genirq: Synchronize only with single thread on free_irq()
authorLukas Wunner <lukas@wunner.de>
Sun, 24 Jun 2018 08:35:30 +0000 (10:35 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Sun, 24 Jun 2018 12:17:27 +0000 (14:17 +0200)
commit519cc8652b3a1d3d0a02257afbe9573ad644da26
treee6809cf5de7b9e7c2bbf1a5ec80dbb1a5a1b9a06
parent836557bd58e5e65c05c73af9f6ebed885dbfccfc
genirq: Synchronize only with single thread on free_irq()

When pciehp is converted to threaded IRQ handling, removal of unplugged
devices below a PCIe hotplug port happens synchronously in the IRQ thread.
Removal of devices typically entails a call to free_irq() by their drivers.

If those devices share their IRQ with the hotplug port, __free_irq()
deadlocks because it calls synchronize_irq() to wait for all hard IRQ
handlers as well as all threads sharing the IRQ to finish.

Actually it's sufficient to wait only for the IRQ thread of the removed
device, so call synchronize_hardirq() to wait for all hard IRQ handlers to
finish, but no longer for any threads.  Compensate by rearranging the
control flow in irq_wait_for_interrupt() such that the device's thread is
allowed to run one last time after kthread_stop() has been called.

kthread_stop() blocks until the IRQ thread has completed.  On completion
the IRQ thread clears its oneshot thread_mask bit.  This is safe because
__free_irq() holds the request_mutex, thereby preventing __setup_irq() from
handing out the same oneshot thread_mask bit to a newly requested action.

Stack trace for posterity:
    INFO: task irq/17-pciehp:94 blocked for more than 120 seconds.
    schedule+0x28/0x80
    synchronize_irq+0x6e/0xa0
    __free_irq+0x15a/0x2b0
    free_irq+0x33/0x70
    pciehp_release_ctrl+0x98/0xb0
    pcie_port_remove_service+0x2f/0x40
    device_release_driver_internal+0x157/0x220
    bus_remove_device+0xe2/0x150
    device_del+0x124/0x340
    device_unregister+0x16/0x60
    remove_iter+0x1a/0x20
    device_for_each_child+0x4b/0x90
    pcie_port_device_remove+0x1e/0x30
    pci_device_remove+0x36/0xb0
    device_release_driver_internal+0x157/0x220
    pci_stop_bus_device+0x7d/0xa0
    pci_stop_bus_device+0x3d/0xa0
    pci_stop_and_remove_bus_device+0xe/0x20
    pciehp_unconfigure_device+0xb8/0x160
    pciehp_disable_slot+0x84/0x130
    pciehp_ist+0x158/0x190
    irq_thread_fn+0x1b/0x50
    irq_thread+0x143/0x1a0
    kthread+0x111/0x130

Signed-off-by: Lukas Wunner <lukas@wunner.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Cc: Mika Westerberg <mika.westerberg@linux.intel.com>
Cc: linux-pci@vger.kernel.org
Link: https://lkml.kernel.org/r/d72b41309f077c8d3bee6cc08ad3662d50b5d22a.1529828292.git.lukas@wunner.de
kernel/irq/manage.c