mirror of
https://github.com/Droid-Deck/DroidDeck.git
synced 2026-11-18 20:36:04 +00:00
fix(hud): GPU clock and load without root on Adreno 8xx and hardware-scheduled kernels (#1981)
On phones where Android blocks the GPU's sysfs files for apps, the HUD
reads the GPU without root (#1969). On Adreno 8xx that showed no GPU
clock and only the partial load from Vulkan timing (OnePlus 13, Adreno
830, reported by iamTooShaar).
- GPU clock: the clock job used the Adreno 6xx/7xx registers. The
register layout is now picked from the KGSL chip id: on 8xx
CP_ALWAYS_ON_COUNTER is 0x8e7 and the RBBM cycle countable is 1.
Unknown later generations read nothing.
- GPU load: hardware-scheduled kgsl kernels (every 8xx, and some 7xx
such as the AYANEO Pocket FIT's) never write the memstore ringbuffer
timestamps. When those stay 0, the reader samples the contexts' own
memstore entries (every context that has submitted work, up to 64,
contexts with new or running work first).
- HUD diagnostics report: a "GPU readers without root" line with the
reason each reader works or fails, "none" instead of "null", and the
per-core CPU % source named honestly.
Tested: iamTooShaar's OnePlus 13 shows GPU clock and load with the PR
test build (06be2c5). Pocket FIT standalone build of the final readers
against the kernel's sysfs numbers: idle 1-2 % vs 3-4 %, GPU job loop
25-29 % vs 28-30 %, clock 1050 MHz vs 1050 MHz. The review fixes after
the test build were not tested on the OnePlus 13.
This commit is contained in:
1 parent
313b6a59d2
commit
55da0afd06
13 files changed
+299
-29
No files matched your search
@@ -1,5 +1,6 @@
|
||||
// Whole-GPU busy share without root on Adreno (KGSL) devices whose kgsl sysfs is blocked for apps: a thread
|
||||
// samples the memstore ringbuffer timestamps (kgsl_busy.h) once a millisecond while running.
|
||||
// samples the memstore timestamps (kgsl_busy.h) once a millisecond while running. Ringbuffer entries are used where
|
||||
// the kernel writes them; on hardware-scheduled kernels, which leave them at 0, the contexts' own entries are used.
|
||||
#include <errno.h>
|
||||
#include <fcntl.h>
|
||||
#include <pthread.h>
|
||||
@@ -15,13 +16,21 @@
|
||||
#include "kgsl_busy.h"
|
||||
|
||||
#define SAMPLE_NS 1000000L
|
||||
// Context mode: how many samples between scans of the context entries.
|
||||
#define SCAN_EVERY 200
|
||||
|
||||
enum { MODE_RINGBUFFER = 1, MODE_CONTEXTS = 2 };
|
||||
|
||||
struct reader {
|
||||
int fd;
|
||||
void *map;
|
||||
size_t bytes;
|
||||
int first;
|
||||
int mode;
|
||||
uint32_t last[KGSL_BUSY_SLOTS];
|
||||
uint32_t *seen, *ctx_last;
|
||||
int live[KGSL_BUSY_MAX_LIVE];
|
||||
int live_count;
|
||||
pthread_mutex_t lock;
|
||||
pthread_t thread;
|
||||
int thread_on;
|
||||
@@ -32,9 +41,18 @@ struct reader {
|
||||
static void *sample_loop(void *arg) {
|
||||
struct reader *r = arg;
|
||||
const struct timespec gap = {0, SAMPLE_NS};
|
||||
int until_scan = 0;
|
||||
while (atomic_load(&r->run)) {
|
||||
int moved = 0;
|
||||
int busy = kgsl_busy_sample(r->map, r->first, r->last, &moved);
|
||||
int moved = 0, busy;
|
||||
if (r->mode == MODE_CONTEXTS) {
|
||||
if (until_scan-- <= 0) {
|
||||
r->live_count = kgsl_busy_scan_contexts(r->map, r->first, r->seen, r->live, KGSL_BUSY_MAX_LIVE);
|
||||
until_scan = SCAN_EVERY;
|
||||
}
|
||||
busy = kgsl_busy_sample_contexts(r->map, r->live, r->live_count, r->ctx_last, &moved);
|
||||
} else {
|
||||
busy = kgsl_busy_sample(r->map, r->first, r->last, &moved);
|
||||
}
|
||||
if (busy) atomic_fetch_add(&r->busy, 1);
|
||||
if (moved) atomic_fetch_add(&r->moves, 1);
|
||||
atomic_fetch_add(&r->samples, 1);
|
||||
@@ -62,6 +80,8 @@ static void reader_close(struct reader *r) {
|
||||
reader_run(r, 0);
|
||||
if (r->map && r->map != MAP_FAILED) munmap(r->map, r->bytes);
|
||||
if (r->fd >= 0) close(r->fd);
|
||||
free(r->seen);
|
||||
free(r->ctx_last);
|
||||
pthread_mutex_destroy(&r->lock);
|
||||
free(r);
|
||||
}
|
||||
@@ -85,7 +105,16 @@ static int reader_open(struct reader **out) {
|
||||
int moved = 0, started = 0;
|
||||
kgsl_busy_sample(r->map, r->first, r->last, &moved);
|
||||
for (int i = 0; i < KGSL_BUSY_SLOTS; i++) started |= r->last[i] != 0;
|
||||
if (!started) { reader_close(r); return -ENODATA; }
|
||||
if (started) {
|
||||
r->mode = MODE_RINGBUFFER;
|
||||
} else {
|
||||
r->seen = calloc((size_t)r->first, sizeof(uint32_t));
|
||||
r->ctx_last = calloc((size_t)r->first, sizeof(uint32_t));
|
||||
if (!r->seen || !r->ctx_last) { reader_close(r); return -ENOMEM; }
|
||||
r->live_count = kgsl_busy_scan_contexts(r->map, r->first, r->seen, r->live, KGSL_BUSY_MAX_LIVE);
|
||||
kgsl_busy_sample_contexts(r->map, r->live, r->live_count, r->ctx_last, &moved);
|
||||
r->mode = MODE_CONTEXTS;
|
||||
}
|
||||
*out = r;
|
||||
return 0;
|
||||
}
|
||||
@@ -103,6 +132,10 @@ JNIEXPORT jlong JNICALL Java_com_droiddeck_launcher_session_KgslBusy_nativeOpen(
|
||||
return e ? 0 : (jlong)(intptr_t)r;
|
||||
}
|
||||
|
||||
JNIEXPORT jint JNICALL Java_com_droiddeck_launcher_session_KgslBusy_nativeMode(JNIEnv *env, jclass cls, jlong handle) {
|
||||
return handle ? ((struct reader *)(intptr_t)handle)->mode : 0;
|
||||
}
|
||||
|
||||
JNIEXPORT void JNICALL Java_com_droiddeck_launcher_session_KgslBusy_nativeRun(JNIEnv *env, jclass cls, jlong handle, jboolean on) {
|
||||
if (handle) reader_run((struct reader *)(intptr_t)handle, on ? 1 : 0);
|
||||
}
|
||||
|
||||
@@ -34,4 +34,44 @@ static inline int kgsl_busy_sample(const volatile uint32_t *memstore, int first,
|
||||
return busy;
|
||||
}
|
||||
|
||||
// Hardware-scheduled kernels (every gen8 GPU and some gen7 ones: .gpudev = *_hwsched_gpudev in adreno-gpulist.h)
|
||||
// leave the ringbuffer entries at 0. The GMU firmware writes each context's own entry instead
|
||||
// (gen8_hwsched_hfi.c send_context_pointers). Context ids run from 1 below the ringbuffer entries, lowest free first,
|
||||
// and kgsl zeroes an entry when its context is created or destroyed.
|
||||
#define KGSL_BUSY_MAX_LIVE 64
|
||||
|
||||
// Collects the contexts to sample: every entry below first with a non-zero start stamp (kgsl zeroes an entry when its
|
||||
// context is created or destroyed). Entries whose start stamp changed since the previous scan (seen[]) or that are
|
||||
// busy now come first, so the max_live cap only drops idle ones. Returns how many ids went into live[].
|
||||
static inline int kgsl_busy_scan_contexts(const volatile uint32_t *memstore, int first, uint32_t *seen, int *live, int max_live) {
|
||||
int n = 0, idle = 0;
|
||||
int idle_ids[KGSL_BUSY_MAX_LIVE];
|
||||
for (int id = 1; id < first; id++) {
|
||||
const volatile uint32_t *entry = memstore + (size_t)id * KGSL_BUSY_ENTRY_WORDS;
|
||||
uint32_t sop = entry[KGSL_BUSY_SOP], eop = entry[KGSL_BUSY_EOP];
|
||||
int changed = sop != seen[id];
|
||||
seen[id] = sop;
|
||||
if ((changed || sop != eop) && n < max_live) live[n++] = id;
|
||||
else if (sop != 0 && idle < KGSL_BUSY_MAX_LIVE) idle_ids[idle++] = id;
|
||||
}
|
||||
for (int i = 0; i < idle && n < max_live; i++) live[n++] = idle_ids[i];
|
||||
return n;
|
||||
}
|
||||
|
||||
// 1 when a live context has a submission between its start and end timestamps. *moved is set when a start
|
||||
// timestamp differs from last[] (indexed by context id), which then holds the new ones.
