jffs2 copy too long after mounting

jffs2 copy too long after mounting

[hejianet]

hi All
my kernel is 2.6.16
after mounting the nand flash with jffs2 partition, the first copy operation
for a 15M file(from jffs2 partition to jffs2 partition) take 9
minutes,but the
second copy operation takes 1 minutes.
the disk usage(df result) is over 85%.
I wonder is there any patch ported to such old kernel?
Thanks

B.R.
Jia

Re: jffs2 copy too long after mounting

[hejianet]

On 2012-08-02 11:04, hejianet wrote:
> hi All
> my kernel is 2.6.16
> after mounting the nand flash with jffs2 partition, the first copy operation
> for a 15M file(from jffs2 partition to jffs2 partition) take 9
> minutes,but the
> second copy operation takes 1 minutes.
> the disk usage(df result) is over 85%.
> I wonder is there any patch ported to such old kernel?
> Thanks
>
> B.R.
> Jia
>
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
besides, how to search all the archives of linux-mtd?
http://lists.infradead.org/pipermail/linux-mtd/ is a link, but it
couldn't be searched totally by date. I have review all pages
month by month :(

Re: jffs2 copy too long after mounting

[hejianet]

On 2012-08-02 11:04, hejianet wrote:
> hi All
> my kernel is 2.6.16
> after mounting the nand flash with jffs2 partition, the first copy operation
> for a 15M file(from jffs2 partition to jffs2 partition) take 9
> minutes,but the
> second copy operation takes 1 minutes.
> the disk usage(df result) is over 85%.
> I wonder is there any patch ported to such old kernel?
> Thanks
>
> B.R.
> Jia
>
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>
>
Is there anyway to show the summary result of jffs2's fragmentation?
Is there anyway to manually make defragmentation?
Thanks

Jia

Re: jffs2 copy too long after mounting

[Brian Norris]

On Wed, Aug 1, 2012 at 8:16 PM, hejianet <hejianet@linux.vnet.ibm.com> wrote:
> besides, how to search all the archives of linux-mtd?
> http://lists.infradead.org/pipermail/linux-mtd/ is a link, but it
> couldn't be searched totally by date. I have review all pages
> month by month :(

You can do a search of this particular archive using a third-party
search engine, although you may not get results sorted by date. For
example, with Google, just include the following in your search term:
site:lists.infradead.org/pipermail/linux-mtd

Also, there are other archives that track the same mailing list. For
example, gmane has a pretty good interface for search or threaded
browsing:
http://dir.gmane.org/gmane.linux.drivers.mtd
(Try "search the list" or browse with the various options, like "On
the web, using frames and threads")

Hope that helps.

Brian

Re: jffs2 copy too long after mounting

[Brian Norris]

On Wed, Aug 1, 2012 at 10:58 PM, hejianet <hejianet@linux.vnet.ibm.com> wrote:
> Is there anyway to show the summary result of jffs2's fragmentation?
> Is there anyway to manually make defragmentation?

I'm not a JFFS2 expert, but I don't believe fragmentation really is an
issue in quite the same way as in spinning-disk, block storage
filesystems like FAT. Access time to all blocks on an MTD is uniform.

Also, I quite doubt there is a tool that displays the low-level
details of file/block mapping.

Regards,
Brian

Re: jffs2 copy too long after mounting

[hejianet]

[-- Attachment #1: Type: text/plain, Size: 1774 bytes --]

Thanks, Brian
  Here is my latest investigation on the copy_too_long_time issue:
I defined the log level of jffs2 as "noisy",and inserted timestamp for
each jffs2 printk log.
I watched that, there are more lines of such printk compared to good
case.(bad case is first copy, takes 9 minutes, good case is second copy,
takes 1 minute)
1344387878.230899000:<7>Check potential deletion dirent at 0081aad8
1344387878.233899000:<7>Check potential deletion dirent at 00819960
1344387878.236899000:<7>Check potential deletion dirent at 00819388
1344387878.239899000:<7>Check potential deletion dirent at 00818d5c
1344387878.242899000:<7>Check potential deletion dirent at 00818a7c
1344387878.245899000:<7>Check potential deletion dirent at 008141d0
1344387878.248899000:<7>Check potential deletion dirent at 0080e9a0
1344387878.251899000:<7>Check potential deletion dirent at 0080defc
1344387878.254899000:<7>Check potential deletion dirent at 0080dc8c

Thanks for any suggestions or instructions.
kernel version:2.6.16
attached the full logs.

On 2012-08-07 07:02, Brian Norris wrote:
> On Wed, Aug 1, 2012 at 10:58 PM, hejianet <hejianet@linux.vnet.ibm.com> wrote:
>> Is there anyway to show the summary result of jffs2's fragmentation?
>> Is there anyway to manually make defragmentation?
> I'm not a JFFS2 expert, but I don't believe fragmentation really is an
> issue in quite the same way as in spinning-disk, block storage
> filesystems like FAT. Access time to all blocks on an MTD is uniform.
>
> Also, I quite doubt there is a tool that displays the low-level
> details of file/block mapping.
>
> Regards,
> Brian
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>



[-- Attachment #2: copy_too_long.zip --]
[-- Type: application/octet-stream, Size: 572140 bytes --]

Re: jffs2 copy too long after mounting

[Artem Bityutskiy]

[-- Attachment #1: Type: text/plain, Size: 1429 bytes --]

On Thu, 2012-08-02 at 11:04 +0800, hejianet wrote:
> hi All
> my kernel is 2.6.16
> after mounting the nand flash with jffs2 partition, the first copy operation
> for a 15M file(from jffs2 partition to jffs2 partition) take 9
> minutes,but the
> second copy operation takes 1 minutes.
> the disk usage(df result) is over 85%.
> I wonder is there any patch ported to such old kernel?

His is because how JFFS2 is designed. When the mount operation is
finished, JFFS2 does not really finish mounting. There is the background
process which keeps working - it basically reads all files in the system
and checks the CRC, which is needed to find out which nodes are new and
which are obsolete.

So what happens is that you mount JFFS2, the background thread keeps
working (see 'top' for example).

While the background thread works, you can read files. But you cannot
write - writes will be blocked. Well, it is a bit more complex - you can
write a little, but from some point your writes will be blocked until
the background thread finishes. Then the write will continue and finish.

In you case, you mount, start writing your 15MiB file, write gets
blocked, waits for the background thread, then continues write. The
background thread takes ages to finish on big partitions.

The second write is fast because you do not wait for the background
thread any longer.

-- 
Best Regards,
Artem Bityutskiy

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: jffs2 copy too long after mounting

[hejianet]

Hi Artem, thanks for your reply
Now I can reproduce it after 2-3 days stress tests:
1.just do lots of small writing(4k bytes each time) radomly in one of 1000
 thousand file
2.remove the 1000 files,now the jffs2 fs is *very* dirty.
3.I dynamically increase the threshold nr_free_blocks to trigger the gc
in jffs2_thread_should_wake():
     if (c->nr_free_blocks + c->nr_erasing_blocks < c->resv_blocks_gctrigger &&
            (dirty > c->nospc_dirty_size))
        ret = 1;
4. I watched after a while, there were tens of  thousands of lines (I enable the
jffs2 log level to be 1)
Check potential deletion dirent at 008ddd24
Check potential deletion dirent at 008ddcf4
Check potential deletion dirent at 008ddcc4
Check potential deletion dirent at 008ddc68
Check potential deletion dirent at 008dca34
Check potential deletion dirent at 008dca04
Check potential deletion dirent at 0088d77c
Check potential deletion dirent at 0088d74c
Check potential deletion dirent at 0088d71c
Check potential deletion dirent at 0088d6ec
Check potential deletion dirent at 0088d6bc
Check potential deletion dirent at 0088c648
Check potential deletion dirent at 0088c618
Check potential deletion dirent at 0088c2a8
Check potential deletion dirent at 0088c278
Check potential deletion dirent at 0088c248
Check potential deletion dirent at 0088c218
Check potential deletion dirent at 0088c1e8
Check potential deletion dirent at 00886b50
Check potential deletion dirent at 00885388
Check potential deletion dirent at 00885358
Check potential deletion dirent at 008852a8
Check potential deletion dirent at 00885278
5. I did find that the free blocks number of jffs2 is very slowly increased

