[PATCH v2 0/3] xfsprogs: Add support for preadv2() and RWF_DONTCACHE

[PATCH v2 0/3] xfsprogs: Add support for preadv2() and RWF_DONTCACHE

[Ritesh Harjani (IBM)]

This adds following support to xfs_io:
- Support for preadv2()
- Support for uncached buffered I/O (RWF_DONTCACHE) for preadv2() and pwritev2()

v1 -> v2:
========
1. Based on pwritev2 autoconf support enable HAVE_PREADV2 support.
2. Check if preadv2/pwritev2 flags are passed w/o -V argument
3. Fixed space before tabs issue.
[v1]: https://lore.kernel.org/linux-xfs/cover.1741087191.git.ritesh.list@gmail.com/#t

Ritesh Harjani (IBM) (3):
  xfs_io: Add support for preadv2
  xfs_io: Add RWF_DONTCACHE support to pwritev2
  xfs_io: Add RWF_DONTCACHE support to preadv2

 include/linux.h   |  5 ++++
 io/Makefile       |  2 +-
 io/pread.c        | 62 ++++++++++++++++++++++++++++++++++-------------
 io/pwrite.c       | 14 +++++++++--
 man/man8/xfs_io.8 | 16 ++++++++++--
 5 files changed, 77 insertions(+), 22 deletions(-)

--
2.48.1

[PATCH v2 1/3] xfs_io: Add support for preadv2

[Ritesh Harjani (IBM)]

This patch adds support for preadv2() to xfs_io.

Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
---
 io/Makefile |  2 +-
 io/pread.c  | 45 ++++++++++++++++++++++++++++++---------------
 2 files changed, 31 insertions(+), 16 deletions(-)

diff --git a/io/Makefile b/io/Makefile
index 8f835ec7..14a3fe20 100644
--- a/io/Makefile
+++ b/io/Makefile
@@ -66,7 +66,7 @@ LLDLIBS += $(LIBEDITLINE) $(LIBTERMCAP)
 endif
 
 ifeq ($(HAVE_PWRITEV2),yes)
-LCFLAGS += -DHAVE_PWRITEV2
+LCFLAGS += -DHAVE_PWRITEV2 -DHAVE_PREADV2
 endif
 
 ifeq ($(HAVE_MAP_SYNC),yes)
diff --git a/io/pread.c b/io/pread.c
index 62c771fb..b314fbc7 100644
--- a/io/pread.c
+++ b/io/pread.c
@@ -162,7 +162,8 @@ static ssize_t
 do_preadv(
 	int		fd,
 	off_t		offset,
-	long long	count)
+	long long	count,
+	int			preadv2_flags)
 {
 	int		vecs = 0;
 	ssize_t		oldlen = 0;
@@ -181,8 +182,14 @@ do_preadv(
 	} else {
 		vecs = vectors;
 	}
+#ifdef HAVE_PREADV2
+	if (preadv2_flags)
+		bytes = preadv2(fd, iov, vectors, offset, preadv2_flags);
+	else
+		bytes = preadv(fd, iov, vectors, offset);
+#else
 	bytes = preadv(fd, iov, vectors, offset);
-
+#endif
 	/* restore trimmed iov */
 	if (oldlen)
 		iov[vecs - 1].iov_len = oldlen;
@@ -195,12 +202,13 @@ do_pread(
 	int		fd,
 	off_t		offset,
 	long long	count,
-	size_t		buffer_size)
+	size_t		buffer_size,
+	int			preadv2_flags)
 {
 	if (!vectors)
 		return pread(fd, io_buffer, min(count, buffer_size), offset);
 
-	return do_preadv(fd, offset, count);
+	return do_preadv(fd, offset, count, preadv2_flags);
 }
 
 static int
@@ -210,7 +218,8 @@ read_random(
 	long long	count,
 	long long	*total,
 	unsigned int	seed,
-	int		eof)
+	int		eof,
+	int		preadv2_flags)
 {
 	off_t		end, off, range;
 	ssize_t		bytes;
@@ -234,7 +243,7 @@ read_random(
 				io_buffersize;
 		else
 			off = offset;
-		bytes = do_pread(fd, off, io_buffersize, io_buffersize);
+		bytes = do_pread(fd, off, io_buffersize, io_buffersize, preadv2_flags);
 		if (bytes == 0)
 			break;
 		if (bytes < 0) {
@@ -256,7 +265,8 @@ read_backward(
 	off_t		*offset,
 	long long	*count,
 	long long	*total,
-	int		eof)
+	int		eof,
+	int		preadv2_flags)
 {
 	off_t		end, off = *offset;
 	ssize_t		bytes = 0, bytes_requested;
@@ -276,7 +286,7 @@ read_backward(
 	/* Do initial unaligned read if needed */
 	if ((bytes_requested = (off % io_buffersize))) {
 		off -= bytes_requested;
-		bytes = do_pread(fd, off, bytes_requested, io_buffersize);
+		bytes = do_pread(fd, off, bytes_requested, io_buffersize, preadv2_flags);
 		if (bytes == 0)
 			return ops;
 		if (bytes < 0) {
@@ -294,7 +304,7 @@ read_backward(
 	while (cnt > end) {
 		bytes_requested = min(cnt, io_buffersize);
 		off -= bytes_requested;
-		bytes = do_pread(fd, off, cnt, io_buffersize);
+		bytes = do_pread(fd, off, cnt, io_buffersize, preadv2_flags);
 		if (bytes == 0)
 			break;
 		if (bytes < 0) {
@@ -318,14 +328,15 @@ read_forward(
 	long long	*total,
 	int		verbose,
 	int		onlyone,
-	int		eof)
+	int		eof,
+	int		preadv2_flags)
 {
 	ssize_t		bytes;
 	int		ops = 0;
 
 	*total = 0;
 	while (count > 0 || eof) {
-		bytes = do_pread(fd, offset, count, io_buffersize);
+		bytes = do_pread(fd, offset, count, io_buffersize, preadv2_flags);
 		if (bytes == 0)
 			break;
 		if (bytes < 0) {
@@ -353,7 +364,7 @@ read_buffer(
 	int		verbose,
 	int		onlyone)
 {
-	return read_forward(fd, offset, count, total, verbose, onlyone, 0);
+	return read_forward(fd, offset, count, total, verbose, onlyone, 0, 0);
 }
 
 static int
@@ -371,6 +382,7 @@ pread_f(
 	int		Cflag, qflag, uflag, vflag;
 	int		eof = 0, direction = IO_FORWARD;
 	int		c;
+	int		preadv2_flags = 0;
 
 	Cflag = qflag = uflag = vflag = 0;
 	init_cvtnum(&fsblocksize, &fssectsize);
@@ -463,15 +475,18 @@ pread_f(
 	case IO_RANDOM:
 		if (!zeed)	/* srandom seed */
 			zeed = time(NULL);
-		c = read_random(file->fd, offset, count, &total, zeed, eof);
+		c = read_random(file->fd, offset, count, &total, zeed, eof,
+						preadv2_flags);
 		break;
 	case IO_FORWARD:
-		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof);
+		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof,
+						 preadv2_flags);
 		if (eof)
 			count = total;
 		break;
 	case IO_BACKWARD:
-		c = read_backward(file->fd, &offset, &count, &total, eof);
+		c = read_backward(file->fd, &offset, &count, &total, eof,
+						  preadv2_flags);
 		break;
 	default:
 		ASSERT(0);
-- 
2.48.1

[PATCH v2 2/3] xfs_io: Add RWF_DONTCACHE support to pwritev2

[Ritesh Harjani (IBM)]

Add per-io RWF_DONTCACHE support flag to pwritev2().
This enables xfs_io to perform uncached buffered-io writes.

e.g. xfs_io -fc "pwrite -U -V 1 0 16K" /mnt/f1

Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
---
 include/linux.h   |  5 +++++
 io/pwrite.c       | 14 ++++++++++++--
 man/man8/xfs_io.8 |  8 +++++++-
 3 files changed, 24 insertions(+), 3 deletions(-)

diff --git a/include/linux.h b/include/linux.h
index b3516d54..6e83e073 100644
--- a/include/linux.h
+++ b/include/linux.h
@@ -237,6 +237,11 @@ struct fsxattr {
 #define RWF_ATOMIC	((__kernel_rwf_t)0x00000040)
 #endif
 
+/* buffered IO that drops the cache after reading or writing data */
+#ifndef RWF_DONTCACHE
+#define RWF_DONTCACHE	((__kernel_rwf_t)0x00000080)
+#endif
+
 /*
  * Reminder: anything added to this file will be compiled into downstream
  * userspace projects!
diff --git a/io/pwrite.c b/io/pwrite.c
index fab59be4..5fb0253f 100644
--- a/io/pwrite.c
+++ b/io/pwrite.c
@@ -45,6 +45,7 @@ pwrite_help(void)
 " -N   -- Perform the pwritev2() with RWF_NOWAIT\n"
 " -D   -- Perform the pwritev2() with RWF_DSYNC\n"
 " -A   -- Perform the pwritev2() with RWF_ATOMIC\n"
+" -U   -- Perform the pwritev2() with Uncached/RWF_DONTCACHE\n"
 #endif
 "\n"));
 }
@@ -285,7 +286,7 @@ pwrite_f(
 	init_cvtnum(&fsblocksize, &fssectsize);
 	bsize = fsblocksize;
 
-	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uV:wWZ:")) != EOF) {
+	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uUV:wWZ:")) != EOF) {
 		switch (c) {
 		case 'b':
 			tmp = cvtnum(fsblocksize, fssectsize, optarg);
@@ -328,6 +329,9 @@ pwrite_f(
 		case 'A':
 			pwritev2_flags |= RWF_ATOMIC;
 			break;
+		case 'U':
+			pwritev2_flags |= RWF_DONTCACHE;
+			break;
 #endif
 		case 's':
 			skip = cvtnum(fsblocksize, fssectsize, optarg);
@@ -392,6 +396,12 @@ pwrite_f(
 		exitcode = 1;
 		return command_usage(&pwrite_cmd);
 	}
+	if (pwritev2_flags != 0 && vectors == 0) {
+		printf(_("pwritev2 flags require vectored I/O (-V)\n"));
+		exitcode = 1;
+		return command_usage(&pwrite_cmd);
+	}
+
 	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
 	if (offset < 0) {
 		printf(_("non-numeric offset argument -- %s\n"), argv[optind]);
@@ -480,7 +490,7 @@ pwrite_init(void)
 	pwrite_cmd.argmax = -1;
 	pwrite_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
 	pwrite_cmd.args =
-_("[-i infile [-qAdDwNOW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
+_("[-i infile [-qAdDwNOUW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
 	pwrite_cmd.oneline =
 		_("writes a number of bytes at a specified offset");
 	pwrite_cmd.help = pwrite_help;
diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
index 59d5ddc5..47af5232 100644
--- a/man/man8/xfs_io.8
+++ b/man/man8/xfs_io.8
@@ -244,7 +244,7 @@ See the
 .B pread
 command.
 .TP
-.BI "pwrite [ \-i " file " ] [ \-qAdDwNOW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
+.BI "pwrite [ \-i " file " ] [ \-qAdDwNOUW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
 Writes a range of bytes in a specified blocksize from the given
 .IR offset .
 The bytes written can be either a set pattern or read in from another
@@ -287,6 +287,12 @@ Perform the
 call with
 .IR RWF_ATOMIC .
 .TP
+.B \-U
+Perform the
+.BR pwritev2 (2)
+call with
+.IR RWF_DONTCACHE .
+.TP
 .B \-O
 perform pwrite once and return the (maybe partial) bytes written.
 .TP
-- 
2.48.1

[PATCH v2 3/3] xfs_io: Add RWF_DONTCACHE support to preadv2

[Ritesh Harjani (IBM)]

Add per-io RWF_DONTCACHE support flag to preadv2().
This enables xfs_io to perform uncached buffered-io reads.

	e.g. xfs_io -c "pread -U -V 1 0 16K" /mnt/f1

Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
---
 io/pread.c        | 17 +++++++++++++++--
 man/man8/xfs_io.8 |  8 +++++++-
 2 files changed, 22 insertions(+), 3 deletions(-)

diff --git a/io/pread.c b/io/pread.c
index b314fbc7..79e6570e 100644
--- a/io/pread.c
+++ b/io/pread.c
@@ -38,6 +38,9 @@ pread_help(void)
 " -Z N -- zeed the random number generator (used when reading randomly)\n"
 "         (heh, zorry, the -s/-S arguments were already in use in pwrite)\n"
 " -V N -- use vectored IO with N iovecs of blocksize each (preadv)\n"
+#ifdef HAVE_PREADV2
+" -U   -- Perform the preadv2() with Uncached/RWF_DONTCACHE\n"
+#endif
 "\n"
 " When in \"random\" mode, the number of read operations will equal the\n"
 " number required to do a complete forward/backward scan of the range.\n"
@@ -388,7 +391,7 @@ pread_f(
 	init_cvtnum(&fsblocksize, &fssectsize);
 	bsize = fsblocksize;
 
