Re: [PATCH v3 4/4] fsx: add IO_URING test

[Brian Foster <bfoster@redhat.com>]

On Sun, Aug 23, 2020 at 02:30:32PM +0800, Zorro Lang wrote:
> New IO_URING test for fsx, use -U option to enable IO_URING test.
> 
> Signed-off-by: Zorro Lang <zlang@redhat.com>
> ---

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...

>  ltp/fsx.c | 158 +++++++++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 144 insertions(+), 14 deletions(-)
> 
> diff --git a/ltp/fsx.c b/ltp/fsx.c
> index 7c76655a..05663528 100644
> --- a/ltp/fsx.c
> +++ b/ltp/fsx.c
...
> @@ -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)
> -#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?

>  const char *replayops = NULL;
>  const char *recordops = NULL;
...
> @@ -2425,13 +2427,131 @@ out_error:
>  	errno = -ret;
>  	return -1;
>  }
> +#endif
> +
> +#ifdef URING

A whitespace line here...

> +struct io_uring ring;
> +#define URING_ENTRIES	1024

... and here would help readability.

> +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?

> +}
>  
> -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()?

>  {
> +	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));
> +		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.

> +			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;
> +			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.

> +		} 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
> +		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

>  	} 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;
> -- 
> 2.20.1
>