[clang] [llvm] [llvm][clang] Sandbox filesystem reads (PR #165350)

[Jan Svoboda via cfe-commits]

================
@@ -347,12 +348,15 @@ llvm::Expected<FileEntryRef> FileManager::getSTDIN() {
   if (STDIN)
     return *STDIN;
 
-  std::unique_ptr<llvm::MemoryBuffer> Content;
-  if (auto ContentOrError = llvm::MemoryBuffer::getSTDIN())
-    Content = std::move(*ContentOrError);
-  else
+  auto ContentOrError = [] {
+    auto BypassSandbox = llvm::sys::sandbox::scopedDisable();
+    return llvm::MemoryBuffer::getSTDIN();
----------------
jansvoboda11 wrote:

I decided to make `MemoryBuffer::getSTDIN()` a sandbox violation because I'd argue that calling it from within a library is always the wrong thing to do. (Here we have `clangBasic` doing it, but multiple LLVM libraries try to do it too.) For now, I decided to bless this particular spot to be allowed to call it, but I think the ideal solution is that stdin gets read explicitly in the client executable (in `cc1_main` let's say) and the `MemoryBuffer` gets passed down to `FileManager` (or other libraries that want **the** input) explicitly. For CAS caching, we'd capture that buffer to make it part of the compilation key.

https://github.com/llvm/llvm-project/pull/165350