[boost] Re: [shared_ptr] PPC implementations updated

From: Alexander Terekhov (terekhov_at_[hidden])
Date: 2005-04-08 11:17:11

Alexander Terekhov wrote:
[...]
> That (preliminary) notice doesn't really say what's the problem
> with respect to "dangling" reservation. (Context switch of course
> must clear reservation; everybody and his dog knows that.)

Just to clarify. Think of simple unconditional increments on UP
(count is zero initially).

   Thread A: <load-reserved>
             <add 1>

   [preemptive context switch]

   Thread B: <load-reserved>
             <add 1>
             <store-conditional> // count = 1
             .
             .
             [another increment, count = 2]
             .
             .
             .
             [begin yet another increment]
             <load-reserved>

   [preemptive context switch]

   Thread A: <store-conditional> // count = 1

And you've just lost one billion dollars.

In any preemptive OS context switch must always clear reservation 
to ensure that you never get into this sort of trouble.

regards,
alexander.