On Thu, Nov 19, 2015 at 10:11 PM, Kouhei Kaigai <kaigai@ak.jp.nec.com> wrote:
> The above two points are for the case if and when extension want to use
> variable length fields for its private fields.
> So, nodeAlloc() callback is not a perfect answer for the use case because
> length of the variable length fields shall be (usually) determined by the
> value of another fields (like num_inner_rels, num_gpu_devices, ...) thus
> we cannot determine the correct length before read.
>
> Let's assume an custom-scan extension that wants to have:
>
> typedef struct {
> CustomScan cscan;
> int priv_value_1;
> long priv_value_2;
> extra_info_t extra_subplan_info[FLEXIBLE_ARRAY_MEMBER];
> /* its length equal to number of sub-plans */
> } ExampleScan;
>
> The "extnodename" within CustomScan allows to pull callback functions
> to handle read node from the serialized format.
> However, nodeAlloc() callback cannot determine the correct length
> (= number of sub-plans in this example) prior to read 'cscan' part.
>
> So, I'd like to suggest _readCustomScan (and other extendable nodes
> also) read tokens on local CustomScan variable once, then call
> Node *(nodeRead)(Node *local_node)
> to allocate entire ExampleScan node and read other private fields.
>
> The local_node is already filled up by the core backend prior to
> the callback invocation, so extension can know how many sub-plans
> are expected thus how many private tokens shall appear.
> It also means extension can know exact length of the ExampleScan
> node, so it can allocate the node as expected then fill up
> remaining private tokens.
On second thought, I think we should insist that nodes have to be
fixed-size. This sounds like a mess. If the node needs a variable
amount of storage for something, it can store a pointer to a
separately-allocated array someplace inside of it. That's what
existing nodes do, and it works fine.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company