Merge branch 'main' into feat/performance-fixes-2

.github/workflows/build.yml

@@ -211,9 +211,13 @@ jobs:
             rm mango-pkgs/pkg.tar.zst
           done
           install -m644 mango-pkgs/usr/bin/mangoapp "$M/mangoapp"
-          for so in libfmt.so.10 libspdlog.so.1.13 libglfw.so.3 libtraceevent.so.1 libtracefs.so.1; do
+          for so in libfmt.so.10 libspdlog.so.1.13 libglfw.so.3 libtraceevent.so.1; do
             cp -L "mango-pkgs/usr/lib/$so" "$M/$so"
           done
+          # Ours, not the package's: GPU memory without tracefs (tools/mangoapp/libtracefs-shim.c).
+          aarch64-linux-gnu-gcc -shared -fPIC -O2 -Wall -Wl,-soname,libtracefs.so.1 \
+            -o "$M/libtracefs.so.1" tools/mangoapp/libtracefs-shim.c
+          aarch64-linux-gnu-strip --strip-unneeded "$M/libtracefs.so.1"
           install -Dm644 tools/mangoapp/mangoapp "$D/usr/local/bin/mangoapp"
           rm -rf mango-pkgs
           ls -l "$M"

app/src/main/java/com/droiddeck/launcher/runtime/LinuxRuntime.java

@@ -231,12 +231,15 @@ public final class LinuxRuntime {
         };
         for (String[] entry : procFiles) {
             File fake = new File(fakeProc, entry[0]);
-            if (fake.isFile() && !new File(entry[1]).canRead()) {
+            // A live stand-in the session binds itself (CpuStatComponent's /proc/stat) wins.
+            boolean live = false;
+            if (extraBinds != null) for (String spec : extraBinds) live |= spec.endsWith(":" + entry[1]);
+            if (!live && fake.isFile() && !new File(entry[1]).canRead()) {
                 bind(cmd, fake.getPath() + ":" + entry[1]);
             }
         }
         bindGpuNode(context, cmd);
-        bindAdrenoStats(cmd);
+        bindAdrenoStats(cmd, extraBinds);
         bindCpuTemps(cmd, root);
         if (extraBinds != null) {
             for (String spec : extraBinds) bind(cmd, spec);
@@ -282,12 +285,71 @@ public final class LinuxRuntime {
             if (!Files.isSymbolicLink(subsystem.toPath())) {
                 Os.symlink("/sys/bus/platform", subsystem.getPath());
             }
+            // What MangoHud names a GPU's driver by; msm_drm is how an Adreno's render node reads
+            // on a mainline kernel, and the driver it reads the load of (bindAdrenoStats).
+            File driver = new File(device, "driver");
+            if (!Files.isSymbolicLink(driver.toPath())) {
+                Os.symlink("/sys/bus/platform/drivers/msm_drm", driver.getPath());
+            }
         } catch (IOException | ErrnoException e) {
             return;
         }
         bind(cmd, new File(base, "sys").getPath() + ":/sys/dev/char");
         bind(cmd, dri.getPath() + ":/dev/dri");
         bind(cmd, KGSL_DEVICE + ":/dev/dri/" + node);
+        bindDrmClass(base, cmd, node, major + ":" + minor);
+    }
+
+    /**
+     * MangoHud (Deck mode's performance overlay, mangoapp) finds the GPU by listing
+     * {@code /sys/class/drm}, and an exception from that listing ends the process: under an
+     * enforcing SELinux policy - every retail phone - an app may not list it, so mangoapp died on
+     * start, gamescope restarted it until the session gave up, and there was no overlay. Where the
+     * app cannot list it, the session's {@code /sys/class/drm} is ours: the render node above,
+     * named as an Adreno's is. MangoHud then reads the GPU's stats from KGSL's sysfs, which such a
+     * device refuses too - SessionService's GpuStatsComponent stands in for that. Where the
+     * listing is allowed it is left alone.
+     */
+    private static void bindDrmClass(File base, List<String> cmd, String node, String devNumbers) {
+        if (new File("/sys/class/drm").list() != null) return;
+        File drmClass = new File(base, "class");
+        try {
+            if (!drmClass.isDirectory() && !drmClass.mkdirs()) return;
+            // Only symlinks are ever made here; deleting them must not follow one into its target.
+            File[] old = drmClass.listFiles();
+            if (old != null) for (File f : old) Files.deleteIfExists(f.toPath());
+            Os.symlink("/sys/dev/char/" + devNumbers, new File(drmClass, node).getPath());
+        } catch (IOException | ErrnoException e) {
+            return;
+        }
+        bind(cmd, drmClass.getPath() + ":/sys/class/drm");
+        android.util.Log.i("LinuxRuntime", "hud: /sys/class/drm is not listable here; the session's lists " + node);
+        // Also refused to list, also asked for by MangoHud twice a second - an error line each
+        // time in the session log - and holding nothing it reads on an Adreno phone.
+        if (new File("/sys/class/powercap").list() == null) {
+            File empty = new File(base, "empty-powercap");
+            if (empty.isDirectory() || empty.mkdirs()) bind(cmd, empty.getPath() + ":/sys/class/powercap");
+        }
+    }
+
+    /** The temperature of a CPU zone the app may read (the first by name), or null. */
+    public static String cpuTempSource() {
+        for (java.util.Map.Entry<String, File> e : new java.util.TreeMap<>(thermalZones()).entrySet()) {
+            File temp = new File(e.getValue(), "temp");
+            if (e.getKey().contains("cpu") && !e.getKey().contains("gpu") && temp.canRead()) return temp.getPath();
+        }
+        return null;
+    }
+
+    /** The file the GPU's temperature reads from, or null where the app may read none. */
+    public static String gpuTempSource() {
+        String kgsl = "/sys/class/kgsl/kgsl-3d0/";
+        return firstReadable(kgsl + "temp", kgsl + "devfreq/temp", thermalZone("gpu"));
+    }
+
+    private static String gpuLoadSource() {
+        String kgsl = "/sys/class/kgsl/kgsl-3d0/";
+        return firstReadable(kgsl + "gpu_busy_percentage", kgsl + "devfreq/gpu_load");
     }
 
