PCI: mvebu: Dispose INTx IRQs before to removing INTx domain
Subsystems (3)
PATCH v2 1/1 PCI: mvebu: Dispose INTx IRQs before to removing INTx domain
From: Pali Rohár <[email protected]>
The documentation for the irq_domain_remove() function says that all mappings within the IRQ domain must be disposed before the domain is removed.
Currently, the INTx IRQs are not disposed in pci-mvebu driver .remove() method, which causes the kernel to crash when unloading the driver and then reading /sys/kernel/debug/irq/irqs/<num> or /proc/interrupts.
Unmapping of the IRQs at this point of the .remove() method is safe, since the PCIe bus is already unregistered, and all its devices are unbound from their drivers and removed. If there was indeed any remaining use of PCIe resources, then it would mean that PCIe hotplug code is broken, and we have bigger problems.
---
Changes since v1:
- added explanation into commit message about why this is safe to do, as suggested by Andy. The explanation originally comes from Pali:
https://lore.kernel.org/linux-arm-kernel/20220809133911.hqi7eyskcq2sojia@pali/ --- drivers/pci/controller/pci-mvebu.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
drivers/pci/controller/pci-mvebu.c +8 -1
@@ -1683,8 +1683,15 @@ static void mvebu_pcie_remove(struct platform_device *pdev)irq_set_chained_handler_and_data (irq,NULL ,NULL );/* Remove IRQ domains. */ -if (port->intx_irq_domain) +if (port->intx_irq_domain) { +for (int j =0 ; j <PCI_NUM_INTX ; j++) { +int virq =irq_find_mapping (port->intx_irq_domain, j); + +if (virq >0 ) +irq_dispose_mapping (virq); + }irq_domain_remove (port->intx_irq_domain); + }/* Free config space for emulated root bridge. */ pci_bridge_emul_cleanup (&port->bridge);
Signature · 1 line
2.44.2
Quoted 16 lines · Marek Behún
From: Pali Rohár <[email protected]>
The documentation for the
irq_domain_remove()function says that all mappings within the IRQ domain must be disposed before the domain is removed.Currently, the INTx IRQs are not disposed in pci-mvebu driver .remove() method, which causes the kernel to crash when unloading the driver and then reading /sys/kernel/debug/irq/irqs/<num> or
/proc/interrupts.Unmapping of the IRQs at this point of the .remove() method is safe, since the PCIe bus is already unregistered, and all its devices are unbound from their drivers and removed. If there was indeed any remaining use of PCIe resources, then it would mean that PCIe hotplug code is broken, and we have bigger problems.
Quoted 7 lines
Fixes ec075262648f ("PCI: mvebu: Implement support for legacy INTx interrupts")Reported-by Hajo Noerenberg <[email protected]>Signed-off-by Pali Rohár <[email protected]>Reviewed-by Marek Behún <[email protected]>[ Marek: refactored a little, added more explanation to commit message ]Signed-off-by Marek Behún <[email protected]>Reviewed-by Manivannan Sadhasivam <[email protected]>
Andrew
[+cc Pali, seems like the author should be included, Thomas, Marc since they actually know about IRQs, unlike me]
Quoted 10 lines · Marek Behún
From: Pali Rohár <[email protected]>
The documentation for the
irq_domain_remove()function says that all mappings within the IRQ domain must be disposed before the domain is removed.Currently, the INTx IRQs are not disposed in pci-mvebu driver .remove() method, which causes the kernel to crash when unloading the driver and then reading /sys/kernel/debug/irq/irqs/<num> or
/proc/interrupts.
Quoted 8 lines
Unmapping of the IRQs at this point of the .remove() method is safe, since the PCIe bus is already unregistered, and all its devices are unbound from their drivers and removed. If there was indeed any remaining use of PCIe resources, then it would mean that PCIe hotplug code is broken, and we have bigger problems.
Fixes ec075262648f ("PCI: mvebu: Implement support for legacy INTx interrupts")Reported-by Hajo Noerenberg <[email protected]>
Is there a URL for this report?
Quoted 19 lines
Signed-off-by Pali Rohár <[email protected]>Reviewed-by Marek Behún <[email protected]>[ Marek: refactored a little, added more explanation to commit message ]Signed-off-by Marek Behún <[email protected]>Reviewed-by Manivannan Sadhasivam <[email protected]>---
Changes since v1:
- added explanation into commit message about why this is safe to do, as suggested by Andy. The explanation originally comes from Pali:https://lore.kernel.org/linux-arm-kernel/20220809133911.hqi7eyskcq2sojia@pali/ --- drivers/pci/controller/pci-mvebu.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)drivers/pci/controller/pci-mvebu.c +0
-0
drivers/pci/controller/pci-mvebu.c static void mvebu_pcie_remove(struct platform_device *pdev) · line 1691
Quoted 10 lines
drivers/pci/controller/pci-mvebu.c +6
-1@@ -1683,8 +1683,15 @@ static void mvebu_pcie_remove(struct platform_device *pdev)irq_set_chained_handler_and_data (irq,NULL ,NULL );/* Remove IRQ domains. */ -if (port->intx_irq_domain) +if (port->intx_irq_domain) { +for (int j =0 ; j <PCI_NUM_INTX ; j++) { +int virq =irq_find_mapping (port->intx_irq_domain, j); + +if (virq >0 ) +irq_dispose_mapping (virq);
I am not an IRQ expert, so all I can really do is compare this to usage in other drivers.
