Message-ID: <20090510083517.GB5794@localhost>
Date:	Sun, 10 May 2009 16:35:17 +0800
From:	Wu Fengguang <fengguang.wu@...el.com>
To:	Ingo Molnar <mingo@...e.hu>
Cc:	Frédéric Weisbecker <fweisbec@...il.com>,
	Steven Rostedt <rostedt@...dmis.org>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Li Zefan <lizf@...fujitsu.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	LKML <linux-kernel@...r.kernel.org>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>,
	Andi Kleen <andi@...stfloor.org>,
	Matt Mackall <mpm@...enic.com>,
	Alexey Dobriyan <adobriyan@...il.com>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>
Subject: Re: [patch] tracing/mm: add page frame snapshot trace

On Sat, May 09, 2009 at 10:05:12PM +0800, Ingo Molnar wrote:
> 
> * Wu Fengguang <fengguang.wu@...el.com> wrote:
> 
> > > ( End even for tasks, which are perhaps the hardest to iterate, we
> > >   can still do the /proc method of iterating up to the offset by 
> > >   counting. It wastes some time for each separate thread as it has 
> > >   to count up to its offset, but it still allows the dumping itself
> > >   to be parallelised. Or we could dump blocks of the PID hash array. 
> > >   That distributes tasks well, and can be iterated very easily with 
> > >   low/zero contention. The result will come out unordered in any 
> > >   case. )
> > 
> > For task/file based page walking, the best parallelism unit can be 
> > the task/file, instead of page segments inside them.
> > 
> > And there is the sparse file problem. There will be large holes in 
> > the address space of file and process(and even physical memory!).
> 
> If we want to iterate in the file offset space then we should use 
> the find_get_pages() trick: use the page radix tree and do gang 
> lookups in ascending order. Holes will be skipped over in a natural 
> way in the tree.

Right. I actually have code doing this, very neat trick.

> Regarding iterators, i think the best way would be to expose a 
> number of 'natural iterators' in the object collection directory. 
> The current dump_range could be changed to "pfn_index" (it's really 
> a 'physical page number' index and iterator), and we could introduce 
> a couple of other indices as well:
> 
>     /debug/tracing/objects/mm/pages/pfn_index
>     /debug/tracing/objects/mm/pages/filename_index
>     /debug/tracing/objects/mm/pages/task_index
>     /debug/tracing/objects/mm/pages/sb_index

How about 

     /debug/tracing/objects/mm/pages/walk-pfn
     /debug/tracing/objects/mm/pages/walk-file
     /debug/tracing/objects/mm/pages/walk-task

     /debug/tracing/objects/mm/pages/walk-fs
     (fs may be a more well known name than sb?)

They begin with a verb, because they are verbs when we echo some
parameters into them ;-)

> "filename_index" would take a file name (a string), and would dump 
> all pages of that inode - perhaps with an additional index/range 
> parameter as well. For example:
> 
>     echo "/home/foo/bar.txt 0 1000" > filename_index

Better to use

     "0 1000 /home/foo/bar.txt"

because there will be files named "/some/file 001".

But then echo will append an additional '\n' to filename and we are
faced with the question whether to ignore the trailing '\n'.

> Would look up that file and dump any pages in the page cache related 
> to that file, in the 0..1000 pages offset range.
> 
> ( We could support the 'batching' of such requests too, so 
>   multi-line strings can be used to request multiple files, via a 
>   single system call.

Yes, I'd expect it to make some difference in efficiency, when there
are many small files.

>   We could perhaps even support directories and do 
>   directory-and-all-child-dentries/inodes recursive lookups. )

Maybe, could do this when there comes such a need.

> Other indices/iterators would work like this:
> 
>     echo "/var" > sb_index
> 
> Would try to find the superblock associated to /var, and output all 
> pages that relate to that superblock. (it would iterate over all 
> inodes and look them all up in the pagecache and dump any matches)

Can we buffer so much outputs in kernel? Even if ftrace has no such
limitations, it may not be a good idea to pin too many pages in the
ring buffer.

I do need this feature. But it sounds like a mixture of
"files-inside-sb" walker and "pages-inside-file" walker. 
It's unclear how it will duplicate functions with the
"files object collection" to be added in:

        /debug/tracing/objects/mm/files/*

For example,

        /debug/tracing/objects/mm/files/walk-fs
        /debug/tracing/objects/mm/files/walk-dirty
        /debug/tracing/objects/mm/files/walk-global
and some filtering options, like size, cached_size, etc.

> Alternatively, we could do a reverse look up for the inode from the 
> pfn, and output that name. That would bloat the records a bit, and 
> would be more costly as well.

That sounds like "describe-pfn" and can serve as a good debugging tool.

> The 'task_index' would output based on a PID, it would find the mm 
> of that task and dump all pages associated to that mm. Offset/range 
> info would be virtual address page index based.

Right.

Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists