On Wed, Mar 23, 2005 at 06:13:26PM +0900, Tejun Heo wrote: > Hi, > > James Bottomley wrote: > >On Wed, 2005-03-23 at 11:14 +0900, Tejun Heo wrote: > > > >> So, basically, SCSI high-level object (scsi_disk) and > >> mid-level object (scsi_device) are reference counted by users, > >> not the requests they submit. Reference count cannot go zero > >> with active users and users cannot access the object once the > >> reference count reaches zero. > > > > > >Actually, no. Unfortunately we still have some fire and forget APIs, so > >the contention that we always have an open refcounted descriptor isn't > >always true.
What API's, and what usage? > Yeap, you're right. So, what we have is > * All high-level users have open access to the scsi high-level > object on issueing requests, but may close it before its requests > complete. > * All mid-layer users do get_device() before submitting requests, > but may put_device() before its requests complete. Any LLDD's issuing requests should be doing a get/put around the request. Any upper level drivers calling scsi_device_put() before a request completes is likely a bug. sg has code in place to handle the post-release/close completion of IO (IMO, a bad design). Are any upper level drivers calling scsi_device_put() while they have outstanding IO? The scan code never calls upper level drivers probe functions via device_add unless we are going to keep the scsi_device (well, there are error paths in scsi_sysfs_add_sdev that look bad - we don't check the result of scsi_sysfs_add_sdev). But for completeness, we could add get/puts to the scan code. As you pointed out, the current get_device() will never return NULL when called via: get_device(&sdev->sdev_gendev) The current code only narrows the window where problems might occur, I don't see how it can completely avoid races with removal. And the patch removes code from the mainline scsi IO paths. -- Patrick Mansfield - To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html