Re: [SeaBIOS] seabios hangs when injecting an MSI-X interrupt in early booting.

[Kevin O'Connor]

On Thu, Mar 22, 2018 at 11:40:46AM +0800, Zhe Liu wrote:
> Hi!
> Recently, I encountered a seabios hanging problem. When I try to inject an
> MSI-X interrupt along with SMI in qemu like below:
> *************** patch below*********************
> diff --git a/target-i386/kvm.c b/target-i386/kvm.c- a/target-i386/kvm.c
>
> ---   a/target-i386/kvm.c
> +++   b/target-i386/kvm.c
>
> @@ -104,6 +104,7 @@ static uint32_t num_architectural_pmu_counters;
>  static int has_xsave;
>  static int has_xcrs;
>  static int has_pit_state2;
> +static int msi_count;
>
> static bool has_msr_mcg_ext_ctl;
>
> @@ -2870,12 +2871, 26 @@ void kvm_arch_pre_run(CPUState *cpu, struct
> kvm_run *run)
> DPRINTF("inject SMI\n");
> MSIMessage msg = {.address = 0x0, .data = 0x4062};
> if (msi_count == 1) {
> ret = kvm_irqchip_send_msi(kvm_state, msg);
> if (ret <0) {
> QEMU_LOG(LOG_ALERT, "MSI lost %s\n", strerror(-ret));
> }
> }
> msi_count++;
>
> ret = kvm_vcpu_ioctl(cpu, KVM_SMI);
> if (ret < 0)
> **************patch above*************************
>
> Then, when the vm will hang when booting.
> There will be only one line in the vnc displayer:
> "Guest has not initialized the display(yet)"
>
> I went through the codes(kvm qemu seabios), and I found the check_irqs
> function in seabios will allow interrupt for a while.
> When the interrupt triggered, the vcpu will lookup the seabios IDT table
> and jump to the interrupt handler.
>
> However, in ivt_init(void), the entries from 0x60 to 0x66 will be cleared,
> which is configured by the patch below.
>
> https://github.com/coreboot/seabios/commit/b164d2c1b8ff2dd764dcf064e2624dd4a...
>
> So the vcpu will jump to an undefined entry instead of the default entry.
>
> =========================================================================
> I removed the patch, and the bug seems to disappear. I wonder if this is an
> approprite solution.

This doesn't seem like a SeaBIOS issue.  The real-mode handlers
installed at 0x60-0x66 are intended for software interrupts; not for
hardware raised interrupts.

I don't know why the machine irq routing would deliver a hardware
interrupt to that vector, but I suspect that is what needs to be
looked at.

-Kevin

seabios hangs when injecting an MSI-X interrupt in early booting.

[Wanzongshun (Vincent)]

> So the vcpu will jump to an un
> This doesn't seem like a SeaBIOS issue.  >The real-mode handlers
> installed at 0x60-0x66 are intended for >software interrupts; not for
> hardware raised interrupts.

seabios hangs when injecting an MSI-X interrupt in early booting.

[Zhe Liu]

Dear Kevin,
    Thanks for getting back to me. Your replay helps me a lot. And there
are two things I want to make clear.
    1. > The real-mode handlers installed at 0x60-0x66 are intended for
software interrupts.
During the booting of the vm seabios, which modules will raise the software
interrupts? (Qemu? KVM? VirtualMachine?)
    2. In src/post.c: init_ivt, the handlers for 0x60-0x66 is set to none
as below.
    for (i=0x60; i<=0x66; i++)
        SET_IVT(i, SEGOFF(0, 0));
    If I change the handlers of 0x60-0x66 to default handler, which is
"iret" directly, is there some side effects?

2018-03-27 0:33 GMT+08:00 Kevin O'Connor :
> On Thu, Mar 22, 2018 at 11:40:46AM +0800, Zhe Liu wrote:
> > Hi!
> > Recently, I encountered a seabios hanging problem. When I try to inject
>
> an
> > MSI-X interrupt along with SMI in qemu like below:
> > *************** patch below*********************
> > diff --git a/target-i386/kvm.c b/target-i386/kvm.c- a/target-i386/kvm.c
> >
> > ---   a/target-i386/kvm.c
> > +++   b/target-i386/kvm.c
> >
> > @@ -104,6 +104,7 @@ static uint32_t num_architectural_pmu_counters;
> >  static int has_xsave;
> >  static int has_xcrs;
> >  static int has_pit_state2;
> > +static int msi_count;
> >
> > static bool has_msr_mcg_ext_ctl;
> >
> > @@ -2870,12 +2871, 26 @@ void kvm_arch_pre_run(CPUState *cpu, struct
> > kvm_run *run)
> > DPRINTF("inject SMI\n");
> > MSIMessage msg = {.address = 0x0, .data = 0x4062};
> > if (msi_count == 1) {
> > ret = kvm_irqchip_send_msi(kvm_state, msg);
> > if (ret <0) {
> > QEMU_LOG(LOG_ALERT, "MSI lost %s\n", strerror(-ret));
> > }
> > }
> > msi_count++;
> >
> > ret = kvm_vcpu_ioctl(cpu, KVM_SMI);
> > if (ret < 0)
> > **************patch above*************************
> >
> > Then, when the vm will hang when booting.
> > There will be only one line in the vnc displayer:
> > "Guest has not initialized the display(yet)"
> >
> > I went through the codes(kvm qemu seabios), and I found the check_irqs
> > function in seabios will allow interrupt for a while.
> > When the interrupt triggered, the vcpu will lookup the seabios IDT table
> > and jump to the interrupt handler.
> >
> > However, in ivt_init(void), the entries from 0x60 to 0x66 will be
>
> cleared,
> > which is configured by the patch below.
> >
> > https://github.com/coreboot/seabios/commit/b164d2c1b8ff2dd76
>
> 4dcf064e2624dd4aa78c112
> > So the vcpu will jump to an undefined entry instead of the default entry.
> >
> > ============================================================
>
> =============
> > I removed the patch, and the bug seems to disappear. I wonder if this is
>
> an
> > approprite solution.
>
> This doesn't seem like a SeaBIOS issue.  The real-mode handlers
> installed at 0x60-0x66 are intended for software interrupts; not for
> hardware raised interrupts.
>
> I don't know why the machine irq routing would deliver a hardware
> interrupt to that vector, but I suspect that is what needs to be
> looked at.
>
> -Kevin