make qdisc_restart more readable
Subsystems (3)
RFC 1/1 make qdisc_restart more readable
This compiles and passes some basic tests - no serious testing. Against net-2.6.
The patch is ugly looking, so i have at the end the re-written qdisc; you can easily tell the rest from the patch.
Please flush out any fluff - I would like to submit this (almost lost it, thanks to an early posting, found it).
cheers,
jamal
include/linux/netdevice.h +3 -0
@@ -83,6 +83,9 @@ struct wireless_dev;#define NETDEV_TX_OK 0/* driver took care of packet */ #define NETDEV_TX_BUSY 1/* driver tx path was busy*/ #define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */ +#define NETDEV_TX_QUEUE -3/* request caller to requeue packet */ +/* * Compute the worst case header length according to the protocols
net/sched/sch_generic.c +114 -93
235 lines not on this page · raw message
intqdisc_restart(struct net_device *dev)
{
struct Qdisc *q = dev->qdisc;
unsigned lockless = (dev->features & NETIF_F_LLTX);
struct sk_buff *skb = NULL;
int ret;
ret = try_get_tx_pkt(&skb, dev, q);
if (ret == 0)
return qqlen(q);
/* we have a packet to send */
if (!lockless) {
if (!netif_tx_trylock(dev))
return handle_tx_locked(skb, dev, q);
}
/* all clear .. */
spin_unlock(&dev->queue_lock);
/* churn baby churn .. */
ret = dev_hard_start_xmit(skb, dev);
if (!lockless)
netif_tx_unlock(dev);
spin_lock(&dev->queue_lock);
q = dev->qdisc;
/* most likely result, packet went ok */
if (ret == NETDEV_TX_OK)
return qqlen(q);
/* only for lockless drivers .. */
if (ret == NETDEV_TX_LOCKED && lockless) {
return handle_tx_locked(skb, dev, q);
}
if (unlikely (ret != NETDEV_TX_BUSY)) {
/* XXX: Do we need a ratelimit? or put a
* BUG_ON((int) ret != NETDEV_TX_BUSY) ?
**/
printk("BUG %s code %d qlen %d\n",dev->name, ret, q->q.qlen);
}
return handle_dev_requeue(skb, dev, q);
}Quoted 6 lines
This compiles and passes some basic tests - no serious testing. Against net-2.6.
The patch is ugly looking, so i have at the end the re-written qdisc; you can easily tell the rest from the patch.
Quoted 6 lines
Please flush out any fluff - I would like to submit this (almost lost it, thanks to an early posting, found it).
cheers,
jamal
In qdisc_restart(), you removed any check for if (!netif_queue_stopped(dev)) before calling dev_hard_start_xmit(). If the underlying queue is stopped and you send the skb, you'll generate a requeue. Is there a reason it was removed?
Thx,
-PJ Waskiewicz
Quoted 5 lines · Peter P
In
qdisc_restart(), you removed any check for if (!netif_queue_stopped(dev)) before callingdev_hard_start_xmit(). If the underlying queue is stopped and you send the skb, you'll generate a requeue. Is there a reason it was removed?
No - its a bug ;->
Thanks for staring at this Peter. I will fix it.
cheers,
jamal
Quoted 7 lines · Peter P
Quoted 3 lines
In
qdisc_restart(), you removed any check for if
(!netif_queue_stopped(dev)) before calling
dev_hard_start_xmit(). IfQuoted 1 line
the underlying queue is stopped and you send the skb,
you'll generate
Quoted 8 lines
Quoted 2 lines
a requeue. Is there a reason it was removed?
No - its a bug ;->
Thanks for staring at this Peter. I will fix it.cheers,
jamal
After thinking about this a bit more: even if the queue is stopped, you'd end up requeueing anyways. Plus, you'd need to re-acquire dev->queue_lock (which is what happens today). I think the best way overall would be to check the queue state before you physically dequeue (whether it's in qdisc_restart() or in the qdisc's ->dequeue()). That way you still hold dev->queue_lock in case the queue is stopped, and haven't yet pulled an skb causing a requeue event.
Just $0.02 on Friday.
Thx,
-PJ Waskiewicz
Quoted 5 lines · Peter P
After thinking about this a bit more: even if the queue is stopped, you'd end up requeueing anyways. Plus, you'd need to re-acquire
dev->queue_lock(which is what happens today). I think the best way overall would be to check the queue state before you physically dequeue (whether it's inqdisc_restart()or in the qdisc's ->dequeue()).
For lockless drivers, I think it could be moved up; only thing is nothing is stopping it from changing again at transmit time. For drivers that are not lockless, i am wondering if its even useful to have that check given we have already grabbed the tx lock.
Quoted 3 lines
That
way you still holddev->queue_lockin case the queue is stopped, and haven't yet pulled an skb causing a requeue event.
It certainly sounds like a useful optimization. Defer that thought - for now i just want to make sure its doing what the previous code did.
I will repost the fixed patch later today after some testing; if you have time i will appreciate some eyeballing. That piece of code is hairy. Hopefully we can make it bald.
cheers,
jamal
Quoted 4 lines
For lockless drivers, I think it could be moved up; only thing is nothing is stopping it from changing again at transmit time. For drivers that are not lockless, i am wondering if its even useful to have that check given we have already grabbed the tx lock.
Good point. I live in an LLTX world, so I don't think tx_lock mode very often.
Quoted 9 lines
Quoted 3 lines
That
way you still holddev->queue_lockin case the queue isstopped, and
Quoted 1 line
haven't yet pulled an skb causing a requeue event.
It certainly sounds like a useful optimization. Defer that thought - for now i just want to make sure its doing what the previous code did.
Agreed.
Quoted 4 lines
I will repost the fixed patch later today after some testing; if you have time i will appreciate some eyeballing. That piece of code is hairy. Hopefully we can make it bald.
