Re: [PATCH] printk: Fix discarding of records

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Sun, Feb 16, 2014 at 11:28:36AM -0800, Linus Torvalds wrote:
> Adding Kay and Greg, since the original code is from commit
> 7ff9554bb578 ("printk: convert byte-buffer to variable-length record
> buffer") and all the "prev" flag tweaks end up building on top of
> that.
> 
> The whole "prev flags" is messed up, and LOG_CONT is done very confusingly.
> 
> Why are *those* particular two "prev = msg->flags" incorrect, when
> every other case where we walk the messages they are required?
> 
> The code/logic makes no sense. You remove the "prev = msg->flags" at
> line 1070, when the *identical* loop just above it has it. So now the
> two loops count the number of characters differently. That makes no
> sense.
> 
> So I don't think this fixes the fundamental problem. I'm more inclined
> to believe that LOG_CONT is wrongly set somewhere, for example because
> a continuation wasn't actually originally printed due to coming from
> different users or something like that.
> 
> Or at the very least I want a coherent explanation why one loop would
> do this and the other would not, and why counting up *different*
> numbers could possibly make sense.
> 
> Because as it is, there clearly is some problem, but the patch does
> not look sensible to me.

Yeah, it doesn't make much sense to me either.

Kay had a printk() test module that would stress these types of paths
out a bunch, Kay, is that module around somewhere that we could maybe
add it to the kernel tree so it could be used to test changes like this?

thanks,

greg k-h