Skip to content

L
linux
Switch branch/tag
  1. 06 Apr, 2021 2 commits
    • Eric Dumazet's avatar
      virtio_net: Do not pull payload in skb->head · 0f6925b3
      Eric Dumazet authored 
      Xuan Zhuo reported that commit 3226b158 ("net: avoid 32 x truesize
under-estimation for tiny skbs") brought  a ~10% performance drop.

The reason for the performance drop was that GRO was forced
to chain sk_buff (using skb_shinfo(skb)->frag_list), which
uses more memory but also cause packet consumers to go over
a lot of overhead handling all the tiny skbs.

It turns out that virtio_net page_to_skb() has a wrong strategy :
It allocates skbs with GOOD_COPY_LEN (128) bytes in skb->head, then
copies 128 bytes from the page, before feeding the packet to GRO stack.

This was suboptimal before commit 3226b158 ("net: avoid 32 x truesize
under-estimation for tiny skbs") because GRO was using 2 frags per MSS,
meaning we were not packing MSS with 100% efficiency.

Fix is to pull only the ethernet header in page_to_skb()

Then, we change virtio_net_hdr_to_skb() to pull the missing
headers, instead of assuming they were already pulled by callers.

This fixes the performance regression, but could also allow virtio_net
to accept packets with more than 128bytes of headers.

Many thanks to Xuan Zhuo for his report, and his tests/help.

Fixes: 3226b158 ("net: avoid 32 x truesize under-estimation for tiny skbs")
Reported-by: default avatarXuan Zhuo <xuanzhuo@linux.alibaba.com>
Link: https://www.spinics.net/lists/netdev/msg731397.htmlCo-Developed-by: default avatarXuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: default avatarXuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: virtualization@lists.linux-foundation.org
Acked-by: default avatarJason Wang <jasowang@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      0f6925b3
    • Lv Yunlong's avatar
      net: broadcom: bcm4908enet: Fix a double free in bcm4908_enet_dma_alloc · b25b343d
      Lv Yunlong authored 
      In bcm4908_enet_dma_alloc, if callee bcm4908_dma_alloc_buf_descs() failed,
it will free the ring->cpu_addr by dma_free_coherent() and return error.
Then bcm4908_enet_dma_free() will be called, and free the same cpu_addr
by dma_free_coherent() again.

My patch set ring->cpu_addr to NULL after it is freed in
bcm4908_dma_alloc_buf_descs() to avoid the double free.

Fixes: 4feffead ("net: broadcom: bcm4908enet: add BCM4908 controller driver")
Signed-off-by: default avatarLv Yunlong <lyl2019@mail.ustc.edu.cn>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      b25b343d
  3. 05 Apr, 2021 6 commits
  5. 02 Apr, 2021 3 commits
  7. 01 Apr, 2021 17 commits
  9. 31 Mar, 2021 12 commits
    • Ong Boon Leong's avatar
      xdp: fix xdp_return_frame() kernel BUG throw for page_pool memory model · 622d1369
      Ong Boon Leong authored 
      xdp_return_frame() may be called outside of NAPI context to return
xdpf back to page_pool. xdp_return_frame() calls __xdp_return() with
napi_direct = false. For page_pool memory model, __xdp_return() calls
xdp_return_frame_no_direct() unconditionally and below false negative
kernel BUG throw happened under preempt-rt build:

[  430.450355] BUG: using smp_processor_id() in preemptible [00000000] code: modprobe/3884
[  430.451678] caller is __xdp_return+0x1ff/0x2e0
[  430.452111] CPU: 0 PID: 3884 Comm: modprobe Tainted: G     U      E     5.12.0-rc2+ #45

Changes in v2:
 - This patch fixes the issue by making xdp_return_frame_no_direct() is
   only called if napi_direct = true, as recommended for better by
   Jesper Dangaard Brouer. Thanks!

Fixes: 2539650f ("xdp: Helpers for disabling napi_direct of xdp_return_frame")
Signed-off-by: default avatarOng Boon Leong <boon.leong.ong@intel.com>
Acked-by: default avatarJesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      622d1369
    • Eric Dumazet's avatar
      Revert "net: correct sk_acceptq_is_full()" · c609e6aa
      Eric Dumazet authored 
      This reverts commit f211ac15.

We had similar attempt in the past, and we reverted it.

History:

64a14651 [NET]: Revert incorrect accept queue backlog changes.
8488df89 [NET]: Fix bugs in "Whether sock accept queue is full" checking

I am adding a fat comment so that future attempts will
be much harder.