Clippers are standing by. :)
-PJ
take 2 WAS (RE: [RFC] make qdisc_restart more readable
Quoted 1 line · Peter P
Clippers are standing by. :)
Nothing as nice as time off to code away;->
Ok, booting fine to me;
if you receive this email, it must have worked. OTOH, if you didnt receive this it probably failed ;->
(as the presenter said "if you cant hear me in the far corner please raise your hand").
clip away ...
cheers,
jamal
include/linux/netdevice.h +3 -0
@@ -83,6 +83,9 @@ struct wireless_dev;#define NETDEV_TX_OK 0/* driver took care of packet */ #define NETDEV_TX_BUSY 1/* driver tx path was busy*/ #define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */ +#define NETDEV_TX_QUEUE -3/* request caller to requeue packet */ +/* * Compute the worst case header length according to the protocols
net/sched/sch_generic.c +121 -95
242 lines not on this page · raw message
RE: take 2 WAS (RE: [RFC] make qdisc_restart more readable
Quoted 6 lines
Ok, booting fine to me;
if you receive this email, it must have worked. OTOH, if you didnt receive this it probably failed ;-> (as the presenter said "if you cant hear me in the far corner please raise your hand").clip away ...
+static inline int
+handle_dev_requeue(struct sk_buff *skb, struct net_device *dev, struct Qdisc *q)
+{
+
+ if (unlikely(q == &noop_qdisc))
+ kfree_skb(skb);
+ else if (skb->next)
+ dev->gso_skb = skb;
+ else
+ q->ops->requeue(skb, q);
+ /* XXX: Could netif_schedule fail? Or is that fact we are
+ * requeueing imply the hardware path is closed
+ * and even if we fail, some interupt will wake us
+ */
+ netif_schedule(dev);
+ return 0;
+}
The current requeue calls in qdisc_restart() returns a value of 1 to the upper layer. If you return 0, the __qdisc_run() won't break from the while() it seems if you have to successfully requeue a packet.
-static inline int qdisc_restart(struct net_device *dev) +int
+qdisc_restart(struct net_device *dev)
{
Why is this no longer defined static?
+ /* we need to refresh q because it may be invalid since we
droped
<nitpick>
droped = dropped
</nitpick>
That's all I have for now. I'll try applying it for kicks and see what happens.
Thanks,
-PJ Waskiewicz
RE: take 2 WAS (RE: [RFC] make qdisc_restart more readable
Quoted 2 lines · Peter P
The current requeue calls in
qdisc_restart()returns a value of 1 to the upper layer.
Which was not meaningful ;->
i.e qdisc run would break as a result.
Note, this is still in conformance with net-2.6 based on a change in the last few days made by Herbert - I was following his lead.
I put that comment there just as something to remember and maybe revisit later.
Quoted 3 lines
If you return 0, the
__qdisc_run()won't break from thewhile()it seems if you have to successfully requeue a packet.
It wont; the check is for < 0
Quoted 8 lines
net/sched/sch_generic.c +2
-1-static inline int qdisc_restart (struct net_device *dev) +int +qdisc_restart (struct net_device *dev) {Why is this no longer defined static?
Good catch - i will restore it to what it was originally.
Quoted 8 lines
(quoted diff) +1
-0+ /* we need to refresh q because it may be invalid since wedroped
<nitpick>
droped = dropped
</nitpick>
Ok, better to fix it now before the spelling police show up ;->
Quoted 2 lines
That's all I have for now. I'll try applying it for kicks and see what happens.
Just make those two fixes in your version; also note, you need Davems net-2.6 tree.
cheers,
jamal
Re: take 2 WAS (RE: [RFC] make qdisc_restart more readable
In reply to jamal* jamal <[email protected]> 2007-05-11 15:29
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 107
Quoted 11 lines
net/sched/sch_generic.c +10
-0+netif_schedule (dev); +return 0 ; +} + +static struct sk_buff * +try_get_tx_pkt (struct net_device *dev,struct Qdisc *q) +{ +struct sk_buff *skb =NULL ; + +if (skb = dev->gso_skb)
This must have caused a warning and you probably want to inline this function.
Re: take 2 WAS (RE: [RFC] make qdisc_restart more readable
Quoted 5 lines · Thomas Graf
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 107
Quoted 3 lines
net/sched/sch_generic.c +3
-0+struct sk_buff *skb =NULL ; + +if (skb = dev->gso_skb)This must have caused a warning
Possibly (dont recall) - i will move the assignment to the first line.
Quoted 1 line
and you probably want to inline this function.
will do.
Thanks Thomas.
cheers,
jamal
take 3 [RFC] make qdisc_restart more readable
In reply to jamalOk, heres the next iteration based on feedback from Thomas and Peter. It seems this is good form in matching what the original code does. I have a few doubts in labelled as XXX in there. Thomas, if you have time, do you mind looking at those?
The usual testing happened i.e "if you can see this email, it must have worked";->
cheers,
jamal
include/linux/netdevice.h +3 -0
@@ -83,6 +83,9 @@ struct wireless_dev;#define NETDEV_TX_OK 0/* driver took care of packet */ #define NETDEV_TX_BUSY 1/* driver tx path was busy*/ #define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */ +#define NETDEV_TX_QUEUE -3/* request caller to requeue packet */ +/* * Compute the worst case header length according to the protocols
net/sched/sch_generic.c +117 -93
238 lines not on this page · raw message
Re: take 3 [RFC] make qdisc_restart more readable
include/linux/netdevice.h struct wireless_dev; · line 86
Quoted 9 lines · jamal
include/linux/netdevice.h +1
-0@@ -83,6 +83,9 @@ struct wireless_dev;#define NETDEV_TX_OK 0/* driver took care of packet */ #define NETDEV_TX_BUSY 1/* driver tx path was busy*/ #define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */
This shouldn't be a NETDEV_TX code since its only handled correctly internally for handle_dev_cpu_collision().
include/linux/netdevice.h struct wireless_dev; · line 87
Quoted 1 line
include/linux/netdevice.h +1
-0+#define NETDEV_TX_QUEUE -3/* request caller to requeue packet */
How will this be used? The driver can simply stop the queue if it doesn't want to receive more packets.
include/linux/netdevice.h struct wireless_dev; · line 90
Quoted 3 lines
include/linux/netdevice.h +3
-0+static inline int handle_dev_cpu_collision (struct net_device *dev) +{ +if (dev->xmit_lock_owner ==smp_processor_id ()) {
unlikely would make sense here.
include/linux/netdevice.h struct wireless_dev;
Quoted 6 lines
include/linux/netdevice.h +6
-0+static inline int +handle_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (q == &noop_qdisc)) +kfree_skb (skb);
Why is this special-casing needed? __qdisc_run already checks for noop_qdisc and noop has a proper requeue function.
include/linux/netdevice.h
Quoted 2 lines
include/linux/netdevice.h +2
-0+else if (skb->next) + dev->gso_skb = skb;
The gso_skb cases could probably also be marked unlikely.
Quoted 15 lines
include/linux/netdevice.h +15
-0+else + q->ops->requeue (skb, q); +/* XXX: Could netif_schedule fail? Or is that fact we are +* requeueing imply the hardware path is closed +* and even if we fail, some interupt will wake us +*/ +netif_schedule (dev); +return 0 ; +} + +static inline struct sk_buff * +try_get_tx_pkt (struct net_device *dev,struct Qdisc *q) +{ +struct sk_buff *skb = dev->gso_skb; +
include/linux/netdevice.h
Quoted 10 lines
include/linux/netdevice.h +10
-0+if (skb) + dev->gso_skb =NULL ; +else + skb = q->dequeue (q); + +return skb; +} + +static inline int +handle_tx_locked (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q)
handle_.* sounds a bit like "I couldn't think of a better name" :) How about dev_tx_locked()?
Quoted 6 lines
include/linux/netdevice.h +6
-0+{ +int ret = (dev); + +if (ret ==NETDEV_TX_DROP ) { +kfree_skb (skb); +return qdisc_qlen (q);
include/linux/netdevice.h
Quoted 10 lines
include/linux/netdevice.h +9
-0+ } + +return handle_dev_requeue (skb, dev, q); +} +if (unlikely (ret !=NETDEV_TX_BUSY )) { +/* XXX: Do we need a ratelimit? or put a +* BUG_ON((int) ret != NETDEV_TX_BUSY) ? +**/ +printk ("BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen);
BUG_ON sounds a bit extreme, net_ratelimit() makes sense.
Re: take 3 [RFC] make qdisc_restart more readable
Quoted 6 lines · Patrick McHardy
include/linux/netdevice.h struct wireless_dev; · line 86
Quoted 4 lines · jamal
include/linux/netdevice.h +1
-0#define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */
Quoted 8 lines
This shouldn't be a
NETDEV_TXcode since its only handled correctly internally forhandle_dev_cpu_collision().Quoted 1 line
+#define
NETDEV_TX_QUEUE-3 /* request caller to requeue packet */How will this be used? The driver can simply stop the queue if it doesn't want to receive more packets.
I can make those part of a different namespace;
put defines in sch_generic sound reasonable?
Quoted 6 lines
include/linux/netdevice.h
Quoted 3 lines
include/linux/netdevice.h +3
-0+static inline int handle_dev_cpu_collision (struct net_device *dev) +{ +if (dev->xmit_lock_owner ==smp_processor_id ()) {unlikely would make sense here.
sure.
Quoted 10 lines
include/linux/netdevice.h
Quoted 6 lines
include/linux/netdevice.h +6
-0+static inline int +handle_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (q == &noop_qdisc)) +kfree_skb (skb);Why is this special-casing needed?
__qdisc_runalready checks fornoop_qdiscand noop has a proper requeue function.
Good catch: That was a patch from Herbert after Thomas' "q refresh" patch. So i was just replicating. I will fix it. Note: I think that change may have made it to -stable?
Quoted 5 lines
include/linux/netdevice.h
Quoted 2 lines
include/linux/netdevice.h +2
-0+else if (skb->next) + dev->gso_skb = skb;The
gso_skbcases could probably also be marked unlikely.
will do.
Quoted 7 lines
include/linux/netdevice.h
Quoted 3 lines
include/linux/netdevice.h +3
-0+ +static inline int +handle_tx_locked (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q)handle_.* sounds a bit like "I couldn't think of a better name" :) How about
dev_tx_locked()?
dev_tx_islocked()?
Quoted 7 lines
include/linux/netdevice.h
Quoted 5 lines
include/linux/netdevice.h +5
-0+if (unlikely (ret !=NETDEV_TX_BUSY )) { +/* XXX: Do we need a ratelimit? or put a +* BUG_ON((int) ret != NETDEV_TX_BUSY) ? +**/ +printk ("BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen);
BUG_ONsounds a bit extreme,net_ratelimit()makes sense.
Ok.
Thanks a lot Patrick.
cheers,
jamal
PATCH 1/1 make qdisc_restart more readable
Ok, I am upgrading this to last call after taking in Patricks feedback. Speak now - or send patches against it later.
Dave, just let it simmer down for a day or two, then if no complaints, go ahead and apply it. It is against net-2.6
Many thanks to Thomas, Peter and Patrick for their reviews.
cheers,
jamal
[NET_SCHED]: Cleanup readability of qdisc restart
Over the years this code has gotten hairier. Resulting in many long discussions and patches that get it wrong.
