fsstress,fsx: add io_uring test and do some fix

IO_URING is a new feature of curent linux kernel, add basic IO_URING read/write into fsstess to cover this kind of IO testing.

Signed-off-by Zorro Lang <[email protected]>

---
 README                 |   4 +-
 configure.ac           |   1 +
 include/builddefs.in   |   1 +
 ltp/Makefile           |   5 ++
 ltp/fsstress.c         | 139 ++++++++++++++++++++++++++++++++++++++++-
 m4/Makefile            |   1 +
 m4/package_liburing.m4 |   4 ++
 7 files changed, 152 insertions(+), 3 deletions(-)
 create mode 100644 m4/package_liburing.m4

ltp/fsstress.c +0 -0

index 709fdeec..7a0e278a 100644

ltp/fsstress.c +44 -0

@@ -2170,6 +2189,108 @@ do_aio_rw(int opno, long r, int flags)
 }
 #endif
 
+#ifdef URING
+void
+do_uring_rw(int opno, long r, int flags)
+{
+	char		*buf;
+	int		e;
+	pathname_t	f;
+	int		fd;
+	size_t		len;
+	int64_t		lr;
+	off64_t		off;
+	struct stat64	stb;
+	int		v;
+	char		st[1024];
+	struct io_uring_sqe	*sqe;
+	struct io_uring_cqe	*cqe;
+	struct iovec	iovec;
+	int		iswrite = (flags & (O_WRONLY | O_RDWR)) ? 1 : 0;
+
+	init_pathname(&f);
+	if (!get_fname(FT_REGFILE, r, &f, NULL, NULL, &v)) {
+		if (v)
+			printf("%d/%d: do_uring_rw - no filename\n", procid, opno);
+		goto uring_out3;
+	}
+	fd = open_path(&f, flags);
+	e = fd < 0 ? errno : 0;
+	check_cwd();
+	if (fd < 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - open %s failed %d\n",
+			       procid, opno, f.path, e);
+		goto uring_out3;
+	}
+	if (fstat64(fd, &stb) < 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - fstat64 %s failed %d\n",
+			       procid, opno, f.path, errno);
+		goto uring_out2;
+	}
+	inode_info(st, sizeof(st), &stb, v);
+	if (!iswrite && stb.st_size == 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - %s%s zero size\n", procid, opno,

ltp/fsstress.c +35 -0

+	lr = ((int64_t)random() << 32) + random();
+	len = (random() % FILELEN_MAX) + 1;
+	buf = malloc(len);
+	if (!buf) {
+		if (v)
+			printf("%d/%d: do_uring_rw - malloc failed\n",
+			       procid, opno);
+		goto uring_out2;
+	}
+	iovec.iov_base = buf;
+	iovec.iov_len = len;
+	if (iswrite) {
+		off = (off64_t)(lr % MIN(stb.st_size + (1024 * 1024), MAXFSIZE));
+		off %= maxfsize;
+		memset(buf, nameseq & 0xff, len);
+		io_uring_prep_writev(sqe, fd, &iovec, 1, off);
+	} else {
+		off = (off64_t)(lr % stb.st_size);
+		io_uring_prep_readv(sqe, fd, &iovec, 1, off);
+	}
+
+	if ((e = io_uring_submit_and_wait(&ring, 1)) != 1) {
+		if (v)
+			printf("%d/%d: %s - io_uring_submit failed %d\n", procid, opno,
+			       iswrite ? "uring_write" : "uring_read", e);
+		goto uring_out1;
+	}
+	if ((e = io_uring_wait_cqe(&ring, &cqe)) < 0) {
+		if (v)
+			printf("%d/%d: %s - io_uring_wait_cqe failed %d\n", procid, opno,
+			       iswrite ? "uring_write" : "uring_read", e);
+		goto uring_out1;
+	}
+	if (v)
+		printf("%d/%d: %s %s%s [%lld, %d(res=%d)] %d\n",

ltp/fsstress.c +19 -0

+}
+#endif
+
 void
 aread_f(int opno, long r)
 {

@@ -5044,6 +5165,22 @@ unresvsp_f(int opno, long r)
 	close(fd);
 }
 
+void
+uring_read_f(int opno, long r)
+{
+#ifdef URING
+	do_uring_rw(opno, r, O_RDONLY);
+#endif
+}
+
+void
+uring_write_f(int opno, long r)
+{
+#ifdef URING
+	do_uring_rw(opno, r, O_WRONLY);
+#endif
+}
+
 void
 write_f(int opno, long r)
 {

m4/Makefile +1 -0

index 7fbff822..0352534d 100644
@@ -14,6 +14,7 @@ LSRCFILES = \
 	package_dmapidev.m4 \
 	package_globals.m4 \
 	package_libcdev.m4 \
+	package_liburing.m4 \
 	package_ncurses.m4 \
 	package_pthread.m4 \
 	package_ssldev.m4 \

m4/package_liburing.m4 new file +4 -0

new file mode 100644
index 00000000..c92cc02a
@@ -0,0 +1,4 @@
+AC_DEFUN([AC_PACKAGE_WANT_URING],
+  [ AC_CHECK_HEADERS(liburing.h, [ have_uring=true ], [ have_uring=false ])
+    AC_SUBST(have_uring)
+  ])

Signature · 1 line

2.20.1

Quoted 15 lines · Zorro Lang

IO_URING is a new feature of curent linux kernel, add basic IO_URING read/write into fsstess to cover this kind of IO testing.

Signed-off-by Zorro Lang <[email protected]>

---
 README                 |   4 +-
 configure.ac           |   1 +
 include/builddefs.in   |   1 +
 ltp/Makefile           |   5 ++
 ltp/fsstress.c         | 139 ++++++++++++++++++++++++++++++++++++++++-
 m4/Makefile            |   1 +
 m4/package_liburing.m4 |   4 ++
 7 files changed, 152 insertions(+), 3 deletions(-)
 create mode 100644 m4/package_liburing.m4

...

Quoted 4 lines

ltp/fsstress.c +0 -0

index 709fdeec..7a0e278a 100644

...

Quoted 53 lines

ltp/fsstress.c +49 -0

@@ -2170,6 +2189,108 @@ do_aio_rw(int opno, long r, int flags)
 }
 #endif
 
+#ifdef URING
+void
+do_uring_rw(int opno, long r, int flags)
+{
+	char		*buf;
+	int		e;
+	pathname_t	f;
+	int		fd;
+	size_t		len;
+	int64_t		lr;
+	off64_t		off;
+	struct stat64	stb;
+	int		v;
+	char		st[1024];
+	struct io_uring_sqe	*sqe;
+	struct io_uring_cqe	*cqe;
+	struct iovec	iovec;
+	int		iswrite = (flags & (O_WRONLY | O_RDWR)) ? 1 : 0;
+
+	init_pathname(&f);
+	if (!get_fname(FT_REGFILE, r, &f, NULL, NULL, &v)) {
+		if (v)
+			printf("%d/%d: do_uring_rw - no filename\n", procid, opno);
+		goto uring_out3;
+	}
+	fd = open_path(&f, flags);
+	e = fd < 0 ? errno : 0;
+	check_cwd();
+	if (fd < 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - open %s failed %d\n",
+			       procid, opno, f.path, e);
+		goto uring_out3;
+	}
+	if (fstat64(fd, &stb) < 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - fstat64 %s failed %d\n",
+			       procid, opno, f.path, errno);
+		goto uring_out2;
+	}
+	inode_info(st, sizeof(st), &stb, v);
+	if (!iswrite && stb.st_size == 0) {
+		if (v)
+			printf("%d/%d: do_uring_rw - %s%s zero size\n", procid, opno,
+			       f.path, st);
+		goto uring_out2;
+	}
+	sqe = io_uring_get_sqe(&ring);
+	if (!sqe) {

ltp/fsstress.c do_aio_rw(int opno, long r, int flags) · line 2245

Quoted 5 lines

ltp/fsstress.c +5 -0

+		if (v)
+			printf("%d/%d: do_uring_rw - io_uring_get_sqe failed\n",
+			       procid, opno);
+		goto uring_out2;
+	}

I'm not familiar with the io_uring bits, but do we have to do anything to clean up this sqe object (or the cqe) before we return?

The original number(128) of aio events for io_setup too big. When try to run lots of fsstress processes(e.g. -p 1000) always hit io_setup EAGAIN error, due to the nr_events exceeds the limit of available events. Due to each fsstress process only does once libaio read/write operation each time. So reduce the aio events number to 1, to make more fsstress processes can do AIO test.

Signed-off-by Zorro Lang <[email protected]>

---

 ltp/fsstress.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

ltp/fsstress.c +3 -2

index 7a0e278a..ef2017a8 100644
@@ -28,6 +28,7 @@
 #endif
 #ifdef AIO
 #include <libaio.h>
+#define AIO_ENTRIES	1
 io_context_t	io_ctx;
 #endif
 #ifdef URING

@@ -699,8 +700,8 @@ int main(int argc, char **argv)
 			}
 			procid = i;
 #ifdef AIO
-			if (io_setup(128, &io_ctx) != 0) {
-				fprintf(stderr, "io_setup failed");
+			if (io_setup(AIO_ENTRIES, &io_ctx) != 0) {
+				fprintf(stderr, "io_setup failed\n");
 				exit(1);
 			}
 #endif

Signature · 1 line

2.20.1

If io_submit or io_getevents fails, the do_aio_rw() won't free the "buf" and cause memory leak.

