Xiumei and Christoph reported the following lockdep splat, complaining of
the qdisc root lock being taken twice:
============================================
WARNING: possible recursive locking detected
6.7.0-rc3+ #598 Not tainted
--------------------------------------------
swapper/2/0 is trying to acquire lock:
ffff888177190110 (&sch->q.lock){+.-.}-{2:2}, at: __dev_queue_xmit+0x1560/0x2e70
but task is already holding lock:
ffff88811995a110 (&sch->q.lock){+.-.}-{2:2}, at: __dev_queue_xmit+0x1560/0x2e70
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0
----
lock(&sch->q.lock);
lock(&sch->q.lock);
*** DEADLOCK ***
May be due to missing lock nesting notation
5 locks held by swapper/2/0:
#0: ffff888135a09d98 ((&in_dev->mr_ifc_timer)){+.-.}-{0:0}, at: call_timer_fn+0x11a/0x510
#1: ffffffffaaee5260 (rcu_read_lock){....}-{1:2}, at: ip_finish_output2+0x2c0/0x1ed0
#2: ffffffffaaee5200 (rcu_read_lock_bh){....}-{1:2}, at: __dev_queue_xmit+0x209/0x2e70
#3: ffff88811995a110 (&sch->q.lock){+.-.}-{2:2}, at: __dev_queue_xmit+0x1560/0x2e70
#4: ffffffffaaee5200 (rcu_read_lock_bh){....}-{1:2}, at: __dev_queue_xmit+0x209/0x2e70
stack backtrace:
CPU: 2 PID: 0 Comm: swapper/2 Not tainted 6.7.0-rc3+ #598
Hardware name: Red Hat KVM, BIOS 1.13.0-2.module+el8.3.0+7353+9de0a3cc 04/01/2014
Call Trace:
<IRQ>
dump_stack_lvl+0x4a/0x80
__lock_acquire+0xfdd/0x3150
lock_acquire+0x1ca/0x540
_raw_spin_lock+0x34/0x80
__dev_queue_xmit+0x1560/0x2e70
tcf_mirred_act+0x82e/0x1260 [act_mirred]
tcf_action_exec+0x161/0x480
tcf_classify+0x689/0x1170
prio_enqueue+0x316/0x660 [sch_prio]
dev_qdisc_enqueue+0x46/0x220
__dev_queue_xmit+0x1615/0x2e70
ip_finish_output2+0x1218/0x1ed0
__ip_finish_output+0x8b3/0x1350
ip_output+0x163/0x4e0
igmp_ifc_timer_expire+0x44b/0x930
call_timer_fn+0x1a2/0x510
run_timer_softirq+0x54d/0x11a0
__do_softirq+0x1b3/0x88f
irq_exit_rcu+0x18f/0x1e0
sysvec_apic_timer_interrupt+0x6f/0x90
</IRQ>
This happens when TC does a mirred egress redirect from the root qdisc of
device A to the root qdisc of device B. As long as these two locks aren't
protecting the same qdisc, they can be acquired in chain: add a per-qdisc
lockdep key to silence false warnings.
This dynamic key should safely replace the static key we have in sch_htb:
it was added to allow enqueueing to the device "direct qdisc" while still
holding the qdisc root lock.
v2: don't use static keys anymore in HTB direct qdiscs (thanks Eric Dumazet)
CC: Maxim Mikityanskiy <maxim@isovalent.com>
CC: Xiumei Mu <xmu@redhat.com>
Reported-by: Christoph Paasch <cpaasch@apple.com>
Closes: https://github.com/multipath-tcp/mptcp_net-next/issues/451
Signed-off-by: Davide Caratti <dcaratti@redhat.com>
Link: https://lore.kernel.org/r/7dc06d6158f72053cf877a82e2a7a5bd23692faa.1713448007.git.dcaratti@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When the mirred action is used on a classful egress qdisc and a packet is
mirrored or redirected to self we hit a qdisc lock deadlock.
See trace below.
[..... other info removed for brevity....]
[ 82.890906]
[ 82.890906] ============================================
[ 82.890906] WARNING: possible recursive locking detected
[ 82.890906] 6.8.0-05205-g77fadd89fe2d-dirty #213 Tainted: G W
[ 82.890906] --------------------------------------------
[ 82.890906] ping/418 is trying to acquire lock:
[ 82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[ 82.890906]
[ 82.890906] but task is already holding lock:
[ 82.890906] ffff888006994110 (&sch->q.lock){+.-.}-{3:3}, at:
__dev_queue_xmit+0x1778/0x3550
[ 82.890906]
[ 82.890906] other info that might help us debug this:
[ 82.890906] Possible unsafe locking scenario:
[ 82.890906]
[ 82.890906] CPU0
[ 82.890906] ----
[ 82.890906] lock(&sch->q.lock);
[ 82.890906] lock(&sch->q.lock);
[ 82.890906]
[ 82.890906] *** DEADLOCK ***
[ 82.890906]
[..... other info removed for brevity....]
Example setup (eth0->eth0) to recreate
tc qdisc add dev eth0 root handle 1: htb default 30
tc filter add dev eth0 handle 1: protocol ip prio 2 matchall \
action mirred egress redirect dev eth0
Another example(eth0->eth1->eth0) to recreate
tc qdisc add dev eth0 root handle 1: htb default 30
tc filter add dev eth0 handle 1: protocol ip prio 2 matchall \
action mirred egress redirect dev eth1
tc qdisc add dev eth1 root handle 1: htb default 30
tc filter add dev eth1 handle 1: protocol ip prio 2 matchall \
action mirred egress redirect dev eth0
We fix this by adding an owner field (CPU id) to struct Qdisc set after
root qdisc is entered. When the softirq enters it a second time, if the
qdisc owner is the same CPU, the packet is dropped to break the loop.
Reported-by: Mingshuai Ren <renmingshuai@huawei.com>
Closes: https://lore.kernel.org/netdev/20240314111713.5979-1-renmingshuai@huawei.com/
Fixes: 3bcb846ca4 ("net: get rid of spin_trylock() in net_tx_action()")
Fixes: e578d9c025 ("net: sched: use counter to break reclassify loops")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Pedro Tammela <pctammela@mojatatu.com>
Tested-by: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://lore.kernel.org/r/20240415210728.36949-1-victor@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
TC filters come in 3 variants:
- no flag (try to process in hardware, but fallback to software))
- skip_hw (do not process filter by hardware)
- skip_sw (do not process filter by software)
However skip_sw is implemented so that the skip_sw
flag can first be checked, after it has been matched.
IMHO it's common when using skip_sw, to use it on all rules.
So if all filters in a block is skip_sw filters, then
we can bail early, we can thus avoid having to match
the filters, just to check for the skip_sw flag.
This patch adds a bypass, for when only TC skip_sw rules
are used. The bypass is guarded by a static key, to avoid
harming other workloads.
There are 3 ways that a packet from a skip_sw ruleset, can
end up in the kernel path. Although the send packets to a
non-existent chain way is only improved a few percents, then
I believe it's worth optimizing the trap and fall-though
use-cases.
+----------------------------+--------+--------+--------+
| Test description | Pre- | Post- | Rel. |
| | kpps | kpps | chg. |
+----------------------------+--------+--------+--------+
| basic forwarding + notrack | 3589.3 | 3587.9 | 1.00x |
| switch to eswitch mode | 3081.8 | 3094.7 | 1.00x |
| add ingress qdisc | 3042.9 | 3063.6 | 1.01x |
| tc forward in hw / skip_sw |37024.7 |37028.4 | 1.00x |
| tc forward in sw / skip_hw | 3245.0 | 3245.3 | 1.00x |
+----------------------------+--------+--------+--------+
| tests with only skip_sw rules below: |
+----------------------------+--------+--------+--------+
| 1 non-matching rule | 2694.7 | 3058.7 | 1.14x |
| 1 n-m rule, match trap | 2611.2 | 3323.1 | 1.27x |
| 1 n-m rule, goto non-chain | 2886.8 | 2945.9 | 1.02x |
| 5 non-matching rules | 1958.2 | 3061.3 | 1.56x |
| 5 n-m rules, match trap | 1911.9 | 3327.0 | 1.74x |
| 5 n-m rules, goto non-chain| 2883.1 | 2947.5 | 1.02x |
| 10 non-matching rules | 1466.3 | 3062.8 | 2.09x |
| 10 n-m rules, match trap | 1444.3 | 3317.9 | 2.30x |
| 10 n-m rules,goto non-chain| 2883.1 | 2939.5 | 1.02x |
| 25 non-matching rules | 838.5 | 3058.9 | 3.65x |
| 25 n-m rules, match trap | 824.5 | 3323.0 | 4.03x |
| 25 n-m rules,goto non-chain| 2875.8 | 2944.7 | 1.02x |
| 50 non-matching rules | 488.1 | 3054.7 | 6.26x |
| 50 n-m rules, match trap | 484.9 | 3318.5 | 6.84x |
| 50 n-m rules,goto non-chain| 2884.1 | 2939.7 | 1.02x |
+----------------------------+--------+--------+--------+
perf top (25 n-m skip_sw rules - pre patch):
20.39% [kernel] [k] __skb_flow_dissect
16.43% [kernel] [k] rhashtable_jhash2
10.58% [kernel] [k] fl_classify
10.23% [kernel] [k] fl_mask_lookup
4.79% [kernel] [k] memset_orig
2.58% [kernel] [k] tcf_classify
1.47% [kernel] [k] __x86_indirect_thunk_rax
1.42% [kernel] [k] __dev_queue_xmit
1.36% [kernel] [k] nft_do_chain
1.21% [kernel] [k] __rcu_read_lock
perf top (25 n-m skip_sw rules - post patch):
5.12% [kernel] [k] __dev_queue_xmit
4.77% [kernel] [k] nft_do_chain
3.65% [kernel] [k] dev_gro_receive
3.41% [kernel] [k] check_preemption_disabled
3.14% [kernel] [k] mlx5e_skb_from_cqe_mpwrq_nonlinear
2.88% [kernel] [k] __netif_receive_skb_core.constprop.0
2.49% [kernel] [k] mlx5e_xmit
2.15% [kernel] [k] ip_forward
1.95% [kernel] [k] mlx5e_tc_restore_tunnel
1.92% [kernel] [k] vlan_gro_receive
Test setup:
DUT: Intel Xeon D-1518 (2.20GHz) w/ Nvidia/Mellanox ConnectX-6 Dx 2x100G
Data rate measured on switch (Extreme X690), and DUT connected as
a router on a stick, with pktgen and pktsink as VLANs.
Pktgen-dpdk was in range 36.6-37.7 Mpps 64B packets across all tests.
Full test data at https://files.fiberby.net/ast/2024/tc_skip_sw/v2_tests/
Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Maintain a count of filters per block.
Counter updates are protected by cb_lock, which is
also used to protect the offload counters.
Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Maintain a count of skip_sw filters.
This counter is protected by the cb_lock, and is updated
at the same time as offloadcnt.
Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is now possible to disable BQL, but that causes the cpsw driver to break:
drivers/net/ethernet/ti/am65-cpsw-nuss.c:297:28: error: no member named 'dql' in 'struct netdev_queue'
297 | dql_avail(&netif_txq->dql),
There is already a helper function in net/sch_generic.h that could
be used to help here. Move its implementation into the common
linux/netdevice.h along with the other bql interfaces and change
both users over to the new interface.
Fixes: ea7f3cfaa5 ("net: bql: allow the config to be disabled")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a qdisc is deleted from a net device the stack instructs the
underlying driver to remove its flow offload callback from the
associated filter block using the 'FLOW_BLOCK_UNBIND' command. The stack
then continues to replay the removal of the filters in the block for
this driver by iterating over the chains in the block and invoking the
'reoffload' operation of the classifier being used. In turn, the
classifier in its 'reoffload' operation prepares and emits a
'FLOW_CLS_DESTROY' command for each filter.
However, the stack does not do the same for chain templates and the
underlying driver never receives a 'FLOW_CLS_TMPLT_DESTROY' command when
a qdisc is deleted. This results in a memory leak [1] which can be
reproduced using [2].
Fix by introducing a 'tmplt_reoffload' operation and have the stack
invoke it with the appropriate arguments as part of the replay.
Implement the operation in the sole classifier that supports chain
templates (flower) by emitting the 'FLOW_CLS_TMPLT_{CREATE,DESTROY}'
command based on whether a flow offload callback is being bound to a
filter block or being unbound from one.
As far as I can tell, the issue happens since cited commit which
reordered tcf_block_offload_unbind() before tcf_block_flush_all_chains()
in __tcf_block_put(). The order cannot be reversed as the filter block
is expected to be freed after flushing all the chains.
[1]
unreferenced object 0xffff888107e28800 (size 2048):
comm "tc", pid 1079, jiffies 4294958525 (age 3074.287s)
hex dump (first 32 bytes):
b1 a6 7c 11 81 88 ff ff e0 5b b3 10 81 88 ff ff ..|......[......
01 00 00 00 00 00 00 00 e0 aa b0 84 ff ff ff ff ................
backtrace:
[<ffffffff81c06a68>] __kmem_cache_alloc_node+0x1e8/0x320
[<ffffffff81ab374e>] __kmalloc+0x4e/0x90
[<ffffffff832aec6d>] mlxsw_sp_acl_ruleset_get+0x34d/0x7a0
[<ffffffff832bc195>] mlxsw_sp_flower_tmplt_create+0x145/0x180
[<ffffffff832b2e1a>] mlxsw_sp_flow_block_cb+0x1ea/0x280
[<ffffffff83a10613>] tc_setup_cb_call+0x183/0x340
[<ffffffff83a9f85a>] fl_tmplt_create+0x3da/0x4c0
[<ffffffff83a22435>] tc_ctl_chain+0xa15/0x1170
[<ffffffff838a863c>] rtnetlink_rcv_msg+0x3cc/0xed0
[<ffffffff83ac87f0>] netlink_rcv_skb+0x170/0x440
[<ffffffff83ac6270>] netlink_unicast+0x540/0x820
[<ffffffff83ac6e28>] netlink_sendmsg+0x8d8/0xda0
[<ffffffff83793def>] ____sys_sendmsg+0x30f/0xa80
[<ffffffff8379d29a>] ___sys_sendmsg+0x13a/0x1e0
[<ffffffff8379d50c>] __sys_sendmsg+0x11c/0x1f0
[<ffffffff843b9ce0>] do_syscall_64+0x40/0xe0
unreferenced object 0xffff88816d2c0400 (size 1024):
comm "tc", pid 1079, jiffies 4294958525 (age 3074.287s)
hex dump (first 32 bytes):
40 00 00 00 00 00 00 00 57 f6 38 be 00 00 00 00 @.......W.8.....
