Обсуждение: Restore CurrentUserId only if 'prevUser' is valid when abort transaction

On Wed, Oct 10, 2018 at 03:37:50PM +0800, Richard Guo wrote:
> This is a follow-up to the issue described in thread
> https://www.postgresql.org/message-id/CAMbWs4-Mys%3DhBQSevTA8Zpd-TYFnb%3DXuHhN2TnktXMsfMUbjiQ%40mail.gmail.com
>
> In short, during the first transaction starting phase within a backend, if
> there is an 'ereport' after setting transaction state but before saving
> CurrentUserId into 'prevUser' in TransactionStateData, 'prevUser' will
> remain as InvalidOid. Then in AbortTransaction(), CurrentUserId is
> restored with 'prevUser'. As a result, CurrentUserId will be
> InvalidOid in the rest of the session.
>
> Attached is a patch that fixes this issue.

I guess that's an issue showing up with Greenplum as you folks are
likely tweaking how a transaction start happens?

It is as easy as doing something like that in StartTransaction() to get
into a failed state:
    s->state = TRANS_START;
    s->transactionId = InvalidTransactionId;    /* until assigned */

+   {
+       struct stat statbuf;
+       if (stat("/tmp/hoge", &statbuf) == 0)
+           elog(ERROR, "hoge invalid state!");
+   }

Then do something like the following:
1) Start a session
2) touch /tmp/hoge
3) Issue BEGIN, which fails and initializes CurrentUserId to InvalidOid.
4) rm /tmp/hoge
3) any DDL causes the system to crash.

Anyway, looking at the patch, I am poked by the comment on top of
GetUserIdAndSecContext which states that InvalidOid can be a possible
value.  It seems to me that the root of the problem is that TRANS_STATE
is enforced to TRANS_INPROGRESS when aborting a transaction in a
starting state, in which case we should not have to reset CurrentUserId
as it has never been set.  The main reason why this was done is to
prevent a warning message to show up.

Tom, eedb068c0 is in cause here, and that's your commit.  Could you
check if something like the attached is adapted?  I am pretty sure that
we still want the sub-transaction part to still reset CurrentUserId
unconditionally by the way.

Thanks,
--
Michael

On Fri, Oct 12, 2018 at 02:28:36PM +0800, Richard Guo wrote:
> I think it is a better idea to avoid adjusting the state to TRANS_INPROGRESS
> from TRANS_START when aborting a transaction, as your patch does, since its
> only purpose is to suppress warning message.

Actually, as StartSubTransaction also switches to TRANS_START for a
savepoint, if there is an error until the state is switched to
TRANS_INPROGRESS then the code would fail to switch back to
CurrentUserId even if it is set, and it should be switched.  So that
solution is not correct either as AtSubStart_ResourceOwner() or such
could fail on memory allocation.  That's unlikely going to happen, but
it could.
--
Michael

On Tue, Oct 23, 2018 at 02:40:30PM +0900, Michael Paquier wrote:
> Actually, as StartSubTransaction also switches to TRANS_START for a
> savepoint, if there is an error until the state is switched to
> TRANS_INPROGRESS then the code would fail to switch back to
> CurrentUserId even if it is set, and it should be switched.  So that
> solution is not correct either as AtSubStart_ResourceOwner() or such
> could fail on memory allocation.  That's unlikely going to happen, but
> it could.

So, I have spent a couple of hours today looking a bit more at the
problem, and I have hacked the attached patch that I am pretty happy
with:
- Normal transactions can rely on TRANS_START to decide if the security
context can be reset or not.
- When defining a savepoint, the subtransaction state is initialized by
PushTransaction() before being switched to its sub-in-progress state
when the query which created the savepoint commits.  In this case, we
should call GetUserIdAndSecContext() just before switching the
transaction context.

The patch includes a set tweaks I used to inject some errors in specific
code paths and trigger failures, checking if a security context which
has been set is correctly reset:
- /tmp/error_start for the end of StartTransaction
- /tmp/error_sub for the end of StartSubTransaction
- /tmp/error_push for the end of PushTransaction.

