[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: Possible transaction issue with LMDB



William Brown wrote:
On Fri, 2018-08-17 at 01:34 +0300, Леонид Юрьев wrote:
Hi,

Shortly, no issue.

LMDB provides MVCC via COW, therefore any read transaction see the
constant snapshot of DB, which is correspond to last committed write
transaction at the time when reading was started.

Any further write transactions will create the new snapshots, which
will be visible for further reading, but not for read transactions
that was started before.



Hi,

I'm quite aware that it is COW - this issue is specific to COW trees.
Transactions must be removed in order, they can not be removed out of
order. It is about how pages are reclaimed for the freelist

Incorrect.

If you have tree state A, you have nodes say 1 through 4. We'll say 1
is the root, and 2 - 4 are the leaves.

We take the transaction at A.

Now we conduct a write with TXN X. We'll copy nodes 1 (r) and 4 to 5
(r) and 6 now and make our update. We now commit. At this point this is
the "last generation" where 1 and 4 exist.

We begin the transaction at B

Now we condact a write with TXN Y. We'll copy nodes 2, 3 and 5 (r). 5
Being the root from X, 2,3 still in use at A. We now commit.

We begin a read transaction C.

At this point we now close transaction B.

B now clears it's resources - Because B is the last location where 2,3
were "alive", they are at this point freed. however, they are *still
required* by TXN A for a valid and complete tree.

Closing read txn B has no effect on the tree. Closing read txns has no
effect on the on-disk structures.

At this point, TXN A now consists of nodes 1, 4, and nodes 2,3 are on
the freelist where they are possibly able to be reused.

False. Closing a read txn doesn't put any nodes on the freelist. Only committing
a write txn can do so.

You begin a new write, and commit many values. This will over-write the
content of nodes 2,3 (as they are in the free list), causing the reader
of TXN A to now percieve invalid data.

False.

I have previously seen this with my in-memory only tree,

but I was able
to recreate this with LMDB.

Impossible. There is no code path in LMDB that can do what you have described.

Unless... you have disabled LMDB's lock management and are doing your own thing,
and you simply got it wrong.

I'm not able to release the POC at this
time. The issue arises because transactions *must* be cleared in order
from oldest to newest, as a fundamental part of COW is the shared
resources, and understanding when node lifetimes expire.

Your assertions have no bearing on reality, as far as LMDB goes.

I hope that this explains the issue more thoroughly.


Regards,
Leonid.


2018-08-16 13:03 GMT+03:00 William Brown <william@blackhats.net.au>:
Hi there,

While doing some integration testing of LMDB I noticed that there
may
be an issue with out of order transaction handling.

The scenario is:

Open Read TXN A
Open Write TXN X, and change values of the DB
Commit X
Open Read TXN B
Open Write TXN Y, and change values of the DB
Commit Y
Open Read TXN C

Abort/Close TXN B.

At this point, because of the page touch between A -> B and B -> C,
B
now believes that the pages of A are the "last time" they are
available
as they were all subsequently copied for TXN C. The pages of A are
then
added to the freelists when B closes. When TXN A is read from the
data
may have been altered due to future writes as LMDB attempts to use
previously allocated pages first.

This situation is more likely to arise on large batch writes, but
could
manifest with smaller series of writes. This would be a silent
issue,
as the over-written pages may be valid, and could cause data to
"silently vanish" inside of the read transaction A leading to
unpredictable results.

I hope that this report helps you to diagnose and resolve the
issue.



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