The register_syscore_ops() function takes a mutex and might sleep. In the IOMMU initialization code it is invoked during irq-remapping setup already, where irqs are disabled. This causes a schedule-while-atomic bug: BUG: sleeping function called from invalid context at kernel/locking/mutex.c:747 in_atomic(): 0, irqs_disabled(): 1, pid: 1, name: swapper/0 no locks held by swapper/0/1. irq event stamp: 304 hardirqs last enabled at (303): [<ffffffff818a87b6>] _raw_spin_unlock_irqrestore+0x36/0x60 hardirqs last disabled at (304): [<ffffffff8235d440>] enable_IR_x2apic+0x79/0x196 softirqs last enabled at (36): [<ffffffff818ae75f>] __do_softirq+0x35f/0x4ec softirqs last disabled at (31): [<ffffffff810c1955>] irq_exit+0x105/0x120 CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.13.0-rc2.1.el7a.test.x86_64.debug #1 Hardware name: PowerEdge C6145 /040N24, BIOS 3.5.0 10/28/2014 Call Trace: dump_stack+0x85/0xca ___might_sleep+0x22a/0x260 __might_sleep+0x4a/0x80 __mutex_lock+0x58/0x960 ? iommu_completion_wait.part.17+0xb5/0x160 ? register_syscore_ops+0x1d/0x70 ? iommu_flush_all_caches+0x120/0x150 mutex_lock_nested+0x1b/0x20 register_syscore_ops+0x1d/0x70 state_next+0x119/0x910 iommu_go_to_state+0x29/0x30 amd_iommu_enable+0x13/0x23 Fix it by moving the register_syscore_ops() call to the next initialization step, which runs with irqs enabled.
Unified · Split
| @@ -2440,11 +2440,11 @@ static int __init state_next(void) · 2439 unmodified lines | ||
| 2440 | 2440 | break; |
| 2441 | 2441 | case IOMMU_ACPI_FINISHED: |
| 2442 | 2442 | early_enable_iommus(); |
| 2443 | − register_syscore_ops(&amd_iommu_syscore_ops); | |
| 2444 | 2443 | x86_platform.iommu_shutdown = disable_iommus; |
| 2445 | 2444 | init_state = IOMMU_ENABLED; |
| 2446 | 2445 | break; |
| 2447 | 2446 | case IOMMU_ENABLED: |
| 2447 | + register_syscore_ops(&amd_iommu_syscore_ops); | |
| 2448 | 2448 | ret = amd_iommu_init_pci(); |
| 2449 | 2449 | init_state = ret ? IOMMU_INIT_ERROR : IOMMU_PCI_INIT; |
| 2450 | 2450 | enable_iommus_v2(); |