10 04 2c 6d 81 88 ff ff 10 04 2c 6d 81 88 ff ff ..,m......,m....
backtrace:
[<ffffffff81c06a68>] __kmem_cache_alloc_node+0x1e8/0x320
[<ffffffff81ab36c1>] __kmalloc_node+0x51/0x90
[<ffffffff81a8ed96>] kvmalloc_node+0xa6/0x1f0
[<ffffffff82827d03>] bucket_table_alloc.isra.0+0x83/0x460
[<ffffffff82828d2b>] rhashtable_init+0x43b/0x7c0
[<ffffffff832aed48>] mlxsw_sp_acl_ruleset_get+0x428/0x7a0
[<ffffffff832bc195>] mlxsw_sp_flower_tmplt_create+0x145/0x180
[<ffffffff832b2e1a>] mlxsw_sp_flow_block_cb+0x1ea/0x280
[<ffffffff83a10613>] tc_setup_cb_call+0x183/0x340
[<ffffffff83a9f85a>] fl_tmplt_create+0x3da/0x4c0
[<ffffffff83a22435>] tc_ctl_chain+0xa15/0x1170
[<ffffffff838a863c>] rtnetlink_rcv_msg+0x3cc/0xed0
[<ffffffff83ac87f0>] netlink_rcv_skb+0x170/0x440
[<ffffffff83ac6270>] netlink_unicast+0x540/0x820
[<ffffffff83ac6e28>] netlink_sendmsg+0x8d8/0xda0
[<ffffffff83793def>] ____sys_sendmsg+0x30f/0xa80
[2]
# tc qdisc add dev swp1 clsact
# tc chain add dev swp1 ingress proto ip chain 1 flower dst_ip 0.0.0.0/32
# tc qdisc del dev swp1 clsact
# devlink dev reload pci/0000:06:00.0
Fixes: bbf73830cd ("net: sched: traverse chains in block with tcf_get_next_chain()")
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The datapath can now find the block of the port in which the packet arrived
at.
In the next patch we show a possible usage of this patch in a new
version of mirred that multicasts to all ports except for the port in
which the packet arrived on.
Co-developed-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Co-developed-by: Pedro Tammela <pctammela@mojatatu.com>
Signed-off-by: Pedro Tammela <pctammela@mojatatu.com>
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit makes tc blocks track which ports have been added to them.
And, with that, we'll be able to use this new information to send
packets to the block's ports. Which will be done in the patch #3 of this
series.
Suggested-by: Jiri Pirko <jiri@nvidia.com>
Co-developed-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Co-developed-by: Pedro Tammela <pctammela@mojatatu.com>
Signed-off-by: Pedro Tammela <pctammela@mojatatu.com>
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Incrementing on Daniel's patch[1], make tc-related drop reason more
flexible for remaining qdiscs - that is, all qdiscs aside from clsact.
In essence, the drop reason will be set by cls_api and act_api in case
any error occurred in the data path. With that, we can give the user more
detailed information so that they can distinguish between a policy drop
or an error drop.
[1] https://lore.kernel.org/all/20231009092655.22025-1-daniel@iogearbox.net
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move drop_reason from struct tcf_result to skb cb - more specifically to
struct tc_skb_cb. With that, we'll be able to also set the drop reason for
the remaining qdiscs (aside from clsact) that do not have access to
tcf_result when time comes to set the skb drop reason.
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the kfree_skb_reason() in sch_handle_{ingress,egress}() can only
express a basic SKB_DROP_REASON_TC_INGRESS or SKB_DROP_REASON_TC_EGRESS reason.
Victor kicked-off an initial proposal to make this more flexible by disambiguating
verdict from return code by moving the verdict into struct tcf_result and
letting tcf_classify() return a negative error. If hit, then two new drop
reasons were added in the proposal, that is SKB_DROP_REASON_TC_INGRESS_ERROR
as well as SKB_DROP_REASON_TC_EGRESS_ERROR. Further analysis of the actual
error codes would have required to attach to tcf_classify via kprobe/kretprobe
to more deeply debug skb and the returned error.
In order to make the kfree_skb_reason() in sch_handle_{ingress,egress}() more
extensible, it can be addressed in a more straight forward way, that is: Instead
of placing the verdict into struct tcf_result, we can just put the drop reason
in there, which does not require changes throughout various classful schedulers
given the existing verdict logic can stay as is.
Then, SKB_DROP_REASON_TC_ERROR{,_*} can be added to the enum skb_drop_reason
to disambiguate between an error or an intentional drop. New drop reason error
codes can be added successively to the tc code base.
For internal error locations which have not yet been annotated with a
SKB_DROP_REASON_TC_ERROR{,_*}, the fallback is SKB_DROP_REASON_TC_INGRESS and
SKB_DROP_REASON_TC_EGRESS, respectively. Generic errors could be marked with a
SKB_DROP_REASON_TC_ERROR code until they are converted to more specific ones
if it is found that they would be useful for troubleshooting.
While drop reasons have infrastructure for subsystem specific error codes which
are currently used by mac80211 and ovs, Jakub mentioned that it is preferred
for tc to use the enum skb_drop_reason core codes given it is a better fit and
currently the tooling support is better, too.
With regards to the latter:
[...] I think Alastair (bpftrace) is working on auto-prettifying enums when
bpftrace outputs maps. So we can do something like:
$ bpftrace -e 'tracepoint:skb:kfree_skb { @[args->reason] = count(); }'
Attaching 1 probe...
^C
@[SKB_DROP_REASON_TC_INGRESS]: 2
@[SKB_CONSUMED]: 34
^^^^^^^^^^^^ names!!
Auto-magically. [...]
Add a small helper tcf_set_drop_reason() which can be used to set the drop reason
into the tcf_result.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Cc: Victor Nogueira <victor@mojatatu.com>
Link: https://lore.kernel.org/netdev/20231006063233.74345d36@kernel.org
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20231009092655.22025-1-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
pfifo_fast prio2band[] is renamed to sch_default_prio2band[]
and exported because we want to share it in FQ.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Dave Taht <dave.taht@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The 'filter_cnt' counter is used to control a Qdisc class lifetime.
Each filter referecing this class by its id will eventually
increment/decrement this counter in their respective
'add/update/delete' routines.
As these operations are always serialized under rtnl lock, we don't
need an atomic type like 'refcount_t'.
It also means that we lose the overflow/underflow checks already
present in refcount_t, which are valuable to hunt down bugs
where the unsigned counter wraps around as it aids automated tools
like syzkaller to scream in such situations.
Wrap the open coded increment/decrement into helper functions and
add overflow checks to the operations.
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Pedro Tammela <pctammela@mojatatu.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This work refactors and adds a lightweight extension ("tcx") to the tc BPF
ingress and egress data path side for allowing BPF program management based
on fds via bpf() syscall through the newly added generic multi-prog API.
The main goal behind this work which we also presented at LPC [0] last year
and a recent update at LSF/MM/BPF this year [3] is to support long-awaited
BPF link functionality for tc BPF programs, which allows for a model of safe
ownership and program detachment.
Given the rise in tc BPF users in cloud native environments, this becomes
necessary to avoid hard to debug incidents either through stale leftover
programs or 3rd party applications accidentally stepping on each others toes.
As a recap, a BPF link represents the attachment of a BPF program to a BPF
hook point. The BPF link holds a single reference to keep BPF program alive.
