diff options
Diffstat (limited to 'results/classifier/105/instruction/1587')
| -rw-r--r-- | results/classifier/105/instruction/1587 | 38 |
1 files changed, 38 insertions, 0 deletions
diff --git a/results/classifier/105/instruction/1587 b/results/classifier/105/instruction/1587 new file mode 100644 index 000000000..04ed2209a --- /dev/null +++ b/results/classifier/105/instruction/1587 @@ -0,0 +1,38 @@ +instruction: 0.949 +graphic: 0.888 +semantic: 0.794 +device: 0.759 +socket: 0.556 +network: 0.547 +vnc: 0.528 +mistranslation: 0.483 +boot: 0.357 +assembly: 0.348 +other: 0.345 +KVM: 0.207 + +Invalid memory access allowed (possibly due to TLB bypassing PMP after mret) +Description of problem: +A load instruction that should be blocked by PMP due to MPRV changing the effective privilege mode to U is allowed. The sequence that I observed was: + + +1. Be in machine mode. +2. Set MPP to U (0). +3. Set MPRV to 1. +4. Enter an ISR, setting MPP to M (3). +5. Load from address xxxx (populating the QEMU TLB). +6. Execute mret, setting MPP back to U (0). +7. Load from address xxxx, which should fail but succeeds without any TLB fill. +Steps to reproduce: +``` +git clone https://github.com/dreiss/qemu_pmp_repro +cd qemu_pmp_repro +./build_and_run.sh +``` +The `build_and_run.sh` script expects `riscv-none-elf-gcc` and `qemu-system-riscv64` on PATH. It will also attempt to run the reproducer with `spike`, the reference RISC-V emulator, which succeeds. +Additional information: +Adding a call to `tlb_flush` to `helper_mret` causes this test to pass in QEMU, but I don't know if that's a valid fix. + +Output from `build_and_run.sh`: + +[output.txt](/uploads/108547bcb160a8f0bfffe72ea77b215f/output.txt) |