Re: [PATCH v3 08/10] vfio/pci: protect cap/ecap_perm bits alloc/free

From: Alex Williamson
Date: Mon Dec 16 2019 - 12:42:54 EST


On Mon, 16 Dec 2019 11:57:54 +0000
"Liu, Yi L" <yi.l.liu@xxxxxxxxx> wrote:

> > From: Alex Williamson [mailto:alex.williamson@xxxxxxxxxx]
> > Sent: Monday, December 16, 2019 6:47 AM
> > To: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > Subject: Re: [PATCH v3 08/10] vfio/pci: protect cap/ecap_perm bits alloc/free
> >
> > On Thu, 21 Nov 2019 19:23:45 +0800
> > Liu Yi L <yi.l.liu@xxxxxxxxx> wrote:
> >
> > > This patch add a user numer track for the shared cap/ecap_perms bits,
> > > and the alloc/free will hold a semaphore to protect the operations.
> > > With the changes, first caller of vfio_pci_init_perm_bits() will
> > > initialize the bits. While the last caller of vfio_pci_uninit_perm_bits()
> > > will free the bits. This is a preparation to have multiple cap/ecap_perms
> > > bits users.
> > >
> > > Cc: Kevin Tian <kevin.tian@xxxxxxxxx>
> > > Cc: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
> > > Suggested-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
> > > Signed-off-by: Liu Yi L <yi.l.liu@xxxxxxxxx>
> > > ---
> > > drivers/vfio/pci/vfio_pci_config.c | 33 +++++++++++++++++++++++++++++++--
> > > 1 file changed, 31 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/vfio/pci/vfio_pci_config.c b/drivers/vfio/pci/vfio_pci_config.c
> > > index f0891bd..274c993 100644
> > > --- a/drivers/vfio/pci/vfio_pci_config.c
> > > +++ b/drivers/vfio/pci/vfio_pci_config.c
> > > @@ -36,6 +36,13 @@
> > > (offset >= PCI_ROM_ADDRESS && offset < PCI_ROM_ADDRESS + 4))
> > >
> > > /*
> > > + * vfio_perm_bits_sem: prorects the shared perm_bits alloc/free
> > > + * vfio_pci_perm_bits_users: tracks the user of the shared perm_bits
> > > + */
> > > +static DEFINE_SEMAPHORE(vfio_perm_bits_sem);
> > > +static int vfio_pci_perm_bits_users;
> > > +
> > > +/*
> > > * Lengths of PCI Config Capabilities
> > > * 0: Removed from the user visible capability list
> > > * FF: Variable length
> > > @@ -995,7 +1002,7 @@ static int __init init_pci_ext_cap_pwr_perm(struct
> > perm_bits *perm)
> > > /*
> > > * Initialize the shared permission tables
> > > */
> > > -void vfio_pci_uninit_perm_bits(void)
> > > +static void vfio_pci_uninit_perm_bits_internal(void)
> > > {
> > > free_perm_bits(&cap_perms[PCI_CAP_ID_BASIC]);
> > >
> > > @@ -1009,10 +1016,30 @@ void vfio_pci_uninit_perm_bits(void)
> > > free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
> > > }
> > >
> > > +void vfio_pci_uninit_perm_bits(void)
> > > +{
> > > + down(&vfio_perm_bits_sem);
> > > +
> > > + if (--vfio_pci_perm_bits_users > 0)
> > > + goto out;
> > > +
> > > + vfio_pci_uninit_perm_bits_internal();
> > > +
> > > +out:
> > > + up(&vfio_perm_bits_sem);
> > > +}
> > > +
> > > int __init vfio_pci_init_perm_bits(void)
> > > {
> > > int ret;
> > >
> > > + down(&vfio_perm_bits_sem);
> > > +
> > > + if (++vfio_pci_perm_bits_users > 1) {
> > > + ret = 0;
> > > + goto out;
> > > + }
> > > +
> > > /* Basic config space */
> > > ret = init_pci_cap_basic_perm(&cap_perms[PCI_CAP_ID_BASIC]);
> > >
> > > @@ -1030,8 +1057,10 @@ int __init vfio_pci_init_perm_bits(void)
> > > ecap_perms[PCI_EXT_CAP_ID_VNDR].writefn = vfio_raw_config_write;
> > >
> > > if (ret)
> > > - vfio_pci_uninit_perm_bits();
> > > + vfio_pci_uninit_perm_bits_internal();
> > >
> > > +out:
> > > + up(&vfio_perm_bits_sem);
> > > return ret;
> > > }
> > >
> >
> > Hi Yi,
> >
> > Sorry for slowness in providing feedback on this series. If we
> > provided a vfio-pci-common module that vfio-pci and vfio-mdev-pci
> > depend on, doesn't this entire problem go away?
>
> I checked previous email, export the common functions out of
> vfio-pci module was proposed in RFC v2. But at that time, I didn't
> propose to have a separate module. So I guess it may be just correct
> way. Below is the reply at that time.
>
> https://lkml.org/lkml/2019/3/19/756
>
> > I played a little bit
> > with this in the crude patch below, it seems to work. To finish this,
> > I think we'd move the function declarations out of the "private" header
> > file and into one under include/linux, then we could also move
> > vfio_mdev_pci.c to the samples directory like we intended originally.
> > I know you had tried to link things from samples and it didn't work,
> > but is the below a better attempt at resolving this? It commits us to
> > exporting a bunch of functions, we'll need to decide whether that's a
> > good idea. Thanks,
>
> I played with your patch and added a crude patch to move
> vfio-mdev-pci to samples directory. I can see below errors with
> either CONFIG_VFIO_PCI_COMMON=y and
> CONFIG_SAMPLE_VFIO_MDEV_PCI=y, or
> CONFIG_VFIO_PCI_COMMON=m and
> CONFIG_SAMPLE_VFIO_MDEV_PCI=m.
>
> But CONFIG_VFIO_PCI_COMMON works well with CONFIG_VFIO_PCI...
>
> Kernel: arch/x86/boot/bzImage is ready (#88)
> ERROR: "vfio_pci_fill_ids" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_err_handlers" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_reflck_attach" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_write" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_disable" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_reflck_put" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_refresh_config" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_set_power_state" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_enable" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_set_vga_decode" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_probe_power_state" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_mmap" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_ioctl" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> ERROR: "vfio_pci_read" [samples/vfio-mdev-pci/vfio-mdev-pci.ko] undefined!
> scripts/Makefile.modpost:93: recipe for target '__modpost' failed
>
> So I'm afraid that it still cannot resolve the problem which we encountered
> when trying to place vfio-mdev-pci in samples/.