This patch helps tame that code so normal people will understand it.
---
commit d5088c3eb28211ac8c4137d6e2e69edf6061a57d
tree 7c683e412072a992ba49bccc34f3e1ab3c42efa3
parent 15cfc8c4efe964593b0246df53e3aa30502d9550
author Jamal Hadi Salim <[email protected]> Sun, 13 May 2007 10:19:52 -0400 committer Jamal Hadi Salim <[email protected]> Sun, 13 May 2007 10:19:52 -0400
net/sched/sch_generic.c | 205 +++++++++++++++++++++++++---------------------- 1 files changed, 111 insertions(+), 94 deletions(-)
net/sched/sch_generic.c +111 -94
@@ -34,6 +34,9 @@#include <net/sock.h> #include <net/pkt_sched.h> +#define SCHED_TX_DROP -2 +#define SCHED_TX_QUEUE -3 +/* Main transmission queue. */ /* Modifications to data participating in scheduling must be protected with
@@ -59,7 +62,74 @@ void qdisc_unlock_tree(struct net_device *dev)spin_unlock_bh (&dev->queue_lock); } +static inline int qdisc_qlen (struct Qdisc *q) +{ +BUG_ON ((int ) q->q.qlen <0 ); +return q->q.qlen; +} + +static inline int handle_dev_cpu_collision (struct net_device *dev) +{ +if (unlikely (dev->xmit_lock_owner ==smp_processor_id ())) { +if (net_ratelimit ()) +printk (KERN_DEBUG +"Dead loop on netdevice %s, fix it urgently!\n" , + dev->name); +return SCHED_TX_DROP ; + } +__get_cpu_var (netdev_rx_stat).cpu_collision++; +return SCHED_TX_QUEUE ; +} + +static inline int +do_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (skb->next)) + dev->gso_skb = skb; +else + q->ops->requeue (skb, q); +/* XXX: Could netif_schedule fail? Or is that fact we are +* requeueing imply the hardware path is closed +* and even if we fail, some interupt will wake us +*/ +netif_schedule (dev); +return 0 ; +} + +static inline struct sk_buff * +try_get_tx_pkt (struct net_device *dev,struct Qdisc *q) +{ +struct sk_buff *skb = dev->gso_skb; + +if (skb) + dev->gso_skb =NULL ; +else + skb = q->dequeue (q); + +return skb; +} + +static inline int +tx_islocked (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ +int ret =handle_dev_cpu_collision (dev); + +if (ret ==SCHED_TX_DROP ) { +kfree_skb (skb); +return qdisc_qlen (q); + } + +return do_dev_requeue (skb, dev, q); +} + +/* +NOTE: Called under dev->queue_lock with locally disabled BH. + +__LINK_STATE_QDISC_RUNNING guarantees only one CPU +can enter this region at a time. +dev->queue_lock serializes queue accesses for this device AND dev->qdisc pointer itself.
@@ -67,114 +137,61 @@ void qdisc_unlock_tree(struct net_device *dev)dev->queue_lock and netif_tx_lock are mutually exclusive, if one is grabbed, another must be free. -*/ +Multiple CPUs may contend for the two locks. -/* Kick device. +Note, that this procedure can be called by a watchdog timer +Returns to the caller: Returns: 0 - queue is empty or throttled. >0 - queue is not empty. - -NOTE: Called under dev->queue_lock with locally disabled BH. */ static inline int qdisc_restart (struct net_device *dev) {struct Qdisc *q = dev->qdisc; +unsigned lockless = (dev->features &NETIF_F_LLTX );struct sk_buff *skb; +int ret; -/* Dequeue packet */ -if (((skb = dev->gso_skb)) || ((skb = q->dequeue (q)))) { -unsigned nolock = (dev->features &NETIF_F_LLTX ); - - dev->gso_skb =NULL ; + skb =try_get_tx_pkt (dev, q); +if (skb ==NULL ) +return qdisc_qlen (q); -/* -* When the driver has LLTX set it does its own locking -* in start_xmit. No need to add additional overhead by -* locking again. These checks are worth it because -* even uncongested locks can be quite expensive. -* The driver can do trylock like here too, in case -* of lock congestion it should return -1 and the packet -* will be requeued. -*/ -if (!nolock) { -if (!netif_tx_trylock (dev)) { - collision: -/* So, someone grabbed the driver. */ - -/* It may be transient configuration error, -when hard_start_xmit() recurses. We detect -it by checking xmit owner and drop the -packet when deadloop is detected. -*/ -if (dev->xmit_lock_owner ==smp_processor_id ()) { -kfree_skb (skb); -if (net_ratelimit ()) -printk (KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n" , dev->name); - goto out; - } -__get_cpu_var (netdev_rx_stat).cpu_collision++; - goto requeue; - } - } - - { -/* And release queue */ -spin_unlock (&dev->queue_lock); - -if (!netif_queue_stopped (dev)) { -int ret; - - ret =dev_hard_start_xmit (skb, dev); -if (ret ==NETDEV_TX_OK ) { -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto out; - } -if (ret ==NETDEV_TX_LOCKED && nolock) { -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto collision; - } - } - -/* NETDEV_TX_BUSY - we need to requeue */ -/* Release the driver */ -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - } - -/* Device kicked us out :( -This is possible in three cases: - -0. driver is locked -1. fastroute is enabled -2. device cannot determine busy state -before start of transmission (f.e. dialout) -3. device is buggy (ppp) -*/ - -requeue: -if (unlikely (q == &noop_qdisc)) -kfree_skb (skb); -else if (skb->next) - dev->gso_skb = skb; -else - q->ops->requeue (skb, q); -netif_schedule (dev); -return 0 ; +/* we have a packet to send */ +if (!lockless) { +if (!netif_tx_trylock (dev)) +return tx_islocked (skb, dev, q); } - -out: -BUG_ON ((int ) q->q.qlen <0 ); -return q->q.qlen; +/* all clear .. */ +spin_unlock (&dev->queue_lock); + + ret =NETDEV_TX_BUSY ; +if (!netif_queue_stopped (dev)) +/* churn baby churn .. */ + ret =dev_hard_start_xmit (skb, dev); + +if (!lockless) +netif_tx_unlock (dev); + +spin_lock (&dev->queue_lock); + +/* we need to refresh q because it may be invalid since +* we dropped dev->queue_lock earlier ... +* So dont try to be clever grasshopper +*/ + q = dev->qdisc; +/* most likely result, packet went ok */ +if (ret ==NETDEV_TX_OK ) +return qdisc_qlen (q); +/* only for lockless drivers .. */ +if (ret ==NETDEV_TX_LOCKED && lockless) +return tx_islocked (skb, dev, q); + +if (unlikely (ret !=NETDEV_TX_BUSY &&net_ratelimit ())) +printk (KERN_DEBUG " BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen); + +return do_dev_requeue (skb, dev, q); }void __qdisc_run (struct net_device *dev)
Quoted 5 lines · jamal
Ok, I am upgrading this to last call after taking in Patricks feedback. Speak now - or send patches against it later.
