| Message ID | 20210104120929.294063-2-phil@raspberrypi.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | A brace of vchiq bulk transfer fixes | expand |
Am 04.01.21 um 13:09 schrieb Phil Elwell: > The addition of the local 'userdata' pointer to > vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor > WAITING modes are used, in which case the value provided by the > caller is replaced with a NULL. > > Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") > > Signed-off-by: Phil Elwell <phil@raspberrypi.com> Tested-by: Stefan Wahren <stefan.wahren@i2se.com>
On Mon, Jan 04, 2021 at 12:09:27PM +0000, Phil Elwell wrote: > The addition of the local 'userdata' pointer to > vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor > WAITING modes are used, in which case the value provided by the > caller is replaced with a NULL. > > Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") > > Signed-off-by: Phil Elwell <phil@raspberrypi.com> > --- > drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > index f500a7043805..2a8883673ba1 100644 > --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > struct vchiq_service *service; > struct bulk_waiter_node *waiter = NULL; > bool found = false; > - void *userdata = NULL; > + void *userdata; > int status = 0; > int ret; > > @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > "found bulk_waiter %pK for pid %d", waiter, > current->pid); > userdata = &waiter->bulk_waiter; > + } else { > + userdata = args->userdata; "args->userdata" is marked as a user pointer so we really don't want to mix user and kernel pointers here. Presumably this opens up a large security hole. regards, dan carpenter
On 04/01/2021 18:31, Dan Carpenter wrote: > On Mon, Jan 04, 2021 at 12:09:27PM +0000, Phil Elwell wrote: >> The addition of the local 'userdata' pointer to >> vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor >> WAITING modes are used, in which case the value provided by the >> caller is replaced with a NULL. >> >> Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") >> >> Signed-off-by: Phil Elwell <phil@raspberrypi.com> >> --- >> drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- >> 1 file changed, 3 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c >> index f500a7043805..2a8883673ba1 100644 >> --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c >> +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c >> @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, >> struct vchiq_service *service; >> struct bulk_waiter_node *waiter = NULL; >> bool found = false; >> - void *userdata = NULL; >> + void *userdata; >> int status = 0; >> int ret; >> >> @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, >> "found bulk_waiter %pK for pid %d", waiter, >> current->pid); >> userdata = &waiter->bulk_waiter; >> + } else { >> + userdata = args->userdata; > > "args->userdata" is marked as a user pointer so we really don't want to > mix user and kernel pointers here. Presumably this opens up a large > security hole. It's an opaque, pointer-sized token that only exists to bereturned to userspace (or not, without this patch) - it's hard to see that as a security hole. Phil
On Mon, Jan 04, 2021 at 07:26:42PM +0000, Phil Elwell wrote: > On 04/01/2021 18:31, Dan Carpenter wrote: > > On Mon, Jan 04, 2021 at 12:09:27PM +0000, Phil Elwell wrote: > > > The addition of the local 'userdata' pointer to > > > vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor > > > WAITING modes are used, in which case the value provided by the > > > caller is replaced with a NULL. > > > > > > Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") > > > > > > Signed-off-by: Phil Elwell <phil@raspberrypi.com> > > > --- > > > drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- > > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > > > diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > index f500a7043805..2a8883673ba1 100644 > > > --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > struct vchiq_service *service; > > > struct bulk_waiter_node *waiter = NULL; > > > bool found = false; > > > - void *userdata = NULL; > > > + void *userdata; > > > int status = 0; > > > int ret; > > > @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > "found bulk_waiter %pK for pid %d", waiter, > > > current->pid); > > > userdata = &waiter->bulk_waiter; > > > + } else { > > > + userdata = args->userdata; > > > > "args->userdata" is marked as a user pointer so we really don't want to > > mix user and kernel pointers here. Presumably this opens up a large > > security hole. > > It's an opaque, pointer-sized token that only exists to bereturned to userspace (or not, > without this patch) - it's hard to see that as a security hole. I