[lldb][AArch64] Do not crash if NT_ARM_TLS is missing

[igorkudrin]

D156118 states that this note is always present, but it is better to check it explicitly, as otherwise lldb may crash when trying to read registers.

[llvmbot]

@llvm/pr-subscribers-lldb

Author: Igor Kudrin (igorkudrin)

Changes

D156118 states that this note is always present, but it is better to check it explicitly, as otherwise lldb may crash when trying to read registers.

---

Full diff: https://github.com/llvm/llvm-project/pull/106478.diff

3 Files Affected:

• (modified) lldb/source/Plugins/Process/Utility/RegisterInfoPOSIX_arm64.cpp (+4-3)
• (added) lldb/test/Shell/Process/elf-core/aarch64-no-NT_ARM_TLS.yaml (+34)
• (added) lldb/test/Shell/Process/elf-core/lit.local.cfg (+1)

diff --git a/lldb/source/Plugins/Process/Utility/RegisterInfoPOSIX_arm64.cpp b/lldb/source/Plugins/Process/Utility/RegisterInfoPOSIX_arm64.cpp
index 054b7d9b2ec575..8ebdd214458c45 100644
--- a/lldb/source/Plugins/Process/Utility/RegisterInfoPOSIX_arm64.cpp
+++ b/lldb/source/Plugins/Process/Utility/RegisterInfoPOSIX_arm64.cpp
@@ -254,9 +254,10 @@ RegisterInfoPOSIX_arm64::RegisterInfoPOSIX_arm64(
       if (m_opt_regsets.AllSet(eRegsetMaskMTE))
         AddRegSetMTE();
 
-      // The TLS set always contains tpidr but only has tpidr2 when SME is
-      // present.
-      AddRegSetTLS(m_opt_regsets.AllSet(eRegsetMaskSSVE));
+      if (m_opt_regsets.AllSet(eRegsetMaskTLS))
+        // The TLS set always contains tpidr but only has tpidr2 when SME is
+        // present.
+        AddRegSetTLS(m_opt_regsets.AllSet(eRegsetMaskSSVE));
 
       if (m_opt_regsets.AnySet(eRegsetMaskSSVE))
         AddRegSetSME(m_opt_regsets.AnySet(eRegsetMaskZT));
diff --git a/lldb/test/Shell/Process/elf-core/aarch64-no-NT_ARM_TLS.yaml b/lldb/test/Shell/Process/elf-core/aarch64-no-NT_ARM_TLS.yaml
new file mode 100644
index 00000000000000..bd8ce225c9e071
--- /dev/null
+++ b/lldb/test/Shell/Process/elf-core/aarch64-no-NT_ARM_TLS.yaml
@@ -0,0 +1,34 @@
+# REQUIRES: aarch64
+
+## Check that lldb does not crash if a core file does not contain an NT_ARM_TLS
+## note while there are notes for other dynamic register sets.
+
+# RUN: yaml2obj %s -o %t
+# RUN: %lldb -c %t -o "re r -a" | FileCheck %s
+
+# CHECK:      Pointer Authentication Registers:
+# CHECK-NEXT:   data_mask =
+# CHECK-NEXT:   code_mask =
+# CHECK-NOT:  Thread Local Storage Registers:
+
+--- !ELF
+FileHeader:
+  Class:        ELFCLASS64
+  Data:         ELFDATA2LSB
+  Type:         ET_CORE
+  Machine:      EM_AARCH64
+ProgramHeaders:
+  - Type:       PT_NOTE
+    FirstSec:   .note
+    LastSec:    .note
+Sections:
+  - Name:       .note
+    Type:       SHT_NOTE
+    Notes:
+      - Name:   CORE
+        Desc:   0b00000000000000000000000b00000000000000000000000000000000000000389300001b930000389300001b930000000000000000000000000000000000000000000000000000e02e0000000000000000000000000000119100000000000000000000000000005ee100000000000000000000000000002f00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000400162e2ffff00005801400000000000200162e2ffff0000240140000000000000000000000000000100000000000000
+        Type:   NT_PRSTATUS
+      - Name:   LINUX
+        Desc:   0000000000007f000000000000007f00
+        Type:   NT_ARM_PAC_MASK
+...
diff --git a/lldb/test/Shell/Process/elf-core/lit.local.cfg b/lldb/test/Shell/Process/elf-core/lit.local.cfg
new file mode 100644
index 00000000000000..8169b9f95e118c
--- /dev/null
+++ b/lldb/test/Shell/Process/elf-core/lit.local.cfg
@@ -0,0 +1 @@
+config.suffixes = ['.yaml']

[DavidSpickett]

Interested in how you ended up with a file without it, perhaps there was some era of kernels that produces that result.

But given we do a check whether to set the regset bit, we should use the result of that later anyway, so this looks ok to me.

[DavidSpickett]

And "don't crash on invalid input" is a good enough reason to fix this regardless of how the core file was generated, even if the core file is corrupted.

[DavidSpickett]

LGTM, thanks!