-	while ((c = getopt(argc, argv, "b:BCFRquvV:Z:")) != EOF) {
+	while ((c = getopt(argc, argv, "b:BCFRquUvV:Z:")) != EOF) {
 		switch (c) {
 		case 'b':
 			tmp = cvtnum(fsblocksize, fssectsize, optarg);
@@ -417,6 +420,11 @@ pread_f(
 		case 'u':
 			uflag = 1;
 			break;
+#ifdef HAVE_PREADV2
+		case 'U':
+			preadv2_flags |= RWF_DONTCACHE;
+			break;
+#endif
 		case 'v':
 			vflag = 1;
 			break;
@@ -446,6 +454,11 @@ pread_f(
 		exitcode = 1;
 		return command_usage(&pread_cmd);
 	}
+	if (preadv2_flags != 0 && vectors == 0) {
+		printf(_("preadv2 flags require vectored I/O (-V)\n"));
+		exitcode = 1;
+		return command_usage(&pread_cmd);
+	}
 
 	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
 	if (offset < 0 && (direction & (IO_RANDOM|IO_BACKWARD))) {
@@ -514,7 +527,7 @@ pread_init(void)
 	pread_cmd.argmin = 2;
 	pread_cmd.argmax = -1;
 	pread_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
-	pread_cmd.args = _("[-b bs] [-qv] [-i N] [-FBR [-Z N]] off len");
+	pread_cmd.args = _("[-b bs] [-qUv] [-i N] [-FBR [-Z N]] off len");
 	pread_cmd.oneline = _("reads a number of bytes at a specified offset");
 	pread_cmd.help = pread_help;
 
diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
index 47af5232..df508054 100644
--- a/man/man8/xfs_io.8
+++ b/man/man8/xfs_io.8
@@ -200,7 +200,7 @@ option will set the file permissions to read-write (0644). This allows xfs_io to
 set up mismatches between the file permissions and the open file descriptor
 read/write mode to exercise permission checks inside various syscalls.
 .TP
-.BI "pread [ \-b " bsize " ] [ \-qv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
+.BI "pread [ \-b " bsize " ] [ \-qUv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
 Reads a range of bytes in a specified blocksize from the given
 .IR offset .
 .RS 1.0i
@@ -214,6 +214,12 @@ requests will be split. The default blocksize is 4096 bytes.
 .B \-q
 quiet mode, do not write anything to standard output.
 .TP
+.B \-U
+Perform the
+.BR preadv2 (2)
+call with
+.IR RWF_DONTCACHE .
+.TP
 .B \-v
 dump the contents of the buffer after reading,
 by default only the count of bytes actually read is dumped.
-- 
2.48.1

Re: [PATCH v2 1/3] xfs_io: Add support for preadv2

[Darrick J. Wong]

On Wed, Mar 05, 2025 at 03:57:46PM +0530, Ritesh Harjani (IBM) wrote:
> This patch adds support for preadv2() to xfs_io.
> 
> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
> ---
>  io/Makefile |  2 +-
>  io/pread.c  | 45 ++++++++++++++++++++++++++++++---------------
>  2 files changed, 31 insertions(+), 16 deletions(-)
> 
> diff --git a/io/Makefile b/io/Makefile
> index 8f835ec7..14a3fe20 100644
> --- a/io/Makefile
> +++ b/io/Makefile
> @@ -66,7 +66,7 @@ LLDLIBS += $(LIBEDITLINE) $(LIBTERMCAP)
>  endif
>  
>  ifeq ($(HAVE_PWRITEV2),yes)
> -LCFLAGS += -DHAVE_PWRITEV2
> +LCFLAGS += -DHAVE_PWRITEV2 -DHAVE_PREADV2
>  endif
>  
>  ifeq ($(HAVE_MAP_SYNC),yes)
> diff --git a/io/pread.c b/io/pread.c
> index 62c771fb..b314fbc7 100644
> --- a/io/pread.c
> +++ b/io/pread.c
> @@ -162,7 +162,8 @@ static ssize_t
>  do_preadv(
>  	int		fd,
>  	off_t		offset,
> -	long long	count)
> +	long long	count,
> +	int			preadv2_flags)
>  {
>  	int		vecs = 0;
>  	ssize_t		oldlen = 0;
> @@ -181,8 +182,14 @@ do_preadv(
>  	} else {
>  		vecs = vectors;
>  	}
> +#ifdef HAVE_PREADV2
> +	if (preadv2_flags)
> +		bytes = preadv2(fd, iov, vectors, offset, preadv2_flags);
> +	else
> +		bytes = preadv(fd, iov, vectors, offset);
> +#else
>  	bytes = preadv(fd, iov, vectors, offset);
> -
> +#endif
>  	/* restore trimmed iov */
>  	if (oldlen)
>  		iov[vecs - 1].iov_len = oldlen;
> @@ -195,12 +202,13 @@ do_pread(
>  	int		fd,
>  	off_t		offset,
>  	long long	count,
> -	size_t		buffer_size)
> +	size_t		buffer_size,
> +	int			preadv2_flags)

Too much indenting here ^^ I think?

