Oops, I forgot to mention that reqfn is reformatted mostly as suggested by Chritoph Hellwig. Sorry. On Tue, Apr 12, 2005 at 07:33:03PM +0900, Tejun Heo wrote: > 03_scsi_reqfn_reimplementation.patch > > This patch rewrites scsi_request_fn(). scsi_dispatch_cmd() is > merged into scsi_request_fn(). Goals are > > * Remove unnecessary operations (host_lock unlocking/locking, > recursing into scsi_run_queue(), ...) > * Consolidate defer/kill paths. > * Be concise. > > The following bugs are fixed. > > * All killed requests now get fully prep'ed and pass through > __scsi_done(). This is the only kill path. > - scsi_cmnd leak in offline-kill path removed > - unfinished request bug in > scsi_dispatch_cmd():SDEV_DEL-kill path removed. > - commands are never terminated directly from blk > layer unless they are invalid, so no need to supply > req->end_io callback for special requests. > * Timer is added while holding host_lock, after all conditions > are checked and serial number is assigned. This guarantees > that until host_lock is released, the scsi_cmnd pointed to > by cmd isn't released. That didn't hold in the original > code and, theoretically, the original code could access > already released cmd. > * For the same reason, if shost->hostt->queuecommand() fails, > we try to delete the timer before releasing host_lock. > > Other changes/notes > > * host_lock is acquired and released only once. > enter (qlocked) -> enter loop -> dev-prep -> switch to hlock -\ > ^---- switch to qlock <- issue <- host-prep <-/ > * unnecessary if () on get_device() removed. > * loop on elv_next_request() instead of blk_queue_plugged(). > We now explicitly break out of loop when we plug and check if > the queue has been plugged underneath us at the end of loop. > * All device/host state checks are done in this function and > done only once while holding qlock/host_lock respectively. > * Requests which get deferred during dev-prep are never > removed from request queue, so deferring is achieved by > simply breaking out of the loop and returning. > * Failure of blk_queue_start_tag() on tagged queue is a BUG > now. This condition should have been catched by > scsi_dev_queue_ready(). > * req->special == NULL test removed. This just can't happen, > and even if it ever happens, scsi_request_fn() will > deterministically oops. > * Requests which gets deferred during host-prep are requeued > using blk_requeue_request(). This is the only requeue path. > > Note that scsi_kill_requests() still terminates requests using > blk layer. The path is circular-ref workaround and soon to be > replaced, so ignore it for now. -- tejun - : 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
