[Qemu-devel] Re: [PATCH] PCI: Bus number from the bridge, not the device

[Gleb Natapov <gleb@redhat.com>]

On Mon, Nov 08, 2010 at 07:08:37PM +0200, Michael S. Tsirkin wrote:
> On Mon, Nov 08, 2010 at 07:00:15PM +0200, Gleb Natapov wrote:
> > On Mon, Nov 08, 2010 at 06:26:33PM +0200, Michael S. Tsirkin wrote:
> > > On Mon, Nov 08, 2010 at 07:52:12AM -0700, Alex Williamson wrote:
> > > > On Mon, 2010-11-08 at 13:22 +0200, Michael S. Tsirkin wrote:
> > > > > On Mon, Oct 04, 2010 at 03:53:11PM -0600, Alex Williamson wrote:
> > > > > > pcibus_dev_print() was erroneously retrieving the device bus
> > > > > > number from the secondary bus number offset of the device
> > > > > > instead of the bridge above the device.  This ends of landing
> > > > > > in the 2nd byte of the 3rd BAR for devices, which thankfully
> > > > > > is usually zero.  pcibus_get_dev_path() copied this code,
> > > > > > inheriting the same bug.  pcibus_get_dev_path() is used for
> > > > > > ramblock naming, so changing it can effect migration.  However,
> > > > > > I've only seen this byte be non-zero for an assigned device,
> > > > > > which can't migrate anyway, so hopefully we won't run into
> > > > > > any issues.
> > > > > > 
> > > > > > Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> > > > > 
> > > > > Good catch. Applied.
> > > > > I don't really see why do we put the dev path
> > > > > in the bus object: why not let device supply its name?
> > > > 
> > > > Because the device name is not unique.  This came about from the
> > > > discussion about how to create a canonical device path that Gleb and
> > > > Markus are again trying to hash out.  If we go up to the bus and get the
> > > > bus address, we have a VM unique name.  Unfortunately, it's difficult to
> > > > define what the bus should print in all cases (ISA), but since they
> > > > don't do hotplug and typically don't allocate ramblocks, we can mostly
> > > > ignore it for this use case.
> > > > 
> > > > > And I think this will affect nested bridges. However they are currently
> > > > > broken anyway: we really must convert to topological names as bus number
> > > > > is guest-assigned - they don't have to be unique, even.
> > > > 
> > > > Yes, nested bridges are a problem.  How can the seg/bus/devfn not be
> > > > unique?
> > > 
> > > Bus numbers for nested bridges are guest assigned. We start with 0 after reset.
> > > 
> > > > > What does fixing this involve? Just changing pcibus_get_dev_path?
> > > > 
> > > > How do you plan to fix it?  Don't forget that migration depends on these
> > > > names, so some kind of compatibility layer would be required.  Thanks,
> > > > 
> > > > Alex
> > > 
> > > Replace bus number with slot numbers of parent bridges up to the root.
> > > This works for root bridge in a compatible way because bus number there
> > > is hard-coded to 0.
> > > IMO nested bridges are broken anyway, no way to be compatible there.
> > > 
> > > 
> > > Gleb, Markus, I think the following should be sufficient for PCI.  What
> > > do you think?  Also - do we need to update QMP/monitor to teach them to
> > > work with these paths?
> > > 
> > 
> > I am no longer use bus's get_dev_path callback for my purpose (I added
> > another one get_fw_dev_path)
> 
> Why? Because get_dev_path is unstable? to avoid changing for migration?
Because to get functional get_dev_path() we need to implemented it
recursively (hint domain should not be part of PCI bus get_dev_path but
its parent bus instead). And since  get_dev_path() is used in migration
code you have all or nothing situation. Either you implement it for all
devices properly or for none.

> So with this patch, you get to use it again.
Unfortunately not.

> IMO two ways to address devices is a bad idea.
> 
> > since it is used for migration
> 
> what is used for migration?
> 
get_dev_path() callback is used to create unique instance id. Actually
it is the only use of it so simple grep will show you this :)

> > it should
> > be done for all buses to work properly. And by properly I mean produce
> > full path from system root to the device itself recursively.
> 
> This is what the code does. Recursion is not needed here though.
> 
Recursion is needed to create full device path. I am not interested in
only PCI here. In fact PCI one is the easiest.

