Re: [KVM PATCH v2 0/2] irqfd: use POLLHUP notification for close()

From: Avi Kivity
Date: Sun Jun 14 2009 - 08:52:10 EST

Gregory Haskins wrote:
Michael S. Tsirkin wrote:
[ Resending with correct address for Davide. Pls don't reply
to the original one, you'll get bounces. ]

On Thu, Jun 04, 2009 at 08:48:02AM -0400, Gregory Haskins wrote:
(Applies to kvm.git/master:25deed73)

Please see the header for 2/2 for a description. This patch series has been
fully tested and appears to be working correctly.

[Review notes:
*) Paul has looked at the SRCU design and, to my knowledge, didn't find
any holes.
*) Michael, Avi, and myself agree that while the removal of the DEASSIGN
vector is not desirable, the fix on close() is more important in
the short-term. We can always add DEASSIGN support again in the
future with a CAP bit.
So, I've been thinking about this, and this approach has another
problem: it depends on pollhup on close which is AFAIK an
eventfd-specific feature.

Thats ok with me, as we are already married to eventfd for other reasons
(see eventfd_fget()).

This will prevent us from supporting polling
other useful file types, such as sockets and pipes, down the road, with
this interface.
I am thinking that we would add explicit support in the future if there
are other fd types that might want to also inject interrupts. For
instance, perhaps POLLHUP is added to pipes if/when they are patched as
a valid transport for irqfd. Or perhaps irqfd is abstracted such that
eventfd_fget/POLLHUP are eventfd specific assign/deassign implementation

Another option is that we s/irqfd/irq-eventfd to leave room for other
interfaces like irq-pollfd, irq-socketfd, etc. IOW, there is no reason
to make the current irqfd code "one-fd-interface to rule them all" per
se. The real abstraction is the kvm_set_irq() + gsi interface anyway. The current irqfd code is a thin shim in front of that. Perhaps each fd
type would be better served with code to specifically handle each type,
for its hard to predict what the requirements for translating, say, a
pipe-write into a gsi-inject will be apriori.

I don't see a reason to avoid a monogamous relationship with eventfd as it exactly captures the essence of an raising an interrupt: events are coalesced and it doesn't block. Since irqfd will rarely work by itself (need a separate data channel), having things like a tcp socket inject an interrupt are, while exotic, fairly useless.

I didn't realise these implications when I suggested deassign on close.
To me, it now looks like we are better off reverting this patch.
We can later add 'deassign on close' support with CAP bit after all :)

Avi, Gregory, what's your take?

I like the design with the single-call close in place, so my vote is to
keep it as it is now.

We could work around it by allocating a gsi private to the eventfd, and when we want to mask the gsi, simply drop all its routes. Hacky.

error compiling committee.c: too many arguments to function

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at