Any device which belongs to an "IOMMU group" (used by vfio) will
have links to all devices of its group listed in
/sys/bus/pci/$device/iommu_group/devices;
/sys/bus/pci/$device/iommu_group is actually a link to
/sys/kernel/iommu_groups/$n, where $n is the group number (there
will be a corresponding device node at /dev/vfio/$n once the
devices are bound to the vfio-pci driver)

The following functions are added:

virPCIDeviceGetIOMMUGroupList

  Gets a virPCIDeviceList with one virPCIDeviceList for each device
  in the same IOMMU group as the provided virPCIDevice (a copy of the
  original device object is included in the list.

virPCIDeviceAddressIOMMUGroupIterate

  Calls the function @actor once for each device in the group that
  contains the given virPCIDeviceAddress.

virPCIDeviceAddressGetIOMMUGroupAddresses

  Fills in a virPCIDeviceAddressPtr * with an array of
  virPCIDeviceAddress, one for each device in the iommu group of the
  provided virPCIDeviceAddress (including a copy of the original).

virPCIDeviceAddressGetIOMMUGroupNum

  Returns the group number as an int (a valid group number will always
  be 0 or greater).  If there is no iommu_group link in the device's
  directory (usually indicating that vfio isn't loaded), -2 will be
  returned. On any real error, -1 will be returned.

The "fix" I pushed a few commits ago would still leak a virPCIDevice
in case of an OOM error. Although it's inconsequential in practice,
this patch satisfies my OCD.

Make a copy of the device and add the copy to the
list. (virPCIDeviceListAdd() adds the original object to the list
instead).

If the device is bound to a stub driver different from what is saved
in the virPCIDevice's stubDriver attribute, update it.

The same strings were being re-created multiple times just to save
declaring a new variable. In the meantime, the use of the generic
variable names led to confusion when trying to follow the code. This
patch creates strings for:

 stubDriverName  (was called "driver" in original args)
 stubDriverPath  ("/sys/bus/pci/drivers/${stubDriverName}")
 driverLink      ("${device}/driver")
 oldDriverName   (the final component of path linked to by
                  "${device}/driver")
 oldDriverPath   ("/sys/bus/pci/drivers/${oldDriverName}")

then re-uses them as necessary.

This function has utility outside of virpci.c, so make it public.

Also the name didn't fit convention, so change it to
virPCIDeviceAddressParse.

I realized after the fact that it's probably better in the long run to
give this function a name that matches the name of the link used in
sysfs to hold the group (iommu_group).

I'm changing it now because I'm about to add several more functions
that deal with iommu groups.

The driver arg to virPCIDeviceDetach is no longer used (the name of the stub driver is now set in the virPCIDevice object, and virPCIDeviceDetach retrieves it from there). Remove it.

virPCIDeviceDetach would previously sometimes consume the input device
object (to put it on the inactive list) and sometimes not. Avoiding
memory leaks required checking beforehand to see if the device was
already on the list, and freeing the device object in the caller only
if there wasn't already an identical object on the inactive list.

This patch makes it consistent - virPCIDeviceDetach will *never*
consume the input virPCIDevice object; if it needs to put one on the
inactive list, it will create a copy and put *that* on the list. This
way the caller knows that it is always their responsibility to free
the device object they created.

virPCIDeviceReattach was making the assumption that the dev object
given to it was one and the same with the dev object on the
inactiveDevs list. If that had been the case, it would not need to
free the dev object it removed from the inactive list, because the
caller of virPCIDeviceReattach always frees the dev object that it
passes in. Since the dev object passed in is *never* the same object
that's on the list (it is a different object with the same name and
attributes, created just for the purpose of searching for the actual
object), simply doing a "ListSteal" to remove the object from the list
results in one leaked object; we need to actually free the object
after removing it from the list.

* virPCIDeviceFindByIDs - find a device on a list w/o creating an object
    This makes searching for an existing device on a list lighter weight.

* virPCIDeviceCopy - make a copy of an existing virPCIDevice object.

* virPCIDeviceGetDriverPathAndName - construct new strings containing
    1) the name of the driver bound to this device.
    2) the full path to the sysfs config for that driver.
    (This code was lifted from virPCIDeviceUnbindFromStub, and replaced
    there with a call to this new function).

Previously stubDriver was always set from a string literal, so it was
okay to use a const char * that wasn't freed when the virPCIDevice was
freed. This will not be the case in the near future, so it is now a
char* that is allocated in virPCIDeviceSetStubDriver() and freed
during virPCIDeviceFree().

add it to the syntax-check list and fix the one offending caller.

Found with 'git grep "= 1"'.

The source code base needs to be adapted as well. Some files
include virutil.h just for the string related functions (here,
the include is substituted to match the new file), some include
virutil.h without any need (here, the include is removed), and
some require both.

virPCIDeviceReattach and virPCIDeviceUnbindFromStub (called by
virPCIDeviceReattach) had previously required the name of the stub
driver as input. This is unnecessary, because the name of the driver
the device is currently bound to can be found by looking at the link:

  /sys/bus/pci/dddd:bb:ss.ff/driver

Instead of requiring that the name of the expected stub driver name
and only unbinding if that one name is matched, we no longer take a
driver name in the arglist for either of these
functions. virPCIDeviceUnbindFromStub just compares the name of the
currently bound driver to a list of "well known" stubs (right now
contains "pci-stub" and "vfio-pci" for qemu, and "pciback" for xen),
and only performs the unbind if it's one of those devices.

This allows virsh nodedevice-reattach to work properly across a
libvirtd restart, and fixes a couple of cases where we were
erroneously still hard-coding "pci-stub" as the drive name.

For some unknown reason, virPCIDeviceReattach had been calling
modprobe on the stub driver prior to unbinding the device. This was
problematic because we no longer know the name of the stub driver in
that function. However, it is pointless to probe for the stub driver
at that time anyway - because the device is bound to the stub driver,
we are guaranteed that it is already loaded, and so that call to
modprobe has been removed.

Given a virPCIDevice, this function returns the path for the device
that controls the vfio group the device belongs to,
e.g. "/dev/vfio/15".

This can be set when the virPCIDevice is created and placed on a list,
then used later when traversing the list to determine which stub
driver to bind/unbind for managed devices.

The existing Detach and Attach functions' signatures haven't been
changed (they still accept a stub driver name in the arg list), but if
the arg list has NULL for stub driver and one is available in the
device's object, that will be used. (we may later deprecate and remove
the arg from those functions).

POSIX says that both basename() and dirname() may return static
storage (aka they need not be thread-safe); and that they may but
not must modify their input argument.  Furthermore, <libgen.h>
is not available on all platforms.  For these reasons, you should
never use these functions in a multi-threaded library.

Gnulib instead recommends a way to avoid the portability nightmare:
gnulib's "dirname.h" provides useful thread-safe counterparts.  The
obvious dir_name() and base_name() are GPL (because they malloc(),
but call exit() on failure) so we can't use them; but the LGPL
variants mdir_name() (malloc's or returns NULL) and last_component
(always points into the incoming string without modifying it,
differing from basename semantics only on corner cases like the
empty string that we shouldn't be hitting in the first place) are
already in use in libvirt.  This finishes the swap over to the safe
functions.

* cfg.mk (sc_prohibit_libgen): New rule.
* src/util/vircgroup.c: Fix offenders.
* src/parallels/parallels_storage.c (parallelsPoolAddByDomain):
Likewise.
* src/parallels/parallels_network.c (parallelsGetBridgedNetInfo):
Likewise.
* src/node_device/node_device_udev.c (udevProcessSCSIHost)
(udevProcessSCSIDevice): Likewise.
* src/storage/storage_backend_disk.c
(virStorageBackendDiskDeleteVol): Likewise.
* src/util/virpci.c (virPCIGetDeviceAddressFromSysfsLink):
Likewise.
* src/util/virstoragefile.h (_virStorageFileMetadata): Avoid false
positive.
Signed-off-by: NEric Blake <eblake@redhat.com>

http://www.uhv.edu/ac/newsletters/writing/grammartip2009.07.01.htm
(and several other sites) give hints that 'onto' is best used if
you can also add 'up' just before it and still make sense. In many
cases in the code base, we really want the two-word form, or even
a simplification to just 'on' or 'to'.

