Comment by haberman | Hacker News Reader

Comment by haberman | Hacker News Reader

Nice article, though I think that any intro-level material on lock-free programming should always include a "don't try this at home for anything important" warning. Until you have some experience with this stuff you will almost certainly make mistakes, but these mistakes might only manifest themselves as crashes in extremely rare circumstances.

I wrote my first lock-free code in 2004 based on reading some papers by Maged Michael from IBM. I wrote a lock-free FIFO in PowerPC assembly, and was convinced it was safe and robust. When I emailed Maged about it, he pointed out that if a thread was suspended on one specific instruction and some specific memory was unmapped before it could run again, the program could crash. I was amazed; I had thought hard about this algorithm, but had completely missed that possibility.

Some other specific notes about the article:

> Basically, if some part of your program satisfies the following conditions, then that part can rightfully be considered lock-free.

The are actually several levels of lock-freedom defined in the literature: lock-freedom, wait-freedom, and obstruction-freedom. For more info see: http://en.wikipedia.org/wiki/Non-blocking_algorithm

> Processors such as PowerPC and ARM expose load-link/store-conditional instructions, which effectively allow you to implement your own RMW primitive at a low level, though this is not often done.

One benefit of load-linked/store-conditional (often abbreviated LL/SC) is that it avoids the ABA problem (http://en.wikipedia.org/wiki/ABA_problem). In practice this doesn't matter that much since x86 doesn't support LL/SC, but I just think it's an interesting factoid to know.

> For instance, PowerPC and ARM processors can change the order of memory stores relative to the instructions themselves, but the x86/64 family of processors from Intel and AMD cannot.

(I've edited my reply here since my original assertion was incorrect). It's true that x86/64 won't reorder stores (see http://en.wikipedia.org/wiki/Memory_ordering for details) but it will reorder loads, so memory barriers are still required in some situations. However I believe that the atomic instructions ("lock cmpxchg", and "lock xadd") imply full barriers on x86.

I totally agree with you about the dangers of lock-free programming, and that every programmer will certainly make mistakes. I feel strongly enough about it that I actually wrote several paragraphs about the importance of stress testing lock-free algorithms, which I eventually decided to break off into a separate post... I might get around to publishing that, we'll see. We're not the first to feel this way :)

> The are actually several levels of lock-freedom defined in the literature: lock-freedom, wait-freedom, and obstruction-freedom. For more info see: http://en.wikipedia.org/wiki/Non-blocking_algorithm

It's true, and I had a hard time reconciling that Wikipedia page with how others had been describing lock-free programming, which is why I spent so much of the post on semantics.

> One benefit of load-linked/store-conditional (often abbreviated LL/SC) is that it avoids the ABA problem

Good point. Though I suspect if you access other cache lines between the LL & SC, it may invalidate the reservation.

> I don't think this is true about x86/64.

It is true about x86/64, at least for the most common, non-SSE instructions and normal, non-write-combined memory. I'll update the post to be more specific. See volume 3, section 8.2.2 of the Intel 64 and IA-32 Developer's Manuals: (http://www.intel.com/content/www/us/en/processors/architectu...)

- Reads are not reordered with other reads. - Writes are not reordered with older reads. - Writes to memory are not reordered with other writes (with a few exceptions for special instructions)

> It is true about x86/64

Yes, sorry I misspoke initially about stores being reordered, but it does still appear that stores can be reordered after loads, see: http://bartoszmilewski.com/2008/11/05/who-ordered-memory-fen... . (Your Intel manual reference is surely a more authoritative source but I'm not able to do a deep dive into it at the moment).

There's no contradiction. Normally, the x86/64 memory model is quite strong, preserving LoadLoad, LoadStore, and StoreStore ordering. But doesn't preserve StoreLoad ordering; eg. stores can be reordered after loads, as you point out. As I understand it, that's mainly because there's a limit to how quickly each store can be seen by other processors.

I'm pretty sure it's because of the store forwarding hardware. Intel CPUs can return a successful load of a recently-stored address with low latency (basically by keeping a special purpose cache of recent store addresses in the pipeline and returning their values before the actual commit). But that means that the "store" is viewed from the issuing CPU to have committed long before loads that might have been filled by caches on the other CPUs. There's no way to preserve both this optimization and a unified order.

> It is true about x86/64, at least for the most common, non-SSE instructions and normal, non-write-combined memory

To embellish further and avoid jargon for those who didn't parse this, it basically means "It is true except for the interfaces which explicitly document where it's not.".

Intel CPUs have had a "uncached/write-combining" mode for years (originally specified through special MTRR registers, then in the Page Attribute Table, and most recently in the "non-temporal" MOV instructions in SSE) which officially relaxes the requirements for store ordering. The reason for having two sets of semantics is performance: when dealing with uncached memory (typically memory used by another device, e.g. a video framebuffer) combining a bunch of writes to the same DRAM burst unit (or set of such units) is much faster than doing each word-sized I/O individually.

To support your first point, I implemented Maged's lock-free allocator so we could compare against it (http://www.scott-a-s.com/files/michael.tar.gz), and I never got it working correctly on Itanium. I spent weeks trying to figure out what was wrong. I eventually concluded that I did not understand the Itanium architecture and the algorithms used in the allocator well enough to be confident I could fix it - and that it wasn't worth improving my understanding of either. (I asked Maged if he had any experience with his allocator on Itanium, and he did not.)

Reply on news.ycombinator.com