On Wed, Jul 08, 2020 at 06:28:05PM +0530, Kanchan Joshi wrote:

The last thing is about the flag used to trigger this processing. Will it be
fine to intoduce new flag (RWF_APPEND2 or RWF_APPEND_OFFSET)
instead of using RWF_APPEND?

New flag will do what RWF_APPEND does and will also return the
written-location (and therefore expects pointer setup in application).

I think it's simpler to understand if it's called RWF_INDIRECT_OFFSET
Then it'd look like:

+	rwf_t rwf = READ_ONCE(sqe->rw_flags);
...
-	iocb->ki_pos = READ_ONCE(sqe->off);
+	if (rwf & RWF_INDIRECT_OFFSET) {
+		loff_t __user *loffp = u64_to_user_ptr(sqe->addr2);
+
+		if (get_user(iocb->ki_pos, loffp)
+			return -EFAULT;
+		iocb->ki_loffp = loffp;
+	} else {
+		iocb->ki_pos = READ_ONCE(sqe->off);
+	}
...
-	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
+	ret = kiocb_set_rw_flags(kiocb, rwf);