Re: Write hang on ARM based target

[Ryusuke Konishi <ryusuke-sG5X7nlA6pw@public.gmane.org>]

Hi Bill,
On Mon, 22 Jun 2009 12:58:22 -0600, "Dunphy, Bill" wrote:
> An update.
> 
> I've just begun some testing on a different reference design board
> utilizing the 88F6281.  This particular board has native support in the
> 2.6.30 kernel which allowed me to give the in-tree version of NILFS a
> go.  This board/kernel version combination ran though the testing
> mentioned below without a hitch this weekend (1 million loops).  I've
> since performed a number of massive simultaneous data transfers without
> any errors.  Performance appeared to be much better from a high level as
> well.  So at this point, it appears to me that there is a NILFS
> sensitivity to the 2.6.22.18 kernel and/or a board oddity (even though
> other file systems worked flawlessly).  My near term plan is to move
> forward with this new board/kernel combination.  However, I will keep
> the original board and it's 2.6.22.18 kernel up and available if you
> would like me to try some other changes - you decide.  In the meanwhile,
> I'll start banging away on this platform and report in if I see any
> strange behavior.
> 
> Bill

Sorry for my late reply.

I found an inconsistent state in the value of sequence counter shown
in your log.  I think some sort of synchronization problem is present.
If so, I think we should resolve the problem because it may occur in
any RISC architectures.

Could you test if the attached patch makes a difference?

The patch adds volatile specifiers to sequence counters which may be
shared among different tasks.

Thanks,
Ryusuke Konishi
 
> -----Original Message-----
> From: Dunphy, Bill 
> Sent: Friday, June 19, 2009 9:01 AM
> To: 'Ryusuke Konishi'
> Cc: users-JrjvKiOkagjYtjvyW6yDsg@public.gmane.org; konishi.ryusuke-Zyj7fXuS5i5L9jVzuh4AOg@public.gmane.org
> Subject: RE: [NILFS users] Write hang on ARM based target
> 
> Thanks.  That patch integrated successfully.
> 
> Ran it again with the following result:
> 
> Jun 19 08:38:26 kernel: ========= NILFS SEGMENT INFORMATION ========
> 
> Jun 19 08:38:26 kernel: full segment: segnum=39, start=79872, end=81919
> 
> Jun 19 08:38:26 kernel: partial segment: start=81162, rest=758
> 
> Jun 19 08:38:26 kernel: ------------------ SUMMARY -----------------
> 
> Jun 19 08:38:26 kernel: nfinfo     = 4 (number of files)
> 
> Jun 19 08:38:26 kernel: nblocks    = 14 (number of blocks)
> 
> Jun 19 08:38:26 kernel: sumbytes   = 312 (size of summary in bytes)
> 
> Jun 19 08:38:26 kernel: nsumblk    = 1 (number of summary blocks)
> 
> Jun 19 08:38:26 kernel: flags      = LOGBGN|LOGEND|SR
> 
> Jun 19 08:38:26 kernel: ============================================
> 
> Jun 19 08:38:26 kernel: NILFS(segment)
> nilfs_segctor_update_payload_blocknr: called                           
> Jun 19 08:38:26 kernel: NILFS(segment)
> nilfs_segctor_update_payload_blocknr: done                             
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_file_bmap:
> called                                
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_file_bmap:
> done                                  
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_checkpoint:
> called                               
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_checkpoint:
> done                                 
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_update_segusage:
> called                                  
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_update_segusage:
> done                                    
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_checksums:
> called                                
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_fill_in_checksums:
> done                                  
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segbuf_write: submitting
> summary blocks                          
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_alloc_seg_bio: allocated
> bio (max_vecs=16)                       
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segbuf_write: submitting
> normal blocks (index=1)                 
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_submit_seg_bio: submitting
> bio (start_sector=649296, size=57344,) Jun 19 08:38:26 kernel:
> NILFS(segment) nilfs_segbuf_write: submitted a segment (err=0,
> pseg_start=81162, #req)
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segbuf_wait: called nbio=1
> 
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segbuf_wait: wait completed
> 
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_complete_write:
> completing segment (flags=0x7)           
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_complete_write:
> completed a segment having a super root )
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_do_construct:
> submitted all segments                     
> Jun 19 08:38:26 kernel: NILFS(segment) nilfs_segctor_construct: end
> (stage=9)                                 


diff --git a/fs/segment.c b/fs/segment.c
index 84201ce..c9c28c2 100644
--- a/fs/segment.c
+++ b/fs/segment.c
@@ -2530,7 +2530,7 @@ void nilfs_segctor_clear_segments_to_be_freed(struct nilfs_sc_info *sci)
 
 struct nilfs_segctor_wait_request {
 	wait_queue_t	wq;
-	__u32		seq;
+	volatile __u32	seq;
 	int		err;
 	atomic_t	done;
 };
@@ -2699,7 +2699,7 @@ int nilfs_construct_dsync_segment(struct super_block *sb, struct inode *inode,
 
 struct nilfs_segctor_req {
 	int mode;
-	__u32 seq_accepted;
+	volatile __u32 seq_accepted;
 	int sc_err;  /* construction failure */
 	int sb_err;  /* super block writeback failure */
 };
diff --git a/fs/segment.h b/fs/segment.h
index 44dca64..8533783 100644
--- a/fs/segment.h
+++ b/fs/segment.h
@@ -163,8 +163,8 @@ struct nilfs_sc_info {
 	wait_queue_head_t	sc_wait_daemon;
 	wait_queue_head_t	sc_wait_task;
 
-	__u32			sc_seq_request;
-	__u32			sc_seq_done;
+	volatile __u32		sc_seq_request;
+	volatile __u32		sc_seq_done;
 
 	int			sc_sync;
 	unsigned long		sc_interval;
-- 
1.6.2