Summary:
Here is an example for class inline member function declared and defined in the same file:

```
Definition:
0x00004c0a:           DW_TAG_subprogram [55] *
                        DW_AT_linkage_name [DW_FORM_strp] ( .debug_str[0x00177124] = "_ZNK5folly10symbolizer4test24ClassWithInlineFunctions9inlineBarERNS0_10FrameArrayILm100EEE")
                        DW_AT_name [DW_FORM_strp] ( .debug_str[0x00089a73] = "inlineBar")
                        DW_AT_decl_file [DW_FORM_data1] (0x1a)
                        DW_AT_decl_line [DW_FORM_data1] (0xa5)
                        DW_AT_declaration [DW_FORM_flag_present]  (true)
                        DW_AT_external [DW_FORM_flag_present] (true)
                        DW_AT_accessibility [DW_FORM_data1] (0x01)
…

Declaration:   -- no DW_AT_decl_file
0x000164c1:   DW_TAG_subprogram [228] *
                DW_AT_specification [DW_FORM_ref4]  (cu + 0x48ea => {0x00004c0a})
                DW_AT_inline [DW_FORM_data1]  (0x01)
                DW_AT_object_pointer [DW_FORM_ref4] (cu + 0x161ac => {0x000164cc})

…

Inline function call
0x0001657b:     DW_TAG_inlined_subroutine [140] *
                  DW_AT_abstract_origin [DW_FORM_ref4]  (cu + 0x161a1 => {0x000164c1})
                  DW_AT_low_pc [DW_FORM_addr] (0x000000000040fa6e)
                  DW_AT_high_pc [DW_FORM_data4] (0x00000055)
                  DW_AT_call_file [DW_FORM_data1] (0x1a)
                  DW_AT_call_line [DW_FORM_data1] (0xe2)
```

Here is an example for class inline member function declared and defined in different files:

```
Definition:  DW_AT_decl_file: 0x01  --> SymbolizerTestUtils.h
0x00004c3d:           DW_TAG_subprogram [55] *
                        DW_AT_linkage_name [DW_FORM_strp] ( .debug_str[0x0017717f] = "_ZNK5folly10symbolizer4test22InlineFunctionsWrapper9inlineBarERNS0_10FrameArrayILm100EEE")
                        DW_AT_name [DW_FORM_strp] ( .debug_str[0x00089a73] = "inlineBar")
                        DW_AT_decl_file [DW_FORM_data1] (0x01)
                        DW_AT_decl_line [DW_FORM_data1] (0x30)
                        DW_AT_declaration [DW_FORM_flag_present]  (true)
                        DW_AT_external [DW_FORM_flag_present] (true)
                        DW_AT_accessibility [DW_FORM_data1] (0x01)

…

Declaration:   DW_AT_decl_file: 0xbf  --> SymbolizerTestUtils-inl.h
0x00017ee8:   DW_TAG_subprogram [231] *
                DW_AT_decl_file [DW_FORM_data1] (0xbf)
                DW_AT_decl_line [DW_FORM_data1] (0x21)
                DW_AT_specification [DW_FORM_ref4]  (cu + 0x491d => {0x00004c3d})
                DW_AT_inline [DW_FORM_data1]  (0x01)
                DW_AT_object_pointer [DW_FORM_ref4] (cu + 0x17bd5 => {0x00017ef5})

…

Inline function call
0x00017fa5:     DW_TAG_inlined_subroutine [158] *
                  DW_AT_abstract_origin [DW_FORM_ref4]  (cu + 0x17bc8 => {0x00017ee8})
                  DW_AT_low_pc [DW_FORM_addr] (0x00000000004104ee)
                  DW_AT_high_pc [DW_FORM_data4] (0x00000055)
                  DW_AT_call_file [DW_FORM_data1] (0x1a)
                  DW_AT_call_line [DW_FORM_data2] (0x0101)
```

Reviewed By: luciang

Differential Revision: D19658560

fbshipit-source-id: 78c1e42d76c19f09e4632c40e0df18bfb49479f9

Summary: Linkage name (DW_AT_linkage_name) contains more information than the identifier name (DW_AT_name).

Reviewed By: luciang

Differential Revision: D19657767

fbshipit-source-id: 87b2293e6f723ceb8b5202db0f3c569c37c6c7ea

Summary:
Actually class inline functions exist in both .debug_info and .debug_types sections. In this diff, only the information in .debug_info is used.

Step 1. Find the DW_TAG_inlined_subroutine debug info entry (DIE) based on the given address.
Step 2. Find the DW_TAG_subprogram DIE based on the DW_AT_abstract_origin attr in inlined_subroutine, but this DIE is just a declaration.
Step 3. Find the actual definition DW_TAG_subprogram DIE based on the DW_AT_specification attr.

```
.debug_info

0x00003657:           DW_TAG_subprogram [56] *   ---- step 3
                        DW_AT_linkage_name [DW_FORM_strp] ( .debug_str[0x0002932d] = "_ZN5folly10symbolizer4test24ClassWithInlineFunctions9inlineBarERNS0_10FrameArrayILm100EEE")
                        DW_AT_name [DW_FORM_strp] ( .debug_str[0x00014064] = "inlineBar")
                        DW_AT_decl_file [DW_FORM_data1] (0x18)
                        DW_AT_decl_line [DW_FORM_data1] (0xa1)
                        DW_AT_declaration [DW_FORM_flag_present]  (true)
                        DW_AT_external [DW_FORM_flag_present] (true)
                        DW_AT_accessibility [DW_FORM_data1] (0x01)

…

0x0000cd74:   DW_TAG_subprogram [203] * .   ---- step 2
                DW_AT_specification [DW_FORM_ref4]  (cu + 0x358f => {0x00003657})
                DW_AT_inline [DW_FORM_data1]  (0x01)
                DW_AT_object_pointer [DW_FORM_ref4] (cu + 0xccb7 => {0x0000cd7f})

…
0x0000cda0:   DW_TAG_subprogram [181] *
                DW_AT_low_pc [DW_FORM_addr] (0x00000000002a4980)
                DW_AT_high_pc [DW_FORM_data4] (0x00001e22)
                DW_AT_frame_base [DW_FORM_exprloc]  (<0x1> 56 )
                DW_AT_object_pointer [DW_FORM_ref4] (cu + 0xccf0 => {0x0000cdb8})
                DW_AT_specification [DW_FORM_ref4]  (cu + 0x312f => {0x000031f7})

…

0x0000ce02:     DW_TAG_inlined_subroutine [157] *               ---- step 1
                  DW_AT_abstract_origin [DW_FORM_ref4]  (cu + 0xccac => {0x0000cd74})
                  DW_AT_low_pc [DW_FORM_addr] (0x00000000002a4f66)
                  DW_AT_high_pc [DW_FORM_data4] (0x000000c0)
                  DW_AT_call_file [DW_FORM_data1] (0x18)
                  DW_AT_call_line [DW_FORM_data1] (0xe6)

.debug_types

0x00020fdf:         DW_TAG_class_type [33] *
                      DW_AT_calling_convention [DW_FORM_data1]  (0x05)
                      DW_AT_name [DW_FORM_strp] ( .debug_str[0x00067094] = "ClassWithInlineFunctions")
                      DW_AT_byte_size [DW_FORM_data1] (0x08)
                      DW_AT_decl_file [DW_FORM_data1] (0x18)
                      DW_AT_decl_line [DW_FORM_data1] (0x9f)

…

0x00020ff5:           DW_TAG_subprogram [56] *
                        DW_AT_linkage_name [DW_FORM_strp] ( .debug_str[0x0002932d] = "_ZN5folly10symbolizer4test24ClassWithInlineFunctions9inlineBarERNS0_10FrameArrayILm100EEE")
                        DW_AT_name [DW_FORM_strp] ( .debug_str[0x00014064] = "inlineBar")
                        DW_AT_decl_file [DW_FORM_data1] (0x18)
                        DW_AT_decl_line [DW_FORM_data1] (0xa1)
                        DW_AT_declaration [DW_FORM_flag_present]  (true)
                        DW_AT_external [DW_FORM_flag_present] (true)
                        DW_AT_accessibility [DW_FORM_data1] (0x01)

```

Reviewed By: luciang

Differential Revision: D19644033

fbshipit-source-id: fa5d6dc46dbb1a417b40c1d06d9eaf9178661d21

Summary:
[Folly] Revise `SignalSafeElfCache` to use `mmap` for dynamic allocation and not to require static preallocation.

While, technically, `mmap` is not documented to be async-signal-safe, in practice it is so at least on Linux. Take advantage.

Prior to this change, the `SignalSafeElfCache` with all of its storage would have to be preallocated before setting the signal handler and must not be destroyed ever. Prior to this change, the preallocation would require at least `sizeof(Path) * capacity`, which defaults to ~2MB.

It is now possible to use stack-allocated SignalSafeElfCache in signal handlers.

This change adds a multi-thread-safe and async-signal-safe `mmap`-based allocator. Details as follows:
* All allocations are rounded up to the nearest power of two in size and alignment.
* Page or larger allocations are handled directly by `mmap` and free by `munmap` and are not tracked.
* Smaller allocations are handled by an `mmap`-backed refcounted arena. Arena sections are 16 pages in size; they are allocated on demand; and they are linked together to allow for `munmap` when the allocator refcounted arena is destroyed.

Reviewed By: davidtgoldblatt

Differential Revision: D18760282

fbshipit-source-id: 56a34abfe39a7108f790537afeda832fd39714d7

Summary: [Folly] Put the current fiber manager behind an accessor function v.s. having it as a static class member.

Reviewed By: andriigrynenko

Differential Revision: D19229507

fbshipit-source-id: 7cf4eab00ee281bc6104f994a71f5e48d9553852

Summary: MSVC informed me that these classes had noexcept move constructors but the move assignment operators were not marked as noexecpt.

Reviewed By: yfeldblum

Differential Revision: D19796949

fbshipit-source-id: d2bdc8ebb6d77709e367a4137fb0817c8006084f

Summary: This logic makes it impossible to reason about order between io events, internal and external events. The comment also seems outdated.

Reviewed By: yfeldblum

Differential Revision: D19785249

fbshipit-source-id: 62eee5427a3c47cccba6cb33c2ca88af82767bed

Summary:
The `Future(T2&&)` constructor is currently constrained to disallow other `Future` and `SemiFuture` types, but it allows types which will cause a hard error when instantiating the body of the constructor. The constructor argument is later used to initialize a `Try<T>`, so the function should properly be constrained with `std::is_constructible<Try<T>, T2>::value`.

Also, add the same constraint to the `SemiFuture(T2&&)` constructor, and to the `FutureBase(T2&&)` constructor for good measure.

Reviewed By: yfeldblum

Differential Revision: D19782313

fbshipit-source-id: 5e1ef5425b0c7ab9cc76c998425e90d0ba0b148f

Summary:
The current implementation does not guarantee atomicity on power
failure or OS crash.  For instance, if a crash happens after writing
the directory to storage but before writing the file data to storage,
then a zero-byte file can result.

Provide an optional parameter to sync and provide atomicity in these
situations.

Adjust comments to reflect the actual guarantees provided.

Differential Revision: D19754974

fbshipit-source-id: 5ddb2f3a1b9e72b523ee13937f648e0913c94c2f

Summary:
[Folly] Use atomic operations on maybe-TLS `SharedMutex` static data  via `atomic_ref`. The fields are thread-local on platforms which support thread-local fields, but global otherwise.

When thread-local is supported, relaxed loads and stores should ordinarily have the same overhead as non-atomic loads and stores on most platforms. When thread-local is not supported and these fields are global, the concurrent atomic loads and stores will change to become defined behavior.

Differential Revision: D19688867

fbshipit-source-id: 51873e12924ecff9d971316658fe71e2de18bdec

Summary:
Instead of fetching a tarball for the vendored Rust crates, let's just use
what's present in third-party/rust.

Reviewed By: wez, fanzeyi

Differential Revision: D19770580

fbshipit-source-id: ca6615a5f5c672d9a5527ad47feab6bdd33a39db

Summary: [Folly] Cut `StaticMetaBase` `[[noreturn]]` dtor, since MSVC C4722 flags it.

Reviewed By: filipecalasans

Differential Revision: D19738476

fbshipit-source-id: d697aad1194b76e7c371aa9a34ea50cb793b5c32

Summary: Mononoke is becoming part of Eden repository, remove it's getdeps manifest until we finalize the move.

Reviewed By: krallin

Differential Revision: D19769090

fbshipit-source-id: 9b471686728e9ff28317f9157f5f8a1834c5f5e4

Summary:
D19417574 added support for snapshotting stack traces for threads our requests are running on. The problem with this approach was that we are racing - stack trace may be recorded AFTER requests has already left the thread. Hence we might be presenting incorrect and confusing picture to the users.

Here we make sure that request root id and stack trace are recorded together, and there is no race.

Because we are using interrupts to get on the thread, sometimes we might not be able to record the information using this approach. We then fallback to the previous way of associating threads and requests (by walking TLS entries).

(Note: this ignores all push blocking failures!)

Reviewed By: andriigrynenko

Differential Revision: D19445100

fbshipit-source-id: 4311f4e521a48852bf99aede03814de0941c7814

Summary:
[Folly] Fix gcc-specific problem in `reentrant_allocator` test.

```
folly/memory/test/ReentrantAllocatorTest.cpp: In member function ‘virtual void ReentrantAllocatorTest_large_Test::TestBody()’:
folly/folly/memory/test/ReentrantAllocatorTest.cpp:68:22: error: no matching function for call to ‘std::vector<ReentrantAllocatorTest_large_Test::TestBody()::type, folly::reentrant_allocator<ReentrantAllocatorTest_large_Test::TestBody()::type> >::push_back(<brace-enclosed initializer list>)’
     vec.push_back(type{i});
                      ^
```

Differential Revision: D19738608

fbshipit-source-id: bbffd1f47749be1aa88bb45bddc401ffca48c4ca

Summary: Pull Request resolved: https://github.com/facebook/folly/pull/1298

Reviewed By: igorsugak

Differential Revision: D19281240

Pulled By: yfeldblum

fbshipit-source-id: 3bf437d49aa85b40923ff1ad93ba62fa97b726af

Summary: [Folly] Use `cacheline_aligned` in `folly/experimental/settings/`.

Reviewed By: aary

Differential Revision: D19657676

fbshipit-source-id: cf31152b6525df2838b5c79f846e1698af17777e

Summary:
The std::move prevents copy elision (NRVO), and gcc9 is warning about
it. Every compiler we use will do NRVO here, and in c++17 and later
the compiler is required to use the move constructor if it (for
whatever reason) fails to elide the copy.

Reviewed By: ot

Differential Revision: D19698783

fbshipit-source-id: 589556a0a88da0d84c607c9894399201dd7d6ae8

Summary:
Using FiberManagerMap is a prerequisite for the gdb extension, but
previously it didn't support using a non-void local type. One option is to
make GlobalCache<EventBaseT, LocalT>, I ended up sharing the cache across local
types to avoid changing much in the gdb side.

The gdb extension is already updated to handle the namespace change.

Reviewed By: andriigrynenko

Differential Revision: D19422428

fbshipit-source-id: 26cb385791a9503906fc69d91039bfa93837b6db

Summary:
Until C++17 types with extended alignment need special handling
in eg. std::vector and also std::unique_ptr. CPUThreadPoolExecutor has a
default queue that requires extended alignment, but also allows the user
to provide their own blocking queue. To handle this we move the private
member to a shared_ptr which supported type erased destructors, and then
use allocate_shared for the default queue type with a custom allocator
that will satisfy alignment constraints.

```
.../unique_ptr.h:825:34: runtime error: constructor call on misaligned address 0x613000000040 for type 'folly::UnboundedBlockingQueue<folly::CPUThreadPoolExecutor::CPUTask>', which requires 128 byte alignment
0x613000000040: note: pointer points here
 02 00 00 1c  be be be be be be be be  be be be be be be be be  be be be be be be be be  be be be be
              ^
    #0 0x75b35e in std::_MakeUniq<folly::UnboundedBlockingQueue<folly::CPUThreadPoolExecutor::CPUTask> >::__single_object std::make_unique<folly::UnboundedBlockingQueue<folly::CPUThreadPoolExecutor::CPUTask> >() .../unique_ptr.h:825:30
    #1 0x75602f in folly::CPUThreadPoolExecutor::CPUThreadPoolExecutor(unsigned long, std::shared_ptr<folly::ThreadFactory>) xplat/folly/executors/CPUThreadPoolExecutor.cpp:66:18
    #2 0x756c6c in folly::CPUThreadPoolExecutor::CPUThreadPoolExecutor(unsigned long) xplat/folly/executors/CPUThreadPoolExecutor.cpp:84:7
```

Differential Revision: D19669832

fbshipit-source-id: 7fb60a06dbb6a04edddb6b619af7a8cef4995fd2

Summary:
D19222635 added a `ReentrantAllocatorTest.cpp` test file but in the
CMakeLists.txt file it listed `MmapAllocatorTest.cpp` instead.

