, for SOCK_STREAM sockets, unix_inq_len() returns the number of
bytes in recvq.
However, unix_inq_len() does not hold unix_state_lock(), and the
concurrent listen() might change the state after checking sk->sk_state.
If the race occurs, 0 is returned for the listener, instead of -EINVAL,
because the length of skb with embryo is 0.
We could hold unix_state_lock() in unix_inq_len(), but it's overkill
given the result is true for pre-listen() TCP_CLOSE state.
So, let's use READ_ONCE() for sk->sk_state in unix_inq_len().
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
sk->sk_state is changed under unix_state_lock(), but it's read locklessly
in many places.
This patch adds WRITE_ONCE() on the writer side.
We will add READ_ONCE() to the lockless readers in the following patches.
Fixes: 83301b5367 ("af_unix: Set TCP_ESTABLISHED for datagram sockets too")
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When a SOCK_DGRAM socket connect()s to another socket, the both sockets'
sk->sk_state are changed to TCP_ESTABLISHED so that we can register them
to BPF SOCKMAP.
When the socket disconnects from the peer by connect(AF_UNSPEC), the state
is set back to TCP_CLOSE.
Then, the peer's state is also set to TCP_CLOSE, but the update is done
locklessly and unconditionally.
Let's say socket A connect()ed to B, B connect()ed to C, and A disconnects
from B.
After the first two connect()s, all three sockets' sk->sk_state are
TCP_ESTABLISHED:
$ ss -xa
Netid State Recv-Q Send-Q Local Address:Port Peer Address:PortProcess
u_dgr ESTAB 0 0 @A 641 * 642
u_dgr ESTAB 0 0 @B 642 * 643
u_dgr ESTAB 0 0 @C 643 * 0
And after the disconnect, B's state is TCP_CLOSE even though it's still
connected to C and C's state is TCP_ESTABLISHED.
$ ss -xa
Netid State Recv-Q Send-Q Local Address:Port Peer Address:PortProcess
u_dgr UNCONN 0 0 @A 641 * 0
u_dgr UNCONN 0 0 @B 642 * 643
u_dgr ESTAB 0 0 @C 643 * 0
In this case, we cannot register B to SOCKMAP.
So, when a socket disconnects from the peer, we should not set TCP_CLOSE to
the peer if the peer is connected to yet another socket, and this must be
done under unix_state_lock().
Note that we use WRITE_ONCE() for sk->sk_state as there are many lockless
readers. These data-races will be fixed in the following patches.
Fixes: 83301b5367 ("af_unix: Set TCP_ESTABLISHED for datagram sockets too")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This list is used to tranfert dst that are handled by
rt_flush_dev() and rt6_uncached_list_flush_dev() out
of the per-cpu lists.
But quarantine list is not used later.
If we simply use list_del_init(&rt->dst.rt_uncached),
this also removes the dst from per-cpu list.
This patch also makes the future calls to rt_del_uncached_list()
and rt6_uncached_list_del() faster, because no spinlock
acquisition is needed anymore.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/20240604165150.726382-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Toke mentioned unrcu_pointer() existence, allowing
to remove some of the ugly casts we have when using
xchg() for rcu protected pointers.
Also make inet_rcv_compat const.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Toke Høiland-Jørgensen <toke@redhat.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20240604111603.45871-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
In case of region creation fail in ipc_devlink_create_region(), previously
created regions delete process starts from tainted pointer which actually
holds error code value.
Fix this bug by decreasing region index before delete.
Found by Linux Verification Center (linuxtesting.org) with SVACE.
Fixes: 4dcd183fbd ("net: wwan: iosm: devlink registration")
Signed-off-by: Aleksandr Mishin <amishin@t-argos.ru>
Acked-by: Sergey Ryazanov <ryazanov.s.a@gmail.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/20240604082500.20769-1-amishin@t-argos.ru
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Paul Barker says:
====================
Improve GbEth performance on Renesas RZ/G2L and related SoCs
This series aims to improve performance of the GbEth IP in the Renesas
RZ/G2L SoC family and the RZ/G3S SoC, which use the ravb driver. Along
the way, we do some refactoring and ensure that napi_complete_done() is
used in accordance with the NAPI documentation for both GbEth and R-Car
code paths.
Much of the performance improvement comes from enabling SW IRQ
Coalescing for all SoCs using the GbEth IP, and NAPI Threaded mode for
single core SoCs using the GbEth IP. These can be enabled/disabled at
runtime via sysfs, but our goal is to set sensible defaults which get
good performance on the affected SoCs.
The rest of the performance improvement comes from using a page pool to
allocate RX buffers, and reducing the allocation size from >8kB to 2kB.
The overall performance impact of this patch series seen in testing with
iperf3 is as follows (see patches 5-7 for more detailed results):
* RZ/G2L:
* TCP TX: +1.8% bandwidth
* TCP RX: +1% bandwidth at 47% less CPU load
* UDP RX: +1% bandwidth at 26% less CPU load
* RZ/G2UL:
* TCP TX: +37% bandwidth
* TCP RX: +43% bandwidth
* UDP TX: -8% bandwidth
* UDP RX: +32500% bandwidth (!)
* RZ/G3S:
* TCP TX: +25% bandwidth
* TCP RX: +76% bandwidth
* UDP TX: -9% bandwidth
* UDP RX: +37900% bandwidth (!)
* RZ/Five:
* TCP TX: +18% bandwidth
* TCP RX: +212% bandwidth
* UDP TX: +2% bandwidth
* UDP RX: +inf bandwidth (test no longer crashes)
There is no significant impact on bandwidth or CPU load in testing on
RZ/G2H or R-Car M3N.
Fixing the crash in UDP RX testing for RZ/Five is a cumulative effect of
patches 1, 2, 5 & 6 so this is very difficult to break out as a bugfix
for backporting.
Changes v4->v5:
* Added Sergey's Reviewed-by tags.
* Improved the commit message for patch 2/7.
* Re-wrapped to 80 cols, except where this would significantly impact
readability.
* Use lower case `skb` consistently in comments.
* Included <net/page_pool/types.h> in ravb.h.
* Moved rx_buffer_size so it is in the same place in ravb_hw_info as
rx_max_desc_use was previously.
* Used reverse xmas tree ordering in variable declarations.
* Split lines after binary operators, instead of before.
* Factor subtraction of sizeof(__sum16) out of the if condition in
ravb_rx_csum_gbeth().
* Add blank lines after variable declarations where needed.
* Used goto instead of break to handle napi_build_skb() failure in
ravb_rx_gbeth(). Break was incorrectly scoped to the surrounding
switch statement, when it's the outer loop we really want to break
out of.
* Used continue instead of break to handle NULL priv->rx_1st_skb in
ravb_rx_gbeth() as we may still be able to process further
descriptors.
* Unconditionally set priv->rx_1st_skb = NULL after processing a
packet in ravb_rx_gbeth(). We don't need to check die_dt as this
will be a no-op for single descriptor packets.
* Moved napi_build_skb() call after dma_sync_single_for_cpu() in
ravb_rx_rcar() to align the order of operations with ravb_rx_gbeth()
and ensure the data is sync'd before it is accessed.
* Moved zeroing of rx_buff->page to the end of packet processing in
ravb_rx_rcar() to align the order of operations with
ravb_rx_gbeth().
Changes v3->v4:
* Dependency patches have merged so this is no longer an RFC.
* Fixed update of stats->rx_packets.
* Simplified refactoring following feedback from Niklas and Sergey.
* Renamed needs_irq_coalesce -> coalesce_irqs.
* Used a separate page pool for each RX queue.
* Passed struct ravb_rx_desc to ravb_alloc_rx_buffer() so that we can
simplify the calling function.