Like on HEAD, this patch still triggers the following WARNING if
injecting an error in PushTransaction as StartSubTransaction has not
switched the status of the transaction yet:
AbortSubTransaction while in DEFAULT state

Another WARNING which can be reached is the following if injecting an
error in StartSubTransaction:
AbortSubTransaction while in START state

Per the set of routines called when starting the subtransaction, I think
that we ought to do as main transactions and silence this warning
equally.  I am attaching the patch for review by others.  Please note
that this includes the error injections to ease tests.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:
> So, I have spent a couple of hours today looking a bit more at the
> problem, and I have hacked the attached patch that I am pretty happy
> with:

I don't like this too much, because it does not scale: there can be
only one action that can rely on executing "just before we switch to
TRANS_INPROGRESS".

I think the real bug here is that a bunch of potentially-failable
operations have been thrown in before we've finished initializing the
TransactionState to minimal sanity.  (Inserting code with the aid of a
dartboard seems to be a chronic disease around here :-(.)  Since
GetUserIdAndSecContext is *not* an operation that can fail, there's
no reason why we need to expend a lot of effort on the possibility that
it hasn't happened.  What we ought to do is move that and the rest of the
"initialize current transaction state fields" stanza up to before we start
doing things like calling RecoveryInProgress().  And put in a comment to
clearly mark where we first allow failure to occur.

I'd be strongly inclined to change the elog(WARNING) at line 1815
to an assertion, because calling elog exposes us to all kinds of
hazards that we don't need here.

            regards, tom lane

On Mon, Nov 12, 2018 at 08:17:29PM -0500, Tom Lane wrote:
> I don't like this too much, because it does not scale: there can be
> only one action that can rely on executing "just before we switch to
> TRANS_INPROGRESS".

Okay.

> I think the real bug here is that a bunch of potentially-failable
> operations have been thrown in before we've finished initializing the
> TransactionState to minimal sanity.  (Inserting code with the aid of a
> dartboard seems to be a chronic disease around here :-(.)  Since
> GetUserIdAndSecContext is *not* an operation that can fail, there's
> no reason why we need to expend a lot of effort on the possibility that
> it hasn't happened.  What we ought to do is move that and the rest of the
> "initialize current transaction state fields" stanza up to before we start
> doing things like calling RecoveryInProgress().  And put in a comment to
> clearly mark where we first allow failure to occur.

When first working on the patch I got to wonder if there were any
intermediate states which relied on the user ID of the security context
flags which could have justified its current position.  Just checking
now it looks safe to move up the call.  I have checked as well my test
cases injecting errors.  What do you think about the attached?

Also, I think that we should backpatch something all the way down.
An ERROR in this code path is perhaps unlikely to happen but having
Postgres to crash if the ERROR shows for the first query of a session is
not nice.  Any thoughts about that?

> I'd be strongly inclined to change the elog(WARNING) at line 1815
> to an assertion, because calling elog exposes us to all kinds of
> hazards that we don't need here.

No objections from here.  I would do that only on HEAD though.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:
> On Mon, Nov 12, 2018 at 08:17:29PM -0500, Tom Lane wrote:
>> I think the real bug here is that a bunch of potentially-failable
>> operations have been thrown in before we've finished initializing the
>> TransactionState to minimal sanity.  (Inserting code with the aid of a
>> dartboard seems to be a chronic disease around here :-(.)

> When first working on the patch I got to wonder if there were any
> intermediate states which relied on the user ID of the security context
> flags which could have justified its current position.  Just checking
> now it looks safe to move up the call.  I have checked as well my test
> cases injecting errors.  What do you think about the attached? 

Looks sane to me.

> Also, I think that we should backpatch something all the way down.

Yes.

>> I'd be strongly inclined to change the elog(WARNING) at line 1815
>> to an assertion, because calling elog exposes us to all kinds of
>> hazards that we don't need here.

> No objections from here.  I would do that only on HEAD though.

Well, if it is an issue then it's an issue for back branches too.
It's fine, I think, as long as the warning stays a warning ...
but there are lots of ways in which trying to print a warning
might turn into an error.

            regards, tom lane

On Tue, Nov 13, 2018 at 04:08:04PM -0500, Tom Lane wrote:
> Well, if it is an issue then it's an issue for back branches too.
> It's fine, I think, as long as the warning stays a warning ...
> but there are lots of ways in which trying to print a warning
> might turn into an error.

At the end I have agreed with this position, and patched back-branches
so as they use an assertion instead of the elog().  So committed.
--
Michael