     /**
@@ -298,11 +360,11 @@ public final class LinuxRuntime {
      * Android PC emulators (GameNative, Winlator forks) read them in. A value found nowhere is left
      * alone.
      */
-    private static void bindAdrenoStats(List<String> cmd) {
+    private static void bindAdrenoStats(List<String> cmd, List<String> extraBinds) {
         String kgsl = "/sys/class/kgsl/kgsl-3d0/";
-        String gpuTemp = firstReadable(kgsl + "temp", kgsl + "devfreq/temp", thermalZone("gpu"));
+        String gpuTemp = gpuTempSource();
         String[][] stats = {
-                {firstReadable(kgsl + "gpu_busy_percentage", kgsl + "devfreq/gpu_load"), "/sys/kernel/debug/dri/0/perf_now"},
+                {gpuLoadSource(), "/sys/kernel/debug/dri/0/perf_now"},
                 {firstReadable(kgsl + "devfreq/cur_freq", kgsl + "gpuclk"),
                         "/sys/devices/platform/soc@0/3d00000.gpu/devfreq/3d00000.gpu/cur_freq"},
                 {gpuTemp, "/sys/class/thermal/thermal_zone28/temp"},
@@ -310,8 +372,21 @@ public final class LinuxRuntime {
                 {thermalZone("ddr"), "/sys/class/thermal/thermal_zone22/temp"},
         };
         for (String[] stat : stats) {
-            if (stat[0] != null) bind(cmd, stat[0] + ":" + stat[1]);
+            // A value the session feeds itself (GpuStatsComponent, where KGSL is refused) wins.
+            boolean fed = false;
+            if (extraBinds != null) for (String spec : extraBinds) fed |= spec.endsWith(":" + stat[1]);
+            if (stat[0] != null && !fed) bind(cmd, stat[0] + ":" + stat[1]);
         }
+        // Which of the overlay's GPU values this device lets the app read: a phone under an
+        // enforcing policy refuses some of them, and the overlay then has no line for that value.
+        android.util.Log.i("LinuxRuntime", "hud sources: gpu load " + orNone(stats[0][0]) + ", gpu clock "
+                + orNone(stats[1][0]) + ", gpu temp " + orNone(gpuTemp) + ", ddr temp " + orNone(stats[4][0])
+                + ", thermal zones " + thermalZones().size()
+                + ", kgsl gpubusy " + (new File(kgsl + "gpubusy").canRead() ? "readable" : "refused"));
+    }
+
+    private static String orNone(String path) {
+        return path != null ? path : "none";
     }
 
     /**

app/src/main/java/com/droiddeck/launcher/session/CpuStatComponent.kt

@@ -0,0 +1,110 @@
+package com.droiddeck.launcher.session
+
+import android.os.SystemClock
+import android.util.Log
+import com.droiddeck.launcher.core.SessionPart
+import java.io.File
+
+/**
+ * A `/proc/stat` that moves.
+ *
+ * Android has refused apps `/proc/stat` since 8.0, and the session's stand-in for it was one
+ * `cpu0` line of zeros - enough for glibc and Steam to start, but every CPU load in the session
+ * read 0 % (the performance overlay showed one core, idle). The kernel still lets apps read each
+ * core's idle residency (`cpuidle/stateN/time`, µs, the same label as the cpufreq files), which is
+ * how overlays on Android work out CPU load: a core's busy time is the time that passed less the
+ * time it spent idle. This writes `/proc/stat` from that, one line per core, refreshed twice a
+ * second. Counters only grow - a sample's idle is never more than the time that passed - so a
+ * reader taking the difference of two reads never sees a negative. Where `/proc/stat` is readable,
+ * or the idle times are not, nothing is done.
+ */
+class CpuStatComponent(val file: File) : SessionPart() {
+    @Volatile private var running = false
+    private var thread: Thread? = null
+    private val cores = (0 until coreCount()).toList()
+    private val busy = LongArray(cores.size)
+    private val idle = LongArray(cores.size)
+    private val lastIdleUs = LongArray(cores.size)
+    private var lastMs = 0L
+    private val bootSeconds = (System.currentTimeMillis() - SystemClock.elapsedRealtime()) / 1000
+
+    /** Writes the first sample; true when it should be bound over `/proc/stat`. */
+    fun prepare(): Boolean {
+        if (File("/proc/stat").canRead() || cores.isEmpty() || idleUs(0) == null) return false
+        lastMs = SystemClock.elapsedRealtime()
+        for (i in cores) lastIdleUs[i] = idleUs(i) ?: 0L
+        write()
+        Log.i(TAG, "hud: /proc/stat from the cores' idle times (${cores.size} cores)")
+        return true
+    }
+
+    override fun start() {
+        running = true
+        thread = Thread({
+            while (running) {
+                try { Thread.sleep(PERIOD_MS) } catch (e: InterruptedException) { break }
+                if (!running) break
+                sample()
+                write()
+            }
+        }, "cpu-stat").apply { isDaemon = true; start() }
+    }
+
+    override fun stop() {
+        running = false
+        thread?.interrupt()
+        thread = null
+    }
+
+    private fun sample() {
+        val now = SystemClock.elapsedRealtime()
+        val elapsed = (now - lastMs) / MS_PER_TICK
+        if (elapsed <= 0) return
+        lastMs += elapsed * MS_PER_TICK
+        for (i in cores) {
+            // An offline core has no idle residency to count and runs nothing: all idle.
+            val us = idleUs(i)
+            val idleTicks = if (us == null) elapsed else ((us - lastIdleUs[i]) / US_PER_TICK).coerceIn(0, elapsed)
+            if (us != null) lastIdleUs[i] = us
+            idle[i] += idleTicks
+            busy[i] += elapsed - idleTicks
+        }
+    }
+
+    /** The whole file at once, renamed into place: a reader opens either the last sample or this one. */
+    private fun write() {
+        val text = StringBuilder()
+        text.append("cpu  ${busy.sum()} 0 0 ${idle.sum()} 0 0 0 0 0 0\n")
+        for (i in cores) text.append("cpu$i ${busy[i]} 0 0 ${idle[i]} 0 0 0 0 0 0\n")
+        text.append("intr 0\nctxt 0\nbtime $bootSeconds\nprocesses 1\nprocs_running 1\nprocs_blocked 0\n")
+        try {
+            val tmp = File(file.parentFile, file.name + ".tmp")
+            tmp.writeText(text.toString())
+            if (!tmp.renameTo(file)) tmp.delete()
+        } catch (e: Exception) {
+        }
+    }
+
+    companion object {
+        private const val TAG = "SessionService"
+        private const val PERIOD_MS = 500L
+        /** /proc/stat counts in USER_HZ, 100 a second on every Linux the guest runs. */
+        private const val MS_PER_TICK = 10L
+        private const val US_PER_TICK = 10_000L
+
+        private fun coreCount(): Int =
+            File("/sys/devices/system/cpu").list()?.count { it.matches(Regex("cpu\\d+")) } ?: 0
+
+        /** The core's total idle residency in µs, or null when it has none to read (offline). */
+        private fun idleUs(core: Int): Long? {
+            val states = File("/sys/devices/system/cpu/cpu$core/cpuidle").listFiles { f -> f.name.startsWith("state") }
+                ?: return null
+            if (states.isEmpty()) return null
+            var total = 0L
+            for (state in states) {
+                total += runCatching { File(state, "time").readText().trim().toLong() }.getOrNull() ?: return null
+            }
+            return total
+        }
+    }
+}