Fixes: f211ac15 ("net: correct sk_acceptq_is_full()")
Cc: iuyacan <yacanliu@163.com>
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      c609e6aa
    • David S. Miller's avatar
      Merge tag 'mlx5-fixes-2021-03-31' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux · 9dc22c0d
      David S. Miller authored 
      Saeed Mahameed says:

====================
mlx5 fixes 2021-03-31

This series introduces some fixes to mlx5 driver.
Please pull and let me know if there is any problem.
====================
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      9dc22c0d
    • David S. Miller's avatar
      Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/klassert/ipsec · c9170f13
      David S. Miller authored 
      Steffen Klassert says:

====================
pull request (net): ipsec 2021-03-31

1) Fix ipv4 pmtu checks for xfrm anf vti interfaces.
   From Eyal Birger.

2) There are situations where the socket passed to
   xfrm_output_resume() is not the same as the one
   attached to the skb. Use the socket passed to
   xfrm_output_resume() to avoid lookup failures
   when xfrm is used with VRFs.
   From Evan Nimmo.

3) Make the xfrm_state_hash_generation sequence counter per
   network namespace because but its write serialization
   lock is also per network namespace. Write protection
   is insufficient otherwise.
   From Ahmed S. Darwish.

4) Fixup sctp featue flags when used with esp offload.
   From Xin Long.

5) xfrm BEET mode doesn't support fragments for inner packets.
   This is a limitation of the protocol, so no fix possible.
   Warn at least to notify the user about that situation.
   From Xin Long.

6) Fix NULL pointer dereference on policy lookup when
   namespaces are uses in combination with esp offload.

7) Fix incorrect transformation on esp offload when
   packets get segmented at layer 3.

8) Fix some user triggered usages of WARN_ONCE in
   the xfrm compat layer.
   From Dmitry Safonov.

Please pull or let me know if there are problems.
====================
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      c9170f13
    • Lv Yunlong's avatar
      net/rds: Fix a use after free in rds_message_map_pages · bdc2ab5c
      Lv Yunlong authored 
      In rds_message_map_pages, the rm is freed by rds_message_put(rm).
But rm is still used by rm->data.op_sg in return value.

My patch assigns ERR_CAST(rm->data.op_sg) to err before the rm is
freed to avoid the uaf.

Fixes: 7dba9203 ("net/rds: Use ERR_PTR for rds_message_alloc_sgs()")
Signed-off-by: default avatarLv Yunlong <lyl2019@mail.ustc.edu.cn>
Reviewed-by: default avatarHåkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      bdc2ab5c
    • Tong Zhu's avatar
      neighbour: Disregard DEAD dst in neigh_update · d47ec7a0
      Tong Zhu authored 
      After a short network outage, the dst_entry is timed out and put
in DST_OBSOLETE_DEAD. We are in this code because arp reply comes
from this neighbour after network recovers. There is a potential
race condition that dst_entry is still in DST_OBSOLETE_DEAD.
With that, another neighbour lookup causes more harm than good.

In best case all packets in arp_queue are lost. This is
counterproductive to the original goal of finding a better path
for those packets.

I observed a worst case with 4.x kernel where a dst_entry in
DST_OBSOLETE_DEAD state is associated with loopback net_device.
It leads to an ethernet header with all zero addresses.
A packet with all zero source MAC address is quite deadly with
mac80211, ath9k and 802.11 block ack.  It fails
ieee80211_find_sta_by_ifaddr in ath9k (xmit.c). Ath9k flushes tx
queue (ath_tx_complete_aggr). BAW (block ack window) is not
updated. BAW logic is damaged and ath9k transmission is disabled.
Signed-off-by: default avatarTong Zhu <zhutong@amazon.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      d47ec7a0
    • Tariq Toukan's avatar
      net/mlx5e: Guarantee room for XSK wakeup NOP on async ICOSQ · 3ff3874f
      Tariq Toukan authored 
      XSK wakeup flow triggers an IRQ by posting a NOP WQE and hitting
the doorbell on the async ICOSQ.
It maintains its state so that it doesn't issue another NOP WQE
if it has an outstanding one already.

For this flow to work properly, the NOP post must not fail.
Make sure to reserve room for the NOP WQE in all WQE posts to the
async ICOSQ.

