exposing the sdev_target in sysfs

The aim of this patch is to give us somewhere to hang the per target timer that emulex needs.

However, it will break the current sysfs layout. devices will now appear something like this:

lrwxrwxrwx 1 root root 0 Sep 3 18:31 /sys/class/scsi_device/0:0:5:0/device -> ../../../devices/parisc8/parisc8:0/pci0000:00/0000:00:13.0/host0/target0:0:5/0:0:5:0

(note the extra target0:0:5 in there).

which may break some of the user land tools (although you're all using the scsi class interface to access your devices, aren't you? in which case you won't notice any problems).

Comments welcome.

James

drivers/scsi/scsi_sysfs.c +27 -0

+int scsi_sysfs_target_initialize(struct scsi_device *sdev)
+{
+	struct scsi_target *starget = NULL;
+	struct scsi_device *sdev_sibling;
+	unsigned long flags;
+
+	spin_lock_irqsave(sdev->host->host_lock, flags);
+	/*
+	 * Search for an existing target for this sdev.
+	 */
+	list_for_each_entry(sdev_sibling, &sdev->same_target_siblings,
+			    same_target_siblings) {
+		if (sdev_sibling->sdev_gendev.parent != NULL) {
+			starget = scsi_target(sdev_sibling);
+			break;
+		}
+	}
+	if (!starget) {
+		struct device *dev;
+		int error;
+
+		starget = kmalloc(sizeof(*starget), GFP_ATOMIC);
+		if (!starget) {
+			printk(KERN_ERR "%s: allocation failure\n", __FUNCTION__);
+			spin_unlock_irqrestore(sdev->host->host_lock,
+					       flags);
+			return -ENOMEM;

drivers/scsi/scsi_sysfs.c +8 -0

+			spin_unlock_irqrestore(sdev->host->host_lock, flags);
+			printk(KERN_ERR "Target device_add failed\n");
+			return error;
+		}
+	}
+	get_device(&starget->dev);
+	sdev->sdev_gendev.parent = &starget->dev;
+	spin_unlock_irqrestore(sdev->host->host_lock, flags);
 	return 0;
 }

device_add calls kobject_add which can sleep.

Quoted 1 line · Mike Christie

device_add calls kobject_add which can sleep.

Yes, I'll move it outside the lock.

I suppose I should add a comment somewhere about our critical dependence on sequential LUN scanning as well just in case someone tries to parallelise it.

sequential on a given bus, rather than globally, right? There's no problem with scanning different busses simultanously, is there?

The aim of this patch is to give us somewhere to hang the per target timer that emulex needs.

However, it will break the current sysfs layout. devices will now appear something like this:

lrwxrwxrwx 1 root root 0 Sep 3 18:31 /sys/class/scsi_device/0:0:5:0/device -> ../../../devices/parisc8/parisc8:0/pci0000:00/0000:00:13.0/host0/target0:0:5/0:0:5:0

(note the extra target0:0:5 in there).

which may break some of the user land tools (although you're all using the scsi class interface to access your devices, aren't you? in which case you won't notice any problems).