On Thu, Feb 27, 2020 at 7:23 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Thu, 27 Feb 2020 16:19:49 +0200
> "Tzvetomir Stoyanov (VMware)" <tz.stoyanov@xxxxxxxxx> wrote:
>
> > From: Tzvetomir Stoyanov <tstoyanov@xxxxxxxxxx>
> >
> > In order to match host and guest events, a mapping between guest VCPU
> > and the host task, running this VCPU is needed. Extended existing
> > struct guest to hold such mapping and added logic in read_qemu_guests()
> > function to initialize it. Implemented a new internal API,
> > get_guest_vcpu_pid(), to retrieve VCPU-task mapping for given VM.
> >
> > Signed-off-by: Tzvetomir Stoyanov <tstoyanov@xxxxxxxxxx>
> > ---
> > tracecmd/include/trace-local.h | 2 +
> > tracecmd/trace-record.c | 77 ++++++++++++++++++++++++++++++++++
> > 2 files changed, 79 insertions(+)
> >
> > diff --git a/tracecmd/include/trace-local.h b/tracecmd/include/trace-local.h
> > index 29f27793..a5cf0640 100644
> > --- a/tracecmd/include/trace-local.h
> > +++ b/tracecmd/include/trace-local.h
> > @@ -247,6 +247,8 @@ void update_first_instance(struct buffer_instance *instance, int topt);
> >
> > void show_instance_file(struct buffer_instance *instance, const char *name);
> >
> > +int get_guest_vcpu_pid(unsigned int guest_cid, unsigned int guest_vcpu);
> > +
> > /* moved from trace-cmd.h */
> > void tracecmd_create_top_instance(char *name);
> > void tracecmd_remove_instances(void);
> > diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
> > index 28fe31b7..4370c964 100644
> > --- a/tracecmd/trace-record.c
> > +++ b/tracecmd/trace-record.c
> > @@ -3035,11 +3035,30 @@ struct guest {
> > char *name;
> > int cid;
> > int pid;
> > + int cpu_max;
> > + int *cpu_pid;
> > };
> >
> > static struct guest *guests;
> > static size_t guests_len;
> >
> > +static int set_vcpu_pid_mapping(struct guest *guest, int cpu, int pid)
> > +{
> > + int *cpu_pid;
> > +
> > + if (cpu < 0 || pid < 0)
> > + return -1;
>
> This check makes the check before its call not needed (see below).
>
> > + if (cpu >= guest->cpu_max) {
> > + cpu_pid = realloc(guest->cpu_pid, (cpu + 1) * sizeof(int));
>
> It is possible that the cpu numbers may be sparse, which means we should we
> should probably initialize the new numbers as...
>
> > + if (!cpu_pid)
> > + return -1;
>
> /* Handle sparse CPU numbers */
> for (i = guest->cpu_max; i < cpu; i++)
> guest->cpu_pid[i] = -1;
>
> Note, the above wont loop at all if there's no sparse CPUs (missing
> numbers).
>
> > + guest->cpu_max = cpu + 1;
> > + guest->cpu_pid = cpu_pid;
> > + }
> > + guest->cpu_pid[cpu] = pid;
> > + return 0;
> > +}
> > +
> > static char *get_qemu_guest_name(char *arg)
> > {
> > char *tok, *end = arg;
> > @@ -3052,6 +3071,46 @@ static char *get_qemu_guest_name(char *arg)
> > return arg;
> > }
> >
> > +static void read_qemu_guests_pids(char *guest_task, struct guest *guest)
>
> Probably should return a status.
>
> > +{
> > + struct dirent *entry;
> > + char path[PATH_MAX];
> > + char *buf = NULL;
> > + size_t n = 0;
> > + long int vcpu;
> > + long int pid;
>
> "int" is not needed. Just "long" is good enough.
>
> long vcpu;
> long pid;
>
> Although, I doubt there will be more than 4 billion of either of these, so
> they should probably be just "int". Which means you don't need the INT_MAX
> checks.
>
The reason I use long is strtol(), which returns long integer. In case
of an error, this API returns LONG_MAX or LONG_MIN.
The set_vcpu_pid_mapping() accepts int, that's why I put these checks
- to filter possible strtol() parsing errors and int overflow.
I'll keep the checks and remove the duplicated ones in set_vcpu_pid_mapping()
> > + DIR *dir;
> > + FILE *f;
> > +
> > + snprintf(path, sizeof(path), "/proc/%s/task", guest_task);
> > + dir = opendir(path);
> > + if (!dir)
> > + return;
> > +
> > + while ((entry = readdir(dir))) {
> > + if (!(entry->d_type == DT_DIR && is_digits(entry->d_name)))
> > + continue;
> > +
> > + snprintf(path, sizeof(path), "/proc/%s/task/%s/comm",
> > + guest_task, entry->d_name);
> > + f = fopen(path, "r");
> > + if (!f)
> > + continue;
> > +
> > + if (getline(&buf, &n, f) >= 0 &&
> > + strncmp(buf, "CPU ", 4) == 0) {
> > + vcpu = strtol(buf + 4, NULL, 10);
> > + pid = strtol(entry->d_name, NULL, 10);
> > + if (vcpu < INT_MAX && pid < INT_MAX &&
> > + vcpu >= 0 && pid >= 0)
>
> The vcpu >= 0 && pid >= 0 are not needed due to the check in the function.
>
> > + set_vcpu_pid_mapping(guest, vcpu, pid);
>
> Probably should check the return status of this function, and report a
> warning if it fails.
>
> > + }
> > +
> > + fclose(f);
> > + }
> > + free(buf);
> > +}
> > +
> > static void read_qemu_guests(void)
> > {
> > static bool initialized;
> > @@ -3115,6 +3174,8 @@ static void read_qemu_guests(void)
> > if (!is_qemu)
> > goto next;
> >
> > + read_qemu_guests_pids(entry->d_name, &guest);
>
> Should probably check the status of the above function. Die on error?
I'll put a warning, as this failure is not critical - the guest
tracing will still work, even without this mapping.
Thanks, Steven!
>
> > +
> > guests = realloc(guests, (guests_len + 1) * sizeof(*guests));
> > if (!guests)
> > die("Can not allocate guest buffer");
> > @@ -3160,6 +3221,22 @@ static char *parse_guest_name(char *guest, int *cid, int *port)
> > return guest;
> > }
> >
> > +int get_guest_vcpu_pid(unsigned int guest_cid, unsigned int guest_vcpu)
> > +{
> > + int i;
> > +
> > + if (!guests)
> > + return -1;
> > +
> > + for (i = 0; i < guests_len; i++) {
> > + if (!guests[i].cpu_pid || guest_vcpu >= guests[i].cpu_max)
>
> As a cpu_pid may be zero (unlikely), and the code I showed above
> initialized sparse cpus as -1, then this should be guests[i].cpu_pid < 0.
>
> -- Steve
>
> > + continue;
> > + if (guest_cid == guests[i].cid)
> > + return guests[i].cpu_pid[guest_vcpu];
> > + }
> > + return -1;
> > +}
> > +
> > static void set_prio(int prio)
> > {
> > struct sched_param sp;
>
--
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center
|