app/src/main/java/com/droiddeck/launcher/session/GpuMemComponent.kt

@@ -0,0 +1,118 @@
+package com.droiddeck.launcher.session
+
+import android.util.Log
+import com.droiddeck.launcher.core.SessionPart
+import java.io.File
+
+/**
+ * The GPU memory in use, for the performance overlay's VRAM and PVRAM lines.
+ *
+ * Valve's mangoapp reads it from the kernel's gpu_mem_total tracepoint, which no Android app may
+ * reach, so both lines read 0. Our libtracefs for mangoapp (tools/mangoapp/libtracefs-shim.c) hands
+ * it what this writes to [file] - `/run/droiddeck-hud/gpu-mem` in the guest - every few seconds, in
+ * bytes, from the best of what KGSL lets the app read:
+ *
+ * - `/sys/class/kgsl/kgsl/page_alloc`, every GPU page on the device - what the tracepoint's total
+ *   counts, where the policy allows it;
+ * - else KGSL's own accounting of each of the session's processes
+ *   (`/sys/class/kgsl/kgsl/proc/<pid>/gpumem_mapped` and `gpumem_unmapped`);
+ * - else the size of the KGSL mappings in each of the session's processes' `maps` - only the
+ *   buffers the CPU has mapped, so less than the GPU holds (about 40 % of KGSL's own count in a
+ *   game on an AYN Thor), but always readable: every process the app can see in `/proc` is one of
+ *   its own. `maps` rather than `smaps`, which gives the same sizes but walks page tables to do it:
+ *   in a game, 180,000 lines and most of a second of CPU a pass.
+ *
+ * Sampled only while the overlay shows (OverlayShown).
+ */
+class GpuMemComponent(private val file: File, private val root: File) : SessionPart() {
+    @Volatile private var running = false
+    private var thread: Thread? = null
+    private var source = ""
+
+    /** True when there is a source to write from. */
+    fun prepare(): Boolean {
+        val bytes = sample() ?: return false
+        write(bytes)
+        Log.i(TAG, "hud: gpu memory from $source (${bytes / (1024 * 1024)} MiB now)")
+        return true
+    }
+
+    override fun start() {
+        running = true
+        thread = Thread({
+            while (running) {
+                try { Thread.sleep(PERIOD_MS) } catch (e: InterruptedException) { break }
+                if (running && OverlayShown.check(root)) sample()?.let { write(it) }
+            }
+        }, "gpu-mem").apply { isDaemon = true; start() }
+    }
+
+    override fun stop() {
+        running = false
+        thread?.interrupt()
+        thread = null
+        file.delete()
+    }
+
+    private fun sample(): Long? {
+        readLong("$KGSL_ROOT/page_alloc")?.let { source = "$KGSL_ROOT/page_alloc"; return it }
+        val pids = sessionPids()
+        var perProcess = 0L
+        var anyProcess = false
+        for (pid in pids) {
+            val mapped = readLong("$KGSL_ROOT/proc/$pid/gpumem_mapped") ?: continue
+            perProcess += mapped + (readLong("$KGSL_ROOT/proc/$pid/gpumem_unmapped") ?: 0L)
+            anyProcess = true
+        }
+        if (anyProcess) { source = "$KGSL_ROOT/proc/<pid>/gpumem_*"; return perProcess }
+        var mappedKb = 0L
+        var anyMap = false
+        for (pid in pids) {
+            val kb = kgslMappedKb(pid) ?: continue
+            mappedKb += kb
+            anyMap = true
+        }
+        if (anyMap) { source = "/proc/<pid>/maps (KGSL mappings)"; return mappedKb * 1024 }
+        return null
+    }
+
+    /** The processes this app may see in /proc: Android hides every other uid's. */
+    private fun sessionPids(): List<String> =
+        File("/proc").list()?.filter { it.isNotEmpty() && it.all(Char::isDigit) }.orEmpty()
+
+    /** Size of a process's mappings of the KGSL device, in kB; null where its maps cannot be read. */
+    private fun kgslMappedKb(pid: String): Long? = runCatching {
+        var bytes = 0L
+        File("/proc/$pid/maps").forEachLine { line ->
+            if (line.endsWith("kgsl-3d0") || line.contains("/dev/dri/renderD")) {
+                val dash = line.indexOf('-')
+                val space = line.indexOf(' ')
+                if (dash > 0 && space > dash) {
+                    bytes += line.substring(dash + 1, space).toLong(16) - line.substring(0, dash).toLong(16)
+                }
+            }
+        }
+        bytes / 1024
+    }.getOrNull()
+
+    /** Whole, renamed into place: the reader opens either the last value or this one. */
+    private fun write(bytes: Long) {
+        try {
+            file.parentFile?.mkdirs()
+            val tmp = File(file.parentFile, file.name + ".tmp")
+            tmp.writeText("$bytes\n")
+            if (!tmp.renameTo(file)) tmp.delete()
+        } catch (e: Exception) {
+        }
+    }
+
+    companion object {
+        private const val TAG = "SessionService"
+        private const val KGSL_ROOT = "/sys/class/kgsl/kgsl"
+        /** GPU memory moves slowly; a pass over every process's maps is ~40 ms of CPU in a game. */
+        private const val PERIOD_MS = 5_000L
+
+        private fun readLong(path: String): Long? =
+            runCatching { File(path).readText().trim().toLong() }.getOrNull()
+    }
+}

app/src/main/java/com/droiddeck/launcher/session/GpuStatsComponent.kt

@@ -0,0 +1,176 @@
+package com.droiddeck.launcher.session
+
+import android.util.Log
+import com.droiddeck.launcher.core.SessionPart
+import com.droiddeck.launcher.runtime.LinuxRuntime
+import java.io.File
+import java.io.RandomAccessFile
+
+/**
+ * The Adreno's stats as mangoapp (Deck mode's performance overlay) reads them.
+ *
+ * MangoHud takes an Adreno's load, clock and temperature from `gpu_busy_percentage`, `clock_mhz`
+ * and `temp` in `/sys/class/kgsl/kgsl-3d0`, and ends the process on one of them it may not read -
+ * not on one that is missing, which it skips. An enforcing SELinux policy (every retail phone)
+ * refuses apps some or all of KGSL's sysfs, each vendor differently, so mangoapp died on start and
+ * the session ran with no overlay. Where any of the three is refused, a directory of our own is
+ * bound there instead, each value fed twice a second from the first source this device lets the
+ * app read - the places other Android overlays (GameNative, WinNative, Bannerlator) read an
+ * Adreno's numbers from, across Snapdragon generations and vendor kernels. A value with no
+ * readable source is left out, and the overlay shows no line for it. Where all three are
+ * readable, nothing is done.
+ *
+ * Valve's build also reads load and clock where a mainline msm kernel has them (debugfs
+ * `perf_now`, the GPU's devfreq `cur_freq` in Hz) - the paths LinuxRuntime.bindAdrenoStats points
+ * at KGSL's own files when it may - so the same values are fed there too, in the same formats
+ * ([binds]).
+ */
+class GpuStatsComponent(val dir: File) : SessionPart() {
+    @Volatile private var running = false
+    private var thread: Thread? = null
+    private val feeds = ArrayList<Feed>()
+
+    /** One of MangoHud's files and where its value comes from. */
+    private class Feed(
+        val file: File, val source: String, val read: () -> Long?, val format: String, val percent: Boolean = false,
+    )
+
+    /** Where Valve's msm reader looks, and the feed for it, as host:guest binds. */
+    fun binds(): List<String> {
+        val list = mutableListOf(dir.path + ":" + KGSL)
+        if (File(dir, "perf_now").exists()) list.add(File(dir, "perf_now").path + ":" + MSM_LOAD)
+        if (File(dir, "cur_freq").exists()) list.add(File(dir, "cur_freq").path + ":" + MSM_CLOCK)
+        return list
+    }
+
+    /** Writes the directory; true when it should be bound over [KGSL]. */
+    fun prepare(): Boolean {
+        if (!File(KGSL).exists() || MANGOHUD_FILES.none { File(KGSL, it).let { f -> f.exists() && !f.canRead() } }) return false
+        return try {
+            dir.mkdirs()
+            dir.listFiles()?.forEach { java.nio.file.Files.deleteIfExists(it.toPath()) }
+            feeds.clear()
+            loadFeed()?.let { load ->
+                feeds.add(load)
+                // KGSL's gpu_busy_percentage reads "37 %"; padded after, so a shorter value
+                // written over a longer one leaves no digits behind.
+                feeds.add(Feed(File(dir, "perf_now"), load.source, load.read, "%-6s\n", percent = true))
+            }
+            pick("clock_mhz", clockSources(), "%5d\n") { mhz(it) }?.let { clock ->
+                feeds.add(clock)
+                feeds.add(Feed(File(dir, "cur_freq"), clock.source, { clock.read()?.times(1_000_000) }, "%-10d\n"))
+            }
+            pick("temp", tempSources(), "%6d\n") { milliCelsius(it) }?.let { feeds.add(it) }
+            feeds.forEach { feed -> feed.read()?.let { write(feed, it) } }
+            Log.i(TAG, "hud: kgsl stats from the session's own: " + MANGOHUD_FILES.joinToString { name ->
+                "$name " + (feeds.firstOrNull { it.file.name == name }?.source ?: "none")
+            })
+            true
+        } catch (e: Exception) {
+            Log.w(TAG, "hud: could not stand in for kgsl stats: $e")
+            false
+        }
+    }
+
+    override fun start() {
+        if (feeds.isEmpty()) return
+        running = true
+        thread = Thread({
+            while (running) {
+                try { Thread.sleep(PERIOD_MS) } catch (e: InterruptedException) { break }
+                if (!running) break
+                for (feed in feeds) feed.read()?.let { write(feed, it) }
+            }
+        }, "gpu-stats").apply { isDaemon = true; start() }
+    }
+
+    override fun stop() {
+        running = false
+        thread?.interrupt()
+        thread = null
+    }
+
+    /** Load: a file that states it in percent, else KGSL's busy and total of the last sample. */
+    private fun loadFeed(): Feed? {
+        pick("gpu_busy_percentage", LOAD_SOURCES, "%3d %%\n") { it.coerceIn(0, 100) }?.let { return it }
+        val gpubusy = "$KGSL/gpubusy"
+        val read = {
+            runCatching {
+                val parts = File(gpubusy).readText().trim().split(Regex("\\s+"))
+                val total = parts[1].toLong()
+                if (total > 0) (parts[0].toLong() * 100 / total).coerceIn(0, 100) else 0L
+            }.getOrNull()
+        }
+        return if (read() != null) Feed(File(dir, "gpu_busy_percentage"), gpubusy, read, "%3d %%\n") else null
+    }
+
+    /** The first of [sources] that reads as a number, its value put in MangoHud's unit by [unit]. */
+    private fun pick(name: String, sources: List<String>, format: String, unit: (Long) -> Long?): Feed? {
+        for (source in sources) {
+            val read = { firstNumber(source)?.let(unit) }
+            if (read() != null) return Feed(File(dir, name), source, read, format)
+        }
+        return null
+    }
+
+    /** In place and at one length, so a reader that keeps the file open sees each value whole. */
+    private fun write(feed: Feed, value: Long) {
+        try {
+            val text = if (feed.percent) String.format(feed.format, "$value %") else String.format(feed.format, value)
+            RandomAccessFile(feed.file, "rw").use { it.seek(0); it.write(text.toByteArray()) }
+        } catch (e: Exception) {
+        }
+    }
+
+    companion object {
+        private const val TAG = "SessionService"
+        const val KGSL = "/sys/class/kgsl/kgsl-3d0"
+        /** Must match LinuxRuntime.bindAdrenoStats. */
+        const val MSM_LOAD = "/sys/kernel/debug/dri/0/perf_now"
+        const val MSM_CLOCK = "/sys/devices/platform/soc@0/3d00000.gpu/devfreq/3d00000.gpu/cur_freq"
+        private const val PERIOD_MS = 500L
+        private val MANGOHUD_FILES = listOf("gpu_busy_percentage", "clock_mhz", "temp")
+        /** Qualcomm's own GPU summary, beside KGSL in vendor kernels, under the plain sysfs label. */
+        private const val QCOM_GPU = "/sys/kernel/gpu"
+        private val LOAD_SOURCES = listOf(
+            "$KGSL/gpu_busy_percentage", "$KGSL/devfreq/gpu_load", "$KGSL/gpuload", "$QCOM_GPU/gpu_busy",
+        )
+        /** The GPU's devfreq node is named by its register address, which differs between Snapdragons. */
+        private val DEVFREQ_NODES = listOf(
+            "kgsl-3d0", "3d00000.qcom,kgsl-3d0", "2c00000.qcom,kgsl-3d0", "5000000.qcom,kgsl-3d0", "5900000.qcom,kgsl-3d0",
+        )
+
+        private fun clockSources(): List<String> {
+            val sources = mutableListOf(
+                "$KGSL/clock_mhz", "$KGSL/gpu_clock", "$KGSL/gpuclk", "$KGSL/devfreq/cur_freq", "$QCOM_GPU/gpu_clock",
+            )
+            DEVFREQ_NODES.forEach { sources.add("/sys/class/devfreq/$it/cur_freq") }
+            File("/sys/class/devfreq").list()?.sorted()?.forEach { node ->
+                if (node.contains("kgsl") || node.contains("gpu")) sources.add("/sys/class/devfreq/$node/cur_freq")
+            }
+            return sources.distinct()
+        }
+
+        private fun tempSources(): List<String> =
+            listOfNotNull("$KGSL/temp", "$KGSL/devfreq/temp", "$QCOM_GPU/temp", LinuxRuntime.gpuTempSource()).distinct()
+
+        /** The first integer in the file ("550", "37 %", "550000000"), or null. */
+        private fun firstNumber(path: String): Long? =
+            runCatching { Regex("-?\\d+").find(File(path).readText())?.value?.toLong() }.getOrNull()
+
+        /** Hz, kHz or MHz to MHz, as the files differ; a stopped clock reads 0. */
+        private fun mhz(raw: Long): Long? = when {
+            raw < 0 -> null
+            raw > 10_000_000 -> raw / 1_000_000
+            raw > 10_000 -> raw / 1_000
+            else -> raw
+        }
+
+        /** °C or m°C to MangoHud's m°C; nothing a GPU reads at is below 1 °C or above 1000. */
+        private fun milliCelsius(raw: Long): Long? = when {
+            raw <= 0 -> null
+            raw < 1000 -> raw * 1000
+            else -> raw
+        }
+    }
+}

app/src/main/java/com/droiddeck/launcher/session/HwmonComponent.kt

@@ -0,0 +1,141 @@
+package com.droiddeck.launcher.session
+
+import android.os.SystemClock
+import android.system.Os
+import android.util.Log
+import com.droiddeck.launcher.core.SessionPart
+import com.droiddeck.launcher.runtime.LinuxRuntime
+import java.io.File
+import java.io.RandomAccessFile
+
+/**
+ * The session's `/sys/class/hwmon`, as mangoapp (Deck mode's performance overlay) looks there.
+ *
+ * MangoHud finds a CPU's temperature by an hwmon sensor named `cpuN_thermal`, and an Adreno's
+ * temperature, fan and power by one whose name has `gpuss` in it - a mainline Arm kernel's names.
+ * An Android kernel has neither (its hwmon holds the PMIC's sensors, or nothing), and where the
+ * app may not list hwmon at all the overlay logged an error twice a second. So the overlay showed
+ * no CPU temperature and no fan. This is two sensors named so:
+ *
+ * - `hwmon0` (`cpu0_thermal`): `temp1_input` linked to a CPU thermal zone the app can read;
+ * - `hwmon1` (`gpuss0_thermal`): `temp1_input` linked to the GPU's zone, and `fan1_input` with
+ *   the device's fan speed where it has a fan the app can read - an AYN handheld's tachometer
+ *   (`/sys/class/gpio5_pwm2/speed`), or a `fan*_input` in the device's own hwmon;
+ * - `hwmon2` (`slg4ax46073v`): the same fan, for Valve's own Qualcomm reader (msm.cpp), which
+ *   looks for its fan controller by that name and halves what it reads - that part reports twice
+ *   the real speed - so this one holds twice it.
+ *
+ * Nothing in a phone's own hwmon is read by the overlay, so standing in for it loses nothing.
+ */
+class HwmonComponent(val dir: File, private val root: File) : SessionPart() {
+    @Volatile private var running = false
+    private var thread: Thread? = null
+    private var fanSource: String? = null
+    private val fanFile = File(dir, "hwmon1/fan1_input")
+    private val valveFanFile = File(dir, "hwmon2/fan1_input")
+    private var lastRpm = 0L
+    private var lastRpmAt = 0L
+
+    /** Writes the sensors; true when the directory should be bound over `/sys/class/hwmon`. */
+    fun prepare(): Boolean = try {
+        dir.mkdirs()
+        clear(dir)
+        val cpu = LinuxRuntime.cpuTempSource()
+        if (cpu != null) sensor("hwmon0", "cpu0_thermal").also { link(cpu, File(it, "temp1_input")) }
+        val gpu = LinuxRuntime.gpuTempSource()
+        fanSource = FAN_SOURCES().firstOrNull { readNumber(it) != null }
+        if (gpu != null || fanSource != null) {
+            val sensor = sensor("hwmon1", "gpuss0_thermal")
+            if (gpu != null) link(gpu, File(sensor, "temp1_input"))
+            if (fanSource != null) {
+                sensor("hwmon2", "slg4ax46073v")
+                writeFan(sampleFan())
+            }
+        }
+        Log.i(TAG, "hud: hwmon cpu temp ${cpu ?: "none"}, gpu temp ${gpu ?: "none"}, fan ${fanSource ?: "none"}")
+        cpu != null || gpu != null || fanSource != null
+    } catch (e: Exception) {
+        Log.w(TAG, "hud: could not write the session's hwmon: $e")
+        false
+    }
+
+    override fun start() {
+        if (fanSource == null) return
+        running = true
+        thread = Thread({
+            while (running) {
+                try { Thread.sleep(FAN_PERIOD_MS) } catch (e: InterruptedException) { break }
+                // A tachometer read busy-waits 100 ms in the kernel: only while the overlay shows.
+                if (running && OverlayShown.check(root)) writeFan(sampleFan())
+            }
+        }, "hwmon-fan").apply { isDaemon = true; start() }
+    }
+
+    override fun stop() {
+        running = false
+        thread?.interrupt()
+        thread = null
+    }
+
+    /**
+     * The fan's speed in RPM. An AYN tachometer counts pulses over a short window per read and
+     * often catches none, reading 0 while the fan turns; a reading is held until a newer one
+     * comes or [FAN_HOLD_MS] passes with only zeros, when the fan is taken to have stopped.
+     */
+    private fun sampleFan(): Long {
+        val now = SystemClock.elapsedRealtime()
+        val rpm = fanSource?.let { readNumber(it) } ?: 0L
+        if (rpm > 0) {
+            lastRpm = rpm
+            lastRpmAt = now
+        } else if (now - lastRpmAt > FAN_HOLD_MS) {
+            lastRpm = 0
+        }
+        return lastRpm
+    }
+
+    /** In place and at one length, so a reader that keeps the file open sees each value whole. */
+    private fun writeFan(rpm: Long) {
+        for ((file, value) in listOf(fanFile to rpm, valveFanFile to rpm * 2)) {
+            try {
+                RandomAccessFile(file, "rw").use { it.seek(0); it.write(String.format("%6d\n", value).toByteArray()) }
+            } catch (e: Exception) {
+            }
+        }
+    }
+
+    private fun sensor(name: String, label: String): File =
+        File(dir, name).apply { mkdirs(); File(this, "name").writeText("$label\n") }
+
+    private fun link(target: String, link: File) {
+        java.nio.file.Files.deleteIfExists(link.toPath())
+        Os.symlink(target, link.path)
+    }
+
+    /** Only files and links are made here; deleting a link must not follow it into sysfs. */
+    private fun clear(root: File) {
+        root.listFiles()?.forEach { f ->
+            if (!java.nio.file.Files.isSymbolicLink(f.toPath()) && f.isDirectory) clear(f)
+            java.nio.file.Files.deleteIfExists(f.toPath())
+        }
+    }
+
+    companion object {
+        private const val TAG = "SessionService"
+        /** The AYN tachometer reads 0 if read again within a couple of seconds. */
+        private const val FAN_PERIOD_MS = 3_000L
+        private const val FAN_HOLD_MS = 10_000L
+
+        private val FAN_SOURCES = {
+            val sources = mutableListOf("/sys/class/gpio5_pwm2/speed", "/sys/class/gpio5_pwm/speed")
+            File("/sys/class/hwmon").listFiles()?.sortedBy { it.name }?.forEach { hwmon ->
+                hwmon.listFiles { f -> f.name.matches(Regex("fan\\d+_input")) }?.sortedBy { it.name }
+                    ?.forEach { sources.add(it.path) }
+            }
+            sources
+        }
+
+        private fun readNumber(path: String): Long? =
+            runCatching { File(path).readText().trim().toLong() }.getOrNull()?.takeIf { it >= 0 }
+    }
+}

app/src/main/java/com/droiddeck/launcher/session/OverlayShown.kt

@@ -0,0 +1,17 @@
+package com.droiddeck.launcher.session
+
+import java.io.File
+
+/**
+ * Whether Deck mode's performance overlay is on screen. The client sets its level by rewriting
+ * mangoapp's config (bannerlator-session makes it, `/tmp/mangohud.*`), and `no_display` is the
+ * level that hides it. The feeds whose reads cost something - the fan's tachometer busy-waits in
+ * the kernel, GPU memory walks every process's mappings - only run while it shows.
+ */
+object OverlayShown {
+    fun check(root: File): Boolean {
+        val config = File(root, "tmp").listFiles { f -> f.name.startsWith("mangohud.") }
+            ?.maxByOrNull { it.lastModified() } ?: return false
+        return runCatching { !config.readText().contains("no_display") }.getOrDefault(false)
+    }
+}

app/src/main/java/com/droiddeck/launcher/session/SessionService.kt

@@ -710,6 +710,33 @@ class SessionService : Service() {
         battery.attach(this)
         components.add(battery)
         binds.add(battery.dir.path + ":/sys/class/power_supply")
+        // The overlay's CPU and GPU temperatures and the fan, as hwmon sensors it knows by name.
+        val hwmon = HwmonComponent(File(filesDir, "session/sys/hwmon"), LinuxRuntime.rootDir(this))
+        if (hwmon.prepare()) {
+            hwmon.attach(this)
+            components.add(hwmon)
+            binds.add(hwmon.dir.path + ":/sys/class/hwmon")
+        }
+        // CPU load for everything in the session that reads /proc/stat, the overlay among them.
+        val cpuStat = CpuStatComponent(File(filesDir, "session/proc-stat"))
+        if (cpuStat.prepare()) {
+            cpuStat.attach(this)
+            components.add(cpuStat)
+            binds.add(cpuStat.file.path + ":/proc/stat")
+        }
+        // The GPU memory in use for the overlay's VRAM lines, which it would read from tracefs.
+        val gpuMem = GpuMemComponent(File(LinuxRuntime.rootDir(this), "run/droiddeck-hud/gpu-mem"), LinuxRuntime.rootDir(this))
+        if (gpuMem.prepare()) {
+            gpuMem.attach(this)
+            components.add(gpuMem)
+        }
+        // The GPU's load and temperature for the performance overlay, where KGSL's sysfs is refused.
+        val gpuStats = GpuStatsComponent(File(filesDir, "session/sys/kgsl-3d0"))
+        if (gpuStats.prepare()) {
+            gpuStats.attach(this)
+            components.add(gpuStats)
+            binds.addAll(gpuStats.binds())
+        }
         // Rumble for the on-screen pad: the fake evdev layer sends force-feedback effects to this
         // listener, which drives the phone's vibrator (see RumbleComponent).
         if (controllersOn) components.add(RumbleComponent().also { it.attach(this) })

tools/build_local.sh

@@ -224,9 +224,13 @@ while read -r package_sha256 package_url; do
     zstd -dc "${package_archive}" | tar -xf - -C "${mango_pkgs}"
 done < <(grep -v '^#' "${repo_root}/tools/mangoapp/packages.txt")
 install -m644 "${mango_pkgs}/usr/bin/mangoapp" "${mango_dir}/mangoapp"
-for library in libfmt.so.10 libspdlog.so.1.13 libglfw.so.3 libtraceevent.so.1 libtracefs.so.1; do
+for library in libfmt.so.10 libspdlog.so.1.13 libglfw.so.3 libtraceevent.so.1; do
     cp -L "${mango_pkgs}/usr/lib/${library}" "${mango_dir}/${library}"
 done
+# Ours, not the package's: GPU memory without tracefs (tools/mangoapp/libtracefs-shim.c).
+docker run --rm --platform linux/amd64 --user "$(id -u):$(id -g)" -v "${repo_root}:/src" -w /src "${image_name}" \
+    aarch64-linux-gnu-gcc -shared -fPIC -O2 -Wall -Wl,-soname,libtracefs.so.1 \
+    -o app/src/main/assets/linuxfs/usr/local/lib/mangoapp/libtracefs.so.1 tools/mangoapp/libtracefs-shim.c
 mkdir -p "${linuxfs_dir}/usr/local/bin"
 install -m644 "${repo_root}/tools/mangoapp/mangoapp" "${linuxfs_dir}/usr/local/bin/mangoapp"
 

tools/mangoapp/libtracefs-shim.c

@@ -0,0 +1,144 @@
+// libtracefs.so.1 for mangoapp: the GPU memory Valve's overlay reads, without the kernel's tracing.
+//
+// Valve's mangoapp (msm.cpp, ftrace_gpumem.c) takes the GPU's memory in use - the VRAM and PVRAM
+// lines - from the gpu_mem_total tracepoint (pid 0, the device's total) through libtracefs. An
+// Android app may not use tracefs at all: the instance could not be made, and both lines read 0.
+// This stands in for the libtracefs calls mangoapp makes, and hands its record walker one
+// gpu_mem_total record carrying what the app writes to /run/droiddeck-hud/gpu-mem (bytes; the
+// session's GpuMemComponent, from KGSL's own accounting). Only mangoapp loads this copy.
+//
+// The walker reads libtraceevent's structures directly, so the ones below are libtraceevent
+// 1.8.2's (event-parse.h), the version tools/mangoapp/packages.txt pins; tep_find_event_by_record
+// and tep_read_number are answered here too, ahead of the real libtraceevent in mangoapp's lookup.
+
+#include <sched.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define EXPORT __attribute__((visibility("default")))
+
+struct tep_handle;
+struct tep_event;
+struct tep_print_arg;
+struct tep_print_parse;
+
+struct tep_format_field {
+    struct tep_format_field *next;
+    struct tep_event *event;
+    char *type;
+    char *name;
+    char *alias;
+    int offset;
+    int size;
+    unsigned int arraylen;
+    unsigned int elementsize;
+    unsigned long flags;
+};
+
+struct tep_format {
+    int nr_common;
+    int nr_fields;
+    struct tep_format_field *common_fields;
+    struct tep_format_field *fields;
+};
+
+struct tep_print_fmt {
+    char *format;
+    struct tep_print_arg *args;
+    struct tep_print_parse *print_cache;
+};
+
+struct tep_event {
+    struct tep_handle *tep;
+    char *name;
+    int id;
+    int flags;
+    struct tep_format format;
+    struct tep_print_fmt print_fmt;
+    char *system;
+    void *handler;
+    void *context;
+};
+
+struct tep_record {
+    unsigned long long ts;
+    unsigned long long offset;
+    long long missed_events;
+    int record_size;
+    int size;
+    void *data;
+    int cpu;
+    int ref_count;
+    int locked;
+    void *priv;
+};
+
+// gpu_mem_total's record: gpu_id, pid, size - as the kernel lays it out after the common fields.
+struct gpu_mem_total {
+    unsigned int gpu_id;
+    unsigned int pid;
+    unsigned long long size;
+};
+
+static struct tep_format_field size_field = {
+    .name = "size", .offset = offsetof(struct gpu_mem_total, size), .size = 8};
+static struct tep_format_field pid_field = {
+    .next = &size_field, .name = "pid", .offset = offsetof(struct gpu_mem_total, pid), .size = 4};
+static struct tep_event gpu_mem_event = {
+    .name = "gpu_mem_total", .system = "gpu_mem", .format = {.nr_fields = 2, .fields = &pid_field}};
+
+// Never dereferenced by mangoapp: they only have to be non-null.
+static char handle;
+static char instance;
+
+static unsigned long long gpu_mem_bytes(void) {
+    unsigned long long bytes = 0;
+    FILE *f = fopen("/run/droiddeck-hud/gpu-mem", "r");
+    if (f) {
+        if (fscanf(f, "%llu", &bytes) != 1) bytes = 0;
+        fclose(f);
+    }
+    return bytes;
+}
+
+EXPORT void *tracefs_instance_create(const char *name) { (void)name; return &instance; }
+EXPORT char *tracefs_instance_get_dir(void *inst) { (void)inst; return strdup("/run/droiddeck-hud"); }
+EXPORT void tracefs_put_tracing_file(char *name) { free(name); }
+EXPORT struct tep_handle *tracefs_local_events(const char *dir) { (void)dir; return (struct tep_handle *)&handle; }
+EXPORT int tracefs_event_enable(void *inst, const char *system, const char *event) { (void)inst; (void)system; (void)event; return 0; }
+EXPORT int tracefs_event_disable(void *inst, const char *system, const char *event) { (void)inst; (void)system; (void)event; return 0; }
+EXPORT bool tracefs_instance_is_new(void *inst) { (void)inst; return false; }
+EXPORT int tracefs_instance_destroy(void *inst) { (void)inst; return 0; }
+EXPORT void tracefs_instance_free(void *inst) { (void)inst; }
+
+EXPORT int tracefs_iterate_raw_events(struct tep_handle *tep, void *inst, cpu_set_t *cpus, int cpu_size,
+                                      int (*callback)(struct tep_event *, struct tep_record *, int, void *),
+                                      void *context) {
+    (void)tep; (void)inst; (void)cpus; (void)cpu_size;
+    if (!callback) return -1;
+    struct gpu_mem_total data = {.gpu_id = 0, .pid = 0, .size = gpu_mem_bytes()};
+    struct tep_record record = {.size = sizeof(data), .record_size = sizeof(data), .data = &data};
+    callback(&gpu_mem_event, &record, 0, context);
+    return 0;
+}
+
+EXPORT struct tep_event *tep_find_event_by_record(struct tep_handle *tep, struct tep_record *record) {
+    (void)tep; (void)record;
+    return &gpu_mem_event;
+}
+
+EXPORT unsigned long long tep_read_number(struct tep_handle *tep, const void *ptr, int size) {
+    (void)tep;
+    switch (size) {
+    case 1: return *(const unsigned char *)ptr;
+    case 2: return *(const unsigned short *)ptr;
+    case 4: return *(const unsigned int *)ptr;
+    case 8: return *(const unsigned long long *)ptr;
+    default: return 0;
+    }
+}
+
+EXPORT void tep_free(struct tep_handle *tep) { (void)tep; }

tools/mangoapp/packages.txt

@@ -1,4 +1,4 @@
-# Valve's aarch64 mangoapp (the Deck performance overlay) and the five libraries it needs that the
+# Valve's aarch64 mangoapp (the Deck performance overlay) and the libraries it needs that the
 # runtime does not have, at the versions it was built against (its .BUILDINFO: fmt 10.2.0,
 # spdlog 1.13.0). CI fetches each, checks the sha256 and stages the files named below into the apk
 # under usr/local/lib/mangoapp; SessionFiles copies them into the runtime at every launch.
@@ -7,5 +7,6 @@
 ef7411229ea4e9b3ab4e36ac356a85b5072f362ee1de68b28d09c43f0b0ecc44  https://steamdeck-packages.steamos.cloud/archlinux-deckard/archlinux/mash-20240705.2/extra/os/aarch64/fmt-10.2.0-1-aarch64.pkg.tar.zst
 c9bfd2c3f15071329d381b9c2061d07503c4787ceb7001b9dfbf72a23c85cb04  https://steamdeck-packages.steamos.cloud/archlinux-deckard/archlinux/mash-20240705.2/extra/os/aarch64/spdlog-1.13.0-1-aarch64.pkg.tar.zst
 f9fae6c84a0382188f1cb20761424051593fd09b1456191cc250bd16e0a1d7a5  https://steamdeck-packages.steamos.cloud/archlinux-deckard/archlinux/mash-20240705.2/extra/os/aarch64/glfw-3.4-2-aarch64.pkg.tar.zst
+# libtracefs is ours (libtracefs-shim.c), and reads libtraceevent's structures as 1.8.2 lays them
+# out: moving libtraceevent off 1.8.2 means checking event-parse.h against the shim.
 0cccdc9c92cb3cae1e37ebeb0ed20b93bb3b5000c8f955b63d0e07bd878c7018  https://steamdeck-packages.steamos.cloud/archlinux-deckard/archlinux/mash-20240705.2/extra/os/aarch64/libtraceevent-1%3A1.8.2-2-aarch64.pkg.tar.zst
-47ffd6a953959d6a0df44539bef30a67d0b8c95cce0053273908d9d70ef7abc4  https://steamdeck-packages.steamos.cloud/archlinux-deckard/archlinux/mash-20240705.2/extra/os/aarch64/libtracefs-1.8.0-1-aarch64.pkg.tar.zst