Keith Bostic wrote:
>>I've been following the thread titled 'Subversion's use of Berkeley DB
>>[#11511]', and the problem seems to be: We have no easy way to
>>determine if we need to perform recovery, and the only solution we have
>>so far is to run recovery every time.
>>
>>
>
>I agree, running recovery every time isn't the right solution.
>
>Your design is pretty much where I ended up as well.
>
The proposal is a classical case of using a mutex to protect a reference
count, so of course it will work.
> A couple
>of differences: I used fcntl instead of flock, and built the
>file more like a registry, with process IDs of the processes
>running in the database environment. I also used an individual
>lock for each process, to detect if a process has died in the
>database environment.
>
>
Could you elaborate a bit? From this description, I can't see how this
will work; I see a number of potential new races, so I'm hoping I
misunderstood. :-)
>>I have mapped out a potential locking procedure that would make it
>>possible to detect if any process exited uncleanly, and determine if
>>recovery needs to be run on the database. It should be implementable
>>using POSIX flock() (meaning that it should work on Win32). It should
>>also be implementable using fcntl locking, which is NFS-friendly, but
>>Win32-hostile.
>>
>>
>
>POSIX 1003.1 requires fcntl, doesn't Win32 support it?
>
>
It does, but Windows also gives you a number of other useful tools. I
don't think this locking scheme can be implemented completely portably.
>I think I see problems with your design, but I can't say for
>sure.
>
>There were places I didn't follow what you were trying to say,
>so I may just be wrong -- specifically, there were two sections
>labelled "startup", and they were different. It wasn't clear
>to me how they were related?
>
>More specifically, I was concerned about the transition from the
>exclusive lock to the shared lock. It seems to me that there
>was a race in there where a process could end up in the database
>environment without the reference set to '1'.
>
>
This worries me, too. I think you'd need a separate file for the shared
lock, and be very careful about the locking order.
>I'm happy to work through this in detail with you, but we might
>as well take it offline.
>
>
I wish you hadn't... or at least, bring me into the loop, please?
-- Brane
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Thu Dec 23 14:31:05 2004