Re: SerializeParamList vs machines with strict alignment - Mailing list pgsql-hackers

Wow, Tom. This is great stuff. Thanks for it.

Lou Picciano

> On Sep 9, 2018, at 11:27 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> I wondered why buildfarm member chipmunk has been failing hard
> for the last little while.  Fortunately, it's supplying us with
> a handy backtrace:
>
> Program terminated with signal 7, Bus error.
> #0  EA_flatten_into (allocated_size=<optimized out>, result=0xb55ff30e, eohptr=0x188f440) at array_expanded.c:329
> 329        aresult->dataoffset = dataoffset;
> #0  EA_flatten_into (allocated_size=<optimized out>, result=0xb55ff30e, eohptr=0x188f440) at array_expanded.c:329
> #1  EA_flatten_into (eohptr=0x188f440, result=0xb55ff30e, allocated_size=<optimized out>) at array_expanded.c:293
> #2  0x003c3dfc in EOH_flatten_into (eohptr=<optimized out>, result=<optimized out>, allocated_size=<optimized out>)
atexpandeddatum.c:84 
> #3  0x003c076c in datumSerialize (value=3934060, isnull=<optimized out>, typByVal=<optimized out>, typLen=<optimized
out>,start_address=0xbea3bd54) at datum.c:341 
> #4  0x002a8510 in SerializeParamList (paramLI=0x1889f18, start_address=0xbea3bd54) at params.c:195
> #5  0x002342cc in ExecInitParallelPlan (planstate=0xffffffff, estate=0x18863e0, sendParams=0x46e, nworkers=1,
tuples_needed=-1)at execParallel.c:700 
> #6  0x002461dc in ExecGather (pstate=0x18864f0) at nodeGather.c:151
> #7  0x00236b20 in ExecProcNodeFirst (node=0x18864f0) at execProcnode.c:445
> #8  0x0022fc2c in ExecProcNode (node=0x18864f0) at ../../../src/include/executor/executor.h:237
> #9  ExecutePlan (execute_once=<optimized out>, dest=0x188a108, direction=<optimized out>, numberTuples=0,
sendTuples=<optimizedout>, operation=CMD_SELECT, use_parallel_mode=<optimized out>, planstate=0x18864f0,
estate=0x18863e0)at execMain.c:1721 
> #10 standard_ExecutorRun (queryDesc=0x188a138, direction=<optimized out>, count=0, execute_once=true) at
execMain.c:362
> #11 0x0023d630 in postquel_getnext (fcache=0x1888408, es=0x1889d68) at functions.c:867
> #12 fmgr_sql (fcinfo=0x701c7c) at functions.c:1164
>
> This is remarkably hard to replicate on other machines, but I eventually
> managed to duplicate it on gaur's host, after which it became really
> obvious that the parallel-query data transfer logic has never been
> stressed very hard on machines with strict data alignment rules.
>
> In particular, SerializeParamList does this:
>
>        /* Write flags. */
>        memcpy(*start_address, &prm->pflags, sizeof(uint16));
>        *start_address += sizeof(uint16);
>
> immediately followed by this:
>
>        datumSerialize(prm->value, prm->isnull, typByVal, typLen,
>                       start_address);
>
> and datumSerialize might do this:
>
>            EOH_flatten_into(eoh, (void *) *start_address, header);
>
> Now, I will plead mea culpa that the expanded-object API doesn't
> say in large red letters that the target address for EOH_flatten_into
> is supposed to be maxaligned.  It only says
>
> * The flattened representation must be a valid in-line, non-compressed,
> * 4-byte-header varlena object.
>
> Still, one might reasonably suspect from that that *at least* 4-byte
> alignment is expected.  This code path isn't providing such alignment,
> and machines that require it will crash.  The only reason we've not
> noticed, AFAICS, is that nobody has been running with
> force_parallel_mode = regress on alignment-picky hardware.
>
>             regards, tom lane
>