Re: [Boost-users] [Thread] Writer/Reader strange scheduling

Subject: Re: [Boost-users] [Thread] Writer/Reader strange scheduling
From: Stefan Strasser (strasser_at_[hidden])
Date: 2010-08-10 10:28:18

Zitat von Gaetano Mendola <mendola_at_[hidden]>:

> Hi all,
> I'm in need to have a reader/writer mutex with no starvation (I'd say a fair
> mutex), before to make a my own implementation I have checked the boost one
> to see if it makes starvation or not, and I'm a bit puzzled about the behave:
>
> I'm experiencing this:
>
> reader 1 start
>  ... writer 1 running ...
> reader 2 start
>  ... writer 2 running ...
> writer 1 start                << blocked
> reader 3 start                << blocked
> reader 4 start                << blocked
>
> until now it seems that reader 3 and reader 4 are waiting because there is a
> writer waiting and then the writer will run before reader 3 and 4,  
> unfortunately
> this is not the case, the execution continues like this:
>
> reader 1 ends
> reader 2 ends
>  ... writer 3 running ...
>  ... writer 4 running ...
> reader 3 ends
> reader 4 ends
>  ... writer 1 running ...
> writer 1 ends
>
> isn't that a waste? Why make reader 3 and 4 waiting and then make  
> those running
> before the writer? The code to reproduce that is at the end.

I don't think Boost.Thread specifies this behaviour, so it is probably  
platform-dependent.
I'm guessing readers 3 and 4 have to wait because the writer thread  
requested an exclusive lock (and otherwise the writer thread would be  
starved), but once all locks have been released it is a decision of  
the OS scheduler which thread is then woken up first and can acquire a  
lock - one of the reader threads in this case apparently.