On Tue, Mar 28, 2023 at 10:44:11PM -0400, Steven Rostedt wrote:

On Wed, 22 Mar 2023 10:22:43 +0000
Vincent Donnefort [off-list ref] wrote:

quoted

+#include <linux/types.h>
+
+struct ring_buffer_meta_page_header {
+#if __BITS_PER_LONG == 64
+	__u64	entries;
+	__u64	overrun;
+#else
+	__u32	entries;
+	__u32	overrun;
+#endif
+	__u32	pages_touched;
+	__u32	meta_page_size;
+	__u32	reader_page;	/* page ID for the reader page */
+	__u32	nr_data_pages;	/* doesn't take into account the reader_page */
+	__u32	data_page_head;	/* ring-buffer head as an offset from data_start */
+	__u32	data_start;	/* offset within the meta page */
+};
+

I've been playing with this a bit, and I'm thinking, do we need the
data_pages[] array on the meta page?

I noticed that I'm not even using it.

Currently, we need to do a ioctl every time we finish with the reader page,
and that updates the reader_page in the meta data to point to the next page
to read. When do we need to look at the data_start section?

This is for non-consuming read, to get all the pages in order.

If we remove this section we would lose this ability ... but we'd also simplify
the code by a good order of magnitude (don't need the update ioctl anymore, no
need to keep those pages in order and everything can fit a 0-order meta-page).
And the non-consuming read doesn't bring much to the user over the pipe version.

This will although impact our hypervisor tracing which will only be able to
expose trace_pipe interfaces. But I don't think it is a problem, all userspace
tools only relying on consuming read anyway.

So if you're happy dropping this support, let's get rid of it.

-- 
Vincent