Re: [RFC PATCH 1/3] Unified trace buffer

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, 24 Sep 2008, Steven Rostedt wrote:
> 
> Right now I have a list of pages that make up the ring buffer. Are you 
> saying that the first entry in the page should be a timestamp?

I think the most straightforward model would be that the "head" of the 
ring buffer (regardless of size in pages) would have that timestamp. 
Making them per-page is an option, of course, I have no strong opinions 
either way. The per-page one could have advantages (ie it would give a 
nice upper limit for just how many entries you have to walk in order to 
convert an entry into a full timestamp), but I certainly don't think 
that's a big decision, more of a detail.

But if we start out with having the full TSC in each entry, that's easily 
going to be painful to fix later. If we start out with a delta system, 
changing the details of where the base is gotten is likely to be exactly 
that - just a detail.

So I'd like the thing to have small headers, and be designed from the 
start to have small headers.

> I will now have a ring_buffer API, which will do basic recording. It will 
> have two modes when allocated. Fixed sized entry mode where you can just 
> put whatever you want in (I'm still aligning everything by 8 bytes, just 
> since memory is cheap). Or you can have variable length mode that will 
> make the following event header:
> 
> struct {
> 	unsigned char length;
> 	unsigned char buff[];
> };

So the only reason I'm not thrilled with this is that I really think that 
timestamping should be inherent, and at the lowest level.

Without timestamping, what's the real point? EVERYBODY eventually wants a 
timestamp. We added it even to the kernel printk()'s. People want them for 
network packets to user space. X wants it for all its events. It's one of 
those things that people never do from the beginning, but that everybody 
eventually wants anyway.

So I certainly don't mind layering, but I *do* mind it if it then means 
that some people will use a broken model and not have timestamps. So I 
think the timestamping code should just be there - without it, a trace 
buffer is pointless.

		Linus