As discovered by Jarkko, taking mm->mmap_sem around EADD can lead to
deadlock as attempting to acquire mmap_sem while holding encl->lock
violates SGX's lock ordering.
Resolving the issue simply by reversing the lock ordering gets ugly due
to the behavior of sgx_encl_grow(), which has a path that drops encl->lock
in order to do EPC page reclaim, i.e. moving mm->mmap_sem up would require
it to be dropped and reacquired as well.
Luckily, the entire flow can be simplified by preventing userspace from
invoking concurrent ioctls on a single enclave. Requiring ioctls to be
serialized means encl->lock doesn't need to be held to grow/shrink the
enclave, since only ioctls can grow/shrink the enclave. This also
eliminates theoretical cases that sgx_encl_grow() has to handle, e.g. the
enclave being initialized while it's waiting on reclaim, since the
protection provided by serializing ioctls isn't dropped to do reclaim.
v2:
- Make encl->flags an atomic, reuse for "in ioctl" detection. [Jarkko]
- Drop smp_{r,w}mb() patch as it is superceded by atomic flags. [Jarkko]
- Tack on two interdependent bug fixes found when auditing encl->flags.
- Rebase to Jarkko's latest master.
Sean Christopherson (5):
x86/sgx: Convert encl->flags from an unsigned int to an atomic
x86/sgx: Reject concurrent ioctls on single enclave
x86/sgx: Take encl->lock inside of mm->mmap_sem for EADD
x86/sgx: Reject all ioctls on dead enclaves
x86/sgx: Destroy the enclave if EEXTEND fails
arch/x86/kernel/cpu/sgx/driver.c | 3 +-
arch/x86/kernel/cpu/sgx/encl.c | 18 ++--
arch/x86/kernel/cpu/sgx/encl.h | 3 +-
arch/x86/kernel/cpu/sgx/ioctl.c | 163 ++++++++++++++++--------------
arch/x86/kernel/cpu/sgx/reclaim.c | 10 +-
5 files changed, 105 insertions(+), 92 deletions(-)
--
2.22.0
