Repository navigation
Open
Labels
area: netnet, net/http, http2, url, and the rest of the network stacknet, net/http, http2, url, and the rest of the network stackchecklistA tracking issue whose body is a checklist of what is leftA tracking issue whose body is a checklist of what is leftmilestoneA tracking issue, one per milestone in the planA tracking issue, one per milestone in the planportPorting a standard library package, or part of onePorting a standard library package, or part of one
Milestone
Description
Tracking issue for P4.
Goal: net, then net/http with HTTP/2. This is the headline, it is the reason most people will look at this project, and it is the most concurrency intensive code in the standard library, which means it is also where P0's scheduler finds out whether it was built correctly.
Gated on P3 for TLS. Small team, deep work, and not a phase that absorbs contributors the way P1 does.
Exit gate: the ordering and conformance tables in docs/design/11-packages-net.md sections 5 and 6, with h2spec passing in full.
net
-
IP,IPMask,IPNet,SplitHostPortandJoinHostPort(net: IP, IPMask and IPNet, with SplitHostPort and JoinHostPort #361) -
netFDand the file descriptor refcounting, which is the part that goes wrong under close during read (net: the poll FD, so a goroutine waits on a socket without a thread #378) - Deadlines, and the interaction between a deadline and a blocked goroutine (net: the poll FD, so a goroutine waits on a socket without a thread #378, with tests on each socket type in net: TCPConn, TCPListener and TCPAddr #383 to net: UnixConn, UnixListener and UnixAddr #385)
- The dual resolver:
getaddrinfothrough the platform layer by default, and Go's pure resolver ported alongside it, switched byGODEBUG=netdns=exactly as Go does (the resolv.conf reader, the hosts file andDNSErrorin net: resolver config, hosts cache and DNSError #391, the DNS client and the lookup functions in net: Resolver, the Lookup functions and the DNS client #395, and the cgo resolver in net: ask the system's resolver the way Go's cgo resolver does #414) - TCP, UDP, Unix domain sockets, raw sockets where the platform has them (TCP, UDP and Unix in net: TCPConn, TCPListener and TCPAddr #383, net: UDPConn and UDPAddr #384 and net: UnixConn, UnixListener and UnixAddr #385,
IPConnfor raw sockets in net: add IPConn #413, and the rest of the socket surface in net: add Dialer, ListenConfig and the Resolve functions #411 to net: add Dialer.DialIP and Multipath TCP #420) -
net.Listener,net.Conn, and the interfaces everything above stands on (net: Conn, Addr, Listener, OpError and Pipe #371)
net/url and net/textproto
These land in P1 because they are pure computation, and they are listed here because P4 is where they get their first real use.
- IDNA, checked against the reference test data rather than eyeballed (xnet/idna: port the IDNA package Go vendors #374)
-
textproto.Dial, the one declaration that had to wait for net, so net/textproto is at 100 percent (net/textproto: Dial #674)
net/http, server side
-
ServeMux, including the Go 1.22 pattern syntax, which is a small parser of its own (net/http: ServeMux patterns and routing #457) -
Serverlifecycle:Serve,Shutdown,Close, and the connection states in between (net/http: the HTTP/1 server #461), andServeTLSandListenAndServeTLS(net/http: ServeTLS and ListenAndServeTLS #599) - The
ResponseWritercontract, all of it, including the parts that are only documented by Go's tests (the server and its first batch of tests are in, net/http: the HTTP/1 server #461, net/http: the rest of the request tests #467 and net/http: more HTTP/2 server tests #475, and HTTP/2 server push withPusher, net/http: add server push for HTTP/2 #716, and the first 42 tests of clientserver_test.go with its harness, net/http: port clientserver_test.go up to TestStarRequestMethod #722 and net/http: port clientserver_test.go from TestTransportDiscardsUnneededConns #723; more of serve_test.go and clientserver_test.go is still to port) - Request parsing, and every request smuggling CVE Go has fixed pinned as a regression test (reading requests and responses, with the Transfer-Encoding and Content-Length checks, went in with net/http: read requests and responses #429, and writing them with net/http: write requests and responses #456; the smuggling tests from serve_test.go are in, net/http: the request smuggling regression tests from serve_test.go #606, and the go1.27.2 fixes are tracked in Port the security fixes from go1.27.2 and go1.27.3 #601)
-
http.FileServer,http.ServeContent, range requests (net/http: FileServer, ServeContent and ServeFile #459) - Cookies, redirects, basic auth, the middleware surface (cookies are in, net/http: add Cookie, ParseCookie and ParseSetCookie #423, basic auth went in with net/http: read requests and responses #429, redirects with the client in net/http: Client, Transport and httptest.Server #462, and the wrapping handlers with the server in net/http: the HTTP/1 server #461)
net/http, client side
-
Transportand its connection pool, which is where the interesting deadlocks live (net/http: Client, Transport and httptest.Server #462), TLS in the Transport (net/http: TLS in the Transport #602),Request.Cancel(net/http: add Request.Cancel #717), andTransport.NewClientConnwithClientConn(net/http: port ClientConn and Transport.NewClientConn #721) - Proxies, redirects, cookie jar (net/http/cookiejar: the cookie jar #460, net/http: Client, Transport and httptest.Server #462)
-
httptrace(net/http/httptrace: client trace hooks #463, with the two TLS handshake hooks in net/http: TLS in the Transport #602) -
httptest, because everything downstream tests against it (net/http/httptest: ResponseRecorder and NewRequest #458, net/http: Client, Transport and httptest.Server #462, and the TLS server in net/http/httptest: the TLS server #603)
HTTP/2
- The framing layer, HPACK, flow control, priority (HPACK is in, xnet/hpack: port golang.org/x/net/http2/hpack #386, the frame layer, net/http/internal/http2: port the frame layer #389, flow control, net/http: HTTP/2 flow control, dataBuffer and pipe #471, and the write schedulers, net/http: HTTP/2 frame writers and write schedulers #472)
- Server and client, both required rather than optional, because an HTTP library without h2 in 2026 is not an HTTP library (net/http: the HTTP/2 server #473, net/http: the HTTP/2 transport #474, with tests in net/http: more HTTP/2 server tests #475 and net/http: HTTP/2 transport tests against a real server #476)
-
h2specpassing in full, which is the gate
The rest of the tier
-
net/smtp,net/mail(net/mail is in, net/mail: parse mail messages, addresses and dates #372; net/smtp is in, net/smtp: the SMTP client #455, and StartTLS, net/smtp: StartTLS #604) -
net/rpc,net/rpc/jsonrpc(net/rpc, net/rpc/jsonrpc: RPC server, client and JSON codec #470) -
net/http/cgi,net/http/fcgi,net/http/httputilincluding the reverse proxy (net/http/httputil: dumps and chunked encoding #464, net/http/httputil: ReverseProxy #466, net/http/cgi: the cgi package #468, net/http/fcgi: the fcgi package #469, and the TLS state in cgi, net/http/cgi: the request's TLS state #605)
Security posture
-
BURROW_HARDENED=1and what it turns on, documented rather than implied - The header and body limits that Go applies by default, applied by default here too (docs: the default HTTP limits in the net guide #584)
Deferred to after 1.0
- HTTP/3 and QUIC, behind a build flag, tracked but not blocking this milestone
Metadata
Metadata
Assignees
Labels
area: netnet, net/http, http2, url, and the rest of the network stacknet, net/http, http2, url, and the rest of the network stackchecklistA tracking issue whose body is a checklist of what is leftA tracking issue whose body is a checklist of what is leftmilestoneA tracking issue, one per milestone in the planA tracking issue, one per milestone in the planportPorting a standard library package, or part of onePorting a standard library package, or part of one
Activity
tamnd commentedon Oct 6, 2026
#361 is merged, which adds the address types from Go's ip.go (
IP,IPMask,IPNet) along withSplitHostPortandJoinHostPort. It came early because crypto/x509 and crypto/tls need it. The ip_test.go port passes on server3 both natively and under ASan and UBSan. TestLookupWithIP is the only test left out, and it will come with the resolver.tamnd commentedon Oct 6, 2026
#369 fixes a runtime bug that the net.Pipe tests ran into. When a channel close woke several goroutines blocked in a select, the select cleanup could put another goroutine's waiter back on the closed channel's queue, and freeing the channel then threw. The new select test fails 20 times out of 20 without the fix and passes 20 out of 20 with it on server3. The net.Conn PR comes next, rebased on top of it.
tamnd commentedon Oct 6, 2026
#371 is in:
NetConn,NetAddr,NetListener, thenet.Errorchecks,OpError,ErrClosedandnet.Pipe. The pipe passes the nettest conformance suite ported from golang.org/x/net, and that suite is what turned up the select bug fixed in #369. TCP, UDP and Unix sockets come next, on top of the syscall socket calls and netpoll.tamnd commentedon Oct 6, 2026
#372 is in: net/mail, with all ten of Go's tests and the four examples, plus two tests of burrow's own for Header.AddressList and the ReadMessage errors. net/smtp is next on that line, once net has TCP to run it over.
tamnd commentedon Oct 6, 2026
IDNA is in with #374.
xnet/idnais a port of the copy of golang.org/x/net/idna that Go 1.27.1 vendors. Its tables are not written by hand:tools/gen-xnet-idna.shreads them out of Go's own package. The tests are Go's from x/net and x/text, including all of IdnaTestV2.txt, plus a differential test against Go and a sweep over every rune that only the long runs do. I ticked the IDNA box.tamnd commentedon Oct 7, 2026
Three more pieces of the HTTP stack are on main.
golang.org/x/net/dns/dnsmessagegolang.org/x/net/http2/hpackgolang.org/x/net/http/httpgutsThe structured field value parser (#388) had a UBSan failure in CI: an empty string with a NULL pointer took a zero offset in one helper. That's fixed on the branch, and it goes in once CI is green again. The HTTP/2 frame layer (#389) is stacked on it.
tamnd commentedon Oct 7, 2026
Merged #388 and #389. #388 ports httpsfv, the RFC 8941 structured field value parser that net/http uses for headers like Priority. #389 ports the HTTP/2 frame layer: the framer that reads and writes every frame type, with the size and stream checks Go's has. HPACK is already in, so flow control and priority are what's left before the HTTP/2 server and client can start. CI passed on all 33 jobs for both, including the big-endian s390x build.
tamnd commentedon Oct 7, 2026
Opened #391, stacked on the Unix socket PR: the pure resolver's groundwork. It has the resolv.conf reader, the cached hosts file lookups, name validation, the SRV and MX ordering, and a public
DNSError. No queries go out yet. Its 16 tests pass under gcc and under ASan and UBSan on Linux, and under wine.The wasip1 job on #384 failed because of an unused IPv6 probe helper that WASI never calls. The fix is on #383 and has been merged up through the stack.
tamnd commentedon Oct 7, 2026
#391 is merged, so the resolv.conf reader, the hosts file cache and
DNSErrorare on main. I ticked the netFD and deadline items, since the poll FD (#378) carries both and the TCP, UDP and Unix PRs test deadlines on each socket type.Still in progress: the DNS client that sends queries, and
Dial,Listen,DialerandListenConfigon top of it, with Happy Eyeballs for "tcp". Both are on branches and waiting for their test runs on Linux and under wine before the PRs go up.tamnd commentedon Oct 9, 2026
The rest of the
netpackage is up for review as a stack on top of #395:Dialer,ListenConfigand the Resolve functionsPacketConnandListenPacketIPConnInterfaces,InterfaceAddrs,ParseMACand the zone cacheListenMulticastUDPReadMsgandWriteMsgfor UDP, IP and UnixBuffers, writev, andTCPConn.ReadFromandWriteToFileConn,FileListener,FilePacketConnand theFilemethodsDialer.DialIPand Multipath TCP#396 (
log) and #397 (thenet/httpheader code) are the first two pieces ofnet/http, which the HTTP server needs. The rest of the HTTP stack follows once these are in.6 remaining items
tamnd commentedon Oct 14, 2026
net/smtp is in with #455: the client, PLAIN and CRAM-MD5 auth, and SendMail, at 25 of 27 items. StartTLS and TLSConnectionState wait for crypto/tls, and until then SendMail refuses a server that offers STARTTLS rather than send in the clear. 13 tests pass under plain, ASan and TSan builds on server3.
tamnd commentedon Oct 14, 2026
The net/http stack is merged, #456 to #476, 21 PRs in order: writing requests and responses, ServeMux, httptest, FileServer, the cookie jar, the HTTP/1 server and client, httptrace, httputil and ReverseProxy, cgi, fcgi, net/rpc, and the HTTP/2 flow control, schedulers, server and transport with their tests.
Coverage on main after the merge:
The stack tip was tested on server3 before the last merges from main, with ok 334 and fail 0 across the full suite, and the amalgamation build and doc check passed. Still open here: the rest of serve_test.go, the request smuggling regression set, h2spec, and the hardened mode.
tamnd commentedon Oct 21, 2026
#584 is merged. It adds a Limits section to docs/guides/net.md that lists the header and body limits Go applies by default, the default for each, and what changes it. They were already enforced in the same places Go enforces them, so this only documents them. I ticked the limits item.
tamnd commentedon Oct 21, 2026
#589 fixes a test-only leak in
http_pprof_test. When TestDeltaProfile skipped because the mutex profile missed the hog, the arena was never freed, and the asan+ubsan job failed on it. Cleanups now free the arena and put the profile rate back. In 40 ASan runs on server3 the test skipped 16 times, and none of the runs had a leak report.net/http: ServeTLS and ListenAndServeTLS (#599)
tamnd commentedon Oct 23, 2026
#599 is in: Server.ServeTLS, Server.ListenAndServeTLS and the package-level ServeTLS and ListenAndServeTLS, with HTTP/2 turned on by itself over TLS, the isBadCipher check, and the request's TLS state for handlers. It also fixed 64 functions in crypto/tls that left the error out-param unset on success, which is what made TestServeTLS hang at first. The Transport side of TLS and the httptest TLS server are next, then StartTLS for net/smtp.
tamnd commentedon Oct 23, 2026
#600, #602 and #603 are in. The Transport now does TLS on its own, with TLSClientConfig, TLSHandshakeTimeout, TLSNextProto and ForceAttemptHTTP2, and HTTP/2 comes on over TLS under Go's rules. httptrace has its two TLS handshake hooks, and httptest has NewTLSServer, StartTLS and Certificate, along with the transport tests that need a TLS server. #600 fixed a use-after-free that ASan found along the way: a retained certificate error now keeps its own copy of the certificates.
The security fixes from go1.27.2 and go1.27.3 are tracked in #601.
tamnd commentedon Oct 23, 2026
Merged today: #604 adds StartTLS to net/smtp, so SendMail now switches to TLS when the server offers it. #605 fills in the request's TLS state in net/http/cgi. #606 adds the six request smuggling regression tests from serve_test.go. #607 is the first of the go1.27.2 security fixes (#601), the limit on Range headers.
tamnd commentedon Oct 24, 2026
The hardened build is in with #644.
make HARDENED=1or-DBURROW_HARDENED=ONturns on the compiler's bounds and object-size checks in trap mode, and lowers the HTTP server's default header limits to 64 KiB and 100 values. docs/guides/building.md has a section on what it turns on, and CI has a job that builds and tests it. The one CI failure on the PR was dudect on arm, which is #579.tamnd commentedon Oct 25, 2026
#666 brings the transport and HTTP/2 server up to go1.27.3: only a body closed early is drained now, its trailer is discarded, and Request.TLS is set on any HTTP/2 request over TLS.
tamnd commentedon Oct 25, 2026
#674 added textproto.Dial, which was waiting for net. net/textproto is now at 64 of 64 and in the done list. The same PR fixes TestPipeline under -test.count, which kept the last run's state.
tamnd commentedon Oct 27, 2026
#716 adds HTTP/2 server push (
Pusher,PushOptions) and the optional ResponseWriter interfaces as typedefs, with Go's server_push_test.go tests passing under ASan+UBSan, TSan and MSan in CI. #718 and #719 fix the amalgamation and the lint job, which #711 and #713 had broken. #717, which addsRequest.Cancel, is waiting on its sanitizer jobs.tamnd commentedon Oct 27, 2026
#721 ports Go 1.27's
ClientConnandTransport.NewClientConnover HTTP/1 and HTTP/2, along with the six tests from clientconn_test.go in both modes. On a Linux box those tests and the transport, client and httptest tests passed plain, under ASan+UBSan, and under TSan. #717 added the deprecatedRequest.Cancel.tamnd commentedon Oct 27, 2026
#722 ports the clientserver_test.go harness and the 33 tests from the top of the file down to TestStarRequestMethod, each in h1 and h2 mode. On server3 the plain, ASan+UBSan and TSan builds all pass, with 69 PASS, 0 SKIP and 0 FAIL in verbose mode. Next is the rest of the file, starting at TestTransportDiscardsUnneededConns.
tamnd commentedon Oct 27, 2026
#723 ports the next nine tests of clientserver_test.go, from TestTransportDiscardsUnneededConns to TestBadResponseAfterReadingBody. All 100 tests and subtests pass on server3, plain and under ASan+UBSan and TSan, with no skips. Next is the rest of the file from TestWriteHeader0. TestClientServerTLSConnWrapper needs a small library change first: the server and transport have to take the TLS state from a conn that isn't a TlsConn.