midx: Support multi-pack-index files in odb_pack.c

This change adds support for reading multi-pack-index files from the
packfile odb backend. This also makes git_pack_file objects open their
backing failes lazily in more scenarios, since the multi-pack-index can
avoid having to open them in some cases (yay!).

This change also refreshes the documentation found in src/odb_pack.c to
match the updated code.

Part of: #5399

src/midx.c

@@ -13,8 +13,6 @@
 #include "odb.h"
 #include "pack.h"
 
-#define GIT_MIDX_FILE_MODE 0444
-
 #define MIDX_SIGNATURE 0x4d494458 /* "MIDX" */
 #define MIDX_VERSION 1
 #define MIDX_OBJECT_ID_VERSION 1
@@ -116,7 +114,7 @@ static int midx_parse_oid_lookup(
 		return midx_error("missing OID Lookup chunk");
 	if (chunk_oid_lookup->length == 0)
 		return midx_error("empty OID Lookup chunk");
-	if (chunk_oid_lookup->length != idx->num_objects * 20)
+	if (chunk_oid_lookup->length != idx->num_objects * GIT_OID_RAWSZ)
 		return midx_error("OID Lookup chunk has wrong length");
 
 	idx->oid_lookup = oid = (git_oid *)(data + chunk_oid_lookup->offset);
@@ -183,7 +181,7 @@ int git_midx_parse(
 
 	GIT_ASSERT_ARG(idx);
 
-	if (size < sizeof(struct git_midx_header) + 20)
+	if (size < sizeof(struct git_midx_header) + GIT_OID_RAWSZ)
 		return midx_error("multi-pack index is too short");
 
 	hdr = ((struct git_midx_header *)data);
@@ -203,7 +201,7 @@ int git_midx_parse(
 	last_chunk_offset =
 			sizeof(struct git_midx_header) +
 			(1 + hdr->chunks) * 12;
-	trailer_offset = size - 20;
+	trailer_offset = size - GIT_OID_RAWSZ;
 	if (trailer_offset < last_chunk_offset)
 		return midx_error("wrong index size");
 	git_oid_cpy(&idx->checksum, (git_oid *)(data + trailer_offset));
@@ -309,6 +307,10 @@ int git_midx_open(
 	idx = git__calloc(1, sizeof(git_midx_file));
 	GIT_ERROR_CHECK_ALLOC(idx);
 
+	error = git_buf_sets(&idx->filename, path);
+	if (error < 0)
+		return error;
+
 	error = git_futils_mmap_ro(&idx->index_map, fd, 0, idx_size);
 	p_close(fd);
 	if (error < 0) {
@@ -325,6 +327,46 @@ int git_midx_open(
 	return 0;
 }
 
+bool git_midx_needs_refresh(
+		const git_midx_file *idx,
+		const char *path)
+{
+	git_file fd = -1;
+	struct stat st;
+	ssize_t bytes_read;
+	git_oid idx_checksum = {{0}};
+
+	/* TODO: properly open the file without access time using O_NOATIME */
+	fd = git_futils_open_ro(path);
+	if (fd < 0)
+		return true;
+
+	if (p_fstat(fd, &st) < 0) {
+		p_close(fd);
+		return true;
+	}
+
+	if (!S_ISREG(st.st_mode) ||
+	    !git__is_sizet(st.st_size) ||
+	    (size_t)st.st_size != idx->index_map.len) {
+		p_close(fd);
+		return true;
+	}
+
+	if (p_lseek(fd, -GIT_OID_RAWSZ, SEEK_END) < 0) {
+		p_close(fd);
+		return true;
+	}
+
+	bytes_read = p_read(fd, &idx_checksum, GIT_OID_RAWSZ);
+	p_close(fd);
+
+	if (bytes_read != GIT_OID_RAWSZ)
+		return true;
+
+	return git_oid_cmp(&idx_checksum, &idx->checksum) == 0;
+}
+
 int git_midx_entry_find(
 		git_midx_entry *e,
 		git_midx_file *idx,
@@ -343,7 +385,7 @@ int git_midx_entry_find(
 	hi = ntohl(idx->oid_fanout[(int)short_oid->id[0]]);
 	lo = ((short_oid->id[0] == 0x0) ? 0 : ntohl(idx->oid_fanout[(int)short_oid->id[0] - 1]));
 
