On Tue 23-05-17 15:25:57, Ross Zwisler wrote: > One of the primary motivations for adding tracepoints to the DAX PMD path > was to allow the user to diagnose whether their system was actually using > PMDs, and if not to help them understand why. For me at least this has > worked okay in some situations, but many times I find myself adding more > debugging to diagnose fallback reasons that aren't immediately obvious, or > situations where the current tracepoints are simply insufficient because > they don't give you enough information. > > This series adds short fallback reason strings to the tracepoints in the > PMD path with the intention of giving the user better information about why > their system is falling back to PTEs. So I don't like adding strings into tracepoints. It just seems too close to debug printk. After all if you need detailed information for debugging some particular bug, you can always add kprobe tracepoint at particular instruction in the function. Probably my objection boils down to the fact that I don't think tracepoints in DAX code are for a normal user. They expose lot of details about the implementation and once users start depending on a particular tracepoint, it becomes part of the kernel API and cannot be changed which will be very inconvenient for DAX implementation. I think these tracepoints are for kernel developers to be able to diagnose what's going on and I'm willing to impose some thinking burden on those ;). Also looking at the reasons you've added most of them can already be detected from the information output by the tracepoint so it seems mostly like a duplicit information. So my opinion is: If you need more information like whether pfn is from devmap, make that visible in the tracepoint (like add pfn flags to the output). Sure the reason for fallback isn't immediately visible but if you really wonder, you go through conditions in the code, verify them against the data from the tracepoint and you should be able to find the reason... If there are some common fallback paths that you'd think are worth detecting and it's not easy to add information to the final tracepoint, then we can add a specific tracepoint to that branch. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
