ubi: Expose the bitrot interface

Using UBI_IOCRPEB and UBI_IOCSPEB userspace can force
reading and scrubbing of PEBs.

In case of bitflips UBI will automatically take action
and move data to a different PEB.
This interface allows a daemon to foster your NAND.
Signed-off-by: Richard Weinberger <richard@nod.at>

drivers/mtd/ubi/cdev.c

@@ -974,6 +974,36 @@ static long ubi_cdev_ioctl(struct file *file, unsigned int cmd,
 		break;
 	}
 
+	/* Check a specific PEB for bitflips and scrub it if needed */
+	case UBI_IOCRPEB:
+	{
+		int pnum;
+
+		err = get_user(pnum, (__user int32_t *)argp);
+		if (err) {
+			err = -EFAULT;
+			break;
+		}
+
+		err = ubi_bitflip_check(ubi, pnum, 0);
+		break;
+	}
+
+	/* Force scrubbing for a specific PEB */
+	case UBI_IOCSPEB:
+	{
+		int pnum;
+
+		err = get_user(pnum, (__user int32_t *)argp);
+		if (err) {
+			err = -EFAULT;
+			break;
+		}
+
+		err = ubi_bitflip_check(ubi, pnum, 1);
+		break;
+	}
+
 	default:
 		err = -ENOTTY;
 		break;

drivers/mtd/ubi/ubi.h

@@ -929,6 +929,7 @@ int ubi_wl_put_fm_peb(struct ubi_device *ubi, struct ubi_wl_entry *used_e,
 int ubi_is_erase_work(struct ubi_work *wrk);
 void ubi_refill_pools(struct ubi_device *ubi);
 int ubi_ensure_anchor_pebs(struct ubi_device *ubi);
+int ubi_bitflip_check(struct ubi_device *ubi, int pnum, int force_scrub);
 
 /* io.c */
 int ubi_io_read(const struct ubi_device *ubi, void *buf, int pnum, int offset,

drivers/mtd/ubi/wl.c

@@ -1440,6 +1440,150 @@ int ubi_wl_flush(struct ubi_device *ubi, int vol_id, int lnum)
 	return err;
 }
 
+static bool scrub_possible(struct ubi_device *ubi, struct ubi_wl_entry *e)
+{
+	if (in_wl_tree(e, &ubi->scrub))
+		return false;
+	else if (in_wl_tree(e, &ubi->erroneous))
+		return false;
+	else if (ubi->move_from == e)
+		return false;
+	else if (ubi->move_to == e)
+		return false;
+
+	return true;
+}
+
+/**
+ * ubi_bitflip_check - Check an eraseblock for bitflips and scrub it if needed.
+ * @ubi: UBI device description object
+ * @pnum: the physical eraseblock to schedule
+ * @force: dont't read the block, assume bitflips happened and take action.
+ *
+ * This function reads the given eraseblock and checks if bitflips occured.
+ * In case of bitflips, the eraseblock is scheduled for scrubbing.
+ * If scrubbing is forced with @force, the eraseblock is not read,
+ * but scheduled for scrubbing right away.
+ *
+ * Returns:
+ * %EINVAL, PEB is out of range
+ * %ENOENT, PEB is no longer used by UBI
+ * %EBUSY, PEB cannot be checked now or a check is currently running on it
+ * %EAGAIN, bit flips happened but scrubbing is currently not possible
+ * %EUCLEAN, bit flips happened and PEB is scheduled for scrubbing
+ * %0, no bit flips detected
+ */
+int ubi_bitflip_check(struct ubi_device *ubi, int pnum, int force)
+{
+	int err;
+	struct ubi_wl_entry *e;
+
+	if (pnum < 0 || pnum >= ubi->peb_count) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	/*
+	 * Pause all parallel work, otherwise it can happen that the
+	 * erase worker frees a wl entry under us.
+	 */
+	down_write(&ubi->work_sem);
+
+	/*
+	 * Make sure that the wl entry does not change state while
+	 * inspecting it.
+	 */
+	spin_lock(&ubi->wl_lock);
+	e = ubi->lookuptbl[pnum];
+	if (!e) {
+		spin_unlock(&ubi->wl_lock);
+		err = -ENOENT;
+		goto out_resume;
+	}
+
+	/*
+	 * Does it make sense to check this PEB?
+	 */
+	if (!scrub_possible(ubi, e)) {
+		spin_unlock(&ubi->wl_lock);
+		err = -EBUSY;
+		goto out_resume;
+	}
+	spin_unlock(&ubi->wl_lock);
+
+	if (!force) {
+		mutex_lock(&ubi->buf_mutex);
+		err = ubi_io_read(ubi, ubi->peb_buf, pnum, 0, ubi->peb_size);
+		mutex_unlock(&ubi->buf_mutex);
+	}
+
+	if (err == UBI_IO_BITFLIPS || force) {
+		/*
+		 * Okay, bit flip happened, let's figure out what we can do.
+		 */
+		spin_lock(&ubi->wl_lock);
+
+		/*
+		 * Recheck. We released wl_lock, UBI might have killed the
+		 * wl entry under us.
+		 */
+		e = ubi->lookuptbl[pnum];
+		if (!e) {
+			spin_unlock(&ubi->wl_lock);
+			err = -ENOENT;
+			goto out_resume;
+		}
+
+		/*
+		 * Need to re-check state
+		 */
+		if (!scrub_possible(ubi, e)) {
+			spin_unlock(&ubi->wl_lock);
+			err = -EBUSY;
+			goto out_resume;
+		}
+
+		if (in_pq(ubi, e)) {
+			prot_queue_del(ubi, e->pnum);
+			wl_tree_add(e, &ubi->scrub);
+			spin_unlock(&ubi->wl_lock);
+
+			err = ensure_wear_leveling(ubi, 1);
+		} else if (in_wl_tree(e, &ubi->used)) {
+			rb_erase(&e->u.rb, &ubi->used);
+			wl_tree_add(e, &ubi->scrub);
+			spin_unlock(&ubi->wl_lock);
+
+			err = ensure_wear_leveling(ubi, 1);
+		} else if (in_wl_tree(e, &ubi->free)) {
+			rb_erase(&e->u.rb, &ubi->free);
+			ubi->free_count--;
+			spin_unlock(&ubi->wl_lock);
+
+			/*
+			 * This PEB is empty we can schedule it for
+			 * erasure right away. No wear leveling needed.
+			 */
+			err = schedule_erase(ubi, e, UBI_UNKNOWN, UBI_UNKNOWN,
+					     force ? 0 : 1, true);
+		} else {
+			spin_unlock(&ubi->wl_lock);
+			err = -EAGAIN;
+		}
+
+		if (!err && !force)
+			err = -EUCLEAN;
+	} else {
+		err = 0;
+	}
+
+out_resume:
+	up_write(&ubi->work_sem);
+out:
+
+	return err;
+}
+
 /**
  * tree_destroy - destroy an RB-tree.
  * @ubi: UBI device description object

include/uapi/mtd/ubi-user.h

@@ -171,6 +171,11 @@
 /* Re-name volumes */
 #define UBI_IOCRNVOL _IOW(UBI_IOC_MAGIC, 3, struct ubi_rnvol_req)
 
+/* Read the specified PEB and scrub it if there are bitflips */
+#define UBI_IOCRPEB _IOW(UBI_IOC_MAGIC, 4, __s32)
+/* Force scrubbing on the specified PEB */
+#define UBI_IOCSPEB _IOW(UBI_IOC_MAGIC, 5, __s32)
+
 /* ioctl commands of the UBI control character device */
 
 #define UBI_CTRL_IOC_MAGIC 'o'