With that fixed,
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

>  {
>  	if (!vectors)
>  		return pread(fd, io_buffer, min(count, buffer_size), offset);
>  
> -	return do_preadv(fd, offset, count);
> +	return do_preadv(fd, offset, count, preadv2_flags);
>  }
>  
>  static int
> @@ -210,7 +218,8 @@ read_random(
>  	long long	count,
>  	long long	*total,
>  	unsigned int	seed,
> -	int		eof)
> +	int		eof,
> +	int		preadv2_flags)
>  {
>  	off_t		end, off, range;
>  	ssize_t		bytes;
> @@ -234,7 +243,7 @@ read_random(
>  				io_buffersize;
>  		else
>  			off = offset;
> -		bytes = do_pread(fd, off, io_buffersize, io_buffersize);
> +		bytes = do_pread(fd, off, io_buffersize, io_buffersize, preadv2_flags);
>  		if (bytes == 0)
>  			break;
>  		if (bytes < 0) {
> @@ -256,7 +265,8 @@ read_backward(
>  	off_t		*offset,
>  	long long	*count,
>  	long long	*total,
> -	int		eof)
> +	int		eof,
> +	int		preadv2_flags)
>  {
>  	off_t		end, off = *offset;
>  	ssize_t		bytes = 0, bytes_requested;
> @@ -276,7 +286,7 @@ read_backward(
>  	/* Do initial unaligned read if needed */
>  	if ((bytes_requested = (off % io_buffersize))) {
>  		off -= bytes_requested;
> -		bytes = do_pread(fd, off, bytes_requested, io_buffersize);
> +		bytes = do_pread(fd, off, bytes_requested, io_buffersize, preadv2_flags);
>  		if (bytes == 0)
>  			return ops;
>  		if (bytes < 0) {
> @@ -294,7 +304,7 @@ read_backward(
>  	while (cnt > end) {
>  		bytes_requested = min(cnt, io_buffersize);
>  		off -= bytes_requested;
> -		bytes = do_pread(fd, off, cnt, io_buffersize);
> +		bytes = do_pread(fd, off, cnt, io_buffersize, preadv2_flags);
>  		if (bytes == 0)
>  			break;
>  		if (bytes < 0) {
> @@ -318,14 +328,15 @@ read_forward(
>  	long long	*total,
>  	int		verbose,
>  	int		onlyone,
> -	int		eof)
> +	int		eof,
> +	int		preadv2_flags)
>  {
>  	ssize_t		bytes;
>  	int		ops = 0;
>  
>  	*total = 0;
>  	while (count > 0 || eof) {
> -		bytes = do_pread(fd, offset, count, io_buffersize);
> +		bytes = do_pread(fd, offset, count, io_buffersize, preadv2_flags);
>  		if (bytes == 0)
>  			break;
>  		if (bytes < 0) {
> @@ -353,7 +364,7 @@ read_buffer(
>  	int		verbose,
>  	int		onlyone)
>  {
> -	return read_forward(fd, offset, count, total, verbose, onlyone, 0);
> +	return read_forward(fd, offset, count, total, verbose, onlyone, 0, 0);
>  }
>  
>  static int
> @@ -371,6 +382,7 @@ pread_f(
>  	int		Cflag, qflag, uflag, vflag;
>  	int		eof = 0, direction = IO_FORWARD;
>  	int		c;
> +	int		preadv2_flags = 0;
>  
>  	Cflag = qflag = uflag = vflag = 0;
>  	init_cvtnum(&fsblocksize, &fssectsize);
> @@ -463,15 +475,18 @@ pread_f(
>  	case IO_RANDOM:
>  		if (!zeed)	/* srandom seed */
>  			zeed = time(NULL);
> -		c = read_random(file->fd, offset, count, &total, zeed, eof);
> +		c = read_random(file->fd, offset, count, &total, zeed, eof,
> +						preadv2_flags);
>  		break;
>  	case IO_FORWARD:
> -		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof);
> +		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof,
> +						 preadv2_flags);
>  		if (eof)
>  			count = total;
>  		break;
>  	case IO_BACKWARD:
> -		c = read_backward(file->fd, &offset, &count, &total, eof);
> +		c = read_backward(file->fd, &offset, &count, &total, eof,
> +						  preadv2_flags);
>  		break;
>  	default:
>  		ASSERT(0);
> -- 
> 2.48.1
> 
> 

Re: [PATCH v2 2/3] xfs_io: Add RWF_DONTCACHE support to pwritev2

[Darrick J. Wong]

On Wed, Mar 05, 2025 at 03:57:47PM +0530, Ritesh Harjani (IBM) wrote:
> Add per-io RWF_DONTCACHE support flag to pwritev2().
> This enables xfs_io to perform uncached buffered-io writes.
> 
> e.g. xfs_io -fc "pwrite -U -V 1 0 16K" /mnt/f1
> 
> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
> ---
>  include/linux.h   |  5 +++++
>  io/pwrite.c       | 14 ++++++++++++--
>  man/man8/xfs_io.8 |  8 +++++++-
>  3 files changed, 24 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux.h b/include/linux.h
> index b3516d54..6e83e073 100644
> --- a/include/linux.h
> +++ b/include/linux.h
> @@ -237,6 +237,11 @@ struct fsxattr {
>  #define RWF_ATOMIC	((__kernel_rwf_t)0x00000040)
>  #endif
>  
> +/* buffered IO that drops the cache after reading or writing data */
> +#ifndef RWF_DONTCACHE
> +#define RWF_DONTCACHE	((__kernel_rwf_t)0x00000080)
> +#endif
> +
>  /*
>   * Reminder: anything added to this file will be compiled into downstream
>   * userspace projects!
> diff --git a/io/pwrite.c b/io/pwrite.c
> index fab59be4..5fb0253f 100644
> --- a/io/pwrite.c
> +++ b/io/pwrite.c
> @@ -45,6 +45,7 @@ pwrite_help(void)
>  " -N   -- Perform the pwritev2() with RWF_NOWAIT\n"
>  " -D   -- Perform the pwritev2() with RWF_DSYNC\n"
>  " -A   -- Perform the pwritev2() with RWF_ATOMIC\n"
> +" -U   -- Perform the pwritev2() with Uncached/RWF_DONTCACHE\n"

I would have just said "...with RWF_DONTCACHE" because that's a lot more
precise.

With that shortened, this looks pretty straightforward.
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

>  #endif
>  "\n"));
>  }
> @@ -285,7 +286,7 @@ pwrite_f(
>  	init_cvtnum(&fsblocksize, &fssectsize);
>  	bsize = fsblocksize;
>  
> -	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uV:wWZ:")) != EOF) {
> +	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uUV:wWZ:")) != EOF) {
>  		switch (c) {
>  		case 'b':
>  			tmp = cvtnum(fsblocksize, fssectsize, optarg);
> @@ -328,6 +329,9 @@ pwrite_f(
>  		case 'A':
>  			pwritev2_flags |= RWF_ATOMIC;
>  			break;
> +		case 'U':
> +			pwritev2_flags |= RWF_DONTCACHE;
> +			break;
>  #endif
>  		case 's':
>  			skip = cvtnum(fsblocksize, fssectsize, optarg);
> @@ -392,6 +396,12 @@ pwrite_f(
>  		exitcode = 1;
>  		return command_usage(&pwrite_cmd);
>  	}
> +	if (pwritev2_flags != 0 && vectors == 0) {
> +		printf(_("pwritev2 flags require vectored I/O (-V)\n"));
> +		exitcode = 1;
> +		return command_usage(&pwrite_cmd);
> +	}
> +
>  	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
>  	if (offset < 0) {
>  		printf(_("non-numeric offset argument -- %s\n"), argv[optind]);
> @@ -480,7 +490,7 @@ pwrite_init(void)
>  	pwrite_cmd.argmax = -1;
>  	pwrite_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
>  	pwrite_cmd.args =
> -_("[-i infile [-qAdDwNOW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
> +_("[-i infile [-qAdDwNOUW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
>  	pwrite_cmd.oneline =
>  		_("writes a number of bytes at a specified offset");
>  	pwrite_cmd.help = pwrite_help;
> diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
> index 59d5ddc5..47af5232 100644
> --- a/man/man8/xfs_io.8
> +++ b/man/man8/xfs_io.8
> @@ -244,7 +244,7 @@ See the
>  .B pread
>  command.
>  .TP
> -.BI "pwrite [ \-i " file " ] [ \-qAdDwNOW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
> +.BI "pwrite [ \-i " file " ] [ \-qAdDwNOUW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
>  Writes a range of bytes in a specified blocksize from the given
>  .IR offset .
>  The bytes written can be either a set pattern or read in from another
> @@ -287,6 +287,12 @@ Perform the
>  call with
>  .IR RWF_ATOMIC .
>  .TP
> +.B \-U
> +Perform the
> +.BR pwritev2 (2)
> +call with
> +.IR RWF_DONTCACHE .
> +.TP
>  .B \-O
>  perform pwrite once and return the (maybe partial) bytes written.
>  .TP
> -- 
> 2.48.1
> 
> 

Re: [PATCH v2 3/3] xfs_io: Add RWF_DONTCACHE support to preadv2

[Darrick J. Wong]

On Wed, Mar 05, 2025 at 03:57:48PM +0530, Ritesh Harjani (IBM) wrote:
> Add per-io RWF_DONTCACHE support flag to preadv2().
> This enables xfs_io to perform uncached buffered-io reads.
> 
> 	e.g. xfs_io -c "pread -U -V 1 0 16K" /mnt/f1
> 
> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
> ---
>  io/pread.c        | 17 +++++++++++++++--
>  man/man8/xfs_io.8 |  8 +++++++-
>  2 files changed, 22 insertions(+), 3 deletions(-)
> 
> diff --git a/io/pread.c b/io/pread.c
> index b314fbc7..79e6570e 100644
> --- a/io/pread.c
> +++ b/io/pread.c
> @@ -38,6 +38,9 @@ pread_help(void)
>  " -Z N -- zeed the random number generator (used when reading randomly)\n"
>  "         (heh, zorry, the -s/-S arguments were already in use in pwrite)\n"
>  " -V N -- use vectored IO with N iovecs of blocksize each (preadv)\n"
> +#ifdef HAVE_PREADV2
> +" -U   -- Perform the preadv2() with Uncached/RWF_DONTCACHE\n"

Same comment as the last patch, but otherwise this looks good;
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

> +#endif
>  "\n"
>  " When in \"random\" mode, the number of read operations will equal the\n"
>  " number required to do a complete forward/backward scan of the range.\n"
> @@ -388,7 +391,7 @@ pread_f(
>  	init_cvtnum(&fsblocksize, &fssectsize);
>  	bsize = fsblocksize;
>  
> -	while ((c = getopt(argc, argv, "b:BCFRquvV:Z:")) != EOF) {
> +	while ((c = getopt(argc, argv, "b:BCFRquUvV:Z:")) != EOF) {
>  		switch (c) {
>  		case 'b':
>  			tmp = cvtnum(fsblocksize, fssectsize, optarg);
> @@ -417,6 +420,11 @@ pread_f(
>  		case 'u':
>  			uflag = 1;
>  			break;
> +#ifdef HAVE_PREADV2
> +		case 'U':
> +			preadv2_flags |= RWF_DONTCACHE;
> +			break;
> +#endif
>  		case 'v':
>  			vflag = 1;
>  			break;
> @@ -446,6 +454,11 @@ pread_f(
>  		exitcode = 1;
>  		return command_usage(&pread_cmd);
>  	}
> +	if (preadv2_flags != 0 && vectors == 0) {
> +		printf(_("preadv2 flags require vectored I/O (-V)\n"));
> +		exitcode = 1;
> +		return command_usage(&pread_cmd);
> +	}
>  
>  	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
>  	if (offset < 0 && (direction & (IO_RANDOM|IO_BACKWARD))) {
> @@ -514,7 +527,7 @@ pread_init(void)
>  	pread_cmd.argmin = 2;
>  	pread_cmd.argmax = -1;
>  	pread_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
> -	pread_cmd.args = _("[-b bs] [-qv] [-i N] [-FBR [-Z N]] off len");
> +	pread_cmd.args = _("[-b bs] [-qUv] [-i N] [-FBR [-Z N]] off len");
>  	pread_cmd.oneline = _("reads a number of bytes at a specified offset");
>  	pread_cmd.help = pread_help;
>  
> diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
> index 47af5232..df508054 100644
> --- a/man/man8/xfs_io.8
> +++ b/man/man8/xfs_io.8
> @@ -200,7 +200,7 @@ option will set the file permissions to read-write (0644). This allows xfs_io to
>  set up mismatches between the file permissions and the open file descriptor
>  read/write mode to exercise permission checks inside various syscalls.
>  .TP
> -.BI "pread [ \-b " bsize " ] [ \-qv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
> +.BI "pread [ \-b " bsize " ] [ \-qUv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
>  Reads a range of bytes in a specified blocksize from the given
>  .IR offset .
>  .RS 1.0i
> @@ -214,6 +214,12 @@ requests will be split. The default blocksize is 4096 bytes.
>  .B \-q
>  quiet mode, do not write anything to standard output.
>  .TP
> +.B \-U
> +Perform the
> +.BR preadv2 (2)
> +call with
> +.IR RWF_DONTCACHE .
> +.TP
>  .B \-v
>  dump the contents of the buffer after reading,
>  by default only the count of bytes actually read is dumped.
> -- 
> 2.48.1
> 
> 

Re: [PATCH v2 1/3] xfs_io: Add support for preadv2

[Ritesh Harjani]

"Darrick J. Wong" <djwong@kernel.org> writes:

> On Wed, Mar 05, 2025 at 03:57:46PM +0530, Ritesh Harjani (IBM) wrote:
>> This patch adds support for preadv2() to xfs_io.
>> 
>> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
>> ---
>>  io/Makefile |  2 +-
>>  io/pread.c  | 45 ++++++++++++++++++++++++++++++---------------
>>  2 files changed, 31 insertions(+), 16 deletions(-)
>> 
>> diff --git a/io/Makefile b/io/Makefile
>> index 8f835ec7..14a3fe20 100644
>> --- a/io/Makefile
>> +++ b/io/Makefile
>> @@ -66,7 +66,7 @@ LLDLIBS += $(LIBEDITLINE) $(LIBTERMCAP)
>>  endif
>>  
>>  ifeq ($(HAVE_PWRITEV2),yes)
>> -LCFLAGS += -DHAVE_PWRITEV2
>> +LCFLAGS += -DHAVE_PWRITEV2 -DHAVE_PREADV2
>>  endif
>>  
>>  ifeq ($(HAVE_MAP_SYNC),yes)
>> diff --git a/io/pread.c b/io/pread.c
>> index 62c771fb..b314fbc7 100644
>> --- a/io/pread.c
>> +++ b/io/pread.c
>> @@ -162,7 +162,8 @@ static ssize_t
>>  do_preadv(
>>  	int		fd,
>>  	off_t		offset,
>> -	long long	count)
>> +	long long	count,
>> +	int			preadv2_flags)
>>  {
>>  	int		vecs = 0;
>>  	ssize_t		oldlen = 0;
>> @@ -181,8 +182,14 @@ do_preadv(
>>  	} else {
>>  		vecs = vectors;
>>  	}
>> +#ifdef HAVE_PREADV2
>> +	if (preadv2_flags)
>> +		bytes = preadv2(fd, iov, vectors, offset, preadv2_flags);
>> +	else
>> +		bytes = preadv(fd, iov, vectors, offset);
>> +#else
>>  	bytes = preadv(fd, iov, vectors, offset);
>> -
>> +#endif
>>  	/* restore trimmed iov */
>>  	if (oldlen)
>>  		iov[vecs - 1].iov_len = oldlen;
>> @@ -195,12 +202,13 @@ do_pread(
>>  	int		fd,
>>  	off_t		offset,
>>  	long long	count,
>> -	size_t		buffer_size)
>> +	size_t		buffer_size,
>> +	int			preadv2_flags)
>
> Too much indenting here ^^ I think?
>

This is how I think git patch is showing. But the indentation is proper
when we apply the patch. In fact the "int fd" param is not properly
aligned to the rest of the params (lacking exactly 1 tab). Otherwise it
would have been easier to compare with "int fd".

> With that fixed,

So I don't think this needs any fixing. 

> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
>

Thanks for the review!

-ritesh

> --D
>
>>  {
>>  	if (!vectors)
>>  		return pread(fd, io_buffer, min(count, buffer_size), offset);
>>  
>> -	return do_preadv(fd, offset, count);
>> +	return do_preadv(fd, offset, count, preadv2_flags);
>>  }
>>  
>>  static int
>> @@ -210,7 +218,8 @@ read_random(
>>  	long long	count,
>>  	long long	*total,
>>  	unsigned int	seed,
>> -	int		eof)
>> +	int		eof,
>> +	int		preadv2_flags)
>>  {
>>  	off_t		end, off, range;
>>  	ssize_t		bytes;
>> @@ -234,7 +243,7 @@ read_random(
>>  				io_buffersize;
>>  		else
>>  			off = offset;
>> -		bytes = do_pread(fd, off, io_buffersize, io_buffersize);
>> +		bytes = do_pread(fd, off, io_buffersize, io_buffersize, preadv2_flags);
>>  		if (bytes == 0)
>>  			break;
>>  		if (bytes < 0) {
>> @@ -256,7 +265,8 @@ read_backward(
>>  	off_t		*offset,
>>  	long long	*count,
>>  	long long	*total,
>> -	int		eof)
>> +	int		eof,
>> +	int		preadv2_flags)
>>  {
>>  	off_t		end, off = *offset;
>>  	ssize_t		bytes = 0, bytes_requested;
>> @@ -276,7 +286,7 @@ read_backward(
>>  	/* Do initial unaligned read if needed */
>>  	if ((bytes_requested = (off % io_buffersize))) {
>>  		off -= bytes_requested;
>> -		bytes = do_pread(fd, off, bytes_requested, io_buffersize);
>> +		bytes = do_pread(fd, off, bytes_requested, io_buffersize, preadv2_flags);
>>  		if (bytes == 0)
>>  			return ops;
>>  		if (bytes < 0) {
>> @@ -294,7 +304,7 @@ read_backward(
>>  	while (cnt > end) {
>>  		bytes_requested = min(cnt, io_buffersize);
>>  		off -= bytes_requested;
>> -		bytes = do_pread(fd, off, cnt, io_buffersize);
>> +		bytes = do_pread(fd, off, cnt, io_buffersize, preadv2_flags);
>>  		if (bytes == 0)
>>  			break;
>>  		if (bytes < 0) {
>> @@ -318,14 +328,15 @@ read_forward(
>>  	long long	*total,
>>  	int		verbose,
>>  	int		onlyone,
>> -	int		eof)
>> +	int		eof,
>> +	int		preadv2_flags)
>>  {
>>  	ssize_t		bytes;
>>  	int		ops = 0;
>>  
>>  	*total = 0;
>>  	while (count > 0 || eof) {
>> -		bytes = do_pread(fd, offset, count, io_buffersize);
>> +		bytes = do_pread(fd, offset, count, io_buffersize, preadv2_flags);
>>  		if (bytes == 0)
>>  			break;
>>  		if (bytes < 0) {
>> @@ -353,7 +364,7 @@ read_buffer(
>>  	int		verbose,
>>  	int		onlyone)
>>  {
>> -	return read_forward(fd, offset, count, total, verbose, onlyone, 0);
>> +	return read_forward(fd, offset, count, total, verbose, onlyone, 0, 0);
>>  }
>>  
>>  static int
>> @@ -371,6 +382,7 @@ pread_f(
>>  	int		Cflag, qflag, uflag, vflag;
>>  	int		eof = 0, direction = IO_FORWARD;
>>  	int		c;
>> +	int		preadv2_flags = 0;
>>  
>>  	Cflag = qflag = uflag = vflag = 0;
>>  	init_cvtnum(&fsblocksize, &fssectsize);
>> @@ -463,15 +475,18 @@ pread_f(
>>  	case IO_RANDOM:
>>  		if (!zeed)	/* srandom seed */
>>  			zeed = time(NULL);
>> -		c = read_random(file->fd, offset, count, &total, zeed, eof);
>> +		c = read_random(file->fd, offset, count, &total, zeed, eof,
>> +						preadv2_flags);
>>  		break;
>>  	case IO_FORWARD:
>> -		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof);
>> +		c = read_forward(file->fd, offset, count, &total, vflag, 0, eof,
>> +						 preadv2_flags);
>>  		if (eof)
>>  			count = total;
>>  		break;
>>  	case IO_BACKWARD:
>> -		c = read_backward(file->fd, &offset, &count, &total, eof);
>> +		c = read_backward(file->fd, &offset, &count, &total, eof,
>> +						  preadv2_flags);
>>  		break;
>>  	default:
>>  		ASSERT(0);
>> -- 
>> 2.48.1
>> 
>> 

