commit | 3dac5b70f053ca2988e9ac77eb64f55b9193acbb | [log] [tgz] |
---|---|---|
author | Ryan Prichard <rprichard@google.com> | Tue Oct 11 13:09:35 2022 -0700 |
committer | Ryan Prichard <rprichard@google.com> | Tue Oct 11 13:26:22 2022 -0700 |
tree | 9571e840ff177e38abccb33c4d7b1e74e364fb6b | |
parent | 59be29d99b8632b1f2e6f1eb7b9e23eb53ba34f5 [diff] |
Refactor ReadAcknowledgements to fix two race conditions Race 1: * Suppose there is a single successful copy to acknowledge. * The early opportunistic ReadAcknowledgements call (read_all==false) won't block. * If the 8-byte ID_OKAY is ready to read soon enough, then the first adb_poll and adb_read calls will bring the entire 8 bytes into acknowledgement_buffer_. * ReadAcknowledgements then reaches "// We don't want to read again yet, because the socket might be empty." and restarts the loop. * The second adb_poll finds no more data and ReadAcknowledgements returns. * The final blocking ReadAcknowledgements call (read_all==true) calls adb_poll and hangs because there are no more bytes to read. Race 2: * Suppose that when ReadAcknowledgements (read_all==true) is called, the socket has less than a complete header available (i.e. 1 <= NUM_BYTES <= 7). * The first adb_read will return the partial header, then ReadAcknowledgements will reach "// Early exit if we run out of data in the socket." We're supposed to block until all the acknowledgements are read, though! * adb_read uses TEMP_RETRY_FAILURE, but that merely restarts an EINTR syscall -- it doesn't ensure that all requested bytes were read. * I found this bug by inspection. I haven't reproduced it. Factor out a "read_until_amount" function that fills acknowledgement_buffer_ to a given number of bytes (or exits early if non-blocking). Use this function to read both the copy response header and the copy response error message. This patch establishes an invariant: when ReadAcknowledgements returns, acknowledgement_buffer_ may be non-empty, but it won't have a complete message in it. Bug: 230428773 Test: treehugger Change-Id: I28a28b6d5f79f9b64913b8b2227e336ced05582e
If you are new to adb source code, you should start by reading OVERVIEW.TXT which describes the three components of adb pipeline.
This document is here to boost what can be achieved within a "window of naive interest". You will not find function or class documentation here but rather the "big picture" which should allow you to build a mental map to help navigate the code.
As outlined in the overview, this codebase generates three components (Client, Server (a.k.a Host), and Daemon (a.k.a adbd)). The central part is the Server which runs on the Host computer. On one side the Server exposes a "Smart Socket" to Clients such as adb or DDMLIB. On the other side, the Server continuously monitors for connecting Daemons (as USB devices or TCP emulator). Communication with a device is done with a Transport.
+----------+ +------------------------+ | ADB +----------+ | ADB SERVER | +----------+ | CLIENT | | | | (USB)| ADBD | +----------+ | | Transport+-------------+ (DEVICE) | | | | +----------+ +----------- | | | | ADB | v + | +----------+ | CLIENT +--------->SmartSocket | (USB)| ADBD | +----------+ ^ | (TCP/IP) Transport+-------------+ (DEVICE) | | | | +----------+ +----------+ | | | | DDMLIB | | | Transport+--+ +----------+ | CLIENT +----------+ | | | (TCP/IP)| ADBD | +----------+ +------------------------+ +----------|(EMULATOR)| +----------+
The Client and the Server are contained in the same executable and both run on the Host machine. Code sections specific to the Host is enclosed within ADB_HOST
guard. adbd runs on the Android Device. Daemon specific code is enclosed in !ADB_HOST
but also sometimes with-in __ANDROID__
guard.
A smart socket is a simple TCP socket with a smart protocol built on top of it. This is what Clients connect onto from the Host side. The Client must always initiate communication via a human readable request but the response format varies. The smart protocol is documented in SERVICES.TXT.
On the other side, the Server communicates with a device via a Transport. adb initially targeted devices connecting over USB, which is restricted to a fixed number of data streams. Therefore, adb multiplexes multiple byte streams over a single pipe via Transport. When devices connecting over other mechanisms (e.g. emulators over TCP) were introduced, the existing transport protocol was maintained.
At the heart of both the Server and Daemon is a main thread running an fdevent loop, which is a platform-independent abstraction over poll/epoll/WSAPoll monitoring file descriptors events. Requests and services are usually served from the main thread but some service requests result in new threads being spawned.
To allow for operations to run on the Main thread, fdevent features a RunQueue combined with an interrupt fd to force polling to return.
+------------+ +-------------------------^ | RUNQUEUE | | | +------------+ | POLLING (Main thread) | | Function<> | | | +------------+ | | | Function<> | ^-^-------^-------^------^^ +------------+ | | | | | ... | | | | | +------------+ | | | | | | | | | | |============| | | | | |Interrupt fd+------+ +----+ +----+ +----+ +------------+ fd Socket Pipe
The asocket, apacket, and amessage constructs exist only to wrap data while it transits on a Transport. An asocket handles a stream of apackets. An apacket consists in a amessage header featuring a command (A_SYNC
, A_OPEN
, A_CLSE
, A_WRTE
, A_OKAY
, ...) followed by a payload (find more documentation in protocol.txt. There is no A_READ
command because an asocket is unidirectional. To model a bi-directional stream, asocket have a peer which go in the opposite direction.
An asocket features a buffer where the elemental unit is an apacket. If traffic is inbound, the buffer stores the apacket until it is consumed. If the traffic is oubound, the buffer stores apackets until they are sent down the wire (with A_WRTE
commands).
+---------------------ASocket------------------------+ | | | +----------------APacket Queue------------------+ | | | | | | | APacket APacket APacket | | | | +--------+ +--------+ +--------+ | | | | |AMessage| |AMessage| |AMessage| | | | | +--------+ +--------+ +--------+ | | | | | | | | | | | | | | ..... | | | | | | | | | | | Data | | Data | | Data | | | | | | | | | | | | | | | | | | | | | | | | | +--------+ +--------+ +--------+ | | | | | | | +-----------------------------------------------+ | +---------------------------------------------------+
This system allows to multiplex data streams on an unique byte stream. Without entering too much into details, the amessage fields arg1 and arg2 are used alike in the TCP protocol where local and remote ports identify an unique stream. Note that unlike TCP which feature an "unacknowledged-send window", an apacket is sent only after the previous one has been confirmed to be received.
The two types of asocket (Remote and Local) differentiate between outbound and inbound traffic.
This pipeline is detailed in daemon/jdwp_service.cpp with ASCII drawings! The JDWP extension implemented by Dalvik/ART are documented in: