| Message ID | f7a8c135-a495-4ce6-bd49-405a45e7ea9a@rowland.harvard.edu (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | f4d1960764d8a70318b02f15203a1be2b2554ca1 |
| Headers | show |
| Series | None | expand |
On Fri, Mar 15, 2024 at 01:06:33PM -0400, Alan Stern wrote: > The show and store callback routines for the "disable" sysfs attribute > file in port.c acquire the device lock for the port's parent hub > device. This can cause problems if another process has locked the hub > to remove it or change its configuration: > > Removing the hub or changing its configuration requires the > hub interface to be removed, which requires the port device > to be removed, and device_del() waits until all outstanding > sysfs attribute callbacks for the ports have returned. The > lock can't be released until then. > > But the disable_show() or disable_store() routine can't return > until after it has acquired the lock. > > The resulting deadlock can be avoided by calling > sysfs_break_active_protection(). This will cause the sysfs core not > to wait for the attribute's callback routine to return, allowing the > removal to proceed. The disadvantage is that after making this call, > there is no guarantee that the hub structure won't be deallocated at > any moment. To prevent this, we have to acquire a reference to it > first by calling hub_get(). > > Signed-off-by: Alan Stern <stern@rowland.harvard.edu> > Cc: <stable@vger.kernel.org> # Needs the previous patch in this series What "previous patch"? I don't see this as a series even on lore.kernel.org. confused, greg k-h
On Tue, Mar 26, 2024 at 10:22:32AM +0100, Greg KH wrote: > On Fri, Mar 15, 2024 at 01:06:33PM -0400, Alan Stern wrote: > > The show and store callback routines for the "disable" sysfs attribute > > file in port.c acquire the device lock for the port's parent hub > > device. This can cause problems if another process has locked the hub > > to remove it or change its configuration: > > > > Removing the hub or changing its configuration requires the > > hub interface to be removed, which requires the port device > > to be removed, and device_del() waits until all outstanding > > sysfs attribute callbacks for the ports have returned. The > > lock can't be released until then. > > > > But the disable_show() or disable_store() routine can't return > > until after it has acquired the lock. > > > > The resulting deadlock can be avoided by calling > > sysfs_break_active_protection(). This will cause the sysfs core not > > to wait for the attribute's callback routine to return, allowing the > > removal to proceed. The disadvantage is that after making this call, > > there is no guarantee that the hub structure won't be deallocated at > > any moment. To prevent this, we have to acquire a reference to it > > first by calling hub_get(). > > > > Signed-off-by: Alan Stern <stern@rowland.harvard.edu> > > Cc: <stable@vger.kernel.org> # Needs the previous patch in this series > > What "previous patch"? I don't see this as a series even on > lore.kernel.org. Ah, found it, you sent it only to me for some reason, and not the lists. thanks, greg k-h
On Tue, Mar 26, 2024 at 10:23:25AM +0100, Greg KH wrote: > On Tue, Mar 26, 2024 at 10:22:32AM +0100, Greg KH wrote: > > On Fri, Mar 15, 2024 at 01:06:33PM -0400, Alan Stern wrote: > > > The show and store callback routines for the "disable" sysfs attribute > > > file in port.c acquire the device lock for the port's parent hub > > > device. This can cause problems if another process has locked the hub > > > to remove it or change its configuration: > > > > > > Removing the hub or changing its configuration requires the > > > hub interface to be removed, which requires the port device > > > to be removed, and device_del() waits until all outstanding > > > sysfs attribute callbacks for the ports have returned. The > > > lock can't be released until then. > > > > > > But the disable_show() or disable_store() routine can't return > > > until after it has acquired the lock. > > > > > > The resulting deadlock can be avoided by calling > > > sysfs_break_active_protection(). This will cause the sysfs core not > > > to wait for the attribute's callback routine to return, allowing the > > > removal to proceed. The disadvantage is that after making this call, > > > there is no guarantee that the hub structure won't be deallocated at > > > any moment. To prevent this, we have to acquire a reference to it > > > first by calling hub_get(). > > > > > > Signed-off-by: Alan Stern <stern@rowland.harvard.edu> > > > Cc: <stable@vger.kernel.org> # Needs the previous patch in this series > > > > What "previous patch"? I don't see this as a series even on > > lore.kernel.org. > > Ah, found it, you sent it only to me for some reason, and not the lists. Oops. My apologies... maybe I'm getting rusty at this. That was definitely a mistake, not intentional. Alan Stern
Index: usb-devel/drivers/usb/core/port.c =================================================================== --- usb-devel.orig/drivers/usb/core/port.c +++ usb-devel/drivers/usb/core/port.c @@ -55,11 +55,22 @@ static ssize_t disable_show(struct devic u16 portstatus, unused; bool disabled; int rc; + struct kernfs_node *kn; + hub_get(hub); rc = usb_autopm_get_interface(intf); if (rc < 0) - return rc; + goto out_hub_get; + /* + * Prevent deadlock if another process is concurrently + * trying to unregister hdev. + */ + kn = sysfs_break_active_protection(&dev->kobj, &attr->attr); + if (!kn) { + rc = -ENODEV; + goto out_autopm; + } usb_lock_device(hdev); if (hub->disconnected) { rc = -ENODEV; @@ -69,9 +80,13 @@ static ssize_t disable_show(struct devic usb_hub_port_status(hub, port1, &portstatus, &unused); disabled = !usb_port_is_power_on(hub, portstatus); -out_hdev_lock: + out_hdev_lock: usb_unlock_device(hdev); + sysfs_unbreak_active_protection(kn); + out_autopm: usb_autopm_put_interface(intf); + out_hub_get: + hub_put(hub); if (rc) return rc; @@ -89,15 +104,26 @@ static ssize_t disable_store(struct devi int port1 = port_dev->portnum; bool disabled; int rc; + struct kernfs_node *kn; rc = kstrtobool(buf, &disabled); if (rc) return rc; + hub_get(hub); rc = usb_autopm_get_interface(intf); if (rc < 0) - return rc; + goto out_hub_get; + /* + * Prevent deadlock if another process is concurrently + * trying to unregister hdev. + */ + kn = sysfs_break_active_protection(&dev->kobj, &attr->attr); + if (!kn) { + rc = -ENODEV; + goto out_autopm; + } usb_lock_device(hdev); if (hub->disconnected) { rc = -ENODEV; @@ -118,9 +144,13 @@ static ssize_t disable_store(struct devi if (!rc) rc = count; -out_hdev_lock: + out_hdev_lock: usb_unlock_device(hdev); + sysfs_unbreak_active_protection(kn); + out_autopm: usb_autopm_put_interface(intf); + out_hub_get: + hub_put(hub); return rc; }
The show and store callback routines for the "disable" sysfs attribute file in port.c acquire the device lock for the port's parent hub device. This can cause problems if another process has locked the hub to remove it or change its configuration: Removing the hub or changing its configuration requires the hub interface to be removed, which requires the port device to be removed, and device_del() waits until all outstanding sysfs attribute callbacks for the ports have returned. The lock can't be released until then. But the disable_show() or disable_store() routine can't return until after it has acquired the lock. The resulting deadlock can be avoided by calling sysfs_break_active_protection(). This will cause the sysfs core not to wait for the attribute's callback routine to return, allowing the removal to proceed. The disadvantage is that after making this call, there is no guarantee that the hub structure won't be deallocated at any moment. To prevent this, we have to acquire a reference to it first by calling hub_get(). Signed-off-by: Alan Stern <stern@rowland.harvard.edu> Cc: <stable@vger.kernel.org> # Needs the previous patch in this series --- drivers/usb/core/port.c | 38 ++++++++++++++++++++++++++++++++++---- 1 file changed, 34 insertions(+), 4 deletions(-)