Taildrop: files sent to the console from the user's other devices are moved from Tailscale's holding area to /data/tailscale/received (a setting), announced on screen and listed on the status page with download links. Updates: the status page can install a newer release when asked to. A release carries tailscale.elf.sig, an Ed25519 signature over its version and SHA-256; the daemon installs only what verifies against the public key built into it, is the version the release claims and is newer than itself. The payload is handed to the ELF loader, and the copy named by the new payloadPath setting is replaced as well. - tsd/relsig, tsd/cmd/signrelease: signing, verifying, and keeping the key (keygen, passphrase-protected backup and restore). - tools/make-release.ps1 builds, signs and checksums a release. - Files are no longer copied to sockets with sendfile, which the PS5 kernel refuses.
14 KiB
Technical notes
How Tailscale, a Go program, runs on a PS5, and what was learned about the
console along the way. Everything here was measured on one console
(firmware 13.42, kern.osreldate 900000); treat it as observations, not a
specification.
Architecture
tailscale.elf is one payload made of two parts.
The launcher (launcher/, C, built with ps5-payload-sdk):
- The SDK's crt runs first. Among other things it patches the process so that system calls may be issued from any address. Without that, the kernel only accepts syscalls made from libkernel, and Go makes its own.
main.craises privileges, moves the process to the lowest scheduling priority (see Scheduling) and setsGOMAXPROCS=4.goload.cmaps the embedded Go program (a position-independent ELF), applies itsR_X86_64_RELATIVErelocations, makes its text executable, builds the argc/argv/envp/auxv block a FreeBSD kernel would pass, switches to a fresh 1 MB stack and jumps to the Go entry point. The embedded copy is then released withmadvise(MADV_FREE).
The launcher keeps a log, /data/tailscale/launcher.log (report.c): the
firmware version, each step that fails, and a last line before it jumps into
the Go program. A failure is also shown as a notification, because a payload
manager does not show what a payload prints. Just before the jump, stderr is
pointed at that log, so that a Go runtime that dies before the daemon has
opened its own log leaves its message there. What cannot be reported this
way is a failure in the SDK's crt, which runs before any of this.
The daemon (tsd/, Go): a tsnet server.
- Inbound: tsnet's fallback TCP handler pipes each tailnet connection to
127.0.0.1:<same port>. It dials the local port before accepting, so ports with no listener are refused properly. - Inbound UDP (
inboundudp.go): tsnet has no catch-all for UDP, so the ports inudpPortsare listened on withtsnet.Server.ListenPacketon the node's tailnet addresses and relayed to127.0.0.1. The default list is PS5 Remote Play's (9295, 9296, 9297, 9302); its service answers clients that arrive from loopback. Both UDP directions shareudprelay.go: one connection to the target per client address, dropped after two idle minutes. - Outbound: local forwards (
localforward.go) listen on localhost and relay TCP and UDP to a tailnet host throughtsnet.Server.Dial. UDP is relayed per client address with an idle timeout. Each Sunshine host is a preset of seven such forwards on the ports that host really uses, derived from Sunshine's port setting (HTTPS -5, HTTP +0, RTSP +21, video +9, control +10, audio +11, microphone +13). The ports cannot be remapped, because the host tells the Moonlight client which ports to use; several hosts can only coexist on 127.0.0.1 if their Sunshine ports differ. - A status page and JSON API on port 8090 (
web.go,settings.go), an optional HTTP proxy, PS5 notifications by writing a request to/dev/notification0. - Password (
auth.go): PBKDF2-SHA256 hash in the config, session cookie, one attempt per second. Requests from loopback are exempt. For that to be safe, connections for the status page that arrive over the tailnet are not piped to localhost like other ports but handed to the page's HTTP server directly, so it sees the tailnet address. - Settings are applied live where possible. The launcher needs one of them,
the priority, before any Go code runs, so the daemon leaves it in
/data/tailscale/priorityfor the next start. - Update notice (
update.go): the latest release tag from the GitHub API, twice a day, compared with the running version. A newer release is shown on the page and announced once on the console; the announced version is kept in/data/tailscale/update-notified. - Installing an update (
selfupdate.go,relsig/): on request only. The release'stailscale.elf.signames a version and a SHA-256 and carries an Ed25519 signature over both. The daemon checks the signature against the public key built into it, that the version is the release's and newer than its own, downloads the payload to/data/tailscale/update/, compares the hash, optionally replaces the copy named bypayloadPath(temporary file, then rename), and writes the payload to the ELF loader on 127.0.0.1:9021. The new instance stops the old one as with any payload sent again, and removes the download when it starts. The connection to the loader is kept open until the old process exits, because it is the new payload's standard output. - Taildrop (
taildrop.go): tsnet does not link Taildrop in; importingtailscale.com/feature/taildropdoes. Received files wait instate/files/<login>-uid-<n>/. The daemon long-polls for them, copies each toreceiveDirunder a name that does not exist yet, and deletes it from the holding area. - Who may connect (
access.go): withallowFromset toown, the TCP handler and the UDP relays ask Tailscale who the sender is (WhoIs) and serve only nodes of the same user as the console, from the console's own tailnet. Answers are kept for a minute per address. Anything that cannot be established is refused. - Key expiry (
keyexpiry.go): the date comes from the node's own status. The page warns from 14 days before, and the console shows a notification at 14, 3 and 1 days. - When a listener reports that it had to reopen its socket (the PS5's network was reconfigured), the daemon asks Tailscale to rebind and re-STUN instead of waiting for its interface polling. See Rest mode for the one time this has been seen.
