It can be reproduced by following steps:
1. virtio_net NIC is configured with gso/tso on
2. configure nginx as http server with an index file bigger than 1M bytes
3. use tc netem to produce duplicate packets and delay:
tc qdisc add dev eth0 root netem delay 100ms 10ms 30% duplicate 90%
4. continually curl the nginx http server to get index file on client
5. BUG_ON is seen quickly
[10258690.371129] kernel BUG at net/core/skbuff.c:4028!
[10258690.371748] invalid opcode: 0000 [#1] SMP PTI
[10258690.372094] CPU: 5 PID: 0 Comm: swapper/5 Tainted: G W 5.0.0-rc6 #2
[10258690.372094] RSP: 0018:ffffa05797b43da0 EFLAGS: 00010202
[10258690.372094] RBP: 00000000000005ea R08: 0000000000000000 R09: 00000000000005ea
[10258690.372094] R10: ffffa0579334d800 R11: 00000000000002c0 R12: 0000000000000002
[10258690.372094] R13: 0000000000000000 R14: ffffa05793122900 R15: ffffa0578f7cb028
[10258690.372094] FS: 0000000000000000(0000) GS:ffffa05797b40000(0000) knlGS:0000000000000000
[10258690.372094] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[10258690.372094] CR2: 00007f1a6dc00868 CR3: 000000001000e000 CR4: 00000000000006e0
[10258690.372094] Call Trace:
[10258690.372094] <IRQ>
[10258690.372094] skb_to_sgvec+0x11/0x40
[10258690.372094] start_xmit+0x38c/0x520 [virtio_net]
[10258690.372094] dev_hard_start_xmit+0x9b/0x200
[10258690.372094] sch_direct_xmit+0xff/0x260
[10258690.372094] __qdisc_run+0x15e/0x4e0
[10258690.372094] net_tx_action+0x137/0x210
[10258690.372094] __do_softirq+0xd6/0x2a9
[10258690.372094] irq_exit+0xde/0xf0
[10258690.372094] smp_apic_timer_interrupt+0x74/0x140
[10258690.372094] apic_timer_interrupt+0xf/0x20
[10258690.372094] </IRQ>
In __skb_to_sgvec(), the skb->len is not equal to the sum of the skb's
linear data size and nonlinear data size, thus BUG_ON triggered.
Because the skb is cloned and a part of nonlinear data is split off.
Duplicate packet is cloned in netem_enqueue() and may be delayed
some time in qdisc. When qdisc len reached the limit and returns
NET_XMIT_DROP, the skb will be retransmit later in write queue.
the skb will be fragmented by tso_fragment(), the limit size
that depends on cwnd and mss decrease, the skb's nonlinear
data will be split off. The length of the skb cloned by netem
will not be updated. When we use virtio_net NIC and invoke skb_to_sgvec(),
the BUG_ON trigger.
To fix it, netem returns NET_XMIT_SUCCESS to upper stack
when it clones a duplicate packet.
Fixes: 35d889d1 ("sch_netem: fix skb leak in netem_enqueue()")
Signed-off-by: Sheng Lan <lansheng@huawei.com>
Reported-by: Qin Ji <jiqin.ji@huawei.com>
Suggested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several conflicts, seemingly all over the place.
I used Stephen Rothwell's sample resolutions for many of these, if not
just to double check my own work, so definitely the credit largely
goes to him.
The NFP conflict consisted of a bug fix (moving operations
past the rhashtable operation) while chaning the initial
argument in the function call in the moved code.
The net/dsa/master.c conflict had to do with a bug fix intermixing of
making dsa_master_set_mtu() static with the fixing of the tagging
attribute location.
cls_flower had a conflict because the dup reject fix from Or
overlapped with the addition of port range classifiction.
__set_phy_supported()'s conflict was relatively easy to resolve
because Andrew fixed it in both trees, so it was just a matter
of taking the net-next copy. Or at least I think it was :-)
Joe Stringer's fix to the handling of netns id 0 in bpf_sk_lookup()
intermixed with changes on how the sdif and caller_net are calculated
in these code paths in net-next.
The remaining BPF conflicts were largely about the addition of the
__bpf_md_ptr stuff in 'net' overlapping with adjustments and additions
to the relevant data structure where the MD pointer macros are used.
Signed-off-by: David S. Miller <davem@davemloft.net>
When testing high-bandwidth TCP streams with large windows,
high latency, and low jitter, netem consumes a lot of CPU cycles
doing rbtree rebalancing.
This patch uses a linear list/queue in addition to the rbtree:
if an incoming packet is past the tail of the linear queue, it is
added there, otherwise it is inserted into the rbtree.
Without this patch, perf shows netem_enqueue, netem_dequeue,
and rb_* functions among the top offenders. With this patch,
only netem_enqueue is noticeable if jitter is low/absent.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Peter Oskolkov <posk@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If sch_fq is used at ingress, skbs that might have been
timestamped by net_timestamp_set() if a packet capture
is requesting timestamps could be delayed by arbitrary
amount of time, since sch_fq time base is MONOTONIC.
Fix this problem by moving code from sch_netem.c to act_mirred.c.
Fixes: fb420d5d91 ("tcp/fq: move back to CLOCK_MONOTONIC")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current implementation of qdisc_destroy() decrements Qdisc reference
counter and only actually destroy Qdisc if reference counter value reached
zero. Rename qdisc_destroy() to qdisc_put() in order for it to better
describe the way in which this function currently implemented and used.
Extract code that deallocates Qdisc into new private qdisc_destroy()
function. It is intended to be shared between regular qdisc_put() and its
unlocked version that is introduced in next patch in this series.
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
An SKB is not on a list if skb->next is NULL.
Codify this convention into a helper function and use it
where we are dequeueing an SKB and need to mark it as such.
Signed-off-by: David S. Miller <davem@davemloft.net>
Extend slotting with support for non-uniform distributions. This is
similar to netem's non-uniform distribution delay feature.
Commit f043efeae2f1 ("netem: support delivering packets in delayed
time slots") added the slotting feature to approximate the behaviors
of media with packet aggregation but only supported a uniform
distribution for delays between transmission attempts. Tests with TCP
BBR with emulated wifi links with non-uniform distributions produced
more useful results.
Syntax:
slot dist DISTRIBUTION DELAY JITTER [packets MAX_PACKETS] \
[bytes MAX_BYTES]
The syntax and use of the distribution table is the same as in the
non-uniform distribution delay feature. A file DISTRIBUTION must be
present in TC_LIB_DIR (e.g. /usr/lib/tc) containing numbers scaled by
NETEM_DIST_SCALE. A random value x is selected from the table and it
takes DELAY + ( x * JITTER ) as delay. Correlation between values is not
supported.
Examples:
Normal distribution delay with mean = 800us and stdev = 100us.
> tc qdisc add dev eth0 root netem slot dist normal 800us 100us
Optionally set the max slot size in bytes and/or packets.
> tc qdisc add dev eth0 root netem slot dist normal 800us 100us \
bytes 64k packets 42
Signed-off-by: Yousuk Seung <ysseung@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we exceed current packets limit and we have more than one
segment in the list returned by skb_gso_segment(), netem drops
only the first one, skipping the rest, hence kmemleak reports:
unreferenced object 0xffff880b5d23b600 (size 1024):
comm "softirq", pid 0, jiffies 4384527763 (age 2770.629s)
hex dump (first 32 bytes):
00 80 23 5d 0b 88 ff ff 00 00 00 00 00 00 00 00 ..#]............
