On 03/07/2016 12:54 PM, Andy Lutomirski wrote:
On Mon, Mar 7, 2016 at 11:44 AM, Khalid Aziz <khalid.aziz@xxxxxxxxxx> wrote:
Consider this scenario:
1. Process A creates a shm and attaches to it.
2. Process A fills shm with data it wants to share with only known
processes. It enables ADI and sets tags on the shm.
3. Hacker triggers something like stack overflow on process A, exec's a new
rogue binary and manages to attach to this shm. MMU knows tags were set on
the virtual address mapping to the physical pages hosting the shm. If MMU
does not require the rogue process to set the exact same tags on its mapping
of the same shm, rogue process has defeated the ADI protection easily.
Does this make sense?
This makes sense, but I still think the design is poor. If the hacker
gets code execution, then they can trivially brute force the ADI bits.
True, with only 16 possible tag values (actually only 14 since 0 and 15
are reserved values), it is entirely possible to brute force the ADI
tag. ADI is just another tool one can use to mitigate attacks. A process
that accesses an ADI enabled memory with invalid tag gets a SIGBUS and
is terminated. This can trigger alerts on the system and system policies
could block the next attack. If a daemon is compromised and is forced to
hand out data from memory it should not be reading (similar to
heartbleed bug). the daemon itself is terminated with SIGBUS which
should be enough to alert system admins. A rotating set of tags would
reduce the risk from brute force attacks. Tags are set on cacheline
(which is 64 bytes on M7). A single regular sized page can have 128 sets
of tags. Allowing for 14 possible values for each set, that is a lot of
possible combinations of tags making it very hard to brute force tags
for more than a cacheline at a time. There are probably other better
ways to make the tags harder to crack.
Also, if this is the use case in mind, shouldn't the ADI bits bet set
on the file, not the mapping? E.g. have an ioctl on the shmfs file
that sets its ADI bits?
Shared data may not always be backed by a file. My understanding is one
of the use cases is for in-memory databases. This shared space could
also be used to hand off transactions in flight to other processes.
These transactions in flight would not be backed by a file. Some of
these use cases might not use shmfs even. Setting ADI bits at virtual
address level catches all these cases since what backs the tagged
virtual address can be anything - a mapped file, mmio space, just plain
chunk of memory.
A process can not just write version tags and make the file inaccessible to
others. It takes three steps to enable ADI:
1. Set PSTATE.mcde for the process.
2. Set TTE.mcd on all PTEs for the virtual addresses ADI is being enabled
on.
3. Set version tags.
Unless all three steps are taken, tag checking will not be done. stxa will
fail unless step 2 is completed. In your example, the step of setting
TTE.mcd will force sharing to stop for the process through
change_protection(), right?
OK, that makes some sense.
Can a shared page ever have TTE.mcd set? How does one share a page,
even deliberately, between two processes with cmd set?
For two processes to share a page, their VMAs have to be identical as I
understand it. If one process has TTE.mcd set (which means vma->vm_flags
is different) while the other does not, they do not share a page.
Thanks,
Khalid
--
To unsubscribe from this list: send the line "unsubscribe linux-arch" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