Dave, just let it simmer down for a day or two, then if no complaints, go ahead and apply it. It is against net-2.6
Many thanks to Thomas, Peter and Patrick for their reviews.
Two final suggestions, sorry for not mentioning this earlier.
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev)
Quoted 6 lines
net/sched/sch_generic.c +5
-0+printk (KERN_DEBUG +"Dead loop on netdevice %s, fix it urgently!\n" , + dev->name); +if (unlikely (ret !=NETDEV_TX_BUSY &&net_ratelimit ())) +printk (KERN_DEBUG " BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen);
For both of these KERN_WARNING or above seems to make sense so the messages don't end up in some debug log. Please also fix whitespace and line length for the second printk if you're going to change this :)
Quoted 1 line · Patrick McHardy
Two final suggestions, sorry for not mentioning this earlier.
np
Quoted 9 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev)
Quoted 6 lines
net/sched/sch_generic.c +5
-0+printk (KERN_DEBUG +"Dead loop on netdevice %s, fix it urgently!\n" , + dev->name); +if (unlikely (ret !=NETDEV_TX_BUSY &&net_ratelimit ())) +printk (KERN_DEBUG " BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen);For both of these
KERN_WARNINGor above
I will make them KERN_WARNING ..
Quoted 3 lines
seems to make sense so
the messages don't end up in some debug log. Please also fix whitespace
What whitespace issue?
Quoted 2 lines
and line length for the second printk if you're
going to change this :)
I will fix the length of the second one.
Thanks Patrick.
cheers,
jamal
Quoted 7 lines · jamal
Quoted 2 lines · Patrick McHardy
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 192
Quoted 2 lines
net/sched/sch_generic.c +2
-0+if (unlikely (ret !=NETDEV_TX_BUSY &&net_ratelimit ())) +printk (KERN_DEBUG " BUG %s code %d qlen %d\n" ,dev->name, ret, q->q.qlen);What whitespace issue?
Space between unlikely and '(', missing space after ',' and dev->name.
Quoted 5 lines
Quoted 2 lines
and line length for the second printk if you're
going to change this :)I will fix the length of the second one.
Thanks.
PATCH 1/1 make qdisc_restart more readable
In reply to jamalIncoporating last comments from Patrick.
Dave, I think this is ready to apply - against net-2.6 from this morning.
cheers,
jamal
[NET_SCHED]: Cleanup readability of qdisc restart
Over the years this code has gotten hairier. Resulting in many long discussions over long summer days and patches that get it wrong. This patch helps tame that code so normal people will understand it.
Thanks to Thomas Graf, Peter J. waskiewicz Jr, and Patrick McHardy for their valuable reviews.
---
commit 4ad20f6340ed13bdfcc5082a2f538af0323c3330
tree 63050b1effc9d30e1c124f52605ce3f543a748e9
parent d831666e98b4f1e19ebdd2349735f47bf37cd293
author Jamal Hadi Salim <[email protected]> Mon, 14 May 2007 09:28:45 -0400 committer Jamal Hadi Salim <[email protected]> Mon, 14 May 2007 09:28:45 -0400
net/sched/sch_generic.c | 206 ++++++++++++++++++++++++++--------------------- 1 files changed, 112 insertions(+), 94 deletions(-)
net/sched/sch_generic.c +112 -94
@@ -34,6 +34,9 @@#include <net/sock.h> #include <net/pkt_sched.h> +#define SCHED_TX_DROP -2 +#define SCHED_TX_QUEUE -3 +/* Main transmission queue. */ /* Modifications to data participating in scheduling must be protected with
@@ -59,7 +62,74 @@ void qdisc_unlock_tree(struct net_device *dev)spin_unlock_bh (&dev->queue_lock); } +static inline int qdisc_qlen (struct Qdisc *q) +{ +BUG_ON ((int ) q->q.qlen <0 ); +return q->q.qlen; +} + +static inline int handle_dev_cpu_collision (struct net_device *dev) +{ +if (unlikely (dev->xmit_lock_owner ==smp_processor_id ())) { +if (net_ratelimit ()) +printk (KERN_WARNING +"Dead loop on netdevice %s, fix it urgently!\n" , + dev->name); +return SCHED_TX_DROP ; + } +__get_cpu_var (netdev_rx_stat).cpu_collision++; +return SCHED_TX_QUEUE ; +} + +static inline int +do_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (skb->next)) + dev->gso_skb = skb; +else + q->ops->requeue (skb, q); +/* XXX: Could netif_schedule fail? Or is that fact we are +* requeueing imply the hardware path is closed +* and even if we fail, some interupt will wake us +*/ +netif_schedule (dev); +return 0 ; +} + +static inline struct sk_buff * +try_get_tx_pkt (struct net_device *dev,struct Qdisc *q) +{ +struct sk_buff *skb = dev->gso_skb; + +if (skb) + dev->gso_skb =NULL ; +else + skb = q->dequeue (q); + +return skb; +} + +static inline int +tx_islocked (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ +int ret =handle_dev_cpu_collision (dev); + +if (ret ==SCHED_TX_DROP ) { +kfree_skb (skb); +return qdisc_qlen (q); + } + +return do_dev_requeue (skb, dev, q); +} + +/* +NOTE: Called under dev->queue_lock with locally disabled BH. + +__LINK_STATE_QDISC_RUNNING guarantees only one CPU +can enter this region at a time. +dev->queue_lock serializes queue accesses for this device AND dev->qdisc pointer itself.
@@ -67,114 +137,62 @@ void qdisc_unlock_tree(struct net_device *dev)dev->queue_lock and netif_tx_lock are mutually exclusive, if one is grabbed, another must be free. -*/ +Multiple CPUs may contend for the two locks. -/* Kick device. +Note, that this procedure can be called by a watchdog timer +Returns to the caller: Returns: 0 - queue is empty or throttled. >0 - queue is not empty. - -NOTE: Called under dev->queue_lock with locally disabled BH. */ static inline int qdisc_restart (struct net_device *dev) {struct Qdisc *q = dev->qdisc; +unsigned lockless = (dev->features &NETIF_F_LLTX );struct sk_buff *skb; +int ret; -/* Dequeue packet */ -if (((skb = dev->gso_skb)) || ((skb = q->dequeue (q)))) { -unsigned nolock = (dev->features &NETIF_F_LLTX ); - - dev->gso_skb =NULL ; + skb =try_get_tx_pkt (dev, q); +if (skb ==NULL ) +return qdisc_qlen (q); -/* -* When the driver has LLTX set it does its own locking -* in start_xmit. No need to add additional overhead by -* locking again. These checks are worth it because -* even uncongested locks can be quite expensive. -* The driver can do trylock like here too, in case -* of lock congestion it should return -1 and the packet -* will be requeued. -*/ -if (!nolock) { -if (!netif_tx_trylock (dev)) { - collision: -/* So, someone grabbed the driver. */ - -/* It may be transient configuration error, -when hard_start_xmit() recurses. We detect -it by checking xmit owner and drop the -packet when deadloop is detected. -*/ -if (dev->xmit_lock_owner ==smp_processor_id ()) { -kfree_skb (skb); -if (net_ratelimit ()) -printk (KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n" , dev->name); - goto out; - } -__get_cpu_var (netdev_rx_stat).cpu_collision++; - goto requeue; - } - } - - { -/* And release queue */ -spin_unlock (&dev->queue_lock); - -if (!netif_queue_stopped (dev)) { -int ret; - - ret =dev_hard_start_xmit (skb, dev); -if (ret ==NETDEV_TX_OK ) { -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto out; - } -if (ret ==NETDEV_TX_LOCKED && nolock) { -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto collision; - } - } - -/* NETDEV_TX_BUSY - we need to requeue */ -/* Release the driver */ -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - } - -/* Device kicked us out :( -This is possible in three cases: - -0. driver is locked -1. fastroute is enabled -2. device cannot determine busy state -before start of transmission (f.e. dialout) -3. device is buggy (ppp) -*/ - -requeue: -if (unlikely (q == &noop_qdisc)) -kfree_skb (skb); -else if (skb->next) - dev->gso_skb = skb; -else - q->ops->requeue (skb, q); -netif_schedule (dev); -return 0 ; +/* we have a packet to send */ +if (!lockless) { +if (!netif_tx_trylock (dev)) +return tx_islocked (skb, dev, q); } - -out: -BUG_ON ((int ) q->q.qlen <0 ); -return q->q.qlen; +/* all clear .. */ +spin_unlock (&dev->queue_lock); + + ret =NETDEV_TX_BUSY ; +if (!netif_queue_stopped (dev)) +/* churn baby churn .. */ + ret =dev_hard_start_xmit (skb, dev); + +if (!lockless) +netif_tx_unlock (dev); + +spin_lock (&dev->queue_lock); + +/* we need to refresh q because it may be invalid since +* we dropped dev->queue_lock earlier ... +* So dont try to be clever grasshopper +*/ + q = dev->qdisc; +/* most likely result, packet went ok */ +if (ret ==NETDEV_TX_OK ) +return qdisc_qlen (q); +/* only for lockless drivers .. */ +if (ret ==NETDEV_TX_LOCKED && lockless) +return tx_islocked (skb, dev, q); + +if (unlikely (ret !=NETDEV_TX_BUSY &&net_ratelimit ())) +printk (KERN_WARNING " BUG %s code %d qlen %d\n" , + dev->name, ret, q->q.qlen); + +return do_dev_requeue (skb, dev, q); }void __qdisc_run (struct net_device *dev)
Quoted 3 lines
Incoporating last comments from Patrick.
Dave, I think this is ready to apply - against net-2.6 from this morning.
Aside from the whitespace changes, I've run this patch over the weekend with no issue under moderate stress. Looks fine to me.
Thx,
-PJ Waskiewicz
Quoted 3 lines · jamal
Incoporating last comments from Patrick.
Dave, I think this is ready to apply - against net-2.6 from this morning.
Dave - hold onto this patch - Sridhar Samudrala <[email protected]> caught a bug i need to fix (GSO wont work).
