Move the responsibility for creating/destroying TupleQueueReaders into
execParallel.c, to avoid duplicative coding in nodeGather.c and
nodeGatherMerge.c. Also, instead of having DestroyTupleQueueReader do
shm_mq_detach, do it in the caller (which is now only ExecParallelFinish).
This means execParallel.c does both the attaching and detaching of the
tuple-queue-reader shm_mqs, which seems less weird than the previous
arrangement.
These changes also eliminate a vestigial memory leak (of the pei->tqueue
array). It's now demonstrable that rescans of Gather or GatherMerge don't
leak memory.
Discussion: https://postgr.es/m/8670.1504192177@sss.pgh.pa.us
Comment the fields of GatherMergeState, and organize them a bit more
sensibly. Comment GMReaderTupleBuffer more usefully too. Improve
assorted other comments that were obsolete or just not very good English.
Get rid of the use of a GMReaderTupleBuffer for the leader process;
that was confusing, since only the "done" field was used, and that
in a way redundant with need_to_scan_locally.
In gather_merge_init, avoid calling load_tuple_array for
already-known-exhausted workers. I'm not sure if there's a live bug there,
but the case is unlikely to be well tested due to timing considerations.
Remove some useless code, such as duplicating the tts_isempty test done by
TupIsNull.
Remove useless initialization of ps.qual, replacing that with an assertion
that we have no qual to check. (If we did, the code would fail to check
it.)
Avoid applying heap_copytuple to a null tuple. While that fails to crash,
it's confusing and it makes the code less legible not more so IMO.
Propagate a couple of these changes into nodeGather.c, as well.
Back-patch to v10, partly because of the possibility that the
gather_merge_init change is fixing a live bug, but mostly to keep
the branches in sync to ease future bug fixes.
Previously, the parallel executor logic did reinitialization of shared
state within the ExecReScan code for parallel-aware scan nodes. This is
problematic, because it means that the ExecReScan call has to occur
synchronously (ie, during the parent Gather node's ReScan call). That is
swimming very much against the tide so far as the ExecReScan machinery is
concerned; the fact that it works at all today depends on a lot of fragile
assumptions, such as that no plan node between Gather and a parallel-aware
scan node is parameterized. Another objection is that because ExecReScan
might be called in workers as well as the leader, hacky extra tests are
needed in some places to prevent unwanted shared-state resets.
Hence, let's separate this code into two functions, a ReInitializeDSM
call and the ReScan call proper. ReInitializeDSM is called only in
the leader and is guaranteed to run before we start new workers.
ReScan is returned to its traditional function of resetting only local
state, which means that ExecReScan's usual habits of delaying or
eliminating child rescan calls are safe again.
As with the preceding commit 7df2c1f8d, it doesn't seem to be necessary
to make these changes in 9.6, which is a good thing because the FDW and
CustomScan APIs are impacted.
Discussion: https://postgr.es/m/CAA4eK1JkByysFJNh9M349u_nNjqETuEnY_y1VUc_kJiU0bxtaQ@mail.gmail.com
The ExecReScan machinery contains various optimizations for postponing
or skipping rescans of plan subtrees; for example a HashAgg node may
conclude that it can re-use the table it built before, instead of
re-reading its input subtree. But that is wrong if the input contains
a parallel-aware table scan node, since the portion of the table scanned
by the leader process is likely to vary from one rescan to the next.
This explains the timing-dependent buildfarm failures we saw after
commit a2b70c89c.
The established mechanism for showing that a plan node's output is
potentially variable is to mark it as depending on some runtime Param.
Hence, to fix this, invent a dummy Param (one that has a PARAM_EXEC
parameter number, but carries no actual value) associated with each Gather
or GatherMerge node, mark parallel-aware nodes below that node as dependent
on that Param, and arrange for ExecReScanGather[Merge] to flag that Param
as changed whenever the Gather[Merge] node is rescanned.
This solution breaks an undocumented assumption made by the parallel
executor logic, namely that all rescans of nodes below a Gather[Merge]
will happen synchronously during the ReScan of the top node itself.
But that's fundamentally contrary to the design of the ExecReScan code,
and so was doomed to fail someday anyway (even if you want to argue
that the bug being fixed here wasn't a failure of that assumption).
A follow-on patch will address that issue. In the meantime, the worst
that's expected to happen is that given very bad timing luck, the leader
might have to do all the work during a rescan, because workers think
they have nothing to do, if they are able to start up before the eventual
ReScan of the leader's parallel-aware table scan node has reset the
shared scan state.
Although this problem exists in 9.6, there does not seem to be any way
for it to manifest there. Without GatherMerge, it seems that a plan tree
that has a rescan-short-circuiting node below Gather will always also
have one above it that will short-circuit in the same cases, preventing
the Gather from being rescanned. Hence we won't take the risk of
back-patching this change into 9.6. But v10 needs it.
Discussion: https://postgr.es/m/CAA4eK1JkByysFJNh9M349u_nNjqETuEnY_y1VUc_kJiU0bxtaQ@mail.gmail.com
If we only need, say, 10 tuples in total, then we certainly don't need
more than 10 tuples from any single process. Pushing down the limit
lets workers exit early when possible. For Gather Merge, there is
an additional benefit: a Sort immediately below the Gather Merge can
be done as a bounded sort if there is an applicable limit.
Robert Haas and Tom Lane
Discussion: http://postgr.es/m/CA+TgmoYa3QKKrLj5rX7UvGqhH73G1Li4B-EKxrmASaca2tFu9Q@mail.gmail.com
This allows us to add stack-depth checks the first time an executor
node is called, and skip that overhead on following
calls. Additionally it yields a nice speedup.
While it'd probably have been a good idea to have that check all
along, it has become more important after the new expression
evaluation framework in b8d7f053c5 - there's no stack depth
check in common paths anymore now. We previously relied on
ExecEvalExpr() being executed somewhere.
We should move towards that model for further routines, but as this is
required for v10, it seems better to only do the necessary (which
already is quite large).
Author: Andres Freund, Tom Lane
Reported-By: Julien Rouhaud
Discussion:
https://postgr.es/m/22833.1490390175@sss.pgh.pa.ushttps://postgr.es/m/b0af9eaa-130c-60d0-9e4e-7a135b1e0c76@dalibo.com
In a followup commit ExecProcNode(), and especially the large switch
it contains, will largely be replaced by a function pointer directly
to the correct node. The node functions will then get invoked by a
thin inline function wrapper. To avoid having to include miscadmin.h
in headers - CHECK_FOR_INTERRUPTS() - move the interrupt checks into
the individual executor routines.
While looking through all executor nodes, I noticed a number of
arguably missing interrupt checks, add these too.
Author: Andres Freund, Tom Lane
Reviewed-By: Tom Lane
Discussion:
https://postgr.es/m/22833.1490390175@sss.pgh.pa.us
Change pg_bsd_indent to follow upstream rules for placement of comments
to the right of code, and remove pgindent hack that caused comments
following #endif to not obey the general rule.
Commit e3860ffa4d wasn't actually using
the published version of pg_bsd_indent, but a hacked-up version that
tried to minimize the amount of movement of comments to the right of
code. The situation of interest is where such a comment has to be
moved to the right of its default placement at column 33 because there's
code there. BSD indent has always moved right in units of tab stops
in such cases --- but in the previous incarnation, indent was working
in 8-space tab stops, while now it knows we use 4-space tabs. So the
net result is that in about half the cases, such comments are placed
one tab stop left of before. This is better all around: it leaves
more room on the line for comment text, and it means that in such
cases the comment uniformly starts at the next 4-space tab stop after
the code, rather than sometimes one and sometimes two tabs after.
Also, ensure that comments following #endif are indented the same
as comments following other preprocessor commands such as #else.
That inconsistency turns out to have been self-inflicted damage
from a poorly-thought-through post-indent "fixup" in pgindent.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
This replaces the old, recursive tree-walk based evaluation, with
non-recursive, opcode dispatch based, expression evaluation.
Projection is now implemented as part of expression evaluation.
This both leads to significant performance improvements, and makes
future just-in-time compilation of expressions easier.
The speed gains primarily come from:
- non-recursive implementation reduces stack usage / overhead
- simple sub-expressions are implemented with a single jump, without
function calls
- sharing some state between different sub-expressions
- reduced amount of indirect/hard to predict memory accesses by laying
out operation metadata sequentially; including the avoidance of
nearly all of the previously used linked lists
- more code has been moved to expression initialization, avoiding
constant re-checks at evaluation time
Future just-in-time compilation (JIT) has become easier, as
demonstrated by released patches intended to be merged in a later
release, for primarily two reasons: Firstly, due to a stricter split
between expression initialization and evaluation, less code has to be
handled by the JIT. Secondly, due to the non-recursive nature of the
generated "instructions", less performance-critical code-paths can
easily be shared between interpreted and compiled evaluation.
The new framework allows for significant future optimizations. E.g.:
- basic infrastructure for to later reduce the per executor-startup
overhead of expression evaluation, by caching state in prepared
statements. That'd be helpful in OLTPish scenarios where
initialization overhead is measurable.
- optimizing the generated "code". A number of proposals for potential
work has already been made.
- optimizing the interpreter. Similarly a number of proposals have
been made here too.
The move of logic into the expression initialization step leads to some
backward-incompatible changes:
- Function permission checks are now done during expression
initialization, whereas previously they were done during
execution. In edge cases this can lead to errors being raised that
previously wouldn't have been, e.g. a NULL array being coerced to a
different array type previously didn't perform checks.
- The set of domain constraints to be checked, is now evaluated once
during expression initialization, previously it was re-built
every time a domain check was evaluated. For normal queries this
doesn't change much, but e.g. for plpgsql functions, which caches
ExprStates, the old set could stick around longer. The behavior
around might still change.
Author: Andres Freund, with significant changes by Tom Lane,
changes by Heikki Linnakangas
Reviewed-By: Tom Lane, Heikki Linnakangas
Discussion: https://postgr.es/m/20161206034955.bh33paeralxbtluv@alap3.anarazel.de
Coverity complained quite properly that commit ea15e1867 had introduced
unreachable code into ExecGather(); to wit, it was no longer possible to
iterate the final for-loop more or less than once. So remove the for().
In passing, clean up a couple of comments, and make better use of a local
variable.
Since 69f4b9c plain expression evaluation (and thus normal projection)
can't return sets of tuples anymore. Thus remove code dealing with
that possibility.
This will require adjustments in external code using
ExecEvalExpr()/ExecProject() - that should neither be hard nor very
common.
Author: Andres Freund and Tom Lane
Discussion: https://postgr.es/m/20160822214023.aaxz5l4igypowyri@alap3.anarazel.de
The previously code worked OK as long as a Gather node was never
rescanned, or if it was rescanned, as long as it got at least as
many workers on rescan as it had originally. But if the number
of workers ever decreased on a rescan, then it could crash.
Andreas Seltenreich
WaitLatch, WaitLatchOrSocket, and WaitEventSetWait now taken an
additional wait_event_info parameter; legal values are defined in
pgstat.h. This makes it possible to uniquely identify every point in
the core code where we are waiting for a latch; extensions can pass
WAIT_EXTENSION.
Because latches were the major wait primitive not previously covered
by this patch, it is now possible to see information in
pg_stat_activity on a large number of important wait events not
previously addressed, such as ClientRead, ClientWrite, and SyncRep.
Unfortunately, many of the wait events added by this patch will fail
to appear in pg_stat_activity because they're only used in background
processes which don't currently appear in pg_stat_activity. We should
fix this either by creating a separate view for such information, or
else by deciding to include them in pg_stat_activity after all.
Michael Paquier and Robert Haas, reviewed by Alexander Korotkov and
Thomas Munro.
This coding pattern creates a race condition, because if an interesting
interrupt happens after we've checked InterruptPending but before we reset
our latch, the latch-setting done by the signal handler would get lost,
and then we might block at WaitLatch in the next iteration without ever
noticing the interrupt condition. You can put the CHECK_FOR_INTERRUPTS
before WaitLatch or after ResetLatch, but not between them.
Aside from fixing the bugs, add some explanatory comments to latch.h
to perhaps forestall the next person from making the same mistake.
In HEAD, also replace gather_readnext's direct call of
HandleParallelMessages with CHECK_FOR_INTERRUPTS. It does not seem clean
or useful for this one caller to bypass ProcessInterrupts and go straight
to HandleParallelMessages; not least because that fails to consider the
InterruptPending flag, resulting in useless work both here
(if InterruptPending isn't set) and in the next CHECK_FOR_INTERRUPTS call
(if it is).
This thinko seems to have been introduced in the initial coding of
storage/ipc/shm_mq.c (commit ec9037df2), and then blindly copied into all
the subsequent parallel-query support logic. Back-patch relevant hunks
to 9.4 to extirpate the error everywhere.
Discussion: <1661.1469996911@sss.pgh.pa.us>
TupleQueueReaderNext() leaks like a sieve if it has to do any tuple
disassembly/reconstruction. While we could try to clean up its allocations
piecemeal, it seems like a better idea just to insist that it should be run
in a short-lived memory context, so that any transient space goes away
automatically. I chose to have nodeGather.c switch into its existing
per-tuple context before the call, rather than inventing a separate
context inside tqueue.c.
This is sufficient to stop all leakage in the simple case I exhibited
earlier today (see link below), but it does not deal with leaks induced
in more complex cases by tqueue.c's insistence on using TopMemoryContext
for data that it's not actually trying hard to keep track of. That issue
is intertwined with another major source of inefficiency, namely failure
to cache lookup results across calls, so it seems best to deal with it
separately.
In passing, improve some comments, and modify gather_readnext's method for
deciding when it's visited all the readers so that it's more obviously
correct. (I'm not actually convinced that the previous code *is*
correct in the case of a reader deletion; it certainly seems fragile.)
Discussion: <32763.1469821037@sss.pgh.pa.us>
The previous display was sort of confusing, because it didn't
distinguish between the number of workers that we planned to launch
and the number that actually got launched. This has already confused
several people, so display both numbers and label them clearly.
Julien Rouhaud, reviewed by me.
Originally, we didn't have nworkers_launched, so code that used parallel
contexts had to be preprared for the possibility that not all of the
workers requested actually got launched. But now we can count on knowing
the number of workers that were successfully launched, which can shave
off a few cycles and simplify some code slightly.
Amit Kapila, reviewed by Haribabu Kommi, per a suggestion from Peter
Geoghegan.
The original coding read tuples from workers in round-robin fashion,
but performance testing shows that it works much better to read enough
to empty one queue before moving on to the next. I believe the
reason for this is that, with the old approach, we could easily wake
up a worker repeatedly to write only one new tuple into the shm_mq
each time. With this approach, by the time the process gets scheduled,
it has a decent chance of being able to fill the entire buffer in
one go.
Patch by me. Dilip Kumar helped with performance testing.
The previous coding attempts to destroy the DSM in this case, but
child nodes might have stored data there and still be holding onto
pointers in this case. So don't do that.
Also, free the reader array instead of leaking it.
Extracted from two different patch versions both by Amit Kapila.
The patches committed as 6e71dd7ce9
and 3a1f8611f2 were developed in
parallel but dependent on each other in a way that I failed to
notice.
This patch to fix the problem was prepared by Amit Kapila.
Commit 4a4e6893aa, which introduced this
mechanism, failed to account for the fact that the RECORD pseudo-type
uses transient typmods that are only meaningful within a single
backend. Transferring such tuples without modification between two
cooperating backends does not work. This commit installs a system
for passing the tuple descriptors over the same shm_mq being used to
send the tuples themselves. The two sides might not assign the same
transient typmod to any given tuple descriptor, so we must also
substitute the appropriate receiver-side typmod for the one used by
the sender. That adds some CPU overhead, but still seems better than
being unable to pass records between cooperating parallel processes.
Along the way, move the logic for handling multiple tuple queues from
tqueue.c to nodeGather.c; tqueue.c now provides a TupleQueueReader,
which reads from a single queue, rather than a TupleQueueFunnel, which
potentially reads from multiple queues. This change was suggested
previously as a way to make sure that nodeGather.c rather than tqueue.c
had policy control over the order in which to read from queues, but
it wasn't clear to me until now how good an idea it was. typmod
mapping needs to be performed separately for each queue, and it is
much simpler if the tqueue.c code handles that and leaves multiplexing
multiple queues to higher layers of the stack.
Commit b0b0d84b3d purported to make it
possible to relaunch workers using the same parallel context, but it had
an unpleasant race condition: we might reinitialize after the workers
have sent their last control message but before they have dettached the
DSM, leaving to crashes. Repair by introducing a new ParallelContext
operation, ReinitializeParallelDSM.
Adjust execParallel.c to use this new support, so that we can rescan a
Gather node by relaunching workers but without needing to recreate the
DSM.
Amit Kapila, with some adjustments by me. Extracted from latest parallel
sequential scan patch.
The original Gather code failed to mark a Gather node as not able to
do projection, but it couldn't, even though it did call initialize its
projection info via ExecAssignProjectionInfo. There doesn't seem to
be any good reason for this node not to have projection capability,
so clean things up so that it does. Without this, plans using Gather
nodes might need to carry extra Result nodes to do projection.
In the previous coding, before returning from ExecutorRun, we'd shut
down all parallel workers. This was dead wrong if ExecutorRun was
called with a non-zero tuple count; it had the effect of truncating
the query output. To fix, give ExecutePlan control over whether to
enter parallel mode, and have it refuse to do so if the tuple count
is non-zero. Rewrite the Gather logic so that it can cope with being
called outside parallel mode.
Commit 7aea8e4f2d is largely to blame
for this problem, though this patch modifies some subsequently-committed
code which relied on the guarantees it purported to make.
A Gather executor node runs any number of copies of a plan in an equal
number of workers and merges all of the results into a single tuple
stream. It can also run the plan itself, if the workers are
unavailable or haven't started up yet. It is intended to work with
the Partial Seq Scan node which will be added in future commits.
It could also be used to implement parallel query of a different sort
by itself, without help from Partial Seq Scan, if the single_copy mode
is used. In that mode, a worker executes the plan, and the parallel
leader does not, merely collecting the worker's results. So, a Gather
node could be inserted into a plan to split the execution of that plan
across two processes. Nested Gather nodes aren't currently supported,
but we might want to add support for that in the future.
There's nothing in the planner to actually generate Gather nodes yet,
so it's not quite time to break out the champagne. But we're getting
close.
Amit Kapila. Some designs suggestions were provided by me, and I also
reviewed the patch. Single-copy mode, documentation, and other minor
changes also by me.
Andres Freund
Tom Lane
Robert Haas
Bruce Momjian