On Tue, May 05, 2015 at 02:20:08PM +0200, Jakub Kiciński wrote: > On Mon, 4 May 2015 16:25:01 +0200, Martin Sperl wrote: > > that makes it much more complicated to add things per spi_master > > implementation (which was my original idea: have a framework that exposes > > the distinct transfer modes of the spi-bcm2835 driver (polling, interrupt > > and dma driven) and also to get some stats on how often we run into those > > “corner” cases - that is also why I want to gather the histogram > > statistics, as it may make it easier to understand the use cases of people > > who complain about “performance” in case they share those infos… > You can also consider adding few strategically placed tracepoints. > It saves the amount of kernel code necessary and is generally more > flexible. We already have a bunch of tracepoints in there (probably sufficiently detailed for what's here) but they solve a different problem to statistics - while you can extract statistics at least as far back as the tracepoints go but it's slightly cumbersome and for longer term tests the buffer will roll round at some point and it's a bit more work for people to set them up. If we can gather a few statistics without too much overhead I don't see a huge problem with doing both.
Attachment:
signature.asc
Description: Digital signature