The game starting in a few holds priority for now. I will send a patch either tonight or tommorow morning.
Otoh, if you have applied it - it is a simple incremental patch.
cheers,
jamal
David Miller
maintainerFrom: jamal <[email protected]>
Date: Wed, 16 May 2007 18:55:04 -0400
Quoted 7 lines · jamal
Quoted 3 lines
Incoporating last comments from Patrick.
Dave, I think this is ready to apply - against net-2.6 from this morning.Dave - hold onto this patch - Sridhar Samudrala <[email protected]> caught a bug i need to fix (GSO wont work).
No problems, I'm busy with something else anyways.
Quoted 1 line
The game starting in a few holds priority for now.
And I can't say whether or not it is related to that :-)
Quoted 7 lines · jamal
Quoted 3 lines · jamal
Incoporating last comments from Patrick.
Dave, I think this is ready to apply - against net-2.6 from this morning.Dave - hold onto this patch - Sridhar Samudrala <[email protected]> caught a bug i need to fix (GSO wont work).
I think your qdisc_restart() cleanup patch is OK. There you are still calling dev_hard_start_xmit() and GSO should work fine.
But the batch xmit patch has some issues.
Thanks
Sridhar
Quoted 6 lines
The game starting in a few holds priority for now. I will send a patch either tonight or tommorow morning.
Otoh, if you have applied it - it is a simple incremental patch.cheers,
jamal
Quoted 3 lines · Sridhar Samudrala
I think your
qdisc_restart()cleanup patch is OK. There you are still callingdev_hard_start_xmit()and GSO should work fine.
Sorry, you are right - I thought i cutnpasted from it. So Dave, false alarm ;-> That patch is clean.
Quoted 1 line
But the batch xmit patch has some issues.
I will resubmit that. I nheed to think about it a little. It should work fine as is if no GSO or nit.
cheers,
jamal
* jamal <[email protected]> 2007-05-10 20:13
Quoted 19 lines
include/linux/netdevice.h +3
-0@@ -83,6 +83,9 @@ struct wireless_dev;#define NETDEV_TX_OK 0/* driver took care of packet */ #define NETDEV_TX_BUSY 1/* driver tx path was busy*/ #define NETDEV_TX_LOCKED -1/* driver tx lock was already taken */ +#define NETDEV_TX_DROP -2/* request caller to drop packet */ +#define NETDEV_TX_QUEUE -3/* request caller to requeue packet */ +/* * Compute the worst case header length according to the protocols net/sched/sch_generic.c +0
-0
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 66
Quoted 9 lines
net/sched/sch_generic.c +5
-1@@ -59,7 +59,79 @@ void qdisc_unlock_tree(struct net_device *dev)spin_unlock_bh (&dev->queue_lock); } -/* +static inline int qqlen (struct Qdisc *q) +{ +BUG_ON ((int ) q->q.qlen <0 ); +return q->q.qlen; +}
Nitpicking but qdisc_qlen() sounds a lot better to me.
Quoted 15 lines
net/sched/sch_generic.c +15
-0+ +static inline int handle_dev_cpu_collision (struct net_device *dev) +{ +if (dev->xmit_lock_owner ==smp_processor_id ()) { +if (net_ratelimit ()) +printk (KERN_DEBUG +"Dead loop on netdevice %s, fix it urgently!\n" , + dev->name); +return NETDEV_TX_DROP ; + } +/* XXX: This maintains original code, but i think we should +* update cpu_collision always +*/ +__get_cpu_var (netdev_rx_stat).cpu_collision++; +return NETDEV_TX_QUEUE ;
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 91
Quoted 10 lines
net/sched/sch_generic.c +10
-0+} + +static inline int +handle_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (q == &noop_qdisc)) +kfree_skb (skb); + +if (skb->next)
Oops, missing else here, this will crash.
Quoted 12 lines
net/sched/sch_generic.c +12
-0+ dev->gso_skb = skb; + else + q->ops->requeue(skb, q); +/* XXX: Could netif_schedule fail? Or is that fact we are +* requeueing imply the hardware path is closed +* and even if we fail, some interupt will wake us +*/ +netif_schedule (dev); +return 0 ; +} + +static inline int
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 113
Quoted 10 lines
net/sched/sch_generic.c +10
-0+try_get_tx_pkt(struct sk_buff **skb, struct net_device *dev, struct Qdisc *q) +{ + if (((*skb = dev->gso_skb)) || ((*skb = q->dequeue(q)))) { + + dev->gso_skb =NULL ; +return -1 ; + } + +return 0 ; +}
Seems simpler to just return the skb or NULL instead
Quoted 62 lines
net/sched/sch_generic.c +36
-10+static inline int +handle_tx_locked (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ +int ret =handle_dev_cpu_collision (dev); + +if (ret ==NETDEV_TX_DROP ) { +kfree_skb (skb); +return qqlen (q); + } + +return handle_dev_requeue (skb, dev, q); +} + + +/* +NOTE: Called under dev->queue_lock with locally disabled BH. + +__LINK_STATE_QDISC_RUNNING guarantees only one CPU +can enter this region at a time. +dev->queue_lock serializes queue accesses for this device AND dev->qdisc pointer itself. @@ -67,114 +139,63 @@ void qdisc_unlock_tree(struct net_device *dev)dev->queue_lock and netif_tx_lock are mutually exclusive, if one is grabbed, another must be free. -*/ +Multiple CPUs may contend for the two locks. -/* Kick device. +Note, that this procedure can be called by a watchdog timer, so that +we do not check dev->tbusy flag here. +Returns to the caller: Returns: 0 - queue is empty or throttled. >0 - queue is not empty. - -NOTE: Called under dev->queue_lock with locally disabled BH. */ -static inline int qdisc_restart (struct net_device *dev) +int +qdisc_restart (struct net_device *dev) {struct Qdisc *q = dev->qdisc; -struct sk_buff *skb; - -/* Dequeue packet */ -if (((skb = dev->gso_skb)) || ((skb = q->dequeue (q)))) { -unsigned nolock = (dev->features &NETIF_F_LLTX ); +unsigned lockless = (dev->features &NETIF_F_LLTX ); +struct sk_buff *skb =NULL ; +int ret; + + ret =try_get_tx_pkt (&skb, dev, q); +if (ret ==0 ) +return qqlen (q); + +/* we have a packet to send */ +if (!lockless) {
