[PATCH] trace-cmd: Trace timesync to find pids that map vCPUs

[PATCH] trace-cmd: Trace timesync to find pids that map vCPUs

[Steven Rostedt]

From: "Steven Rostedt (VMware)" <rostedt@goodmis.org>

If qemu is not found, then there's no mapping between the vCPUs of the
guest and the threads that run them. As the time sync has the agent run on
all the guest's CPUs, trace it, looking for all kvm_entry, which state
which vCPU the host thread is entering on the guest, and then use that to
create the mapping for the data files.

Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
---
 tracecmd/include/trace-local.h |   2 +
 tracecmd/trace-record.c        | 196 ++++++++++++++++++++++++++++++++-
 2 files changed, 197 insertions(+), 1 deletion(-)

diff --git a/tracecmd/include/trace-local.h b/tracecmd/include/trace-local.h
index b3997d00..4c5669c9 100644
--- a/tracecmd/include/trace-local.h
+++ b/tracecmd/include/trace-local.h
@@ -298,11 +298,13 @@ void update_first_instance(struct buffer_instance *instance, int topt);
 void show_instance_file(struct buffer_instance *instance, const char *name);
 
 struct trace_guest {
+	struct tracefs_instance *instance;
 	char *name;
 	int cid;
 	int pid;
 	int cpu_max;
 	int *cpu_pid;
+	int *task_pids;
 };
 struct trace_guest *trace_get_guest(unsigned int cid, const char *name);
 bool trace_have_guests_pid(void);
diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
index 5dd8be4a..143ece6a 100644
--- a/tracecmd/trace-record.c
+++ b/tracecmd/trace-record.c
@@ -3230,6 +3230,38 @@ static int do_accept(int sd)
 	return -1;
 }
 
+/* Find all the tasks associated with the guest pid */
+static void find_tasks(struct trace_guest *guest)
+{
+	struct dirent *dent;
+	char *path;
+	DIR *dir;
+	int ret;
+	int tasks = 0;
+
+	ret = asprintf(&path, "/proc/%d/task", guest->pid);
+	if (ret < 0)
+		return;
+
+	dir = opendir(path);
+	free(path);
+	if (!dir)
+		return;
+
+	while ((dent = readdir(dir))) {
+		int *pids;
+		if (!(dent->d_type == DT_DIR && is_digits(dent->d_name)))
+			continue;
+		pids = realloc(guest->task_pids, sizeof(int) * (tasks + 2));
+		if (!pids)
+			break;
+		pids[tasks++] = strtol(dent->d_name, NULL, 0);
+		pids[tasks] = -1;
+		guest->task_pids = pids;
+	}
+	closedir(dir);
+}
+
 static char *parse_guest_name(char *gname, int *cid, int *port)
 {
 	struct trace_guest *guest;
@@ -3250,10 +3282,12 @@ static char *parse_guest_name(char *gname, int *cid, int *port)
 	} else if (is_digits(gname))
 		*cid = atoi(gname);
 
-	read_qemu_guests();
 	guest = trace_get_guest(*cid, gname);
 	if (guest) {
 		*cid = guest->cid;
+		/* Mapping not found, search for them */
+		if (!guest->cpu_pid)
+			find_tasks(guest);
 		return guest->name;
 	}
 
@@ -3712,6 +3746,162 @@ static int open_guest_fifos(const char *guest, int **fds)
 	return i;
 }
 
