Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 160
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 13 9月, 2017 2 次提交
    • C
      net_sched: carefully handle tcf_block_put() · 1697c4bb
      由 Cong Wang 提交于 
      As pointed out by Jiri, there is still a race condition between
tcf_block_put() and tcf_chain_destroy() in a RCU callback. There
is no way to make it correct without proper locking or synchronization,
because both operate on a shared list.

Locking is hard, because the only lock we can pick here is a spinlock,
however, in tc_dump_tfilter() we iterate this list with a sleeping
function called (tcf_chain_dump()), which makes using a lock to protect
chain_list almost impossible.

Jiri suggested the idea of holding a refcnt before flushing, this works
because it guarantees us there would be no parallel tcf_chain_destroy()
during the loop, therefore the race condition is gone. But we have to
be very careful with proper synchronization with RCU callbacks.
Suggested-by: NJiri Pirko <jiri@mellanox.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      1697c4bb
    • C
      net_sched: fix reference counting of tc filter chain · e2ef7544
      由 Cong Wang 提交于 
      This patch fixes the following ugliness of tc filter chain refcnt:

a) tp proto should hold a refcnt to the chain too. This significantly
   simplifies the logic.

b) Chain 0 is no longer special, it is created with refcnt=1 like any
   other chains. All the ugliness in tcf_chain_put() can be gone!

c) No need to handle the flushing oddly, because block still holds
   chain 0, it can not be released, this guarantees block is the last
   user.

d) The race condition with RCU callbacks is easier to handle with just
   a rcu_barrier(). Much easier to understand, nothing to hide. Thanks
   to the previous patch. Please see also the comments in code.

e) Make the code understandable by humans, much less error-prone.

Fixes: 744a4cf6 ("net: sched: fix use after free when tcf_chain_destroy is called multiple times")
Fixes: 5bc17018 ("net: sched: introduce multichain support for filters")
Cc: Jiri Pirko <jiri@mellanox.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      e2ef7544
  3. 08 9月, 2017 1 次提交
  5. 26 8月, 2017 1 次提交
    • W
      net_sched: remove tc class reference counting · 143976ce
      由 WANG Cong 提交于 
      For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:

1) For class modification and dumping paths, we already hold RTNL lock,
   so all of these ->get(),->change(),->put() are atomic.

2) For filter bindiing/unbinding, we use other reference counter than
   this one, and they should have RTNL lock too.

3) For ->qlen_notify(), it is special because it is called on ->enqueue()
   path, but we already hold qdisc tree lock there, and we hold this
   tree lock when graft or delete the class too, so it should not be gone
   or changed until we release the tree lock.

Therefore, this patch removes ->get() and ->put(), but:

1) Adds a new ->find() to find the pointer to a class by classid, no
   refcnt.

2) Move the original class destroy upon the last refcnt into ->delete(),
   right after releasing tree lock. This is fine because the class is
   already removed from hash when holding the lock.

For those who also use ->put() as ->unbind(), just rename them to reflect
this change.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      143976ce
  7. 23 8月, 2017 2 次提交
    • J
      net: sched: don't do tcf_chain_flush from tcf_chain_destroy · 30d65e8f
      由 Jiri Pirko 提交于 
      tcf_chain_flush needs to be called with RTNL. However, on
free_tcf->
 tcf_action_goto_chain_fini->
  tcf_chain_put->
   tcf_chain_destroy->
    tcf_chain_flush
callpath, it is called without RTNL.
This issue was notified by following warning:

[  155.599052] WARNING: suspicious RCU usage
[  155.603165] 4.13.0-rc5jiri+ #54 Not tainted
[  155.607456] -----------------------------
[  155.611561] net/sched/cls_api.c:195 suspicious rcu_dereference_protected() usage!

Since on this callpath, the chain is guaranteed to be already empty
by check in tcf_chain_put, move the tcf_chain_flush call out and call it
only where it is needed - into tcf_block_put.

Fixes: db50514f ("net: sched: add termination action to allow goto chain")
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      30d65e8f
    • J
      net: sched: fix use after free when tcf_chain_destroy is called multiple times · 744a4cf6
      由 Jiri Pirko 提交于 
      The goto_chain termination action takes a reference of a chain. In that
case, there is an issue when block_put is called tcf_chain_destroy
directly. The follo-up call of tcf_chain_put by goto_chain action free
works with memory that is already freed. This was caught by kasan:

[  220.337908] BUG: KASAN: use-after-free in tcf_chain_put+0x1b/0x50
[  220.344103] Read of size 4 at addr ffff88036d1f2cec by task systemd-journal/261
[  220.353047] CPU: 0 PID: 261 Comm: systemd-journal Not tainted 4.13.0-rc5jiri+ #54
[  220.360661] Hardware name: Mellanox Technologies Ltd. Mellanox switch/Mellanox x86 mezzanine board, BIOS 4.6.5 08/02/2016
[  220.371784] Call Trace:
[  220.374290]  <IRQ>
[  220.376355]  dump_stack+0xd5/0x150
[  220.391485]  print_address_description+0x86/0x410
[  220.396308]  kasan_report+0x181/0x4c0
[  220.415211]  tcf_chain_put+0x1b/0x50
[  220.418949]  free_tcf+0x95/0xc0

So allow tcf_chain_destroy to be called multiple times, free only in
case the reference count drops to 0.

