Using `row::move_as` while iterating `rowset` yields segmentation fault

[Krzmbrzl]

Originally reported by @avpalienko in #1141 (comment)

When applying the following patch to the common-tests.h file, the move_as test case will crash due to a segmentation fault:

diff --git a/tests/common-tests.h b/tests/common-tests.h
index d2b8fb42..ffc80fc6 100644
--- a/tests/common-tests.h
+++ b/tests/common-tests.h
@@ -6779,7 +6779,7 @@ TEST_CASE_METHOD(common_tests, "BLOB", "[core][blob]")
         }
         SECTION("move_as")
         {
-            soci::rowset< soci::row > rowSet = (sql.prepare << "select b from soci_test where id=:id", soci::use(id));
+            soci::rowset< soci::row > rowSet = (sql.prepare << "select b from soci_test where id=:id union all select b from soci_test where id=:id", soci::use(id, "id"));
             bool containedData = false;
             for (auto it = rowSet.begin(); it != rowSet.end(); ++it)
             {

The issue was originally reported to happen with Oracle, but I was also able to reproduce it with SQLite3, indicating that the issue is not backend-specific.

[Krzmbrzl]

The problem seems to boil down to this:

• the row object inside the rowset seems to be reused across different data fetches (loop iterations)
• the row object is (seemingly) only initialized when first creating the rowset object (by binding the contained statement to the row in rowset's constructor)
• when using move_as the respective entry inside row's holders gets into an invalid state. In case of a blob that corresponds to an uninitialized blob object (one where the backend pointer is nullptr)
• In the next iteration, the statement wants to write data into the previously bound objects, including those that have been moved from and are thus in an invalid state
• Trying to write data into an uninitialized blob dereferences the nullptr backend resulting in the segmentation fault

Therefore, the fix for this issue would be to find a way to let the row object know about invalidated holders, which should cause it to reinitialize them before trying to use them for data binding.

[avpalienko]

May be we could use shared_ptr inside the "row" class and eliminate move_as altogether ?

[Krzmbrzl]

Nah, the get function cant return a pointer. That'd be inconsistent.
Besides, this wouldn't really solve the underlying problem. If the user moves from the dereferenced pointer, we're back at square one with the row object being broken.