[webkit-dev] Atomic read/write operations and thread safety

[Mikhail Naganov]

In MSDN there is a widely cited article about lockless programming on
x86 and Xbox: http://msdn.microsoft.com/en-us/library/ee418650(v=vs.85).aspx

In short: using boolean as a flag for thread termination is OK. Using
boolean as a guard for multi-threaded access to data isn't OK, as read
/ write reordering can be made both by C++ compiler and CPU.

On Thu, Dec 16, 2010 at 00:34, Alexey Proskuryakov <ap at webkit.org> wrote:
>
> 15.12.2010, в 10:36, Stephan Aßmus написал(а):
>
>> The only thing one needs to watch out for is to declare such a boolean volatile (which I believe this code does, if memory serves). Otherwise the thread which polls the condition may read from a cached location and miss the change. Worst that can happen on a hypothecial architecture where writing a byte is not atomic is that the changed condition takes affect one loop cycle later.
>
>
> The behavior of volatile depends on the compiler - there is no guarantee that a memory barrier will be emitted. Historically, volatile existed not for multiprocessor systems, but for special addresses in address space (like memory mapped I/O), where it is important to actually access the address each time, and not than keep a local variable in a boolean.
>
> There is code in WebKit that doesn't care about the other thread immediately seeing the change of a boolean value, and just hopes that it will propagate soon enough in practice. I remember implementing worker thread termination in that way. This code is of course formally wrong.
>
> - WBR, Alexey Proskuryakov
>
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org
> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>