Adding new option to ktrace

[Robert Watson]

On Tue, 6 Sep 2005, Peter Jeremy wrote:

> On Tue, 2005-Sep-06 10:33:53 +0530, Nikhil Dharashivkar wrote:
>>     Thanks for replying me. Basically what happend, while testing
>> scsi driver on freebsd, at  some point it crashes. So, there is no way
>> to know how much IO is performed. To know the IO state just before the
>> driver fails, i selected ktrace to print IO information whatever i ll
>> get from dastrategy routine.
>
> It's not clear how ktrace is going to help here.  The ktrXXX(9) 
> functions place ktr_request events in a queue.  A kernel thread then 
> dumps the queue entries into a file via the normal buffer cache.  The 
> data on disk is typically about 30 seconds behind real time.  If the 
> system crashes, you will lose any events that are still in the buffer 
> cache or ktr_todo queue.
>
> Another problem is that since ktrace generates disk I/O, it is likely to 
> disturb your testing.
>
> A better approach would seem to be to build a circular buffer and store 
> the I/O requests in the buffer.  When the system crashes, you can look 
> at the last entries in the buffer.

ktrace is indeed probably the wrong layer to do this analysis, if it is 
firmly believed to be a SCSI layer problem, since process level I/O events 
often map weakly to device level I/O events -- i.e., directory operations 
are substantially divorced from the block operations that implement them 
due to soft updates, write buffering, write combining, etc.

KTR(9) supports tracing of GEOM-layer I/O events to an in-memory buffer 
which can be retrieved from a coredump using ktrdump.  Adding additional 
instrumentation, say in CAM or a device driver, is pretty straight forward 
and probably a useful way to approach this problem.

Robert N M Watson