Use of copy_from_user in msm_gem_submit.c while holding a spin_lock

Hello,

I was wondering about the call to copy_from_user in function submit_lookup_objects for drive /gpu/drm/msm/msm_gem_submit.c It calls copy_from_user[1] in a spin_lock, which is not normally allowed, due to the possibility of a deadlock.

Is there some reason that I am overlooking why it is OK in this case? Is there some code in the same file which ensures that page fault will not occur when we are calling the function holding spin_lock?

Thank you.

[1] http://lxr.free-electrons.com/source/drivers/gpu/drm/msm/msm_gem_submit.c#L85

--
Vaishali

Quoted 6 lines · Vaishali Thakkar

Hello,

I was wondering about the call to copy_from_user in function submit_lookup_objects for drive /gpu/drm/msm/msm_gem_submit.c It calls copy_from_user[1] in a spin_lock, which is not normally allowed, due to the possibility of a deadlock.

Quoted 3 lines

Is there some reason that I am overlooking why it is OK in this case? Is there some code in the same file which ensures that page fault will not occur when we are calling the function holding spin_lock?

hmm, probably just that it isn't typical to use a swap file on these devices (and that lockdep/etc doesn't warn about it).. I guess we probably need some sort of slow-path where we drop the lock and try again in case there would be a fault..

Quoted 8 lines · Rob Clark

Quoted 6 lines · Vaishali Thakkar

Hello,

I was wondering about the call to copy_from_user in function submit_lookup_objects for drive /gpu/drm/msm/msm_gem_submit.c It calls copy_from_user[1] in a spin_lock, which is not normally allowed, due to the possibility of a deadlock.

Quoted 8 lines

Quoted 3 lines

Is there some reason that I am overlooking why it is OK in this case? Is there some code in the same file which ensures that page fault will not occur when we are calling the function holding spin_lock?

hmm, probably just that it isn't typical to use a swap file on these devices (and that lockdep/etc doesn't warn about it).. I guess we probably need some sort of slow-path where we drop the lock and try again in case there would be a fault..

Sigh... Folks, you don't need swap *at* *all* for copy_from_user() to block.

        /* get a zero-filled 64K buffer */
        addr = mmap(NULL, 65536, PROT_READ | PROT_WRITE,
                    MAP_ANONYMOUS | MAP_SHARED, -1, 0);
        if (addr < 0)
                piss off
        buffer = (void *)addr;
        ....
        pass buf to a syscall

and copy_from_user() in that syscall will have to allocate pages (and possibly page tables as well). Which can block just fine, no swap involved. Moreover, if you modify some parts of the buffer first, you will get the pages containing those modifications already present, but anything still untouched will
a) act as if it had been zeroed first and
b) possibly block on the first dereference, be it from kernel or from
userland. Worse yet, there's nothing to stop libc from using the above for calloc() and its ilk, with your application having no way to tell. As far as application is concerned, it has asked a library function to allocate and zero a piece of memory, got one and yes, it does appear to be properly zeroed.

The bottom line is, copy_from_user() can realistically block, without anything fishy going on in the userland setup.

Quoted 19 lines · Al Viro

Quoted 8 lines · Rob Clark

Quoted 6 lines · Vaishali Thakkar

Hello,

I was wondering about the call to copy_from_user in function submit_lookup_objects for drive /gpu/drm/msm/msm_gem_submit.c It calls copy_from_user[1] in a spin_lock, which is not normally allowed, due to the possibility of a deadlock.

Quoted 8 lines

Quoted 3 lines

Is there some reason that I am overlooking why it is OK in this case? Is there some code in the same file which ensures that page fault will not occur when we are calling the function holding spin_lock?

hmm, probably just that it isn't typical to use a swap file on these devices (and that lockdep/etc doesn't warn about it).. I guess we probably need some sort of slow-path where we drop the lock and try again in case there would be a fault..

Sigh... Folks, you don't need swap *at* *all* for copy_from_user() to block.

Quoted 8 lines

        /* get a zero-filled 64K buffer */
        addr = mmap(NULL, 65536, PROT_READ | PROT_WRITE,
                    MAP_ANONYMOUS | MAP_SHARED, -1, 0);
        if (addr < 0)
                piss off
        buffer = (void *)addr;
        ....
        pass buf to a syscall

Sure, I know that.. but if you pass random garbage cmstream to the gpu, it will crash (the gpu) too and/or result in corrupt rendering on screen, etc. GPU submit APIs don't exist for random end users, they exist for one user that knows what it is doing (ie. mesa).

I'm not saying that I shouldn't fix it (although not quite sure how yet.. taking/dropping the spinlock inside the loop is not a good option from a performance standpoint). What I am saying is that this is not something that can happen accidentally (as it could in the case of swap). But I agree that I should fix it somehow to avoid issues with an intentionally evil userspace.

If there is a copy_from_user() variant that will return an error instead of blocking, I think that is really what I want so I can implement a slow-path that drops the spin-lock temporarily.

BR,
-R

Quoted 13 lines

and copy_from_user() in that syscall will have to allocate pages (and possibly page tables as well). Which can block just fine, no swap involved. Moreover, if you modify some parts of the buffer first, you will get the pages containing those modifications already present, but anything still untouched will
a) act as if it had been zeroed first and
b) possibly block on the first dereference, be it from kernel or from
userland. Worse yet, there's nothing to stop libc from using the above for calloc() and its ilk, with your application having no way to tell. As far as application is concerned, it has asked a library function to allocate and zero a piece of memory, got one and yes, it does appear to be properly zeroed.

The bottom line is, copy_from_user() can realistically block, without anything fishy going on in the userland setup.

I'm not saying that I shouldn't fix it (although not quite sure how yet.. taking/dropping the spinlock inside the loop is not a good option from a performance standpoint). What I am saying is that this is not something that can happen accidentally (as it could in the case of swap). But I agree that I should fix it somehow to avoid issues with an intentionally evil userspace.

If there is a copy_from_user() variant that will return an error instead of blocking, I think that is really what I want so I can implement a slow-path that drops the spin-lock temporarily.

Quoted 6 lines · Rob Clark

I'm not saying that I shouldn't fix it (although not quite sure how yet.. taking/dropping the spinlock inside the loop is not a good option from a performance standpoint). What I am saying is that this is not something that can happen accidentally (as it could in the case of swap). But I agree that I should fix it somehow to avoid issues with an intentionally evil userspace.

I wouldn't count on that not happening by accident. With zero changes in mesa itself - it can be as simple as change of allocator in the bowels of libc or throwing libdmalloc into the link flags, etc. And most of the time it would've worked just fine, but the same call in a situation when most of the memory is occupied by dirty pagecache pages can end up having to wait for writeback.

Quoted 3 lines

If there is a copy_from_user() variant that will return an error instead of blocking, I think that is really what I want so I can implement a slow-path that drops the spin-lock temporarily.

*shrug*

pagefault_disable()/pagefault_enable() are there for purpose, so's __copy_from_user_inatomic()... Just remember that __copy_from_user_inatomic() does not check if the addresses are userland ones (i.e. the caller needs to check access_ok() itself) and it is *NOT* guaranteed to zero what it hadn't copied over. Currently it does zero tail on some, but not all architectures; come next cycle it and it will not do that zeroing on any of those.

hmm, looks like, at least on arm (not sure about arm64),

#define __copy_from_user_inatomic __copy_from_user

ie. copy_from_user() minus the access_ok() and memset in the !access_ok() path.. but maybe what I want is just the pagefault_disable() if that disables copy_from_user() being able to block..

Quoted 3 lines · Rob Clark

If there is a copy_from_user() variant that will return an error instead of blocking, I think that is really what I want so I can implement a slow-path that drops the spin-lock temporarily.

*shrug*

pagefault_disable()/pagefault_enable() are there for purpose, so's __copy_from_user_inatomic()... Just remember that __copy_from_user_inatomic() does not check if the addresses are userland ones (i.e. the caller needs to check access_ok() itself) and it is *NOT* guaranteed to zero what it hadn't copied over. Currently it does zero tail on some, but not all architectures; come next cycle it and it will not do that zeroing on any of those.

Quoted 8 lines · Rob Clark

hmm, looks like, at least on arm (not sure about arm64),

#define __copy_from_user_inatomic __copy_from_user

ie. copy_from_user() minus the access_ok() and memset in the !access_ok() path.. but maybe what I want is just the pagefault_disable() if that disables copy_from_user() being able to block..

On a bunch of platforms copy_from_user() starts with might_sleep(); again, that'll spread to all of the pretty soon.

Right now those primitives are very badly out of sync; this will change, but let's not add more PITA sources.

Quoted 7 lines · Al Viro

Quoted 3 lines · Rob Clark

If there is a copy_from_user() variant that will return an error instead of blocking, I think that is really what I want so I can implement a slow-path that drops the spin-lock temporarily.

*shrug*

Quoted 7 lines

pagefault_disable()/pagefault_enable() are there for purpose, so's __copy_from_user_inatomic()... Just remember that __copy_from_user_inatomic() does not check if the addresses are userland ones (i.e. the caller needs to check access_ok() itself) and it is *NOT* guaranteed to zero what it hadn't copied over. Currently it does zero tail on some, but not all architectures; come next cycle it and it will not do that zeroing on any of those.

Ok, this is what I came up with.. let me know what you think. The first hunk was just to see the problem in the first place (no idea if other places on arm would have problems w/ that hunk so it wouldn't be part of my fix+cc-stable patch.. but it seems like I good idea, I would have discovered this issue much sooner if we had it)

------

arch/arm/include/asm/uaccess.h +1 -0

index 35c9db8..ce2e182 100644
@@ -542,6 +542,7 @@ __clear_user(void __user *addr, unsigned long n)
 
 static inline unsigned long __must_check copy_from_user(void *to,
 const void __user *from, unsigned long n)
 {
+    might_fault();
     if (access_ok(VERIFY_READ, from, n))
         n = __copy_from_user(to, from, n);

else /* security hole - plug it */

drivers/gpu/drm/msm/msm_gem_submit.c +8 -0

index 5cd4e9b..3cca013 100644
@@ -66,6 +66,14 @@ void msm_gem_submit_free(struct msm_gem_submit *submit)
     kfree(submit);
 }
 
+static inline unsigned long __must_check
+copy_from_user_inatomic(void *to, const void __user *from, unsigned long n)
+{
+    if (access_ok(VERIFY_READ, from, n))
+        return __copy_from_user_inatomic(to, from, n);
+    return -EFAULT;
+}
+
 static int submit_lookup_objects(struct msm_gem_submit *submit,
         struct drm_msm_gem_submit *args, struct drm_file *file)
 {

drivers/gpu/drm/msm/msm_gem_submit.c +9 -4

     for (i = 0; i < args->nr_bos; i++) {
         struct drm_msm_gem_submit_bo submit_bo;

@@ -86,10 +95,15 @@ static int submit_lookup_objects(struct
 msm_gem_submit *submit,
          */
         submit->bos[i].flags = 0;
 
-        ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo));
-        if (ret) {
-            ret = -EFAULT;
-            goto out_unlock;
+        ret = copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo));
+        if (unlikely(ret)) {
+            pagefault_enable();
+            spin_unlock(&file->table_lock);
+            ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo));
+            if (ret)
+                return -EFAULT;
+            spin_lock(&file->table_lock);
+            pagefault_disable();
         }
 

         if (submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) {

drivers/gpu/drm/msm/msm_gem_submit.c +1 -0

@@ -130,6 +144,7 @@ static int submit_lookup_objects(struct
 msm_gem_submit *submit,
 
 out_unlock:
     submit->nr_bos = i;
+    pagefault_enable();
     spin_unlock(&file->table_lock);
 

     return ret;

------

danvet suggested the doubleplus-super-evil variant of the test program, using an unfaulted but mmap'd gem bo passed in to submit ioctl for the bo's table, which has the additional fun of wanting to acquire the already held struct_mutex in msm_gem_fault(). Which needed the further fix:

------

drivers/gpu/drm/msm/msm_gem.c +0 -0

index 6cd4af4..4502e4b 100644

drivers/gpu/drm/msm/msm_gem.c +1 -0

+
     /* Make sure we don't parallel update on a fault, nor move or remove
      * something from beneath our feet

      */

drivers/gpu/drm/msm/msm_gem.h +14 -0

index b2f13cf..160b635 100644
@@ -122,4 +122,18 @@ struct msm_gem_submit {
     } bos[0];
 };
 
+static inline bool mutex_is_locked_by(struct mutex *mutex,
+        struct task_struct *task)
+{
+    if (!mutex_is_locked(mutex))
+        return false;
+
+#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
+    return mutex->owner == task;
+#else
+    /* Since UP may be pre-empted, we cannot assume that we own the lock */
+    return false;
+#endif
+}
+
 #endif /* __MSM_GEM_H__ */

drivers/gpu/drm/msm/msm_gem_shrinker.c +0 -14

index 283d284..39429bb 100644
@@ -18,19 +18,6 @@
 #include "msm_drv.h"
 #include "msm_gem.h"
 
-static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
-{
-    if (!mutex_is_locked(mutex))
-        return false;
-
-#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
-    return mutex->owner == task;
-#else
-    /* Since UP may be pre-empted, we cannot assume that we own the lock */
-    return false;
-#endif
-}
-
 static bool msm_gem_shrinker_lock(struct drm_device *dev, bool *unlock)
 {
     if (!mutex_trylock(&dev->struct_mutex)) {
------

And yes, I know mutex_is_locked_by() is not super-awesome.. for the shrinker I plan to get rid of this with finer grained locking, although I'm not sure that totally helps with the doubleplus-evil-userspace vs submit ioctl case.. which I think could still cause mischief by passing an unfaulted bo in the bos table, which is processed (and with a finer-grained lock scheme, locked) first, and then using it's vaddr for the cmds table (where it would be faulted after already locked).. perhaps I can just disallow a gem bo to be used for either bos or cmds table in the submit ioctl (since there is no legit reason to allow that)

Note that there is actually no hardware with this gpu which is non-SMP so I don't mind kconfig to 'depends on SMP', at least for a short-term solution. I think any other solution isn't going to be something that candidate for stable.

BR,
-R

Quoted 5 lines · Rob Clark

drivers/gpu/drm/msm/msm_gem.c +0 -0

index 6cd4af4..4502e4b 100644

drivers/gpu/drm/msm/msm_gem.c int msm_gem_fault(struct vm_area_struct *vma, · line 210

Quoted 10 lines

drivers/gpu/drm/msm/msm_gem.c +6 -0

@@ -201,6 +201,13 @@ int msm_gem_fault(struct vm_area_struct *vma,
 struct vm_fault *vmf)
     pgoff_t pgoff;
     int ret;
 
+    /* I think this should only happen if userspace tries to pass a
+     * mmap'd but unfaulted gem bo vaddr into submit ioctl, triggering
+     * a page fault while struct_mutex is already held
+     */
+    if (mutex_is_locked_by(&dev->struct_mutex, current))
+        return VM_FAULT_SIGBUS;

This is an ok (well still horrible) heuristics for the shrinker, but for correctness it kinda doesn't cut it. What you need to do instead is drop all the locks, copy relocations into a temp memory area and then proceed in the msm command submission path above.

Also reentrant mutexes are evil ;-)

Quoted 22 lines · Daniel Vetter

Quoted 5 lines · Rob Clark

drivers/gpu/drm/msm/msm_gem.c +0 -0

index 6cd4af4..4502e4b 100644

drivers/gpu/drm/msm/msm_gem.c int msm_gem_fault(struct vm_area_struct *vma, · line 210

Quoted 10 lines

drivers/gpu/drm/msm/msm_gem.c +6 -0

@@ -201,6 +201,13 @@ int msm_gem_fault(struct vm_area_struct *vma,
 struct vm_fault *vmf)
     pgoff_t pgoff;
     int ret;
 
+    /* I think this should only happen if userspace tries to pass a
+     * mmap'd but unfaulted gem bo vaddr into submit ioctl, triggering
+     * a page fault while struct_mutex is already held
+     */
+    if (mutex_is_locked_by(&dev->struct_mutex, current))
+        return VM_FAULT_SIGBUS;

This is an ok (well still horrible) heuristics for the shrinker, but for correctness it kinda doesn't cut it. What you need to do instead is drop all the locks, copy relocations into a temp memory area and then proceed in the msm command submission path above.

Quoted 1 line

Also reentrant mutexes are evil ;-)

Please note that this is not a reentrant mutex in the fault path, it bails with VM_FAULT_SIGBUG!

There is never a legit reason to use a gem bo for the bos (or cmds) table in the ioctl, so while this may not be pretty, I believe it is an acceptable solution.

Quoted 23 lines · Rob Clark

Quoted 22 lines · Daniel Vetter

Quoted 5 lines · Rob Clark

drivers/gpu/drm/msm/msm_gem.c +0 -0

index 6cd4af4..4502e4b 100644

drivers/gpu/drm/msm/msm_gem.c int msm_gem_fault(struct vm_area_struct *vma, · line 210

Quoted 10 lines

drivers/gpu/drm/msm/msm_gem.c +6 -0

@@ -201,6 +201,13 @@ int msm_gem_fault(struct vm_area_struct *vma,
 struct vm_fault *vmf)
     pgoff_t pgoff;
     int ret;
 
+    /* I think this should only happen if userspace tries to pass a
+     * mmap'd but unfaulted gem bo vaddr into submit ioctl, triggering
+     * a page fault while struct_mutex is already held
+     */
+    if (mutex_is_locked_by(&dev->struct_mutex, current))
+        return VM_FAULT_SIGBUS;

This is an ok (well still horrible) heuristics for the shrinker, but for correctness it kinda doesn't cut it. What you need to do instead is drop all the locks, copy relocations into a temp memory area and then proceed in the msm command submission path above.

Quoted 4 lines

Quoted 1 line

Also reentrant mutexes are evil ;-)

Please note that this is not a reentrant mutex in the fault path, it bails with VM_FAULT_SIGBUG!

Except on UP it totally deadlocks ;-)

-Daniel

Quoted 3 lines

There is never a legit reason to use a gem bo for the bos (or cmds) table in the ioctl, so while this may not be pretty, I believe it is an acceptable solution.

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch