Summary: This is pretty much the same as D65605100, which was reverted due to a cycle detection ([link](https://fb.workplace.com/groups/mobile.sheriffs/posts/27342968238658432/?comment_id=27343054638649792)). It turns out that the issue was that I didn't run `xplat/cross_plat_devx/somerge_maps/compute_merge_maps.py`.
Reviewed By: jbeshay
Differential Revision: D65761171
fbshipit-source-id: fc852ec13e54ea6ea45b4fda6b4556a78f38fc7f
Summary: This is my second attempt at D61871891. This time, I ran `xplat/cross_plat_devx/somerge_maps/compute_merge_maps.py`, which generated quic/somerge_defs.bzl
Reviewed By: kvtsoy
Differential Revision: D61975459
fbshipit-source-id: bec62acb2b400f4a102574e8c882927f41b9330e
Summary: `PacketEvent` is a very inaccurate and misleading name. We're basically using this as an identifier for cloned packets, so `ClonedPacketIdentifier` is a much better.
Reviewed By: kvtsoy
Differential Revision: D61871891
fbshipit-source-id: f9c626d900c8b7ab7e231c9bad4c1629384ebb77
Summary:
Timely reaction to congestion requires relaying any CE marks to the sender as soon as possible.
This change schedules an ack to be sent whenever incoming packets are received with CE marks. This will only happen when the readEcnOnIngress option is enabled.
Reviewed By: mjoras
Differential Revision: D58423959
fbshipit-source-id: 30f8cf8b11d0446985c2d87d7df67c24c0d5afdf
Summary:
The idea here is to make it so we can swap out the type we are using for optionality. In the near term we are going to try swapping towards one that more aggressively tries to save size.
For now there is no functional change and this is just a big aliasing diff.
Reviewed By: sharmafb
Differential Revision: D57633896
fbshipit-source-id: 6eae5953d47395b390016e59cf9d639f3b6c8cfe
Summary:
Remove cmsgs from each outstanding packet. Storing this as a map is excessive. Instead store it as a packed enum. This supports one kind of mark per packet.
If we ned to we could in the future support multiple types of marks per packet by using a bitset-style flag, e.g. each bit index representing a different mark. For now just use an enum.
Reviewed By: jbeshay
Differential Revision: D57397865
fbshipit-source-id: 6d34215c9d7e39537c44c6c304a8ce3a5883541e
Summary:
The existing batch writers do not handle failed writes to the AsyncUDPSocket. A packet that fails to be written is detected as a packet loss later when feedback is received from the peer. This negatively impacts the congestion controller because of the fake loss signal, and artificially inflates the number of retransmitted packets/bytes.
This change adds a new batch writer (SinglePacketBackpressuretBatchWriter) that retains the buffers when a write fails. For subsequent writes, the writer retries the same buffer. No new packets are scheduled until the retried buffer succeeds.
Notes:
- To make sure that retry writes are scheduled, the write callback is installed on the socket when a buffer needs to be retried.
- The retries are for an already scheduled packet. The connection state reflects the timing of the first attempt. This could still have an impact on rtt samples, etc. but it this is a milder impact compared to fake losses/retranmissions.
- Any changes outside of the batch writer only impact the new batch writer. Existing batch writers do not use the fields and are not affected by the changes in this diff.
Reviewed By: kvtsoy
Differential Revision: D57597576
fbshipit-source-id: 9476d71ce52e383c5946466f64bb5eecd4f5d549
Summary:
std::deque by default allocates a large block on the heap for managing its state. This has a fixed memory cost both per connection (because of the crypto streams) and per stream. CircularDeque by comparison does not have this overhead and is only 32 bytes per structure.
For example:
```
"size": 656,
"name": "readBuffer",
"typePath": ["a0", "conn_", "ptr_val", "cryptoState", "ptr_val", "initialStream", "readBuffer"],
"typeNames": ["std::deque<quic::StreamBuffer, std::allocator<quic::StreamBuffer>>"],
```
This should save about 6kB per connection with no streams, and additional memory per stream.
Reviewed By: jbeshay, hanidamlaj
Differential Revision: D56578219
fbshipit-source-id: ab2b529fa9a4169bea6862b11ccbf178c6f5abb1
Summary:
std::deque by default allocates a large block on the heap for managing its state. This has a fixed memory cost both per connection (because of the crypto streams) and per stream. CircularDeque by comparison does not have this overhead and is only 32 bytes per structure.
For example:
```
"size": 656,
"name": "readBuffer",
"typePath": ["a0", "conn_", "ptr_val", "cryptoState", "ptr_val", "initialStream", "readBuffer"],
"typeNames": ["std::deque<quic::StreamBuffer, std::allocator<quic::StreamBuffer>>"],
```
This should save about 6kB per connection with no streams, and additional memory per stream.
