Acelogic
1510d8c6ce
An EL2 monitor/hypervisor and toolchain that boots macOS 27 on the ASUS A16 (Snapdragon X2 Elite), with an Adreno Metal plug-in and IOGPU kext, and air2ir3, an AIR to Mesa ir3 shader compiler. MIT, with GPL-2.0 components listed in LICENSE and THIRD_PARTY_NOTICES.md. No Apple, Qualcomm, ASUS or Microsoft binaries are included.
171 lines
9.1 KiB
Python
Executable File
171 lines
9.1 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
# SPDX-License-Identifier: MIT
|
|
"""tools/perf_counters.py decodes bytes a real C writer produced.
|
|
|
|
A host program includes compat/perf-state.h, records traps through the same
|
|
census and fast-trap counters the monitor uses, fills the index/canonical
|
|
counters with distinct values and writes the raw struct. The decoder, using
|
|
the TARGET layout of xnuboot.c, must recover every value; truncated or
|
|
inconsistent captures must be refused, and every build (with or without the
|
|
perf defines) must place the state at the offset the define-free layout
|
|
readers compute.
|
|
"""
|
|
from pathlib import Path
|
|
import runpy
|
|
import subprocess
|
|
import tempfile
|
|
import unittest
|
|
|
|
ROOT = Path(__file__).resolve().parents[1]
|
|
perf = runpy.run_path(str(ROOT / 'tools/perf_counters.py'))
|
|
DEFINES = ['-DSNAP_FAST_VCPU_TRAPS=1', '-DSNAP_SYSREG_CENSUS=1', '-DSNAP_INCREMENTAL_PUBLISH=1',
|
|
'-DSNAP_ROOT_DIRTY_PUBLISH=1', '-DSNAP_TXN_DIRTY_PUBLISH=1',
|
|
'-DSNAP_CANONICAL_VIEW_FAST=1']
|
|
BOUNDARY = ['-DSNAP_FAST_VCPU_TRAPS=1', '-DSNAP_VECTOR_BRANCH=1', '-DSNAP_FAST_CROSSINGS=1',
|
|
'-DSNAP_NATIVE_WFE=1', '-DSNAP_NATIVE_CALLS=1',
|
|
'-DSNAP_NATIVE_WRITER_ADMISSION=1', '-DSNAP_NATIVE_WRITER_BREAKPOINTS=1',
|
|
'-DSNAP_ERET_SNAPSHOT=1', '-DSNAP_FAST_TIMER_WRITE=1', '-DSNAP_FAST_IDLE_POLL=1']
|
|
WRITER = r'''
|
|
#include <stdio.h>
|
|
#include <string.h>
|
|
#include "perf-state.h"
|
|
#include "fast-vector.h"
|
|
static struct perf_state p;
|
|
int main(int argc,char **argv)
|
|
{
|
|
(void)argc;
|
|
for(unsigned cpu=0;cpu<GUEST_CPU_COUNT;cpu++) {
|
|
for(unsigned k=1;k<FAST_TRAP_CLASSES;k++)p.fast[cpu].handled[k]=cpu*100+k;
|
|
p.fast[cpu].sysreg=cpu*1000+7;p.fast[cpu].declined=cpu+3;p.fast[cpu].first=cpu+1;
|
|
p.fast[cpu].logged=0xff; /* PACGA and the seven ignored registers */
|
|
/* PACGA, then CONTEXTIDR writes and PMC0 reads through the census. */
|
|
sysreg_census_note(&p.census[cpu],(UINT64_C(0x09)<<26));
|
|
for(unsigned n=0;n<=cpu;n++)
|
|
sysreg_census_note(&p.census[cpu],(UINT64_C(0x18)<<26)|FAST_TRAP_SYSREG(3,0,13,0,1)|(5u<<5));
|
|
for(unsigned n=0;n<2*cpu+1;n++)
|
|
sysreg_census_note(&p.census[cpu],(UINT64_C(0x18)<<26)|FAST_TRAP_SYSREG(3,2,15,0,0)|(9u<<5)|1);
|
|
}
|
|
for(unsigned cpu=0;cpu<GUEST_CPU_COUNT;cpu++) {
|
|
p.fast[cpu].branches=cpu*7+1;p.fast[cpu].entries=cpu*1000+11;p.fast[cpu].returns=cpu*1000+13;
|
|
p.fast[cpu].logged|=fast_cross_log_bit(FAST_CROSS_TIMER)|fast_cross_log_bit(FAST_CROSS_USER_RETURN);
|
|
}
|
|
uint64_t *x=&p.root_index.count;for(unsigned i=0;i<14;i++)x[i]=500+i;
|
|
p.canonical.mapped.calls=64;p.canonical.mapped.passed=63;p.canonical.mapped.leaves=61440;
|
|
p.canonical.mapped.walks=15360;p.canonical.mapped.span_walks=60;
|
|
p.canonical.sampled=1;p.canonical.mismatches=0;
|
|
FILE *f=fopen(argv[1],"wb");fwrite(&p,sizeof p,1,f);fclose(f);
|
|
printf("%zu\n",sizeof p);return 0;
|
|
}
|
|
'''
|
|
|
|
|
|
class PerfCounters(unittest.TestCase):
|
|
def test_decode_real_writer(self):
|
|
with tempfile.TemporaryDirectory() as directory:
|
|
d = Path(directory)
|
|
(d / 'writer.c').write_text(WRITER)
|
|
subprocess.run(['cc', '-std=gnu11', '-O1', '-Icompat', '-Ishadow', str(d / 'writer.c'),
|
|
'-o', str(d / 'writer')], cwd=ROOT, check=True)
|
|
size = int(subprocess.run([str(d / 'writer'), str(d / 'raw.bin')], check=True,
|
|
capture_output=True, text=True).stdout)
|
|
raw = (d / 'raw.bin').read_bytes()
|
|
lay = perf['layout'](DEFINES)
|
|
self.assertEqual(lay['size'], size)
|
|
self.assertGreater(lay['context_offset'], 4_000_000)
|
|
out = perf['decode'](raw, lay=lay)
|
|
for cpu, row in enumerate(out['fast']):
|
|
self.assertEqual(row['handled'], {k: cpu * 100 + i + 1 for i, k in
|
|
enumerate(perf['CLASSES'][1:])})
|
|
self.assertEqual((row['sysreg'], row['declined'], row['first'], row['logged']),
|
|
(cpu * 1000 + 7, cpu + 3, cpu + 1, 0xff | (1 << 10) | (1 << 13)))
|
|
self.assertEqual((row['branches'], row['entries'], row['returns']),
|
|
(cpu * 7 + 1, cpu * 1000 + 11, cpu * 1000 + 13))
|
|
for cpu, row in enumerate(out['census']):
|
|
self.assertEqual((row['total'], row['other'], row['pacga']), (3 * cpu + 2, 0, 1))
|
|
self.assertEqual(row['top'][0], dict(register='PMC0 read', count=2 * cpu + 1))
|
|
self.assertIn(dict(register='CONTEXTIDR_EL1 write', count=cpu + 1), row['top'])
|
|
self.assertEqual(out['summary']['vector_branches'], sum(c * 7 + 1 for c in range(6)))
|
|
self.assertEqual((out['summary']['fast_entries'], out['summary']['fast_returns']),
|
|
(sum(c * 1000 + 11 for c in range(6)), sum(c * 1000 + 13 for c in range(6))))
|
|
self.assertEqual(out['root_index']['count'], 500)
|
|
self.assertEqual(out['root_index']['txn_scanned'], 513)
|
|
self.assertEqual(out['canonical_view'], dict(calls=64, passed=63, leaves=61440,
|
|
walks=15360, span_walks=60, sampled=1,
|
|
mismatches=0))
|
|
with self.assertRaises(ValueError):
|
|
perf['decode'](raw[:-8], lay=lay)
|
|
bad = bytearray(raw)
|
|
bad[lay['census'] + lay['census_total']] ^= 1 # cpu0 census total
|
|
with self.assertRaises(ValueError):
|
|
perf['decode'](bytes(bad), lay=lay)
|
|
# An occupied slot whose count was lost, and a wrong distinct-key word.
|
|
occupied = next(i for i in range(lay['census_slots'])
|
|
if any(raw[lay['census'] + lay['census_key'] + 8 * i:][:8]))
|
|
for at in (lay['census_count'] + 8 * occupied, lay['census_total'] + 24):
|
|
bad = bytearray(raw)
|
|
bad[lay['census'] + at:lay['census'] + at + 8] = (
|
|
bytes(8) if at != lay['census_total'] + 24 else (99).to_bytes(8, 'little'))
|
|
with self.assertRaises(ValueError):
|
|
perf['decode'](bytes(bad), lay=lay)
|
|
# class 0, a stray log bit, first > declined, the reserved word
|
|
for at, flip in ((0, 1), (lay['fast_logged'] + 3, 0x08), (lay['fast_first'], 0xff),
|
|
(lay['fast_reserved'], 1)):
|
|
bad = bytearray(raw)
|
|
bad[lay['fast'] + lay['fast_stride'] + at] ^= flip
|
|
with self.assertRaises(ValueError):
|
|
perf['decode'](bytes(bad), lay=lay)
|
|
|
|
def test_late_register_is_named(self):
|
|
# Protects docs/SYSREG-CENSUS-20261016.md failure mode 1: a register
|
|
# first trapped after many others must still get its own slot. The
|
|
# 48-slot first-come list counted it only as `other` (31% of the
|
|
# speed-test runs' system-register traps), so this fails against it.
|
|
writer = r'''
|
|
#include <stdio.h>
|
|
#include "perf-state.h"
|
|
static struct perf_state p;
|
|
int main(int argc,char **argv)
|
|
{
|
|
(void)argc;
|
|
struct sysreg_census_cpu *s=&p.census[2];
|
|
for(unsigned op1=0,n=0;op1<7;op1++)
|
|
for(unsigned crm=0;crm<16 && n<100;crm++,n++)
|
|
sysreg_census_note(s,(UINT64_C(0x18)<<26)|FAST_TRAP_SYSREG(3,op1,11,crm,2)|(3u<<5)|1);
|
|
for(unsigned n=0;n<5000;n++) /* CNTV_TVAL_ALIAS write, first seen last */
|
|
sysreg_census_note(s,(UINT64_C(0x18)<<26)|FAST_TRAP_SYSREG(3,1,15,15,4)|((n%31)<<5));
|
|
FILE *f=fopen(argv[1],"wb");fwrite(&p,sizeof p,1,f);fclose(f);
|
|
return 0;
|
|
}
|
|
'''
|
|
with tempfile.TemporaryDirectory() as directory:
|
|
d = Path(directory)
|
|
(d / 'late.c').write_text(writer)
|
|
subprocess.run(['cc', '-std=gnu11', '-O1', '-Icompat', '-Ishadow', str(d / 'late.c'),
|
|
'-o', str(d / 'late')], cwd=ROOT, check=True)
|
|
subprocess.run([str(d / 'late'), str(d / 'raw.bin')], check=True)
|
|
raw = (d / 'raw.bin').read_bytes()
|
|
row = perf['decode'](raw, lay=perf['layout'](DEFINES))['census'][2]
|
|
self.assertEqual((row['total'], row['other']), (5100, 0))
|
|
self.assertEqual(row['top'][0], dict(register='CNTV_TVAL_ALIAS write', count=5000))
|
|
self.assertEqual(len(row['top']), 101)
|
|
|
|
def test_one_layout_for_every_build(self):
|
|
# Layout readers compute struct context without build defines; the
|
|
# appended state must sit at the same place in every build.
|
|
plain = perf['layout']([])
|
|
for defines in ([], DEFINES, ['-DSNAP_SYSREG_CENSUS=1'], ['-DSNAP_CANONICAL_VIEW_FAST=1'],
|
|
BOUNDARY):
|
|
lay = perf['layout'](defines)
|
|
self.assertEqual((lay['context_offset'], lay['size']),
|
|
(plain['context_offset'], plain['size']))
|
|
layout = runpy.run_path(str(ROOT / 'tools/monitor_layout.py'))['MonitorLayout']()
|
|
self.assertEqual(layout.offset('context', 'perf'), plain['context_offset'])
|
|
self.assertLessEqual(plain['context_offset'] + plain['size'], layout.size('context'))
|
|
# The next append, the SKS request capture, follows it directly.
|
|
self.assertEqual(layout.offset('context', 'sep_sks_request'),
|
|
plain['context_offset'] + plain['size'])
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|