On hot-pluggable ports, simultaneous surprise removal of multiple PCIe endpoints whether by pulling the card, powering it off, or dropping the link can trigger a system deadlock. Example: two PCIe endpoints are bound to vfio-pci and opened by the same process (fdA for device A, fdB for device B). 1. A PCIe-fault brings B’s link down, then A’s. 2. The PCI core starts removing B: - pciehp_unconfigure_device() takes pci_rescan_remove_lock - vfio-pci’s remove routine waits for fdB to be closed 3. While B is stuck, the core removes A: - pciehp_ist() takes the read side of reset_lock A - It blocks on pci_rescan_remove_lock already held by B 4. Killing the process closes fdA first. vfio_pci_core_close_device() tries to hot-reset A, so it needs the write side of reset_lock A. 5. The write request sleeps until the read lock from step 3 is released, but that reader is itself waiting for B’s lock -> deadlock. The stuck thread’s backtrace is as follows: /proc/1909/stack [<0>] vfio_unregister_group_dev+0x99/0xf0 [vfio] [<0>] vfio_pci_core_unregister_device+0x19/0xb0 [vfio_pci_core] [<0>] vfio_pci_remove+0x15/0x20 [vfio_pci] [<0>] pci_device_remove+0x3e/0xb0 [<0>] device_release_driver_internal+0x19b/0x200 [<0>] pci_stop_bus_device+0x6d/0x90 [<0>] pci_stop_and_remove_bus_device+0xe/0x20 [<0>] pciehp_unconfigure_device+0x8c/0x150 [<0>] pciehp_disable_slot+0x68/0x140 [<0>] pciehp_handle_presence_or_link_change+0x246/0x4c0 [<0>] pciehp_ist+0x244/0x280 [<0>] irq_thread_fn+0x1f/0x60 [<0>] irq_thread+0x1ac/0x290 [<0>] kthread+0xfa/0x240 [<0>] ret_from_fork+0x209/0x260 [<0>] ret_from_fork_asm+0x1a/0x30 /proc/1910/stack [<0>] pciehp_unconfigure_device+0x43/0x150 [<0>] pciehp_disable_slot+0x68/0x140 [<0>] pciehp_handle_presence_or_link_change+0x246/0x4c0 [<0>] pciehp_ist+0x244/0x280 [<0>] irq_thread_fn+0x1f/0x60 [<0>] irq_thread+0x1ac/0x290 [<0>] kthread+0xfa/0x240 [<0>] ret_from_fork+0x209/0x260 [<0>] ret_from_fork_asm+0x1a/0x30 /proc/6765/stack [<0>] pciehp_reset_slot+0x2c/0x70 [<0>] pci_reset_hotplug_slot+0x3e/0x60 [<0>] pci_reset_bus_function+0xcd/0x180 [<0>] cxl_reset_bus_function+0xc8/0x110 [<0>] __pci_reset_function_locked+0x4f/0xd0 [<0>] vfio_pci_core_disable+0x381/0x400 [vfio_pci_core] [<0>] vfio_pci_core_close_device+0x63/0xd0 [vfio_pci_core] [<0>] vfio_df_close+0x48/0x80 [vfio] [<0>] vfio_df_group_close+0x32/0x70 [vfio] [<0>] vfio_device_fops_release+0x1d/0x40 [vfio] [<0>] __fput+0xe6/0x2b0 [<0>] task_work_run+0x58/0x90 [<0>] do_exit+0x29b/0xa80 [<0>] do_group_exit+0x2c/0x80 [<0>] get_signal+0x8f9/0x900 [<0>] arch_do_signal_or_restart+0x29/0x210 [<0>] exit_to_user_mode_loop+0x8e/0x4f0 [<0>] do_syscall_64+0x262/0x630 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e Since the device is already disconnected, a hot-reset serves no purpose and risks generating additional PCIe link errors during the unplug sequence. Fix the issue by skipping hot-reset on Link-Down. Signed-off-by: Jinhui Guo --- drivers/vfio/pci/vfio_pci_core.c | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/drivers/vfio/pci/vfio_pci_core.c b/drivers/vfio/pci/vfio_pci_core.c index 3a11e6f450f7..f42051552dd4 100644 --- a/drivers/vfio/pci/vfio_pci_core.c +++ b/drivers/vfio/pci/vfio_pci_core.c @@ -678,6 +678,16 @@ void vfio_pci_core_disable(struct vfio_pci_core_device *vdev) if (!vdev->reset_works) goto out; + /* + * Skip hot reset on Link-Down. This avoids the reset_lock + * deadlock in pciehp_reset_slot() when multiple PCIe devices + * go down at the same time. + */ + if (pci_dev_is_disconnected(pdev)) { + vdev->needs_reset = false; + goto out; + } + pci_save_state(pdev); } -- 2.20.1