Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Em sáb., 30 de dez. de 2023 19:19, Tomas Vondra <
tomas.von...@enterprisedb.com> escreveu:

> On 12/29/23 18:02, Ranier Vilela wrote:
> >
> > ...
> >
> > As I wrote, the new patch version was for consideration.
> > It seems more like a question of style, so it's better to remove it.
> >
> > Anyway +1 for your original patch.
> >
>
> I've pushed my original patch. Thanks for the report.
>
Thank you.

Best regards,
Ranier Vilela

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Tomas Vondra]

On 12/29/23 18:02, Ranier Vilela wrote:
>
> ...
> 
> As I wrote, the new patch version was for consideration.
> It seems more like a question of style, so it's better to remove it.
> 
> Anyway +1 for your original patch.
> 

I've pushed my original patch. Thanks for the report.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Em sex., 29 de dez. de 2023 às 08:53, Ranier Vilela 
escreveu:

> Em qui., 28 de dez. de 2023 às 22:16, Tomas Vondra <
> tomas.von...@enterprisedb.com> escreveu:
>
>>
>>
>> On 12/27/23 12:37, Ranier Vilela wrote:
>> > Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
>> > mailto:tomas.von...@enterprisedb.com>>
>> > escreveu:
>> >
>> >
>> >
>> > On 12/26/23 19:10, Ranier Vilela wrote:
>> > > Hi,
>> > >
>> > > The commit b437571
>> > > > > I
>> > > think has an oversight.
>> > > When allocate memory and initialize private spool in function:
>> > > _brin_leader_participate_as_worker
>> > >
>> > > The behavior is the bs_spool (heap and index fields)
>> > > are left empty.
>> > >
>> > > The code affected is:
>> > >   buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
>> > > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
>> > > - buildstate->bs_spool->index = buildstate->bs_spool->index;
>> > > + buildstate->bs_spool->heap = heap;
>> > > + buildstate->bs_spool->index = index;
>> > >
>> > > Is the fix correct?
>> > >
>> >
>> > Thanks for noticing this.
>> >
>> > You're welcome.
>> >
>> >
>> > Yes, I believe this is a bug - the assignments
>> > are certainly wrong, it leaves the fields set to NULL.
>> >
>> > I wonder how come this didn't fail during testing. Surely, if the
>> leader
>> > participates as a worker, the tuplesort_begin_index_brin shall be
>> called
>> > with heap/index being NULL, leading to some failure during the
>> sort. But
>> > maybe this means we don't actually need the heap/index fields, it's
>> just
>> > a copy of TuplesortIndexArg, but BRIN does not need that because we
>> sort
>> > the tuples by blkno, and we don't need the descriptors for that.
>> >
>> > Unfortunately I can't test on Windows, since I can't build with meson on
>> > Windows.
>> >
>> >
>> > In any case, the _brin_parallel_scan_and_build does not actually
>> need
>> > the separate heap/index arguments, those are already in the spool.
>> >
>> > Yeah, for sure.
>> >
>> >
>> > I'll try to figure out if we want to simplify the tuplesort or
>> remove
>> > the arguments from _brin_parallel_scan_and_build.
>> >
>>
>> Here is a patch simplifying the BRIN parallel create code a little bit.
>> As I suspected, we don't need the heap/index in the spool at all, and we
>> don't need to pass it to tuplesort_begin_index_brin either - we only
>> need blkno, and we have that in the datum1 field. This also means we
>> don't need TuplesortIndexBrinArg.
>>
> With Windows 10, msvc 2022, compile end pass ninja test.
>
> But, if you allow me, I would like to try another approach to
> simplification.
> Instead of increasing the arguments in the call, wouldn't it be better to
> decrease them
> and this way all arguments will be passed in the registers instead of on a
> stack?
>
> bs_spool may well contain this data and will probably be useful in the
> future.
>
> I made a v1 version, based on your patch, for your consideration.
>
As I wrote, the new patch version was for consideration.
It seems more like a question of style, so it's better to remove it.

Anyway +1 for your original patch.

Best regards,
Ranier Vilela

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Tomas Vondra]

On 12/29/23 14:53, Ranier Vilela wrote:
> 
> 
> Em sex., 29 de dez. de 2023 às 10:33, Tomas Vondra
> mailto:tomas.von...@enterprisedb.com>>
> escreveu:
> 
> 
> 
> On 12/29/23 12:53, Ranier Vilela wrote:
> > Em qui., 28 de dez. de 2023 às 22:16, Tomas Vondra
> >  
>  >>
> > escreveu:
> >
> >
> >
> >     On 12/27/23 12:37, Ranier Vilela wrote:
> >     > Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
> >     >  
> >      >
> >      
> >       >     > escreveu:
> >     >
> >     >
> >     >
> >     >     On 12/26/23 19:10, Ranier Vilela wrote:
> >     >     > Hi,
> >     >     >
> >     >     > The commit b437571
> >     >      
> >      >
> >     >      
> >       >     >     > think has an oversight.
> >     >     > When allocate memory and initialize private spool in
> function:
> >     >     > _brin_leader_participate_as_worker
> >     >     >
> >     >     > The behavior is the bs_spool (heap and index fields)
> >     >     > are left empty.
> >     >     >
> >     >     > The code affected is:
> >     >     >   buildstate->bs_spool = (BrinSpool *)
> >     palloc0(sizeof(BrinSpool));
> >     >     > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> >     >     > - buildstate->bs_spool->index =
> buildstate->bs_spool->index;
> >     >     > + buildstate->bs_spool->heap = heap;
> >     >     > + buildstate->bs_spool->index = index;
> >     >     >
> >     >     > Is the fix correct?
> >     >     >
> >     >
> >     >     Thanks for noticing this.
> >     >
> >     > You're welcome.
> >     >  
> >     >
> >     >     Yes, I believe this is a bug - the assignments
> >     >     are certainly wrong, it leaves the fields set to NULL.
> >     >
> >     >     I wonder how come this didn't fail during testing.
> Surely, if
> >     the leader
> >     >     participates as a worker, the tuplesort_begin_index_brin
> shall
> >     be called
> >     >     with heap/index being NULL, leading to some failure
> during the
> >     sort. But
> >     >     maybe this means we don't actually need the heap/index
> fields,
> >     it's just
> >     >     a copy of TuplesortIndexArg, but BRIN does not need that
> >     because we sort
> >     >     the tuples by blkno, and we don't need the descriptors
> for that.
> >     >
> >     > Unfortunately I can't test on Windows, since I can't build with
> >     meson on
> >     > Windows.
> >     >
> >     >
> >     >     In any case, the _brin_parallel_scan_and_build does not
> >     actually need
> >     >     the separate heap/index arguments, those are already in
> the spool.
> >     >
> >     > Yeah, for sure.
> >     >
> >     >
> >     >     I'll try to figure out if we want to simplify the
> tuplesort or
> >     remove
> >     >     the arguments from _brin_parallel_scan_and_build.
> >     >
> >
> >     Here is a patch simplifying the BRIN parallel create code a
> little bit.
> >     As I suspected, we don't need the heap/index in the spool at
> all, and we
> >     don't need to pass it to tuplesort_begin_index_brin either -
> we only
> >     need blkno, and we have that in the datum1 field. This also
> means we
> >     don't need TuplesortIndexBrinArg.
> >
> > With Windows 10, msvc 2022, compile end pass ninja test.
> >
> > But, if you allow me, I would like to try another approach to
> > simplification.
> > Instead of increasing the arguments in the call, wouldn't it be better
> > to decrease them 
> > and this way all arguments will be passed in the registers instead
> of on
> > a stack?
> >
> 
> If this was beneficial, we'd be passing everything through structs and
> not as explicit arguments. But we don't. If you're arguing it's
> beneficial in this case, it'd be good to see it demonstrated.
> 

