https://review.whamcloud.com/#/c/41905/
EX-14494 mdt: check object exists in mdt_close_handle_layouts()
In mdt_close_handle_layouts() the client supplied FID may not identify
an existing object. So check for this before calling lu_object_attr().
Signed-off-by: John L. Hammond <jhammond@whamcloud.com>
Change-Id: Ib1710ca4bf7587e0496b3a37a2afb65f81250455
Oleg Drokin (green@whamcloud.com) merged in patch https://review.whamcloud.com/41905/
Subject: LU-14494 mdt: check object exists in mdt_close_handle_layouts()
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 075bea805efe8a7ef1a3aabd8dd2c166bb52115b
Landed for 2.15
Alex,
I think John has raised this issue elsewhere (though I can't find it), so I will do it again here - should we remove this LASSERT(LOHA_EXISTS) and just return -ENOENT in this case? Jira is littered with tickets where the LOHA_EXISTS assertion is triggered (41 different tickets at last count, I just hit another one in LFSCK over the weekend), and in many cases the solution is just to add extra code to check that the object does not exist before calling lu_object_attr->lu_object_exists(), which isn't really simplifying the code, and causes customer outages. I think this LASSERT has outlived its usefulness if we are adding lines of code just to avoid triggering the LASSERT.
Andreas, I'm not really a defender of that LASSERT(), I didn't introduce that at least 
but I see some logic for that assertion - there is no point to ask for attributes on non-existing object (which has no attributes by definition). OK, say we change lu_object_attr() to return -ENOENT, but then we have to check/fix >60 callers as normally they do like:
if (S_ISDIR(mdd_object_type(obj)))
would that be really different from checking all the places to add if (*_object_exists())... ?
I think at some point John suggested to have a wrapper that checks for existance internally and use that everywhere except few places when we can expect non-existing object as a valid case?