Skip to contentSkip to conversation

partition sysfs OOPS in current GIT

DISCUSSION David Miller 4 participants 30 messages Feb 6, 2008 – Feb 7, 2008

lore
Share & tools
lore
#0 lore raw

I get the following OOPS from udevd during bootup on sparc64:

[    0.982046]               \|/ ____ \|/
[    0.982054]               "@'/ .. \`@"
[    0.982058]               /_| \__/ |_\
[    0.982063]                  \__U_/
[    0.982482] udevd(1305): Kernel illegal instruction [#1]
[    0.982550] TSTATE: 0000004411001602 TPC: 00000000007ddb80 TNPC: 00000000007ddb84 Y: 00000000    Not tainted
[    0.982647] TPC: <0x7ddb88>
[    0.982684] g0: 0000000000000000 g1: 00000000007ddb80 g2: 0000000000000000 g3: 0000000000000000
[    0.982760] g4: fffff803fb332a00 g5: fffff8000f8a6000 g6: fffff803fb3a8000 g7: fffff803fc893724
[    0.982831] o0: fffff803fc8bc010 o1: 00000000007ddb58 o2: fffff803fbba0000 o3: fffff803fb3abd48
[    0.982907] o4: 0000000000040001 o5: 0000000000000000 sp: fffff803fb3ab391 ret_pc: 00000000005b7524
[    0.982984] RPC: <dev_attr_show+0x24/0x30>
[    0.983016] l0: 0000000000000000 l1: 0000000000000000 l2: fffff803fc9a5308 l3: 000000000005a000
[    0.983074] l4: 0000000000000004 l5: fffff803fb340b00 l6: 0000000000000168 l7: fffff803fbd32700
[    0.983133] i0: fffffffffffffffb i1: 00000000007ddb58 i2: fffff803fbba0000 i3: 0000000000000000
[    0.983197] i4: 00000000005002a4 i5: 0000000000000008 i6: fffff803fb3ab451 i7: 0000000000500620
[    0.983269] I7: <sysfs_read_file+0x80/0x104>
[    0.983299] Caller[0000000000500620]: sysfs_read_file+0x80/0x104
[    0.983368] Caller[00000000004bedc0]: vfs_read+0x78/0x10c
[    0.983556] Caller[00000000004bf114]: sys_read+0x34/0x60
[    0.983622] Caller[0000000000406294]: linux_sparc_syscall32+0x3c/0x40
[    0.983707] Caller[0000000000016d90]: 0x16d98
[    0.983873] Instruction DUMP: 007ddb80  00000000  00000000 <00000000> 00000000  00000000  00000000  00000000  007ddb98 

This is dev_attr_show() calling attr->show() which points to 'part_attr_group' struct instead of a function. :-)

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200
    Driver core: convert block from raw kobjects to core devices
    
    This moves the block devices to /sys/class/block. It will create a
    flat list of all block devices, with the disks and partitions in one
    directory. For compatibility /sys/block is created and contains symlinks
    to the disks.
    
      /sys/class/block
      |-- sda -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda
      |-- sda1 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda1
      |-- sda10 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda10
      |-- sda5 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda5
      |-- sda6 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda6
      |-- sda7 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda7
      |-- sda8 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda8
      |-- sda9 -> ../../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/sda9
      `-- sr0 -> ../../devices/pci0000:00/0000:00:1f.2/host1/target1:0:0/1:0:0:0/block/sr0
    
      /sys/block/
      |-- sda -> ../devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda
      `-- sr0 -> ../devices/pci0000:00/0000:00:1f.2/host1/target1:0:0/1:0:0:0/block/sr0
    
    Signed-off-by: Kay Sievers <[email protected]>
    Signed-off-by: Greg Kroah-Hartman <[email protected]>
#1 lore raw
Quoted 28 lines · David Miller

I get the following OOPS from udevd during bootup on sparc64:

[    0.982046]               \|/ ____ \|/
[    0.982054]               "@'/ .. \`@"
[    0.982058]               /_| \__/ |_\
[    0.982063]                  \__U_/
[    0.982482] udevd(1305): Kernel illegal instruction [#1]
[    0.982550] TSTATE: 0000004411001602 TPC: 00000000007ddb80 TNPC: 00000000007ddb84 Y: 00000000    Not tainted
[    0.982647] TPC: <0x7ddb88>
[    0.982684] g0: 0000000000000000 g1: 00000000007ddb80 g2: 0000000000000000 g3: 0000000000000000
[    0.982760] g4: fffff803fb332a00 g5: fffff8000f8a6000 g6: fffff803fb3a8000 g7: fffff803fc893724
[    0.982831] o0: fffff803fc8bc010 o1: 00000000007ddb58 o2: fffff803fbba0000 o3: fffff803fb3abd48
[    0.982907] o4: 0000000000040001 o5: 0000000000000000 sp: fffff803fb3ab391 ret_pc: 00000000005b7524
[    0.982984] RPC: <dev_attr_show+0x24/0x30>
[    0.983016] l0: 0000000000000000 l1: 0000000000000000 l2: fffff803fc9a5308 l3: 000000000005a000
[    0.983074] l4: 0000000000000004 l5: fffff803fb340b00 l6: 0000000000000168 l7: fffff803fbd32700
[    0.983133] i0: fffffffffffffffb i1: 00000000007ddb58 i2: fffff803fbba0000 i3: 0000000000000000
[    0.983197] i4: 00000000005002a4 i5: 0000000000000008 i6: fffff803fb3ab451 i7: 0000000000500620
[    0.983269] I7: <sysfs_read_file+0x80/0x104>
[    0.983299] Caller[0000000000500620]: sysfs_read_file+0x80/0x104
[    0.983368] Caller[00000000004bedc0]: vfs_read+0x78/0x10c
[    0.983556] Caller[00000000004bf114]: sys_read+0x34/0x60
[    0.983622] Caller[0000000000406294]: linux_sparc_syscall32+0x3c/0x40
[    0.983707] Caller[0000000000016d90]: 0x16d98
[    0.983873] Instruction DUMP: 007ddb80  00000000  00000000 <00000000> 00000000  00000000  00000000  00000000  007ddb98 
Quoted 2 lines

This is dev_attr_show() calling attr->show() which points to 'part_attr_group' struct instead of a function. :-)

Ick, that's not good :)

How is it working for anyone else then? sparc64 isn't doing anything "odd" with it's block devices, is it?

Quoted 6 lines

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200
Quoted 6 lines

Driver core: convert block from raw kobjects to core devices

This moves the block devices to /sys/class/block. It will create a flat list of all block devices, with the disks and partitions in one directory. For compatibility /sys/block is created and contains symlinks to the disks.

So I'm guessing if you revert this it works?

Do you have CONFIG_SYSFS_DEPRECATED=Y or N?

thanks,

greg k-h

#2 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:31:17 -0800

Quoted 7 lines · David Miller
Quoted 6 lines

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200
Quoted 8 lines
Quoted 6 lines

Driver core: convert block from raw kobjects to core devices

This moves the block devices to /sys/class/block. It will create a flat list of all block devices, with the disks and partitions in one directory. For compatibility /sys/block is created and contains symlinks to the disks.

So I'm guessing if you revert this it works?

Unfortunately that's not so simple to test because afterwards there are changesets where all the files under block/ got renamed and as such I can't just revert that single patch from the tip.

Quoted 1 line

Do you have CONFIG_SYSFS_DEPRECATED=Y or N?

I have SYSFS_DEPRECATED=y

#3 lore raw

From: David Miller <[email protected]>
Date: Wed, 06 Feb 2008 15:37:14 -0800 (PST)

Quoted 6 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:31:17 -0800

Quoted 1 line

Do you have CONFIG_SYSFS_DEPRECATED=Y or N?

I have SYSFS_DEPRECATED=y

I tested with SYSFS_DEPRECATED disabled, it made no difference.

#4 lore raw
In reply to David Miller
Quoted 17 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:31:17 -0800

Quoted 14 lines · David Miller
Quoted 12 lines

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200

Driver core: convert block from raw kobjects to core devices

This moves the block devices to /sys/class/block. It will create a flat list of all block devices, with the disks and partitions in one directory. For compatibility /sys/block is created and contains symlinks to the disks.

Quoted 5 lines
Quoted 1 line

So I'm guessing if you revert this it works?

Unfortunately that's not so simple to test because afterwards there are changesets where all the files under block/ got renamed and as such I can't just revert that single patch from the tip.

Understood :(

but since you've bisected it down to this, that's a pretty big suggestion that this is the problem.

What block drivers are you using for sparc? Scsi? Or something else? What could make sparc64 different from x86 in regards to block device structure, odd...

thanks,

greg k-h

#5 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:59:02 -0800

Quoted 3 lines

What block drivers are you using for sparc? Scsi? Or something else? What could make sparc64 different from x86 in regards to block device structure, odd...

Only Fusion SAS on this system, therefore scsi.

#6 lore raw
Quoted 8 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:59:02 -0800

Quoted 3 lines

What block drivers are you using for sparc? Scsi? Or something else? What could make sparc64 different from x86 in regards to block device structure, odd...

Only Fusion SAS on this system, therefore scsi.

Can you send me the output of 'tree /sys/block/' on 2.6.24?

Are there a lot of partitions here? Anything different you can think of from x86 that I can have a chance to try to narrow things down with? :)

I don't know if you can boot without udev, but if you could, can you send the 'tree' output of the offending kernel too?

thanks,

greg k-h

#7 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 16:09:59 -0800

Quoted 8 lines · David Miller
Quoted 7 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 15:59:02 -0800

Quoted 3 lines

What block drivers are you using for sparc? Scsi? Or something else? What could make sparc64 different from x86 in regards to block device structure, odd...

Quoted 6 lines
Quoted 1 line

Only Fusion SAS on this system, therefore scsi.

Can you send me the output of 'tree /sys/block/' on 2.6.24?

Are there a lot of partitions here? Anything different you can think of from x86 that I can have a chance to try to narrow things down with? :)

It's a pretty simple partitioning scheme.

[    0.307823]  sda: sda1 sda2 sda3 sda4
Quoted 2 lines

I don't know if you can boot without udev, but if you could, can you send the 'tree' output of the offending kernel too?

It actually is able to continue booting after udevd gets killed off by the OOPS.

I'm going to run a bisect again and get some other info for you...

Here is the tree output with current GIT:

/sys/block/

|-- loop0 -> ../devices/virtual/block/loop0
|-- loop1 -> ../devices/virtual/block/loop1
|-- loop2 -> ../devices/virtual/block/loop2
|-- loop3 -> ../devices/virtual/block/loop3
|-- loop4 -> ../devices/virtual/block/loop4
|-- loop5 -> ../devices/virtual/block/loop5
|-- loop6 -> ../devices/virtual/block/loop6
|-- loop7 -> ../devices/virtual/block/loop7
|-- ram0 -> ../devices/virtual/block/ram0
|-- ram1 -> ../devices/virtual/block/ram1
|-- ram10 -> ../devices/virtual/block/ram10
|-- ram11 -> ../devices/virtual/block/ram11
|-- ram12 -> ../devices/virtual/block/ram12
|-- ram13 -> ../devices/virtual/block/ram13
|-- ram14 -> ../devices/virtual/block/ram14
|-- ram15 -> ../devices/virtual/block/ram15
|-- ram2 -> ../devices/virtual/block/ram2
|-- ram3 -> ../devices/virtual/block/ram3
|-- ram4 -> ../devices/virtual/block/ram4
|-- ram5 -> ../devices/virtual/block/ram5
|-- ram6 -> ../devices/virtual/block/ram6
|-- ram7 -> ../devices/virtual/block/ram7
|-- ram8 -> ../devices/virtual/block/ram8
|-- ram9 -> ../devices/virtual/block/ram9
|-- sda -> ../devices/pci0000:02/0000:02:00.0/0000:03:02.0/0000:0a:00.0/host0/port-0:0/end_device-0:0/target0:0:0/0:0:0:0/block/sda

`-- sr0 -> ../devices/pci0000:02/0000:02:00.0/0000:03:01.0/0000:04:00.0/0000:05:01.0/0000:06:00.0/0000:07:00.2/usb3/3-2/3-2:1.0/host1/target1:0:0/1:0:0:0/block/sr0

26 directories, 0 files

#8 lore raw
In reply to Greg KH

Greg, I'm pretty sure I know what's happening.

For whatever reason we're invoking dev_attr_show() on attribute_group objects.

The reason it probably only crashes on sparc64 is because perhaps at that dev_attr->show offset on x86 there are zero bytes there instead of a pointer, so the NULL check here in dev_attr_show() masks the bug.

The problem with all of this "container_of() this", "container_of() that" is that we lose real type checking. So unless we add magic cookies to verify or other hacks, functions never really know if the container they are being passed really is a subset object of the type they expect.

Can you read the code instead of asking more information from me to try and figure out why the attribute showing paths might be misconfigured for these block device objects after the changeset in question? I can do this, but you're more likely to find the problem quickly than I am.

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks.

#9 lore raw
Quoted 5 lines · David Miller

Greg, I'm pretty sure I know what's happening.

For whatever reason we're invoking dev_attr_show() on attribute_group objects.

Ugh, that makes sense.

Quoted 9 lines

The reason it probably only crashes on sparc64 is because perhaps at that dev_attr->show offset on x86 there are zero bytes there instead of a pointer, so the NULL check here in dev_attr_show() masks the bug.

The problem with all of this "container_of() this", "container_of() that" is that we lose real type checking. So unless we add magic cookies to verify or other hacks, functions never really know if the container they are being passed really is a subset object of the type they expect.

We are supposed to be careful about this, but bad things are known to happen :)

Quoted 5 lines

Can you read the code instead of asking more information from me to try and figure out why the attribute showing paths might be misconfigured for these block device objects after the changeset in question? I can do this, but you're more likely to find the problem quickly than I am.

Yes, I'll look into it tonight if I get this -stable push out in time, or if not, first thing in the morning.

Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks for doing that, I'll let you know when I have a patch to test.

greg k-h

#10 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 21:47:38 -0800

Quoted 5 lines · David Miller
Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks for doing that, I'll let you know when I have a patch to test.

I found the problem, it's the "whole_disk" partition attribute.

Look in fs/partitions/check.c:add_partition() where it tests the ADDPART_FLAG_WHOLEDISK flag.

I think that code block needs to be updated to match the rest of the changes in the offending changeset we are discussing.

#11 lore raw
Quoted 9 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 21:47:38 -0800

Quoted 5 lines · David Miller
Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks for doing that, I'll let you know when I have a patch to test.

Quoted 7 lines

I found the problem, it's the "whole_disk" partition attribute.

Look in fs/partitions/check.c:add_partition() where it tests the ADDPART_FLAG_WHOLEDISK flag.

I think that code block needs to be updated to match the rest of the changes in the offending changeset we are discussing.

Ah, great catch. That code just looks wrong.

I'm guessing that you have a partition that is the whole disk? That would make sense why I and most others haven't seen this yet.

I'll make up a patch...

thanks a lot for still digging into this,

greg k-h

#12 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:38:03 -0800

Quoted 2 lines

I'm guessing that you have a partition that is the whole disk? That would make sense why I and most others haven't seen this yet.

It's an attribute used by Sun disk labels, usually it's the third partition.

#13 lore raw
In reply to David Miller
Quoted 10 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 21:47:38 -0800

Quoted 5 lines · David Miller
Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks for doing that, I'll let you know when I have a patch to test.

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

you are in a maze of kobject pointers, all alike...

#14 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 8 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

you are in a maze of kobject pointers, all alike...

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

It's existence just means that the partition is a "whole disk" partition type.

#15 lore raw
Quoted 10 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 7 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

Quoted 7 lines
Quoted 1 line

you are in a maze of kobject pointers, all alike...

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

It's existence just means that the partition is a "whole disk" partition type.