Reviewed By: kvtsoy
Differential Revision: D56496459
fbshipit-source-id: ec4049614939f885f64481bd374c81e74ad47c66
Summary:
Allow the server/client transport to initiate periodic key update. It's defaulted to being disabled.
The new logic for initiating and verifying a key update was handled correctly by the peer is consolidated in QuicTransportFunctions.
Reviewed By: mjoras
Differential Revision: D53109624
fbshipit-source-id: 0c3a944978fc0e0a84252da953dc116aa7c26379
Summary:
This stack adds key update support to Mvfst client and server. This diff adds the main logic for detecting key updates in the QuicReadCodec. When an update is successful, the server transport reacts to it by updating the write phase and cipher.
The high level design is as follows:
- The QuicReadCodec is responsible for detecting incoming key update attempts by the peer, as well as tracking any ongoing locally-initiated key updates.
- Upon detecting a successful key update, the QuicReadCodec updates its state. The Server/Client transport reacts to this change by updating its write phase and cipher.
- A locally initiated key update starts with updating the write phase and key, and signaling the read codec that a key update has been initiated.
- The read codec keeps this in a pending state until a packet is successfully received in the new phase.
- Functions for syncing the read/write phase on incoming key updates, as well as initiating and verifying outgoing key updates are abstracted in QuicTransportFunctions and are used by both the client and server transports.
- Common handshake functions used for rotating the keys are now in HandshakeLayer that is shared by both client and server handshakes.
Reviewed By: mjoras
Differential Revision: D53016559
fbshipit-source-id: 134e965dabd62917193544a9655a4eb8868ab7f8
Summary:
This is the major transition that updates mvfst code to use the new interfaces. The new Folly implementations of the interfaces maintain all the existing behavior of folly types so this should not introduce any functional change. The core changes are:
- Update the BatchWriters to use the new interfaces.
- Update the FunctionLooper to use the new interfaces.
- Change QuicServerTransport to take the folly types and wrap them in the new types for use in the QuicTransportBase.
The rest of the diff is for updating all the existing uses of the QuicTrasnport to initialize the necessary types and pass them to the QUIC transport instead of directly passing folly types.
Reviewed By: mjoras
Differential Revision: D51413481
fbshipit-source-id: 5ed607e12b9a52b96148ad9b4f8f43899655d936
Summary:
- Remove setCustomTransportParameter, which (based on the quic v19 rfc), verifies whether a parameter is within the private range [0xff00, 0xffff]
> Values with the first byte in the range 0x00 to 0xfe (in hexadecimal) are assigned via the Specification Required policy [RFC8126].
- Consolidating adding MaxStreamGroups transport parameter into all other transport parameters extension.
More specifically, `QuicClientTransport::maybeEnableStreamGroups()` logic is now moved into `QuicClientTransport::setSupportedExtensionTransportParameters()`
Reviewed By: mjoras
Differential Revision: D50461610
fbshipit-source-id: 802b546c8364586cdcf36a230b156ca140c57ce4
Summary:
This diff changes `QuicAsyncUDPSocketWrapper` so that it is an abstraction layer that inherits from `QuicAsyncUDPSocketType`, instead of simply being a container with aliases.
- Key changes in `QuicAsyncUDPSocketWrapper.h`, the rest of the updates switch us from using `QuicAsyncUDPSocketType` to `QuicAsyncUDPSocketWrapper`.
- It's difficult to mock the UDP socket today given that we expose the entire `folly::AsyncUDPSocket` type to the higher layers of the QUIC stack. This complicates testing and emulation because any mock / fake has to implement low level primitives like `recvmmsg`, and because the `folly::AsyncUDPSocket` interface can change over time.
- Pure virtual functions will be defined in `QuicAsyncUDPSocketWrapper` in a follow up diff to start creating an interface between the higher layers of the mvfst QUIC stack and the UDP socket, and this interface will abstract away lower layer details such as `cmsgs` and `io_vec`, and instead focus on populating higher layer structures such as `NetworkData` and `ReceivedPacket` (D48714615). This will make it easier for us to mock or fake the UDP socket.
This diff relies on changes to `folly::MockAsyncUDPSocket` introduced in D48717389.
--
This diff is part of a larger stack focused on the following:
- **Cleaning up client and server UDP packet receive paths while improving testability.** We currently have multiple receive paths for client and server. Capabilities vary significantly and there are few tests. For instance:
- The server receive path supports socket RX timestamps, abet incorrectly in that it does not store timestamp per packet. In comparison, the client receive path does not currently support socket RX timestamps, although the code in `QuicClientTransport::recvmsg` and `QuicClientTransport::recvmmsg` makes reference to socket RX timestamps, making it confusing to understand the capabilities available when tracing through the code. This complicates the tests in `QuicTypedTransportTests`, as we have to disable test logic that depends on socket RX timestamps for client tests.
- The client currently has three receive paths, and none of them are well tested.
- **Modularize and abstract components in the receive path.** This will make it easier to mock/fake the UDP socket and network layers.
- `QuicClientTransport` and `QuicServerTransport` currently contain UDP socket handling logic that operates over lower layer primitives such `cmsg` and `io_vec` (see `QuicClientTransport::recvmmsg` and `...::recvmsg` as examples).
- Because this UDP socket handling logic is inside of the mvfst transport implementations, it is difficult to test this logic in isolation and mock/fake the underlying socket and network layers. For instance, injecting a user space network emulator that operates at the socket layer would require faking `folly::AsyncUDPSocket`, which is non-trivial given that `AsyncUDPSocket` does not abstract away intricacies arising from the aforementioned lower layer primitives.
- By shifting this logic into an intermediate layer between the transport and the underlying UDP socket, it will be easier to mock out the UDP socket layer when testing functionality at higher layers, and inject fake components when we want to emulate the network between a mvfst client and server. It will also be easier for us to have unit tests focused on testing interactions between the UDP socket implementation and this intermediate layer.
- **Improving receive path timestamping.** We only record a single timestamp per `NetworkData` at the moment, but (1) it is possible for a `NetworkData` to have multiple packets, each with their own timestamps, and (2) we should be able to record both userspace and socket timestamps.
Reviewed By: jbeshay, hanidamlaj
Differential Revision: D48717388
fbshipit-source-id: 4f34182a69ab1e619e454da19e357a6a2ee2b9ab
Summary: Turns out these functions have been returning 0 for bytes written for a while. Nothing was using them (yet), so there wasn't any functional breakage.
Reviewed By: kvtsoy
Differential Revision: D46653336
fbshipit-source-id: 765b3363c1fc0729e8239d1293ddcc4ae4bb9c79
Summary:
Adds a prewrite function in packet processors that can be used to set PreWrite Reuqests that will apply apply to the next write loop in the QuicSocket.
Currently, the PrewriteRequest only contains a list of cmsgs but it can be expanded when needed.
This replaces the setAdditionalCmsgs callback that was previously passed down directly to AsyncUDPSocket, allowing multiple processors to affect the additionalCmsgs sent to the AsyncUDPSocket.
Reviewed By: bschlinker
Differential Revision: D44233700
fbshipit-source-id: 799357ee5df89b2a526611bd83d0b4a9a13b71d5
Summary:
We don't need to carry these states after the handshake is confirmed, so make them pointers instead. This will facilitate adding a structure to the AckState for tracking duplicate packets.
(Note: this ignores all push blocking failures!)
Reviewed By: hanidamlaj
Differential Revision: D41626895
fbshipit-source-id: d8ac960b3672b9bb9adaaececa53a1203ec801e0
Summary:
- PTOs should not be subject to congestion control limits
- Quickly recover from PTOs being writableBytesLimited by calling onPTOAlarm() as soon as we become unblocked
Reviewed By: mjoras
Differential Revision: D35480409
fbshipit-source-id: 51500db6fff17a7badefea8bda7f63141e97f746
Summary: - logging number of bytes the server sent during the handshake for insight.
Reviewed By: mjoras
Differential Revision: D33069800
fbshipit-source-id: e7e8f25183ee30de99e2971bae3c6b93882f6e63
Summary: Otherwise we can end up in a situation where the non-DSR scheduler was limited by the loop time while the DSR scheduler is not, leading to an inconsistency.
Reviewed By: lnicco
Differential Revision: D32570027
fbshipit-source-id: f43d2517589c22bac0f2bb76626cc55c2a21fa5d
Summary:
These are either no longer relevant, are unlikely to be done, or are spculative enough that they don't deserve code space.
Hope here is to make our search for TODOs higher signal.
Reviewed By: lnicco
Differential Revision: D29769792
fbshipit-source-id: 7cfa62cdc15e72d8b7b0cd5dbb5913ea3ca3dc5a
Summary: Right now we are running the handshakeConfirmed code a lot on the client. This is excessive. We only need to run the code if we haven't already dropped the cipher.
Reviewed By: yangchi
Differential Revision: D27725974
fbshipit-source-id: ca325c132debdd280e447ca30876488b879ff13c
Summary:
As before we will now aggressively send probes on all spaces with probes available when the PTO timer fires.
This time with more unit tests and some bug fixes.
Reviewed By: yangchi
Differential Revision: D27338523
fbshipit-source-id: 8a9ccb90ed691e996fab4afa2f132c0f99044fbc
Summary:
Previously, we maintained state and counters to count both, header and body
bytes together. This commit introduces additional counters and state to keep
track of just the body bytes that were sent and acked etc. Body bytes received
will be implemented later.
Reviewed By: bschlinker
Differential Revision: D27312049
fbshipit-source-id: 33f169c9168dfda625e86de45df7c00d1897ba7e
Summary:
(1) Check if stream has DSR data
(2) Write them to the packet builder
(3) Guard against flow control and congestion control limit
(4) Add the instruction to the DSRPacketizationRequestSender
(5) Flush
(6) update connection state
Reviewed By: mjoras
Differential Revision: D26851557
fbshipit-source-id: 850958995ef10744826ee892fdea211ac441b078
Summary: try to land this again without the compiler flags change
Differential Revision: D26958681
fbshipit-source-id: d00659aaf819dbb2942da8b41deab3d108a19f0f
Summary:
Update retransmission queue in the stream after BufferMeta is written,
for both new data and lost data cases.
Reviewed By: mjoras
Differential Revision: D26260987
fbshipit-source-id: 15a4fa17826426b4b972b63cf370fe791b3101ff
Summary:
This is the transport function that sends d6d probe to socket. It
- checks if there's the pending event `sendD6DProbePacket`, which will be set by upon d6d's probe timer expiration.
- uses the current probeSize to instantiate a D6DProbeScheduler, which will be changed by a small d6d state machine.
Reviewed By: yangchi
Differential Revision: D23193967
fbshipit-source-id: dfe6ce831cfd6ff0470e801644b95d4e8da34e87
Summary:
The problem with Ping being a simple frame:
(1) All SimpleFrames are in the same scheduler. So sending ping means we may
also send other frames which can be problematic if we send in Initial or
Handshake space
(2) Ping isn't retranmisttable. But other Simple frames are. So we are
certainly setting this wrong when we send pure Ping packet today.
That being said, there are cases where we need to treat Ping as retransmittable.
One is when it comes to update ack state: If peer sends us Ping, we may want to
Ack early rather than late. so it makes sense to treat Ping as retransmittable.
Another place is insertion into OutstandingPackets list. When our API user sends
Ping, then also add a Ping timeout. Without adding pure Ping packets into OP list,
we won't be able to track the acks to our Pings.
Reviewed By: mjoras
Differential Revision: D21763935
fbshipit-source-id: a04e97b50cf4dd4e3974320a4d2cc16eda48eef9
Summary:
On loss timer, currently we knock all handshake packets out of the OP
list and resend everything. This means miss RTT sampling opportunities during
handshake if loss timer fires, and given our initial loss timer is likely not a
good fit for many networks, it probably fires a lot.
This diff keeps handshake packets in the OP list, and add packet cloning
support to handshake packets so we can clone them and send as probes.
With this, the handshake alarm is finally removed. PTO will take care of all
packet number space.
The diff also fixes a bug in the CloningScheduler where we missed cipher
overhead setting. That broke a few unit tests once we started to clone
handshake packets.
The writeProbingDataToSocket API is also changed to support passing a token to
it so when we clone Initial, token is added correctly. This is because during
packet cloning, we only clone frames. Headers are fresh built.
The diff also changed the cloning behavior when there is only one outstanding
packet. Currently we clone it twice and send two packets. There is no point of
doing that. Now when loss timer fires and when there is only one outstanding
packet, we only clone once.
The PacketEvent, which was an alias of PacketNumber, is now a real type that
has both PacketNumber and PacketNumberSpace to support cloning of handshake
packets. I think in the long term we should refactor PacketNumber itself into a
real type.
Reviewed By: mjoras
Differential Revision: D19863693
fbshipit-source-id: e427bb392021445a9388c15e7ea807852ddcbd08
Summary:
Now we won't have a zero PTO and we will properly clear out the outstanding packets.
Note that this cipher dropping is not what the draft prescribes, instead dropping both the initial and handshake ciphers when we know 1-rtt communication is functioning.
Reviewed By: yangchi
Differential Revision: D20388737
fbshipit-source-id: 0b89eb80c8faa796ab09eda3eaa10a00dcf7bae9
Summary:
To prepare for another configuration of this chain. With this, now we
can land all the outstanding GSO optimization diffs without having to worry
about breaking the production code.
Reviewed By: mjoras
Differential Revision: D20838453
fbshipit-source-id: 807a0c546305864e0d70f8989f31d3de3b812278
Summary:
The fizz `encrypt` API will opportunistically use inplace encryption if the `IOBuf` you pass it is correctly sized. If it is not it will manually create the output buffer.
It turns out doing inplace versus encrypting with a given target output is a bit faster in microbenchmarks, and it will also allow experimenting with caching `IOBuf`s on the send path.
Reviewed By: knekritz
Differential Revision: D19506291
fbshipit-source-id: 3ef41290538ceac34a344114badbd167e2c25a50