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace:
[<00000000d8a19b9d>] __alloc_skb+0xc9/0x520
[<000000001709b32f>] skb_segment+0x8c8/0x3710
[<00000000c7b9bb88>] tcp_gso_segment+0x331/0x1830
[<00000000c921cba1>] inet_gso_segment+0x476/0x1370
[<000000008b762dd4>] skb_mac_gso_segment+0x1f9/0x510
[<000000002182660a>] __skb_gso_segment+0x1dd/0x620
[<00000000412651b9>] netem_enqueue+0x1536/0x2590 [sch_netem]
[<0000000005d3b2a9>] __dev_queue_xmit+0x1167/0x2120
[<00000000fc5f7327>] ip_finish_output2+0x998/0xf00
[<00000000d309e9d3>] ip_output+0x1aa/0x2c0
[<000000007ecbd3a4>] tcp_transmit_skb+0x18db/0x3670
[<0000000042d2a45f>] tcp_write_xmit+0x4d4/0x58c0
[<0000000056a44199>] tcp_tasklet_func+0x3d9/0x540
[<0000000013d06d02>] tasklet_action+0x1ca/0x250
[<00000000fcde0b8b>] __do_softirq+0x1b4/0x5a3
[<00000000e7ed027c>] irq_exit+0x1e2/0x210
Fix it by adding the rest of the segments, if any, to skb 'to_free'
list. Add new __qdisc_drop_all() and qdisc_drop_all() functions
because they can be useful in the future if we need to drop segmented
GSO packets in other places.
Fixes: 6071bd1aa1 ("netem: Segment GSO packets on enqueue")
Signed-off-by: Alexey Kodanev <alexey.kodanev@oracle.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In Kernel 4.15.0+, Netem does not work properly.
Netem setup:
tc qdisc add dev h1-eth0 root handle 1: netem delay 10ms 2ms
Result:
PING 172.16.101.2 (172.16.101.2) 56(84) bytes of data.
64 bytes from 172.16.101.2: icmp_seq=1 ttl=64 time=22.8 ms
64 bytes from 172.16.101.2: icmp_seq=2 ttl=64 time=10.9 ms
64 bytes from 172.16.101.2: icmp_seq=3 ttl=64 time=10.9 ms
64 bytes from 172.16.101.2: icmp_seq=5 ttl=64 time=11.4 ms
64 bytes from 172.16.101.2: icmp_seq=6 ttl=64 time=11.8 ms
64 bytes from 172.16.101.2: icmp_seq=4 ttl=64 time=4303 ms
64 bytes from 172.16.101.2: icmp_seq=10 ttl=64 time=11.2 ms
64 bytes from 172.16.101.2: icmp_seq=11 ttl=64 time=10.3 ms
64 bytes from 172.16.101.2: icmp_seq=7 ttl=64 time=4304 ms
64 bytes from 172.16.101.2: icmp_seq=8 ttl=64 time=4303 ms
Patch:
(rnd % (2 * sigma)) - sigma was overflowing s32. After applying the
patch, I found following output which is desirable.
PING 172.16.101.2 (172.16.101.2) 56(84) bytes of data.
64 bytes from 172.16.101.2: icmp_seq=1 ttl=64 time=21.1 ms
64 bytes from 172.16.101.2: icmp_seq=2 ttl=64 time=8.46 ms
64 bytes from 172.16.101.2: icmp_seq=3 ttl=64 time=9.00 ms
64 bytes from 172.16.101.2: icmp_seq=4 ttl=64 time=11.8 ms
64 bytes from 172.16.101.2: icmp_seq=5 ttl=64 time=8.36 ms
64 bytes from 172.16.101.2: icmp_seq=6 ttl=64 time=11.8 ms
64 bytes from 172.16.101.2: icmp_seq=7 ttl=64 time=8.11 ms
64 bytes from 172.16.101.2: icmp_seq=8 ttl=64 time=10.0 ms
64 bytes from 172.16.101.2: icmp_seq=9 ttl=64 time=11.3 ms
64 bytes from 172.16.101.2: icmp_seq=10 ttl=64 time=11.5 ms
64 bytes from 172.16.101.2: icmp_seq=11 ttl=64 time=10.2 ms
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds extack support for graft callback to prepare per-qdisc
specific changes for extack.
Cc: David Ahern <dsahern@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Alexander Aring <aring@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds extack support for change callback for qdisc ops
structtur to prepare per-qdisc specific changes for extack.
Cc: David Ahern <dsahern@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Alexander Aring <aring@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds extack support for init callback to prepare per-qdisc
specific changes for extack.
Cc: David Ahern <dsahern@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Alexander Aring <aring@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix compilation on 32 bit platforms (where doing modulus operation
with 64 bit requires extra glibc functions) by truncation.
The jitter for table distribution is limited to a 32 bit value
because random numbers are scaled as 32 bit value.
Also fix some whitespace.
Fixes: 99803171ef ("netem: add uapi to express delay and jitter in nanoseconds")
Reported-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since times are now expressed in nanosecond, need to now do
true 64 bit divide. Old code would truncate rate at 32 bits.
Rename function to better express current usage.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Slotting is a crude approximation of the behaviors of shared media such
as cable, wifi, and LTE, which gather up a bunch of packets within a
varying delay window and deliver them, relative to that, nearly all at
once.
It works within the existing loss, duplication, jitter and delay
parameters of netem. Some amount of inherent latency must be specified,
regardless.
The new "slot" parameter specifies a minimum and maximum delay between
transmission attempts.
The "bytes" and "packets" parameters can be used to limit the amount of
information transferred per slot.
Examples of use:
tc qdisc add dev eth0 root netem delay 200us \
slot 800us 10ms bytes 64k packets 42
A more correct example, using stacked netem instances and a packet limit
to emulate a tail drop wifi queue with slots and variable packet
delivery, with a 200Mbit isochronous underlying rate, and 20ms path
delay:
tc qdisc add dev eth0 root handle 1: netem delay 20ms rate 200mbit \
limit 10000
tc qdisc add dev eth0 parent 1:1 handle 10:1 netem delay 200us \
slot 800us 10ms bytes 64k packets 42 limit 512
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
netem userspace has long relied on a horrible /proc/net/psched hack
to translate the current notion of "ticks" to nanoseconds.
Expressing latency and jitter instead, in well defined nanoseconds,
increases the dynamic range of emulated delays and jitter in netem.
It will also ease a transition where reducing a tick to nsec
equivalence would constrain the max delay in prior versions of
netem to only 4.3 seconds.
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Upgrade the internal netem scheduler to use nanoseconds rather than
ticks throughout.
Convert to and from the std "ticks" userspace api automatically,
while allowing for finer grained scheduling to take place.
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Geeralize private netem_rb_to_skb()
TCP rtx queue will soon be converted to rb-tree,
so we will need skb_rbtree_walk() helpers.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While running TCP tests involving netem storing millions of packets,
I had the idea to speed up tfifo_reset() and did experiments.
I tried the rbtree_postorder_for_each_entry_safe() method that is
used in skb_rbtree_purge() but discovered it was slower than the
current tfifo_reset() method.
