qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH] Flush pending AIO on reboot and shutdown.


From: Anthony Liguori
Subject: Re: [Qemu-devel] [PATCH] Flush pending AIO on reboot and shutdown.
Date: Wed, 13 Aug 2008 14:03:13 -0500
User-agent: Thunderbird 2.0.0.16 (X11/20080723)

Gleb Natapov wrote:
On Wed, Aug 13, 2008 at 10:53:57AM -0500, Anthony Liguori wrote:
Now that I think about it, I think your fixing the wrong problem. The issue isn't that the IO requests need to be completed, but that they *will* complete which means that the IDE driver will receive a callback for a request that it no longer knows about (because it was reset). So
Not exactly. IDE still knows about the request after reset (actually currently
IDE doesn't know about reset), but the IDE code is written in such a way that there can be only one outstanding request in progress. When guest issues
another request before previous request is completed global data is
modified and first request start to use wrong data and consequences are
unpredictable. It may be crash, image corruption, infinity recursion.
The fact that IDE code allows to issue another request wile IDE is still
busy is also a bug.

Yeah, let's fix this properly. I worry that this could be a DoS on the part of the guest (or even worse).

what we really need to do is modify the IDE device such that when it is reset, it cancels any pending requests.

The fact that this reset happens as a consequence of a system reset is really just a coincidence.
That will also solve the problem of cause, but what bother me is that we
consciously drop user data that we can easily save. Why? Real HW tries
hard to save every bit of user data and we just decided to drop it. The
difference between cancel or complete a request may be corrupted or not
corrupted file system after a crash.

I don't think of this as "saving user data". From the guest's perspective, data is only written to disk after a write completion has be issued. I think it's worse for data to end up being written to disk without that completion ever being seen by the guest.

Regards,

Anthony Liguori

I'll send updated patch.

--
                        Gleb.







reply via email to

[Prev in Thread] Current Thread [Next in Thread]