|
||||
static inline int kgsl_busy_sample_contexts(const volatile uint32_t *memstore, const int *live, int count, uint32_t *last, int *moved) {
|
||||
int busy = 0;
|
||||
for (int i = 0; i < count; i++) {
|
||||
const volatile uint32_t *entry = memstore + (size_t)live[i] * KGSL_BUSY_ENTRY_WORDS;
|
||||
uint32_t sop = entry[KGSL_BUSY_SOP], eop = entry[KGSL_BUSY_EOP];
|
||||
if (sop != eop) busy = 1;
|
||||
if (sop != last[live[i]]) {
|
||||
*moved = 1;
|
||||
last[live[i]] = sop;
|
||||
}
|
||||
}
|
||||
return busy;
|
||||
}
|
||||
|
||||
#endif
|
||||
@@ -16,21 +16,18 @@
|
||||
#include <unistd.h>
|
||||
|
||||
#include "../adrenotools/src/hook/kgsl.h"
|
||||
#include "kgsl_layout.h"
|
||||
|
||||
#define CP_TYPE7_PKT 0x70000000u
|
||||
#define CP_NOP 0x10u
|
||||
#define CP_REG_TO_MEM 0x3eu
|
||||
#define CP_ALWAYS_ON_COUNTER 0x980u
|
||||
#define RBBM_ALWAYS_COUNT 0u
|
||||
// gen7 CP protection marks registers 0x000-0x4ff readable by command buffers (adreno-gpulist.h); a counter
|
||||
// kgsl places anywhere else is not read.
|
||||
#define READABLE_LAST 0x4ffu
|
||||
#define XO_HZ 19200000.0
|
||||
#define NOP_DWORDS 0x3fffu
|
||||
#define NOP_BLOCKS 4
|
||||
|
||||
struct sampler {
|
||||
int fd;
|
||||
const struct kgsl_layout *layout;
|
||||
unsigned context;
|
||||
unsigned counter_reg;
|
||||
struct kgsl_gpumem_alloc_id ib, out;
|
||||
@@ -60,7 +57,7 @@ static void sampler_close(struct sampler *s) {
|
||||
if (s->out.id) { struct kgsl_gpumem_free_id f = {.id = s->out.id}; ioctl(s->fd, IOCTL_KGSL_GPUMEM_FREE_ID, &f); }
|
||||
if (s->context) { struct kgsl_drawctxt_destroy d = {.drawctxt_id = s->context}; ioctl(s->fd, IOCTL_KGSL_DRAWCTXT_DESTROY, &d); }
|
||||
if (s->counter_reg) {
|
||||
struct kgsl_perfcounter_put p = {.groupid = KGSL_PERFCOUNTER_GROUP_RBBM, .countable = RBBM_ALWAYS_COUNT};
|
||||
struct kgsl_perfcounter_put p = {.groupid = KGSL_PERFCOUNTER_GROUP_RBBM, .countable = s->layout->cycles};
|
||||
ioctl(s->fd, IOCTL_KGSL_PERFCOUNTER_PUT, &p);
|
||||
}
|
||||
close(s->fd);
|
||||
@@ -69,7 +66,7 @@ static void sampler_close(struct sampler *s) {
|
||||
}
|
||||
|
||||
static void emit_reads(struct sampler *s, size_t *n, unsigned pass) {
|
||||
const unsigned regs[2] = {CP_ALWAYS_ON_COUNTER, s->counter_reg};
|
||||
const unsigned regs[2] = {s->layout->always_on, s->counter_reg};
|
||||
for (unsigned k = 0; k < 2; k++) {
|
||||
uint64_t dst = s->out.gpuaddr + (uint64_t)(pass * 2 + k) * sizeof(uint64_t);
|
||||
s->ib_map[(*n)++] = pkt7(CP_REG_TO_MEM, 3);
|
||||
@@ -85,10 +82,16 @@ static int sampler_open(struct sampler **out) {
|
||||
if (!s) return -ENOMEM;
|
||||
s->fd = open("/dev/kgsl-3d0", O_RDWR | O_CLOEXEC);
|
||||
if (s->fd < 0) { int e = -errno; s->fd = -1; sampler_close(s); return e; }
|
||||