* Explained the calculation of rx_desc->ds_cc.
* Added handling of nonlinear SKBs in ravb_rx_csum_gbeth().
* Used Niklas' suggested commit message for patch 2/7.
* Added Sergey's Reviewed-by tags to patches 5/7 and 6/7.
Changes v2->v3:
* Incorporated feedback on RFC v2 from Sergey.
* Split out bugfixes and rebased. This changed the order of what was
the first 5 patches of v2 and things look a little different so I've
not picked up Reviewed-by tags from v2.
* Further refactoring and tidy up of RX ring refill and
ravb_rx_gbeth().
* Switched to using a page pool to allocate RX buffers.
* Re-tested and provided updated performance figures.
Changes v1->v2:
* Marked as RFC as the series depends on unmerged patches.
* Refactored R-Car code paths as well as GbEth code paths.
* Updated references to the patches this series depends on.
====================
Link: https://lore.kernel.org/r/20240604072825.7490-1-paul.barker.ct@bp.renesas.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This patch makes multiple changes that can't be separated:
1) Allocate plain RX buffers via a page pool instead of allocating
SKBs, then use build_skb() when a packet is received.
2) For GbEth IP, reduce the RX buffer size to 2kB.
3) For GbEth IP, merge packets which span more than one RX descriptor
as SKB fragments instead of copying data.
Implementing (1) without (2) would require the use of an order-1 page
pool (instead of an order-0 page pool split into page fragments) for
GbEth.
Implementing (2) without (3) would leave us no space to re-assemble
packets which span more than one RX descriptor.
Implementing (3) without (1) would not be possible as the network stack
expects to use put_page() or page_pool_put_page() to free SKB fragments
after an SKB is consumed.
RX checksum offload support is adjusted to handle both linear and
nonlinear (fragmented) packets.
This patch gives the following improvements during testing with iperf3.
* RZ/G2L:
* TCP RX: same bandwidth at -43% CPU load (70% -> 40%)
* UDP RX: same bandwidth at -17% CPU load (88% -> 74%)
* RZ/G2UL:
* TCP RX: +30% bandwidth (726Mbps -> 941Mbps)
* UDP RX: +417% bandwidth (108Mbps -> 558Mbps)
* RZ/G3S:
* TCP RX: +64% bandwidth (562Mbps -> 920Mbps)
* UDP RX: +420% bandwidth (90Mbps -> 468Mbps)
* RZ/Five:
* TCP RX: +217% bandwidth (145Mbps -> 459Mbps)
* UDP RX: +470% bandwidth (20Mbps -> 114Mbps)
There is no significant impact on bandwidth or CPU load in testing on
RZ/G2H or R-Car M3N.
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
NAPI Threaded mode (along with the previously enabled SW IRQ Coalescing)
is required to improve network stack performance for single core SoCs
using the GbEth IP (currently the RZ/G2L SoC family and the RZ/G3S SoC).
This patch gives the following improvements during testing with iperf3.
* RZ/G2UL:
* TCP TX: +32% bandwidth (638Mbps -> 841Mbps)
* TXP RX: +8.8% bandwidth (667Mbps -> 726Mbps)
* UDP RX: +104% bandwidth (53Mbps -> 108Mbps)
* RZ/G3S:
* TCP TX: 29% bandwidth (529Mbps -> 681Mbps)
* UDP RX: +1290% bandwidth (6.46Mbps -> 90Mbps)
* RZ/Five:
* UDP RX: Test no longer crashes (0 -> 20 Mbps)
This patch gives the following reductions in performance in the same
testing:
* RZ/G2UL:
* UDP TX: -7.5% bandwidth (594Mbps -> 549Mbps)
* RZ/G3S:
* UDP TX: -5% bandwidth (625Mbps -> 594Mbps)
These losses are considered acceptable given the benefits shown above.
If UDP TX bandwidth must be maximised for a particular use case, NAPI
threaded mode can be disabled at runtime via sysfs writes.
The improvement of UDP RX bandwidth for the single core SoCs (RZ/G2UL &
RZ/G3S) is particularly critical.
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Software IRQ Coalescing is required to improve network stack performance
in the RZ/G2L SoC family and the RZ/G3S SoC, i.e. the SoCs which use the
GbEth IP.
This patch gives the following improvements during testing with iperf3:
* RZ/G2L:
* TCP RX: same bandwidth with -6% CPU load (76% -> 71%)
* UDP RX: same bandwidth with -10% CPU load (99% -> 89%)
* RZ/G2UL:
* UDP RX: +4200% bandwidth (1.23Mbps -> 53Mbps)
* RZ/G3S:
* UDP RX: +425% bandwidth (1.23Mbps -> 6.46Mbps)
The improvement of UDP RX bandwidth for the single core SoCs (RZ/G2UL &
RZ/G3S) is particularly critical.
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We can reduce code duplication in ravb_rx_gbeth().
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
To reduce code duplication, we add a new RX ring refill function which
can handle both the initial RX ring population (which was split between
ravb_ring_init() and ravb_ring_format()) and the RX ring refill after
polling (in ravb_rx()).
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Align ravb_poll() with the documentation in
`Documentation/networking/kapi.rst` and
`Documentation/networking/napi.rst`.
The documentation says that we should prefer napi_complete_done() over
napi_complete(), and using the former allows us to properly support busy
polling. We should ensure that napi_complete_done() is only called if
the work budget has not been exhausted, and we should only re-arm
interrupts if it returns true.
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We don't need to pass the work budget to ravb_rx() by reference, it's
cleaner to pass this by value and return the amount of work done. This
allows us to simplify the ravb_poll() function and use the common
`work_done` variable name seen in other network drivers for consistency
and ease of understanding.
This is a pure refactor and should not affect behaviour.
Signed-off-by: Paul Barker <paul.barker.ct@bp.renesas.com>
Reviewed-by: Sergey Shtylyov <s.shtylyov@omp.ru>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Tariq Toukan says:
====================
net/mlx5e: SHAMPO, Enable HW GRO once more
This series enables hardware GRO for ConnectX-7 and newer NICs.
SHAMPO stands for Split Header And Merge Payload Offload.
The first part of the series contains important fixes and improvements.
The second part reworks the HW GRO counters.
Lastly, HW GRO is perf optimized and enabled.
Here are the bandwidth numbers for a simple iperf3 test over a single rq
where the application and irq are pinned to the same CPU:
+---------+--------+--------+-----------+-------------+
| streams | SW GRO | HW GRO | Unit | Improvement |
+---------+--------+--------+-----------+-------------+
| 1 | 36 | 57 | Gbits/sec | 1.6 x |
| 4 | 34 | 50 | Gbits/sec | 1.5 x |
| 8 | 31 | 43 | Gbits/sec | 1.4 x |
+---------+--------+--------+-----------+-------------+
Benchmark details:
VM based setup
CPU: Intel(R) Xeon(R) Platinum 8380 CPU, 24 cores
NIC: ConnectX-7 100GbE
iperf3 and irq running on same CPU over a single receive queue
====================
Link: https://lore.kernel.org/r/20240603212219.1037656-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When doing hardware GRO (SHAMPO), the driver puts each data payload of a
packet from the wire into one skb fragment. TCP Zero-Copy expects page
sized skb fragments to be able to do it's page-flipping magic. With the
current way of arranging fragments by the driver, only specific MTUs
(page sized multiple + header size) will yield such page sized fragments
in a high percentage.
This change improves payload arrangement in the skb for hardware GRO by
coalescing payloads into a single skb fragment when possible.
To demonstrate the fix, running tcp_mmap with a MTU of 1500 yields:
- Before: 0 % bytes mmap'ed
- After : 81 % bytes mmap'ed
More importantly, coalescing considerably improves the HW GRO performance.
Here are the results for a iperf3 bandwidth benchmark:
+---------+--------+--------+------------------------+-----------+
| streams | SW GRO | HW GRO | HW GRO with coalescing | Unit |
|---------+--------+--------+------------------------+-----------|
| 1 | 36 | 42 | 57 | Gbits/sec |
| 4 | 34 | 39 | 50 | Gbits/sec |
| 8 | 31 | 35 | 43 | Gbits/sec |
+---------+--------+--------+------------------------+-----------+
Benchmark details:
VM based setup
CPU: Intel(R) Xeon(R) Platinum 8380 CPU, 24 cores
NIC: ConnectX-7 100GbE
iperf3 and irq running on same CPU over a single receive queue
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-15-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add back HW-GRO to the reported features.
As the current implementation of HW-GRO uses KSMs with a
specific fixed buffer size (256B) to map its headers buffer,
we reported the feature only if the NIC is supporting KSM and
the minimum value for buffer size is below the requested one.
iperf3 bandwidth comparison:
+---------+--------+--------+-----------+
| streams | SW GRO | HW GRO | Unit |
|---------+--------+--------+-----------|
| 1 | 36 | 42 | Gbits/sec |
| 4 | 34 | 39 | Gbits/sec |
| 8 | 31 | 35 | Gbits/sec |
+---------+--------+--------+-----------+
A downstream patch will add skb fragment coalescing which will improve
performance considerably.
Benchmark details:
VM based setup
CPU: Intel(R) Xeon(R) Platinum 8380 CPU, 24 cores
NIC: ConnectX-7 100GbE
iperf3 and irq running on same CPU over a single receive queue
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-14-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
KSM Mkey is KLM Mkey with a fixed buffer size. Due to this fact,
it is a faster mechanism than KLM.
SHAMPO feature used KLMs Mkeys for memory mappings of its headers buffer.
As it used KLMs with the same buffer size for each entry,
we can use KSMs instead.
This commit changes the Mkeys that map the SHAMPO headers buffer
from KLMs to KSMs.
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-13-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Count the number of header-only packets and bytes from SHAMPO.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-12-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After modifying rx_gro_packets to be more accurate, the
rx_gro_match_packets counter is redundant.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-11-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Don't count non GRO packets. A non GRO packet is a packet with
a GRO cb count of 1.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-10-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
SHAMPO SKB can be flushed in mlx5e_shampo_complete_rx_cqe().
If the SKB was flushed, rq->hw_gro_data->skb was also set to NULL.
We can skip on flushing the SKB in mlx5e_shampo_flush_skb
if rq->hw_gro_data->skb == NULL.
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-9-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
mlx5e_fill_skb_data() used to have multiple callers. But after the XDP
multibuf refactoring from commit 2cb0e27d43 ("net/mlx5e: RX, Prepare
non-linear striding RQ for XDP multi-buffer support") the SHAMPO code
path is the only caller.
Take advantage of this and specialize the function:
- Drop the redundant check.
- Assume that data_bcnt is > 0. This is needed in a downstream patch.
Rename the function as well to make things clear.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Suggested-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-8-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The function that releases SHAMPO header pages (mlx5e_shampo_dealloc_hd)
has some complicated logic that comes from the fact that it is called
twice during teardown:
1) To release the posted header pages that didn't get any completions.
2) To release all remaining header pages.
This flow is not necessary: all header pages can be released from the
driver side in one go. Furthermore, the above flow is buggy. Taking the
8 headers per page example:
1) Release fragments 5-7. Page will be released.
2) Release remaining fragments 0-4. The bits in the header will indicate
that the page needs releasing. But this is incorrect: page was
released in step 1.
This patch releases all header pages in one go. This simplifies the
header page cleanup function. For consistency, the datapath header
page release API (mlx5e_free_rx_shampo_hd_entry()) is used.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-7-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When HW GRO is enabled, forwarding of packets is broken due to gso_size
being set incorrectly on non GRO packets.
