[dpdk-stable] [PATCH v4] test: optimise fd closing in forked test process

[dpdk-stable] [PATCH v4] test: optimise fd closing in forked test process

[David Marchand]

From: Krzysztof Kanas <kkanas@marvell.com>

Caught while investigating timeouts on a ARM64 server.

Stracing a test process running the eal_flags_autotest, we can see that
the fork helper is checking all possible file descriptors from
getdtablesize() to 2, and close the existing ones.
We can do better by inspecting this forked process /proc/self/fd
directory.

Besides, checking file descriptors via /proc/self/fd only makes sense for
Linux. This code was a noop on FreeBSD.

Fixes: af75078fece3 ("first public release")
Cc: stable@dpdk.org

Signed-off-by: Krzysztof Kanas <kkanas@marvell.com>
Signed-off-by: David Marchand <david.marchand@redhat.com>
---
Changelog since v3:
- rewrote commit log message,
- restricted the fix to Linux only,

---
 app/test/process.h | 50 ++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 44 insertions(+), 6 deletions(-)

diff --git a/app/test/process.h b/app/test/process.h
index 128ce41..191d279 100644
--- a/app/test/process.h
+++ b/app/test/process.h
@@ -11,6 +11,7 @@
 #include <stdlib.h> /* NULL */
 #include <string.h> /* strerror */
 #include <unistd.h> /* readlink */
+#include <dirent.h>
 #include <sys/wait.h>
 
 #include <rte_string_fns.h> /* strlcpy */
@@ -40,7 +41,7 @@ process_dup(const char *const argv[], int numargs, const char *env_value)
 {
 	int num;
 	char *argv_cpy[numargs + 1];
-	int i, fd, status;
+	int i, status;
 	char path[32];
 #ifdef RTE_LIBRTE_PDUMP
 	pthread_t thread;
@@ -56,13 +57,50 @@ process_dup(const char *const argv[], int numargs, const char *env_value)
 		argv_cpy[i] = NULL;
 		num = numargs;
 
-		/* close all open file descriptors, check /proc/self/fd to only
-		 * call close on open fds. Exclude fds 0, 1 and 2*/
-		for (fd = getdtablesize(); fd > 2; fd-- ) {
-			snprintf(path, sizeof(path), "/proc/" exe "/fd/%d", fd);
-			if (access(path, F_OK) == 0)
+#ifdef RTE_EXEC_ENV_LINUX
+		{
+			const char *procdir = "/proc/" self "/fd/";
+			struct dirent *dirent;
+			char *endptr;
+			int fd, fdir;
+			DIR *dir;
+
+			/* close all open file descriptors, check /proc/self/fd
+			 * to only call close on open fds. Exclude fds 0, 1 and
+			 * 2
+			 */
+			dir = opendir(procdir);
+			if (dir == NULL) {
+				rte_panic("Error opening %s: %s\n", procdir,
+						strerror(errno));
+			}
+
+			fdir = dirfd(dir);
+			if (fdir < 0) {
+				status = errno;
+				closedir(dir);
+				rte_panic("Error %d obtaining fd for dir %s: %s\n",
+						fdir, procdir,
+						strerror(status));
+			}
+
+			while ((dirent = readdir(dir)) != NULL) {
+				errno = 0;
+				fd = strtol(dirent->d_name, &endptr, 10);
+				if (errno != 0 || endptr[0] != '\0') {
+					printf("Error converting name fd %d %s:\n",
+						fd, dirent->d_name);
+					continue;
+				}
+
+				if (fd == fdir || fd <= 2)
+					continue;
+
 				close(fd);
+			}
+			closedir(dir);
 		}
+#endif
 		printf("Running binary with argv[]:");
 		for (i = 0; i < num; i++)
 			printf("'%s' ", argv_cpy[i]);
-- 
1.8.3.1

Re: [dpdk-stable] [dpdk-dev] [PATCH v4] test: optimise fd closing in forked test process

[Kevin Traynor]

On 12/11/2019 20:31, David Marchand wrote:
> From: Krzysztof Kanas <kkanas@marvell.com>
> 
> Caught while investigating timeouts on a ARM64 server.
> 
> Stracing a test process running the eal_flags_autotest, we can see that
> the fork helper is checking all possible file descriptors from
> getdtablesize() to 2, and close the existing ones.
> We can do better by inspecting this forked process /proc/self/fd
> directory.
> 
> Besides, checking file descriptors via /proc/self/fd only makes sense for
> Linux. This code was a noop on FreeBSD.
> 
> Fixes: af75078fece3 ("first public release")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Krzysztof Kanas <kkanas@marvell.com>
> Signed-off-by: David Marchand <david.marchand@redhat.com>

Seems to be an improvement on what is existing.

Acked-by: Kevin Traynor <ktraynor@redhat.com>

[dpdk-stable] [dpdk-dev] [PATCH v4] test: optimise fd closing in forked test process

[David Marchand]

On Thu, Nov 14, 2019 at 8:32 PM Kevin Traynor <ktraynor@redhat.com> wrote:
>
> On 12/11/2019 20:31, David Marchand wrote:
> > From: Krzysztof Kanas <kkanas@marvell.com>
> >
> > Caught while investigating timeouts on a ARM64 server.
> >
> > Stracing a test process running the eal_flags_autotest, we can see that
> > the fork helper is checking all possible file descriptors from
> > getdtablesize() to 2, and close the existing ones.
> > We can do better by inspecting this forked process /proc/self/fd
> > directory.
> >
> > Besides, checking file descriptors via /proc/self/fd only makes sense for
> > Linux. This code was a noop on FreeBSD.
> >
> > Fixes: af75078fece3 ("first public release")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Krzysztof Kanas <kkanas@marvell.com>
> > Signed-off-by: David Marchand <david.marchand@redhat.com>

Tested-by: Krzysztof Kanas <kkanas@marvell.com>

> Seems to be an improvement on what is existing.
>
> Acked-by: Kevin Traynor <ktraynor@redhat.com>
>

Applied, thanks.


--
David Marchand