docs/performance.md compares master (C) with the C++ engine on macOS
arm64: zero heap allocations per connection or datagram (from 27-89),
TCP throughput within noise, and about 20 MiB more RSS at 1,000 open
connections from the inline per-connection buffers. docs/cpp-core.md
and CLAUDE.md now describe the finished layout.
Sizes the resolver's query pool, the last of the capacity settings the
no-heap-after-startup design calls for. Defaults are unchanged: the
connection limit (plus 64 on ss-server, which also resolves for its
statistics).
The C cache (now used only by the reference codecs) moves to
tests/reference and src/uthash.h is deleted. ACL rules and ss-manager's
servers are ListNodes on IntrusiveLists; destroying an element unlinks
it, so the free paths no longer have to remember to. core.h keeps only
ss_ip. test_rule is ported to C++ and also covers literal rules.
The programs now use only the C++ codecs. The C stream/AEAD/AEAD-2022
implementations stay as the reference the differential tests and the C
unit tests compare against. Key derivation and the cipher tables used by
the doc snippets move to src/crypto/keys.{hpp,cpp}.
ss-manager keeps its servers in an ss_list (at most MAX_PORT_NUM,
looked up only on control messages) instead of a uthash table, and
tests/test_manager.py covers add/ping/list/remove end to end.
The loop is created before the programs fork into the background, and
kqueue/epoll descriptors do not survive fork(), so ss-server -f (which
ss-manager always uses) died as soon as it ran. Call uv_loop_fork() in
the child.
Each program's loop, signals, plugin watcher and stats timer now go
through ss::event::Program (src/event/program.hpp), held on main()'s
stack; ss-manager uses it too. ss-server resolves the manager address
with the relay resolver. With no C caller left, ss_event.cpp/.h,
resolv.cpp/.h and their C tests go; test_resolver_cpp.cpp replaces the
resolver tests against the same delayed local DNS server, and ss_ev is
now an interface target over libuv.
<shadowsocks/control.h> starts ss-local on a library-owned thread,
stops it from any thread through a libuv async handle and waits for it;
<shadowsocks/shadowsocks.hpp> wraps that in a move-only, header-only
shadowsocks::LocalServer whose destructor stops and waits. The library
exports only C symbols, shadowsocks.h is unchanged, and the shared
library version becomes 2.1.0 (SOVERSION 2). test_local_api_cpp covers
the lifecycle; the installed-package consumer gains a C++ program.
The TCP connection pool and the UDP session table are sized from these
options (or max_connections / max_udp_sessions in the configuration
file) at startup instead of fixed constants; the DNS query pool follows
the connection limit. test_limits.py checks that a connection beyond
the limit is refused and that accepting resumes once one closes.
GCC enables -Wsign-compare with -Wall for C++ (clang does not), which
turned a few signed/unsigned comparisons in base64, netutils, server and
manager into errors; udp.cpp needed <ctime>.
The programs point libuv, c-ares and Mbed TLS (MBEDTLS_PLATFORM_MEMORY in
the bundled build) at their own FixedHeap before anything runs, mark the
end of startup with heap_seal(), and in verbose mode report heap use and
any allocation after it at exit. The library build keeps the host's
allocator.
measure_alloc.py gains --udp, and the alloc_steady_programs_* tests now
fail on any malloc per relayed TCP connection (by name and by address)
or UDP association, for every cipher family.
musl's CMSG_NXTHDR compares a size_t with a ptrdiff_t, which Clang's
-Wsign-compare rejects in C++. Call it from one helper that suppresses
the warning there.
ss-redir's TCP side moves onto the relay engine (per-listener TOS for the
DSCP ports). src/relay/udp replaces udprelay.cpp: sessions keyed by
client address own their upstream socket and live in a fixed LRU,
AEAD-2022 server sessions are shared through RefPtr between the
session-ID table and the relays using them, and datagrams waiting for
DNS sit in pool slots. Packets are received into and built in
relay-owned buffers instead of a 64 KiB allocation per datagram. The
programs no longer call crypto_init(); Method sets up libsodium and the
replay filter.
crypto.h's min/max macros broke <algorithm> in C++ sources that include
it first, netutils.h's linux/tcp.h clashed with the netinet/tcp.h libuv
includes, and GCC rejects memcpy into crypto_t. Apple clang accepted all
three.