-	pos = git_pack__lookup_sha1(idx->oid_lookup, 20, lo, hi, short_oid->id);
+	pos = git_pack__lookup_sha1(idx->oid_lookup, GIT_OID_RAWSZ, lo, hi, short_oid->id);
 
 	if (pos >= 0) {
 		/* An object matching exactly the oid was found */
@@ -399,6 +441,24 @@ int git_midx_entry_find(
 	return 0;
 }
 
+int git_midx_foreach_entry(
+		git_midx_file *idx,
+		git_odb_foreach_cb cb,
+		void *data)
+{
+	size_t i;
+	int error;
+
+	GIT_ASSERT_ARG(idx);
+
+	for (i = 0; i < idx->num_objects; ++i) {
+		if ((error = cb(&idx->oid_lookup[i], data)) != 0)
+			return git_error_set_after_callback(error);
+	}
+
+	return error;
+}
+
 int git_midx_close(git_midx_file *idx)
 {
 	GIT_ASSERT_ARG(idx);
@@ -416,6 +476,7 @@ void git_midx_free(git_midx_file *idx)
 	if (!idx)
 		return;
 
+	git_buf_dispose(&idx->filename);
 	git_midx_close(idx);
 	git__free(idx);
 }

src/midx.h

@@ -14,6 +14,7 @@
 
 #include "map.h"
 #include "mwindow.h"
+#include "odb.h"
 
 /*
  * A multi-pack-index file.
@@ -49,6 +50,9 @@ typedef struct git_midx_file {
 
 	/* The trailer of the file. Contains the SHA1-checksum of the whole file. */
 	git_oid checksum;
+
+	/* something like ".git/objects/pack/multi-pack-index". */
+	git_buf filename;
 } git_midx_file;
 
 /*
@@ -66,11 +70,18 @@ typedef struct git_midx_entry {
 int git_midx_open(
 		git_midx_file **idx_out,
 		const char *path);
+bool git_midx_needs_refresh(
+		const git_midx_file *idx,
+		const char *path);
 int git_midx_entry_find(
 		git_midx_entry *e,
 		git_midx_file *idx,
 		const git_oid *short_oid,
 		size_t len);
+int git_midx_foreach_entry(
+		git_midx_file *idx,
+		git_odb_foreach_cb cb,
+		void *data);
 int git_midx_close(git_midx_file *idx);
 void git_midx_free(git_midx_file *idx);
 

src/odb_pack.c

@@ -11,11 +11,12 @@
 #include "git2/repository.h"
 #include "git2/indexer.h"
 #include "git2/sys/odb_backend.h"
+#include "delta.h"
 #include "futils.h"
 #include "hash.h"
-#include "odb.h"
-#include "delta.h"
+#include "midx.h"
 #include "mwindow.h"
+#include "odb.h"
 #include "pack.h"
 