Fixes: 8d94b590 ("net/mlx5e: Turn XSK ICOSQ into a general asynchronous one")
Fixes: 1182f365 ("net/mlx5e: kTLS, Add kTLS RX HW offload support")
Fixes: 0419d8c9 ("net/mlx5e: kTLS, Add kTLS RX resync support")
Signed-off-by: default avatarTariq Toukan <tariqt@nvidia.com>
Reviewed-by: default avatarMaxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      3ff3874f
    • Dima Chumak's avatar
      net/mlx5e: Consider geneve_opts for encap contexts · 929a2fad
      Dima Chumak authored 
      Current algorithm for encap keys is legacy from initial vxlan
implementation and doesn't take into account all possible fields of a
tunnel. For example, for a Geneve tunnel, which may have additional TLV
options, they are ignored when comparing encap keys and a rule can be
attached to an incorrect encap entry.

Fix that by introducing encap_info_equal() operation in
struct mlx5e_tc_tunnel. Geneve tunnel type uses custom implementation,
which extends generic algorithm and considers options if they are set.

Fixes: 7f1a546e ("net/mlx5e: Consider tunnel type for encap contexts")
Signed-off-by: default avatarDima Chumak <dchumak@nvidia.com>
Reviewed-by: default avatarVlad Buslov <vladbu@nvidia.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      929a2fad
    • Daniel Jurgens's avatar
      net/mlx5: Don't request more than supported EQs · a7b76002
      Daniel Jurgens authored 
      Calculating the number of compeltion EQs based on the number of
available IRQ vectors doesn't work now that all async EQs share one IRQ.
Thus the max number of EQs can be exceeded on systems with more than
approximately 256 CPUs. Take this into account when calculating the
number of available completion EQs.

Fixes: 81bfa206 ("net/mlx5: Use a single IRQ for all async EQs")
Signed-off-by: default avatarDaniel Jurgens <danielj@mellanox.com>
Reviewed-by: default avatarParav Pandit <parav@nvidia.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      a7b76002
    • Tariq Toukan's avatar
      net/mlx5e: kTLS, Fix RX counters atomicity · 6f4fdd53
      Tariq Toukan authored 
      Some TLS RX counters increment per socket/connection, and are not
protected against parallel modifications from several cores.
Switch them to atomic counters by taking them out of the RQ stats into
the global atomic TLS stats.

In this patch, we touch 'rx_tls_ctx/del' that count the number of
device-offloaded RX TLS connections added/deleted.
These counters are updated in the add/del callbacks, out of the fast
data-path.

This change is not needed for counters that increment only in NAPI
context, as they are protected by the NAPI mechanism.
Keep them as tls_* counters under 'struct mlx5e_rq_stats'.

Fixes: 76c1e1ac ("net/mlx5e: kTLS, Add kTLS RX stats")
Signed-off-by: default avatarTariq Toukan <tariqt@nvidia.com>
Reviewed-by: default avatarMaxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      6f4fdd53
    • Tariq Toukan's avatar
      net/mlx5e: kTLS, Fix TX counters atomicity · a51bce96
      Tariq Toukan authored 
      Some TLS TX counters increment per socket/connection, and are not
protected against parallel modifications from several cores.
Switch them to atomic counters by taking them out of the SQ stats into
the global atomic TLS stats.

In this patch, we touch a single counter 'tx_tls_ctx' that counts the
number of device-offloaded TX TLS connections added.
Now that this counter can be increased without the for having the SQ
context in hand, move it to the mlx5e_ktls_add_tx() callback where it
really belongs, out of the fast data-path.

This change is not needed for counters that increment only in NAPI
context or under the TX lock, as they are already protected.
Keep them as tls_* counters under 'struct mlx5e_sq_stats'.

Fixes: d2ead1f3 ("net/mlx5e: Add kTLS TX HW offload support")
Signed-off-by: default avatarTariq Toukan <tariqt@nvidia.com>
Reviewed-by: default avatarMaxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      a51bce96
    • Maor Dickman's avatar
      net/mlx5: E-switch, Create vport miss group only if src rewrite is supported · e929e3da
      Maor Dickman authored 
      Create send to vport miss group was added in order to support traffic
recirculation to root table with metadata source rewrite.
This group is created also in case source rewrite isn't supported.

Fixed by creating send to vport miss group only if source rewrite is
supported by FW.

Fixes: 8e404fef ("net/mlx5e: Match recirculated packet miss in slow table using reg_c1")
Signed-off-by: default avatarMaor Dickman <maord@nvidia.com>
Reviewed-by: default avatarVlad Buslov <vladbu@nvidia.com>
Reviewed-by: default avatarRoi Dayan <roid@nvidia.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@nvidia.com>
      e929e3da
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7