Обсуждение: Problem with txid_snapshot_in/out() functionality
Hackers, the Slony team has been getting seldom reports of a problem with the txid_snapshot data type. The symptom is that a txid_snapshot on output lists the same txid multiple times in the xip list part of the external representation. This string is later rejected by txid_snapshot_in() when trying to determine if a particular txid is visible in that snapshot using the txid_visible_in_snapshot() function. I was not yet able to reproduce this problem in a lab environment. It might be related to subtransactions and/or two phase commit (at least one user is using both of them). The reported PostgreSQL version involved in that case was 9.1. At this point I would find it extremely helpful to "sanitize" the external representation in txid_snapshot_out() while emitting some NOTICE level logging when this actually happens. I am aware that this does amount to a functional change for a back release, but considering that the _out() generated external representation of an existing binary datum won't pass the type's _in() function, I argue that such change is warranted. Especially since this problem could possibly corrupt a dump. Comments? Jan -- Jan Wieck Senior Software Engineer http://slony.info
On 04/12/2014 12:07 AM, Jan Wieck wrote: > Hackers, > > the Slony team has been getting seldom reports of a problem with the > txid_snapshot data type. > > The symptom is that a txid_snapshot on output lists the same txid > multiple times in the xip list part of the external representation. This > string is later rejected by txid_snapshot_in() when trying to determine > if a particular txid is visible in that snapshot using the > txid_visible_in_snapshot() function. > > I was not yet able to reproduce this problem in a lab environment. It > might be related to subtransactions and/or two phase commit (at least > one user is using both of them). The reported PostgreSQL version > involved in that case was 9.1. It's two-phase commit. When preparing a transaction, the state of the transaction is first transfered to a dummy PGXACT entry, and then the PGXACT entry of the backend is cleared. There is a transient state when both PGXACT entries have the same xid. You can reproduce that by putting a sleep or breakpoint in PrepareTransaction(), just before the "ProcArrayClearTransaction(MyProc);" call. If you call txid_current_snapshot() from another session at that point, it will output two duplicate xids. (you will have to also commit one more unrelated transaction to bump up xmax). > At this point I would find it extremely helpful to "sanitize" the > external representation in txid_snapshot_out() while emitting some > NOTICE level logging when this actually happens. I am aware that this > does amount to a functional change for a back release, but considering > that the _out() generated external representation of an existing binary > datum won't pass the type's _in() function, I argue that such change is > warranted. Especially since this problem could possibly corrupt a dump. Hmm. Do we snapshots to be stored in tables, and included in a dump? I don't think we can guarantee that will work, at least not across versions, as the way we handle snapshot internally can change. But yeah, we probably should do something about that. The most straightforward fix would be to scan the array in txid_current_snapshot() and remove any duplicates. - Heikki
On 04/12/14 03:27, Heikki Linnakangas wrote: > On 04/12/2014 12:07 AM, Jan Wieck wrote: >> Hackers, >> >> the Slony team has been getting seldom reports of a problem with the >> txid_snapshot data type. >> >> The symptom is that a txid_snapshot on output lists the same txid >> multiple times in the xip list part of the external representation. This >> string is later rejected by txid_snapshot_in() when trying to determine >> if a particular txid is visible in that snapshot using the >> txid_visible_in_snapshot() function. >> >> I was not yet able to reproduce this problem in a lab environment. It >> might be related to subtransactions and/or two phase commit (at least >> one user is using both of them). The reported PostgreSQL version >> involved in that case was 9.1. > > It's two-phase commit. When preparing a transaction, the state of the > transaction is first transfered to a dummy PGXACT entry, and then the > PGXACT entry of the backend is cleared. There is a transient state when > both PGXACT entries have the same xid. > > You can reproduce that by putting a sleep or breakpoint in > PrepareTransaction(), just before the > "ProcArrayClearTransaction(MyProc);" call. If you call > txid_current_snapshot() from another session at that point, it will > output two duplicate xids. (you will have to also commit one more > unrelated transaction to bump up xmax). Thanks, that explains it. > >> At this point I would find it extremely helpful to "sanitize" the >> external representation in txid_snapshot_out() while emitting some >> NOTICE level logging when this actually happens. I am aware that this >> does amount to a functional change for a back release, but considering >> that the _out() generated external representation of an existing binary >> datum won't pass the type's _in() function, I argue that such change is >> warranted. Especially since this problem could possibly corrupt a dump. > > Hmm. Do we snapshots to be stored in tables, and included in a dump? I > don't think we can guarantee that will work, at least not across > versions, as the way we handle snapshot internally can change. At least Londiste and Slony do store snapshots as well as xids in tables and assuming that the txid epoch is properly bumped, that information is useful and valid after a restore. > > But yeah, we probably should do something about that. The most > straightforward fix would be to scan the array in > txid_current_snapshot() and remove any duplicates. The code in txid_snapshot_in() checks that the xip list is ascending. txid_snapshot_out() does not sort the list, so it must already be sorted when the snapshot itself is created. That scan would be fairly simple. Jan -- Jan Wieck Senior Software Engineer http://slony.info
On 2014-04-12 10:27:16 +0300, Heikki Linnakangas wrote: > On 04/12/2014 12:07 AM, Jan Wieck wrote: > >The symptom is that a txid_snapshot on output lists the same txid > >multiple times in the xip list part of the external representation. This > >string is later rejected by txid_snapshot_in() when trying to determine > >if a particular txid is visible in that snapshot using the > >txid_visible_in_snapshot() function. > It's two-phase commit. When preparing a transaction, the state of the > transaction is first transfered to a dummy PGXACT entry, and then the PGXACT > entry of the backend is cleared. There is a transient state when both PGXACT > entries have the same xid. Which I find to be a pretty bad idea independent of this bug. But I think that's nothing fixable in the back branches. > Hmm. Do we snapshots to be stored in tables, and included in a dump? I don't > think we can guarantee that will work, at least not across versions, as the > way we handle snapshot internally can change. Hm. I don't think we'll earn much love changing that - there's at the very least slony and londiste out there using it... IIRC both store the result in tables. > But yeah, we probably should do something about that. The most > straightforward fix would be to scan the array in txid_current_snapshot() > and remove any duplicates. Since it's sorted there, that should be fairly straightforward. Won't fix already created and stored datums tho. Maybe _in()/parse_snapshot() should additionally skip over duplicate values? Looks easy enough. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 04/12/14 08:38, Andres Freund wrote: > Since it's sorted there, that should be fairly straightforward. Won't > fix already created and stored datums tho. Maybe _in()/parse_snapshot() > should additionally skip over duplicate values? Looks easy enough. There is the sort ... missed that when glancing over the code earlier. Right, that is easy enough and looks like an acceptable fix for back branches too. Thanks, Jan -- Jan Wieck Senior Software Engineer http://slony.info
Heikki Linnakangas <hlinnakangas@vmware.com> writes:
> On 04/12/2014 12:07 AM, Jan Wieck wrote:
>> the Slony team has been getting seldom reports of a problem with the
>> txid_snapshot data type.
>> The symptom is that a txid_snapshot on output lists the same txid
>> multiple times in the xip list part of the external representation.
> It's two-phase commit. When preparing a transaction, the state of the
> transaction is first transfered to a dummy PGXACT entry, and then the
> PGXACT entry of the backend is cleared. There is a transient state when
> both PGXACT entries have the same xid.
Hm, yeah, but why is that intermediate state visible to anyone else?
Don't we have exclusive lock on the PGPROC array while we're doing this?
If we don't, aren't we letting other backends see non-self-consistent
state in regards to who holds which locks, for example?
I'm worried that the proposed fix is just band-aiding one particular
symptom of inadequate locking.
regards, tom lane
On 2014-04-12 09:47:24 -0400, Tom Lane wrote: > Heikki Linnakangas <hlinnakangas@vmware.com> writes: > > On 04/12/2014 12:07 AM, Jan Wieck wrote: > >> the Slony team has been getting seldom reports of a problem with the > >> txid_snapshot data type. > >> The symptom is that a txid_snapshot on output lists the same txid > >> multiple times in the xip list part of the external representation. > > > It's two-phase commit. When preparing a transaction, the state of the > > transaction is first transfered to a dummy PGXACT entry, and then the > > PGXACT entry of the backend is cleared. There is a transient state when > > both PGXACT entries have the same xid. > > Hm, yeah, but why is that intermediate state visible to anyone else? > Don't we have exclusive lock on the PGPROC array while we're doing this? It's done outside the remit of ProcArray lock :(. And documented to lead to duplicate xids in PGXACT. EndPrepare():/* * Mark the prepared transaction as valid. As soon as xact.c marks * MyPgXact as not running our XID (whichit will do immediately after * this function returns), others can commit/rollback the xact. * * NB: a side effect ofthis is to make a dummy ProcArray entry for the * prepared XID. This must happen before we clear the XID from MyPgXact,* else there is a window where the XID is not running according to * TransactionIdIsInProgress, and onlookers wouldbe entitled to assume * the xact crashed. Instead we have a window where the same XID appears * twice in ProcArray,which is OK. */MarkAsPrepared(gxact); It doesn't sound too hard to essentially move PrepareTransaction()'s ProcArrayClearTransaction() into MarkAsPrepared() and rejigger the locking to remove the intermediate state. But I think it'll lead to bigger changes than we'd be comfortable backpatching. > If we don't, aren't we letting other backends see non-self-consistent > state in regards to who holds which locks, for example? I think that actually works out ok, because the locks aren't owned by xids/xacts, but procs. Otherwise we'd be in deep trouble in CommitTransaction() as well where ProcArrayEndTransaction() clearing that state. After the whole xid transfer, there's PostPrepare_Locks() transferring the locks. Brr. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
<p dir="ltr"><br /> On 12 Apr 2014 08:35, "Jan Wieck" <<a href="mailto:jan@wi3ck.info">jan@wi3ck.info</a>> wrote:<br/> ><br /> > On 04/12/14 03:27, Heikki Linnakangas wrote:<br /> >><br /> >> On 04/12/2014 12:07AM, Jan Wieck wrote:<br /> >>><br /> >>> Hackers,<p dir="ltr">>> Hmm. Do we snapshots to bestored in tables, and included in a dump? I<br /> >> don't think we can guarantee that will work, at least not across<br/> >> versions, as the way we handle snapshot internally can change.<br /> ><br /> ><br /> > At leastLondiste and Slony do store snapshots as well as xids in tables and assuming that the txid epoch is properly bumped,that information is useful and valid after a restore.<p dir="ltr">As I understand it the epoch increments wheneverthe xid wraps. <p dir="ltr">A physical restore would continue the same xid space in the same epoch which should workfine as long as no system stores any txids outside the database from the "future".<p dir="ltr">A pg_restore would starta new xid space from FirstNormalXid which would obviously not work with any stored txids.
On 04/12/14 10:09, Greg Stark wrote: > A pg_restore would start a new xid space from FirstNormalXid which would > obviously not work with any stored txids. > http://www.postgresql.org/docs/9.3/static/app-pgresetxlog.html Regards, Jan -- Jan Wieck Senior Software Engineer http://slony.info
On 2014-04-12 11:15:09 -0400, Jan Wieck wrote: > On 04/12/14 10:09, Greg Stark wrote: > > >A pg_restore would start a new xid space from FirstNormalXid which would > >obviously not work with any stored txids. > > > > http://www.postgresql.org/docs/9.3/static/app-pgresetxlog.html Using that as part of any sort of routine task IMNSHO is a seriously bad idea. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 04/12/14 11:18, Andres Freund wrote: > On 2014-04-12 11:15:09 -0400, Jan Wieck wrote: >> On 04/12/14 10:09, Greg Stark wrote: >> >> >A pg_restore would start a new xid space from FirstNormalXid which would >> >obviously not work with any stored txids. >> > >> >> http://www.postgresql.org/docs/9.3/static/app-pgresetxlog.html > > Using that as part of any sort of routine task IMNSHO is a seriously bad > idea. Nobody is advocating doing so. Jan -- Jan Wieck Senior Software Engineer http://slony.info
On 04/12/14 10:03, Andres Freund wrote: > On 2014-04-12 09:47:24 -0400, Tom Lane wrote: >> Heikki Linnakangas <hlinnakangas@vmware.com> writes: >> > On 04/12/2014 12:07 AM, Jan Wieck wrote: >> >> the Slony team has been getting seldom reports of a problem with the >> >> txid_snapshot data type. >> >> The symptom is that a txid_snapshot on output lists the same txid >> >> multiple times in the xip list part of the external representation. >> >> > It's two-phase commit. When preparing a transaction, the state of the >> > transaction is first transfered to a dummy PGXACT entry, and then the >> > PGXACT entry of the backend is cleared. There is a transient state when >> > both PGXACT entries have the same xid. >> >> Hm, yeah, but why is that intermediate state visible to anyone else? >> Don't we have exclusive lock on the PGPROC array while we're doing this? > > It's done outside the remit of ProcArray lock :(. And documented to lead > to duplicate xids in PGXACT. > EndPrepare(): > /* > * Mark the prepared transaction as valid. As soon as xact.c marks > * MyPgXact as not running our XID (which it will do immediately after > * this function returns), others can commit/rollback the xact. > * > * NB: a side effect of this is to make a dummy ProcArray entry for the > * prepared XID. This must happen before we clear the XID from MyPgXact, > * else there is a window where the XID is not running according to > * TransactionIdIsInProgress, and onlookers would be entitled to assume > * the xact crashed. Instead we have a window where the same XID appears > * twice in ProcArray, which is OK. > */ > MarkAsPrepared(gxact); > > It doesn't sound too hard to essentially move PrepareTransaction()'s > ProcArrayClearTransaction() into MarkAsPrepared() and rejigger the > locking to remove the intermediate state. But I think it'll lead to > bigger changes than we'd be comfortable backpatching. > >> If we don't, aren't we letting other backends see non-self-consistent >> state in regards to who holds which locks, for example? > > I think that actually works out ok, because the locks aren't owned by > xids/xacts, but procs. Otherwise we'd be in deep trouble in > CommitTransaction() as well where ProcArrayEndTransaction() clearing > that state. > After the whole xid transfer, there's PostPrepare_Locks() transferring > the locks. Since it doesn't seem to produce any side effects, I'd think that making the snapshot unique within txid_current_snapshot() and filtering duplicates on input should be sufficient and eligible for backpatching. The attached patch adds a unique loop to the internal sort_snapshot() function and skips duplicates on input. The git commit is here: https://github.com/wieck/postgres/commit/a88a2b2c25b856478d7e2b012fc718106338fe00 Regards, Jan -- Jan Wieck Senior Software Engineer http://slony.info
Вложения
On Sat, Apr 12, 2014 at 02:10:13PM -0400, Jan Wieck wrote: > Since it doesn't seem to produce any side effects, I'd think that > making the snapshot unique within txid_current_snapshot() and > filtering duplicates on input should be sufficient and eligible for > backpatching. Agreed. > The attached patch adds a unique loop to the internal > sort_snapshot() function and skips duplicates on input. The git > commit is here: > > https://github.com/wieck/postgres/commit/a88a2b2c25b856478d7e2b012fc718106338fe00 > static void > sort_snapshot(TxidSnapshot *snap) > { > + txid last = 0; > + int nxip, idx1, idx2; > + > if (snap->nxip > 1) > + { > qsort(snap->xip, snap->nxip, sizeof(txid), cmp_txid); > + nxip = snap->nxip; > + idx1 = idx2 = 0; > + while (idx1 < nxip) > + { > + if (snap->xip[idx1] != last) > + last = snap->xip[idx2++] = snap->xip[idx1]; > + else > + snap->nxip--; > + idx1++; > + } > + } > } I think you need to do SET_VARSIZE also here. Alternative is to move SET_VARSIZE after sort_snapshot(). And it seems the drop-double-txid logic should be added also to txid_snapshot_recv(). It seems weird to have it behave differently from txid_snapshot_in(). -- marko
