Skip to content

K
Kernel

openeuler / Kernel
1 年多 前同步成功

通知 8
Star 0
Fork 0
K
Kernel
切换分支/标签
  1. 25 3月, 2021 5 次提交
  3. 24 3月, 2021 1 次提交
    • D
      net: make unregister netdev warning timeout configurable · 5aa3afe1
      由 Dmitry Vyukov 提交于 
      netdev_wait_allrefs() issues a warning if refcount does not drop to 0
after 10 seconds. While 10 second wait generally should not happen
under normal workload in normal environment, it seems to fire falsely
very often during fuzzing and/or in qemu emulation (~10x slower).
At least it's not possible to understand if it's really a false
positive or not. Automated testing generally bumps all timeouts
to very high values to avoid flake failures.
Add net.core.netdev_unregister_timeout_secs sysctl to make
the timeout configurable for automated testing systems.
Lowering the timeout may also be useful for e.g. manual bisection.
The default value matches the current behavior.
Signed-off-by: NDmitry Vyukov <dvyukov@google.com>
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=211877
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      5aa3afe1
  5. 23 3月, 2021 3 次提交
    • E
      net: set initial device refcount to 1 · add2d736
      由 Eric Dumazet 提交于 
      When adding CONFIG_PCPU_DEV_REFCNT, I forgot that the
initial net device refcount was 0.

When CONFIG_PCPU_DEV_REFCNT is not set, this means
the first dev_hold() triggers an illegal refcount
operation (addition on 0)

refcount_t: addition on 0; use-after-free.
WARNING: CPU: 0 PID: 1 at lib/refcount.c:25 refcount_warn_saturate+0x128/0x1a4

Fix is to change initial (and final) refcount to be 1.

Also add a missing kerneldoc piece, as reported by
Stephen Rothwell.

Fixes: 919067cc ("net: add CONFIG_PCPU_DEV_REFCNT")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Reported-by: NGuenter Roeck <groeck@google.com>
Tested-by: NGuenter Roeck <groeck@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      add2d736
    • V
      net: move the ptype_all and ptype_base declarations to include/linux/netdevice.h · 744b8376
      由 Vladimir Oltean 提交于 
      ptype_all and ptype_base are declared in net/core/dev.c as non-static,
because they are used by net-procfs.c too. However, a "make W=1" build
complains that there was no previous declaration of ptype_all and
ptype_base in a header file, so this way of declaring things constitutes
a violation of coding style.

Let's move the extern declarations of ptype_all and ptype_base to the
linux/netdevice.h file, which is included by net-procfs.c too.
Signed-off-by: NVladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      744b8376
    • V
      netdev: add netdev_queue_set_dql_min_limit() · f57bac3c
      由 Vincent Mailhol 提交于 
      Add a function to set the dynamic queue limit minimum value.

Some specific drivers might have legitimate reasons to configure
dql.min_limit to a given value. Typically, this is the case when the
PDU of the protocol is smaller than the packet size to used to
carry those frames to the device.

Concrete example: a CAN (Control Area Network) device with an USB 2.0
interface.  The PDU of classical CAN protocol are roughly 16 bytes but
the USB packet size (which is used to carry the CAN frames to the
device) might be up to 512 bytes.  Wen small traffic burst occurs, BQL
algorithm is not able to immediately adjust and this would result in
having to send many small USB packets (i.e packet of 16 bytes for each
CAN frame). Filling up the USB packet with CAN frames is relatively
fast (small latency issue) but the gain of not having to send several
small USB packets is huge (big throughput increase). In this case,
forcing dql.min_limit to a given value that would allow to stuff the
USB packet is always a win.

This function is to be used by network drivers which are able to prove
through a rationale and through empirical tests on several environment
(with other applications, heavy context switching, virtualization...),
that they constantly reach better performances with a specific
predefined dql.min_limit value with no noticeable latency impact.
Signed-off-by: NVincent Mailhol <mailhol.vincent@wanadoo.fr>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      f57bac3c
  7. 20 3月, 2021 1 次提交
    • E
      net: add CONFIG_PCPU_DEV_REFCNT · 919067cc
      由 Eric Dumazet 提交于 
      I was working on a syzbot issue, claiming one device could not be
dismantled because its refcount was -1

unregister_netdevice: waiting for sit0 to become free. Usage count = -1

It would be nice if syzbot could trigger a warning at the time
this reference count became negative.