 #include "git2/odb_backend.h"
@@ -25,6 +26,8 @@
 
 struct pack_backend {
 	git_odb_backend parent;
+	git_midx_file *midx;
+	git_vector midx_packs;
 	git_vector packs;
 	struct git_pack_file *last_found;
 	char *pack_folder;
@@ -47,36 +50,43 @@ struct pack_writepack {
  *	Initialization of the Pack Backend
  *	--------------------------------------------------
  *
- *	# git_odb_backend_pack
- *	| Creates the pack backend structure, initializes the
- *	| callback pointers to our default read() and exist() methods,
- *	| and tries to preload all the known packfiles in the ODB.
+ * # git_odb_backend_pack
+ * | Creates the pack backend structure, initializes the
+ * | callback pointers to our default read() and exist() methods,
+ * | and tries to find the `pack` folder, if it exists. ODBs without a `pack`
+ * | folder are ignored altogether. If there is a `pack` folder, it tries to
+ * | preload all the known packfiles in the ODB.
  * |
- *	|-# packfile_load_all
- *	 | Tries to find the `pack` folder, if it exists. ODBs without
- *	 | a pack folder are ignored altogether. If there's a `pack` folder
- *	 | we run a `dirent` callback through every file in the pack folder
- *	 | to find our packfiles. The packfiles are then sorted according
- *	 | to a sorting callback.
- * 	 |
- *	 |-# packfile_load__cb
- *	 | | This callback is called from `dirent` with every single file
- *	 | | inside the pack folder. We find the packs by actually locating
- *	 | | their index (ends in ".idx"). From that index, we verify that
- *	 | | the corresponding packfile exists and is valid, and if so, we
- *	| | add it to the pack list.
- *	 | |
- *	 | |-# packfile_check
- *	 |		Make sure that there's a packfile to back this index, and store
- *	 |		some very basic information regarding the packfile itself,
- *	 |		such as the full path, the size, and the modification time.
- *	 |		We don't actually open the packfile to check for internal consistency.
- *	|
- *	|-# packfile_sort__cb
- *		Sort all the preloaded packs according to some specific criteria:
- *		we prioritize the "newer" packs because it's more likely they
- *		contain the objects we are looking for, and we prioritize local
- *		packs over remote ones.
+ * |-# pack_backend__refresh
+ *   | The `multi-pack-index` is loaded if it exists and is valid.
+ *   | Then we run a `dirent` callback through every file in the pack folder,
+ *   | even those present in `multi-pack-index`. The unindexed packfiles are
+ *   | then sorted according to a sorting callback.
+ *   |
+ *   |-# refresh_multi_pack_index
+ *   |   Detect the presence of the `multi-pack-index` file. If it needs to be
+ *   |   refreshed, frees the old copy and tries to load the new one, together
+ *   |   with all the packfiles it indexes. If the process fails, fall back to
+ *   |   the old behavior, as if the `multi-pack-index` file was not there.
+ *   |
+ *   |-# packfile_load__cb
+ *   | | This callback is called from `dirent` with every single file
+ *   | | inside the pack folder. We find the packs by actually locating
+ *   | | their index (ends in ".idx"). From that index, we verify that
+ *   | | the corresponding packfile exists and is valid, and if so, we
+ *   | | add it to the pack list.
+ *   | |
+ *   | # git_mwindow_get_pack
+ *   |   Make sure that there's a packfile to back this index, and store
+ *   |   some very basic information regarding the packfile itself,
+ *   |   such as the full path, the size, and the modification time.
+ *   |   We don't actually open the packfile to check for internal consistency.
+ *   |
+ *   |-# packfile_sort__cb
+ *       Sort all the preloaded packs according to some specific criteria:
+ *       we prioritize the "newer" packs because it's more likely they
+ *       contain the objects we are looking for, and we prioritize local
+ *       packs over remote ones.
  *
  *
  *
@@ -84,48 +94,66 @@ struct pack_writepack {
  *	A standard packed `exist` query for an OID
  *	--------------------------------------------------
  *
- * # pack_backend__exists
- * | Check if the given SHA1 oid exists in any of the packs
- * | that have been loaded for our ODB.
+ * # pack_backend__exists / pack_backend__exists_prefix
+ * | Check if the given SHA1 oid (or a SHA1 oid prefix) exists in any of the
+ * | packs that have been loaded for our ODB.
  * |
- * |-# pack_entry_find
- *	| Iterate through all the packs that have been preloaded
- *	| (starting by the pack where the latest object was found)
- *	| to try to find the OID in one of them.
- *	|
- *	|-# pack_entry_find1
- *		| Check the index of an individual pack to see if the SHA1
- *		| OID can be found. If we can find the offset to that SHA1
- *		| inside of the index, that means the object is contained
- *		| inside of the packfile and we can stop searching.
- *		| Before returning, we verify that the packfile behing the
- *		| index we are searching still exists on disk.
- *		|
- *		|-# pack_entry_find_offset
- *		| | Mmap the actual index file to disk if it hasn't been opened
- *		| | yet, and run a binary search through it to find the OID.
- *		| | See <http://book.git-scm.com/7_the_packfile.html> for specifics
- *		| | on the Packfile Index format and how do we find entries in it.
- *		| |
- *		| |-# pack_index_open
- *		|	| Guess the name of the index based on the full path to the
- *		|	| packfile, open it and verify its contents. Only if the index
- *		|	| has not been opened already.
- *		|	|
- *		|	|-# pack_index_check
- *		|		Mmap the index file and do a quick run through the header
- *		|		to guess the index version (right now we support v1 and v2),
- *		|		and to verify that the size of the index makes sense.
- *		|
- *		|-# packfile_open
- *			See `packfile_open` in Chapter 3
+ * |-# pack_entry_find / pack_entry_find_prefix
+ *   | If there is a multi-pack-index present, search the SHA1 oid in that
+ *   | index first. If it is not found there, iterate through all the unindexed
+ *   | packs that have been preloaded (starting by the pack where the latest
+ *   | object was found) to try to find the OID in one of them.
+ *   |
+ *   |-# git_midx_entry_find
+ *   |   Search for the SHA1 oid in the multi-pack-index. See
+ *   |   <https://github.com/git/git/blob/master/Documentation/technical/pack-format.txt>
+ *   |   for specifics on the multi-pack-index format and how do we find
+ *   |   entries in it.
+ *   |
+ *   |-# git_pack_entry_find
+ *     | Check the index of an individual unindexed pack to see if the SHA1
+ *     | OID can be found. If we can find the offset to that SHA1 inside of the
+ *     | index, that means the object is contained inside of the packfile and
+ *     | we can stop searching. Before returning, we verify that the
+ *     | packfile behing the index we are searching still exists on disk.
+ *     |
+ *     |-# pack_entry_find_offset
+ *       | Mmap the actual index file to disk if it hasn't been opened
+ *       | yet, and run a binary search through it to find the OID.
+ *       | See <https://github.com/git/git/blob/master/Documentation/technical/pack-format.txt>
+ *       | for specifics on the Packfile Index format and how do we find
+ *       | entries in it.
+ *       |
+ *       |-# pack_index_open
+ *         | Guess the name of the index based on the full path to the
+ *         | packfile, open it and verify its contents. Only if the index
+ *         | has not been opened already.
+ *         |
+ *         |-# pack_index_check
+ *             Mmap the index file and do a quick run through the header
+ *             to guess the index version (right now we support v1 and v2),
+ *             and to verify that the size of the index makes sense.
  *
  *
  *
  *	Chapter 3: The neverending story...
  *	A standard packed `lookup` query for an OID
  *	--------------------------------------------------
- *	TODO
+ *
+ * # pack_backend__read / pack_backend__read_prefix
+ * | Check if the given SHA1 oid (or a SHA1 oid prefix) exists in any of the
+ * | packs that have been loaded for our ODB. If it does, open the packfile and
+ * | read from it.
+ * |
+ * |-# git_packfile_unpack
+ *     Armed with a packfile and the offset within it, we can finally unpack
+ *     the object pointed at by the SHA1 oid. This involves mmapping part of
+ *     the `.pack` file, and uncompressing the object within it (if it is
+ *     stored in the undelfitied representation), or finding a base object and
+ *     applying some deltas to its uncompressed representation (if it is stored
+ *     in the deltified representation). See
+ *     <https://github.com/git/git/blob/master/Documentation/technical/pack-format.txt>
+ *     for specifics on the Packfile format and how do we read from it.
  *
  */
 
@@ -140,6 +168,8 @@ static int packfile_sort__cb(const void *a_, const void *b_);
 
 static int packfile_load__cb(void *_data, git_buf *path);
 
+static int packfile_byname_search_cmp(const void *path, const void *pack_entry);
+
 static int pack_entry_find(struct git_pack_entry *e,
 	struct pack_backend *backend, const git_oid *oid);
 
@@ -163,6 +193,14 @@ static int pack_entry_find_prefix(
  *
  ***********************************************************/
 
+static int packfile_byname_search_cmp(const void *path_, const void *p_)
+{
+	const git_buf *path = (const git_buf *)path_;
+	const struct git_pack_file *p = (const struct git_pack_file *)p_;
+
+	return strncmp(p->pack_name, git_buf_cstr(path), git_buf_len(path));
+}
+
 static int packfile_sort__cb(const void *a_, const void *b_)
 {
 	const struct git_pack_file *a = a_;
@@ -198,20 +236,20 @@ static int packfile_load__cb(void *data, git_buf *path)
 	struct pack_backend *backend = data;
 	struct git_pack_file *pack;
 	const char *path_str = git_buf_cstr(path);
-	size_t i, cmp_len = git_buf_len(path);
+	git_buf index_prefix = GIT_BUF_INIT;
+	size_t cmp_len = git_buf_len(path);
 	int error;
 
 	if (cmp_len <= strlen(".idx") || git__suffixcmp(path_str, ".idx") != 0)
 		return 0; /* not an index */
 
 	cmp_len -= strlen(".idx");
+	git_buf_attach_notowned(&index_prefix, path_str, cmp_len);
 
-	for (i = 0; i < backend->packs.length; ++i) {
-		struct git_pack_file *p = git_vector_get(&backend->packs, i);
-
-		if (strncmp(p->pack_name, path_str, cmp_len) == 0)
-			return 0;
-	}
+	if (git_vector_search2(NULL, &backend->midx_packs, packfile_byname_search_cmp, &index_prefix) == 0)
+		return 0;
+	if (git_vector_search2(NULL, &backend->packs, packfile_byname_search_cmp, &index_prefix) == 0)
+		return 0;
 
 	error = git_mwindow_get_pack(&pack, path->ptr);
 
@@ -228,22 +266,26 @@ static int packfile_load__cb(void *data, git_buf *path)
 
 }
 
-static int pack_entry_find_inner(
-	struct git_pack_entry *e,
-	struct pack_backend *backend,
-	const git_oid *oid,
-	struct git_pack_file *last_found)
+static int pack_entry_find(struct git_pack_entry *e, struct pack_backend *backend, const git_oid *oid)
 {
+	struct git_pack_file *last_found = backend->last_found, *p;
+	git_midx_entry midx_entry;
 	size_t i;
 
+	if (backend->midx &&
+		git_midx_entry_find(&midx_entry, backend->midx, oid, GIT_OID_HEXSZ) == 0 &&
+		midx_entry.pack_index < git_vector_length(&backend->midx_packs)) {
+		e->offset = midx_entry.offset;
+		git_oid_cpy(&e->sha1, &midx_entry.sha1);
+		e->p = git_vector_get(&backend->midx_packs, midx_entry.pack_index);
+		return 0;
+	}
+
 	if (last_found &&
 		git_pack_entry_find(e, last_found, oid, GIT_OID_HEXSZ) == 0)
 		return 0;
 
-	for (i = 0; i < backend->packs.length; ++i) {
-		struct git_pack_file *p;
-
-		p = git_vector_get(&backend->packs, i);
+	git_vector_foreach(&backend->packs, i, p) {
 		if (p == last_found)
 			continue;
 
