Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On 08/28/2016 04:59 PM, Tom Lane wrote: Tomas Vondrawrites: I'm however pretty sure this will have absolutely no impact on profiles. Yeah, I do not believe that either. Also, I think that the intent of the existing code is to defend against bad parameters even in non-assert builds. Which might argue for turning these into test-and-elog rather than asserts. I agree. There's a fair number of contrib modules, and when they compile just fine people are unlikely to test them with assert-enabled builds. So they would not notice any issues, but it'd get broken as we've removed the code that fixes the values for them. -- Tomas Vondra http://www.2ndQuadrant.com PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Tomas Vondrawrites: > I'm however pretty sure this will have absolutely no impact on profiles. Yeah, I do not believe that either. Also, I think that the intent of the existing code is to defend against bad parameters even in non-assert builds. Which might argue for turning these into test-and-elog rather than asserts. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On 08/28/2016 04:41 PM, Tom Lane wrote: Robert Haaswrites: Also, I think we ought to replace this code in aset.c: initBlockSize = MAXALIGN(initBlockSize); if (initBlockSize < 1024) initBlockSize = 1024; maxBlockSize = MAXALIGN(maxBlockSize); With this: Assert(initBlockSize >= 1024 && initBlockSize == MAXALIGN(initBlockSize)); Assert(maxBlockSize == MAXALIGN(maxBlockSize)); Good idea --- if we'd had it that way, these errors would never have gotten committed in the first place. I'm for doing that only in HEAD though. regards, tom lane Then maybe also add Assert(initBlockSize <= maxBlockSize); And perhaps also an assert making sure the minContextSize value makes sens with respect to the min/max block sizes? I'm however pretty sure this will have absolutely no impact on profiles. It might save a few cycles, but this only runs when creating the memory context and all profiles I've seen are about palloc/pfree. It might matter when creating many memory contexts, but then you probably have other things to worry about. regards -- Tomas Vondra http://www.2ndQuadrant.com PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On 08/28/2016 04:41 PM, Tom Lane wrote: Robert Haaswrites: Also, I think we ought to replace this code in aset.c: initBlockSize = MAXALIGN(initBlockSize); if (initBlockSize < 1024) initBlockSize = 1024; maxBlockSize = MAXALIGN(maxBlockSize); With this: Assert(initBlockSize >= 1024 && initBlockSize == MAXALIGN(initBlockSize)); Assert(maxBlockSize == MAXALIGN(maxBlockSize)); Good idea --- if we'd had it that way, these errors would never have gotten committed in the first place. I'm for doing that only in HEAD though. Then maybe also add Assert(initBlockSize <= maxBlockSize); And perhaps also an assert making sure the minContextSize value makes sens with respect to the min/max block sizes? I'm however pretty sure this will have absolutely no impact on profiles. It might save a few cycles, but this only runs when creating the memory context and all profiles I've seen are about palloc/pfree. It might matter when creating many memory contexts, but then you probably have other things to worry about. regards -- Tomas Vondra http://www.2ndQuadrant.com PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Robert Haaswrites: > Also, I think we ought to replace this code in aset.c: > initBlockSize = MAXALIGN(initBlockSize); > if (initBlockSize < 1024) > initBlockSize = 1024; > maxBlockSize = MAXALIGN(maxBlockSize); > With this: > Assert(initBlockSize >= 1024 && initBlockSize == MAXALIGN(initBlockSize)); > Assert(maxBlockSize == MAXALIGN(maxBlockSize)); Good idea --- if we'd had it that way, these errors would never have gotten committed in the first place. I'm for doing that only in HEAD though. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On Sat, Aug 27, 2016 at 2:08 PM, Tom Lanewrote: > The standard calling pattern for AllocSetContextCreate is > > cxt = AllocSetContextCreate(parent, name, > ALLOCSET_DEFAULT_MINSIZE, > ALLOCSET_DEFAULT_INITSIZE, > ALLOCSET_DEFAULT_MAXSIZE); > > or for some contexts you might s/DEFAULT/SMALL/g if you expect the context > to never contain very much data. I happened to notice that there are a > few calls that get this pattern wrong. After a bit of grepping, I found: > > hba.c lines 389, 2196: >ALLOCSET_DEFAULT_MINSIZE, >ALLOCSET_DEFAULT_MINSIZE, >ALLOCSET_DEFAULT_MAXSIZE); > > guc-file.l line 146: >ALLOCSET_DEFAULT_MINSIZE, >ALLOCSET_DEFAULT_MINSIZE, >ALLOCSET_DEFAULT_MAXSIZE); > > brin.c line 857: > > ALLOCSET_SMALL_INITSIZE, > ALLOCSET_SMALL_MINSIZE, > ALLOCSET_SMALL_MAXSIZE); > > autovacuum.c line 2184: > ALLOCSET_DEFAULT_INITSIZE, > ALLOCSET_DEFAULT_MINSIZE, > ALLOCSET_DEFAULT_MAXSIZE); > > typcache.c lines 757, 842: > > ALLOCSET_SMALL_INITSIZE, > ALLOCSET_SMALL_MINSIZE, > ALLOCSET_SMALL_MAXSIZE); > > In all of these cases, we're passing zero as the initBlockSize, which is > invalid, but but AllocSetContextCreate silently forces it up to 1K. So > there's no failure but there may be some inefficiency due to small block > sizes of early allocation blocks. I seriously doubt that's intentional; > in some of these cases it might be sane to use the SMALL parameters > instead, but this isn't a good way to get that effect. The last four > cases are also passing a nonzero value as the minContextSize, forcing > the context to allocate at least that much instantly and hold onto it > over resets. That's inefficient as well, though probably only minorly so. I noticed this a while back while playing with my alternate allocator, but wasn't sure how much of it was intentional. Anyway, +1 for doing something to clean this up. Your proposal sounds OK, but maybe AllocSetContextCreateDefault() and AllocSetContextCreateSmall() would be nice and easier to remember. Also, I think we ought to replace this code in aset.c: initBlockSize = MAXALIGN(initBlockSize); if (initBlockSize < 1024) initBlockSize = 1024; maxBlockSize = MAXALIGN(maxBlockSize); With this: Assert(initBlockSize >= 1024 && initBlockSize == MAXALIGN(initBlockSize)); Assert(maxBlockSize == MAXALIGN(maxBlockSize)); This might save a few cycles which wouldn't be unwelcome given that AllocSetContextCreate does show up in profiles, but more importantly I think it's completely inappropriate for this function to paper over whatever errors may be made by callers. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Andres Freundwrites: > On 2016-08-27 15:46:26 -0400, Tom Lane wrote: >> (Or in other words, the fact that "DefaultContextCreate" is spelled >> "AllocSetContextCreate" is a bit of a historical artifact, but I do >> not see why changing the spelling is a useful exercise.) > Well, if you're going through nearly all of the instances where it's > referenced *anyway*... I am not touching, and cannot touch, all the instances in third-party code. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On 2016-08-27 15:46:26 -0400, Tom Lane wrote: > (Or in other words, the fact that "DefaultContextCreate" is spelled > "AllocSetContextCreate" is a bit of a historical artifact, but I do > not see why changing the spelling is a useful exercise.) Well, if you're going through nearly all of the instances where it's referenced *anyway*... -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Andres Freundwrites: > On 2016-08-27 15:36:28 -0400, Tom Lane wrote: >> What is actually of interest, IMO, is making *some* contexts have a >> different allocator, and that is going to require case-by-case decisions >> anyway. A blanket switch seems completely useless to me. > Don't think aset.c is actually that good, so I'm not convinced of > that. And even if it's just to verify that a special case allocator > actually works with more coverage... I'm not exactly following. If you have a new allocator that dominates aset.c on all cases, why would we not simply replace aset.c with it? Mass substitution implemented at the caller end seems like the hard way. (Or in other words, the fact that "DefaultContextCreate" is spelled "AllocSetContextCreate" is a bit of a historical artifact, but I do not see why changing the spelling is a useful exercise.) regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
On 2016-08-27 15:36:28 -0400, Tom Lane wrote: > Andres Freundwrites: > > On 2016-08-27 14:08:44 -0400, Tom Lane wrote: > >> Barring objection, I propose to make these changes in HEAD and 9.6. > > > I think we might also / instead want to think about replacing a lot of > > those AllocSetContextCreate with a wrapper function. Currently is really > > rather annoying to experiment with switching the default allocator > > out. And if we're touching all that code anyway ... > > I do not see why you'd think that "switching the default allocator out" > requires anything except making AllocSetContextCreate do something else. Well yea. But given it's explicitly tied to aset.c that doesn't seem like a sensible thing. DefaultContextCreate() or something (in mcxt.c), which then calls AllocSetContextCreate, seems better to me. > What is actually of interest, IMO, is making *some* contexts have a > different allocator, and that is going to require case-by-case decisions > anyway. A blanket switch seems completely useless to me. Don't think aset.c is actually that good, so I'm not convinced of that. And even if it's just to verify that a special case allocator actually works with more coverage... Andres -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Andres Freundwrites: > On 2016-08-27 14:08:44 -0400, Tom Lane wrote: >> Barring objection, I propose to make these changes in HEAD and 9.6. > I think we might also / instead want to think about replacing a lot of > those AllocSetContextCreate with a wrapper function. Currently is really > rather annoying to experiment with switching the default allocator > out. And if we're touching all that code anyway ... I do not see why you'd think that "switching the default allocator out" requires anything except making AllocSetContextCreate do something else. What is actually of interest, IMO, is making *some* contexts have a different allocator, and that is going to require case-by-case decisions anyway. A blanket switch seems completely useless to me. regards, tom lane -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] Bogus sizing parameters in some AllocSetContextCreate calls
Hi, On 2016-08-27 14:08:44 -0400, Tom Lane wrote: > The standard calling pattern for AllocSetContextCreate is > Barring objection, I propose to make these changes in HEAD and 9.6. > I don't feel a great need to adjust the back branches --- although there > might be some argument for adding these new calling-pattern macros to the > back branches so as not to create a back-patching hazard for patches > that add new AllocSetContextCreate calls. I think we might also / instead want to think about replacing a lot of those AllocSetContextCreate with a wrapper function. Currently is really rather annoying to experiment with switching the default allocator out. And if we're touching all that code anyway ... Andres -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers