blktrace and xfs_db block offset mismatch

blktrace and xfs_db block offset mismatch

[Sushma R]

[-- Attachment #1.1: Type: text/plain, Size: 678 bytes --]

Hi,

I ran blktrace while reading a file on XFS partition and here's it's output.

8,17   3        1    22.065944956  5489  D   R 20973896 + 8 [cat]

Using xfs_db, I dumped the inode for the same file and it shows the
following as its data extents.

u.bmx[0] = [startoff,startblock,blockcount,extentflag] 0:[0,2621481,1,0]

I understand that blktrace reports in 512-byte sectors while xfs_db in
filesystem block size (which is the default 4KB on my setup).
Converting blktrace offset to fs block = 2621737 which is close to the data
extent startblock (2621481), but still off by 256 fs blocks.

Could someone please help understand the reason for this mismatch?

Thanks,
Sushma

[-- Attachment #1.2: Type: text/html, Size: 915 bytes --]

[-- Attachment #2: Type: text/plain, Size: 121 bytes --]

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: blktrace and xfs_db block offset mismatch

[Brian Foster]

On Sat, Sep 27, 2014 at 12:32:21AM -0700, Sushma R wrote:
> Hi,
> 
> I ran blktrace while reading a file on XFS partition and here's it's output.
> 
> 8,17   3        1    22.065944956  5489  D   R 20973896 + 8 [cat]
> 
> Using xfs_db, I dumped the inode for the same file and it shows the
> following as its data extents.
> 
> u.bmx[0] = [startoff,startblock,blockcount,extentflag] 0:[0,2621481,1,0]
> 
> I understand that blktrace reports in 512-byte sectors while xfs_db in
> filesystem block size (which is the default 4KB on my setup).
> Converting blktrace offset to fs block = 2621737 which is close to the data
> extent startblock (2621481), but still off by 256 fs blocks.
> 
> Could someone please help understand the reason for this mismatch?
> 

I believe you are comparing a raw read of a device to an internal XFS
fsb, the latter of which is encoded to account for the internal
structure of XFS (e.g., allocation groups). You might want to take a
look at xfs_bmap on the file to see the actual disk blocks.

The fsb to daddr conversion depends on the geometry of the fs, but you
can get the various values with xfs_db as well. E.g.:

convert fsb 2621481 agno	# ag number
convert fsb 2621481 agbno	# relative ag block nr
convert fsb 2621481 daddr	# disk address

Also, I'm not familiar enough with blktrace output off the top of my
head so perhaps not an issue, but it's not clear to me whether the
output above could have been converted/adjusted to a raw address based
on a device partition.

You might also want to verify you're looking at the actual file read vs.
a directory/inode lookup or block mapping lookup or something else. I
see you're running cat, a direct I/O read via dd might ensure you aren't
just pulling file data from cache.

Brian

> Thanks,
> Sushma

> _______________________________________________
> xfs mailing list
> xfs@oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: blktrace and xfs_db block offset mismatch

[Sushma R]

[-- Attachment #1.1: Type: text/plain, Size: 2944 bytes --]

Thanks Brian.

I'm trying to understand how the internal XFS structures affect this
mapping (http://oss.sgi.com/projects/xfs/papers/xfs_filesystem_structure.pdf
)

xfs_bmap on the file matches that reported by extents in the file inode
using xfs_db.

Other conversions under xfs_db
xfs_db> convert fsb 2621481 agno
0x0 (0)
xfs_db> convert fsb 2621481 agbno
0x280029 (2621481)
xfs_db> convert fsb 2621481 daddr
0x1400148 (20971848)

I tried dd command with directio option and still blktrace reports the same
raw block number as earlier.
8,17   3        1     0.000000000  6999  D   R 20973896 + 1 [dd]

I'm using xfs_db with the partition i.e. "xfs_db -r /dev/sdb1"
However, if I use it on the whole disk "xfs_db -r /dev/sdb", it gives the
following error
xfs_db: cannot init perag data (117)

Does this suggest something?

Thanks,
Sushma

On Sat, Sep 27, 2014 at 7:38 AM, Brian Foster <bfoster@redhat.com> wrote:

> On Sat, Sep 27, 2014 at 12:32:21AM -0700, Sushma R wrote:
> > Hi,
> >
> > I ran blktrace while reading a file on XFS partition and here's it's
> output.
> >
> > 8,17   3        1    22.065944956  5489  D   R 20973896 + 8 [cat]
> >
> > Using xfs_db, I dumped the inode for the same file and it shows the
> > following as its data extents.
> >
> > u.bmx[0] = [startoff,startblock,blockcount,extentflag] 0:[0,2621481,1,0]
> >
> > I understand that blktrace reports in 512-byte sectors while xfs_db in
> > filesystem block size (which is the default 4KB on my setup).
> > Converting blktrace offset to fs block = 2621737 which is close to the
> data
> > extent startblock (2621481), but still off by 256 fs blocks.
> >
> > Could someone please help understand the reason for this mismatch?
> >
>
> I believe you are comparing a raw read of a device to an internal XFS
> fsb, the latter of which is encoded to account for the internal
> structure of XFS (e.g., allocation groups). You might want to take a
> look at xfs_bmap on the file to see the actual disk blocks.
>
> The fsb to daddr conversion depends on the geometry of the fs, but you
> can get the various values with xfs_db as well. E.g.:
>
> convert fsb 2621481 agno        # ag number
> convert fsb 2621481 agbno       # relative ag block nr
> convert fsb 2621481 daddr       # disk address
>
> Also, I'm not familiar enough with blktrace output off the top of my
> head so perhaps not an issue, but it's not clear to me whether the
> output above could have been converted/adjusted to a raw address based
> on a device partition.
>
> You might also want to verify you're looking at the actual file read vs.
> a directory/inode lookup or block mapping lookup or something else. I
> see you're running cat, a direct I/O read via dd might ensure you aren't
> just pulling file data from cache.
>
> Brian
>
> > Thanks,
> > Sushma
>
> > _______________________________________________
> > xfs mailing list
> > xfs@oss.sgi.com
> > http://oss.sgi.com/mailman/listinfo/xfs
>
>

[-- Attachment #1.2: Type: text/html, Size: 4810 bytes --]

[-- Attachment #2: Type: text/plain, Size: 121 bytes --]

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: blktrace and xfs_db block offset mismatch

[Sushma R]

[-- Attachment #1.1: Type: text/plain, Size: 3543 bytes --]

Please ignore my last comment. It doesn't make sense since file system is
mounted on a disk partition.

>>Also, I'm not familiar enough with blktrace output off the top of my
>>head so perhaps not an issue, but it's not clear to me whether the
>>output above could have been converted/adjusted to a raw address based
>>on a device partition.

Yes, it appears that blktrace adds device partition offset.

Thanks!!


On Sat, Sep 27, 2014 at 8:40 AM, Sushma R <gsushma@gmail.com> wrote:

> Thanks Brian.
>
> I'm trying to understand how the internal XFS structures affect this
> mapping (
> http://oss.sgi.com/projects/xfs/papers/xfs_filesystem_structure.pdf)
>
> xfs_bmap on the file matches that reported by extents in the file inode
> using xfs_db.
>
> Other conversions under xfs_db
> xfs_db> convert fsb 2621481 agno
> 0x0 (0)
> xfs_db> convert fsb 2621481 agbno
> 0x280029 (2621481)
> xfs_db> convert fsb 2621481 daddr
> 0x1400148 (20971848)
>
> I tried dd command with directio option and still blktrace reports the
> same raw block number as earlier.
> 8,17   3        1     0.000000000  6999  D   R 20973896 + 1 [dd]
>
> I'm using xfs_db with the partition i.e. "xfs_db -r /dev/sdb1"
> However, if I use it on the whole disk "xfs_db -r /dev/sdb", it gives the
> following error
> xfs_db: cannot init perag data (117)
>
> Does this suggest something?
>
> Thanks,
> Sushma
>
> On Sat, Sep 27, 2014 at 7:38 AM, Brian Foster <bfoster@redhat.com> wrote:
>
>> On Sat, Sep 27, 2014 at 12:32:21AM -0700, Sushma R wrote:
>> > Hi,
>> >
>> > I ran blktrace while reading a file on XFS partition and here's it's
>> output.
>> >
>> > 8,17   3        1    22.065944956  5489  D   R 20973896 + 8 [cat]
>> >
>> > Using xfs_db, I dumped the inode for the same file and it shows the
>> > following as its data extents.
>> >
>> > u.bmx[0] = [startoff,startblock,blockcount,extentflag] 0:[0,2621481,1,0]
>> >
>> > I understand that blktrace reports in 512-byte sectors while xfs_db in
>> > filesystem block size (which is the default 4KB on my setup).
>> > Converting blktrace offset to fs block = 2621737 which is close to the
>> data
>> > extent startblock (2621481), but still off by 256 fs blocks.
>> >
>> > Could someone please help understand the reason for this mismatch?
>> >
>>
>> I believe you are comparing a raw read of a device to an internal XFS
>> fsb, the latter of which is encoded to account for the internal
>> structure of XFS (e.g., allocation groups). You might want to take a
>> look at xfs_bmap on the file to see the actual disk blocks.
>>
>> The fsb to daddr conversion depends on the geometry of the fs, but you
>> can get the various values with xfs_db as well. E.g.:
>>
>> convert fsb 2621481 agno        # ag number
>> convert fsb 2621481 agbno       # relative ag block nr
>> convert fsb 2621481 daddr       # disk address
>>
>> Also, I'm not familiar enough with blktrace output off the top of my
>> head so perhaps not an issue, but it's not clear to me whether the
>> output above could have been converted/adjusted to a raw address based
>> on a device partition.
>>
>> You might also want to verify you're looking at the actual file read vs.
>> a directory/inode lookup or block mapping lookup or something else. I
>> see you're running cat, a direct I/O read via dd might ensure you aren't
>> just pulling file data from cache.
>>
>> Brian
>>
>> > Thanks,
>> > Sushma
>>
>> > _______________________________________________
>> > xfs mailing list
>> > xfs@oss.sgi.com
>> > http://oss.sgi.com/mailman/listinfo/xfs
>>
>>
>

[-- Attachment #1.2: Type: text/html, Size: 6584 bytes --]

[-- Attachment #2: Type: text/plain, Size: 121 bytes --]

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs