tcg: add perfmap and jitdump

Add ability to dump /tmp/perf-<pid>.map and jit-<pid>.dump.
The first one allows the perf tool to map samples to each individual
translation block. The second one adds the ability to resolve symbol
names, line numbers and inspect JITed code.

Example of use:

    perf record qemu-x86_64 -perfmap ./a.out
    perf report

or

    perf record -k 1 qemu-x86_64 -jitdump ./a.out
    DEBUGINFOD_URLS= perf inject -j -i perf.data -o perf.data.jitted
    perf report -i perf.data.jitted

Co-developed-by: Vanderson M. do Rosario <vandersonmr2@gmail.com>
Co-developed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
Message-Id: <20230112152013.125680-4-iii@linux.ibm.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit is contained in:
Ilya Leoshkevich 2023-01-12 16:20:13 +01:00 committed by Richard Henderson
parent 7c10cb38cc
commit 5584e2dbe8
10 changed files with 505 additions and 0 deletions

View File

@ -13,6 +13,7 @@ tcg_ss.add(when: 'CONFIG_USER_ONLY', if_true: files('user-exec.c'))
tcg_ss.add(when: 'CONFIG_SOFTMMU', if_false: files('user-exec-stub.c'))
tcg_ss.add(when: 'CONFIG_PLUGIN', if_true: [files('plugin-gen.c')])
tcg_ss.add(when: libdw, if_true: files('debuginfo.c'))
tcg_ss.add(when: 'CONFIG_LINUX', if_true: files('perf.c'))
specific_ss.add_all(when: 'CONFIG_TCG', if_true: tcg_ss)
specific_ss.add(when: ['CONFIG_SOFTMMU', 'CONFIG_TCG'], if_true: files(

375
accel/tcg/perf.c Normal file
View File

@ -0,0 +1,375 @@
/*
* Linux perf perf-<pid>.map and jit-<pid>.dump integration.
*
* The jitdump spec can be found at [1].
*
* [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/tools/perf/Documentation/jitdump-specification.txt
*
* SPDX-License-Identifier: GPL-2.0-or-later
*/
#include "qemu/osdep.h"
#include "elf.h"
#include "exec/exec-all.h"
#include "qemu/timer.h"
#include "tcg/tcg.h"
#include "debuginfo.h"
#include "perf.h"
static FILE *safe_fopen_w(const char *path)
{
int saved_errno;
FILE *f;
int fd;
/* Delete the old file, if any. */
unlink(path);
/* Avoid symlink attacks by using O_CREAT | O_EXCL. */
fd = open(path, O_RDWR | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR);
if (fd == -1) {
return NULL;
}
/* Convert fd to FILE*. */
f = fdopen(fd, "w");
if (f == NULL) {
saved_errno = errno;
close(fd);
errno = saved_errno;
return NULL;
}
return f;
}
static FILE *perfmap;
void perf_enable_perfmap(void)
{
char map_file[32];
snprintf(map_file, sizeof(map_file), "/tmp/perf-%d.map", getpid());
perfmap = safe_fopen_w(map_file);
if (perfmap == NULL) {
warn_report("Could not open %s: %s, proceeding without perfmap",
map_file, strerror(errno));
}
}
/* Get PC and size of code JITed for guest instruction #INSN. */
static void get_host_pc_size(uintptr_t *host_pc, uint16_t *host_size,
const void *start, size_t insn)
{
uint16_t start_off = insn ? tcg_ctx->gen_insn_end_off[insn - 1] : 0;
if (host_pc) {
*host_pc = (uintptr_t)start + start_off;
}
if (host_size) {
*host_size = tcg_ctx->gen_insn_end_off[insn] - start_off;
}
}
static const char *pretty_symbol(const struct debuginfo_query *q, size_t *len)
{
static __thread char buf[64];
int tmp;
if (!q->symbol) {
tmp = snprintf(buf, sizeof(buf), "guest-0x%"PRIx64, q->address);
if (len) {
*len = MIN(tmp + 1, sizeof(buf));
}
return buf;
}
if (!q->offset) {
if (len) {
*len = strlen(q->symbol) + 1;
}
return q->symbol;
}
tmp = snprintf(buf, sizeof(buf), "%s+0x%"PRIx64, q->symbol, q->offset);
if (len) {
*len = MIN(tmp + 1, sizeof(buf));
}
return buf;
}
static void write_perfmap_entry(const void *start, size_t insn,
const struct debuginfo_query *q)
{
uint16_t host_size;
uintptr_t host_pc;
get_host_pc_size(&host_pc, &host_size, start, insn);
fprintf(perfmap, "%"PRIxPTR" %"PRIx16" %s\n",
host_pc, host_size, pretty_symbol(q, NULL));
}
static FILE *jitdump;
#define JITHEADER_MAGIC 0x4A695444
#define JITHEADER_VERSION 1
struct jitheader {
uint32_t magic;
uint32_t version;
uint32_t total_size;
uint32_t elf_mach;
uint32_t pad1;
uint32_t pid;
uint64_t timestamp;
uint64_t flags;
};
enum jit_record_type {
JIT_CODE_LOAD = 0,
JIT_CODE_DEBUG_INFO = 2,
};
struct jr_prefix {
uint32_t id;
uint32_t total_size;
uint64_t timestamp;
};
struct jr_code_load {
struct jr_prefix p;
uint32_t pid;
uint32_t tid;
uint64_t vma;
uint64_t code_addr;
uint64_t code_size;
uint64_t code_index;
};
struct debug_entry {
uint64_t addr;
int lineno;
int discrim;
const char name[];
};
struct jr_code_debug_info {
struct jr_prefix p;
uint64_t code_addr;
uint64_t nr_entry;
struct debug_entry entries[];
};
static uint32_t get_e_machine(void)
{
Elf64_Ehdr elf_header;
FILE *exe;
size_t n;
QEMU_BUILD_BUG_ON(offsetof(Elf32_Ehdr, e_machine) !=
offsetof(Elf64_Ehdr, e_machine));
exe = fopen("/proc/self/exe", "r");
if (exe == NULL) {
return EM_NONE;
}
n = fread(&elf_header, sizeof(elf_header), 1, exe);
fclose(exe);
if (n != 1) {
return EM_NONE;
}
return elf_header.e_machine;
}
void perf_enable_jitdump(void)
{
struct jitheader header;
char jitdump_file[32];
void *perf_marker;
if (!use_rt_clock) {
warn_report("CLOCK_MONOTONIC is not available, proceeding without jitdump");
return;
}
snprintf(jitdump_file, sizeof(jitdump_file), "jit-%d.dump", getpid());
jitdump = safe_fopen_w(jitdump_file);
if (jitdump == NULL) {
warn_report("Could not open %s: %s, proceeding without jitdump",
jitdump_file, strerror(errno));
return;
}
/*
