Skip to contentSkip to conversation

Do not deprecate binary semaphore or do allow mutex in software interrupt contexts

DISCUSSION Matti Linnanvuori 3 participants 3 messages Sep 11, 2007

lore
Share & tools
lore
#0 lore raw

Arjan van de Ven:

Quoted 1 line

what do you do if the trylock fails?

Just do not read the status variable now but modify the timer to run later.

Quoted 2 lines

to be honest, the scenario describe really smells of broken locking, in fact it really sounds like it wants to use spinlocks instead

No, I don't think it is broken.
Spinlocks can be used, but I don't see them being obviously better in all cases. If access takes a long time, it is better to sleep during it.
And if you sleep, you might just end up creating a new mutex implementation with a spinlock.

Alan Cox:

Quoted 7 lines

For polling and timer based code its often simpler to do

   del_timer_sync(&my_timer);
   FrobStuff
   add_timer(&my_timer);

especially if "FrobStuff" is likely to change when you next need to poll.

In the scenario I presented, the timer modifies itself to run later.
Therefore, simply calling del_timer_sync is not enough but you have to set an atomic variable to prevent the timer from adding itself again. Again, you end up creating a new mutex implementation, which is not good.

       __________________________________ 

Yahoo! Clever - Der einfachste Weg, Fragen zu stellen und Wissenswertes mit Anderen zu teilen. www.yahoo.de/clever

#1 lore raw

[re-adding CCs, please do not drop these]

Quoted 9 lines · Matti Linnanvuori

Arjan van de Ven:

Quoted 1 line

what do you do if the trylock fails?

Just do not read the status variable now but modify the timer to run later.

Quoted 2 lines

to be honest, the scenario describe really smells of broken locking, in fact it really sounds like it wants to use spinlocks instead

No, I don't think it is broken.

Yes it is.

Quoted 4 lines

Spinlocks can be used, but I don't see them being obviously better in all cases. If access takes a long time, it is better to sleep during it.
And if you sleep, you might just end up creating a new mutex implementation with a spinlock.

If you have to wait a long time in an atomic context you've done something wrong. If you're only reading it from an atomic context you might consider storing a copy that can be quickly updated and protect that using a spinlock.

do_update ()
{
mutex_lock(&my_device_mutex); my_device_frob_state(&my_state); /* <-- this takes a _long_ while */ spin_lock_irq(&my_shadow_state_lock);
my_shadow_state = state; /* <-- this is a quick memcopy */ spin_unlock_irq(&my_shadow_state_lock); mutex_unlock(&my_device_mutex);
}

do_read()
{
spin_lock_irq(&my_shadow_state); do_something_with_shadow_state(&mt_shadow_state); spin_unlock_irq(&my_shadow_state);

return foo;
}

Quoted 7 lines

Alan Cox:

Quoted 6 lines

For polling and timer based code its often simpler to do

   del_timer_sync(&my_timer);
   FrobStuff
   add_timer(&my_timer);
Quoted 5 lines
Quoted 1 line

especially if "FrobStuff" is likely to change when you next need to poll.

In the scenario I presented, the timer modifies itself to run later. Therefore, simply calling del_timer_sync is not enough but you have to set an atomic variable to prevent the timer from adding itself again.

Not being too familiar with the timer stuff, it smells wrong what you say.

As for the whole polling method, consider what Alan said, don't do it if you don't need to. You'll annoy people at no end. Try to push state changes where possible.

#2 lore raw
In reply to Matti Linnanvuori
Quoted 10 lines · On Tue, 11 Sep 2007 09:20:23 -0700 (PDT) Matti Linnanvuori…

Arjan van de Ven:

Quoted 1 line

what do you do if the trylock fails?

Just do not read the status variable now but modify the timer to run later.

Quoted 3 lines

to be honest, the scenario describe really smells of broken locking, in fact it really sounds like it wants to use spinlocks instead

Quoted 5 lines

No, I don't think it is broken.
Spinlocks can be used, but I don't see them being obviously better in all cases. If access takes a long time, it is better to sleep during it. And if you sleep, you might just end up creating a new mutex implementation with a spinlock.

at this point the discussion has gone so theoretical that I think it's better to go with a real example. What actual source code do you think is a legit case for this?

I still think that whatever case you have in mind is better served with something else, but until we see the actual complete drier we're both talking air.