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. 17 10月, 2017 1 次提交
  3. 22 9月, 2017 1 次提交
  5. 01 9月, 2017 1 次提交
    • C
      net_sched: add reverse binding for tc class · 07d79fc7
      由 Cong Wang 提交于 
      TC filters when used as classifiers are bound to TC classes.
However, there is a hidden difference when adding them in different
orders:

1. If we add tc classes before its filters, everything is fine.
   Logically, the classes exist before we specify their ID's in
   filters, it is easy to bind them together, just as in the current
   code base.

2. If we add tc filters before the tc classes they bind, we have to
   do dynamic lookup in fast path. What's worse, this happens all
   the time not just once, because on fast path tcf_result is passed
   on stack, there is no way to propagate back to the one in tc filters.

This hidden difference hurts performance silently if we have many tc
classes in hierarchy.

This patch intends to close this gap by doing the reverse binding when
we create a new class, in this case we can actually search all the
filters in its parent, match and fixup by classid. And because
tcf_result is specific to each type of tc filter, we have to introduce
a new ops for each filter to tell how to bind the class.

Note, we still can NOT totally get rid of those class lookup in
->enqueue() because cgroup and flow filters have no way to determine
the classid at setup time, they still have to go through dynamic lookup.

Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      07d79fc7
  7. 31 8月, 2017 1 次提交
    • C
      net/sched: Change cls_flower to use IDR · c15ab236
      由 Chris Mi 提交于 
      Currently, all filters with the same priority are linked in a doubly
linked list. Every filter should have a unique handle. To make the
handle unique, we need to iterate the list every time to see if the
handle exists or not when inserting a new filter. It is time-consuming.
For example, it takes about 5m3.169s to insert 64K rules.

This patch changes cls_flower to use IDR. With this patch, it
takes about 0m1.127s to insert 64K rules. The improvement is huge.

But please note that in this testing, all filters share the same action.
If every filter has a unique action, that is another bottleneck.
Follow-up patch in this patchset addresses that.
Signed-off-by: NChris Mi <chrism@mellanox.com>
Signed-off-by: NJiri Pirko <jiri@mellanox.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      c15ab236
  9. 17 8月, 2017 1 次提交
  11. 12 8月, 2017 1 次提交
  13. 08 8月, 2017 6 次提交
  15. 05 8月, 2017 1 次提交
  17. 08 6月, 2017 1 次提交
  19. 05 6月, 2017 1 次提交
  21. 25 5月, 2017 1 次提交
  23. 01 5月, 2017 1 次提交
  25. 25 4月, 2017 1 次提交
  27. 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
  29. 14 4月, 2017 1 次提交
  31. 18 2月, 2017 2 次提交
  33. 04 2月, 2017 1 次提交
  35. 31 1月, 2017 1 次提交
  37. 20 1月, 2017 1 次提交
    • A
      net/sched: cls_flower: reduce fl_change stack size · 39b7b6a6
      由 Arnd Bergmann 提交于 
      The new ARP support has pushed the stack size over the edge on ARM,
as there are two large objects on the stack in this function (mask
and tb) and both have now grown a bit more:

net/sched/cls_flower.c: In function 'fl_change':
net/sched/cls_flower.c:928:1: error: the frame size of 1072 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]

We can solve this by dynamically allocating one or both of them.
I first tried to do it just for the mask, but that only saved
152 bytes on ARM, while this version just does it for the 'tb'
array, bringing the stack size back down to 664 bytes.

Fixes: 99d31326 ("net/sched: cls_flower: Support matching on ARP")
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Acked-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      39b7b6a6
  39. 17 1月, 2017 1 次提交
  41. 12 1月, 2017 1 次提交
  43. 29 12月, 2016 1 次提交
  45. 24 12月, 2016 1 次提交
  47. 17 12月, 2016 2 次提交
  49. 09 12月, 2016 2 次提交
  51. 06 12月, 2016 1 次提交
  53. 03 12月, 2016 3 次提交
  55. 30 11月, 2016 1 次提交
  57. 28 11月, 2016 1 次提交
    • D
      net, sched: respect rcu grace period on cls destruction · d9363774
      由 Daniel Borkmann 提交于 
      Roi reported a crash in flower where tp->root was NULL in ->classify()
callbacks. Reason is that in ->destroy() tp->root is set to NULL via
RCU_INIT_POINTER(). It's problematic for some of the classifiers, because
this doesn't respect RCU grace period for them, and as a result, still
outstanding readers from tc_classify() will try to blindly dereference
a NULL tp->root.