Can you try this patch to see if it solves the oops, and that the file is still there and works properly?

thanks,

greg k-h

---
 fs/partitions/check.c |    8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
fs/partitions/check.c +3 -5
@@ -353,11 +353,9 @@ void add_partition(struct gendisk *disk,
 	partition_sysfs_add_subdir(p);
 	p->dev.uevent_suppress = 0;
 	if (flags & ADDPART_FLAG_WHOLEDISK) {
-		static struct attribute addpartattr = {
-			.name = "whole_disk",
-			.mode = S_IRUSR | S_IRGRP | S_IROTH,
-		};
-		err = sysfs_create_file(&p->dev.kobj, &addpartattr);
+		static DEVICE_ATTR(whole_disk, S_IRUSR | S_IRGRP | S_IROTH,
+				   NULL, NULL);
+		err = device_create_file(&p->dev, &dev_attr_whole_disk);
 	}
 
 	/* suppress uevent if the disk supresses it */
#16 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 23:00:44 -0800

Quoted 13 lines · David Miller
Quoted 12 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 8 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

you are in a maze of kobject pointers, all alike...

Quoted 8 lines
Quoted 5 lines

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

It's existence just means that the partition is a "whole disk" partition type.

Can you try this patch to see if it solves the oops, and that the file is still there and works properly?

It doesn't crash, but the file returns -EIO instead of zero when read.

#17 lore raw
Quoted 19 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 23:00:44 -0800

Quoted 16 lines · David Miller
Quoted 15 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 8 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

you are in a maze of kobject pointers, all alike...

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

It's existence just means that the partition is a "whole disk" partition type.

Can you try this patch to see if it solves the oops, and that the file is still there and works properly?

It doesn't crash, but the file returns -EIO instead of zero when read.

Picky, picky, picky :)

It's odd that the original one didn't also do that, but I'll fix that up properly, let me try again...

thanks,

greg k-h

#18 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 23:20:30 -0800

Quoted 1 line

It's odd that the original one didn't also do that,

In the old code it would vector through part_sysfs_ops (look at how it works before the changeset), which returns 0 if the attribute lacked a ->show method.

#19 lore raw
Quoted 8 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 23:20:30 -0800

Quoted 1 line

It's odd that the original one didn't also do that,

In the old code it would vector through part_sysfs_ops (look at how it works before the changeset), which returns 0 if the attribute lacked a ->show method.

Ah, yeah, what a fun hack... Makes sense though. The updated patch I send you should do the same thing, just with an additional function needed.

thanks,

greg k-h

#20 lore raw
In reply to David Miller
Quoted 19 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 23:00:44 -0800

Quoted 16 lines · David Miller
Quoted 15 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 8 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

you are in a maze of kobject pointers, all alike...

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

It's existence just means that the partition is a "whole disk" partition type.

Can you try this patch to see if it solves the oops, and that the file is still there and works properly?

It doesn't crash, but the file returns -EIO instead of zero when read.

How about this attempt?

thanks for your patience,

greg "Kay owes me a beer" k-h

---
 fs/partitions/check.c |   17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)
fs/partitions/check.c +10 -7
@@ -319,6 +319,14 @@ void delete_partition(struct gendisk *di
 	put_device(&p->dev);
 }
 
+static ssize_t whole_disk_show(struct device *dev,
+			       struct device_attribute *attr, char *buf)
+{
+	return 0;
+}
+static DEVICE_ATTR(whole_disk, S_IRUSR | S_IRGRP | S_IROTH,
+		   whole_disk_show, NULL);
+
 void add_partition(struct gendisk *disk, int part, sector_t start, sector_t len, int flags)
 {
 	struct hd_struct *p;
@@ -352,13 +360,8 @@ void add_partition(struct gendisk *disk,
 	device_add(&p->dev);
 	partition_sysfs_add_subdir(p);
 	p->dev.uevent_suppress = 0;
-	if (flags & ADDPART_FLAG_WHOLEDISK) {
-		static struct attribute addpartattr = {
-			.name = "whole_disk",
-			.mode = S_IRUSR | S_IRGRP | S_IROTH,
-		};
-		err = sysfs_create_file(&p->dev.kobj, &addpartattr);
-	}
+	if (flags & ADDPART_FLAG_WHOLEDISK)
+		err = device_create_file(&p->dev, &dev_attr_whole_disk);
 
 	/* suppress uevent if the disk supresses it */
 	if (!disk->dev.uevent_suppress)
#23 lore raw
In reply to David Miller
Quoted 10 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:42:10 -0800

Quoted 7 lines · David Miller
Quoted 1 line

I found the problem, it's the "whole_disk" partition attribute.

I don't understand that code at all, on 2.6.24, what does reading that file give you? At first glance, I don't see how that file would spit out anything and not give you the same kind of oops.

Quoted 7 lines
Quoted 1 line

you are in a maze of kobject pointers, all alike...

It's supposed to just exist, and be an empty zero length file. That's why it's given no ->show method pointer.

It's existence just means that the partition is a "whole disk" partition type.

Ah, ok, that's a bit wierd, thanks for explaining it, I'll go make it work...

thanks,

greg k-h

#24 lore raw
In reply to David Miller
Quoted 7 lines · David Miller

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 21:47:38 -0800

Quoted 4 lines · David Miller
Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Quoted 6 lines
Quoted 1 line

Thanks for doing that, I'll let you know when I have a patch to test.

I found the problem, it's the "whole_disk" partition attribute.

Look in fs/partitions/check.c:add_partition() where it tests the ADDPART_FLAG_WHOLEDISK flag.

So, if you just comment out that whole "if (flags & ADDPART_FLAG_WHOLEDISK)" chunk, does the oops go away? I think that is the real solution here as I don't see what this attribute is supposed to be showing.

thanks,

greg k-h

#25 lore raw

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 22:44:07 -0800

Quoted 12 lines · David Miller
Quoted 11 lines

From: Greg KH <[email protected]>
Date: Wed, 6 Feb 2008 21:47:38 -0800

Quoted 5 lines · David Miller
Quoted 2 lines

I redid the bisect to make sure it absolutely was that specific changeset, and it is.

Thanks for doing that, I'll let you know when I have a patch to test.

I found the problem, it's the "whole_disk" partition attribute.

Quoted 7 lines
Quoted 2 lines

Look in fs/partitions/check.c:add_partition() where it tests the ADDPART_FLAG_WHOLEDISK flag.

So, if you just comment out that whole "if (flags & ADDPART_FLAG_WHOLEDISK)" chunk, does the oops go away? I think that is the real solution here as I don't see what this attribute is supposed to be showing.

See my other reply, it's supposed to just exist and be a zero length file. udev looks for it to determine which partition is the "whole disk" partition

#26 lore raw
In reply to Greg KH
Quoted 29 lines · On Wed, 6 Feb 2008 15:31:17 -0800 Greg KH <[email protected]> wrote:
Quoted 28 lines · David Miller

I get the following OOPS from udevd during bootup on sparc64:

[    0.982046]               \|/ ____ \|/
[    0.982054]               "@'/ .. \`@"
[    0.982058]               /_| \__/ |_\
[    0.982063]                  \__U_/
[    0.982482] udevd(1305): Kernel illegal instruction [#1]
[    0.982550] TSTATE: 0000004411001602 TPC: 00000000007ddb80 TNPC: 00000000007ddb84 Y: 00000000    Not tainted
[    0.982647] TPC: <0x7ddb88>
[    0.982684] g0: 0000000000000000 g1: 00000000007ddb80 g2: 0000000000000000 g3: 0000000000000000
[    0.982760] g4: fffff803fb332a00 g5: fffff8000f8a6000 g6: fffff803fb3a8000 g7: fffff803fc893724
[    0.982831] o0: fffff803fc8bc010 o1: 00000000007ddb58 o2: fffff803fbba0000 o3: fffff803fb3abd48
[    0.982907] o4: 0000000000040001 o5: 0000000000000000 sp: fffff803fb3ab391 ret_pc: 00000000005b7524
[    0.982984] RPC: <dev_attr_show+0x24/0x30>
[    0.983016] l0: 0000000000000000 l1: 0000000000000000 l2: fffff803fc9a5308 l3: 000000000005a000
[    0.983074] l4: 0000000000000004 l5: fffff803fb340b00 l6: 0000000000000168 l7: fffff803fbd32700
[    0.983133] i0: fffffffffffffffb i1: 00000000007ddb58 i2: fffff803fbba0000 i3: 0000000000000000
[    0.983197] i4: 00000000005002a4 i5: 0000000000000008 i6: fffff803fb3ab451 i7: 0000000000500620
[    0.983269] I7: <sysfs_read_file+0x80/0x104>
[    0.983299] Caller[0000000000500620]: sysfs_read_file+0x80/0x104
[    0.983368] Caller[00000000004bedc0]: vfs_read+0x78/0x10c
[    0.983556] Caller[00000000004bf114]: sys_read+0x34/0x60
[    0.983622] Caller[0000000000406294]: linux_sparc_syscall32+0x3c/0x40
[    0.983707] Caller[0000000000016d90]: 0x16d98
[    0.983873] Instruction DUMP: 007ddb80  00000000  00000000 <00000000> 00000000  00000000  00000000  00000000  007ddb98 
Quoted 4 lines
Quoted 2 lines

This is dev_attr_show() calling attr->show() which points to 'part_attr_group' struct instead of a function. :-)

Ick, that's not good :)

gregkh-driver-block-device.patch strikes again.

Quoted 9 lines

How is it working for anyone else then? sparc64 isn't doing anything "odd" with it's block devices, is it?

Quoted 6 lines

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200
Quoted 8 lines
Quoted 6 lines

Driver core: convert block from raw kobjects to core devices

This moves the block devices to /sys/class/block. It will create a flat list of all block devices, with the disks and partitions in one directory. For compatibility /sys/block is created and contains symlinks to the disks.

So I'm guessing if you revert this it works?

Going offtopic here...

The patch was committed to mainline last week and it has a git timestamp from eight months ago. When you received the original email from Kay.

But the patch changed in that time period. This doesn't seem right?

#27 lore raw
Quoted 16 lines · Andrew Morton
Quoted 16 lines

How is it working for anyone else then? sparc64 isn't doing anything "odd" with it's block devices, is it?

Quoted 12 lines

I'm pretty sure the following changeset is to blame:

commit edfaa7c36574f1bf09c65ad602412db9da5f96bf

Author: Kay Sievers <[email protected]>
Date:   Mon May 21 22:08:01 2007 +0200

Driver core: convert block from raw kobjects to core devices

This moves the block devices to /sys/class/block. It will create a flat list of all block devices, with the disks and partitions in one directory. For compatibility /sys/block is created and contains symlinks to the disks.

Quoted 8 lines
Quoted 1 line

So I'm guessing if you revert this it works?

Going offtopic here...

The patch was committed to mainline last week and it has a git timestamp from eight months ago. When you received the original email from Kay.

But the patch changed in that time period. This doesn't seem right?

The patch did change over time, but not that much, minor bugfixes for it. I didn't think to update the original date in the quilt file, sorry. It was in -mm for quite a while, so I thought it got a good enough testing period.

I'll try to remember to update the timestamp on patches that get updated, it's a pretty rare thing for my patchflow, shouldn't be hard to remember.

thanks,

greg k-h

#28 lore raw

Oops figures

In reply to David Miller
Quoted 10 lines · David Miller

Subject: partition sysfs OOPS in current GIT

I get the following OOPS from udevd during bootup on sparc64:

[    0.982046]               \|/ ____ \|/
[    0.982054]               "@'/ .. \`@"
[    0.982058]               /_| \__/ |_\
[    0.982063]                  \__U_/

Just what is this creature that one sees on sparc dumps? Looks like a smiley ball with two blinkenlights (symbolizing the sun?) on each side...

#29 lore raw

Re: Oops figures

From: Jan Engelhardt <[email protected]>
Date: Thu, 7 Feb 2008 00:42:52 +0100 (CET)

Quoted 8 lines · David Miller
Quoted 6 lines

Subject: partition sysfs OOPS in current GIT

I get the following OOPS from udevd during bootup on sparc64:

Quoted 8 lines
Quoted 4 lines
[    0.982046]               \|/ ____ \|/
[    0.982054]               "@'/ .. \`@"
[    0.982058]               /_| \__/ |_\
[    0.982063]                  \__U_/

Just what is this creature that one sees on sparc dumps? Looks like a smiley ball with two blinkenlights (symbolizing the sun?) on each side...

It's ascii art I took it from someone's signature 12 years ago, it's meant to be the guy on the cover of some of the editions of the Hitchhikers Guide to the Galaxy by Douglas Adams.

"Don't Panic!" :-)