Re: [PATCH v2] iommu/vt-d: fix system hang on reboot -f

[Ethan Zhao]

在 2025/2/25 22:26, Jason Gunthorpe 写道:
> On Tue, Feb 25, 2025 at 04:54:54PM +0800, Ethan Zhao wrote:
> > > On 2025/2/25 14:48, Yunhui Cui wrote:
> > > > We found that executing the command ./a.out &;reboot -f (where a.out
> > > > is a
> > > > program that only executes a while(1) infinite loop) can
> > > > probabilistically
> > > > cause the system to hang in the intel_iommu_shutdown() function,
> > > > rendering
> > > > it unresponsive. Through analysis, we identified that the factors
> > > > contributing to this issue are as follows:
> > > >
> > > > 1. The reboot -f command does not prompt the kernel to notify the
> > > > application layer to perform cleanup actions, allowing the
> > > > application to
> > > > continue running.
> > > >
> > > > 2. When the kernel reaches the intel_iommu_shutdown() function, only the
> > > > BSP (Bootstrap Processor) CPU is operational in the system.
> > > >
> > > > 3. During the execution of intel_iommu_shutdown(), the function
> > > > down_write
> > > > (&dmar_global_lock) causes the process to sleep and be scheduled out.
> Why does this happen? If the kernel has shutdown other CPUs then what
> thread is holding the other side of this lock and why?
>
> > > > 4. At this point, though the processor's interrupt flag is not cleared,
> > > >    allowing interrupts to be accepted. However, only legacy devices
> > > > and NMI
> > > > (Non-Maskable Interrupt) interrupts could come in, as other interrupts
> > > > routing have already been disabled. If no legacy or NMI interrupts occur
> > > > at this stage, the scheduler will not be able to run.
> > > > 5. If the application got scheduled at this time is executing a
> > > > while(1)-
> > > > type loop, it will be unable to be preempted, leading to an infinite
> > > > loop
> > > > and causing the system to become unresponsive.
> If the schedular doesn't run how did we get from 4 -> 5?
>
> Maybe the issue is the shutdown handler here is running in the wrong
> time and it should not be running after the scheduler has been shut
> down.
>
> I don't think removing the lock is a great idea without more
> explanation.

Seems it is not so simple job to explain why there is no race window between
this iommu_shutdown() and following dmar_global_lock holders.

1. PCIe hotplug dmar_pci_bus_notifier()

2. mm_core_init detect_intel_iommu()

3. late_initcall dmar_free_unused_resources()

4. acpi attach dmar_device_hotplug()

5. pci_iommu_init intel_iommu_init() init_dmars()

6. rootfs_initcall ir_dev_scope_init()

though here is the last stage of reboot. then how about we turn back to v1

Just repalce with own_write() with down_write_trylock().

Thanks,

Ethan


> Jason

--
"firm, enduring, strong, and long-lived"