1 files changed, 34 insertions, 0 deletions
diff --git a/results/classifier/108/permissions/1421 b/results/classifier/108/permissions/1421
new file mode 100644
index 00000000..eabbbdd7
--- /dev/null
+++ b/results/classifier/108/permissions/1421
@@ -0,0 +1,34 @@
+permissions: 0.976
+graphic: 0.843
+device: 0.804
+debug: 0.803
+performance: 0.739
+semantic: 0.534
+vnc: 0.443
+PID: 0.401
+other: 0.385
+socket: 0.372
+boot: 0.317
+network: 0.296
+files: 0.161
+KVM: 0.028
+
+GDB memory reads fail on Cortex-M33
+Description of problem:
+GDB fails to read memory from the guest.  There appear to be at least two problems:
+
+1. In `arm_cpu_get_phys_page_attrs_debug`, `arm_is_secure(env)` returns false, because the implementation doesn't seem to know about Armv7-M or Armv8-M secure states.  However, `arm_mmu_idx(env)` does know how to check `env->v7m.secure`, so it returns `ARMMMUIdx_MSPriv` (the S stands for secure).  The mismatch between an apparently non-secure access to a secure MMU seems to cause the read to fail laster.
+2. With the MPU enabled (not the case in this repro, but I can provide one), `cpu_memory_rw_debug` computes `page = addr & TARGET_PAGE_MASK`, and uses the page to compute permissions.  However, TARGET_PAGE_MASK is based on 4K pages on this platform, but the MPU granularity is 32 bytes.  So the wrong page is used for checking.
+Steps to reproduce:
+```
+# Sorry for the large clone.  It's mostly unused files in CMSIS.
+git clone --recursive -b qemu-repro-1 https://github.com/dreiss/mpu_experiments
+cd mpu_experiments
+git checkout origin/qemu-repro-1
+cmake -S . -B build -DBOARD=qemu-mps2-an505 -DAPP=mpu_stacktrace -DCMAKE_BUILD_TYPE=Debug
+cmake --build build
+/path/to/qemu-system-arm -machine mps2-an505 -nographic -kernel build/kernel.elf -s -S -d int
+# Open a separate terminal and cd into mpu_experiments
+gdb build/kernel.elf -ex 'target remote :1234' -ex 'break base_case' -ex continue -ex backtrace -ex quit
+# Note the memory read failures in the backtrace.
+```