Hi Ted,


On 05/25/2011 12:42 AM, Ted Ts'o wrote:

On Wed, May 11, 2011 at 10:10:41AM +0300, Surbhi Palande wrote:

quoted

While the fsstress background writes are busy dirtying the page cache, if a
fsfreeze happens then the background writes should stall. A sync should then
not have any data to sync to the FS. If it does have any data to sync then
sync will cause a deadlock by holding the s_umount write semaphore and waiting
in the wait queue for the FS to thaw, whereas the F.S can never thaw without
getting the s_umount write semaphore.

Signed-off-by: Surbhi Palande<redacted>

Hi Surbhi,

Have you tried out Jan Kara's patches?

[1/3] fs: Create __block_page_mkwrite() helper passing error values back
[2/3] vfs: Block mmapped writes while the fs is frozen
[3/3] ext4: Rewrite ext4_page_mkwrite() to return locked page

Yes! We have tried these patches and we still see the same 
deadlock/hang. The following is the reason for it:


// lets assume the inode is clean and so are its pages.
P1: process that tries mmap write
t1) __do_fault()
   t2) ext4_page_mkwrite()
     t3) block_page_mkwrite()
       t4) vfs_check_frozen()
// filesystem is not frozen so control falls through.
       t5) __block_page_mkwrite()
         t6) set_page_dirty()
           t7) __set_page_dirty()
	    t8) radix_tree_tag_set(PAGECACHE_TAG_DIRTY)
// page is dirtied, but inode is yet clean.
---------------------- Pre-empted-----------------
P2: freeze process

t9) freeze_super()
   t10) sync_filesystem()
  // page cache now clean! no inode is dirty.
// however we have a dirty page belonging to a clean inode.
----------------------Freeze process finishes, filesystem frozen!----


P1: process that tries mmap write gets control.
t11) __set_page_dirty() // gets control back
     t12) __mark_inode_dirty()v
    // inode is now dirty and it has a dirty page.
    // though in reality there is no write which has occured.
t13)   if (inode->i_sb->s_frozen != SB_UNFROZEN)
     // __block_page_mkwrite() gets control back
t14) unlock_page()
t15) __block_page_mkwrite() returns -EAGAIN
t16) block_page_mkwrite() returns VM_FAULT_RETRY

---------------------------
// now we see the original deadlock reported.
P3: sync a filesystem
t17) down_read(s_umount)
  t18) sync_filesystem()
   t19) sb->s_op->sync_fs() // =ext4_sync_fs()
    t20) vfs_check_frozen() // now blocks for thaw.
// so thaw cannot happen because sync process sleeps with s_umount!

This deadlock can occur whenever the freeze happens after the 
vfs_check_frozen() but before the __mark_inode_dirty().

We see blocked sync processes every time we do the following:

1) executing iozone on multipath and
2) I modified the script that Toshiyuki sent, attaching it here. This 
script reproduces the bug faster when executed with iozone.
(Note, that since this is a race, this script _may not_ always produce 
it on its own)


I also found one more missing piece in the "Add support to freeze and 
unfreeze journal":
1) Call jdb2_journal_thaw() from ext4_unfreeze() to restart the 
transactions.

I shall send a patch for the same as a reply to this email again.

Thanks!

Warm Regards,
Surbhi.












P3: sync

Do these patches fix the problem you've been trying to fix with your
patches?  I believe they should, but I would appreciate confirmation
that with these patches, you're no longer able to reproduce the
problem you've been concerned about.

Thanks, regards,

						- Ted