Message ID | 20090709215606.526259917@linutronix.de |
---|---|
State | Changes Requested, archived |
Delegated to: | David Miller |
Headers | show |
From: Thomas Gleixner <tglx@linutronix.de> Date: Thu, 09 Jul 2009 21:59:22 -0000 > The hrtimer callback cbq_undelay() is not serialized against > cbq_ovl_delay(). That affects at least q->pmask and q->delay_timer. > > Lock it proper. > > Signed-off-by: Thomas Gleixner <tglx@linutronix.de> The problems here are even much deeper than it appears. First of all, I am to understand that hrtimers run from hardware interrupt context, right? If so, all of these datastructures are softirq safe only. And it is not merely the immediate things you see being modified in this hrtimer, such as ->pmask etc., it is also the q->active[] pointers, the list state for the classes, just about everything in the qdisc state is referenced in this hrtimer code path. I wonder how many queer unexplainable bugs we see because of this. What should probably happen is that the hrtimer merely fires off work at software interrupt context (perhaps a tasklet or similar), and that software interrupt code take the qdisc's root lock throughout it's execution. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
David, On Sun, 12 Jul 2009, David Miller wrote: > From: Thomas Gleixner <tglx@linutronix.de> > Date: Thu, 09 Jul 2009 21:59:22 -0000 > > > The hrtimer callback cbq_undelay() is not serialized against > > cbq_ovl_delay(). That affects at least q->pmask and q->delay_timer. > > > > Lock it proper. > > > > Signed-off-by: Thomas Gleixner <tglx@linutronix.de> > > The problems here are even much deeper than it appears. > > First of all, I am to understand that hrtimers run from hardware > interrupt context, right? If so, all of these datastructures are > softirq safe only. > > And it is not merely the immediate things you see being modified in > this hrtimer, such as ->pmask etc., it is also the q->active[] > pointers, the list state for the classes, just about everything in the > qdisc state is referenced in this hrtimer code path. That's what I was worried about. > I wonder how many queer unexplainable bugs we see because of this. > > What should probably happen is that the hrtimer merely fires off work > at software interrupt context (perhaps a tasklet or similar), and that > software interrupt code take the qdisc's root lock throughout it's > execution. Sigh, I almost expected that the removal of the callback modes will fire back some day. Thanks, tglx -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Thomas Gleixner <tglx@linutronix.de> Date: Tue, 14 Jul 2009 10:55:14 +0200 (CEST) > David, > > On Sun, 12 Jul 2009, David Miller wrote: > >> What should probably happen is that the hrtimer merely fires off work >> at software interrupt context (perhaps a tasklet or similar), and that >> software interrupt code take the qdisc's root lock throughout it's >> execution. > > Sigh, I almost expected that the removal of the callback modes will > fire back some day. Well this makes hrtimers decidedly less useful for networking and we have a ton of bugs right now, basically in every hrtimer used by the networking currently. The only way we can use them, as things currently stand, is as triggers for softirq work. Is it really that troublesome to provide this kind of facility generically, rather than having various subsystems replicate such code where they want to use hrtimers and are restricted to softirqs? -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, 2009-07-14 at 09:00 -0700, David Miller wrote: > From: Thomas Gleixner <tglx@linutronix.de> > Date: Tue, 14 Jul 2009 10:55:14 +0200 (CEST) > > On Sun, 12 Jul 2009, David Miller wrote: > > > >> What should probably happen is that the hrtimer merely fires off work > >> at software interrupt context (perhaps a tasklet or similar), and that > >> software interrupt code take the qdisc's root lock throughout it's > >> execution. > > > > Sigh, I almost expected that the removal of the callback modes will > > fire back some day. > > Well this makes hrtimers decidedly less useful for networking and we > have a ton of bugs right now, basically in every hrtimer used by the > networking currently. > > The only way we can use them, as things currently stand, is as > triggers for softirq work. > > Is it really that troublesome to provide this kind of facility > generically, rather than having various subsystems replicate such code > where they want to use hrtimers and are restricted to softirqs? Linus really hated the softirq mode, which is what prompted me to change that. Now, it might be he only hated the particular interface and the resulting code, but I think to remember he simply thought the whole thing daft. I can look into adding it back if we can agree on the interface and code impact, but looking at: # git grep hrtimer_init net/ | sort -u net/can/bcm.c: hrtimer_init(&op->thrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); net/can/bcm.c: hrtimer_init(&op->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); net/sched/sch_api.c: hrtimer_init(&wd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); net/sched/sch_cbq.c: hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); I wonder if its worth the impact on the core kernel code, or whether its better for these few timers to kick off a tasklet or the like. Further, I don't think a lot of subsystems would need this, as the general trend is away from softirqs/tasklets and towards threads/workqueues as most people want to schedule. And for those hardirq hrtimers are good enough as a wakeup source. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, 14 Jul 2009, Peter Zijlstra wrote: > > Linus really hated the softirq mode, which is what prompted me to change > that. > > Now, it might be he only hated the particular interface and the > resulting code, but I think to remember he simply thought the whole > thing daft. Yes. And I hated the bugs it had. Don't make something as core as timers any more complicated. Don't take locks in timers and then complain about deadlocks. If your locking is broken, don't make the core timers be idiotically broken. Because it was. The code was a total mess to follow, and had bugs. Linus -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
David Miller wrote: > From: Thomas Gleixner <tglx@linutronix.de> > Date: Tue, 14 Jul 2009 10:55:14 +0200 (CEST) > >> David, >> >> On Sun, 12 Jul 2009, David Miller wrote: >> >>> What should probably happen is that the hrtimer merely fires off work >>> at software interrupt context (perhaps a tasklet or similar), and that >>> software interrupt code take the qdisc's root lock throughout it's >>> execution. >> Sigh, I almost expected that the removal of the callback modes will >> fire back some day. > > Well this makes hrtimers decidedly less useful for networking and we > have a ton of bugs right now, basically in every hrtimer used by the > networking currently. The CAN stuff is clean in this topic. See below. > > The only way we can use them, as things currently stand, is as > triggers for softirq work. > > Is it really that troublesome to provide this kind of facility > generically, rather than having various subsystems replicate such code > where they want to use hrtimers and are restricted to softirqs? Indeed this had been my concerns also, when i moved the hrtimer usage in a CAN protocol to use tasklets. ("can: update can-bcm for hrtimer hardirq callbacks") http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=6e5c172cf7ca1ab878cc6a6a4c1d52fef60f3ee0 due to ("hrtimer: removing all ur callback modes") http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=ca109491f612aab5c8152207631c0444f63da97f I was not very amused that time and wanted to NACK that change, but Linus said: "Quite frankly, your NAK doesn't matter. We've had too many bugs in hrtimers. They _will_ get simplified." (http://lkml.indiana.edu/hypermail/linux/kernel/0812.1/00218.html) Thomas, is there chance to get this nice simple possibility back to invoke at least a hrtimer for SOFT_IRQ context additional to the current functionality?? I would expect this to save lot's of tasklet code that is - and will be - created due to the lack of the hrtimers softirq capability ... FWIK there were really many callback modes before (per-cpu stuff and so), that were probably too much. But having a SOFT_IRQ callback mode again would really help. Regards, Oliver -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Index: linux-2.6/net/sched/sch_cbq.c =================================================================== --- linux-2.6.orig/net/sched/sch_cbq.c +++ linux-2.6/net/sched/sch_cbq.c @@ -163,6 +163,7 @@ struct cbq_sched_data psched_time_t now_rt; /* Cached real time */ unsigned pmask; + spinlock_t lock; struct hrtimer delay_timer; struct qdisc_watchdog watchdog; /* Watchdog timer, started when CBQ has @@ -503,6 +504,9 @@ static void cbq_ovl_delay(struct cbq_cla cl->undertime = q->now + delay; if (delay > 0) { + unsigned long flags; + + spin_lock_irqsave(&q->lock, flags); sched += delay + cl->penalty; cl->penalized = sched; cl->cpriority = TC_CBQ_MAXPRIO; @@ -518,6 +522,7 @@ static void cbq_ovl_delay(struct cbq_cla hrtimer_restart(&q->delay_timer); cl->delayed = 1; cl->xstats.overactions++; + spin_unlock_irqrestore(&q->lock, flags); return; } delay = 1; @@ -599,6 +604,7 @@ static enum hrtimer_restart cbq_undelay( now = psched_get_time(); + spin_lock(&q->lock); pmask = q->pmask; q->pmask = 0; @@ -623,6 +629,7 @@ static enum hrtimer_restart cbq_undelay( time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay)); hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS); } + spin_unlock(&q->lock); sch->flags &= ~TCQ_F_THROTTLED; __netif_schedule(qdisc_root(sch)); @@ -1396,6 +1403,7 @@ static int cbq_init(struct Qdisc *sch, s q->link.avpkt = q->link.allot/2; q->link.minidle = -0x7FFFFFFF; + spin_lock_init(&q->lock); qdisc_watchdog_init(&q->watchdog, sch); hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); q->delay_timer.function = cbq_undelay;
The hrtimer callback cbq_undelay() is not serialized against cbq_ovl_delay(). That affects at least q->pmask and q->delay_timer. Lock it proper. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> --- net/sched/sch_cbq.c | 8 ++++++++ 1 file changed, 8 insertions(+) -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html