Skip to contentSkip to conversation

sis190

DISCUSSION Pascal CHAPPERON 2 participants 2 messages Jun 26, 2005 – Jun 30, 2005

lore
Share & tools
lore
#0 lore raw
Quoted 2 lines

Message du 22/06/05 01:02
De : "Francois Romieu" <[email protected]>

[...]

Quoted 2 lines

I have copied the sis190 patches at http://www.zoreil.com/~romieu/sis190
if someone wants to hack them in the meantime.

1) sis190 freezes the box when kernel PREEMPT is enabled :

I made many tries, but i could not solve it;
- it does not occur while receiving huge files.
- it does not occur when only a few packets are transmitted (remote connection, ls, find)
- it occurs only while transmiting huge files AND trying to do someting else (open a new term,...)
- I could transfer a huge file (700MB) several times as i was at the console (and i could switch to another
console to perform find, ls,... during the transfer).

I managed the system so the sis190 had its own IRQ, but it made no difference.

As i suspected nvidia driver, i switched to nv driver : no result.

It seems to me that a task inside the sis190_tx_interrupt() is not protected against preemption (and it is probably the same on a SMP not prempted).

I tried to play with spinlocks, but with no result :
@@ -621,6 +621,7 @@ static irqreturn_t sis190_interrupt(int

        void __iomem *ioaddr = tp->mmio_addr;
        int handled = 0;
        int boguscnt;

+ unsigned long flags;

        for (boguscnt = max_interrupt_work; boguscnt > 0; boguscnt--) {
                u32 status = SIS_R32(IntrStatus);
@@ -651,9 +652,9 @@ static irqreturn_t sis190_interrupt(int
                        sis190_rx_interrupt(dev, tp, ioaddr);

                if (status & TxQ0Int) {
-                       spin_lock(&tp->lock);
+                       spin_lock_irqsave(&tp->lock, flags);
                        sis190_tx_interrupt(dev, tp, ioaddr);
-                       spin_unlock(&tp->lock);
+                       spin_unlock_irqrestore(&tp->lock, flags);
                }
        }

or :

@@ -581,6 +581,7 @@ static void sis190_tx_interrupt(struct n
struct sis190_private *tp, void __iomem *ioaddr)

 {
        unsigned int tx_left, dirty_tx = tp->dirty_tx;

+ unsigned long flags;

        for (tx_left = tp->cur_tx - dirty_tx; tx_left > 0; tx_left--) {
                unsigned int entry = dirty_tx % NUM_TX_DESC;
@@ -604,10 +605,12 @@ static void sis190_tx_interrupt(struct n
                dirty_tx++;
        }
+       spin_lock_irqsave(&tp->lock, flags);
        if (tp->dirty_tx != dirty_tx) {
                tp->dirty_tx = dirty_tx;
                netif_wake_queue(dev);
        }

+ spin_unlock_irqrestore(&tp->lock, flags);
}

/*
@@ -651,9 +654,7 @@ static irqreturn_t sis190_interrupt(int

                        sis190_rx_interrupt(dev, tp, ioaddr);

                if (status & TxQ0Int) {
-                       spin_lock(&tp->lock);
                        sis190_tx_interrupt(dev, tp, ioaddr);
-                       spin_unlock(&tp->lock);
                }
        }

In fact, i don't know where are the critical sections...

2) sis190 freezes the box when the link partner is a r8169 forced in 10 full autoneg off (preempted or not preempted kernel) :

it is probably for the same reason, as it occurs only when transmitting a big file.

3) poor TX performances when link partner is forced to autoneg off (not preempted kernel):

Once again it seems that something goes wrong with TX tasks...

Regards
Pascal

#1 lore raw

Pascal CHAPPERON <[email protected]> :
[...]

