Re: clone() <-> getpid() bug in 2.6?

[Kalin KOZHUHAROV <kalin@ThinRope.net>]

Linus Torvalds wrote:
> 
> On Sun, 6 Jun 2004, Kalin KOZHUHAROV wrote:
> 
>>Well, not exactly sure about my reply, but let me try.
>>
>>The other day I was debugging some config problems with my qmail instalation and I ended up doing:
>># strace -p 4563 -f -F
>>...
>>[pid 13097] read(3, "\347\374\375TBH~\342\233\337\220\302l\220\317\237\37\25"..., 32) = 32
>>[pid 13097] close(3)                    = 0
>>[pid 13097] getpid()                    = 13097
>>[pid 13097] getpid()                    = 13097
>>[pid 13097] getuid32()                  = 89
>>[pid 13097] getpid()                    = 13097
>>[pid 13097] time(NULL)                  = 1086497450
>>[pid 13097] getpid()                    = 13097
>>[pid 13097] getpid()                    = 13097
>>[pid 13097] getpid()                    = 13097
> 
> 
> qmail is a piece of crap. The source code is completely unreadable, and it 
> seems to think that "getpid()" is a good source of random data. Don't ask 
> me why.
> 
> It literally does things like
> 
> 	random = now() + (getpid() << 16);
> 
> and since there isn't a single comment in the whole source tree, it's
> pointless to wonder why. (In case you wonder, "now()" just does a
> "time(NULL)" call - whee.).
> 
> I don't understand why people bother with it. It's not like Dan Bernstein
> is so charming that it makes up for the deficiencies of his programs.
Well, it just works once you set it up right. And many people use it and you can get community support to a certain extent.

> But no, even despite the strange usage, this isn't a performance issue.  
> qmail will call "getpid()" a few tens of times per connection because of
> the wonderful quality of randomness it provides, or something.
> 
> This is another gem you find when grepping for "getpid()" in qmail, and 
> apparently the source of most of them:
> 
> 	if (now() - when < ((60 + (getpid() & 31)) << 6))
> 
> Don't you love it how timeouts etc seem to be based on random values that 
> are calculated off the lower 5 bits of the process ID? And don't you find 
> the above (totally uncommented) line just a thing of beauty and clarity?
:-) DJB is (in)famous for its "code clarity".

> Yeah. 
> 
> Anyway, you did find something that used more than a handful of getpid() 
> calls, but no, it doesn't qualify as performance-critical, and even 
> despite it's peyote-induced (or hey, some people are just crazy on their 
> own) getpid() usage, it's not a reason to have a buggy glibc.

I definately agree that getpid() should not be cached as it gives inconsistent results.
That is why I just reported I case of "more than a handful of getpid() calls" that struck me recently.

Ok, I think I/we have no more to say about the above getpid() usage.

Is there anybody insisting on getpid() caching?
And can/will anydoby fix that in glibc?

Kalin.


-- 
||///_ o  *****************************
||//'_/>     WWW: http://ThinRope.net/
|||\/<" 
|||\\ ' 
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^