* chore(i18n): translate every missing string in all locales
Adds the 87 strings the bottom-screen Dashboard and in-game stats pages
introduced to all 32 locales, plus the 51 older settings strings Japanese
and the 50 Chinese locales were still missing. Every locale now has all
3456 translatable strings with the same placeholders as English.
Also translates 36 entries that were still English where the language
does not use the English word, and makes es-419 touch_app_id use the same
%1$s placeholder as English.
* chore(patches): drop comments from the lines downstream patches add
Removes comment lines and trailing comments from lines our patches add,
with hunk headers recounted. Lines that a later patch uses as context
and the upstream backport 0123 are left as they are.
The gamescope, FEX, proot and xodus series still apply to their pinned
sources, and the patched trees differ only in those comments.
test_hud_frames now finds the HUD writer by its first declaration.
* docs: credit Greenlight, Better xCloud, viture-v2, GameNative and DiRTTelemetryErrorFix
These were credited only in source headers and NOTICE files.
* chore: remove source comments and unused functions
Removes explanatory comments from first-party Kotlin, Java, C, C++,
Rust, TypeScript, Python, shell, Gradle, CMake and workflow files.
License and attribution headers stay, and vendored third-party trees,
generated files and license files are untouched. A shellcheck directive
is kept, and udevmon's fallthrough comments became fallthrough attributes.
Syntax trees and token streams match the previous sources apart from the
comments and Python docstrings. Module docstrings that argparse prints
are kept, and device-smoke.sh prints the same --help text from a heredoc.
Also removes 37 functions and 4 fields that nothing referenced.
* chore: remove comments from the Syncthing integration and the latest main sources
* docs: credit Syncthing and the syncthing-android fork
* chore(i18n): translate the Syncthing strings in all locales
The 107 strings the Syncthing integration added are now in all 32 locales.
sync_web_path_locked names the Syncthing tab by its label.
* docs: credit @xXJSONDeruloXx for the HUD metric readers
The CPU, GPU load, thermal and battery readers come from their
PerformanceHudView pull requests #777, #877, #991, #1406 and #1573.
The FusionHUD credit line is unchanged.
* docs: credit the HUD metric readers to @xXJSONDeruloXx's work
* fix: a read-only file on a USB drive no longer fails to close the first time it is read
On the RedMagic Astra 2, Android's NTFS USB drive returns EIO from close() for a file opened read-only when that first read updates its access time; the update itself succeeds and the descriptor is released. The x86 loader under FEX treats a failed close as fatal, so every Linux game newly installed or updated on such a drive failed to start once (libfreetype.so.6: cannot close file descriptor: Input/output error). The session library's close() now reports success for EIO on read-only descriptors; errors on descriptors opened for writing are still returned.
* Check the read-write close still reports EIO
* fix(desktop): video wallpapers no longer race ahead after DroidDeck returns from the background
* test: give the vrclient loader harness the clock syscall hook
* perf(fex): answer raw stat, readlink and statx syscalls through the proot fast path
FEX issues newfstatat, readlinkat and statx as raw syscall() calls, which skipped the libc wrappers the fast path hooks and fell through to ptrace. Route them through libblfastpath the same way raw openat already is, and keep the fast path's own /proc/self/exe lookup off the hooked syscall().
* test(fex): keep the raw lookup test passing on kernels without statx
A capture format or buffer change now closes the app's decoder so the
next packet reopens it with the new stride and height, and a capture
slot whose remap fails no longer keeps its old mapping.
Steam's arm64 streaming_client only decodes through a V4L2 stateful
decoder and libdrm dumb buffers, neither of which exist on Android, so
streams from a PC never showed a picture. The preload now offers it a
/dev/video-dec0 and a libdrm stand-in backed by dma-heap buffers, and
each frame is decoded by the app's MediaCodec (Qualcomm low-latency
mode) over the encoder socket, falling back to the FFmpeg in Steam's
runtime.
On a GPU without KGSL, Setup offers an experimental Mali bridge switch
(off by default; ignored on Adreno). With it on, the session starts the
bridge server, points the Vulkan loader at the bridge ICD, names Mali's
device as the stand-in render node, runs the bridge self-test into the
session log, and caps vkd3d-proton's feature level at 12_0. Flatpak
sandboxes get the bridge ICD and socket. The DRM preload falls back to
flock in a bridge session when kcmp is refused, and a clock/thermal log
is kept beside the bridge logs.
