"Hen, Shmulik" wrote: > Is there a time limit on holding a lock ? None other than "not long." Spinlocks --should not-- be held for a long period of time. If you do this, IMHO that is a clear design bug that should be fixed. One RULE though: never, ever, call a function that might sleep while holding a spinlock. If you are locking all your functions with spin_lock_bh or similar, it sounds like you are doing this. > We had to choose spin_lock_bh inside do_ioctl since it is run as user > context and if hard_start_xmit was called the same time and tried to lock > too we would be in a deadlock (actually ,this is how we found out about > spin_*_bh :-). Why is spin_lock_bh appropriate for this case? Anyway, for preventing Tx: spin_lock(&dev->xmit_lock); netif_stop_queue(dev); spin_unlock(&dev->xmit_lock); For preventing Rx, if this is a hardware device, you should disable interrupts (using the card's interrupt mask register not cli...) to prevent Rx packets from arriving while the topology is changing. > We have to use a single lock since we don't know in advance what topology > changes could happen during do_ioctl so we need to halt all Tx/Rx until we > know for sure we can forward a packet to one of the base drivers we're bound > to etc. (I didn't want to bore everyone with code excerpts, but if you think > it would help, then...). It still sounds to me like a single spinlock here is wrong. Read the existing netdev code, especially net/core/dev.c and include/linux/netdevices.h. Jeff -- Jeff Garzik | Building 1024 | Would you like a Twinkie? MandrakeSoft | - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