struct kgsl_perfcounter_get get = {.groupid = KGSL_PERFCOUNTER_GROUP_RBBM, .countable = RBBM_ALWAYS_COUNT};
|
||||
struct kgsl_devinfo info;
|
||||
memset(&info, 0, sizeof(info));
|
||||
struct kgsl_device_getproperty prop = {.type = KGSL_PROP_DEVICE_INFO, .value = &info, .sizebytes = sizeof(info)};
|
||||
if (ioctl(s->fd, IOCTL_KGSL_DEVICE_GETPROPERTY, &prop)) { int e = -errno; sampler_close(s); return e; }
|
||||
s->layout = kgsl_layout_for(info.chip_id);
|
||||
if (!s->layout) { sampler_close(s); return -ENOTSUP; }
|
||||
struct kgsl_perfcounter_get get = {.groupid = KGSL_PERFCOUNTER_GROUP_RBBM, .countable = s->layout->cycles};
|
||||
if (ioctl(s->fd, IOCTL_KGSL_PERFCOUNTER_GET, &get)) { int e = -errno; sampler_close(s); return e; }
|
||||
s->counter_reg = get.offset;
|
||||
if (!get.offset || get.offset > READABLE_LAST) { sampler_close(s); return -ERANGE; }
|
||||
if (!get.offset || get.offset > s->layout->readable_last) { sampler_close(s); return -ERANGE; }
|
||||
struct kgsl_drawctxt_create ctx = {.flags = KGSL_CONTEXT_NO_GMEM_ALLOC | KGSL_CONTEXT_PREAMBLE};
|
||||
if (ioctl(s->fd, IOCTL_KGSL_DRAWCTXT_CREATE, &ctx)) { int e = -errno; sampler_close(s); return e; }
|
||||
s->context = ctx.drawctxt_id;
|
||||
|
||||
@@ -0,0 +1,27 @@
|
||||
#ifndef DROIDDECK_KGSL_LAYOUT_H
|
||||
#define DROIDDECK_KGSL_LAYOUT_H
|
||||
#include <stddef.h>
|
||||
|
||||
// Register layout per GPU generation. A counter kgsl places outside the range the kernel's CP protection lets
|
||||
// command buffers read (adreno-gpulist.h *_protected_regs) is not read.
|
||||
struct kgsl_layout {
|
||||
unsigned always_on; // CP_ALWAYS_ON_COUNTER
|
||||
unsigned cycles; // RBBM countable that counts GPU cycles
|
||||
unsigned readable_last; // last register of the readable range starting at 0
|
||||
};
|
||||
// A6xx/A7xx: CP_ALWAYS_ON_COUNTER 0x980, RBBM countable 0 = ALWAYS_COUNT, 0x000-0x4ff readable (gen7 list).
|
||||
static const struct kgsl_layout KGSL_LAYOUT_A7XX = {0x980u, 0u, 0x4ffu};
|
||||
// A8xx: CP_ALWAYS_ON_COUNTER 0x8e7 (gen8_reg.h), RBBM countable 1 = US_ALWAYS_COUNT (0 is NEVER_COUNT, Mesa
|
||||
// a8xx_perfcntrs.xml), 0x000-0x3a3 readable (gen8_0_0_protected_regs).
|
||||
static const struct kgsl_layout KGSL_LAYOUT_A8XX = {0x8e7u, 1u, 0x3a3u};
|
||||
|
||||
// The KGSL chip id's top byte: 0x44 on every A8xx (Adreno 810/829/830/840, Mesa freedreno_devices.py),
|
||||
// 0x43 or the generation number on A6xx/A7xx. Later generations are not known, so nothing is read there.
|
||||
static inline const struct kgsl_layout *kgsl_layout_for(unsigned chip_id) {
|
||||
unsigned core = chip_id >> 24;
|
||||
if (core == 0x44u) return &KGSL_LAYOUT_A8XX;
|
||||
if (core > 0x44u) return NULL;
|
||||
return &KGSL_LAYOUT_A7XX;
|
||||
}
|
||||
|
||||
#endif
|
||||
Reference in new issue
Block a user