Re: (ITS#7853) Improve safety of mdb_env_close

[hyc@symas.com]

dw@botanicus.net wrote:
> Full_Name: David Wilson
> Version: LMDB 0.9.11
> OS:
> URL: https://raw.githubusercontent.com/dw/py-lmdb/master/misc/readers_mrb_env.patch
> Submission from: (NULL) (178.238.153.20)
>
>
> Currently if a user (wilfully or accidentally, say, through composition of
> third party libs) opens the same LMDB environment twice within a process, and
> maintains active read transactions at the time one mdb_env_close() is called,
> all reader slots will be deallocated in all environments due to the logic
> around line 4253 that unconditionally clears reader slots based on PID.
>
> I'd like to avoid this in py-lmdb, and seems there are a few alternatives to
> fix it:
>
>
> 1) Tell users not to do that (doc fix, nobody reads docs)

This is already in the doc - don't open the same env twice within one process.

> 2) Externally maintain a set of (st_dev, st_ino) keys representing already
> opened environments, and check any attempt to open an environment against this
> list, failing if data.mdb's key already appears.

That would require global library state, which is definitely unwanted.

> 3) Disable the mdb_env_close() cleanup loop when MDB_NOTLS is enabled. Since no
> reader slot will ever exist in TLS in this mode, the existing mdb_env_close()
> doc "All transactions, databases, and cursors must already be closed before
> calling this function." ensures no readers exist at close time, and the loop
> need not run.

Could do this, but since it's an incomplete solution, not sure there's any point.

> 4) Modify lock.mdb to include MDB_env* address within the process, and update
> mdb_env_close() to invalidate only readers associated with the environment
> being closed. I dislike using the MDB_env* as an opaque publicly visible
> cookie, but perhaps storing this field might be reused for other things in
> future.

Hmm. I guess this could work. I think storing the MDB_env* is harmless since 
the pointer is meaningless to any other process.

> Option 3 lets the user wilfully mix py-lmdb Environment objects within the
> process, since py-lmdb always enables NOTLS, but it does not fix the case where
> the user is integrating a Python application with a C library with some opaque
> interface, and both the application and the library independently have need to
> access the environment.

Allowing multiple independent libraries to open the same environment would 
consume redundant chunks of address space. For large enough DBs this may 
simply fail.

Unfortunately the semantics of fcntl locks prevents us from even detecting 
that this problem has occurred.

> This kind of scenario regularly crops up in huge software projects, the user
> may not even be aware the library uses LMDB internally. In that case, the user
> is again exposed to having their reader slots silently become deallocated
> through no fault of their own.
>
>
> Option 4 is the least aesthetic, but has all the benefits of option 3 in
> addition to preventing the "accidental integration" scenario. I've attached a
> patch for it (against 0.9.11), although I completely understand if it does not
> get applied. :)
>
> One alternative would be instead of using MDB_env* address, using a
> monotonically increasing counter based on a static global, say, "mrb_token",
> but this might break in the case two copies of LMDB were somehow linked in the
> application. (It's easy to argue the user is an idiot in this case ;)

Global state is verboten.
>
>
> David
>
>


-- 
   -- Howard Chu
   CTO, Symas Corp.           http://www.symas.com
   Director, Highland Sun     http://highlandsun.com/hyc/
   Chief Architect, OpenLDAP  http://www.openldap.org/project/