So could you give me any ideas, it is normal or abnormal?
Thanks a lot
On 2012-08-24 15:53, Artem Bityutskiy wrote:
> On Thu, 2012-08-02 at 11:04 +0800, hejianet wrote:
>> hi All
>> my kernel is 2.6.16
>> after mounting the nand flash with jffs2 partition, the first copy operation
>> for a 15M file(from jffs2 partition to jffs2 partition) take 9
>> minutes,but the
>> second copy operation takes 1 minutes.
>> the disk usage(df result) is over 85%.
>> I wonder is there any patch ported to such old kernel?
> His is because how JFFS2 is designed. When the mount operation is
> finished, JFFS2 does not really finish mounting. There is the background
> process which keeps working - it basically reads all files in the system
> and checks the CRC, which is needed to find out which nodes are new and
> which are obsolete.
>
> So what happens is that you mount JFFS2, the background thread keeps
> working (see 'top' for example).
>
> While the background thread works, you can read files. But you cannot
> write - writes will be blocked. Well, it is a bit more complex - you can
> write a little, but from some point your writes will be blocked until
> the background thread finishes. Then the write will continue and finish.
>
> In you case, you mount, start writing your 15MiB file, write gets
> blocked, waits for the background thread, then continues write. The
> background thread takes ages to finish on big partitions.
>
> The second write is fast because you do not wait for the background
> thread any longer.
>

Re: jffs2 copy too long after mounting

[hejianet]

Hi all
  I think I find the reason why there are so many similar logs("Check
potential deletion dirent").
  It seems in jffs2_garbage_collect_deletion_dirent(), jffs2 tries to 
find "whether this deletion dirent is still needed to actively delete
a 'real' dirent with the same name that's still somewhere else on the 
flash."
  I wonder why there is such a necessary handling procedure? There must
be a newest valid direntry. Why we need to remove the older one and 
"write out a new deletion dirent to replace it"? What if jffs2 just 
did not search the older deletion dirent allover the flash?
Thanks.
B.R.
Jia


