[Tarantool-patches] [PATCH 1.10 11/16] WIP: module api: add box_key_def_new_ex()

[Alexander Turenko <alexander.turenko@tarantool.org>]

Unlike box_key_def_new() it allows to set nullability, collation and
JSON path.

Note: JSON paths are not supported in the backported version of the
patch for 1.10.

Provided public non-opaque key part definition structure to create a key
definition. The next commit will also use this structure to dump a key
definition.

There are several techinal points around the box_key_part_def_t
structure. They are mainly about providing stable ABI.

- Two uint32_t fields are placed first for better aligning of next
  fields (pointers, which are usually 64 bit wide).

- A padding is used to guarantee that the structure size will remain the
  same across tarantool versions. It allows to allocate an array of such
  structures.

- The padding array is not a field of the structure itself, but added as
  a union variant (see the code). It allows to get rid of manual
  calculation of cumulative fields size, which is hard to do in a
  platform independent way.

- A minimal size of the structure is guaranteed by the union with
  padding, but a static assert is required to guarantee that the size
  will not overrun the predefined value.

- PACKED is added as an extra remedy to make the structure layout
  predictable.

- A bit flag is used for is_nullable. bool is considered as too
  expensive (it requires 8 bits). bitfields (int:1 and so on) do no
  guarantee certain data layout (it is compiler implementation detail),
  while a module is compiled outside of tarantool build and may use
  different toolchain. A bit flag is the only choice.

- A collation is identified using a string. Different IDs may be used on
  different tarantool instances for collations. The only 'real'
  identifier is a collation name, so using it as identifier in the API
  should be more convenient and less error-prone.

- A field type is also identified using a string instead of a number. We
  have <enum field_type> in the module API, but it cannot be used here,
  because IDs are changed across tarantool versions. Aside of this, size
  of a enum is compiler defined. Anyway, we can expose field types as
  numbers and implement number-to-name and name-to-number mapping
  functions, but IMHO it would just add extra complexity.

XXX: Add a module API test.

Part of #5273

(backported from commit ae27cccc2f967282e05fd76d4d5624ecb9bbb8b2)
---
 extra/exports         |   2 +
 src/box/key_def_api.c | 123 ++++++++++++++++++++++++++++++++++++++++++
 src/box/key_def_api.h | 103 +++++++++++++++++++++++++++++++++++
 3 files changed, 228 insertions(+)

diff --git a/extra/exports b/extra/exports
index 7feaaa05e..b70560db7 100644
--- a/extra/exports
+++ b/extra/exports
@@ -147,6 +147,8 @@ box_txn_alloc
 box_txn_id
 box_key_def_new
 box_key_def_delete
+box_key_def_new_ex
+box_key_part_def_create
 box_tuple_format_default
 box_tuple_new
 box_tuple_ref
diff --git a/src/box/key_def_api.c b/src/box/key_def_api.c
index 104ec5d5c..f19f30950 100644
--- a/src/box/key_def_api.c
+++ b/src/box/key_def_api.c
@@ -30,6 +30,67 @@
  */
 #include "key_def_api.h"
 #include "key_def.h"
+#include "small/region.h"
+#include "coll_id_cache.h"
+#include "tuple_format.h"
+#include "field_def.h"
+#include "coll_id_cache.h"
+#include "fiber.h"
+
+/* {{{ Helpers */
+
+static int
+key_def_set_internal_part(struct key_part_def *internal_part,
+			  box_key_part_def_t *part)
+{
+	*internal_part = key_part_def_default;
+
+	/* Set internal_part->fieldno. */
+	internal_part->fieldno = part->fieldno;
+
+	/* Set internal_part->type. */
+	if (part->field_type == NULL) {
+		diag_set(IllegalParams, "Field type is mandatory");
+		return -1;
+	}
+	size_t type_len = strlen(part->field_type);
+	internal_part->type = field_type_by_name(part->field_type, type_len);
+	if (internal_part->type == field_type_MAX) {
+		diag_set(IllegalParams, "Unknown field type: \"%s\"",
+			 part->field_type);
+		return -1;
+	}
+
+	/* Set internal_part->is_nullable. */
+	internal_part->is_nullable =
+		(part->flags & BOX_KEY_PART_DEF_IS_NULLABLE_MASK) ==
+		BOX_KEY_PART_DEF_IS_NULLABLE_MASK;
+
+	/* Set internal_part->coll_id. */
+	if (part->collation != NULL) {
+		size_t collation_len = strlen(part->collation);
+		struct coll_id *coll_id = coll_by_name(part->collation,
+						       collation_len);
+		if (coll_id == NULL) {
+			diag_set(IllegalParams, "Unknown collation: \"%s\"",
+				 part->collation);
+			return -1;
+		}
+		internal_part->coll_id = coll_id->id;
+	}
+
+	return 0;
+}
+
+/* }}} Helpers */
+
+/* {{{ API functions implementations */
+
+void
+box_key_part_def_create(box_key_part_def_t *part)
+{
+	memset(part, 0, sizeof(*part));
+}
 
 box_key_def_t *
 box_key_def_new(uint32_t *fields, uint32_t *types, uint32_t part_count)
@@ -52,6 +113,66 @@ box_key_def_new(uint32_t *fields, uint32_t *types, uint32_t part_count)
 	return key_def;
 }
 
+box_key_def_t *
+box_key_def_new_ex(box_key_part_def_t *parts, uint32_t part_count)
+{
+	struct region *region = &fiber()->gc;
+	size_t region_svp = region_used(region);
+	size_t internal_parts_size;
+	struct key_part_def *internal_parts =
+		region_alloc_array(region, typeof(internal_parts[0]),
+				   part_count, &internal_parts_size);
+	if (parts == NULL) {
+		diag_set(OutOfMemory, internal_parts_size, "region_alloc_array",
+			 "parts");
+		return NULL;
+	}
+	if (part_count == 0) {
+		diag_set(IllegalParams, "At least one key part is required");
+		return NULL;
+	}
+
+	/*
+	 * It is possible to implement a function similar to
+	 * key_def_new() and eliminate <box_key_part_def_t> ->
+	 * <struct key_part_def> copying. However this would lead
+	 * to code duplication and would complicate maintanence,
+	 * so it worth to do so only if key_def creation will
+	 * appear on a hot path in some meaningful use case.
+	 */
+	uint32_t min_field_count = 0;
+	for (uint32_t i = 0; i < part_count; ++i) {
+		if (key_def_set_internal_part(&internal_parts[i],
+					      &parts[i]) != 0) {
+			region_truncate(region, region_svp);
+			return NULL;
+		}
+		bool is_nullable =
+			(parts[i].flags & BOX_KEY_PART_DEF_IS_NULLABLE_MASK) ==
+			BOX_KEY_PART_DEF_IS_NULLABLE_MASK;
+		if (!is_nullable && parts[i].fieldno > min_field_count)
+			min_field_count = parts[i].fieldno;
+	}
+
+	struct key_def *key_def = key_def_new(internal_parts, part_count);
+	region_truncate(region, region_svp);
+	if (key_def == NULL)
+		return NULL;
+
+	/*
+	 * Update key_def->has_optional_parts and function
+	 * pointers.
+	 *
+	 * FIXME: It seems, this call should be part of
+	 * key_def_new(), because otherwise a callee function may
+	 * obtain an incorrect key_def. However I don't know any
+	 * case that would prove this guess.
+	 */
+	key_def_update_optionality(key_def, min_field_count);
+
+	return key_def;
+}
+
 void
 box_key_def_delete(box_key_def_t *key_def)
 {
@@ -73,3 +194,5 @@ box_tuple_compare_with_key(const box_tuple_t *tuple_a, const char *key_b,
 	return tuple_compare_with_key(tuple_a, key_b, part_count, key_def);
 
 }
+
+/* }}} API functions implementations */
diff --git a/src/box/key_def_api.h b/src/box/key_def_api.h
index 82899eed6..6025afe67 100644
--- a/src/box/key_def_api.h
+++ b/src/box/key_def_api.h
@@ -44,11 +44,84 @@ typedef struct tuple box_tuple_t;
 
 typedef struct key_def box_key_def_t;
 
+/** Key part definition flags. */
+enum {
+	BOX_KEY_PART_DEF_IS_NULLABLE_SHIFT = 0,
+	BOX_KEY_PART_DEF_IS_NULLABLE_MASK =
+		1 << BOX_KEY_PART_DEF_IS_NULLABLE_SHIFT,
+};
+
+/**
+ * It is recommended to verify size of <box_key_part_def_t>
+ * against this constant on the module side at build time.
+ * Example:
+ *
+ * | #if !defined(__cplusplus) && !defined(static_assert)
+ * | #define static_assert _Static_assert
+ * | #endif
+ * |
+ * | (slash)*
+ * |  * Verify that <box_key_part_def_t> has the same size when
+ * |  * compiled within tarantool and within the module.
+ * |  *
+ * |  * It is important, because the module allocates an array of key
+ * |  * parts and passes it to <box_key_def_new_ex>() tarantool
+ * |  * function.
+ * |  *(slash)
+ * | static_assert(sizeof(box_key_part_def_t) == BOX_KEY_PART_DEF_T_SIZE,
+ * |               "sizeof(box_key_part_def_t)");
+ *
+ * This snippet is not part of module.h, because portability of
+ * static_assert() / _Static_assert() is dubious. It should be
+ * decision of a module author how portable its code should be.
+ */
+enum {
+	BOX_KEY_PART_DEF_T_SIZE = 64,
+};
+
+/**
+ * Public representation of a key part definition.
+ *
+ * Usage: Allocate an array of such key parts, initialize each
+ * key part (call <box_key_part_def_create>() and set necessary
+ * fields), pass the array into <box_key_def_new_ex>() function.
+ *
+ * The idea of separation from internal <struct key_part_def> is
+ * to provide stable API and ABI for modules.
+ *
+ * New fields may be added into the end of the structure in later
+ * tarantool versions. Also new flags may be introduced within
+ * <flags> field. <collation> cannot be changed to a union (to
+ * reuse for some other value), because it is verified even for
+ * a non-string key part by <box_key_def_new_ex>().
+ *
+ * Fields that are unknown at given tarantool version are ignored.
+ */
+typedef union PACKED {
+	struct {
+		/** Index of a tuple field (zero based). */
+		uint32_t fieldno;
+		/** Flags, e.g. nullability. */
+		uint32_t flags;
+		/** Type of the tuple field. */
+		const char *field_type;
+		/** Collation name for string comparisons. */
+		const char *collation;
+	};
+	/**
+	 * Padding to guarantee certain size across different
+	 * tarantool versions.
+	 */
+	char padding[BOX_KEY_PART_DEF_T_SIZE];
+} box_key_part_def_t;
+
 /**
  * Create key definition with given field numbers and field types.
  *
  * May be used for tuple format creation and/or tuple comparison.
  *
+ * \sa <box_key_def_new_ex>().
+ *
  * \param fields array with key field identifiers
  * \param types array with key field types (see enum field_type)
  * \param part_count the number of key fields
@@ -57,6 +130,28 @@ typedef struct key_def box_key_def_t;
 API_EXPORT box_key_def_t *
 box_key_def_new(uint32_t *fields, uint32_t *types, uint32_t part_count);
 
+/**
+ * Initialize a key part with default values.
+ *
+ * All trailing padding bytes are set to zero.
+ *
+ * All unknown <flags> bits are set to zero.
+ */
+API_EXPORT void
+box_key_part_def_create(box_key_part_def_t *part);
+
+/**
+ * Create a key_def from given key parts.
+ *
+ * Unlike <box_key_def_new>() this function allows to define
+ * nullability, collation and other options for each key part.
+ *
+ * <box_key_part_def_t> fields that are unknown at given tarantool
+ * version are ignored. The same for unknown <flags> bits.
+ */
+API_EXPORT box_key_def_t *
+box_key_def_new_ex(box_key_part_def_t *parts, uint32_t part_count);
+
 /**
  * Delete key definition
  *
@@ -94,6 +189,14 @@ box_tuple_compare_with_key(const box_tuple_t *tuple_a, const char *key_b,
 
 /** \endcond public */
 
+/*
+ * Size of the structure should remain the same across all
+ * tarantool versions in order to allow to allocate an array of
+ * them.
+ */
+static_assert(sizeof(box_key_part_def_t) == BOX_KEY_PART_DEF_T_SIZE,
+	      "sizeof(box_key_part_def_t)");
+
 #if defined(__cplusplus)
 } /* extern "C" */
 #endif /* defined(__cplusplus) */
-- 
2.25.0