- A payload that is sent again stops the running instance (through the status page, or failing that by the pid it recorded) and takes over.
The icon helper (appicon/, C) is a second, tiny payload embedded in the
launcher. The first time tailscale.elf runs, the launcher sends it to the
ELF loader on 127.0.0.1:9021, where it runs as a process of its own,
registers a home screen app whose param.json has a deeplinkUri to the
status page, reports the result and exits. The launcher then writes
/data/tailscale/icon-installed and never does it again. It is a separate
payload so that the system libraries it needs are never loaded into the
long-running daemon process, where their threads could receive signals meant
for the Go runtime.
The same helper removes the icon (sceAppInstUtilAppUnInstall). A payload
sent to the ELF loader gets no arguments, so the mode is a byte in the file
after the marker TSICON-MODE=. The launcher leaves a copy of the helper in
/data/tailscale/icon-helper.elf; for Uninstall the daemon flips that byte
and sends it to the loader (tsd/homeicon.go).
There is no installer. Nothing is copied anywhere and no payload autoloader is touched: the payload is run from wherever the user keeps it.
The PS5 as a Go target
Go is built for GOOS=freebsd GOARCH=amd64 with cgo off and
-buildmode=pie, from a Go 1.27.1 tree with
patches/go1.27.1-ps5.patch. What the patch changes and why:
| File | Change | Reason |
|---|---|---|
internal/platform/supported.go |
Allow internally linked PIE on freebsd/amd64. | The image must load at any address and no C toolchain is involved. |
runtime/defs_freebsd_amd64.go |
48 bytes of padding in ucontext before uc_mcontext. |
The PS5's signal context has extra fields there. Without it, nil-dereference panics and preemption read garbage. |
runtime/defs_freebsd_amd64.go, sys_freebsd_amd64.s |
kevent 363 with the 32 byte struct. |
The FreeBSD 12 kevent (560) does not exist. |
runtime/sys_freebsd_amd64.s, os_freebsd.go |
pipe2 built from pipe and fcntl. |
pipe2 (542) does not exist. |
runtime/os_freebsd.go |
CPU count from hw.ncpu. |
kern.smp.maxcpus does not exist. |
syscall/asm_unix_amd64.s, asm9_unix2_amd64.s, ps5_freebsd_amd64.go |
Syscall emulation layer. | See below. |
internal/routebsd/interface_freebsd.go |
Find the link address using if_data's own length field. |
Sony's if_data is 0xb0 bytes, so interface names came back empty. |
os/executable_sysctl.go |
os.Executable falls back to a name from argv[0]. |
The kernel has no path for a payload; tsnet treats the error as fatal. |
The syscall table
The kernel implements the FreeBSD 9 system call numbers. Every number from
532 up is a Sony syscall with an unrelated meaning, so nothing FreeBSD
added later exists: pipe2, accept4, ppoll, utimensat, futimens,
getrandom, the 64-bit inode fstat/fstatat/getdirentries/statfs
family, the extended kevent. Issuing those numbers would call something
else entirely.
Go's Syscall, Syscall6, RawSyscall, RawSyscall6 and Syscall9 are
patched to divert numbers >= 532 to ps5_freebsd_amd64.go, which rebuilds
them from FreeBSD 9 calls and converts the structures (old stat, dirent
and statfs layouts). Anything not emulated returns ENOSYS. Because
golang.org/x/sys/unix and internal/syscall/unix go through the same entry
points, they are covered too. socket and socketpair are also diverted, to
apply SOCK_NONBLOCK/SOCK_CLOEXEC with fcntl.
Other kernel behaviour
- Page size is 16384. The Go linker aligns segments to 4096; the loader applies protection per 16 KiB page.
- Working as on FreeBSD:
thr_new,_umtx_op,sigaction,sigaltstack,thr_kill, kqueue includingEVFILT_USER,mmapreservations withPROT_NONE(64 MB and 4 GB tried),MAP_FIXED,madvise(MADV_FREE),sysctlforhw.pagesize,hw.ncpu,kern.arandom,kern.boottime,NET_RT_IFLISTandNET_RT_DUMP,AF_ROUTEsockets. - Directory reads:
/data(nullfs) and/user(bfs) only accept a buffer of exactly 65536 bytes forgetdirentries. The emulation always reads 64 KiB blocks and hands the converted entries out across calls. - Unix domain sockets cannot be bound on
/data. - There is no
/etc/resolv.confand no CA bundle. Go's resolver falls back to127.0.0.1:53; the daemon importsx509roots/fallbackfor TLS roots. - A payload's stdin, stdout and stderr are the ELF loader's TCP connection.
Go kills a process whose write to fd 1 or 2 fails with
EPIPE, so the daemon moves that connection to another descriptor and points 1 and 2 at its log file. - The SDK's
kernel_mprotect()rewrites the protection of the whole kernel map entry that contains the address. The loader first splits the text range off with an ordinarymprotect(). sendfileon a socket fails with "socket is not connected". Go uses it whenever a file is copied straight to a TCP connection (io.Copy(conn, file),http.ServeContentwith a file), so the daemon hides the file behind a plain reader in those places.- When the network is reconfigured (connection settings changed, Wi-Fi to
Ethernet), listening sockets fail with errno 163, a Sony-specific code, and
do not recover. The daemon's listeners reopen themselves
(
tsd/listener.go).
Scheduling
This is the part to read before running any new test on a console.
Payload threads start as FIFO threads at priority 700 (rtprio_thread
reports class 10, priority 700; the range is 256 to 767 and lower runs
first). A FIFO thread that never blocks is never descheduled. A Go test with
a busy loop on each of the 16 logical cores froze the console completely:
the kernel still answered ping, but no user process ran, and only a
power-cycle recovered it.
rtprio_thread(RTP_SET) silently ignores priorities outside that range, so
the usual RTP_PRIO_NORMAL, 0 does nothing and reports success. These are
accepted and read back: class 3 (time-sharing) at 767, class 2 (round-robin)
at 767, class 10 (FIFO) at 767. New threads inherit the setting.
The launcher sets class 3, priority 767 before Go starts, reads it back, and
refuses to start if it did not take. With that, Go's signal-based preemption
works across cores (4 spinning goroutines, 5 garbage collections in about
350 ms). Test builds can add a watchdog thread that kills the process after
a fixed time (build-payload.ps1 -Watchdog).
The "high" priority setting uses class 2 (round-robin) at priority 700 instead: equal to games and system threads, but equal-priority round-robin threads take turns. With it, the same test passes (5 collections in about 300 ms) and the console stays responsive: the status page answered within 60 ms throughout while four goroutines spun.
Rest mode
Observed once, for a rest of about a minute on Ethernet. The process is not killed: it is frozen with the rest of the console and continues afterwards.
- Going to sleep, the network is taken down first. Every socket fails with errno 163 at the same moment: the status page listener, Tailscale's relay connection and its connection to the coordination server. The listener reopened at once, the daemon asked for a rebind, and Tailscale saw "all links down" and paused.
- Nothing is logged while the console sleeps, and it is not reachable on the tailnet.
- On waking, Tailscale's monitor noticed the jump in the clock, rebound its sockets, reconnected to its relay and had its endpoints back within about 300 ms. The status page, forwarded TCP ports and the Remote Play UDP ports answered through the tailnet address afterwards without anything being restarted.
A rest of hours has not been tried, nor one on Wi-Fi.
Home screen icon
/user/app/TSCL00001/sce_sys/param.json with applicationCategoryType 65536
and a deeplinkUri, plus icon0.png, registered with
sceAppInstUtilAppInstallTitleDir. The console opens the link in its
browser.
Linking libSceAppInstUtil alone leaves the payload stopped before it runs.
It needs -lSceIpmi -lSceAppInstUtil -lSceUserService -lSceSystemService, in
that order, as in the SDK's install_app sample.
Known problems
- Once, switching the console from Wi-Fi to Ethernet during a Moonlight stream through the forwards froze the console. At that moment two daemon instances were running because of a bug that has since been fixed; whether that, the Moonlight client or something else caused the freeze is not known.
- Once, a first login completed in the browser but the daemon's follow-up request was answered with "auth path not found". The daemon now requests a new link when it sees that error; the recovery path has not been observed in practice.
- Tailscale's sockets recovered after rest mode took the network down and brought it back. A change of interface (Wi-Fi to Ethernet or back) while the daemon runs has not been observed since the rebind request was added.