Quoted 11 lines

1) sis190 freezes the box when kernel PREEMPT is enabled :

I made many tries, but i could not solve it;
- it does not occur while receiving huge files.
- it does not occur when only a few packets are transmitted (remote connection, ls, find)
- it occurs only while transmiting huge files AND trying to do someting else (open a new term,...)
- I could transfer a huge file (700MB) several times as i was at the console (and i could switch to another
console to perform find, ls,... during the transfer).

Are you saying that PREEMPT and X are both needed to freeze the box ?

If so, could you try preempt + console + rsync (+ dd from disk) ?

Quoted 4 lines

I managed the system so the sis190 had its own IRQ, but it made no difference.

As i suspected nvidia driver, i switched to nv driver : no result.

Please keep binary modules out of the loop until the things are stable. I do not want to try and diagnose a bug in a system wherein such an amount of unknown code was loaded (even if it was unloaded later).

Quoted 14 lines

It seems to me that a task inside the sis190_tx_interrupt() is not protected against preemption (and it is probably the same on a SMP not prempted).

I tried to play with spinlocks, but with no result :

(quoted diff) +1 -0
@@ -621,6 +621,7 @@ static irqreturn_t sis190_interrupt(int
        void __iomem *ioaddr = tp->mmio_addr;
        int handled = 0;
        int boguscnt;
+       unsigned long flags;
 
        for (boguscnt = max_interrupt_work; boguscnt > 0; boguscnt--) {
                u32 status = SIS_R32(IntrStatus);

Review static irqreturn_t sis190_interrupt(int · line 657

Quoted 8 lines
(quoted diff) +2 -2
@@ -651,9 +652,9 @@ static irqreturn_t sis190_interrupt(int
                        sis190_rx_interrupt(dev, tp, ioaddr);
 
                if (status & TxQ0Int) {
-                       spin_lock(&tp->lock);
+                       spin_lock_irqsave(&tp->lock, flags);
                        sis190_tx_interrupt(dev, tp, ioaddr);
-                       spin_unlock(&tp->lock);
+                       spin_unlock_irqrestore(&tp->lock, flags);

Afaik the irq handler is already protected against reentrancy (see kernel/irq.c::__do_IRQ) and softirq (see arch/xxx/kernel/irq.c::irq_exit). So this change should not make a difference.

[...]

Quoted 9 lines
(quoted diff) +1 -0
@@ -581,6 +581,7 @@ static void sis190_tx_interrupt(struct n
                                struct sis190_private *tp, void __iomem *ioaddr)
 {
        unsigned int tx_left, dirty_tx = tp->dirty_tx;
+       unsigned long flags;
 
        for (tx_left = tp->cur_tx - dirty_tx; tx_left > 0; tx_left--) {
                unsigned int entry = dirty_tx % NUM_TX_DESC;

Review static void sis190_tx_interrupt(struct n · line 614

Quoted 10 lines
(quoted diff) +2 -0
@@ -604,10 +605,12 @@ static void sis190_tx_interrupt(struct n
                dirty_tx++;
        }
 
+       spin_lock_irqsave(&tp->lock, flags);
        if (tp->dirty_tx != dirty_tx) {
                tp->dirty_tx = dirty_tx;
                netif_wake_queue(dev);
        }
+       spin_unlock_irqrestore(&tp->lock, flags);
 }

The irqsave/restore should not be needed for the same reason as above.

Quoted 1 line

In fact, i don't know where are the critical sections...

In the Tx path the critical section is related to netif_{start/stop}_queue.

Quoted 3 lines

2) sis190 freezes the box when the link partner is a r8169 forced in 10 full autoneg off (preempted or not preempted kernel) :

The r8169 driver will not necessarily do what you would expect when you autoneg it off and it faces an unstable driver. I'll send an update for it.

Any TX timeout message on the console ? Freeze == sysrq has no effect and keyboard leds do not blink any more ?

There is an updated version at
http://www.zoreil.com/~romieu/sis190/20050630-2.6.13-rc1-sis190-test.patch

It would be nice to know how it behaves wrt preempt (no need to experiment with the media management), especially if you can describe the freeze more specifically.

--
Ueimor