qemu/contrib/plugins/lockstep.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

357 lines
9.7 KiB
C
Raw Normal View History

plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
/*
* Lockstep Execution Plugin
*
* Allows you to execute two QEMU instances in lockstep and report
* when their execution diverges. This is mainly useful for developers
* who want to see where a change to TCG code generation has
* introduced a subtle and hard to find bug.
*
* Caveats:
* - single-threaded linux-user apps only with non-deterministic syscalls
* - no MTTCG enabled system emulation (icount may help)
*
* While icount makes things more deterministic it doesn't mean a
* particular run may execute the exact same sequence of blocks. An
* asynchronous event (for example X11 graphics update) may cause a
* block to end early and a new partial block to start. This means
* serial only test cases are a better bet. -d nochain may also help.
*
* This code is not thread safe!
*
* Copyright (c) 2020 Linaro Ltd
*
* SPDX-License-Identifier: GPL-2.0-or-later
*/
#include <glib.h>
#include <inttypes.h>
#include <unistd.h>
#include <sys/socket.h>
#include <sys/un.h>
#include <stdio.h>
#include <errno.h>
#include <qemu-plugin.h>
QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION;
/* saved so we can uninstall later */
static qemu_plugin_id_t our_id;
static unsigned long bb_count;
static unsigned long insn_count;
/* Information about a translated block */
typedef struct {
uint64_t pc;
uint64_t insns;
} BlockInfo;
/* Information about an execution state in the log */
typedef struct {
BlockInfo *block;
unsigned long insn_count;
unsigned long block_count;
} ExecInfo;
/* The execution state we compare */
typedef struct {
uint64_t pc;
unsigned long insn_count;
} ExecState;
typedef struct {
GSList *log_pos;
int distance;
} DivergeState;
/* list of translated block info */
static GSList *blocks;
/* execution log and points of divergence */
static GSList *log, *divergence_log;
static int socket_fd;
static char *path_to_unlink;
static bool verbose;
static void plugin_cleanup(qemu_plugin_id_t id)
{
/* Free our block data */
g_slist_free_full(blocks, &g_free);
g_slist_free_full(log, &g_free);
g_slist_free(divergence_log);
close(socket_fd);
if (path_to_unlink) {
unlink(path_to_unlink);
}
}
static void plugin_exit(qemu_plugin_id_t id, void *p)
{
g_autoptr(GString) out = g_string_new("No divergence :-)\n");
g_string_append_printf(out, "Executed %ld/%d blocks\n",
bb_count, g_slist_length(log));
g_string_append_printf(out, "Executed ~%ld instructions\n", insn_count);
qemu_plugin_outs(out->str);
plugin_cleanup(id);
}
static void report_divergance(ExecState *us, ExecState *them)
{
DivergeState divrec = { log, 0 };
g_autoptr(GString) out = g_string_new("");
bool diverged = false;
/*
* If we have diverged before did we get back on track or are we
* totally loosing it?
*/
if (divergence_log) {
DivergeState *last = (DivergeState *) divergence_log->data;
GSList *entry;
for (entry = log; g_slist_next(entry); entry = g_slist_next(entry)) {
if (entry == last->log_pos) {
break;
}
divrec.distance++;
}
/*
* If the last two records are so close it is likely we will
* not recover synchronisation with the other end.
*/
if (divrec.distance == 1 && last->distance == 1) {
diverged = true;
}
}
divergence_log = g_slist_prepend(divergence_log,
g_memdup(&divrec, sizeof(divrec)));
/* Output short log entry of going out of sync... */
if (verbose || divrec.distance == 1 || diverged) {
g_string_printf(out, "@ 0x%016lx vs 0x%016lx (%d/%d since last)\n",
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
us->pc, them->pc, g_slist_length(divergence_log),
divrec.distance);
qemu_plugin_outs(out->str);
}
if (diverged) {
int i;
GSList *entry;
g_string_printf(out, "Δ insn_count @ 0x%016lx (%ld) vs 0x%016lx (%ld)\n",
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
us->pc, us->insn_count, them->pc, them->insn_count);
for (entry = log, i = 0;
g_slist_next(entry) && i < 5;
entry = g_slist_next(entry), i++) {
ExecInfo *prev = (ExecInfo *) entry->data;
g_string_append_printf(out,
" previously @ 0x%016lx/%ld (%ld insns)\n",
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
prev->block->pc, prev->block->insns,
prev->insn_count);
}
qemu_plugin_outs(out->str);
qemu_plugin_outs("too much divergence... giving up.");
qemu_plugin_uninstall(our_id, plugin_cleanup);
}
}
static void vcpu_tb_exec(unsigned int cpu_index, void *udata)
{
BlockInfo *bi = (BlockInfo *) udata;
ExecState us, them;
ssize_t bytes;
ExecInfo *exec;
us.pc = bi->pc;
us.insn_count = insn_count;
/*
* Write our current position to the other end. If we fail the
* other end has probably died and we should shut down gracefully.
*/
bytes = write(socket_fd, &us, sizeof(ExecState));
if (bytes < sizeof(ExecState)) {
qemu_plugin_outs(bytes < 0 ?
"problem writing to socket" :
"wrote less than expected to socket");
qemu_plugin_uninstall(our_id, plugin_cleanup);
return;
}
/*
* Now read where our peer has reached. Again a failure probably
* indicates the other end died and we should close down cleanly.
*/
bytes = read(socket_fd, &them, sizeof(ExecState));
if (bytes < sizeof(ExecState)) {
qemu_plugin_outs(bytes < 0 ?
"problem reading from socket" :
"read less than expected");
qemu_plugin_uninstall(our_id, plugin_cleanup);
return;
}
/*
* Compare and report if we have diverged.
*/
if (us.pc != them.pc) {
report_divergance(&us, &them);
}
/*
* Assume this block will execute fully and record it
* in the execution log.
*/
insn_count += bi->insns;
bb_count++;
exec = g_new0(ExecInfo, 1);
exec->block = bi;
exec->insn_count = insn_count;
exec->block_count = bb_count;
log = g_slist_prepend(log, exec);
}
static void vcpu_tb_trans(qemu_plugin_id_t id, struct qemu_plugin_tb *tb)
{
BlockInfo *bi = g_new0(BlockInfo, 1);
bi->pc = qemu_plugin_tb_vaddr(tb);
bi->insns = qemu_plugin_tb_n_insns(tb);
/* save a reference so we can free later */
blocks = g_slist_prepend(blocks, bi);
qemu_plugin_register_vcpu_tb_exec_cb(tb, vcpu_tb_exec,
QEMU_PLUGIN_CB_NO_REGS, (void *)bi);
}
/*
* Instead of encoding master/slave status into what is essentially
* two peers we shall just take the simple approach of checking for
* the existence of the pipe and assuming if it's not there we are the
* first process.
*/
static bool setup_socket(const char *path)
{
struct sockaddr_un sockaddr;
int fd;
fd = socket(AF_UNIX, SOCK_STREAM, 0);
if (fd < 0) {
perror("create socket");
return false;
}
sockaddr.sun_family = AF_UNIX;
g_strlcpy(sockaddr.sun_path, path, sizeof(sockaddr.sun_path) - 1);
if (bind(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr)) < 0) {
perror("bind socket");
close(fd);
return false;
}
/* remember to clean-up */
path_to_unlink = g_strdup(path);
if (listen(fd, 1) < 0) {
perror("listen socket");
close(fd);
return false;
}
socket_fd = accept(fd, NULL, NULL);
if (socket_fd < 0 && errno != EINTR) {
perror("accept socket");
close(fd);
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
return false;
}
qemu_plugin_outs("setup_socket::ready\n");
close(fd);
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
return true;
}
static bool connect_socket(const char *path)
{
int fd;
struct sockaddr_un sockaddr;
fd = socket(AF_UNIX, SOCK_STREAM, 0);
if (fd < 0) {
perror("create socket");
return false;
}
sockaddr.sun_family = AF_UNIX;
g_strlcpy(sockaddr.sun_path, path, sizeof(sockaddr.sun_path) - 1);
if (connect(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr)) < 0) {
perror("failed to connect");
close(fd);
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
return false;
}
qemu_plugin_outs("connect_socket::ready\n");
socket_fd = fd;
return true;
}
static bool setup_unix_socket(const char *path)
{
if (g_file_test(path, G_FILE_TEST_EXISTS)) {
return connect_socket(path);
} else {
return setup_socket(path);
}
}
QEMU_PLUGIN_EXPORT int qemu_plugin_install(qemu_plugin_id_t id,
const qemu_info_t *info,
int argc, char **argv)
{
int i;
g_autofree char *sock_path = NULL;
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
for (i = 0; i < argc; i++) {
char *p = argv[i];
g_autofree char **tokens = g_strsplit(p, "=", 2);
if (g_strcmp0(tokens[0], "verbose") == 0) {
if (!qemu_plugin_bool_parse(tokens[0], tokens[1], &verbose)) {
fprintf(stderr, "boolean argument parsing failed: %s\n", p);
return -1;
}
} else if (g_strcmp0(tokens[0], "sockpath") == 0) {
sock_path = tokens[1];
} else {
fprintf(stderr, "option parsing failed: %s\n", p);
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
return -1;
}
}
if (sock_path == NULL) {
fprintf(stderr, "Need a socket path to talk to other instance.\n");
return -1;
}
if (!setup_unix_socket(sock_path)) {
fprintf(stderr, "Failed to setup socket for communications.\n");
return -1;
}
plugins: new lockstep plugin for debugging TCG changes When we make changes to the TCG we sometimes cause regressions that are deep into the execution cycle of the guest. Debugging this often requires comparing large volumes of trace information to figure out where behaviour has diverged. The lockstep plugin utilises a shared socket so two QEMU's running with the plugin will write their current execution position and wait to receive the position of their partner process. When execution diverges the plugins output where they were and the previous few blocks before unloading themselves and letting execution continue. Originally I planned for this to be most useful with -icount but it turns out you can get divergence pretty quickly due to asynchronous qemu_cpu_kick_rr_cpus() events causing one side to eventually run into a short block a few cycles before the other side. For this reason I've added a bit of tracking and I think the divergence reporting could be finessed to report only if we really start to diverge in execution. An example run would be: qemu-system-sparc -monitor none -parallel none -net none \ -M SS-20 -m 256 -kernel day11/zImage.elf \ -plugin ./tests/plugin/liblockstep.so,arg=lockstep-sparc.sock \ -d plugin,nochain with an identical command in another window in the same working directory. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Robert Foley <robert.foley@linaro.org> Tested-by: Robert Foley <robert.foley@linaro.org> Cc: Richard Henderson <richard.henderson@linaro.org> Cc: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> Message-Id: <20200610155509.12850-3-alex.bennee@linaro.org>
2020-06-10 18:55:05 +03:00
our_id = id;
qemu_plugin_register_vcpu_tb_trans_cb(id, vcpu_tb_trans);
qemu_plugin_register_atexit_cb(id, plugin_exit, NULL);
return 0;
}