The tp->root object is strictly private to the classifier implementation
and holds internal data the core such as tc_ctl_tfilter() doesn't know
about. Within some classifiers, such as cls_bpf, cls_basic, etc, tp->root
is only checked for NULL in ->get() callback, but nowhere else. This is
misleading and seemed to be copied from old classifier code that was not
cleaned up properly. For example, d3fa76ee ("[NET_SCHED]: cls_basic:
fix NULL pointer dereference") moved tp->root initialization into ->init()
routine, where before it was part of ->change(), so ->get() had to deal
with tp->root being NULL back then, so that was indeed a valid case, after
d3fa76ee, not really anymore. We used to set tp->root to NULL long
ago in ->destroy(), see 47a1a1d4 ("pkt_sched: remove unnecessary xchg()
in packet classifiers"); but the NULLifying was reintroduced with the
RCUification, but it's not correct for every classifier implementation.

In the cases that are fixed here with one exception of cls_cgroup, tp->root
object is allocated and initialized inside ->init() callback, which is always
performed at a point in time after we allocate a new tp, which means tp and
thus tp->root was not globally visible in the tp chain yet (see tc_ctl_tfilter()).
Also, on destruction tp->root is strictly kfree_rcu()'ed in ->destroy()
handler, same for the tp which is kfree_rcu()'ed right when we return
from ->destroy() in tcf_destroy(). This means, the head object's lifetime
for such classifiers is always tied to the tp lifetime. The RCU callback
invocation for the two kfree_rcu() could be out of order, but that's fine
since both are independent.

Dropping the RCU_INIT_POINTER(tp->root, NULL) for these classifiers here
means that 1) we don't need a useless NULL check in fast-path and, 2) that
outstanding readers of that tp in tc_classify() can still execute under
respect with RCU grace period as it is actually expected.

Things that haven't been touched here: cls_fw and cls_route. They each
handle tp->root being NULL in ->classify() path for historic reasons, so
their ->destroy() implementation can stay as is. If someone actually
cares, they could get cleaned up at some point to avoid the test in fast
path. cls_u32 doesn't set tp->root to NULL. For cls_rsvp, I just added a
!head should anyone actually be using/testing it, so it at least aligns with
cls_fw and cls_route. For cls_flower we additionally need to defer rhashtable
destruction (to a sleepable context) after RCU grace period as concurrent
readers might still access it. (Note that in this case we need to hold module
reference to keep work callback address intact, since we only wait on module
unload for all call_rcu()s to finish.)

This fixes one race to bring RCU grace period guarantees back. Next step
as worked on by Cong however is to fix 1e052be6 ("net_sched: destroy
proto tp when all filters are gone") to get the order of unlinking the tp
in tc_ctl_tfilter() for the RTM_DELTFILTER case right by moving
RCU_INIT_POINTER() before tcf_destroy() and let the notification for
removal be done through the prior ->delete() callback. Both are independant
issues. Once we have that right, we can then clean tp->root up for a number
of classifiers by not making them RCU pointers, which requires a new callback
(->uninit) that is triggered from tp's RCU callback, where we just kfree()
tp->root from there.

Fixes: 1f947bf1 ("net: sched: rcu'ify cls_bpf")
Fixes: 9888faef ("net: sched: cls_basic use RCU")
Fixes: 70da9f0b ("net: sched: cls_flow use RCU")
Fixes: 77b9900e ("tc: introduce Flower classifier")
Fixes: bf3994d2 ("net/sched: introduce Match-all classifier")
Fixes: 952313bd ("net: sched: cls_cgroup use RCU")
Reported-by: NRoi Dayan <roid@mellanox.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: Roi Dayan <roid@mellanox.com>
Cc: Jiri Pirko <jiri@mellanox.com>
Acked-by: NJohn Fastabend <john.r.fastabend@intel.com>
Acked-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      d9363774
  59. 10 11月, 2016 1 次提交