I found a serious performance bug in packet schedulers using hrtimers.
sch_htb and sch_fq are definitely impacted by this problem.
We constantly rearm high resolution timers if some packets are throttled
in one (or more) class, and other packets are flying through qdisc on
another (non throttled) class.
hrtimer_start() does not have the mod_timer() trick of doing nothing if
expires value does not change :
if (timer_pending(timer) &&
timer->expires == expires)
return 1;
This issue is particularly visible when multiple cpus can queue/dequeue
packets on the same qdisc, as hrtimer code has to lock a remote base.
I used following fix :
1) Change htb to use qdisc_watchdog_schedule_ns() instead of open-coding
it.
2) Cache watchdog prior expiration. hrtimer might provide this, but I
prefer to not rely on some hrtimer internal.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For classifiers getting invoked via tc_classify(), we always need an
extra function call into tc_classify_compat(), as both are being
exported as symbols and tc_classify() itself doesn't do much except
handling of reclassifications when tp->classify() returned with
TC_ACT_RECLASSIFY.
CBQ and ATM are the only qdiscs that directly call into tc_classify_compat(),
all others use tc_classify(). When tc actions are being configured
out in the kernel, tc_classify() effectively does nothing besides
delegating.
We could spare this layer and consolidate both functions. pktgen on
single CPU constantly pushing skbs directly into the netif_receive_skb()
path with a dummy classifier on ingress qdisc attached, improves
slightly from 22.3Mpps to 23.1Mpps.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The same macros are used for rx as well. So rename it.
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: David S. Miller <davem@davemloft.net>
tc code implicitly considers skb->protocol even in case of accelerated
vlan paths and expects vlan protocol type here. However, on rx path,
if the vlan header was already stripped, skb->protocol contains value
of next header. Similar situation is on tx path.
So for skbs that use skb->vlan_tci for tagging, use skb->vlan_proto instead.
Reported-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Standard qdisc API to setup a timer implies an atomic operation on every
packet dequeue : qdisc_unthrottled()
It turns out this is not really needed for FQ, as FQ has no concept of
global qdisc throttling, being a qdisc handling many different flows,
some of them can be throttled, while others are not.
Fix is straightforward : add a 'bool throttle' to
qdisc_watchdog_schedule_ns(), and remove calls to qdisc_unthrottled()
in sch_fq.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Validation of skb can be pretty expensive :
GSO segmentation and/or checksum computations.
We can do this without holding qdisc lock, so that other cpus
can queue additional packets.
Trick is that requeued packets were already validated, so we carry
a boolean so that sch_direct_xmit() can validate a fresh skb list,
or directly use an old one.
Tested on 40Gb NIC (8 TX queues) and 200 concurrent flows, 48 threads
host.
Turning TSO on or off had no effect on throughput, only few more cpu
cycles. Lock contention on qdisc lock disappeared.
Same if disabling TX checksum offload.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ktime_get_ns() replaces ktime_to_ns(ktime_get())
ktime_get_real_ns() replaces ktime_to_ns(ktime_get_real())
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The DRR scheduler requires that items on the active list are work
conserving, i.e. do not hold on to skbs for throttling purposes, etc.
Attaching e.g. tbf renders DRR useless because all other classes on the
active list are delayed as well.
So, warn users that this configuration won't work as expected; we
already do this in couple of other qdiscs, see e.g.
commit b00355db3f
('pkt_sched: sch_hfsc: sch_htb: Add non-work-conserving warning handler')
The 'const' change is needed to avoid compiler warning ("discards 'const'
qualifier from pointer target type").
tested with:
drr_hier() {
parent=$1
classes=$2
for i in $(seq 1 $classes); do
classid=$parent$(printf %x $i)
tc class add dev eth0 parent $parent classid $classid drr
tc qdisc add dev eth0 parent $classid tbf rate 64kbit burst 256kbit limit 64kbit
done
}
tc qdisc add dev eth0 root handle 1: drr
drr_hier 1: 32
tc filter add dev eth0 protocol all pref 1 parent 1: handle 1 flow hash keys dst perturb 1 divisor 32
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 6da7c8fcbc ("qdisc: allow setting default queuing discipline")
added the ability to change default qdisc from pfifo_fast to say fq
But as most modern ethernet devices are multiqueue, we cant really
see all the statistics from "tc -s qdisc show", as the default root
qdisc is mq.
