Re: [PATCH v3 2/3] uio_pci_generic: add MSI/MSI-X support

[Greg KH <gregkh@linuxfoundation.org>]

On Mon, Oct 05, 2015 at 02:09:32PM +0300, Avi Kivity wrote:
> On 10/05/2015 01:57 PM, Greg KH wrote:
> >On Mon, Oct 05, 2015 at 01:48:39PM +0300, Vlad Zolotarov wrote:
> >>
> >>On 10/05/15 10:56, Greg KH wrote:
> >>>On Mon, Oct 05, 2015 at 10:41:39AM +0300, Vlad Zolotarov wrote:
> >>>>>>+struct msix_info {
> >>>>>>+	int num_irqs;
> >>>>>>+	struct msix_entry *table;
> >>>>>>+	struct uio_msix_irq_ctx {
> >>>>>>+		struct eventfd_ctx *trigger;	/* MSI-x vector to eventfd */
> >>>>>Why are you using eventfd for msi vectors?  What's the reason for
> >>>>>needing this?
> >>>>A small correction - for MSI-X vectors. There may be only one MSI vector per
> >>>>PCI function and if it's used it would use the same interface as a legacy
> >>>>INT#x interrupt uses at the moment.
> >>>>So, for MSI-X case the reason is that there may be (in most cases there will
> >>>>be) more than one interrupt vector. Thus, as I've explained in a PATCH1
> >>>>thread we need a way to indicated each of them separately. eventfd seems
> >>>>like a good way of doing so. If u have better ideas, pls., share.
> >>>You need to document what you are doing here, I don't see any
> >>>explaination for using eventfd at all.
> >>>
> >>>And no, I don't know of any other solution as I don't know what you are
> >>>trying to do here (hint, the changelog didn't document it...)
> >>>
> >>>>>You haven't documented how this api works at all, you are going to have
> >>>>>to a lot more work to justify this, as this greatly increases the
> >>>>>complexity of the user/kernel api in unknown ways.
> >>>>I actually do documented it a bit. Pls., check PATCH3 out.
> >>>That provided no information at all about how to use the api.
> >>>
> >>>If it did, you would see that your api is broken for 32/64bit kernels
> >>>and will fall over into nasty pieces the first time you try to use it
> >>>there, which means it hasn't been tested at all :(
> >>It has been tested of course ;)
> >>I tested it only in 64 bit environment however where both kernel and user
> >>space applications were compiled on the same machine with the same compiler
> >>and it could be that "int" had the same number of bytes both in kernel and
> >>in user space application. Therefore it worked perfectly - I patched DPDK to
> >>use the new uio_pci_generic MSI-X API to test this and I have verified that
> >>all 3 interrupt modes work: MSI-X with SR-IOV VF device in Amazon EC2 guest
> >>and INT#x and MSI with a PF device on bare metal server.
> >>
> >>However I agree using uint32_t for "vec" and "fd" would be much more
> >>correct.
> >I don't think file descriptors are __u32 on a 64bit arch, are they?
> >
> >And NEVER use the _t types in kernel code, the namespaces is all wrong
> >and it is not applicable for us, sorry.
> 
> Wasn't the real reason that they aren't defined (or reserved) by C89, and
> therefore could clash with a user identifier, rather than some inherent
> wrongness?

Kind of, my memory is vague.  There's a great rant from Linus about why
they don't work in the kernel somewhere in the lkml archives...