MDP major registration

MDP major registration

[Hubert Verstraete]

Neil,

What is the status of the major for the partitionable arrays ?
I see that it is 254, which is in the experimental section, according to 
the official Linux device list (http://www.lanana.org/docs/device-list/).
Will there be an official registration ?

Regards,
Hubert

Re: MDP major registration

[Neil Brown]

On Thursday March 13, hubskml@free.fr wrote:
> Neil,
> 
> What is the status of the major for the partitionable arrays ?

automatically determined at runtime.

> I see that it is 254, which is in the experimental section, according to 
> the official Linux device list (http://www.lanana.org/docs/device-list/).
> Will there be an official registration ?

No.  Is there any need?

NeilBrown

Re: MDP major registration

[Hubert Verstraete]

Neil Brown wrote:
> On Thursday March 13, hubskml@free.fr wrote:
>> Neil,
>>
>> What is the status of the major for the partitionable arrays ?
>
> automatically determined at runtime.
>
>> I see that it is 254, which is in the experimental section, according to 
>> the official Linux device list (http://www.lanana.org/docs/device-list/).
>> Will there be an official registration ?
> 
> No. Is there any need?

I got this question in mind when I saw that mkfs.xfs source code was 
referring to the MD major to tune its parameters on an MD device, while 
it ignores MDP devices.
If there were reasons to register MD, wouldn't they apply to MDP too ?

Regards,
Hubert

Re: MDP major registration

[Luca Berra]

On Tue, Mar 25, 2008 at 05:57:06PM +0100, Hubert Verstraete wrote:
>Neil Brown wrote:
>>On Thursday March 13, hubskml@free.fr wrote:
>>>Neil,
>>>
>>>What is the status of the major for the partitionable arrays ?
>>
>>automatically determined at runtime.
>>
>>>I see that it is 254, which is in the experimental section, according to 
>>>the official Linux device list (http://www.lanana.org/docs/device-list/).
>>>Will there be an official registration ?
>>
>>No. Is there any need?
>
>I got this question in mind when I saw that mkfs.xfs source code was 
>referring to the MD major to tune its parameters on an MD device, while 
>it ignores MDP devices.
>If there were reasons to register MD, wouldn't they apply to MDP too ?

i don't think so:
bluca@percy ~ $ grep mdp /proc/devices
253 mdp

L.


-- 
Luca Berra -- bluca@comedia.it
        Communication Media & Services S.r.l.
 /"\
 \ /     ASCII RIBBON CAMPAIGN
  X        AGAINST HTML MAIL
 / \

Re: MDP major registration

[Bill Davidsen]

Luca Berra wrote:
> On Tue, Mar 25, 2008 at 05:57:06PM +0100, Hubert Verstraete wrote:
>> Neil Brown wrote:
>>> On Thursday March 13, hubskml@free.fr wrote:
>>>> Neil,
>>>>
>>>> What is the status of the major for the partitionable arrays ?
>>>
>>> automatically determined at runtime.
>>>
>>>> I see that it is 254, which is in the experimental section, 
>>>> according to the official Linux device list 
>>>> (http://www.lanana.org/docs/device-list/).
>>>> Will there be an official registration ?
>>>
>>> No. Is there any need?
>>
>> I got this question in mind when I saw that mkfs.xfs source code was 
>> referring to the MD major to tune its parameters on an MD device, 
>> while it ignores MDP devices.
>> If there were reasons to register MD, wouldn't they apply to MDP too ?
>
> i don't think so:
> bluca@percy ~ $ grep mdp /proc/devices
> 253 mdp

Why is it important to have XFS tune its parameters for md and not for 
mdp? I don't understand your conclusion here, is tuning not needed for 
mdp, or so meaningless that it doesn't matter, or that XFS code reads 
/proc/devices, or ??? I note that device-mapper also has a dynamic 
major, what does XFS make of that?

I don't know how much difference tuning makes, but if it's worth doing 
at all, it should be done for mdp as well, I would think.

-- 
Bill Davidsen <davidsen@tmr.com>
  "Woe unto the statesman who makes war without a reason that will still
  be valid when the war is over..." Otto von Bismark 

Re: MDP major registration

[Hubert Verstraete]

Bill Davidsen wrote:
> Luca Berra wrote:
>> On Tue, Mar 25, 2008 at 05:57:06PM +0100, Hubert Verstraete wrote:
>>> Neil Brown wrote:
>>>> On Thursday March 13, hubskml@free.fr wrote:
>>>>> Neil,
>>>>>
>>>>> What is the status of the major for the partitionable arrays ?
>>>>
>>>> automatically determined at runtime.
>>>>
>>>>> I see that it is 254, which is in the experimental section, 
>>>>> according to the official Linux device list 
>>>>> (http://www.lanana.org/docs/device-list/).
>>>>> Will there be an official registration ?
>>>>
>>>> No. Is there any need?
>>>
>>> I got this question in mind when I saw that mkfs.xfs source code was 
>>> referring to the MD major to tune its parameters on an MD device, 
>>> while it ignores MDP devices.
>>> If there were reasons to register MD, wouldn't they apply to MDP too ?
>>
>> i don't think so:
>> bluca@percy ~ $ grep mdp /proc/devices
>> 253 mdp
> 
> Why is it important to have XFS tune its parameters for md and not for 
> mdp? I don't understand your conclusion here, is tuning not needed for 
> mdp, or so meaningless that it doesn't matter, or that XFS code reads 
> /proc/devices, or ??? I note that device-mapper also has a dynamic 
> major, what does XFS make of that?

It reads from /proc/devices.

> I don't know how much difference tuning makes, but if it's worth doing 
> at all, it should be done for mdp as well, I would think.

Same thought. I wrote the patch for mkfs.xfs but did not publish it for 
two reasons:
1) MD is registered but not MDP. Now I understand, it's not a problem, 
we just need to read /proc/devices as device-mapper does.
2) Tuning XFS for MDP can be achieved through the mkfs.xfs options. With 
a few lines in shell, my XFS on MDP now has the same performance as XFS 
on MD.

Hubert

[PATCH] XFS tuning on software RAID5 partitionable array; was: MDP major registration

[Hubert Verstraete]

[-- Attachment #1: Type: text/plain, Size: 2224 bytes --]

Hi XFS list,

please find attached a patch for libdisk/mkfs.xfs which tunes XFS on 
software partitionable RAID arrays, also called mdp.

Hubert Verstraete

Hubert Verstraete wrote:
> Bill Davidsen wrote:
>> Luca Berra wrote:
>>> On Tue, Mar 25, 2008 at 05:57:06PM +0100, Hubert Verstraete wrote:
>>>> Neil Brown wrote:
>>>>> On Thursday March 13, hubskml@free.fr wrote:
>>>>>> Neil,
>>>>>>
>>>>>> What is the status of the major for the partitionable arrays ?
>>>>>
>>>>> automatically determined at runtime.
>>>>>
>>>>>> I see that it is 254, which is in the experimental section, 
>>>>>> according to the official Linux device list 
>>>>>> (http://www.lanana.org/docs/device-list/).
>>>>>> Will there be an official registration ?
>>>>>
>>>>> No. Is there any need?
>>>>
>>>> I got this question in mind when I saw that mkfs.xfs source code was 
>>>> referring to the MD major to tune its parameters on an MD device, 
>>>> while it ignores MDP devices.
>>>> If there were reasons to register MD, wouldn't they apply to MDP too ?
>>>
>>> i don't think so:
>>> bluca@percy ~ $ grep mdp /proc/devices
>>> 253 mdp
>>
>> Why is it important to have XFS tune its parameters for md and not for 
>> mdp? I don't understand your conclusion here, is tuning not needed for 
>> mdp, or so meaningless that it doesn't matter, or that XFS code reads 
>> /proc/devices, or ??? I note that device-mapper also has a dynamic 
>> major, what does XFS make of that?
> 
> It reads from /proc/devices.
> 
>> I don't know how much difference tuning makes, but if it's worth doing 
>> at all, it should be done for mdp as well, I would think.
> 
> Same thought. I wrote the patch for mkfs.xfs but did not publish it for 
> two reasons:
> 1) MD is registered but not MDP. Now I understand, it's not a problem, 
> we just need to read /proc/devices as device-mapper does.
> 2) Tuning XFS for MDP can be achieved through the mkfs.xfs options. With 
> a few lines in shell, my XFS on MDP now has the same performance as XFS 
> on MD.
> 
> Hubert
> -- 
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

[-- Attachment #2: xfsprogs_mdp.patch --]
[-- Type: text/x-patch, Size: 2082 bytes --]

diff -u -r xfsprogs-2.8.11/libdisk/md.c xfsprogs-2.8.11-mdp/libdisk/md.c
--- xfsprogs-2.8.11/libdisk/md.c	2006-06-26 07:01:15.000000000 +0200
+++ xfsprogs-2.8.11-mdp/libdisk/md.c	2008-03-26 20:12:38.000000000 +0100
@@ -24,8 +24,12 @@
 	dev_t		dev)
 {
 	if (major(dev) == MD_MAJOR)
-		return 1;
-	return get_driver_block_major("md", major(dev));
+		return MD_IS_MD;
+	if (get_driver_block_major("md", major(dev)))
+		return MD_IS_MD;
+	if (get_driver_block_major("mdp", major(dev)))
+		return MD_IS_MDP;
+	return 0;
 }
 
 int
@@ -37,12 +41,32 @@
 	int		*sectalign,
 	struct stat64	*sb)
 {
-	if (mnt_is_md_subvol(sb->st_rdev)) {
+	char *pc, *dfile2 = NULL;
+	int is_md;
+
+	if ((is_md = mnt_is_md_subvol(sb->st_rdev))) {
 		struct md_array_info	md;
 		int			fd;
 
+		if (is_md == MD_IS_MDP) {
+			if (!(pc = strrchr(dfile, 'd'))
+			|| !(pc = strchr(pc, 'p'))) {
+				fprintf(stderr,
+					_("Error getting MD array device from %s\n"),
+					dfile);
+				exit(1);
+			}
+			dfile2 = (char *) malloc(pc - dfile + 1);
+			if (dfile2 == NULL) {
+				fprintf(stderr,
+					_("Couldn't malloc device string\n"));
+				exit(1);
+			}
+			strncpy(dfile2, dfile, pc - dfile);
+			dfile2[pc - dfile + 1] = '\0';
+		}
 		/* Open device */
-		fd = open(dfile, O_RDONLY);
+		fd = open(dfile2 ? dfile2 : dfile, O_RDONLY);
 		if (fd == -1)
 			return 0;
 
@@ -50,10 +74,11 @@
 		if (ioctl(fd, GET_ARRAY_INFO, &md)) {
 			fprintf(stderr,
 				_("Error getting MD array info from %s\n"),
-				dfile);
+				dfile2 ? dfile2 : dfile);
 			exit(1);
 		}
 		close(fd);
+		if (dfile2) free(dfile2);
 
 		/*
 		 * Ignore levels we don't want aligned (e.g. linear)
diff -u -r xfsprogs-2.8.11/libdisk/md.h xfsprogs-2.8.11-mdp/libdisk/md.h
--- xfsprogs-2.8.11/libdisk/md.h	2006-06-26 07:01:15.000000000 +0200
+++ xfsprogs-2.8.11-mdp/libdisk/md.h	2008-03-26 20:12:10.000000000 +0100
@@ -20,6 +20,9 @@
 #define MD_MAJOR		9 /* we also check at runtime */
 #endif
 
+#define MD_IS_MD		1
+#define MD_IS_MDP		2
+
 #define GET_ARRAY_INFO          _IOR (MD_MAJOR, 0x11, struct md_array_info)
 
 #define MD_SB_CLEAN		0

Re: MDP major registration

[Bill Davidsen]

Hubert Verstraete wrote:
> Bill Davidsen wrote:
>> Luca Berra wrote:
>>> On Tue, Mar 25, 2008 at 05:57:06PM +0100, Hubert Verstraete wrote:
>>>> Neil Brown wrote:
>>>>> On Thursday March 13, hubskml@free.fr wrote:
>>>>>> Neil,
>>>>>>
>>>>>> What is the status of the major for the partitionable arrays ?
>>>>>
>>>>> automatically determined at runtime.
>>>>>
>>>>>> I see that it is 254, which is in the experimental section, 
>>>>>> according to the official Linux device list 
>>>>>> (http://www.lanana.org/docs/device-list/).
>>>>>> Will there be an official registration ?
>>>>>
>>>>> No. Is there any need?
>>>>
>>>> I got this question in mind when I saw that mkfs.xfs source code 
>>>> was referring to the MD major to tune its parameters on an MD 
>>>> device, while it ignores MDP devices.
>>>> If there were reasons to register MD, wouldn't they apply to MDP too ?
>>>
>>> i don't think so:
>>> bluca@percy ~ $ grep mdp /proc/devices
>>> 253 mdp
>>
>> Why is it important to have XFS tune its parameters for md and not 
>> for mdp? I don't understand your conclusion here, is tuning not 
>> needed for mdp, or so meaningless that it doesn't matter, or that XFS 
>> code reads /proc/devices, or ??? I note that device-mapper also has a 
>> dynamic major, what does XFS make of that?
>
> It reads from /proc/devices.
>
>> I don't know how much difference tuning makes, but if it's worth 
>> doing at all, it should be done for mdp as well, I would think.
>
> Same thought. I wrote the patch for mkfs.xfs but did not publish it 
> for two reasons:
> 1) MD is registered but not MDP. Now I understand, it's not a problem, 
> we just need to read /proc/devices as device-mapper does.
> 2) Tuning XFS for MDP can be achieved through the mkfs.xfs options. 
> With a few lines in shell, my XFS on MDP now has the same performance 
> as XFS on MD.

Hopefully the following patch will be picked up by vendors. Now from the 
linux-kernel list there have been two recent changes in similar areas, 
the ability of loop to support partitionable devices, and the ability to 
have nbd handle partitionable devices. Are XFS users so lucky that this 
change would also improve things for those media?

-- 
Bill Davidsen <davidsen@tmr.com>
  "Woe unto the statesman who makes war without a reason that will still
  be valid when the war is over..." Otto von Bismark