]> git.ipfire.org Git - thirdparty/git.git/commitdiff
reftable: write reftable files
authorHan-Wen Nienhuys <hanwen@google.com>
Thu, 7 Oct 2021 20:25:06 +0000 (20:25 +0000)
committerJunio C Hamano <gitster@pobox.com>
Fri, 8 Oct 2021 17:45:48 +0000 (10:45 -0700)
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Makefile
reftable/reftable-writer.h [new file with mode: 0644]
reftable/writer.c [new file with mode: 0644]
reftable/writer.h [new file with mode: 0644]

index 2e5a9f40ed1365db515fafaae7f2dfdf163961a5..05be43355a6a1501df85bd99213ed859e3978c02 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -2463,6 +2463,7 @@ REFTABLE_OBJS += reftable/blocksource.o
 REFTABLE_OBJS += reftable/publicbasics.o
 REFTABLE_OBJS += reftable/record.o
 REFTABLE_OBJS += reftable/tree.o
+REFTABLE_OBJS += reftable/writer.o
 
 REFTABLE_TEST_OBJS += reftable/basics_test.o
 REFTABLE_TEST_OBJS += reftable/block_test.o
diff --git a/reftable/reftable-writer.h b/reftable/reftable-writer.h
new file mode 100644 (file)
index 0000000..af36462
--- /dev/null
@@ -0,0 +1,148 @@
+/*
+Copyright 2020 Google LLC
+
+Use of this source code is governed by a BSD-style
+license that can be found in the LICENSE file or at
+https://developers.google.com/open-source/licenses/bsd
+*/
+
+#ifndef REFTABLE_WRITER_H
+#define REFTABLE_WRITER_H
+
+#include "reftable-record.h"
+
+#include <stdint.h>
+#include <unistd.h> /* ssize_t */
+
+/* Writing single reftables */
+
+/* reftable_write_options sets options for writing a single reftable. */
+struct reftable_write_options {
+       /* boolean: do not pad out blocks to block size. */
+       unsigned unpadded : 1;
+
+       /* the blocksize. Should be less than 2^24. */
+       uint32_t block_size;
+
+       /* boolean: do not generate a SHA1 => ref index. */
+       unsigned skip_index_objects : 1;
+
+       /* how often to write complete keys in each block. */
+       int restart_interval;
+
+       /* 4-byte identifier ("sha1", "s256") of the hash.
+        * Defaults to SHA1 if unset
+        */
+       uint32_t hash_id;
+
+       /* boolean: do not check ref names for validity or dir/file conflicts.
+        */
+       unsigned skip_name_check : 1;
+
+       /* boolean: copy log messages exactly. If unset, check that the message
+        *   is a single line, and add '\n' if missing.
+        */
+       unsigned exact_log_message : 1;
+};
+
+/* reftable_block_stats holds statistics for a single block type */
+struct reftable_block_stats {
+       /* total number of entries written */
+       int entries;
+       /* total number of key restarts */
+       int restarts;
+       /* total number of blocks */
+       int blocks;
+       /* total number of index blocks */
+       int index_blocks;
+       /* depth of the index */
+       int max_index_level;
+
+       /* offset of the first block for this type */
+       uint64_t offset;
+       /* offset of the top level index block for this type, or 0 if not
+        * present */
+       uint64_t index_offset;
+};
+
+/* stats holds overall statistics for a single reftable */
+struct reftable_stats {
+       /* total number of blocks written. */
+       int blocks;
+       /* stats for ref data */
+       struct reftable_block_stats ref_stats;
+       /* stats for the SHA1 to ref map. */
+       struct reftable_block_stats obj_stats;
+       /* stats for index blocks */
+       struct reftable_block_stats idx_stats;
+       /* stats for log blocks */
+       struct reftable_block_stats log_stats;
+
+       /* disambiguation length of shortened object IDs. */
+       int object_id_len;
+};
+
+/* reftable_new_writer creates a new writer */
+struct reftable_writer *
+reftable_new_writer(ssize_t (*writer_func)(void *, const void *, size_t),
+                   void *writer_arg, struct reftable_write_options *opts);
+
+/* Set the range of update indices for the records we will add. When writing a
+   table into a stack, the min should be at least
+   reftable_stack_next_update_index(), or REFTABLE_API_ERROR is returned.
+
+   For transactional updates to a stack, typically min==max, and the
+   update_index can be obtained by inspeciting the stack. When converting an
+   existing ref database into a single reftable, this would be a range of
+   update-index timestamps.
+ */
+void reftable_writer_set_limits(struct reftable_writer *w, uint64_t min,
+                               uint64_t max);
+
+/*
+  Add a reftable_ref_record. The record should have names that come after
+  already added records.
+
+  The update_index must be within the limits set by
+  reftable_writer_set_limits(), or REFTABLE_API_ERROR is returned. It is an
+  REFTABLE_API_ERROR error to write a ref record after a log record.
+*/
+int reftable_writer_add_ref(struct reftable_writer *w,
+                           struct reftable_ref_record *ref);
+
+/*
+  Convenience function to add multiple reftable_ref_records; the function sorts
+  the records before adding them, reordering the records array passed in.
+*/
+int reftable_writer_add_refs(struct reftable_writer *w,
+                            struct reftable_ref_record *refs, int n);
+
+/*
+  adds reftable_log_records. Log records are keyed by (refname, decreasing
+  update_index). The key for the record added must come after the already added
+  log records.
+*/
+int reftable_writer_add_log(struct reftable_writer *w,
+                           struct reftable_log_record *log);
+
+/*
+  Convenience function to add multiple reftable_log_records; the function sorts
+  the records before adding them, reordering records array passed in.
+*/
+int reftable_writer_add_logs(struct reftable_writer *w,
+                            struct reftable_log_record *logs, int n);
+
+/* reftable_writer_close finalizes the reftable. The writer is retained so
+ * statistics can be inspected. */
+int reftable_writer_close(struct reftable_writer *w);
+
+/* writer_stats returns the statistics on the reftable being written.
+
+   This struct becomes invalid when the writer is freed.
+ */
+const struct reftable_stats *writer_stats(struct reftable_writer *w);
+
+/* reftable_writer_free deallocates memory for the writer */
+void reftable_writer_free(struct reftable_writer *w);
+
+#endif
diff --git a/reftable/writer.c b/reftable/writer.c
new file mode 100644 (file)
index 0000000..3ca721e
--- /dev/null
@@ -0,0 +1,690 @@
+/*
+Copyright 2020 Google LLC
+
+Use of this source code is governed by a BSD-style
+license that can be found in the LICENSE file or at
+https://developers.google.com/open-source/licenses/bsd
+*/
+
+#include "writer.h"
+
+#include "system.h"
+
+#include "block.h"
+#include "constants.h"
+#include "record.h"
+#include "tree.h"
+#include "reftable-error.h"
+
+/* finishes a block, and writes it to storage */
+static int writer_flush_block(struct reftable_writer *w);
+
+/* deallocates memory related to the index */
+static void writer_clear_index(struct reftable_writer *w);
+
+/* finishes writing a 'r' (refs) or 'g' (reflogs) section */
+static int writer_finish_public_section(struct reftable_writer *w);
+
+static struct reftable_block_stats *
+writer_reftable_block_stats(struct reftable_writer *w, uint8_t typ)
+{
+       switch (typ) {
+       case 'r':
+               return &w->stats.ref_stats;
+       case 'o':
+               return &w->stats.obj_stats;
+       case 'i':
+               return &w->stats.idx_stats;
+       case 'g':
+               return &w->stats.log_stats;
+       }
+       abort();
+       return NULL;
+}
+
+/* write data, queuing the padding for the next write. Returns negative for
+ * error. */
+static int padded_write(struct reftable_writer *w, uint8_t *data, size_t len,
+                       int padding)
+{
+       int n = 0;
+       if (w->pending_padding > 0) {
+               uint8_t *zeroed = reftable_calloc(w->pending_padding);
+               int n = w->write(w->write_arg, zeroed, w->pending_padding);
+               if (n < 0)
+                       return n;
+
+               w->pending_padding = 0;
+               reftable_free(zeroed);
+       }
+
+       w->pending_padding = padding;
+       n = w->write(w->write_arg, data, len);
+       if (n < 0)
+               return n;
+       n += padding;
+       return 0;
+}
+
+static void options_set_defaults(struct reftable_write_options *opts)
+{
+       if (opts->restart_interval == 0) {
+               opts->restart_interval = 16;
+       }
+
+       if (opts->hash_id == 0) {
+               opts->hash_id = GIT_SHA1_FORMAT_ID;
+       }
+       if (opts->block_size == 0) {
+               opts->block_size = DEFAULT_BLOCK_SIZE;
+       }
+}
+
+static int writer_version(struct reftable_writer *w)
+{
+       return (w->opts.hash_id == 0 || w->opts.hash_id == GIT_SHA1_FORMAT_ID) ?