Moreover, hook points do not reference a BPF link, only the application's
fd or pinning does. A BPF link holds meta-data specific to attachment and
implements operations for link creation, (atomic) BPF program update,
detachment and introspection. The motivation for BPF links for tc BPF programs
is multi-fold, for example:
- From Meta: "It's especially important for applications that are deployed
fleet-wide and that don't "control" hosts they are deployed to. If such
application crashes and no one notices and does anything about that, BPF
program will keep running draining resources or even just, say, dropping
packets. We at FB had outages due to such permanent BPF attachment
semantics. With fd-based BPF link we are getting a framework, which allows
safe, auto-detachable behavior by default, unless application explicitly
opts in by pinning the BPF link." [1]
- From Cilium-side the tc BPF programs we attach to host-facing veth devices
and phys devices build the core datapath for Kubernetes Pods, and they
implement forwarding, load-balancing, policy, EDT-management, etc, within
BPF. Currently there is no concept of 'safe' ownership, e.g. we've recently
experienced hard-to-debug issues in a user's staging environment where
another Kubernetes application using tc BPF attached to the same prio/handle
of cls_bpf, accidentally wiping all Cilium-based BPF programs from underneath
it. The goal is to establish a clear/safe ownership model via links which
cannot accidentally be overridden. [0,2]
BPF links for tc can co-exist with non-link attachments, and the semantics are
in line also with XDP links: BPF links cannot replace other BPF links, BPF
links cannot replace non-BPF links, non-BPF links cannot replace BPF links and
lastly only non-BPF links can replace non-BPF links. In case of Cilium, this
would solve mentioned issue of safe ownership model as 3rd party applications
would not be able to accidentally wipe Cilium programs, even if they are not
BPF link aware.
Earlier attempts [4] have tried to integrate BPF links into core tc machinery
to solve cls_bpf, which has been intrusive to the generic tc kernel API with
extensions only specific to cls_bpf and suboptimal/complex since cls_bpf could
be wiped from the qdisc also. Locking a tc BPF program in place this way, is
getting into layering hacks given the two object models are vastly different.
We instead implemented the tcx (tc 'express') layer which is an fd-based tc BPF
attach API, so that the BPF link implementation blends in naturally similar to
other link types which are fd-based and without the need for changing core tc
internal APIs. BPF programs for tc can then be successively migrated from classic
cls_bpf to the new tc BPF link without needing to change the program's source
code, just the BPF loader mechanics for attaching is sufficient.
For the current tc framework, there is no change in behavior with this change
and neither does this change touch on tc core kernel APIs. The gist of this
patch is that the ingress and egress hook have a lightweight, qdisc-less
extension for BPF to attach its tc BPF programs, in other words, a minimal
entry point for tc BPF. The name tcx has been suggested from discussion of
earlier revisions of this work as a good fit, and to more easily differ between
the classic cls_bpf attachment and the fd-based one.
For the ingress and egress tcx points, the device holds a cache-friendly array
with program pointers which is separated from control plane (slow-path) data.
Earlier versions of this work used priority to determine ordering and expression
of dependencies similar as with classic tc, but it was challenged that for
something more future-proof a better user experience is required. Hence this
resulted in the design and development of the generic attach/detach/query API
for multi-progs. See prior patch with its discussion on the API design. tcx is
the first user and later we plan to integrate also others, for example, one
candidate is multi-prog support for XDP which would benefit and have the same
'look and feel' from API perspective.
The goal with tcx is to have maximum compatibility to existing tc BPF programs,
so they don't need to be rewritten specifically. Compatibility to call into
classic tcf_classify() is also provided in order to allow successive migration
or both to cleanly co-exist where needed given its all one logical tc layer and
the tcx plus classic tc cls/act build one logical overall processing pipeline.
tcx supports the simplified return codes TCX_NEXT which is non-terminating (go
to next program) and terminating ones with TCX_PASS, TCX_DROP, TCX_REDIRECT.
The fd-based API is behind a static key, so that when unused the code is also
not entered. The struct tcx_entry's program array is currently static, but
could be made dynamic if necessary at a point in future. The a/b pair swap
design has been chosen so that for detachment there are no allocations which
otherwise could fail.
The work has been tested with tc-testing selftest suite which all passes, as
well as the tc BPF tests from the BPF CI, and also with Cilium's L4LB.
Thanks also to Nikolay Aleksandrov and Martin Lau for in-depth early reviews
of this work.
[0] https://lpc.events/event/16/contributions/1353/
[1] https://lore.kernel.org/bpf/CAEf4BzbokCJN33Nw_kg82sO=xppXnKWEncGTWCTB9vGCmLB6pw@mail.gmail.com
[2] https://colocatedeventseu2023.sched.com/event/1Jo6O/tales-from-an-ebpf-programs-murder-mystery-hemanth-malla-guillaume-fournier-datadog
[3] http://vger.kernel.org/bpfconf2023_material/tcx_meta_netdev_borkmann.pdf
[4] https://lore.kernel.org/bpf/20210604063116.234316-1-memxor@gmail.com
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/r/20230719140858.13224-3-daniel@iogearbox.net
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
mini_Qdisc_pair::p_miniq is a double pointer to mini_Qdisc, initialized
in ingress_init() to point to net_device::miniq_ingress. ingress Qdiscs
access this per-net_device pointer in mini_qdisc_pair_swap(). Similar
for clsact Qdiscs and miniq_egress.
Unfortunately, after introducing RTNL-unlocked RTM_{NEW,DEL,GET}TFILTER
requests (thanks Hillf Danton for the hint), when replacing ingress or
clsact Qdiscs, for example, the old Qdisc ("@old") could access the same
miniq_{in,e}gress pointer(s) concurrently with the new Qdisc ("@new"),
causing race conditions [1] including a use-after-free bug in
mini_qdisc_pair_swap() reported by syzbot:
BUG: KASAN: slab-use-after-free in mini_qdisc_pair_swap+0x1c2/0x1f0 net/sched/sch_generic.c:1573
Write of size 8 at addr ffff888045b31308 by task syz-executor690/14901
...
Call Trace:
<TASK>
__dump_stack lib/dump_stack.c:88 [inline]
dump_stack_lvl+0xd9/0x150 lib/dump_stack.c:106
print_address_description.constprop.0+0x2c/0x3c0 mm/kasan/report.c:319
print_report mm/kasan/report.c:430 [inline]
kasan_report+0x11c/0x130 mm/kasan/report.c:536
mini_qdisc_pair_swap+0x1c2/0x1f0 net/sched/sch_generic.c:1573
tcf_chain_head_change_item net/sched/cls_api.c:495 [inline]
tcf_chain0_head_change.isra.0+0xb9/0x120 net/sched/cls_api.c:509
tcf_chain_tp_insert net/sched/cls_api.c:1826 [inline]
tcf_chain_tp_insert_unique net/sched/cls_api.c:1875 [inline]
tc_new_tfilter+0x1de6/0x2290 net/sched/cls_api.c:2266
...
@old and @new should not affect each other. In other words, @old should
never modify miniq_{in,e}gress after @new, and @new should not update
@old's RCU state.
Fixing without changing sch_api.c turned out to be difficult (please
refer to Closes: for discussions). Instead, make sure @new's first call
always happen after @old's last call (in {ingress,clsact}_destroy()) has
finished:
In qdisc_graft(), return -EBUSY if @old has any ongoing filter requests,
and call qdisc_destroy() for @old before grafting @new.
Introduce qdisc_refcount_dec_if_one() as the counterpart of
qdisc_refcount_inc_nz() used for filter requests. Introduce a
non-static version of qdisc_destroy() that does a TCQ_F_BUILTIN check,
just like qdisc_put() etc.
Depends on patch "net/sched: Refactor qdisc_graft() for ingress and
clsact Qdiscs".