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Em sex., 29 de dez. de 2023 às 10:33, Tomas Vondra <
tomas.von...@enterprisedb.com> escreveu:

>
>
> On 12/29/23 12:53, Ranier Vilela wrote:
> > Em qui., 28 de dez. de 2023 às 22:16, Tomas Vondra
> > mailto:tomas.von...@enterprisedb.com>>
> > escreveu:
> >
> >
> >
> > On 12/27/23 12:37, Ranier Vilela wrote:
> > > Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
> > >  > 
> >  > >>
> > > escreveu:
> > >
> > >
> > >
> > > On 12/26/23 19:10, Ranier Vilela wrote:
> > > > Hi,
> > > >
> > > > The commit b437571
> > >  > 
> > >  > >> I
> > > > think has an oversight.
> > > > When allocate memory and initialize private spool in
> function:
> > > > _brin_leader_participate_as_worker
> > > >
> > > > The behavior is the bs_spool (heap and index fields)
> > > > are left empty.
> > > >
> > > > The code affected is:
> > > >   buildstate->bs_spool = (BrinSpool *)
> > palloc0(sizeof(BrinSpool));
> > > > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> > > > - buildstate->bs_spool->index = buildstate->bs_spool->index;
> > > > + buildstate->bs_spool->heap = heap;
> > > > + buildstate->bs_spool->index = index;
> > > >
> > > > Is the fix correct?
> > > >
> > >
> > > Thanks for noticing this.
> > >
> > > You're welcome.
> > >
> > >
> > > Yes, I believe this is a bug - the assignments
> > > are certainly wrong, it leaves the fields set to NULL.
> > >
> > > I wonder how come this didn't fail during testing. Surely, if
> > the leader
> > > participates as a worker, the tuplesort_begin_index_brin shall
> > be called
> > > with heap/index being NULL, leading to some failure during the
> > sort. But
> > > maybe this means we don't actually need the heap/index fields,
> > it's just
> > > a copy of TuplesortIndexArg, but BRIN does not need that
> > because we sort
> > > the tuples by blkno, and we don't need the descriptors for
> that.
> > >
> > > Unfortunately I can't test on Windows, since I can't build with
> > meson on
> > > Windows.
> > >
> > >
> > > In any case, the _brin_parallel_scan_and_build does not
> > actually need
> > > the separate heap/index arguments, those are already in the
> spool.
> > >
> > > Yeah, for sure.
> > >
> > >
> > > I'll try to figure out if we want to simplify the tuplesort or
> > remove
> > > the arguments from _brin_parallel_scan_and_build.
> > >
> >
> > Here is a patch simplifying the BRIN parallel create code a little
> bit.
> > As I suspected, we don't need the heap/index in the spool at all,
> and we
> > don't need to pass it to tuplesort_begin_index_brin either - we only
> > need blkno, and we have that in the datum1 field. This also means we
> > don't need TuplesortIndexBrinArg.
> >
> > With Windows 10, msvc 2022, compile end pass ninja test.
> >
> > But, if you allow me, I would like to try another approach to
> > simplification.
> > Instead of increasing the arguments in the call, wouldn't it be better
> > to decrease them
> > and this way all arguments will be passed in the registers instead of on
> > a stack?
> >
>
> If this was beneficial, we'd be passing everything through structs and
> not as explicit arguments. But we don't. If you're arguing it's
> beneficial in this case, it'd be good to see it demonstrated.
>
Please see the https://www.agner.org/optimize/optimizing_cpp.pdf
Excerpt:
"Use 64-bit mode
Parameter transfer is more efficient in 64-bit mode than in 32-bit mode,
and more efficient in 64-bit Linux than in 64-bit Windows. In 64-bit Linux,
the first six integer parameters and the first eight floating point
parameters are transferred in registers, totaling up to fourteen register
parameters. In 64-bit Windows, the first four parameters are transferred in
registers, regardless of whether they are integers or floating point
numbers."

With function:
_brin_parallel_scan_and_build(buildstate, buildstate->bs_spool,
brinshared, sharedsort,  heapRel, indexRel, sortmem, false);
We have:
Linux -> six first parameters in registers and two parameters in stack
Windows -> four parameters in registers and four parameters in stack


> > bs_spool may well contain this data and will probably be useful in the
> > future.
> >
> > I made a v1 version, based on your patch, for your consideration.
> >
>
> I did actually consider doing it this

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Tomas Vondra]

On 12/29/23 12:53, Ranier Vilela wrote:
> Em qui., 28 de dez. de 2023 às 22:16, Tomas Vondra
> mailto:tomas.von...@enterprisedb.com>>
> escreveu:
> 
> 
> 
> On 12/27/23 12:37, Ranier Vilela wrote:
> > Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
> >  
>  >>
> > escreveu:
> >
> >
> >
> >     On 12/26/23 19:10, Ranier Vilela wrote:
> >     > Hi,
> >     >
> >     > The commit b437571
> >      
> >      >> I
> >     > think has an oversight.
> >     > When allocate memory and initialize private spool in function:
> >     > _brin_leader_participate_as_worker
> >     >
> >     > The behavior is the bs_spool (heap and index fields)
> >     > are left empty.
> >     >
> >     > The code affected is:
> >     >   buildstate->bs_spool = (BrinSpool *)
> palloc0(sizeof(BrinSpool));
> >     > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> >     > - buildstate->bs_spool->index = buildstate->bs_spool->index;
> >     > + buildstate->bs_spool->heap = heap;
> >     > + buildstate->bs_spool->index = index;
> >     >
> >     > Is the fix correct?
> >     >
> >
> >     Thanks for noticing this.
> >
> > You're welcome.
> >  
> >
> >     Yes, I believe this is a bug - the assignments
> >     are certainly wrong, it leaves the fields set to NULL.
> >
> >     I wonder how come this didn't fail during testing. Surely, if
> the leader
> >     participates as a worker, the tuplesort_begin_index_brin shall
> be called
> >     with heap/index being NULL, leading to some failure during the
> sort. But
> >     maybe this means we don't actually need the heap/index fields,
> it's just
> >     a copy of TuplesortIndexArg, but BRIN does not need that
> because we sort
> >     the tuples by blkno, and we don't need the descriptors for that.
> >
> > Unfortunately I can't test on Windows, since I can't build with
> meson on
> > Windows.
> >
> >
> >     In any case, the _brin_parallel_scan_and_build does not
> actually need
> >     the separate heap/index arguments, those are already in the spool.
> >
> > Yeah, for sure.
> >
> >
> >     I'll try to figure out if we want to simplify the tuplesort or
> remove
> >     the arguments from _brin_parallel_scan_and_build.
> >
> 
> Here is a patch simplifying the BRIN parallel create code a little bit.
> As I suspected, we don't need the heap/index in the spool at all, and we
> don't need to pass it to tuplesort_begin_index_brin either - we only
> need blkno, and we have that in the datum1 field. This also means we
> don't need TuplesortIndexBrinArg.
> 
> With Windows 10, msvc 2022, compile end pass ninja test.
> 
> But, if you allow me, I would like to try another approach to
> simplification.
> Instead of increasing the arguments in the call, wouldn't it be better
> to decrease them 
> and this way all arguments will be passed in the registers instead of on
> a stack?
> 

If this was beneficial, we'd be passing everything through structs and
not as explicit arguments. But we don't. If you're arguing it's
beneficial in this case, it'd be good to see it demonstrated.

> bs_spool may well contain this data and will probably be useful in the
> future.
> 
> I made a v1 version, based on your patch, for your consideration.
> 

I did actually consider doing it this way yesterday, but I don't like
this approach. I don't believe it's faster (and even if it was, the
difference is going to be negligible), and parameters hidden in some
struct increase the cognitive load. I like explicit arguments.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Em qui., 28 de dez. de 2023 às 22:16, Tomas Vondra <
tomas.von...@enterprisedb.com> escreveu:

>
>
> On 12/27/23 12:37, Ranier Vilela wrote:
> > Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
> > mailto:tomas.von...@enterprisedb.com>>
> > escreveu:
> >
> >
> >
> > On 12/26/23 19:10, Ranier Vilela wrote:
> > > Hi,
> > >
> > > The commit b437571
> >  > > I
> > > think has an oversight.
> > > When allocate memory and initialize private spool in function:
> > > _brin_leader_participate_as_worker
> > >
> > > The behavior is the bs_spool (heap and index fields)
> > > are left empty.
> > >
> > > The code affected is:
> > >   buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
> > > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> > > - buildstate->bs_spool->index = buildstate->bs_spool->index;
> > > + buildstate->bs_spool->heap = heap;
> > > + buildstate->bs_spool->index = index;
> > >
> > > Is the fix correct?
> > >
> >
> > Thanks for noticing this.
> >
> > You're welcome.
> >
> >
> > Yes, I believe this is a bug - the assignments
> > are certainly wrong, it leaves the fields set to NULL.
> >
> > I wonder how come this didn't fail during testing. Surely, if the
> leader
> > participates as a worker, the tuplesort_begin_index_brin shall be
> called
> > with heap/index being NULL, leading to some failure during the sort.
> But
> > maybe this means we don't actually need the heap/index fields, it's
> just
> > a copy of TuplesortIndexArg, but BRIN does not need that because we
> sort
> > the tuples by blkno, and we don't need the descriptors for that.
> >
> > Unfortunately I can't test on Windows, since I can't build with meson on
> > Windows.
> >
> >
> > In any case, the _brin_parallel_scan_and_build does not actually need
> > the separate heap/index arguments, those are already in the spool.
> >
> > Yeah, for sure.
> >
> >
> > I'll try to figure out if we want to simplify the tuplesort or remove
> > the arguments from _brin_parallel_scan_and_build.
> >
>
> Here is a patch simplifying the BRIN parallel create code a little bit.
> As I suspected, we don't need the heap/index in the spool at all, and we
> don't need to pass it to tuplesort_begin_index_brin either - we only
> need blkno, and we have that in the datum1 field. This also means we
> don't need TuplesortIndexBrinArg.
>
With Windows 10, msvc 2022, compile end pass ninja test.

But, if you allow me, I would like to try another approach to
simplification.
Instead of increasing the arguments in the call, wouldn't it be better to
decrease them
and this way all arguments will be passed in the registers instead of on a
stack?

bs_spool may well contain this data and will probably be useful in the
future.

I made a v1 version, based on your patch, for your consideration.

best regards,
Ranier Vilela


v1-brin-parallel-create-simplify.patch
Description: Binary data

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Tomas Vondra]

On 12/27/23 12:37, Ranier Vilela wrote:
> Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra
> mailto:tomas.von...@enterprisedb.com>>
> escreveu:
> 
> 
> 
> On 12/26/23 19:10, Ranier Vilela wrote:
> > Hi,
> >
> > The commit b437571
>  > I
> > think has an oversight.
> > When allocate memory and initialize private spool in function:
> > _brin_leader_participate_as_worker
> >
> > The behavior is the bs_spool (heap and index fields)
> > are left empty.
> >
> > The code affected is:
> >   buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
> > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> > - buildstate->bs_spool->index = buildstate->bs_spool->index;
> > + buildstate->bs_spool->heap = heap;
> > + buildstate->bs_spool->index = index;
> >
> > Is the fix correct?
> >
> 
> Thanks for noticing this.
> 
> You're welcome.
>  
> 
> Yes, I believe this is a bug - the assignments
> are certainly wrong, it leaves the fields set to NULL.
> 
> I wonder how come this didn't fail during testing. Surely, if the leader
> participates as a worker, the tuplesort_begin_index_brin shall be called
> with heap/index being NULL, leading to some failure during the sort. But
> maybe this means we don't actually need the heap/index fields, it's just
> a copy of TuplesortIndexArg, but BRIN does not need that because we sort
> the tuples by blkno, and we don't need the descriptors for that.
> 
> Unfortunately I can't test on Windows, since I can't build with meson on
> Windows.
> 
> 
> In any case, the _brin_parallel_scan_and_build does not actually need
> the separate heap/index arguments, those are already in the spool.
> 
> Yeah, for sure.
> 
> 
> I'll try to figure out if we want to simplify the tuplesort or remove
> the arguments from _brin_parallel_scan_and_build.
> 

Here is a patch simplifying the BRIN parallel create code a little bit.
As I suspected, we don't need the heap/index in the spool at all, and we
don't need to pass it to tuplesort_begin_index_brin either - we only
need blkno, and we have that in the datum1 field. This also means we
don't need TuplesortIndexBrinArg.

I'll push this tomorrow, probably.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Companydiff --git a/src/backend/access/brin/brin.c b/src/backend/access/brin/brin.c
index 23f081389b2..a58f662f2cf 100644
--- a/src/backend/access/brin/brin.c
+++ b/src/backend/access/brin/brin.c
@@ -56,8 +56,6 @@
 typedef struct BrinSpool
 {
 	Tuplesortstate *sortstate;	/* state data for tuplesort.c */
-	Relation	heap;
-	Relation	index;
 } BrinSpool;
 
 /*
@@ -1144,8 +1142,6 @@ brinbuild(Relation heap, Relation index, IndexInfo *indexInfo)
 	   RelationGetNumberOfBlocks(heap));
 
 	state->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
-	state->bs_spool->heap = heap;
-	state->bs_spool->index = index;
 
 	/*
 	 * Attempt to launch parallel worker scan when required
@@ -1200,8 +1196,7 @@ brinbuild(Relation heap, Relation index, IndexInfo *indexInfo)
 		 * factor.
 		 */
 		state->bs_spool->sortstate =
-			tuplesort_begin_index_brin(heap, index,
-	   maintenance_work_mem, coordinate,
+			tuplesort_begin_index_brin(maintenance_work_mem, coordinate,
 	   TUPLESORT_NONE);
 
 		/*
@@ -2706,8 +2701,6 @@ _brin_leader_participate_as_worker(BrinBuildState *buildstate, Relation heap, Re
 
 	/* Allocate memory and initialize private spool */
 	buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
-	buildstate->bs_spool->heap = buildstate->bs_spool->heap;
-	buildstate->bs_spool->index = buildstate->bs_spool->index;
 
 	/*
 	 * Might as well use reliable figure when doling out maintenance_work_mem
@@ -2736,8 +2729,8 @@ _brin_leader_participate_as_worker(BrinBuildState *buildstate, Relation heap, Re
 static void
 _brin_parallel_scan_and_build(BrinBuildState *state, BrinSpool *brinspool,
 			  BrinShared *brinshared, Sharedsort *sharedsort,
-			  Relation heap, Relation index, int sortmem,
-			  bool progress)
+			  Relation heap, Relation index,
+			  int sortmem, bool progress)
 {
 	SortCoordinate coordinate;
 	TableScanDesc scan;
@@ -2751,9 +2744,7 @@ _brin_parallel_scan_and_build(BrinBuildState *state, BrinSpool *brinspool,
 	coordinate->sharedsort = sharedsort;
 
 	/* Begin "partial" tuplesort */
-	brinspool->sortstate = tuplesort_begin_index_brin(brinspool->heap,
-	  brinspool->index,
-	  sortmem, coordinate,
+	brinspool->sortstate = tuplesort_begin_index_brin(sortmem, coordinate,
 	  TUPLESORT_NONE);
 
 	/* Join parallel scan */
@@ -2763,8 +2754,8 @@ _brin_parallel_scan_and_build(BrinBuildState *state, BrinSpool *brinspool,
 	scan = table_beginscan_parallel(heap,
 	Parall