Reviewed By: yfeldblum

Differential Revision: D19650216

fbshipit-source-id: 794ced255de6781a05da7114eb7f8b3258f90a8c

Summary:
The context initialization test was testing undefined behavior by asserting that
OpenSSL will return 0 (and consequently, result in SSLContext throwing) after
a SSL_CTX is freed. That is to say, it is exercising control flow based on
memory that is already freed.

This diff alters the logic to detect an SSL_CTX free via EX_DATA.

Reviewed By: igorsugak

Differential Revision: D19306728

fbshipit-source-id: cc3a94dbb0b2ab614fa72b099c585467b7156613

Reviewed By: aary

Differential Revision: D19403185

fbshipit-source-id: 02ebbda5b9500413920d7225b6fbe29cf5d73237

Summary: In stack-constrained environments, primarily fibers, the context state allocated by getStackTrace can be problematic. getStackTraceHeap gives the opposite guarantee from getStackTraceSafe - there is no significant stack allocation and we explicitly heap allocate instead.

Reviewed By: terrelln

Differential Revision: D19575634

fbshipit-source-id: 2ec86eccec55d1044bd2532324c648c791df5a07

Summary: s/assiciated/associated

Reviewed By: ericniebler

Differential Revision: D19645433

fbshipit-source-id: 844f3f782a083030dbab5b8196c98fb562f0f7de

Summary: add common interface for loop controllers with publicly accessible executor

Reviewed By: A5he

Differential Revision: D19626108

fbshipit-source-id: 07884726d7fce273dd56bdcf454be70051a4f62b

Summary:
[Folly] Fix gcc-specific problem in `reentrant_allocator` test.

```
folly/memory/test/ReentrantAllocatorTest.cpp: In member function ‘virtual void ReentrantAllocatorTest_large_Test::TestBody()’:
folly/folly/memory/test/ReentrantAllocatorTest.cpp:68:22: error: no matching function for call to ‘std::vector<ReentrantAllocatorTest_large_Test::TestBody()::type, folly::reentrant_allocator<ReentrantAllocatorTest_large_Test::TestBody()::type> >::push_back(<brace-enclosed initializer list>)’
     vec.push_back({i});
                      ^
```

Reviewed By: simpkins

Differential Revision: D19652353

fbshipit-source-id: ca72a80002096e1d35c61a219f9af271fa79a4a7

Reviewed By: yfeldblum

Differential Revision: D19641278

fbshipit-source-id: 2631056c6852de3b1e8b972e8b20b26283ea1070

Summary:
Add a static method to Subprocess that allows creating a `Subprocess` object
from an existing child process ID.  This allows creating `Subprocess` objects
for children processes that were originally spawned through other APIs than
Subprocess itself.

This allows using the Subprocess `wait()` and `poll()` APIs for these child
processes, and allows passing them to existing parts of the code that expect
to work with `Subprocess` objects.

Reviewed By: yfeldblum

Differential Revision: D19588079

fbshipit-source-id: e29b56be4def659cedfd007ddb3a85e826c0a44c

Summary:
Since 60da8ef5, typed tests have been moved to a separate unit (`AsyncBaseTestLib`), which is not the part of `folly` or `folly_test_support`.

This commit adds `AsyncBaseTestLib` to `async_io_test`'s sources.

Alternatively, the test lib could be added to `folly_test_support` with an additional check on `LIBAIO_FOUND`, but `AsyncBaseTestLib` has a specific `TemporaryFile` implementation that conflicts with the one in `experimental/TestUtil.h`.
Pull Request resolved: https://github.com/facebook/folly/pull/1313

Reviewed By: simpkins

Differential Revision: D19648091

Pulled By: yfeldblum

fbshipit-source-id: 27881b464dc3c84aa20ffee1ffd3320e0a67eb27

Summary: [Folly] Use `cacheline_aligned` in small-locks benchmark.

Reviewed By: aary

Differential Revision: D19614282

fbshipit-source-id: d5ec3e9d970654cc532a880de951ac20c898c90a

Summary: std::memcmp is only meaningful for trivially-copyable objects with no padding. Before this diff, `func` is empty class and the check can fail since padding byte is undetermined.