Non GRO packets have a skb GRO count of 1. mlx5 always sets gso_size on
the skb, even for non GRO packets. It leans on the fact that gso_size is
normally reset in napi_gro_complete(). But this happens only for packets
from GRO'able protocols (TCP/UDP) that have a gro_receive() handler.
The problematic scenarios are:
1) Non GRO protocol packets are received, validate_xmit_skb() will drop
them (see EPROTONOSUPPORT in skb_mac_gso_segment()). The fix for
this case would be to not set gso_size at all for SHAMPO packets with
header size 0.
2) Packets from a GRO'ed protocol (TCP) are received but immediately
flushed because they are not GRO'able (TCP SYN for example).
mlx5e_shampo_update_hdr(), which updates the remaining GRO state on
the skb, is not called because skb GRO count is 1. The fix here would
be to always call mlx5e_shampo_update_hdr(), regardless of skb GRO
count. But this call is expensive
The unified fix for both cases is to reset gso_size before calling
napi_gro_receive(). It is a change that is more effective (no call to
mlx5e_shampo_update_hdr() necessary) and simple (smallest code
footprint).
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-6-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For the following scenario:
ethtool --features eth3 rx-gro-hw on
ethtool --features eth3 rx-fcs on
ethtool --features eth3 rx-fcs off
... there is a firmware error because the driver enables HW GRO first
while FCS is still enabled.
This patch fixes this by swapping the order of HW GRO and FCS for this
specific case. Take LRO into consideration as well for consistency.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-5-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When all the strides in a WQE have been consumed, the WQE is unlinked
from the WQ linked list (mlx5_wq_ll_pop()). For SHAMPO, it is possible
to receive CQEs with 0 consumed strides for the same WQE even after the
WQE is fully consumed and unlinked. This triggers an additional unlink
for the same wqe which corrupts the linked list.
Fix this scenario by accepting 0 sized consumed strides without
unlinking the WQE again.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-4-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Under the following conditions:
1) No skb created yet
2) header_size == 0 (no SHAMPO header)
3) header_index + 1 % MLX5E_SHAMPO_WQ_HEADER_PER_PAGE == 0 (this is the
last page fragment of a SHAMPO header page)
a new skb is formed with a page that is NOT a SHAMPO header page (it
is a regular data page). Further down in the same function
(mlx5e_handle_rx_cqe_mpwrq_shampo()), a SHAMPO header page from
header_index is released. This is wrong and it leads to SHAMPO header
pages being released more than once.
Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240603212219.1037656-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jacob Keller says:
====================
Intel Wired LAN Driver Updates 2024-05-29 (ice, igc)
This series includes fixes for the ice driver as well as a fix for the igc
driver.
Jacob fixes two issues in the ice driver with reading the NVM for providing
firmware data via devlink info. First, fix an off-by-one error when reading
the Preserved Fields Area, resolving an infinite loop triggered on some
NVMs which lack certain data in the NVM. Second, fix the reading of the NVM
Shadow RAM on newer E830 and E825-C devices which have a variable sized CSS
header rather than assuming this header is always the same fixed size as in
the E810 devices.
Larysa fixes three issues with the ice driver XDP logic that could occur if
the number of queues is changed after enabling an XDP program. First, the
af_xdp_zc_qps bitmap is removed and replaced by simpler logic to track
whether queues are in zero-copy mode. Second, the reset and .ndo_bpf flows
are distinguished to avoid potential races with a PF reset occuring
simultaneously to .ndo_bpf callback from userspace. Third, the logic for
mapping XDP queues to vectors is fixed so that XDP state is restored for
XDP queues after a reconfiguration.
