[Runtime] Change MetadataLookup section vectors to use ConcurrentReadableArray.

[mikeash]

This fixes a deadlock that can occur when dyld loads new images into the process. MetadataLookup.cpp takes a lock to iterate over sections and while the lock is held it can call code which eventually calls _dyld_register_func_for_add_image which takes a lock inside dyld. Elsewhere, a function registered with _dyld_register_func_for_add_image gets called with that same dyld lock held, and tries to acquire the same lock in MetadataLookup.cpp. This inconsistent lock ordering results in a deadlock when the stars align just so.

The fix here is to switch to ConcurrentReadableArray which does not take a lock for reading. It still takes a lock for writing, but since that's only one half of the deadlock, that's OK.

Relevant stack trace excerpts for the curious:

1000  void (anonymous namespace)::addImageCallback<&((anonymous namespace)::TypeFieldRecordSection), &(swift::addImageTypeFieldDescriptorBlockCallback(void const*, unsigned long))>(mach_header const*, long) + 48 (libswiftCore.dylib + 3658928) [0x110ff34b0]
  1000  swift::addImageTypeFieldDescriptorBlockCallback(void const*, unsigned long) + 38 (libswiftCore.dylib + 3720902) [0x1110026c6]
    1000  swift::MutexPlatformHelper::lock(_opaque_pthread_mutex_t&) + 11 (libswiftCore.dylib + 3738827) [0x111006ccb]

1000  swift_getFieldAt + 685 (libswiftCore.dylib + 3723501) [0x1110030ed]
  1000  swift_getFieldAt::$_3::operator()(swift::reflection::FieldDescriptor const&) const + 237 (libswiftCore.dylib + 3724429) [0x11100348d]
    1000  swift_getFieldAt::$_2::operator()(swift::reflection::FieldDescriptor const&) const + 411 (libswiftCore.dylib + 3724075) [0x11100332b]
      1000  swift::_getTypeByMangledName(llvm::StringRef, llvm::function_ref<swift::TargetMetadata<swift::InProcess> const* (unsigned int, unsigned int)>) + 659 (libswiftCore.dylib + 3722195) [0x111002bd3]
        1000  swift::Demangle::TypeDecoder<(anonymous namespace)::DecodedMetadataBuilder>::decodeMangledType(swift::Demangle::Node* const&) + 112 (libswiftCore.dylib + 3726800) [0x111003dd0]
          1000  swift::Demangle::TypeDecoder<(anonymous namespace)::DecodedMetadataBuilder>::decodeMangledType(swift::Demangle::Node* const&) + 207 (libswiftCore.dylib + 3726895) [0x111003e2f]
            1000  swift::Demangle::TypeDecoder<(anonymous namespace)::DecodedMetadataBuilder>::decodeMangledNominalType(swift::Demangle::Node* const&, (anonymous namespace)::DecodedMetadataBuilder::BuiltNominalTypeDecl&, swift::TargetMetadata<swift::InProcess> const*&) + 1873 (libswiftCore.dylib + 3732193) [0x1110052e1]
              1000  dispatch_once_f + 41 (libdispatch.dylib + 5419) [0x7fff660e552b]
                1000  _dispatch_client_callout + 8 (libdispatch.dylib + 5490) [0x7fff660e5572]
                  1000  _dyld_register_func_for_add_image + 44 (libdyld.dylib + 19614) [0x7fff66123c9e]

rdar://problem/40230581

[jckarter]

Awesome, thanks for catching and fixing this!

[xedin]

Thank you! For some reason I didn't even stop to think that Mutex might not be reentrant but instead just blindly used the same pattern found in other places in this file while working on swift_getFieldAt :(

[mikeash]

@xedin I don't think it would have occurred to me that _dyld_register_func_for_add_image would take a lock that would also be taken when invoking the callbacks, had I not seen it directly in the stack trace, so I certainly can't blame you for missing this.

[mikeash]

@swift-ci please test

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - b3d1030

[mikeash]

@swift-ci Please test OS X platform

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - b3d1030

[mikeash]

@swift-ci please test

[swift-ci]

Build failed
Swift Test OS X Platform
Git Sha - b3d1030

[swift-ci]

Build failed
Swift Test Linux Platform
Git Sha - b3d1030