Re: [PATCH v2 2/2] PCI: pciehp: Prevent deadlock on disconnect

[Mika Westerberg <mika.westerberg@linux.intel.com>]

On Tue, Oct 22, 2019 at 06:00:06PM -0500, Bjorn Helgaas wrote:
> On Mon, Aug 12, 2019 at 05:31:33PM +0300, Mika Westerberg wrote:
> > If there are more than one PCIe switch with hotplug downstream ports
> > hot-removing them leads to a following deadlock:
> > 
> >  INFO: task irq/126-pciehp:198 blocked for more than 120 seconds.
> >  "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >  irq/126-pciehp  D    0   198      2 0x80000000
> >  Call Trace:
> >   __schedule+0x2a2/0x880
> >   schedule+0x2c/0x80
> >   schedule_timeout+0x246/0x350
> >   ? ttwu_do_activate+0x67/0x90
> >   wait_for_completion+0xb7/0x140
> >   ? wake_up_q+0x80/0x80
> >   kthread_stop+0x49/0x110
> >   __free_irq+0x15c/0x2a0
> >   free_irq+0x32/0x70
> >   pcie_shutdown_notification+0x2f/0x50
> >   pciehp_remove+0x27/0x50
> >   pcie_port_remove_service+0x36/0x50
> >   device_release_driver_internal+0x18c/0x250
> >   device_release_driver+0x12/0x20
> >   bus_remove_device+0xec/0x160
> >   device_del+0x13b/0x350
> >   ? pcie_port_find_device+0x60/0x60
> >   device_unregister+0x1a/0x60
> >   remove_iter+0x1e/0x30
> >   device_for_each_child+0x56/0x90
> >   pcie_port_device_remove+0x22/0x40
> >   pcie_portdrv_remove+0x20/0x60
> >   pci_device_remove+0x3e/0xc0
> >   device_release_driver_internal+0x18c/0x250
> >   device_release_driver+0x12/0x20
> >   pci_stop_bus_device+0x6f/0x90
> >   pci_stop_bus_device+0x31/0x90
> >   pci_stop_and_remove_bus_device+0x12/0x20
> >   pciehp_unconfigure_device+0x88/0x140
> >   pciehp_disable_slot+0x6a/0x110
> >   pciehp_handle_presence_or_link_change+0x263/0x400
> >   pciehp_ist+0x1c9/0x1d0
> >   ? irq_forced_thread_fn+0x80/0x80
> >   irq_thread_fn+0x24/0x60
> >   irq_thread+0xeb/0x190
> >   ? irq_thread_fn+0x60/0x60
> >   kthread+0x120/0x140
> >   ? irq_thread_check_affinity+0xf0/0xf0
> >   ? kthread_park+0x90/0x90
> >   ret_from_fork+0x35/0x40
> >  INFO: task irq/190-pciehp:2288 blocked for more than 120 seconds.
> >  irq/190-pciehp  D    0  2288      2 0x80000000
> >  Call Trace:
> >   __schedule+0x2a2/0x880
> >   schedule+0x2c/0x80
> >   schedule_preempt_disabled+0xe/0x10
> >   __mutex_lock.isra.9+0x2e0/0x4d0
> >   ? __mutex_lock_slowpath+0x13/0x20
> >   __mutex_lock_slowpath+0x13/0x20
> >   mutex_lock+0x2c/0x30
> >   pci_lock_rescan_remove+0x15/0x20
> >   pciehp_unconfigure_device+0x4d/0x140
> >   pciehp_disable_slot+0x6a/0x110
> >   pciehp_handle_presence_or_link_change+0x263/0x400
> >   pciehp_ist+0x1c9/0x1d0
> >   ? irq_forced_thread_fn+0x80/0x80
> >   irq_thread_fn+0x24/0x60
> >   irq_thread+0xeb/0x190
> >   ? irq_thread_fn+0x60/0x60
> >   kthread+0x120/0x140
> >   ? irq_thread_check_affinity+0xf0/0xf0
> >   ? kthread_park+0x90/0x90
> >   ret_from_fork+0x35/0x40
> > 
> > What happens here is that the whole hierarchy is runtime resumed and the
> > parent PCIe downstream port, who got the hot-remove event, starts
> > removing devices below it taking pci_lock_rescan_remove() lock. When the
> > child PCIe port is runtime resumed it calls pciehp_check_presence()
> > which ends up calling pciehp_card_present() and pciehp_check_link_active().
> > Both of these read their parts of PCIe config space by calling helper
> > function pcie_capability_read_word(). Now, this function notices that
> > the underlying device is already gone and returns PCIBIOS_DEVICE_NOT_FOUND
> > with the capability value set to 0. When pciehp gets this value it
> > thinks that its child device is also hot-removed and schedules its IRQ
> > thread to handle the event.
> 
> I can't remember if there was a reason why 8c0d3a02c130 ("PCI: Add
> accessors for PCI Express Capability") reset *val to 0 if
> pci_read_config_word() fails.  It doesn't seem like the right thing;
> it seems like it would be better for it to be consistent with a plain
> pci_read_config_word().
> 
> > The deadlock happens when the child's IRQ thread runs and tries to
> > acquire pci_lock_rescan_remove() which is already taken by the parent
> > and the parent waits for the child's IRQ thread to finish.
> > 
> > We can prevent this from happening by checking the return value of
> > pcie_capability_read_word() and if it is PCIBIOS_DEVICE_NOT_FOUND stop
> > performing any hot-removal activities.
> > 
> > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > ---
> > No changes from the previous version.
> > 
> >  drivers/pci/hotplug/pciehp.h      |  6 +++---
> >  drivers/pci/hotplug/pciehp_core.c | 11 ++++++++---
> >  drivers/pci/hotplug/pciehp_ctrl.c |  4 ++--
> >  drivers/pci/hotplug/pciehp_hpc.c  | 32 +++++++++++++++++++++++--------
> >  4 files changed, 37 insertions(+), 16 deletions(-)
> > 
> > diff --git a/drivers/pci/hotplug/pciehp.h b/drivers/pci/hotplug/pciehp.h
> > index 8c51a04b8083..81c514ab9518 100644
> > --- a/drivers/pci/hotplug/pciehp.h
> > +++ b/drivers/pci/hotplug/pciehp.h
> > @@ -173,10 +173,10 @@ int pciehp_query_power_fault(struct controller *ctrl);
> >  void pciehp_green_led_on(struct controller *ctrl);
> >  void pciehp_green_led_off(struct controller *ctrl);
> >  void pciehp_green_led_blink(struct controller *ctrl);
> > -bool pciehp_card_present(struct controller *ctrl);
> > -bool pciehp_card_present_or_link_active(struct controller *ctrl);
> > +int pciehp_card_present(struct controller *ctrl);
> > +int pciehp_card_present_or_link_active(struct controller *ctrl);
> >  int pciehp_check_link_status(struct controller *ctrl);
> > -bool pciehp_check_link_active(struct controller *ctrl);
> > +int pciehp_check_link_active(struct controller *ctrl);
> >  void pciehp_release_ctrl(struct controller *ctrl);
> >  
> >  int pciehp_sysfs_enable_slot(struct hotplug_slot *hotplug_slot);
> > diff --git a/drivers/pci/hotplug/pciehp_core.c b/drivers/pci/hotplug/pciehp_core.c
> > index e9f82afa3773..4c032d75c874 100644
> > --- a/drivers/pci/hotplug/pciehp_core.c
> > +++ b/drivers/pci/hotplug/pciehp_core.c
> > @@ -134,10 +134,15 @@ static int get_adapter_status(struct hotplug_slot *hotplug_slot, u8 *value)
> >  {
> >  	struct controller *ctrl = to_ctrl(hotplug_slot);
> >  	struct pci_dev *pdev = ctrl->pcie->port;
> > +	int ret;
> >  
> >  	pci_config_pm_runtime_get(pdev);
> > -	*value = pciehp_card_present_or_link_active(ctrl);
> > +	ret = pciehp_card_present_or_link_active(ctrl);
> >  	pci_config_pm_runtime_put(pdev);
> > +	if (ret < 0)
> > +		return ret;
> > +
> > +	*value = ret;
> >  	return 0;
> >  }
> >  
> > @@ -153,13 +158,13 @@ static int get_adapter_status(struct hotplug_slot *hotplug_slot, u8 *value)
> >   */
> >  static void pciehp_check_presence(struct controller *ctrl)
> >  {
> > -	bool occupied;
> > +	int occupied;
> >  
> >  	down_read(&ctrl->reset_lock);
> >  	mutex_lock(&ctrl->state_lock);
> >  
> >  	occupied = pciehp_card_present_or_link_active(ctrl);
> > -	if ((occupied && (ctrl->state == OFF_STATE ||
> > +	if ((occupied > 0 && (ctrl->state == OFF_STATE ||
> >  			  ctrl->state == BLINKINGON_STATE)) ||
> >  	    (!occupied && (ctrl->state == ON_STATE ||
> >  			   ctrl->state == BLINKINGOFF_STATE)))
> > diff --git a/drivers/pci/hotplug/pciehp_ctrl.c b/drivers/pci/hotplug/pciehp_ctrl.c
> > index 631ced0ab28a..5a433cc8621f 100644
> > --- a/drivers/pci/hotplug/pciehp_ctrl.c
> > +++ b/drivers/pci/hotplug/pciehp_ctrl.c
> > @@ -221,7 +221,7 @@ void pciehp_handle_disable_request(struct controller *ctrl)
> >  
> >  void pciehp_handle_presence_or_link_change(struct controller *ctrl, u32 events)
> >  {
> > -	bool present, link_active;
> > +	int present, link_active;
> >  
> >  	/*
> >  	 * If the slot is on and presence or link has changed, turn it off.
> > @@ -252,7 +252,7 @@ void pciehp_handle_presence_or_link_change(struct controller *ctrl, u32 events)
> >  	mutex_lock(&ctrl->state_lock);
> >  	present = pciehp_card_present(ctrl);
> >  	link_active = pciehp_check_link_active(ctrl);
> > -	if (!present && !link_active) {
> > +	if (present <= 0 && link_active <= 0) {
> >  		mutex_unlock(&ctrl->state_lock);
> >  		return;
> >  	}
> > diff --git a/drivers/pci/hotplug/pciehp_hpc.c b/drivers/pci/hotplug/pciehp_hpc.c
> > index bd990e3371e3..1f918b043adb 100644
> > --- a/drivers/pci/hotplug/pciehp_hpc.c
> > +++ b/drivers/pci/hotplug/pciehp_hpc.c
> > @@ -201,13 +201,16 @@ static void pcie_write_cmd_nowait(struct controller *ctrl, u16 cmd, u16 mask)
> >  	pcie_do_write_cmd(ctrl, cmd, mask, false);
> >  }
> >  
> > -bool pciehp_check_link_active(struct controller *ctrl)
> > +int pciehp_check_link_active(struct controller *ctrl)
> >  {
> >  	struct pci_dev *pdev = ctrl_dev(ctrl);
> >  	u16 lnk_status;
> > -	bool ret;
> > +	int ret;
> > +
> > +	ret = pcie_capability_read_word(pdev, PCI_EXP_LNKSTA, &lnk_status);
> > +	if (ret == PCIBIOS_DEVICE_NOT_FOUND)
> > +		return -ENODEV;
> >  
> > -	pcie_capability_read_word(pdev, PCI_EXP_LNKSTA, &lnk_status);
> >  	ret = !!(lnk_status & PCI_EXP_LNKSTA_DLLLA);
> >  
> >  	if (ret)
> > @@ -373,13 +376,17 @@ void pciehp_get_latch_status(struct controller *ctrl, u8 *status)
> >  	*status = !!(slot_status & PCI_EXP_SLTSTA_MRLSS);
> >  }
> >  
> > -bool pciehp_card_present(struct controller *ctrl)
> > +int pciehp_card_present(struct controller *ctrl)
> >  {
> >  	struct pci_dev *pdev = ctrl_dev(ctrl);
> >  	u16 slot_status;
> > +	int ret;
> >  
> > -	pcie_capability_read_word(pdev, PCI_EXP_SLTSTA, &slot_status);
> > -	return slot_status & PCI_EXP_SLTSTA_PDS;
> > +	ret = pcie_capability_read_word(pdev, PCI_EXP_SLTSTA, &slot_status);
> > +	if (ret == PCIBIOS_DEVICE_NOT_FOUND)
> > +		return -ENODEV;
> 
> Isn't this racy?
> 
>                                         # pdev is present
>   pci_read_config_word
>     if (pci_dev_is_disconnected(pdev))  # currently false
>                                         # pdev is removed
>     pci_bus_read_config_word            # fails, returns ~0
>   slot_status = ~0
> 
> I think pci_read_config_word() checks pci_dev_is_disconnected() merely
> as an optimization.  Obviously it can't guarantee that the subsequent
> config access will succeed.
> 
> If pci_dev_is_disconnected() was false but the config read fails, I
> think we'll get ~0 data and return 1, i.e., "PDS was set".

Yes, it is racy. If a device can be hot-removed by user anytime we would
need to check each and every access and based on that stop touching the
hardware. That of course affects performance.

However, in practise this situation happens almost always when the
chained child hotplug port is resumed after the parent hotplug port
notices the downstream link is down. So first access to the now missing
hardware returns ~0 for the child in its resume hook.

> Shouldn't we check for slot_status being an error response (~0)
> instead of looking for PCIBIOS_DEVICE_NOT_FOUND?  There are 7 RsvdP
> bits in Slot Status, so ~0 is not a valid value for the register.
> 
> All 16 bits of Link Status are defined, but ~0 is still an invalid
> value because the Current Link Speed and Negotiated Link Width fields
> only define a few valid encodings.

Indeed that's a good point. I'll try that.

> > +	return !!(slot_status & PCI_EXP_SLTSTA_PDS);
> >  }
> >  
> >  /**
> > @@ -390,10 +397,19 @@ bool pciehp_card_present(struct controller *ctrl)
> >   * Presence Detect State bit, this helper also returns true if the Link Active
> >   * bit is set.  This is a concession to broken hotplug ports which hardwire
> >   * Presence Detect State to zero, such as Wilocity's [1ae9:0200].
> > + *
> > + * Returns: %1 if the slot is occupied and %0 if it is not. If the hotplug
> > + *	    port is not present anymore returns %-ENODEV.
> >   */
> > -bool pciehp_card_present_or_link_active(struct controller *ctrl)
> > +int pciehp_card_present_or_link_active(struct controller *ctrl)
> >  {
> > -	return pciehp_card_present(ctrl) || pciehp_check_link_active(ctrl);
> > +	int ret;
> > +
> > +	ret = pciehp_card_present(ctrl);
> > +	if (ret)
> > +		return ret;
> > +
> > +	return pciehp_check_link_active(ctrl);
> 
> The names of these functions seem misleading to me: all they can
> really tell us is "the card *was* present" or "the link *was* active"
> at some time in the past.  But the names make it so tempting to
> pretend that "the card *is* present" or "the link *is* active", and
> that may no longer be true.
> 
> I think names like "pciehp_card_absent()" and "pciehp_link_down()"
> would make it easier to think about these situations.

Fair enough. I'll create a separate patch for renaming them.