- Handlers installed with SA_RESETHAND are no longer wrapped, so a deferred one-shot signal can't fall through to the default action.
- Deferred signals are replayed with every signal blocked, so a signal landing mid-replay can't send an entry twice.
- Taking a deferral slot and holding a signal are safe when one signal interrupts another, and a held signal is released only at the nesting level that held it, so a handler's return can't block it again.
- A fifth distinct signal inside a critical section is now held pending with the thread's signal mask and released once the section ends, instead of running its handler while an object lock is held.
- A wait from a signal handler during another wait's sleep uses its own queue and, for a second device, a reference for that call only, so the outer wait keeps its queue and can't return the inner wait's result.
Wine asks for userfaultfd with the i386 syscall number on arm64, so it
never gets one and falls back to mprotect and a SIGSEGV per first write,
which under proot is a ptrace stop each time. libblsession hands it the
real userfaultfd and, where Android refuses the PAGEMAP_SCAN ioctl, answers
it from /proc/self/pagemap and UFFDIO_WRITEPROTECT.
* perf: FEX's in-root path lookups go through the fast path instead of proot
Android has no openat2, so libblsession walks RESOLVE_IN_ROOT lookups for
FEX one component at a time. Every fstatat, readlinkat and openat in that
walk was a raw syscall that stopped in proot. When libblfastpath is loaded,
the walk now calls its fstatat, readlinkat and openat, which answer without
proot and fall back to the raw syscall themselves. Without it the walk uses
the raw syscalls as before.
One lookup drops from 250-940 us to 26-86 us. A Graveyard Keeper launch
makes 37.5k proot stops instead of 54.0k and reaches its first scene
0.31 s sooner.
* test: count fstatat64 where the preload asks for it
* perf: FEX's own file opens go through the fast path instead of proot
FEX opens host files for the game with syscall(SYS_openat), so every one
stopped in proot even when the fast path could answer it. libblsession's
syscall() now sends SYS_openat to the fast path's openat as well, the same
way the in-root walk does, and makes the raw syscall when the fast path is
not loaded.
A Graveyard Keeper launch makes 28.0k proot stops instead of 37.5k; FEX's
openat stops go from 6,739 to none.
* fix: the fast path's start-up reads don't come back through syscall()
libblsession's syscall() now sends SYS_openat to the fast path. A process
named FEX starts the fast path by reading /proc/self/cmdline with
syscall(SYS_openat), which came straight back into the fast path and
waited on its own pthread_once. Its start-up reads now call libc's
syscall directly.
A process's status read for another process is served read-only, so
Chromium's ReadOnlyFile check no longer kills the browser. Chromium-based
AppImages without chrome-sandbox get --no-sandbox, and Flathub Chromium and
Electron apps skip zypak's spawn portal, which cannot work here.
* feat: run SteamOS PipeWire and WirePlumber for gamescope screen capture
* feat: hardware Remote Play encoding through MediaCodec behind Steam's V4L2 stage
* feat: deliver Remote Play touches to gamescope as touches
* feat: route game audio through PulseAudio while a Remote Play client is connected
* fix: keep the Remote Play marker when a forked child closes the encoder
* fix: allow standalone preload tests without V4L2 encoder hook
fix: let Wine write hooks into code FEX has translated
Wine 11's wineserver writes another process's memory with process_vm_writev,
which the kernel refuses on a page that is not writable. FEX's WoW64 layer
keeps the pages it has translated read-only to catch self-modifying code and
re-protects them right after a program unprotects them, so WriteProcessMemory
and NtWriteVirtualMemory into ntdll failed with a partial copy and Ubisoft
games stopped at "Error at hooking API NtProtectVirtualMemory". Wine only
falls back to /proc/<pid>/mem on ENOSYS or EPERM.
libblsession now completes a single-buffer write that wineserver was refused
with EFAULT through /proc/<pid>/mem, the way Wine wrote before Wine 11. Other
programs and every other error keep the kernel's answer. The iovecs are read
with process_vm_readv, so a bad pointer still returns EFAULT, and offsets are
64-bit in the i386 build.
Valve's FEX opens rootfs paths with openat2(RESOLVE_IN_ROOT), which the
preload emulates under proot. droiddeck-fex's rootfs links to the Steam
runtimes with absolute paths outside the rootfs, so every lookup missed
and FEX fell back to the arm64 host files: x86 bash, and with it Stardew
Valley's launcher, failed on libdl.so.2. An absolute link whose target
is missing in the root but present on the host now finishes there.
An opt-in second Steam client (Steam settings, Steam client: x86-64): Valve's
x86-64 Linux client, unchanged, runs under FEX in its own home with its own
game library and the same sign-in. Linux builds run directly, so titles such
as Team Fortress 2 and Half-Life use their Linux versions; Windows games use
DroidDeck-Proton as on arm64. The arm64 client stays the default.
- droiddeck-steam-x64: FEX and the x86-64 rootfs from DroidDeck-Components
(sha256-pinned), client install and sync from Valve's CDN, sign-in shared
with the arm64 client, its web helper on the GPU (ANGLE GL over FEX's GL
thunk, zink, Turnip), Proton wrappers for the arm64 client's tools,
DroidDeck direct for Linux builds, SD-card library (x86-64/ beside the
arm64 one), repair.
- FEX 2610 built by build-fex.yml with thunks and three patches: an openat2
fallback without openat2 (proot), newer-clang thunkgen, MAP_32BIT
exhaustion returning ENOMEM. libblsession resolves RESOLVE_IN_ROOT paths
inside the root.
- Session: x86-64 gating, shortcuts and compatibility tools per client,
logs from the client that ran.
- App: client setting with the FEX version, an x86-64 chip on the Games
header, the active client's library, art, storage, downloads, repair,
Offline Mode, Decky, desktop mode, optional VAC diagnostics; per-client
state for art, the content log, repair and leftovers.
- Strings in every locale.
- Review fixes: the library cache is saved only by the newest scan; an
abandoned session's Steam logs come from the client it recorded; distinct
VAC console names; FEXInterpreter in CPU pinning; clearer VAC hint in
pl, ru, uk, cs and bg.
install() replaces an existing cache entry by calling detach(old), which
destroys the listing when no reader holds it, and then reads old->wd to
decide whether to release the watch. When the old listing has no readers,
that read happens after the free.
Two threads that enumerate the same uncached directory at the same time
reach this path: both build a listing, the first installs an entry with
no readers, and the second replaces it.
Save the watch descriptor before detaching, as drop() already does.
* Add opt-in zRAM compression for the Steam client while a game runs
A "zRAM compression" toggle under Steam > Settings > Session, below
"Keep downloads running". Off by default, kept once set, and disabled
on devices without swap or with a kernel older than 5.4.
When on, the session watches for Steam's reaper. 20 s into a game the
client and its web helpers page their private anonymous memory out to
the device's zram with MADV_PAGEOUT from inside their own process
(libblsession zram.c, SIGURG), again every 5 minutes, and read every
swapped page back once the game exits. No root is needed: only a
process itself may do this without CAP_SYS_NICE. With the toggle off
nothing is installed or started.
RedMagic NP06J, Palworld: Steam's resident memory fell from ~1.7 GB to
~1.1 GB with ~560 MB in zram, available memory during play rose by
~0.5-0.8 GB, and ~2 MB was left in zram after the read-back.
* Turn zRAM compression on by default below 12 GB of RAM
Until the toggle is changed, it follows the device: on when the kernel
supports it and the device has less than 12 GB of RAM, read from the
advertised memory size on Android 14+ and from total memory before that.
Turning it off is kept.
---------
Co-authored-by: xXJsonDeruloXx <danielhimebauch@gmail.com>
* feat: add Cardboard mode with side-by-side lens-distorted eyes and 3DoF head tracking
* fix: derive the Cardboard field of view from the panel and recentre fully when far from upright
* feat: share the Cardboard head pose with the guest and add a SteamVR driver for it
* feat: run SteamVR's compositor on Turnip/KGSL with a display-extension layer and a driver-direct HMD
* feat: show SteamVR frames in Cardboard mode through a driver-side presenter and a fake direct-mode display
* feat: install and start SteamVR for Cardboard mode from the session drawer
* docs: describe Cardboard mode and the SteamVR path
* feat: drive SteamVR hands from the game pad and keep the Cardboard headset out of standby
* docs: record where the SteamVR dashboard stops
* fix: map Cardboard head tracking to the screen's axes and keep turns about the real vertical as yaw
* fix: give Cardboard its previous picture back when SteamVR stops
* fix: report sem_otime from emulated SysV semaphores so SteamVR's web helper can start
* fix: bridge SteamVR's shared semaphores across processes and keep Steam alive while SteamVR runs
* feat: present SteamVR straight to DroidDeck's compositor at panel resolution
* docs: update Cardboard VR notes for the dashboard, semaphore bridge and direct presenter
* fix: keep Wine and FEX signal handlers in the VR layer; add semaphore tracing
* fix: clear a stale SteamVR active marker on status
* fix: clear only an old SteamVR active marker
* fix: unblock SteamVR clients on Turnip and keep Steam's web helper alive across SteamVR runs
* fix: judge shared waits before a submit's own signals and on Vulkan 1.0 devices
* docs: record VR games, SteamVR Home and OpenXR results on the Thor
* docs: add Cardboard VR learnings from the Thor bring-up
* docs: add Cardboard VR screenshots from the Thor
* fix: keep the VR layer's display stand-in to SteamVR runs and clear a leftover marker at session start
* fix: report SteamVR start and stop failures from the helper's exit code
* Run programs, scripts and AppImages from the desktop
Windows programs open with the Proton Steam has chosen, and a Steam game's own
launch program is handed to Steam. x86-64 Linux programs and game scripts run
under FEX, and an exec the kernel refuses goes to droiddeck-open through
libblsession's binfmt.c. Programs on noexec shared storage pass KIO's checks; a
folder of Linux programs there, and sharun AppImages, are added by the app once
and run from the copy. Executable scripts run without KIO's prompt, mpv plays
on the software desktop, FEXServer's socket moves under XDG_RUNTIME_DIR, and
Proton runtimes no longer show up as games.
* Switch to Desktop from Big Picture without Deck mode
Big Picture's Switch to Desktop asks the client's SteamOS manager, which the
client only has when started as SteamOS (Deck mode). Without it the request
failed and the screen stayed on "Switching to Desktop..." for good. The live
compat helper now sees the request on its way to the client and hands it to
steamos-session-select, which opens the KDE desktop as Deck mode already did.
* Run x86 programs from shared storage where they lie
FEX asks for the guest's PROT_EXEC when it maps a program or library, which
the kernel refuses on noexec shared storage; FEX never runs those pages
itself, so libblsession maps them without it in FEX's processes. An x86
game, its libraries and its start script now run in place instead of being
copied in; only native code on noexec storage is still copied.
A Steam library's games are never copied, and a Linux game's launch program
opened from the desktop goes to Steam like a Windows one.
* Front end: draw the capsule wall without clipping it every frame
The tilted wall clipped each capsule to a rounded rect and recomposed its columns on every drift
step, and a rotated rounded clip is rasterised in software on the render threads: the idle main
page held about 1.7 cores. The art is now loaded at the capsule's size with its corners already
rounded, each column draws its capsules itself in one layer moved by the drift, and the drift
advances at most every 15 ms. The idle page takes about a third of a core.
* libfakeinput: let calls on descriptors it does not serve skip its lock
Every read, write, ioctl, close and poll of every process in a Proton session went through the
controller lock and the three map lookups, Wine's wineserver pipes and the game's files alike.
A bitmap of the descriptors the maps hold, set before one is handed out and cleared when it is
closed, lets everything else go straight to the kernel.
* Pad motion: ask the display for its rotation four times a second
The rotation was looked up for every sensor reading, every 2 ms, and cost more than the rest of
the reading; it is now refreshed every 250 ms, and the readings no longer allocate pairs.
* Session: keep the library passes out of a game's launch
The redist seeder stamped each prefix with the checksum of the whole set of names, which differs
between the session's pass and the game launch's (the game's own names are added), so every
launch re-appended the markers to every prefix in the library while the game loaded. Each
prefix's marker now lists the names it was given and only the missing ones are added.
The beat's seed and compatibility passes, and the Lossless Scaling ownership probe (which loads
steamclient.so), also wait while a wineserver runs: a launch rewrites the manifest and made them
due at exactly that moment. The game's own launch seeds its prefix regardless.
* Agent: keep the focus watcher's X connection open while nothing changes
The connect timeout stayed on the socket, so with no property change the read timed out, and the
watcher dropped the connection, reconnected and took a fresh snapshot every few seconds.
* Game environment: start Cyberpunk 2077 without REDlauncher
The prelauncher opens REDlauncher, a Qt WebEngine window that hands over to the game half a
minute later and then stays up beside it, well over a core busy for as long as the game runs; its
own --launcher-skip started the game twice. The game's executable is started directly instead.
