Re: Valgrind updates

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Jan 29, 2009, 14:22 UTC
Message-ID: <alpine.DEB.1.00.0901291514240.3586@pacific.mpi-cbg.de>
In-Reply-To: <alpine.LFD.2.00.0901281751580.3123@localhost.localdomain>

Hi,

On Wed, 28 Jan 2009, Linus Torvalds wrote:

Show 5 quoted lines

> On Tue, 27 Jan 2009, Johannes Schindelin wrote:
> > 
> > To help ye Gods, I put together this almost minimal C program:
> 
> This one is buggy.

Not exactly buggy.  Underexplained.

Show 8 quoted lines

> > 	out = fopen("/dev/null", "w");
> > 	fwrite(compressed + 51, 51, 1, out);
> > 	fwrite(compressed + 51, 1, 1, stderr);
> > 	fflush(out);
> > 	fclose(out);
> 
> The problem is that the first argument to that first "fwrite()" is simply 
> wrong. It shouldn't be "compressed + 51", it should be just "compressed".

Nope. It should be "compressed + 51" to narrow down the issue, as valgrind does not complain about _any other_ offset.

Not even when that is _well_ after the 58 bytes deflate() says are available.

> As it is, you're writing 51 bytes, starting at 51 bytes in, and that's 
> obviously not correct (you only got 58 bytes from deflate()).

It is not, granted. But I left it in for a purpose: to show that valgrind does not even bother to mention bytes we think should be invalid.

I thought that there might be a shortcut for /dev/null, so I changed the outfile to a real file, and it _still_ does not complain.

> So valgrind does complain about it, but for a perfectly valid reason.

Only it does not. It complains about the write of 1 byte, not the write of 51.

But I know why: "out" is opened buffered, so it shows the error (well delayed, I might add, and not in a helpful manner) when fflush() is called.

The real issue, namely that an access of offset 51 triggers a valgrind error, is demonstrated by my small test case.

Ciao, Dscho