X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=drivers%2Fblock%2Fub.c;h=2098eff91e14721641f81c47b3c510481fbb6bcc;hb=refs%2Fheads%2Fvserver;hp=acea9f08ea4475cc305445f64f05f64f542fd3e0;hpb=6a77f38946aaee1cd85eeec6cf4229b204c15071;p=linux-2.6.git

diff --git a/drivers/block/ub.c b/drivers/block/ub.c
index acea9f08e..2098eff91 100644
--- a/drivers/block/ub.c
+++ b/drivers/block/ub.c
@@ -8,49 +8,106 @@
  * and is not licensed separately. See file COPYING for details.
  *
  * TODO (sorted by decreasing priority)
- *  -- Do resets with usb_device_reset (needs a thread context, use khubd)
  *  -- set readonly flag for CDs, set removable flag for CF readers
  *  -- do inquiry and verify we got a disk and not a tape (for LUN mismatch)
- *  -- support pphaneuf's SDDR-75 with two LUNs (also broken capacity...)
- *  -- special case some senses, e.g. 3a/0 -> no media present, reduce retries
  *  -- verify the 13 conditions and do bulk resets
- *  -- normal pool of commands instead of cmdv[]?
- *  -- kill last_pipe and simply do two-state clearing on both pipes
- *  -- verify protocol (bulk) from USB descriptors (maybe...)
- *  -- highmem and sg
+ *  -- highmem
  *  -- move top_sense and work_bcs into separate allocations (if they survive)
  *     for cache purists and esoteric architectures.
+ *  -- Allocate structure for LUN 0 before the first ub_sync_tur, avoid NULL. ?
  *  -- prune comments, they are too volumnous
- *  -- Exterminate P3 printks
  *  -- Resove XXX's
- *  -- Redo "benh's retries", perhaps have spin-up code to handle them. V:D=?
+ *  -- CLEAR, CLR2STS, CLRRS seem to be ripe for refactoring.
  */
 #include <linux/kernel.h>
 #include <linux/module.h>
 #include <linux/usb.h>
+#include <linux/usb_usual.h>
 #include <linux/blkdev.h>
-#include <linux/devfs_fs_kernel.h>
 #include <linux/timer.h>
 #include <scsi/scsi.h>
 
 #define DRV_NAME "ub"
-#define DEVFS_NAME DRV_NAME
 
 #define UB_MAJOR 180
 
 /*
- * Definitions which have to be scattered once we understand the layout better.
+ * The command state machine is the key model for understanding of this driver.
+ *
+ * The general rule is that all transitions are done towards the bottom
+ * of the diagram, thus preventing any loops.
+ *
+ * An exception to that is how the STAT state is handled. A counter allows it
+ * to be re-entered along the path marked with [C].
+ *
+ *       +--------+
+ *       ! INIT   !
+ *       +--------+
+ *           !
+ *        ub_scsi_cmd_start fails ->--------------------------------------\
+ *           !                                                            !
+ *           V                                                            !
+ *       +--------+                                                       !
+ *       ! CMD    !                                                       !
+ *       +--------+                                                       !
+ *           !                                            +--------+      !
+ *         was -EPIPE -->-------------------------------->! CLEAR  !      !
+ *           !                                            +--------+      !
+ *           !                                                !           !
+ *         was error -->------------------------------------- ! --------->\
+ *           !                                                !           !
+ *  /--<-- cmd->dir == NONE ?                                 !           !
+ *  !        !                                                !           !
+ *  !        V                                                !           !
+ *  !    +--------+                                           !           !
+ *  !    ! DATA   !                                           !           !
+ *  !    +--------+                                           !           !
+ *  !        !                           +---------+          !           !
+ *  !      was -EPIPE -->--------------->! CLR2STS !          !           !
+ *  !        !                           +---------+          !           !
+ *  !        !                                !               !           !
+ *  !        !                              was error -->---- ! --------->\
+ *  !      was error -->--------------------- ! ------------- ! --------->\
+ *  !        !                                !               !           !
+ *  !        V                                !               !           !
+ *  \--->+--------+                           !               !           !
+ *       ! STAT   !<--------------------------/               !           !
+ *  /--->+--------+                                           !           !
+ *  !        !                                                !           !
+ * [C]     was -EPIPE -->-----------\                         !           !
+ *  !        !                      !                         !           !
+ *  +<---- len == 0                 !                         !           !
+ *  !        !                      !                         !           !
+ *  !      was error -->--------------------------------------!---------->\
+ *  !        !                      !                         !           !
+ *  +<---- bad CSW                  !                         !           !
+ *  +<---- bad tag                  !                         !           !
+ *  !        !                      V                         !           !
+ *  !        !                 +--------+                     !           !
+ *  !        !                 ! CLRRS  !                     !           !
+ *  !        !                 +--------+                     !           !
+ *  !        !                      !                         !           !
+ *  \------- ! --------------------[C]--------\               !           !
+ *           !                                !               !           !
+ *         cmd->error---\                +--------+           !           !
+ *           !          +--------------->! SENSE  !<----------/           !
+ *         STAT_FAIL----/                +--------+                       !
+ *           !                                !                           V
+ *           !                                V                      +--------+
+ *           \--------------------------------\--------------------->! DONE   !
+ *                                                                   +--------+
  */
 
-/* Transport (despite PR in the name) */
-#define US_PR_BULK	0x50		/* bulk only */
-
-/* Protocol */
-#define US_SC_SCSI	0x06		/* Transparent */
+/*
+ * This many LUNs per USB device.
+ * Every one of them takes a host, see UB_MAX_HOSTS.
+ */
+#define UB_MAX_LUNS   9
 
 /*
  */
-#define UB_MINORS_PER_MAJOR	8
+
+#define UB_PARTS_PER_LUN      8
 
 #define UB_MAX_CDB_SIZE      16		/* Corresponds to Bulk */
 
@@ -65,7 +122,7 @@ struct bulk_cb_wrap {
 	u32	Tag;			/* unique per command id */
 	__le32	DataTransferLength;	/* size of data */
 	u8	Flags;			/* direction in bit 0 */
-	u8	Lun;			/* LUN normally 0 */
+	u8	Lun;			/* LUN */
 	u8	Length;			/* of of the CDB */
 	u8	CDB[UB_MAX_CDB_SIZE];	/* max command */
 };
@@ -85,8 +142,6 @@ struct bulk_cs_wrap {
 
 #define US_BULK_CS_WRAP_LEN	13
 #define US_BULK_CS_SIGN		0x53425355	/* spells out 'USBS' */
-/* This is for Olympus Camedia digital cameras */
-#define US_BULK_CS_OLYMPUS_SIGN	0x55425355	/* spells out 'USBU' */
 #define US_BULK_STAT_OK		0
 #define US_BULK_STAT_FAIL	1
 #define US_BULK_STAT_PHASE	2
@@ -99,7 +154,7 @@ struct bulk_cs_wrap {
  */
 struct ub_dev;
 
-#define UB_MAX_REQ_SG	1
+#define UB_MAX_REQ_SG	9	/* cdrecord requires 32KB and maybe a header */
 #define UB_MAX_SECTORS 64
 
 /*
@@ -129,27 +184,16 @@ enum ub_scsi_cmd_state {
 	UB_CMDST_CLR2STS,		/* Clearing before requesting status */
 	UB_CMDST_STAT,			/* Status phase */
 	UB_CMDST_CLEAR,			/* Clearing a stall (halt, actually) */
+	UB_CMDST_CLRRS,			/* Clearing before retrying status */
 	UB_CMDST_SENSE,			/* Sending Request Sense */
 	UB_CMDST_DONE			/* Final state */
 };
 
-static char *ub_scsi_cmd_stname[] = {
-	".  ",
-	"Cmd",
-	"dat",
-	"c2s",
-	"sts",
-	"clr",
-	"Sen",
-	"fin"
-};
-
 struct ub_scsi_cmd {
 	unsigned char cdb[UB_MAX_CDB_SIZE];
 	unsigned char cdb_len;
 
 	unsigned char dir;		/* 0 - none, 1 - read, 3 - write. */
-	unsigned char trace_index;
 	enum ub_scsi_cmd_state state;
 	unsigned int tag;
 	struct ub_scsi_cmd *next;
@@ -160,18 +204,23 @@ struct ub_scsi_cmd {
 
 	int stat_count;			/* Retries getting status. */
 
-	/*
-	 * We do not support transfers from highmem pages
-	 * because the underlying USB framework does not do what we need.
-	 */
-	char *data;			/* Requested buffer */
 	unsigned int len;		/* Requested length */
-	// struct scatterlist sgv[UB_MAX_REQ_SG];
+	unsigned int current_sg;
+	unsigned int nsg;		/* sgv[nsg] */
+	struct scatterlist sgv[UB_MAX_REQ_SG];
 
+	struct ub_lun *lun;
 	void (*done)(struct ub_dev *, struct ub_scsi_cmd *);
 	void *back;
 };
 
+struct ub_request {
+	struct request *rq;
+	unsigned int current_try;
+	unsigned int nsg;		/* sgv[nsg] */
+	struct scatterlist sgv[UB_MAX_REQ_SG];
+};
+
 /*
  */
 struct ub_capacity {
@@ -180,28 +229,6 @@ struct ub_capacity {
 	unsigned int bshift;		/* Shift between 512 and hard sects */
 };
 
-/*
- * The SCSI command tracing structure.
- */
-
-#define SCMD_ST_HIST_SZ   8
-#define SCMD_TRACE_SZ    63		/* Less than 4KB of 61-byte lines */
-
-struct ub_scsi_cmd_trace {
-	int hcur;
-	unsigned int tag;
-	unsigned int req_size, act_size;
-	unsigned char op;
-	unsigned char dir;
-	unsigned char key, asc, ascq;
-	char st_hst[SCMD_ST_HIST_SZ];	
-};
-
-struct ub_scsi_trace {
-	int cur;
-	struct ub_scsi_cmd_trace vec[SCMD_TRACE_SZ];
-};
-
 /*
  * This is a direct take-off from linux/include/completion.h
  * The difference is that I do not wait on this thing, just poll.
@@ -252,25 +279,49 @@ struct ub_scsi_cmd_queue {
 };
 
 /*
- * The UB device instance.
+ * The block device instance (one per LUN).
+ */
+struct ub_lun {
+	struct ub_dev *udev;
+	struct list_head link;
+	struct gendisk *disk;
+	int id;				/* Host index */
+	int num;			/* LUN number */
+	char name[16];
+
+	int changed;			/* Media was changed */
+	int removable;
+	int readonly;
+
+	struct ub_request urq;
+
+	/* Use Ingo's mempool if or when we have more than one command. */
+	/*
+	 * Currently we never need more than one command for the whole device.
+	 * However, giving every LUN a command is a cheap and automatic way
+	 * to enforce fairness between them.
+	 */
+	int cmda[1];
+	struct ub_scsi_cmd cmdv[1];
+
+	struct ub_capacity capacity; 
+};
+
+/*
+ * The USB device instance.
  */
 struct ub_dev {
-	spinlock_t lock;
-	int id;				/* Number among ub's */
+	spinlock_t *lock;
 	atomic_t poison;		/* The USB device is disconnected */
 	int openc;			/* protected by ub_lock! */
 					/* kref is too implicit for our taste */
+	int reset;			/* Reset is running */
 	unsigned int tagcnt;
-	int changed;			/* Media was changed */
-	int removable;
-	int readonly;
-	int first_open;			/* Kludge. See ub_bd_open. */
-	char name[8];
+	char name[12];
 	struct usb_device *dev;
 	struct usb_interface *intf;
 
-	struct ub_capacity capacity; 
-	struct gendisk *disk;
+	struct list_head luns;
 
 	unsigned int send_bulk_pipe;	/* cached pipe values */
 	unsigned int recv_bulk_pipe;
@@ -279,10 +330,6 @@ struct ub_dev {
 
 	struct tasklet_struct tasklet;
 
-	/* XXX Use Ingo's mempool (once we have more than one) */
-	int cmda[1];
-	struct ub_scsi_cmd cmdv[1];
-
 	struct ub_scsi_cmd_queue cmd_queue;
 	struct ub_scsi_cmd top_rqs_cmd;	/* REQUEST SENSE */
 	unsigned char top_sense[UB_SENSE_SIZE];
@@ -291,46 +338,66 @@ struct ub_dev {
 	struct urb work_urb;
 	struct timer_list work_timer;
 	int last_pipe;			/* What might need clearing */
+	__le32 signature;		/* Learned signature */
 	struct bulk_cb_wrap work_bcb;
 	struct bulk_cs_wrap work_bcs;
 	struct usb_ctrlrequest work_cr;
 
-	struct ub_scsi_trace tr;
+	struct work_struct reset_work;
+	wait_queue_head_t reset_wait;
+
+	int sg_stat[6];
 };
 
 /*
  */
-static int ub_bd_rq_fn_1(struct ub_dev *sc, struct request *rq);
-static int ub_cmd_build_block(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
-    struct request *rq);
-static int ub_cmd_build_packet(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
-    struct request *rq);
+static void ub_cleanup(struct ub_dev *sc);
+static int ub_request_fn_1(struct ub_lun *lun, struct request *rq);
+static void ub_cmd_build_block(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_scsi_cmd *cmd, struct ub_request *urq);
+static void ub_cmd_build_packet(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_scsi_cmd *cmd, struct ub_request *urq);
 static void ub_rw_cmd_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
-static void ub_end_rq(struct request *rq, int uptodate);
+static void ub_end_rq(struct request *rq, unsigned int status);
+static int ub_rw_cmd_retry(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_request *urq, struct ub_scsi_cmd *cmd);
 static int ub_submit_scsi(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
-static void ub_urb_complete(struct urb *urb, struct pt_regs *pt);
+static void ub_urb_complete(struct urb *urb);
 static void ub_scsi_action(unsigned long _dev);
 static void ub_scsi_dispatch(struct ub_dev *sc);
 static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
+static void ub_data_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
 static void ub_state_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd, int rc);
-static void __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
+static int __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
 static void ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
+static void ub_state_stat_counted(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
 static void ub_state_sense(struct ub_dev *sc, struct ub_scsi_cmd *cmd);
 static int ub_submit_clear_stall(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
     int stalled_pipe);
 static void ub_top_sense_done(struct ub_dev *sc, struct ub_scsi_cmd *scmd);
-static int ub_sync_tur(struct ub_dev *sc);
-static int ub_sync_read_cap(struct ub_dev *sc, struct ub_capacity *ret);
+static void ub_reset_enter(struct ub_dev *sc, int try);
+static void ub_reset_task(struct work_struct *work);
+static int ub_sync_tur(struct ub_dev *sc, struct ub_lun *lun);
+static int ub_sync_read_cap(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_capacity *ret);
+static int ub_sync_reset(struct ub_dev *sc);
+static int ub_probe_clear_stall(struct ub_dev *sc, int stalled_pipe);
+static int ub_probe_lun(struct ub_dev *sc, int lnum);
 
 /*
  */
+#ifdef CONFIG_USB_LIBUSUAL
+
+#define ub_usb_ids  storage_usb_ids
+#else
+
 static struct usb_device_id ub_usb_ids[] = {
-	// { USB_DEVICE_VER(0x0781, 0x0002, 0x0009, 0x0009) },	/* SDDR-31 */
 	{ USB_INTERFACE_INFO(USB_CLASS_MASS_STORAGE, US_SC_SCSI, US_PR_BULK) },
 	{ }
 };
 
 MODULE_DEVICE_TABLE(usb, ub_usb_ids);
+#endif /* CONFIG_USB_LIBUSUAL */
 
 /*
  * Find me a way to identify "next free minor" for add_disk(),
@@ -341,118 +408,12 @@ MODULE_DEVICE_TABLE(usb, ub_usb_ids);
  */
 #define UB_MAX_HOSTS  26
 static char ub_hostv[UB_MAX_HOSTS];
-static DEFINE_SPINLOCK(ub_lock);	/* Locks globals and ->openc */
-
-/*
- * The SCSI command tracing procedures.
- */
-
-static void ub_cmdtr_new(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
-{
-	int n;
-	struct ub_scsi_cmd_trace *t;
-
-	if ((n = sc->tr.cur + 1) == SCMD_TRACE_SZ) n = 0;
-	t = &sc->tr.vec[n];
-
-	memset(t, 0, sizeof(struct ub_scsi_cmd_trace));
-	t->tag = cmd->tag;
-	t->op = cmd->cdb[0];
-	t->dir = cmd->dir;
-	t->req_size = cmd->len;
-	t->st_hst[0] = cmd->state;
-
-	sc->tr.cur = n;
-	cmd->trace_index = n;
-}
-
-static void ub_cmdtr_state(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
-{
-	int n;
-	struct ub_scsi_cmd_trace *t;
-
-	t = &sc->tr.vec[cmd->trace_index];
-	if (t->tag == cmd->tag) {
-		if ((n = t->hcur + 1) == SCMD_ST_HIST_SZ) n = 0;
-		t->st_hst[n] = cmd->state;
-		t->hcur = n;
-	}
-}
-
-static void ub_cmdtr_act_len(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
-{
-	struct ub_scsi_cmd_trace *t;
 
-	t = &sc->tr.vec[cmd->trace_index];
-	if (t->tag == cmd->tag)
-		t->act_size = cmd->act_len;
-}
+#define UB_QLOCK_NUM 5
+static spinlock_t ub_qlockv[UB_QLOCK_NUM];
+static int ub_qlock_next = 0;
 
-static void ub_cmdtr_sense(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
-    unsigned char *sense)
-{
-	struct ub_scsi_cmd_trace *t;
-
-	t = &sc->tr.vec[cmd->trace_index];
-	if (t->tag == cmd->tag) {
-		t->key = sense[2] & 0x0F;
-		t->asc = sense[12];
-		t->ascq = sense[13];
-	}
-}
-
-static ssize_t ub_diag_show(struct device *dev, char *page)
-{
-	struct usb_interface *intf;
-	struct ub_dev *sc;
-	int cnt;
-	unsigned long flags;
-	int nc, nh;
-	int i, j;
-	struct ub_scsi_cmd_trace *t;
-
-	intf = to_usb_interface(dev);
-	sc = usb_get_intfdata(intf);
-	if (sc == NULL)
-		return 0;
-
-	cnt = 0;
-	spin_lock_irqsave(&sc->lock, flags);
-
-	cnt += sprintf(page + cnt,
-	    "qlen %d qmax %d changed %d removable %d readonly %d\n",
-	    sc->cmd_queue.qlen, sc->cmd_queue.qmax,
-	    sc->changed, sc->removable, sc->readonly);
-
-	if ((nc = sc->tr.cur + 1) == SCMD_TRACE_SZ) nc = 0;
-	for (j = 0; j < SCMD_TRACE_SZ; j++) {
-		t = &sc->tr.vec[nc];
-
-		cnt += sprintf(page + cnt, "%08x %02x", t->tag, t->op);
-		if (t->op == REQUEST_SENSE) {
-			cnt += sprintf(page + cnt, " [sense %x %02x %02x]",
-					t->key, t->asc, t->ascq);
-		} else {
-			cnt += sprintf(page + cnt, " %c", UB_DIR_CHAR(t->dir));
-			cnt += sprintf(page + cnt, " [%5d %5d]",
-					t->req_size, t->act_size);
-		}
-		if ((nh = t->hcur + 1) == SCMD_ST_HIST_SZ) nh = 0;
-		for (i = 0; i < SCMD_ST_HIST_SZ; i++) {
-			cnt += sprintf(page + cnt, " %s",
-					ub_scsi_cmd_stname[(int)t->st_hst[nh]]);
-			if (++nh == SCMD_ST_HIST_SZ) nh = 0;
-		}
-		cnt += sprintf(page + cnt, "\n");
-
-		if (++nc == SCMD_TRACE_SZ) nc = 0;
-	}
-
-	spin_unlock_irqrestore(&sc->lock, flags);
-	return cnt;
-}
-
-static DEVICE_ATTR(diag, S_IRUGO, ub_diag_show, NULL); /* N.B. World readable */
+static DEFINE_SPINLOCK(ub_lock);	/* Locks globals and ->openc */
 
 /*
  * The id allocator.
@@ -478,66 +439,128 @@ static int ub_id_get(void)
 
 static void ub_id_put(int id)
 {
+	unsigned long flags;
 
 	if (id < 0 || id >= UB_MAX_HOSTS) {
 		printk(KERN_ERR DRV_NAME ": bad host ID %d\n", id);
 		return;
 	}
+
+	spin_lock_irqsave(&ub_lock, flags);
 	if (ub_hostv[id] == 0) {
+		spin_unlock_irqrestore(&ub_lock, flags);
 		printk(KERN_ERR DRV_NAME ": freeing free host ID %d\n", id);
 		return;
 	}
 	ub_hostv[id] = 0;
+	spin_unlock_irqrestore(&ub_lock, flags);
+}
+
+/*
+ * This is necessitated by the fact that blk_cleanup_queue does not
+ * necesserily destroy the queue. Instead, it may merely decrease q->refcnt.
+ * Since our blk_init_queue() passes a spinlock common with ub_dev,
+ * we have life time issues when ub_cleanup frees ub_dev.
+ */
+static spinlock_t *ub_next_lock(void)
+{
+	unsigned long flags;
+	spinlock_t *ret;
+
+	spin_lock_irqsave(&ub_lock, flags);
+	ret = &ub_qlockv[ub_qlock_next];
+	ub_qlock_next = (ub_qlock_next + 1) % UB_QLOCK_NUM;
+	spin_unlock_irqrestore(&ub_lock, flags);
+	return ret;
+}
+
+/*
+ * Downcount for deallocation. This rides on two assumptions:
+ *  - once something is poisoned, its refcount cannot grow
+ *  - opens cannot happen at this time (del_gendisk was done)
+ * If the above is true, we can drop the lock, which we need for
+ * blk_cleanup_queue(): the silly thing may attempt to sleep.
+ * [Actually, it never needs to sleep for us, but it calls might_sleep()]
+ */
+static void ub_put(struct ub_dev *sc)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&ub_lock, flags);
+	--sc->openc;
+	if (sc->openc == 0 && atomic_read(&sc->poison)) {
+		spin_unlock_irqrestore(&ub_lock, flags);
+		ub_cleanup(sc);
+	} else {
+		spin_unlock_irqrestore(&ub_lock, flags);
+	}
 }
 
 /*
  * Final cleanup and deallocation.
- * This must be called with ub_lock taken.
  */
 static void ub_cleanup(struct ub_dev *sc)
 {
+	struct list_head *p;
+	struct ub_lun *lun;
+	request_queue_t *q;
 
-	/*
-	 * If we zero disk->private_data BEFORE put_disk, we have to check
-	 * for NULL all over the place in open, release, check_media and
-	 * revalidate, because the block level semaphore is well inside the
-	 * put_disk. But we cannot zero after the call, because *disk is gone.
-	 * The sd.c is blatantly racy in this area.
-	 */
-	/* disk->private_data = NULL; */
-	put_disk(sc->disk);
-	sc->disk = NULL;
+	while (!list_empty(&sc->luns)) {
+		p = sc->luns.next;
+		lun = list_entry(p, struct ub_lun, link);
+		list_del(p);
+
+		/* I don't think queue can be NULL. But... Stolen from sx8.c */
+		if ((q = lun->disk->queue) != NULL)
+			blk_cleanup_queue(q);
+		/*
+		 * If we zero disk->private_data BEFORE put_disk, we have
+		 * to check for NULL all over the place in open, release,
+		 * check_media and revalidate, because the block level
+		 * semaphore is well inside the put_disk.
+		 * But we cannot zero after the call, because *disk is gone.
+		 * The sd.c is blatantly racy in this area.
+		 */
+		/* disk->private_data = NULL; */
+		put_disk(lun->disk);
+		lun->disk = NULL;
+
+		ub_id_put(lun->id);
+		kfree(lun);
+	}
 
-	ub_id_put(sc->id);
+	usb_set_intfdata(sc->intf, NULL);
+	usb_put_intf(sc->intf);
+	usb_put_dev(sc->dev);
 	kfree(sc);
 }
 
 /*
  * The "command allocator".
  */
-static struct ub_scsi_cmd *ub_get_cmd(struct ub_dev *sc)
+static struct ub_scsi_cmd *ub_get_cmd(struct ub_lun *lun)
 {
 	struct ub_scsi_cmd *ret;
 
-	if (sc->cmda[0])
+	if (lun->cmda[0])
 		return NULL;
-	ret = &sc->cmdv[0];
-	sc->cmda[0] = 1;
+	ret = &lun->cmdv[0];
+	lun->cmda[0] = 1;
 	return ret;
 }
 
-static void ub_put_cmd(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
+static void ub_put_cmd(struct ub_lun *lun, struct ub_scsi_cmd *cmd)
 {
-	if (cmd != &sc->cmdv[0]) {
+	if (cmd != &lun->cmdv[0]) {
 		printk(KERN_WARNING "%s: releasing a foreign cmd %p\n",
-		    sc->name, cmd);
+		    lun->name, cmd);
 		return;
 	}
-	if (!sc->cmda[0]) {
-		printk(KERN_WARNING "%s: releasing a free cmd\n", sc->name);
+	if (!lun->cmda[0]) {
+		printk(KERN_WARNING "%s: releasing a free cmd\n", lun->name);
 		return;
 	}
-	sc->cmda[0] = 0;
+	lun->cmda[0] = 0;
 }
 
 /*
@@ -596,119 +619,103 @@ static struct ub_scsi_cmd *ub_cmdq_pop(struct ub_dev *sc)
  * The request function is our main entry point
  */
 