Sasha fixes reporting of Energy Efficient Ethernet support via ethtool in
the igc driver.
v1: https://lore.kernel.org/r/20240530-net-2024-05-30-intel-net-fixes-v1-0-8b11c8c9bff8@intel.com
====================
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-0-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The commit 01cf893bf0 ("net: intel: i40e/igc: Remove setting Autoneg in
EEE capabilities") removed SUPPORTED_Autoneg field but left inappropriate
ethtool_keee structure initialization. When "ethtool --show <device>"
(get_eee) invoke, the 'ethtool_keee' structure was accidentally overridden.
Remove the 'ethtool_keee' overriding and add EEE declaration as per IEEE
specification that allows reporting Energy Efficient Ethernet capabilities.
Examples:
Before fix:
ethtool --show-eee enp174s0
EEE settings for enp174s0:
EEE status: not supported
After fix:
EEE settings for enp174s0:
EEE status: disabled
Tx LPI: disabled
Supported EEE link modes: 100baseT/Full
1000baseT/Full
2500baseT/Full
Fixes: 01cf893bf0 ("net: intel: i40e/igc: Remove setting Autoneg in EEE capabilities")
Suggested-by: Dima Ruinskiy <dima.ruinskiy@intel.com>
Signed-off-by: Sasha Neftin <sasha.neftin@intel.com>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-6-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ice_pf_dcb_recfg() re-maps queues to vectors with
ice_vsi_map_rings_to_vectors(), which does not restore the previous
state for XDP queues. This leads to no AF_XDP traffic after rebuild.
Map XDP queues to vectors in ice_vsi_map_rings_to_vectors().
Also, move the code around, so XDP queues are mapped independently only
through .ndo_bpf().
Fixes: 6624e780a5 ("ice: split ice_vsi_setup into smaller functions")
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Signed-off-by: Larysa Zaremba <larysa.zaremba@intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-5-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 6624e780a5 ("ice: split ice_vsi_setup into smaller functions")
has placed ice_vsi_free_q_vectors() after ice_destroy_xdp_rings() in
the rebuild process. The behaviour of the XDP rings config functions is
context-dependent, so the change of order has led to
ice_destroy_xdp_rings() doing additional work and removing XDP prog, when
it was supposed to be preserved.
Also, dependency on the PF state reset flags creates an additional,
fortunately less common problem:
* PFR is requested e.g. by tx_timeout handler
* .ndo_bpf() is asked to delete the program, calls ice_destroy_xdp_rings(),
but reset flag is set, so rings are destroyed without deleting the
program
* ice_vsi_rebuild tries to delete non-existent XDP rings, because the
program is still on the VSI
* system crashes
With a similar race, when requested to attach a program,
ice_prepare_xdp_rings() can actually skip setting the program in the VSI
and nevertheless report success.
Instead of reverting to the old order of function calls, add an enum
argument to both ice_prepare_xdp_rings() and ice_destroy_xdp_rings() in
order to distinguish between calls from rebuild and .ndo_bpf().
Fixes: efc2214b60 ("ice: Add support for XDP")
Reviewed-by: Igor Bagnucki <igor.bagnucki@intel.com>
Signed-off-by: Larysa Zaremba <larysa.zaremba@intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-4-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Referenced commit has introduced a bitmap to distinguish between ZC and
copy-mode AF_XDP queues, because xsk_get_pool_from_qid() does not do this
for us.
The bitmap would be especially useful when restoring previous state after
rebuild, if only it was not reallocated in the process. This leads to e.g.
xdpsock dying after changing number of queues.
Instead of preserving the bitmap during the rebuild, remove it completely
and distinguish between ZC and copy-mode queues based on the presence of
a device associated with the pool.
Fixes: e102db780e ("ice: track AF_XDP ZC enabled queues in bitmap")
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Signed-off-by: Larysa Zaremba <larysa.zaremba@intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Chandan Kumar Rout <chandanx.rout@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-3-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The ice driver reads data from the Shadow RAM portion of the NVM during
initialization, including data used to identify the NVM image and device,
such as the ETRACK ID used to populate devlink dev info fw.bundle.
Currently it is using a fixed offset defined by ICE_CSS_HEADER_LENGTH to
compute the appropriate offset. This worked fine for E810 and E822 devices
which both have CSS header length of 330 words.
Other devices, including both E825-C and E830 devices have different sizes
for their CSS header. The use of a hard coded value results in the driver
reading from the wrong block in the NVM when attempting to access the
Shadow RAM copy. This results in the driver reporting the fw.bundle as 0x0
in both the devlink dev info and ethtool -i output.
The first E830 support was introduced by commit ba20ecb1d1 ("ice: Hook up
4 E830 devices by adding their IDs") and the first E825-C support was
introducted by commit f64e189442 ("ice: introduce new E825C devices
family")
The NVM actually contains the CSS header length embedded in it. Remove the
hard coded value and replace it with logic to read the length from the NVM
directly. This is more resilient against all existing and future hardware,
vs looking up the expected values from a table. It ensures the driver will
read from the appropriate place when determining the ETRACK ID value used
for populating the fw.bundle_id and for reporting in ethtool -i.
The CSS header length for both the active and inactive flash bank is stored
in the ice_bank_info structure to avoid unnecessary duplicate work when
accessing multiple words of the Shadow RAM. Both banks are read in the
unlikely event that the header length is different for the NVM in the
inactive bank, rather than being different only by the overall device
family.
Fixes: ba20ecb1d1 ("ice: Hook up 4 E830 devices by adding their IDs")
Co-developed-by: Paul Greenwalt <paul.greenwalt@intel.com>
Signed-off-by: Paul Greenwalt <paul.greenwalt@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-2-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The ice_get_pfa_module_tlv() function iterates over the Type-Length-Value
structures in the Preserved Fields Area (PFA) of the NVM. This is used by
the driver to access data such as the Part Board Assembly identifier.
The function uses simple logic to iterate over the PFA. First, the pointer
to the PFA in the NVM is read. Then the total length of the PFA is read
from the first word.
A pointer to the first TLV is initialized, and a simple loop iterates over
each TLV. The pointer is moved forward through the NVM until it exceeds the
PFA area.
The logic seems sound, but it is missing a key detail. The Preserved
Fields Area length includes one additional final word. This is documented
in the device data sheet as a dummy word which contains 0xFFFF. All NVMs
have this extra word.
If the driver tries to scan for a TLV that is not in the PFA, it will read
past the size of the PFA. It reads and interprets the last dummy word of
the PFA as a TLV with type 0xFFFF. It then reads the word following the PFA
as a length.
The PFA resides within the Shadow RAM portion of the NVM, which is
relatively small. All of its offsets are within a 16-bit size. The PFA
pointer and TLV pointer are stored by the driver as 16-bit values.
In almost all cases, the word following the PFA will be such that
interpreting it as a length will result in 16-bit arithmetic overflow. Once
overflowed, the new next_tlv value is now below the maximum offset of the
PFA. Thus, the driver will continue to iterate the data as TLVs. In the
worst case, the driver hits on a sequence of reads which loop back to
reading the same offsets in an endless loop.
To fix this, we need to correct the loop iteration check to account for
this extra word at the end of the PFA. This alone is sufficient to resolve
the known cases of this issue in the field. However, it is plausible that
an NVM could be misconfigured or have corrupt data which results in the
same kind of overflow. Protect against this by using check_add_overflow
when calculating both the maximum offset of the TLVs, and when calculating
the next_tlv offset at the end of each loop iteration. This ensures that
the driver will not get stuck in an infinite loop when scanning the PFA.
Fixes: e961b679fb ("ice: add board identifier info to devlink .info_get")
Co-developed-by: Paul Greenwalt <paul.greenwalt@intel.com>
Signed-off-by: Paul Greenwalt <paul.greenwalt@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20240603-net-2024-05-30-intel-net-fixes-v2-1-e3563aa89b0c@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
