McKenney: So You Want to Rust the Linux Kernel? [LWN.net]

McKenney: So You Want to Rust the Linux Kernel?

Posted Oct 11, 2021 0:19 UTC (Mon) by neilbrown (subscriber, #359)
In reply to: McKenney: So You Want to Rust the Linux Kernel? by PaulMcKenney
Parent article: McKenney: So You Want to Rust the Linux Kernel?

Hi Paul,
yes that does help - particularly in providing me with specific points to disagree with - or at least to discuss.

While use of a refcount does provide the opportunity to read the counter value (kref_read()) it is my understanding that the code has no business doing that. The value could change between the moment when it is read and the moment when it is used, making it unreliable at best. However Linux does have a surprisingly large number of calls to kref_read(). Many are in debugging messages, which would be expected to be racy. I haven't analysed the others - probably many are justified. In any case I don't think this ability is core to refcounts in general, just useful in some particular refcount patterns.

While RCU cannot fail to claim the bulk-reference, rcu_dereference() can still return NULL. This can be roughly equivalent to kref_get_unless_zero() failing.

While the CONFIG_PREEMPT=n RCU implementation has no explicit reference counter, it seems likely to me that there is some implicit counter. At any moment there is some set of events which need to happen before a grace period can end. I think they are "a CPU calls schedule() or returns to user-space". The size of this set is the implicit ref-count.

While I accept that replacing reader/writer locking may be a common path to RCU, I don't think that at all justifies presenting such a flawed analogy. Describing RCU as "a way of waiting for already-started things to finish" is equally unjustified, though for completely different reasons. I personally prefer "exploiting deferred destruction" as the description, but even that misses the core idea.

The core idea of any RCU-solution is to use write-side atomics and read-side barriers. They are what replaces the various sorts of spinlocks. The use of rcu_read_lock() provides a context where these barriers and atomics can be used safely - particularly when combined with pointer dereference. I guess you could argue that this parallels a spinlock which provides a context where normal loads and stores can be used safely. Whether that parallel is enough to justify the analogy .... I'm not sure.

I was surprised by the strong contrast you suggested between RCU and hazard-pointers. They are both deferred-destruction mechanisms which seems to me to have similar use cases. In fact your perfbook suggests it may be possible for an implementation to switch between them based on dynamic assessment of performance needs. This points to substantial similarity.
I guess the distinction you are highlighting is that RCU can only hold its bulk-reference for a bounded time (or a least, there are costs that are at least linear in hold time). Hazard-pointers can protect a single pointer for an arbitrary time for comparatively little cost. This allows hazard pointers to *replace* refcounts in a way that RCU cannot.

This suggests I could clarify my analogy to say that "RCU provides an effective TRANSIENT reference-count on *all* objects...".
Or something like that.

The heart of my argument is that a spinlock typically prevents an object from changing, while a refcount prevents it from being destroyed. Surely RCU is more like the latter than the former.

Behind these thoughts is the observation that I read from time to time that RCU is difficult. Similarly that memory-barriers.txt is confusing. These comments remind me of the famous "you are not expected to understand this" in the Unix Edition 6 scheduler. It has always bothered me.
All computing is difficult, but we build abstraction and tools to make it manageable. "While" loops are better than goto's. Recursion is rendered transparent by invariants. Resource management can be greatly simplified by ownership tracking.
Telling people that something is hard, or that it does not need to be understood, is self-defeating as some of them might believe you and not even try.

If RCU is perceived as difficult then we (collectively) would be well served by efforts to find the weak points in our presentation of it and to strengthen them (and you are typically at the forefront of this). I genuinely think that "Like a lock" is one of those weak points.
For myself, I've found that thinking about RCU as a transient refcount on everything (which makes use of atomics safe) made it a lot easier to think about. Similarly the introduction of "read_acquire" and "store_release" terminology makes it much easier to think about memory barriers than "read-memory-barrier" and "write-memory-barrier" ever did.

Maybe the refcount analogy wouldn't work as well for others, and maybe we should avoid analogies altogether. Certainly "a lock that doesn't provide exclusion" doesn't seem like a winning description.

"RCU supports find-grained concurrency control using atomics and memory barriers and particularly allows pointers to be dereferenced without holding any lock" ??
(Some might say "supports lockless programming" but I hate that term as there still are locks - the CPU/memory controller manage them internally)

Thanks.

to post comments

McKenney: So You Want to Rust the Linux Kernel?

Posted Oct 11, 2021 17:15 UTC (Mon) by PaulMcKenney (✭ supporter ✭, #9624) [Link] (2 responses)

Hello, Neil, and thank you for your thoughts!

