lkml.org 
[lkml]   [2013]   [Jan]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] pci-sysfs: replace mutex_lock with mutex_trylock to avoid potential deadlock situation
[+cc Yinghai]

On Fri, Jan 25, 2013 at 3:02 AM, Gu Zheng <guz.fnst@cn.fujitsu.com> wrote:
> Hi Bjorn,
> Thanks for your review and comments! Please refer to inlined comments below.
>
> On 01/25/2013 07:12 AM, Bjorn Helgaas wrote:
>
>> On Thu, Dec 27, 2012 at 12:42 AM, Lin Feng <linfeng@cn.fujitsu.com> wrote:
>>> There is a potential deadlock situation when we manipulate the pci-sysfs user
>>> interfaces from different bus hierarchy simultaneously, described as following:
>>>
>>> path1: sysfs remove device: | path2: sysfs rescan device:
>>> sysfs_schedule_callback_work() | sysfs_write_file()
>>> remove_callback() | flush_write_buffer()
>>> *1* mutex_lock(&pci_remove_rescan_mutex)|*2* sysfs_get_active(attr_sd)
>>> ... | dev_attr_store()
>>> device_remove_file() | dev_rescan_store()
>>> ... |*4* mutex_lock(&pci_remove_rescan_mutex)
>>> *3* sysfs_deactivate(sd) | ...
>>> wait_for_completion() |*5* sysfs_put_active(attr_sd)
>>> *6* mutex_unlock(&pci_remove_rescan_mutex)
> ...snip...
>>> Reported-by: Taku Izumi <izumi.taku@jp.fujitsu.com>
>>> Signed-off-by: Lin Feng <linfeng@cn.fujitsu.com>
>>> Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
>>> ---
>>> drivers/pci/pci-sysfs.c | 42 ++++++++++++++++++++++++++----------------
>>> 1 files changed, 26 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/drivers/pci/pci-sysfs.c b/drivers/pci/pci-sysfs.c
>>> index 05b78b1..d2efbb0 100644
>>> --- a/drivers/pci/pci-sysfs.c
>>> +++ b/drivers/pci/pci-sysfs.c
>>> @@ -295,10 +295,13 @@ static ssize_t bus_rescan_store(struct bus_type *bus,
> const char *buf,
>>> return -EINVAL;
>>>
>>> if (val) {
>>> - mutex_lock(&pci_remove_rescan_mutex);
>>> - while ((b = pci_find_next_bus(b)) != NULL)
>>> - pci_rescan_bus(b);
>>> - mutex_unlock(&pci_remove_rescan_mutex);
>>> + if (mutex_trylock(&pci_remove_rescan_mutex)) {
>>> + while ((b = pci_find_next_bus(b)) != NULL)
>>> + pci_rescan_bus(b);
>>> + mutex_unlock(&pci_remove_rescan_mutex);
>>> + } else {
>>> + return 0;
>> What are the semantics of returning 0 from a sysfs store function?
>> Does the user's write just get dropped? I would think we'd return
>> "count" for that case.
>
> Oh, yes, return "count" seems suitable here, although we did not reach the
> user's target goal(rescan the bus), but the user's write has been flushed yet.
> But the user still can not judge whether pci_rescan_bus() was successfully done
> only by the return value. Shall we return a suitable error here to tell the user
> that his write was written, but pci_rescan_bus() was not done ?
>
>> Is there some sort of automatic retry in libc
>> or something if we return zero?
>
> No, there is not any extra operations in libc if we return zero indeed.
>
>> Are you relying on the user code to
>> notice that nothing was written and do its own retry?
>>
>
> Yes, but it seems impractical.
>
>> The last seems most likely, but that seems like it complicates the
>> user's life unnecessarily.
>>
>>> + }
>>> }
>>> return count;
>>> }
>>> @@ -319,9 +322,12 @@ dev_rescan_store(struct device *dev, struct device_attribute *attr,
>>> return -EINVAL;
>>>
>>> if (val) {
>>> - mutex_lock(&pci_remove_rescan_mutex);
>>> - pci_rescan_bus(pdev->bus);
>>> - mutex_unlock(&pci_remove_rescan_mutex);
>>> + if (mutex_trylock(&pci_remove_rescan_mutex)) {
>>> + pci_rescan_bus(pdev->bus);
>>> + mutex_unlock(&pci_remove_rescan_mutex);
>>> + } else {
>>> + return 0;
>>> + }
>>> }
>>> return count;
>>> }
>>> @@ -330,9 +336,10 @@ static void remove_callback(struct device *dev)
>>> {
>>> struct pci_dev *pdev = to_pci_dev(dev);
>>>
>>> - mutex_lock(&pci_remove_rescan_mutex);
>>> - pci_stop_and_remove_bus_device(pdev);
>>> - mutex_unlock(&pci_remove_rescan_mutex);
>>> + if (mutex_trylock(&pci_remove_rescan_mutex)) {
>>> + pci_stop_and_remove_bus_device(pdev);
>>> + mutex_unlock(&pci_remove_rescan_mutex);
>>> + }
>> In the other cases, I think the user will at least get some
>> indication, e.g., a write() that returns zero, when we abort. But
>> here, we silently skip the pci_stop_and_remove_bus_device(). That
>> sounds wrong to me. What actually happens here, and why is it OK to
>> skip it?
>
> Yeah, the hasty skip seems not suitable. We should give out some information
> here, if we can not do pci_stop_and_remove_bus_device().
>
>> Can we avoid the deadlock by queuing these in a workqueue instead of
>> using the mutex_trylock() approach?
>>
>
> No, I think use a workqueue to queue the rescan routine into workqueue as the
> remove is not suitable.
> After we queue the scan-bus work into workqueue, the rescan routine can
> return directly(case1) or wait until work is completed(case2).
> case1:
> If we return directly after we queue the scan-bus work into workqueue.
> Maybe this work will be scheduled some time later. If there is a
> user's routine want to use a new-added device before the scan-bus work is
> successfully done will fail. It can avoid the deadlock, but the rescan routine
> is executed asynchronously.
> case2:
> If we wait until work is completed after we queue the scan-bus work into
> workqueue. The s_active of the sys_dirent is still hold by us, so this approach
> could not avoid the deadlock.

I don't think the asynchronous nature of case 1 is a fatal problem.
I'm not sure we can guarantee that a specific device is added and
ready for use when the sysfs write completes, so I'm dubious about
user space making assumptions about a device being ready. It should
probably use a different mechanism, like udev, to learn about a new
device.

I'm sorry that you tripped over this deadlock, because now I'm worried
about related locking issues outside sysfs :) The mutex you're
fiddling with is only in sysfs, but the routines *protected* by that
mutex are used in other places, too. So what happens when a hotplug
driver does a rescan at the same time a user does a rescan or remove
via sysfs? I don't even know what the rules are for protecting
scan/remove, but I don't have confidence that the issue you're fixing
is the only one.

Bjorn

>>> }
>>>
>>> static ssize_t
>>> @@ -366,12 +373,15 @@ dev_bus_rescan_store(struct device *dev, struct
> device_attribute *attr,
>>> return -EINVAL;
>>>
>>> if (val) {
>>> - mutex_lock(&pci_remove_rescan_mutex);
>>> - if (!pci_is_root_bus(bus) && list_empty(&bus->devices))
>>> - pci_rescan_bus_bridge_resize(bus->self);
>>> - else
>>> - pci_rescan_bus(bus);
>>> - mutex_unlock(&pci_remove_rescan_mutex);
>>> + if (mutex_trylock(&pci_remove_rescan_mutex)) {
>>> + if (!pci_is_root_bus(bus) && list_empty(&bus->devices))
>>> + pci_rescan_bus_bridge_resize(bus->self);
>>> + else
>>> + pci_rescan_bus(bus);
>>> + mutex_unlock(&pci_remove_rescan_mutex);
>>> + } else {
>>> + return 0;
>>> + }
>>> }
>>> return count;
>>> }
>>> --
>>> 1.7.1
>>>
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>


\
 
 \ /
  Last update: 2013-01-25 19:41    [W:0.064 / U:0.756 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site