On Sun, Jul 01, 2007 at 09:35:58PM +0400, Oleg Nesterov wrote:
Jarek Poplawski wrote:
quoted
#1
Until kernel ver. 2.6.21 (including) cancel_rearming_delayed_work()
required a work function should always (unconditionally) rearm with
delay > 0 - otherwise it would endlessly loop. This patch replaces
this function with cancel_delayed_work(). Later kernel versions don't
require this, so here it's only for uniformity.
But 2.6.22 doesn't need this change, why it was merged?
One bad reason is given above. Should I look for another one?
In fact, I suspect this change adds a race,
You are right!
quoted
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -72,7 +72,8 @@ static void queue_process(struct work_struct *work)
netif_tx_unlock(dev);
local_irq_restore(flags);
- schedule_delayed_work(&npinfo->tx_work, HZ/10);
+ if (atomic_read(&npinfo->refcnt))
+ schedule_delayed_work(&npinfo->tx_work, HZ/10);
return;
}
netif_tx_unlock(dev);
@@ -785,9 +786,15 @@ void netpoll_cleanup(struct netpoll *np)
if (atomic_dec_and_test(&npinfo->refcnt)) {
skb_queue_purge(&npinfo->arp_tx);
skb_queue_purge(&npinfo->txq);
- cancel_rearming_delayed_work(&npinfo->tx_work);
+ cancel_delayed_work(&npinfo->tx_work);
flush_scheduled_work();
Suppose that ->refcnt == 1, and queue_process() was preempted just after
atomic_read(&npinfo->refcnt).
netpoll_cleanup() comes, cancel_delayed_work() does nothing, flush_scheduled_work()
sleeps.
queue_process() gets CPU, re-schedules ->tx_work, and returns.
flush_scheduled_work() completes, netpoll_cleanup() frees npinfo and returns
while ->tx_work is pending.
No?
No no. (Yes!)
I had some doubts about this, and you found very good reason
for this.
I'll soon send a patch to restore cancel_rearming_delayed_work
in 2.6.22.
So, 2.6.21 needs something better (maybe you've found it btw.?),
but they weren't too interested, anyway.
Thanks very much & regards,
Jarek P.
On 07/02, Jarek Poplawski wrote:
quoted
quoted
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -72,7 +72,8 @@ static void queue_process(struct work_struct *work)
netif_tx_unlock(dev);
local_irq_restore(flags);
- schedule_delayed_work(&npinfo->tx_work, HZ/10);
+ if (atomic_read(&npinfo->refcnt))
+ schedule_delayed_work(&npinfo->tx_work, HZ/10);
return;
}
[...snip...]
So, 2.6.21 needs something better (maybe you've found it btw.?),
but they weren't too interested, anyway.
We can do a double flush trick. If queue_process() checks ->refcnt before
schedule_delayed_work() like above, netpoll_cleanup() can do
flush_scheduled_work();
// the next invocation of queue_process()
// must see ->refcnt == 0
if (!cancel_delayed_work(&npinfo->tx_work)) {
/* may be queued, wait for completion */
flush_scheduled_work();
}
Jarek, I don't understand net/, a silly question. Why do we need the #2 chunk?
Isn't it better to move skb_queue_purge(&npinfo->txq) after cancel_..._work()
instead?
Oleg.
On Mon, Jul 02, 2007 at 01:24:08PM +0400, Oleg Nesterov wrote:
On 07/02, Jarek Poplawski wrote:
quoted
quoted
quoted
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -72,7 +72,8 @@ static void queue_process(struct work_struct *work)
netif_tx_unlock(dev);
local_irq_restore(flags);
- schedule_delayed_work(&npinfo->tx_work, HZ/10);
+ if (atomic_read(&npinfo->refcnt))
+ schedule_delayed_work(&npinfo->tx_work, HZ/10);
return;
}
[...snip...]
So, 2.6.21 needs something better (maybe you've found it btw.?),
but they weren't too interested, anyway.
We can do a double flush trick. If queue_process() checks ->refcnt before
schedule_delayed_work() like above, netpoll_cleanup() can do
flush_scheduled_work();
// the next invocation of queue_process()
// must see ->refcnt == 0
if (!cancel_delayed_work(&npinfo->tx_work)) {
/* may be queued, wait for completion */
flush_scheduled_work();
}
I'll try to think about it later, but I don't plan to do next patch,
so feel free to send this. I didn't plan to fix netpoll at all
(I never didn't use nor studied this before...). But couldn't stand
this stupid lockup stays in 2.6.21. Now, I see it probably doesn't
annoy more than 2 or 3 people around...
Jarek, I don't understand net/, a silly question. Why do we need the #2 chunk?
Isn't it better to move skb_queue_purge(&npinfo->txq) after cancel_..._work()
instead?
I've thought about this too, but because I don't know netpoll/netconsole
enough I didn't want to change more than minimum needed.
skb_queue_purge() uses heavy locking (irqsave) and I don't remember now
if I've found the reason or simply believed somebody had to have a reason
to do this there, anyway, if moved after cancel_ it could be done without
this locking, and something like while () instead of my if () should be
enough.
But there was not much interest about this patch, and I'm not currently
interested to be the main netconsole expert too, so maybe you would
like to try...
Cheers,
Jarek P.