There are 20+ drivers in drivers/pci/controller, and I don't see irq_dispose_mapping() usage similar to this elsewhere. Does that mean most or all of the other drivers have a similar defect?
Quoted 9 lines
drivers/pci/controller/pci-mvebu.c +2
-0+ }irq_domain_remove (port->intx_irq_domain); + }/* Free config space for emulated root bridge. */ pci_bridge_emul_cleanup (&port->bridge);Signature · 1 line
2.44.2
Quoted 13 lines · Bjorn Helgaas
drivers/pci/controller/pci-mvebu.c static void mvebu_pcie_remove(struct platform_device *pdev) · line 1691
Quoted 8 lines · Marek Behún
drivers/pci/controller/pci-mvebu.c +6
-1/* Remove IRQ domains. */ -if (port->intx_irq_domain) +if (port->intx_irq_domain) { +for (int j =0 ; j <PCI_NUM_INTX ; j++) { +int virq =irq_find_mapping (port->intx_irq_domain, j); + +if (virq >0 ) +irq_dispose_mapping (virq);I am not an IRQ expert, so all I can really do is compare this to usage in other drivers.
Quoted 3 lines
There are 20+ drivers in
drivers/pci/controller, and I don't seeirq_dispose_mapping()usage similar to this elsewhere. Does that mean most or all of the other drivers have a similar defect?
Right.
But the real question is why is such a mapping not torn down by the entity (device, bridge, whatever) which set it up in the first place?
Thanks,
tglx
Quoted 14 lines · Thomas Gleixner
Quoted 13 lines · Bjorn Helgaas
drivers/pci/controller/pci-mvebu.c static void mvebu_pcie_remove(struct platform_device *pdev) · line 1691
Quoted 8 lines · Marek Behún
drivers/pci/controller/pci-mvebu.c +6
-1/* Remove IRQ domains. */ -if (port->intx_irq_domain) +if (port->intx_irq_domain) { +for (int j =0 ; j <PCI_NUM_INTX ; j++) { +int virq =irq_find_mapping (port->intx_irq_domain, j); + +if (virq >0 ) +irq_dispose_mapping (virq);I am not an IRQ expert, so all I can really do is compare this to usage in other drivers.
Quoted 8 lines
Quoted 3 lines
There are 20+ drivers in
drivers/pci/controller, and I don't seeirq_dispose_mapping()usage similar to this elsewhere. Does that mean most or all of the other drivers have a similar defect?Right.
But the real question is why is such a mapping not torn down by the entity (device, bridge, whatever) which set it up in the first place?
Marek/Pali, the commit log mentions a crash when unloading. Do you have a pointer to any details? Maybe there's a driver there that we can fix?
Bjorn
Quoted 5 lines · Bjorn Helgaas
Quoted 4 lines · Thomas Gleixner
Quoted 3 lines
There are 20+ drivers in
drivers/pci/controller, and I don't seeirq_dispose_mapping()usage similar to this elsewhere. Does that mean most or all of the other drivers have a similar defect?
Quoted 8 lines
Quoted 4 lines
Right.
But the real question is why is such a mapping not torn down by the entity (device, bridge, whatever) which set it up in the first place?
Marek/Pali, the commit log mentions a crash when unloading. Do you have a pointer to any details? Maybe there's a driver there that we can fix?
Pali already explained it in his reply to me, but for some stupid reason (my fault hitting reply instead of reply-all) this did not make the list.
Let me replay the discussion:
Quoted 14 lines
Quoted 13 lines
Quoted 13 lines
Pali:
I'm not expert neither. But IIRC it is because endpoint drivers do not callirq_dispose_mappingat their own for shared IRQs. And this is how shared PCI INTA-D interrupts are implemented in the kernel. First endpoint driver (e.g. wifi card) request for shared interrupt and kernel automatically creates irq mapping if it does not exist. Second endpoint driver (e.g. second wifi card) request for shared interrupt and kernel just returns existing mapping. And when the first endpoint driver is releasing its own irq handler it do not dispose irq mapping because it may be shared with other endpoint drivers (in this case with second wifi card). Seems that the owner of these shared mappings is the PCI controller driver and it has to do dispose them on its own removal (when releasing the domain for shared PCI IRQs).
Quoted 6 lines
Quoted 6 lines
tglx:
Working around this in a particular PCI controller driver is just wrong then. This wants a common cleanup function so all affected drivers which remove the INTX irqdomain can be fixed up without copy & pasta of the very same code. Something like the below and then change theirq_domain_remove()call for the INTX domain to use that function.
There was some additional back and forth but the actual patch which can be used for both INTX and other, e.g. error reporting, domains is below.
Sorry for taking this offlist unintenitonally.
The background of all this is that initially PCI[e] controllers were not removable/modular. Later on the whole modularization effort created this problem.
Thanks,
tglx
---
drivers/pci/irq.c +16 -0
@@ -278,3 +278,19 @@ int __weak pcibios_alloc_irq(struct pci_void __weak pcibios_free_irq (struct pci_dev *dev) { } + +#ifdef CONFIG_IRQDOMAIN +void pci_remove_irqdomain (struct irqdomain *domain,unsigned int nr_irqs) +{ +/* +* Comment explaining the oddity of this. +*/ +for (unsigned int i =0 ; j < nr_irqs; i++) { +int virq =irq_find_mapping (domain, i); + +if (virq >0 ) +irq_dispose_mapping (virq); + } +irq_domain_remove (domain); +} +#endif