On Wed, 2014-04-02 at 04:18 +0000, Gonglei (Arei) wrote:
> > > Hi,
> > >
> > > I have a problem about SR-IOV pass-through.
> > >
> > > The PF is Emulex Corporation OneConnect NIC (Lancer)(rev 10),
> > > and the VF pci config is as follow:
> > >
> > > LINUX:/sys/bus/pci/devices/0000:04:00.6 # hexdump config
> > > 0000000 ffff ffff 0000 0010 0010 0200 0000 0080
> > > 0000010 0000 0000 0000 0000 0000 0000 0000 0000
> > > 0000020 0000 0000 0000 0000 0000 0000 10df e264
> > > 0000030 0000 0000 0054 0000 0000 0000 0000 0000
> > > 0000040 0000 0000 0008 0000 0000 0000 0000 0000
> > > 0000050 0000 0000 6009 0008 2b41 c002 0000 0000
> > > 0000060 7805 018a 0000 0000 0000 0000 0000 0000
> > > 0000070 0000 0000 0000 0000 8411 03ff 4000 0000
> > > 0000080 3400 0000 9403 0000 0000 0000 0000 0000
> > > 0000090 0000 0000 0010 0002 8724 1000 0000 0000
> > > 00000a0 dc83 0041 0000 0000 0000 0000 0000 0000
> > > 00000b0 0000 0000 0000 0000 001f 0010 0000 0000
> > > 00000c0 000e 0000 0000 0000 0000 0000 0000 0000
> > > 00000d0 0000 0000 0000 0000 0000 0000 0000 0000
> > >
> > > We can see the msix_max is 0x3ff and msix_table_entry is 0x4000 (4 pages).
> > But QEMU
> > > only mmap MSIX_PAGE_SIZE memory for all pci devices in funciton
> > assigned_dev_register_msix_mmio,
> > > meanwhile the set the one page memmory to zero, so the rest memory will
> > be random value
> > > (maybe etnry.data is not 0).
> > >
> > > In function assigned_dev_update_msix_mmio maybe occur the issue of
> > entry_nr > 256,
> > > and the kmod reports the EINVAL error.
> > >
> > > My patch fix this issue which alloc memory according to the real size of
> > > pci
> > device config.
> > >
> > > Any ideas? Thnaks.
> >
> > Isn't this already fixed if you use vfio-pci? pci-assign is not well
> > supported anymore.
> >
> No, I haven't tried it use vfio-pci. Maybe I will have a try.
>
> > > Signed-off-by: Gonglei <arei.gong...@huawei.com>
> > > ---
> > > hw/i386/kvm/pci-assign.c | 24 +++++++++++++++++++-----
> > > 1 files changed, 19 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/hw/i386/kvm/pci-assign.c b/hw/i386/kvm/pci-assign.c
> > > index a825871..daa191c 100644
> > > --- a/hw/i386/kvm/pci-assign.c
> > > +++ b/hw/i386/kvm/pci-assign.c
> > > @@ -1591,10 +1591,6 @@ static void
> > assigned_dev_msix_reset(AssignedDevice *dev)
> > > MSIXTableEntry *entry;
> > > int i;
> > >
> > > - if (!dev->msix_table) {
> > > - return;
> > > - }
> > > -
> >
> > How would this not result in a segfault if the assigned device doesn't
> > support msix?
> >
> Actually I move the judge in function assigned_dev_register_msix_mmio.
> Because assigned_dev_register_msix_mmio do not address the return value,
> if dev->msix_table is null, this will result a segfault. Right?
I see the confusion, there is a bug there but I think it should be fixed
by setting msix_table to NULL in the MAP_FAILED case. The intent of
assigned_dev_msix_reset() is to put the msix table in the state it would
be in at reset. Without a memset() that doesn't happen. I believe the
reason we test msix_table in assigned_dev_msix_reset() is so that the
function could be called from anywhere, such as reset_assigned_device()
even though it's currently not called from there. So, if the memset()
gets removed, then the whole function should be dissolved. I'd prefer
to keep it and store or recalculate the size for memset.
> > > memset(dev->msix_table, 0, MSIX_PAGE_SIZE);
> >
> > This memset may no longer cover the entire table
> >
> Yeah, this memset is useless. Do it in assigned_dev_register_msix_mmio.
Not useless, see above.
> > >
> > > for (i = 0, entry = dev->msix_table; i < dev->msix_max; i++,
> > > entry++) {
> > > @@ -1604,13 +1600,31 @@ static void
> > assigned_dev_msix_reset(AssignedDevice *dev)
> > >
> > > static int assigned_dev_register_msix_mmio(AssignedDevice *dev)
> > > {
> > > - dev->msix_table = mmap(NULL, MSIX_PAGE_SIZE,
> > PROT_READ|PROT_WRITE,
> > > + int nr_pages;
> > > + int size;
> > > + int entry_per_page = MSIX_PAGE_SIZE / sizeof(struct MSIXTableEntry);
> > > +
> > > + if (dev->msix_max > entry_per_page) {
> > > + nr_pages = dev->msix_max / entry_per_page;
> > > + if (dev->msix_max % entry_per_page) {
> > > + nr_pages += 1;
> > > + }
> > > + } else {
> > > + nr_pages = 1;
> > > + }
> > > +
> > > + size = MSIX_PAGE_SIZE * nr_pages;
> >
> > Agree with the ROUND_UP comments:
> >
> > size = ROUND_UP(dev->msix_max * sizeof(struct MSIXTableEntry),
> > MSIX_PAGE_SIZE);
> >
> Nice. I don't know the macro before. Thank you so much!
> > > + dev->msix_table = mmap(NULL, size, PROT_READ|PROT_WRITE,
> > > MAP_ANONYMOUS|MAP_PRIVATE, 0, 0);
> > > if (dev->msix_table == MAP_FAILED) {
> > > error_report("fail allocate msix_table! %s", strerror(errno));
> > > return -EFAULT;
> > > }
> > > + if (!dev->msix_table) {
> > > + return -EFAULT;
> > > + }
> >
> > This is a bogus test for mmap(2)
> >
> > >
> > > + memset(dev->msix_table, 0, size);
> >
> > This is unnecessary when assigned_dev_msix_reset() is fixed to memset
> > the whole mmap.
> >
> > > assigned_dev_msix_reset(dev);
> > >
> > > memory_region_init_io(&dev->mmio, OBJECT(dev),
> > &assigned_dev_msix_mmio_ops,
> >
> > This memory_region_init_io also requires a size parameter that isn't
> > updated for the new size.
> >
> Nice catch.
> > As noted by other comments, the munmap() size isn't addressed.
> >
> Get it.
> > IMHO, the correct way to fix this would be to update pci-assign to use
> > the msix infrastructure, but legacy KVM assignment is being phased out
> > and replaced by VFIO. Is there something that ties you to pci-assign
> > instead of using vfio-pci? Thanks,
> >
> I will have a try. Alex, What's your opinion about my former letter about the
> MSI-X maximum entry.
>From other thread:
> BTW, do you think the KVM should upsize the max support MSI-X entry to 2048.
> Because the MSI-X supports a maximum table size of 2048 entries, which is
> descript in
> PCI specification 3.0 version 6.8.3.2: "MSI-X Configuration".
>
> The history patch about downsize the MSI-X entry size to 256:
> http://thread.gmane.org/gmane.comp.emulators.kvm.devel/38852/focus=38849
No, I think we should deprecate KVM device assignment and use VFIO.
Thanks,
Alex