This patch adds the calls to qdisc_list_add() to mq and mqprio
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
By default, the pfifo_fast queue discipline has been used by default
for all devices. But we have better choices now.
This patch allow setting the default queueing discipline with sysctl.
This allows easy use of better queueing disciplines on all devices
without having to use tc qdisc scripts. It is intended to allow
an easy path for distributions to make fq_codel or sfq the default
qdisc.
This patch also makes pfifo_fast more of a first class qdisc, since
it is now possible to manually override the default and explicitly
use pfifo_fast. The behavior for systems who do not use the sysctl
is unchanged, they still get pfifo_fast
Also removes leftover random # in sysctl net core.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are a mix of function prototypes with and without extern
in the kernel sources. Standardize on not using extern for
function prototypes.
Function prototypes don't need to be written with extern.
extern is assumed by the compiler. Its use is as unnecessary as
using auto to declare automatic/local variables in a block.
Reflow modified prototypes to 80 columns.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tbf will need to schedule watchdog in ns. No need to convert it twice.
Signed-off-by: Jiri Pirko <jiri@resnulli.us>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use of "unsigned int" is preferred to bare "unsigned" in net tree.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Define three helpers to manipulate QDISC_STATE_RUNNIG flag, that a
second patch will move on another location.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One of my test machine got a deadlock during "tc" sessions,
adding/deleting classes & filters, using traffic estimators.
After some analysis, I believe we have a potential use after free case
in est_timer() :
spin_lock(e->stats_lock); << HERE >>
read_lock(&est_lock);
if (e->bstats == NULL) << TEST >>
goto skip;
Test is done a bit late, because after estimator is killed, and before
rcu grace period elapsed, we might already have freed/reuse memory where
e->stats_locks points to (some qdisc->q.lock)
A possible fix is to respect a rcu grace period at Qdisc dismantle time.
On 64bit, sizeof(struct Qdisc) is exactly 192 bytes. Adding 16 bytes to
it (for struct rcu_head) is a problem because it might change
performance, given QDISC_ALIGNTO is 32 bytes.
This is why I also change QDISC_ALIGNTO to 64 bytes, to satisfy most
current alignment requirements.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds an additional queuing strategy, called pfifo_head_drop,
to remove the oldest skb in the case of an overflow within the queue -
the head element - instead of the last skb (tail). To remove the oldest
skb in congested situations is useful for sensor network environments
where newer packets reflect the superior information.
Reviewed-by: Florian Westphal <fw@strlen.de>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This cleanup patch puts struct/union/enum opening braces,
in first line to ease grep games.
struct something
{
becomes :
struct something {
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
dev_queue_xmit enqueue's a skb and calls qdisc_run which
dequeue's the skb and xmits it. In most cases, the skb that
is enqueue'd is the same one that is dequeue'd (unless the
queue gets stopped or multiple cpu's write to the same queue
and ends in a race with qdisc_run). For default qdiscs, we
can remove the redundant enqueue/dequeue and simply xmit the
skb since the default qdisc is work-conserving.
The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the
default fast queue. The controversial part of the patch is
incrementing qlen when a skb is requeued - this is to avoid
checks like the second line below:
+ } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
>> !q->gso_skb &&
+ !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) {
Results of a 2 hour testing for multiple netperf sessions (1,
2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are
aggregate Mb/s across iterations tested with this version on
System-X boxes with Chelsio 10gbps cards:
----------------------------------
Size | ORG BW NEW BW |
----------------------------------
128K | 156964 159381 |
256K | 158650 162042 |
----------------------------------
Changes from ver1:
1. Move sch_direct_xmit declaration from sch_generic.h to
pkt_sched.h
2. Update qdisc basic statistics for direct xmit path.
3. Set qlen to zero in qdisc_reset.
4. Changed some function names to more meaningful ones.
Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Let's use TICKS instead of US, so PSCHED_TICKS2NS and PSCHED_NS2TICKS
(like in PSCHED_TICKS_PER_SEC already) to avoid misleading.
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Change PSCHED_SHIFT from 10 to 6 to increase schedulers time
resolution. This will increase 16x a number of (internal) ticks per
nanosecond, and is needed to improve accuracy of schedulers based on
rate tables, like HTB, TBF or CBQ, with rates above 100Mbit. It is
assumed this change is safe for 32bit accounting of time diffs up
to 2 minutes, which should be enough for common use (extremely low
rate values may overflow, so get inaccurate instead). To make full
use of this change an updated iproute2 will be needed. (But using
older iproute2 should be safe too.)
This change breaks ticks - microseconds similarity, so some minor code
fixes might be needed. It is also planned to change naming adequately
eg. to PSCHED_TICKS2NS() etc. in the near future.
Reported-by: Antonio Almeida <vexwek@gmail.com>
Tested-by: Antonio Almeida <vexwek@gmail.com>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use PSCHED_SHIFT constant instead of '10' in PSCHED_US2NS() and
PSCHED_NS2US() macros to enable changing this value later.
Additionally use PSCHED_SHIFT in sch_hfsc SM_SHIFT and ISM_SHIFT
definitions. This part of the patch is based on feedback from
Patrick McHardy <kaber@trash.net>.
Reported-by: Antonio Almeida <vexwek@gmail.com>
Tested-by: Antonio Almeida <vexwek@gmail.com>
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Patrick McHardy <kaber@trash.net> suggested:
> How about making this flag and the warning message (in a out-of-line
> function) globally available? Other qdiscs (f.i. HFSC) can't deal with
> inner non-work-conserving qdiscs as well.
This patch uses qdisc->flags field of "suspected" child qdisc.
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current check wrongly uses the state of one (currently the first)
tx queue for all tx queues in case of non-default qdiscs. This check
mainly prevented requeuing loop with __netif_schedule(), but now it's
controlled inside __qdisc_run(), while dequeuing. The wrongness of
this check was first noticed by Herbert Xu.
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since some qdiscs call qdisc_tree_decrease_qlen() (so qdisc_lookup())
without rtnl_lock(), adding and deleting from a qdisc list needs
additional locking. This patch adds global spinlock qdisc_list_lock
and wrapper functions for modifying the list. It is considered as a
temporary solution until hfsc_dequeue(), netem_dequeue() and
tbf_dequeue() (or qdisc_tree_decrease_qlen()) are redone.
With feedback from Herbert Xu and David S. Miller.
Signed-off-by: Jarek Poplawski <jarkao2@gmail.com>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Based upon a bug report by Andrew Gallatin on netdev
with subject "CPU utilization increased in 2.6.27rc"
In commit 37437bb2e1
("pkt_sched: Schedule qdiscs instead of netdev_queue.")
the test of the queue being stopped was erroneously
removed from qdisc_run().
When the TX queue of the device fills up, this omission
causes lots of extraneous useless work to be queued up
to softirq context, where we'll just return immediately
because the device is still stuffed up.
Signed-off-by: David S. Miller <davem@davemloft.net>
Add size table functions for qdiscs and calculate packet size in
qdisc_enqueue().
Based on patch by Patrick McHardy
http://marc.info/?l=linux-netdev&m=115201979221729&w=2
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we have shared qdiscs, packets come out of the qdiscs
for multiple transmit queues.
Therefore it doesn't make any sense to schedule the transmit
queue when logically we cannot know ahead of time the TX
queue of the SKB that the qdisc->dequeue() will give us.
Just for sanity I added a BUG check to make sure we never
get into a state where the noop_qdisc is scheduled.
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently it is associated with a netdev_queue, but when we have
qdisc sharing that no longer makes any sense.
Signed-off-by: David S. Miller <davem@davemloft.net>
This effectively "flips the switch" by making the core networking
and multiqueue-aware drivers use the new TX multiqueue structures.
Non-multiqueue drivers need no changes. The interfaces they use such
as netif_stop_queue() degenerate into an operation on TX queue zero.
So everything "just works" for them.
Code that really wants to do "X" to all TX queues now invokes a
routine that does so, such as netif_tx_wake_all_queues(),
netif_tx_stop_all_queues(), etc.
pktgen and netpoll required a little bit more surgery than the others.
In particular the pktgen changes, whilst functional, could be largely
improved. The initial check in pktgen_xmit() will sometimes check the
wrong queue, which is mostly harmless. The thing to do is probably to
invoke fill_packet() earlier.
The bulk of the netpoll changes is to make the code operate solely on
the TX queue indicated by by the SKB queue mapping.
Setting of the SKB queue mapping is entirely confined inside of
net/core/dev.c:dev_pick_tx(). If we end up needing any kind of
special semantics (drops, for example) it will be implemented here.
Finally, we now have a "real_num_tx_queues" which is where the driver
indicates how many TX queues are actually active.
With IGB changes from Jeff Kirsher.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Convert packet schedulers to use the netlink API. Unfortunately a gradual
conversion is not possible without breaking compilation in the middle or
adding lots of casts, so this patch converts them all in one step. The
patch has been mostly generated automatically with some minor edits to
at least allow seperate conversion of classifiers and actions.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since hardware header operations are part of the protocol class
not the device instance, make them into a separate object and
save memory.
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The behaviour of NET_CLS_POLICE for TC_POLICE_RECLASSIFY was to return
it to the qdisc, which could handle it internally or ignore it. With
NET_CLS_ACT however, tc_classify starts over at the first classifier
and never returns it to the qdisc. This makes it impossible to support
qdisc-internal reclassification, which in turn makes it impossible to
remove the old NET_CLS_POLICE code without breaking compatibility since
we have two qdiscs (CBQ and ATM) that support this.
This patch adds a tc_classify_compat function that handles
reclassification the old way and changes CBQ and ATM to use it.
This again is of course not fully backwards compatible with the previous
NET_CLS_ACT behaviour. Unfortunately there is no way to fully maintain
compatibility *and* support qdisc internal reclassification with
NET_CLS_ACT, but this seems like the better choice over keeping the two
incompatible options around forever.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since we're now holding the rtnl during the entire dump operation, we
can remove qdisc_tree_lock, whose only purpose is to protect dump
callbacks from concurrent changes to the qdisc tree.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that all packet schedulers have been converted to hrtimers most users
of PSCHED_JIFFIE2US and PSCHED_US2JIFFIE are gone. The remaining users use
it to convert external time units to packet scheduler clock ticks, so use
PSCHED_TICKS_PER_SEC instead.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Get rid of the manual clock source selection mess and use ktime. Also
use a scalar representation, which allows to clean up pkt_sched.h a bit
more and results in less ktime_to_ns() calls in most cases.
The PSCHED_US2JIFFIE/PSCHED_JIFFIE2US macros are implemented quite
inefficient by this patch, following patches will convert all qdiscs
to hrtimers and get rid of them entirely.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
In PSCHED_TADD and PSCHED_TADD2, if delta is less than tv.tv_usec (so,
less than USEC_PER_SEC too) then tv_res will be smaller than tv. The
affectation "(tv_res).tv_usec = __delta;" is wrong. The fix is to
revert to the original code before
4ee303dfea and change the 'if' in
'while'.
[Shuya MAEDA: "while (__delta >= USEC_PER_SEC){ ... }" instead of
"while (__delta > USEC_PER_SEC){ ... }"]
Signed-off-by: Guillaume Chazarain <guichaz@yahoo.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