-static void ub_bd_rq_fn(request_queue_t *q)
+static void ub_request_fn(request_queue_t *q)
 {
-	struct ub_dev *sc = q->queuedata;
+	struct ub_lun *lun = q->queuedata;
 	struct request *rq;
 
 	while ((rq = elv_next_request(q)) != NULL) {
-		if (ub_bd_rq_fn_1(sc, rq) != 0) {
+		if (ub_request_fn_1(lun, rq) != 0) {
 			blk_stop_queue(q);
 			break;
 		}
 	}
 }
 
-static int ub_bd_rq_fn_1(struct ub_dev *sc, struct request *rq)
+static int ub_request_fn_1(struct ub_lun *lun, struct request *rq)
 {
+	struct ub_dev *sc = lun->udev;
 	struct ub_scsi_cmd *cmd;
-	int rc;
+	struct ub_request *urq;
+	int n_elem;
+
+	if (atomic_read(&sc->poison)) {
+		blkdev_dequeue_request(rq);
+		ub_end_rq(rq, DID_NO_CONNECT << 16);
+		return 0;
+	}
 
-	if (atomic_read(&sc->poison) || sc->changed) {
+	if (lun->changed && !blk_pc_request(rq)) {
 		blkdev_dequeue_request(rq);
-		ub_end_rq(rq, 0);
+		ub_end_rq(rq, SAM_STAT_CHECK_CONDITION);
 		return 0;
 	}
 
-	if ((cmd = ub_get_cmd(sc)) == NULL)
+	if (lun->urq.rq != NULL)
+		return -1;
+	if ((cmd = ub_get_cmd(lun)) == NULL)
 		return -1;
 	memset(cmd, 0, sizeof(struct ub_scsi_cmd));
 
 	blkdev_dequeue_request(rq);
 
-	if (blk_pc_request(rq)) {
-		rc = ub_cmd_build_packet(sc, cmd, rq);
-	} else {
-		rc = ub_cmd_build_block(sc, cmd, rq);
+	urq = &lun->urq;
+	memset(urq, 0, sizeof(struct ub_request));
+	urq->rq = rq;
+
+	/*
+	 * get scatterlist from block layer
+	 */
+	n_elem = blk_rq_map_sg(lun->disk->queue, rq, &urq->sgv[0]);
+	if (n_elem < 0) {
+		/* Impossible, because blk_rq_map_sg should not hit ENOMEM. */
+		printk(KERN_INFO "%s: failed request map (%d)\n",
+		    lun->name, n_elem);
+		goto drop;
 	}
-	if (rc != 0) {
-		ub_put_cmd(sc, cmd);
-		ub_end_rq(rq, 0);
-		blk_start_queue(sc->disk->queue);
-		return 0;
+	if (n_elem > UB_MAX_REQ_SG) {	/* Paranoia */
+		printk(KERN_WARNING "%s: request with %d segments\n",
+		    lun->name, n_elem);
+		goto drop;
 	}
+	urq->nsg = n_elem;
+	sc->sg_stat[n_elem < 5 ? n_elem : 5]++;
 
+	if (blk_pc_request(rq)) {
+		ub_cmd_build_packet(sc, lun, cmd, urq);
+	} else {
+		ub_cmd_build_block(sc, lun, cmd, urq);
+	}
 	cmd->state = UB_CMDST_INIT;
+	cmd->lun = lun;
 	cmd->done = ub_rw_cmd_done;
-	cmd->back = rq;
+	cmd->back = urq;
 
 	cmd->tag = sc->tagcnt++;
-	if ((rc = ub_submit_scsi(sc, cmd)) != 0) {
-		ub_put_cmd(sc, cmd);
-		ub_end_rq(rq, 0);
-		blk_start_queue(sc->disk->queue);
-		return 0;
-	}
+	if (ub_submit_scsi(sc, cmd) != 0)
+		goto drop;
+
+	return 0;
 
+drop:
+	ub_put_cmd(lun, cmd);
+	ub_end_rq(rq, DID_ERROR << 16);
 	return 0;
 }
 
-static int ub_cmd_build_block(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
-    struct request *rq)
+static void ub_cmd_build_block(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_scsi_cmd *cmd, struct ub_request *urq)
 {
-	int ub_dir;
-#if 0 /* We use rq->buffer for now */
-	struct scatterlist *sg;
-	int n_elem;
-#endif
+	struct request *rq = urq->rq;
 	unsigned int block, nblks;
 
 	if (rq_data_dir(rq) == WRITE)
-		ub_dir = UB_DIR_WRITE;
+		cmd->dir = UB_DIR_WRITE;
 	else
-		ub_dir = UB_DIR_READ;
+		cmd->dir = UB_DIR_READ;
 
-	/*
-	 * get scatterlist from block layer
-	 */
-#if 0 /* We use rq->buffer for now */
-	sg = &cmd->sgv[0];
-	n_elem = blk_rq_map_sg(q, rq, sg);
-	if (n_elem <= 0) {
-		ub_put_cmd(sc, cmd);
-		ub_end_rq(rq, 0);
-		blk_start_queue(q);
-		return 0;		/* request with no s/g entries? */
-	}
-
-	if (n_elem != 1) {		/* Paranoia */
-		printk(KERN_WARNING "%s: request with %d segments\n",
-		    sc->name, n_elem);
-		ub_put_cmd(sc, cmd);
-		ub_end_rq(rq, 0);
-		blk_start_queue(q);
-		return 0;
-	}
-#endif
-
-	/*
-	 * XXX Unfortunately, this check does not work. It is quite possible
-	 * to get bogus non-null rq->buffer if you allow sg by mistake.
-	 */
-	if (rq->buffer == NULL) {
-		/*
-		 * This must not happen if we set the queue right.
-		 * The block level must create bounce buffers for us.
-		 */
-		static int do_print = 1;
-		if (do_print) {
-			printk(KERN_WARNING "%s: unmapped block request"
-			    " flags 0x%lx sectors %lu\n",
-			    sc->name, rq->flags, rq->nr_sectors);
-			do_print = 0;
-		}
-		return -1;
-	}
+	cmd->nsg = urq->nsg;
+	memcpy(cmd->sgv, urq->sgv, sizeof(struct scatterlist) * cmd->nsg);
 
 	/*
 	 * build the command
@@ -716,10 +723,10 @@ static int ub_cmd_build_block(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
 	 * The call to blk_queue_hardsect_size() guarantees that request
 	 * is aligned, but it is given in terms of 512 byte units, always.
 	 */
-	block = rq->sector >> sc->capacity.bshift;
-	nblks = rq->nr_sectors >> sc->capacity.bshift;
+	block = rq->sector >> lun->capacity.bshift;
+	nblks = rq->nr_sectors >> lun->capacity.bshift;
 
-	cmd->cdb[0] = (ub_dir == UB_DIR_READ)? READ_10: WRITE_10;
+	cmd->cdb[0] = (cmd->dir == UB_DIR_READ)? READ_10: WRITE_10;
 	/* 10-byte uses 4 bytes of LBA: 2147483648KB, 2097152MB, 2048GB */
 	cmd->cdb[2] = block >> 24;
 	cmd->cdb[3] = block >> 16;
@@ -729,30 +736,13 @@ static int ub_cmd_build_block(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
 	cmd->cdb[8] = nblks;
 	cmd->cdb_len = 10;
 
-	cmd->dir = ub_dir;
-	cmd->data = rq->buffer;
 	cmd->len = rq->nr_sectors * 512;
-
-	return 0;
 }
 
-static int ub_cmd_build_packet(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
-    struct request *rq)
+static void ub_cmd_build_packet(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_scsi_cmd *cmd, struct ub_request *urq)
 {
-
-	if (rq->data_len != 0 && rq->data == NULL) {
-		static int do_print = 1;
-		if (do_print) {
-			printk(KERN_WARNING "%s: unmapped packet request"
-			    " flags 0x%lx length %d\n",
-			    sc->name, rq->flags, rq->data_len);
-			do_print = 0;
-		}
-		return -1;
-	}
-
-	memcpy(&cmd->cdb, rq->cmd, rq->cmd_len);
-	cmd->cdb_len = rq->cmd_len;
+	struct request *rq = urq->rq;
 
 	if (rq->data_len == 0) {
 		cmd->dir = UB_DIR_NONE;
@@ -762,42 +752,107 @@ static int ub_cmd_build_packet(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
 		else
 			cmd->dir = UB_DIR_READ;
 	}
-	cmd->data = rq->data;
-	cmd->len = rq->data_len;
 
-	return 0;
+	cmd->nsg = urq->nsg;
+	memcpy(cmd->sgv, urq->sgv, sizeof(struct scatterlist) * cmd->nsg);
+
+	memcpy(&cmd->cdb, rq->cmd, rq->cmd_len);
+	cmd->cdb_len = rq->cmd_len;
+
+	cmd->len = rq->data_len;
 }
 
 static void ub_rw_cmd_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
-	struct request *rq = cmd->back;
-	struct gendisk *disk = sc->disk;
-	request_queue_t *q = disk->queue;
-	int uptodate;
+	struct ub_lun *lun = cmd->lun;
+	struct ub_request *urq = cmd->back;
+	struct request *rq;
+	unsigned int scsi_status;
 
-	if (blk_pc_request(rq)) {
-		/* UB_SENSE_SIZE is smaller than SCSI_SENSE_BUFFERSIZE */
-		memcpy(rq->sense, sc->top_sense, UB_SENSE_SIZE);
-		rq->sense_len = UB_SENSE_SIZE;
+	rq = urq->rq;
+
+	if (cmd->error == 0) {
+		if (blk_pc_request(rq)) {
+			if (cmd->act_len >= rq->data_len)
+				rq->data_len = 0;
+			else
+				rq->data_len -= cmd->act_len;
+		}
+		scsi_status = 0;
+	} else {
+		if (blk_pc_request(rq)) {
+			/* UB_SENSE_SIZE is smaller than SCSI_SENSE_BUFFERSIZE */
+			memcpy(rq->sense, sc->top_sense, UB_SENSE_SIZE);
+			rq->sense_len = UB_SENSE_SIZE;
+			if (sc->top_sense[0] != 0)
+				scsi_status = SAM_STAT_CHECK_CONDITION;
+			else
+				scsi_status = DID_ERROR << 16;
+		} else {
+			if (cmd->error == -EIO) {
+				if (ub_rw_cmd_retry(sc, lun, urq, cmd) == 0)
+					return;
+			}
+			scsi_status = SAM_STAT_CHECK_CONDITION;
+		}
 	}
 
-	if (cmd->error == 0)
+	urq->rq = NULL;
+
+	ub_put_cmd(lun, cmd);
+	ub_end_rq(rq, scsi_status);
+	blk_start_queue(lun->disk->queue);
+}
+
+static void ub_end_rq(struct request *rq, unsigned int scsi_status)
+{
+	int uptodate;
+
+	if (scsi_status == 0) {
 		uptodate = 1;
-	else
+	} else {
 		uptodate = 0;
-
-	ub_put_cmd(sc, cmd);
-	ub_end_rq(rq, uptodate);
-	blk_start_queue(q);
+		rq->errors = scsi_status;
+	}
+	end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
+	end_that_request_last(rq, uptodate);
 }
 
-static void ub_end_rq(struct request *rq, int uptodate)
+static int ub_rw_cmd_retry(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_request *urq, struct ub_scsi_cmd *cmd)
 {
-	int rc;
 
-	rc = end_that_request_first(rq, uptodate, rq->hard_nr_sectors);
-	// assert(rc == 0);
-	end_that_request_last(rq);
+	if (atomic_read(&sc->poison))
+		return -ENXIO;
+
+	ub_reset_enter(sc, urq->current_try);
+
+	if (urq->current_try >= 3)
+		return -EIO;
+	urq->current_try++;
+
+	/* Remove this if anyone complains of flooding. */
+	printk(KERN_DEBUG "%s: dir %c len/act %d/%d "
+	    "[sense %x %02x %02x] retry %d\n",
+	    sc->name, UB_DIR_CHAR(cmd->dir), cmd->len, cmd->act_len,
+	    cmd->key, cmd->asc, cmd->ascq, urq->current_try);
+
+	memset(cmd, 0, sizeof(struct ub_scsi_cmd));
+	ub_cmd_build_block(sc, lun, cmd, urq);
+
+	cmd->state = UB_CMDST_INIT;
+	cmd->lun = lun;
+	cmd->done = ub_rw_cmd_done;
+	cmd->back = urq;
+
+	cmd->tag = sc->tagcnt++;
+
+#if 0 /* Wasteful */
+	return ub_submit_scsi(sc, cmd);
+#else
+	ub_cmdq_add(sc, cmd);
+	return 0;
+#endif
 }
 