+                            1 :
+                            2;
+}
+
+static int writer_write_header(struct reftable_writer *w, uint8_t *dest)
+{
+       memcpy(dest, "REFT", 4);
+
+       dest[4] = writer_version(w);
+
+       put_be24(dest + 5, w->opts.block_size);
+       put_be64(dest + 8, w->min_update_index);
+       put_be64(dest + 16, w->max_update_index);
+       if (writer_version(w) == 2) {
+               put_be32(dest + 24, w->opts.hash_id);
+       }
+       return header_size(writer_version(w));
+}
+
+static void writer_reinit_block_writer(struct reftable_writer *w, uint8_t typ)
+{
+       int block_start = 0;
+       if (w->next == 0) {
+               block_start = header_size(writer_version(w));
+       }
+
+       strbuf_release(&w->last_key);
+       block_writer_init(&w->block_writer_data, typ, w->block,
+                         w->opts.block_size, block_start,
+                         hash_size(w->opts.hash_id));
+       w->block_writer = &w->block_writer_data;
+       w->block_writer->restart_interval = w->opts.restart_interval;
+}
+
+static struct strbuf reftable_empty_strbuf = STRBUF_INIT;
+
+struct reftable_writer *
+reftable_new_writer(ssize_t (*writer_func)(void *, const void *, size_t),
+                   void *writer_arg, struct reftable_write_options *opts)
+{
+       struct reftable_writer *wp =
+               reftable_calloc(sizeof(struct reftable_writer));
+       strbuf_init(&wp->block_writer_data.last_key, 0);
+       options_set_defaults(opts);
+       if (opts->block_size >= (1 << 24)) {
+               /* TODO - error return? */
+               abort();
+       }
+       wp->last_key = reftable_empty_strbuf;
+       wp->block = reftable_calloc(opts->block_size);
+       wp->write = writer_func;
+       wp->write_arg = writer_arg;
+       wp->opts = *opts;
+       writer_reinit_block_writer(wp, BLOCK_TYPE_REF);
+
+       return wp;
+}
+
+void reftable_writer_set_limits(struct reftable_writer *w, uint64_t min,
+                               uint64_t max)
+{
+       w->min_update_index = min;
+       w->max_update_index = max;
+}
+
+void reftable_writer_free(struct reftable_writer *w)
+{
+       reftable_free(w->block);
+       reftable_free(w);
+}
+
+struct obj_index_tree_node {
+       struct strbuf hash;
+       uint64_t *offsets;
+       size_t offset_len;
+       size_t offset_cap;
+};
+
+#define OBJ_INDEX_TREE_NODE_INIT    \
+       {                           \
+               .hash = STRBUF_INIT \
+       }
+
+static int obj_index_tree_node_compare(const void *a, const void *b)
+{
+       return strbuf_cmp(&((const struct obj_index_tree_node *)a)->hash,
+                         &((const struct obj_index_tree_node *)b)->hash);
+}
+
+static void writer_index_hash(struct reftable_writer *w, struct strbuf *hash)
+{
+       uint64_t off = w->next;
+
+       struct obj_index_tree_node want = { .hash = *hash };
+
+       struct tree_node *node = tree_search(&want, &w->obj_index_tree,
+                                            &obj_index_tree_node_compare, 0);
+       struct obj_index_tree_node *key = NULL;
+       if (node == NULL) {
+               struct obj_index_tree_node empty = OBJ_INDEX_TREE_NODE_INIT;
+               key = reftable_malloc(sizeof(struct obj_index_tree_node));
+               *key = empty;
+
+               strbuf_reset(&key->hash);
+               strbuf_addbuf(&key->hash, hash);
+               tree_search((void *)key, &w->obj_index_tree,
+                           &obj_index_tree_node_compare, 1);
+       } else {
+               key = node->key;
+       }
+
+       if (key->offset_len > 0 && key->offsets[key->offset_len - 1] == off) {
+               return;
+       }
+
+       if (key->offset_len == key->offset_cap) {
+               key->offset_cap = 2 * key->offset_cap + 1;
+               key->offsets = reftable_realloc(
+                       key->offsets, sizeof(uint64_t) * key->offset_cap);
+       }
+
+       key->offsets[key->offset_len++] = off;
+}
+
+static int writer_add_record(struct reftable_writer *w,
+                            struct reftable_record *rec)
