Re: (take 2)[PATCH] JBD: positively dispose the unmapped data buffers in journal_commit_transaction

Hi,

Jan Kara wrote:

quoted

Hi,

Jan Kara wrote:

quoted

 Hi,

quoted

I updated my patch and introduction article for it by reflecting
the comment of Andrew's.

<SNIP>

quoted

quoted

Signed-off-by: Toshiyuki Okajima <redacted>

 I agree with the change. It's true that we can leave some anonymous
pages behind and it's nicer to the MM to release them earlier when we
know they will be never needed again. The patch looks fine to me, you
can add
 Acked-by: Jan Kara [off-list ref]

Thank you for confirming.

  Please keep me CCed (use group reply), thanks. I sometimes don't have
time for reading mailing lists or just skim through them so I can easily
miss replies...

OK, I will keep you CCed next time.

quoted

quoted

 How much have you stressed the patched kernel? I suggest you use
fsxlinux and put some memory pressure to the system...

I have stressed it for 72 or more hours.
Stresser does:
 - allocates/frees big memory(1.7GB) which was almost system
   memory size(2GB) repeatedly.

  OK, I suppose you also wrote something to the memory (otherwise it
won't be really allocated).

I forgot to explain. You are right.
It writes something to each of all pages.

quoted

Confirmation of integrity of patched Filesystem(jbd) does:
 - creates files, and copies 3 files from created each file
   (3 copies run concurrently), and confirms whether there is
   no difference between created files and copied files.
 (20 processes runs these works concurrently and repeatedly.)
Above 2 jobs run concurrently.

  This sounds reasonable. fsxlinux does actually something similar but
it also stresses mmaped accesses and truncate patch. In this case, what
you did should be enough.

								Honza

I wrote only abstract of my long run test at previous mail,
but my test does also mmaped accesses.
So, it seems my test works really the same as fsxlinux...

Thanks,
---
Toshiyuki Okajima