On 04/13/14 08:27, Marko Kreen wrote: > On Sat, Apr 12, 2014 at 02:10:13PM -0400, Jan Wieck wrote: >> Since it doesn't seem to produce any side effects, I'd think that >> making the snapshot unique within txid_current_snapshot() and >> filtering duplicates on input should be sufficient and eligible for >> backpatching. > > Agreed. > >> The attached patch adds a unique loop to the internal >> sort_snapshot() function and skips duplicates on input. The git >> commit is here: >> >> https://github.com/wieck/postgres/commit/a88a2b2c25b856478d7e2b012fc718106338fe00 > >> static void >> sort_snapshot(TxidSnapshot *snap) >> { >> + txid last = 0; >> + int nxip, idx1, idx2; >> + >> if (snap->nxip > 1) >> + { >> qsort(snap->xip, snap->nxip, sizeof(txid), cmp_txid); >> + nxip = snap->nxip; >> + idx1 = idx2 = 0; >> + while (idx1 < nxip) >> + { >> + if (snap->xip[idx1] != last) >> + last = snap->xip[idx2++] = snap->xip[idx1]; >> + else >> + snap->nxip--; >> + idx1++; >> + } >> + } >> } > > I think you need to do SET_VARSIZE also here. Alternative is to > move SET_VARSIZE after sort_snapshot(). > > And it seems the drop-double-txid logic should be added also to > txid_snapshot_recv(). It seems weird to have it behave differently > from txid_snapshot_in(). > Thanks, yes on both issues. Will create another patch. Jan -- Jan Wieck Senior Software Engineer http://slony.info
On 04/12/2014 05:03 PM, Andres Freund wrote: >> >If we don't, aren't we letting other backends see non-self-consistent >> >state in regards to who holds which locks, for example? > I think that actually works out ok, because the locks aren't owned by > xids/xacts, but procs. Otherwise we'd be in deep trouble in > CommitTransaction() as well where ProcArrayEndTransaction() clearing > that state. > After the whole xid transfer, there's PostPrepare_Locks() transferring > the locks. Right. However, I just noticed that there's a race condition between PREPARE TRANSACTION and COMMIT/ROLLBACK PREPARED. PostPrepare_Locks runs after the prepared transaction is already marked as fully prepared. That means that by the time we get to PostPrepare_Locks, another backend might already have finished and removed the prepared transaction. That leads to a PANIC (put a breakpoint just before PostPrepare_Locks): postgres=# commit prepared 'foo'; PANIC: failed to re-find shared proclock object PANIC: failed to re-find shared proclock object The connection to the server was lost. Attempting reset: Failed. FinishPrepareTransaction reads the list of locks from the two-phase state file, but PANICs when it doesn't find the corresponding locks in the lock manager (because PostPrepare_Locks hasn't transfered them to the dummy PGPROC yet). I think we'll need to transfer of the locks earlier, before the transaction is marked as fully prepared. I'll take a closer look at this tomorrow. - Heikki
On 04/13/14 14:22, Jan Wieck wrote: > On 04/13/14 08:27, Marko Kreen wrote: >> I think you need to do SET_VARSIZE also here. Alternative is to >> move SET_VARSIZE after sort_snapshot(). >> >> And it seems the drop-double-txid logic should be added also to >> txid_snapshot_recv(). It seems weird to have it behave differently >> from txid_snapshot_in(). >> > > Thanks, > > yes on both issues. Will create another patch. New patch attached. New github commit is https://github.com/wieck/postgres/commit/b8fd0d2eb78791e5171e34aecd233fd06218890d Thanks again, Jan -- Jan Wieck Senior Software Engineer http://slony.info
Вложения
On 04/13/2014 11:39 PM, Heikki Linnakangas wrote: > However, I just noticed that there's a race condition between PREPARE > TRANSACTION and COMMIT/ROLLBACK PREPARED. PostPrepare_Locks runs after > the prepared transaction is already marked as fully prepared. That means > that by the time we get to PostPrepare_Locks, another backend might > already have finished and removed the prepared transaction. That leads > to a PANIC (put a breakpoint just before PostPrepare_Locks): > > postgres=# commit prepared 'foo'; > PANIC: failed to re-find shared proclock object > PANIC: failed to re-find shared proclock object > The connection to the server was lost. Attempting reset: Failed. > > FinishPrepareTransaction reads the list of locks from the two-phase > state file, but PANICs when it doesn't find the corresponding locks in > the lock manager (because PostPrepare_Locks hasn't transfered them to > the dummy PGPROC yet). > > I think we'll need to transfer of the locks earlier, before the > transaction is marked as fully prepared. I'll take a closer look at this > tomorrow. Here's a patch to do that. It's very straightforward, I just moved the calls to transfer locks earlier, before ProcArrayClearTransaction. PostPrepare_MultiXact had a similar race - it also transfer state from the old PGPROC entry to the new, and needs to be done before allowing another backend to remove the new PGPROC entry. I changed the names of the functions to distinguish them from the other PostPrepare_* functions that now happen at a different time. The patch is simple, but it's a bit scary to change the order of things like this. Looking at all the calls that now happen after transferring the locks, I believe this is OK. The change also applies to the callbacks called by the RegisterXactCallback mechanism, which means that in theory there might be a 3rd party extension out there that's affected. All the callbacks in contrib and plpgsql are OK, and it's questionable to do anything complicated that would depend on heavy-weight locks to be held in those callbacks, so I think this is OK. Warrants a note in the release notes, though. - Heikki
Вложения
On Sun, Apr 13, 2014 at 05:46:20PM -0400, Jan Wieck wrote: > On 04/13/14 14:22, Jan Wieck wrote: > >On 04/13/14 08:27, Marko Kreen wrote: > >>I think you need to do SET_VARSIZE also here. Alternative is to > >>move SET_VARSIZE after sort_snapshot(). > >> > >>And it seems the drop-double-txid logic should be added also to > >>txid_snapshot_recv(). It seems weird to have it behave differently > >>from txid_snapshot_in(). > >> > > > >Thanks, > > > >yes on both issues. Will create another patch. > > New patch attached. > > New github commit is https://github.com/wieck/postgres/commit/b8fd0d2eb78791e5171e34aecd233fd06218890d Looks OK to me. -- marko
On 04/12/2014 05:03 PM, Andres Freund wrote: > On 2014-04-12 09:47:24 -0400, Tom Lane wrote: >> Heikki Linnakangas <hlinnakangas@vmware.com> writes: >>> On 04/12/2014 12:07 AM, Jan Wieck wrote: >>>> the Slony team has been getting seldom reports of a problem with the >>>> txid_snapshot data type. >>>> The symptom is that a txid_snapshot on output lists the same txid >>>> multiple times in the xip list part of the external representation. >> >>> It's two-phase commit. When preparing a transaction, the state of the >>> transaction is first transfered to a dummy PGXACT entry, and then the >>> PGXACT entry of the backend is cleared. There is a transient state when >>> both PGXACT entries have the same xid. >> >> Hm, yeah, but why is that intermediate state visible to anyone else? >> Don't we have exclusive lock on the PGPROC array while we're doing this? > > It's done outside the remit of ProcArray lock :(. And documented to lead > to duplicate xids in PGXACT. > EndPrepare(): > /* > * Mark the prepared transaction as valid. As soon as xact.c marks > * MyPgXact as not running our XID (which it will do immediately after > * this function returns), others can commit/rollback the xact. > * > * NB: a side effect of this is to make a dummy ProcArray entry for the > * prepared XID. This must happen before we clear the XID from MyPgXact, > * else there is a window where the XID is not running according to > * TransactionIdIsInProgress, and onlookers would be entitled to assume > * the xact crashed. Instead we have a window where the same XID appears > * twice in ProcArray, which is OK. > */ > MarkAsPrepared(gxact); > > It doesn't sound too hard to essentially move PrepareTransaction()'s > ProcArrayClearTransaction() into MarkAsPrepared() and rejigger the > locking to remove the intermediate state. But I think it'll lead to > bigger changes than we'd be comfortable backpatching. Hmm. There's a field in GlobalTransactionData called locking_xid, which is used to mark the XID of the transaction that's currently operating on the prepared transaction. At prepare, that ensures that the transaction cannot be committed or rolled back by another backend until the original backend has cleared its PGPROC entry. At COMMIT/ROLLBACK PREPARED, it ensures that only one backend can commit/rollback the transaction. I wonder why we don't use a VirtualTransactionId there. AFAICS there is no reason for COMMIT/ROLLBACK PREPARED to be assigned an XID of its own. And if we used a VirtualTransactionId there, prepare could clear the xid field of the PGPROC entry earlier. - Heikki
On 2014-04-14 12:15:30 +0300, Heikki Linnakangas wrote: > Hmm. There's a field in GlobalTransactionData called locking_xid, which is > used to mark the XID of the transaction that's currently operating on the > prepared transaction. At prepare, that ensures that the transaction cannot > be committed or rolled back by another backend until the original backend > has cleared its PGPROC entry. At COMMIT/ROLLBACK PREPARED, it ensures that > only one backend can commit/rollback the transaction. > > I wonder why we don't use a VirtualTransactionId there. I wondered about that previously as well. My bet it's because the 2pc support arrived before the virtualxact stuff... Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
Heikki Linnakangas <hlinnakangas@vmware.com> writes:
>> I think we'll need to transfer of the locks earlier, before the
>> transaction is marked as fully prepared. I'll take a closer look at this
>> tomorrow.
> Here's a patch to do that. It's very straightforward, I just moved the
> calls to transfer locks earlier, before ProcArrayClearTransaction.
> PostPrepare_MultiXact had a similar race - it also transfer state from
> the old PGPROC entry to the new, and needs to be done before allowing
> another backend to remove the new PGPROC entry. I changed the names of
> the functions to distinguish them from the other PostPrepare_* functions
> that now happen at a different time.
Why didn't you also move up PostPrepare_PredicateLocks? Seems like its
access to MySerializableXact is also racy.
> The patch is simple, but it's a bit scary to change the order of things
> like this.
Yeah. There are a lot of assumptions in there about the order of resource
release, in particular that it is safe to do certain things because we're
still holding locks.
I poked around a bit and noticed one theoretical problem sequence: if the
prepared xact drops some relation that we're still holding buffer pins on.
This shouldn't really happen (why are we still pinning some rel we think
we dropped?) but if it did, the commit would do DropRelFileNodeBuffers
which would end up busy-looping until we drop our pins (see
InvalidateBuffer, which thinks this must be an I/O wait situation).
So it would work, more or less, but it seems pretty fragile. I'm afraid
there are more assumptions like this one.
The whole thing feels like we are solving the wrong problem, anyway.
IIUC, the complaint arises because we are allowing COMMIT PREPARED
to occur before the source transaction has reported successful prepare
to its client. Surely that does not need to be a legal case? No
correctly-operating 2PC xact manager would do that.
I'd prefer to leave the prepare sequence alone and instead find a way
to reject COMMIT PREPARED until after the source transaction is safely
clear of the race conditions. The upthread idea of looking at vxid
instead of xid might help, except that I see we clear both of them
in ProcArrayClearTransaction. We'd need some state in PGPROC that
isn't cleared till later than that.
regards, tom lane
On 2014-04-14 12:51:02 -0400, Tom Lane wrote: > The whole thing feels like we are solving the wrong problem, anyway. > IIUC, the complaint arises because we are allowing COMMIT PREPARED > to occur before the source transaction has reported successful prepare > to its client. Surely that does not need to be a legal case? No > correctly-operating 2PC xact manager would do that. I agree here. This seems somewhat risky, just to support a case that shouldn't happen in reality - as somewhat evidenced by the fact that there don't seem to be field reports around this. > The upthread idea of looking at vxid > instead of xid might help, except that I see we clear both of them > in ProcArrayClearTransaction. We'd need some state in PGPROC that > isn't cleared till later than that. I wonder if the most natural way to express this wouldn't be to have a heavyweight lock for every 2pc xact 'slot'. ResourceOwnerRelease(RESOURCE_RELEASE_LOCKS) should be scheduled correctly to make error handling for this work. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
Andres Freund <andres@2ndquadrant.com> writes:
> I wonder if the most natural way to express this wouldn't be to have a
> heavyweight lock for every 2pc xact
> 'slot'. ResourceOwnerRelease(RESOURCE_RELEASE_LOCKS) should be scheduled
> correctly to make error handling for this work.
That seems like not a bad idea. Could we also use the same lock to
prevent concurrent attempts to commit/rollback the same already-prepared
transaction? I forget what we're doing to forestall such cases right now.
regards, tom lane
On 2014-04-14 13:47:35 -0400, Tom Lane wrote: > Andres Freund <andres@2ndquadrant.com> writes: > > I wonder if the most natural way to express this wouldn't be to have a > > heavyweight lock for every 2pc xact > > 'slot'. ResourceOwnerRelease(RESOURCE_RELEASE_LOCKS) should be scheduled > > correctly to make error handling for this work. > > That seems like not a bad idea. Could we also use the same lock to > prevent concurrent attempts to commit/rollback the same already-prepared > transaction? I forget what we're doing to forestall such cases right now. GlobalTransaction->locking_xid is currently used. If it points to a live transaction by another backned "prepared transaction with identifier \"%s\" is busy" will be thrown. ISTM if there were using a lock for every slot, that logic couldbe thrown away. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 04/14/2014 07:51 PM, Tom Lane wrote: > I'd prefer to leave the prepare sequence alone and instead find a way > to reject COMMIT PREPARED until after the source transaction is safely > clear of the race conditions. The upthread idea of looking at vxid > instead of xid might help, except that I see we clear both of them > in ProcArrayClearTransaction. We'd need some state in PGPROC that > isn't cleared till later than that. Hmm. What if one of the post-cleanup action fails? We can't bail out of the prepare sequence until we have transfered the locks to the new PGPROC. Otherwise the locks are lost. In essence, there should be a critical section from the EndPrepare call until all the critical cleanup actions like PostPrepare_Locks have been done, and I don't think we want that. We might be able to guarantee that the built-in post-cleanup operations are safe enough for that, but there's also CallXactCallbacks in there. Given the lack of reports of that happening, though, perhaps that's not an issue. - Heikki
On 04/14/2014 09:48 PM, Heikki Linnakangas wrote: > On 04/14/2014 07:51 PM, Tom Lane wrote: >> I'd prefer to leave the prepare sequence alone and instead find a way >> to reject COMMIT PREPARED until after the source transaction is safely >> clear of the race conditions. The upthread idea of looking at vxid >> instead of xid might help, except that I see we clear both of them >> in ProcArrayClearTransaction. We'd need some state in PGPROC that >> isn't cleared till later than that. > > Hmm. What if one of the post-cleanup action fails? We can't bail out of > the prepare sequence until we have transfered the locks to the new > PGPROC. Otherwise the locks are lost. In essence, there should be a > critical section from the EndPrepare call until all the critical cleanup > actions like PostPrepare_Locks have been done, and I don't think we want > that. We might be able to guarantee that the built-in post-cleanup > operations are safe enough for that, but there's also CallXactCallbacks > in there. > > Given the lack of reports of that happening, though, perhaps that's not > an issue. I came up with the attached fix for this. Currently, the entry is implicitly considered dead or unlocked if the locking_xid transaction is no longer active, but this patch essentially turns locking_xid into a simple boolean, and makes it the backend's responsibility to clear it on abort. (it's not actually a boolean, it's a BackendId, but that's just for debugging purposes to track who's keeping the entry locked). This requires a process exit hook, and an abort hook, to make sure the entry is always released, but that's not too difficult. It allows the backend to release the entry at exactly the right time, instead of having it implicitly released by ProcArrayClearTransaction. If we error during prepare, after having written the prepare WAL record but before the locks have been transfered to the dummy PGPROC, the locks are simply released. This is wrong, but it's always been like that and we haven't heard any complaints of that from the field, so I'm inclined to leave it as it is. We could use a critical section to force a panic, but that cure could be a worse than the disease. I considered Andres' idea of using a new heavy-weight lock, but didn't like it much. It would be a larger patch, which is not nice for back-patching. One issue would be that if you run out of lock memory, you could not roll back any prepared transactions, which is not nice because it could be a prepared transaction that's hoarding the lock memory. - Heikki
Вложения
Hi,
On 2014-05-05 13:41:00 +0300, Heikki Linnakangas wrote:
> I came up with the attached fix for this. Currently, the entry is implicitly
> considered dead or unlocked if the locking_xid transaction is no longer
> active, but this patch essentially turns locking_xid into a simple boolean,
> and makes it the backend's responsibility to clear it on abort. (it's not
> actually a boolean, it's a BackendId, but that's just for debugging purposes
> to track who's keeping the entry locked). This requires a process exit hook,
> and an abort hook, to make sure the entry is always released, but that's not
> too difficult. It allows the backend to release the entry at exactly the
> right time, instead of having it implicitly released by
> I considered Andres' idea of using a new heavy-weight lock, but didn't like
> it much. It would be a larger patch, which is not nice for back-patching.
> One issue would be that if you run out of lock memory, you could not roll
> back any prepared transactions, which is not nice because it could be a
> prepared transaction that's hoarding the lock memory.
I am not convinced by the latter reasoning but you're right that any
such change would hardly be backpatchable.
> +/*
> + * Exit hook to unlock the global transaction entry we're working on.
> + */
> +static void
> +AtProcExit_Twophase(int code, Datum arg)
> +{
> + /* same logic as abort */
> + AtAbort_Twophase();
> +}
> +
> +/*
> + * Abort hook to unlock the global transaction entry we're working on.
> + */
> +void
> +AtAbort_Twophase(void)
> +{
> + if (MyLockedGxact == NULL)
> + return;
> +
> + /*
> + * If we were in process of preparing the transaction, but haven't
> + * written the WAL record yet, remove the global transaction entry.
> + * Same if we are in the process of finishing an already-prepared
> + * transaction, and fail after having already written the WAL 2nd
> + * phase commit or rollback record.
> + *
> + * After that it's too late to abort, so just unlock the GlobalTransaction
> + * entry. We might not have transfered all locks and other state to the
> + * prepared transaction yet, so this is a bit bogus, but it's the best we
> + * can do.
> + */
> + if (!MyLockedGxact->valid)
> + {
> + RemoveGXact(MyLockedGxact);
> + }
> + else
> + {
> + LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE);
> +
> + MyLockedGxact->locking_backend = InvalidBackendId;
> +
> + LWLockRelease(TwoPhaseStateLock);
> + }
> + MyLockedGxact = NULL;
> +}
Is it guaranteed that all paths have called LWLockReleaseAll()
before calling the proc exit hooks? Otherwise we might end up waiting
for ourselves...
> /*
> * MarkAsPreparing
> @@ -261,29 +329,15 @@ MarkAsPreparing(TransactionId xid, const char *gid,
> errmsg("prepared transactions are disabled"),
> errhint("Set max_prepared_transactions to a nonzero value.")));
>
> - LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE);
> -
> - /*
> - * First, find and recycle any gxacts that failed during prepare. We do
> - * this partly to ensure we don't mistakenly say their GIDs are still
> - * reserved, and partly so we don't fail on out-of-slots unnecessarily.
> - */
> - for (i = 0; i < TwoPhaseState->numPrepXacts; i++)
> + /* on first call, register the exit hook */
> + if (!twophaseExitRegistered)
> {
> - gxact = TwoPhaseState->prepXacts[i];
> - if (!gxact->valid && !TransactionIdIsActive(gxact->locking_xid))
> - {
> - /* It's dead Jim ... remove from the active array */
> - TwoPhaseState->numPrepXacts--;
> - TwoPhaseState->prepXacts[i] = TwoPhaseState->prepXacts[TwoPhaseState->numPrepXacts];
> - /* and put it back in the freelist */
> - gxact->next = TwoPhaseState->freeGXacts;
> - TwoPhaseState->freeGXacts = gxact;
> - /* Back up index count too, so we don't miss scanning one */
> - i--;
> - }
> + before_shmem_exit(AtProcExit_Twophase, 0);
> + twophaseExitRegistered = true;
> }
It's not particularly nice to register shmem exit hooks in the middle of
normal processing because it makes it impossible to use
cancel_before_shmem_exit() previously registered hooks. I think this
should be registered at startup, if max_prepared_xacts > 0.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services
On 05/06/2014 02:44 PM, Andres Freund wrote:
> On 2014-05-05 13:41:00 +0300, Heikki Linnakangas wrote:
>> +/*
>> + * Exit hook to unlock the global transaction entry we're working on.
>> + */
>> +static void
>> +AtProcExit_Twophase(int code, Datum arg)
>> +{
>> + /* same logic as abort */
>> + AtAbort_Twophase();
>> +}
>> +
>> +/*
>> + * Abort hook to unlock the global transaction entry we're working on.
>> + */
>> +void
>> +AtAbort_Twophase(void)
>> +{
>> + if (MyLockedGxact == NULL)
>> + return;
>> +
>> + /*
>> + * If we were in process of preparing the transaction, but haven't
>> + * written the WAL record yet, remove the global transaction entry.
>> + * Same if we are in the process of finishing an already-prepared
>> + * transaction, and fail after having already written the WAL 2nd
>> + * phase commit or rollback record.
>> + *
>> + * After that it's too late to abort, so just unlock the GlobalTransaction
>> + * entry. We might not have transfered all locks and other state to the
>> + * prepared transaction yet, so this is a bit bogus, but it's the best we
>> + * can do.
>> + */
>> + if (!MyLockedGxact->valid)
>> + {
>> + RemoveGXact(MyLockedGxact);
>> + }
>> + else
>> + {
>> + LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE);
>> +
>> + MyLockedGxact->locking_backend = InvalidBackendId;
>> +
>> + LWLockRelease(TwoPhaseStateLock);
>> + }
>> + MyLockedGxact = NULL;
>> +}
>
> Is it guaranteed that all paths have called LWLockReleaseAll()
> before calling the proc exit hooks? Otherwise we might end up waiting
> for ourselves...
Hmm. AbortTransaction() will release locks before we get here, but the
before_shmem_exit() callpath will not. So an elog(FATAL), while holding
TwoPhaseStateLock would cause us to deadlock with ourself. But there are
no such elogs.
I copied this design from async.c, which is quite similar, so if there's
a problem that ought to be fixed too. And there are other more
complicated before_shmem callbacks that worry me more, like
createdb_failure_callback(). But I think they're all all right.
>> /*
>> * MarkAsPreparing
>> @@ -261,29 +329,15 @@ MarkAsPreparing(TransactionId xid, const char *gid,
>> errmsg("prepared transactions are disabled"),
>> errhint("Set max_prepared_transactions to a nonzero value.")));
>>
>> - LWLockAcquire(TwoPhaseStateLock, LW_EXCLUSIVE);
>> -
>> - /*
>> - * First, find and recycle any gxacts that failed during prepare. We do
>> - * this partly to ensure we don't mistakenly say their GIDs are still
>> - * reserved, and partly so we don't fail on out-of-slots unnecessarily.
>> - */
>> - for (i = 0; i < TwoPhaseState->numPrepXacts; i++)
>> + /* on first call, register the exit hook */
>> + if (!twophaseExitRegistered)
>> {
>> - gxact = TwoPhaseState->prepXacts[i];
>> - if (!gxact->valid && !TransactionIdIsActive(gxact->locking_xid))
>> - {
>> - /* It's dead Jim ... remove from the active array */
>> - TwoPhaseState->numPrepXacts--;
>> - TwoPhaseState->prepXacts[i] = TwoPhaseState->prepXacts[TwoPhaseState->numPrepXacts];
>> - /* and put it back in the freelist */
>> - gxact->next = TwoPhaseState->freeGXacts;
>> - TwoPhaseState->freeGXacts = gxact;
>> - /* Back up index count too, so we don't miss scanning one */
>> - i--;
>> - }
>> + before_shmem_exit(AtProcExit_Twophase, 0);
>> + twophaseExitRegistered = true;
>> }
>
> It's not particularly nice to register shmem exit hooks in the middle of
> normal processing because it makes it impossible to use
> cancel_before_shmem_exit() previously registered hooks. I think this
> should be registered at startup, if max_prepared_xacts > 0.
<shrug>. async.c and namespace.c does the same, and it hasn't been a
problem.
I committed this now, but please let me know if you see a concrete
problem with the locks.
- Heikki
On 2014-05-15 17:21:28 +0300, Heikki Linnakangas wrote: > >Is it guaranteed that all paths have called LWLockReleaseAll() > >before calling the proc exit hooks? Otherwise we might end up waiting > >for ourselves... > > Hmm. AbortTransaction() will release locks before we get here, but the > before_shmem_exit() callpath will not. So an elog(FATAL), while holding > TwoPhaseStateLock would cause us to deadlock with ourself. But there are no > such elogs. > I copied this design from async.c, which is quite similar, so if there's a > problem that ought to be fixed too. And there are other more complicated > before_shmem callbacks that worry me more, like createdb_failure_callback(). > But I think they're all all right. Perhaps we should enforce that LWLockReleaseAll() is called first? E.g. in shmem_exit()? It'll happen in ProcKill() atm, but that's normally pretty much at the bottom of the stack. > >It's not particularly nice to register shmem exit hooks in the middle of > >normal processing because it makes it impossible to use > >cancel_before_shmem_exit() previously registered hooks. I think this > >should be registered at startup, if max_prepared_xacts > 0. > > <shrug>. async.c and namespace.c does the same, and it hasn't been a > problem. Well, it doesn't seem unreasonable to have C code using PG_ENSURE_ERROR_CLEANUP/PG_END_ENSURE_ERROR_CLEANUP around a 2pc commit to me. That'll break with this. Perhaps we should just finally make cancel_before_shmem_exit search the stack of callbacks. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 04/14/2014 11:55 AM, Marko Kreen wrote: > On Sun, Apr 13, 2014 at 05:46:20PM -0400, Jan Wieck wrote: >> On 04/13/14 14:22, Jan Wieck wrote: >>> On 04/13/14 08:27, Marko Kreen wrote: >>>> I think you need to do SET_VARSIZE also here. Alternative is to >>>> move SET_VARSIZE after sort_snapshot(). >>>> >>>> And it seems the drop-double-txid logic should be added also to >>>> txid_snapshot_recv(). It seems weird to have it behave differently >>> >from txid_snapshot_in(). >>>> >>> >>> Thanks, >>> >>> yes on both issues. Will create another patch. >> >> New patch attached. >> >> New github commit is https://github.com/wieck/postgres/commit/b8fd0d2eb78791e5171e34aecd233fd06218890d > > Looks OK to me. Ok, committed. - Heikki
On Thu, May 15, 2014 at 10:38 AM, Andres Freund <andres@2ndquadrant.com> wrote: >> <shrug>. async.c and namespace.c does the same, and it hasn't been a >> problem. > > Well, it doesn't seem unreasonable to have C code using > PG_ENSURE_ERROR_CLEANUP/PG_END_ENSURE_ERROR_CLEANUP around a 2pc commit > to me. That'll break with this. > Perhaps we should just finally make cancel_before_shmem_exit search the > stack of callbacks. Yes, please. And while we're at it, perhaps we should make it Trap() or fail an Assert() if it doesn't find the callback it was told to remove. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company