I'm not sure what we're doing differently (with your patch applied):

$ grep VFIO ~/build/.config | grep -v ^#
CONFIG_KVM_VFIO=y
CONFIG_VFIO_IOMMU_TYPE1=m
CONFIG_VFIO_VIRQFD=m
CONFIG_VFIO=m
CONFIG_VFIO_PCI_COMMON=m
CONFIG_VFIO_PCI=m
CONFIG_VFIO_PCI_VGA=y
CONFIG_VFIO_PCI_MMAP=y
CONFIG_VFIO_PCI_INTX=y
CONFIG_VFIO_PCI_IGD=y
CONFIG_VFIO_MDEV=m
CONFIG_VFIO_MDEV_DEVICE=m
CONFIG_SAMPLE_VFIO_MDEV_PCI=m

$ make O=~/build -j36 > /dev/null && echo $?
0

$ grep VFIO ~/build/.config | grep -v ^#
CONFIG_KVM_VFIO=y
CONFIG_VFIO_IOMMU_TYPE1=y
CONFIG_VFIO_VIRQFD=y
CONFIG_VFIO=y
CONFIG_VFIO_PCI_COMMON=y
CONFIG_VFIO_PCI=y
CONFIG_VFIO_PCI_VGA=y
CONFIG_VFIO_PCI_MMAP=y
CONFIG_VFIO_PCI_INTX=y
CONFIG_VFIO_PCI_IGD=y
CONFIG_VFIO_MDEV=m
CONFIG_VFIO_MDEV_DEVICE=m
$ make O=~/build -j36 > /dev/null && echo $?
0

$ grep VFIO ~/build/.config | grep -v ^#
CONFIG_KVM_VFIO=y
CONFIG_VFIO_IOMMU_TYPE1=y
CONFIG_VFIO_VIRQFD=m
CONFIG_VFIO=y
CONFIG_VFIO_PCI_COMMON=m
CONFIG_VFIO_PCI=m
CONFIG_VFIO_PCI_VGA=y
CONFIG_VFIO_PCI_MMAP=y
CONFIG_VFIO_PCI_INTX=y
CONFIG_VFIO_PCI_IGD=y
CONFIG_VFIO_MDEV=m
CONFIG_VFIO_MDEV_DEVICE=m
$ make O=~/build -j36 > /dev/null && echo $?
0

> ================= [the crude patch] =================
> From fa860ff15ab188481141f7bd2b9cb3a1d500f24d Mon Sep 17 00:00:00 2001
> From: Liu Yi L <yi.l.liu@xxxxxxxxx>
> Date: Sun, 15 Dec 2019 19:16:54 +0800
> Subject: [PATCH] vfio/pci/sample: move vfio-pci-mdev to samples
>
> Signed-off-by: Liu Yi L <yi.l.liu@xxxxxxxxx>
> ---
> drivers/vfio/pci/Makefile | 2 -
> drivers/vfio/pci/vfio_mdev_pci.c | 421 ----------------------------------
> drivers/vfio/pci/vfio_pci.c | 2 +-
> drivers/vfio/pci/vfio_pci_common.c | 2 +-
> drivers/vfio/pci/vfio_pci_config.c | 2 +-
> drivers/vfio/pci/vfio_pci_igd.c | 2 +-
> drivers/vfio/pci/vfio_pci_intrs.c | 2 +-
> drivers/vfio/pci/vfio_pci_nvlink2.c | 2 +-
> drivers/vfio/pci/vfio_pci_private.h | 228 ------------------
> drivers/vfio/pci/vfio_pci_rdwr.c | 2 +-
> include/linux/vfio_pci_private.h | 228 ++++++++++++++++++

Of course this would not be "private" if this is the direction we
decide to go. Potentially there are still things private to
vfio-pci-common that we'd leave in the drivers directory though. I'm
not entirely thrilled to expose the objects outside of vfio-pci, but
potentially if vfio-pci had a mechanism to choose an alternate
vfio_device_ops that provided vendor mediation extension when calling
vfio_add_group_dev(), and common code was sharable to vendor drivers,
it might clean up the interface Yan is proposing for adding migration
to non-mdev devices. Thanks,

Alex