On 2012-10-23 15:13, hejianet wrote:
> Hi Artem, thanks for your reply
> Now I can reproduce it after 2-3 days stress tests:
> 1.just do lots of small writing(4k bytes each time) radomly in one of 1000
>  thousand file
> 2.remove the 1000 files,now the jffs2 fs is *very* dirty.
> 3.I dynamically increase the threshold nr_free_blocks to trigger the gc
> in jffs2_thread_should_wake():
>      if (c->nr_free_blocks + c->nr_erasing_blocks < c->resv_blocks_gctrigger &&
>             (dirty > c->nospc_dirty_size))
>         ret = 1;
> 4. I watched after a while, there were tens of  thousands of lines (I enable the
> jffs2 log level to be 1)
> Check potential deletion dirent at 008ddd24
> Check potential deletion dirent at 008ddcf4
> Check potential deletion dirent at 008ddcc4
> Check potential deletion dirent at 008ddc68
> Check potential deletion dirent at 008dca34
> Check potential deletion dirent at 008dca04
> Check potential deletion dirent at 0088d77c
> Check potential deletion dirent at 0088d74c
> Check potential deletion dirent at 0088d71c
> Check potential deletion dirent at 0088d6ec
> Check potential deletion dirent at 0088d6bc
> Check potential deletion dirent at 0088c648
> Check potential deletion dirent at 0088c618
> Check potential deletion dirent at 0088c2a8
> Check potential deletion dirent at 0088c278
> Check potential deletion dirent at 0088c248
> Check potential deletion dirent at 0088c218
> Check potential deletion dirent at 0088c1e8
> Check potential deletion dirent at 00886b50
> Check potential deletion dirent at 00885388
> Check potential deletion dirent at 00885358
> Check potential deletion dirent at 008852a8
> Check potential deletion dirent at 00885278
> 5. I did find that the free blocks number of jffs2 is very slowly increased
>
> So could you give me any ideas, it is normal or abnormal?
> Thanks a lot
> On 2012-08-24 15:53, Artem Bityutskiy wrote:
>> On Thu, 2012-08-02 at 11:04 +0800, hejianet wrote:
>>> hi All
>>> my kernel is 2.6.16
>>> after mounting the nand flash with jffs2 partition, the first copy operation
>>> for a 15M file(from jffs2 partition to jffs2 partition) take 9
>>> minutes,but the
>>> second copy operation takes 1 minutes.
>>> the disk usage(df result) is over 85%.
>>> I wonder is there any patch ported to such old kernel?
>> His is because how JFFS2 is designed. When the mount operation is
>> finished, JFFS2 does not really finish mounting. There is the background
>> process which keeps working - it basically reads all files in the system
>> and checks the CRC, which is needed to find out which nodes are new and
>> which are obsolete.
>>
>> So what happens is that you mount JFFS2, the background thread keeps
>> working (see 'top' for example).
>>
>> While the background thread works, you can read files. But you cannot
>> write - writes will be blocked. Well, it is a bit more complex - you can
>> write a little, but from some point your writes will be blocked until
>> the background thread finishes. Then the write will continue and finish.
>>
>> In you case, you mount, start writing your 15MiB file, write gets
>> blocked, waits for the background thread, then continues write. The
>> background thread takes ages to finish on big partitions.
>>
>> The second write is fast because you do not wait for the background
>> thread any longer.
>>
>
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>
>

Re: jffs2 copy too long after mounting

[hejianet]

Sorry,maybe my question is long and bothering:(
I try to make it more easily and clearly
in jffs2_garbage_collect_deletion_dirent(),
what does it mean, why in nand, we can't actually mark nodes obsolete
pernamently
    /* On a medium where we can't actually mark nodes obsolete
       pernamently, such as NAND flash, ..... */

On 2012-10-25 09:41, hejianet wrote:
> Hi all
>   I think I find the reason why there are so many similar logs("Check
> potential deletion dirent").
>   It seems in jffs2_garbage_collect_deletion_dirent(), jffs2 tries to 
> find "whether this deletion dirent is still needed to actively delete
> a 'real' dirent with the same name that's still somewhere else on the 
> flash."
>   I wonder why there is such a necessary handling procedure? There must
> be a newest valid direntry. Why we need to remove the older one and 
> "write out a new deletion dirent to replace it"? What if jffs2 just 
> did not search the older deletion dirent allover the flash?
> Thanks.
> B.R.
> Jia
>
>
> On 2012-10-23 15:13, hejianet wrote:
>> Hi Artem, thanks for your reply
>> Now I can reproduce it after 2-3 days stress tests:
>> 1.just do lots of small writing(4k bytes each time) radomly in one of 1000
>>  thousand file
>> 2.remove the 1000 files,now the jffs2 fs is *very* dirty.
>> 3.I dynamically increase the threshold nr_free_blocks to trigger the gc
>> in jffs2_thread_should_wake():
>>      if (c->nr_free_blocks + c->nr_erasing_blocks < c->resv_blocks_gctrigger &&
>>             (dirty > c->nospc_dirty_size))
>>         ret = 1;
>> 4. I watched after a while, there were tens of  thousands of lines (I enable the
>> jffs2 log level to be 1)
>> Check potential deletion dirent at 008ddd24
>> Check potential deletion dirent at 008ddcf4
>> Check potential deletion dirent at 008ddcc4
>> Check potential deletion dirent at 008ddc68
>> Check potential deletion dirent at 008dca34
>> Check potential deletion dirent at 008dca04
>> Check potential deletion dirent at 0088d77c
>> Check potential deletion dirent at 0088d74c
>> Check potential deletion dirent at 0088d71c
>> Check potential deletion dirent at 0088d6ec
>> Check potential deletion dirent at 0088d6bc
>> Check potential deletion dirent at 0088c648
>> Check potential deletion dirent at 0088c618
>> Check potential deletion dirent at 0088c2a8
>> Check potential deletion dirent at 0088c278
>> Check potential deletion dirent at 0088c248
>> Check potential deletion dirent at 0088c218
>> Check potential deletion dirent at 0088c1e8
>> Check potential deletion dirent at 00886b50
>> Check potential deletion dirent at 00885388
>> Check potential deletion dirent at 00885358
>> Check potential deletion dirent at 008852a8
>> Check potential deletion dirent at 00885278
>> 5. I did find that the free blocks number of jffs2 is very slowly increased
>>
>> So could you give me any ideas, it is normal or abnormal?
>> Thanks a lot
>> On 2012-08-24 15:53, Artem Bityutskiy wrote:
>>> On Thu, 2012-08-02 at 11:04 +0800, hejianet wrote:
>>>> hi All
>>>> my kernel is 2.6.16
>>>> after mounting the nand flash with jffs2 partition, the first copy operation
>>>> for a 15M file(from jffs2 partition to jffs2 partition) take 9
>>>> minutes,but the
>>>> second copy operation takes 1 minutes.
>>>> the disk usage(df result) is over 85%.
>>>> I wonder is there any patch ported to such old kernel?
>>> His is because how JFFS2 is designed. When the mount operation is
>>> finished, JFFS2 does not really finish mounting. There is the background
>>> process which keeps working - it basically reads all files in the system
>>> and checks the CRC, which is needed to find out which nodes are new and
>>> which are obsolete.
>>>
>>> So what happens is that you mount JFFS2, the background thread keeps
>>> working (see 'top' for example).
>>>
>>> While the background thread works, you can read files. But you cannot
>>> write - writes will be blocked. Well, it is a bit more complex - you can
>>> write a little, but from some point your writes will be blocked until
>>> the background thread finishes. Then the write will continue and finish.
>>>
>>> In you case, you mount, start writing your 15MiB file, write gets
>>> blocked, waits for the background thread, then continues write. The
>>> background thread takes ages to finish on big partitions.
>>>
>>> The second write is fast because you do not wait for the background
>>> thread any longer.
>>>
>>
>> ______________________________________________________
>> Linux MTD discussion mailing list
>> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>>
>>
>>
>
>
>
>
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
>
>

Re: jffs2 copy too long after mounting

[Artem Bityutskiy]

[-- Attachment #1: Type: text/plain, Size: 327 bytes --]

On Tue, 2012-10-23 at 15:13 +0800, hejianet wrote:
> Hi Artem, thanks for your reply
> Now I can reproduce it after 2-3 days stress tests:

You should use perf or oprofile and find out where the kernel actually
spends time. Logs is not the best source of information in this case.

-- 
Best Regards,
Artem Bityutskiy

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: jffs2 copy too long after mounting

[Artem Bityutskiy]

[-- Attachment #1: Type: text/plain, Size: 665 bytes --]

On Fri, 2012-10-26 at 09:45 +0800, hejianet wrote:
> Sorry,maybe my question is long and bothering:(
> I try to make it more easily and clearly
> in jffs2_garbage_collect_deletion_dirent(),
> what does it mean, why in nand, we can't actually mark nodes obsolete
> pernamently
>     /* On a medium where we can't actually mark nodes obsolete
>        pernamently, such as NAND flash, ..... */

Because in NAND you cannot change the contents of a NAND page which has
already been written, because you have ECC for the data in the OOB. So
you cannot just mark a node as obsolete "in-place". In NOR you can do
this.

-- 
Best Regards,
Artem Bityutskiy

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]