On Wed, 2005-09-14 at 17:33 -0400, Alan Stern wrote: > On Wed, 14 Sep 2005, James Bottomley wrote: > > Yes ... really the only case for unprep is when we've partially released > > the command (like in scsi_io_completion) where we need to tear the rest > > of it down. > > In other words, in scsi_requeue_command and nowhere else. Pretty much, yes. > Or will be prepared for the first time, as in scsi_execute. As far as I > can tell, a new struct request is not set to all 0's. So if you queue a > request with REQ_SPECIAL set and you fail to clear req->special, you're in > trouble. Do you have any idea why this hasn't been causing errors all > along? That's true, it's not. However ll_rq_blk.c:rq_init() clears req- >special (and initialises all other important fields). > Okay, then how does this patch look (moved the routine over to where it > gets used, plus two real changes)? Well ... under pressure to fix this in -mm, I already commited a version to rc-fixes. What I did was fully reverse the changes to the scsi_insert_queue() [the patch I sent Anton]. We can move the unprep function if you feel strongly about it, but I'm also happy to keep it where it is. James - : send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