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 175
Quoted 10 lines
net/sched/sch_generic.c +8
-1+if (!netif_tx_trylock (dev)) +return handle_tx_locked (skb, dev, q); + } + +/* all clear .. */ +spin_unlock (&dev->queue_lock); - dev->gso_skb =NULL ; +/* churn baby churn .. */ + ret =dev_hard_start_xmit (skb, dev);
Check whether queue is stopped got lost here
Quoted 55 lines
net/sched/sch_generic.c +3
-51-/* -* When the driver has LLTX set it does its own locking -* in start_xmit. No need to add additional overhead by -* locking again. These checks are worth it because -* even uncongested locks can be quite expensive. -* The driver can do trylock like here too, in case -* of lock congestion it should return -1 and the packet -* will be requeued. -*/ -if (!nolock) { -if (!netif_tx_trylock (dev)) { - collision: -/* So, someone grabbed the driver. */ - -/* It may be transient configuration error, -when hard_start_xmit() recurses. We detect -it by checking xmit owner and drop the -packet when deadloop is detected. -*/ -if (dev->xmit_lock_owner ==smp_processor_id ()) { -kfree_skb (skb); -if (net_ratelimit ()) -printk (KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n" , dev->name); - goto out; - } -__get_cpu_var (netdev_rx_stat).cpu_collision++; - goto requeue; - } - } +if (!lockless) +netif_tx_unlock (dev); - { -/* And release queue */ -spin_unlock (&dev->queue_lock); - -if (!netif_queue_stopped (dev)) { -int ret; - - ret =dev_hard_start_xmit (skb, dev); -if (ret ==NETDEV_TX_OK ) { -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto out; - } -if (ret ==NETDEV_TX_LOCKED && nolock) { -spin_lock (&dev->queue_lock); - q = dev->qdisc; - goto collision; - } - } +spin_lock (&dev->queue_lock);
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 182
Quoted 10 lines
net/sched/sch_generic.c +1
-8-/* NETDEV_TX_BUSY - we need to requeue */ -/* Release the driver */ -if (!nolock) { -netif_tx_unlock (dev); - } -spin_lock (&dev->queue_lock); - q = dev->qdisc; - } + q = dev->qdisc;
Maybe comment why q needs to be refreshed so nobody removes it by accident.
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 189
Quoted 7 lines
net/sched/sch_generic.c +7
-0+/* most likely result, packet went ok */ +if (ret ==NETDEV_TX_OK ) +return qqlen (q); +/* only for lockless drivers .. */ +if (ret ==NETDEV_TX_LOCKED && lockless) { +return handle_tx_locked (skb, dev, q); + }
Your style of placing parentheses seems to vary randomly :-)
Quoted 11 lines · Thomas Graf
* jamal <[email protected]> 2007-05-10 20:13
Quoted 9 lines
* Compute the worst case header length according to the protocols
net/sched/sch_generic.c +0
-0@@ -59,7 +59,79 @@ void qdisc_unlock_tree(struct net_device *dev)spin_unlock_bh (&dev->queue_lock); }
Quoted 9 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 66
Quoted 6 lines
net/sched/sch_generic.c +5
-1-/* +static inline int qqlen (struct Qdisc *q) +{ +BUG_ON ((int ) q->q.qlen <0 ); +return q->q.qlen; +}Nitpicking but
qdisc_qlen()sounds a lot better to me.
I thought nitpicking was my department ;->qdisc_qlen is more descriptive; i will change it.
Quoted 11 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 91
Quoted 8 lines
net/sched/sch_generic.c +8
-0+static inline int +handle_dev_requeue (struct sk_buff *skb,struct net_device *dev,struct Qdisc *q) +{ + +if (unlikely (q == &noop_qdisc)) +kfree_skb (skb); + +if (skb->next)Oops, missing else here, this will crash.
tabarnaq. I will fix - thanks.
Quoted 2 lines
Seems simpler to just return the skb or NULL instead
sure.
Quoted 5 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 175
Quoted 2 lines
net/sched/sch_generic.c +2
-0+/* churn baby churn .. */ + ret =dev_hard_start_xmit (skb, dev);Check whether queue is stopped got lost here
Also caught by Peter - will fix.
Quoted 4 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 182
Quoted 1 line
net/sched/sch_generic.c +1
-0+ q = dev->qdisc;Maybe comment why q needs to be refreshed so nobody removes it by accident.
Makes sense ..
Quoted 9 lines
net/sched/sch_generic.c void qdisc_unlock_tree(struct net_device *dev) · line 189
Quoted 7 lines
net/sched/sch_generic.c +7
-0+/* most likely result, packet went ok */ +if (ret ==NETDEV_TX_OK ) +return qqlen (q); +/* only for lockless drivers .. */ +if (ret ==NETDEV_TX_LOCKED && lockless) { +return handle_tx_locked (skb, dev, q); + }Your style of placing parentheses seems to vary randomly :-)
There may have been a debug in there that i removed. Frankly i dont mind having braces even for a one liner, but i know that is not popular with many folks.
Any other spot you noticed?
Thanks a lot Thomas.
cheers,
jamal