Re: [PATCH 1/2] mpt2sas: Refcount sas_device objects and fix unsafe list usage

[Calvin Owens <calvinowens@fb.com>]

On Monday 07/13 at 11:05 -0400, Joe Lawrence wrote:
> On 07/12/2015 12:24 AM, Calvin Owens wrote:
> > These objects can be referenced concurrently throughout the driver, we
> > need a way to make sure threads can't delete them out from under each
> > other. This patch adds the refcount, and refactors the code to use it.
> > 
> > Additionally, we cannot iterate over the sas_device_list without
> > holding the lock, or we risk corrupting random memory if items are
> > added or deleted as we iterate. This patch refactors _scsih_probe_sas()
> > to use the sas_device_list in a safe way.
> > 
> > Cc: Christoph Hellwig <hch@infradead.org>
> > Cc: Bart Van Assche <bart.vanassche@sandisk.com>
> > Signed-off-by: Calvin Owens <calvinowens@fb.com>
> > ---
> >  drivers/scsi/mpt2sas/mpt2sas_base.h      |  22 +-
> >  drivers/scsi/mpt2sas/mpt2sas_scsih.c     | 434 ++++++++++++++++++++-----------
> >  drivers/scsi/mpt2sas/mpt2sas_transport.c |  12 +-
> >  3 files changed, 315 insertions(+), 153 deletions(-)
> 
> [ ... snip ... ]
> 
> > @@ -2078,7 +2150,7 @@ _scsih_slave_configure(struct scsi_device *sdev)
> >  	}
> >  
> >  	spin_lock_irqsave(&ioc->sas_device_lock, flags);
> > -	sas_device = mpt2sas_scsih_sas_device_find_by_sas_address(ioc,
> > +	sas_device = __mpt2sas_get_sdev_by_addr(ioc,
> >  	   sas_device_priv_data->sas_target->sas_address);
> >  	if (!sas_device) {
> >  		spin_unlock_irqrestore(&ioc->sas_device_lock, flags);
> > @@ -2116,13 +2188,14 @@ _scsih_slave_configure(struct scsi_device *sdev)
> >  	if (!ssp_target)
> >  		_scsih_display_sata_capabilities(ioc, handle, sdev);
> >  
> > -
> >  	_scsih_change_queue_depth(sdev, qdepth);
> >  
> >  	if (ssp_target) {
> >  		sas_read_port_mode_page(sdev);
> >  		_scsih_enable_tlr(ioc, sdev);
> >  	}
> > +
> > +	sas_device_put(sas_device);
> >  	return 0;
> >  }
> 
> Hi Calvin,
> 
> Any reason why this sas_device_put is placed outside the sas_device
> lock?  Most other instances in this patch were called just before unlocking.

Thanks for looking at this.

I guess I thought that something below where we drop the sas_device_lock
referenced it, but it looks like nothing does. I'll move it up in v3.

I don't think it's strictly necessary that the put() happen under the
lock: the only way this could be the final put() is if both ->hostdata
and the sas_device_list had dropped their references, and in that case
it would be impossible to have a concurrent get(), since those are the
only two ways to lookup/get a sas_device. But absent any reason not to,
let's make it more consistent.

I'm really glad you pointed this out, because I realized I flubbed this
in _scsih_target_alloc() and forgot to eliminate the sas_device_put()
from before the ->hostdata lookup was added. I'll fix this in v3.

> BTW I attempted testing, but needed to port to mpt3 and ended up with a
> driver that didn't boot :(   Hopefully I can retry later this week, or
> find an older mpt2 box lying around.

More testing would be fantastic if that's possible :)

Thanks very much,
Calvin

> -- Joe