+struct trace_mapping {
+	struct tep_event		*kvm_entry;
+	struct tep_format_field		*vcpu_id;
+	struct tep_format_field		*common_pid;
+	int				*pids;
+	int				*map;
+	int				max_cpus;
+};
+
+static void start_mapping_vcpus(struct trace_guest *guest)
+{
+	char *pids = NULL;
+	char *t;
+	int len = 0;
+	int s;
+	int i;
+
+	if (!guest->task_pids)
+		return;
+
+	guest->instance = tracefs_instance_create("map_guest_pids");
+	if (!guest->instance)
+		return;
+
+	for (i = 0; guest->task_pids[i] >= 0; i++) {
+		s = snprintf(NULL, 0, "%d ", guest->task_pids[i]);
+		t = realloc(pids, len + s + 1);
+		if (!t) {
+			free(pids);
+			pids = NULL;
+			break;
+		}
+		pids = t;
+		sprintf(pids + len, "%d ", guest->task_pids[i]);
+		len += s;
+	}
+	if (pids) {
+		tracefs_instance_file_write(guest->instance, "set_event_pid", pids);
+		free(pids);
+	}
+	tracefs_instance_file_write(guest->instance, "events/kvm/kvm_entry/enable", "1");
+}
+
+static int map_vcpus(struct tep_event *event, struct tep_record *record,
+		     int cpu, void *context)
+{
+	struct trace_mapping *tmap = context;
+	unsigned long long val;
+	int type;
+	int pid;
+	int ret;
+	int i;
+
+	/* Do we have junk in the buffer? */
+	type = tep_data_type(event->tep, record);
+	if (type != tmap->kvm_entry->id)
+		return 0;
+
+	ret = tep_read_number_field(tmap->common_pid, record->data, &val);
+	if (ret < 0)
+		return 0;
+	pid = (int)val;
+
+	for (i = 0; tmap->pids[i] >= 0; i++) {
+		if (pid == tmap->pids[i])
+			break;
+	}
+	/* Is this thread one we care about ? */
+	if (tmap->pids[i] < 0)
+		return 0;
+
+	ret = tep_read_number_field(tmap->vcpu_id, record->data, &val);
+	if (ret < 0)
+		return 0;
+
+	cpu = (int)val;
+
+	/* Sanity check, warn? */
+	if (cpu >= tmap->max_cpus)
+		return 0;
+
+	/* Already have this one? Should we check if it is the same? */
+	if (tmap->map[cpu] >= 0)
+		return 0;
+
+	tmap->map[cpu] = pid;
+
+	/* Did we get them all */
+	for (i = 0; i < tmap->max_cpus; i++) {
+		if (tmap->map[i] < 0)
+			break;
+	}
+
+	return i == tmap->max_cpus;
+}
+
+static void stop_mapping_vcpus(struct buffer_instance *instance,
+			       struct trace_guest *guest)
+{
+	struct trace_mapping tmap = { };
+	struct tep_handle *tep;
+	const char *systems[] = { "kvm", NULL };
+	int i;
+
+	if (!guest->instance)
+		return;
+
+	tmap.pids = guest->task_pids;
+	tmap.max_cpus = instance->cpu_count;
+
+	tmap.map = malloc(sizeof(*tmap.map) * tmap.max_cpus);
+	if (!tmap.map)
+		return;
+
+	for (i = 0; i < tmap.max_cpus; i++)
+		tmap.map[i] = -1;
+
+	tracefs_instance_file_write(guest->instance, "events/kvm/kvm_entry/enable", "0");
+
+	tep = tracefs_local_events_system(NULL, systems);
+	if (!tep)
+		goto out;
+
+	tmap.kvm_entry = tep_find_event_by_name(tep, "kvm", "kvm_entry");
+	if (!tmap.kvm_entry)
+		goto out_free;
+
+	tmap.vcpu_id = tep_find_field(tmap.kvm_entry, "vcpu_id");
+	if (!tmap.vcpu_id)
+		goto out_free;
+
+	tmap.common_pid = tep_find_any_field(tmap.kvm_entry, "common_pid");
+	if (!tmap.common_pid)
+		goto out_free;
+
+	tracefs_iterate_raw_events(tep, guest->instance, NULL, 0, map_vcpus, &tmap);
+
+	for (i = 0; i < tmap.max_cpus; i++) {
+		if (tmap.map[i] < 0)
+			break;
+	}
+	/* We found all the mapped CPUs */
+	if (i == tmap.max_cpus) {
+		guest->cpu_pid = tmap.map;
+		guest->cpu_max = tmap.max_cpus;
+		tmap.map = NULL;
+	}
+
+ out_free:
+	tep_free(tep);
+ out:
+	free(tmap.map);
+	tracefs_instance_destroy(guest->instance);
+	tracefs_instance_free(guest->instance);
+}
+
 static int host_tsync(struct common_record_context *ctx,
 		      struct buffer_instance *instance,
 		      unsigned int tsync_port, char *proto)
@@ -3724,11 +3914,15 @@ static int host_tsync(struct common_record_context *ctx,
 	if (guest == NULL)
 		return -1;
 
+	start_mapping_vcpus(guest);
+
 	instance->tsync = tracecmd_tsync_with_guest(top_instance.trace_id,
 						    instance->tsync_loop_interval,
 						    instance->cid, tsync_port,
 						    guest->pid, instance->cpu_count,
 						    proto, ctx->clock);
+	stop_mapping_vcpus(instance, guest);
+
 	if (!instance->tsync)
 		return -1;
 
-- 
2.29.2

Re: [PATCH] trace-cmd: Trace timesync to find pids that map vCPUs

[Tzvetomir Stoyanov]

On Fri, May 7, 2021 at 8:56 PM Steven Rostedt <rostedt@goodmis.org> wrote:
>
> From: "Steven Rostedt (VMware)" <rostedt@goodmis.org>
>
> If qemu is not found, then there's no mapping between the vCPUs of the
> guest and the threads that run them. As the time sync has the agent run on
> all the guest's CPUs, trace it, looking for all kvm_entry, which state
> which vCPU the host thread is entering on the guest, and then use that to
> create the mapping for the data files.
>
> Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
> ---
>  tracecmd/include/trace-local.h |   2 +
>  tracecmd/trace-record.c        | 196 ++++++++++++++++++++++++++++++++-
>  2 files changed, 197 insertions(+), 1 deletion(-)
>
> diff --git a/tracecmd/include/trace-local.h b/tracecmd/include/trace-local.h
> index b3997d00..4c5669c9 100644
> --- a/tracecmd/include/trace-local.h
> +++ b/tracecmd/include/trace-local.h
> @@ -298,11 +298,13 @@ void update_first_instance(struct buffer_instance *instance, int topt);
>  void show_instance_file(struct buffer_instance *instance, const char *name);
>
>  struct trace_guest {
> +       struct tracefs_instance *instance;
>         char *name;
>         int cid;
>         int pid;
>         int cpu_max;
>         int *cpu_pid;
> +       int *task_pids;
>  };
>  struct trace_guest *trace_get_guest(unsigned int cid, const char *name);
>  bool trace_have_guests_pid(void);
> diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
> index 5dd8be4a..143ece6a 100644
> --- a/tracecmd/trace-record.c
> +++ b/tracecmd/trace-record.c
> @@ -3230,6 +3230,38 @@ static int do_accept(int sd)
>         return -1;
>  }
>
> +/* Find all the tasks associated with the guest pid */
> +static void find_tasks(struct trace_guest *guest)
> +{
> +       struct dirent *dent;
> +       char *path;
> +       DIR *dir;
> +       int ret;
> +       int tasks = 0;
> +
> +       ret = asprintf(&path, "/proc/%d/task", guest->pid);
> +       if (ret < 0)
> +               return;
> +
> +       dir = opendir(path);
> +       free(path);
> +       if (!dir)
> +               return;
> +
> +       while ((dent = readdir(dir))) {
> +               int *pids;
> +               if (!(dent->d_type == DT_DIR && is_digits(dent->d_name)))
> +                       continue;
> +               pids = realloc(guest->task_pids, sizeof(int) * (tasks + 2));
> +               if (!pids)
> +                       break;
> +               pids[tasks++] = strtol(dent->d_name, NULL, 0);
> +               pids[tasks] = -1;
> +               guest->task_pids = pids;
> +       }
> +       closedir(dir);
> +}
> +
>  static char *parse_guest_name(char *gname, int *cid, int *port)
>  {
>         struct trace_guest *guest;
> @@ -3250,10 +3282,12 @@ static char *parse_guest_name(char *gname, int *cid, int *port)
>         } else if (is_digits(gname))
>                 *cid = atoi(gname);
>
> -       read_qemu_guests();

I would suggest to leave the legacy logic for reading existing qemu
VMs. We can reimplement it - instead of parsing the qemu command line,
virhs can be used to get that information, as Stefano Garzarella
suggested in other thread. We can even implement similar auto
discovery logics for other VM emulators. The VMs auto discovery will
be a best effort, if the user passes "-A <guest>" to the command line,
and a guest with that name is not known - trace-cmd will die,
suggesting  to use the more robust "-A <guest>@cid:port" method.

[...]


--
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center