Fixes: 5bc17018 ("net: sched: introduce multichain support for filters")
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      744a4cf6
  9. 19 8月, 2017 1 次提交
  11. 10 8月, 2017 1 次提交
  13. 09 8月, 2017 1 次提交
  15. 08 8月, 2017 2 次提交
  17. 05 8月, 2017 3 次提交
  19. 26 5月, 2017 2 次提交
  21. 23 5月, 2017 2 次提交
  23. 18 5月, 2017 10 次提交
  25. 22 4月, 2017 1 次提交
    • W
      net_sched: move the empty tp check from ->destroy() to ->delete() · 763dbf63
      由 WANG Cong 提交于 
      We could have a race condition where in ->classify() path we
dereference tp->root and meanwhile a parallel ->destroy() makes it
a NULL. Daniel cured this bug in commit d9363774
("net, sched: respect rcu grace period on cls destruction").

This happens when ->destroy() is called for deleting a filter to
check if we are the last one in tp, this tp is still linked and
visible at that time. The root cause of this problem is the semantic
of ->destroy(), it does two things (for non-force case):

1) check if tp is empty
2) if tp is empty we could really destroy it

and its caller, if cares, needs to check its return value to see if it
is really destroyed. Therefore we can't unlink tp unless we know it is
empty.

As suggested by Daniel, we could actually move the test logic to ->delete()
so that we can safely unlink tp after ->delete() tells us the last one is
just deleted and before ->destroy().

Fixes: 1e052be6 ("net_sched: destroy proto tp when all filters are gone")
Cc: Roi Dayan <roid@mellanox.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Acked-by: NDaniel Borkmann <daniel@iogearbox.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      763dbf63
  27. 18 4月, 2017 1 次提交
  29. 14 4月, 2017 1 次提交
  31. 15 2月, 2017 1 次提交
  33. 11 2月, 2017 6 次提交
  35. 27 12月, 2016 1 次提交
    • D
      net, sched: fix soft lockup in tc_classify · 628185cf
      由 Daniel Borkmann 提交于 
      Shahar reported a soft lockup in tc_classify(), where we run into an
endless loop when walking the classifier chain due to tp->next == tp
which is a state we should never run into. The issue only seems to
trigger under load in the tc control path.

What happens is that in tc_ctl_tfilter(), thread A allocates a new
tp, initializes it, sets tp_created to 1, and calls into tp->ops->change()
with it. In that classifier callback we had to unlock/lock the rtnl
mutex and returned with -EAGAIN. One reason why we need to drop there
is, for example, that we need to request an action module to be loaded.

This happens via tcf_exts_validate() -> tcf_action_init/_1() meaning
after we loaded and found the requested action, we need to redo the
whole request so we don't race against others. While we had to unlock
rtnl in that time, thread B's request was processed next on that CPU.
Thread B added a new tp instance successfully to the classifier chain.
When thread A returned grabbing the rtnl mutex again, propagating -EAGAIN
and destroying its tp instance which never got linked, we goto replay
and redo A's request.

This time when walking the classifier chain in tc_ctl_tfilter() for
checking for existing tp instances we had a priority match and found
the tp instance that was created and linked by thread B. Now calling
again into tp->ops->change() with that tp was successful and returned
without error.

tp_created was never cleared in the second round, thus kernel thinks
that we need to link it into the classifier chain (once again). tp and
*back point to the same object due to the match we had earlier on. Thus
for thread B's already public tp, we reset tp->next to tp itself and
link it into the chain, which eventually causes the mentioned endless
loop in tc_classify() once a packet hits the data path.

Fix is to clear tp_created at the beginning of each request, also when
we replay it. On the paths that can cause -EAGAIN we already destroy
the original tp instance we had and on replay we really need to start
from scratch. It seems that this issue was first introduced in commit
12186be7 ("net_cls: fix unconfigured struct tcf_proto keeps chaining
and avoid kernel panic when we use cls_cgroup").

Fixes: 12186be7 ("net_cls: fix unconfigured struct tcf_proto keeps chaining and avoid kernel panic when we use cls_cgroup")
Reported-by: NShahar Klein <shahark@mellanox.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Tested-by: NShahar Klein <shahark@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      628185cf
  37. 03 12月, 2016 1 次提交
    • H
      net/sched: cls_flower: Add offload support using egress Hardware device · 7091d8c7
      由 Hadar Hen Zion 提交于 
      In order to support hardware offloading when the device given by the tc
rule is different from the Hardware underline device, extract the mirred
(egress) device from the tc action when a filter is added, using the new
tc_action_ops, get_dev().

Flower caches the information about the mirred device and use it for
calling ndo_setup_tc in filter change, update stats and delete.

Calling ndo_setup_tc of the mirred (egress) device instead of the
ingress device will allow a resolution between the software ingress
device and the underline hardware device.

The resolution will take place inside the offloading driver using
'egress_device' flag added to tc_to_netdev struct which is provided to
the offloading driver.
Signed-off-by: NHadar Hen Zion <hadarh@mellanox.com>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      7091d8c7