Re: [PATCH 25/30] mds: bring back old style backtrace handling

["Yan, Zheng" <zheng.z.yan@intel.com>]

On 05/24/2013 06:58 AM, Sage Weil wrote:
> On Thu, 23 May 2013, Yan, Zheng wrote:
> [snip]
>> +
>> +void CInode::store_backtrace(Context *fin)
>> +{
>> +  dout(10) << "store_backtrace on " << *this << dendl;
>> +  assert(is_dirty_parent());
>> +
>> +  auth_pin(this);
>> +
>> +  int64_t pool;
>> +  if (is_dir())
>> +    pool = mdcache->mds->mdsmap->get_metadata_pool();
>> +  else
>> +    pool = inode.layout.fl_pg_pool;
>> +
>> +  inode_backtrace_t bt;
>> +  build_backtrace(pool, &bt);
>> +  bufferlist bl;
>> +  ::encode(bt, bl);
>> +
>> +  // write it.
>> +  SnapContext snapc;
>> +  object_t oid = get_object_name(ino(), frag_t(), "");
>> +  object_locator_t oloc(pool);
>> +  Context *fin2 = new C_Inode_StoredBacktrace(this, inode.backtrace_version, fin);
>> +
>> +  if (!state_test(STATE_DIRTYPOOL)) {
>> +    mdcache->mds->objecter->setxattr(oid, oloc, "parent", snapc, bl,
>> +				     ceph_clock_now(g_ceph_context),
>> +				     0, NULL, fin2);
>> +    return;
>> +  }
>> +
>> +  C_GatherBuilder gather(g_ceph_context, fin2);
>> +  mdcache->mds->objecter->setxattr(oid, oloc, "parent", snapc, bl,
>> +				   ceph_clock_now(g_ceph_context),
>> +				   0, NULL, gather.new_sub());
>> +  for (set<int64_t>::iterator p = bt.old_pools.begin();
>> +       p != bt.old_pools.end();
>> +       ++p) {
>> +    object_locator_t oloc2(*p);
>> +    mdcache->mds->objecter->setxattr(oid, oloc2, "parent", snapc, bl,
>> +				     ceph_clock_now(g_ceph_context),
>> +				     0, NULL, gather.new_sub());
>> +  }
> 
> I think for both of theese operations we need an ObjectWriteOperation that 
> does a touch() and then tsetxattr to ensure the object actually exists.
> 
will add it

> Also, if one mds has a backtrace write in flight, exports teh inode, and 
> the second mds needs to update it, we need to make sure they don't race 
> and overwrite a newer trace with an older one.  That could be done with a 
> parent_version xattr with the backttrace_version in it and a generic rados 
> cmpxattr guard, I believe.  Even then we may race with an unlink, but that 
> may be something we just tolerate...
> 
my code calls auth_pin() in CInode::store_backtrace(). I think it also avoid
the race.

Regards
Yan, Zheng