The request header, outbound ACL, target resolution through the relay's
resolver, the anti-probing stop stage and the manager traffic counters
are engine code now; the connmark-to-TOS tracker becomes a pair of relay
hooks. server.cpp keeps option parsing, nftables and the stats timer.
The SOCKS5 handshake, the ACL bypass decision (with its own resolver for
host-name rules) and the short wait for a first payload are now engine
code; local.cpp keeps only option parsing and startup. The library entry
points start the same engine.
One engine (src/relay/tcp) for the four TCP relays: connections live in a
fixed pool owned through PoolPtr, handshake, cipher state and buffers are
inline, timeouts are intrusive deadline lists on one timer and DNS answers
come back through generation-checked handles. A full pool refuses new
connections and pauses accepting until a slot frees.
ss-tunnel is the first user; its hand-written relay callbacks are gone.
tests/tunnel_interop.py covers every method against ss-server and
shadowsocks-rust.
Queries live in pool slots with their first A and AAAA answer inline,
c-ares sockets are watched from a fixed table, and answers reach the
waiter through a callback plus opaque words instead of a back-pointer.
resolv.cpp keeps the C API as a shim. The event Loop can now adopt the
ss_event loop, and a Poll can stop watching a socket it does not own.
TCP request/response streams with the salt echo between the two
directions, an exact-match salt pool on FixedLru, and UDP packets sealed
and opened in place with caller-owned session state. Differentially
tested against aead2022.cpp in both roles; the steady phase allocates
nothing.
The session table and the relay contexts each hold a reference; evicting a
session from the full table no longer frees memory a relay still uses.
Ported from the parked fix/aead2022-udp-session-eviction work.
src/crypto/stream.hpp implements the legacy stream methods this build
supports (AES CFB/CTR, Camellia CFB, Salsa20, ChaCha20, ChaCha20-IETF)
as value types. The AES and Camellia contexts are held by value through
Mbed TLS's block-cipher APIs instead of an allocated
mbedtls_cipher_context_t, and the libsodium ciphers track their
keystream position, finishing partial 64-byte blocks from a stack block
instead of the padding copies stream.cpp makes.
test_crypto_stream_cpp runs TCP streams (with fragments that split
keystream blocks) and UDP packets between the C and C++
implementations in both directions for all twelve methods and checks
IV replay; alloc_steady_crypto_stream requires zero allocations for new
connections and packets.
cache.cpp keeps its C API but replaces uthash and the two mallocs per
entry with one slab of nodes holding their keys inline, linked in
last-use order, and an open-addressing index of node ids that hashes
through the slab with a per-cache random seed. Both blocks are reserved
by cache_create() and touched only as entries are used, so a large,
mostly empty cache costs no resident memory.
Semantics are unchanged: an insert evicts the least recently used entry
once the cache holds capacity entries (the AEAD-2022 session table's
4095/4096 boundary), lookups refresh recency and age, cache_clear()
drops stale entries, and free_cb may re-enter the cache. Re-inserting a
key now replaces its data instead of shadowing it. cache_create_keyed()
sizes keys per cache; the AEAD-2022 salt pool and session table use it.
test_cache_cpp covers the boundary, ordering, replacement, ageing,
re-entrant callbacks and limits; alloc_steady_cache requires zero
allocations across 100,000 inserts on a full cache. AEAD-2022 ss-server
drops from 84 to 80 allocations per relayed connection.
Mbed TLS 2 returns void from mbedtls_sha1_starts/update/finish (the int
versions are the _ret variants), which the system-library build on
Ubuntu (2.28) rejects. crypto/mbedtls_compat.hpp gives one int-returning
spelling for both.
src/crypto/aead.hpp implements the SIP004/SIP007 AEAD ciphers as value
types: AeadEngine holds the libsodium or Mbed TLS GCM state by value,
AeadTcpEncryptor/AeadTcpDecryptor frame the chunk stream with a fixed
reassembly buffer, and aead_udp_seal/open handle packets. HKDF-SHA1 runs
on stack SHA-1 state instead of Mbed TLS's allocating HMAC API.
The bundled Mbed TLS config now enables MBEDTLS_BLOCK_CIPHER_C, which
keeps the AES context inside mbedtls_gcm_context, so GCM key setup no
longer allocates. That alone saves the C programs two allocations per
connection.
test_crypto_aead_cpp checks HKDF against RFC 5869 and crypto_hkdf, runs
TCP streams and UDP packets between the C and C++ implementations in
both directions for all five AEAD methods, and covers tampering,
framing and salt replay. alloc_steady_crypto_aead runs it under the
malloc counter and requires zero allocations across 600 new
connections and 600 packets (tests/alloc_audit/check_steady.py).
ss_strict_warnings() now applies -Wextra per source file so targets
mixing new C++ with converted C modules keep the latter at -Wall, and
Result gains move assignment.
src/event/loop.hpp wraps uv_loop/poll/timer/signal/async handles that
live by value inside their owner and dispatch to an owner member
function chosen at compile time. A Poll owns its socket and carries
read and write interest on one handle, closing the descriptor in
libuv's close callback. CloseGroup collects the asynchronous closes of
an owner's handles and reports completion afterwards, which is how
pooled owners implement AsyncClosable. DeadlineList drives the idle
timeouts of many owners from one timer.
test_event_cpp covers readiness, closing a peer from inside a
callback, recycling a pooled owner only after its handles closed,
repeating timers, deadline ordering and cross-thread wakeups.
ss-redir only builds on Linux, so the macOS conversion checks never
compiled it: getsockname() wants a struct sockaddr *, and memcpy's
void * result needs a cast.
Rename utils, ss_event, udprelay, winsock, local, server, tunnel, redir
and manager from .c to .cpp. Inside the moved files only the
\snippet{doc} file references and header comments change; snippet
references in doc/*.md, tests/test_cli_docs.py, the build files and
CONTRIBUTION.md follow the new names.
Also cast the Windows-only string literal returned by
get_default_conf(), found by the zig x86_64-windows-gnu build.
Every module of src/ except the unbuilt android.c and the vendored
blake3 is now compiled as C++20. The new .cpp files join the no-heap
ratchet baseline.
Cast allocator results and void* callback data explicitly, cast string
literals stored in char * variables, replace compound literals in
ss_event.c with zeroing plus member assignment, and give ss-server's
listener array its MAX_REMOTE_NUM bound instead of a VLA.
Where a goto CLEAN_UP jumps over an initialised declaration (C++
rejects that), split the declaration from its initialisation in place.
The variables were already indeterminate at the label in C, and the
cleanup code does not read them. Checked for every MODULE_* variant of
udprelay.c, plus LIB_ONLY for local.c. Still valid C.
Rename netutils, jconf, stream, aead, aead2022, aead2022_udp, crypto,
acl, plugin, plugin_disabled and resolv from .c to .cpp without changing
their contents, and update source lists, snippet references, test
comments and debian/copyright.
The Doxygen source list and scripts/check_cli_docs.py now accept either
extension for each module, so the remaining programs can move one at a
time. The new .cpp files are added to the no-heap ratchet baseline.
socketserver listens with a backlog of 5. With concurrent connections
through the proxy on a loaded machine (parallel ctest), the backlog
overflows, the kernel resets the origin connection, and the proxy
correctly drops the client, which surfaced as intermittent
alloc_audit_report failures ("getpeername: Invalid argument" in
ss-local, "Connection reset by peer" in ss-server). Reproduced
reliably at 64 concurrent connections; gone with a backlog of 256.
Cast allocator and memchr results explicitly, pass mbedtls_operation_t
to mbedtls_cipher_setkey, and declare the two locals in
ss2022_encrypt_first() before the gotos that jump past them. Drop the
unused RESOLV_MODE global that resolv.c defined by accident through an
anonymous enum declaration. Still valid C.
GCC's C++ -Wall includes -Wsign-compare and -Wclass-memaccess, which C's
-Wall does not, so the converted modules failed -Werror on GCC (Linux,
UCRT64) although nothing in them changed. They keep the warning set they
were written against; code under ss_strict_warnings() re-enables both.
Once the C headers are included from C++ translation units,
bugprone-dynamic-static-initializers reports each extern variable
declaration in them; none is defined in a header, so disable the check.
Mark the two bounded memchr searches in ssurl.cpp, which
bugprone-not-null-terminated-result misreads as string copies.
Cast allocator results and the thread argument explicitly and
value-initialise JOBOBJECT_EXTENDED_LIMIT_INFORMATION with {}. Found by
the zig x86_64-windows-gnu cross build.
Max Lv