This reminds me of a queuing-theory class I took some decades back, though more than 50 years after the passing of Agner Karup Erlang, who laid the foundations for modern queuing theory. The professor explained one aspect of queuing theory (I forget which, sorry!). Then explained it again. And again. At which point, I raised my hand and asked why we were going over this so many times. The professor complimented me on understanding that aspect so quickly and then asked me to be quiet while he continued getting it across to the remainder of the class. I eventually figured out that he was explaining that aspect of queuing theory from different viewpoints, so it could be understood by different mindsets. I just happened to be lucky in that he happened to pick my mindset first.

I thought nothing of this incident at the time, but have had many occasions to recall it since Jack and I invented RCU.

And this is because of the wide gap between RCU's base-level semantics and typical use cases. And the base semantic really is "waiting for pre-existing things to finish", as you can see from the LKMM formalization of both RCU and SRCU. And because of this wide gap, I am not at all surprised at your discomfort with this description. It is a good place to start only for those few people who like building up from the fundamentals (not me!), and is in many cases worse than useless for the many people who prefer thinking in terms of tools to take on particular use cases.

To your later "deferred destruction" point, Maged Michael and I group Hazard Pointers and RCU under the "deferred reclamation" heading, along with reference counting. But not all RCU use cases involve reclamation. Plus this loses much of the connection to replacements for reader-writer locking. (The reason we don't say "deferred destruction" has to do with C++ destructors, in case you were wondering.)

So it is necessary to present multiple views of RCU, just as it was necessary for that professor to present multiple views of that aspect of queuing theory, whatever it might have been.

And your reference-coiunting view is a most excellent view, which is why it has its own section in perfbook. More than one, in fact: 9.5.4.3 "RCU is a Restricted Reference-Counting Mechanism" and 9.5.4.4 "RCU is a Bulk Reference-Counting Mechanism". But to your point, it does not seem to be well represented in Documentation/RCU. Would you like to add it? (Not as a replacement for the reader-writer-locking analogy, but as an additional view.)

Yes, RCU and hazard pointers overlap quite a bit. See Tables 9.6 and 9.7 in perfbook for Maged's and my view of the advantages and drawbacks of each. The point I am trying to make is that if your code already uses reference counts and already handles reference-count acquisition failure, the software-engineering disadvantages of hazard pointers don't matter so much to you. Yes, you might simplify your code by switching to RCU, but there might not be so much incentive to do so. Similarly, if your code already uses reader-writer locking, you will have kept your read-side critical sections reasonably short, so the memory-footprint disadvantages of RCU don't matter so much to you. But it is all just software, and thus can all be changed. In theory, anyway. ;-)

A common RCU use case resembling reader-writer locking avoids changing data (as opposed to list pointers). This use case has readers seeing constant data, just as they would with classic reader-writer locking (but not with some of the more esoteric use cases that cause some people to not unreasonably object to the name "reader-writer locking"). And to your point, one reason why the C++ committee was interested in both RCU and hazard pointers was that they saw them as ways of taming many classes of lockless algorithms. And the C++ community is one reason why I shy away from describing RCU in terms of reference counting. Once someone in that community dives down the smart-pointer rabbit hole, it is extremely difficult for them to understand RCU. And almost impossible to haul them back out of that hole.

To people claiming that RCU is difficult, I go back further in time, to a late 1970s lecture by the late Edsger Dijkstra that I had the privilege of attending. He claimed that "while" loops were difficult and that the typical programmer could not be trusted to write one correctly. This claim resonated with much of the audience (though not with my then-young and arrogant self), but probably would not resonate today. So part of the perception of RCU's difficulty is cultural, and the culture is changing. Me, I have had quite a number of people tell me that they had heard that RCU was really difficult, and were surprised to find that not only was it reasonably easy, but that it simplified things (deadlock avoidance being the source of simplification that is usually cited).

So from what I can see, the trick is to get a given developer pointed at the view of RCU that best matches the use case at hand. Thoughts on attacking that problem?

And again, can I interest you in adding an "RCU as reference counter" .rst file to Documentation/RCU?

McKenney: So You Want to Rust the Linux Kernel?

Posted Oct 12, 2021 22:09 UTC (Tue) by neilbrown (subscriber, #359) [Link] (1 responses)

Hi Paul,
thanks for sharing your experience and wisdom!

> to a late 1970s lecture by the late Edsger Dijkstra that I had the privilege of attending

...envy...

> And again, can I interest you in adding an "RCU as reference counter" .rst file to Documentation/RCU?

Ahhh... the gentle art of delegation. :-) Stay tuned.

McKenney: So You Want to Rust the Linux Kernel?

Posted Oct 13, 2021 0:38 UTC (Wed) by PaulMcKenney (✭ supporter ✭, #9624) [Link]

I very much look forward to seeing what you come up with!