[ELF] Don't crash when parsing a file with external version definition references

[arichardson]

Summary

We were crashing when linking telnetd in FreeBSD because lld was emitting
corrupted output files for --norosegment. In this file the version index of some symbols
was set to 9 but lld only found 8 version definitions.

I am not sure how to create a minimal .so file that also exposes this behaviour so I just added the one that initially caused the error to Inputs/

This partially addresses https://bugs.llvm.org/show_bug.cgi?id=34705

[ruiu]

Even with this patch, you cannot link telnetd in FreeBSD because this patch is to just prevent lld from crashing, right? I wonder why you don't add a feature so that you can link telnetd.

[arichardson]

I was able to link telnetd with these changes, possibly it doesn't use any of the symbols with the external version reference.

I don't really know anything about ELF versions so I thought as a first step not crashing would be useful. Also I don't think LLD should crash when reading an invalid input file but rather reject it.

[arichardson]

The issue that caused the corrupted input file appears to be fixed by https://reviews.llvm.org/D38579 and https://reviews.llvm.org/D38580. However we should still never access out-of-bounds memory, so turned this into a hard error

[arichardson]

fix typos in test case

[ruiu]

LGTM

ELF/InputFiles.cpp

782–784 ↗

(On Diff #117825)

Since this is presumably a rare error (we wouldn't have noticed it without a bug of lld itself), I'd make it slightly shorter.

error: corrupt input file: version definition index 10 for symbol bar is out of range
>>> defined in foo.so

lld/trunk/ELF/InputFiles.cpp

for (const Elf_Sym &Sym : Syms) {
     if (Sym.isUndefined()) {
       Undefs.push_back(Name);
       continue;
     }
 
     // Ignore local symbols.
     if (Versym && VersymIndex == VER_NDX_LOCAL)
       continue;
-
-    const Elf_Verdef *V =
-        VersymIndex == VER_NDX_GLOBAL ? nullptr : Verdefs[VersymIndex];
+    const Elf_Verdef *V = nullptr;
+    if (VersymIndex != VER_NDX_GLOBAL) {
+      if (VersymIndex >= Verdefs.size()) {
+        error("corrupt input file: version definition index " +
+              Twine(VersymIndex) + " for symbol " + Name +
+              " is greater than the maximum value " +
+              Twine(Verdefs.size() - 1) + "\n>>> symbol is defined in " +
+              toString(this));
+        continue;
+      }
+      V = Verdefs[VersymIndex];
+    }
 
     if (!Hidden)
       Symtab->addShared(Name, this, Sym, V);
 
     // Also add the symbol with the versioned name to handle undefined symbols
     // with explicit versions.
     if (V) {
       StringRef VerName = this->StringTable.data() + V->getAux()->vda_name;

lld/trunk/test/ELF/corrupted-version-reference.s

+# RUN: llvm-mc -triple=mips64-unknown-freebsd %s -filetype=obj -o %t.o
+# RUN: not ld.lld %t.o %S/Inputs/corrupt-version-reference.so -o %t.exe 2>&1 | FileCheck %s
+
+# CHECK: error: corrupt input file: version definition index 9 for symbol __cxa_finalize is greater than the maximum value 8
+# CHECK: >>> symbol is defined in {{.+}}/corrupt-version-reference.so
+
+.globl __start
+__start:
+    dla $a0, __cxa_finalize
+    nop