 /*
@@ -808,8 +863,6 @@ static void ub_end_rq(struct request *rq, int uptodate)
  * No exceptions.
  *
  * Host is assumed locked.
- *
- * XXX We only support Bulk for the moment.
  */
 static int ub_submit_scsi(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
@@ -855,7 +908,7 @@ static int ub_scsi_cmd_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 	bcb->Tag = cmd->tag;		/* Endianness is not important */
 	bcb->DataTransferLength = cpu_to_le32(cmd->len);
 	bcb->Flags = (cmd->dir == UB_DIR_READ) ? 0x80 : 0;
-	bcb->Lun = 0;			/* No multi-LUN yet */
+	bcb->Lun = (cmd->lun != NULL) ? cmd->lun->num : 0;
 	bcb->Length = cmd->cdb_len;
 
 	/* copy the command payload */
@@ -866,7 +919,6 @@ static int ub_scsi_cmd_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 	sc->last_pipe = sc->send_bulk_pipe;
 	usb_fill_bulk_urb(&sc->work_urb, sc->dev, sc->send_bulk_pipe,
 	    bcb, US_BULK_CB_WRAP_LEN, ub_urb_complete, sc);
-	sc->work_urb.transfer_flags = URB_ASYNC_UNLINK;
 
 	/* Fill what we shouldn't be filling, because usb-storage did so. */
 	sc->work_urb.actual_length = 0;
@@ -875,7 +927,6 @@ static int ub_scsi_cmd_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 
 	if ((rc = usb_submit_urb(&sc->work_urb, GFP_ATOMIC)) != 0) {
 		/* XXX Clear stalls */
-		printk("ub: cmd #%d start failed (%d)\n", cmd->tag, rc); /* P3 */
 		ub_complete(&sc->work_done);
 		return rc;
 	}
@@ -884,7 +935,6 @@ static int ub_scsi_cmd_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 	add_timer(&sc->work_timer);
 
 	cmd->state = UB_CMDST_CMD;
-	ub_cmdtr_state(sc, cmd);
 	return 0;
 }
 
@@ -896,9 +946,10 @@ static void ub_urb_timeout(unsigned long arg)
 	struct ub_dev *sc = (struct ub_dev *) arg;
 	unsigned long flags;
 
-	spin_lock_irqsave(&sc->lock, flags);
-	usb_unlink_urb(&sc->work_urb);
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_lock_irqsave(sc->lock, flags);
+	if (!ub_is_completed(&sc->work_done))
+		usb_unlink_urb(&sc->work_urb);
+	spin_unlock_irqrestore(sc->lock, flags);
 }
 
 /*
@@ -908,7 +959,7 @@ static void ub_urb_timeout(unsigned long arg)
  * the sc->lock taken) and from an interrupt (while we do NOT have
  * the sc->lock taken). Therefore, bounce this off to a tasklet.
  */
-static void ub_urb_complete(struct urb *urb, struct pt_regs *pt)
+static void ub_urb_complete(struct urb *urb)
 {
 	struct ub_dev *sc = urb->context;
 
@@ -921,10 +972,9 @@ static void ub_scsi_action(unsigned long _dev)
 	struct ub_dev *sc = (struct ub_dev *) _dev;
 	unsigned long flags;
 
-	spin_lock_irqsave(&sc->lock, flags);
-	del_timer(&sc->work_timer);
+	spin_lock_irqsave(sc->lock, flags);
 	ub_scsi_dispatch(sc);
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_unlock_irqrestore(sc->lock, flags);
 }
 
 static void ub_scsi_dispatch(struct ub_dev *sc)
@@ -932,20 +982,19 @@ static void ub_scsi_dispatch(struct ub_dev *sc)
 	struct ub_scsi_cmd *cmd;
 	int rc;
 
-	while ((cmd = ub_cmdq_peek(sc)) != NULL) {
+	while (!sc->reset && (cmd = ub_cmdq_peek(sc)) != NULL) {
 		if (cmd->state == UB_CMDST_DONE) {
 			ub_cmdq_pop(sc);
 			(*cmd->done)(sc, cmd);
 		} else if (cmd->state == UB_CMDST_INIT) {
-			ub_cmdtr_new(sc, cmd);
 			if ((rc = ub_scsi_cmd_start(sc, cmd)) == 0)
 				break;
 			cmd->error = rc;
 			cmd->state = UB_CMDST_DONE;
-			ub_cmdtr_state(sc, cmd);
 		} else {
 			if (!ub_is_completed(&sc->work_done))
 				break;
+			del_timer(&sc->work_timer);
 			ub_scsi_urb_compl(sc, cmd);
 		}
 	}
@@ -955,12 +1004,12 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
 	struct urb *urb = &sc->work_urb;
 	struct bulk_cs_wrap *bcs;
-	int pipe;
+	int len;
 	int rc;
 
 	if (atomic_read(&sc->poison)) {
-		/* A little too simplistic, I feel... */
-		goto Bad_End;
+		ub_state_done(sc, cmd, -ENODEV);
+		return;
 	}
 
 	if (cmd->state == UB_CMDST_CLEAR) {
@@ -968,11 +1017,9 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 			/*
 			 * STALL while clearning STALL.
 			 * The control pipe clears itself - nothing to do.
-			 * XXX Might try to reset the device here and retry.
 			 */
-			printk(KERN_NOTICE "%s: "
-			    "stall on control pipe for device %u\n",
-			    sc->name, sc->dev->devnum);
+			printk(KERN_NOTICE "%s: stall on control pipe\n",
+			    sc->name);
 			goto Bad_End;
 		}
 
@@ -988,14 +1035,8 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 
 	} else if (cmd->state == UB_CMDST_CLR2STS) {
 		if (urb->status == -EPIPE) {
-			/*
-			 * STALL while clearning STALL.
-			 * The control pipe clears itself - nothing to do.
-			 * XXX Might try to reset the device here and retry.
-			 */
-			printk(KERN_NOTICE "%s: "
-			    "stall on control pipe for device %u\n",
-			    sc->name, sc->dev->devnum);
+			printk(KERN_NOTICE "%s: stall on control pipe\n",
+			    sc->name);
 			goto Bad_End;
 		}
 
@@ -1009,83 +1050,74 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 
 		ub_state_stat(sc, cmd);
 
-	} else if (cmd->state == UB_CMDST_CMD) {
+	} else if (cmd->state == UB_CMDST_CLRRS) {
 		if (urb->status == -EPIPE) {
+			printk(KERN_NOTICE "%s: stall on control pipe\n",
+			    sc->name);
+			goto Bad_End;
+		}
+
+		/*
+		 * We ignore the result for the halt clear.
+		 */
+
+		/* reset the endpoint toggle */
+		usb_settoggle(sc->dev, usb_pipeendpoint(sc->last_pipe),
+			usb_pipeout(sc->last_pipe), 0);
+
+		ub_state_stat_counted(sc, cmd);
+
+	} else if (cmd->state == UB_CMDST_CMD) {
+		switch (urb->status) {
+		case 0:
+			break;
+		case -EOVERFLOW:
+			goto Bad_End;
+		case -EPIPE:
 			rc = ub_submit_clear_stall(sc, cmd, sc->last_pipe);
 			if (rc != 0) {
 				printk(KERN_NOTICE "%s: "
-				    "unable to submit clear for device %u"
-				    " (code %d)\n",
-				    sc->name, sc->dev->devnum, rc);
+				    "unable to submit clear (%d)\n",
+				    sc->name, rc);
 				/*
 				 * This is typically ENOMEM or some other such shit.
 				 * Retrying is pointless. Just do Bad End on it...
 				 */
-				goto Bad_End;
+				ub_state_done(sc, cmd, rc);
+				return;
 			}
 			cmd->state = UB_CMDST_CLEAR;
-			ub_cmdtr_state(sc, cmd);
 			return;
-		}
-		if (urb->status != 0) {
-			printk("ub: cmd #%d cmd status (%d)\n", cmd->tag, urb->status); /* P3 */
+		case -ESHUTDOWN:	/* unplug */
+		case -EILSEQ:		/* unplug timeout on uhci */
+			ub_state_done(sc, cmd, -ENODEV);
+			return;
+		default:
 			goto Bad_End;
 		}
 		if (urb->actual_length != US_BULK_CB_WRAP_LEN) {
-			printk("ub: cmd #%d xferred %d\n", cmd->tag, urb->actual_length); /* P3 */
-			/* XXX Must do reset here to unconfuse the device */
 			goto Bad_End;
 		}
 
-		if (cmd->dir == UB_DIR_NONE) {
+		if (cmd->dir == UB_DIR_NONE || cmd->nsg < 1) {
 			ub_state_stat(sc, cmd);
 			return;
 		}
 
-		UB_INIT_COMPLETION(sc->work_done);
-
-		if (cmd->dir == UB_DIR_READ)
-			pipe = sc->recv_bulk_pipe;
-		else
-			pipe = sc->send_bulk_pipe;
-		sc->last_pipe = pipe;
-		usb_fill_bulk_urb(&sc->work_urb, sc->dev, pipe,
-		    cmd->data, cmd->len, ub_urb_complete, sc);
-		sc->work_urb.transfer_flags = URB_ASYNC_UNLINK;
-		sc->work_urb.actual_length = 0;
-		sc->work_urb.error_count = 0;
-		sc->work_urb.status = 0;
-
-		if ((rc = usb_submit_urb(&sc->work_urb, GFP_ATOMIC)) != 0) {
-			/* XXX Clear stalls */
-			printk("ub: data #%d submit failed (%d)\n", cmd->tag, rc); /* P3 */
-			ub_complete(&sc->work_done);
-			ub_state_done(sc, cmd, rc);
-			return;
-		}
-
-		sc->work_timer.expires = jiffies + UB_DATA_TIMEOUT;
-		add_timer(&sc->work_timer);
-
-		cmd->state = UB_CMDST_DATA;
-		ub_cmdtr_state(sc, cmd);
+		// udelay(125);		// usb-storage has this
+		ub_data_start(sc, cmd);
 
 	} else if (cmd->state == UB_CMDST_DATA) {
 		if (urb->status == -EPIPE) {
 			rc = ub_submit_clear_stall(sc, cmd, sc->last_pipe);
 			if (rc != 0) {
 				printk(KERN_NOTICE "%s: "
-				    "unable to submit clear for device %u"
-				    " (code %d)\n",
-				    sc->name, sc->dev->devnum, rc);
-				/*
-				 * This is typically ENOMEM or some other such shit.
-				 * Retrying is pointless. Just do Bad End on it...
-				 */
-				goto Bad_End;
+				    "unable to submit clear (%d)\n",
+				    sc->name, rc);
+				ub_state_done(sc, cmd, rc);
+				return;
 			}
 			cmd->state = UB_CMDST_CLR2STS;
-			ub_cmdtr_state(sc, cmd);
 			return;
 		}
 		if (urb->status == -EOVERFLOW) {
@@ -1093,14 +1125,54 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 			 * A babble? Failure, but we must transfer CSW now.
 			 */
 			cmd->error = -EOVERFLOW;	/* A cheap trick... */
+			ub_state_stat(sc, cmd);
+			return;
+		}
+
+		if (cmd->dir == UB_DIR_WRITE) {
+			/*
+			 * Do not continue writes in case of a failure.
+			 * Doing so would cause sectors to be mixed up,
+			 * which is worse than sectors lost.
+			 *
+			 * We must try to read the CSW, or many devices
+			 * get confused.
+			 */
+			len = urb->actual_length;
+			if (urb->status != 0 ||
+			    len != cmd->sgv[cmd->current_sg].length) {
+				cmd->act_len += len;
+
+				cmd->error = -EIO;
+				ub_state_stat(sc, cmd);
+				return;
+			}
+
 		} else {
+			/*
+			 * If an error occurs on read, we record it, and
+			 * continue to fetch data in order to avoid bubble.
+			 *
+			 * As a small shortcut, we stop if we detect that
+			 * a CSW mixed into data.
+			 */
 			if (urb->status != 0)
-				goto Bad_End;
+				cmd->error = -EIO;
+
+			len = urb->actual_length;
+			if (urb->status != 0 ||
+			    len != cmd->sgv[cmd->current_sg].length) {
+				if ((len & 0x1FF) == US_BULK_CS_WRAP_LEN)
+					goto Bad_End;
+			}
 		}
 
-		cmd->act_len = urb->actual_length;
-		ub_cmdtr_act_len(sc, cmd);
+		cmd->act_len += urb->actual_length;
 
+		if (++cmd->current_sg < cmd->nsg) {
+			ub_data_start(sc, cmd);
+			return;
+		}
 		ub_state_stat(sc, cmd);
 
 	} else if (cmd->state == UB_CMDST_STAT) {
@@ -1108,62 +1180,55 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 			rc = ub_submit_clear_stall(sc, cmd, sc->last_pipe);
 			if (rc != 0) {
 				printk(KERN_NOTICE "%s: "
-				    "unable to submit clear for device %u"
-				    " (code %d)\n",
-				    sc->name, sc->dev->devnum, rc);
-				/*
-				 * This is typically ENOMEM or some other such shit.
-				 * Retrying is pointless. Just do Bad End on it...
-				 */
-				goto Bad_End;
+				    "unable to submit clear (%d)\n",
+				    sc->name, rc);
+				ub_state_done(sc, cmd, rc);
+				return;
 			}
-			cmd->state = UB_CMDST_CLEAR;
-			ub_cmdtr_state(sc, cmd);
+
+			/*
+			 * Having a stall when getting CSW is an error, so
+			 * make sure uppper levels are not oblivious to it.
+			 */
+			cmd->error = -EIO;		/* A cheap trick... */
+
+			cmd->state = UB_CMDST_CLRRS;
 			return;
 		}
+
+		/* Catch everything, including -EOVERFLOW and other nasties. */
 		if (urb->status != 0)
 			goto Bad_End;
 
 		if (urb->actual_length == 0) {
-			/*
-			 * Some broken devices add unnecessary zero-length
-			 * packets to the end of their data transfers.
-			 * Such packets show up as 0-length CSWs. If we
-			 * encounter such a thing, try to read the CSW again.
-			 */
-			if (++cmd->stat_count >= 4) {
-				printk(KERN_NOTICE "%s: "
-				    "unable to get CSW on device %u\n",
-				    sc->name, sc->dev->devnum);
-				goto Bad_End;
-			}
-			__ub_state_stat(sc, cmd);
+			ub_state_stat_counted(sc, cmd);
 			return;
 		}
 
 		/*
 		 * Check the returned Bulk protocol status.
+		 * The status block has to be validated first.
 		 */
 
 		bcs = &sc->work_bcs;
-		rc = le32_to_cpu(bcs->Residue);
-		if (rc != cmd->len - cmd->act_len) {
+
+		if (sc->signature == cpu_to_le32(0)) {
 			/*
-			 * It is all right to transfer less, the caller has
-			 * to check. But it's not all right if the device
-			 * counts disagree with our counts.
+			 * This is the first reply, so do not perform the check.
+			 * Instead, remember the signature the device uses
+			 * for future checks. But do not allow a nul.
 			 */
-			/* P3 */ printk("%s: resid %d len %d act %d\n",
-			    sc->name, rc, cmd->len, cmd->act_len);
-			goto Bad_End;
-		}
-
-#if 0
-		if (bcs->Signature != cpu_to_le32(US_BULK_CS_SIGN) &&
-		    bcs->Signature != cpu_to_le32(US_BULK_CS_OLYMPUS_SIGN)) {
-			/* Windows ignores signatures, so do we. */
+			sc->signature = bcs->Signature;
+			if (sc->signature == cpu_to_le32(0)) {
+				ub_state_stat_counted(sc, cmd);
+				return;
+			}
+		} else {
+			if (bcs->Signature != sc->signature) {
+				ub_state_stat_counted(sc, cmd);
+				return;
+			}
 		}
-#endif
 
 		if (bcs->Tag != cmd->tag) {
 			/*
@@ -1173,18 +1238,20 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 			 * commands and reply at commands we timed out before.
 			 * Without flushing these replies we loop forever.
 			 */
-			if (++cmd->stat_count >= 4) {
-				printk(KERN_NOTICE "%s: "
-				    "tag mismatch orig 0x%x reply 0x%x "
-				    "on device %u\n",
-				    sc->name, cmd->tag, bcs->Tag,
-				    sc->dev->devnum);
-				goto Bad_End;
-			}
-			__ub_state_stat(sc, cmd);
+			ub_state_stat_counted(sc, cmd);
 			return;
 		}
 
+		len = le32_to_cpu(bcs->Residue);
+		if (len != cmd->len - cmd->act_len) {
+			/*
+			 * It is all right to transfer less, the caller has
+			 * to check. But it's not all right if the device
+			 * counts disagree with our counts.
+			 */
+			goto Bad_End;
+		}
+
 		switch (bcs->Status) {
 		case US_BULK_STAT_OK:
 			break;
@@ -1192,34 +1259,74 @@ static void ub_scsi_urb_compl(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 			ub_state_sense(sc, cmd);
 			return;
 		case US_BULK_STAT_PHASE:
-			/* XXX We must reset the transport here */
-			/* P3 */ printk("%s: status PHASE\n", sc->name);
 			goto Bad_End;
 		default:
 			printk(KERN_INFO "%s: unknown CSW status 0x%x\n",
 			    sc->name, bcs->Status);
-			goto Bad_End;
+			ub_state_done(sc, cmd, -EINVAL);
+			return;
 		}
 
 		/* Not zeroing error to preserve a babble indicator */
+		if (cmd->error != 0) {
+			ub_state_sense(sc, cmd);
+			return;
+		}
 		cmd->state = UB_CMDST_DONE;
-		ub_cmdtr_state(sc, cmd);
 		ub_cmdq_pop(sc);
 		(*cmd->done)(sc, cmd);
 
 	} else if (cmd->state == UB_CMDST_SENSE) {
 		ub_state_done(sc, cmd, -EIO);
 
-	} else {
-		printk(KERN_WARNING "%s: "
-		    "wrong command state %d on device %u\n",
-		    sc->name, cmd->state, sc->dev->devnum);
-		goto Bad_End;
+	} else {
+		printk(KERN_WARNING "%s: "
+		    "wrong command state %d\n",
+		    sc->name, cmd->state);
+		ub_state_done(sc, cmd, -EINVAL);
+		return;
+	}
+	return;
+
+Bad_End: /* Little Excel is dead */
+	ub_state_done(sc, cmd, -EIO);
+}
+
+/*
+ * Factorization helper for the command state machine:
+ * Initiate a data segment transfer.
+ */
+static void ub_data_start(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
+{
+	struct scatterlist *sg = &cmd->sgv[cmd->current_sg];
+	int pipe;
+	int rc;
+
+	UB_INIT_COMPLETION(sc->work_done);
+
+	if (cmd->dir == UB_DIR_READ)
+		pipe = sc->recv_bulk_pipe;
+	else
+		pipe = sc->send_bulk_pipe;
+	sc->last_pipe = pipe;
+	usb_fill_bulk_urb(&sc->work_urb, sc->dev, pipe,
+	    page_address(sg->page) + sg->offset, sg->length,
+	    ub_urb_complete, sc);
+	sc->work_urb.actual_length = 0;
+	sc->work_urb.error_count = 0;
+	sc->work_urb.status = 0;
+
+	if ((rc = usb_submit_urb(&sc->work_urb, GFP_ATOMIC)) != 0) {
+		/* XXX Clear stalls */
+		ub_complete(&sc->work_done);
+		ub_state_done(sc, cmd, rc);
+		return;
 	}
-	return;
 
-Bad_End: /* Little Excel is dead */
-	ub_state_done(sc, cmd, -EIO);
+	sc->work_timer.expires = jiffies + UB_DATA_TIMEOUT;
+	add_timer(&sc->work_timer);
+
+	cmd->state = UB_CMDST_DATA;
 }
 
 /*
@@ -1231,7 +1338,6 @@ static void ub_state_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd, int rc)
 
 	cmd->error = rc;
 	cmd->state = UB_CMDST_DONE;
-	ub_cmdtr_state(sc, cmd);
 	ub_cmdq_pop(sc);
 	(*cmd->done)(sc, cmd);
 }
@@ -1240,7 +1346,7 @@ static void ub_state_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd, int rc)
  * Factorization helper for the command state machine:
  * Submit a CSW read.
  */
-static void __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
+static int __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
 	int rc;
 
@@ -1249,21 +1355,20 @@ static void __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 	sc->last_pipe = sc->recv_bulk_pipe;
 	usb_fill_bulk_urb(&sc->work_urb, sc->dev, sc->recv_bulk_pipe,
 	    &sc->work_bcs, US_BULK_CS_WRAP_LEN, ub_urb_complete, sc);
-	sc->work_urb.transfer_flags = URB_ASYNC_UNLINK;
 	sc->work_urb.actual_length = 0;
 	sc->work_urb.error_count = 0;
 	sc->work_urb.status = 0;
 
 	if ((rc = usb_submit_urb(&sc->work_urb, GFP_ATOMIC)) != 0) {
 		/* XXX Clear stalls */
-		printk("%s: CSW #%d submit failed (%d)\n", sc->name, cmd->tag, rc); /* P3 */
 		ub_complete(&sc->work_done);
 		ub_state_done(sc, cmd, rc);
-		return;
+		return -1;
 	}
 
 	sc->work_timer.expires = jiffies + UB_STAT_TIMEOUT;
 	add_timer(&sc->work_timer);
+	return 0;
 }
 
 /*
@@ -1272,11 +1377,30 @@ static void __ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
  */
 static void ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
-	__ub_state_stat(sc, cmd);
+
+	if (__ub_state_stat(sc, cmd) != 0)
+		return;
 
 	cmd->stat_count = 0;
 	cmd->state = UB_CMDST_STAT;
-	ub_cmdtr_state(sc, cmd);
+}
+
+/*
+ * Factorization helper for the command state machine:
+ * Submit a CSW read and go to STAT state with counter (along [C] path).
+ */
+static void ub_state_stat_counted(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
+{
+
+	if (++cmd->stat_count >= 4) {
+		ub_state_sense(sc, cmd);
+		return;
+	}
+
+	if (__ub_state_stat(sc, cmd) != 0)
+		return;
+
+	cmd->state = UB_CMDST_STAT;
 }
 
 /*
@@ -1286,6 +1410,7 @@ static void ub_state_stat(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 static void ub_state_sense(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 {
 	struct ub_scsi_cmd *scmd;
+	struct scatterlist *sg;
 	int rc;
 
 	if (cmd->cdb[0] == REQUEST_SENSE) {
@@ -1294,20 +1419,25 @@ static void ub_state_sense(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 	}
 
 	scmd = &sc->top_rqs_cmd;
+	memset(scmd, 0, sizeof(struct ub_scsi_cmd));
 	scmd->cdb[0] = REQUEST_SENSE;
 	scmd->cdb[4] = UB_SENSE_SIZE;
 	scmd->cdb_len = 6;
 	scmd->dir = UB_DIR_READ;
 	scmd->state = UB_CMDST_INIT;
-	scmd->data = sc->top_sense;
+	scmd->nsg = 1;
+	sg = &scmd->sgv[0];
+	sg->page = virt_to_page(sc->top_sense);
+	sg->offset = (unsigned long)sc->top_sense & (PAGE_SIZE-1);
+	sg->length = UB_SENSE_SIZE;
 	scmd->len = UB_SENSE_SIZE;
+	scmd->lun = cmd->lun;
 	scmd->done = ub_top_sense_done;
 	scmd->back = cmd;
 
 	scmd->tag = sc->tagcnt++;
 
 	cmd->state = UB_CMDST_SENSE;
-	ub_cmdtr_state(sc, cmd);
 
 	ub_cmdq_insert(sc, scmd);
 	return;
@@ -1342,7 +1472,6 @@ static int ub_submit_clear_stall(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
 
 	usb_fill_control_urb(&sc->work_urb, sc->dev, sc->send_ctrl_pipe,
 	    (unsigned char*) cr, NULL, 0, ub_urb_complete, sc);
-	sc->work_urb.transfer_flags = URB_ASYNC_UNLINK;
 	sc->work_urb.actual_length = 0;
 	sc->work_urb.error_count = 0;
 	sc->work_urb.status = 0;
@@ -1361,14 +1490,9 @@ static int ub_submit_clear_stall(struct ub_dev *sc, struct ub_scsi_cmd *cmd,
  */
 static void ub_top_sense_done(struct ub_dev *sc, struct ub_scsi_cmd *scmd)
 {
-	unsigned char *sense = scmd->data;
+	unsigned char *sense = sc->top_sense;
 	struct ub_scsi_cmd *cmd;
 
-	/*
-	 * Ignoring scmd->act_len, because the buffer was pre-zeroed.
-	 */
-	ub_cmdtr_sense(sc, scmd, sense);
-
 	/*
 	 * Find the command which triggered the unit attention or a check,
 	 * save the sense into it, and advance its state machine.
@@ -1379,17 +1503,20 @@ static void ub_top_sense_done(struct ub_dev *sc, struct ub_scsi_cmd *scmd)
 	}
 	if (cmd != scmd->back) {
 		printk(KERN_WARNING "%s: "
-		    "sense done for wrong command 0x%x on device %u\n",
-		    sc->name, cmd->tag, sc->dev->devnum);
+		    "sense done for wrong command 0x%x\n",
+		    sc->name, cmd->tag);
 		return;
 	}
 	if (cmd->state != UB_CMDST_SENSE) {
 		printk(KERN_WARNING "%s: "
-		    "sense done with bad cmd state %d on device %u\n",
-		    sc->name, cmd->state, sc->dev->devnum);
+		    "sense done with bad cmd state %d\n",
+		    sc->name, cmd->state);
 		return;
 	}
 
+	/*
+	 * Ignoring scmd->act_len, because the buffer was pre-zeroed.
+	 */
 	cmd->key = sense[2] & 0x0F;
 	cmd->asc = sense[12];
 	cmd->ascq = sense[13];
@@ -1397,68 +1524,124 @@ static void ub_top_sense_done(struct ub_dev *sc, struct ub_scsi_cmd *scmd)
 	ub_scsi_urb_compl(sc, cmd);
 }
 
-#if 0
-/* Determine what the maximum LUN supported is */
-int usb_stor_Bulk_max_lun(struct us_data *us)
+/*
+ * Reset management
+ * XXX Move usb_reset_device to khubd. Hogging kevent is not a good thing.
+ * XXX Make usb_sync_reset asynchronous.
+ */
+
+static void ub_reset_enter(struct ub_dev *sc, int try)
 {
-	int result;
-
-	/* issue the command */
-	result = usb_stor_control_msg(us, us->recv_ctrl_pipe,
-				 US_BULK_GET_MAX_LUN, 
-				 USB_DIR_IN | USB_TYPE_CLASS | 
-				 USB_RECIP_INTERFACE,
-				 0, us->ifnum, us->iobuf, 1, HZ);
-
-	/* 
-	 * Some devices (i.e. Iomega Zip100) need this -- apparently
-	 * the bulk pipes get STALLed when the GetMaxLUN request is
-	 * processed.   This is, in theory, harmless to all other devices
-	 * (regardless of if they stall or not).
-	 */
-	if (result < 0) {
-		usb_stor_clear_halt(us, us->recv_bulk_pipe);
-		usb_stor_clear_halt(us, us->send_bulk_pipe);
+
+	if (sc->reset) {
+		/* This happens often on multi-LUN devices. */
+		return;
 	}
+	sc->reset = try + 1;
 
-	US_DEBUGP("GetMaxLUN command result is %d, data is %d\n", 
-		  result, us->iobuf[0]);
+#if 0 /* Not needed because the disconnect waits for us. */
+	unsigned long flags;
+	spin_lock_irqsave(&ub_lock, flags);
+	sc->openc++;
+	spin_unlock_irqrestore(&ub_lock, flags);
+#endif
 
-	/* if we have a successful request, return the result */
-	if (result == 1)
-		return us->iobuf[0];
+#if 0 /* We let them stop themselves. */
+	struct list_head *p;
+	struct ub_lun *lun;
+	list_for_each(p, &sc->luns) {
+		lun = list_entry(p, struct ub_lun, link);
+		blk_stop_queue(lun->disk->queue);
+	}
+#endif
 
-	/* return the default -- no LUNs */
-	return 0;
+	schedule_work(&sc->reset_work);
+}
+
+static void ub_reset_task(struct work_struct *work)
+{
+	struct ub_dev *sc = container_of(work, struct ub_dev, reset_work);
+	unsigned long flags;
+	struct list_head *p;
+	struct ub_lun *lun;
+	int lkr, rc;
+
+	if (!sc->reset) {
+		printk(KERN_WARNING "%s: Running reset unrequested\n",
+		    sc->name);
+		return;
+	}
+
+	if (atomic_read(&sc->poison)) {
+		;
+	} else if ((sc->reset & 1) == 0) {
+		ub_sync_reset(sc);
+		msleep(700);	/* usb-storage sleeps 6s (!) */
+		ub_probe_clear_stall(sc, sc->recv_bulk_pipe);
+		ub_probe_clear_stall(sc, sc->send_bulk_pipe);
+	} else if (sc->dev->actconfig->desc.bNumInterfaces != 1) {
+		;
+	} else {
+		if ((lkr = usb_lock_device_for_reset(sc->dev, sc->intf)) < 0) {
+			printk(KERN_NOTICE
+			    "%s: usb_lock_device_for_reset failed (%d)\n",
+			    sc->name, lkr);
+		} else {
+			rc = usb_reset_device(sc->dev);
+			if (rc < 0) {
+				printk(KERN_NOTICE "%s: "
+				    "usb_lock_device_for_reset failed (%d)\n",
+				    sc->name, rc);
+			}
+
+			if (lkr)
+				usb_unlock_device(sc->dev);
+		}
+	}
+
+	/*
+	 * In theory, no commands can be running while reset is active,
+	 * so nobody can ask for another reset, and so we do not need any
+	 * queues of resets or anything. We do need a spinlock though,
+	 * to interact with block layer.
+	 */
+	spin_lock_irqsave(sc->lock, flags);
+	sc->reset = 0;
+	tasklet_schedule(&sc->tasklet);
+	list_for_each(p, &sc->luns) {
+		lun = list_entry(p, struct ub_lun, link);
+		blk_start_queue(lun->disk->queue);
+	}
+	wake_up(&sc->reset_wait);
+	spin_unlock_irqrestore(sc->lock, flags);
 }
-#endif
 
 /*
  * This is called from a process context.
  */
-static void ub_revalidate(struct ub_dev *sc)
+static void ub_revalidate(struct ub_dev *sc, struct ub_lun *lun)
 {
 
-	sc->readonly = 0;	/* XXX Query this from the device */
+	lun->readonly = 0;	/* XXX Query this from the device */
 
-	sc->capacity.nsec = 0;
-	sc->capacity.bsize = 512;
-	sc->capacity.bshift = 0;
+	lun->capacity.nsec = 0;
+	lun->capacity.bsize = 512;
+	lun->capacity.bshift = 0;
 
-	if (ub_sync_tur(sc) != 0)
+	if (ub_sync_tur(sc, lun) != 0)
 		return;			/* Not ready */
-	sc->changed = 0;
+	lun->changed = 0;
 
-	if (ub_sync_read_cap(sc, &sc->capacity) != 0) {
+	if (ub_sync_read_cap(sc, lun, &lun->capacity) != 0) {
 		/*
 		 * The retry here means something is wrong, either with the
 		 * device, with the transport, or with our code.
 		 * We keep this because sd.c has retries for capacity.
 		 */
-		if (ub_sync_read_cap(sc, &sc->capacity) != 0) {
-			sc->capacity.nsec = 0;
-			sc->capacity.bsize = 512;
-			sc->capacity.bshift = 0;
+		if (ub_sync_read_cap(sc, lun, &lun->capacity) != 0) {
+			lun->capacity.nsec = 0;
+			lun->capacity.bsize = 512;
+			lun->capacity.bshift = 0;
 		}
 	}
 }
@@ -1471,12 +1654,11 @@ static void ub_revalidate(struct ub_dev *sc)
 static int ub_bd_open(struct inode *inode, struct file *filp)
 {
 	struct gendisk *disk = inode->i_bdev->bd_disk;
-	struct ub_dev *sc;
+	struct ub_lun *lun = disk->private_data;
+	struct ub_dev *sc = lun->udev;
 	unsigned long flags;
 	int rc;
 
-	if ((sc = disk->private_data) == NULL)
-		return -ENXIO;
 	spin_lock_irqsave(&ub_lock, flags);
 	if (atomic_read(&sc->poison)) {
 		spin_unlock_irqrestore(&ub_lock, flags);
@@ -1485,27 +1667,7 @@ static int ub_bd_open(struct inode *inode, struct file *filp)
 	sc->openc++;
 	spin_unlock_irqrestore(&ub_lock, flags);
 
-	/*
-	 * This is a workaround for a specific problem in our block layer.
-	 * In 2.6.9, register_disk duplicates the code from rescan_partitions.
-	 * However, if we do add_disk with a device which persistently reports
-	 * a changed media, add_disk calls register_disk, which does do_open,
-	 * which will call rescan_paritions for changed media. After that,
-	 * register_disk attempts to do it all again and causes double kobject
-	 * registration and a eventually an oops on module removal.
-	 *
-	 * The bottom line is, Al Viro says that we should not allow
-	 * bdev->bd_invalidated to be set when doing add_disk no matter what.
-	 */
-	if (sc->first_open) {
-		if (sc->changed) {
-			sc->first_open = 0;
-			rc = -ENOMEDIUM;
-			goto err_open;
-		}
-	}
-
-	if (sc->removable || sc->readonly)
+	if (lun->removable || lun->readonly)
 		check_disk_change(inode->i_bdev);
 
 	/*
@@ -1513,12 +1675,12 @@ static int ub_bd_open(struct inode *inode, struct file *filp)
 	 * under some pretty murky conditions (a failure of READ CAPACITY).
 	 * We may need it one day.
 	 */
-	if (sc->removable && sc->changed && !(filp->f_flags & O_NDELAY)) {
+	if (lun->removable && lun->changed && !(filp->f_flags & O_NDELAY)) {
 		rc = -ENOMEDIUM;
 		goto err_open;
 	}
 
-	if (sc->readonly && (filp->f_mode & FMODE_WRITE)) {
+	if (lun->readonly && (filp->f_mode & FMODE_WRITE)) {
 		rc = -EROFS;
 		goto err_open;
 	}
@@ -1526,11 +1688,7 @@ static int ub_bd_open(struct inode *inode, struct file *filp)
 	return 0;
 
 err_open:
-	spin_lock_irqsave(&ub_lock, flags);
-	--sc->openc;
-	if (sc->openc == 0 && atomic_read(&sc->poison))
-		ub_cleanup(sc);
-	spin_unlock_irqrestore(&ub_lock, flags);
+	ub_put(sc);
 	return rc;
 }
 
@@ -1539,16 +1697,10 @@ err_open:
 static int ub_bd_release(struct inode *inode, struct file *filp)
 {
 	struct gendisk *disk = inode->i_bdev->bd_disk;
-	struct ub_dev *sc = disk->private_data;
-	unsigned long flags;
+	struct ub_lun *lun = disk->private_data;
+	struct ub_dev *sc = lun->udev;
 
-	spin_lock_irqsave(&ub_lock, flags);
-	--sc->openc;
-	if (sc->openc == 0)
-		sc->first_open = 0;
-	if (sc->openc == 0 && atomic_read(&sc->poison))
-		ub_cleanup(sc);
-	spin_unlock_irqrestore(&ub_lock, flags);
+	ub_put(sc);
 	return 0;
 }
 
@@ -1576,20 +1728,14 @@ static int ub_bd_ioctl(struct inode *inode, struct file *filp,
  */
 static int ub_bd_revalidate(struct gendisk *disk)
 {
-	struct ub_dev *sc = disk->private_data;
-
-	ub_revalidate(sc);
-	/* This is pretty much a long term P3 */
-	if (!atomic_read(&sc->poison)) {		/* Cover sc->dev */
-		printk(KERN_INFO "%s: device %u capacity nsec %ld bsize %u\n",
-		    sc->name, sc->dev->devnum,
-		    sc->capacity.nsec, sc->capacity.bsize);
-	}
+	struct ub_lun *lun = disk->private_data;
+
+	ub_revalidate(lun->udev, lun);
 
 	/* XXX Support sector size switching like in sr.c */
-	blk_queue_hardsect_size(disk->queue, sc->capacity.bsize);
-	set_capacity(disk, sc->capacity.nsec);
-	// set_disk_ro(sdkp->disk, sc->readonly);
+	blk_queue_hardsect_size(disk->queue, lun->capacity.bsize);
+	set_capacity(disk, lun->capacity.nsec);
+	// set_disk_ro(sdkp->disk, lun->readonly);
 
 	return 0;
 }
@@ -1605,9 +1751,9 @@ static int ub_bd_revalidate(struct gendisk *disk)
  */
 static int ub_bd_media_changed(struct gendisk *disk)
 {
-	struct ub_dev *sc = disk->private_data;
+	struct ub_lun *lun = disk->private_data;
 
-	if (!sc->removable)
+	if (!lun->removable)
 		return 0;
 
 	/*
@@ -1619,12 +1765,12 @@ static int ub_bd_media_changed(struct gendisk *disk)
 	 * will fail, then block layer discards the data. Since we never
 	 * spin drives up, such devices simply cannot be used with ub anyway.
 	 */
-	if (ub_sync_tur(sc) != 0) {
-		sc->changed = 1;
+	if (ub_sync_tur(lun->udev, lun) != 0) {
+		lun->changed = 1;
 		return 1;
 	}
 
-	return sc->changed;
+	return lun->changed;
 }
 
 static struct block_device_operations ub_bd_fops = {
@@ -1648,7 +1794,7 @@ static void ub_probe_done(struct ub_dev *sc, struct ub_scsi_cmd *cmd)
 /*
  * Test if the device has a check condition on it, synchronously.
  */
-static int ub_sync_tur(struct ub_dev *sc)
+static int ub_sync_tur(struct ub_dev *sc, struct ub_lun *lun)
 {
 	struct ub_scsi_cmd *cmd;
 	enum { ALLOC_SIZE = sizeof(struct ub_scsi_cmd) };
@@ -1659,27 +1805,25 @@ static int ub_sync_tur(struct ub_dev *sc)
 	init_completion(&compl);
 
 	rc = -ENOMEM;
-	if ((cmd = kmalloc(ALLOC_SIZE, GFP_KERNEL)) == NULL)
+	if ((cmd = kzalloc(ALLOC_SIZE, GFP_KERNEL)) == NULL)
 		goto err_alloc;
-	memset(cmd, 0, ALLOC_SIZE);
 
 	cmd->cdb[0] = TEST_UNIT_READY;
 	cmd->cdb_len = 6;
 	cmd->dir = UB_DIR_NONE;
 	cmd->state = UB_CMDST_INIT;
+	cmd->lun = lun;			/* This may be NULL, but that's ok */
 	cmd->done = ub_probe_done;
 	cmd->back = &compl;
 