@@ -253,20 +295,6 @@ static int pack_entry_find_inner(
 		}
 	}
 
-	return -1;
-}
-
-static int pack_entry_find(struct git_pack_entry *e, struct pack_backend *backend, const git_oid *oid)
-{
-	struct git_pack_file *last_found = backend->last_found;
-
-	if (backend->last_found &&
-		git_pack_entry_find(e, backend->last_found, oid, GIT_OID_HEXSZ) == 0)
-		return 0;
-
-	if (!pack_entry_find_inner(e, backend, oid, last_found))
-		return 0;
-
 	return git_odb__error_notfound(
 		"failed to find pack entry", oid, GIT_OID_HEXSZ);
 }
@@ -281,22 +309,35 @@ static int pack_entry_find_prefix(
 	size_t i;
 	git_oid found_full_oid = {{0}};
 	bool found = false;
-	struct git_pack_file *last_found = backend->last_found;
+	struct git_pack_file *last_found = backend->last_found, *p;
+	git_midx_entry midx_entry;
+
+	if (backend->midx) {
+		error = git_midx_entry_find(&midx_entry, backend->midx, short_oid, len);
+		if (error == GIT_EAMBIGUOUS)
+			return error;
+		if (!error && midx_entry.pack_index < git_vector_length(&backend->midx_packs)) {
+			e->offset = midx_entry.offset;
+			git_oid_cpy(&e->sha1, &midx_entry.sha1);
+			e->p = git_vector_get(&backend->midx_packs, midx_entry.pack_index);
+			git_oid_cpy(&found_full_oid, &e->sha1);
+			found = true;
+		}
+	}
 
 	if (last_found) {
 		error = git_pack_entry_find(e, last_found, short_oid, len);
 		if (error == GIT_EAMBIGUOUS)
 			return error;
 		if (!error) {
+			if (found && git_oid_cmp(&e->sha1, &found_full_oid))
+				return git_odb__error_ambiguous("found multiple pack entries");
 			git_oid_cpy(&found_full_oid, &e->sha1);
 			found = true;
 		}
 	}
 
-	for (i = 0; i < backend->packs.length; ++i) {
-		struct git_pack_file *p;
-
-		p = git_vector_get(&backend->packs, i);
+	git_vector_foreach(&backend->packs, i, p) {
 		if (p == last_found)
 			continue;
 
