[ltt-dev] [patch 9/9] LTTng instrumentation - swap

[mingo@elte.hu (Ingo Molnar)]

* Mathieu Desnoyers <mathieu.desnoyers at polymtl.ca> wrote:

> +DECLARE_TRACE(swap_in,
> +	TPPROTO(struct page *page, swp_entry_t entry),
> +		TPARGS(page, entry));
> +DECLARE_TRACE(swap_out,
> +	TPPROTO(struct page *page),
> +		TPARGS(page));
> +DECLARE_TRACE(swap_file_open,
> +	TPPROTO(struct file *file, char *filename),
> +		TPARGS(file, filename));
> +DECLARE_TRACE(swap_file_close,
> +	TPPROTO(struct file *file),
> +		TPARGS(file));

These are more complete than the pagecache tracepoints, but still 
incomplete to make a comprehensive picture about swap activities.

Firstly, the swap_file_open/close events seem quite pointless. Most 
systems enable swap during bootup and never close it. These 
tracepoints just wont be excercised in practice.

Also, to _really_ help with debugging VM pressure problems, the 
whole LRU state-machine should be instrumented, and linked up with 
pagecache instrumentation via page frame numbers and (inode,offset) 
[file] and (pgd,addr) [anon] pairs.

Not just the fact that something got swapped out is interesting, but 
also the whole decision chain that leads up to it. The lifetime of a 
page how it jumps between the various stages of eviction and LRU 
scores.

a minor nit:

> +DECLARE_TRACE(swap_file_open,
> +	TPPROTO(struct file *file, char *filename),
> +		TPARGS(file, filename));

there's no need to pass in the filename - it can be deducted in the 
probe from struct file.

a small inconsistency:

> +DECLARE_TRACE(swap_in,
> +	TPPROTO(struct page *page, swp_entry_t entry),
> +		TPARGS(page, entry));
> +DECLARE_TRACE(swap_out,
> +	TPPROTO(struct page *page),
> +		TPARGS(page));

you pass in swp_entry to trace_swap_in(), which encodes the offset - 
but that parameter is not needed, the page already represents the 
offset at that stage in do_swap_page(). (the actual data is not read 
in yet from swap, but the page is already linked up in the 
swap-cache and has the offset available - which a probe can 
recover.)

So this suffices:

 DECLARE_TRACE(swap_in,
	TPPROTO(struct page *page),
		TPARGS(page));

 DECLARE_TRACE(swap_out,
	TPPROTO(struct page *page),
		TPARGS(page));

And here again i'd like to see actual meaningful probe contents via 
a TRACE_EVENT() construct. That shows and proves that it's all part 
of a comprehensive framework, and the data that is recovered is 
understood and put into a coherent whole - upstream. That makes it 
immediately useful to the built-in tracers, and will also cause 
fewer surprises downstream.

	Ingo