diff mbox series

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

Message ID 57bd6d327ac8ed2f8e9859f1e42775622a8b9d09.1741170031.git.ritesh.list@gmail.com (mailing list archive)
State New
Headers show
Series xfsprogs: Add support for preadv2() and RWF_DONTCACHE | expand

Commit Message

Ritesh Harjani (IBM) March 5, 2025, 10:27 a.m. UTC 
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(-)

Comments

Darrick J. Wong March 5, 2025, 6:10 p.m. UTC | #1 
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
> 
>
Ritesh Harjani (IBM) March 5, 2025, 7:11 p.m. UTC | #2 
"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
>> 
>>
diff mbox series

Patch

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