Summary:
The current MergeChannel public interface has a few problems that have become more apparent with usage.

1. The output receiver does not contain the key associated with the input receiver that geneated the value. This often means that the user must transform input receivers before adding them to the merge channel to include the key, which wastes memory.
2. There is no way for a consumer of the output receiver to know if an input receiver was closed without an exception.
3. If an input receiver was closed with an exception, the entire merge channel is closed.

This diff changes the public API to fix these problems. Instead of returning an output receiver that just has values, the output receiver contains MergeChannelEvent objects. Each such event object has the key of the corresponding input receiver associated with the event, along with the event contents. There are four types of events:

1. A value was received through an existing input receiver.
2. A new input receiver was added.
3. An existing input receiver was removed.
4. An existing input receiver was closed (with or without an exception).

Note that we could theoretically avoid genering events for 2 and 3 (as they correspond with calls that the user makes to add and remove), but there are cases where it is useful to have 2 and 3. This way, an event is always received when an existing input receiver is removed, regardless of whether it was removed explicitly (through a call to removeReceiver) or implicitly (if the input receiver was closed).

Reviewed By: aary

Differential Revision: D33033374

fbshipit-source-id: 34f74cba3f765c76a80d448647d733ad38a2195c

Summary: This diff changes MergeChannel's template parameter names to adhere to folly's style guidelines. It also reverses the order of the template parameters, as it conceptually makes more sense when the key parameter precedes the value parameter.

Reviewed By: Gownta

Differential Revision: D33033369

fbshipit-source-id: 0b31d18baa26778ba9e8d17b9419cb180b1939d9

Summary: It is pure syntax - so, dependency-free - and we may like to use in other places in the same header.

Reviewed By: iahs, luciang

Differential Revision: D32946539

fbshipit-source-id: 2843b73af178c6d7636a4431e61017f35a48681e

Summary: By marking their constructors and conversions as `constexpr`.

Reviewed By: dstechenko

Differential Revision: D33176575

fbshipit-source-id: bb650e768caac787ea7faf249e9c5e80b67c03a1

Summary: Standard containers are expected to have cr iterator variants.

Reviewed By: yfeldblum

Differential Revision: D33165480

fbshipit-source-id: fa6ca8899afc22fed3502ef7ed03723047c129e4

Summary: Describing the common use of keeping test/benchmark check functions.

Reviewed By: Gownta, luciang

Differential Revision: D33146116

fbshipit-source-id: 00b2ee2a02eaa5fd23ab3fa5547dce1828b66db2

Summary:
The history of `sk_*_free` and `sk_*_pop_free`:

* OpenSSL 1.0.2 - These were macros
* OpenSSL 1.1.0 - These changed to static inline functions
* OpenSSL 1.1.1 - These remained as static inline functions (for ABI compatibility with OpenSSL 1.1.0)
* OpenSSL 3.0.0 - These changed to macros.

Give up on trying to use the public interface for deleting `STACK_OF(T)`. This diff switches the implementation
to just invoke the underlying `OPENSSL_sk_free` and `OPENSSL_sk_pop_free`.

If a future version of OpenSSL changes the implementation such that `STACK_OF(T)` is not simply an alias of `OPENSSL_STACK`, we'll deal with it then.

Reviewed By: yfeldblum

Differential Revision: D33071607

fbshipit-source-id: ef37c984a120652c5224410cbafa634d4dc763d1

Summary: This doesn't compile on platform010

Reviewed By: yfeldblum

Differential Revision: D33143441

fbshipit-source-id: b91d08ffb06acfc1ee70399405bca978016da204

Summary:
Create a benchmark to measure subprocess creation time.

This can be used as a baseline to measure future changes to process creation time. Specifically I am interested in optimizing the clean-up of open file descriptors in spawned process.

Reviewed By: yfeldblum

Differential Revision: D33128497

fbshipit-source-id: fa9b69877e6ee77a4a0e3e612e875c88d6bf4e88

Summary: It is SSE < 4 code in a block marked only for SSE >= 4. So, unused.

Reviewed By: Gownta

Differential Revision: D33117027

fbshipit-source-id: d6957d6c73a002cc93623116f79a24f8ff599d79

Reviewed By: simpkins

Differential Revision: D33062040

fbshipit-source-id: a4d36cc558fc14de5929aa4ba278d394e5ca6225

Summary: fb303 contains rust thrift build, but it doesn't run it from its cmake CI, so we can remove this dependency providing that its rust using consumers have the dependency

Reviewed By: Imxset21, fanzeyi

Differential Revision: D33091506

fbshipit-source-id: fc738602104249b5de6c656d59b2eb82987bf1a0

Summary:
For some reasons, I only allow my compiler (gcc 10.0.1) to use x86 SSE, SSE2 and SSE3 instructions. When I compile, I get the following error:

```
In file included from /opt/1A/toolchain/x86_64-v20.0.12/lib/gcc/x86_64-1a-linux-gnu/10.0.1/include/smmintrin.h:32,
                 from /opt/1A/toolchain/x86_64-v20.0.12/lib/gcc/x86_64-1a-linux-gnu/10.0.1/include/nmmintrin.h:31,
                 from /home/docker/opensource/folly/folly/GroupVarint.h:43,
                 from /home/docker/opensource/folly/folly/test/GroupVarintTest.cpp:17:
/home/docker/opensource/folly/folly/GroupVarint.h: In function 'folly::GroupVarint<unsigned int>::decode(char const*, unsigned int*, unsigned int*, unsigned int*, unsigned int*)':
/opt/1A/toolchain/x86_64-v20.0.12/lib/gcc/x86_64-1a-linux-gnu/10.0.1/include/tmmintrin.h:136:1: error: inlining failed in call to 'always_inline' '_mm_shuffle_epi8(long long __vector(2), long long __vector(2))': target specific option mismatch
  136 | _mm_shuffle_epi8 (__m128i __X, __m128i __Y)
      | ^~~~~~~~~~~~~~~~
```

After some searches, it seems the requested instructions are defined in SSE4.1 as we can see in `nmmintrin.h`:
```
#ifndef _NMMINTRIN_H_INCLUDED
#define _NMMINTRIN_H_INCLUDED

/* We just include SSE4.1 header file.  */
#include <smmintrin.h>

#endif /* _NMMINTRIN_H_INCLUDED */
```
This PR fix the code accordingly.
Pull Request resolved: https://github.com/facebook/folly/pull/1330

Reviewed By: Orvid

Differential Revision: D20323400

Pulled By: yfeldblum

fbshipit-source-id: 9913cc5eb378d180403094589b852d45e70978bc

Reviewed By: Mizuchi

Differential Revision: D33008210

fbshipit-source-id: 1615b4a0a268343d79b7b83c558c0b60d65be616

Summary: Title

Reviewed By: yfeldblum

Differential Revision: D32573163

fbshipit-source-id: c34c4de9795dcb2796045f5c464757b8f3d544d9

Summary: The current approach is not nothrow-copy-constructible, which violates assumptions about exception types.

Differential Revision: D33009493

fbshipit-source-id: 323e0cd8b8a6808dd79c6b8a0f066d46e2293289

Summary:
While working on adding allocator support for thrift reference data types (cpp.ref) I realized that allocate_unique always expects the allocator to be of the same type. This means we cannot allocate a unique_ptr<T> with a Alloc<char>. Other allocate methods like std::allocate_shared allows this and it is quite useful. This is critical since it is almost impossible to pass a scoped_allocator of a different type for all the containers in a struct if this kind of allocator type conversion magic is not supported.

This diff adds support to folly::allocate_unique to use an allocator for different type.

Reviewed By: yfeldblum

Differential Revision: D33026177

fbshipit-source-id: 88c43374c2482b76697485c1667064441a057117

Summary:
Rather than demanding precisely a random access iterator, allow any base of it.

https://en.cppreference.com/w/cpp/iterator/iterator_tags

Reviewed By: yfeldblum, Gownta

Differential Revision: D33039206

fbshipit-source-id: 7b62c03704a70340a605df4d7885dce29726d8ee

Summary: All usages in fbsource have been removed.

Reviewed By: yfeldblum, philippv

Differential Revision: D33007671

fbshipit-source-id: 74c5f2dad098c6b7cf7085aa1a2e01a48fa81781

Summary:
Drop all the remaining instances of `IOBufQueue::clear()`.

(Note: this ignores all push blocking failures!)

Reviewed By: yfeldblum

Differential Revision: D32941992

fbshipit-source-id: c7d983e215a78014ffda74a944b628e1886c35c8

Summary:
These checkers do verify that a user is on the VPN, or that their certificates
are not expired. Since we didn't bundle these, we weren't checking these
things...

Reviewed By: chadaustin

Differential Revision: D33007790

fbshipit-source-id: e4e7f8b625443b4f2f9bfd0fe181b7a1858cd97c

Summary:
Split the SmartExceptionTracer so that code can depend on the reading
functionality without also registering the throw hooks at the same time. This
is because hooks during throwing an exception can be expensive for some users.
Splitting it this way allows a library to try to print an async trace if it's
available, but still letting the end application decide if it wants to enable
the feature or not.

Reviewed By: yfeldblum, ot

Differential Revision: D32766294

fbshipit-source-id: 1ffd86d026fc2c18dfa547cf2d169f6491a877cc

Summary: `IOBuf` capacity can be rounded up if supported by the allocator, and the round-up is done including the size of the shared info struct in `createCombined()`, so requesting 8 and 16 can return buffers with the same capacity, breaking the test.

Reviewed By: yfeldblum

Differential Revision: D33004319

fbshipit-source-id: 2bfbeaf7249e89325f9d2b11aa844c1ae1b1ceca

Summary:
I noticed mvfst and proxygen started breaking on my other getdeps changes, turns out it wasn't my diffs, it was that the the packaged cmake was now too old for them.

Fix it by packaging 3.20.2 which matches the system version on centos stream 8, which I found works

Reviewed By: bigfootjon

Differential Revision: D32805140

fbshipit-source-id: adfeb625c1df112bb28a743531cf34473cb734c3

Summary: When building fizz with system packages zlib is required

Reviewed By: HarveyHunt

Differential Revision: D32987670

fbshipit-source-id: 7974125707f201418c96c735262ca56d02ab76f4

Summary: The system packages necessary on linux can vary even within .deb or .rpm space.  Add a getdeps.py feature to allow this, and use it in boost to mark that only centos_stream 8 has/needs the special boost169 packages

Reviewed By: HarveyHunt

Differential Revision: D32987046

fbshipit-source-id: 37f771159128e3a61fe4d6bb3d1444e95efd8a62

Reviewed By: simpkins

Differential Revision: D32998904

fbshipit-source-id: ae6d2e5cdfb4c8cbf2863e8f1fad2df3eb33568d

Summary:
## Optimizations:
- Avoid unnecessary creation of empty segments when iterating.
- Keep track of the lowest and highest nonempty segments.

## Improvements:
- Construction/destruction:
```
CHM ctor/dtor -- empty              213 ns    200 ns    196 ns
CHM ctor/dtor -- 1 item             395 ns    388 ns    373 ns
```
```
CHM ctor/dtor -- empty               99 ns     96 ns     94 ns
CHM ctor/dtor -- 1 item             218 ns    217 ns    216 ns
```
- begin() before and after:
```
CHM begin() -- empty                705 ns    689 ns    660 ns
```
```
CHM begin() -- empty                  6 ns      6 ns      6 ns
```
- Iteration before and after:
```
CHM iterate -- empty                722 ns    714 ns    707 ns
CHM iterate -- 1 item               621 ns    601 ns    587 ns
CHM iterate -- 10 items              89 ns     88 ns     84 ns
CHM iterate -- 100 items             55 ns     51 ns     48 ns
CHM iterate -- 1K items              15 ns     14 ns     13 ns
CHM iterate -- 10K items             13 ns     13 ns     12 ns
CHM iterate -- 100K items            13 ns     12 ns     11 ns
CHM iterate -- 1M items              36 ns     29 ns     27 ns
CHM iterate -- 10M items             36 ns     35 ns     34 ns
```
```
CHM iterate -- empty                  7 ns      6 ns      6 ns
CHM iterate -- 1 item                33 ns     32 ns     31 ns
CHM iterate -- 10 items              27 ns     26 ns     26 ns
CHM iterate -- 100 items             34 ns     33 ns     33 ns
CHM iterate -- 1K items              10 ns     10 ns     10 ns
CHM iterate -- 10K items             10 ns      9 ns      8 ns
CHM iterate -- 100K items             9 ns      7 ns      7 ns
CHM iterate -- 1M items              24 ns     19 ns     17 ns
CHM iterate -- 10M items             27 ns     24 ns     23 ns
```
- empty() before and after:
```
CHM empty() -- empty                106 ns    104 ns    102 ns
```
```
CHM empty() -- empty                  1 ns      1 ns      1 ns
```
- size() before and after:
```
CHM size() -- empty                 104 ns    101 ns     99 ns
CHM size() -- 1 item                103 ns    101 ns     96 ns
CHM size() -- 10 items              107 ns    106 ns    104 ns
CHM size() -- 100 items             111 ns    100 ns     95 ns
```
```
CHM size() -- empty                   1 ns      1 ns      1 ns
CHM size() -- 1 item                  2 ns      2 ns      2 ns
CHM size() -- 10 items                8 ns      8 ns      8 ns
CHM size() -- 100 items              88 ns     71 ns     68 ns
```

Reviewed By: Gownta

Differential Revision: D32684489

fbshipit-source-id: fb58277d6e58de0d95cca3c4274e1b08d182a45b

Summary: Add microbenchmarks.