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Em ter., 26 de dez. de 2023 às 19:07, Tomas Vondra <
tomas.von...@enterprisedb.com> escreveu:

>
>
> On 12/26/23 19:10, Ranier Vilela wrote:
> > Hi,
> >
> > The commit b437571  I
> > think has an oversight.
> > When allocate memory and initialize private spool in function:
> > _brin_leader_participate_as_worker
> >
> > The behavior is the bs_spool (heap and index fields)
> > are left empty.
> >
> > The code affected is:
> >   buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
> > - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> > - buildstate->bs_spool->index = buildstate->bs_spool->index;
> > + buildstate->bs_spool->heap = heap;
> > + buildstate->bs_spool->index = index;
> >
> > Is the fix correct?
> >
>
> Thanks for noticing this.

You're welcome.


> Yes, I believe this is a bug - the assignments
> are certainly wrong, it leaves the fields set to NULL.
>
> I wonder how come this didn't fail during testing. Surely, if the leader
> participates as a worker, the tuplesort_begin_index_brin shall be called
> with heap/index being NULL, leading to some failure during the sort. But
> maybe this means we don't actually need the heap/index fields, it's just
> a copy of TuplesortIndexArg, but BRIN does not need that because we sort
> the tuples by blkno, and we don't need the descriptors for that.
>
Unfortunately I can't test on Windows, since I can't build with meson on
Windows.


> In any case, the _brin_parallel_scan_and_build does not actually need
> the separate heap/index arguments, those are already in the spool.
>
Yeah, for sure.


> I'll try to figure out if we want to simplify the tuplesort or remove
> the arguments from _brin_parallel_scan_and_build.
>
Thank you for your work.

best regards,
Ranier Vilela

Re: Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Tomas Vondra]

On 12/26/23 19:10, Ranier Vilela wrote:
> Hi,
> 
> The commit b437571  I
> think has an oversight.
> When allocate memory and initialize private spool in function:
> _brin_leader_participate_as_worker
> 
> The behavior is the bs_spool (heap and index fields)
> are left empty.
> 
> The code affected is:
>   buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
> - buildstate->bs_spool->heap = buildstate->bs_spool->heap;
> - buildstate->bs_spool->index = buildstate->bs_spool->index;
> + buildstate->bs_spool->heap = heap;
> + buildstate->bs_spool->index = index;
> 
> Is the fix correct?
> 

Thanks for noticing this. Yes, I believe this is a bug - the assignments
are certainly wrong, it leaves the fields set to NULL.

I wonder how come this didn't fail during testing. Surely, if the leader
participates as a worker, the tuplesort_begin_index_brin shall be called
with heap/index being NULL, leading to some failure during the sort. But
maybe this means we don't actually need the heap/index fields, it's just
a copy of TuplesortIndexArg, but BRIN does not need that because we sort
the tuples by blkno, and we don't need the descriptors for that.

In any case, the _brin_parallel_scan_and_build does not actually need
the separate heap/index arguments, those are already in the spool.

I'll try to figure out if we want to simplify the tuplesort or remove
the arguments from _brin_parallel_scan_and_build.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Fix Brin Private Spool Initialization (src/backend/access/brin/brin.c)

[Ranier Vilela]

Hi,

The commit b437571  I
think has an oversight.
When allocate memory and initialize private spool in function:
_brin_leader_participate_as_worker

The behavior is the bs_spool (heap and index fields)
are left empty.

The code affected is:
  buildstate->bs_spool = (BrinSpool *) palloc0(sizeof(BrinSpool));
- buildstate->bs_spool->heap = buildstate->bs_spool->heap;
- buildstate->bs_spool->index = buildstate->bs_spool->index;
+ buildstate->bs_spool->heap = heap;
+ buildstate->bs_spool->index = index;

Is the fix correct?

best regards,
Ranier Vilela


001-fix-brin-private-spool-initialization.patch
Description: Binary data