* netif: answer Wine's network driver from a reading of the link
Mortal Kombat X has Wine's network driver enumerate every interface, address and route table
hundreds of times a second. Each enumeration read /etc/droiddeck-net several times, asked the
kernel for its interface dump again (a netlink socket bound under proot, refused each time) and
looked the real functions up with dlsym. The link file is now read at most once a second, a
refused dump is not asked for again within that second, and the real functions are looked up
once.
* Fast path: create memfds without a stop in proot
proot looks at a memfd's name only to refuse the few that are then executed or locked; every other
memfd_create is now issued from the trampoline. The session's /proc/net and status copies made
one per read, a stop each. Without memfd in the kernel the call still goes to proot's ashmem.
* Logs: run each redaction rule only on a line that could match it
A session's end copies some 70 MB of the Steam client's logs and puts every line through all
twenty-odd rules, most of a minute of a core on the phone after every session. Each rule now runs
only on a line holding what it cannot match without: its keyword, the five separators of a MAC,
the four dashes of a GUID, a long enough hex or token run, the learned account or address text.
Lines with non-ASCII text still meet every case-insensitive rule. Over 895,000 lines of real
client logs, with the device's accounts and addresses learned, the output is identical and the
pass takes a twentieth of the time.
* feat: Steam's touch controller as on-screen controls (beta)
* fix: Steam touch controls appear once the session makes the device; tidier default layout
* fix: Steam touch layout decoding, saved layouts and new sessions
* docs: Steam's touch controls
* feat: Steam's touch icons, labels and colours on the touch controls; pick and save them like Steam
* feat: Steam touch icon picker as in-session panels; layout colour; docs
* Give each physical controller its own player slot
Every controller fed the one exported pad, so two pads played as one.
The first controller keeps slot 0 (the Deck in a Steam session); further
ones get slots 1-3 as separate Xbox 360 pads, created on first input and
removed on disconnect. In Deck mode only the Steam client sees those
pads and hands them to games through Steam Input. Rumble follows the
slot it names.
* feat: Steam Link's touch menu, per-game input modes, action set editing, add/copy/paste, oval controls, paste control
* feat: Steam touch controls drawn as Steam Link draws them; Steam Link's touch menu and layout tray
* feat: Steam touch controls in DroidDeck's style, D-mark guide button, shake to hide
* refactor: Steam touch controls become an On-screen controls choice instead of a second dropdown
---------
Co-authored-by: maxjivi05 <maxjividen@gmail.com>
* perf: fast path resolves links, /proc and glue as proot does, and answers uname, isatty and passwd
The fast path handed back to proot whatever crossed a symlink, /proc or a directory proot
only shows as glue, which is most of what Steam and Wine ask. It now walks a path the way
proot's canonicalize() does, so those calls stay in-process:
- links are followed in guest terms and their text detranslated as detranslate_path() does,
including readlink into a buffer the kernel cut short and /proc's exe, cwd and fd links;
- a binding proot dropped at start is dropped here too, glue above a binding is ENOENT, and a
last component proot cannot stat is left to the kernel;
- pipe and socket dirfds are ENOTDIR, rmdir("dir/") removes dir, and "." and ".." go to proot;
- a blocking accept() with a connection already waiting no longer goes to proot.
Also answered: uname() and gethostname() (proot's utsname never changes), isatty() (glibc
2.43 asks with TCGETS2, which proot rewrites) and getpwnam/getpwuid when nsswitch.conf
sends passwd to files first and the entry is there (Xwayland asks for every client).
equiv.py, equiv2.py and equiv3.py match line for line with and without the fast path
(222, 507 and 924 lines). Per call on the RedMagic: realpath 240 us -> 1.0 us,
realpath /proc/self/fd/N 408 -> 9.9, getpwuid 41.5 -> 4.6, uname 28 -> 0.006.
* perf: LD_AUDIT module skips library lookups proot would fail
ld.so looks each library up along the game's search path, and Proton puts the game's
directory, Steam's runtime and video directories and its own library directories ahead of
/usr/lib: 72% of ld.so's opens in an Among Us launch failed, each a round trip to proot.
libblaudit.so answers la_objsearch() for a candidate proot would certainly fail with ENOENT
(its directory walked as canonicalize() walks it, the name or a directory on the way missing).
