Re: [PATCH 1/4] mm/damon/dbgfs: Implement recording feature

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 11 Oct 2021 09:30:57 +0000 SeongJae Park <sj@kernel.org> wrote:

> Hello Andrew,
> 
> 
> Thank you for great questions!
> 
> On Sun, 10 Oct 2021 15:01:40 -0700 Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > On Fri,  8 Oct 2021 09:45:06 +0000 SeongJae Park <sj@kernel.org> wrote:
> > 
> > > The user space can get the monitoring results via the 'damon_aggregated'
> > > tracepoint event.  For simplicity and brevity, the tracepoint events
> > > have some duplicated information such as 'target_id' and 'nr_regions',
> > > though.  As a result, its size is greater than really needed.  Also,
> > > dealing with the tracepoint could be complex for some simple use cases.
> > > To provide a way for getting more efficient and simple monitoring
> > > results to user space, this commit implements 'recording' feature in
> > > 'damon-dbgfs'.
> > > 
> > > The feature is exported to the user space via a new debugfs file named
> > > 'record', which is located in '<debugfs>/damon/' directory.  The file
> > > allows users to record monitored access patterns in a regular binary
> > > file in a simple format.
> > 
> > Binary files are troublesome.
> > 
> > Is the format of this file documented anywhere?
> 
> No.  I intended the Python script in the following patch[1] and the user space
> tool[2] to be used as such documents.  I will write up one before the next
> spin.
> 
> [1] https://lore.kernel.org/linux-mm/20211008094509.16179-3-sj@kernel.org/
> [2] https://github.com/awslabs/damo/blob/v0.0.5/_damon_result.py#L38
> 
> > 
> > I assume that the file's contents will have different representations
> > depending on host endianness and word size and I further assume that
> > the provided python script won't handle this very well?
> 
> You're right.  I will make the script properly handle the cases in the next
> spin.

Well, rather than messing with the different file formats, you could
make the binary file machine-independent.  Decide on the endianness and
word size, implement them and document them.  Things like cpu_to_le32
are zero-cost on little-endian machines.