This patch adds CONFIG_PCPU_DEV_REFCNT options which defaults
to per cpu variables (as before this patch) on SMP builds.

v2: free_dev label in alloc_netdev_mqs() is moved to avoid
    a compiler warning (-Wunused-label), as reported
    by kernel test robot <lkp@intel.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      919067cc
  9. 19 3月, 2021 3 次提交
  11. 17 3月, 2021 1 次提交
  13. 04 3月, 2021 1 次提交
  15. 25 2月, 2021 3 次提交
  17. 13 2月, 2021 1 次提交
  19. 12 2月, 2021 1 次提交
    • C
      net: fix dev_ifsioc_locked() race condition · 3b23a32a
      由 Cong Wang 提交于 
      dev_ifsioc_locked() is called with only RCU read lock, so when
there is a parallel writer changing the mac address, it could
get a partially updated mac address, as shown below:

Thread 1			Thread 2
// eth_commit_mac_addr_change()
memcpy(dev->dev_addr, addr->sa_data, ETH_ALEN);
				// dev_ifsioc_locked()
				memcpy(ifr->ifr_hwaddr.sa_data,
					dev->dev_addr,...);

Close this race condition by guarding them with a RW semaphore,
like netdev_get_name(). We can not use seqlock here as it does not
allow blocking. The writers already take RTNL anyway, so this does
not affect the slow path. To avoid bothering existing
dev_set_mac_address() callers in drivers, introduce a new wrapper
just for user-facing callers on ioctl and rtnetlink paths.

Note, bonding also changes slave mac addresses but that requires
a separate patch due to the complexity of bonding code.

Fixes: 3710becf ("net: RCU locking for simple ioctl()")
Reported-by: N"Gong, Sishuai" <sishuai@purdue.edu>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: NCong Wang <cong.wang@bytedance.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      3b23a32a
  21. 10 2月, 2021 2 次提交
  23. 09 2月, 2021 1 次提交
  25. 29 1月, 2021 1 次提交
    • J
      net: adjust net_device layout for cacheline usage · 28af22c6
      由 Jesper Dangaard Brouer 提交于 
      The current layout of net_device is not optimal for cacheline usage.

The member adj_list.lower linked list is split between cacheline 2 and 3.
The ifindex is placed together with stats (struct net_device_stats),
although most modern drivers don't update this stats member.

The members netdev_ops, mtu and hard_header_len are placed on three
different cachelines. These members are accessed for XDP redirect into
devmap, which were noticeably with perf tool. When not using the map
redirect variant (like TC-BPF does), then ifindex is also used, which is
placed on a separate fourth cacheline. These members are also accessed
during forwarding with regular network stack. The members priv_flags and
flags are on fast-path for network stack transmit path in __dev_queue_xmit
(currently located together with mtu cacheline).

This patch creates a read mostly cacheline, with the purpose of keeping the
above mentioned members on the same cacheline.

Some netdev_features_t members also becomes part of this cacheline, which is
on purpose, as function netif_skb_features() is on fast-path via
validate_xmit_skb().
Signed-off-by: NJesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/r/161168277983.410784.12401225493601624417.stgit@firesoulSigned-off-by: NJakub Kicinski <kuba@kernel.org>
      28af22c6
  27. 23 1月, 2021 1 次提交
    • M
      sch_htb: Hierarchical QoS hardware offload · d03b195b
      由 Maxim Mikityanskiy 提交于 
      HTB doesn't scale well because of contention on a single lock, and it
also consumes CPU. This patch adds support for offloading HTB to
hardware that supports hierarchical rate limiting.

In the offload mode, HTB passes control commands to the driver using
ndo_setup_tc. The driver has to replicate the whole hierarchy of classes
and their settings (rate, ceil) in the NIC. Every modification of the
HTB tree caused by the admin results in ndo_setup_tc being called.

After this setup, the HTB algorithm is done completely in the NIC. An SQ
(send queue) is created for every leaf class and attached to the
hierarchy, so that the NIC can calculate and obey aggregated rate
limits, too. In the future, it can be changed, so that multiple SQs will
back a single leaf class.

ndo_select_queue is responsible for selecting the right queue that
serves the traffic class of each packet.

The data path works as follows: a packet is classified by clsact, the
driver selects a hardware queue according to its class, and the packet
is enqueued into this queue's qdisc.

This solution addresses two main problems of scaling HTB:

1. Contention by flow classification. Currently the filters are attached
to the HTB instance as follows:

    # tc filter add dev eth0 parent 1:0 protocol ip flower dst_port 80
    classid 1:10

It's possible to move classification to clsact egress hook, which is
thread-safe and lock-free:

    # tc filter add dev eth0 egress protocol ip flower dst_port 80
    action skbedit priority 1:10

This way classification still happens in software, but the lock
contention is eliminated, and it happens before selecting the TX queue,
allowing the driver to translate the class to the corresponding hardware
queue in ndo_select_queue.

Note that this is already compatible with non-offloaded HTB and doesn't
require changes to the kernel nor iproute2.

2. Contention by handling packets. HTB is not multi-queue, it attaches
to a whole net device, and handling of all packets takes the same lock.
When HTB is offloaded, it registers itself as a multi-queue qdisc,
similarly to mq: HTB is attached to the netdev, and each queue has its
own qdisc.

Some features of HTB may be not supported by some particular hardware,
for example, the maximum number of classes may be limited, the
granularity of rate and ceil parameters may be different, etc. - so, the
offload is not enabled by default, a new parameter is used to enable it:

    # tc qdisc replace dev eth0 root handle 1: htb offload
Signed-off-by: NMaxim Mikityanskiy <maximmi@mellanox.com>
Reviewed-by: NTariq Toukan <tariqt@nvidia.com>
Signed-off-by: NJakub Kicinski <kuba@kernel.org>
      d03b195b
  29. 20 1月, 2021 1 次提交
    • J
      bonding: add a vlan+srcmac tx hashing option · 7b8fc010
      由 Jarod Wilson 提交于 
      This comes from an end-user request, where they're running multiple VMs on
hosts with bonded interfaces connected to some interest switch topologies,
where 802.3ad isn't an option. They're currently running a proprietary
solution that effectively achieves load-balancing of VMs and bandwidth
utilization improvements with a similar form of transmission algorithm.

Basically, each VM has it's own vlan, so it always sends its traffic out
the same interface, unless that interface fails. Traffic gets split
between the interfaces, maintaining a consistent path, with failover still
available if an interface goes down.

Unlike bond_eth_hash(), this hash function is using the full source MAC
address instead of just the last byte, as there are so few components to
the hash, and in the no-vlan case, we would be returning just the last
byte of the source MAC as the hash value. It's entirely possible to have
two NICs in a bond with the same last byte of their MAC, but not the same
MAC, so this adjustment should guarantee distinct hashes in all cases.

This has been rudimetarily tested to provide similar results to the
proprietary solution it is aiming to replace. A patch for iproute2 is also
posted, to properly support the new mode there as well.

Cc: Jay Vosburgh <j.vosburgh@gmail.com>
Cc: Veaceslav Falico <vfalico@gmail.com>
Cc: Andy Gospodarek <andy@greyhouse.net>
Cc: Thomas Davis <tadavis@lbl.gov>
Signed-off-by: NJarod Wilson <jarod@redhat.com>
Link: https://lore.kernel.org/r/20210119010927.1191922-1-jarod@redhat.comSigned-off-by: NJakub Kicinski <kuba@kernel.org>
      7b8fc010
  31. 19 1月, 2021 1 次提交
  33. 10 1月, 2021 1 次提交
  35. 08 1月, 2021 1 次提交
  37. 17 12月, 2020 1 次提交
  39. 02 12月, 2020 1 次提交
  41. 01 12月, 2020 1 次提交
    • B
      net: Introduce preferred busy-polling · 7fd3253a
      由 Björn Töpel 提交于 
      The existing busy-polling mode, enabled by the SO_BUSY_POLL socket
option or system-wide using the /proc/sys/net/core/busy_read knob, is
an opportunistic. That means that if the NAPI context is not
scheduled, it will poll it. If, after busy-polling, the budget is
exceeded the busy-polling logic will schedule the NAPI onto the
regular softirq handling.

One implication of the behavior above is that a busy/heavy loaded NAPI
context will never enter/allow for busy-polling. Some applications
prefer that most NAPI processing would be done by busy-polling.

This series adds a new socket option, SO_PREFER_BUSY_POLL, that works
in concert with the napi_defer_hard_irqs and gro_flush_timeout
knobs. The napi_defer_hard_irqs and gro_flush_timeout knobs were
introduced in commit 6f8b12d6 ("net: napi: add hard irqs deferral
feature"), and allows for a user to defer interrupts to be enabled and
instead schedule the NAPI context from a watchdog timer. When a user
enables the SO_PREFER_BUSY_POLL, again with the other knobs enabled,
and the NAPI context is being processed by a softirq, the softirq NAPI
processing will exit early to allow the busy-polling to be performed.

If the application stops performing busy-polling via a system call,
the watchdog timer defined by gro_flush_timeout will timeout, and
regular softirq handling will resume.

In summary; Heavy traffic applications that prefer busy-polling over
softirq processing should use this option.

Example usage:

  $ echo 2 | sudo tee /sys/class/net/ens785f1/napi_defer_hard_irqs
  $ echo 200000 | sudo tee /sys/class/net/ens785f1/gro_flush_timeout

Note that the timeout should be larger than the userspace processing
window, otherwise the watchdog will timeout and fall back to regular
softirq processing.

Enable the SO_BUSY_POLL/SO_PREFER_BUSY_POLL options on your socket.
Signed-off-by: NBjörn Töpel <bjorn.topel@intel.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Reviewed-by: NJakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/bpf/20201130185205.196029-2-bjorn.topel@gmail.com
      7fd3253a
  43. 25 11月, 2020 1 次提交
  45. 24 11月, 2020 2 次提交
  47. 18 11月, 2020 1 次提交
  49. 10 11月, 2020 1 次提交
  51. 01 11月, 2020 2 次提交
  53. 14 10月, 2020 1 次提交