David Hoyer reports that powering pciehp slots up or down via sysfs may
hang: The call to wait_event() in pciehp_sysfs_enable_slot() and
_disable_slot() does not return because ctrl->ist_running remains true.
This flag, which was introduced by commit 157c1062fc ("PCI: pciehp: Avoid
returning prematurely from sysfs requests"), signifies that the IRQ thread
pciehp_ist() is running. It is set to true at the top of pciehp_ist() and
reset to false at the end. However there are two additional return
statements in pciehp_ist() before which the commit neglected to reset the
flag to false and wake up waiters for the flag.
That omission opens up the following race when powering up the slot:
* pciehp_ist() runs because a PCI_EXP_SLTSTA_PDC event was requested
by pciehp_sysfs_enable_slot()
* pciehp_ist() turns on slot power via the following call stack:
pciehp_handle_presence_or_link_change() -> pciehp_enable_slot() ->
__pciehp_enable_slot() -> board_added() -> pciehp_power_on_slot()
* after slot power is turned on, the link comes up, resulting in a
PCI_EXP_SLTSTA_DLLSC event
* the IRQ handler pciehp_isr() stores the event in ctrl->pending_events
and returns IRQ_WAKE_THREAD
* the IRQ thread is already woken (it's bringing up the slot), but the
genirq code remembers to re-run the IRQ thread after it has finished
(such that it can deal with the new event) by setting IRQTF_RUNTHREAD
via __handle_irq_event_percpu() -> __irq_wake_thread()
* the IRQ thread removes PCI_EXP_SLTSTA_DLLSC from ctrl->pending_events
via board_added() -> pciehp_check_link_status() in order to deal with
presence and link flaps per commit 6c35a1ac3d ("PCI: pciehp:
Tolerate initially unstable link")
* after pciehp_ist() has successfully brought up the slot, it resets
ctrl->ist_running to false and wakes up the sysfs requester
* the genirq code re-runs pciehp_ist(), which sets ctrl->ist_running
to true but then returns with IRQ_NONE because ctrl->pending_events
is empty
* pciehp_sysfs_enable_slot() is finally woken but notices that
ctrl->ist_running is true, hence continues waiting
The only way to get the hung task going again is to trigger a hotplug
event which brings down the slot, e.g. by yanking out the card.
The same race exists when powering down the slot because remove_board()
likewise clears link or presence changes in ctrl->pending_events per commit
3943af9d01 ("PCI: pciehp: Ignore Link State Changes after powering off a
slot") and thereby may cause a re-run of pciehp_ist() which returns with
IRQ_NONE without resetting ctrl->ist_running to false.
Fix by adding a goto label before the teardown steps at the end of
pciehp_ist() and jumping to that label from the two return statements which
currently neglect to reset the ctrl->ist_running flag.
Fixes: 157c1062fc ("PCI: pciehp: Avoid returning prematurely from sysfs requests")
Link: https://lore.kernel.org/r/cca1effa488065cb055120aa01b65719094bdcb5.1584530321.git.lukas@wunner.de
Reported-by: David Hoyer <David.Hoyer@netapp.com>
Signed-off-by: Lukas Wunner <lukas@wunner.de>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Cc: stable@vger.kernel.org # v4.19+
3e487d2e4a