There is an issue with the hotplug operation when it's done on the
bridge/switch slot. The bridge-port and devices behind the bridge, which
become offline by hot-unplug operation, don't get hot-plugged/enabled by
doing hot-plug operation on that slot. Only the first port of the bridge
gets enabled and the remaining port/devices remain unplugged. The hot
plug/unplug operation is done by the hotplug driver
(drivers/pci/hotplug/pnv_php.c).
Root Cause Analysis: This behavior is due to missing code for the DPC
switch/bridge. The existing driver depends on pci_hp_add_devices()
function for device enablement. This function calls pci_scan_slot() on
only one device-node/port of the bridge, not on all the siblings'
device-node/port.
The missing code needs to be added which will find all the sibling
device-nodes/bridge-ports and will run explicit pci_scan_slot() on
those. A new function has been added for this purpose which gets
invoked from pci_hp_add_devices(). This new function
pci_traverse_sibling_nodes_and_scan_slot() gets all the sibling
bridge-ports by traversal and explicitly invokes pci_scan_slot on them.
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@kernel.org>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Cc: Gaurav Batra <gbatra@linux.ibm.com>
Cc: Nathan Lynch <nathanl@linux.ibm.com>
Cc: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Krishna Kumar <krishnak@linux.ibm.com>
---
Command for reproducing the issue :
For hot unplug/disable - echo 0 > /sys/bus/pci/slots/C5/power
For hot plug/enable - echo 1 > /sys/bus/pci/slots/C5/power
where C5 is slot associated with bridge.
Scenario/Tests:
Output of lspci -nn before test is given below. This snippet contains
devices used for testing on Powernv machine.
0004:02:00.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
0004:02:01.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
0004:02:02.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
0004:02:03.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
0004:08:00.0 Serial Attached SCSI controller [0107]:
Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3 [1000:00c9] (rev 01)
0004:09:00.0 Serial Attached SCSI controller [0107]:
Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3 [1000:00c9] (rev 01)
Output of lspci -tv before test is as follows:
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
| | +-01.0-[08]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | +-02.0-[09]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | \-03.0-[0a-0e]--
| \-00.1 PMC-Sierra Inc. Device 4052
C5(bridge) and C6(End Point) slot address are as below:
# cat /sys/bus/pci/slots/C5/address
0004:02:00
# cat /sys/bus/pci/slots/C6/address
0004:09:00
Hot-unplug operation on slot associated with bridge:
# echo 0 > /sys/bus/pci/slots/C5/power
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--
| \-00.1 PMC-Sierra Inc. Device 4052
From the above lspci -tv output, it can be observed that hot unplug
operation has removed all the PMC-Sierra bridge ports like:
00.0-[03-07], 01.0-[08], 02.0-[09], 03.0-[0a-0e] and the SAS devices
behind the bridge-port. Without the fix, when the hot plug operation is
done on the same slot, it adds only the first bridge port and doesn't
restore all the bridge-ports and devices that it unplugged earlier.
Below snippet shows this.
Hot-plug operation on the bridge slot without the fix:
# echo 1 > /sys/bus/pci/slots/C5/power
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
| \-00.1 PMC-Sierra Inc. Device 4052
After the fix, it restores all the devices in the same manner how it
unplugged them earlier during the hot unplug operation. The below snippet
shows the same.
Hot-plug operation on bridge slot with the fix:
# echo 1 > /sys/bus/pci/slots/C5/power
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
| | +-01.0-[08]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | +-02.0-[09]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | \-03.0-[0a-0e]--
| \-00.1 PMC-Sierra Inc. Device 4052
Removal of End point device behind bridge are also intact and behaving
correctly.