I measured time taken to release skbs with three occupation levels :
10^4, 10^5 and 10^6 skbs with three methods :
1) (current 'naive' method)
while ((p = rb_first(&q->t_root))) {
struct sk_buff *skb = netem_rb_to_skb(p);
rb_erase(p, &q->t_root);
rtnl_kfree_skbs(skb, skb);
}
2) Use rb_next() instead of rb_first() in the loop :
p = rb_first(&q->t_root);
while (p) {
struct sk_buff *skb = netem_rb_to_skb(p);
p = rb_next(p);
rb_erase(&skb->rbnode, &q->t_root);
rtnl_kfree_skbs(skb, skb);
}
3) "optimized" method using rbtree_postorder_for_each_entry_safe()
struct sk_buff *skb, *next;
rbtree_postorder_for_each_entry_safe(skb, next,
&q->t_root, rbnode) {
rtnl_kfree_skbs(skb, skb);
}
q->t_root = RB_ROOT;
Results :
method_1:while (rb_first()) rb_erase() 10000 skbs in 690378 ns (69 ns per skb)
method_2:rb_first; while (p) { p = rb_next(p); ...} 10000 skbs in 541846 ns (54 ns per skb)
method_3:rbtree_postorder_for_each_entry_safe() 10000 skbs in 868307 ns (86 ns per skb)
method_1:while (rb_first()) rb_erase() 99996 skbs in 7804021 ns (78 ns per skb)
method_2:rb_first; while (p) { p = rb_next(p); ...} 100000 skbs in 5942456 ns (59 ns per skb)
method_3:rbtree_postorder_for_each_entry_safe() 100000 skbs in 11584940 ns (115 ns per skb)
method_1:while (rb_first()) rb_erase() 1000000 skbs in 108577838 ns (108 ns per skb)
method_2:rb_first; while (p) { p = rb_next(p); ...} 1000000 skbs in 82619635 ns (82 ns per skb)
method_3:rbtree_postorder_for_each_entry_safe() 1000000 skbs in 127328743 ns (127 ns per skb)
Method 2) is simply faster, probably because it maintains a smaller
working size set.
Note that this is the method we use in tcp_ofo_queue() already.
I will also change skb_rbtree_purge() in a second patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb->rbnode shares space with skb->next, skb->prev and skb->tstamp
Current uses (TCP receive ofo queue and netem) need to save/restore
tstamp, while skb->dev is either NULL (TCP) or a constant for a given
queue (netem).
Since we plan using an RB tree for TCP retransmit queue to speedup SACK
processing with large BDP, this patch exchanges skb->dev and
skb->tstamp.
This saves some overhead in both TCP and netem.
v2: removes the swtstamp field from struct tcp_skb_cb
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Wei Wang <weiwan@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
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: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are many code paths opencoding kvmalloc. Let's use the helper
instead. The main difference to kvmalloc is that those users are
usually not considering all the aspects of the memory allocator. E.g.
allocation requests <= 32kB (with 4kB pages) are basically never failing
and invoke OOM killer to satisfy the allocation. This sounds too
disruptive for something that has a reasonable fallback - the vmalloc.
On the other hand those requests might fallback to vmalloc even when the
memory allocator would succeed after several more reclaim/compaction
attempts previously. There is no guarantee something like that happens
though.
This patch converts many of those places to kv[mz]alloc* helpers because
they are more conservative.
Link: http://lkml.kernel.org/r/20170306103327.2766-2-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com> # Xen bits
Acked-by: Kees Cook <keescook@chromium.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Andreas Dilger <andreas.dilger@intel.com> # Lustre
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com> # KVM/s390
Acked-by: Dan Williams <dan.j.williams@intel.com> # nvdim
Acked-by: David Sterba <dsterba@suse.com> # btrfs
Acked-by: Ilya Dryomov <idryomov@gmail.com> # Ceph
Acked-by: Tariq Toukan <tariqt@mellanox.com> # mlx4
Acked-by: Leon Romanovsky <leonro@mellanox.com> # mlx5
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Anton Vorontsov <anton@enomsg.org>
Cc: Colin Cross <ccross@android.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: Ben Skeggs <bskeggs@redhat.com>
Cc: Kent Overstreet <kent.overstreet@gmail.com>
Cc: Santosh Raspatur <santosh@chelsio.com>
Cc: Hariprasad S <hariprasad@chelsio.com>
Cc: Yishai Hadas <yishaih@mellanox.com>
Cc: Oleg Drokin <oleg.drokin@intel.com>
Cc: "Yan, Zheng" <zyan@redhat.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pass the new extended ACK reporting struct to all of the generic
netlink parsing functions. For now, pass NULL in almost all callers
(except for some in the core.)
