Speculative Data Oblivious Execution (SDO) is a safe mechanism for improving speed of delay execution schemes (such as Speculative Taint Tracking (STT) (MICRO'19)) by executing leaky long-latency operations (such as loads) speculatively in a data-oblivious manner. More details can be found in our ISCA'20 paper here. Here is a sample format for citing our work:

@inproceedings{yu2020sdo,
  title={Speculative Data-Oblivious Execution: Mobilizing Safe Prediction For Safe and Efficient Speculative Execution},
  author={Yu, Jiyong and Mantri, Namrata and Torrellas, Josep and Morrison, Adam and Fletcher, Christopher W},
  booktitle={Proceedings of the ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA)},
  pages={707---720},
  year={2020}
}

We implement SDO on top of Gem5 simulator, which is a cycle-accurate simulator with both system call emulation and full system modes. SDO is implemented on an early version of Gem5 (commit:38a1e23). SDO is fully implemented on Gem5's O3 processor and Ruby memory subsystem.

SDO relies on STT to identify leakage (via tainting/untainting) and explicit/implicit channels. The major changes of SDO are:

• implementing a variety of location predictors
• adding logic in load-store unit for issuing Obl-Ld requests for protecting unsafe loads
• adding logic in memory subsystem for implementing safe Obl-Ld operations at each cache level

We use X86 architecture and Ruby memory model. To build the Gem5 executable (such as gem5.opt):

scons build/X86_MESI_Three_Level/gem5.opt

We provide a sample script in './sample_script' to run different configurations, including Unsafe baseline, STT and SDO.

SDO supports different modes, and different predictor variants and configurations. Here we enumerate all important options:

• --scheme=[string]: different protection schemes

  • UnsafeBaseline: unmodified processor without any protection
  • DelayExecute: after identifying unsafe operations, delay the execution of those operations
  • SDO: after idenfitying unsafe operations, using SDO to execute those operations safely

• -- mem_model=[string]: memory consistency model

  • TSO: Total Store Ordering (TSO) model
  • RC: Released consistency (RC) model

• -- threat_model=[string]: attacker model

  • Spectre: Spectre threat model (covering control-flow speculation)
  • Futuristic: Futuristic threat model (covering all types of speculations, exceptions, interrupts)

• --STT=[int]: whether STT is enabled

  • 0: disable STT (all speculative transmitters are unsafe)
  • 1: enable STT (STT's taint tracking determines which transmtters are unsafe and requires protection)

• --impChannel=[int]: whether STT's implicit channel protection is enabled

  • 0: disable STT's implicit channel protection
  • 1: enabling STT's implicit channel protection

• --pred_type=[string]: location predictor type

  • static: static location predictor. Predict a constant level.
  • greedy: Greedy location predictor
  • hysteresis: Hysteresis location predictor
  • local: Local location predictor
  • loop: Loop location predictor
  • random: Random location predictor. Predict a random level.
  • perfect: Perfect location predictor. Predict the correct level.
  • tournament_2way: combining two predictors (e.g., greedy + hysteresis)
  • tournament_3way: combining three predictors (e.g., greedy + hysteresis + local)

• --pred_option=[number]: parameters for location predictor

  • (for static): 0 - always predicting L1; 1 - always predicting L2; 2 - always predicting L3 ; 3 - alwasy predicting DRAM
  • (for greedy, hysteresis, local, loop, perfect, tournament_2way/3way): 0 - can predict any level; 1 - don't predict DRAM; 2 - don't predict DRAM+LLC

• --subpred1_type=[string]: the first component of a tournament_2way/tournament_3way predictor

  • choose from any --pred_type options

• --subpred2_type=[string]: the second component of a tournament_2way/tournament_3way predictor

  • choose from any --pred_type options

• --subpred3_type=[string]: the third component of a tournament_3way predictor

  • choose from any --pred_type options

• --TLB_defense=[string]: TLB defense option

  • No: No TLB defense
  • SDO: SDO's TLB defense: on TLB misses, proceed with predicted page number
  • UnsafeDelay: on TLB misses, delay the memory access