On Thu, Jan 12, 2023 at 07:20:48AM -0800, Paul E. McKenney wrote: > On Thu, Jan 12, 2023 at 08:24:16AM +0000, David Woodhouse wrote: > > On Wed, 2023-01-11 at 13:30 -0500, Paolo Bonzini wrote: > > > > > > +- ``synchronize_srcu(&kvm->srcu)`` is called inside critical sections > > > + for kvm->lock, vcpu->mutex and kvm->slots_lock. These locks _cannot_ > > > + be taken inside a kvm->srcu read-side critical section; that is, the > > > + following is broken:: > > > + > > > + srcu_read_lock(&kvm->srcu); > > > + mutex_lock(&kvm->slots_lock); > > > + > > > > "Don't tell me. Tell lockdep!" > > > > Did we conclude in > > https://lore.kernel.org/kvm/122f38e724aae9ae8ab474233da1ba19760c20d2.camel@xxxxxxxxxxxxx/ > > that lockdep *could* be clever enough to catch a violation of this rule > > by itself? > > > > The general case of the rule would be that 'if mutex A is taken in a > > read-section for SCRU B, then any synchronize_srcu(B) while mutex A is > > held shall be verboten'. And vice versa. > > > > If we can make lockdep catch it automatically, yay! > > Unfortunately, lockdep needs to see a writer to complain, and that patch > just adds a reader. And adding that writer would make lockdep complain > about things that are perfectly fine. It should be possible to make > lockdep catch this sort of thing, but from what I can see, doing so > requires modifications to lockdep itself. > Please see if the follow patchset works: https://lore.kernel.org/lkml/20230113065955.815667-1-boqun.feng@xxxxxxxxx "I have been called. I must answer. Always." ;-) > > If not, I'm inclined to suggest that we have explicit wrappers of our > > own for kvm_mutex_lock() which will do the check directly. > > This does allow much more wiggle room. For example, you guys could decide > to let lockdep complain about things that other SRCU users want to do. > For completeness, here is one such scenario: > > CPU 0: read_lock(&rla); srcu_read_lock(&srcua); ... > > CPU 1: srcu_read_lock(&srcua); read_lock(&rla); ... > > CPU 2: synchronize_srcu(&srcua); > > CPU 3: write_lock(&rla); ... > > If you guys are OK with lockdep complaining about this, then doing a Actually lockdep won't complain about this, since srcu_read_lock() is always a recursive read lock, so it won't break other srcu_read_lock(). FWIW if CPU2 or CPU3 does write_lock(&rla); synchronize_srcu(&srcua); it's a deadlock (with CPU 1) Regards, Boqun > currently mythical rcu_write_acquire()/rcu_write_release() pair around > your calls to synchronize_srcu() should catch the other issue. > > And probably break something else, but you have to start somewhere! ;-) > > Thanx, Paul