[1] To illustrate, the syzkaller reproducer adds ingress Qdiscs under
TC_H_ROOT (no longer possible after commit c7cfbd1150 ("net/sched:
sch_ingress: Only create under TC_H_INGRESS")) on eth0 that has 8
transmission queues:
Thread 1 creates ingress Qdisc A (containing mini Qdisc a1 and a2),
then adds a flower filter X to A.
Thread 2 creates another ingress Qdisc B (containing mini Qdisc b1 and
b2) to replace A, then adds a flower filter Y to B.
Thread 1 A's refcnt Thread 2
RTM_NEWQDISC (A, RTNL-locked)
qdisc_create(A) 1
qdisc_graft(A) 9
RTM_NEWTFILTER (X, RTNL-unlocked)
__tcf_qdisc_find(A) 10
tcf_chain0_head_change(A)
mini_qdisc_pair_swap(A) (1st)
|
| RTM_NEWQDISC (B, RTNL-locked)
RCU sync 2 qdisc_graft(B)
| 1 notify_and_destroy(A)
|
tcf_block_release(A) 0 RTM_NEWTFILTER (Y, RTNL-unlocked)
qdisc_destroy(A) tcf_chain0_head_change(B)
tcf_chain0_head_change_cb_del(A) mini_qdisc_pair_swap(B) (2nd)
mini_qdisc_pair_swap(A) (3rd) |
... ...
Here, B calls mini_qdisc_pair_swap(), pointing eth0->miniq_ingress to
its mini Qdisc, b1. Then, A calls mini_qdisc_pair_swap() again during
ingress_destroy(), setting eth0->miniq_ingress to NULL, so ingress
packets on eth0 will not find filter Y in sch_handle_ingress().
This is just one of the possible consequences of concurrently accessing
miniq_{in,e}gress pointers.
Fixes: 7a096d579e ("net: sched: ingress: set 'unlocked' flag for Qdisc ops")
Fixes: 87f373921c ("net: sched: ingress: set 'unlocked' flag for clsact Qdisc ops")
Reported-by: syzbot+b53a9c0d1ea4ad62da8b@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/r/0000000000006cf87705f79acf1a@google.com/
Cc: Hillf Danton <hdanton@sina.com>
Cc: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Peilin Ye <peilin.ye@bytedance.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
For drivers to support partial offload of a filter's action list,
add support for action miss to specify an action instance to
continue from in sw.
CT action in particular can't be fully offloaded, as new connections
need to be handled in software. This imposes other limitations on
the actions that can be offloaded together with the CT action, such
as packet modifications.
Assign each action on a filter's action list a unique miss_cookie
which drivers can then use to fill action_miss part of the tc skb
extension. On getting back this miss_cookie, find the action
instance with relevant cookie and continue classifying from there.
Signed-off-by: Paul Blakey <paulb@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
syzbot reported a nasty crash [1] in net_tx_action() which
made little sense until we got a repro.
This repro installs a taprio qdisc, but providing an
invalid TCA_RATE attribute.
qdisc_create() has to destroy the just initialized
taprio qdisc, and taprio_destroy() is called.
However, the hrtimer used by taprio had already fired,
therefore advance_sched() called __netif_schedule().
Then net_tx_action was trying to use a destroyed qdisc.
We can not undo the __netif_schedule(), so we must wait
until one cpu serviced the qdisc before we can proceed.
Many thanks to Alexander Potapenko for his help.
[1]
BUG: KMSAN: uninit-value in queued_spin_trylock include/asm-generic/qspinlock.h:94 [inline]
BUG: KMSAN: uninit-value in do_raw_spin_trylock include/linux/spinlock.h:191 [inline]
BUG: KMSAN: uninit-value in __raw_spin_trylock include/linux/spinlock_api_smp.h:89 [inline]
BUG: KMSAN: uninit-value in _raw_spin_trylock+0x92/0xa0 kernel/locking/spinlock.c:138
queued_spin_trylock include/asm-generic/qspinlock.h:94 [inline]
do_raw_spin_trylock include/linux/spinlock.h:191 [inline]
__raw_spin_trylock include/linux/spinlock_api_smp.h:89 [inline]
_raw_spin_trylock+0x92/0xa0 kernel/locking/spinlock.c:138
spin_trylock include/linux/spinlock.h:359 [inline]
qdisc_run_begin include/net/sch_generic.h:187 [inline]
qdisc_run+0xee/0x540 include/net/pkt_sched.h:125
net_tx_action+0x77c/0x9a0 net/core/dev.c:5086
__do_softirq+0x1cc/0x7fb kernel/softirq.c:571
run_ksoftirqd+0x2c/0x50 kernel/softirq.c:934
smpboot_thread_fn+0x554/0x9f0 kernel/smpboot.c:164
kthread+0x31b/0x430 kernel/kthread.c:376
ret_from_fork+0x1f/0x30
Uninit was created at:
slab_post_alloc_hook mm/slab.h:732 [inline]
slab_alloc_node mm/slub.c:3258 [inline]
__kmalloc_node_track_caller+0x814/0x1250 mm/slub.c:4970
kmalloc_reserve net/core/skbuff.c:358 [inline]
__alloc_skb+0x346/0xcf0 net/core/skbuff.c:430
alloc_skb include/linux/skbuff.h:1257 [inline]
nlmsg_new include/net/netlink.h:953 [inline]
netlink_ack+0x5f3/0x12b0 net/netlink/af_netlink.c:2436
netlink_rcv_skb+0x55d/0x6c0 net/netlink/af_netlink.c:2507
rtnetlink_rcv+0x30/0x40 net/core/rtnetlink.c:6108
netlink_unicast_kernel net/netlink/af_netlink.c:1319 [inline]
netlink_unicast+0xf3b/0x1270 net/netlink/af_netlink.c:1345
netlink_sendmsg+0x1288/0x1440 net/netlink/af_netlink.c:1921
sock_sendmsg_nosec net/socket.c:714 [inline]
sock_sendmsg net/socket.c:734 [inline]
____sys_sendmsg+0xabc/0xe90 net/socket.c:2482
___sys_sendmsg+0x2a1/0x3f0 net/socket.c:2536
__sys_sendmsg net/socket.c:2565 [inline]
__do_sys_sendmsg net/socket.c:2574 [inline]
__se_sys_sendmsg net/socket.c:2572 [inline]
__x64_sys_sendmsg+0x367/0x540 net/socket.c:2572
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x3d/0xb0 arch/x86/entry/common.c:80
entry_SYSCALL_64_after_hwframe+0x63/0xcd
CPU: 0 PID: 13 Comm: ksoftirqd/0 Not tainted 6.0.0-rc2-syzkaller-47461-gac3859c02d7f #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 07/22/2022
Fixes: 5a781ccbd1 ("tc: Add support for configuring the taprio scheduler")
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When adding optional new features to Qdisc offloads, existing drivers
must reject the new configuration until they are coded up to act on it.
Since modifying all drivers in lockstep with the changes in the Qdisc
can create problems of its own, it would be nice if there existed an
automatic opt-in mechanism for offloading optional features.
Jakub proposes that we multiplex one more kind of call through
ndo_setup_tc(): one where the driver populates a Qdisc-specific
capability structure.
First user will be taprio in further changes. Here we are introducing
the definitions for the base functionality.
Link: https://patchwork.kernel.org/project/netdevbpf/patch/20220923163310.3192733-3-vladimir.oltean@nxp.com/
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Co-developed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The memory allocated by using kzallloc_node and kcalloc has been cleared.
Therefore, the structure members of the new qdisc are 0. So there's no
need to explicitly assign a value of 0.
