cvs commit: src/sys/cam cam_ccb.h cam_periph.c cam_periph.h cam_sim.c cam_sim.h cam_xpt.c cam_xpt.h cam_xpt_periph.h src/sys/cam/scsi scsi_cd.c scsi_ch.c scsi_da.c scsi_low.c scsi_pass.c scsi_pt.c scsi_sa.c scsi_ses.c scsi_sg.c ...

[Scott Long]

Nate Lawson wrote:
[...]
> 
> Amazing work, I owe you a beer for following through where none of us
> did.  I see you went with the SIM lock approach, which seems reasonable.

I was surprised when I measured the contention on the SIM lock.  It was 
much lower than I expected it to be.  So this approach is looking to be
pretty good.  It can be improved a little bit in CAM, but the focus will
now be on efficient locking strategies for the drivers.  I have simple
locking implementations for MPT, AHC, and AHD right now, and I'll have
CISS and maybe SYM shortly.  I suspect that the easiest approach to an
efficient strategy will be to make the driver ithreads be a lock-free as
possible.

I'm planning on writing a few pages of design docs on this in the next 
few days to give people a guide for locking down their drivers.

Scott