[PATCH] lib/igt_drm_clients: lazy stat process
Tvrtko Ursulin
tursulin at ursulin.net
Thu May 2 10:20:48 UTC 2024
On 01/05/2024 18:38, Lucas De Marchi wrote:
> Only try to get task data like pid and name when that is required.
> Typically most of the processes in a system don't have a drm fd open.
> So only open and use /proc/<pid>/stat when it's needed.
>
> For the data below, system is running with just a single drm client:
>
> Before:
> Performance counter stats for './build/tools/gputop -d 0.1 -n 10':
>
> 507.19 msec task-clock # 0.334 CPUs utilized
> 29 context-switches # 57.178 /sec
> 4 cpu-migrations # 7.887 /sec
> 75 page-faults # 147.874 /sec
> (...)
> 1.518253637 seconds time elapsed
> 0.007989000 seconds user
> 0.503351000 seconds sys
>
> After:
> Performance counter stats for './build/tools/gputop -d 0.1 -n 10':
>
> 384.01 msec task-clock # 0.275 CPUs utilized
> 17 context-switches # 44.270 /sec
> 1 cpu-migrations # 2.604 /sec
> 74 page-faults # 192.705 /sec
> (...)
> 1.397640800 seconds time elapsed
> 0.008106000 seconds user
> 0.385048000 seconds sys
>
> Looking only at the execution time of igt_drm_clients_scan() over 10
> runs.
> Before: 47138890 nsec
> After: 33978388 nsec
>
> So in both methods it shows a ~25% speedup. It is expected that in a
> system with more clients this won't be as high.
>
> Signed-off-by: Lucas De Marchi <lucas.demarchi at intel.com>
> ---
>
> Extracted as is from https://patchwork.freedesktop.org/series/132059/ as
> not really related.
>
> lib/igt_drm_clients.c | 50 +++++++++++++++++++++----------------------
> 1 file changed, 25 insertions(+), 25 deletions(-)
>
> diff --git a/lib/igt_drm_clients.c b/lib/igt_drm_clients.c
> index 40d494441..ab0c2cec2 100644
> --- a/lib/igt_drm_clients.c
> +++ b/lib/igt_drm_clients.c
> @@ -368,26 +368,33 @@ static size_t readat2buf(int at, const char *name, char *buf, const size_t sz)
> }
> }
>
> -static bool get_task_name(const char *buffer, char *out, unsigned long sz)
> +static void get_task_data(int pid_dir, unsigned int *pid, char *task, size_t tasksz)
> {
> - char *s = index(buffer, '(');
> - char *e = rindex(buffer, ')');
> - unsigned int len;
> + char buf[4096];
> + char *s, *e;
> + size_t len;
>
> + if (!readat2buf(pid_dir, "stat", buf, sizeof(buf)))
> + return;
> +
> + s = strchr(buf, '(');
> + e = strchr(s, ')');
> if (!s || !e)
> - return false;
> - assert(e >= s);
> + return;
>
> len = e - ++s;
> - if(!len || (len + 1) >= sz)
> - return false;
> + if (!len)
> + return;
>
> - strncpy(out, s, len);
> - out[len] = 0;
> + if (len + 1 > tasksz)
> + len = tasksz - 1;
>
> - return true;
> + strncpy(task, s, len);
> + task[len] = 0;
> + *pid = atoi(buf);
> }
>
> +
> static bool is_drm_fd(int fd_dir, const char *name, unsigned int *minor)
> {
> struct stat stat;
> @@ -481,13 +488,11 @@ igt_drm_clients_scan(struct igt_drm_clients *clients,
> return clients;
>
> while ((proc_dent = readdir(proc_dir)) != NULL) {
> - unsigned int client_pid, minor = 0;
> + unsigned int client_pid = 0, minor = 0;
> int pid_dir = -1, fd_dir = -1;
> struct dirent *fdinfo_dent;
> char client_name[64] = { };
> DIR *fdinfo_dir = NULL;
> - char buf[4096];
> - size_t count;
>
> if (proc_dent->d_type != DT_DIR)
> continue;
> @@ -499,17 +504,6 @@ igt_drm_clients_scan(struct igt_drm_clients *clients,
> if (pid_dir < 0)
> continue;
>
> - count = readat2buf(pid_dir, "stat", buf, sizeof(buf));
> - if (!count)
> - goto next;
> -
> - client_pid = atoi(buf);
> - if (!client_pid)
> - goto next;
> -
> - if (!get_task_name(buf, client_name, sizeof(client_name)))
> - goto next;
> -
> fd_dir = openat(pid_dir, "fd", O_DIRECTORY | O_RDONLY);
> if (fd_dir < 0)
> goto next;
> @@ -542,6 +536,12 @@ igt_drm_clients_scan(struct igt_drm_clients *clients,
> minor, info.id))
> continue; /* Skip duplicate fds. */
>
> + if (!client_pid) {
> + get_task_data(pid_dir, &client_pid, client_name,
> + sizeof(client_name));
> + assert(client_pid > 0);
> + }
> +
> c = igt_drm_clients_find(clients, IGT_DRM_CLIENT_PROBE,
> minor, info.id);
> if (!c)
Nice improvement and well spotted!
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at igalia.com>
Regards,
Tvrtko
More information about the igt-dev
mailing list