> > But what I
> > learned is that by changing get_dev_path output you will break migration
> > from older guest to newer once (something we have to support).
> 
> Well I think migration for sytems with nested buses are broken anyway:
> it's just a new feature where we simply did not figure out the migration
> andgle before the last release.
> Without nesting my code returns exactly the existing output
> so no, it's not broken.
> 
It does not drops bus id part? If without bridges the output is exactly
same then migration is OK with the patch.


> > And of
> > course, as Alex said already, you need to traverse bridges recursively
> 
> Well this is an implementation detail.  loop is functionally equivalent,
> only cleaner and easier to understand.
> 
With that I can agree. I am referring more to recursion down device
tree.

> > and
> > domain does not provide any meaningful information.
> 
> I dont understand yet, sorry. Code I posted returns exactly
> what's there today. If domain does not provide any meaningful
> information, How do things work then? And how do you address
> roots on a system with multiple roots?
> 
The current code is broken. Look at my patch series.

> > > This is on top of Alex's patch, completely untested.
> > > 
> > > 
> > > pci: fix device path for devices behind nested bridges
> > > 
> > > We were using bus number in the device path, which is clearly
> > > broken as this number is guest-assigned for all devices
> > > except the root.
> > > 
> > > Fix by using hierarchical list of slots, walking the path
> > > from root down to device, instead. Add :00 as bus number
> > > so that if there are no nested bridges, this is compatible
> > > with what we have now.
> > > 
> > > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > 
> > > diff --git a/hw/pci.c b/hw/pci.c
> > > index 7d12473..fa98d94 100644
> > > --- a/hw/pci.c
> > > +++ b/hw/pci.c
> > > @@ -1826,13 +1826,45 @@ static void pcibus_dev_print(Monitor *mon, DeviceState *dev, int indent)
> > >  
> > >  static char *pcibus_get_dev_path(DeviceState *dev)
> > >  {
> > > -    PCIDevice *d = (PCIDevice *)dev;
> > > -    char path[16];
> > > -
> > > -    snprintf(path, sizeof(path), "%04x:%02x:%02x.%x",
> > > -             pci_find_domain(d->bus), pci_bus_num(d->bus),
> > > -             PCI_SLOT(d->devfn), PCI_FUNC(d->devfn));
> > > -
> > > -    return strdup(path);
> > > +    PCIDevice *d = container_of(dev, PCIDevice, qdev);
> > > +    PCIDevice *t;
> > > +    int slot_depth;
> > > +    /* Path format: Domain:00:Slot:Slot....:Slot.Function.
> > > +     * 00 is added here to make this format compatible with
> > > +     * domain:Bus:Slot.Func for systems without nested PCI bridges.
> > > +     * Slot list specifies the slot numbers for all devices on the
> > > +     * path from root to the specific device. */
> > > +    int domain_len = strlen("DDDD:00");
> > > +    int func_len = strlen(".F");
> > > +    int slot_len = strlen(":SS");
> > > +    int path_len;
> > > +    char *path, *p;
> > > +
> > > +    /* Calculate # of slots on path between device and root. */;
> > > +    slot_depth = 0;
> > > +    for (t = d; t; t = t->bus->parent_dev)
> > > +        ++slot_depth;
> > > +
> > > +    path_len = domain_len + bus_len + slot_len * slot_depth + func_len;
> > > +
> > > +    /* Allocate memory, fill in the terminating null byte. */
> > > +    path = malloc(path_len + 1 /* For '\0' */);
> > > +    path[path_len] = '\0';
> > > +
> > > +    /* First field is the domain. */
> > > +    snprintf(path, domain_len, "%04x", pci_find_domain(d->bus));
> > > +
> > > +    /* Leave space for slot numbers and fill in function number. */
> > > +    p = path + domain_len + slot_len * slot_depth;
> > > +    snprintf(p, func_len, ".%02x", PCI_FUNC(d->devfn));
> > > +
> > > +    /* Fill in slot numbers. We walk up from device to root, so need to print
> > > +     * them in the reverse order, last to first. */
> > > +    for (t = d; t; t = t->bus->parent_dev) {
> > > +        p -= slot_len;
> > > +        snprintf(p, slot_len, ":%x", PCI_SLOT(t->devfn));
> > > +    }
> > > +
> > > +    return path;
> > >  }
> > >  
> > 
> > --
> > 			Gleb.

--
			Gleb.