On Thu, 3 Sep 2009, OGAWA Hirofumi wrote: > > If I'm not missing, I think it doesn't have big change with old > code. But I would need to check more deeply. The thing is, the old pty code pushed _directly_ to the receiving ldisc, with no buffering. I'm not entirely sure why Alan felt it needed changing, but moving over to the generic tty buffering code did get rid of some duplicate logic, and the locking is now done in one place, so that's probably the main reason. Anyway, the old pty code would be entirely synchronous, and would do the ld->ops->receive_buf(to, buf, NULL, c); to push the data all the way to the receive side frm pty_write(). So with the old code, the destination "receive_room" was always accurate, because both the reading side and the writing side basically accessed it directly. With the new code, it all goes through tty_buffer.c, and the bugs have been mostly about the receiving side not seeing all the data in the buffers. And those buffers simply didn't use to exist before. > Um.., If "receive_room == 0 && tty->read_cnt == 0" is possible, I wonder > why reverting buffer handling fixes the problem. In the old code, if 'receive_room' was zero, then the writer would simply stop writing (no buffers in between). So in the old code, you could never get into a situation where receive_room was zero and there was still pending data. At least that's how I read the situation. If I'm right, I'm hoping that the patch I sent out fixes it, and if so, we'll do that for 2.6.31 (and then after that maybe re-think whether the extra buffering is worth all this pain). And if it _doesn't_ fix it, then I think we'll just have to revert the commits in question. We won't have time to root-cause it if the above isn't it. Linus -- To unsubscribe from this list: send the line "unsubscribe kernel-testers" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
