Re: [PATCH v3 4/5] pci: designware: remove pci_common_init_dev()

From: Fabrice Gasnier
Date: Mon May 25 2015 - 10:31:00 EST


Hi Bjorn,

On 05/06/2015 09:22 PM, Bjorn Helgaas wrote:
It's not completely obvious that replacing pci_common_init_dev() with
dw_pcie_setup() is equivalent. Here are my notes from trying to convince
myself that this is correct. I think your changelog could stand some
enhancement. It would be ideal if you could break this into a few smaller
patches that were more obviously correct, but I suspect it's too
intertwined to do that.

Thanks you for your review !

Sorry for the late reply, from your detailed analysis bellow, yes, it looks like pci_common_init_dev isn't
completely equivalent.
I'm wondering about PCI_PROBE_ONLY flag...

The idea in the first place, to move to generic probing directly, instead of
pci_common_init_dev() was to avoid being assigned default I/O (e.g. in bios32.c).
Please refer to discussion with Arnd :
http://lists.infradead.org/pipermail/linux-arm-kernel/2015-January/317726.html
But, I don't see how to be fully compatible (e.g. ARM specific option like PCI_PROBE_ONLY)
without calling pci_common_init_dev() or duplicating code from bios32.c.

Maybe should I fall back to the first idea of using specifc handling of an "empty" IO space, and keep pci_common_init_dev() ?

Please advise,

BR,
Fabrice

Here's an outline of pci_common_init_dev():

pci_common_init_dev(parent, hw)
pci_add_flags(PCI_REASSIGN_ALL_RSRC) # [1]
if (hw->preinit)
hw->preinit # [2]
pcibios_init_hw
for (nr = 0; nr < hw->nr_controllers; # [3]
sys = kzalloc # [4]
sys->msi_ctrl = hw->msi_ctrl # [5]
sys->busnr = busnr # [6]
sys->swizzle = hw->swizzle # [7]
sys->map_irq = hw->map_irq # [8]
sys->align_resource = hw->align_resource # [9]
INIT_LIST_HEAD(&sys->resources) # [10]
sys->private_data = hw->private_data # [11]
hw->setup # [12]
pcibios_init_resources # [13]
hw->scan # [14]
if (hw->postinit)
hw->postinit # [15]
pci_fixup_irqs # [16]
list_for_each_entry(sys, &head, # [17]
if (!pci_has_flag(PCI_PROBE_ONLY)) # [18]
pci_bus_size_bridges # [19]
pci_bus_assign_resources
pci_bus_add_devices
list_for_each_entry(sys, &head,
...
pcie_bus_configure_settings # [20]

[1] You don't set PCI_REASSIGN_ALL_RSRC; have you verified that nobody
cares about that?

[2] dw_pci.preinit was NULL, so skipping this doesn't matter; looks OK.

[3] dw_pci.nr_controllers was always 1, so skipping the loop doesn't
matter; looks OK.

[4] You added struct pci_sys_data sysdata as a member of struct
pcie_port, so it is now allocated by dw_pcie_host_init() callers, e.g.,
st_pcie_probe(); looks OK.

[5] You set pp->sysdata.msi_ctrl in dw_pcie_host_init() instead of
pcibios_init_hw(); looks OK.

[6] Since dw_pci.nr_controllers was always 1, sys->busnr was always 0.
You don't set sys->busnr, so it will retain its initial zero value. OK, I
guess. It's still a little broken that we have both pp->busn and
pp->sysdata.busnr, but you didn't break it and it's OK that you didn't
change anything in this regard.

[7] dw_pci.swizzle was NULL, so pcibios_swizzle() would default to
pci_common_swizzle(), which is what you use when you call pci_fixup_irqs()
in dw_pcie_scan_bus(); looks OK.

[8] dw_pci.map_irq was dw_pcie_map_irq() (which used
of_irq_parse_and_map_pci()) and sys->map_irq was used by pcibios_map_irq().
You call pci_fixup_irqs() and pass a pointer to of_irq_parse_and_map_pci();
looks OK.

[9] dw_pci.align_resource was NULL, and I assume
pcie_port.sysdata.align_resource was initialized to zeroes; looks OK.

[10] You call INIT_LIST_HEAD() in dw_pcie_host_init() instead of
pcibios_init_hw(); looks OK.

[11] You set sys->private_data in dw_pcie_host_init() instead of
pcibios_init_hw(); looks OK.

[12] dw_pci.setup was dw_pcie_setup(), and you call that from
dw_pcie_host_init(); looks OK.

[13] You no longer call pcibios_init_resources(). You don't want the
default I/O space, which makes sense; looks OK (but you need to audit other
users and make sure they don't need it either).

[14] dw_pci.scan was dw_pcie_scan_bus(), and you call that from
dw_pcie_host_init(); looks OK.

[15] dw_pci.postinit was NULL, so skipping this doesn't matter; looks OK.

[16] You call pci_fixup_irqs() from dw_pcie_scan_bus() instead of
pci_common_init_dev(); looks OK.

[17] "head" is a local list in pci_common_init_dev(), and you don't need
anything similar because dw_pcie_host_init() is called once per host
bridge; looks OK.

[18] You don't test PCI_PROBE_ONLY; it looks like it can still be set by
booting with "pci=firmware". So previously, "pci=firmware" prevented
resource assignment, but it no longer does. Is that right? Is that what
you intend?

[19] Instead of pci_bus_size_bridges() and pci_bus_assign_resources(), you
call pci_assign_unassigned_bus_resources() from dw_pcie_scan_bus(). That
seems like an improvement because it holds pci_bus_sem and supplies
add_list; looks OK.

[20] I think you no longer call pcie_bus_configure_settings(). That
configured MPS settings, and I think you still want to do that, don't you?

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/