Reviewed By: Gownta

Differential Revision: D32684480

fbshipit-source-id: bf262404d799e79d8dfe05fd5cf5db5da203bbbe

Summary: As title

Reviewed By: Gownta

Differential Revision: D32382113

fbshipit-source-id: 88d06c821306f9452c6b74f4a49fd90631e372f8

Summary: When using multiple libraries in multiple `.so`s, it is sometimes expected that a type be marked `FOLLY_EXPORT` for them to be understood as the same type in RTTI contexts. This turns out to be the case with exceptions. And `StaticSingletonManager` is an RTTI context as well.

Reviewed By: simpkins

Differential Revision: D32979714

fbshipit-source-id: 92b0032428d7a1a184ef886b67034216ee245e64

Summary:
SSLCommonOptions and SSLServerOptions contain ciphers that modern
clients/servers should not be using.

This diff introduces `SSLOptions2021`, containing an updated recommended list of
ciphers and signature algorithms. It renames the old versions into `SSLOptionsCompatibility`
and `SSLServerOptionsCompatibility` and type aliases `SSLOptions` and `SSLServerOptions`
respectively.

At some point in the future (pending rollouts) SSLOptions and SSLServerOptions
will point to the updated versions.

Reviewed By: yfeldblum

Differential Revision: D32969347

fbshipit-source-id: 887572b2037f8eec07a62dbbce10a9ee7a9f4db6

Summary: Their contents may be preprocessed away if the corresponding platform headers are unavailable. But this removes some of the build-time and build-config complexity.

Differential Revision: D32629035

fbshipit-source-id: a2265ceac44b5022ed53d1c7163f591e779da651

Summary: GTest has changed some APIs, requiring that some _CASE flavors become _SUITEs. Codemodded those changes in.

Reviewed By: aary

Differential Revision: D32930335

fbshipit-source-id: e90e0c857fa683d5d08506045393d48de8cc9740

Summary: EventBaseTest/EventBaseTest/0.IdleTime fails under load. Remedy this.

Reviewed By: yfeldblum

Differential Revision: D32965908

fbshipit-source-id: 7f592b6867b9904fe2e2ae83b53e63398d39f9eb

Summary:
We may both improve portability and simplify impleementation by using `std::filesystem` in this case, which is just constructing a path and reading the corresponding symlink.

Pure runtime performance is not critical here - this codepath is mostly used for logging - so the extra allocations are not really a problem.

Reviewed By: Gownta

Differential Revision: D32942561

fbshipit-source-id: 6d24b9ba900dcfd8653a843cce539627f7f57735

Summary: This change was supposed to be in D32607012 (https://github.com/facebook/folly/commit/fc80a3aa0ad74179e3ce1808a18ff104d60d86ca), but didn't make it in.

Reviewed By: praihan

Differential Revision: D32928488

fbshipit-source-id: 1ae34297f42fe2cefc7e58a41787b7b3c5f158ae

Summary:
`IOBufQueue::clear()` is a footgun: while the comment implies that the buffers in the chain can be reused, in reality only the last one is reused, as all methods in `IOBufQueue` only deal with the tail. Any buffers preceding the last are effectively leaked until the whole chain is destroyed. Furthermore, the tail buffer may not be reused either, if it's shared.

Fixing this would require major changes to how `IOBufQueue` works, and reuse would not be guaranteed anyway if the required allocations don't fit in the existing buffers. It is safer to just remove the method.

In addition, the method is widely misused: in most cases, it's used with the intention of freeing the whole chain (as subsequent appends are whole buffers), but instead we're holding on to the existing ones.

Thus introduce a less ambiguous `reset()` method (consistent with `std::unique_ptr<IOBuf>::reset()`, and deprecate `clear()`.

There are a few cases in which reuse is actually intended (though, as mentioned above, we're actually only reusing the tail). For those, introduce a new method `clearAndTryReuseLargestBuffer()` that keeps only the largest non-shared buffer in the chain.

Reviewed By: Gownta

Differential Revision: D32649882

fbshipit-source-id: c12ffcd0809dae2a4421156d25bcf9ac31a337b1

Summary:
`clone()` preserves headroom and tailroom, so what is happening right
now when deserializing from non-managed buffer (such as wrapped range), is that
every lazy field would result in allocation equivalent in size to the input
buffer!

Apply that to a container and that becomes a combinatorial explosion.

(Note: this ignores all push blocking failures!)

Reviewed By: ot

Differential Revision: D32942233

fbshipit-source-id: b81d8389178b27068dce85b414883d75b1f50ca3

Summary:
This code is not used anywhere, and does not currently build.
`operator delete()` was calling `folly_coro_async_malloc()` with the wrong
number of arguments.  Presumably it meant to actually call
`folly_coro_async_free()`.

Reviewed By: yfeldblum, iahs

Differential Revision: D32824148

fbshipit-source-id: bf4266cd33ecc39539dc5fdafd761d8f56d6060d