Re: SerializedSnapshotData alignment

On Mon, Feb 27, 2017 at 5:13 AM, Noah Misch <noah(at)leadboat(dot)com> wrote:
> Dear 7b4ac19 authors,
>
> Field ps_snapshot_data usually receives four-byte alignment within
> ParallelIndexScanDescData, but it contains the eight-byte whenTaken field.
> The select_parallel test dies with SIGBUS on "Oracle Solaris 10 1/13
> s10s_u11wos_24a SPARC", building with gcc 4.9.2. Some credible fixes:
>
> 1. Move the SerializedSnapshotData declaration from snapmgr.c to snapmgr.h and
> declare the ps_snapshot_data field to be of type SerializedSnapshotData.
> Probably also add a field "TransactionId xids[FLEXIBLE_ARRAY_MEMBER]" to
> SerializedSnapshotData, to assert the variable-length nature.
>
> 2. Change "char ps_snapshot_data[FLEXIBLE_ARRAY_MEMBER]" to "int64 ...". I
> have attached this in SerializedSnapshot-int64-v1.patch.
>
> 3. Change no declarations, and make snapmgr.c memcpy() the
> SerializedSnapshotData through a local buffer. I have attached this as
> SerializedSnapshot-memcpy-v1.patch.
>
> I like (2) well enough, but I don't see that technique used elsewhere in the
> tree. (1) is more typical of PostgreSQL, though I personally like it when
> structs can stay private to a file. (3) is also well-attested, particularly
> in xlog replay code. I am leaning toward (2). Other opinions?

In my imagination, we were already doing #3, so I'd probably favor
that approach.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company