* docs/hacking.html.in: Use correct 'on to'.
* python/libvirt-override.c: Likewise.
* src/lxc/lxc_controller.c: Likewise.
* src/util/virpci.c: Likewise.
* daemon/THREADS.txt: Use simpler 'on'.
* docs/formatdomain.html.in: Better usage.
* docs/internals/rpc.html.in: Likewise.
* src/conf/domain_event.c: Likewise.
* src/rpc/virnetclient.c: Likewise.
* tests/qemumonitortestutils.c: Likewise.
* HACKING: Regenerate.
Signed-off-by: NEric Blake <eblake@redhat.com>

Though they are the same thing, mixed use of them is uncomfortable.
"unsigned" is used a lot in old codes, this just tries to change the
ones in utils.

Members of struct virPCIDevice are changed together.

Commit 9a3ff01d (which was ACKed at
the end of January, but for some reason didn't get pushed until during
the 1.0.4 freeze) fixed the logic in virPCIGetVirtualFunctions().
Unfortunately, a typo in the fix (replacing VIR_REALLOC_N with
VIR_ALLOC_N during code movement) caused not only a memory leak, but
also resulted in most of the elements of the result array being
replaced with NULL. virNetDevGetVirtualFunctions() assumed (and I think
rightly so) that virPCIGetVirtualFunctions() wouldn't return any NULL
elements in the array, so it ended up segfaulting.

This was found when attempting to use a virtual network with an
auto-created pool of SRIOV VFs, e.g.:

    <forward mode='hostdev' managed='yes'>
      <pf dev='eth4'/>
    </forward>

(the pool of PCI addresses is discovered by calling
virNetDevGetVirtualFunctions() on the PF dev).

virPCIGetVirtualFunctions returns 0 even if there is no "virtfn"
entry under the device sysfs path.

And virPCIGetVirtualFunctions returns -1 when it fails to get
the PCI config space of one VF, however, with keeping the
the VFs already detected.

That's why udevProcessPCI and gather_pci_cap use logic like:

if (!virPCIGetVirtualFunctions(syspath,
                               &data->pci_dev.virtual_functions,
                               &data->pci_dev.num_virtual_functions) ||
    data->pci_dev.num_virtual_functions > 0)
    data->pci_dev.flags |= VIR_NODE_DEV_CAP_FLAG_PCI_VIRTUAL_FUNCTION;

to tag the PCI device with "virtual_function" cap.

However, this results in a VF will aslo get "virtual_function" cap.

This patch fixes it by:
  * Ignoring the VF which has failure of getting PCI config space
    (given that the successfully detected VFs are kept , it makes
    sense to not give up on the failure of one VF too) with a warning,
    so virPCIGetVirtualFunctions will not return -1 except out of memory.

  * Free the allocated *virtual_functions when out of memory

And thus the logic can be changed to:

    /* Out of memory */
    int ret = virPCIGetVirtualFunctions(syspath,
                                        &data->pci_dev.virtual_functions,
                                        &data->pci_dev.num_virtual_functions);

    if (ret < 0 )
        goto out;
    if (data->pci_dev.num_virtual_functions > 0)
        data->pci_dev.flags |= VIR_NODE_DEV_CAP_FLAG_PCI_VIRTUAL_FUNCTION;

Commits 20253560 and ba72cb12 introduced typos.

* src/util/virpci.c (virPCIIsVirtualFunction) [!__linux__]: Fix
function name.
* src/util/virutil.c (virGetDeviceID): Fix attribute spelling.

To allow modifications to the lists to be synchronized, convert
virPCIDeviceList and virUSBDeviceList into virObjectLockable
classes. The locking, however, will not be self-contained. The
users of these classes will have to call virObjectLock/Unlock
in the critical regions.
Signed-off-by: NDaniel P. Berrange <berrange@redhat.com>

Rename all the pciDeviceXXX and pciXXXDevice APIs to have a
fixed virPCIDevice name prefix

Pass stub driver name directly to pciDettachDevice and pciReAttachDevice to fit
for different libvirt drivers. For example, qemu driver prefers pci-stub, but
Xen prefers pciback.
Signed-off-by: NChunyan Liu <cyliu@suse.com>

The pciWrite32 function assembled the array of data to be written to the
fd with a bad offset on the last byte. This issue was probably caused by
a typo (14, 24).

Directly open and close PCI config file in the APIs that need it rather
than keeping the file open for the whole life of PCI device structure.

In order to be able to steal PCI device by its index in the list.

The libvirt coding standard is to use 'function(...args...)'
instead of 'function (...args...)'. A non-trivial number of
places did not follow this rule and are fixed in this patch.
Signed-off-by: NDaniel P. Berrange <berrange@redhat.com>