tcg: add perfmap and jitdump
PATCH 1/1 tcg: add perfmap and jitdump
In reply to Ilya LeoshkevichAdd 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 perf inject -j -i perf.data -o perf.data.jitted perf report -i perf.data.jitted
--- accel/tcg/debuginfo.c | 108 +++++++++++++ accel/tcg/debuginfo.h | 54 +++++++ accel/tcg/meson.build | 2 + accel/tcg/perf.c | 333 ++++++++++++++++++++++++++++++++++++++ accel/tcg/perf.h | 28 ++++ accel/tcg/translate-all.c | 3 + docs/devel/tcg.rst | 20 +++ linux-user/elfload.c | 3 + linux-user/exit.c | 2 + linux-user/main.c | 15 ++ linux-user/meson.build | 1 + meson.build | 8 + qemu-options.hx | 20 +++ softmmu/vl.c | 11 ++ tcg/tcg.c | 2 + 15 files changed, 610 insertions(+) create mode 100644 accel/tcg/debuginfo.c create mode 100644 accel/tcg/debuginfo.h create mode 100644 accel/tcg/perf.c create mode 100644 accel/tcg/perf.h
accel/tcg/debuginfo.c new file +108 -0
@@ -0,0 +1,108 @@ +/* +* Debug information support. +* +* SPDX-License-Identifier: GPL-2.0-or-later +*/ + +#include "qemu/osdep.h" + +#include <elfutils/libdwfl.h> + +#include "debuginfo.h" + +static QemuMutex lock; +static Dwfl *dwfl; +static const Dwfl_Callbacks dwfl_callbacks = { + .find_elf =NULL , + .find_debuginfo = dwfl_standard_find_debuginfo, + .section_address =NULL , + .debuginfo_path =NULL , +}; + +__attribute__((constructor)) +static void debuginfo_init (void ) +{ +qemu_mutex_init (&lock); +} + +bool debuginfo_report_elf (const char *image_name,int image_fd, +target_ulong load_bias) +{ +qemu_mutex_lock (&lock); + +if (dwfl ==NULL ) { + dwfl =dwfl_begin (&dwfl_callbacks); + }else { +dwfl_report_begin_add (dwfl); + } + +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + +dwfl_report_elf (dwfl, image_name, image_name, image_fd, load_bias, true); +dwfl_report_end (dwfl,NULL ,NULL ); +qemu_mutex_unlock (&lock); +return true; +} + +bool debuginfo_get_symbol (target_ulong address, +const char **symbol,target_ulong *offset) +{ +Dwfl_Module *dwfl_module; +GElf_Off dwfl_offset; +GElf_Sym dwfl_sym; + +qemu_mutex_lock (&lock); + +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_module =dwfl_addrmodule (dwfl, address); +if (dwfl_module ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + *symbol =dwfl_module_addrinfo (dwfl_module, address, &dwfl_offset, + &dwfl_sym,NULL ,NULL ,NULL ); +if (*symbol ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + *offset = dwfl_offset; +qemu_mutex_unlock (&lock); +return true; +} + +bool debuginfo_get_line (target_ulong address, +const char **file,int *line) +{ +Dwfl_Module *dwfl_module; +Dwfl_Line *dwfl_line; + +qemu_mutex_lock (&lock); + +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_module =dwfl_addrmodule (dwfl, address); +if (dwfl_module ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_line =dwfl_module_getsrc (dwfl_module, address); +if (dwfl_line ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + *file =dwfl_lineinfo (dwfl_line,NULL , line,0 ,NULL ,NULL ); +qemu_mutex_unlock (&lock); +return true; +}
accel/tcg/debuginfo.h new file +54 -0
54 lines not on this page · raw message
accel/tcg/meson.build +2 -0
@@ -1,5 +1,6 @@ tcg_ss = ss.source_set() tcg_ss.add(files( + 'perf.c', 'tcg-all.c', 'cpu-exec-common.c', 'cpu-exec.c',
@@ -11,6 +12,7 @@ tcg_ss.add(files( 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')) specific_ss.add_all(when: 'CONFIG_TCG', if_true: tcg_ss) specific_ss.add(when: ['CONFIG_SOFTMMU', 'CONFIG_TCG'], if_true: files(
accel/tcg/perf.c new file +333 -0
333 lines not on this page · raw message
accel/tcg/perf.h new file +28 -0
@@ -0,0 +1,28 @@ +/* +* 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 + +#include <stddef.h> +#include <stdint.h> + +/* 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 (void *start,size_t size); + +/* Add information about JITted guest code to profiler maps. */ +void perf_report_code (void *start,size_t size,int icount,uint64_t pc); + +/* Stop writing perf-<pid>.map and/or jit-<pid>.dump. */ +void perf_exit (void ); + +#endif
accel/tcg/translate-all.c +3 -0
@@ -62,6 +62,7 @@#include "tb-hash.h" #include "tb-context.h" #include "internal.h" +#include "perf.h" /* #define DEBUG_TB_INVALIDATE */ /* #define DEBUG_TB_FLUSH */
@@ -1492,6 +1493,8 @@ TranslationBlock *tb_gen_code(CPUState *cpu, } tb->tc.size = gen_code_size; +perf_report_code (gen_code_buf, gen_code_size, tb->icount, tb->pc); +#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);
docs/devel/tcg.rst +20 -0
@@ -188,3 +188,23 @@ 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 $REMAINING_ARGS + perf inject -j -i perf.data -o perf.data.jitted + perf report -i perf.data.jitted
linux-user/elfload.c +3 -0
@@ -19,6 +19,7 @@#include "qemu/selfmap.h" #include "qapi/error.h" #include "target_signal.h" +#include "accel/tcg/debuginfo.h" #ifdef _ARCH_PPC64#undef ARCH_DLINFO
@@ -3261,6 +3262,8 @@ static void load_elf_image(const char *image_name, int image_fd,load_symbols (ehdr, image_fd, load_bias); } +debuginfo_report_elf (image_name, image_fd, load_bias); +mmap_unlock ();close (image_fd);
linux-user/exit.c +2 -0
@@ -17,6 +17,7 @@* along with this program; if not, see <http://www.gnu.org/licenses/>. */ #include "qemu/osdep.h" +#include "accel/tcg/perf.h" #include "exec/gdbstub.h" #include "qemu.h" #include "user-internals.h"
@@ -38,4 +39,5 @@ void preexit_cleanup(CPUArchState *env, int code)#endif gdb_exit (code);qemu_plugin_user_exit (); +perf_exit (); }
linux-user/main.c +15 -0
@@ -53,6 +53,7 @@#include "signal-common.h" #include "loader.h" #include "user-mmap.h" +#include "accel/tcg/perf.h" #ifdef CONFIG_SEMIHOSTING #include "semihosting/semihost.h"
@@ -423,6 +424,16 @@ static void handle_arg_abi_call0(const char *arg) }#endif +static void handle_arg_perfmap (const char *arg) +{ +perf_enable_perfmap (); +} + +static void handle_arg_jitdump (const char *arg) +{ +perf_enable_jitdump (); +} +static QemuPluginList plugins =QTAILQ_HEAD_INITIALIZER (plugins);#ifdef CONFIG_PLUGIN
@@ -493,6 +504,10 @@ static const struct qemu_argument arg_table[] = { {"xtensa-abi-call0", "QEMU_XTENSA_ABI_CALL0", false, handle_arg_abi_call0, "", "assume CALL0 Xtensa ABI"}, #endif + {"perfmap", "QEMU_PERFMAP", false, handle_arg_perfmap, + "", "Generate a /tmp/perf-${pid}.map file for perf"}, + {"jitdump", "QEMU_JITDUMP", false, handle_arg_jitdump, + "", "Generate a jit-${pid}.dump file for perf"}, {NULL, NULL, false, NULL, NULL, NULL} };
linux-user/meson.build +1 -0
@@ -22,6 +22,7 @@ linux_user_ss.add(files( 'uname.c', )) linux_user_ss.add(rt) +linux_user_ss.add(libdw) linux_user_ss.add(when: 'TARGET_HAS_BFLT', if_true: files('flatload.c')) linux_user_ss.add(when: 'TARGET_I386', if_true: files('vm86.c'))
meson.build +8 -0
@@ -1631,6 +1631,12 @@ if libbpf.found() and not cc.links(''' endif endif +# libdw +libdw = dependency('libdw', + method: 'pkg-config', + kwargs: static_kwargs, + required: false) + ################# # config-host.h # #################
@@ -1897,6 +1903,7 @@ config_host_data.set('CONFIG_DBUS_DISPLAY', dbus_display) config_host_data.set('CONFIG_CFI', get_option('cfi')) config_host_data.set('CONFIG_SELINUX', selinux.found()) config_host_data.set('CONFIG_XEN_BACKEND', xen.found()) +config_host_data.set('CONFIG_LIBDW', libdw.found()) if xen.found() # protect from xen.version() having less than three components xen_version = xen.version().split('.') + ['0', '0']
@@ -3937,6 +3944,7 @@ summary_info += {'libudev': libudev} # Dummy dependency, keep .found() summary_info += {'FUSE lseek': fuse_lseek.found()} summary_info += {'selinux': selinux} +summary_info += {'libdw': libdw} summary(summary_info, bool_yn: true, section: 'Dependencies') if not supported_cpus.contains(cpu)
qemu-options.hx +20 -0
@@ -4799,6 +4799,26 @@ SRST Enable synchronization profiling. ERST +#ifdef CONFIG_TCG +DEF("perfmap", 0, QEMU_OPTION_perfmap, + "-perfmap generate a /tmp/perf-${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:)
softmmu/vl.c +11 -0
@@ -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"
@@ -2900,6 +2903,14 @@ void qemu_init(int argc, char **argv)case QEMU_OPTION_DFILTER :qemu_set_dfilter_ranges (optarg, &error_fatal);break ; +#ifdef CONFIG_TCG +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 ;
tcg/tcg.c +2 -0
@@ -61,6 +61,7 @@#include "exec/log.h" #include "tcg/tcg-ldst.h" #include "tcg-internal.h" +#include "accel/tcg/perf.h" #ifdef CONFIG_TCG_INTERPRETER #include <ffi.h>
@@ -749,6 +750,7 @@ 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),
Signature · 1 line
2.37.2
cover tcg: add perfmap and jitdump
Hi,
This is a rework of jitdump and perfmap patches from Vanderson and Alex:
- jitdump: https://wiki.qemu.org/Features/TCGCodeQuality
v1: https://lists.gnu.org/archive/html/qemu-devel/2019-08/msg02676.html
v2: https://lore.kernel.org/qemu-devel/[email protected]/
v9: https://www.mail-archive.com/[email protected]/msg650269.html
- perfmap
rfc: https://lists.nongnu.org/archive/html/qemu-devel/2014-03/msg05386.html
v2: https://lists.gnu.org/archive/html/qemu-devel/2014-07/msg02061.html
v3: https://lists.gnu.org/archive/html/qemu-devel/2014-07/msg04289.html
v4: https://lists.gnu.org/archive/html/qemu-devel/2015-08/msg00095.html
It addresses some of the feedback (among other things: symlink attacks, Elf64_Ehdr on 32-bit hosts, flockfile, license headers, style) and adds debuginfo support.
Best regards,
Ilya
Ilya Leoshkevich (1):
tcg: add perfmap and jitdump
accel/tcg/debuginfo.c | 108 +++++++++++++ accel/tcg/debuginfo.h | 54 +++++++ accel/tcg/meson.build | 2 + accel/tcg/perf.c | 333 ++++++++++++++++++++++++++++++++++++++ accel/tcg/perf.h | 28 ++++ accel/tcg/translate-all.c | 3 + docs/devel/tcg.rst | 20 +++ linux-user/elfload.c | 3 + linux-user/exit.c | 2 + linux-user/main.c | 15 ++ linux-user/meson.build | 1 + meson.build | 8 + qemu-options.hx | 20 +++ softmmu/vl.c | 11 ++ tcg/tcg.c | 2 + 15 files changed, 610 insertions(+) create mode 100644 accel/tcg/debuginfo.c create mode 100644 accel/tcg/debuginfo.h create mode 100644 accel/tcg/perf.c create mode 100644 accel/tcg/perf.h
Signature · 1 line
2.37.2
Quoted 12 lines · Ilya Leoshkevich
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 reportor
Quoted 7 lines
perf record -k 1 qemu-x86_64 -jitdump ./a.out perf inject -j -i perf.data -o perf.data.jitted perf report -i perf.data.jitted
Co-developed-by Vanderson M. do Rosario <[email protected]>Co-developed-by Alex Bennée <[email protected]>Signed-off-by Ilya Leoshkevich <[email protected]>
I think I remember this, and the question that was never answered was:
accel/tcg/translate-all.c TranslationBlock *tb_gen_code(CPUState *cpu, · line 1496
Quoted 5 lines
accel/tcg/translate-all.c +1
-0@@ -1492,6 +1493,8 @@ TranslationBlock *tb_gen_code(CPUState *cpu, } tb->tc.size = gen_code_size; +perf_report_code (gen_code_buf, gen_code_size, tb->icount, tb->pc);
When do_tb_flush is called, everything that is recorded in perfmap is invalidated. How do you tell perf about that?
r~
Quoted 31 lines · Richard Henderson
Quoted 14 lines · Ilya Leoshkevich
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 reportor
Quoted 7 lines
perf record -k 1 qemu-x86_64 -jitdump ./a.out perf inject -j -i perf.data -o perf.data.jitted perf report -i perf.data.jitted
Co-developed-by Vanderson M. do Rosario <[email protected]>Co-developed-by Alex Bennée <[email protected]>Signed-off-by Ilya Leoshkevich <[email protected]>I think I remember this, and the question that was never answered was:
Quoted 5 lines
accel/tcg/translate-all.c +0
-0@@ -1492,6 +1493,8 @@ TranslationBlock *tb_gen_code(CPUState *cpu, } tb->tc.size = gen_code_size;
+ perf_report_code(gen_code_buf,gen_code_size,tb->icount, tb-
Quoted 9 lines
Quoted 1 line
Quoted 1 line
pc);
When
do_tb_flushis called, everything that is recorded in perfmap is invalidated.
How do you tell perf about that?r~
I don't think we should handle this altogether. Perf already knows how to handle overlapping addresses:
$ cat main.c
#include <dlfcn.h>
#include <stdio.h>
int main()
{
void *d = dlopen("./lib1.so", RTLD_NOW);
void (*f)(void) = dlsym(d, "f");
printf("%p\n", f);
f();
dlclose(d);
d = dlopen("./lib2.so", RTLD_NOW);
f = dlsym(d, "f");
printf("%p\n", f);
f();
dlclose(d);
}$ cat 1.c static volatile int i = 0;
void f(void) {
while (i < 1000000000) {
i++;
}
}$ gcc -o main -O3 -g main.c -ldl $ gcc -o lib1.so -shared -fPIC -O3 -g 1.c $ cp 1.c 2.c $ gcc -o lib2.so -shared -fPIC -O3 -g 2.c
$ perf record ./main 0x7f47925eb100 0x7f47925eb100
$ perf report 49.96% main lib1.so [.] f 49.92% main lib2.so [.] f
Note that there is "load" event - PERF_RECORD_MMAP2, but no "unload" event. The old mappings are simply discarded by maps__fixup_overlappings() when something new is mapped.
In our case this means that as soon as we JIT new code, perf will see that it replaced the one that was flushed. Until then it will live with stale mappings, but there is no harm in that, since they are not executed.
Am 12.10.22 um 07:18 schrieb Ilya Leoshkevich:
Quoted 12 lines
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 reportor
Quoted 7 lines
perf record -k 1 qemu-x86_64 -jitdump ./a.out perf inject -j -i perf.data -o perf.data.jitted perf report -i perf.data.jitted
Co-developed-by Vanderson M. do Rosario <[email protected]>Co-developed-by Alex Bennée <[email protected]>Signed-off-by Ilya Leoshkevich <[email protected]>
I think this would be awesome to have. I know our performance people do use this for Java a lot.
Quoted 68 lines · Ilya Leoshkevich
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 reportor
perf record -k 1 qemu-x86_64 -jitdump ./a.out perf inject -j -i perf.data -o perf.data.jitted perf report -i perf.data.jitted
Co-developed-by Vanderson M. do Rosario <[email protected]>Co-developed-by Alex Bennée <[email protected]>Signed-off-by Ilya Leoshkevich <[email protected]>--- accel/tcg/debuginfo.c | 108 +++++++++++++ accel/tcg/debuginfo.h | 54 +++++++ accel/tcg/meson.build | 2 + accel/tcg/perf.c | 333 ++++++++++++++++++++++++++++++++++++++ accel/tcg/perf.h | 28 ++++ accel/tcg/translate-all.c | 3 + docs/devel/tcg.rst | 20 +++ linux-user/elfload.c | 3 + linux-user/exit.c | 2 + linux-user/main.c | 15 ++ linux-user/meson.build | 1 + meson.build | 8 + qemu-options.hx | 20 +++ softmmu/vl.c | 11 ++ tcg/tcg.c | 2 + 15 files changed, 610 insertions(+) create mode 100644 accel/tcg/debuginfo.c create mode 100644 accel/tcg/debuginfo.h create mode 100644 accel/tcg/perf.c create mode 100644 accel/tcg/perf.haccel/tcg/debuginfo.c new file +21
-0@@ -0,0 +1,108 @@ +/* +* Debug information support. +* +* SPDX-License-Identifier: GPL-2.0-or-later +*/ + +#include "qemu/osdep.h" + +#include <elfutils/libdwfl.h> + +#include "debuginfo.h" + +static QemuMutex lock; +static Dwfl *dwfl; +static const Dwfl_Callbacks dwfl_callbacks = { + .find_elf =NULL , + .find_debuginfo = dwfl_standard_find_debuginfo, + .section_address =NULL , + .debuginfo_path =NULL , +}; +
accel/tcg/debuginfo.c line 31
Quoted 10 lines
accel/tcg/debuginfo.c new file +10
-0+__attribute__((constructor)) +static void debuginfo_init (void ) +{ +qemu_mutex_init (&lock); +} + +bool debuginfo_report_elf (const char *image_name,int image_fd, +target_ulong load_bias) +{ +qemu_mutex_lock (&lock);
You can wrap this up with a QEMU_LOCK_GUARD(&lock) { and avoid having to catch all your exit cases.
Quoted 46 lines
accel/tcg/debuginfo.c new file +46
-0+ +if (dwfl ==NULL ) { + dwfl =dwfl_begin (&dwfl_callbacks); + }else { +dwfl_report_begin_add (dwfl); + } + +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + +dwfl_report_elf (dwfl, image_name, image_name, image_fd, load_bias, true); +dwfl_report_end (dwfl,NULL ,NULL ); +qemu_mutex_unlock (&lock); +return true; +} + +bool debuginfo_get_symbol (target_ulong address, +const char **symbol,target_ulong *offset) +{ +Dwfl_Module *dwfl_module; +GElf_Off dwfl_offset; +GElf_Sym dwfl_sym; + +qemu_mutex_lock (&lock); + +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_module =dwfl_addrmodule (dwfl, address); +if (dwfl_module ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + *symbol =dwfl_module_addrinfo (dwfl_module, address, &dwfl_offset, + &dwfl_sym,NULL ,NULL ,NULL ); +if (*symbol ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + *offset = dwfl_offset; +qemu_mutex_unlock (&lock);
accel/tcg/debuginfo.c line 87
Quoted 10 lines
accel/tcg/debuginfo.c new file +10
-0+return true; +} + +bool debuginfo_get_line (target_ulong address, +const char **file,int *line) +{ +Dwfl_Module *dwfl_module; +Dwfl_Line *dwfl_line; + +qemu_mutex_lock (&lock);
ditto.
Quoted 486 lines
accel/tcg/debuginfo.c new file +21
-0+ +if (dwfl ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_module =dwfl_addrmodule (dwfl, address); +if (dwfl_module ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + + dwfl_line =dwfl_module_getsrc (dwfl_module, address); +if (dwfl_line ==NULL ) { +qemu_mutex_unlock (&lock); +return false; + } + *file =dwfl_lineinfo (dwfl_line,NULL , line,0 ,NULL ,NULL ); +qemu_mutex_unlock (&lock); +return true; +}accel/tcg/debuginfo.h new file +54
-0@@ -0,0 +1,54 @@ +/* +* Debug information support. +* +* SPDX-License-Identifier: GPL-2.0-or-later +*/ + +#ifndef ACCEL_TCG_DEBUGINFO_H +#define ACCEL_TCG_DEBUGINFO_H + +#include "exec/cpu-defs.h" + +#ifdef CONFIG_LIBDW +/* +* Load debuginfo for the specified guest ELF image. +* Return true on success, false on failure. +*/ +bool debuginfo_report_elf (const char *image_name,int image_fd, +target_ulong load_bias); + +/* +* Find a symbol name associated with the specified guest PC. +* Return true on success, false if there is no associated symbol. +*/ +bool debuginfo_get_symbol (target_ulong address, +const char **symbol,target_ulong *offset); + +/* +* Find a line number associated with the specified guest PC. +* Return true on success, false if there is no associated line number. +*/ +bool debuginfo_get_line (target_ulong address, +const char **file,int *line); +#else +static inline bool debuginfo_report_elf (const char *image_name,int image_fd, +target_ulong load_bias) +{ +return false; +} + +static inline bool debuginfo_get_symbol (target_ulong address, +const char **symbol, +target_ulong *offset) +{ +return false; +} + +static inline bool debuginfo_get_line (target_ulong address, +const char **file,int *line) +{ +return false; +} +#endif + +#endif accel/tcg/meson.build +2
-0@@ -1,5 +1,6 @@ tcg_ss = ss.source_set() tcg_ss.add(files( + 'perf.c', 'tcg-all.c', 'cpu-exec-common.c', 'cpu-exec.c',@@ -11,6 +12,7 @@ tcg_ss.add(files( 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')) specific_ss.add_all(when: 'CONFIG_TCG', if_true: tcg_ss) specific_ss.add(when: ['CONFIG_SOFTMMU', 'CONFIG_TCG'], if_true: files(accel/tcg/perf.c new file +333
-0@@ -0,0 +1,333 @@ +/* +* 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 "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)); + } +} + +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 ]; +#ifdef CONFIG_LINUX +void *perf_marker; +#endif + +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 ; + } + +#ifdef CONFIG_LINUX +/* +* `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 ; + } +#endif + + 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 (void *start,size_t size) +{ +if (perfmap) { +fprintf (perfmap,"%" PRIxPTR" %zx tcg-prologue-buffer\n" , + (uintptr_t )start, size); + } +} + +/* +* Append a single line mapping to a JIT_CODE_DEBUG_INFO jitdump entry. +* Return 1 on success, 0 if there is no line number information for guest_pc. +*/ +static int append_debug_entry (GArray *raw,void *host_pc, +target_ulong guest_pc) +{ +struct debug_entry ent; +const char *file; +int line; + +if (!debuginfo_get_line (guest_pc, &file, &line)) { +return 0 ; + } + + ent.addr = (uint64_t )host_pc; + ent.lineno = line; + ent.discrim =0 ; +g_array_append_vals (raw, &ent,sizeof (ent)); +g_array_append_vals (raw, file,strlen (file) +1 ); +return 1 ; +} + +/* Write a JIT_CODE_DEBUG_INFO jitdump entry. */ +static void write_jr_code_debug_info (void *start,size_t size,int icount) +{ +GArray *raw =g_array_new (false, false,1 ); +struct jr_code_debug_info rec; +struct debug_entry ent; +target_ulong guest_pc; +void *host_pc; +int insn; + +/* Reserve space for the header. */ +g_array_set_size (raw,sizeof (rec)); + +/* Create debug entries. */ + rec.nr_entry =0 ; +for (insn =0 ; insn < icount; insn++) { + host_pc = start; +if (insn !=0 ) { + host_pc += tcg_ctx->gen_insn_end_off[insn -1 ]; + } + guest_pc = tcg_ctx->gen_insn_data[insn][0 ]; + rec.nr_entry +=append_debug_entry (raw, host_pc, guest_pc); + } + +/* Trailing debug_entry. */ + ent.addr = (uint64_t )start + size; + ent.lineno =0 ; + ent.discrim =0 ; +g_array_append_vals (raw, &ent,sizeof (ent)); +g_array_append_vals (raw,"" ,1 ); + rec.nr_entry++; + +/* Create header. */ + rec.p.id =JIT_CODE_DEBUG_INFO ; + rec.p.total_size = raw->len; + rec.p.timestamp =get_clock (); + rec.code_addr = (uint64_t )start; +memcpy (raw->data, &rec,sizeof (rec)); + +/* Flush. */ +fwrite (raw->data, raw->len,1 , jitdump); +g_array_unref (raw); +} + +/* Write a JIT_CODE_LOAD jitdump entry. */ +static void write_jr_code_load (void *start,size_t size, +const char *symbol,const char *suffix) +{ +static uint64_t code_index; +struct jr_code_load rec; +size_t suffix_size; +size_t name_size; + + name_size =strlen (symbol); + suffix_size =strlen (suffix) +1 ; + rec.p.id =JIT_CODE_LOAD ; + rec.p.total_size =sizeof (rec) + name_size + suffix_size + size; + rec.p.timestamp =get_clock (); + rec.pid =getpid (); + rec.tid =gettid (); + rec.vma = (uint64_t )start; + rec.code_addr = (uint64_t )start; + rec.code_size = size; + rec.code_index = code_index++; +fwrite (&rec,sizeof (rec),1 , jitdump); +fwrite (symbol, name_size,1 , jitdump); +fwrite (suffix, suffix_size,1 , jitdump); +fwrite (start, size,1 , jitdump); +} + +void perf_report_code (void *start,size_t size,int icount,uint64_t pc) +{ +char suffix[32 ] ="" ; +char symbol_buf[32 ]; +const char *symbol; +target_ulong offset; + +/* Symbolize guest PC. */ +if (perfmap || jitdump) { +if (!debuginfo_get_symbol (pc, &symbol, &offset)) { +snprintf (symbol_buf,sizeof (symbol_buf),"subject-%" PRIx64, pc); + symbol = symbol_buf; + offset =0 ; + } +if (offset !=0 ) { +snprintf (suffix,sizeof (suffix),"+0x%" PRIx64, (uint64_t )offset); + } + } + +/* Emit a perfmap entry if needed. */ +if (perfmap) { +flockfile (perfmap); +fprintf (perfmap,"%" PRIxPTR" %zx %s%s\n" , + (uintptr_t )start, size, symbol, suffix); +funlockfile (perfmap); + } + +/* Emit jitdump entries if needed. */ +if (jitdump) { +flockfile (jitdump); +write_jr_code_debug_info (start, size, icount); +write_jr_code_load (start, size, symbol, suffix); +funlockfile (jitdump); + } +} + +void perf_exit (void ) +{ +if (perfmap) { +fclose (perfmap); + perfmap =NULL ; + } + +if (jitdump) { +fclose (jitdump); + jitdump =NULL ; + } +}accel/tcg/perf.h new file +28
-0@@ -0,0 +1,28 @@ +/* +* 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 + +#include <stddef.h> +#include <stdint.h> + +/* 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 (void *start,size_t size); + +/* Add information about JITted guest code to profiler maps. */ +void perf_report_code (void *start,size_t size,int icount,uint64_t pc); + +/* Stop writing perf-<pid>.map and/or jit-<pid>.dump. */ +void perf_exit (void ); + +#endif accel/tcg/translate-all.c +1
-0@@ -62,6 +62,7 @@#include "tb-hash.h" #include "tb-context.h" #include "internal.h" +#include "perf.h" /* #define DEBUG_TB_INVALIDATE */ /* #define DEBUG_TB_FLUSH */
accel/tcg/translate-all.c TranslationBlock *tb_gen_code(CPUState *cpu, · line 1497
Quoted 5 lines
accel/tcg/translate-all.c +2
-0@@ -1492,6 +1493,8 @@ TranslationBlock *tb_gen_code(CPUState *cpu, } tb->tc.size = gen_code_size; +perf_report_code (gen_code_buf, gen_code_size, tb->icount, tb->pc); +
I think the recent code re-factoring means this needs updating. If its the guest pc I think that is already in a local variable.
Quoted 12 lines
accel/tcg/translate-all.c +0
-0#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);docs/devel/tcg.rst +1
-0@@ -188,3 +188,23 @@ 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
docs/devel/tcg.rst memory areas instead calls out to C code for device emulation. · line 201
Quoted 10 lines
docs/devel/tcg.rst +10
-0+--------------------- + +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.
I think this needs to be a bit clearer. Does jitdump only make sense if the guest has debug information. Running:
perf record -k 1 ./qemu-system-aarch64 -jitdump -monitor none \
-display none -chardev stdio,id=output \
-M virt -cpu max -display none \
-semihosting-config enable=on,target=native,chardev=output \
-kernel tests/tcg/aarch64-softmmu/memory
perf inject -j -i perf.data -o perf.data.jitted
perf report -i perf.data.jittedgives me pretty much the same report as the -perfmap. I think this is because we only properly look at the guest elf data for linux-user?
Quoted 213 lines
docs/devel/tcg.rst +9
-0+ +.. code:: + + perf record $QEMU -perfmap $REMAINING_ARGS + perf report + + perf record -k 1 $QEMU -jitdump $REMAINING_ARGS + perf inject -j -i perf.data -o perf.data.jitted + perf report -i perf.data.jittedlinux-user/elfload.c +3
-0@@ -19,6 +19,7 @@#include "qemu/selfmap.h" #include "qapi/error.h" #include "target_signal.h" +#include "accel/tcg/debuginfo.h" #ifdef _ARCH_PPC64#undef ARCH_DLINFO@@ -3261,6 +3262,8 @@ static void load_elf_image(const char *image_name, int image_fd,load_symbols (ehdr, image_fd, load_bias); } +debuginfo_report_elf (image_name, image_fd, load_bias); +mmap_unlock ();close (image_fd);linux-user/exit.c +2
-0@@ -17,6 +17,7 @@* along with this program; if not, see <http://www.gnu.org/licenses/>. */ #include "qemu/osdep.h" +#include "accel/tcg/perf.h" #include "exec/gdbstub.h" #include "qemu.h" #include "user-internals.h" @@ -38,4 +39,5 @@ void preexit_cleanup(CPUArchState *env, int code)#endif gdb_exit (code);qemu_plugin_user_exit (); +perf_exit (); }linux-user/main.c +15
-0@@ -53,6 +53,7 @@#include "signal-common.h" #include "loader.h" #include "user-mmap.h" +#include "accel/tcg/perf.h" #ifdef CONFIG_SEMIHOSTING #include "semihosting/semihost.h" @@ -423,6 +424,16 @@ static void handle_arg_abi_call0(const char *arg) }#endif +static void handle_arg_perfmap (const char *arg) +{ +perf_enable_perfmap (); +} + +static void handle_arg_jitdump (const char *arg) +{ +perf_enable_jitdump (); +} +static QemuPluginList plugins =QTAILQ_HEAD_INITIALIZER (plugins);#ifdef CONFIG_PLUGIN @@ -493,6 +504,10 @@ static const struct qemu_argument arg_table[] = { {"xtensa-abi-call0", "QEMU_XTENSA_ABI_CALL0", false, handle_arg_abi_call0, "", "assume CALL0 Xtensa ABI"}, #endif + {"perfmap", "QEMU_PERFMAP", false, handle_arg_perfmap, + "", "Generate a /tmp/perf-${pid}.map file for perf"}, + {"jitdump", "QEMU_JITDUMP", false, handle_arg_jitdump, + "", "Generate a jit-${pid}.dump file for perf"}, {NULL, NULL, false, NULL, NULL, NULL} };linux-user/meson.build +1
-0@@ -22,6 +22,7 @@ linux_user_ss.add(files( 'uname.c', )) linux_user_ss.add(rt) +linux_user_ss.add(libdw) linux_user_ss.add(when: 'TARGET_HAS_BFLT', if_true: files('flatload.c')) linux_user_ss.add(when: 'TARGET_I386', if_true: files('vm86.c'))meson.build +8
-0@@ -1631,6 +1631,12 @@ if libbpf.found() and not cc.links(''' endif endif +# libdw +libdw = dependency('libdw', + method: 'pkg-config', + kwargs: static_kwargs, + required: false) + ################# # config-host.h # #################@@ -1897,6 +1903,7 @@ config_host_data.set('CONFIG_DBUS_DISPLAY', dbus_display) config_host_data.set('CONFIG_CFI', get_option('cfi')) config_host_data.set('CONFIG_SELINUX', selinux.found()) config_host_data.set('CONFIG_XEN_BACKEND', xen.found()) +config_host_data.set('CONFIG_LIBDW', libdw.found()) if xen.found() # protect from xen.version() having less than three components xen_version = xen.version().split('.') + ['0', '0']@@ -3937,6 +3944,7 @@ summary_info += {'libudev': libudev} # Dummy dependency, keep .found() summary_info += {'FUSE lseek': fuse_lseek.found()} summary_info += {'selinux': selinux} +summary_info += {'libdw': libdw} summary(summary_info, bool_yn: true, section: 'Dependencies') if not supported_cpus.contains(cpu)qemu-options.hx +20
-0@@ -4799,6 +4799,26 @@ SRST Enable synchronization profiling. ERST +#ifdef CONFIG_TCG +DEF("perfmap", 0, QEMU_OPTION_perfmap, + "-perfmap generate a /tmp/perf-${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:)softmmu/vl.c +11
-0@@ -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" @@ -2900,6 +2903,14 @@ void qemu_init(int argc, char **argv)case QEMU_OPTION_DFILTER :qemu_set_dfilter_ranges (optarg, &error_fatal);break ; +#ifdef CONFIG_TCG +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 ;tcg/tcg.c +2
-0@@ -61,6 +61,7 @@#include "exec/log.h" #include "tcg/tcg-ldst.h" #include "tcg-internal.h" +#include "accel/tcg/perf.h" #ifdef CONFIG_TCG_INTERPRETER #include <ffi.h> @@ -749,6 +750,7 @@ 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),
Signature · 1 line
Alex Bennée