Signed-off-by Zorro Lang <[email protected]>

---

 ltp/fsstress.c | 31 ++++++++++++++++---------------
 1 file changed, 16 insertions(+), 15 deletions(-)

ltp/fsstress.c +16 -15

index ef2017a8..17b024b5 100644
@@ -2099,8 +2099,7 @@ do_aio_rw(int opno, long r, int flags)
 	if (!get_fname(FT_REGFILE, r, &f, NULL, NULL, &v)) {
 		if (v)
 			printf("%d/%d: do_aio_rw - no filename\n", procid, opno);
-		free_pathname(&f);
-		return;
+		goto aio_out3;
 	}
 	fd = open_path(&f, flags|O_DIRECT);
 	e = fd < 0 ? errno : 0;

@@ -2109,16 +2108,13 @@ do_aio_rw(int opno, long r, int flags)
 		if (v)
 			printf("%d/%d: do_aio_rw - open %s failed %d\n",
 			       procid, opno, f.path, e);
-		free_pathname(&f);
-		return;
+		goto aio_out3;
 	}
 	if (fstat64(fd, &stb) < 0) {
 		if (v)
 			printf("%d/%d: do_aio_rw - fstat64 %s failed %d\n",
 			       procid, opno, f.path, errno);
-		free_pathname(&f);
-		close(fd);
-		return;
+		goto aio_out2;
 	}
 	inode_info(st, sizeof(st), &stb, v);
 	if (!iswrite && stb.st_size == 0) {

@@ -2150,6 +2146,12 @@ do_aio_rw(int opno, long r, int flags)
 	else if (len > diob.d_maxiosz)
 		len = diob.d_maxiosz;
 	buf = memalign(diob.d_mem, len);
+	if (!buf) {
+		if (v)
+			printf("%d/%d: do_aio_rw - memalign failed\n",
+			       procid, opno);
+		goto aio_out2;
+	}
 
 	if (iswrite) {
 		off = (off64_t)(lr % MIN(stb.st_size + (1024 * 1024), MAXFSIZE));

@@ -2166,27 +2168,26 @@ do_aio_rw(int opno, long r, int flags)
 		if (v)
 			printf("%d/%d: %s - io_submit failed %d\n",
 			       procid, opno, iswrite ? "awrite" : "aread", e);
-		free_pathname(&f);
-		close(fd);
-		return;
+		goto aio_out1;
 	}
 	if ((e = io_getevents(io_ctx, 1, 1, &event, NULL)) != 1) {
 		if (v)
 			printf("%d/%d: %s - io_getevents failed %d\n",
 			       procid, opno, iswrite ? "awrite" : "aread", e);
-		free_pathname(&f);
-		close(fd);
-		return;
+		goto aio_out1;
 	}
 
 	e = event.res != len ? event.res2 : 0;
-	free(buf);
 	if (v)
 		printf("%d/%d: %s %s%s [%lld,%d] %d\n",
 		       procid, opno, iswrite ? "awrite" : "aread",
 		       f.path, st, (long long)off, (int)len, e);
-	free_pathname(&f);
+ aio_out1:
+	free(buf);
+ aio_out2:
 	close(fd);
+ aio_out3:
+	free_pathname(&f);
 }
 #endif
 

Signature · 1 line

2.20.1

New IO_URING test for fsx, use -U option to enable IO_URING test.

Signed-off-by Zorro Lang <[email protected]>

---

 ltp/fsx.c | 158 +++++++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 144 insertions(+), 14 deletions(-)

ltp/fsx.c +0 -0

index 7c76655a..05663528 100644

ltp/fsx.c +2 -2

@@ -176,21 +179,17 @@ int	integrity = 0;			/* -i flag */
 int	fsxgoodfd = 0;
 int	o_direct;			/* -Z */
 int	aio = 0;
+int	uring = 0;
 int	mark_nr = 0;
 
 int page_size;
 int page_mask;
 int mmap_mask;
-#ifdef AIO
-int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
 #define READ 0

ltp/fsx.c +0 -0

 const char *replayops = NULL;
 const char *recordops = NULL;

ltp/fsx.c +31 -0

 {
+	struct io_uring_sqe	*sqe;
+	struct io_uring_cqe	*cqe;
+	struct iovec		iovec;
 	int ret;
+	int res, res2 = 0;
+	char *p = buf;
+	unsigned l = len;
+	unsigned o = offset;
+
+
+	/*
+	 * Due to io_uring tries non-blocking IOs (especially read), that
+	 * always cause 'normal' short reading. To avoid this short read
+	 * fail, try to loop read/write (escpecilly read) data.
+	 */
+ uring_loop:
+	sqe = io_uring_get_sqe(&ring);
+	if (!sqe) {
+		fprintf(stderr, "uring_rw: io_uring_get_sqe failed: %s\n",
+		        strerror(errno));
+		return -1;
+        }
+
+	iovec.iov_base = p;
+	iovec.iov_len = l;
+	if (rw == READ) {
+		io_uring_prep_readv(sqe, fd, &iovec, 1, o);
+	} else {
+		io_uring_prep_writev(sqe, fd, &iovec, 1, o);
+	}
+
+	ret = io_uring_submit_and_wait(&ring, 1);

ltp/fsx.c +16 -0

+			fprintf(stderr, "uring %s: no events available\n",
+			        rw == READ ? "read":"write");
+		else {
+			fprintf(stderr, "errcode=%d\n", -ret);
+			fprintf(stderr, "uring %s: io_uring_wait_cqe failed: %s\n",
+			        rw == READ ? "read":"write", strerror(-ret));
+		}
+		goto uring_error;
+	}
+	res = cqe->res;
+	io_uring_cqe_seen(&ring, cqe);
+
+	res2 += res;
+	if (len != res2) {
+		if (res > 0) {
+			o += res;

ltp/fsx.c +22 -0

+		} else {
+			fprintf(stderr, "uring %s bad io length: %d instead of %u\n",
+			        rw == READ ? "read":"write", res2, len);
+		}
+	}
+	return res2;
+
+ uring_error:
+	/*
+	 * The caller expects error return in traditional libc
+	 * convention, i.e. -1 and the errno set to error.
+	 */
+	errno = -ret;
+	return -1;
+}
+#endif
+
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
+{
+	int ret = -1;
 
 	if (aio) {
+#ifdef AIO
 		ret = __aio_rw(rw, fd, buf, len, offset);
+#elif

ltp/fsx.c +13 -3

 	} else {
 		if (rw == READ)
 			ret = read(fd, buf, len);

@@ -2441,8 +2561,6 @@ int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
 	return ret;
 }
 
-#endif
-
 #define test_fallocate(mode) __test_fallocate(mode, #mode)
 
 int

@@ -2496,7 +2614,7 @@ main(int argc, char **argv)
 	setvbuf(stdout, (char *)0, _IOLBF, 0); /* line buffered stdout */
 
 	while ((ch = getopt_long(argc, argv,
-				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:WXZ",
+				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:UWXZ",
 				 longopts, NULL)) != EOF)
 		switch (ch) {
 		case 'b':

@@ -2604,6 +2722,9 @@ main(int argc, char **argv)
 		case 'A':
 		        aio = 1;
 			break;
+		case 'U':
+		        uring = 1;
+			break;
 		case 'D':
 			debugstart = getnum(optarg, &endp);
 			if (debugstart < 1)

@@ -2694,6 +2815,11 @@ main(int argc, char **argv)
 	if (argc != 1)
 		usage();
 
+	if (aio && uring) {
+		fprintf(stderr, "-A and -U shouldn't be used together\n");
+		usage();
+	}
+
 	if (integrity && !dirpath) {
 		fprintf(stderr, "option -i <logdev> requires -P <dirpath>\n");
 		usage();

@@ -2784,6 +2910,10 @@ main(int argc, char **argv)
 	if (aio) 
 		aio_setup();
 #endif
+#ifdef URING
+	if (uring)
+		uring_setup();
+#endif
 
 	if (!(o_flags & O_TRUNC)) {
 		off_t ret;

Signature · 1 line

2.20.1

Quoted 4 lines · Zorro Lang

New IO_URING test for fsx, use -U option to enable IO_URING test.

Signed-off-by Zorro Lang <[email protected]>

---

Note that this one doesn't compile if one of the ifdefs doesn't evaluate true:

fsx.c:2551:6: error: #elif with no expression
2551 | #elif
| ^
[CC] fsx
fsx.c: In function 'fsx_rw':
fsx.c:2551:6: error: #elif with no expression
2551 | #elif
| ^
gmake[2]: *** [Makefile:52: fsx] Error 1 gmake[1]: *** [include/buildrules:30: ltp] Error 2
make: *** [Makefile:53: default] Error 2

I suspect you want to replace both of those with #else. Otherwise mostly some aesthetic comments...

Quoted 8 lines

 ltp/fsx.c | 158 +++++++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 144 insertions(+), 14 deletions(-)

ltp/fsx.c +0 -0

index 7c76655a..05663528 100644

...

Quoted 19 lines

ltp/fsx.c +2 -6

@@ -176,21 +179,17 @@ int	integrity = 0;			/* -i flag */
 int	fsxgoodfd = 0;
 int	o_direct;			/* -Z */
 int	aio = 0;
+int	uring = 0;
 int	mark_nr = 0;
 
 int page_size;
 int page_mask;
 int mmap_mask;
-#ifdef AIO
-int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
 #define READ 0
 #define WRITE 1
-#define fsxread(a,b,c,d)	aio_rw(READ, a,b,c,d)
-#define fsxwrite(a,b,c,d)	aio_rw(WRITE, a,b,c,d)
-#else
-#define fsxread(a,b,c,d)	read(a,b,c)

ltp/fsx.c int integrity = 0; /* -i flag */ · line 193

Quoted 5 lines

ltp/fsx.c +2 -2

-#define fsxwrite(a,b,c,d)	write(a,b,c)
-#endif
+#define fsxread(a,b,c,d)	fsx_rw(READ, a,b,c,d)
+#define fsxwrite(a,b,c,d)	fsx_rw(WRITE, a,b,c,d)
 

Could we do the refactoring that introduces fsx_rw and shuffles around some of the existing AIO in an initial refactoring patch?

Quoted 3 lines

ltp/fsx.c +0 -0

 const char *replayops = NULL;
 const char *recordops = NULL;

...

ltp/fsx.c out_error: · line 2432

Quoted 7 lines

ltp/fsx.c +3 -0

@@ -2425,13 +2427,131 @@ out_error:
 	errno = -ret;
 	return -1;
 }
+#endif
+
+#ifdef URING

A whitespace line here...

ltp/fsx.c out_error: · line 2434

Quoted 2 lines

ltp/fsx.c +2 -0

+struct io_uring ring;
+#define URING_ENTRIES	1024

... and here would help readability.

ltp/fsx.c out_error: · line 2446

Quoted 12 lines

ltp/fsx.c +12 -0

+int
+uring_setup()
+{
+	int ret;
+
+	ret = io_uring_queue_init(URING_ENTRIES, &ring, 0);
+	if (ret != 0) {
+		fprintf(stderr, "uring_setup: io_uring_queue_init failed: %s\n",
+                        strerror(ret));
+                return -1;
+        }
+        return 0;

Looks like some whitespace damage here.

Also, the fsstress patch has a io_uring_queue_exit() call but I don't see one in this patch. Is that not needed?

ltp/fsx.c out_error: · line 2450

Quoted 6 lines

ltp/fsx.c +3 -1

+}
 
-int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
+int
+__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)

Do we still need the __ in the function names here and for __aio_rw()?

Quoted 38 lines

ltp/fsx.c +35 -0

 {
+	struct io_uring_sqe	*sqe;
+	struct io_uring_cqe	*cqe;
+	struct iovec		iovec;
 	int ret;
+	int res, res2 = 0;
+	char *p = buf;
+	unsigned l = len;
+	unsigned o = offset;
+
+
+	/*
+	 * Due to io_uring tries non-blocking IOs (especially read), that
+	 * always cause 'normal' short reading. To avoid this short read
+	 * fail, try to loop read/write (escpecilly read) data.
+	 */
+ uring_loop:
+	sqe = io_uring_get_sqe(&ring);
+	if (!sqe) {
+		fprintf(stderr, "uring_rw: io_uring_get_sqe failed: %s\n",
+		        strerror(errno));
+		return -1;
+        }
+
+	iovec.iov_base = p;
+	iovec.iov_len = l;
+	if (rw == READ) {
+		io_uring_prep_readv(sqe, fd, &iovec, 1, o);
+	} else {
+		io_uring_prep_writev(sqe, fd, &iovec, 1, o);
+	}
+
+	ret = io_uring_submit_and_wait(&ring, 1);
+	if (ret != 1) {
+		fprintf(stderr, "errcode=%d\n", -ret);
+		fprintf(stderr, "uring %s: io_uring_submit failed: %s\n",
+		        rw == READ ? "read":"write", strerror(-ret));

ltp/fsx.c out_error: · line 2493

Quoted 6 lines

ltp/fsx.c +6 -0

+		goto uring_error;
+	}
+
+	ret = io_uring_wait_cqe(&ring, &cqe);
+	if (ret < 0) {
+		if (ret == 0)

That doesn't look right since we only get here if ret < 0.

Quoted 17 lines

ltp/fsx.c +16 -0

+			fprintf(stderr, "uring %s: no events available\n",
+			        rw == READ ? "read":"write");
+		else {
+			fprintf(stderr, "errcode=%d\n", -ret);
+			fprintf(stderr, "uring %s: io_uring_wait_cqe failed: %s\n",
+			        rw == READ ? "read":"write", strerror(-ret));
+		}
+		goto uring_error;
+	}
+	res = cqe->res;
+	io_uring_cqe_seen(&ring, cqe);
+
+	res2 += res;
+	if (len != res2) {
+		if (res > 0) {
+			o += res;

ltp/fsx.c out_error: · line 2519

Quoted 10 lines

ltp/fsx.c +10 -0

+			l -= res;
+			p += res;
+			if (l > 0)
+				goto uring_loop;
+		} else if (res < 0) {
+			ret = res;
+			fprintf(stderr, "errcode=%d\n", -ret);
+			fprintf(stderr, "uring %s: io_uring failed: %s\n",
+			        rw == READ ? "read":"write", strerror(-ret));
+			goto uring_error;

Can we elevate the error checks into the top level rather than nesting logic like this? It's a little confusing to read and it looks particularly odd since we've already done res2 += res before we get here.

Also I'm wondering if this whole function would read a little better as a do {} while() loop rather than using a label and goto.

Quoted 26 lines

ltp/fsx.c +22 -0

+		} else {
+			fprintf(stderr, "uring %s bad io length: %d instead of %u\n",
+			        rw == READ ? "read":"write", res2, len);
+		}
+	}
+	return res2;
+
+ uring_error:
+	/*
+	 * The caller expects error return in traditional libc
+	 * convention, i.e. -1 and the errno set to error.
+	 */
+	errno = -ret;
+	return -1;
+}
+#endif
+
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
+{
+	int ret = -1;
 
 	if (aio) {
+#ifdef AIO
 		ret = __aio_rw(rw, fd, buf, len, offset);
+#elif

ltp/fsx.c out_error: · line 2554

Quoted 10 lines

ltp/fsx.c +10 -0

+		fprintf(stderr, "io_rw: need AIO support!\n");
+		exit(111);
+#endif
+	} else if (uring) {
+#ifdef URING
+		ret = __uring_rw(rw, fd, buf, len, offset);
+#elif
+		fprintf(stderr, "io_rw: need IO_URING support!\n");
+		exit(111);
+#endif

I think the ifdefs would be cleaner if used to define stubbed out variants of the associated functions. E.g.:

#ifdef URING
int
__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset) {
<do uring I/O>
}
#else
int
__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)

{
	fprintf(stderr, "io_rw: need IO_URING support!\n");
	exit(111);
}

#endif

Brian

Quoted 57 lines

ltp/fsx.c +13 -3

 	} else {
 		if (rw == READ)
 			ret = read(fd, buf, len);

@@ -2441,8 +2561,6 @@ int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
 	return ret;
 }
 
-#endif
-
 #define test_fallocate(mode) __test_fallocate(mode, #mode)
 
 int

@@ -2496,7 +2614,7 @@ main(int argc, char **argv)
 	setvbuf(stdout, (char *)0, _IOLBF, 0); /* line buffered stdout */
 
 	while ((ch = getopt_long(argc, argv,
-				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:WXZ",
+				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:UWXZ",
 				 longopts, NULL)) != EOF)
 		switch (ch) {
 		case 'b':

@@ -2604,6 +2722,9 @@ main(int argc, char **argv)
 		case 'A':
 		        aio = 1;
 			break;
+		case 'U':
+		        uring = 1;
+			break;
 		case 'D':
 			debugstart = getnum(optarg, &endp);
 			if (debugstart < 1)

@@ -2694,6 +2815,11 @@ main(int argc, char **argv)
 	if (argc != 1)
 		usage();
 
+	if (aio && uring) {
+		fprintf(stderr, "-A and -U shouldn't be used together\n");
+		usage();
+	}
+
 	if (integrity && !dirpath) {
 		fprintf(stderr, "option -i <logdev> requires -P <dirpath>\n");
 		usage();

@@ -2784,6 +2910,10 @@ main(int argc, char **argv)
 	if (aio) 
 		aio_setup();
 #endif
+#ifdef URING
+	if (uring)
+		uring_setup();
+#endif
 
 	if (!(o_flags & O_TRUNC)) {
 		off_t ret;

Signature · 1 line

2.20.1

Quoted 21 lines · Brian Foster

Quoted 4 lines · Zorro Lang

New IO_URING test for fsx, use -U option to enable IO_URING test.

Signed-off-by Zorro Lang <[email protected]>

---

Note that this one doesn't compile if one of the ifdefs doesn't evaluate true:

fsx.c:2551:6: error: #elif with no expression
2551 | #elif
| ^
[CC] fsx
fsx.c: In function 'fsx_rw':
fsx.c:2551:6: error: #elif with no expression
2551 | #elif
| ^
gmake[2]: *** [Makefile:52: fsx] Error 1 gmake[1]: *** [include/buildrules:30: ltp] Error 2
make: *** [Makefile:53: default] Error 2

Quoted 2 lines

I suspect you want to replace both of those with #else. Otherwise mostly some aesthetic comments...

Sorry, that's truely a mistake, I'll fix it :)

Quoted 28 lines

Quoted 8 lines

 ltp/fsx.c | 158 +++++++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 144 insertions(+), 14 deletions(-)

ltp/fsx.c +0 -0

index 7c76655a..05663528 100644

...

Quoted 19 lines

ltp/fsx.c +2 -6

@@ -176,21 +179,17 @@ int	integrity = 0;			/* -i flag */
 int	fsxgoodfd = 0;
 int	o_direct;			/* -Z */
 int	aio = 0;
+int	uring = 0;
 int	mark_nr = 0;
 
 int page_size;
 int page_mask;
 int mmap_mask;
-#ifdef AIO
-int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset);
 #define READ 0
 #define WRITE 1
-#define fsxread(a,b,c,d)	aio_rw(READ, a,b,c,d)
-#define fsxwrite(a,b,c,d)	aio_rw(WRITE, a,b,c,d)
-#else
-#define fsxread(a,b,c,d)	read(a,b,c)

Quoted 8 lines

ltp/fsx.c int integrity = 0; /* -i flag */ · line 193

Quoted 5 lines

ltp/fsx.c +2 -2

-#define fsxwrite(a,b,c,d)	write(a,b,c)
-#endif
+#define fsxread(a,b,c,d)	fsx_rw(READ, a,b,c,d)
+#define fsxwrite(a,b,c,d)	fsx_rw(WRITE, a,b,c,d)
 

Could we do the refactoring that introduces fsx_rw and shuffles around some of the existing AIO in an initial refactoring patch?

May I save this pre-patch, if you don't insist on that :-P

Quoted 34 lines

Quoted 3 lines

ltp/fsx.c +0 -0

 const char *replayops = NULL;
 const char *recordops = NULL;

...

ltp/fsx.c out_error: · line 2432

Quoted 7 lines

ltp/fsx.c +3 -0

@@ -2425,13 +2427,131 @@ out_error:
 	errno = -ret;
 	return -1;
 }
+#endif
+
+#ifdef URING

A whitespace line here...

ltp/fsx.c out_error: · line 2434

Quoted 2 lines

ltp/fsx.c +2 -0

+struct io_uring ring;
+#define URING_ENTRIES	1024

... and here would help readability.

ltp/fsx.c out_error: · line 2446

Quoted 12 lines

ltp/fsx.c +12 -0

+int
+uring_setup()
+{
+	int ret;
+
+	ret = io_uring_queue_init(URING_ENTRIES, &ring, 0);
+	if (ret != 0) {
+		fprintf(stderr, "uring_setup: io_uring_queue_init failed: %s\n",
+                        strerror(ret));
+                return -1;
+        }
+        return 0;

Looks like some whitespace damage here.

Quoted 2 lines

Also, the fsstress patch has a io_uring_queue_exit() call but I don't see one in this patch. Is that not needed?

There's not aio_destroy() either. I think due to fsstress is a multi-process test, so it'd like to destroy io_uring or aio at each process end. But fsx is a pure single process test, the io_uring or aio will destroyed when fsx exit. I can add io_uring_queue_exit() and aio_destroy() if you think it would be better.

Quoted 8 lines

ltp/fsx.c out_error: · line 2450

Quoted 6 lines

ltp/fsx.c +3 -1

+}
 
-int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
+int
+__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)

Do we still need the __ in the function names here and for __aio_rw()?

I don't think it's needed. I use the "__" just due to the old __aio_rw() has. I can remove both "__" of __aio_rw and __uring_rw.

Quoted 38 lines

Quoted 38 lines

ltp/fsx.c +35 -0

 {
+	struct io_uring_sqe	*sqe;
+	struct io_uring_cqe	*cqe;
+	struct iovec		iovec;
 	int ret;
+	int res, res2 = 0;
+	char *p = buf;
+	unsigned l = len;
+	unsigned o = offset;
+
+
+	/*
+	 * Due to io_uring tries non-blocking IOs (especially read), that
+	 * always cause 'normal' short reading. To avoid this short read
+	 * fail, try to loop read/write (escpecilly read) data.
+	 */
+ uring_loop:
+	sqe = io_uring_get_sqe(&ring);
+	if (!sqe) {
+		fprintf(stderr, "uring_rw: io_uring_get_sqe failed: %s\n",
+		        strerror(errno));
+		return -1;
+        }
+
+	iovec.iov_base = p;
+	iovec.iov_len = l;
+	if (rw == READ) {
+		io_uring_prep_readv(sqe, fd, &iovec, 1, o);
+	} else {
+		io_uring_prep_writev(sqe, fd, &iovec, 1, o);
+	}
+
+	ret = io_uring_submit_and_wait(&ring, 1);
+	if (ret != 1) {
+		fprintf(stderr, "errcode=%d\n", -ret);
+		fprintf(stderr, "uring %s: io_uring_submit failed: %s\n",
+		        rw == READ ? "read":"write", strerror(-ret));

Quoted 8 lines

ltp/fsx.c out_error: · line 2493

Quoted 6 lines

ltp/fsx.c +6 -0

+		goto uring_error;
+	}
+
+	ret = io_uring_wait_cqe(&ring, &cqe);
+	if (ret < 0) {
+		if (ret == 0)

That doesn't look right since we only get here if ret < 0.

Thanks, it should be (ret <= 0)

Quoted 34 lines

Quoted 18 lines

ltp/fsx.c +16 -0

+			fprintf(stderr, "uring %s: no events available\n",
+			        rw == READ ? "read":"write");
+		else {
+			fprintf(stderr, "errcode=%d\n", -ret);
+			fprintf(stderr, "uring %s: io_uring_wait_cqe failed: %s\n",
+			        rw == READ ? "read":"write", strerror(-ret));
+		}
+		goto uring_error;
+	}
+	res = cqe->res;
+	io_uring_cqe_seen(&ring, cqe);
+
+	res2 += res;
+	if (len != res2) {
+		if (res > 0) {
+			o += res;

ltp/fsx.c out_error: · line 2519

Quoted 10 lines

ltp/fsx.c +10 -0

+			l -= res;
+			p += res;
+			if (l > 0)
+				goto uring_loop;
+		} else if (res < 0) {
+			ret = res;
+			fprintf(stderr, "errcode=%d\n", -ret);
+			fprintf(stderr, "uring %s: io_uring failed: %s\n",
+			        rw == READ ? "read":"write", strerror(-ret));
+			goto uring_error;

Can we elevate the error checks into the top level rather than nesting logic like this? It's a little confusing to read and it looks particularly odd since we've already done res2 += res before we get here.