@@ -319,6 +360,141 @@ static int pack_entry_find_prefix(
 		return 0;
 }
 
+/***********************************************************
+ *
+ * MULTI-PACK-INDEX SUPPORT
+ *
+ * Functions needed to support the multi-pack-index.
+ *
+ ***********************************************************/
+
+/*
+ * Remove the multi-pack-index, and move all midx_packs to packs.
+ */
+static int remove_multi_pack_index(struct pack_backend *backend)
+{
+	size_t i, j = git_vector_length(&backend->packs);
+	struct pack_backend *p;
+	int error = git_vector_size_hint(
+			&backend->packs,
+			j + git_vector_length(&backend->midx_packs));
+	if (error < 0)
+		return error;
+
+	git_vector_foreach(&backend->midx_packs, i, p)
+		git_vector_set(NULL, &backend->packs, j++, p);
+	git_vector_clear(&backend->midx_packs);
+
+	git_midx_free(backend->midx);
+	backend->midx = NULL;
+
+	return 0;
+}
+
+/*
+ * Loads a single .pack file referred to by the multi-pack-index. These must
+ * match the order in which they are declared in the multi-pack-index file,
+ * since these files are referred to by their index.
+ */
+static int process_multi_pack_index_pack(
+		struct pack_backend *backend,
+		size_t i,
+		const char *packfile_name)
+{
+	int error;
+	size_t cmp_len = strlen(packfile_name);
+	struct git_pack_file *pack;
+	size_t found_position;
+	git_buf pack_path = GIT_BUF_INIT, index_prefix = GIT_BUF_INIT;
+
+	error = git_buf_joinpath(&pack_path, backend->pack_folder, packfile_name);
+	if (error < 0)
+		return error;
+
+	/* This is ensured by midx__parse_packfile_name() */
+	if (cmp_len <= strlen(".idx") || git__suffixcmp(git_buf_cstr(&pack_path), ".idx") != 0)
+		return git_odb__error_notfound("midx file contained a non-index", NULL, 0);
+
+	cmp_len -= strlen(".idx");
+	git_buf_attach_notowned(&index_prefix, git_buf_cstr(&pack_path), cmp_len);
+
+	if (git_vector_search2(&found_position, &backend->packs, packfile_byname_search_cmp, &index_prefix) == 0) {
+		/* Pack was found in the packs list. Moving it to the midx_packs list. */
+		git_buf_dispose(&pack_path);
+		git_vector_set(NULL, &backend->midx_packs, i, git_vector_get(&backend->packs, found_position));
+		git_vector_remove(&backend->packs, found_position);
+		return 0;
+	}
+
+	/* Pack was not found. Allocate a new one. */
+	error = git_mwindow_get_pack(&pack, git_buf_cstr(&pack_path));
+	git_buf_dispose(&pack_path);
+	if (error < 0)
+		return error;
+
+	git_vector_set(NULL, &backend->midx_packs, i, pack);
+	return 0;
+}
+
+/*
+ * Reads the multi-pack-index. If this fails for whatever reason, the
+ * multi-pack-index object is freed, and all the packfiles that are related to
+ * it are moved to the unindexed packfiles vector.
+ */
+static int refresh_multi_pack_index(struct pack_backend *backend)
+{
+	int error;
+	git_buf midx_path = GIT_BUF_INIT;
+	const char *packfile_name;
+	size_t i;
+
+	error = git_buf_joinpath(&midx_path, backend->pack_folder, "multi-pack-index");
+	if (error < 0)
+		return error;
+
+	/*
+	 * Check whether the multi-pack-index has changed. If it has, close any
+	 * old multi-pack-index and move all the packfiles to the unindexed
+	 * packs. This is done to prevent losing any open packfiles in case
+	 * refreshing the new multi-pack-index fails, or the file is deleted.
+	 */
+	if (backend->midx) {
+		if (!git_midx_needs_refresh(backend->midx, git_buf_cstr(&midx_path))) {
+			git_buf_dispose(&midx_path);
+			return 0;
+		}
+		error = remove_multi_pack_index(backend);
+		if (error < 0) {
+			git_buf_dispose(&midx_path);
+			return error;
+		}
+	}
+
+	error = git_midx_open(&backend->midx, git_buf_cstr(&midx_path));
+	git_buf_dispose(&midx_path);
+	if (error < 0)
+		return error;
+
+	git_vector_resize_to(&backend->midx_packs, git_vector_length(&backend->midx->packfile_names));
+
+	git_vector_foreach(&backend->midx->packfile_names, i, packfile_name) {
+		error = process_multi_pack_index_pack(backend, i, packfile_name);
+		if (error < 0) {
+			/*
+			 * Something failed during reading multi-pack-index.
+			 * Restore the state of backend as if the
+			 * multi-pack-index was never there, and move all
+			 * packfiles that have been processed so far to the
+			 * unindexed packs.
+			 */
+			git_vector_resize_to(&backend->midx_packs, i);
+			remove_multi_pack_index(backend);
+			return error;
+		}
+	}
+
+	return 0;
+}
 
 /***********************************************************
  *
@@ -340,9 +516,16 @@ static int pack_backend__refresh(git_odb_backend *backend_)
 	if (p_stat(backend->pack_folder, &st) < 0 || !S_ISDIR(st.st_mode))
 		return git_odb__error_notfound("failed to refresh packfiles", NULL, 0);
 
-	git_buf_sets(&path, backend->pack_folder);
+	if (refresh_multi_pack_index(backend) < 0) {
+		/*
+		 * It is okay if this fails. We will just not use the
+		 * multi-pack-index in this case.
+		 */
+		git_error_clear();
+	}
 
 	/* reload all packs */
+	git_buf_sets(&path, backend->pack_folder);
 	error = git_path_direach(&path, 0, packfile_load__cb, backend);
 
 	git_buf_dispose(&path);
@@ -478,9 +661,11 @@ static int pack_backend__foreach(git_odb_backend *_backend, git_odb_foreach_cb c
 	backend = (struct pack_backend *)_backend;
 
 	/* Make sure we know about the packfiles */
-	if ((error = pack_backend__refresh(_backend)) < 0)
+	if ((error = pack_backend__refresh(_backend)) != 0)
 		return error;
 
+	if (backend->midx && (error = git_midx_foreach_entry(backend->midx, cb, data)) != 0)
+		return error;
 	git_vector_foreach(&backend->packs, i, p) {
 		if ((error = git_pack_foreach_entry(p, cb, data)) != 0)
 			return error;
@@ -562,6 +747,7 @@ static int pack_backend__writepack(struct git_odb_writepack **out,
 static void pack_backend__free(git_odb_backend *_backend)
 {
 	struct pack_backend *backend;
+	struct git_pack_file *p;
 	size_t i;
 
 	if (!_backend)
@@ -569,11 +755,13 @@ static void pack_backend__free(git_odb_backend *_backend)
 
 	backend = (struct pack_backend *)_backend;
 
-	for (i = 0; i < backend->packs.length; ++i) {
-		struct git_pack_file *p = git_vector_get(&backend->packs, i);
+	git_vector_foreach(&backend->midx_packs, i, p)
+		git_mwindow_put_pack(p);
+	git_vector_foreach(&backend->packs, i, p)
 		git_mwindow_put_pack(p);
-	}
 
+	git_midx_free(backend->midx);
+	git_vector_free(&backend->midx_packs);
 	git_vector_free(&backend->packs);
 	git__free(backend->pack_folder);
 	git__free(backend);
@@ -584,7 +772,12 @@ static int pack_backend__alloc(struct pack_backend **out, size_t initial_size)
 	struct pack_backend *backend = git__calloc(1, sizeof(struct pack_backend));
 	GIT_ERROR_CHECK_ALLOC(backend);
 
+	if (git_vector_init(&backend->midx_packs, 0, NULL) < 0) {
+		git__free(backend);
+		return -1;
+	}
 	if (git_vector_init(&backend->packs, initial_size, packfile_sort__cb) < 0) {
+		git_vector_free(&backend->midx_packs);
 		git__free(backend);
 		return -1;
 	}

src/pack.c

@@ -481,6 +481,9 @@ int git_packfile_resolve_header(
 	off64_t base_offset;
 	int error;
 
+	if (p->mwf.fd == -1 && (error = packfile_open(p)) < 0)
+		return error;
+
 	error = git_packfile_unpack_header(&size, &type, &p->mwf, &w_curs, &curpos);
 	if (error < 0)
 		return error;
@@ -631,6 +634,9 @@ int git_packfile_unpack(
 	size_t stack_size = 0, elem_pos, alloclen;
 	git_object_t base_type;
 
+	if (p->mwf.fd == -1 && (error = packfile_open(p)) < 0)
+		return error;
+
 	/*
 	 * TODO: optionally check the CRC on the packfile
 	 */

tests/pack/midx.c

@@ -27,3 +27,19 @@ void test_pack_midx__parse(void)
 	git_repository_free(repo);
 	git_buf_dispose(&midx_path);
 }
+
+void test_pack_midx__lookup(void)
+{
+	git_repository *repo;
+	git_commit *commit;
+	git_oid id;
+
+	cl_git_pass(git_repository_open(&repo, cl_fixture("testrepo.git")));
+
+	cl_git_pass(git_oid_fromstr(&id, "5001298e0c09ad9c34e4249bc5801c75e9754fa5"));
+	cl_git_pass(git_commit_lookup_prefix(&commit, repo, &id, GIT_OID_HEXSZ));
+	cl_assert_equal_s(git_commit_message(commit), "packed commit one\n");
+
+	git_commit_free(commit);
+	git_repository_free(repo);
+}