It has no libc and maps the fast path's trampoline itself. Each lookup's first candidate is
still opened, since a skipped one fails with ld.so's errno as it stands, and the default
directories are never skipped, so dlerror() reads the same. droiddeck-proton and the extra
wrapper add it to LD_AUDIT when the fast path is on.
* perf: droiddeck-proton checks a tool's components with one sha256sum
Each program the launcher starts pays a dozen traced library lookups under the game's search
path before it does anything. One sha256sum over every listed file replaces one per file;
the check itself is unchanged.
* perf: build a locale archive for C.UTF-8
Every program loads the locale LANG names before main(). glibc ships C.UTF-8 as a directory
it looks for once per category under two spellings, before reading gconv's module lists:
about 30 opens, each a stop in proot, in every process of the session. droiddeck-session
builds /usr/lib/locale/locale-archive (and gconv's cache) once per glibc package, so that is
one open.
localedef names a finished archive by hard link, which Android refuses; with
DROIDDECK_LINK_RENAME=1 libblsession turns a refused link into a rename that will not replace
an existing name.
* perf: library beat starts its helpers only when what they read changed
Every 15 s the session ran droiddeck-seed-redists (some forty programs under proot) and, when
a library directory had changed, steam-compatibility: a burst of thousands of proot stops each
quarter minute, game loads included. The seeder now runs when a library changed or a prefix
(or Proton's template for them) is unmarked, the compat pass when a manifest, depot, tool or
config.vdf changed, both checked from bash without starting anything, and each every five
minutes regardless. The seeder reads its stamp and the time without a subshell each.
* perf: fakeinput opens pad.log on its first line
Every process the library is loaded in opened the session's pad.log at start, a stop in proot
on shared storage, and most never write to it. The file is now opened when there is a line to
write.
* docs: the fast path as it ships, the audit module and measured launch times
* fix: a game the app launches after an idle spell comes to the front
The Steam client brings a game it launched forward only while it counts someone as present,
which it reads from gamescope's input counter. A game the app launched after four or more
minutes without a touch started, mapped its window, and stayed behind the library page until
the next input (device: Palworld, Stardew Valley, Fallout: New Vegas, Mortal Kombat X; the same
on main). The session now watches a requested launch, and when the game's window is focusable
but the interface keeps focus for two polls, counts one input on gamescope's counter, as a key
would. Nothing changes for a launch the client already brings forward.
* feat: replace the LXQt desktop with SteamOS-style KDE Plasma
* feat: leaving the KDE desktop logs out through Plasma and returns to gaming mode
* feat: DroidDeck logo wallpaper as the KDE desktop default
* feat: build the KDE wallpaper from the DroidDeck mark artwork
* ci: keep the built desktop package as a short-lived artifact for device testing
* fix: the desktop loading screen no longer promises Steam on the desktop
* docs: the output resize comments name the desktop's KWin
* fix: the desktop package verifies every package's Arch Linux ARM signature and never replaces a published tag
* fix: programs sent to the GPU from the desktop drop its software rendering overrides
* fix: an interrupted desktop install no longer counts as installed
* feat: pin the desktop package to its DroidDeck-Components release instead of the catalog
* fix: keep Steam's directory listings on FUSE libraries
Before it creates, moves or deletes a content file the Steam client lists
the whole destination directory about ten times to match names
case-insensitively. On Android's FUSE storage every listing goes through
the media provider, so reserving space grew quadratically with the number
of files: 33 minutes for Haunted Paws Demo (3,938 files) in a Games folder.
The client's listings under steamapps/ on FUSE are now kept while the
directory's stat is unchanged, for at most five seconds, and patched by
the client's own creates, renames and removals. Other changes fail the
stat check and the directory is read again. BL_NO_DIR_CACHE=1 turns it off.
Haunted Paws Demo into the internal shared Games folder: reserving space
33m08s -> 25-52s, committing 6m15s -> 19-30s, install 39m50s -> 1m12s-1m51s.
Internal storage is unchanged (5s reserve).
* fix: drop stale directory listings on coarse-timestamp volumes
NTFS reports whole-second directory times and a fixed size, so a change
made outside the client's own calls could leave a cached listing stale.
Listings now carry an inotify watch and are dropped on any event the
client's own patched changes don't account for. A forked child stops
caching, since it shares the parent's queue.
Covers rewinddir, telldir and seekdir on cached listings, and directories
opened by the fast-path preload's opendir, which goes through fdopendir.
Kurt Himebauch