Re: [PATCH 1/2] ring-buffer: Introducing ring-buffer mapping functions

[Vincent Donnefort <vdonnefort@google.com>]

On Tue, Mar 21, 2023 at 11:40:47AM -0400, Steven Rostedt wrote:
> On Tue, 21 Mar 2023 15:17:15 +0000
> Vincent Donnefort <vdonnefort@google.com> wrote:
> 
> > On Mon, Mar 20, 2023 at 09:45:16PM -0400, Steven Rostedt wrote:
> > > On Fri, 17 Mar 2023 14:33:09 +0000
> > > Vincent Donnefort <vdonnefort@google.com> wrote:
> > >   
> > > > Also, the meta-page being... a single page, this limits at the moment the
> > > > number of pages in the ring-buffer that can be mapped: ~3MB on a 4K pages
> > > > system.  
> > > 
> > > I hate this limitation, so I fixed it ;-)  
> > 
> > Thanks a lot for having a look. Do you mind if I fold this in my patch for a V2?
> 
> Hold off, I found some bugs that I'm fixing ;-)
> 
> > 
> > > 
> > > I added a meta_page_size field to the meta page, and user space can do:
> > > 
> > > 	meta = mmap(NULL, page_size, PROT_READ, MAP_SHARED, fd, 0);
> > > 	if (meta == MAP_FAILED)
> > > 		pdie("mmap");
> > > 
> > > 	map = meta;
> > > 	meta_len = map->meta_page_size;
> > > 
> > > 	if (meta_len > page_size) {
> > > 		munmap(meta, page_size);
> > > 		meta = mmap(NULL, meta_len, PROT_READ, MAP_SHARED, fd, 0);
> > > 		if (meta == MAP_FAILED)
> > > 			pdie("mmap");
> > > 		map = meta;
> > > 	}
> > > 
> > > This appears to work (but I'm still testing it).
> > > 
> > > -- Steve
> > > 
> > > diff --git a/include/uapi/linux/trace_mmap.h b/include/uapi/linux/trace_mmap.h
> > > index 24bcec754a35..12f3f7ee33d9 100644
> > > --- a/include/uapi/linux/trace_mmap.h
> > > +++ b/include/uapi/linux/trace_mmap.h
> > > @@ -18,6 +18,7 @@ struct ring_buffer_meta_page {
> > >  	__u32	reader_page;
> > >  	__u32	nr_data_pages;	/* doesn't take into account the reader_page */
> > >  	__u32	data_page_head;	/* index of data_pages[] */
> > > +	__u32	meta_page_size;	/* size of the meta page */  
> > 
> > Do we want a specific field here? That could be deduced from nr_data_pages()
> > quite easily?
> 
> I rather not have too much implementation detail knowledge in user space.
> It only removes a single entry, and it makes user space easier. In fact,

Ack.

> I'm thinking we should not include "__u32 data_pages[]" but instead add a:
> "__u32 data_start" where user space does:
> 
> 	__u32 *data_pages = (_u32 *)meta_page + meta_page->data_start;
> 
> That way we could extend the data provided by the meta_page in the future.

That'd be nice. Couldn't we keep both to simplify the code for the kernel side?

> 
> -- Steve
>