x86: ioapic needs check attr when programmed more than once
PATCH 1/1 x86: ioapic needs check attr when programmed more than once
When programming ioapic pinX more than once, current code does not check whether the later attr (trigger&polarity) is the same as the former or not. This causes a broken semantic.
Fix it by reporting -EBUSY, when attr is different.
--- arch/x86/kernel/acpi/boot.c | 5 ++++- arch/x86/kernel/apic/io_apic.c | 7 ++++++- 2 files changed, 10 insertions(+), 2 deletions(-)
arch/x86/kernel/acpi/boot.c +4 -1
@@ -1113,6 +1113,7 @@ int mp_register_gsi(struct device *dev, u32 gsi, int trigger, int polarity)int ioapic;int ioapic_pin;struct io_apic_irq_attr irq_attr; +int ret;if (acpi_irq_model !=ACPI_IRQ_MODEL_IOAPIC )return gsi;
@@ -1142,7 +1143,9 @@ int mp_register_gsi(struct device *dev, u32 gsi, int trigger, int polarity)set_io_apic_irq_attr (&irq_attr, ioapic, ioapic_pin, trigger ==ACPI_EDGE_SENSITIVE ?0 :1 , polarity ==ACPI_ACTIVE_HIGH ?0 :1 ); -io_apic_set_pci_routing (dev,gsi_to_irq (gsi), &irq_attr); + ret =io_apic_set_pci_routing (dev,gsi_to_irq (gsi), &irq_attr); +if (ret <0 ) + gsi =INT_MIN ;return gsi; }
arch/x86/kernel/apic/io_apic.c +6 -1
@@ -3375,12 +3375,17 @@ int io_apic_setup_irq_pin_once(unsigned int irq, int node, {unsigned int ioapic_idx = attr->ioapic, pin = attr->ioapic_pin;int ret; +struct IO_APIC_route_entry orig;/* Avoid redundant programming */ if (test_bit (pin, ioapics[ioapic_idx].pin_programmed)) {pr_debug ("Pin %d-%d already programmed\n" ,mpc_ioapic_id (ioapic_idx), pin); -return 0 ; + orig =ioapic_read_entry (attr->ioapic, pin); +if (attr->trigger == orig.trigger && + attr->polarity == orig.polarity) +return 0 ; +return -EBUSY ; } ret =io_apic_setup_irq_pin (irq, node, attr);if (!ret)
Signature · 1 line
1.8.1.4
Quoted 5 lines · Liu Ping Fan
When programming ioapic pinX more than once, current code does not check whether the later attr (trigger&polarity) is the same as the former or not. This causes a broken semantic.
Fix it by reporting
-EBUSY, when attr is different.
Was this observed in real life somehow, and if yes, what is the before/after behavior?
Thanks,
Ingo
Quoted 12 lines · Ingo Molnar
Quoted 5 lines · Liu Ping Fan
When programming ioapic pinX more than once, current code does not check whether the later attr (trigger&polarity) is the same as the former or not. This causes a broken semantic.
Fix it by reporting
-EBUSY, when attr is different.Was this observed in real life somehow, and if yes, what is the before/after behavior?
Using qemu q35 machine, I found the ioapic's ioredtbl[] will never has low-active, even the hpet driver registered it. After tracing, I found it shared a high-level active IRQ line with other device. So in fact, the acpi_register_gsi(, ACPI_ACTIVE_LOW) in hpet driver fail, but it did not detect it.
The effect in qemu: when hpet-dev assert low-level, the kernel can not respond.
Thanks,
Pingfan
Quoted 13 lines · Liu ping fan
Quoted 12 lines · Ingo Molnar
Quoted 5 lines · Liu Ping Fan
When programming ioapic pinX more than once, current code does not check whether the later attr (trigger&polarity) is the same as the former or not. This causes a broken semantic.
Fix it by reporting
-EBUSY, when attr is different.Was this observed in real life somehow, and if yes, what is the before/after behavior?
Quoted 7 lines
Using qemu q35 machine, I found the ioapic's ioredtbl[] will never has low-active, even the hpet driver registered it. After tracing, I found it shared a high-level active IRQ line with other device. So in fact, the
acpi_register_gsi(, ACPI_ACTIVE_LOW)in hpet driver fail, but it did not detect it.
The effect in qemu: when hpet-dev assert low-level, the kernel can not respond.
After changing, I can observe the low-active be set in ioredtbl[x], and with some bug fix in qemu's hpet, the kernel can work.
Quoted 3 lines
Thanks,
Pingfan
Quoted 21 lines · Liu ping fan
Quoted 13 lines · Liu ping fan
Quoted 12 lines · Ingo Molnar
Quoted 5 lines · Liu Ping Fan
When programming ioapic pinX more than once, current code does not check whether the later attr (trigger&polarity) is the same as the former or not. This causes a broken semantic.
Fix it by reporting
-EBUSY, when attr is different.Was this observed in real life somehow, and if yes, what is the before/after behavior?
Quoted 7 lines
Using qemu q35 machine, I found the ioapic's ioredtbl[] will never has low-active, even the hpet driver registered it. After tracing, I found it shared a high-level active IRQ line with other device. So in fact, the
acpi_register_gsi(, ACPI_ACTIVE_LOW)in hpet driver fail, but it did not detect it.
The effect in qemu: when hpet-dev assert low-level, the kernel can not respond.
Quoted 2 lines
After changing, I can observe the low-active be set in ioredtbl[x], and with some bug fix in qemu's hpet, the kernel can work.
Okay, so because in practice this kind of information is much more important to users than anything else in the changelog please put all this into the changelog and re-send.
Thanks,
Ingo