Re: [PATCH v2 2/3] xfs_io: Add RWF_DONTCACHE support to pwritev2

[Ritesh Harjani]

"Darrick J. Wong" <djwong@kernel.org> writes:

> On Wed, Mar 05, 2025 at 03:57:47PM +0530, Ritesh Harjani (IBM) wrote:
>> Add per-io RWF_DONTCACHE support flag to pwritev2().
>> This enables xfs_io to perform uncached buffered-io writes.
>> 
>> e.g. xfs_io -fc "pwrite -U -V 1 0 16K" /mnt/f1
>> 
>> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
>> ---
>>  include/linux.h   |  5 +++++
>>  io/pwrite.c       | 14 ++++++++++++--
>>  man/man8/xfs_io.8 |  8 +++++++-
>>  3 files changed, 24 insertions(+), 3 deletions(-)
>> 
>> diff --git a/include/linux.h b/include/linux.h
>> index b3516d54..6e83e073 100644
>> --- a/include/linux.h
>> +++ b/include/linux.h
>> @@ -237,6 +237,11 @@ struct fsxattr {
>>  #define RWF_ATOMIC	((__kernel_rwf_t)0x00000040)
>>  #endif
>>  
>> +/* buffered IO that drops the cache after reading or writing data */
>> +#ifndef RWF_DONTCACHE
>> +#define RWF_DONTCACHE	((__kernel_rwf_t)0x00000080)
>> +#endif
>> +
>>  /*
>>   * Reminder: anything added to this file will be compiled into downstream
>>   * userspace projects!
>> diff --git a/io/pwrite.c b/io/pwrite.c
>> index fab59be4..5fb0253f 100644
>> --- a/io/pwrite.c
>> +++ b/io/pwrite.c
>> @@ -45,6 +45,7 @@ pwrite_help(void)
>>  " -N   -- Perform the pwritev2() with RWF_NOWAIT\n"
>>  " -D   -- Perform the pwritev2() with RWF_DSYNC\n"
>>  " -A   -- Perform the pwritev2() with RWF_ATOMIC\n"
>> +" -U   -- Perform the pwritev2() with Uncached/RWF_DONTCACHE\n"
>
> I would have just said "...with RWF_DONTCACHE" because that's a lot more
> precise.
>

Yes, probably I was just overthinking and trying to give a reason here of
choosing -U (Uncached). 

> With that shortened, this looks pretty straightforward.

Sure will fix in v3.

> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

Thanks!

-ritesh

>
> --D
>
>>  #endif
>>  "\n"));
>>  }
>> @@ -285,7 +286,7 @@ pwrite_f(
>>  	init_cvtnum(&fsblocksize, &fssectsize);
>>  	bsize = fsblocksize;
>>  
>> -	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uV:wWZ:")) != EOF) {
>> +	while ((c = getopt(argc, argv, "Ab:BCdDf:Fi:NqRs:OS:uUV:wWZ:")) != EOF) {
>>  		switch (c) {
>>  		case 'b':
>>  			tmp = cvtnum(fsblocksize, fssectsize, optarg);
>> @@ -328,6 +329,9 @@ pwrite_f(
>>  		case 'A':
>>  			pwritev2_flags |= RWF_ATOMIC;
>>  			break;
>> +		case 'U':
>> +			pwritev2_flags |= RWF_DONTCACHE;
>> +			break;
>>  #endif
>>  		case 's':
>>  			skip = cvtnum(fsblocksize, fssectsize, optarg);
>> @@ -392,6 +396,12 @@ pwrite_f(
>>  		exitcode = 1;
>>  		return command_usage(&pwrite_cmd);
>>  	}
>> +	if (pwritev2_flags != 0 && vectors == 0) {
>> +		printf(_("pwritev2 flags require vectored I/O (-V)\n"));
>> +		exitcode = 1;
>> +		return command_usage(&pwrite_cmd);
>> +	}
>> +
>>  	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
>>  	if (offset < 0) {
>>  		printf(_("non-numeric offset argument -- %s\n"), argv[optind]);
>> @@ -480,7 +490,7 @@ pwrite_init(void)
>>  	pwrite_cmd.argmax = -1;
>>  	pwrite_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
>>  	pwrite_cmd.args =
>> -_("[-i infile [-qAdDwNOW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
>> +_("[-i infile [-qAdDwNOUW] [-s skip]] [-b bs] [-S seed] [-FBR [-Z N]] [-V N] off len");
>>  	pwrite_cmd.oneline =
>>  		_("writes a number of bytes at a specified offset");
>>  	pwrite_cmd.help = pwrite_help;
>> diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
>> index 59d5ddc5..47af5232 100644
>> --- a/man/man8/xfs_io.8
>> +++ b/man/man8/xfs_io.8
>> @@ -244,7 +244,7 @@ See the
>>  .B pread
>>  command.
>>  .TP
>> -.BI "pwrite [ \-i " file " ] [ \-qAdDwNOW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
>> +.BI "pwrite [ \-i " file " ] [ \-qAdDwNOUW ] [ \-s " skip " ] [ \-b " size " ] [ \-S " seed " ] [ \-FBR [ \-Z " zeed " ] ] [ \-V " vectors " ] " "offset length"
>>  Writes a range of bytes in a specified blocksize from the given
>>  .IR offset .
>>  The bytes written can be either a set pattern or read in from another
>> @@ -287,6 +287,12 @@ Perform the
>>  call with
>>  .IR RWF_ATOMIC .
>>  .TP
>> +.B \-U
>> +Perform the
>> +.BR pwritev2 (2)
>> +call with
>> +.IR RWF_DONTCACHE .
>> +.TP
>>  .B \-O
>>  perform pwrite once and return the (maybe partial) bytes written.
>>  .TP
>> -- 
>> 2.48.1
>> 
>> 

Re: [PATCH v2 3/3] xfs_io: Add RWF_DONTCACHE support to preadv2

[Ritesh Harjani]

"Darrick J. Wong" <djwong@kernel.org> writes:

> On Wed, Mar 05, 2025 at 03:57:48PM +0530, Ritesh Harjani (IBM) wrote:
>> Add per-io RWF_DONTCACHE support flag to preadv2().
>> This enables xfs_io to perform uncached buffered-io reads.
>> 
>> 	e.g. xfs_io -c "pread -U -V 1 0 16K" /mnt/f1
>> 
>> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
>> ---
>>  io/pread.c        | 17 +++++++++++++++--
>>  man/man8/xfs_io.8 |  8 +++++++-
>>  2 files changed, 22 insertions(+), 3 deletions(-)
>> 
>> diff --git a/io/pread.c b/io/pread.c
>> index b314fbc7..79e6570e 100644
>> --- a/io/pread.c
>> +++ b/io/pread.c
>> @@ -38,6 +38,9 @@ pread_help(void)
>>  " -Z N -- zeed the random number generator (used when reading randomly)\n"
>>  "         (heh, zorry, the -s/-S arguments were already in use in pwrite)\n"
>>  " -V N -- use vectored IO with N iovecs of blocksize each (preadv)\n"
>> +#ifdef HAVE_PREADV2
>> +" -U   -- Perform the preadv2() with Uncached/RWF_DONTCACHE\n"
>
> Same comment as the last patch, but otherwise this looks good;

Sure will do in v3.

> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

Thanks!
-ritesh

>
> --D
>
>> +#endif
>>  "\n"
>>  " When in \"random\" mode, the number of read operations will equal the\n"
>>  " number required to do a complete forward/backward scan of the range.\n"
>> @@ -388,7 +391,7 @@ pread_f(
>>  	init_cvtnum(&fsblocksize, &fssectsize);
>>  	bsize = fsblocksize;
>>  
>> -	while ((c = getopt(argc, argv, "b:BCFRquvV:Z:")) != EOF) {
>> +	while ((c = getopt(argc, argv, "b:BCFRquUvV:Z:")) != EOF) {
>>  		switch (c) {
>>  		case 'b':
>>  			tmp = cvtnum(fsblocksize, fssectsize, optarg);
>> @@ -417,6 +420,11 @@ pread_f(
>>  		case 'u':
>>  			uflag = 1;
>>  			break;
>> +#ifdef HAVE_PREADV2
>> +		case 'U':
>> +			preadv2_flags |= RWF_DONTCACHE;
>> +			break;
>> +#endif
>>  		case 'v':
>>  			vflag = 1;
>>  			break;
>> @@ -446,6 +454,11 @@ pread_f(
>>  		exitcode = 1;
>>  		return command_usage(&pread_cmd);
>>  	}
>> +	if (preadv2_flags != 0 && vectors == 0) {
>> +		printf(_("preadv2 flags require vectored I/O (-V)\n"));
>> +		exitcode = 1;
>> +		return command_usage(&pread_cmd);
>> +	}
>>  
>>  	offset = cvtnum(fsblocksize, fssectsize, argv[optind]);
>>  	if (offset < 0 && (direction & (IO_RANDOM|IO_BACKWARD))) {
>> @@ -514,7 +527,7 @@ pread_init(void)
>>  	pread_cmd.argmin = 2;
>>  	pread_cmd.argmax = -1;
>>  	pread_cmd.flags = CMD_NOMAP_OK | CMD_FOREIGN_OK;
>> -	pread_cmd.args = _("[-b bs] [-qv] [-i N] [-FBR [-Z N]] off len");
>> +	pread_cmd.args = _("[-b bs] [-qUv] [-i N] [-FBR [-Z N]] off len");
>>  	pread_cmd.oneline = _("reads a number of bytes at a specified offset");
>>  	pread_cmd.help = pread_help;
>>  
>> diff --git a/man/man8/xfs_io.8 b/man/man8/xfs_io.8
>> index 47af5232..df508054 100644
>> --- a/man/man8/xfs_io.8
>> +++ b/man/man8/xfs_io.8
>> @@ -200,7 +200,7 @@ option will set the file permissions to read-write (0644). This allows xfs_io to
>>  set up mismatches between the file permissions and the open file descriptor
>>  read/write mode to exercise permission checks inside various syscalls.
>>  .TP
>> -.BI "pread [ \-b " bsize " ] [ \-qv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
>> +.BI "pread [ \-b " bsize " ] [ \-qUv ] [ \-FBR [ \-Z " seed " ] ] [ \-V " vectors " ] " "offset length"
>>  Reads a range of bytes in a specified blocksize from the given
>>  .IR offset .
>>  .RS 1.0i
>> @@ -214,6 +214,12 @@ requests will be split. The default blocksize is 4096 bytes.
>>  .B \-q
>>  quiet mode, do not write anything to standard output.
>>  .TP
>> +.B \-U
>> +Perform the
>> +.BR preadv2 (2)
>> +call with
>> +.IR RWF_DONTCACHE .
>> +.TP
>>  .B \-v
>>  dump the contents of the buffer after reading,
>>  by default only the count of bytes actually read is dumped.
>> -- 
>> 2.48.1
>> 
>>