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
qdisc_reset() is clearing qdisc->q.qlen and qdisc->qstats.backlog
_after_ calling qdisc->ops->reset. There is no need to clear them
again in the specific reset function.
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Link: https://lore.kernel.org/r/20220824005231.345727-1-shaozhengchao@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
In qdisc_run_end(), the spin_unlock() only has store-release semantic,
which guarantees all earlier memory access are visible before it. But
the subsequent test_bit() has no barrier semantics so may be reordered
ahead of the spin_unlock(). The store-load reordering may cause a packet
stuck problem.
The concurrent operations can be described as below,
CPU 0 | CPU 1
qdisc_run_end() | qdisc_run_begin()
. | .
----> /* may be reorderd here */ | .
| . | .
| spin_unlock() | set_bit()
| . | smp_mb__after_atomic()
---- test_bit() | spin_trylock()
. | .
Consider the following sequence of events:
CPU 0 reorder test_bit() ahead and see MISSED = 0
CPU 1 calls set_bit()
CPU 1 calls spin_trylock() and return fail
CPU 0 executes spin_unlock()
At the end of the sequence, CPU 0 calls spin_unlock() and does nothing
because it see MISSED = 0. The skb on CPU 1 has beed enqueued but no one
take it, until the next cpu pushing to the qdisc (if ever ...) will
notice and dequeue it.
This patch fix this by adding one explicit barrier. As spin_unlock() and
test_bit() ordering is a store-load ordering, a full memory barrier
smp_mb() is needed here.
Fixes: a90c57f2ce ("net: sched: fix packet stuck problem for lockless qdisc")
Signed-off-by: Guoju Fang <gjfang@linux.alibaba.com>
Link: https://lore.kernel.org/r/20220528101628.120193-1-gjfang@linux.alibaba.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In qdisc_run_begin(), smp_mb__before_atomic() used before test_bit()
does not provide any ordering guarantee as test_bit() is not an atomic
operation. This, added to the fact that the spin_trylock() call at
the beginning of qdisc_run_begin() does not guarantee acquire
semantics if it does not grab the lock, makes it possible for the
following statement :
if (test_bit(__QDISC_STATE_MISSED, &qdisc->state))
to be executed before an enqueue operation called before
qdisc_run_begin().
As a result the following race can happen :
CPU 1 CPU 2
qdisc_run_begin() qdisc_run_begin() /* true */
set(MISSED) .
/* returns false */ .
. /* sees MISSED = 1 */
. /* so qdisc not empty */
. __qdisc_run()
. .
. pfifo_fast_dequeue()
----> /* may be done here */ .
| . clear(MISSED)
| . .
| . smp_mb __after_atomic();
| . .
| . /* recheck the queue */
| . /* nothing => exit */
| enqueue(skb1)
| .
| qdisc_run_begin()
| .
| spin_trylock() /* fail */
| .
| smp_mb__before_atomic() /* not enough */
| .
---- if (test_bit(MISSED))
return false; /* exit */
In the above scenario, CPU 1 and CPU 2 both try to grab the
qdisc->seqlock at the same time. Only CPU 2 succeeds and enters the
bypass code path, where it emits its skb then calls __qdisc_run().
CPU1 fails, sets MISSED and goes down the traditionnal enqueue() +
dequeue() code path. But when executing qdisc_run_begin() for the
second time, after enqueuing its skbuff, it sees the MISSED bit still
set (by itself) and consequently chooses to exit early without setting
it again nor trying to grab the spinlock again.
Meanwhile CPU2 has seen MISSED = 1, cleared it, checked the queue
and found it empty, so it returned.
At the end of the sequence, we end up with skb1 enqueued in the
backlog, both CPUs out of __dev_xmit_skb(), the MISSED bit not set,
and no __netif_schedule() called made. skb1 will now linger in the
qdisc until somebody later performs a full __qdisc_run(). Associated
to the bypass capacity of the qdisc, and the ability of the TCP layer
to avoid resending packets which it knows are still in the qdisc, this
can lead to serious traffic "holes" in a TCP connection.
We fix this by replacing the smp_mb__before_atomic() / test_bit() /
set_bit() / smp_mb__after_atomic() sequence inside qdisc_run_begin()
by a single test_and_set_bit() call, which is more concise and
enforces the needed memory barriers.
Fixes: 89837eb4b2 ("net: sched: add barrier to ensure correct ordering for lockless qdisc")
Signed-off-by: Vincent Ray <vray@kalrayinc.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20220526001746.2437669-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
commit 56b765b79e ("htb: improved accuracy at high rates") broke
"overhead X", "linklayer atm" and "mpu X" attributes.
"overhead X" and "linklayer atm" have already been fixed. This restores
the "mpu X" handling, as might be used by DOCSIS or Ethernet shaping:
tc class add ... htb rate X overhead 4 mpu 64
The code being fixed is used by htb, tbf and act_police. Cake has its
own mpu handling. qdisc_calculate_pkt_len still uses the size table
containing values adjusted for mpu by user space.
iproute2 tc has always passed mpu into the kernel via a tc_ratespec
structure, but the kernel never directly acted on it, merely stored it
so that it could be read back by `tc class show`.
Rather, tc would generate length-to-time tables that included the mpu
(and linklayer) in their construction, and the kernel used those tables.
Since v3.7, the tables were no longer used. Along with "mpu", this also
broke "overhead" and "linklayer" which were fixed in 01cb71d2d4
("net_sched: restore "overhead xxx" handling", v3.10) and 8a8e3d84b1
("net_sched: restore "linklayer atm" handling", v3.11).
"overhead" was fixed by simply restoring use of tc_ratespec::overhead -
this had originally been used by the kernel but was initially omitted
from the new non-table-based calculations.
"linklayer" had been handled in the table like "mpu", but the mode was
not originally passed in tc_ratespec. The new implementation was made to
handle it by getting new versions of tc to pass the mode in an extended
tc_ratespec, and for older versions of tc the table contents were analysed
at load time to deduce linklayer.
As "mpu" has always been given to the kernel in tc_ratespec,
accompanying the mpu-based table, we can restore system functionality
with no userspace change by making the kernel act on the tc_ratespec
value.
Fixes: 56b765b79e ("htb: improved accuracy at high rates")
Signed-off-by: Kevin Bracey <kevin@bracey.fi>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Jiri Pirko <jiri@resnulli.us>
Cc: Vimalkumar <j.vimal@gmail.com>
Link: https://lore.kernel.org/r/20220112170210.1014351-1-kevin@bracey.fi
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
BPF layer extends the qdisc control block via struct bpf_skb_data_end
and because of that there is no more room to add variables to the
qdisc layer control block without going over the skb->cb size.
Extend the qdisc control block with a tc control block,
and move all tc related variables to there as a pre-step for
extending the tc control block with additional members.
Signed-off-by: Paul Blakey <paulb@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently rcu_barrier() is used to ensure that no readers of the
inactive mini_Qdisc buffer remain before it is reused. This waits for
any pending RCU callbacks to complete, when all that is actually
required is to wait for one RCU grace period to elapse after the buffer
was made inactive. This means that using rcu_barrier() may result in
unnecessary waits.
To improve this, store the current RCU state when a buffer is made
inactive and use poll_state_synchronize_rcu() to check whether a full
grace period has elapsed before reusing it. If a full grace period has
not elapsed, wait for a grace period to elapse, and in the non-RT case
use synchronize_rcu_expedited() to hasten it.
Since this approach eliminates the RCU callback it is no longer
necessary to synchronize_rcu() in the tp_head==NULL case. However, the
RCU state should still be saved for the previously active buffer.
Before this change I would typically see mini_qdisc_pair_swap() take
tens of milliseconds to complete. After this change it typcially
finishes in less than 1 ms, and often it takes just a few microseconds.
Thanks to Paul for walking me through the options for improving this.
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Signed-off-by: Seth Forshee <sforshee@digitalocean.com>
Link: https://lore.kernel.org/r/20211026130700.121189-1-seth@forshee.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
__QDISC_STATE_RUNNING is only set/cleared from contexts owning qdisc lock.
Thus we can use less expensive bit operations, as we were doing
before commit f9eb8aea2a ("net_sched: transform qdisc running bit into a seqcount")
Fixes: 29cbcd8582 ("net: sched: Remove Qdisc::running sequence counter")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ahmed S. Darwish <a.darwish@linutronix.de>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Tested-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For non TCQ_F_NOLOCK qdisc, qdisc_run_begin() tries to set
__QDISC_STATE_RUNNING and should return true if the bit was not set.
test_and_set_bit() returns old bit value, therefore we need to invert.
Fixes: 29cbcd8582 ("net: sched: Remove Qdisc::running sequence counter")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ahmed S. Darwish <a.darwish@linutronix.de>
Tested-by: Ido Schimmel <idosch@nvidia.com>
Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Tested-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The Qdisc::running sequence counter has two uses:
1. Reliably reading qdisc's tc statistics while the qdisc is running
(a seqcount read/retry loop at gnet_stats_add_basic()).
2. As a flag, indicating whether the qdisc in question is running
(without any retry loops).
For the first usage, the Qdisc::running sequence counter write section,
qdisc_run_begin() => qdisc_run_end(), covers a much wider area than what
is actually needed: the raw qdisc's bstats update. A u64_stats sync
point was thus introduced (in previous commits) inside the bstats
structure itself. A local u64_stats write section is then started and
stopped for the bstats updates.
Use that u64_stats sync point mechanism for the bstats read/retry loop
at gnet_stats_add_basic().
For the second qdisc->running usage, a __QDISC_STATE_RUNNING bit flag,
accessed with atomic bitops, is sufficient. Using a bit flag instead of
a sequence counter at qdisc_run_begin/end() and qdisc_is_running() leads
to the SMP barriers implicitly added through raw_read_seqcount() and
write_seqcount_begin/end() getting removed. All call sites have been
surveyed though, and no required ordering was identified.
Now that the qdisc->running sequence counter is no longer used, remove
it.
Note, using u64_stats implies no sequence counter protection for 64-bit
architectures. This can lead to the qdisc tc statistics "packets" vs.
"bytes" values getting out of sync on rare occasions. The individual
values will still be valid.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The only factor differentiating per-CPU bstats data type (struct
gnet_stats_basic_cpu) from the packed non-per-CPU one (struct
gnet_stats_basic_packed) was a u64_stats sync point inside the former.
The two data types are now equivalent: earlier commits added a u64_stats
sync point to the latter.
Combine both data types into "struct gnet_stats_basic_sync". This
eliminates redundancy and simplifies the bstats read/write APIs.
Use u64_stats_t for bstats "packets" and "bytes" data types. On 64-bit
architectures, u64_stats sync points do not use sequence counter
protection.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The not-per-CPU variant of qdisc tc (traffic control) statistics,
Qdisc::gnet_stats_basic_packed bstats, is protected with Qdisc::running
sequence counter.
This sequence counter is used for reliably protecting bstats reads from
parallel writes. Meanwhile, the seqcount's write section covers a much
wider area than bstats update: qdisc_run_begin() => qdisc_run_end().
That read/write section asymmetry can lead to needless retries of the
read section. To prepare for removing the Qdisc::running sequence
counter altogether, introduce a u64_stats sync point inside bstats
instead.
Modify _bstats_update() to start/end the bstats u64_stats write
section.
For bisectability, and finer commits granularity, the bstats read
section is still protected with a Qdisc::running read/retry loop and
qdisc_run_begin/end() still starts/ends that seqcount write section.
Once all call sites are modified to use _bstats_update(), the
Qdisc::running seqcount will be removed and bstats read/retry loop will
be modified to utilize the internal u64_stats sync point.
Note, using u64_stats implies no sequence counter protection for 64-bit
architectures. This can lead to the statistics "packets" vs. "bytes"
values getting out of sync on rare occasions. The individual values will
still be valid.
[bigeasy: Minor commit message edits, init all gnet_stats_basic_packed.]
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The gnet_stats_queue::qlen member is only used in the SMP-case.
qdisc_qstats_qlen_backlog() needs to add qdisc_qlen() to qstats.qlen to
have the same value as that provided by qdisc_qlen_sum().
gnet_stats_copy_queue() needs to overwritte the resulting qstats.qlen
field whith the caller submitted qlen value. It might be differ from the
submitted value.
Let both functions use gnet_stats_add_queue() and remove unused
__gnet_stats_copy_queue().
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The code for handling active queue changes is identical
between mq and mqprio, reuse it.
Suggested-by: Cong Wang <cong.wang@bytedance.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
mq / mqprio make the default child qdiscs visible. They only do
so for the qdiscs which are within real_num_tx_queues when the
device is registered. Depending on order of calls in the driver,
or if user space changes config via ethtool -L the number of
qdiscs visible under tc qdisc show will differ from the number
of queues. This is confusing to users and potentially to system
configuration scripts which try to make sure qdiscs have the
right parameters.
Add a new Qdisc_ops callback and make relevant qdiscs TTRT.
Note that this uncovers the "shortcut" created by
commit 1f27cde313 ("net: sched: use pfifo_fast for non real queues")
The default child qdiscs beyond initial real_num_tx are always
pfifo_fast, no matter what the sysfs setting is. Fixing this
gets a little tricky because we'd need to keep a reference
on whatever the default qdisc was at the time of creation.
In practice this is likely an non-issue the qdiscs likely have
to be configured to non-default settings, so whatever user space
is doing such configuration can replace the pfifos... now that
it will see them.
Reported-by: Matthew Massey <matthewmassey@fb.com>
Reviewed-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
TC action ->init() API has 10 parameters, it becomes harder
to read. Some of them are just boolean and can be replaced
by flags. Similarly for the internal API tcf_action_init()
and tcf_exts_validate().
This patch converts them to flags and fold them into
the upper 16 bits of "flags", whose lower 16 bits are still
reserved for user-space. More specifically, the following
kernel flags are introduced:
TCA_ACT_FLAGS_POLICE replace 'name' in a few contexts, to
distinguish whether it is compatible with policer.
TCA_ACT_FLAGS_BIND replaces 'bind', to indicate whether
this action is bound to a filter.
TCA_ACT_FLAGS_REPLACE replaces 'ovr' in most contexts,
means we are replacing an existing action.
TCA_ACT_FLAGS_NO_RTNL replaces 'rtnl_held' but has the
opposite meaning, because we still hold RTNL in most
cases.
The only user-space flag TCA_ACT_FLAGS_NO_PERCPU_STATS is
untouched and still stored as before.
I have tested this patch with tdc and I do not see any
failure related to this patch.
Tested-by: Vlad Buslov <vladbu@nvidia.com>
Acked-by: Jamal Hadi Salim<jhs@mojatatu.com>
Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: Cong Wang <cong.wang@bytedance.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Trivial conflict in net/netfilter/nf_tables_api.c.
Duplicate fix in tools/testing/selftests/net/devlink_port_split.py
- take the net-next version.
skmsg, and L4 bpf - keep the bpf code but remove the flags
and err params.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As MISSED and DRAINING state are used to indicate a non-empty
qdisc, qdisc->empty is not longer needed, so remove it.
Acked-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com> # flexcan
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK
flag set, but queue discipline by-pass does not work for lockless
qdisc because skb is always enqueued to qdisc even when the qdisc
is empty, see __dev_xmit_skb().
This patch calls sch_direct_xmit() to transmit the skb directly
to the driver for empty lockless qdisc, which aviod enqueuing
and dequeuing operation.
As qdisc->empty is not reliable to indicate a empty qdisc because
there is a time window between enqueuing and setting qdisc->empty.
So we use the MISSED state added in commit a90c57f2ce ("net:
sched: fix packet stuck problem for lockless qdisc"), which
indicate there is lock contention, suggesting that it is better
not to do the qdisc bypass in order to avoid packet out of order
problem.
In order to make MISSED state reliable to indicate a empty qdisc,
we need to ensure that testing and clearing of MISSED state is
within the protection of qdisc->seqlock, only setting MISSED state
can be done without the protection of qdisc->seqlock. A MISSED
state testing is added without the protection of qdisc->seqlock to
aviod doing unnecessary spin_trylock() for contention case.
As the enqueuing is not within the protection of qdisc->seqlock,
there is still a potential data race as mentioned by Jakub [1]:
thread1 thread2 thread3
qdisc_run_begin() # true
qdisc_run_begin(q)
set(MISSED)
pfifo_fast_dequeue
clear(MISSED)
# recheck the queue
qdisc_run_end()
enqueue skb1
qdisc empty # true
qdisc_run_begin() # true
sch_direct_xmit() # skb2
qdisc_run_begin()
set(MISSED)
When above happens, skb1 enqueued by thread2 is transmited after
skb2 is transmited by thread3 because MISSED state setting and
enqueuing is not under the qdisc->seqlock. If qdisc bypass is
disabled, skb1 has better chance to be transmited quicker than
skb2.
This patch does not take care of the above data race, because we
view this as similar as below:
Even at the same time CPU1 and CPU2 write the skb to two socket
which both heading to the same qdisc, there is no guarantee that
which skb will hit the qdisc first, because there is a lot of
factor like interrupt/softirq/cache miss/scheduling afffecting
that.
There are below cases that need special handling:
1. When MISSED state is cleared before another round of dequeuing
in pfifo_fast_dequeue(), and __qdisc_run() might not be able to
dequeue all skb in one round and call __netif_schedule(), which
might result in a non-empty qdisc without MISSED set. In order
to avoid this, the MISSED state is set for lockless qdisc and
__netif_schedule() will be called at the end of qdisc_run_end.
2. The MISSED state also need to be set for lockless qdisc instead
of calling __netif_schedule() directly when requeuing a skb for
a similar reason.
3. For netdev queue stopped case, the MISSED case need clearing
while the netdev queue is stopped, otherwise there may be
unnecessary __netif_schedule() calling. So a new DRAINING state
is added to indicate this case, which also indicate a non-empty
qdisc.
4. As there is already netif_xmit_frozen_or_stopped() checking in
dequeue_skb() and sch_direct_xmit(), which are both within the
protection of qdisc->seqlock, but the same checking in
__dev_xmit_skb() is without the protection, which might cause
empty indication of a lockless qdisc to be not reliable. So
remove the checking in __dev_xmit_skb(), and the checking in
the protection of qdisc->seqlock seems enough to avoid the cpu
consumption problem for netdev queue stopped case.
1. https://lkml.org/lkml/2021/5/29/215
Acked-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com> # flexcan
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
qdisc->running seqcount operation is mainly used to do heuristic
locking on q->busylock for locked qdisc, see qdisc_is_running()
and __dev_xmit_skb().
So avoid doing seqcount operation for qdisc with TCQ_F_NOLOCK
flag.
Acked-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com> # flexcan
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The spin_trylock() was assumed to contain the implicit
barrier needed to ensure the correct ordering between
STATE_MISSED setting/clearing and STATE_MISSED checking
in commit a90c57f2ce ("net: sched: fix packet stuck
problem for lockless qdisc").
But it turns out that spin_trylock() only has load-acquire
semantic, for strongly-ordered system(like x86), the compiler
barrier implicitly contained in spin_trylock() seems enough
to ensure the correct ordering. But for weakly-orderly system
(like arm64), the store-release semantic is needed to ensure
the correct ordering as clear_bit() and test_bit() is store
operation, see queued_spin_lock().
So add the explicit barrier to ensure the correct ordering
for the above case.
Fixes: a90c57f2ce ("net: sched: fix packet stuck problem for lockless qdisc")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Lockless qdisc has below concurrent problem:
cpu0 cpu1
. .
q->enqueue .
. .
qdisc_run_begin() .
. .
dequeue_skb() .
. .
sch_direct_xmit() .
. .
. q->enqueue
. qdisc_run_begin()
. return and do nothing
. .
qdisc_run_end() .
cpu1 enqueue a skb without calling __qdisc_run() because cpu0
has not released the lock yet and spin_trylock() return false
for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb
enqueued by cpu1 when calling dequeue_skb() because cpu1 may
enqueue the skb after cpu0 calling dequeue_skb() and before
cpu0 calling qdisc_run_end().
Lockless qdisc has below another concurrent problem when
tx_action is involved:
cpu0(serving tx_action) cpu1 cpu2
. . .
. q->enqueue .
. qdisc_run_begin() .
. dequeue_skb() .
. . q->enqueue
. . .
. sch_direct_xmit() .
. . qdisc_run_begin()
. . return and do nothing
. . .
clear __QDISC_STATE_SCHED . .
qdisc_run_begin() . .
return and do nothing . .
. . .
. qdisc_run_end() .
This patch fixes the above data race by:
1. If the first spin_trylock() return false and STATE_MISSED is
not set, set STATE_MISSED and retry another spin_trylock() in
case other CPU may not see STATE_MISSED after it releases the
lock.
2. reschedule if STATE_MISSED is set after the lock is released
at the end of qdisc_run_end().
For tx_action case, STATE_MISSED is also set when cpu1 is at the
end if qdisc_run_end(), so tx_action will be rescheduled again
to dequeue the skb enqueued by cpu2.
Clear STATE_MISSED before retrying a dequeuing when dequeuing
returns NULL in order to reduce the overhead of the second
spin_trylock() and __netif_schedule() calling.
Also clear the STATE_MISSED before calling __netif_schedule()
at the end of qdisc_run_end() to avoid doing another round of
dequeuing in the pfifo_fast_dequeue().
The performance impact of this patch, tested using pktgen and
dummy netdev with pfifo_fast qdisc attached:
threads without+this_patch with+this_patch delta
1 2.61Mpps 2.60Mpps -0.3%
2 3.97Mpps 3.82Mpps -3.7%
4 5.62Mpps 5.59Mpps -0.5%
8 2.78Mpps 2.77Mpps -0.3%
16 2.22Mpps 2.22Mpps -0.0%
Fixes: 6b3ba9146f ("net: sched: allow qdiscs to handle locking")
Acked-by: Jakub Kicinski <kuba@kernel.org>
Tested-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow a policer action to enforce a rate-limit based on packets-per-second,
configurable using a packet-per-second rate and burst parameters.
e.g.
tc filter add dev tap1 parent ffff: u32 match \
u32 0 0 police pkts_rate 3000 pkts_burst 1000
Testing was unable to uncover a performance impact of this change on
existing features.
Signed-off-by: Baowen Zheng <baowen.zheng@corigine.com>
Signed-off-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: Louis Peens <louis.peens@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
