Re: [PATCH] habanalabs: use correct variable to show fd open counter

From: Oded Gabbay
Date: Mon Jul 08 2019 - 10:28:50 EST


On Mon, Jul 8, 2019 at 3:20 PM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, Jul 08, 2019 at 02:51:33PM +0300, Oded Gabbay wrote:
> > On Mon, Jul 8, 2019 at 2:43 PM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Jul 08, 2019 at 02:30:13PM +0300, Oded Gabbay wrote:
> > > > On Mon, Jul 8, 2019 at 2:21 PM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, Jul 08, 2019 at 01:43:55PM +0300, Oded Gabbay wrote:
> > > > > > The current code checks if the user context pointer is NULL or not to
> > > > > > display the number of open file descriptors of a device. However, that
> > > > > > variable (user_ctx) will eventually go away as the driver will support
> > > > > > multiple processes. Instead, the driver can use the atomic counter of
> > > > > > the open file descriptors which the driver already maintains.
> > > > > >
> > > > > > Signed-off-by: Oded Gabbay <oded.gabbay@xxxxxxxxx>
> > > > > > ---
> > > > > > drivers/misc/habanalabs/sysfs.c | 2 +-
> > > > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > >
> > > > > > diff --git a/drivers/misc/habanalabs/sysfs.c b/drivers/misc/habanalabs/sysfs.c
> > > > > > index 25eb46d29d88..881be19b5fad 100644
> > > > > > --- a/drivers/misc/habanalabs/sysfs.c
> > > > > > +++ b/drivers/misc/habanalabs/sysfs.c
> > > > > > @@ -356,7 +356,7 @@ static ssize_t write_open_cnt_show(struct device *dev,
> > > > > > {
> > > > > > struct hl_device *hdev = dev_get_drvdata(dev);
> > > > > >
> > > > > > - return sprintf(buf, "%d\n", hdev->user_ctx ? 1 : 0);
> > > > > > + return sprintf(buf, "%d\n", atomic_read(&hdev->fd_open_cnt));
> > > > > > }
> > > > >
> > > > > Odds are, this means nothing, as it doesn't get touched if the file
> > > > > descriptor is duped or sent to another process.
> > > > >
> > > > > Why do you care about the number of open files? Whenever someone tries
> > > > > to do this type of logic, it is almost always wrong, just let userspace
> > > > > do what it wants to do, and if wants to open something twice, then it
> > > > > gets to keep the pieces when things break.
> > > > >
> > > > > thanks,
> > > > >
> > > > > greg k-h
> > > >
> > > > I care about the number of open file descriptors because I can't let
> > > > multiple processes run simultaneously on my device, as we still don't
> > > > have the code to do proper isolation between the processes, in regard
> > > > of memory accesses on our device memory and by using our DMA engine.
> > > > Basically, it's a security hole. If you want, I can explain more in
> > > > length on this issue.
> > >
> > > But the issue is that you can't "force" this from the kernel side at
> > > all. Trying to catch this at open() time only catches the obvious
> > > processes.
> > >
> > > As I said, how do you check for:
> > > fd = open(...);
> > > fd_new = dup(fd);
> > >
> > > write(fd, ...);
> > > write(fd_new, ...);
> > >
> > > or, pass the fd across a socket? Or other fun ways of sending file
> > > descriptors around a system.
> > >
> > > You have to trust userspace here, sorry. If someone wants to do
> > > multiple accesses, they can, but again, they deserve the pieces when
> > > things fall apart.
> >
> > I see what you are saying, but from *security* perspective, I don't
> > think I really care about the scenarios above, right ?
> > Because that would mean a user duplicated his *own* fd. Sure, things
> > won't work for him, but what do I care about that, as you rightly
> > said.
> > I'm only concerned about the security risk, where there is a
> > legitimate user and a malicious one. Because I can't isolate between
> > them, I want to be able to allow only one of them to run.
>
> But how can you tell if the first one is the "malicious" one and the
> second one is "legitimate"? You do that by the "normal" file
> permissions and the like, you don't try to do a "first one wins!" type
> of policy, that's crazy :)

I don't care who is malicious and who is not. Of course I can't count
on "first one wins" to do that...
What I care about, is that two different processes won't be able to
send jobs to the device at the same time. That's it.
But I see your point about not using file-descriptors to enforce this
limitation.
I will change my code, but it will take a bit of time. It's not a
trivial change.

Thanks,
Oded

>
> > I don't care if one of them duplicates his own FD, right ?
>
> If you are trying to keep someone from having multiple FD per device,
> then yes, you should care. The point is, you can't know.
>
> > Please tell me if my assumption here is correct or not, because this
> > has implications.
>
> Don't rely on "first one wins!" as a security policy to prevent bad
> things from happening. That's never going to work, let userspace police
> these things, as that is the only place you can do it properly (or with
> a selinux/apparmor/lsm policy).
>
> > > > We have the H/W infrastructure for that, using MMU and multiple
> > > > address space IDs (ASID), but we didn't write the code yet in the
> > > > driver, as that is a BIG feature but it wasn't requested by anyone
> > > > yet.
> > > >
> > > > So the current solution is to block the ability to open multiple file
> > > > descriptors.
> > > >
> > > > Regarding this specific sysfs property, I don't really care about it.
> > > > I simply saw that it is shown in other drivers and I thought it may be
> > > > nice for a system admin utility to show it.
> > >
> > > What drivers show the number of open file descriptors? Time to go
> > > delete them as well :)
> > hehe
> > I tried to grep for it but I couldn't find any. Strange because I was
> > sure I saw this in some driver.
>
> If you run across it again, please let me know. It's a common
> "anti-pattern" that I have been guilty of in the past.
>
> thanks,
>
> greg k-h