Regression in 2.6.28-rc1: When I added the new state_mutex which
prevents corruption of raw1394's internal state when accessed by
multithreaded client applications, the following possible though
highly unlikely deadlock slipped in:
Thread A: Thread B:
- acquire mmap_sem - raw1394_write() or raw1394_ioctl()
- raw1394_mmap() - acquire state_mutex
- acquire state_mutex - copy_to/from_user(), possible page fault:
acquire mmap_sem
The simplest fix is to use mutex_trylock() instead of mutex_lock() in
raw1394_mmap(). This changes the behavior under contention in a way
which is visible to userspace clients. However, since multithreaded
access was entirely buggy before state_mutex was added and libraw1394's
documentation advised application programmers to use a handle only in a
single thread, this change in behaviour should not be an issue in
practice at all.
Since we have to use mutex_trylock() in raw1394_mmap() regardless
whether /dev/raw1394 was opened with O_NONBLOCK or not, we now use
mutex_trylock() unconditionally everywhere for state_mutex, just to have
consistent behavior.
Reported-by: Johannes Weiner <[email protected]>
Signed-off-by: Stefan Richter <[email protected]>
---
Background: That new state_mutex went only in because raw1394_ioctl()
already head some weak protection by the Big Kernel Lock, which I
removed for the general reasons pro BKL removal (get better performance
with local locks; make the locking clearer, easier to debug, more
reliable).
drivers/ieee1394/raw1394.c | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)
Index: linux/drivers/ieee1394/raw1394.c
===================================================================
--- linux.orig/drivers/ieee1394/raw1394.c
+++ linux/drivers/ieee1394/raw1394.c
@@ -2268,7 +2268,8 @@ static ssize_t raw1394_write(struct file
return -EFAULT;
}
- mutex_lock(&fi->state_mutex);
+ if (!mutex_trylock(&fi->state_mutex))
+ return -EAGAIN;
switch (fi->state) {
case opened:
@@ -2548,7 +2549,8 @@ static int raw1394_mmap(struct file *fil
struct file_info *fi = file->private_data;
int ret;
- mutex_lock(&fi->state_mutex);
+ if (!mutex_trylock(&fi->state_mutex))
+ return -EAGAIN;
if (fi->iso_state == RAW1394_ISO_INACTIVE)
ret = -EINVAL;
@@ -2669,7 +2671,8 @@ static long raw1394_ioctl(struct file *f
break;
}
- mutex_lock(&fi->state_mutex);
+ if (!mutex_trylock(&fi->state_mutex))
+ return -EAGAIN;
switch (fi->iso_state) {
case RAW1394_ISO_INACTIVE:
--
Stefan Richter
-=====-==--- =-=- ==-=-
http://arcgraph.de/sr/
* Stefan Richter <[email protected]> wrote:
> Regression in 2.6.28-rc1: When I added the new state_mutex which
> prevents corruption of raw1394's internal state when accessed by
> multithreaded client applications, the following possible though
> highly unlikely deadlock slipped in:
>
> Thread A: Thread B:
> - acquire mmap_sem - raw1394_write() or raw1394_ioctl()
> - raw1394_mmap() - acquire state_mutex
> - acquire state_mutex - copy_to/from_user(), possible page fault:
> acquire mmap_sem
>
> The simplest fix is to use mutex_trylock() instead of mutex_lock() in
> raw1394_mmap(). This changes the behavior under contention in a way
> which is visible to userspace clients. However, since multithreaded
> access was entirely buggy before state_mutex was added and libraw1394's
> documentation advised application programmers to use a handle only in a
> single thread, this change in behaviour should not be an issue in
> practice at all.
>
> Since we have to use mutex_trylock() in raw1394_mmap() regardless
> whether /dev/raw1394 was opened with O_NONBLOCK or not, we now use
> mutex_trylock() unconditionally everywhere for state_mutex, just to have
> consistent behavior.
>
> Reported-by: Johannes Weiner <[email protected]>
> Signed-off-by: Stefan Richter <[email protected]>
> ---
>
> Background: That new state_mutex went only in because raw1394_ioctl()
> already head some weak protection by the Big Kernel Lock, which I
> removed for the general reasons pro BKL removal (get better performance
> with local locks; make the locking clearer, easier to debug, more
> reliable).
>
> drivers/ieee1394/raw1394.c | 9 ++++++---
> 1 file changed, 6 insertions(+), 3 deletions(-)
>
> Index: linux/drivers/ieee1394/raw1394.c
> ===================================================================
> --- linux.orig/drivers/ieee1394/raw1394.c
> +++ linux/drivers/ieee1394/raw1394.c
> @@ -2268,7 +2268,8 @@ static ssize_t raw1394_write(struct file
> return -EFAULT;
> }
>
> - mutex_lock(&fi->state_mutex);
> + if (!mutex_trylock(&fi->state_mutex))
> + return -EAGAIN;
>
> switch (fi->state) {
> case opened:
> @@ -2548,7 +2549,8 @@ static int raw1394_mmap(struct file *fil
> struct file_info *fi = file->private_data;
> int ret;
>
> - mutex_lock(&fi->state_mutex);
> + if (!mutex_trylock(&fi->state_mutex))
> + return -EAGAIN;
>
> if (fi->iso_state == RAW1394_ISO_INACTIVE)
> ret = -EINVAL;
> @@ -2669,7 +2671,8 @@ static long raw1394_ioctl(struct file *f
> break;
> }
>
> - mutex_lock(&fi->state_mutex);
> + if (!mutex_trylock(&fi->state_mutex))
> + return -EAGAIN;
So we can return a spurious -EAGAIN to user-space, if the state_mutex
is taken briefly by some other context?
Ingo
Ingo Molnar wrote:
> * Stefan Richter <[email protected]> wrote:
>
>> Regression in 2.6.28-rc1: When I added the new state_mutex which
>> prevents corruption of raw1394's internal state when accessed by
>> multithreaded client applications, the following possible though
>> highly unlikely deadlock slipped in:
>>
>> Thread A: Thread B:
>> - acquire mmap_sem - raw1394_write() or raw1394_ioctl()
>> - raw1394_mmap() - acquire state_mutex
>> - acquire state_mutex - copy_to/from_user(), possible page fault:
>> acquire mmap_sem
>>
>> The simplest fix is to use mutex_trylock() instead of mutex_lock() in
>> raw1394_mmap().
[...]
>> Since we have to use mutex_trylock() in raw1394_mmap() regardless
>> whether /dev/raw1394 was opened with O_NONBLOCK or not, we now use
>> mutex_trylock() unconditionally everywhere for state_mutex, just to have
>> consistent behavior.
[...]
>> Background: That new state_mutex went only in because raw1394_ioctl()
>> already head some weak protection by the Big Kernel Lock, which I
>> removed for the general reasons pro BKL removal (get better performance
>> with local locks; make the locking clearer, easier to debug, more
>> reliable).
>>
>> drivers/ieee1394/raw1394.c | 9 ++++++---
>> 1 file changed, 6 insertions(+), 3 deletions(-)
>>
>> Index: linux/drivers/ieee1394/raw1394.c
>> ===================================================================
>> --- linux.orig/drivers/ieee1394/raw1394.c
>> +++ linux/drivers/ieee1394/raw1394.c
>> @@ -2268,7 +2268,8 @@ static ssize_t raw1394_write(struct file
>> return -EFAULT;
>> }
>>
>> - mutex_lock(&fi->state_mutex);
>> + if (!mutex_trylock(&fi->state_mutex))
>> + return -EAGAIN;
>>
>> switch (fi->state) {
>> case opened:
>> @@ -2548,7 +2549,8 @@ static int raw1394_mmap(struct file *fil
>> struct file_info *fi = file->private_data;
>> int ret;
>>
>> - mutex_lock(&fi->state_mutex);
>> + if (!mutex_trylock(&fi->state_mutex))
>> + return -EAGAIN;
>>
>> if (fi->iso_state == RAW1394_ISO_INACTIVE)
>> ret = -EINVAL;
>> @@ -2669,7 +2671,8 @@ static long raw1394_ioctl(struct file *f
>> break;
>> }
>>
>> - mutex_lock(&fi->state_mutex);
>> + if (!mutex_trylock(&fi->state_mutex))
>> + return -EAGAIN;
>
> So we can return a spurious -EAGAIN to user-space, if the state_mutex
> is taken briefly by some other context?
.write() and .mmap() were not serialized against each other and against
.ioctl() at all in raw1394 before 2.6.28-rc1. Only .ioctl() was
(somewhat) serialized against itself due to traditionally being
implemented as .ioctl() instead of .unlocked_ioctl(). If there were
ever concurrent userspace contexts accessing the same file handle, they
would have corrupted raw1394's internal state eventually.
Application developers have been advised all the time to access the file
handle only from a single thread, ever.
So, before: Concurrent access resulted in hard to debug driver
malfunction, without clear error indication.
After: Concurrent access results in -EAGAIN.
I.e. since concurrent access was so broken before that hopefully no
raw1394/libraw1394 client exists which did use a single handle in more
than one thread, we are quite free how we fix that.
Or at least that's how I understand the situation. Would be good to
hear opinions from those who are more involved with libraw1394.
--
Stefan Richter
-=====-==--- =-=- ==-=-
http://arcgraph.de/sr/
I wrote:
> .write() and .mmap() were not serialized against each other and against
> .ioctl() at all in raw1394 before 2.6.28-rc1.
PS: There is a need for serialization to some degree because the client
registers itself with a controller via .write() (among many other things
that are implemented through .write()), manages isochronous I/O contexts
on this controller via .ioctl() and maps DMA buffers for isochronous I/O
via .mmap().
The raw1394 driver tracks respective state by means of two state
variables and some other variables, and accesses of the state variables
is not reentrant within one opener of /dev/raw1394. AFAICS the issue
exists between .write() and .write(), and independently of that between
.ioctl() and .ioctl() and between .ioctl() and .mmap().
Local mutex protection is the simplest way to fix that --- except that
there is this obscure issue of locking order between the driver's mutex
and the mmap semaphore outside the driver.
--
Stefan Richter
-=====-==--- =-=- ==-=-
http://arcgraph.de/sr/