+{
+       struct strbuf key = STRBUF_INIT;
+       int err = -1;
+       reftable_record_key(rec, &key);
+       if (strbuf_cmp(&w->last_key, &key) >= 0) {
+               err = REFTABLE_API_ERROR;
+               goto done;
+       }
+
+       strbuf_reset(&w->last_key);
+       strbuf_addbuf(&w->last_key, &key);
+       if (w->block_writer == NULL) {
+               writer_reinit_block_writer(w, reftable_record_type(rec));
+       }
+
+       assert(block_writer_type(w->block_writer) == reftable_record_type(rec));
+
+       if (block_writer_add(w->block_writer, rec) == 0) {
+               err = 0;
+               goto done;
+       }
+
+       err = writer_flush_block(w);
+       if (err < 0) {
+               goto done;
+       }
+
+       writer_reinit_block_writer(w, reftable_record_type(rec));
+       err = block_writer_add(w->block_writer, rec);
+       if (err < 0) {
+               goto done;
+       }
+
+       err = 0;
+done:
+       strbuf_release(&key);
+       return err;
+}
+
+int reftable_writer_add_ref(struct reftable_writer *w,
+                           struct reftable_ref_record *ref)
+{
+       struct reftable_record rec = { NULL };
+       struct reftable_ref_record copy = *ref;
+       int err = 0;
+
+       if (ref->refname == NULL)
+               return REFTABLE_API_ERROR;
+       if (ref->update_index < w->min_update_index ||
+           ref->update_index > w->max_update_index)
+               return REFTABLE_API_ERROR;
+
+       reftable_record_from_ref(&rec, &copy);
+       copy.update_index -= w->min_update_index;
+
+       err = writer_add_record(w, &rec);
+       if (err < 0)
+               return err;
+
+       if (!w->opts.skip_index_objects && reftable_ref_record_val1(ref)) {
+               struct strbuf h = STRBUF_INIT;
+               strbuf_add(&h, (char *)reftable_ref_record_val1(ref),
+                          hash_size(w->opts.hash_id));
+               writer_index_hash(w, &h);
+               strbuf_release(&h);
+       }
+
+       if (!w->opts.skip_index_objects && reftable_ref_record_val2(ref)) {
+               struct strbuf h = STRBUF_INIT;
+               strbuf_add(&h, reftable_ref_record_val2(ref),
+                          hash_size(w->opts.hash_id));
+               writer_index_hash(w, &h);
+               strbuf_release(&h);
+       }
+       return 0;
+}
+
+int reftable_writer_add_refs(struct reftable_writer *w,
+                            struct reftable_ref_record *refs, int n)
+{
+       int err = 0;
+       int i = 0;
+       QSORT(refs, n, reftable_ref_record_compare_name);
+       for (i = 0; err == 0 && i < n; i++) {
+               err = reftable_writer_add_ref(w, &refs[i]);
+       }
+       return err;
+}
+
+static int reftable_writer_add_log_verbatim(struct reftable_writer *w,
+                                           struct reftable_log_record *log)
+{
+       struct reftable_record rec = { NULL };
+       if (w->block_writer &&
+           block_writer_type(w->block_writer) == BLOCK_TYPE_REF) {
+               int err = writer_finish_public_section(w);
+               if (err < 0)
+                       return err;
+       }
+
+       w->next -= w->pending_padding;
+       w->pending_padding = 0;
+
+       reftable_record_from_log(&rec, log);
+       return writer_add_record(w, &rec);
+}
+
+int reftable_writer_add_log(struct reftable_writer *w,
+                           struct reftable_log_record *log)
+{
+       char *input_log_message = NULL;
+       struct strbuf cleaned_message = STRBUF_INIT;
+       int err = 0;
+
+       if (log->value_type == REFTABLE_LOG_DELETION)
+               return reftable_writer_add_log_verbatim(w, log);
+
+       if (log->refname == NULL)
+               return REFTABLE_API_ERROR;
+
+       input_log_message = log->value.update.message;
+       if (!w->opts.exact_log_message && log->value.update.message) {
+               strbuf_addstr(&cleaned_message, log->value.update.message);
+               while (cleaned_message.len &&
+                      cleaned_message.buf[cleaned_message.len - 1] == '\n')
+                       strbuf_setlen(&cleaned_message,
+                                     cleaned_message.len - 1);
+               if (strchr(cleaned_message.buf, '\n')) {
+                       /* multiple lines not allowed. */
+                       err = REFTABLE_API_ERROR;
+                       goto done;
+               }
+               strbuf_addstr(&cleaned_message, "\n");
+               log->value.update.message = cleaned_message.buf;
+       }
+
+       err = reftable_writer_add_log_verbatim(w, log);
+       log->value.update.message = input_log_message;
+done:
+       strbuf_release(&cleaned_message);
+       return err;
+}
+
+int reftable_writer_add_logs(struct reftable_writer *w,
+                            struct reftable_log_record *logs, int n)
+{
+       int err = 0;
+       int i = 0;
+       QSORT(logs, n, reftable_log_record_compare_key);
+
+       for (i = 0; err == 0 && i < n; i++) {
+               err = reftable_writer_add_log(w, &logs[i]);
+       }
+       return err;
+}
+
+static int writer_finish_section(struct reftable_writer *w)
+{
+       uint8_t typ = block_writer_type(w->block_writer);
+       uint64_t index_start = 0;
+       int max_level = 0;
+       int threshold = w->opts.unpadded ? 1 : 3;
+       int before_blocks = w->stats.idx_stats.blocks;
+       int err = writer_flush_block(w);
+       int i = 0;
+       struct reftable_block_stats *bstats = NULL;
+       if (err < 0)
+               return err;
+
+       while (w->index_len > threshold) {
+               struct reftable_index_record *idx = NULL;
+               int idx_len = 0;
+
+               max_level++;
+               index_start = w->next;
+               writer_reinit_block_writer(w, BLOCK_TYPE_INDEX);
+
+               idx = w->index;
+               idx_len = w->index_len;
+
+               w->index = NULL;
+               w->index_len = 0;
+               w->index_cap = 0;
+               for (i = 0; i < idx_len; i++) {
+                       struct reftable_record rec = { NULL };
+                       reftable_record_from_index(&rec, idx + i);
+                       if (block_writer_add(w->block_writer, &rec) == 0) {
+                               continue;
+                       }
+
+                       err = writer_flush_block(w);
+                       if (err < 0)
+                               return err;
+
+                       writer_reinit_block_writer(w, BLOCK_TYPE_INDEX);
+
+                       err = block_writer_add(w->block_writer, &rec);
+                       if (err != 0) {
+                               /* write into fresh block should always succeed
+                                */
+                               abort();
+                       }
+               }
+               for (i = 0; i < idx_len; i++) {
+                       strbuf_release(&idx[i].last_key);
+               }
+               reftable_free(idx);
+       }
+
+       writer_clear_index(w);
+
+       err = writer_flush_block(w);
+       if (err < 0)
+               return err;
+
+       bstats = writer_reftable_block_stats(w, typ);
+       bstats->index_blocks = w->stats.idx_stats.blocks - before_blocks;
+       bstats->index_offset = index_start;
+       bstats->max_index_level = max_level;
+
+       /* Reinit lastKey, as the next section can start with any key. */
+       w->last_key.len = 0;
+
+       return 0;
+}
+
+struct common_prefix_arg {
+       struct strbuf *last;
+       int max;
+};
+
+static void update_common(void *void_arg, void *key)
+{
+       struct common_prefix_arg *arg = void_arg;
+       struct obj_index_tree_node *entry = key;
+       if (arg->last) {
+               int n = common_prefix_size(&entry->hash, arg->last);
+               if (n > arg->max) {
+                       arg->max = n;
+               }
+       }
+       arg->last = &entry->hash;
+}
+
+struct write_record_arg {
+       struct reftable_writer *w;
+       int err;
+};
+
+static void write_object_record(void *void_arg, void *key)
+{
+       struct write_record_arg *arg = void_arg;
+       struct obj_index_tree_node *entry = key;
+       struct reftable_obj_record obj_rec = {
+               .hash_prefix = (uint8_t *)entry->hash.buf,
+               .hash_prefix_len = arg->w->stats.object_id_len,
+               .offsets = entry->offsets,
+               .offset_len = entry->offset_len,
+       };
+       struct reftable_record rec = { NULL };
+       if (arg->err < 0)
+               goto done;
+
+       reftable_record_from_obj(&rec, &obj_rec);
+       arg->err = block_writer_add(arg->w->block_writer, &rec);
+       if (arg->err == 0)
+               goto done;
+
+       arg->err = writer_flush_block(arg->w);
+       if (arg->err < 0)
+               goto done;
+
+       writer_reinit_block_writer(arg->w, BLOCK_TYPE_OBJ);
+       arg->err = block_writer_add(arg->w->block_writer, &rec);
+       if (arg->err == 0)
+               goto done;
+       obj_rec.offset_len = 0;
+       arg->err = block_writer_add(arg->w->block_writer, &rec);
+
+       /* Should be able to write into a fresh block. */
+       assert(arg->err == 0);
+
+done:;
+}
+
+static void object_record_free(void *void_arg, void *key)
+{
+       struct obj_index_tree_node *entry = key;
+
+       FREE_AND_NULL(entry->offsets);
+       strbuf_release(&entry->hash);
+       reftable_free(entry);
+}
+
+static int writer_dump_object_index(struct reftable_writer *w)
+{
+       struct write_record_arg closure = { .w = w };
+       struct common_prefix_arg common = { NULL };
+       if (w->obj_index_tree) {
+               infix_walk(w->obj_index_tree, &update_common, &common);
+       }
+       w->stats.object_id_len = common.max + 1;
+
+       writer_reinit_block_writer(w, BLOCK_TYPE_OBJ);
+
+       if (w->obj_index_tree) {
+               infix_walk(w->obj_index_tree, &write_object_record, &closure);
+       }
+
+       if (closure.err < 0)
+               return closure.err;
+       return writer_finish_section(w);
+}
+
+static int writer_finish_public_section(struct reftable_writer *w)
+{
+       uint8_t typ = 0;
+       int err = 0;
+
+       if (w->block_writer == NULL)
+               return 0;
+
+       typ = block_writer_type(w->block_writer);
+       err = writer_finish_section(w);
+       if (err < 0)
+               return err;
+       if (typ == BLOCK_TYPE_REF && !w->opts.skip_index_objects &&
+           w->stats.ref_stats.index_blocks > 0) {
+               err = writer_dump_object_index(w);
+               if (err < 0)
+                       return err;
+       }
+
+       if (w->obj_index_tree) {
+               infix_walk(w->obj_index_tree, &object_record_free, NULL);
+               tree_free(w->obj_index_tree);
+               w->obj_index_tree = NULL;
+       }
+
+       w->block_writer = NULL;
+       return 0;
+}
+
+int reftable_writer_close(struct reftable_writer *w)
+{
+       uint8_t footer[72];
+       uint8_t *p = footer;
+       int err = writer_finish_public_section(w);
+       int empty_table = w->next == 0;
+       if (err != 0)
+               goto done;
+       w->pending_padding = 0;
+       if (empty_table) {
+               /* Empty tables need a header anyway. */
+               uint8_t header[28];
+               int n = writer_write_header(w, header);
+               err = padded_write(w, header, n, 0);
+               if (err < 0)
+                       goto done;
+       }
+
+       p += writer_write_header(w, footer);
+       put_be64(p, w->stats.ref_stats.index_offset);
+       p += 8;
+       put_be64(p, (w->stats.obj_stats.offset) << 5 | w->stats.object_id_len);
+       p += 8;
+       put_be64(p, w->stats.obj_stats.index_offset);
+       p += 8;
+
+       put_be64(p, w->stats.log_stats.offset);
+       p += 8;
+       put_be64(p, w->stats.log_stats.index_offset);
+       p += 8;
+
+       put_be32(p, crc32(0, footer, p - footer));
+       p += 4;
+
+       err = padded_write(w, footer, footer_size(writer_version(w)), 0);
+       if (err < 0)
+               goto done;
+
+       if (empty_table) {
+               err = REFTABLE_EMPTY_TABLE_ERROR;
+               goto done;
+       }
+
+done:
+       /* free up memory. */
+       block_writer_release(&w->block_writer_data);
+       writer_clear_index(w);
+       strbuf_release(&w->last_key);
+       return err;
+}
+
+static void writer_clear_index(struct reftable_writer *w)
+{
+       int i = 0;
+       for (i = 0; i < w->index_len; i++) {
+               strbuf_release(&w->index[i].last_key);
+       }
+
+       FREE_AND_NULL(w->index);
+       w->index_len = 0;
+       w->index_cap = 0;
+}
+
+static const int debug = 0;
+
+static int writer_flush_nonempty_block(struct reftable_writer *w)
+{
+       uint8_t typ = block_writer_type(w->block_writer);
+       struct reftable_block_stats *bstats =
+               writer_reftable_block_stats(w, typ);
+       uint64_t block_typ_off = (bstats->blocks == 0) ? w->next : 0;
+       int raw_bytes = block_writer_finish(w->block_writer);
+       int padding = 0;
+       int err = 0;
+       struct reftable_index_record ir = { .last_key = STRBUF_INIT };
+       if (raw_bytes < 0)
+               return raw_bytes;
+
+       if (!w->opts.unpadded && typ != BLOCK_TYPE_LOG) {
+               padding = w->opts.block_size - raw_bytes;
+       }
+
+       if (block_typ_off > 0) {
+               bstats->offset = block_typ_off;
+       }
+
+       bstats->entries += w->block_writer->entries;
+       bstats->restarts += w->block_writer->restart_len;
+       bstats->blocks++;
+       w->stats.blocks++;
+
+       if (debug) {
+               fprintf(stderr, "block %c off %" PRIu64 " sz %d (%d)\n", typ,
+                       w->next, raw_bytes,
+                       get_be24(w->block + w->block_writer->header_off + 1));
+       }
+
+       if (w->next == 0) {
+               writer_write_header(w, w->block);
+       }
+
+       err = padded_write(w, w->block, raw_bytes, padding);
+       if (err < 0)
+               return err;
+
+       if (w->index_cap == w->index_len) {
+               w->index_cap = 2 * w->index_cap + 1;
+               w->index = reftable_realloc(
+                       w->index,
+                       sizeof(struct reftable_index_record) * w->index_cap);
+       }
+
+       ir.offset = w->next;
+       strbuf_reset(&ir.last_key);
+       strbuf_addbuf(&ir.last_key, &w->block_writer->last_key);
+       w->index[w->index_len] = ir;
+
+       w->index_len++;
+       w->next += padding + raw_bytes;
+       w->block_writer = NULL;
+       return 0;
+}
+
+static int writer_flush_block(struct reftable_writer *w)
+{
+       if (w->block_writer == NULL)
+               return 0;
+       if (w->block_writer->entries == 0)
+               return 0;
+       return writer_flush_nonempty_block(w);
+}
+
+const struct reftable_stats *writer_stats(struct reftable_writer *w)
+{
+       return &w->stats;
+}
diff --git a/reftable/writer.h b/reftable/writer.h
new file mode 100644 (file)
index 0000000..09b8867
--- /dev/null
@@ -0,0 +1,50 @@
+/*
+Copyright 2020 Google LLC
+
+Use of this source code is governed by a BSD-style
+license that can be found in the LICENSE file or at
+https://developers.google.com/open-source/licenses/bsd
+*/
+
+#ifndef WRITER_H
+#define WRITER_H
+
+#include "basics.h"
+#include "block.h"
+#include "tree.h"
+#include "reftable-writer.h"
+
+struct reftable_writer {
+       ssize_t (*write)(void *, const void *, size_t);
+       void *write_arg;
+       int pending_padding;
+       struct strbuf last_key;
+
+       /* offset of next block to write. */
+       uint64_t next;
+       uint64_t min_update_index, max_update_index;
+       struct reftable_write_options opts;
+
+       /* memory buffer for writing */
+       uint8_t *block;
+
+       /* writer for the current section. NULL or points to
+        * block_writer_data */
+       struct block_writer *block_writer;
+
+       struct block_writer block_writer_data;
+
+       /* pending index records for the current section */
+       struct reftable_index_record *index;
+       size_t index_len;
+       size_t index_cap;
+
+       /*
+        * tree for use with tsearch; used to populate the 'o' inverse OID
+        * map */
+       struct tree_node *obj_index_tree;
+
+       struct reftable_stats stats;
+};
+
+#endif