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
I recently reported on the netem list that iperf network benchmarks
show unexpected results when a bandwidth throttling rate has been
configured for netem. Specifically:
1) The measured link bandwidth *increases* when a higher delay is added
2) The measured link bandwidth appears higher than the specified limit
3) The measured link bandwidth for the same very slow settings varies significantly across
machines
The issue can be reproduced by using tc to configure netem with a
512kbit rate and various (none, 1us, 50ms, 100ms, 200ms) delays on a
veth pair between network namespaces, and then using iperf (or any
other network benchmarking tool) to test throughput. Complete detailed
instructions are in the original email chain here:
https://lists.linuxfoundation.org/pipermail/netem/2017-February/001672.html
There appear to be two underlying bugs causing these effects:
- The first issue causes long delays when the rate is slow and no
delay is configured (e.g., "rate 512kbit"). This is because SKBs are
not orphaned when no delay is configured, so orphaning does not
occur until *after* the rate-induced delay has been applied. For
this reason, adding a tiny delay (e.g., "rate 512kbit delay 1us")
dramatically increases the measured bandwidth.
- The second issue is that rate-induced delays are not correctly
applied, allowing SKB delays to occur in parallel. The indended
approach is to compute the delay for an SKB and to add this delay to
the end of the current queue. However, the code does not detect
existing SKBs in the queue due to improperly testing sch->q.qlen,
which is nonzero even when packets exist only in the
rbtree. Consequently, new SKBs do not wait for the current queue to
empty. When packet delays vary significantly (e.g., if packet sizes
are different), then this also causes unintended reordering.
I modified the code to expect a delay (and orphan the SKB) when a rate
is configured. I also added some defensive tests that correctly find
the latest scheduled delivery time, even if it is (unexpectedly) for a
packet in sch->q. I have tested these changes on the latest kernel
(4.11.0-rc1+) and the iperf / ping test results are as expected.
Signed-off-by: Nik Unger <njunger@uwaterloo.ca>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tc_from field fulfills two roles. It encodes whether a packet was
redirected by an act_mirred device and, if so, whether act_mirred was
called on ingress or egress. Split it into separate fields.
The information is needed by the special IFB loop, where packets are
taken out of the normal path by act_mirred, forwarded to IFB, then
reinjected at their original location (ingress or egress) by IFB.
The IFB device cannot use skb->tc_at_ingress, because that may have
been overwritten as the packet travels from act_mirred to ifb_xmit,
when it passes through tc_classify on the IFB egress path. Cache this
value in skb->tc_from_ingress.
That field is valid only if a packet arriving at ifb_xmit came from
act_mirred. Other packets can be crafted to reach ifb_xmit. These
must be dropped. Set tc_redirected on redirection and drop all packets
that do not have this bit set.
Both fields are set only on cloned skbs in tc actions, so original
packet sources do not have to clear the bit when reusing packets
(notably, pktgen and octeon).
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Extract the remaining two fields from tc_verd and remove the __u16
completely. TC_AT and TC_FROM are converted to equivalent two-bit
integer fields tc_at and tc_from. Where possible, use existing
helper skb_at_tc_ingress when reading tc_at. Introduce helper
skb_reset_tc to clear fields.
Not documenting tc_from and tc_at, because they will be replaced
with single bit fields in follow-on patches.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ktime is a union because the initial implementation stored the time in
scalar nanoseconds on 64 bit machine and in a endianess optimized timespec
variant for 32bit machines. The Y2038 cleanup removed the timespec variant
and switched everything to scalar nanoseconds. The union remained, but
become completely pointless.
Get rid of the union and just keep ktime_t as simple typedef of type s64.
The conversion was done with coccinelle and some manual mopping up.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
To make the code clearer, use rb_entry() instead of container_of() to
deal with rbtree.
Signed-off-by: Geliang Tang <geliangtang@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This change replaces sk_buff_head struct in Qdiscs with new qdisc_skb_head.
Its similar to the skb_buff_head api, but does not use skb->prev pointers.
Qdiscs will commonly enqueue at the tail of a list and dequeue at head.
While skb_buff_head works fine for this, enqueue/dequeue needs to also
adjust the prev pointer of next element.
The ->prev pointer is not required for qdiscs so we can just leave
it undefined and avoid one cacheline write access for en/dequeue.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
After previous patch these functions are identical.
Replace __skb_dequeue in qdiscs with __qdisc_dequeue_head.
Next patch will then make __qdisc_dequeue_head handle
single-linked list instead of strcut sk_buff_head argument.
Doesn't change generated code.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
A followup change will replace the sk_buff_head in the qdisc
struct with a slightly different list.
Use of the sk_buff_head helpers will thus cause compiler
warnings.
Open-code these accesses in an extra change to ease review.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several cases of overlapping changes, except the packet scheduler
conflicts which deal with the addition of the free list parameter
to qdisc_enqueue().
Signed-off-by: David S. Miller <davem@davemloft.net>
If skb_unshare() fails, we call qdisc_drop() with a NULL skb, which
is no longer supported.
Fixes: 520ac30f45 ("net_sched: drop packets after root qdisc lock is released")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Qdisc performance suffers when packets are dropped at enqueue()
time because drops (kfree_skb()) are done while qdisc lock is held,
delaying a dequeue() draining the queue.
Nominal throughput can be reduced by 50 % when this happens,
at a time we would like the dequeue() to proceed as fast as possible.
Even FQ is vulnerable to this problem, while one of FQ goals was
to provide some flow isolation.
This patch adds a 'struct sk_buff **to_free' parameter to all
qdisc->enqueue(), and in qdisc_drop() helper.
I measured a performance increase of up to 12 %, but this patch
is a prereq so that future batches in enqueue() can fly.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If the packet was dropped by lower qdisc, then we must not
access it later.
Save qdisc_pkt_len(skb) in a temp variable.
Fixes: 2ccccf5fb4 ("net_sched: update hierarchical backlog too")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: WANG Cong <xiyou.wangcong@gmail.com>
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
rtnl_kfree_skbs() can be used in tfifo_reset()
It would be nice if we could iterate through rb tree instead
of removing one skb at a time, and build a single skb chain.
But this is left for a future patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
__QDISC_STATE_THROTTLED bit manipulation is rather expensive
for HTB and few others.
I already removed it for sch_fq in commit f2600cf02b
("net: sched: avoid costly atomic operation in fq_dequeue()")
and so far nobody complained.
When one ore more packets are stuck in one or more throttled
HTB class, a htb dequeue() performs two atomic operations
to clear/set __QDISC_STATE_THROTTLED bit, while root qdisc
lock is held.
Removing this pair of atomic operations bring me a 8 % performance
increase on 200 TCP_RR tests, in presence of throttled classes.
This patch has no side effect, since nothing actually uses
disc_is_throttled() anymore.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Looks like it is only there as some optimization attempt.
Since __QDISC_STATE_THROTTLED set/unset is way too expensive,
and netem is the last user, just remove this check.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
after removal of TCA_CBQ_OVL_STRATEGY from cbq scheduler, there are no
more callers of ->drop() outside of other ->drop functions, i.e.
nothing calls them.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
After the removal of TCA_CBQ_POLICE in cbq scheduler qdisc->reshape_fail
is always NULL, i.e. qdisc_rehape_fail is now the same as qdisc_drop.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
net/ipv4/ip_gre.c
Minor conflicts between tunnel bug fixes in net and
ipv6 tunnel cleanups in net-next.
Signed-off-by: David S. Miller <davem@davemloft.net>
This was recently reported to me, and reproduced on the latest net kernel,
when attempting to run netperf from a host that had a netem qdisc attached
to the egress interface:
[ 788.073771] ---------------------[ cut here ]---------------------------
[ 788.096716] WARNING: at net/core/dev.c:2253 skb_warn_bad_offload+0xcd/0xda()
[ 788.129521] bnx2: caps=(0x00000001801949b3, 0x0000000000000000) len=2962
data_len=0 gso_size=1448 gso_type=1 ip_summed=3
[ 788.182150] Modules linked in: sch_netem kvm_amd kvm crc32_pclmul ipmi_ssif
ghash_clmulni_intel sp5100_tco amd64_edac_mod aesni_intel lrw gf128mul
glue_helper ablk_helper edac_mce_amd cryptd pcspkr sg edac_core hpilo ipmi_si
i2c_piix4 k10temp fam15h_power hpwdt ipmi_msghandler shpchp acpi_power_meter
pcc_cpufreq nfsd auth_rpcgss nfs_acl lockd grace sunrpc ip_tables xfs libcrc32c
sd_mod crc_t10dif crct10dif_generic mgag200 syscopyarea sysfillrect sysimgblt
i2c_algo_bit drm_kms_helper ahci ata_generic pata_acpi ttm libahci
crct10dif_pclmul pata_atiixp tg3 libata crct10dif_common drm crc32c_intel ptp
serio_raw bnx2 r8169 hpsa pps_core i2c_core mii dm_mirror dm_region_hash dm_log
dm_mod
[ 788.465294] CPU: 16 PID: 0 Comm: swapper/16 Tainted: G W
------------ 3.10.0-327.el7.x86_64 #1
[ 788.511521] Hardware name: HP ProLiant DL385p Gen8, BIOS A28 12/17/2012
[ 788.542260] ffff880437c036b8 f7afc56532a53db9 ffff880437c03670
ffffffff816351f1
[ 788.576332] ffff880437c036a8 ffffffff8107b200 ffff880633e74200
ffff880231674000
[ 788.611943] 0000000000000001 0000000000000003 0000000000000000
ffff880437c03710
[ 788.647241] Call Trace:
[ 788.658817] <IRQ> [<ffffffff816351f1>] dump_stack+0x19/0x1b
[ 788.686193] [<ffffffff8107b200>] warn_slowpath_common+0x70/0xb0
[ 788.713803] [<ffffffff8107b29c>] warn_slowpath_fmt+0x5c/0x80
[ 788.741314] [<ffffffff812f92f3>] ? ___ratelimit+0x93/0x100
[ 788.767018] [<ffffffff81637f49>] skb_warn_bad_offload+0xcd/0xda
[ 788.796117] [<ffffffff8152950c>] skb_checksum_help+0x17c/0x190
[ 788.823392] [<ffffffffa01463a1>] netem_enqueue+0x741/0x7c0 [sch_netem]
[ 788.854487] [<ffffffff8152cb58>] dev_queue_xmit+0x2a8/0x570
[ 788.880870] [<ffffffff8156ae1d>] ip_finish_output+0x53d/0x7d0
...
The problem occurs because netem is not prepared to handle GSO packets (as it
uses skb_checksum_help in its enqueue path, which cannot manipulate these
frames).
The solution I think is to simply segment the skb in a simmilar fashion to the
way we do in __dev_queue_xmit (via validate_xmit_skb), with some minor changes.
When we decide to corrupt an skb, if the frame is GSO, we segment it, corrupt
the first segment, and enqueue the remaining ones.
tested successfully by myself on the latest net kernel, to which this applies
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
CC: Jamal Hadi Salim <jhs@mojatatu.com>
CC: "David S. Miller" <davem@davemloft.net>
CC: netem@lists.linux-foundation.org
CC: eric.dumazet@gmail.com
CC: stephen@networkplumber.org
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When the bottom qdisc decides to, for example, drop some packet,
it calls qdisc_tree_decrease_qlen() to update the queue length
for all its ancestors, we need to update the backlog too to
keep the stats on root qdisc accurate.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove nearly duplicated code and prepare for the following patch.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Only left enqueue_root() user is netem, and it looks not necessary :
qdisc_skb_cb(skb)->pkt_len is preserved after one skb_clone()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Sheng Lan