Re: [PATCH][v2] blktrace: conversion to tracepoints

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Oct 29 2008, Arnaldo Carvalho de Melo wrote:
> Hi Jens,
> 
> 	Now that the tracepoints infrastructure is merged I updated the
> patch, please take a look.
> 
> 	One suggestion I got was to have things like:
> 
>   trace_block_unplug_io(q, q->rq.count[READ] + q->rq.count[WRITE]);
> 
> 	That was:
> 
>   blk_add_trace_pdu_int(q, BLK_TA_UNPLUG_IO, NULL,
> 	                q->rq.count[READ] + q->rq.count[WRITE]);
> 
> 	To be:
> 
>   trace_block_unplug_io(q, q->rq.count[READ], q->rq.count[WRITE]);
> 
> 	Or even:
> 
>   trace_block_unplug_io(q);
> 
> 	And on  blk_add_trace_unplug_io tracepoint do the math and feed
> it to __blk_add_trace.
> 
>         So that the information on the number of types of requests
> instead of the sum, what do you think? Overengineering? For blktrace it
> would end up being preserved as is in, say:
> 
> static void blk_add_trace_unplug_io(struct request_queue *q,
> 				    unsigned int rd, unsigned int wr)
> {
>         struct blk_trace *bt = q->blk_trace;
> 
>         if (bt) {
>                 __be64 rpdu = cpu_to_be64(rd + wr);
> 
>                 __blk_add_trace(bt, 0, 0, 0, BLK_TA_UNPLUG_IO, 0,
>                                 sizeof(rpdu), &rpdu);
>         }
> }
> 
> 	Perhaps doing it as 'trace_block_unplug_io(q)' would be the best
> scenario, as the tracepoint user can look at struct_request queue at
> will anyway and the code gets cleaner :-)
> 
> 	Feel free to point any disgusting aspect, perhaps there is at
> least one to warn me about fixing 8-)

You my as well pass the members separately now that it's a specific call
anyway, to avoid doing the calculation when tracing is disabled.

Patch looks straight forward. Perhaps it would be cleaner to use an
atomic type for the reference?

> @@ -237,6 +243,10 @@ static void blk_trace_cleanup(struct blk_trace *bt)
>  	free_percpu(bt->sequence);
>  	free_percpu(bt->msg_data);
>  	kfree(bt);
> +	mutex_lock(&blk_probe_mutex);
> +	if (--blk_probes_ref == 0)
> +		blk_unregister_tracepoints();
> +	mutex_unlock(&blk_probe_mutex);
>  }

Then this would be

        if (atomic_dec_and_test(&blk_probes_ref))
                blk_unregister_tracepoints();

>  int blk_trace_remove(struct request_queue *q)
> @@ -428,6 +438,14 @@ int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
>  	bt->pid = buts->pid;
>  	bt->trace_state = Blktrace_setup;
>  
> +	mutex_lock(&blk_probe_mutex);
> +	if (!blk_probes_ref++) {
> +		ret = blk_register_tracepoints();
> +		if (ret)
> +			goto probe_err;
> +	}
> +	mutex_unlock(&blk_probe_mutex);
> +

And this would be

        if (atomic_add_return(&blk_probes_ref, 1) == 1) {
		ret = blk_register_tracepoints();
		if (ret)
			goto probe_err;
        }

-- 
Jens Axboe