Hot-unplug operation on Endpoint device C6:
# echo 0 > /sys/bus/pci/slots/C6/power
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
| | +-01.0-[08]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | +-02.0-[09]--
| | \-03.0-[0a-0e]--
| \-00.1 PMC-Sierra Inc. Device 4052
Hot-plug operation on Endpoint device C6:
# echo 1 > /sys/bus/pci/slots/C6/power
# lspci -tv
+-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
| | +-01.0-[08]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | +-02.0-[09]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
| | \-03.0-[0a-0e]--
| \-00.1 PMC-Sierra Inc. Device 4052
arch/powerpc/include/asm/ppc-pci.h | 4 ++++
arch/powerpc/kernel/pci-hotplug.c | 5 ++---
arch/powerpc/kernel/pci_dn.c | 32 ++++++++++++++++++++++++++++++
3 files changed, 38 insertions(+), 3 deletions(-)
diff --git a/arch/powerpc/include/asm/ppc-pci.h b/arch/powerpc/include/asm/ppc-pci.h
index a8b7e8682f5b..83db8d0798ac 100644
--- a/arch/powerpc/include/asm/ppc-pci.h
+++ b/arch/powerpc/include/asm/ppc-pci.h
@@ -28,6 +28,10 @@ struct pci_dn;
void *pci_traverse_device_nodes(struct device_node *start,
void *(*fn)(struct device_node *, void *),
void *data);
+
+void pci_traverse_sibling_nodes_and_scan_slot(struct device_node *start,
+ struct pci_bus *bus);
+
extern void pci_devs_phb_init_dynamic(struct pci_controller *phb);
#if defined(CONFIG_IOMMU_API) && (defined(CONFIG_PPC_PSERIES) || \
diff --git a/arch/powerpc/kernel/pci-hotplug.c b/arch/powerpc/kernel/pci-hotplug.c
index 0fe251c6ac2c..639a3d592fe2 100644
--- a/arch/powerpc/kernel/pci-hotplug.c
+++ b/arch/powerpc/kernel/pci-hotplug.c
@@ -106,7 +106,7 @@ EXPORT_SYMBOL_GPL(pci_hp_remove_devices);
*/
void pci_hp_add_devices(struct pci_bus *bus)
{
- int slotno, mode, max;
+ int mode, max;
struct pci_dev *dev;
struct pci_controller *phb;
struct device_node *dn = pci_bus_to_OF_node(bus);
@@ -129,8 +129,7 @@ void pci_hp_add_devices(struct pci_bus *bus)
* order for fully rescan all the way down to pick them up.
* They can have been removed during partial hotplug.
*/
- slotno = PCI_SLOT(PCI_DN(dn->child)->devfn);
- pci_scan_slot(bus, PCI_DEVFN(slotno, 0));
+ pci_traverse_sibling_nodes_and_scan_slot(dn, bus);
max = bus->busn_res.start;
/*
* Scan bridges that are already configured. We don't touch
diff --git a/arch/powerpc/kernel/pci_dn.c b/arch/powerpc/kernel/pci_dn.c
index 38561d6a2079..bea612759832 100644
--- a/arch/powerpc/kernel/pci_dn.c
+++ b/arch/powerpc/kernel/pci_dn.c
@@ -493,4 +493,36 @@ static void pci_dev_pdn_setup(struct pci_dev *pdev)
pdn = pci_get_pdn(pdev);
pdev->dev.archdata.pci_data = pdn;
}
+
+void pci_traverse_sibling_nodes_and_scan_slot(struct device_node *start, struct pci_bus *bus)
+{
+ struct device_node *dn;
+ int slotno;
+
+ u32 class = 0;
+
+ if (!of_property_read_u32(start->child, "class-code", &class)) {
+ /* Call of pci_scan_slot for non-bridge/EP case */
+ if (!((class >> 8) == PCI_CLASS_BRIDGE_PCI)) {
+ slotno = PCI_SLOT(PCI_DN(start->child)->devfn);
+ pci_scan_slot(bus, PCI_DEVFN(slotno, 0));
+ return;
+ }
+ }
+
+ /* Iterate all siblings */
+ for_each_child_of_node(start, dn) {
+ class = 0;
+
+ if (!of_property_read_u32(start->child, "class-code", &class)) {
+ /* Call of pci_scan_slot on each sibling-nodes/bridge-ports */
+ if ((class >> 8) == PCI_CLASS_BRIDGE_PCI) {
+ slotno = PCI_SLOT(PCI_DN(dn)->devfn);
+ pci_scan_slot(bus, PCI_DEVFN(slotno, 0));
+ }
+ }
+ }
+
+}
+
DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pci_dev_pdn_setup);
--
2.45.0On Tue, May 14, 2024 at 11:54 PM Krishna Kumar <krishnak@linux.ibm.com> wrote:
>
> There is an issue with the hotplug operation when it's done on the
> bridge/switch slot. The bridge-port and devices behind the bridge, which
> become offline by hot-unplug operation, don't get hot-plugged/enabled by
> doing hot-plug operation on that slot. Only the first port of the bridge
> gets enabled and the remaining port/devices remain unplugged. The hot
> plug/unplug operation is done by the hotplug driver
> (drivers/pci/hotplug/pnv_php.c).
>
> Root Cause Analysis: This behavior is due to missing code for the DPC
> switch/bridge.
I don't see anything touching DPC in this series?
> *snip*
>
> Command for reproducing the issue :
>
> For hot unplug/disable - echo 0 > /sys/bus/pci/slots/C5/power
> For hot plug/enable - echo 1 > /sys/bus/pci/slots/C5/power
>
> where C5 is slot associated with bridge.
>
> Scenario/Tests:
> Output of lspci -nn before test is given below. This snippet contains
> devices used for testing on Powernv machine.
>
> 0004:02:00.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
> 0004:02:01.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
> 0004:02:02.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
> 0004:02:03.0 PCI bridge [0604]: PMC-Sierra Inc. Device [11f8:4052]
> 0004:08:00.0 Serial Attached SCSI controller [0107]:
> Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3 [1000:00c9] (rev 01)
> 0004:09:00.0 Serial Attached SCSI controller [0107]:
> Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3 [1000:00c9] (rev 01)
>
> Output of lspci -tv before test is as follows:
>
> # lspci -tv
> +-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--+-00.0-[03-07]--
> | | +-01.0-[08]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
> | | +-02.0-[09]----00.0 Broadcom / LSI SAS3216 PCI-Express Fusion-MPT SAS-3
> | | \-03.0-[0a-0e]--
> | \-00.1 PMC-Sierra Inc. Device 4052
>
> C5(bridge) and C6(End Point) slot address are as below:
> # cat /sys/bus/pci/slots/C5/address
> 0004:02:00
> # cat /sys/bus/pci/slots/C6/address
> 0004:09:00
Uh, if I'm reading this right it looks like your "slot" C5 is actually
the PCIe switch's internal bus which is definitely not hot pluggable.
I find it helps to look at the PCI topology in terms of where the
physical PCIe links are. Here we've got:
- A link between the PHB (0004:00:00.0) and the switch upstream port
(0004:01:00.0)
- A link from switch downstream port 0 (0004:02:00.0) to nothing
- A link from switch downstream port 1 (0004:02:01.0) to a SAS card
- A link from switch downstream port 2 (0004:02:02.0) to a SAS card
- A link from switch downstream port 2 (0004:02:03.0) to nothing
Note that there's no PCIe link between the switch upstream port
(0004:01:00.0) and the downstream ports on bus 0004:02. The connection
between those is invisible to us because it's custom bus logic
internal to the PCIe switch ASIC. What I think has happened here is
that system firmware has supplied bad PCIe slot information to OPAL
which has resulted in pnv_php advertising a slot in the wrong place.
Assuming this following the usual IBM convention I'd expect the bridge
device for C5 to be the PHB's root port and the bus should be 0004:01.
It might be worth adding some logic to pnv_php to verify the PCI
bridge upstream of the slot actually has the PCIe slot capability to
guard against this problem.
> Hot-unplug operation on slot associated with bridge:
> # echo 0 > /sys/bus/pci/slots/C5/power
> # lspci -tv
> +-[0004:00]---00.0-[01-0e]--+-00.0-[02-0e]--
> | \-00.1 PMC-Sierra Inc. Device 4052
Yep, "powering off" C5 doesn't remove the upstream port device. This
would create problems if you physically removed the card from C5 since
the kernel would assume the switch device is still present.
> *snip*
> diff --git a/arch/powerpc/kernel/pci_dn.c b/arch/powerpc/kernel/pci_dn.c
> index 38561d6a2079..bea612759832 100644
> --- a/arch/powerpc/kernel/pci_dn.c
> +++ b/arch/powerpc/kernel/pci_dn.c
> @@ -493,4 +493,36 @@ static void pci_dev_pdn_setup(struct pci_dev *pdev)
> pdn = pci_get_pdn(pdev);
> pdev->dev.archdata.pci_data = pdn;
> }
> +
> +void pci_traverse_sibling_nodes_and_scan_slot(struct device_node *start, struct pci_bus *bus)
> +{
> + struct device_node *dn;
> + int slotno;
> +
> + u32 class = 0;
> +
> + if (!of_property_read_u32(start->child, "class-code", &class)) {
> + /* Call of pci_scan_slot for non-bridge/EP case */
> + if (!((class >> 8) == PCI_CLASS_BRIDGE_PCI)) {
> + slotno = PCI_SLOT(PCI_DN(start->child)->devfn);
> + pci_scan_slot(bus, PCI_DEVFN(slotno, 0));
> + return;
> + }
> + }
> +
> + /* Iterate all siblings */
> + for_each_child_of_node(start, dn) {
> + class = 0;
> +
> + if (!of_property_read_u32(start->child, "class-code", &class)) {
> + /* Call of pci_scan_slot on each sibling-nodes/bridge-ports */
> + if ((class >> 8) == PCI_CLASS_BRIDGE_PCI) {
> + slotno = PCI_SLOT(PCI_DN(dn)->devfn);
> + pci_scan_slot(bus, PCI_DEVFN(slotno, 0));
> + }
> + }
> + }
If you're going to iterate over all the DT nodes why not just scan all
of them rather than special casing bridges? IIRC current logic is the
way it is because PowerVM only puts single devices under a PHB and in
the PowerNV (pnv_php) case the PCIe spec guarantees that only device 0
will be present on the end of a link. If you want to handle the more
generic case then feel free, but do it properly.
© 2016 - 2026 Red Hat, Inc.