Quoted 2 lines

Also I'm wondering if this whole function would read a little better as a do {} while() loop rather than using a label and goto.

Sure, I'll try to change that.

Quoted 49 lines

Quoted 26 lines

ltp/fsx.c +22 -0

+		} else {
+			fprintf(stderr, "uring %s bad io length: %d instead of %u\n",
+			        rw == READ ? "read":"write", res2, len);
+		}
+	}
+	return res2;
+
+ uring_error:
+	/*
+	 * The caller expects error return in traditional libc
+	 * convention, i.e. -1 and the errno set to error.
+	 */
+	errno = -ret;
+	return -1;
+}
+#endif
+
+int fsx_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
+{
+	int ret = -1;
 
 	if (aio) {
+#ifdef AIO
 		ret = __aio_rw(rw, fd, buf, len, offset);
+#elif

ltp/fsx.c out_error: · line 2554

Quoted 10 lines

ltp/fsx.c +10 -0

+		fprintf(stderr, "io_rw: need AIO support!\n");
+		exit(111);
+#endif
+	} else if (uring) {
+#ifdef URING
+		ret = __uring_rw(rw, fd, buf, len, offset);
+#elif
+		fprintf(stderr, "io_rw: need IO_URING support!\n");
+		exit(111);
+#endif

I think the ifdefs would be cleaner if used to define stubbed out variants of the associated functions. E.g.:

#ifdef URING
int
__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset) {
<do uring I/O>
}
#else
int
__uring_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)

Quoted 5 lines

{
	fprintf(stderr, "io_rw: need IO_URING support!\n");
	exit(111);
}

#endif

Sure, will do that.

Thanks for your review, Brian!
Zorro

Quoted 60 lines

Brian

Quoted 57 lines

ltp/fsx.c +13 -3

 	} else {
 		if (rw == READ)
 			ret = read(fd, buf, len);

@@ -2441,8 +2561,6 @@ int aio_rw(int rw, int fd, char *buf, unsigned len, unsigned offset)
 	return ret;
 }
 
-#endif
-
 #define test_fallocate(mode) __test_fallocate(mode, #mode)
 
 int

@@ -2496,7 +2614,7 @@ main(int argc, char **argv)
 	setvbuf(stdout, (char *)0, _IOLBF, 0); /* line buffered stdout */
 
 	while ((ch = getopt_long(argc, argv,
-				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:WXZ",
+				 "b:c:dfg:i:j:kl:m:no:p:qr:s:t:w:xyABD:EFJKHzCILN:OP:RS:UWXZ",
 				 longopts, NULL)) != EOF)
 		switch (ch) {
 		case 'b':

@@ -2604,6 +2722,9 @@ main(int argc, char **argv)
 		case 'A':
 		        aio = 1;
 			break;
+		case 'U':
+		        uring = 1;
+			break;
 		case 'D':
 			debugstart = getnum(optarg, &endp);
 			if (debugstart < 1)

@@ -2694,6 +2815,11 @@ main(int argc, char **argv)
 	if (argc != 1)
 		usage();
 
+	if (aio && uring) {
+		fprintf(stderr, "-A and -U shouldn't be used together\n");
+		usage();
+	}
+
 	if (integrity && !dirpath) {
 		fprintf(stderr, "option -i <logdev> requires -P <dirpath>\n");
 		usage();

@@ -2784,6 +2910,10 @@ main(int argc, char **argv)
 	if (aio) 
 		aio_setup();
 #endif
+#ifdef URING
+	if (uring)
+		uring_setup();
+#endif
 
 	if (!(o_flags & O_TRUNC)) {
 		off_t ret;

Signature · 1 line

2.20.1

This patchset tries to add new IO_URING test into fsstress [1/4] and fsx [4/4]. And then do some changes and bug fix by the way [2/4 and 3/4].

fsstress and fsx are important tools in xfstests to do filesystem I/Os test, lots of test cases use it. So add IO_URING operation into fsstress and fsx will help to cover IO_URING test from fs side.

I'm not an IO_URING expert, so cc io-uring@ list, please feel free to tell me if you find something wrong or have any suggestions to improve the test.

V2 did below changes:
1) 1/4 change the definition of URING_ENTRIES to 1
2) 2/4 change the difinition of AIO_ENTRIES to 1, undo an unrelated changed line
3) 4/4 turn to use io_uring_prep_readv/io_uring_prep_writev, due to old liburing(0.2-2) doesn't support io_uring_prep_read/io_uring_prep_write.

V3 changed io_uring_submit(&ring) to io_uring_submit_and_wait(&ring, 1). I'm not sure if this's the real mean of Jens Axboe's review point, please check.
https://marc.info/?l=fstests&m=159811932808057&w=2

Thanks,
Zorro