* `perf inject` will see that the mapped file name in the corresponding
* PERF_RECORD_MMAP or PERF_RECORD_MMAP2 event is of the form jit-%d.dump
* and will process it as a jitdump file.
*/
perf_marker = mmap(NULL, qemu_real_host_page_size(), PROT_READ | PROT_EXEC,
MAP_PRIVATE, fileno(jitdump), 0);
if (perf_marker == MAP_FAILED) {
warn_report("Could not map %s: %s, proceeding without jitdump",
jitdump_file, strerror(errno));
fclose(jitdump);
jitdump = NULL;
return;
}
header.magic = JITHEADER_MAGIC;
header.version = JITHEADER_VERSION;
header.total_size = sizeof(header);
header.elf_mach = get_e_machine();
header.pad1 = 0;
header.pid = getpid();
header.timestamp = get_clock();
header.flags = 0;
fwrite(&header, sizeof(header), 1, jitdump);
}
void perf_report_prologue(const void *start, size_t size)
{
if (perfmap) {
fprintf(perfmap, "%"PRIxPTR" %zx tcg-prologue-buffer\n",
(uintptr_t)start, size);
}
}
/* Write a JIT_CODE_DEBUG_INFO jitdump entry. */
static void write_jr_code_debug_info(const void *start,
const struct debuginfo_query *q,
size_t icount)
{
struct jr_code_debug_info rec;
struct debug_entry ent;
uintptr_t host_pc;
int insn;
/* Write the header. */
rec.p.id = JIT_CODE_DEBUG_INFO;
rec.p.total_size = sizeof(rec) + sizeof(ent) + 1;
rec.p.timestamp = get_clock();
rec.code_addr = (uintptr_t)start;
rec.nr_entry = 1;
for (insn = 0; insn < icount; insn++) {
if (q[insn].file) {
rec.p.total_size += sizeof(ent) + strlen(q[insn].file) + 1;
rec.nr_entry++;
}
}
fwrite(&rec, sizeof(rec), 1, jitdump);
/* Write the main debug entries. */
for (insn = 0; insn < icount; insn++) {
if (q[insn].file) {
get_host_pc_size(&host_pc, NULL, start, insn);
ent.addr = host_pc;
ent.lineno = q[insn].line;
ent.discrim = 0;
fwrite(&ent, sizeof(ent), 1, jitdump);
fwrite(q[insn].file, strlen(q[insn].file) + 1, 1, jitdump);
}
}
/* Write the trailing debug_entry. */
ent.addr = (uintptr_t)start + tcg_ctx->gen_insn_end_off[icount - 1];
ent.lineno = 0;
ent.discrim = 0;
fwrite(&ent, sizeof(ent), 1, jitdump);
fwrite("", 1, 1, jitdump);
}
/* Write a JIT_CODE_LOAD jitdump entry. */
static void write_jr_code_load(const void *start, uint16_t host_size,
const struct debuginfo_query *q)
{
static uint64_t code_index;
struct jr_code_load rec;
const char *symbol;
size_t symbol_size;
symbol = pretty_symbol(q, &symbol_size);
rec.p.id = JIT_CODE_LOAD;
rec.p.total_size = sizeof(rec) + symbol_size + host_size;
rec.p.timestamp = get_clock();
rec.pid = getpid();
rec.tid = qemu_get_thread_id();
rec.vma = (uintptr_t)start;
rec.code_addr = (uintptr_t)start;
rec.code_size = host_size;
rec.code_index = code_index++;
fwrite(&rec, sizeof(rec), 1, jitdump);
fwrite(symbol, symbol_size, 1, jitdump);
fwrite(start, host_size, 1, jitdump);
}
void perf_report_code(uint64_t guest_pc, TranslationBlock *tb,
const void *start)
{
struct debuginfo_query *q;
size_t insn;
if (!perfmap && !jitdump) {
return;
}
q = g_try_malloc0_n(tb->icount, sizeof(*q));
if (!q) {
return;
}
debuginfo_lock();
/* Query debuginfo for each guest instruction. */
for (insn = 0; insn < tb->icount; insn++) {
/* FIXME: This replicates the restore_state_to_opc() logic. */
q[insn].address = tcg_ctx->gen_insn_data[insn][0];
if (TARGET_TB_PCREL) {
q[insn].address |= (guest_pc & TARGET_PAGE_MASK);
} else {
#if defined(TARGET_I386)
q[insn].address -= tb->cs_base;
#endif
}
q[insn].flags = DEBUGINFO_SYMBOL | (jitdump ? DEBUGINFO_LINE : 0);
}
debuginfo_query(q, tb->icount);
/* Emit perfmap entries if needed. */
if (perfmap) {
flockfile(perfmap);
for (insn = 0; insn < tb->icount; insn++) {
write_perfmap_entry(start, insn, &q[insn]);
}
funlockfile(perfmap);
}
/* Emit jitdump entries if needed. */
if (jitdump) {
flockfile(jitdump);
write_jr_code_debug_info(start, q, tb->icount);
write_jr_code_load(start, tcg_ctx->gen_insn_end_off[tb->icount - 1],
q);
funlockfile(jitdump);
}
debuginfo_unlock();
g_free(q);
}
void perf_exit(void)
{
if (perfmap) {
fclose(perfmap);
perfmap = NULL;
}
if (jitdump) {
fclose(jitdump);
jitdump = NULL;
}
}

49
accel/tcg/perf.h Normal file
View File

@ -0,0 +1,49 @@
/*
* Linux perf perf-<pid>.map and jit-<pid>.dump integration.
*
* SPDX-License-Identifier: GPL-2.0-or-later
*/
#ifndef ACCEL_TCG_PERF_H
#define ACCEL_TCG_PERF_H
#if defined(CONFIG_TCG) && defined(CONFIG_LINUX)
/* Start writing perf-<pid>.map. */
void perf_enable_perfmap(void);
/* Start writing jit-<pid>.dump. */
void perf_enable_jitdump(void);
/* Add information about TCG prologue to profiler maps. */
void perf_report_prologue(const void *start, size_t size);
/* Add information about JITted guest code to profiler maps. */
void perf_report_code(uint64_t guest_pc, TranslationBlock *tb,
const void *start);
/* Stop writing perf-<pid>.map and/or jit-<pid>.dump. */
void perf_exit(void);
#else
static inline void perf_enable_perfmap(void)
{
}
static inline void perf_enable_jitdump(void)
{
}
static inline void perf_report_prologue(const void *start, size_t size)
{
}
static inline void perf_report_code(uint64_t guest_pc, TranslationBlock *tb,
const void *start)
{
}
static inline void perf_exit(void)
{
}
#endif
#endif

View File

@ -62,6 +62,7 @@
#include "tb-hash.h"
#include "tb-context.h"
#include "internal.h"
#include "perf.h"
/* Make sure all possible CPU event bits fit in tb->trace_vcpu_dstate */
QEMU_BUILD_BUG_ON(CPU_TRACE_DSTATE_MAX_EVENTS >
@ -406,6 +407,12 @@ TranslationBlock *tb_gen_code(CPUState *cpu,
}
tb->tc.size = gen_code_size;
/*
* For TARGET_TB_PCREL, attribute all executions of the generated
* code to its first mapping.
*/
perf_report_code(pc, tb, tcg_splitwx_to_rx(gen_code_buf));
#ifdef CONFIG_PROFILER
qatomic_set(&prof->code_time, prof->code_time + profile_getclock() - ti);
qatomic_set(&prof->code_in_len, prof->code_in_len + tb->size);

View File

@ -188,3 +188,26 @@ memory areas instead calls out to C code for device emulation.
Finally, the MMU helps tracking dirty pages and pages pointed to by
translation blocks.
Profiling JITted code
---------------------
The Linux ``perf`` tool will treat all JITted code as a single block as
unlike the main code it can't use debug information to link individual
program counter samples with larger functions. To overcome this
limitation you can use the ``-perfmap`` or the ``-jitdump`` option to generate
map files. ``-perfmap`` is lightweight and produces only guest-host mappings.
``-jitdump`` additionally saves JITed code and guest debug information (if
available); its output needs to be integrated with the ``perf.data`` file
before the final report can be viewed.
.. code::
perf record $QEMU -perfmap $REMAINING_ARGS
perf report
perf record -k 1 $QEMU -jitdump $REMAINING_ARGS
DEBUGINFOD_URLS= perf inject -j -i perf.data -o perf.data.jitted
perf report -i perf.data.jitted
Note that qemu-system generates mappings only for ``-kernel`` files in ELF
format.

View File

@ -17,6 +17,7 @@
* along with this program; if not, see <http://www.gnu.org/licenses/>.
*/
#include "qemu/osdep.h"
#include "accel/tcg/perf.h"
#include "exec/gdbstub.h"
#include "qemu.h"
#include "user-internals.h"
@ -38,4 +39,5 @@ void preexit_cleanup(CPUArchState *env, int code)
#endif
gdb_exit(code);
qemu_plugin_user_exit();
perf_exit();
}

View File

@ -53,6 +53,7 @@
#include "signal-common.h"
#include "loader.h"
#include "user-mmap.h"
#include "accel/tcg/perf.h"
#ifdef CONFIG_SEMIHOSTING
#include "semihosting/semihost.h"
@ -423,6 +424,16 @@ static void handle_arg_abi_call0(const char *arg)
}
#endif
static void handle_arg_perfmap(const char *arg)
{
perf_enable_perfmap();
}
static void handle_arg_jitdump(const char *arg)
{
perf_enable_jitdump();
}
static QemuPluginList plugins = QTAILQ_HEAD_INITIALIZER(plugins);
#ifdef CONFIG_PLUGIN
@ -493,6 +504,10 @@ static const struct qemu_argument arg_table[] = {
{"xtensa-abi-call0", "QEMU_XTENSA_ABI_CALL0", false, handle_arg_abi_call0,
"", "assume CALL0 Xtensa ABI"},
#endif
{"perfmap", "QEMU_PERFMAP", false, handle_arg_perfmap,
"", "Generate a /tmp/perf-${pid}.map file for perf"},
{"jitdump", "QEMU_JITDUMP", false, handle_arg_jitdump,
"", "Generate a jit-${pid}.dump file for perf"},
{NULL, NULL, false, NULL, NULL, NULL}
};

View File

@ -4838,6 +4838,26 @@ SRST
Enable synchronization profiling.
ERST
#if defined(CONFIG_TCG) && defined(CONFIG_LINUX)
DEF("perfmap", 0, QEMU_OPTION_perfmap,
"-perfmap generate a /tmp/perf-${pid}.map file for perf\n",
QEMU_ARCH_ALL)
SRST
``-perfmap``
Generate a map file for Linux perf tools that will allow basic profiling
information to be broken down into basic blocks.
ERST
DEF("jitdump", 0, QEMU_OPTION_jitdump,
"-jitdump generate a jit-${pid}.dump file for perf\n",
QEMU_ARCH_ALL)
SRST
``-jitdump``
Generate a dump file for Linux perf tools that maps basic blocks to symbol
names, line numbers and JITted code.
ERST
#endif
DEFHEADING()
DEFHEADING(Generic object creation:)

View File

@ -96,6 +96,9 @@
#include "fsdev/qemu-fsdev.h"
#endif
#include "sysemu/qtest.h"
#ifdef CONFIG_TCG
#include "accel/tcg/perf.h"
#endif
#include "disas/disas.h"
@ -2926,6 +2929,14 @@ void qemu_init(int argc, char **argv)
case QEMU_OPTION_DFILTER:
qemu_set_dfilter_ranges(optarg, &error_fatal);
break;
#if defined(CONFIG_TCG) && defined(CONFIG_LINUX)
case QEMU_OPTION_perfmap:
perf_enable_perfmap();
break;
case QEMU_OPTION_jitdump:
perf_enable_jitdump();
break;
#endif
case QEMU_OPTION_seed:
qemu_guest_random_seed_main(optarg, &error_fatal);
break;

View File

@ -61,6 +61,7 @@
#include "exec/log.h"
#include "tcg/tcg-ldst.h"
#include "tcg-internal.h"
#include "accel/tcg/perf.h"
/* Forward declarations for functions declared in tcg-target.c.inc and
used here. */
@ -913,6 +914,7 @@ void tcg_prologue_init(TCGContext *s)
#endif
prologue_size = tcg_current_code_size(s);
perf_report_prologue(s->code_gen_ptr, prologue_size);
#ifndef CONFIG_TCG_INTERPRETER
flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf),