EDMM user space support for kernel interfaces #751
Conversation
Signed-off-by: Zhang Lili <lili.z.zhang@intel.com>
Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com>
This avoids a 3-way deadlock for later EMM implementation. The utility thread is inside enclave setting up TCS, waiting on mm_lock to commit SSA, etc. Another thread #PF inside sgx_mm functions with mm_lock held, trigger vDSO user handler which tries to get back into enclave via Enclave::ecall, which calls acquire_thread trying to get the m_thread_mutex of CThreadPool A newly created pthread making ecall also tries to acquire_thread holding the CThreadPool::m_thread_mutex will wait for the utility thread to finish allocating TCS. For user handler already has the TCS for current thread, directly use it to enter enclave. Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com>
Add a runtime agnostic enclave memory manager implementation in sdk/emm. The EMM APIs are based on design proposed in this PR: openenclave/openenclave#3991 Enclave common loader changes are in psw/enclave_common/sgx_mm_ocalls.cpp, which are basically OCall support needed for EMM. Intel SDK runtime specific support are implemented as runtime abstraction layer: sdk/trts/ema_rt.c Current limitations, steps to build and test are documented in sdk/emm/README.md Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com> Signed-off-by: Xiaofeng Xu <20158212+xxu36@users.noreply.github.com>
|
Another concern is about the dealloc function, in this case: |
|
per my understanding, the ema node should only exist during the enclave's lifetime, i.e., created/intiliazed on enclave initialization, and released on enclave_destroy, check the code, is there anywhere "destory_ema_root" invoked when enclave is destoryed? |
It's similar to how mmap works, if user does not track and remove, EMAs(like VMAs) will be kept until enclave dies (like process dies). |
Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com>
The ioctl numbers for EDMM functions are shifted in upstream patches: https://patchwork.kernel.org/project/intel-sgx/cover/cover.1638381245.git.reinette.chatre@intel.com/ Change those definitions accordingly in isgx_user.h Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com>
Signed-off-by: Haitao Huang <4699115+haitaohuang@users.noreply.github.com>
|
Kernel patch updated to v2: https://github.com/rchatre/linux/tree/sgx/sgx2_submitted_v2_plus_rwx |
This implementation includes:
Please refer to sdk/emm/README.md for a quick guide on build and test as well as a summary of current limitations: