zvol_write() can use dmu_tx_hold_write_by_dnode()

We can improve the performance of writes to zvols by using
dmu_tx_hold_write_by_dnode() instead of dmu_tx_hold_write().  This
reduces lock contention on the first block of the dnode object, and also
reduces the amount of CPU needed.  The benefit will be highest with
multi-threaded async writes (i.e. writes that don't call zil_commit()).

Reviewed-by: Jorgen Lundman <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Tony Nguyen <[email protected]>
Signed-off-by: Matthew Ahrens <[email protected]>
Closes openzfs#10184

module/os/linux/zfs/zvol_os.c

@@ -143,7 +143,7 @@ zvol_write(void *arg)
 		if (bytes > volsize - off)	/* don't write past the end */
 			bytes = volsize - off;
 
-		dmu_tx_hold_write(tx, ZVOL_OBJ, off, bytes);
+		dmu_tx_hold_write_by_dnode(tx, zv->zv_dn, off, bytes);
 
 		/* This will only fail for ENOSPC */
 		error = dmu_tx_assign(tx, TXG_WAIT);