was assuming the bug here was a NULL dereference... Apparently that's not the case? The commit message needs to be updated to be more clear about how the bug looks like to the user. Are we using the "&waiter->bulk_waiter" as a "token to be returned to userspace" as well? It looks like maybe it is in vchiq_put_completion(). That defeats KASLR and is a different sort of security problem. Mixing __user pointers and regular pointers is dangerous and has lead to security problems in this driver in the past. But also mixing mixing tokens with pointers just makes the code hard to read. Instead of undoing Arnd's work where he split the user space and kernel pointers apart we should go ahead and spit it up even more. At least add a giant FIXME comment and an item in the TODO list so we don't forget to do this before removing the code from staging. regards, dan carpenter
On Tue, 5 Jan 2021 at 11:04, Dan Carpenter <dan.carpenter@oracle.com> wrote: > > On Mon, Jan 04, 2021 at 07:26:42PM +0000, Phil Elwell wrote: > > On 04/01/2021 18:31, Dan Carpenter wrote: > > > On Mon, Jan 04, 2021 at 12:09:27PM +0000, Phil Elwell wrote: > > > > The addition of the local 'userdata' pointer to > > > > vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor > > > > WAITING modes are used, in which case the value provided by the > > > > caller is replaced with a NULL. > > > > > > > > Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") > > > > > > > > Signed-off-by: Phil Elwell <phil@raspberrypi.com> > > > > --- > > > > drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- > > > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > index f500a7043805..2a8883673ba1 100644 > > > > --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > > struct vchiq_service *service; > > > > struct bulk_waiter_node *waiter = NULL; > > > > bool found = false; > > > > - void *userdata = NULL; > > > > + void *userdata; > > > > int status = 0; > > > > int ret; > > > > @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > > "found bulk_waiter %pK for pid %d", waiter, > > > > current->pid); > > > > userdata = &waiter->bulk_waiter; > > > > + } else { > > > > + userdata = args->userdata; > > > > > > "args->userdata" is marked as a user pointer so we really don't want to > > > mix user and kernel pointers here. Presumably this opens up a large > > > security hole. > > > > It's an opaque, pointer-sized token that only exists to bereturned to userspace (or not, > > without this patch) - it's hard to see that as a security hole. > > I was assuming the bug here was a NULL dereference... Apparently that's > not the case? The commit message needs to be updated to be more clear > about how the bug looks like to the user. > > Are we using the "&waiter->bulk_waiter" as a "token to be returned to > userspace" as well? It looks like maybe it is in vchiq_put_completion(). > That defeats KASLR and is a different sort of security problem. > > Mixing __user pointers and regular pointers is dangerous and has lead to > security problems in this driver in the past. But also mixing mixing > tokens with pointers just makes the code hard to read. Instead of > undoing Arnd's work where he split the user space and kernel pointers > apart we should go ahead and spit it up even more. At least add a giant > FIXME comment and an item in the TODO list so we don't forget to do this > before removing the code from staging. Those all sound like valid comments to have made against the original patch, but that seems to have received little attention. I'll just leave this here - perhaps Arnd has the patience to finish the job. Phil
On Tue, Jan 05, 2021 at 11:53:32AM +0000, Phil Elwell wrote: > On Tue, 5 Jan 2021 at 11:04, Dan Carpenter <dan.carpenter@oracle.com> wrote: > > > > On Mon, Jan 04, 2021 at 07:26:42PM +0000, Phil Elwell wrote: > > > On 04/01/2021 18:31, Dan Carpenter wrote: > > > > On Mon, Jan 04, 2021 at 12:09:27PM +0000, Phil Elwell wrote: > > > > > The addition of the local 'userdata' pointer to > > > > > vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor > > > > > WAITING modes are used, in which case the value provided by the > > > > > caller is replaced with a NULL. > > > > > > > > > > Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") > > > > > > > > > > Signed-off-by: Phil Elwell <phil@raspberrypi.com> > > > > > --- > > > > > drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- > > > > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > > > > > > > diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > > index f500a7043805..2a8883673ba1 100644 > > > > > --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > > +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c > > > > > @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > > > struct vchiq_service *service; > > > > > struct bulk_waiter_node *waiter = NULL; > > > > > bool found = false; > > > > > - void *userdata = NULL; > > > > > + void *userdata; > > > > > int status = 0; > > > > > int ret; > > > > > @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, > > > > > "found bulk_waiter %pK for pid %d", waiter, > > > > > current->pid); > > > > > userdata = &waiter->bulk_waiter; > > > > > + } else { > > > > > + userdata = args->userdata; > > > > > > > > "args->userdata" is marked as a user pointer so we really don't want to > > > > mix user and kernel pointers here. Presumably this opens up a large > > > > security hole. > > > > > > It's an opaque, pointer-sized token that only exists to bereturned to userspace (or not, > > > without this patch) - it's hard to see that as a security hole. > > > > I was assuming the bug here was a NULL dereference... Apparently that's > > not the case? The commit message needs to be updated to be more clear > > about how the bug looks like to the user. > > > > Are we using the "&waiter->bulk_waiter" as a "token to be returned to > > userspace" as well? It looks like maybe it is in vchiq_put_completion(). > > That defeats KASLR and is a different sort of security problem. > > > > Mixing __user pointers and regular pointers is dangerous and has lead to > > security problems in this driver in the past. But also mixing mixing > > tokens with pointers just makes the code hard to read. Instead of > > undoing Arnd's work where he split the user space and kernel pointers > > apart we should go ahead and spit it up even more. At least add a giant > > FIXME comment and an item in the TODO list so we don't forget to do this > > before removing the code from staging. > > Those all sound like valid comments to have made against the original > patch, but that seems to have received little attention. > > I'll just leave this here - perhaps Arnd has the patience to finish the job. I kind of have a headache today so maybe I shouldn't be sending emails. But really, all I'm asking is for is two fairly reasonable things: 1) The commit message needs to say what the bug looks like to the user. Up to now, I still have no idea the answer to this question. 2) Put a note in the TODO which says: "Clean up Sparse warnings from __user annotations. See vchiq_irq_queue_bulk_tx_rx(). Ensure that the the address of "&waiter->bulk_waiter" is never disclosed to userspace." regards, dan carpenter
On Tue, Jan 5, 2021 at 12:53 PM Phil Elwell <phil@raspberrypi.com> wrote: > On Tue, 5 Jan 2021 at 11:04, Dan Carpenter <dan.carpenter@oracle.com> wrote: > > > > Mixing __user pointers and regular pointers is dangerous and has lead to > > security problems in this driver in the past. But also mixing mixing > > tokens with pointers just makes the code hard to read. Instead of > > undoing Arnd's work where he split the user space and kernel pointers > > apart we should go ahead and spit it up even more. At least add a giant > > FIXME comment and an item in the TODO list so we don't forget to do this > > before removing the code from staging. > > Those all sound like valid comments to have made against the original > patch, but that seems to have received little attention. > > I'll just leave this here - perhaps Arnd has the patience to finish the job. I don't really have an interest in this driver. I did a larger cleanup in order to kill off copy_in_user() from the kernel, and then cleaned it up some more for good measure, but I would hope someone else can finish the address space mismatch. Arnd
diff --git a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c index f500a7043805..2a8883673ba1 100644 --- a/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c +++ b/drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c @@ -958,7 +958,7 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, struct vchiq_service *service; struct bulk_waiter_node *waiter = NULL; bool found = false; - void *userdata = NULL; + void *userdata; int status = 0; int ret; @@ -997,6 +997,8 @@ static int vchiq_irq_queue_bulk_tx_rx(struct vchiq_instance *instance, "found bulk_waiter %pK for pid %d", waiter, current->pid); userdata = &waiter->bulk_waiter; + } else { + userdata = args->userdata; } /*
The addition of the local 'userdata' pointer to vchiq_irq_queue_bulk_tx_rx omitted the case where neither BLOCKING nor WAITING modes are used, in which case the value provided by the caller is replaced with a NULL. Fixes: 4184da4f316a ("staging: vchiq: fix __user annotations") Signed-off-by: Phil Elwell <phil@raspberrypi.com> --- drivers/staging/vc04_services/interface/vchiq_arm/vchiq_arm.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)