Re: FS hangups.

cmpilato@collab.net wrote:

>I just reproduced the FS hangs locally using Philip magnifi-script
>(dude, you are to be publically praised for that beaut). The hang in
>the committing svn (the others are stuck updating) occurs here:
>
> #0 0x401d217e in __select () from /lib/libc.so.6
>
> [... Berkeley stuff ...]
>
> #12 0x808c626 in svn_fs__get_txn (txn_skel=0xbffff640, fs=0x80f9140,
> svn_txn=0x8107930 "385", trail=0x8134ee0)
> at subversion/libsvn_fs/txn-table.c:225
> #13 0x808c7d5 in svn_fs__get_txn_ids (root_id_p=0xbffff66c,
> base_root_id_p=0xbffff668, fs=0x80f9140, svn_txn=0x8107930 "385",
> trail=0x8134ee0) at subversion/libsvn_fs/txn-table.c:292
> #14 0x808be8c in txn_body_txn_root (baton=0xbffff6c0, trail=0x8134ee0)
> at subversion/libsvn_fs/tree.c:2758
> #15 0x8088592 in svn_fs__retry_txn (fs=0x80f9140,
> txn_body=0x808be34 <txn_body_txn_root>, baton=0xbffff6c0,
> pool=0x81348c0) at subversion/libsvn_fs/trail.c:127
> #16 0x808bf07 in svn_fs_txn_root (root_p=0xbffff6f4, txn=0x8109818,
> pool=0x81348c0) at subversion/libsvn_fs/tree.c:2778
> #17 0x808a727 in txn_body_merge (baton=0xbffff768, trail=0x81098e0)
> at subversion/libsvn_fs/tree.c:1634
> #18 0x8088592 in svn_fs__retry_txn (fs=0x80f9140,
> txn_body=0x808a5a8 <txn_body_merge>, baton=0xbffff768, pool=0x8109700)
> at subversion/libsvn_fs/trail.c:127
> #19 0x808ab83 in svn_fs_commit_txn (conflict_p=0xbffff7d8,
> new_rev=0xbffff7e0, txn=0x8109818) at subversion/libsvn_fs/tree.c:1826
>
> [... libsvn_repos stuff and so on]
>
>Not that we enter through the public FS API twice, once at
>svn_fs_commit_txn, and once at svn_fs_txn_root.
>
>Now, when I was working on turning off deltification of directory
>entries in the filesystem code last week, I posted to the list my
>patch and a request for help because with my patch in place, Berkeley
>would hang in __select(). I determined that the hangs would disappear
>if I didn't re-enter the public FS API (more specifically, if I didn't
>try to create a Berkeley transaction from within a Berkeley
>transaction), my hangs disappeared. The exact same bug (and solution)
>hit me later last week when I was implementing the recursive tree
>(un-)deltification.
>
>I wonder...could the same thing be happening here?
>

Let me guess ... the transactions created by the two calls to the
top-level API are not related in any way -- i.e., one is not a child of
the other. Therefore they aren't nested. Therefore they can deadlock.

Now why do I get the feeling this can't be right?

I think we need a lower-level API that mimics the public one, except
that every function accepts a trail parameter. Calling any public FS api
function from within the FS implementation should be absolutely prohibited.

-- 
Brane Čibej   <brane_at_xbc.nu>   http://www.xbc.nu/brane/
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org