Re: [PATCH RFC v7 00/23] DEPT(Dependency Tracker)

[Boqun Feng <boqun.feng@gmail.com>]

On Fri, Jan 20, 2023 at 10:51:45AM +0900, Byungchul Park wrote:
> Boqun wrote:
> > On Thu, Jan 19, 2023 at 01:33:58PM +0000, Matthew Wilcox wrote:
> > > On Thu, Jan 19, 2023 at 03:23:08PM +0900, Byungchul Park wrote:
> > > > Boqun wrote:
> > > > > *Looks like the DEPT dependency graph doesn't handle the
> > > > > fair/unfair readers as lockdep current does. Which bring the
> > > > > next question.
> > > > 
> > > > No. DEPT works better for unfair read. It works based on wait/event. So
> > > > read_lock() is considered a potential wait waiting on write_unlock()
> > > > while write_lock() is considered a potential wait waiting on either
> > > > write_unlock() or read_unlock(). DEPT is working perfect for it.
> > > > 
> > > > For fair read (maybe you meant queued read lock), I think the case
> > > > should be handled in the same way as normal lock. I might get it wrong.
> > > > Please let me know if I miss something.
> > > 
> > > From the lockdep/DEPT point of view, the question is whether:
> > > 
> > >	read_lock(A)
> > >	read_lock(A)
> > > 
> > > can deadlock if a writer comes in between the two acquisitions and
> > > sleeps waiting on A to be released.  A fair lock will block new
> > > readers when a writer is waiting, while an unfair lock will allow
> > > new readers even while a writer is waiting.
> > > 
> > 
> > To be more accurate, a fair reader will wait if there is a writer
> > waiting for other reader (fair or not) to unlock, and an unfair reader
> > won't.
> 
> What a kind guys, both of you! Thanks.
> 
> I asked to check if there are other subtle things than this. Fortunately,
> I already understand what you guys shared.
> 
> > In kernel there are read/write locks that can have both fair and unfair
> > readers (e.g. queued rwlock). Regarding deadlocks,
> > 
> > 	T0		T1		T2
> > 	--		--		--
> > 	fair_read_lock(A);
> > 			write_lock(B);
> > 					write_lock(A);
> > 	write_lock(B);
> > 			unfair_read_lock(A);
> 
> With the DEPT's point of view (let me re-write the scenario):
> 
> 	T0		T1		T2
> 	--		--		--
> 	fair_read_lock(A);
> 			write_lock(B);
> 					write_lock(A);
> 	write_lock(B);
> 			unfair_read_lock(A);
> 	write_unlock(B);
> 	read_unlock(A);
> 			read_unlock(A);
> 			write_unlock(B);
> 					write_unlock(A);
> 
> T0: read_unlock(A) cannot happen if write_lock(B) is stuck by a B owner
>     not doing either write_unlock(B) or read_unlock(B). In other words:
> 
>       1. read_unlock(A) happening depends on write_unlock(B) happening.
>       2. read_unlock(A) happening depends on read_unlock(B) happening.
> 
> T1: write_unlock(B) cannot happen if unfair_read_lock(A) is stuck by a A
>     owner not doing write_unlock(A). In other words:
> 
>       3. write_unlock(B) happening depends on write_unlock(A) happening.
> 
> 1, 2 and 3 give the following dependencies:
> 
>     1. read_unlock(A) -> write_unlock(B)
>     2. read_unlock(A) -> read_unlock(B)
>     3. write_unlock(B) -> write_unlock(A)
> 
> There's no circular dependency so it's safe. DEPT doesn't report this.
> 
> > the above is not a deadlock, since T1's unfair reader can "steal" the
> > lock. However the following is a deadlock:
> > 
> > 	T0		T1		T2
> > 	--		--		--
> > 	unfair_read_lock(A);
> > 			write_lock(B);
> > 					write_lock(A);
> > 	write_lock(B);
> > 			fair_read_lock(A);
> > 
> > , since T'1 fair reader will wait.
> 
> With the DEPT's point of view (let me re-write the scenario):
> 
> 	T0		T1		T2
> 	--		--		--
> 	unfair_read_lock(A);
> 			write_lock(B);
> 					write_lock(A);
> 	write_lock(B);
> 			fair_read_lock(A);
> 	write_unlock(B);
> 	read_unlock(A);
> 			read_unlock(A);
> 			write_unlock(B);
> 					write_unlock(A);
> 
> T0: read_unlock(A) cannot happen if write_lock(B) is stuck by a B owner
>     not doing either write_unlock(B) or read_unlock(B). In other words:
> 
>       1. read_unlock(A) happening depends on write_unlock(B) happening.
>       2. read_unlock(A) happening depends on read_unlock(B) happening.
> 
> T1: write_unlock(B) cannot happen if fair_read_lock(A) is stuck by a A
>     owner not doing either write_unlock(A) or read_unlock(A). In other
>     words:
> 
>       3. write_unlock(B) happening depends on write_unlock(A) happening.
>       4. write_unlock(B) happening depends on read_unlock(A) happening.
> 
> 1, 2, 3 and 4 give the following dependencies:
> 
>     1. read_unlock(A) -> write_unlock(B)
>     2. read_unlock(A) -> read_unlock(B)
>     3. write_unlock(B) -> write_unlock(A)
>     4. write_unlock(B) -> read_unlock(A)
> 
> With 1 and 4, there's a circular dependency so DEPT definitely report
> this as a problem.
> 
> REMIND: DEPT focuses on waits and events.

Do you have the test cases showing DEPT can detect this?

Regards,
Boqun

> 
> > FWIW, lockdep is able to catch this (figuring out which is deadlock and
> > which is not) since two years ago, plus other trivial deadlock detection
> > for read/write locks. Needless to say, if lib/lock-selftests.c was given
> > a try, one could find it out on one's own.
> > 
> > Regards,
> > Boqun
> >