Reviewed By: yfeldblum

Differential Revision: D19460552

fbshipit-source-id: 1b98876a12053c1e7d342e4d8a5fcb25d94f77f3

Summary:
The AsyncGenerator was not clearing its Executor::KeepAlive when it ran to completion.

This meant that when you call `blockingWait(gen.next())` and this would result in the end-of-stream (error or done) then `blockingWait()` would never return because the destructor of the local executor created in the call was blocked waiting for all KeepAlive instances to be destroyed - but this would not happen until the AsyncGenerator was destroyed.

AsyncGenerator now clears the caller context (which includes the KeepAlive) when the coroutine runs to completion.

Reviewed By: kirkshoop

Differential Revision: D19625741

fbshipit-source-id: f3254181af14828b32bfa8f7ef5d400d53c04efa

Summary:
[Folly] `reentrant_allocator`, a multi-thread-safe and async-signal-safe allocator.

Based on `mmap` for dynamic allocation. While, technically, `mmap` is not documented to be async-signal-safe, in practice it is so at least on Linux. Take advantage.

Details:
* Large allocations are handled directly by `mmap` and deallocations by `munmap`, where large-size is 2^12 (1 page) by default; they are not tracked; they are aligned only to page boundaries.
* Small allocations are handled by an `mmap`-backed refcounted arena list. Arena sections are block-size bytes less overhead by default, where block-size is 2^16 (16 pages) by default; they are allocated on demand; and they are linked together to allow for `munmap` when the allocator refcounted arena list is destroyed.

Reviewed By: nbronson, luciang

Differential Revision: D19222635

fbshipit-source-id: adf30580c1bf3dd7f8dab13b1d4ac1e10b2f5c52

Summary:
clang has the warning `-Wnullability-extension` to avoid unportable code. Folly correctly handles it, so let's
ignore the warning if it is set.

(Note: this ignores all push blocking failures!)

Reviewed By: stepancheg

Differential Revision: D19580470

fbshipit-source-id: 59ddb5377a98d907a62df5d480b1ea0a65645a6b

Summary:
MSVC seems to have a bug when handling private classes that
are both part of the inheritence hierarchy (private superclass of a
superclass) and available as a template parameter, incorrectly claiming
that the class is inaccessible due to private inheritance. This diff
works around the issue. It also adds an explicit case to avoid a narrowing
warning generated by MSVC.

(Note: this ignores all push blocking failures!)

Reviewed By: yfeldblum, mjoras

Differential Revision: D19561643

fbshipit-source-id: 494d1611ce662a0eb6c441c64479d3c5d2bbd5b6

Summary: It's unsafe to destroy the object before keepAlive reaches 0. We can still skip running the scheduled functions if those are unsafe to run.

Reviewed By: nanshu

Differential Revision: D19605152

fbshipit-source-id: b99c28dc65da2a5621ce5260491c13754fca03ad

Summary: Produce a -Wdeprecated-declarations warning in code that uses the old SYNCHRONIZED macros.

Reviewed By: yfeldblum

Differential Revision: D19567088

fbshipit-source-id: b0637c06cbd125d09e1b39eec1e82a71ddca2324

Summary:
`-Wrange-loop-analysis` complains about the unnecessary copy.

```
In file included from servicerouter/common/InstanceCount.h:9:
folly/experimental/SingletonRelaxedCounter.h:77:21: error: loop variable 'kvp' of type 'const std::pair<std::atomic<long> *const, unsigned long>' creates a copy from type 'const std::pair<std::atomic<long> *const
, unsigned long>' [-Werror,-Wrange-loop-analysis]
    for (auto const kvp : tracking->locals) {

...

folly/experimental/SingletonRelaxedCounter.h:77:10: note: use reference type 'const std::pair<std::atomic<long> *const, unsigned long> &' to prevent copying
    for (auto const kvp : tracking->locals) {
```

Reviewed By: stepancheg

Differential Revision: D19554423

fbshipit-source-id: 742dd4e7d144dd2dd0c888236b85f222373e9e54

Reviewed By: yfeldblum

Differential Revision: D19579433

fbshipit-source-id: ab63d21e679a8f73547d27200e738558fc69a15b