On Thu, Aug 29, 2019 at 05:17:03PM -0700, Sean Christopherson wrote: > Except for ENCLAVE_SET_ATTRIBUTE, all SGX ioctls() must be executed > serially to successfully initialize an enclave, e.g. the kernel already > strictly requires ECREATE->EADD->EINIT, and concurrent EADDs will result > in an unstable MRENCLAVE. Explicitly enforce serialization by returning > EINVAL if userspace attempts an ioctl while a different ioctl for the > same enclave is in progress. > > The primary beneficiary of explicit serialization is sgx_encl_grow() as > it no longer has to deal with dropping and reacquiring encl->lock when > a new VA page needs to be allocated. Eliminating the lock juggling in > sgx_encl_grow() paves the way for fixing a lock ordering bug in > ENCLAVE_ADD_PAGE without having to also juggle mm->mmap_sem. > > Serializing ioctls also fixes a race between ENCLAVE_CREATE and > ENCLAVE_SET_ATTRIBUTE, as the latter does not take encl->lock, e.g. > concurrent updates to allowed_attributes could result in a stale > value. The race could also be fixed by taking encl->lock or making > allowed_attributes atomic, but both add unnecessary overhead with this > change applied. > > Note, manipulating the list of VA pages (encl->va_pages) during EADD > must be done while holding encl->lock, the list of VA pages is visible > to the reclaimer after the first EADD completes. > > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx> Took a few rounds but I think it is now right :-) Was tricky to squash between three patches but I think I got that one right too. Can you check driver, provisioning and reclaimer patches and see if you see anything weird? /Jarkko