-	spin_lock_irqsave(&sc->lock, flags);
+	spin_lock_irqsave(sc->lock, flags);
 	cmd->tag = sc->tagcnt++;
 
 	rc = ub_submit_scsi(sc, cmd);
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_unlock_irqrestore(sc->lock, flags);
 
-	if (rc != 0) {
-		printk("ub: testing ready: submit error (%d)\n", rc); /* P3 */
+	if (rc != 0)
 		goto err_submit;
-	}
 
 	wait_for_completion(&compl);
 
@@ -1697,9 +1841,11 @@ err_alloc:
 /*
  * Read the SCSI capacity synchronously (for probing).
  */
-static int ub_sync_read_cap(struct ub_dev *sc, struct ub_capacity *ret)
+static int ub_sync_read_cap(struct ub_dev *sc, struct ub_lun *lun,
+    struct ub_capacity *ret)
 {
 	struct ub_scsi_cmd *cmd;
+	struct scatterlist *sg;
 	char *p;
 	enum { ALLOC_SIZE = sizeof(struct ub_scsi_cmd) + 8 };
 	unsigned long flags;
@@ -1711,40 +1857,40 @@ static int ub_sync_read_cap(struct ub_dev *sc, struct ub_capacity *ret)
 	init_completion(&compl);
 
 	rc = -ENOMEM;
-	if ((cmd = kmalloc(ALLOC_SIZE, GFP_KERNEL)) == NULL)
+	if ((cmd = kzalloc(ALLOC_SIZE, GFP_KERNEL)) == NULL)
 		goto err_alloc;
-	memset(cmd, 0, ALLOC_SIZE);
 	p = (char *)cmd + sizeof(struct ub_scsi_cmd);
 
 	cmd->cdb[0] = 0x25;
 	cmd->cdb_len = 10;
 	cmd->dir = UB_DIR_READ;
 	cmd->state = UB_CMDST_INIT;
-	cmd->data = p;
+	cmd->nsg = 1;
+	sg = &cmd->sgv[0];
+	sg->page = virt_to_page(p);
+	sg->offset = (unsigned long)p & (PAGE_SIZE-1);
+	sg->length = 8;
 	cmd->len = 8;
+	cmd->lun = lun;
 	cmd->done = ub_probe_done;
 	cmd->back = &compl;
 
-	spin_lock_irqsave(&sc->lock, flags);
+	spin_lock_irqsave(sc->lock, flags);
 	cmd->tag = sc->tagcnt++;
 
 	rc = ub_submit_scsi(sc, cmd);
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_unlock_irqrestore(sc->lock, flags);
 
-	if (rc != 0) {
-		printk("ub: reading capacity: submit error (%d)\n", rc); /* P3 */
+	if (rc != 0)
 		goto err_submit;
-	}
 
 	wait_for_completion(&compl);
 
 	if (cmd->error != 0) {
-		printk("ub: reading capacity: error %d\n", cmd->error); /* P3 */
 		rc = -EIO;
 		goto err_read;
 	}
 	if (cmd->act_len != 8) {
-		printk("ub: reading capacity: size %d\n", cmd->act_len); /* P3 */
 		rc = -EIO;
 		goto err_read;
 	}
@@ -1758,7 +1904,6 @@ static int ub_sync_read_cap(struct ub_dev *sc, struct ub_capacity *ret)
 	case 2048:	shift = 2;	break;
 	case 4096:	shift = 3;	break;
 	default:
-		printk("ub: Bad sector size %u\n", bsize); /* P3 */
 		rc = -EDOM;
 		goto err_inv_bsize;
 	}
@@ -1778,7 +1923,7 @@ err_alloc:
 
 /*
  */
-static void ub_probe_urb_complete(struct urb *urb, struct pt_regs *pt)
+static void ub_probe_urb_complete(struct urb *urb)
 {
 	struct completion *cop = urb->context;
 	complete(cop);
@@ -1790,6 +1935,126 @@ static void ub_probe_timeout(unsigned long arg)
 	complete(cop);
 }
 
+/*
+ * Reset with a Bulk reset.
+ */
+static int ub_sync_reset(struct ub_dev *sc)
+{
+	int ifnum = sc->intf->cur_altsetting->desc.bInterfaceNumber;
+	struct usb_ctrlrequest *cr;
+	struct completion compl;
+	struct timer_list timer;
+	int rc;
+
+	init_completion(&compl);
+
+	cr = &sc->work_cr;
+	cr->bRequestType = USB_TYPE_CLASS | USB_RECIP_INTERFACE;
+	cr->bRequest = US_BULK_RESET_REQUEST;
+	cr->wValue = cpu_to_le16(0);
+	cr->wIndex = cpu_to_le16(ifnum);
+	cr->wLength = cpu_to_le16(0);
+
+	usb_fill_control_urb(&sc->work_urb, sc->dev, sc->send_ctrl_pipe,
+	    (unsigned char*) cr, NULL, 0, ub_probe_urb_complete, &compl);
+	sc->work_urb.actual_length = 0;
+	sc->work_urb.error_count = 0;
+	sc->work_urb.status = 0;
+
+	if ((rc = usb_submit_urb(&sc->work_urb, GFP_KERNEL)) != 0) {
+		printk(KERN_WARNING
+		     "%s: Unable to submit a bulk reset (%d)\n", sc->name, rc);
+		return rc;
+	}
+
+	init_timer(&timer);
+	timer.function = ub_probe_timeout;
+	timer.data = (unsigned long) &compl;
+	timer.expires = jiffies + UB_CTRL_TIMEOUT;
+	add_timer(&timer);
+
+	wait_for_completion(&compl);
+
+	del_timer_sync(&timer);
+	usb_kill_urb(&sc->work_urb);
+
+	return sc->work_urb.status;
+}
+
+/*
+ * Get number of LUNs by the way of Bulk GetMaxLUN command.
+ */
+static int ub_sync_getmaxlun(struct ub_dev *sc)
+{
+	int ifnum = sc->intf->cur_altsetting->desc.bInterfaceNumber;
+	unsigned char *p;
+	enum { ALLOC_SIZE = 1 };
+	struct usb_ctrlrequest *cr;
+	struct completion compl;
+	struct timer_list timer;
+	int nluns;
+	int rc;
+
+	init_completion(&compl);
+
+	rc = -ENOMEM;
+	if ((p = kmalloc(ALLOC_SIZE, GFP_KERNEL)) == NULL)
+		goto err_alloc;
+	*p = 55;
+
+	cr = &sc->work_cr;
+	cr->bRequestType = USB_DIR_IN | USB_TYPE_CLASS | USB_RECIP_INTERFACE;
+	cr->bRequest = US_BULK_GET_MAX_LUN;
+	cr->wValue = cpu_to_le16(0);
+	cr->wIndex = cpu_to_le16(ifnum);
+	cr->wLength = cpu_to_le16(1);
+
+	usb_fill_control_urb(&sc->work_urb, sc->dev, sc->recv_ctrl_pipe,
+	    (unsigned char*) cr, p, 1, ub_probe_urb_complete, &compl);
+	sc->work_urb.actual_length = 0;
+	sc->work_urb.error_count = 0;
+	sc->work_urb.status = 0;
+
+	if ((rc = usb_submit_urb(&sc->work_urb, GFP_KERNEL)) != 0)
+		goto err_submit;
+
+	init_timer(&timer);
+	timer.function = ub_probe_timeout;
+	timer.data = (unsigned long) &compl;
+	timer.expires = jiffies + UB_CTRL_TIMEOUT;
+	add_timer(&timer);
+
+	wait_for_completion(&compl);
+
+	del_timer_sync(&timer);
+	usb_kill_urb(&sc->work_urb);
+
+	if ((rc = sc->work_urb.status) < 0)
+		goto err_io;
+
+	if (sc->work_urb.actual_length != 1) {
+		nluns = 0;
+	} else {
+		if ((nluns = *p) == 55) {
+			nluns = 0;
+		} else {
+  			/* GetMaxLUN returns the maximum LUN number */
+			nluns += 1;
+			if (nluns > UB_MAX_LUNS)
+				nluns = UB_MAX_LUNS;
+		}
+	}
+
+	kfree(p);
+	return nluns;
+
+err_io:
+err_submit:
+	kfree(p);
+err_alloc:
+	return rc;
+}
+
 /*
  * Clear initial stalls.
  */
@@ -1816,7 +2081,6 @@ static int ub_probe_clear_stall(struct ub_dev *sc, int stalled_pipe)
 
 	usb_fill_control_urb(&sc->work_urb, sc->dev, sc->send_ctrl_pipe,
 	    (unsigned char*) cr, NULL, 0, ub_probe_urb_complete, &compl);
-	sc->work_urb.transfer_flags = 0;
 	sc->work_urb.actual_length = 0;
 	sc->work_urb.error_count = 0;
 	sc->work_urb.status = 0;
@@ -1876,9 +2140,9 @@ static int ub_get_pipes(struct ub_dev *sc, struct usb_device *dev,
 	}
 
 	if (ep_in == NULL || ep_out == NULL) {
-		printk(KERN_NOTICE "%s: device %u failed endpoint check\n",
-		    sc->name, sc->dev->devnum);
-		return -EIO;
+		printk(KERN_NOTICE "%s: failed endpoint check\n",
+		    sc->name);
+		return -ENODEV;
 	}
 
 	/* Calculate and store the pipe values */
@@ -1900,19 +2164,23 @@ static int ub_probe(struct usb_interface *intf,
     const struct usb_device_id *dev_id)
 {
 	struct ub_dev *sc;
-	request_queue_t *q;
-	struct gendisk *disk;
+	int nluns;
 	int rc;
 	int i;
 
+	if (usb_usual_check_type(dev_id, USB_US_TYPE_UB))
+		return -ENXIO;
+
 	rc = -ENOMEM;
-	if ((sc = kmalloc(sizeof(struct ub_dev), GFP_KERNEL)) == NULL)
+	if ((sc = kzalloc(sizeof(struct ub_dev), GFP_KERNEL)) == NULL)
 		goto err_core;
-	memset(sc, 0, sizeof(struct ub_dev));
-	spin_lock_init(&sc->lock);
+	sc->lock = ub_next_lock();
+	INIT_LIST_HEAD(&sc->luns);
 	usb_init_urb(&sc->work_urb);
 	tasklet_init(&sc->tasklet, ub_scsi_action, (unsigned long)sc);
 	atomic_set(&sc->poison, 0);
+	INIT_WORK(&sc->reset_work, ub_reset_task);
+	init_waitqueue_head(&sc->reset_wait);
 
 	init_timer(&sc->work_timer);
 	sc->work_timer.data = (unsigned long) sc;
@@ -1921,25 +2189,25 @@ static int ub_probe(struct usb_interface *intf,
 	ub_init_completion(&sc->work_done);
 	sc->work_done.done = 1;		/* A little yuk, but oh well... */
 
-	rc = -ENOSR;
-	if ((sc->id = ub_id_get()) == -1)
-		goto err_id;
-	snprintf(sc->name, 8, DRV_NAME "%c", sc->id + 'a');
-
 	sc->dev = interface_to_usbdev(intf);
 	sc->intf = intf;
 	// sc->ifnum = intf->cur_altsetting->desc.bInterfaceNumber;
-
 	usb_set_intfdata(intf, sc);
 	usb_get_dev(sc->dev);
-	// usb_get_intf(sc->intf);	/* Do we need this? */
+	/*
+	 * Since we give the interface struct to the block level through
+	 * disk->driverfs_dev, we have to pin it. Otherwise, block_uevent
+	 * oopses on close after a disconnect (kernels 2.6.16 and up).
+	 */
+	usb_get_intf(sc->intf);
 
-	/* XXX Verify that we can handle the device (from descriptors) */
+	snprintf(sc->name, 12, DRV_NAME "(%d.%d)",
+	    sc->dev->bus->busnum, sc->dev->devnum);
 
-	ub_get_pipes(sc, sc->dev, intf);
+	/* XXX Verify that we can handle the device (from descriptors) */
 
-	if (device_create_file(&sc->intf->dev, &dev_attr_diag) != 0)
-		goto err_diag;
+	if (ub_get_pipes(sc, sc->dev, intf) != 0)
+		goto err_dev_desc;
 
 	/*
 	 * At this point, all USB initialization is done, do upper layer.
@@ -1952,8 +2220,10 @@ static int ub_probe(struct usb_interface *intf,
 	 * This is needed to clear toggles. It is a problem only if we do
 	 * `rmmod ub && modprobe ub` without disconnects, but we like that.
 	 */
+#if 0 /* iPod Mini fails if we do this (big white iPod works) */
 	ub_probe_clear_stall(sc, sc->recv_bulk_pipe);
 	ub_probe_clear_stall(sc, sc->send_bulk_pipe);
+#endif
 
 	/*
 	 * The way this is used by the startup code is a little specific.
@@ -1970,65 +2240,92 @@ static int ub_probe(struct usb_interface *intf,
 	 * has to succeed, so we clear checks with an additional one here.
 	 * In any case it's not our business how revaliadation is implemented.
 	 */
-	for (i = 0; i < 3; i++) {	/* Retries for benh's key */
-		if ((rc = ub_sync_tur(sc)) <= 0) break;
+	for (i = 0; i < 3; i++) {  /* Retries for the schwag key from KS'04 */
+		if ((rc = ub_sync_tur(sc, NULL)) <= 0) break;
 		if (rc != 0x6) break;
 		msleep(10);
 	}
 
-	sc->removable = 1;		/* XXX Query this from the device */
-	sc->changed = 1;		/* ub_revalidate clears only */
-	sc->first_open = 1;
+	nluns = 1;
+	for (i = 0; i < 3; i++) {
+		if ((rc = ub_sync_getmaxlun(sc)) < 0)
+			break;
+		if (rc != 0) {
+			nluns = rc;
+			break;
+		}
+		msleep(100);
+	}
 
-	ub_revalidate(sc);
-	/* This is pretty much a long term P3 */
-	printk(KERN_INFO "%s: device %u capacity nsec %ld bsize %u\n",
-	    sc->name, sc->dev->devnum, sc->capacity.nsec, sc->capacity.bsize);
+	for (i = 0; i < nluns; i++) {
+		ub_probe_lun(sc, i);
+	}
+	return 0;
+
+err_dev_desc:
+	usb_set_intfdata(intf, NULL);
+	usb_put_intf(sc->intf);
+	usb_put_dev(sc->dev);
+	kfree(sc);
+err_core:
+	return rc;
+}
+
+static int ub_probe_lun(struct ub_dev *sc, int lnum)
+{
+	struct ub_lun *lun;
+	request_queue_t *q;
+	struct gendisk *disk;
+	int rc;
+
+	rc = -ENOMEM;
+	if ((lun = kzalloc(sizeof(struct ub_lun), GFP_KERNEL)) == NULL)
+		goto err_alloc;
+	lun->num = lnum;
+
+	rc = -ENOSR;
+	if ((lun->id = ub_id_get()) == -1)
+		goto err_id;
+
+	lun->udev = sc;
+
+	snprintf(lun->name, 16, DRV_NAME "%c(%d.%d.%d)",
+	    lun->id + 'a', sc->dev->bus->busnum, sc->dev->devnum, lun->num);
+
+	lun->removable = 1;		/* XXX Query this from the device */
+	lun->changed = 1;		/* ub_revalidate clears only */
+	ub_revalidate(sc, lun);
 
-	/*
-	 * Just one disk per sc currently, but maybe more.
-	 */
 	rc = -ENOMEM;
-	if ((disk = alloc_disk(UB_MINORS_PER_MAJOR)) == NULL)
+	if ((disk = alloc_disk(UB_PARTS_PER_LUN)) == NULL)
 		goto err_diskalloc;
 
-	sc->disk = disk;
-	sprintf(disk->disk_name, DRV_NAME "%c", sc->id + 'a');
-	sprintf(disk->devfs_name, DEVFS_NAME "/%c", sc->id + 'a');
+	sprintf(disk->disk_name, DRV_NAME "%c", lun->id + 'a');
 	disk->major = UB_MAJOR;
-	disk->first_minor = sc->id * UB_MINORS_PER_MAJOR;
+	disk->first_minor = lun->id * UB_PARTS_PER_LUN;
 	disk->fops = &ub_bd_fops;
-	disk->private_data = sc;
-	disk->driverfs_dev = &intf->dev;
+	disk->private_data = lun;
+	disk->driverfs_dev = &sc->intf->dev;
 
 	rc = -ENOMEM;
-	if ((q = blk_init_queue(ub_bd_rq_fn, &sc->lock)) == NULL)
+	if ((q = blk_init_queue(ub_request_fn, sc->lock)) == NULL)
 		goto err_blkqinit;
 
 	disk->queue = q;
 
-        // blk_queue_bounce_limit(q, hba[i]->pdev->dma_mask);
+	blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
 	blk_queue_max_hw_segments(q, UB_MAX_REQ_SG);
 	blk_queue_max_phys_segments(q, UB_MAX_REQ_SG);
-	// blk_queue_segment_boundary(q, CARM_SG_BOUNDARY);
+	blk_queue_segment_boundary(q, 0xffffffff);	/* Dubious. */
 	blk_queue_max_sectors(q, UB_MAX_SECTORS);
-	blk_queue_hardsect_size(q, sc->capacity.bsize);
-
-	/*
-	 * This is a serious infraction, caused by a deficiency in the
-	 * USB sg interface (usb_sg_wait()). We plan to remove this once
-	 * we get mileage on the driver and can justify a change to USB API.
-	 * See blk_queue_bounce_limit() to understand this part.
-	 *
-	 * XXX And I still need to be aware of the DMA mask in the HC.
-	 */
-	q->bounce_pfn = blk_max_low_pfn;
-	q->bounce_gfp = GFP_NOIO;
+	blk_queue_hardsect_size(q, lun->capacity.bsize);
 
-	q->queuedata = sc;
+	lun->disk = disk;
+	q->queuedata = lun;
+	list_add(&lun->link, &sc->luns);
 
-	set_capacity(disk, sc->capacity.nsec);
-	if (sc->removable)
+	set_capacity(disk, lun->capacity.nsec);
+	if (lun->removable)
 		disk->flags |= GENHD_FL_REMOVABLE;
 
 	add_disk(disk);
@@ -2038,27 +2335,29 @@ static int ub_probe(struct usb_interface *intf,
 err_blkqinit:
 	put_disk(disk);
 err_diskalloc:
-	device_remove_file(&sc->intf->dev, &dev_attr_diag);
-err_diag:
-	usb_set_intfdata(intf, NULL);
-	// usb_put_intf(sc->intf);
-	usb_put_dev(sc->dev);
-	spin_lock_irq(&ub_lock);
-	ub_id_put(sc->id);
-	spin_unlock_irq(&ub_lock);
+	ub_id_put(lun->id);
 err_id:
-	kfree(sc);
-err_core:
+	kfree(lun);
+err_alloc:
 	return rc;
 }
 
 static void ub_disconnect(struct usb_interface *intf)
 {
 	struct ub_dev *sc = usb_get_intfdata(intf);
-	struct gendisk *disk = sc->disk;
-	request_queue_t *q = disk->queue;
+	struct list_head *p;
+	struct ub_lun *lun;
 	unsigned long flags;
 
+	/*
+	 * Prevent ub_bd_release from pulling the rug from under us.
+	 * XXX This is starting to look like a kref.
+	 * XXX Why not to take this ref at probe time?
+	 */
+	spin_lock_irqsave(&ub_lock, flags);
+	sc->openc++;
+	spin_unlock_irqrestore(&ub_lock, flags);
+
 	/*
 	 * Fence stall clearnings, operations triggered by unlinkings and so on.
 	 * We do not attempt to unlink any URBs, because we do not trust the
@@ -2066,6 +2365,11 @@ static void ub_disconnect(struct usb_interface *intf)
 	 */
 	atomic_set(&sc->poison, 1);
 
+	/*
+	 * Wait for reset to end, if any.
+	 */
+	wait_event(sc->reset_wait, !sc->reset);
+
 	/*
 	 * Blow away queued commands.
 	 *
@@ -2075,14 +2379,13 @@ static void ub_disconnect(struct usb_interface *intf)
 	 * and the whole queue drains. So, we just use this code to
 	 * print warnings.
 	 */
-	spin_lock_irqsave(&sc->lock, flags);
+	spin_lock_irqsave(sc->lock, flags);
 	{
 		struct ub_scsi_cmd *cmd;
 		int cnt = 0;
-		while ((cmd = ub_cmdq_pop(sc)) != NULL) {
+		while ((cmd = ub_cmdq_peek(sc)) != NULL) {
 			cmd->error = -ENOTCONN;
 			cmd->state = UB_CMDST_DONE;
-			ub_cmdtr_state(sc, cmd);
 			ub_cmdq_pop(sc);
 			(*cmd->done)(sc, cmd);
 			cnt++;
@@ -2092,33 +2395,32 @@ static void ub_disconnect(struct usb_interface *intf)
 			    "%d was queued after shutdown\n", sc->name, cnt);
 		}
 	}
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_unlock_irqrestore(sc->lock, flags);
 
 	/*
-	 * Unregister the upper layer, this waits for all commands to end.
+	 * Unregister the upper layer.
 	 */
-	if (disk->flags & GENHD_FL_UP)
-		del_gendisk(disk);
-	if (q)
-		blk_cleanup_queue(q);
+	list_for_each (p, &sc->luns) {
+		lun = list_entry(p, struct ub_lun, link);
+		del_gendisk(lun->disk);
+		/*
+		 * I wish I could do:
+		 *    set_bit(QUEUE_FLAG_DEAD, &q->queue_flags);
+		 * As it is, we rely on our internal poisoning and let
+		 * the upper levels to spin furiously failing all the I/O.
+		 */
+	}
 
 	/*
-	 * We really expect blk_cleanup_queue() to wait, so no amount
-	 * of paranoya is too much.
-	 *
-	 * Taking a lock on a structure which is about to be freed
-	 * is very nonsensual. Here it is largely a way to do a debug freeze,
-	 * and a bracket which shows where the nonsensual code segment ends.
-	 *
 	 * Testing for -EINPROGRESS is always a bug, so we are bending
 	 * the rules a little.
 	 */
-	spin_lock_irqsave(&sc->lock, flags);
+	spin_lock_irqsave(sc->lock, flags);
 	if (sc->work_urb.status == -EINPROGRESS) {	/* janitors: ignore */
 		printk(KERN_WARNING "%s: "
 		    "URB is active after disconnect\n", sc->name);
 	}
-	spin_unlock_irqrestore(&sc->lock, flags);
+	spin_unlock_irqrestore(sc->lock, flags);
 
 	/*
 	 * There is virtually no chance that other CPU runs times so long
@@ -2132,21 +2434,10 @@ static void ub_disconnect(struct usb_interface *intf)
 	 * and no URBs left in transit.
 	 */
 
-	device_remove_file(&sc->intf->dev, &dev_attr_diag);
-	usb_set_intfdata(intf, NULL);
-	// usb_put_intf(sc->intf);
-	sc->intf = NULL;
-	usb_put_dev(sc->dev);
-	sc->dev = NULL;
-
-	spin_lock_irqsave(&ub_lock, flags);
-	if (sc->openc == 0)
-		ub_cleanup(sc);
-	spin_unlock_irqrestore(&ub_lock, flags);
+	ub_put(sc);
 }
 
-struct usb_driver ub_driver = {
-	.owner =	THIS_MODULE,
+static struct usb_driver ub_driver = {
 	.name =		"ub",
 	.probe =	ub_probe,
 	.disconnect =	ub_disconnect,
@@ -2156,21 +2447,21 @@ struct usb_driver ub_driver = {
 static int __init ub_init(void)
 {
 	int rc;
+	int i;
 
-	/* P3 */ printk("ub: sizeof ub_scsi_cmd %zu ub_dev %zu\n",
-			sizeof(struct ub_scsi_cmd), sizeof(struct ub_dev));
+	for (i = 0; i < UB_QLOCK_NUM; i++)
+		spin_lock_init(&ub_qlockv[i]);
 
 	if ((rc = register_blkdev(UB_MAJOR, DRV_NAME)) != 0)
 		goto err_regblkdev;
-	devfs_mk_dir(DEVFS_NAME);
 
 	if ((rc = usb_register(&ub_driver)) != 0)
 		goto err_register;
 
+	usb_usual_set_present(USB_US_TYPE_UB);
 	return 0;
 
 err_register:
-	devfs_remove(DEVFS_NAME);
 	unregister_blkdev(UB_MAJOR, DRV_NAME);
 err_regblkdev:
 	return rc;
@@ -2180,8 +2471,8 @@ static void __exit ub_exit(void)
 {
 	usb_deregister(&ub_driver);
 
-	devfs_remove(DEVFS_NAME);
 	unregister_blkdev(UB_MAJOR, DRV_NAME);
+	usb_usual_clear_present(USB_US_TYPE_UB);
 }
 
 module_init(ub_init);