On 8/16/2021 8:21 PM, Bjorn Helgaas wrote:
On Sun, Aug 15, 2021 at 02:27:13AM +0300, Max Gurtovoy wrote:
On 8/13/2021 8:44 PM, Bjorn Helgaas wrote:
On Fri, Aug 13, 2021 at 02:21:41AM +0300, Max Gurtovoy wrote:
On 8/12/2021 11:26 PM, Bjorn Helgaas wrote:
On Thu, Aug 12, 2021 at 04:51:26PM -0300, Jason Gunthorpe wrote:
On Thu, Aug 12, 2021 at 10:57:07AM -0500, Bjorn Helgaas wrote:
On Thu, Aug 12, 2021 at 10:27:28AM -0300, Jason Gunthorpe wrote:
On Wed, Aug 11, 2021 at 02:07:37PM -0500, Bjorn Helgaas wrote:
On Thu, Aug 05, 2021 at 09:23:57PM -0300, Jason Gunthorpe wrote:
Do the other bus types have a flag analogous to
PCI_ID_F_VFIO_DRIVER_OVERRIDE? If we're doing something similar to
other bus types, it'd be nice if the approach were similar.
They could, this series doesn't attempt it. I expect the approach to
be similar as driver_override was copied from PCI to other
busses. When this is completed I hope to take a look at it.
I think this would make more sense as two patches:
- Add a "PCI_ID_DRIVER_OVERRIDE" flag. This is not VFIO-specific,
since nothing in PCI depends on the VFIO-ness of drivers that use
the flag. The only point here is that driver id_table entries
with this flag only match when driver_override matches the driver.
This would require using two flags, one to indicate the above to the
PCI code and another to indicate the vfio_pci string to
file2alias. This doesn't seem justified at this point, IMHO.
I don't think it requires two flags. do_pci_entry() has:
if (flags & PCI_ID_F_VFIO_DRIVER_OVERRIDE)
strcpy(alias, "vfio_pci:");
I'm just proposing a rename:
s/PCI_ID_F_VFIO_DRIVER_OVERRIDE/PCI_ID_DRIVER_OVERRIDE/
- Update file2alias.c to export the flags and the "vfio_pci:" alias.
This seems to be the only place where VFIO comes into play, and
putting it in a separate patch will make it much smaller and it
will be clear how it could be extended for other buses.
Well, I don't want to see a flag called PCI_ID_DRIVER_OVERRIDE mapped
to the string "vfio_pci", that is just really confusing.
Hahaha, I see, that's fair :) It confused me for a long time why you
wanted "VFIO" in the flag name because from the kernel's point of
view, the flag is not related to any VFIO-ness. It's only related to
a special variety of driver_override, and VFIO happens to be one user
of it.
In my original patch I used
#define PCI_ID_DRIVER_OVERRIDE PCI_ID_F_VFIO_DRIVER_OVERRIDE
and in the pci core code I used PCI_ID_DRIVER_OVERRIDE in the "if" clause.
So we can maybe do that and leave the option to future update of the define
without changing the core code.
In the future we can have something like:
#define PCI_ID_DRIVER_OVERRIDE (PCI_ID_F_VFIO_DRIVER_OVERRIDE |
PCI_ID_F_MY_BUS_DRIVER_OVERRIDE)
The file2alias.c still have to use the exact PCI_ID_F_VFIO_DRIVER_OVERRIDE
flag to add "vfio_" prefix.
Is that better ?
I don't think it's worth having two separate #defines. If we need
more in the future, we can add them when we need them.
I meant 1 #define and 1 enum:
enum {
PCI_ID_F_VFIO_DRIVER_OVERRIDE = 1 << 0,
};
#define PCI_ID_DRIVER_OVERRIDE PCI_ID_F_VFIO_DRIVER_OVERRIDE
Basically the same thing. Doesn't seem worthwhile to me to have both.
When reading the code, it's not at all obvious why you would define a
new name for PCI_ID_F_VFIO_DRIVER_OVERRIDE.
because we need the "vfio_" prefix in the alias.
And the match can use PCI_ID_DRIVER_OVERRIDE that in the future cab be
(#define PCI_ID_DRIVER_OVERRIDE (PCI_ID_F_VFIO_DRIVER_OVERRIDE |
PCI_ID_F_SOME_OTHER_ALIAS_DRIVER_OVERRIDE)
What if we renamed "flags" to be specifically for this override case,
e.g., "override_only"? Then the flag could be
PCI_ID_F_VFIO_DRIVER_OVERRIDE, which would trigger a "vfio_" prefix in
file2alias.c, but pci_match_device() could just check for it being
non-zero, without caring whether the reason is VFIO or something else,
e.g.,
pci_match_device(...)
{
...
if (found_id->override_only) {
if (dev->driver_override)
return found_id;
...
Jason suggested something like this:
static const struct pci_device_id *pci_match_device(struct pci_driver *drv,
struct pci_dev *dev)
{
struct pci_dynid *dynid;
const struct pci_device_id *found_id = NULL, *ids;
/* When driver_override is set, only bind to the matching driver */
if (dev->driver_override && strcmp(dev->driver_override, drv->name))
return NULL;
/* Look at the dynamic ids first, before the static ones */
spin_lock(&drv->dynids.lock);
list_for_each_entry(dynid, &drv->dynids.list, node) {
if (pci_match_one_device(&dynid->id, dev)) {
found_id = &dynid->id;
break;
}
}
spin_unlock(&drv->dynids.lock);
if (found_id)
return found_id;
for (ids = drv->id_table; (found_id = pci_match_id(ids, dev));
ids = found_id + 1) {
/*
* The match table is split based on driver_override. Check the
* flags as well so that any matching
* PCI_ID_F_VFIO_DRIVER_OVERRIDE entry is returned.
*/
if (!(found_id->flags & PCI_ID_F_VFIO_DRIVER_OVERRIDE) ||
dev->driver_override)
return found_id;
}
/*
* if no static match, driver_override will always match, send a dummy
* id.
*/
if (dev->driver_override)
return &pci_device_id_any;
return NULL;
}
It looks good to me as well.
I missed your point. Isn't the above basically the 09/12 patch [1] we're
talking about?
Yes, I see the code structure is slightly different, but the question
we're talking about here is the name of the "flags" field and the enum
or #define for the VFIO bit.
I guess the renaming of "__u32 flags" to "__u32 driver_override" is ok
from my perspective.
The enum for vfio should stay.
The prefix we want in the alias is "vfio_" and not "driver_override_".
This will allow a clean uAPI. "driver_override_" prefix will be too
generic for userspace tools like libvirt that would like to find a
*VFIO* driver not something else.
Thus we need alias to be "vfio_".
In the future if some other driver will use this flag, it will create an
alias also. In your suggestion, the alias will be the same and the
userspace tool won't be able to distinguish between the two.
But in the original solution, for non vfio driver override drivers, one
can use new enum PCI_ID_F_SOME_OTHER_ALIAS_DRIVER_OVERRIDE and add its
own alias prefix for recognition "my_prefix_".
I prefer the "flags" naming since its more generic and easy to extend.
We don't need to worry about "flags" being generic or extensible until
we need to extend it. It's easy to fiddle with it at that point.
can we continue with the above suggestion for V2 ?
I don't see what really changed with the above suggestion.
The point I'm trying to make is that using PCI_ID_F_VFIO_DRIVER_OVERRIDE
in pci_match_device() suggests that the code there has some connection
or dependency on VFIO, but it does not.
This is why I suggested a "#define PCI_ID_DRIVER_OVERRIDE
PCI_ID_F_VFIO_DRIVER_OVERRIDE"
[1] https://lore.kernel.org/r/20210721161609.68223-10-yishaih@xxxxxxxxxx
