There are cases where during execution a Memory Intensive Operator (MI)
may not use all the memory that is allocated to it. This means that this
extra memory (quota - allocated) can be relinquished for other MI nodes
to use during execution of a statement. For example

->  Hash Join
         ->  HashAggregate
         ->  Hash
In the above query fragment the HashJoin operator has a MI operator for
both its inner and outer subtree. If there ever is the case that the
Hash node used much less memory than was given as its quota it will now
call MemoryAccounting_DeclareDone() and the difference between its
quota and allocated amount will be added to the allocated amount of the
RelinquishedPool. Doing this will enable HashAggregate to request memory
from this RelinquishedPool if it exhausts its quota to prevent spilling.

This PR adds two new API's to the MemoryAccounting Framework

MemoryAccounting_DeclareDone(): Add the difference between a memory
account's quota and its allocated amount to the long living
RelinquishedPool

MemoryAccounting_RequestQuotaIncrease(): Retrieve all relinquished
memory by incrementing an operator's operatorMemKb and setting the
RelinquishedPool to 0

Note: This PR introduces the facility for Hash to relinquish memory to
the RelinquishedPool memory account and for the Agg operator
(specifically HashAgg) to request an increase to its quota before it
builds its hash table. This commit does not generally apply this
paradigm to all MI operators
Signed-off-by: NSambitesh Dash <sdash@pivotal.io>
Signed-off-by: NMelanie Plageman <mplageman@pivotal.io>

Before this commit all memory allocations made by ORCA/GPOS were a
blackbox to GPDB. However the ground work had been in place to allow
GPDB's Memory Accounting Framework to track memory consumption by ORCA.
This commit introduces two new functions
Ext_OptimizerAlloc and Ext_OptimizerFree which
pass through their parameters to gp_malloc and gp_free and do some bookeeping
against the Optimizer Memory Account. This introduces very little
overhead to the GPOS memory management framework.
Signed-off-by: NMelanie Plageman <mplageman@pivotal.io>
Signed-off-by: NSambitesh Dash <sdash@pivotal.io>

* resgroup: provide helper funcs for memory usage updates.

We used to have complex and duplicate logic to update group & slot
memory usage under different context, now we provide two helper
functions to increase or decrease memory usage in group and slot.

Two bad named functions `attachToSlot()` and `detachFromSlot()` are
retired now.

* resgroup: provide helper function to unassign a dropped resgroup.

* resgroup: move complex checks into helper functions.

Many helper functions were added with descriptive names to increase
readability of lots of complex checks.

Also added a pointer to resource group slot in self.

* resgroup: add helper functions for wait queue operations.

The intend of this extra configuration file is to control the
synchronization between primary and mirror for WALREP.

The gp_replication.conf is not designed to work with filerep, for
example, the scripts like gp_expand will fail since it directly modify
the configuration files instead of going through initdb.
Signed-off-by: NXin Zhang <xzhang@pivotal.io>

We already had a hack for the EXECUTE ON ALL SEGMENTS case, by setting
prodataaccess='s'. This exposes the functionality to users via DDL, and adds
support for the EXECUTE ON MASTER case.

There was discussion on gpdb-dev about also supporting ON MASTER AND ALL
SEGMENTS, but that is not implemented yet. There is no handy "locus" in the
planner to represent that. There was also discussion about making a
gp_segment_id column implicitly available for functions, but that is also
not implemented yet.

The old behavior was that a function that if a function was marked as
IMMUTABLE, it could be executed anywhere. Otherwise it was always executed
on the master. For backwards-compatibility, this keeps that behavior for
EXECUTE ON ANY (the default), so even if a function is marked as EXECUTE ON
ANY, it will always be executed on the master unless it's IMMUTABLE.

There is no support for these new options in ORCA. Using any ON MASTER or
ON ALL SEGMENTS functions in a query cause ORCA to fall back. This is the
same as with the prodataaccess='s' hack that this replaces, but now that it
is more user-visible, it would be nice to teach ORCA about it.

The new options are only supported for set-returning functions, because for
a regular function marked as EXECUTE ON ALL SEGMENTS, it's not clear how
the results should be combined. ON MASTER would probably be doable, but
there's no need for that right now, so punt.

Another restriction is that a function with ON ALL SEGMENTS or ON MASTER can
only be used in the FROM clause, or in the target list of a simple SELECT
with no FROM clause. So "SELECT func()" is accepted, but "SELECT func() FROM
foo" is not. "SELECT * FROM func(), foo" works, however. EXECUTE ON ANY
functions, which is the default, work the same as before.

In this commit, we add more detailed memory metrics to the 'memory_usage'
column of gp_resgroup_status include current/available memory usage in
a group, current/available memory usage for a slot, current/available
memory usage for the shared part.

Previously, waiters waiting on a dropped resource group need to be
reassigned to a new group, to achieve it, ResGroupSlotAcquire is
modified to be complicated and not easy to understand, this commit
refines it.

Author: Gang Xiong <gxiong@pivotal.io>

Allow CREATE RESOURCE GROUP and ALTER RESOURCE GROUP to set concurrency
to 0, so there will eventually be no running queries after some time, so
the resource group can be dropped. On drop all pending queries will be
moved to the new resource group assigned to the role; but if the role is
also dropped the pending queries will all be canceled. Another thing is
we do not allow setting concurrency of admin group to zero, superuser is
under admin group and only superuser can alter resource group, so once
concurrency of admin group is set to zero, there will be no chance to set
it again.
Signed-off-by: NNing Yu <nyu@pivotal.io>

This commit does two changes:
1. Remove the restriction that sum of memory_spill_ratio and memory_shared_quota
must be no larger than 100.
2. Change the range of memory_spill_ratio to be [0, 100].

Function FaultInjectorIdentifierStringToEnum(faultName) pass a const
string to a non-const parameter, which cause a build warnig. But on the
second thought, we have supported injecting fault by fault name without
corresponding fault identifier, so it's better to use faultname instead
of fault enum identifier in the ereport.

Signed-off-by: NAbhijit Subramanya <asubramanya@pivotal.io>

New API: void set_gp_replication_config(const char *name, const char *value)

This function is inspired by the upstream ALTER SYSTEM command
AlterSystemSetConfigFile() from commit
7dfab04a.

Once we merged the upstream changes, we can remove this function and
directly use the AlterSystemSetConfigFile().
Signed-off-by: NAbhijit Subramanya <asubramanya@pivotal.io>

We use this file to store the GUC value `sychronous_standby_names` in
order to control the blocking behavior between primary and mirrors as
used in upstream. When this GUC is on, the primary is blocked and
waiting for the commits propagated to mirrors regardless of mirror
status. When this GUC is off, primary just archive and won't wait for
mirrors.

gp_replication.conf is now read unconditionally by the GUC parsing logic
and needs to be set up by initdb. Refactor set_null_conf() to take a
filename so that we don't copy-paste more code.
Signed-off-by: NJacob Champion <pchampion@pivotal.io>

Using fault name instead of enum as the key of fault hash table

GPDB fault injector uses fault enum as the key of fault hash table.
If someone wants to inject fault into gpdb extensions(a separate repo),
she has to hard code the extension related fault enums into gpdb core
code, this is not a good practice.
So we simply use fault name as the hash key to remove the need of hard
code the fault enum. Note that fault injector API doesn't change.

In GPDB, we have so far used a WindowFrame struct to represent the start
and end window bound, in a ROWS/RANGE BETWEEN clause, while PostgreSQL
uses the combination of  a frameOptions bitmask and start and end
expressions. Refactor to replace the WindowFrame with the upstream
representation.

Historically this function was used to special case a few operators that
were not considered to be MemoryIntensive. However, now it always
returns true. This commit removes the function and also moves the case
for T_FunctionScan in IsMemoryIntensiveOperator into the group that
always returns true, as this is its current behavior

The bzip2 library is only used by the gfile/fstream code, used for external
tables and gpfdist. The usage of bzip2 was in #ifndef WIN32 blocks, so it
was only built on non-Windows systems.

Instead of tying it to the platform, use a proper autoconf check and
HAVE_LIBBZ2 flags. This makes it possible to build gpfdist with bzip2
support on Windows, as well as building without bzip2 on non-Windows
systems. That makes it easier to test the otherwise Windows-only codepaths
on other platforms. --with-libbz2 is still the default, but you can now use
--without-libbz2 if you wish.

I'm sure that some regression tests will fail if you actually build the
server without libbz2, but I'm not going to address that right now. We have
similar problems with other features that are in principle optional, but
cause some regression tests to fail.

Also use "#ifdef HAVE_LIBZ" rather than "#ifndef WIN32" to enable/disable
zlib support in gpfdist. Building the server still fails if you use
--without-zlib, but at least you can build the client programs without
zlib, also on non-Windows systems.

Remove obsolete copy of bzlib.h from the repository while we're at it.

Using gp_segment_configuration catalog table easily can find if mirrors exist or
not, do not need special table to communicate the same. Earlier
gp_fault_strategy used to convey 'n' for mirrorless system, 'f' for replication
and 's' for san mirrors. Since support for 's' was removed in 5.0 only purpose
gp_fault_strategy served was mirrored or not mirrored system. Hence deleting the
gp_fault_strategy table and at required places using gp_segment_configuration to
find the required info.

Although I'm not too familiar with SystemTap, I'm pretty sure that recent
versions can do user space tracing better. I don't think anyone is using
these hacks anymore, so remove them.

* resgroup: move MyResGroupSharedInfo into MyResGroupProcInfo.

MyResGroupSharedInfo is now replaced with MyResGroupProcInfo->group.

* resgroup: retire resGranted in PGPROC.

when resGranted == false we must have resSlotId == InvalidSlotId,
when resGranted != false we must have resSlotId != InvalidSlotId,
so we can retire resGranted and keep only resSlotId.

* resgroup: rename sharedInfo to group.

in resgroup.c there used to be both `group` and `sharedInfo` for the
same thing, now only use `group`.

* resgroup: rename MyResGroupProcInfo to self.

We want to use this variable directly so a short name is better.

This is the spec-compliant spelling, but GPDB has only allowed
"agg OVER (window)" so far. With this commit, the parens are still allowed,
for backwards-compatibility.

Change deparsing code to also use the non-parens syntax in view definitions
and EXPLAIN. Adjust expected output of regression tests accordingly.

It was noticed in CentOS 6 that the entries in the char* list would
become blank after freeing the raw string that is obtained from
setting the GUC wal_consistency_checking. We should free this string
only after we finish using it.

Authors: Jimmy Yih and Taylor Vesely

The view gp_toolkit.gp_resgroup_status collect resgroup status
information on both QD and QEs, also before and after a 300ms delay to
measure the cpu usage. When there is concurrently resgroup drops it
might fail due to missing resgroups.

This is fixed by holding ExclusiveLock which blocks the drop operations.
Signed-off-by: NNing Yu <nyu@pivotal.io>

Previously in ResGroupSlotRelease(), we mark the resWaiting flag to
false out of the ResGroupLock, a race condition can happen as that
the waitProc that meant to be wake up is canceled just before it's
resWaiting is set to false and in the window resWaiting is set, the
waitProc run another query and wait for a free slot again, then
resWaiting is set and the waitProc will get another unexpected free slot.

eg: rg1 has concurrency = 1

s1: BEGIN;
s2: BEGIN; -- blocked
s1: END; -- set breakpoint just before resWaiting is set in ResGroupSlotRelease()
s2: cancel it;
s3: BEGIN; -- will get a free slot.
s2: BEGIN; -- run begin again and be blocked.
s1: continue;
s2: s2 will got unexpected free slot

To avoid leak of resgroup slot, we set resWaiting under ResGroupLock so the
waitProc have no window to re-fetch the slot before resWaiting is set.
It's hard to add a regression test, adding breakpoint in ResGroupSlotRelease
may block all other operations because it holds the ResGroupLock.

tqual.c is quite a long file, and there are plenty of legitimate
MPP-specific changes in it compared to upstream. Move these functions to
separate file, to reduce our diff footprint. This hopefully makes merging
easier in the future.

Run pgindent over the new file, and do some manual tidying up. Also, use
CStringGetTextDatum() instead of the more complicated dance with
DirectFunctionCall and textin(), to convert C strings into text Datums.

In the upstream, two different structs are used to represent a window
definition. WindowDef in the grammar, which is transformed into
WindowClause during parse analysis. In GPDB, we've been using the same
struct, WindowSpec, in both stages. Split it up, to match the upstream.

The representation of the window frame, i.e. "ROWS/RANGE BETWEEN ..." was
different between the upstream implementation and the GPDB one. We now use
the upstream frameOptions+startOffset+endOffset representation in raw
WindowDef parse node, but it's still converted to the WindowFrame
representation for the later stages, so WindowClause still uses that. I
will switch over the rest of the codebase to the upstream representation as
a separate patch.

Also, refactor WINDOW clause deparsing to be closer to upstream.

One notable difference is that the old WindowSpec.winspec field corresponds
to the winref field in WindowDef andWindowClause, except that the new
'winref' is 1-based, while the old field was 0-based.

Another noteworthy thing is that this forbids specifying "OVER (w
ROWS/RANGE BETWEEN ...", if the window "w" already specified a window frame,
i.e. a different ROWS/RANGE BETWEEN. There was one such case in the
regression suite, in window_views, and this updates the expected output of
that to be an error.

This adds the 'winstar' field from the upstream. Also bring in the 'winagg'
field while we're at it, although it's only used for an assertion in
nodeWindow.c so far.

FaultInjector_UpdateHashEntry() was using FaultInjector_InsertHashEntry(), which
ends-up adding entry if not present without incrementing
`faultInjectorShmem->faultInjectorSlots`. This causes inconsistency, plus also
sometimes encounters "FailedAssertion(""!(faultInjectorShmem->faultInjectorSlots
== 0)""," during fault inject reset, as goes negative. Fixing the same by using
FaultInjector_LookupHashEntry() instead as that's what
FaultInjector_UpdateHashEntry() needs.

Scenario the Assertion was hitting:
gpfaultinjector -f all -m async -y resume -r primary -H ALL
gpfaultinjector -f all -m async -y reset -r primary -H ALL

Use the available macro for detecting FTP protocol and use pstrdup
instead of palloc+strncpy. Also fix a spacing issue with an ereport
call.

These functions as inline functions were producing warnings, based on discussion
https://groups.google.com/a/greenplum.org/d/msg/gpdb-dev/6fgKvN9QpV4/zjysjqIZAgAJ
converting them to macro as upstream. Adding explicit type casting wherever
needed now that DatumGetPointer() returns (char *) instead of (void *).

It hasn't been implemented, but there is basic support in the grammar,
just enough to detect the syntax and throw an error or ignore it. All the
rest was dead code.

* In ruleutils.c, the ereport() was broken. Use elog() instead, like in
  the upstream. (elog() is fine for "can't happen" kind of sanity checks)

* Remove a few unused local variables.

* Add a missing cast from Plan * to Node *.

We will be less conservative and enable by default recursive CTE on
master, while keeping recursive CTE hidden as we progress on developing
the feature.

This reverts the following two commits:
* 280c577a "Set gp_recursive_cte_prototype GUC to true in test"
* 4d5f8087 "Guard Recursive CTE behind a GUC"
Signed-off-by: NHaisheng Yuan <hyuan@pivotal.io>
Signed-off-by: NJesse Zhang <sbjesse@gmail.com>

Previously we were setting the value of
`optimizer_join_arity_for_associativity_commutativity` to very large
number and so ORCA would spend a very long time evaluating all possible n_way_join
combinations to come up with the cheapest join tree to use in the plan.

We are reducing this value to `7` as it does not prove to be beneficial
to spend time and resources to evaluate any more than 7_way_joins in
trying to find the cheapest join tree.

Plus minor corrections in spelling and comments.
Signed-off-by: NSam Dash <sdash@pivotal.io>

While Recurisve CTE is still being developed it will be hidden from users by
the guc gp_recursive_cte_prototype
Signed-off-by: NSambitesh Dash <sdash@pivotal.io>

get_name_for_var_field didn't have enough context to interpret a reference to
a CTE query's output.  Fixing this requires separate hacks for the regular
deparse case (pg_get_ruledef) and for the EXPLAIN case, since the available
context information is quite different.  It's pretty nearly parallel to the
existing code for SUBQUERY RTEs, though.  Also, add code to make sure we
qualify a relation name that matches a CTE name; else the CTE will mistakenly
capture the reference when reloading the rule.

In passing, fix a pre-existing problem with get_name_for_var_field not working
on variables in targetlists of SubqueryScan plan nodes.  Although latent all
along, this wasn't a problem until we made EXPLAIN VERBOSE try to print
targetlists.  To do this, refactor the deparse_context_for_plan API so that
the special case for SubqueryScan is all on ruleutils.c's side.

(cherry picked from commit 742fd06d)

Planner generates plan that doesn't insert any motion between WorkTableScan and
its corresponding RecursiveUnion, because currently in GPDB motions are not
rescannable. For example, a MPP plan for recursive CTE query may look like:
```
Gather Motion 3:1
   ->  Recursive Union
         ->  Seq Scan on department
               Filter: name = 'A'::text
         ->  Nested Loop
               Join Filter: d.parent_department = sd.id
               ->  WorkTable Scan on subdepartment sd
               ->  Materialize
                     ->  Broadcast Motion 3:3
                           ->  Seq Scan on department d
```

For the current solution, the WorkTableScan is always put on the outer side of
the top most Join (the recursive part of RecusiveUnion), so that we can safely
rescan the inner child of join without worrying about the materialization of a
potential underlying motion. This is a heuristic based plan, not a cost based
plan.

Ideally, the WorkTableScan can be placed on either side of the join with any
depth, and the plan should be chosen based on the cost of the recursive plan
and the number of recursions. But we will leave it for later work.

Note: The hash join is temporarily disabled for plan generation of recursive
part, because if the hash table spills, the batch file is going to be removed
as it executes. We have a following story to enable spilled hash table to be
rescannable.

See discussion at gpdb-dev mailing list:
https://groups.google.com/a/greenplum.org/forum/#!topic/gpdb-dev/s_SoXKlwd6I

If a prepared statement, or a cached plan for an SPI query e.g. from a
PL/pgSQL function, contains stable functions, the stable functions were
incorrectly evaluated only once at plan time, instead of on every execution
of the plan. This happened to not be a problem in queries that contain any
parameters, because in GPDB, they are re-planned on every invocation
anyway, but non-parameter queries were broken.

In the planner, before this commit, when simplifying expressions, we set
the transform_stable_funcs flag to true for every query, and evaluated all
stable functions at planning time. Change it to false, and also rename it
back to 'estimate', as it's called in the upstream. That flag was changed
back in 2010, in order to allow partition pruning to work with qual
containing stable functions, like TO_DATE. I think back then, we always
re-planned every query, so that was OK, but we do cache plans now.

To avoid regressing to worse plans, change eval_const_expressions() so that
it still does evaluate stable functions, even when the 'estimate' flag is
off. But when it does so, mark the plan as "one-off", meaning that it must
be re-planned on every execution. That gives the old, intended, behavior,
that such plans are indeed re-planned, but it still allows plans that don't
use stable functions to be cached.

This seems to fix github issue #2661. Looking at the direct dispatch code
in apply_motion(), I suspect there are more issues like this lurking there.
There's a call to planner_make_plan_constant(), modifying the target list
in place, and that happens during planning. But this at least fixes the
non-direct dispatch cases, and is a necessary step for fixing any remaining
issues.

For some reason, the query now gets planned *twice* for every invocation.
That's not ideal, but it was an existing issue for prepared statements with
parameters, already. So let's deal with that separately.